Vous êtes sur la page 1sur 552

Readings

Volume

in

Mathematical

Psychology

Readings

Volume

in

New

York

and

London

\
Mathematical

EDITED

R.

Psychology

BY

Duncan

Luce,

Bush,

Robert

R.

Eugene

Galanter,

John

Wiley

and

University of Pennsylvania

University of Pennsylvania

Universityof Washington

Sons,

Inc.

"

Copyright

All

1963

by

John

Wiley

"

Sons,

Inc.

rights reserved.

This

book

must

not

without

or

be
the

Library

of

Printed

in

any

part

reproduced
written

thereof
in

any

the

United

of

permission

Congress Catalog
States

form

Card
of

the

publisher.

Number:

America

63-14066

Preface

"

The

first,

of

volumes

two

designed

are

Mathematical

as

Readings
The

Psychology.

references

that

articles

suggestions

and

limitations

our

to

appearing

authors

the

evaluations,

own

the

asked

were

Readings
took

we

which

this

three-volume

particularlyimportant
in

of

Psychology,

accompany

Handbook

considered

they

the

Mathematical

in

materials

source

in

to

their

fields;

considerable

from

Because

liberty in

of

journal

suggest

selected.

were

is the

Handbook

of

these

space

the

selection

and

learning.

process.
focuses

volume

This
Part

consists

and

Part

II

of

topics.

These

Volume

II of the

and

mathematical
after

the

Of

the

from

35

I of

and

of

papers

National

from

of Symbolic Logic,
of

the

papers

The

35

interest

to

U.S.

The

volume,

Journal

11

time,

statistical

the

Handbook.

chapters.
as

Decision

esses
Proc-

Theory (Stanford,
It

is

view

our

1959),

that

They

of

Transactions

of

are

every
listed

are

Psychometrika,

PacificJournal

the

10

are

3 from

the

of Mathematics,

Biophysics, the Proceedings of

the

Radio

the

the

of

Institute

Mathematical
is

from

of Experimental Psychology,
2 from

Gratitude

Force.

of

Handbook

his bookshelf.

on

Mathematical

Annals

the

8-10

and

Statistics, and

expressed

for

Engineers,
private

permissions

ment
docu-

reproduce

to

here.

represent

papers
note

statisticians,

published

Air

this

of

of Sciences,

and

other

Readings.

books

such

the

Bulletin

mathematical

Learning

present

reaction

Handbook.

in

3 from

and

publications, such

Society of America,

the

Academy

Journal

these

the Acoustical

each

one

the

reproduced

Psychological Review,

Journal

the
have

1-6
to

Mathematical

in

from

Volume

preface to

relevant

hard-cover

psychologist should

related

Chapters

papers

Studies

intentionallyexcluded

were

in

in

appeared

(Wiley, 1954)

or

referenced

psychophysics

psychophysics,

learning and

on

contains

Readings

have

psychology:

measurement,

papers

are

papers

that

Papers

21

of

areas

on

papers

of

consists

main

two

on

14

that
3

in 1947,

and

compilation
For

of

17

are

of

the

these

are

engineers,
the
a

handling

others
book
this

of

work

are

of
and

30

different

professional psychologists,8
and

are

rather
this
other

philosophers.

uniformly spread

sort

requires

details, the

It

contributors.
are

of

One
over

the

wish

to

1950-1962.

of

thank

spondence.
corre-

Miss

Katz.
R.

Luce

Duncan

Philadelphia,Pennsylvania

Robert

R.

March,

Eugene

Galanter

1963

of

was

papers

the years

surprising amount

editors

be

may

mathematicians

Bush

Ada

Contents

PART

MEASUREMENT,

An

TIME

Formulation

Axiomatic

Intervals

Successive

by

Ernest

Decision

Adams

and

and
Psychoacoustics

by

David

Some

M.

Generalization

Time

3
Messick

Relations

Duncan

Detection

Luce

Theory

and

by

M.

Correction

R.

Duncan

William

Random

by

67

Laws
Psychophysical

69

Luce

Multivariate

by

of
Theory''

Green

the Possible

On

41

Green

Comments

David

in
17

R.

and Detection
'''Psychoacoustics

by

of

Behavior

S. Christie

Lee

Samuel

and

Structure

and

Scaling

and

Simple Choice
by

AND

PSYCHOPHYSICS,

REACTION

InformationTransmission

84

J. McGill

Fluctuations

William

of Response

104

Rate

J. McGill

to Changes in the Intensity


Sensitivity
of White
and
Loudness
to Masking

Noise

and

Its Relation

by George

A.

Miller

The

Magical

Number

Some

Limits

by George
Remarks
I. The

Least

Frederick

on

Our

or

Minus

Two:

Capacityfor ProcessingInformation

135

Miller

the Method

on

Deviations

by

A.

Seven, Plus

119

Squares

and

of Paired
Solution

Assuming Equal Standard

Equal Correlations

Mosteller

Comparisons:
152

CONTENTS

Vlll

Theoretical

Some

Relationships
among

Measures

of

Conditioning
by Conrad G. Mueller
The

by

159

Theoryof SignalDetectability
W,

and
Peterson, T. G. Birdsall,

W.

167

W.

C. Fox

Aspectsof Theories of Measurement


Scott and Patrick Suppes

Foundational

by

Dana

for Choice-Reaction

Models

by Mervyn

PART

II

by T.

AND

about
Inference

Stochastic

by

STOCHASTIC

Anderson

W.

R. J.

and

Model

by

Robert

Model

R. Bush

Two-Choice

by

R. Bush

K.

Choice

Behavior

263

K.

for SimpleLearning
Frederick

Frederick

Thurlow

278

Mosteller

and

Discrimination

Fish

300

R. Wilson

TheoryofLearning

308

Estes

Theoryof SpontaneousRecoveryand Regression

Estes and

332

in Learning
Variability

C. J. Burke

Verbal

Situation
Conditioning

in Terms

of

Statistical LearningTheory

An

by
A

K. Estes and

J. H.

Investigation
of Some
Curt

F.

by

343

Straughan
Mathematical

Models

for Learning

353

Fey

Functional
Laveen

322

Estes

Analysis
of a
by W.

289

Mosteller

ofParadise

and

Theoryof Stimulus

W.

241

A. Goodman

Generalization

and

Statistical

K.

Statistical

by W.

and

Behavior

Robert

W.

Chains

for Individual

for Stimulus

Toward

PROCESSES

Markov

Leo

Model

R. Bush

by Robert

by

228

Audley

Mathematical

Time

Stone

LEARNING

Statistical

212

EquationAnalysis
of Two

Kanal

LearningModels

360

CONTENTS

The

Distribution

Asymptotic

Samuel

Walks

Learning Theory
by John Lamperti and
Markov

A.

by George

Statistical
A.

by George
Ultimate

by

Miller

Patrick

in

Frank

Role

Suppes
429

Psychology

of the

Shannon-

Wiener
448

William

and

and

Miller

between
a

G.

Two

Madow

Verbal

William

470

Learning

J. McGill

Attractive

Goals:
498

Model

Mosteller

and

Maurice

Tatsuoka

515

Learning

Restle

of Observing Responses

Benjamin Wyckoff,

in

Discrimination

Learning,
524

404

Applicationto

Estimate

Theory of Discrimination

L.

Model

413

Descriptionof

from

Frederick

Part

Their

Likelihood

Choice

Predictions

The

Learning

of Information

Measure

by

Beta

Miller

the Maximum

381

Suppes

and

Processes

A.

by George

by

Patrick

Order
of Infinite

Finite

Models

Karlin

Lamperti and

John

Chains

Arising in Learning

Asymptotic Propertiesof Luce's

Some

On

Two-Absorbing-Barrier

Kanal

Random

Some

by

the

376

Laveen

by

for

Model

Beta

by

IX

Jr.

Part

PSYCHOPHYSICS,
MEASUREMENT,

REACTION
AND

TIME

AXIOMATIC

AN

SCALING*

INTERVALS

SUCCESSIVE

OF

GENERALIZATION

AND

FORMULATION

Adams

Ernest
university

of

berkeley

california,
AND

Samuel
educational

Messick
service

testing

A formal set of axioms is presented for the method


of successive intervals,
of the scaling assumptions are
derived.
directlytestable consequences
by a systematic modification of basic axioms the scalingmodel is generalized
stimulus distributions of both specifiedand unspecified
to non-normal

and
Then

form.

scaling models

Thurstone's

comparisons [17,24] have

been

of successive

intervals

[7, 21] and

severelycriticized because

paired

of their

dependence
untestable
of
This
assumption
apparently
normahty.
objection
upon
was
recently summarized
by Stevens [22],who insisted that the procedure
of a psychologicalmeasure
of using the variability
to equalize scale units
of a kind of magic
trick for climbing the hierarchy of scales.
"smacks
a rope
in this case
The
is the assumption that in the sample of individuals
rope
in
tested the trait
question has a canonical distribution,(e.g.,'normal')
There
those who
believe that the psychologistswho make
are
tions
assump."
whose
validityis beyond test are hoist with their own
petard
Luce [13]has also viewed these models as part of an "extensive and unsightly
have
literature which
has been largelyignored by outsiders,who
correctly
condemned
the ad hoc nature
of the assumptions."
Gulliksen [11],on the other hand, has explicitly
discussed the testability
of these models
and
has suggested alternative procedures for handling data
an

"

"

"

"

"

which
also

do

not

satisfy the checks. Empirical

mentioned

implied

or

in

other

several

Random

deviations
*This

paper

from

was

M'hile the

authors

were

were

[e.g.,

presented [8, 18],

satisfactory scaling within

scaling assumptions, are

written

been

methods

well

sampling fluctuations,as

and

errors

fit have

scaling theory

of the

accounts

8, 9, 12, 15, 21, 25]. Criteria of goodness of


which, if met
by the data, would indicate

acceptable error.

of the

tests

"

"

thereby evaluated
attending

the

1957

as

by

an

tematic
sys-

these

Social Science

in Social Science.
Applications of Mathematics
171-034
The
research
was
supported in part by Stanford University under Contract NR
with
Research, by Social Science Research
Group Psychology Branch, Office of Naval
Council, and by Educational
Testing Service. The authors wish to thank Dr. Patrick Suppes
for his interest and
throughout the writing of the report and Dr. Harold
encouragement
the manuscript.
Gulliksen for his helpful and instructive
comments
on
Research

This

Council

article

Summer

appeared

Institute

in

on

Psychomelrika, 1958, 23,


3

355-368.

Reprinted

with

permission.

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

over-all internal

checks. However, tests of the scaling


consistency
assumptions,
been
and in particularthe normality hypothesis,have
not
yet
explicitly
derived in terms of the necessary
and sufficientconditions required to satisfy
the model. Recently Rozeboom
and Jones
[20] and Hosteller [16] have
of
successive
intervals
and pairedcomparisons,
t
he
investigated
sensitivity
that departuresfrom
to a normality requirement,indicating
respectively,
normality in the data are not too disruptiveof scale values with respect to
of the assumptions of the
goodness of fit,but direct empiricalconsequences
not specified
model were
such.
as
The present axiomatic characterization of a well-established scaling
model
was
(a) an
attempted because of certain advantages which might accrue:
of
that
follows from a preciseknowledge of formal propease
generalization
erties
in
and
parisons
by systematically
modifying axioms,
(6)an ease
making comof different models. The next section deals
between the properties
with the axioms for successive intervals and serves
as the basis for the ensuing
i
n
stimulus
to non-normal
tions.
distribusection, which the model is generalized
which
One outcome
of the followingformalization
should again be
sequences
highlightedis that the assumption of normality has directlyverifiable conand should not be characterized as an untestable supposition.

Thurstone's Successive Intervals

ExperimentalMethod

The

In
set of
to

ScalingModel

some

of successive

method

the

stimuli and asked


attribute.

placedin category

The

intervals subjects are

to sort them

into k ordered

proportionof times f,i that

i is determined

from

the responses.

presented with a
with respect
categories
a
given stimulus s is
If it is assumed

category actuallyrepresents a certain interval of stimulus values for


then

the relative

frequency with

category should
the stimulus
This

value

to

probabilityis in

which

represent the
lie within

turn

that
probability

the interval

simply the

area

the interval. So far scale values for the end


but
areas

if the observed
under

for
probabilities

normal

category boundaries
Scale
and

values

interval widths

curve,

and
for

the

given stimulus

is

the

pointsof the intervals

given

then scale values

may

ticular
par-

subject estimates
the

under the distribution

are

subject,

placed in

correspondingto

stimulus

that

taken

be obtained

category.

curve

are

to

inside

unknown,
represent

for both

the

stimulus.

interval boundaries
not assumed

are

determined

in the method

by
of

this

model,

equalappearing
obtainingsuccessive intervals
scale values have been presentedby Saffir [21],Guilford [10],Hosier
[15],
Attneave
Burros
Garner
and
Edwards
Hake
Bishop [3],
[5],and
[2],
[7],
[9],
Rimoldi [19].The basic rationale of the method
outlined
had been previously
by Thurstone in his absolute scalingof educational tests [23,26].Gulliksen
are

equal,

as

intervals. Essentially
equivalentprocedures for

ADAMS

ERNEST

MESSICK

SAMUEL

AND

[6],and Bock [4]have described least


and
and
Rozeboom
Jones [20]
solutions for successive intervals,
square
presenteda derivation for scale values which utilized weights to minimize
sampling errors. Most of these papers contain the notion that the assumption
than one
stimulus. Although
of normality can be checked by considering
more
distribution of relative frequencies
can
always be converted to a normal
one
it is by no
means
always possibleto normalize simultaneouslyall
curve,
and variances,on
of the stimulus
distributions,
allowing unequal means
of exact conditions under which this is
the same
base line. The specification
the problem of sampling
be attempted.In all that follows,
possiblewill now
fluctuations is largelyignored,and the model is presentedfor the errorless
Tucker

[12],Diederich,Messick, and

case.

The

Model

Formal

set of

The

limit upon

testingthe
s

in

which

from

admissible

each

have

category i

stimulus

stimuli,although

"

"

it will
specifically,

placed in category i
product of "S X {1,2,

be

the

that

members.

"

"

There

"

is no

for the purpose


of
For each stimulus

k, the relative frequency/","with


is given.Formally / is a function

"

is

the Cartesian

r, s, u, v,

least two

at

1, 2,

elements

of

number

S must

model,

S, and

S, has

denoted
stimuli,

the

"

"

k}

"

More

in

S, f^ will be a probability
the set {1,2,
distribution over
statement
k}. For the sake of an explicit
of the assumptions of the model, this fact will appear
as
an
axiom, although
it must
be satisfiedby virtue of the method
of determiningthe values of /^,,case

"

"

for each

to the real numbers.

"

A;}into the real numbers


/ is a function mapping *S X {1,
in
is
d
istribution
s
over
k];'i.e.,
{1,
S, f^ a probability
each sin. S and i
k, 0 " j,^i" 1 and Xli=i
f^.i 11,
The
the function
set S and
/ constitute the ohservables of the model.
which
remain to be introduced.
not directlyobserved
more
are
concepts
Axiom

1.

"

"

"

such that for each

"

"

"

for

"

"

"

Two
The

first of these is

of the

ti

"

"

t^k-l) which

"

are

corresponding to the categories.It is


the entire
adjacent and that they cover

simply be assumed

that

^i

"

"

^(^-1)are

"

the end

are

intervals

intervals
it will

set of numbers

assumed

that

real line.

points
these

Formally,

series of real
increasing

an

numbers.
Axiom
i

,{k
Finally,the
2,

"""

by
Axiom
over

2. Interval

boundaries

ti

"

"

^(i-d are

"

1), ^(.-1)"U.
correspondingto

distribution

normal
3. iV is

distribution function
a

function

mappmg

each

real

stimulus

numbers, and
s

for

*S is represented

A'^,
.

S into normal

distribution functions

the real line.


Axioms

1-3

do not

state

fullythe mathematical

properties
requiredfor

READINGS

IN

the set

ti
S, the numbers
of completeness,these will
"

MATHEMATICAL

formal

stated in the

be

0. S is

In the interests

followingAxiom

should be referred to instead of Axioms

purposes

Axiom

t(^k-i)and the functions N,

"

PSYCHOLOGY

for

0, which

1-3.

set. A;is

a
positiveinteger.
/ is a function
k] into the closed interval [0,1],such that for each s
mapping S X {1,
i^ S) ^i=i fs.i
1- For i
{k
1),ti is a real number, and for
I,
"
i
N
is
a function
(k
ti+i
2), ti
1,
mapping S into the set of
"

non-empty

"

"

"

"

"

"

"

"

"

"

distribution functions

normal

Axioms

2 and

3 state

only the set-theoretical character of the elements

ti and

N,

of the

theory states the connection

and

have

/,.," and the assumed


Axiom

4.

intuitive

no

Axioms

between

1-4 state the formal

hypothesis
relative frequencies

in *Saudi

=!,"""

k,

f' NXcc) da.

1, ta-D is set equal to

central

the observed

(Fundamental hypothesis)For each


=

if i

empiricalcontent. The

underlyingdistributions A^,

/,.,"

(Note that

the real numbers.

over

oo, and

"

ii i

k, ti

oo.)

assumptionsof the theory although,because

the fundamental
ti , it is not
wiU

hypothesis(Axiom 4) involves the unobservables N, and


testable in these terms. The questionof testing
the model
directly

be discussed in the next

yet been introduced. These


N,

and

hence

are

section. Scale values for the stimuli have


defined to be

are

equal to the

easilyderived. The function

means
v

not

of the distributions

will represent the

scale values of the stimuli.


Definition
that for each

1.

in

is the function

S, v, is the
V,

mapping S into the real numbers


of N, ; i.e..

mean

such

aNXoi) da.

Testingthe Model
The

model

of Axioms

1-4

will be said to fit exactlyif all of the testable consequences


will be
of these axioms
verified. Testable consequences
are

those consequences

concepts S and
If

which

/, or

are

formulated

of concepts which

further

about

solelyin
are

terms

of the observable

definable in terms

of S and

independent determination

/.
of

an
assumptions are made
then
the
and
testable
are
just those which
ta-i)
N,
consequences
follow about
the assumption that there exist numbers
/ and *S from
1-4. In this model,
^1
t(k-i)and functions N^ which satisfyAxioms
it is possibleto give an exhaustive description
of the testable consequences;
hence this theory is axiomatizable in the sense
that it is possible
to formulate

ti

no

"

"

"

"

"

"

observable conditions which

are

necessary

and sufficientto insure the existence

For all
i

be linear functions of each other in the

(7),is that all 3, ."

and

I,

r
"""

and

in

S,

there exist real numbers

2,..-

requirednumbers

a,

Zr
-^

ar.sZr.i

and "r

,,

(9)

6^,,such that for each

independent of

i and

"

exist if and

"

only if for each

and s, the ratio

i
=

a,,r

Gr,,

Z,j

"

"r."

."

Zr

"

Zs,i

is

a^., and

followingsense.

,k,

(8)
The

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

j.

(8) exist,then they are related to


a^., and 6^,.satisfying
the scale values v,. and the standard deviations o-^ in a simpleway. For each
If constants

r,

in

S,

(10)

ttr.s

(^r/(rs
,

and

"r."

(11)

(i^r y.)/0-a
-

"

Clearlythe arbitrarychoice of the constants Vr and o-^ in (5),(6),and (7)


represents the arbitrary choice of originand unit in the scale. Since scale
once
values of U and v, are uniquely determined
chosen,the
v^ and
o-^ are
interval scale
an
scale values are unique up to a linear transformation;
i.e.,
of measurement
does

not

has

been

determined.

requireequalityof standard

It should

deviations

be

noted

(or

what

that

this model

Thurstone

has

dispersions[25]) but provides for their determination


in its possible
from the data by equation (6).This adds powerfulflexibility
applications.
remark
about
the necessary
and sufficient
It remains only to make
a
fulfill in
condition which a set of observed relative frequencies/,."must
and sufficient condition is simply
order to satisfythe model. This necessary
called discriminal

that

the

numbers

z,,i

which

are

defined in terms

of the observed

relative

be determined
related as expressedin (8),This can
be linearly
frequencies,
by seeing if the ratios computed from (9) are independent of i and j,or by
of the plotsof z,,i againstZr,i
Hence
evaluatingfor aU s, r the linearity
for this model there is a simple decision procedurefor determiningwhether
not a given set of errorless data fits.
or
related for all s, r in S, the assumpIf z, ,i and Zr ," are found to be linearly
tions
of the scalingmodel are verified for that data. If the z's are not linearly
then assumptionshave been violated. For example,the normal
curve
related,
other
distribution function for the stimuli and some
not be an appropriate
may
function might yielda better fit [cf.11, 12].Or perhaps the responses cannot
be summarized
unidimensionallyin terms of projectionson the real line
representingthe attribute [11].If the stimuli are actuallydistributed in a
.

ERNEST

multidimensional
be

may

does not

This

if the

that
A

satisfactorily
by

model

does

multidimensional

such

never

be

in several

varying

dimensions

may

rather

might be operating.
more
appropriate in

effects

[14] might

sions.
dimen-

but
intervals,

of successive

distortion

scaling model

practicethe

prove

points (2;^,,(k
1) will
Zs,i)ior i
2,
exactly fit the straightline of (8) but will fluctuate about it. It remains
this fluctuation
decided whether
represents systematic departure

the model

from

or

set

still be

is not

size of the

In

"

"

"

"

obtained

the

only by eye.
by the method

minimum

of

absence

precise,although the

if

straightline

variance.

error

evaluated, even

pointsto

of

the decision
linearity,

the

stimuli

fit,such

not

of variations in other

the presence

the method

of the attributes

one

cases.

In

to

that

mean

MESSICK

on
judgments of projections

then

space,

SAMUEL

AND

distorted by
differentially

be scaled

not

ADAMS

One

statistical test

linearityof the
and

of least squares

[4,6, 12].In

error

plots may

is to fit the

approach

then

obtained
evaluate

event, the

any

for

test

of

of a suitable
in the errorless case, and the incorporation
sampling theory would
provide decision criteria for direct experimental

the

model

is exact

applications.
Generalization of the Successive

be

generalizedin

of ways.

number

previous section can


treated in detail
generalization,
in the

discussed

model

successive mtervals

The

Intervals Model

One

of a
by Torgerson [27],considers each interval boundary f,-to be the mean
approach toward
subjective distribution with positivevariance. Another
distributions
the requirement of normal
generalizingthe model is to weaken
amounts
to enlarging
of stimulus scale values. Formally, this generalization
of specifying
the class of admissible distribution functions. Instead
exactly
which

distribution

functions

an
assume
generalization,
real line,to which it is requiredthat

allowed

are

in the

arbitraryset \pof distributions over the


the stimulus distributions belong.In formalizingthe model, \pis characterized
be
3 may
the real line. Axiom
simply as a set of distribution functions over
of the class 4/ and stating
the nature
axiom specifying
replaced by a new
for each s in S, C,
of \p;i.e.,
that C is a function mapping S into elements
the distribution

(interpretedas
final

One

if \p contains

formations
defined
C by

stretch
on

as

of

assumption

other

any

shift of

along

about

the

distribution function
C. A

linear

C,

class \{/needs
then

transformation

distribution function

originand

the horizontal

of

be

function.

form. Let D and D'

added:

be

to

namely,

distribution function

which

scale transformation
axis must

of \l/.

member

contain all linear trajis-

it must

be

can

of the

compensated

the vertical axis in order that the transformed

density

s) is

of the stimulus

function

Algebraically,these transformations
then D' is a
be distribution functions,

obtamed

horizontal

for

by

also be
have

is

from
axis. A

contraction

a
a

probability
following

the

linear transformation

10

of D

if there exists

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

positivereal number

and

real number

h such that

for all X,

D'{x)

This is not truly a linear transformation


for lack of

h).

aD{ax +

of

because

multiplication
by a on the
is
for
used.
The
reason
phrase

this

better term

a
ordinate,but
of distribution functions be closed under
class
the
that
^
requiring

in any
possibleto convert them

it will be

values

to form

in

by

of stimulus

scale

linear transformation

into

the stimulus values obtained


values;i.e.,

set of scale

admissible

another

determination

that

insure

is to

transformations

linear

are

interval scale. If the set ^ is not closed under linear transformations,

an

will not

generalit

be

possibleto alter the scale by

an

arbitrarylineaf

transformation.
3'. ^ is

Axiom

C is a function
and

b is

set of distribution functions

S into \p.For all D

mapping

real number, then the function D' such

D'{x)
is a member

aD(ax +

the real

over

numbers, and
positivereal number

in ^, if a is a

that for all x,

6)

of ^.
that the set of normal

It is to be observed

distributions has the

required

property of being closed under linear transformations. This set is in fact a


that aU normal distribution functions
minimal class of this type, in the sense
can

be

generatedfrom

the

points ti

4'. For each

Axiom

in *S and

before to be the

for each

1'.
s

in

and

f*

0"

1,

"

S,

is the function
v, is

the

problem

\p.Each

now

is to

"

k,

stimulus

.) The

mean

C,

are

defined

as

such

specifythe

distributions. If the

values

mapping S into the real numbers


of C,
i.e..
dx.

class of admissible distribution functions

of this class amounts


specification

stimulus

C,

C.ix) dx.

=/:xC,{^)
The

"

of the distribution functions

means

Definition
that

""

"

/",." the distribution functions

formations.
trans-

f,.i
(Here again ^o

by linear

which specifies
generalization

obvious

an

the observed

between

interval end

distribution function

singlenormal

4 is replacedby

Axiom
Finally,
the connection
and

to

theory about

the underlying

hypothesis of normality is altered

or

ADAMS

ERNEST

weakened,

what

of all distribution functions


is made

of C,

C.{x)

Non-normal
It is
be

to

of f,-can

scale values

1 "

Every

set of

be determined

only

members

transformations,has

the
the

when

restrictions

to

linear transformation.

fits. For

model

minimal

class of distribution functions.


1. There

exists

an

h, such

that

in

S,

k,

It will next
of

sense

that

assume

\p is

such that for all


and

exists

h).
scale values

3' and

positivereal number

obtained

are

Assumption
a, and

imply

real number

for all x,

CXx)
the function

(13)

on

the

/.,."
=

is the cumulative

distributions p,

(14)

follows. Axiom

aMa.x

,"

are

defined

p,,i

as

1,

aMa.x

distribution

(12) is

functions

the
=

5J,

+
of

rightside

form
linearlyrelated to all
specified
Axiom
4',then, for each s in S, and 2

present

1 is satisfied the

proceed as

there

(12)
where

positivereal number

aD{ax +

Assumption

interval scale,we

that for all

"

that for all x,

if

that

on

distribution function D

D'{x)
show

the

D' in \}/there exists

distribution functions

To

"

are

values

h such

"

if the scale values


i/-

some

stimulus

real number

1,

by linear
generatedfrom a singlemember
desired property of generatinga linear scale of

class all of whose

Assumption

class of distribution functions,in the

minimal

any

z,

Form
oj Specified

clearlynecessary to make
determined
uniquely up
that

"

a;

othenvise.

Distributions

shown

If

weak.

then the theory is very

on

no

construct

to

necessary

If

always possibleto determine distribution functions


this it is only
for arbitrarily
ti To show
specified
with the followingdefinition.
them
in accordance

4'

Axiom
C, satisfying

be

amount

ordinal scale. It is

an

are

would

real numbers.

over

theory, and the

will fit the

data
on

the forms

about

assumption about
to letting
\^ be the set
assumption whatever

replaceit? Omitting any

assumptions can

of the distribution functions

the form

11

MESSICK

SAMUEL

AND

"

"

"

D'

Tr(aJ,+

5,

some

According

to

6,)dx.

correspondingto D,

a,D(a,x +

in ^.

of

,k,

before,then

fixed function

b,)dx

and

the

cumulative

12

READINGS

Assuming

that

the function

of function

form

z,

such that for each

,,"

is

is

Ps.i

(16)

z,,i

that

to assume

based

on

Vs

tt

generalbe

in

ti ,

in S and

and

a,

1,

is

"

Trfe.i).

"

k. It is clear from

"

"

be

not

in the normal

If it were
increasing.

Zs,i determined

not, there would

by (15);hence the scale values

unique. It is also seen that (4),relatingZs,i to


distribution model, is simply a particular
of
case

between

connection

0-,

(15)why it is necessary

unique

h,

aji +

monotone
strictly

z,,i would

(16)here. The

"

(15) imply immediately that

Equations (14) and

for all

"

(15)

not

strictlymonotone

increasing,
then,knowing
possibleto determine uniquely the numbers
in aS and i
1,
k,
x

Z), it

the

PSYCHOLOGY

MATHEMATICAL

IN

a^

1/a,

b, and
Vs

and

a-,

"hja,

is

v,

(15),as in the correspondingset of equations obtained from the


the left are known, and the numbers
on
normalityassumption,the numbers
As before,if two numbers
the rightare unknown.
arbitrarily
on
a^ and 6^ are
determined for a fixed stimulus r, then the ti are uniquely determined by the
followingequation.
In

(17)

U
scale values

The

the

from

for the

coefficients Zs,i

the

mean

of C^

is determined
,

(18)

Vs

Both

the

tts and

is the

cannot

h^ without

and

a^

I,

stimuli,however,

of the basic distribution D. If


as

{zr,i ".)M

mean

of

,k.

"""

then v^

D,

v^

which
,

was

mean

defined

by
(m

6,)/a,

"

(17)can be determined in terms


in
is immediately determinable

the 6, in

(19) and (20);hence


by (18).
quantities

directlydetermined

be

first specifyingthe

of z,,i a^ and 6, ,
of just these
terms
,

''"'""'''"'

(19)

a,

a.

^"\zr,"
(^'''
h).
~

(20)

hs=Zs.i-

It is clear then

that

transformation.

that

set of data

the

related.

rightin

ti and

y, are

determined

up

to

sufficient conditions
and
Furthermore, necessary
of differences in
the
that
ratios
model
are
simply
that the z's be linearly
(19)be independentof i and j;i.e.,

linear

g's on

the scale values

fit the

ERNEST

The

final generalizationto

holds, but where


it is assumed
i.e.,

the form
that the

that the

class,but

class

obtain

than

more

assumed

that

it is

still possibleto
Since
will
D

is

tt

obtain

never

This

zero.

if and

only

in S and

\i

i,j

(21)

1,

"

ps.i

Therefore

from

"

"

about

all x,

minimal

one

family

"

through
Now,

by solving (15),but
it
increasing,

numbers

{a,ti+

the

case

is

6 J,

however,
increasing;

explicitin

is made

D{x)

up
is also unknown.

the

D.
and

values. If it is

strictlymonotone

increasing in

assumption

it

distribution function

Assumption

2.

0.
it follows that

then
increasingj

(14) holds, then

it will be

the

wix)

that

case

"

ir{y)

for all r,

k,

"

only if aJi +

if and

Pr,j

the numbers

ordering on

an

scale

z^,,-

distribution it is monotone

y. If

"

tt

numbers

is

information

if TT is strictly
monotone

Now,

the function

case

function

some

2. For

Assumption

the

about

the model

test

is unknown, all of the deductions

in this

only be strictlymonotone

is

Assumption

distribution function

any

distributions all belong to


but D

the

cumulative

information

impossible to discover the

postulatedthat

in which

one

it is still possibleto

case

ordinal

generated by a
D,
(14) go through, although
of course,

generated by

be

stimulus

the

is

considered

be

generatingfunction D is not specified;


underlyingdistributions all belong to one minimal

can

function

if it is

13

MESSICK

of the

in this

Interestinglyenough,
to

SAMUEL

of the Distributions Unspecified

Forms
A

AND

ADAMS

63

p^.i

"

Qrt, +

one

can

6,

obtain

system of

h^ and i, If it is further specified


involving the constants
inequalities
a^
(as is requiredfor the conditions of the problem) that a, " 0 for all S, then
this set of inequalities
solution.
will not in general have
a
,

However,
The
a,

ti and

this set of
taken

to be

whether

and

necessary

not

or

set of data

fits the model

the data. This

stillbe determined.

for fitis that there exist numbers

sufficient condition

a,

"

to

construct

is done

in the

function

distribution function
A

6J

which

can

represent

differentiable monotone

following way.
is constructed
7r(.T)
by connecting the
T(aJi +

discrete set of

creasing
in-

points

ps,i

increasingcurve.
smooth, strictlymonotone
of stimuli,then such a curve
can
only a finite number
D
defined
is
the
function
distribution
by
Finally,
(22)

may

the system of inequalities


(21).If
0) satisfying
be
inequalitieshas a solution,then the interval boundaries
may
the ti satisfying(21).To determine
the scale values of the stimuli

h, (where

it is first necessary

with

any

D{x)=-~Tix).

If,as is usual, there is


always be constructed.

14

READINGS

Then, if the
are

determined

the

V, is

of the

mean

determined

arbitrary

h and

and

can
on

in the

constant

PSYCHOLOGY

distribution D

by (18),v,

concerned,it

MATHEMATICAL

IN

(m

be

seen

the

mean

is m,

the values

of the

v,

stimuli

6J/a, As far as the determination of


that they depend solelyon the previously
"

which

m,

determination

be regarded

can

of the

v,

as

an

additional

remaining point of discussion for this model is the determination


of the degree of uniqueness of the scale values. Finding the set of all possible
solutions to the inequalities
(21) presents, in general,extreme
difficulty.
be simply determined
is the class of what might be called
One thing that can
the universal transformationsof the solutions of the system of inequalities.
is one
A universal transformation
which, applied to a solution of any set of
another
solution
set of inequalities.
to the same
yields
inequalities,
By noting
between
the theory of the inequalities
close connection
a
(21) and a twodimensional affine geometry with a distinguished
set of horizontal and vertical
shown
of
universal transformations
for this
[1]that the class
lines,it can be
The

model

is

subset of the affine transformations.

of the interval boundaries

by

and

hence

arbitraryconstant
is also

so

universal transformations

the linear ones, and of the


The
6, also are determined
up

tf are

positiveconstant.

The

the

are

also enters

scale values
into their

a,

to

are
a

tions
multiplicalinear transformation,

(although the additional


determination).
in which, even
though there is
y,

interesting
specialcase
the scale values of the ti are determined
only a finite number of observations,
to
linear
transformation.
This
a
might be called the specialcase of equal
up
intervals,in which differences in successive tiare all the same.
If,for example,
there exist stimuli with such relations among
correspondingp's as p^,,
mine
etc.,it is possibleto deterVx,i+i
Pv,i+2
Py,i
Pz.i+i
Vz.i-^2
Vy.i+i
that successive intervals are
equal [1].
The fact that scale values obtained in this model, at least under certain
circumstances,are unique up to a linear transformation has two interesting
based
for the originalsuccessive intervals model
the noron
mality
consequences
model
If
in
errorless
the
t
the
case
original
hypothesis,(i)
fits,hen
bution
other successive intervals model which assumes
a different
no
form for the distriwill fit.
The reason
for this is that the forms of the distribution
functions
determined
cumulative
functions (or the
are
distributions)
by the values of
determined
the point ti Hence, if the ti are
p,,i lying above
up to linear
(ii)Where the normality assumption
transformation,so are the curves
p..,does not fitthe data it is theoretically
possibleto use the present generalization
There

an

to obtain

scale. Then

the deviation

of the scale values from

those obtained

normality requirement can be evaluated. This, at least in principle,


provides a second kind of goodness of fit besides the usual least squares
methods
regression
employed where the data do not exactlyfit the Thurstone

under

model.

16

READINGS

[24]

Thurstone,

L.

[25]

Thurstone,

L.

L.

[26]

Thurstone,

L.

L.

The

1927,
[27]

Manuscript

MATHEMATICAL

Psychophysical
of

law

Amer.

analysis.
comparative
of

unit

PSYCHOLOGY

judgment.
in

measurement

J.

Psychol,

1927,

Psychol.
educational

38,

Rev.,

1927,

scales.

J.

368-389.

34,
educ.

424-432.

Psychol.,

505-524.

Torgerson,
behavior.

Revised

18,

L.

IN

W.
New

received

manuscript

S.

York:

law
New

of
York

2/12/58
received

4/21/58

categorical
Univ.

judgment.
Press,

1954.

In

L.

S.

Clark

(Ed.),

Consumer

STRUCTURE

DECISION

AND

CHOICE

SIMPLE

S. Christie^

Lee

Group

The

It

Laplace

transform

attempts

to

problems

the

of

fit

further

in

considered

In

this

beings organize

the

I. Introduction.

into
such

that

describe

of

Control

Systems
of

carried

was

Laboratory,

that

laboratory

should

the

for

empirical
for

University

of

April

issued

the

authors

the

it

were

Illinois.

1954,

tions
situathesis

our

be

may

studies,
model

sible
pos-

is

must

we

not

firmly

consultants
It is

supported

the

in

distribution.

reaction-time

the

from

when

that

man
hu-

way

reflected

be

must

and, therefore,
from

It is

decisions.

speculative,
out

model

propose

decisions

derives
as

analysis

the
which

required by simple choice

organization

proposal

work

R-53

of

decision

made.

are

component

times

thinking

our

this

This

of

reaction
the

infer

Although

we

organization

an

distribution
to

paper

of
data

plex
com-

csises

evaluating

of

treating

features

provide

to

model,

decisions

collection

essential

the

on

work

is

proposals,
the

that

to

special

problem

shown

earlier

leads

Two

the

of

use

negated

model

solved.

statistical
It

leaves

which

model

decisions.

the

by

general

not

the

discussed.

is

experimental

Two

but

and

out,

time-discrete

as

unchanged.
be

model

the

processing

worked

are

The

formulated

are

elementary

difficulties

overcome

of

terms

treated

be

can

reactions.

choice

which

model

the

of

as

so

analyze

in

hypothetical

data

to

considered

is

from

decisions

reaction-time

that

Technology

of

decisions

simple

such

composes

argued

is

Electronics,

of

Institute

of

Luce^

Duncan

Laboratory,

Laboratory

structure

which

R.

and

Massachusetts

IN

BEHAVIOR*

Networks

Research

RELATIONS

TIME

under

the

to

of

revision

port
re-

contract

DA-36-039-SC-56695.
'Present

Research

Operations

address:

Maryland.
Chevy
University,
Chase,
^Present
for
address:
Center

Sciences,
This

Stanford,

article

appeared

Advanced

Office,

Study

The

in

Johns

the

Hopkins

Behavioral

California.
in

Bull.

math.

Biophysics,

permission.
17

1956, 18, 89-112.

Reprinted

with

18

READINGS

leased
has

such

on

led

to

us

to

determine

to

decide

experiments which

two

suggest

These

merit it has.

what

it is desirable
model

modify the

PSYCHOLOGY

However, the development of the model

studies.

whether

to

MATHEMATICAL

IN

better with

details
particular

little hope that the

further work

pursue

help
help

may

will also

experiments

to

to accord

believe

we

in

for
reality,

of the present model

tempt
at-

an

have

we

have

any

lasting value.
U.
a

fixed

at

type

time

known

is

that

clear

the

as

mean

reaction

unwanted

stimuli
and

response,

of

and

that the
no

the stimulus
If the

time.
and

choice

type of

and

the

subject is

sponse
re-

sented
pre-

of response

In either

time.

than

more

interaction

time

fixed

of

tingent
con-

it is

case,

tions
readilyanalyzable time distribustimulus
be simple enough so that
a

between

intervene

may

stimulus

required the corresponding time interval

stable

time is

t with

time

at

disjunctive reaction

the

the

of stimuli

is

it is necessary
the

responds

set

obtain

to

subjectreceives

simple reaction

stimulus

the

on

of

one

interval,i, between

the

is called
with

0 and

time

The

response.

that

Suppose

Times.

Reaction

will

which

distribution

stimulus
will

stimuli
be

Otherwise

two.

or

the test

the

among

second

and
a

cause

the
tortion
dis-

difficult to

very

analyze.
study of reaction times, including disjunctivereaction times,
long history in the literature of psychology (cf.Woodworth,

The
has

1938, chap. xiv). In


has
this loss
a

been

recent

evident

of interest

to

in reaction-time
related

two

this separation involved

to make
to

stimulus

this

time

with

the

when

from

If the

reason.

motor

"

from
times

We

use

stimulus
of

the
motor

same

technique has
his

lags involved

been

no

time

reaction

One

to

was

the

subject'sresponse

to

decision

was

subtracting

and

made

be

attempt

the

stimulus

same

involved.

This

unsatisfactory for the following


he is able to bring
to make
decision

considered
no

for the

time

presentation

distinguished

required to respond

subject has

readiness

First, there has been


decision (decisionlatency*)

measuring

when

attribute

may

in the total process.

decision

action

We

studies.

causes.

failure to separate the time to make

from the other time

terest
however, relativelylittle in-

years,

when
to

specified
referring
motor

parts of such

response

to

response;
a

process.

the

time

to

of

latency when

much

process

higher
timed

referring

to

S,

LEE

he

than

pitch

determined,

time

if at

make

distributions

if these

(base times) from

one

been

obtained

would

We

these

latencies

human

based

being

At

the

our

its

of

the

be

when

the

decision
that

in

that

of

elementary
latencies

in itself.

end

an

for it in

primitive

more

next

section

with

familiar
to

we

Let

Laplace Transform.
t such

that

F{t)

L{F) of the real variable

that
that

be

the

of information

Laplace transform
about

in

certain

mathematical

and

to

of

one

every
we

have

of

list of those

shall need.

be

real-valued

for ^

defined

of F.

making

concerned.

are

we

it

or

way

"

0.

by

The

function

real-valued

this

There

of

tion
func-

equation

the

(1)

e-''F{t)dt

the

tirely
en-

an

the

Laplace transform,

the

position
decom-

from

unlikely that

it is

decompose

employed usefully in

be

may

its definition

elementary propertieswhich

may

decision

model

the idea

Since

times.

to

describe

to

observed
finarlly
a

non-choice

empiricalcorrelate

no

purports

such

able
unreason-

Such

distribution

with

not

adapted,

or

primitive terms.

Laplace transform
be

is called

If,

of

were

the

out

tease

to

used,

proposal is

our

will

variable

be

more

It is with

readers

III. The
real

the

reaction

the

will

It is true

non-choice

of

might

used

which

model

study of

devoted

to

time

pure

class

to account

process

composes

heart

technique of
the

on

elementary ones.

more

the

what?

some

reaction-time

observed

mathematical

formal
be

taken

for it is
difficulties,

also

into

latencies

of the

may

base

latency distribution

method

(base times) can

decision

the

equated

extremely simple,

be

to

separation

related

as

that the

suppose

latencies

the

situation

remain.

describe
to

that

then

"

approximated by

complex character,the challenge


terms

be

cannot

"

another

or

way

found

decision

choice

measurements

resulting decision

the

however,

itself

from

in

to react

action;
disjunctive re-

decision.

functions, the

mathematical

make

to

conclude

may

were

well

be

they could

We

all, only

has

latency distribution

time

decision

that in

suppose

required

the

"

19

LUCE

DUNCAN

R.

is

the

time.

subject is required to
Second,

AND

he

for

simple reaction

any
be

base

time

the

excluding

when

can

the

thus,

CHRISTIE

is

essentially no

transformation

loss

[see equation

20

READINGS

(4)],but
there

because

is

of

We

for

they

which

we

known

well

are

shall

L(F)

with

F and

shall

formed
working v/ith transelementary properties

to

of the

few

need

later;no proofs will


(cf. Churchill,1944).

be

given

(2)

L(G), then

the

sL(F)

property

(3)

N is

where

N,

f^ N(t)dt

continuous, the

are

F(0)

0 for all r"

A' is continuous

(4)

some

0. If it is known
and

so

0,i.e.,

G.
iv.

If

and

constants,

are

L(aF
If

V.

F(i)

\e~

bG)

where

The

aL(F)

A is

bL(G)

(5)

constant, then

-^"

UF)

IV.

f^y

iii. If

List

transform

ii.

that

advantage

the

F^(T)F^(t~r)dr\
L(F^)L(F^).
l\J

1.

function

PSYCHOLOGY

special propertiesof

distinct

functions.

MATHEMATICAL

of the

some

sometimes

of the transform

IN

(6)

=
.

Our

proposal is based on assumptions which are


do not appear
to be
intuitivelyacceptable, but which at the moment
It is our
susceptible of direct verification.
impression that any
Model.

deal with the full set of


must
empirical verification of the model
assumptions rather than with each in isolation.
Assumption I. It is possible, for a given experimental situation,
to divide the observed
reaction
time t into two latency components

t^and t^,called
1.

t+

2.

The

base

time and

respectively,such

time

choice

that:

value
and

on

it is

not

of

t^ depends only

the

required of

actions

motor

mode

the

on

directlydependent

on

the

of stimulus

entation
pres-

subject. Specifically,

the

character

the

choice

of the

choice

demanded.
3. The

value

of i

it is not
or

Let

the

depends only

on

directlydependent

the motor

actions

distributions

respectively. Since

of

t,

on

the

mode

demanded.
of stimulus

cificall
Speentation
pres-

required.

t^,and

conditions

and

be
3

denoted

by f, /^,and /^

imply that

the

two

com-

LEE

S.

latencies

ponent

it follows

CHRISTIE

R.

DUNCAN

for

independent

are

from

AND

21

LUCE

fixed

tion,
situa-

experimental

1 that

condition

(7)

Our
and

second

major assumption

requires the distribution


The

distributions.

by

in

such

exists

structure

the

of

structure

which

be

may

which

set

of

simpler decisions

of

decision

to that

organization of these

complex

is

one

of the nature

of

the

In

addition, the

and

restricted

shall

we

breakdown
to

of

decision

of the

machine,

both

analogous

of

set

by

decisions

engineer. The

the

This

at least

is true

of human

complex decision

more

machine

or

man

it is true

where

form

to

individual

the

serial process

from

being made.

If

machine, i.e.,the elementary

suppose

of

are,

computing machine,

machine

the

which

it is

elementary decisions

function

in

made

built into him.

making,

composed

as

tary
elemen-

the final decision

elementary decisions

process

latencies

more

that

capabilitiesbuilt into

actual

from

composed

is

decision

thought

choice

idea

in human

be

to

elementary relative

are

decision

and

into

organized

appropriate sense,

some

to

is

person

basic

only the

concerns

beings.

is not, in general,

elementary decision

one

portions may be
simultaneously employed on different parts of the problem. There
this is also true in a human
to suppose
seems
being.
every reason
by another, for in

is followed

We

shall

describe

graph, (The
in the
used

Lines

of

of

finite set

between

It is

line to connect
in

more

that

there

In this

may

paper,
that

shall

the

directed

Lines

when

we

Line.

that

in

of

with
use

these

between

in the
as

in

term

for

nodes^

we

and

directions

as

directed

pair of

nodes

ployed
em-

directed
in

directed

one

have

may
in
in

is

sists
con-

shown

are

graph, we
allowed,
there

graph

with

than

Figure 2a,

possibilitiesis
any

directed

that

been

flow diagram

term

more

directed

also

examples

sense

opposite
the

the

Several

by

have

called

are

general,

direction

be

decisions

coding.) A

points which

points, both

same

neither

suppose

computer

possible,

two

and

literature

pairs of them.

some

different

graph and network

mathematical
with

macKine

organization of

oriented

terms

in connection

Figure 1.
or

the

the

two

sense

Figure 2b.
shall
that

is at

pose
sup-

is,

most

we

one

22

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

Figure

employ a directed graph to represent the organizationof


At each node we shall assume
that
decisions
in the followingway:
an
"elementarydecision** will take place, the latency distribution
governingthe decision at node i being denoted by /^,The decision
have
process is initiated at node i when, and only when, decisions
We

been

shall

made

each

at

line from ; to

begin making

of those

We

his

may

; such

nodes

that there is

think of the "demon**

decision

until he

has

at node

received

waiting to

the decisions

precede him in the directed graph.


For the directed graphs we
shall consider, there will be
one
node, possibly more, which is the terminal point of

all the demons

these

will be

and

of

who

the

stimulus

node,

directed

decision

points which

are

activated

0,

There

will also

again possibly more,

which

initiates

at

time

Figure

be
no

by
at

at least

line;

no

the

least

directed

perimental
exone

line.

24

READINGS

Even

if

we

for certain

able

were

wide

MATHEMATICAL

into

process

hope is
relatively small

But

practicallyidentical.
II.3

formal

that

the

the

he

others

and

VVe should

want

When

is

as

For

the set

with

is

required to signal

is

but

of the

is

not

colored

sis
analy-

an

sumption
that As-

tary
elemen-

one

the

if

part of the

situations

the

will

naturally

we

subject is

differentlyfrom

location

the

set

of

that

one.

tions
situa-

of these

integers. VVe should

smaller

the

over

applied

be

put in the
find

to

the model

set

same

other

less

tuitivel
in-

held.

to

following sections
to the

fM)

form of

shall

we

make

the

following

/:

\e-'^\
t "

0,

[0

0,

^5"

There

grounds for supposing


this might be an appropriateassumption. First,let us suppose
that when
reached
been
has
decision
no
by time t following
A

stimulation
reached

are

ones.

explicitassumption as

where

to

probable

suppose

"similar"

as

ranges

model

of "similar"

some

the

experimentaldata we anticipate
of the directed
graph being a single point will be
the intuitively
"simplest" choice situation within

case

identified

is

stimulus

example,

of which

consider
n

model

the

the

forced

demands

which

they could not


if by great ingenuity we
able
were
simple sets of situations for which

even

positive constant.

at

time

between

proportional
case,

to be

want

experimental situations

points, one

probably reject the

In

those

to

generated

that

subprocesses which

S of "similar"

being similar.

as

presented with

S,

of

thinking,although

our

sets

subsets

as

of

Equally well, if

firings. It

situation

implicit in

model,

think

an

for its response.

also

include

not

neurone

stimulus

decision
It is

as

effectivelyprevents this extremity by requiringthe

of

existence

are

accept the model

process.

set

do

we

of individual

in terms

doing
extremely
so

excessively complex we should reject


that it is possible to subdivide
the total

The

model.

/ which

should

we

decision

adequate descriptionof the


directed graphs required are
the

distributions

that

is doubtful

complicated, it

be met

assumptions can
experimental data, but that in

of

classes

PSYCHOLOGY

that these

show

to

elementary decision

obtain

we

IN

to

it is not

0 then

t and

A^, with

+
a

are

the

that
probability

A^,

where

constant

difficult to show

that

two

the decision

small, is approximately

A^

is

of

A.
proportionality

the

will be

distribution

In this

of decisions

LEE

CHRISTIE

AND

R.

DUNCAN

is

correct

the

an

virtue

of

made

better

observation

and

more

that

certain

as

simple, the

more

is

assumption

observed

decision

situations
better

latency is

and

placed
exponential distribution slightly disorigin (Christie,1952b; Luce, 1953). The main

approximated by
from

this

that it has
empirical problem, but it must be admitted
simplicity. Second, and probably more
relevant,it is

relativelycommon

are

25

LUCE

Whether

exponential (Christie,1952a, b).

is

S.

the

an

is

plicity
generally on the rising limb. If this change toward simdirected
is actually toward
a
graph consisting of one point,
and if our
other assumptions hold, then it seems
plausible that the
elementary decision
latency is actually exponential but that the
error

observed

distribution

time distribution

VI.

presumed

to

satisfy the

then

to

graphs /V^,where
when
f ^cr)
(fiy

solutions

choice

that

appear

where

f^and /^are

which

compose

general, there

/ and

/' are

from the

same

triples with
stimulus
These

one

are

two
set

the

situations

distributions

appear

/. and

should

to

be

years,

of

be

no,

an

appropriate

or

one,

in any

degree

indirectly.

plausible to

It may

present

suppose

and

that,
if

problem. However,

this

to

with

associated

to

in

situations

choice

accept only those

both

to restrict

attacked, let alone

Further

cases.

the

possibilities.

solved, in this

difficulty. We

direction, but

electrical

of Section

satisfy the assumptions

considerable

in this

triples

following problem: Given a


of all triples (/j^,
set
f^,A/),

further

be

by

The

S.

the

of

solved

is to be

very

serve

not

important lead

it. In recent

that

set

of directed

set

of course,

S, then it will be necessary


same

of

each

somewhat

solutions

many

and

is

are

solution.

problem

seems

which

assumptions

hopes

one

first the

/. It

problems will
they

paper;

form

the

to

continuous, which

to

base-

the reaction-

member

that

may,

/, find the

distribution

/^ denote

/^and /

attacked

solve

appropriate to

continuous

only

be

must

the

situations

typical

exactly one

if the

of

model, i.e.,S

II. Let

S, such

over

problem, but

will be

generality,it

prove

in

with

ranges

to the

of S there

It would
of

Assumption

find distributions
a

of the

assumptions

in

convolution

of choice

set

yields the distribution /^. There

many

the

distribution.

composed according

"

IV

by

associated

distribution

problem is

"S be

Let

of the type described


time

smeared

the decision-time

and

Problem.

The

is

engineers

we

have

have

been

not

know

of

gated
investi-

concerned

26

READINGS

with

the

problem

networks

to have

with

there

systematic

manner

functions.

If

with

distribution

the

electrical

identifythe

functions, the

/ with component

the two

analogy between

an

we

transfer

and

network,

electrical
is

transfer

preassigned

the

PSYCHOLOGY

MATHEMATICAL

synthesizing in

of

given reaction-time
graph

IN

acteristic
char-

problems. This

is

but it is almost
certain
that solving
probably worth investigation,
our
problem will prove to be a major research
undertaking.

To

of

some

assumptions

our

then

by Assumption

and

another

consists

of

one

point. Let

in the latter

the

and

includes

times

pose

II

times

stimulus

From

case.

Let

exists
whose

situation

set

Assumption

we

the

tion
situaS which

directed

distribution

the

f^be

given stimulus

there

/j denote

simplifiedby using

transform.

for

know

we

be

may

Laplace

of reaction

distribution

observed
a^

problem we

the

extent

some

graph

of reaction

write

may

(8)

f ft,(r)fjt-r)dr.
/i(^)=
Taking
(2),

the

Laplace

in each

transform

L(0

If

we

divide

the

first

and

case

applying equation

L(4)L(4)

the

equation by

^g^

second

in

equation (9),we

obtain

Up
L{A)
This
is
is

is

seen

fairlycrucial

that all mention

(10)

UL)
of

consequence
of the base

this

point

Empirically,one
rather

we

should

does

not

approximations

raise
obtain

an

to

eliminated.

been

/ and

it

It

A'^.

important practicalproblem.

estimates

to the cumulative

F(t)

assumptions, for

our

has

time

equation relatingthe empirical data

an

At

but

L{Q
_

of

the

distribution

distribution

f f(r)dr

/,

LEE

(Throughout
while

Now,
well

whether

into

about

statements

(3) we

to

denote

their

be

reasonably accurate,

results,

our

cumulative

the

of data

So

cumulatives.)

tends

magnify

to

question

the

it is

arises

particular equation (10),

in

From

distributions.

equation

have

LU)

sL{F)

are

we

so

F{Qi)

speaking of empirical
equation (10) becomes

Since
and

to

avoided.

be

to

translate

can

may

tions
distribu-

denote

letters

differentiation

is, therefore,

we

to

numerical

that

and

errors

corresponding capitals

approximations

known

Latin

small

use

27

LUCE

DUNCAN

R.

AND

CHRISTIE

shall

we

the

and

S.

data

we

F(0)

assume

may

H41.

^^
eliminated

/^ from
Since

it remains.

/,

is

the

from

any

consists

in the

same

of

several

LiQ

As

an

example

exponential with

of

constant

suffice

is the

course,

case

to

assumes

determine
the

where

it

graph

case

L(F,)

UU)
^

.
_

^llilA
may

be

used, suppose

by equation (6)

Then

(12)

equation (12)

how

time

it will

cases,

point, in which

one

equation (11)

in

division

simplest, of

The

cne.

mining
discussion, the problem of deter-

our
our

0,

(11)

Having

/^is

s
1

and

If

so

we

equation (12) becomes

make

the

equations (3) and

reasonable

(5) we

find

assumption

that

/j(0) 0,
=

then

from

28

READINGS

Assuming

that

that

and

continuous

f^ is

PSYCHOLOGY

MATHEMATICAL

IN

f^ has

rivative
de-

continuous

equation (4)implies
1

df^

or

integratingfrom

0 to ^

Since

be

f^ must

determined

from

equation (13) that considerable

VII. Serial
the

of

estimates

accurate

Decision

F^

graph

An

discussed

in Section

but

they

may

such

extra

assumptions

with

hope that they

this
the
case

VI

is

choose

We

shall examine

extreme

forms

section, is
second, which
shown

in

be

may

two

cases

of the directed

the

general serial

will be discussed

the

about

have

relevant
which

for

The

shown

in Section

"

""

^-"-

Stimulus

FIGURE

directed

solution

of

heuristic

sense,

grounds,
the two

first,the topic of
in Figure 3a, and

VIII,is

-"

the

experimental

some

are, in

graph N,
case

intuitive

on

solving

this alternative

considerable

Figure 3b,
Stimulus

obtain

discover

to

may

most

to

program

explicitassumptions

value. We

data.

alternative

the

general problem,
the

to

necessary

The results of
elementary latency /
will,unfortunately,be much weaker than

A' and

program
the

be

clear from

is

certain

of

will

data

Process.

general problem

consequences

empirical data, it

"^

the

parallel

LEE

It follows
observed

S.

AND

CHRISTIE

Assumptions

immediately from

distribution

/ of

serial process

29

LUCE

DUNCAN

R.

II.2. b that

I and

having

nodes

is

the

given

by

rr 4^)4^^ -^i). -4^^


=['"""

4^^)

Applying the Laplace


(2) we

transform

to

equation (14) and using equation

have

dividingby

or

tjdt.dt,
...dt^. (14)

the

case

1,

"^.,(/)-.^:l^.

(16)

Equation (16) is the explicit form of equation (11) for


data we
case.
Clearly,if we have given numerical
may
(possiblynumerically)/ for each value of n.
As

general
A

example

an

of

form

this

might

be

when

done

/ is exponential with

suppose

the

serial

determine

know

we

time

the

constant

equation (16)becomes

In that case,

how

of

the

UF)

(17)
71-1

M^.)-(A.i)
1

In

Figure

have

we

presented plots

of

/1
values
A

of

equation

may

serial process

\X

second

of
^j(n),
is

be

obtained

with

vs.

"

for small

"

'^

1 r

by observing

that

exponential elementary

the

mean,

decisions

given by

^j(n) ^j(6)+^,

(18)

where

is
ii^{h)

the

base

mean

time.

Thus,

^.

(19)

,z,(n)-Mi(l)
=

equations (17) and (19) to attempt to decide


whether
a
given set of data is adequately fit by the assumptions of
the model, plus the added
assumptions of a serial directed graph
We

may

and

exponential elementary

now

questions

use

as

to

how

latencies.

this

may

best

There
be

are

serious

cal
statisti-

done, but the following

30

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

LO

0.8

0.6

0.4

0.2

O.Q

FIGURE

method

ready

s; this

we

assume

may

For

plot A"
5

We

know

find in

is in the form of

plot,which

of

and

We

thus enter

left side

for

of

function

shall call

we

value

some

of

correspondingvalue

if the

plot A

at this

selected

we

will be
of

the

and

correct

are

Since

the value

Figure 4

equation (17) that this

from

(19) presents
which

as

lated
formu-

are

of

y*

^,

chosen.

compute

we

value
(reasonable)

each

assumptions
s

the data

say

of

From

solved.

and

problems

statistical

until the

suffice

may

3.0

2.5

2.0

1.5

1.0

0.5

0.0

relation

satisfied
such

2s

our

that the

be

equation (19)]and (n

of

value

correct

has

our

observed

between

1)/A is

are

valid.

the observed
a

But

means,

been

the value

point and determine

the

if

to

HP,)

assumptions

error

(x^r

equal

this determines

between
if

must

of

equation
A, and

We
means

choose

[the

minimum; this yields

32

READINGS

To

the

evaluate

above

PSYCHOLOGY

MATHEMATICAL

IN

the

consider

sum,

function

n~\

t"(x)

-x'^-(i
-xy-'
2] C'-k')(~^)^cc^-x^

k=0

Observe

that
'

i\^k)
'

dx

^0

^0

=0

fc=o
A;
=0

T-

/^ +

and

L{nfFr').

that

^(x)dx

nl

x^^"^(1

dx
x)''-'^

i)r(n)
nr(|+

n) is
S(;72,

where

From

these

the

Beta

results

we

function

and

r(n)

Gamma

is the

function.

easily obtain

nBjY

1, n\

(22)

r(^
(l
+
^

n
^

+
-^

^)
i

n\r(j2J
+

In

Figure 4

for small

have

we

values

of

n.

also

presented

plots of

"

-.

iA

r-

vs.

LEE

The

of

mean

/i,(n) ^j(6)+
=

S.

the

AND

DUNCAN

R.

parallelprocess
and

"

T-

second

CHRISTIE

thus

be

can

have,

we

33

LUCE

shown

in

as

be

to

given by

serial

the

case,

relation which

be

must

met

^j(n)-^j(l)
=

^^

(23)

J,

1=2

The

procedure

for

fittingis

curve
s

serial

that

except

case

y
X

the

enter

IX.

graph

assumptions

there

there

is

but

it is

is
that

other

such

other

directed

between

to

use

first to do
to

The

that

but

another

to choose

will

be

such

selection

value

in

ad

an

be

there

Thus,

cases.

the

are

from
is

stants,
con-

an

of

is

sumptions
as-

Figure

in the

Presumably, any
some

shape of
the

extremely revealing of

cated
indi-

the

sets

seen

which, in

curves

have

n,

procedure

different

can
one

hoc

fairlysimilar.

produce

We

determine

to

among
for it

of

are

unfortunate
there

VIl)

and

as

of

set

one

any

them.

graph

of similar

set

such

constants,

that

curves

extreme
not

an

directed

serial

addition, within

to how

as

lie

sense,

the

proper

empirical
directed

situation.
a

number

in all Likelihood

of difficult

it will

prove

to

statistical
be

lems
prob-

efficient

more

experimental exploring using subjectivejudgments


and
the
to solve
goodness-of-fitbefore trying to formulate

statistical

X.

two

"

here,

two

place

general problem

problems

as

whether

best

certain

small

any

will

It is clear

as

In

how

as

almost

graph

these

the

to

such

undetermined

are

that the

curves

graph

and

evidently quite serious

for almost

favorable

more

arise statistical

data

optimal.
The difficultyof making

data

solution

procedure (end of Section

one

for the

not.

or

question

be

better

is

there

fit the

exponential /

assumptions

not

particularset of assumptions,

and

VI,

described

as

Without

in Section

described

and

is

Selection,

Model

well

than

same

to

seems

to

the

some

problems.

Perceptual Moment.

reaction-time

studies

the

mean

In

Section

reaction

II

time

we

should

remarked
be

that

in

of the order

34

READINGS

of

if unwanted

second

one

This

avoided.

have

peculiar phenomena

very

such

one

termed

PSYCHOLOGY

with

interactions

that the data

means

rapidly at certain discrete


The
them.
period between

and

times

who

to

adapted to cope
simple hypothesis as to

belief

Let

the

that while

and
that

is

timing of the moment,


to

assumption
only

be

may

able

to

shall return

we

The

we

where

took

it to

moment

formal

vestigat
in-

are

exist,

probably
sis
analy-

with

not

any

general
8

duration, say

stimulus

at

tures
fea-

seconds,

time

any

the

at

Since

we

end

tary)
(elemen-

is

during

cording
ac-

This

5.

to

that

happen

may

the

presented
0

in the interval

information

that

assume

may

the stimulus
h

during

of the period.

presentation and

the stimulus

person

part of the moment;

point later.
form

exponential, and
that if

no

so

shall

we

decision

has

decision

in the ith moment

by

response

the

ith moment

(1

-X5)P,_, +A5

reached

time

X5

iS
If

we

sume
ascase

discrete

by
then

the
the

call the

then

P.
,

^i =^i-,+ [l-P"-.]^S
=

the

use

been

is

should

we

continuous

the

following the presentation,i.e., at

probabilityof a
of a
probability

this

shall make

that the

information

inappropriate,for it

assume

the

we

of the moment,

assume

may

assimilate

be

it most

that all intermediate

assume

distribution

to this

give

to

that it does

arises
to the discrete
as
question now
In
for the elementary decision
process.

analogue. We
ith

multiples of

we

uniform

as

between

correlation

no

fractory
re-

on

Indeed, there

to indicate

receive

may

at

occur

time

In this section

is of fixed

moment

person

unchanged.

only serve
Furthermore,we will

there

is

only

period it will
decisions

conducted

case

situations

to

the nature

analysis remain

assume

us

In the

with it.

that it is correct, but


of the

is in

is, therefore,of interest whether

be

can

It

effect.

an

tion
informa-

that he

been

not

applied

servatio
ob-

period from the beginning of


the beginning of the next has been
(Stroud,1949a,b). Unfortunately,

its existence.

will be

analysis

will have

doubt

certain

explain these

and

possible at this
propertiesof the moment.

of the

statement

our

it is

so

where

range

to be

are

subject processes

relativelylittle direct experimentation has

problem,

stimuli

To

observed.

been

hypotheticalevent
the perceptual moment

other

will be in

proposed that

been

it has

MATHEMATICAL

IN

(24)

LEE

With

the

solved

CHRISTIE

S.

AND

initial condition

P^

the

equation (24) is

difference

by

p^
The

35

LUCE

DUNCAN

R.

probabilityof

xsy

in the ith moment

decision

(1

is

obviously

[1^P._JA5;
hence,

have

we

X5(l-A5)'-i
as

distribution

our

If

replace this

we

continuous

^
the limit

as

the observed

f^(t)fUrn
=

-"o

data

point i8

the

in the discrete

then
,

width

and

distribution.

be

by /^ as

denoted

serial

case

is

height

it is clear

the discrete

distribution

base-time

the

rectangles of

this becomes

"

"

has

about

centered

"

distribution,equation (25), by

discrete

"^^which

one

"

Let

(25)

that in

before, then

given by

/ f,(t,)h(t2-^i)%(h-i2)-"
/'"""/'/
...

(26)

Applying
L

(fj

the
=

Laplace

transform

L{h) L {^^^
(/j)

Urn L

using equation (2)

and
=

lim L
L(h)\
(f^)

($,)]" (27)
.

Observe,

.2.

_i

f
""

8"

2] X8(l-X8y-'e'

A5(l-A5)-'
se

"

f^l(l-X8)e-'^'

36

IN

READINGS

MATHEMATICAL

PSYCHOLOGY

But,

lim

so,

Lim

(0^)
^

5-

-(l_A5)e-^^

Substitutingin equation (27) and

dividingby

the

case

1, we

have

(28)

=[l_(l-.A5)e-^sJ
Z(^)
'

which

is the

crucial

of the discrete

mean

equation for
distribution

^
Thus,

the
is

discrete

observed

The

case.

given by

i5A5(l -kSy-^

the relation between

serial

=j

(29)

is

means

//j(n)-^,(l)=^^.
if

Now,
then

these

fashion

know

we

two

as

were

theoretical

no

the value

5, i.e-,the length of the moment,


be used in exactly the same
of equations may
sets
We have
equations (17) and (19) of Section VU.

value

of 5

is

real

of

measurements
moment

of

it will be

so

phenomenon

it will be

71

further comment

1, the

such

as

with

f^

convolution

that shown
,

of h and
in

for reasonable

than

1, the

amount

0^, when

Figure 5.
/. will
,

The

serve

Smearing
depending on the

If
"

0, is

convolution
to

the

smear

will also
value

of

ascertain

its

ignore /^ and let

we
""

perceptual

time.

reaction

on

to

dependen
perform in-

of this function

steps but it will

result
n.

step function

if

is

Thus, if

larger
our

sumption
as-

roughly correct, we should expect,


least for comparatively simple situations,to find the observed
as

to

the moment

that if the

important

interest:

some

utterlydestroy them.

not

at

of

to

necessary

It is clear

it.

propertiespriorto analyzing experiments


One

(30)

is

latencydistribution somewhat
lumpy. Indeed, in the Literature (cf.
Woodworth, 1988) it has been remarked not only that the data are
lumpy but that there is an oscillation superimposed on the distribu-

S.

LEE

CHRISTIE

DUNCAN

R.

AND

3"

28

to

were

to

This

curve.

effect could
h uniform

assume

55

4f

FIGURE

tion

easily be obtained
only

over

is

trulya refractoryperiod during which

information.
need

for

These

we

assume

considerations

bringout

comprehensive experiments

we

portion of the interval


the vast majority of the moment

in other

if

analyticallyif

small

5,

words,

37

LUCE

there is
even

of

strongly the

more

determine

to

intake

no

propertiesof

the

the moment.
We

shall not

reasons

with

are

attempt,

that the

worthwhile

of the

opinion

that

carry

it is

explanationof
dimensions*'

an

effect

nature

the

out

with

other

of

than

complex
it

moment

the

is

moment

one

for

hint

number

of

may

and

hardly

accepted in

serially. It

possible

changing

information

of the

information

accepted in
is

rather

analysis. Furthermore, we

unlikelythat

latter remark

the

in

the

on

that the information

parallel. The

problem is

mathematical

is dealt

moments

however,

to

The

before, to study the parallelcase.

little information

so

seems

in

as

are

ferent
dif-

happen,

processed

developing

an

"psychological

display.

ExperimentalProposals. The key assumption in our analysis


be found of such a sort
is that elementarydecision
can
processes
which
that complex decisions
be built up from them in a way
can
XI,

leaves

present

their characteristic

A value

experimentalsubjects

with

invariant.
stimuli

One

which

should
vary

in

like to
several

38

READINGS

but

dimensions
identical

several

we

values.

different

in every

other

respect,

If

possibly introducing
several
objects with the

with

and

the

have

we

have

conceptually different

uses

use

we

for each

relevant

dimension

same

into

one

of the dimensions

difficultyof

the

run

each

on

If

characteristics.

time

dimensions,

decisions

for which

PSYCHOLOGY

MATHEMATICAL

IN

identical characteristics

that
difficulty

the

reception of

into several
unitary, but broken down
parts separated by receptor orienting acts such as eye movements.
The
first of the two
following proposals suffers from the latter
the

stimulus

1st

of

row

3"

of

with

cards

pairs per card

position. Cards
n

of

triple-spacedtyped, horizontal
The
digits, 0 and 1, on each.

from

sixteen.

to

one

On

each

card

no

series

the

pair in
to

to vary

pairs will be unlike digits,i.e.,(0,1)or (1,0);


like pairs,i.e.,(1,1)or (0,0). The place of the unlike

pair or

one

the remainder

one

5"

verticallyaligned pairs

number
either

the former,

Experiment: Digit Difference Perception

White

Stimuli:

from

second

the
difficulty;

be

not

may

with
be

will

of

pairs

the

unlike
in

included

to

the initial to the final

from

vary

pair in each
the

with

set

of the

positions from

equal frequency,and

frequency.
pairwill be included with the same
the
The
remaining places will be
assignment of (1,1)or (0,0) to
made
an
on
equiprobable random basis, and the choice of (0,1)or
basis.
(1,0)for the unlike pairwill be made on the same
cards

with

no

unlike

Responses'. Experimenter
presentation how
will

respond

not

bear

or

announce

pairs the card

subject will be

told

to
no

will be instructed

and

be

each

shown

bears.

depending on whether the card does


pair, by pressing the appropriate one
that

possible positions, including in


events,

to

prior to

no,

unlike

an

The

keys.

yes

many

will

to

Subject
or

does

of two

pair in each of the


equally likely
position, are
unlike

an

no

stimulus

read

the lines

of

pairs from left

primary interest will be the latencies of the


tencies
unlike pair and the lawhich
bear no
to the cards
response
bear an unlike pair
of the yes response
to the cards which

right. The

in the

nth

data

of

position.

Apparatus: 1.
2.

Stimulus

cards

as

described

Light projectorwith

fast

above,

shutter,

40

IN

READINGS

data

of

primary

set-stimulus

Same

Apparatus'.

for

calling

pairs

for

as

be

will

interest

PSYCHOLOGY

MATHEMATICAL

the

latencies

the

the

to

response

response.

yes

first

of

experiment

for

except

the

stimulus

cards.

LITERATURE

S.

L.

Christie,
Psychol,

59,

S.,

L.

Christie,

in

Cambridge:

R.

Discriminative

Behavior,"

Macy,

Jr.

Groups."
of

1952b.

"Communication

(Technical

Report

Electronics,

231.)

No.

Massachusetts

stitute
In-

J.

Modem

Macy,

Jr.,

L.S.

Christie,

Groups."

Task-Oriented

in

Mathematics

Operational

Engineering,

in

Hill.

Research
of

J.

Laboratory

McGraw

D.,

Flow

and

Task-Oriented

1944.

V.

York:

Luce,

of

Technology.
R.

Churchill,
New

Luce,

D.

Research
of

Measvirement

443-52.

R.

Learning

and

"The

1952a.

Rev.,

Laboratory

and

H.

(Technical
of

1953.

Hay.

"Information

Report
Massachusetts

Electronics,

bridge:
Cam-

264.)

No.

Institute

Technology.

Stroud,

J.

Steinford

Stroud,

1949a.

B.

J.

1949b.

B,

"The
Sixth

Ed.

Forrester,

Moment

Hypothesis."

Function

M.

S.

Thesis.

Psychological
Conference

Moment
on

Perception."

in

Cybernetics,

Josiah

H.

Macy,

27-63,

Foundation,

Woodworth,

"The

University.

R,

S.

1938.

Experimental

Psychology,

New

York:

RECEIVED

Holt.

5-5-55

von

Jr.

PSYCHOACOUSTICS

DETECTION

AND

David
massachusetts

This

institute

how

process.

By

statistical

to

to be
appears
of signal and

with

theory

and

the

theory

has

been

threshold

noise

of the
held

are
on

as

process

independent

special emphasis

of

concept
used

theory

as

ideal

the

applied
of

two

The

observer.

analyze

to

it is

as

combination

paper
threshold

auditory

of

the

instance

an

psychophysical
The

constant.

of

assumptions

when

procedure
of

concept

derivation.

the

the

ideal

physical
is

observer

viewed
re-

usefulness

The

of

the
by considering the shape of the psychophysical function
relatingthe detectability of the signal to its intensity. A rather general model
is illustrated

this concept
function
based

of detection
is treated

theory

hypothesis testing,two
of the
(2)
are
(1) the detectability of the signal and
recognized:
process
of analysis which
The
level of the observer.
theory provides a technic
The
tectability
factors.
of signal deof both
obtain
measure
a quantitative estimate

criterion

parameters

technology,

review

Detection

decision

treating the

determinants

one

of

massachusetts

fairly complete

decision

structures:

discusses

allows

data.

psychoacoustic

theoretical

the

presents

paper

certain

Green

M.

cambridge,

to

THEORY*!

the

on

"

of

concept

is

signal uncertainty

which

presented

to

attempts

this

explain

relationship.
Introduction
There
is the

two

are

breadth

from

techniques range
of vowel

that

which

to

second
view

deficit

consensus

is the

This

recent

interest.

t This

by

the

Hanscom

Note.

Its

Other

paper

data
have

not

of

Electronics,

M.I.T.

is

"

any

to

in J. Acoust.

This

is the

Soc.

first of

papers
was

will

This

in

is Tech.

Rept.

No.

part by

follow
the

Force

in

If

basic

some

AFCCDD

41

papers

U.S.

Air

on

grant from
under

Research

by

TR-60-20.

Psycho-

reflection
a

of

general

complete

of the

Reprinted

latter

with

mission.
per-

aspects of acoustics
the

National

of

Science

issues.

Force

Cambridge

force

example

1189-1203.

theoretical

old.

the

may

recent

subsequent

administered

and

Massachusetts,

32,

from

integrative

paper
A

series of tutorial

supported by
partially

Bedford,

new

1960,

Amer.,

publication is supported
of this kind

field,a

procedure.

measurement

difficult.

structure

complete comprehensive theory.


where
on
methodology. Often, even
of the

tion
percep-

it reduces

most

areas

integratedwith

easily be

The

being overlooked.

is

these

of any

since

one

area

some

entire

analysisof the

system

sensory

lack

study hearing.

to

fortunate

integrationof

makes

Operational Applications Office, Air


Field,

of the

field is the

might

consensus

exist in

appeared

Editor's

Foundation.

lack

of the

article

of the

new

does

to

seems

re-examination

cochlea

rapidlyexpanding experimental literature.

the

acoustics, however,
this

diversitywhich

characteristic

existed, these

structure

of the

studies

psychoacoustics. One

techniques used

approach
multidisciplinary

reallysignificant
aspect

any

it creates

However,
A

This

field of

of the

skills and

hydrodynamic

forms.

the chances

strikingcharacteristics

very

varietyof research

and

the

contract,

Center,

Research

monitored
Laurance

Laboratory

G.
of

42

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

scale of
exchangesof Garner^ and Stevens^ on the quantitative
loudness.
Such a situation compounds the problem of integration.
This paper, therefore,makes
no
attempt at broad coverage. The author hopes
that by concentrating
limited
rather
contribution
one
be
on
can
topicsome
positive
made.
This topicis the detection of signals
in noise. In recent years a general
tical
theorestructure
(detection theory)has been used to analyzesuch experiments.Unfortunatel
confusion
there appears to be some
both about the theoryitselfand the
of its application.
The main objective
of this paper will be to clarify
these two
manner
Fart
of
confusion
the
about
the
arises
from
the
fact
that
detection
questions.
theory
of two distinct theoretical structures:
decision theoryand the
theoryis a combination
of
ideal
observers.
Before
detailed
discussion
of these two
we
a
theory
begin
aspects
of detection theory,we
will briefly
outline them
and relate them
to psychoacoustic
problems.
Decision theoryprovidesan analysis
of the process which generates the dichotomy
between
stimuli the subjectreports he does and does not hear.
The theory
that
and
of
and
incorrect
a
correct
costs
values,
decisions,
recognizes
prioriprobabilities,
this
as well as the physical
parameters of the signal,
playa decisive role in establishing
criterion.
dichotomy. We will find that this dichotomy is determined by an adjustable
The theoryshows how a quantitative
estimate of the criterion can be obtained from the
be found

may

in the

data.
There

are

many

whose
psychoacousticians

parameter from

constant

which

obtain

to

parameters, for example,the absolute


the

only interest

substantive

threshold

in this criterion is as

relations between

energy

as

function

two

of

physical

or
frequency,

justdetectable

change in power as a function of power (A/ vs /). To them this


if
theory will be of methodologicalinterest only. Yet clearly,
factors such as a prioriprobability,
and
do
in
role
the
costs
values,
playa
determining
is imperative.
threshold, their control in substantive experiments
The
second
related to substantive
directly
part of detection theory is more
it is the theoryof ideal observers.
the theoryprovides
matters
a collection of
Briefly,
ideal mathematical
models
which
relates the detectability
of the signalto definite
characteristics of the stimulus.
There is a collection of such models
because
physical
aspect of detection

"

different restrictions
make
one
may
theoretical observers
are
rarelyused
Most

often,theyare

the ideal observer

comparison,in

used

of the

suggestseither

mechanism,
hearing
This
discrepancy. will be

or

new

models

and

and

detection

of the

comparing human

specifythe nature

to

of the

nature

actual

as

for the sake of

in order

turn,

the

on

amount

new

illustrated in

These

hearingmechanism.

performancewith that of
of discrepancy.
This
accurate
representation

more
hopefully
further
to clarify
experiments
a

device.

the exact

nature

of the

later section of the paper.

Decision

Theory

We shall demonstrate, under quite


how a transformation
of
generalassumptions,
utilized
determine
both
the
criterion
and
the
be
to
can
subject's
subject's
responses
of the signal.This analysis
detectability
requiresan understandingof several basic
and start,
concepts which are rather complex. We mightskipover these fundamentals
the

W.

S. S. Stevens,J. Acoust.

R.

Garner,

/. Acoust.

Soc. Am.,
Soc. Am.,

1958, 30, 1005.


1959, 31, 995.

as

assumptionsabout Gaussian distributions


would
be unfortunate
because
a procedure
impliesthat strong assumptionsare needed to

have, with
previousexpositions

some

and

parameters of these

it robs the

of
analysis

distributions.

some

Such

its generahtyand

its applicability.
Such
justify

is not

43

GREEN

M.

DAVID

the

case.

psychoacousticians
Typically,
by making
responses
try to analyzethe subject's
the
the
about
in
is
the
which
sound
some
way
assumptions
processedby
hearing
mechanism.
One assumes,
for example,that the cochlea either makes a frequency
sis
analyof the waveform
that it does not, etc. We
wish to postpone temporarily
such
or
substantive
issues. Let us, for the present, merely assume
that each sound
may be
These numbers
by a series of numbers.
represented
might be the values of a series of
the representation,
let
attributes,or various states of the nervous
system. Whatever
us

call this abstraction


The

alternative should

observation.

an

wish

problem we

consider

to

be chosen?

is this: Given

is

What

observation, what

an

choice

response

how can
we
good
analyzethese
choices?
these questions
We shall attempt to answer
a single
by considering
example.
the generality
The example is obviouslyspecific;
rests in the concepts. The
single
motive in presenting
this example is to enable us to discuss these concepts
likelihood
with some
and yet avoid formalism.-^
ratio, decision rule, and criterion
precision
of these concepts
After this theoretical discussion,we shall investigate
the applicability
to a psychoacoustic
experiment.
a

and

"

"'

"

exampleof decision theory

An

Let

us

three numbers

[Xj

observations.

Given

of

instance

H^

{x^,x^, x^)],and

an

shall

observation

extended

reader

should

of the observation.

The

(integers

or

completeinformation
hypothesis.

observations

with

work

can

we

is an

about

the

probabilities

x^) could
{x-^,
x.-^,

numbers

have

been

ality
is independentof the dimensioncould

{x) of the observation

variables

observation

the

whether

have

we

10

{Xj) represented
by
//j,H.y,about the
hypotheses,

two

Everythingthat follows

real numbers)

or

observation

decide

to

that the three

note

three hundred.

to

wish

we

assume

to

have

we

giveneach

the example
By limiting
The
directly.

that

observation,

//g.^ We

or

of each
probability

10 observations, each

have

we

assume

(red, blue, or
qualitative

be

green).They

are

quantitative
scription
simply de-

of the observation.
Likelihood
numbers

In

ratio.

corresponding

These

inference.

concepts

Most

which
principle
*

A.

J.

"

I,

have

we

observation.

from

come

of the

Table

each

to

theorems

the

topicof

key
with Neyman
originated

listed the
The

statistical decision

and

Neyman

For

a concrete

theoryand

E. S. Pearson, Phil. Trans.


of the
interpretation

the

extended

who

data

theoryof
the

basic

Pearson.

Wiley,1950.

York:

New
Wald, Statistical decision functions.
and

the three

providethe

columns

first presented
by Wald,

were

and

observations

two

next

Roy.

Sac.

1933. A231, 289.

London,

example,the reader

mightthink of the observation

and the
width, and depth of the package,
package,the three numbers as the length,
animal.
The
whether
the package contains a toy car or
as
problem,then, is this:
hypothesis

as

sealed

Given

the measurements

attributes.

consonant

package, guess

mightthink of the observation


The
problem is: Decide

one
or

of

or

vowel.

as

whether
a

from

sound

it contains
which

the three

can

be

numbers

car

or

an

animal.

specified
by three
whether

the

natively,
Alternumbers

sound

is

44

on

READINGS

the

of each
probabilities

observation

the ratio of the fifth column


likelihood

H^

divided

that

H^

that
by the probability

is correct.

that this number


observation
variable

to

on

PSYCHOLOGY

each

If

Note
is

which

is

we

have

representsthe
a

itresulted from

{X^) we

should

that the likelihood

function

The
hypothesis.

the sixth and

that
ratio,then, is the probability

call the "odds."

some

MATHEMATICAL

IN

specified
by

three

values

likelihood

observation
particular

H^.
be

The

likelihood

to
willing

ratio is

of three variables

final column

number,

ratio.

The

resulted from
ratio

giveswhat

nine cents

wager
not

is simply

to one

and
probability,

have taken an
{x-^^,
x^, x^. Thus we
related
and
it to a single
(x^,x^, x^,

l{x-^,
x^, x^.

The

reason

make

we

have

decisions

performed this

if we

transformation

is

simplystated:

We

can

the likelihood

ratio. We
have not stated what we
use
optimum
by optimum, but let us take up this pointa littlelater. First,let us show how we
might use the likelihood ratio in making decisions.
rule. If someone
Decision
asks us to make
decision about
a
a
particular
observation,whether it is an instance of H^ or H^, we would probablyguess it was H^
if the probability
of that observation was
greateron H^ than on H^. Such a statement
is called a decision rule. In terms
of likelihood ratio this decision can be expressed
as
follows:
Choose
In
if
decision
rule by
1{X) " 1.
our
effect,we have specified
H^
choosingone number; in this case, the number "one." This number is called a criterion
a likelihood-ratio criterion.
or, more
precisely,
ten times as
observation, //g was
Suppose that, independentof any specific
would
without
maintain
not
a
s
criterion;
w
e
our
even
likely H^. Clearly,
previous
mean

knowingthe
It turns
choose

and

out

characteristics of the observation, the odds

in this

case

H^ only if,in

that

we

should

choose

are

H^ only\U{X)

ten

"

to

one

in favor of

10. That

is,we

i/g-

should

observation
is X
(4,3, 3).
example,the specific
if we
Similarly,
placeasymmetricalvalues and costs on the various correct
incorrect decisions,we should change our criterion or likelihood ratio accordingly.
our

DAVID

Monotonic

functions
of likelihood

in terms

of likelihood

While

ratio.

ratio,there

45

GREEN

M.

we

state

can

decision

our

cedure
pro-

exactlyequivalent
ways of stating
the decision rules. In the example, it so happens that the producta\ times ."".,minus .T3
is also an optimum decision quantity.This is true because this quantity
is monotonic
with the likelihood
on

which
we

ratio. The

likelihood-ratio

decisions

criterion number

the criterion number

alternative

usingthe

In

is not

scale,but there is always some

correspondsto

select the

other

are

on

likelihood

if

" 1.25;
/(-^i,
H^
-^2,.T3)
decision rule,select H^ if {x\ .rg
"

such

the

same

number

x^

"

use

scale

example,suppose

would

we

would

we

this monotonic

ratio. For

then
"

that

as
on

identical

make

5.00.

of this theoryto psychoacoustics,


the
application
many
decision axis is unobservable, and hence we
interested
in
are
decision
only
equivalent
procedures.To say the observer uses an optimum decision proceduremeans
only that
he is usinga monotonic
transformation
of likelihood ratio.
Optimum nature of likelihood ratio. We turn now to the very importantquestion
of the optimum nature
of likelihood ratio. Clearly
a decision procedurebased
on
likelihood ratio is only optimum if it best attains some
Let us list
specific
objective.
of these objectives
to indicate their generality:
the expectedvalue
some
(1) maximize
cases,

the

as

decisions,^
(2) minimize risk,^
(3) estimate a posteriori
(4) maximize
probability,^
and (5) set the error
decision
rate
decisions,^
on
some
percentage of correct
of

at

and

constant

some

maximize

the number

of correct

decisions

the
native
alter-

for the other

alternative.^ The

fact is that a decision criterion based on likelihood ratio is


impressive
optimum
objectives.
Naturallythis criterion may be different for
different objectives.
The references listed with the objectives
contain a more
detailed
of
each
and
how
decision
rule
based
likelihood
on
a
ratio,
explanation
objective prove
monotonic
transformation
of
that
the best
or
some
quantity,
may be used to make
under

all the above

decisions.^"
Distribution

of likelihood

of the number

quantity

We

ratio.

of attributes included

likelihood

ratio.

Likelihood

have

seen

how

each

observation, independent

in the observation,can
ratio is

be reduced

function

to

single

of several variables

simply
We
singleobservation is simplya number.
may then properlyconsider a
defined
variable
likelihood
ratio.
Let
the
the
us
consider, in particular,
on
probability
that
shall
obtain
of
likelihood
ratio
value
under H^ and Ho
a particular
probability we
of the precedingexample. Table II shows
these probabilities
and the corresponding
cumulative
distributions for both hypothesesof our example. The likelihood ratio is
of the ROC
curve.^-^
ranked from largest
to smallest to facilitatethe explanation
"

and

for any

ROC

and

curves

their

(Receiver OperatingCharacteristic)curve.
is to accept

H^

if l{x^,
x^,

shall

properties.We

x^) ^ k.

If ^

14

'

W.

W.

Peterson, T. G. Birdsall,and W.

T.

W.

Anderson,

M.

Woodward,

An

introduction

to

Table

do this,let

To
=

use

we

us

find that the

C. Fox, Trans.
multivariate

II to construct

the decision

assume

of
probability

IRE,

ROC

an

rule

accepting
171.

1954, PGIT-4,

York:

statistical analysis,New

Wiley,1958.
^

New

P.

York:
^"

estimate
"

McGraw-Hill,

of

criterion is involved.
no
probability
posteriori
transformation
monotonic
is a simple
posteriori
probability

estimate

To
a

Note

and informationtheorywith applications


to radar.
Probability
1955.

that since two

under
probabilities

both

observations

to
hypotheses

yielda

obtain

the

likelihood

In

ratio of 0.50,

of that
probability

this

the best

case

of likelihood
we

have

likelihood

ratio.

added

ratio.

the

46

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

TABLE
Each

under
Probability

KX)

Will Have

Hi when it is true [Pjj(Hj)]is 0.14 and


k, we
[Pjj(Hi)] is 0.01. By decreasing

II

Hypothesisthat

Certain

Value.

of accepting
H^ when it is false
probability
The upper curve
change both probabilities.

the

probabilities
change as a function of k, and is called an
matrix
The two
ROC
curve.
probabilities
completelyrepresent the stimulus-response
in a two-alternative detection task since the complements of Pjj (Hj) and Pjj (Hi) are
the two remainingcells in the stimulus-response
matrix.
shown

in

Fig.1

shows

how

the

1.00

0.80

55

0.60

a:

:z

0.40

0.20

0.00

0.20

0.40
Error

0.80

1.00

rate-PH2(H])
Figure

The

0.60

receiver

characteristic (ROC) curve


of
operating
is the probability
of responding
if
the
observation
H^
of respondingH^ if the observation
was
probability
Table

II.

the

example.The

was

from

axes

from

H^, and

H^.

The

are

which
Pji_J^H^),

which
PgJ^H^),

pointswere

is the

plottedfrom

48

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

III

TABLE
Calculation

of the
in

likelihood

ratio

of
Probability
Forced-Choice

Correct

Response

Test.

largerlikelihood was in
fact produced by H^ and the smaller was
in fact produced by //g. The probability
of
this occurrence
is Pjj^[l^iX)]
where
In
if
the
"
fact,
l^iX)
l^iX).
PjiJIzi^)]
larger
likelihood ratio is equal to k, the probability
of a correct
choice is simply:Pjj [1{X)
all
k] ^iPfjUii^) " k]}^ To obtain the final result we need only summate
over
the values of k, since any of these values mightbe the largest,
value
the
lowest
except
was

produced by H^, we

shall be correct

ifthe

"

"

of likelihood
Table

ratio itself.

(0.8042). While the


givesthese calculations and the final answer
of calculating
method
this probability
is straightforward,
in psychooften, especially
acoustic experiments,
does not have numerical
distributions on a likelihood-ratio
one
scale. Two
and the safest,
approachescould be used in these situations. The first,
since it makes
additional assumptions,
would
be to compute the probability
from
no
determined
If
III
ROC
look
Table
at
curve.
an
experimentally
closely,
you
you will
used in the calculation are simplyAf^ (//J times [1
see that the quantities
Pfj (H^)]
for each successive pointon the ROC
1
t
he
curve
(Fig. ). Obviously, accuracy of such a
estimate of the
procedureis heavilydetermined by the accuracy of the experimental
ROC
The merit of the techniqueis that no assumptionsbeyond that of the
curve.
decision rule are necessary to predictforced-choice
data.
behavior
from
the ROC
A second procedure,
which has often been used, is to make
one
some
tions
assumpabout the distributions which generated
and then use these assumpthe ROC
curve
tions
in predicting
behavior in the forced-choice
experiment.The most popularset of
assumptionsis that the distribution of observations on the likelihood-ratio axis, or
monotonic
function of that axis,is normal
under both hypotheses.
some
or Gaussian
The distributions are assumed
to differ only in their means
and, sometimes, in their
standard deviations.
for simplicity,
that standard
deviations are equal
Let us assume,
under both hypotheses,
then the ROC
be characterized
curve
can
by one parameter;
III

"

"

If

more

than

two, say M, alternatives

are

used

in the forced-choice

becomes

P(correct)

2^^/^
k

^)

" k)
^Pu^ili

test,the

equation

M.

DAVID

the difference in the

usuallydenoted

divided

means

by the

49

GREEN

standard

deviation

by

d'

^.Mja.

calculations

(AM/rr). This parameter

of a correct
probability
detection in a two-alternative forced-choice situation ifthese assumptions
made are
are
that
The
likelihood
is
than
another
is
the
quitesimple.
probability one
larger
ability
probthat the difference is greaterthan zero.
tion
transformaSince, by assumption,
some
is

The

of the

of /(A')is normal, the difference distribution is normal


variance

equalto

decision

is

the

sum

of the

variances.
original

PCcorrect,2 alternative)
The

of beingcorrect
probability

reference

of AM

mean

of
probability

and

correct

-|- 0%)^] 0[^7(2)'].


a)[AM/(CTf

for any

number

of alternatives is

givenin

footnote

have

reviewed

now

all the essential aspects of how

theoryin

the process of detection.


analyzing
these notions
results and see to what
extent

Let
are

detection

theoryuses
mental
experisupported.Followingthis

us

now

turn

to some

studies,we shall conclude this section with a discussion


experimental
studies
for psychoacoustic
of
these
implications
proceduresin general.

review
the

the

15.

We

decision

Hence

with

of the

of

Experimentalresults
ROC

One

curve.

the

of the earlier studies^^ simplysought to

shape of the ROC

in

curve

determine

task.
simplepsychoacoustic

The

experimentally
signalwas

1000-cpssinusoid. White noise,the masking stimulus, was present


occurred to mark the obsession. A light
servation
continuously
throughoutthe experimental
added to the noise (SN)
interval. During this interval either the signal
was
these
the
two
were
or simplythe noise was
(N):
hypothesesof the detection
presented
button if he
of two
task. The subject
possible
gave one
responses; he pressedone
believed no
second
if
he
believed the signal
button
a
was
or
present("yes") pressed
noise
of
The
the
situation,including
signalwas present ("no").
physicalparameters
the probability
and signal
held constant.
The independentvariable was
levels,were
selected
of
of
Five
levels
were
a prioriprobability
(0 priori) a signalbeingpresent.
(0.1,0.3, 0.5, 0.7, 0.9) and the one used for a givensession of 300 observations was
announced
to the subject.After the subjectresponded,he was
given immediate
The
information
the signalhad in fact been presented.
to whether
not
as
or
subject
and fined an equalamount
fraction of a cent for each correct answer
awarded
was
some
much
as
He
for each incorrect answer.
instructed to make
as
was
possible.
money
The results for one
of the subjects
are
ability
presentedin Fig.2. [Py;{A)is the probof
the
data
trend
The
of saying"yes"when
noise alone was
general
presented.]
is generated
drawn
The curve
by assumingthe
analysis.
supports the decision-theory
distributions on likelihood ratio are normal
under both hypotheses.The normalized
a

1/10second

of

difference between
Threshold

the
model

is 0.92.

means

and

the ROC

subjects
adopted the
consider
^=

No.

the
whether
not
or
considering
let
their
maximize
payoff, us
actually
Before

so
as
proper
alternative explanation
of the data.

one

P.

curve.

criterion

B.

Elliott,Electronic

Defense

to

This is the so-called threshold

of Michigan,Technical
Group, University

model.

Report

97, 1959.
"

W.

P. Tanner, J. A. Swets, and

Michigan,Technical

Report No.

30,

D. M.

1956.

Green, Electronic

Defense

of
Group, University

50

READINGS

IN

0.00

0.20

0.40

0.60

Figure
A

sample of the ROC

is the

The

0.80

1.00

experiment.See footnote 16. P^{A)


was
Psn(^) is the probapresented.
bility
when
These probabilities
estimated
was
were
saying"yes"
signal-plus-noise
presented.
from the stimulus-response
matrix.
See text for details of the experiment.

probabihtyof
of

from

PSYCHOLOGY

MATHEMATICAL

curve

auditorydetection

responding"yes"when

essentials of this model

process within

an

are

that the

noise alone

when
signal,

added

to

the noise,augments

some

organism,such that if the increment reaches a critical level called


the threshold, the signalis heard and can
be correctly
detected.
So far,we
note
no
analysis
greatdifference with the decision-theory
except in semantics. If one calls the
criterion a threshold
and the hypothetical
decision-theory
process likelihood ratio,
is
differences
between
the models appear when
the correspondence complete.The
one
considers
"subthreshold"
and the proceduresused to deal with these events.
events
fail to reach the threshThe threshold model
that should the signal
increment
old,
assumes
is present.
the subject
can
only make a pure guess as to whether or not the signal
This is surelytrue since anythingbelow the threshold is justthat. If orderingis preserved
below the threshold, the word
has no meaning. The difference in terminology
criterion and threshold is important,
for to say the subject
between
adoptsa criterion
rule.
continuum
is
used
is to simplysay an arbitrary
cut pointon
a
as the decision
Given that the subject
guesses about events which are "subthreshold," he may,
if blanks are ever
employed,report the signalis presentwhen it is not (falsepositive
both consistent with the threshold assumption,
might be
response).Two techniques,
instruct
the
t
o
ifthis
used
is
to
One
occurs.
subject be more
employed
procedurewidely
be interpreted
an
as
careful; this can
attempt to instruct the subjectto respond
cussed
of this procedurewill be disThe
all
"subthreshold"
to
events.
implication
negatively
of
this
from
the
valid
in a later section. Another
assumptions
equally
procedure,
sumes
model, would be to employ a correction for guessing.This correction procedureasThe
the guessingmechanism
and the sensory mechanisms
are
independent.
the inadequacy
I believe,to show
excellent experiments
the first,
of Smith and Wilson^'' were
of this second procedure.This fact led them to reconsider the entire notion
1'

M.

the

Smith

and

E. A. Wilson,

Psychol.
Monogr.,1953, 67, Whole

No.

359.

M.

DAVID

of the threshold

51

GREEN

and

alternative model, one


as an
theypresented,
very similar to that
footnote
Sec.
(
See
17.) Munson
suggestedby decision-theory
analysis.
especially IV,
and Karlin,*'^
the detection process
usingan information-theory
analysis,
investigated
under "absolute threshold conditions."
In order to deal with false positive
responses,
This model
is also very similar to that
they proposed a "discriminant level model."
suggestedby decision-theory
analysis.
The threshold model could stillattempt to account
for the data shown
in Fig.2.
The argument would
run
achieves some
hit and falseas follows:
Suppose the subject
If the situation is changed in some
alarm rate.
modify his behavior by
way, he can
simplygivingmore
"yes" responses. Since this guessingrate is independentof the
stimulus conditions (both noise and signal-plus-noise
below the threshold)
events
are
will
this
relative amounts,
both the hit and false-alarm rates.
increase,by the same
In short,a linear function will result. In the extreme, the subject
says "yes" all the time,
hence this linear function must
the
in
corner
point the upper right-hand
go through
\.00, P^^(A)
for the data is a collection
1.00]. Thus the threshold prediction
[P^{A)
of lines havingthe upper right-hand
the
and
as
corner
common
intercept, a slope
of the signal.No
linear function which
has this
depending upon the detectability
fit
value
than
few
of
the
data
for
of the
value
one
can
more
a
as
intercept
points
any
conflict with this version of
then, seriously
slope.The results of this first experiment,
of support to the decision-theory
the threshold model and givesome
measure
analysis.
=

conflict between

The

version

some

of the threshold

model

and

the

decision

analysishas been the subjectof considerable experimentaleffort. There are other


experimentalresults more
damaging to the threshold position.These experiments
attack the threshold
because they suggest that orderingbelow
the
concept directly
^^
threshold value is indeed possible.
We shall drop this conflict and proceedto other
questions.
Actual
in

Fig.2

wishes

one

to

optimum

discuss the

and

select

P(N)IP{SN), where
the

and

criterion

an

questionof the

optimum

obtain

criterion

P is the criterion value

of noise
prioriprobabilities

of course,

Let

criterion.

alone

on

on

us

now

optimum

likelihood

likelihood

return

to

played
the results dis-

criterion. It turns

out

that if

ratio, it is equal to ji
P(N ) and /*(SN) are
=

ratio and

We
can,
respectively.
signal-plus-noise,
criterion
the
slopeof
subject's
by measuring
data point.This rough comparison
experimental

and

of the

rough
pointnearest the
the
is displayedin Fig. 3. Note
that while there is a strong relation between
from
consistent
an
estimated and optimalcriterion values, there is also a
departure
the
summarized
The
trend
be
exact
subjects
saying
by
correspondence.
general
might
1
f!
conservative; they tend to adopt criteria which are not as diff"erentfrom
are
of the procedure.
be. This result is almost an inevitable consequence
as they should
of the
The way in which expectedvalues change for various criterion levels is the crux
A.
detail in Appendix
problem. This topicis discussed in more
have been utilized to vary the
Since these earlier investigations,
other procedures
successful
and is certainly
criterion. One which seems
more
straightforward
subject's
is simplyto instruct the subject
to adopt diff'erentcriteria such as lax or very
verbally
for Py{A)}^
maintain
instruct
the
a certain value
to
to
strict,or even
subject
the ROC

curve

at

measure

the

18
1^

W.

A. Munson

J. P.

Egan, A.

and

J. E. Karlin, J. Acoust.

I. Schulman,

and

G.

Soc. Am.,

1956, 26, 542.

J. Acoust.
Z. Greenberg,

Soc. Am.,

1959, 31, 768.

52

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

4.00

2.00
"aa
T3
"U

"E

LOO

0.50

0.20
0.10

0.20

0.50

1.00

2.00

10.00

5.00

Optimum /3
Figure
of the

Comparison
of the
equivalent
normal
assuming

optimum

and

criterion level

on

statistics for both

/"(SN) is the

obtained
likelihood

criterion
ratio.

levels.

This

The

criterion

level,/5,is the

criterion is obtained

optimum
by
where
hypotheses.It is equal to [1 i'(SN)]//'(SN),
of the signal.
a priori
probability
"

from the questionof the criterion


Let us turn now
of detectability.
the measure
of
of
another
to
analysis,
adjustment
detection-theory
aspect
this measure
remains
whether
and more
not
or
relatively
specifically,
detectability,
different
different experimental
invariant over
can
one
procedures.How
compare
is
different
obtained
measurements
using
experimental
procedures an important
question,
but for any scientific enterprise.
Let us review
not
only for psychoacousticians
has permittedsuch a
to which
the evidence
the extent
on
analysis
detection-theory
comparison. If we make the usual assumptionthat the distribution of likelihood is
with equalvariance on both hypotheses,
normal
as in the situation outlined in the first
is
d'.
of detectability
then the measure
experiment,
index
of this detectability
the applicability
A paper by Swets^" has considered
obtained
and
also
he
has
for yes-no and forced-choice procedures
;
compared predicted
results using
two, three,four,six,and eightalternatives in the forced-choice procedure.
failure
based on d' hold up remarkablywell. The worst
these predictions
In general,
1 db; no
trend is evident in the data.
consistent error
to be about
reportedseems
first
ROC
of generating
Another
method
suggestedby Swets et al}^ has
curves,
with the standard
al}^
method
this
tested and compared
been employed. Egan et
the decision-theory
observation
or yes-no procedure,
yes-no procedure.In the single
determines
and
this
criterion
claims
the
that
a "yes"
subjectadopts a single
analysis
"no"
then, is employing the subjectas a threshold
or
response. The experimenter,
after
could have the subject
the experimenter
device. Alternatively,
reporta number
could
such as likelihood ratio; from these numbers, the experimenter
each observation
ROC
construct
curve
an
by placingvarious criteria on the likelihood ratios reported.
The subject
The rating
procedureis a compromise between these two extremes.
in the rating
procedureis asked to placeeach observation in one of several categories;
Measure

and

of

its

2"

J. A. Swets, /. Acoust.

^^

J. A. Swets, W.

Michigan,Technical

Soc. Am., 1959, 31, 511.

P. Tanner, and

Report No.

T. G. Birdsall,Electronic

40, 1955.

Defense

Group, University

DAVID

M.

53

GREEN

signal's
presence, the next for a lesser degreeof
and so forth. ROC
constructed.
One can then comare
curves
pare
sureness,
subsequently
the measure
of signaldetectability
obtained
from these two
procedures,
yes-no
differed for his three subjects
and rating.Egan et alP found these two
measures
by
error.
0.3,0.4, and 0.1 db, differences probablywell within the experimental
In summary
decision analysis
have seen
how
allows one
to predict
then, we
within a fairly
wide range of psychoacoustic
procedures.The forced-choice procedures
using two to eightalternatives and a single-interval
procedureusing two to four
of
summarized
of detectability,
a
categories response can be
by a singlemeasure
is
invariant.
measure
which, for practical
purposes,
of
methods.
The
traditional methods
more
for psychoacoustic
Implications
utilize some
psychoacoustics
parameter of the signalsuch as the threshold energy.
This value is obtained by an analysis
of the subject's
responses. Many of these methods
allow
determine
do not
the subject's
criterion and in most
methods
it
to
one
directly
is presumed to be constant.
how variation in the subject's
Let us investigate
if it occurs, will affect
criterion,
the estimate of the threshold energy. Variation
of the subject's
criterion affects the
false-alarm rate P^{A). Figure4 shows
how
the probability
distribution for signalbe varied as the false-alarm rate P^{A) is changed to maintain
must
a constant
plus-noise
value of signal
detection P^^{A). We
have assumed
Gaussian
distribution and
The insert displays
the essentials
equalvariance to construct the solid line of the figure.
of the calculations and shows how a change in P^{A) of from 0.10 to 0.01 necessitates
of the signaldistribution from
1.3 to 3.1 in order to maintain
a change in the mean
the top

^sn(^)

one

beingused

of

sureness

This value oi

0-50.

"

for

P^^{A)

is a reasonable

since it is often

one

used

the

as

10"

10"^

10"

10

of how

change

P-^iA) is the false-alarm


constant
rate

at

0.5.

for various

The

rate;

mean

values

in criterion
a

will influence

"yes" response
signaldistribution
to

no

of the

oi P^{A).

The

constant,

10"

10

Figure
Evaluation

,-3

C,

the size of the "threshold"

signal.The

hit rate, P^^(A),

signal.

was

varied (see insert)to achieve

was

chosen

so

that 10

log 1.3

held

was

this hit

-|-C

0.

54

READINGS

estimate

of "threshold."

methods

this

control

IN

Very

PSYCHOLOGY

MATHEMATICAL

small values of false-alarm


of

parameter to the extent

rate

were

used

because

most

keepingit very low.

how this change in the mean


of the signal
distribution
generally
for
sinusoidal
in
signals noise, d' is
any signalparameter. However,
to signal
roughlyproportional
energy ; thus the "estimated threshold" may vary over a
the criterion of the subject.(In other experimentsd'
6-db range depending on
cussion.)
varies with signal
voltage hence the range might be 12 db. See Fig.7 and the disWe

cannot

is related

say

to

"

ject's
change in the estimated threshold, of say 6 db, will only occur if the subthat it is approximately
to assume
criterion changes. One may be willing
constant
^^
of the experiment.
Then this number, 6 db, could be interpreted
the course
over
The theory,
sets of different measurements.
as a tolerable difference in comparing two
view in psychoacoustics
then, is consistent with the rather wide-spread
; namely,that
results obtained
gruence.
usingdifferent methods should not be expectedto show exact conthese differences are largeenough to warrant
Whether
concern
dependsboth
of the problem and the precision
desired.
nature
the particular
on
The use of ROC
and speechresearch.
and the measure
Decision
curves
analysis
confusion
has been
d' has not been limited to detection experiments.Since some
of
d'
this
issue
deserves
attention.
the
some
multiplicity measures,
by
generated
taken from a reportby Egan.^^The similarity
ROC
an
curve
Figure5 displays
and Fig.2 is apparent,even
between
this figure
though measures
employed to construct
is presentedin
The procedurehere is as follows:
A word
this graph differ greatly.
This

noise to

writes down

listener who

he thinks

the word

was

He
presented.

then checks

The
conditional
he believes this identification response is correct.
whether
not
or
those words where he in fact was,
correct
of the receiver sayinghe was
on
probabilities
and

not

was

correct,

Egan's ROC

define the ordinate


curve,

rather than
contingencies

and

abscissa

then, is constructed

from

from

of Fig.5.
respectively

of response-response
contingencies,was the ROC curve
stimulus-response
a

table

as

is by no
the standpointof analysis,
means
decision
the
a
First,
really
by Egan
two-stage
process.
the
he
must
several
most
select
observer has to
word; second,
(from
possibilities)
likely
Such a process produces
evaluate this decision with respectto all other possibilities.
ful
doubtto evaluate exceptunder the most
mathematical
impossible
virtually
expressions
set of simplifying
assumptions.
does not, of course,
This difficulty
prevent one from summarizing the data
presentedin Fig.5 by a singleparameter. The line drawn to the data pointsis that
generatedby moving a criterion alongtwo normal deviates of the same variance which
labeled d' because of its
This measure
differ onlyin means.
initially
was, unfortunately,
d'
because the detection measure
It is unfortunate
analogyto the detection measure.
of signal
and noise. No
measurements
related to physical
has often been specifically
^^

Obviouslyone
of the order
probabilities
false-alarm
discussed
d'

rate

in the

1, could
-^

Report

J. P.
under

to

is

used

method

trivial. The

difference,from

This

earlier.
presented

10"^. If one

measurable

be used

as

The

because

constant

willing
to

make

10 \

signalenergy

or

this
use

Communication

measure
directly
raise the
must
assumption,
of
other
the
one
techniques
one

cannot

one

necessary

the counterpart of the threshold

Egan,Hearingand
contract, 1957.

is not

value, Py(A) "

previoussection.

then

it is

only assume

can

to

obtain

certain

d', say

energy.

Indiana
Laboratory,

Technical
University,

56

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

Theory of Ideal
In the most

ideal observer

is

simply a

function

an
relating
an
alreadyspecified
since
I
Table
this task. This is not
ideal observer for our simpleexample,
accomplishes
because
the
observations
in terms
were
however,
an interesting
alreadyspecified
example,
A
under
each
of
of the probabilities
ideal
an
hypothesis. more
interesting
example

observation

observer

the observations

arises where

differ under

are

ratio

Thus

waveforms

The
hypothesis.

each

likelihood

calculate

waveform,

an

of that observation.

the likelihood

to

the waveform
a

generalsense,

Observers

or

have

we

and

where

the characteristics of

task of the ideal observer


monotonic

some

is,then, given

transformation

of that

quantity.
need
ideal observer,strictly
speaking,

The

problem of what

ratio is computed, the

not

decision

any decisions.

make

rule to

If likelihood

is determined

employ
by the
have
cussed
been dismaking
possible
objectives
in the previous
could be
sections,where it was pointedout that these objectives
attained by usinga decision rule based on likelihood ratio. Although the calculation of
the ideal observer for a givenproblem,such information
is of
likelihood ratio specifies
of
littlevalue unless we can evaluate this observer's performance.One general
method
/?0C curves, but to obtain
the ideal observer's performanceis to determine
evaluating
Thus
evaluate completely
the
to
calculate two
ROC
must
we
an
curve
probabilities.
likelihood is calculated but the
ideal observer we actually
have to specify
not only how
likelihood
ratio
both
distribution
of
on
hypotheses.
probability
Having established the generalbackground of this problem,let us consider a
specific
example: the ideal observer for conditions of a signalwhich is known exactly.
in
objective
specific

Ideal observer

must

Hi

known
for the signal

Various

exactly(SKE)

define this special


case
hypothesesactually
of the following
select one
hypotheses:

Two
one

the decisions.

"

the

waveform

bandwidth

//g
"

is

{W) and

the waveform

sample

of white

noise power

Gaussian

noise

givena waveform,

n(t) with

specified

density(A'^q)-

sit).Everything
specified
signalwaveform
its starting
time, duration, and phase. It
known
sine wave
as
i.e.,
long as it is specified,

is n(t)plussome

is known

about

need

be

not

in which,

s{t)if it occurs:

segment

of

exactly.
derive
hypotheseswe wish to calculate likelihood ratio,and, if possible,
distribution of likelihood ratio on both hypotheses.Obviouslysuch
the probability
calculations will be of littleuse unless the final results can be fairly
simplysummarized
such isthe
and noise. Happily,
of signal
in terms
measurement
of some
simplephysical
From

these two

case.

We

shall not

present the derivation

here

since it is not

in itself particularly

assumptionof the derivation will,


raised;
an
assumptionhas been recently
objection
of applyingthis result to any
which
an
objection
seriously
questionsthe legitimacy
native
the alterconducted.
been
which
has
Unfortunately,
psychoacoustic
experiment
yet
assumptionsuggestedhas a different but equallyserious flaw.

instructive and

can

be obtained

however, be discussed,since

elsewhere.'^ One
to

this

DAVID

57

GREEN

M.

of the waveform
Representation
The

assumptionconcerns

likelihood

ratio,one

must

Since the waveform


hypothesis.
a

with
probability

various

we
probabilities

of waveforms

we

must

limited.

set

somehow

of

measures

to

In

order

pute
com-

each

on

associate

from

the

form
wave-

measures.

to

these

compute

assumptionsabout
very specific

some

Fox^

If the waveform

assumed

that the waveforms

is of this class It

where

ff is the "bandwidth"

in terms
representation

of course,

of the waveform?

nature

make

obtain

these

series

measures,

In order

waveform

the class

will consider.

Peterson, Birdsall,and
band

somehow

or

with
probability

exactly is the

what

But

of the waveform.
representation
of a certain
probability
is simplya function of time,one must
find the

this waveform,

associate

and

the

can

of the noise and


of sine and

be

representedby

Tis the duration

cosine

of

Fourier

were

be

might
the
identify

//

series-

of the

used.

2WT

form.
wave-

There

are

n parameters, but
equivalent
writingthis series to
ways
if the original
waveform
is indeed Fourier series-band limited,
the waveform
in the interval (0, T). Acceptingthis assumptheywill reproduceexactly
tion,
find that a monotonic
transformation
of likelihood ratio (the logarithm)
is
we
under both hypotheses.
normal

these

are

many

all unique,and

is normal
H^: log/(x)

with

mean

"EJNq,

variance

EjN^^,

is normal
H2,'.
log/G^)

with

mean

+EINq,

variance

EJNq,

where
(IE/Nq)^

is the signal
dt, and Nq is the noise
energy, J^[5(/)]^
if
this
about
waveform
is not made
the
the
density.Naturally,
assumption
power
David^"
result
is
invalid.
Mathews
and
have
considered
different
a slightly
preceding
Fourier integral-band
the waveforms
limited. The
are
assumption. They assumed
conclusion
from
this
is
that
the
is
detectable
in the
resulting
assumption
signal perfectly
In short, d' is infinite
noise independentof the ratio "'/A^0'
^s
long as it is not zero.
for any nonzero
value of E/Nq. Which
of these assumptionsis the more
reasonable or
to a psychoacoustic
applicable
experiment?
all psychoacoustic
In almost
Neither assumptioncan
be completely
justified.
noise
tube.
The
t
he
is
a
experiments,
voltage
voltage actually
produced by special
and filtered. Such noise is not Fourier series-band
produced by this tube is amplified
^^
not
limited,for the noise is clearly
periodic.
Although a Fourier series might serve
d'

AMjo

3"

M.

^^

It is somewhat

V. Mathews

and
unfair

E. E. David, J. Acoust.
to

Soc. Am.,

1959, 31, 834(A).

implythat Peterson, Birdsall,and


that each waveform

Fox

assumed
could

be

the noise

was

represented
by
assumption,
strictly
speaking,
is througha samplingplan,which
The way theyobtained these numbers
a finite set of numbers.
discuss in detail. It was
cannot
not
we
a simpleFourier
expansionin terms o^ sine and cosine.
footnote 7;
This is a difficultand complex topic;for a discussion of the details in this area
see
in
Gaussian
Trans.
detection
of
Gaussian
noise,"
D. Slepian,
the
"Some
comments
on
signals
L.
Random
in
W.
Root,
IRE, PGIT-4, 65 (1958); and W. B. Davenport and
signals noise. New
of the situation where the noise is filtered,i.e.,
York:
McGraw-Hill, 1958. Precise analysis
in principle.
The
be worked
out
where the power spectrum of the noise is a polynomial,
can
One can
be obtained
is complex and exact
can
answers
only in certain simplecases.
analysis
of the signal
is finite.
situations the detectability
show
in general,however, that for practical
(See Davenport and Root.)

Their
periodic.

was

58

READINGS

as

an

be

an

IN

MATHEMATICAL

PSYCHOLOGY

excellent

in the interval (0, T), it would


not
approximationto these waveforms
of the waveform.
an
representation
Similarly,
assumptionof a Fourier

exact

limitation of the bandwidth


be correct, because the waveform
cannot
does not
integral
have a sharp cutoff in the Fourier integral
If
it
the
waveform
would
be
sense.
did,
If it were
the ideal observer could sampleat one
analytic.
analytic,
pointin time, obtain
all the derivatives at that point,
and know
the exact form of the wave
for all time. Such
result
leads
the
conclusion
that
the
ideal
to
a
observer, by observing
one
sample of the
waveform
at any time can, immediately,
in principle,
his decision about all the
make
waveforms
the experimenter
has presented
in the pastand all those he may ever
decide
to produce. This approach is therefore of littlepractical
use.
The
has indicated
issue,while obviouslyonly an academic
one
one,
very
of
the
important
problem. The ideal observer is,like all ideal concepts,only as
aspect
good as the assumptionsthat generateit. Clearly,
any such idealization of a practical
situation is based on certain simplifying
assumptions.It is alwaysextremelyimportant
to understand
what
these assumptionsare and even
more
importantto realize the
of a changein these assumptions.In short,there are many ideal observers,
implications
each generatedby certain key assumptionsabout the essential nature
of the detection
task.
For

the discussion

approach and
number
As

follows,we

that the waveforms

assume

of measurements.

shall
can

A similar treatment

progress is made

more

which

with the

theoryof

be

use

completelyrepresented
by

is givenby Van
ideal observers

we

Meter

detection will vary if certain definite restrictions


the observer operates.Peterson, Birdsall,
and Fox

in which

several such

and

their results. Each

and

should

how
quiteprecisely
manner

Fox'

the Peterson, Birdsall,and


a

finite

Middleton.^^

be able to state

imposedon

are

the

have, in fact,considered

providesus with a framework


parison
performanceof the subject.Such a commay
both
aud
provides
qualitative
quantitative
guides for further research. ^^
There are several areas
we
might select to illustrate this approach. The one we have
selected was
chosen because it is a general
what
sometopicand because it has been slighted
in psychoacoustics.
from

which

cases

evaluate

we

and

assess

case

the

Shape of the psychophysical


function
The
of

function is generally
defined as the curve
the perpsychophysical
relating
centage
detections of the signal
correct
of the
(the ordinate)to some
measure
physical

variant of the constant


stimuli method
is used, the curve
signal(the abscissa). If some
rises monotonically
from zero
hundred
level
is increased.
the
to one
signal
percentas
Generally,
hypothesesabout the form of this function arise from assumptions
about the process of discrimination.
Often these assumptionsare sufficient to allow
deduce
the
form
of
function to within two or three paramto
the psychophysical
one
eters
which are then determined
Obviously,it is extremelyimportant
experimentally.
for the model to specify
stimulus which is used
the exact transformation
of the physical
the theory
as the abscissa of the psychophysical
function; without such specification,
is

incomplete.
In

there
psychoacoustics,

this function.
"-

D. Van

33

W.

Most
Meter

P. Tanner

has been

theories of the
and
and

D.

littleconcern
comparatively
auditoryprocess have been content

with

Trans. IRE, 1954, PGIT-4, 119.


/. Acoust. Soc. Am., 1958, 30, 922.
Birdsall,

Middleton,

T. G.

with the form

of

attempting

DAVID

M.

59

GREEN

or
curve,
usuallythe mean
parameter of the psychophysical
to obtain from the literature information
result,it is nearlyimpossible

predictonly one

to

threshold.

As

the actual form

of the

function.
psychophysical
is the neural-quantum
exceptionto the precedingstatement
The authors of this theorysay that it "enables us to predict
the form and
hypothesis.^*
from the model
the slopeof certain psychometric
functions."
It can
be demonstrated
to
that the form of the function should be linear and this linear function is specified
within one parameter. The physical
is never
mentioned
in the derivation of the
measure
that sound pressure and frequency
theoryand we find onlyafter the data are presented
The authors remark
in their paper that "strictly
are the appropriate
measures.
physical
rectilinear
functions
data
when
againstsound
speaking,
yielding
psychometric
plotted
when
expressedin terms of sound energy,
pressure do not show absolute rectilinearity
but calculation shows that the departure
from rectilinearity
is negligible."
It is certainly
and indeed pressure cubed, are all nearlylinear
true that pressure, pressure squared,
for small values of pressure
but that is not entirely
the point.
function
that
It is the location of this
playsa crucial role in the theory.If the
physical
subjectemploys a two-quantum criterion then, accordingto the theory,the psychoon

The

notable

"

function
hundred

percent at

Where

the

be

must
two

breaks

curve

zero

up

to

quantum unit,show

one

linear increase to

one

quantum units.
quantum units,and maintain this level for more
hundred
it
reaches
and
where
from zero
one
percentreports

if the subject
requires
specified
by the theory.In general,
percentreportsis precisely
extend
from
function
must
linear
the
to
a
increasing
produce positive
report,
quanta
be
to
units.
what
Now
io n + \ quantum
a two-quantum subject
n
clearly,
appears
(0% at one pressure unit, 100% at two pressure units),when the data are plottedin
n

in energy units.
as a two-quantum subject
interpreted
This is true no
as an
any-number-of-quantumsubject.
interpreted

pressure units,cannot
he cannot
how

be

be

small the values of pressure.


This criticism of the rather post hoc

In

fact,

matter

physicalscale is by no
limited to the neural-quantumhypothesis.Many hypothesesabout the shape
means
formulations ofthe Gaussian
ofthe psychophysical
function,including
some
hypothesis,
this rather crucial factor.
neglect
based
with these theories. Models
contrast
Detection theorystands in marked
function
the ideal observer concept predict
the form of the psychophysical
on
exactly.
free
and
there
no
are
The proper physical
dimensions
are
eters.
paramspecified
completely
of the

treatment

less
observers somewhat
to find human
Obviously,one would not be surprised
least
function
at
of
the
the shape
might
psychophysical
optimum, but hopefully,
Often however, the obtained
from the model.
to that obtained
physical
psychoparallel
this
and
model
the
function does not parallel
that predicted
discrepancy
by

than
be

deserves

discussion.

some

Signaluncertaintyand
of correct
3*

^^

to

some

detections

S. S. Stevens, C. T.
The
ideas

the essentials
conversations

ideal detectors.^'" In
in

two-alternative

are

the

same.

on

this

topic.

The

we

have

forced-choice

plottedthe percentage
versus
procedure

1941, 54, 315.


Psycho!.,
isverysimilar
data from
uncertainty
signal
of
the
details
differ,
P. Tanner.
analysis
Althoughseveral
for many
author is indebted to Dr. Tanner
long and lively

Morgan, and

of detection
analysis
expressed
by Dr. W.

Fig.6,

J.

Volkmann,

Am.

the viewpointof

J.

60

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

1.00

"13

40 spl

Frequency:1000
Duration: 100
I

0.50

-4-2

theoretical

The
signals.

parameter

test.

The

obtained

and
typical
subject

is simplyto detect
We

12

14

16

sinusoidal

of

ordinate

data

are

10

"

1 of M-orthogonal
detecting
The ideal detector
possible
orthogonalsignals.
it.
abscissa
The
is
times
the logarithmof signal
ten
identify

is the number

not
only detect the signal,
The
density.
energy to noise-power

"^-^ofor

functions for the ideal observer


psychophysical

need

forced-choice

10

Figure
The

cps
ms

is the percent correct

compared with the


db to the right.

series of mathematical

signaladded

to

detection in

two-alternative

theoretical function

models.

The

shifted about

problem in

background of white

all cases

noise.

subject"because
"typical

the shape of this function is remarkablyinvariant


and a range of physical
durations
For
subjects
signal
parameters.
of 10 to 1000 msec^^ and signal
from 250 to 4000 cps,^^
there appears to be
frequencies
no
greatchange in the shape of the function when plottedagainstthe scale shown in
the exact location of the curve
Fig.6. Naturally,
dependson the exact physical
parameters
of the signal,
but exceptfor this constant, which is a simpleadditive constant
in
of
this
function
form, the shape is remarkablystable. The striking
logarithmic
aspect
is its slope.We
notice the slopeof the observed
function is steeperthan most
of the
theoretical functions depicted
in Fig.6.
The class of theoretical functions is generatedby assuming the detector has
^^
various uncertainties about the exact nature
of the signal.
Each function is generated
the
detector
knows
will
that
the
be
of M-orthogonalsignals.
one
by assuming
only
signal
If the signal
is known
For sinusoidal signals,
1)there is no uncertainty.
exactly(M
the nature
of the uncertainty
be
time
of
of the signal,
occurrence
or
might
phase,
The
of
reflected
this
is
the
As
signalfrequency.
degree uncertainty
by
parameter M.

say

both

over

was
assume

38

D.

"

D. M.

^^

The

M.

Green, /. Acoust. Soc. Am., 1959, 31, 836(A).


Green, M. J. McKey, and J. C. R. Licklider,J. Acoust. Soc. Am., 1959,31,1446.

details of this model

selected because

it has

been

may

be found

presentedin

but which
signaluncertainty
functions producedby these

the value of the parameter (M) would

in footnote

the literature.

differ in details about


models
be

are

7, p. 207.
There

the decision

similar to those

changed somewhat.

This

are

model
particular

other
rule.

models
The

which

psychophysical
in Fig.6, although
displayed

DAVID

M.

61

GREEN

function
increases,the psychophysical
uncertainty

slope. It therefore
with sufficient uncertainty
about the signalto
appears that there may exist a model
function
which
is
similar
the
human
observer.
that
to
a
displayedby
generate
very
for
the
of
the
human
the
that
the
moment
extreme
Accepting
assumption
slope
observer's psychophysical
function is due to some
about the
degreeof uncertainty
to
this
various
w
e
signal, might try
manipulate
slopeby
experimentalprocedures.
Preview
technique.One generalclass of procedureswould attempt to reduce
the uncertainty
form of cueingor
by supplyingthe missinginformation throughsome
for
previewtechnique.If, example,the observer is uncertain about the frequencyof the
the signal
at a
signalwe might attempt to reduce this uncertainty
briefly
by presenting
level
the
observation
interval.
of
to
if
the
time
of
occurrence
high
justprior
Similarly,
the signal
is uncertain we might increase the noise duringthe observation
interval. If
the noise was
increased for all trials,
whether
the signal
it would
or not
was
presented,
tion
provideno information about the signal's
presence but would convey direct informatime
and
duration.
about the signal's
Both
of
have
been
these
starting
techniques
utilized with only partial
While
it
is
there
was
no
success.
impossibleto assert that
the amount
of change was
in
the
small,
change (the null hypothesis)
although
very
increases

in

direction. ^^

proper

procedureswhich has been utilized to attempt to reduce the


the signal
parameters involves changingthe detection task
that some
information
is directly
so
supplied.The proceduresare like the preceding
interval. For example,to
but actually
include the information
in the observation
continuous
sine
add
to the noise. The
a
wave
remove
we
might
frequencyuncertainty,
continuous
evident in the noise.
is adjusted
sine wave
to a level such that it is clearly
ment.
The signal
is an increment
added to this sine wave
and the task is to detect this increThe
psychophysical
proceduredefinitely
changes the slope of the subject's
function
it becomes
is easier to detect.^''
less steepand the signal
sine wave
This procedureof making the signal
increment
to a continuous
an
but does not
remove
temporal uncertainty.
providesgood frequencyinformation
minimizes
is
Another
which
all
procedure
practically uncertainty in fact a modification
A two-alternative
of a standard
the j'.n.d.
for intensity.
procedureused to investigate
at
in noise, one
sinusoids
forced-choice
Two
is
occur
gated
procedure employed.
task is to select
The subject's
standard
level,the other at this level plusan increment.
to a power
is
the interval containing
If the standard
the increment.
adjusted
signal
function
level about
the psychophysical
actually
equal to the noise-powerdensity,
case."*^ It is from 3 to 6 db off
that expectedfor the signal-known-exactly
parallels
optimum in absolute value, depending on the energy of the standard. (See Fig.7.
Note the change in scale between
Figs.6 and 7.)
Let us, at least tentatively,
acceptas the conclusion of these last results that the
to various uncertainties
function is in fact due primarily
shape of the psychophysical
about the signal
parameter. If this is true, then we stillhave the problem of explaining
Another

class of

about
subject's
uncertainty

"

Unpublishedwork of the author. Also see T. Marill, Ph.D. thesis, Massachusetts


logical
Technology,1956, and J. C. R. Licklider and G. H. Flanagan, "On a methodoin
problem audiometry,"unpublished.
W. P. Tanner, J. Bigelow,and D. M. Green, unpublished.
"
of Michigan,Technical Report
W. P. Tanner, Electronic Defense
Group, University
3*

Institute of

*"

No.

47. 1958.

62

READINGS

the lack of

IN

evidenced

MATHEMATICAL

when

the

PSYCHOLOGY

previoustechniqueswere

employed. Should
not
a
certainty?
signal,
precedingan observation, serve to reduce frequencyunThe answer
but not
might be that such proceduresdo reduce uncertainty,
stillremaining.From
enough relative to the uncertainty
Fig.6 we note that, as we
introduce
the slopeof the psychophysical
function increases very
signaluncertainty,
in
for
small
the
t
hen,
as
increases,the slope
rapidly
changes uncertainty;
uncertainty
value.
A
in
M
from
256 to 64 may
change uncertainty
approachessome asymptotic
function.
This fact also probablyexplains
hardlyaffect the psychophysical
why the
functions
do
not
for
to
much
of
a
psychophysical
change very
variety signal
appear
duration and signal
frequency.Undoubtedly,as the signal
parameters, such as signal
duration increases,the uncertainty
about the time of occurrence
of the signalis reduced.
Due
the
initial
this
is
to
large
uncertainty, change too small to be detected in
success

previewof the

the data.
of checkingthis general
signalfrequency.Still another manner
model is to vary the uncertainty
of the signal
and determine how this affectsthe subject's
One
for
select
several different sinusoidal signals
and
performance.
might,
example,
select one at random
used on a particular
trial. The subject
is simplyasked
as the signal
it. Depending on the frequencyseparation
to detect a signal,
not
and the
identify
of signals
number
used, one can directly
manipulatesignaluncertainty.
Uncertain

1.00

0.90

0O.8O

0.70
(2!

0.60

0.50
-10

-5

10

Figure
Observed
The

data in the A/

abscissa

abscissa.

The

and
two

of the

two

curves
curves

are

the

differ
at

same

by

low

level of the noise; the lines show

6 db

values

as

at

in
each

observer (M
signal-known-exactly
but
the
note
Figure6,
change in scale

/experimentand

versus

ordinate

15

the

value

of percent

of

percent correct.

correct

the level of / in power,

is
and

The

The
illusory.
the maximum

1).

of the

apparent

insert shows

vergence
con-

the

/ -)-A/ power

64

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

functions obtained with this type of signal


also slightly
are
psychophysical
steeperthan
those predicted
time uncertainty
stillremains or signal
by the model.^'' Either partial
alone is not a sufficient explanation.
The author feels that a better model
uncertainty
would

that the human

assume

assumption,
coupledwith
the results obtained

the

thus far.

observer

utilizes

some

nonlinear

detection rule.

This

could probablyexplain
of
most
uncertainty
explanation,
The mathematical
of such devices,is however,
analysis

complex.
Internal
Often

it is

Before

noise.

summarizing, one

final

be

point must

considered.

cussing
temptationto invoke the concept of internal or neural noise when disbetween
ideal model
and the human
observer.
an
There are
discrepancy
for avoidingthis temptation.While
it would
take us too far afield to
good reasons
this pointin detail,
the following
remarks
will illustrate the point.
cover
Only if the model is of a particularly
simpleform can one hope to evaluate the
effects of the assumptionof internal noise. The signal-known-exactly
observer
specific
is of this type. Here one
show how a specific
of
internal
noise
can
can
simplybe
type
treated as addingnoise at the inputof the detection device. Thus one can evaluate the
function and it will be shifted to the right
psychophysical
by some number of decibels
due
internal
the
to
noise. But, of course, such an assumptioncan immedi(see Fig.6)
ately
be rejected
since no shift in the psychophysical
function can account
for the data
in the figure.
displayed
With more
difficultto say exactly
what internal
models, itis usually
complicated
noise will do. While it will obviously
lower discrimination,the specific
effects of the
effects can
be
assumption are often impossibleto evaluate. Unless these specific
the
evaluated, the assumption
of
the
simplyrephrases
original
problem
discrepancy.
I am
not suggesting
that the human
observer is perfect
in any sense, nor attempting
the importanceof the concept of internal noise. What
to minimize
I am
ing
emphasiza

the

is that the

concept must

importanceit must

be made

what

this noise is,i.e.,


that

what

way

it interacts with

what
specifically
out

be used

of the

the detection
on

If the

or

discrimination

performance. Unless

concept

is to have

any

process, and
these

steps

can

(3) evaluate
be carried

assumptionvitiates its usefulness.


Summary

The

great care.

This impliesthat we
have to (1) state exactly
specific.
have to characterize it mathematically,
in
we
(2) specify

effect it will have

the ad hoc nature

with

and

Conclusion

main

emphasisin this paper has been to explaindetection theoryand to


such a theory has been appliedto certain areas
illustratehow
of psychoacoustics.
This method
of analysis
is simplyone of many that are currently
beingused in an attempt
to understand
the process of hearing.
Two
main aspects of this approach have been distinguished.
The first,
decision
that
the
criterion
well
the
of the
as
as
emphasizes
theory,
subject's
properties
physical
stimulus playa major role in determining
the subject's
responses. The theoryindicates
both

the class of variables

which

determines

the level of the criterion,and,

importantly,
techniquefor removingthis source
suggestsan analytic
techniqueleaves
invariance

of this

relatively
pure
measure

over

measure

several

D.

M.

Green, /. Acoust. Soc

This

of the signal.The
detectability
psychophysical
procedureshas alreadybeen
of the

demonstrated.
*'

of variation.

more

Am., 1960, 32, 121.

M.

DAVID

65

GREEN

1.00

0.80

Si 0.60

1 0.40

O0.20

0.00

0.20

0.40

0.60

Figure
The

normalized

curve

based

on

expectedvalue as a function
the data presented
in Figures
2
to

The
detail. The

second

aspect,the theoryof

usefulness of such

function.
psychophysical
model

even

The

No

of

discussed in

illustrated by considering
the form

some

of the

providesa completeor comprehensive

that
psychoacoustics

hypothesesand

we

have discussed in this

standard

which experimental
against
It is too earlyto attempt any completeevaluation of
results can be evaluated.
and the application
this approach. The mathematical
of these
models are relatively
new
models to a sensory process began with Tanner
and Swets'*^only about five years ago.
There remain
and experimental
of a mathematical
be
solved
both
to
many problems
should
become
As more
nature.
more
specific
progress is made in both areas, the theory
with the research
and concrete, then perhaps it will be able to interact more
directly
from several other areas
in psychoacoustics.

paper.

providesa

source

of

curve.

ideal observers,has also been

analysis
areas

changesin criterion. This is a theoretical


3. The appendixliststhe assumptions
used

the

ideal observer

for the rather limited


model

and

was

an

1.00

of

construct

0.80

Appendix A
of comparing the optimum criterion value and that
difficulty
function. Let us investigate
i
s
the
employed by
subject
shape of the expected-value
in detail a typical
situation. We have assumed
that the distribution on likelihood ratio
is normal
under
both hypotheses,
that the mean
separationis one sigma unit, and
The

inherent
the

that the values and

costs

of the various

decision

alternatives

are

all the

same.

From

assumptionswe have constructed Fig.9. This figureshows how the expected


of signal
PCSN) and false-alarm rate
changes in a prioriprobability
values oi a priori
that for extreme
P^{A). We see immediately
probability,
e.g..P(SN)
behavior
0.004] and
0.10, the difference between optimum expected-value
{P^{A)
these

value varies with

"

W.

P. Tanner

and

J. A. Swets,

Rev., 1954, 61,


Psychol.

401.

66

READINGS

tends

force

to

to

the

the

On

to

instructed

are

0.50
Thus

will

any

correspondence

June

23,

achieve

attempt

1960.

one

if

least

at

to

3 %.

strategies

moderate

see

we

90%

of

and

in

moderate

more

any

the

maximum

any

in

curves

the

of

of

criteria

than

Since

for

were

most

are

in

employed
a

this
tions.
condi-

extreme

within

P^(A)

expected
more

figure

experiments,

P^(A)

probabilities

value

the

maximum.

psychoacoustic

priori

optimum

the
of

values

that

in

fact,

In

location

the

see

investigate,

obtained

PSYCHOLOGY

than

to

pure

more

0.50],

less

adopt

to

hand,

between

is

allow
avoid

subject

other

0.000]

to

[e.g., P(SN)

experiment

Received

exaggerated

subjects

0.15

[P^(A)

strategy

pure

somewhat

MATHEMATICAL

IN

range

the

from

payoff.
correlational

appears

extremely

sense,

the

difficult.

SOME

COMMENTS

AND

"PSYCHOACOUSTICS

AND
David

department

of

economics

massachusetts

Dr. S. S. Stevens
acoustics

misunderstand!

ng

by

laboratory

of

electronics,

massachusetts

cambridge,

the percentageof
relating
stimulus

often called the

more

research

technology,

items in my

paper,

"Psycho-

in order

comment

to

avoid

of the
physicalintensity
is

Green

kindlypointedout two
Theory,"^that requirefurther

I called the function

function

THEORY"*

has very

Detection

and

of

OF

DETECTION
M.

and

institute

CORRECTION

the

detection responses to the


It is true that this

correct

function.
psychophysical

psychometricfunction, a

probablyintroduced

term

in 1908.2

Urban

added

successive

just-noticeable-differences
(jnd's)to
and the physical
intensity
relation is commonly called the psychophysical
of the stimulus. The resulting
function.
Since Fechner's
other techniques
for determining
this relation have been
time many
functions (e.g.,
devised and the results are also called psychophysical
Stevens' power
methods
do not involve determining
law^). The newer
jnd'sand are not obtained by
classical
methods
of
variant
of
the
usingany simple
psychophysics.We are therefore
obtained
faced with the anomaly that psychometricfunctions
are
by usingpsychophysical
functions are now
determined
different
methods
and psychophysical
other,
by
techniques.
used in vision
curve
PersonallyI find the designation
frequency-of-seeing
function.
minology
distasteful
than
the
term
Some
even
more
psychometric
change in terwould
welcome.
be most
I am
open for suggestions.
Fechner
Originally

determine

the relation between

up

the

magnitudeof sensation

"

The

second

item

quantum theory. I

is

crucial and

more

concerns

asserted that data that appear

"

remarks

my
to

indicate

about

the

two-quantum

neuralobserver

be interpreted
as
plottedagainstpressure units cannot
any kind of quantum
observer when
units.
There
is,however, a very straightforward
plottedagainstenergy
this assertion incorrect.
of the scales of pressure and energy that makes
interpretation
and
this
occurred
to
I therebydid injustice
had
never
me,
Unfortunately, interpretation
and
to the authors
of the neural-quantumtheory. Let me
explainthis interpretation
when

the scale of pressure and energy units that I had in mind when
have a continuous
In the neural-quantumprocedure we
itthe standard).
(call
and

the observer's

At

times
specific

measure

of this letter
preparation

Amer.,

Office and
(Operational
Applications

Naval

Research.

F. M.

M.

This is Technical

Green,^ /icoM5^

this sinusoid

the pressure of the


the
increment, call it
plus

measure

we

Reprintedwith permission.
supportedby the U.S. Army SignalCorps,the

was

Force

D.

stimulus

1961, 33, 965.

/. Acoust.

If

the pressure of the standard

From

Soc.

remarks.

the amplitudeof
briefly

task is to detect these increments.

standard, call \ip,and


The

increase

we

my

sinusoidal

I made

Note

Office of Scientific Research), and

No.

Soc. Am.,

ESD

TN

Air

the Office of

61-56.

1960, 32, 1189.

Urban, The application


of statistical methods

Clinic Press, 1908, p. 107.


Philadelphia:
Psychological
S. S. Stevens, Psychol.
Rev., 1957, 64, 153.
3

67

to

the

problemsofpsychophysics,

68

READINGS

p + A^, then by
values of A^. We

IN

former

the
subtracting
call this

may

if we
Similarly,

PSYCHOLOGY

MATHEMATICAL

from

the latter

we

the increment
quantity/S.p

obtain

on

pressure scale

of pressure.

the power of the standard, a quantity


to
proportional
increment
standard
the
a
to
plus
quantityproportional

measure

of the
and the power
p"^,
{p + ^pT, we might subtract

the former

from

obtain

the latter,
and

(since the

constants

of

quantity,(/"^ 2^pp
A^^ ^2j
to energy, since the increment
proportional
is of constant
the increment
of energy. The
duration, and we may call this quantity
isthat
these
result
the
increment
in
two
and
the increment
quantities,
important
pressure
in energy, are nearlylinear for values of Ap much
less than p. If some
data are exactly
of the neural quantum theoryon one scale,theywould
consistent with the predictions
very nearlybe consistent on the other scale.
When
data plottedon
I made
a scale of signal
my remarks, I had in mind
I
the
waveform
added
the
to
standard
or
that
signalenergy. By signal mean
pressure
the observers are asked to detect. In this terminology,
the pressure of the signalis
and
the
of
t
o
i
s
to that quantity
proportional Ap
signal proportional
squared,
energy
data
scale
of
defined it are in
Ap^. Only
plottedon a
signalpressure as I have now
of neural quantum theory.
agreement with the predictions
for my oversight
Part of the reason
undoubtedlyarose from the fact that this
of signal
of the data
measure
some
energy Ap^ is the quantityI used in presenting
later
in
There
inherent
for
is,however,
no
reason
reported
usingmy particular
my paper.
the

proportionality
(lApp + Ap^).The

are

would

be

some

the two

cases

exactlythe
be

an

same.

increment

and

I should

have

made

my

reference

clear.

different scales of energy obtained from the pressure scale


This would
herent;
happen if the standard and signalare inco-

that is,if the middle


this would

the

is also
quantity

of the stimulus

measure

In

latter

same)

term

in the square of {Ap + p) is zero.


An example of
In the case at hand, this is not true and the

in white noise.

that I called signal


quantitythat I have called increment in energy and the quantity
different.
are
quite
energy
The general
is that the neural-quantum
to make
pointI was trying
theorydoes
in advance
how
the physical
stimulus should
be measured.
not specify
It was
my
it
is
for
t
hat
of
how
the
t
o
a
position
theory psychophysics specify
important
physical
scale is related to the expectedpsychological
results. This position
is apparently
not
endorsed.
I
with
the
of
number
theories
that
am
widely
particularly
impressed
suggest
that the psychometricfunction
is Gaussian, log-Gaussian,
or
Poisson, rectilinear,
but cannot
in advance
what particular
transformation
of the physical
logistic,
specify
scale will yield
these results. It is not hard to envision different circumstances
in which
all these assertions

experimental

error.

are

true

Somehow

at

least in the
there

never

sense
seems

that deviations
to be

are

within the range of


to these different

any resolution

findings.
One

can, of course,

simplyignoreall this and go on measuringonlyone arbitrary


psychometricfunction such as the "threshold" value. While this
parameter
positionobviouslyhas the merit of convenience, it would also appear importantto
demonstrate
how all of these different results might come
about from one single
general
To
latter
task one
have a theorywhich carefully
must
theory.
accomplishthe
specifies
the physical
theory.
part of the psychophysical
of the

Received

April14, 1961.

THE

ON

POSSIBLE

PSYCHOPHYSICAL
R- DUNCAN

LUCE

University

Harvard

This

is concerned

paper

century-old effort
functional

relations that hold

subjectivecontinua
continua
them.

that
The

are

It rests

upon

the

physical

not

the

mathematical

to

easilythe most

specifythe

interval

an

what
By
problem someessentiallyby replacing the
what
jnd assumption with the someequalthat
condition
"equally
stronger

sible
pos-

recastinghis

often noticed differences

sity
inten-

physicalcontinuum

generate

"

made

varies with

sufficient to

scale.

by Fechner.
empiricalknowledge of
was

discrimination

along the

between

(1958) have pointedout that Fechner's


reasoningwas not sound.
other
Among
things,his assumption is

the

presumed to underlie

influential,
attempt

how

and

and
first,

relations

with

determine

to

LAWS

when

and

always

or

are

never

able to show

cept
equal,ex-

noticed"

"

an
they
matical
scale results,and to present a matheassumption that jnd's are
tinuum.
subjectively
expressionfor it. Their work
equal throughoutthe conhas no practical
nation
When, for example, discrimiimport when Weber's
is proportional to
intensity law, or its linear generalizationA-r
the loga+ b, is true, because
ax
(Weber's law), Fechner claimed that
rithm
is
the
the equalleads
still
but
their
to
a
solution,
jnd assumption
jnd
relation (Fechner'slaw). scale differs from
Fechner's
logarithmic
integral
law is replacedby some
This idea has always been subject when Weber's
other
but
function
attacks
to controversy,
recent
relatingstimulus jnd's
upon
it have been particularly
At
to intensity.
severe.
the theoretical level,
Luce and Edwards
At
the
empirical level, Stevens
(1956, 1957) has argued that jnd'sare
^ This
work
has been supported in part
size on intensive,
unequal in subjective
by Grant M-2293 from the National Institute
calls
what
he
continua
of Mental
a
or
Health
and
in part by Grant
prothetic,
NSF-G
5544 from
the National
Science
contention
supported by considerable
were

interval

that

the

upon

"

Foundation.

data

"

and that the relation between

the

Ward

Edwards, E. H. Galanter, Frederick


Mosteller,Frank
Restle, S. S. Stevens, and
Warren
Torgerson have kindly given me
their thoughtful comments
drafts of this
on
of which
are
incorporated into
paper,
many
this version.
I am
particularlyindebted to
S. S. Stevens
and

for his very

and
subjective

function ax^, not

Using

such

"direct"

estimation
he

and

Stevens

two

and

article

appeared

in

"

Psychol. Rev.,1959, 66, 81-95.


