{"id":"cb11ceeb-d75a-4736-96d1-ffa29ed94cd6","arxiv_id":"2411.11228","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Positive-P simulations efficiently validate GBS experiments with photon-number-resolving detectors and show Borealis data deviates from the ideal squeezed-state ground truth.","lead":"This paper uses a fast phase-space simulation method to check the outputs of Gaussian boson sampling (GBS) quantum experiments. Applied to the Borealis experiment's public data, it finds the measured photon counts deviate strongly from the ideal theoretical distribution, though a two-parameter noise model fits some one-dimensional statistics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The positive-P validation in §IV.B only covers uniform-loss Haar networks; Borealis's structured time-domain-multiplexed T-matrix may have different sampling-error convergence, so the reported within-error agreement of the modified model is not yet established.","rationale":"The paper's method is exact in the infinite-sample limit, so the only numerical risk is whether the sampling-error estimates are reliable for the actual structured Borealis network. This is load-bearing for the 'within sampling error' statement about the modified model (Table II), but not for the primary 'discrepancies with perfect squeezing' claim, because ZEI values of 66-167 are far beyond plausible sampling-error inflation. The reader's weakest assumption names exactly this transfer-of-validation problem; I agree with it. A truncated-network exact comparison would settle it. I also note the 288-mode ideal ground truth relies on private communication [53], a reproducibility gap, but the overall conclusion survives without that data set. Accordingly the verdict should remain conditional: accept with conditions on verifying structured-network convergence and making the 288-mode ground-truth correction public.","tokens_in":24853,"tokens_out":5953,"duration_ms":63671,"concrete_test":"Truncate the actual Borealis T-matrix to M=30 modes (keeping the corresponding squeezing vector and loss structure), compute the d=1 total-count GCP exactly for counts m≤10 using a standard Hafnian routine for PNR GBS, and compare with positive-P simulations at the same ES=1.2×10^6 used in Section V. If the normalized differences exceed the estimated σT,i by more than 2σ, the sampling-error transfer to structured networks fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim—that the two-parameter thermalized model brings the Borealis d=1 GCPs within sampling error—relies on the positive-P sampling errors σT,i estimated by sub-ensemble splitting (Appendix B). Section IV.B validates the method only against exact uniformly-lossy Haar-random networks (tU), while the actual Borealis network uses a time-domain-multiplexed transmission matrix with mode-dependent losses and strong per-mode structure documented in Section V.D. Nothing in Sections IV or V checks that the stochastic convergence and the Gaussian sub-ensemble variance estimate remain accurate for this structured, non-random network. If σT,i is underestimated, the χ2 values for the modified model are inflated and the claimed |ZET|≈1 agreement is not meaningful; conversely, the large ZEI discrepancies are robust to this issue. A secondary support gap is the 288-mode ideal ground truth, which is corrected on the basis of private communication [53] rather than a public reproducible source; this affects one of the two quantum-advantage data sets, though not the overall discrepancy conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends the positive-P phase-space representation method to compute grouped count probabilities (GCPs) for Gaussian boson sampling with photon-number-resolving (PNR) detectors. The method is validated against exact total-count distributions for Haar-random unitary networks with uniform amplitude loss, and its numerical scaling is measured for d=1,2,3 GCPs on up to 480 modes. The method is then applied to the Borealis experimental data (16, 72, 216-LS, 216-HS, and 288 modes). The authors report that the experimental data are far from the ideal pure-squeezed ground truth (Z_EI between 66 and 167 for the quantum-advantage-relevant sets), and that a two-parameter modification of the ground truth—a thermalization fraction ϵ and a transmission-matrix correction t—brings the one-dimensional total-count GCPs within sampling error for the 72, 216-LS, 216-HS, and 288-mode data sets. However, the same model fails to describe two-dimensional GCPs and per-mode photon-number moments, which the authors attribute to additional systematic errors.","tokens_in":25072,"tokens_out":7729,"duration_ms":79073,"significance":"If the claims hold, the paper provides a scalable and computationally efficient validation toolkit for GBS experiments with PNR detectors, with claimed speedups of order 10^18 over direct Hafnian-based simulation. The central discrepancy result—that the experimental data deviate strongly from the ideal pure-squeezed ground truth—is parameter-free, consistent across multiple data sets, and robust to the sampling-error concerns. The paper also provides an open-source implementation (xqsim) and makes concrete falsifiable predictions, namely that a two-parameter thermalized model can capture the d=1 total-count distributions but not the d=2 or per-mode statistics. These strengths make the paper a useful contribution to the validation methodology for photonic quantum advantage experiments.","major_comments":[{"comment":"The positive-P sampler is validated only against Haar-random unitary networks with uniform loss (tU) in Section IV.B. The Borealis simulations in Section V use the actual structured time-domain-multiplexed T-matrix with mode-dependent losses described in