Bell's Theorem and the CHSH Inequality

For thirty years after Einstein, Podolsky and Rosen argued that quantum mechanics must be incomplete, the dispute looked hopelessly metaphysical. If measuring one particle of an entangled pair lets you predict its partner's outcome with certainty, EPR reasoned, the partner must have carried that answer all along — a pre-existing "element of reality." Add the assumption that a measurement here cannot instantly disturb a particle far away, and you have the local hidden variable program: outcomes are fixed in advance by some variable $\lambda$ the particles share. How could anyone ever test a claim about variables that are, by construction, hidden? John Bell's 1964 answer was to stop arguing about interpretation and start counting. He proved that any theory built on locality and realism must obey a quantitative ceiling on how strongly the outcomes of separated measurements can be correlated — and that quantum mechanics predicts correlations that break through that ceiling. The question became experimental. The Clauser–Horne–Shimony–Holt (CHSH) form, introduced in 1969, packages Bell's idea into a single number $S$ built from four correlation measurements. Every local realistic world is trapped below $|S|\le 2$ ; a quantum singlet, measured at the right angles, reaches $2\sqrt2\approx 2.828$ . This article derives both walls. We enumerate every local strategy to show the classical bound is exactly $2$ , we compute the singlet's correlation $E(\theta)=-\cos\theta$ from the Pauli algebra, and we prove the Tsirelson bound $2\sqrt2$ from an operator identity — repairing a sign that the standard textbook shorthand often gets wrong. The Bell states, the EPR argument, hidden-variable theories, contextuality, and the loophole-free experiments each own a neighbouring piece of this landscape; here we own the inequality itself.

Key takeaways

Bell's theorem proves no local hidden variable theory can reproduce all of quantum mechanics: locality plus realism imposes an inequality that entangled correlations violate.
The CHSH quantity $S=E(a,b)-E(a,b')+E(a',b)+E(a',b')$ obeys $|S|\le 2$ in every local realistic theory — provable by checking all $2^4=16$ deterministic strategies.
A quantum singlet gives $E(\theta)=-\cos\theta$ ; at detector offsets of $45^\circ$ this reaches $S=2\sqrt2\approx 2.828$ , the Tsirelson bound.
The bound follows from the operator identity $\hat S^2=4\mathbb 1+[A,A']\otimes[B,B']$ , whose commutator term has norm at most $4$ , so $\|\hat S\|\le 2\sqrt2$ .
The violation is empirical, not metaphysical: loophole-free experiments (2015) measure $|S|>2$ , ruling out local realism in nature; the 2022 Nobel honoured this line of work.
Key idea: for $\pm1$ outcomes fixed locally by $\lambda$ , the algebra $A(B-B')+A'(B+B')=\pm2$ caps $|S|$ at $2$ ; non-commuting quantum observables spread correlation across all four terms and reach $2\sqrt2$ .

The core insight

Local realism forces the four CHSH correlations to satisfy $|S|\le 2$ because the per- $\lambda$ bracket $A(B-B')+A'(B+B')$ is algebraically $\pm2$ ; quantum mechanics replaces the definite $\pm1$ values with non-commuting operators whose singlet correlations $E=-\mathbf a\cdot\mathbf b$ reach $S=2\sqrt2$ .

Where this lives today

Bell's inequality has quietly become a piece of engineering. In device-independent quantum key distribution and certified random-number generation, a measured CHSH value above $2$ certifies that genuine entanglement is present and that the outputs contain fresh, unpredictable randomness — without trusting the internal workings of the devices at all. The violation itself is the security guarantee: an eavesdropper who could predict the outcomes would, by Bell's theorem, have driven $S$ back below $2$ . Commercial and government randomness beacons now run continuous Bell tests for exactly this reason. The same number is a workhorse diagnostic for quantum hardware. Because the Werner-state relation $S=2\sqrt2\,p$ turns a measured $S$ into a lower bound on entanglement fidelity, CHSH doubles as an entanglement witness for benchmarking noisy qubits — superconducting, photonic, or trapped-ion. And Bell's theorem still shapes foundations: together with the Kochen–Specker theorem, it partitions the surviving hidden-variable programs, forcing any realistic completion of quantum theory to be explicitly non-local (as in Bohmian mechanics) or to abandon single-outcome definiteness altogether.

