Tsirelson's Bound

Tsirelson's bound stopped being a foundational curiosity the moment people realised it could certify hardware. The landmark loophole-free Bell tests of 2015 — the Delft nitrogen-vacancy experiment of Hensen and colleagues, and the photonic experiments of the Vienna and NIST groups — all report CHSH values between $2$ and $2\sqrt2$ (Delft measured $S\approx 2.42$), closing the detection and locality loopholes at once and confirming that nature respects the quantum ceiling while breaking the classical one. Because a Bell violation rules out local hidden-variable explanations under the protocol assumptions, its magnitude can be used to bound an adversary's information and certify device-independent randomness or secret-key rates. Device-independent quantum key distribution was demonstrated experimentally in 2022 with trapped ions and separately with distant trapped atoms, with the security analysis relying on the observed Bell violation together with the protocol's statistical and physical assumptions. The same rigidity powers self-testing, where nearly saturating $2\sqrt2$ certifies that a black box really contains a Bell pair and anticommuting measurements — the basis for verifying untrusted quantum computers. And the gap between $2\sqrt{2}$ and $4$ remains a fundamental distinction between quantum and general nonsignalling correlations. The 2020 proof that $\mathrm{MIP}^*=\mathrm{RE}$ gave a negative answer to Tsirelson's problem by showing that the closure of the tensor-product quantum correlation set is strictly smaller than the quantum commuting-operator correlation set.

Picture CHSH as a cooperative game. A referee hands Alice a random bit $x\in\{0,1\}$ and Bob a random bit $y\in\{0,1\}$; each returns a $\pm1$ answer, scored as $S=\langle A_0B_0\rangle+\langle A_0B_1\rangle+\langle A_1B_0\rangle-\langle A_1B_1\rangle$. The lone minus sign is the whole drama: three terms want positive correlation, one wants negative, and no deterministic assignment can satisfy all four wishes at once. Classical players sharing any pre-agreed randomness top out at $S=2$; players sharing a Bell pair do strictly better — but the same non-commutativity that lets them win also caps their winnings. The first figure lays out the three regimes on one axis.

Why does the quantum ceiling fall at $2\sqrt{2}$ and not at $4$? The engine of the violation is that Alice's two observables need not commute, and likewise for Bob. In a Bell experiment, $A_0$ and $A_1$ are alternative measurement settings rather than sequential measurements, so the violation should not be interpreted as one measurement physically disturbing the other. But that non-commutativity is also the governor: squaring the Bell operator reduces every term to a product of commutators strictly limited in size for $\pm1$ observables. The geometry that extracts the maximum is clean: Alice measures along two orthogonal Bloch axes, Bob along their two bisectors, so each of the four correlators equals $\cos 45^\circ = 1/\sqrt{2}$ and they add coherently to $4/\sqrt{2}=2\sqrt{2}$. The second figure shows that optimal configuration on the Bloch circle.

The gap between $2\sqrt{2}$ and $4$ is the space of theories that are "too nonlocal" — perfectly causal, yet unreachable by any quantum state. That such theories are logically consistent but physically absent is one of the deepest hints that quantum mechanics has special internal structure beyond mere no-signalling, and it is why so much effort has gone into deriving $2\sqrt{2}$ from an information-theoretic axiom rather than the formalism itself.

The starting object is the Bell operator: a single Hermitian operator whose expectation value is the CHSH combination, built from four dichotomic observables, two per party, each squaring to the identity.

where: $A_0,A_1$ — Alice's two dichotomic observables, $A_i=A_i^\dagger$, $A_i^2=\mathbb{1}$, eigenvalues $\pm1$; $B_0,B_1$ — Bob's two dichotomic observables, $B_j=B_j^\dagger$, $B_j^2=\mathbb{1}$; $\otimes$ — tensor product encoding spatial separation, so every $A_i$ commutes with every $B_j$. Reading it: $\langle\mathcal{B}\rangle$ scores the game, and the single minus sign forbids any deterministic strategy from winning all four terms. Tsirelson's bound states that this score, over all quantum states and valid observables, never exceeds $2\sqrt{2}$ — a ceiling on an operator's expectation value, not a property of any one experiment.

