Optimum unambiguous discrimination between linearly independent symmetric states

28 December

1998

PHYSICS

LETTERS

A

ELSEWIER

Physics Letters A 250 ( 1998) 223-229

Optimum unambiguous discrimination between linearly independent symmetric states Anthony Chefles I, Stephen M. Barnett Department of Physics and Applied Physics. University of Strathclyde, Glasgow G4 ONG. Scotland, UK Received

13 July 1998; accepted for publication 23 October Communicated by P.R. Holland

1998

Abstract The quantum formalism permits one to discriminate sometimes between any set of linearly independent pure states with certainty. We obtain the maximum probability with which a set of equally likely. symmetric, linearly independent states can be discriminated. The form of this bound is examined for symmetric coherent states of a harmonic oscillator or field mode. @ 1998 Published by Elsevier Science B.V. PAC.9 03.65.B~;

03.67.-a;

03.67.Hk

1. Introduction

It is possible to manipulate the state of a quantum system in far more interesting ways than can be achieved by carrying out unitary operations and von Neumann measurements on the system of interest alone. Consideration of the effects of interactions with other systems has led to the development of the quantum operations formalism [ 11, which allows any completely positive, trace-preserving map to represent, in principle, a realisable transformation of the density operator 6. One particularly interesting type of operation is a probabilistic operation. This is an operation which, with some probability less than 1, will transform the state of the system in a manner which cannot be brought about by any deterministic process. Although such operations generally have a non-zero failure probability, one generally knows whether or

not the desired transformation has taken place. An important class of probabilistic operations are those which allow one to discriminate unambiguously between non-orthogonal states, that is, with zero probability of error. When carried out on a quantum system prepared in one of the non-orthogonal states I@j), such an operation will, with some probability, transform the state into a corresponding member of an orthonormal set l+j). The latter states can be discriminated without error using a simple von Neumann measurement. Although such an operation cannot have unit probability of success, we can always tell whether or not the desired transformation has taken place. When the attempt fails, we obtain an inconclusive result. The subject of unambiguous state discrimination was pioneered a decade ago by Ivanovic [2], Dieks [ 31 and Peres [4], and has recently undergone interesting further developments. While earlier work concentrated only on the problem of discriminating between two states, the problem of discriminating

’ E-mail: [email protected]. 0375-9601/98/$ - see front matter @ 1998 Published PIf SO3759601(98)00827-5

by Elsevier Science B.V. All rights reserved

224

A. Chefie..r, SM.

Bamei?/Physics

between multiple states has since been addressed. In particular, one of us [5] has shown that the necessary and sufficient condition for a set of states It./Qj) to be amenable to unambiguous state discrimination is that they must be linearly independent. More recently, the problem of discriminating between three states has been examined in detail by Peres and Terno [6]. We have shown that unambiguous discrimination is intimately related to other well-known types of probabilistic operation, such as entanglement concentration [5,7] and exact cloning [ 81. The latter connection has also been examined by Guan and Duo [ 91. Unambiguous discrimination between two non-orthogonal states has been demonstrated in the laboratory by Huttner et al. [ IO]. In this experiment, weak pulses of light were prepared in non-orthogonal polarisation states, a fraction of which were converted into orthogonal ones by a loss mechanism. It is clearly of interest to find the optimum strategy for discriminating unambiguously between a set of known states, that is, to determine the maximum probability of obtaining a definite result. As with problems in conventional quantum detection theory [ 111, where the aim is to find the absolute maximum of the discrimination probability or mutual information for a given source, few analytic solutions for optimum strategies and their respective figures of merit are known. The complete solution for unambiguous discrimination between two states with arbitrary a priori probabilities has been found by Jaeger and Shimony [ 121. Peres and Terno [ 61 have explored the geometry of the optimisation problem for three states, and obtained useful insight into the general N state case. As yet, however, no analytical solutions have been found for more than two states. Such a solution is given in this paper. We determine the maximum probability with which N symmetric states can be unambiguously discriminated, assuming they have equal prior probabilities. We then apply our result to examine the maximum probability of discriminating between N symmetric coherent states

Letters

A 250

(1998)

2.23-229

If the states are non-orthogonal, no quantum operation can deterministically discriminate between them. It is, however, possible to devise a strategy which, with some probability, will reveal the state with zero error probability. To see how this may be done, it is convenient to employ the Kraus representation of quantum operations [ 11. Each of the possible, distinguishable outcomes of an operation, which are labelled by the index ,u, is associated with a linear transformation operator ApCL. These form a resolution of the identity

c A$4p= f.