The physical picture

Picture two particles born together and flying to distant labs run by Alice and Bob. In a local realistic universe, each particle carries an instruction sheet $\lambda$ that already specifies how it will answer every question either experimenter might ask. Alice and Bob each pick one of two questions (analyzer settings), and each records a $\pm1$ answer. Bell's insight was that with definite instruction sheets there simply are not enough consistent ways to make all four question-pairs correlate strongly at once. The correlations are budget-limited. The CHSH combination $S=E(a,b)-E(a,b')+E(a',b)+E(a',b')$ is engineered so three terms reinforce and one fights them. With definite $\pm1$ answers, whatever instruction sheet a particle carries, exactly one of the two brackets $B\pm B'$ collapses to zero and the other becomes $\pm2$ — the bracket $A(B-B')+A'(B+B')$ is always exactly $\pm2$ . Average over any distribution of instruction sheets and $|S|$ can never exceed $2$ . That is the wall. Quantum mechanics vaults over it. Because a single spin has non-commuting components, the same particle cannot carry a definite answer to every question simultaneously; instead the singlet distributes its perfect anti-correlation smoothly across relative angles as $E(\theta)=-\cos\theta$ . Near $\theta=0$ this cosine is steeper than any straight line an instruction sheet can draw, and that extra steepness is exactly what overshoots the classical bound. The lesson is not the tired "measurement disturbs the system": it is that the joint assumption of predetermined, local outcomes is flatly incompatible with the observed statistics.

Figure

Two questions each, one shared source — the CHSH setup

The payoff figure is the one Bell taught us to read: plot the CHSH value as the measurement geometry rotates. The classical ceiling is a flat line at $2$ . The quantum curve rises above it, peaking at $2\sqrt2$ exactly where the detectors sit $45^\circ$ apart. Every loophole-free experiment since 2015 has planted its measured $S$ in that forbidden zone above the line.

Figure

CHSH value vs geometry — quantum curve breaks the classical ceiling

The core equations

The whole argument turns on a single quantity built from four correlation measurements, so we write it first and then read it. Alice and Bob each choose one of two dichotomic ( $\pm1$ ) measurements, and $E(\mathbf a,\mathbf b)$ denotes the average product of their outcomes over many entangled pairs when Alice uses setting $\mathbf a$ and Bob uses $\mathbf b$ .

S=E(\mathbf a,\mathbf b)-E(\mathbf a,\mathbf b')+E(\mathbf a',\mathbf b)+E(\mathbf a',\mathbf b') \qquad (1)

where: $E(\mathbf a,\mathbf b)=\langle AB\rangle$ is the correlation of the two $\pm1$ outcomes; $\mathbf a,\mathbf a'$ are Alice's two analyzer settings and $\mathbf b,\mathbf b'$ are Bob's. The asymmetric sign pattern — one minus, three pluses — is deliberate: it is the combination for which three terms can cooperate while the fourth is forced to interfere, which is exactly what makes the classical and quantum ceilings differ. In a local hidden variable theory every correlation must factorize through the shared variable, because Alice's outcome is not allowed to depend on Bob's distant setting. This single equation encodes both realism (definite outcomes given $\lambda$ ) and locality (each outcome depends only on the local setting).

E(\mathbf a,\mathbf b)=\int \mathrm d\lambda\,\rho(\lambda)\,A(\mathbf a,\lambda)\,B(\mathbf b,\lambda) \qquad (2)

where: $\lambda$ is the hidden variable shared by both particles; $\rho(\lambda)\ge 0$ with $\int\rho\,\mathrm d\lambda=1$ is its distribution; $A(\mathbf a,\lambda)\in\{+1,-1\}$ is Alice's outcome, depending only on her setting and $\lambda$ ; $B(\mathbf b,\lambda)\in\{+1,-1\}$ likewise for Bob. The crucial structural fact is what is absent: $A$ has no $\mathbf b$ , $B$ has no $\mathbf a$ . This factorized form is the mathematical statement of the local-realism premises that The EPR Paradox and Local Realism lays out, and the exact target that Hidden Variable Theories identifies for Bell's attack. Feed the factorized correlations into $S$ and a hard ceiling appears, independent of the form of $\rho(\lambda)$ or the dimension of $\lambda$ .

|S_{\mathrm{LHV}}|\le 2 \qquad (3)

where: $S_{\mathrm{LHV}}$ is the CHSH quantity in any local hidden variable theory. This is the Bell–CHSH inequality; a measured value above $2$ falsifies the entire class of theories at once. Quantum mechanics computes its own correlations from the state and the observables. For the spin singlet the answer is a clean dot product, which is where the extra strength comes from.

