What Does It Cost to Find a Bell Violation? Adaptive Measurement and the Entanglement–Nonlocality Gap

Entanglement and Bell nonlocality are often treated as the same thing measured two ways, but they differ in an important respect. Entanglement belongs to a quantum state alone. Nonlocality (correlations between two distant labs that no classical "local hidden variable" model can reproduce) belongs to a state together with the measurement settings each lab chooses. An entangled state measured in badly chosen directions looks perfectly classical. So an experimenter handed copies of an unknown entangled state cannot simply run a Bell test. They first have to find settings that expose the violation, and each copy spent searching is a copy not spent testing. This note asks how large that search, or alignment, cost is. It also asks whether adaptive measurement, meaning choosing the next copy's settings based on earlier outcomes, reduces the cost in a way that fixed or randomized measurement schedules provably cannot.

The central bet is that the copy cost of certifying nonlocality splits into two pieces. One is a test cost that depends only on how large the violation is and does not grow with the dimension. The other is an alignment cost that does grow with dimension and is a learning problem in its own right. Adaptivity should do little in two qubits, where alignment is cheap. In higher dimensions it may matter substantially, because there the cheapest non-adaptive route to good settings is full state reconstruction, which costs about $d^3$ copies. Separately, entanglement certification with trusted devices pays no alignment cost at all. That gives the entanglement–nonlocality distinction an operational form: a gap in copy complexity, rather than only the static gaps in measurement locality catalogued so far.

Core Claim

Take $n$ copies of an unknown bipartite state $\rho$ of local dimension $d$. On each copy, Alice and Bob each perform a projective measurement, and the settings for later copies may depend on earlier outcomes. The protocol must output "nonlocal" with probability at most $\delta$ whenever the devices follow any local hidden variable model, including one with memory across rounds. When $\rho$ admits a Bell violation of margin $\Delta$ under its best settings, we conjecture that the expected number of copies behaves as

$$ n(\rho) \approx n_{\text{align}}(d, \Delta) + c\,\frac{\log(1/\delta)}{\Delta^2}, $$

where $n_{\text{align}}$ is the number of copies needed to find settings achieving at least, say, $\Delta/2$ of the margin, and $c$ is a constant. The second term follows from concentration of bounded per-round Bell scores and does not depend on $d$.

Two observations determine where adaptivity can help. First, at the optimal settings the Bell value is stationary. A misalignment of angle $\theta$ therefore loses only $O(\theta^2)$ of violation. Learning settings to precision $\theta$ costs about $1/\theta^2$ copies, so $n_{\text{align}} = O(1/\Delta)$ in fixed small dimension, which is negligible next to the $1/\Delta^2$ test cost. Our prediction for two qubits is accordingly modest: adaptive and explore-then-commit strategies should be within constant factors of each other, and the main loser should be fixed-setting tests on misaligned states. Second, in dimension $d$ the settings are $O(d^2)$ real parameters per side. The question becomes whether any non-adaptive schedule can avoid paying something like full tomography, or the dilution of violation under random settings, while an adaptive protocol pays only for the parameters the Bell value actually depends on.

Soundness does not depend on trusting the search. If the Bell score for each round is fixed before that round's outcomes arrive, the running product of likelihood-ratio-type scores is a nonnegative supermartingale under any local model with memory. Settings can therefore be chosen by an arbitrarily clever, device-dependent learner without weakening the device-independent guarantee of the final test. We will verify this numerically against adversarial local simulators.

Relation to Prior Work

Three results on adaptive measurement set the expectations here. Tight Bounds for Quantum State Certification with Incoherent Measurements proves that for mixedness testing, adaptivity gives no advantage over randomized single-copy measurements, at $\Theta(d^{3/2}/\varepsilon^2)$ copies. The role of randomness in quantum state certification with unentangled measurements shows that randomness alone closes the gap to deterministic schemes. Both results constrain this proposal and warn that adaptivity often buys nothing. Tight Bounds for State Tomography with Incoherent Measurements shows that whether adaptivity helps depends on the metric: there is no gain for trace distance, but a quadratic gain for infidelity. We reuse its decision-tree and posterior-tilt lower-bound technique. Our question is which side of that dichotomy nonlocality certification falls on.

On Adaptivity in Quantum Testing finds a worst-case factor-of-4 gain from sequential stopping in binary testing. It also finds a polynomial gap on an artificial hypothesis-selection problem, and it explicitly leaves open whether gaps appear for natural properties such as entanglement. Our hidden-alignment problem is a natural relative of that artificial problem: the state is "diagonal in one of many bases," with the bases being local rotations. Its Learn-then-Refine design is the most important baseline we have to beat.

