Journal of Philosophy, Inc.

Second Thoughts about Church's Thesis and Mathematical Proofs

Author(s): Elliott Mendelson
Source: The Journal of Philosophy, Vol. 87, No. 5 (May, 1990), pp. 225-233
Published by: Journal of Philosophy, Inc.
Stable URL: http://www.jstor.org/stable/2026831 .
Accessed: 14/10/2014 09:20

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Journal of Philosophy, Inc. is collaborating with JSTOR to digitize, preserve and extend access to The Journal
of Philosophy.

http://www.jstor.org

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
 THE JOURNAL OF PHILOSOPHY
VOLUME LXXXVII, NO. 5, MAY 1990

SECOND THOUGHTS ABOUT CHURCHi'S THESIS

AND MATHEMATICAL PROOFS
I intend to renounce the standardviewsconcerningthe natureof
Church's thesis. In order to put things in proper perspective, I
shall give a brief introduction to the relevant notions and a
resume of the history of the thesis.
The central notion is that of algorithm. An algorithm is an effec-
tive and completely specified procedure for solving a whole class of
problems. Examples are the well-known algorithms for adding and
multiplying any two natural numbers. In logic, the truth-table tech-
nique of testing a statement form to see whether it is a tautology is an
algorithm. An algorithm does not require ingenuity; its application is
prescribed in advance and does not depend upon any empirical or
random factors.
An algorithm that is applied to natural numbers and yields a natu-
ral number as a value is called an effectively computable function. In
general, the objects that an algorithm is applied to and the objects
that it yields can be represented by natural numbers. Therefore, for
the sake of simplicity, it is customary to confine attention to effec-
tively computable functions.
It is crucial to understand certain things about effectively comput-
able functions. First of all, we do not mean actual human computa-
bility or empirically feasible computability. A function can be com-
putable even though the instructions for it may require more sym-
bols than there are atoms in the universe. Likewise, a value of an
effectively computable function may require more time for its com-
putation than the entire past and future history of the human race.
When we talk about computability, we ignore any limitations of
space, time, or resources. Second, functions are treated extension-
ally. A function is determined by the ordered pairs of its arguments

0022-362X/90/8705/225-233 ?) 1990 The Journal of Philosophy, Inc.

225

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
226 THE JOURNAL OF PHILOSOPHY

and values, not by a particular way in which it is defined. For exam-

ple, consider the following function:
f 1 if Fermat's Last Theorem is true
{O if Fermat'sLastTheoremis false
This function is an effectively computable function. It is either the
constant function 1 (which is effectively computable) or it is the
constant function 0 (which is also effectively computable). At the
present time, we happen not to know how to compute any values of
f (x). Thus, we are adopting a completely classical, nonintuitionistic
stance. '
Many concrete algorithms have long been known in the history of
mathematics and logic. In all such cases, the fact that the alleged
algorithm actually is an algorithm is an intuitively obvious fact. (For
example, it was so easy to see that our method of adding two natural
numbers is an algorithm that we never bothered to prove it. There-
fore, it never entered our mind that the notion of algorithm required
any further elucidation.) This was not felt to be a difficulty until
mathematicians and logicians began to encounter important classes
of problems for which there seemed to be no algorithmic solutions.
(Examples: (a) Is a given first-order formula logically valid? (b) Is a
given arithmetic statement derivable from the axioms of Peano arith-
metic? (c) Is a given mathematical statement true? (d) Is a given
equation true in a given finitely presented group?) To show the
nonexistence of algorithms for solving these problems seemed to call
for a mathematically precise equivalent of the notions of algorithm
or effectively computable function. The "pressure" of these prob-
lems finally became so great that in the middle of the 1930s several
quite different explications were independently proposed for the
notion of effectively computable function. These proposals were all
published in the same year, 1936, by Alonzo Church, Emil Post, and
Alan Turing.2 It turned out that they (and many other later ones) are
mutually equivalent. Let us look at Turing's approach, which is the
most convincing.

1
It was from this classical viewpoint that Church's thesis was originally proposed
and discussed, and it is still commonly used with such an interpretation. By keeping
to this classical setting, I am not denying the tenability of a constructivist philosophy
of mathematics.
2 Church, "An Unsolvable Problem of Elementary Number Theory," Amer. J. of
Mathematics, LVIII (1936): 345-363; Post, "Finite Combinatory Processes. For-
mulation I," Journal of Symbolic Logic, I (1936): 103-5; Turing, "On Computable
Numbers with Applications to the Entscheidungsproblem," Proceedings of the
London Mathematical Society, XLII (1936): 230-265.