E_{\mathrm{QM}}(\mathbf a,\mathbf b)=\langle\Psi^-|(\boldsymbol\sigma\cdot\mathbf a)\otimes(\boldsymbol\sigma\cdot\mathbf b)|\Psi^-\rangle=-\,\mathbf a\cdot\mathbf b \qquad (4)

where: $|\Psi^-\rangle=\tfrac1{\sqrt2}(|{\uparrow\downarrow}\rangle-|{\downarrow\uparrow}\rangle)$ is the singlet, $\boldsymbol\sigma$ the vector of Pauli matrices, and $\mathbf a\cdot\mathbf b=\cos\theta_{ab}$ . For coplanar settings this reads $E(\theta)=-\cos\theta$ : perfect anti-correlation at $\theta=0$ , no correlation at $90^\circ$ , perfect correlation at $180^\circ$ . The perfect anti-correlation along any shared axis is the "classical-correlation ceiling" established in Quantum Entanglement and Non-Separability; here we push it to unequal angles and extract the violation. Choosing the geometry to maximize $S$ gives the quantum answer, larger than the classical $2$ but strictly short of the algebraic maximum $4$ .

S_{\mathrm{QM}}^{\max}=2\sqrt2\approx 2.828 \qquad (5)

where: $S_{\mathrm{QM}}^{\max}$ is the largest CHSH value quantum mechanics allows for any state and any observables. This is the Tsirelson bound. It sits a factor $\sqrt2$ above the classical ceiling and a factor $\sqrt2$ below the logical maximum — quantum correlations are stronger than classical but weaker than the no-signalling limit permits. A convenient one-parameter family interpolates between a pure singlet and white noise, and tells you how much noise a Bell test can survive.

\rho_W=p\,|\Psi^-\rangle\langle\Psi^-|+\frac{1-p}{4}\,\mathbb 1 \qquad (6)

where: $p\in[0,1]$ is the singlet fraction and $\mathbb 1$ is the $4\times4$ identity (the maximally mixed two-qubit state). The optimal CHSH value scales linearly, $S_{\max}(p)=2\sqrt2\,p$ , so violation requires $p>1/\sqrt2\approx0.707$ ; below that, an explicit local model reproduces every statistic.

Derivation AThe CHSH inequality from local realism

We want to show that any factorizable, $\pm1$ -valued correlation model obeys $|S|\le 2$ , with no assumption whatsoever about the form of $\rho(\lambda)$ or how many components $\lambda$ has. The strategy is to freeze the hidden variable at a single value, bound the resulting number algebraically, and only then average — because for fixed $\lambda$ everything is an ordinary product of $\pm1$ numbers, and the bound is elementary. Watch that the whole argument uses only two facts: outcomes are $\pm1$ , and they factorize.

Step 1. Abbreviate the four local outcomes at a fixed $\lambda$ as $A\equiv A(\mathbf a,\lambda)$ , $A'\equiv A(\mathbf a',\lambda)$ , $B\equiv B(\mathbf b,\lambda)$ , $B'\equiv B(\mathbf b',\lambda)$ . Each is $+1$ or $-1$ .

Step 2. Assemble the CHSH integrand for this $\lambda$ and regroup it:

AB-AB'+A'B+A'B' = A(B-B')+A'(B+B').

Why regroup this way. It isolates the two combinations $B-B'$ and $B+B'$ , whose values are tightly constrained because $B$ and $B'$ are each $\pm1$ .

Step 3. Note that since $B,B'\in\{\pm1\}$ , either they are equal (then $B-B'=0$ and $B+B'=\pm2$ ) or they differ (then $B+B'=0$ and $B-B'=\pm2$ ). Exactly one bracket vanishes and the other is $\pm2$ .

Step 4. Therefore the integrand equals either $\pm2A$ or $\pm2A'$ , and since $|A|=|A'|=1$ its magnitude is exactly $2$ :

\big|A(B-B')+A'(B+B')\big| = 2 \quad\text{for every }\lambda.

Step 5. Average over $\lambda$ with the triangle inequality and normalization $\int\rho\,\mathrm d\lambda=1$ :

|S|=\left|\int \rho(\lambda)\big[A(B-B')+A'(B+B')\big]\mathrm d\lambda\right|\le\int\rho(\lambda)\cdot 2\,\mathrm d\lambda=2.