Section V.D, and the error bars entering the χ² tests are the sub-ensemble estimates σ_T,i from Appendix B (Eq. 29). If σ_T,i is underestimated for this non-uniform structured network, the reported |Z_ET| ≈ 1 agreement in Table II would be inflated. Please provide a direct convergence test on a Borealis-like structured matrix—for example, by computing exact GCPs for a small sub-network with the experimentally measured T-matrix and squeezing vector and comparing them to positive-P results—or explicitly discuss how sensitive the ET agreement is to the accuracy of the sampling-error estimates.","section":"§IV.B, §V.D, Appendix B"},{"comment":"The ideal ground truth for the 288-mode data set is corrected based on private communication (Ref. [53]), and the corrected distribution is reported to be closer to the experimental data than the distribution originally published in the Borealis paper. This correction is not publicly verifiable, yet it directly affects the Z_EI = 167 result in Table I and the associated discrepancy claim for the largest quantum-advantage data set. Please make the corrected ground-truth distribution publicly available (e.g., as a data file in the repository) or provide a detailed reproducible derivation so that this result can be independently checked.","section":"§V.B, Table I, Ref. [53]"},{"comment":"The two-parameter (ϵ, t) modified ground truth is fitted separately to each data set using a Nelder-Mead simplex algorithm, but the paper does not specify the cost function minimized, the initial values and convergence criteria, or how the quoted ±0.0005 uncertainties on ϵ and t were obtained. Since the ET agreement in Table II is a central claim, please provide these details, along with confidence intervals or a χ² surface plot, so the reader can judge whether the fit is well constrained and whether the agreement is a meaningful test or an over-parameterized post-hoc fit.","section":"§V.B, Table II, §V.C"}],"minor_comments":[{"comment":"The statement that the 288-mode distribution is '98σN further' from the expected mean than the 216-mode HS distribution is arithmetically inconsistent with the reported Z_EI values of 70 and 167; the difference is 97σN, so '97σN' would be correct.","section":"§V.B"},{"comment":"The abstract states a speedup of '~10^18 times faster' without specifying the baseline; the Conclusion clarifies that this is relative to direct Hafnian computation on Fugaku. Please state the baseline in the abstract for clarity.","section":"Abstract and Conclusion"},{"comment":"Several LaTeX/OCR artifacts appear in the text, including '/dispiint' in Eq. (15) and Eq. (21), and '/radicaltp' in Appendix B. These should be corrected to the intended integral and square-root symbols.","section":"§III.B, Eq. (15), Eq. (18), Eq. (21), Appendix B"},{"comment":"The caption states that ϵ and t each have error bars of ±0.0005 for all data sets, but the origin of these error bars is not described. Please clarify whether these are fit uncertainties, estimated from the spread over the ten simulation runs, or derived from the χ² surface.","section":"Table II"},{"comment":"The notation f_c = 2F1(a,b;c;z) is used without defining the hypergeometric parameters a, b, c explicitly; Eq. (24) would be easier to follow if the arguments of the hypergeometric function were written out.","section":"§IV.A, Eq. (24)"}],"recommendation":"major_revision","confidential_remarks":"This paper addresses a timely and important problem—validation of GBS experiments—and the parameter-free discrepancy result is convincing. The main risks are (i) the validation gap for structured transmission matrices, which affects the reliability of the ET sampling-error agreement, and (ii) the reliance on private communication for the 288-mode ground truth, which is a reproducibility concern. The fitted two-parameter model is appropriately framed as a modification, but the ET agreement should not be oversold as a predictive test. I believe the paper is within the journal's scope and that the issues can be addressed with additional validation and data/code release, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper extends the positive-P grouped-count-probability (GCP) method from threshold to photon-number-resolving detectors, and applies it to Borealis data. It's a practical step forward: PNR detectors measure more information, and the method avoids the inverse Fourier transforms needed for threshold detectors. The core computational claim is well supported: validation against exact uniform-loss Haar networks gives errors below 1e-3 for the relevant loss regime, and run-times are minutes on a desktop. The central experimental finding—Borealis count distributions are 66 to 167 sigma from the ideal pure-squeezed ground truth—is parameter-free and robust.\n\nThe paper is also refreshingly honest. It explicitly reports that the thermalized two-parameter model, while bringing d=1 total-count distributions within sampling error, fails for d=2 GCPs and for per-mode photon number moments. It does not oversell the 16-mode dataset, which remains far from both ground truths. That transparency is real credit.\n\nThe soft spots are the ones the stress-test flags, and they are real but not fatal. First, the positive-P sampler is validated only on Haar-random uniform-loss networks; the actual Borealis T-matrix is structured (time-domain multiplexed) with mode-dependent losses. Sub-ensemble error estimates are standard, but no independent check shows that sampling-error convergence is unchanged for that structured network. That's a gap the referee should ask to close—for example, a smaller exactly solvable network with the same T-matrix structure. Second, the 'small modification greatly improves agreement' is a two-parameter fit to the same d=1 data per dataset, so the reported chi-square values don't account for degrees of freedom lost. The paper acknowledges the model's limitations, but a chi-square with proper df would be more honest. Third, the 288-mode ideal ground truth rests on private communication [53]; that's a reproducibility issue for one dataset, not a load-bearing flaw.