Active learning for quantum mechanical measurements reallocates shots across a fixed set of observables to estimate fidelities. It does not change which observables exist, and it never targets Bell quantities. Few Single-Qubit Measurements Suffice to Certify Any Quantum State uses adaptivity within a single copy to certify a known pure state. That kind of adaptivity is incompatible with spacelike-separated Bell parties, so we restrict ourselves to adaptivity across copies.

On the nonlocality side, Entanglement as upper bound for the nonlocality of a general two-qubit system gives the closed-form CHSH value $2\sqrt{\lambda_1^2+\lambda_2^2}$ from the two largest singular values of the correlation matrix $T$, together with the ceiling $2\sqrt{1+C^2}$ in terms of concurrence $C$. This supplies both the two-qubit oracle for the optimal margin $\Delta$ and the ceiling states we use as test cases. Maximal Violation of the Collins-Gisin-Linden-Massar-Popescu Inequality for Infinite Dimensional States shows that the states maximally violating CGLMP (a Bell inequality for $d$-outcome measurements) are not maximally entangled, and that the violation shrinks as $d$ grows. Our high-dimensional study therefore has to separate growing alignment cost from shrinking $\Delta$. Violations of Bell inequalities from random pure states reports that random pure states violate CGLMP on average once $N\approx 8$, but only under optimized settings. Our experiments ask how much of that violation survives when settings must be learned from data.

Algorithmic construction of local models for entangled quantum states: Optimization for two-qubit states builds explicit local models for entangled two-qubit states. We use such states as negative controls, since a sound protocol must never certify them, and its strategy-sampling idea seeds our adversarial local simulators. Detecting entanglement and nonlocality with minimum observable length orders entanglement below nonlocality by how many parties must be measured jointly, using static witnesses and no sample analysis. We propose a complementary ordering by number of copies.

Research Program

The first part is theoretical. We formalize the sequential certification problem above and prove the soundness of predictable-score supermartingale tests against memoryful local models. For two qubits, we make the stationarity argument rigorous. We bound the excess copy cost of explore-then-commit with doubling over the unknown $\Delta$ against an oracle that knows the optimal settings. We then check whether any non-adaptive schedule, fixed or randomized, matches it. Our expectation is that randomized settings lose a constant factor and fixed settings fail outright on rotated states.

For dimension $d$, we study the family $(U_A\otimes U_B)|\psi^\ast_d\rangle$, where $|\psi^\ast_d\rangle$ is the CGLMP-optimal state and $U_A, U_B$ are unknown local unitaries. We attempt a lower bound for non-adaptive schedules using the tilt/likelihood-ratio method of the tomography and certification papers. In parallel, we build an adaptive protocol that learns only the local bases and phases the CGLMP value depends on, using partial tomography of the reduced eigenbases followed by sequential local refinement. Its alignment cost is targeted at $\widetilde{O}(d^2)$ copies, compared with the $d^3$ of full incoherent tomography; the $d^2$ figure comes from parameter counting and is a conjecture, not a result.

The second part is an operational comparison with entanglement. For each test state we measure the copies needed to certify entanglement with trusted devices: two-qubit correlators, or a fidelity-type witness after alignment. We then compute the ratio of nonlocality cost to entanglement cost as a function of $d$ and concurrence. The prediction is that this ratio stays bounded for states on the Su et al. ceiling and blows up for states with high concurrence but poorly expressed nonlocality, and for growing $d$.

Baselines

The naive reference is the standard fixed-setting Bell test: textbook CHSH settings, or CGLMP Fourier-basis settings, with shots split uniformly. It costs nothing to set up. On aligned states it is the best possible protocol. On rotated states it should certify nothing, which is the failure our method must repair. We run it on exactly the same states, confidence level, and copy budget.

The strongest competitor is a two-stage measure-first pipeline in the Learn-then-Refine style of On Adaptivity in Quantum Testing. Stage one spends a fraction of copies on fixed local Pauli or mutually unbiased basis tomography, using maximum-likelihood or the neural reconstruction of Neural networks for quantum state tomography with constrained measurements. Stage two optimizes settings offline and runs a fresh-data sequential Bell test. To keep the comparison fair, we tune its split fraction with a doubling schedule over unknown $\Delta$ and give it the same tuning budget and stopping rule as our method. Its only difference is that it refuses to re-aim after stage one. The file contains no reported copy cost for certifying CHSH or CGLMP nonlocality under this or any pipeline; I found no reported result on Lacuna for that quantity. The nearest reported adaptive gains are the worst-case factor of 4 from sequential testing in that paper and the measurement reduction of up to a factor of about 3 on 6-qubit fidelity estimation in Active learning for quantum mechanical measurements. Beating this baseline means a lower median copy count at matched soundness that grows with $d$. A constant-factor win in two qubits would not count as a result.