69

the

logarithm.

methods
ratio

is the

as

nitude
mag-

production,

others

(Stevens: 1956, 1957;


Galanter, 1957) have accumulated

considerable

drafts.
This

power

detailed substantive

criticisms of the last


stylistic

physicalcontinua

evidence

Reprinted with

to

but-

permission.

70

IN

READINGS

MATHEMATICAL

empiricalgeneralityof

the

for

the

PSYCHOLOGY

is too

the

function.

relations

are

exponents

power
be

can

it not

Were

functions

sumptio

whose

predictedfrom

specialto

be

acceptable
a
a
theory.
power
Elsewhere
fact that some
easy
unare
(Luce, in press),I have
psychophysicists
seem
about these methods, which
iom,
suggested another approach. An axof
wide
or
law,
our
possible
to rest
experience
bility
applicaheavily upon
in the study of choice behavior,
number
the
with
system, the point
In an
be taken in conjunctionwith the
to be established.
would
seem
may
linear generalization
law to
Stevens
of Weber's
effort to bypass these objections,
demonstrate
the
of
scale
existence
had
has
a
recently
subjects
(1959)
that is a power
function of the physical
between
values
match
pairs of concontinuum.
finds
the
that
that
he
and
resulting
Although
theory
tinua,
tress

the

ables.
magnitude
separate varimuch
remains
Thus, although
the "direct" methods
to be learned about
scales of the

as

basic axiom

leads to what

in

appears

deductive

to be

the correct

criticisms.
form, it is open to two
criminati
First,the exponent predictedfrom disdata

is at

least

an

order

magnitudelargerthan that obtained


Second, the
by direct scalingmethods.
teresting
inis
based
summarize
functions appear
to
an
assumptions
theory
upon
and
these are
about discriminability,
body of data.
is
mined
not obviouslyrelevant to a scale deterGiven these empiricalresults,
one
method.
of
Scales
another
suitable
formal
to
by
challenged develop a
be related to
apparent magnitude may
theory from which they can be shown
There
be littledoubt
jnd.scales,but it would be unwise to
to follow.
can
take it for granted that they are.
monly
that, as a startingpoint,certain comThe purpose
of this paper is to outline
made
priate:
assumptions are inapprostillanother approach to the problem,
equalityof jnd's,equallyoften
that is not subjectto the last
one
Thurstone's
noticed
differences,and
The
results have applicabilcriticism.
ity
assumption. Since,
equal variance
of psychofar beyond the bounds
however, differences stand in the same
the general
physics,for they concern
logarithmic relation to ratios as
of
relation
between
the
urement
measquestion
tion,
funcFechner's law does to the power
and substantive theories.
a reasonable
startingpointmight
the
the
that
be
to
seem
assumption
Types
Scales
of
ratio of stimuli one jnd apart
subjective
lus
is a constant
independentof the stimuby now
Although familiarity
may
the
dulled
of
Obvious
have
its
as
our
sense
intensity.
importance,
dure
procein my
opinion it will Stevens' (1946, 1951) stress upon the
seem,
may
that leave certransformation
not do.
tain
Although generationsof psychologists
groups
convince
invariant
have
scale
to
managed
properties
specified
of the
themselves that the equaljnd assumpmust, I think,be considered one
tion
is plausible,
if not
cussion
obvious, it is more
strikingcontributions to the disbeen
in the past few
and
has
of measurement
not
never
particularly
ers
Prior to his work, most writdecades.
compelling; and in this respect, an
ferent.
difhad put extreme
emphasis upon the
equal-ratio
assumption is not much
is a
jective property of "additivity,"
which
This is not to deny that subhave
the equal- characteristic of much
continua may
urement
measphysical
if the power
ratio property
they must
(Cohen " Nagel, 1934). It
of

scaling,the

of

resultingpower

"

"

"

law

is correct

but

rather

and
to

Weber's

argue

that

law

such

holds
an

"

as-

wks

held
to

mental
that this property is fundascientific measurement
and,

72

more

theories are usuallystated in terms


the scales
of fimctional relations among
several
from
the
ment
measurethat result
theories for the variables involved.
For

number

is much

type

of purposes, the scale


the
crucial than
more

theory from

details of the measurement

For

the scale is derived.

which

that

limitations

the

scale

the

sibly
placesupon the statisticsone may senof
the
If
interpretation
employ.

statisticaltest

statistic or
particular

In

finds

physics one

at

least

two

of basic

assumptions: specific
empiricallaws, such as the universal
Ohm's
law of gravitation
or
law, and
o
f
tion,
construca prioriprinciples
theory
such as the requirement that the

classes

should

of mechanics

laws

of the coordinate

laws, such

to

have

the

to
empirical

system.

changed

tions
rota-

Other
of

the conservation

as

seem

be invariant

translations and

uniform

under

type

Theory

of

Construction

paid to

been

attention has

much

ple,
exam-

Principle

relations among
or
two
tive
substanIn practice,
variables.

involves

ter

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

ergy,
en-

from

the

ing
prioricategory dur-

development of physics. In
admissible scale transformations
is altered when
has been put
stress
psychology more
stantive
subthen our
are
applied,
of
the discovery empiricallaws than
on
conclusions will depend upon
the formulation
of guiding princion
ples,
of the
which
arbitraryrepresentation
relations
and the search for empirical
culations.
calhave used in making our
scale we
tends to be pursued without the
when
Most
scientists,
they benefit of
about
explicitstatements
the problem, feel that they
understand
ory.^
what is and is not an acceptabletheshun
such statistics and rely
should
have been
Since such principles
only upon those that exhibit the appropriate
in physicsto limit the
used effectively
invariances for the scale type
metic
the geometric
and arith-

Both

at hand.

in this sense
legitimate
for ratio scales (unit arbitrary),
only
the latter is legitimatefor interval
and
scales (unit and zero
arbitrary),
means

are

Stevens:

discussions,see
1955 ; for

somewhat

For

scales.

for ordinal

neither

fuller

1946, 1951,

second
group

mation
place where the transforlimitations
is in
imposes

the construction of substantive theories.


far

limitations

less

seem

attention

questions,even
more

of the
the

formulate

to have

received

the

statistical

than

though they
fundamental.

are

doubtedly
un-

The

mainder
re-

will attempt

paper
relation between

to

scale

are

possible.As alreadypointed

out, these

issues have

scientific relevance

beyond psychophysics.

of

errors

be

the behavioral

hand, if
what
limits

The

observation.

acute
problem is particularly

sciences.

to

us

On

an

some

laws,
possible

then

in

the other

prioriconsideration

constitutes

about

theory
acceptable
rather

small

fairlycrude

set

of

obser-

attempts to introduce and use such


the
in behavioral problems are
statements
combining of classes condition in stochastic
2

Two

Thomplearning theory (Bush, Hosteller," son,


work
the form of
on
1954) and some
which is based
function for money
the utility
that certain game
theory
the demand
upon
a
solutions should remain unchanged when

swer
types and functional laws, and to anthe questionwhat psychophysical

laws

not

in psychology.
possible
Without
such principles,
practically
and
relation
is
a
prioripossible,
any
is difficultto pin down
the correct one
because of the ever
means
by empirical

error

Mos-

wonders

one

something similar may

whether

present

teller,1958.

These

possiblephysical laws,

pretation
less strict inter-

of the conclusions,see
A

the

constant

sum

is added
of money
" Thompson,

to

all the

1957). In
larly
particudo the conditions seem
neither case
compelling.
payoffs (Kemeny

vations may
which law

suffice to decide

sometimes

substantive

or

able,
dependent variis simplysayingthat the
then one
strictures imposed by the measurement
theories are
incompatiblewith those
imposed by the substantive theory.
Such
a
logicalinconsistencymust, I
be
think,
interpretedas meaning that
retical
something is amiss in the total theoof

transformation

ble

actuallyobtains.
to be suggestedappears
principle
be
of one
used in
to
a
generalization
physics. It may be stated as follows.
The

73

LUCE

DUNCAN

R.

theory relatingtwo

variables and

more

theories

for

the

urement
meas-

bles
these varia-

structure.

should

be

that:

such

invariance

The

1.

substantive

(Consistencyof
transformations

of

one

controversial.

should

or

laws

independentvariables

of the

more

and

Admissible
theories)

and measurement

values of parameters that


reflect the effect on the dependent
of

admissible

the mathematical

of

the

independent

variables.
this

In

the

to

are

For

example,

one

of units is used.

physicsthat

sort of violation
particular

lows, of these until


fol-

principle
pendent
deas

ent
differ-

be viewed

can

of

Although this
are
examples

plausible,there

from

not

set

one

when

units is used and another

do

we

way,

law when

however, let us postpone

and
in what
principle,
terms
independentand

used

that

another

have

to

theory shall be
formations
trans-

to

be able to say that Ohm's

Put

current.

structure

admissible

reference

states

seems

independentof

we

the

set

of the substantive

field without

to

mations
transfor-

ables,
independent vari-

that

substantive

that voltageis proportional


product of resistance and current
without specifyingthe units that are
used to measure
voltage,resistance,or

law

to

want

of the

the

the variables.

want

we

dependentvariables.
tive
2. (Invariance
of the substanmerical
theory) Except for the nu-

subtle

more

It asserts
state

to

particularscales

measure

of the

variables

be able

of the

the

ory,
shall lead,via the substantive themations
only to admissible transfor-

part is

as

2:

of Part

the discussion

of the
consequences
have
been derived.

some

stated

meaning of the principlemay


the variables to which arbitrary, be clarified by examples that violate it.
distinguish
Suppose it is claimed that two ratio
admissible transformations
are
law.
related by a logarithmic
scales
are
the
imposed from those for which
variables

transformations
substantive
in

some

cases

in the

theorycan

are

used

are

only

determined

by the

theory. As will be seen,


the labeling
is trulyarbitrary
sense

that the substantive

be written

so

that any

either in the

appears

there is
that

true

asymmetry

variables must

some

in the
be

able
vari-

admissible

An

positiveconstant

by

of

unit.

kx

However,

log

cases

of

transformation

independentvariable

k -\-log

the

is multiplication

a
k, i.e.,
change
the fact that log

means

that

an

admissibl
in-

transformation, namely,

dependent

independentrole,but in other

or

The

to

change

of zero,

is effected

variable.

sense

fails to

dependent

Hence,
the

meet

on

the

pendent
de-

the

logarithm
consistency

tive
nential
requirement. Next, consider an expoindependentif any substanthe
transformation
then
all.
relates
them
law,
at
theory
This
be
e*=* (^)^
leads
can
to
One
sistency
can
hardly question the conviolation of coneither as
sistency
viewed
a
part of the principle.If an

and

others

admissible

of

transformation

variable leads

to

an

an

pendent
inde-

inadmissi-

or

is

of invariance.

then
exponential,

the

If the

law

dependent vari-

74

READINGS

is raised

able

to

which

power,

is

inconsistent ^^^th its being a ratio scale.


the
Alternatively,

taken

be

may
then

is

law

the

it is

dependent

exponentialraised

an

to

power
indethe unit of the pendent

variable.
APPLICA.TION

An

in

PRINCIPLE

THE

OF

of the

tering
enphysicalmeasures
into psychophysicsare idealized
physicaltheories in such a way that

Most

time

interval

or

durations

measured

are

on

i.e.,
kx", where
name
applied

logarithmicinterval scales.
Because
this topic is more
general
than psychophysics,I shall refer to
the variables as independent and dependent
rather than
assumed

Of course, differences and


of interval scale values

time
and

having
"

Although

on

attempted

scales that

either ratio

are

the

variable,where

often

assumption

and

the

has

estimation

Examples:

"law
lead

of

to

argued
methods

conditions
from

to

I have

should

us

the

data.

about

should

unit

of

the

sult
re-

law

is to say,

That

be unaffected.

dependent

depend upon k, but it


depend upon x, so we denote
matical
it by K{k). Casting this into mathemay

terms,

ratio

we

the

obtain

tional
func-

equation

Our

the relations

pendent
de-

of the functional

the form

changed

the

tion
multiplica-

shall not

ficient
given sufa

of

positiveconstant,

and

interval

that nitude
magresult in

derive

discrimination

tell

by

parative
com-

question here, however, is not how


in
have succeeded
well psychologists
other,
perfectingscales of one type or anbut what a knowledge of scale
among
In

unit of the

variable,namely

variable

belief); and
plausible

can

variables

both

transformation

difference

ory
theratio scales (but no measurement
has been offered in support of this

types

known
un-

have

closely related

of Thurstone's

scales; Stevens

or

noticed

judgment"

scale

is the

ratio scales. If the

form

pendent
de-

the

relatingthem.

law

Suppose, first,that

u{x)"
of

independent variable is changed by


multiplying all values by a positive
arrive at
constant
k, then according to the
interval,
missible
above
stated
only an adprinciple

to

former.

preferably
the equally
V

sidered
con-

of the

measurement
psychological

theories have

Case

be

ordinal,those who

to be

worked

best

at

and

corresponding value

stitute
con-

Let

point.

one

variable

functional

can

be

continua

typical value

independent

rivatives
dethe

psychologicalscales

most
use

chological.
psy-

will

numerical

than

more

denote

ratio scales.

in current

physical and
variables

form

to

interval

on

0. The

"

scale goes
scale,since the transformed
will consider
into c log X + log k. We
all combinations
of ratio,interval,and

measured

0 and

this scale type reflects


the fact that log x is an interval

ratio

physical time (not


ordinarytemperature,
durations),
are

^ "
to

Both

scales,and
entropy
scales.

by positive constants
multiplications
and raisingto positivepowers,

are

and

length, pressure,

Mass,

scales.

ratio

either

form

they

interval scales (Stevens,1957). In this


the admissible
transformations
case

not

depends upon

that

variable

ratio scale,but
invariant because

be

to

PSYCHOLOGY

MATHEMATICAL

IN

u{kx)
where

Q and

k"

cases

are

K{k)

arrived at in

They
The

K{k)u{x)
"

0.

equations for

Functional

scales.

are

summarized

questionis : What

the other

similar

ner.
man-

in Table

to these two
common
functional equations,each of which
is
interest
there
of
some
scales,
types
the principle,
in what have been called logarithmic embodies
imply about

addition

do these nine

TABLE
Functional

The

Equations

Principle

of m?

form

the

consideration

to

continuous,

of

X.

Theorem

shall limit

We

theories

for

our

where

u{k)u(x)/u{l).Lety

log["/"(!)],

then

function

nonconstant

v{kx)

log[_u{kx)/u{l)^

u(k)u{x)

If the independentand

1.

the

Construction

is

Satisfying

Laws

the

Theory

of

75

LUCE

DUNCAN

R.

"

"^m(1)"(1)

dependentcontinua are both ratio scales,


then u (x)
ax^,where /3is independent
of the units of both variables.^
=

Set:x;

Proof.

u{k)

1 in

constant

we

Equation

Because

K{k)u{l).

choose

may

non-

that

so

0, and because K(k)


follows that m(1) " 0,soK{k)
M (1) Thus, Equation 1 becomes

u{k)

"

0, it

"

u(k)/
u

In

the

in the

this and

statement

can

continuous, so is v, and it is
that the only continuous
well known
Since

is

solutions to the last functional

(kx)

tion
equa-

of the form

are

following theorems,
general*if
more

v{x)

/3log jc

log x^

be made

replacedby x + 7, where 7 is a constant


unit as
independent of x but having the same
of
The effect of this is to place the zero
X.
at some
u
point different from the zero of x.
be reIn psychophysics the constant
garded
7 may

log [m(x)/m(1)]

v(k) -{-v(x)

is

1, then

is

as

such

the

constant

The

threshold.

of

u(x)

where

of course, that a plot


will not in general be a

ae'^'^
ax^

u{l).

means,

of

log M vs. log X


straight line. If, however,

variable

presence

Thus,

is measured

in

terms

the

We

observe

that

since

independent

u{kx)

of deviations

become
the threshold, the plot may
straight. Such nonlinear plots have been
instances the
observed, and in at least some
to be correlated
degree of curvature
seems
ther
Furwith the magnitude of the threshold.
empiricalwork is needed to see whether
this is a correct explanationof the curvature.

ak^x^

a'x^

from

/3is independent of the unit of


it is
of

clearly independent

of

x,

the

and
unit

u.

Theorem
is

tinuum
If the independent conratio scale and the depend-

2.
a

76

READINGS

continuum

ent

either

(x)

IN

MATHEMATICAL

interval scale, then


log x -\-^, where a is

an

PSYCHOLOGY

It is easy

to

that 5 is

see

of the unit of

and

jS is

independent
independent

pendent of both units.


independent of the unit of the indeax^ + 8,
A much
variable,or u(x)
rem
simpler proof of this theowhere /8 is independent of the units of
be
if
can
that u
given we assume
is dififerentiable in addition to being
both variables and 5 is independent of
the unit of the independentvariable.
continuous. Since the derivative of an
interval scale is a ratio scale,it follows
Proof.
In solvingEquation 2, there
=

immediately that
du/dx satisfies
to consider.
possibilities
1. li K{k)
e".
1, then define v
Equation 1 and so, by Theorem
1,
dx
2
becomes
v (^x)
D(k)v(x),
Equation
c^.
we
e^^''^" 0 and v is conIntegrating,
where D(k)
get
tinuous,
and nonconstant
cause
bepositive,
u is.
By Theorem
l,v(x) Sx",
x^+i + 5 if /8 5^
1
u{x)
where a is independent of the unit of
j |8+ 1
where 5 " 0 because, by definiI a logx+S
tion,
X and
if fl
1
0.
V "
Taking logarithms,u(x)
3. If the independentconTheorem
tinuum
a log X+/8, where
0= log 5.
two

are

K{k) ^ I, then

2. U

be two

let

different solutions to the

and

define

u*

lem,
probIt follows

u.

"

is

u*

and

immediately from Equation 2


must
satisfythe functional
equation w{kx)
K{k)w(x). Since
that

both

and

u*

K(k) ^ I,

be

it is clear

impossiblesince
to be different.

and

that

the

0, and

u*

Since

constant.

solution isw

constant

w(x)

continuous, so is w;

are

however, it may

only

the

ent
depend-

"

units of both variables and 8 is independent


of the unit of the independent

dependent
variable,or u(x)
ax^, where /8is inthe
both
variables.
units of
of
=

this is
Proof.

Take

3 and

Thus, by Theorem

Substitutingthis into the


functional equation for w, it follows
that K{k)
".
Then
0
settingx
in Equation 2, we obtain C{k)
w(0)
observe
that
k^). We
now
X(l
ax^ -\-8, where 8
u(x)
u(0), is a
solution to Equation 2 :
=

the

let v

1,

ax^.

is

interval
logarithmic
where a
Se'"^,
scale,then either u{x)
is independent of the unit of the dependent
is
variable,
^
independentofthe

chosen

were

ratio scale and

continuum

vikx)

logarithm of Equation
log u :

K*ik)

K*ik)

where

Cik)v(x)

logK(k).

rem
By Theo-

2, either

"

u(kx)

v(x)

ax^ -\-8*

v(x)

13log x -{-a*

either
Taking exponentials,

ak^x^-\-8

u(x)

a"x^+u(0)k^-\-uiO)-u(P)k^

k^u(x)+u{0)il-k^)

where

8
a

8e'"^

or

u{x)

ax^

"** and, in the second

tion,
equa-

e"*.

K(k)uix)-\-C(k)
Theorem

Any

or

other solution is of the

same

because

u*{x)

u(x) + w(x)

ax^ -\-8 -{-ax^

(a-\-a)x^ +

form

interval scale,then it is
an
impossible
for the dependent continuum
to be

tinuum
If the independentcon-

4.

is

Proof.

by

ratio scale.

Let

Theorem

Equation 4, then
ax^.
know
u{x)

0 in
we

DUNCAN

R.

Now

set

1 and

in

5^ 0

tion
Equa-

77

LUCE

Proof.

Take

6 and

a(x + cy

K{l,c)axP

the

let v

v{kx + c)

tion
logarithm of Equalog u:

K*{k,c) +

C(k,c)v{x)

SO

-^

where

K(l,cy/^x

K*{k,c)

Theorem
which

impHes
to

have

Theorem

than

more

trary
con-

that

v(x)

both

point.

one

continua

scales,then u{x)

-\-^, where

ax

independent of the

interval

both

are

unit

of

the

u(x)
where

pendent
inde-

Theorem

5 reduces

7.

is

let

we

e"*.

0, then

tion
Equa-

Equation 2 and so
Theorem
2 applies. If "(x)
a log x
\ and c 7^ 0
-\-jS,then choosing k
in Equation 5 yields

it is

tinuum
If the independent coninterval
a
scale,
logarithmic
impossiblefor the dependent

continuum

to be

Proof.

Let

c) + (S

By taking the derivative with respect

log x)

X(l,c)/3+ C(l,c) Thus, log X


ratio

u{x), i.e.,
v{y)

7 becomes

K{k,c)u{\ogx)

K{\,c)a logx

to

u(logx)

w(e^),then Equation

Z'(log^ +
log (x +

ratio scale.

to

ae^^

/3is

then

If

/3x-fa*

so

variable.

Proof.

By

5,

If the independentand

5.

dependent

constant,

assumption

our

continua

is

log K{k,c).

it is easy to see that x must


be
which
is
impossible.
constant,
conclude
that u{x)
Thus, we must
ax^ -f (S. Again, set k
1 and

is an

interval scale and

scale,which

by Theorem

is

4 is

impossible.

X,

5^

0,

Theorem

tinuum
If the independent coninterval
scale
is a logarithmic
and the dependent continuum
is an
terval
in-f
scale, then u(x)
x
log
a
/3,
where a is independent of the unit of the
independent variable.

8.

a(x -f cY

i^(l,c)ax"

X(l,c)|3+

-f

C(1,C)
Proof.

If 5 ?^ 1, then
to

differentiate with

spect
re-

v(\og x)

Let

Equation
V

8 becomes

(logk +

log x)

K{l,c)a5x^-'

K(k,c)v {logx) +

must

that

is

impliesx
conclude

m(x)

ox

1.

constant,
It is easy

so

we

to

see

so

j8satisfiesEquation

5.

log X

and

By Theorem

is

6.

If the

an

interval

independent
scale

and

tinuum
con-

both

are

interval

C{k,c)
scales.

5,

u(x)
Theorem

then

aS(x + cy-'
which

u{x),

i'(logx)

log X -f /3

the

Theorem
9.
If the independent and
logarithmic
both logarithmic
continua
where
are
dependent
ae^",
u(x)
is
pendent interval scales,then u{x)
ax^, where
a
independent of the unit of the indevariable and
/S is independent /3is independent of the units of both the
of the unit of the dependent variable.
independent and dependent variables.

dependent

continuum

interval scale, then

is

78

READINGS

Proof.

Take

9 and

PSYCHOLOGY

MATHEMATICAL

logarithm of Equation
:
log u

the

let

IN

form
course,

As

v(kx')

K*{k,c)

where
Theorem

C{k,c)vix)

K*ik,c) +

log K(k,c).

By

v(x)

+
iSlog:*;

a"

example

an

u(x)

""(*'

ax^

used to

is

interval

an

added

constraint

with
of

source

No

of them.

some

examples

idealized

are

or

as

attempt

have
The

variables

law
gravitation
in each

are

no

lomb's
Cou-

law, and Newton's


all ratio scales,and

Additional

1.
1

of the law

can

is

of

of

on

square

scales,and
as

we

can

illustrate Theorem
constant

then

relation

mass

its energy

variables
If

is moving at

If-the temperature
constant, then as

of

as

they will
body of

velocityv,

is of the form

av^ +

discussed

from

matical
mathe-

not

are

8.

of sure
a
presp the entropy of the gas is of the

between

that it is

new

"

the

may

cases.
specific
to challenge

and

physics,such
of radioactive

law

ables
vari-

two

empirical,not

theorems, they have


from

seem

be in

this view

support

an

to ascertain what

matter
theoretical,

cited
as

amples
ex-

the

ponential
ex-

decay

or

sinusoidal function of time, which


stated
violate the theorems
to
We

above.

therefore,examine
these examples bypass

must,
in which

the ways
the rather

strong conclusions of
the present theory.
have
All physicalexamples which
been

suggested

perfect gas is

function

hold

can

and

are

anticipate

therefore
2.

such

I have

which

"

had
strong misgivings about
is that
interpretation
; the feeling
substantive
of
nature
a
something
have
been
must
smuggled into the
formulation
of the problem. They
functional
that practically
argue
any

To

interval

form

whom

have

ratio

its side

variables

entropy

dependent

with

point of view

some

important
and

energy

that

Some

the function

illustrations.
Other

one.

these theorems

since

the

area

rem
Theo-

dependency of the
sphere upon its radius or

scales; thus
of the

in geometry
volume
are

and

length, area,
volume

examples of

be found

is

Discussion

rem
function,as called for by Theo-

power

scale, as

their

seems

entering into

the form

case

to

scales, because

law. Ohm's

is the linear

concerninglogarithmic

of scales of this type


been made.

temperature

interval

an

tion
usuallyassumed, then the only rela5
possibleaccording to Theorem

interval

use

If the

any

given
freezing
a

con-

will be made

illustrate the results


interval

is classical

of water.

scale is also

of the fundamental

most

either ratio

tinua that form

to

choose

may

point

variables

actual

we

Illustrations

physics,where

scales.

(subjectto the
that the length is

initial length to correspond to


such
the
as
temperature,

that accord
best

The

we

temperature

measure

C*.

be useful,priorto discussing
It may
these results,to cite a few familiar
laws

scale

since
positive),
where

Theorem

ordinary temperature,
which is frequentlymeasured
in terms
of the length of a column
of mercury.
forms a
Although lengthas a measure
ratio scale,the length of a column
of
mercury

so

of

of

consider

may

8,

logp -\-/8. No examples,


4.
are
possiblefor Theorem

to

form

common

is

to

the
:

the

me

as

theorems

examples
counter-

have

able
independent vari-

ratio scale, but

it enters

into

80

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

tion may
reflect the actual state
not
of affairs in the empiricalworld.
It

and

that

they

related

tion
by a functhat is nonnegative, nonconu
is certainly
true that, in detail,cal
but
physi- stant, and monotonic
increasing,
mathematical
continua
c
ontinuous.
not
We
not
are
necessarily
now
need
that u cannot
be
continua, and there is ample reason
only show
to

that

suspect

the

holds

same

psychologicalvariables.

But

that stimuli and


both

continua

form

are

the

for

bounded

idealizations

difficult to

are

but

it is doubtful

be of much

would

are

tinuity
disconvariable.

if this

helpby itself. The

solutions to, say, Equation


manifold and extremelywild

M
.

this

"(^) ^

They are so wild


that it is difficultto say anything precise
about
them
at all (see Hamel,
1905; Jones: 1942a, 1942b),and it is
in their behavior.

T^/Lx,

,.

implies w

0, con-

assumption. Thus, for all


1, which by Equation
u(kx)
1, means
u(x), for all x and
1.
This in turn
^"
implies " is a
which
again is contrary to
constant,
have
lished
estabassumption. Thus, we
to

trary

k"

1, K(k)

that such solutions represent


empiricallaws.
doubtful

Second, casual observation

the
the

discontinuous
1

that

exist in

Suppose, therefore,that it is bounded


and that the bound
is M.
By Equation
1, u{kx)
K{k)u(x)" M, so
For k"\, the monu (x)"
M/Kik).
otonicity of u implies that u{x)
" u {kx)
K(Ji)u{x) so choosingu {x)
"0
1.
that K{k)"
If for
we
see
"
^
some
1, K{k) " 1, then K can be
made
arbitrarily
large since, for any
integern, K^k")
Kik)"",but since

responses

give up; to do
would
mean
so
casting out much
tively,
of psychophysical theory. Alternacould drop the demand
that
we
them
be
the function
tinuous,
conrelating
that

show

to
must

sumptions
as-

are

suggests

that it might be

appropriateto assume
claim that some
our
that at least the dependent variable is
reside in the
must
bounded, e.g., that there is a psychologicallythe variables.
maximum

be

though
Al-

loudness.

boundedness
plausible,

two

absolute

scale

with

or

in the formulation

tinuity
discon-

some

of the

lem,
prob-

of the
possibly in the nature
function
variables or possiblyin the
tablish
Actually, one can esrelatingthem.
that it must
of

the

variables.
form

we

hold all but

in

our

quency
intensityand freUsually

loudness.

variable constant

one

but
empirical investigations,

the fact remains


that

their presence
difference in the

and

there
make

some

of

range

variables, x

and

y,

ratio scales and

form

total

For

two

are

are

may

example,
independent

possiblelaws.
there

suppose

the others

that

both

of which

that the

ent
depend-

ratio scale,
of
then the analogue
Equation 1 is
variable

is also

u(kx,hy)

K{k,h)u(x,y)

be in the nature

Suppose,

contrary, that the variables


scales that

determine

appropriatefunctional

if the functions
equations are unbounded
be
must
a
s
they
are
increasing,
It seems
clear
for empiricalreasons.
of the dependent
that boundedness
variable is intimatelytied up either
with introducinga reference level so
that the independent variable is an

more

or

theorems, all the continuous

solutions to the

situations,there are
independent variables;

example, both

for

of

nature

many

cannot

imposed by itselfsince,as is shown

in the

Third, in

tinuity
discon-

numerical

on

are

the
ratio

continua

where
We
hold

k "

know
one

0, h

"

0, and

by Theorem
variable,say

K(k,h)
1 that

y, fixed at

0.

"

if

we

some

R.

value

and

must

be of the form

leth=

DUNCAN

1, then the solution

81

LUCE

the

dimensions

variables.

of

This

the

independent

be parto
appears
ticularly
if the dependent

appropriate
u{x,y)

a(j)x^^"^

variable has
The

But

holding

1,

also

we

and

constant

know

that

letting

it must

be

tinua

they

u's

to

partialderivatives of both

having

the

solutions only of the form

{x,y)

Thus, the
admit

ax^y'-^"^

than

when

we

be

must

that

emphasized

in Footnote
If a function

here.

3 does

not

u{x,y)

that

+ tCv)]^^"^
a(y)\ix

theorems.

The

first

can

be

rejectionof Part

ratio scale
dimensional

numerical

this

In

constants.

of the form
functions
which

provides

use

it involves
constants

is

the

to

arrive

is either

for

theories exist

being obtained.

ratio

scale,
the

not

of
For

tend

interval

there

is

these

methods

Stevens

terval
logarithmicin-

or

interval

an

results

matchings

from
to

scale.

cross-modality
the logarithmic
as
a possibility,

eliminate

scale

presumptive evidence

has

that

yield ratio scales,

as

claimed.

Summary
The

at

following problem
What

of

result of the present work.

chophysical
psy-

methods

least indirect evidence

at

of
the presence
that cancel out

of this argument

of the

by

measurement

no

2 of the

factory
satismuch
more
u{x,y) seems
and
the heuristic
convincing than
development given in Section 2.C of Luce (in
press), and the empirical suggestions given
there should gain correspondingly in interest
a

may

of the psychophysical
is determined
e.xcept for

viewed

or as the creation of a dimenprinciple


sionless independent variable from a

form

we

of scales

the form

functions
some

Since

In sum,
there appear
to be two ways
around
the restrictions set forth in the

as

known,

physical
psycho-

that

types
Once

being obtained.

solution possibilities
wholly new
(see Section 2.C.3 of Luce pn

The

ment
measure-

example, the magnitude methods


seem
result in power
to
which
functions,
ure
suggests that the psychological meas-

rather

what

the type of scale

press])

the

laws, but

in order

for certain

apply

other, e.g.,

as

mine
deter-

empirically testable

the

depends upon
independent variable is added to

either

of the

the

to

not

meantime, however, experimental determinati

variable.*
remark

that

argue

independent

one

can

theories for the several


know

stricts
principleagain severely re-

more

one

methods

the possiblelaws, even

then

hand, if the theorems

the forms

'"8

that

to create

are

exist

other

important question is

variables,

this equation can


be shown
tion
(seeSec2.C.2 of Luce [inpress])to have

the

that

bounded.

are

limited
restrict ourselves

we

one

con-

assume

applicable,then the possible psychophysical


(and other) laws become
severely limited. Indeed, they are so

a(y)x^^''^ 8{x)y'^'^

It

numerical

as

are

Thus,

If

tion
reject the idealiza-

and, possibly, to

On

8(x)y''-^

is to

of the variables

of the form

u{x,y)

true, well-defined bound.

second

of

substantive

dependent
manner

Each

to

are

theory

variable
an

variable

the

in

was

sidered.
con-

possibleforms
that
a

relates

continuous

independent variable?
is idealized

as

nu-

82

IN

READINGS

TABLE

Possible Laws

The

""

The

a/x

notation

Satisfying

the

and is restricted

merical continuum

Principle

variable

that

are

to

admissible under

theory shall

measurement

Theory

Construction

REFERENCES

by

beingeither
logarithmic
a
ratio,an
a
of theory
interval scale. As a principle
construction,it is suggested that
of the independent
transformations
theory
interval,or

its measurement

of

of the unit of x.'

"a is independent

means

PSYCHOLOGY

MATHEMATICAL

its

F., " Thompson,


for multiplechoice situations. In R. M. Thrall, C. H.
Coombs, " R. L. Davis (Eds.), Decision
York:
Wiley, 1954. Pp.
processes. New

Bush,

formal

structure

99-126.
tion
R., " Nagel, E. An introducNew
method.
logic and scientific
Harcourt, Brace, 1934.

M.

Cohen,

to

result

not

in inadmissible transformations of the


and
dependent variable (consistency)

R., Mosteller,

R.

G. L.

York:

Hamel, G. Eine Basis aller Zahlen und die


unstetigen Losungen der Funktionalgleition
Math.
that the form of the functional rela/(x + y)=/(jr) +/(3;).
chung:
variables shall
Annalen. 1905, 60, 459-462.
the two
between
and disconnected
Connected
formation Jones, F. B.
be altered by admissible transnot
functional
the
equation f{x)
and
sets
plane
of the independent variable
Math.
Bull. Amer.
+ fiy) =fix + y).
limits sigThis principle
nificantly
(invariance).
Soc, 1942, 48, 115-120. (a)
and other properties
the possiblelaws relating Jones, F. B. Measure
Math. Soc,
Bttll.Amer.
basis.
of
Hamel
2.
Table
a
in
shown
the two continua, as
472-481.
(b)
48,
1942,
portant
These results do not hold in two imThe
G. L.
J. G., 8l Thompson,
Kemeny,
First, if the
circumstances.
comes
attitudes on the outpsychological
In M. Dresher, A. W.
of games.
Tucker, " P. Wolfe (Eds.), Contributions
Princeton:
III.
to the theory of games.

effect of

is

independent variable
that

is

at

variables

upon

the

by
having

pendent
those of the inde-

Princeton

Univer.

Press, 1957.

Pp. 273-

298.

how

one

wishes

to

Second, if the

matter.

tinuous,
discrete rather than conif the functional relation is

are
or

discontinuous, then
those

constant

prin- Luce, R. D. Individual choice behavior: A


variable,then either the ciple
it is violated,
York:
or
Wiley,
content
no
theoretical analysis. New

depending
look

ratio scale

dimensionless

rendered

multiplying it by
to
units reciprocal
has

given in Table

than

laws

other

possible.

are

in press.

Luce, R. D.,
of

"

Edwards,

differences.

The

W.

tion
deriva-

able
just noticePsychol. Rev., 1958, 65,

scales
subjective

from

222-237.

M0STEI.LER, F.
corpus.

The

mystery

of the missing

Psychometrika,1958, 23, 279-289.

R.

S.

S.

Stevens,

On

S.

S.

Stevens,
and

the

1946,

New

.sensory

S.

Stevens,

1956,
S.

Stevens,
subjective

S.

S.

S.

estimation

Thurstone,

loudness.

of

Anter.

J.

the

1957,

psychophysical

Galanter,
scales

loudness,

L.

vibration.

E.
for

exp.

L.

law

Psychol.

Rev.,

J.,

theory

of

(2nd
Univer.

validation

/.

Neumann,

VON

law.

153-181.

Cross-modality
for

"

H.

Ratio

dozen

Psychol.,

ceptual
per-

1957,

of

comparative

1927.

34.

273-

286.

The

64,

1959,

377-411.

54,

data.

1-25.

On

scales

1-49.

Psychol.,

exp.

category
continua.

"

69,

Rev.,

Psychol.

of

S.,

and

judgment.

direct

The

magnitudes

Psychol.,

scales

113-116.

121,
S.

Pp.

/.

shock.

S.

Stevens,
Stevens

averaging

the

electric
201-209.

57,

ogy.
psychol-

1951.

Wiley,
On

1955,

Science,
Stevens,

S.

and

677-680.

S.

experimental

of
York:

S.

S.

83

LUCE

of

scales

measurement

In

Handbook

Stevens,

103,

Mathematics,

psychophysics.

(Ed.),

of

theory

Science,

measurement.

DUNCAN

Press,

"

Morgenster.n-,
and

games

Princeton:

ed.)
1947.

of

(Received

December

2,

1958;

economic
Princeton

O.
havior.
be-

MULTIVARIATE

INFORMATION
William

MASSACHUSETTS

J. McGill

INSTITUTE

A multivariate
analysis based
It is shown
that
sample transmitted
for

measuring

and

tables. Relations

Several

from

communication
the

This

theory
human

It

consider

us

will

statistical tests

Basic

(10)

transmitted

system.

leads

information
and

we

of transmitted

measure

appropriate

to

of information

some

uncertainty

transmitted

and

are

the

at
some

input variable,
values

or

receiver
not

not

in

of association

event

the

if

hand,

Naturally

these

most

There

extremes.

information

Some

sent.

input and
is

is

get through.
what

1, 2, 3,
and

is sent

(k,m).

This

"

"

Suppose

that

variable,

y.

"

with

1, 2, 3,
m

is

joint

"

"

"

is

in

discrete

discrete, it

probabilitiesindicated
F with
probabilities
,

received, we
event

have

we

Since

,
=

is, but

information

transmitted

the

output

values
k

other

transmitted.

was

output.

between

perfectlycorrelated,

are

between

what

about

discrete
=

the

is

found

its

output

On

transmitted.

assumes

happens that

joint input-output

amount

input and

are

does

signalsk

by p{k). Similarly,y
it

no

information
a;, and

and

information

transmission

interested
of

amount

If

input and

channel

is transmitted.

independent,

cases

p{m). If

not

or

by

Definitions

the

measures

of the channel.

output

are

on

psychology.
available

whether

contingency data,

communication

input information

We

as

how

shown

in

communications

of Shannon's
that

of

data

analyzing

shown

be

analysis

information

input and

takes

and

tests.

Transmitted

the

pointed out,

described

extension

an

1.

Let

for

is best

organism

simple multivariate

output

multi-dimensional

psychological journals have

useful

are

will present

paper

statistical

all the

in
are

communication

information.
a

theory are being applied


widely understood, however, that the tools made

It is not

to

is presented.
simple method
contingency

information

provides

information

articles in the

recent

derived

believe

transmitted

on

testing association
analysis of variance

with

TECHNOLOGY

OF

described.

are

ideas

TRANSMISSION*!

has

can

speak

of the

probability p{k,m).

work
*This
Human
Factors
Operations
was
supported in part by the Air Force
under
Research
and
Air
Force
Laboratories, and in part jointly by the Army, Navy,
with
the Massachusetts
contract
Institute
of Technology.
dependentl
been
and
fSeveral of the indices
in this paper
have
discussed
developed intests
by J. E. Keith Smith (11) at the University of Michigan, and by W. R. Garner
at Johns
Hopkins University.

This

article

appeared

in

Psychometrika,

1954, 19, 97-116.

Reprinted

with

permission.

J.

WILLIAM

The

rules

governingthe selection of signalsat either end

be constructed

so

J2 P{m)

*-l

successive

independent,the

signals are

signalis defined

in "bits" per

H(x) + H(y)

1.

k.m

transmitted

T{x;y)

T.p{k,m)

conditions,assuming

of information

amount

must

m-Y

JLvik)
these

of the channel

that
k-X

Under

85

MCGILL

as

H{x,y),

(1)

where

Hix)

H(y)

Zp(k) log,p{k),

H(x,y)

2 pM

log2pirn),

-Y^pik,m) \og2pik,m).
k ,m

equal to "logs (^)and represents the information conveyed by


two equallyprobable alternatives. Our development will use
a choice between
in information
the bit as a unit, since this is the convention
theory,but
unit
convenient
be
substituted
of
the
the
base
by changing
logarithm.
any
may
and y, H(x) + H{y) " H{x,y) and
If there is a relation between
x
the size of the inequalityis just T{x;y).On the other hand, if x and y are
H(x,y)
H{x) + H{y) and T(x;y) is zero. It can be shown
independent,
that T{x;y)is never
negative.
of
The presentation to this point has been an outline of the properties
the measure
of transmitted information as set forth by Shannon
(10).These
of
that
the
be
information
amount
summarized
propertiesmay
by stating
the association
that
transmitted is a bivariate,
measures
positivequantity
between input and output of a channel. There are, however, very few restrictions
how
be defined. The input-output relations that
channel may
a
on
sequently
in
occur
psychologicalcontexts are certainlypossiblechannels. Conmany
One

"bit" is

we

can

information

transmitted

measure

in these

contexts

and

anticipatethat the results will be interesting.


2.

Our
on

measures

of information

as

the number

we

make

of

information,i.e.,
constructed from relative frequencies.
of events
observations
{k,m). We identify
n

development will be based

Suppose that
Ukm

Sample Information

of times that k

sample

on

sent

was

that

n*.
k

E
k tTn

and

measures

was

received. This

means

86

where
that

and
received,

was

experiment
and

entries n^m

was

representedby

of times

is the number

sent, n"

of observations.

is the total number

be

then

can

that

of times

is the number

Uk

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

contingencytable

particular
cells

XY

with

"

p(w), and p{k,m) with Uk/n,


p(fc),
probabilities,
is
Sample transmitted information,T'{x',y),
n"/n, and Uk^/n, respectively.
We

may

defined

the

estimate

as

H'{y)

H'{x) +

T'ix;y)

(2)

H'{x,y),

H'{x,y) are constructed from relative frequencies


from probabilities.
[Throughout the paper a prime is used over a

H'(x), H'{y) and

where

instead of

quantity

the

without

prime,

e.g., T'{x;y)is

estimator for

an

As
T{x\y).]

in "bits"

signal.

per

we
manipulate logs of relative frequencies,

it is difficult to

Since

introduce

an

Sm

^A-n

-^n"

10g2nkm

log2n"

logsn.

of information
Expressionsinvolving sample measures
handle in this notation. For example, T'{x;y)becomes

T'{x;y)

write equations like

(3) is

s-notation. Thus

let

two

sources,

v, that

and

in equation (2) with

u,

are

has been
i

transmit

and

riu,v;y)
where

subdivided

1,2, 3,

"

"

s"

Si"

(3)

Information

the definition of transmitted

extend

us

Transmitted

3. Three-Dimensional

Now

Si

easier to

are

(3) are equivalentexpressions for T'{x;y).When


(3),we shall say that these equationsare written in
(2) in s-notation.

Equations (2) and


we

will

easier notation:

Sfc-

of

quantity

same

before,T'(x;y)

(inthe sample) measured

information

of transmitted

is the amount

of the

likelihood estimator

to indicate the maximum

"

we

y. To

to

to include

information

accomplish

this

replacex

we

find that

H'{u,v)+ H'(y)
into two

U, while

classes,u and

assumes

(4)

H'(u,v,y),

values j

v.
=

The

possiblevalues

1,2, 3,

"

"

F. The

"

88

READINGS

mitted

information

IN

MATHEMATICAL

will be called

T[{u\y),where

T.'^m{u;y)l

T'Xu;y)
=

and

is
T'j{u;y)

of V,

namely j. It

information
is

transmitted

readily shown

T'Xu;y)
that

see

is added

subscriptj
There

and

to each

are

three

table.

For

Finallywe

example, the

T'{v;y)

TL{v;y)

s,

these results in mind


and

T[{u;y) ^

y. If

has

T'{u;y).One

singlevalue

(8)

T'{u;y)except that the

as

to

A'{uvy)

"

Si

three-dimensional

tingency
con-

between

of this symmetry,
see

that

between

we

any

v,

i.e.,
(11)

Siy"

between

transmitted
and

y, then

the size of the effect is by


-

T'{u;y),
Sii

"

Si",

"

"

Sy" +

s.y"

(13)

that

T'Xu;y)

T'{u;y),

TL{v;y)

T'{v;y),

Tl{u;v)

r{u;v).

may

(12)

the information

transmission

on

(10)

A'{uvy) is the
two

s,,"

reconsider

Si -\-Sm

A\uvy)

(9)

and

Sir,

T'Xu;y)

substitutions will show

more

measure

s,"

s, -h Sii

Si^

effect

an

way

Si

+
s,"

between

let us

A'{uvy)

We

s^^

way

s"

s,,-

s^

In view

equations for transmission

s;

T'y{u;v)

few

same

two

transmission

r{u;v)

s,",

different pairs of variables in

study

may

y for

of the s-terms.

between

s^

in the

With

and

written

are

T[{u;y)is written

between

(7)

that

sy

We

PSYCHOLOGY

call

A'{uvy)the

u-v-y

(14)

interaction information.

gain (or loss)in sample information

of the variables,due to additional

mitted
trans-

knowledge of

the

third variable.
Now
from

we

u,v to y,

can

the

express

i.e.,T'(u,v;y),
as
r(:u,v;y)
=

three-dimensional
a

T{u;y) +

+ A'iuvy),
T'{v',y)

T'{u,v;y) T'Xu;y)-j- Ti{v;y)


=

information

transmitted

function of its bivariate components,

A'{uvy).

for

(15)
(16)

WILLIAM

J.

Equations (15)and (16) taken togethermean


by a diagram with overlappingcircles as
what

assumes

shall call

we

89

MCGILL

can
T'{u,v\y)
in Figure
shown

that

"positive" interaction

be

represented

1. The

diagram

u,v and

between

y. Inter-

"""v^^iy)^ ^Tu(v;y)

T'(u,v,y)
Figure
Schematic
of

of the components

diagram

three-dimensional
The

three-dimensional

analyzed

into

plus
meanings

The

transmitted

diagram

formation.
in-

shows

transmission

that
be

can

missions
pair of bivariate trans-

interaction

an

the

of

symbols

term.
are

plained
ex-

in the text.

action is positivewhen
is to

constant

This
if

of these

one

increase

that

means

the effect of
the

holding one

of the interactingvariables

of association

amount

T'{u;y)and
inequalitiesholds,both
T'Xu;y)"

is

that interaction may


be negative. When
and
the interactingvariables are reversed,

happens, relations
the diagram in Figure 1
this

correct.
longer strictly

no

Components of Response Information

4.

The
because
those
to

two.

[Because of (14),
TL(v;y)" T'(v;y).
Later
must
hold.]
on, however, we

shall show
between

the other

between

multivariate

model

the situations treated by communication

we

deal with

to

knows

the statistical

means

random

experiments

propertiesof the

The

kind

generally do

transmittinginformation.
statistical noise

with

the

bivariate model

theory tells us

from

transmission

noise. This
we

to

us

the

same
as
theory are
engineer is usually able

not

appUcations.The
psychological

in

restrict himself

is useful

transmission

of information

of
not

We

source,

He
source.
single information
of
he
noise
he speaks
and when

is seldom
precision
in advance

know

must

therefore

be

available
how

many

careful

to

In

us.

sources

not

to

our
are

confuse

experimenter'signorance.
of transmitted

to attribute

information

to random

provided by

noise whatever

cation
communi-

uncertainty there

90

READINGS

is in

the
specifying

if several
will

can

transmission

H\y)
H'(y)

see

and

s"

"

that

HUy)

hand,

sources

from

the multivariate

model

For

transmittingsources.

T'{u;y) +

H'u^(y)

s,,-

H'(y), the

uncontrolled

to

model

bivariate

example,

find that

we

(1).Consequently,
the

responses,

the other

the effects due to the various

in three-dimensional

We

to

is known

effects due

On
variability.

to random

measure

where

the stimulus

discriminate

to

PSYCHOLOGY

information

transmit

certainlyfail

MATHEMATICAL

when

response

sources

those due

IN

r{v;y) +
Si,",

"

A'{uvy),

information,has

response

(17)

been

into

analyzed

plus a set of correlation terms due to the input variables. The


error
term, HiXv), is the residual or unexplainedvariabilityin the output,
due to the inputs,u and v, has been removed.
In
y, after the information
bivariate information transmission,
the response
information is analyzed less
an

term

error

precisely.For

example,

we

have

may

Ii'{y)

H'M

In

this

the

case

Shannon

(10) showed

In other words

if we

the

also control

Hi{y)

because

H'M

term, when

Equation (19)
we

have

error

inputs,namely,
term
seen

well

as

to be

noise.

is

u.

HUy)

information

Multivariate

can
on

kind

be increased

cannot
controlled,

(19)

n{v;y).

transmitted

sides in s-notation.

information

from

extractingthe association

transmitted

information

of
a

only

still smaller

and

y,

of the

one

Controlling v

v.

between

Thus

via responses,

error

is thus

y from

the

essentially information

is

the noise part of bivariate transmission.


5. An

The

is

variables that transmit

equivalent to

analyzed from

only

keep track
term; H^{y), provided we
contains
term
However, this error

the

as

HUy)-

proved by expanding both

stimulus

are

an

is recorded.

input, u,

one

In fact

HL{y)

if u and

only

(18)

that

error

v.

is

term

error

T\u;y).

Example

of analysis that multivariate

be illustrated by

set of data obtained

information

from

one

transmission

subject in

an

yields

experiment

frequencyjudgment.
equally loud tones, 890, 925, 970, and 1005 cycles per second
presented to the subject one at a time in random order. Each tone was

Four
were

I second long

and

separatedby

about

3 seconds

from

the next

tone.

During

preliminarytrainingthe subject learned to identifythe tones by pairing them


with four response
keys. In experimental sessions,a loud masking noise was
turned

on

and

random

sequence

of 250

tones

was

presentedagainstthe

WILLIAM

background.A flashing
lighttold

noise
and

J.

he

the stimulus

subjectwhen

the

if in doubt

instructed to guess

was

91

MCGILL

which

about

occurred,

of the four tones

one

it was.

object of the experimentwas to find weightsfor both the frequency


in determining which
and the immediately precedingresponse
key

One
stimulus

subject would

the
The

data

press.

presentedhere

close to the masked


order

In

Tests
were

weights,we

of three-dimensional

example

considered
stimuli. The

consider

can

in which

odd-numbered

stimuli

the
are

the experiment as an
analysis is based on the

Our

stimuli. The

the context

as

the

when

transmission.

to the 125 even-numbered

responses

ratios.
signal-to-noise
ratio
was
signal-to-noise

several

at

run

threshold.

calculate

to

were

obtained

odd-numbered

responses

are

subject judged the even-numbered


ignoredin this analysis.

designatedas the variable u. Last previousresponses


called "presponses"and they will be indicated by the variable v. These
are
the inputs. Current
are
are
representedby y. This is the output
responses
of
the
Thus
variable.
can
we
identify
joint event (i,j,m)as the occurrence
The

stimuli will be

to

response

stimulus

j. Failure

i, following presponse

possibleresponse. Consequently
and five response
categories.
as

The

subject's responses

from

to

contingency table. Two

4X5X5

table

this master

TABLE

Stimulus-Response

are

the

125

test

there

stimuli

of the reduced

reproduced here

are

respond is
stimulus

were

tables that

in order

to

Presponse-Hesponse

Table

Frequency

sorted

our

12

into

com-

Frequency

Presponse
3

sidered
con-

gories
cate-

obtained

were

illustrate

TABLE

Stimulus

to

four

Table

92

READINGS

The

for s,", goes

calculation

-^-[1 log2 1

s,"

374.05750/125,

sv.

2.99246.
same

s," is

way,

Response table,Table

^n

log21 +

s^^

372.38710/125,

s,^

2.97910.

The

for

log21 +

from

s,

[31 log2
yI^

31

Si

620.83188/125,

Si

4.96665.
is based

It

is evident

"

"

"

figuresfor

the

that

log2 2 +

logs30 +

these

"

"

33

entries

log^7+10

log^10],

in the

Presponse-

nj"

log29 +

log,3],

marginal of Table

log^33 +

the total number

on

log2125

"

the n. in the bottom

30

s,

computation for

log2 12 +

computed from

obtain the value

1 has

2:

sy"

We

Table

follows:

as

log25+12

s."

In the

PSYCHOLOGY

example, the Stimulus-Response plot in

putations. For
w,"

MATHEMATICAL

IN

31

1:

log^31],

of measurements:

6.96579.

calculations

are

wishes, the reader may

performed
also make

very

easily with

the

computations
(8),and Dolansky (3).
of p log,p tables for analyzing discrete data is not recommended,
The use
that the table of n log,n avoids.
however, because it leads to rounding errors
The complete set of s-terms in the experiment on frequencyjudgment worked
a

table of

with

out

log, n.

tables of p

as

If he

like those

log,p

follows:
=

1.45211

s,

4.96665

s.,

2.91389

s,

4.79269

Sir,

2.99246

4.93380

s,"

2.97910

6.96579

s.,"

In

Newman

preparedby

section

it

shown

was

that

s"

response

information,H'(y), can

be

analyzedinto components
H'{y)

HUy)

T\u;y) +

{v;y)+ A'{uvy).

(17)

WILLIAM

H'{y)

Since
had

2 bits. The

at most
not

respond.

and

2. The

This

extra

bits. This
for either

is the

Some

from

auditory

or

72

per

between

s.,-

part of the

by

cent

the

We

s,,"

"

did

the

or

of the

response

that
sequently,
Con-

presponses.

information

information

is

therefore

must

and the two

subject'sresponses

that

see

from

information

response

stimuli

of the response

28 per cent

to associations

be due

been

in Tables

the

the

1.46178/2.03199

unanalyzed error.

subject

right-handmarginals
equation (17) are easilycomputed

of the

accounted

bits. If the

2.03199

keys equally often, this figurewould have


shows
information
that the subject sometimes

be verified from

can

is 1.46178

HlXy)

H\y)

that

see

quantitiesin
example, Hl,{y) is computed

rest

For

s-terms.

we

the four response

used

is not

s"

93

MCGILL

J.

predicting

variables.
If

consider the association

we

T'{u;y)

T'{u;y)
only

.058 bits

transmitted

are

frequency stimuli,accounting for

the

from

3 per cent

consider

we

This

near

because
surprising

is not

threshold

masked

the

association

the

find

{y),we

responses

r(v;y)
value

response

The

of .218

bits

and

the

s,-

s"

stimuli

current

information:

4- s,"

.21840.
to

11

some

cent

per

of the

information.
last element

in

equation (17) is the

A'{uvy)

"s

A'iuvy)

.29401.

that about

14

interaction. Knowledge

is

computed

-\-Si +

per

s, +

cent

s"

stimulus

Sii

"

of the

s,".

"

response

also

of the interaction

"

s,-

Sa

permits

T'{u;y)+
.35181.

s," +

Sy" +

information

is:

Ti{u;y)

response

presponse

from

from
while
measuring transmission
inputs constant
stimuli
from
to
example, the transmission
responses
constant

(v) and

presponses

transmitted, amounts

interaction,A'{uvy).This

see

between

transmitted

little more

T'(v;y)

We

difficult to hear.
If

This

Srr. -\-s,"

information.

of the response
set
ratio was
the signal-to-noise
less than

Si

.05780.

were

responses

have

(y),we

Thus

auditory stimuli (w) and

between

s,i"

A'{uvy)

us

s,-,,",

is due

to hold

the
with

other

one

to

of the

input.

presponses

the

For

held

94

READINGS

3, 11 and
These

parts of the

the

14

per

to 28

figuressum

stimuli,presponses

for

cent

information

response

model, lead

the three-dimensional

analyze with

can

for

calculations

Our

PSYCHOLOGY

MATHEMATICAL

IN

cent, the

per

weights of approximately
interaction respectively.

of transmitted

amount

predictedfrom the size of the noise term. We can


weight directlyby computing the information transmitted
together.We have
s

divide

now

we

There

several

are

theory

this

to

component
to

the

sum

be

can

of squares
out
worked

second

for

process

analysis of variance
information

are

analysis of variance
multivariate

not.

are

distribution-free in the
test

square

The

of

number

discrimination
immediate
discussed

of

matter

that

is

of transmitted

No

exact.

the

partition

fact,a

notation

exactly parallelto the


(4).

transmission

information

is made

information

are

order

to

when

zero

to the restriction

categories,while methods

assumptions about

No

transmission.

linearityare

Furthermore,

later section,it will be shown


that

sense

they

are

extensions

when

based

on

introduced
statistical

that these tests


of the familiar

transmitted

of information

of amount

advantages. Garner

that the amount


of the

to

are

chi-

independence.

measure

inherent

As

similar

transmission

in discrete

information

developed in

are

measure

analysisis additive.
of error
(ornoise)

contingency-sense (asopposed
analysisof variance).In addition,the analysis

independence in

designedfor frequency data

tests

the

is very

by the

cent.

Furthermore, the analysis is

in

point

information

independent in the

to linear

in

plus the

analysisof variance.

is that

inputs

tion
applicationof informa-

our

first is that

in

contingency tables. Measures

variables

is

The

28 per

noting about

The

information.
involved.

both

Sii^

transmitted

figureof

of association

s-notation in multivariate
The

our

pointsworth

measures

approximations are
of

get back

experiment.

response

Sii +

"

three-dimensional

this

information, we

response

Sm

"

from

.57021.

T'(u,v;y)

sum

tion
informa-

also obtain this total

T'{u,v\y)

The

we

and

we

If

that

to

and

of information

Hake

also

has

certain

(5) have pointed out


approximately the logarithm

(2) and Miller

transmitted

is

perfectlydiscriminated input-classes.In experiments on


have
Like the one
we
discussed,the measure
provides an
of

picture of the subject's discriminative

apphcations of this property

in mental

abihty. Miller has also


testingand in the general

theory of measurement.
6.
It

is

Independencein Three-Dimensional

evident

from

the

definition

of

Transmission

transmitted

information

that

96

Now

ri(w;y)

Si

Ti{u;y)

s,-

T'Xu',y)

rXu;y)

T'{u',y).

kinds of

Both

Sii

s^

"

Si

"

Si,^

Sj

Sii"

s^,,-\-s

s"

"

equation (8).

Sim

between

and

Uiim

of classes in

V is the number
studied

have

the

Si

"

that

is the

happens

we

is not

do

not

only input variable


and

v).As might
generated from a single

between
be

F'
v.

independent of y. We could have


independent of v. The results are analogous to
where

case

independent of y, or
have presented.
those we

had

this

y. When

(providedthat no information is transmitted


be expected,both kinds of independence can
restriction on the data, namely

We

Si^

transmission,since

three-dimensional

where

s^

independence,(21)and (22),togethermean

in transmission

involved
have

s,-,".in

substitute for s,". and

we

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

is

7. Correlated Sources

of Information

for
information, T'{u,v;y),accounts
three-dimensional
a
contingency
only part of the total amount
table. It does not exhaust all the association in the table because it neglects
this association is considered,
i.e.,
the inputs. When
the association between
led
to
are
we
all the relations in the contingencytable are represented,
when
equation that is very useful for generatingthe components of multivariate
an

transmitted

Three-dimensional

of association in

transmission. Consider

C'{u,v,y) H'iu) +
=

If

we

add

and

H\vj + H'iy)

H'{u,v),we

subtract

(23)

H'{u,v,y).

obtain

+ r{u,v;y),
C'iu,v,y) T'{u',v)
=

C'{u,v,y) T'{u]v)-f T'{u;y)-f T'{v;y)+


=

We

see

that

C'(u,v,y)
generates all possiblecomponents

information-sources,u,

v, and

8. Four-Dimensional
It will be

instructive to extend

transmitted information
results

can

be

with

of the three

lated
corre-

y.

Transmitted
our

measures

Information
one

to
step further,i.e.,

three input variables,since

generaUzed easily to

(24)

A'{uvy).

an

from

iV-dimensional input.For

that point

simplicity

WILLIAM

we

shall restrict

input and
M

outputs does

with

us

not

present

add

us

We

u,v,w.

four

that

suppose

of information

sources

interaction

sends signalsh
and

u,v,w,

with

A^

multivariate

inputs and

be

can

constructed

input, u,v. The


1,2,3,

y. We

jointinput

W.

"""

This

proceed to define

can

gives
four-

follows:

Al{uvy)

already defined A'{uvy).The

have

case

with

clear.

information, A'{uvwy), as

A'iuvwy)
We

general

more

to the bivariate

w
w

channel

specialproblems, and

any

variable

new

of

case

the rules become

once

now

way

output. The

difficulty

no

Let
is

univariate

97

MCGILL

to the

development

our

J.

A'{uvy).

"

definition of

Ai{uvy) will be similar


except that the subscript w indicates that A'(uvy) is to be averaged over
w.
As we
have alreadynoted, this is accomplished by adding the subscripth to
each of the s-terms
that make
up A'(uvy).Consequently
Ai{uvy)

It is readilyshown
which

matter

"Sh

Shi

that

Ski +

Sh"

Shii

Sa,"

"

A'{uvwy) is symmetrical

variable is chosen

for

in the

Sa,-" +
sense

Snum

(25)

that it does not

averaging, i.e..

(26)

We

that

of information
A'{uvwy) is the amount
when
a fourth variable
by controlling
any

see

gained (or lost)in


three

mission
trans-

of the variables

are

alreadyknown.
If

table,we

all

examine

we

possibleassociations

in

four-dimensional

contingency

obtain

C\u,v,w,y)

T'(u;v)+

T'{u;w) +

T'{u;y)+

A'{uvw) +

A'{uvy) +

A'{uwy) +

T'{v;w) +

T'(v;y)+

A'{vwy) +

r(iv;y)

A'{uvwy),

(27)

where

C'{u,v,w,y) H'{u) + H'{v) + H'iw) + H'^y)


=

Equation
It turns

out

(27)

can

be

both

proved by expanding

that in the

H'{u,v,w,y).
sides in s-notation.

y) is expanded by writing
general case, C'(u,v,w,
for all possiblepairs of variables,and .4-terms
for all possible
"

"

"

down

T-terms

combinations

of

three,four variables

Four-dimensional
can
T'{u,v,w;y),

be

transmitted

written

as

and

so

on.

u,v,w

to

y,

i.e.,

follows:

T'(u,v,w]y) H'{y) + H\u,v,io)


=

from

information

H'{u,v,w,y).

(28)

98

READINGS

The

arguments

same

used

PSYCHOLOGY

MATHEMATICAL

to

justify(28)as
To

transmission.

in three-dimensional
we

are

IN

find the

used in the

were

components

of

(4)
of T'{u,v,w]y),
case

that

note

T'{u,v,w;y) C'{u,v,w,y) C\u,v,w).


=

This

that

means

correlations

the

are
T'(u,v,w;y)

all the components

contains
T'(u,v,w',y)
among

the

(29)

inputs. Consequently

C'{u,v,w,y)
except

of
the

components

of

"

T'{u,v,w;y) T'{u;y)+ T'{v;y)+ r{w;y)


=

A'{uwy) +

A^uvy) +

shown

components of T'(u,v,w;y)are

The

A'{vwy) + A'{uvwy).

in schematic

form

in

(30)

Figure 2.

T'(u,v,wiy)
Figure

Schematic
of

diagram

four-dimensional

2
of the

components

transmitted

with three transmitters


a

formation,
in-

and

singlereceiver.

If it happens that
rihii^

where
are

nii"/W,

that
of classes in w, all the components of C'(u,v,w,y)
In similar fashion,
drop out and C'(u,v,w,y) C'{u,v,y).

is the number

functions

of

C'{u,v,y)can be reduced to C'{u,y).This is preciselywhat we did in the


of independencein three-dimensional transmitted information. Since
analysis
of transmission with multivariate
T'(u;y),we see that all cases
C'{u,y)
inputs can be related to the bivariate case.
are
three inputs controlled,
we
ready to extend the analysisof
With
=

WILLIAM

response

in section

information

H'{y)
Equation (31) says that

we

4, a step further. We

HLM

can

due to the three inputs. This

from

mUy)

(31)

the effects in response


information
the fact that (30) tells us how

measure

is evident

HUy)

have

T'{u,v,w;y).

in its components.
expand T'{u,v,w;y)

to

99

MCGILL

J.

In addition

know

we

that

TUw;y),

(32)

where

TUw,y)
We

that

see

T'{w;y)+

in addition

w
controlling

transmitted

A'{uwy) +

between

to

and

A'{vwy) +
and

A'{uvwy).

v, enables

y from

to

us

(33)

rescue

the

noise,and to replace
with a better estimate of noise information,namely H^,^{y).
H'uviy)
The
evident. In general,
transition to an A^'-dimensional input is now

information

the

have

we

H'(y)

HL^...Xy) + T{u,v,w,

,z;y).

"""

(34)

z;y)can
(N -\- 1)-dimensional transmitted information,T'(u,v,w,
in the manner
that we
have described.
then be expanded in its components
The

"

"

"

9.

Miller and
to the likehhood

AsymptoticDistributions

Madow

(6) have shown that sample information is related


show that the
ratio. Following Miller and Madow, we can

mate
largesample distribution of the likelihood ratio may be used to find approxiinvolved in multivariate transmission.
distributions for the quantities
Consider,for example, three-dimensional sample transmitted-informatest the hypothesis that T(u,v;y)is equal to zero.
We
can
tion,T'(u,v;y).
the
This is equivalentto
hypothesisthat
(35)

V(i,j,in) PihJ)'Pi'"n),
=

since

independent.This hypothesis
the likelihood ratio [seereference (7)],

T{u,v;y)is zero

leads to

input and output

when

ri-"

n {n,r' n W"11 inurn)

If

we

take

logs,we

are

obtain
-2

log.X
_

1.3863
-2

n~

log.X

^'"

^'' "^

""'"'"

1.3863nT'iu,v;y).

'

(37)

100

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

For

large samples, "2 log. X has approximately a x distribution with


{UV
1)(F
1) degreesof freedom when the null hypothesis(35)is true.
is distributed approximately like x
if T(ti,v;y)
Thus
1.3863 nT'{u,v',y)
is
"

equal

"

to

zero.

in

Suppose

sources.

three-dimensional

our

P(hj"m)
hypothesis leads

This

three-dimensional

testing suspected information


that
example, we assume

involves

problem

important

more

p(i)-vU) p{m)

(38)

"

to the likelihood ratio for

complete independence in

contingency table,
n-'"
X

n (nr

(nr

U (n^'"

==r^

'

(39)

"

11 (ni,J"'""

After

take logs we

we

find that

loge\

"2

Ss

H\u)

"2

large samples

log,

{U
1)
I)
{UVY
null hypothesisis true.
-

also know

We

Si

"

Sj

"

Sm

"

"

H'(v) -f Wiy)

s,-,-"

H'(u,v,y)

(40)

l.S8QSnC'{u,v,y).

For

approximately a x^ distribution with


1)
(Y
1) degreesof freedom when the

has

(V

that

C'{u,v,y) T'{u;y)+

r{v;y) +

(41)

T'Mv)^

nT'(u',y)and
1.3863 nT'{v;y)are asymptotically distributed like x with {U
1)
1){Y
and {V
if T{u;y) and T{v;y)
1) degreesof freedom, respectively,
\){Y
find
of
To
the
distribution
make the following
we
zero.
are
Ty{u;v),
asymptotic
hypothesis:
likelihood ratio

The

used

be

can

show

to

that

1.3863

"

"

p^ij) is the conditional probabilityof j given


Now

we

have

(42)

Vihm)-Pm{j),

p(i,j,m)
where

m.

the ratio

'"'""^"'^

'=

li

"

^^

_
~

^""

*'"'""^ ^'"''"'

log.X

(43)

{ni,J

log,X

1.3863
-2

(H""

n (n..)"- n

"'

1.3863nri(M;").

.V

WILLIAM

In this

of

log. \ has Y{U

"2

case

(41) we

J.

1)(V

"

101

MCGILL

1) degreesof freedom.

"

write

can

+ T'{v;y)
+
1.3863nC'(w,v,i/)1.3863n[T'(w;y)
the

quantitieson

{UVY

"

freedom

"

rightside of (45)have degrees of freedom


F

"

2), Since this is the

that

to

sum

degreesof

of

(45),the quantitieson the right side of


asymptotically independent,if the null hypothesis,

as

(45) are

the left hand

number

same

(45)

T'y{u;v)l

The

In view

on

side of

is true.
This

that

means

as

approximation

an

we

under
Tl{u;v)simultaneouslyfor significance
stated. The

test

is very

similar to

tests
significance

this

we

need

will be made

data

in the

stimulus

2s +

C'{u,v,y)

.69055.

also need

with responses

predictthe

the information
Ty{u;v),
held constant.

This

expect much

transmitted

Sij^

Sj

"

transmitted
how

measures

information

T^(w;y)

Si

"

Since stimuU

auditorystimuU.

table. We

preponse

C'(u,v,y)
=

We

response

here. The
s,"

s," +

+ A\uvy),
T'{u',v)

"

5.

(45).To do

terms

not

were

of

amount

ciation
asso-

presponses

at

the

see

have

random,

computation

s"

can

in section

successfullythe

chosen

were

s"

"

from

"

example

our

Ty{u;v),since these
is the total
C'(u,v,y)

that

note

we

from

the quantitiesin equation

on

and
C'(u,v,y)

to compute

discussed in section 5. First

the null

analysis of variance. We

an

by applying the test to the


similarity
The

T'{u',y),
T'{v;y)and
hypothesiswe have

test

can

goes

to stimuli
presponses
we

do not

as

follows:

s,-,",,

.41435.

T'{u;y),T'{v;y) and
computed values for C'{u,v,y),
summarized
results
The
and
the
tests.
are
(45)
perform
x
level of
in Table 3. We
have not attempted to calculate the significance
do not have enough data to sustain the 88 degrees of
because we
C'{u,v,y)

We

may

now

put

our

Ty{u;v)into equation

freedom.
In any

The
case

same

Table

probably be leveled at our test for T'y{u;v).


effect in the experiment
that the only significant

criticism
3 shows

can

is the presponse-response
association.
One interestingfact that the analysis brings out

clearly,is that we
of transmitted information is big or small
decide whether
cannot
amount
an
without knowing its degreesof freedom. In our example we find that Ty{u;v)
and T'y{u;v)
.414 bits,while T'{v;y)
.218 bits. Yet
T'{v;y)is significant
=

is not.

The

reason

lies in the

difference

in

degrees of freedom.

Miller

and

102

TABLE

Table

Madow
measures

In

(6) have

discussed

of

Transmitted

Information

of statistical bias

the amount

in information

degreesof freedom, and have suggestedcorrections.


the association between
3, we tested T'y{u;v),
presponses

Table

held

responses

constant.
TABLE

Table

Probability

further in Table

not

are

of

Transmitted

This

association is broken

down

and
still

Information

estimated.

4. No

term, A'{uvy), because


A-terms

due to

stimuli with

**

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

probabilityis estimated
its

distributed hke

in Table

asymptotic distribution
the

difference of two

4 for the interaction

is not

chi-square.All

variables each

of which

chi-squaredistribution. The distribution of this difference is evidently


not
chi-square because the difference can be negative. Its densityfunction
has been derived by Pearson, Stouffer,and David
(9),but the writer has
In some
the problem can
been unable to find a table of the integral.
cases
T-terms.
be circumvented
by combining A-terms with T-terms to make new
[See,for example, equation (33).]However, in other cases, the interactions
in their own
are
right and should be tested directly.
genuinelyinteresting
available.
These cases
be treated when adequate tables become
can
has the

RANDOM

FLUCTUATIONS

OF

William

RESPONSE

RATE*

J. McGill

COLUMBIA

UNIVERSITY

A simple model
for fluctuating interresponse times is developed and
studied. It involves a mechanism
that generates regularlyspaced excitations,
each of which can
triggeroff a response after a random
delay. The excitations
not
are
observable, but their periodicityis reflected in a regular patterning
of responses.
The
is
probability distribution of the time between
responses
derived
and
its properties are
also
are
analyzed. Several limiting cases
examined.

number

of

that

behavioral

optic

trains of

horseshoe

preciselytimed

is illuminated

by
of

occurrence

found

in studies

The

is not

Under

close

the

example
for

is

is the

the

long

its visual receptor


with

comparable

conditioningwhen

the

rate

by reinforcingpaced responding;

point

bear

to

of intervals

ticking of

are

responses

in mind

in each

between

responses

watch

of these
is not

than

more

the

regular
ir-

random

components.
a

periodic system.
to be

wide
a

the

We

us.

for

is

so

article

of

in

intervals

distribution

model
the

with
model

but less stable than


the

capacity

these

to

found
it is

in

between
of these

both

periodic

will

then

take

be

completely
on

Since

extremes.

biologicalprocesses,

one

any

this

seems

physicalsystems

surprisingthat the

Httle attention.
to

examine

the writer

Psychometrika,

was

1962, 27,
104

properties of

the

producing noisy fluctuations

appeared

generated by

sequence,

*This paper
was
completed while
Lincoln Laboratory, Lexington, Mass.
This

to construct

periodic response

the

The

amounts.

of the type of randomness

attempt

an

of these

many

perfect, and

they will have


of possibilities
between

problem has received


mechanism

want

purely random

extension

paper

than

Intervals

orderly behavior

more

This

less

of

Moreover

range

natural

be

change in small

to

interests

stable than

of the

of electrons.

stream

itself to

seen

changes is what

more

of

scrutiny the timing

reveals

systems

to

resemble

heart

of the

tribution
Consequently the Poisson diswhich
is sometimes
proposed [11] to deal with rate fluctuations,
likelyto be very helpful.
fluctuations

and

essential

sponses
pulse-likere-

is famous

sequences

of operant

that the sequence

intervals

Another

limulus, which

is stabilized

response

is the fact

random.

mind.

potentialsit produces when

action

(see [5],pp. 498-502). The


examples

crab,

beating

constant
to

of

sequences

steady light [6].Response

also

periodicityare
of

springs immediately

the

of

nerve

generate

regularly in time. The

recur

illustration that

an

systems

in
a

3-17.

otherwise

an

elementary

constant

visitingsummer

Reprinted with

time

scientist at

the

permission.

WILLIAM

J.

105

MCGILL

intervals. Despite its

the mechanism
can
simplicity,
duplicatea variety of
observed
butions
phenomena, ranging from sharplypeaked and symmetrical distriof interresponse times to highly skewed
distributions,and even
completely random responding.Moreover, all these behaviors can be elicited
from the same
mechanism
by alteringthe rate at which it is excited.
Periodic Excitation
We

begin by examining interresponsetimes that are nearly constant.


is some
sort of periodicexcitatory
we
key to this regularity,
assume,
that triggers
after a short random
a response
delay. Even when the
process
excitations are not observable their effects are seen
in the regularintervals
between
The periodic mechanism
they impose
proposed here is
responses.
in
which
also
illustrates
notation.
our
diagrammed
Fig. 1,
The

"^1

"^2

"^S

Figure
Stochastic latency mechanism

yieldingvariable interresponsetimes

with a periodiccomponent.
regularintervals t, but are subject to random
delays before producing responses. Heavy line is the time axis.

Excitations

and

(not observable)

come

at

denote excitation and response

The
respectively.

time

interval

successive responses
variable and is called t. The
is a random
excitations is a fixed (unknown)
analogous interval (or period) between
between

two

constant

t.

and

Excitation
response

its distance

The

response

will almost

from

each

almost

always

excitation

be
can

never

coincide in time.

located between
be

expressed

as

two
two

quently
Conse-

excitations,
location

these,r, is the delay from a response to the next


excitation. The
second, s, is the corresponding interval between
and the excitation that immediately precedesit.
The

response

and

first of

basic random

the distribution

of t when

quantity in Fig. 1 is s, and


the distribution of

our

is known.

ordinates.
co-

lowing
fola

problem is to deduce
Accordingly,
suppose

106

that
times

READINGS

an

were

PSYCHOLOGY

MATHEMATICAL

exponentialdistribution
completelyrandom. Let

has

IN

Ks)

(1)

would

as

be the

if interresponse

case

xe-

/(s)is the frequency function of s, and X is a positiveconstant, i.e.,


constant. Equation (1) then describes a very simple delay process
of a response
in which the probability
during any short interval of time As
and equal to XAs
(see Feller [4],p. 220).
followingexcitation is constant
This defines what
mean
we
by "completely" random; the instantaneous
of response
is independent of time.
probability
in view
in trouble immediately,for (1)is not strictly
We
legitimate
are
be seen
from the fact that
of the requirementsjust set down. This may
"
"
mum
whereas the maxithe
interval
0
distributed
is
s
the exponential
on
where

the time

oo

value

of

If Xr is

in

Fig. 1

is

t.

circumstance, the

excitation and

delay between

average

because, in this

real trouble is encountered

sufficiently
large,no

is small

response

to Ei
excitations. Hence
the response
compared with the period between
before E2 comes
certain to occur
is practically
along, and the tail of the
distribution of s never
reallygets tangledwith the next followingexcitation.
it happens that Xr is not large,
When
a simple adjustment of (1)is required
tion
and r, without changing its characterizain order to bound
zero
s between
interval.
as a completelyrandom

With

Distribution of InterresponseTimes
Our

main

results

are

probabihty
and picturedin

describe the

outlined in the first section

the distribution
densityfunction describing

The

between

Periodic Component

given in (2)and (3),which

distribution of the mechanism

Fig. 1.

of the time interval t

is

tw^o successive responses

Xj/
1

sinh \t

t "

T,

-X(

t "

T,

(2)

1 +

J'
.

Xe
2p

in which
skewed

z/

is

and has

Whenever

constant
a

Xr

the distribution of

given by

i^

well-defined maximum

over

distribution is

evidently

t.

small,
large enough so that v is negligibly
to
time
in
(2)
interresponse
simplifies

happens

to

be

(3)

e"^^ The

Ki

r)

Equation (3)is the well-known


and sharplypeaked over
t

2'

-xu-co

"

"

cal
Laplace densityfunction [1].It is symmetriof
the latency
the behavior
t, and describes

WILLIAM

mechanism

by

the

the intervals between

when

component
periodic
then

must

be

J.

small

r.

"Noise"

dominated

successive responses
are
introduced by the random

comparison with

in

107

MCGILL

the

component
the

generatedby
periodicity

excitatory
process.
approximation in (3) is easilyrationalized if 1/X is considered as
1/Xr
component. In that case
measuring the magnitude of the random
the size of the noisy perturbationrelative to the period between
measures
The

excitations. Hence

the parameter

toward

will go

the ratio

whenever

zero

the random
small.
whenever
1/Xr gets small, i.e.,
component is effectively
It is not obvious that (2)approaches the Laplace distribution as v disappears,
but a brief study of (2)shows that this is in fact what happens.

Proof of the Distribution*


shall

We

show

now

interresponsetimes when
shown

in

(2) is the

that

triggered
by periodicexcitations

are

responses

of the distribution of

form

correct

as

Fig.1.
be

First of all,(1) must


handled.
is easily

beginwith

We

hold

adjustedto

and

zero

t.

This

simplycyclethe exponential

excitation and

an

between

lettingthe distribution
until it reaches t again,and repeating the process
continue
down
to run
to any
ad infinitum. The ordinate corresponding
point s between zero and
distribution back

will then

be

to the

originas

will be distributed

reaches

r,

/(")
=

distribution

is now
excitation,
in (la)yields

^^^"'^

of r, the interval from

/w

Evidently r and
excitation period.On

are

0"s"r.

the response

determined,since,from

(4)

excitations.

Fig. 1, s

Hence, when

writer is indebted

original
proof.

to

r.

o"r"r.

only

intervals

one

between

referee for

can

response

to

in the

same

between

occur
are

responses

belong to one excitation period and s will belong


making r and s independent for determining t.
of r and
It should be clear that t is not just the sum
*The

"

correlated
perfectly(and inversely)

the other hand

following
Substituting

the next

to
=

^^^'^

will

the

excitations

in the interval between

response

as

(la)

two

given by

of the
Consequentlythe position

The

as

soon

analyzed, r

later one,

althougha

suggestingseveral excellent ways

to

thus

cursory

simplify

108

READINGS

examination

IN

PSYCHOLOGY

MATHEMATICAL

The trouble with the impression


Fig. 1 leaves that impression.

of

periodsmay separate Ri from R2 In other words,


excitation may
not forced by excitations. A new
are
come
along
responses
1
before the response
is emitted. We have drawn Fig. as though response R2
reflection suggests
fellinto the excitation periodfollowing
Ri but a moment's
that thingsmight not happen so neatly.To deal with this nasty eventuality,
is that several excitation

shall define t

we

as

(5)

where

is taken

the time

as

is the

i.e. the number

Equation (5) is
whose
of

distributions
and

s,

interval between

analogousinterval between

s
excitation,
precedingit,and

occurs,

kr +

of

now

are

Ri and

the next

R2 and

mediatel
the excitation im-

periodsin which no response


empty excitation periodsbetween Ri and R2
of quantities
of t in terms
unique specification
k is the number

of

known

as

soon

as

X and

are

fixed. The

in (4) and (la).Our


alreadybeen specified

have

following

distributions

step is

next

to

find the distribution of kr.

beginningwith

interval

An

and terminating in

excitation

an

response

This latency is
occurs.
may
span several excitations before the response
denoted by kr + 5, where k takes on values 0, 1,2, 3, etc. It is evident from
bution
leadingup to (la) that kr -{"s has the exponentialdistriout to infinity.
the delay from an excitation to the first
Accordingly,
be resolved into two
can
subsequent response
independent components:
the location of response
(i)the number of excitation periodspassed,and (ii)
R2 in the periodbetween the last two excitations. In view of the independence
the arguments

of k and s,

we

can

write
Xe

-\(.kr a)
\e~''''^''
+

T^/7

where

Pikr) is the probabiUtyof

(6)

Pikr)

In other

words, the distribution

at successive

multiplesof

All three components


r

and

can

form

of

The

(7)

"

find that

of kr. We

v).

of kr is geometricwith ordinates

spaced out

independent.Moreover, the variables


Consequently,(5)

for each value of k.

same

to read
i

where 0 "

/(I

(5) are

unit that is the

be amended

value
particular

t.

(5a)

some

Pikr)

-\-s "

kr +

y,

2t,and kr has the geometricdistribution givenby (6).

distribution of y is obtained

from

the convolution

of

and

s.

After

find that
we
simplification

=\csinh\y
[csinhX(2r

"

y)

"

"y

f(y)

"

"

r,

2r,

WILLIAM

J.

109

MCGILL

where
\v
c

(1
The

the

distribution

pair of

in the interval. The

of excitations

/(I

/.(y)

(8)

it

separatelyby linking
density function

of

component

fixes k, and

of the distribution

component

new

f{t)will be indicated
t. Equations (5a) and (6) yield

A;th harmonic

number

each component

to describe

of t. It will be convenient
it to the number

determine

k will define

change in

of excitations between

the number

on

each interval. This

that bound

responses

that each

follows

depend

of t will

^r

fk(y),since

as

of the

and

v)/(y).

times with a
example, /o(?/)is the density function of interresponse
of f{t)
each pair of responses.
This component
singleexcitation between
For

has k equal to

and is defined

zero

the interval 0 "

over

2t. The

^ "

average

interresponsetime in the interval is r.


The

first harmonic

fiiy)spans values of
is 2t. Higher harmonic
The
function

each

between

excitations

just two

refers
fi{ij)

component
t between

pair

of responses.

and

St. The

components

are

that for values

has contributions from

two

harmonic

unity and
time
interresponse

Hence

average

defined in the

it evident

foregoing makes

with

interresponsetimes

to

same

k is

way.

oi t "

components

the

in each

density
interval

lengthof an excitation period.The pair of contributors


change as we proceed away from the originin multiples of r, but every
out to have two
element of densityin f(t)after /
t will turn
components.
Specifically,
corresponding to

the

will

U(y) + h.^iy

/(/)

(9)

If the

densities

the

on

expressionsdetermined

r)

2r.

r"y"

right-handside of (9) are replacedby equivalent


(8),it is easilyshown that

from

(9a)

/(/)

^Xe"

'.

simply another way of writing t,and it is apparent


that
in a way
of f(t)interlace themselves
that the harmonic
components
for the distribution of / :
simple expression
produces a surprisingly
Now

recall that kr -\-y is

sinh \t

Xe-''

t "

~,

i
^ +

"

2^

This is (2) and

the

proof is complete.

r.

110

112

This

zero.

and

we

are

can

necessary

of
Refer

lapses
geometric distribution in (6) colof r and s,
t in Fig. 1 will be just preciselythe sum
0. Hence
of
excitation
Two
periods.
responses
ignore the possibihty empty
there must
also be two independent occurrences
to define t. Hence
Call them
R2
respectively.
S2 corresponding to Ri and
Si and

follows

when

s.

and

Fig. 1

and

S2

we

can

m.g.f.of
given in Table
for M ,{B),
we

the

write

the

as

"

S2

difference

Si

"

exponential

of two

generatingfunction

its moment

as

MXe)MX-d).

exponential distribution

1 for the variable

Si,

is distributed

"

that

observe

M,_.(^)
The

the

Consequently,
variables

the fact that

from

to

now

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

kr +

s.

is,of course, very familiar and is


Substitutingthis exponentialm.g.f.

obtain

Mt-r{e)-'

(1 -^A)(i

^/x)

ieW~,

is,as we have already shown, the m.g.f.of the Laplace distribution.


Evidently the Laplace density function (3) is in fact simply the distribution
two
of the difference between
exponential variables. This simple point is
ignored in most texts on statistics because, perhaps, no one imagines why
estabhshes
be interested. Our
else would
a
good
argument
very
anyone

which

for

reason

being

provides a characterization
periodicexcitation.

under

of the

between

also behaves

latency mechanism
excitations
has

of the mechanism

another

gets
a

very

small.

in

error

the

Laplace

timing

device

in

We

an

now

fairlyslow response,
rapid succession.

in

interestingway
suppose

the

that

but is bombarded
The

Vim

fit)

the

as

that

is

period

delay part

by

excitations

restriction is achieved

following one
by fixingthe delay time constant, X, while allowing r
zero.
Equation (2) ior f(t)immediately leads to:
(14)

bution,
distri-

Excitation

Continuous
The

hence

difference,and

The

interested.

to

bolically
sym-

approach

Xe"".

T-.0

0 and t
t
t
portion of f{t)between
is
the
constant
approaching
v
must
disappearas t approaches zero. Meanwhile,
of (2)
the
of
falls
out
the
Umit
for
portion
right
f{t)
unity. Consequently
defined for t " t. The
same
exponential limit can be obtained by studying
with X fixed,
the behavior
of the m.g.f.for f(t)in (12) as t approaches zero
The

or

result is almost

by analyzing

the

obvious.

The

variance

of

fit) under

these

same

limitingconditions.

12
INTERRESPONSE

General

distribution
is

curve

of interresponse times

(standard

units)

with

plotof equation (2) in the text, with

and

arbitrary random
X

and

1. Dashed

periodic parts.

lines

are

harmonic

of the distribution.

components

given in Table 1. The


harmonics
the variance
(i.e.,

Variances

TIME

Figure

The

113

MCGILL

J.

WILLIAM

formulas

are

establish

that

the component

vanishes, and the


y) disappears as
harmonics.
b
etween
entire variance
becomes
differences
This imphes that the probabihty distribution of f{t)must
congeal around
its harmonic
peaks (seeFig. 2) when
goes to zero, and that each peak then
contributes
"line" of density to the resulting exponential distribution.
a
be contributed
that no
by
delay can
Intuitively,the limit in (14) means
within

of

in the

concentrated

the
is
is

latency

between

and

response

the

next

excitation.

in Fig. 1 vanishes
r
instantly available. Hence
consumed
by the latency between excitation and

assumed

to be

and

the

response,

That

excitation

entire
which

interval
we

have

exponential.
Applications

compiled from a long .series


of limidus.
of action potentials recorded
a
on
singlefiber of the optic nerve
that the data are
periodic and the
The
distribution demonstrates
narrow
fiber.
to originatein the refractory period of the nerve
periodicityseems
The
mechanism, however, is not well understood. In this particularcase,
achieved
the regular sequence
by dissecting out a
of action potentialswas
Fig.

presents

frequency

distribution

114

IN

READINGS

PSYCHOLOGY

MATHEMATICAL

0+20+40

-20

-40

FROM

DEVIATION

PERIOD

Figure

nerve

when

continuouslyto

the

milliseconds.

291

of 303

the eye

intervals

Measured

is

fiber of the
the

nerve

fiber

produced

into the control gate of

pairs of responses

were

deviations

Laplace

shining a

beam

of discrete

barrage

was

out
a

onto

1000

of

attached.

was

magnetic tape.
digitalcounter, and

read

from

linear

the

drift. Smooth

curve

distribution.

on

the gate

passed through

are

the fiber

to which

recorded

amplified and

and

optic nerve,

ommaiidium,

the

interresponsetimes observed in a single fiber of the optic


fiber adapted
illuminated
was
by a steady light.The nerve
linear
increase
in
slow
in
period from 261 to
illumination,
resulting a

Frequency distribution
of limulus

( milliseconds

cps

Later
time

permanent
sine

wave

light on

the

receptor, i.e.,
steady illumination,

Under
responses
on,

which

the

tape

was

intervals between
record. The

then

were

played
alternate

timing signal

generated by

calibrated

is of the order

of the measurements

tuning fork oscillator. Over-all accuracy


2 milliseconds,
due to variations in the speed of the tape recorder.*
The
fiber adapted continuously to steady illumination,resulting
nerve
about
261
milliseconds
to 291
in a slow increase in the basic period from
data
the
in
blocks
isolated
of
milliseconds. This change was
by averaging
which
and fitting
25 intervals,
a line to the averages,
quite
fortunatelywere

of "

linear. Measured
from
is

between

this

responses

line,yieldingthe frequency
Laplace density function.
*The

and

intervals

the writer

with

distribution

made
the assistance

preparation and recording were

analyzed by

were

in

Fig. 3.

by C. G. Mueller.
of Michael

into deviations

converted

S.

The

The

data

Kennedy.

fitted

were

curve

recovered

WILLIAM

normal

A
a

data

taken

in which
last
in

payoff

Fig.

normal

was

dashed

on

the

the

distribution

normal

Fig.

and

obtained
a

main

and

interrespbnsetime

in

was

by

approximation

class interval

immediately
effect. In

evident,and

white

to

as

is taken

by mea.suring
rat.

Hill's data

The

data

schedule
from

the

is shown

used for this purpose


because
after reinforcement
believed to be unare
related
the

event

any

not

w^ere

leptokurticcharacter

of Hill's

it suggests that the

long regimen of training(184 hours)


Hill's rat into a fairly accurate
problem made

discrimination

time

of

high shoulders

frequency distribution in the background. This


and variance
by matching mean
sponses
to the data. Re-

fitted

of responses
to

is

distribution

bar-pressesmade

The

response.

curve

have

day of conditioning with a reinforcement


contingent on delaying at least 21 seconds

in the 0-3 second


bursts

the

would

93rd

was

by the

data

illustration is provided

successive

the

on

previous

data

defines

1 15

MCGILL

same

reported by Hill [7].The


between

intervals
were

fact then

leptokurtic.Another

being
from

fitted to the

curve

flat top. This

J.

Laplace-type clock. We are led naturally to conjecture about how the rat
Does it happen internallyvia some
constructs
t.
type of neurologicalclock or
via
of
movements?
a stereotyped sequence
externally

.50

.40

"o
z
UJ
z"

.30

o
UJ

UJ

.20

.10

s,
0

12

15

18

INTERRESPONSE

TIME

Figure

Distribution

of

conditioningon
seconds

from
and

interresponse times
a

last

schedule

previous

demonstrates

in which
response.

27

30

36

33

(seconds)

produced by
reinforcement
Dotted

24

21

curve

bar-pressingrat
is

contingent

is best

peaking of empiricaldistribution.

on

after

long period of

delaying

fittingnormal
(Data from

at

least 21

approximation
Hill).

116

READINGS

Skewed

with

interresponse times

of

distributions

PSYCHOLOGY

MATHEMATICAL

IN

the

appearance

of

usually in connection with


(2) (seeFig. 2) are found often in the literature,
studied
[2] who
high speed responding. Fig. 5 is taken from Brandauer
small
illuminated
at
a
generated by a pigeon pecking
sequences
response
and the bird
controlled
Reinforcement
a
was
high
speed
flip-flop
by
target.
reinforced whenever
a particular
a peck happened to coincide in time with
was
forcement
of the two states of the flip-flop.
Consequently, the probabilityof reinone
determined

was

in that

the net

state, and

per

1000

result

of the

tail which
In this

interval

in turn
it is

case

rhythmic
use

faster than

even

that every

between

reflects

likelythat

to

second.

This

flip-flop
spent
same
(low)

the

generate high

distribution
Intervals

of 1000

excitations

follows because
the

by

were

the average

exponential

excitation.
varying degrees
the period t is constructed
by a pre-programmed
that humans
head something like the mechanism
rates

of

20

(second!

TIME

.23 seconds

"

10"^

interresponsetimes

longer than

follow

tapping.

Figure

rate.

that

of failure to

INTERRESPONSE

high

the
had

is increased

responses

10

Frequency

time

response

conclude

would

we

5.3 times

oscillation of the

in order

was

of

The

reinforcement.

periodicexcitatorymechanism,
coming
length

proportion

rate of 5.3
pigeon generated an average
second
mately
approxiduring the run shown in Fig. 5 which covers
created
is
in
5
the
in
fact
If
by a
Fig.
sharp peak
responses.

probabilityof
responses

the

by

not

recorded
shown.

from

pigeon pecking

(Data from

Brandauer).

at

WILLIAM

In

recent

interresponse times
in the

Hunt

paper,

spinal cord

and

recorded
of the

J.

117

MCGILL

Kuno

several

[9]present

distributions

of

during spontaneous

cat. The

data

activity of single fibers


gamut from the Laplace to

the

run

the

exponential,includingseveral examples of what appears to be our skewed


distribution
(Fig. 2). The effect is exactly what might be expected, if the
same
general response
subjected to varying rates of periodic
system were
excitation.

Discussion
be

It would
fiber

activity and

times

When

by

study

be

can

neurons

give

organization may
provide

constructs

have

r, we

of the mechanism
The
into

as

delineation

of the

have

to

device

and

the

excitation

for

summating

For

one

in

the

is

the

response

are

for

or

closelyrelated to
required,and

number

[3].These

it is not

noise

how
earlier,

kind

with

in

an

thus

the animal

conception

our

of noise

our

presents

insights

of the

general

higher

whether

the

barely scratches the


has indistinguishable
there is

new

is

simply

parallelchanneling present
distributions
show

response

replaces
by

blocked

suggestions
themselves.

of interresponse
response
inter-

clusteringsof

period, and
components.

entirelyclear yet what

excitation

this is clear other

distributions

the

[1].

before

trigger,or

of harmonic

centers

interestingproblems

very

mechanism

appears

Once

similar

detecting periodicityin

in this paper

multiples of a fundamental
(2) which also has harmonic

else is

system looks

organization,and

its parameters

that

already working.

Hterature
at

out

difference

no

excitations
there

times

so, the

compatible

for

excitation

new

reactivates

that

example,

times

estimate

to

considered

it makes

earlier one,
old

did

what

suggests possible ways

It turns
possibilities.

the

that

we

is

indicates

mechanism

limited

the noise.

by

mechanism

excitations. Whenever
an

that

way

in the

of

attempting

latency

surface x"fthe

to

dictated

ask, as

we

single
interresponse

operating in overt
complicated systems of

Fig.

of the

nature

instance, the Laplace distribution

noise. For
to

coding

face, and

The

to find

and

affords
simple periodic mechanisms
fibers. Ejiowledge
coding in single nerve

information

form

responses,

than

the

to

of

neither

simple jobs. Even

very

it. When

study

to

way

clue

distributions

like

more

no

do

to

apart than

the other.

mechanisms

means

organized

between

organized than

stochastic

responding, it probably

the

further

applicable to both,

seem

of the time

better

find

we

responding. Yet

overt

complicated or

more

find levels of behavior

to

in this paper

presented
afforded

view

hard

that

hence
But

seem

to

something

something

else is.

REFERENCES

[1] Arley, N.
New

York:

and

Buch,

WUey,

K.

1950.

R.

Introduction

to the

theory of probabilityand

statistics.

118

[2]

of

rate

[3]

[5]

the

F.,

Bronk,
1946,

C.

Ferster,

and

H.

Hartline,

1940,

Amer.,

[7]

Vol.

[8]
[9]

Hoel,
Hunt,

K.

G.

P.

[11]

Mood,

[12]

Widder,

[13]

Wilks,

Manuscript
Revised

C.
Nat.

and

S.

received

manuscript

N.

Ann.

nerve.

Y.

of

ed.)

New

York:

Wiley,

1950.

York:

New

reinforcement.

Appleton-

fibers

the

of

the

visual

J.

pathway.

Soc.

opt.

Differential

mathematical

to

Physiol,

Mathematical

the

low

responding.

rate

N.

River,

Y.,

P.

theory

36,

calculus.

and

New

York:

New

evoked

York:

Wiley,
of

response

of

statistics.

some

among

McGraw-Hill,

E.

New

1954.

spinal

1950.
of

measures

conditioning.

123-130.

statistics.

8/28/61

ed.)

discharge
364-384.

relationships

1950,

12/14/59

received

147,

1959,
to

Sci.,

Pearl

Co.,

(2nd

statistics.

Background

M.

Theoretical

G.

of

reinforcement

Cyanamid

American

383-415.

Kuno,

Advanced

D.

in

behavior.

pp.

Acad.

S.

Schedules

messages

nerve

Introduction

Mueller,
Proc.

(1st

applications.

its
F.

Laboratories,

/.

A.

B.

Introduction
C.

C.

of

239-247.

30,

interneurones.

[10]

The

11/10/59,

4,

excitation

response

1958.

1957.

Lederle

of

Report

and

theory

Operant

T.

R.

Hill,

Chemical

the

upon

University,

457-485.

Skinner,

Century-Crofts,
[6]

reinforcement

Columbia

thesis.
M.

Larrabee,

and

D.,
47,

Probability

W.

Feller,

doctoral

Unpublished

of

probabilities

uniform

of

effects

The

pigeon.

Sci.,

Brink,
Acad.

[4]

C.

Brandauer,

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

York:
Princeton:

Prentice-Hall,
Princeton

1947.
Univ.

Press,

1943.

p Sound^

treated

Figure

diagramof

Schematic

equipmentwith

the

room

circuit used in
equivalent
intensity.

the

computationof

the

size of the increment in

AP/P

AP
140

10

2000

10

fc-1000

0.006
0.008

120

0.01

"

0.1"-

kr

100

1.0^^1.0

't-

0.02

100
ir- 10

0.3

-h:

^-

0.04

0.1^^0.1
~^

0.5

0.06

5;

1.0

a,
--

"

0.08
60

"0.1

0.1

0.01 ^ir

0.2

0.01

40
^0.01

3 -^-

0.4

0.001

20
-

5"-

0.6

0.001

0.001

0."
0

1.0
Figure

Nomogram

to convert

0.0002

values of AP/P to AP

120

0.0002

when

P is known.

0.0002

GEORGE

Resistance

Decade
The

attenuators

Rq and

Box.
used

were

schematic

keep

to

Rj^, surroundingthe shunt

into the

in

shown

Fig. 1.

whole

the

of the

equipment

values

of

constant

resistances,Ri and

system

which
can

this circuit,the

For

121

MILLER

diagram

of the increment

computation

resistance, 7?2- The

A,

source

by the insertion

produced
representedby

size of

AEl

the

increment

an

in

Fig. 1.
impedance,

load

R2, since these values

is

be

is shown

and

must

enter

of the variable

equivalentcircuit,also
voltage^El is given by

in

R^R^Rl
'

RiiRoiRi

El

If

the

does

system

introduce

not

R2+

R3)

of its sensation
for the

level

noise.

the number

"

R2)]

after the
increments
amplitude distortion
be taken
pressure, expressed in decibels, can

in sound
produced, the increment
+
AElIEl).
201ogio(l
Throughout the followingdiscussion
terms

Rl(Ri

the

of
intensity

of decibels

above

the

noise

are
as

will be stated

the listener's absolute

in

old
thresh-

If the

sound-pressurelevel of the noise is taken to be the level


the
generatedby a moving-coilearphone (Permoflux PDR-10) when the voltageacross
is
the
the
for
soidal
sinusame
as
a
earphone (measured by a thermocouple)
voltagerequired
to generate the givensound
in
of
volume
6
wave
then
(1000 cycles)
a
cc,
pressure
for the noise correspondsto a sound
the absolute
threshold
of
approximately
pressure
10 db

level

pressure
sensation
150

by

level.

and

7000

c.

Once

p. s.

sensation

the

discussed

into

noise

and

by

value

passes

to

of the

spectrum

into

converted
the

adding
uniform
was
relatively

the

dynes/cm^will

be

can

10 db

noise

value

sound-

given

for

the

db) between

("5

transduced

the

by

Hawkins.^

relative

of the increment

line which
straight

scale,and

level

in detail

sound-pressurelevel
value

of

and

measurement

been

decibels

convenience.

of

spectrum
The

the absolute

convertingthe

the

simple procedure

has

the

known,

hand

the

The

PDR-10

earphone
are

dyne/cm^. Thus

0.0002

re

size of the increment

be

find the

through

in decibels

Those

computed.
of Fig.2
nomogram

can

value

of the sound

interested
a

in

considerable

of A/ in decibels

the left-

on

the middle

scale,will intersect
through a
pressure on
When
the stimulus
is a
right-handscale at the appropriatevalue of AP in dyne/cm'^.
sound
its
acoustic
in
is
plane progressive
wave,
intensity watts/cm^ proportionalto the

the

of the

square

The

pressure

peak

kp^.

"

amplitudesin the

of

wave

white

noise

expect, therefore, that the size of the justnoticeable

to

function

of the

aspect of the

distribution

stimulus,

passed through
increments
altered

were

by

uniform

the

square-wave

in

amplitudesin the

peak
experiment .was

second

difference
In

wave.

conducted.

generator (Hewlett Packard,

square-wave
introduced.

level. The

randomly

of

and

The

order

The
Model

wave
resulting
frequency.

might

might
to

vary

evaluate

noise

210-A)

be described

as

before
are

"squared oflF"

square-

wave

as

this

voltagewas

qualityof the noise


subjective

spectrum
generator, but the peak amplitudes are
form

It is reasonable

constant.

not

are

the
not
at

modulated

masking of pure tones and of speech by white noise," in a report


Report No. 5387 (Psycho-Acoustic
by noise, OSRD
Laboratory,
masking ofsignals
Harvard
Services,
1, 1945) (available through the Office of Technical
University,October
U.S. Department of Commerce,
Washington, D.C.).
^

entitled

J. E.

The

Hawkins,

"The

122

READINGS

The
as

PSYCHOLOGY

differential sensitivity
procedure for determining
was
experimental
and

of

increment.

an

noise

the

The

observer, seated

monaurally through

listener heard

continuous

the

percentage

heard

which

to

(1.5 sec.

Four

Decibels

Two

Time,

of

5 to

points on

threshold
two

obtained

was

series

as

Noise.

Listeners

Could

Function

of Sensation

Hear

used

were

to

used

The

A
periodically.

presented,

to determine

determine

each

Increments

50

each

Percent

in

of the

Level.

psychometricfunction, and from


Thus
500
by linear interpolation.

listeners
experienced

listened

room,

added

were

the

impendingpresentation

sound-treated

was

same

was

for Intensity
of
Sensitivity

Which

intervals of 4.5 sec.) was

at

such

TABLE

DiflFerential

in

difference

only

indicate the

increment

an

duration

tabulated.

was

alone

The

the

high quality,
dynamic earphone (PDR-10).

noise,

series of 25 identical increments


and

MATHEMATICAL

Volkmann.^
employed by Stevens, Morgan,
of a signallight
which
theysometimes used to

that

omission

to

IN

this function
to 800

the

differential
each

of

at the

16

judgmentsby

differential

threshold

different intensities.
Results
The
time
be

are

noted

increments

in decibels

presentedin Table

as

which
a

the

function

two

listeners

of the

could

sensation

hear

50

level of the

percent
noise.

of the
It will

for "square-wave noise"


is not significantly
sensitivity
in
the
noise.
the
fluctuations
peak amplitude
greaterthan that for random
Apparently
do not influence the size of the justnoticeable
of the wave
The
increment.
response of
the ear
And
since the
is probably too sluggishto follow these brief fluctuations.
the two
difference between
of the phase relations
forms is essentially
matter
wave
a
*

that

S. S.

discrimination

the

differential

Stevens, C.
of loudness

T.

Morgan,
and

and

Am.
pitch.

J.

J.

Volkmann,

Theory of

Psychol,1941, 54,

the neural

315-335.

quantum

in the

123

MILLER

A.

GEORGE

3.5

3.0

:9 2.5

2.0

.5,L5
LO

0.5

20

10

40

30

50

60

Sensation-level

Figure

in decibels

of the noise

of

purposes

the

among

components,

differential

eflfecton

The

data

comparison.

we

value

0.048

or

energy,

the

to

is

in

curve

of

Fig.3. The

obtained

by

For

pressure.

Karlin^

with
of

The

range

is indicated

of the

intensity

presentedfor

are

have

important

no

group

At

which

for sound

of 0.099

is proportional

increment

the

portion of

horizontal

the

threshold,

highestintensities

the

Weber-fraction

over

by

the absolute

above

more

of intensities agree

this range

over

or

constant.
approximately

corresponds to

stimulation

values

purposes

Knudsen^

db, which

0.41

for sound

level

tones

phase relations

that these

that, for intensities 30 db

indicate

is about

function

sensitivity.

the relative differential threshold


the

are

solid line representsEq. (2).

The

conclude

may

100

plottedas a
hearing. Data for

of

threshold

the

above

90

50 percent of the time

heard
in intensity

Increments

70

in decibels

quitewell with

the

solid

the values

of 50 listeners.

comparison,Fig.3

includes

obtained

data

from

differ

by

Riesz^

those

and

obtained

by
for

markedly
intensities.
at low
Possibly
quite different, especially
introduced
the
data represent sensitivity
Knudsen's
to the "noise"
by
abrupt onset of
of
of his use
because
intensities
his tones, or possiblyRiesz's data at low
are
suspect
for
obtained
in intensity.
Data
beats to produce increments
by Stevens and Volkmann^
with
of
four
intensities
to
more
tone
seem
listener at
a
closely
a single
1000-cycle
agree
mine
the present results than with Riesz's, but their data are not complete enough to deter-

noise, but

800

Riesz's

function.

c.p.s. which

these

studies

R.

R.

results do

Knudsen's

for tones.

data

are

King, and
favorablywith

Churcher,
compare

indicate

not

that

the

Davies^
the

for

with

reported data
Riesz.

Taken

intensityis

of the

of

function

limen

difference

have

of the
sensitivity
Riesz, Differential intensity

ear

for pure

tone

of

together,all
order

same

tones,

of

P/iys.Rev.,

1928, 31, 867-875.


"

V. O.

and frequency,
Knudsen,Thesensibilityoftheeartosmalldifferencesinintensity

Phys. Rev., 1923, 21,


"^

S.

1940, 92,
*

S.

Stevens

84-103.
and

J.

Volkmann,

The

quantum

of sensory

discrimination.

Science,

583-585.

B.

of
intensity

G.
a

Churcher,
pure

A.

tone, Phil.

J.

King,

and

H.

Mag., 1934, 18,

Davies, The
927-939.

minimum

change
perceptible

of

124

READINGS

magnitude for

noise

it is for tones, at least at the

as

for

intensities the discrimination

lower

PSYCHOLOGY

MATHEMATICAL

IN

higher levels

stimulus

noise

be

may

At the
intensity.^

of

somewhat

acute

more

for tones.

than

Implications
for a Quanta! Theoryof Discrimination

is not

units

neural

Only

It is

new.

neural

limen

difference

the

evidence

has

discreteness

to

of discrete

selves.
sensory cells themthe

support

discrimination

mediatinga

processes

of the

obtained

been

activation

the

depends upon
the

suggestedby

however,
recently,

basic

the

that

that

notion

The

of

are

assumption
all-or-none

an

character.

evidence
principal

The

Volkmann^"

Stevens, Morgan, and


We

sensory
excites

these

number

quanta
This

bring into

the
activity

added

left-over stimulation

fluctuation

evident

that

the

over

When

the

increments

added

are

neural

to

in effect, a
The

constant

error

of

that

"

with

the

increment,

to

expect?

an

individual

difference

of

If [the over-all

quantum,

it is

equally

occur

is added

taneously
instanthis fraction

itself.
the listener

stimulus,however,
stimulus

from

the

changes
In order to
sensitivity.
ignoreall one-quantum changes.
in his

conditions

these

to

fraction of the time, and

of fluctuations
forced

A/,

much

How

to

changesin the

must

activate

at least two

and reported.
Thus,
perceived
function.
the
to
psychometric
be described
line of reasoning
can

will be

is added

quantum

one

psychometricfunction

it will
little to spare
additional
some

excite

surplusstimulation

continuous

distinguish
one-quantum

units in order

which

stimulus

that, if the increment

the size of the increment

judgments,the listener is
under
Consequently,a stimulus increment
additional

we

of the

it follows

reliable

make

are

the size of

to

all values

constantlyoccurringbecause

are

with

a certain
perceived

stimulus, it will be

finds it difficult to
which

of time

considerations

The
.

so

perceptionof

in the

units.

for discrimination.

surplus excitation

to
directly
proportional

in

contribute,along

needed

to the

is

course

these

From
.

will

quantum

or

psychometricfunction.
the following
way :

surplusinsufficient to

is largecompared
sensitivity]

in

often.

ordinarilydo

will

small

the

involved
initially

distinct
functionally

quanta

leave

shape of

the

structures

into

surplusstimulation

quantum.
this

of

and

from

present the argument

neural

the

divided

are

certain

excite

that

assume

continuum

derives

predicted
by

this

less than
to a steady sound
are
followingway.
value A/q, they are never
some
reported,and over the range of increments from 0 to
remains
the
function
at 0 percent. Between
A/q and 2AIq the
A/q
psychometric
with the size of the increment, and
proportionof the increments reportedvaries directly
reaches
100 percent at 2 A/q. Such a function is illustrated by the solid line of Fig.4.
is reported50 percent of the time is
It will be noted that the difference which
If we
take this value as defininga unit
to 1.5 times the quantalincrement.
equivalent
functions
increment
in the stimulus,all the psychometric
obtained for the two listeners

in the

can

When

into

be combined

scales
intensity

coincide
^

at the

Of

the

50

"

stimulus

singlefunction.

which
against

the functions

increments

In other
are

percent point.In Fig.4

the

See reference 4, p. 317.

Soc.

we

can

adjustthe

to make

Am., 1941, 12,

with
disagreesstrikingly
and

R. M.

517-525.

individual

all the functions

size of the relative increment

investigations,
only Dimmick's
higherintensities. F. L. Dimmick

in audition,/, acoust.

words,

plottedin order

modern

reportedhere for the


limen

the

Olson, The

in sound
the values

ence
intensive differ-

A.

GEORGE

125

MILLER

2Q

IQ

AP/P adjusted to 1.5Q

psychometricfunctions

32

of the time

to

this value.

pressure, AP/P, has been


the

time

is

Each

the

classical

difference

attributable

to

Is there

is, for

white

means

that

and
value

quantal increment.
quantal function was
of
described
by the phi-function
the

dashed

that

by

an

level

any

noise
the

is

amount

the

of the

the cumulative

When

chance.

time

increment
the

be

in the
The
is

hypothesis.Any
produce

the

S-shaped normal

It should
one

of the

"

function:
467.

source

G. A.

Miller and

Implicationsfor

the

W.
a

trolled
con-

reasoning

statistical

surprisingindeed

to

of the

nature

noise

obscure

in view

And
if the

of the fluctuating

rigorousexperimental

situation

predictedby

the

the

demonstrates
the

step-wiseresults

quantal
and

to

probability
integral.
shape of the psychometricfunction is only
the slope
quantalargument. According to the hypothesis,

be noted, however,

of
implications

If this

are

are

probable value,
merely the most
will depart from
this probable

increment

tends
variability

which

variables

quantalhypothesisshould
the listening
situation.
experiment? Certainlythere

fulfilled. This
requirementsof the quantal hypothesiswere
functions
in obtaining the rectilinear
practicaldifficulty
of

variables

these

the

sufficient to affect the discrimination.

stimulus, it would

function
probability

into
variability

of randomness.

of the

portionof the

in this

(the normal

gamma

is revealed."

Fig.4 from

of random

paradigm
value

pointsin

of randomness

source

calculated

certain

of

obvious

percent of

obtained

not

of small, indeterminate

according to

introduction

the

50

line.

applicationof

number

combine

the deviations

heard

was

the

assumes

which

which

eliminated,the step-wise,
"quantal" relation

or

is correct, then
be

that the increment

for the

argument

limen

and
independent,

the characteristic

that

experiment.The data are better


indicated
probabilityintegral)
by
to

in

adjustedso

1.5 times

plottedas

Figure4 shows

The

singlegraph. Values of APjP heard 50 percent


the datum
pointson each function are plottedrelative
point represents 100 judgments.

combined

designatedas 1.5Q, and

are

3Q
j.n.d.

Figure
The

R.

that the

Garner,

the psychometric
on
presentation
1944, 57, 451discrimination.Am. J. Psychol.,

Effect

quantaltheory of

of random

126

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

of the

psychometricfunction is determined
by the size of the difference limen for all
of stimulus-intensity.
The
with
this second
present data accord
prediction.
The
standard
deviations
of the probability
the
which
describe
data
are
integrals
one-third
the
0.5
all
for
the
thresholds
measured
for both
means
(or
approximately
AIq)
subjects.This invariance in the slopeof the function is necessary but not sufficient
for a neural quantum, and it makes
evidence
of the results
possiblethe representation
in Fig.4.
in the form
shown
values

order

In

to

Symbolic Representationof

the

represent the experimentalresults

in

symbols will be used

Data
the

symbolicform,

following

numerical

constant

numerical

constant

==

1.333,
0.066

A/^// when

" /",

difference limen

noticeable difference)expressed
in decibels,
(just
in
second,
/ frequency cyclesper
sound
/
intensity
(energyflow),
/~
sound intensity
per cycle,
sound
which is justaudible in quiet,
/q
intensity
sound
which is justmasked
in noise,
I^
intensity
in
sound
increment
AIq quantal
intensity 0.667^1^,^,
heard
50 percent of the time,
A/gQ increment in sound intensity
DL

in sones,

loudness

masking in decibels,

of

number

quantal increments above threshold,


ratio per cycleat
signal-to-noise
any frequency,

effective level of noise at any

Nq

An

empirical
equation
where

the

variable
time

we

A/g^
"

the solid curve

in

the

Fig.3

are

"Weber's

Law,"

a
1-

German
181.

the

H.

Helmholtz,

ed..Vol.

II. The

be

can

developed from

the

(1)

I"

which

increment

have

to

be heard

can

101ogio[l+

justnoticeable

know

the

only

computations
The

obtained.
obtain
that

which

of the law
Treatise

at

fixed

50

percent

and

of the

on

sensations

fit of this

that

the

values

size of

it is added.
at

to the form

low

in decibels
I and

through,the
curve

to

of the

in
expressed

as

/q and

function
not

their

values indicated

the data

is

by
good enough

function.

(1) is

to
equivalent

justnoticeable

Differential

intensities,and

(2)

1.5b(IJI)].

ratio between

high intensities Eq.

which
Law

1.5c

increment

carried

smoothed

states

to
proportionalto the intensity
departs from Weber's

modification

I "

is assumed
stimulus-energy

are

of Eq. (2) to
the use
justify
It is interesting
to note

known

Table

write

can

level,althoughwe
When

in

bl"

101ogio(l+ A4o//)

values.

absolute

data

to compute the
(2) it is possible

of sensation

to

Since

component.

the

in the

quantalincrement

DL
From

M^

equals1.5A/q,

"

frequency.

adequate descriptionof

the

well-

difference

is

teristically
characsensitivity

Fechner

Eq. (1).^^The

long ago suggested


essential feature

physiological
optics(translatedby P. C. Southall from
of vision, 1911),Optical
Societyof America, 1924, pp.

3rd
172-

128

READINGS

IN

MATHEMATICAL

White

Noise.

PSYCHOLOGY

TABLE

of White

Masking
of

Values

of the

Masking
Masking

It is

givenin

Noise

by

Obtained
Noise.

Table

I, and

function

values

of

determine

into

of the sensation

level of the

computed

are

intensities

intense

than

able

to

obvious
when

with

used

over

for
wide

masks

this noise

less energy

is needed

than

the

of the

noise

step is to
to

purposes

when

the

when

Eqs.

two

which

results of

Fig.5 where

masking noise.

(3a)

converting

thresholds

are

masking is plotted

addition,Table

In

(1) and

the information

II contains

combined:

are

(4)

masking noise

threshold, the

is about

12 db

more

ask whether
mask

these

tones

or

results

human

correspondto

the functions

speech. Fortunately,we

are

masking effects of noise on


equipment directly
comparable
the

here.

which

forms

in

The

Eq. (4).

to

into masked

then

the

Level

noise.

is used

1000-cycletone

when

Eq. (3a).
and

question.Hawkins^ has measured


and
with
speech
experimentalconditions

those

find

into

Iq from

and

A/q

and

+bl
101ogio[(f///o)

above

more

of

shown

are

in Decibels

of the Sensation

Masking According

this

Suppose,
We

next

noise

answer

and

tones

or

masked

the

The

obtained

25 db

Increments
Function

values

values

listeners,and

masking which

of

quantalincrements

M
For

the

substitute these

to

as

Values

II for the two

givenin Table
a

Listeners

Computed

possibleto

now

the differential thresholds

as

for Two

II

Quantal

energy

range

of

is 20 db
we
itself,

for
is

comparison,

of intensities

less intense.
conclude

when
audibility
the
over
spread

Since

the
entire

to

mask

white
particular

the

that, for this

masking functions, therefore,we


the 1000-cycle
tone.

masks

choose

we

this

that

energy

tone.
1000-cycle

noise

justmasks

corresponding value
is concentrated
In

subtract

order
8 db

at
to

from

1000

compare
the

is 12 db

noise spectrum,
specific

spectrum.
can

8 db

c.p.s.
the

level of

GEORGE

A.

129

MILLER

90
80

:^7o
a"

I 60
"

50

if

5 40
00

.E30
(a

S20
10

-10

10

20

30

40

Sensation-level

60

50

of noise

in

increments

experiments. Solid

of white
intensity

line represents function


and

When

and
tone
1000-cycle
obtain
we
intensity,

is

from

taken

curve,

remarkably

function

to

warrant

noise

by

of

by

of

the ratio

at that

data, and

the

as

Fig. 5.

for

noise

of the corrected

function

The

data

level for Hawkins'

in the noise

this
correspondence between
in the present experiment

pointsobtained

computed from Eq. (4) falls


in Fig.5.
separate presentation

close

too

to

Hawkins'

of

choice

Munson^''

function

in

noise.

lOOOc.p.s.is not crucial to this correspondence. As Fletcher


pointed out, a singlefunction is adequate to describe the masking
is a
noise is corrected by a factor which
of
pure tones, if the intensity the
This factor is givenat any frequency/
the frequency of the masked
tone.
tone
of the masked
to the intensity
R of the intensity
per cycleof the noise

The
and

white

function

The
its

analogous to masking
plottedin a manner
the masking of tones
Hawkins
for
by

of this tone

solid line shown

Hawkins'

close.

100

obtained

of 8 db

plotthe masking
the

noise

speech by

this correction

make

we

90

Figure

Discriminable

80

70

(///qin decibels)

have

frequency: R

intensities well

JJI'^.

above

is

threshold"

determined
experimentally
rectilinear portionof

the

on

for all
the

at
frequencies

function

in

shown

Fig. 5.
noises

For
attributed

to

the

with

continuous

noise

band

in the

it is convenient

to

spectra,the masking of
of

relate the

tone

of

frequency/ can

be

immediately adjacent f}^


frequencies
of frequency/tothe intensity
of
tone
a
masking
to

sequently,
Con-

of
in decibels re the threshold
and to express this intensity
/',
be
which
can
regarded
hearingat any frequency. This proceduregives10 logio(/~//o)'
of the
level
Z
effective
The
noise.
of
band
the
sensation
level
at /of a one-cycle
as
noise at that frequencyis then defined as

per

cycleof

the

noise

at

13

Soc.

Am.,
"

H.

Fletcher

and

W.

A.

(5)

+
lOlogio/?.
10log,o(/~//o)

Munson,

Relation

between

loudness

and

1937, 9, 1-10.
H.

Fletcher,Auditorypatterns.Rev.

Mod.

Plivs.,1940, 12, 47-65.

masking, J.

cicoiis.

130

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

masking of pure tones is plottedas a function of Z, the relation between M


function
of frequency.A single
to be independent
expresses the relation
and Z for all frequencies.
between
M
M
in
When
obtained
to Z with the function
we
relating
compare the function
find
that
the
sensation
of
level
the noise is equivalent
the present experiment,we
to
When

and

the

is found

Therefore,

ll.Sdb.

15.14/?(/~//o).

///o

this expression
into Eq. (4) gives
Substituting
M

equation,
along with the
compute the masking of pure
10 logio(/~//o)
is greater than
be
computed more
masking can
This

Hawkins'

show

results

+6].
101ogio[i?(/~//o)

functions
tones

by

random

any

enables
Iq to frequency,

15

noise

of known

for
db, b is negligible
10

to

us

spectrum. When

and
frequencies,
+
logio(/~//o)simplyas
log^,/?
that the function
of Eq. (4) can
also be adaptedto
about

of human

the

and

R
relating

(6)

all

the

10

white

scribe
de-

noise.

speech by
masking
the correspondence
the masking and the masked
seems
complete. When
sounds
to changes in intensity
are
are
identical,masking and sensitivity
equivalent.
obtained
with
identical
and
masked
The results
noises are directly
masking
comparable
Thus

to

with

results obtained

different

the determination

that

of

masked

sounds.

to
sensitivity

It is reasonable

changes

in

to

is a
intensity

fore,
conclude, there-

case
special

of the

generalmasking experiment.
of masking is also applicable
It is worth
to visual
noting that this interpretation
of white
and
to changes in the intensity
sensitivity
light.Data obtained by Graham
of the similarity
of their
Bartlett^^ providean excellent basis for comparison, because
and
because
they used homogeneous,
procedure to that of the masking experiment,
more

rod-free, foveal

plottedas
function
and

we

retina.

of visual

measures

that

of the

areas

have

used

When

these

masking,the result
to

express

the

data

are

substituted

be described

can

auditorymasking by

by

noise

into
the

Eq. (3) and

same

of tones,

general
speech,

noise.
Relation

to

Loudness

adopted the just noticeable difference as the unit for sensory


is stillalive today : Are equally-often-noticed
a controversy which
scales,he precipitated
?
the
In
to be
of auditoryloudness, the answer
differences subjectively
case
seems
equal
much
noticeable
differences
intensities
at high
are
subjectively
(j.n.d.'s)
negative.Just
at low intensities.
largerthan j.n.d.'s
When

^^

Fechner

C. H. Graham

eye: III. The

and

influence of

The
N. R. Bartlett,

area

on

foveal

relation of stimulus

and

in the human
intensity
1940, 27,
Psychol.,

discrimination, J. exper.
intensity

149-159.

Crozier

has

used

detectable

similar

visual data

to

of the just
reciprocal
lightintensity
by a normal probability

demonstrate

that the

increment
is related to the logarithmof the
is determined
by the not-alreadyintegral.This is deduced on the assumption that sensitivity
excitable neural effects. Crozier's equaexcited portionof the total populationof potentially
tions
here. W. J. Crozier, On the
of the auditorydata presented
givean excellent description
Proc. Nat. Acad.
law for minimal
discrimination
of intensities. IV. A/ as a function of intensity,

Sci.,1940, 26, 382-388.

GEORGE

In order
need

we

to

the

two

be

can

in

is the

need

We

If these

as

the

threshold, and

Tone

as

Sones.

for

Data

Listeners.

12

of

The

tones,

intensity
loudness
subjective
was
rightand j.n.d.'s

not

agree, Fechner

was

III

Sensation-Level

of Sensation-Level

Function

of Quanta.

Number

for pure

to the

correspond,Fechner
If.they do
loudness-scale.
on
a subjective
pictureis more
complex than he imagined.

the

as

noise
relating

functions

two

TABLE
and

well

as

the functions

units

and

Loudness

for noise

case

to know

distinguishable
steps above

sones.

used

wrong,

that such

of information.

of

number

of the noise

to demonstrate

kinds

131

MILLER

A.

of

of

Noise, with

The

Last

Column

Quantal

Units

Above

Gives

1000-Cycle
Equally Loud
Loudness
in
Corresponding
the Corresponding Number

Threshold.

of differential

level
to a givensensation
quanta Nq corresponding
of
scale
a
readilyobtained
by "steppingoff" the quant.alincrements
against
consists
of
the
number
of
increments
The
unit
procedure
finding
quantal
per
number

of noise is
decibels.

of

and
intensity

then

integrating:

NQ=(llMQ-dI.
If

increment
substitute for the size of thee quantal
qu

we

When

we

solve

convert

in terms

We

assume

loudness

in

of the noise with


to the
alternately

equations were
The

of

that the number

Eq. (4) indicates that M


Eq. (9) are given in Table
The

base

logarithmsto the
masking M, we find
Nq

same

made

result of this

III, and

sones

was

the loudness
ear,

by

and

experiment
"

10, insert the values

for the

and

constants,

that

3.49M

(9)

K.
is

Therefore, K

zero

when

"5.1.

Values

/",and

at

of

obtained

Nq

this

point
by

plottedin Fig. 7.
ness
determined
by requiringlisteners to equate the loudThe two sounds
of a 1000-cycle
tone.
were
presented
of the tone.
Five
the listener adjustedthe intensity

of twelve

each

(8)

-JnMQ+C.
-

quantalincrements

1.46 db.

6/o

to

of

accordingto Eq. (1),

-dl
77-rTr-dr
"TT
cl

^Q-

(7)

the

listeners

level of the

for

the

six noise-intensities

1000-cycletone

which

sounds

studied.

equal in

132

READINGS

IN

MATHEMATICAL

20

40

PSYCHOLOGY

60

Figure

and

Observed

defines

"

III and

in Table

tabulated

values

in

from

sones

givesthe

Stevens'

standard

indicated

are

by

the

been

loudness

for the

constructed

loudness-scale^^

deviations

are

Standard

of

deviations

lengthsof the vertical bars.

plottedin Fig. 6, the


has

noise.

level of the noise.

the loudness

which

the loudness-scale

from

also

noise

to the

loudness
are

computed

values for 15 listeners

the

level of white

of the loudness

values

100

80

of noise in decibels

Sensation-level

included

With
in

these

is determined

sones

1000-cycletone.
in Table

of loudness

of the distributions

data, which

III.

Table

levels obtained

The
III

for

the 12 listeners.
Loudness
for

also be

can

loudness
calculating

computed.

from

the

Fletcher

masking

and

which

Munson
the

sound

developeda procedure
this
produces. When

300

225

75

150

50

25

75

20

40

Figure

Comparison

of the number
noticeable

!'"' S.

S. Stevens

and

of

80

60

discriminatory
quanta
increments
of intensity
are
H.

Davis, Hearing,New

100

of noise

Sensation-level

with
not

York:

the loudness

of white

equal.
subjectively

Wiley,1938, p.

118.

noise.

Just

GEORGE

procedureis appliedto
the

computed

shown
in Fig.6.
quitesatisfactory.

The

the

masking

equippedto present the


number
of quantalunits as

functions

two
a

by noise, we get
computed and experimental

of pure tones

between

agreement

now

are

shows

curve

for the

data

values

results is
We

Hawtcins'

133

MILLER

A.

in

shown

of sensation

function

Fig.7.
level.

solid

The

The

dashed

100

10

1.0

0.1

0.01
6789

45678

"

100

Masking (A7/7o in decibels)


8

Figure

Relation

curve

the

shows
error

the

loudness

of Fechner's
are

When,

not
as

S. S.

in sones.

The

and

masking

for white

discrepancybetween

assumption. Loudness

and

noise.

these

two

the

number

of

are

both

related

curves

affirms

justnoticeable

ferences
dif-

related.
linearly
in the

to determine
possible
"

loudness

between

present

case,

their relation

Stevens, A

two

to each

variables
other.

scale for the measurement

Rev., 1936, 43, 405-416.


Psychol.

of

Stevens^^ has used


a

to

third, it is

Riesz's

data

for

psychological
magnitude: loudness,

134

the

arrive

to

tones

pure

in

tone

sones,

is the

in Table

of the

most

the

is

and

loudness.

noise

an

alternative

In

the

itself. Let

on

Fig. 7

Table

the

expression L
in

form
than

KM^

loud

is related

fits the
the

to

of the

power
the

cast

we

Fig.

relation

data

third

rather

of the

power

logarithm

of the

that

j.n.d.'s

over

why

that

it

along

of

in

masking

to

be

can

creases
in-

noise

itself,

on

whatever

intensity.In

faint

that

seen

white

the

level;

the noise

smaller

are
j.n.d.'s

of loudness.

scale

the

of

defined, and

of

in

sensitivity

to
sensitivity

relation

loudness

it is obvious

equal

noise

masking

so

Fig. 8

increment

units

tinguishabl
dis-

but

sensation

to

The
In

The

quantal

not

are

the

masking produced by

empiricalequation, however,

j.n.d.'sand

that

masking,

level.

well.

functions,

computed

functions.

two

for

well

differential

is related

masking

data

rather

notion

the

between

of

of

apparent.

between

we

sensation

to

these

and

the

power

is not

developed

II

with

loudness

number

relation

be

to

is the

is the

the

relation

masking,

the

out

tones

the

Table

by combining

proportion

third

i.e.,the

of

turn

and

we

where

computation
describes

both

state

kN^-^,

first step, and

kN^

section

5 and

loudness

III

to

way

examine

now

In

is obtained

loudness

specialcase

the

for noise

preceding

us

subjectiveloudness.
and

that

different

be

intensityis

in

changes

sones

that
It is interesting

should

There

find

we

of

parallelStevens'

we

III,

range.

exponent

in

size

When

steps.

presented

empirical equation

the

at

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

Acknowledgment

in

valuable

criticism

and

advice

who

S. S. Stevens,

preparation of

the

during

who

Shirley Mitchell,

his gratitude to
Miss
author
wishes
to
express
obtaining the experimental data, and to Professor

The
assisted

this

tributed
con-

manuscript.

Correction
In

1963

is written

D.

as

added

pressures
pressures,

from

recomputed
above

noise

12

alone

in

Osman,
and

M.

acoust.

to

the

and
V.
Soc.
When

obtains.

db

there

therefore,
able

Table

I
the

threshold,

(not

were

of

20

noise

of the

Rich;

page

1962,

(cf. Fig. 1),

effect

for

At

db

25

then

would

noise.

This

sinusoids

model

instead
have

fact
has

for

In

of

of

been

was

been
monaural

masking
25
the

than

about

below

verified

if

db

is
or

masked
25

db,

listeners

masking;

inaudible

sound

summed

their

intensities

intense
levels

their

so

amount

for

more

sensation

which

Eq. (3), which

of

square
the

log^o(AP/Pq),
is about

Energy-detection

the

When

pressures.

in

error

an

generated independently.

was

(negative masking)

"masking"

similar

Mathews,

noticed
been

power

128).

in-phase increments

absence

Am.,

on

facilitation

was

hear

squared

masking

stated

as

combined

using

had

perfectlycorrelated

were

their

Rich

E.

noises

masking

their

phase;
sum

and

Osman,

and
noises

two

in
the

not

E.

Raab,

masked

the

fact, however,

more

H.

if the

presented

directlyby Raab,

reported by

S.

auditory

M.

Pfafflin

detection,

/.

34, 1842-1853.

this correction

is

made,

of course,

the

relation

shown

in

Fig. 8

no

longer

136

READINGS

the variance.

MATHEMATICAL

it also enables

And

based
are

the

different

on

good

some

So

ing
adopt-

for

reasons

The

of variance and amount


similarity
of information
might be explained this
When
have a large variance,
we
way:
we
are
ignorant about what is govery
ing
If we are very ignorant,
to happen.
then

when

make

we

gives us a lot of
other hand, if the
know

we

must

system,
what

into

goes

great

deal

comes

out.

their

the

variance

If it is

their

(or

say

this

what
That

is to

upon
with the

to the

of

So

circle the amount

are

Whenever

that

the
is

see

in and
the

say,

is

the

measure

you

whenever

transmitted

understand

"amount

to

of

will understand

I refer to "amount

information," you

"covariance"

or

If

be

will

by

two

The

and

the
to

of transmitted
observer's

will be

absolute
then
will

recoverable

If he makes

transmitted

formatio
in-

errors,

information

may

considerablyless than the input. We


increase the amount
expect that,as we
of input information,the observer will
and more
more
begin to make
errors;
we

reasonable

amount

kind

then

of

server
ob-

If the human

judgments.
system,

the

of his

test the limits of accuracy

can

absolute

of communication

when

input

increase

we

the

information
will increase

information

at

eventuallylevel off at some


value.
This asymptoticvalue
asymptotic
take to be the channel
we
capacity of
firstand

will

it represents the

of information

amount
us

about

an

absolute

the

stimulus

judgment.

is the upper
to which

can

the

by

information.

be

tion."
"correla-

graphically
partiallyoverlapping cir-

situation

stimulus-

measured

as

the

transmitted

formation,"
inthe observer:

ance."
"vari-

overlap the

amount

transmitted
a

in his

judgments are
quite accurate,
all
the
of
nearly
input information

is

measure

amount

experimentalproblem is to increase
and
of input information
amount

much

mission.
trans-

the

stimuli,the right

of transmitted

"noise"

simply
input-outputcorrelation.
two
simple rules to follow.
refer

the

correlation

the

during

information

we

And

The

and

Then

of information

then

how
or

in the

be

to

represent the

of information

system,

variance

input and

channel.

left circle would

ment,
judg-

absolute

is considered

communication

atic
system-

of the

There

his responses.

the

fluctuations
the system

by

of transmitted
measure

experiments on

the observer

from

goes

of the output

to random

introduced

of

also

input, or
input. If we
then we
can
correlation,

much

attributable
is due

that

in terms

some

depend

will be correlated

how

the

In

tion.
informa-

of transmitted

amount

amount

information).

be

must

out.

will

measure

and

communication

there

comes

output

system

relation between
what

cation
communi-

realize

described

good

however,

the

as

response

variabilityabout what
The
input and the output
be

taken

variance of the input,


rightcircle the variance of the output,
and the overlap the covariance
of
and
input
output. I shall speak of the
left circle as the amount
of input information,
the right circle as the amount
of output, information,and the overlap

of

therefore

can

observation.

about
great deal of variability

be

can

the

formation
get little inresponses,

we

will

you

the

is very small,
tion
observaour

how
so

it

On

information.

making the
now
imagine

will

If you

observation

variance

out,

come

from

there is

the

in advance

the left circle

cles.

metrics.

concept.

newer

Then

ferent to represent the


dif-

in quite
experimental situations where it
ances
varibe meaningless to compare

would

PSYCHOLOGY

to

us

results obtained

compare

there

IN

be described

Now

that he

can

to the

stimuli

just a

brief word

we

can

give

basis

channel

The

limit

the observer

the

on

greatest

on

of
pacity
ca-

the extent

match

his

sponses
re-

give him.
about

the bit

GEORGE

and

we

One

bit of information

begin

can

information
decision

that

man

than

six feet tall and

the

chances
bit

one

if

or

more

know

we

information.

to

way

does

unit

that
refer

not

of

"

Two

bits of information

decide

four

among

Three

among

decide

Four

five among
lyl,and
12
if
there
are
say,

so

know
So

five successive

every
is increased

There

rule
of

two

are

could

of

give information
would
time
the

rect.
cor-

simple:
bit

one

to

the

Or

increase.
variable

stimuli.

data
the

is the
in

of

from

ordinate

more

With

per

14, the

from

Fig.

stimulus.
tones

was

mation
input infor-

1 to 3.8 bits.

plotted the

1.

of input

amount

bits

alternative

2 to

is

in

plotted

bottom

increased
the

or

listeners

the

tones

are

the number

increased

five

frequent.

were

information
As

confusions

tones

with

were

them.

On
of

amount

at

which

observer,so

that

unit time

per

could

we

we

ignore the

completely and increase


by
input information
number

the

In
we

of

absolute

We

give

alternative

judgment

interested

are

alternative.

periment
ex-

in the second

the

observer

as

Fig.

time
we

must

but

tones

mistakes.

many

Along

correct

confused

never

different

different

These

three

or

confusions

the

tone.

of

increasing the

of

two

quite rare,

made

told

was

made

input information.

the rate

of information

amount

much

the

in

crease
in-

might

we

ways

increase

the amount

is

is

spect
re-

is added.

the amount
We

four

alternatives

factor of two,

only

With

fore
each, be-

alternative

number

by

of information

bit

general

the

is to

covered

listener had

of the

the listeners

alternatives, fourteen

one

which

the

time

When

the

he

response

tones

That

and

8000

After

identification

natives,
equally likelyalter-

worth
binary decisions,
we

his

bits of

mation
infor-

to

different with

were

100

numeral.

were

on.

make

must

able
en-

to

to
equal
range
cps
logarithmicsteps. A tone was sounded
and
the listener responded by giving a

to

us

tones

from

happens
judgments of

(17) asked listeners


by assigningnumerals

frequency,

to

eight equally
16

among

The

what

absolute

Pollack

them.

Uni-

of

stimuli

consider

us

make

tones.

natives. used
equally likelyalter-

likelyalternatives.

we

enable

bits of information

decide

to

us

let
we

need

we

length that
we
use
feet,inches, centimeters,etc.
However
the man's height,
measure
you
stillneed
bit
of information.
we
justone
any

Now

identifytones

that

Notice

of information
the

whether

50-50, then

are

of

this unit
in

decide

six feet tall

Judgments

dimensional

ternatives.
equally likelyalwhen

must

we

Absolute

of

make

to

137

MILLER

data.

some

need

two

is less than

at

is the amount

we

between
If

look

to

A.

as

he wants

will appear

callinghis

begin
near

increase

stimuli

discriminate

confusions

his

sponse;
re-

1.

and

the number
which

among

look

to

see

Confusions

to

occur.

the

point that
capacity."

"channel

he

where

we

are

Data

from

Pollack

(17, 18)

on

the

by
of
judgments
of input information
auditory pitch. As the amount
2
is increased
by increasing from
the number
of different pitches to be
14
to
tion
informaof transmitted
judged, the amount
limit a channel
approaches as its upper
capacity of about 2.5 bits per judgment.
amount

simply

alternative

to make

listeners

of information
who

make

that

is transmitted

absolute

138

READINGS

transmitted
of

information.

transmitted

much

IN

The

MATHEMATICAL

amount

behaves

information

PSYCHOLOGY

in

the way
would expect a comwe
munication
channel to behave; the mitted
transinformation

to about
an

increases

2 bits and

asymptote

then

linearly
up

bends

about

at

off toward
2.5 bits.
LOUDNESSES

This

is what
value,2.5 bits,therefore,
we
are
callingthe channel capacityof
the listener for absolute judgments of
pitch.
So
bits.
that

now

What

have

we

does

the

it mean?

number

2.5

First,note

15-110

12
INPUT

DB

INFORMATION

Fig. 2.

Data from Garner (7) on the channel


audicapacity for absolute judgments of tory

2.5 bits

loudness.
correspondsto about six
equallylikelyalternatives. The result
that we
Next
than
ask how
cannot
means
can
pick more
reproducible
you
this result is. Does
six different pitchesthat the listener will
it depend on
the
confuse.
ferently,
never
spacingof the tones or the various conditions
Or, stated slightlydifhow
tive
alternaof judgment?
Pollack
varied
matter
no
many
these conditions in a number
ask him to judge,the best
of ways.
tones we
be changed
can
we
can
expect him to do is to assign The range of frequencies
them to about six different classes without
by a factor of about 20 without changing
the amount
of information
mitted
transerror.
Or, again,if we know that
there were
than a small percentage.
alternative stimuli,then
N
more
Different groupings of the pitchesdecreased
his judgment enables us to narrow
down
the particularstimulus to one
the transmission,
but the loss
of
out
small.
For example, if you
was
can
N/6.
Most
five
that
the
discriminate
high-pitchedtones in
people are surprised
number
series and five low-pitchedtones in
is as small as six. Of course,
one
there is evidence
another
that a musically sopect
phisticated
series,it is reasonable to exall ten into
that you could combine
with absolute pitch
person
a
singleseries and still tell them all
can
identifyaccuratelyany one of SO
When
I
60 different pitches. Fortunately,
or
apart without error.
try it,
you
nel
chandoes
work.
The
do not have time to discuss these reit
not
markable however,
be
about
for
to
tunate
capacity
pitchseems
exceptions. I say it is for-

because

I do

not

know

how

to

six and

that is the best you

can

do.

While
we
are
on
tones, let us look
explaintheir superiorperformance. So
I shall stick to the more
pedestrianfact next at Garner's (7) work on loudness.
data
for loudness
Garner's
that most
of us can
are
marized
sumidentifyabout one
in Fig.2. Garner went to some
out of only five or six pitchesbefore we
trouble to get the best possiblespacing
begin to get confused.
the intensityrange
of
his tones over
It is interesting
to consider that psychologists
He used 4, 5, 6, 7,
have been using seven-point from 15 to 110 db.
10, and 20 different stimulus intensities.
rating scales for a long time, on the
results shown
in Fig. 2 take into
The
intuitive basis that trying to rate into
the differences among
add much
does not really
account
finer categories
subjects
mediately
influence of the imlack's and the sequential
to the usefulness of the ratings. Polresults indicate that,at least for
judgment.
Again
preceding
to be a limit.
find that there seems
we
pitches,this intuition is fairlysound.

GEORGE

A.

139

MILLER

and

Garner

markers.

Fig.
two
^_l.9
BITS

(8) asked observers


two
visually between
Their

They

4.

In

ways.

observer
and

use

100

to

TA$TE$
JUDGMENTS

OF

INPUT

Fig.
and

Data

3.

(1)
absolute judgments

results

on

Beebe-Center, Rogers,
the channel

of

observers

five

perfectlydiscriminable

capacity for

values

of loudness

Since

these

different

is 2.3

studies

two

laboratories

absolute

ments
judg-

bits,or

about

alternatives.
done

were

in

were

to

to

channel

with

capacityfor

saltiness.

The

shown

are

in

the

experiment in
version they let the

between
zero
any number
describe
the position,although

this

unlimited

technique are shown by the


the graph. In the other
on

INFORMATION

from

O'Connell

one

scale

they presented stimuli at only


5, 10, 20, or SO different positions.The

SALINE

CONCENTRATION
12

results
did

terpolate
in-

to

limited

in

version

the

their

sponses
re-

lus
reportingjust those stimuthat were
That
is
possible.

in the second

say,

response

filled circles

version

the

ber
num-

of different responses
that the observer
could make
the
was
exactly
same
the number

as

the

of different stimuli

experimenter might present.

that
The

with

ferent results with this limited response


nique
techslightlydifand
methods
of
shown
the
circles on
are
techniques
sis,
analyby
open
in a good position to
the graph. The
not
functions
we
are
two
so
are
whether
five loudnesses is signififair to conclude
cantly similar that it seems
argue
different from six pitches. Probably
that the number
available
of responses
the difference is in the rightdirection,
had nothing to do with
to the observer
and
absolute
judgments of pitch the channel capacity of 3.25 bits.
than absolute
The Hake-Garner
accurate
are
slightlymore
experiment has been
loudness.
The
though
Aland Klemmer.
judgments of
important repeatedby Coonan
point,however, is that the two answers
they have not yet published
order of magnitude.
of the same
their results,they have given me
are
mission
perThe
nel
chanto say that they obtained
experiment has also been done
for taste intensities. In Fig. 3 are the
capacitiesranging from 3.2 bits for
results obtained
by Beebe-Center, Rogers,
and
O'Connell
(1) for absolute
of salt
judgments of the concentration
concentrations
solutions. The
ranged
from
cc.

0.3

to

34.7
in

tap water

They

used

1.9

bits,which

concentrations.
seem

per

100

equal subjectivesteps.

3, 5, 9, and
The

NaCl

gm.

17

different

channel
is about
Thus

centrations.
con-

capacity
four

taste

little less distinctive

is

distinct

intensities
than

tory
audi-

stimuli,but again the order of


magnitude is not far off.
On
the other hand, the channel capacity
for judgments of visual position
to be significantly
seems
larger. Hake

IMP'J"^

Fig.
on

the

INFORMATION

Garner
Hake
and
(8)
ments
judgcapacity for absolute
of the position of a pointer in a linear
4.

interval.

