{"id":"52045540-a718-4afa-a299-6e3f7f4db7fb","arxiv_id":"2506.07775","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A two-photon GPT-tomography experiment finds no evidence against tomographic locality in quantum theory, while its real-amplitude sub-sector shows the predicted failure.","lead":"An experiment on pairs of polarized photons tested whether composite quantum systems are fully characterized by local measurements, the principle of tomographic locality. No violation was found in full quantum data, while a control test on the real-amplitude sector produced the expected signature of failure.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central comparison relies on exactly factorizing transformations, yet the paper never quantifies the validity of that assumption; the existing conditional verdict is appropriate.","rationale":"The strongest claim is the null result on tomographic locality. For that claim to hold, the comparison of D(S,E) with D(~Sprod,~Eprod) must be testing the right notion of 'product'. The paper is unusually careful in acknowledging that preparation and measurement procedures are not exactly product, and in defining the tensor product structure through transformations (Sec. III C). This makes the exact factorization of transformations the single load-bearing assumption. The manuscript does not quantify how close the realized transformations are to product form; the phrase 'much smaller' is not a bound. Without such a bound, a reasonable scenario exists in which small alignment or coupling imperfections introduce a small interaction between the two modes, and this coupling changes which GPT vectors count as 'product' in the inferred secondary set. The positive control is impressive but cannot rule this out by itself, since both analyses share the same assumption. I therefore agree with the reader's identification of the weakest assumption. I do not see an additional independent flaw that would require moving the verdict: the model-selection results are consistent with quantum expectations, the effective-rank thresholds are empirically justified, and the real-sector benchmark behaves exactly as predicted. The appropriate disposition remains conditional on supplying either a quantitative characterization of the factorization assumption (e.g., a sensitivity analysis) or the raw data and code that would allow the re-analysis proposed above.","tokens_in":15433,"tokens_out":8536,"duration_ms":120185,"concrete_test":"Re-analyze the full dataset with the factorization constraint relaxed: retain the d=16 model and the same 90/10 train-test protocol of Sec. III, but allow each waveplate transformation to include a small non-product term, T = T_A⊗T_B + ε C with normalized C, and select ε by cross-validation. If the optimal ε is statistically distinguishable from 0, or if the inferred effective rank of D(~Sprod,~Eprod) changes for ε at the estimated waveplate misalignment level, the factorization assumption is falsified and the reported null result is not robust. As a complementary stability check, rerun the secondary-procedure optimization in Eq. (1) with many random initializations and verify that the resulting rank of D(~Sprod,~Eprod) is unchanged.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's verdict on tomographic locality is obtained by comparing rank D(S,E) with effective rank D(~Sprod,~Eprod) (Secs. III E and IV). This comparison is meaningful only relative to a tensor product structure, and that structure is fixed in Sec. III C by imposing exact factorization on the transformations, T = T_A ⊗ T_B. This imposition is the operational definition of 'product' used in the secondary-procedure construction of Sec. III D, so any non-factorizing component in the realized transformations can deform the inferred product subspace and distort the rank comparison in either direction. The paper's only defense is the assertion that deviations from factorization are 'much smaller' for transformations than for preparations or measurements; no quantitative bound, calibration, or sensitivity analysis is provided. The real-amplitude control does not settle this, because both the full and the restricted analyses share the same factorization assumption; a systematic non-product component could in principle generate or mask a rank gap while preserving the internal consistency of the control. A related, more technical weakness is that the secondary-procedure objective in Eq. (1) uses an ordinary 2-norm in the GPT vector space, although the GPT representation is defined only up to the gauge freedom D_M = SΛΛ^(-1)E noted in Sec. II; the inferred closest product vectors, and hence the effective ranks, are not shown to be gauge-independent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports an experimental test of tomographic locality using the polarization degrees of freedom of a pair of photonic modes in a prepare-and-measure scenario. The data are analyzed within the framework of generalized probabilistic theories (GPTs): a GPT tomography fit is performed with the tensor product structure defined by imposing exact factorization on the transformation procedures, T = T_A ⊗ T_B. The authors reconstruct sets of GPT state and effect vectors, use secondary procedures to infer strictly factorizing states and effects from the convex hull of the reconstructed sets, and compare the effective ranks of the full and product matrices. For the full 60×60 dataset they find effective rank 16 for both D(S,E) and D(˜Sprod,˜Eprod), concluding that there is no evidence for a failure of tomographic locality. For the 24×24 real-amplitude submatrix they find effective ranks 10 and 9 respectively, reproducing