(2.1)

P

If the system is prepared with the initial density operator b, the probability Pfi of the ,uth outcome is TrbALA,.

The final density

operator

corresponding

to this result is Ap,5AfL/Pp. The state discrimination operation will have N + I distinct outcomes, corresponding to detection of each of the states, and an additional answer which gives no information about the state. This is the inconclusive result. The operator which corresponds to the detection of the state 1ti.j) is d,i, where j = 0,. . . , N - 1, and we let A, be the operator which leads to a failure of the discrimination attempt. Clearly, we have

A~lLi,+ C

A)A,j

=

(2.2)

i.

The zero-errors condition

takes the form

where F’j is the conditional probability, given that the system was prepared in the state /$j), that this state will be identified. This zero-errors condition can only be met if the I$j) are linearly independent, and we find that the Alj have the form [5]

~2.4)

lC?j).

2. Unambiguous

state discrimination

Consider a quantum system prepared in one of N pure quantum states II,!J~),where j = 0,. . . , N - 1. These states span an N-dimensional Hilbert space ti.

where the ]4j) form an orthonormal basis for X. Here we have also introduced the reciprocal states I$;‘). The reciprocal state ]$,A) is defined as that which lies in ‘7-fand is orthogonal to all i$jJ) for j#j’. The set of states I+,+) is simply the N-dimensional complex generalisation of the set of (normalised) reciprocal

A. Chejles. SM. Barnett/Physics

vectors in crystallography [ 131 with respect to the unit cell basis vectors, these being the I$,j). A complete set of reciprocal states exists if, and only if, the 1fi.j) are linearly independent. The reciprocal states are also necessarily linearly independent, as is shown in Ref.

[51. Given that the states 1fi.j) have a priori probabilities q,j, the total probability of correctly identifying the state is

(2.5) It convenient to proceed using the language of positive operator-valued measures (POVMs) [ 11. The measurement can be expressed as an N + 1 element POVM operation by defining the ositive Hermitian operators -Pand &,c = A,A,r. It is also useful to dekDj

=

A,jAj

fine _&o = c,; a,;aj.

The discrimination

probability

PD is constrained by the fact that 8, must be positive. This, together with the decomposition of the identity, 8, + ED = i, means that none of the eigenvalues of fro may exceed unity. It has been shown that the optimum measurement corresponds to the maximum eigenvalue of i?o being equal to 1 [ 51.

3. Maximum discrimination symmetric states

probability

for

In this section we derive the maximum attainable value of the unambiguous discrimination probability PD for symmetric states with equal a priori probabilities. A set of quantum states I+j) spanning a Hilbert space ‘7-Lis symmetric [ 141 if there exists a unitary transformation I!? on ‘7-Lsuch that 1ti.j)

= ol$.j-1)

I$o) =

@h--IL

fiN = i.

= ~jltiO)t

(3.1)

Letters A 250 (1998) 223-229

22s

discrimination probability or mutual information for all possible measurements and not just the subset defined by the no-errors constraint, the maximum discrimination probability for a set of symmetric states with equal a priori probabilities can be obtained exactly [ 11,141. The optimum strategy uses the so-called “square root” [ 11,141 or “pretty good” [ 151 measurement. Considerable progress has also been made towards maximising the mutual information for these states [ 161, in particular in connection with symmetric quantum channels [ 171. As has been shown in Ref. [ 51, unambiguous discrimination between linearly independent symmetric states arises naturally in connection with entanglement concentration, that is, transforming a fraction of an ensemble of systems all prepared in the same imperfectly entangled state into a maximally entangled state, using only local operations and classical communication. The protocol given there, a generalisation of the “Procrustean” technique due to Bennett et al. [ 181, will maximise the entanglement of a pair of subsystems with probability equal to the probability of discriminating between a certain set of linearly independent symmetric states. Prior to solving for the maximum unambiguous discrimination probability for equally probable linearly independent symmetric states, we shall obtain a representation of them which simplifies our analysis. We use the fact that the operator i? can be expanded as follows: N-l 0=

C

f+lyk)(ykl,

(3.4)

k=O

where (YklYk’) = a&‘. The real angles 4k may be taken to lie in half-open interval [ 0,29~). It follows from Eq. (3.3) that 2r.fk

(3.2)

5f’k=N’

(3.3)

where fk is an integer satisfying 0 < fk < II- 1. It is convenient to arrange the fk in increasing order, so that fk > fk’ for k > k’. Clearly, we can expand I&) as xkck]Yk), for some ck satisfying ck /ck12 = 1. Together with Eqs. (3.1) and (3.4), this leads to

Eq. (3.3) follows from the fact that any state in X can be written as a superposition of the l#j), and from QNI$,j) = I$,,). Such states are also said to be covariant with respect to 0 and have been found to have a preferential status with regard to problems in quantum detection theory. In conventional quantum detection theory, where the aim is to maximise the

(3.5)

N-l 1e.j)

=

C k=O

Cke2vijfklNIyk).

