The Communication Cost of Simulating Bell Correlations

B. F. Toner∗ and D. Bacon†

Institute for Quantum Information, California Institute of Technology, Pasadena, CA 91125 and
Department of Physics, California Institute of Technology, Pasadena, CA 91125

What classical resources are required to simulate quantum correlations? For the simplest and
most important case of local projective measurements on an entangled Bell pair state, we show
that exact simulation is possible using local hidden variables augmented by just one bit of classical
communication. Certain quantum teleportation experiments, which teleport a single qubit, therefore
admit a local hidden variables model.

PACS numbers: 03.65.Ud,03.65.Ta,03.67.-a,03.67.Hk,03.67.Dd

arXiv:quant-ph/0304076v1 10 Apr 2003

Recent theoretical research into quantum algo- spin states | ↑i, | ↓i are defined with respect to a local
rithms [1], quantum communication complexity [2], and set of coordinate axes: | ↑i (resp. | ↓i) corresponds to
quantum cryptography [3] has shown that quantum de- spin along the local +ẑ (resp. −ẑ) direction. Alice and
vices are more powerful than their classical counterparts. Bob each measure their particle’s spin along a direction
Indeed, the flourishing field of quantum information the- parameterized by a three-dimensional unit vector: Alice
ory [4] aims to provide an information-theoretic quantifi- measures along â, Bob along b̂. Alice and Bob obtain
cation of the power underlying quantum resources. One results, α ∈ {+1, −1} and β ∈ {+1, −1}, respectively,
important feature of quantum theory lies in the statisti- which indicate whether the spin was pointing along (+1)
cal correlations produced by measurements on local com- or opposite (−1) the direction each party chose to mea-
ponents of a quantum system. Almost forty years ago, sure. Locally, Alice and Bob’s outcomes appear random,
John Bell showed that such correlations cannot be ex- with expectation values hαi = hβi = 0, but their joint
plained by descriptions based on realistic properties of probabilities are correlated such that that hαβi = −â · b̂.
local subsystems [5]. To experimentally distinguish quan- We refer to these correlations as Bell correlations.
tum correlations from those produced by local hidden
variables theories, Bell introduced the notion of Bell in- It is not possible to reproduce these correlations us-
equalities, with subsequent experimental evidence falling ing a protocol which draws on random variables shared
squarely on the side of quantum theory [6]. Bell inequal- between Alice and Bob, but does not allow communi-
ities, whilst usually considered relevant only to founda- cation after they have selected measurements [5]. So
tional studies of quantum theory, answer a fundamental how much communication is required to exactly simu-
information-theoretic question: what correlations can be late them [11, 12, 13, 14, 15, 16, 17]? Naively, Alice
produced between separate classical subsystems, which can just tell Bob the direction of her measurement â
have interacted in the past, if no communication be- (or vice versa) but this requires an infinite amount of
tween the subsystems is allowed? Violation of a Bell in- communication. The question of whether a simulation
equality, however, does nothing to quantify what classical can be done with a finite amount of communication was
information-processing resources are required to simulate raised independently by Maudlin [11], Brassard, Cleve,
a particular set of quantum correlations. and Tapp [12], and Steiner [13]. Their approaches dif-
The simplest and most important example of quantum fer in how the communication cost of the simulation is
correlations involves the correlations produced by pro- defined: Brassard et al. take the cost to be the num-
jective measurements on a Bell pair. Bell pairs are the ber of bits sent in the worst case; Steiner, the average.
maximally-entangled states of two quantum bits (qubits) (Steiner’s model is weaker because the amount of commu-
and are the basic resource currency of bipartite quantum nication in the worst case can be unbounded.) Brassard
information theory. Various equivalences are known: one et al. present a protocol which simulates Bell correla-
shared Bell pair plus two bits of classical communication tions using exactly eight bits of communication (since
can be used to teleport one qubit [7] and, conversely, one improved to six bits [14]). Surprisingly [15], the only
shared Bell pair plus a single qubit of communication can lower bound for the amount of communication in the
be used to send two bits of classical communication via worst case is given by Bell’s theorem: at least a single
superdense coding [8]. bit of communication is needed. Here we present a sim-
Consider the gedanken experiment of Einstein, Podol- ple protocol that achieves that bound.
sky, and Rosen [9] (EPR), as reformulated by Bohm [10]. We first note three simple properties of Bell correla-
Two spatially separate parties, Alice and Bob, each have tions: (i) if â = b̂, then we must have α = −β: Alice
a spin- 12 particle, or qubit. The global spin wave func- and Bob must output perfectly anticorrelated bits; (ii)
tion is the entangled Bell singlet state (also known as either party can reverse their measurement axis and flip
an EPR pair) |ψi = √12 (| ↑iA | ↓iB − | ↓iA | ↑B i). The their output bit; and (iii) the joint probability is only
 2

