16January1995

PHYSICS LETTERS A

ELSEVIER Physics Letters A 197 (1995) 83-87

Optimal distinction between two non-orthogonal quantum states

Gregg Jaeger ", Abner Shimony b
a Department of Physics and Collegeof GeneralStudies, Boston University, Boston, MA 02215, USA
b Departments of Physics andPhilosophy, Boston University, Boston, MA 02215, USA
Received 14 September 1994; revised manuscript received 9 November 1994; accepted for publication 9 November 1994
Communicated by P.R. Holland

Abstract

Two procedures are developed for classifying an individual system as IP) or Iq), non-orthogonal, given an ensemble with
respective proportions r and 1 - r. One (generalizing Ivanovic, Dieks, and Peres) infallibly classifies some systems, leaving others
unclassified. The second is statistically optimum, allowing individual errors.

Ivanovic [ 1 ], D i e k s [ 2 ], and Peres [ 3 ] consider 1 - P = probability of no classification

the problem of determining what quantum state an
individual system was prepared in, when it is given
that it was prepared either in IP) or I q ) , in general
nonorthogonal. The problem, thus stated, is vague
and can be clarified in (at least) two distinct ways:
(1) What procedure yields on the average a maxi-
m u m number of correct classifications, in an ensem-
ble of such cases, assuming that for each member of
the ensemble a definite classification is made? (2)
What procedure enables one in a m a x i m u m number
of cases to infer with certainty whether the system was
prepared in IP) or I q), leaving a m i n i m u m number
o f cases undecided?
The three authors mentioned are primarily inter-
ested in the problem (2), though Dieks briefly con-
siders problem ( 1 ). Under the assumption that half
o f the ensemble is prepared in IP) and half in I q),
all three obtain the same evaluation of the m a x i m u m
probability o f correct classification and m i n i m u m
probability of no decisions: namely,
P = probability of correct classification
=l-I(plq) l, (la)
= I(Plq) I • (lb)
The procedures suggested by Dieks and Peres are es-
sentially the same: prepare an auxiliary system in a
state Iso) and choose a unitary evolution (equiva-
lent to a choice of interaction) which yields
IP So ) -~a [P~ s~ ) +fliP2 sa ) , (2a)
[qso)~lq, sl ) +61q2 s2 ) , (2b)
where Isl), Is2), [p,), IP2), Lql), Iqz) are allnor-
malized, and
(Pl ]ql ) = 0 , (2c)
(st Is2 ) = 0 , (2d)
and I q2) is identical with IP2) except for a phase fac-
tor. With this unitary evolution there is a measure-
ment on the auxiliary system that decides unequivo-
cally between Is1 ) and Is2); if the former occurs, then
a measurement between IPl) and Iq~) decides be-
tween IP) and I q ) , and if the latter occurs the state
of the primary system is inevitably IP2) (or equiva-
lently I q2 ) ), leaving the choice between IP ) and I q)

0375-9601/95/$09.50 © 1995 Elsevier Science B.V. All rights reserved

SSDI 0375-9601 ( 94 )00919-8
84 G. Jaeger, A. Shimony /Physics Letters A 197 (1995) 83-87