(3.6)

226

A. Chejles. SM.

Burnett/Physics

Note that the linear independence of the I$,i) implies that all of the ck are non-zero. Linear independence means that no superposition of the I++;) can vanish, so consider .

i

N-l

N-l

c

e-2ni,rijr/N

1fi.i)

= 1

,=o

Ck’%f,, b’k).

(3.7)

k=O

If the I$i) are linearly independent, this must be nonzero for all r = 0, . . N - 1. Therefore, the fk must take every value in this range of integers. As we have arranged these integers in increasing order, we find that fk is simply equal to k, so that linearly independent symmetric states necessarily have the form N-i

I$,) =

c

ck

e2Ti’ik’NIyk).

(3.8)

k=O

That having this form is also a sufficient condition for linear independence is proven in Ref. [ 51. For these states, the corresponding reciprocal states are given by N-l

I$;‘} = z-112

c

c;-l

e*rir/NIyr),

Letms

A 250 (1998)

223-229

which we denote by h+ ( J!?D) , is equal to I. We then define k(l) = fi’~,@l D

These operators clearly have the same eigenvalues as ,!?o and give the same value of PD. In fact, J!?:’ can be obtained from i?D by cycling the probabilities Pi. Writing explicitly the dependence of these operators on the c,, we see that ,!?$‘(l’,) = !?D(e,_[), where CifN = pi. Consider now the operator

The second equation here is true if the a priori probabilities vj are all equal to 1/N, which we take to be the case. The operator ED ^‘w is invariant under the similarity transformation @e+ir&‘eot and gives the same transformation probability as kg’. Forming @T from i?o amounts to replacing ah of the cj by their average value PO, the quantity we wish to maximise. However, its maximum eigenvalue A+ ( PGe) satisfies

(3.9)

(3.14)

r=O

where Z = C, lc,.-‘.

Note that the I$,:)

are also

symmetric, with respect to the same transformation fi as the i+,i). For symmetric states, the operator i?o has the explicit form

ED =

-$ c

qiC;,-lc;l

e2~i.i(r-r’)lNIYr,)(Yr~.

i.r.r’

(3.10) Let us denote by 2:’ an operator of this form which gives the maximum value of PD. We do not assume this operator to be unique, that is, we do not assume the optimum Pi to be unique. We can however, show that there exists an optimal operator @’ which possesses the symmetry jpt = fi&“P’i)+ D D

.

(3.11)

We prove this by contradiction, by first supposing that no Pop’ satisfies Eq. (3.11) . Consider now any operator i?D of the form (3.10). If ED corresponds to the maximum value of PO, then its maximum eigenvalue,

(3.12)

This is a consequence of the fact the maximum eigenvalue is convex on the space of Hermitian operators on 7-t. A simple proof of this is given in Appendix A. Let us define .!?b = _@/A+( ~$2). The maximum eigenvalue of this operator is 1, so it is physically admissible. However, the success probability for this operator is Pb = PD/h+ (tic) > PO, from (3.14). Therefore, we can obtain from any &D another operator i?b which is invariant under the similarity transformation and whose associated discrimination probability Pb is at least as high as PD. Thus, the premise that there is no 8, which gives the highest value of PD and has the specified symmetry is false. It follows from Eq. (3.13). and from the orthogonality of the Irr), that the eigenvalues of 8; are simply PD/NIc~I*. The optimum symmetric operator FE’ is simply that whose maximum eigenvalue is 1. The desired least upper bound on PO is then given by PD < Nxmin(c,12.

