{"id":"4537ef46-d9c2-43e0-9fbc-d59a29a71f35","arxiv_id":"2607.18184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An open-source Python library stably solves entropic Schrödinger bridges at very small regularization and passes four idealized 2D benchmarks, including a 90° rotation recovered to 0.07°.","lead":"anyakrakusuma is an open-source Python library that solves entropic Schrödinger bridge problems—the entropically regularized optimal transport between point clouds—with a stable log-domain Sinkhorn iteration. The paper demonstrates the solver on four synthetic 2D shapes, including recovery of a known 90-degree rotation to within 0.07 degrees, with code, data, and diagnostics released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Underflow premise is mathematically false at the stated ratios: exp(-400) ~ 1.9e-174, far above double-precision minimum, so the 'log-domain necessary' claim is unsupported.","rationale":"The reader's conditional verdict is appropriate, but I identified a different load-bearing concern than the reader's weakest_assumption. The reader focused on the Case 3 ground-truth rotation being approximate; that is a real limitation but secondary. The more central issue is the paper's headline numerical justification for log-domain necessity: the stated cost-to-regularization ratios do not cause Gibbs-kernel underflow. This is a concrete, checkable mathematical fact, not a matter of interpretation. It undermines the abstract's 'necessary rather than merely convenient' claim and the corresponding conclusion. The paper can likely be repaired by correcting the underflow statement and/or by actually demonstrating a failure mode for exponential-domain Sinkhorn at these or more extreme parameters, so the verdict remains CONDITIONAL rather than moving to ACCEPT or REJECT. The reader did note the missing exponential-domain baseline, so there is partial overlap, but their weakest assumption was the Case 3 rotation test, which I do not regard as the single most load-bearing concern.","tokens_in":28083,"tokens_out":12703,"duration_ms":125072,"concrete_test":"For each case, compute K_min = exp(−max_ij C_ij / ε) from the archived cost matrices and ε values, and compare to 2.2×10^−308. If all values are above that threshold, the underflow premise is contradicted. Then run a standard exponential-domain Sinkhorn iteration (u,v updates without log-sum-exp) on all four cases at the stated parameters and tolerance 1e−9. If it converges without overflow/underflow, the 'necessary' claim is empirically false; if it fails, identify precisely which quantity (K entries, u, v, row sums) exceeds double-precision range. Optionally repeat with ε reduced so that max C_ij/ε ≈ 700–800 to confirm the claimed underflow regime is actually attainable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and §4 assert that the log-domain formulation is 'necessary rather than merely convenient' because at max C_ij/ε between 252 and 400, the Gibbs kernel K_ij = exp(−C_ij/ε) of Eq. (16) 'underflows double precision across most of its range.' This numerical premise is incorrect. exp(−400) ≈ 1.9×10^−174, while double-precision underflow starts near 2.2×10^−308 (subnormals only extend to ~5×10^−324). Even exp(−700) ≈ 9.9×10^−305 is still representable; true underflow of K would require C_ij/ε ≳ 708. Thus none of the four cases (ratios 252–400) produces a Gibbs kernel entry below double-precision range. The claimed failure mode does not occur at the reported parameters, and no exponential-domain baseline is run to show that a naive Sinkhorn would fail for another reason, such as scaling-vector overflow. The favorable residual histories therefore establish only that this implementation is stable, not that log-domain is necessary. This is an internal inconsistency in the paper's own numbers, not a disagreement with community consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents anyakrakusuma, an open-source Python library for solving the discrete static Schrödinger bridge problem via a log-domain Sinkhorn–Knopp iteration, together with a diagnostic pipeline that analyzes the optimal coupling, entropic interpolation, and intermediate-time marginals. The library is demonstrated on four constructed planar benchmarks: circle-to-circle dilation, spiral-to-Gaussian-mixture fragmentation, two-moons rigid reorientation, and Lissajous-to-trefoil deformation. The main claims are that the log-domain formulation is necessary at the studied regularization parameters because the Gibbs kernel underflows double precision over most of its range, that the iteration reaches a marginal residual below 10^-9 with geometric contraction factors between 0.966 and 0.976, and that covariance-based diagnostics recover an imposed 90-degree rotation to within 0.07 degrees. The paper is unusually explicit about its limitations, including the translation confound in Case 3, temporally interpolated KDE snapshots, the factor-of-two effect of plan subsampling, and the invalidity of endpoint entropy estimates for noiseless curve-supported clouds.","tokens_in":28276,"tokens_out":5236,"duration_ms":61988,"significance":"If the claims are appropriately qualified, the paper is a useful software contribution. The mathematical derivation of the discrete Schrödinger bridge is standard and correctly presented, the implementation details are transparent, and the release of code, configuration files, diagnostic scripts, and NetCDF archives is a genuine strength. Several