You are on page 1of 4

6 Local hidden variable (LHV) theories

and Bell inequalities


Aim. Description in terms of classical probability theory + physically reasonable as-
sumptions.
Assumptions.
(i) One can assign definite values to quantities even before they are measured. These
values may, however, be ‘hidden’ or unknown to the observer so that a probabilistic
description may be necessary.
(ii) The properties of one subsystem should not immediately depend on what happens
to a very distant other subsystem.
Example scenario. A source emits pairs of particles to distant observers (Alice and
Bob) who perform ±1 valued measurements each.

Assume that both observers choose one out of m 2 N measurement devices denoted by
Ax , Bx , with x, y 2 {1, . . . , m}.
Let p(a, b | x, y) be the probability that Alice obtains outcome a 2 {±1} while Bob
obtains outcome b 2 {±1} in the same run of the experiment, if they have used devices
Ax and By , respectively. P
Consider the expectation value of their product, i.e., hAx By i := ab p(a, b | x, y).
a,b2{±1}

LHV Ansatz. Z
hAx By i = Ax (!)By (!) dP (!) (6.1)

with random variables Ax , By : ⌦ ! {±1} and a probability measure P .
! 2 ⌦ is the hidden variable and locality is expressed in the fact that Ax does not depend
on y and By does not depend on x.
Bell inequalities.
Definition 29. Consider an m⇥m tuple C 2 Rm⇥m of empirically obtained expectation
values hAx By i =: Cxy for pairs of ±1 valued measurements. Let C ✓ Rm⇥m be the set
of all such C’s for which there exists an LHV description as in (6.1).
An inequality for C is called Bell inequality if it holds for all C 2 C.

23
Remarks.
• C is a closed convex polytope, so for a complete description of C a finite set of
linear inequalities suffices.
• There are trivial Bell inequalities of the form |hAx By i|  1.
• Stochastic LHV theories allow the random variables in (6.1) to have ranges in
[ 1, 1] instead of { 1, +1}. It turns out, however, that this does not change C (!
exercise).
Corollary 30. Let ⇢ 2 B(HA ⌦HB ) be classically correlated and Mx : B({±1}) ! B(HA )
and similarly My0 : B({±1}) ! B(HB ) be POVMs for x, y 2 {1, . . . , m}.
Then C 2 C holds for
X ⇥ ⇤
Cxy := ab tr ⇢ Mx (a) ⌦ My0 (b) .
a,b2{±1}
P P
Proof. Define Ax := a Mx (a) = Mx (1) Mx ( 1) and By := b My0 (b).
a2{±1} b2{±1}
Then  Ax , B y  .
P (!) (!)
Suppose ⇢ = p! ⇢A ⌦ ⇢B is a convex combination of product states.
!
P h i h i
(!) (!)
Then Cxy = tr[⇢ Ax ⌦ By ] = p! tr ⇢A Ax tr ⇢B By is a stochastic LHV description
! | {z } | {z }
=:Ay (!) =:By (!)
with discrete probability space.
Consequently, unentangled quantum states can never violate any Bell inequality.
The simplest and most famous non-trivial Bell inequality is the Causer-Horne-Shimony-
Holt (CHSH) inequality:
Theorem 31 (CHSH). Every LHV theory for the description of ±1 valued measure-
ments satisfies:
hA1 B1 i + hA1 B2 i + hA2 B1 i hA2 B2 i  2. (6.2)
R
Proof. L.h.s. = ⌦ A1 (!) B1 (!) + B2 (!) + A2 (!) B1 (!) + B2 (!) dP (!) .
For a fixed ! 2 ⌦ we can distinguish two cases:
1. B1 (!) = B2 (!). Then B1 (!) B2 (!) = 0 and B1 (!) + B2 (!) 2 {±2}.
2. B1 (!) 6= B2 (!). Then B1 (!) + B2 (!) = 0 and B1 (!) B2 (!) 2 {±2}.
In either case the integrand is in { 2, 2} so that the average is in [ 2, 2].
The remarkable thing is, that this can be violated within quantum theory. For the
quantum description, we assign POVMs Mx and My0 to the measurement devices and
introduce again Ax := Mx (1) Mx ( 1), By := My0 (1) My0 ( 1).
With ˆ := Ah1 ⌦i(B1 + B2 ) + A2 ⌦ (B1 B2 ) 2 B(HA ⌦ HB ) the l.h.s. of (6.2) is then
given by tr ⇢ ˆ .