The third baseline is randomized non-adaptive settings: Haar-random local bases on each copy, scored with the best Bell functional for each drawn setting. The certification bounds in the file show this protocol is often optimal, so it directly tests whether nonlocality behaves like certification, where adaptivity is useless, or like infidelity tomography, where adaptivity helps.

The ablations separate the ingredients. "Allocation without aiming" applies Zhu et al.-style active shot allocation over a fixed setting set. "Aiming without sequential stopping" uses adaptive settings with a fixed-$n$ test. Comparing these two separates the gain from re-choosing settings from the factor that sequential stopping alone provides.

Why not just…?

Why not just do tomography first, then a Bell test on fresh copies? In two qubits this is a serious threat, and our own stationarity argument predicts it is within a constant of anything adaptive. We will report that outcome honestly as a negative result if it holds. The open case is dimension $d$. There, Tight Bounds for State Tomography with Incoherent Measurements shows that incoherent reconstruction to trace-distance accuracy needs $\Omega(d^3/\varepsilon^2)$ copies. The hidden-alignment experiment checks whether the Bell value needs far less than full reconstruction. If the two-stage cost tracks the adaptive cost as $d$ grows, the conjecture fails.

Why not just use randomized measurements, which are already optimal for certification? Tight Bounds for Quantum State Certification with Incoherent Measurements and The role of randomness in quantum state certification with unentangled measurements do support this view for distance-to-a-reference questions. A Bell score, however, is a maximum over settings, not an average. The violation Violations of Bell inequalities from random pure states reports requires optimized settings, and we expect the score under random settings to concentrate near the local bound. Experiment two measures this dilution directly as a function of $d$.

Why not just keep standard settings and allocate shots cleverly? Allocation cannot fix misalignment. By Entanglement as upper bound for the nonlocality of a general two-qubit system, a rotated state with high concurrence can give a CHSH value at or below 2 in fixed settings, and no reallocation of shots across those settings changes that. The allocation-only ablation quantifies the effect.

Why not just certify entanglement with a static witness? A witness does not certify nonlocality. The local-model states of Algorithmic construction of local models for entangled quantum states: Optimization for two-qubit states are entangled yet fully classical in Bell tests. Comparing the cost of witnessing entanglement against the cost of certifying nonlocality on the same states is one of our outputs, not a substitute for it.

Proposed Experiments

All experiments are exact-probability simulations with sampled outcomes, at confidence $\delta=10^{-3}$, over many independent runs per state.

Experiment one covers two qubits. We use Su et al. ceiling states, generic mixed states, and states with high concurrence but low CHSH value, each under random local rotations, with the optimal margin computed from the singular-value formula. Werner and Bell-diagonal states inside the Fillettaz et al. local-model region serve as negative controls. The metrics are median and 90th-percentile copies to certify as functions of $\Delta$, false-positive rate on the controls, and false-positive rate against adversarial memoryful local simulators. The decisive check is whether every adaptive variant and the two-stage baseline collapse onto the same $1/\Delta^2$ curve up to constants, with fixed settings failing on rotated states.

Experiment two measures dilution under random settings. For $|\psi^\ast_d\rangle$ and Hilbert–Schmidt random pure states with $d$ from 2 to roughly 16, we compute the distribution of the CGLMP score under Haar-random local settings. As sanity anchors under optimized settings, we reproduce the reported values from Violations of Bell inequalities from random pure states: a mean target value of about 1.083 at $N=2$, and a crossover below 1 near $N\approx 8$.

Experiment three is the hidden-alignment scaling study. On $(U_A\otimes U_B)|\psi^\ast_d\rangle$ we compare the adaptive protocol, the two-stage pipeline, randomized settings, and the ablations. We fit the exponent of copy cost against $d$ after normalizing by the known optimal $\Delta(d)$, which corrects for the shrinking violation reported by Zohren and Gill. The decisive outcome is a statistically separated exponent: near 2 for adaptive versus near 3 for the tomography pipeline, or growing without bound for randomized settings. Matching exponents would refute the claim.

Experiment four repeats experiments one and three with trusted-device entanglement certification. It reports the cost ratio between certifying nonlocality and certifying entanglement as a function of $d$ and concurrence, giving a copy-complexity counterpart to the detection-length hierarchy of Detecting entanglement and nonlocality with minimum observable length.