Bell Inequalities and the CHSH Game

Two players, Alice and Bob, sit in separate labs light-years apart. Once the game starts, they can never communicate again. A referee sends each of them a random bit. Each must instantly reply with a bit of their own. Classical physics puts a hard ceiling on their success rate. This applies to any theory built on shared randomness fixed in advance. The ceiling limits how often their answers satisfy the referee's winning condition. Quantum mechanics smashes that ceiling. It does so with a mechanism you can build in a lab. The mechanism uses two entangled particles and a handful of angled measurements. This is not a mathematical curiosity. The gap between the classical ceiling and the quantum score is measurable. It is a reproducible fact about the universe. It has been confirmed in loophole-free experiments with photons, electrons, and superconducting qubits. It is the sharpest experimental refutation of an idea championed by Einstein, Podolsky, and Rosen. That idea claims quantum mechanics is an incomplete description patched over by hidden variables. The CHSH game is named for Clauser, Horne, Shimony, and Holt. It turns a philosophical debate about the nature of reality into a two-line arithmetic inequality. Experiments can violate that inequality. The big picture is this. Consider any theory where measurement outcomes are determined, possibly probabilistically, by variables that existed before measurement. Such a theory must not allow faster-than-light influence. It must then obey a bound called a Bell inequality. Quantum entanglement violates that bound. This tells us that no local hidden-variable theory can reproduce quantum predictions. Nature is genuinely, provably non-classical at the level of correlations. This holds true, not just at the level of individual outcomes. This article derives the CHSH inequality from first principles. It computes the quantum violation exactly using the singlet state. It explains precisely why the argument is airtight. It builds directly on Quantum Entanglement Revisited. It sets up the resource-theoretic and cryptographic uses of entanglement. These uses are covered in Entanglement as a Physical Resource and Quantum Key Distribution Principles.

Key takeaways

The CHSH game tests whether measurement outcomes can be explained by pre-agreed classical strategies (local hidden variables).
Any local hidden-variable strategy wins the CHSH game with probability at most 75%, equivalently a correlation score $S \le 2$ .
Quantum mechanics, using an entangled pair and the right measurement angles, achieves $S = 2\sqrt{2} \approx 2.83$ — the Tsirelson bound.
The derivation of the classical bound uses nothing but Boolean algebra; the quantum violation uses nothing but $2\times2$ matrix expectation values.
Violation of CHSH is experimentally confirmed and rules out all local hidden-variable theories, closing the door EPR opened.
The same entangled correlations that violate CHSH are the resource behind device-independent quantum key distribution and entanglement verification.

The core insight

Quantum entanglement produces correlations between distant measurement outcomes that no classical, pre-shared randomness can reproduce — and the CHSH game measures this gap with a single number.

The idea

A Bell inequality is a constraint on the statistics of a two-party experiment that must hold if the outcomes are generated by a local hidden-variable (LHV) model: a shared random variable $\lambda$ , distributed according to some fixed probability density $\rho(\lambda)$ , that was created before the parties separated and that locally determines each party's output as a function of $\lambda$ and their own measurement choice only. "Local" means Alice's output cannot depend on Bob's measurement choice or vice versa — encoding the assumption that no signal faster than light passes between them during the experiment. The CHSH game (Clauser–Horne–Shimony–Holt, 1969) is the cleanest operational version of a Bell test. A referee draws two independent random bits $x, y \in \{0,1\}$ and sends $x$ to Alice, $y$ to Bob. Alice must respond with a bit $a \in \{0,1\}$ , Bob with $b \in \{0,1\}$ , without any further communication between them once $x,y$ are issued. They win if

a \oplus b = x \wedge y .

That is: they must agree ( $a=b$ ) unless both inputs are $1$ , in which case they must disagree. Before the game starts, Alice and Bob may agree on any strategy and may share any correlated randomness (classical or, as we will see, quantum) — the only restriction imposed during the round is no communication. The central object is the CHSH correlation function. Define $\pm1$ -valued outcomes $A = (-1)^a$ , $B = (-1)^b$ for measurement settings $x, y \in \{0,1\}$ , and let $\langle A_x B_y\rangle$ denote the expectation of the product of outcomes when Alice uses setting $x$ and Bob uses setting $y$ . The CHSH quantity is

S = \langle A_0 B_0\rangle + \langle A_0 B_1\rangle + \langle A_1 B_0\rangle - \langle A_1 B_1\rangle .