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
 CHURCH'S THESIS AND MATHEMATICAL PROOFS 227

Turing wanted to capture the essence of computability by setting

up mathematical "models" that would generate all functions, and
only such functions, that are computable. These models are what we
now call Turing machines. When calculating we require a field of
computation, such as a piece of paper or a blackboard. A Turing
machine uses for this purpose a tape, divided into squares. At any
one moment, this tape consists of finitely many squares, but, when
necessary, additional squares can be added to the right and left. At
any given moment, the squares of the tape may be occupied by
certain symbols, at most one symbol to a square. These symbols are
chosen from a finite set of symbols, the alphabet of the machine. The
Turing machine has a reading device or scanner which, at any given
moment, is observing the content of one of the squares of the tape.
At any one moment, the scanner can be in any one of a fixed finite
collection of internal states. (These states correspond to the presum-
ably finite number of states that characterize the structure of a per-
son or computing machine at any given time.) The Turing machine
also has a program, a finite list of instructions that tell it what to do
under all circumstances. If the scanner is in a certain state and is
observing a certain symbol, these instructions tell it to do one of four
things: (1) replace the observed symbol by another (possibly the
same) specific symbol; (2) move one square to the right; (3) move one
square to the left; (4) stop. After the appropriate action is per-
formed, the new internal state of the scanner is determined by the
program. A Turing machine computes a functionf (xl, . . . , x,n)of
n arguments in the following manner. At the beginning of the com-
putation, the n arguments are written on the tape, separated by
blank squares. (Each argument, a natural number m, is represented
by a sequence of m + 1 strokes, I II . . 1, written in consecutive
squares.) The machine starts off in a special state, called the initial
state, and proceeds to operate according to its instructions. If the
machine stops and the only thing left on the tape is the representa-
tion of a certain number k, then k is taken to be the valuef (xI,. . . .
xn). A function computable by some Turing machine is called a Tur-
ing-computable function. Such a function can be partial, that is, it
need not be defined for all n-tuples.
Notice that Turing made several normalizing assumptions. Al-
though we usually use at least a two-dimensional field of computa-
tion, he restricts the computation to a one-dimensional tape. This
really is not an essential limitation, since it is well-known how to
encode higher dimensional arrays of symbols in terms of a string of
symbols. Second, although we usually move freely around our field
of computation, Turing restricts his machine to movements to an

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
228 THE JOURNAL OF PHILOSOPHY

adjacent square. Here, one has to verify that this is not a serious
limitation; the effect of a "jump" from a square to a nonadjacent
square can be accomplished by a suitable sequence of movements to
adjacent squares. Third, the fact that the alphabet is finite is not a
problem. Actual computing systems use finite alphabets; moreover, a
denumerable alphabet a,, a2,. . . can be replaced by a two-symbol
alphabet {a, '} by letting a' stand for a,, a' for a2, and so on. Fourth,
the restriction to a finite set of internal states seems to be acceptable.
Turing justifies this as follows: "If we admitted an infinity of states of
mind, some of them will be 'arbitrarily close' and will be confused."
(If there is any weak link in Turing's presentation, this is probably it.)
Church's thesis can now be formulated.
Church'sthesis(CT):A functionis effectivelycomputableif and only if it
is Turing-computable.
The Turing-computable functions have been shown to be identical
with the so-called partial recursive functions (defined by S. C.
Kleene3). Church's thesis is often stated in terms of partial recursive
functions instead of Turing-computable functions, because the
former are easier to work with from a technical standpoint.
Here are some of the arguments in favor of CT.
(a) The host of notions proposed as equivalentsof effectivelycomput-
able function have all turned out to be equivalent.This seems to
show that the underlyingintuitivetarget notion has been hit.4
(b) It is easyto see that one half of CT is true:everyTuring-computable
function is effectivelycomputable.
(c) The other half of CT, that every effectivelycomputablefunction is
Turing-computable,has been confirmedfor the very large number
of cases in which it has been tested.
(d) The argumentsgiven by Turing show that the special methods of
computationused by Turing machinesinvolve no essential restric-
tion on the generalmethodsof computationemployedby (idealized)
human beings or computers.
CT is now almost universally accepted by mathematicians and lo-
gicians. In the theory of recursive functions, it is constantly resorted
to without explicit mention.5 Moreover, all the limitative theorems of
modern logic, such as Godel's incompleteness theorem, depend for
their significance on the validity of CT. For example, Church's
theorem states that there is no recursive decision procedure for

3 Introduction to Metamathematics (Princeton: Van Nostrand, 1952).

4 Of course, this is not conclusive. It is conceivable that all the equivalent notions
define a concept that is related to, but not identical with, effective computability.
5 See the standard reference work by Hartley Rogers, Jr., Theory of Recursive

Functions and Effective Computability (New York: McGraw-Hill, 1967).