24
Theorem 32 (CHSH violation & Cirelson bound). If ˆ is as above, then for all density
operators ⇢ 2 B(HA ⌦ HB ): h i p
tr ⇢ ˆ  2 2.
p
That is, quantum theory violates the CHSH inequality at most by a factor 2.
Moreover, there exists a pure state on C2 ⌦ C2 and observables Ax , By 2 B(C2 ) with
eigenvalues ±1, s.t. equality holds in this equation.
h i
Proof. Consider the map A1 7! tr ⇢ ˆ . Since this is an affine functional the supremum
and infimum over the convex set  A1  is attained for some extreme point for
which spec(A1 ) = {±1} and thus A21 = .
By the same reasoning we can assume that A1 , A2 , B1 , B2 are all such that their square
is . Using this property, direct computation leads to
ˆ2 = 4 ⌦ + [A2 , A1 ] ⌦ [B1 , B2 ].

We exploit this via positivity of the variance and obtain


h i2 h i ⇥ ⇤
tr ⇢ ˆ  tr ⇢ ˆ2
= 4 + tr ⇢ [A2 , A1 ] ⌦ [B1 , B2 ] (6.3)
 4 + [A2 , A1 ] ⌦ [B1 , B2 ] (6.4)
 8, (6.5)

where the last inequality uses that k[A1 , A2 ]k  kA1 A2 k + kA2 A1 k  2kA1 kkA2 k = 2
and similarly for the B’s.h i
p
(6.3)-(6.5) shows that tr ⇢ ˆ  2 2 as claimed.
In order to prove that equality can be achieved, assume that ⇢ = | ih | where | i is an
eigenvector of ˆ with eigenvalue ⌫. Then equality holds in (6.3).
Now take A1 = B1 = 1 and A2 = B2 = 2 Pauli matrices, h iso that ˆ2 = 4 + 4 3 ⌦ 3
2
has eigenvalues 0 and 8. Hence, ⌫ can be chosen s.t. tr ⇢ ˆ = ⌫ 2 = 8.

Remarks.
p
• In the early 80’s, the violation of CHSH by a factor 2 has been verified ex-
perimentally. This was done using down conversion in a non-linear crystal, which
produces entangled pairs of photons, whose polarization degrees of freedom violate
CHSH.

• This is a remarkable step in the history of science, since a debate (initially mainly
between Einstein and Bohr) that was originally considered metaphysical has even-
tually been decided by an experiment.

The argumentation can be generalized to more than two observables:

25
Consider hAx By i =: Cxy , x, y 2 {1, . . . , m} as before, 2 Rm⇥m and define

X
k kLHV := sup xy ax by
a,b2{±1}m x,y

X
k kquantum := sup xy tr[⇢ Ax ⌦ By ]
⇢,{Ax ,By } x,y

where ⇢ 2 B(HA ⌦ HB ) is a density operator and  Ax , B y  .


The Cirelson bound for the CHSH inequality then reads:
✓ ◆
k kquantum p 1 1
⌫( ) := = 2 for = .
k kLHV 1 1

Theorem 33 (General Cirelson bounds).


p
(i) 2 R2⇥2 ) ⌫( )  2.

(ii) 2 Rm⇥m , m 2 N ) ⌫( )  KG < ⇡ p


2 ln(1+ 2)
⇡ 1.782.

Remark.

• (i) means that the choice of coefficients in CHSH is optimal.

• (ii) is a non-trivial statement whose proof is based on a deep result of Grothendieck.


KG is called Grothendieck’s constant, which is unknown but equal to the supremum
of ⌫( ) over all 2 Rm⇥m and all m 2 N.

26

You might also like