The CHSH Game, Played

Imagine two friends, Anaya in Kochi and Bharat in Guwahati, agreeing to play a game over the weekend. The rules are simple. A referee will call each of them independently, once per round. When the referee calls Anaya, she is asked a yes/no question — say the referee flips a fair coin to pick between question 0 and question 1, and tells Anaya which one it is. Anaya must answer with a single bit: 0 or 1. The referee calls Bharat the same way — flips a fair coin, asks him question 0 or question 1, he replies with a single bit. Anaya and Bharat cannot talk to each other during the round. They have no time to compare notes, coordinate, or signal.

The referee has a scoring rule. Let x be the bit the referee sends Anaya and y the bit sent to Bharat. Let a be Anaya's answer and b be Bharat's. The referee declares the round won if

where \oplus is XOR (1 if exactly one bit is 1, else 0) and \wedge is AND (1 if both bits are 1, else 0).

Spelled out:

• If (x, y) = (0, 0), then x \wedge y = 0, so they win if a \oplus b = 0 — their answers agree.
• If (x, y) = (0, 1), then x \wedge y = 0 — agree.
• If (x, y) = (1, 0), then x \wedge y = 0 — agree.
• If (x, y) = (1, 1), then x \wedge y = 1 — disagree.

Three cases out of four want agreement; the fourth, specifically when both questions are 1, wants disagreement. The referee picks (x, y) uniformly at random from the four possibilities, so each case occurs 25\% of the time. Question for Anaya and Bharat, before they play: how often can we win?

Anaya and Bharat are allowed to prepare in any way they want before the game begins. They can plan, share lists of random numbers, synchronise watches, share Bell pairs of particles prepared by a trusted lab in Bangalore. The only restriction is that once the game starts, they cannot communicate. Their answers depend only on the question they receive and whatever they brought with them.

This chapter does two things. First, it proves that if Anaya and Bharat only have classical resources — random bit strings, fixed lookup tables, any conspiracy short of live communication — they win at most 75\% of rounds. Second, it shows that if they share a Bell pair and measure it along a specific pair of angles, they win \cos^2(\pi/8) \approx 85.4\% of rounds. The gap is real, it is measurable, and it is the CHSH inequality violation expressed as something you can literally play on a phone with a referee app. The 2022 Physics Nobel prize was awarded, in part, for experiments that played exactly this game with entangled photons and measured the quantum advantage to many standard deviations.

The game, drawn

Start with the setup. The referee at the centre, Anaya and Bharat far apart, one 1-bit input flying to each, one 1-bit output flying back. The referee tallies the outcomes over many rounds.

Two small things to notice.

First, agreement is the dominant requirement. In three out of the four possible (x, y) pairs — specifically (0, 0), (0, 1), (1, 0) — the win condition is a \oplus b = 0, i.e. the two bits match. Only when (x, y) = (1, 1) does the game want the bits to differ. A naive strategy — Anaya and Bharat both always output 0 — wins three cases and loses one. That is 75\% right out of the gate, and we are about to prove it cannot be improved classically.

Second, the inputs are drawn uniformly. Each of the four question pairs occurs with probability 1/4. The "win probability" is the average over this uniform distribution:

If, on input (x, y), the pair (a, b) satisfies a \oplus b = x \wedge y with probability p_{xy}, then P_{\text{win}} = \tfrac{1}{4}(p_{00} + p_{01} + p_{10} + p_{11}). A strategy that wins deterministically on three cases and loses on the fourth gives p_{00} = p_{01} = p_{10} = 1, p_{11} = 0, and so P_{\text{win}} = 3/4.

The question is whether any cleverness beats this.

The classical bound — the always-output-zero strategy and why it is optimal

Anaya and Bharat's most general classical strategy is this: before the game starts, they agree on a shared random variable \lambda (any distribution; any amount of entropy; any lookup table). During the round, Anaya sees her input x and \lambda, and outputs a bit a = f_A(x, \lambda). Bharat sees y and \lambda and outputs b = f_B(y, \lambda). The functions f_A and f_B can be anything — deterministic, stochastic (via \lambda), any algorithm. The locality constraint is structural: Anaya's output does not see y; Bharat's output does not see x.

Reducing to deterministic strategies

First, reduce the problem. A stochastic strategy is a convex combination of deterministic strategies — you roll \lambda and then play deterministically given its value. Therefore the maximum win probability over stochastic strategies equals the maximum over deterministic strategies (the max of a convex combination is at most the max of its components, and is equal for the deterministic component that realises the max). So without loss of generality, \lambda is a fixed constant — we just need Anaya's function a = f_A(x) and Bharat's b = f_B(y), each a function from \{0, 1\} to \{0, 1\}.