Classical (local hidden-variable) strategies obey $S \le 2$ ; quantum strategies can reach $S = 2\sqrt2$ . This gap is what the rest of the article proves.

How it works

Step 1 — the classical bound

Suppose Alice's outputs $A_0, A_1 \in \{-1,+1\}$ and Bob's outputs $B_0, B_1 \in \{-1,+1\}$ are all fixed numbers for a given value of the hidden variable $\lambda$ (deterministic strategies suffice to bound $S$ , because any probabilistic LHV strategy is a mixture of deterministic ones and the bound is linear, hence inherited by convex combinations). Consider the algebraic identity

A_0B_0 + A_0B_1 + A_1B_0 - A_1B_1 = (A_0+A_1)B_0 + (A_0 - A_1)B_1 .

Since $A_0, A_1 \in \{-1,+1\}$ , one of $(A_0+A_1)$ and $(A_0-A_1)$ is $0$ and the other is $\pm2$ . Hence the right-hand side is always $\pm2$ , so for every fixed $\lambda$ ,

A_0B_0 + A_0B_1 + A_1B_0 - A_1B_1 = \pm2 .

Averaging over the distribution of $\lambda$ (and over any local randomness Alice and Bob use) gives

S = \mathbb E[A_0B_0 + A_0B_1 + A_1B_0 - A_1B_1] \le 2 .

This is the CHSH inequality. It is a purely combinatorial fact about $\pm1$ numbers — no physics has entered yet. Physics enters only in the claim that any local, pre-agreed strategy must produce outcomes of exactly this form (a function of $\lambda$ and the local setting), which is precisely the definition of local realism. Translating $S \le 2$ into a winning probability: it is a short exercise to show $S \le 2$ is equivalent to

\Pr[\text{win}] \le \tfrac34 = 75\% .

An easy classical strategy — always output $a=b=0$ — already wins $75\%$ of the time (it fails only when $x=y=1$ , which happens with probability $1/4$ ), and no LHV strategy, however clever, can beat this.

Step 2 — the quantum strategy

Alice and Bob share one Bell pair, the singlet state

|\Psi^-\rangle = \frac{1}{\sqrt2}\bigl(|01\rangle - |10\rangle\bigr) .

On input $x$ , Alice measures her qubit in a basis rotated by angle $\theta_x$ from the $Z$ -axis in the $X$ – $Z$ plane; on input $y$ , Bob measures in a basis rotated by angle $\phi_y$ . Concretely, each measures the observable

M(\theta) = \cos\theta\, Z + \sin\theta\, X,

a Hermitian, $\pm1$ -eigenvalue operator representing "spin along direction $\theta$ ." Quantum mechanics predicts, for the singlet state,

\langle M(\theta)\otimes M(\phi)\rangle = -\cos(\theta - \phi).

This is the key physical fact (derivable directly from the singlet's rotational symmetry): the correlation between the two measurement outcomes depends only on the relative angle between the two measurement bases, and it can be as strongly correlated or anti-correlated as $\pm1$ , unlike any classical local model with the same marginal statistics. Choose angles

\theta_0 = 0,\quad \theta_1 = \pi/2,\quad \phi_0 = \pi/4,\quad \phi_1 = -\pi/4 .

Then

\langle A_0B_0\rangle = -\cos(0-\tfrac\pi4) = -\tfrac{1}{\sqrt2},\qquad \langle A_0B_1\rangle = -\cos(0+\tfrac\pi4) = -\tfrac1{\sqrt2},

\langle A_1B_0\rangle = -\cos(\tfrac\pi2-\tfrac\pi4) = -\tfrac1{\sqrt2},\qquad \langle A_1B_1\rangle = -\cos(\tfrac\pi2+\tfrac\pi4) = \tfrac1{\sqrt2}.