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
 CHURCH'S THESIS AND MATHEMATICAL PROOFS 229

first-order logical validity.6 It is CT that then allows us to draw the

important conclusion that there is no effective procedure for decid-
ing whether a first-order formula is logically valid.
CT is generally regarded as an assertion of the extensional equiva-
lence of the notions of effectively computable function and partial
recursive function (or Turing-computable function). Because the
former notion is intuitive and the latter is alleged to be a precise
mathematical concept, it is further asserted that it is impossible to
prove CT; hence, CT is just a conjecture or a thesis, and is not
susceptible of a rigorous proof. This is what almost everybody be-
lieves about CT; for example, Church and Kleene:
This definition is thought to be justified by the considerations which
follow, so far as positive justification can ever be obtained for the selec-
tion of a formal definition to correspond to an intuitive notion (Church,
op. cit., ?7).
Since our original notion of effective calculability of a function is a
somewhat vague intuitive one, [CT] cannot be proved . . . While we
cannot prove [CT], since its role is to delimit precisely a hitherto vaguely
conceived totality, we require evidence that it cannot conflict with the
intuitive notion which it is supposed to complete; i.e. we require evi-
dence that every particular function which our intuitive notion would
authenticate as effectively calculable is [recursive]. The thesis may be
considered a hypothesis about the intuitive notion of effective calcula-
bility, or a mathematical definition of effective calculability; in the latter
case, the evidence is required to give the theory based on the definition
the intended significance (Kleene, op. cit., pp. 317-9).
These standard views about CT hold it to be a rational recon-
struction, in the sense of Rudolf Carnap and Carl Hempel. A ratio-
nal reconstruction is a precise, scientific concept that is offered as an
equivalent of a prescientific, intuitive, imprecise notion. In all stan-
dard cases in which the original intuitive notion is definitely known
to apply or not to apply, the rational reconstruction should yield the
same outcome. The rational reconstruction may go beyond the origi-
nal notion, however, in cases where the latter notion is not deter-
minate. Moreover, the rational reconstruction need not, and usually
does not, have the same psychological effect as the original notion;
expert users of the language generally do not think of the rational
reconstruction when they use the original notion. Confirmation of
the correctness of a rational reconstruction apparently must involve,
at least in part, an empirical investigation. The correctness of the
reconstruction cannot be proved.
6
More precisely, iff is the function such thatf (n) = 1 if n is the Godel number of
a logically valid first-order formula and f(n) = 0 otherwise, thenf is not recursive.

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
230 THE JOURNAL OF PHILOSOPHY

I would like to challenge the standard conception of CT as an

unprovable thesis. My viewpoint can be brought out clearly by argu-
ing that CT is another in a long line of well-accepted mathematical
and logical "theses," and that CT may be just as deserving of accep-
tance as those other theses. Of course, these theses are not ordinarily
called "theses," and that is just my point.7
The following are a few obvious cases of "theses" that have never
been called "theses," but probably merit this term as much or as little
as CT.
I. Functions. In contemporary logic and mathematics, a functionf
is defined to be a set of ordered pairs satisfying the following condi-
tion: (*) If (x, y) ef and (x, z) ef, theny = z. In classical mathematics,
a function was tied to a rule for calculating it, generally by means of a
formula of a standard type. In the course of the nineteenth century,
mathematicians reluctantly came to the conclusion that this concep-
tion of function was too narrow; now, with the exception of con-
structivists, everyone accepts the definition given above. The earliest
source that I can find that contains this definition is: G. Peano,
"Sulla definizione di funzione," Atti della Reale Accademia dei
Lincei, Classe di scienze fisiche, matematiche e naturali, XX (1911):
3-5. Hence, we might call the identification of this definition with
the intuitive notion of function "Peano's thesis." It is clear that an
intuitively understood function determines a unique set of ordered
pairs satisfying condition (*). Conversely, such a setf can be thought
of as determining a rule for calculating a function: to each x assign
that object y such that (x, y) ef. (Since this "rule" is not necessarily
constructively definable, constructivists would reject this part of
Peano's thesis.) It appears that, in contradistinction to CT, Peano's
thesis has no important mathematical consequences and is useful
only for foundational purposes: (a) in an axiomatic presentation of
mathematics, we do not need to assume the concept of function as
primitive; (b) together with other axioms of ZF (Zermelo-Fraenkel
set theory), this definition allows us to prove that, for any sets A and
B, the set AB of all functions from B into A exists.
II. Tarski's definition of truth. Consider a first-order language L
7 A search of the literature has turned up only one paper with explicit references
to 'theses' in the history of logic and mathematics: Stewart Shapiro, "Understanding
Church's Thesis," Journal of Philosophical Logic, x (1981): 353-365. Shapiro
seems to accept the standard view of the inherent unprovability of Church's thesis,
however. While preparing the final form of the present paper, another such refer-
ence came to my attention: R. L. Epstein & W. A. Carnielli, Computability: Com-
putable Functions, Logic, and the Foundations of Mathematics (Pacific Grove,
CA: Wadsworth, 1989), ch. 25. They consider theses concerning the notions of
"circle" and "continuous function." But they also accept the orthodox view of CT:
"(CT) is a nonmathematical thesis: it equates an intuitive notion . . . with a precise,
formal one" (p. 86).