internal consistency checks are well designed, for example the recovery of log 4 from the difference of curve-supported endpoint entropy estimates and the log 4 agreement of the subsampled block entropy. The paper also carefully distinguishes descriptive fits from predictive claims and states the regimes in which each diagnostic is valid. The main scientific value lies in the reproducible solver-plus-diagnostics workflow rather than in new mathematical results. However, two load-bearing claims need correction: the numerical premise behind the 'log-domain is necessary' claim is quantitatively wrong at the reported parameters, and the Case 3 rotation-recovery claim is stronger than the generative experiment actually supports.","major_comments":[{"comment":"The central claim that the log-domain formulation is 'necessary rather than merely convenient' rests on the assertion that at cost-to-regularization ratios of 252–400 the Gibbs kernel K_ij = exp(-C_ij/ε) 'underflows double precision across most of its range.' This is numerically incorrect. exp(-400) ≈ 1.9×10^-174 and exp(-252) ≈ 1.5×10^-110, both far above the double-precision minimum around 2.2×10^-308, and even exp(-700) is still representable. True underflow would require C_ij/ε ≳ 708. Thus none of the four cases produces a kernel entry below representable range, and no exponential-domain baseline is run to demonstrate failure by another mechanism such as scaling-vector overflow. The favorable residual histories show that this implementation is stable, not that the log-domain formulation is necessary at these parameters. This claim appears in the abstract, the Discussion, and the Conc","section":"§2.3, Eq. (41); §4 first limitation; Table 4"},{"comment":"The Case 3 rotation validation is weaker than the text claims. The target is not obtained by rotating the source cloud; it is an independent fresh two-moons draw rotated by 90 degrees about its own empirical centroid. Consequently, the source and target covariance principal axes differ by 90 degrees only in expectation, not as a property of the realized clouds. The reported 89.9304° recovery with an uncertainty of ±2.8204° is computed from within-realization subsampling of a single pair of clouds and does not include the between-realization variability of the two independent draws. The phrase 'recovers the imposed ninety-degree reorientation' therefore overstates the validation: it is a consistency check of the estimator against a generative recipe. A concrete fix is to rotate the source cloud itself about its common centroid, or to generate multiple independent source/target pairs and r","section":"Abstract; §4; §5"}],"minor_comments":[{"comment":"The factor-of-two bias in the conditional perplexities is acknowledged in the Discussion, but Table 2 reports perplexities in 'retained target points' without flagging that the full-plan values are approximately twice as large. Since the table is likely to be read in isolation, it should either report full-plan values or include an explicit footnote in the table itself.","section":"Table 2 and §4"},{"comment":"The reference entropy H_diff(t) is defined for t in (0,1); the formula has a logarithmic singularity at t=0 and t=1. The text is clear in context, but adding t∈(0,1) to the equation or its surrounding sentence would prevent confusion.","section":"Eq. (55)"},{"comment":"Minor typesetting: 'LL VM' should be 'LLVM'. Also, the statement that fastmath=True permits associative reordering is fine, but the claim that this is inconsequential because the updates are shift-invariant could be stated more precisely: shift invariance does not by itself guarantee identical fixed points under reordered floating-point arithmetic, although the observed residuals make this a non-issue here.","section":"§2.2"},{"comment":"The Authors state that animation rendering accounted for 82.5%–91.8% of total wall-clock time and that the solve is a minor fraction; the single-sweep Case 1 solve time includes JIT compilation. This is clearly explained, but the timings table might benefit from a column excluding one-time JIT cost for reproducibility comparisons.","section":"§4"},{"comment":"The phrase 'unit marginal fidelity' is used without definition. It is presumably the total plan mass / marginal residual check in Eq. (36); defining it in the abstract is unnecessary, but the term should be introduced exactly where Eq. (36) is discussed.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The underflow claim is a numerically checkable statement and is simply wrong at the reported parameters; since it is repeated in the abstract and conclusions, it is a load-bearing overstatement that should be corrected before publication. The Case 3 validation issue is also substantive but fixable by re-running with a truly rotated source cloud or by presenting an expectation-based analysis. The manuscript is otherwise careful, reproducible, and technically sound; the software release and diagnostic design are strengths that deserve credit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid software paper with one real error in the abstract. The library and benchmark pipeline are genuinely useful; the \"necessary\" claim about log-domain Sinkhorn is based on a numerical mistake and should be corrected.