A sign convention absorbs into how $a,b$ are read off the $\pm1$ eigenvalues; choosing consistent output labels makes all four terms contribute with the correct sign so that

S = \left|-\tfrac1{\sqrt2}\right| + \left|-\tfrac1{\sqrt2}\right| + \left|-\tfrac1{\sqrt2}\right| + \left|\tfrac1{\sqrt2}\right| = 4\cdot\tfrac1{\sqrt2} = 2\sqrt2 .

This achieves $S = 2\sqrt2 \approx 2.828$ , strictly greater than the classical bound of $2$ . Converted to a winning probability,

\Pr[\text{win}] = \frac12 + \frac{S}{8} = \frac12+\frac{2\sqrt2}{8} = \cos^2(\pi/8) \approx 85.36\% ,

beating the classical ceiling of $75\%$ by a wide, experimentally resolvable margin.

The protocol as an algorithm

``` Setup (once, before any rounds): Prepare entangled pair |Ψ⁻⟩ = (|01⟩ - |10⟩)/√2 Distribute qubit 1 to Alice, qubit 2 to Bob Agree in advance on measurement angles: θ0 = 0, θ1 = π/2 (Alice's bases) φ0 = π/4, φ1 = -π/4 (Bob's bases) Each round: Referee draws x, y independently and uniformly from {0,1} Referee sends x to Alice, y to Bob // no further communication allowed Alice measures her qubit along M(θx) -> gets outcome a' in {+1,-1} Bob measures his qubit along M(φy) -> gets outcome b' in {+1,-1} Alice reports a = (1 - a')/2 // convert ±1 to bit Bob reports b = (1 - b')/2 Referee checks: win iff a XOR b == x AND y Repeat over N rounds, estimate empirical win rate p_win Classical bound: p_win <= 0.75 (LHV theories) Quantum optimum: p_win = cos²(π/8) ≈ 0.8536 (this strategy) If observed p_win significantly exceeds 0.75, local realism is falsified. ```

Figure

CHSH game: setup, one round, and the classical-vs-quantum score

Complexity & correctness

Computational cost. The CHSH game itself is not an algorithm solving a computational problem — it is a communication/statistics protocol — so "complexity" here means resource cost per round and the number of rounds needed for statistical confidence, not asymptotic time complexity of a Turing machine.

Quantum resources per round: one shared Bell pair (created once, or one fresh pair per round), one single-qubit measurement per party, $O(1)$ classical bits exchanged with the referee. Generating and distributing an entangled pair costs $O(1)$ physical operations independent of any problem size, since CHSH involves a fixed, small Hilbert space ( $2\times 2$ qubits).
Classical resources per round: $O(1)$ bits of shared randomness (or none) and $O(1)$ local computation to pick an output bit.
Statistical cost: to distinguish an LHV theory ( $p \le 0.75$ ) from the quantum prediction ( $p \approx 0.8536$ ) with confidence, one needs $N = O(1/\varepsilon^2)$ rounds, where $\varepsilon$ is the gap you want to resolve relative to statistical noise (by a standard Hoeffding/Chernoff concentration argument for a sum of $N$ i.i.d. $\pm1$ -bounded random variables). Since the CHSH gap $2\sqrt2 - 2 \approx 0.828$ is a constant, $O(1/\delta^2)$ rounds suffice for any fixed target failure probability $\delta$ — practically, a few thousand rounds give overwhelming statistical significance, which is why tabletop Bell tests are feasible.

Correctness of the classical bound (proof recap). The derivation in "How it works" is a complete, self-contained proof: for every deterministic assignment of $A_0, A_1, B_0, B_1 \in \{-1,+1\}$ , the identity

A_0B_0+A_0B_1+A_1B_0-A_1B_1 = (A_0+A_1)B_0 + (A_0-A_1)B_1 \in \{-2,+2\}

holds by exhaustive case check (only four sign patterns for $(A_0,A_1)$ exist, and in every case one bracket vanishes and the other is $\pm2$ , giving product $\pm2$ either way). Averaging this bounded random variable over any distribution of shared randomness $\lambda$ (which only reweights which deterministic strategy is used) cannot push the expectation outside $[-2,2]$ , because expectation is a convex combination of values already inside that interval. This is the entire proof; no physics assumption beyond "outcomes are functions of local setting and pre-shared $\lambda$ " is used.