the expected failure signature and serving as a positive control.","tokens_in":15694,"tokens_out":7012,"duration_ms":83265,"significance":"If the central result is sound, this is the first direct experimental test of tomographic locality, a central axiom in GPT reconstructions of quantum theory. The analysis is theory-agnostic, uses cross-validation to select the model dimension, and includes a positive control that correctly detects a known violation (real-amplitude quantum theory); these are genuine strengths. The main null result is also robust in the sense that the relevant singular values are large and well separated, so the conclusion does not hinge on a fine-tuned threshold. However, the validity of the result depends on the empirical justification of the factorization assumption that defines the tensor product structure, and on the gauge invariance of the secondary-procedure objective; these points need to be established before the headline claim can be fully accepted.","major_comments":[{"comment":"The tensor product structure that defines 'product' is fixed by imposing exact factorization on the transformations, T = T_A ⊗ T_B. The paper states that deviations from factorization are 'much smaller' for transformations than for preparations or measurements, but no quantitative bound, calibration, or sensitivity analysis is provided. Since the secondary-procedure construction and the inferred product subspace are defined relative to this structure, a non-factorizing component in the realized transformations could deform the inferred product subspace and shift the effective ranks in either direction. The real-amplitude control shares the same factorization assumption, so it cannot by itself rule out such a systematic distortion. Please provide a quantitative test of the factorization assumption, for example by independently characterizing the realized transformations or by injecting non-factorizing perturbations into the fit and showing that the rank comparison is stable.","section":"Sec. III C"},{"comment":"The secondary-procedure objective minimizes the 2-norm |s_sec − s| in the GPT vector space, but the GPT representation is fixed only up to the gauge freedom D_M = SΛΛ^{-1}E noted in the footnote to Eq. (1). The closest factorizing vector, and hence the matrices D(˜Sprod,˜Eprod) and their effective ranks, are not shown to be gauge-invariant. Because the factorization constraint on transformations is also representation-dependent, the reported rank comparison may depend on an arbitrary gauge choice. The authors should either fix a canonical gauge and justify it, or replace the 2-norm in Eq. (1) with a gauge-invariant measure based directly on the data matrix.","section":"Eq. (1) and Sec. II"},{"comment":"The optimization in Eq. (1) is nonconvex, because the factorizing constraint s_sec = s_A ⊗ s_B is bilinear, and the paper reports no information about multiple random restarts, convergence diagnostics, or global optimality for the SLSQP runs. If the optimizer returns local minima for some of the secondary states or effects, the effective rank of D(˜Sprod,˜Eprod) could be underestimated or overestimated. Please report the distribution of objective values over restarts and the sensitivity of the final ranks to the initialization, at least for the secondary procedures used in the central comparison.","section":"Sec. III D"}],"minor_comments":[{"comment":"The caption of Fig. 5 says that the shaded regions denote singular values below 10^{-1.6}, which are treated as effectively zero, while the text in Sec. IV states that the threshold is 10^{-1}. These numbers should be reconciled.","section":"Fig. 5 caption vs Sec. IV"},{"comment":"In the paragraph describing the real-amplitude analysis, the sentence 'We can then compute D(S,E) and D(˜Sprod,˜Eprod)' should presumably refer to D(S_real,E_real) and D(˜S_real_prod,˜E_real_prod), consistent with the notation used elsewhere.","section":"Sec. IV"},{"comment":"There are minor typographical errors: 'for the the insight' appears in the Introduction, and 'factorizng' appears in Sec. III D. These should be corrected.","section":"Introduction and Sec. III D"},{"comment":"The parameter count for the structured parameterization is written as '2(d^2 − 1) + 13d^2 + 13d^2 + 2d^2'; the duplication of the 13d^2 term is confusing and should be explained or corrected.","section":"Footnote 4"}],"recommendation":"major_revision","confidential_remarks":"The reader's stress-test concern about the factorization assumption is valid and should be the primary focus of the revision; the gauge-invariance issue raised in my second major comment is also important and may require more than a local fix. The paper is within scope for this journal, and the positive control is encouraging, but the central claim needs additional support before I can recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is the first experimental test of tomographic locality, and it is largely convincing. The real-amplitude control is the strongest part—it shows the method can detect a known violation, which makes the null result for complex quantum theory meaningful.\n\nThe paper extends GPT tomography to a bipartite system for the first time; earlier work covered single qubit and qutrit. It also introduces a principled way to define the tensor product structure operationally via factorizing transformations rather than preparations or measurements. That is a real methodological contribution, and it is clearly explained.