\n\nNet: this is a useful paper for anyone working on GBS validation, noise models, or classical simulation boundaries. The central discrepancy claim holds up. It should go to peer review; with the structured-network check and a public ground-truth specification, I'd accept it.","headline":"A practically useful, honest extension of GCP validation to PNR detectors; the central discrepancy claim is robust, while the fitted thermal model is clearly labeled and fails in higher dimensions.","tokens_in":25608,"tokens_out":3298,"would_cite":true,"duration_ms":30974,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that grouped count probabilities for photon-number-resolving Gaussian boson samplers can be simulated efficiently with the positive-P phase-space method, and that applying the test to large recent data shows disagreement…","keywords":["Gaussian boson sampling","positive-P phase-space representation","grouped count probabilities","photon-number-resolving detection","quantum advantage validation","thermalized squeezed states","statistical hypothesis testing"],"falsifier":"Compute exact grouped count probabilities for the actual 16-mode transmission matrix using exact Hafnian evaluation, and compare them with positive-P GCPs; a mismatch beyond the quoted sampling error would show the method does not transfer to the structured network, and this check is feasible because the 16-mode data set is small enough for exact probabilities.","tokens_in":24612,"feed_emoji":"🎲","tokens_out":9059,"duration_ms":85879,"temperature":0.7,"pith_summary":"This paper claims that the positive-P phase-space representation can compute grouped count probabilities for Gaussian boson sampling with photon-number-resolving detectors far faster than direct classical simulation, making partial validation of large experiments practical. Applied to recent 72-, 216-, and 288-mode data, the method shows that the experimental counts as a whole disagree with the ideal pure-squeezed-state ground truth. Adding two corrections, a thermalization fraction and a transmission-matrix measurement correction, brings the one-dimensional total-count distribution into agreement within sampling error for the larger data sets. The same corrections fail for two-dimensional grouped counts and per-mode photon-number moments, indicating further systematic errors that the small number of experimental samples cannot resolve. The authors propose using such tests as feedback to improve Gaussian boson sampling hardware.","feed_headline":"A fast phase-space test shows GBS data diverges from ideal theory","feed_subtitle":"Thermal noise and a transmission correction rescue one-dimensional counts, not higher-order statistics.","key_machinery":"The engine is the positive-P phase-space distribution paired with grouped count probabilities, or GCPs. The positive-P representation expands any density operator as a positive distribution over pairs of independent coherent-state amplitudes, so normally ordered expectation values, including photon-counting projectors of all orders, become ordinary moments that can be estimated by random sampling. A GCP is the probability that the total photon number in each of $d$ subsets of output modes takes specified values; in one dimension it is the total count distribution, and as $d$ approaches the mode number the GCP converges to the full Hafnian distribution, where the Hafnian is the matrix function whose evaluation for a count pattern is #P-hard. The positive-P samples are generated from Gaussian inputs through a quadrature-variance formula, with thermalized squeezed states parameterized by $\\epsilon$, and propagation through the lossy transmission matrix is done at the amplitude level.","core_discovery":"The central claim is that binning a photon-number-resolving Gaussian boson sampling output into grouped count probabilities turns an exponentially hard validation problem into an efficiently samplable one, provided the normally ordered positive-P distribution is used. The paper generalizes Mandel's binning of classical photon statistics to arbitrary quantum states: each stochastic sample of the positive-P distribution directly estimates a multidimensional grouped count probability, so the method captures arbitrarily high-order correlations without computing a single Hafnian. In the loss-dominated regime of current experiments the sampled moments converge to exact distributions with errors below $10^{-3}$. When the same simulation is run on the published transmission matrix of a 216- and 288-mode processor, the ideal squeezed-state ground truth is rejected by chi-square and Z-statistic tests, while a thermalized ground truth with two fitted parameters ($\\epsilon$ and $t$) reproduces the one-dimensional total count distribution within sampling error. The paper therefore establishes that current data are inconsistent with the ideal target distribution, that low-dimensional tests can be passed by a decoherent effective theory, and that higher-dimensional and per-mode tests expose residual systematic errors.","pith_inferences":["Because one-dimensional counts can be fitted by two parameters while two-dimensional counts cannot, the real error model is likely underdetermined by total-count data alone; mode-dependent losses or phase noise would be needed to explain the higher-order discrepancies. (Editorial inference.)","The periodic structure visible