Correctness of the quantum violation. The formula $\langle M(\theta)\otimes M(\phi)\rangle_{\Psi^-} = -\cos(\theta-\phi)$ follows from direct computation: writing $M(\theta) = \cos\theta Z + \sin\theta X$ and using the singlet's defining property that it is invariant (up to sign structure) under simultaneous rotation of both qubits, one computes

\langle\Psi^-| M(\theta)\otimes M(\phi) |\Psi^-\rangle = -\bigl(\cos\theta\cos\phi + \sin\theta\sin\phi\bigr) = -\cos(\theta-\phi),

using $\langle\Psi^-|Z\otimes Z|\Psi^-\rangle = -1$ , $\langle\Psi^-|X\otimes X|\Psi^-\rangle=-1$ , and $\langle\Psi^-|Z\otimes X|\Psi^-\rangle=\langle\Psi^-|X\otimes Z|\Psi^-\rangle=0$ , all verifiable by direct matrix computation on the four-dimensional Hilbert space. Substituting the four chosen angles into this closed-form expression and summing with the CHSH signs gives exactly $2\sqrt2$ , an exact — not approximate — quantum prediction.

Why $2\sqrt2$ and not $4$ . One might ask why entanglement doesn't achieve the algebraically maximal value $S=4$ (all four correlators equal to $\pm1$ ). This is forbidden even quantum mechanically: Tsirelson's bound states that for any quantum state and any choice of $\pm1$ -valued observables (Hermitian, unit-eigenvalue operators) on a bipartite system, $S \le 2\sqrt2$ . The proof uses the operator identity

S^2 = 4 \cdot I - [A_0,A_1]\otimes[B_0,B_1] \preceq 4\cdot I + \|[A_0,A_1]\|\,\|[B_0,B_1]\| \cdot I \le 4I + 4I = 8I,

(using $\|[A_0,A_1]\|\le 2$ for $\pm1$ -eigenvalue Hermitian operators), so $S \le \sqrt8 = 2\sqrt2$ . This shows the quantum violation found above is not merely a good strategy — it is optimal, and quantum mechanics itself imposes a strictly-more-permissive-than-classical, but still bounded, ceiling on non-local correlations. This intermediate status (more than classical, less than the algebraic maximum of "no-signaling" theories) is itself a deep structural fact about quantum theory.

Tradeoffs and caveats. The clean $2\sqrt2$ figure assumes a perfect singlet state and perfect, lossless, unbiased measurements. Real experiments must close the detection-efficiency loophole (low detector efficiency can let a local model mimic quantum statistics), the locality loophole (measurement choices and outcomes must be spacelike separated), and the freedom-of-choice loophole (the input bits $x,y$ must be generated by processes uncorrelated with $\lambda$ ). Loophole-free experiments (2015 onward, using electron spins in diamond, entangled photons over kilometers, and superconducting qubits) closed all three simultaneously, and all confirmed violation consistent with quantum predictions and no LHV model.

Worked example

Trace the protocol through several rounds with the optimal quantum angles, showing the correlator computation explicitly. Recall

\langle A_xB_y\rangle = -\cos(\theta_x - \phi_y), \qquad \theta_0=0,\ \theta_1=\tfrac\pi2,\ \phi_0=\tfrac\pi4,\ \phi_1=-\tfrac\pi4.

Setting pair $(x,y)$	Angle difference $\theta_x - \phi_y$	$\langle A_xB_y\rangle = -\cos(\theta_x-\phi_y)$	Sign used in $S$	Contribution
$(0,0)$	$0 - \pi/4 = -\pi/4$	$-\cos(\pi/4) = -1/\sqrt2$	$+$	$-1/\sqrt2$
$(0,1)$	$0-(-\pi/4)=\pi/4$	$-\cos(\pi/4)=-1/\sqrt2$	$+$	$-1/\sqrt2$
$(1,0)$	$\pi/2-\pi/4=\pi/4$	$-\cos(\pi/4)=-1/\sqrt2$	$+$	$-1/\sqrt2$
$(1,1)$	$\pi/2-(-\pi/4)=3\pi/4$	$-\cos(3\pi/4)=+1/\sqrt2$	$-$	$-1/\sqrt2$

Summing the contribution column with CHSH's prescribed signs $(+,+,+,-)$ :

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

The overall sign is a labeling artifact (which outcome we call " $+1$ "); flipping the sign convention for one party's outputs converts $\langle A_xB_y\rangle \to -\langle A_xB_y\rangle$ uniformly and yields the standard $S=+2\sqrt2$ . What matters physically is $|S| = 2\sqrt2$ , unambiguously exceeding the classical bound of $2$ . Now trace individual rounds at the level of bits, using the correlator to get per-setting outcome probabilities. For setting pair $(x,y)=(1,1)$ , correlator $+1/\sqrt2 \approx 0.707$ means Alice and Bob's $\pm1$ outcomes agree with probability

p(\text{agree}) = \frac{1+\langle A_1B_1\rangle}{2} = \frac{1+0.707}{2} \approx 0.854,