Data

channel

from

140

READINGS

short

IN

PSYCHOLOGY

MATHEMATICAL

Curvature
pointer position for the long exposure.
was
harder
When
to
the
longer exposures.
apparently
judge.
values
than
are
slightlyhigher
length of the arc was constant, the result
and
must
clude
conGarner's, so we
at the short exposure
duration was

very

These
Hake

bits for

that there
distinct

of the

exposures

3.9

to

between

are

and

10

linear

positionsalong
is the largest channel
a

This

that

has

unidimensional
At

been

measured

absolute

unidimensional

stimulus

for any

2.2

bits, but

chord

that

the

length

constant, the result

was

bits.

1.6

when

This

last value

has

anyone

of

the

only

was

is the lowest

measured

to

date.

should

these four

judgments

stimuli

in the

appeared
However,

pacity
ca-

variable.

the present time


on

IS

val.
inter-

experiments
simple,

of

all that have

are

add, however, that these values


low because
to be slightlytoo
data from all subjectswere
pooled
apt

are

the

before

the

transmitted

information

was

psychologicaljournals. computed.

great deal of work

variables

has

other

on

yet appeared

not

Now

let

where

we
are.
First,
to be a
capacity does seem
notion
for describinghuman
servers.
obthe
channel
Second,
capacities
us

see

the channel

in the

journals. For example, Eriksen


valid
that
the
(6) have found
channel
capacity for judging the sizes measured
for these unidimensional
ables
variof squares
about
five
is 2.1 bits, or
1.6 bits for curvature
from
range
perimental
categories,under a wide range of exinterval.
to 3.9 bits for positionsin an
conditions.
In a separate
Although there is no question that the
(5) found 2.8 bits
experiment Eriksen
the variables are real
differences among
l.Z bits
for size,3.1 bits for hue, and
and
the more
meaningful,
impressive
for brightness. Geldard
has measured
fact to me
is their considerable
larity.
simithe channel
capacity for the skin by
If I take the best estimates
I
placing vibrators on the chest region.
channel
all
of
the
for
can
capacities
get
A good observer can
identifyabout four
the stimulus variables I have mentioned,
five durations, and
about
intensities,
and

Hake

about
One

the

locations.

seven

of the most

in this

active groups

the Air Force

has been

area
Operational
Applications Laboratory. Pollack has
been
kind
with
enough to furnish me

the

results

of

their

several aspects of visual


made

displays. They

for

measurements

for

measurements

and

area

for

the

curvature, length,and direction of


lines. In one
set of experiments they
used
"

very

%o

second
2.6

bits

v/ith

bits with

the

length of

bits with

For

or

the
the

the

long exposure.
they got about

of

bits for the short

exposure

5-

and

exposure

line

angle

peated
re-

they got

area

short

the short

bits with

with

measurements

exposure.

2.7

3.0

they

"

the

is

standard

deviation

and

exposure

and

3.3

the

from

total

range

to

is

to

wide

that
be

the

There
built

be

to

seems

into

either

some

limitation

by
design of our nervous
systems, a
limit that keeps our
channel
capacities
in this general range.
On
the basis of
us

by learning or

the

the

present

evidence

2.6

say

that

possess

small

we

one

it

safe

seems

finite and

to

rather

mensional
capacity for making such unidiand
that
this
pacity
cajudgments

does not

2.8

bits

standard

15 categories. Considering
variety of different variables
have
been studied,I find this to
remarkably narrow
range.

from

tion,
Direc-

long exposure.
inclination,
gave

includes

categories,and

10

For

about

the

only 0.6 bit. In terms of


this mean
distinguishablealternatives,
one
corresponds to about 6.5 categories,

of the stimulus

short exposure
second
and
then

bits and

is 2.6

mean

deviation

vary

simple sensory

great deal from

attribute

to another.

GEORGE

Absolute

Judgments

A.

dimensional
Multi-

of

suits.

have

may

been

noticed

that

increased
that

means

You

the channel

Now

have

to

Stimuli

141

MILLER

of

one

any

careful

people

have

capacity seems
bits, which
rately
identify accupositionsin the
4.6

to
can

24

that this magical


say
square.
appliesto one-dimensional
The
positionof a dot in a square is
judgments. Everyday experienceteaches
clearly a two-dimensional
proposition.
that we
can
us
identifyaccuratelyany
Both
its horizontal and its vertical position
of several hundred
one
faces,any one
be identified. Thus
must
it seems
of several thousand
words, any one of
natural to compare
the 4.6-bit capacity
several thousand
The story
etc.
objects,
for a square
with the 3.25-bit capacity
certainlywould not be complete if we
for the positionof a point in an
val.
interhave
stopped at this point. We must
The
in
the
point
requires
square
some
understanding of why the onetwo judgments of the interval type.
If
dimensional
variables we
judge in the
have a capacity of 3.25 bits for estimating
we
laboratory give results so far out of
intervals and we
do this twice,
line with what
do constantlyin our
we
should
6.5
bits as our
we
get
capacity
behavior outside the laboratory. A possible
for locatingpoints in a square.
Adding
of
explanationlies in the number
the second independent dimension
gives
independently variable attributes of the
increase from 3.25 to 4.6, but it
an
us
stimuli that are being judged. Objects,
falls short of the perfect addition
that
faces,words, and the like differ from
would
give 6.5 bits.
another
in many
whereas
the
one
ways,
Another
example is providedby Beebestimuli
have
considered
thus
we
simple
When
Center, Rogers, and O'Connell.
far differ from one
another
in only one
they asked people to identifyboth the
number

to

seven

respect.

saltiness and

Fortunately,there are a few data on


what
absolute
make
we
happens when
judgments of stimuli that differ from
another

one

look
Frick

in

several

in
the

first at

Let

ways.

results

Klemmer

us

and

lute
(13) have reported for the absojudgment of the positionof a dot
In Fig. 5 we
their resee
square.

the

containingvarious concentrations of salt


and sucrose,
nel
they found that the chanpacity
capacitywas 2.3 bits. Since the cafor salt alone was
1.9, we might
expect about
of

3.8

bits if the two

the

the second

dimension

capacity but
conceivablymight.
A third example

the

IN A

.03 SEC.
3

4
INPUT

Fig. S.
on

the

Data

from

channel
of the

SQUARE

GRID
EXPOSURE

second

INFORMi^TION

Klemmer

and

dimension

capacity but
Frick

not

adds
as

much

little
as

it

lack
provided by Pollisteners to judge
both the loudness and the pitch of pure
Since pitch gives 2.5 bits and
tones.
loudness
gives 2.3 bits,we might hope
4.8 bits for pitch and
to get as much
as
loudness together. Pollack obtained
3.1
bits, which
again indicates that the

(18), who

NO

aspects

compound stimuli were


judged
with
As
tions,
independently.
spatial locato

POINTS

of solutions

sweetness

(13)

capacity for absolute


ments
judgposition of a dot in a square.

fourth

the work
on

not

is

asked

much

example
of Halsey

confusions

the channel

augments
so

among

as

it

might.

be drawn

can

and

from

Chapanis (9)

colors of

equal

142

READINGS

luminance.

IN

Although they

MATHEMATICAL

did not

their results in informational

they estimate
15

that there

are

about

3.6 bits.

in both

hue

1 1 to

our

terms,

dimensional
with

Since these colors varied

and

saturation,it is probably
regard this as a twojudgment. If we compare
to

correct

this

lyze
ana-

terms,

about

identifiable colors,or, in

PSYCHOLOGY

Eriksen's

bits

3.1

for

hue

(which
a questionablecomparison to
draw), we again have something less
than
second
perfect addition when
a
is

dimension

NUMBER

is added.

It is still a

Fig. 6.

The

OF

VARIABLE

general form

ASPECTS

of the relation

tween
be-

channel capacity and the number


of independentl
however, from
variable attributes of the stimuli.
these two-dimensional
examples to the
multidimensional
stimuli
provided by
It is interesting
decreasingrate.
to
faces,words, etc. To fill this gap we
that
the
note
channel
is
creased
inhave only one
capacity
experiment, an auditory
when
the several variables
even
study done by Pollack and Picks (19).
not independent. Eriksen
are
(5) reports
They managed to get six different acoustic
when
and
that,
size,brightness,
variables that they could change:
hue all vary togetherin perfectcorrelation,
of interrupfrequency, intensity,
rate
tion,
the transmitted
information
is 4.1
on-time
total
fraction,
duration,
bits
with
of
as
compared
and spatiallocation. Each
an
of these
average
one

long

six variables could


five different

5^, or

were

they
a

could

way,

assume

of

one

any

values,so altogetherthere
15,625 different tones that

present.

The

the transmitted

Under

one

was

ent
corresponds to about 150 differcategoriesthat could be absolutely
Now

error.

up

we

are

into the range

time.

that

the

was

function

increased

input
input

without

in

the
creasing
in-

tion;
informa-

increase in channel

an

capacity of about
the dotted

are

By confounding

of

amount

the result

bits,

which

beginningto get

the

of these

7.2

at

one

these attributes

three attributes,
Eriksen

these conditions

information

identified without

2.7 bits when

varied

of
dimensionality

listeners made

separate rating for each

six dimensions.

about

the

that

amount

Fig. 6 would

lead

to expect.

us

The
add

point

increase

seems

to

be

that, as

we

variables to the display,we

more

the

total

crease
capacity,but we defor
particular
any
expect.
accuracy
Suppose that we
plot these data, variable. In other words, we can make
relativelycrude judgments of several
fragmentaryas they are, and make
a
about
how
the channel
things
simultaneously.
guess
capacity
with
We
the
of
the
of
changes
might argue that in the course
dimensionality
stimuli.
The
result is given in Fig. 6.
evolution
those organisms were
most
In a moment
of considerable
successful that were
daring I
responsiveto the
sketched
the dotted line to indicate
widest
of stimulus
energies in
range
the
trend
that
the data seemed
their environment.
In order to survive
roughly
to be taking.
in a constantlyfluctuating
world, it was
Clearly,the addition of independently better to have a little information about

ordinary experience would

lead

us

to

the

variable

attributes

the channel

to

the

stimulus

capacity,but

creases
ina lot of
at

thingsthan
about

to have

small

lot of information

segment

of the

144

READINGS

IN

SUBITIZING
out
generalarea witht
he
periments
mentioning,however briefly, exconducted
at Mount
Holyoke
ber
College on the discrimination of numIn
(12).
experimentsby Kaufman,
random
Lord, Reese, and Volkmann
leave this

I cannot

patterns of dots
for
to

of

second.

than

more

200

the pattern.

report how
The

flashed

were

The

on

first point to note

The

of dots

was

in
to

was

can

be
the
are

This

is

on

make

errors.

small

bers
num-

the

dots that it was


performance with more
Below
the
seven
given a specialname.
said
to
subjects were
subitize;above
said to estimate.
This
seven
they were
will
what
we
once
as
is, you
recognize,
called
"the span of attention."
optimistically

one

number

magical

Here

me.

of

numerousness

seven

as

yet when

of the ways
the
in which
has persecuted
seven
have

we

closelyrelated

two

experiments, both of which


of the number
significance
limit on our capacities.And
examine

we

the matter

to be a
closely,there seems
that
it
is
nothing
suspicion
a

discontinuityat seven
is, of
suggestive. Is this the same

The

Span

Let

course,

basic process
that limits our
sional judgments to about
The

unidimengories?
cate-

seven

is tempting,
generalization
in my
opinion. The data

than

more

the

summarize

me

this way.

Memory

Immediate

of

There

is

limit to the accuracy

more

reasonable

coincidence.

unidimensional

would

with

which

absolute
that

maintain

call

definite
we

limit

judgment,
is

can

of

variable.

this

I
the

and

for unidimensional

this span

in

magnitude

stimulus
to

propose

of

span

situation

clear and

identify absolutely the


This

are

the

point to
patterns

six dots

or

of

the subject
density. When
and
area
subitize,
density may not
the significant
variables,but when
subjectmust estimate perhaps they
parison
significant.In any event, the comis not
so
simple as it might
at first thought.

were.

different from

so

and

appear

is that

containingup
subjectssimply did not
performance on these

dimensions

kinds

dots there

many

to five

the

from

subject'stask

two
area

seem

screen

Anywhere

dots could

PSYCHOLOGY

MATHEMATICAL

ments
judg-

usually somewhere

We
are
neighborhood of seven.
this
the
of
at
completely
mercy
on
lyzed
analimited span, however, because we have
in informational
terms; but on
a
variety of techniques for getting
the basis of the published data I would
around
it and increasingthe accuracy
that
transmitted
the
subjects
guess
portant
of
judgments. The three most imthan
four bits of information our
something more
of these devices
are
(a) to
about
the number
of dots.
absolute
relative
rather
than
make
ments;
judgUsing the same
arguments as before,we
that
is not possible,(b)
would conclude that there are
about 20
or, if
but not sound
number

estimates

have

distinguishable
categoriesof nuThis is considerablymore
merousness.
or

30

information

qet from

is,as

than

we

of

matter

two-dimensional

is not
in the

Prick's

expect

random

dot

entirelyclear,these
same

for

of dots

range

their
in

as

two-dimensional
square.

Perhaps

increase

along
(c)

to

that

we

terns
patand
play
disthe

The
one

the

which

It

results

Klemmer

to

to

display.
fact,very much like a
display. Although the

dimensionalityof the
are

would

unidimensional

in the

not

not been

stimuli

the task

arrange

of

number

the

dimensions

can

differ;or

in such

way

solute
of several absequence
in
a
row.
judgments
make

study

of

of the oldest

psychology,and

relative judgments is
topicsin experimental
I will not

now.

The

second

pause

to

view
re-

creasing
device, inhave just
the dimensionality,
we
It seems
that by adding
considered.
it

GEORGE

dimensions

more

requiringcrude, a lot of
tributethis span
judgments on each at-

extend

can

judgment
Judging

150.

the

145

MILLER

and

binary, yes-no
we

A.

of absolute

the span

from

at

to

seven

from

least

everyday
probably in

is

of

span

havior,
be-

our

limit

I have

the

there

different kinds
is about

is

encompass
What
is

but I must

This

objectiveevidence
This

of

that there is

at once
to

support

picion.
sus-

ing
question sadly needexperimentalexploration.
Concerning the third device,the use
successive judgments, I have quite a

bit to say

because

this device

introduces

the handmaiden

as

memory

is

this

no

And,

of discrimination.

since mnemonic

items

in length.
just shown you that there is a
absolute judgment that can
tinguish
disabout seven
categoriesand that

thousands,if indeed there is a limit. In


opinion,we cannot
ing
my
go on compounddimensions
indefinitely.I suspect
that there is also a span of perceptual
is
dimensionality and that this span
in the neighborhood of ten,
somewhere
add

of test materials

seven

span

about

of attention

that

six

objectsat

natural

more

than

all three of these spans


of

And

that

is

I shall be at

will

glance.

think

to

that

different

are

pects
as-

single underlying process?


fundamental
a
mistake, as
some
pains to demonstrate.

mistake

is

of

one

the

malicious

persecutions that

the magical number


subjected me to.
mistake went
something like this.

has

seven

My

have

We

is the

that

seen

in the

span

of

amount

observer

the

invariant

of absolute
information

that

There

transmit.

can

ture
fea-

judgment
is

the
real

the absolute
operationalsimilaritybetween
least
at
and
as
the
are
as complex
are
perceptual
judgment experiment
immediate
mediate
we
can
anticipatethat their
processes,
experiment. If immemory
interactions
will not
be easily disentangled.
is like absolute judgment,
memory
processes

then

to

tending
Suppose that we start by simply exslightlythe experimental procedure
that we
have
been
using. Up
this point we
have presented a single

stimulus
it

and

asked

the observer

to withhold

have

given

him

of the sequence
stimuli he then makes
his response.
still have

the

situation

that

of
We

of

is

transmitted

tion.
informa-

But
have
now
we
passed from
experiment on absolute judgment to
what
is traditionally
called an
ment
experian

on

Before

immediate
we

look

at

warning

to

help

associations

Everybody
span

knows

of immediate

data

any

give you

you

that

avoid
can

that
memory

that

be

of

ous
obvi-

some

there

this

on

word

the span

confusing.
is

and

finite

that for

observer

an

and

the span

the items contain

them,

for

about

amount

of

constant

at

able

remember

can

information

at

only
random.

amount

item

in the

test

seven

formatio
inare

total

should

two

In

how

be

three

or

this way
the span
vary

as

of information
materials.
of memory

are

i.2"

remain

to

we

should

memory

measurements

For

bits of

23

is

bits, then

of the

literature

when

worth

English words
apiece. If the

function

the

long

recall about

generateda theory about

The

be

digitsare

10 bits

of immediate

the individual

lot of information

total of

worth

chosen

of

span

littleinformation.

2?)

If the

constant, then

should

Isolated

to

retain.
in the

be short when

contain

example, decimal
bits apiece. We

per

of immediate
of information

can

is

memory

should
items

variant
that the in-

in the span

of information

amount

words

memory.

topic I feel I must

memory

of

follow

is also the amount

server
ob-

input-output
required for the

sort

same

of

measurement

tend
ex-

his response
until we
several stimuli in succession.

the end

At

feature

immediate

to name

immediately thereafter. We can


this procedure by requiringthe

it should

span

suggestive on

in
this

146

READINGS

question,but
necessary

see.

Hayes

do

to

it

so

experiment

tried it out

(10)

different kinds

the

PSYCHOLOGY

MATHEMATICAL

And

definitive.

not

was

IN

with

of test materials:

to

five

binary

digits,decimal

phabet,
digits,letters of the aldecimal
and
digits,
plus
The
1,000 monosyllabic words.
letters

with

listswere
item

per

much

read

aloud

second

and

at

the

the rate

they needed
A
procedure
(20) was used

time

as

responses.

Woodworth

of

one

subjectshad
give

to

described
to

as

their

by
the

score

INFORMATION

PER

ITEM

BITS

IN

responses.

The

results

shown

are

the

filled

dotted

line

by

circles in Fig. 7.

Here

indicates

should
span
of information

been

what

the

if the amount

span

The

constant.

were

represent the data.

experiment using
different

the

in the

solid

from

Data

8.

of

Pollack

information

presentation plotted
of

amount

as

information

(16)

retained

on

after
of

the

the

test

function

item

per

in

the
one

materials.

curves

Hayes repeated the


test

sizes but

have

Fig.
amount

vocabularies

of

There

is

with
wrong
Pollack

nothing
experiment, because

Hayes's
(16) repeated
elaboratelyand

all

containing only
it much
more
English monosyllables (open circles in
result.
lack
Polgot essentiallythe same
Fig. 7). This more
homogeneous test
took pains to measure
the amount
material did not change the picturesignificantly.
of information
and did not
transmitted
With
items
the
binary
span
rely on the traditional procedure for
is about nine and, although it drops to
His results are
scoring the responses.
about
five with
monosyllabic English plotted in Fig. 8. Here it is clear that
words, the difference is far less than
the amount
of information
transmitted
the hypothesis of constant
information
is not a constant, but increases almost
would require.
of information
linearlyas the amount
per
INARY
BINARY
DIGITS
ilGITS

DECIMAL

DIGITS

f-^
[-LETTERS

rUETTERS
rL

ft DIGITS

in the

so

In

of

40
\
" CONSTANT
V INFORMATION

increased.

input is

the outcome

is

perfectlyclear.

coincidence

the

spite
magical number
places, the span

50

a.

item

And

iggo
WORDS

seven

of

the

that

appears
absolute

in both

judgment

and

the span
of immediate
memory
quite different kinds of limitations

30
\

are

\
\
\

imposed

on

our

Absolute

information.

are

that

abilityto process
ited
judgment is lim-

20

by

the amount
memory

10

of information.
is limited

INFORMATION

test

7.

Data

immediate

of the

ber
num-

amount

materials.

from
memory

PER

ITEM

of information

per

hits

of

information

BITS

IN

Hayes (10)
plotted as

of distinguishing

fallen into the custom


between

of

the

tinction
In order to capture this disin somewhat
picturesqueterms,

of items.

I have

Fig.

by

mediate
Im-

and

chunks

say

that the number

the

on
a

span
function

item

in the

and

Then

of information.

can

tion
of bits of informa-

is constant

for absolute

the number

of chunks

judgment

of informa-

GEORGE

tion is constant
The

span

of

bits

has

also
we

constitutes

drawn

and
fact

definite about

very

chunk

bit

terms

what
For

of information.

memory

of five words

span

obtained

each

when

at random

from

word

span

of

had

about

IS

been

called

phonemes,
phonemes

three

recallingfive

in it.

the

here

subjectswere
phonemes,

words,
logicaldistinction

with

process

familiar

into

input
grouping
or
chunks, and a great deal
has

the

into
gone
familiar units.

of

units

learning

formation

into

sequence

Since

the memory

of

units
is

span

chunks, we

can

fixed

ber
num-

the

ber
num-

increase

that

it

than

just beginning
telegraphiccode hears each
as

to

to

man

separate chunk.

organizethese

then

he

deal

can

chunks.

sounds
with

dit and

Soon

he

many

The

into

is able

letters

operator

bits per

the

of communication

chunks

code

with

many

with

more

ways

that

and

group,
rather

then

than

Since
a

the

that

bits per
the input

contains

fewer

There
bits per chunk.
do
this
to
recoding,but

probably the simplest is


input events, apply a new

to

the

group
to

name

the

remember

the

name

new

originalinput events.

I am

convinced

very

general and

psychology,I

that

this process

important

to tell you

want

one

about

experiment that should


I am
ing
talkwhat
explicit
perfectly
This
about.
ducted
conexperiment was
and
was
ported
reby Sidney Smith
chologic
by him before the Eastern Psy-

demonstration

make

but
a

with

we

the observed

recalled

suspect
of

be

and

in these
that

at once

used

immediate

digits. In Table

as

fact that people

eightdecimal digits,
nine
only
binary digits. Since there
of
largediscrepancy in the amount

could
span

in 1954.

repeat back

can

is

ory,
the-

code

few

operator recodes

another

chunks

dah

into letters and


the

increase

to

this process
would
be
The
input is given in

Begin

radio-

the

ing.
called recod-

information

learn

am

jargon

chunk.

tains
con-

before.

the

that

correspondingly.
proposing to use, the

Association

simply by building larger and


larger chunks, each chunk
containing
information

contains

chunks.

or

of bits of information

more

the

In

precisely,
speak more
must
we
therefore,
recognizethe importance
of grouping or organizing the input

ized
organ-

chunk.

is

to

are

increases

operator learns

Recoding
order

dahs

that

message

remember

for

In

dits and

In the terms

these

of

the

of

amount
can

mediately
imare

are

of

the

is not

dealing
organizing or

We

apparent.

tively,
Intui-

15

not

as

propriately
ap-

word

overlap

by learninginto patterns and


these
larger chunks
emerge

memory

each

since

it is clear that the

but

and

rates

the

during
learningprocess.
simply pointing to the obvious

am

fact that

set of 1000

English monosyllablesmight just as


have

different

at

other

date.

to

highlightthe

to

serves

example, the
that Hayes
was

the

of

not

are

examined

been

contrast

chunk
that

per

that

range
The

each

seems

memory

147

MILLER

achieved

memory.

independent of the number


the
chunk, at least over

be almost

to

for immediate

of immediate

A.

renaming

cases,

recoding procedure

to

memory
1

two

increase
for

the

binary

method

for ing
groupis illustrated.
Along
a

of 18 binary digits,
able to
subject was
and
he
hear
recall
after
larger chunks,
a
single presentation. In
begins to
whole phrases. I do not mean
that each
line these same
the next
binary digits
that plaor
step is a discrete process,
teaus are
grouped by pairs. Four possible
00 is renamed
in his learningcurve,
must
0, 01 is
pairs can occur:
appear
11 is
for surely the levels of organizationare
and
renamed
renamed
10
is
2,
1,

themselves

Then
as

the

letters

words,

which

organize

are

still

the top is
far

more

a sequence
than any

148

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

TABLE
Ways

Binary

4:1

10

10

Recoding

220

Chunks

from
four

remember,

111
7

Chunks

1010

0010

0111

0011

10

10100

span

is to

same

the

In

to

this is almost

from
into

and

7.

Now

of

sequence

this is well within


In

memory.

within

the

binary digits

of three.
of

There

tween
be-

name

recoded

binary digits
octal digits,and

binary digitsare grouped by fours and


by fives and are
given decimal-digit
from

names

It is
of

to

15

and

from

reasonablyobvious

to

11001

31.

110

25

recoding schemes increased their


case.
binary digits in every
span
But the increase was
not
as
large as we
had expected on
the basis of their span
for octal digits. Since the discrepancy
increased as the recodingratio increased,
for

reasoned

we

that

subjects had

had
the

Apparently
code

few

been

ing
recod-

sufficient.

translation

from

one

almost

be

must

the

minutes

learning the
not

the other

to

matic
auto-

subjectwill lose part of the


while he is trying to remember

the

or

next

the

spent

schemes

18

the span
of ate
immedithe last two lines the

10

The

three,so

have

we

of

sequence

quence
se-

In the next

memory.
of

sequence

eightpossiblesequences
a
give each sequence
new

we

receded

just nine digitsto

now

is regrouped into chunks


are

base-

110

16

recede

we

say,

10

001

01001

20

arithmetic

of immediate

line the

11

are

and

00

100

arithmetic.
there

11

13032

That

3.

01

000

base-two

10

Recoding

renamed

00

101

Chunks

Digits

Recoding

Recoding
5

Binary

of

101000100111001110

Chunks

(Bits)

Digits

Sequences

Recoding

of

group

the translation
Since

the 4

of the last group.


5 : 1 ratios require

1 and

considerable

decided

imitate

study. Smith
Ebbinghaus and do

the

himself.

on

that this kind

With

he drilled himself

to

ment
experi-

Germanic
on

each

tience
pa-

ing
recod-

recoding increases the bits per chunk,


into
packages the binary sequence

obtained
the reand
sults
successively,
the data
shown
in Fig. 9.
Here
form that can
be retained within the
follow along rather nicely with the rea
sults
of immediate
So Smith
would
span
predict on the basis of
memory.
you
assembled
20
could rehis span
for octal digits. He
member
subjects and measured
their spans
for binary and octal digits.
the 2:1
12 octal digits. With
The spans
9 for binaries and
7 for
worth
were
were
recoding, these 12 chunks
octals.
Then
he gave
ing
each
recoding 24 binary digits. With the 3:1 recodscheme
36
to five of the subjects. They
worth
binary
digits.
were
they
studied
the recoding until they said
5 : 1 recodings,they
With
the 4 : 1 and
it for about
5 or
about 40 binary digits.
10
worth
they understood
were
and

"

minutes.

binary
use

he tested their span


for
digitsagain while they tried to

the

studied.

Then

recoding

schemes

they

had

It is

to watch

little dramatic

digitsin

binary
get
back
then repeat them
40

However,

if you

think

row

without

of this

son
per-

and
error.

merely

as

GEORGE

A.

149

MILLER

eyewitnessesis

50

but
PREDICTED

40

FROM

into

3:l

2:l

4:|

RECOCTING

5:i

RATIO

of immediate

span

SUBJECT

PRACTICED

HIGHLY

ONE

The

for

memory

of the
binary digits is plotted as a function
tion
The
predicted funcrecoding procedure used.
for
is obtained
by multiplying the span
octals by 2, 3 and
3.3 for recoding into base
4, base 8, and base 10, respectively.

mnemonic

memory

trick

span,

extending

for

will miss

you

the

the

more

important point that is implicit in


devices.
The
nearly all such mnemonic
that
is
is
an
point
extremely
recoding
for
powerful weapon
increasing the
of

amount

that

information

deal with.

In

form

one

we

can

another

or

recoding constantly

use

of

mony
testifollow

in

our

"

the

few

rich

chunks

mation.
infor-

in

suspect that

imagery is a
form of recoding,too, but images seem
and
much
harder to get at operationally
the more
to study experimentallythan
symbolic kinds of recoding.
zation
memoriIt seems
probable that even
I

9.

random

naturally from

30

Fig.

legalpsychology,

they
particularrecoding
that the witness used, and the particular
recoding he used depends upon his
whole life history. Our language is tremendously
useful for repackaging material

FOR
DIGITS

\\

in

distortions

not

are

SPAN

OCTAL

well known

the

studied

be

can

The

these

in

terms.

of

be simply

memorizing may
of chunks, or groups
of items that go together,until there
that we
few enough chunks
can
are
so
recall all the items.
The work by Bousiield and Cohen
(2) on the occurrence
is
of clusteringin the recall of words
in this respect.
especiallyinteresting
process

the formation

Summary
I have

we

that

daily

like

come

to

I wanted

to

end

of the

present,

make

to

now

the

so

data

would

summarizing

some

marks.
re-

behavior.
In

opinion the most


customary
my
of recoding that we do all the time

kind

is to translate into
there

is

idea that

try
words."

story

we

verbal

or

want

remember,

to

When

it "in

witness

we

to

our

recreate

we

by secondary

details

the

particularverbal
to have

and

verbalization.

the

that

we
our

Upon

member
re-

recall

elaboration

well-known

with

into several
into

have
one

on

demonstration
The

inaccuracy

deserves
than
the

that

people

do

of
at

to

By

ceive,
re-

ganizing
or-

ously
simultaneand

dimensions

cessively
suc-

chunks, we

least stretch)

bottleneck.

much

it has
kind

able

input

(or

pose
im-

amount

remember.

Second, the process


important one in
very
and

the

are

sequence

break

this informational

periment
ex-

of the process.
of the testimony of

to

manage

and

Walter

and

judgment

memory

on

we

stimulus

the

pen
hap-

by Carmichael, Hogan,
(3) on the influence that names
the recall of visual figuresis

that

process,

verbal

then

limitations

of information

event
a

of immediate

severe

own

recoding we

The

an

of absolute

span

the span

ally
usu-

consistent

seem

made.

or

some

make

remember, we
descriptionof the event
want

we

and

When

argument

an

rephrase

to

code.

First,the

of
seems

lifeblood of the

of

recoding is a
ogy
psycholattention
explicit

human
more

received.

In

ticular,
par-

linguistic
recoding
to

me

to

be

the

thought processes.
very
constant
a
Recoding procedures are
social psycholoto
concern
clinicians,

150

IN

READINGS

PSYCHOLOGY

MATHEMATICAL

and
for absolute
and anthropologists
jects
objudgment, the seven
gists,linguists,
less
is
in
the
of
and
the
recoding
attention,
yet, probably because
span
accessible to experimental
manipulation seven
digitsin the span of immediate
T mazes, the
than nonsense
For the present I propose
to
or
syllables
memory?
has
withhold
traditional experimentalpsychologist
there
is
judgment. Perhaps
little or
contributed
something deep and profound behind all
nothing to their
these sevens, something just callingout
Nevertheless,
experimental
analysis.
of
be
for us
it. But
I suspect
to discover
reused, methods
techniquescan
dicantsthat it is only a pernicious,
behavioral inPythagorean
coding can be specified,
And
be found.
I anticipate coincidence.
can
will find a very orderly set of
that we
REFERENCES
relations describingwhat
now
seems
an
uncharted

wilderness

of individual

ferences.1. Beebe-Center,
difJ. G., Rogers, M.
Transmission
O'CoNNELL, D. N.

Third, the concepts and measures


provided by the theory of information
provide a quantitativeway of gettingat
of these questions. The
some
theory
brating
caliwith
for
a
provides us
yardstick
stimulus

our

measuring

materials and
of

about

of
through the sense
Psychol., 1955, 39, 157-160.
2.

J.
frequencies-of-usage.

jects.
sub-

the

performance

the

tion
interests of communica-

our

I have
of

suppressed the

information

technical

tails
de-

useful

lead you to think they are


in research.
Informational

have

already proved

the study of discrimination

valuable
and

number

of

seven?

What

about

the

of

of the week?

5.

1954, No.
6.

point ratingscale,the

about
seven

the

seven-

categories

the

of

reproduction
J.

exp.

effects of

terminate
de-

Aufgaben.

Amer.
J. Psychol, 1932, 44, 163-174.
Multidimensional
lus
stimuEriksen, C. W.
and

nation.
of discrimi-

accuracy

Tech.

WADC

USAF,

Rep.,

54-165.

Eriksen, C. W., "


judgments
stimulus

range

stimulus

and

H.

Hake,
as

lute
Abso-

W.

of

function

the

and
response

exp. Psychol., 1955, 49,

the

number

of

categories. J.
323-332.

8.

sis
analyjudgments of loudness.
/. exp. Psychol., 1953, 46, 373-380.
fect
efThe
Hake, H. W., " Garner, W. R.

9.

reading accuracy.
J. exp. Psychol.,1951, 42, 358-366.
ChroHalsey, R. M., " Chapanis, A.

7.

Garner,

W.

of

of
discrete

the

What

the

on

indeterminate

differences

seven

world, the seven


seas,
the seven
ters
daughdeadly sins,the seven
of Atlas in the Pleiades,the seven
levels of hell,
the seven
of man,
ages
t
the seven
primary colors,he seven notes
of the musical scale,and the seven
days

wonders

P., " Walter,

Relative

W.
and

in

great deal in the


and it
memory;

H.

experimental study

D.

Chapman,

guage;
of lan-

they promise a
learning and
been proposed that they can
has even
be useful in the study of concept formation.
A lot of questionsthat seemed
fruitless twenty or thirtyyears ago may
be worth
another look.
In fact,I
now
feel that my
stop just
story here must
it begins to get reallyinteresting.
as
the magical
And
what
about
finally,
study

4.

not
cepts
con-

An

Psychol.,

gen.

of
visually perceived form.
Psychol., 1932, 15, 73-86.

have

will not

A.

effect of language

and

measurement

1955, 52, 83-95.


Carmichael, L., Hogan,
A.

tried to express
the ideas in more
familiar terms; I hope this paraphrase

The
A., " Cohen, B. H.
clusteringin the recall of
of different
arranged words
of

randomly

3.

In

/.

taste.

BousFiELD, W.
occurrence

for

"

saline solutions

and

sucrose

S.,

of information

An

R.

informational

absolute

presentingvarious
steps

on

maticity-confusioncontours
/.
viewing situation.
Amer., 1954, 44, 442-454.
10.

Hayes,

J. R. M.
vocabularies
size.

numbers

of

scale

Memory
as

In

in

plex
com-

Opt. Soc.

for
span
function
of

eral
sevcabulary
vo-

Quarterly Progress

Acoustics
Report, Cambridge, Mass.:
of
Institute
Laboratory, Massachusetts
Technology, Jan.-June, 1952.

ON

REMARKS
I. THE

METHOD

THE

SQUARES

EQUAL

STANDARD

DEVIATIONS

CORRELATIONS*

EQUAL

Mosteller

Frederick
HARVARD

Case

Thurstone's
stimuli

and

to

an

that

the

assumption

of

in method.
stimulus

is shown

positions

method

of

There

(1)

lead

to

The

stimuli

pairs of

continuum

least

sumes
as-

to

sensations.

estimate

squares

notions

fundamental

of

set

comparisons
corresponding

paired

of

the

scale.

sensation

paired comparisons
is

of

sensations

of

between

to

the

on

Introdtiction.

1.

method

the

deviations

correlations
be relaxed
of zero
can
assumption
with
between
correlations
no
change
pairs
equal
of paired comthe usual
parisons
approach to the method

Further
Case

UNIVERSITY

correlations

zero

It is shown

of

standard

equal

ASSUMING

SOLUTION

LEAST

AND

COMPARISONS:

PAIRED

OF

(4)
which

stimuli

these:

are

be

can

located

jective
sub-

on

scale,usually not having

(a sensation

stone's
Thur-

underlying

urable
meas-

physical characteristic)

Each

(2)
to

sensation

for

Stimuli
rise

to

(5)

It is

(6)

Our
for

*This
the

sensations

sensations

article

and

is to space

was

available
of

the

from

particular stimulus

individual, thus

an

The

stimulus.

reports which

stimuli

individual

pares
com-

is greater.

paired sensations

the

ing
giv-

to be

(the sensation

correlated.

means),

cept
ex-

transformation.

linear

gives rise

is normal.

each

for

possible for these

research
made

Department
This

of

sensation

task

individual

an

individual.

presented in pairs to

are
a

these

grant

presented to

population of individuals

(4)

in the

distribution

The

(3)

when

stimulus

performed
to

Air

in

the

Laboratory

Harvard

University by the
Force, Project RAND.

appeared in Psychometrika,

1951, 16, 3-9.

152

of
RAND

Social

Relations

Corporation

Reprinted with

under
under

permission.

There

example,

analysis" for

the

only

but

individuals

several

need

there

Case

stone's

correlations
discuss
V

data

rather

of

been

has

deviations
the

of

for

number

table.

and

ith and

objects

of

Xj

jth

distributed
mean

stimuli, then
for

of Xi

of Zi

Marginal
in

Thurstone's

will

does

empirical

stick

not

not

stimuli, Oi

or

lie

on

evoked

sensations
we

O2

"rMXi)

of Xi and

to

to be

seem

distributions

pij

of

the

Distributions
Case

of
of

the

the

there

These

stimuli

continuum

S
.

individual

an

by the

with

(i

(i

(i,j

assume

Xj to be jointly normally

="r2

Xj

sensation

in

and

Xi

assume

On.

"""

single

We

Data.

Si

marginal

Error-Free

Figure

Stimuli

this

population of individuals

the

correlation

The

V.

equal standard

of
We

approximation.

tvith

which

single

are

variance

The

first

Case

original proportions

assumption

The

shall

We

to fit

the

that

this

for

seems

the

reproducing

of

sense

Stimuli

give rise to sensations


If Xi

stimuli

the

ordering

the

zero.

are

Thurard
stand-

V.

Case

Ordering

2.

tion
sensa-

as

and

equal

sensations

correlations, because

zero

the

of

that

case

are

quite frequently and

employed

reasonable

assumption

essential

are

is

distributions

of

paired comparison

the

assume

we

fully is known

in this

assumed

has

method

well in the

shall

or

example,

for

deviations

most

pairs of stimulus

standard

Case

discussed

been

sensation

between

dividuals
in-

not?

or

has

of the

deviations

equal,

standard

the

Thurstone

V.

; or

the

times;

several

discussed"

be

to

intercorrelations

equal

which

case

"cases"

assume

we

distributions
The

times

several

comparisons

comparisons

all

are

all the

Shall

zero?

them

all

individuals,

different

in

used

be people.

not

assume

we

make

may

Furthermore,
shall

makes

who

individual

one

have

not

may

we

materials

basic

the

of

variations

numerous

are

153

MOSTELLER

FREDERICK

Xi's

"

"

"

w)

,2

,""

,n)

2
,

appear

"

"

"

as

n)

(1)
.

in Figure

1.

Sensations

Method

of

Produced

Paired

by the
Comparisons.

Separate

154

In

more

than

best

can

the

For

of the

(1)

Our
of the

time

we

Sa

"

to build

are

thing
any-

and

Oj

whether

reports

of the method
in the

the problem
data

exceeds

assume

we

know

the

true

conditions

the

that

and

Xj

data.

of nontallible

case

we

stimuli

ordering the

for

tion
propor-

above

given

fulfilled.
is to find the

sensations

1). Clearly

for

Xi exceeds

the

Xj

data

say

this

do

to

reported

pa

the

by them,

hope

cannot

stimuli

of the

spacing

produced

we

transformation,
times

and

Oi

the tenor

Xi

problem

mean

if

time

though 5i

even

scale.

compares

through

exactly

are

part

of nonlallible

case

the

of

Zi

"

allowed).

are
see

first work

we

happen
rank-order
/

(no ties

Xj
We

if

to

individual

An

Xi

has

this

fact

possibilitythat X2

the

figure indicates

The

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

Si

ure
Sn points in Fig-

"""

within

except
the

merely

are

(or the spacing

linear
of

percentages

-ldij-{Si-Sj)y
Pij

PiXi"Xj)=

V27la{dij)
where

da

loss in

Xi

Xj

"

generality

in

and

J,

2aHdij)

aHdtj)

assigning

ddiji2)

the

2a2(l

scale

p). There

"

factor

will be

so

2aHl-p)=l.
It is at this point that
Case

However,

his

derivations

(and
equal

easing
items

the

the

assume

We

can

continuum

same

and

another.
that

Past

items

stated

model
has

it is difficult to

equation

the

model

unable

was

noticed
With

not

were

that

this
the

(2)

of

uses

imagine

this

really

method

have

assumed

assumed

to take

them

the

cognizance

zero

plicitly
ex-

set

is

of

great

i.e.,nearly

.42
,

attitudinal

all correlated

zero

with

one

all benefited

from

the

fact

to be
items

.38
,

equal. But

this

Actually
.34

are

implicitly with

through

readily imagine
correlated

acterized
char-

correlations.

zero

correlations

carried

are

(not necessarily zero).

conditions.

on

only

who

Thurstone,

variances

equal

but
artificially),

correlations
of

having

as

(3)

depart slightly from

we

no

that

uncorrelated.
were

of the

It

was

uncorrelated, but
Guttman

statement.

only
the

(2)

independently.

scale

factor

chosen

in

equation

(3),

we

can

rewrite

155

MOSTELLER

FREDERICK

/""

(4)

-^y'dy.

/,._.,

V2(4), given

From

of

table

normal

matrix

0
find

can

we

all other

Si

as

Thus

location

the

given

of

use

is

fallible data

with

problem

The

Si

the

arbitrarily assign

we

compute

can

we

if

Then

by

"{SiSj)

for

solve

can

we

pa

areas.

"S\=

parameter

any

pa

more

complicated.

fallible data,

have

p'a which

have

we

equation

to

Analogous

Scaling with

Comparison

Paired

3.

(4)

where

the D'i, are

normal

deviate

further

notice

p'ii

"z:^

estimates

of

the

Dij

were

corresponding
that

not

hold

We

conceive

for

Djh

the D'a

Da
to

Si

=^

we

pa

(5)

e-'y'dy.

We

Si"Sj

p'a

to get the

not

be

look

merely

of D'ij

matrix

consistent

in the

the

up

We
.

that

sense

Sj

"

as

(S'i

[D'ij
"

Sk^^ Dik

"

follows:

of the Si's called

Sj

S'i,such

the

D'ij

to construct

that

is to be

S'j)y

"

from

minimum.

(6)

i,i

It will

data.

One

help
can

to

indicate

set up

the

another

Si

Totals
Means

S-i

On

2'Si-nSx
5

"

S,

OF

S2

On

25i
S

of

form

solution

nonfallible

for

Sj matrix:

"

MATRIX

the problem

set of estimates

When
the true

i.e.,
Dij

does

estimates

of

/""

need

the D'a

Data.

have

we

are

Fallible

"

WS2
S2

Si

"

Sn

25i
S

S3

"

"

Si

w^a
S3

Sn

lSi
S

(Sn

"

"

nS"
Sn

156

READINGS

by setting Si

Now

(S

Si)

"

the

S'i

(S

"

(5

S^)

"

will

We

on.

so

to minimize

respect

(S

"

this

use

the

(S'i

"

with

Since

S'i

"

of squares

sum

D'^

S'i

S2), Ss

"

plan shortly

for

respect to S'i

need

above

from

for

S'i

only

the

main

which

i "

rivative
partial de-

the

and

"D'ji

we

take

we

"

S'j) matrix, i.e.,terms

"

(6)

expression

S'i

to

DH

and

S'i)

"

with

D'ij

S3), and

"

wish

we

with

(S'j

get S2

we

If

"

PSYCHOLOGY

MATHEMATICAL

IN

S'j

"

concern

selves
our-

diagonal
j

^=

in the

ing
Differentiat-

get:

we

9(2/2)
dS'i

S'j + S'i)-^

(D'ji

S'i

(D'ij
-

S'j)

(7)

j=i+l

j=l

{i=l,2,---,n).
this

Setting

partial

+S'i

S'.

derivative

+S'i-i"

"""

equal

to

(n"l)S'i

zero

have

we

S'i+i +

+S'n

"""

(8)
i-l

-Lf

"

j-l

but

D'ij

D'ji

"

and

D'a

0 ; this makes

be

can

have

and
to

is to

not

of the

be

suggested
in the

by

the

matrix

(8)

choose

from

S'i

of Si

S'i

"

of

we

the

left side

the

have

There

are

This
.

solution
of

(9)

means

S'

ishes.
van-

scale

our

various

by setting

try the

left side

of

only chosen

example,
=

(9)

ways

we

(10) which

(9) to the

or

will
is

umn
total col-

^D'ji/n
j=i

we

for

Sj

of

to set S'i

Then

similarity

l,2,---,n).

parameter.

parameter,

distances

measure

because

location

We

coefficients

expected

assign this location

by setting 5"i

of

j=l

(i

^D'j

assigned

side

written

determinant

This

right

2 D'ji

i=i+l

^S'j-nS'i

The

the

D'ji

j=l

(8)

j=i+l

2 D'ji + D'a

Thus

{i=l,2,--,n)

-Lf ij

ji

"

^D'ji/n.
j=i

(10)

i^l

when

that

Notice

S'i

157

MOSTELLER

FREDERICK

0 and

that

because

which

because

happens

double

;'

and

term

every

Therefore, substituting

sum.

negative

its

of

left side

in the

(10)

this

in

appear

(9)

we

have

^D'j,/n-:ZD'ji/n

2i5';i-n

(11)

^D'r.,

;=l

is

linear

transformation

an

The
the

of

this

solution
of

sense

not

to

known

condition

be

This
the

p'ij tend

This

to

and

zero

S'i will be

the

After

table.

in

in

solution

the

seems

been

have
lines.

these

entirely satisfactory because


stimuli are
extreme
compared.

when

numbers

in the

order, the

proper

say,

of columns

preliminary arrangement

approximately

D'a table.

beyond,

all numbers

by excluding

customary

it may

closely along

of

is not

unity

met

the

solution

although

worked

he

that

squares

for

assumption

the

least squares

unsatisfactorily large

difficultyis usually
the

least

literature

solution

squares

introduces

from

since

(3),

that

show

to

is

this

the

in

mentioned

least

is

That

(6).

to Horst

is

and

comparisons

paired

to

any

course,

equally satisfactory.
to provide a background

to indicate

is unnecessary,

Of

solved.

are

is

presentation

paired comparisons,

correlations

zero

equations
solutions

of the

of

point

theory

the

identity, and

which

so

This
2.0
that

quantity

^{D'ij-D\,i,,)/k
is

where

computed

entries
such

in both

appear

discussion

method

which

therefore

take

reasonable
the

of

account

of the

It should

D'a

also

one

original y'a

the

because
In
.

to be

seem

be

the

remarked
we

other

The

differing

this

words,

want
a

more

ford's
Guil-

(see for example

This

to

for methods

computations

variabilities

unmercifully

really

between

paired comparisons).

results.

that

which

of i for

differences

Then

of

method

k values

separations

scale

give reasonable

to

seems

of

the

over

y+1

j and

the

as

(1)

is

summation
column

taken

are

means

the

of

the

p' a

and

extensive.

solution
check

is not
our

reasonable

entirely

results
solution

against
might

158

READINGS

such

be

one

by

p"i;

that

the

once

minimize,

and

MATHEMATICAL

IN

S'i

PSYCHOLOGY

computed

are

we

can

the

estimate

p'a

say,

'2ip'ij

or

p"ijy

"

perhaps

Such

thing

can

do

attempts
method

not
be

to

sin

(arc

VP

doubt

no

differ

to

seem

be

sin

arc

"

ii

but

done,

^/p"ii)^.

the

from

enough

results

the

of

results

author's

the

of

the

present

pursuing.

worth

REFERENCES

1.

Guilford,
1936,

2.

3.

Guttman,
Annals

Horst,

P.

stimulus

Thurstone,

P.

New

Book

McGraw-Hill

York:

math.

of
method

Stat.,
for

situations.
L.

L.

received

J.

educ.

Psychophysical

8/22/50

1946,

determining

paired

quantifying

for

approach

389.

Manuscript

Methods.

Psychometric

An

L.

order.

of
4.

J.

Co.,

227-8,

17,
the

Psychol.,
analysis.

and

comparisons

rank

144-163.
absolute

1932,
Amer.

affective

23,

values

of

series

418-440.

J.

Psychol.,

1927,

38,

368-

160

READINGS

interval Pyt is 0.5 and

time

interval.

one

in

response

in

responses

obvious

Equations (1) and

probabilityof

and

Assume,

however,

that

at

response

time

probabilityof
rate

tions
equa-

gettingno

into

measure

continuous

intervals

interval

in the

bility
proba-

distribution
is zero,

able.
specifiand

stimulus

the

but

is finite and

between

conditions

stimulus

some

between

particulartime

time

directlyconsidered
the

within

getting exactly

relation

the

with

any

given

the

to

is not

thus

probabilityof

transform

dealing

within

0.69/r. Equation

/ is

of responses

Po is e~''^.

case

to

us

are
a

response

response

the

consider

we

In this

we

usually refers

Latency

of

when

probabilityof

(time),the

median

the

or

numbers

example,

(2) permit

Since

measure.

rate

For

interval, T.

an

loge0.5

The
interval, T, is {rT)e~'^'^.

an

(2) is

(1) and

is

"rl

PSYCHOLOGY

probabilityof various

(2) gives the

specifiedtime

MATHEMATICAL

IN

previous development.

are

of the

determinant

one

responding,that is,that the rate has different values for different


This assumption is consistent
with the discussions
conditions.
by

stimulus
Skinner

and

assumption

others

who

the

of rate;
emphasized the measurement
be
an
presumably
elementary requirement for

would

have

any

measure.

the

Under

circumstances

latency,
of the

beginning

since

of the

the

observation

the first response.

Thus,

determinant

of

of rate
are

responses

of the
the

on

delay

distributed

distribution

value, say

the

but

on

the

in time,

length / between

of latencies

time

employed

interval

between

the

presented) and
conditions

previousassumption
a

in discussing

of the

statement

that

rate

are

of

the
sponding
re-

conditions

of the first response.

occurrence

the

be

may

period (when a stimulus was


the assumption that stimulus

specifiedstimulus
of

be

would

responding and

randomly

under

latter

assumption, t

also

the
This

of the

median

ment
implies a probabilitystateand
of
the
stimulus
presentation

tells

statement

not

us

relationshipbetween

latency, and

the

rate

of

only of
some

the

sentative
repre-

responding;

for

latency,
probabilityof a response
greater than the median
tmd, is 0.5; and from
rima
equation (1) we see that
loge 0.5 or that
the median
latency equals 0.69/r.
The
ditioning
precedingdevelopment does not imply any particulartheory of conbut may
be incorporatedinto a large class of theories.
For example,
is combined
with
that
if the foregoing discussion
states
a
theory
that rate of responding is proportionalto the number
of responses
that remain
in
of number
of responses
to be given in extinction, the measure
In
extinction is immediately related to our latency and probabilityterms.

example,

the

"

other

words, if
r

where

A^ is the

number

of responses

(3)

k(N-n),
in extinction,

is the

number

of

re-

CONRAD

already given, and ^


equation (1) and obtain

is

sponses
r

in

G.

constant,

P"t
This

equation

terms

N,

n,

and

number

the

distribution

distribution

through
for direct evidence
above
The
course

outlined
data
of

figure 1

in

were

this interval

are

this

as

taken

distributed

getting
is the rate

of

text.

The

in

The

data

bar-press

line

drawn

equation (1).

well

from
The

session

as

for evidence

relating to

between

obtained

measurements

data

the

rate

the

is whether

responses
the

of

responding
the responses
that

greater than
same

the
a

ditioning.
periodicrecon-

of

Equation (1) states

responding expressed in

during

represent the responses

of "three-minute"

period
question at issue

randomly.

interval

an

intervals

time

rats

observation
The

constant.

the

data

examine

to

SECONDS)

white

plot of

of

assumption

25

in seconds.

time

the

of randomness.

diu-inga 20-minute

approximately

is

data

the

20

inter-response

of randomness

Within

where

periodic reconditioning.^

singleanimal

of

described

as

the

consequences

in

FIGURE

experiment with

an

situation

IN

t is

t, where

than

from

15

predict

tion.
stages of extinc-

at various

from

to

strength and

constant

mainly

(TIME

of

percentage

greater

for

responses

10

the

latency, rate

be used

in time, it is of interest

The

equation (4) may

follows

for

(4)

relationshipsamong

in extinction

of responses

k{N"n)

substitute

may

existing among
relationships

the

to

in extinction,

present argument

are

for

intervals between

distribution

random

examined

of responses

of time

the

be

we

-k(N"n)t

"

In addition

/.

of responses

and

Since

then

may

161

MUELLER

units

as

the
/

/.

was

in

bility
probais

e^",

In the

162

session, 238

20-minute
and

the

that

responses

measiu'ed

the

data

intervals,theory specifies

uniquely. In this
greater than t (in seconds) is
of the

percentage

the various

time

the

figure 1

date

consistent

are

per

corded,
re-

Thus,

in

with

intervals

values

the

theory

found

be

may

tween
be-

on
specified

represents

assumption

the

that

occurred

from

deviations

example,

through

were

second.

randomly in time.
of the agreement
be representative
data of figure1 may
certain cases
of
theory under the conditions specified,

and

data

for

The

responses
the

Although

greater than

were

function.

theoretical

the

shows

figure1

solid line

The

abscissa.

the

of

ordinate

g-o.2oj. xhe

response

intervals

responses

interval

time

time

237

of time

intervals between

probabihtyof gettinga

the

case

the distribution

to

of time

the distribution

is 0.20

session

in this

direct reference

without

made,

were

responses

rate

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

in

be

may

where

cases

One

noted.
animals

tween
betematic
sys-

class of deviations,

show

marked

"holding"

seconds.
depressed and held down for many
in
is
the usual
not
a
"refractory"period^
Although the "holding" period
in
the
data
similar
it
affects
a
of the term,
During
obviously
sense
way.
is zero.
One
the "holding" period,the probabilityof response
occurrence
complicating feature in analyzing responses characterized by "holding" is
the bar is

where
behavior, i.e.,

variable

"holding" is of

the fact that

length.

The

data available

at

present

of this

problem, but the simplicity


the factor of
result from apparatus changes designed to eliminate
that may
from the additional
"holding" and the advantages that may accrue
response
do

warrant

not

extensive

an

be shown.

specification
may
An

shown

example of a distribution showing systematicdeviations from theory is


The
in figure 2.
computations and plot are similar to those in
ordinate

figure 1.

The

responses

greater

is correct

except

the

appHed

to

time
the

in the
and

when
the

of

as

solid

line

in the

reference

in

case

the

to

the function

is

tribution
dis-

appears

analysisleading to equation (1) and applied to


when
period
applied to all portions of the observation
be
additional
An
then
in
test
"holding."
may
spent

data

the

which

from

one

the rate

response

time

term,
minus

of

form

in the

figm-e 2

of the

beginning of

abscissa
specified

total

determined,

responding without
fitis obviously poor;

The

meastu-ement

meastuements

the end

was

The

values.

between

the

that

assume

1
figtu-e

Hne

intervals

asymmetric.

and

us

percentage of

of

rate

intervals.

of time

Let

of the

the

the

specifiedabscissa

the

constant

figure1, directlyfrom
sigmoid

represents

than

The

theoretical.

the

treatment

values.

obtained.

interval between

Figure 3

shows

plot of the percentage of


beginning of the next that
The

Now

we

the end
the

are

of

ested
interone

sponse
re-

results of such

intervals between
were

solid line through the data

equal to the ratio of the number


"holding" time, i.e.,to the number

r, is set

the

the next.^

the

and

time

was

greater than
is theoretical
of responses

to

of responses

CONRAD

G.

163

MUELLER

10
t

plot similar

to

15

(TIME

IN

FIGURE

figure 1 showing the.deviation


The

line drawn

through

(TIME

The

time

of

data

that

in

between

response.

described

figure2

"corrected

figiu-es1 and
the

The
in the

end
line
text.

2, except
of

one

drawn

data

IN

theory
is

in

plot of

cases

of

"holding"

equation (1).

SECONDS)
3

FIGURE

to

SECONDS)

from

the

20

for
that

holding."
the

The

measured

response

and

the

through

the

data

plot is similar
interval

beginning
is

is the

of the next

theoretical

one

havior.
be-

164

READINGS

time.

per unit of "available"


the

independently of

to the

relevant

Data

The

agreement

percentage

between

of time

and

Yamaguchi

In the

straightline

of intervals.

evaluate

case

are

the

Hull^ is shown

In the

of

is evaluated

in

ous.
numer-

reported

figure4,

where

plotted.

are

of the

latency data under


tion
independently of the distribu-

figure4

fitted to

not

data

specifiedabscissa values

curve.

possibleto

intervals.

slope of

the

2 the constant

present theory and

of latencies greater than

it is not

consideration

and

distribution

the

solid line is the theoretical

The

by

of the

figures1

present analysisof latency measures

by Felsinger, Gladstone,
the

in

As

shape

PSYCHOLOGY

MATHEMATICAL

IN

case

the constant

was

determined

plot of logeP"/ againstt.

0123456789
IN

(TiME

SECONDS)

FIGURE

The

of latencies

percentage

than

greater

may
tests

be made.
that

have
One

promising.
of time

intervals

tried

animals,

will be

and

the

curve

between

as

from

are

say, response

distributed

found

for

in

manner

In other

in figure 1 will

of the

of the formulation

tests
not

sampling

available.
data

and
the

of animals

at

For

the

theory

is

distribution

comparable

tion
stage in extincspecified
of
i?" and Rn+i^ for a largenumber
in figure1
similar to that shown
at

(thereforethe steepness
n.

further

concerns

number

expectationis that

The

constant,

tested

been

figure 1 of Felsinger,

Hull.*

between

agreement
has

responses

systematicallywith
such

the

predictionthat

the intervals between,

vary

data
and

be

been

stages in extinction.

that

The

gained at this time by


additional
of equation (1),but many
tests appropriatedata are
For some

little is to

consequences

t.

Yamaguchi

Gladstone,

Probably

of

drop

of the

curve)

will

words, the steepness of the drop of

depend

on

where

in extinction

the

inter-

CONRAD

vals

measured.

are

measured,

cases

extinction

fact

In

this

the

although

is not

G.

expectation

number

of the

features

of the

distributions
of the

many

values
one

may

changes

in another,

changes

say

using different

data

functions

only

that

Summary.

"

responses
obtain

we

directlya

(or of

any

of the

function
obtain

intervals

number

length

the

that

probability

the

these

where
between

various
between

terms

values, the
time

of

of

and

are

responding

and

to

In

for

of

addition

has

we

that

assume

in time,

occurrence

of

interval

responding

inter-responsetime
specified
out

be

fixed

yet

responses

to

to

the

consequences

and

about

the

the

occur,

latency

statements

for

to

bility
probaaverage

distribution

of

and, by extension, for the distribution

responses

also

we

sponse
reas

responding,it turns
various lengths may

of

be related

may

rate.

present formulation
between

to

measures

that, for any


of

ing.
test-

probabilityvalue,
be specified. (3) Finally,
may
theory specifying the relationship

number

in extinction
as

rate

some

rate

theory

distributed

of

assume

responding, or,
added

of

occurrence

we

If

specifiedtime

of the rate

of latencies

occurrence

aid to

(1)

probabilityof

If

Therefore,

multiplicityof
descriptivestatistics not

randomly

of

the

relationships
among

are

and

corresponds

median.

and

considered.

of the

latency

well

as

intervals

latency

been

in terms

mean,

actually an

some

probabilityof

rates

of responses

number

of

the

or

is

but

of

account

considerations

rate

mean

of representative

preceding equations,

arithmetic

of different

of the interval

condition, there

relation

the

comparable

use

varying lengths. (2)

specifiedfor

basis

of responses) within

stimulus
the

of the

geometric

of the

statement

of

arbitraryselection

made

Since

eliminate

possibleto

the

the

statement

it is

the

are

has

account

data.

experimental

On

"free-response"situation

of

stage

quency
possibilityof specifyingthe fre-

statistic,say

theoretical

present

with

difficult problem

of

one

of strength of conditioning has


in

the

by

out

each

at

of the

discussed

data.

arise from

may

form

is the

measures

statistics

to pose

ceases

in

the

treatment

associated

summarizing

state

borne

of measurements

that

account

of the

problems
in

be

to

seems

large.

Finally,it may be pointed out


for the
important consequences
one

165

MUELLER

of

measures.

'

Hull, C. L., Principles of Behavior, D.

Skinner,

B.

F., The

Behavior

Appleton-Century

of Organisms,

D.

Co., New

Appleton-Century

York,

1943.

Co., New

York,

1938.
^

slightly different

i.e.,that
getting
period

after

immediately
a

to

response

randomness

equation

is

zero.

is

we

we

there

response

If

if

results

is

that

assume

instantaneous,

the

that

assume

"refractory"

the

transition

probability of getting

t is

P"t

period during which

-r{t-to)

the

from
an

the

interval

period exists,
probability of
"refractory"
greater

than

166

where

treated
*

as

The

"refractory"

the

is

to

gradual

of

Laboratories
is

This

for

useful
*

37,

merely

is

intervals

at

first

measure

make

with

this

in

analyses
of

the

transition

is

Psychological

is

next

that

that

show

may

the

indicate

the

the

independent

not

is

approximation

distribution
the

of

expression

an

and

Hull

fact

the

does

that

0.5

at

the

have

not

The

point

reasons.

intervals

step

of

use

normal.

the

two

the

starting

zero,

The

more

of

deviation

with

than

greater

appropriate.

Psychol.,

Exptl.
for

distribution

were

more

[/.

Hull

formulation

our

frequency

be

and

of

test

latency

for
is

Yamaguchi

in

exponential

Yamaguchi

the

notion

the
After

has

from

the

responses

short

zero

as

of
in

distribution

step
lower

method

second

maximum

marizing
sum-

figure

4.

reported

frequency

by
the

at

many

of
under

for

of

stimulus

and
same"

light

[J.
is

would

the
stimulus

first

of

conditions.

stays

more

and

on

until

the

and

long

to

subsequent

sure
in-

would
the

notion
used

procedure
the

sponse
re-

one

procedures

closely

of

measurement

cedure
pro-

(1948)]

96-123

sponse,
re-

theory

present

sufficiently

The

responding.

response

time

the

experimental

experimental
parallel

the

permitting

26,
and

of

the

tinuous
con-

extend

to

occurrence

of

an

Psychol.,

presented

Such

of

the

the

possible

is

test

from

period

fixed

rate

before

obtained

Prick

and

end

it

equivocal

less

(Frick).

stimuli

advantage

"the

cit.).

some

responses

determine

the

which
A

latencies

light,

Although

stimuli
than

role

important

more

bar.

the

duration

(op.

on

transient

additional

onset

stays

or

of

of

discriminative

transient

many

play

may

of

no

say,

the

required.

Skinner

of,

period

bar

presence

are

by

used

sort

the

that

distribution

conditions
the

of

stimuli

to

unspecified
stimulus

of

assumed

the

with

occurrence

minimize

be'

may

exposure

(Skinner)

occurs

the

it

assumptions

expected

be

others.

between

an

would

point

associated

present

of

between

the

length.

Kaplan

beginning

optimal

shortest

account

may

additional

Frick

variable

Michael

procedure

our

reported

the

an

place,

second

ones

that

if

complex

interval.

the

may

of

the

Gladstone,

are

If

of

associated

the

data

Gladstone,

step

In

response

provide

not

lowest

data

deviation

Felsinger,

and

results

Felsinger,

zero.

easily
the

The

one

The

may

the

the

at

could

limit

that

begin

which

more

Subsequent

approximation.

of

by

(1947)

first

Mr.

by

is

has

present.

214-228

The

first
end

experiment

The

period

recorded

were

period.

the

"refractory"

the

University.

the

"holding"

the

if

or

here

Columbia

between

interval
of

one

formulation

The

period.

reported

data

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

time

interval
intervals

by

168

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

Any equipment which the observer uses


Therefore
the voltagewith which

"receiver."
"receiver

The

input."

"receiver"

the

of

this

by

who

the

is

placed

receiver.

here

and

out

instructions

fact that

have

been

various

these
the

actual

performanceis evaluated

its

called the

presentedis called

the

primarilyin specifying

of statisticalmethods
applications

intended

are

minimum

possess

subsequent sections

In

carried

the

on

consist

may

judgment is

is

observer.

signaldetectability.
They

problem
to
subjectfor those
definitions of "optimum"
to

the

this

observer

sections of this article survey the

first three

The

instructions

optimum

be used

to

make

to

the

to

serve

as

of mathematical

proposed by
definitions

other

lead

introduction

an

Several
training.

authors.

Emphasis

the same
essentially
optimum receiver is
of practical
cases

to

of the
specification
for some
numerically

interest [17].
1.1

and

SN
Population

signalplusnoise may be capableof producingmany


of all possible
receiver inputswhen
noise alone
diff'erentreceiver inputs.The totality
the collection of all receiver inputs
when
is present is called "Population N''; similarly,
with a
signalplusnoise is present is called "PopulationSN.'' The observer is presented
from
of
but
he
does
know
receiver
the
from
which
not
two
one
input
populations,
know
the probability
that it arose
from a
indeed, he may not even
populationit came;
the
The
observer
from
which
receiver
must
population
particular
population.
judge
inputcame.
noise alone

Either

1.2

or

the

Samplingplans

the
on
sampling plan is a system of making a sequence of measurements
interval in such a way that it is possible
to rereceiver input duringthe observation
construct
the receiver input for the observation
interval from
the measurements.
functions of time as sequences
Mathematically,a samplingplan is a way of representing
The
list a few examples.
of numbers.
this
idea
is
describe
to
to
simplest
way
that
the
interval.
observation
interval begins
A:
Fourier
series on an
Suppose
at time t^ and is T seconds
long,and that each function in the population6'A'^and A'^can
A

be

in

expanded

for

series

Fourier

receiver input can


particular

each

input,which

in turn

can

ao +

process

(oq,ai, bi,

InntjT

On,

in the

population the
(^0,"^i,bi,
.

ay,

of

bn ) is

that

on

by

the

formula

t^ "

-"-

"

to +

(1)

T.

function

.) is

Fourier

x(t) by

the

samplingplan
series which

sequence

in the

involve

sense

the

of its Fourier

described
cosine

and

efficients
co-

above.
sine

of

For such a
simply "series-bandlimited."
the
finite
representingeach receiver input x(t) by
sequence
a finite sample plan.*

series

sense

sampling plan is finite if there


inputsin the population.
A

measurements

population
cyclesper second. Suppose that for a particular
the population
of frequencygreaterthan njTa.ve zero; i.e.,

Fourier

process

bn,

in the

frequencynjT
inputsthe terms

is bandlimited

receiver

is of

of receiver

pairof terms

The

Zj"sm

each
representing

by making

coefficients

Fourier

The

l-rrnt
+

""

interval.

these measurements

from

lirnt
^71 COS

Z
n

the

obtained

be

be reconstructed

^
x(t)
Thus

observation

the

on

or

is a finite maximum

lengthfor

the sequences

for all

W.

Shannon''s

B:
all time

and

that

frequencies
greater than

(.

x{t^

this

x{t) by its

njlW),

the

case

\J2W\

The

instants

0 and

of time

a^

^0

receiver

from

input

is

populationis to

of the

represent

to)

apart,

.). In

inputis

receiver

77-[2
W(t

called

are

zero

/?]

(2)

2H^/

?(,+ n/lW

includes

band

spaced 1/2[^seconds
XjlW),
x(to+ nj2W),

x{t^ +
x(to),

sin

interval
to

every

this

samplingplan
at times
amplitudemeasured

[2] for the reconstruction

x(t)

of
for

169

FOX

observation

the

transform

W.

x(to

...,

formula

that

C.

W,

"transform-bandlimited"

Fourier

cyclesper second, i.e.,the

AND

BIRDSALL,

populationsare

to

function

G.

samplingplan. Suppose
the

for
each

T.

PETERSON,

W.

7T[2Wit

to)

n]

Each
sampling-times.
If the

all

different

observation

choice

of

/q between

samplingplan.
againincludes
from
transform-bandlimited
to a frequencyband
time, but
populationsare
then each receiver input
^/2 to/o + W/l which does not contain zero frequency,
/o
be considered
and
modulated
as
an
waveform,
x(t) can
x(t)
amplitude
frequency
ous
+ B{t));r(t)is the amplitude of the envelopeand
6(t)is the instantaner(t)cos{iTrfot
is obtained in this
phase of the carrier. A sampling plan employingsampling-times
each
receiver
the
K^o + "1^)^
case
by representing
inputby
sequence (. K^o)'^(^o)'
a
1/2ff yields

interval

the

"

6(tQ+ njlV), .) of envelope amplitudes and carrier phases


times spaced by 1/ff seconds
apart [1]. The reconstruction of
this
sequence is givenby
.

n=

that

the

^0 +

''

77; I

CO

"

2-/o?+ 0|ro+-

cos

"

Fourier

have

all time.

observation

If the

populationsare

series-bandlimited,then

times

similar

are

populationsand
beginningof

populations
plan for this

rrlW^t

can

in

from

obtained

be

of its amplitudes measured

0 to

inputfrom

n]

to)

to)

n]

interval.

therefore

and

appliesonly

the
the

when

Only
hypothesis

observation

lengthand if the
are
sampling plansutilizing
samplingfor
transform-bandlimited
B
paragraph
interval

interval.

which
interval,

the observation

situation

there

described

those

series-bandlimited

are

sequence

to

infinite observation

an

transforms,

transform-bandlimited

populationsare

interval includes

the

sampling-

at

(3)

''I

for all times

known

which

"

the receiver

Sampling plan using sampling-times


for a finiteobservation

C:
functions

"

sin TT[W{t

"t)

"

measured

is of

Suppose

is T seconds
f^

finite

that

time

long,and

cyclesper second.

by representingeach

seconds
l/2P'f^

is measured

that the

suppose
A

finite

receiver

from

sampling

input by

the

apart [1]

x{to),x\to+

fo +

(4)

T2W

2W

and

the

reconstruction

of the

2Trr-i
^1

receiver

input from
sin7r[2W{t

this sequence

is

n]

t^)
-

,0 "
2W(t

-to)

"

T.

(5)

-n

2^Frsin
2WT

Again

each

choice

samplingplan.

In

of the
a

(initial)
sampling-timeto between

similar

fashion,

if the

observation

0 and
interval

1/2^F yieldsa
is

unchanged

different
but

the

170

READINGS

MATHEMATICAL

IN

series-bandlimited
on
are
populations
which
include
does
not
to/o + W/2

PSYCHOLOGY

this interval

to

frequency,then

zero

from

frequencyband

each

receiver

/q

Wjl

"

be

inputcan

+ ^1^), ^(^o+ V^),


''(''o
representedby a finite sequence [K^o)'^(''o)'
fUo + T
at sample
IjW), 6(t()
+ T
l/fV)]of envelopeamplitudesand carrier phases measured
points 1/^F seconds apart; /q is againused to denote the initial samplingtime which
of the receiver inputfrom
may be chosen anywhere from 0 to 1/ff The reconstruction
-

"

"

"

"

this

of measurements

sequence

is

givenby
"

Infy

dlto+

these

From

examples

that

seen

kind, e.g., instantaneous

same

kinds, e.g., envelope amplitude and


the

common

made

property

role which
of

primarilyone
representedas
methods.

that

the receiver

the

an

can

in order

to

ferences
important dif-

from

the

of different
have

in

ments
measure-

Grenander

back

the

theory presentedin

The

populationsN
samplingplansin order

concerning an
to the

available for
that

SN

paper is
will

be

applystatistical

receiver, it is often

languageof receiver inputs.


of the theory,then
particular
application

parameters of the "optimum"


Both
for this reason
and
plans.
approximatedby usingfinite-sampling
the
the
here
where
is
restricted
to cases
simpify exposition, theorypresented

by

[3] shows

to

this

familiar

more
a

"optimum"

and

the

desired

be

available.

finite-sampling
plansare
2.
2.1

of

use

is obtained

answer

to translate this answer


possible
a finite-sampling
plan is not

work

of

be reconstructed

inputcan

convenience.

throughthe

sequences

Once

receiver

number

or
amplitude measurements,
phase. However, they all

carrier

sampling plan plays in

mathematical

If

recent

are

it.

on

The

there

samplingplans such as (a) the length of the observation


the measurements
are
employed, and (c) whether
sampling-times

be of the

all to

be

can

various

between

interval,(b) whether
are

it

(6)

"T.

"t

Optimum

Tests

Fixed

on

Observation

Intervals

Probability
density
functions
This

for
requires
of the

part
raw

of the

paper is concerned

data

finite sequence

with

method

of numbers

(x-^,
x^,

of statistical analysis
which
.

which
x."),

is the result

finiteinput accordingto some


particular
of
from
which
the
a
population
"sample"
samplingplan.
sequence
it arose, and is denoted
by a singleletter; thus, if the receiver input is x(t),and the
x"),then this sequence is called the sample
sampling plan yieldsa sequence (x^,x^,
X.
The theory to be developed here is intended
to specify
an
optimum receiver and is
couched
in the languageof samples,X
a receiver
x^. If n is very large,
(x^,x^,
which
had to make
called for by a samplingplan would
the measurements
certainly
this practical
avoided
when
the
be impractical.However,
is
of
difficulty
specification
from
the language of samples to the language of the
back
the receiver is translated
receiver inputs; this can
it is possible
the inputsfrom
because
be done
to reconstruct
made

measurements

The

at

the

receiver

is often

called

the

samples.

W.

AND

BIRDSALL,

G.

171

FOX

C.

W.

finite

purposes of the subsequent development any

the

For

T.

PETERSON,

W.

samping plan may

of the associated
known
are
sample X so
providedenough properties
tions
the probability
densityfuncthat certain probabilities
may be calculated. Specifically,
from
is
drawn
when
X
for
the
variable
X
of
the
cases
sample
fs{X) and/sA'CA')
of
The
basic properties
two
be known.*
must
A'^and
populations
^A^, respectively,
be considered,

densityfunctions

are

\fy{X)dX^\,

/ivW"0

(7)

and

^fs^iX)dX^\

fsdX)"Q
where

taken
integration
symbol representsthe multipleintegral

the

of the

range

sample

2.2

The

The

observer's

of a

concept
Consider

variable

(x-^^,
x^,

x^.

criterion
an

now

entire

the

over

has

who

observer

job is to judgefor each

sampleX

available data the

as

sample whether

not

or

{x^,

taken

it was

from

x").

tion
popula-

the (probably
subconscious) criterion
to determine
possible
of it. Ideally
to find an external manifestation
used by the observer, it is quitepossible
his
and
the
observer
to record
to
is
is
each
all that
to submit
sample
possible
necessary
decided
the
observer
judgment. This will yielda tabulation of those sampleswhich
is giventhis tabulation and
If
drawn
from populationSN.
were
any other observer
SN.

instructed

base

to

the

Thus,

itis not

Although

his decisions

tabulation

of these

employed by
denoted
by the

"Acceptingthe
denoted

2.3

by

the

observer

to

replacethe
a

criterion

mental

criterion and
in

the

will be

statistics of

phraseologycommon
signalis present."The tabulation of the remaining
from
drawn
concluded
were
populationA^,will be

refers to
a

used

will also be called

B.

associated

Probabilities
There

all

that
hypothesis

be

can

tabulation

which

letter A,

which

samples,those

responses

Such

the observer.

did the first observer.

exactlyas

it,he will behave

on

are,

of course,

with criteria
as

many

criteria it is necessary
possible

diflFerent criteria

to select those

that

as

are

there

are

observers.

best for various


with

each

Among

purposes.

criterion.

To

It will be

associated
must
quantities
will be
of the populations
that a sample from
one
probability
necessary
definitions,these probcriterion A. According to the standard
abilities
listed in a particular
are
givenby
be

do so, certain numerical


to

the

know

"

(8)

and

-J.,
where

the
*

from
Also

is
multipleintegral

population SN"
event

of

over

all

samples listed in

the criterion A.

kept in mind that "the event of the sample being drawn


corresponds signaland noise being present at the receiver input.
the same
thing.
populationSN being sampled" means

In this discussion

"the

taken

it should

be

to

172

MATHEMATICAL

IN

READINGS

PSYCHOLOGY

sample plan might have


example,a particular

For

function
density

of the form

would
+
Kexp [" {x^+ x^ +
lie outside
consist of those
a
,x^) which
(a^i,
samples X
x^,
sphereof radius 1
Then
the integral
would
be taken over
the exterior of this sphere.
centered at the origin.
These
have
a special
P\(A) is the conditional
ability
probprobabilities
significance.
in
that a sample from
criterion
that
will
be
TV
will
be
listed
A
is,
;
population
Thus
SN.
from
F
is
the
conditional
false
PyiA)
population
judged as a sample
alarm
of a certain kind of
Also, Pgy{A) is the conditional probability
probability.
that
called
correct
a hit (that of judgingcorrectly
a sample is from
population
response
of judgingfalsely
that a sample is from population
SN). The conditional
probability
of a miss.
^'A'^ is,therefore,givenby 1
M, the conditional
Pgy{A)
probability
alarms
and
misses
their
conditional
The only errors
which
false
can
occur
are
;
abilities
prob-

fyi^i,X2,

.r")

x^)]. A

possible criterion

"

called

F and

M,

reader

familiar

are

with

the
briefly
the formal

probabilities.

error

content

of

probability
theory should

note

that

these

the sample
on
are
true conditional probabilities
; the firstis conditional
quantities
from
drawn
from
second
is
conditional
drawn
the
its
on
SN;
being
being
population
them
from a prioriprobabilities
(the probabilities
populationN. This is to distinguish
that a certain population
will be sampled,for example)which are not as yet assumed
known.

2.4

Likelihood

ratio and

the ratio criteria

It is convenient
defined

as

the ratio

the likelihood
drawn

was

that

from

to

introduce

function

new

fgy(X)lfy(X)for sample points X


the

sample

was

from

drawn

SN

called the likelihood


=

(x^,
.

N.

in fact drawn

from

populationSN, i.e.,that Xshould

was

"best"

criterion. Thus, for each

/? "

number

0, a

ratio,KX),

a:")
; l(X) represents

relative to the likelihood

it would
Hence, if l(X)is sufficiently
large,

that A'

be reasonable

it

that

to conclude

be listed in the desired

certain criterion A{(i)will be selected;

each sample Jffor which l(X) " p. The problem then reduces
by listing
of ;6;that is,to determine
how
wise
choice
large"
making a
large"sufficiently

/4(^)is chosen
to that

of

is. Criteria of the

form

A{(^)will

be called

ratio criteria.

presentedvarying definitions of a criterion being


as a ratio
"optimum." It turns out that each of these optimum criteria can be expressed
criterion,so that a receiver designedto yieldlikelihood ratio as output could be used
A

with

2.5

any

number

of writers have

of them.

Weightedcombination

criteria

factor repas
a weighting
to assigna certain number
w
Suppose it is possible
resenting
relative to a hit. Since Pgy{A) is the probthe importance of a false alarm
ability
then be reasonable
of a false alarm, it would
of a hit,and Py(A) the probability
to

find

criterion A

which

maximizes

the

Psy(A)
But

this

quantitycan

be written

quantity
wPy(A).

(9)

as

I'^(X)

wfyiX)]dX,

(10)

W.

where

PETERSON,

W.

would

173

FOX

C.

W.

the

over

list in A

one
integral,
for
Solvingthat inequality

AND

BIRDSALL,

this
sample pointsX listed in ^4. To maximize
the
for
which
not
integrandwas
negative.
every sample
contain
those sample pointsX for
that A should
sees
one

is taken
integration

the

G.

T.

w,

l(X)='f^"w.
the

Thus

2.6

certain

[4] as

it would

maximizes
of

type

be reasonable

of
probability

the

optimum

(1) Py{A^)

it is

so

ratio criterion.

choose

to

hit.

from

Thus

and

below

criteria that

such

among

Neyman

Py{A)

proposed

Pearson

Ai^ for which

criterion

criterion any

false alarm

k, and

"

(2) Pgy{A,^)is a
The

and

of
important to keep the probability
critically

level k, then

which

one

simplyA{w),

criteria

Neyman-Pearson
If it is

is

criterion A

desired

(11)

criterion

A^. type

for all the criteria A

maximum

also

can

be

with

expressedas

the

This

criterion.

ratio

"

property P\{A)

k.
be

can

follows.
To beginwith, it is necessary to consider
as
only those criteria
plausible
in order to meet
will
taken
because
be
A
as
as
k,
large possible
Py{A)
the
consider
the curve
condition
(2). Now
by
equations
givenparametrically
made

for which

X(^)

Py[A(li)]

(12)

and
Y
This
for

will be

curve
a

receiver

output

PsylAiP)].

ROC)
OperatingCharacteristic (briefly,

Receiver

called the

whose

y(iS)

is likelihood

ratio and

ratio criteria

which

with

curve,

being

are

used.
The
the second

ROC
at

iS

samples.Thus

passes through the


0. At /S
0, liX) " /3
=

the

be certain to make
not

observer

will

false alarm

from
exclusively

be drawn

points(0, 0)

curve

basic property of the

0 for all X,

(1, 1), the first

and

^(0) consists

so

report that every sample is drawn


and to make
a hit. (This assumes

one

of the

populations.)This

densityfunctions expressedby

the

from

^3

at

of all

SN,

oo,

possible

so

he will

sampleswill

that the

be verified,using the

can

followingequations:

Ps^AA{0)]=jfsd^)dX^l
(13)

and

Py[Am
where

X(0)

the

y(0)

samples
will

is taken
integration

never

These
curve

1.

"

oo;

all

X{oo)

Moreover,

I{X)

can

nor

he

make

y(co)

i.e.,A{od) contains

hit.

Thus

can

somewhat

0, because
no

the

operator
=

as

those
in

of the

Fig.1.

0 and
next

at

that

mean
equations

for /^

samples

Pgy[A((X))]

considerations, togetherwith
be sketched

These

samples X.
possible

report a signalis present. Therefore,

false alarm

ROC

with

over

(fy(X)dX
\,

oo

all and

there
the

are

no

operator

possiblymake
0.
P\[Aico)]

cannot

section, show

that

the

174

IN

READINGS

MATHEMATICAL

XW)

PSYCHOLOGY

P^\A(0)]

Figure

TypicalROC
To
one,
curve

determine

that

so

the desired

Px(Ak)
lies

which

Pj^[A(^)]

k and

lie between
probabilities
there is a pointQ of

clear that
the

the ratio criterion

weightingfactor

usingthe
usingany

criterion A{^^

is
A((ij^)

both

case

is greater than

are

equalto

k.

the

same

(14)
is substituted

into

"

Psn(A).

(15)
be chosen

to

be this

ular
partic-

curve

It is desirable

to

digressfor

Its value

lies in the fact that

is

Aifi),then
criterion,

i?^,the

hPdA).

If this value

Therefore, the desired Neyman-Pearson criterion Aj^should


ratio criterion,
A(fij^).
ROC

to

obtains

one

PsdAi^,)]

2.7

if

equal

or

criterion A, i.e..

other

and PyiA)
Pjs;[A(P,c)]

above,
inequality

optimum weighted-

an

^^. Therefore,

the

the ROC

the

" Psn(A)
Psn[A("(^ic)]^icPnIA{Pic)]

In this

and

zero

Then

criterion with

weighted combination

all

one.

paragraph 2.5, it is

weightedcombination

and

(X, Y) of Q are
point (k, 0). The coordinates
Y
will be written /S^. Now
/S,which
Psn[MP)1 for some
and
will be the
therefore
(1) because
A((ijc)
Pff[A((i;.)]k,
for
Pg^[A{(i,^)]
any criterion with the property that P^vC/l) k.

desired Aj, if Psj^iA)"

combination

recall that

zero

above
vertically

A(^k) satisfies condition


From

Aj"

k is between

curve.

ratio
can

be read

off the

ROC

coordinate

is the conditional

conditional

probabilityof

being correct
of

when

miss.

noise

if the

moment

to

study the

type of criterion chosen

ROC
for

of the
completedescription

alone

It will be shown

is

present,and
in

moment

ROC

(1

"

Y)

that the

curve,

F coordinate

X) is the conditional

"

closely.

By the very definition of the


F, of false alarm, and the
probability
(1
Similarly

more

particularapplication
detection system'sperformance

curve.

hit.

curve

the

is the

of
probability
ability
prob-

is the conditional

operatinglevel ^

for the ratio

176

In the

same

of
probability

the

way

MATHEMATICAL

IN

READINGS

miss

PSYCHOLOGY

is

givenby

P(SN)M.
Because

an

the

of these

sum

error

can

in

occur

exactlythese

these

already been

substituted

are

ways, the

two

of
probability

is

error

quantities
PiE)

It has

(21)

into the

PiN)F

PiSN)M.

pointed out that F


for PiE)
expression

(22)

1
P^yiA) and M
Pgj"^(A).If
simplealgebraicmanipulationgives
=

"

PiN)
PiE)

It is desired

PiSN)

minimize

to

PiSN)

But

PiE).

(23)

PsdA)-^^yP^,iA)
from

the last

equationthis

is

to
equivalent

maximizingthe quantity
PiN)

^sn(A)

and, of
which

2.9 Maximum

decidingthat

optimum
such

assigna weightingfactor

criterion. This
and

Vd be the value

values, Vm,

to

way

known,

associated

if numerical

of detection

noise

alone

the value

of

with

is

criterion

criterion

as

values

criterion

with

PiN)lPiSN),

can

if the

the

"expected

PiSN)
prioriprobabilities

be

can

to

V^, the value

of

be determined.

now

which

one

depends on knowing

and

the four alternatives.

Let
assigned
and
that is,of correctly
Vq the value of being"quiet,"
The
other
also assigned
alternatives
two
are
present.

miss, and

be determined

can

maximizes

the

false alarm.

In this

case

The

expectedvalue

it is natural

expectedvalue.

It

can

to

define

be shown

an

that

criterion maximizes

Psn(A)

By

(24)

'

criteria
expected-vahie

of each
are

yielda weightedcombination
simplya ratio criterion Aiw).

be

to

Another

PiN)

Py(A\

^^y

this will

course,

is known

value"

definition

PiN)

Vn

P(SN)

Vn

Vp

Pn(A).

(25)

Vm

(see paragraph2.5),this criterion is a

weightedcombination

criterion with

weightingfactor
PiN)
w

Vq

PiSN)
and
case

hence

likelihood

for which

2.10/4

Vq

Vp

"

Vd

Vj,
(26)^
^

Vd-Vm'

"Ideal
Siegert's

Observer"

criterion

is the

special

Vm-

"

and signal
detectability
probability
posteriori

Heretofore
noise

ratio criterion.

"

"

is

the

present" or

best of his

observer

"noise

knowledge,is the

has

been

limited to two

"signal
plus
answers,
possible
present."Instead he may be asked what, to the
is present. This approach has the
that a signal
probability

alone

is

W.

W.

T.

PETERSON,

G.

AND

BIRDSALL,

W.

C.

177

FOX

information
from
the receiving
ward
more
advantage of getting
equipment. In fact, Woodand Davies
that
if
the
observer
makes
best
estimate
of this
the
out
point
possible
for each possibletransmitted
probability
message, he is supplyingall the information
which
his equipment can
givehim [6]. A good discussion of this approach is found in
the original
and Davies
for the a posteriori
[6,7]. Their formula
papers by Woodward
probability,
PxiSN), becomes, in the notation of this paper,

^^

l^^iX)P(SN)

(1

P(SN)}f^iX)

'

or

1{X)P(SN)

^^^"^^^

"

l{X)P(SN)

If

receiver

which

has

calculation
of

could

make

-P(SN)

as

its

output

be

can

the

receiver

calculated

calibration, since (28) is

receiver

receiver

optimum

an

if the

built, and

be

posterioriprobabilitycan

be built into the

l(X) ; this would

ratio

likelihood

is known,

probabilityPiSN)

^^^^

for

a priori
easily.The

tion
func-

monotonic

obtaininga

teriori
pos-

probability.
3.
3.1

SequentialTests

with

Minimum

Duration

Average

Sequentialtesting
idea

The

input; if the

is this : make
measurement
one
sequential
testing
x-^ on the receiver
is
decide
whether
the receiver
as
to
persuading,
x-^^ sufficiently
from
populationSN or from populationA'^. If the evidence is not so

drawn

inputwas
strong,

of

evidence

make

second

to

make

at

evenly spaced

x^ and

measurement

until

consider

the

the

evidence

Continue

x^.
(a-^,

of

is

measurements
measurements
sufficiently
resulting
sequence
Obviously this involves the
population or the other.
persuadingin favor of one
theoretical
of making arbitrarily
before a final decision
measurements
possibility
many
is made.
This does not mean
that infinitely
be made
in an
measurements
must
many
it
actual
does
that
the
entail
nor
mean
an
operation might
application,
necessarily
taken
measurements
are
long interval of time. If,in a particular
arbitrarily
application,

another

However,
?

1/2,

such

times

then

the

plan might
(rt
"

"time

base"

of such

call for measurements

l)/",and

as

these

times

plan

measurement

to be

all lie in the

made

at

time

the

is infinite.

instants

interval

from

/
zero

0,
to

plan would have a time base of only one unit of time.


If the measurement
plan has been carried out to the stage where n measurements
have
been
the variable
made,
a-")is called the "th
X^
{x-^,
x^, x^,
,x^
x^,
A
for
will
be
considered
measurements
specific
only if for
stage sample variable.
plan
each
possiblestage n, the two densityfunctions /^^CA",,)
and/^(A'")of the //th stage
variable
the
first
of
these
is applicable
when
functions
X" are known;
sample
density
when
population ^'A'^is being sampled and the second is applicable
population A' is
diff"er
different
These
well
functions
at
beingsampled.
density
stages,so that
may very
should
be
however, the // appearingin the argument
they
writtenyj^(^")
and/ifv(A'");
functions
the density
on
Xn should always make the situation clear, and the superscript
one,

measurement

themselves

will be

omitted.

178

3.2

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

Sequentialtests
A

assignmentof

(2) An

(A)

Signalplus noise

(B)

Noise

(C)

Another

measurement

first

of

At

the

present,i.e. the sample


should

the

the

present is drawn

This

entire

alone

noise
another

of

is present, and

againthe

will be made,

measurement

(real)number

system, which
the
in

test

will be written

S^

criteria A^, B^,


first-stage
conclusion
that a signal
is
/4i,the

If

is terminated.

the test

all could

three

If it is listed in

and

at

that the firststage sample

means

number

the test is terminated.

and

populationSN,
populationTV,

from

comes

plan, any

(.r^)
ranges through
for the first stage sample space.
to stand
Suppose
If the sample X^ is listed
and Q,have been chosen.
X-^

from

comes

be made.

measurement

the first measurement.

from

result
theoretically
variable

stage

conclusions:
possible

present,i.e. the sample

is

alone

three

is

things:

plan with densityfunctions/;v(J!f")


andy^;y(A'"),
criteria to each
plan.
stage of the measurement

three

representthe

criteria

three

of two

measurement
(infinite)

An

(1)
These

will consist

test
sequential

moves

By, the conclusion


X^

should

to the

on

is that

be listed in

second

Cy,

stageinstead

terminating.
When

space
way

stage criteria

the first

have

been

chosen,

limitation

is

5*2,the

placedon

(x^,X2) ranges. The only


through which the second stage sample variable X^
the
second
is
for
be listed in Q.
fore,
Thereto
to
can
Xy
(xy)
proceed
stage
but
all
second
does
contain
not
(x^,a^g) onlythose
^'2
possible
stage samples X2
"

the test

(xy)is listed

for which
chosen
there

from

when
The

An, B", and


those
in

C".

Three

A2, B2, and


stage criteria,

second

They
sample

that

signalis or

no

the

C", have

When

from
an

"

S2

is omitted.

that

the

test should

now

way

These

be

be
that

criteria

is,the three

continued, are

C2 respectively.

stage criteria
been chosen, then the next stage's
sample space Sn+i consists of
which
for
X^
x^) was listed
(x^,X2,
(x^,X2,
x^, x^+y)

selection of criteria

samplesXn+i
Then

is not

in

C2, must

in such

chosen

in the first stage. That

chosen

present,or
sample X2 is listed in A2, B2, or
a

be

must

no

conclusions
drawn

Q.

in the listings
and
duplications
the
those
as
same
exactly
significance

are

carry

in

samples X^ listed in S2.

those

proceeds in

the

same

way.

If the

"th

S^+i

are

drawn

the three

("

1) stage criteria ^"+i, B"+y, and

C"+i.

entire sequence

(A" By, Ci),


'^2'

^2)5

"2'

(^rn "m C"),

of criteria is selected,a
course

that

the

test

test"
"sequential

will

be
necessarily

a
possibleways of selecting
be
which
ones
particular
may

sequence
are

has

been

determined.

useful.
particularly
of

criteria and

very useful.

hence

This

does not

However,
a

among

mean

of

all the

test, there
sequential

3.3

associated

Probabilities

is any

If Qn

T.

PETERSON,

W.

W.

/?th

with

G.

AND

BIRDSALL,

C.

W.

179

FOX

tests
sequential

stage criterion,then

the

quantities*

(29)

and

represent the (A^ or

listed in the criterion

(1) The

A?th

conditional

SN)

Q,j.

Conditional

stage conditional

that an A?th stage sample Xn


probabilities
interest are:
of particular
probabilities
error
probabilities:

will be

population A'^ is sampled, then the probabilitythat the sample variable Xn


of a false alarm.
listed in A^ is P^(A"). This is the A^-conditional probability
the
that
then
the
If population^'A'^is sampled,
sample variable X^
probability
is the 5A^-conditional
listed in B" is Pg^{B"). This
probabilityof a miss.
If

will be

will

be

conditional

(2) The

of the
probabilities

error

entire

test

00

2 PN^'^n),the

of
probability

A^-conditional

(30)

false alarm, and

oo

are

'S'A^-conditional

th^
2 "PsA'(^n)'

M=

merely the

of the

sums

conditional

(3) The

of
probability

(31)

miss,

all stages.
over
probabilities
of
probabilitiesterminatingat stage n are
same

error

PNi^n)

(32)

Pn^B,,),

and

T"N
These
terminate

at

equationscan

mutually exclusive

is the

sample

variable

of the

sum

since Xn

conditional

(4) The

PsNi^n)

(33)

PsN(Bn).
The

by a simpleargument.
justified

is for the

stage n

of this event
probability
are

be

can

X"

to

be listed in either

of the
probabilities

be listed in at most

that
probabilities

one

the entire

test

the test

only way
A"

component
of A"

or

B".

events

and

The

which

B".

will terminate

are

1 n,

Tn=

can

(34)

"=i

and

3.4.

(35)

Ts\.

Average sample numbers


There

are

sequentialtest

is that

samplingprocess when
*

The

pointslisted

notation
in

Q".

be introduced.
which
must
quantities
it affords an
opportunityof arrivingat a
the

data

happen

indicates that the

^Qn

to

be

feature

One

other

two

decision

unusuallyconvincing. Thus

is
integration

to

be

carried

out

over

of the

earlyin the
one
might
all

sample

180

READINGS

the

expect that, on

IN

PSYCHOLOGY

stage of termination

the

average,

MATHEMATICAL

than

lower

be

could

of

achieved

be

well-constructed

sequential

by an
equal,good standard
It is therefore
test.
important to obtain expressionsfor the average or expectedvalue
other probabilities,
As with
there will be two
of the stage of termination.
of these
conditional on population
N beingsampled; the other conditional
on
one
quantities:
^A^
are
givenby
beingsampled. They
population
test

would

otherwise

00

En=

nn

(36)

^Tly.

(37)

and
00

EsN

The

is used

letter E
called

are

the

refer to the term

to

termination"

the

take

may

It should

be

be

less than

that

on

the

for

sample

average
so

if it is known

that

are

test

made,

3.5

Tgy

each

become

Sequentialratio

that

the
this

The

take

which

the

but

be

can

E^j^^r

justified

"stage of

variable

the

sense

not

situation would

mean

strictly

are

will

times
some-

will also be,


upon

occasion, much

numbers

finite would

assumption,*it can
of

numbers

stages of termination
are

to be considered

only ones

(in the

sample

not

those

are

that

be shown

Ty

with

be
finite

Tgy

On the other
probability).

hand,

always follow that the average sample


only that if a sequence of runs of the

would

probably terminate, but the


made.
runs
were
largeas more
arbitrarily
run

and

fy
quantities

formulas

that
(conditional)probability

average

sample

1 it does

the

numbers

average

to terminate

finite. Such

numbers

would

Ty

the

n,

test, the
sequential

whose

Under

that the test is certain

were

Therefore
applications.
numbers.

value,

weightedby

sample

average

test
larger. Any sequential

useless

of

runs

each

these

value.

heavilyemphasized

In actual
figures.

average

be

must

form

The

that

grounds

on

will in fact take

variable

"expected value."

average sample numbers.

freely)on

(somewhat

average

tion
stage of termina-

tests

found that the best


tests usingfinite samples it was
studyingnon-sequential
it may
be
be
in
of
likelihood
ratio.
terms
Therefore,
always
expressed
The
nth
likelihood
ratios
of
infinite
useful to introduce
at each stage
an
sample plan.
ratio
is
the
ratio
likelihood
function
defined
as
liX")
/g^7(A'")//v(^n)Optimum
stage
turned
be
criteria
all
criteria in the finite-sample
tests
out
to
listing samples X for
which
be possible
It should
to choose
l(X) is greaterthan or equalto a certain number.
in
numbers
the
criteria
For
each
two
same
(/4",5", "")
fl" and
sequential
stage
way.
In

criterion could

bn

bn

with

numbers

"

fl" could

be

bn

An

lists all

samples X^

Bn

lists all

samples X^, of

Cn

lists all

samples X^

the Xi.

Remember

would

chosen.

a" and

that

the

Then

the

criteria (A", B^, C") determined

by

the

be

of the

sample space S^

for which

/(A'")"

a",

b^,

the

sample

space

5" for which

l^X^)

of the

sample

space

Sn for which

b^

process

Is

assumed

sampling

not

to

"

"

l{Xn)

"

a^.

yieldindependenceamong

W.

PETERSON,

W.

in this way

be finite,then

resulting
sequential

meet

is called

test

181

FOX

C.

W.

requirementsthat

the

If criteria selected
the

AND

BIRDSALL,

G.

T.

the average

sample

numbers

ratio test."
"sequential

tests
Optimum sequential

3.6

sample

the average
fixed

addition

sample

average

"'^v ^^"^

minimum

givenin

Section

numbers

all

among

that

as

for which

one

with

tests
sequential

M.

formulas

the

to

test
optimum sequential

an

"'v and

numbers

F and
probabilities

error

In
the

define

customary [8] to

It is

formulas

3.4, alternative

[9] for

are
00

"-v

2 P,v(Q)

1 +
1

(38)

and

2 PsxiCi).

1 +

^-^.Y
=

if

Thus,

test, then

set

its

(39)

criteria {A*, B*, C*) is presentedas a possibleoptimum


sequential
whether
the inequalities
is decided
character
by ascertaining
optimum
of

iPxiC*) "J,Py(Q)

(40)

lPssiC*)"2Psx(Q)

(41)

and

I
for every

hold

other

of

set

criteria {{A^, B^, C")}


sequential

with the

same

abilities
prob-

error

i.e.,with

^P^{A*)=J^P^-{Ad

(42)

lPssiBt)^2Psx^Bi).

(43)

and

the
test is difficult because
an
optimum sequential
problem of constructing
is
when
there
satisfied
and
be
no
even
(43) can
(42)
apparent term-by-term
equalities
has proposed as optirelation between
the sequences {P^(C*)} and
mum
{Py{C,)}. Wald
The

the tests in which


and

fli

a"

the

case

noise."
of

for

the

densityfunctions

example

when

at least
signaldetectability,

is not
in the

successive

and
met

sense

stages are

tests

independent,as

Z)"
are
can

signalplus noise consist of "random


of the theory
with in most
applications
hypothesesof

that the

Wald

and

witz
Wolfo-

satisfied.

are

Consider
F and

not

at

that is,6^

[10] proved that these

Wolfowitz

noise

both

this "randomness"

However,

and

Wald

{6"}is constant,

{o"} and

of the sequences

each

Moreover

n.

whenever

optimum
be

for all

test of fixed

Although
generally
requiresless
M.

the

lengthas

described

in Section

test with
sequential

optimum

2, with

these

same

error
error

probabilities
probabilities

it has the disadvantagethat it will sometimes


with the
length test requires.In a conversation
the
that
Professor
Mark
or
Kac
of
Cornell
authors,
Universitysuggested
dispersion,
usefulness
aff"ect
the
be so largeas seriously
to
variance,of the sample numbers
may
of the sequentialtests in applications
to signaldetectability.
Certainlythis matter
should
be investigated
before
final decision is reached
a
concerning the merits of
use

much

more

tests
sequential
matter

to

time

time

than

on

the

relative to tests

calculate

the

the average,
fixed

on

variance

fixed observation
of

the

sample

interval.

numbers.

However
Therefore

it is a difficult
an

electronic

182

PSYCHOLOGY

of Michigan which will simulate


beingbuilt at the University
for
data
ROC
of both types as well as the
curves
provide
sample numbers.
(sequential)

both

is

simulator

of tests and

of the

MATHEMATICAL

IN

READINGS

4.

Cases
for Specific

Detection

Optimum

types

distribution

will

4.1 Introduction
The

chief conclusion

presentedin Section

input is

receiver

ratio for each

obtained

2 of this paper

from
is that

the

generaltheoryof signaldetectability

the
a

receiver which
receiver

optimum
TABLE

for

calculates the likelihood

detecting
signalsin

of SignalEnsemble
Description

Section

Signalknown

4.4

Application
Coherent

exactly^

known

except for phase"

Signalknown

4.5

radar

with

range

and

white

Signala sample of

4.7

output of

Detector

broad

Gaussian

noise

time
a

radar

video

radar

case

train

(A

incoherent

with

of

pulses

Signalone of
signals

other

and

pulsefrom
a crystal-

type

is at

with

range
radar

Coherent

orthogonal

ing
start-

with

Ordinary pulseradar

phase)

known
a

broad

receiver

known

4.10

(such
beacon)

as

or

band

4.8

target

signals;
speechsounds in

of

Detectinga pulseof

band

no

of noise-like

detection

receiver

with

range and character

Detection

sian
Gaus-

noise

target of

character

Ordinary pulse radar


and
with
integration
of known

4.6

noise.

of

one

target of
character

and

where
a

gration
inte-

with
a

the

target
of

finite number

positions
non-overlapping
4.11

Signalone
known

of M

orthogonalsignals

except for phase

Ordinary pulse radar


integrationand with
which

may

finite

number

appear
of

with
a

at

no

target

one

of

lapping
non-over-

positions
*

Our

of these two

treatment

work, but here they are

fundamental

treated in terms

cases

of likelihood

is based

upon

ratio,and hence

Woodward

apply

and
to

Davies'

criterion type

probability
type receivers. These first two cases have been
posteriori
in
solved for the more
which
the noise is Gaussian
but has an arbitrary
generalproblem
of
infinite
Those
solutions
the
use
an
require
sampling plan and are
spectrum [11,12].
derivations in this report.
involved than the corresponding
considerablymore

receivers

as

well

as

to

184

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

described
distributions are
by givingthe probability
probability
receiver
the
for
x.
/sv^-'^')
and/^,(a:)
inputs
densityfunction for the receiver inputswith noise alone are
probability

plus noise. The


densityfunctions
The

assumed

be

to

..2

/ivC^) IT

exp

WittN

"

x^
IN

(46)

or

\M/2

/-(^"

I2^^

2N^/']

exp
.

where

and

is 2 WT

"

A''is the noise power.

It

be verified

can

of noise
which
densityfunction is the description
is
and
has
the
at
stationary,
amplitude every time,

Fourier

Thus

components.

Gaussian

we

shall

refer

to

it

has
same
as

that
easily
Gaussian

this

probability

distribution

average power

of

of its

in each

white

"stationarybandlimited

noise."
functions

The

are
rpjjit)
orthogonaland

have

energy 1/2PF, and

therefore

[x{t)fdt.

(47)

0
so

that
1

^^^^^

where

Nq

Nl

\"/2

x(tf dt

"'^P

noise power

IV is the

'^

(48)

per unit bandwidth.

information
is givenabout the signals
as
they would
practical
application,
the signal
noise at the receiver input,rather than about
plusnoise
appear without
and
the
this
information
from
Then
be
calculated
must
density.
fs^i^)
probability
will be
for the noise.
The
noise and the signals
probability
densityfunction f;^(x)
of
each
other.
assumed
independent
the
the noise, then
of the signaland
If the input to the receiver is the sum
receiver inputx{t) could have been caused
s{t)and noise n{t) x(t) s(t).
by any signal
in
the
The
densityfor
input
signalplus noise is thus the probability
probability
that s(t)and x(t)
^(O.
s{t)will occur
averagedover all possible
(density)
together,
then
function y^(5),
If the probability
is described by a density
of the signals
In

"

"

sYs(^)ds,
J/yvC-^

/s.v(^)
=

where
form

the

is over
integration

is used

when

P,^; the formula

the

case

the

is

fsd^)

This

general
sample variable s. A more
is described by a probability
measure
signals

the entire range of the

of
probability

in this

(49)

\fN(^

and
is a Lebesgueintegral,
integral

(50)

s)dPs{s).

is essentially
an

"average"of f^{x

"

s) over

W.

of

all values

the

weighted by

G.

T.

PETERSON,

W.

AND

BIRDSALL,

is
/^rCi)

If

Pg.
probability

185

FOX

C.

W.

Eq. (46), this

from

taken

becomes

=).-

f^ix -s)dPsis)

/.w(^)

Y^'

"

dPgis)

\2WV/
(51)
1

exp

2^N

2/V
1

^2

exp

2N

J J

'

-A-

dPgis),

=1

"/2

fsyi^)

J/vC-

^) ^^sW

i:^]f f -^i| [.r(r)

exp

2777V

s(t)fdt\dPgis)

^n

(52)

The

2/2
f'^

factor

f^

x^ dt

exp

exp

-^''(0^?

-(l/7Vo)

exp

If

"

exp

f^

^^ ^f

tt"

exp

"

x|]can

S
[ -(1/2A'^)

r^

xs

be

dt

dPsis).

brought out

of the

since
integral

it does

s, the variable

depend on

not

of

that the

Note
integration.

J ^(0'^?=^2^?

(53)

=E{s)

is the

energy

expectedsignal,while

of the

'
\

is the

4.3

correlation

cross

between

ratio with

Likelihood

Likelihood

by (46) and

x{t)s{t)dt=^2^iSi
the

Gaussian

white

the

expectedsignaland

receiver

(54)

input.

noise

is defined

ratio

With
fj^^{x).

fgyix)and

integral

as

Gaussian

ratio

the
noise

of

densityfunctions
probability

the

by dividingEq. (51) and (52)

it is obtained

:
(48) respectively

E(s)

/(.r)

exp

exp

^x

(55)

dPsis),

TV

E(s)'
l{x)
If the

signalis

signalis unity,and
zero.

Then

exp

the

exp

x(t)sit)dt

dPsis).

(56)

L^OJO

for that
the probability
exactlyor completelyspecified,
s is
probabilityfor any set of possiblesignalsnot containing

known

the

likelihood

ratio becomes

E(sy

Ip)

exp

No.

"

(57)

exp

or

E(sf

/,(.xO
-

Thus

the
*

generalformulas
This

assumes

exp

(55) and

that the circuit

exp

dt
x{t)sit)

(58)

N,

(56) for likelihood

ratio state that

impedanceis normalized

to

one

is the
/(.r)

ohm.

weighted

186

READINGS

over
average of //.r)

set of all

the

i.e.,
signals,

l{x)
An

equipment which

PSYCHOLOGY

MATHEMATICAL

IN

(l^{x)dPsis).

calculates

likelihood

the

(59)

ratio l(x)for each

receiver

input

is the

optimum receiver. The form of equation(58) suggests one form which this
lihood
expected signals, the individual likeequipment might take. First,for each possible
ratio lg(x)is calculated.
Then
these numbers
are
averaged.Since the set of
is
is
this
direct
It is freoften
method
quently
infinite,
expected signals
usuallyimpractical.
in
obtain
cases
to
on
possible particular
by mathematical
Eq. (58) a
operations
X

different form

for

equipment,simplerthan
this which

the
as
recognized
response
by the direct
equipmentspecified
foUov/ing
paragraphs.

l(x)which

is done

in the

If the distribution

function

Pgis)depends

and
phase,signalenergy, or carrier frequency,
are
independent,the expressionfor likelihood

these parameters
functions

indicated

are

denoted

are

by

r^, r^,

hyf-^{r-^),
/^{r^,
.

The

likelihood

of

be

can

the

"

"

various

on

if the

be

can

It is

essentially

parameters such
in these

distributions

ratio

carrier

as

parameters
If

somewhat.
simplified

the associated

r^, and

realizable electronic

method.

probability
density

then
,fn{rn),

ratio becomes

l{x)

"

"

Is(^)f
lifi)

"

"

drj^- drn
-fnii-n)
"

"

"

(60)
=

Thus

4.4

the likelihood

77?^

likelihood

can

be found

4.3

with respectto
lg(x)
by averaging

the parameters

exactly

ratio for the

presentedin Section

"

"

known
of a signal

case

The

been

ratio

] dr^.
\fiiri)lsi^)dr,
\\fnirn)---

case

when

signalis

the

known

has already
exactly

Fi

"

(61)

(62)
As

the first

the distribution
step in finding

find the distribution

(.^i,
x^,
Xf has

x.")is due

distribution

normal

(5i,52,

...

SilN times the


zero

and

(IjN) S

to white

independent.Because

are
s

for

Sr,),each

mean

the

x^s^ when

Gaussian
with
Sj

zero

are

summand

of x^, and

Because
5'f/A'^
respectively.

independent,each

with

normal

noise.
mean

constants

variance
the x^

for

is noise

alone.

It

are

can

be

A'^

dependingon
a

(sjNf

normal
times

Then

and

signalto

be

sum

each
the x^

detected,

with

mean

ofx^, which

are

{siX^jN are

summands

their

to

inputs
that

distribution

the variance

therefore

the

Eq. (46) that

WNq

"

the

independent,the

distribution,and

l(x),it is convenient
from

seen

variance

and

{x^Si)jN has

with

functions

there

has

normal

W.

distribution with

the

mean

G.

T.

PETERSON,

W.

of the

sum

AND

BIRDSALL,

means

zero
i.e.,

"

C.

W.

and

"

187

FOX

variance

the

sum

of the

variances.

5?
".,

IWEis)

IE

N.

H
(1/A'^)

x^Si

SignalEnergy
(63)

N
distribution

The

variance

for

lEjN^.

that

sees

defined

with

Recallingfrom
l{x)

one

Noise

the

noise

Power

alone

Per

Unit

is thus

Bandwidth
with

normal

zero

and

mean

Eq. (61)

for

distribution

(64)

exp

N-

TVn

21 x^s^
(1/A'^)

be used

can

by introducing
directly
a

by
E
ha

exp

or

to {XjN) S
l{x) " /?is equivalent
inequality

The

Fn(P)

47tE

distribution for the

The

which

Because

(65)

therefore

dy.

llE-^

(66)

be found

can

by usingEq. (19),

that

states

these

likelihood

a, and

signalplus noise

of

case

In iS

"

x^si "

exp

are
equal to
probabilities

ratio,this

be

can

written

the

distribution functions

complementary

for

as

(68)

^ dFs{(i).

dFsNil^)

Eq. (66),
Differentiating
da..
^^A'(^)=-./^exp

(69)

AE

obtains

combining (65),(68),and (69),one

and

^0

dFs^iP)=

-J^exp

Fsyii^)

ue].

exp

/3,have

normal

1- a

(70)

doL.

AE

Thus,
=

In summary,
well

with

as

of the

any

is

and
*

In

In

alone; the variance

2EJNq.
operatingcharacteristic

with

In / has

normal

plusnoise.
signal

Fig. 3, the

paper.

On

receiver

this paper

erf (x)

"

the receiver

with

parameter

axes

are

linear

(1/ V277)
"

in

and

3*

variance

same

are

both

plottedfor
with

noise

figureis equal to the square of

curves

in the

signalplus noise as
lEjN^, and the difference
with

Figs.2

the

d in this

exp

"

J
this makes

curves

operatingcharacteristic
both

A^o/.

distribution is

distribution
The

distributions

of each

receiver

in which

case

alone

noise

means

The

therefore

and

(71)

dy.
4

error

plottedon

are

function

[-fil] dt;

CO

lines.
operatingcharacteristic straight

bility"
"double-proba-

188

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

1,0

0.9

0.8

0.7

0.6

S 0.5

0.4

0.3

0.2

0.1

0.1

0.2

0.3

0.4

0.5

Figure

Receiver

operatingcharacteristic.

the difference of the means,

apply

curves

to

Eq. (62) describes

operationin

divided

the

the receiver is

can

means

of

be taken

care

obtainingcross

If the

correlation

form

of the

1.0

normal

deviate

with

teristic
by the variance. These receiver operatingcharacknown
d
with
IEJNq.
exactly,
signal
=

the ideal receiver should

do for this

obtainingthe correlation,

adding a constant,
simplyin the calibration

constant,
of

0.9

In / is

correlation

have

been

and

case.

^(0^(0 dt-

Jo0

tions,multiplying
by

0.8

of the

case

what

0.7

0.6

The

The

essential

other

opera-

tion,
takingthe exponentialfunc-

of the receiver

output. Electronic
[13].
developed recently

signalis simple,there is a simpleway to obtain this cross


impulseresponse of a filter. The response e^it)

[6, 7]. Suppose h{t)is the

of the filter to

voltagex{t)is
e^it)

x{t)hit
J

"

"

"

t) dr.

(72)

W.

W.

If

filter can

be

G.

T.

PETERSON,

W.

AND

BIRDSALL,

C.

189

FOX

that

so
synthesized

h{t)

s{T

h(t)

0,

t),

"t

"T

(73)
otherwise,

then

eoiT)

so

that the response

ideal receiver

consists

0-

of this filter at time

simplyof

(74)

xir)s(r)dT,

Tis

filter and

correlation

the cross

the

required.Thus,

amplifiers.
99.9

99.5
99

98
97
96
94
92
90

80

70

Vd-

60

50

d
o

40

30

20

10
8

6
4
3
2

1
0.5

0.1
0.1

0.3 0.5

12

810

20

30

40

50

60

70

90

yb

100F^(0
Figure

Receiver

operatingcharacteristic.

In / is

normal

deviate,

a\y

cr^-,
(A/^y

"

Mjy)-

da\-.

190

when
specified

power
4.5

for the filter which

asks

one

PSYCHOLOGY

that this filteris the same,

be noted

It should

MATHEMATICAL

IN

READINGS

except for

maximizes

constant

factor,as that

peak signal

average noise-

to

ratio [14].

except for carrier phase

Signalknown

considered in this section consists of all signals


which differ
ensemble
signal
from
and frequency
modulated
a givenamplitude
signalonlyin their carrier phase,and
all carrier phasesare assumed
equallylikely.
The

s{t)=f{t)

phase angle6

Since the unknown

has

uniform

(75)

[oit+ "i"{t) 6].

cos

distribution,
(76)

ATT

The

likelihood

E{s) is the

ratio

for all values

same

by applyingEq. (56),and
the carrier phase 6,

be found

can

of

since the

signalenergy

l{x)

Expanding s

exp

into the coefficients of

s(t) =f{t)

cos

exp
d and

cos

cj"{t)]
cos

[oit+

^i^i

sin 6 will be

dPsis).
:
helpful

sin [cot+

+f{t)

(77)

sin 6,
"f"(t)]

(78)

and
1

1
^

Tr

2.^1^1

COS

NT

Z^ifih)

"

[ojti+ "i"{t,)]

cos

1
+

Because

easiest to introduce

"

Z^if(fi) sin [oiti+ ^(ti)].'

(79)

with respect to 6 to find the likelihood ratio,it is


integrate
parameters similar to polarcoordinates (r,6^)such that

wish

we

sin

to

0"

r cos

"

^
=

[oJti+ 4"{ti)]

T^Z^ifih)
COS

(80)
1
-

and

00

r sm

]^Z *i/(0

[wti + "}"{t,)],

sin

therefore
1

r
^

^Z^i^i =;^cos(0
-

Using this

form

the likelihood

(81)

^o)-

ratio becomes
dd

-cos(0

-0o)

2^
(82)

where

/q is the
*

Bessel function

ti denotes

the /th

of

zero

order

samplingtime, i.e.,ti

and
=

pure

ijlW.

imaginaryargument.

192

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

1.0

0.9

0.8

0.7

0.6

s;

0.5

0.4

0.3

0.2

0.1

0.1

0.2

0.3

0.4

0.5

0.7

0.8

0.9

1.0

Figure

Receiver

0.6

except for phase.

characteristic. Signalknown
operating

Eqs. (85) and (89) yieldthe receiver operatingcharacteristic in parametricform, and


levels [15]. These
are
graphedin Fig.4 for some
Eq. (84) givesthe associated operating
used
as
were
signalenergy to noise power per unit bandwidth
the
that
2
of
effect
and
known
3,
so
knowing
exactly,
Figs.
phaseangle
the phase can
be easilyseen.
the
If the signal
to match
is sufficiently
simpleso that a filtercould be synthesized
known
in
0
the
of
for
a
case
a
exactly,
signal
expected signal
givencarrier phase as
likelihood ratio. For simplicity
then there is a simple
to
a receiver to obtain
design
way
let us consider only amplitudemodulated
signals\^{i) 0] in Eq. (75). Let us also
the filter has impulse
0=0.
choose
(Any phase could have been chosen.) Then
of the

when

same

the

values

of

was

response

h{t) =f{T
=

0,

t) cos

{oj{T

t)l

"t

"T,

(90)
otherwise.

W.

output of

The

?o(0

filter in response

the

x(r)h{t

t) ^T

G.

W.

AND

BIRDSALL,

193

FOX

C.

T(t) is then

to

x(r)f{T+

C"(t +

t) cos

t) dr

Jt-T

-co

C0(T

cos

T.

PETERSON,

W.

x{T)f{T +

t)

"

t) cos

"

dr

WT

Jt~T

"

of the

and
integrals,*

x{T)f{T+

t) sin

"

the filter output will be the square

envelope of

The

-I

sin co{T

the

time

envelopeat

Twill

(91)

dr.

ojT

of the

root

of the squares

sum

proportionalto rjN, since

be

fT

x{r)f{T)COS

dT

lOr

sin
!:(t)/(t)

dr

COT

(92)

2W

which

be

can

identified

x(t)passes through
linear

square of the envelope of ^qCO at time T. If the input


an
impulse response givenby Eq. (90),then through a

the

as

the filter with

known

Eq.
the

likelihood

4.6

Signalconsisting
of a sample of white

ratio of the

Suppose
Gaussian

monotone

to

Gaussian

with

?i/2

2tt{N +

S)

a, defined

the

by

!1 y

exp

of

presence

variance, and
of freedom,

;8 is

l(x) "

noise

they

squares of these

2N

(93)

are

alone

the

will exceed

can

^,2

(94)

'

to

introduce

the

or

zero

be shown

have

at zero

that these

(\IN) 2 xf
zero

that
probability

ox

frequencyand at
contain
integrals

with

In

cr.

and

mean

the

"

unit

of the

sum
n

degrees
(96)

K,,{rf?).
,.

xJVn
the

that

distribution
chi-square

is the

^^.

S2

condition

the

variables

random

If the line spectrum o^ s(t)is


it

S^

for /,it is convenient

\N

i.e.,

Ico/ln,then

(95)

exp
""^

to
equivalent

greaterthan

2N

independent. Therefore,

variables

1 __}__ y

V"/2

FsiP)

co/Itt.

equation

,A^ + SJ

'

functions

TV

condition

^2

2N^

distribution

determiningthe

parameter

the

since

1-f

exp

nli

l{x)^

the

The
signalpower.
is
noise
plus
signal

S, the

Gaussian,

is also

independent

sample pointsare

ratio is

likelihood

The

Then

read

1 WT.

In

ratio,
to

variables:
1

the

variance

and

mean

zero

signalplus noise

/s.v(^)
where

output

calibrated

be

can

noise

signalvoltageat

random

Gaussian

of two

sum

of the

variables

probabilitydensitydue

the likelihood

Because

input.

the values

random

T.

the
/-/TV,

of

function

(82), is

the

at time

detector, the output will be (Nj2)rlN

all frequencies
equal to
no

or

as high as
frequencies

194

READINGS

in
Similarly,

[IJiN +

The

unit variance.

S)] S

xf

[Nl(N

"

MATHEMATICAL

PSYCHOLOGY

signalplus noise

presence of

the

and

mean

zero

that

IN

the

5)]a2,and

For

(97)

distribution
chi-square
a^ " 0,
for
[16],
precisely

K^io^')

requiring
distribution,
chi-square
as

is

approximatelynormal

Fd^)

same

of the

use

S have

K"

of n, the

largevalues
portion;more

a^ is the

"

again making

the center

xj Vn

variables

iljN) E xf

condition

FsdP)

random

exp

VItt

dy,

2^

over

(98)

"v/2a2-\/2w-l
and
\

exp

If the
and

this

signalenergy

case

is small
have

distributions

both

too, with the value

noise, VnI(N

of the

compared to that
nearlythe same
of d givenby

Then

variance.

dy.

--r

(99)

S) is nearlyunity

Figs.2 and

apply

to

TV
d

{ln

1) 1

(100)

N
For
relation

index

signalto noise ratios and largesamples,there is a simple


of samples,and
the detection
signalto noise ratio, the number
small

these

between

d.
1 5

TV
1

(101)

and

Two

signalto

noise ratios,{SjN\ and

characteristic

if the

the
{SlN)^,will giveapproximately
of

correspondingnumbers

sample points,Wj

ing
operat-

same

and

n^,

satisfy

A^

(102)
^X2
N}

By Eq. (94),the

likelihood is a monotone

function

of

xf. But

the

output of

an

energy detector.

{"t)fdt=-:^J^xf
o(0=J
is

proportionalto S a;?. Therefore


ratio,and

hence

can

be used

an

as

energy
an

detector

optimum

can

be calibrated

receiver in this

case.

to

(103)

hood
read likeli-

W.

Video

4.7

output of
a

band

Gaussian

BIRDSALL,

AND

W.

C.

195

FOX

receiver

problem considered
signalsand noise are assumed

The

G.

T.

band

broad

designof a

The

to

PETERSON,

W.

in this section

is

in Fig.5.
representedschematically
and at the
to have
passedthrough a band pass filter,
the filter,
A
the
assumed
be
limited
in
to
on
are
diagram,they
point
spectrum
of width
^Fand
center
to be
frequencycoJItt" H^j2. The noise is assumed

noise

pass

through

and

noise

with

uniform

spectrum

linear detector.

the

over

band'.

output of the detector

The

The
is the

and
signals

noise

envelope of the

then

signals

they appeared at pointA; all knowledge of the phase of the receiver input
is lost at pointB. The
ceiver
resignalsand noise as they appear at point B are considered
and
the
of
is
these
video inputsto ascertain
inputs,
theory signaldetectability
appliedto
the best video design and the performance of such a system. The mathematical
of the signals
and noise will be given for the signals
and noise as they
description
appear
The envelope functions,which
at pointA.
appear at point B, will be derived,and the
will be found
likelihood ratio and its distribution
for these envelope functions.
The only case
here is the case
in which
which will be considered
the amplitudeof
it would
the signal
function of time.
as
appear at point A is a known
A
be
band
function
will
limited
at
of width
W and
to
a band
center
Any
point
frequencyajjln " W/l. Any such function/(0 can be expanded as follows:
as

f(t)
where

x(t) and

themselves*

be

y(t)are

band

x(t)cos

y(t)sin cot,

(105)

limited to

no
frequencies
higherthan
expanded by samplingplan C, yielding

I-^ IVM

CO?

cos

2/ (

sin
Iy"i{t)

"

PF/2,and

hence

can

(106)

cot

The

amplitudeof

the

function

r(t)
and

thus

the

amplitude at

the

/th

/(r)is
V[xit)f + [yiOr,

(107)

samplingpoint is

^^""^
^^*'
'''Tf)'"'*'

(108)

"

The

angle
arctan

might be
might be

considered
described

the

phase of/(0

by givingthe

Input from

Band

antenna
or

at the

r^ and

arccos

"

/th

samplingpoint.The

d^ rather

than

of

amplifier

Point B-

FlGURE

diagram

function/(0then

y^.

Video

detector

Point A-^

Block

the x^ and

Linear

pass

filter

mixer

(109)

"

broad

band

receiver.

Because
-\-{WjD, the
frequency greater than {collir)
any function /(/) at A has no
samplingplan C might have been used on/(/). However, the distribution in noise alone,
fxi^i),would probablynot be applicable.
usual

196

Let

denote

us

filter (i.e.,
at the

by

it would

the

signalas
hence
signal,

or

x^, y^,

pointA

in

r^,

Oi,the sample

Fig.5). Let

Oi,

values

"

no

"

"

with
signal,

"

the

known.

Let

phase sample

there

is white

denote

us

values

Gaussian

the
for
the

by

"^j. The
noise and

is

2 WT,

W2

for

inputafter
sample values
The
envelopeof

the

noise.

no

assumed

""

fN(-,y) =\j;;^)
and

are

receiver

denote
"^j,

were

Z^,
amplitudesample
^n/2)the distribution function of
densityfunction for the inputat A when
probability
4"2'
-fs(^i5

the

for

values

bi,or/,

if there

pointA

at

appear

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

w/2

nj2

(110)

exp

signalplusnoise, it is

fsd-, y)

=[^) J/xp[--^2/-.cid'+l(yi
-

Expressedin

(r,6) sample values,Eq. (110) and

of the

terms

\n/2 w/2

bd'

dPsiaAl

(111)

Eq. (Ill) become

m/2

(112)
and
\n/2 w/2

fsNir,6)

2777V
1

exp

'"^
J

w/2

I [rl+fl

"

cos
(6,
2r,/;.

...,

from

factors

because

introduced

the

they are

Jacobian

"Pn,2)-

of the transformation

samplingplan to the r, d samplingplan [16].*


probability
densityfunction for r alone, i.e.,the densityfunction for the
the densityfunctions
for r
the detector,is obtained
simplyby integrating

the x, y
The

output of
and

Ilrj are

^,)]
(113)

dFsi4"i,
The

2N,

Q with

respectto

Q.
"277

fdr)

/"277

/'27;

Jo

Jo

Mri,Oddd,dd,---dd
(114)

or

I \w/2 n/2

1
"

27V

"/2

.fi

and

fsyi'-)
=

r r- f"Vs.v(^"
Gi)d6x dd,--"

Jo

Jo

"

dB

Jo

or

2 \n/2 1^ w/2

W./2

W2

/,,;;.

n/o(^K^('^i,"/'2---

/s'v('-)
^"""4=1

(115)

or

1 \w/2 w/2

For

example,in

two

/ff.

w/2

y) dx dy =f^{r,B)r dr dd.
dimensions, /^(a;,

W.

that

Notice
which

PETERSON,

the

probability
densityfor

the

all information

had;
"^,:

is

about

197

FOX

C.

W.

AND

BIRDSALL,

G.

T.

W,

tribution
completelyindependentof the dishas
been
the phase of the signals

lost.

ratio for

likelihood

The

input,r{t),is

video

nl2

(116)
A^

Again

it is

convenient

more

with

work

to

w/2

""

ratio. Thus,

the likelihood

logarithmof

the

/r f\

(118)
which

is

approximately
E

In /[KO]

"'^0

function

The

is nearlylinear for

and
be

In

[^

W\

-"

[r{t)f{t)'
'

(119)

dt.

\nl,
Jo

of

values

for small
the parabolaa'2/4
Iq{x)is approximately
for likelihood
largevalues of x. Thus, the expression

ratio

might

approximatedby
IHt)]

In

and
signals,

for small

(120)

by
ln/[KO]

where
largesignals,

for

I V{tmf{t)fdt

TTro

and

Ci

Cg

chosen

are

(121)

r{t)f{t)dt

approximateIn Iq best

to

in the desired

range.
The
Thus

the

correlator
the square

ratio,the
its

output

in Eqs. (120) and


(121) can
integrals
receiver
for
weak
signalsis
optimum
which

finds

the

correlation

cross

be
a

interpretedas

square law
the

between

detector, followed

detector

envelope of the expectedsignal.For the case


optimum receiver is a linear detector, followed by a
cross

of the

correlation

detector

output

of

of the

the

correlations.

cross

by

the

[f{i)T,

largesignalto

noise

has

for

amplitudeof

the

which

correlator

output and/(0,

and

expectedsignal.
The

distribution

function

approximation developed here


ratio,since this is the
for the
First

logarithmof

we

will

of most

case

largesignalto
shall find the

the likelihood

for

apply to
ratios
and

mean
as

shown

designedfor

the receiver

would

standard

be

easilyin

found

be

interest in detection

noise

ratio

/(r)cannot

studies.
even

deviation

easier

An
to

for the

this

case.

The

low

signalto noise
analogous approximation
derive.

distribution

of the

above,

^nm^-^^lff^^Jrftt

(122)

198

for the

of small

case

signalto

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

ratio. The

noise

probability
densityfunctions

for each

r,- are

(123)

and

The

these from the jointdistributions


^;v(^i)
and^^^Cr^)is used to distinguish
called
of each term
and
were
f^{r)
fsN^^)-The mean
previously

notation

of all the

ri which

in
rff^l4N^

the

in

sum

Eq. (122) is

(124)
or

(124)

Similarly,
"

The

second

of each

moment

f^SN

dn

zm

is
rff^jAN^

term

167V

or

r_m\

ji_ r"

^^f^lM\"in

"I67v0o
^'^""^
^^^'^167V^j
^

2N

(125)

Similarly,

N^^''^'''^'^'"
~\6N^]o
^""{leN^j
rift\

or

16N'

The

r^ "5

ft

00

for
integrals

the

case

16N

of noise alone

Ara^xp
Jo N^
.

can

be evaluated

dri
2N

:
easily

t^N

(126)

and

M-N

of signal
for the case
plusnoise can
integrals
which
for
turns
out
f
unction,
hypergeometric

The

be evaluated
the

cases

in terms

above

to

of the confluent

reduce

to

simple

200

READINGS

of the

Thus

MATHEMATICAL

there
(fi),

numbers

WT

IN

consecutive

are

PSYCHOLOGY

which

ones

are

not

These

zero.

givenby

are

=7^

U
where

is the

pulseenergy

in

pointA

at

(132)

Fig.5

in the absence

of noise.

For

this case,

Eq. (130) and Eq. (131) become


1 "2

/i^;vDn/(r)]
="

^,

=0,
/.^[ln/(r)]

(133)
""2 /

0|^[ln/(r)]=-

"-""'WI-mFjI'+m^
and

normal
distribution of In /(/-)
is approximately

The

for,by

the central
with

variables

limit theorem, the distribution of


distribution

common

large.The

becomes

actual

nearlynormal
are

than

The

its distribution

The

usingthe

noise

than^v('*i)operatingcharacteristic for the case


distribution as approximationto the true

normal

the

case

distributions

distribution then

have

leads to the

radar

that the

and
a

in

of times

signal
plusnoise is more

uniform

16

is

plottedin Fig.6

distribution.

In many

cases

approximatelythe
of Figs.2 and
curves

same

variance.

Assuming

3, with

(2Ef

case

This section deals with


assume

itselfany number

2E

as

be calculated

can

that
1

4.8

distribution

norma)

alone, since the distributions gsN^.^'i)

receiver

it will be found

normal

the

of noise alone

case

distribution of In /(r)for
with

than one,
larger
independentrandom

nearlynormal

more

In such

is much

of M

with
for Xh^gj^ixi)
integral

expressedin closed form.

be

can

if M

sum

approach

must

distribution for the

this case, since the convolution

givenrange. That is,we shall


rence
pulseswhose time of occurbe
assumed
The
known.
carrier phase will
to have
of
all
the
are
others,
i.e.,
independent
pulses
pulse

at
a radar target
detecting

if it occurs,
signal,

envelopeshape

are

distribution for each

consists

of

train of M

incoherent.
The

set

of

can
signals

be described

as

follows

M-l

^(0

2 /(^ +

where

the

^^^

cos

{(ot+ 0,.),

(136)

angles6^ have independentuniform

and
distributions,

the function

/,

W.

W.

PETERSON,

T.

G.

AND

BIRDSALL,

W.

201

FOX

C.

99.9

99.5

20

0.5

0.1
0.1

0.3 0.5

8 10

20

30
100

40

characteristic.
operating

envelope of

is the

which

Broad

70

60

90

95

F;,^/;
6

Figure

Receiver

50

band

singlepulse,has

receiver with

the

video

optimum

property

16.

design,M

that

IE

(137)

j\t +ir)f{t +jr)dt


Jo
where
The

is
(5,;
time

enough so
f(t)is also
*

because

the

Kronecker

delta

is

function, which

zero

if / ^

between

states

factor 2 appears

the total energy

function

that

The

/.
j, and unity if /
far
are
spaced
pulses

that the

pulses.Eq. (137)
The
and that the total signalenergy is E.
they are orthogonal,
as
as
assumed
loJItt.
no
to have
high
frequency components
is the interval

"

in

is M

(137) because /(O is the pulseenvelope; the factor


times

the energy

of

singlepulse.

appears

202

ratio

likelihood

The

MATHEMATICAL

IN

READINGS

be

can

obtained

PSYCHOLOGY

by applyingEq. (56).
r

"2

/(.i)

exp

dt

exp

A^n

Then

Jos(t)x(t)

(138)

dPsis)

'

No

or
r-lir

l{x)

exp

J'2tt

TVn

exp
The

be

M-l

Jo

N"

can
integral

Jo

0
T

?)i

+ 0,")cf/ dd,---dB,j_^.
f{t + tm)x{t) cos {(x"t

(139)

0'

evaluated,

in Section

as

4.5, yielding
M-\

l{x)

exp

No.

(140)

h
Q

where
"

fit + mT)x{t) cos

CO?

t/r

/(/ + mT)x{t) sin

\NoJo

ojt

dt

0 JO

(141)
This

of the

discussion

of the

case

with

quantityTq is connected
filter for the

signalknown
in

obtained

be

r"i could

each

identical with

is almost

quantityr"j

the first

described

pulse;it could

be obtained

by designing

the

fact,

section.
an

will be

(Nj2)rjN

delayof the filter. The


the pulses
which come
It is convenient

at

other
later.
to

instant

some

The

output
the

of time
differ

r^
quantities

have

{cot+

The

ideal

(142)

e)

puttingthe output through a

phase angleQ, and

of the

value

output

in that

In

signal
=

for any

appearedin

phase,Section 4.5.

manner

Soit) fit)cos

The

which

quantityr

for carrier

except
in the

receiver

the

r^which

only

in that

of the filter at time

receiver

calculate

for each

r,", and

the

linear detector.

is determined

they are

t^ +

nn

by

the time

with

associated

will be

logarithmof the

iNj2)rmlN.
likelihood

ratio.

Thus

In

the

Io(r"jN)must

be found

will

usuallybe

previoussection, r"jN
^(r"J^f
approximatedby x^l4. The quantities
As

in the

than

rather

detector

times

to, t^ +

consists of
detector

an

(for
We

T,

IF
the

shall

logarithmof the

linear

detector, and

tQ + (M

"

1)t

then

the
must

be added.
must
quantities
that
In
so
Iq(x)can be
enough
found
by usinga square law

these

small
can

be

outputs
be

of the

added.

at
square law detector
ideal system thus

The

passband matched to a singlepulse,*a square


device.
threshold
signalcase), and an integrating
find normal
approximationsfor the distribution functions of
likelihood ratio usingthe approximation
with
amplifier

its

In/nlT^'l^^,

law,

the

(144)

47V2
*

of the IF

It is usuallymost

amplifier.

convenient

to

make

the ideal filter(or an

approximationto it)a part

W.

T.

PETERSON,

for small

is vahd

which

W.

values

of

the

for the

distributions
individual

of the

discussion

same

known
signal

except

for

of

TT

{cut +

cos

distribution

the

as

W.

for

phase;

the

203

FOX

yields
(145)

"

0,")are

the

C.

(144) into (143)

follows

independent;this

are

+ wt)
pulsefunctions/(/'
is the

for each

Substitution

-77+1

quantities
r"j

AND

BIRDSALL,

r.,jN.*

In/^

The

G.

from

orthogonal.

quantity
r

which

the
The

tion
distribu-

appears

in the

both

analysisapplies

to

same

fact that

cases.

Thus, by Eq. (83)t

^"'F

IE

Pv|^'^")=exp
and

by (89),
1*00

NMr,
exp

SN

IE

AV

The

(146)

A^
be

can

^N

dy.,

obtained

Io(a)da.

oexp

NM

4E

(146) and
by differentiating

MN.
is\

This

is the

in the

same

\N

and

the

and

manner,

"

AE

as
situation,mathematically,
appeared

same

(147):

yVoM

2E

deviation

standard

(147)

\N

IE

exp

IE

densityfunctions

IE

exp

for the

mean

in the

(148)

previoussection.

logarithmof the likelihood

ratio

can

The

be found

they are

MNl
/)
/i^v(ln

0,

"2

a|^-(ln/)

2E

4(ln /)

and

If the
their means

distributions
and

previous section.
receiver
*

can

variances.
The

be

These

problem

operatingcharacteristic
See the footnote

t The
rather than

A/ appears
E.

below
in the

MNl

MM

MN^

assumed

normal, they are

formulas
is the
curves

(149)

1 +

are

same,
at

the

identical with

completelydetermined
the formulas

mathematically,and
end

of Section

4.7

the

discussion

apply to

by

(133) of the

both

and
cases.

equation (131).
equationsbecause
following

the energy

of

singlepulseis E'M

204

READINGS

4.9

Approximate

of an optimum

evaluation

PSYCHOLOGY

MATHEMATICAL

IN

receiver

remainingtwo cases, the assumption


the receiver operating
characteristic
be approxithat in these cases
is made
can
mated
of Figs.2 and
that the logarithmof the likelihood
ratio is
3, i.e.,
by the curves
This section discusses the approximationand
method
for
a
approximatelynormal.
of
the
and
the
2
3.
characteristic
receiver
to
curves
Figs.
operating
fitting
In order

approximateresults

obtain

to

can
By (68),Fg^(l)

the

Fgy{l). Hence,
variance

the

between
noise

the

ratio with

Probably

noise

is

alone

noise

a%,which

alone

is

a%, the second


1 -I- a'^.Thus

this number

is the variance

characterizes

of the

be

seen

if

unity, and

with

moment

the

likelihood

detect

to
ability

it can

of the distribution

l)th moment

"

signalplusnoise,is

with

mean

is equalto

means

alone.

any other

hence

is the {n
Fj^{l)

Furthermore,

of the likelihood ratio with

mean

of the likelihood

noise alone, and

if Fj^{l)
is known.

of the distribution

that the nih. moment

the

be calculated

for the

difiFerence
ratio

with

signalsbetter

than

singlenumber.

Suppose

logarithmof

the

the

ratio has

likelihood

normal

with

distribution

noise alone, i.e.,


1
=

"th

is the

and

mean

d the variance

of the likelihood

moment

ratio

'00

J
the

square

the likelihood

(x

exp

in the

has

made.

been

exponent and

The

usingthe
dx

exp

ratio. The

nif-

"

dx,

(151)

Id

VlndJ-oD

(150)

follows:

as

exp [nx] exp

'

logarithmof

/"oo

substitution

completingthe

of the

l^^dFyil)

where

dx.
2d

be found

can

nif

"

exp

Jmi
Jin

Vlird
where

(x

J'

^.v(/)

can
integral

be

evaluated

by

fact that

\/2tt d.

2d

(152)

Thus,
'n^d

fJ-.wil'')
exp

In

the
particular,

mean

of l(x),which

unity,is

be

must

mn

~d

l^dO
and

therefore

The

variance

of l(x)with noise alone

is

(154)

"

a^, and

therefore the

f^Nil') l/'NiOf+ o%(l)


=

and

this must

1 +

second

moment

of

l(x)is

o%(l),

(155)

agree with (152). It follows that

/*iv(^^)^
~

and

(153)

-|-m

exp

"n

"

6xp [2d +

2m]

(156)

exp [d].

therefore
^
The

distribution

of likelihood

ln(l

a%).

ratio with

signalplus noise

(157)
can

be

found

by

W.

W.

T.

PETERSON,

G.

AND

BIRDSALL,

W.

C.

205

FOX

applyingEq. (68). Thus


dFsdn

=IciFJl),

(158)

Fsx(l)
If

dF^{l) is

obtained

from

ldFj,{l).

Eq. (150) and

/ is

exp .r, then

by
replaced.

f'Oj

f'sNil)

Jlni

77-

dx

[.r]exp

exp

V2

(159)

f^

Vl-ndJlnl
Thus
with

^/2 and

mean

variance

abilityto

detect

likelihood

ratio has

that with

with

in both
The

the

are

the

of Section

cases

to

when

distribution

the

cases

approximate

results

the

has

could

followingsections for
characteristic curves
operating

Suppose
which

have

that

the

same

one

of M

the

set

plottedin Fig.2,
(160)

is the

distribution

which

4.8 this distribution

sample

pointsis

In

occurs.

to be the

is found

large.Certainlyin

it seems
Thus
reasonable
that
general form.
for
a%
a
by calculating
only
given case
is
signals

two
are

the
approximately

distribution.

normal

in the

of

is also the diflFerence

those

are

and

distribution

most

useful
and

obtained

ratio has

is

is d, which

curves

logarithm of

this distribution

ajfis given.The

cases

this
exactly,
Section

of the

Signal which

case

measures

a%).

of

this
be

ln(l

4.7, and

to detect
ability

4.10

if

in both

number

assumingthat the
likelihood

this

t"ythe equation

cry

4.6, Section

distribution
limiting

alone, then

operatingcharacteristic

signalknown

noise

ratio

If the

singlenumber.

with

Its variance

cases.

d related

of

case

signalplus noise,in

of the likelihood

oy

other

any

d
In the

is

there

completelydetermined

receiver

parameter

than

distribution

normal

plusnoise
signal

means.

when

d.

signalsbetter

In l(x)is normal

also

it is probable that the variance

In summary,

of the

of In / is normal

distribution

the

and

cases,

the

this

On

assertion
of

approximatedby those

as

same

if the

logarithm

basis, orv(/)is calculated


is made

Fig.2

with

that
d

"

the

receiver

ln(l

o^^).

signals
orthogonal
of

includes justM
expected signals

the
probability,

same

E, and

energy

S"

dt
s,(t)s,(t)

are

functions

5^(0, all of

orthogonal.That

is,

(161)

Ed,^.

Jo

Then

the likelihood

ratio

K^)

can

be found

^ 1
2 77

from

exp

be

Eq. (56) to

A^n

exp
z

(162)
I

/(.)=-I^exp
J^

where

s^t

are

the

sample

values

of the

function

"'^OJ

s^it).

206

READINGS

With

noise alone, each

PSYCHOLOGY

MATHEMATICAL

IN

of the form

term

has

normal

distribution

with

mean

and

zero

IE

_M

variance

"
~

Furthermore, the

different

N,0

quantities

X.^"-^^^
yy-l
i
are

independent,since

the

functions

follows

orthogonal.It

^^^COare

that

the

terms

exp
1

are

independent.
Since the logarithmof
Z

has

normal

each

distribution with

distribution

be found

can

WJi"^''"~A^oj
( "EJNq)

mean

from

term

exp

"

Eq. (152).

variance

and

lElN^, the

of the

is

"th moment

The

moments

n(n

fi.yiZ'') exp
=

It follows

that the

of each

mean

A^.
oj

and
is unity,

term

(163)

1)

is

variance

the

'2E'

a%{Z)
The
the

variance

of

of

sum

KZ^)

[^iiZf]

exp

independentrandom

1.

(164)

N, oJ

variables is the

sum

of the variances

of

Therefore

terms.

/2E]

G%iMl)
and

it follows

that the variance

exp I

of the likelihood

^-(^^4
curves

pointedout in
those
approximately

was

are

Section
of

IE

4.9, that

This

equation can

ln(l

be solved

a%)

for

(166)

the

receiver

operatingcharacteristic

Fig.2, with
1

(165)

is

ratio

exp

It

In

^-M

2E

M^^PliVo/J

(167)

IEJNq.

IE

ln[l + Mie'^

1)].

No
*

The

reasoningis the

same

as

that in Section

4.4.

(168)

208

The

READINGS

value of each

mean

MATHEMATICAL

IN

PSYCHOLOGY

is

term

'00

2E

exp

This

be evaluated

can

as

No'

N,oJ

Jo

page 174 of Threshold

on

d(x.

exp

Signals[5],and

(176)

the result is that

/."^)(i3)1.
=

The

second

of each

moment

is

term

-J.

/.",^)(|S2)=
fi^dF'-J^Hfi),
(177)

or

2E

"f

/4'^
(^')
The

exp

L 'No_\

be evaluated

can
integral

IE

in

as

Appendix

of Part

da.

exp

II of reference

[17],and

the

result is
IE
M
M'r(^')

h\TF

(178)

No

The

variance

of each

in

term

Eq. (172) is
IE

["rW)f
It follows

^^'^(^')

that the variance

of Ml

1/^''-'
m'

io\j^^

(179)

(180)

(181)

1.

is
"

(IE

4(M/)
and

therefore
IE

o\{l)

since the

variance

for the

of

sum

Nn
variables

independentrandom

is the

sum

of the

variances.
If the

characteristic

approximationdescribed in Section 4.9 is used,


those of Fig.2, with
curves
are
approximately
1

4.12

The
A

4.1.

Two

broad

band

ln(l

receiver

and

al)
the

the receiver

/2E

(182)

In

optimum

receiver

in
of the results of Section 4 are suggested
applications
further
from
examples of practical
knowledge obtainable

few

presentedin this section and

operating

Table
the

I, Section

theoryare

in the next.

in a frequencyband
of width
detecting
pulsesignals
B is to build a receiver which covers
this entire frequency
Such a receiver with a
band.
time is studied in Section 4.7. This is not a trulyoptimum
pulsesignalof known starting
it with an optimum receiver. We
to compare
receiver; it would be interesting
of a signal
have been unable to find the distribution of likelihood ratio for the case
which is a pulseof unknown
is distributed evenlyover
a
carrier phase if the frequency
that the frequencyis restricted
band.
so
However, if the problem is changed slightly,
One

common

method

of

W.

to

W.