dependent on â and b̂ via the combination â · b̂. In his

original paper, Bell gave a local hidden variables model (a) Alice’s output
that reproduces these three properties for all possible λ̂1 λ̂2
axes, but his model fails to reproduce Bell correlations
because the statistical correlations when â 6= b̂ are not as
strong as those of quantum mechanics [5]. The protocol output −1
we describe below is inspired by Bell’s original protocol.
output +1
Property (iii) implies that we may restrict attention to
rotationally invariant protocols, for which all probabili-
ties depend only on â · b̂ and not â and b̂ separately, by
randomizing over all inputs with the same dot product.
More precisely, suppose P is any protocol that simulates
the correlations. Then define a new protocol P ′ whose (b) Communication
hidden variables consist of (i) those required by P , and
A sends +1 A sends −1
(ii) a random rotation R ∈ SO(3). Protocol P ′ then con-
sists of running protocol P on Râ and Rb̂ in place of â λ̂1 λ̂2 λ̂1 λ̂2
and b̂.
We now describe our protocol. Alice and Bob share
two random variables λ̂1 and λ̂2 which are real three-
dimensional unit vectors. They are chosen independently
and distributed uniformly over the unit sphere.
The protocol proceeds as follows:
1. Alice outputs α = −sgn(â · λ̂1 ).
2. Alice sends a single bit c ∈ {−1, +1} to Bob where
(c) Bob’s output
c = sgn(â · λ̂1 )sgn(â · λ̂2 ).
λ̂1 λ̂1 + λ̂2 λ̂2
h i
3. Bob outputs β = sgn b̂ · (λ̂1 + cλ̂2 ) , λ̂1 λ̂2

where we have used the sgn function defined by sgn(x) =

+1 if x ≥ 0 and sgn(x) = −1 if x < 0. A geometric
description of our protocol is given in Fig. 1. We note
immediately that Bob obtains no information about Al- λ̂1 − λ̂2
ice’s output from the communication.
We now prove the protocol reproduces the correct ex-
pectation values. Each party’s output changes sign under
the symmetry λ̂1 ↔ −λ̂1 , λ̂2 ↔ −λ̂2 , so hαi = hβi = 0
because λ̂1 and λ̂2 are uniformly distributed. The joint
expectation value hαβi can be calculated using FIG. 1: The protocol: The shared unit vectors λ̂1 and λ̂2
( described in the text divide the Bloch sphere into four quad-
hαβi = E − sgn(â · λ̂1 ) × rants, as shown. Alice and Bob’s actions depend on which
quadrant their respective measurement axes lie in, and in
) Bob’s case, the bit he receives from Alice. (a) Alice’s out-
X (1 + cd) h i put: if â lies in the shaded region, Alice outputs −1; in the
sgn b̂ · (λ̂1 + dλ̂2 ) (1) unshaded region, she outputs +1. (b) The communication:
2
d=±1 Alice sends c = +1 if her measurement axis lies in the N or S
1
R R quadrants, and −1 otherwise. (c) Bob’s output: this depends
where E {x} = (4π)2 dλ̂1 dλ̂2 x, c = sgn(â · λ̂1 ) sgn(â ·
on the bit received from Alice. The shading is as for (a).
λ̂2 ) and we have used the trick that (1+cd)/2 = 1 if c = d
and 0 if c 6= d. After substituting for c and expanding
Eq. (1), we obtain the sum of four terms (because each
This integral may be evaluated with the help of the two
term inside the summation sign is itself the sum of two
diagrams shown in Fig. 2, with the result that hαβi =
terms) and, using sgn(â · λ̂1 ) c = sgn(â · λ̂2 ), we note that
−â · b̂, as required.
the four terms are related by the symmetries λ̂1 ↔ λ̂2
or λ̂2 ↔ −λ̂2 , so each has the same expectation value. Our protocol exactly simulates the quantum mechani-
Hence cal probability distribution for projective measurements
n h  io on the singlet Bell pair state. If a large number of sim-
hαβi = E 2 sgn(â · λ̂1 ) sgn b̂ · λ̂2 − λ̂1 . (2) ulations are performed in parallel, the communication
 3

b̂ = (0, 0, 1) â = (0, 0, 1)

λ̂1 = (sin t, 0, cos t)

b̂ = (sin r, 0, cos r)
t r
θ θ

cos t = b̂ · λ̂1 cos r = â · b̂

λ̂2 = (sin θ cos φ, sin θ sin φ, cos θ) λ̂1 = (sin θ cos φ, sin θ sin φ, cos θ)
1 1
Z h i Z
(a) dλ̂2 sgn b̂ · (λ̂2 − λ̂1 ) = −b̂ · λ̂1 (b) dλ̂1 sgn(â · λ̂1 )b̂ · λ̂1 = â · b̂
4π 2π