i.e., $a=b$ about $85.4\%$ of the time — but the game requires $a\ne b$ exactly when $x=y=1$ . So on this particular setting the quantum strategy "loses" $85.4\%$ of the time locally — yet this is precisely the correct trade: sacrificing this one setting-quadrant lets the other three succeed at rate $85.4\%$ each, and the overall weighted win rate (each pair equally likely) is

p_{\text{win}} = \frac14\Bigl[p_{00}(\text{agree}) + p_{01}(\text{agree}) + p_{10}(\text{agree}) + p_{11}(\text{disagree})\Bigr] = \frac14\bigl[0.854\times3 + 0.854\bigr] = 0.854,

matching $\cos^2(\pi/8)$ computed earlier. Compare a hypothetical LHV strategy that tries to mimic this by fixing $A_0=A_1=+1$ , $B_0=+1,B_1=-1$ with probability $1/2$ and the flipped assignment otherwise — direct enumeration shows the best such classical mixture still tops out at exactly $75\%$ , a full $10.4$ percentage points below the quantum strategy, illustrating that the CHSH gap is not a rounding artifact but a robust, exploitable statistical signature.

Common pitfalls & traps

Pitfall	Why it happens	Fix
Confusing "no communication" with "no shared randomness"	LHV models are allowed unlimited pre-shared randomness $\lambda$ ; the bound $S\le2$ already accounts for this	Remember: the classical bound is proven for all distributions over deterministic strategies, not just fixed strategies
Assuming $S=4$ is achievable quantum-mechanically	Extrapolating "quantum beats classical" without checking the algebra	Apply Tsirelson's bound; $2\sqrt2$ is the true quantum ceiling, not $4$
Treating a single round's outcome as proof of violation	$S$ is a statistical expectation; single trials are just $\pm1$	Violation is only meaningful after averaging over many rounds and computing confidence intervals
Ignoring detector inefficiency in real experiments	Low-efficiency detectors let "no-detection" events be discarded, opening a loophole for local models	Use fair-sampling assumptions explicitly, or run loophole-free protocols with high detection efficiency
Misplacing the "no-signaling" boundary	Believing violation of CHSH implies faster-than-light communication	CHSH violation requires no communication whatsoever during the run; correlations, not causal signals, are what's non-classical
Picking arbitrary measurement angles and expecting maximal violation	The $2\sqrt2$ bound requires the specific $45°$ -offset angle choice; generic angles give less violation	Always verify the angle choice against $\langle A_xB_y\rangle=-\cos(\theta_x-\phi_y)$ before claiming optimality

Common trapThe single most dangerous trap: mistaking correlation for communication

CHSH violation shows that Alice and Bob's outcomes are correlated in a way no pre-agreed classical strategy can match — it does not mean Alice's measurement choice instantaneously changes Bob's marginal outcome statistics. The no-signaling theorem guarantees Bob's outcome distribution is exactly the same (uniformly random, $50/50$ ) regardless of what Alice measures or even whether she measures at all. It is only the joint statistics, compared after the fact by bringing results together, that show the non-classical correlation. Treating CHSH violation as evidence for superluminal signaling is the most common — and most damaging — misinterpretation of Bell's theorem, and it directly contradicts relativity and the mathematics of quantum channels alike.

Pioneers & history

Origins · 1 minJohn Stewart Bell and how this was discovered

John Stewart Bell

1928–1990 · Theoretical physicist, CERN

Bell trained as a particle physicist but is remembered above all for a 1964 paper written almost as a side project, addressing the EPR paradox directly. Motivated by David Bohm's explicit hidden-variable reformulation of quantum mechanics, Bell asked whether any local hidden-variable theory — not just Bohm's — could reproduce all quantum predictions. His answer, expressed as an inequality obeyed by all such theories but violated by quantum entangled states, transformed a philosophical dispute into an experimentally decidable question. Bell's original inequality applied to idealized perfectly-correlated settings; Clauser, Horne, Shimony, and Holt generalized it in 1969 into the experimentally practical CHSH form used throughout this article, permitting realistic, imperfectly-correlated measurement settings and detector inefficiencies.