There are 4 such functions on each side (00, 01, 10, 11 as truth tables), so 16 total deterministic strategies. Enumerate them.

One cell computed — always-0

Pick the simplest strategy: f_A(x) = 0 and f_B(y) = 0 for all inputs. Compute the win probability case by case.

• (x, y) = (0, 0): x \wedge y = 0. a \oplus b = 0 \oplus 0 = 0. Win.
• (x, y) = (0, 1): x \wedge y = 0. a \oplus b = 0. Win.
• (x, y) = (1, 0): x \wedge y = 0. a \oplus b = 0. Win.
• (x, y) = (1, 1): x \wedge y = 1. a \oplus b = 0. Loss.

Three wins, one loss. P_{\text{win}} = 3/4.

Proving no strategy exceeds 75% — the contradiction argument

Suppose, for contradiction, that some deterministic strategy (f_A, f_B) wins all four cases. Then for every (x, y) \in \{0, 1\}^2, we have f_A(x) \oplus f_B(y) = x \wedge y. Write out the four equations:

XOR all four equations together. On the left, each of f_A(0), f_A(1), f_B(0), f_B(1) appears exactly twice, and XOR of any bit with itself is 0, so the entire left side vanishes. On the right:

So 0 = 1. Contradiction.

Why this tells you the bound: perfect winning is impossible, so any deterministic strategy must lose at least one of the four cases. Losing at least one of four uniformly random cases means P_{\text{win}} \leq 3/4. Stochastic strategies are convex combinations, so they are bounded by the same 3/4.

So the classical bound is P_{\text{win}}^{\text{cl}} \leq 3/4 = 75\%, and it is tight — the always-0 strategy achieves it. Done.

Same calculation through the correlation S

The CHSH inequality chapter (ch.19) stated the bound as |S| \leq 2, where

and \langle A_x B_y\rangle is the expectation of A \cdot B when Alice uses setting x and Bob uses y, with A, B \in \{+1, -1\} (the \pm 1 encoding where +1 means output bit 0 and -1 means output bit 1). The conversion from win-probability to S is straightforward:

Why this formula: \langle A_x B_y\rangle = \Pr[a = b] - \Pr[a \neq b] = 2\Pr[a \oplus b = 0] - 1. For (x, y) \neq (1, 1), winning means a \oplus b = 0, so \Pr[\text{win} \mid x, y] = \tfrac{1 + \langle A_x B_y\rangle}{2}. For (x, y) = (1, 1), winning means a \oplus b = 1, so \Pr[\text{win} \mid 1, 1] = \tfrac{1 - \langle A_1 B_1\rangle}{2}. Averaging over the four uniform cases gives P_{\text{win}} = \tfrac{1}{4}\bigl[\tfrac{1 + \langle A_0 B_0\rangle}{2} + \tfrac{1 + \langle A_0 B_1\rangle}{2} + \tfrac{1 + \langle A_1 B_0\rangle}{2} + \tfrac{1 - \langle A_1 B_1\rangle}{2}\bigr] = \tfrac{1}{2} + \tfrac{S}{8}.

So S = 2 corresponds to P_{\text{win}} = \tfrac{1}{2} + \tfrac{2}{8} = 3/4, and the classical-bound statements "S \leq 2" and "P_{\text{win}} \leq 3/4" are identical. If you know one, you know the other.

The quantum strategy — share a Bell pair, measure at specific angles

Now change the resource. Before the game, a trusted lab in Bengaluru prepares pairs of entangled qubits in the Bell state |\Phi^+\rangle = \tfrac{1}{\sqrt 2}(|00\rangle + |11\rangle). For each round, Anaya receives one qubit of a fresh pair, Bharat receives the other. They cannot communicate during the round — but they can each do a measurement on their qubit.

Here is the quantum strategy.

• On input x = 0, Anaya measures her qubit along the Z axis of the Bloch sphere (computational basis). On input x = 1, she measures along the X axis.
• On input y = 0, Bharat measures along the axis halfway between +z and +x — an angle of \pi/4 from the Z axis. On input y = 1, he measures along the axis halfway between +z and -x — an angle of -\pi/4 from Z.

The measurement outcomes are \pm 1. The players output bit 0 for outcome +1 and bit 1 for outcome -1. Then compute how often the game wins.