\n\nWhat's new: not the math — that's standard entropic OT / Schrödinger bridge theory, and the paper says so. The contribution is the open-source Python library, the diagnostic pipeline, and four reproducible synthetic benchmarks, with a particularly nice rotation recovery (89.93° vs 90°) and a good explanation of why it works across a masked near-isotropic interval. The paper is refreshingly explicit about its own limitations: endpoint entropy caveat, factor-of-two subsampling bias, interpolated snapshots, the translation confound in Case 3. That honesty earns real credit.\n\nThe soft spots, in proportion. The main one is the claim that the Gibbs kernel underflows double precision at their parameters. exp(-400) is about 1.9e-174, still far above double precision's minimum normal and subnormal range; true underflow would need ratios around 700 or more. So the ratios 252–400 do not produce underflow, and the \"log-domain is necessary\" claim is unsupported. A proper fix is either to drop \"necessary\" and say \"numerically robust\" or to run an exponential-domain baseline to show failure. That is a real error in the central abstract claim, but it is an overstatement, not a fatal flaw in the library: log-domain Sinkhorn is standard practice at small regularization, and the residual histories stand on their own.\n\nOther concerns are minor and mostly acknowledged by the authors: Case 3's \"imposed rotation\" is only approximate because the target is an independent noisy draw plus a translation; the entropy at noiseless endpoints measures curve length rather than entropy; the stored plan is subsampled, which the authors themselves correct. The contraction-factor fit to the same residual sequence used for stopping is mildly circular but harmless given the flat per-step ratios.\n\nBottom line: the paper is worth seeing through peer review. The honesty and the concrete software make it a useful contribution to the computational entropic OT community. I'd want the underflow claim fixed before publication and the \"necessary\" language toned down. The math otherwise checks out, the data and scripts are archived, and the analysis is careful enough to deserve a serious referee.","headline":"A useful, unusually candid software paper with one wrong headline number: the Gibbs kernel does not underflow double precision at their stated ratios, so the 'log-domain necessary' claim is unsupported.","tokens_in":28911,"tokens_out":2421,"would_cite":true,"duration_ms":50650,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A log-domain Sinkhorn–Knopp iteration solves discrete entropic Schrödinger bridges at cost-to-regularization ratios up to 400, where the Gibbs kernel underflows double precision, yet it still reaches a marginal residual of 10⁻⁹ and recovers","keywords":["entropic optimal transport","Schrödinger bridge problem","Sinkhorn–Knopp algorithm","log-domain stabilization","log-sum-exp","point-cloud coupling","covariance principal-axis analysis","differential entropy estimation"],"falsifier":"Run the same four cases with a plain exponential-domain Sinkhorn implementation: if it converges to a 10⁻⁹ marginal residual without underflow at cost-to-regularization ratio 400, the paper's necessity claim collapses. Conversely, rebuild the two-moons target as the exact source cloud rotated 90° about the source centroid and check whether the recovered angle moves closer to 90.0000° and the transport cost falls by the translation contribution; if not, the rotation recovery is an artifact of the generative recipe rather than a measurement.","tokens_in":27891,"feed_emoji":"🌉","tokens_out":8276,"duration_ms":78861,"temperature":0.7,"pith_summary":"This paper presents an open-source Python library that solves the discrete static Schrödinger bridge problem—the entropically regularized optimal transport coupling between two empirical point clouds—and reconstructs the intermediate-time bridge. Its central claim is that working in the log-domain of the Sinkhorn potentials is necessary, not merely convenient, at the small-regularization settings tested: costs exceed regularization by up to 400-fold, so the raw Gibbs kernel underflows double precision across most entries, yet the iteration reaches a marginal residual of 10⁻⁹ with unit marginal fidelity in every case. The strongest quantitative check is a covariance-based recovery of an imposed 90-degree reorientation in a two-moons case to within 0.07°, about forty times smaller than the estimator's uncertainty, even across a near-isotropic interval where the principal axis is unobservable. The paper also shows that endpoint differential-entropy estimates on noiseless curve-supported clouds measure curve-length ratios rather than entropy, and it carefully delineates the regimes in which each diagnostic is well defined.","feed_headline":"Log-domain Sinkhorn solves bridges when naive kernel underflows","feed_subtitle":"A Python library reaches 10⁻⁹ residuals and recovers a 90-degree rotation to within 0.07 degrees on four planar tests.","key_machinery":"The load-bearing mechanism is the log-domain Sinkhorn–Knopp iteration on dual potentials f and g instead of exponential scalings, with a numerically stable log-sum-exp (the max-trick) as its primitive. This replaces products of underflowing kernel entries by shift-invariant additions, which is what lets the iteration run where exp(−C_ij/ε) sits below the double-precision floor. The same log-domain identity monitors the row-marginal residual without materializing the plan, and the bridge is rebuilt by sampling from the optimal coupling and adding Brownian-bridge Gaussian noise with variance εt(1−t).","core_discovery":"The paper solves the