(3.15)

The bound here is clearly less than 1 unless all lc,12 are equal to N-‘, in which case the 1G.i) are orthogo-

A. Chejies, S.M. Burnett/Physics

nal. Although this inequality gives the analytic maximum discrimination probability, in practice it may be necessary to employ computational techniques to determine the smallest of the lc,/‘, as we will see in the next section. To find the maximum value of PO for a specific set of states, it is useful to have an expression for the 1~~1’which explicitly exhibits their dependence upon the states. Using Eq. (3.8), we find that (3.16) .i# The simplest set of linearly independent symmetric states comprises just two states. The problem of finding the maximum value of PO for a pair of states has been solved by Ivanovic [2], Dieks [3] and Peres [4] for equal a priori probabilities and generalised by Jaeger and Shimony [ 121 to the case of unequal probabilities. Denoting the two states by I$*), the Ivanovic-Peres-Dieks limit for the probability of error-free state discrimination is

Letters A 250 (1998) 223-229

coherent states of the harmonic a boson field. These states are

or mode of

(4.1)

where j = 0,. . . , N - 1 and a,j = cr e2Vi,jlN, where cy = LYOmay be any complex number. The In) are the usual boson number states. The magnitudes of the complex arguments aj are all equal to 1~1. However, their phases are distributed around the circle at regular intervals of 21r/N. Let us denote by & the projector onto X, the subspace spanned by the 1a.j). The unitary transformation 0 which maps each state onto its successor is c = & e?nifiIN&,

(4.2)

where A is the boson number operator. The quantities of interest if we wish to determine the maximum of PD are the square-moduli of the c,. One can show using Eq. (3.16) that lc,12 = $ C

It is interesting to see how this limit arises as a special case of the bound in (3.15). Up to an irrelevant phase difference, the states I@*) may be represented as I-),

(3.18)

where the angle 8 lies in the range [0, r/4] and the states 1%) constitute an orthogonal basis for the space spanned by I$*). The system may be represented as a spin- l/2 particle, and 14~)taken to be the eigenstates of 8, with eigenvalues i 1. Note that c?~ is the unitary operator fi relating the states we aim to distinguish, since I$*) = 8, IQT). We find that the corresponding reciprocal states are I$,‘) =sinBI+)&cos+).

(3.19)

The expansion coefficients ck are given by C+ = cos 0 and c- = sin 0. Within the specified range of 8, /c+12 3 Ic-12, so that the maximum value of PO is 2lc_l’, which is easily seen to be equal to PIDP.

4. Symmetric

oscillator

(3.17)

fiDP = 1 - I((cl+I$-)I.

I$+) = cos0 I+)fsinB

221

coherent states

In this section, we apply the bound (3.15) on PD to the problem of discriminating between symmetric

e-2~i.jr/Nela12(eZm”N-l),

(4.3)

.I Unfortunately, this summation seems to resist significant simplification, and in general must be carried out numerically. A further complication arises if we wish to determine the maximum value of PO, which entails finding the smallest of the lc,/‘. This is the fact that for general N, none of the /c,12 remains the smallest for all values of [al’. This can be seen in Fig. 1, which shows the variation of the lc,I’ as functions of lay1’ for N = 10. The behaviour seen here for N = 10 is typical of what happens for all N except for N = 2. In this simplest case, Ical* = ( 1 + ee21ru12) /2 and Ici I2 = (1 - e-‘la1*)/2, so that IcaJ’ 3 /c112 for all 1~~1’.For general N, each lc,12 is less than the others for some range of 1~~1~.At 1~~1~= 0, we find that lcoi’ = 1 and all of the other cr are zero. Our numerical results for various values of N indicate that, as /aI* increases, the smallest of the lc,12 is successively IcN-1 12,then (~~-21~ and so on until it is lcol’ then the cycle repeats itself indefinitely. It is evident from the figure that the point at which the minimum coefficient changes to a new one occurs when the derivative of the latter is zero. This can be understood when we observe that the /clj2 obey the relation

A. Chejies. SM. Burnett/Physics

1

Letters A 250 (1998) 223-229

I~~*+cQ, the /c,12 tend to l/N. Here, the overlaps between the states becomes indefinitely small, and the maximum discrimination probability approaches 1. The maximum discrimination probability for N = 10 is shown in Fig. 2. We see that it is an increasing function of IcyI*, although its derivative is discontinuous whenever a new Ic,I’ becomes the smallest.

5. Discussion

I1 1,

0.0 0

1, 20

10

30

lal' Fig. I. Dependence

of the jcrI’ on /aI2

for IO symmetric coherent

states.

d( Icrl’) = IC,_] /* - lc,12. d( H2) It follows to lay1’is at which function

(4.4)

that when the derivative of lc,12 with respect zero, we have /c,l* = /c,-I 12.This is the point these functions cross and thus the smallest ceases to be lc,12 and becomes lc,_~[*. As

0.6 1

a”