The correlation on each question pair

From the CHSH chapter, the correlation of measuring |\Phi^+\rangle along directions \vec{a} and \vec{b} on the Bloch sphere is

where \theta_{ab} is the angle between the two axes. This is a standard result derived in ch.19; the key piece is that the Bell state |\Phi^+\rangle has a rotational structure that forces the correlation to depend only on the angle, not on absolute orientation.

For the strategy above, the four relevant angles are:

• \theta(A_0, B_0) = 45° — between Z and the +45° axis. Cosine: 1/\sqrt 2.
• \theta(A_0, B_1) = 45° — between Z and the -45° axis. Cosine: 1/\sqrt 2.
• \theta(A_1, B_0) = 45° — between X and the +45° axis. Cosine: 1/\sqrt 2.
• \theta(A_1, B_1) = 135° — between X and the -45° axis. Cosine: -1/\sqrt 2.

Why the fourth angle is 135°: X is at 90° from Z in the xz-plane; the -45° axis is at -45° from Z; the angle between them is 90° - (-45°) = 135°. Cosine of 135° is -1/\sqrt 2.

Plug into the win-probability formula:

Compute: \sqrt 2 / 4 \approx 0.3536. So P_{\text{win}} \approx 0.5 + 0.3536 = 0.8536, which is 85.36\%.

Why it equals \cos^2(\pi/8)

The trig identity \cos^2(\pi/8) = (1 + \cos(\pi/4))/2 = (1 + 1/\sqrt 2)/2 = \tfrac{2 + \sqrt 2}{4}. Compare to the expression above: \tfrac{1}{2} + \tfrac{\sqrt 2}{4} = \tfrac{2}{4} + \tfrac{\sqrt 2}{4} = \tfrac{2 + \sqrt 2}{4}. Same number. So

This is the Tsirelson bound for the CHSH game — the maximum win probability achievable by any quantum strategy (shown in ch.19; no state or set of measurements does better). The classical gap is 85.36\% - 75\% = 10.36 percentage points.

The circuit you can run on real hardware

The quantum strategy, as a circuit, is straightforward. Prepare |\Phi^+\rangle using the H + CNOT routine from the Bell-states chapter. Then, depending on the input, rotate the measurement basis before measuring in the computational basis.

Why the gap matters — physical, operational, foundational

The CHSH game makes the quantum advantage operational in a way that the bare CHSH inequality does not. Three different ways to appreciate the ten-point gap:

Physically. No classical experiment, no matter how elaborate, can produce a correlated bit-pair stream with agreement statistics better than 75\% on the CHSH question pattern. A shared Bell pair plus the right measurements does 85.4\%. The two regimes are operationally distinguishable: after enough rounds, a referee can tell with arbitrarily high confidence which regime the players are in.

Operationally. A CHSH violation is a certificate. Observing P_{\text{win}} > 0.75 on a source of bit-pairs implies that source is distributing genuine quantum entanglement. This observation powers device-independent quantum key distribution (DIQKD), in which the security guarantee depends only on the observed CHSH win rate, not on any trust in the quantum hardware. Think of it as the "quantumness" equivalent of a surprise inspection: no need to trust the factory; just check the behaviour.

Foundationally. The classical bound comes from XOR arithmetic on four bit-variables. The quantum bound comes from the anticommutation structure of Pauli operators and the inner-product geometry of the Bloch sphere. These are deeply different pieces of mathematics, and the fact that the latter beats the former by a specific amount (2\sqrt 2 vs. 2, or \cos^2(\pi/8) vs. 3/4) is the cleanest fingerprint of quantum theory in the physical world.

India's experimental quantum programme is in the thick of this. The Raman Research Institute (RRI) in Bangalore has been running Bell-inequality experiments on polarisation-entangled photon pairs since the 2010s, and the S.N. Bose National Centre for Basic Sciences in Kolkata has a theory group active in quantum foundations and pseudo-telepathy research. ISRO's 2022 Bengaluru–Hyderabad satellite-to-ground QKD demonstration used BB84, which does not rely on Bell violation — but the roadmap of the National Quantum Mission explicitly includes Bell-inequality-based entanglement certification as a target for future satellites.

Hype check. Winning the CHSH game more often than 75\% does not mean Anaya and Bharat are signalling faster than light. It means their measurement outcomes are correlated more tightly than any classical story can reproduce. Every CHSH game demonstration relies on classical post-processing — the referee collects Anaya's a's and Bharat's b's and compares them after the round. The no-communication theorem (see ch.18) rigorously forbids using Bell pairs to transmit information. The game is a demonstration of correlation, not causation. Pop-science headlines saying "quantum players communicate telepathically" are wrong; they are playing a cooperative game better than any classical team could, but nothing is being communicated during the round.