T.

PETERSON,

G.

BIRDSALL,

AND

W.

C.

209

FOX

of the pulse width


pointsspacedapproximatelythe reciprocal
apart, then pulses
and the case
of the signalwhich
are
frequencies
approximately
orthogonal,
of M
known
be applied. Eq. (182)
one
orthogonal signals
except for phase can

at different

is

be

should
band

used

width.

width, the

IEJNq

as

with

Since

the band

parameter
function

equal to

width
used

of ^

are

the

of the

frequency band width B to the pulse


of its pulse
pulseis approximatelythe reciprocal
Curves
Section 4.7 also has this value.
showing

of

in

ratio
a

givenin Fig.7

for both

the

approximateoptimum

receiver

100

90
80
70

60

50

40
2"

No
30

20

10

14

12

16

18

d
Figure

Comparison

of

optimum

and

broad

band

receivers.

20

22

24

210

READINGS

and

the broad

Eq.

(135) and

4.13

for several values

receiver

band

(182),which

Eq.

Uncertainty and

PSYCHOLOGY

MATHEMATICAL

IN

hold

for

of M.

In the

largevalues

d is calculated
figure,

from

of M.

signaldetectability
where

the

signalconsidered

is one

of M

the
orthogonalsignals,
t
o
provides opportunity study the
In the approximateevaluation of the optieffect of uncertainty
on
signaldetectability.
mum
the signal
is one
of M
the
of
receiver when
ROC
functions,
curves
orthogonal
Figs.2 and 3 are used with the detection index d givenby
In the two

cases

signalis

uncertaintyof the

of M.

function

This

This

(IE

ln

be solved

equationcan

an

(167)

for the

signalenergy, yielding

IE
=

"

ln[l

Me^]

In M

\n{e^

1),

(175)

approximationholdingfor largeIE/Nq. From this equationit can be seen that


the detection index d,
the signal
of In M when
a linear function
energy is approximately
and
hence
the ability
detect
is
that
It might be suspected
to
signals, kept constant.
of the
2EINq is a linear function of the entropy, S ^^ Inp^,where /",:is the probability
ith signal.The
linear relation holds
when
all
the
The
only
expression
pi are equal.
which occurs
in this more
generalcase is :
the

"

2E
^

"

-ln(2pf)+ln(e^-l).

(176)

REFERENCES

[1] S.Goldman.
is devoted

[2] C.

to

^ew
Information
theory,
samplingplans.
Communication

E. Shannon.

York:

ChapterII,pp. 65-84,

Prentice-Hall,1953.

in the presence of noise.

Proc.

IRE, January,1949, 37,

10-21.

[3] U. Grenander.
Bd

nr

Stochastic

and statisticalinference. v^r^/j;/or Mor/zewor/^, 1950,

processes

17, p. 195.

[4] J. Neyman, and

E. S. Pearson.

problems of the most efficient tests of statistical


ofthe RoyalSocietyofLondon, 1933,231, Series A,

the

On

Transactions
hypotheses.Philosophical

289.
and

[5] J. L. Lawson
[6] P. M.

Woodward

G. E. Uhlenbeck.

Threshold

and

Information

Proc.

[7]

I. L. Davies.

On

I. L. Davies.

I.E.E.

New
signals.
theoryand

York:

McGraw-Hill,

inverse

1950.

communicati
probabilityin tele-

(London), March, 1952, 99, Part III,37-44.

determiningthe

presence

of

signalsin

noise.

Proc.

I.E.E. (London),

March, 1952, 99, Part III, 45-51.


[8] A. Wald.

analysis.New York:
Wiley, 1947.
Sequential
A unified description
of statistical methods
employing
Signaldetectability:
of
fixed and sequentialobservation
Defense
Electronic
Group,
University
processes.
Michigan, Technical Report No. 19 (unclassified).
ratio test.
[10] A. Wald and J. Wolfowitz.
Optimum character of the sequential
probability
[9] W.

C. Fox.

Ann.
*

Math.

Stat.,September,1948, 19, 326.

If lEjN^

"

3, the

error

is less than

10%.

W.

[11]

Reich

E.

and

Physics,
[12]

C.

R.

J.

V.

Y.

Conf.

1952,
North.

See

[15]

periodic
J.

W.

in

wave

C.

211

FOX

Gaussian

noise,

Journal

Applied

in

noise.

Applied

Journal

Physics,

January,

Signal-to-noise

October,

1950,
A

device

improvement

38,

1197.

for

computing

through

functions.

correlation

in

integration

Rev.

Sci.

Jr., and

J.

in

signals

Reintjes.

of

the

Wiesner.

B.

noise.

five

Applications

Proc.

channel

of

October,

I.R.E.,
electronic

correlation

1950,

analysis

38,

correlator.

analog

to

1165.
El.

Nat.

Proc.

8.
An

analysis
RCA

systems.
[5],
of

values

values

large

F.

reference
of

Graphs

sine

signals

sure

Meade.

E.

Cheatham,

carrier

also

J.

P.

and

D.

pulsed

of

Rogers.

F.

of

T.

Levin

O.

of

AND

347.

23,

Lee,

T.

I.R.E.,

and

1952,

J.

BIRDSALL,

detection

detection

and
Proc.

detection

M.

G.

289.

24,

the

On

Harting

W.

the

[14]

1953,

tube,

E.

The

Swerling.

Harrington

Instr.,

T.

PETERSON,

76-82.

25,

storage
A.

P.

March,
Davis.

1954,
[13]

W.

of

Rpt

(89)

integral
in

appear

which

determine

signal-noise

discrimination

in

1943.

PTR-6C,

206.

p.

the

factors

Laboratory

S.

with

along

approximate

Mathematical

Rice,

O.

for

expressions
of

analysis

random

and

small

noise,

B.

for

S.T.J.
,

1944-1945,
J.

Marcum

Project
[16]

P.

[17]

The

G.

Hoel.

of

Defense
Part

W.

be

may

noise,

theory

that

of

Journal

detection.

Trans.

noise.

in

Birdsall,

D.

of

No.

G.

the

this

Rand

function

have

been

Table

Corporation,

the

material

November

Applied

2,
p.

of

371,

26,

Physics,

signals

1953;

March,

'statistical

criteria

January,

for

in

this
the
D.

1953;

April,
of

signal

No.

of

criteria

Wiley,
reference

of

theory

Section

Statistical

24,

from

Report
in

detection

York:

New

drawn

The

Birdsall,

Physics,

PGIT-3,

Discussion

is

paper

Michigan,Technical

Optimum
35,

IRE.,

I, 11', Journal

of

statistics.

this

Middleton,

Applied
I

of

T.

contains

detection.

Report

G.

and

University
report

Technical

T.

and

Peterson,

found

of

report

mathematical

to

Sections

Group,
of

II

field

W.

Tables

compiled
of

by:

^-functions.

RM-399.

Introduction
of

46-156.

24,

unpublished

an

Report

material

Part

in

in

Rand

and

282-332

23,

noise,

D.

1954;
for

1954,

the

25,

paper.

Middleton,
detection
128-130.

Electronic

July,

Other

work

of

pulsed

Middleton,
M.I.T.

The

Lincoln

W.

of

1953.
in

this

carriers
statistical

Laboratory.
theory

W.

from

and

(unclassified),

detection

246.

p.

above

detectabiHty,

Statistical

Middleton,

D.

13

1947,
[9]

of

Peterson,

pulsed

carriers

signal
and
in

ASPECTS

FOUNDATIONAL

control
precise

be neither

OF

PATRICK

MEASUREMENT

SUPPES

It is

scientific

platitudethat there
ment.
predictionof phenomena without measure-

of measurement.

1. Definition
can

and

SCOTT

DANA

THEORIES

OF

nor

Disciplinesas diverse as cosmology and social psychology provide


evidence that it is nearly useless to have an exactly formulated
quantitative
of measurement
feasible methods
cannot
be developed
theory if empirically
for a substantial portion of the quantitative
concepts of the theory. Given
or
a psychologicalconcept like that of
a physical
concept like that of mass

strength,the pointof

habit

of

of

characteristic
a

of

consists

From

collection

an

of relations

additional relation,are
of the

measurement

The

major

must

be

shall

so

to

as

guarantee

Nevertheless

the

the

measure

this

specifiedobjects.For

physicalobjectsare

set; additional

easily

data, and

quantitative
yielda satisfactory
adequate theory of
reasonable

of mass,

measurement

that

basic set of

the

postulated have

often

ment
measure-

numerical

for instance,

objectsmeasured

acceptable numerical

empirically.
Conversely,as
be

we

of relations which

the structure

cannot

numerical

succinctlycharacterized

interpretation.
difficultywill

of

major
varietyof empiricalcontexts.
source

and

exact

an

of this paper,

desired

set of

utterly unsuitable

are

empiricalmeaning

sound

the structure

pretation.
technicallypracticalempirical inter-

relations

the

last section

in the

have

embarrassing consequence
infinite. Here

see

between

to

analyses of the

classical

but
interpretations,

have

to

objects.

have

yet also

and
interpretation

the

may

in providing an
difficulty

relations which

is to construct

have

needed

of the

masses

of

source

The

be used

corresponding to the concept. Why


abstract standpoint a set of empirical

example, data on the relative weights of


representedby an ordering relation on
fortiori an

laybare

empiricalphenomena

collection of relations?

data

is to

measurement

empiricalrelations which

collection of

theory of

be

carefully

point of the present


be grounded
analysesof measurement
may
paper is to show how foundational
in the generaltheory of models, and to indicate the kind of problems relevant
scrutinized

in

which

measurement

to

then

may

stated

be

The

here

not

main

(and perhaps answered)

in

precisemanner.
Received
^

We

clear

and

like

here

special cases

related, and

our

of the

indebteness
mathematical

presentation (see [7]).Although

our

Office

record

to

precise formulation

constitute

research

1957.

September 24,

would

we

was

of Naval

have

of the

made

supported

use

under

classes

arithmetical
of results and
Contract

our

NR

methods

to

Professor

Alfred

theory of models
theories
of

171-034,

the

the

Group

notions

article

are

fluenced
innot

closely
This

Psychology Branch,

appeared in J. symbolic Logic, 1958,23, 113-128.


212

do

theory of models.

Research.
This

whose

greatly

of measurement

Tarski,
from

Taxski
has

DANA

measurement,

want

we

To

involved.

notions

relational
(finitary)

"^, Ri,

The

give

to

relational

is to

is

of

sets

R^,

R^

.,

of
the

to

being

as

of the

of elements

form

called the

finitaryrelations

are

relational system 31 is called finiteif the set A

meaning

empirical data

set

non-empty

of theories

say, finite sequences

91, and

system

construction

with

treat

we

systems, that
^

213

SUPPES

PATRICK

precise set-theoretical

begin with,

R"y, where

.,

of the

domain
on

AND

turning to problems connected

Before

51

SCOTT

is finite ; otherwise,

are
infinite.It should be obvious from this definition that we
mainly
data.
think
of each
consideringqualitative
empirical
Intuitivelywe may
r
elation
R^ (an w^-ary relation,say) as representinga complete
particular
to a question asked
"no"
set of "yes" or
of every Wj-termed seanswers
quence
of objectsin A. The point of this paper is not to consider that aspect

analysisof

the
If
a

"Wi,

relational
.

is
i?,-

relation

the

all the

S""

Ri{a^,

.,

then

one,

31 is

and

is

Within

possibleto

are

framework

give

an

Although
isomorphism

in

s
a

Then

w"".

.,

onto

"%,...,

a^"

However,
with

image

outlines

of

domain

9'?is

s.

only if

.,

"

elements

S3
,

of elements

homoeach

of

/ is

A,
one-

isomorphic.

are

w, the

relation

in S3 if

numerical

that, for

function

R^

system
sub-

some

relational system

is the

Re

set

of all

relational system 3t with respect


function

which

of

imbeds

31 in 5Z.

theory

contexts

of

for

imbeddability
theories

homomorphism

isomorphic imbeddability

it is

now

of measurement.

theory are determined


by fixing a
relational
only considering
systems

homomorphism,
notion

of

relational system 91 of type

mathematical

the

of 3t. ^ A

positiveintegersand

than

of type

$8 is

such

31 is imheddahle

characterization

exact

n
.,

not

numerical

most

rather

both

determined

required to be one-one.
of the preceding formal
definitions

of

2.

S,-to ^.

system

finite sequence

Section

uniquely

-S and, for each

relational

general

connected

they

similar

are

.,

assignmentfor

First of all the

Next

numerical

assignment is

the

restrictive.

/ from

sequence

relational system whose

numerical

s.

"Wi,

function

homomorphic

numerical

of type

real numbers.
a

systems

a"^) if and only if Si(/(ai), f(a^)).If the


S3 is an isomorphicimage of 31,or simply 3t and

of S3 is

to

that

each

relational

is

subsystem of SS if ^

simply

if for each

avoiding of this ambiguity is not


relational systems 31
"yl,i?i,
R^} and

is the restriction of the relation

is

sequence

the

of type

are

for

rather

interpretations.
of positiveintegers,then

R"y is of type s

relational system

two

morphicimage of 31 if there
i=\,...,n

non-empty;

are

.,

of data, but

their numerical

positiveintegerssuch

Suppose that

"-B,5i,

type of

relations

worthwhile.

S3

sequence

that the

Notice

collection

relation. Two

w^-ary

of

actual

w-termed

an

"^, R^,

an

the

systems and

is

w""

relational system 31

if there is

with

connected

of measurement

and

the

of

is selected which

is defined

here
are

is

in

terms

this

measurement

used
facts

of

is too

actually closely

explained

in

detail

in

214

READINGS

corresponds

intended

the

to

PSYCHOLOGY

MATHEMATICAL

IN

numerical

imbeddable

relational

of
interpretation
in

the

'j!fl
are

theory,and

Moreover

the
permitted.
systems
only
all
relational
of
not
imbeddable
need
in
'^
concern
systems
type s
theory
subclass. Since it is reasonable
that no special
but only a distinguished
set of
we
require that the distinguishedsubclass be closed
objectsbe preferred,
of
under
isomorphism. We thus arrive at the following characterization
of measurement

theories

exists

finite sequence

system 9^ of t5^e

be

is

that

all relational systems in K

object that

may

have

of measurement

of
a

the definition

than

rather
linguistic

ordinarilythought

not

and

as

models

in character, since

linguistic
entity.To

natural

formalization

of every

be

sure,

in first-order

first-order axioms

that

infinite relational

one

and
cardinality,

infinite

by

theory

theories

many

predicatelogic
themselves

system

as

are

model

it is difficult to

then

they

how

be established between
numerical
models
can
any natural connection
Even
models
of arbitrary cardinality.
neglectingthis criticism first-

and
order

have

of theories of measurement

set-theoretical

identity.Notice, however,
adequate, for if they admit

with

of type

are

in SfJ.

readers

Some

such

imbeddable

should

is
theory of measurement
systems closed under isomorphism for which there
numerical
relational
a
s of positiveintegersand

of relational

class K

definite entities:

as

axioms

ordering

it has

theories which

will

Archimedean

wish

to

In fact,we

give

use

relational system is finite

and

these

metamathematics.

of

involved

On

in

the other

of the

some

hand,

do

we

linguisticquestions.
pointof departure for asking

reject any

we

set-theoretical definition

our

as

generalpropertieswould

more

immediately

be

that

or

linguisticdefinition

properties.Any

impressionthat

the

propertiesinvolving arbitrary

express

permitexpressionof

require extensive machinery


deepest problems of modem
not

to

numbers, for example, that

natural
an

adequate

not

are

see

as

just such questions.


On

definition

basis of the

the

questions naturally arise, to


first place,is a given class of
in the

And
it be

second

not
a

of which

devote

we

relational systems

place,given a theory

theory

of measurement,

adopted,
a

two

In

section.

the

of measurement?
in what

sense

can

of measurement.

simple counterexample

shows

that

every class of relational systems of a given type closed under isomorphism


Let O be the class of all relational systems of
theory of measurement.

type "2"
3

each

of measurement

axiomatized?

2. Existence

is

of theories

In

that

contexts

some

relative

are

to

'^. Notice

of measurement,

then

simple orderings.Let ",A,Ry


we

shall say

that
so

that

the

consequence

is every

the class of all systems imbeddable

subclass

class K
of this
of K

in members

is

be

system

theory of

of K

under

is also

isomorphism.
a

where

theory

of type

measurement

is that, if K

definition

closed

in O

is

theory

Moreover,

of measurement

216

READINGS

numerical

relations in various

covered

method

such
of all

morass

PSYCHOLOGY

MATHEMATICAL

IN

unifying
possiblenumerical

that

relational systems is the most


has

found.

been

yet

relational systems

the
among
few
of
are
very
in terms
of the

only
only those definable
From
an
empirical standpoint
a

computational value, indeed

any

portant
im-

But

most
sets
ordinary arithmetical notions.
find numerical
of qualitativedata can
interpretationby relations defined
of example we
in terms
of addition and ordering alone. By way
cite
may

the

of masses,

measurement

Frequently
probabilities.
also the

given in this paper

of

consideration

and subjective
intensities,

weighted averages

multiplicationof numbers.

of the

use

distances, sensation
the

shall restrict ourselves

we

However,

in the

to the notions

requires
examples

of addition

and

ordering.
No

natural

scientific situation

would

strictlyto require the

seem

of sets of infinite data. This state

of affairs suggests that theories

containing only finite relational systems would

of measurement

The

empiricalpurposes.

problem

sideration
con-

suffice for

is delicate,however, for the measurement

meteorologicalquantity such as temperature by an automatic recording


device is usually treated as continuous
both in its own
scale and in time.
lie in the correct
does not really
Yet the important problem of measurement
of such recordingdevices but rather in their initial calibration,a process
use
of qualitativedecisions. Because
of the
proceedingfrom a finite number
of

awkwardness
shall not

of the

applicationof

uniform

finite relational

systems,

we

this restriction.

generallymake

are
establishingthe existence of measurement
In
reference
to
concrete
recent
best motivated
a
a
example.
by
paper [4],
Luce
has introduced
a
generalizationof simple orderings which he calls

Further

remarks

semiorders.

satisfies the

about

semiorder

is

relational

followingaxioms

"^A P"

system

for all x, y, z,

51.

Not

52.

// xPy and

zPw, then either xPw

53.

// xPy and

zPx, then either wPy

of type

"2" which

A:

xPx.
or
or

zPy.
^

zPw.

in situations

where

objectsare to
two
be arranged in order and where
it is difficult to say exactly when
objectsare indifferent. For example, to say that xPy might be interpreted
as
higherthan the pitch
meaning that the pitch of the sound x is definitely
of y, or that the hue of color x is definitely
than the hue of color y,
brighter
Such

relations

are

most

likelyto

occur

greater than that of y, etc.


weight of the object x is noticeably
two
Indifference between
objects x and y (in symbols: xly) is defined as
that

or

not

the

xPy,

See

of those

and

not

[4], Section

yPx. The

2, p.

given by Luce.

181.

is that the relation / of

point of Luce's axioms

The

axioms

given here

are

actually

simplification

DANA

SCOTT

AND

PATRICK

217

SUPPES

always transitive,a fact easilyappreciated for each of


the intuitive interpretations
given above.
In his paper
Luce
interpretationfor certain
gives a certain numerical
show
that
kinds of semiorders, but he does not
particularclass of
any
indifference

is not

is

semiorders

of measurement

theory
not
are
interpretations
finite

the

the

case

only if X

situation

Clearly,if

y+1-

"

it is fair to say that

"Re, " " is

system

defined
and

exercise

simple

Further

followingresult:
The
class of finitesemiorders
numerical

real

are

the

by

condition

is

such

that

give

in

that
and

if

that

x"^

y,

is noticeably

better,x

prove

shall

we

theory of

or

y,

be

x^

numbers

to

his

relation. However,

relativelysimple. Let

becomes

semiorder.

fixed numerical

here, because

used

sense

is definitely
greaterthan

y. It is in fact

greaterthan

in the

relative to

real numbers

between

relation

then

the relational

the

of the

proof

relative to the

measurement

relational system "Re, ^".

presentingthe proof of the above,


in proofs of the existence of
general method

it would

Before

call the method


of type

of

Let

cosets.

"Wi, ...,m"".

31

"yl,R^,

which

measurement

i?"" be

.,

we

relational

shall

system

equivalence relation

determined

uniquely

be well to outline

is

1
n
xEy ifand only iffor each i
and each pair (^z-^^, z^y, (Wi,
w^^.y oim^-termed sequences of elements
1
of A, ifZj 7^ Wj implies{z^,
w^
[x,y] for /
w^, then Riiz^,
z^^)
and
if
only if Ri{w-^,
w^^).
Even
the
above
definition is complicated to state in general,the
though

introduced

into 91

the condition

by
.

.,

The

they
R^. ^

notion

of

xEy is simple : elements


perfectsubstitutes for

are

.,

the relation

the relations

where

.,

meaning of
E just when

weak

orderingcan

relation

serve

as

is connected

and

each

an

in the relation

y stand

other

with

example.

respect

Let

transitive.

and

9t

all

to

"^, i?"

Then

binary
xEy is
-equivalentto the condition : For all z A, xRz if and only ifyRz, and zRx
if and only if zRy. However, this simplifiesfinallyto: xRy and yRx.
to the general case,
define, for each x
A, [x] to be the
Returning now
the

class of all y such


of all

[x]ior X
permissibleto
x^^eA,
system

A.

define w-ary

relations

Rfiixj],
if
-...[x^^])
21*

It is at
is

"

xEy. [x]is called the coset


Directlyfrom the definition of

once

of

that

iA*,R*,
obvious

and

R*y

Rf

over

A*

x.

we

such

can

21* is

deduce

only if Ri{x-^,
...,x^).

is called
a

the reduction

homomorphic

be

the

class

that it is

that, for all Xj^,


The

image

'

The

authors

are

indebted

suggested

this

to

the

referee

for

general definition.

pointing

out

the

of 21 and

work

.,

cosets.

that

2t**

is the following:
is not quite obvious
isomorphicwith 21*. What
// S3 is a homomorphic image of 21,then 21* is a homomorphic image of

in [3] which

relational

of 2t by

that

^*

Let

33.

by Hailperin

218

By

proof,let /

of

way

that if

for

f{x)
simplicitythat

We

must

show

that

91 and

homomorphism
[x]
[y].Instead

if

from

that

show

g is

of type "2" and

are

91

wish

to

generalcase,

assume

"^, i^", $8

"B, 5".

to

have

that

therefore

[x] tor

g{f{x))
=

of 39 onto

shown
x

following relation between


subsystem : if 93 is a homomorphic

subsystem

function

any

of 91. For
B

from

of 91 to the

restriction

let

/ be

into

there

is

function

A. It is trivial to

verifythat

the

concepts of homomorphic
phic
image of 91,then 93 is isomorof 91 onto

homomorphism

such

of g

range

that

91*.

the

and

image

that

f{g{y))

for all y

35. Let g

of 91

yields the subsystem

B.

The

isomorphic

33.
the

Using

above

remarks

in

35 if and

91 is imheddahle

Further, it follows

that

of 91* onto

isomorphism

we

Let

function

subsystem

shown

have

(i)K
SfJif and

in addition

relative to

is closed under

be

remarked

satisfyexactly
notion

of

the

that for

K". In effect

without

of

one-one

subsystems, then if *
numerical
ments.
assign-

class of weak

proof in

the

relational system

orders, then
the

first

relational system 91,91 and


of first-order

logic not

identity,then

K*

is

always

involving the

satisfying

class of all systems

satisfyingthe

in addition

satisfyingthe axioms for K and


(*) If xEy, then x
y.
The applicationof this remark

is the

K*

paragraph
91*

is the class of all relational systems

if K

Hence,
identity.

tirst-order axioms

formulas

same

numerical

possessingonly

our

It should

images

isomorphism.

91* for 91

some

the formation

example again, if K is the


that
the class of simple orders. Notice
of this section is a specialcase
of (ii).
use

always an

above:

class of all systems in K

is the

91* in 35 is

of all homomorphic

isomorphic to

(ii)If K

equivalence:

in 35.

of 91*.

theory of measurement
only if K* is also.

is

imbedding

of 35, and

the

once

class of relational systems closed under

any

be the class of all systems

K*

To

be

now

at

only if 91* is imbeddahle


any

establish

can

of 91 this property is characteristic

we

of the

33. We

"

homomorphism

with

such

A*

onto

Notice

Let

of 91 onto

35

clearlysymmetric.We

be

be

f{x) f[y),then xEy, or in other words, for all 2: ^,


if and only if zRy. Assume
xRz. It follows
if and only if yRz, and zRx
hence
S
w
hich
that
a
nd
f{y) f{z),
implies
yRz. The argument
f{x)S f{z),

xRz

is

MATHEMATICAL

f{y),then

PSYCHOLOGY

IN

READINGS

axiom:

^=

to weak

simple orderings

orderings and

is left to the reader.


Consider
semiorders.
defined
of E

again
For

above.

the

any
In terms

xEy if and

of semiorders.

case

"^

P"

of /

S, consider

one

only forif all z

can
e

Let

the

establish

A, xlz if and

be

the class

relation

of all finite

of indifference

simplifiedcharacterization
only ifylz.

DANA

Introduce

(*) as

SCOTT

axiom

new

AND

S4. The

just the class S*. Notice that


orderingsand simpleorderings,the
of subsystems even
though S is.

unlike

is

For

semiorder

any

only iffor all z,

leave

We
a

weak

if

"^A,Py

and

reader

S*, then

"

is clearer

x^Pyi, and
Now

the

to

^A

has

an

Let

in

assignment

{xq,

.,

the

uniquely by
If

(1)

the formation

relation

and

weak

then

follows

as

xPz.

of the

fact that

xPy impliesxRy,

that, if xRx^^,

and

xPy.
fixed member

a.

". Under

"Re, "

sequence

then

a,-

of S*. We

the

XiRXi_-^and

wish

relation R, A

to

is

Xi_^. Define

x^ ^

rational

of

Uq, ...,""

show

that 9t

simply ordered.
of
by a course

numbers

determined

conditions:

followingtwo

xJxq,

closed under

with

that

x^} where

recursion

values

then

notices that

Py he

satisfyingS4

pleasant situation

further

.S

xEy if and only if xRy and yRx. Thus,


simple ordering of A. The connection between

a.

one

then

yiRy,

let SH

if

all 21

if yPz
zPy,
verification
elementary

and

is

of

class S* is not

if zPx
the

ordering of A,

is

the

introduce

"y4 P"

219

SUPPES

class

if and

xRy

PATRICK

i+1
i
If

(2)

and

xJXj

XiPXj_^where

0, then

^"

a^

a^ +

i-\-\
Notice

that

show

Passing
Let

such

have

always

while

that

the

we

"

choice
that

If k ^

have

aj_-^ and
then

the induction

such

k ",

course

It follows

that

a^_-^. \i k

ai "

flj_i+

to

that

must

implies i "
that

the

j. By

a^. "

"

0.

Clearly

(I ) this is obvious.
If Xi_j^IxQ,
then
1
XiPXj_-^^.
fl,_i "
in other words
or
Xi_j^IxQ,
Xi_^PxQ.
case

and Xi_-i^Pxj^_-^.
Xi_-i^Ix,^
By definition
^. there

is

no

then

problem. Assurhe

xJXj

xjx^.

show

1.

and
so
XjIXi_-^,
by our
it
induction
follows
hypothesis on i,
j, the requiredinequalityis obvious.

Similarly"i_i

a,_i, and

and

aj_i ^

hence

We
x^Px^^^^.
a,^.But

that, if a^ "

k. Assume
Let

hence

x^ be

by
the

way

a^ +

1 "

then

1, but

again, by

a^.j.

Let x^ be the
a^.-\-\.
^"1 " k, and, in view

a^ "

have

aj_i+
aj.+

"

a^ "

that, if x^Pxj^,then

prove

precedingargument,

Converselywe
of

not

1- If ?

hypothesis,a^. ^
step is

next

first element
of the

such

i. For

on

and xjx^,
XiRXi_-i^,
Xi_-i^RXj,

we

/" I,

The

0, /

of k

"
fl,.

and

assume

rcik-i

"

i. Now

"

cases:

1
aj^ -\

"

that

induction

xjx^

can

first element

i-\
-

by

ai_i
that

assume

1. Hence

two

are

0.

a^ ^

(2),

to

flj "

x^Ix^,there

first that a^ "

x,^ be

"!_!

in

every
first element

We

that

we

a^^^ +

(2) the hypothesis implies that / ^ i, while in the case


formula
for a^ simplifiesto a^.
Notice
further
flj._^-f z+1.
either under
element
(1) or (2);for lettingXj be the
x^ comes

i the

i-\-^

a^, whence

The
x^Px,,.

of contradiction

first element

such

that

that

a^ "

aj.+ 1

hypothesis
not

x^Px,..

xjx.;

then

220

READINGS

and

'^

then

0 ": a^ "

a,. "

We
aj^-\-\.

and

Uj. If

Uj^ ^

thus

xJxq

a^. "

1, which

which
cij^-\-\,

that

is

again
is

the argument

covered, and

0, then

0 ^

conclude

can

"

a^

and

on

of cosets,

Let

us

given

class, K

a^. ^

a,,
been

have

cases

that

a^. We

been

have

actually

that

proved

S*

is

and, by the general remarks


S is also

theory of

ment
measure-

in the infinite

also work

the steps in

The

upon.

Second, if the proof that K

existence

semiorders.

systems, the

numerical
the

ment
measure-

First, after

one

relational

nummcal

systems in K, and

theory of

as

marked,
re-

all the

orderingof

and

be

was

where

systems

of addition
a

of

relational system should

numerical

is

as

co.

the
establishing

of relational

say,

be decided

long

case

real

is not

measurement

at

tion.
obvious, the cardinalityof systems in K should be taken into considera-

once

restriction to countable

The

and
justified,

of K,

member

helpfulin

the

convenient
of

After

the

existence

is

Consider

relational

introduce

that

of interest:

Let H

For

and

such

that

xyM^uv

be the class of all such

for every

x, y, z, u,

Al.

// xyDzw

and

A2.

xyDzw

zwDxy.

or

v,

zwDuv,

and

given

then

We

or

method
for each

feasible

so

relational

present

an

In

is K.

subsystems.

established, there is

of type "4". For

one

system, what

example.
such

systems

if and

only if xyDyy. xyM^zw


if and only if there
zRy. xyM^+hw
xRy

uvM^zw.

A:

then

be

systems. This plan was

relational systems which


w

by the
assignments

plan : cosets

been

iA,Dy

the followingdefinitions:

only if xyDzw, zwDxy, yRz

exist u,

either

assignments?

systems %

in K

of measurement,

has

of measurement

finite

considers some
is to say, one
element
of K" is a subsystem

every

used

have

to

often

can

of cosets, it is sometimes

theory

class of all its numerical

if and

the reduced

Instead

such

could

we

if often

question which

axioms

If K'
.

of semiorders

case

is the

K'

in K'

system

some

the

subclass

empirically

restriction

of measurement

find numerical

only on

of semiorders.

case

relational system

trying to

concentrates

one

existence

imbedding by subsystems. That

consider

to

of

Then, instead

are

the

always seem

possiblewith

of each

the reduction

simplified
by
of cosets.

proof of

the

systems would

results

adequate

systems. Third,

we

it has

"Re, ^",

of
suggested naturallyby the structure
it is most
practicalto consider
be simply defined in terms
relations can
numbers.

All

examples simple orderings and

as

system should

a^ "

f{Xi)

well-orderingof type

summarize

now

using
is

is

ordering i?

the

as

that

Thus

proofwould

the above

that

inequahty

"Re, ^".

relative to
Notice

x^Rxj^.But

the

a,-fl, but

contradiction.

conclude

we

x,JXfj,because
contradicts

0. Now

"

such

relative to

of measurement
method

the

and

complete.

Finally define a function / on ^


% in "Re, ^".
that / imbeds
shown
theory

PSYCHOLOGY

MATHEMATICAL

IN

xyDuv.

satisfythe following

DANA

A3.

then

A5.

// xRy

A6.

There

is

A7.

//

xyDzw and
not uRy.

and

xRu,

not

A8.

and

These

axioms

between

relational

w.

not

xzDzy

and

then

there

xRy,
then

xRy,

that

there

are

zyDxy.
is

u,

such

and

between

Making heavy
H is a theory

that

system 9t

that zwDxu,

an

such

use

and

be formulated

y is not

of the

is the

only if x"y

"

last three

that

weak

and

yRx

case

the interval

existence

axioms, it

relative to the numerical

quaternary relation defined


for all x, y, z,

z"w

property of the

in first-order

in

greater than

of measurement

Archimedean

the

the relation i? is

in H,

interpretationof xyDzw

the intuitive

xy/!"izw
if and

cannot

for

"Re, A" where

system

be stressed

in A8

xzDuv.

that

such

not

interval

that

the condition
must

and

and

and

be shown

can

imply

the

is that

wzDyx.
yzDuv, then

zuDxy.

ordering of A,
wRz

221

SUPPES

PATRICK

then

and

// xyDzw

zuM'^vw

AND

xzDyw.

A4.

// xyDzw,
// xyDzw,

not

SCOTT

logic,because

by

Re.

It

ordering embodied
it

implies that

all

than the power


of the continuum.
systems in H* have cardinalitynot more
In addition, it can
be shown
that, if 31 is in H, and / and g are two numerical
assignments of 91 relative to "Re, A", then / and g are related by a positive

linear

transformation

that, for

all

^"

Re, f{x)

that

is, there

exist

/5e

a,

+ ^. This givesin
(xg{x)

with

Re

certain

sense

"

the

such

answer

numerical
one
assignment for 21, we
question above: If we know
all. Except for very specialsystems in H, nothing more
specific
be
can
really
expected.
Notice that all relational systems in H are
necessarilyinfinite. In the
in detail the theory of measurement
F
next
section we
shall consider
in "Re, A". Here
the
consistingof all finite relational systems imbeddable
situation is quite hopeless.There simply is no apparent general statement
be made
about
that can
the relation between
as
assignments. In as much
formation
imbeds
"Re, A" in itself is necessarilya linear transany function (p which
the

to

know

them

conversely,it follows that, if 91 is a system in F and / is an


is also an
assignment for 9t,then / composed with a linear transformation
with F is that two
assignment. The main
difficulty
assignments for the
same
system in F need not be related by a linear transformation.
and

3.

it is natural
Axiomatizability. Given a theory of measurement,
to
ask
various
questions about its axiomatizability,for the axiomatic
considerable
analysis of any mathematical
theory usually throws
light
characterithe structure
on
of the theory. In particular,
extrinsic
given an
^"

The

proofs

proofs of both
in

Suppes

and

these
Winet

facts

[6].

about

are

very

similar

to

the

corresponding

222

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

via a particularnumerical
relational
theory of measurement
characterization
system, it is quite desirable to have an intrinsic axiomatic
of the theory to be able better to recognizewhen
a relational system actually
of

zation

results
belongs to the theory. In view of the paucity of metamathematical
of higher-ordertheories, we
shall restrict ourselves
concerning the axiomatics
in firstto the problem of axiomatizing theories of measurement
order

logic.

It is

the

then

model,

infinite

result that, if

well-known

most

part

we

it has

of real numbers,

set

is

first-order axioms

is not

strong

Even

hardly

an

models

are

them,

That

asset.

of

To

of

set

such

remove

simply restrict the

we

restriction to finite cardinalities

difficult

questions.Thus for the


theories of measureshall consider only finitary
ment,

we

rather

some

is called axiomatizable,if there exists


of the

for

is to

given

i.e.,theories containing only finite relational systems. Such

(the axioms

one

cardinalities. Since

measurement.

understand

to

leads to

and

of this section

remainder

theory of

consideration.

cardinalities under

are

systems that
a

has

assignments with values in the

one-one

cardinalities

usually not
having

difficulties without

too

in

unbounded

class of all relational

say, the

of unbounded

models

interested

are

of first-order axioms

set

that

theory) such

of sentences

set

of first-order

finite relational

theory
logic

system is

in

the

in the set. A
theory if and only if the system satisfies all the sentences
theory is finitelyaxiomatizable if it has a finite set of axioms. A theory
versal
axiomatizable if it has a set of axioms
each of which is a uniis universally

normal

in prenex

a sentence
(i.e.,

sentence

form

with

only universal

quantifiers)
.

It should

observed, first,that

be

finitarytheory of

any

measurement

deeper than saying that in first-order logicwe


write down
can
a sentence
completelydescribingthe isomorphism type of
each finite relational system not in the given theory, and clearlythe negations
This

is axiomatizable.

of these

for

closed

relational

that

of whether

recursive

the

system

is

recursive

enumeration

imbeddable

the

as

in each

case

or

in

in

may

theories

sentences

problem

is

true
or

as

theory

universal

consistingof

tences,
sen-

all finite

relational system, then


is simply the

effective axiomatization

recursivelyenumerable

this last

be taken

given numerical

number

also that if the

Notice

considers

one

give an

effective method

clearlya continuum

are

the axioms

It is of

required set of axioms.


instance

measurement.

class of universal

that

relational

then

systems imbeddable

establish

not

serve

cannot

we

subsystems
conversely.In

problem of

relational

can

finitarytheories of

under
and

no

the axioms, since there

writing down

of distinct

the

sentences

quiteobvious

course

is

in

not.

the

given

It is not

problem

numerical

difficult to

equivalent to the problem of giving a

of all the relation types of finite relational systems


the

systems whose

given numerical
relations

are

relational

definable

system. For

in first-order

numerical

logicin

terms

224

under

closed

We

are

the

to

turn

now

numerical

the

relational

for

qualifyas candidates
the

copcerned with
of the

loudness

sounds

systems of type "4" imbeddable

with

would

other

and

wide

of loudness

be obtained

every

"Re, A".

system

of

sounds, the

by asking subjectsto

then

to

compare

the difference of

compare

pretations
pair of sounds with every other. More elaborate interrequired to obtain appropriate data on utilitydifferences for

in every
are

individuals

social groups

or

[6]).It may

Winet

case.

subjectivemeasurement

appropriateempiricaldata

be of

this choice

to make

they vacillate
chosen.

in their

be

y may

interest to mention

some

to choose

choice, and

estimated

only

from

the

interpretation
probabilistic

one

of

the relative

p^y

pairedcomparisons.

objects,but

between
are

they

will be chosen

frequency with
".

p^^

we

asked

are

situations in which

many

probabilityp^y that

of the form
inequalities

From

Siegel[2],Suppes and

scalingmethod

of times. There

number

and

Davidson, Suppes
(cf.

to the classical

closelyrelated
Subjects are asked

over

such

one

variety of sets of
In fact, all sets of psychological
data based upon
of sensation intensities or of differences in utility
membership in F. For example, in an experiment

empiricaldata are in F.
judgments of differences

each

proof for

relational
F be the class of all finitary

Let
in

theories of measurement
interesting
examples of finitary
tence.
subsystems which are not axiomatizable
by a universal sen-

there

sentence,

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

which
obtain

may

is

x
a

set

so

of

empirical data, that is, a finite relational system of type "4", which is a
for membership in F. The
intended
interpretationis that, if
sation
P^y " I and p^^ ^ |, then p^y ^ p^^ if and only if the difference in sencandidate

intensityor difference
that

than

closer

between

and

together than

by
system

axiomatize

to

the

on

F. Let

y is

equal

to

or

less

is closer to one-half

than

that of

over

w.

F is not axiomatizable
theory of measurement
relational
sentence, we
intuitivelyindicate for a

universal
elements

and

being, of course, that if x and y are


in the subjective scale, then the relative

frequencyof choice of x over y


Before formallyproving that
of ten

utilitybetween

the idea

w,

and

in

the

the kind
the

of

elements

ten

following diagram with

arises in any

which
difficulty

atomic

be

a^,

intervals

u^q ordered

attempt
as

shown

designations

given the

indicated.

I"1 I

I"3 I

"2

"4

I/^lI/^2I

/^3

i^4

(a^,a^),let /5be the interval (ag,a^,),and let y be


or
largerthan
/?.We suppose further that a^, a.^, a.^, a.^ is equal in size to
^^
but
is less than /3.
/Sg,
Pi,Pi,^3, respectively,
Let

be

the

interval

"

to

Essentiallythis example

show

that

was

first

particular set of axioms

given

in another

is defective.

context

by

Herman

Rubin

DANA

size

The
that

now

elements

is

225

SUPPES

intervals

remaining

is imbeddable

in

chosen

so

"Re, A", whereas

clearlynot.

using the criterion derived

and

be

may

from

Vaught's

prove

The

Theorem.

PATRICK

elements

Generalizingthis example
we

the

of nine

of ten

the full system

theorem

AND

relationships
among

subsystem

any

SCOTT

theory of

is not

measurement

axiomatizable

by

versal
uni-

sentence.

In order to

Proof.

sentence,

need

we

is in F but

show

to

% of type "4" such

apply

make

F.) To
as

specifyingthe
"i for

We

then

now

1,

set

a^

1,

I.

and

a^m

and

^^

for i

2*

a^

2.

we

have

two

Then

m"\

for k

be

may

atomic

1,

to

cases

to

consider

take

we

one

10.

easily described

by

We

for

and

m"\,

.,

2m

"

than

intervals.
"^+i

I elements

numerical

greater

or

most

a^+.+i"

compact,

disrupt exactly

{"!,
.

define

1,

a^+i

depending

.,

a^

the

1.

w"

In

2^^.

"

on

ag^} defined,

then

we

parity

".x,y, z, w} is not

and

if and

only

(1)

nine

let

b
a-^^,

permutations

(These nine

^^

a^/2

set

a^,

of "a,

"

b, c, dy

relation

of

have

a^,

as

a^,

the

y ^
""+!,

"

a^^. Then

we

12,

set

and

merical
nu-

flgm-

I^

a^^-^, a^m),

the

following

in D:

"a, c, d, by

"c, d, a, by

"J),
d, c,

"a, d, c, by

"c, d, b, ay

the

2m

expected

put

ib,d, a, c"

d"c.

2,

w.

"c, b, d, a}

"

if

"", b, d, c"

b"a

for i

the

a^^j

ib, a, d, c"
ay

2,

permutation of ".a-^,
a^,

some

we

the

for i

.,

permutations correspond exactly

followingfrom
of iA, Dy

a^/g

m"\.

define

now

except for permutations

ix, y, z.wyeD

Moreover,

relation
X, y, z,

^^

m"2.
Thus if w
1,3,
and/Sj
c(.(^^i)i^iori
/Sg,a.^
i^^,ag
^i, cx.^ jSg,(x^
p^. With
.

set

we

"

a^
,

is odd, and

i
I,3,
a-(rn+i-i)/2 ior
Then
is even,
and
m"\

is odd.

.,

have

even.

^^

.,

m"\
,

is

m"2

Case

4,

k elements
and

finite

2m"

integerequal

even

construct

we

m.

Case

4,

system

in its domain

subsystem of

definite

and

size of the

..,m"\

an

10

of 2m"

both

domain

a^,

numerical

we
fixingthe size of /S,-,

of

be

of numbers

selection

"i+i"

elements

every

subsystem

every

of the

elements

2m

that

construction

the

Let
relationship.

The

universal

finite relational

of 9t with

integer n

even

(A fortiori

is in F.

numbers

subsystem

51 of type "4" such

relational system

is in

axiomatizability
by

there is

% is not.

this end, for every

To

for every

that

that every

of

the criterion

All nine

are

needed

appropriate properties.)

to
to

the
make

strict
the

inequalities
subsystems

226

the choice

From

subsystem of

that every
Case

be

in the

The

2.

Case

2a.

(2)are

merely

in D

imbeddable

to

show

naturallyarise.

cases

or

a^^^

restricted to the

in "Re, A", but

a^ "

this situation

For

a^.

assignment the function

a^, a^,

subsystem. There

in the

not

is neither

omitted

element

element

the

is not

length equal
but by hypothesis
to

is a^, a^,

subsystem

not

in F;

not

a^m-

subsystem,
(1)

virtue of

by

of it.

subsystem

Case

up

is in F. Two

1 elements

"

omitted

element

subsystem

add

must

a^

the

and

is

The

2m

it is obvious

a^+j and a^^,


the interval {a^^-^,
a^.^.It remains

the nine permutationsof

Then

a^ and

(a^,a^ is less than

the interval

the definition of D

intervals between

of the atomic

sum

and

"Re, A", that is,that iA, Dy is

in

intervals between

for the atomic


to the

in A

of the numbers
imbeddable

is not

that iA, Dy

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

we

a^ra- Let

nor

a^+i

are

two

cases

may

use

for

a^

consider.

to

numerical

our

.,i"\,
1,
/ defined by /(a,_j) "t-i+l for /
to
It is straightforward but tedious
1. ...,n"i.
=

CLi+j for ?
that
/ is a numerical
verify

f{ai^j)

"

lation
assignment, that is, that it preserves the reand
D
to
observations
crucial
two
defined
are
as
(2).Only
by (1)
intervals (in the full system), if
this verification. First, regarding atomic

"i_,+i-a,_,.

(a^_j"1)
chosen

(^x,
y) 7^ {z,w)
from

the

the

defined,

as

the

1,

of

would
.

proof of

A,

z"w,

that

a^-\-\.

ai"

{z,w) is

and

then

not

numbers

no

so

it is clear

f(w). (Note that the

"

distinct

two

were

interval, and

Then

z"w.

f{y)^ f{z)

"

in A

atomic

an

x"y-\-2

f{x)

result that

the

non

atomic

intervals

size.)

same

2b.

"

x"y

definition

Case

for

implies the weaker

above
have

and

/(",_,+i)
-/(",_,") (a^.^+i-l)-

Second,
f{cik+i)"f{^k)-

^fc+i" aj.
that, if x, y, z,
=

then

k"i,

for

a^+i-a^

.,

be

i"\,

Here

expected

use

we

may
from
the

fia-i+j)^j+j+l
=

for

numerical

previous
/

1,

assignment /
f{ai_"j ai_j
by
case,
n"i.
This completes
=

.,

the theorem.

pleasant to report that we could prove a stronger result about


the theory of measurement
axiomatizable.
F, namely, that it is not finitely
for studying
of
tools
available
to
there
be
seems
a paucity
Unfortunately,
It would

be

such

questions for classes


state a conjecturewhich

of relational

to

if true

the

finite

are

closed under

would

systems. However,

provide one

axiomatizabilityof finitarytheories
say that two
satisfied by the same

submodels.

We

would

useful tool for

of measurement

sentences

we

are

like

studying

like F which

fmitelyequivalent

finite relational systems, and


only if they are
we
by a finitemodel
conjecture:// S is a sentence such that if it is satisfied
it is satisfied
by every submodel of the finitemodel, then there is a universal
sentence
finitelyequivalentto S. If this conjecture is true, it follows that
is finitely
submodels
closed under
finitarytheory of measurement
any
if and

by a universal sentence.
only if it is axiomatizable
proof (ordisproof)of this conjectureappears difficult. It easilyfollows

axiomatizable
The

if and

DANA

that,

sense

wider

sentences.
is

the

[7]

sense

and

Since

logical

by

logically equivalent
of

consequence

full set, it also

the

implies

submodels,

under

is

universal

class

of

set

of

it

because

equivalent

is thus

and

subset

sentences,

but

them;

in

universal

of universal

set

subset

finite

the

implies

this

to

finite

some

class

the

in

of relational

the

denumerable

wider

throughout

for

is true;

conjecture

is axiomatizable

is

removed

are

the

in

classes

(arithmetical)

restrictions

modified

227

SUPPES

PATRICK

universal

on

finitistic

thus

AND

satisfying S, being closed

systems
the

if
the

conjecture,

it

results

Tarski's

from

SCOTT

it.

to

Our

is

conjecture

its

pertinence

is not

we

should

like

to

special

area

in the

definable in
theory of

of

measurement

is

problem

Let
in

manner

all

for

negative

be

arise

which

relation

finitary

the

Is

finitely axiomatizable?

the

of

theory
the

that

shows

relations

quaternary

less than.

and

then

is true,
and

conclusion

binary numerical

in

imbeddable

models

In

typical of those
any

and

models

of

theory

measurement.

of plus

terms

systems

finite

about

of

problem

finitely axiomatizable

is not

an

general

theories

to

unsolved

of measurement.

conjecture

(If our

restricted

mention

elementary

an

the

concerning

one

in

definable

ment
measure-

of

terms

this

to

answer

plus and

less than.)

REFERENCES

[1]

G.

[2]

D.

this

R.

vol.

[6]

D.

19

W.
P.

[7]

A.

[8]

R.

and

and

PRINCETON

Hypothese

du

M.

An

Winet,

systems,

Econometrica,

theory of utility discrimination,

vol.

16
Remarks

Vaught,

vol.

UNIVERSITY

16

to

(1954), pp.
on

AND

vol.

the

STANFORD

582-588,

classes

589-591.

UNIVERSITY

of

I,

and
relational

1934,
based

on

192

the

pp.

notion

259-270.

(1955), pp.

of models,

theory

572-581,

Lwow

and

of utility

axiomatization

universal

(1954), pp.

Warsaw

continu,

science,

Contributions

Tarski,

mathematicae,

pp.

axiom

14-20.

(1954), pp.

utility differences. Management

mathematicae,

first-order

178-191.

SiERPiNSKi,
SuppES

in

121

1957,

Press),

description

experimental

An

making:

Decision

University

identity and

on

colloquium

Society

pp.

(Stanford

Semiorders

Luce,

(1956), pp.

24

[5]

Remarks

Hailperin,

283

S. Siegel,

and

California

Stanford,

T.

xiv

Mathematical

American

theory,

(1948),

Suppes

P.

Davidson,

Journal,

[4]
vol.

ed.

25, revised

approach,
[3]

Lattice

BiRKHOFF,

series, vol.

II,

vol.

III,
17

Indagationes

(1955),

systems,

pp.

56-64.

Indagationes

of

^lODELS

CHOICE-REACTION

FOR

Stone

jMervyn
medical

In

the

TIME

council*

research

two-choice

situation, the Wald


sequential probability ratio
variance
of the decision
and
applied to relate the mean
alternative
and
the ratio of
rates
times, for each
separately, to the error
the frequencies of presentation of the alternatives.
For
situations
involving
than
two
more
choices, a fixed sample decision procedure (selectionof the
alternative
with
highest likelihood) is examined, and the relation is found
between
the decision
time
(or size of sample), the error
rate, and the number
decision

is

procedure

of alternatives.

This
models

of

is made

no

is

and

with

data

is

that

need

designed with

to be

The

models

subject (S) is given


identify some
after
for

signal and
given

independent
The

present

models

dynamic

run

and

their

in

time, the

the

following

about

or

The

response.

time.

time

Thus

appeared

taken

the

in

may

choice

of

procedures, but
Also

no

(i)the

to avoid

parisons
com-

paucity

premature

orthogonal

at

of
a

for the

onset

change

stream

signal, a

into

rate

S. After

S's

reaches

Chaucer

recorded

Road,

1960, 25, 251-260.

up

signal
that

is,

mutually
with

time.

of three

Reprinted

certain

taking

time, S makes

will be

Cambridge,

of

stream

decision

time, the decision


is made

are

to

The

They will be hydro-

each

of

required

sequence;

not

the

reaction.

signals

do

which

presented with

is

of response.

time

228

is

random

response

15

signal and
S

to be

Unit,

in

appropriate

an

uniform

further

choice-reaction

Psychometrika,

mode

this

situation

of different

the

At

or

presentation

settled

quick

The

reasons:

the

to

form

attributes

After

Research

decision

pendices
ap-

that

hoped

in directions

ia made.

attributes

flows

"computer."

making

have.

be kept open

make

reaction

sense.

the

*Applied Psychology
article

signal

signal
input time, the front

mechanism

This

S has

in

to

in mind.

and

probabilitiesof

that

It is

for several

applying

as

the

successive

text.

he

confined

powerful discrimination,experiments will

most

of signals, the

assume

information

motor

the

in the

summarized

specificmodels

until

are

statistical

data

often

are

of the

details

reader

time-stationary stimulus

attribute

signal remains

data

field should

envisaged

are

the

mathematical

several

psychologically unreasonable.

the

interests;(iii)for the

our

appear

assist

experimental

means

rejections;(ii)published data
to

results

by analogy with

presented which

are

working

analysis of the

mainly

made

of available

The

presentation will

"calculated-observed"

model

point of usefulness

time.

only definitions

method

models

the

to

choice-reaction

for
and

this

develops

paper

called

the

components:
England.

with

permission.

MERVYN

the

the decision

input time, T", ;

apply

which

to Td

makes

Ti and

models

"

(the number

variables

will be related to the environmental

frequencies of presentation) and

their

the

Tm

are

It is assumed

that
that

commences;

information

the

the

paced
is supposed

there

that

patterns of information

may

"S's computer

which

variables

at

time

either

arise from
is

the

need

no

We

sources.

for

of

occurred) until

or

Sq

or

If there

Si

of
the
self-

is, some
is

that

variable

the

certainty
un-

at

has

variable
which

on

on

random

independent

random

the

information

(dependent

certainty
un-

added

noise

series of

each

no

The

that

will suppose

t and

random

Si

stream

statistical computer.

situation,from

external

intervals

(stationary)distribution

the

information; that

the

in

operates is equivalent to

short

signal (eitherSq

examining

preparatory warning signal is given. It

overlap

may

both

from

or

to

some

arise from

start

is "noisy" until
(This stream
it.)This assumption holds in the

some

there

in this sense,

input stage,

added

is

to

computer.

also when

and

when

knows

the

when

implied

it is not

Situation

Two-Choice

jar the

subject knows

is,he

signal is

condition

Models

the

arriving at

from

stream

Ratio

which

at

rate

By concentratingon T^ in this way,


necessarilyindependent of these factors.

Likelihood

say)

The

time, Tm

motor

incorrect responses.

that

has

] the

time, T^

signals and

of

229

STONE

the

signal

is made.

the response

Signal
Xi

Poix) and

Let

respectively.If

the

x's

of information

stream

auto-correlation
the

be

Pi(x)

x's

are

the

probabilitiesof

are

instantaneous

then

to

the

integralsof

the

stream

samples

over

not

the

stored

the
of

an

So and

less than

time

lags (or

at

zero

apart. If

intervals,then

the

least for

transforms

computer

Si

continuous

almost

an

successive

the

in

signal is

independence implies

for all time

t).Suppose

is then

quantity c(x) which

of

stream

auto-correlation
with

when

assumption

parts of the

between

assumption requires zero


those not small compared
x

X3

X2

each

adder.

SequentialCase
The
A

and

makes

computer
with

log B

the appropriate motor


the
been

total has
made

not

for Si

"

a
are

running

total of

preselectedso

c(xi),0(^2),

that

"

S decides

"

Constant

"
.

log

for So

(and makes
log B, provided

the total falls below


as
action) as soon
the decision would
exceeded
log A when
previously
facilitates
(The odd way of expressing the constants

have
later

230

READINGS

references.)If

IN

decision

the

is made

of the function

c(x)

S is familiar

with

different

that
situation

of

number
So

Si

distributions
then

[1]in

on

xi

X2

and

the

decisions

"

of

thought
S

by

let no
when

who

have

may

to

ceivable
con-

of

structure

of

(1) is
of the

averages

other

decision

incorrect

be

trained

are

cedure
pro-

response

this

possible that

is

signals presented

for any

It is

it is

optimality

the

and

exploratory,

as

the

be the

ni

learning,

task

However

from

The
,

averages

ni

of

process

equal probabilitiesof

n*

S,

to

Even

decision

information

The

that

initiation

T"

the

implies that

and

incorrect

of

Without

to

form

of

it

use

by

to

ways

variable

assume

called

the

much

T"

assumption

affect the

may

therefore

tion
assump-

is the

same

is made

for

T" and

deliberatelyshortened.

However,

Tj

in the

uncertainty

may

distribution

for So

(or of

be

without
about

for

(and

So
a

hold
a

reasonable
needed

to

the

n's,and

natively,
; alterit does

therefore

leading to Si) is

With

the

above

the

of

same

assumption,
of the correct

for

a comparison
comparison of those leading to Si).

test

about

of the

those

incorrect.

or

examining

"condition

The
.

affect Ta

result should

of

for So

that Ti +
possibility

and

is

requires only the

of Ti +

same

which

processed since the


by information
is
shown
1
it
Appendix
that, with mild

In

assumptions
of

it is T

influenced

correct

would

errors

be

decision

basis

the

time, Ti

Pi{x), the

same

decision

(The

exclude

action.

leading to

the

leading to

that
which

the distribution

wrong.

may

are

making

more

intervening

the

T's

provides

proportion

decisions

this

of

be

cannot
motor

is available

the computer

to

remembered

not

length

Td's,leading to

of

does

po{x) and

on

that

right or

long, T"

of the

restrictions

whether

is

be

test

trials

of Tj

presented
if Td is

assume

Consider

.) This

for Si

it must
so,

the value

correlated.

unwise

the

is found.

one

decision

are

fio and

"

Td

not

is,given
whether

think

n%

be

his computer.

on

for

appeal

not

be

etc., with

n*

followingassumption.

This

choice

probabilitydistributions

discrimination

deduced

following terms:

testing the model,

measured

the

theory

optimum

reward.

Before

be

the

optimality does
suitable

the

result of

may

optimal

can

necessary

then

Si

the

imposed

respectively.If n*o

and

So

the

samples

based
to

c(x)'suntil

and

by Wald

the

the

of the

of results. *S's computer

given knowledge

stated

that

nt. The

logpo(a:).

be
Po(x) and Pi(x).Such familiaritymay
S
has
trials
provided
performed many

the

sample T^

[1] shows

test

\ogp,(x)

implies that

function

trying out

nth

c(x) is

(1)
a

PSYCHOLOGY

the

at

sequential probability ratio

of the

Such

MATHEMATICAL

give

of the
a

Po(x)

model.

powerful
and

validity of

However,

fair

test.

Pi(x),it

the

model.

is difficult to
Since

is

an

operational definition,it would


clearly be
and
is
there
one
Pi(x).However,
Po{x)
tion,
assumpof

symmetry,"

which

in

some

discrimination

232

in
in

READINGS

IN

Appendix 3 that, provided lOe "


the followingrelation between
Ta
Td

The

J(e,e)

cc

PSYCHOLOGY

MATHEMATICAL

po "

and

e
,

lOe, the minimization

"

results

po

J(po,Po).

Kon-Sequeniial Fixed-Sample Case


If S has

incentive

then the advantage


quickly and correctly,
of the sequentialdecision procedure is that those discriminations which
by
chance
be
made
and
time
saved.
is
However
it
are
happen to
quickly
easy
is possiblethat S may
less efficient strategy which
is to
adopt a different,
an

to react

"

fix Tj for all trials at

Let the sample size


ratio

procedures are

decide

and

for Sj if

procedure based
procedures with
sequential case
on
po(x) and
making

case,

will

correspondingto
as

"

"

certain

procedures

be

n.

rate.

error

likelihood

The

c{x") " log C;


log Pi{x)
log po{x)
"

log C; c{x)
optimal in the sense

are

accepted

time

for So if c(xi)+

c(x") "

"

give

this decision

follows: decide

c{xi) +

0. These

"

which

value

"

"

"

that,if

other

any

used, there exists one of the likelihood ratio


remarkable
that in the
It was
error
probabilities.
useful predictionswere
obtainable
under
mild restrictions
piix).Unfortunately this does not hold for the fixed-sample
on

a;i

"

"

"

a;" is

smaller

difficult the

more

problem

of

testing whether

such

model

holds.
If there

input storage, it is possiblethat the results of the


imposed strategy just outlined are equivalent to those obtainable when
time Td
experimenter himself cuts off the signalsafter an exposure
this is the

is

of situation

type

of these

emphasis

self-

no

authors

considered

by

Peterson

is

and

Birdsall

mainly on the external parameters


supposed intervening variable. They

the
But
.

[2].The
(such as

define a
energy) rather than on any
situations for auditory discrimination in terms
of a parameter
set of phj^sical
normal
the means
of two
d, which is equivalent to the difference between
populations with unit variance. (For, in the cases
considered,it happens
that the logarithm of the likelihood ratio of the actual physical random
variables for the two
alternatives is normally distributed with equality of
This parameter sets a limit to the various
alternatives.)
discriminator
to So and Si)of any
using the
performances (errorprobabilities
bound
the
whole
of the physical information.
therefore
It
sets an
on
upper
performance of S who can only use less than the whole. In [2]the authors
S makes
make
the assumption that the information
the basis of which
on
likelihood
his discrimination
nevertheless gives normality of logarithmof the
ratio. They examine
data to "3e
whether
S is producing error
frequencies
that lie on a curve
defined hy a d greater than that in the external situation.
variance

For

under

the two

alternatives

variable

More

than

there

are

(which

may

be

Two

Alternatives

vening
probabilitydistributions for the interthat is,signals,- induces an
multivariate);
m

with

the

of

the consequences
where

pi(x) ior

distribution

probabUity

233

STONE

MERVYN

1,

"

fixed-sampledecision procedure

will consider

We

m.

"

"

based

on

.Xi

"

"

"

x"

is fixed.

signalsare presented independently with probabilities


Pm
Pi
the
when
the
of
and
is
to
if
signals.
q:.(D)
probability error
(adding to unity)
decision procedure D (based on a;i
.t")is used, then the probability
of error
to a singlepresentationis
If the

"

"

"

"

"

"

J2pM^)-

minimizing e is that which effectively


selects the
posteriorprobability.In this section, this
minimum
(or T^/t) and m when distributions are normal.
e will be related to n
it might be necessary
to supplement
in the validation of the model
However
Td with a time T^ representingthe time the computer requiresto examine
is the largest.For, although it
the m posteriorprobabilities
to decide which
It is shown

Appendix 4 that
signal with maximum

the

in

3D

might

be reasonable

expect Ta
that Ta

with

to vary

the

We

"

the

will state

relation between
take

Pi

x{l),

"

P2

"

inputs
is readilyseen

on

the

to

and

is held

is all-round

with

channel

of

x{l),

"

"

"

mean

a;(l),

"

x(m)

variance,
ances.
unit vari-

regarded

be

can

is stimulated

iih channel

The

unit
and

means

"

independent

are

0 and

zero
"

variable

random

x(m)

"

with

normal

in the

[3],who stated the


by the experimenter):we

under

Si

ing
signalcorrespondAppendix 5 that,

the

optimal procedure is to choose


the largesttotal. It is shown

that the

two

any

is constant

multivariate

with

symmetry.

similar channels.

suppose

Birdsall

constant

that

are

when

and

and

1/m

of

components

there

the

with

p"

to

larger.

Peterson

would

one

be

to compare

necessary

Thus

It

when

m
=

"

the other

while

"

'

between

x(m). Under
"."
suppose
x(i) is normally distributed

"

that

and

which

relation

and

is the

(treatedby

followingspecialcase

as

t' is the time

where
1)^',
(m
and decide
probabilities
=

for Ta would

simplest model

The

m.

Ti -f T^ is independent of w,

that

to suppose

in

this procedure,

nn'
for those

{1 +

for which

distribution

have

been

n(j.^
and

[0.64(m
e

results

nearly linear

are

(1/m). $"'
The

values

of

ni/

for certain

against log

m,

which

agrees

be

can

with

values

ized
standardof

and

Td is proportional to

then

is

^-\l/m)]'

is the inverse of the normal

independent of m,
plotted in Figure 1. It

If

e)

"

function.

calculated.

the

"

[$"'(1
l)"'''
-f-0.45]'}

that

seen

T^ is

very

experimental findings

some

in this field.
The

question may

be

raised
model.

whether

symmetry

condition

of the

to the

where

auditory signalis

case

an

Peterson

any

w-choice
and

task

Birdsall

presented in

one

can

obey
the

the

model

apply
equal periods

of four

234

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

LOGg^
Figure

of

an

(Ta) for Error

Time

Decision

The

present, but

expect the model

apply to
lightsarranged

discriminable

positional.However,
on

is

one

of which

the

to

low

in the

in

where

intensityvisual

positionaleffect

the
difficult,

would

to

of

one

not

fairlyeasily
be highly

display,for the noise would

some

case

superficially

upset it. We

of response

case

is

symmetry

case

would

difficulties of S

memory

any

Equally Likely Alternatives (m)

of

noise. In this

of S to "white"

exposure

1
Number

(e) and

Rate

lightsare patches
signalis superimposed so
the

not

may

be

important

and

noise

of white

that response
be
there may

symmetry.

Appendix

Let n,y be the sample size for a decision in favor of ","when


s, is presented.
by its characteristic function,
distribution of n,^ is completely determined

The

rPijFrom

A5.1

of

[1],if
"f"iii)
=

12Piix)[pi(x)/poix)Y,

then

a)JBVoo[-log0o(O] + aAVio[-log0o(O]

(6)

(1

(7)

/35Voi[-log0i(O] +

Now

^)AVn[-log"/.,(0]

1,

1,

F" defined in Appendix 2 are small. If a "


quantitiesEi
/8/(l
^)/a and B
(I
iS " 0.1 then to a good approximation A
and
in
t
1
+
(7),
w
(6)
+
w)
"/"o(1
^i(w);so, putting

provided
and

the

0.1

/35Voo[-log0i(i^)]+
(1

"

(1

^)AV,o[-log"^i(w)]

a)SVo:[-log0o(w)] + aAVn[-log"^o(w)]

1,

"

a).

MERVYN

By

comparing these equations with

and

^10

of

\pn

and

nio

(7), it is found that i^oo 'A"i


of noo and noi (and similarly those

(6) and

the distributions

Therefore
.

uu)

identical.

are

Appendix
In the

235

STONE

of symmetry,

case

X Po{x)log [po{x)/pi(x)] J2 Pi(x)log [pi(x)/Po(x)]

E,

and
var

log [Pq{x)/p\{x)]under
oi[l],\i E and

A:12

From

paix)

(8)

no

V.

small,

are

J{a, ^)/E;

Pi{x)

log [Pi{x)/pQ{x)]under

var

n^

J(/3,a)/E.

Therefore

/((8,a)/J (a,/3).

ni/flo
=

By

with

differentiating(6) twice

(8) and

the fact that

a(l

VJ{oc,/3)
'"

t and

to

respect

lAyis the characteristic

substituting t

function

oc)[^n\ (ni

0, using

of riij
,

n,y]

E'

{I

"y

-a-

By symmetry
^(1

F/(/3,g)
""'

^)ml

(no

fiO^]

E'

a-

py

a-

Hence

J{a, P)vi

/(/3,a)yo

a)[4n? (fii no)2]


niy]]/(l
J(", /3)/3(l |3)[4n^ (no

a)"(l
{,1(13,

^y

Appendix 3
If

"

Keeping
lOe

"

po

the usual

"

and

0.1

[or po
1

^
+

"

"

(1

then, by (8), Ta

0.1

Po)/3]constant

"

lOe, the condition

"

methods

on

that the minimum

at

set of

be

the

for which

set
a

of all

decision

value

in the

a).
po)/(i3,

range

Td is proportional to J{e, e)

"

by

J{po, Po)-

possiblevalues
is made

given by

/3will be satisfied. It is found

and

Appendix
Let

(1

p(J{a, /3)+

oc

for Si

of

a:

(xi

"

"

"

Xn) and

Xi the

Then
.

TO

"

Z) Pi
1

Suppose Xi and

Xj have

common

S
xfX

"

p.(^)Xi

boundary; then, for

to be

minimum,

236

READINGS

it will not

IN

changed by

be

PSYCHOLOGY

MATHEMATICAL

small

displacements in this boundary. Hence, on


the boundary, PiPi{x)
PiPi(x); that is, the posterior probability of s"
equals that of s,- Consideringall possibleboundaries,the solution is that
=

is the

Xi

other

set of x's for which

s,- has

than
greater posteriorprobability

the

signals.
Appendix

Write

^(*)
under

Then,

Si

1] xXi)/n.

\/nx{l)is A^(\/n/i,
1) and

\/nx{i)is A^(0, 1) for

i t^

1.

Therefore,
ai(2D)

"

arrk!^

"

"

On

mi)r-'

[-Kw

exp

aA fjif]
du.

integrationby parts,

(9)

Z pM^)

(w

f ^ur-'^(u
l)(27r)-'''

6
say, where
their tabulation.
"

while
The

ei{6)

same

that oiv

as

-\-w, where

of

not

are

addition

of

to

determine
"

will

the

and

"

as

as

the basis of
5

-^

oo

"

for 6.

the distribution

to

(wj

max

"

"

"

variables.

of

of 6 turns

v^^i)and

v,Vi

out
"

Referring to Graph
second

(9),e"(0)

moment

distribution.

normal

and

of Peterson

o-^as follows. From

e"(^)

"

"probabilitydensity function"

those

d is

"

"

be

y"_i

4.2.2(7)
quotients
the

Also

approximately normal,
we
N(v, o-^),
(1/m). Also e"(0)

Birdsall. If 6 is
=

"

Therefore
(" v/cr).
=

0-^

[0.64(m

-^

and

oo

improve normality. Hence

calculations

(-iw') du

exp

this form

[3]use

20, the first and

"

v/aAlso

different from

very

agreeing with

normal

as

hence

and

m
independent
[4],it can be seen that,for

of

"

"

is
\ei{d)\

standard

are

Birdsall

and

e"(0)

However

0. Therefore

"

Vnix)

"00

-s/nn.
Peterson

characteristic function

the

'^^

(27r)-

the constant

nti"
=

var

var

1)~*+ 0.45]'for
error

{1 +

and

w
m

"

-$"'(l/m).
from

w
4.2.2(6) of [4], var
o"^
determines
Putting e"{d)

Graph

20, which

rate,

[0.64(m

ly'''+ 0.45f}[$-'(! e)
-

*-'(l/m)f.

e,

MERVYN

237

STONE

REFERENCES

[1]

Wald,

[2]

Peterson,
No.

[3]

13,

Gumbel,

Revised

among

E.

Birdsall,

and

York:
T.

Group,

Birdsall,

alternatives.

T.

Wiley,

G.

The

Univ.
G.

The

Quarterly

1947.

signal

of

theory
Michigan,

of

probability
Prog.

detectai^lity.

Tech.

J.

Statistics

of

extremes.

10/^6/69
received

1/4/60

New

York:

Rep.

1953.

Rep.

No.

decision

correct

10,

Electronic

Defense

1954.

received

manuscript

and

Defense

W.

W.

Michigan,

Univ.

Manuscript

W.

Electronic

Peterson,
choice

[4]

W.

New

analysis,

Sequential

A.

Columbia

Univ.

Press,

1958.

in

forced

Group,

I
STATISTICAL

INFERENCE

T.

W.

Anderson

Columbia

obtained

are

order when

there

A.

Goodman^

Universityof Chicago

likelihood estimates

for the transition

CHAINS

MARKOV

Leo

and

Universityand

Maximum

Summary.

ABOUT

and

in
probabilities

asymptotic distribution
chain of arbitrary

their

Markov

repeated observations of the chain. Likelihood ratio tests


and x^-testsof the form used in contingency tables are obtained for testingthe
of a first order chain
followinghypotheses:(a) that the transition probabilities
are
are
constant, (b) that in case the transition probabilities
constant, they are
that
the
chain against
is a wth order Markov
specified
numbers, and (c)
process
the alternative it is rth but not wth order. In case
0 and r
u
1, case (c)
results in tests of the null hypothesisthat observations at successive time points
are
independent againstthe alternate hypothesisthat observations
statistically
chain. Tests of several other hypotheses are also
from a firstorder Markov
are
of a long
considered. The statistical analysisin the case of a singleobservation
are

"

chain

is also discussed. There


ratio criteria and

1. Introduction.
for certain time

which

is

discussion of the relation between

some

x'-tests of the form


Markov

used in contingencytables.

chain is sometimes

series in which

hood
likeli-

the observation

suitable probabilitymodel

at

given time

is the category

individual falls. The

chain is that in which


simplest Markov
of equiand
finite number
a
categories
distant
time points at which observations
are
made, the chain is of first-order,
the same
for each time interval. Such a
and the transition probabilities
are
chain is described by the initial state and the set of transition probabilities;
mediately
namely, the conditional probabilityof going into each state, given the imWe
inference
shall
of
statistical
state.
consider
methods
preceding
into

there

are

an

for this model


and

the

an

when

there

are

set of transition

same

to estimate

of states

finite number

or test
probabilities
asymptotic theory for these methods

increases.

We

for each

time

The

RAND

of inference

the transition

use

of

some

when

wish
velop
de-

the number

of inference

need
probabilities

of the statisticalmethods

given in detail [2].The data


study" on vote intention. Preceding
number
of potentialvoters
asked
was

This work

initial states

example, one may


hypotheses about them. We

shall also consider methods

has been

'

of the

for
not

of

more

be the

interval.

illustration of the

Received

in each

probabilities
operate. For

generalmodels, for example, where


An

observations

many

the transition

observations

same

or

for this illustration


the

1940

came

described herein
from

election
presidential

his party

or

candidate

Math.

Stat.,1957, 28, 89-110.


241

of

preferenceeach

August 29, 1955; revised October 18, 1956.


carried out under the sponsorship of the Social Science Research
was
Corporation, and the Statistics Branch, Office of Naval Research.

This article appeared in Ann.

"panel

each

Council,

Reprinted with permission.

242

month

from

May

(6 interviews)

October

to

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

At

each

interview

intention

at

constant

over

It

and

This

to

both

use

Some
in

the

herein

present

paper.

for

that,

The
of

Hoel

[10] in the

they consider
increases. We
a
a

and

of view
the

of

by

present

and

further

ordered

these methods

of state

"

"

as

the

be

time

and

states

so

discuss

Wi(0)

the

case

as

Bartlett

by

is observed;

time

points
where
that

is

likelihood ratio

the

to

ratio

are

t, given

likelihood

ratio

1, 2,

is referred to

Markov

first-order

tests

"

1 to m,

1,

i at

state

i is

the

Though
actual

no

of

is made

use

"

"

"

time

m;
^

"

1,

"

"

"

T)

are

the probability

he

with

shall deal both

We

1
.

for each

ni(0) individuals
they

[7].

chain.

-jm.

"

[2] or

a political
party, a
(a, b), etc. Let the times of observation

state

same

though they

where

x^-tests,

ordinary contingency

some

between

relation

and

criteria

that i might be, for example,

Pij(t){i,j

with (b) nonstationary


probabihties need not be the
there

ing
test-

hypothesis

pij

ior t

transition probabilities(that is,where

the

of the

of states
of

from

treat

for

given

(a) stationary transition probabilities(that is,Pij{t)

section that

their

5 of the present paper,

likelihood

integer running from

an

at

studied

number

chains, the reader


of

T. Let

"

presented

final section.

of Markov

arrangement,

0, 1,

=^

the

is

related

are

geographical place, a pair of numbers


he

given

and

motivation

been

considered

was

of the

parameters

Let

usually thought of
this

has

as

tables

present both

discussion

of the

model.

also

[2] are

single sequence

theory

contingency

we

discussion

2. Estimation
2.1. The

only

asymptotic
in

in the

x^-tests appears

explains

potheses
general hy-

more

[10].

paper,

procedures.

For

chain

of the

where

hypothesis that

how

it is shown

table

order

the

the

Hoel

given

[2],and

and
transition probabilities,

of the

situation

used

form

generahzation
In

their

methods,

shall discuss this situation in Section

x"-test of the

point

some

these

hypothesis.

methods

new

pothesis
hy-

first mentioned

were

[1] and

in

the

was

for all

simpler.
estimation

of fit and

goodness

of

this

for dealing Avith

of

the im-

at

the methods

in [1] and

given

or

this null

to

differed from

appearing
of

advantage
users

many

of the

problem

[3] and

An

to be

application seem

methods

new

party

probabiUty

the theory and

those

the probability of

conformed

methods, which

newer

residual

probabilitieshold

same

the data

of formulas

corrections

is

some

old and

the

the

extends

different from

somewhat

on

the data

how

see

develops and

paper

are

to

that

specificways

[2].It also presents

in [9],that

how

also in what

present

in [1] and

of interest

was

only

that

on

his intention

(first-ordercase), that such

and
(stationarity)
,

time

individuals.

one

interview

mediatelj''preceding

depended

interview

decided

was

person

the latter being

Democrat, or
Know,"
of
who
had not
primarily
people
consisting
category
candidate. One of the null hypotheses in the study was
voter's

each

"Don't

classified as RepubUcan,

were

time

in state

i Sit t

nonrandom,

random

interval).We
=

while

variables. An

0.

In

I,

"

"

"

T)

the

transition

assume

in this

this

in Section

observation

section,we
4,
on

we
a

shall

given

T.

iiO),i(\),i{2),

T, namely

"

These

possiblesequences.

transition

the

nij{t)denote
show

shall

of rnT

form

the

is

sum

probability, in

for the

individuals

be

i "t
t

m;

"

and

"

1,

"

"

T),

"

observed

for the

whose

t.
set

sequences.

of states

sequence

at
a

is

2(0),

i(t

1)

"

and

dimensions), of

nT

are

i(t)

j. The

for all

all sequences

given ordered

is

"

"

"

"

Pi,T-.u"rATr''-'''''''"

"

\i(T-l).i(T)

^^p."(0"-"'^
products

the

in the

0,

"

"

of

distribution
factorials.

of

"

"

lines

first two

nij{t)form

of numbers

set

actual

Let

1)

"

"

"

set

all values

over

of sufficient

of the

"

,m,

1)

2J7=i nij{t).Then

given ni(t

"

-{- 1 indices.

announced.
as
statistics,

nij(t)is (2.3) multiplied by

the

ni(t

1,

are

appropriate

an

the

conditional

1) (or given nk{s),k

1,

"

"

"

is

'^4^^^^-Updtr'''.

is the
on

numbers

(2.5)

"

PminAir'"''''''')
( n

( U

the

tions

pm-DUD

"""

(2.4)

This

"

describing

space

there

initial state

distribution of nij{t),
j
m;

abilities
prob-

should

(n[p.-(.-i).-(r)(r)r'"^^"''--(nb.(o)m)(i)]"^'"^^''^-"^"0

probabilities

transition

the

form

in state

z's with

of the

\i(0),i{l)

are

dimensional

nmT

"

function

"

m'"

^
[p.(0).-a)(l)
p.-(i)i(2)(2)p.(r-i)i(r)(r)r^-""'^"'-"

The

"

Z)n,-(0),(i)...,(7-)

all values

over

the

of sequences

Thus,

of the

1,

of individuals

the number

(for each

individuals

where

stationary. (When

nij(t){i,j

n,y(0

(2.3)

with

events

"

i(0),there

initial state

of sufficient statistics

set

(2.2)

set

i(T). Then

"

where

0,

PHT-l)i(T)

"

of individuals

of

the set

ni(o)i(i)...i(r) be
"

"

is in at "

throughout.)

the number

that

numbers,

"

"

probabilitiesare

replaced by pm-Diwit)
Let

the

necessarilystationary,symbols

not

are

i(l),

243

GOODMAN

individual

the

i{T). Given

"

Pi{0)ia) PiWHi)

when

Let

of states

A.

represent mutually exclusive

(2.1)

We

"

LEO

AND

of the sequence

consists

individual

ANDERSON

W.

same

distribution

one

distribution

multinomial

nij(t).The

as

distribution

would
with

of the

obtain

if

one

had

probabilitiesPij(t)and

nij(t)(conditionalon

'"4^^^UpAtr''''

n ndt)
L

J=i

ni{t

J=l

"

1)

with

obsei-va-

resulting

the ^^(0))is

244

READINGS

For

with

chain

Markov

IN

concerning
form

of

set

from

nnp;r''

(2.6)

necessarily stationary

not

minimal

Maximum

likelihood

estimated

be

can

subject of

transition

estimates.

The

the

are

Uij

actual

factor

does

This

depend

not

{i

pi,

on

1, 2,

probabilitiespij (i,j

and

the

easily verified that

that

as

"

1,2,

"

,w,

obtained

m) consists of

"

1, 2,

"""

precisely of the

is

probability

the ith. sample

trials with

samples, it is well-known
for pij

that

1,

to

respect

0 and

observations.

samples, where

multinomial
zli'i^ii

estimates

probabilities

stationary transition

probability (2.6) with

the restrictions pa

to

course

independent

such

probabilities Vai^), the nij{t) are

the

by maximizing

form, except for

same

nf

written

be

can

the

npr.

T^pij
the

when

^J=i Uijit)

that, when

i,j

(2.7)

for

n,;

fact

of sufficient statistics.

set

2.2.

Pij

set

the

probability (2.3)

o,i

'=1

Pij

from

result

form

in the

For

follows

stationary, the

are

the

(2.3); namely,

sufficient statistics. This

probabiHties

transition

PSYCHOLOGY

stationarytransition probabilities,a stronger

sufficiencyfollows

MATHEMATICAL

"

"

"

m).

For

likelihood

the maximum

are
T

pij

riij/n* X) nij(t)/^

53 w*(0

A:=l

(=1

/=1

(2-8)

and

hence

this is also

probabilityis of the
the Pij

on

the parameters

pij

approach used
likelihood

in the

the

are

in

estimates

form

same

other

except

in which

the elementary

for parameter-free factors, and

particular,it applies

In

same.

distribution

the

strictions
re-

of

the estimation

to

(2.6).

the transition

When

for any

true

probabilitiesare

preceding paragraph

applied,and the maximum

stillbe

can

found

for the Pij{t)are

necessarily stationary, the general

not

to

be

/mZ

nikit).

k=l

The

same

consider

the conditional

distribution
are

the
on

numbers

likelihood

maximum

same
a

as

one

multinomial

nij{t).

would

obtain

"

"

"

nij{T)

if for each

distribution

with

is used.
i and

obtained

Pait) are

of nij{t)given ni(t

distribution

of the niy(l),nij(2),

for the

estimates

one

1)

"

as

Formally
had

"

we

the joint

when
these

n,(i

probabilitiesPij{t)and

when

estimates

tions
1 ) observa-

with

resulting

estimates

The

can

estimate

to

tables

way

ptj

for

for

The
Yltriijit).

the

1,

"

"

by the

following
m

is the

of pa

of

estimate

is the

pait)

entries in the ith

of the

sum

entries nij{t)

the

Let

way:

The

table.

order

In

row.

i,

the corresponding entries in the two-

T, obtaining

"

of the

entries

with

table

two-way

i,jth entry

table

given further

n,j

of n,y'sdivided

by

find the

shall

n,y(i).We

the

presented

estimates

in

on.

nij(t).To

of

behavior

likelihood

maximum

of the

structure

consider

first

the

245

GOODMAN

A.

entries in the ith. row.

Asymptotic

2.3.
Pij

will be

this section

in

two-way

LEO

AND

stationary chain, add

covariance

The

described

estimate

of the

sum

be

divided

table

in the

jih.entry

W.

in

for given t be entered

ANDERSON

T.

nk{0)/2l wy(0)

that

assume

the

of

behavior

asymptotic

"

"

77^

(Vk

0,^

"

Pi(0)i(i) Pi(i)i(2)

the

"

"

variables

variables, and
P

of sequences
seek

we

time

including

the low

order

given

n,(o)(0)

size

n,(0),(i)..w(r)

and

are

parameters

asymptotically normally distributed


combinations

linear

are

nomial
multi-

of these

asymptotically normally distributed.


P'. Then
p]'/is the
of the matrix

state

time

0,

i at

nk-ij(t)be

0. Let

i at time
time

the

number

and j at time

"

t.

Then

of

moments

Uijit)
=

probability associated

set

the elements

k at

(2.10)

The

are

also

are

state

state

hence

nvXO

be
p,-y'

j at

probability of

sample

The

i{0), the

each

For

00.

hence
let

(pij)and

-"

with

and

Pi{T-\)i{T)

"

size increases.

sample

Let

Zl nj{0)

as

multinomial

simply

as

1)

r]k

^nk;ij{t)pi'"^'
pa

nk-ij{t)is

with

with

size of

sample

nt(0).

Thus

Var{n,.;,XO}

(2.12)
(2.13)

Cov

since the

set

us

now

follows
of nk;ij{t)

given

were

examine

procedures. The
to

be

n,(0)pir"po[l Pkf'^P.A,

-n,(0)?)ir'VoWr'V(7A,

Z!y i^k;ij{t);
they
seen

{nk.,i
At), nk;gK{t)\

other variables
Let

nk{Q)pkf'^Pii

"nk;ij{t)

(2.11)

in

will

be

of nk;ij{t)
in

probabilitiespa
-

8{n4;,X0

nk:i{t
-

between

"

theory for
"

1)

Thus,

PijUk-iit

1),

l)pij]

(2.15)
=

SS{[w,.;,;(0 nk-i{t
-

l)po] Ink:i{t
-

1)!

1)

nk-i{t

of nk-ij{t)given nk-i{t

IUk-.iit 1)}
Z[nk;ii{t)

(2.14)

"

obtaining the asymptotic

distribution
the

\)pii where

rik-iH

"

needed

multinomial, with

^).
(S'.

[1].

moments

conditional

Covariances

distribution.

multinomial

ih J)

0.

is

test

easily

246

The

READINGS

of this

variance

is

quantity

rik.iit 1)

"[nk;ij{t)

PSYCHOLOGY

MATHEMATICAL

IN

pijf

88{[n,;,v(0

"nk;i{t

1) pijf\ ",;,(^

nk-i{t

1)}

(2.16)

rikiO)pir^'Pij(l

of pairs of such

covariances

The

"[nk-ijit)
"

(2.17)

iik-iit

Z[-nk-i{t

"

"[nk-ij{t)
"

"

l)Pik]\nk.,,{t l)\

-Wfc(O) pkf^^PijPih

nk-i{t

\)pij\{nk-gh{t) nk-g{t

h,

1) pg,]

"

"

S8{[Wi-;i"(0

l)pij][nk;gh(t) rik-git

nk;i(t
=

1) p.u]

"

l)pi,][nk;ih(t) n,;,("

1) PijPih]

"

Pij).

1) Pij\[nk-ih{t) nk-i{t

"

Pij)

quantities are

""{[nk",^i{t) Uk-iit

1) Pij{l

"

\)pg}]

(2.18)

Ink-,{t

1),Uk-git

0,

8[Mi.;o(0
=

l)j

+ r)
nk;i{,t l)p^j\\nk-gh{t

8g{K-;iX0

nk;r(t

nk-g(t+

+ r)
l)pij][nfc;0A(^

l)pgi,]

+
Mi.;g(i

^.

l)pg/J

(2.19)

IWt;g(^+
=

To

nk;gh(s)

Since

"

we

and

nu-gis

size

"

ii t ^

1,

variables

Uk-ijit) Uk-jit

variables

uncorrected

Wi(0) fixed, nk-ait) and

assume

"

of multinomial

covariances

The
nkiO)pki~^\

l)pgh are

"

l)pijfor^

nk;ij(t) nk;i{t

variables

variances

probabilitiespij and sample


and

"

0 and

means

1),ni-;i(^ l), ni;iy(0|

0,

summarize, the random

have

"

"

i ^

or

ni-gh(t)are

0.

"

"

"

"

with

l)pij

g.

independent

if A- ^

l.

Thus

8[niy(0

(2.20)

Uiit

l)pij]

0,

(2.21)

8[n,,(0
8[wi,(0

Uiit

Uiit

l)pijf T. nk(0)pl'~'^
pijd
=

l)pij][nih{i)riiit
-

pij),

l)pih]

(2.22)
=

"

Hnk(Q)pki ^^ PijPih,

7^

h,

k=l

(2.23)

8[Wij(0

n,{t

l)pij][ngh{s) ng(s
-

l)pgh]

0,

t 9^

or

i 9^ g.

248

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

Let
T

E Z

(2.30)

fr=l

Then

limiting variance

the

limiting covariance

the

the

Because

pir"

Vk

of the numerator

between

of (2.27) is "i"i
Pij{l

linear

limiting normal

distribution

n'^ {pa

Since

n"^ {pa
Pij(l
have

are

other terms,

the

Q,

(i

"

"

"

1, 2,

Pij from

often

for

for

shall also make

hence

have

the

of the

samples with

the

This
this section.

values

fact

Hence,

can

in

of i
be

or

two

with

sample size

in the iih. state

for

different values

of i

(i.e.,the

the

about

probabilities

1, 2,

pij in

pij{t)

"

"

"

terms

m).
of

methods

of t

It
m

in

terms

standard

the variables Pij{t)

those

to

of
test

1)
of

pendent.
asymptotically inde-

are

similar

"

the estimates

as

1), and

"

nij(t)/ni(t

testinghypotheses concerning the Pij{t)it

reformulate

limiting

same

trials.

different values

proved by

the

of multinomial

sample sizes "ni{t

and
trials,

used

earlier in

will sometimes
T"- independent

procedures

may

then

applied.
3. Tests

of

3.1. Tests

On

covariances

limiting joint
limiting joint distribution as

same

hypotheses

hypotheses

be

pa) for

"

estimates

the

possible to

Ui

sample sizes n({)i(i

samples consistingof multinomial

be

i has

asymptotic distribution

same

probabilitieswith

different

for two

given

of the fact that the variables

use

t have

i and

given

multinomial

pa)

"

limiting

pa) and

"

probabilitiespa

independent samples consistingof multinomial


We

Pij(l

pij(l

pij) has

"

of observations

{ru^if^
{pa

possible to reformulate

be

pa)

"

number

functions

independent

0, variances

means

0, variances

means

asymptotically independent

are

similar

from

obtained

m)

(2.27),the variables

as

{ntf^ {pa

set

of multinomial

variables

\. The

"

with

0, variances

means

^Jlo ni(t).
set {n(t"if
{pa

factors),and

distribution

will

"

with

expected total

"

"

(see,e.g.,

"

"

covariances

The variables (ri4^i)^'^(pij


"5i(,pijP9ft/"/"i.
pij)

Also, the

the estimates

as

is the

and

rK/"i, which

of this limit

covariances

distribution with

distribution

"digPaPgh

n*

distribution

and

limit distribution

same

the covariances

where

"^igPiiPoh
In

distribution

joint normal

the variances

limitingjoint normal

limiting joint normal

covariances

and

has the

pa)

"

and
pij)/0i

"

pgh

nomial
multi-

[4]).

Pij)have

"

4"i pij

of normalized

increasing sample size,they have

limits of the respectivevariances

2, p. 5 in

Theorem

and

distribution

the

Pij),and

"

is "dig

combinations

variables,with fixed probabilitiesand


a

different numerators

two

of (2.27) are

numerators

0i

(=1

of

the basis
derive

can

every

pij "

First

we

hypotheses

and

hypotheses

of the

certain

confidence
about

regions.

asymptotic distribution theory


methods

confidence

specificprobabilitiesand

of statistical

inference.

in the
Here

regions.

preceding section,we
we

shall

assume

that

0.

consider

testing the hypothesis

that

certain transition

probabiUties

T.

have

p.j

specifiedvalues

and

{n*f

the
zero,

determine

or

We

1,

"

"

for

m,

"

given

i. Under

test

for

one

the usual
of

/A

which
m

the
A

of the set pij for which

in the denominator
for different i

be

can

instance

(3.1) over
The

use

all

i,resultingin

of the

of

-test

there is

for

good reason
as
goodness of fit,described

of
were

borrowed

3.2.
In

Testing

the

is that the
the

null

the

region

less than

and

"

Markov
1 moves

hence

(i,j

1, 2,

"

"

"

goodness

p,j

ph j

freedom

critical

of the

region
pij for

set

coefficient

sists
con-

n*{pij

are

x^-variables.

other
obtained

be

can

po

pijf

"

(3.1)for different

to obtain

m)

to

x^-distribution
with

variables

forms

by adding

1) degrees of freedom.

"

in the

as

the

the

of fit is discussed

in this section

(according

significancepoint. (The

adopting the tests, which

chain,
to

the

is the

pij

state

transition

t. A

at

(t

pij

of the transition

1,

in
are

[5].We

believe

analogous

situation

from

to

that

x^-tests

which

probabilities are

probability
general

t; let

on

"

"

"

that

they

us

say

T). Under
time

the
t are

"

function

1)

the null hypothesis is

IlUPir''-

(3.3)

likelihood function

under

maximized

maximized

under

the alternative

UllP.itr^''.
t

i.j

is

to

it is

""'^^^^

pijit)

constant.

individual

an

alternative

probabiUties for

riiit

(3.4)

be added

can

transition probability depends

(3.2)

The

that

of confidence

the

with m(m
x^-variable

hypothesis H:pij(t)

hkelihood

more

consists

replaced by pij.) Since the

hjrpothesis that

the estimates

The

or

one

(see [5]).

stationary

state i at time

about

variables).Thus

significancepoint of

confidence

for all pij

test

pij

degrees of

"

asymptotically independent, the

are

tains
ob-

as

way

\2

significancelevel

(3.1) is

asymptotically independent,
For

of multinomial

than

of freedom.

degrees

with

of this hypothesis at

(3.1) is greater

"

means

Pij

-distribution

asymptotic theory

test

one

more

same

^P^LZ^

n*

i=i

asymptotic

or

testing the hypothesis

an

the

hypothesis

with

asymptotic theory for

standard

use

to

p?jin

on

the null hypothesis

distribution

limiting normal

the null hypothesis,

(3.1)

has

of the fact that under

can

region

consider

specificexample

We

249

GOODMAN

A.

depending

distributions

confidence

LEO

use

covariances

As

make

estimates.

normal

or

AND

pij)have

"

and

for multinomial

ANDERSON

pij

(pij

variances

multinomial
Pij

W.

in

this assumption

We
Pij{t).
alternate

test

pothesis
hy-

250

READINGS

ratio is the likelihood ratio

The

PSYCHOLOGY

MATHEMATICAL

IN

criterion

(3.5)
PiiWJ

of

shght extension

"

is distributed

log X

theorem

as

of Cramer

with

(T

1) [m(m

"

of Neyman

[6] or

[11] shows

that

1)]degrees of freedom

"

when

the null hypothesis is true.


ratio (3.5) resembles

Hkelihood

The

in

of homogeneity

tests

further this similarityto usual

equivalent
For

those

to

has

table, which
used

1, 2,

"

have

rows

"

"

"

jth.column

the

to test

the

Pirn

is

equivalent

^y

with
,

equal

pij

in Section

given

as

i and

given

for j

be

can

1, 2,

6.

mates
the esti-

contingency table,

from

multinomial

that

is,in order

random

that

"

"

"

1, such

that

variables represented by the

the

data

the hypothesis that

to

homogeneous

are

there

are

constants

the probability associated

that is,Pij(t)

for t

pa

this

in

1,2,

"

pn

with
"

"

the
T.

"

appropriate here ([6],p. 445); that is,in order

seems

calculate

22 ni(t

l)[pij(t) pijfI
-

is true, Xi has

Pij ;

the usual limitingdistribution with

(m

"

1)

of freedom.

of the

test

the

in all T rows;

to pa

hypothesis

1) degrees

Another

for
pij{t)

distribution,so

same

is

"

as

appearance

interest is that

this hypothesis, we

if the null

asymptotically

are

asymptotic distribution

same

proof

r.

"

(3.6)

(T

formal

same

of homogeneity
x^-test

The

the

respect. This
Pi2

"

tables.

velop
de-

now

probabilitiesPij{t)for T independent samples. An

hypothesisof

The

pij(t)has

represent the joint estimates

to

and

procedures for contingency

by this contingency table approach

set

of multinomial

for standard
shall

presentedearlier in this section will be

given i, the

obtained

ratios

(see [6],p. 445). We

contingency

that the results obtained

likelihood

tables

hypothesis

trials
test

to

can

of

homogeneity

be obtained

by

for

independent samples

of the likelihood ratio criterion;

use

this hypothesis for the data

given

in the

table,

calculate

(3.7)

X,

IliPij/PiAt)^''''.
t,j

which

is

formally

distribution of

"

similar

log

Xi is

to

the

x^ with

likelihood

(m

"

1){T

ratio
"

criterion. The

asymptotic

1) degrees of freedom.

T.

of i.

value

given

AND

relating
the

Hence,

LEO

the

to

A.

251

GOODMAN

table

contingency

hypothesis

be

can

approach

dealt

tested separately for each

of i.

value

Let

us

consider the jointhypothesis that

now

1, 2,

"

"

of i

values

the set of
the

m,

"

I,

the fact that

directlyfrom

and

ANDERSON

preceding remarks

The

with

m,

W.

"

"

T. A

"

the

random

variables

calculated for each i

1,2,

"

"

for all i

Pij{t)and

1, 2,

the

null

"

"

"

follows

hypothesis

pij for two

under

different

hypothesis,

asymptotically independent,

are

,m

"

pa

joint null

asymptotically independent. Hence,

are

Xi

Vait)

of this

test

sum
TO

x'

(3.8)

E X?

E E riiit

with

based

Similarly,the test criterion

Vi^'I

Va

t,j

limiting distribution

has the usual

DlMt)

m{m

(3.5) can

on

1)(7'

"

1) degrees of freedom.

"

be written

Ttl

(3.9)

log Xi

-2

log

-2

X.

i=l

of the

3.3. Test

second-order

hypothesis that the


chain. Given

Markov

of being in state
Pijk for t

2,

"

t. When

k at
"

T.

"

{j,k)

at

composite
t

given

of state

form

composite

(h,k), h
chain

9^

with

to

This

representation is useful

chains
Now
and

be

can

let

t, and

the ni(0) and


the

ni(0)

riijkii)
(i,j,
the

let nij{t

1,

different sequences

nij{t
"

because

"

"

1)

m;

of states.

"

of

the

first-

i and

states

composite

state

the probabihty

course,
states

easily

are

transition probabiUties 0.

of the results for first-order Markov

"

The

is well-

as

comphcated

more

composite

certain

in state

the n^Xl)
"

probabihty

special second-order

at

T)

is

conditional

were

set

2, in

"

j at

in this section

assume

extending the idea of

2,

2 and

"

stationary,pukit)

Of
Pijk{t).

The

zero.

i "t t

2.

and
=

as

1 is

"

We
E^ ''^Hkit).

nonrandom,

"

first a

the other hand,

probability

of individuals

"

is

i. On

on

with

some

Section

nonrandom

were

and

states

number

Wij(l) are

the

(i,j) at

state

from

over

be the
riijkit)

in k at

where

carried

is

Consider

the

this,let the pair of successive

j, given (i,j), is

seen

chain

represented

(i,j). Then

state

the

be

can

do

order.

is in state

2,3,---,T)he

depend

not

chain

(see,e.g. [2]).To

chain
a

individual

an

the second-order

Pijk(t)does

known, the second-order

given

first-order stationary chain

for which

chain, one

define

is of

1, let pukit) (i,j,k=l,---,m;t

injatt"

order

chain

that

the

earUer

random

"

1,
that

sections

variables.

The

of sufficient statistics for

distribution

of nijk{t),
given

1),is

(When

the transition probabiUties need

the symbols pijk should, of course,

be

not

be

the

same

for each

time

interv^al,

replaced by the appropriate Pukit)through-

252

READINGS

the

of

set

riijk

is the

obtained

be

^r=2

nijk{t)form

of pijk for

likelihood estimate

of

for

was

'^m

;=i

Now

let

against

the

Pi"fc

that

"

T,

"

t.

stronger result

ing
concern-

the

maximum

1).

"

"=2

hypothesis that the chain

it is second-order.

1,

==

"

"

is first-order

hypothesis is

null

The

A:

'Pmik
Vik,
say, fory,
testing this hypothesis is^

"""

ratio criterion for

and

"

null

2,

is

(=2

testing the

alternative

V'2.jk

consider

us

i,

21 Wiyt(0/ 2Z nait

(3.10) over

/m

and

"

first-order chains; namely,

stationary chains

"

of sufficient statistics. The

set

"

probabiUties, a

it

as

I,

product

transition

stationary

sufficiencycan
numbers

nijk{t)for i,j, k

n,;(l) is given,

with

chains

For

PSYCHOLOGY

MATHEMATICAL

of

joint distribution

out). The
when

IN

The

m.

"

that

likelihood

(3.12)

/ p^i.)"'"'^
(PiA.

i,i,k=\

where
m

(3.13)

p^vi

maximum

is the

from

likelihood

(2.8).This
random

nij(l) were
were

We

1)

"

1)

"

m{m

the

given j,

estimates

2,

"

"

"

m).

n'^ {puk
X

for t, /e

2,

"

"

"

m,

1, 2,

and

the

"

used

be
"

has

table, which

contingency table, can


and

probabilitiesfor

develop further

now

formal

same

the

distribution

seems

that

for

p^*

Vjk

yuk

appropriate.

To

as

appearance

estimates

hypothesis is

of homogeneity
x^-test

tained
ratios ob-

independent samples {i

the

represent

null

The

m.

"

to

of freedom.

likelihood

shall

the nij{l)

asymptotic x"-

1)^degrees

"

(3.12) resembles

ratio

that

an

"

of multinomial
An

has

procedures for contingency tables.


asymptotic
puk) have the same

this similarityto standard


the

what
some-

earlier section the

assumed

we

log

"2

pjk differs

that

in the

for problems relatingto contingency tables. We

For

T"\

here

see

the fact that

to

hypothesis,

likelihood

the

We
"

in this section

m{m

"

Z) ^7Vfc(0
/ 23 w/O

of y^k

while

the null

rn(m

that

observe

'^m

difference is due

Under

with

distribution

estimate

variables

nonrandom.

w,yfc /HI]

as

1,
a

given j

for

1,

this hypothesis

test

calculate

(3.14)

2Z n*j(pijk pjkf/pjk

Xi

i,k

where

(3.15)

n*j

2]

riijk

23 22 nijk(t) 22 ^ijit
=

If the hypothesis is true,

degrees
2

The

Xi

1)

-=

1=2

t=2

has

the

usual

"

2Z nij(t).
(=1

limiting distribution

with

(w

of freedom.

criterion

was
(.3.12)

written

incorrectly in (6.35) of [1]and

(4.10) of [2].

"

1)

T.

analogy with

111 continued

for

homogeneity
by

ANDERSON

W.

use

is

The

formally similar
of "2

of

given value
of

Let

us

2,

"

"

"

log

the

to

of the

test

trials

multinomial
We

hypothesis of
be

can

tained
ob-

calculate

(pjkI ViiLT''\
likelihood

ratio

asymptotic

The

criterion.

with (m
1)^degrees
relatingto the contingency table approach dealt

with

for

each

X" is %

remarks

preceding

value

from

253

GOODMAN

A.

3.2, another

independent samples

X,

distribution

Section

LEO

of the likelihood ratio criterion.

(3.16)
which

AND

the

j. Hence,

of freedom.

"

hypothesis

tested

be

can

separately

j.
A

m.

the

consider

now

of this joint hypothesis

test

that

joint hypothesis

p,yt

can

for all i, j, k

pjk

be obtained

1,

by computing the

sum

X^

(3.17)

Hxj

which

Similarly the

criterion based

test

Vik

j.i.k

with

Umiting distribution

the usual

has

Z) n*j(pijk VikfI

m(m

of freedom.

written

be

(3.12) can

on

1)^degrees

"

-2

log Xy

log

-2

-=

X)

(3.18)

/ p"J
log [pijk

nijk

''"

^='

X)

[logpijk

nijk

log pjk].

ijk

preceding remarks

The
Let

{i,j,

pij...ki

I at

state

and

state

time

"

"

"

i at time

is

(t

"

chain

directlygeneralized

be

can

k,

"

"

1, 2,

"

"

r,r

-{- 1,

of order

"

"

"

"

and

"

state

it is not

m) against the alternate hypothesis that

chain

of order

at

time

r.

an

for i

Pj-ki

"

-{- I

"

hypothesis

shall test the null

(that is,Pij...ki

"

"

T). We

"

for

the transition probability of

denote

m)

"

k at time

t, given state

that the process

2,

"

1 but

I,

r-order

an

chain.
denote
nij...ki{t)

Let

respectivetimes

the

We
^7=1 'nij...ki{t).

r, t

here

assume

(3.19)
where

-\- I,

"

Pij...ki

nij...ki

"

"

"

the

"

.ki

For

as

2,

"

"

"

set

j,

estimates

m), and

"

"

k,

"

let nij...k(t
"

are

nonrandom.

I at

1)

The

nij...ki/n*j...k
,

and
Xr=r nij...ki(t)

n*j...k X

given

1)

i, j,

is

r-i

(3.20)

states

I, t, and

"

the nij...k(r

that

of pij..

likelihood estimate

maximum

frequency of

the observed
"

"

"

"

nij...ki

fc,the

be

^ nij...k(t 1)

set Pij...ki will have

of multinomial

may

"

the

probabilitiesfor

represented by

an

same

n,j...k{t).
asymptotic

tion
distribu-

independent samples {i
table.

If the null hypothesis

254

('P:j-ki

for

Pj--ki

1, 2,

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

"

"

is true, then

m)

"

the

of homogeneity

-test

appropriate, and

seems

xl-k

(3.21)

Pj...kif
1^ n*j...k(pij...ki
I Pj...ki

where

(3.22)

the usual

has

that

here

assumed

in this

differs somewhat

from

the

the

m~^

2Zi, -,* Xy

sum

"

"

"

"

"

Another

and

sets

j,

"

"

""

all values

test of the null

from

1)

"

(j

1, 2,

see

for

1 to

"

"

m;

"

"

; k

"

1, 2,

"

k)

m),

"

"

if

"

for i

is true.

obtained

be

can

"

"

ni'~\m

Pj-.-ki

"

while

Pj...ki

fixed.

are

hypothesis {pij...ki
of j,

hypothesis

"

1)-order chain,

"

hmiting distribution with

the joint null

under

We

estimate

parameters

that the nj...ki{r

",

(r

an

with

variables

the usual

will have
-fc

degrees of freedom
1,2,

random

assumed

have

we

are

1), for

"

be multinomial

of freedom.

likelihood

maximum

the fact that the nj...ki(r

to

paragraph

1)^degrees

"

This
?^y..-i^(0/2Z^=r-2
nj...k(t)).
(viz.,2ir=r-i

l)-orderchain

"

to

Since there

(m

difference is due
are

21 n*j...k 22 nj...kiit)
/ 2^ nj...k(t),
with

(r

an

I
n.ij...ki

limiting distribution

pj...ki

for

Pj...ki

pj...ki

by

of the likelihood

use

ratio criterion

(3.23)

\j...k
=

",
{Pr..ki/Pij...ir'

i.l

"2

where

asymptotically as x^ with

log Xj.-.k is distributed

of freedom.

2Z {-21ogXy...t}

nij...ki\og{pij...u/pj-ki)

i,j,-'',k,l

j,---,k

with m~^{m
limiting x^-distribution
joint null hypothesisis true (see [10]).
a

In

where

the special case

successive

at

time

1, the
points are
=

alternate hypothesis that observations

is

the process
is of order

order

will note

reader

The

chain

be

can

of order

is of the

test

to

these statistics

from

test

when

section, we

for each

used

to

hypothesis that

test

null

the

servations
ob-

hypothesis

in

it is of order

(m

this section, we

asymptoticallyas

"

can

ratio and

x^ with

[m''
"

that

r).By

an

compute

"

proach
ap-

the
that

observe

m"]im

it

is of

the process

the null hypothesis that

earUer

hypothesis that

grees
1) de-

the null hypothesis is true.


have

assumed

that

the

transition

probabihtiesare

interval,that is,stationary. It is possible to


has
the rth order chain
stationary transition

time

the

first-order chain.

1 against the alternate

"

when

hypothesis that

null

the logarithm of the Hkelihood

distributed

are

of freedom
this

times

presented

of freedom

statisticallyindependent against the

are

method

the

generalizedto
that

"2
X^-criterion
or

same

that

\f degrees

"

against the alternate hypothesis that


similar

In

if degrees

"

Also,

(3.24)
has

(m

test

the

the null

probabihties

256

READINGS

where

naff,^v

When
2Jf=iiia^.^vit)-

hkelihood

Under

of

the number

{A

\){B

\){AB

modified

4. A

-qi

is

imum
max-

where
,

has

asymptotic x"-distribution,and

an

AB^AB

1)

"

A(A

"

"

1)

B{B

"

1)

"

-^ A+B).

An

probability

with

log

"2

of freedom

In

model.

nonrandom.

were

is

hypothesis,

degrees

the null hypothesis is assumed, the

of Pa".iiv is qa^ r^y

estimate

null

the

PSYCHOLOGY

MATHEMATICAL

IN

the

is that

the ni(0) are

size

Then

alternative

and

assumed

preceding sections,we

sample

n.

the

(2.5)multiphed by the marginal distribution

distributed

distribution

of the

the

that

multinomially

of the

/i/(0)which

set

ni(0)
n,j(0

set

is

i^^^^nr,.r'^r

(4.1)

11^^.(0)!'=^
J=l

In

this model, the maximum


likelihood

maximum

likelihood
of

estimate

-m

(4.2)

estimate

of p.y

is

again (2.8),and

the

is

rii

=
.

The

variances, and

means,

taking

the

expected

nfc(O)replaced by

nrik

Since nk{Q)/n estimates


of n'^ {pij

"

pij) are

asymptotic theory
The
starts

of (2.20) to

values
Also
.

T]k
as

n,XO

~"

(2.23); the

^i(^

in Section

of the tests

given
and

2.4. It follows
in Section

stationary state; that is,if

k=l

uncorrelated
variances
from

Vi-

found

l)pij are

these

and

with

by
with

apply

^.^(0).

covariances

facts

for this modified

simplify somewhat

covariances

^VkPki

3 hold

"

formulas

same

^)Pij are

(4.3)

ni{t

"

consistently,the asymptotic

asymptotic variances
from

of nij{t)

covariances

that

the

model.

if the

chain

T.

then

For
from

/_,

of

available.
known

efficient

have

for the
rji

and

Tj{

ignored.

subject

the

of

case

likelihood

Pij

Lagrange

multipliers can

subject

observation

asymptotic

results

00, while

has

when

shown

"

"

"

the

I and

"

m)

from

for

Section

2.4. This
we

stationary

the

7?i

log

the

zljPii
the

not

because

relevant
estimates

/mAO)

1, Pa

the

obtained
=

maximizing
Vj

for

equations

known,

are

th

by

zlt ViPa

1)

is

mates
esti-

likelihood

where

are

it

are

state

pa

case

zL^ijlog p,_, +
Sj Vj
Vj

"

are

likelihood

state

maximum

same

for

as

likelihood

of the

chain

is

the likelihood

0.

Pij

maximum

the

the process

is of

[10],and

Hoel

the

that

time

t, for t

"

rii

probabiUties

an

"

that

0 and

means

in

proved
Wi(0)
as

on

"

"

the

is of order

r{u

r), which

"

alternative

to

Bartlett's

Pij...ki

Pa-.-ki

(i

argument
have

pa)

"

and

same

given

Gardner
=

test

1 is

and

are

that

essentially

For

tests

the

to

(specified).

is

case

3.1

the order

parallelto
x

-test

the

hypothesis

can

3.3,
be

are

example,

where

in Section

in Section
ratio

3.3. The

the

alternate

presented

likelihood

it is

in

[8].

procedures

chains.

the
like

limiting

covariances
L. A.

1,

that

m), and

"

in Section

given

the generalizationsof the

n,j/nt

the null hypothesis that

test

criterion for this test

order

"

possibly nonstationary

also

given by the usual

0. An

by

against the alternate hypothesis

the process

"

and

co

Hence,

="

pa

observations

are

variances

different way

asymptotically

estimates

of i

in

was

(see [3],p. 91). He

1, 2,

is

T,

"

(n*)^'^
(pij

the

1)

oo). Bartlett

-^

sequence

covariances

(j

pij

the variables

criterion [10] to

the null hypothesis that

"

different values

applicablefor large T. Also, the x"-test presented


provide

"

(T

observed

the

independent

to

(n

they consider the asymptotic

appropriate

tests

ratio

sequence

likelihood

maximum

of states

increases

times

at

z27=i^i(O)

hence

asymptotic variances

was

such

ratio

is that

one

previous sections,

the asymptotic theory for T

this hypothesis, and

hypothesis

and

oo

'positivelyregular'situation

fixed and

for

except

-^

observed

of observation
of

Uij

Wi(0)

of

the

and

with

that

see

for

case

for two

result

Hoel's

to

in the

where

case

general

more

great length. In

that

2.4 shows

distribution

valid

for the

of

multinomial

of Section

Thus

starts

use

of the

maximum

chain

in state

in the

covariances

normal

the

the

obtain

number

formulas

asymptotic
that

general

to

of times

([3],p. 93)

multinomial

2,

more

stationary

[3]and

Bartlett

z27=i^ij) have

even

used

fixed. The

normally distributed

(n*

in

presented

were

by

that

i at time

has

the number

shown

state

or

the

1, "^i ViPa

restrictions

be

on

was

studied

been

theory

the

chain

additional

some

estimates

used

in

'^jPa

of

be

by maximizing

estimates

to

the

estimates.

5. One

has

have

chain

chain

that

obtained

the restrictions

to

hood

of

obtained

are

pij

and

-m

with

the special case,

In

257

GOODMAN

A.

of the

estimates

special case

is

maximum

(4.3) holds,

zLnijlog

LEO

Ti)i If it is known

"f)^

in this paper

The

case.

in the

0- In

rji

the

AND

knowledge

dealt

whether

information

log

and

Vi

when

-pn

We

for this

0,

ANDERSON

stationary state, equations (4.3) should

estimation

not

pi!"

Vk

W.

for
null

is that

are

also

generalized

criterion [3] for testing

258

READINGS

ratio criteria. The

likelihood

6. x^-tests and

PSYCHOLOGY

MATHEMATICAL

IN

in this paper

x^-testspresented

equivalent, in

to the corresponding likelihood


a certain sense,
are
asymptotically
to follow
ratio tests, as will be proved in this section. This fact does not seem

general theory of

from

the

from

those

that
x^-tests
in each

individuals

2.1) as
need

consider

not

herein
For

small
are

the

to

based

are

may

samples,

chain

preferred (see
asymptotic distributions

samples

known.

will be

power

when

shall

now

that

prove

the

p's have

relevant

this will be
statistics
ratio

the

of the form
but

to

under

null

here

can

be

herein, and
Let

us

under

applied

consider
From

mi{t)

p,j(l
by
X

usual

"

the

formally

are

show

well

similar

[nmi{t

tingency
con-

that

the

In

particular,

to

prove

that

the likelihood

appropriate

of proof presented

method

-distribution

lence
question of the equiva-

the

in

the

other

sections

oo

x^-statistic
(3.8) under

that n'^ (pijit)


"

Pii(l

variances
or

Va)

Vij]have

are

the

'Pij)/mi{t
are

null

pothesis
hy-

asymptotically
"

"

different i, they

1)]^'^
[piXO

"

whenever

the

asymptotic

an

certain

log \i is asymptotically

"2

it has

mogeneity
ho-

of the

0 and

to

ratios, have

where

as

different t

"ni{t)/n. For

in

used

their motivation

distribution.

hypothesis. The

see

means

form

used

be

can

shall discuss

we

as

2.4, we

with

be preferred

asymptotic x^-distribution

appropriate statistics given


=o

to

are

(see (3.6)).In order

x"

therefore

distribution

Section

1), etc.,

asymptotically

ances
asymptotically vari-

2Zt f^iit
1). Then
p*1) pij(t)/^t'ini{t
has
^nrwi
p*jf/p*i
x^-theory,
asymptotic
an
(t
l)[pij{t)
Ptj),etc. Let

-distribution under

(6.1)

the
T

the

independent. Then
the

then

the alternate

to

also where

normally distributed
where

shall

hypothesis. Then

of the tests

small

x^-tests
(tests of

the

an

of proof

actually hkelihood

and
the Xi-statistic,

equivalent

for

specificalternate

methods,

and

ratio

has
x^-statistic

method

X, (see (3.7)),which
are

tests

of the

are

which

approach

tests

asymptotically equivalent in

are

of the form

not

of

tive
related to the rela-

there is

appropriate limiting normal

asymptotic distribution,we

the

3.2

the

that

for statistics

true

criterion

likelihood

hypothesis. The

null

and

simpler.

the

in Section

shall show

First,we

be

of the

deciding which

of these

decide

to

somewhat

large and

users

many

to

chain

relative rate

power

which
x^-tests,

of the

advantage

presented
sense.

relative

for

tentatively suggested

their application seem

under

the

accumulated

[5]).The

this section,a method

In

tables, is that, for


We

and

been

in

the sample size is moderately


An

hypothesis.
and

has

comments

is not

on

Markov

model.

data

enough

not

be

to

from

extremely general, while the x^-tests


presented

be

Markov

obtained

of

(see Section
m^ categories

sequences

x'^-tests
based

of interest. The

having been

as

different

are

directlyby considering the number

possiblemutually exclusive

the data

on

presented herein

-tests

obtained

be

variables

hypothesis

the alternate

tests

of the

the multinomial

-tests;the

can

"

"

"

the

null

"

hypothesis. But
p

lim

(p*

pij)

T.

ANDERSON

W.

LEO

AND

A.

259

GOODMAN

because

(6.2)
the

From
and

(^

p lim

in probabilityof (p*
pa) and {niiit) ni{t)/n),
it follows that
n'^ (piAt)
Pa) has a limitingdistribution,

convergence

the fact that

"

"

"

[nZ^-^^ ^^^f^ ^*^^' ^ rjMj^JliMLlljl]


o.
"

(6.3)

miit))0.

p lim

"

Pij

Pij

has the same


asymptotic distribution as zLnniiit 1)
Hence, the x^-statistic
This proof also indicates that the
that is,a x^-distribution.
[Pijii) ptjf/Pij}
show
that
shall now
We
X?-statistics
(3.6)also have a limitingx^-distribution.
is asymptoticallyequivalentto Xi under the null hypothesis;
2 log Xi (see(3.7))
and hence will also have a limitingx -distribution.
We first note that for |a;|
" ^
"

"

(1 + x) log (1 + x)

(1 + x)ix

xl2 + xjZ

xV4

"

"

"

(6.4)

x72

(.tV6)(1 x/2
-

"

"

),

"

and

x)

(6.5) I(1 + x) log {\^


(seep.

217 in

We
[6]).

-2

log Xi

(6.6)

-2

I(a;V6)(l x/2 +

"

"

)\^\x\

"

also that

see

x^/2I

X-

X) nijit)
log [pij/pij{t)]

E nit

1)pi,(^)
log [pii{t)/vd

j,t

Z Wi(^

Dp.ytl+ ^vi(^)]
log [1 + Xij(t)],

i.t

where

PiMpa

[paii)

a;iy(0
=

"

"

The

difference A between

"2

log Xi and the

is
Xi-statistic
A

1ogX.

-2

(6.7)

Y.i.i
riiit i)piA[i+
-

2l7=i
Pa^ijit)

Since

Xi

[xiXOf/2}.
E riiit l)piA[l+ Xiiii)]
log [1 + a;iy(0]-a;.y(i)
-

i.e.for any
probability;
tends
e, under the null hypothesis,
probabilitysatisfies the relation

that A converges
of
the
relation |A | "
probabiUty
shall show

We

Ei ni{t)

"^

Pr{ IA I "
(6.9)

[x.mm.

0,

(6.8)

log [1 + xiM

xiM

The

00
.

e} ^ Pr-{IA 1 "
^

Pr{ 12

Pr{2n

to 0 in

to

"

0, the

unity as

IXij{t)
\" U

and

and |XiXO \ "


1"
J^j,t
riiit l)pii[Xii{t)f
"

Y.j.t
I'"
IXiiit)

"

and

\"h].
]Xij{t)

h\

260

READINGS

It is therefore

only

necessary

[pait)

Xij(t)

Since

the null

hypothesis,and

(6.10)

v^^;Mn x,m

to

that

prove

n[Xij{t)]converges

pijl/pij
converges

"

to

in

zero

in

to

ability.
prob-

probability under

{[^"%~
^'']
[-^-^^^]}

v^^;m^

PSYCHOLOGY

MATHEMATICAL

IN

'

it follows that

n[xiM'

(6.11)
to

converges

Since
and

the

has

"

-statistic has

log X,

method

The

log Xi and

in

converges

null

Xi

also
"2

under

probability to

zero,

log

would

It

apply

to

in Section

we

the likelihood ratio criterion under

as

have

the alternate

suppose

t, s, i,j. It is

to

easy

under

increases,the

may

Pait)

also

moved

hypothesis,
This

be

can

be

can

We

when
we

shall

alternate

Since

(6.12)

the

to

null

that

the two

the

and

tests

hypothesis given
is

statistic

in

the

approach

(x^or

"2

some

to

as

some

ratio test

are

of Pij(t)
as

close to 1 in large

hypothesis

If the

closer

move

values
the

to

of
null

again asymptotically equivalent.


of the

proof of asymptotic

in this section

(see also [5],p. 323).

the

null

preferred

of these

is

rejected

specifiedcritical

the likelihood

to

"2

when

tests

hypothesis

(stochastically)larger than

sense

for

[11]).

increases.

is true.

kept fixed,then

are

is not

log X) exceeds
be

Pais)

tests, the alternate

are

kept fixed. Since

is
x^-test

limiting

hypothesis

the comparison

to

hkeUhood

same

the Hkelihood

fixed but

the

null

null

is,Pij(t)^

shght modification

1)

"

the null hypothesis.

the power

not

X,

m(m

actually Ukelihood

not

are

to

"2

of statistics

that, under

other words, if the values

hypothesis

hypothesis are
seen

value,

ratio

log

test

under

hypothesis.

ni(t) is

converges

test

applied

1 (see [5]and

by

x^ is

ni(t)/n converges
/n

to

hypothesis

statistic

the alternate

tends

deduced

an
appropriate
might decide that

if the

-test

In

suggest another

now

comparisons between

make

to

alternate

it

is true; that

the situation in which

closer

equivalence under
the

of each

3.2

-statistic has the

hypothesis.
hypothesis and the significancelevel

power

for the

the

3.2

where

case

hypothesis

that both

see

order to examine

samples and

the

alternate

any

for the alternate

be

refer to

remarks

previous

In

another

in Section

Xt since they

proof that the

distribution

consistent

x? +

asymptotic equivalence

proved

ratios.)Hence,

Now

log X,

"2

with
limiting x^-distribution

(The proof

not

the

was

"

has

show

to

and

of freedom.

ratio criterion,but

The

Q.E.D.

the null hypothesis,

the null hypothesis.

used

be

log

hypothesis, "2

1) degrees

"

-distribution

presented herein for showing the asymptotic equivalence of


could

of the form

(T

x"

"

the null hypothesis is true.

Hmiting

under
limiting x^-distribution

the

probability when

in

zero

Xij{t)f
[ixiAt)ny''

in
in

linear

combination

probability to its

probabiUty

"

of multinomial

expected value

variables, we

to

m,(t

see

that

"[ni{t)/'n] mi{t).Hence,
=

l)[pii(t) pijlVpii
-

T.

and

log X)/n

"

ANDERSON

W.

in

converges

(6.13)

LEO

AND

261

GOODMAN

A.

probability to

Mt

Dpiiit)log [pijit)/pi,],

i,},t

where

(6.14)

pij

XI Pijit)mi(t 1)/ Xl niiit

(6.15)

is

alternative

tests

are

for the

applied,
simple alternatives

(see p.
decrease
the

as

alternate

reducing the

hypothesis

be

(b)

of the

statistic
has

form

be

where

en,

limit

quite different from

approximate computations

it will tend

and

the
can

we

suggest which

points
enough

limit

information

the
seen

methods

The

x^ and

to

the
for

least it

test

to

I, or

(b)

this

of

the

methods

certain sense,
alternate

determining

the

of

the

test

herein

critical value
the

critical

question

some

from

see

will tend

test

to

is greater than

of the

the power
ratio

test, and

if

the

some

preferred.

the significancelevel

of

asymptotic theory

so

II

particular Type

some

as

the

-test

does

not

error

give
limits

of stochastic

comparison

the

significance

can

also be

used

study of

in the

ordinary contingency tables. We

for

x^ and

hypothesis
which

the

hypothesis.

powers.

comparison discussed
ratio

0 if

is to be

vary

usual

problem,

null

desirable sequences

that

seems

en). While

handle

comparison of

hkelihood

that, in
when

method

c' and

lie between

suggest

may

(or at

hkelihood

of the

power

to

find that

appealing approach is to
However, a more
ratio of significancelevel to the probability
approaches

made

When

be

comment

be

can

really suitable),we
of

of type

the

increases.

?i

the power

that

Hence, by this approach

hmit.

error

to

(there may
is

critical value

for the

as

constant

an

compared

hypothesis.

(a) is used, then

will increase
is

of

steadily closer

moved

preceding paragraph

the stochastic

is less than

and

computed

alternate

when

case

related to Cochran's

chance

[3].If method

in

log X)

"

this form

stochastic
can

(x^or

in the

remarks

can

discussed

was

the

whether

to

in the

limits. When

usually the

be

can

is somewhat

tests

increases,thus

Method

value

included

are

as

[5])that either (a) the significanceprobability

in

323

that

comparing

for

hmits

stochastic

these

then

method

This

composite hypothesis,

some

0,

Pij)/pijis

{pij{t)
"

the two

between
is

differ from

limits
If

is better.

test

difference

small

stochastic

two

suggests which

only

Pijf/i^pD-

hypothesis, these

will be

small, then there


the

l)[piAt)

of them

computation

and

Zwii (t

pij

n-"c!0

(6.13)is approximately

alternate

the

Under

lim

(6.12) and

between

difference

The

1)

likelihood ratio methods


and

is true
is to be

fixed, and

we

are

have

not

have

lent
equiva-

suggested

preferred.

REFERENCES

[1] T.

W.

Anderson,
RAND

"Probability

Research

Memorandum

models
No.

for

analyzing

455, 1951.

time

changes

in

attitudes,"

262

[2]

READINGS

T.

Mathematical
The

[3]

[4]

M.

H.

S.

W.

"The

Cambridge

Philos.

[6]

H.

Soc,

47

changes

edited'

Sciences,

Paul

by

in

attitudes,"

F.

Lazarsfeld,

1954.

of

goodness

Vol.

time

analyzing

Social

Illinois,

frequency

"Large-sample

(1951),

theory:

fit

for

test

probability

Proc.

chains,"

86-95.

pp.

parametric

Math.

Ann.

case,"

Vol.

Stat.

27

1-22.

pp.

"The

Cochran,
pp.

for

the

Glencoe,

Press,

Free

Chernopf,

G.

in

Thinking

Bartlett,

(1956),
[5]

models

"Probability

Anderson,

W.

PSYCHOLOGY

MATHEMATICAL

IN

of

x^-test

goodness

fit,"

of

Math.

Ann.

Vol.

Stat.,

23

(1952),

315-345.
Mathematical

Cramer,

Methods

of

Princeton

Statistics,

University

ton,
Prince-

Press,

1946.

[7]

W.

Feller,
Wiley

[8]

L.

A.

L.

A.

[10]

P.

G.

[11]

J.

Neyman,

Hoel,

Symposium
Press,

"On

the

Vol.

26

Stat.,
"A

test

for

on

Berkeley,

p.

Mathematical

the

in

problems

1,

John

information

1954.

Library,

of

analysis

theory
Statistics

pp.

Vol.

Applications,

Markov

chains"

(abstract),

Ann.

771.

chains,"

to

Its

distribution

University

statistical

Markoff

1949,

and

Columbia

(1955),

"Contribution

and

Theory

1950.

estimation

Essay,

Master's

Goodman,
Math.

York,

"Some

Jr.,

Probability

to

New

Sons,

Gardner,
theory,"

[9]

Introduction

An
and

239-274.

Biometrika,
of

the
and

Vol.

x^-test,"
Probability,

41

(1954),

Proceedings
University

430-433.

pp.

of

the
of

Berkeley
California

264

READINGS

which

those

permit

kind

of the second

be also included

to

in

descriptionof

unitary

chastic
sto-

behavior.

choice

PSYCHOLOGY

MATHEMATICAL

IN

(t,t +

time

"implicit"response
which

Particular

re

Considered
It is believed

the

that
which

hypotheses upon
description is based
choice

most

applicable to
However, the

situations.
of

derivation

mathematical

model

be
can
hypotheses
readily applied to experimental data
additional
without
assumptions is
for a cerachieved
tain
more
conveniently
which

these

from

may,

in

taken

to

certain

equivalent to
usually classified

there

VTEs

are

the
as

partial
VTE.

situations in which

some

observed

not

are

be

circumstances,

be

response

But

an

with

specificinterpretationis given
term
"implicit response." It

underlying

the stochastic

are

the

to

of

occur

kind

the

is associated.

the parameter

No
Sihiations Which

Choice

tJierc will

At),

and

would

unlikely to be present. In these


the "implicit" response
be
cases
may
regarded as a tendency to make a given
or
might perhaps be given
response,
some
physiologicalinterpretation.
seem

probabilities of the various


"implicit" responses
ring
occurconsists of experiments where
edge
knowlconsidered to be independent
are
of the outcome
of
correctness
or
of one
another.
So that for given conditions,
is not
available to the 5
a
response
of
each
kind
implicitresponses
until after the choice has been made.
intervals unare
affected
appearing at random
junctive
Thus, for example, most
ordinarj^disof other
by the appearance
time studies are
reaction
not
It follows from
implicit responses.
considered because the 5 in these experiments
the firstassumption that the distribution
class

of

situations.

This

The

class

kinds

match

can
a

known

the

requirement.

class

of situations
is not

considered

of

with

of the

Nevertheless,
which

trivial

can

one.

be
is
It

others

tion
{a) Discriminaventional
conexperiments, including most
psychophysical procedures
this category,
{b)Studies of preference
and
conflict,
(c) Investigations
learning in choice situations.

includes among

in

his response

The

next

of the

section

of

is

paper

mainly concerned with the events supposed


be
to
taking place during a
singleexperimental trial.

intervals

between

sive
succes-

implicitresponses
of a given kind
and
is
determined
tirely
enexponential
by the response
parameter

see
[e.g.,

Feller, 1950,

p.

220].

that
Assumption 2. It is assumed
final choice response is made when a
run
of K implicitresponses of a given
kind appears, this run
rupted
being unintersponses
of implicit reby occurrences
either
other
kinds.
K
of
may
be assumed
to take a particularvalue
be regarded as a further paramor can
eter,

which

can

be estimated

from

perimental
ex-

data.
The

Stochastic

Model

Assumption
The

notions

is based

which

upon

the model

simple and involve


only two assumptions :
Assumption 1. It is first assumed
that, for given stimulus and organismic
conditions,there is associated with each
possiblechoice response a singleparameter.
are

This

very

parameter

probabilitythat

in

determines
small

interval

the

of

before.

Mueller

has been
employed
(1950) has used this

tween
approach to describe the intervals betioning
bar-pressesin an operant condiwhere
one
only
experiment
response

is involved.

For

the

same

"
(1950) and Bush
used an
Mosteller
sumption
as(1951) have
which
is very
similar, the
only difiference being that their models

situation, Estes

J.

R.

used

rather

discontinuous

distribution

continuous

than

265

AUDLEY

of responses

keep the expositionas brief

in order to
as

consideration
possible,

in this paper
to situations
involving

Christie (1952) in discussing will be limited


choice
between
of response
a
abilities
prob-

in time.

only

determination

the

in

discrimination

the

used

has

experiment,
assumption

same

for situations where

two

are

responses

Finally,the author of the


(Audley: 1957, 1958)
present paper
notions
has previously used the same

competing.

combine

to

response

in

Bush

1.

in

an

been

has

it

all

and

analysis of

that

(1955),

Mosteller

with

in

times

response

model

continuous

amples,
ex-

assumed

situation, have

runway

these

considered
K

"

There

several

are

Firstly,when

of

for all alternatives

identicallythe

same,

be

shown

and

will be

and

have

elaborated

possible

two

and

called A

and

the

Let
the

with

jS. Assumption

and

that

means

occurring
+ A/) is

an

{t,t

interval

time

small

parameters
be
responses

two

of
p{a),the probability
in

be

will

responses

will be labelled

respectively.

2.

B, and implicitresponses

kinds

of the two
b

situation with K

given by:

p{a)

aM

[la]

1.

p{h)

^M

[lb]

Similarly

1,

times

response

can

presented.
general case

more

two-choice

sults
Re-

be

can

distributions

be

which

reasons

for assuming that K "


i^
1, but not if i^ "

be advanced

the

in

is relatively
only this

that

so

elsewhere.

associated

to

the

Furthermore,
2,

will

derived

1, but

generalizationdoes not appear


have
been previously employed
situation involving choice.
this

for the
been

havior.The
learning be-

individual

However,

sponse
re-

simple when
special case

tives,
alterna-

two

problem

mathematical

scription The
stochastic de-

in a
probabilities

of

and

times

i.e.,m

2.

are

plicit
im-

an

but

of either kind

response

ponential
ex-

of

probabilityp {a or h)

The

to

not

terval
Audley, 1958). both, occurring in the small time inis
of these
Neither
properties is in
with
experimental findings.
agreement
p {a or b)=p (a)+p (b) 2p {a)p(b)
Secondly, when K " 1, the sequence
{a-\-0)M-2a^{MY
fore
of "implicit"responses
occurring be-

(e.g.,see

final

ofifer a

of

various sequences

"implicit"choice

suggests an approach to descriptorsof


the second kind.
For example, "perfect

might be
identified with sequences
consistingof
of one
kind only.
"implicit"responses
confidence"

Derivation
No

Hence

including VTE's
havior.
the description of choice bethe
of
classification
Thirdly,
if
of

possible means
within

is made

choice

in

p{aorb)
of order

terms

This

the distribution

choice

of the Stochastic

Model

further assumptions are required


of the model, which

the

to

{a-\-^)^t
(A^^

continuous

transition

case

when

of implicitresponses

in time

follows that of

[Ic]

ignored.

are

possibleif

becomes

is made

Poisson

ess
proc-

220).
Therefore
probability,pin, t),of
in the
obtaining n implicit responses
time interval
(o,t) is (e.g.,again see
Feller, 1950, p. 221):
Feller, 1950,

(e.g.,see

p.

the

in the derivation
can

any

be applied

number,

m,

to

situations

of choices.

involving
However,

p{n,t)

(g + ,g)n^"e-("+g)^
[2J

266

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

series of sequence
p{o, t),
particularthe probability,
probabilities.
of
obtaining no implicit response
Thus,
either kind in time t is given by :
Pa
p^-\-p\ + p\ + p^q^ +
[5]
e-("+^)'
p(o,t)
[3]
Whence, simplifying,and substituting
for p and q from Equations 4a and 4b
that
the
The
first
Pa,
probability:
is
is
to occur
a
implicitresponse
an
In

nite

of

/*

=1

Pa

a2[a + 2/3]
[a + ^][(a + 0Y

00

"

t=o

/3T2a -f /3]

[4a]

[" + /3][(a+ ^y
Similarly,for implicitb

Equation 6a
following form

responses

(8
+

Since

be

may

of

occurrences

follow

implicit

Poisson

[4b]

Pa

written

in

the

[(g + ^y

say,

[6b]

aiS]

say,

[6a]

Similarly
a

Pb

"

a^-\

p(o, t)adt

"

/32]

a-f/3'[(a+ |S)2-a^]

sponses
re-

process.

that

SO

when

/S,Pa

"

and

"

Equations 4a and 4b also give the


/3
probabilitythat, startingat any given
+ /3
a
the next
to
implicitresponse
moment,
the
Thus
the difference between
5 respectively.
will be an
a
occur
or
probabilitiesof the various
implicit
Therefore, ignoring for the moment
in
occurring is accentuated
responses
vals
questions concerning the time interthe expressionsfor the probabilities
of
successive
between
sponses,
implicit rechoice
overt

the
to
a

final choice

of

sequence

events

of independent

sequence

trials,with
Pb, of the

increases

be treated

can

as

that

there

binomial

two

4a

and

is believed

4b.

Probability,Pa, That
is an
A Response

the

possible sequences

which

Final

Choice

with
can

For

the

of

occurrence

an
"

between
a's

occur.

and

b, until

The

two

and

clearly: p'^,p'^q,p^q, p^q^ etc.


over-all probability.Pa, that the
an

A, is the

sum

The
final

of this infi-

be

Trial

b, with

them.

This
which

property

and

VTEs,

Error

obtained.
to

the moments

of VTEs

distribution

cessive
suc-

early members

determine

of the

can

here

Attention
the

choice

any

of the

readily be
will be

fined
con-

of VTEs

number

mean

preceding (a)

of this class of sequences


: aa,
baa,
are
The respectiveprobabaa, babaa, etc.
abilities
The
of these various
is
sequences

is

in the

identify alternating appearances


"implicit" responses,

we

minate
ter-

implies

organisms exhibit.

Vicarious
If

K
be easily classified when
2.
all be simple alternations
they must

choice

than

to

ation
accentu-

and

certainty in the
underlying

more

which

processes

many

The

with

the

Equations
The

is

choices

overt

Pa and
probabilities,
given by
types of event

The

responses.

ing
lead-

(b)

ticular
par-

choice.
Mean

Any
There

of

Number

VTEs

Preceding

Choice, V
are

no

VTEs

of implicitresponses

if the sequence
is

aa

or

bb.

J.

R.

is 1 VTE

There
baa

267

AUDLEY

if the sequence

is

if the sequence

is

Va

yields the following results

choice

abb.

or

There

2 VTEs

are

2afi

/?

(a + /3)2

a^

abaa

babb, and

or

Dividing the

so

on.

with

into those

responses

of impHcit

sequences

odd

an

with

those

number

even

an

(a + j8)2

of

following probabiHties are


VTEs,
Since
found (lettingP(F
w) be the probability
of obtaining n VTEs) :

P(V

p' + q'

2)

p'q + pq'

4)

pY

P{V

1)

p'q + pq'

P(F

3)

p3g2 ^ p2g3

P{V

5)

p'q' -i-p'q*

given

moment,

Pb, Va

"

SP{V

"""

latencies

algebraicmanipulation
some
again substituting for p and q
from Equation 4a and 4b.

La and

[7]

{a + /3)2

a^

Let

then

be

7 may

Equation

re-

q:

written

as

P{a, t) be

time

at

If 7

sideration
con-

the

mean

and
all responses
B
and
for A

Latencies

Mean

3a0
=

of

the
sponses
re-

and

Lb

respectively.
The

all the

distribution

separately.La

taken

after

and

Choice

Here, however,

latency, L, of
mean

and

Final

will be limited to

3) +

of

time

of the

2)

i.e.,if

possibleto determine

It is

2P{V

1) +

VTEs
is dominant

Vb.

"

Distribution

Time

the

on

fewer

response

final responses.

P{V

be

would

any

moments

that

seen

at

The

Now

can

which

etc.

respec-

precedingthe
Pa

re-

p
^+2

be

there

average

be

may

iS

tively, it

etc.

/3

and ^

as

-+2

PY

0)

2a
1

written
=

and

-f 2/3

P(V

2a

a^

13

the

p{V

[8b]

=
-

and

2/3

2aj8
Vb

ber
num-

[8a]

/,

for

and

sponses,
Re-

Lb

probabilitythat,

the

a's

consecutive

two

no

and

or

last

the

that

appeared,
Let
an
a.
was
implicit response
P{a, t;n) be the probabilitythat, at
6's have

37

Line

7)==

have

appeared,

there

was
response
been
have

/,

no

a's

consecutive

two

or

b's

(1 +

V is dependent only on the ratio


maximum
becomes
of /3 to a, and
a
Therefore
when
1,
/3.
i.e.,
a
7

Thus

responses.

and

that
an

plicit
the last imalso that

a, and

exactly

implicit

Thus

00

the number
when

Pa

be

would

of VTEs

Pb

mum
maxi-

P{a,t)

h-

and

of VTEs
Responses, Va

Preceding

Number

Mean
B

and

P{a,t;n)

and

Vb
the

an

and

the

method

combined.
Let

mean

P{a,t;n), Equation 2
employed to find Pa

determine

are

Separate consideration of
number
of VTEs
preceding

E
n=l

To
The

be the probabilitythat

P{a;n)

sequence

of

events

ends

with

an

a,

268

READINGS

consecutive

two

no

as

IN

b's

or

PSYCHOLOGY

MATHEMATICAL

having

Similarly it

be

may

Clearly,

occurred.

determined

^.

,_

that

,"

-l)
P(^0=e-(-+^)'[(^
^

P(a;l)

P(a;2)

P{a;Z)

V^l

JJ M

W + ^y
Now
etc.
,

(a + ^y

P(a, t)atdt
being respectivehprobabilities

these

with

associated

aba,

the

a,

t=o

2(a + /3)

ba,

etc.

(a + ^y

P{n,t). P{a\n),
P(a,t;n)
2
gives P{n, t),so that
Equation

Now
and

sequences;

P(o, Oa^^

S=o

a/3

i8
+

(a + /?)("+

[10a]

2/3)

P{a,t; 1)
and

similarly

P(l,t)-P{a; 1)

2(a + i3)
Ln

a/3

a-\- (3
+

P{a,t;2)

By
al3

kind

same

latency for

a^th--{a+")t

argument
that the mean

all responses,

is given

L,

2(a+i3)''+a/3
L

[a+^][(a+^)2-ai8]

a+/3

3al3
-{a+")l

P{a,t;2")

[11.]

[a+/3][(a+/3)2-a/3]

a^^t^e3!

Returning

and

10a

Equations

to

Hence

/3

be

10b it can
=

by

Similarly

PiaJ)

it

2!

etc.

of

be demonstrated

may

{a + ^y

the

[10b]

(a + /3)(2a+ ^)

V.

that

seen

(a+^)(a + 2^)

P{a,t;n)
and

(a+^)(2a+/3)

be written

may

as

1
and

1!

re-

2)
2) (a+^)(^
(a+^)(^
+

"*"

which,

upon

3l

"^

spectively. Thus
will, on

simplification,
gives
shorter

choice

i.e.,if Pa
2

~V

response

time

Pb, La

"

In order

the
the

to compare

time

dominant

than
"

the

other,

Lb.
the

theoretical

distribution to observed

data, the probabilityP(0, t)of


response

sponse
re-

have

average,

having occurred

by

no

time

final
/ is

J.

R.

is

also given. This

269

AUDLEY

situation, then generallyspeaking

clearly

ical
would

one

P(0, 0

P{o, i) +

P{a, /) + P{b, t)

of

P{o, t) is given by Equation 3 and


7
{a
{a,t)and P{b,t) by Equations 9a and
9b so that, upon
some
simplification,
7

the

T7^

"

the

probability

"

-rrz

Hence, the probability,Pc,

being

for this type

correct

Pc
and

Model

vance
present, it is only possibleto ad-

At

is

[13]

^2

the

Descriptors of

Kind

Second

fident
con-

i.e.,choosing A,

a2 +
The

of

given by

,_

,_

bb,

/3)2
"choice,"

a+/3

of

/3)'

{a +

of

/3. The
ability
probwould
be
aa

"

sequence

and

"

"

expect

speculations concerning
dence"
as
"degree of confiin the correctness
of a given
worth
choice.
Nevertheless, it seems
there
these
since
considering
appears

Comparing this probabilitywith the


overall probability of an
A response,
Pa given by Equation 6a,

some

variables

be

to

such

definite relation between

Pc-P.

a2(a+ 2i3)
a2+/32

[a+/3][(a+/3)2-a/3]
a^^^ja-^)

the
~

kind of descriptorand the


indices
of choice
conventional
second

(1911),whose

Henmon

be

will

in

considered

havior.
be-

paper

detail

more

that
choices
regarded
later, showed
with
confidence
are
by an 5
generally
than others.
accurate
quicker and more

This

result

discrimination

psychophysical
where

definite

in

demonstrated

was

choice

existed.

be

"

^ and

Since

only

Pc

"

Pa-

for these "confident"

implicitresponses

two

responses

fore
be-

occur

final choice, it is clear that

mean

hence

positivewhen

14 is

Clearly,Equation

tion
situa-

correct

[a2+/32][a+j8][(a+/3)=^-a/3]
[14]

more

is shorter

time

response

their
the

than

This

time.

over-all average
response
approach consists essentiallyin ing
equat"degree of confidence" with some

possibleways
of the reciprocalof the number
tributed function
might be atof VTEs
The
preceding the final choice.
to a particularchoice.
The second suggested approach to
tion
classificafirst of these involves some
is based
plicit judgmental confidence
of imof the various
upon
sequences
sponse,
the fact that these appraisalsof a repreceding a final
responses
There

choice.

vacillation

no

bb, might be regarded

confident"

than

large number
abababaa.

kind

all,such

at

as

sequences

It will be shown
sequence

suppose

be the correct

the incorrect choice in

has

the

data.
and

psychophys-

implicit responses

with
to

as

that this

itself. Degree

response

of confidence, therefore, might be

involving a

propertiesrequired by Henmon's
For,

"more

of vacillations, such

of "confident"

after the

as

low
instructions,fol-

normal

under

which

For example, sequences

involve
or

two

"confidence"

in which

aa,

to

seem

response

has

response
occurs

in

after

occur

has

been

the

time

of confidence

might be taken

to

overt

an

made,
before
is

lead

choice

If,after

occurred.

an

further
the

ment
state-

produced,
to

sociated
as-

tinuing
con-

greater

this
con-

270

fidence than ifnothing

Indeed, it might be

or

appeared.
develop

tion
of the

making the primary


choice
and
mate
giving an estiresponse
of
confidence
from
for degree
this kind of assumption.
Other approaches to the second kind
of descriptorare undoubtedly possible

under

another
be

these

assumed

This

data

where

it

that

there
in

can

one

reasonably
are

no

tematic
sys-

6"s behavior.

an

therefore

can

trials

closely resemble

changes

various

hypotheses quite
predict how often

be

regarded

further

special assumptions.
before examining Henmon

The

with

between

Properties

of

that

set

of very

which

simple

might be expected

which

observed

variables
In

situation.

it is not

it

is
sumptions
as-

in

among

choice

exposition of this
possible to examine, in
an

of the model in
detail,the success
describing the results of experiments
which
relevant.
For one
are
thing,
only the particularcase arising when
K
in
2 has been presented,whereas
be
it
t
practice may
more
profitableo
i^ as
treat
a
Also, the
parameter.
far presented is concerned
so
argument
with the events
at
supposed to occur
The
a
single experimental trial.
in which
the model
is applied
manner
to
a
experimental data based upon
of trials will depend
number
very

's

sults,
re-

exhibit

the model

seems

be examined

cannot

preferableto

seems

about

test

here

hypotheses

vidual
indirelations upon
A brief argument
for this

functional
data.

derive relations

to

to

about
empirical evidence
choice behavior in general.
In effectinga general appraisal of
is hindered
the model, one
by the
general lack of individual results in
the experimental literature. For reasons

Model

of this paper

be used

can

in which

manner

Data

principal aim

worthwhile

seems

ever,
How-

match

to

the

the

Empirical

AND

show

the

dence
level of confi-

each

Agreement

The

it

be determined.

can

kind

and

the conditions

individual

portant as appropriate for testing the model


The impresent scheme.
without there being any need to make
point is that it is possible to

associated

the

which

conducted

were

easily. They each


a
given level of confidence would be
Also
the expected distribution
employed.
kind
of
the
first
of descriptors

to

(1911), in which

Henmon

the

within

test

of quantitative evidence
will be
confined
to
mainly
an
experiment by

possibleto

between

times

PSYCHOLOGY

MATHEMATICAL

distribution

for the

model

IN

READINGS

has been presented by


(1955) and for the study of
by Audley and
learning behavior
ferred
Jonckheere (1956). The reader is reof view

point
Bakan

to

any

these papers

However,
taken

stand
clear

on

for further

tails.
de-

irrespectiveof the
this question, it is

is concerned
present model
with individual results and that

that

the

much

upon

the way

trials resemble
may

be actual

conditions
may
upon

be

from
a

in which

one

separate
There
another.

variations

in stimulus

trial to trial,or

direct dependence

earlier

trials,as
For

in

this reason,

able.
generally availthe following
For this reason,
mental
comparison of the model with experievidence is largelyqualitative,
although, given appropriate data,
quantitative comparisons would have

such

results

been

possible.

not

Psychophysical Discrimination
In

there

of later

are

tions
Situa-

physical
considering results from psychoexperiments, say using the

constant

method,

it is necessary

to

separately the comparison of


periments.consider
learning exconsidera-

each

variable with

the standard.

