What Is Bell’s Theorem?

John Bell spent his working life on accelerator design at CERN. Quantum foundations was his hobby, pursued partly on sabbatical, in a field that regarded the whole subject as a career mistake. In 1964 he published four pages in a journal called Physics, which folded after four years. Bell’s theorem is those four pages. It is not a result about quantum mechanics, which never appears in the argument. It is a result about what any theory of a certain shape must obey, and it is the reason we now know the world is not that shape.

The question Bell was actually asking

By 1964 the debate about whether quantum mechanics was complete had gone quiet, and not because anyone had settled it.

John von Neumann had published a proof in 1932 that hidden variables were impossible. Everybody cited it. Almost nobody read it. Bell read it, and showed it assumed something it was not entitled to assume. Meanwhile David Bohm had done in 1952 what von Neumann said could not be done: built a working hidden-variable theory reproducing quantum predictions. Bohm’s theory had one conspicuous feature. It was flagrantly nonlocal, with distant parts of a system directly influencing one another.

That is where Bell started. Was Bohm’s nonlocality a defect of Bohm’s particular construction, something a cleverer theorist might engineer away? Or was it forced?

Bell proved the nonlocality is unavoidable. Any theory that reproduces the correlations of entangled particles while keeping distant events independent will fail. He did not set out to overturn realism. He set out to check a loose end in someone else’s theory.

The derivation, in about six lines

The modern form comes from Clauser, Horne, Shimony and Holt in 1969. It is short enough to follow.

Two particles fly apart. Alice picks setting a or a′, Bob picks b or b′, and each gets an outcome of +1 or -1. Now suppose some variable λ carries whatever the particles brought with them, distributed however you like. Write Alice’s outcome as A(a, λ) and Bob’s as B(b, λ). That notation is the entire assumption of locality: Alice’s result mentions her own setting and λ, and never Bob’s setting.

Take the quantity A(a,λ)[B(b,λ) – B(b′,λ)] + A(a′,λ)[B(b,λ) + B(b′,λ)]. Every term is +1 or -1. So one bracket is zero and the other is ±2, which means the whole expression is always +2 or -2. Average over λ and an average of numbers bounded by 2 in size cannot exceed 2. Written with correlations, that is S ≤ 2.

Quantum mechanics predicts up to 2√2, about 2.828, and experiments measure it.

Everything the theorem quietly assumes

A violation refutes a conjunction, so the assumptions are worth listing individually. There are four, plus one nobody states.

Hidden variables. Something, λ, exists and sets or weights the outcomes. Locality. This splits in two: Alice’s outcome does not depend on Bob’s setting, and Alice’s outcome does not depend on Bob’s outcome. Measurement independence. The distribution of λ does not depend on which settings get chosen. Single outcomes. Each measurement produces one result. And implicitly, no retrocausation: nothing in the future influences the past.

This list is the map of the whole interpretation landscape. Every surviving account of quantum mechanics abandons exactly one item on it. Bohmian mechanics gives up the first half of locality. Many-worlds gives up single outcomes. Superdeterminism gives up measurement independence, and retrocausal models give up the last one. The experiments tell you the list cannot all be true. They do not tell you which line to cross out.

Why “local realism” is a misleading phrase

Textbooks say Bell tests refute local realism, leaving you to drop locality or realism, and most physicists cheerfully drop realism. Bell thought that reading was wrong, and he said so repeatedly.

His argument was structural. Einstein, Podolsky and Rosen did not assume hidden variables. They derived them, from locality plus the perfect correlations. Determinism was their conclusion, not their premise. Bell then showed that local hidden variables fail. Chain the two together and locality is what breaks, whatever you think about realism. In his 1981 essay on Bertlmann’s socks he wrote that it was remarkably hard to get across that “determinism is not a presupposition of the analysis.”

Tim Maudlin, Travis Norsen and Sheldon Goldstein have pressed this line since. It is not the consensus. Many-worlds supporters answer that their picture denies single outcomes rather than realism, and claim locality survives intact on that reading. Competent people disagree here. Anyone who tells you the matter is settled is reporting a preference.

GHZ: the version with no statistics at all

Bell’s argument is statistical. You need many runs, you accumulate correlations, and you compare an average against a bound. In 1989 Daniel Greenberger, Michael Horne and Anton Zeilinger found a version that needs none of that.