Why this is airtight. Nothing was assumed about $\rho$ beyond being a normalized probability distribution, and nothing about $\lambda$ at all — so the bound holds for every local hidden variable theory, deterministic or stochastic. Verified: exhaustive enumeration of all $2^4=16$ deterministic $\pm1$ strategies gives values only in $\{-2,0,+2\}$ (see the verifier), so $\max|S|=2$ and the per- $\lambda$ bracket is identically $\pm2$ ; reproduces CHSH 1969.

Go deeper · Derivation B · 1 minThe singlet correlation E = minus a dot b

Derivation BThe singlet correlation E = minus a dot b

Our goal is to compute the quantum correlation $E_{\mathrm{QM}}(\mathbf a,\mathbf b)$ for the singlet directly from the Pauli algebra, showing it collapses to the simple $-\mathbf a\cdot\mathbf b$ . The clean route is to first prove the elementary two-qubit fact $\langle\Psi^-|\sigma_i\otimes\sigma_j|\Psi^-\rangle=-\delta_{ij}$ and then contract with the unit-vector components — the rotational invariance of the singlet guarantees the answer can only depend on the relative angle, which the $-\delta_{ij}$ result makes explicit.

Step 1. Work in the basis $|00\rangle,|01\rangle,|10\rangle,|11\rangle$ with $|\Psi^-\rangle=\tfrac1{\sqrt2}(|01\rangle-|10\rangle)$ . Compute the nine expectation values $\langle\Psi^-|\sigma_i\otimes\sigma_j|\Psi^-\rangle$ directly.

Step 2. The result is diagonal and negative:

\langle\Psi^-|\sigma_i\otimes\sigma_j|\Psi^-\rangle=-\delta_{ij}.

Why negative and diagonal. The singlet is the total-spin-zero state, invariant under joint rotations $U\otimes U$ ; the only invariant rank-2 tensor is $\delta_{ij}$ , and the sign is fixed by anti-correlation along any single axis ( $\langle\sigma_z\otimes\sigma_z\rangle=-1$ ).

Step 3. Expand each observable in components, $\boldsymbol\sigma\cdot\mathbf a=\sum_i a_i\sigma_i$ , and contract:

E_{\mathrm{QM}}(\mathbf a,\mathbf b)=\sum_{i,j}a_ib_j\langle\sigma_i\otimes\sigma_j\rangle=\sum_{i,j}a_ib_j(-\delta_{ij})=-\sum_i a_ib_i=-\mathbf a\cdot\mathbf b.

Step 4. For coplanar settings separated by $\theta$ , $\mathbf a\cdot\mathbf b=\cos\theta$ , so $E(\theta)=-\cos\theta$ : anti-correlated at $\theta=0$ , uncorrelated at $90^\circ$ , correlated at $180^\circ$ . Verified: the full $3\times3$ correlation matrix computes to $-\mathbb 1$ numerically, and $E_{\mathrm{QM}}(\mathbf a,\mathbf b)=-\mathbf a\cdot\mathbf b$ holds to machine precision over 200 random axis pairs (max error $3\times10^{-16}$ ).

Go deeper · Derivation C · 1 minThe Tsirelson bound from an operator identity

Derivation CThe Tsirelson bound from an operator identity

We now prove that quantum mechanics itself cannot exceed $2\sqrt2$ , for any state and any $\pm1$ observables — a bound on the theory, not just the singlet. The elegant route is to form the CHSH operator $\hat S$ , square it, and show the square is bounded by $8$ ; the key is a clean operator identity whose sign is easy to get wrong. Let $A,A',B,B'$ be Hermitian with eigenvalues $\pm1$ (so each squares to the identity), and let Alice's operators commute with Bob's (distinct subsystems).

Step 1. Define the CHSH operator, so that $S=\langle\hat S\rangle$ :

\hat S=A\otimes B-A\otimes B'+A'\otimes B+A'\otimes B'=A\otimes(B-B')+A'\otimes(B+B').

Step 2. Square it. Using $A^2=A'^2=\mathbb 1$ one finds $(B-B')^2+(B+B')^2=4\,\mathbb 1$ for the diagonal terms, while the cross terms give $A A'\otimes(B-B')(B+B')+A'A\otimes(B+B')(B-B')$ . Since $(B-B')(B+B')=[B,B']$ and $(B+B')(B-B')=-[B,B']$ , the cross terms collapse to $(AA'-A'A)\otimes[B,B']=[A,A']\otimes[B,B']$ :

\hat S^2=4\,\mathbb 1+[A,A']\otimes[B,B'].