where: $|\psi\rangle$ — any bipartite quantum state (pure here; mixed states follow by convexity); $S$ — the CHSH expectation value; $2\sqrt{2}\approx 2.8284$ — the maximal quantum value. Reading it: a value above $2$ rules out local hidden-variable models; a value above $2\sqrt{2}$ is impossible for quantum theory. The whole proof pivots on one identity. Squaring the Bell operator makes the four diagonal contributions add to $4\,\mathbb{1}$, while the cross terms combine into a single product of commutators.

where: $[A_0,A_1]=A_0A_1-A_1A_0$ — the commutator on Alice's side; $[B_0,B_1]$ — the commutator on Bob's side. Reading it: if either pair of observables commutes, the commutator correction vanishes and $\|\mathcal{B}\|=2$ — the quantum value cannot exceed the classical CHSH bound. Nonlocality is enabled by the noncommutativity of the two observables on each side. Each of those commutators is small. For observables bounded by $1$ in norm, the triangle inequality caps a commutator at twice the product of the norms.

where: $\|\cdot\|$ — the operator (spectral) norm, the largest singular value. Reading it: each product $A_0A_1$, $A_1A_0$ has norm $1$, so their difference has norm at most $2$, saturated exactly when $A_0$ and $A_1$ anticommute. Chaining the identity and the commutator bound caps the norm of $\mathcal{B}^2$, and taking the square root of this positive Hermitian operator caps $\mathcal{B}$ itself.

where: $\|\mathcal{B}\|^2=\|\mathcal{B}^2\|$ — valid because $\mathcal{B}$ is Hermitian; the tensor-product norm factorises, $\|M\otimes N\|=\|M\|\,\|N\|$. Reading it: the top eigenvalue of $\mathcal{B}$ is at most $\sqrt{8}=2\sqrt{2}$, and every expectation value obeys $\langle\psi|\mathcal{B}|\psi\rangle\le\|\mathcal{B}\|$. Finally, the configuration that saturates the bound is explicit: a maximally entangled Bell pair, Alice measuring orthogonal axes, Bob their bisectors.

where: $|\Phi^+\rangle$ — the maximally entangled two-qubit state; $Z,X$ — Pauli observables; the $45^\circ$ offset of Bob's axes is what turns each correlator into $\cos 45^\circ=1/\sqrt{2}$. Reading it: substitute into (1) and the score is $4/\sqrt{2}=2\sqrt{2}$ exactly, so the bound is attained and cannot be improved.

The CHSH value $S=\langle\mathcal{B}\rangle$ lives inside nested ceilings that depend entirely on which theory generates the correlations.

1 · The optimal Bell-state configuration (fully worked)

A source emits the maximally entangled state $|\Phi^+\rangle=\tfrac{1}{\sqrt2}(|00\rangle+|11\rangle)$. Alice measures $A_0=Z$ or $A_1=X$; Bob measures $B_0=\tfrac{Z+X}{\sqrt2}$ or $B_1=\tfrac{Z-X}{\sqrt2}$. Compute the CHSH value $S$ and confirm it saturates Tsirelson's bound.

Step 1. Record the four Pauli correlators on $|\Phi^+\rangle$: $\langle Z\otimes Z\rangle=+1$, $\langle X\otimes X\rangle=+1$, $\langle Z\otimes X\rangle=\langle X\otimes Z\rangle=0$. These are the only inputs needed.

Step 2. Expand each CHSH correlator by linearity: $\langle A_0B_0\rangle=\tfrac{1}{\sqrt2}(\langle ZZ\rangle+\langle ZX\rangle)=\mathbf{+1/\sqrt2}$, and identically $\langle A_0B_1\rangle=\langle A_1B_0\rangle=\mathbf{+1/\sqrt2}$, while $\langle A_1B_1\rangle=\tfrac{1}{\sqrt2}(\langle XZ\rangle-\langle XX\rangle)=\mathbf{-1/\sqrt2}$.

Step 3. Assemble with the CHSH signs: $S=\tfrac{1}{\sqrt2}+\tfrac{1}{\sqrt2}+\tfrac{1}{\sqrt2}-\big(-\tfrac{1}{\sqrt2}\big)=\tfrac{4}{\sqrt2}=\mathbf{2\sqrt2\approx 2.828}$. This equals Tsirelson's bound exactly, so it is an optimal CHSH strategy — and by self-testing, any configuration reaching $2\sqrt2$ is equivalent to this one up to a local change of basis.