\n\nThe main caveat is the one the reader flagged: the tensor product structure is fixed by assuming T = T_A ⊗ T_B for the transformations, and the paper only argues qualitatively that this is a much better approximation than for preparations or measurements. No quantitative bound or sensitivity analysis is provided. The real-amplitude control shares this assumption, so it does not independently validate the factorization. I do not see this as a load-bearing flaw: the transformations are implemented with physically separated waveplates, and the main null result (rank 16 vs 16) rests on large singular values, so small non-factorizing components would have to be sizable to overturn it. Still, the authors should either quantify the factorization error or show that the verdict is stable under plausible deviations.\n\nA second, minor technical point: the secondary-procedure objective in Eq. (1) uses a 2-norm in the GPT vector space, and because of the gauge freedom D_M = SΛΛ^{-1}E, the inferred closest product vectors are not shown to be gauge-invariant. This is a gap in the proof, not a demonstrated error in the result.\n\nOn the plus side, the analysis is theory-agnostic, uses cross-validation to select model dimension, and the noise threshold for effective rank is empirically justified. The paper does not ship raw data or code, which limits independent verification but is not unusual for this kind of letter.\n\nWho is this for: foundations readers and anyone doing theory-agnostic tomography. It deserves a serious referee. My own verdict is conditional: the central claim holds up, but the factorization assumption should be addressed in revision, ideally with a sensitivity analysis or a calibration measurement.","headline":"First direct test of tomographic locality, with a clean positive control; the transformation-factorization assumption is the main caveat, but it does not sink the result.","tokens_in":16198,"tokens_out":1538,"would_cite":true,"duration_ms":18669,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.65.Ta","42.50.Xa"],"model":"deepseek-v4-flash","headline":"A direct two-photon experiment finds no violation of tomographic locality—the principle that local measurements fully determine a composite system's state—while a real-amplitude control reproduces the predicted failure signature.","keywords":["tomographic locality","generalized probabilistic theories","GPT tomography","real-amplitude quantum theory","tensor product structure","stabilizer states","two-photon polarization","effective rank"],"falsifier":"Take the same two-photon setup and intentionally make the two modes' transformations non-factorizing, for example by inserting a single waveplate that acts jointly on both modes before the measurement stage, and re-run the GPT fit; if the effective rank of the product-state matrix rises toward the full rank even though the physical state space is unchanged, then the factorization assumption, rather than tomographic locality, is carrying the verdict.","tokens_in":15236,"feed_emoji":"🔬","tokens_out":10070,"duration_ms":104005,"temperature":0.7,"pith_summary":"This paper reports an experimental test of tomographic locality, the principle that the state of a composite system is fully determined by the statistics of measurements on its parts. Using generalized probabilistic theory (GPT) tomography, a fitting method that does not assume quantum mechanics, the authors analyze polarization measurements on pairs of photonic modes and compare the dimension spanned by all realized states with the dimension spanned by strictly product states. For the full quantum data, the two effective ranks are both 16, so the experiment provides no evidence for a violation of tomographic locality. For a control subset restricted to real-amplitude quantum states, the effective ranks are 10 and 9, reproducing the predicted signature of failure. This matters because tomographic locality is a central axiom in reconstructions of quantum theory, and a direct test shows what that axiom buys at the precision frontier.","feed_headline":"Quantum tomographic locality passes a direct two-photon test","feed_subtitle":"A theory-agnostic fit finds no rank gap; the real-amplitude control shows the predicted gap.","key_machinery":"The central object is the matrix of GPT probabilities and its effective rank. States are vectors in a real vector space $V_A \\otimes V_B$, effects are vectors in the dual space, and measurement probabilities are inner products collected in a matrix $D_M = S E$. The tensor product structure is defined operationally through the transformations: each mode's waveplates are modeled as acting independently, so $T = T_A \\otimes T_B$. For preparations and measurements that target product states, the authors use the secondary-procedure technique to find the closest strictly factorizing state or effect inside the convex hull of all fitted states or effects. They then form the matrices $D(S,E)$, $D(\\tilde{S}_{\\mathrm{prod}}, \\tilde{E}_{\\mathrm{prod}})$ and their real-amplitude analogues, and compare effective ranks by counting singular values above a noise-calibrated threshold. The stabilizer formalism supplies the 60 states and 60 effects, with 36 product, 24 entangled, and a 24-element real-amplitude subset.","core_discovery":"On the paper's own terms, the central discovery is a method and a null result: tomographic locality can be tested in a prepare-and-measure experiment without presupposing quantum mechanics, and when the test is run on the polarization of two photonic modes, the effective rank of the full state-effect matrix and the effective rank of the strictly product state-effect matrix both come out to 16, so there is no evidence that non-separable measurements reveal anything beyond the local statistics. The same analysis applied to the real-amplitude subset yields effective ranks 10 and 9, a