Bell Test Frontiers

Bell experiments have established violations of Bell inequalities under carefully controlled conditions. The useful frontier is no longer a generic request to “test quantum mechanics”: it concerns stronger causal control, weaker auxiliary assumptions, network structures, and finite-data certification. Reviewed 11 October 2026, this page uses the cited 2014–2023 literature as its explicit baseline rather than attempting a complete 2026 experiment census.

Helpful background. Bell experiments and Bell loopholes explain experimental conditions; CHSH gives the standard inequality.

Bell inference is a conditional exclusion

A local hidden-variable model with measurement independence has

$$p(a,b\mid x,y)=\int d\lambda\,q(\lambda) p(a\mid x,\lambda)p(b\mid y,\lambda).$$

The setting-independent distribution $q(\lambda)$ is a substantive premise. Local response functions by themselves do not imply observable no-signaling if the distribution of hidden variables depends on the chosen settings.

For binary outcomes the CHSH bound is $|S|\le2$. An experimental violation excludes the tested model class subject to the event definition, setting-generation assumptions, spacetime arrangement, and statistical procedure. It does not establish a unique ontology or operational faster-than-light communication.

Hensen and colleagues’ 2015 event-ready experiment is a concrete primary example of addressing locality and detection together. Brunner and colleagues’ review supplies the mathematical landscape. These are established reference points; the open questions below concern more demanding inference and applications.

A sequential test that tolerates device memory

A concrete inferential baseline is more useful than reporting how many standard deviations $S$ exceeds two. Let $W_i$ be the indicator of winning the CHSH game at trial $i$. With fresh uniform settings independent of the devices and a valid trial definition, every local model obeys $\mathbb E[W_i\mid\mathcal F_{i-1}]\le\beta,\qquad\beta=\frac34,$ where $\mathcal F_{i-1}$ contains the complete past. The device may change its strategy using that past.

Choose $q\in(\beta,1)$ before analyzing the trials and define $M_n=\prod_{i=1}^{n} \left(\frac q\beta\right)^{W_i} \left(\frac{1-q}{1-\beta}\right)^{1-W_i}.$ If the conditional winning probability is $p_i\le\beta$, the expected next factor is $p_i\frac q\beta+(1-p_i)\frac{1-q}{1-\beta}\le1.$ Thus $M_n$ is a nonnegative test supermartingale with $M_0=1$. The probability that a local model ever reaches $M_n\ge1/\alpha$ is at most $\alpha$. Monitoring repeatedly or stopping when the threshold is crossed preserves this guarantee. Predictable choices of $q_i$ based only on previous trials also work; optimizing $q$ retrospectively on the full data does not inherit the same argument.

For 850 wins in 1,000 trials and a prespecified $q=0.85$, $\log M_n\approx29.76$, giving an anytime-valid bound of about $1.18\times10^{-13}$. This is an illustrative synthetic dataset. The number is conditional on the settings and event-selection premises; it cannot rescue a compromised trial definition. Elkouss and Wehner supply sharper finite-data methods and requirements for sound data collection.

If each input bit is independent of the devices and the other input, but its probability is within $\epsilon$ of one half, a simple worst-case local bound is $\beta_\epsilon=1-(1/2-\epsilon)^2.$ Every deterministic local CHSH strategy loses on at least one input pair, whose probability is at least $(1/2-\epsilon)^2$. Replacing $\beta$ by this conservative bound makes the cost of imperfect setting randomness explicit. Correlated or leaked inputs require their own conditional model.

Where a research contribution can be precise

Memory and drift. Independent identically distributed trials are convenient but can be false. A robust analysis conditions each trial’s null bound on the complete past and uses a valid sequential test. Choosing a stopping time after looking at a conventional fixed-sample significance can invalidate the reported error probability.

Networks. Multiple sources introduce causal assumptions about source independence. Network inequalities can detect correlations not captured by a single-source description, but shared clocks, control systems, or common environmental disturbances must be represented honestly.

Certification. A statistically significant violation is not automatically sufficient for a useful randomness or key rate. Loss, finite sample size, leakage, extraction overhead, and the adversary model matter. Device-independent foundations owns this conversion question.

Relaxed assumptions. Quantifying limited setting dependence or limited communication is more informative than asserting an unrestricted loophole has been “closed.” A useful result gives an explicit bound on the departure and shows how the Bell bound changes.

An entry project: a Bell analysis that survives drift

Baseline. Simulate a CHSH game with uniformly sampled settings. A local strategy wins at most $3/4$ of trials in conditional expectation under the stated assumptions. Reproduce the ideal quantum winning probability $\cos^2(\pi/8)$, then add setting-independent detector noise.

Resources. A laptop, elementary probability, and the ability to read a martingale or test-supermartingale argument are sufficient. Keep the simulator, analysis, and random seed generation separate.

Deliverable. Implement the displayed supermartingale in log space and reproduce its analytic expectation bound. Compare it with an applicable method from Elkouss and Wehner under an identical event definition. Then compare a fixed-sample analysis with a sequentially valid test under controlled drift and predetermined stopping rules. Report false-positive behavior on local strategies, power on quantum data, and the number of accepted trials. Calibrate over many independent simulated experiments, attaching binomial uncertainty to the estimated false-positive rate; a single favorable trace is not validation. At very small target error probabilities, simulation cannot establish coverage economically: the proof supplies the guarantee, while simulation checks implementation at less extreme thresholds.

Failure criterion. If a local model exceeds the declared false-positive rate, stop and inspect conditioning, postselection, and optional stopping. Reject any improvement that discards outcome-dependent failures without including them in the null model.

Research extension. Optimize a valid witness against time-varying but bounded setting bias, or quantify robustness to imperfect source independence in a small network. Novelty requires comparison with the strongest applicable existing bound, not merely a higher raw violation.

A falsifiable extension

A tractable question is whether a predictable adaptation rule improves expected stopping time for a measured drift model without increasing trusted input assumptions. Report stopping-time distributions under several drift patterns, including ones not used to tune the rule. A gain over the fixed-$q$ example is a training result; originality requires a comparison with existing adaptive Bell analyses and a proof that the new rule remains valid.

What would change the assessment?

A new platform matters when it opens a distinct regime of causal separation, efficiency, integration, or repeatability. A new analysis matters when its assumptions are weaker or its finite-data guarantee materially improves at realistic noise. Record the full experimental contract before comparing numerical $S$ values: unlike-for-unlike rankings can reward different postselection.

References

• N. Brunner, D. Cavalcanti, S. Pironio, V. Scarani, and S. Wehner, “Bell nonlocality,” Reviews of Modern Physics 86, 419–478 (2014). arXiv:1303.2849.
• D. Elkouss and S. Wehner, “(Nearly) optimal P values for all Bell inequalities,” npj Quantum Information 2, 16026 (2016). Finite-data analysis — memory, setting bias, and event-ready statistics.
• B. Hensen et al., “Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres,” Nature 526, 682–686 (2015). Primary report.
• V. Zapatero, T. van Leent, R. Arnon-Friedman, W.-Z. Liu, Q. Zhang, H. Weinfurter, and M. Curty, “Advances in device-independent quantum key distribution,” npj Quantum Information 9, 10 (2023). Review.

Entry points

Reproduce a two-setting local bound first, then inspect an experimental supplement’s event-definition and statistical sections. The bridge from a Bell witness to a security statement is a separate reading task.