2 · How much noise a Bell pair can survive (Werner state)

Model an imperfect source as a Werner state $\rho=V\,|\Phi^+\rangle\langle\Phi^+|+(1-V)\tfrac{\mathbb{1}}{4}$, visibility $V\in[0,1]$. Using the optimal measurements of Example 1, find (a) the CHSH value versus $V$, (b) the minimum $V$ to violate the classical bound, and (c) the minimum $V$ to violate a local model at all.

Step 1. The maximally mixed part $\tfrac{\mathbb{1}}{4}$ gives zero correlation for traceless observables: $\mathrm{Tr}\big[\tfrac{\mathbb{1}}{4}\,A_i\otimes B_j\big]=\tfrac14\mathrm{Tr}(A_i)\mathrm{Tr}(B_j)=0$ since Pauli observables are traceless.

Step 2. So only the coherent part survives, scaled by $V$: every correlator is $V$ times its Bell-state value, hence $S(V)=\mathbf{2\sqrt2\,V}$ — visibility multiplies the CHSH value directly.

Step 3. (b) Violation needs $S>2$, i.e. $2\sqrt2\,V>2\Rightarrow V>\tfrac{1}{\sqrt2}=\mathbf{0.7071}$. Below $71\%$ the optimal strategy no longer beats $2$.

Step 4. (c) CHSH is the strongest two-setting test, so a violation with these settings needs the same $V>1/\sqrt2$ — sufficient for entanglement but not necessary: Werner states are entangled for $V>1/3$, leaving a window $1/3<V\le 1/\sqrt2$ that is entangled yet admits a local model here. Endpoints check: $V=1\Rightarrow S=2\sqrt2$, $V=1/\sqrt2\Rightarrow S=2$, $V=0\Rightarrow S=0$. A CHSH test needs visibility above about 71 percent to certify nonlocality.

3 · Reading a real experiment

A loophole-free Bell test reports $S=2.42\pm 0.20$. What can and cannot be concluded?

Step 1. Against the classical bound, $2.42>2$ by more than one standard deviation, so the data are inconsistent with any local model — nonlocality, and hence entanglement, is certified without state tomography.

Step 2. Against Tsirelson's bound, $2.42 < 2\sqrt{2}\approx2.828$, comfortably inside the quantum region. It does not saturate the bound, so the source is not the ideal self-testing configuration. The reduced value can arise from imperfect state preparation, measurement imperfections, or residual experimental noise; in a loophole-free experiment, detector inefficiency is accounted for in the analysis rather than being treated as an open detection loophole.

Step 3. The equivalent visibility is $V\approx S/(2\sqrt2)=2.42/2.828\approx\mathbf{0.856}$, an $86\%$ Bell pair. A value between 2 and 2√2 certifies nonlocality and entanglement and quantifies randomness, but never implies super-quantum physics.

Tsirelson's bound names a physical ceiling after the mathematician who found it — not an experimentalist chasing data, but a functional analyst asking how large a certain operator norm could be.

Texts. Nielsen and Chuang, Quantum Computation and Quantum Information, section 2.6 · Preskill, Lecture Notes on Quantum Computation, chapter 4 · Watrous, The Theory of Quantum Information, chapter on nonlocal games · Wilde, Quantum Information Theory, 2nd ed. · Mermin, Quantum Computer Science. The key checkpoints of the argument are: (1) the identity $\mathcal{B}^2=4\,\mathbb{1}-[A_0,A_1]\otimes[B_0,B_1]$; (2) the norm chain $\|\mathcal{B}^2\|\le 8\Rightarrow\|\mathcal{B}\|\le 2\sqrt2$; (3) substitution of $A_0=Z,A_1=X,B_{0,1}=(Z\pm X)/\sqrt2$ on $|\Phi^+\rangle$ giving $S=2\sqrt2\approx 2.8284$; (4) the local bound $|S|\le 2$ from the sixteen $\pm1$ assignments; (5) the Werner scaling $S=2\sqrt2\,V$ with threshold $V=1/\sqrt2$.

Prerequisites. The CHSH Inequality and Its Quantum Violation · Bell's Theorem and Local Hidden Variables · Bell States and EPR Pairs · Nonlocality and the No-Signalling Principle. Leads to. PR Boxes and Super-Quantum Correlations · Separable versus Entangled States · Entanglement Witnesses · Entanglement Entropy.

Ask the tutor: Ask me to walk through squaring the Bell operator line by line, or to show exactly how a noisy Werner state loses its CHSH violation as the visibility drops toward $1/\sqrt2$.