in per-mode photon-number moments points to the time-multiplexed network layout, and a classical sampler exploiting that structure might pass one-dimensional validation tests even where the full distribution is nonclassical. (Editorial inference.)","The same generalized binning machinery applies to any nonclassical input state, not only squeezed states, so the validation protocol could be exported to other photon-counting devices beyond Gaussian boson sampling. (Editorial inference.)","A natural next step would be to scan a wider parameter space, including mode-dependent thermalization and per-mode transmission corrections, and test whether any higher-dimensional model can simultaneously fit the one- and two-dimensional GCPs before claiming further systematic errors. (Editorial inference.)"],"forward_implications":["One-dimensional GCP validation runs in minutes on a desktop for 288 modes, a speed-up the paper estimates at more than $10^{18}$ over Hafnian-based direct simulation.","Every analyzed data set is at least 66 standard deviations from the ideal pure-squeezed ground truth in the one-dimensional total-count test.","A thermalized state with $\\epsilon \\approx 0.05$ and a transmission correction $t \\approx 0.98$--$0.99$ brings the total-count distribution within sampling error for the 72-, 216-LS, 216-HS, and 288-mode data sets.","The same corrections fail to reconcile two-dimensional grouped counts and per-mode photon-number moments, and the 16-mode data set with the most samples remains far from both ideal and corrected theory.","The authors propose using these validation tests as feedback to adjust experimental parameters, which they suggest may be a more practical route to quantum advantage than hardware improvements alone."],"supporting_citations":[{"why":"Supplies the experimental photon-count data and transmission matrices used for all experimental comparisons.","marker":"[5]"},{"why":"Introduces the positive-P phase-space simulation method for Gaussian boson sampling validation that this paper extends to photon-number-resolving detectors.","marker":"[9]"},{"why":"Provides the Hafnian formula for Gaussian boson sampling count probabilities, the exact benchmark the simulation bypasses.","marker":"[7]"},{"why":"Origin of the binning method for photon-counting distributions that this paper generalizes to nonclassical states.","marker":"[11, 12]"},{"why":"Prior validation tests of threshold-detector Gaussian boson sampling using a decoherent target distribution, setting the trend and permutation-test methodology.","marker":"[17]"},{"why":"Supplies exact lossy Gaussian boson sampling distributions and the estimated Hafnian computation cost used to quantify the speed-up.","marker":"[32]"},{"why":"Defines the generalized positive-P representation used for all nonclassical state sampling.","marker":"[42]"},{"why":"Corrects the ideal ground-truth distribution for the 288-mode data set; the discrepancy claim for that data set depends on it.","marker":"[53]"}],"fun_headline_variants":["Positive-P binning makes GBS validation tractable","Fast simulation shows GBS data fails ideal test","Phase-space method exposes GBS mismatch with ideal","Ideal GBS fails, thermal model rescues low-order stats","Binning photon counts with positive-P speeds validation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on assuming the positive-P sampling errors measured on uniform-loss random networks transfer to the real processor's structured, mode-dependent-loss network—and, for the 288-mode data, that a privately corrected ideal ground truth is the right comparison.","fun_headline_variants_meta":{"raw":{"variants":["Positive-P binning makes GBS validation tractable","Fast simulation shows GBS data fails ideal test","Phase-space method exposes GBS mismatch with ideal","Ideal GBS fails, thermal model rescues low-order stats","Binning photon counts with positive-P speeds validation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00085,"raw_usage":{"total_tokens":3714,"prompt_tokens":977,"completion_tokens":2737,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":2661}},"tokens_in":593,"tokens_out":2737,"duration_ms":18570,"temperature":1.0,"reasoning_tokens":2661,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:46:46.004331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute exact grouped count probabilities for the actual 16-mode transmission matrix using exact Hafnian evaluation, and compare them with positive-P GCPs; a mismatch beyond the quoted sampling error would show the method does not transfer to the structured network, and this check is feasible because the 16-mode data set is small enough for exact probabilities.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the positive-P phase-space simulation method for Gaussian boson sampling validation that this paper extends to photon-number-resolving detectors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior validation tests of threshold-detector Gaussian boson sampling using a decoherent target distribution, setting the trend and permutation-test methodology."},{"cited_title":"Deshpande, A","cited_arxiv_id":null,"evidence_quote":"Supplies exact lossy Gaussian boson sampling distributions and the estimated Hafnian computation cost used to quantify the speed-up."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the generalized positive-P representation used for all nonclassical state sampling."},{"cited_title":"Photon counting distribution for the ideal ground truth of the 288-mode data was inde- 19 pendently cross-validated","cited_arxiv_id":null,"evidence_quote":"Corrects the ideal ground-truth distribution for the 288-mode data set; the discrepancy claim for that data set depends on it."}],"review_version":1}