Randomness from Nonlocality

If a Bell test certifies that your boxes contain genuine entanglement, it also certifies something operationally priceless: their outputs are unpredictable, even to an adversary who built the hardware. This lesson explains why a Bell violation lower-bounds the randomness of the outcomes, device-independently.

Why a violation implies randomness

Suppose, toward contradiction, that Alice's outcome $a$ were deterministic given the inputs — fixed in advance, perhaps by a hidden variable the manufacturer pre-loaded. Then the model is a local deterministic one, and we proved such models obey $S\le 2$ . So observing $S>2$ rules out determinism: the outcomes cannot be a pre-set function of the settings. Some genuine randomness must be generated at measurement time. The larger the violation, the more randomness is forced.

Quantifying it: the guessing probability

The right measure is the min-entropy. Let $P_{\text{guess}}(a\mid x)$ be the maximum probability with which an adversary Eve — who may hold a quantum system correlated with the boxes — predicts the outcome $a$ for a fixed input $x$ . The certified randomness is

H_{\min}(a\mid x) = -\log_2 P_{\text{guess}}(a\mid x).

The central DI result bounds $P_{\text{guess}}$ using only the observed CHSH value $S$ . For the CHSH scenario,

P_{\text{guess}} \le \frac{1}{2}\left(1 + \sqrt{2 - \frac{S^2}{4}}\right),

a function that decreases monotonically as $S$ grows from $2$ to $2\sqrt2$ .

The two extremes

At the classical boundary $S=2$ , the bound gives $P_{\text{guess}}\le 1$ : no randomness is certified, consistent with a deterministic local model. At maximal violation $S=2\sqrt2$ , we have $2 - S^2/4 = 2 - 2 = 0$ , so

P_{\text{guess}} \le \frac12, \qquad H_{\min}(a\mid x) \ge 1\ \text{bit}.

A single maximally-nonlocal round certifies a full bit of fresh, private randomness in Alice's outcome — the strongest possible for one binary output. Eve, despite having manufactured the devices and holding any correlated quantum side information, can do no better than a coin flip.

From one round to a protocol

A randomness-expansion protocol interleaves many rounds, uses a few random rounds to estimate $S$ , and feeds the rest through a classical randomness extractor to distil near-uniform bits. Security proofs against quantum adversaries accumulate the per-round min-entropy via the entropy accumulation theorem, yielding a net output rate set by the observed violation. This is how the 2010 experiment of Pironio et al. produced the first numbers "certified by Bell's theorem."

The takeaway

A Bell violation is incompatible with determinism, so it forces randomness. Quantitatively, $P_{\text{guess}}$ is bounded by a decreasing function of $S$ , hitting $1/2$ — a full certified bit — at $S=2\sqrt2$ . Randomness becomes a measurable consequence of nonlocality, secure against an adversary who controls the devices.