A Bell experiment can produce correlations that no local shared-randomness model explains. Exhibiting one violation, however, does not show that it is the strongest quantum violation. The CHSH operator gives an upper bound valid for arbitrary finite-dimensional states and local binary measurements. Comparing that bound with a no-signaling example then separates two different restrictions on correlations.
The experiment and the classical bound
Alice and Bob receive input bits and produce signs .
Definition 1.8.1
(CHSH value and CHSH game)
Write for the correlation of the outputs given inputs . The CHSH value is
In the CHSH game the inputs are uniform and independent, and the players win when .
For a deterministic local strategy, Alice's signs depend only on her input and Bob's signs only on his. Then
Indeed, one of and is zero and the other is .
Proposition 1.8.2
(Classical CHSH bound)
For every local shared-randomness strategy, with inputs independent of the shared variable, . Under uniform CHSH-game inputs the winning probability is , so it is at most .
Proof
(Averaging deterministic strategies)
A shared-randomness strategy is a mixture of deterministic ones, and is linear in the conditional correlations, so after averaging. Communication after the inputs arrive is not part of this local model.
For the game, the conditional winning probability for inputs is . Averaging over the four equally likely inputs gives
The game statement uses uniform inputs; the correlation inequality itself is stated in terms of the four conditional correlations.
Worked example 1.8.3
(A deterministic strategy that wins three rounds in four)
Let Alice always output and Bob always output . Then for all four input pairs, so and the winning probability is . Checking directly: the players win when , which holds for and fails for , where but is required. This is an extreme case, since no local strategy can exceed .
The quantum operator bound
Let be any finite-dimensional bipartite density operator. First suppose and are Hermitian observables with , acting locally on their respective systems.
Definition 1.8.4
(Bell operator)
For such observables, the Bell operator is
For local effects , the joint probability is . Summing times these probabilities gives , where and . Hence . We first prove the bound for the stated observables and extend it to general binary POVMs in the section on binary POVMs.
Theorem 1.8.5
(Tsirelson's CHSH bound)
For any finite-dimensional bipartite state and local binary quantum measurements, including two-outcome POVMs,
With uniform CHSH-game inputs, the success probability is at most .
Proof
(Squaring the Bell operator)
We prove the bound for observables squaring to the identity; the section on binary POVMs removes that assumption. Expanding and using gives
Each observable has operator norm one, so and similarly for Bob. The tensor-product norm bound gives . Because is Hermitian, , and therefore
The trace inequality follows by diagonalizing as a convex combination of unit vectors. The game bound follows from the formula above, since .
Proof X-ray
Incompatibility appears in the commutators
The proof has one real idea: bound rather than . The identity is used to make the squared terms add up to , and the cross terms then combine into a single product of commutators. Locality enters because Alice's and Bob's operators commute across parties, as they act on different tensor factors; without it the cross terms would not factor. Within one party the two settings need not commute, and their commutators are exactly the correction to . If either pair did commute, that correction would vanish and , giving , the classical value. The ceiling arises from the largest possible commutator norm, on each side, so no qubit or particular state is assumed. The argument closes by passing from the operator norm to the expectation in .
Common mistake
(A binary POVM need not square to the identity)
The computation of uses . A two-outcome POVM with unsharp effects , , has sign observable , and . Applying the squared formula to such without further work is therefore invalid. The section on binary POVMs shows how to repair it.
Why binary POVMs are included
A two-outcome POVM on Alice's system has effects with . Its sign expectation is represented by the Hermitian contraction , so . It need not satisfy .
Lemma 1.8.6
(Dilation of a binary POVM to a squared-identity observable)
Let be a Hermitian matrix with and let . Then
is Hermitian, satisfies , and has as its upper-left block.
Proof
(Checking the block matrix)
The eigenvalues of lie in , so the finite-dimensional spectral theorem gives and is defined. Since commutes with (both are functions of ), is Hermitian and
The upper-left block is by construction.
Apply this construction separately to each setting on each side, using the same local two-dimensional ancilla for both settings. Initialize both ancillas in . Compressing to the ancilla state gives , so every correlation of the original POVMs is reproduced by squared-identity observables on the enlarged bipartite state. The preceding proof therefore applies. This is a mathematical dilation for the finite-dimensional bound; it does not assume the original devices physically used those ancillas.
Quantum correlations are also no-signaling: for fixed shared , summing Bob's effects gives , independent of . Summing Alice's effects proves the analogous statement for Bob.
Worked example 1.8.7
(Dilating an unsharp measurement)
Take with . Then and
The diagonal blocks of are , and the off-diagonal blocks are , so . In the edge case we have and the dilation is just , with no mixing; for the observable is replaced by the swap-type matrix , whose ancilla- compression is indeed .
Achievability and a noisy example
Take the Bell state and observables
Worked example 1.8.8
(Attaining the Tsirelson bound)
Each observable squares to the identity. On the Bell state, and both cross expectations vanish. Therefore and , so and the bound is attained. The corresponding game success is .
For , replace the state by .
Worked example 1.8.9
(Adding white noise)
The observables above are traceless, so the maximally mixed part contributes zero to each correlation. These measurements give and violate the classical bound when . At , they give ; at they give . This calculation concerns these specified measurements; it is not a noise-tolerance claim for every experimental device.
Common mistake
(Entangled does not mean a violation with these measurements)
For the state is entangled, because the partial transpose of has eigenvalues , so the partial transpose of has smallest eigenvalue . Yet the measurements above give , which does not violate the classical bound. Entanglement of the state is necessary for a violation, but a particular choice of measurements may not reveal it.
Counterexample mode
No signaling alone permits stronger correlations
A natural guess is that forbidding signaling already forces . Consider a conditional distribution on output bits : for each input pair , assign probability to each of the two pairs satisfying , and zero otherwise. Set , . Each individual output is uniform for every input pair, so neither marginal depends on the other party's input. Yet always, giving and game success one.
This is a mathematically valid no-signaling distribution, often called a PR box. The quantum bound proves that no finite-dimensional local quantum strategy of the stated kind realizes it. No signaling is therefore weaker than the quantum restriction; it does not by itself enforce the Tsirelson bound.
Practice with solutions
Quick check
(If , what is the success probability under uniform CHSH inputs?)
Use .
Solution · Answer
.
Quick check
(For the noisy state above with , what value of do the specified observables give?)
Use .
Solution · Answer
, below for these measurements.
Quick check
(Why does imply that has spectrum in ?)
Write in terms of alone.
Solution · Answer
Positivity of both effects gives , since . So the eigenvalues of lie in .
Quick check
(Does violation of the classical CHSH bound imply that the players can signal through their marginals?)
Compare the Bell-state strategy with the PR box.
Solution · Answer
No. Bell violation and signaling are different properties; the Bell-state strategy has no-signaling marginals, and the PR-box example has uniform marginals despite .