I

We have obtained the least upper bound on the unambiguous discrimination probability for linearly independent symmetric states with equal a priori probabilities. The corresponding detection operators ED, are the simplest possible ones. The only free parameters in these operators for general states are the conditional probabilities Pj.For equally probable symmetric states, the maximum value of PO is obtained when all Pjare equal, and the one remaining free parameter is set by A+ (Pi’) = 1, which is a necessary condition for the maximum value of PO for any linearly independent set. We thus find that for equally probable symmetric states, the solution for the maximum of PO is the simplest N state solution. This appears also to be the case for optimisation problems in other areas of quantum detection theory, in particular the determination of the unconstrained maximum discrimination probability [ 11,141. However, one issue we have not addressed here is whether or not it is only for equiprobable symmetric states that setting all P, to the same value gives the optimum measurement. We have also examined the behaviour of the maximum Ps for N symmetric coherent states, in particular its dependence on the parameter 1aI*, which has many interesting features for N > 2. Most notably, the derivative of the maximum discrimination probability is not continuous, owing to the fact that no single /c,l* is smallest for all values of 1~~1~.While this makes perfect sense from a mathematical point of view, the physical reasons for this phenomenon are by no means obvious and further work may clarify the matter.

Acknowledgement 0

10

20

30

la? Fig. 2. Maximum value of the probability PD of distinguishing between ten symmetric coherent states as a function of 1~~1~.

We gratefully acknowledge financial support by the UK Engineering and Physical Sciences Research Council (EPSRC) .

A. Chejk,

Appendix A. Convexity eigenvalue

SM. Bnmetr/Physics

of the maximum

Letters A 250 (1998) 223-229

Rearranging h+($+r)

Consider a set of Hermitian operators I?[ on a finitedimensional Hilbert space ‘Ft, where I = 0,. . . , n - 1. Let us denote by A+ (&) the maximum eigenvalue of 8,. Then, as we show here, h+ is convex, that is, for any real, positive constants al,

A+

6

&VA+(G).

(A.11

I

This admits the following simple proof. Let us write 2, = c;zO’ al& and let 1~~) be any eigenstate of $ corresponding Then,

to the maximum

eigenvalue

A+ ($-).

+ a,h+(&),

(A.2)

where the equality is satisfied only if 3, and $ have a simultaneous eigenvector corresponding to the maximum eigenvalues of both of these operators. From the definition of s,, we see that II-

I

aoA+L@o’o) + C[A+(S+d -

A+(S)1

/=I n-l

=

A+

(

-&k, i=o

>.

(A.31

- A+(G)

in (A.2)

gives

< QA+(‘&).

Substituting this inequality into (A.3) gives (A. 1) , completing the proof.

(A.4) immediately

References 11] K. Kraus, [2] 131 141 [5] [6] [7] [ 81 [9]

[IO] 6 A+($)

the inequality

229

States, Effects and Operations: Fundamental Notions of Quantum Theory (Springer, Berlin. 1983). I.D. Ivanovic, Phys. Lett. A 123 (1987) 257. D. Dieks, Phys. Lett. A 126 ( 1988) 303. A. Peres, Phys. Lett. A 128 (1988) 19. A. Chefles, Phys. Len. A 239 ( 1998) 339. A. Peres, D.T. Temo, LANL Report No. quant-ph/9804031. A. Chefles, SM. Bamett, J. Mod. Opt. 45 (1998) 1295. A. Chefles, S.M. Bamett, submitted to Phys. Rev. Lett. L-M. Duan, G-C. Guo, Phys. Rev. Lett 80 (1998) 4999. B. Huttner, A. Muller J.D. Gautier, H. Zbinden, N. Gisin,

Phys. Rev. A 54 ( 1996) 3783. [ 1 I ] C.W. Helstrom, Quantum Detection and Estimation Theory (Academic

Press, New York, 1976).

[ 121 G. Jaeger, A. Shimony, Phys. Lett. A 197 (1995) 83. [ 131 See any introductory solid state physics text,

for example, A. Guinier, R. Jullien, The Solid State: From Superconductors to Superalloys (Oxford Univ. Press, Oxford, 1989). [ 141 M. Ban, K. Kurokawa, R. Momose, 0. Hirota, Int. J. Theor. Phys. 36 (1997) 1269. [ 151P. Hausladen. W.K. Wootters, J. Mod. Opt. 41 (1994) 2358. [ 161 M. Osaki, 0. Hirota, M. Ban, J. Mod. Opt. 45 ( 1998) 269. [ 171 M. Sasaki, S.M. Bamett, R. Josza, M. Osaki, 0. Hirota, in preparation. [ 181 C.H. Bennett, H.J. Bernstein, S. Popescu, B. Schumacher, Phys. Rev. A 53 (1996) 2046.