This

272

stimulus

conditions

the

are

Nevertheless,

all trials.

for

same

general
predictionscan be advanced.
Here, "degree of confidence" will be
function of the
equated with some
number
of VTEs
reciprocalof the
ceding
preVTEs

of

can,

course,

number

range

from

of
zero

infinity. Generally speaking,confidence

to

is rated

scale from

some

upon

unity. Let C, be the degree of


with
confidence
associated
a
given
of VTEs
choice, and, V, the number
ing
Determinpreceding this choice act.
to

zero

suitable relation between

would, in fact,be one of the experimental


problems suggested by the
For the moment,
present approach.
however, it will be assumed
that.

V+
that when

1
1 ; and

0.
00, C
It will be recalled from
=

when

with

concerned

If the stimulus

VTEs

section

that the

number

of these will,when
less than the number

i^

mean

2, be

see

e.g.,

varied

are

diff^erentsets of trials,
for
as
in
the
method
discussed
constant
example
in the previous section, general
conclusions

are
again possible. For
discussingEquation 7, it was shown

in

that

the

number

mean

of VTEs

the ratio of

only upon
Again assuming

that

pends
de-

/S.

to

(a -+-j8) is

constant,

proximatel
ap-

be

would

roughly symmetrical function of the


magnitude of the variable, having a
maximum

the

at

PSE.

Thus

the

degree of confidence,C, would


ing
roughly U shaped function hav-

average
a

minimum

maximum

the

at

been

has

time

PSE.

Since

shown
PSE

the

at

have

to

and

crease
de-

to

side of this point,


again vary inversely.
This agrees
with experimental data
(seeGuilford, 1954).
either

upon

the

conditions

dence
confi-

between

choice

0, C

between

judgment time,
Guilford
again
(1954).

[16]

relation

and

be

so

of the

C and

This hyperbolic function is in ment


agreewith
tions
experimental determina-

some

The

final choice.

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

and

would

of implicit
choice.
final
preceding
a
responses
Preferenceand ConflictSituations
be
Now
it can
demonstrated,
easily
In this kind of situation,a number
using Equation Ic, that the mean
of objectsare
paired and the subject
choice time when
n implicitresponses
makes
choice indicating the prea
ferred
Tn, is given by
occur,
object of each pair. For any
given pair of objects, say A and B,
T
two

Whence,

choice

is

possible to

time

T,

the
n

is eliminated

17, it

since

[17]

2, and

"

from

express

the

V,
to

given by
V

-\-2

{a-\-^)C

+
a

|3

To.

[19]

of ence
preferthere are

measure

Because

of objects, it is more
venient
conto label the r objectspresented
the subject as Xi, and to let the
associated

"absolute

preference" for each,

the

(i

1, 2,

"

"

"

r).

equations will
and

say

aj

the

^'th and

This, of

ak,

now

with

kind

parameter

ai

Substituting for V from Equation 16


and adding an arbitrary constant.
To,
choice time possible,
for the minimum

be taken

/3 can

number

mean

of

and

represent some
for A and B.

to

Equation

function

as

cause
be-

parameters

The

and

/S of

for the comparison

is to

be

be replaced by,

jth. objects,Xj and

course,

of

make

of
Xk.

the very

strong assumption that the a/s

are

in-

J.

R.

dependent of the particularcomparison


in which they are involved.
This
be
could
tested
assumption
readily
by
using the model appropriately, and is
accepted here only in order to simplify
notation.
The results of the following
would
be qualitativelythe
argument
if

even

same,

contextual

there

effects

to

comparison.
Variation

in choice

comparisons.

time

The

among

of

set

i be

a paired comparison
usually be ranked.
individual's ranking of

object,

X^"X2"
This

on.

"

a,

that

so

we

write

may

"Xi"Xi+,"
X,- is preferred to X2
"

meaning
"

an

"

"

"

that

means

tti+i

"

"

"

"

"

ai

and

0:2

"

Consider

ar.

can

is given by Equation
be rewritten as
now

11, and

that

the

scale

Number

"

"

be

any

this

have

tained.
main-

the

was

lute
abso-

an

basis, so

relative

values

of VTEs
It

"

would

than

by
comparisons.

should

be

inclusion

of

affected
unnew

for differentcomparisons.
shown, in discussing

number
Equation 7, that the mean
of
in a given situation,depends
the ratio of a
entirely upon
to
(3.
the
Using
present notation this would
VTEs

be the ratio of
and

Xk.

The

maximum

aj

to ajt,

number

when

ay

for

objectsXj

of VTEs
=

ak,

has

and

creases
de-

become

the values of the parameter


the
more
disparate. Thus

number

of VTEs

as

Lu.k)

values

so

pair of parameters, say uj and ak, and


let these be the a and /3of the earlier
choice
the mean
equations. Then
time

scale

rather

"Xr,

"

"

absence

an

be

to

the basis of

on

an

of

interestingto determine

far the assumption of


contextual
effects can

jects, the
ob-

technique, can
Let

It will be
how

If the assumption turns


out
approximately true, then the
parameters, at, would provide a means
of scaling the stimulus objects for a
In essence,
fact, given individual.
such an
each
approach would resemble that adopted
by Bradley and Terry (1952), but
ferentwould have the added advantage that
dif-

in

were

peculiar

273

AUDLEY

"

ocj

ak

upon

SajUk
+ oikY
[aj -\-a/t][(ay
"

ClearlyL(^j,k)
depends

upon

aya*]
two

[20]

things ;

the

the

should

depend

tirely
en-

differences in preference

and not upon


the general level of
preference for the two paired objects.
Thus
for adjacent objects, Xi and
of VTEs
before a
Xi+1, the number

of the parameters
(ay + a^)
sum
final choice will not rise with
choice
and, secondly, the product of the rameters,
patime
from
as one
proceeds
preferred to
Other
things being
ujUk.
nonpreferred objects. This is slightly
equal, the choice time will decrease

(ay -f-ak) increases.


Again, with
will increase
(ay -j-ah) constant, L(j,k)
with the product, reaching a maximum
as

when

ay

aA:.

Choice

time

will therefore

(o) depend upon the general level


preferencefor objects,being quicker
for preferredobjects,(b)will be quicker
the greater the difference in preference
for the two
paired objects. This in
with
experimental finding,
agreement
for
children
choosing among
e.g.,
liquidsto drink, Barker (1942), for
aesthetic preferences,Dashiell (1937).
of

complicated by differences in "preference


distance" between
jects,
adjacent obbut the predictionis again found
with experimental
to be in agreement
evidence, e.g., see Barker (1942).
Learning in choice situations. It is
in considering learningbehavior
that
the

need

for

individual

results

is

greatest (Audley " Jonckheere, 1956).


full advantages of the present
The
variables can
approach to response
only be gained by incorporating the
for
assumption in a stochastic model

274

in which
this
learning. The
way
1, has
might be contrived, when K
already been outlined and illustrated
elsewhere
(Audley: 1957, 1958). On
the whole, therefore, the experimental
literature does not provide results in
which
enable the predictionsof
a way
be falsified,
the model
to
at a
even
level.
The
that
most
can
qualitative
=

be

done

is to show

here

that

the

be good mations
approxithe properties of learning

to

data.

particulartheory of

of course,

it would,
anchor

the theory
variables

more

by

be

possibleto

closelyto
identifying

of the choice model

parameter

sponse
re-

the

with

the ratio of
will

rise to

Pb

fall

Let

with

A, the correct
B, the incorrect

way

in which

with

and

and

response,
response.

with

vary

/3

The
ward
re-

punishment is naturally a
for investigation and
would
matter
form
condition
the
of
the
certainly
made.
be
would
prediction which
Nevertheless,it is not unreasonable to
and

that

assume

will be

some

monoincreasingfunction,and ^ some
tonic decreasing function of practice
and of punishments and rewards.

Let

supposed that the S has


strong tendency to produce

it be

first a

incorrect

choice, i.e.,a

is small

this

upon

(a -f /3), and if the levels of,


punishment and reward, are such
then

of
say
as

of this quantity,
tion
accentua-

an

of
flattening of the curve
latency as a function of practice. The
or

monotonic

decline

in

when

tencies
la-

response
an

is introduced

learning situation for the


firsttime does not counter
this prediction.
For, then, it is to be expected
that (a -f /3)will be initially
small and
the effect of increasinga, and, hence,
{a + /8)will be reinforced by the growing
difference in magnitude between
a
and /?. In originallearning,therefore,
factors work
the two
together and
decrease
in
produce the monotonic
latency.
a

number

The

be expected
a

from

of VTEs,
to be

7, is seen

function

to /8. Thus

to

rise to

13, i.e..Pa

would

maximum

Pb

VTEs

tion
Equaonly of

until

0.5, and

the

decline.
These

predictionsare

applicable
choice

to

the

situations

very
so

probably only
simple two-

far

considered.

studies,the problem
the
in
be
the
is
to
to
complicated
by
might
expected
happen
way
which
the
over-all latency L, and the latencies
relevant
being
cues
are
of A and B, La and Lb respectively. utilized by the organism and there is
In discussing Equations 10a and 10b
no
point in reviewing the controversy
relative

it

was

to

what
/3. Consider, firstly,

then

influence

the

the constancy
there will be

tency,
la-

constant,
until Pa

Superimposed

the ratio of

monotonic

over-all

|8) and

fall will be

observed

be associated

(i.e.,a

again.

into

and
properties of the model
Consider,
simple learning behavior.
for example, learning in a simple two-

dependent
(a -f ^) and

are

sum

maximum

0.5

Lb.

than

/S. The

to

La

0.5, when

L, if (a + j8)remains

an

The

choice situation.

latencies

factors,the

two

upon

exceeds

generally shorter

All of the

ing
learn-

appropriate theoretical construction.

the

will be

to disturb

Given

at

Pa, reaches and

dictions rise and


pre-

well

might

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

shown

that

the

dominant

sponse,
re-

For

over

discrimination

this

matter.

It does

however

pointing out that, in


behavior, it is very
place it will be expected that La will
probablethat there appears something
until the probbe greater than Lability like the problem of the use of the third
in psychophysical procedof making the correct
choice, category
on

will have
the
the average,
Thus in the first

shorter choice time.

seem

worthwhile

discrimination

R.

That

ures.

be

is,a distinction
between, on

necessary

hand,

definite

because

in the

done

one

This

situation.

of the parameters
may
behavior
upon

the

exert

in

an

two

be

about
be

tive
specula-

because

possiblethat both of
accounted

these

size
ence
influ-

which

few

suitable

of confidence

advanced

speculations

in

the

present

The

important point, it seems


the
the author, is that
to
general
stochastic
model
is capable of dealing
paper.

with

ways.

the

only

been

have

differences

for by

analysis of judgments

on

to

275

AUDLEY

may

occurs

has

something

is raised

point

the

which

the other hand, behavior

simply

to

seems

of choice and,

act

J.

this

kind

of issue, rather

than

the
at
Firstly, by determining
ability
probof making a particularresponse
present time.
when
"true" choice is made
a
and,
Henmon
gives the distribution of
the
ability all choice times for each individual.
probsecondly, by determining
that

"true"

choice

that

is made.

Since

Henmon's

experiment. The experiment


Henmon
conducted
(1911) is
by
of particular interest, because it provides
data
from individual 5s, in a
situation

stimulus

where

be assumed

can

to

therefore,are
concerned

conditions

be fairlyconstant

important for any model


with
the
properties of

choice behavior.
Henmon

required 5s,

horizontal

shorter

than

of the
20.3

5s

lines

was

other.

the

lines

always
respectively.

mm

in

each

of

one

of

confidence
The

in each

to

choice

largerthan

In

responses

each

indicate

time
that

are

category

their

5s

There

that

this

in

the parameters
of choice times

it was

However,

it would

matter,

the distribution

decided

that

be
from

alone.

perhaps

if
could be made
out
stronger case
the only time datum
used to estimate

model, the wrong


relatively quicker in

is

was
are

equations

if values

correct

some

appears

indication
there

is

with longer
general decline in accuracy
choice times, again predicted by the
model, there is also a slight rise in
in going from
short to
accuracy
very
It is
moderately short choice times.

the present

no

of

must,

can

to

determine

6a

This

from

tables of results,because
in

For

the

assumed
and

following
minimum

jS

were

is not

Henmon's
the data

intervals
this

possibletime
way.

was

of

best fitwas

entirelya
with

The

are

200
the

reason,

estimated

For

various

times, estimates of
the
determined, and

theoretical distribution of choice

computed.

some

before which

occur.

already grouped

be

course,

time

response

response

easy

mates
esti-

Equations

upon

for the

11.

There

in

for

chosen

was

based

mined,
be deter-

to

probability of

Accordingly

minimum

function

the

Pa,

latency.
required

mean

course

/3 are

response,

second.
and

the

of

and

milliseconds.

The

confidence.

although

of

minimum

age
aver-

the

of

the parameters

Two

are

is
for wrong
responses
for correct
choices, as

qualitative difference
a
as
examining accuracy

of time.

from

of the

usual to estimate

judgment.

second

some

addition,

is qualitativelyin ment
agreeHenmon's
data, except in

predicted by

derived

two
comparison
distributions should give further indications
the
to
as
adequacy of the
present approach to choice behavior.
In testing the goodness of fit of the

and

mm

things. Firstly,although

two

also be

can

model,

and

model
with

in

longer or
The lengths
20

were

instructed

were

this

in all details

1,000 trials,to decide whether


two

the

model

observations,

trial to trial. The

from

it succeeds

times

leading to the
then adopted. This is not
satisfactoryprocedure, but

assumed

value

to

be

2, and

with

no

276

READINGS

TABLE

observations

ignored

indication

of

These

direct

the

the

Bl and

Br

be

to

this basis

3.19 and

time

giving a
comparison
=

1.28,these

scale

13

mum
minisec.

4.28.

and

times
of response
The
1.
agreement

given

between

in Table
model

data

and

seems

be

to

reasonably good.
Concluding
the

looseness

there

whole,
in the

hypotheses

are

linked

variables.

therefore,
whether
related

and
to

It
to

in which

way

theories

contemporary

is

needs

In

indicate

try

observed

model
the

seems

to

of the

the

to

one

another

share

of
potentialities
than

be

to

is to
the

make

proach,
ap-

specific
=

Each

with

choice
to

situation
certain

have

to

be

does

important

behavior

will

and

into

count.
ac-

to

seem

properties
therefore

reasonable

doubtedly
un-

tions
unique condi-

be taken

model

the

certain

appears

these variables might not

general

intention

case

which

while,
worthdetermine

not

K
2.
arisingwhen
It is not to be expected that the two
simple assumptions will alone account
for the relations existing between
sponse
revariables in a wide diversityof
tests

But

sponse
re-

most

way,

be tested.

to

rather

many

local

this

this presentation of the

stochastic

certain

even

In

are

have
On

in

operate

havior
descriptions of choice beconsiderably simplified, but
better ways
of formulating and testing
theories
model
suggested. The
are
itself is naturally also a theory about
certain aspect of behavior, and
a
as

only

situations.

Remarks

which

situations.

pected
ex-

distributions
is

choice

such

ured
meas-

be 0.34

to

observed

the

On

sec.

Br, the

and

6.68
of

/3

For
taken

was

0.40

time

in seconds.

5s

194)

p.

possible time

about

referringto

values

for

below.

considered

are

taken

was

2,

Bl, the minimum

For

in

results

(1911, Table

Henmon's

simple laws

minimum

The

circumstances.

in calculations.

best available

the

time, it seemed

PSYCHOLOGY

MATHEMATICAL

IN

it

initial

working hypothesis. It can be tested


the
in great detail against data, and

by relatively parameters

are

of

kind

which

could

J.

R.

either

with

identified

be

R.

Bradley,

psychological

of

Methods

estimating

statistical

will

For

neither

of

method

metrika,

of

goodness

elsewhere.

model,

present

parameters

of

tests

discussed

be

fit

example,

Rev.,

of

of

one

and

the

time,

R.

R.,

solved

easily

and

D.

Cartwright,

mean

sponse
re-

L.

R.

as

behaviour

J.

Quart.

ject.
sub-

1957,

Psychol.,

exp.

S.

The

J.

criminative
dis-

of

measurement

Rev.,

Psychol.

Affective

F.

J.
determinant
J.

of

1952,

value-distances
esthetic

1937,

Psychol.,
K.

W.

EsTES,

individual

an

Amer.

174-196.

54,

judgments.
57-67.

50,

of

description

of

decision-time

of

response.

443-452.

59,

stochastic

J.

of

behaviour.

Amer.

learning

Wiley,

may

priate
appro-

REFERENCES

the

Stochastic

York:

Relation

1941,

Psychol.,

11

the

give

New

categories

the

to

Dashiell,

AuDLEY,

F.

Mosteller,

learning.

1955.

values.

parameter

chol.
Psy-

of

alternative

over-all

to

matical
mathe-

learning.

313-323.

58,

"

for

Christie,
be

F.

simple

problems.

the

Equations

I.
Bio-

comparisons.

324-345.

for

1951,

rank

cedures
pro-

probability

the

given

responses

The

designs.

Mosteller,

"

model

models

occurrence

E.

block

paired

39,

R.,

the

these

novel

any

of

1952,

R.

Bush,

Bush,
involves
For

H.

Terry,

"

incomplete

constructs.

The

and

A.,

of

analysis

physiological

or

277

AUDLEY

Toward

learning.

statistical

Psychol.

Rev.,

of

theory

1950,

57,

94-107.

9,
W.

An

theory

and

its

Wiley,

1950.

Feller,

introduction

probability

to

12-20.

R.

AuDLEY,
times

within

inclusion

The

J.

response

description

stochastic

of

of

individual

of

Attitude

S.

S.

in

1958,

Psychometrika,
J.,

JoNCKHEERE,

"

A.

Brit.

learning.

statist.

A
"

The

Skills,
R.

the

resolution

Q.

McNemar

Studies

G.

Hill,

of

5,

M.

of

Perceptual

by

New

Rev.,
of

study

In

children.

Merrill

A.

V.

Henmon,

aggregate:

211-212.

conflict

personality.
1942.

the

experimental

An

"
in

and

distinction.

1955,

P.

J.

York:

(Eds.),
Mc-

1911,

New

18,

Nat.

relation

of

the

time

Psychol.

accuracy.

Theoretical

relationships
of

measures

Acad.
R.

York:

its

186-201.

G.

some

WoODWORTH,

The
to

C.

among

Proc.

C.

A.

New

1954.

Hill,

judgment

Mueller,

Methods.

Psychometric

McGraw

87-94.

9,

general

methodological
Motor

Barker,

Graw

1956,

Psychol.,
D.

physical
psychoPsychol.,

1-37.

28,

J.

York:

Bakan,

J.

chastic
Sto-

R.

Guilford,
for

processes

to

Amer.

25-31.

23,

1917,
R.

relation

subjects.
judgment.

AuDLEY,

York:

the

George,
behaviour

learning

New

applications.

Sci.,
S.

Henry

1950,

conditioning.
36,

Experimental
Holt,

1938.

123-130.

psychology.

BUSH

R.

ROBERT

BY

FOR

MODEL

MATHEMATICAL

FREDERICK

AND

experimental arrangements
and Skinner
straightrunways

for empirical
though
development of a

models

aid the

sufficient body of titative


quaninformation has been accumulated.

science when

This accumulation

be used

can

point the direction in which

to

general;
in

to

believe the model

we

plan

we

order

later

possiblewe
correspondence between

the

and the

model

more

shall

Wherever

papers.
discuss

the adequacy of such models in their


our
quently
interim states.
Models, in turn, fre-

in

discrimination

and

test

is

the model

multiple-choice
zation
experimentsin generali-

and

problems

extend

to

describe

to

models

and

constructed

be

should

using
boxes,

in

Introduction

phenomena

HOSTELLER

University 2

Harvard

Mathematical

LEARNING

SIMPLE

being developed

one

since striking parallels


of the
though many

(2)
by
organizingand
do
exist even
experimentaldata and in basic
interpreting
premises differ. Our model is
directions for experimental discussed and
suggestingnew
developed primarilyin
research

for

necessary

for

quantitativemodels
The

most

is that of Estes

this paper

learning

situations.

attention

on

This

and

We

simple

some

shall focus

We

Fellow

Post-dcx;toral

in the

research

was

oratory
supported by the Lab-

Relations, Harvard

part of a program
Project on Mathematical
as

authors

helpful advice and


but especiallyto
R.

sity,
Univer-

Laboratory's

constant

Drs.

W.

Jenkins,

L. Solomon.

of

appeared in Psychol. Rev., 1951, 68,

313-323.

as

follows:

(1)perception

stimulus, (2) performance

response

occurrence

278

the type of
mental
called "instru-

with

concerned

learningwhich has been


conditioning" (5), "operant
behavior"
"type R conditioning"
or
tioning"
(10), and not with "classical condi(5),"Pavlovian conditioning"
"type S conditioning" (10). We
or
shall follow Sears (9) in dividing up
of

ment,
encourage-

O.

mathematical

theory.
are

persons

many

cepts.
conmon
com-

example is the particle


of modern
interpretations

the chain of events

Models.

gratefulto

are

R. R. Sears, and

of the

most

An

wave

atomic

Social Sciences.

of Social

article

and

of

of

sets

of affairs is

state

feature

models.

(2).

acquisitionand extinction

iSSRC-NRC
Natural

This

tion
contribu-

of other

in terms

new

stems

formalize association theory. Both


be
preted
re-intermodels, however,
may

learningphenomena.

recent

designedto describe

model

This

Estes' model

construct

to

concepts while
from
an
attempt

to

shall present the


mathematical

we

of

basic structure

for

ing
learn-

building.
provided by

attempts

numerous

The

as

model

of this fact is

Evidence

'

as

are

of reinforcement

terms

rich

quantityand variety of available

data

In

the branches

Among

of psychology,few

the

useful in

are

in

"

Estes

of

or

instrumental

act, (3)

an

environmental

event,

Reprinted with

permission.

280

write

to

operator in the form

our

Qp
This is

be used

(2) will

equation

basic operator and

our

"

ment
equation (2) corresponds to an increin p which is proportional
to the
maximum
p).
possibleincrement, (1
the term,
Moreover, since b is positive,
in p
bp, correspondsto a decrement
is
the
maximum
which
proportionalto
Therefore,
possibledecrement, "p.
"

"

factors

those

Continuous

Up to this point,we have discussed


of a
only the effect of the occurrence
the
of
that
probability
response
upon
Since probability must
be
response.
since in a time interval
conserved, i.e.,
h an organism will make
some
response
must
we
investigate
or
no
response,
effect of the

the

upon

response

of

occurrence

one

the probability of
In a later paper,

response.

shall discuss

this problem

in

other
anwe

detail,

the eter
parammust
but for the present purpose
we
crease
always deinclude the followingassumption. We
the probability. It is for these
that there are
conceive
two
general
that

reasons

rewrote

we

given

in the form

reward

or

operator

our

equation (2).

in

associate the event

We

of

overt.

presenting

into classes

reinforcingstimulus

other

with

the

that

in

0 when

no

we

assume

is

reward

overt

With

b,

work

the

(See
making the response.
review
[11] of the
by Solomon

required in
the

of work

influence

respects,

many

corresponds to

on
our

an

behavior.) In
a(l
term,
p),
in

increment

make

further

no

our

of reward, amount

of work,

strength of motivation, etc.


theoretical

our

experimental
values

fit.

of

and

B is not

it is neither rewarded

which

on

any

response

probapunished. Since the total bility


of all responses
be
must
unity,

nor

data,

we

b which

In other words,

our

In

results
will

paring
com-

with

lost

at the

by
by

response

be compensated

must

corresponding

probabilityof

in

the

loss

or

non-overt

gain
sponses.
re-

tant
assumption is imporof
experiments
analysis

This
in

which

the
a

use

example.
singleclass of
for study, but
and

can

later

do

paper

such

In

other

the

Skinner

overt

We

or

responses

defer until

discussion
two

box,

experiments a
is singled out

responses

occur.

in which
are

or

runway

for

choose

give the best


model

the probabilitygained

it follows that
citatory
"ex-

parameters,
b, to experimental variables such

amount

is

punished, then

"

potential" in Hull's theory


(6) and our term, "bp, corresponds to
in Hull's
increment
"inhibitory
an
potential."

and

If

and

mutually exclusive
changed.
Nevertheless, the probability of that
A is changed after an occurrence
response

or

In this paper,
we
relate
to
attempt

divided
sub-

etc.

occurs

nor

probability of

the

overt

associate events

we

punishment and

as

response

non-

are

A, B, C,
A

rewarded

neither

given

experimental extinction.

the parameter
such

a, and

parameter

and

of responses,
overt
The overt
responses

kinds

an

as

Reinforcementand
Extinction

factors which

b those

as

for particularconditions
particularorganisms.

and

with

probabilityand

the

with the parameter a


which
always increase

associate

can

of parameters

1,

also lie
the parameters a and b must
Since a is positive,
0 and 1.
between
that the term,
p), of
a(l
see
we

we

only with
present time is concerned
the form of conditioning and extinction
with
the
not
precisevalues
curves,

tain
main-

0 and

probabilitybetween

the

for

the cornerstone

as

theoretical development. To

our

(2)

hp.

a{\-p)-

p +

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

more

of

ments
experi-

responses

reinforced differentially.
mathematical
the aid of our
With

ROBERT

BUSH

R.

operator of equation (2) we

AND

Q of equation (2) with

now

may

change in the
progressive
of
ment
probability a response in an experisuch as the Graham-Gagn^ runway
(3) or Skinner box (10) in which

describe the

the

environmental

same

follow

events

We

of the response.
only apply our operator

each
need

occurrence

to

set

equal to

zero:

Ep

It follows

bp

(i

b)p.

(5)

directlythat if we apply
for n
successively

this operator E to p
have
times we

peatedly
re-

E^p

of the

initial value

some

281

MOSTELLER

FREDERICK

(1

(6)

byp.

Each
applicationof This equationthen describes a curve
the operator correspondsto one occurrence
extinction.
of experimental
sequent
and the subof the response
The
environmental
events.
Latent Time, and Rate
Probability,
p.
probability

tions
algebra involved in these manipulaisstraightforward.For example,
if we
apply Q to p twice, we have
Q'P

QiQp)

a-\- (I

b)Qp

ous
Before the above results on continuand
extinction
reinforcement
can
be compared with empirical results,
we

first establish

must

between
=

our

(1 -a-")

a+

relationships
perimental
probability,
p, and exsuch

measures

[a + (1

b)pj

(3)

time and rate of


to

Moreover, it

be

may

that if we

apply Q

times, we

have

to

readilyshown
n
p successively

(1

X
Provided

and

are

both

not

both unity,the quantity(1


tends to an asymptotic value
"

as

increases.

(4)

b)\

zero

or

b)"

"

of

zero

Therefore, Q"p

then

that the response


or class of responses
first is p.
will
studied
occur
being
Since

have

we

already assumed

that

of

other

occurrences

affect p, one
may
of
number
the
calculate
mean
easily
the
will
before
which
occur
responses
do

not

limitingvalue of a/ (a + b)
becomes
large. Equation (4)
of acquisition.
describes a curve
response

no

response

we

have

assumed
is

reward
occurs,

we

that

given
may

after

describe

extinction trial by a special


operator
E which is equivalentto our operator

mean

number

ing
includwill occur,
of responses
the one beingstudied,is simply \lp.

which

were

all

that

assumed

In that derivation it was


the responses

independent of
abilities
prob-

one

that transition
another, i.e.,

the

same

between

the
an

that the

shown

and

Since

being studied takes place.


tion
(2) has presented this calcula-

Estes

either zero or unity. For


necessarily
ii
b (speakingroughly
a
example,
of rethis impliesthat the measures
ward
the
ultimate
and work are equal)
in time h of
probabilityof occurrence
the response
studied
is 0.5.
being

purpose,
to be a

non-reinforced

It should be noticed that the totic


asympvalue of the probabilityis not

when

order

model.

pendent
of responses which are indeof one
another.
(For this
consider doing "nothing"
we
response.)The probability

proaches
apresponses

as

have

simple and useful model is the one


described by Estes (2). Let the activity
of an organism be described by a

(iTT-.-^)

a-f "

must

sequence

(2-/,=

responding. In

this, we

do

latent

as

for all

is a bold
think

of overt

follow

one

us

that

one

pairsof responses are


pairs. This assumption
indeed (itis easy to

responses

that

cannot

another),but it appears to
other assumption would
any

282

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

of the
of Grindley (4) on chicks and the data
require a detailed specification
of
in
each
perimental
exCrespi (1) on white rats.
possible
responses
many
Since
ered.
considequation (7) is an expression
arrangement being
have
time between
and
for the mean
Frick
the end of
[8]
(Miller
of the type being studied
attempted such an analysis for a parresponse
ticular one

experiment.)
that

assumed
the

performance.

and

The

the

which

of responses
"trial":
L

time,

latent

mean

then, is simply h times

mean

the
a

(7)

of

from

some

then

each

strength of

drive

likely

pressed
ex-

in terms

latent

time

as

of trials.

expression is

of the

function

It turns
identical

At

If the increments
is

latent

mean

an

follows:

its

upon

time, L,
of the probability,p,
by equation (7),while this probability
is given in terms
of
of the number
Hence
we
trials,n, by equation (4).
obtain an expressionfor the mean
may
The

of the response
in i
increment

occurrence

the organism

on

motivation.

or

measured

responses,

arbitrary starting point,

time, h, required for each response

depend, of course,
and
involved
very

culate
cal-

now

may

of responding in

being studied adds


as

will

of the

If i reparrangement.
resents
time required for the

mean

occurrence

response

we

rate

mean

Skinner-box
the

on

occur

of the next

being studied,

type

ber
num-

P
The

the end

of time, h, for its

amount

same

requires

response

every

further

is

It

out
to

ber
num-

this

that

equation (4)

we

obtain

small,
sufficiently

are

write

them

for the

mean

may

as

differentials and

rate

of

responding

dn

l/h.

We

"activity level"

and

where

co

is the

maximum

which

ences
(2) except for differnotation.
T in
(Estes uses

shall call

rate

when

occurs

(10)

Oip,

It

the

by definition "
of responding
=

\ obtains.

of Estes' paper
in

place of our n; our use of a difference


equation rather than of a differential
a
b)
equation gives us the term (1
of
Estes'
Estes
fitted
instead
e~^.)
his equation to the data
of Graham
and
results
differ
(3). Our
Gagne
"

from

Estes'

in

"

respect, however:
latent time
in

one

the asymptotic mean


is simply h, while
Estes' model

we

obtain

-(^)

L.

This
mean

amount

equation suggests that


latent
time
depends
of reward

of required work,
that

two

(8)

and

and

on

since

the

final

on

the

the amount
we

have

sumed
as-

depend on those
variables,respectively. This conclusion
seems

to

agree

with

the data

The

Free-Responding Situation

free-respondingsituations,such
box experiments, one
rate of responding or
usually measures
In

as

that in Skinner

the
versus

cumulative
time.

of responses
obtain
theoretical

number
To

first
we
expressionsfor these relations,
for
obtain an
the
expression
bility
proba/" as

function

of time.

From

equation (2), we see that if the response


being studied occurs, the change
in probabilityis A^
a{\
p)
bp.
We
have already assumed
that ifother
and are not reinforced,
occur
responses
in
the
no
change
probabilityof occurrence
of the response
being studied will
Hence
the expected change in
ensue.
probability during a time interval h
is merely the change in probability
times the probability p that the re=

"

"

ROBERT

studied

being

sponse

interval

time

BUSH

R.

AND

in

occurs

that

The

of

rate

by the time

as

derivative

of

change

As

the

dt

(12)

hp\

1/^ is
where, as already defined, w
the activity level. This 'equation is
plicit
easily integrated to give p as an ex=

of time

function

Since equation

t.

(10) states that the mean


responding, dn/dt, is co times
p,

and

is

(13)

and

and

gration
inte-

oipo

(13)
have

we

of

initial rate
Fo

the

time

/ is

The

h/a.

responding at

co^o, and

long

very

let

of responses,

and

results

0 is

ours

above,

of work

between

is the

Estes'

dependence,

of the final rate

and

upon

of reward

amount

trial.

extend

analysisto give
our
expressions for rates and cumulative
during extinction. Since we
responses
have
assumed
that a
0 during extinction,
in place of equahave
we
tion
(12)
may

after

final rate

number

form as
respectively,have the same
the analogous equations derived by
Estes (2) which were
fitted by him to
data
on
a
bar-pressing habit of rats.

We

Pq{\+u)-\-[\ -^0(1+")"-""'

function

with

cumulative

per

where

linear

equation
agrees
that
the
asymptotic
says
Both
constant.
equations
(15) for rate of responding

which

rate

large,the
equation (15)approach

This

amount

H^

(15)

very

becomes

discussed

dn

/ becomes

ability The essential difference


prob-

the

after the

obtain

we

of

rate

time

time.

(14)

dp

o:p{a{l-p)

e-""')+ e-""']

exponentials in
zero

have

we

(1

bility
proba-

Writing this

h.

of

3";

(11)

with time is then this expression

rate

result is

\uit-\+ u)
log [j"o(l

"

1 +

p{a{\-p)-hp].

expected

divided

The

/.

Expected (Ap)
=

time

283

MOSTELLER

FREDERICK

(16)
dt

[
(14)
dt\t==oc~'i.+u~l
+ bla
dvA

Foo

ci)

which

when

by
Equation (13) is quite
expression obtained
by
for

similar

to

Estes

inclusion of the ratio

per

response,

hence

with

the

follow from

expression

number

do

conclusions
results
for

of responses

reinforcement

of work

amount

Estes'

b and

the

where

the

(17)

Oibpjt

we

extinction

is Ve
cope.
write equation (17) in

ma}/

form

not

dm
V

cumulative

during

tiplied
mul-

ginning
pe is the probability at the beThe rate at the
of extinction.

(2).

is obtained

beginning of
Hence

per

Oipe

It

decreases with

These

response.

An

and

dm

except

and

the

b/a.
The final rate of responding according
with a and
to equation (14),increases
of reward
hence with the amount
given
our

CO

integrated for p
gives

Ve

(18)

dt

Vebt

ous
continu-

grating
by inte-

equation (13) with respect

to

An

integration of this equation gives

for the cumulative

number

of extinc-

284

IN

READINGS

where

tion responses

^log [1 +

PSYCHOLOGY

MATHEMATICAL

have

we

VebO

let

^log
(^")(19)
to the empirical
log t,used by Skinner

is similar

result

This

K
equation m
in fitting
experimentalresponse
=

equation
(10).
gin
advantage of passingthrough the ori-

if

of

number

extinction

sponses,
re-

limit.
Thus,
m, has no upper
result is correct, and
indeed if

our

equation (21) into equation

(20),we

is

no

of extinction

sponses.
re-

all

of the time

for large values

slow

the

to

practicalpurposes,
is
the logarithmicvariation
For

however,

limit

upper

size of the "reserve"

to
justified

We

use

some

t,

arbitrary

"completion" of

criterion for the

1,
niT

tinction.
ex-

(1

responses

the

be about

5, and

assumed
ofa
in

responses

data,

unity. Values

0.026

result is shown

to
was

chosen

were

the data.

to

in the

figure.

Fixed Ratio

and

Ratio

Random

Reinforcement
In

from

guage,
psychological lan-

present day
the term

1,

m.=-log_

Ve

(20)

"fixed ratio" (7)refers


the procedure of rewarding every
in

ation
free-respondingsitu")" In a "random
2, 3,

^th response

(^

to

express

of extinction

this "total"
responses,

mr,

explicitfunction of the number


of preceding reinforcements, n.
The
in
which
only quantity
equation (20)
depends upon n is the rate, Ve, at the
If we assume
beginning of extinction.
that this rate is equal to the rate at
the end of acquisition,we
have from
equations (4) and (10)
an

shall
rates

over

now

of

numbers
types

an

"

animal

is rewarded

but the
after k responses
of responses
per reward

the average
on
actual number
varies

"

ratio" schedule,

as

"

fittingequation (23)

The

to

number

the

the ratio Fo/F/

and

0.014

tinction
ex-

estimated

was

be about

to

From

Fmax/F/

ratio

of

number

after 5, 10, 30 and

reinforcements.

90

this criterion is

wish

(23)
")"|

shall consider extinction

of extinction

now

measuring the "total"

be

number

Fol

This result may


be compared with the
data
of Williams
(12) obtained
by

of

We

rF",ax

F^ax

-7

rate
"complete" when the mean
responding V has fallen to some
value, F/. Thus, the "total"
specified
to

obtain

empiricalequation is correct,

there

it is

beginning

that the logarithmic


of equation (19)impliesthat

Skinner's

so

co^o is the rate at the


acquisition. If we now
=

substitute

be noted

total

then

Fo
of

it must.

as

It may
character
the

where

curves

has the additional

Our

and

h'

(22)

"

'max

specifiedrange.

some

We

derive expressionsfor mean


responding and cumulative
of

responses

of reinforcement

for these
schedules.

two

If

of

equation
apply our operator Q,
(2),to a probabilityp, and then apply
our
operator E, of equation (5),to Qp
1) times, we obtain
repeatedly for (^
we

"

^^
Tr

(E''-'Q)p=il-b)''-'[p+a{l-p)-bp^

at

(F^ax-

Fo)(l -a

-by

(21)

p-^a'(l-p)-b'p

(24)

ROBERT

R.

20

lO

BUSH

A/ur?76e.r
"Total"
Curve

70

00

so

/^"//7/oro"/?7ey7A3
as

a
=

0.014, F"u"x

/?

of reinforcements.

of the number

function

0.026,

5Fo, V,

Fo.

This equation is identical to our result


for continuous
reinforcement, except

where

that
=

(25)

a\\-{k-\)h^---\^a

and

We
the

a'/k replacesa
obtain

may

b'/kreplacesb.

and
a

schedule

Q operates on
the probability that E
l/k). Hence
p IS (\
change in p per response
"

symbol
equal to."

"approximately

means

the

In

approach

would

present

case

be to retain

on

and

operates
the

on

expected

is

Expected (Ap)

tQP

-f (1
After
and
from

\/k)Ep

inserted

equations (2)and (5)are


the

we
simplified,

result

(28)

p.

obtain

equation (28)

Expected (Ap)

-"

P)

bp.

(29)

b'
,

-(l

p)

This

-jp
(27)

result is identical to the

p) -bp.

in

result shown

the fixed ratio

^-(l

bility
proba-

p is l/k and

a'

Ap

lows:
fol-

the

the

ever
throughout; howthe approximationsprovide a link
The
the
with
previous discussion.
approximations on the right of these
two
justifiedif kh is
equations are
the
small compared
to unity. Now
mean
change in p per response will be
the second and third terms of equation
(24) divided by k:

primes

as

the

response,

any

that

The

result for

similar

ratio"

"random

After

exact

no

100

(12).

Williams

from

Data

of

of extinction responses
equation (23) with

number

plotted from

60

-so

AO

30

285

MOSTELLER

FREDERICK

AND

(27)

and

mate
approxi-

equation (27) for

case.

(29)

Since both tions


equahave
the same

286

form

as

we

case,

for the

write

the

result for the continuous

our

reinforcement

at

may

mean

equationidentical to equation (13),


that a is replaced by a'/k.
except
Similarly,we
the
to

of responding identical
placed
equation (14) except that a is reis
This
result
meant
by a'/k.
final rate

fixed ratio and

to both

apply

to

expression for

an

ratio schedules

comparing

In

with

dom
ran-

result

random

comparing

those under

different

ratios

of

ment
reinforce-

fixed ratio and

ratios

difficultyof

large). The

are

comparing

no

rates

of comparing rates
(unless both

ratio, nor

under

see

we

under

rates

schedules

various

does not

forcement
reinto

seem

of

of reinforcement.

the above

of

responding under continuous

an

obtain

time,

recovery

meaningfulway

once

of responding

rate

mean

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

for

be a weakness
our
model, but rather
natural consequence
of the experia
mental
portance
imthe
However,
procedure.

of these considerations hinges


equation
the orders of magnitude involved,
(14) for continuous reinforcement,we
upon
and such questions are empiricalones.
be careful about
must
equating the
^activitylevel, w, for the three cases
(continuous, fixed ratio and random
Aperiodic and Periodic Reinforcement
ratio reinforcements). Since 1/corepresents
Many experiments of recent years
the minimum
time between
mean
designed so that an animal was
were
it
includes
successive responses,
reinforced
at a rate aperiodicor periodic
both the eating time and a "recovery
in time (7). The usual procedure
time."
the
mean
By the latter we
the

asymptotic

time

or

with

rates

for the animal

necessary

to

organize
re-

itselfafter eating and get in


make
another bar press

position to
key peck.

presumably

In the fixed ratio case,


the animal
learns to look

after each

for food not


as

both
the
are

the

fore
There-

eating time

mean

time

recovery

for the continuous


ratio

In the

case.

it seems

reasonable

random

the

ratio than

mean

reward

would

To

the

was

The

analyze this situation

consider

the

or

random

be lower

of responses

for

per

Moreover,

same.

k, the

first

per

reward,
interval

mean

we

may

of

number

mean

time

to

be

equal

to

sponses
re-

the

multipliedby
responding:

rate

of

T-t:
at

conclude

for fixed ratio when

number

of time

sequence

rewards.

to occur.

to

for either fixed ratio


that

of this set is

rewarded.

dft

be
the activity level,co, would
smaller
for continuous
reinforcement

ratio, and

value

mean

arrangement

intervals between

that
than

have

the actual

as

dom
ran-

expect

similar but smaller difference

Hence,

Tn, which

"

mean

would

one

case,

"

Some

used

and

per response
less for the fixed ratio case
than
mean

"

T.

intervals,

or

k response.

every

Ti,

of time

set

peck, response
after one
of
press
which
occurs
but
ideally
case,
intervals has elapsed is
these time

in the continuous

only after

choose

is to

we

(30)

Tc^p.

Equation (29)for the expected change


is stillvalid
probability
per response

in

if we
as

consider ^ to be

now

variable

expressedby equation(30). Thus,

the time

rate

of change of p is

should
with

increase
expect that co would
the number
of responses
ward,
per reif eating time were
k.
Even

subtracted
expect
Without

out

these
a

in all

cases

arguments

we

should

to

apply.

quantitative estimate

of

(31)
f f;(l-,)-"6,l
=

With

little effort, this differential

equationmay

be

integratedfrom

0 to

288

for

continuous

all

other

(7).

However,

story.

For

it

thing,

one

it is

that

is just

this

for

easier

ent
pointed

of

work

(2)

Estes

out.

REFERENCES

of the

part

clear

seems

1.

L.

Crespi,

forcement
rein-

variation

Quantitative

and

in the

performance

to
Amer.

rat.

continuous

p.

incentive

organism

the

between

discriminate

the

and

results

experimental

with

is consistent

is

deduction

This

model

for

than

reinforcement

schedules.

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

J.

55, 467-

1942,

Psychol,

of

white

517.
.

and
handled

extinction

this

have

; we

not

2.

of

here.

effect

W.K.

Estes,

Toward

statistical

Psychol.

learning.

Rev.,

theory

1950, 57,

94-107.
3.

Summary

Graham,

C,

is

learning

of

in

of

of

the

aid

The

parameters

mathematical

with

are

the

in

appear

related

to

Grindley,
influence

operators.

which

variables

such

5.

the

as

imental
exper-

the

of

measures

time

latent

and

probability
of

rate

tween
be-

empirical

responding

defined.

are

operant

1940,

Psychol.,

26,

C.

of

the

Experiments

in

Psychol.,

1929-30,

Hilgard,

E.

reward

chickens.

young

R.,

on

Brit.

J.

173-180.

20,

D.

Marquis,

and

and

the

on

of

amount

New

learning.

Appleton-Century

G.

York

1940.

Co.,

amount

Relations

work.

C.

Conditioning

6.

and

reward

The

ous
spontane-

conditioned

exp.

learning

D.

of

M.

and

251-280.

the

response

described

are

equations

operator

in

J.

response.

4.

of

occurrence

time

small

simple

Changes

presented.

probability

for

R.

extinction,

recovery

model

mathematical

Gagn6,

and

acquisition,

C.

Hull,

L.

York:
7.

W.

Jenkins,
Partial

and

O.,

New
1943.

Stanley,

and

reinforcement:

critique.

Acquisition

of behavior.

Principles

Appleton-Century-Crofts,

C.

J.

review

and

Psychol. Bull., 1950, 47,

193-

234.

of

extinction

and

discussed
for

for

the

behavior

the

habits

simple

Skinner

are

runway

box.

8.

G.

Miller,

and

Prick,

and

F.

and

Psychol.

responses.

of

Equations

A.,

behavioristics

tical
Statis-

C.

of

sequences

Rev.,

56,

1949,

311-324.

latent

mean

number

time

derived

are

problem;

as

and

function
for

for

equations

of responding

of trial

the

the

R.

Sears,

rate

10.

Skinner,

responses

for

the

Skinner

box

is made

attempt

to

derived

are

experiments.
the

analyze

11.

An

with

various

type
the

reinforcement

experiment.
correspondence

in

Wherever
between

the

versity,
Uni-

behavior

of

isms.
organ-

Appleton-Century-

1938.

L.

R.

The

influence

Psychol.

of

Bull.,

work

1948, 45,

Williams,

S.

B.

Resistance

to

tion
extinc-

of

Skinner

pres-

function

reinforcements.

possible,
the

The

York:

behavior.

as

partial

Harvard

at

1949.

1-40.

ing
learn-

schedules

Solomon,
on

12.

process

F.

B.

numbers

time

versus

Lectures

New

Crofts,
of

R.

Summer,

runway

mean

cumulative

9.

23,

[MS.

of
J. exp.

the

number

506-521.

Received

September

21,

of

Psychol., 1938,

1950]

MODEL

FOR

STIMULUS

AND
BY

ROBERT

GENERALIZATION

DISCRIMINATION

R.

BUSH"

AND

FREDERICK

Harvard

University^

Introduction
The

and

discrimination
behavior

to

simple mechanisms
and

eralization and

of stimulus

processes

extinction

Whether

not

or

to

seem

there

can

previous

the

as

distinction

be littledoubt

how

tween
be-

defined

as

assumed

was

Stated

is a useful

our

the stimulus

tinction
ex-

tion
situa-

by the experimenter

constant.

in the
is the

that few

ther,
Furmodel

this

work

(1),where

reinforcement

learningand behavior
one,

as

discrimination.
show

the basic postulates of


on
acquisitionand

generates
fundamental

learning theory.

this

shall

we

zation
generali-

theory

of

are

MOSTELLER

simplest terms, generalization


in which
phenomenon
strength of a response

increase
an
applicationsof behavior theory
learned
in one
plies
stimulus situation imbe made
to practical problems can
in
of
increase
an
strength response
without a clear exposition of the phenomena
in a somewhat
stimulus
different
uation.
sitof generalizationand discrimination.
When
the
this
two
occurs,
It is our
impression that few
though
Alsituations are said to be similar.
crucial experiments in this area
have
several
there
intuitive
are
been reported compared with the number
notions as to what
is meant
larity,"
"simiby
of important experiments on simple
the
one
usually means
ties
properconditioning and extinction. Perhaps
which
rise
to generalization.
give
for this is that there
part of the reason
We
alternative to using the
see
no
few theoretical formulations
too
are
of generalizationas an operaamount
tional
is to say,
available.
That
we
ceive
conof degree of "simidefinition
larity."
that explicitand quantitative
In the model, however, we
theoretical structures
useful in
are
shall give another
definition of the
guiding the direction of experimental
of
but this definition
degree
similarity,
research and in suggesting the type of
will be entirely consistent
with the
data which are needed.
above-mentioned
tion.
operational definiIn this paper
describe a model,
we
based
elementary concepts of
upon
We
also wish
to clarify what
we
mathematical
set theory. This model
In
by stimulus discrimination.
mean
for
provides one
possible framework
of
all
is
the term,
sense
one
learning a
analyzing problems in stimulus genof
discrimination.
Our usage
process
in

if any

"SSRC-NRC
Natural
*

Post-doctoral

and

This

of the

in the

research

oratory process
supported by the Lab-

was

Project on

Mathematical
indebted

are

to

Models.
many

persons

for

sistance
as-

make

to

F. R.

F.

We

us

This

refer

by which

an

in

and

restricted

more

one,

to the
specifically

animal
one

learns

stimulus

to

ation
situ-

A
(or response
response
different
in
ent
differwith
"strength")
a
B

stimulus
at

the

moment

situation.

We

are

not

with, for

concerned

D.

and
R. L. Solomon.
Sheffield,
ing
gratefulto W. K. Estes for sendof his paper(2).
pre-publication
copy

also

are

We

response

and

but in particular
encouragement,
gomery,
Brush, C. L Hovland, K. C. Mont-

is

term

however.

Social Sciences.

of Social Relations, Harvard


sity,
Univerof the Laboratory's
part of a program

as

We

Fellow

article

example,
animal

appeared in Psychol. Rev.,1%1, 58,


289

the

process

which

by

learns to discriminate

413-423.

Reprinted with

an

between

permission.

290

READINGS

possibleresponses

various

IN

MATHEMATICAL

in

fixed

PSYCHOLOGY

apply it

the

to

above

described

periments.
ex-

stimulus situation.
of the more
general
prototypes
stimulus
of
generalization
problems
shall consider
and discrimination, we
As

of experiments

kinds

the followingtwo

The

shall employ some


of the elementary
notions
of mathematical
set

We

theory
(i) An animal is trained
by the
particular response,
procedure, in an

tally
experimen-

strength

insofar

in

new

same

the

possible. One then asks about


strength or probability of occurrence

is

the

in this

response
and how

it

similarityof

the

In

similar.

an

one,

response
that
response
rewarded

or

the
in

process
of rewards

effects

strength in

response

animal

the
but

one

in the

not

the other, but

learns

tually
even-

respond

to

other,

the

influence

at

or

least

in

probabilityof

of

tion
in each situathe response
of training
the number

to

of
properties

individual
such

of

introduce

of

the notion

of such

all these

sists
con-

of elements,

set

is the

of

sum

Intuitively,the

numbers.

weight associated with an element is


of the potentialimporthe measure
tance
in

element

of that

organism's behavior.

the

we

define

can

the set; the

over

influencing
More

erally,
gen-

density function

gral
inte-

is the

measure

of that function

not

that these two

consider

bridge the

To

and

reward.

of

responses,

basic

the

over

between

stimuli

shall borrow

some

gap

we

the set.

notions

of

(2).

Estes

will

of reinforcement
(The concept
close to exhausting
experiments come
role,however.)
integral
an
play
the problems classified under
elements
stimulus
that
assumed
crimination,
and disthe heading of generalization

kinds

of

but

they
model

we

do

fundamental.

are

to

we

positive
its "weight";

denote

to

ments.
ele-

If the set

set.

finite number

measure

any

final results

our

numbers
next

number

trials,with the degree of similarityof the


of
two
situations, and with the amount

do

rise

give

associate with each element

may

the

we

with

varies

We

nor

measure

to

(rates
respond with different probabilities
the
strengths). One then asks how
or

ments
ele-

and

their number.

on

not

neither
We

of the

first.

non-rewards

and

situation

stimulus

one

warded
re-

generalization,

of

does

elements

the

not

in the

less than

the

Through

either

is

in

and

The

undefined

are

serious difficultiessince

are

fined
experimentallyde-

is rewarded,

other

which

situations

stimulus

two

presented alternately

is part of the

of stimuli.

restriction

no

universe.

this situation

lack of definition of the stimulus

involve
is

animal

(ii)An

of the

rest

universe

elements

situation to the old

new

the

we

This

degree of

the

on

an

specificproperties

set of stimuli which

place

tion
situa-

stimulus

new

depends

of

with

shall denote

of this set

situation.

stimulus
with

by

entire

response,

ticular
paras

(geometrical,optical,acoustical,
tinct
etc.) is regarded as separate and dis-

experimentally defined,

it is

as

the

which

in

and

one

experimental box

Thus,

situation similar to the training

stimulus

model.

our

situation,such

occurrence.

"tested"

is then

animal

The

certain

stimulus

from

probability of

or

the

At

has

end of training,the response

define

to

forcement
rein-

usual

situation.

stimulus

defined

make

to

Model

be

described

believe

that

Thus,

the

has

been
of

signed
dethese

to
permit analysis
experiments. In the next section we
will present the major features of the

model, and

in later sections

we

shall

in

of

one

two

states

as

far

It is

exist
as

the

organism involved is concerned; since


the elements are undefined,these states
do not require definition but merely
need

speak
state

labelling. However,
of elements
as

which

shall

we
are

being "conditioned"

in
to

one

the

ROBERT

state

of

FREDERICK

291

MOSTELLER

being "non-conditioned."
particular trial or occurrence

as

On

AND

in the other

of elements

and

response,

BUSH

R.

in the

response

that

it is conceived

of the

sub-set

learning process,
organism perceives

an

stimuli

total

It is postulated that the


of the response
probabilityof occurrence
available.
in

the

to

given time interval

is equal

of the elements

in the

measure

had been

sub-set which

ditioned,
previouslycon-

of
by the measure
sub-set.
the entire
Speaking roughly,
divided

the

ratio of the

probabilityis the

of the conditioned

elements

the importance of all the


sumed
perceived. It is further as-

perceivedto
elements

that

the

conditioned

to

the

if that

response

is rewarded.

response

be wrong

conditioned

and

to

that the

suppose

ments
ele-

spatially separated in the


actual situation as Fig. 1 might suggest;
are

the

spread

smoothly

out

conditioned

notation,
of

elements

conditioned

and

any

m{ ) denotes

the

sub-set named

set or

parentheses,

and

between

where

Xr\

of equal
that
so

or

overlap
an

of

the conditioned

miS)

of the

of

and

sub-set X

in

Heuristically, this assumption of


arise from
a
equal proportions can
fluid model.
Suppose that the total
situation is represented by a vessel
containing an ideal fluid which is a

fluids

if the

Thus

in

that

the

fraction of

the

in

fraction

the

well

are

fraction of alcohol
much

will be

thimbleful

as

sponds
corre-

of the mixture

mixed, the volumetric


same

the

to

of conditioned

sub-set

thimbleful

course

vol-

C (totalset minus
set),and the volume

The

stimuli.

the

(2)

The

corresponds

alcohol

the

the

to

"

stimuli in X

ures
meas-

equal

of non-conditioned

of the sub-set

stimuli, S

tion
assump-

m(C)

m{X)

the

is

corresponds to the

of the water

cates
indi-

C (also

proportionsin the

m{Xr\ C)
P

and

the intersection of X
called set-product,meet,
X and C). We
then make

of

measure

fact, that the

of the partialvolumes.

sum

"ume

to a

where

to

contrary
of the mixture

measure

m{X)

do

alcohol

and

water

assume,

volume

(1)

be

the substances

of

m(Xr\ C)
P

which

pletely
chemically interact but are comFor discussion let
miscible.

measure

of the response

occurrence

substances

of two

non-

for the probability

then have

we

defined in the text.

are

set-theoretic

In

ones.

are

the

among

mixture
not

non-conditioned

and

the various sub-sets involved

particular trial. C is the sub-set of


set
elements
previously conditioned,X the subof S perceived on the trial. The sub-sets
in

situation is illustrated in Fig. 1.

The

It would

diagram of the singlestimulus

Set

1.

is

perceived

sub-set

Fig.

situation S with
importance

whole

the

vessel.
of

measure

will be

whole

ditioned
con-

equal

set

S,

to
as

nition
expressedby equation (2). Our defiof p is essentiallythat of Estes
he speaks of
(2) except that where
of elements, we
number
speak of the
measure

We

of the elements.
next

consider

situation which
this

general
disjunct from
In

we

new

another

set

stimulus
a

set

S'.

S' will

not

be

denote

by

the set S, i.e.,S and

S'

292

READINGS

will intersect

We

overlap

or

can

shown

as

Sn

(3)

define

now

similarityof S'

to 5

of

index

an

Generalization

The

An

animal
in

(4)

=
'

m{S')

is trained

that the

index

this definition says


of similarityof S' to

by the

of the set

measure

notation

makes

made

tacit

of

measure

clear

S'.

that

assumption
element

an

we

has been

2.

Fig.

But

whose

the set S,
been

part of the sub-set


in the second
form

tioned
condiin

shown

as

response

elements

C is

situation

the

set

S';

we

this part by Cr\ S'.


From
the discussion precedingequations

denote

is

5 to S'

the

have

independent of the set in


is
which it
measured.) Definition (4)
also gives the index
of similarityof

form

C of 5 will have
to

ments
of ele-

set

or

sub-set

also contained

the

sponse
re-

trained in the firstsituation

elements

(Our

that

situation and

then his response


strength is measured
in a similar situation.
After the animal

is the
divided

of their intersection

measure

make

to

stimulus

one

whose
In words

Problem

We
in a position to say
are
now
something about the firstexperimental
problem described in the Introduction.

by
mil)

rj(5'to S)

PSYCHOLOGY

in

the intersection by

denote

We

Fig. 2.

MATHEMATICAL

IN

(1) and
that

the

(2), we

easily see

can

probabilityof the

respjonse

occurring in S' is

as

mjCnS')
P'-

m(I)
7i{Sto S')

m{S)

(5)
m(S')
_
~

m{S)

We

now

use

the assumption of

to S).
7,(5'

m{Cr\S')
this last equation it is clear that
be
the similarityof S' to 6" may
not
In

the similarityof 5
of the
fact, if the measure
same

is not

zero,

section
inter-

indices

the two

S'.

to

as

equal

proportionsso that

From
the

(6)

are

m(Cnl)
"m(I)

m{I)
The

first equality in

follows

m(C)

from

the

part of C which

w(5)

(7)

this equation

fact that

the

only

is in S' is in the inter-

of 5 and S'
equal only if the measures
regrettablethat
equal. It seems
are
similarity,by our definition,is nonsymmetric. However, we do not care
the general assumption that
to make
of all situations are
(a) the measures
the
the
time make
and
at
same
equal
of
an
assumption that (b) measures
element

For

elements,
to

say

the
2' X

situation,

lightbulb,
in

same

2' X
say

as

ballroom.

set

FiG.

of

have

would

small

2' box,

same

it appears.

importance of

the
say

is the

in which

situation

then

be

of elements

set

or

in each

situation,
in

large

Further

this pair of assumptions, (a) and


leads to conceptual difficulties.

(b),

2.

Diagram

of

two

similar stimulus

of them.
situations after conditioning in one
situation in which
The
training occurred is
denoted
by the set S; the sub-set C of 5
represents
to

the portion of 5 which


The new
the response.

was

ditioned
con-

stimulus

situation in which the response


strength is to
is represented by the set S',
be measured
and the intersection of S' and 5 is denoted by

ROBERT

BUSH

R.

AND

Fig. 2. The
second equahty
equation (,7)is an
applicationof our assumption that the
of C is uniformly distributed
measure

section

in

shown

as

in

5 and

over

the

the entire

does

If

of

fraction

same

and

the intersection

so

now

equations (6)

m(C)

m{I)

P'

such

to

sound

physicaldimensions

lightor
as
In fact,our
intensity,frequency, etc.
model
such
that
general
no
suggests
relation

is

i.e.,that
possible,

sible
sen-

any

similarity is

of

measure

very

fore,
Thereorganism determined.
of our
the point of view
model, experiments such as those of
Hovland
stration
serve
only as a clear demonthat stimulus
generalization
such
exists. In addition, of course,
tions,
experiments provide empirical relacharacteristic of the organism
studied,between the proposed index of
sions,
similarityand various physical dimen-

much

S.

set

above

similarity defined

as

obtain

(7),we

(8)

m(S')'m(S)
"

equation (4) we

From

of C

measure

combine

we

contains

of

293

MOSTELLER

FREDERICK

note

the

that

firstratio in equation (8) is the index


of similarity of S' to S, while from
ond
equation (2) we observe that the secratio in equation (8) is merely
in S.
the probabilityp of the response

from

these relations

but

of

the scope

outside

are

model.

our

conclude, therefore, that our


this point has made
to
(9) model
p'
v(S' to S)p.
up
no
quantitative predictionsabout the
Equation (9) now
provides us with
shape of generalizationgradientswhich
the necessary
operational definition of
with
be compared
can
experiment.
the index of similarity,7j(6"to S), of
Nevertheless, the preceding analysis
The probathe set S' to the set 5.
bilities
of generalizationdoes provide us with
in S
of
the
and
p'
p
response
to discuss experiments on
a framework
be measured
and S', respectively,can
In the following
stimulus discrimination.
either directly or
ments
through measureshall extend
We

Hence

sections

of latent

of

of

rate

or

sponding
re-

model

so

to

as

(1). Therefore, with equation


experiments.
(9),we have an operational way
determining the index of similarity.
As

direct
of

draw

Any

change
where

that

does

of

consequence

our

equal proportions,we
made

conditioned

was

provided
stimuli

introduce

This

conclusion

the

change

which

been previously conditioned

In this section

had

to that

follows

sponse.
re-

from

that

the

model

word

needs

to

be

said

develop

about

the

correspondence between our result and


the experimental results such as those
of Hovland
predicts
(3). Our model
nothing about the relation of the index

some

sults
re-

later and

of the

present

show
paper

postulates used in our previous


shall examine
(1). We
paper
the step-wise change in probability of
in a single stimulus
tion
situaa
response

generates

S.

equation (9) and the fact that we have


defined our similarityindex in such a
that it is never
greater than unity.
way
A

we

that will be used

of occurrence
probability

tinction
Ex-

and

Operators

can

tion
situa-

stimulus

response

the

response,

not

in

our

we

permit analysisof such

Reinforcement

The

sumption
as-

the followinggeneral conclusion.

will reduce

of

time

We

generalize the notions


already presented as follows : Previous
of
to a particular trial or
occurrence
the
have

response,

been

conditioned.

in

question a
perceived as
to

sub-set

our

sub-set
shown

of S

On
X

oi S

the

will

trial

will be

in Fig. 1.

cording
Ac-

previous assumptions,

294

probabilityof the

the

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

is

response

P
We

m(Xr\ C)

m{C)

m{X)

m{S)

the

that

assume

will be conditioned

the response

to

result of the reward

the

will

of A

measure

of reward,

amount

motivation,

on

We

etc.

non-conditioned

and B

that A

be shown

to

and

of A

will become

oi X

of

result

the

making

sponse.
re-

to

assume
error

compared

small

are

extend

We

assumption of equal

our

proportionsso that

we

m{Br\C)

m(C)

m{A)

m{B)

m{S)

results

is part

the

to

b.

"

apply

experiment

discrimination

stimulus

on

described

in

sidered,
con-

of the

conditional

sub-set while sub-set

part of the

new

non-conditioned

Thus, the change in the

in

position to treat
scribed
experimental problem de-

now

are

the second

in the

new

is

set.
sub-

is

An

Introduction.

presentedalternatelywith

mal
anitwo

S' which

situations 5 and

stimulus

similar, i.e.,which

measure

Problem

Discrimination

The

We
sub-set A

b) and g=l"a
p^, a/(a-\shall
In the next
section we

where

(11)

the end of the trial being

at

pidg", (16)

(poo

P""-

obtain

we

the Introduction.

have

m{Ar\C)

Now

presented

initial probability
po,

an

these

that of S.)

to

have

we

the basic

Q^'po pn

this last assumption


be small if the measures

our

postulates of our
previous model which we applied to
other types of learning problems (1).
When
the operator Q is appliedn times

set-theoretic model

can

the general

model
postulated
in
a fixed
acquisitionand extinction
stimulus
situation (1). Hence, the

operator

generates

assume

disjunct. (The

are

resulting from

in

further

simplicity we

For

is identical to

operator

(15)

bp.

given and that


depend on the
the strength of

as

in

required

the work

as

This

a(l-p)-

p +

trial:

for

sub-set

that another

oi X

sub-set A

Qp

Qp effective
probability

of the next

start

(10)

now

value of

new

at

have

are

non-zero

of C is

^m{C)

[m{A)

m(A)il

C)~\
m{An
m{Bn C) (12)
m(B)p.
p)
-

This

last form

of writing equation

(12)

equalitiesgiven in
equations (10) and (11). If we then
the

from

results

let

b^m,

m.(A)
a

=
'

m{S)

m{S)

(13)

Fig.

and

divide

equation (12) through by

m{S), we have finallyfor the change


in probability:
Am(C)
^p

a{\ -p)-

m{S)

bp.

(14)

and

S'.

diagram

Set

3.

trainingin

The

S' includes

various
Set

numbered.

thus define

Q which

when

mathematical

ator
oper-

appliedto p gives

S
situations,
disjunct sub-sets are

includes

2, 4, 5, and

6.

1, 3, 5, and
The

6;

intersection I

is denoted
by 5 and 6. T, the complement
plement
of / in S, is shown
by 1 and 3; T', the comof / in S', is shown
by 2 and 4. C,
the conditioned sub-set in S, is representedby

6, while the conditioned sub-set in S',


Tc is denoted by 3,
is representedby 4 and 6.

3 and

We

discrimination

for

similar stimulus

two

To' by 4, and

Ic

by

6.

296

READINGS

"discrimination

MATHEMATICAL

IN

operator,"denoted

by

PSYCHOLOGY

S which

are

not

in S'

D, which operates on the similarity Tc divided


index r; each time the environmental
The
second

by

the

following the

event

from

e.g., from

reward

changes
another,

response

of event

type

one

to

non-reward.

to

thus alternate
associated

events

with

the operators
of the

Q'. So if tj, is the ratio

and

of I

measure

that

to

of S

of le divided
S and

Our

Dr^i.

task is to postulate the form

next

nor

pn' [""'
=

For

choose

choice of such

our

mathematical

wish

simplicitywe

have

to

operator

an

which

always decreases 77 ^or holds it


lead to
fixed),but which will never
values
of
negative
Therefore, we
tj.
postulate that
Dr]
where

^ is

in the range
then have

ki],

(25)

which
is
new
parameter
between
and 1. We
zero

"

value

D^riQ

of

v(S' to S)

m(S)

(29)

m(S')
We

shall

consider

now

k^no.

it

reasonable

seems

the

"operant"
the

in

to

m{Tc)

m{I)
m{TJ)
mjC)

(2""o(l D-rio)+ (Q'Qy^oD-m


[""" (a" ao)g"](l ^"t?o)

Hence, from

/)o

ao

ao'

/So.

is

variation

C^oc-(^oo-^o)/"""r7o.
(27)

response

our

final expression for the

of pn, the probabilityof the


in situation S, as a function

of the trial number


is composed

first term

n.

This

of two

major
corresponds to

of the

inspectionof equation (27)

Moreover,
shows
have

measures

'

m{S')

equations (17),(18),and

have

we

(30)

_
~

m(T')

+
This

assume

may

m(Ic)

Pn

same.

assumptions

m{T)

(19),

the

are

Moreover, in view of our


of equal proportions,we
that initially:
m(C)

that

assume

levels of performance

situations

two

(26)

Combining equations (20), (21),(22),


and
(26),we have

special

some

examples for which certain simplifying


assumptions can be made.
nation
(a) No conditioningbeforediscrimitraining. If no previous conditioning
took place in either S or S',

m(S)
=

ao')g'"](lk'^no)

[^ao-(^oo-)8o)/''"V,(28)

Vn

help

operator which represents


transformation
over,
Moreon
ij.

we

of S).

measure

a'

late.
postu-

an

linear

in

a' /(a'-f b'),and g'


a"'
1
is the initial
b',and where t/o'

where

v'

(aJ

experimental

intuition is of much

our

in guiding

by the

m(I)

of the operator D.
find that neither
We
data

(24)

elements
S' (the measure

in S':

"

Tji^i

the

of the symmetry
between
S', we may write for the probability

after the

ith presentation of S, the ratio after


the (i-|-l)th presentation is

S).

to

Because

of the

occurrences

of the

measure

of
of

measure

corresponds

term

the intersection of S and

In

the present problem, we


ing
considerare
alternate presentationsof 5 and S'
and

relative

(the measure

stimulus

equation
The

terms.

the

relative

elements

of

that, except
p^

ax,

and

when
in

like

1,

we

manner

equation (28) for k 9^ 1, we have


In Fig. 4 we
have plotted
aj.
pj
equations (27) and (28)with the above
0.12,
assumptions. The values a
b
W
0.03, /"o
0.05, 170
0.50,
from

0.95

chosen

were

As

can

for these calculations.

be seen,

the proba-

ROBERT

R.

AND

BUSH

These

is learned.

i.o

297

MOSTELLER

FREDERICK

describe the

curves

by
general sort
Woodbury for auditory discrimination
in dogs (4).
have
We
argued (1) that the mean
latent
time varies
inversely as the
in
Thus
Fig. 5 we have
probability.
result

of

obtained

p" and pn
exhibit
curves

of
plotted the reciprocals

Fig. 4. These
perimental
general property of the exof
time
on
running
curves
(5)
rats obtained
by Raben
(b) Complete conditioningin S before
discrimination training. Another
spein

given

the

same

ao

4.
Curves
of probability,p (in S),
trial number, n, for
p' (in S'), versus
discrimination trainingwithout previous conditioning.

Fig.

and

It

assumed

was

in S

rewarded

was

but

not

that the response


in S'.
rewarded

Equation (27),equation (28),and


p^'
0.05, a
0.12, o'
0.03, ijo
no' 0.50, and k

p^

the values

0,
0.95

6'

were

used.

bilityof the response in 5 is a monoerated


tonicallyincreasing,negatively accelfunction of the trial number,
creases
while the probability in S' first indue to generalization,
but then
decreases

to

zero

as

the discrimination

Curves
of probability,p, and its
FiG. 6.
trial number, n, for the case
reciprocalversus
criminatio
of complete conditioning in S before the distraining. Equation (27)with the
values /".o
0,90,
0.80, k
1, /3co 0, 7/0
0.50 were
used.
and /
=

cial
the

of interest is that

case

set

in which

5 is completely conditioned to
before the discrimination

the response

experiment is performed. In this case,


In Fig. 6 we
have
px/So ^0
ao
tions
plottedpn and \/pn with these condi=

the values

and
2.0

7?o

The
Fig.

5.

Reciprocals of probability,p, of
in S',
in S, and p',of the response

0.80,

curve

of

p"

0.90,

1/p

the

versus

1, i8""

and

/
n

0,

0.50.

is similar

experimental latency
Solomon
(6) from
by
for discrimination
n,
with
rats.
experiment
a jumping
training without previous conditioning. In
latent
the model described earlier (1), mean
cal.
(c) Limiting case of S and S' identitime is proportionalto the reciprocalof probability.
kind
of
the
Another
limitingcase
The
were
curves
plotted from the
of discrimination
experiment being
values of probabilityshown
in Fig. 4.

the response
trial number,
versus

in

shape

curve

to

obtained

298

READINGS

here obtains when

considered
the

two

into

one

where,

S'

type of partialreinforcement
is refor example, an animal
warded
on

S' is of

I oi S and

S'.

5 and

trial in

second

every

The

situation.

stimulus
both

S and

problem degenerates

The

identical.

make

we

situations

stimulus

MATHEMATICAL

IN

fixed

intersection

the

of

measure

of 5.

equal the measure


equation (5),we have
/ must

From

the

(31)

'

ni(S)

while according to

postulate about
equation (26), the
our

D,

operator

similarity index

from

varies

trial to

trial:

For

S and

S' identical,
the above

equations

are

take

I.

cues

depends
that k

way

available.

are

how

on

to

the problem.
The
usual
procedure is to select a block of {j -f-/)
trials during which 5 is presented j
times and S' presented / times.
The
is determined
actual sequence
ing
by draw"6" balls" at random

"5 balls" and


from

"6"

containing j "S

urn

an

and

balls."

I when

none

since / and

Moreover,

are

trials

describe

we

applying Q

its

to

index

ij

by the
the

number

For

mean

of times

number

mean

S'.

to

the

we

have

just argued that for 5


have

t}

I.

Thus

the

of shifts from
of j

of shifts is j.

f,

Since

(Q'Q)%D"'r,i. (36)

(34)

of the analysis exactly allels


parof
that given above for the case

rest

alternate

The
P

to

previously,we applied D tori for each


write for the (i4-l)th
we
pair of shifts,
block of {2j) trials

The
But

determined

specialcase

number

pected
ex-

operand j times, Q'

f times, and

its operand

or

of probability by

(33)

^V-

and S' identical,


we

is

sequence

effective

an

value

new

+
=

This

balls"

repeated throughout training.


In our
describe the
model, we
can
effects on the probabilityof a known
by an appropriate application
sequence
of our
and
D for
Q,
Q',
operators
of
S, presentations of
presentations
S', and shifts from one to the other,
less cumbersome
respectively. A
reasonable
method
provides a
mation
approxifor each
block of (j + j')
:

to

many

of T, the
identical,the measure
be zero.
complement of / in S, must
Since Tc is a sub-set of T, the measure
of Tc must
also be zero.
Therefore,
equations (17) and (19) give in place
of equation (20)
5

simple
analysis

will handle

we

available for discrimination

are

in such

two

incompatible, unless
Thus, we are forced

that k

assume

(32)

k^VO-

Vn

then

m(I)
"^

the basis of temporal order.


generalization of the above

identical to

course

Thus

PSYCHOLOGY

for the

S'.

presentations of 5 and

results will
value

be

of k

identical

except

involved

in

the

operator D.

Equation (22) gives us then

Summary

Pn

(Q'Qypo

P.-

(poo po)f".(35)
-

This equation agrees with our


result on partialreinforcement

(1).
(d) Irregularpresentationsof S and
S'.
In most
experiments, ^ and S'
not
are
presented alternately,but in
that the animal
an
irregularsequence
so
cannot

learn to

mathematical

model

previous generalization and

discriminate

on

described

in terms

for stimulus

discrimination

is

retic
of simple set-theo-

An index of similarity
concepts.
of the model but is
is defined in terms
in

related to measurements
The

tion
generaliza-

mathematical

experiments.
for acquisition and

operators

extinc-

ROBERT

in

discussed

tion,

earlier

an

from

derived

are

R.

the

BUSH

AND

(1),

paper

C.

HovLAND,

set-theoretic

here.

presented

model

The

The

I.

conditioned

applied

the

to

stimulus

on

of

analysis

1937,

17,

received

II.

periments
ex-

discrimination.

October

/.

C.

Psychol.,

B.

by

of

R.

R.,

"

Mosteller,

model
Psychol.
2.

ESTES,
of

W.

Rev.,
K.

learning.

94-107.

F.

for

1951,

Toward
Psychol.

58,

matical
mathe-

simple

rat's
in

Rev.,

1949,

Solomon,

313-323

statistical

theory
1950,

measured
/.

response.

learning.

57,

42,
R.

L.

56,

"

physiol.

of

Latency
learning

discrimination.

1943,

by

comp.

nation
discrimiof

intensity
running
Psychol.,

254-272.

of

measure

ulus
stimcamp.

29-40.

white

differences

illumination

of
J.

dogs.

35,
The

W.

279-291.

51,

learning

References

Bush,

eralization
gen-

responses:

1937,

The

1943,

M.

Raben,

1.

gen.

The

125-148;

Psychol.,

genet.

Woodbury,

1950]

13,

of

/.

conditioned

patterns

[MS.

I.

is
of

finally

generalization

responses:

Psychol.,
model

299

MOSTELLER

FREDERICK

Amer.
422-432.

in

response

'single
J.

as

door'

Psychol.,

TWO-CHOICE

BEHAVIOR

ROBERT

BUSH

R.

OF

AND

PARADISE

THURLOW

Harvard

FISH

WILSON

R.

*
University

Our

individua
S's. In general, most
Ss
problem stems
principallyfrom
served in a contingent experiment are
found
two
experiments. Brunswik
(1) obthe
bution
acquisition of a position to have an asymptotic choice distrifood was
discrimination
of
the
selection
of
by rats when
100%
in
box.
favorable
gent
Noncontinalternative.
placed more
frequently
one
Research
situations
by Humphreys
(9) was
give rise to other kinds
of choice distributions;in such expericomparable in that S had two choices
ments,
with
of
reinforcement
both.
the
partial
asymptotic proportion of
He required college students
choices
of the
favorable
alternative
to
guess
trial whether
match
the proon
not
to
or
a
light has been observed
every
portion
would
scheduled
of reinforcements
flash, and then in accordance
with
for the alternative.^
a
predetermined schedule, the
did
flash.
The
We
the nonlight did
not
or
attempted to obtain
Ss.
Humphreys
study exemplifies a noncontingent results with nonhuman
for
confronted
two-choice
Red
fish
contingent procedure
were
paradise
by
with partial
learning since the flash of the light a position discrimination
did not
the choice made
side was
in which
reinforcement
depend upon
one
random
and
the
other
Brunswik's
faced a conrats
correct
by S.
tingent
a
75%
situation

since
the
side correct
for the remaining 25%.
mental
environdiscrimination
change, presentation of food, The
was
a
apparatus
box with adjacent goal compartments.
was
contingent in part on S's response.
A contingent two-choice
For the experimental Ss, E placed the
research
on
humans
has been performed by Goodin
the
food
correct
compartment
Ss decided regardless of whether
entered
now
S had
(2, pp. 294-296). Her
trial
which
the
the
division
of
on
two
correct
goal box;
every
buttons
If the choice
between
the
to
two
was
was
goal boxes
press.
for
the
they earned
a
experimental
poker chip, transparent
correct,

otherwise

learning
has

Human

not.

with

been

two-choice

partial reinforcement

further

observed

Bush

and

Mosteller

two

associated

of

types

with

(2) suggest
procedures are

different

forms

food
when

under

"contingent procedure (3) and under


noncontingent procedure (3, 4, 5, 6,
7, 8, 10).
that these

the

see

that

so

group

these
in the

they

The

Ss

correct

had

control

able

were

to

ment
compart-

chosen

rectly.
incorwas

group

run

separatingthe
in order to produce
goal compartments
conditions
comparable to those used
with

by

an

opaque

divider

Brunswik.

of
Theory

asymptotic choice distribution


distribution
after learning)
^

This research
of

We
and

and
This

are

to

(choice
for

We

the

to
attempt
data within

describe

the

the framework

mental
experiof the

tory
supported by the Labora^
Social Relations,Harvard
Besides
University.
contingent and noncontingent procedure,
indebted to W.
other kinds of factors,such as a gambling
S. Verplanck for suggesting
that we
fish in learningexperiments
use
versus
a problem-solving
orientation,have been
F. Mosteller for numerous
suggestions related to asymptotic choice distribution (6)
was

criticisms.
article appeared

We
in

J. exp.

shall

Psychol.,1956, 51,
300

not

315-322.

deal with these other factors.

Reprinted with

permission.

ROBERT

On

Mosteller

(2).
1,2,.. .) there

that

each

trial and
of

of four

One

side.

the

that

to

favorable

more

events

leads

As

pn+i-

assume

choose

will

this

on

occurs

different value

in

similar

the

effect

analyses,

we

feeding

of

is

lyzing
previously used for anatwo-choice
experiments using the
from
contingent procedure is obtained
the above table by imposing the further

The

model

restriction

that

This

1.

ai

event

an

301

WILSON

R.

the

symmetrical for

two
goal boxes; we
assumption for nonthat a
feeding. In addition, we assume
side
of
on
one
feedings
long sequence
the probability of
make
would
tend
to
special assumptions
going there unity. These
the general model
reduce
to
about
the following statements
pn+i-

make

similar

information

This

schedules.

model

is

by Bush
equivalentto
and
Estes
Mosteller
and
(2)
by
(4) for
human
with
the
experiments
describing
the models

used

non-contingent procedure.
The
other specificmodel

tion
assump-

implies that nonfeeding is

THURLOW

AND

and
given by Bush
trial " (where k
0,
exists a probability /""

model

Stochastic

BUSH

R.

for

the

perimental
ex-

the
group,
bilities.
probais
obtained
It was
secondary reinforcementmodel,
expected that this model
X
1
from
the additional
restrictions,
describe learning by the control
would
that
This
model
and Q!2 " ai.
assumes
in the present
experiment. Given
group
S enters
when
one
goal box and sees food
be shown
that
this specificmodel, it can
in the other
goal box it is secondarily
the asymptotic p for each S will be either
the response
reinforced
for
just made.
1.0 or
0; for the 75:25 schedule it is

which

does

herein

alter the response

not

called

predicted that
will tend

towards

depends
We

the

exact

of

value

of

the

models

ai.

for the

present

p of 1
towards

of additional

that

than

of

the values
We

The

i's will

tend

precise

portion
pro-

1 depends

towards

and

ai

more

0.

about

the

dicts
pre-

asymptotic

an

on

ai-

with

chiefly concerned

are

tions
restric-

this model

S will have

0 and

or

that

shown

each

that tend

foregoing table by imposing


sets

been

that

periment.
ex-

obtained

are

It has

centage
per-

specificmodels

group

different

two

the

upon

These
from

The

1.0.

two

propose

experimental

of i's

high percentage

of the

forms

dictions
preptotic
asym-

of the
of choices
distributions
suggested by two
side.
favorable
These
predictionscould
different theories of learning. The first
be
tested
experimentally by running
formation
specific model, herein called the intaining
trials in the experiment and obmany
by taking
model, is obtained
for
of
choices
each
a
proportion
X
and
0.
As
a
result, the
ai
ai
the last 100 trials. The
S during, say,
in the forefirst and fourth listed events
going
thus
obtained
would
form
proportions
table have the
effect on
in

which

turn

are

p";

same

they correspond
the favorable
and

third

to

side.

listed

food

being placed on
Similarly,the second
have

events

effect; they correspond


placed on the unfavorable

to

the
food

side.

same

Karlin

These

each

trial may
providing information

be

about

described

(11)
of

models

could

be

is very

many

experiment

distributions

slow.

to

obtain

In view
we

are

that

suggests

the

trials would

as

the payoff

which

distribution

being

restrictions appear
arise most
to
readily
from a cognitivelearning point of view,

because

compared

with the predicted ones.


Unfortunately,
the mathematical
analysis presented by

forced

the
of

vergence
con-

these

Therefore, a great
required in the

be

the desired

bution.
distri-

of these considerations,
examine
the "nearto

302

predicts that

model

clustered

will be

distribution

mation
infor-

The

distributions.

asymptotic"

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

such

around

START

BOXES
STARTGATE

the
.75, whereas
model
predicts
that it will be U-shaped with a peak near
smaller peak near
0.
somewhat
1 and
a
The model for the control group
(a2
1)
also predictsa U-shaped near-asymptotic
0 should
distribution,but the peak near
be very
small compared to that for the
These
model.
secondary reinforcement
with
data
compared
predictions are
below

just

point

GOAL
DIVIDER

CHAMBER

"/"\z"

reinforcement

secondary

Fig.

Sketch of the discrimination

apparatus.

below.

Method
suction

by
Subjects.The Ss

49 red

27
paradisefish,
and 22 in the experimental
in the control group
The
red
paradise fish (Macropodus
group.
is
a
hardy tropicalfish about 2 in.
opercularis)
of its small demands
in length selected because
housed
The
Ss were
for care.
separatelyin
"

with

tanks
This

water

the

was

turned
each

day

fish has

diurnal

Apparatus.
box

of

in

J-in.

white

1"F.

by

our

matically
auto-

were

12-hr. period

activity cycle. (This


of

activity.)

apparatus

for parts of the


had

group

the

The

"

which

standard

rhythm

shown

as

constructed
except

for

on

control

to

indicated

"

appetite. Lighting

fixtures

fluorescent

by

of 80"

temperature

temperature
for maximum

feedingstudies
was

were

The

maze

white

opaque

nation
discrimi-

was

Fig. 1.

control

divider,whereas

opaque

side of the other


could

box

was

white

opaque.

These

interchanged. (Exploratory
studies indicated that a positiondiscrimination
with identical goal boxes is learned very slowly
by these fish.)
The apparatus
was
placed in a 10-gal.tank
shielded from
room
lights. Lighting came
largelyfrom a 75-w. spotlight2 ft. above the
sides

and

maze
was

taken

be

focused
to

that

experimentaltank
those

of the home

the

on

ensure

were

start
water
as

Care

of this

close

tanks of Ss.

as

possibleto

to

mm.

the

in

mm.

opening of

the

not

made

an

In the

eat

in

error

Procedure.

All

"

apparatus

Ss

received

75%

on

for

scheduled
trials.

On

were

that

than

two.

The

runs

selected

of Incorrect

All fish had

right,
yellowside

was

140

goal was
remaining
goal box was

the

one

restricted

by

The

could

the

of

the favorable side

of 20.

blocks

total

goal box (the

on

trials for which

within
was

for reinforcement

given trial only

incorrect

was

because

the other

reinforcement

The

correct.

trials while

One

20 trials a day or less.


trials,
scheduled
favorable side) was
of the

or

procedure.)

same

domization
ran-

restriction

not

be

longer

schedule.

favorable for about

white for one-fourth;


one-fourth of the Ss; right,
and left and white
leftand yellowfor one-fourth,
for one-fourth.

The
follows.

procedure for
The

fish

was

the

control

group

released from

was

the

as

start

down
to the goal boxes.
chamber, and it swam
experimental food was
pared
preinto the goal box which
If the fish poked Its nose
an
inexpensive(10 cents an
E lowered
for that trial,
correct
a medicine
("Lumpfish caviar" packed by
was
the
Hansen
with
fish
into
Caviar Co., New
a
York, N. Y.). These
dropper
compartment
egg
found
food of
secured to an arm) to allow
to be a highlypreferred
(the dropper was
eggs were
If the fish entered the Incorrect
the paradisefish and were
convenient
the fish to feed.
to obtain
In
and
The
store.
were
presented singly; goal box, no food was placedin the goal box.
eggs
fish
chased
he
held
the
into
the
the
end
of
medicine
back
either case,
t
was
on
a
dropper
egg was

Feeding.
"

The

chamber.
conditions

larger than

"

for

this divider was


parent.
transexperimentalgroup
For one
goal box the side oppositethe
formed
from a pieceof
to the box
entrance
was
lightyellow plastic;the corresponding
opaque
the

was

egg

the
the
fish was
dropper). To secure
egg,
obligedto pullit from the dropper. A fish was
all of its food by solvingthe
requiredto earn
discrimination problem.
Pretraining.The pretrainingtook two
or
the first day the fish was
three days. On
fed
(10 or 20) by eye dropper in its home tank.
eggs
For the next
two
one
or
days the fish underwent
forced trials (10 or 20) in the maze.
Half of the
the right-side
forced trials were
to
goal box.
one-third
of the fish were
About
rejectedfrom
the experiment at the end of pretraining
after
or
two
or
one
days of discrimination
training
rejectedbecause they
leaving49 Ss. (Fish were

Plexiglas, would
The

goal boxes.

was

(the
and

diameter

fish eggs
ounce) caviar

from

304

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

predict the
of

relative
of

groups

Ss

framework

of

the

learning are

determined

from

the

the

of

values

be

must

The

mated
esti-

models

other

considered

data

Within

models,
by the

data.

predict,however,

different

rates

which

parameters

learning

under

run

conditions.

experimental

of

of

rates

do

properties of
the following

in

sections.
The
The

near-asymptoticdistributions.
for the experimodels
specific
mental
"

two

group

and
model

140

120

100

80

60

about

the

of
Fig.

2.

parallelthe

is

for

22

the

of the

the
The

spread

of

152-160)

pp.

from

likelihood

maximum
used

was

the data, giving .7


The

during 10
The

distribution

trials

results

ratio

gives

test

considered

be

shown

are

of Table

column

of fish.

control
than
little

the

goal

box

slowed

Just

can

sight of
when
down

how

this

determined

analysisof
We
the

be

can

One

hasten
models

food

was

the

learning

only by
the
to

we

two

the

this

conjecture,of course,
food in the opposite
obtained

not

about

comes
a

more

process.
can

be

detailed

data.
note

described

at

model

this
above

point that
do

between

is

the

ber
num-

on

the

.22.

The

first

formation
in-

predicts a clustering

TABLE
Distribution

was

but

from

inferred

the

last

rapidly

group,

second

show

we

choices

trials

10

last

Trials
AND

Side)
for
for

not

Successes

of

Group

(Choices

During

Two

the
the

Parallel

that

more

experimental

more

figure.
that

and

in

of the

is clear

learned

group

the

It

successes

of 10 trials for each

groups

of favorable

Favorable

Fig.

"

coefficient

satisfactory.

In
Learning curves.
show
the proportion of
blocks

fit

the

In

Table

likelihood

This

.4.

of

learning is

frequenciesof successes
during the last
49
of successes
trials (the number
varies from 0 through 49). The
served
obtribution
U-shaped near-asymptotic disis not
determined
by initial
the
rank-order
relation
corpreferences alone;

mate.
esti-

the

butions
distri-

the

after

successes

computed.

the

tions
predic-

of

shape

successes

in

1 and
P

the

as

of

then

can

(12,

estimate

to

different

complete.

portion
pro-

side is plotted

block of 10 trials.

method

nearly
column

Mean

of the favorable

which
determines
parameter
the
distribution.
of

two

stat-fish which

experimental group.

of choices

for each

for each

Learning curve

of fish and

groups

very

model

reinforcement

secondary

make

"

TRIALS

the information

"

the

22
the
of

the

Groups

Stat-Fish
Experimental
Real

Fish

of

Last
of

Fish

Which

the

49

ROBERT

37

around

this

but

confirmed

clearly

not

data.

group

BUSH

R.

prediction is
mental
by the experiThe
secondary

the other
reinforcement
model, on
bution
hand, predictsa U-shaped distriwith greater density at the
high end than at the low end. This
On this basis
is confirmed.
prediction
choose the secondary
alone we
can
model

reinforcement

in favor

of the

Detailed

tions
ques-

model.

information

goodness of fit are


followingsections.

The

the control

assumption that
nonreward
has no effect (ao
1) and
it predictsthat the near-asymptotic
the

involves

group

the model

to the data in a detailed way.


that
the control group,
we
assume
the same
mate
model appliesand then esti-

For

both
whether
=

a2

of

U-shaped

but with

the

end.

at

low

small

very

with the data shown

density

indeed

This

agrees

of 27 fish stabilized

at

side

2 ; one
out
the unfavorable

during the

side 46 times
other

the

favorable

the

trials

a-z

yet
In

run.

the basic
in the

1 made

on

estimates.

three

of

secondary

we
experimental group,
the primary reward

estimate

for

model
need

parameter,

than

of

secondary reinforcement
(the smaller the

Estimates
FOR

Each

of

the
of

Two
the

is

nonreward

for the control


ai

as

control

fact that it is not

does

control
1.0

near

for the

model

reward
the

pectation
ex-

large

the

group,

in the

assumed

but

group,

the
that

quite 1.0 suggests

slightlyreinforcingeven
The

group.

result that

is less for the

experimentalgroup than
(primary reward
group
is
not
effective)
predictedby any

more

control

of the models.
The

effects of

relative

as

follows.

942

We

this

and
is

reward

secondary

primary

for each group

secondary reward

and
be

can

that

note

that

means

about

60%

as

for the
effective
as
primary reward
primary and
Similarly,
in this experiexperimental
(.956)-^^
group.
ment

value

of a,

the

.986,and

Parameters

Groups

Obtained
of

Fish

so

is about

secondary reward

effective for the

These

Two

is

a2

than

event

the

For

value.)
of

confirms

that primary reward


is more
(A small value of a impliesa

effective

more

30%
TABLE

This

ai.

effective.

(915).69

effects

of 10

in

used

are

correspondingprimary

the

parameter,

a^.

relative

butions
distri-

block

that

estimated

the

observed

be
shown in Table 3. It can
the secondary reward
rameter,
pais larger for both groups
a2,

are

noted

a\, and

two

scribed
de-

the first

uses

conjunctionwith
for moments
of the p-value
formulas
and
distributions derived
by Bush
however,
Mosteller (2, p. 98). The results,

to

the secondary reward parameter,


These
estimates are
required for
reasons:
{a) we wish to measure

It

in each

successes

be

cannot

of the

moments

for the

Having chosen

"

reinforcement

the

to

parameters
in detail here.

value

for the control group.

Parameter

the

not

consider

we

that

did

or

the

during

assumption
model
the

side

section

next

last 49 trials.

26 fish either stabilized

The

stabilize

it chose that

"

estimate

procedure used

in the last column

of Table

that

reward

will also be

successes

assumption

1 is tenable.

The
two

determine

and

parameters
the
not

or

distribution

and {b) the estimates


effect),
measuring goodness of fit ot

greater the
used in
are

these
trials;

proposedfor

model

305

WILSON

R.

considered

of
in the

THURLOW

AND

percentages
because

may

control
be

in

group.

error

of the sampling

in the parameter
estimates,but
indicate roughly the effects.

preciably
aperrors

they

do

of comparing
Stat-fish.A convenient
way
model
w
ith
is to
data
predictions
Carlo computations or "statMonte
"

run

fish"

as

described elsewhere

131, 251-252).
140 trialseach

One

were

(2, pp. 129-

hundred
carried out

of

runs
on

IBM

306

IN

READINGS

machines^

using the
Table

given in
sample of
the

22

initial

stratified

such

drawn

was

runs

TABLE
Comparison
Data

that

22

probabilities
symmetric beta

the

with

from

Experimental

the

Fish

Statistics Computed

of
for

of

distribution

approximate

would

experimental

these 100 runs,

From

group.

values

parameter

for the

PSYCHOLOGY

MATHEMATICAL

and

from

Obtained

from

the

Group

of

Sequences

the

22 Stat-Fish

the

the parameter
.7.
s
These 22 stat-fish can
then be compared
directlywith the 22 paradise fish in the
distribution

experimental group.
The
of the stat-fish
"learning curve"
in Fig. 2 along with those of the
is shown
fish.

real

fish

It

is

curve

the

be

can

seen

between
it is

model
the

the model

curve

parameters

for

discrepancy

of

Rather

how

well

estimated

were

stat-

should

the data.

and

indication

some

the

the

This

experimental group.
be
interpreted as

not

that

slightlyabove

the
from

mates
Loosely speaking, the estiobtained
by requiring that
the learning rates
of the model
lation
popuand
of the experimental sample
be equal. To
measure
goodness of fit
look at other properties of the
we
must
data.

were

these

All

chose

the

others

chose

four

The

of the 22 stat-fish

in the
The

distribution

near-asymptotic
same

manner

results

of Table

shown

are

for the real fish.

as

in the third column

the

that

for goodness

consider

we

of fit would

side

and

two

each.

once

initial

These

bilities
proba-

success

real fish is five.

.85, respectively.
of failures

This

suggests
have
been
would
agreement
if the initial distribution of bilities
probahad
had
less density in the
the

extremes;

formal

be

only

never

better

that

are

group
tests

by

found

close to the
sufficiently
corresponding frequenciesof the experimental
2 and

it

of .95, .95, .85, and


The smallest number

of

obtained

was

unfavorable

of

result

are

of the stat-fish

two

stat-fish had

data.

successes

discrepancies

the fact that

the

to

fluous.
super-

beta

symmetric
only as

used

was

distribution.

initial

true

bution
distri-

mation
approxi-

an

learning during the first


the distribution
10 trials tends
to spread out
of response
probabilitiesand so
the true initial distribution probably had
than
the symmetric beta
less variance
Furthermore,

Many

sequentialpropertiesof the data


the corresponding
to
in
properties of the stat-fish "data"
be

can

order

compared

obtain

to

further

goodness of fit.

Thus

the distribution

of

of

measures

have

we

tabulated

(of successes
for the experimental group
failures)
for the
the

stat-fish.
and

mean

In

and

show

we

of the total number

SD

of the number

runs,

Table

of

distribution

of

The

given

distributions

stat-fish
manner

lengths,as

well
S.

per

but

of

one

the

as

It

the

be

can

slightlysmaller for the


the stat-fish,
and
that
these
'

We

measures

are

number

tabulated

to

of

cesses
suc-

that

seen

be

can

real fish than

B. P. Cohen

and

Seymour for making these computations.

of iSs.

The

and

we

so

the

putations.
com-

compare

distributions

are

statistics listed
fish.

P. D.

greater
the

that

much
of the data.

the

same

two

groups

not

normal

the Mann-Whitney

used

(13). Comparison
values

fish and

in

compared
to

statistics

real

of each

of the

for

the variabilityof
real

used

as

of

for the

all
are

means

is less for the

indebted

Table

in

of various

runs

stat-fish

the

in

used

and

runs

in

than

Table
.3.

model

led

Thus,

we

test
seven

to

clude
con-

adequately

scribes
de-

of the fine-graincharacter

ROBERT

R.

BUSH

AND

THURLOW

Summary

307

WILSON

R.

multiple-choice

behavior.

/.

chol,
Psy-

exp.

1955, 49, 97-104.


A

two-choice
with

Ss

of

each

Ss

The
two

experiment

choice
49

were

conventional

the

given

were

the

Both

maze.

75%

side

Two

of

of

placed

the

on

that

predicts

the

side

of

The

data

the
reward
detailed
and

box

predicts

100%

the

choice

Hake,

which

9.

H.

W.,

of

binary

the

between

and

model

It

then

is

reinforcement

much

of

the

of

with

differing
/.

exp.

1-5.
R.

of

Perception

of

structure

L.

random

/.

G.

series

Psychol,

exp.

M.

E.

11.

fine-grain

S.

data.

in

alternative

Some

random

models

learning

learning

and

the

serial

symbols.

/.

1951, 41, 291-297.

Psychol,

Karlin,

/.

294-301.

effect

of

anticipation
exp.

tinction
ex-

in

conditioning.

to

Probability

recency

and

expectations

1939, 25,

Psychol,

Jarvik,

Acquisition

verbal

analogous

negative

cluded
con-

model

ation.
situ16-22.

Hornseth,

"

reinforcement.

symbols.

Humphreys,

exp.
10.

predictions

made.

are

secondary

describes

data

bility
Proba-

extinction

Hyman,

"

statistical

W.,

and
response

of

situation

ure
meas-

secondary

H.

1951, 42,

the

/.

1953, 45, 64-74.

secondary

and

primary

results

the

the

and

8.

sumes
as-

goal

opposite
made

approach

from

comparisons

of

the

hand,

Parameters
of

experimental

structure

"secondary

L.

1955, 49,

Psychol,

conditioned

Psychol,

in

theory.

problem-solving

Acquisition

model
proaches
ap-

sis
Analy-

situation

Postman,

"

A., Hake,

D.

P.

verbal

was

choices

other

support

estimated

that

J.

of

model.

adequately

Grant,

other.

effectiveness
are

This

The

the

in

obtained

reinforcement

food

of

just
will

7.

L.

New

H.

J.

learning

in

exp.

percentages

on

food

fish
the

or

trial.

an

choosing

which

on

all fish.

response

individual

one

for

iish

distribution

.75

sight

that

that

side

model,"

reinforces

assumed

of

preceding

about

of

the

the

reinforcement

model"

R.

"

processes.

1954, 47, 225-234.

learning
/.

in

theory.

Coombs,

conditioning

J. J.,

cussed.
dis-

are

certain
un-

1954.

Psychol,

exp.

H.

statistical

GooDNOw,

the

predicting
group

probability

trial

particular

for

of

C.

Straughan,

"

verbal

terms

the

K.,

in

association

Decision

Wiley,
W.

of

one

side

Thrall,
(Eds.),

Estes,

behavior

interpretation

an

statistical

M.

R.

of

on

Individual

of

York:

6.

experimental

the

sides

other

K.

Davis

5.

time.

"information

in

the

on

In

the

mental
experi-

both

on

W.

terms

observe

to

rewarded

models

the

The

increment

and

the

stochastic

behavior

on

time

25%

food

were

groups

the

of

remaining

the

Estes,

situations:

into

with

run

whereas

of

4.

comes
out-

divided

were

opportunity

an

absence

or

presence

the

is described.

fish

Ss

procedure

Ss

trial

paradise

control

provide

to

about

each

on

red

the

groups;

designed

information

complete

I.

walks

arising

Pacific

J.

in

Math.,

1953, 3, 725-756.
12.

References

Mood,

A.

M.

Introduction

the

to

York:

New

statistics.

theory of

McGraw-Hill,

1950.
1.

Brunswik,
of

rat

E.

Probability

behavior.

/.

exp.

as

determiner

Psychol.,

13.

1939,

25, 175-197.
2.

Bush,

R.

quantitative
(Ed.),

R.,

"

Mosteller,

Stochastic

F
.

models

New

for learning.

York

Wiley,

F.,

Mosteller,

"

Bush,

techniques.

Handbook

Cambridge,

of
Mass.

In
social

Detambel,

M.

H.

test

of

model

for

(Received

May

26,

G.

LIndzey

psychology.

Addison-Wesley,

1954.

1955.
3.

Selected

R.

R.

1955)

TOWARD

STATISTICAL
BY

THEORY

WILLIAM

K.

OF

LEARNING

ESTES

University

Indiana

Improved experimental techniques studies developinga statisticaltheoryof


and simple elementary learningprocesses.
From
study of conditioning
discrimination learningenable the presthe
definitions
and
which
assumptions
ent
for this kind of formulation,
day investigatorto obtain data
appear
necessary
which
shall attempt to derive
we
are
sufficiently
producible
orderlyand reto support exact
relations among
ures
commonly used meastive
quantitaof behavior and quantitative
of behavior. Analogy
expressions
predictions
with other sciences suggests that full
describing various simple
utilization of these techniquesin the
learningphenomena.
pend
analysisof learningprocesses will dePreliminary
Considerations
to some
extent upon
a comparable
Since propositions concerning psyrefinement of theoretical concepts and
chological
verifiable
events
are
methods.
only
The necessary interplaybetween
to the extent that they are reducible to
theory and experiment has been
of behavior under specified
predictions
hindered,however, by the fact that
environmental
conditions,it appears
of the many
theories of
none
current
that
and consistency
greatest economy
likely
learningcommands
general agreement
for the

among

researchers.

progress toward
reference will be
theories

It

frame

common

slow

so

built around

in

likelythat

seems

result from

of

theoretical
the statement

structure

will

mental
of all funda-

laws in the form

long as most
fined
verballyde-

R
f(S),
which are
constructs
hypothetical
R
where
and 5 represent behavioral
not susceptible
tion.
to unequivocalverificaenvironmental
variables respectively.
and
While awaitingresolution of the
Response-inferred
laws,as for example
peting
comapparent disparities
many
among
those of differential psychology,should
it may
be advantageous
theories,
of this
be derivable from relationships
to systematize well established empirical
The
form.
reasoningunderlyingthis
at a peripheral,
tical
statisrelationships
has been developedin a recent
position
level of analysis.The possibility
oped
paper by Spence (8). Although develof agreement on
theoretical framework,
a
within this generalframework,the
at
least in certain intensively
tent
expresent formulation departsto some
studied areas, may
be maximized
by
from traditional definitions of 5
definingconcepts in terms of experimentally and R variables.
manipulable variables,and
Many
apparent differences among
of assumpdevelopingthe consequences
tions
to
contemporary learningtheories seem
by strict mathematical reasoning. be due in part to an oversimplified
nition
defi-

This
*

For

are

essay will introduce


continual

reinforcement

series of

of his efforts

of stimulus
view

of stimulus

and
and

The

response.
response

as

mentary,
ele-

units has
reproducible

as
always
theory construction,as
many
the writer
specificcriticisms and suggestions,
had considerable appeal because of its
is indebted to his colleaguesat Indiana
versity,
UniThis simplicity
is deceptive,
simplicity.
especiallyCletus J. Burke, Douglas
entails
it
the
since
postulation
however,
G. Ellson,Norman
Guttman, and William S.
of
various
hypothetical
processes to acVerplanck.

well

at

This

for

article appeared in Psychol. Rev.,1950,57, 94-107.


308

Reprinted with permission.

W.

K.

309

ESTES

observed

in beThe pointof view to be developedhere


havior.
variability
In the present formulation,
will adopt as a standard
we
conceptual
shall follow the alternative approach of
model a closed system of behavioral and
variables. In any spein
environmental
includingthe notion of variability
cific
the definitions of stimulus and response,
the environmental
behavior-system,
and investigating
the theoretical conseinclude either the enquences
tire
component may
of these definitions.
stimuli
available
of
in
population
It will also be necessary to modify
the situation or some
specified
portion
the traditional practice
of that population. The
of statinglaws
behavioral
in terms of relations between
of learning
clusive
component will consist in mutually exfor

count

isolated

stimuli

at a

and

responses.

of
description
quantitative

and
learning

extinction of operant behavior


led the writer to believe that

have
a

self-consistenttheorybased

classical S-R

model

may

be

the

upon

if
difficult,

to extend over
impossible,
any very
wide range of learning
out
phenomena withthe continual addition of ad hoc
not

hypothesesto
A

described

handle

recurrent
as

follows.

every

new

ation.
situ-

difficulty
might be
In most

classes of responses, defined in


these classes
objective
criteria;

tempts
At-

tions
formula-

terms

of

will be exhaustive in the


will include
be

evoked

Given

sense

all behaviors

by

that

they

which

that stimulus

may
situation.

of the
the initial probabilities

various responses available to an organism


in a givensituation,
shall expect
we
the laws of the

theoryto enable predictions


of changes in those probabilities
function of changes in values of
as
a
independentvariables.

of

the organism
simplelearning,
Definitions
Assumptions
and
to "do nothing" in
originally
the presence
1. R-variables. It will be assumed
of some
stimulus;during
t
he
that
of
movement
make
or
to
learning,
organism comes
sequence
any
be analyzedout of an
movements
some
predesignated
response in the presmay
ence
of the stimulus;then during exand
tinction,organism'srepertory of behavior
the response
treated
various
as
a
"response,"
graduallygives
erties
propbe treated as dependent
of which can
to a state of "not
responding"
way
variables subject
to all the laws
again. But this type of formulation
does not define a closed or conservative
shall abof the theory. (Hereafter
breviate
we
with
in
the
word
In
order
derive
to
sense.
as
system
R,
response
any
and extinction
sary.)
necesappropriate subscriptswhere
propertiesof conditioning
from the same
In order to avoid a common
set of generallaws,
it is necessary to assignspecific
of confusion,
it will be necessary
source
ties
properclear
distinction
make
between the
to
of not
to the state
a
responding
is said

which

is the alternative to

occurrence

terms

of the

One solution
designatedresponse.
is to assignproperties
needed
as
by special
as has been done,
hypotheses,
for example,in the Pavlovian
tion
concepof inhibition.
of
simplicity
shall avoid
The

role of

the interest of

theoretical structure,we
procedure so far as

this

possible.

In

i?-class and

The
to

term

always refer
produce
effects within a specified

class of behaviors

environmental
range

/^-occurrence.

i?-class will

of values.

This

which

definition is not

without

objection(c/.4) but has the


the actual pracadvantageof following
tice
It may
be
of most experimenters.
possibleeventuallyto coordinate R-

classes defined in terms


mental
of environcompeting reactions has
effects
defined
with
i?-classes
in
b
ut
emphasizedby some
writers,
effector
activities.
usuallyneglectedin formal theorizing. terms of

been

310

READINGS

IN

shall mean
i?-occurrence
we
By
particular,unrepeatable behavioral
All

event.

defining

which

occurrences

of

criteria

meet

7?-class

an

PSYCHOLOGY

MATHEMATICAL

point of view,
predictionsof

present

cording
of

learning enable

the

changes

are

function

in

of

probabilityof response
as
a
time under
mental
environgiven

conditions.
class,and
such are experimentallyinterchangeA stimulus,or stimulatingsituation,
able.
as
In fact,various instances of an
will be regarded as a finite population
IJ-class are
ordinarilyindistinguishable of relativelysmall, independent, environmental
of an
in the record
which
of
experiment even
only
a
events,
with
effective
is
time.
though they may
actually vary
at any
sample
given
not
In the followingsections we
shall desigare
respect to properties which
nate
the total number
of elements
picked up by the recordingmechanism.
sociated
asIndices of tendency to respond, e.g.,
with a given source
of stimulation
probability as defined below, always
5 (with appropriatesubscripts
as
counted

as

instances

of that

the

to

laws

refer to i?-classes.
These

distinctions

where
may

be

by an illustration. In the Skinner-type


conditioningapparatus, bar-pressingis
usually treated as an i^-class. Any
of the organism which
movement
sults
rein sufficient depressionof the bar
the recording mechanism
is
to actuate

than

more

clarified

be

must

and

of stimulation

source

one

considered

the number

in

ment),
experi-

an

of elements

fective
ef-

given time as s. It is
tions
experimentalcondithe
stimulation
involve
repeated
of an organism by the "same
stimulus,"
ments
that is by successive
samples of elecounted
instance
of the class.
as
an
from an 5-population,each sample
-R-class may
The
be subdivided
into
be
treated
an
as
independent
may
finer classes by the same
kind of criteria.
random
pected
sample from S. It is to be expression
We
could, if desired,treat dethat sample size will fluctuate
of a bar by the rat's right
somewhat
from one moment
to the next,
and
forepaw
depression of the bar by
treated as the
in which
case
s will be
the left forepaw as
instances
of two
number
of elements per sample
average
different classes provided that we
have
over
a given period.
will be
which
a
recording mechanism
In applying the theory, any portion
affected differently
by the two kinds of
and

movements

different relations

mediate

to stimulus

at

of

the

environment
is

input (as for example

the presentation of discriminative

exposed

tion.

the

uniform

ganism
or-

tions
condi-

5-popula-

an

of different S's said

number

The

which

to

under

considered

be

may

uli
stim-

reinforcingstimuli). If probability
is increased
by reinforcement,
then reinforcement
of a right-forepawwill
the probaincrease
bar-depression
bility

any

that when

assumed

or

that
will

occur,

instances
will

and

of
also

that

subclass

increase

the

probabilitythat instances of the broader


will occur.
class,bar-pressing,
2. S-variables.
For
analytic purposes
it is assumed

is conditional

upon

It is not
responses

can

stimuli
eliciting

be

that

lation.
appropriate stimu-

can

be

identified.

Ac-

present in
the

behavior

constant

experiment,

and

of

If the experimenter

to hold

the

during
then

the

stimulating

the

course

entire

of

tion
situa-

single5. If
a
conditioningexperiment, a light
shock are to be independently manipulated
the CS and US, then each
as
will be

in

of

to be made.

attempts

an

number

are

situation

all behavior

implied,however, that
predicted only when

situation will

depend
perimental
exindependent
upon
a
nd
the
degree of
operations,
predictions of
specificitywith which
to be

these

treated

sources

as

of stimulation

will be

312

READINGS

Elicitation

Controlled

BY

tegrated to yield

ment
Reinforce-

Conditioning:

Simple

PSYCHOLOGY

MATHEMATICAL

IN

Let
of

consists

described

to be

The

tem
sys-

of

sub-

elements, So,
be
which
ently
independmanipulated
may
of the remainder of the situation,
S, and a class,R, of behaviors defined
properties. By
by certain measurable
controlled
of
originalstimulus,
a
means
which
has initially
that is,one
a high
probabilityof evoking R, it is ensured
stimulus

of

population

will

occur

contiguouslywith

the

of

instance

an

trial

every
of stimulus

which

elements

on

is present.
ditioning
con-

experiment, for example. So


lus
would represent the populationof stimufrom
the
sound
elements
emanating
R

of

limb

flexion response

5c which
time

as

be

may

conditioned

conditioned

on

If the remainder

as

function

(1

po)e-qt

than

tively
change in x per trial is relasmall,and the process is assumed
continuous, the right hand portion of
rate
be taken as the average
(1) may
of change of x with respect to number
of trials,
T, at any moment, giving

cial
spe-

any

original(or
other

than

conditioned)
un-

that

the

accumulation
other

in

of

ditional
con-

situations

classical conditioning,provided
tion
experimentaloperationsfunc-

ensure

learned
every
tion.

will

sample

that the response


the

in

occur

drawn

from

to

be

of
presence
the 5-popula-

If the

(Sc

dx_
_
~

dT

''

Operant

x)
'

Sc

In

differential

equation

may

be

the

more

common

by

different
to

be

controlled

some

ment
Reinforce-

Stimulation

mental
type of experivarious

instrumental,trial and

operant,

in-

Contingent

arrangement,

This

Conditioning:
BY

etc.
'

(30

assumed

not

relations

(1)

'

'

regularlyevoking the response to be


conditioned,it is to be expected that
the equations developed in this section
will describe

any

x)

Sc

of rein-

of

to
=

of the number

have

we

that other

Ax

has

trials.

new

trial will

any

"

of the situation

propertiesfor the

trial

at

ap-

experimentally neutralized,the
probabilityof R in the presence of a
sample from 5c will be given by the
ratio x/Sc. Representing this ratio by
the singleletter p, and making appropriate
substitutions in (3), we have the
of
followingexpressionfor probability

Since

from"

of

to

limitingvalue,Sc,in

later section.

of elements

to

will

be

(Sc

its initial value

stimulus

expected number

x, the

of x, and

proach
a negaA method
of
tivelyaccelerated curve.
evaluating x in these equations from
of response latency,
empiricalmeasures
reaction time, will be developed in a
or

tion
administra-

on
any one
of elements

the number
are

elements

is

from
the

evoked

number

mean

So effective

from
So, and

increase

electric shock.

an

Designating the
as

and

which

jCq is the initial value

speci- forced
fications
amplitude;

training trial by

each
of

(3)

represents the ratio Sc/So. Thus

i?

ments
move-

conditioned

be

to

where

certain

meeting

direction

of

all

include

would

and

the
t5^ically,

on

-qt
Xo)e

"

been

ple
sam-

buzz-shock

familiar

the

In

source

(Sc

"

consider first tliesimplesttjrpe

us

conditioningexperiment.

that

Sc

termed
error,

the
investigators,

learned

is not

elicited

originalstimulus,but

initial strengthin the

sponse
re-

by
has

experimental

W.

situation and

K.

as
originally
part
activity."Here

occurs

of so-called "random

be evoked concurthe response


cannot
rently
of each new
with the presentation
stimulus
effects

sample,but
in

Let

of the

occurrences

consider

us

same

by making
stimulatingsituation

the

changes
contingentupon

some

secured

be

can

of the

sponse.
re-

situation of

this sort,assuming that the activities of


the organismhave been cataloguedand
all
classifiedinto two categories,
sequences characterized by

313

ESTES

trial lasts until the

R,

discrimination

common

animal
the

sponse,
repre-designated
example,in a

For

occurs.

is placedon

beginningof

stand;

trial in

jumping stand

each trial and

until the

continues

animal

so

on.

at

the trial

leaves

the

experiment

runway

reaches

the end

Typicallythe

ing
stimulat-

lasts until the animal

box,and

apparatus the

situation present at the beginning


of a trial is radically
changed, if not

rence
completelyterminated,by the occurof the response in question
; and a
trial beginsunder the same
tions,
condiset of properties
new
being assignedto
for
class R and all others to the class Re,
except
sampling variations,
interval. The
after some
of class R are to be
and that members
pre-designated
learned.
pattern of movement-produced stimuli
be changed
If changesin the stimulus sample are
present during a trial may
i?
the
after
of
evocation
occurrences
behavior,
by
independentof the organism's
we

should

tain
cer-

expect instances of the two

classes to occur,

response

ment
move-

on

the

age,
aver-

to their initial
proportional
For if x elements from
probabilities.
conditioned
the 5-populationare
originally
of R
to R, then the probability
of
ments
elewill be x/S; the number
new
at rates

of

bit of behavior

uniform

some

eatingor

drinking;in

some

such

cases

as

the

utilized for this purpose must


be established by special
training
prior

behavior

learning experiment. In the


box, for example,the animal is
trained to respond to the sound of the
conditioned to R if an instance
magazine by approachingit and eating
will be ^[(S
or
s again
a!;)/5],
drinking. Then when operationof
occurs
of
stimulus
number
of
the
the
magazine follows the occurrence
representing
tioning
condimatically a bar-pressing
elements in a sample; and the matheduring
response
the animal's reof the latter,
sponse
expected increase in x will
it
to the magazine will remove
be the product of these quantities,
stimuli
the
o
f
the
from
the
in
the
same
time,
vicinity
sx[(S x)/S^]. At
that for an interval of
of Rg will be (5
the probability
x)/S, bar and ensure
to

Skinner

"

"

"

of

the number

and

to Re if

be

an

new

elements

instance

occurs

tioned
condiwill

time

thereafter the animal

exposed

sx/S; multiplyingthese quantities,therefore

to

most

the

of the

sample

of

will not

be

5-population;
elements

to

matically which the animal will next respondmay


sx[{S x)/5^] as the matherandom
be considered very nearlya new
Thus
decrease
in
x.
expected
S.
from
should predictno
we
change sample
average
In the simplestoperant conditioning
in X under these conditions.
be
ing
In the acquisition
possible to
experiments it may
phase of a learnthe
almost
entire
stimulus
tions
restrictwo
ple
samchange
important
experiment
each
jR
tend
after
of
occurrence
plete
(comimposed by the experimenter
while in other
to force a correlation between
reinforcement),
changes
the sampling of only some
in the stimulus sample and occurrences
stricted
recases
of
the
duced
of R.
The organism is usuallyintroportion
5-populationis
ment).
correlated with R
into the experimentalsituation
(partialreinforceshall consider the former
We
and the
at the beginningof a trial,

we

have

"

314

in

case

of

detail in the remainder

some

shall

which

any

trial to

actuallyoccur

instances

conditioned

be

elements

present

first movement

conditioned

the

the

to

until

be

be

so

x/S; if this value


we
can
trial,
readily
of reprobable number
sponses

the

the

trial is

will

occur

terminated.

next

the

of

recurrence

In practice,

sample of effective
will change somewhat

the

and

some
responses
fail to octrial may
cur
The
only response

one

on

the

the

on

of movements.

composition,
occur

stimulating
maintained, the most

next.

is R, since
be omitted
never
may
This
the trial continues until R occurs.

instance

an

the first response

on,

the

of

the

elements

on

and

response,
trial.
the

of events

course

however,
in

value

(of all classes)that

be

If

sequence

which

the

compute

before

R,

trial would

stimulus

will have

if the

terminates

could

probable
same

experimental
ple
representedby 5, the sameffective on
trial
one
by s, and
any
the ratio s/S by q. The probability,
p,
of class R at the beginning of any trial

tioned
condi-

predesignated

complete constancy
situation

ments
ele-

be

situation

varies littlewithin

trial;

stimulus

of

The

not

and

occurs

the

cues,

the first movement,

the

of class Re.

the

in

completely constant
probabilitythat
trial,and to proprioceptive

is

from

cues

cues

be

will

external

some

situation

during

of

be

will

occur

beginning

movement

next

trial.

that

on

to

on

ulus
stim-

to

environmental

the

to

at

present

the total number


available

all i?-classes

expect

that

to

proxima

conditional

definition of the

from

The

equal

Let

this section.

By our
relation,we

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

trial in

to

question is p;

that it will be

will

of R

occur

the

on

the

probability
is p (l"p);

the second

the

probabilitythat it will be the third


p)^; etc. If we imagine an inp{l
of trials run undefinitely
largenumber
der identical conditions,and
represent
is

"

the

number

any

trial

of response

by

value

we

n,

of

occurrences

"

on

weight each possible


probability(i.e.,

may

by its
expected relative frequency) and obtain
In S3mia
mean
expected value of n.
n

bolic notation

have

we

which

argument
detail

by

verify

the

need

we

been

has

Guthrie
line

of

developed in greater
(4). In order to
reasoning involved,
ideas down

to set these

now

investigatethe

and

mathematical

form

of
possibility

deriving functions

will describe

in

which

Znpil

Since
we

need

expression inside the


sign will be recognized as
of

term

the

duration

trial lasts until R

of

trial in

categorized
which
in

all

to be

are

that

movement

counted

convenience
we

duration

shall

in the
assume

of instances

general

the

infinite series with

well-known

Then

we

have, by substitution,

p/{i

(1

p)y

1/p.

occurs,

probable
of
we

the
have

sequences
as

summation

l/{l-{l-p)y.

sum

L, the

Then
be

the

of

responses

of the

and

will
trial,
number
expected

time

average

product

the

per

time

mean

per

response.

"responses"
L

mum
and that the minigiven situation,
for completion of a
time needed

response-occurrence is,on
For

the

terms

Suppose

strength of R.

py-K

ing.
of learn-

empirical curves

expressionfor

an

pZnil

The

each

p)"-i

the average,

h.

opment,
followingdevelthat the

of class R

mean

is ap-

Since

nh

will be

h/p

from

may

Sh/x.

conditioned

stimulus elements
we

substitute

present
for

on

its

to

all

new

each trial,
equivalent

equation (3), dropping

the

sub-

W.

So and

scripts from

K.

obtaining

So, and

315

ESTES

in the

figurerepresents the equation


2.5

Sh
L

(S

xo)e-^''

where

(Lo
1

h^e-^"^

(4)

values

estimated

.9648e-^2^'

of

from

the

give

to

appears

Thus,

will

value of Lo
the
h

over

from

decline

initial

an

(equal to Sh/x^)
asymptotic minimum

and

value

preliminary test of the validityof


this development may
be obtained
by
data
to
learning
applying equation (4)
from a runway
experiment in which the
A

in the derivation

assumed

conditions

are

tion.
degree of approximaIn Fig. 1 we
have plottedacquisition
Graham
and
data
reported by
Gagne (3). Each empiricalpoint represents
the geometric mean
latency for

realized to

of

group

with
elevated

fair

21

rats

food

for

which

traversinga simple
theoretical

The

runway.

noted, is very similar in form to the


developed
acquisitioncurve
Graham
and
The
by
Gagne.
present
differs from
formulation
theirs chiefly
in including the time of the first response
of
the
ing
learnas
an
integralpart
The
tion
quantitativedescripprocess.
of extinction
be

presented in

in this situation

to

shall have

we

to

between
situations

be

may

treated

defined

Then

if

from

the
to

the

two

any

beginning of
given occurrence

of

(T+l)

may

of

be

write

the

data

by

of
a

Graham

theoretical

and
curve

from

response

At

fore
there-

(and

considered

during
identity
Ax

runway

of R, and

have

from

Sh/x.

Ax
^

Latencies

elapsed
learningperiod

the

occurrences

in time
15

the

of R.

occurrences

reinforcements)of R, we
preceding development

Since

1.

''trials"

above, will represent

Fig.

as

let t represent time

we

the number

during conditioning, obtained

tervals
in-

reinforced.

40

TRIAL

the

in those

analyticalpurposes.
Making this
derive
we
an
assumption,
may
sion
expresfor rate of change of conditioned
strengthas a function of time
response
in the experimentalsituation,
during a
in
which
all
class R
of
period
responses

time between

10

that

assume

reinforcements

for

curve

L, as

will

forthcoming paper.
apply the present theory
the
to experimental situations such
as
Skinner
in
which
the
box,
learningperiod
is not divided into discrete trials,
order

In

are

curve

be

forced
rein-

were

This

proach theoretical
ap-

series of trials.

data.

satisfactorygraduation
it
pointsand, might

of the obtained
L

been

q have

and

Lq, h,

AT

as

the

we
trial,

crement
incan

AT

'~At'

published

Gagne

(3),

derived

in the

are

fitted
text.

its equivalent
Substitutingfor Ax/ AT
from
(1), without subscripts,and for

316

IN

MATHEMATICAL

the

ing
preced-

READINGS

AT/At its equivalentfrom


equation, we have

PSYCHOLOGY

Since

(or
s(S

Ax

At^

x)

"

s(S

"

x)x
(5)

occurrence

hS

new

to

small

in

be

may

respect

tinuous,
con-

portion of

taken

time

is

begun)

of R,

are

we

the

of i?

tion
equathe value of

as

and

after each
in

now

expected
a

function

of

occurrence

as

Representing rate
dR/dt, and
by y
=

dx/dt

to

reinforcement
is assumed

the process
the righthand

derivative

the

per

and

(5)
with

"trial" is

express

occurrence

change

If the

is administered

reinforcement

situation

consideringa

are

we

in which

the ratio

tion
posiof

rate

of time.
of

l/h by

w,

have

we

integrated

dT

dR

wx

{S

dt

dt

"

Xo)

"

1 +
Xo

5
X

(6)

(S
1

Xo)

"

and

if

take

we

of

Xo

Tq

wXq/S

the rate

the

upon

ginning
at the be-

experimental period

as

this relation becomes

s/Sh. In general,this equation


with
the
defines a logisticcurve
of initial acceleration dependamount
ing

where

of R

the

value

of Xq.

Curves

(W

"

(7)

To)

1 +

of

ro

probability(x/S) vs. time for S


100,
B
0.25, and several different values
of Xq are illustrated in Fig. 2.
=

To

illustrate

plotted in

this

Fig.

have

function, we
measures

of

rate

1.00 r

Fm.

2.

Illustrative curves

of
are

probabilityvs.
the

same

time

during conditioning;
parameters

except for the initial a;-values.

of the

curves

of

K.

W.

317

ESTES

10

"

10
IN

TIME
Fig.

Number

3.

15

20

MINUTES

minute
during conditioningof a bar-pressinghabit
per
is derived in the text.
singlerat; the theoretical curve

in

of responses

tive
and fittingthe cumulamost
respondingduring conditioningof a barpurposes,
of
records
with
The
the
rat.
integral
a
single
pressingresponse by
equaSkinner
tion
tion(7):
a
box; motivaapparatus was
the animal had
24 hours thirst;
was
w
been
trained
drink out of
to
previously
i?
w^+-^log
w
\w
/
lustrated
the magazine, and during the period il.

reinforced

was

all

with

at

various

times

for

were

an

estimate

of the

obtained

responses
per minute
The
theoretical curve
the

in terms

rate
at

the

in the

represents the number

made

of responses
by counting the number
half-minute
made
the
before
during
and
the half-minute
after the point
being considered,and taking that value
as

where

Measures

bar-pressing responses.

of rate

water

of

midpoint.
figurerepresents

after any

of

sponses
re-

interval of

time,
learning

beginning of the
sponses
original record of reperiod. The
vs.
time, from which the data
of Fig. 3 were
obtained,is reproduced
in Fig. 4. Integrationof the rate equation
for this animal
yields
t, from

i?

the

13/ +

125

logio(.038 + .962e--240.

equation

Magnitudes of R computed from this


equation for several values of t have
13
been plotted in Fig, 4 to indicate the
1 + 25e--24'
has
goodness of fit;the theoretical curve
A considerable
the
drawn
it
of
been
in
the
since
variability not
part
figure
of the empiricalpoints in the figureis
would
completely obscure most of the
due to the inaccuracy of the method
In an
experimental
empirical record.
of estimatingrates.
in press
In order to avoid
(2), equation (8)
report now
is fitted to several mean
this loss of precision,the writer has
conditioning
four
tive
cumulafor
of
of
the
curves
adopted
rats; in all
using
practice
groups
curves

of

responses

vs.

time

for

cases,

the theoretical

curve

accounts

for

318

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

elements

of stimulus
-100

is

trolled
con-

so

lation

by experimentalconditions that
from
drawn
sample of elements
of
contiguous with an occurrence

each

it is

The
sort of derivation
the response.
handle
this
kind of partial
needed to

in this section.

MINUTES

IN

TIME

sketched

will be

reinforcement

briefly

detailed treatment

more

vant
given,togetherwith releexperimentalevidence,in a paper
in preparation. It should be emphasized
now
that we
are
using the term
"partial"to refer to incomplete change
of the stimulus sample on each occurof a given response, and not to
rence
periodic,or intermittent reinforcement.
will be

Fig.

record
3

equation given in the text.

an

than

more

99

comparing

the

present formulation

Consider

study by

that

from

derived

Further

values.

verification of the
been

of the variance

cent

per

observed

the

has

the

which

from

obtained.

were

from

of

mulative
originalcuof
Fig.
points
Solid circles are
computed

the

Reproduction of

4.

of successively
curves
acquisition
habits,
bar-pressing
ing
Skinner-type condition-

behavior

ing
involv-

system

competing behaviors,

classes of

two

tion,
in a situaRg, which may occur
5, composed of two independently
obtained in a
Sr and Sg.
manipulable sub-populations,
apparatus which included two bars
Experimental conditions are to ensure
differingonly in position. It has been
lating
that of the sample, s, of elements stimuand
found
that the parameters w
s/S
ments
the
at
time, elebe evaluated from the conditioning
organism
any
can

learned

of

curve

used

one

bar

to

conditioning of

and

response,

predict the
a

detailed

then
of

course

learned

second

and

sponse.
re-

while

The

overall accuracy
in

describingthe
of bar-pressingand

should
that

fact

the

that

runway

be allowed to obscure

not

due

are
disparities

to the

experimentalconditions

do

fact
not

usuallyfullyrealize the assumption that


ment
only one i?-class receives any reinforceA
the
period.
learning
during
the
of
theory,
generalformulation
more
which
does not requirethis assumption,
in the next

will be discussed

Rg.

by

section.

the

entire stimulus

sample

by

of

occurrence

for

box

might

the

terminated

sponse;
bar-pressingre-

luminate
is il-

box

if the

visual stimulation

the

lustrate
il-

be

in which

is not

the

example,

of

occurrence

of system
Skinner

R,

effective

Se remain

by

kind

of

occurrence

from

terminated

systematic
initial portion
It is believed

by

This

but

small

is present in the
error
of the curves.
of most
that these

sponses
re-

the

elements

until
of these tions
equationing
of condirate

effective until

Sr remain

from
terminated

will

unaffected by bar-pressing
relatively

be
but

its

closes

Let

elements

given time,
from
the

Xr

the

of

behavior

Rg).
of

total number

S conditioned
number

Sr conditioned
numbers

of

instance

represent the
from

latter

(the

eyes

being,then, an

animal

if the

terminated

will be

to

occurrences

to

i? at

of elements

R, Tr and
of R

Tg
and

Rg prior to the time in question,and q


Partial
It
may

can

be

situation

Reinforcement

be shown

"learned"

that
in

a
a

providedthat

given response
trial and

some

s/S. By reasoning similar to


utilized in derivingequations (2)
age
(5), we may obtain for the averrate of change of Xr with respect to

the ratio

error

sub-popu-

that
and

320

READINGS

crease

IN

decrease in probabilityof

or

PSYCHOLOGY

MATHEMATICAL

currence
oc-

theory

be

to

seems

series of trials

an

inevitable developmen

the

at
depending
present stage of the
ability science of behavior; agreement on this
probis less than or greater than the
writers of
point may be found among
value
for
those
otherwise
conditions.
diverse
equilibrium
widely
viewpoints,
(1), Hoagland
(5),
e.g., Brunswik
Discussion
Skinner (7),and Wiener
(9). It is to
The
be expected that with increasingrigor
foregoingsections will suffice to
illustrate the manner
in which problems
of definition and
continued
interplay
of learningmay
be handled
within the
between
theory and
experiment, the
framework
of a statisticaltheory. The
various
formulations
of learning will
over

extent

the momentary

whether

upon

which

to

the

here

formal

velopedtend
de-

system

to converge

upon

set of

common

fruitfully
applied
may
to
interpret experimental phenomena
can
only be answered by a considerable
of research.
A study of conprogram
curi-ent conditioningand
extinction of
simple skeletal responses which realizes
quite closelythe simplifiedconditions

concepts.

assumed

in press.
Other papers
in preparation will apply this tion
formula-

rigidityand oversimplification
traditional
of
stimulus-response
theory without abandoning its principal
tion
advantages. We have adopted a defini-

extinction,
spontaneous recovery,
and
discrimination, related phenomena.

Skinner's

(7) concept

response similar to
of genericclasses,

and

given it

statistical

be

in the derivations
has

paper

report is

of the ent
prescompleted, and a

been

now

be
It may
the point of

relation of this program


to contemporary
theories of learningrequires

little comment.
made

No

been

certain

on

the

implied by

versial
contro-

present

analysis.
Stimulus-response terminology.
has

attempt

been

made

to

An

overcome

of the

some

of stimulus

have

and

Laws
within

of

tation.
interprelearning developed

this framework
defined

refer to behavior
in the

introductory
theory. It
systems (as
section of this paper) rather than
to
investigationto
between
relations
isolated
stimulusof
ing
learnconceptions
correlations.
and
discrimination
portant response
by stating imThis
gation
investiThe
learning curve.
concepts in quantitativeform
is
another
their
intended
be
to
ships
interrelationnot
investigating
search
for
"the
learning function."
by mathematical
analysis. Many
to

present

"new"

is the

of
purpose
of the
clarifysome

and

attempt has

view

issues

to

The

helpful to outline briefly

our

similarities will be noted

developed here

between

tions
func-

and

"homologous"
lations
expressionsin the quantitativeformuof Graham
and Gagne (3) and
of Hull (6). A thorough study of those
theories

has

thinking
than

in

build

influenced

directlyon

I have
formulations,

explore
based

writer's

either

Rather
of

point

of

statistical definition
and

those

felt it desirable to

alternative

an

on

the

respects.

many

behavior

and

view
of

vironment
en-

doing

greater justiceto the theoretical views


of Skinner

and

Guthrie.

statistical

The

writer

simple

does

function

not

will

believe
be

found

that
to

any
count
ac-

learning independently of
On
particularexperimentalconditions.
the other hand, it does seem
quite possible
small set of
that from a relatively
definitions and assumptions we
be
may
able to derive
expressions describing
learning under various specificexperimental
for

arrangements.
Measures

of
has

behavior.
been

indicate

the

of

as
mary
priAnalyses presented
that simple rela-

responding
dependent variable.
above

Likelihood

taken

K.

ESTES

probability
experimentally
rate of responding

common

W.

tions

derived

be

can

and
obtained

such

common

measures

as

between

and

latency.
able
of contiguityand effect.Availexperimental evidence on simple
to the writer to
learninghas seemed
require the assumption that temporal
contiguityof stimuli and behavior is a
Laws

321

pendent
quantitativeproperties;indevariables

environmental

of
Laws

the

of

theory

relations between
behavioral

statistical distributions

are

and

state

momentary
environmental

events.

probability
changes in
variables.

point of view it has been


tween
derive
simple relations bepossibleto
eral
probabilityof response and seving,
learnof
used
condition
the
formation
measures
for
commonly
necessary
and
to develop mathematical
pressions
exrelations.
At the level
of conditional
that is of laws
of differential analysis,
describinglearning in both
classical conditioningand instrumental
ior
relatingmomentary changes in behavditions.
conto changes in independentvariables, learningsituations under simplified
other assumption has proved necesno
sary
From

No

gation.
present stage of the investi-

at the

In

order

accumulation

to

account

of conditional

for

the

relations in

the

this

made

been

effort has

to

defend

lation
assumptions underlyingthis formu-

by verbal
"really"happens

analyses
inside

of

what

the

organism
proposed
arguments.
pealed
apbe
evaluated solelyby
that the theory
of experimentalopto a group
erations
its fruitfulness in generatingquantitative
which
are
usually subsumed
various
functions
ena
phenomunder the term "reinforcement" in currelating
rent
and
discrimination.
of
Both
learning
experimental literature.
Guthrie's
(4) verbal analyses and the
REFERENCES
writer's mathematical
dicate
ininvestigations

favor

of

others

one

in any

i"!-classat the expense


have
we
situation,

of

or

It is

similar

E.
1. Brunswik,
Probability as a determiner
essential property of reinforcement
/. exp. Psychol.,1939,
of rat behavior.
that sucis that it ensures
cessive
25, 175-197.
of a given R will be
occurrences
K.
Effects of competing reactions
2. EsTES, W.
contiguouswith different samples from
for
the
conditioning curve
on
the available populationof stimuli. We
bar-pressing. /. exp. Psychol, (in

that

have

an

made

assumptions concerning
specialpropertiesof certain

press)

no

the role of

3.

The acquisiC. H., " Gagne, R.


tion,
covery
reextinction, and
spontaneous

Graham,

such as driveafter-effects of responses,


in
affective
reduction,
changes
tone, etc.
Thus

the
here

of

oped
quantitativerelations develmay

prove

useful to investigators

the

nature

learningphenomena regardless
beliefs as to the
investigators'
of underlying processes.

conditioned

exp.

operant

Psychol., 1940,

sponse.
re-

26,

251-280.

Psychological facts and


psychologicaltheory. Psychol Bull.,
1946, 43, 1-20.

Guthrie,

5.

Hoagland,

H.

the

and

R.

E.

4.

of
of

/.

The

law

Weber-Fechner

all-or-none

theory.

/.

gen.

Psychol, 1930, 3, 351-373.


Summary
An
some

attempt has been


issues in current

by givinga
the

6.

made

clarify
learningtheory
and

response

7.

Skinner, B. F.
New

8.

New

Principlesof behavior.
Appleton-Century, 1943.

C. L.

York:

to

statistical interpretationto

concepts of stimulus

Hull,

York:

The

behavior
of organisms.
Appleton-Century, 1938.
The

Spence, K. W.
in

nature

contemporary

of

theory

struction
con-

psychology.

Psychol.
Rev., 1944, 51, 47-68.
by derivingquantitativelaws that
N.
9.
Cybernetics. New
Wiener,
pendent
simple behavior systems. Degovern
1948.
WUey,
in this formulation,
variables,
classes of behavior
are
samples with
[MS. received July 15, 1949]

and

York:

THEORY

STATISTICAL

SPONTANEOUS

OF
AND

REGRESSION

W.

K.

ESTES

Indiana

the

From
in

viewpoint of

constructing

would

be

habits
any

if

convenient

of

responding

given

situation

situation.

of

that

In

unreasonable,

be

that

could

be

much

tendencies

do

forgetting
"

How

construct

law

measure

as

has

to

situation

very

or

the

candidate
has
or

usually

paper

tenure

Social

as

Science

article

the

neural
varies

The

appeared

in

clearly

sition
po-

or

purely

of

Rev.,1955,
322

pothetic
hy-

now

area

gests
sug-

have

this type

of

Until
ables
vari-

explanatory
to

or

define

which

problems

of

evaluate

to

that

quire
re-

the

have

proposed.
parsimonious"

"more

By

explanation,

intrinsic

to

situation

and

play

this

special hypotheses

thor's
au-

Council.

Psychol.

class

refer

ordinarilystimulus

spontane-

which

"

possible either

be

of

fully explored, it will

explanation

been

array

prematurely.

scene

been

the

vorite various
fa-

postulated

prepared during the


fellow
faculty research
Research

scarcely

ronment
envi-

the

postulates

have

some

in

parsimonious

an

intervening

been

which

was

in

organism.

either

process,

hypothetical,
This

for

of

trace

attention

the

more

temporal

events

that

entered

never
as

for

The

events.

e.g., set, reactive

"

memory

den
bur-

statistical

to

the

of

the

hypothesized

organism

constructs

time,

interval

inferred,

or

in

of

the

extensiveness

havioral
be-

some

The

be filled with

sort, observed

This

properties

satisfying

variable.

explanatory

from

of environmental

compete

with

illustrated

shift

to

attempts

in

enough

cured
se-

out

events

will be

explanation

inhibition,

to

hoc assumptions.
learntime-dependent ing

to

processes

function

permanently

remains

the

expressing

temporal

unfilled

an

the

face

cannot

intervals

It is easy

for?

accounted

to

state

recovery,

than

having

once

which

paper

of

ways
it is al-

ad

few

approach

in this

with
difficulty

empirical

of

phenomena

separated.
these "spontaneous" changes

be

gap

and

they

the

hypothetical entities

foothold

of

for

postulate

to

Few

turn

whatever

is that

that

aid

response

in

The

ill-favored

new

The

are

to

but

the

ogy
psychol-

during

organism

the

well

are

each

tions
rela-

of

spontaneous

occur

so
a

certain, however,

more

e.g.,

"

are

mental
environ-

in

Nothing

unpostulate.

in

account

construct

easier

the

ing
learn-

of

and

much

hope

to

terms

orderly changes

that

when

in

laws

behavioral

variables.
is

facie,

prima

stated

between

than

it would

of

not

to

exposure

to

changes.

this type

modifiable

case,

empirical

the

all of

behavioral

to

respect

intervals

rest

is required

manner

organism's

an

with
were

only during periods

ously during

it

theory,

learning

University

interested

one

RECOVERY

role

62, 145-154.

any

variables,

variables, which

given type
thus

in

the

to

must

of

be

of

sources

are

behavioral

expected

to

interpretive schema.

Reprinted

with

permission.

W.

the

present instance

K.

ested
inter-

323

ESTES

carried

during a given period,some


stimulus
ments
elenewly conditioned
way
^
tendencies
will be replacedbefore the next
change during rest
response
have
which
intervals following experimental perinot
period by elements
ods.
ing.
conditionAnd
that there are
been
available
for
note
two
we
previously
lowing
ables
principalways in which stimulus variSimilarly,during the interval folcould lead to modification
in reextinction
an
sponse
period, random
tendencies
fluctuation will lead to the replacement
during rest intervals.
The first is the direct effect that changes
of some
lus
of the just extinguishedstimuin the stimulus
characteristics
of exelements by others which were
pled
samperimental
situations from trial to trial
during conditioningbut have not
In
been
available
or
sponse
period to period may have upon reduring extinction.
the
result
will
be
either case, the
a
probability. The second is
gressive
probetween
ods
perichange in response probability
learningthat may occur
if the stimulatingsituations obtaining
of the rest
function
of duration
as
a
have
within
and
interval.
between
periods
The
former catethese ideas testable,
In order to make
elements in common.
gory
state more
can
plicitly
again be subdivided according we must
formally and exenvironmental
the
variation
is
and
the
as
assumptions
tematic
sysconcepts
will
random.
involved.
Once
this is done, we
or
The
have in effect a fragmentary theory,or
random
lected
component has been sefirst subject of investigafor certain
account
as
our
tion
model, which may
for several reasons.
One
that
it
is
sponse
apparently spontaneous changes in rehas received little attention
heretofore
tendencies.
At a minimum, this
in learningtheory. Another
will enable
is that in
formal
model
to derive
us
other sciences apparently spontaneous
of the concept
the logicalconsequences
environmental fluctuation so
changes in observables have frequently of random
turned out to be attributable to random
be tested againstexperithat they may
mental
level.
the
data.
If
molecular
at
correspondence
a
more
processes
wish to
considerable
out to be good, we
turns
Perhaps not surprisingly,
may
in order
to
analysis has been needed
incorporatethis model into the conceptual
environmental
random
ascertain
how
of S-R
structure
learning theory,
intervals
of
limited
fluctuations
rest
counts
during
theory which acviewing it as a
pendent
for a specificclass of time-depected
followinglearningperiods would be exto influence response
ties.
probabiliphenomena.
of
It will require the remainder
shall reMost
of the assumptions we
quire
In

in
specifically

this paper
to summarize
and
results of this one
over-all

we

are

the

learned

the

of

methods
of

phase

the

investigation.

(8)
for

Theory

Stimulus

of

our

Even
can

prior to a
anticipatethat

for

fluctuation
of

response

analysis,we
mental
environ-

whenever
occurs,
at

the

the end

only

ability in
probof

For

present purposes.
environmental

experimental period will not be the


the probabilityat the beginsame
ning
as
If conditioning is
of the next.

terms

that

of

the
It
the

could

fluctuation

continua.

shall

concepts

within

will

be

Hullian
be

simplicity

mathematical

of

reasons

theory.
argument
the

convenience

one

briefly

situation,as
given time, determines
given organism a population of

and

restated

be

Any

constituted at

Fluctuation
detailed

need

elsewhere

discussed

been

have
and

a.

General

out

the

worked

of stimuli

develop
of

ideas

these

ing
statisticallearn-

however,

apparent,
system
out

in

similar
terms

of

along generahzation

324

stimulus

which

from

events

organism'sbehavior

affects the

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

mental

at

constant

any

instant; in statisticallearning theories


a
the populationis conceptualizedas

of which

all
situation,

sample

random
let

Now

in which

type of

the

consider

us

undergo

fluctuation.
periment
ex-

is

run
organism
period in the same
ior
random
sample is drawn on each trial. apparatus. In dealing with the behavthat
extinction
and
mental
experioccur
occurs
b. Conditioning
during any given
pled
period, the total population S*
only with respect to the elements samon

exhaustive

an

population

and

ment
ele-

is conditioned

at
can

any
be

the

available

and
are

derived by various

Under

not.

the

periment
ex-

two

elements

of

during

5' of

that

period
which

elements

ered
conditions consid-

probabilityof a
time
during the
given
any
ments
equal to the proportionof elethe

in this paper,
response
periodis

subset

subset

the

the

in

during the
partitionedinto

which

are

available

time

portions:

of these response classes.


exactly
of these assumptions,
basis
the
On
been

an

one

elements

stimulus

situation

one

functions have

than

more

be

stimulus

time, each

any
in the

ganism
or-

classes.

response

d. At

for

of

available to

behaviors

in a given situation may


categorizedinto mutually exclusive

to

trial.