Why the plus sign matters. A widely-copied shorthand writes this with a minus sign; that is wrong. The correct sign is $+$ , as the cross-term algebra above shows and as the verifier confirms operator-by-operator. The final bound is the same either way — the term is bounded in norm regardless of its sign — but the identity itself must be stated correctly.

Step 3. Bound the commutators. For Hermitian $M$ with $\|M\|=1$ , $\|[M,M']\|\le 2\|M\|\|M'\|=2$ , so $\|[A,A']\otimes[B,B']\|\le 2\cdot2=4$ .

Step 4. Combine: $\|\hat S^2\|\le 4+4=8$ , hence the largest eigenvalue magnitude of $\hat S$ satisfies

|S|\le\|\hat S\|\le\sqrt8=2\sqrt2.

Step 5. The optimal singlet configuration saturates this: the $45^\circ$ geometry makes $[A,A']$ and $[B,B']$ align to reach the maximal $4$ , so $\hat S$ attains eigenvalue $2\sqrt2$ . Quantum mechanics stops exactly here, short of the algebraic maximum $4$ . Verified: for the optimal observables, $\hat S^2=4\mathbb 1+[A,A']\otimes[B,B']$ holds to machine precision, the minus-sign version fails, $\|[A,A']\otimes[B,B']\|=4.000$ , and $\max|\mathrm{eig}(\hat S)|=2\sqrt2$ saturates the bound.

Limiting cases

Case	Result	Meaning
Local hidden variables	$\vert S\vert \le 2$	The Bell/CHSH bound for any local, factorizable theory — deterministic or stochastic.
Quantum singlet, optimal angles	$\vert S\vert = 2\sqrt2 \approx 2.828$	Tsirelson bound, reached with detector offsets of $45^\circ$ .
Algebraic / PR-box maximum	$\vert S\vert \le 4$	Logical ceiling if outcomes were unconstrained; no-signalling "super-quantum" Popescu–Rohrlich boxes saturate it but are not realized in nature.
Separable (product) state	$\vert S\vert \le 2$	Unentangled states never violate CHSH — entanglement is necessary for violation.
Werner state, fraction $p$	$\vert S\vert = 2\sqrt2\,p$	Violation requires $p>1/\sqrt2\approx0.707$ ; noisier mixtures admit an explicit local model.
$\theta=0$ (aligned detectors)	$E=-1$	Perfect anti-correlation; no CHSH violation from any single angle alone.

Common trapBell violation does not let you signal faster than light

The most common misreading is that beating $|S|\le 2$ transmits information superluminally. It does not. The no-signalling theorem guarantees Alice's marginal outcome distribution is completely independent of Bob's setting — locally she sees fair coin flips no matter what Bob does. Only the correlations between the two records exceed the classical bound, and extracting them requires bringing the two records together through ordinary (subluminal) classical communication. A second, subtler trap: the CHSH derivation assumes outcomes are strictly $\pm1$ . Real detectors miss particles, and a "no-click" third outcome opens the detection loophole; unless the detection efficiency is high enough (or the analysis handles non-detection correctly) a local model can fake an apparent violation. This is why loophole-free tests were such a milestone.

Worked examples

Example 1 — The classical ceiling versus the quantum peak. Compute $S$ for the canonical singlet geometry and compare it to the best any local theory can do.

Step 1. Choose coplanar settings $\mathbf a=0^\circ$ , $\mathbf a'=90^\circ$ , $\mathbf b=45^\circ$ , $\mathbf b'=135^\circ$ . The relative angles are $\theta_{ab}=45^\circ$ , $\theta_{ab'}=135^\circ$ , $\theta_{a'b}=45^\circ$ , $\theta_{a'b'}=45^\circ$ .

Step 2. Insert $E(\theta)=-\cos\theta$ with $\cos45^\circ=\tfrac1{\sqrt2}$ and $\cos135^\circ=-\tfrac1{\sqrt2}$ :

E(\mathbf a,\mathbf b)=-\tfrac1{\sqrt2},\quad E(\mathbf a,\mathbf b')=+\tfrac1{\sqrt2},\quad E(\mathbf a',\mathbf b)=-\tfrac1{\sqrt2},\quad E(\mathbf a',\mathbf b')=-\tfrac1{\sqrt2}.