Use three entangled particles instead of two. In three particular measurement configurations, quantum mechanics predicts perfect correlations, with no randomness left over. Feed those three certainties into the local hidden-variable framework and it predicts, with certainty, the outcome of a fourth configuration. Quantum mechanics predicts the opposite result, also with certainty. There is no inequality, no averaging, and no room to negotiate.

Jian-Wei Pan and colleagues in Zeilinger’s group ran it on three photons and published in Nature in 2000. The fourth measurement agreed with quantum mechanics. A single run, in principle, refutes local realism outright.

Why nature stops at 2.828

Here is a question that sounds idle and is not. The classical limit on S is 2. Quantum mechanics reaches 2.828. But the arithmetic maximum of S is 4. Why does nature stop where it does?

You might hope relativity enforces it. Sandu Popescu and Daniel Rohrlich killed that hope in 1994 by inventing a hypothetical gadget, now called a PR box, that reaches the full value of 4 and still cannot send a signal. Relativistic causality alone permits a universe far more nonlocal than ours. Ours declines the offer.

The best answer so far is information causality, proposed by Marcin Pawłowski and colleagues in Nature in 2009. Send someone m classical bits and they should gain at most m bits about your data. Any correlation stronger than the Tsirelson bound breaks that rule. The principle recovers 2.828 from an information-theoretic axiom rather than from quantum mechanics, which is elegant. It also fails to carve out the quantum set exactly, and physicists dispute its claim to be fundamental. The question remains open.

The theorem turned into a security product

In 1991 Artur Ekert saw a use for the no-go theorem. If your correlations violate a Bell inequality, no local description of your devices exists, so no eavesdropper holding a classical record of what they will do can exist either. The violation itself certifies the secrecy.

The idea goes by the name device-independence, and it takes a strange, powerful stance. You do not trust the manufacturer. You do not trust the hardware. You trust an inequality. Two groups demonstrated the full protocol in 2022. An Oxford-led team using entangled strontium ions extracted 95,628 key bits from 1.5 million Bell pairs over eight hours, though their measurements were not space-like separated, which they state plainly. A Munich-led group ran it between independently operated stations with neutral atoms.

It sits in a different category from post-quantum cryptography, which hardens classical channels with harder mathematics. Here the guarantee comes from a theorem about what the world cannot do.

People are still finding new loopholes

Three groups shut the famous loopholes together in 2015: detection, locality and freedom of choice. Coverage often treats that as the end of the story. It is not.

In 2025 Armin Tavakoli and colleagues identified a new one. Bell tests using high-dimensional entanglement need detectors that distinguish many outcomes, but experiments usually fake this with a bank of click or no-click detectors. That substitution, it turns out, opens a gap a local hidden-variable model can crawl through. A 2026 experiment closed it using four-dimensional path-mode entanglement with genuine multi-outcome detection, though that result is still a preprint.

Sixty-two years after four pages in a dying journal, people are still discovering ways the conclusion might have been wrong, and then shutting them. That is not a sign of weakness in the result. It is what taking a theorem seriously looks like.

What a Bell violation does not prove

The theorem attracts more nonsense than any other result in physics, so it is worth being blunt about the limits.

It does not permit faster-than-light signalling. Whatever Alice does, Bob’s own statistics are unchanged, and the correlation only appears once the two records meet. It does not show that consciousness affects matter. It does not prove reality does not exist, and it does not prove the universe is random. It does not prove quantum mechanics is correct, since a future theory could reproduce the same correlations and would face the same constraint. And it does not select an interpretation, which is precisely why the interpretation argument is still running.

Established, contested, unproven

Established. The derivation, which is arithmetic and not in dispute. That quantum mechanics predicts violations and that experiments find them, loophole-free, across photons, ions, atoms, diamond defects and superconducting circuits. That local hidden-variable theories are dead. That the violation cannot carry a signal.

Contested. Which assumption fails. Whether “local realism” even names the right thing. Bell’s own reading, that locality is the casualty regardless of your view on realism, is a serious minority position rather than the textbook one. Why the Tsirelson bound sits at 2.828 also remains genuinely open.

Unproven. Superdeterminism, which no experiment can exclude, since it denies the premise that makes experiments informative. Whether any information-theoretic principle uniquely singles out quantum correlations. And whether device-independent cryptography can be made practical, given that the 2022 demonstrations produced a modest number of key bits over many hours.