The

c.

which

from

elements

of stimulus

set

at

scribe
in the available set S that are
investigators(2, 5, 8, 16, 21) to deof learning predicted conditioned to that response.
the course
Owing to
the
in
which
for an
idealized situation
there is some
environmental fluctuation,
is

physicalenvironment

idealized situations

trial. No

each

populationon

stimulus

stant
perfectlycon-

organism samples the

the

and

testingpurposes, but the theory seems


pirical
to
give good approximationsto emin
obtained
functions
learning

an

the

unavailable,

terval
into S', during any given inthat
an
;'
A^, and a probability

i.e.,go

in S' will enter

element
are

in

element

available set 5 will become

for

available

are

probability; that

5.

These

illustrated in Fig. 1 for

ideas

cal
hypotheti-

situation.

experimentalperiodsunder well-

short

controlled conditions.
In the present paper

behavioral

from

tention
at-

our

changes

that

experimentalperiods

within

occur

turn

we

the

changes that

the

intervals between

as

to

of

spondingly,
periods. Correreplace the simplifying

we

of

assumption

occur

function

perfectlyconstant

domly
assumption of a ransituation.^
cally,
Specififluctuating

the

with

situation

ability
that the avail-

it will be assumed

during a
given learning period depends upon a
of independently variable
large number
stimulus

of

components

or

elements

aspects of the environFig.

It is

possiblenow

to

functions

derived

the
for
be

this
able

paper.

random
to

go

go

back
earlier

variation, but
into

this

point

we

in

and

rect"
"cor-

to

allow

wUl

not

the present

1.

Fluctuations

in stimulus

sets

ing
dur-

regression (upper panel) and


from extinction (lower
recovery

spontaneous

spontaneous
panel). Circles represent elements connected
Values
of p represent probA.
to response
abilities
of response

in the

available

set S.

W.

The

relevance

of

the

learning phenomena
fact that

both

K.

scheme

for

from

the

arises

conditioned

and

ditioned
uncon-

will constantly be
elements
fluctuatingin and out of the available
set S.
During an experimental period
in which
curs,
conditioningor extinction octhe proportion of conditioned
ments
elein

will

increase

decrease

or

relative to
proportion in 5'. But
these
during a subsequent rest interval,
will
tend
toward
proportions
equalityas

325

ESTES

rium

in which

result of the fluctuation

Recovery

The

essentials of

the

at

any

period
in

curve

The

is

the

in
an

that will be

in

ing
condition-

given by the topmost


panel of Fig. 2.
upper

equation of

In

equal.

proportion

followingthe

time

are

tioned
condi-

spontaneous

the

will be

curve

rived
de-

later section.

analogous fashion

the essentials

of the spontaneous

recovery process
in the lower panel of

are

Fig.

We

tion
begin at the left with a situamaximal
following
conditioningso

that all elements


A.

of

treatment

our

of

of

Regression

and

of

elements

Spontaneous

of

S'

of conditioned

1.

Interpretation

and

predicted course
regressionin terms

schematized

process.

densities

in 5

The

the

the

elements

conditioned

are

During

to

singleperiod

sponse
re-

of

all elements in the available


and
spontaneous
regression extinction,
recovery
5
conditioned
set
to the class of
are
will be clear from
an
inspection of
competing responses A and the probability
Fig. 1. The upper panel illustrates a
A
of
to
zero.
temporarily
in which, starting from
goes
case
a
zero
Then
during a recovery interval,the
level,conditioningof a given response
random
interchangeof conditioned and
A is carried out during one
period until
unconditioned
elements
between
5 and
probabilityof A in the available
5'
results
in
increase
in
the
a
gradual
situation represented by the set S is
elements in
unity. At the end of the conditioning proportion of conditioned
5
until
the
final
is
state
equilibrium
will have, neglecting any
period we
reached.
The
of
predictedcourse
fluctuation
that
occurred
taneous
sponhave
may
the

during
in 5

the

period, all

conditioned

to

of

the

and

elements

all of

the

temporarily unavailable elements in S'


unconditioned.
val
During the first interof
the
the ensuing rest interval,
At
.6 of the conditioned
proportion }
elements will escape
from 5, being replaced
the
the
of
by
proportion;' .2
=

unconditioned

elements

from

5'.

ing
Dur-

as

recovery

function

of time

is

in the
given by the topmost curve
lower panel of Fig. 3.
According to this analysis,spontaneous
regressionand recovery are to be
regarded as two aspects of the same
process.
process

In
is

each

curve

the

case

given by
with

form

of the

celerated
negatively ac-

the relative rate of

acteristics
interchange change depending solelyupon the charof the physicalsituation embodied
will continue, at a
creasing
progressivelydein
the
and
Rates
;'.
parameters
;
rate, until the system arrives
and recovery should,then,
at the final state of statisticalequilib- of regression
the variability
together whenever
vary
* The
term
regression will be
spontaneous
of the stimulatingsituation is modified.
used here to refer to any
in response
decrease
It cannot
be assumed, however, that
is attributable
probability which
solely to
of regression and
amounts
recovery
stimulus fluctuation. It is assumed
that over
short time intervals,the empirical phenomeshould be equal and opposite in all exnon
periments.
of forgettingmay
be virtually identified
The
illustrative example of
with regression,but that over
longer intervals
Fig. 1 meets two specialconditions that
forgettingis influenced to an increasing extent
do
not always hold: (a) the conditionby effects of interpolatedlearning.
further

intervals the

326

IN

READINGS

MATHEMATICAL

PSYCHOLOGY

in the upper
given by the top curve
of
2.
If
in
the
tion,
situapanel
Fig.
same
conditioninghas been carried only
to a probabilitylevel of, say, .67,then
the

Pit)

total

number

will be

conditioned

of

smaller

and

the

ments
eleof

curve

regressionwill not only start at a lower


value, but will run to a lower asymptote,
and
so
on.
Similarly,if in the
situation representedby the lower panel
of Fig. 1, response
probability
goes to
the
extinction
zero
during
period, the
of spontaneous
covery
repredicted course
is given by the lowest curve
in
the upper
if
extinction
of
panel
Fig. 3;
terminates
at higher probabilitylevels,
the successivelyhigher reobtain
covery
we
the
in
shown
curves
figure.
Number
of preceding learning periods.
of preceding
Increasingthe number
acquisitionperiods would tend to
Fig.

the

In

upper
elements

conditioned
is
the

regression
panel the proportion of

of spontaneous

Families

2.

curves.

parameter.

in 5' at

In

the

of conditioned

ing

and

extinction

ditioning
con-

in

elements

series

the

at

the

tion
propor-

from

start

initial response

zero

and

probabilitiesof
unity, respectively;and (b)

ditioning
con-

and

condi-

proportion in 5 is
lower
panel the proportion

of conditioning is unity and


in S' is the parameter.

end

of

the end

of

total number

the

and

zero

the

increase

extinction

are

carried to

comparable criteria within the experimental


terval.
inthe
rest
period preceding

Predictions

Concerning

Experimental

OF

Terminal

level
If

fixed,the
attained

period
value

at

of conditioningor

other

conditions

tinction.
ex-

remain

level of response
probability
the end of a singlelearning

will determine

and

Effects
Variables

both

the

FiG.

initial

the

as3miptote

of

the

curve

the
For
regression or recovery.
the
situation
represented by
upper
ing
of conditionpanel of Fig. 1, the curve
the predicted
goes to unity, and
of
course
regression is
spontaneous

Families

3.

In

curves.

conditioned

of

the

of

elements

in 5' at

is unity and
elements

in

spontaneous

recovery

panel the proportion of

upper

the
5

at

the

end

proportion
the

end

of
of

tion
of extinc-

In the lower
parameter.
proportion of conditioned elements
is the

the
at

the

the

end

of extinction is the

proportion in 5 is

zero.

tinction
ex-

ditioned
con-

parameter

panel,
in 5'
and

328

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

elementary probabilitytheory we have


element is
for the probabilitythat an
the
val
of
in 5 at the end
(^ + l)st interA*
following an experimental period:

us
designate by p(t) and p'(t)
proportionsof conditioned elements,
and therefore the response probabilities,
lowing
in 5 and 5' respectively
at time t folan
experimentalperiod. The set

Let

the

of conditioned

1)

/(/ +
This

[!-/(/)]/ +/(0(1

difference

equation

-J).
solved

be

can

(2, 12) to yield


by standard methods
of t and
in
for
formula
terms
a
/(^
the

will

elements

in 5

the conditioned
in part from
in number, that were

come

elements,p{0)N
in S at the end
and

of the

experimentalperiod,

time

are

these

two

from

obtained

[1]

/-C/-/(0)]a'

/(/);
j'; and
'}"']')"

"

between

is bounded

is in S

will settle down

value /

to the constant

sufficiently
long interval

spectively.
re-

relations

at

of

p{t)=^\:piO){J-(J+

p'(0)/(l -a')N'2

the

/,

by the definition of ; and


probabilitythat any element"
after

and

hand
are
we
ready to write the general
expressionfor spontaneous recovery and
regression:

and

"

to

the initialvalue of

/ represents the fraction ;'/;+


a
represents the quantity (1

Since

at

Equation

these

With

in 5

sources

by setting/(O) equal

/(O) is

conditioned

the

in part from

from

where

elements,p'{0)N' in number, that were


of findingelein 5'. The
ments
probabilities

parameters:

time

at

p{0)lJ

{J

l)aq

-f-^'(0)(l-aO(l-/).

[3]

time,

having been
Equation 2.
values N
and S' will stabilize at mean
The functions illustrated by the curve
and N', respectively,
which satisfythe
families of Fig. 2 and 3 are all special
relation.
the upper
of Equation 3.
In
cases
N
[2] panel of Fig. 2, p'{0) has been set
J(N -f N').
equal to 0; in the lower panel, /"(0)
Spontaneous recovery and regression.
has been set equal to 1.
In the upper
and reCurves
of spontaneous
gression
recovery
has
been
of
set
panel
Fig. 3, P'{0)
be obtained by apcan
now
propriate
equal to 1; in the lower panel, p{0)
application of Equation 1,
has been set equal to 0.
and

of elements

the total numbers

in 5

the parameters
eliminated by

N'

and

means

of

simplicity,it has been

For

all of the

paper

that

same

values

some

situations

to

of

that

assume

associated

with

data
that
elements

it

;'.

and

be

might

in
In

different

reasonable

more

in the

Skinner

should

be
while

For

(11)

Homme

box

regarded
the

values

elements.

by

as

in this

5* have the
dealing with

different parameter

obtained

available

assumed

elements

portion of
fixed

remainder

and

gest
sug-

the
ways
al-

fluctuate.

from

those given in this paper.

Relevance

General

sections

and

considerations.

developments

are
ample,
ex-

described
Application of an analytic method
(8) shows that conclusions in the
will differ only quantitatively
general case
elsewhere

Empirical

present

two

of

Adequacy
The

the

aspects,

retical
theo-

preceding
one

eral
gen-

which
are
by no
specific,
with
the
regard
same
footing
means
on
to
It will be necessary
to testability.
discuss separately the general concept
of stimulus fluctuation and the specific
and

one

mathematical

model

utilized for

pur-

W.

of

deriving

its

testable

K.

329

ESTES

to contemporary
assumptions common
statistical learning theories
(2, 5, 8,
The
reason
matical
why the fluctuation concept
16), the result of the union is a mathehad to be incorporatedinto a formal
model which yieldsa largenumber
of predictionsconcerning changes
theory in order to be tested was,
of course, the difficulty
of direct observational
in response
tervals.
probabilityduring rest incheck.
Thus
for the presOnce
formulated, this model
ent
this concept must
be treated with
is readilysubject to experimental test.
the same
and
reserve
even
suspicion Its adequacy as a descriptivetheory of
which
as
appeals to
spontaneous recovery and regressioncan
any interpretation
unobservable
This
be evaluated quite independentlyof the
events.
remoteness
from
direct observation
merits
of the underlying idea of stimulus
however,
may,
fluctuation.
represent only a transitorystage in the
developmentof the theory. Relatively
Spontaneous recovery.
Space does
direct attacks upon
certain aspects of
the
detailed
not
discussion
of
permit
the stimulus element concept are
experimentalstudies,and we shall have
vided
proby recent experiments (1, 21) in
of
to limit ourselves to a brief summary
which the sampling of stimulus populafrom
derivable
tions
empirical relationships
has been modified
experimentally the theory, together with appropriate
and the outcome
retical references
ture.
compared with theoto the
experimental literaTo
the best of my
expectation. Further, it should
knowledge,
poses

be

noted

that

the

idea

of

quences.
conse-

stimulus

fluctuation is well grounded in physical


considerations.
Surelyno one would

the

references cited include

which

all studies

able
provide quantitativedata suitfor comparison with
predicted

functions.
deny that stimulus fluctuation must occur
continuously; the only question is
The
of recovery
is exponential in
a.
curve
whether
fluctuations are
large enough
form
(3, 9, 17) with the slope independent
under ordinary experimental conditions
of the initial value
(3).
effects upon
ior.
behavto yield detectable
b. The
of recovery
is inversely
asymptote
The
surmise
that they are
is not
related to the degree of extinction
(3, 11).
The
of
is
idea
of
the
c.
directly
asymptote
a new
vironmental
enrecovery
fluctuating
one;

components

has

been

used

related

to

the

number

given prior

in

to

of

conditioning periods
(11).

extinction

an
explanatory sense
by a number
d. The
of recovery
is directly
asymptote
with parin connection
investigators
ticular related to the spacing of preceding conditioning
problems: e.g., by Pavlov
(19)
periods (11).
and
Skinner
Amount
creases
of recovery
e.
(22) in accounting for
progressively detion
during a series of successive extincof conditioning
perturbationsin curves
periods (4; 13; 19, p. 61).
or
counting
extinction,by Guthrie
(10) in acfor the effects of repetition,
be noted that items c and d
It may
and recentlyby Saltz (20) in accounting
represent empiricalfindingsgrowing out
for disinhibition and
reminiscence.
of a study conducted
expresslyto test
Considered
in isolation,
the concept
the
certain aspects of
theory. Many
of stimulus fluctuation is not even
directly
inadditional
from
predictions derivable
be incorpotestable;it must
rated
the theory must
remain
unevaluated
into some
broader body of theory
until appropriateexperimental evidence
before empirical consequences
be
can
becomes
lation
available,e.g., the inverse re-

of

derived.
found
in

In the present paper

that when

this concept

conjunctionwith

other

we

have

is taken

concepts and

between
and

as3miptote

spacing of extinction
and

the

of

recovery

trials

or
ods,
peripredictionsconcerning"ex-

330

in

mentioned

zero"

tinction below

previous section.
Spontaneous regression. Predictions
tween
beconcerning functional relationships
spontaneous regressionand such
experimentalvariables as trial spacing
or
degree of learning parallel those
given
but

for

above

in the

of

in amount

Summary
In

are

of

purposes

predictedexponential

verification. The
decrease

for

available

progressivelyreduced if we
successful in bringingother relevant
are
cal
independent variables into the theoretiof the
fold by further applications
here.
analyticalmethod illustrated

regressionas

this paper

be

extinction,may

from

recovery

for in terms

ing
of preceding learnbeen
observed in several

investigated

have

we

that certain apparently


the possibility
behavioral
changes, e.g.,
spontaneous

in stimulus

function of number

periodshas

be

it may

recovery,

regressionthere

of

case

data

fewer

spontaneous

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

of random

conditions.

counted
ac-

tion
fluctuaTaken

the concept of random


isolation,

in

lus
stimu-

has proved untestable,


fluctuation
tions
(6, 11, 13, 14). Predicit
into a model
but when
incorporated
concerning regressionin relation
of
have
led
not
has
descriptions
of
to
quantitative
to spacing
learningperiods
pirical
tested in conditioningsituations, a variety of already established embeen
to be in agreement with
but they seem
relationshipsconcerning spontaneous
and
to
and
lationships
regression
rather widely established empirical rerecovery
of
and
tion
retenthe
determination
new
ones.
some
between
spacing
A forthcoming paper in which the same
in human
learning (15, pp. 156tribution
model is appliedto the problem of dis158; 18, p. 508).
ther
furwill
be
raised
of
provide
the
practice
question may
Finally,

studies

there

whether

that would

are

no

embarrass

ness
evaluation of its scope and usefulof
in the interpretation
learning

facts
experimental
the

ory.
present the-

comprehensiveness phenomena.
the theory, then
made
for
had
been
REFERENCES
negative instances would be abundantly
K., " Hexlyer,
1. BxjRKE, C. J., EsTES, W.
available. Under
conditions,for
some
lation
of verbal conditioning in reS. Rate
regressionfails to
or
example, recovery
to stimulus
variability. /. exp.
at all following extinction or
appear
Psychol, 1954, 48, 153-161.
F.
Stochastic
ever,
2. Bush, R. R., " Mosteller,
conditioning,respectively.Since, howmodels
for learning. New York: Wiley,
dealing with a theory that
are
we
in press.
pendent
is limited to effects of a single inde3. EixsoN, D. G.
Quantitativestudies of the
variable,stimulus fluctuation,
ery
interaction of simple habits: I. Recovfects
instances of that sort are of no special
from
specificand generalizedef/. exp. Psychol.,
of extinction.
limited
Like
theory,
significance.
any
1938, 23, 339-358.
tions
be tested only in situathis one
can
If

where

claim

of

suitable

measures

are

taken

4.

G.

D.

Ellson,

Successive

extinctions of

in rats.
bar-pressingresponse
Psychol.,1940, 23, 283-288.
a

and

where

the

represented
negligibleor

effects of variables not

in

And

the
else

model

are

5.

these

evidence

seems

cations,
qualifito

K.

Toward

statisticaltheory

learning. Psychol. Rev., 1950, 57,

94-107.
6.

be

danger
uniformly confirmatory. The
dence
of continually evading negative eviables
by ad hoc appeals to other varibe entirelyobviated, but
cannot

EsTES, W.
of

quantitatively
predictable.

subject to

available

either

/. gen.

EsTES, W.

K.

Effects

of competing

tions
reac-

for bar
conditioning curve
pressing. /. exp. Psychol., 1950, 40,
on

the

200-205.
7.

EsTES, W.

K.

Statistical theory of

phenomena
Psychol. Rev., in press.

in

tributiona
dis-

learning.

K.

W.

8.

K.,

W.

EsTES,
of

C.

Graham,

H.,

16.

The

A.,

and

extinction,
conditioned

operant

Psychol.,

exp.

17,

1952,

369-396.

sponse.
re-

17.
/.

tistical
sta-

learning.

spontaneous

Psychometrika,
of

recovery

J.

verbal

of

description
acquisition,

The
New

1952.
W.

McGill,

"

L.

learning.
Green,

Longmans,
G.

Miller,

A.

Irion,

"

human

of

psychology
York:

M.

R.

A.,

J.

McGeoch,

276-286.

Gagne,

"

15.

theory
learning.

in

60,

1953,

Rev.,

J.

variability

Psychol.
9.

C.

Burke,

"

stimulus

331

ESTES

N.

Miller,

E.,

S.

Stevenson,

"

tated
Agi-

S.

26,

1940,

behavior

of

during

rats

mental
experi-

251-280.
extinction
10.

and

York:

New

of

Spontaneous

extinction

in

reinforcements,

13.

of

D.

"

I.

Pavlov,

diana
In-

E.

W.

1927.

Press,

cence,
reminis-

for

theory

other

and

regression,
Psychol.

sive
Succes-

K.

reflexes.
London:

Anrep.)

single

act

1950.

EsTES,

V.

Univer.

Oxford

Saltz,

1953.

Conditioned
G.

perimental
ex-

York:

New

Press,

P.

by

(Trans,

20.

differences.

finite

of

Chelsea,
W.,

Univer.

Oxford
19.

in

theory

and

psychology.

tion,
acquisi-

thesis,

Ph.D.

Method

E.

C.

Osgood,

number

extinction

initial

of

Calculus

York:

Lauer,

spacing

18.

1953.

Univer.,
C.

Jordan,
New

to

Unpublished

period.

12.

relation

duration

and

recovery

205-231.

21,

1936,

from

taneous
spon-

Psychol.,

comp.

1952.

Harper,

E.

L.

Homme,

of

curve

/.

recovery.

11.

ing.
learn-

of

psychology

The

R.

E.

Guthrie,

Rev.,

1953,

nomena.
phe159-

60,

171.
extinctions

and

acquisitions

of

21.

habit

jumping

in

relation

to

M.

Schoeffler,

compounds

to

of

/.

reinforcements.

48,

1955,

of

Probability
of

comp.

/.

exp.

Psychol.,

48,

1954,

8-13.
323-329.

14.

Lauer,

D.

W.,

"

W.

Estes,

K.

of

Rate

22.

successive

learning

discrimination

F.

B.

Skinner,

in

relation

to

trial

1953,

8,

York:

spacing.
Crofts,

Psychologist,

The

384.

behavior

of

versals
re-

New

Amer.

sponse
re-

discriminated

physiol.
stimuli.

Psychol.,

S.

schedule

1938.

stract)
(Ab(Received

April

18,

1954)

Appleton-Century-

isms.
organ-

VARIABILITY

STIMULUS

OF

THEORY

W.

K.

AND

ESTES
Indiana

There

are

by

ments
learningexperirecognizedas important
in

stimulatingsituation

J. BURKE

University

of aspects of the

number

C.

LEARNING

IN

others.

Statistical theories of learning

differ from

Hull

in

making

way
1

are

to

the necessary
tions
condito
as
ble
plausigenerally appear
for
drawn
from
are
learning
tiguity
conbut they have not gained wide acceptance
theories
from
reinforcement
or
ing,
of learninvestigators
among
certain
characteristics
of the
theories,
possiblybecause Guthrie's assumptions
invariant
with
spect
relearning process are
been
formalized in a
have
not
while
other
stimulus
to
properties
them
make
that would
easilyused
characteristics depend in specificways
This paper
is based upon
a
reported upon
paper
the
of the
nature
stimulating

by the writers
Institute

and

at

the

Boston

of Mathematical
The

lines has

has

meetings of the

situation.

ber
Statistics in Decem-

writers' thinking along these

1951.
related

research

been

been

stimulated

facDitated

by

and

their

The

participation

in mathematical
interuniversityseminar
at
for
behavior
met
theory which
of 19S1 and
Tufts
College during the summer
sponsored by SSRC.
was

in

central

stimulus

be used
variability
concept
for explanatory purposes
than
rather
points
by theorists of otherwise diverse viewit
and
of
as
a
but which
source
treating
they
explicit
error,
require
resentation
repand Guthrie
in attempting
model
for effectivego beyond Skinner
in a
formal
that
One
to construct
utilization.
a formalism
find, for
may
will permit unambiguous statements
of
example, in the writings of Skinner,
about
stimulus
variables
of
clear
and
Guthrie
assumptions
recognition
Hull,
and
of
the conthe statistical character of the stimulus
sequences
rigorous derivation
these
of
All
conceive
a
assumptions.
stimulating
concept.
It has
been
shown
in a
of many
situation as made
previous
nents
compoup
that
several
pendently.
less inde(7)
which
pects
or
quantitativeasmore
paper
vary
of learning,for example the exthis locus of agreeFrom
ponential
ment,
of habit growth regucurve
strategies
diverge. Skinner (17)
larly
obtained
in
certain conditioning
incorporates the notion of variability
of
into
his stimulus-class
experiments, follow as consequences
concept, but
statistical assumptions and need not be
little use
of it in treatingdata.
makes
accounted
for by independent postuHull states the concept of multiplecomponents
lates.
All of the derivations were
ried
carexplicitly(13) but proceeds to
tions
write postulatesconcerning the condiout, however, under the simplifying
of a
of single assumption that all components
of learning in terms
the
stimulatingsituation are equally likely
leaving a gap between
components,
trial. By removing
to
fined
deand
occur
on
formal
experimentally
any
theory
that
in a posivariables.
Guthrie
now
we
are
restriction,
tion
(11) gives
the theory
to generalizeand extend
verbal
nomena,
interpretationsof various phein several respects. It will be possible
e.g., effects of repetition,in
show
that regardlessof whether
to
sumptions
asthese interpretations
of stimulus variability;
terms
that

Set

Generalized

Assumptions

and

Model:

Notation

an

models

This

article

appeared

in

The

exposure

of

stimulating situation

Psychol. Rev., 1953, 60, 276-286.


332

an

organism
determines

Reprinted with

to
a

set

permission.

W.

of

referred

events

ESTES

K.

AND

C.

that any change in the


shall attempt to deal

It is assumed

collectivelyas

to

333

BURKE

J.

situation (and we
constitute
plines
ing
only with controlled changes correspondspecial disciof
to
experimental
concerned with vision,audition,
manipulations
tion
distribumodel
a
wish to formulate
new
variables) determines
We
our
etc.
of values of the 6i. By repeating
mation
of the stimulus situation so that inforthe same
these special disciplines the "same"
mean
from
we
situation,
described
in
as
be fed into the theory, although
physical terms, and we
can
will depend
utilization of that information
speaking,repetition
recognizethat,strictly
refers
of
the
situation
to an
demands
of
the
exsame
learning
upon
which
be
idealized state of affairs
can
pferiments.
shall make
approached by increasingexperimental
For the present we
only
completely
the followingvery
general assumptions control but possibly never
realized.
the stimulating situation:
about
(a)
It is recognizedthat some
of
sources
The effect of a stimulus situation upon
internal
the
stimulation
to
made
are
be
ism.
organregarded as
an
organism may
This means
that in order to have
events,
of many
(b)
component
up
a
reproduciblesituation in a learning
When
a series
a situation is repeated on
to control
these
of
experiment it is necessary
of trials,
one
component
any
These

stimulation.
data

the

of

stimulus

events

trials and
first

events

various

the

occur

may

fail to

others; as a
tive
least,the rela-

occur

when

the

stimulus

(as

situation

same

experimentally)occurs

on

fined
de-

series

pendent
be representedby indetrials,
may
formulate
probabilities.We
lows:
these assumptions conceptually as folof

and

preceding

any given organism


elements.^
set 5* of N*

elements

N*

of 5*

stimulus

of the
in that

each

represent all

that

events

organism

with

to

are

can

of these

values

by di
event
of S*

N,
and
the

In

we

assume

have

noted

distribution

stimulus

the

of

6; we represent
probabilitythat the stimulus
corresponding to the i^^ element
occurs
on
given trial.
any
the sequel, various sets will be designated
by the letter 5, accompanied by priate
approsubscripts and superscripts. The letter
the

same

arrangement

set.

denotes

"trial"

the behavior

on

situation

certain

to be

elements
trial.

that

sampled

probability

those

elements

upon

the

of

of the parameter

superscripts,always

that

given trial is assumed

possibleevents

the

with

use

trial.
on

function

which

If in

elements
of

have

behavior

correspondingto an element of the set.


this reason
For
(b) For any reproduciblestimulating specificsituation
situation

present

the term

of 5*

are

given
have

occur

situation whatever

in any

the

In

ganism
or-

immediately

sociate
as-

The

trial.

the

the

of

extended
to necessense
sitate
sufficiently
the
distribution
in
6
including
movement-produced-stimulation arising
the
from
the responses
occurring on
in

we

also activities

shall not

we

paper

We

(a) With

schedule

maintenance

on

approximation, at
frequenciesof the various

events

the

some

on

set 5.

An

being sampled,
negligibleeffect

in

we

by

that

often
means

element

situation.

represent
of
of 5*

a
duced
re-

is in 5

value of
only if it has a non-zero
the given situation.
These
sets
nection,
representedin Fig. 1. In this conthat a probmust
note
we
ability
of zero
for a given event
does

if and
6

in

are

not

mean

that

the

event

can

never

"accidentally";this probability
has the weaker
tive
meaning that the relaof subscripts
of the
frequency of occurrence
the size of
is zero
in the long run.
For a
event
occur

334

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

3 +.1

Fig.
5

set

1.

schematic

representation of stimulus elements,the stimulus space 5*, the reduced


with non-zero
and the response
0 values for a given stimulatingsituation,

containingelements
classes A

partitionof

more

and

detailed

the reader

A.

The

S into Sa and

joiningelements of S

arrows

explicationof

this

is referred to Cramer

It should

events.

be

different

many
For

associated

example,

with

The

point

(5).

clearlyunderstood that
the probability,
6, that a given stimulus
trial may
event
occurs
on
a
depend
upon

to the response

classes represent the

S^

environmental
stimulus

The

response

previous paper

for its

model

(7)

formulated

will be

used

in

here

We
of

important modification.
any
shall deal only with the simple case
tive
two
mutually exclusive and exhaus-

event

may

Model

without

response

visual stimulation

Response

classes.

class being recorded

in

The
a

response
tion
given situa-

probabilityupon several
will be designatedA and the complementary
in the environment.
light sources
The
dependent
class,A.
lus
Suppose that for a given stimuvariable of the theory is the probability
the
associated
element,
probability that the
occurring on a given
response
^ in a given situation
depends only trial is a member
of class A.
It is recognized
two
separatelymanipulable components
upon
that in a learning experiment
of the environment, a and
b, the behaviors available to the organism
and
that the probabilities
ment's
of the eledifferent
be
classified in many
may
being drawn if only a or b alone
the
interests
of
depending upon
ways,
were
present are 6a and ^s,respectively. the
experimenter. The response class
Then
the probabilityattached
to this
be anyselected for investigation
thing
may
element
in the situation with both components
from the simplestreflex to a complex
present will be
chain of behaviors
involvingmany
different groups of effectors. Adequate
6
Oa-\- Ob
daQhdepend

different

"

336

READINGS

correspond

to

and

system,

speak

of

case

this

of

parameter

we

evaluated

of

the

not

where

havior
beto

cease

the
be

it cannot

relative frequency. It

as

MATHEMATICAL

probabilityin

as

situation

do

IN

PSYCHOLOGY

that
specifythe probabilities
element
that is sampled

on

trial will become

to

A.

For

shall

connected

convenience

limit

ourselves

lus
stimu-

any

given
to

or

in

exposition,we

in

this paper

to

a homogepreviouspaper (7) the simplestspecialcase, i.e.,


neous
trials with
series of discrete
be related in a simple manthat p can
ner
ments
that all eleto rate or latency of responding in
probabilityequal to one
nected
tions
occurringon a trial become consituations;thus in all applicamany
A.
to response
of the theory, p is evaluated in
with the rules prescribedby
We
accordance
begin by asking what can be said
about
the course
of learningduring a
the theory,either from frequency data
and
trials
tribution
from
other
of
or
appropriate data,
regardlessof the dissequence

has

in

shown

been

matical
is treated for all mathe-

evaluated

once

as

purposes

probability.

of
be

shown

define
Representation

Learning

of

stimulus

that

our

familyof

events.

It

will

generalassumptions
mathematical

tors
opera-

describinglearningduring any prescribed


t
he
member
of
trials,
sequence
the family applicablein a given situation
depending upon the 6 distribution.
shall first inquire into the characteristics

Processes

gradual of
of learning in most
course
situations,
number
the
earlier
of
a
quantitative We
of a
to all members
common
theories,e.g., those of Hull (13), Gulliksen and
Wolfie
family, and then into the conditions
(10), Thur stone
under
which
the operators can
be apassumed
individual
have
that
nections
con(18)
proximate
formed
adequatelyby the relatively
gradually over
are
a
series of learningtrials. Once we adopt
simple functions that have been found
convenient
for
statistical
view
the
of
a
representing learning
stimulating
data in previouswork.
be
shown
situation,however, it can
Let us consider the course
of learning
rigorouslythat not only the gradual
the
in
but
the
form
the
of
trials
of
of
course
during a sequence
learning
be accounted
simplifiedsituation. Each trial in the
can
typicallearningcurve
series is to begin with the presentation
for in terms
of probabilityconsiderations
that
of a certain stimulus
if
assume
even
we
tions
conneccomplex. This
In

order to account

are

basis.
to

This

be

would

no

formed

on

an

all-or-none

being the case, there


evidence
whatsoever

require a

formation

for the

of

seems

that

postulate of gradual

individual

connections.

situation defines
5*

so

that each

distribution

of 6

element in 5* has

over
some

di,of occurringon any trial,


probability,
and
we
represent by S the subset of
elements
with non-zero
6 values; any

element
that occurs
on
a trial becomes
Psychologicallyan all-or-none assumption
connected
has the advantage of enabling us
to ^
(or remains connected
has
drawn
A
if
it
been
for
the
that
to
fact
to
account
on
a previous
readily
reader
the
situations
in some
learning is sudden
trial). For concreteness
and gradual in others;mathematically, might think
of a simple conditioning
it has the advantage of great simplicity. experiment with the CS
preceding the
For
these
statistical US
recent
by an optimal interval,and with
reasons,
is
theories of learninghave adopted some
conditions
arranged so that the UR
each
trial
and
decremental
form of the all-or-none assumption (3, evoked
on
the situation repfactors are negligible;
7, 15).
resented
Under
all-or-none theory,we must
an
by S is that obtainingfrom the

W.

of the CS

onset

The' number

of elements

US,

shall

connected

at the

This

p equal to

zero.

"*"*element

after the

Sj

it is not

in 5 will stillremain
w**' trial if and

only

"

element

is connected

(3)
The

to A

F,{n)

(1

to

E[Ni{n)],

will be

ed-.

from

individual

[1

(1

e.)"]

now

in

iv

E (1

positionto

^i)".

express

p, the probabilityof response ^, as a


function of the number
of trials in this
for the term
By substituting
of
its
Fi{n)
equation (1)
equivalent
from
equation (3), we obtain the relation

curve

reduces to

(6)

p{n)

which, except
is the

(1

to

and
no

for

equal

^)"

notation,
previously

in

change

function

same

for the

(7)

have

derived

and

6 case*

sponds
corre-

linear operator used by


Mosteller (2) for situations
the

factor

decremental

is

volved.
in-

the

relation

same

observed

to

Hull's
probability
responding
theory as in the present formulation.
tion
Except where the distribution funcof the 9i either is known, or can
be assumed
theoretical grounds to
on
be approximated by some
pression,
simple exwill
be
not
venient
conequation (5)
with.
In practicewe
to work
are
apt to assume
equal di and utilize
describe
to
equation (6)
experimental
data.

The

in

nature

of the

of approximatio

error

involved in doing this

can

be stated

generally. Immediately after


first trial,
the curve
for the general

the

must

case

4
=

can

situation.

(5) P{n)

$i. It

verified

of

(4) E[_NA{n)'] i:Fi{n)

are

independent

the

in 5

elements:

We

there is

where

of elements

contributions

of

by substitution that
fixed point at /"
a
1, and this
the asymptote approached by
of /"(") vs. n as " increases

will be
the

possible

number

In mathematical
form, equation
Hull's wellas
(6) is the same
after the
n^^ trial, known
expression for growth of habit
the sum
of these expected
strength, but the function does not

expected number

connected

easilybe

Bush

distribution of

after the n'''

obtain:

we
trial,

the

on

trials;the
{\
6iY. Hence, if Fi{n) represents
the expected probabilitythat this

is

of

of

family

each

if

any of the first n


likelihood that this occurs

sampled

Equation (5) defines


for
learning curves, one
6 distribution,
and it has
simple propertiesthat are

this simplification
negativelyaccelerated curves, approaching
a
easilybe extended
simple negative growth function
initial
the
Bi tend toward
as
arbitrary
equality. If all
of the 9i are
equal to 6, equation (5)

condition.
The

that the learning

337

BURKE

J.

in

results may
of any
case

the

to

are

C.

all bounds.
all begin with N^.
Members
of
the
over
will
No
loss of erality
be monotonicallyincreasing,
family
gen-

is involved
our

in S

beginning of the

means

obtained

curves

and

of the elements

to A

experiment.

simplicitywe
tions
following deriva-

the

in

suppose
that none

in S will

For

N.

designatedby

in

of the

to the onset

AND

the response probabilityp will refer


tion.
of ^ in this situato the probability

and

be

ESTES

K.

This

j^^^el\-{\-e:)--]

is
for

lie above

the

the
essentially
the

equal $

that

paper.

function

same
case

(7) ; the terms 9 and


correspond to the terms
q

paper

for the

curve

in

of

s/S,

veloped
de-

previous

equation (6)
and

of

338

PSYCHOLOGY

MATHEMATICAL

of trials for the numerical exin S, as a function of number


amples
Response probability,
is the exact solution for a population of elements,
presentedin the text. The solid curve
The dashed
describes the equal $ approximacurve
of which have ^ = 0.1 and half ^ = 0.3.
tion

Fig.

half

IN

READINGS

2.

with

0.2.

Initiallyno

elements

of S

are

to A.

conditioned

involved

ple
of a simby means
numerical example. Imagine that
increases for a few trials,
two
curves
then decreases until they cross
structing we
are
dealing with some
particular
(in condistributions
of
in
which
the
6
conditioning experiment
hypothetical
have
CS can be representedby a set 5, comdiverse
forms we
posed
usually found
ments,
of two
this crossing in the neighborhood of
subsets of stimulus eleand
the
after
of
sizes
the fourth to eighth trial)
ing,
crossSj,
5i
N^
N^
;
of elements
the curves
tent
diverge to a smaller exN/2, where N is the number
Assume
that for all elements
than before,then come
in S.
togetheras
the
both
the
at
in
of
to
same
probability
being
5i
as3niiptote
go
be
drawn
I.
It can
0.3 and for
proved that the
on
p
any trial is ^^
those in Sg, $2
wish to
for the general and specialcase
^"^- Now
we
curves
from
the
n
one
as
cross
exactly once
predicted learning curve
compute
goes
make
cannot
to infinity.We
sponses
during a series of trials on which A reeral
any genabout
the maximum
statement
ror
erare
reinforced,assuming that
involved
in approximating expreswe
begin with all elements connected to
sion
(5) with expression(6), but after /4. Equation (S) becomes
of specialcases, we
studying a number

equal

6 case;

the difference between

the

siderat
=

inclined to

are

believe that

the

error

p(n)

approximation will
be too small to be readily detectable
ing
experimentally for most
simple learnintroduced

by

situations

the

that

do

not

involve

l-^lNr{0.3)
X(l-0.3)"+iV2(0.1)(l-0.1)'^]

pounding
com-

l-^[0.3(0.7)"+0.

of stimuli.
The

development of equations (5)


(6) has necessarilybeen given in
rather general terms, and
be
it may
of the conhelpful to illustrate some

and

Plotting
from
curve

Now

this

numerical

computed

values

equation,we
given in Fig. 2.
let us approach

obtain

the

the solid

same

prob-

W.

ESTES

K.

AND

supposing this time that we


the
different 6
know
nothing about
values in the subsets S^ and 5, and are
We
tain
ob0.2.
now
given only that d
under
predicted learning curves
the equal 9 approximation. Equation
lem,

but

(6) becomes:

p{n)

values

numerical

and

(1

yield the dashed


Inspectionof Fig.

this

computed from
of Fig. 2.
curve
that

shows

leads

treatment

exact

0.2)"

to

the

higher values

early trials but to lower


the difference
values on
the later trials,
The
for
large n.
becoming negligible
in brief,for the steeper curvature
reason,
of p{n)

of the exact
with

high
drawn, and

is that elements

curve

values

conditioned

therefore

earlier in the learningprocess


low

with

values,and

will tend

to

be

likely to

are

recur

A,

to

than

ments
ele-

then

cause
be-

frequently

J.

339

BURKE

selves
theory developed above, limiting our6
the
to
equal case.
As written,equation (6) represents
of conditioningfor
the predictedcourse
sponse
a
single organism with an initial reWe
can
probabilityof zero.
that an experiallow for the possibility
ment
value of ^(0)
begin at some
may
other than zero
by rewriting(6) in the
form
more
general

(7)

p{n)

which

the

on

they

C.

=:

has

the

of

course

to

(8) Pin)

ey

(6)

as

consider

conditioningin

organisms with
varying initial
need simply
we
the group

form

same

wish

we

pmn

cept
ex-

initial value.

for the
If

[1

like

mean

of

group

values

of

but

probabilities,

response

equation (7) over


by m, obtaining

sum

divide

and

the

^Zpin)

in successive

high
late

the

of

would

be
an

that

have

early trials will

unconnected

with

values

lead

p.

elements

not

During
been

contribute
per

tively
rela-

to

stages of learning,elements

low 6 values
on

samples, to

with
drawn

the
The

[1

standard

p(o)](i

deviation

these circumstances

is

ey.

of p{n)
simply

der
un-

more

trial than

appearing at the same


stage
equal 6 distribution and will

(9) "tM

P'(n) pKn)
\Jl^i:
-

^)Vp(O)
(1
depress the value of p below the curve
where
for the equal 6 approximation.
o-p(O) is the dispersionof the
bility
Variainitial p values for the group.
be
It should
emphasized that the
around
the
curve
mean
learning
the
to
of
generality
present approach
in
decreases
manner
to zero
a
simple
troduced
learningtheory lies in the concepts inand
the methods
developed as learningprogresses.
ing,
The
of counter-conditiontreatment
for operating with
them, not in the
extinguishingone response by
i.e.,
tion
Equaparticular equations derived.
to a competing
(5), for example, can be expected giving uniform reinforcement
follows
automatically
to apply only to an
response,
extremely narrow
from
of the acquisition
the
account
class of learningexperiments. On
our
riving process.
utilized in deother hand, the methods
Returning to equation (6)
that the probabilities
of A
and
to
recalling
equation (5) are applicable a
and
A
terest
to unity, we
wide varietyof situations.
For the inmust
always sum
that while response
A undergoes
note
of the experimentally oriented
with
accordance
in
will
few
of
indicate
(6),
conditioning
brieflya
reader,we
extinction
in
the
of
A
the
obvious
extensions
must
most
undergo
response
=

340

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

between

with the function

accordance

of

terms

the

respond

theory and experimentalvariables,will


1
(1
e)\
pA{n)
plin)
be
experimental evaluation
possible.
If,then, we begin with any arbitrary Limitations of space preclude a detailed
theoretical analysisof individual learning
p(0) and arrange conditions so that A
=

conditioned

and

is evoked

drawn

to

to

in this paper.
In order
how
the model
will be utilized

indicate
and

will be

to

suggest

of

some

its

planatory
ex-

shall

function

simple decay

situations

all elements

tinction
trial,the exgiven by

each

on

of response
the

we
clude
conpotentialities
few
a
general remarks
of learning
concerning the interpretation
within
the
theoretical
phenomena
framework
have developed.
we
Applicationof the model to any one
isolated experiment will always involve
element
for information
of circularity,
an
about a given 6 distribution must

with

p{n)

(10)

p{0){l

e)\

and standard
deviation
Again the mean
for a
be
of p{n) can
computed
easily
like
values
of
of organisms with
group
6 but varying values of p{0):

(11)

p{n)

(12)
As

crM

in the

around
zero

case

the

in

p(0){l-dy

(1

of

acquisition,
variability

mean

curve

simple manner

be

^)Vp(O).

decreases
over

to

series of

obtained

from

behavioral

This

disappears as
circularity

data

are

related

available

from
The

data.
soon

number

as

of

utilityof the

experiments.
is expected to lie in the possibility
of predictinga variety of facts

concept

trials.
Since

due to variation in
variability
the parameters of the 0 distribution
once
ing
p{0) is reduced during both conditionhave
been
evaluated
for a situation.
and counter-conditioning,
it will be
involved
has
The
methodology
that in general we
should
seen
expect
been illustrated on a small scale by an
of reless variabilityaround
a
curve
6
experiment (6) in which the mean
of original
learningthan around a curve
value for an operant conditioningsituation
learning for a given group of subjects.
estimated
from the acquisition
was

bar-pressinghabit and then


of
predicting the course
bar-pressing
acquisition of a second
under
habit
animals
by the same
modified
conditions.
slightly
curve

Application

of

Model

to

the

Statistical

Learning

Experiments
Since
been

our

with

in this paper
has
the development of a stimulus
concern

of

utilized

When

in

the

statistical model

is taken

tion
togetherwith an assumption of associahas been necessary
have
the
in the interests of
tials
essenby contiguity,we
clear exposition to omit
reference
of a theory of simple learning.
to
of the empirical material
The
most
learning functions (5), (6), and
upon
which
theoretical
our
(10) derived above should be expected
assumptions are
based.
The
evaluation
of the model
to provide a descriptionof the course
detailed interpretation of learning in certain elementary exmust
rest upon
periments
of specificexperimental situations.
It
in the areas
of conditioning
is clear,however, that the statistical
be emand verbal association.
It must
phasized,
model
developed here cannot be tested
however, that these functions
in isolation;
will not constitute an
gether alone
adequate
only when it is taken toof
with
for a number
assumptions as to how
theory of conditioning,
and with rules of correlevant variables,
especiallythose conlearningoccurs
model

of considerable

it
generality,

W,

been

not

tions.
deriva-

our

factors

decremental

(1,4, 9, 14, 16)

that the

mized,
mini-

are

evidence

is considerable

there

of

curve

C.

AND

conditioning experiments

In

where

in

into account

taken

have

decrement,

trollingresponse

ESTES

K.

ditioning
con-

argument

equation (5) and can


by the equal

The

that

fact

the model
curves

from

fitted to certain

be
is

(7).

case

pirical
em-

desirable outcome,

of

be

regarded as providing
of
the
test
a
exacting
very
theory; probably any
contemporary
to accomplish
quantitativetheory will manage
but

course,

cannot

this

much.

the

On

detail in the

be noted

that the

tain
theory yieldscerspecificpredictionsconcerning the
effects of past learningupon
the course
situation.
of learningin a new
In general,

the

increment

extent

end

in p

depends to a certain
immediately preceding
trials. Suppose that we

the

upon

of
sequence
identical
have
two
which

decrement

or

trial

during any
proximated
ap-

well

derived

functions

can

be
6

in mathematical

present paper, it may


statisticalassociation

has the principalpropertiesof


our

341

BURKE

J.

animals

each

of

/"(")equal,say, to 0.5 at the


trial n of an experiment, and

has
of

that for each animal


suppose
trial n +
A
is reinforced
on
histories

of

the

two

animals

to differ in that

other

response
The
I.

are
sumed
prethe first animal

has
arrived
via a seat
quence
p{n)"O.S
hand, the fact that the propertiesof our
reinforced
trials
while
the
of
tistical
learning functions follow from the stasecond animal has arrived at this value
tion
of the stimulatingsituanature
of unreinforced
trials.
is of some
interest;in this respect via a sequence
animal
will
On
trial n 4- 1, the second
the structure
of the present theory is
receive the greater increment
to p (except
simpler than certain others, e.g., that
the
9
in
the
pendent
indewhich
of Hull
reason
equal case);
is,
(13),
require an
animals
the
in brief, that for both
for the
postulate to account
stimulus elements
most
form of the conditioningcurve.
likelyto occur
\
those
with high 6
tions
trial
It should also be noted that devian
+
are
on
ments
from
the exponentialcurve
form
values; for the first animal these eleoccurred
will
have
of
be
instances
frequently
as
a
s
significant
may
the present model
quence
during the immediately preceding sewe
good fit. From
of trials and thus will tend to
kind of deviation
must
predicta specific
be preponderantly connected to A prior
when
the stimulatingsituation contains
of the secelements
values.
of widely varying 6
to trial " -h 1 ; in the case
ond
of conditioning
animal, the high 6 elements will
If, for example, curves
to A
stimuli taken
to two
during the
separatelyyield have been connected
and
different values of 6, then
immediately preceding sequence
significantly
trial
is reinforced
thus
when
A
the curve
of conditioning to a comon
pound
of the two
to deviate

the

separate

growth

one

(16); Miller's

theory as
relevant

this
to

The

only

results appear
analysis,but

substantiated

Although

become
we

pected
ex-

simple

relevant

discovered

this

be

either of
a

in

reported by

regard

data

than

from

have

we

literature is

hesitate

further

curves

function.

line with

stimuli should

to
we

aspect

periment
ex-

the

Miller
be

in

would
of

the

until additional

develop

-\-\, the

second

the

animal

will receive

in weight of conthe greater increment


nected
elements.
this analysisit
From

that,other thingsequal,a curve


reconditioning will approach its
rapidly than the curve
asymptote more
tion
of originalconditioningunless extinchas actuallybeen carried to zero.
How
important the role of the unequal
follows

of

distribution will prove

to

be

in

counting
ac-

empirical phenomena of
relearningcannot be adequately judged
for

available.

shall not

342

until

further

means

for

has

research

provided

estimating the orders

of the effects

3.

R.

Bush,
for

of magnitude

R.,

"

stimulus

F.
Mosteller,
generalization and

Psychol. Rev.,

mentioned

have

we

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

model

nation.
discrimi-

1951, 58, 413-

423.

here.

.4.

Decremental

Calvin, J. S.

factors in

con-

ditioned-respo^nselearning. Unpublished
Summary
statistical

Earlier

Ph.D.

learning

associative

of

treatments

have

5.

ple
sim-

thesis,Yale
H.

Cramer,

statistics.

been

Univer.,

1939.

Mathematical
Princeton:

methods
Princeton

of

Univer.

fined
re-

Press, 1946.
and
stimulus

generalized by analyzing

heretofore
fact

in greater detail

concept
and

by taking

different

that

the
6.

of

components

corresponding

to

bar

statistical model
relative

The

set.

which

various

variable

7.

with

given
set

model,

assumption
contiguity, provides

of
a

9.

1950.

D.

A.,

Grant,

10.

together
association
by
limited
theory

E.

Guthrie,

1946, 43,
Hilgard,

characteristics

the

of

R.

E.

C.

York:
14.

here

R.,

"

other

quantitative
learning.

Brogden,
In

J.

Miller,

G. A.,

of

16.

Miller,

S. S. Stevens

of

human

York

R.

R.,

"

model

Skinner,
Crofts,

18.

Bush,

Hosteller,
for

F.

W.
verbal

J.

313-323.

/.

tistical
sta-

learning.

in press.
rate

subjects

to

stimuli.

of conditioning of
single and
multiple
/.

gen.

Psychol.,

399-408.

B.

F.

The
York:

behavior

of

isms.
organ-

Appleton-Century-

1938.

Thurstone,
/. gen.

matical
mathe-

L.

L.

The

learning function.

Psychol., 1930, 3,

469-493.

simple learning. Psychol.

Rev., 1951, 58,

tinction
ex-

in

Wiley, 1951.
2.

and

ing.
learn-

(Ed.), Handbook
New

New

1943.

conditioning.

to

McGill,

"

The

J.

New

of experimental psychology.

Ap-

expectations

description of

1939, 20,
studies

tioning
Condi-

York:

Acquisition

verbal

analogous

conditioned

Animal

G.

exp. Psychol, 1939, 25, 294-301.


15.

17.
1.

Bull.,

1940.

G.

L.

of

REFERENCES
W.

New

Appleton-Century,

with

formulations

D.

Marquis,

learning.

Humphreys,
situation

facts and

Psychol.

Principles of behavior.

L.

D.

model

compared

are

127-149.

theory.

Psychometrika,
elaborated

Psychol., 1951,

exp.

Psychological

pleton-Century,
Hull,

forcement
rein-

eyelid

1-20.

and

13.

tion
adap-

H.,
of

psychological

12.

Dark
random

in human

chometrika, 1938, 3,
11.

quisition,
ac-

extinction,and relearning,are
compared with experimental findings.
Salient

/.

York:

A
D.
L.
"
Wolfle,
transfer.
learning and
Psy-

Gulliksen,

taken

theory concerning

W.

Humphreys

conditioning.
42, 417-423.

periment
ex-

conditioning phenomena.
theory it has been possible
of the learning
to distinguish aspects
that
depend
properties of
upon
process
situation
from
those
the
stimulating
that
do
Certain
not.
general predictions
the

H.

Hake,

"

the

probability

to

applications. New

phenomenon

tions
opera-

1950, 57,

Rev.,

introduction
its

Wiley,

this

from

statistical theory

Psychol.

An

and

certain

Within

Toward

theory and

theory

an

of

K.

learning.

Feller, W.

functions.

statistical

The

8.

stimulus

represented by

are

and

in

actions
re-

for

curve

94-107.

events

the

of

behavior

competing

pressing. J. exp. Psychol., 1950, 40,

EsTES, W.
of

by a mathematical
frequencies with

aspects

affect

of

conditioning

200-205.

perimental
independent exis represented in

an

variable
the

stimulus

of

population

Effects
the

on

have
ent
differstimulating situation may
probabilitiesof affecting behavior.
The

K.

than
of the

account

W.

EsTES,

[MS.

received

November

12, 1952]

344

READINGS

trial. The

dependence of

the

upon

S's responses

stimulating situation

is

pressed
ex-

theory by defining a
such that each
relationship

conditional

in Sc is conditioned

that

evokes

when

from

which

it evokes

Then

on

expect

(1) that
Sc on the

which

of

If

Ei

relation

of Ss begins

group

with

the value

of Trial

-\- I will be

7r[(l-5)p(")+ 0]
(l-,r)(l-0)p(n)
"(1-6) Pin) +"T.

experiment

an

p(0), then

the end

at

have

would

we

(3)

E2

p(i)

of class Aj.

response

occurs

we

the end

at

will

trial will become

be

that

so

on

ii-e)viO)+6T,

of Trial 2

trials

-e)viO)

A2.

to

Sc

from

are

+ dir
e-ir']

ditioned
con=

and

so

that

[t

successive

be shown

can

the end

at

p(0)](i

for

on

sufficientlygeneral it

are

samples

(i-0)[(i

E2 trial the

an

conditioned

if successive

discrete

ciples
prinp(2)
sampled

all elements

Ai while

to

sample

p(n+l)

bility
proba-

average

the basis of association

on

from

the response
that when
an

trial on

given by the

to

is

compatible
predictingEi but

of

and

predictingE2,

Now

belonging

response

the

of Ai after Trial

it

occurs

interferes with

occurs

(1"jt);then

class Ai, i.e.,


one
with the response
which

Ei

an

situation.Equation 1 will
applicableon the proportion tt of
trials and Equation 2 on
the proportion
be

(tends

to

to evoke) either Ai or
A2. In order to
of
interpretthe formal model in terms
verbal conditioning experiment, we
a
assume

PSYCHOLOGY

forcement

the

in

element

MATHEMATICAL

IN

e)\

trials;in

by

induction

of the nth. trial

statistically
independent, the probability
of

Ai after Trial

an

p(n), is

defined

ated
n, abbrevi-

in the

the proportion of elements

model

as

Since

(1

6)

"

in Sc that

Ai,

to

and

for the

similarly that
Equation

of an A2, [.l"p{n)^.
probability
accelerated
these definitions the rule for calculating
initial value
the change in response
ability
prob-

With

value

on

Ei trial

an

be stated

may

formallyas

and

on

E2 trial

an

Pin

1)

(1)

{l-e)p(n)+d

as

(2)

(l-e)p(n).

This

genesis of these equationswill be


The proportion (1-^)
fairlyobvious.
of stimulus

the

and

sampled

elements

on

trial does

the proportion 6 is
elements

are

is not

of elements

status

not

all conditioned

not

are

change;
and

these

either

Ai or to A2 accordingly as an Ei
Now
in a random
E2 occurs.^
Consequently
should be expectedto apply only to

paper

the

be

be

seen

negatively

running from

curve

p(0)

to

the

the

asymptotic

of the statisticallearning

surprisingat first
since it makes
asymptotic response
the
probabilitydepend solely upon
reinforcement.
It
o
f
probability
seems,
however,

is rather

be in excellent

agreement
results
of Grant,
experimental
to

(3) and Jarvik


Hake, and Hornseth
(5). The question that interests us

or

to
an

rein-

ing situations which

functions derived in this


leam-

are

lowing
symmetrical in the fol-

each response class there must


condition which, ifprescorresponda reinforcing
ent
To

sense.

on

any

to

the

that

ensures
trial,

class will terminate

response

ing
belong-

the trial. These

functions should,for example, be applicableto


discrimination with
learningof a simpleleft-right
discrimination
left-right
in the
free
responding
correction,
Skinner box, or to Pavlovian conditioning.

correction ; but
*

fraction between

it will

sampled,

that

sampled

4 must

outcome

with the
The

one,

(4)

T.

model

p(n-\-l)

be

must

and

zero

conditioned

are

p(O)](l-0)-.

Pin)=^-{jr-

without

not

to

to

regarded
as

or

We

as

estimate

cannot

conducted

not

were

the theory, and


that we would be

to

the

since

for the latter conclusion

test

TABLE
Design

Experimental

in

Terms

Series

Each

DURING

bility
Proba-

of

(ttValue)

Reinforcement

OF

level

confidence

345

STRAUGHAN

H.

J.

theory.

the

of

confirmation

AND

is to be
this agreement
coincidence
remarkable

is whether

now

ESTES

K.

W.

periments
ex-

cally
specificannot

we

alert to

as
guarantee
the theory
to
notice results contrary
in the literature
which might appear

in the

been

have

we

a"

of these

case

It has

decidedly positiveinstances.
seemed

to

out

of

of

predictionfor
situation,it

can

testable

one

given experimental
to
generally be made
In the

more.
yieldmany
to be reportedwe

experiment

tried to set

have

situation similar in essentials

separate
from that of

to

value
event

groups

to

1 has

in Table

been

into four subgroups of four

; within

treatment

group,

say

subgroups have the same


Group
but each receives a separate
value
"K
of Ei's and
randomly drawn
sequence
I, all

Ea's.
Method

that

Humphreys, Grant, and others


experimental design which
permit testing of a variety of
Each

of the theory.

Ss each

orders
particular

each of the three

occurrences,

indicated

up

used by
with
an
would

the effect of over-all

subdivided

theory,namely,

if it will generate

that

IT

of

tures
fea-

convenient

of the

one

mathematical

is to

ing
experiments,mak-

new

some

use

able
objection-

of this impasse

out

way

carry

that the least

us

In

differs.

reinforcement

preceding
order

in a
run
Apparatus. ^The experiment was
containinga 2-ft. square signalboard and
were
four booths.
Upon the signal board
12 12-v., .25-amp. lightbulbs spaced
mounted
"

room

The bulbs
evenly in a circle 18 in. in diameter.
occupied the half-hour positionsof a clock face.
used
fication Only the top two lightson the board were
of 120 trialsin an individualized modiin this experiment. The signalboard
signals
as
situation
of the Humphreys
table 40 in. high
mounted
a
on
vertically
was
with the schedule of ir values shown in
Ss'
booths.
of
front
in
5
ft.
about
and

consequences

through

run

was

successive series

two

was

Table

will be able
and

from
to

learning rates

compare

different

be able
different
same

of
probabilities
the second

ment
reinforce-

series

startingat

to

groups
compare
initial values but

will

we

exposed to

of reinforcement.
probabilities
Comparison of Group I with

the other groups

over

permit
learning rate
series when

{B value)
the

tt

series will

both

of
stability

of the

evaluation

value

from

change. Series Ia and


will provide a comparison

not

initial response
values are the same

series

does

or

which

probabilitiesand
but the

to

does

series IIb
in

amount

booths

The

we

of groups
starting
asymptotes
similar initial values but exposed
; within

the

to

first series

the

Within

1,

ir

of

from

made

were

two

30 X

60 in.

30 in. high,placedend to end but meeting


tables,
behind them would
at an
angle so that Ss sitting
toward the signal
board,
be facingalmost directly
about

7 ft. in front of Ss'


The four Ss

eyes.

table.

each

were

Two

Ss

sat

at

separated from

another by panels2 ft.high and 32 in. wide.


mounted
verticallyon the
panels were

one

These

table tops

edge of

so

as

extend

to

14^

in.

beyond the

the seated Ss.

the table between

each booth, 18 in. back from S's edge of


wooden
panel 12 in. high
a
table, was
ing
the table top and extendmounted
on
vertically
In

the

across

the width

of this panel facingS


of the

same

covered

lightswere

size

as

of the booth.

those

On

the side

lights
reinforcing
the signalboard but

two

were
on

by white, translucent lenses. These


in front of S, 4 in. apart and
directly

8 in. above
the table top. On the table below
each reinforcing
lightwas a telegraphkey.
The orders of presentationand the durations

of the signallightsand

lightswere
reinforcing

346

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

"Are you sure


understand
all of the instructions
corder
Esterline-Angusreby a modified
you
far.'' The
of the trials will
rest
so
using a punched tape and a system of
The
recorder
off without
electrical pick-up brushes.
have to be run
conversation
was
or
any
the
table
the
board.
behind
other
choice on
on
signal
a
placed
interruptions.Please make
activated by deprestrial even
Recorder pens which were
if it seems
difficult. Make
a
sion
every
of the telegraphkeys in Ss' booths
the first trial,
then try to improve your
were
on
guess
between
the brushes.
mounted
as
Thus, the preas
rect
corsentationsguesses
you
go along and make
many
of the lightsand Ss' responses
choices as possible."
were
the same
recorded
answered
on
Questions were
by rereading or
tape. A panel lightwas
the Esterline-Angus
mounted
above
recorder so
tions.
paraphrasingthe appropriatepart of the instructhe signalboard, could
If there were
that E, seated behind
questionsabout tricks
any
the followingadditional paragraph was
watch the operation of brushes and pens during
read.
"We
have
the experiment.
told you
everything that will
in the experimentalroom
Windows
were
ered
covhappen. There are no tricks or catches in this
with opaque
material and the experiment
how well
to
see
experiment. We simply want
in darkness except for lightthat came
cult
was
run
can
profitfrom experiencein a rather diffiyou
from the apparatus.
problem-solvingsituation while working
under time pressure."
48 students obtained
Subjects.The Ss were
in psychology
from beginning lecture courses
The recorder was
started again and the
now
of 1952 and assignedat
240 experimentaltrials were
off in a conrun
tinuous
during the fall semester
with no break or other indication
random
to experimentalgroups.
sequence
Procedure.
At the beginningof a session,Ss
S at the transition from
Series A to
to
were
brought into the room, asked to be seated, Series B. On each trial,the signallamps were
and read the followinginstructions:
lightedfor approximately2 sec. ; 1 sec. later the
"Be sure
seated comfortably;it will
are
lightin each S's booth
appropriatereinforcing
you
be necessary
hand
to
keep one
restinglightly lightedfor .8 sec; then after an interval of .4
the next
beside each of the telegraphkeys throughout the
sec.
ready signalappeared; and so on.
used
The high rate of stimulus presentation
was
experiment and to watch both the largeboard in
controlled

"

"

the front of the

will be
each

at
or
trial,
ready signalon
two
on
top lights

the

small

two

Your

compartment.

own

your

and

room

task

lightsin

in this

periment
ex-

sure

to

make

your

choice

as

soon

will give you


four practicetrials."
guished
extinpoint the overhead
lightswere

and

the

recorder
made

started.

If any

ous
obvi-

during the four


by
practicetrials,they were
pointed out by E.
During the four practice trials the reinforcing
lightswere
always given in the order: Ei, Ei,
E2, E2. After the practicetrials the following
instructions

were

were

Terminal

read;

be

Discussion

and

It
probabilities.

response

clear

for each

of

Tf

have

from

"

of

Equation 4 that the predicted asymptote

we

mistakes

the part of

the

is on.
"Now

on

left

the

At this

verbalization

will

ready signalappears, press the proper key


down
firmly,then release the key before the
ready signalgoes off. It is important that you
press either the left or the rightkey, never
both,
each trial,
and that you
make
decision
on
your
and indicate your choice while the signallight
as

minimize

Results

to

big board. About a second


the rightlamp in your
or
will
for
As soon
a moment.
light
compartment
the ready signal flashes you
to
as
are
guess
whether the left or the rightlamp will lighton
choice by pressing
that trial and indicate your
the proper
key. If you expect the left lamp to
light,press the left key; if you expect the right
not
lamp to light,
are
press the rightkey; if you
Be

to

Ss.

outguess the experimenteron


The
least as often as you can.
each trial will be a flash from the

later either the

sure, guess.

in order

series will be
the

obtaining during
taken

discussion

our

the

the

value

series.

We

proportion of Ai

mean

during the last 40 trials of


responses
of terminal
each series as an estimate
response
are

both

and
probability,

summarized

for

series in Table

Mean
FOR

Response

Each

values
and

TABLE
Terminal

these

all groups
2.

Series

Probabilities

W.

ESTES

K.

For the firstseries a simpleanalysis


of variance
yields an F significant
.001
level for diff'erences
the
beyond
estimate

variance

groups

value for the standard

each

between

series the

.6"
"*

falls

Group

series.

/^'

means

"

Group I in the
tests
computed
the last

significantly

I
2

I I
3 4

I
5

I
7

I
6
OF

BLOCKS

I I I I
9 10 II 12

I
8

(m)

TRIALS

20

in

the

theoretical

series,but

level
probability

same

P(m)" 300

"
yt
^
Pimh .850-260(.92)20"'n-'"

both

in

second

of

in the

had
t

Fig. 1.

series.

asymptote
reaches the

"p(m)-300*J73O82)

and theoretical

curves
Empirical
resenting
repapproximates it in
of Ei predictions
mean
(Ai
proportion
falls
II
Group
nificantly
sigresponses)
per 20-trialblock for each series

the first series but


short

iiT

"

short of the theoretical asymptote


the second

For

mean.

between-groups

asymptote

"

the

seems
forward.
straightinterpretation
III
approximates
Group

theoretical

"

20(m-l)

t test

and

differences among
subgroup
at the .05 level.
significant

the

"

the .05
between
F has a probability
and .10 levels. In neither series were

The

group

in the

appropriatetheoretical
the second

of

mean

group

P(mJ-.300*280(.982)^"

"^T

withinobtain

we

error

this is used

and

mean,

the

From

means.

among

347

STRAUGHAN

J. H.

AND

firstseries.

Of the

for differences
blocks

two

as

tween
be-

Taking the theoretical "r


ally
equal to "V407r(l tt),which is actuof the true
underestimate
a slight
value,we find that approximatelyhalf
distribution.

"

of 20 trials

of the scores
in each series fall within
all
yielded
probabilities
series,
and
one
a of the theoretical asymptote
greater than .10 except thef for Series
only one score in each series deviates
the .02
at
IIb which was
significant
It
than three a.
more
in each

level.

concerning
Evidently the predictions
mean
asymptotic values are

correct, but

the

rate

of

approach to

asymptote

is faster with

than

the other conditions.

under

Group III

According to theory,not only group


means,

but

also

should approach
To obtain evidence

individual
tt

curves

by
then,that except

appears,

for

ant
few widelydevi-

the p values of individual Ss


theoretical asymptote.

cases

approach the

One might raise


what

is meant

curve
empirical

kind.

asymptotically.the

Ss

Naturallyone
to

questionas

to

just

by the asymptote of an
in a situation of this

perform

would
at

not

constant

expect
rates

the

tenability indefinitely.It does not seem


that
have
of this aspect of the theory we
of
sort
breaking point was
proached
apany
examined
vidual
the distributions of indiever;
in the present study, howAi response proportionsfor the
one
subgroup of Group I was
last 40 trials of Series IIIa,IIIb,and
for an additional 60 trials beyond
run
^B. If allindividual p values approxiTrial 240 and maintained
mate
an
average
as

to

the theoretical asymptotes over


then for each of the series
these trials,

the

individual

an

.304 Ai responses

the

mean

over

these trials.

Mean
proportions
learning curves.
data
mean
are
value,
plottedin
of Ai responses
approximately binomial
proportion
response

should cluster around


TT, with

proportionof

"

In
terms

Fig.

of the

per block of

348

READINGS

trials. The

20

which

readilyobtained

is

curves

empirical
obtained
tion
Equa-

ordinal

Trial

by

20

Ai responses
=

p(0)](l

[7r

Equation

m,

obtaining for K

[1

(1

all values

over

of

blocks of trials

should

(5)

each

describe

of

Fig. 1

mean

of the

numerical

mean

curves

values

once

for the eters


param6; furthermore,the
differ
of 6 required should not
substituted

are

and
X, p{0),

groups

should be

from

constant

The

group.

series to series
values

of

are

of
Equation 7 to the sum
proportionsfor a given
solve for 6.
For Group I

then equate
the observed
series and

the estimate

we

obtain

for

IIIa,6

.08.

values

.018 and

Using these

eter
param-

have

computed the
theoretical curves
for Group I and for
the first series of Group III, which
be

may

we

in Fig. 1.

seen

In this analysis

find agreement
between
we
theory in one respect but

either series and

within

er-]

expressionbeing simply the

for each

be

procedure.
used is

P(i)](i
^)20('"-i)
Kir-lirP(l)]
(1
[1
fl)2o^]
(7)
1
(1
6y

g)a''("'-i)

the wth block of 20


value of p (n) over
trials. According to theory, Equation

among

pirical
em-

can

have

we

T. -^-b^-

20 6

value

sum

method

to

""=1

this

only

these

the

P(m)

P(m)

lack

we

of 6 and

simple statistical

The

ber
num-

n
+
inclusive,
expected proportionof
in the block,we can write

to

respectively.Now
estimates

block of 20 trials running from

Trial_"+
and

function

PSYCHOLOGY

from

Letting m be the

4.
of

theoretical

describe these

should

MATHEMATICAL

IN

and

another.

The

theoretical

data
in

not

curves

vide
pro-

reasonably good descriptionsof


by the experimental
in the
the observed
points,especially
procedure. The values of p (0) in the
the
of
but
values
for
6
case
Group
I,
in
the neighborfirst series should be
hood
the
two
are
no
means
by
equal.
of size 16
groups
of .50,but for groups

of

fixed

course

sampling deviations
large

so

in favor

p{0)

measured
this

could

be

quite

best to get rid of


of P(l) which
be
can

it will be

accurately. To do
1
write Equation 5 for m
more

we

The

latter finding does

in

asymptote

Group III
P(l)

[1

20 6
solve for

then

did

We

Cx-p(O)]

[x

(1

p(0)]

20e[x-P(l)]
(1

the

not

try

substitute

this result

to

and

6y

into

estimate

values

for

value

without

using

Series II b
used

above,

interest

more

predictedcurves
from

estimate

to

IIIb by the method

tion
Equa-

5 giving

P{m)

series,while

but it will be of
and

first

Group II since the


is virtually
horizontal
empiricalcurve
and
closely approximates the line
.50. We
could proceed
r
P(w)

as

not.

was

Lr

come

for the first series of

en

p(0)]]

"

not

had found in
as
we
surpriseinasmuch
the previoussection that Group I was
short of its theoretical
significantly

to

struct
con-

for these series

tion
additional informa-

any

the data.

According

the

to

[tt

Observed
be .58 and

P(l)](l
values

(6)
")2''('"-i).

of P(l)

theory,it should
those

turn

.59 for Series Ia and

out

to

IIIa,

already
values

at

be

from

curves
our

possible

to

disposal. The

in the second

pute
com-

information

p(0)

series should

be

W.

TABLE
Predicted

THE

Ai

ESTES

AND

J.

Observed

and
OF

K.

Mean

Second

be

will

Response

expected

except insofar

Frequencies

p(0), so

the

in

349

STRAUGHAN

H.

Series

as

except

.018

we
d, respectively,

theoretical

III, respectively.The
difference between

lies in the

number

"

80

160

cedural
pro-

lis

to

the

this

1"''
2400

tical
statis-

utilized

in

variable

80

160

have

p{0),it, and
computed

Similarly,the

for Series III

should

Series B

"

80

curve

160

have

fitting,the

been
spondence
corre-

the theoretical and

between

240

data

240

"III

80

160

240

TRIALS
Fig. 2.

Empiricaland theoretical cumulative

and

also to

in the

forcements;
preceding rein-

of

according
model, however,

only
Ia and

I should

IIIb, and we have used


this value, .08,together with .30 for ir
and
.85 for p(0) to compute
the predicted
for IIIb shown
in Fig. 1.
curve
dom
Considering that no degrees of free-

apply

and

of

Fig. 1.

estimated

d value

in

the

error

for Series II b,

curve

this is plottedin

change

Using .50,.30,

the values

as

to

effect

no

for Group

applicableto lis-

be
and

of the first
the theoretical asymptotes
and
for
.85
.50
series,or
Groups II

have

for sampling

estimated

6 value

to

it leads

response

curves

for individual Ss of Group I

350

empiricalcurves
The

for

reason

does

seem

bad.

of the irregularities

some

brought

will be
section.

IN

not

in the

out

statistical

PSYCHOLOGY

MATHEMATICAL

READINGS

increase

no

Slopes

of

similar

and

in resistance

the

two

change.

to

curves

are

very

the

totals do not
response
differ significantly.This
result is in

next

of

test

to

one

line with predictions from


be
the statistical
correspondence can
retical
model, but a little surprising,
by calculatingfor each theototal
curve
a
predicted mean
perhaps,from the viewpoint of Thornin the second
ory
thereinforcement
of Ai responses
series, dikian or Hullian
since partial reinforcement
has
of Equation 5, and
paring
comby means
held
increase
these values with the observed
to
generally (6) been

aspect
obtained

of the

This has been

totals.

mean

the

is

comparison

The

low.
satisfactorily
In order to give an
which

to

the

to

form

Fig.

for the
the

of

data

of

the

then

to

effect, some
curves
being too irregularfor curveThe
theoretical
fitting purposes.
2
in
Fig. represent Equation 7
curves
with 6 values obtained
by a method of

approximation.

of the

are

curves

require other values for this parameter,


viz.,.075,.45,.24,and .18,respectively.
3 and
the

from

4 deviate

theoretical

siderably
con-

form.

that the empirIn general,it appears


ical
individual Ss can
for most
curves

quite satisfactorilyby
the theoretical function, and this fact
basis for inferringthat
gives us some
in this situation mean
learning curves
be

of Ss reflect the

learning uncomplicated by

individual
any

IT

value

effect of 120 reinforcements


of

comparing
response

We

.50 may
curve

be

forms

find that

predict the

(b) that

learning
the

rate

response

of reinforcement
the

in

between

initial

the

bility
proba-

probabilityand

obtaining

ing
dur-

series.

Sequence effects.The mean


curves
studied
in the preceding section may
reflect adequately all of the learning
not
that went
on
during the experiment.
of
The
in some
irregularities
the mean
of Fig. 1 might be
curves
for if there is a significant
accounted
to
tendency for Ss' response sequences
follow the vagaries of the sequences
of
El's and

E2's. To

we

have

check

proportionsof Ai

mean

trial block
In

Ss,

there

blocks and

Ai

to

per

block.

IIb-

in

lead

blocks

they werd

in which

into

no

12

cessive
suc-

were

these

of

trial

classified according
of Ei

Then

120 trials

Since there
576

the number

vs.

10-

in Series B.

divided

were

3 the

responses

preparingthis graph, the


were

bility
possi-

Fig.

occurrences

for all groups

of Series B
48

this

on

plotted in

frequencies of Ei

by

the reinforcements

of

course

the difference

upon

evaluated

totals for Series Ia and

the
parameters
series of learning trials and

one

and

mean

ate
evalu-

from

blocks of 10.
at

seem

stances
circum-

some

learning curve
depends,
approaches its asymptote
an
as
ner,
yet incompletely specifiedman-

artifacts of averaging.

gross

The

of

trend

would

possibleto

series ; and
a
new
the
which
mean

described

for groups

of

study

our

"

fitted quite well by this function with


.30 as the asymptote
T
parameter.
Four
Numbers
2,
11, IS, 16,
curves,

Curves

in
at

smoothing

this

in

theoretical

function,

noncumulative

Ten

least it is

at

vidual
indi-

from

learning curves
{a) that under

mean

be

Ss

curves
response
I.
The
cumulative

chosen

conclusions

extent

for

Group

was

The

seem

extinction

to

situation.

the

of individual

theoretical

cumulative
all Ss of

3.

between

idea of the

plotted in

have

we

in Table

resistance

to

the behavior

conforms

and

theoretical values

and

observed

given

differences

for

values

done

for

occurrences

the

set

of all

Ei's occurred, the

352

READINGS

independent of tt, as requiredby the


values
all
are
theory, the numerical
the 6 estimates

larger than
from

of reinforcement;and
that response
should change in accordance with
probabilities
exponentialfunctions,
learningrates (as measured
by slopeparameters)being independentof
and probability
both initial condition
forcement.
of rein-

bility

obtained
The

curves.

response

mean

PSYCHOLOGY

MATHEMATICAL

IN

most

of this
straightforward interpretation

disparitywould
short

that,owing to the
interval, successive

intertrial

trials

are

independent in the

not

the

required by

Nonindependence
as

theoretical
would

the present

First,stimulus

at

least

learning that

occurred

affect behavior
than

on

cessive
suc-

the

overlap,and

trial

on

one

the

next

on

to

end

statisticalcriterionfor

asymptote
of the second

both

model.

drawn

samples

trials would

would

have

sense

in so far
consequences
isconcerned.
experiment

immediate

two

The

be

first and

series,
Group

II

met

was

series

second
was

retical
approach to theoby Group I by the
and by Group III in

series. In

the second

short of theoretical asymptote

but

reached the same


response probability
Group I duringthe firstseries.
Learning rates were
virtuallyidentical for
and Group II,second series,
Group I,firstseries,
that resistance of response
indicating
probability
forcement
reinaltered by 50% random
to change is not
in this situation. Learning rates differed
within both
significantly
among
groups
series. In general,learningrate
was
directly
as

had

random

related to difference between


initial response
sampling
greater
and
of
reinforcement
probability
probability
would allow for,thus increasingPAiEj,
It was
that this relationship
a series.
during
suggested
and decreasingPa^e^- Second, the reinforcing
depend upon temporal massing of
may
stimulus of one
trial,Ei or
but individual
trials. Not only group
means,
plex
E2, would be part of the stimulus comcould be described satisfactorily
learningcurves
effective at the beginning of the
by theoretical functions.
extent

is
interpretation
widely spaced trials

correct, then more


should result in better
the alternative

of

tendency was
as

whole.

for Ss

to

respond

On

creased
(Ei and E2 occurrences)inas
a function of trials.
significantly

nonreinforcements

of d

estimates

observed

tivity
the contrary, sensieffectsof individual reinforcements and

series
to

tween
be-

agreement

learning rate

mean

to

of the

also in reduction

and

No

If this

trial.

next

dependence
ability
probupon

References

of reinforcement.
1.

Summary

2.

behavior, and

Learning rates, asymptotic


in
of response
sequential
properties
situation

a verbal conditioning
studied in relation to

were

from statistical learningtheory.


predictions
in
run
Forty-eightcollegestudents were
individualized modification

of the "verbal

3.

an

tioning"
condi-

experiment originatedby Humphreys


4.
(4). Each trial consisted in presentationof a
followed
left-hand
a
or
right-hand
by
signal
"reinforcing"
light;S operated an appropriate
key to indicate his predictionas to which light
would
the

appear

on

each trial. For each

selected randomly, was


lights,

El, the

other

as

Eo.

On

one

of

the first series of 120

probability
.30,.50,and
On
.85 for Groups I,II,and III,respectively.
the second
120 trials,
Ei occurred with probability

El occurred
trials.

5.

designatedas

with

6.

.30 for all groups.

Theoretical

that mean
were
bility
probapredictions
predictingEi should tend asymptotically
of Ei, both during
to the actual probability
original
learningand followinga shift in probaof

7.

EsTEs, W. K. Toward a statisticaltheoryof


Rev.,1950,57,94-107.
learning.Psychol.
EsTES, W. K., " Burke, C. J. A theory of
in learning.Psychol.
stimulus variability
Rev.,1953,60, 276-286.
Grant, D. A., Hake, H. W., " Hornseth,
and extinction of a verbal
J.P. Acquisition
conditioned
with differing
response
/. exp.
percentages of reinforcement.
1951,42, 1-5.
Psychol.,
tion
Humphreys, L. G. Acquisitionand extincof verbal expectations
In a situation
chol.,
analogousto conditioning./. exp. Psy1939,25, 294-301.
Jarvik, M. E. Probabilitylearningand a
pation
negativerecency effectin the serialanticiof alternative symbols. /. exp.
1951,41,291-297.
Psychol.,
Jenkins, W. O., " Stanley, J. C. Partial
and
review
reinforcement:
a
critique.
193-234.
1950,47,
Psychol.Bull.,
statistics.New
McNemar, Q. Psychological
York: Wiley, 1949.
(ReceivedJuly 10, 1953)

'

LEARNING

FOR

MODELS

MATHEMATICAL

SOME

OF

INVESTIGATION

AN

CURT

F.

FEY

Universityof Pennsylvania

this

In

determine

to

is

is made
attempt
the results of
whether

study

an

of

merit

The

lies in

model

invariance

parameter
The

learning experiments
be described by stochastic models
can
Mosteller
and
proposed by Bush
Luce
(1959) without
(1955) and
changing the model parameters.
different

two

this

investigate

to

of

question

in

tail.
de-

greater

experimental design

was

that of Galanter
and
improved over
Bush
(1959) by running only one
rather than three trials per day, and
extended
it was
to
parison
provide a combetween
100% reinforcement
reinforcement.
and 75% random

its

ability to describe and predict data


mum
successfullywith the aid of a miniof free parameters.
For any
this can
be done with

Models

The

experiment

one

models.

several

Consequently

stringent test
to predict the
data with

of

such

of the

values

in

can

Galanter

and

studied
Mosteller
showed

in

it is not

clear

the

P/a

and

analysis

invariance

basic

or

was

errors

of

and

attributable

was

mechanism
consequence

in

of

difficulties in

the

model

sampling
estimating

of the

present

gets

rewarded,

Based

on

the author's

PhD

dissertation

This

article

appeared

in /. exp.

the

let

ai,

given
one

on

which

it

probability
the

on

models

of

next

specify the

changes.
probability of going
right-hand side (probabilityof an
the

trial

on

such
with

associated

with

can

n;

;8i and

a2,

parameters
associated
then

be

let

that

reward

and

ai

and

in

pn',

nonnegative
and

02

The

nonreward.
defined

qn

jSa be

the

jSi are
/S2are
models

following

way:

Psychol.,1961, 61,
353

for

response

The

be

pn

"error")

supervisedby R. R. Bush, and read by R. D.


Luce
and
The
data analysis was
J. Beck.
of the
Center
performed at the Computer
tance
University of Pennsylvania with the assisof S. Corn and P. Z. Ingerman.
' Now
General
at
Dynamics/Electronics,
Rochester, New York.

same

any
if

that

the

then

either

turn

can

the

on

outcome

of these

Let
to

only
the

right on
state

ing
mak-

pathprobability
are

response

that

making

maze

the

trial increases.

study

makes

stochastic,

are

models

models

trial

beta

quantity

probabilitiesof

to

or

The

manner

purpose

in

left

trial.

parameters.
The

with
The

animal

the

to

parameter
to

lack

models

these

linear

the

the

independent: the response


on
a
given trial depends
probability and
response
the previous trial.
on
An

the

of

i.e.,they deal
responses.

T-maze

whether

in

probability p;
is applied to
P).

Both

model

uses

alpha model
is applied to the

response
it
model

invariance
Bush

beta

the

the linear transformation


same

may

(Bush

alpha

In

transformations.

lack of parameter
situation, but

apparent

an

invariance

in this paper
model

used
the

as

Mosteller, 1955) and the


of them
(Luce, 1959). Each

predict
experiment.
(1959) previously

Their

(1955).

"

to

parameter
of
model

linear

the

Bush

of the

models

two

designated

mined
deter-

these

used

be

of another

ability

be

once
are

experiment

the outcome

in

that

way

The

more

of parameters

set

parameters

one

parameters

structure

invariant

one

in

fine

is its

model

455-461.

Reprinted with

permission.

354

Response

Model

Alpha
pn+\

CLypn

left tum

pn+\

a^pn

right turn

g"+i

aig"

right turn

5"+i

a^qn

left turn

properties of

Mathematical

model

listed by

were

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

the

Galanter

alpha
Bush

and

male hooded
rats
Subjects. The 5s were
the Long-Evans
Rockland
strain, from
York
York.
Farms, New
City, New
They
weighed about 75 gm. on arrival. Eight rats
used for the preliminary experiment. In
were
the main experiment 63 rats were
used, but
the final A^
50, because 13 died during the
experiment.
"

(1959,

pp.

condition

272-273)

that

the

for

left response

is

special
always

of

is
right-hand response
rewarded, (100:0). For the beta
never
model
the mathematical
properties have
determined
been
(1960, 1961),
by Kanal
Bush, Galanter, and Luce (1959, p. 387),
and
Bush
(1960).
rewarded

and

'

TABLE
Period
Comparison

Statistics

of

Model

Corresponding
aa

.955

FOR

from

the

Values
ALPHA

/32 =

Note.

"

Standard

error

of the mean

was

MODEL

computed

1960.

Fey,

Group

35

Trials

Calculated
Beta

from

of

with

pi

AND

.642, FOR

see

2, 100:0
First

details

For

range

Group

Experimental

pi

.97,^i

Model

approximation.

.858, and
1, ai
.952, AND
=

with

CURT

F.

TABLE

Model

of

Statistics

Values
Alpha

75 : 25 Experimental

of

Calculated
Model

pi

and

The
model
approximation.

Note.

"

parameters

were

from

Corresponding

with

Beta

estimated

Group

.858,and
1, ai
.97,j3i .952, and ^2
^i

with

FOR

range

2, 75:25 Group

Period

Comparison

355

FEY

a2
=

.955

for

.642

Model

the 100.0

Standard

group,

error

was

computed

from

was
a
Apparatus. The T maze
replica mash to balance olfactorycues, and the top
used by Galanter and
Bush
(1959). contained the reward pellet.
This experiment consisted of
It consisted of a straightalley runway
Procedure.
for
a
nd
for
the
main
three
T
maze
a
pretraining
ment.
experiparts: (a) preliminary handling; (")
The T maze
built in such a way
was
straightalley pretraining;and (c) T-maze
of the
that the crossbar and the start arm
learning.
T could be separatedand a goalbox could be
The 5s were
kept in the laboratoryfor 23
and were
hooked to the stem of the T, therebychanging
days at ad lib. food and water
5s were
the maze
into a straightrunway.
The maze
handled daily. Then
deprived of
built of plywood with a removable
wire
food for 18, 21, 21i 21 1, and 22 hr. on Days
was
and
wood
doors.
The
mesh
24, 25, 26, 27, and 28, respectively.
pressed
top
The
inside of the stem and the attachable goalbox
pretrainingstarted on Day 29. For
under
the remainder of the experiment5s were
were
paintedmedium gray, the rightarm was
18 hr. food deprivationat the beginning of
painted lightgray, and the left,dark gray.
fed 4 hr. later
The length of the cross
60 in.,the
each daily run.
arm
was
They were
for a 2-hr. period.Water
able
was
always available
lengthof the stem was 26 in.,and the attachgoalbox was 10 in. The alleyswere 4 in. in the cages.
8 in. high. The
trial per day of
wide and the walls were
The 5s were
given one
10 in. long with
was
pretrainingon the straightalley runway.
startingcompartment
side and
a
a
Pretraininglasted for three days.
guillotinedoor on the maze
tine
During the 30 days of Period 1 of the Thinged door on the outside. Another guillodoor was
maze
at the choice point. The goal
learning,the followingprocedure was
adhered
to:
were
deposited
placed at the end of each arm.
.038-gm. pelletwas
cups
in the rightgoal cup; nothing was
The metal goal cups had double
floors,the
placed in
"

of that

"

bottom

part contained inaccessiblewet

food

the left goalcup.

The

was

placedin

the

356

READINGS

The

lowered.

door

startbox

left in the

was

door

until it

maze

min.

up,

whichever

occurred

the end

of Period

1 5s

were

At

into

random

on

2, and the other

was

following schedule

rewarded

according to

obtained

from

P(L)

table with
number
LLRLLRLLLLLRLRLLLLLR

Period

LLRRRRLLLLL.
35

0.75:

until
by exploration of the parameter
space
Carlo
probabilities (Monte
response
similar
the experito
computations) were
mental
of the 100:0

criteria

at

the

random
L

for finding the alpha model


estimates were
modified

These

data

divided

used

the

first.

were

One
was
group
groups.
the left side during Period

two

always rewarded

those

to

parameters.

was

pellet,until it investigated the goal


side), or until 3
(on the nonrewarded

ate

used

were

The following
group.
total number
of

the

generated by the model had to match


a
mean
plot of trial-by-trial
probabilities
produced by the model
response
had
similar to the corresponding
to appear
plot of the data.

errors

the

data, and

2 lasted for

Results

days.
of parameters.

Estimation
of the

alpha model

followingway
taken

"

The parameters
in the
estimated

other

two

the Period

of trials before

observed

mean

the first

success

total number

their respectiveexpected values.


Initial estimates of the beta model
were

determined

by

methods

of

The

mean

and
errors

parameters
similar

results of the

and

experiment

are

Fig. 1 and 2 and Tables


the
Figure 1 presents

in

summarized

was

parameters
2 data of the

by equating the observed

group

number
the

The

from

estimated

100:0

were

initial probabilitypi

The

be 1.00.

to

were

to

raised.

was

the

cup

PSYCHOLOGY

MATHEMATICAL

point, its

choice

the

passed

As

the

and

startbox

IN

2.

of the 100 : 0
proportions of R response
2
Period
and
the corduring
responding
group
curves
generated by the
models.
Figure 2 depicts the same
for the
75:25
data
during
group
2
Tables
and
2.
1 give
Period

"

EXPERIMENT

"" MODEL

/9 MODEL

1"

25

-T"

30

35

TRAILS
made
Trial by trial proportions of L responses
by 25
2, Group 100:0.
(smooth line) computed with
(filledcircles); generated by alpha model
Carlo analogs
0.955; and generated by 500 beta model Monte
0.858, and at
1.00, ai
pi
and
0.647.
with
0.952,
0.97,
02
0i
(open circles)computed
pi
Fig.

1.

Period
5s

experimental
=

CURT

F.

357

FEY

I.Ql

p."

EXPERIMENT

.o"K MODEL

"."3 MODEL

"

2Y

20

TRAILS
Fig.

2.

Period 2, Group 75:75.

Trial

R R
1"
30

35

(N)

by trial proportionsof

made

responses

by

25

Carlo analogs(open circles)


Monte
puted
comcircles)
(filled
; by 100 alpha model
0.955 ; and by 200 beta model Monte
Carlo analogs
with pi
1.00,ai
0.858,and ai
food reward is in
0.97,/3i 0.952, and /Sj 0.647. (R
(triangles)
computed with pi

experimental5s

"

right

maze

arm,

is

otherwise the leftarm

baited.)

This
experiment indicates that the
comparative results of this experiment
models
under
consideration fitthe Period
A
and correspondingmodel values.
100:0
data, from which their
2,
group
detailed analysisof results is
more
estimated, quite well,
were
parameters
presentedby Fey (1960).

The

Discussion
of

(using parameters estimated from


Period 2, 100:0 group) is less successful.

of learninglies not
the data of any one
aid of parameters

model

mathematical
much

25 group

the fit to the

the
merit

Period 2, 75

but

data

Both

models

show

an

apparent

lack

in

describing of parameter invariance of approximately


experiment with the
equal magnitude.
estimated
from that
Tables 1 and 2 might give the impression
particularexperiment as in its ability
that the alpha model fits the data
to
represent accurately the learning slightlybetter than
beta
the
model.
of a varietyof different experimental This conclusion is hardly warranted
process
if
set of
situations using the same
the magnitudes of the differences and the
parameters.
parameters

In
are

so

other

estimated

words

for

once
one

the

methods

mental
experi-

of estimating the

considered.

The

parameters

alpha model parameters


situation the model
should be
determined analytically
were
; those
able to predict the course
of learning
beta model
the
estimated
of
were
by
in other
which
experiments. Models
Thus
the
Monte
Carlo
procedures.
will handle
of
a
experimental
variety
alpha model parameters were determined
situations with the same
set of parameters
more
are called parameter invariant.
exactlythan those of the beta model.
are

358

READINGS

The
basic

IN

lack of long runs


seems
of the models.
difficulty

little consequence
but it does

in 100:0

well

as

(Derks, 1960). This

generally manifested
but
only
curves,
the
of
data.
analysis

of

The

fact

serious

as

runs.

than

change
will

parameters

"stat

rats"

experimental
lack of long

the

as

in the

ing
learn-

sequential

the

size of the
the

correct

is

runs

mean

75:25

the

slowly

more

is not

^s

that

in

in

mals
ani-

behavior

long

in

not

ing,
learn-

for

choice

lack

PSYCHOLOGY

is of

important

schedules

in human

as

be

to

This

animal

be

to

seem

partial reinforcement

learn

MATHEMATICAL

model

former

ficiency.
deBAITED

"

The

reducing
by

the

about

match

for

those

of

by

ARM

parameters

total

number

of

model

analogs
experimental

the

ARM

UNBAITED

These

will
5s

reduced

the

fit

learning of

the

75:25

the

to

group.

The

slow

could

analogs

100:0

be

handled

in which

manner

modified

lack

the
the

that

group.
decrease

parameters

the

model

75:25

the

100:0

indicate

beta

25%,
by

made

errors

data

when
to

75:25.

the

With

parameter

values

fit of model

to

Galanter

no

would

slightlyduring
phenomenon

the

to

respect
small

increase

it
occurs

5s of Period

first few

began
also

to

noted

the

in

that

the

more

quently
fre-

situations

acquisition

in other

This

experi-

(Gibson "
Walk,
Lach"
Jensen, 1960; Kendler
1958). In the present experiment,
initial dip is hardly noticeable.

1956;
man,

the
A

to decrease

rise.

look

the

baited

the

maze

ARM

(Fig. 3

the

baited

in

3
TRIALS

Fig.

3.

maze

on

the

are

(N)

5s

of the

unbaited
the
up,

that

Galanter
removed

were

than
later

maze;

removal

occurred

the

the

of

arms

unbaited

of the

arm

from

until it

pellet on
min.

The

side

either

for removing

the

those

in

spent

4) indicates

for this is found

reason

in the
cup

and

^s and

rats

unbaited

experiment,
approximately

interval
The

the

Ill*

the

after

ARM

time
in the

(1959) Exp.
quickly from the

from
in

the

and

Bush

more

UNBAITED

at

initially
our
and

BAITEO

1.

in
mental

side tended
the

Trial by trial distribution of time


baited, left (filled
circles)and unbaited, right (open circles)maze
arm
by
4.

in

spent
50

change

Bush

rewarded

before

are

changes from

(1959)
three
of their
experiments
probability of turning to the

trials

Fig.

parameters

data.

and

TRIALS

by specifying

the schedule

of perseverance,

model

time

same

of

the

maze.

in the criteria
5

maze:

is left

the
investigates

food

side, until it eats


baited
side, or until

whichever

investigate the food

first.

occurs

cup

on

the

Trial

by trial distribution of time


spent in baited, left (filledcircles)and unbaited, right (open circles)maze
arm
by 20
.Ss of Galanter
and
Bush
(1959) Exp. III.

The

data

obtained
plotted in Fig. 3 were
originalprotocolsof the experiment
reported by Galanter and Bush (1959).
from

the

FUNCTIONAL

EQUATION

ANALYSIS

OF

TWO

LEARNING

MODELS*

La

KANALf

VEEN

DYNAMICS/eLECTRONICS

GENERAL

NEW

ROCHESTER,

beta

One-absorbing barrier
for learning and

random

model

alpha model)

are

derived

the

walks

linear

YORK

Luce's

arising from

model

commuting-operator

nonlinear

(called the

Functional
statistics are
equations for various
defined
Solutions
models.
branching processes
by the two
to general- functional
equations, satisfied by statistics of the alpha and beta
obtained.
The
methods
models, are
presented have
application to other
from

learning

The

models.

two-response,

number

considered.

of stochastic

models

...

contingent version

path-independent,

two-event,
for

of

learning is given by the equations

\QiPn

with

probability

W2?"n

with

probability (1

p"

where

Qi and

Q2 represent

respectively, the
model
in

discussed

(1) are

transition

probabilities of
by

defined

Bush

and

this paper,

of the

operators

^^^

"beta"

defined

are

^'^-^
terms

of the

*Abstracted
June
for the
many

valuable

A2

trial

on

[8] is obtained

(1

"

n.

the

when

p")
A

are,

linear

operators

from

^iPn

(0

"

QzPn

CX2Pn

(0

"

is called

v"

author

and

helpful discussions

"alpha" model.

the

QJl

Q!2

"

1),

"

1)

when

the

^'^l'^;

'

p"/{l

is indebted

encouragement
and

/3,"0;

Pn) the transition

"

to

doctoral
Robert

received

for partial support


School
of Electrical

R.

from
from

equations for this

sylvania,
dissertation. University of Pennsupervisor
Bush, his dissertation
him

an

9^p,9^1.

NSF

and

to

R.

Duncan

Luce

for

grant.

fFormerly at the Moore


Engineering, University of Pennsylvania,
Philadelphia, Pa. The author is grateful to J. G. Brainerd, S. Gorn, and C. N. Weygandt
of the Moore
School, and N. F. Finkelstein, D. Parkhill and A. A. Wolf of General
Dynamics
for their
This

encouragement.

article

appeared

"

specialization

[13] is obtained

by Luce

proposed

portions of the author's

The

help

QlP"

by the equations:

variable

1960.

model
model

(T-Dp.

1 +

In

and

p"

by the equations

this linear

nonlinear

and

Ai

responses

Mosteller

/2)

In

operators, and

Pn),

"

in Psychometrika,

1962, 27,
360

89-104.

Reprinted

with

permission.

361

In the

model

beta

linear transformations

methods

the

trial to

from
the

into

inevitably enter

propertiesof the model,


propertiesof linear learning models

of stochastic

derivation

generally used

derive

to

than

of choice
probabilities

the

trial. Since

rather

nonlinear

undergo
probabilities

response

apply to the beta model.


applicable to both the alpha
Analytical methods
presented in this paper. The approach used is to consider
defined by the decision rules of the two
models, and
do not

functional

and
the

models

beta

from

it to

functional

formulate

and

statistics of interest. Tatsuoka

equations for various

are

branching process
ler
Hostel-

statistics for

obtain

some
equation approach
somewhat
those presented
differ
from
techniques
alpha
of attack for
here; the approach developed here leads to a unified method
and can
to others.
the alpha and beta models
be extended

[151 used

model.

the

Their

Random

Some
In
and

(4),/?, "

nonreward

Aa is

always
1, 182 "

(81 "
shown

1 and

Walks

^i

1. If neither

these

If response

nature

of the

in this paper.

model

either response

model

is

are

responses

rewarded

response

always

rewarded

/3i "

1. It is

1, 182 "

to

and

barriers for these

OAB

the

model

beta
case

when

probabilityof

random

other
model

beta

two-outcome

for the

Except

diminishes

reward

and

one-absorbing-barrier(OAB),
(TRB) walks. Rigtwo-refiecting-barrier
orous

resultingfrom the two-alternative,


Lamperti and Suppes [12].Only the
considered

ever

lead

cases

and

two-absorbing-barrier(TAB),
proof of the

is

rewarded

Aj is never

1- If both

response

three

Model

with
be identified,respectively,

1, (82 "

"

the Beta

Arising from

1 may

/3, "

of the response.
rewarded

[11] that

in

to

(j8i"
a,-

walks

is given

by

/Sz "

1) is

1, in the

alpha
alpha

1,

response

Ai

; the

model.
one-absorbing-barrier
Functional

Equations for Statistics of the

One-Absorbing Barrier Models


The

OAB

p" which

has

on

the

part of organisms

eventually
processes

and

alpha and beta


all its density at

learn

is obtained

Sternberg [9]on

are

0.

an

asymptotic distribution

(Considering response

learning, this

means

that

A. 1

as

an

of

error

all organisms

information
about
the
errors).Additional
various
statistics. Following the work
of Bush
from
the
statistics
considered
m
odel,
simple single-operator
to

not

which

lead to

models

make

362

READINGS

those

are

mean,

which

describe

weighted
the first

statistics

and

Ai

are

derived

of

variance

oi

of the

A2

an

to the

approach

of responses;

runs

occurrence

an

PSYCHOLOGY

MATHEMATICAL

the rate

mean,

statistics concerning

of

IN

approach; sequential
such as those describing
statistics,
(success)and the last occurrence

other
response

by consideringthe

the

as

of

rate

Functional
(failure).

response

asymptote, such

equations
branching processes

satisfied
shown

by these
in Fig. 1.

V-^A2v
Pz
V-"-/3,V

/9|2)92V
\-Pc

i^zv

V-*.^

i-Pi
/-""

V-^^

/3|02^V

2^

l-Pi
V-"/3 2^V
1-P3

/92^V

V-*'

TRIAL
Z

TRIAL

TRIAL

Figure
The

Beta

la

Model

Lattice

p"-a|P

P"0|P
2

P*a| 02?
l-a.p

1-0, a,p

2"

l-p

,02 p

I-O2P
2

a"p
l-a-^p

h
TRIAL
I

TRIAL

TRIAL

Figure
The

Alpha

Model

lb
Lattice

TRIAL
4

LAVEEN

analysiswhich

the

For

is defined

variables

363

KANAL

follows, a

sequence

a;i

0:2

"

"

"

a;" of random

that

such

fl if response

Ai

occurs

on

trial

\0 if response

occurs

on

trial

n.

random

The
The

trials is

E(Xn)

the

decrease

variables

In

(|Si ^2) and

obtained.

are

equal

to the

of Ai

Fig. 1

from

N)
"t"^{v,

to

responses

A. 1 response.

an

functional

the

E(X)

in A^"trials starting

A2

an

by /3

in the two

in

of Ai responses

result of trial 1 is

responses,

for

responses

Xat-i

number,

models

of the

bounds

X^ of Ai

Xjv-i if the result of trial 1 is


number

expectation

a:" with

\imE(X^)

02) finite upper

trial 2 if the

trials/starting from

obtained

the number

Now

trial 1 will be

equal to 1 +
the expected

^^=1

Aj, and

response

(ai

max

from

X^v

fact, by replacing the parameters

is of interest. In

models

one-absorbing barrier models, both

the

probabilityof

of Ai responses

the total number

x"

variable

E{X)

max

responses

the random

given by

]^^=ip"

expectationsp"

of the random

In terms
in N

of Ai

number

mean

have

variables

(N

"

and

response

1)
be

Letting 0 denote

equations

for "t"are

be
A^
"t",W,v,

pM

r^v ^^^^^'''

(1

1)] +

p^)M^2V, N

^
Y+~v '^'^^'''

1] +

1) +

1)

1)'

and

N)
"f",{p,
When

-^

uv)

(6)
the
=

above

1] +

(1

N
p)"i"a{cc2P,
-

1).

functions

must,

r+^
(1

of course,

"^
'^^^^^''^

r+^

p)"t"M2P)+

'

p.

satisfythe boundary

condition

0.

second

then,

as

E{X_^) the

denote

A''

"

of Ai

of the number

moment

Letting 6

(7)

Y^y '^^(^i^)+

0(0)

are

1) +

"t"a{p) p"t"MiP) +

Both

The

equations become

these

(5)

p["f"M.P,iV

"

responses

functional

equations for

the second

moment

"o
,

9,iv)
=

Y^^ e^(M

]-qr^ UM

y^^ [i +

20^(^1^)]^

364

READINGS

eM

(8)
6{0)

0 is

V"MiV)

(1

P)0.{a2p)+ p[l

condition.

boundary

PSYCHOLOGY

MATHEMATICAL

IN

Finite

2"f"MiP)]-

bounds

upper

and
^^3(2;)

exist for

max
max
{^1 jSa)and a
(ai aa)
da{p)',
replacingthe parameters by jS
if
of X^ is y^Li Pn(l
the variance
Pn) which remains finite as iV *
y^T Pn does. Functional
are
equations for higher moments
easily obtained
=

"

in this

manner.

The

functional
of Ai

number
and

Tatsuoka

been

X2

of A

the random

Yq.n represents
weighting function

"

"

as

"

a^i

the

^2

"

(10)

^"(p)

boundary condition

Number

Fi -\- 1 denote

first time

equal

to

at trial

so

that

if A2

zero

F2 denotes

2, if trial

p,(v) pMM

equal

(1 +

Fi,Ar_i

to

Yi,^m-i))

infinite number

P)^("2P) +

to

of trials,

Mv),

-r

(t"a(p).

(success)occurs

response

trial number

the

an

an

responses,

Fq.atis equal

is

7-4-MM

(1

the

is equal to

the first A2 response


denote

Letting v
equations for

response.

[(1

response

of trials before

F, the functional

1] +

which

on

first trial and

of trials,before

variable

0.

on

For

response.

Fi is the number

1 results in

variables

the

occurs

the number

of the random

(11)

^(0)

and

response

P^PMP)
is

variables

of Ai

number

that

by noting

of trials beforethe firstA

Let
the

^1

MM

T~"
i -f- V

the random

weighted

the

A2

an

if the result of the first trial is an

Mv)

weighted

relabelingthe random

represented by

expectation of
equations are obtained

(9)

by

From

n.

trials with

trial 2 on, the

if the result of the first trial is

which

be

can

"

in N

of Ai responses

trial number

is ^^=2 ^^n
"

number

weighted

being the

for the

the functional

[14]

is somewhat

of derivation

responses

If }p stands

by Tatsuoka

the

variable

of Ai responses

X3

method

of

moment

presented here.

Then

number

second

obtained

previously

the

and

mean

[15].Their

Mosteller

that

weightednumber
Define

have

responses

the

for

equations

and

different from
The

""

"

pM

A2

first A2

(1 +

for

occurs

Fi is

7^2),where

starting
expectation

occurs,

the

are

+ ]r^~v
YJr-y'^(M
'

LA

the

denotes

equations for

second

of the

moment

V-

random

F, the functional

variables

p are

pM

(13)

Y^

[1 +

(14)

Pa(p)
number

'

1 +v

PPaioCi V) + P[l + 2j/(aip)].

last Ai response

at which

P^(M]

2"',(M +

'^"^"^

1 +

TnaZ

365

KANAL

+
VVaiotiV)

v"{p)

(12)
If

VEEN

occurs

Let

fO if no
L"

"^

Ai

1 if the last

[(N +
Then

Ai
on

1)

random

the

Ai

any

the

trial and

and

by definition

of A

occurrence

by Ai

/x

equation
fiaip)
=

for

Ma

on

(1

+
=

[(1

3(1

(1

the

may

be deduced

p){l

(1

p)oi2p+

p)(l

p)(l

(1

response

last

occurs

A2A2A1

of responses

by Ao

the

the second

on

that

random

L, the functional

variables

1. For

infinite number

an

of trials

+ 2]
p)a2p[(Jia{aia2p)

(1

p)a2p

2(1

p){\

a2p){l

"

p)(l

"

"

"

+
ci2p)alp

"

"

"

a2p)a2p

a2p)alp+

"

"

(1

"

p)a2PHa(.ocia2p)

+
"].
a2p)a2piJLc(oiia2p)
"

"

in the

Ai

no

n.

on.

last

expression is just (1

from, the expression for

p)a2p

if

zero

"

which

at

the sequence

+ 3] +
a2p)al'p[iJLMiOclp)

in brackets

term

(1

(1

1] +

pnaiotip)+

But

so

expectation of the
developed from Fig.

p[Ma(aip)+

trial N

on

the first trial followed

the
is

occurs

the third trial. It is evident

on

denote

Li is

development,

and

Letting

trial

on

L^ represents the trial number

trial. In the following

denotes

occurs

response

after trial

or

on,

if the last A^ response

"

variable

occurs,

response

occurs

response

p){l

"

p)Ma("2?")as

Ma(p)- Also

a2p)a2p +

"

"

"

JI

(1

"

"2P)-

366

READINGS

similar

development

(15)

,x,(v)
=

(16)

-^
1 +

MATHEMATICAL

IN

for

n,iM

T-^
M^(M
1 +

'

'^''^'

'

equations

+
p)Ma("2?")

^^\ (1 +

+ (1
PfJLaiotlP)

Ma(?")

in the functional

np{v)results

^''^^^

PSYCHOLOGY

11 (l

"

^\v)

'

^2^)

"

1=0

with

n(0)

0, since for

for the
the

mean

model

alpha

the

For

will be found

Y^^y"(M

at which

second
for 7

Hosteller
to

in

occurs

[15].

consider the expectationsof


random

of the

moment

variables

are

y"m
:i-^

tion
different deriva-

occurs.

the last A, response

and

it is necessary

the

equations

ever

response

in Tatsuoka

expectation of Ll

7 the functional

(17) 7,"

Ai

Ono

1)^,etc. Denoting

(L2 +
by

of the trial number

[2,Mf\^
+

l]
,

and

(18) 7a(p)
with
be

7(0)

v^Mv)

the above

n (1
[2m.(p)

equations

for

higher

"2?")

of Lj

moments

can

easily

manner.

of lengthj, of A^

of runs,

The

vhMv)

0. Functional

generated in

Number

(1

responses

of responses

sequence

Ai^jTix

"

"

"

A.iA.2

"

'

j trials
is termed
of

The

process

of

Kronecker
of

runs

"

total number

Fig.
delta

1 it is

of
seen

function.

"

runs

that

Fig.

1 (a) For
.

5;t,,is the Kronecker

the expression

of

an

delta

to

;, and

and

o-,

i?",,+

denote

equation

the

for aj^

infinite number

function.

the

the

From
.

5",,+2

number

length greater
number

termination

the

length j is then i^i

Ri,j

of

Let /?",,denote

,,

Letting

length j, the functional

lattice of

where

length j.Statistics concerning the

length exactly equal

runs

model

of

responses

of

"),are of interest.
1, 2
equal to j (j
of length j, which
between
trial n
occur

or

process.

of

of Ai responses,

run,

of A^

runs

than
of

of the

branching

5",,+2 is the

where
,

expectation of the
is developed from

number
the beta

of trials

for part
Substituting "tj^{^iv)

of

gives,
i

+ (1
"Jif"{v)Pi"r,p{^,v)
=

+ IIPi[(l
Pi)(Tie{fi2V)
-

P/

i)

P/

i(l

P"+2)].

368

READINGS

(25) Kv, /?! ^2)

IN

MATHEMATICAL

Y^^ mv,

^1

PSYCHOLOGY

^2) +

pqr^ mv,

/3x ^2) +
,

g{v,^,

0.)

where
0 "

0"y"oo,
The

term

13, "

0 "

I,

g{v,^j, ^2) is,in general, different

132 "

for each

1.

statistic considered.

all except the

For

statistics

run

giv,/5i ^2) "

0,

^(0, ^1

0,

(26)

Umg{v,0,
these

For

^2)

,|S2)"

1.

statistics

/(O,^1

(27)
lim

^2)

0,

J(V,;Si 182)

"

Equation (26) does


for the

statistics have

run

functional

The
to have

the

defined

to be

equations

statistics and

run

for

the

boundary conditions

separately.

statistics of the

alpha model

are

seen

general form

y(p,"!

(28)

hold for the

not

az)

pyioiip,
Q!i

(1

aa) +

p)y(a2P,a,

az) + z(p,a,

a^)

where
0 "

0"p"l,
For

the statistics of the

alpha

/29N

and

the

boundary conditions

0 "

1,

a2

"

1.

model

^(0,"i

z(l,ai

"2)
02)

0,

"

0,

for all the statistics considered

y(0,ai

(30)

"

"!

a2)

are

and
lim

The

02)

is finite.

equations for the run statistics of the beta model differ


from the functional equations for the other statistics considered. A
statistics is presented in [11].
of the functional equations for the run
sections which
follow present formal
solutions to (25) and
(28)

functional

in nature
discussion
The

under

y(p,"!

the

boundary

conditions

(27) and

(30) respectively.Theorems

con-

369

KANAL

LAVEEN

cerning existence,uniquenessand other propertiesof the solutions have


Some of these
been proved in [11]by methods similar to those of Bellman [3].
theorems
On

stated here without

are

proof.
beta model

the junctional
equationfor the OAB

Writing f{v,0^

(31)

/Sa)simply as f{v),
(25)takes the form

r-"-^ mv)

mv)

+ -q"

giv),

where

giv)"0,

^(0)

lim^(i;)"l,

0,

and

/(O)

lim

0;

Further,let 0 " ^Si" 1;0 " /32" 1. The cases (/3i 1,jSs" 1) and {0, "
1) can be considered separately.
/Sz
Existence of solution. For any function r(v)define the operator T by
=

1,

(32)

"^
1 +

T-r{v)
=

Theorem

+ g(v).
:p^ri^-.v)

+
r(0,v)

1. -r

1.

r-'giv)

lim

j{v)
=

the limit exists.

when

Theorem

If g{v)is a

2.

increasingfunctionof v, then

monotone

solution

f{v)exists if

Z gi^'v)
1=0

is finite
for 0 "

oo, where 0 "

"

/3

(^Si (82)"

max

1-

g{v)occurringin the beta model first-moment equations


the conditions of Theorem
functions of v which satisfy
increasing

As almost allthe
are

monotone

2, the existence of the


the OAB

"

beta model

From

is a monotone

of most

mean

of the random

variables introduced

is assured.

proof similar

to

that for Theorem

2 it follows that when

increasingfunction of v,

j^giM,
i:Sf(/3")
"f(v)
"

(33)

i=0

,=0

Bwheni

0,,

for

max

(iSi^2)
,

and

0^.

min

(0, ^2),

g{v)

370

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

Conti7iuity. If | g{v) \ is bounded

"

in 0

"

the

oo,

solution

j{v)is

continuous.
If

Monotonicity.
1^2, then

/5i "

j{v)is

the

solution

oo

OAB

Existence.

alpha models

(1

p-y(a,p)+

+ z{p)
p)y{ot2p)

of existence,uniqueness and

is similar to that

(31).Some

for
For

any

(1

pQic^^v)+

other

propertiesof the

propertiesof y{p) are stated


Q{p), define the operator

function

miv)

(35)

if

equation

yip)

development

of v, and

the functional

(34)

and

increasing function
increasingfunction of v.
/(y)is unique in 0 " y "
monotone

junctionalequationfor the

For

the

monotone

The

Uniqueness.
On

g{v) is

solution

without

proof.

+ z(p)
p)Q(c^2p)

let

A^"^-z(p)
Ip^i

lim

Theorem

c(ai ag)-

3.

y(p)

A^"^-z(p).

lim

n-"co

Theorem

// z{p) is

4.

increasingin

monotone

p, then

00

CO

Z) ^("nP)

yip)

"

Zl zialp)

"

i=0

"=0

where

If

Monotonicity.
If

Solution

zip) is

and

is convex
2(7?)
Functional

of the

solution

model

derived

to

Theorem

(36)

is shown

here. A

min

(ai a2).
,

p, and

increasingin

monotone

(31) is
in

detailed
5.

Along
fiv)

==

ai

The

solution

Fig.
presentationwill

aj

then
,

yip) is

OAB

aj

"

0:2

then
,

be

parameter
for

Model
solutions

space

of the

specialparameter
found in [11].

(1) and (2) of Fig. 2,

E dii^Tm n

convex.

Beta

by generalizingfrom

values.

2. One

sides

"

Equation for the

obtained

equation for special parameter


beta

oi"

increasing in p.

yip) is monotone
Convexity.

The

("! 0:2),

max

a"

^2^lV

f/o(1 + /?2/3"

of the
OAB

values

is

371

KANAL

LAVEEN

"

/3
Figure
Parameter

Proof.
summation

side

Along
over

is

n,

(1),182

Model

Beta

1, and

0, |8i"

The

nonzero.

of OAB

only

the

resultingexpression is

equation, for in this

the functional

from

Space

the

0 term
one

of the

obtained

case

mv)
1 +

g(v),

givmg

^7v

/(/?"
for

0, 1,

substitution.

"

"

which

from

"

(1 +

K^r'v) +

m
the

desired

g{(3:v),

result is obtained

Along

z gm

side

(2),/3i

z
^0

1, and

1, jSa "
02V

f-4 (1 + m

rr

Ed

by

(36) becomes
+

02v)g{02v),

successive

372

the last expressionbeing also the


for the

Kv)

to

/3i

case

KM

Note

exists

case

the functional

equation
equation reduces

the functional

v)g{v).Q.E.D.

"

"

the parameter

be obtained.

resultingfunctional
in the form
of g-difference
be written
equations for which
extensive body of literature [1,2].
of the solutions for various specialparameter values suggests

an

the form
the

1, for which

from

point (1,1) of the parameter space the solutions diverge.


/S*(fc 1,2,
/Sg jSg
"),solutions along arcs of the form

Examination

by

(1 +

obtained

one

that at the

By letting/81
(6) and (7) of
equations may
there

1, 182 "

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

of the

space

can

general solution.

The

The

general solution

(31)is given

to

followingtheorem.

Theorem

6.

m=0

n=0

where

Ao,o(v)

1,

Ao.M=Jlr^r^
A^.M

X) Ao.MA,.o(02v)A"_,,,,_,{^,^lv)(m,n

Proof.

Z
m

E
0

(n=h2,-..),

Substitution

A","iv)9(^Tm

in

rxT,
1

Z
V

T-XT

E A^JMgi^T^r'v)

that

(37)
which

E A^,o{v)g{^7v) g(v)+
=

y^^

E A^_,,o(M9(^:v),

gives
m

Ao.oiv)

(38)

1;

A",o(v)

E Ao,Mgi02v)
n=l

").

E A",sM9{^r'm

m=o

so

""

(31) gives

-]-

n=0

1,2,

YJr,A^-..o(M

-A-,E
1

~\-

"=!

Ao,n-,{M9(m,

yo

"

YfW'v'

g(v)

LAVEEN

which

373

KANAL

gives,

"1

and

i:i: A_,."(^ii;)^(CT2t^)

i:i:^"."(i;)^(/3r^^?;)
7^

(39)

m-l

n=l

-t

"=i

"=i

1 +

which

from

.4^1

^2V

JJ

^rv)

(1 +

"

-i

TT

(1 +

to the OAB

Solution

Alpha

Model

solution

similar
is

Theorem

/3i/3r*-V)
'

to

that

for the

used

beta

Functional

Equation

gives the parameter space of


be derived
for (34) can
[11]in
functional
model
equation. The

Replacing /Sgand /3iby az and ai in Fig.


OAB
alpha model. The general solution

manner

equation

General

nk

132V

yp

difference

"=i

[11]

follows

the

expressionsatisfythe

coefficients in this last

The

^1

given by
7.

yiv)

Z) X) 6","(p)-2(a"2p),

Tn

where

K.oip)

1,

^^V

h^.oip)

p-'ar

"o."(p)

fl(l-"rp)

^^

(m=l,2,

(n=

...),

1,2, """),

7=1
n

"m,"(p)

(m,n
Z) hi,o{a2p)bo,k{p)hrn-i.n-kiaia2p)

1,2, """)"

374

READINGS

Proof.
Details

The

proof is similar
given in [11].

are

PSYCHOLOGY

MATHEMATICAL

IN

to that

used

model

beta

for the

equation.

Discussion

Analytical techniques applicable to


presented. Functional
equations for
models, viz., Luce's nonlinear beta model
been

model

called the

alpha model,

have

class

and

statistics of two
linear

derived

been

have
learning models
learning

of

various

commuting-operator
the branching processes

from

defined by the models.


The
the

results

alpha model,

first and

second

number

last A^

the

the

[15].However,

power
power

series solutions
first and

functional

to the

of A

been

are

obtained

For

new.

for the

equations

the

and

responses

expanding

of

techniques

trial
and

Tatsuoka

by
the

in

functions

fails,as is illustrated by the fact that the

often

(obtainedby

second

had

occurs

model

beta

Luce's

of the total number

moments

series in the variable

for the

propertiesof

series solutions

power

which

at

Hosteller

stochastic

on

the functional

Tatsuoka

[14])to

of the

total number

moments

of Ai

equations

responses

for

the

one-absorbing-barrier(OAB) beta model are not valid for y " 1.


By investigatingtwo general equations, the problem of solving the
individual
functional
models was
simplified.The
equations for the OAB

functional
do not

have

the

in this paper,

could

be

the OAB
for
to

one

of the

models,

beta

Furthermore,

use

beta

model

methods

Some

method

some

beta

model

used

to

has also been

obtain

methods

upper

derivation

derived

bounds

failed for

for

in
a

bound

upper

[11],mainly

few

statistics

of the statistics.

number

of close

which

for statistics of

bounds

lower

presented in ([11],ch. 5). An

for the

remains

and

upper

find

to

been

These

model.

made

was

statistics

close bounds

lower

for the

bounds

to be found.

Empirical tests and comparisons of the beta model with other models
Luce
been presented by Bush, Galanter, and
[6] and Fey [10].The
of statistics such

parameters
Bush

attempt

an

easilycomputed.
alpha model have

illustrate the

have

requireadditional investigation.
complexity of the expressionsobtained for the

their solutions

statistic of the OAB

of the OAB

OAB

same

and

Because
of the OAB

beta model
sequential statistics of the OAB
boundary conditions as the generalequation presented

equations for the

and

and

as

those

the

in this paper

goodness of
[8],Bush, Galanter, and Luce

for measuring

Hosteller

derived

for the

fit has

been

[6]and

estimation

of

discussed

by others

by
(see[5]).

REFERENCES

[1] Adams,

C. R.

On

the linear ordinary g-differenceequation.Ann.

Math., 2nd

ser.,

30,

1929, 195-205.
[2] Adams,

C.

R.

1931, 361-400.

Linear

g-differenceequations. Bull. Anier.

math.

Soc, 2nd

ser.,

37,

DISTRIBUTION

ASYMPTOTIC

THE

FOR

BARRIER

MODEL*

BETA

La

TWO-ABSORBING-

THE

KanalI

VEEN

DYNAMICS/eLECTRONICS

GENERAL

NEW

ROCHESTER,

YORK

beta
For
the two-absorbing-barrier
learning
specialization of Luce's
sity
of the response
model, the asymptotic distribution
probability has all its denthe
of
the
The
functional
for
0 and
1.
amount
at p
equation
p
1 is investigated in this paper.
density at p
=

beta

Luce's

is

case

learning

given by

"'""" '''*

(1)

"...

and

p"

for

/?2 "

where

p"/(l

1.

derived.

are

model

In

this

(TAB)
for the

two-reflecting-barrierbeta
the

For
of p"

has

all its

1 is

"

"

^,

"

density

useful

at

"particle" starting

at

functional

for

equation

(2)

is

for the

1, /Sz "

1 is

model

2.
,

the

in

absorbed

[3]statistics

when

/3i

1,

statistics

distribution

of the

amount

the

-foo,

at

"

[4].

asymptotic

f{v) is

If

Ai and

response

presented. Some

1. The

models.

eventually

0,

two-absorbing-barrier

considered

are

beta

0 and

"

paper

obtained

statistic

model

statistic for these

companion

model

beta

/Sj "

probabilitiesof

p"). In

two-absorbing-barrier

tingent
con-

p"

"

"

paper

arising when

beta

0 "

respectively the

y"

two-event,

two-response,

equations

one-absorbing-barrier (OAB)

the

the

p.

probability
are

p"

"

and
2

transition

P"bability

with

response

[5] for

W2Vn

where

model

the

density

at

that

probability
i.e.,at

1, the

j(v) is

Y^^mv),

YT~,i(M

1,

1,

where
0

"

"

*Abstracted

Pennsylvania,
Bush,
from

his

/3i "

00

June

from

1960.

dissertation

portion
The

author

supervisor,

/32 "

of

the

author's

is indebted
for

the

/(O)

doctoral

to Prof.

valuable

B.

help

lim

0,

f(v)

1.

of

dissertation, University

Epstein
and

and

to Prof.

encouragement

Robert

R.

received

them.

sylvania,
School
of Electrical
the
Moore
University of Pennat
Engineering,
School
for the support
author
is grateful to the Moore
Philadelphia, Pa. The
D. Parkhill and
studies.
also wishes
to thank
extended
He
to him
during his doctoral
of his work.
N. Finkelstein
of General
for their encouragement
Dynamics

fFormerly

This

article

appeared

in Fsychometrika,

1962, 27, 105-109.


376

Reprinted

with

permission.

VEEN

LA

The

of

solution

(2) is
the

monotonicity of
of Bellman
[1].
Solution
For

the

subject of

solution

the

and

(3)

solution

of

log,^i

fix)

The

methods

[4]by

two-absorbing-barriersymmetric

log"V

in

similar

to

those

for the symmetric model

^.
Let

Existence, uniqueness, and

this paper.

shown

are

377

KANAL

Then

model

1.

"

(2) becomes

+ 6) +
Y^rrJix

(3) is given by

beta

Theorem

rxT^-^^^

^)-

1.

1*.

Theorem

1)6]j
|-^[x
Z^exp|-^[a:-(/c
^)6r|

Z
%)

exp

(k -t-

Proof.
one

(3), letting g(x)

From

gets h(x)

-}- h)

h{x

"

Assuming

x.

f(x)

f{x

"

h(x)

Cq

h), h(x)

"

log, g(x),

and
-\-CaX^,

-\-CiX

tuting
substi-

gives

g(x)
where

p(x) is

fix +h)

periodicfunction

Kx)

g{x

-^,
(x
26

p{x) exp

of

period b.

Kx)

h),

6/2)^]

As

f(x +nh)

Y, 9(x +

Then

as

-^

f(x

^
,

Kx)

nb)

"

"

p{x)

kh).

1 and

|-^[x

f:exp

m"j-

(k-

Furthermore

Kx
and

lettingn

"

n6)
"

"

"

p(x)

^E^

exp

|-^

[x +

(k

1)6]

gives

p(x)

(^-i)6]j
E^exp|-^[:r
+

*Prof. B. Epstein pointed out


Theorem
1 presented by Bush
[2].

the

error

in taking limits

in

an

earlier version

of

378

SO

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

that

|:exp|-^[:r
(fc-i)"]j
+

/(.r)

from

which

1 follows.

Theorem

problem mdicates.
Corollary

that

/(O)

of the

the symmetry

as

Q.E.D.
When

1.

f{x)
as

Note

^^

for large negative values of


i.e.,

"^,

-56 (^

p{x) exp

p(.t)is of period h and the

"Z'^']

'

/c

corresponding to

term

x,

0 dominates

in the

numerator.

Corollary

2.

fix)
Corollary

large positivex,

For

[-^(^

4 the

denominator

6 "

When

3.

p(x) exp

by performing

Corollary

for then

sum

fourier series
6 "

When

4.

and

constant

from

zero

to

given by

'

analysis.

4,

by Corollary 3, the denominator

by
the

1 is

of Theorem

J2^

_1
pix) ^\
obtained

W^fj-

the numerator

infinityby

may

be

closelyapproximated

approximated by replacing

integralfrom

an

1 is

of Theorem

"1/2

to

Using
infinity.

the transformation

Vb

[i-'-i]

gives the corollary.


Solution

for the generalTAB

For the general case


in

terms

of

the

|Si"

solution

beta model

1, (82 "
for

the

1, it is convenient

symmetric
1 be

symmetric model given in Theorem


is given by Theorem
solution for the general model

for the

to

model.
denoted
2.

obtain
Let

the solution
the

by R(v).

solution
Then

the

LAVEEN

Theorem

For

2.

/3i "

^\Pl)

1, /Sa "

it

m=-0

379

KANAL

w/iere

5"(/3,)

(1

/32)

(n+l)/2

n
i

Proof.

Define

(1

"v)

the transform

F{s)

dv.
f(v)v-'-'

Jo

Writing (2) in

the form

')
mv)
(l i)/(.
+

and

applying the

transform

If

R{s)

gives

F{s +

F{s) +
is the transform

of

from

The

inverse

transform

of

transform

1)

the

of

that

[4]that

and

1).

denominator

terms

2.

in the

^^'''^rRi.s)
taking
being ^(/3i~*/3"i;),

F(s) gives Theorem

It is noted

in

numerator

+ ^'r''F{s+
filF{s)

R{v), it is shown

which, by expanding
product, one gets

^\Pl)

lm.v),

the

inverse

Q.E.D.

the coefficients in the series of Theorem

2 tend

to

zero

rather rapidly.
REFERENCES

[1] Bellman, R. On
and
memo.

H.

certain

class of functional

Fuvctional

equations.In

T.

E.

in decision

equations occurring
Shapiro (Eds.),
RM-878, Rand
Corp., Santa Monica, Calif.,1952.

Hams,

R.

processes.

Bellman,
Research

380

[2]

READINGS

R.

Bush,

P.

methods

[3]

Kanal,
1962,

[4]

in

Kanal,

Luce,

Manuscript
Revised

social

of

the

analysis

for
First

Calif.:

Stanford,

sciences.

equation

model

beta

Proceedings

for

learning.

Stanford

learning

two

In

Stanford

K.

J.

Symposium
Univ.

models.

Press,

S.

Arrow,
on

matical
mathe-

1960.

Psychometrika,

27,

89-104.

R.

of

Analysis

L.

doctoral

[5]

the

functional

L.

(Eds.),

Suppes

Luce's

of

properties

Some

R.
and

Karlin,

PSYCHOLOGY

MATHEMATICAL

IN

D.

received

manuscript

thesis,

Individual

stochastic

some

Univ.

choice

12/3/60

received

7/23/61

processes

Peimsylvania,
behavior.

New

arising

from

1960.
York:

Wiley,

1959.

learning

model.

published
Un-

SOME

RANDOM

WALKS

LEARNING

MODELS

Samuel

IN

ARISING
I

Karlin

Introduction
The

models
learning

some

makes

organism
there is
that the
n, and

by

Bush

of responses

sequence

Mosteller

and

fixed finite set of alternatives

among

and

n that
probability
They suppose
/?"at moment
response 5 will occur.
the
determined
are
/j^""*"^^
probabilities
bythe/?^, response 5" made aftermoment

the outcome

or

will

which

follows

as

r^ and

outcomes

r" that follows

event

models

simplestform

which

introduced

in
operatorsarising
[2]. They suppose that the

certain transition

further

one-dimensional
in

of
analysis

present paper presentsan

r^,,for each

apply where

in their

occur

exist two

There

made

/ was

models

theory. These

and
A-^^

alternatives

experiment. There

choice

shall examine

response s". We

and

exists

set of

outcome

A^,

Vj

in detail the

can

be described

and

two

Markofi^

possible

matrices

F^j

Let/?representthe

occurs.

of choosing
of choosing alternative A^, and
1
probability
p the probability
by the
A^. Depending on the choice and outcome, the vector {p,1
/?)is transformed
initial

"

"

vector
a new
F^
probabilities
representsthe new
appropriate
probability
is interested
of preference
of A^ and A^, respectively,
by the organism. The psychologist
in knowing the limiting
form
of the probability
choice vector {p,1
p).

into

which

"

mathematical

The

follows

as

two

impulses.If

and
(j){x),

"

behavior

of

change of

ulated
of the simplest
description
process of this type can be formwalk subject
to
the unit interval executes
a random
on
particle

it is located
-^

F^x

the

at

"

F-^x ax
pointx, then x
The
with
oca;
"ji{x).
probability

of

the nature

depends on

-^

The
"j"{x).

[1

dF
"t"{t)]

introduce

an

limiting

is givenby
particle

dF.

Jo

Jo
We

probability

actual

/'(:r-l + a)/a

rxjo

{TF) {x)

the

with

the
operator representing

transition

of
the position
describing

the distribution

additional

continuous

operator, actingon

functions, and

givenby
U7T{t)

[1

+ "t"{t)TT{\a
"p{t)]TT{at)
-

at).

conjugateto U; hence knowing the behavior of U one obtains


This
much
information
about
T.
considerably.The
interplayshall be exploited
it
does
pactness
is
continuous
nor
not weakly completely
possess any kind of comoperator T
theorems
of the classical ergodic
apply to this type [3].
property; thus none
the
of
on
The
r"F
behavior
assumptionsmade
depends very sensitively
limiting

It turns

out

that

T is

about the operators Fj and


This article is from

the

(t"{x).
probabilities

J. Math., 1953, 3, 725-756.


Pacific

381

with permission.
Reprinted

382

READINGS

Section
to

IN

the

1 treats

where

case

absorbing states, and

be

MATHEMATICAL

thus

PSYCHOLOGY

(f)(x)

shown

the

knowledge of

the

of U^tt

convergence
and
out

H.

is

"

of the

any continuous

In this

between
of r"Fare

the convergence

the

different
entirely

proofs

are

obtain much

to

"

"

Additional

tt.

remark
finally

R.

are

needed

Bellman, T. Harris,

independently.
They did not point
The methods
theyused to establish

case

Tand

arguments

that

U.

paper in " 1 overlapswith theirs in some


results subsume
15; our
theirs,and their

from

and

Section

ours.

2 considers

the

case

is
"f"{x)

where

and
increasing

monotone

\"/"{x)
-cp{y)\
This

only
By examining
additional
knowledge.

concentrates

Our
probabilistic.

notably 6, 8, 9, 12,

theorems,

at these

continuouslydiflFerentiable then
1. It is worth
emphasizingthat
distributions
does not imply the uniform

we

operators

0 and

the initial distribution.

function

connection,

boundaries

times

Shapiro[1] have analyzed onlythis

N.

the connection

of the

each

convergence

for

for this conclusion.

if

that

For
example, we
(L'^"77)''''
converges uniformlyfor

the

causes

hmitingdistribution

dependson
points.However, the concentration
t/in detail,we
have been able
the corresponding
have

This

x.

the

leads

""

1.

"

the

situation,where the limiting


ergodicphenomonon, or steady-state
independentof the startingdistributions.
examine
the situation "^(.t) I
This corresponds
In " 3, we
to completely
x.
the ergodic
boundaries, and of course
phenomenon holds. Other interesting
reflecting
the
also
consider
in " 4 the case
where
We
of
are
(f)(x)
developed.
properties
operators
linear
and
monotonic
Section
further
where
5 introduces a
is
decreasing.
possibility
allow the particle
still with certain probability.
tically
This type has been statisto stand
we
examined
Flood
the
M.
M.
In
6
we
[5].
"
by
investigate generalergodictype
where
both abstract analysis
linear. The arguments here combine
^(x) is not necessarily
it is worth
and probabilistic
r
ecurrent
event
Furthermore,
theory.
reasoninginvolving
where
the
the
in
6
without
case
emphasizing, proofsgiven " apply
any modifications to
allow
of impulsesactingon
the particle.
In a future paper we
we
any finite number
shall present the extension
where
of this model
to the circumstance
changes in time
infinite
motion
of the particle
has a continuous
or
occur
continuouslyand the possible
to

distributions

are

"

range of values.

discrete

The

last section

studies

of the

some

of
properties

the

in the

distribution
limiting

in all circumstances
that the limiting
distribution is either
ergodictypes. It is shown
the value of
and
the
actual
form
+ a.
on
or
absolutelycontinuous,
singular
depends
models
where
of the analysis
carries over
Most
more
to higherdimensional
a

alternatives

allowed.

are

In

subsequent paper

that this
We
note
generalizations.
finally
and
it
is
analysis
probability; hoped that
of this type.
investigations
It has

[8],and

been

brought

[9] relate closelyto

to

my

the

represents a combination

paper

attention
content

present this theory with

shall

we

methods

the

by

will be

used

useful for future

of

the referee that the material


Their

of this paper.

other

of abstract

techniques

seem

[6],[7],
to

be

different.
1

A
.

law:
X,

and

walk
a random
particle
undergoes
is
after
unit
then
at
particle
x,

If the
x

cumulative

-*

with

1
probability
the
distribution describing
ax

"

x,

where

location

on

the unit interval

time
0

of

"
x

a,

a.

-\-{\

"

at the

to
subject

If

1
.

ing
the follow-

(x)xwith probability

"

F{x) representsthe

beginningof

the time

interval

384

READINGS

this

operator

U.

ambiguityarises

To

be

we

shall

IN

MATHEMATICAL

complete,we should denote


Let
drop the subscripts.
W^rit)

ClearlyW~^

We

W.

PSYCHOLOGY

observe

now

the

the

where

no

isometry

t).

the

denote

77(1

but
operator by t/"j^,

identity

^l-cc.l-. ^^a..^-

(3)

The
of

(1
mapping {a, a)
mapping the triangleof
-*

other

located
triangle

restrict

attention

our

for the other


and

unit

to the case

where

"

"

above

otherwise, we

shall

that

The

two

theorems,

Theorem

2.

77?^

operator

Theorem

3.

The

operator

assume

which

a.

"

0,

the effect

ct=0

into the

"

"

enables

Corresponding theorems
easilyby virtue

deduced

are

From

in

on

now

us

to

valid

of (3)

" 2, unless explicitly

"0.

"

for

state

we

0.

"

cr

"

of this section.

at the end

byl

itself has

isomorphism property

"

"

into

parameter space

This

square.

circumstances, where

next

of the

a)

"

the unit square bounded

in the

will be summarized

stated

a, 1

"

from

immediate

are
completeness,

(2).

if ni{t)"
particular,

Theorem

4.

Proof.

If tt,

([/,,)(") (1
=

all t, then

7r""" "

...

0 and

1.

functions.

for
ir^it),

77',

at

tinuous
positiveconpositive; that is, it transforms

is

into positive
continuous
functions
In

the values

U preserves

0, then

t/77j"

U-n^.

Utt, (Urr)',
.

" 0.
(C/77)""'

simple calculation yields


0(T"77"")(aO
+ ti\

nil

a)"7T"")(a
+ (1

a)f)

a)"-i77""-i'(a
+ (1

a)/)

na''-^J''-^\ot).(4)

Since
"

at

conclude

we

assumption that \
that (UttY'^^
"

it follows
for 0

(y.

is monotonic
7j-*""^'(0

since

77("-i)(a
+ (1
The

"

a"

0.

The

(l

a.)t,

that
increasing

a)0

" 0.
77"''-l"(0r)

"

impliesthat

(1

conclusion

same

0^-1.

a)"-i "

and

argument

" 0,
77-"'*'(0
to
(UnY^^
apply

As

"/""-!.

functions
into functions
U transforms
convex
particular,
positivemonotonic
of the same
the existence
of
kind.
4 we
assumed
Although in the proof of Theorem
the argument can
be carried throughroutinelyat the expense of elegance,
derivatives,
by use of the generaldefinitions of convexityand monotonicity.
In

Theorem
0

"i

5.

and hence

"n

Proof.
true

for /

/ "

"

"

If c " tt^HO "


(WTrY^Xt) " Ki.

for

The
0.
1

"

"i

"

n,

proof is by induction.
By Theorem
have
established
the result
Suppose we
Equation (4) yields

2, the theorem
for

(t/77)""'(l)77"""(1) Ci(a)77""-l)(l)
+ [(1
C2(a)77("-l)(ff)
-

" Ki for
("/^7r)'^"(l)

then

the

is

trivially

/th derivative

")"

with

1]77"")(1),
(5)

where

Ci(a)and CgCc)are

385

KARLIN

SAMUEL

depending only on

constants

and

c;

and
respectively,

on

n.

If

7r"""(l)
" M(a,
where

is

then
suflficiently
large,

constant

c),

a,

(5) yields

"
(t/77)""'(l)
77""'(1).

Since

and
c-^{a)

c^{a)do

depend

not

k, and

on

by

\{U^7tY-\x)\"

uniformlyin

and

find

we

x,

in

generalthat

the induction

hypotheses

M
becomes
(t/''77)""'(l)

when

largerthan

a, c), then

M(a,

" (f/^-77)"")(l).
((7^-l77)""'(l)
+

the
Consequently,

for k
(t/*^77)""*(l)

iterates

M(a,
This

impliesthe
trivially
The

conclusion

of the

proof
completeness.

it for

Theorem

77(0)=0a"^77(l)

Ci(a)M

exists at

77(g
we

deduce

77(0)

impliesthat
Theorem

uniformlyas

Proof.

and
^r-^^
=

ti^

Tig denote

of Urr

maximum

solutions

two

then

Tig;

"

present

for which

with

the prescribed

0.

77-q(0) 77q(1)
"

"

Let

?q

Since

+ to7r(cc
+ (1
t^jniatf,)

a)/o),

find

we
point. Iterating,

of 7r(r).A

value

77j^

7.

(1

0?o is ^^^o

0 is the maximum

which

that

solution

continuous

We

1.

(By contradiction.) Let


Put
boundary conditions.
ttq
point where tt-qachieves its maximum.

Bellman.

R.

to
originally

one

Proof.

be

5.

is due

most

by

c^{a)M.

of Theorem

theorem

next

There

6.

c)

a,

bounded

k^ are

"

similar

shows

argument

by continuitythat
that 0

"

min

7r(r),

n^.

"

For

any

function Tr{t)
"

t^

with

"

oo

"

I, U'"{t^) converges

co.

Clearlyt

"

"

p{t),where
'0,

for 0

"

"

/q;

pit)
for tn "t
1
and

?o is close

0 and

to

1 with

fixed.

find that

fixed, we

are

Since

t "

and
of

lim
U^t

theorem

hand,

VH
at

are

Bit)for every t. Since


uniformly bounded,

the convergence
if t^ is close

to

of U"t
1 then

to

Ut

is

by

convex

Theorem

4, and

the

values

at

Hence

Ut.

jjnf

"\;

^n

"

ijn^l

6(0 is
we

f "

convex,

conclude

6(t)is uniform.

0,
and
that

by

6{t) is continuous.

Obviously, Ud

(C^)'(l)"/''(!)(see the

5 the derivatives

Theorem

proof

of

6.

Theorem

By
On

Dini's

the other

5). Since

386

READINGS

Theorem

Up

"

p,

hence

U^f

"

"

U^p,

PSYCHOLOGY

convexityof Up,
therefore
U'^+'^p;

"

U^p
point,and therefore

fixed

of U"t

the

guarantees

and

MATHEMATICAL

IN

that

lim

the

lim

Theorem

by

deduce

we

and

U'^p

(f"(t).
Again,

infer that

we

U"t^

is
"f"{t)

(f)(t) d(t). On
the

that

tinuous
con-

account

4"{t)with

0 is 0, it follows

slope at

being

convergence

uniform.
denote

We

unique fixed point of

this

whenever
or
by "f}^
by 4"it)
J^t),

no

biguity
am-

arises.

for any

77?^ iterates

8.

Theorem

The

functions

constant

fixed

are

uniformlyfor any

Theorem

7, U^q converges
by the functions (1, T). The

i|"/"||=1,

as

when

appliedto
The

uniformly

converges

functiontt).

continuous
Proof.

{that is,U^n
strongly

U'^ converge

by

actual

well-known

limit is
lim

easilyseen

U"qit)

function

to

of

of continuous

Banach,

spanned

functions.

over,
More-

strongly

U'^q converges

q(t).
be givenby

qilU.Jt)

Consequently by

U^.

q(t)in the linear space

in the space

theorem

function

continuous

any

is dense

set L

points of

q{0)[1

(6)

cf"^Jt)].

n-*oo

This

is

immediate

an

two

dimensional

two

functions

of the

consequence

fact that

space spanned by the function


^o which

q^ and

0 and

at

agree

points of

the fixed

1 and
1 have

^"
the

consist

of the

Equation (6) shows

^.

limit.

same

This

that

enables

us

to show:

Theorem

9.

IfqiO

bounded

is any

functioncontinuous

1, then

and

at

U^q

strongly.

converges

Let

Proof.

derivatives

q(t),in

0 and

at

1.

addition

Then

being continuous

to

clearlythere

exist

and

at

functions

continuous

two

finite

1, possess

h^it)and

/?2(0with
Ihit)"q{t)
where

h^iO)

argument
then
first

we

find

can

of the

part

theorem

/rgCO)and

"

of Theorem

"

/ "

any

proof

Theorem
0

for

follows

now

h^il)
a

with

by

\q{t) qjit)\"
"

standard

from

result

this

continuous

q{t)is

only
propertiesassumed

As

e.

the

about

0 and

1,

q{t)in the
of the

conclusion

1, the

||f/"||

using the

at

argument.

for

"

i "

Wrr^'Kt)]

then

m,

"

for

Ci

m.

The

proof is by induction.
and the constant
functions
preserves positivity,
Proof.

established

result for

the

f/TT-*"'* (1
--

\.

"

t)a'''7T'''^\at)
+ t{\

+
This

now

the
q^{t)satisfying

// W'\t)\"Ci

10.

conclude

h^{\). We

equation(6). If

7 and

h^it),

"

m(l

We

note

For
are

0, the

result is trivial since

fixed pointsof U.

Suppose we

that

a)('"'77"")[a
+ (1

a)'"-i7T('"-i'[a
+ (1

a)r]
a)r]

mo'''-'^TT^"'-'^\at).

easilyyieldsthat
max

"
"*'(/)!
11/77"

max

have

U^'^Kt)]+

Cmax

^''"-I'COI
,

387

KARLIN

SAMUEL

where
I

[(1

max

t{\

1)0"^ +

a)"]

1.

"

Therefore,
" Amax|(t/'=-i77)('")(0l
|(t/*77)("'(0l
+ Cmax

max

KC/^-V^-HOI
t

Amax|(C/"'^-i"77)"""'(0l
+ K

"

by

this
hypothesis.Iterating

induction

our

inequahty yields

last

k-l

"
|(t/'^77)""'(0l
^

max

establishes

This

^'K

"
|77"""(0I

A^max

i=0

M.

the theorem.
C"

Ifqit)belongsto

U.

Theorem

{n

then
derivatives),

continuous

[U'''q{t)T^

lim
m"^co

converges

uniformly
for 0
We

Proof.
On

the

prove

"

I.

"

only for

theorem

continuityof U"q^^K Thus we can


Ijm^a)are also uniformly bounded.

1, for the other

select

cases

are

similar.

{Lf^qf^^
impliesthe equi-

of

boundedness

10, the uniform

of Theorem

account

"

subsequence converginguniformly since

Let

T(0

t/"'^^*!'.

lim

i"*co

lim

Since
As

V^'q converges
d'(t)is independent of

uniformly to a unique limit 6{t),we


the subsequence chosen, the conclusion

obtain

^(t).

6'(t)

of the theorem

easily

follows.

Let

Proof.

^(0)

0, /7(1)
=

The

12.

Theorem

is analytic
for 0
fixedpoint (fi^,^

p(t)denote

1.