Step 3. Assemble the CHSH quantity with its one minus sign:

S=\left(-\tfrac1{\sqrt2}\right)-\left(+\tfrac1{\sqrt2}\right)+\left(-\tfrac1{\sqrt2}\right)+\left(-\tfrac1{\sqrt2}\right)=-\tfrac{4}{\sqrt2}=-2\sqrt2.

Step 4. So $|S|=2\sqrt2\approx\mathbf{2.828}$ . The best any local realistic theory can reach is $|S|=2$ (Derivation A). The quantum surplus is $2\sqrt2-2\approx\mathbf{0.828}$ — a violation of the classical bound by a factor $\sqrt2$ , and the number every Bell test is built to measure.

Example 2 — How much noise can a Bell test survive? A source produces a Werner state $\rho_W=p|\Psi^-\rangle\langle\Psi^-|+\tfrac{1-p}{4}\mathbb 1$ with singlet fraction $p=0.75$ . Does it violate CHSH, and where is the breakeven point?

Step 1. Because $\langle\hat S\rangle$ is linear in the state and the maximally mixed part contributes zero (its CHSH value is $0$ ), the optimal CHSH value scales as $S_{\max}(p)=2\sqrt2\,p$ .

Step 2. Evaluate at $p=0.75$ : $S_{\max}=2\sqrt2\times0.75=1.5\sqrt2\approx\mathbf{2.121}$ . Since $2.121>2$ , this state does violate CHSH — but only by about $0.12$ , so a real experiment would need good statistics to resolve it.

Step 3. Find the breakeven fraction where $S_{\max}=2$ : solve $2\sqrt2\,p=2$ , giving $p=\tfrac1{\sqrt2}\approx\mathbf{0.707}$ . For $p$ below this the state admits an explicit local hidden variable model and cannot violate CHSH at all — the boundary between the quantum and classical worlds sits at $\approx71\%$ singlet purity.

Origins · 1 minJohn Clauser and how this was discovered

John Clauser

1942– · Pasadena and Berkeley, California

John Francis Clauser is the American experimental physicist who did as much as anyone to drag Bell's theorem out of philosophy and into the laboratory. As a graduate student he found Bell's 1964 paper largely ignored, and together with Michael Horne, Abner Shimony and Richard Holt he recast it in 1969 into the CHSH inequality — a form robust enough for real detectors, requiring only two dichotomic settings per side. In 1972, with Stuart Freedman, he performed the first experimental Bell test using photons from a calcium cascade and measured a violation of the classical bound, at a time when most physicists doubted the effort was worthwhile. His insistence that the foundations were worth testing helped launch the entire field of experimental quantum information. He shared the 2022 Nobel Prize in Physics with Alain Aspect and Anton Zeilinger for this line of work.

History

1935

— Einstein, Podolsky and Rosen argue quantum mechanics is incomplete

1964

— John Stewart Bell proves local hidden variables obey an inequality quantum mechanics violates

1966

— Bell exposes the flaw in von Neumann's 1932 no-hidden-variables proof

1969

— Clauser, Horne, Shimony and Holt publish the experimentally testable CHSH inequality

1972

— Freedman and Clauser perform the first Bell test, measuring a violation

1980

— Tsirelson proves the quantum bound is exactly 2√2

1982

— Aspect, Dalibard and Roger use rapidly switched analyzers, tightening the locality condition

2015

— Loophole-free Bell tests close the locality and detection loopholes simultaneously

2022

— Aspect, Clauser and Zeilinger share the Nobel Prize for Bell-test experiments

Going further

References. Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., §12.2 (Bell's theorem) · Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., §3.10 · Nielsen & Chuang, Quantum Computation and Quantum Information, §2.6 · J. S. Bell, Physics 1, 195 (1964) · Clauser, Horne, Shimony, Holt, Phys. Rev. Lett. 23, 880 (1969) · B. S. Cirel'son (Tsirelson), Lett. Math. Phys. 4, 93 (1980) · Aspect, Dalibard, Roger, Phys. Rev. Lett. 49, 1804 (1982).