undecided. The procedure of Ivanovic is more cum-

bersome, requiring possibly an infinite number of P=r(1-[(Plq) [2) . (9b)
steps. Note that if r = s = ½, then Eq. (7a) holds and hence
The present note extends the work of Ivanovic, (9a), which in this case agrees with Eq. ( l a ) .
Dicks, and Peres in several ways. The first is to allow A second extension of the results of Ivanovic, Dieks,
a proportion r of systems of the ensemble to be in ]p } and Peres is to assume that the dimensionality of the
and s= 1 - r to be in [q}, instead of requiring r = s = ½. Hilbert space of the system of interest is greater than
With no loss of generality, two. It is then possible to parallel the strategy of Dicks
r>s. (3) and Peres without introducing an auxiliary system.
We show that because of the dimensionality of the
The procedure for solving problem ( 1 ) is the same as Hilbert space, we can express [p) and Iq) in the form
that of Dieks and Peres, except for the evaluation of
the coefficients on the r.h.s, of (2a) and (2b). From IP} = a IP~ } +fliP2 ) , (lOa)
Eqs. (2a) and (2b), and the assumed proportions r
Iq} =YlqL } -t-61P2 } , (10b)
and s, we have
P=r]ol]iq-s]y]2= 1-- rl/?12-sI6l 2 . (4) where Ip~), Iqt) and IP2) are orthonormal and/?
and 6 satisfy Eqs. (5b) and (6a), and oe and 7 are
By unitarity and (2c), (2d) real. To achieve this expression we first write
I ( p l q ) l=r/?l 161 I ( p e l q 2 ) l , (5a) Iq} = el°Nip} + ( 1 - N 2) 1/21m } , ( 1 la)
and hence
N=l(plq)[, (lib)
I/?1 161 ~>l(Plq) l • (5b)
where Im) is normalized and orthogonal to IP).
The maximum of P in Eq. (4), subject to the con- Then Eq. (10b) is equivalent to
straint of Eq. (5b) and to 0 ~ I/?12~<1, 0 ~< 1612~<1, is
ei°N[p)+(1-N 2)t/2 I m ) = y l q l ) + 6 l p 2 ) . (10e)
achieved when
There exists in the Hilbert space a normalized vector
I / ? [ 2 = m a x { l ( P l q } l ( s / r ) '/2, I(P[q)[2} (6a)
I/} orthogonal to ]p) and I m ) , and we shall express
and [Pl ), Iq,) and [P2) explicitly in terms of IP), I m )
and I/). At this juncture we must treat the cases of
IP2 ) = Iq2 ) • (6b)
(7a) and ( 7b ) separately.
(Note: I/?12is not permitted to be less than I (Pl q) 12, If (7a) holds, make the following identifications of
for if it were, then 161 would exceed unity. ) If /?, & o~, y, IPl), Iql), and IP2},
1/712= I ( p l q ) l(s/r)l/Z>_-l(plq)12 , (7a) /?=N W2(s/r)'/4, (12a)
then 6= eiON l/2(r/s ) i/4, (12b)
1612= I ( p l q } l (r/s) 1/2 (8a) oe = [ 1 - (s/r) '/2N] 1/2 (12c)
and 7= [ 1 - (r/s)l/2N] l/i, (12d)
P= 1-2(rs)t/21 ( P l q ) I , (%) IPl ) = [ 1 - (s/r) 1/2N] 1/2 IP}
but if e -iO
( 1 - - N 2) 1/2 N[ 1 - (s/r)I/2N] l/2lm }
1/712= I (P[q)12> l ( P l q ) l (s/r) 1/2 , (7b)
then 1
( 1 - N 2) l/5 N l / 2 ( s / r ) , / 4 [ I -- ( r / s ) I / 2 N ] 1/21l),
1612=1 (8b) (12e)
and
 G. Jaeger, A. Shimony / Physics Letters A 197 (1995) 83-87
1
Iql ) - - (l__N2)I/2 [ 1 - ( r / s ) l / Z N ] l / 2 l m )
e i0
+ (1 - - N 2) ' / s N ' / Z ( r / s ) l / 4 [ 1 -- ( s / r ) l / 2 N ] ~/Zll)
12f)
11)2 ) = N ' / Z ( s / r ) 1 / 4 IP)
e -io
+ ( 1 - N 2) l / : i N l / 2 ( r / s ) l / 4 [ 1 -- ( s / r ) l / 2 N ] m)
1 if 2, then there is no decision. The probability of a
( 1 - N 2) ,/2 [ 1 - ( r / s ) l / 2 N ] ,/2
X [ 1 - ( s / r ) ' / 2 N l '/21l ) . 12g)
Then eqs. (10a), (10b) are satisfied with the requi-
site orthonormality of IP~ ), I q~ ), and IP2), while fl
and ~ satisfy Eqs. ( 5b ) and (6a) and a and ~ are real.
Hence we can optimally decide with certainty (but
sometimes abstaining from a decision) whether the
system was initially in IP) or in Iq) by a procedure
analogous to that given in Eqs. ( 4 ) - ( 9 a ) . Specifi-
cally, measure
A=lpl)(pll+2lql)(qll+3lp2)(P2l . (13)
If the measured value of A is 1, we know with cer-
tainty that the system was prepared in IP) ; if the value
is 2, we know with certainty that it was prepared in
Iq ) ; and if the value is 3, we do not know and abstain
from deciding. The probability of a correct classifi-
cation is given by Eq. (9a). This agreement shows
that the essence o f the procedure is to increase the
dimensionality o f the relevant Hilbert space beyond
the two dimensions determines by IP) and I q)- If
the Hilbert space o f the system o f interest is intrinsi-
cally greater than two, then one need not introduce
an auxiliary system in the manner of Dieks and Peres.
In the case of Eq. (7b), I~1 is unity and therefore
Eqs. (10a), ( 10b ) reduce to
IP) -- c~Ip, ) +fliP2 ) , (14a)
Iq) = ~ l p 2 ) , (14b)
where
IPl = N ~ I < P l q > l , (15a)
l a l = ( 1 - N 2 ) 1/2 (15b)
85
Now IP ) can be rewritten
[p) = e~°NI q) + ( 1 - N 2) 1/21/7 ) , ( 16 )
where I n ) is normalized and orthogonal to Iq)- An
o p t i m u m procedure for classifying a system as being
in IP) or Iq) with certainty is to measure the
observable
B = I n ) ( n l + 2 1 q ) (ql • (17)
If the measured value is 1, then the system can be
inferred with certainty to have been prepared in IP);
correct classification is
P=rlal2=r( 1- I(plq)12 ) , (18)
in agreement with Eq. (9b). Note that the procedure
can be carried out in a Hilbert space o f dimension
two and hence is more economical than the proce-
dure in case (7a).
A third extension is previous work is to solve prob-
lem ( 1 ), by finding a procedure for maximizing on
the average a correct classification. Again, we con-
sider the general preparation in which a proportion r
is prepared in IP) and s - 1 - r in I q). We seek a bi-
valent procedure, which prescribes choosing IP)
upon one outcome and Iq) upon the other. There is
no loss o f generality in restricting the bivalent proce-
dure to measuring a projection operator E on the Hil-
bert space of the system, with eigenvalues ' 1 and 0.
If the measured value of E is 1 choose IP), if0 choose
I q). The probability o f a correct choice is
P=r(plEIp) + s ( 1 - ( q l E I q) ) • (19)
Decompose I q) into a superposition o f o r t h o n o r m a l
IP) and I m ) , as in Eq. ( 1 la), and consider the ac-
tion of the projection operator E on these two vectors,
E I p ) = c l p ) +c' l m ) +c" ll) , (20a)
El m ) = d i p ) +d' I m ) + d " [ l ' ) , (20b)
where II) and I I ' ) are o r t h o n o r m a l t o both IP) and
[ m ) . ( O f course, if the Hilbert space is two-dimen-
sional, there are no terms in I l) and I I' ). ) Then
(pIEIp) =c (21a)
The procedure used here was presented in Section 2 ofRef. [ 4 ]
by Jaeger et al. In that paper the procedure was used for predic-
tion, whereas here it is used for retrodiction.
86 G. Jaeger, A. Shimony / Physics Letters A 197 (1995) 83-8 7