FIG. 2: Construction
 used to evaluate
 Eq. (2): (a) We first integrate over λ̂2 , taking b̂ to point along the positive z-axis [18].
Observe that sgn b̂ · (λ̂2 − λ̂1 ) is positive in the top spherical cap (shaded) and negative otherwise. The area of the top
Rt R  
spherical cap is A+ = 2π 0 sin θdθ = 2π(1 − cos t) where cos t = b̂ · λ̂1 , hence dλ̂2 sgn b̂ · (λ̂2 − λ̂1 ) = A+ − (4π − A+ ) =
−4π cos t = −4π b̂ · λ̂1 . (b) We now take â to point along the positive z-axis [19], set b̂ = (sin r, 0, cos r), and integrate over λ̂1 ,
R Rπ R 2π
obtaining dλ̂1 sgn(â · λ̂1 )b̂ · λ̂1 = 0 sin θdθ 0 dφ sgn(cos θ) (cos θ cos r + sin θ cos φ sin r) = 2π cos r = 2πâ · b̂.

may be compressed. To see this, assume Alice’s mea- The protocol is as follows:
surement vector â is uniformly distributed (if not, we
randomize, as outlined above). Then, if λ̂1 · λ̂2 = cos η, 1. Alice sends c1 = sgn(â · λ̂1 ) and c2 = sgn(â · λ̂2 ) to
Alice sends −1 with probability η/π and 1 with probabil- Bob.
ity 1−η/π, so that the communication can be compressed h  i
R π/2 2. Bob outputs β = sgn b̂ · c1 λ̂1 + c2 λ̂2 .
to 0 sin η dηH(η/π) ≈ 0.85 bits, where H(η/π) is the
Shannon entropy. This encoding depends on the shared
unit vectors λ̂1 and λ̂2 : a third party without access to It is easy to verify that hβi = â · b̂, as required. Us-
the hidden variables will observe Alice sending uniformly ing this protocol, we obtain a (not necessarily optimal)
distributed bits to Bob. protocol to simulate joint projective measurements on
partially entangled states of two qubits, which uses two
Our protocol is easily modified to simulate joint mea- bits of communication: Alice first simulates her mea-
surements on any maximally entangled state of two surement and determines the post-measurement state of
qubits, because every such state is related to the sin- Bob’s qubit; Alice and Bob then execute the classical
glet by a local change of basis and thus may be sim- teleportation protocol.
ulated by rotating and/or reflecting the input vectors The classical teleportation protocol also shows that
â and b̂, before running our protocol. It may also be certain quantum teleportation experiments admit a local
converted to a two bit protocol for “classical telepor- hidden variables description. We first note that quan-
tation” of a qubit [16]. In classical teleportation of a tum teleportation experiments do not purport to test
qubit, Alice is given a description of a qubit and Bob whether quantum mechanics allows a local hidden vari-
is given a description of a projective measurement on a ables model; rather they aim to distinguish quantum
qubit. The parties are allowed to communication clas- teleportation from other protocols Alice and Bob might
sically and to share randomness (but no entanglement) carry out using classical communication, but no entangle-
and the goal is for Bob to produce the correct output ment [20]. From this point of view, teleportation experi-
of his measurement on Alice’s state. It is sufficient to ments represent “investigations within quantum mechan-
consider the case where Alice is given the description of ics” [21], rather than comparisons of quantum mechanics
a pure state, which we suppose has spin aligned along with classical local hidden variables models. With this
the axis â. We suppose Bob’s measurement is along the distinction in mind, it is still interesting to ask whether
axis b̂. Alice and Bob share uniformly distributed ran- teleportation experiments can be explained by a local
dom three-dimensional unit vectors λ̂1 and λ̂2 , as before. hidden variables model.
 4