History

1935

Einstein, Podolsky, and Rosen argue quantum mechanics must be incomplete, introducing the entangled-pair thought experiment later formalized as the EPR paradox.

1964

John Bell derives the first Bell inequality, proving local hidden-variable theories make testable predictions distinct from quantum mechanics.

1969

Clauser, Horne, Shimony, and Holt publish the CHSH inequality, making Bell tests experimentally feasible with real, imperfect detectors.

1972–1982

Clauser and Freedman, then Alain Aspect's group, perform increasingly rigorous Bell tests with entangled photons, observing violations consistent with quantum predictions.

2015

Independent groups (Delft, NIST, Vienna) report loophole-free Bell tests closing detection, locality, and freedom-of-choice loopholes simultaneously.

Where this matters

Bell inequality violation is not confined to foundational physics seminars — it underwrites entire technology stacks built on trusting entanglement as a genuine physical resource. Device-independent quantum key distribution protocols use CHSH-style violations to certify security without trusting the internal workings of the devices — if an eavesdropper's intervention could be modeled as a local hidden variable, the observed CHSH score would drop toward $2$ , so a measured score near $2\sqrt2$ certifies genuine, un-tampered entanglement. This idea is developed further in Quantum Key Distribution Principles.

Random number generation exploits the same fact: because no LHV model can reproduce quantum correlations, a device that reliably violates CHSH is provably producing outcomes that cannot have been predetermined, giving "Bell certified randomness" used in some cryptographic and lottery infrastructure. Entanglement verification and benchmarking in quantum computing hardware (superconducting qubits, trapped ions, photonic chips) routinely runs CHSH-style tests as a certification step — a quantum processor's two-qubit gates are validated in part by demonstrating they can prepare states violating Bell inequalities close to the Tsirelson bound, connecting directly to the resource-theoretic view developed in Entanglement as a Physical Resource. Beyond cryptography, Bell tests inform the foundations of quantum sensing and metrology, where entangled probes outperform classical ones precisely because of the non-classical correlations Bell inequalities quantify, and they remain the gold-standard experimental benchmark separating genuinely quantum information processing from classical simulation with shared randomness.

Ask the tutor: How does the CHSH game's security guarantee actually get used inside a device-independent quantum key distribution protocol?

Does violating the CHSH inequality mean information travels faster than light?

No. The no-signaling theorem, provable directly from quantum mechanics, guarantees that Bob's local measurement statistics are completely independent of Alice's choice of measurement setting — his marginal outcome distribution is always uniformly random regardless of what she does. What Bell violation shows is that the joint statistics, compared only after both results are brought together classically, cannot be explained by any pre-shared classical randomness. The correlation is real and non-classical, but no signal, message, or causal influence propagates between the parties during the experiment. Confusing joint correlation with communication is the most common misunderstanding of Bell's theorem, and relativity remains fully intact.

Why is 2√2 the quantum limit rather than 4?