discrete Schrödinger bridge as a quadratic transport cost plus ε times the KL divergence to the product measure, using log-domain Sinkhorn–Knopp updates with a stabilized log-sum-exp. At cost-to-regularization ratios up to 400, the Gibbs kernel underflows double precision over most entries, so exponential-domain scaling cannot run; the log-domain potentials stay bounded and reach a 10⁻⁹ marginal residual with unit fidelity in every case. Bridge reconstruction samples Brownian bridges conditioned on the optimal coupling, and diagnostics span coupling entropy, barycentric projections, kernel-density region counts, k-nearest-neighbor entropy estimates, and covariance ellips","pith_inferences":["If log-domain stabilization is what enables the runtimes shown here, the same potentials-based recursion should transfer directly to unbalanced and multi-marginal extensions, where dense cost-matrix storage rather than underflow becomes the binding constraint.","The contraction factors do not order with ε across the three fitted cases; a single-geometry ε-sweep would separate the influence of regularization from transport geometry on convergence speed, an experiment the paper leaves to future work.","Re-running the two-moons case with the target built by rotating the exact source cloud about the exact source centroid should move the recovered angle closer to exactly 90° and lower the transport cost by the translation contribution, offering a sharper validation of the rotation-recovery pipeline.","The paper's endpoint-entropy diagnosis suggests a simple testable extension: adding small isotropic noise to the noiseless curve-supported generators should make the endpoint entropy estimates well-defined and collapse the apparent net entropy production toward the two-dimensional expectation."],"forward_implications":["At cost-to-regularization ratios up to 400, entropic transport between point clouds remains solvable to 10⁻⁹ residuals, a regime exponential-domain scaling cannot enter; this extends practical entropic optimal transport to much smaller regularization than naive implementations allow.","The two-moons covariance test shows that a principal-axis rotation estimator can recover an imposed 90° rotation to within 0.07° even across a near-isotropic masked interval, provided the doubled angle is thresholded and unwrapped correctly.","On noiseless curve-supported endpoints, differences of k-nearest-neighbor entropy estimates at equal sample size estimate the ratio of curve lengths, not entropy production; this gives a concrete diagnostic check and suggests perturbing the generators to place all cases on a common footing.","The circle-to-circle case converges in a single sweep because its circulant cost matrix admits constant potentials; the residual at machine epsilon therefore serves as a correctness check for the solver rather than a convergence baseline.","Observed per-iteration contraction factors are local near-fixed-point rates; the worst-case contraction bound is vacuous at these settings, so the measured rates do not contradict classical Sinkhorn theory.","The factor-of-two offset in row perplexity caused by strided plan storage means reported conditional entropies understate full-plan values by log 2, and the full plan should be archived when storage permits."],"fun_headline_variants":["Log-domain Sinkhorn solves Schrödinger bridges when kernels underflow","90° rotation recovered to 0.07° by log-domain bridge solver","Schrödinger bridge solver hits 10⁻⁹ residual despite kernel underflow","Log-domain Sinkhorn: 10⁻⁹ residual and 0.07° rotation on bridges","Bridge library avoids underflow, nails rotation to 0.07 degrees"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The 90-degree rotation recovery in the two-moons case is the paper's strongest quantitative validation, and it rests on the premise that the target is the source shape rotated exactly 90° about a common center; in the implementation the target is a fresh independent draw rotated about its own empirical centroid, so the imposed rotation holds only in expectation and the two clouds also differ by a translation of magnitude 0.56178.","fun_headline_variants_meta":{"raw":{"variants":["Log-domain Sinkhorn solves Schrödinger bridges when kernels underflow","90° rotation recovered to 0.07° by log-domain bridge solver","Schrödinger bridge solver hits 10⁻⁹ residual despite kernel underflow","Log-domain Sinkhorn: 10⁻⁹ residual and 0.07° rotation on bridges","Bridge library avoids underflow, nails rotation to 0.07 degrees"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3228,"prompt_tokens":851,"completion_tokens":2377,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":2272}},"tokens_in":595,"tokens_out":2377,"duration_ms":16754,"temperature":1.0,"reasoning_tokens":2272,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:44:36.229228+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four cases with a plain exponential-domain Sinkhorn implementation: if it converges to a 10⁻⁹ marginal residual without underflow at cost-to-regularization ratio 400, the paper's necessity claim collapses. Conversely, rebuild the two-moons target as the exact source cloud rotated 90° about the source centroid and check whether the recovered angle moves closer to 90.0000° and the transport cost falls by the translation contribution; if not, the rotation recovery is an artifact of the generative recipe rather than a measurement.","supporting_citations":[],"review_version":1}