Provenance. Equations and numbers are standard results, each recomputed in the verifier: (1) the per- $\lambda$ bracket is $\pm2$ over all 16 deterministic strategies, giving $|S|\le2$ ; (2) the singlet correlation matrix is $-\mathbb 1$ , hence $E=-\mathbf a\cdot\mathbf b$ (200 random axes, error $\sim10^{-16}$ ); (3) the canonical angles give $S=-2\sqrt2$ ; (4) the operator identity is $\hat S^2=4\mathbb 1+[A,A']\otimes[B,B']$ — the base article's minus sign was corrected to a plus here, verified operator-by-operator (the base's $-$ version fails numerically); the norm bound $\|\hat S^2\|\le 8$ and the saturation $\max|\mathrm{eig}(\hat S)|=2\sqrt2$ hold either way; (5) the Werner relation $S=2\sqrt2\,p$ with threshold $p>1/\sqrt2$ .

Builds on: The EPR Paradox and Local Realism · Hidden Variable Theories · The Bell States and Maximally Entangled Pairs · Quantum Entanglement and Non-Separability. Leads to: Experimental Tests and Violation of Bell Inequalities · The Kochen–Specker Theorem and Contextuality · The GHZ State and All-or-Nothing Nonlocality.

Is Bell's theorem about locality or about realism — which one has to go?

Bell's theorem shows the conjunction of locality and realism is untenable; it does not by itself tell you which to drop. Standard quantum mechanics keeps a form of locality (no-signalling holds) but abandons realism in the EPR sense — outcomes are not predetermined. Bohmian mechanics keeps realism (definite particle positions) but is explicitly non-local. Both reproduce the $2\sqrt2$ violation. So the theorem is a fork in the road, not a verdict for one branch; what it definitively kills is the specific combination "local and realistic," which is exactly the local hidden variable class.

Why the specific sign pattern in S — one minus and three pluses?

The pattern is chosen so that, for any assignment of definite $\pm1$ outcomes, three terms can agree while the fourth is forced to disagree. That is what caps the classical value at $2$ : you can never make all four cooperate. Quantum mechanics, using non-commuting observables, spreads correlation across all four terms more evenly and reaches $2\sqrt2$ . Any of the four terms can carry the minus (the labeling is a convention); what matters is that exactly one does, so that the two brackets $B\pm B'$ appear.

How does the singlet correlation

-\cos\theta

beat a straight line?

A local hidden variable model, integrating $\pm1$ outcomes against $\rho(\lambda)$ , can only produce correlation functions that are, in a precise sense, "no steeper than linear" in the relevant combination — that linearity is what the $|S|\le2$ bound encodes. The quantum $E(\theta)=-\cos\theta$ has slope $\sin\theta$ , which near $\theta=45^\circ$ is steeper than the best straight-line interpolation between the fixed points at $0^\circ$ and $90^\circ$ . Summing four such terms at $45^\circ$ spacing accumulates that excess steepness into the factor $\sqrt2$ .

Does entanglement alone guarantee a CHSH violation?

No. Entanglement is necessary but not sufficient for CHSH violation. Every CHSH-violating state is entangled (product states give $|S|\le2$ ), but there exist entangled states — certain Werner states with $1/3<p\le1/\sqrt2$ , and some higher-dimensional entangled states — that admit local hidden variable models for projective measurements and never violate CHSH. So a measured $S>2$ certifies entanglement, but $S\le2$ does not certify separability. Detecting the entanglement of such "local" entangled states requires other witnesses.

What is a PR box and why isn't nature one?

A Popescu–Rohrlich (PR) box is a hypothetical device that reaches the algebraic maximum $|S|=4$ while still respecting no-signalling — Alice's marginals stay independent of Bob's setting. It is perfectly consistent with relativity's ban on superluminal signalling, yet quantum mechanics only reaches $2\sqrt2$ . Why nature stops at Tsirelson's bound rather than going all the way to $4$ is a deep open question; proposed explanations invoke principles like "information causality" and "macroscopic locality." PR boxes show that no-signalling alone does not single out quantum correlations.

Is the Tsirelson bound

2\sqrt2

specific to two qubits?

The value $2\sqrt2$ is the maximum of the two-setting, two-outcome CHSH expression over all quantum states and observables in any dimension — Tsirelson proved it holds universally, not just for qubits, and it is already saturated by a single two-qubit singlet. Using more settings, more outcomes, or higher-dimensional systems defines different Bell inequalities (CGLMP, chained inequalities, etc.) with their own quantum bounds. But for the CHSH combination itself, no state anywhere in quantum theory beats $2\sqrt2$ .