By

virtue

At
2

this

through

Theorem

12 for the

case

deduce

we

0 and

that

" 0.
("/"/?)"'"'
"l"i%
=

and

hence, by

desirable to summarize

pointit seems

Theorem

1 1 and

of Theorem

monotonic
is absolutely
"f)^".

Therefore

" 0.
"f"'Jl^

I with

"

"
diflferentiable with p^^\t)
infinitely

function

lim

"

where

"x

the

theorem,

well-known

analogous results
We

"

I.

0, \,2,

the

enumerate

is analytic.

of Theorems

ing
correspond-

theorems.
Theorem

(-iy-\U7ryiKt)
In

"
If (-ly-'^rr^^Kt)

for

same

Theorem

n,

and

0, then

7T(t)"

"0.
functions

particular,
positiveincreasingconcave

of the

then 0 "

4'.

are

transformed

into

tions
func-

kind.
5'.

If

"

Tr(t)"

" Ki, and


(-l)^-i(C/'"77)"^"(0)

and

hence

" 0 for 1
(-iy-'^TT^^\t)
" Ktfor 1 " i " n.
TT^'Kt)\

"

"

i "

n,

388

READINGS

with

7 holds

Theorem

the

lie in the

providedonly they

a,

MATHEMATICAL

unchanged and

6 remains

Theorem
and

IN

PSYCHOLOGY

is valid
unit

open

modification

independentof

the conditions

on

interval.

of

the

proof where

p{t)is replacedby

function

concave

'

\,

for 1 "

"

?o

for 0

"

?o

pit)-

y.

-t,

"

^0
the

and

functions

replacedby 1
function, a familyof functions
of Theorem
to infer the validity
us

constant

enables

changes

are

in their statements

modifications
appropriate

solution

^5

which

for this situation,where

analytic.In

the

remainder

Theorem

The

13.

8.

cr

of

also constitute,with

the

in

9, 10, and

above

11, with

for Theorem

7.

completelymonotonic

the theorems

suitable

by simple
The
unique
and

established

are

the

C[0, 1]. This

established

reader, are

1, is

"

is dense

span

Theorems

for

of this section

the value

as to
specification

linear

that indicated

to

These

ty.

"

whose

leave

we

similar

(1

"

hence

without

any

(t.

functions
00

"^m(0

to
geometrically

converge

from

(6) that

U^m
tends

uniformlyto
by Theorem

Since

zero.

conclude

11

A "

Let
.

knQ

the derivative

that for

denote

the last

"

at

0 and

t)] " Xt{\

k
integer

2
t

T,(0

1 of

sufficiently
largethere

"
"
"i"J,t)
"^,,"(0

"

t)]

t/^o^l
with

t)]

0.

It is immediate

Proof.

U^m
m

0 is

"

exists

an

"

We

1 and

"1,

n^iX)such

we

that

t)

for which

UV{\

/(I

kn^

t)] " Ck^

"

m.

obtain

" Cp",
Cp""o+i)fc

1=0

where
=

Theorem

converges

14.

We

first establish

simple calculation

shows

"

m,

we

obtain

the

result

for

specialfunctions

-t)"

U{t') -f"

continued

upon

Ct{l

of
application

conclusion

now

with

1 "

"

oo.

summation

that

t/'[r(l
-

t)] "

U%t')

"/""(r)"

i=m

The

U^[q(t)]

t).

and

-C

oo, then Urn

that

-Ctil
For

1.

is continuous, \q\\)\
"
"
Ifq{t)
ooand\q'(0)\
geometrically.

Proof.
A

?}nno+l) "

U\t{\

t)).

i=m

follows

from

Theorem

13.

The

generalfunction

satisfying
q{t),

SAMUEL

Theorem

hypothesisof

the

and

Pi(0

PaCO

the first

this fact and

observe

We

14,

which

part

at 0 and

agree

of this

from

bounded

be

can

389

KARLIN

above

The

below

and

result

by two polynomials
from
directly

follows

now

proof.

easilythe identity
Ut

U and
Applyingsuccessively

={:^

\)t{\

t).

obtain

adding, we

00

"f",^, lim

C/-?

W"

the

consider

2 K^^

1)

(7)

0-

of calculation.

the dependence of (f)^,^


on
describing

remarks

Some

"-00

is useful for purposes

This

(a +

and

in order.

are

We

:
followingidentity
w

Ula

"

U^a,.{U,,. U,;.'W::r\

Va-,0.'
=

(8)

"

1=0

If

/(O

function

is any

with

bounded

derivatives,then

the

by

mean-value

that

theorem

" 1(1
Kt^a.a f/a',a')/l

[/(^O -/("t'0]+ ?[/("-^ (1

"

-/(a' +(1
C{\o

"

Applyingequation8
by

obtain

we

Theorem

2,

f(t)

to

-aOOll

a'\ +\oL

a'|)r(l

t).

that

^^^
remembering
"f"a',a',

"x)t)

are
preserved
inequalities

obtain

we

\U^^"l"a'
a'

4"a',J C{\a
"

ja

a'\ +

"

a'l) ^

U^td

t)).

i=0

Allowing "

to

go to

easilythat

have

we

oo,

" ^(k
\i"a,a 4"a\a'\

K(rj) is finite,providedthat

where

It is worthwhile
of the unit

square.
Next

^ff,a(^) ^=

observe
0 and

by

(f,(x)
=

at

0.

and

"

(7

I and

0, 0(1)

of the unit square


"

1 is

arbitraryand

only

1, is (f"(x) 1 for 0

"

TT

while when

"

0, 0

"

"

(1

^'"^)"

\, then
00

"^a,a TT
=

a;

calculated
easily

are

00

the

boundary

1, then

L^'X,

have

we

"

1.
and

that

is continuous
(f"

provided
continuous
fixed point "f"

'f"a,l 1

when

1, then for

then

that

that

the

1 then

"

the solutions
a

0(1)

"^(0)=0(0"a?"l)

0(a;)

I.

t] "

"

(a, a) lyingon

for
^

verification

0 and

hence

"

+ x"f)(x).
x)(f"{ax)

"

point with "^(0)

fixed

when
Similarly,
satisfying0(0)

boundaries
If 0

is

and
(f"(Gx),

^"

"'!),

then

I;

"

of

direct

a'cr,o'

a,

nature

1/(1, (1
Therefore, if ^

r] "

"

the

discuss

to

First, we
let

^^'l+

"

On

the

turn

out

other
as

at

two

follows

390

L"

where
0

on

a")
(o-",

the

'Pa ,a (0)

(^^

0.

Also, for any


0

as

loss of

are

We

to

tend

to the

in

"

1, and

"

a;

any

"

0"x"l

we

^(t)uniformly,for

"

with
positive,

are

interval of the form

any

1.

(t"

"

over
equi-continuous
which
be
subsequence
may

^^

select

can

8 "

"

that 1

I, the first derivatives

"

"

I,

a^ "

1 otherwise.
(f"a,a(^)

assume

0"a:;"l

0,

limit

We

boundary.
(ap,0) with

-^

interval

this impliesthe

" 1,
"f"a",a^

(a, a)

generality may
monotonic
and
increasing

interval

Since

"

where

case

the
investigate

now

we

convex,

interior

"

"

As'

1.

get T(l)

1 and

T(0)
similarly

*F at

of
continuity

the

"f)"
^

"!"" convergingto

as

(5 "

we

the

subinterval,and

denoted

and

a.)x. Finallyfor

"

identitymapping.

is uniform

without

uniformlybounded.

are

5"

the

(1

pointwiseto 0 for 0

convergence

0); then
(o-Q,

Therefore

0.

"

a"

gives

studyingthe

to

-*

to

allow

we

as

"^o converges

Moreover,
Let

and

the

to

PSYCHOLOGY

3,

that

show

we

0^

reduces

for definiteness

attention

and

operationL appUed
U

operator

dependenceof
our

the

I and

1 the

MATHEMATICAL

IN

READINGS

uniform

The

0.

of

convergence

guarantees
i^a^ar

zero.

Put

Ur
the

consider

We

take

We

U^^^Q,and

"j"r

"i"a^,a^.

:
followingidentity

C/oT

Uq
Ua^^oL,,

(T

fixed

^,)

(t/,T

t/oT)

|/i| |T
trivially

\; then

"

{"i", t/,^0

^^| "

"

/i

when

/g.

is

sufficiently

large.Also
I/2I 10,
=

t/,T|

\U,4r

Ur'y]

1(1

+ (1
a:[";6,(a,

+
for

But

ic

a^o "

xMria^x)

observe

fixed,we

that

a.^)x) T(a,

"

varies
oi.j)xq

"

(1

"

(1

a,

-T(cT,a;)]

in

"

an

a,)a;)]|.
interval

T
appliesto
convergence
for r large. Thus
S yieldsl/gl" e.
inside 0 " a; " 1
By construction, \I^\"
and
verification
for x
I.
T
for
"
0
direct
infer
the
a^
1,
"
we
t/^T
by
equality
T
T with T(0)
1 and
the fixed pointto the equationC/o^
0, T(l)
However,
"

"

as

a,

0, and

the

same

of "^,

uniform

Oj-x. The

"

continuous
is the

at 0 is

appliesto

the

furthermore

established

the

rj "

a,

"

(a, a)

The
I and

1 and

T(l)

and

hence

-^

0)
(o-p,

with

(1,a^) with

Summarizing,we
yieldto simpleranalysis.

Finally,a

word

values

lie

parameter

on

have

the following
satisfy
continuity
properties:
fixedpoints (ji^^a
0

CTq "
a^

function

4"a",a"
converges

"

a,

a'

"

rj,

"

a)
If((y,

that

deduce

we

the hmit

Thus

is

" ^(^)[k
l"^a,a "^a',a'l

If{a, a)

1.

independent of CTq " 1- ^ similar analysis


of the
(1, a) (a " 0). The continuityproperties

boundaries

two

15.
a.' "

"

a;

that T

followingtheorem

Theorem
"

note

where

case

for the other

solution

1 for 0

subsequence of a"

for every

same

pointwise.We

IfO

^{x)

"

1, then
0, then

then
"

f^'l+ !"

"'!].

"

0 pointwisefor0 "
"f"a,r".(^)
1 pointwise
for 0
'i"a,rj"x)
-^

-^

concerningconvergence
the boundary. When

of
a

V^tt
=

0,

for
o-

"

tt

a;

"

1 and

"
a;

"

continuous

1, then

(f"aj^l)
=

1.

1.

V^tt

when

the

converges

392

READINGS

model

2. In this second
is at
1

a:,.then

(1

-^

the random

a)x with

"

PSYCHOLOGY

MATHEMATICAL

IN

walk

is described

follows:

as

and
"f"ix)
probability

If the

particle
probability

with

ax

-*

where
"i"{x),

"

"ix"\.
\"f,{x) cf"{y)\
-

analogous transition operator

The

(1) becomes

to

/-(ai-aVd-a)

V/ff

(1

the same

In

this

are

section,

considered

operate

the

are

We

"

1 ; the

"

a,

but

operator

similar

T is

conjugateto

^(0

is monotonic

that

"f"{t)

where

case

A +

a)/].

The

(10)

values

spaces

for

and

which

on

Theorem

to

manner

they

obtain:

1, we

the operator U.
This
increasing.

A +

fxt,where

Let

boundary

of great interest.

not

Again, in

1.

assume

where

important
A "

The

further

now

"

+ "^(077[a+ (1
i"{t)]7riat)

handle
in

as

17.

case

then

to

easy

same

[l

take

we

is

Theorem

the

before.

understandingconcerningF applyingas
Un

(9)

Jo

with

4"(t)dF(t),

+
i"{t)){dF{t))

1 ; and

^" "

model

whenever

includes
A +

0.

Theorem

The

18.

operator

and
preserves positivity

positivemonotonic

creasing
in-

functions.

the

Since

the

0(1)

Theorem

converges

UiT

TT

This

complete the
Theorem
Proof.

functions

by

spans

The

Since

well-known

that

assume

if

can

"^(0)

"

treated

"

in

0,

analyze
analogous

we

an

exists and
"^'(0)

0, then

increasingbounded

monotonic

^(0)

or

be

and

is finite.

then
positive,

carried

operators U^n

\\U"\\

dense
theorem

out

subset

1, and

converge

the

space

uniformly
forany
of all monotonic

of the set of all continuous

continuous

function.

continuous
positive
follows

functions, the theorem

of Banach.

tions
distribu-

21.
For any distribution F, the distributions T^F
converge as
TG
G
G
which
which
distribution
is
independent
of F.
for
unique

Theorem
to

Proof.

Theorem

U^tt

constant.

easilyusing the techniquesemployed above.


fixed
(f){t)
easilyyieldsthe fact that the only continuous
in
functions.
similar
the
used
The
is
to
constant
are
proof
proof
fact directly
with the result of Theorem
21 below.
connects
First,
function
of
for
continuous
of
V^tt
77(0.
proof
any
convergence

20.

circumstance

other

now

proof can be
hypothesison
6.

Theorem

we

The

If Tr{t)is

uniformly

The

pointsof

19.
to

The

we

1.

"

Furthermore,

manner.

^(0 implieseither ^(1)

hypothesison

where

case

verification.

Direct

Proof.

1 6.

The
To

weak*convergence

complete the proof

of r"F
we

must

follows

from
directly

establish

that

Theorem

if lim

T'^F

20

and

and

SAMUEL

lim

T^H

K, then

Indeed, let T

K.

393

KARLIN

denote

function.

continuous

any

have

We

that

(T,

K)

lim

91"

(T, T%F

H))

lim

"-C0

M"

(t/'^T,F

H)

]
"(1"^-/'
a{ \dF

\ dH

^0

"-00

(11)
as

F and

/f

distributions.

are

Hence

{
W{t)dF{t)
=

for any

function

continuous
It

fixed distribution.

Theorem

if{G^,a")

T"^,a"

T for

about

1, then

"

cr, a

K.

the

complete

it in

at
F^^^^^ F^^rt
-^

the

distribution

7t(/)denote

any

at

every

denote

We

function of

continuous

of this

nature

later section.

{a, a); by Helly'stheorem

-"

Let

T^,^.

determine

more

say

"

(cr",a")

Let

and

therefore
to

distribution F^r^ is

Fa^ ,0L" converging to

the

The

with

((T,a)

-^

shall

We

22.

Proof.

Fr

T, and

extremely difficult

seems

a,

unique
it by
that

a;

is,

point of continuity of
choose

a subsequence
point. Write Tj. for
continuity
we

every

can

fixed continuous

function.

consider

We

quantity
(tt,F

Since

F^

Now

-"

note

we

TF)

C/

(77,F^)

find for

(77,TF^)

(77,FF^

TF).

F
sufficiently
largethat [(77,

F^)! "

"

e.

that

U"^ ^^

F,)

we
distributions,

as

1(77,
Fr)
Since

(77,F

C/^77
converges

(77,TFr)\

f/

to
strongly

converges

U-n.

uniformly to
\{Ur-n

1(77,
F,F,)

"

(77,FF,)|

as

F^

it follows
verify,

we
distributions,

are

lt/^77

max

|(t/,77 C/77,F,)l.

is trivial to

U^^^,as

Whence,

Utt, Fr)\

"

Utt\

"

"

that

infer that

I
when

is chosen

largeenough. Evidently,with
1(77,
T{Fr

Therefore
Since

obtain

we

is any

77

Theorem

21.

for

function,

Consequently, as
22

is

any

FF)|

infer

we

limit

largewe

\iUn, Fr

F
largethat 1(77,

continuous

of Theorem

F))\

get as

F)\

"

distribution

hence

and

TF

of F"

that

e.

3e, and

"

before

(77,F)

therefore
must

(77,TF).

F^,^ by

be

F^ ^

In

this case

the

clusion
con-

immediate.

now

m
"

3.

The

monotonic

model

considered

The
decreasing.

operator

Urrit)

Note
and

that

does

the ends

we

not

0 and

have

1 the

"

a.

In
generality.

^{x)

=z

x.

"

is
"f"

becomes

tTT{at)+ (1

replaceda by

restrict any

is with

in this section

This

t)7T(\
-

is

only for

this model

there
greater probability

is of

the

moving

at).

convenience
closer
back

the

(12)
in Theorem

particlemoves

28,
to

into the interior. The

394

READINGS

situation

IN

MATHEMATICAL

PSYCHOLOGY

described here is of

boundaries.
completelyreflecting
Again it is easy to show
the constant
function. Therefore,
are
points Un
shall find as in " 2 that the distributions describingthe positionof the particle
we
converge to a limit distribution independentof the initial distribution. We firstproceed
to analyzeconvergence
it is no
In this case
propertiesof W^n.
longer true that U
the
of
class
m
onotonic
functions.
is conserved
positive
preserves
Only positivity
by the
U.
described
in
Theorem
23
here
well.
a
new
as
However,
serves
mapping
quality
this
section in order to avoid trivial changes of proof and different
Throughout
that

the

only continuous

results at times,

suppose that 0

we

Theorem

fixed

23.

If Tr{t)has

"

"

cr

a,

then
derivative,

continuous

"
|(^7r)'(0l

max

|77'(r)|,

max

with

equality
ifand
holding
Proof.

By

direct

onlv

obtain

t)aTT'(l

theorem

\tan'{at)+ (1

a^l

(13)

7r((7t) 77(1
-

(at

(1

a)

"

[ta + (l

max

t)a +

"

"

(a

"

(I

"

a)t]max

"

a)

"

at

"

\Tr'{t)\ max

then

let t^ denote

UXOI-

equalityholds,

at)

at)

Gt

If

a/).

get

we

t)o!.n'(l a

-rriat) Tr{l

a/) +

the aid of the mean-value

" max
|t/77'(/)|

max

is linear.
if-^it)

computation,we

U-rr'it) tav'(ot) + (1
Hence, with

pointwhere
|77'(0I k'(?o)l-

max

It follows

easilyfrom

(13) that
77-(l

TT{ot^
=

"

Q'

(14)

o'^o)!
=

atQ
This

yieldsthat 77(0

between
chord
a

ct/qand

"

is linear
a

cr^o"

Theorem
max^

Theorem

converges
Remark.

and

at^,

or

a)

ato

otherwise
than

greater magnitude
also
these points. Equation(14) shows

the

somewhere

slopeof

the

24.

The

proofis similar

25.

to

// 7t(?)possesses
uniformlyto a constant.
The

reason

so, will be
necessarily

that of Theorem
two

why the two


explainedlater.

"

be linear.

// 7r{t)belongsto C"* [7r(;')


possesses m
in
b
ounded
is uniformly
r (0
n for each
|(t/"77-)''"'(/)|

Proof.

U^TT

"

slope has

in (13) requires
7T{t)to
impliesthat equality

then

"

(1

that at^ and (1


by Tr(t)at
then
maximum
points of 7T'(t).Repeating this argument successively

subtended

a/g) are

for

at^ the

at(^

"

"

l^'(l
\TT'{t)\ |7r'(cT^o)l

max

"

derivatives],

m).

10.
and
derivatives,

continuous

cases

continuous
"

and

are

5^ a,

then

distinguished,

SAMUEL

In view

Proof.

of

L/"v

of functions.

We

Thus
thus

can

by

virtue

of Theorem

derivatives

second

(U^tt)' constitute

equicontinuous

subsequence rij such that U"'tt converges


that
It follows
trivially
uniformly to "l"'(t).
a

23,

KC/^'TrVI

max

and

U^tt

select

and
(U^'tt)' converges
uniformly to (j)(t),
ifrii+i^ tends uniformly to Ucf)and

Moreover,

24, the first and

Theorem

23 and

uniformlybounded.

are

famiUes

of Theorem

395

KARLIN

"

|(C/"'+M'l"

max

|([/"'+i77)'|.

max

(15)

Hence
lim

\{U'''tt)'\ lim

max

i"fCO

("""00

Therefore, by the uniform

|"^'(0I

23

i(t/20)'(O|.

max

secure

and
U4"{t)are linear. However, if a 5^ rr and "f"{t)
yieldsthat 4"{t)
forces
U4" is quadratic. This impossibility
t, then
4"{t)to be
that
Let / be chosen
sufficiently
largeso

with

term

derivatives, we

|((7"A)'(0l

max

InvokingTheorem

i""-oo

of the

Kf/^'+V)'!.

max

convergence

max

contains

Kf/^'+^Tr)'!lim

max

a constant.
identically

|t/"'7r

c\

"

e.

Then

lU'^'^'^TT c\

Repeating

this

c\ +

t\U"'7T(ot)

"

shows

argument

(1

This

establishes

Theorem

26.

linearly
a

we

obtain

dense

case

where

subset

next

1 "

polynomial does

cr

"

0.

Theorem
and

is

Pn-\

of

If P{t)

27.

The

constant

degree " "


it
proof, is enough

proof
c

"

to

then

"

e.

with

ol,

then

two

all continuous
a

establish
in this

note

its

j^

c.

V^-n

continuous

derivatives

functions.

Since

theorem

the uniform

convergence

the

case

uniformly.

converges

well-known

spans

||t/"||
=

1,

of Banach.

fact that
interesting

of V^tt
t/

for the

appliedto

degree. Particularly,

[a"

"a"-i(l

a)]x" + P"_i(a;),

polynomial of degree n

the convergence

Proof.
if P

and

and

25
we

We

increase

not

of

space

theorems

two

"

converges uniformly to

continuous

using Theorem

P"_i(x) denotes

constant

U^tt

C\

all functions

of the

Ux""
where

of

space

the result

In the

that

If 7t(/)is

The

Proof.

a/)

that

IW'^^Tr
for any/7.

t) |C/"'7r(l

\.

"

polynomial, then

is any

U^P

converges

uniformlyto

is geometric.
is

by

induction

U^P

c.

[/%"
=

a"

the

degree of
have

1 that the iterates

verifythat

on

Suppose we
t/''P"-i
converge

converges
-

the

shown

for

any polynomial
uniformly. To complete the

uniformly.Let

rta"-i(l a);

polynomial.Clearly

396

READINGS

then

1 since

"

1 "

IN

0.

"

MATHEMATICAL

obtain

We

f/x"

get, for A:

Repeating,we

PSYCHOLOGY

Ax-" +

P^^iC.r).

"

k-l

This

last

form

is of the

sum

with

"
|fl,.|

00,

and

Hm

b^X^)exists.

*=

function.

constant

Finallywe

speak,is the

to

U^'x^

uniformly.Thus,

converges

any

It is

well-known

theorem

that

lim

c^{x)

00

uniformlywhenever

exists

so

uniformlyto

converges

in the

that

note

of the
regardless

same

polynomial converges

where

case

of the

outcome

The
geometrically.

fixed

a.

be

must

(the rate of learning,

experiment),then

be

proof can

pointwhich

carried

for

C/"P

through by using

induction.
This

fact

yieldsthe

that

that

Proof.

Similar

We

note

V^TT

now

Theorem

unit

that

not

for the circumstance


for
a

1 "

the

U
1
(T

"

can

0,

return

Theorem
r

"

or

conclude
(7=0

when
particular,

77

is

"

0,

"

The

only
boundary

"

case

where

the

1 it is not

of f/" is trivial.

However,

V^r^-n converges

For

when

even

produces
"

for every

of the

and

77,
=

that

to show

hard

function
cr

of

argument
mention

cr

longer true

no

C/^"+^7r converge

We

traverse

dense.

1 and
0

and

continuous

now

29.

to

the

0
hypothesis

"

a,

then

If -nit)belongsto C",

"

(t/*77-)"^'(/)
converges

uniformly
for

m.

Proof.

This follows

easilyfrom

Theorems

f'x/o

TF=\

rix

dF(t)

Jo
This

is

continuous

convergence.

then

0 it is

formly.
uni-

converges

polynomialsis

otherwise.

convergence

of

lack

again a
1 "

the

V^tt

and
U'^'^tt

1 and

polynomial.The

for which

and

case

a,

for every

necessarily
converge
where

when

occurs

-n.

We

"

In

identityoperator

we

function

behavior

of all

periodicphenomenon
A
in this case
as the quantity

square for this model.


does
V^ry^TT

when

occurs

0, then

"

set

in this

geometricallyto

highermoments.

the

important example that

difficult convergence

that

26, since

It is easily
verified that

down

breaks

27

and

continuous

Theorem

to

the

converges.

separatelybut
other

is
If -rrit)

28.

expected positionconverges

similar results valid for

with
limitingexpectedposition
Theorem

the

represents the transition

ix

"

24, 26, and

28.

Let

D/a.

(1

/)dFit).

Jo
law

for the

distribution

the position
of
describing

the

SAMUEL

for
particle

this model.

sections, we
Tand

between

those

to

theorems,
following

in the

employed

using the

preceding
conjugaterelationship

For any distribution F the distributions T'^F


tions
converge as distribudistribution
which
which
F" ^ for
is independent
unique
F^^",
TF^^
of F.

in the

31.

of

sense

it

Again

The

distributions

Theorem

seems

F"^

constitute

tions
family of distribu-

continuous

11.

difficult to

very

determine

explicitinformation

more

any

iv,aThe

4.

model

at least 1 "

and

analogous

30.

Theorem

about

the

U.

Theorem
to

By arguments

establish

can

397

KARLIN

examined

Utt

Of course,

or

before,0

as

elementary in

this

"

The

//.

{Xx

is such

1.

"

cr

a,

that

he

"

bounded

immediate

An
to

-\- i.i,with

X +

fi

"\

"

which

ax).

V^tt
is

(16)

turn

out

to be

very

easilyproven.

then
derivative,

\Tr'{x)\

max

\.

"

/h)tt{\

"

with

he

the form

followingtheorem

"
\{Utt)'{x)\

max

4"{^)

"

Convergence questionsfor

of the

If Tr{x)has

(1

has

operator

iu)tt(ox)+

in view

case

32.

Theorem

"

here

0.

Let

the standard

we

way,

Theorem

of Theorem

consequence
denote

33.

distribution F^^^ which

the

obtain

For
is

transition

operator

{V^tt)'converges geometrically

is that

of distributions

for

this model.

In

distribution

any
a

32

the

distributions

functionofio,

continuous

a), and

T^F

TF"^

converge

to

the

F^,,^.Moreover,

is independentof F.
"F(j
o;
5.
feature

77?/^ section

added

the two

fixed

is devoted

first is that

points0

and

F-^x
the

of
possibility

to the

two

new

towards

the transformations

by

ax

preceding models.
impulses of motions

of the

variations

some

in addition

and

models

where

F^^

"

ax

stands stillwith certain probability.


particle
statistical
particularly
important
learningproblems, and much
this type has been done
on
by M. M. Flood [5]. They are referred to as
investigation
of this type is as
the pure models.
The
mathematical
descriptionof the first model
three
random
A
unit
interval
is
follows:
impulses: (1)
particlex on the
subjectto
with probability-n-^iX x)\ {1) x
\
x
ax
a.
-\- olx with probability-n^x; and
a; with
(3) X
tt^x,where 0 " w^, wg " 1. This
probability(1
rr-^il x) + {\
is similar to model
I where
absorption takes place at the boundaries 0 and 1. The
These

third

to

allow

we

motion

the

in

are

-^

"

-*

"

"

"

"

operator analogousto (2) becomes


Utt

TT^il
"

x)TT(ax)+ [(1

"

tt-^){\ x)
"

-f (1

"

TT2)x]7r(x)
-I- Tr^x-niX
"

-I- ax).

(17)

398

IN

MATHEMATICAL

transition

operator

READINGS

denote

Again, let T
into
particle

setup, and

increasingfunctions.

If '^,

Uv

over

changed

into

"

convexity,and

analogues of
and

methods,

obtain

we

[1

that

computation.
4 does

Theorem

to

noting that

for tt-^

773

not

1 the

carry

have

we

here

condition

on.

easilyextend to
converges uniformlyto a

U^n

this model
limit

the

by

given by
(18)

+ "^a,a,.".3(^)^(l),
"^",a,.".,(^)]'^(0)

of concavity.

Moreover,

5, 6, 7, and

Theorems

that
verify

property of monotone

the

direct

obtain

we

so

0,

"

analogue
34.

" 2,

to

It is easy to

Theorem

ifand onlyif

"

through by

of the

compared

as

experiment.

U.

0 the property

tt' "

in Theorem

stated

of
preservation

The

and

-n

remainder

the

the condition

of " 1 for

same

that

remark

We
under

be carried

proofcan

The

Proof.

1)

+ TTaCa
(t)77i

the
locating

0, then {Un)"

"

distribution

of the

also preserves

obtain

we

n"

'^'^nd

with

preserves

3 and

2 and

(1
otherwise

the

maps
end

conjugateto
consequently

Furthermore,

34.

Theorem

T'\s

of Theorems

fulfillsthe conditions

and

which
at the

corresponding distribution

the

for this

is valid

the

PSYCHOLOGY

0 and
fixed point of U"f"
(f"with ^(0)
unique continuous
of
function
of
"f){\) 1.
theory
geometricconvergence, continuity ^ as a
established
of a, a, 77i, and 772, and the form of the limitingdistribution of the particle
The
in
the
valid with slight
of " 1 remains
clusion
for the model
changes
proofs.
generalconstillhas no effect on the convergence
of standing
is that introducing
a probability
form
its limiting
of the distributions
or
providedonly the essential feature of absorbing
ary
boundstillprevails.
boundaries
Finally,in this connection we remark that for special
is
(ftaa,-^
,-n

where

The

the

parameters tt^ and


points; for example, n^

^2 the

of the

values
of the end

6.

H^e

"

but
those

in the unit interval.

includes
cases

and

so

The

on.

drift to

followinggeneralnonlinear
with
from
"f){x)
probability

to

this

function

The

ax.

that

case

the

examples

"
"f"{x)

types of

is

one

or

other

about

to

a;

0 and

"

discussed

models

the rate

one-dimensional
1

cc

"

""

2 and

4.

"
(f"ix)

in

ccx

with

only continuous

6 "

in
investigated

[1

"f"(t)]
dF{t)

""

the

d "

1 and

However,

of convergence

4-

jo
T is

become

3,
in

of derivatives,

transition operators become

TF=\
and

may

0.

stronger results

much

obtained

motion

TTg "

excludes

This

of the

subcases

some
we

0,

in this section, the

treat

The
particlemoves
learningmodel.
1
from
and with probability
"f"{x)
for
additional
important requirement
for all

entire

dF{t),

(20)

aO.

(21)

Jo

adjointto
(t/77)(0
=

(1

mX"^t)

^(0^(1

function nit). The


uniformly for any continuous
proof of this fact shall be based on the followinghighlyintuitive proposition.Let
failure at each
or
success
an
outcomes,
experiment be repeatedwith only two possible
the
of success
trial. Suppose further that the probability
/?" at the "th trial dependson

We

shall show

that

U"'tt converges

400

READINGS

a
Consequently,

as

/^

00,

-*

trials times

I\

that

of

run

success

MATHEMATICAL

IN

lengthr

is certain to

0, since /g is bounded

K.

the

On

other

twice

by

in view

hand,

PSYCHOLOGY

happen in finitetime. In particular


of no success
in n
run
probability
and equation(22) we
lemma
secure

the

of the

CA*-. Therefore,

"

iim |C/"77(a;) t/"7r(2/)|


" CA'-,
-

n"

which

be

can

exists for

for every

made

"CO

small
arbitrarily

as

shows

if

Hence,

oo.

lim

"/"77(2/)
=

lim

C/"77(x)

Since

x.

be found

subsequencecan

lim

one

^-

y, then
single

i"

for

and

for all x,

hence

If^'Trix)

"-oo

argument

an

that

so

used in the close of the

proofof

Theorem

25

that
lim

t/"77(x)

a.

W-"00

The

lemma

easilyimpliesthat

||"/"|| 1, we

can

35.

Theorem
to

the

up

Theorem

// "f"{t)
belongsto C",

uniformin

Theorem

of F
independent
This

37.

For

with

TF^^

Finally,
random

note

walks

states
at the

in

two

""

with

distributions F,

any

any

F^,^ and

ends

0 and

in

C", then

we

distribution Fg,^
to
a
converges
with respect to o, a.

T^F

of the

used
of

of T and U.
conjugaterelationship
be
can
employed to analyze

in this section

impulses.

(1

a,)w,-+

cc^x.

the
investigate

nature

In

the

case

where

the

of the

distribution
limiting
were
absorbing

boundaries

distribution is discrete and concentrates


limiting
1
distribution F and
at
weight
depends on the starting

find that the

The

1
.

is

nit)is

continuous
F(j,a

account

on

number

models.

5, we

and

that the method

various

1 and

uniformlyconverging

""

the present section

in the

as

exists

fact that

"oo

FiX
7. In

V^v

the

follows:

t.

follows

last theorem
we

lim

as

0
(C/'^7r)"""(/)

lim

obtained

then

Using
model

""**'

W-"

the

for this nonhnear

If v^t) is continuous,

36.

with convergence

is uniform.

convergence

the conclusions

limit.

constant

sum

given by
"1

i"g^^(x)dF(x),
J.
where

Many

is the

with ^(0)
uniquecontinuous fixed pointoi U"f) "f"
of "^"j,a
are
developed in those sections. In all the
properties

^^^

1.

0 and

^(1)

other

types the

ergodicproperty

Let

deal

us

and

the

with

-*

"

relevant

"

examine

Fj

case

and

(0, 1

(b) first. We
the

F^ applied to
a). Any

"

(ct^,
(1

intervals

note

that the union

unit

interval does

a)a) and

"

(a(l

the limit of the total set covered


Cantor

set

It is

C.

set

a), (1

"

by

F^^

and

overlap with

must

"

the

the

(/

givenby
Let

a.

"

us

F2[0, 1] of

subinterval

additional

this way,

its full

concentrate

with

aa;

are

open

two

2) in

is

by F^^^.

(b) ct

empty

the

walk
+

"

image sets ^^[0,1]

leave

Fg

operators

a)^).Proceedingin

"

random

denoted

"

of Fi
applications

that

seen
easily

be

of the
not

of F^ and
applications

two

The

the

with
ax
F^x
givenby
w
here
^ " j"{x)" (5 " 0. The
1
"^{x),
probability
distribution
equations(20) and (21). Let the limiting
We
1
two
"
now
cases:
(a)
distinguish
a;

independentof

limitingdistribution was
followinggeneraltype.
and x
(pix),
F^
probability1

to hold

seen

was

initial distribution.

401

KARLIN

SAMUEL

open

find

we

that

any arrangement is
on
probability

this

C.
let

Now

..

{X) =\

fl,ifx=fo

77-,

We
^

show

that

except

then

at most

neither

is zero
V^nfjx)converges uniformly to zero. Note that UTT^J^t)
value
of
if
Of
/
" t^
one
; namely, F^^/qor F~^?q.
course,

for

exists for that t^; and

inverse

lUntJ "

otherwise

1
[cf"(x),

max

exists and

only one

"l
cf"(x)]

every

\ "a,

"

cr

8.

L/^tt^ " (1
Similarly,

from
5)'*,

"

the probability
of
Consequently,

up,

have

we

established

distribution
We

of the two

one

denote
0 for
at

If
zero
{probability

now

any

this for

of F~^
application

operators obtained

every

t^

A "

"

"

F~^

F^^t^
a

total

of

so

{h

"

^q

"

Summing

Let

A"

F""

"

/;; then

1 "

^ "

i"ix)"6"0,

so

/^ in the
the

TT{t)"

that
or
same

we

F^^

get that

"

exists

way

order

nit)
^

order
specific

reverse

A",

for every

as
Consequently,

first that at least

that

"

singular

set.

least F~^

denotes

denote

note

^'^\X -y\

"

at

t^ from

F"

is

in the unit interval. Let

0). Since

construct

times.

F^,^

note

the unit interval

on

"

We

a.

"

for every

t^ to /". We

largethat

"

F^^t^, where

fny^

t^. We

by passingfrom

defined

t^

steps,obtainingt^

Choose

t^ with 0

that

observe

now

limitingdistribution
point)spread on a Cantor-like

F~^ is defined

or

"

obtain

or

We

the

then

(a) where

case

ipn^
where

a,

positivefunction

/(,(sayF^^),we

continue

at

mappings F~^

continuous

for any

/q is zero

at
a

"

examine

to

subinterval

some

"

turn

F^

follows.

the assertion

38.

Theorem

which

and

of

of the

402

IN

READINGS

TT{t)"

7? "

since

shown

thus

could

defined

Also, U^nf

{[1

max

has

spread

by

out

entire

the

unit

term

in

interval

and

involving

t/"

establish that

we

U^nt

converges

has at most
values
two
possible
U-n-f^
and
0
while
(f"{F~^tQ)
(f"(F~^to),
respectively, Utt^

"

that

four

at most

possiblevalues

and

the

maximum

value

that

"f"{F-\)"i"{F-\),
"f"iF^\)l

[1

cf"{F-\)mF-^F-\)

for the

"t"{F-\)[\ 4"{F-^F-\)]}

of

^^^ ^^
consider
the same
repeatedU^ttiq^
section.
The
conditional
of
set up
previous
probabilities
trial satisfy
the uniform
1 " 1 "??"/?""
success
^ " 0,
inequalities
pn at the "th
in
is
taken
be
of
the
the
where
this
to
to
case
an
success
application
impulseF^
particle.
that the probability
k {k "n)
of securing
It is readily
seen
by standard inequalities

To

secure

bound

observe

we

Again

is
U^tt,
'0

for

cf"(F-\m

before.

as

this end

To

zero.

achieved

be

the

/;]covers

let TTf (t) be

elsewhere.

is

"

givenby
F^^tQand -F^^^o

at

h, t^

"

a, the operator U is strictly


If " \
positive;that is,for each
that U^-n is strictly
t
here
exists
Tr(t)
an
n
so
function
dependingupon

positivecontinuous
positive.

uniformly to

PSYCHOLOGY

39.

Theorem

Now

F-"[tQ

this initial interval which

on

have

We

F".

all

positivefor

t/"7r is

MATHEMATICAL

successes

converges

deduce

as

of F

into

measures

Let

We

Proof.
denote
-nit)

/, and

on

deduce

have

Fg, where

transition

Fj

either

function

F^

f/^-n-^
-^

Thus

t.

directlythat

the

and

We

cumulative

absolutely
into

singular

measures

F^. However,
F^ vanishes.

TF^

0.

continuous
absolutely

operator transforms

measures

or

as

fixed distribution

the

uniquedistribution F"j is either absolutely


in every open interval.
positivemeasure
^^

has
^

bounded

/' of /.

We

hence

Fj is

all the conclusions

demonstrated

subinterval

and

the

"

0 for all t.

U^iTf

for every

+
F-^^

the

TF^
that

continuous

closed
6 "

C/"77- "

that

and

a, then
If " \
F
urthermore,
F^
singular.

or

for

bound

it follows

Moreover,

co.

-"

continuous
absolutely

find

40.

Theorem
continuous

"

probabilityzero

"

is continuous.

we
measures,
singular
is unique,we

that

has

F^

as

Fg is singular.Observing that

and

continuous

Let

zero

successes) is

that

before

the

uniformly to

of k
maXfc(probability

distribution

maximum

in

experimentmodel

By

by

virtue

and

of the theorem
zero

outside

39 there

of Theorem

an

but the last.


open

exists

an

interval
n

such

that

note

But

1dF,^^

"

and

the

proof of
close

We

the

with

continuous.
cT

1/2

"

theorem

the
An

a, where

is

(n, F,^^)"8"0,

complete.

conjecturethat
example where
F" J^x)
x.
=

when

cr

this is the

"

1
case

"

a,

then

solutely
F^^^ is always abby "f)(x) 1 12,

is furnished

SAMUEL

403

KARLIN

REFERENCES

[1]

Bellman,

R.

Harris, and

T.

decision processes.

[2]

R.

R.

Bush

and

RM

C.

H.

N.

878, RAND
Mosteller.

F.

Shapiro. Studies on functionalequationsoccurringin


Corporation,July, 1952.
for simplelearning.Psych. Rev.,
mathematical
model

1951, 58, 313-323.

[3] J.

L. Doob.

[4]

W.

of
Asymptotic properties

Markoff

transition

Amer.

Trans.
probabilities.

Math.

1948, 63, 393-421.

Soc,

Feller.

introduction

An

to

probability
theory and

its

York:

applications.New

Wiley,

1950.

[5] M.
[6]

O.

[7]

W.

M.

Flood.

On

and

Onicescu

theory. RM
learning

game
G.

Mihoc.

Sur

853, RAND

les chaines

de variables

Corporation, May 30, 1952.


Bull. Sci. Math.,
statistiques.

1935, 2, 59, 174-192.


Doeblin

and

R.

Fortet.

Sur

les chaines

liaisons

completes.Bull.

Soc.

Math.

France,

1937, 65, 132-148.

[8]

R.

Fortet.

variables.

[9]

lonescu

(These) Sur

I'iteration des

424, Afio

Revista, No.

Tulcea

and

G.

Marinescu.

Sci. Paris, 1948, 227, 667-669.


Received

December

19,

1952.

substitutions

40, Lima,
Sur

lineaires
algebrique

une

infinite de

1938.

certaines

chaines

liaisons

completes. C.R.

Acad.

SOME

ASYMPTOTIC

PROPERTIES

BETA

LEARNING

Lamperti

John

MODEL*

and

statistics

This

given for the two-operator


reinforcement.
noncontingent

application

and

collaborators,Bush
which

the

and

of the

Some

learning

behavior

response

postulate
trial to the

next, with

of their
in

the

linear

behavior

the

shown

viewpoint

,"

response

the r"{i)are

of

that

of

of

probability
depending
no

on

of

do

not

the

that

overt

like that

of

which

models
from

response

the

beta

one

reinforcing event,

fication
general psychological justi-

more

certain

of

construct

learning

in

linear

theoretical

believe

Spence

his

model

not

are

trial. Both

T-maze

is evidence

there

empirical standpoint

an

models

and

stochastic

the

offer

they

research

Research

article

some

with

experiments

yield good

simple postulates

very

there

exists

probability of
trial

on

n.

transformed
on

process

*This
of Naval

and

learning

development

in terms

transformation

stochastic

preceding
the

Hull

this

as

tingent
con-

of

predictions

rats,
actual

ratio scale

over

[7]
the

choice

on

behavior,
with

of responses

set

that

is the

p,

stochastic

This

like

postulates. From

basis

property

of this
that

the

has

where

on

explained

far

of

[1,7].

On

Luce

so

totic
Asymp-

cases

trial to trial

from

the

be

experiments, particularly

some

that

the

in

unsatisfactory

are

From

considered

motivated

transformation

linear

theorists
best

may

strength.

response

have

considerations

have

response

probability of response

empirical

model.

Galanter, [1, 7]

probability of

in

changes

functions

model.

beta

simple learning situations, Luce

various

to

Luce's

four-operator

and

are

and

For

asymptotic properties of

studies

paper

laboratories

university

stanford

results

Suppes

Patrick

and

mathematics

applied

LUCE'S

OF

was

and

appeared

A^

Additional

supported

strengths

on

trial n, and

Vn{i)is

simple postulates

linearlyfrom

response

in part by

response

trial to
then

trial,and

determines

in part by the Group


Psychology
Rockefeller
Foundation.

lead

the

to

strength
result

the

this unobservable
stochastic
Branch

process

of the

Office

the

in Psychometrika,

1960, 25, 233-241.


404

Reprinted

with

permission.

JOHN

in the
way

study

subject of
to

the

infer by

in connection

asymptotic
of the

means

This
probabilities.

course

mathematical

numerous

behavior

asymptotic

interest

determine

then

the

and

nearly

behavior

of the

have

is made.

Let p" be the

El be the
response

Luce's
if Aj and

of

event

situations

to

behavior

[7]and

Luce
the

on

the

responses,

Ai

trial n, and

let

behavior.

of two

Ai

response

E2 the

of

path

transformations

one

and

encounters

alternative

asymptotic

of the response

on

of

event

reinforcing

Ek occurred

trial n, then

on

for j

by the followingtransformations:
1, 2 and k

1, 2,

2n

^,k

0. Luce

"

1 and

/3,i "

want

it is assumed
most

example,
occurrences

it is

easilyshown

fact about

(1) is

first

63

that

9^ 1.

in the

of A2E1

(Generally,we
general formulation.
effects
of
reinforcement;
primary
"
"
"
/812
/322.)
Throughout
1821
/Sn

that 0 9^ pi

important

suppose

more

ordinarilyassumed

this paper

|S,-,(1 Pn)

/)" +

1, to reflect the

/3,2"

moreover,

The

[7]gives a

it is

that

trials there

occurrences

the operators commute.

61

are

of A1E2

64

For

of AiEi

occurrences

of A2E2

occurrences

62

; then

that

CX)

ri

Pi +

The

The

their

characterized

is then

Pno

where

for

"

beta model

(1)

beta

would
be
learning data
strengths v"(i)and
response

nonlinear

Ai

simplest
"

any

in which

reinforcingresponse

the

probabilities a

response

taken

probabilityof

that

seem

equation given above the


is pursued rather far by

restrict ourselves

A2

the

of

studying directly the properties of the


to obtain results on
probabilities
response
We

would

with

difficulties. We

405

SUPPES

PATRICK

it
probabilities.
Superficially,

response

to

AND

LAMPERTI

aim

of the

present

paper

/3ii/32i/3i2/822(l
Pi)
-

is to

study asymptotic propertiesof

the

probabilisticschedules of reinforcement.
of attack
[4]and by Lamperti and Suppes [6]
do not directly apply to the nonlinear beta model.
linear learningmodels
The basis of our
approach is to change the state space (the probability
model

for

Pn is the

standard

certain

methods

used

state)from

the unit

that the transformations

by Karlin

interval

to the

whole

real line in such

translations.

way

The

noncontingent
simply
of independent random
variables;
(thenext section)then reduces to sums
the contingentcases
also be studied by "comparing" the resultingrandom
can
variables. The probabilistic
tool for
walks with the case
of sums
of random
this is developed and applied in later sections. The general conclusion
to be
ment
drawn
of noncontingent reinforcefrom our results is that for all but one
case
(1) become

case

individual
which

probabilitiesare ultimately
to corresponding results
contrast

response

is in marked

either

zero

for linear

or

one,

learning

406

READINGS

models.

Absorption

at

zero

PSYCHOLOGY

MATHEMATICAL

IN

or

also

one

for many,

occurs

but

of

all,cases

not

contingent reinforcement.
Noncontingent Reinforcement with Two
If the
trial
Let

probability of

number,
TT

be

have

we

the

reinforcement

what

probability of

is called

an

/?2i

/3i2

1S22

"

j8

"

1.

for

random

of the

simplicity let

/?,
7,

1,

"

variable

and
x, |8,

numbers
is defined

rj^

and

reinforcement.

expression for the asymptotic probability distribution

probabilitiesin terms
The

of response

simple noncontingent

/3n

I
seek

independent

is

Ei reinforcement, and

an

(3)

We

Operators

of response

7.

recursively as follows:

1(3

with

prob

[y

with

prob (1

r,

\r)"^with

prob

[r]"ywith

prob (1

tt);

"

TT,

The

random

variable

X" is defined

follows:

as

X"

tt).

"

log

??"

Then
,

Jz"+

log /3 with prob

tt,

[X" +
It is clear from

(4} and

what

identicallydistributed

By the strong law


/c^

Define

now

-^

Xn

""

for any

variables

F, defined

I^*^S^

with

prob

[log7

with

prob (1

tt).

"

of

sum

independent

by

tt,

"

of large

X"

prob (1

with

preceded that X" is the

has

random

log 7

if

00

"

numbers, with

"^

7r

if

log /3 +
TT

real number

tt).

probability one
(1

log /3 +
x

"

w)

(1

"

log 7

"

tt)log 7

as

0,
"

0.

-^

0"

408

READINGS

where

0 "
with

(and

1
b, (fix),

a,

the

same

and

// for all x

Lemma.
then

Pr(X"

Pr(F"

Let

Proof.

{^"} be

uniformly distributed

"

type

same

the transition

as

6, and if Pr(F"

(p(x)"
oo)

-^

"

0,

"

6 and

M, "p{x)"

random

independent
{X"} process

[0, 1].The

0")

-^

if

0.

of

sequence

on

6 and

the other hand, for x

0, then Pr(X"

of the

process

(p(x).

"

has

M, one
0. //, on

be another

constants

"

)"

oo

oo)

b) but with

of (p(x)and

in place
probabilities

{F"}

(p(x).Let

"

PSYCHOLOGY

MATHEMATICAL

IN

variables,each
referred to {^"}

will be

by letting
(10)

'X" +

if

lX"

otherwise.

This

does

lead

Choose

Fo

that

F"+i

for all n, the

so

X"

F"

"

Pr(F"
n

on

-^

S, X"

of

{X"]

and

is

X"

the

6.

0 there
"

Xq

assume

the property

the

(10) to

(p(x)"
that

that

assumption

6 and

"

that

since

F"+i

F" +

F"

(p{x)

is

is

Yq

"

our

"

impossible,

set

in the

F"

-^

in

proved

Hence

oo.

F"

sample

Pr(X"

of the

F"

oo)

using the

"

"

0. The

{^"};
oo

second

construction

same

is

for all

sequence

in the set X"

-^

oo
,

space

is contained

similar way,

oo) is positive,so

Pr(-F"^4-

the event

for all n). But

as

b/{a

2. Let

exist. Then

if a

"

and

Pr(limsupX"

while if a

"

(")

(13)

c, and

"p{x)
=

fi "
+

and

6)

lim

(11)

since

part

linking

"

Pr

(14)
0

"

and

(X"

6 "

-^

1
.

lim

^(.r)

/3

c,

lim

oo

inf X"

jS "

"

that

suppose

and

(")c,

oo)

"

({X"|ts recurrent),

=1

then

Pr(X"-^-a"

Finally,if a

for some

"

{F"}.

Theorem

(12)

if ^"+i

since

therefore

m;

and

positiveprobability,and

of the lemma

only

of Xq

{F" } with

processes,

X"+i

F" and

"

if and

of

manner

increase.

considered

be
set

the

{F"}

The

seen.

F" is also valid for all n. This follows from

proof,note

"

F"

00
,

may

"S is

only

can

complete the

To

easily be

may

some

sequences

the

the transition

M;

"

a;

"linking"

as

value

Yq for

"

inequalityX"

construction

(9)

F" +

the

"

that for those

assert

now

law

Whatever

positiveprobabilitythat X"

for

"p{X"),

{X"} by referringit after

to

M.

"

"

transition

{^"},so

sequence

same

the

to

be linked

can

process

We

^,,,1"

X".,

(4-co))

1.

c,

4- oo)

5,

Pr

(Z"

-^

oo)

"

JOHN

Proof.
be

d "

"

with

process

The

c.

Yn

transition

constant

{F" } process

J2Zi

Yo +

be

may

"

{F" } (as in

Let

c.

probabiUties

regarded as

where

Pr

(Z,-

a)

e.

and

the

where
variables

and

lemma)

1"6

of random

sums

409

SUPPES

PATRICK

instance,that

for

Suppose,

AND

LAMPERTI

(15)
Pr(Z,E(Z,)

But

ad

cd)
Pr(F"
Pr(Z"-^ 4- oo)
I
Similarly,if a
-^

"

if j8 "

while

this
the

Consider

the

0, since

"

of

law

large

"

this

c;

the

From

numbers.

that

impUes

lemma,

0.
"

it follows that

for convergence
obtain in the

6),we

"

^)

by

also holds

6(1

-h)

to

when

"

"

"

and

"p

that /3 "

way

probabilityis

case

(with

oo

"

same

Pr(X"

replaced by

makes

the lemma

0. Since

"

Pr(X"

-^

"

ex?)"

and

/? "

c; there is then

positive probability
-^

oo

0"

"

-^

the

remains

to

in

But
whose

oo

"

N. Now

0,

zero.

It is not hard to see that X,


but not at +
absorption at
+
with probabilityone; the idea is roughly as follows. Since X"
close to 1 for
often with probabilityarbitrarily
have Xn " N infinitely

of

and

(p

"

probabilitythat from
the left ol N

"

or

infinite sequence
probability on each

walk

some

and

goes

) " 0.
positivesince Pr(X"
event
an
necessarilyindependent trials,
be

must

of not

an

the random

to the left of N

we

oo

"

trial is bounded

from

away

zero

"

"

oo

is certain

to

and remain
walk will eventuallybecome
M, the random
with probabilityarbitrarily
to the left oi N
M, and therefore X"
other
The
close to 1 (and so equal to one).
are
cases
similar;one can think
is an absorbing or
the conditions under which
+
of a "
"
c as
c or
reflecting
barrier,etc., and the process behaves accordingly.
for any

Hence

occur.

-^

"

oo

"

oo

The

{X"}

be

the four-operator

generalizationto
real Markov

(17)

X"+i

where

.Oi

"

aa

0 "

lim

(18)

and

"pXx)

(Pi(x)"

0.

then
and

v',(a;),

"pi{x) /S,=

let

M+

methods

it is

Let

""03

I"

By

be described.

Suppose

lim

and

ai

now

x, then

with prob

a,

"+a3

2"

exist,and

a:

a;,, at

that if X"

such

process

will

case

entirely similar

o,q:,

the

Theorem

the process

m-

used

to those

possibleto prove
3. For

and

Zl

above,

(^i^i

but

"

rather

more

involved,

following.

(12)holds;ifijl+ " (")0


"
0, (14) is valid.
M-

and

[X"}
n-

"

described

(")0

above,if /i+ " 0 and /x_


(13)applies;while if m+

then

"

"

410

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

ContingentReinforcement with
probability of
(on the
preceding response

Two

Operators

reinforcement

If the

depends only on the immediately


has {simple)contingentreinforcesame
one
trial),
ment.
and let the two operators
Let Pr("'i |^i)
Pr("'i |^2)
tti and
7r2
variable X"
i8 and 7 be specifiedas in (3).Using (6), define the random
s
ince
and
0
"
recursively. (Note that log 7 appears
first,
log 7
log /8 " 0,
in order most
directlyto apply Theorem
2.)
=

X,. +

(19)

X",i

FxM{l

with

log ^

prob [1

tti)

(1

,X" +

FxM){\

TT^)

^(X"),

^(X")].

that

Observe

lim

(20)

(p{x)

2,

with

then

one

"

Ti

and

lim sup

has

"

if 1
(iii)

"

tti

"

immediately

Theorem

4.

of the two-operator model, let c

then

and

Pn

"

tti

"

lim inf p"

(ii)if 1

"p{x)

one
probability

"

lim

the contingent case

log /3/log{y/fi)Then
(i) if \

and

7r2

"

Theorem

4. For

Theorem

(20) and

Combining

"

with prob

log 7

0,

T2

"

and

^2

"

and

tti

"

tti

"

"

then p^

"

then p^

1,
0.

Moreover,
(iv)if \

TT2

"

in (i)and
then
are

both

intuitive

and

1)

-^

an

"

^i

(p"-^ 0)

and

if 1

5.

ttz "

"

xa

"

and

and
of

probabilityone

"

5 "

the results

between

be clear. If 1

response

reflecting
barriers,whereas

5 with 0 "

then for some

Pr

5,

should

(iv)of this theorem


of

tti

"

of the distinction

character

probability zero
both

(p"

Pr
The

"

"

expressed
"

^1

an

xi

"

or

c,

tti

"

c,

response

they

are

absorbing barriers.
It is also to be

Theorem

noticed

that

except when

"

tti

ttj

"

for the

c,

contingent case. It can


shown
methods
that
1
be
if
c) then
c (or 1
[5]by deeper
7r2
tti
is again a reflecting
probabilityone
(respectivelyzero) of an A^ response
with those given by Luce
barrier. These
results agree
([7],p. 124) and in
Detailed
addition settle most
of the open questionsin his Table 6.
comparison
differs considerablyfrom ours
is tedious because
his classification of cases
as
4

covers

all values

of jS,7,

tti

and

7r2

"

given

in the above

theorem.

"

LAMPERTI

JOHN

We

formulated

four-operator model

(21)

3 to the contingent

finallyto apply Theorem

want

X"+i

in

Operators

Four

ContingentReinforcement with

411

SUPPES

PATRICK

AND

(1). Analogous

FxXpi))

7r2)(l

log i322

with

prob (1

Z"

logjSi2

with

prob (1

X"

log (S21

with

prob 7r2(l

Z"

log /3n

with

prob x,Fy"(pi)

"

"

general

(19),

to

Xn

of the

case

^22(^0,

7ri)Fx"(pi) "Pi2{Xr,),
=

"

-"

Fx"(P^))

""2i(X"),

(Pn{X,).

Also,
(P22ix)

lim

1"1:2

lim

"P22

lim

(pi2ix)

lim

^21(2^)

0,

lim

"Pu(.x)

TTi

0,

I-4+CO

(Pj2(x) 0,

lim

TTi

"

(22)
lim

^21(3-)

TTa

"pn{x)

0,

lim

Then

(23)

M+

M-

2 log /3,i lim tpikix)

TTz

log /321+

(1

S log /3yfclim "Pik{x)

tt,

log 13^

(1

ttz)log /322

i
and

(24)

To

apply Theorem

On

this

and

assumption,

"Theorem

For

5.

assumes

that

utilizing(23)

and

also

one

ttJ log /3i2

^^22 /3i2 "

"

the contingent

1821 /?n
5.

of the four-operator model,

case

0.

"

infer Theorem

(24), we

with

one
probability

and

"

0 and

(ii)if n+

"

0 and

/x_

"

0 then 7)0=

(iii)
if /z+

"

0 and

m-

"

0 then

0 and

m-

if m+

"

(iv)Pr(p"
I

0 then lim sup

(i) if n+

-^

Specializationof

1)

"

"

n-

pa"

0, then for some


5, Pr(p"

this theorem

to

-^

0)
cover

and

"

lim

inf p"

0,

1,

0;

Pn

the

with 0

"

"

5.

noncontingent

case

is immediate.

412

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

REFERENCES

[1]

Bush,

K.

[3]

Mem.
J.

Hodges,
J.

[4]

and

L.

variables.

Ch.

1959.

Press,

Chung,

Math.,

1953,

of

Tests

the

"beta

learning

model."

In

theory.

Stanford:

values

of

R.

R.

Bush

Stanford

W.

H.

Math.

On

J.

Soc,
M.

of

distribution

the

1951,

of

sums

random

1-12.

6,

time

Recurrence

in

moments

walks.

random

Pac.

127-136.

3,

Some

S.

Karlin,

D.

mathematical

in

18.

Rosenblatt,

and

R.

Luce,

Studies

Fuchs,
Amer.

L.

and

E.,

(Eds.),

Estes

K.

Univ.

[2]

Galanter,

R.,

R.
W.

and

arising

walks

random

in

models

learning

I.

Pac.

J.

Math.,

1953,

3,

725-756.

[5]

Lamperti,
math.

[6]

[7]

Applications,

Lamperti,

Luce,

Manuscript

Revised

R.

and

J.
Pac.

theory.

Criteria

J.

Anal.

D.

J.

manuscript

the

recurrence

(in

1959,

Chains

9,

choice

4/^7/59

received

or

of

transience

stochastic

processes

I.

press).

P.

Suppes,
Math.,

Individual

received

for

11/10/59

of

infinite

order

and

their

739-754.

behavior.

New

York:

Wiley,

1959.

application

to

learning

J.

CHAINS

Lamperti

John

used
a

The

Introduction.

1.

behavior

of

models

of

learning experiments.

as

Namely,

completes."
history of the

process,

in order

employ; however,
found

it necessary

for

the

and

details

quite in

the

The

form

simple.

trial he
of

makes

of

finite

response,

which

This

response

the

which

upon

models

may

Suppes

[6].

lines for

will

We

definitions

give

references

thorough

under

that

Received
Statistics

of

involving

Branches

linear functions

two

asymptotic
linear

20, 1958.
the

the

research

This

Office

of

in

properties.

Naval

has

of

type

Suppes

teractio
in-

[1].

in "3.

learning models
time

much

and

below

do not, except

and

Estes

along similar

constructed

given

pends
desuch

about

[4],and

Atkinson

the

of

functions

the

results

Many

ject's
sub-

the

is that

subjects and

and

are

of

[2], Estes

more

9]

form

models

or

above
of

after

the

occurred.

processes

treatment

of

are

[6, Section

these

reinforcement (again

trial

next

each

finite set

model

here

also study

from

are

on

the

Hosteller

and

trials, and

choice

models

these

of

has

mentioned

is, that

November

not

"linear

called

are

by

the

approximately stationary and

and

but

is followed

trial, where

in Bush

them

of

assumption

general conditions

behavior; that

some

[7] and

Harris

arguments,

tools

these

series of

consists

reinforcement

experiments

The

The

present

found

be

between
Precise

presented

probabilitieson

probabilitieson

and

aration
prep-

in

only

us

psychological standpoint

is

number).

response

of "2.

as

serve

by

papers

results

shall study with


a

subject

possible actions.

one

[3],

to

to

original with

it is

addition

we

From

4, and

tained
self-con-

results is the content

section is included

this

we

of learning models

cases

their hypotheses.

closely related

which

close to that

require.

we

processes

certain

past
rems
theo-

Such

past.

form

chastic
for sto-

entire

the

on

remote

in

additional

some

"

In

very

earning models."
very

of

extensions.

[3]

somewhat

that

theorems

[8] contain

Karlin

and

emphasize

should

We

relax

to

Fortet

accomodate

to

of these

discussion

and

Doeblin

given by

were

the

only slightlyon

but

liaisons

theorems

limit

transition probabilitiesdepend

whose

processes

"chaines

or

certain

employ

shall

we

order"

been

applying

this by

do

will

ptotic
asym-

have

which

processes

We

infinite

of

theory of so-called "chains

we

stochastic

of

the

study

is to

this paper

of

purpose

large

Suppes

Patrick

and

class

THEORY

LEARNING

TO

APPLICATION

THEIR

AND

ORDER

INFINITE

OF

very

special cases,
shall

We
exhibit

passed these

prove

"ergodic"

processes

come
be-

infiuence of the initial distributions


was

supported

Research

by

under

the

Group

contracts

Psychology

with

Stanford

University.
This

article

appeared in the PacificJ. Math., 1959, 9, 739-754.


413

Reprinted with

permission.

414

READINGS

This

is not

goes

to

used

in experimental

zero.

proved by
it.

Our

work,

method

our

MATHEMATICAL

IN

theorems

the

case

but

it

in almost
this

to

PSYCHOLOGY

for

all models

seems

as

all the

in which

cases

eifect, their

which

have

been

if ergodic behavior

proofs and

might

one

be

can

expect

corollaries

some

are

given in "4.
The

major

Karlin

[8], who

models.
do

work

obtains

However,
apply

not

to

of

far

so

many

the

detailed

linear model

is impractical arise
is selected
in the

estabhshed
treats

under

depend

is reinforced

right.

two

the

process

or

above.

On

the

on

rat's

probabilities. The

has

been

been

initial response

thoroughly studied

generalized

We

hope

Chains

2.
of

of

non-Markov

influenced
theorems

they

stochastic

this

given

type

here

weaker

hypotheses

it is in

[3], but the

has

also

that

his

subject.

studied

numerable
Let

the

integers

remote

proofs

chains;
point

chain); we
from

of

the

shall

out

/.

The

the

left turn

upon

have

form

shall

that

the

2.1

be

can

of

Fortet

2.2).
complicated than
T.

Harris

E.

remark

but

results

background
to

[3];
The

and

the

on

of

finite number

extended

to

the

de-

methods.

1 to

notation

"m"

and

restriction

from

subscript

his

references

theorems

2.1

more

use

convergence

and

affected.

much
not

the

original

(^Theorems

theory

probabilitiesare

Doeblin

to

future.

present

we

The

Lemma
not

are

we

integers

in the

other

chains."

order

transition

due

and

results

"infinite

section
the

are

change

much

use

left

goes

results

these

detailed
of

past.

the proof of

without

/ consist

of

we

this

generahzed

other

For

of this experiment

model

more

where

process

essential,and

case

of

make

these

Finally

is not

states

in

In

[7] gives additional

paper

he

it is depends

this development

to

processes

of

work

approach.

our

whether

2], and

ideas

the

order.

only slightly by the


for

are

further

infinite

both

that

possible using

contribute

to

these

[9].

comment

we

applications seem

(ii)

subject (a rat, say)

the

which

in [8, Section

by Kennedy

In conclusion

nearly 1, and

or

and

situations

ergodic behavior

appropriate linear model, the probability of

the

forcement
rein-

the

responses,

of

scope

trial regardless of

eventually is either nearly

the

hand, Karlin's

other

the

sentation
repreare

tion
representa-

which

Both

in which

experiment
each

such

of

paper

states

chains, and

outside

cases

when

previous

more

mentioned

is

starting point
whose

is

classes

Karlin's

of

probabilitieswith

restrictions.

(rewarded)

In

the

on

T-maze

Markov

certain

for

techniques

His

using infinite order

mild

consider

even

typical situations

interesting non-ergodic

example,

or

as

situations

many-person

theorems
the

interest.

(i)when

(and will)be studied

can

of

probabilities. Two

response

limit

the results and


cases

of learning models

limiting behavior

on

on

for
x",

(to represent
a

finite sequence

merely adds

the

the

states

ia,h,

"

"

"

specifica-

416

READINGS

of

then
to

"m'

define

We

Proof.

eS,!^

course

12

and

in k

\v^^\x + x')

PSYCHOLOGY

instead of
quantites ";,f^by using y"^*'

"""

uniformly

-*

MATHEMATICAL

IN

the

as

conclusion

^^

of

lemma

the

in (2.4);

jPi

is equivalent

Now

co.

+ x") I
v')!'\x

Wi'~''\3 +

+ x')
x')'Pj{x

i"

pl*~'Xy +

"

+ a?")}
|
cc")pj(a;

a:

.7

S ^'X^'+ x')\v?~'\3+

contains

Suppose

that

estimate

is less than

but if j
"'-m~^\

than

of (2.3) and

(In

case

Uq

"^.

series

these

1,

be

be

can

of the

the

above
is less

first term

"";/'. Taking

to

obtain

we

8el^;,'"
+ (1

idea

not

improved

term

in the

account

estimate

S)""i-"

carried

out; the

details

are

more

given.)

iterated

obtain

to

of

estimate

an

in
"S,f'

of

terms

computation the result is

some

"'"f^
^

If the

el'^^ Ne,,
same

be

x")\

second

the
value

absolute

can
=

be

(2.7) can

After

that

will

and

this
^o

1, the

"

cumbersome
Now

j^

Then

times.

Ne,,,. The

assuming

(2.7)

j^

x')- vl'-'^ij+

-^

Ne,Jza

8y

are

S)^
I ^)(1
'g'('

Ne^,,S'

extended

Ns.n,./"ii
+ i)(i

to

"

"

"

iVS^-^",..._,
.

true;

ing
call-

have

A^--^we

"

"

"

inequality remains

infinity,the

series A^, A^,


(infinite)

sy

e':^^N^e,,,,8^A,.
1

But

it

can

be

shown

A,+i

AijS.

Since

A^

"

A,

S''

(1

8)A^,,

obtain

we

Ai

and
8-'-'-'^\

hypothesis (2.5),the uniform

convergence

(2.8).
Lemma

2.2.

lim

(2.9)
and

hence

"^r ^S-S",,,.

(2.8)
RecalHng

that
diflficulty

much

without

A,,,
or

the convergence

is

\p'r\x') p\"\x")\

uniform

in

x' and

x"
.

of

"5"^follows

from

JOHN

clarity

For

Proof.

LAMPERTI

shall

we

purely analytic rephrasing is


operating independently with
In

of

view

two

least

at

being

take

can

will

and

this

occupying

this proves
n

be

can

Theorem

The

by

both

to

less

two

than

"

at

of

state

s/2.

For

bilities
probae,

and

2.1

that

most

Lemma

j^.

state

have

processes

(2.3) and

lows
it fol-

sometime

"run"

differ by

from

see

includes

in

remain

simultaneous

which

will

there

if the

But

e/2.

most

at

not

x"

their probabilities

then

n,

processes

this

the

with

period which

time

probability

x"
.

shall

now

pl"'(a;)

t^i

we

the

prove

first theorem:

quantities
lim

independent

are

with

that

such

processes

other

the

probability one,

in x' and

(2.10)
exist,

differ

that

preparation

2.1.

an

for

before

time

i at

uniformly

this much

With

is

therefore, the

n,

and

states

It is also easy

(2.9).
chosen

"

which

so

of

state

there

with

with

values

"

same

during

time

time

time

large enough

all greater
of

i at

to

sometime

(2.3) that

before

occur

the

ends

state

period of length

We

in

condition

from

be

and

up

any

stochastic

two

probabilitiesPiix),one

transition

occupied

times

of

io

have

Consider

hard.

history

2.1, for

Lemma

processes

ja

its past

x' for

sequence

although

probabilisticarguments,

use

not

417

SUPPES

PATRICK

AND

of

satisfy X

and

x,

^i

1/ ^^^^

"

is

convergence

in

uniform

Applying

Proof.

where

x.

(2.2) repeatedly,

have

we

+
+ x)"'
p,^(i,
Pi^_^{x)p,^_J,i^-,

"

x^

i^,i^,

"

"

"

"

in-^ + x)p'r\x,,+ x)

"

i^-^. Therefore

"

\p[^^'^"(x) p["\x)\
-

and

by

Lemma

absolute

value

Pi,^J^)
'

2.2, for

signs

+
^'m-i(^) ^io(^i
*

'

"

"

"

"

any

"

the

on

"

"

+ x)im-i + x)\p','^\x^

p'r\x)\

that

within

there

is

an

right

is

less

'^m-i+ ^)

"

+
Piji'^i

"

sum

to

than

so

number

p'"\x)

has

(uniform

in

x) limit

".

"

tt^.

each

term

Since

the

Since

there

of states,

E
i

TTj

2 lim pl"\x)
i

m-"oo

lim "
n-""i

weights

have

we

one,

\p[''^"'\x)p["'ix)\"
and

such

p'^\x)

1
,

are

finite

418

READINGS

and

IN

this completes the


Next

(2.11)

x,r,

If

joint probabilities.

is \,\,

x^

""",i^_j, let

v.S''') ^^::(^')
=

is, of

This

proof.

shall define

we

PSYCHOLOGY

MATHEMATICAL

starting with

probabilitityof

the

course,

past history x'

the

executing

We

define

can

of states

sequence

also

the

higher joint

probabilities:
+ ";').
pL"'(a;')S??Xa;')Pl."""(i

(2.12)

of

Analogues
the

Lemmas

arguments

same

Theorem

and

2.1

used

The

2.2.

independent

Remark.
stochastic

on

not

have

quantities
this

will

we

useful

the

past history

in

^^

imply

the

existence

"

prove

convergence

is itself

PzJ,^)is

Theorem

2.3.

the

The

Pi(x) for

the

at

state

time

variable, and

process

so

"moments"

idea is that
transition

if

we

bilities,
proba-

is i given the

it makes

sense

to

+ x)p^ (x)
p]'(r",

V,

Thus
"

al(m, x)
tz^ exists.

is the

We

as

same

shall

now

p["'\x).
prove

quantities
\m\.a\(m,x)

positive integer

is independent

data.

functions

lim")(m, x)

(2.15)
exist for every

stationary
certain

for

theorems

random

by (2.11).

that

The

measure

of members

formally, define

defined

states

This

idea

here.

with

process

probability

extended.

be

stationary

infinite sequences

in studying experimental

a]{m, x)

2.1

space

then

probability Pi(x,J that


x^

of

of

probabilities. The

define

to

convergence

prove

(2.14)

limit

the

further

study E{p]{x,J). More

Theorem

Xi

used

be

can

tt^

can

us

stochastic

where

1/ ^^^

satisfy

Pi{x) for transition

the

measure

concern

are
a

difficult to

theorems

two

with

process

Finally
which

it is not

quantities by

p':''{x') TT,

of x', and

"cylinder sets"

of /, and
need

These

the

the

for these

in x'.

uniform

is that

in this way

quantities
Urn

exist, are

be proved

can

already;

(2.13)

is

2.2

of

x.

v;

a)

convergence

is

uniform

in

and

the

LAMPERTI

JOHN

We

Proof.

use

419

SUPPES

to show

simple estimate

PATRICK

AND

that

is

a]{m, x)

Cauchy

sequence:

k +

\a'^{m-h

^vi +

^m

is chosen

(2.3) and
times

the

nothing

this proves

hmit

made

/c-"

oo,

we

Theorem

case

exist

ji

'

"

'

"

by Theorem

the

the

that

at

states

k, x')\^

limit
limit

once;

in

the

along much

of

a]{m

is the

for

same

"

"

all

"cross"

accordingly
"

k, x) exists

additional

some

h, and

for all

as

x.

moments

define

we

+ x)P:,J,x)
p]l(,x,,
.

2.3, which

generalization of Theorem

quantities
lim

uniformly

in

jk " /, and

the limits

used

x)
oc]'^'Zl
aj^\';.'j*(m,
=

all

non-negative integers

are

independent

for

of
2.3

in proving Theorem

kThe

only trivial changes, and

that moments

yields

1:

The

2.4.

argument

with

Since

is then

(2.17)
'

ing
carry-

0,

"

"

any

a}(m

"

consider

to

for

the

treats

by

many

is uniform

limit

estimate

+ x)
+ x)p;](,x,,
S p}jix,"
x)
a}Jf.'.]lim,

following theorem

The

k, x)

conclude

involving Pi{x,")for several


(2.16)

that

large.

desirable

exist; the
Another

Cauchy.
show

a]{n, x)\ is small

"

(2.15) must

to

are

can

also

It is

jo

in x^,; this

all x) if k is large enough,

fc, \al{n + h, x)

and

from

for

(and

\al{m
m

rewritten

be

may

arbitrarily

contains

sequence

last term

be

(resulting

the

be

can

conditions

all the indices except those

since a\{m, x) is uniformly


line

will

terms

the

long

two

than

ix)

I ^ S 1pC''(^) ^^'J^)
1
(x) p'^'J^ix))
xXp'J^J''^

that

provided

first

more

over

for all h

if

Thus

same

the

0, and that

-^

e^

summation

is small

2.2.

-^m +

high probability. The

IS p}{x^

which

pKa^,"+ x)\p^

large enough,

(2.5))that

with

out

k, x)\

k + li

involves

this

small;

nl{m

"

+ x)
S lP'(a^"+^

If

h, x)

need

not

involving several values

can

"

"

"

v^

all

and

x.

works

be repeated.
of

Vj

be

in this

Finally we

case

also

remark

considered, and

it

420

READINGS

be

can

tion of Theorem

apply

to

trial the

followed

by

by

the

choice

the

value

actual

of

j ^

r,

and

i +

is

sequence

event

reinforcing

reinforcing

the

numbering

Our

single
Axiom

(3.1)
where
We

0 ^

ing
representtrial

each

responses

pairs of integers
thus

represents

the

of

is

theory
variable

random

for

these

is from

here

our

the

on

to

when
inforcement
re-

is the

If En
=

"

of

in

sequences
to

present,

in

Chapter

different,

not

j\x,)

(1

1 and

P{x")

e,)P{A"

S ^jk

linear model

"

ments.
reinforce-

a;".

Thus

order, but

note

notation

the

as

reverse

in

then

j\x,^,)+ e,Xj,

studied

intensitivelyin [6] by setting:


fc 9^ 0

1^

for

1^

for /c

==

\^jj
=

Xj^
t
"

foYJ-^k

0
r

1-

trial

and

responses

this

on

response

following linearityassumption:
k and

[6].

the

use

we

and

responses

is somewhat

Suppes

sequence

formulation

the

probability of

the

past

with

notation

and

Estes

preceding

write

to

6^, Xj,,^
the

general

and

aim

model

following

are

[2], although

entire

P{A",,

obtain

(3.2)

we

[4] and

Estes

axiom
L.

on

distributions, is imposed

this preceding sequence

(It is convenient

data

these

response

linear

general

the

events

given the
For

literature:

j representing the
a

variables

general

class of

theory is formulated

The

of

sequence

The

the

number

in

envisage

random

of the

in

pair (j, k) of integers with

ordered

Any

is

variables, where

relevant

The

is, we

t, that

is

which

represented

be

may

number

of E"

n.

an

response,

variable.

Hosteller

and

being closer to

by

established
is

of trials.

sequence

of random

")

"

the value

outcome.

dealing with

of Bush

"

conventions

A"

makes

consider

we

experiment

an

probability distribution of the


random

t -\- 1

generahza-

models

consists of

A^, E^,

"

trial

particular distribution,or
In

"

events.

of values

possible experimental

"

on

k ^

0 ^

Thus

and

represented

and

predict the

provides

The

experiment

variable

trial n,

on

be

1 ^

the

follows

random

the

reinforcing

then

may

letters

response

the

of

subject

(^i,E^, A., E^,

of

This

learning models.

reinforcing event.

sequence

of linear

experimental situation which

an

each

On

exist also.

2.2.

Definition

3.

PSYCHOLOGY

MATHEMATICAL

their Hmits

that

shown

IN

that

"2.)

linear model

satisfying (3.2) we

models

be

such

(3.1) may

shall term

3\Xn)

(1

0)P{A"

j\x"-,) +

ilx"_0

HE,

if

(p(A"ib"-0
Hosteller.
formulation

also

interest
The

and

dependence

whose

(3.4)

And

a),,

if the

S P^(A"

a)

The

they

in

enter

natural

easily observed
P{A"+i
to
a

j, An

We

j). (For other

in

situation

of

ii ^

represents

We

where

in Estes

1,

""-,

[5]

and

1 and

tests

in

s,

(1

on

Atkinson

"

"

and

subjects

as

his

on

and

inforcements
re-

prior

own

be

then

trial may

for

event

responses

in

presented
re-

j\,k^) of

integers

with

sequence

of such

tuples

any

Let

variables

if El^^

E^n^ be the

A'n^ and

for the

ith. subject

on

trial

to:

and

P(a;")"

then

0^;')P{A^:' j\xn-,)+ e^^y^fi


=

2 ^jV

of Axiom

each

outcome.

random

^i

s,

well

as

2s-tuple {ji,k^,

ior

j\x")

6';*\X'A^^

Experimental

preceding

on

data

The

have

we

model

linear

the

particular reinforcing

subjects

"

of

that

suppose

generalize Axiom

For

P{A^:i,
0 ^

[6].)

see

extensions

in general

reinforcement

M.

(3.6)

ki ^ti,

then

may

quantities which are


joint probability
instance, the

examples,

possible experimental

Axiom

other

however,

way;

expression of

for

may

depend

ordered

an

and

response

the

0 ^

Ti,

We

unsymmetrical

studying

reinforcements.

by
1 ^

will

subject

responses

define

an

that the probability of

such

and

in

"

are:

the

experimentally

also interested

are

one

in

way

multiperson situations.

any

n.

formed

(3.4) are

moments

investigate subsequently.

we

lim a)^"

experimental

of

are

j 1a;"_OP(x"-i)

appropriate limits exist, we

(3.5)

which

probabilitiesat trial

response

a;".

sequence
moments

properties

asymptotic

a)^, of the

moments

the

and

Bush

of

(3.1) essentiallytheir general


plicitly
obtained, although they do not ex-

is

on

0.

in

certain

here

define

by

model

linear

the

of

indicate
We

replacing

Upon

ki^O, ki^j

E^^k,

condition

classes

combining

satisfies the

^3

if^"

Axiom

for

and

Model,

Estes

an

replaced by the simpler condition:

f{l-e)P{A"
(3.3) P(^".,

421

SUPPES

PATRICK

AND

LAMPERTI

JOHN

and

1for two-person

Suppes

[1],

situations
Let

are

be
xl,'Li

reported

just

the

422

READINGS

of first

sequence
a

Axiom

of

consequence'

it is in terms

and

their

Axiom

all the

the

in the

exist.

trial

on

with

these

reasonable

and

responses

ments
reinforce-

statistically
independent:

are

work
additional

the

preceding

logous
ana-

joint moments

To

need

we

"

responses

If P(x"_i) "

I.

exactly
(x['j,n

moments

interested

of Axiom

when

given,

define

we

if they
J]^....,J

in terms
that

assumption
are

that
xl^'l^

asymptotes

latter moments

that

shall also be

We

(3.4).

to

and

of

It is

of subject i.

reinforcements

and

1 responses

"

PSYCHOLOGY

MATHEMATICAL

IN

then

P\An

.?!"

^=

'

"

"

"

An

Js\^n-l)

iA

^^

restriction

experimental

The

4.

of notation,

matters

some

of

asymptotic

broad

In

it will be

we

write

may

to

indicate

Xm

x' is

for counting back


To

interested

are

just the

notion

of the

on

all x^, x' and

the

other
1

Proof

if, and

/Ck"

that

It is to

of

place

of

in

of

terms

We

Xn-i.

given trial

probability of
of

P(A"

the

reserve

ilx"-i)
"sum"

The

x^-i.

of

notation

the

subscript

n.

to

define

in

reinforcing

an

exact

pending
de-

event

and

trial outcomes

past

has

only

reinforcement schedule

if,for all

x^

be

X')

includes
noticed

P{En,

the

k, n

that

fact

is

analogous

/CK

and

n' with

the

of

use

of

to that of Theorem

X")

pendent
inde-

with

past

n, n' "

precedes E^.,^^

Aj," which

response

side of (4.1) yields independence


of this

the

x"

P{En

trial n.)

some

it is desirable

conditional

linear model

dependence of length

(It is understood

some

trial number.

Definition.

(4.1)

use

x')

last

finite number

to

sequence

trials from

of the

only

on

and

j\x^

in the

combined

give

begin with

We

L.

clarify the general theorem

the

way

we

existence

the

on

theorems

ergodic behavior.

convenient

P(A"

dealing with

After

the

of

hypotheses

guarantee

satisfied in

been

theorems

general

satisfying Axiom

this section

Thus

"2,

which

models

one-person

The

moments.

conditions

has

learning models.

state

we

"

linear model.

the

for

theorems

Asymptotic

Ji\^n-l)

Axiom

implied by

multiperson studies employing

the

^\^n
l

on

one

side and

trial number.
4.8 of

[6].

The

n'

on

term

424

READINGS

(2.5),consider

establish

To

|P(A",.i

(4.3)

j,E",

=:

where
contains

least

at

i|^",

1P(A",.,

(1

do

We

not

the

obtain

6,)\P{A,,

6"

factor

(1

6^*)at

(1

inequality is
for

"

least

e,r\P{An,-u

length

not

of

times,

1,

that from

so

equality

get, ignoring

x")\

k, x

x")\

repeatedly,

j, E,"

P{A".-Ax")\

diiference

The

we

that

so

P(A"",i

sequence

The

x")

apply Axiom

we

J\^')

x\

than

more

k\x

j\x

x")\

of (i)

j\E^"

P(A""

as

/clx + X')

i, "'",

is the

where

the

x")\

k, x

m*.

"

P{E,,"

P(A"",i

X')

0, but

that
"

x')

j\x

know

(4.4) 1P(A".,,
^

A;,a;

with

by virtue

k\x

right-hand side of (4.3) we

the

to

once

where

inequalities:

=^

j\E^"

and

x,

k*,

for

k\x-\-x')

Axiom

of

E",,

3,

P(A""+i

of

terms

hypothesis,

;r,.,^^, P{E",
Applying

m*

P{A"".,

x')

occurrences

(i)of the

from

follows

last

the

means

x,^.

X')

k,x

=:

following equalitiesand

the

k\x

7r,,,JP(A",^ij\E",

PSYCHOLOGY

MATHEMATICAL

IN

term

k\x

x")\

on

the

the

obtain

(4.4) we

right of this
estimate

m*

(1

"", ^

e,.r

whence

is (2.5).

which

On

so

for

the

ting

the

order

are

allows

trials without

can

chain
that

I:P(A"

^f

which

they

original condition

inclusion

are

of

independent

of

cases

reinforcement).

where

as

insures

do

not

to

be

existence

the

depend

upon

about

made

of the

of the

sums

given in [3] would


some

k\xn-,)

i,"'"_i

be expressed

remarks

several

If all 0* -V= 0, the

(2.5)

a)^"

(3.5) and

moments

There

ilx"_,)

moments

infinite

jS^" exist and

2.4 that

Theorem

from

know

the

the

initial

But

of responses.

P(A"

(2.5) we

of

cross-moments

distribution

and

(2.3) and

of

basis

the

asymptotic

e^

be
are

cross-moments
of

the

limit-

initial conditions.
the

theorem

satisfied;our
0 (i.e.where

just

weaker
there

dition
concan

be

JOHN

First,

proved.

observe

we

that

min

(4.5)

on

Contingent

k\An-"

x)

3,

for

with

case

that

(i)of
need

(4,5) we
An

and

responses

cases

(ii)aof

trials which

general apply
On

the

that

the

2.2, conclude

Theorem

(I) and

satisfy

Experimental

their

is

that

they

method

(I) excludes

of

if

(I) only

may

not

in

virtue

of

0.

4.1

[8] do

Karlin

are

moments

response

constructive

in

Theorem

proof of

0.

It
computational difficulties.)
theorems

to

j', tt^-jj,^

[5] test

Estes

(ii)aand

for

although

known

no

(The

the

apply

satisfied,and

asymptotic

is

(II)

4.1, there

Ti^jjr

for all j and

0 and

convergence

basis of the

To

0.

"

P("'"

"

immediately

cause

(II),and

to

j, x)

for all j.

j',x)

j',x)

actual asymptotes.

the

noted

be

^^

about

fact

P(^"-"

let

model

Estes

[5].

A"_i

3,

j, A^-^

that

the

Let

basis of Theorem

the

on

non-reinforced

we

by

may,

of successive
the asymptotic jointprobabilities

that

also exist:

responses

Corollary

for

positive

(4.5),(i)and

of

k, 7Zj^{v)
i^ 0
Estes

experimentally and

test

to

that

(4.1) is

such

interesting

such

case.

Theorem

computing

also

P(A"

in

given

are

contingent

such

Then

for

has

interesting experimental

In

some

P{E^ =^k\A^

exist

of

x^*

in terms

v.

lag

only that for


0, 1, 2

Double

II.

simple

sequence

of

described

for all
Tt^^{v),

"

need

(4.5),we

all

what

k*

event

4.1.

I.

for

reinforcing

number

be

can

the

matter

no

preceded.
model

hnear

Theorem

data

trial

every

reinforcements
the

dition
con-

7t,c'.x.^

interpretation of (4.5) is that

probabihty
of

simple sufficient (but not necessary)

(ii)bis

for

The

425

SUPPES

PATRICK

AND

LAMPERTI

every

1.

//

limit

the

-i

as

(-^n+m

hypothesis of

the

^^

^-

Theorem

4.1

is

satisfied, then

of

co

3mf "'^n
+ m-l

"

3m-l"

'

'

'

"

-^n

"

3o)

exists.
We

random

1, and
of

the

may

regard

the

probability vector
distribution
existence

probabilities.

F^
of

on

the

quantities P{A"
with

trial
moments

an
n.

jlx^.^), for

1 "

arbitrary joint distribution


The

following corollary is

a) independent

of

the

j ^
F^

r
on

as

trial

consequence

initial response

426

READINGS

Corollary
is

there

IN

the

//

2.

PSYCHOLOGY

MATHEMATICAL

hypothesis of Theorem

unique asymptotic distribution

is satisfied,then

4.1

F", independent

of F^

which

to

the distributions

have
of

F" converge.
multiperson situation characterized

the

For

analogous

theorem

this theorem

we

such

have
and

schedule

all x"" x' and

^.(1)...,.(.o,

Theorem

(i)

exactly

m,

For

in the

use

of

reinforcement

as

we

i ^

if for all /c, 1 ^

in

did

s, all

we

hypothesis

schedule

with

(4.1), namely,

and

M,

n' with

we

n, n' "

x"

PiEir

P(Ell'

Let

4.2.

has

^//

4.1.

notion

the

of length

past dependence

Theorem

to

define

/ and

by Axioms

k"'\

k^'"\x^ + X')

/c^')\x,,+

k^'\ "',"[:'

be

^//

linear

s-person

an

schedule

reinforcement

Eir

...,

X")

model

with

that

such

dependence

past

of

length m*,

(ii) there
(a)

^1H,*^0,

(b)

there

all integers

is

8*

and

an

i ^

1 ^

integers k^^^\ for

are

that

for all

that

such

m^

such

s,

the

of the initial distribution


The

Proof.
"

"

",

that

k^'^ is the

Using

for which

establish

we

to

(2.3) and

P{Al,'l, i",

X)
Pn.AJ''"\k,

we

PiAi'l,

omit

the

verify (2.3) we

applying

"

"

",

subject

"

"

preceding

be such
j'^'"*
the

state

notation, it

simpHfy

(2.5). To

let

as
k^''"*)

",

by the ith

the

on

hypothesis,

k^'^*,
i^'^*,

2s-tuples

as

made

response

that

k'-'^*of the

(j^'"\

are

is

jo

venient
con-

define:

P..,{j,k\x)

To

take

We

defined

now

are

for

and

responses.

j^'" is the

reinforcement

reinforcements

the

^ 0.
^^(^,.^(0-

Moreover,

chain

the

j^'\ k'^^\""",A;("), where

subject and
trial.

of

states

of

ctZi exist

7\i). (3),...,,(s)of -/^

moments

"

asymptotic

independent

(i*^\

and

^ r
P{E\r.n., k^'^\ "-,E\:U,^ k^'^*\x,,)
Then

sequences

now

Axioms

.-.,

j^^\ E^:'

A'"n,

i^^^l^^''
-

k^'\

superscript notation
proceed exactly
I and

instead

as

Elp

.,

from
in

A;",

the

of L, and

""",

k^^",x)

and

^^'^

X.

proof of Theorem
obtain

we

k\x") ^n
mo+l(y,

6*^(0
A^(f)*^(t).5*
.

k^''\x),

2?n+

that

4.1,

JOHN

For

first observe

(2.5),we

\Vn'^i{3,k\x+ x')

AND

that

PATRICK

the

hypothesis

riVn^'^APAk,

a^')

a;

notice

the

-np""UJ^'^k,x

x")\

Up^"^,{j^'^\k,x

+ x')-p^.,,,{j^'^k,x
+ x")\
x")\p".,,{j^'"\k,x

development,

same

x
E \Pn'.AJ^'^\k,

X')

Continuing this

x")\

right-hand side is

+ x')\Up",^,iJ^'"
\k,x
W^\k,
^^.r.*\pn'^,

i=i

that

next

x")\

1=1

(i)of

of

by virtue

Vn"AJ, k\x

II Vn'AP^\k,
TT,,"*!

We

427

SUPPES

Axiom

and

LAMPERTI

obtain:

we

x +
x")\
x')~ Pr,"^.,U'''Ak,

1=1

And

Hne

by the

contains

sequence

reasoning used

of

in the

proof

of

at
least
(i*^^^',", A:"^'^*)

state

"

4.1, if the

Theorem

"

times

of

""j

last

the

is

quantity

I
i

Provided

conclude

we

follows

then
A

w*

"

pair

(2.5) holds.

from

the

Finally, we
which

want

of

this

linear functions

these

in the

as

from

remark

of

the

Theorem

section
are

from

asymptotic

of Theorem

case

which

moments

Q.E.D.

4.1.

which

just proved

are

4.1.

Axiom

that

the

theorem

apply

involves

diminishing, i.e., have

asymptotic results
which

to

estimate

an

existence

theory of "2
given after

two

distance

are

The

corollaries follow

like the

exactly

this inequality yields

that

of

'^
=

slope less
to

linear functions
than

The

one.

learning models

many

replaced by non-linear

functions

in

having

this property.

References
1.

2.

C.

Richard

in terms
Robert

R.

and Patrick
analysis of two-person
Suppes, An
learning theory, J. of Experimental
Psychology, 55

Atkinson,

of statistical
Bush,

and

Frederick

Mosteller,

Stochastic

Models

game

situations

(1958), 369-378.

for Learning,

New

York,

1955.
3.

W.

France,
4.

W.

Doeblin,
65

K.

and

R.

Fortet, Sur

des

chaines

liaisojis

completes. Bull.

Soc.

Math.

(1937), 132-148.
Estes, Theory of learning with

reinforcement, Psychometrika,
5.
,

Of models

and

22
men,

constant, variable,

or

of
contingent probabilities

(1957), 113-132.
Amer.

Psychologist,

12

(1957), 609-617.

428

6.

READINGS

W.

Model

Sirnple

for

Mathematics

and

7.

T.

8.

Samuel

(1953),
9.

E.,

Stanford

chains

On

Harris,

of

Some

of

Studies

of
random

of

Contract

Nonr

1957.

Learning

Pacific

order,

arising

16,

Theory,
225(17),

edited

by

sion
ver-

R.

in

J.

learning

Math.,
models

(1955),
I,

Pacific

Kennedy,
Math.,

University

(1957),

convergence

1107-1124.

theorem

R.

707-724.

J.

Math.,

725-756.

J.

The

plied
Ap-

abridged

An

Tlieory,

I.

1959.

Press,

infinite
u'alks

Learning

University,

Mathematical

in

University

Statistical

No.

Report
Stanford

Laboratory,

Stanford

Estes,

Karlin,

Maurice

Pacific

K.

Foundations
Technical

Learning,

Chapter

as

W.

PSYCHOLOGY

MATHEMATICAL

Suppes,

Statistics

and

appears

Bush

Patrick

and

Estes,

K.

Linear

IN

for

certain

class

of

Markoff

processes.

MARKOV

FINITE

PROCESSES

George
MASSACHUSETTS

Finite

Markov

A.

Miller

INSTITUTE

TECHNOLOGY

OF

reviewed

are

processes

PSYCHOLOGY*

IN

considered

and

usefulness

for their

alternative
in
data.
The
various
description of behavioral
responses
and
define a vector
changes in the probabilities
experimental situation
an
space,
in this space.
ods
Methof these alternatives
represented by movements
are
considered.
of fitting the theory to experimental data
are
of transitional
with
matrix
The
constant
probabilities
a
simplest process,
the effect of successive
that
to
is applied repeatedly
trials, seems
represent
function
that may
be useful for
A matrix
learning data.
inadequate for most
learning theory is presented.
the

in

In
as

the

general

two

where

areas

probabilistic considerations
of

these

two

in time.
effects
This

areas,

The
of

basic

based

model
the

the

has
very

of

misrepresent

statistical

the

of

assumption

probabilities are
are

examined

possible

familiar

use

because

ignored

sequential
randomized.

or

probability

of

For

and

ask

failures

models

help

their

to

attitude

In

either

fail

rejection

abandon

is to

had

in

pendent;
inde-

of

the

from

dependent

incorporating

dependent

be

can

models

processes.

and

may

not

must

lead

familiar

it is intrinsic
are

variables

proper

what

Markov

usefulness

example,

measurements

independent

more

this

psychology, however,

mathematical

finite

for their

of

Such

process.

simplest
the

to

relatively invariant

length

be

construction,

test

It is characteristic

are

at

can

successive

inadequate;

as

and

variables.

that

independence

probabilities. The

it

theory

basic

concepts

and

profitable results.

to

use

explored

relatively successful

worth.

observations

problems

learning

to

attempts
or

led

often

not

be

random

dynamic

more

notion

can

makes

independent

upon

With

the

secondary

are

situation

their

proved

ago

that

parameters

measurement

fortunate

long

however,

been

psychology

quantitative science, i.e.,sensory

has

psychology

this

limitations

such

paper

for

processes

describing psychological

data.

1.

Simple

experiments
in
order

in
*This

Jersey, while
This

the

time

which

article

the

usually

was

author

appeared

written
was

in

on

Often

choices
at

the

in

the

Psychometrika,

for

Advanced
Harvard

1952, 17, 149-167.


429

to

purpose

from

data

from
of

sequences

possible
The

occur.

leave

of

form

it is

Institute

sabbatical

The

Alternatives.

Two

come

continuum.

alternative

article

with

Chains

Markov

ignore
of

logical
psycho-

choices
the

this

temporal
discussion,

in Princeton,
Study
University.

Reprinted

with

bedded
em-

New

permission.

430

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

should
the temporal sequence
situations in which
however, is to examine
model
of dependent probshall adopt the Markovian
be ignored. We
abilities
not
with
the
We
such
discuss
simplest
begin, therefore,
to
sequences.
of

possibleexample
Consider

and

letters A

the

produce

are

its future
We

shall

the

words,

where
.

the

might

durations

this sequence

is

produced

of

of the

state

present

and

number

of the trial: 0, 1, 2,

the two

alternative

p*"^(^) probabilityof
p(A)

asymptotic

d"

the

system

governs

plT^B)

at

at n, the

conditional

probabilityoi

at

2, 3,

on

as

trial n,

or

of two

1.

-\-m,

probabilities.

of the matrix

roots

it follows
ways.

at trial

occur

can

"

of transitional

X,-

1 in either

trial n, considered

at
probabilities

of
probability

characteristic

an

n.

"""

"

conditional

at trial

at n, the

matrix

Alternative

p'"'
(^4)as

of

set of absolute

given A
given A

Va{B)

responses.

alternative

value

...

vector; [p'"'(^),p'"'(B)].

assume

of trials

sequence
.

are

responses

alternatives. If the

two

development.
adopt the followingnotation:

at

that

the distribution

that
i.e.,

In other

n.

follows

one

ABBAAABA

of responses

process;

of trial

outcome

alternative

two

of these

choices,then

these

ignored. We

only

of

at trial n + 1
probabilities
of trial n. However, the knowledge of outcomes
the outcome
the
does not change our
descriptionof the system if we know

by a Markov
depends upon
prior to

choice

designate

sequence

latencies

and

of

which

in

experiment

an

trial consists

possible.A

chain.

Markov

1 in either

trial

on

n.

of two

Either

ways.

Similarly,B

fact leads

obvious

This

T.

to

can

the

it

occur

following

equations:

p'"\A)pM)

p'"'{B)ps(A) p'"*^'(A)
=

^j^

p'"\A)pAB) +p"'\B)ps(B) =p'"''\B).


In matrix

notation

these

be written

equations can

ipM)

PB{A))y"\A)}W''''\A)l
^2)

]p,(B)
psiBMp^'HB)) y-^'HB))
The

reader

is assumed

If the distribution
vectors

d" and

of

to

be

familiar

probabilities

d"+i in

on

with
trials

two-dimensional

the
n

and

space,

of matrix

elements
n

then

1 is

the

theory.
regarded as the

square

matrix

of

432

similarlyfor
convenient
det
The

{T

\I) =\'

of this

equation

equation

and
of T

y,-

we

obtain
the

(r

roots

PsiA)]

Pb(A).
that

note

into Tvi

the

columns

X.y,-and

S,

solving

[1, Pa{B)/pb(A)]

vectors

of

unity is always

and

so

from

Eq.

+ ^=(^)
-if)owA)-p.u)rf^^(^)

PsiA)

be written

Eq. (6) can

Pb(A)

(6)

conveniently

more

\pb{A)pBiA)}

,
__

(pAB)

Pa{B)+Pb{A)

pAB))
[pM)
pb{A)T ^ ^^^^^
i-pAB)
pAB)+pBiA)
-

|Pa{A)

Since
to

zero

as

With
A

0.

of the matrix:

(1, "1). These vectors comprise


(5) we obtain,after invertingS,

"

unity,we

are

and

rpn

in the

be written

can

roots

Pa(A)

Substitutingthese

vectors,

[pM)

Pb(5)]X+

Xa

of all the columns

sums

the characteristic

are

characteristic

for the

\pM)

of these matrices.

root

determinantal

the
subscripts),

Xi

PSYCHOLOGY

form

roots

Since the

MATHEMATICAL

IN

READINGS

on

00
J

the

so

Eq. (7) we

successive

p^B)

can

PBiA)

right of Eq. (7) goes


first term
represents the asymptotic form of T".
and so obtain the probabilityof
calculate T"do

Pb(A) \ "

"

^'^M. (7)

1, the second

term

on

the

trials:

Pb(A)

p''\B)Pb(A)
,.,^1n?"^"^(A)?".(^)
-

The

value

r^/,^

of

Pa(B) + PsiA)
It is apparent that

Eq. (8) can

be written

p^-'iA) ail
=

be-n,

(9)

433

MILLER

A.

GEORGE

where

pAB)
b

Eq. (9) is

The

from

are

term

the matrix

subject

2^

[pM)

p^^\B),

Pb(A)].

function

"

follow such

may

form

that

Suppose

of the

use

alternative

two

scribe
frequently used to dethat
be noted, however,

learning function, the individual


do not

generating stationary time series that


"learning" probably should be reserved
operator changes on successive trials.

shall illustrate the

We

'

learning experiments. It should

the average

subjects

-In

exponential growth

an

data
while

-p''\A)

Pb{A)
+ PsiA)

with

chain

Markov

in which

cases

numerical

called

are

responses

represent learning.

for those

ample.
ex-

and

right (R)

by the percentage of
(W), that p''"\R)and p^"\W) are measured
subjects in a large sample that choose R and W on trial n, and that the
constant.
observed
successive pairs of trials are
transitional probabilities
on
values for T do
Assume
the followingnumerical
di :
wrong

time;

follows wrong

wrong

that the successive

73 per

values

of

Eq. (8) we calculate


0, .27,.46, .59,.68,etc., approaching

p^"\R) are

we

know

that

probabilityof R

on

cases

defined

on

we

Function.

wish

to

will mention

consider

tedious

next

it

0, 1, 2,

.)

from

the

for the

function

matrix

chains

Markov

because

now

the autocorrelation

compute

to

of such

simple parameter

We

function.

is most

or

(n

.7")

particulartrial a W occurred,this equation gives the


succeeding trial.

is the autocorrelation

complex

the nth

2. Autocorrelation

not

of the

cent

per

of the time. From

cent

p'-'iR) .9(1
If

97

equation is

of .90. The

the asymptote

right response

by another

is followed

right response

j-27
1.73

more

is either

of transitional

probabilities.
The

autocorrelation

displaced 0, 1, 2,
series with
responses

steps. With
+1.

itself is, of course,


on

is the correlation

function

trials 1, 2, 3,

"

"

"

are

zero

of

time

series with

itself

displacement the correlation of


a
displacement of one step,

With

correlated

with

the

responses

on

the
the

trials

434

READINGS

2, 3, 4,
after

IN

binary choices is fairlylong,the autocorrelation


step is given by

one

We

that

note

Vi

More

is

pM)

PsiA).

characteristic root

(10)

of the matrix

of transitional

abilities
prob-

generally,
r"

Pa'^A)

where

PSYCHOLOGY

.If the series of

...

displacement of

MATHEMATICAL

p'rXA)

p]r\A) are

and

that these elements

of T""

^^^

^'

p's'-XA),

elements

(11)

of T"*. From

Eq. (7) we

observe

are

Va{B) + pM)

and

Vb{A)

Vb(-)(^)

ps{A)[pM)

PBiA)T

When

these values

short,for
and

If

simple Markov

is the wth

[p^{A)

in

pBiA)r

(12)

rT.

chain, the autocorrelation


of the autocorrelation

power

simple example

noted

that

the

is

between

as

Markov

Consonants
function
one

is

easilycomputed from

letter the

.06, -.03,
The

value

of the

.51

in written

zero.

Samoan.

this matrix.

correlation

function

the determinant

is the determinant
When

1.

For

autocorrelation

The

successive

displacements

coefficient is 1, ".49, .24, ".12,

etc.

autocorrelation
as

Py{V)\ h

follow consonants

never

positions
and

|0.49/

\pc(C) py(C))
iPciV)

language. E. B. Newman
(C) and vowels {V) in Samoan
chain with the. followingmatrix

of consonants

sequence

between

by the Samoan

provided

writingis adequately described


of transitional probabilities:

of

obtain

Eq. (11),we

Iri I " 1, then |r" |declines monotonically toward


A

has

substituted

are

r"

In

Pa{B) + ps(A)

the

for this

of T". Thus

can

process

is the determinant

Tq

of T, r'2is the determinant

distribution

simple
of

also be
of T"

scribed
de-

I, Vi

T^,etc.

of

1 depends upon
events
at n +
probabilities
prior to n as well as upon n itself,
Eq. (10) stillholds as a definition of the
than two
autocorrelation
but
more
function,
Eq. (11) does not hold. When
unsealed alternatives are used, the autocorrelation
function is not defined.

3. Extension

to

More

than Two

Alternatives.
than

The

equations
straightforward.
Designate the alternatives A, B, C,
to

experiments involving more

two

extension
alternative
...

of the matrix
responses

A^. Then
,

we

isj

havei

A.

GEORGE

435

MILLER

Vn{A)
Va{B)

vb{B)

(n

l)

{A)

Vn{B)
\ =\

[va{N) Vb{N)
General

solutions

considerable

[p"'''\N)}

V.{N)]W\N)]
known

are

(13)

for certain types of operators. These

interest in

and

initial distribution

five

genetics,where

the

of

are

of T

elements

are
physics
scriptive
given by theory. The present use of such operators is almost purely dehowever, for we do not know what specialtypes of matrices will
of
the
be
greatest psychologicalinterest.
It is not always necessary
to find a general solution. A qualitativeunderstanding
of an experimental situation is often provided by simply transforming

the

or

direct

steps by

ten

matrix

plication.
multi-

three
analyzed
might
kinds of responses:
correct
(G).
(C), slightlywrong
(S), and grossly wrong
of learning a subject begins by making gross mistakes,
During the course
then slightmistakes, and finally
Such a
correct
to make
responses.
manages
situation could produce a matrix equation like the following:
For

example,

(PciQ

learning situation

psiQ

into

be

p,

.(5)""19""'(*S)"".l
=

(pciG)
W"\G)) (o
ps(G) PoiG))
It is tedious

.1

.7)
(l)

general solution of T", and it is easy to see


happens. The proportion of grossly wrong

successive

on

.3""0".

to find the

multiplicationwhat
declines steadily:1, .7, .52, .40, .32, .26,
errors

.6

responses

proportion of small

.08. The

"

first increases,then

trials at

by direct

decreases:

0, .3, .39, .40,

gives a roughly Sresponses


.69. This situation is analogous
shaped function: 0, 0, .09,.20,.30,.38,.45,
to pouring water
from one
vessel into a second,which in turn pours the water
into a third. The
asymptotic distribution can always be found by solving
.38, .35,

"

"

.23. The

"

proportion of

correct
.

the equation Td"


The

form

of

distinct roots, as

d"

general solution

"

can

be

indicated,for finite matrices

X,- represent the N

follows. Let

polynomial det {T

X/). We

define

XJ)

(T

characteristic

set of matrices

roots

with
of the

/.(T) by

fm
(r

(X,

J)(T
xo(x,
-

X2)

'""

"

"

"

X,-J){T

(X,

x.-_0(x,
-

X,. J)

x,.o

jT

"""

"

"

"

(X,

\^I)
x.v)

(14)

436

READINGS

In terms

of these

matrices,T
T

If

g{\) is a

rational

g{T)
In

\jm

transformation

the roots

X,-

we

other

know

"

\^f^{T),

"

(15)

"

"

gMf^T).

"

(16)

have

is

that

fall between

roots

g{X2)UT) +

xr/iCr) +

expressed

polynomial, then

X",we

PSYCHOLOGY

\,f,{T) +

9(\;)MT) +

2X2

be

can

scalar

if g{X)
particular,

The

MATHEMATICAL

IN

x:/.(70 +

expressed in

Xi

be

can

and

"1

"

"

xif^T).

"

this form

Eq. (7).Concerning

in

assigned the value


the

Thus

+1.

(i7)

1, and

that

value

asymptotic

all the
of T"

is

givenby /i(r).
The

solution for

particularmatrix

always
polynomial, det(r

be obtained

by (a)finding
X7); (b) determining the
fi{T) according to Eq. (14);(c) substitutinginto Eq. (17);and (d) solving
This procedure has the advantage
T"do for the given boundary conditions of do
of avoiding the problem of inverting a large matrix, but if two
or
roots are
more
nearly the same, the computations may be quite difficult.
the roots

of the characteristic

can

"

The

autocorrelation

because
alternatives,
the

to

various

alternatives.

2X2

much

case

same

inconvenience
the

prior

to

as

to N

that

of the

state

of T",

as

coefficient varies

for

predictingthe
What

case.

system

in order

we

to

to be

no

outcome

We

is to

such

is

must

known,

1. We

are

"1,

possibleusefulness
explored.

trial
at

do

must

make

and

+1

in

in
periodicities

memory.

of the

outcome

coefficient

psychologicalpurposes

have

processes

The

abilities
prob-

and
coefficient,

reveal

can

For

Responses.

different

the

to

of n, lies between

needs

according

of transitional

matrix

function.

transformations

that, if the

values

and

processes,

unordered

two

autocorrelation

function

than

more

correlation

the

autocorrelation

Markov

irrelevant

of

and

Markov

Compound

to

of the

characteristics

an

the non-Markovian
of

for

of the correlation

determinant

restriction

are

the

way

4. Extension
an

defined

determinant

of the

0 for the

of this extension

remove

the

determinant

toward

the

is not

possible assignments of numerical

many

identical. The
declines

the value

However,

has
the

function

expand

must

now

events
sider
con-

definition

the

Markovian

systems

it is

in

largerspace.
and

If the

trial

of events

"

at
probabilities
but
1,
knowledge

for
to

be

Markovian
the

we

-f- 1,

we

have

by changing
state

of the

priorto

the

system.

definition

by

characterize it by pairs of responses.

the

1 does

"

non-Markovian

system

the outcomes

-\- 1 depend upon

of

an

are

This

of
two

of trials

change

our

system

is made

Instead

event.

occurrence

If there

not

diction
pre-

acterizing
of char-

single response,
alternatives,

atomic

A.

GEORGE

437

MILLER

and

B, in the originalsystem, then there are four compound alternatives,


AA, AB, BA, and BB, in the new
system. Thus we must define a distribution
four
a
nd
is
matrix of fourth order:
7"
d" over
a square
alternatives,

Td"

Paa(AA)

Paa{AB)

Vba{AA)
Vba{AB)

V'^'\AA)

V'"'{AB)
p'"'
(BA)

Pab(BA)

Vbb{BA)

Vab{BB)

Vbb{BB)] [p'^Xbb)J

^p'^^^'^AAy
p'^^'^AB)

(n+l)

d",i

(18)

{BA)

yp'""^''
{BB)]
Note

that

for the

AB

"

"

to

system
the

AA

of the

many

"

BB

made

been

vowels

{B)

can

as

E.

B.

AA

from

move

others

to

state

some

BB

to

it is not

zero;

in

singlestep. For

in less than

two

and

of vowels
Newman.

in written

consonants

The

by
sequence
be adequately representedby a matrix

of consonants

ample,
ex-

steps:

of the form

Hebrew

{A) and
of

be applied iteratively
T can
to
before,the transformation
distribution.
stable
into a final,
unique,

As

possible

AABJB.

in the sequence

of sequences

Tabulations
have

cannot

system
"

from

move

are
probabilities

transitional

Eq. (18):

carry

any

initial distribution
This

merit. For

to

seem

requires
tray. In

animal

an

order

define

could

responses

of the Markov

extension

and

to

be

similar

of the

approaches

manner.

system

of

to

length m

the transformation

The

include

into account

far

verbal

case

is

as

much

of the past

so

possible

there

would

be of order

would

in human

verbal

all the

-\- \. Thus

adequately discussed in this paper.


In principle it is possibleto extend

to take

as

as

the data

conditioning
example, fixed-ratio reinforcement
then
times in one
to respond m
approach the food
way,
we
keep track of the sequential aspects of this behavior

sequential dependencies arise


in

be carried

in operant

state

states,and

can

process

behavior

2'"*\ More
and

complex, however,
the

Markov

sequences

complex
be treated

can

that

definition

historyof the system

of

be 2'"'^^alternative

as

it cannot

indefinitely
one

desires.

438

IN

READINGS

Cases

known, however, in which

are

PSYCHOLOGY

MATHEMATICAL

the extension

far into the past in order for the Markov


infinitely
Such

the information.

likelythat

seems

cases

better handled

are

methods, and that Markov

other

are
probabilities

model

in other

learning situations will need

most

when

to be

summarize

to

At

ways.

all

present, it

be described

to

carried

by these

using a singlematrix of transitional

processes

valuable

most

need

would

the behavior

has settled into

relatively

stable pattern.

5.

Fit
Leastr-Squares
describes the

one

T. We

give the best estimate


fairlynatural

the

assumption that

trial can

wish

for T from

be the most

not

may

Under

behavior,every

singletransformation

of the
will

to Data.

be considered

to find

formation
singletrans-

measurement

least-squaressolution that

the available

data. The

followingprocedures

efficient for Markov

extension

of the

processes, but they represent


used
with more
familiar statistical
procedures

represent the observed

problems.
introduce

We
is formed

by placing
from trial
trials,

trials,then

matrix

is

an

(n

placing in successive

trials from

through

the elements

to the observed

wish

in iVto

to determine

have

and
the

Eq. (20) we

Thus

is also

A^ and

the

observed

known

matrix

distributions
an

the

(n

obtain

an

is formed

observed
"

on

1) matrix.

distributions:

corrections

that

must

be added

N.

of the transformation.
a

From

singleoperator throughout

C.

(20)

-N

+
be

must

CC

MC

TM.

(21)
minimum.

to T to

zero

This

is obtained

for C

from

0.

(22)

Eq. (21) into Eq. (22) and obtain

TM)'

-MN'

MM'T'

by

dT

M{-N

cessive
suc-

expressionfor C\

least-squaressolution,CC

substitute

tains
con-

for

equation :

-1
now

on

(19)

assumption of

putting the partialderivative with respect

We

matrix

C,

are

are

This

distribution

of the successive

T, the best estimate

C
For

the

give the best estimate

TM
From

1) matrix.

"

distributions
1. If each

"

columns

n.

of the matrix

values

the definition of M

learning,we

data.

distributions

A^ represents the best estimate


N

We

the

trial

through
quantities,and n such

the successive

where

colunms

alternative

analogously by
The

to

in successive

successive
a

matrix

0.

440

READINGS

fit of this

IN

to the observed
function,p^"'(i2),

fitfor the transformation


From

7^,^

PSYCHOLOGY

MATHEMATICAL

Eq. (21)we

can

calculate the corrections that

added

are

to A^:

.708

.708

.814

.814

.814

.761

.814

.867

(.345

.239

.292

.292

.186

.186

.186

.239

.186

.133

.814

.761

.814

.867

.920

.920

.814

.814

.867)

.186

.239

.186

.133

.080

.080

.186

.186

.133)

.114

-.039

.045

.086

.108

.161

-.092

-.108

-.161
.067

-.086

.092

.114

-.067

.014

-.039

-.014

-.014

-.114

-.086

-.133

-.080

.039

-.114

.014

.086

.014

-.120

squared deviations

are

of the

-.086

-.014

.086

-.033/
.033)

givenby

(-.144
best estimate

.039

.080

.133

.120

The

least-squares

(.655.761

(-.045

The

have

data; we

T.

.144)

dispersionof the calculated from

the observed

values is

""'
=

/
In

"

matrix
The.variance-covariance
V

From

Eq. (25) we
p^{A) and VeiB):

compute

same

procedurecan

-092.

(24)

-^H-

(25)

17

'"''

1-2.91

the standard

data matrices M

\fW

V is givenby

a[p,{B)]
The

1
1

AMMr'=^\

c[pM)]

The

"

be

.092

J~~

.092

^^

11.99)

deviations

of the estimates

.04

.132.

appliedto the data from

and A'^then have either 0

of

or

on

singleanimal.
trials;
e.g.,

successive

GEORGE

^^jlOl

441

MILLER

A.

11001011

(01000110100
(O
N

(10
In order

110

to solve for T

10

determine

we

h^^'^^"^^^M, MM'

NM'=

(m(l,0)m(0,0))
The

symbol m{i,j) represents

i^j;m{i)

represents the

1, where

"

of

the number
of

number

of i; and

m(0)

invert MM'

and

occurrences

of trials. Next

is the number

of the ordered

occurrences

we

(mil,l)m(0,l)^
(_}_

NM'iMM'Y"

solve for T:

^
\

\hri{l)

pair

m{l)

(m(l,0)
m(0,0);

^)

(26)

m(0,l)\

/m(M)
m(l)

(^

m(0)

m(0,0) i
)m(l,0)

is the

Eq. (26)

transitional

would

that

result

m(0)

m(l)

be

from

expected

to

estimate

the

fM

m(0,l)

) "^^^^ "^^^^
m(0,0)
)m(l,0)

m(l,l)\

m(l,l)
"^^^^

"^^^^

m{\,G)i

m(l,0)

TM-N

'

"

m(l)

m(l)

-m(0,0)

m(l,l)

) "^^^^ "^^^^
)-m(l,l)m(0,0)

-m(l,0)
"^^^^

-m(l,0)\
'

squared

deviations

m(0)

mil)

are

given by
CC

"

'

"^^^^
V

m{\,0)i

m(l,0)
'

The

find

Eq. (21) we
/

m(0)

m{\)

'.

"

'

of the

calculate

dispersionwe

/m(l,l)

from

definition

the

probabilities.

In order

Then

m(l)

'

"

m(l)

442

READINGS

IN

MATHEMATICAL

PSYCHOLOGY

where

4"^]

m(l

[m(l,0) +

The

m(l)

m(l

M^^^i

M'm]

[m(l,l)m(l,0)']
+ [m(0,0) +

^i

'

m(l)

w(l)

w(0)

m(0)

m(l,0)1 ".^Jm(0,l)
m(0,0)1
[^(1,1)
+ m(0)
L

"

m(l)

m(l)

"^

'

^1

m(0)

m(0)

J'

dispersionis,therefore,
m(l)
\n

"

"

m(l,l)

m(l,0)
w(l)

m(l)

"

r m((
(0)
+
The

"m(0,l)m(0,0)1
m(0,l)]

variance-covariance

matrix

"

Ln

"

m(0,l)
m(0)

"

m(0,0)iy^
(27)
m(0)

Jj

is

|m(l)
F

/^-l

a\MMO

"

m(0).
and

from

this matrix

compute

we

'

a[pA(A)]

and

a^

Although these examples are worked


the same
alternatives,
procedures can be
or

with

Markov

o-[ps(5)]

m(l)

processes

defined

out

used

""VmCO)

for the Markov


with

more

than

case

two

with

(28)
two

alternatives

for

be
It should
compound responses.
of
Markov
chains
neither
are
properties
techniques will undoubtedly develop as
widely applied.

that the statistical

stressed,however,
simple nor well understood.
the Markov

process

becomes

Better
more

the explicit
Transformations. Up to this point we have made
could describe the successive
assumption that a singletransformation
alternative seof the alternative responses
or
quences
changes in the probabilities
theoretical
of responses.
This
the
assumption greatly simplifies
the data hint that it might be true.
landscape and should be made whenever
Simplicity is not, however, an intrinsic property of the behavior of living
organisms, and so we must be prepared to deal with situations that obviously
violate the assumption.
6.

Variable

GEORGE

A.

443

MILLER

is adequate means
that the
assumption that a singletransformation
transitional probabilities
fixed from the first through the last trial. Since
are
determine
the transitional probabilities
the sequences
of responses
that are
that
the
animal's
of action
course
probable or improbable, we are assuming
is
fixed
the
In
certain
or
throughout
a
experiment.
strategy
sense, therefore,
that there is no
such an
the
assumption means
learning at all;as soon
as
situation
is
encountered
for the first time, the subject adopts
experimental
the set of transitional probabilities
that will later describe the statistical
propertiesof his behavior after he has had long experiencein the situation.
The
would
be justified,
for exassumption of a singletransformation
ample,
after a long series of alternate conditioningand extinction. In this
for the reexperiment the subject is able to evolve a singletransformation
inforcement
The

conditions
animal

has

adopted

distracted

temporarily

is removed
But

of the

in most

priori reason
there

several

are

In

to

order

to

and

stable
in

situations

to

reasons

consider

The

same,

and

from

follow

situation
normal

to

Or

and

when

if

an

then
the

is
pediment
im-

singletransformation.
that are
studied experimentally there is no
a
will be adequate, and
singletransformation

illustrate what

data

expected

to

expect that it will

has

in

his return

conditions.

is involved

been

another

10 rats

the

to

assumption

show

one

assumption

consecutive

20

on

be.

in the

prepared

where

not

where

case

is wrong.

choices

of

in

Once
a

single
the
more

T-maze.

choice,and 0 represents an incorrect choice.


lA and IB the numbers
of rats making the correct
choice are
the
the
both are the same
in
the
fitted
section.
as
example
preceding

symbol

In Tables

the

extinction

of behavior

way,

be

expect that

for the

mode

some

might

transformation, Table I
assumption is correct and
we

another

represents a

correct

TABLE

Hypothetical Data

for Ten

Rats

1
on

Twenty

Trials in

T-Maze

444

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

TABLE

(Continued)

Variable Transformation

IB.

Trial
Rat

12

successive
lA

There

to be

seems

secure

no

trend
more

trials by

10

we

can

11

12

13

14

15

16

17

18

19

20

10

10

clear trend

IB

in IB

for

Po(0)

Pi(l)

Trial

pi(l)to

pi(l) is observable in
reliable estimates,we get
for

Pi(l)and Po(0)

of

the values

estimate

[m(i,j)]/m{i):
Po(0)

Pi(l)

whereas
trials,
to

in Table

pairsof

Trial

by fives

the data

From
on

increase

lA. If

we

on

group

successive
the trials

GEORGE

Comparisons
transformation
lA

and

IB

such

these

as

that

show

identical in this respect. The

are

analysis of short

constant

distributions alone, for

assumption

is

if
justified

relativelyconstant

of trials shows

sequences

of

assumption

the

by the successive

be checked

cannot

445

MILLER

A.

the

transitional

definite trend,
in lA. If the transitional frequenciesshow
as
a
frequencies,
in
the
as
IB,
assumption is not justified.
transformations.
The
face variable
question is what to do when
we
Whatever
do cannot
PQRST
we
do, the situation will not be simple.If
translated
be
into
TTTTT
plex.
do the matrix products may
get quite comIf we
could choose P, Q, R, S, T as commutative
matrices, it would be
solution for all of them; all matrices would
possibleto find a simultaneous
.

"

have

the

characteristic

same

however,

not

different characteristic roots.

but

vectors

it does

possiblein

seem

general to

choose

fortunatel
Unmutative
com-

by the data.
propertiesdemanded
If the complexity of the problem is admitted
inevitable,we can still
as
reasonable
look for a matrix function of n, T(n), that changes in some
way
successive trials.The followingargument illustrates one
approach.
on
possible
We assume
that at the beginning of the experiment the subjectsare equipped
with transitional preferencesgiven by the matrix
U. After long experience
in the situation the subjectsdevelop transitional preferencesgiven by the
matrices

matrix

the

by U are
represented
slowly strengthened.

experiment progresses the tendencies


slowly extinguished and those represented by V are
Consider the followingsequence
of equations:

where
the

V. As

with

0 "

the

T(0)

T{1)

wT{0) +

T(2)

wT(l) -^ {1

T(n)

wT{n

perseverationof the

If the extinction

Eq. (29) can

w)V

w)V

(1

1) +

tendencies

represents the abilityto adopt the


the old pattern

(1

rationale for this set of

1. The

"

the

on

mode

new

(29)

w)V,

equations is that w represents


w)
preceding trial,and (1
of response
symbolized by V.
"

of the old pattern of responses

extinguishesrapidly,w
be written

in terms

T{0)

T{\)

wU

T[2)

w'U

+{l

T{n)

vfU

-f- (1

is

{\

is

near

unity; if

V:

w\U

V)+V

w\U

V) +

V.

w')V

w\U

V)

V.

w'')V

w\V

w)V

zero.

of U and

near

is slow,

F) -h F.

(30)

446

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

creases.
that, since 0 " w; " 1, T(n) approaches F as n inThe importance of U becomes
progressivelysmaller as the subject
and more
has more
experiencein the experimentalsituation. This formulation
has the advantage that it is relatively
easy to compute the successive values
of T{n), given U and V. The initial and final matrices,U and V, can be given
from data obtained
be determined
or
can
theoretically
priorto the first trial
and after the learned behavior has stabilized again in the new
of action.
course
it is clear

In this form

For

illustrative purposes,

that

assume

and

and that the weight

Then

on

have

known

are

to be

Eq. (30)gives

to be 0.8. Then

learningtrials we

successive
n:

is calculated

and

.9

.4)

.1

.6i

10---

Pa{A):

.5 .58

.644

.695

.736

.768

.796

.816

.832

.846

.857

PsiB):

.5 .52

.536

.549

.559

.567

.574

.579

.583

.587

.589

Next

we

proportionsof rightand
is given by the equation:

calculate the

trials. This

wrong

responses

no)rfo

d,

T{i)d^

d2

Ta)no)do

T{2)d,

ds

T{2)T{\)T{G)do

T{n)d,

rf",i
=

"""

"""

successive

on

(31)

n T{i)do.
n

It is assumed

n:

p{R) :

.5

.53

Considerable
errors

are

from

Assume

givesthe

mentation.
preliminary experidirect
Then
putation
com(.5,.5).

known
U and do are
T{Qi)
the boundary condition d'o

that

values

.559

.587

.614

.639

.662

.683

.700

taken

with

care

must

be

such

10
.716

"

"

"

"

"

"

"

.800

iterated computation, for the

cumulative.

It should

be noted

that if it;

0, the variable

case

reduces to the constant

MAXIMUM

THE

ON

LIKELIHOOD

ESTIMATE

MEASURE

SHANNON-WIENER

OF

A.

George

OF

THE

INFORMATION

Miller

AND

G.

William

form
limiting

The

the first two

and

of the maximum
of information

from

Also,

the

drawn
per observation
approximationsto the bias and

of the

asymptotic moments

estimate

likelihood

Madow

of the Shannon-Wiener
multinomial

a
mean

square

error

bution
sampling distriof amount

measure

distribution

are

determined.

of the estimate

given.

are

Preface
statisticdefined

The

information

in

from

drawn

event

an

(3)and by Wiener (4)to

by Shannon

multinomial

the amount

measure

distribution has been

certain aspects of stimulus

of

adoptedby

in
psychologists
response events
these
In
the
is
however,
applications,
psychologist
psychological
experiments(2).
small samplesand the samplingdistribution of the
forced to work with relatively
usually
to

some

of real interest. In the present paper

becomes

measure

measure

asymptoticdistribution
explored.
1
.

LimitingDistribution of the

The

an

Pi "

per

has
or
experiment
operation

0, i

\,

performanceof

the
results,
possible
Wiener

,k, the Shannonthis

Likelihood

Maximum

of the

moments

samplesis

Estimate

of Information

of Amount
If

the first two

the bias of the statisticfor small

derived and

are

and

or
operation

measure

/th of which

of the amount

has

ability
prob-

tion
of informa-

is

event
k

J,Pi ^og2Pi.

likelihood
estimate
of the maximum
properties
H' of H obtained from
n independent
performancesof the operation.Since if is a
values of the probcontinuous and diff'erentiablefunction of/?,,
abilitie
/(^ for all positive
We

propose

to

consider

the

it follows that the maximum

likelihood estimate,H', is

where

n^ is the
This

Research

with
frequency

article is from

the

Center, Air Research

AFCRC-TR-54-75,

contract

AF

which

"

the /th of the k

outcomes
possible

occurs

in the

Laboratory,Air Force Cambridge


Operational
Applications
and DevelopmentCommand,
BoilingAir Force Base, 1954,
18(600)-322. Reprintedwith permission.
448

GEORGE

performances,and
will

show:

now

distribution; and

H'

pi

one

be

to

Ijk,

or

sponding
corre-

0.
all

not

are

I,

define the

of /, we

values

more

equal,then

k, then

has

H'
has

H'

normal

limiting

chi-squarelimiting

freedom.

degreesof

"

obtain

we
preliminary,

0 for

(a) If the pi

if

(b)
k

distribution with
As

if "j

where,

449

MADOW

G.

WILLIAM

AND

MILLER

(njn) logg",/"of

terms

We

A.

in

H'

"

further

the
simplifies

that

form

calculations.
The

Lemma.

H
difference

is given by the

H'

"

equations:
following

Let

.Ci

"

npi

and

K=%^\j^
-PiJH^Pi.

(1)

Then
-H'

where

0, i

"

p^

vanish,but

k.

\,

that

"i

n j

to do

0 but

is

ni "

i\n

follows

as

0 are

n-

^"^2Pi

Suppose, for example,


and

definitions of H

the

from

Pi
i

stated.

as

-^Pi\og^Pi +2
=

definedto

Then

themselves

are

logg/?,.

H-H'

"

expand H'

can

the eff'ectsof rt^

otherwise.

ni

-'ly^-Pi)lOgaA

npi

verifythat

we

i^i

10g2TiZ

have

0 stillyield "pi

npi

-17

need

Vn

in C/" that

Terms

By simplesubstitutions

we

in Vn that have

terms

Proof.

All

Un

H'

have

we

"

-floga-:
2

"

"

and

Un

=1
i

=2

Tioga"
=

npi

rloga- -I
=

"

"

-^og^Pi
=

"

77-

J^n

=2
i

so

that

if

we

Theorem
a.

with

mean

the

combine

Ifthe
0 and

values

of Un

-10g2/'i +^.
=

"

and

V^,

we

verifythat

H'

Un

V^.

1.

Pi

are

not

all

equal,then Vn(H

variance

H')

has

normal

distribution
limiting

450

b.

IfPi

distribution with k

"

the

precedinglemma,

problem
equivalent

^ninjn

variables

distribution with

\,

,k

"

with

pd,

I,

Therefore,

^nVn

has

(k

a.

be

replacedby

the

the

V" F"

has

proof: The

normal
random

normal
l)-variatelimiting

"

of the random

linear combination

^nEj^ I-

distribution with

normal
limiting

VnEVr,

variables.

value

mean

pA loga/?,

^n^\og2PiE\-^-pi) =0,

and

sketch

"

the

1
values 0, variances p^qi, (qi
pi),and covariances -piPj,
mean
the
\, (i 7^ p. Since
log pi are constant weightsappliedto these

variables,it is clear that Vn F" is a

random

a^. We

\, have

of

Because

all equalthen

not

are

variance

,k

cally
asymptoti-

^nVn.

pi

are

almost

of the calculations

most

0 and

mean

estimates

since

asymptoticmoments.
H') can
evaluating V"(//

if the

that

estimates

for the

evaluating"^nUn

of

distribution
limiting

of

problem

chi-square
limiting

likelihood

partsof the theorem

case

any

has

likelihood

Also, maximum

both

first note

We

Proof.

in

H')

"

for maximum

1 holds

Theorem

needed

be

would

then

,k

will prove

efficient. We

i,j

[1],p. 500).
(e.g.
exception

without

made

\,

part of

first

The

e) {H
(2n/log2
degreesoffreedom.

Xjk,i

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

variance

ff2

-/'Jloga/'^
t/"\^

Var

2(log2/'^)'Var

+
n

i^Og2Pj)PiPj
2 i^O^iPifPi^i 2.(l0g2/'t)

Pi

(Piloga/?,)
0og2/'i)^ 2 (Piloga/^i)
-

lPi(^og,Pi)'-fJ'

1
We

that

Vn\i-p,

Cow
^Oog2 Pi)(iog^Pj)
i^j

i\og.Pi +Hf

Pi

next

2nUjlog2 e
Let

us

that

show
has

"f/" converges

in

to
probability

chi-squarelimitingdistribution with

define

"

=
,

nPi

1
,

K.
,

zero

"

as

increases, and

degreesof

freedom.

GEORGE

A.

MILLER

AND

WILLIAM

G.

451

MADOW

Then

if 1

npi

2 Ml

and

1, it follows

since /7j "


A.2*

Lemma

and

that

Xi "

^t)log2(1

.-c.^=

(2)
1.

"

Hence

we

apply

can

obtain

we

Un

l/np^"

1 +

"

^i),

"i

log:

4r i-'^y i^i

"Pi

=1

np"i

Pi
f^P

K^

1) \

^;+i

(3)

"P

where

j+1

"/'i

\ni

Pi

""^"^^-JiyxTn)
"/'i

npi\J+^

(npiY

ijXj+l)

since

Furthermore,

2
i

("t

npi)

0,

have

we

(4)

log2^

'2v{v

(npi)

l)i^i

where
*

J^'Ui"l

1) iti

y(y +

It follows
rtj

in the

from

(2) that

do

we

|",

not

"/7,|^+l

(npiY-^

need

of terms

specialtreatment

any

(3), since

the

with

of "j

a
as
approximations to t/" yieldedby
appearance
of (2) to vanish
the corresponding term
when
of terms
it possiblefor the re0 has made
elimination
mainder
involving",
did not requirerif logg"; to vanish
when
to be bounded, for if we
0,
"j

will automaticallycause
multiplier
Hi

"

The

0.

terms

it would

follow

that

Furthermore,

there

from

would

be

that
positive
probability

and

B.l

Lemma

C.2

Lemma

it

1
"
Pr(R';+,

e)=0

be

can

would

H'

seen

be

minate.
indeter-

that

1
^

,2;-2-j-l /

\ fjj-3

and

ER''

where

O,

rj

Actually,it is

easy to

(-iv+^
^i+l

from

see

4, (",

(y + i)/\=i
The

Appendix

A.

letter "A"

in

(4) that

we

"/;,v+i

itjis odd

"

ifyis even.

have

(-1P+2

(".

(j + 2)(j+l)if,

(my
"Lemma

"

A.2"

indicates

that

this

np,y+^

{npd^
lemma

will

be

found

in

452

and

MATHEMATICAL

IN

READINGS

PSYCHOLOGY

symbolically,

hence,

^;^:=o(jzj)
o(i)o(;l5)=o(;lj).
Eq. (5) shows that
unnecessarily
large,but
0

converge

fast

as

that
0(l/"-'"^)

the upper

bound

of

the

device

is sufficient to

Thus

to

(5)

above

have

we

that

prove

found

is
R'j^^
ER'I^^

for

and
R'^+i

as

(npiy

i=i

and

^it'i (my
2n
of

the first term

Now

Un
is

logge

^
i

which

to

whereas

all other

C.2.

Lemma

have

2n

Hence

npif

f^Pi

I degreesof
distribution with k
chi-square
limiting
in probability
to zero
of 2" t/"/log2
e converge
by
distribution.
On
the other
UJlog2 e has a limiting
chi-square

is well known

freedom,

(n;

"

terms

hand, since

\log2^/\2V"/
productof

is the

that ^

0, it follows

to

converges

Thus, if the/7j
are

other

in

converges

the

probability

to

has

to
probability

H')

"

VnVn

^n{H

Hence

0.

^nU"

is the

of

sum

the

whereas
distribution,
limiting

normal

zero.

hand, if the pi

are

distribution
limiting

same

degreesof

in

variable that

H') has the

"

same

limiting

^nV^.

as

the other

On

equal,Vn{H

all

distribution by
limiting

U" converges

of which

one

distribution

has

not

variables,

random

two

variable that has

random

as

e)](H
[2/;/(log2
with
e)]U", namely, chi-square
[2"/(log2

all equal,
then

K"

0 and

H')

"

"

freedom.
2.

The

Limiting First

of H

Moment

"

H'

By(l)
EH'

Since
we

EVn

now

0, it follows

that

"EUn

EUn

EV^.

is the bias of H'.

In order

to

evaluate

this bias

approximate "(/".
From

u"
_

loga^

have

(4) we

J_ Y
2n

"^"/

i^i

"/^^)"
_

npi

J- y
^n

^"'

f^^
1

"^

I2n

"P'^^

(PPif

^ (n,-

iti

np,f

inp.f

1
20/7

^(n,-np,f1
{i^

{np,f

"'
n

GEORGE

From

B.l

Lemma

J_ y

EUn
^

we

A.

WILLIAM

AND

MILLER

J_ y npiqlqi p,)

J_

453

MADOW

that

see

npiqi

G.

|.lOn^pfqfiqiPi) +

J_

+ npiqlX
3"^/?|^|

npM^J

-Pi)0-

^piq,)

12/;,^^) ^

A:-l

or,

\qMi-pd

combining terms,

v!?i

have

we

k-l

EU^

-/'f

^1

estimate

an

and

estimate

an

of H

of H

that is unbiased

J_

to

^
^zj'

Thus,

have

we

Theorem

^og2g

e)(k
(iogg

"

l)/2",

is
1/rt^

^og2g

^
V

+(i"g^^"^;r-T2;;^ i2;;^,?,^'
^

+,

theorem
proved the following
Under

2.

1 /" is H'

of order

terms

/1\

of order

to terms

is unbiased

that

_/l
_^

Hence,

^/1\

the stated

[k

conditions

-\

1\

/I

7/-^i/'=log,.(^^-"
,+^,I-J+0(;^
Furthermore, ifwe

let
k
H"

and

=H'

+{\og^e)"^

let
H'"

f/"_!^S2f l^g2fy 1
,

if

EH'

0(l/"),

EH"

0{\ln%

EH'"

0{\ln%

and
H
Theorem

H"

has

2 enables

us

n"^,namely, (/c

of order

bias of lower

are
{(^Ijpd l]/12rt^
"

of order
of course,

""^ for all


so

that

make

order
both

than

or

H'

not

depend

for all values

on

EH"

of the /?,. (Terms

may

be

greater than

about

the bias:

(1)the

term

the

of the

H'
quantities,
positive

possiblevalues

EH"

several observations

does
l)/2rt,

"

to

probabilities
/?, and hence
Since
{k
\)j2nand
/?,;.(2)

is biased

"

downward

of

higherorder

for small

to

even

may

values

be
of

terms

negative,

n.) (3) An

454

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

TABLE

when

Case

increase
omits

in bias

pi

results

H', H", and

of the Estimators

ExpectedValues
=

and

0.50

if one

(4) When
(L\jpi)ll2n^.

uses

when

[(k

pi

all the pi

"

are

/k

equal,H"

for the

H'

Binomial

Sample Sizesupto

1/12"^

l)/2"]

"

for

0.05

as

an

overall

20

correction

and

becomes

k^ -l\

-I

^'4-log,e^^^+-^^j,
which

is

lower

In order
case,

we

state

illustrate the

to

the

to

terms

Ifk

of order
=

(that is

use

of the

say, if the pi

to

bias

corrections

are

of

" k^).
unequal,21//?^
Theorem
2 for a simple

:
following

Corollary.
H

for H'"

bound

2 and

For

the binomial

case,

2,

we

estimates
obtain the following

"~^:
Pi

0.5, then

//"'=//-+

e)(i+^).

(log,

of

456

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

Hence

B. 1 it is clear that
1

Ei,ni-np,r
(7)

to

all terms

retain

consider

need

we

shall want

We

(-ir

+ /f
10g2/7,:

logPi

Pi

logaA

onlyterms

^n^pW

npiqiiqi p,)

to

4.

Hence,

"/^/^Xl 6/?,^,)

lower.

(7) in order (!/")or

of

we

From

Lemma

beginby evaluating

XOn^plq^qjpd

_
_

'~

2
omit

the

where

we

order

l/"^.Then

second

in (7) we
By substituting

of "j since it will

yielda

term

of

+4-(^^^-^^^+4^-

obtain

P. (log./.,
,.2
+

"="?

12

of the fifth moment

term

^3^

n^pl

npi

")

g;;

^4iaog.,,^)+|-i!2iii^i"^,
+

(8)

since

2 A0og2/7i
i
c.

first three

the

details

(nj

are

npif

(";"
-

6jfi

npi

very tedious,

they have

"/?,)'^
J_

(tti
-

(^/^i)'

(/7/7,)2
been

put

in

npif

Appendix

D.

Here

result.
Theorem

3.

terms
Including

nUr,^

k^-\

Vlogae/
and

?7"givenby (4)will be used,namely,

approximationto

2ifi

loga^
the

of the

terms

nUn

Since

=0.

+if)

of EU^

Evaluation
The

ifk

=2,

p^

=}/2

we

of order \jn we
Ik-WjL

have
\

ii^pi

12"

have

"(:^T-?+'
logaej

An

91c'

10k
Un

we

state

GEORGE

Finally,from

(6),(8), and
4.

Theorem

E{H

MILLER

AND

Theorem

Hf

114,

(log2e)2

/I

equal,then

are

H'f

1)
(log2e)2(A:2
^'

i\og,ef

E{H

^^i(log2/7,
H)

i\og,e)W
21ogae|^log2/7,
(lop^eflk

the Pi

457

MADOW

obtain

ifall

G.

In (general
6"

but

WILLIAM

3, we

=\tpiaog2/',

H'f

A.

'2

^
=

.T

("7^

ll)'^'

(logse)^
Furthermore,

and

ifk=2

p^

Theorem

For

any

random

have

we

variable

H'

3(log2e)7"+ 1\

where

o\' is

the

variance

{EH'

"

is

H)

the

H'f

where

E(H

the

mean

E(H'

Theorem

given by

2,

By
Then

we

way
have

mean

the

H.

Hf

EH'f.

we

error

E{H

square

error

approximate a^- by using

can

4.

\/

1
+

1 2"^

4n^

H'f

"

is the

of illustration,consider

3
2

(log2ef

square

estimatingH, the

about

H'

of H', i.e.

H'f

of

error

square

g\, + {EH'

'{k -if

a^,

mean

have

we

a\.
Since

^/l
,

approximated

E(H

3^, //zen

In

will

binomial

Theorem

4.

For

quantity.

where

case

from

fundamental

more

the

^'f
1 p

obtained

be

0,

2 and

p^

0.5.

approximation
(logaef/n

(loggef

1\

1\

(n +

{^og^/n

^'

4/72

Appendix
We

(1

begin with

an

A.

4"2

Expansion for (1

An

expansionof log (1

2n^

a;)log (1 + x)

and

x)

(-1)^-1

then

derive

the

expansion

of

x)\og{\ +x).
Lemma

A. 1.

"1

Let

log (1

^)

Then

"Xq"x.

a;

"

"

"

i?,.+i,

(A.l)

458

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

where

and hence, ifx^ "

while,ifXq

"

0,

1^13+1

l","l
Proof.

and, if we

If

1 "

expand 1/(1 + t),we

Hxq

(A.l) and (A.2) hold.


"

Lemma

=2

Then

(A.4) follows from


A.2.

(A.4)

obtain

0 and

(-^)

x, then

logo +:r)
Thus

"

Let

(-1)^-1

the fact that

c//.

"

'

the fact that

1/(1 + ?) "

1/(1 + t) " 1/(1+ x^)

0 if a;,,"

0.

Then

"x.

(1 +x)\og{\

"

Jo

(A.3) follows from

1 "Xq

"

+(-iy

-vj^7-"^

+x)=x

R]+v

^-"dudt,
) 1

fl"c/hence, ifx^ "

(A.6)

0, ?Ae"

w/f//^,
//"Q " 0, then

Proof.

If

"

1 "x,

then

1 +

Jo
and

I +t

also,integrating
by parts,

I Y^t

"^^

so

^
^
dt

^^^ "^ ^^^"^^^

"^

'^^"
"

f'^"^
^^

"^

'^'^^'

that

(1
From

Lemma

a;)log(1

a;)

A.l, it follows that if a;

Jo

logd

+t)dt=2

log(1

t)dt.

"1, then

"

7-^-+
0
1"
-

i=2

(-1)'

jo

(A.5)

IJ'

""dudt,

jo

1 +

"

GEORGE

and

hence

and

(A.5)

A.

MILLER

(A. 6) hold.

AND

If a^o "

WILLIAM

G.

then, from

that

so

(A.7) is proved.

times
n

and

Then

(A.8)

follows

B.

in

+ 1)
^o)/'0-

similar

Multinomial

let n^ be the number

of occurrences

fashion

from

(A.4).

Moments

operationhaving k possibleoutcomes

an

that

Appendix
Let

it follows

(A.3)

~Jo(1 +^o)/^(1 +

'^'+''

459

MADOW

be

independently
performed n

of the /th of the

in the

possibleoutcomes

performances.Then,
n\
\

nJ
"i! "2!
""""*!

is the
"j.

]Pl"'P2"^~---Plc'

of obtainingany specified
values
probability

^ 0,/?!+
+/?fc
in
possibleoutcomes
it
is
Then,
possible,by

n,pi

"

"

"

1, and/j^ is the

/th of the

of "i,

"

"

W/, where

"

probabiHty of the

each

operation,i

easy

but

\,-

"

"

"""

of the

occurrence

,k.

calculations

tedious

"!

to

prove

the

following

lemma.
Lemma

The

B.L

Eni

six
first

of n^

moments

are

given by

the

following
equations.

npi

Eirii

np^)^

npiqi

Eiiti

np,f

npiqlqi p,)

Eiiii
-

np^^

^n^p\qi+ np^qlX

E{ni

np^^

lOn^plqliqip,) + npiqi(qip,)(\np^q,)

np,f

+ \20pfqf].
+ "/7,^,[l 30/?,^,
\5nYiq\+ 5n^pfqf[5 26/7,^,]

E{ni

In

(where q^

p^)

epiq,)

general,
ifm

is

then
integer,

an

E(ni

npif^

Oin"^)

and

E(ni
The

proof of
We

Lemma

not

is omitted.

B.l

shall need

0("").
npi)'^'^^^

only the

of Wj about

moments

its

but

mean

also certain

of the

productmoments
E(ni
The

followinglemma

from

the

Lemma

fact that

npjf.

conditional

needed

will

moments

Its usefulness

moments.

be

obtainable

results

easilyfrom

B.l.
Lemma

E{x'

helpfulin derivingthese

will be
the

npi^inj

ExJ
Proof.

B.2.

Let

be

random

variable

and

let A' be

'

y
^

In

x'

E{\E{x'

IA')

is any

random

Ex'Y'^'EiVx'

random

event.

Then

E{x' I A')Y 1A')}. (B.l)

^,

if u
general,

Eu

E{E{u

variable,then

IA')}

(B.2)

460

where

\s

A'

random

generalformula

and

event

put

we

"

Cx'

Ex'y

a!

Since

E[{x'

E[{x'

Let

us

("

a)!

A')
[E{x' I
'

will

apply(B.l) for

IA']

Ex'Y

for

t-

E{x' I ^')f

Ex'Y-'^ix'

"

\A')-

E{x'

Ex'f IA']

[E(x' IA')

Ex'f IA'\

[E(x' IA')

evaluate

now

the random

event

E[{nj

will be

IVe

in

Ex'

"

Ex'Y''',

(B.l).