Consider then the following experiment: Alice prepares sium on the Foundations of Computer Science, edited by
a qubit in a state unknown to Bob. She then teleports the S. Goldwasser (IEEE Computer Society, Los Alamitos,
qubit to Bob, who performs a projective measurement on CA, 1994), pp. 124–134; L. Grover, in Proceedings of the
it, along a direction unknown to Alice [22]. If one allows 28th Annual ACM Symposium on the Theory of Compu-
an infinite amount of classical communication from Al- tation (ACM Press, 1996), pp. 212–219.
[2] R. Raz, in Proceedings of the 31st ACM Symposium on
ice to Bob, then there is a trivial local hidden variables Theory of Computing (ACM Press, 1999), pp. 358–367.
model, for Alice can just send a classical description of [3] C. H. Bennett and G. Brassard, in Proceedings of IEEE
the state to Bob, who then simulates his measurement. International Conference on Computers, Systems, and
In contrast, our protocol for classical teleportation is a Signal Processing (IEEE, 1984), pp. 175–179.
local hidden variables model for this experiment that re- [4] M. A. Nielsen and I. L. Chuang, Quantum Computation
quires only two bits of communication, which is the same and Quantum Information (Cambridge University Press,
amount as the quantum teleporation protocol. (We also New York, 2000).
[5] J. S. Bell, Physics 1, 195 (1964).
note that the two bits sent are completely random, just as [6] A. Aspect, P. Grangier, and G. Roger, Phys. Rev. Lett.
in quantum teleportation.) Thus there is a local hidden 49, 91 (1982); A. Aspect, J. Dalibard, and G. Roger,
variables description of this experiment. Phys. Rev. Lett. 49, 1804 (1982); G. Weihs, T. Jen-
Are there quantum teleportation experiments which do newein, C. Simon, H. Weinfurter, and A. Zeilinger,
not have such a description? One obvious possibility is Phys. Rev. Lett. 81, 5039 (1998); W. Tittel, J. Bren-
an experiment that teleports entanglement itself. But del, H. Zbinden, and N. Gisin, Phys. Rev. Lett. 81, 3563
there is a more subtle possibility. If we allow Bob to (1998).
[7] C. H. Bennett, G. Brassard, C. Crépeau, R. Jozsa,
measure the qubit using elements of a positive operator- A. Peres, and W. K. Wootters, Phys. Rev. Lett. 70, 1895
valued measure, then there may not a local hidden vari- (1993).
ables description which respects the two bit classical com- [8] C. H. Bennett and S. J. Wiesner, Phys. Rev. Lett. 69,
munication bound. More generally, if Alice teleports n 2881 (1992).
qubits (which requires 2n bits of communication) and [9] A. Einstein, P. Podolsky, and N. Rosen, Phys. Rev. 41,
Bob makes a joint measurements on them, then it is 777 (1935).
[10] D. Bohm, Quantum Theory (Prentice-Hall, New York,
known that any exact local hidden variables theory re-
1951).
quires that Alice send at least a constant times 2n bits [11] T. Maudlin, in PSA 1992, Volume 1, edited by D. Hull,
of communication in the worst case [12]. Whether this M. Forbes, and K. Okruhlik (Philosophy of Science As-
holds for protocols with bounded error is an important sociation, East Lansing, 1992), pp. 404–417.
open question. [12] G. Brassard, R. Cleve, and A. Tapp, Phys. Rev. Lett.
The results presented here offer an intriguing glimpse 83, 1874 (1999).
into the nature of correlations produced in quantum the- [13] M. Steiner, Phys. Rev. A 270, 238 (2000).
ory. If we interpret Bell inequality violation to mean that [14] J. A. Csirik, Phys. Rev. A 66, 014302 (2002).
[15] D. Bacon and B. F. Toner, Phys. Rev. Lett. (to be pub-
some communication is necessary to simulate Bell corre- lished), quant-ph/0208057 (2003).
lations, then our results prove that the minimal amount, [16] N. Cerf, N. Gisin, and S. Massar, Phys. Rev. Lett. 84,
one bit, is all that is necessary for projective measure- 2521 (2000).
ments on Bell pairs. Is our straightforward protocol an [17] S. Massar, D. Bacon, N. Cerf, and R. Cleve, Phys. Rev. A
indication of a deep structure in quantum correlations? 63, 052305 (2001); A. Coates (2002), quant-ph/0203112.
We hope that our protocol and the development of a [18] K. H. Schatten, Phys. Rev. A 18, 103 (1993).
general theory of the communication cost of simulating [19] N. Gisin and B. Gisin, Phys. Lett. A 260, 323 (1999).
[20] D. Boschi, S. Branca, F. De Martini, L. Hardy, and
quantum correlations will help shed light on this funda- S. Popescu, Phys. Rev. Lett. 80, 1121 (1998)
mental question. [21] S. L. Braunstein, C. A. Fuchs, H. J. Kimble, and P. van
Acknowledgements.—We thank John Preskill, Andrew Loock, Phys. Rev. A 64, 022321 (2001)
Doherty, Patrick Hayden, and Andre Methot for useful [22] It is usual in teleportation experiments to have (i) a third
suggestions. This work was supported in part by the party Victor supply Alice with an unknown quantum
National Science Foundation under grant EIA-0086038, state, and (ii) Bob return the teleported state to Vic-
tor to measure, rather than measuring it himself. Such a
through the Institute for Quantum Information.
distinction is not important for the question we address,
however, because local hidden variables are hidden: al-
though it is convenient to describe a local hidden vari-
ables model as if Alice and Bob had access to the hidden
∗
variables, the model still exists even if the hidden vari-
Electronic address: [email protected] ables are inaccessible to them.
†
Electronic address: [email protected]
[1] P. W. Shor, in Proceedings of the 35th Annual Sympo-