Common confusions

• "85\% is just a small improvement over 75\%. Why does it matter?" Scale it up. Over 10{,}000 rounds, a classical strategy wins approximately 7{,}500 times with standard deviation \sim 43; a quantum strategy wins \sim 8{,}536 times with standard deviation \sim 35. The gap is more than 20 standard deviations apart — an astronomically significant difference. Every modern loophole-free Bell test measures this gap to many standard deviations. The 10.4-point gap is not "small" — it is the tightest experimental signature of quantum advantage we have.

• "If they could communicate, they could win 100\% trivially." True — they just need to share (x, y) and output agreeing bits when required and disagreeing bits when needed. The no-communication constraint is the whole point. The CHSH game is interesting precisely because it measures how much cooperation is possible without communication, with shared resources that can depend only on a pre-established distribution (classical randomness) or a pre-shared quantum state (Bell pair).

• "Shared classical randomness gives no advantage over deterministic strategies." Correct — a stochastic strategy is a convex combination of deterministic ones, and the maximum of a convex combination of values is at most the maximum of those values. So sharing a random bit string before the game does not improve the 75\% cap. Only shared quantum entanglement breaks the classical barrier.

• "Doesn't the quantum strategy cheat by using the Bell-pair's correlation to pre-encode answers?" No. The Bell pair does not encode answers; it carries no classical information whatsoever. Each of Anaya's and Bharat's qubits, viewed locally, is maximally mixed — any local measurement outcome is uniformly random regardless of the pair state. The game-winning correlation is extracted jointly, not locally. Every attempt to treat the Bell pair as a shared lookup table fails, because there is no table to read.

• "The quantum win rate only exists on paper — real hardware has noise." Real hardware has noise, but the violation is robust. Current IBM superconducting chips prepare |\Phi^+\rangle with around 95-99\% fidelity, and measured CHSH values are typically S \approx 2.6-2.8 (translating to P_{\text{win}} \approx 0.82-0.85) — still well above the classical 0.75. The 2015 loophole-free Bell tests (Hensen, Giustina, Shalm) were tighter than this, with clean statistical confidence.

• "The \cos^2(\pi/8) formula is a magical number." It is the Tsirelson bound for CHSH, provable algebraically from two facts: measurement operators on |\Phi^+\rangle have eigenvalues \pm 1 (so A^2 = I, B^2 = I), and Alice's observables commute with Bob's (they act on disjoint qubits). The one-line calculation \hat S^2 = 4I - [A_0, A_1][B_0, B_1] gives \|\hat S\| \leq 2\sqrt 2, which translates to P_{\text{win}} \leq \tfrac{1}{2} + \tfrac{\sqrt 2}{4} = \cos^2(\pi/8). The square-root-of-2 comes from the specific algebra of Pauli operators — not magic, just arithmetic.

Going deeper

Tsirelson's bound from the game formulation

The identity P_{\text{win}} = \tfrac{1}{2} + S/8 converts a game statement into a correlation statement. Tsirelson's bound S \leq 2\sqrt 2 can be proved from three quantum-mechanical facts: Alice's and Bob's measurement operators have eigenvalues in \{-1, +1\} (so each squares to I), they commute between the two sides (they act on disjoint tensor factors), and the CHSH combination \hat S = A_0(B_0 + B_1) + A_1(B_0 - B_1) has an operator-theoretic norm constraint.

Compute \hat S^2 directly:

where \{A_0, A_1\} = A_0 A_1 + A_1 A_0 is the anticommutator and we used [A_x, B_y] = 0. Since A_0^2 = A_1^2 = I and B_0^2 = B_1^2 = I:

and

Assemble:

Both commutators satisfy \|[A_x, A_{x'}]\| \leq 2 (operator norm), so \|\hat S^2\| \leq 4 + 4 = 8, giving \|\hat S\| \leq 2\sqrt 2. The Tsirelson bound. Plug into the game formula: P_{\text{win}} \leq \tfrac{1}{2} + \tfrac{2\sqrt 2}{8} = \cos^2(\pi/8).

PR-boxes — the hypothetical 100\%-winning device

Is \cos^2(\pi/8) the end of the story? The CHSH game has a trivial upper bound of 100\%: just win every round. Popescu and Rohrlich (1994) asked whether a non-signalling physical theory could exist that attains 100\%.