and I _sN 2
( qlEIq) = N 2 c + 2 cos(0+ ¢~)N(1-N 2) t/21c' I x= [ ( ~ _ s N Z ) 2 + s Z N Z ( l _ N 2 ) ] I / 2 (28)

+(1-NZ)d ' , (21b) and

where P = ½ + ½ ( 1 - 4 r s l (p[q)12 ) 1/2 (29)

d*=c' = [c' Iei° . (21c) The procedure for achieving this optimum probabil-
ity of a correct classification is to measure the projec-
Since E2=E, ( x [ E [ x ) = ( E x [ E x ) for any Ix), and tion operator E determined by the x of Eq. (28), and
therefore c and d" are real and choose IP) if the outcome is 1, Iq) if the outcome is
c=c2+ [c' [2+ [c" ]2, (22a) 0. We emphasize that this procedure works whether
the dimension of the Hilbert space of the system of
d ' = [d12+d'2+ [d" [2. (22b) interest is two or greater than two.
For the case of r = s = ½, Eq. (29) implies
Hence
P = ½ + ½ ( 1 - I(plq)[2)1/2 (30)
P=rc+s[1-Nac - (1-N2)d '
For this case, Dicks considers (without claiming op-
- 2 cos(0+ (~)N( 1 - N 2) 1/21c' [ l • (23) timality) a procedure which yields a probability PD
When c, d', [c'] are fixed, P is maximized by of correct classification,
cos(0+ ~) = - 1, P D = I - - ½ 1 ( p I q ) I2. (31)
P= ( r - s N 2 ) c + s - s ( 1 - N Z ) d ' But
+2sN(1 - N 2) 1/21c' I . (24) P - P D = ½[(1 - I ( P l q ) [ 2) 1/2_ (1 - [ ( P l q ) 12)]
By (22a), (22b) >/0, (32)