mismatch that is exactly the signature of a failure of tomographic locality. The paper takes this control result as confirmation that the method is sensitive enough to detect the phenomenon it is designed to test, and as support for the meaningfulness of the full-data null result.","pith_inferences":["The rank-gap diagnostic could be applied to other foil theories, such as quaternionic or fermionic quantum theory, where the failure of tomographic locality should produce its own characteristic dimensional deficit.","Because the verdict depends on a singular-value threshold calibrated from noise, replacing the threshold with a Bayesian or likelihood-based rank comparison could turn the binary pass/fail into a continuous bound on how much the product sector can deviate from the full sector.","A calibration experiment that deliberately introduces coupling between the two modes' transformations and watches how the inferred product rank responds would turn the factorization assumption from a postulate into a measured quantity."],"forward_implications":["A direct experimental route is now open for testing any proposed failure of tomographic locality: prepare a tomographically complete set of states and effects, fit a GPT model, construct strictly product secondary states, and compare effective ranks.","The equal 16-versus-16 ranks in the full data set mean that, at the achieved precision, local measurements are sufficient to characterize the photon-pair state, just as complex quantum theory predicts.","The 10-versus-9 rank gap in the real-amplitude control shows the method would catch a genuine failure of tomographic locality, lending weight to the null result in the full data.","Defining the tensor product structure through factorizing transformations rather than through preparations or measurements gives an operational way to identify bipartite structure when states and effects are slightly non-product, which is also needed for experimental entanglement assessment."],"supporting_citations":[{"why":"Supplies the GPT tomography framework for prepare-measure experiments and the secondary-procedure technique for constructing strictly factorizing states and effects.","marker":"[16]"},{"why":"Extends GPT tomography to a photonic three-level system and demonstrates the theory-agnostic fitting approach this bipartite experiment builds on.","marker":"[17]"},{"why":"Provides the train-and-test model-selection methodology used to choose the embedding dimension and to diagnose overfitting.","marker":"[18]"},{"why":"Provides the SLSQP optimizer used to fit the GPT models to the experimental data.","marker":"[19]"},{"why":"Defines the stabilizer states and effects that form the tomographically complete set of preparations and measurements.","marker":"[20]"},{"why":"Defines real-amplitude quantum theory, the foil theory whose predicted tomographic-locality failure the control data are designed to simulate.","marker":"[7]"}],"fun_headline_variants":["Two-photon test finds no violation of tomographic locality","Tomographic locality survives a direct two-photon probe","Real-amplitude control shows the predicted locality failure","Quantum locality principle passes two-photon experiment","No rank gap in quantum test, but real-amplitude fails"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the transformations applied to the two photonic modes factorize across the modes as $T = T_A \\otimes T_B$, since this factorization is what operationally defines the tensor product structure that separates product from entangled states, and a significant violation of it would distort the inferred product subspace and the rank comparison.","fun_headline_variants_meta":{"raw":{"variants":["Two-photon test finds no violation of tomographic locality","Tomographic locality survives a direct two-photon probe","Real-amplitude control shows the predicted locality failure","Quantum locality principle passes two-photon experiment","No rank gap in quantum test, but real-amplitude fails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1463,"prompt_tokens":936,"completion_tokens":527,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":449}},"tokens_in":552,"tokens_out":527,"duration_ms":6498,"temperature":1.0,"reasoning_tokens":449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:25:18.450408+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same two-photon setup and intentionally make the two modes' transformations non-factorizing, for example by inserting a single waveplate that acts jointly on both modes before the measurement stage, and re-run the GPT fit; if the effective rank of the product-state matrix rises toward the full rank even though the physical state space is unchanged, then the factorization assumption, rather than tomographic locality, is carrying the verdict.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extends GPT tomography to a photonic three-level system and demonstrates the theory-agnostic fitting approach this bipartite experiment builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the train-and-test model-selection methodology used to choose the embedding dimension and to diagnose overfitting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SLSQP optimizer used to fit the GPT models to the experimental data."},{"cited_title":"Virtanen, R","cited_arxiv_id":null,"evidence_quote":"Defines the stabilizer states and effects that form the tomographically complete set of preparations and measurements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines real-amplitude quantum theory, the foil theory whose predicted tomographic-locality failure the control data are designed to simulate."}],"review_version":1}