Boson sampling verification protocol

Answer

The supplied evidence does not establish a single universally accepted, model-independent pass/fail protocol that confirms photonic boson-sampling quantum advantage. Practical experiments use complementary statistical checks against specified alternatives and separate arguments about classical simulation cost. Passing a benchmark is evidence under its assumptions; it is not automatically proof that every classical sampler is excluded 2b32594b; 1d594cd1; 31fc4126; 3968a3b1; 9d4380ff.

There are nevertheless rigorous results with defined scope. The newly included Shchesnovich result gives a distance-based way to distinguish noisy boson sampling from classical approximations built from low-order interference, using a coarse-grained no-count observable and density-dependent sample costs. Chabaud's protocol verifies output states using single-mode Gaussian measurements. Neither should be erased by saying all verification is impossible; neither is documented here as a universal standard adopted by every large-scale photon-counting experiment a4644174; 22c9fb13.

Summary

The practical answer is a toolkit with explicit target distributions, measurement access, null hypotheses, noise assumptions and computational budgets. Likelihood tests, correlations, heavy-output/cross-entropy scores, binning and structural tests provide different checks. The evidence includes classical algorithms that outperform experimental benchmark scores, which limits standalone score-based claims; this does not prove that all protocols fail or that every experiment is classically easy 9a3e4f68; 31fc4126; 4a3ad3cb; 9d4380ff.

Evidence

| Protocol or result | What the supplied evidence establishes | Scope and limitation | |---|---|---| | Review toolkit | Bayesian/hypothesis tests, heavy-output tests, correlations and cross-entropy are described as common practice 2b32594b; 1d594cd1 | Two records of one review; common use is not a universal hardness certificate. | | Likelihood-based experimental validation | Comparison against specified uniform/distinguishable-photon alternatives 9a3e4f68 | Exact outcome-probability calculations limit the likelihood component's scale; untested null models remain open. | | Exact-probability computation | Computing relevant permanents is hard in general; specialized algorithms improve sparse or limited-bandwidth cases cade9b8f | Special-structure speedups are not generic efficient simulation, and worst-case hardness alone is not a device-specific runtime lower bound. | | Loss-aware classical simulation | A tensor-network sampler outperformed experimental benchmark scores 31fc4126 | A counterexample to treating those scores as sufficient evidence; not a theorem that every future device is easy. | | LXE against an ideal reference | Modified linear cross-entropy supplies ideal-model reference values 65ec88d2; 7f376a73 | These records concern the same primary work; a calibrated reference score does not by itself establish resistance to every classical spoof. | | Experimental detector binning | Binned photon-count distributions diagnose distinguishability and chip imperfections 4a3ad3cb | The reported small-chip demonstration is not itself an advantage-scale proof; the record also documents cross-entropy spoofing. | | Grouped statistics on Borealis | Efficient phase-space/binned comparisons find discrepancies from pure-squeezed predictions and partial improvement with a modified noise/network model 506eeafa | Agreement was reported for one-dimensional but not two-dimensional grouped probabilities; computing validation statistics is distinct from generating samples. | | Pattern recognition | Clustering distinguishes specified photon-statistics pathologies c14ab5f0 | Training and held-out performance differ; the source calls it evidence rather than a definitive test. | | Graph features and kernels | Graph-structural signatures benchmark Gaussian boson sampling 4c54c3da | The correspondence uses hafnians/graph matchings; it should not be called permanent-counting certification of arbitrary output distributions. | | Large-scale experiment | Multiple classical mockup comparisons are accompanied by an estimated cost for a specified classical method 3968a3b1 | Rejecting those mockups and estimating that method's cost do not give an unconditional lower bound for every possible classical algorithm. | | Goodman et al. 2026 | A new classical approximator and benchmark comparisons expose the role of errors beyond loss 9d4380ff | The source treats full-system total variation distance as the rigorous but impractical reference; individual practical proxies are spoofable, without claiming every joint collection is provably easy to spoof. | | Gaussian-measurement verification | Efficient output-state verification for a class of continuous-variable circuits 22c9fb13 | Requires its specified measurement protocol; it is not simply post-processing arbitrary existing click-count data. | | Bressanini detector-binning framework | Binned photon-number-resolving distributions have a computable characteristic-function framework and sensitivity to higher-order interference f2132898 | Bayesian outcome-probability tests become unscalable; low-order correlators can be efficient yet insufficiently discriminating. Neither is a universal hardness certificate. | | Direct cross-entropy spoofing result | Oh, Jiang and Fefferman construct a heuristic classical sampler that beats experimental XE scores in the intermediate, verifiable regime a43d3dbc | This demonstrates a score vulnerability. The paper does not experimentally verify its spoofer at quantum-advantage system sizes. | | Shchesnovich distance-based discrimination | Analytical separation of noisy quantum output from the specified low-order-interference classical family, using no counts in selected output ports a4644174 | Sample cost depends on boson density. This is a scoped guarantee against that family, not a universal certificate against unrestricted classical algorithms. |

The new binning record also qualifies an overbroad reading of the Borealis record: low-order correlations can be practical and efficiently computable, while being insufficiently sensitive to high-order interference. It is therefore incorrect to infer that every correlation test requires exact individual-outcome probabilities. The Borealis analysis concerns its specified tests and grouped-count implementation; its partial fit remains a result about that data/model comparison f2132898; 506eeafa.

Similarly, the direct PRL spoofing record confirms that the experimental XE score can be exceeded classically in a verifiable regime. It does not convert a claimed extrapolation to larger systems into a verified advantage-scale demonstration. Common use of XEB and insufficiency of an XEB score are compatible statements about different roles of the metric a43d3dbc.

Methodology

All seventeen supplied verified records are cited. Duplicate primary sources were grouped, not counted as independent confirmations. I separated distribution validation, state certification, rejection of named classical models and classical-cost estimates. The update retains Shchesnovich's output-statistics result and adds the newly supplied detector-binning framework and direct PRL cross-entropy spoofing record. It corrects the predecessor's blanket description of every protocol as only heuristic and its graph-permanent wording: the cited Gaussian graph correspondence is through hafnians. Source checking used primary arXiv abstracts, Goodman's full-text comparison sections and the Chabaud journal abstract in the preceding preparation. This update reopened the Bressanini main-text validation discussion, the direct PRL preprint abstract and Shchesnovich's abstract. Secondary claims carried inside the corpus were not presented as independently reread results. No scores, model sizes or runtime estimates are pooled, and the platform supplies no structured measurement conflicts.

Limitations

• The two Hangleiter–Eisert extracts retain different secondary year attributions for the cross-entropy critique and different specificity about Jiuzhang versions. They share one primary DOI, so this metadata discrepancy is disclosed rather than treated as independent scientific measurements 2b32594b; 1d594cd1.
• The corpus does not benchmark every proposed test against every documented spoofing family under matched resource and noise conditions 31fc4126; 9d4380ff; a4644174.
• Borealis's partial grouped-statistics fit, clustering's held-out performance and the large experiment's mockup comparisons retain their study-specific limitations; the corpus provides no independent replication that removes those caveats 506eeafa; c14ab5f0; 3968a3b1.
• New protocols or classical algorithms can change the evidential assessment. This report identifies the scope of the supplied literature, not a standards-body declaration or an exhaustive proof of nonexistence of a standard 22c9fb13; 9d4380ff.