1, 2, 3 in the

Lemma,
following

we

write

now

(B.3)

Ex'

Ex'f

E{[x'

Ex'f

3[E(x' \A')

E{[x'

for

jointmoments

some

B.3.

Lemma

i*

x'

{E{x' \A')

1, 2, 3.

I^']

Ex')

we

Ex']'-''\A'}

for
by substituting

follows

Lemma

E[(x'

"

\A')

E{[E{x'

E[{x'

that

note

since

Then,

out

To applyto
expectation.

conditional

denotes
and

Ex'y

"

^
"

^^0

the

"

I''

"

the

(x'

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

""j

has

assume

Ex']E{[x'

multinomial

E(x' \A')f \A'}

IA')f IA'}.

E(x'

(B.4)

(B.5)
In all cases,

distribution.

value."
specified

population and

multinomial

\A')f \A'}

E{x'

suppose

Then

j^j.

Pi

E(nj I",)

npj)I"J

(Hi

npj

npi)

Pi

("i

np,)

Hi

E\{nj

np,f

np,fI",] =-\{ni-

"[("," npjfI"J
-

np,)+

Hi

(",

-^

Hi

Hi

npif

(",;
-

Pi

PfJHi Pi)
^
3

"

.2
np,f

{rii
-

/^|(^"Pi)t
-

+,

3"

^Pi^Hi Pi)iHi ^Pi)

PiiHi Pi)(Hi ^Pi)

("i

^
H
f^Pi)

size of

sample is "

",)

"

np.
Hi

Pi

{rii

Hi

Hi

Pi

n,)

(rt^

npj

Hi

npj)I",] =(n

"j and

the conditional

Pi

Pi

"

is pjiqi.Hence

outcome
/th possible

E[inj

"

Hi

If ", is fixed the conditional

E(nj ni) ={n

Hi

of the
probability

np,)

Hi

Hi

Proof.

(","

npi).

npi).

GEORGE

A.

MILLER

WILLIAM

AND

G.

461

MADOW

Also

E{nj I",)

Erij

{tii

np,).

1i
Hence

E[(n,

^'^^' ^'''

npjfIAz,]
=

("^

("t

"

Wif

("

"^")

"

"Pi)

"

rii

nqi

(jti

"

"

("i

"ji
since

"Pi) +

qi

qi

npi).Finally,

qi

qi

qi

.Pj(qi-Pi\(qi-^p^

,r

("i

"Pi)

("i

qi

pfiqi Pj)

Pl
=

qi \
3

"PiT

qi

pfiqt-Po)
("^

3rt

"Pi)

qi

pAqi

Po)(qi ^Pi)

(rtj

npi)

npAq-; Pj)(qi ^pi)


-

+
Hence

E(ni

npi)(nj np^)

E{ni

np,)

qi
Pnpiqt

E(ni

-npjCii

npifitij npjf
-

'P'
np,f % (Hi
J

Eiiti

pMi -Pi)

,2
npiY

qt

"PMi -Pi)
.

("i

"Pi)+
^

qi

qi

qi

pMi

Pi)

"Piqiiqi Pi)

"

"Pi^qi~Pi)
,

"Piqi

"

qi

qi

npip]iy epiqd
-

^"Yiq! +

npipj

(qi-pi)iqi-pi)
qi

(-"i
-

V
2
i ^j

^J,

"^PipAqi Pi)'
-

"Pif("j "Pi)^
-

-^i^rz
"
PiPi

+ (qi
2 n[^i/'i^r
-

M
Pi)^

i "=j

2
~jL

"/!

"qi

^Piqi) (qi Pi)(qi Pi)


-

"qi

462

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

Now

Pj

Pi

2 ^"Ji Po)

qi,

(^

=(k

l)^i-qi

l)qi.

Hence

np,)\nj npjf

Xrii

IE'

n'^PiPi

i"=j

^lPiqi +(^

Appendix

Order

C.

3piqi+ik

-1X^-2)

of Convergence

epiqi

(k

2)(qi Pi)
-

2)qi +

k-62piqi-(k-2f

+-

and

Convergence

in

Probability

If
lim
W"

is

bounded,

we

say

that/(")is

n"^f{n)

"-oo

at most

of order

l/w" and

nj(n)

0,

write

If
lim

n"i-co

we

say

that/(")is of

sequence
for every " " 0, we

lower

of

order

random

and
l/""'

than

variables

u^, u^-

PKI"nl

"

"

becomes

"

in

converges

to
probability

if,

If la.

"

e)

0.

fi,then

converges

in

to
probability

as

infinite.
and
simplemanipulations

Using some

Proof.

the

we
Tchebycheffinequality,

have

PH^^i-=^".|=P.
so

npj)^

(rtj
C.l.

"

have
lim

Lemma

write

that convergence

in

"

el/^"("'/"-l"

if

occurs
probability

2a
1 "0,

if 2a
i.e.,

"

Lemma

/3.
C.l.

IfloL" ^, then
l"i

converges in

to
probability

as

becomes

npi

infinite.

Piqi

"^2/^"(2a/^)-l

464

READINGS

Now

the first term

since

that in

it follows
terms

that

From

Lemma

E[{nh

B.3

obtainingthe expectedvalues
d

-^

("i

np,Y

this in the

("j

qh
so

of

of the terms

shall do

We

^2.

(6) we

can

ignoreall

followingevaluations.

have

we

npnf \ni\

PSYCHOLOGY

of, say,

0(l/"^)where

are

MATHEMATICAL

IN

np,)+

qi

qt

that

nyipuqi

nyiPhqt
+

^("i

2"

-T^ V"nYiq1+

"Pi)

npiqlX

epiq,)-]

P^Hi

"

+
/'.O]
["/'^"9^"(^^

"/'^^^

"

i"

2-

""

and

hence,

"^

^A2W'z2

H
3/?^^^"

V (K

Ijq^.

Also,

'^

PiPhq

"
I

PiPh qi

nPhiRi-Ph)
+

iqi-Pn)

"
""

"

"Pi)'

.]

"2"2^

"

^("t

npiqlqi p"),

where

2^2 +

"2,
^3^
n^^iPiPh

stands

for

that
quantities

I0p^(qi-pi)
""^

will

Oiljn^)or higher.Hence,

yieldterms

3(qi-p^)

(qi Pn)iqi Pi)


-

'^

npi

GEORGE

A.

WILLIAM

AND

MILLER

465

MADOW

G.

and
E

"^m^i^

3ik

lOqiiqi-pi)
;;
n

(k

2)qi

2)qi{qi pd
-

+
n

npi

Also,

Pl
Ew^^Wii

Phiqi -Ph)
E{ni

=
o

npif

'^'''n^plPuq'
3^

.3

E(ni

Ph
"

"

npd

nPiPnqi

[l5n-pU +

"""]-

%:^

["""]+

npiY

%J;^ \ln-pU
ny-qi

f^pm

nyiqi

Eirii

r^

npi

Hence,

\5q1

3(k

"

l)q\

Wi

Finally,
i^h
Ew,,-"Wi^

Whfi^i

Wi^^

and

First,we

that

note

Ewf^

3qt +

npi
and

that
1
E

Z
h

^m^i2

^Piqi+ik

2)qi +

6piqi-{k

-2) (qi

Pi)
:

"

"

"]

466

MATHEMATICAL

IN

READINGS

PSYCHOLOGY

that

so

32 ^1 +

3(^

1)

32 ^1 +

(A:

2)(A:

1)

1
+

62^1

^i

;=1

ir

{k

\)ik

1)

1)

6(/c

62 9l

i=l

2)^

(A:

+k

-6{k

-If

-{k

-\)

n\_i iPi
=

k^ -\

k^ -2k

+2

n\_i=iP
for
Similarly,

the second

term,

^12^13

^m^iS

npi

-pi) -(k
lOqiiqi

U=i

(,^",

nPi

^^ptq,

iiAnpi

2
=

-/'D =2
-f-(^f
fiPi
i

(^^-p^)"

"

^/'i

term,

k
"

/15^

^hz^i^

^ii

ii,,^s,_3

i=l

^1

term,

U'

the final

"/'f

+ 2
2 M^fa

For

2)qi{4pi qj)

(,^B),|-(4.+2),.,,

^1
the third

npi

_y

For

2)qlqi p,)

1=1

i=

(k

Kk-2)qi

^^qlqi-Pi)

I0qi(qi-pi)

^ii

2
Z^h2-

^ii

\5qf I5q!

L "Pi

ijth

^'[%

2
i

2)ql

"

f^Pi
^-

0^2

k
=

3(k
,

Wi

5a

Kk

2)]

=2
1

^^2

^(3^ !)"
-

"/'i

GEORGE

A.

MILLER

AND

WILLIAM

k^ -2k

G.

2
6"

A: +2

15

(/c

1) +

36//
4

1^1
+

7-"

A: +

5(A: + 2){k

6/2

6/2

_"_

l^^l3(3^-1)

/?,"

z-

6"

i=i/7i

^1

^/7/

12/2

12/2iri/?i

1 2//

1^^

A:+8^1

1)

rfi

fC

1)
j

77
4/2

4"i'^i/?j

+
12//

~k^

-2k

1)

36/2

/7j

3(3A:
(^

"

^
/?i

1=1

15
^

"

6"

yc2-l

(k +8)

"

ii=iPi

467

MADOW

Finally,
A:2

/2t/"\2

"L-^

="

logae/

/ 4^

1
V

+Pr

4^

(3 -2^-16+9^+2)

'

12//\,fi/?

-3k^

2^2

5A: +

5 +

[6 -ek

igyr. ^

jQyr,2

12/2

lOA:

A:2

check, E\

r-

18A:]

12/2

iPi

computed

was

"

5A:

9A:2-20A:+7

"

-^-

12/2

As

7^-11^

20

for

specialcase.

Let

=2so

that

\log2ej
n-i^+

n2

np2

"

n^

"

n^

n,

til

"

np^

ii^,

"

-("i

"jPi).

"

Hence,

l(/2i-np^fl

nU^
_

log2e

/2

^2

l/^i ^1

1
l(/2i -/2/;i)7
6

//^

Z'?
1

(/2i-/2/7i)V1

l;,3^3
_

12

/23

468

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

and
1

("!

np^'^

("i

np-^y

nY^q\

'^'

77*/7^^f

36
1

("i

"Tl^l

4:j-(^i+/'i)-

"ViM

12
_...2"2.2

6nY^q\
l5nY,ql{cjl+p'^
nnYq\

6/^1^1 5{q^-p^f

[3

1 /

log,e/ ^4
3

qi

20{q^ p^f
-

15(9^+ p%

15

"^T

3^V

/12

""4
Pi

4"/7i^l

that

1/2,so

2 and

3"/7i^l

\Sp,q,+ 5iq, p,f

^3

12/7/7i^i

12"/7i^i

^1

Let/?!

5(ql+ pD

p,f

5(q,
_

4/7/7i^i

If k

j"(.qi -pi)

"3-3

np^)

".^!^f

(/7i /7/?i)

iqi-pxf^^'

18

15

3^

An

'

1/2,

1/

4+"ll-4

logse

(from square of

1st

term)

of 2nd

(from square

term)

(from productof

1st

by 2nd)

(from product of

1st

by 3rd).

5
4n
From

generalformula
3
1st term

2nd

term

squared

15/

1/4

15

15

15

36

ITr

1/2

36n

36"

36/7

10

product of

1st

product of

1st

by

2nd

by

3rd

6"

6n

each

part checks.

1\

term

12"

5x4

3x5/1

Hence

4+4^^4-

squared

\2

^2/

15
_

5
_

^\2n^4n

GEORGE

A.

MILLER

WILLIAM

AND

G.

469

MADOW

REFERENCES

[1]

Mathematical

H.

Cramer,

methods

of

Princeton:

statistics.

Princeton

Press,

University

1946.

[2]

Miller,

[3]

Shannon,
27,

[4]

G.

What

A.
C.

E.

is
A

information

mathematical

Amer.

measurement?

theory

of

communication.

379-423.

Wiener,

N.

Cybernetics.

New

York:

Wiley,

1948.

Psychologist,
Bell

1953,

System

8,
Tech.

3-11.

J.,

1948,

OF

DESCRIPTION

STATISTICAL

George

A.

Miller

OF

INSTITUTE

LEARNING*

J. McGill

William

and

MASSACHUSETTS

VERBAL

TECHNOLOGY

of a probability model.
verbal
Free-recall
learning is analyzed in terms
that the probability of recalling a word
on
general theory assumes
any
of times
the word
has been
determined
trial is completely
by the number
of this general theory are
recalled
particular cases
on
previous trials. Three
the
three
examined.
In these
placed upon
specific restrictions
are
cases,
of previous recalls.
The
relation
between
probability of recall and number
data
is illustrated.
to typical experimental
application of these special cases
of set theory is suggested but is not
in terms
An
interpretation of the model
The

essential

The

the

to

argument.

learning considered

verbal

end

list of words

following experiment:

procedure is repeated through


to extend

prepared

not

series of

presented

all the

statistical

the

is

down

presentation he writes

of the

is the

in this paper

theory

to

he

At

in the

learner.

the present
of

range

the

At

This

remember.

can

wider

observed

the

to

words

trials.

kind

time

we

are

experimental

procedures.
The
We
has

shall

learned

been

been

recalled

word

will be

it has

Then

Tk

all the
times

1"

word

shall say

that

has

previous

so

function

research

seminar

Social

is to
the

from

so

state

remains

This

different

article

has

probability that

of times

k, the number

previous recalls be symbolized

by

their

meanings

are

recall

exactly k times
Thus

Ak

before

first trial

Aq

is

preceding
first trial

the

they have

say,

the

on

word

the

to

proportion

On

tq

The

Ai

state

^o

recalled

been

second

the

zero

of these

proportion
trial the

"

by the

Science

authors'

Research

Behavior

many

word

in

in state

College, June
Theory, held at Tufts
for advice
especially grateful to Dr. F. Mosteller

for

the

hsted

is in state

; that

passes

facilitated

was

of the

of

and

recalled

Ideally, on

and

recalled

is not

*Thi8

and

words, the

material

test

of times

probability of failing to

been

Aq

trials.

in the

paper.)

word

the

in state

are

is recalled
To

Summer

words
on

words

(Symbols

previously.
of the

1 is

trials,we

word

any

other

In

corresponding

the
When

Tk

"

which

to

probability of recall after

trial

on

end

the

at

degree

Model

completely specii"edby the number

recalled

the

preceding trials.

on

the

Let

is

recalled

been

Appendix

that

assume

General

membership

Council,

entitled

28-August
and

in

the

24, 1951.

criticism

Inter-University

Mathematical
that

The

Models
authors

proved helpful

are
on

occasions.

appeared

in

Psychometrika,

1952, 17, 369-396.


470

Reprinted

with

permission.

472

IN

READINGS

The

generalsolution
p{Aq

of

(1

PSYCHOLOGY

all the

(1)when

n)

MATHEMATICAL

for

To)",

"

n)

ToTi

"

"

T,_i

"""

A;

denominator

differences

of each

(r,

of the form

The

of the

fractions

of times

expected
including trial n, is,by definition,
E(k,n)

in the

for the

r.) except

"

number

word

(2)

0.

r.)

The

for A; "

n (r,-

A) :

0,

^"^"

[^
~

p(A,

different is (seeAppendix

are

t^

summation
difference

zero

is

includes

all

(t,- t,).
told, up to and

recalled,all

"

i2kpUk,n).

(3)

A-O

The

E{k,

This

difference

Thus

-i- 1)

we

have

is the
the

p"+i

An
the

0,

E{k, n+\)

alternative

word

If these
to k
n

theoretical

1.

word

The
.

will both

have

we

That

and

score

1 is the

values

E{k, n),

difference,
trials.

successive

on

symbolize

we

product

be in ^4^ on

the total

is to say,

p{Ak

"

trial

forn

it

by

p"+i

and

(4)

0.

1 "

follows.

as

On

trial

The

n).

also be recalled
all the

over

probabilitythat

0,

probabilityof
the probability
n) is,therefore,

A^'is piA^

summed

have

we

r*

forn

be obtained

can

p"+i

is in state

joint probabilitiesare
n,

recall

expression for

probabilitythat
a

trial

on

the cumulative

general relation

recall in state A^ is r^
that

recalled

n), between

E(k,

"

Po

of words

expected proportion

trial

-4.^from

will be recalled

word

states

on

k
on

1.
0

trial

p"+i
n

Pn+i

The
follows.

np(At

From

(5) are equivalent,which


(4) together we have

(3) and

first summation

be shown

as

1)

J2 kp(At

n).

t-0

the

on

can

2 kpi^k

*-0

n+l)

(5)

n),

expressions(4) and

two

Pn+1

The

by substitutingfor p(At

be rewritten

right can

according to (1):

n-H

22 kp(A,

1)

i-O

2 kpiAk

ti

n)(l

r^) +

k-O

n)Tt-i

2 kp{Ak ,n)

t-0

A-0

2 ^(^*-i

J2 kp{Ak

n)n

t-0

Z(^+

l)p(A*,n)r.

GEORGE

A.

MILLER

this result is substituted

When

Pn+i

AND

WILLIAM

into the

473

MCGILL

expressionfor

J2 kp(Ak ,n)Tk+

J.

p"+i

have

we

J2 (k -{- l)p(Ak n)Tk


,

H TkpiAk n),

which

is the desired

be deduced

of the model

general solution

of the transitional

the

have

words

as

increases

the

with

transitional

case

zero

to

can

in which

All the words

zero.

positiveprobabilityof moving along

limit

without

(2). First consider

probabilitiesr^ is

to the first state, A^

etc.,up

There

result.

the

Ao and

in state

A2

from

more

or

asymptotic behavior

The

one

start

A^

states

probability,r^

0.

trapped; eventually all the words are recalled exactly


h times and
be recalled again. This fact can
cannot
be seen
from (2): If
all
the
in
to
then
terms
Thus p(Ak
r.)" (2) go
(1
zero
asn"^02.
n)
0,
Ti "
for
k
h.
For
k
the
in
of
summation
"
"
front
the
to
zero
product
h,
goes
are

"

must

include

(1

ThY

go

"

to

0, and

r^

zero.

lim

p(Ah

n)

recall score,

then

X)

ft

since

the

This

0.

th

asymptote
shall be

follows,we
different and

If all the

as

Tfc[limp{Ak

'

(Ta_i

"

1-

Th)

"

asymptote; from

an

n)]

not

(5),

0,

n-"m

is concentrated

is at

learning curve

the

only with

concerned

greater than

zero

'

(2) does

h,

at state

is of little interest for

case

of the

Th)

"

h, however,

of

approaches

p"+i

""'"V

"

Th){Ti

"

probabilityat the asymptote

this state

0 for i "

"

r^

When

h.

in the summation

'^"^'

n-"oo

since the

this term

0 and

lim

0 for k "

(To

p"+i

n)

so

n^co

The

p(Ak

so

0)"
(1
1, and
when
Instead,
ta

"

an

zero.

case

and

Ah

for

acquisitiontheory,
Therefore, in what
all the

in which

are

r^

zero.

probabilities
r^

transitional

are

greater than

then

zero,

from

approaches infinityall the terms in the summation


go
of the p{A^
toward
for all finite values of k. Consequently the sum
n)
zero
be made
please for any finite k by selectinga large
we
can
near
zero
as
as
(2) we

see

that

as

enough value
number
almost

of

In

n.

of recalls is

the

Since

zero.

all the probability

for the limit when

all

t^

limit, therefore, the probability of

"

comes

the
to

sum

of the

be concentrated

0,

p{A^

piAk

oo)

1.

n)

must

in state

any

finite

equal unity,

A^, and

we

have

474

READINGS

We

are

moving

able

now

to state

to show

Ak+i

that

if the
,

PSYCHOLOGY

MATHEMATICAL

IN

word

Ak has probabihty

in state

is continued

learning process

happens because almost all words eventually reach


write,for the probabilityof leaving state A^ on some

one

of

indefinitely.This
A^

state

Thus
.

we

can

trial,

00

2] Tkp(Ak ,n)

1,

n=k

or,

2 p(Ak
n

In all the
an

cases

asymptote

for

"

0.

"

Tk

Tk

We

"oo.

"

shall consider

we

A;

as

,n)

in this paper

interested

are

in

the value

of

t^

will

approach

tions
placing the followingrestric-

the Tk'.

on

lim

Tk

Tk

"

0,

Tk

Tj

"

1.

k-'co

The

first two

conditions

insure

large n.

The

condition

that

p(Ak
n) goes toward zero for finite k
provides the asymptotic value of r^ for
,

and

infinite k.
out

to

third

In the summation

and
infinity,

so

for the

limitingvalue of

all terms

p"+i

are

zero

have

we

lim

mp{Ao.

p"+i

^)

(5')

m.

n-"oo

In other
then

value

words, if we

that

assume

is also the

value

asymptotic
In the special cases
discussed
m

of

Tk

in the form

that

0 "
Tk+i

"

that
acquisition,
so

Consider

the

of the linear difference

1 and

is bounded

asymptotic value of
as

p"+i

below,

Tk+i

where

is the
of

"

"

between
r^+i

"

"

zero

t^

(XTk

a.

The

and

one

7^

as

/c ^"

oo

-^oo
,

restriction

is

placed

upon

the

equation,*

(6)

limits for

and, since

have
we

are

been

chosen

so

interested in

followingdevelopment of (5):
n+'l
Pn

+ 2

2
k

*We

have

TkP(Ak ,n-\- I),


0

tried to observe
the convention
that parameters
are
representedby Greek
statistical estimates
of a and
letters. In the case
are
represented by Roman
have violated this convention
in order to make
m, however, we
our
symbols coincide with
those used by other workers.
The
symbols m, o, a, and p were
originallyproposed by
Bush
and Mosteller.

letters and

GEORGE

substitute

Now

for

A.

p(Ak

MILLER

n+l

Pn+2

n+1

Z) np(^i

n)(l

Z) TtP(^*-i n)n-i

Ti) +

Pn+1

substitute

Pn+3

"'(7*

Ti+iTtpCAi n).
,

a)p"+i

(1 +

a)p",x

(1
(1

1) is the second

a)

S (a + a7-0Ttp(Ai n)
,

S 7-iP(^A w)
,

a)E(T^

of the

moment

raw

(7)

1),

r*

is the first

(as p"+i

for trial n+l.

moment)

raw

J2 rlp(Ak,n)+

(1 +

according to (6):

t^+i

Pn+1

where

for

n) +

we

S np(Ak

"

":

Next

475

MCGILL

J.

1) according to (1):

WILLIAM

AND

(6) brings the system into direct correspondence with a


nology,
In their termiand Hosteller.
of the theory developed by Bush
p, to give
operator Qi is applied to the probabilityof response,
A second
trial is successful.
the new
a
as
probabilitywhenever

Restriction

specialcase
an

"i

ocip

operator Q2 is applied to give aa +


the present
restriction
say,

is

az

applicationof
and

unsuccessful

upon

its

unity, so

to

probability of

simple assumption

that

assume

the

occurrence

is will be

seen

trial is unsuccessful.

In

omission

of the

trial consists

reasonable

seems

is

az

generaltheory, Qi is preservedintact by
is to
to be the identity operator. That
In the present application,an
Q2P
p.

more

Q2 is assumed

(6), but
zero

this

whenever

aaP

of the

word
of

non-occurrence

the

on

when

trial.

next

examine

we

during
a

word

recall.

has

effect

no

this

successful

How

It

the data.

Analysis of the Data


At
word

the

end

lists

the

"

of the

experiment
recalled

words

the

experimenter

by the learner

on

has

collected

successive

trials.

of

set

These

that did not

occur
of words
usually contain a small number
of
learner
the
some
are
in the presentation. These spontaneous additions by
in the present discussion.
shall ignore them
interest in themselves, but we

recall lists will

We
estimate
we

the

would
of

suppose,

like to

p"+i

to all of the

transitional

the data

contained

in the word

lists to obtain

an

There

are,
(5). We shall refer to the estimate as r"+i
material
in
learning
words
provided by the experimenter as

in

experiment.
these

use

words
words

It

seems

are

reasonable

A^

we

be considered

may

probabilityof recall,t^

that

assume

By this

homogeneous.

in state

to

imply
as

under
that

estimates

certain

ditions
con-

the responses
of the

same

476

READINGS

We

define

then

can

IN

convenient

statistic,

numbers, Xi^^.n+i,
the same
meaning

The
have

indicate

that

we

words, with

recall

occurs

various

are

zero

or

that

(8)

"

The

one.

attached

event

an

have

we

first summation
that

determine

them

occurs

zero

trial

This

summation

on

1 for any

word

only

goes

correct

in state

to

up

is trial

are

Ak when

Ak

provided

that

k,

over
our

trial

on

the

reference

rules determine

These

n.

state.

straightforward.
summing

occur

because

the A^

to

responses

one

extends

summation

of states

number

or

in each

if a recall fails to

second

The

1.

in state A^

zero

word

experimental

of words

in state

not

1 to

i, the

over

Xi^k.n+iis

an

for all words

previously. They

trial

on

out

the number

count

we

and

subscriptsk

to

is carried

whether

are
X,-,i."+i

states.

experimental

words

on

1.

To

show

that r"+i is unbiased

E(r..O

the

either

that

point for determining the


r"+i as the proportion of

The

i^i

They are zero for any word


1. Lastly the X,,i,"+iare

i.

trial

ic^o

The
on

iV

l~f

k fixed to show

rules that

The

2^ Xi^k.n

looking at

are

The

Aj,

in state

2^

AT

J
^n

PSYCHOLOGY

MATHEMATICAL

expectation of
sum

in the

any

t Z.,
[e{

"

Thus
the expectation of
Xi^^.n+iin state A^ is t^
is N-Tk-p{Ak
n). Substituting this into the ex~
,
.

brackets

we
pression for "'(r"+i),

i
=

that

observe

we

find
n

E(r"+i)

E(r"+0
The

sampling variance of
X,,i,"+iaround

of the various

Var

The

variance

The

variance

(r"^i)
=

Tkp{Ak

p"+i

(9)
p"+i

the transitional

^2

n),

around

r"+i

Var

is determined

by
probabilities,
t*

the variances
.

^J2 ^Y,.*,"+,].

of any

X,,a,"+iin state Ak'is binomial and is given by ri(l


r*).
thus
becomes
A^
Xl^-i^..it.n+i
ing
p{Ak n)Tk (1
n). Substitutthe expressionfor Var (r"+i),
obtain
we
"

of

"

this into

"

Var

It should

be noted

(r"+j)
=

S Pi^k

t;

that this variance

"t^Pn

is never

"

(1

"

w)n(l

largerthan
Pn+l),

n).
the binomial

(10)
variance

GEORGE

since the binomial


the variance

on

Var

order

In

of the

includes
around

r^

p"+i

(10) a

to

477

MCGILL

J.

in addition

^^'^
"^'^^^

WILLIAM

AND

MILLER

variance

(r",0

that

term

depends

"

""^j-

''^^^^ ""^

(100

'

obtain estimates
general theory we must
Now
the
is
probabilities,
probabilityof moving
Tk
r^

of the

trial.

After

the

apply

to

transitional

A.

from

Ak

state

trial n
some

up

obtain

we

remain

some

are
,

an

of

r^

in state Aj,
trial

^1^ on

in

trial to

from

constant

estimate

estimate

an

be

to

words, Nk,n

Ak+i provides

to

trial

of

Aft+1 and

to

on

go

every

is assumed

certain number

moves

Ak+i and

to

of

-\- 1.

that

on

r^

Of these N^.n

Call these

The

trial.

estimates

words,

fraction that

Therefore, on
Then

tk,n+i

"

is zero,
If A^fc,"
Next

we

transitional

the

estimate

no

wish

is

possible.
the

combine

to

4,n+i to obtain
The

probability,Tk

"

upon
to

are

respects the

which

For
6

example,

accuracy

the 4,7.+i

by

estimate

"

it places undue

of A^,,," We
.

of the various

This

trial 8.

on

values

This

of
,

emphasis
prefer,therefore,

estimate,

after trial 7 there may

recalled

are

small

on

the maximum-likelihood

use

10,

based

least-squares solution, obtained

of
is the direct average
r^)^,
minimizing (4,"+i
variance
because
is unbiased, but it has too large a
the 4."+i that

singleestimate,

4.n+i

"

be 10 words

gives the

estimate

in state

^3,8

A3

Of these
.

6/10. Every

final estimate
provides a similar estimate, ^s.n+i
of T3 is obtained
by weighting each of these separate estimates according to
it is based and then averaging. This procedure
the size of the sample on which
data
permit.
is repeated for all the t^ individually as far as the
dependen
basic
the
assumption that r^ is inThe
4."+i are also useful to check
trend, this basic assumption
If the 4,"+i show a significant
of n.

trial on

which

The

i^g.^5^

"

is violated.

The

computation

The
tedious
relation

as

and

among

of

k become
the

r*

Simplest Case:
p(Ak

n) from

One

(2)for

Parameter
the

of

is exceedingly

look,therefore,for
restriction (6). The first case

moderately large. We

of the form

general case

simple

that

we

478

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

shall consider is
To

Tk

In this form

the model

of the difference

,1

a,

(1

contains only the

(12)

d)Tk

The

singleparameter, a.

solution

equation(12)is
1

Tk

(1

ay.

(13)

The

follows: On
of (13)in set-theoretical terms runs
as
interpretation
the firstpresentation
of the list a random
sample of elements is conditioned
for each word.
The measure
of this sample is a, and it representsthe probability,
state
If
word
is
not
of
from
to
state
a
A(,
Ai
no
going
recalled,
To
When
a
change is produced in the proportionof conditioned elements.
random
word is recalled,
the
effect
is
condition
another
to
however,
sample
of elements,drawn
a to that
independentlyof the first sample, of measure
word.
Since some
of the elements sampled at recall will have been previously
pendence
a
fter
recall we
have (becauseof our
one
conditioned,
assumption of indebetween successive samples)
:
.

/Elements conditioned\
\

duringpresentation/

A2

we

a^

"

the firstto the second

independentrandom

elements

\
/

a)^.

recall. The

sample of

second

measure

time

is drawn

word
and

to

is recalled

conditioned,

have

[1

^2

(1

a)']+

Continuingin this way


With

-fl)

a[l

(1

in

(1

a)'.

p(Ak

generaldifference equation(1)becomes

n)(l

a)'^'
+ p(Ak-^ n)[l

The solution of this difference equationcan


outlined

a)']

generatesthe relation (13),

this substitution the

p(Ak

The

(1"

"

/ Common

Ai
quantitygivesus the transitional probability
ti of going from
from

another
so

duringthe recall

This

/Elements conditioned\

Appendix

or

by

the

be obtained

(1

a)'].

by the generalmethod
in (2).
t*

appropriatesubstitution for

solution is

p(Ao

n)

(1

p(Ak

n)

(1

a)",

ar' n [1
1

From

definition

(1

a)"-].

(14)

(5) it is possibleto obtain the followingrecursive

ex-

480

READINGS

reach

asymptote

an

introduction
with

the

of

an

PSYCHOLOGY

MATHEMATICAL

IN

somewhat

below

the

less than

asymptote

theoretical

value

unity will be discussed

unity. The

at

in connection

three-parameter case.

0=0.22

NUMBER

TRIAL

Figure

Comparison
As

shows
k

the

of Theoretical

further

and

check

Values

of

piAt

n) for the One-Parameter

Case

correspondence of theory and data, Figure 2


observed
values of p(Aa
n) as a function of n, for

on

piredictedand

Observed

the

0,1, 2, 3.
Case:

Second
In the one-parameter

Parameters.

Two

that the proportheory it is assumed


tion
the
the
list
is
of
same
as
presentation
data are
recall. Most
not adequately

of the

form

the

sampled during
proportion sampled during each
to
described by such a simple model.
At the very least,then, it is necessary
different.
In
these two sampling constants
consider the situation when
are
restriction
order to introduce
the second
(6) in the
we
phrase
parameter,
of elements

the

followingform:
To

Tt+i

Po
a

-h (1

"

a)Tk ,

(17)

GEORGE

po is the

where

MILLER

proportion of

On

the

first

conditioned

to

measure

and

T2

[1

(1

1)

,n

po)(l

solution

p(Ao

n)

p{Ak ,n)

After

d)] +

Po

(1

{l-po)

n)(l

po)(l

subject

K^.-i

for the recall

form
=

was

given.

(r"+i)

found

that

Figure 3

(1

(1

"

po)(l

po)(l

a)]

a)'-']. (19)

^^""'

a)

"

Po)[l

(seeAppendix B)

becomes

now

(1

0.10

data

are

and
of

p"+i

(p"+2

-vr

was

(21)

ar]pr.

was

these

read aloud.
he could
A

(22)

Pn+i)-

"

equations we
and

collected by Bruner

reading.

every

analysisof

the values

The

n)[l

applicationof

first set

The

before

the

From

(1

Po +

all of the words

wrote

scrambled

(14).

to

monosyllabic words

was

(1

-7T

to illustrate the

sets of data.

Ust of 32

is

a)"

^
1

Var

two

a[l

a).

of r"+i is

variance

In order

elements

11

Pn.i

The

po)(l

general difference equation (1) becomes

the

(20) reduces

recursive

The

of

poT,

a,

is

(19) is

of

(l

a-

.=0

When

one

sample of measure
po
recalled,a random
sample
recall,
therefore,the measure

a)^

+
The

(18)

generates the relation (18).

p{A^

a)\

is

of conditioned

this substitution

With

presentation.

random

word

apo

"

Po)(l

in this way

Continuing

When

measure

(1

the list

the

be written

po)(l

during

is

Po

recalls the

two

(1

conditioned.

elements

Ti

p{Ak

word.

every

of conditioned

After

can

481

MCGILL

J.

equation

presentation of

is drawn

WILLIAM

conditioned

T.

AND

elements

of this difference

solution

The

A.

At the end
The

remember.
total of 32

have

selected
A

Zimmerman.

of each
order

reading the
of the words

presentationsof

the fist

particularsubject it was
good descriptionof the data. In
tion.
(21) are shown by the solid func-

the tk calculated for this


po

0.27 gave

computed

given by

from

the open

circles.

The

dotted

lines

are

drawn

482

READINGS

"

standard

one

MATHEMATICAL

deviation from

check,Figure 4 shows
function of

IN

for A;

the

another

computed from
observed

predictedand

(22). As a further
p(Ak n) as a,

values of

0, 1,2, 3.

distribution of cumulative

The

as

p"+i

PSYCHOLOGY

recalls on

given trial providesstill

any

viewing the data. In Figure 5, the cumulative distribution


of recalls,
is shown for trials 5, 10, 15, 20. The proportion
of test words recalled k times or less is plottedfor comparison on each trial.
The second set of data was
collected by M. Levine.
He read aloud a
100-word anecdote.
At the end of the reading,the subjectwrote down
all
he could remember.
order
The
of
the
words
Four such trials were
given.
trials.
not scrambled duringthe interval between
was
From
the analysisof the data for this particular
subjectit was found
that a
of the results. Figure6
0.87 and po
0.61 gave a good description
shows the comparisonof theoryand experimentboth for p"+i and for p{Ak ,n)
of

way

of

k, the number

for A:

0, 1, 2.

have noted that when the order of the words


we
generalobservation,
is not scrambled between
the parameter a is relatively
large. This
trials,
is to say, when the words are not scrambled,there is a much higherprobability
As

that the

at

will be recalled

successive trials. This effect is related

on

to be recalled.

rate determined

by

subjectrecalls words

The

curve.
serial-position

continue
at

words

same

beginningand
the end of the list. If these words remain in their favored positions,
they

to the

words

New
po

to

those recalled at the ends

learningworks from the

the

so

added

are

at the

two

This effect has been

the

middle,which is the last to be learned.


listsof randomly selected Englishwords as well
Third Case:

as

ends toward
noted

with

with anecdotes.

Three Parameters

and two-parameter cases we have assumed that after sufficient


reach perfect
the subjectshould eventually
performance. Some data,
practice
In the

one-

to consider
to evade this simpleassumption and so it is necessary
however, seem
Such a parameter
what happens when a lower asymptote is introduced.
be necessary when, for example, the periodof time allowed for recall is
may

limited.
To introduce the third parameter
To

Tk+i

The

solution of

Po
a

adopt the generalrestriction (6)

we

-\-aTk

(23)can
r.

where

0"a"l"

a"l.

(23)

be written

j^"
1
"

(y^- Po)a\
-

\1

"

(24)

(24)reduces to (18). From (24)we see that as k increases


without limit,
asymptote. From (5')we know
t^ approachesa/(l "a) as an
When

1 "a,

GEORGE

100

''

'

''

A.

'

MILLER

'I

AND

'I

WILLIAM

483

J, MCGILL

a-'

80

"OO.

_Q^"

^oo

^o

lu
_i

"

60

"J
tJ

cc

o,

^^o
UJ

MONOSYLLABLES

32

40

"L,-"

o.

liJ

CL

20

=0.10

Po=0.27

'

'

'

"

30

26

20

IS

10

NUMBER
TRIAL
Figure
3
Values of p" for a Two-Parameter
standard deviation from p"

Comparison of Theoretical and Observed


line is drawn

"

Dotted

20

NUMBER

TRIAL

Figure

Comparison of Theoretical

Case.

IB

10

one

and

Observed

Values

4
of p(^*

n) for

Two-Parameter

Case

484

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

number

cumulative

of Theoretical and

Comparison

recalls,

of

Figure

Distribution

Observed

Two-Parameter

5
of Recalls
Case

on

Four

Different

Trials in

lOO-WORO
ANECDOTE

0.87

a=

Po-0.61
1

-J

L_

TRIAL

TRIAL

Figure

Comparison

that

Tk

of Theoretical

and

p"+i

and

approach

Observed

the

Values
Case

of

and

p"

p(4*

n) for

asymptotic value, m.

same

So

Two-Parameter

we

have

the

equation
a

lim

Since
"
m

0,
"

"

"

cannot

po

a,

be

for if po "

cannot

we

In

obtain

(25)

unity; and since both a " 0 and


general,we are interested in cases
forgettingrather than acquisition.

exceed

negative.
w,

p"+i

"

where

set-theoretical
of the

conditioned

material

Of

remainder,
as
before,but
1

now

the

of

sample of

measure

measure

of the elements

measure

"

"

sentation
pre-

is

po

is conditioned

{1
extinguishedduring recall,i.e.,

and

measure

"

the

and

the conditioned

add

We

the

elements

conditioned

a) po.

Thus

have

the second

recall the

Continuing
When

its solution

First,we

take

to

variable

new

(18), Math

is

"

"

into

"

"

Tk/m

that
of

(1

the

now

our

(1

d)po

"

is

"

(m

repeated:
a) Ti

"

po)a

"

obtain

the

solution

same

as

appropriate difference

the

is

case

simplest

hardly
to

way

"

forpo

in the

for po and

of two

case
a

less cumbersome
with

work

of the two-parameter

probability,ri

Po/m)a,

po/m

of

know

(1),we

transitional

new

for (1

"

such

these

case.

that

m.'

(26)

parameters given in
a). Therefore, from

that

(1
"

0,T\

appear

substitution

(20),we

"

Po)(x.

"

aTi

advantage

introduce

ri

{I

for the three-parameter

(2). It would

equations is

(2) and

(24) is substituted

than

"

generates the relation (24).

in this way

equation, but

This

(m

~\~O,

apo

"

sampling procedure

Tl

"

same

T2

Po

ToT]

first recall

the

On

follows.

as

of elements

sample

subtract

must

we

Ti

At

At

runs

485

MCGILL

J.

(24)

elements, a portion of
extinguished.
a, are

"

during presentation
we

word.

these
"

for

random

WILLIAM

AND

rationalization

for every

is drawn.

MILLER

A.

GEORGE

tQ"

2^

Ta_
A-l

1=0

ml

p'iAk

When

Thus

we

a^'

n).
,

ri is substituted

of the summation

to

(2), the factor m'

into

in the

cancels the factor m'' in the denominator


know

the

in front
tion.
summa-

that

1^
~

p(A,

product

under

n)

T'or[

"

"

"

tL.

j=0

''^"

(28)
,

486

READINGS

which

is the

same

summation.

(1

This

T,)"

[(1

(1

p\Ak

as

m) +
my

When

term.

consider

we

m)""'w(l

this sequence

substitute

we

"

(1

Z-J

Tk-1

"

m"(l

in

(29)

term

sum

by

have

we

'^i)

J.

'""

tQ".

(28) and

of this sequence

/
"

rd' +

for the numerator

,
'

tQ

the last term

TqTi

the

T^T

n(l

under

be written

can

m(l

for the numerator

(27) except

my-Va
(2)(1

Now

in

n)

numerator

PSYCHOLOGY

MATHEMATICAL

IN

n (r;

r!)

j=0

which

know

we

from

(27) is equal

to

m"p' (A^

n).

The

to last term

next

gives
A

n(l

r[r'

m)"-'(l

(r;

rO

1=0

which

know

we

in this
we

know

from

manner

the term

p{Ak

n)

(27) is equal

ni'p'{Ak n) + n(l

the

parameter
We
in

m)w"~^ p'(^-t

"

brings us eventually to the case where


is zero.
Consequently, we can write

asymptote

"

m)m'~^p'{Ak

"

k, and

"

1) +

Wm''(l^)"~y(^^
-

is

unity (m

ing
Proceed-

1).

"

[n-k}^^ "^y'^^'ViA,

When

n{\

to

"

"

"

k)

(30)

-i)-

(30) reduce

1), (29) and

then

to

the

two-

case.

recall that

(1),(30) can

of the

because

be written

viA, ,n)

Z
i=o

way

in which

our

were
probabilities

as

('')m\l
my-y(A,
-

\1/

i).

fined
de-

488

READINGS

It is of interest to observe
in

(33)for
is,

and

p"+2

p"+i

the
,

lim

reflects the

This
the

variance

that when

t^

the

limitingvalue,m, is substituted

limitingvariance

Var

around

m(l

(r"+i)

is found

to be binomial.

That

m)

"

fact,established

of the

PSYCHOLOGY

MATHEMATICAL

IN

earlier in
to

goes

that
(5'),

as

grows

very

large

zero.

example, we have taken the data from


another
Sixty-four
subject in the experiment by Bruner and Zimmerman.
words
the
and
the
order
of
read aloud
was
monosyllabic English words were
scrambled before every presentation. A visual inspectionof the data led us
to choose an asymptote in the neighborhood of 0.7. This asymptote is drawn
the plot of the 4 in Figure 7 and on the plot of the r" in Figure 8. Then
we
on
In order

to

obtain

numerical

I .00

o
"

0.80

"

"

0.60

a.

"

0.40

"

0.20

"

10

CUMULATIVE

NUMBER
Figure

12

16

RECALLS,

OF

14

Function
of Number
of Recalls
a
as
Transitional
Probability of Recall, ta
circles. The curve
Values of tk are indicated by open
Case.
Parameter
the tk is Tk
0.7
0.57 (0.83)*.
,

in the Threefitted to

in
consideringall the trials on which words were
of the to,n+ifor all those
state Ao and calculating
po as the weighted average
This was
estimated
the sampling parameter a
0.83.
trials. Next
we
done by obtainingthe estimates,4 for successive values of k; these estimates,
used the
We
together with (24),give us a set of equations estimating a.
of these estimates
(ignoringnegative values). Then we
weighted average
the
m(l
from
on
0.12
obtained a
equation a
a). We shall comment
the estimation problems later.

estimated

po

0.13 by

"

GEORGE

'00

"

(III

'

"

"

"

AND

MILLER

A,

WILLIAM

J.

"

^^^

"

489

MCGILL

"

r-r

^O

80

60

40

"S"'

"

Ui
u

q:

^^"^^o

9.

MONOSYLLABLES

64

.^cr.

UJ

20

"

"

"

"

'

"

'

and

Observed

'

'

0.12

oc

0.83

Po

0.13

'

'

'

'

'

30

25

NUMBER

Figure
of Theoretical

"

20

TRIAL

Comparison

15

10

"

8
Values

of

for Three-Parameter

pn

Case

"

"

10

"

trial

20

Figure

Comparison

of Theoretical

and

2S

90

number

Observed

Values

9
of

pC^A

n) for Three-Parameter

Case

490

these parameter

When
function
for A;

Tk

shown

1, 2, 3, 4, we

When

for

they

in

values

Figure 7.

obtained

substituted

were

MATHEMATICAL

IN

READINGS

PSYCHOLOGY

substituted

were

When

the values

the functions

into

(31) we

for

(24) we

were

substituted

p(Ak

obtained

obtained

into

shown

into

the

(28)

in

n)
Figure 9.
the function,for p" shown
,

Figure 8. In Figure 8 the dotted lines are drawn " one standard deviation
from p,
as
computed from (33)
,
A comparison of the values of p" computed from
(31) and from (32) is
given for the first eighteen trials in Table 1. With this choice of parameters
the Bush-Mosteller
highly satisfactory.
approximation seems
in

TABLE

Comparison

and

of Exact

1
Values

Approximate

of

p"

for First

18 Trials

Discussion
In the

preceding pages
being memorized

we

several words

explicitassumption that the


izing
simultaneously are independent, that memorhave

made

the

probabilityof recallinganother word on the


be justified
list. The assumption can
only by its mathematical
convenience,
contradict
it.
The
learner's
the data uniformly
because
introspectivereport
one

is that

word

groups

affect the

does not

of words

go

together

to

form

associated

clusters,and

this

pairs of words
impression is supported in the data by the fact that many
successive
trials.
If the theory
recalled together or omitted
together on
are
is
behavior
of
reasonable
50
describe
the
used
is
a
to
rats, independence
assumption. But when the theory describes the behavior of 50 words in a
list that a singlesubject must
sumption.
learn, independence is not a reasonable asthe
examine
of
ducing
introto
It is important, therefore,
consequences
covariance.
The
the

difference between

theory can
The

independent and

the

dependent versions of
interpretation

of the set-theoretical

Imagine that we have a large ledger with 1000


presentationof the list is equivalent to writingeach of the words

of the two-parameter
pages.

the

best be illustrated in terms


case.

GEORGE

at random

100 pages.

on

at random.
are

On

the
sure

the

that

word
for

written

and

could

simply make
words. A, B, and C, on
recalled again it would
we

The

the

rule is that
Thus

on

for C.

With

sample of

same

be

A, B,

each
a

and

of these

50/1000

likelythat

random

at

50 pages.

and

C.

page

These

words

must

0.05.

With

and

make

dependen
in-

pages

dependent model, however,

selection of 50 pages

the

words

select

we

first select 50 pages


at random
all of them, then select 50 more

more

one

probabilitythat

of the elements

same

and

Then

C would

does

the theoretical

of the estimates

give a
paid to

of

fair
the

write all three

whenever

also be recalled

was

at the

the

subjects
The

may

to

ledger on
written

are

which

describe

each

combined

surprisingthat
even
though no

the equations

of pairs of words.

tive
Associa-

not

scores

attention

the rate, of memorization.

from

word

the linear difference

in the

to estimate

word,

to

is

p"+i

the

on

tion
equa-

list.

the various

only

Thus
r*

from

If the parameters

approximation

an

data

of the

mean

by averaging the recall probabilities of all

be expected
Similarly,the expressions given for p"+i cannot
result of averaging several subjects'data together unless
to have

general theory,
(6). The data
for

applicable,though

the

the

or

is not

theory

For

xt

values

same

of course,

of

oi

recall

obtained

measure

in this way
does not
effect is to increase the variance

not
variability,

and

a, po

word

functions

values

affect the

be

known

are

the form

the

of recall determined

words.

other words

only
words, it is

the

upon

in the

of covariance

The

assumed

vary

introduction

descriptionof

from

probabihty

what

upon

depends

of pages

of jointoccurrences
probabilities

different words

describe

depend

not

In other

p"+i

parameters

(6), are

(the number

recall,
p"+i

clusteringshould
The

will be recalled

to it

Therefore, the

pages.

change

word

conditioned

it is inscribed)and

the

50

Now

0.1.

491

MCGILL

time.

same

was

The

would

J.

100/1000

find written

we

we

was

B,

po

WILLIAM

selected at random.

model

independent

AND

first trial.

50 pages

on

MILLER

Thus

this page

on

responses

be written

A.

to use, and

p{Ak n).
a
descriptivemodel

limited

us

the

will enable

linear

to

force

cases

tedious

of the parameters.

may

all such

to

all

restrictions

consider

more

of

plicated
com-

general solution

to

us

to

(2)

compute

the necessary

used

tease

is

Once

parameters, the

necessary

and

to

step, however,
the

data.

As

yet

we

have

has

been
the

vary

these parameters.

efficient methods

need

we

step is to

next

the effects upon

observe

of this sort

found

no

to

experimental

In order

to take

the

out

conditions
this next

of

estimating the parameters from


to the estimation
satisfactoryanswers

problem.
There
functions
a

is made

example

is

p(Ak
at

of

sizeable amount
,

n) and

p"

If
.

the outset, it takes

in the

computation
choice

poor

several

preceding section,

we

hours

estimated

involved

in

determiningthe

of the parameters
to discover

the

the parameters

a, po

fact.

and
,

In the

successively

492

READINGS

used

and
had

been

too

low.

not

different

parts of the data

computed it seemed

particularlygood

to

ourselves

that the

problem

leave the estimation


with

problem

estimates

our

have

used

have

considered

beyond
with

the

of po and

to fit the

all parameters

was

the mathematical

that

us

different estimates.

we

one.

to estimate

for the

We

Clearly,the method

all of the data

use

PSYCHOLOGY

MATHEMATICAL

IN

pious hope

is

in order

simultaneously. We

to

convinced

Consequently, we

that it will

p"

both

were

to the data

least squares

abihties.

our

theory

After

appeal

to

must
one
some-

to solve it.

competence

Appendix A
Solution
The

solution

enumerated
below

our

has

n) in the General Case

equation (1) with

of

been

obtained

method

own

for p(Ak

several

of solution

in the past

times

because

conditions

boundary

the

(4,5). We

present

procedures involved

the

have

we

may

be

of interest in other

applications.
follows:
as
Equation (1) may be written explicitly
(1

This
1

Top(Ao ,n) + (l-

Ti)p(Ai , n)

Tip(Ai

T2)p(A2 , n)

7i)+

system of equations
-

To

Ti

72

The

T,

infinite matrix

infinite column

we

be written

can

0
T2

p{Ao
p(Ai
p(^2

in matrix

This

(1

To

To)p(Ao , n)

p{A,

p(^3

T3

1)

1)

1)

notation

p{Ao

P(^2

as

follows:

1)

1)

p{A2

,n

\)

p{A^

n)

p{Ao

n)

p(i4i

n)
n)

4- 1)

probabilitieswe shall call T, and the


trial n and
of the state probabilities
on

of transitional

vectors

made

shall call d" and

d"+i

initial distribution of state

up

So
.

we

can

write

do is the infinite column


probabilities,
,

vector

{1, 0, 0, 0,

"

"

"

The

state

trial two

on
probabilities

trial

on
probabilities

state

Tdo
The

WILLIAM

AND

MILLER

A.

GEORGE

J.

MCGILL

one

are

493

then

given by

rfi

given by

are

Tdy

d,

by substitution,

so

Td^
this

Continuing

T{Tdo)

T'do

d2

procedure gives the general relation


rdo

rf"

Therefore, the problem of determining d" can be equated to the problem


determining T".
that it can
be expressedas
know
Since 7" is a semi-matrix,we
T
where
as

are

is

arbitrary,so

SDS-\

its diagonal
elements
on
diagonal matrix with the same
of S are
elements
The
diagonal
2).
diagonal of T (e.g.,

infinite

an

the main

on

of

let Sa

we

S,,

Now

1.

can

we

write

Ti

it is

Now

simple

solve for "S2iwe

which

2 and

(from row

construct

To

for "S,,term

to solve

matter

(1

"

For

example,
equation

term.

1) the

column

To),

Szxil

To/(ti

Ti)"S2i

"

gives
S21

To

by

solve for S31

we

use

T1S21 +

the

(1

"

To).

"

equation
=

*S3i(1

Ti*S2i/(t2 To)

ToTi/(ti

T2)*S3i
S31

To)

"

"

To)(t2
"

To).

to

494

READINGS

this

Proceeding in

IN

gives the

manner

PSYCHOLOGY

MATHEMATICAL

of

elements

necessary

we

have
0

and

S,

""

To

(ti

For

1.

column

times
we

"

"

example,

one

To) (tz

To)(t3

of S~^

elements

The

*S"S~^

To)(t2

"

of *S~^ :

"

obtained

be

can

the element

to/(ti

by

term

term

T2)

"

from

aSziof S~^ is given by

tq) +

"

Tj) (ts

Ti)(t3

"

S21

equation
two

row

Continuing

0.

the

of S

in this way

have

s-

(to

T3)(ri
"

"

matrices

These

"

permit

T3) (ti

T'

T3)(r2
"

T3) (t2

simple representation of

in

(SDS-'XSDS-')

the

"

T3)

powers

of the

SD{S''S)DS-'

SD'S-\

general,
r

Since

"

T., Thus,

matrix

and

T3)(t2

Z) is

diagonalmatrix,D"

SD^S-'.

is obtained

by taking the

nth power

of every

496

READINGS

The

right side

of this

PSYCHOLOGY

MATHEMATICAL

IN

(5) and

equation is,from

The

(18),p"+,

left side

can

be rewritten

LI

k=i

which

becomes

trial

on

\^

O,)

know

we

subtractingp(Ao

we

ft-O

A-O

")V(^* n)

pn

[1

(1

ay]pn

(1

Po)

Z (1

")V(^* ^);

obtain

Rearranging

(1 -Po){l

[1

(1 -a)"]p"}.

gives

terms

Pn.i

is the desired
From

a)V(^* ,n)

P".i

which

n),

~l

23 (1

so

"

that

Pn

and

''".

"

2Z Pi^ic,n)Z (1

Now

and
"

1),

"

[1I [II a|.


?('!".")]

have, by adding

now

a)

"

(with n

g
We

(1

"

(1

po

Po)[l

(1

(21)

a)"]p"

result.

this result

(15) is

obtained

directlyby equating
C

Appendix
List of
a

parameter.

Ak

state

parameter.

d"

infinite column

infinite

number

that

word

is in after

of times
value

asymptotic

number

total number

Nk.n

number

Po

probabilityof

Their

being

t^

n)

its elements.

as

similar to T.

word
of

Meanings

recalled k times.

vector, having p(Ak

diagonal matrix

and

Symbols

has been
and

p"

recalled.

of trial.
of test words

of words

in state

recallinga

po and

to be learned.

A,, on
word

trial

n.

in state

Aq

a.

GEORGE

p(Ak

A.

probability

,n)

that

observed

r"

MILLER

word

recall

WILLIAM

AND

will

score

trial

on

in

be

Ak

state

estimate

n;

497

MCGILL

J.

trial

on

of

n.

p"
.

p"

probability

Sij

elements

S'ij

elements

infinite

tk

estimate

U,n

observed

of

recall

of

*S.

of

aS~\

matrix

on

used

of

trial

n.

transform

to

into

matrix.

diagonal

similar

Xk
"

fraction

Tk

probability

infinite

of

words

of

recalling

in

Ak

state

word

in

that

recalled

are

trial

on

n.

Ak

state

matrix

of

probabilities

transition

r^
.

Var

variance

(r")

of

estimate

the

of

p"
.

"X^i.it.n+i

random

variable

equal

to

0.

or

REFERENCES

1.

Bush,

R.

presented
2.

Cooke,

3.

Estes,

4.

Feller,

and

R.,

the

to

R.
W.

G.

the

Proceedings

5.

Woodbury,

M.

A.

received

the

theory

Berkeley

of

Psychol.

learning.

Symposium

for

with

particular
on

learning.

December

London:

spaces,

processes

model

Boston,

Statistics,
sequence

stochastic

operator

(Paper

27,

MacMillan,
Rev.,

1950,

57,

reference

Mathematical

1951.)
1950.

to

94-107.

tions.
applicaand

Statistics

403-432.
On

313.

Manuscript

of

of

1949,

and

statistical

theory

linear

Mathematical

matrices

Toward
On

Probability,

for

Institute
Infinite

K.

W.

Frederick.

Mosteller,

3/ 11/

52

probability

distribution.

Ann.

math.

Statist,

1949,

20,

311-

CHOICE

ULTIMATE

BETWEEN

TWO

PREDICTIONS

ATTRACTIVE

FROM

GOALS:

MODEL*

MosTELLERf

Frederick
HARVARD

UNIVERSITY

AND

Maurice

Tatsuoka

university

mathematical

A
choices

model

for two-choice

is discussed.

desirable

are

hawaii

of

behavior

According

in situations

to the

model,

one

where

both

the

other

or

tion
ultimately preferred,and a functional
equation is given for the fracsolution
population ultimately preferring a ^ven choice. The
and
the initial probabilitiesof the
the learning rates
depends upon
upon
choices. Several
techniques for approximating the solution of this functional
One
of these leads to an
that gives
equation are described.
explicitformula
This
solution
be generalized to the two-armed
bandit
can
good accuracy.
in each
the equivalent T-maze
or
problem with partial reinforcement
arm,
the calculations
for a highto program
speed
problem. Another
suggests good ways
computer.
choice

is

of

The

mock

has

Buridan's

seemed

always

equihbrium

an

equilibrium

in

of

theory

for

model

for

model

the

initiallyshifts its choices

who

starved
No

this paper

In

to

doubt

behavior.

One

situations

approach-approach

mathematical
In

ass,

unreasonable.

goals will be chosen.

attractive
a

of

immobility

haystacks,
to

the

death

be

unstable

in these

from

to

one

was

such

any

of

one

"

the

properties that flow from

some

behavior

repetitive approach-approach
behavior

two

invented

that

expects

will

between

the story

choice

situations, an
after

another, but

while

cussed.
dis-

are

organism

settles upon

single choice.

Thus

in the

expression

some

different

on

early part of the learning the theoretical


to

the

trials,but

notion

of

eventually

organism

equilibrium by making

an

this behavior

even

different

vanishes

give

may

choices

for the

single

Science
Foimdation
*Support for this research has been received from the National
Health
of Mental
Institute
the National
(Grant M-2293), and the
(Grant NSF-G2258),
University.
Laboratory of Social Relations, Harvard
ance
and express
our
appreciation for the cooperation and assistfWe wish to acknowledge

given by Phillip J. Rulon,

Albert

Beaton,

and executed
numerous
up, programmed,
method
of solution, and by Cleo Youtz
work.
We
also wish to thank
Ray Twery and
3

some

Illiac

calculations

for extensive
Robert

R.

Donald

connected

Bush

article

P.

Nash,

appeared

in

the

at

unpublished results of their calculations. Those calculations


Laboratory of
through the cooperation of the Digital Computer
John

Spearritt, who

with

every
for permission to

of the

Illinois,Dr.
This

Ho, and

Wai-Ching
calculations

set

were

the

linear

tions
equa-

stage of the
use

made

in Table
on

Director.

Psychometrika,

1960, 25, 1-18.


498

Reprinted

with

the

University of

permission.

MOSTELLER

FREDERICK

AND

MAURICE

499

TATSUOKA

ultimately choose one


organisms may
organism. On the other hand, some
goal and others another, so that a notion of equilibriumor balance could be
a
recaptured across
population of organisms. The quantitativeaspects of a
cussed
behavior
for
such
model
are
investigated.The model employed is one disby Bush and Hosteller [1].
then the mathematical
A simple situation will be discussed first,
problem
bandit
encountered
there will be related to the more
complicated two-armed
each
reinforcement
each
that
with
Suppose
on
arm.
on
partial
problem
infinite sequence
trial of an
organism may
an
respond (or choose) in one
of exposition,specify the ways
R and L (for
of two ways.
For purposes
as
for
think
of
that
concreteness
rat choosing
one
can
a
right and left,say), so
the left-hand or right-hand side in a T-maze, or a person
choosing the leftin a two-armed
bandit situation.
hand
the right-hand button
However,
or
intended
for a general pair of attractive objects or
R and L are
to stand
exclusive
lead to attractive
and
exhaustive, which
mutually
responses,
goals.

Suppose that on a given trial


that of choosing L is 1
p, where as
of
time
next
R
probability choosing
"

probabilityof choosing R is p,
usual 0 " p " 1. If i2 is chosen,then

and

is increased

if L

the

to aip

-|- 1

"

but

ai

the

where
to a2p,
probability of choosing R next is reduced
choice is made,
1- The point is that when
0 " ui " 1,0"q;2^
a reinforcing
of
chosen
choice
increased
has an
next
that
being
probability
time, and
The asymmetry
in the formulas
both R and L are
regarded as reinforcing.
the probabilityof choosing R,
from
the fact that the notation
uses
comes
and not the probabilityof choosing the particularside chosen on each trial.
discussed
The
are
by Bush and
operators used to change the probabilities
Mosteller
([1],
p. 154 ff.).
Suppose the organism continues making the choices and that his probabihties are adjusted after every trial according to the rules just given. Then
later the organism stops making one
of the
that sooner
be shown
it can
or
extreme
choices and thereafter chooses only the other. An
example occurs
then the organism chooses forever what he chooses
if both "! and aa are zero

is chosen

the

"

first

learning).
(one-trial
One

mathematical

eventuallychooses

problem is to discover
R rather than

the

probabilitythat

the organism

L all the time. If he does choose

"ultimatelyattracted

R all the

is "ultimately
by R," or
time, then he is said to be
should
be
The
desired
as
expressible a function of
probability
attracting."
coefficients ai and 012 (the
the initial probabilityp and of the attractiveness
attractive the side). For convenience, this will be
smaller an
a, the more
the
called
cated
complisimple approach-approach problem, in contrast to the more
problems.
partialreinforcement
Consider
an
now
as
experiment with paradise fish
example a T-maze
of this experiment a fish
Wilson
On
each
trial
and
described
[2].
by Bush

500

READINGS

IN

PSYCHOLOGY

MATHEMATICAL

other,where the left or right


side could be chosen. When
the right-hand side was
chosen, the fish was
rewarded
75 percent of the trials. When
the left side was
on
chosen, the fish
rewarded
of
the
trials.The
25 percent
was
on
operationwas to placethe reward
started

on

at

side

one

reward

end

one

of

tank

the other

or

through

and

time.

every

to the

swam

In

transparent divider

fish

one

group

when

he

chose

In the other

divider was
used. The
an
group
opaque
that the fish tended to stabilize on one
side

showed

Within

the

framework

of the

was

data
or

see

from

the
side.

unrewarded

the

operators described

able to

these groups

the other.
earlier in this paper,

probabilityof choosing the right-hand side on a given trial,and


right-hand side is chosen and rewarded, the new
probabilityof choosing

if p is the
if the
the
were

be

If the left-hand side


expressed as ap + 1
a.
chosen and rewarded, the new
probabilityof choosing the right might
reduced
to ay. The
parallelwith the previous descriptionsis very close.
three
But
the side chosen
is not rewarded.
Then, essentially,
suppose

right-hand side might

exist.
possibilities
(a) The side chosen

be

"

before. The
likelyto be chosen than it was
explanationmight be, for example, that the organism is building up a habit
pattern, or that he is secondarilyreinforced for being in a place that earlier
was
rewarding.
planation
(b) The side chosen is less likelyto be chosen than before. The exhas been received that
might be, for example, that information
this side is not
Whatever
and

to

is

more

paying off.
the explanation

(b) make

quite different

probability associated
reward
is given or not.

with

corresponding to
be, the models
predictions.The model for (a) says that
may

the

side

chosen

This

is

increased

always

(a)
the

whether

^for the operators described

ultimatelyimplies
every time, that is,that eventuallythe organism
stabilizes on
for (b) would
side. On the other hand, the model
imply
one
that an organism
that the organism does not stabilize. To see this,suppose
choose
the
side
that
is certain (p
1) to
right-hand
is,he has stabilized
of partialreinforcement
the organism will exthe right.Then
because
perience
on
will reduce
trials on the right-hand side. These
nonrewarded
some
left-hand
side
the probabilityof choosing the right-hand side,and
the
so
shows
that the organism
will be chosen
sometimes.
A similar argument

here

that

"

"

side is chosen

one

cannot

"

stabilize

assumption

on

the

(b) would

left. Thus

partialreinforcement,a model for


A subject does
asymptotic instability.
the other,nor
does he finally
acquire a

under

typicallyhave

by one side or
fixed probabilityp of choosing R. Instead,his value of p drifts up and down,
stable way.
model
Thus
though in a stochastically
(a) has attractingand
barriers.
(b) has reflecting
absorbing barriers,while model
The
then
is
everything
(c)
probability
unchanged by a nonreward

not

become

attracted

"

depends

upon

the rewarded

trials.

AND

MOSTELLER

FREDERICK

501

TATSUOKA

MAURICE

(a). In
paradise fish the data suggest model
shall deal with the type (a) model. On the basis of the model,
this paper
we
would like to know
we
(interms of the learningrates,the initial probabilities,
of reward on the two sides)what fraction of the organisms
and the probabilities
In

the

will stabilize

on

given

the

Because
and

to

problem
general problem

the

because

what

know

side.

numerical

work, we will sketch


is time-consuming in
want

with

experiment

has

turned

has

some

solutions

various

that

of R

on

{aip"
-I- 1
P"+i

ai

"

will

worker

the

of

,,.

(1)

research

of them

simpleapproach-approach
probability that an organism is
Let /(pi ; ai
aa) be the probabiUty
on

choosing R. The transition rules


trial n, then the probabilityof R on the

probabiUty

been

in choices

of trials ends

infinite sequence

an

by previous

tried. Each

Work

of previouswork

facilitate discussion

shown

as

development and testing,so a


ground has already been plowed.

problem, a functional equation for the


ultimatelyattracted to R will be derived.
that

have

its

Previous
To

interest

some,
trouble-

rather

be

to

out

of R.

Here,

pi is the initial

are:

if p" is the

next

trial is

if

is chosen

on

trial n,

if

is chosen

on

trial

probabiUty

"{

[asPn

n.

usually no advantage in referringto the trial number


for the
with p, so the subscript on
associated
p stands
pi is dropped and
that the desired
initial probability.
Similarlyit is always to be understood
full
notation
the
is needed,
function / depends upon
ai and
Ui ; so except when
In the

sequel there

is

f{p)will be used.
quantity f(p) may

the notation
The

the

to

of

member
R

with

the first choice


of R

of 72

of L

or

the

same

the fraction p

is uip

-f- 1

ai

"

of two

composed

"

the

parts

initial trial. Assume

the

on

parts

sponding
corre-

that

each

p of

choosing
probabiUty
on
Then,
simple approach-approach problem.
of the individuals choose R, and the new
abiUty
prob-

largepopulation has

is faced

and

choice

be

the

initial

same

for any

of this group.

member

This

means

being ultimately attracted

by R is
probabiUty
the
contributes
portion
f{aip H- 1
ai). Consequently this group
those organisms choosing L
manner
ai) to f(p).In the same
p fiuip 4- 1
first contribute
(1
p) jiazp)to f(p).Thus one derives the basic functional
equation for the simple approach-approach problem:
that

in this group,

of

the

"

"

"

(2)
The

/(p)

boundary

because

if p

conditions

0, then L

p/(a,p -fl
are

/(O)

occurs,

and

aO +
0 and
the

new

(1

p)/(a.p).

conditions
1. These
/(I)
for
R is az-O
probability
=

hold
=

0.

502

READINGS

Therefore

L is

and

Thus

if

them

certain

(A

satisfies it

/(I)

and

after

terms

derived

always

for the function

conditions

the

new

chosen.
needed

are

(2)only determines / to within a


(2),direct substitution shows

that

Af -{- B

also

constants).

are

have

four

had

parts if

related

we

the desired

probability
generally 2"

after

two
occurring
trials,or more
all
are
equations
equivalent,but they
applicationsof (2) to the /'sappearing on

terms

successive

can

all be

the

right-

side.

hand

properties of f{p) have been


Shapiro ([3],Parts II and III),and by
The

all of their results

not

is

trials. These

by

satisfies

four

the

1. These

and

occurs,

Therefore

1.

ai

"

1, then R

linear transformation.

Equation (2) could


to

PSYCHOLOGY

SimilarlyUp

i2 is oii-l +

without

chosen.

always

probabilityfor
Thus
0
/(O)
because

MATHEMATICAL

IN

useful here

solution

of the solution.
the

once

tonicityis

Equation (2) has a unique, monotone, analytic


conditions
are
given. With our boundary conditions

boundary

the solution

before

below.

given

are

i. Nature

and
by Bellman
by
Karlin [4](c.f.[1],p. 163-4). Since
those propertiesof f(p)especially
readilyaccessible,

are

studied

is

with

"

ai

0:2

for

concave
,

cui

"

a2

The
.

mono-

ing
probabilityinterpretationgiven by the learnthe
of
the
larger
given a^
probability
choosing R
az
the more
initially,
likelythat R is ultimatelyattracting.
ii. Solutions under
the
follows, suppose
special conditions. In what
relevant boundary conditions
and
hold.
1 to
The
0
special
/(O)
/(I)
model

consistent

for

convex

for

"

the

and

conditions

(a)

have
=

"!

f(p) is both
(b) ai
p

or

to do
5^

az

0,

1. The
and

convex
=

with

a2

the values

assumed

solution

f{p)

the

is

and

concave

by

by

probabilityof

one

or

both

of the a's.

implied by the fact that


boundary conditions.
defined in our
problem unless
as

p,

the

function / is not

I. The

because

changes

never

and

no

attraction

occurs.

(c) "!

1,

a2

5^

because

1. The

of R

occurrence

leaves

the

probabilityof

toward

unchanged
can
only move
ai
a^p
p, so
L's unless p
1. Thus
choosing more
f{p] 1, az)
0, a2 9^ I, p ^ 1, and
1.
/(I;1, "2)
?^
(d) az
1)
1, ai 9^ 1. Similarlyj{p; ai
1, p 5^ 0,
1, a,
and /(O;aj
=0.
1)
0. Here, the only way
to be ultimatelyattracted to L is always
(e) "!
to choose
L. The
probabilityof the latter behavior is
-{- I

the process

"

"

(3)

Sf(p,
a^)

(1

p){\

a2p){\

alp)

"

"

"

fl(1

1=0

Therefore

(4)

the

probabilityof

ultimate

/(p;0,a2)

attraction
1

by

9(p,cx2).

is

a^p)
.

504

READINGS

close to the true

IN

MATHEMATICAL

PSYCHOLOGY

of this paper,

In the remainder

ones.

techniques for

several

approximating f(p),are
A
then
be

provided.
designed for high-speed

method
an

excellent

approximation

considered,and

calculation
from

obtained

the unit

grid of

interval,and

of these

values

notation

develop

no

the

numbers

longer correspond
set of equations

(7)

=0

KPi)

there

methods

of

bandit

approximating

Equations
,

P2

"

"

"

"

(= Pn+i) in
appliesto each
1

Pn

with

earlier

as

they

but the subscripts


probabilities,
in earlier sections.)
Then
has
one

did

+/(0),
PiKdiPi +

/(P2)

P2K(XiP2+

KPn)

+
PnfiaiPn

first and

to the two-armed

equation (2) as it
(Lest confusion

"

ai) +

(1

ai) +

(1

^l) +

(1

/(I) =/(l)
The

first,
equation will

Pi stillrefers to

to trials

/(O)

considered

variables.

independent
that

note

(= po), Pi

the functional

write

of the

other

Simultaneous

Approximation hy
a

that result will be extended

then

of some
problem. Finally, brief mention
this functional equation will be given.

Consider

will be

differential

Pi)Ka2Pi),
p^i{a2P^

"

Pr)1{oC2P^
,

+0.
of this set of

last members

equations are,

of course,

tautologies;

only n nontrivial equations.


The
right-hand sides of the n nontrivial equations of the set (7) each
involves the values of j{p) at points that do not ordinarilycoincide with any
of the chosen grid points.However, by using an interpolation
formula, both
be approximated by
+ 1
ai) and f(a2Pi),i
j{oLiPi
1, 2,
n, may
linear combinations
consecutive
of the values of f{p) at two
more
or
grid
The
number
of grid points required depends upon
points Pi
Pi+i
are

"

"

"

"

"

"

"

interpolation(two grid points),interpolationwith


second differences (threepoints),third differences (fourpoints),and so forth.
Whatever
the number
of points may
be, each equation of the set (7)
be
can
just the
replaced by an approximate equalityinvolving as unknowns
values
of /(p) at several predetermined grid points, and
these unknowns
Thus
occur
mately
only linearly.
a system of n linear equations is obtained,approxisatisfied by the n unknown
f(Pn)-The
quantities,f(pi),fipz),
idea of deriving a system of linear equations whose
roots approximate /(p"),
whether

one

linear

uses

"

'

"

1, 2,

"

"

"

n,

was

first

unpublished memorandum,

f(aiPi +

"

ai)

and

suggested to

in which

linear

fiazPi).

by J. Arthur
interpolationwas

us

Greenwood
used

to

in

an

mate
approxi-

FREDERICK

this and

In
which

"!

example

in the

.75, ttz

has

fairlyeasy
so
easy to fit,

the

to

attained

MOSTELLER

.80 is used

of

from

compute
the reader

not

Taking

pi

0.25,

short,in accordance

is illustrated

standard

example,

the

precision

ai

grid

0.75, as

of five
0.80.

p)m.80p).

and

0.75

for
=

writing /(p,)

fi

for
,

/i

0.25/(0.4375)+

0.75/(0.20),

/2

0.50/(0.6250)+

0.50/(0.40),

0.75/(0.8125)+

0.25/(0.60).

First,linear interpolationwill be
of linear
/(0.6250),etc., by means
,/3and/4(=

(1

0.25) +

0.50, pa
equations (7)

pz

with

(9)

A ,U

thinking that

described

pK0.75p +

into

be misled

obtainable.

method
just
Example.
equally spaced points,using the
Here, the functional equation is

Kp)

This

methods.

various

in

are
being easily displayed;further,numbers
the disadvantage of being relatively

The

(8)

illustrate the

to

example

it. It has

should

always

numerical

standard

505

TATSUOKA

MAURICE

following sections

advantage

for it is

AND

used

to

approximate /(0.4375),/(0.20),
of the five /'s:/o(= 0),

combinations

1).Thus,
0.5000

,rn
Ao^r^
^(0-^3^^)

0.75/2

0.25/i +
0.25

0.20
,

0.80A

0:^500

^^

^" +

~0:25-^^

0.25
=

0.2500

0.20

^-

^^ +

0^500
=

,,",..
^^^"'

0.4375

0.4375

and, similarly,
/(0.6250)

0.50/2 +

0.50/3

/(0.40)

0.40/1 +

O.6O/2

/(0.8125)

0.75/3 +

0.25/4

/(0.60)

O.6O/2 +

0.40/3

0.75/3 +

Substituting these approximate expressionsfor the


in the right-hand sides of (9) and collecting
all terms
into

the

left-hand

sides,one
0.3375/1

(10)

-0.2000/1

+
-

several

functional

involvingthe

obtains

0.1875/2

0.25,

0.4500/2

0.1500/2 +

0,

0.2500/3 c-i 0,
0.3375/3 c-

0.1875.

values

unknowns

506

READINGS

Replacing the

by

~;

in the

the

resultingequations, one
(The best available values are

PSYCHOLOGY

MATHEMATICAL

IN

set

obtains

of

the

approximations (10) and solving


following approximations to /,"
for comparison.)
.

also shown

only fair.
use
approximating the non-gridin the right-hand sides of (9).The
general
point values of j{p) that occur
formula
(with equally spaced grid points)is
The

with

agreement

Now

the

best

second-order

Kx, +

e)

available

values

is

interpolationfor

i--

Ax

'" +

"

Ax

fA^- icP-'

(11)
e

i ^^- I 1
Ax

Note

Axr^'

'

that (11) givesthe

interpolatedvalue as a weighted
of the three adjacent tabled values instead of using differences.
average
mate
Applying (11) to the problem at hand and substitutingthese approxiinto
the
of
obtains
the
sides
following
one
right-hand
(19),
expressions

where

x.

a*.

"

system of approximations.

0.2410A
(12)

-0.1400/1 +
-

whose

These

roots

0.1744/2 +

0.0235/3

0.3925/2

0.3150/3

0.1369/3

0.0497/2 +

0,
-0.0625,
0.0872,

yield the followingapproximations.

results

are

definite

improvement

over

those

obtained

by

linear

interpolation.
to indicate that a considerable
improvement
example seems
of the approximation can
be expected when
higher differences are used in the
of f{p) in
for expressing the non-grid-pointvalues
interpolationformula
of the grid-pointvalues. However, the interpolation
formulas become
terms
and more
cumbersome
with numerically as higher differences
more
to work
included. It therefore is pertinent to see how
much
are
improvement can
be gained by increasingthe number
alone.
of grid points

The

above

FREDERICK

MOSTELLER

Improvement
Points

Obtained

in

Grid,

Entries

AND

by

using

Increasing

Linear

the

Values

507

TATSUOKA

Number

Interpolation

Approximate

are

MAURICE

of

only;

of

f.

Using only linear interpolation,


approximations from grids of 4, 5, 6,
obtained. These pointswere
not equallyspaced because
11, and 21 pointswere
it was
hoped that better results would be obtained by spacing the grid so
that the functional
values would
be approximately equally spaced. Information
needed
for such spacing was
available from other methods
described
later.
Linear

in the

made
interpolationswere
above
to obtain
approximate

The

0.90.

numbers

shown

are

values

in Table

results for the


at

five

scribed
grids de-

0.10, 0.25, 0.50, 0.75,

1, together with

the

known

best

values.

Using

the difference between

the best value

and

the cell entry for

a given
the
decreases
error
that,
roughly,
Pi
very
linearlywith the spacing.On the other hand, with a five-pointgrid,changing
from linear to second-order
alent
interpolationgives improvement roughly equivof points to 21 and using hnear
to that given by increasingthe number
interpolationonly. Since simultaneous
equations are expensive to solve,it
that second-order
interpolationis well worth the effort,contrary to
appears
as

usual

of error, it will be noted

measure

advice.

Calculations,with the
points and second-difference
have
obtained

been

made.

using third-order

they

could

In

be

more

labeled

The

an

results

electronic

are

well

are

differences,though in
values"

computer,

summarized

differences

useful. The
"best

of

interpolationas

by using second-order

those

numbers

aid

third-order

using

21

third-difference

as

in Table

2. The

grid
polation
inter-

results

hardly distinguishablefrom

sharply curved example


interpolationcolumn
provided

more

throughout this
degree of accuracy

paper.

be attained by using
can
principle,any desired
finer grids,but the cost of the calculations increases roughly as the square
of the number
of grid points used.
A
could
be
high-speed computer
to write its own
programmed
equations and solve them, but such a program

508

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

TABLE

Approximations
With

21

not

written.

this section

are

If

Grid

good

and

Order

Third-Order
the

is

accuracy

Interpolations

Approximation

Differential

Equation

required,the techniques proposed in

recommended.

Approximation by
An

and

Points

Second

By

was

Second-

Using

essential

of

feature

the

Differential
Equation
approximation
simultaneous-equations

replacement of non-grid-point
continuous
of grid-pointvalues. The
values of f{p) by linear combinations
variable analogue of this procedure is the expansion of fiocip-f- 1
ai) and
will
be
of
This
now
approach
as
fioczp)
Taylor's series in the neighborhood
p.
solution yieldsan approximation
used to derive a differential equation whose
discussed

in the

preceding

section

the

was

"

to the desired

function,f(p).
aO
Rewriting f{aip -f 1
the latter as a Taylor'sseries.
"

Kp -H (1

(13)

a.)(l

P))

as

f(p +

(1

aO (1

"

f(p) -h (1
(1
+

ai)(l

a^yq
2!

"

p)),and expanding

p)/'(p)

p)

f'iv) +

FREDERICK

where

/' and /"

MOSTELLER

the first and

are

Kv

(1

509

TATSUOKA

derivatives

of

with

respect

to p.

follows:

a,)v)

MAURICE

second

Similarly,expand /(aap) as
/(a.p)

AND

f(p)

(1

a,)pf(p)

(14)

^^^^np)-

Using only through the term in f'(p) in the two series (13) and
substitute these expressionsfor the functions in the right-hand side
functional
equation (2).The result is a differential equation
^^^

(15)

^f^^^^+

^^

(1

in

By rearranging terms

M(l

ai)'

[(1

"^^^^

+ ^(^
P'^^'^P^

p)[f(p)

(1

"')'(! P)'/"(P)]

+ Ki
a,)pr(j))

(14),
of the

a,)yr'(p)].

(15),

"i)'

(1

a,y]p}r'(p)+ {a,

aOfip)

0.

Hence,
f'ip)
(16)

2(a,

I'iv)

which

is

"2

"i)'

r/1

[(1

constant

Integratingboth

of

Ci and

sides of

C2

are

aO^

1, the final form

(19)

is an

abbreviation

for

+ a2)/2.
{oci

(17),
"]l+l/(l-a)

^2

of

C2 from
of

"|i/(i-a)

^^

/v

"

constants

new

Determining Ci and
=

and
integration,

[n
where

x2

,-,

(1

..

integratedto yield

Ci is a

where

^2^"
cc,Y]p
(1

[n

/(I)

gQ

_
-

^//^^

integration.

the

boundary

conditions

f{p) is

/(?")"."
A^

,.a.

..

{A

ly

where

(1

^=.

(1

a,r

-^y

(1

",)=

and

^
1

(ai

+a.)/2"^^-

/(O)

0 and

510

READINGS

Example: Taking
occurring in (19).

0.75,aa

ai

0.80, as before,calculate the

(0.20)^

from

(19),
260.42
ifr^
Kp)

f9(^^

(20)

be

f(p)in

compared

of 0.05 for p

difference

Two-Armed

reward
xi

of i2
New

out

given

If R

follows

trial,the

new

with

a2p

"

"

"2

nonreward

if L and

reward

if L and

nonreward

in

[1].

brieflyon

and

reward

if -R and

results represent
discussed

21

grid points

with

those

and

third-

whichever

and

L,

follows

reward

If p is the
follows.

is

with

as

occurs

probability
ability
prob-

Probability
of happening

]" R

a2P

and

desk

probabilityfor R

ai

aip

These

with

probabilityts

for R
+

and

occurs,

probability

aip

they may
mate
approxicalculators,the

the various

Among

can

responses

follows.
reward

occurs,

on

two

are

nonreward

; if L

2, where

equally easilybe applied to the


bandit
problem with
appropriate to the two-armed
each arm
experiment).
(or the equivalent T-maze
on

model

or

0.75.

Bandit

differential equationapproach

partialreinforcement
Suppose that there
a

far obtained.

so

equations using

by the simultaneous
interpolation.

general

more

in Table

shown

are

0.25,0.50,and

yields results in closest agreement

The
The

best values

easily carried

be

can

differential equation method


obtained

f(p)for

fi (approx.)

the best values

which

methods

of

the values

intervals

with

p)^""

^^f;^

Pi

of

(2.7778

Using (20),calculate

Values

constants

^"^'^^^'

"

(0.25)^

Hence,

PSYCHOLOGY

MATHEMATICAL

IN

specialcase
p. 287

in the

occur

wip

(1

iri)p

"

^2(1
(1

"

p)
^2) (1
"

"

p)

presented in ([1],
p. 118, 286)
paragraph followingequation (13.22)

of those

512

READINGS

of

Twery and Bush


/(0.50)for two-armed

made

bandit

combinations

used

calculate

the value

the stated

parameter

For

for Two-Armed

series of Monte

experiments
The

of a-values.

various
to

PSYCHOLOGY

Calculations

Carlo

Monte

MATHEMATICAL

IN

Carlo
with

of

case

/(0.50)from

of

Bandits

ai

ti

calculations

0.75, ^2

0.90, 0:2

"

on
=

lUiac

0.25 for

0.95 will be

(23).

values.

(0.75)(0.10)^
+ (0.25)(0.05)^
=

(0.50)[(0.10)' (0.05)']

2.1667,

+
1

Hence, (23) in

this

65015.7

From

f(p)
this

(2.1667

65006.6

formula,
0.977,

The

of alpha

p)"'^'

/(0.50)
compared

14.3333.

becomes

case

(24)

(1.85/2)

with
values
values

Twery and Bush's result,0.970.


of /(0.50),calculated from
(23) for the various combinations
shown
in Table
and Bush, are
3 along with
used by Twery
3

TABLE

Comparison

of

With

Obtained
At

Those
from

the

Differential

800th

of

the

Twery

And

Level

Probability

Mean

Trial

{second

Bush

and

for

IT,

Various
1

a,,
"1'
-

(first

Results

Equation

of
a,,
"2'

0.75

entry)

entry)
100
for

Sequences
p

0.5

FREDERICK

the

Monte

obtained
were

in

run

The

100

result

AND

obtained

by

random

numbers.

The

trial 800.

at

sequences

MAURICE

100

it has

The

between

agreement

the Monte

numbers

of 800

sequences

were

trials each
of p

value

average

variation

random

some

pre-asymptotic to the extent that 800 trials


agreement is quite encouraging for the use

method.

Their

entry itself is the

Thus

513

TATSUOKA

authors.

these

in which

pseudo-experiment

with

for the
is

Carlo

MOSTELLER

and

is not

an

of the

differential-equation

Carlo

infinite number.

results and

the differential

used
equation is surprisinglyclose,consideringthat only 100 sequences
were
that the differential equation is only an
approximation. On the other
hand, both learning parameters are near
unity in these examples; in that
neighborhood the differential equation should be quite a good approximation.
and

T-maze

Experiment

In

the

Wilson

first section

[2]using

for response
and
and
a

was

when

in which

they
(estimatedfrom

probabiHty for response R


varied considerablyfrom
trials)
0.496,
of p approximately

nearly 0.50. Bush

or

followed

(25)

7/

This

initial distribution

attracted

by

R. The

relative

Wilson

the
=

the

of reward

rate

of

to be

results

on

symmetrical

Beta

0.75
0.75

0.916

through
initial

side. The

the first 10 of the 140

another, the

report that

ai

the reward

see

unrewarded

and

was

model, tti

our

estimated

were

fish to

one

and

The

the fish could

chose

Bush

experiment by

L. In the notation

learning-rate
parameters

divider

transparent

T-maze

described.

0.25 for response

for the group

0.942

ttz

R and

Fish

of this paper,

paradise fish

0.25. The

TTj

with Paradise

value

average

the

being

initial distribution

distribution

3.61[p(l -p)f-\
was

areas

used
under

to

calculate

the

curve

the

expected

(25) in

the ten

fraction
intervals

[0,0.1],[0.1,0.2],"".,[0.9,1.0]
found, the

were

and
In the

to

f(p) at

the

midpoints

was

obtained.

Wilson

found

of these intervals
The
15

result

was

of the 22

lated,
calcu-

were

f{p)

0.800.

fish in the

perimental
ex-

This
after about
100 trials,
making nearly all R responses
for
0.68
the proportion ultimately attracted to the R
result is only about one
standard
from the fitted
error
away

the estimate

response.

value

of

weighted average
experiment, Bush and

group

leads

values

their

That

0.80. That

of the
unreliability

small

deviation

does

originalestimates

not

even

take

any

account

of the

of the a's.

Other Methods
Several
One

/(I

that
"

p;

other methods

of

approximating the function have been explored.


rather
successful
was
employed the function f(p; a, 0) or
the iterate change very
0, a), choosing a value of a that made

514

READINGS

little.

This

fo(p)

method

MATHEMATICAL

IN

superior

was

PSYCHOLOGY

to

iteration

an

technique

p.

Since

special

knows

one

case

a^

the

exactly
the

a2

notion

solution
of

to

the

expanding

functional

f(p;

in

tta

the

as

the

series

power

of

neighborhood
developed

note,

in

equation
az)

ai

in

with

beginning

such

itself.

suggests

ax

Robert

R.

in

Bush,

an

published
un-

technique.

REFERENCES

[1]

Bush,

R.

R.

and

Mosteller,

[2]

Buah,

R.

R.

and

Wilson,

1956,

51,

[3]

Harris,
in

[4]

R.

E.,

Bellman,

processes.
S.

Some

R.,
Res.

random

and
Memo.

walks

725-756.

Manuscript

Revised

Stochastic

models

Two-choice

for

behavior

of

York:

New

learning.

paradise

fish.

Wiley,
/.

exp.

1955.

Psychol.,

315-322.

T.

decision

Karlin,

F.

T.

received

manuscript

1/9/59

received

6/29/59

Shapiro,

H.

arising

N.

The

P-382,
in

Studies
RAND

learning

in

functional

Corp.,
models

Santa
I.

Pacific

occurring

equations
Monica,
J.

Calif.,
Math.,

1953.

1953,

3,

THEORY

OF

DISCRIMINATION
FRANK

LEARNING

RESTLE

Stanford University^
This

presents

paper

two-choice

similar

Though

of

theories
and

(5)
this

theory of
learning.

discrimination

form

in

earHer

to

simple learning by

Bush

"relevant"

Estes

introduces

(2,3),
powerful

definite
assumption which makes
tain
quantitative predictions easier to oband
Several such
test.
tions
predicfer
dealing with learning and transderived
from
the
are
theory and
tested against empirical data.
The
stimulus
situation facing a subject
new

in
is

trial of discrimination

of

thought

as

of

set

be used

can

be obtained.

by the

how

or

is

cue

ject
sub-

reward

For

is

if food

example,
black
card
always found behind
a
in
rat
a
experiment, then
cues
aroused
vant.
by the black card are releA
aroused
cue
by an object
to

is

uncorrelated

with

For
is

always
black

from

cues.

"irrelevant."

predict where

to

the

ing
learn-

or

if it

relevant

Mosteller

and

system

In problems to be analyzed by this


theory, every individual cue is either

behind
card

left

reward

example,

to

the

is "irrelevant."

if the
black

is

reward

card

but

randomly
right, then
"position"
moved

These
irrelevant.
cues
are
correspond
concepts
discussed
Lawrence
are
by
abstract,
(6).
or
thing
any
scription In experiments to be
deconsidered,
present, past, or future, of any
which
the subject can
the subject has just two
choice reto
sponses.
subset

of these

to

cues

may

concrete

"

"

learn
In

make

to

differential

this definition

whether

the

learn

of

learned

the

used

way

the

capacity

of

set

to

is

cue

in

to

set

sidered
con-

the
Theory

parts of it.
"cue" will occasionally

refer to any

are

testing the theory. Any


consistent method
of describing these
be applied
which
two
can
responses
throughout a complete experiment is
acceptable in using this theory.

cannot

responses

term

all of which
same

the

different

to

Informally, the
cues,

to

other activities

in

matter

"indivisible"

as

different

be

not

individual

An

that

sense

he has

as

one.

thought
be

response

long

as

No

response.

subject actually makes

differential
cues

it does

of

manipulated in
during a whole experiment.
are

In

tion
solving a two-choice discriminathe
problem
subject learns to
relate his responses
correctly to the
relevant

cues.

responses

irrelevant

become
cues.

At

the

same

time

his

independent of the
These

two

aspects

of discrimination

This

Ph.D.

paper

is

adapted from

dissertation

submitted

to

part of a
Stanford

debted
author
is especially inUniversity. The
to Dr. Douglas H.
Lawrence
and to
Dr. Patrick Suppes for encouragement
and

criticism.

Thanks

Estes

loaned

who

are

also due

Dr.

W.

K.

prepublicationmanuscripts
and
Dr. R. R. Bush
who
pointed out some
the present theory and
relations between
the
model
Bush- Mosteller
(3).
*

Now

at

OfiRce,The
This

the

Human

Resources

Research

c{k,n+l)=c(k,n)-\-e[_l-c(k,n)'][1]

George Washington University.

article appeared

in

learning are represented


by two hypothesized processes,
"conditioning" and "adaptation."
Intuitively,a conditioned cue is one
how
which
the subject knows
to use
If ^ is a relevant
in getting reward.
and c(k,n) is the probability that
cue
conditioned
k has been
at the beginning
of the wth trial,then

Psychol. Rev., 1955, 62,


515

11-19.

Reprinted

with

permission.

516

READINGS

MATHEMATICAL

IN

PSYCHOLOGY

fraction of unadapted cues


probability that it will be conditioned
adapted
of
each
the
the
trial.
next
on
by
beginning
The
trial. On each trial of a given probperformance function
lem
p(n),
constant
a
proportion, d, of unconditioned
representing the probability of a correct
the
relevant
becomes
nth
is
cues
on
trial, in
response

is the

accord

conditioned.
To

the

that

extent

to

correct

unconditioned

an

relevant

equally
incorrect

an

only,

response

to

cue

tributes
con-

and

correct

The

the

is in the form

total number

in the denominator

cues

of conditioned

to

the number

times

response.

given above.
of a ratio,
of unadapted

adapting

function

with

whereas

tioning
the definitions of condi-

and

conditioned

afifectsperformance, it contributes

cue

with

the

and

plus

cues

of other

ber
num-

one-half
in the

cues

Thus

conditioned
numerator.
cues
Intuitively,an adapted cue is one
contribute their whole effect toward
the subject does
consider
a
not
correct
in deciding upon
his choice response.
adapted
cues
tribute
conresponse,
either reis thought of as a "possible
nothing toward
If a cue
sponse,
and other cues
contribute their
solution" to the problem, an adapted
effect equally toward
and incorrect
correct
is a possible solution
which
the
cue
Formally,
responses.
subject rejects or ignores. If a{k,n)
is the probability that irrelevant cue
k has been adapted at the beginning
Ec(M)+IECi-c(^,w)]
of the wth trial,
then

which

a{k,n+\)=a(k,n)-]-d[\-a{k,n)'][2]

-hhJ:ii-a(k,n)-]

P(n)

[4]

r+j:Ll-a(k,n)-]
is

probability that

the

it will

be
r

adapted

by

trial.

next

On

trial of

each

in

the

an

incorrect response.
It will be noticed

relevant

that

it

Some

to

the

cues

is the

[3]

-\-i'

in the

number

of

cues.

i is the

Thus,

ber
num-

the

fraction of unconditioned

conditioned

on

each

cues.

Regarding

subject is

naive

the beginning

at

of training,so

vant
that for any releand
for
k, c(k,l)
0,
any

cue

irrelevant
receives

can

cue

by

k, a(k,l)
0, and if he
on
a
given problem,
=

trials

mathematical

be shown

induction

it

that if k is relevant,
=

(1

e)"

and

[5]

if k is irrelevant,

6 is the

relevant

in the
cues
proportion
problem. This proportion is the same
as

taken

of relevant

problem and

of irrelevant

Learning

c{k,n -f 1)
r

sum

Consequences

If the

same

then

where

is the

the

tributes
con-

nor

over

2Z

Simple
that

taken

functional
non-

equations
2.
1 and
The fundamental simplifying
this
of
assumption
theory deals
with 6. This assumption is that
e

and

cues

the i irrelevant

in both

6 appears

constant

sum
i

over

is

cue

is the

correct

23

Here

becomes

cues

sense

to

the

given
proportion of

problem a constant
unadapted irrelevant
adapted. An
adapted
neither

of

beginning

the

trial,and

cues

the

a(k,n +
Under

1)

these

substitute

equation

(1

circumstances
5

and

d)\
we

[6]
can

into

equations
and, taking advantage of

FRANK

the simplifying efifects of equation 3,

517

RESTLE

An

Empirical

have

we

Test

Learning

of

Combination
Consider

three

53, all of which

shows

that

Plotting equation
p
an
S-shaped function of n with an
(for0 " 0) at 1.00. Also,
asymptote
p{l)
|. Since p{n) is a monotonic
mate
estican
increasing function of 6 we
is

from

of

observations

If

we

want

know

to

theoretical proportion of relevant

cues

Cues
Si, S2, and

problems,

involve

the

vant
irrele-

same

of the

problems, 5i
entirely separate and
different relevant cues, while in problem
and

have

52,

all the relevant

53

and

know

di and

since

cues

relevant.

present
ri -\-^2 and

are

formance.^3
per-

the

"

of

Two

cues.

"

Simple

the

Theory

ii

62

i^

we

by equation

of Si and
That

52

is,

i^. If

we

63,

compute

can

problem for a particularsubject,


Bi
YxKry + i)
have the subject work
the probwe
on
lem,
02
r2/(r2+ i)
record
his performance
curve,
and
solve
for
03
{ri+ r2)/iri-\-r2-^i).
6.
This
equation 7
the simplifying
result depends directlyupon
Solving these equations for ^3 in terms
assumption of equation 3.
of 01 and 02 we
get
Since
the instabilityof individual
makes
it difficult to
learning curves
03
(01+ 02- 20102)
/{I
0A).
[9]
in

fit curves
6

to

them, it is fortunate

be determined

can

in

the

can

This

dififerentway.

E errors
in
subject makes
of solving the problem to
it is
a
rigorous criterion and
very
assumed
for practical purposes
that
he has made
all the errors
he is going
make.
to
Theoretically, the total
made
number
of errors
on
a problem

Suppose

that

course

theorem

errors'

made

are

differential
used

to

errors

cue

learn

used

know

in

learning

how

how

many
to

in

are

many

learning a problem

either X
and

use

many

Y, then how

will be made

(if X

following

we

and

cue

in which

the

answers

question : Suppose

or

be

can

crete)
entirely dis-

are

be written

Eninger (4) has run


which tests equation 9.
of white

rats

were

experiment

an

Three
in

run

groups

maze

n=l

successive

on

Under

the conditions

satisfyingequation
be evaluated
can
mately
approxitime
by using the continuous
variable t in place of the discrete trial
variable
and
integrating. The
n,
result of this integration is that

The

discrimination

first group

7, this

^^l

i
(1

6) log (1

ey

[8]

black-white, the

number

of

errors

made

relates

d, it is possible to make
stable estimates of 0.
to

second

group

and
the
third
tone-no-tone,
had
available
and
both
cues

an

auditory

tion,
discrimina-

relevant.

rigorous
used

are

each

was

group

criterion, total
to

estimate
The

run
error

0i and

values

to

scores

02 by equation

estimated

are

the total
'

on

crimination
dis-

learned

8.^

By equation 8, which

visual

group

Since

log 9

learned

problems.

Total

in

do
not
error
scores
appear
problem
are
no
Eninger's original publication and
relatively longer known.
trials-to-criterion
However,
scores

were

reported. Total

error

scores

were

518

di

READINGS

.020, based

average

made

errors

and

auditory-cue problem,
based
64.5

these

Putting
9

values

two

.029,

"

of

average

the visual-cue

on

errors

the

on

62

estimated

an

on

estimated

an

on

of 98.5

tion
equa-

thereafter

.029

(.020)(.029)/
1
(.020)(.029)

.049.

This

of 63 substituted

value

8 leads to the

equation
about

33 total

errors

expectation of

adapted

an

irrelevant

new

wise,
Like-

appearing as
new
problem

cue

in

cue

if

However,

is made

an

is

conditioned

irrelevant it is

obviously

longer conditioned, since it cannot


larly,
Simias
a
serve
predictor of reward.
it is assumed
that if an adapted

no

is made

cue

relevant in

it becomes

problem,

new

and

unadapted

available

According to the present definition


conditioning, a conditioned
cue

of

average

in

cue

for conditioning.

the combined

In fact,an

problem.

cues

on

into

appears

relevant

in

immediately

problem, it is stillconditioned.

adapted.
.020

as

is conditioned

cue

problem and

cue

^3

that if a

assumed
one

problem.

into

get

we

PSYCHOLOGY

MATHEMATICAL

IN

of

jects
by the four subcontributes
to
correct
a
The
tion
predicresponse.
will
Therefore
the
above
is not very
assumptions
accurate.
However,
the
if
hold
relation
between
not
a cue
only 14 animals were
employed in the
is reversed
in changing
of five, and a reward
entire experiment, in groups
This theory cannot
the problem.
be
five,and four. Individual differences
used
reversal
and
to
animals
analyze
learning,
within
were
among
groups
is applicableonly in cases
in which
is taken of
If account
considerable.
26

errors

made

was

this problem.

on

single- relevant cues maintain


significance.
If two
problems are
prediction is

sampling variabilityof the


cue

groups

group
not

of subjects, the
significantly
wrong.
is needed

whether

Further
to

the proposed law

conditions

same
perimentation
ex-

and

determine
is tenable.

It is easily seen

that 63 will always


larger than di or 62 if all three
problems are solved.
Learning will
always be faster in the combined-cues
problem. Eninger (4) in his paper
ment
points out that this qualitativestateis a consequence
of Spence's
theory of discrimination.
However,
Spence's theory gives no quantitative
be

and

differ

(as where
white

and

Transfer

of

apply this theory


transfer-of-training
experiments in
which more
than one
problem is used,
certain assumptions are
made.
It is
to

estimated

from
trials-to-cr iterion scores
by
using other, comparable data collected by
Amsel
(1). Dr. Amsel
provided detailed
results in a personal communication.

same

the

ratus,
appa-

the discriminanda

problem

one

the other

is

dark

black-

gray-light

discrimination), it is assumed
that both problems involve the same

gray

cues

be

the greater the difference to


discriminated, the more
cues
are
; but

relevant

and

Empirical

Training

to

under

run

in the

the

Tests

less

of

of-Training

order

unchanging

only in the degree of

difference between

law.

In

an

two

of the combined-cue

and

As

Lawrence

it seems
is

that

Transfer-

the

Theory

(7) has pointed out,


a

difficultdiscrimination

easily established

more

irrelevant.

are

are

firsttrained

of

the

same

on

type

an

if the subjects
easy

than

lem
probif all

cult
training is given directlyon the diffiThe
discrimination.
tal
experimenevidence on
this point raises the
question of predicting transfer per-

520

ance
relativelyaccurate, though performis higher than predicted.
also considered the possibility
Lawrence
that

TABLE
Prediction
Rats

Transfer

of

After

Series

Performance
of

of

Pretraining

Problems*

from

gradual transition

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

lems
through successivelyharder probin
result
would
rapid mastery of
He tested this
the difificultproblem.
another
by
giving
group
proposition
of
three
series
of subjects a
pretest

easy

problems before the final test problem.


ing
problems in order of ease of learn-

The

the problem learned by


first,

were,

1 with

No.

ATG

^i

.14,

which

problem

mediate
inter-

an

wise
other-

not

was

Data

Lawrence

from

(7).

used, the difficultpretest problem


with

6z

problem

with

di

test

.04.

02 in Lawrence's

estimate

To

finally the

.07, and

ing procedure, subjects,


equation adopted is

ment
experi-

was
problem ^2 never
separately in simple learning,

.09881ogio(.4J)

The

etc.

[11]

where

used

discriminanda

between

whose

Si

problems

is

are

6 is

data, made

available

with

and

cues

It

zero.

that this assumption, along

found

was

foot-candles,there

zero

relevant

no

are

the stimulus

properly controlled,and
difference

Si, S3,

known.

are

if the

that

know

We

values

it

possible

tentative empirical function


6
relating to the difference between
to write

foot-candles.

in

discriminanda

This

holds only in the


apparatus, train-

equationpresumably
of Lawrence's

case

TABLE
The

Relation

STIMtTLl"

AND

2
Between

"Difference

of

VaLUE

OF

in foot-candles.

that

in apparent

for problems

foot-candles
and

d is the difference between

ences
differ-

the relation of d to

notice

we

where

PROBLEM*

theoretical

criminanda
dis-

It is

this equation has

and
significance

phasized
em-

no

is

merely
equation 11 it is
expedient. From
the 6 value of
possible to determine
the intermediate
pretraining problem
by interpolation. Table 2 gives the
data and results of this interpolation.
Ten trials were
given on each of the
first three problems and
fifty trials
the final test problem.
on
Using the
2 it is possible to
in Table
6 values
predict the test problem performance
have
of subjects who
through
gone
gradual transition pretraining.^ This
predictionis compared with observed

performance

in Table

noted

the

that

case

very

be

It may

tween
be-

correspondence

predictionand
in this

3.

close.

observation

Again,

is

ever,
how-

the prediction is consistentlya


than observed
little lower
ance.
perform-

The

generalpredictionfor transfer through

series of problems which


get successively
be derived by following
difficult can
more

*
**

Data
from
Estimated

Lawrence
(7).
by interpolation from

empirical

16.

t Theoretical

"

see

text

for

explanation.

tion
equa-

note
through and repeating the reasoning in foot4.
Since the resultingequations are
rather
be derived
extremely large and can
easily,they are not given here.

FRANK

worked

Data

New

521

RESTLE

to

we

theory has thus far been tested


Its
of rats.
against the behavior
tested with college
generahty is now
in a simple discrimination
students
learning task.
The

high criterion

assume

can

in

pretraining,
p{n) is

that

the
at
negligibly different from one
of pretraining. Then
end
by equation
7 we
that (1
see
0i)"~^is small,
and equation 10 simplifiesto
"

Subjects and procedure. The subjects in


p{n+j)
.[12]
in the ele23 students
mentary
02+(l-^2)'-H^l-02)
experiment were
versity.
Uniat Stanford
psychology course
This theoretical function of j is compared
The
seated at one
end of a
5 was
=

this

table

"5".

or

On

singlestimulus, which
white

circular

used

squares
In

size.

could
responses
each
trial 5 saw

his

told that

and

either "^"

problem

the

Si

on

two

trials differed in

at

For half the


the

problem

"5"

and

square
^s had

to

was

the

the

Stimuli
A
ten

trials and

the

correct

The

law

well

as

as

then transferred

the

the

same

to

The

he

what

problem

Transfer

"Easy-Hard

5s

The

other

then

transferred

11

^s

of task.

was

thought
was,

of the

and

Certain

sort.

same

in

relevant

are

This

si.

52

responses
and run
to

made

up

called

Group"
trained

were

to

correct

problem

These

criterion.

first
was

on

the

the
EH.

S2

the

and

which

These

one.

in

the

hard

to

be

The
Group" called HE.
were
approximately equated for
known
specialvisual skills.

two

groups
sex, and

age,

proportion

of

estimated

group,

relevant

the

age
aver-

cues,

di,

.254

by equation 8.
the
Using
pretraining performance of
the HE
the average
proportion
group,
of relevant
in problem 52 was
cues
was

estimated
The

at

at

62

.138.

transfer performance of group


first learned the easy and

EH,

which

then

the

hard

by equation

problem,
10.

Since

is

be

harder

identified

transfers from

be

must

cues

when

Therefore,

Prediction

Using the pretrainingperformance


EH

the

problem. For performance


perfect in the easier problem

all relevant

Easier

of the

in

cannot

cues

problem

easy

irrelevant

were

cues

the

the hard

to

fied.
identi-

subject

the easier

"Hard-

Easy Transfer

Results.

be applied
performance

can

Using the line of reasoning which


can
developed
equation 10 we
duce
prosize discrimination.
a
to
an
equation
predict transfer
alternated
domly.
rancalled after each
performance from hard to easier problems

criterion of 15 successive

and

the

other

curred
possible solutions which had ocof questioning
This method
to him.
is a modification of Prentice's method
(8).
5s were
trained first on problem 5i
Twelve
a

on

confirmation

rat

outline

to

to

based

rence's
predicted Law-

This

that the

this type

the smaller

to

error.

is

also

data.

rat

on

asked
to

which

"A"

were

was

solution

formula

experimental group,

period was

rest

negligibleconstant
This
prediction

human

largerone.
problem.
problem was

converse

to

the

to

told that

never

say

formance
per-

that

seen

the correspondence is quite close with

suggests

6 ft.

5s in each

transfer
It is

4.

viewed

were

squares

of about

distance

The

3 in.

was

observed

in Table

differed in

squares

height by \ in., in problem 52 they differed


height of each pair of
by \ in. The mean
squares

with

square

The

background.
alternate

on

black

was

be

predictable

these subjects

of

TABLE

Transfer

of

to

Human

Harder

Training

Problem

Subjects

from
in

522

READINGS

TABLE
Prediction

to

MATHEMATICAL

Transfer

of

Harder

IN

of

Easier

Training

Problem

from

these elements need not be of the


of "pointsof color" or "elementary
nature
tones."

in

Subjects

Human

PSYCHOLOGY

learn

If

subject can

consistent response

to a certain

configurationdespite changes in its


constituents, then

by definition

the

configuration

separate from
its constituents. The intention is to
strated
accept any cue which can be demonis

to

be

cue

possiblebasis

for

differentialresponse.
of conditioningdescribed
process
is formally
in this paper

The

problem
number

small
should expect some
of errors
On the
to be made.
we

similar

to

the

processes

of

tioning
condi-

(5) and Bush and


(2,3). In the present

of Estes

assumption that the hard problem Mosteller


was
completelylearned in pretraining, theory conditioning takes place at
not
only on "reinforced"
the formula for transfer performance each trial,
the easy

on

conditiontrials. In earlier theories ing


is said to occur
forced
only on such rein-

problem is

e2-\-(di-e2)(i-diy-'
=

where
cues

nation
trials. In two-choice discrimi-

[13]

P(n-\-j)

di is the proportionof relevant


problem and 62 is
easy

in the

the proportionof relevant cues in the


harder problem. The
proof of this
theorem is similar to that of equation

the incorrect response


has a
cause
behigh initialprobability
(one-half)
of the physical
of the nature
the way

of recording
Therefore, a theory of
for
learningmust account

situation and
responses.

two-choice

the consistent weakening of such responses


above, and is not given here.
through consistent nonreinEquation 13 yields the prediction forcement.
for transfer performance of the HE
The notion of adaptationused here
subjects. In Table 5 the prediction is formallyanalogous to the operation
12

compared with
performance.
is

Despite the

very

observed

transfer

and

of Bush

Mosteller's Discrimination

Operator
small

"Z""

(3). However,

frequencies whereas

Bush and Mosteller's operator


tion
predictedand observed, the predicis appliedonly on trials in which the
is quite accurate.
In all,seven
reward condition is reversed for a cue,
made by eleven subjects,
errors
were
the present theory indicates that this
whereas a total of eightwere
expected.
takes place each trial. In
process
This is an
of .64 errors
average
per
while
the Discrimination

addition,

subjectobserved,and .73 predicted.

Operator and the


both

are

Discussion
The

definition of

Bush

and

process

of adaptation

exponentialin form.

Mosteller introduce

new

k for this purexponentialconstant


pose
in terms
the
and
the
uses
theory
present
is selected because

"cue"

of possibleresponses
the theoretical results do

not

conditioningconstant 6.
the
The major point differentiating
earlier
from
similar
present theory

the nature
of
depend critically
upon
the stimulating agent.
While
cues
are
thought of as stimulus elements, theories

is the

use

of the strong sim-

FRANK

plifying

assumption

exponential

identifying

of

relevant
may

with

constant

the
portion
pro-

This

the

of

sixth

range

within

was

reasonable

the
tion.
devia-

sampling

sumption
as-

intuitively

appear

and

rate,

the

cues.

523

RESTLE

likely,
un-

REFERENCES

but
further

if

should

it

experiment

predictive

be

be

to

of

power

shown

by

tenable,

the

1.

a.

Amsel,

of

learning

is

theory

enhanced.

be

to

useful

so

no

discriminanda

assumption

an

results

abandoning
unless

2.

Bush,

R.

R.,

1952,

model

Psychol.
3.

Bush,

R.

for

J.

1951,

stimulus

simple
58,

matical
mathe-

learning.

313-323.

F.

Mosteller,

"

comp.

341-346.

F.

for

Rev.,
R.,

function

45,

Mosteller,

"

mental
experi-

it.

require

Psychol,

visual

as

durations.

There

for

reason

learning

discrimination

physiol.
seems

of

Rate

brightness

discrimination

generalization

model
crimination.
dis-

and

Summary

two-choice

4.

theory

earlier

of

Mosteller

been

but

to

and

summation

U.

learning

physiol.
5.

Bush

Estes,
of

differs

M.

selective

similar

(5)

Estes

(3)

Eninger,

presented.

formally

is

theories

and

Habit

1951,

58,

tion
discrimina-

has

learning
The

Rev.,

413-423.

of

theory

Psychol.

Psychol.,

W.

K.

1952,

Toward

learning.

in

problem.

/.

45,

511-516.

statistical

Psychol.

comp.

Rev.,

theory
1950,

57,

what
some-

94-107.

in

basic

new

From
laws

uses

6.

D.

of

three

theory

derived

empirical
with

dealing

one

in
exp.

the
7.

combination

Lawrence,

assumption.

this
are

and

concepts

simplifying

of

relevant

and

cues,

H.

Lawrence,

Selective

stimulus

constant

1950,

Psychol,
D.

H.

of

with

quantitative

of

type

The

laws

transfer

four

groups

of

permitted

rats

these

six

human

predictions

and

subjects.
were

of

crimination
dis-

continuum.

havior
be-

Psychol,

8.

Prentice,

W.

C.

learning.

1952,

Continuity

H.
/.

two

Five

quite

physiol.

J.

45,

511-

516.

the

39,

of

groups

transfer

of

predictions

/.

175-188.

two

comp.

These

training.

of

special

association
situation.

40,

along

dealing

ness
distinctive-

Acquired
II.

cues:

exp.

187-194.

of
accu-

(Received

January

14,

1954)

Psychol,

in

man
hu-

1949,

Part
L.

'

LEARNING

DISCRIMINATION

BY

IN

RESPONSES

OBSERVING

OF

ROLE

THE

I
WYCKOFF,

BENJAMIN

JR.

University of Wisconsin

in the

Theorists

literature, largelybecause

the

in

tion
of discrimina-

area

often had

learning have

became

occasion

necessary

which
situations
the
to
of 5
refer to a set or predisposition
to a parto learn differential responses
ticular theory is intended to apply.
Such
of
stimuli.
a
pair
disposition
pre-

Spence's

to

discrimination states

Spence'stheory of

attributed

has often been

it

clearly

delimit

to

are
stimulus-response connections
tending
reaction of 5 such as an atto some
nation
strengthenedor weakened during discrimiorienting response,
response,
the same
trainingin essentially
way
tional
perceivingresponse, sensory organizationing
would
these
occur
during condias
changes

implement

To

activity, etc.
the discussion

that

of the role of such

or

actions
re-

learning we
sponse"
"observing re-

in discrimination

shall adopt the

(Ro) to

term

refer to any

all

aspects of the

will be

which
of

probabilityof

The

observing

will be denoted

response

These

po.

the

reinforcement

which

are

responses

from

of

occurrence

to

be

is not

the

at

stimulus

situation

pinging
im-

time

weakened

reinforced.

when

the

response

Certain implicationsof

questionedby Krechevsky
theory were
the
and became
(11) and other theorists,
nuity"
subjectmatter of the "continuity-disconti-

this
an

by

tinguished
dis-

controversy.
reviewed a number

upon

responses

response is
it and

between

the response
curred
ocon
will be strengthened.These connections
S

response

to the pair
results in exposure
stimuli involved.
discriminative

extinction. When

reinforced the connections

is based; that is,

running, turning right or left,lever


etc., which, for convenience,
pressing,

need

not

This
of

repeated in

be

material has been


times

(2, 5) and

detail here.

One

aspect of the controversy is pertinentto


the present discussion. Krechevsky (12)
.

cated
responses." presentedexperimentalfindingswhich indiwith
that
learned
rats
nothing
respect
Spence (19) has proposed a theory
stimulus patterns during the first 20
of discrimination which is specificallyto two

shall term

we

intended

to

"effective

observing response

no

discrimination experiment even


a
reinforced
systematically
though they were
ing
for approaching a particularpattern dur-

trials of

deal with situations where


is required of S,

that is to say, to situations in which 5


tablished
esthis interval. Failure to learn was
criminative
is certain to be exposed to the disof
interference
lack
a
showing
by
each trial or
stimuli on
criminati
reversed distested on
5s were
when
a

(po
prior to each effective response
tion
discrimina1). The fact that in some
experiments this condition has

These

been satisfied has become

not

an

in which

issue
used

is submitted

This

paper
of the

in

ment
partialfulfill-

requirements for the degree of


Doctor
of Philosophy, in the Department of
writer
University. The
Psychology, Indiana
to

Burke

for his invaluable

This

express

article

his

Here

were

terferenc
in-

obtained, indicatingthat

learninghad occurred in
the earlyportion of the experiment.
these results,
In interpreting
Spence (20.
that
the stimuli (patterns)
p. 277) argued
not
suflSciently
used by Krechevsky were

appeared in Psychol. Rev.,1952, 59,


524

as

weights
differing

discriminative stimuli.
was

some

to Dr. C. J.
appreciation
lation.
guidance and stimu-

wishes

in apparent

findingswere

disagreementwith the data obtained by


and Pratt (13) in a similar exMcCuUoch
periment

cumulative

431-442.

Reprinted with

permission.

L.

WYCKOFF,

BENJAMIN

to providea legitimate
test of
conspicuous
that
had
He
5s
not
suggested
theory.

will be

his

learned to orient toward


trails. He

the first 20
such

cases,

"...

the stimuli within

pointsout

the animal

orient and fixate its head and


receive the criticalstimuli."

must

525

JR.

made

tensive
exto develop a more
theory of discrimination which

will include situations

that in

learn to

eyes so as to
He then gests
sug-

in which

some

referred
observingresponse (hereafter
is
5
is exbefore
posed
to as Ro)
required
to

the

discriminative

stimuli.

example of such a situation would


in which
this learningmay
a
way
be
an
experiment in which stimulus
reactions are learned
"These
occur.
In this
cards
were
placed overhead.
because they are followed within a short
the response
of raisingthe head
sponse." case
temporal interval by the final goal rewould be the Ro).
This interpretation
was
If we accept the notion that changes
experiput to an mental
.

test by
the
exp"eriment

Ehrenfreund

(5). In his
in po can
be accounted for within the
receiving framework
of reinforcement learning
manipulated by
to devise
theory,it should be

likelihood of 5's

critical stimuli

the

An

was

possible

right a
changingthe positionof the stimuli (uptheory of discrimination which will
with respect
and inverted triangles)
where
include those cases
Ro is
some
to the landing
platformof a jumping stand.
The purpose
of this paper
necessary.
The
tially
designof the experimentwas essensuch
shall
is
outline
to
a theory. We
the same
as
sults
Krechevsky's.The rediscrimination
that
see
by analyzing
conform
to Spence'sinterpretation.
When
the stimuli were
placed relatively learningin this way it will be possible
for stimulus generalization
high,no learningoccurred within the first to account
ing
durwhereas
when
40 trials,
and also changes in generalization
the stimuli were
ing
discrimination learning without
placedcloser to the landingplatformlearndid occur.
ured
tween
Learningwas again measpostulatingany direct interaction bein terms
of

ing
of interference in the learn-

subsequentreversed discrimination.

stimuli.

Several

hypotheses

will be derived from this theory which

have been tested in an experiment by


ations
analysisof discrimination situwhere
the author presented in detail elsein w^hich some
sponse
observing re(22). Finallywe shall outline
is required is of interest for
in which the present theory can
a way
several reasons.
First,discrimination
with existingquantitabe integrated
tive
in
atory
learning situations other than labortinction
theories of conditioningand exexperiments, such as human
theory
to form a quantitative
of every
day
learningin the course
of discrimination.
ondly,
Secevents, is largelyof this kind.
To
simplify this discussion let us
in
the most
trolled
even
closelyconconsider a hypotheticalexperiment
laboratory experiments it is
using a situation similar to that used
seldom, if ever, possibleto say with
by Wilcoxon, Hays, and Hull (21),
criminative
certaintythat 5 is exposed to the discriminatio
and later used by Hull (10) for a distive
stimuli priorto each effecperiment
experiment. In this exIn the case of pattern
response.
in a small
rat
was
a
placed
strated
demondiscriminations it has been
with
a
single exit
that
tively compartment
relaEhrenfreund
(5)
by
ment.
small dififerencesin the position through a door into a goal compart-

The

stimuli will
of the discriminative
indicating
effectdiscrimination learning,
of
fixation
that relatively
precise
the stimulus
In

the

is

required.

present

paper

an

attempt

of the latency of
of running through this
tive
discriminaThe
obtained.
door was
stimuli consisted of a black or a
A

measure

the response

white

door, either

one

of which

was

526

READINGS

present

each

on

color

other color

present.

was

an
present on
of the trials.

was

cent

For

During

of

purposes

the

let us consider

crimination
dissimilarly to

situation in which
stimuli

Each

lus
stimu-

average

of 50

the discriminative

If

each trial,when

lookingup
S

occur

black

will

or

fails to

be

white

When

occur.

When

S will not

occur

Ro does

exposed either
card.

either card, but

rather

to

the Ro

whenever
present

In

not.

or

information
We
the

are

now

the white

card

actually looks

is

if we

sense

between

observing

we

as

can

give

see

rise to

generalizationbetween the
More
specifically,
that po will increase during

assume

tial
learning (differentween
reinforcement),generalizationbethe

sponses
re-

discriminative

decrease.

Similarly,

stimuli will

might

we

that po will decrease


a procedure in which
is reinforced

of

if we

the subject

equally often

either

sume
as-

duce
intro-

in the presence

ential
(non-differreinforcement). This decrease
in po would
give rise to an increase in
the stimuli.
generalizationbetween
In the

stimulus

of the

case

periment
hypotheticalex-

suggested above, generalization


will be

shown

effect between
trials
but

in

"crossover"

positive and

Reinforcements

trials.

on

(positivestimulus

card

negative
positive
present
will

not

necessarily observed)

to

strengthen the effective

tend

sponse
re-

while unreinnegative trials,

on

up

then, S gains only


by making the Ro.
in a positionto examine

relation

changes

involved.

neutral

population of stimuli (walls, floor,


etc.). Note that in this situation S
does not improve its chances
of ultimate
reinforcement
by making the Ro.
The
food is placed in the goal compartment
whether

in

stimuli

be exposed to
to

that po

assume

learningprocesses
that these changes would

5 is

in the apparatus,
there will be
certain
that
the Ro of
probability
will

we

discrimination

placed
a

differential

result of

overhead

if
raising the head, will be necessary
5 is to be exposed to the tive
discriminaOn

learn

in

po increases.

as

rather than directlyin front of S.


In
this
case
an
observing response,

stimuli.

stimuli, or

two

of failure to

terms

cussionchanges
dis-

present

placed

are

the

responses

the

different
slightly

PSYCHOLOGY

readily. Thus we can


see
forcement
reinthat stimulus
degeneralization will crease

present, whereas
withheld
when

was

was

per

trial.

MATHEMATICAL

training the running response


reinforced with food when

was

one

IN

forced responses
tend to weaken

on

negativetrials will

the effective response


If 5"s tendency to

positivetrials.

on

look

during differential

increases

up

and
stimulus
generalization. reinforcement, this "crossover" effect
will decrease.
If during non-differential
In general it is apparent
that if po has
the
reinforcement
to
tendency
be exposed
a low value, S will seldom
look
the
"crossover"
decreases,
up
the
to
discriminative
stimuli

black
will

On

white

have

learn
any

(the
cards). S therefore,
minimum
opportunity to

and

discrimination

discrimination

manifest

the

manifest

hand, if po has a high


opportunity to learn or

decreases

generalization between
stimuli
is usually defined either
of 6"s tendency to respond
terms

emphasized that these


regarding increases and
sumptions
po are, at this point, as-

be

in

which
true

in

may

or

may

not

be

ation.
particularexperimental situ-

We

discrimination will be large.

Stimulus
two

to

It should
statements

the other

value,

in

or

already learned.

effect will increase.

shall present experimental


sumptions
suggest that these as-

findingswhich

are

below.

quite generally

true

528

READINGS

PSYCHOLOGY

MATHEMATICAL

IN

5s' tendencies

to

stimulus

discriminative

stimuli

is exposed to the neutral


population, since, on positive
is reinforced
trials, the running response

when

though 5 does

even

up.

is reinforced

effective response

The

5 is
consistently when
stimulus.
the
positive

most

exposed

to

as

the two

to

the

"degree

of discrimination."

look

not

respond

Earlier it

pointed

was

probabilityof

that the

out

of Ro is

occurrence

one

of the factors determining the rate of


ing
Accordformation of discrimination.

hypothesis the opposite


The
relationshipis also true.
of a circular
the positive and negative
to
resulting picture is one
i
n
which
Ro affects
stimuli will have
net
a
reinforcing interrelationship,
it

Therefore,

effect

still

is

the

postulate that

that, before

expectedto

be

of these

any

of its effect

nisms
mecha-

increase

in po

6" must

learn

occur,

that
differential effective responses,
learn to respond
is to say,
5 must

differentlyto the
stimuli.

discriminative

two

of the "jumping
5 does not have
if
experiment,
In the

cause
be-

of discrimination

formation

the
of both

the present

to

posure
ex-

Rg.

on

It is true

can

plausible to

intermittent

on

criminativ
dis-

to

exposure

stimuli,while the degree


affects Ro

through
involving either

of discrimination
mechanism

another

secondary reinforcement
We

case

changes

or

probabilityof reinforcement.

in the

present four propositions

now

implied by this general


was
hypothesis
differential jumping
ward
tohypothesis. The
of
the
basis
formulated
the discriminative
partly on
stimuli, the
able,
already availwill always experimental evidence
probabilityof reinforcement
which
suggested that these
be 50 per cent, and will not be
true
(22). At presof Rowere
propositions
ent
improved by the occurrence
shall consider them as specific
forcement
we
When
we
apply the secondary reinhypotheses. The first two of these
principlewe can see that
stand"

which

are

tendencies

the positive stimulus

must

temporal relation

the proper

number

in

appear

of times

before

acquire secondary
of
reinforcing properties. In terms
and

Schoenfeld,

Notterman,

to

responses

2. po will decrease

are

high)

under

of

conditions

(or remain

low)

non-differential

reinforcement.

discriminative

secondary reinforcing
acquired by these

before

properties

the

(or remain

forcement.
of differential rein-

conditions

essary
nec-

ive
for 5 to learn differential effect-

stimuli

1. po will increase

under

Dins-

interpretation it will be

moor's

sumptions.
as-

as

forcement
rein-

to

will

this stimulus

introduced

already been

have

It is apparent
are

consistent

stimuli.

that these hypotheses


with

the

pothesis
general hy-

nation
the degree of discrimiwill tend to increase (or remain
since

In view

of these considerations

we

pothesis:
following general hyhigh)under differential reinforcement,
decrease
discriminative
while
it will tend
to
(or
to
Exposure
under
nondifferential
will
effect
stimuli
have a reinforcing
remain
low)
introduce

on

the

the

observing response

that

differentlyto

has
the

to

learned
two

to

the

tent
ex-

respond

discriminative

reinforcement.

but

magnitude

we

shall

refer

to

the

of the difference between

other

words,

will learn to respond differentlyto the


forcement,
differential reinstimuli under
two

stimuli.
Hereafter

In

in the

same

way

will learn to
to

them

differential reinforcement.

respond

under

non-

Additional

hypotheses of interest can be derived


from this general hypothesis.

529

JR.

WYCKOFF,

BENJAMIN

L.

Krechevsky

terval
that during the inresponding approximately

also noted

while

was

accordingto chance with respect to


a
stimuU, he showed
deis reversed
po will crease
These
findings
strong positionpreference.
to
temporarily and then return
in complete agreement with hypotheses
are

3. When

established

well

crimination
dis-

discriminative

the

high value.

this change in po
gree
because, following a reversal,the dewill decrease
of discrimination

originaldiscrimination

the

as

It

will then

discrimination

new

4.

If at

increase

the

is formed.

point in

some

ishes.
vanas

an

of discrimination

experiment

to

be retarded

interval,but finallyto

some

hypotheses presented so
in an
experiment by
were
is presented in
the writer (22) which
In this experiment
detail elsewhere.
four

The

far

tested

direct

the degree of discrimination is low and


time po is low (but greater
at the same
tion
than zero), we shall expect the formafor

and 4 in the present formulation.

shall expect

We

obtained

were

non-

differential reinforcement

ing
dur-

used

were

in

which

The

Skinner-box

the

and

reversal.

discrimination

striking

quite rapidly.

Ro

an

during differential reinforcement,

in

occur

of

measures

Pigeons
situation

effective response

was

single translucent

stimuli

discriminative

key.
ored
col-

were

lights(red and green) projected


hypothesis arises from the fact
the back of the key one
at a time.
on
crimination,
that increases in the degree of diswithheld
and
The
colored lightswere
and increases in po, are
the key was
lighted white until the
Early in
dependent upon each other.
of
The
occurred.
Ro consisted
This

Ro

will be

5
the process
discriminative

stimuli

exposed
only

to

the

small

stepping on

pedal on the floor of the

for using
The reasons
compartment.
proportion of the time and hence the
this response
observing response
as an
crease
incannot
degree of discrimination
discussed in detail elsewhere (22).
are
time
At
the
same
po
rapidly.
that this response
Here
it will suffice to
will not

because

increase

of the

say

low

Then, as
degree of discrimination.
becomes
the degree of discrimination
sufficientlygreat to bring about an
increase in po the entire learning process
will be accelerated.

fallswithin

observing

exposure
As
stimuli.

Krechevsky (11) presents data obtained


experimentsin a jumping
stand situation which
correspond in some
respects to the predictionsof the present
formulation.

Curves

for individual 5s

relatively
abrupt discrimination
In

general the

curves

also

mation.
for-

effect

for discrimination reversal

at

as

the
any

potheti
hy-

of the

discussed
had

response

no

ment
probabilityof reinforce-

given

moment.

above
hypotheses were
of this exresults
the
periment.
by
supported
first
three
the
Concerning
All

of the

hypotheses,po

was

ential
higherunder differthan

under
When

differential reinforcement.
were

shifted from

differential to

differential reinforcement

non-

5s
non-

marked

tion
degree of discriminaAll of these
decrease in po occurred.
terval
level,followed by an inat a 5 per
differences were
significant
duringwhich improvement was much
better.
confidence
of
or
cent level
rapid. Finallythe process accelerated
to

less

shows

on

reinforcement

slight improvement in discrimination


prior to the abrupt change. A curve
sented
pre-

discriminative
case

experiment

show

rapid decrease

the

above, the observing

in discrimination

show

the

to

in

an

in that it resulted

response

in

definition of

our

the

in the

chance

reversed

discrimination

formed.

The

fourth

hypothesisdoes

not

ap-

530

READINGS

ply unless

at

MATHEMATICAL

IN

point in the experiment

some

degree of discrimination
both low.
This condition

the

Po

are

was

satisfied consistently since the


(or base) level of the pedal
operant
turned

high for 5s.

out

this condition

when

learning

satisfied and

fected
af-

We

shall

now

derive

to supquantitative statements
plement
the
above
analysis. We

some

several

in

be

may

observingresponse

some

is requiredof S.

relatively shall attempt

be

to

However,

discrimination

and

not

response

PSYCHOLOGY

in such

that the

way

be

grated
readily intein these cases
the results conformed
into existingquantitative theories
of learning such as Hull's (9),
to the hypothesis.
We
illustrate some
Estes' (6) or
Bush
and
Mosteller's
can
now
ways
in which
this theory might be useful
potential applications of
(3). The
in interpretingbehavior in other exthis development could proceed along
periments.

cases

was

present theory

the relationships

down

to set

involved

two

1. If this

theory is applied

in which

discriminative
can

make

some

more

than

discriminations

pair of

one

different lines.

First,we could attempt to state the


relationships between
sponses
observing reand

stimuli is involved

of 5 to form

on

aspects of the
in such a way
that

effective responses
po could be estimated
where

based

measurable

we

predictionsregarding

in the readiness

changes

ations
situ-

to

can

some

direct

situations

in

of Ro

measurement

is

ticular
parnot

pair of stimuli.

feasible.

This

might be the

case,

ing
for example, if the Ro involved focus2. It has been demonstrated
that
If
of
the
the
we
apply
eye.
ent
preswhen
is reversed rea discrimination
peatedly
in this way,
development
po
5s tend to learn the reversed
would
become
able,
varian
intervening
discrimination
and

rapidly

more

more

which could be used to account


(15, 8). According to the present
for and predictbehavior in situations
theory, during discrimination reversal
the observing response
is partially where (1) the apparent generalization
between stimuli changes,or (2) where
reconditioned.
extinguished and

Thus, during repeated reversals,the


tently.
Ro is, in effect,reinforced intermit-

ease

of discrimination

of formation

changes

as

function

of

training.

Studies of intermittent reinforcementBerlyne (1) suggests that "attention"


be treated in a similar way.
have indicated that when a
could
criminati
predict disSecondly, we
is intermittently extinguished

response

and

the

reconditioned,

strength of the response


a

the

tends

to

tain
at-

relativelyconstant

high value
the first reversal po might

(18). On
drop to a low value, and recover
slowly, but with repeated reversals
would expect this drop to become
we
less prominent, and finally,
po would
remain
high throughout the reversal.
It is apparent

that

high, a

discrimination

reversed

be learned

more

if po remained
would

rapidly than

wise.
other-

by
learning functions
of
set
garding
assumptions readopting some
the component
learning processes
involved.
These
assumptions
could be adopted from some
existing
theory which treats the simpler processes
of

examined

precedingdiscussion
some

of the ways

we

main

the

moment

is the

absence

quantitativefunction

for

at

of any

predicting

shall
we
po. However,
down
the
be able to set
relationships

changes

in

in such
function

have

in which

extinction.

obstacle to this endeaver

The

involved
In the

conditioningand

inserted.

a
can

way

that any

ceptable
ac-

immediately be

BENJAMIN

L.

Quantitative

Analysis

of the Ro.

experiment discussed above.


it was
pointed out that we must

There

take into consideration

positiveor
To

three different

stimulus

populations which may effect


Si
5's*behavior.
We shall adopt the following
notation

stimuli.

the

Ro

S is exposed
card

on

on

52

Sz

when

(white) is

the stimulus
represent
the Ro octrials when
curs

S2

and

and

occurs

positive stimulus

present,

when

the negative stimulus


(black) is present, and 5$ the

stimulus

population

which

to

of the effectiveresponse
at any
given moment
during a trial.
This
variable can
be related to the
occurrence

variable of response
Estes (6)has shown
be

can

expected

probabilityat
trial,the

to

any

during

the present

case

probabilityof

we

must

us

response

Si, and
of

account

to

posed
ex-

the value of p when

5 is exposed

53

to

the net
which

of p for a trial on
positive stimulus is

value

the

present
the net value of p for a trial on
which
the negative stimulus is

the

present

In

probabilityof
given

Ro at any
trial.
a
We

the bols
symrepresent the
of the effective

5 is exposed to Si,

refer to the

occurrence

is

of p

52

to

consider the

Sz, respectively.

also wish

Si

to

of

occurrence

during

moment

net

We

shall

ity
probabil-

of the effective

sponse
re-

on
a
given trial,taking into
that S may
be exposed to

shall

tional
certain func-

express

now

these variables.

relationshipsamong

adopt

pi, pi, and pz to


probabilityof occurrence
when

when

value

for each of three stimulus

response

any

trial

posed
ex-

of p

of the effective

occurrence

populations. Let

at

is

the

ive
the effect-

occur

used.

will

when

p2

occur.

during

value

mean

units of measurement

ent.
pres-

given moment
(= k/L)
the

latency of the response


will be proportional to the reciprocal
of the probability; that is to say,
pL
k/p,where L is the mean
latency,
of
p the probability,and k a constant
will
which
depend on
proportionality
po
the

fails to

the probability that

p+

are

population of stimuli to
S is exposed if (1) the
positivestimulus is present and
(2) the Ro occurs.
the population of stimuli to
which
S is exposed if (1) the
is present
negative stimulus
and (2) the Ro occurs.
the population of stimuli to
which 5 is exposed if the observing

pi

occur

moment

stimuli

the

the

response

p3

represent
trials when

summarize:

latency as follows.
that if a response
with a given

the

to

negative

response

is

exposed when the Ro fails to occur.


In this analysis we
shall use
the
symbol p to represent the probability
of

the

use

which

to

population to which

population

these
represent
the
stimulus
Si represent

Let

trials when

shall

currenc

purposes

return

card

We

symbols /"+ and p^


of this analysis let us
net
probability on
of the hypoto consideration
thetical

For

the

531

JR.

WYCKOFF,

shall express
we
p+ and
functions of the variables

First

P-.

as

two

pi, p2, pz, and


which
variables

po.

and

p+

can

be

p- are
evaluated

such as
experimental measures,
out
latency of the effective response, withfrom

reference

Ro.

different stimuli during the trial depending


the
on
occurrence
or
non-oc-

They

to

direct

correspond

of

measures

to

the

ures
meas-

tained
of response
tendency usually obin discrimination
experiments.

532

PSYCHOLOGY

MATHEMATICAL

IN

READINGS

in

this assumption

However, in the present framework


to be the net
p+ and p- are assumed
result of the operation of the variables

we

pi, p2, pz, and po. Our task will be to


this dependence as a pair of

present framework, since, as we have


already pointed out, stimulus ization
general-

express

functional
done

relationships. This

making

forfeit the abilityto handle


the
generalizationwithin

do not

stimulus

be

can

follows.

as

However,

be accounted

can

such

postulatingany

a selected moment
during
5
positive trial. At this moment
will be exposed to either ^i, with a
probabilityof po, or to Sz, with a
posed
probabilityof (1
p^. If 5 is ex-

for without

direct

tion.
interac-

Consider

the

In

adopt a
conditioning and

Si he will make
with

response

the effective

proababilityof pi.

in such

way

of functions

effective response
and Ro are
independent of each other the probability

not

of functions

for

tempt
extinction, but atthe relationships

down

set

to

do

we

paper

particularset

"

to

present

that any acceptableset


serted.
be immediately incan

If the

that both
will

response

If S

pipo.

Ro and

will be the

occur

is

exposed

product

^3 he will

to

the

effective response
with a
probabilityof pz, and the probability

make

that both

(1

will

po)pz. The

"

that

the

of these

total probability

effective

this moment

at

occur

will be the product

occur

rect
assumption of "negligibledithat
implies
changes
of
the probabilityof occurrence

The

interaction"

the effective

will

response

will be the

in

with respect to
the effective response
stimulus
a
population 5,particular

only during
exposed to Si,
of change with
that the rate
and
will
time
depend on :
respect to
(i

1, 2,

P+

po) pz.

is reinforced.

2. The

(1)

If we

By exactly parallelreasoning with


respect
a

to

selected

moment

obtain

negative trial we

during

let r,- represent

of the time

Si, the

to

of pi at the time.

value

The

(1

P0P2 +

step will be to derive


for predictingthe values

next

pi, p2, and

The

(2)

Po) Pz.

follows

of

rate

of

dpi/dt

change of pi

and

It
readily ascertained.
to
possible
predict changes

if the

be

values of pi, p2, and


learning functions
if

we

assume

in

that

respect

to

each

respect

to

the others.

learningwith

implies
stimuli

will have

This

interaction
a

dPi/dt

(3)

fifoipi)

is reinforced,

(4)

TifeiPi)

effective response

tion
assump-

between

negligibleeffect.

functions

The

is not

forced.
rein-

/" represent

change
an

of probability of

rate

of

occurrence

tioning
effective response
during condiand extinction, respectively.

It will be

value

the

approximate

which
of

/" and

tions
acceptable set of analytic func-

any

tion
extinc-

of these stimuli, ceeds


proindependently of learningwith
that

as

the

pz on the basis of
for the simpler

of conditioning and

processes

be

will

be

can

functions

two

if the effective response

Pz.
contingenciesfor the effective response
in the presence
of ^i, 52, and Sz can

proportion

S is exposed

pressions
ex-

reinforcement

the

during which

approximated by
/)_

effective

the

not

or

response

(1

Popl +

occur

S is

sum

1. Whether

products. Thus
=

3) will

or

in which

the time

noted

of 1 to

that

we

for simple

if

we

will obtain

cases

assign

sions
expres-

of conditioning

L.

of

values

the

present model
be expressed as
each

are

to

of the time.
per cent
time
the subject will be

this

5i

to

Si with

or

Hence

give the
On

Ti

r3

as

of p+
follows.

ability
prob-

of S2

of the

Popi +

(1

p.

P0P2 +

(1

are

in the

of

presence

average

an

of

one-

non-reinforced

and

dpjdt

the

in

Using the above


appropriate functions

obtain

+
We

would

dp+/dt

.5poJc(pi)
.5{l-Po)mps)+fe(Pz)l

.5(1

function

.5poJe{p2)
.5{l-Po)'Uc{Pz)He{Ps)l

(7)

Po)Up3).

of time.

be derived

Such

Equations 1, 2, 10, and 11 represent


simultaneous
equations. By
can
combining these equations we
four

the

singleexpression:

empirically

fi{t). Thus we can obtain


pi
values of pi, p2, ps, and po for any
be
These
values can
point in time.
=

in

equations

1 and

ment
statethrough some
regarding the factors which
bring about changes in po. U po can
be expressed as a function of time we
rewrite equations 5, 6, and 7 to
can
obtain
expressions involving only
dt,
dpi,
pi and t. If these differential
will obtain
be solved we
equations can

substituted

of

pi, pz, and pz as functions


obtain
and
variables
other

express

dpo/dt

G(p+, P-,

theoretical

or

{pi-pz){dPo/dt), (10)

-\-{p2-pz){dpo/dt).(11)

(measurable aspects of effective


responses)if we can predict the values
could

respect to time

po)Mpz)

now

(2)

Substituting values for dpJdt, dpz/


dt and dpz/dtfrom equations 5, 6, and
obtain :
7 and rearranging terms
we

(6)

p-

of po as
function

po)pz.

(5)

outline the steps which


be necessary
to predict p+ and

can

(1)

{\-Po){dp^/dt)-Pi{dPo/dt). (9)

dpJdt

.5(1

po)p3,

po{dp2/dt)
+p2{dpo/dt)

dP'i/dt .Spofeipi)
dpz/dt

proceed

poidpi/dt)+pi{dpo/dt)

sponses
re-

.SpoUpi)

can

to

known

-f (1-Po){dp3/dt)-pz{dpo/dt), (8)

sponses
re-

forced,
rein-

time.
and

obtain

dpi/dt

wish

2 state:

Differentiatingwith

reinforced, and

not

are

for reinforced
we

/?_, we

p+

dp+/dt

that all effective

effective responses
^3 are reinforced

and

we

of po from

1 and

we

po).

of Si
in the presence
effective
responses

of

if

hand

Equations

.5po.

(1

also know

values

other

the values

values

"

half

the

estimate

probability

S will be exposed to S3 with


of (1
po)- Hence:

presence

and

Ti

We

prediction of p+

desired

P-

tive
posi-

50

present

During
exposed

The

follows.

as

negative stimuli

and

of po.

the

can

r^

of po

functions
be

In

extinction.

or

533

JR.

WYCKOFF,

BENJAMIN

to

dp+/dt,dpJdt, po), (12)


G will depend on
/e and/" adopted for the

the function

where

the functions

conditioning and extinction functions.


representing the
Now, if the curves
values

of /"+ and

p-

experimentally,

determined
express

these

analytic functions of time.

variables

as

We

also

can

are

can

we

obtain

dp+/dt and dp-/dtas


SulDstituting the

expressionsfor
functions of time.

functions

for

p+,

534

READINGS

dpj^/dtand dp-/dt in equation

p_,

obtain

we

3. When

12

(13)

G'{t,po).

obtain

be

can

temporarilyand then recover;


4. If the degree of discrimination
po are both low, the formation
will be retarded
discrimination

Po
This

=fo{t).

(14)

will

the

equation

give

us

quite rapidly.
in support of these specific
in an
obtained
hypotheses was

value

experiment in which
directly.

Summary
In

situations

This

learning

discrimination

many

such

response,

some

an

as

will be required of

orienting response,
S before
he is exposed
responses

and

some

"observing responses" {R^,

changes

as

their probabilityof

as

Increases

po.

result in increased

opportunity for 5

results
to

stimulus

learn

or

increases

generalizationbetween
stimuli.

The

of

training.

discriminations

more

sented
pre-

offers

mulation
for-

phenomenon.

opposite
alent
operationallyequiv-

are

function

the

of stimuli,

particularset
a

of

ease
on

relativelysimple and
of this
testable
interpretation
readily
This

effect.

decreases

discriminative

or

Decreased

the

will have

These

to

the

rapidly if reversals are


more
repeatedly. The present

creased
in-

hence

stimuli, and

po

in po will

and

where

and

5s learn reversed

criminative
to the dis-

exposure

discrimination.

currence
oc-

where

cases

discrimination

of

basis of

in

generalizationbetween

in

occur,

call these

indicate

manifest

stimuli

ured
meas-

was

be useful for

may

interpreting behavior

changes

Ro

an

formulation

formation
discriminative

the

to

We

stimuli.

for

will finally occur

but

interval

Evidence

of po for any
point in
during the experiment.

time

of

and

some

desired

ination
discrim-

is reversed, po will decrease

If this differential equation


we

well established

dpo/dt

solved

PSYCHOLOGY

MATHEMATICAL

IN

in

the

lends itself to precise

formulation

A quanquantitative statement.
titative
in two
be
used
could
analysis
quantitative pre(1) to make
dictions
ways:

of behavior

ing
followset

of theoretical

based

on

some

ing
regard-

statements

general hypothesis regarding


the component
learning processes,
changes in po can be derived from the
tions
and
to evaluate
(2)
po from observaof
reinforcement.
secondary
principle
of
effective
of
measurable
aspects
native
Hypothesis: Exposure to discrimifor
The
required
steps
stimuli will have a reinforcing responses.
such an analysis are outlined.
effect on the observing response
to the
that S has learned to respond
extent
REFERENCES
differentlyto the two discriminative
stimuli.

1.

general hypothesis we
derive the followingspecifichypotheses:
From

this

Berlyne,

58,
2.

under

2.

under

conditions

(or remain

Some
M. E., " CoATE, W. B.
crimination
of disexperiments on the nature

conditions

reinforcement

learning in
comp.

forcement
of differential rein-

of

nondifferential

3.

Bush,

R.

Psychol.,1950, 43,
R., " Hosteller,
model

4.

Rev., 1951,

Psychol.

BiTTERMAN,

high)

(discrimination training)
;
low)
po will decrease (or remain

Attention,perceptionand

theory.

137-146.

new

1. po will increase

D. E.

behavior

the

rat.

/.

198-210.
F.

for learning.

matical
mathe-

Psychol.

Rev., 1951, 58, 313-323.


son
DiNSMOOR, J. A. a quantitativecompariof the discriminative

and

reinforc-

Vous aimerez peut-être aussi