A PR-box is an imagined device that, on input (x, y), outputs a uniformly random pair (a, b) \in \{0, 1\}^2 satisfying a \oplus b = x \wedge y deterministically. The marginals (the probability that a = 0 alone, ignoring b) are exactly 1/2 — no information leaks locally — so the device respects the no-signalling principle (Alice cannot tell Bharat's input from her output alone). But it achieves P_{\text{win}} = 1 — above the quantum Tsirelson bound.

PR-boxes are not physical: no known theory realises them. But they are a useful thought-experiment. They show that non-signalling alone does not pick out quantum correlations. Something stronger constrains nature at \cos^2(\pi/8). Exactly what that constraint is remains an open foundational question (candidates include information causality, macroscopic locality, and others — none definitive).

Device-independent QKD using CHSH

Ekert's 1991 protocol (E91) proposed using Bell inequalities as the security audit of a quantum key distribution scheme. The modern device-independent QKD protocol, developed by Mayers-Yao and later Pironio et al., extracts a secret key from CHSH-game rounds: if the observed CHSH win rate is above a threshold, the secret key is proven secret against any adversary, even one who built Anaya's and Bharat's devices.

The first full experimental demonstration of DIQKD was reported by Nadlinger et al. (2022) using trapped ions separated by a few metres. The rate was low (bits per second, not megabits) but the principle — certified security from CHSH violation alone — was verified. The India roadmap anticipates eventually moving to this kind of security model for high-value channels.

Experimental snapshots — what the numbers look like

For reference, here are some representative CHSH game win rates reported by landmark experiments:

• Aspect 1982 (Orsay, photon pairs): S = 2.697 \pm 0.015, giving P_{\text{win}} \approx 0.837. Above 0.75 by many standard deviations. Loopholes not closed.

• Hensen et al. 2015 (Delft, electron spins in diamond NV centres, 1.3 km separation): S = 2.42 \pm 0.20, P_{\text{win}} \approx 0.803. All three loopholes (detection, locality, freedom-of-choice) closed.

• Giustina et al. 2015 (Vienna, polarised photons) and Shalm et al. 2015 (NIST, photons): P_{\text{win}} \approx 0.79-0.82, loophole-free, with statistical significance above 10 standard deviations.

• IBM Quantum (small superconducting chips, 2020s): routine student exercises on IBM's cloud platform reproduce P_{\text{win}} \approx 0.80-0.84 on two-qubit runs. Noise reduces the gap but does not erase it.

The 2022 Physics Nobel Prize went to Aspect, Clauser, and Zeilinger for their pioneering experimental work on Bell-inequality violations; the CHSH game is the modern framing of what they measured.

Parallel repetition and XOR games

The CHSH game is a specific XOR game — the win condition is an XOR of the outputs against a function of the inputs. General XOR games have a tight theory: the classical win probability is related to a linear-programming optimum, and the quantum win probability is computable by a single semidefinite programme (Tsirelson 1987; Cleve, Høyer, Toner, Watrous 2004). This makes CHSH the simplest non-trivial member of a large, well-understood family.

If you play CHSH many times in parallel — the referee sends Anaya n independent questions at once, she answers n bits — the quantum advantage compounds. The quantum win rate remains \cos^{2n}(\pi/8) (exponentially small for large n), while the classical win rate drops faster. The gap persists. This is the basis of amplified Bell tests, used in modern certification protocols where you want extremely high statistical confidence in a single round.

Where this leads next

• Bell's theorem and CHSH — the inequality form of the same result, with its derivation and Tsirelson bound.
• Quantum pseudo-telepathy — games where quantum wins with probability 1 but classical cannot, the Mermin-Peres magic-square being the canonical example.
• Device-independent QKD — using CHSH violation as a cryptographic audit.
• No-cloning theorem — the structural cousin result that prevents a classical team from simulating the quantum strategy by copying Bell pairs.
• Bell states — the four states on which the CHSH game's quantum strategy rests.

References

1. Wikipedia, CHSH inequality — the inequality form of the game, with explicit derivations.
2. R. Cleve, P. Høyer, B. Toner, J. Watrous, Consequences and limits of nonlocal strategies (2004) — the definitive treatment of XOR games and their quantum values — arXiv:quant-ph/0404076.
3. B. Hensen et al., Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres (2015) — arXiv:1508.05949.
4. John Preskill, Lecture Notes on Quantum Computation, Ch. 4 — theory.caltech.edu/~preskill/ph229.
5. Wikipedia, Tsirelson's bound — the quantum maximum 2\sqrt 2, proof and context.
6. D. P. Nadlinger et al., Experimental quantum key distribution certified by Bell's theorem (Nature, 2022) — arXiv:2109.14600.