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
 CHURCH'S THESIS AND MATHEMATICAL PROOFS 231

and a structure A based on this language. For any sentence 13 of L,

we understand what it means for 1 to be true in A. In the 1930s,
Tarski gave a set-theoretic definition of '13 is true in A' that is now
universally acknowledged to be equivalent to the corresponding in-
tuitive notion. By mathematical induction, using some basic proper-
ties of the intuitive notion and of Tarski's notion, we can convince
ourselves of the validity of "Tarski's thesis." Like the notion of re-
cursiveness, Tarski's definition has found important applications,
particularly in connection with proofs of consistency and inde-
pendence.
III. Logical validity. The currently accepted semantical or
model-theoretic definition of logical validity is that a first-order sen-
tence is logically valid if it is true in all structures. (Notice that this
definition attains greater precision if we understand 'true' in the
Tarskian sense.) This definition probably goes back to the 1920s,
although I cannot identify its exact origin. In any case, Gottlob
Frege, Bertrand Russell, A. N. Whitehead, and other earlier logicians
certainly understood the notion of logical validity, although it is
highly unlikely that they had thought of, or would have assented to,
the model-theoretic definition. (Frege thought of his individual vari-
ables as ranging over arbitrary objects in an unrestricted universe,
and predicate letters as ranging over arbitrary properties and rela-
tions on the universe. Hence, his conception of logical validity, if he
had ever formulated it, probably would have differed from the
model-theoretic definition.) Thus, we are dealing here again with a
thesis, one that asserts the equivalence of the model-theoretic defini-
tion with a previous intuitive notion. There are plausible arguments
in favor of the thesis. A theorem of model theory tells us that a
sentence that is true for all structures of a given cardinality a is also
true for all structures of cardinality less than a. Hence, if we think (at
our peril) of a as the cardinality of the universe, then Frege's con-
ception of logical validity would imply the model-theoretic version.
Conversely, if a sentence fB is true for all structures, then G6del's
"completeness theorem" implies that 1 is provable in the first-order
predicate calculus. But all the axioms of that calculus would be uni-
versally accepted as logically valid and its rules of inference would be
universally acknowledged to lead from logically valid formulas to
logically valid formulas. Hence, 13 should be accepted as logically
valid in the Fregean (or any other) sense.
IV. Limits. The intuitive notion of limit was widely used in mathe-
matical analysis in the eighteenth century, and, in the first half of the
nineteenth century, A. L. Cauchy showed how limits could be used to
give a firm foundation to the basic ideas of calculus, such as continu-
ity and differentiation. It seems to have been only in the 1860s,