The algebraic maximum of $S$ , if outcomes could be arbitrary $\pm1$ values with no constraint, is $4$ . Quantum mechanics restricts measurement outcomes to eigenvalues of Hermitian operators acting on a fixed-dimensional Hilbert space, and Tsirelson's bound (an operator inequality derivable from the commutator structure of $\pm1$ -eigenvalue observables) proves $S\le2\sqrt2$ for any quantum state and any such observables. This shows quantum correlations are a genuinely intermediate class: strictly stronger than anything classical local hidden variables can produce, yet strictly weaker than the abstract "no-signaling" correlations allowed by relativity alone. That gap between $2\sqrt2$ and $4$ is itself an active research area in quantum foundations.

Can CHSH be violated with more than two measurement settings per party?

Yes — CHSH is the simplest case (two settings, two outcomes per party), and it is the tightest Bell inequality in that minimal scenario. Generalizations exist with more settings (e.g., the CGLMP inequality) or more outcomes, and some multi-setting or multi-party inequalities (like the Mermin inequalities for GHZ states) show violations that grow with the number of parties, sometimes exponentially. These generalized inequalities are used in multipartite entanglement certification and in some device-independent protocols, but CHSH remains the canonical, pedagogically simplest, and experimentally most common test.

What role does the choice of measurement angle play in the quantum violation?

Everything. The correlator formula $\langle A_xB_y\rangle=-\cos(\theta_x-\phi_y)$ shows the violation depends entirely on the relative angles between Alice's and Bob's measurement bases. The specific choice $\theta_0=0,\theta_1=\pi/2,\phi_0=\pi/4,\phi_1=-\pi/4$ is not arbitrary — it is the angle configuration that maximizes $S$ , achieving exactly $2\sqrt2$ . Suboptimal angle choices (e.g., aligned or nearly-aligned bases) give smaller or even zero violation, so an experimenter must carefully calibrate measurement bases; a common experimental error is a angle-calibration drift that silently reduces the observed violation below the theoretical maximum.

How many trials are needed to statistically confirm a CHSH violation?

Since each round's contribution to $S$ is a bounded random variable (each term is a product of two ±1 outcomes), concentration inequalities like Hoeffding's bound guarantee the empirical average converges to the true expectation at rate $O(1/\sqrt N)$ for $N$ rounds. To resolve the gap between the classical bound (2) and the quantum prediction ( $2\sqrt2\approx2.83$ ), which is a constant-sized gap of about $0.83$ , only a modest number of rounds — typically a few thousand to tens of thousands depending on detector noise — are needed to achieve statistical significance well beyond five-sigma confidence, which is why tabletop and even classroom-scale Bell tests are feasible today.

What loopholes must a real experiment close to make the result convincing?

Three classic loopholes threaten Bell test conclusions. The locality loophole requires that Alice's and Bob's measurement events be spacelike separated so no signal could travel between them in time to influence the outcome. The detection loophole arises when imperfect detectors let low-efficiency post-selection mimic quantum correlations using local hidden variables (fair-sampling assumptions must be justified or avoided). The freedom-of-choice loophole requires that the random settings $x,y$ be generated by a process statistically independent of whatever hidden variable might exist. Experiments since 2015 have closed all three simultaneously, giving the strongest available evidence against local realism.

How does the CHSH game connect to quantum key distribution?

Device-independent QKD protocols use a CHSH-style test as a built-in security check: an eavesdropper who tries to gain information about the shared key by intercepting or cloning quantum states necessarily disturbs the entangled correlations, which manifests as a reduced CHSH score closer to the classical bound of $2$ . By continuously monitoring the CHSH violation during key exchange, the parties can certify security without trusting the internal design or manufacture of their quantum devices — a property impossible in purely classical cryptography and central to the device-independent security proofs covered under Quantum Key Distribution Principles.