c=½ +_½[ 1 - 4 ( I c ' 12+ Ic" [2)]1/2,
d ' = ½ + ½ [ 1 - 4 ( l c ' 1 2 + l d " l 2)]1/2
and since r>~s we shall take the + for c, - for d'.
Moreover, when Ic' I is fixed, c is maximized and d'
is minimized when c" = d" = 0, in other words, when
E is a projection operator on the two-dimensional
Hilbert space spanned by IP) and Iq). Then
c=½(l+x)
d'=~(l-x),
where
x = ( 1 - 4 1 c ' 12) 1/2
Ic'l = ~ ( 1 - x 2 ) ~ / 2 .
Hence
P= ½+x(½ - s N 2) +sN( 1 - N 2) 1/2( 1 -to 2) 1/2.
Maximum P yields
(25a) and therefore our procedure is preferable to the one
considered by Dicks.
(25b) As a final remark, we consider a class of practical
situations in some of which the procedure of problem
(1) is appropriate, and in others the procedure of
problem (2). Suppose that a wager can be made about
each system of the ensemble, with a gain g> 0 if the
classification is correct, a loss l> 0 if the classifica-
tion is incorrect, and neither gain nor loss if the sub-
(26a) ject refrains from betting. The average gain using the
optimum procedure of problem (2) is
(26b)
G2 = g [ 1 - 2 ( r s ) l / 2 1 (p[ q) [] (33a)
ifEq. (7a) holds, and
(26c)
G2=gr(l-I(Plq) [2) (33b)
(26d)
if Eq. (7b) holds. The average gain using the opti-
mum procedure of problem ( 1 ) is
GI =g[½ + ½( 1 - 4 r s [ (Plq)12) 1/2]
(27) - l [ ½ - ½ ( 1 - 4 r s ] (p[q)[2)1/2] . (34)
If 2 ( rs ) 1/21 (P Iq ) I is less than unity, then clearly for
 G. Jaeger,A. Shimony / PhysicsLettersA 197 (1995)83-87 87

l sufficiently larger than g we have G~ < G2. In other + r 2 ( ] ( p l q ) ]2_ I ( P [ q ( I 4 )

words, if the penalty for an incorrect choice is suffi-
>~1 for ½~<r~<l.
ciently severe, then practical reason dictates a proce-
dure which makes no incorrect classifications and as In general, there exist real numbers Pa and Pb, both
many correct classifications as possible. Again if greater than unity, such that procedure 2 is prefera-
2(rs)~/21 ( P l q ) l is less than unity but g=l, then ble to procedure 1 if and only if
G2~< G~, as we can check for the two cases ofEq. (7a)
1/g~pa (ifEq. (7a) h o l d s ) ,
and (7b). I f E q . (7a) holds, then
1/g>~pb ( i f E q . ( 7 b ) h o l d s ) .
( G~-GZ2) /g2= ( 1 - 4 r s l (p[q) [2)
- [1-4(rs)~/21 ( p l q ) l+4rs[ (p[q)[2] The research for this paper was supported by the
=4(rs)'/Zl ( P l q ) I - S r s l ( P l q ) 12 National Science Foundation, grant No. PHY-
9321992.
=4(x-2x2)>~ 1 ,

where x = (rs) 1/21 (Plq> I <~½.

If Eq. (7b) holds, then References

( G 2 - G ~ ) /gZ= ( 1 - 4 r s l ( P l q ) 12 [ 1] I.D. Ivanovic, Phys. Lett. A 123 (1987) 257.

[2] D. Dicks, Phys. Lett. A 126 (1988) 303.
- (r2-2r21 ( P l q ) I 2 + r a f ( P l q ) 1 4 ) [3] A. Peres, Phys. Lett. A 128 (1988) 19.
>/(1-4r+4rZ)l(plq) 12 [4] G. Jaeger, A. Shimony and U Vaidman, Phys. Rev. A 50
(1994).