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
232 THE JOURNAL OF PHILOSOPHY

however, that Karl Weierstrass was the first one to give the well-
known e-6 type of definition of a limit of a function (and a corre-
sponding definition for a limit of a sequence). The Weierstrassian
definitions can be seen to capture exactly the meaning of the preced-
ing intuitive notions; the "Weierstrass thesis" ranks among the most
firmly established mathematical theses. When we ordinarily work
with limits and when we introduce the idea to students, we still rely
on the original intuitive picture, but we resort to the more precise
definition in difficult cases.
These four examples form just a small sample of a large body of
logical and mathematical theses. A few other examples are the notion
of measure as an explication of area and volume, the definition of
dimension in topology, the definition of velocity as a derivative, the
definition of logical implication and logical equivalence in first-order
logic, and the definitions of circle, triangle, interior of an angle, and
many other geometric concepts.
Now that we have other theses with which to compare CT, here is
the main conclusion I wish to draw: it is completely unwarranted to
say that CT is unprovable just because it states an equivalence be-
tween a vague, imprecise notion (effectively computable function)
and a precise mathematical notion (partial-recursive function). My
argument is based on three points.
(1) The concepts and assumptions that support the notion of par-
tial-recursive function are, in an essential way, no less vague and
imprecise than the notion of effectively computable function; the
former are just more familiar and are part of a respectable theory
with connections to other parts of logic and mathematics. (The no-
tion of effectively computable function could have been incorpo-
rated into an axiomatic presentation of classical mathematics, but the
acceptance of CT made this unnecessary.) The same point applies to
theses 1-111. Functions are defined in terms of sets, but the concept
of set is no clearer than that of function and a foundation of mathe-
matics can be based on a theory using function as primitive notion
instead of set. Tarski's definition of truth is formulated in set-theo-
retic terms, but the notion of set is no clearer than that of truth. The
model-theoretic definition of logical validity is based ultimately on
set theory, the foundations of which are no clearer than our intuitive
understanding of logical validity.
(2) The assumption that a proof connecting intuitive and precise
mathematical notions is impossible is patently false. In fact, half of
CT (the "easier" half), the assertion that all partial-recursive func-
tions are effectively computable, is acknowledged to be obvious in all
textbooks in recursion theory. A straightforward argument can be

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions
 CHURCH'S THESIS AND MATHEMATICAL PROOFS 233

given for it. (The so-called initial functions are clearly effectively
computable; we can describe simple procedures to compute them.
Moreover, the operations of substitution and recursion and the
least-number operator lead from effectively computable functions to
effectively computable functions. In each case, we can describe pro-
cedures that will compute the new functions.) This simple argument
is as clear a proof as I have seen in mathematics, and it is a proof in
spite of the fact that it involves the intuitive notion of effective
computability. The fact that it is not a proof in ZF or some other
axiomatic system is no drawback; it just shows that there is more to
mathematics than appears in ZF.
(3) Another difficulty with the usual viewpoint concerning CT is
that it assumes that the only way to ascertain the truth of the equiva-
lence asserted in CT is to prove it. In mathematics and logic, proof is
not the only way in which a statement comes to be accepted as true.
Of course, this is a consequence of the truism that not all truths can
be proved; proofs must assume certain axioms and rules of infer-
ence. But, over and above this, we have observed in our four exam-
ples above that equivalences between intuitive notions and appar-
ently more precise mathematical notions often are simply "seen" to
be true without proof, or are based on arguments that are a mixture
of such intuitive perceptions and standard logical and mathematical
reasoning. (Notice also that these perceptions and arguments seem
to be of a nonempirical nature.)
That CT is true follows, I believe, from Turing's analysis of the
essential elements involved in computation. (Further, even more
sophisticated, Turing-like analyses have been given.8) But this is not
what I have tried to establish. The point I have attempted to make is
that equivalences between intuitive notions and "precise" notions
need not always be considered unprovable theses. Such equivalences
sometimes can be directly perceived to be valid, sometimes they can
be justified by a combination of directly perceived truths and logical
argument, and, in other cases, it is possible to discover entirely con-
vincing proofs.'
ELLIOTT MENDELSON
City University of New York/Queens College
8 A.N. Kolmogorov and V.A. Uspenski, "On the definition of an algorithm,
Uspekhi Mat. Nauk, VIII (1953): 125-176 (Amer. Math. Soc. Translations, xxix
(1963): 217-245). R. Gandy, "Church's Thesis and Principles for Mechanisms,"
The Kleene Symposium (Amsterdam: North-Holland, 1980), pp. 123-148.
9 After completion of this paper, it came to my attention that G. Kreisel had made
a somewhat similar point in his paper, "Informal Rigour and Completeness
Proofs," Problems in the Philosophy of Mathematics, I. Lakatos, ed. (Amsterdam:
North-Holland, 1967), pp. 138-186, esp. p. 176.

This content downloaded from 121.215.28.51 on Tue, 14 Oct 2014 09:20:21 AM

All use subject to JSTOR Terms and Conditions