{"id":"6d1641ec-8e59-439b-834d-bc17582f3930","arxiv_id":"1909.02373","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A semi-supervised squared-loss mutual information estimator that couples unpaired samples through optimal transport matches full-data estimates using only tens of paired examples.","lead":"This paper proposes a method for estimating mutual information when only a handful of matched examples are known but many unmatched examples are available. The method learns a soft matching between unmatched examples and uses it to estimate the density ratio that defines the mutual information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Π-coupling is learned by maximizing the same density-ratio cross-term it is meant to estimate, so without bias/consistency analysis the reported SMI values may be inflated; a bias test on known-SMI data is needed.","rationale":"The reader identified exactly the same weakest assumption: Π as a surrogate for p(x,y) with no consistency analysis. I agree. The paper's own derivation shows that the coupling matrix is not an innocent plug-in: it is optimized to maximize the current density-ratio model's cross term (Eq. 6), hence the alternating scheme is a self-reinforcing loop. This is the load-bearing step for the central claim, because the SMI estimate is computed from r_α over all unpaired pairs (Section 2), and any inflation of r_α from self-selected couplings directly raises \\hat{SMI}. The monotone-convergence result is about the surrogate loss, not about the true SMI. The experiments lack error bars and do not report a bias benchmark with known ground-truth SMI, so the empirical support cannot distinguish accurate estimation from a bias shared with the reference method. I therefore recommend keeping the reader's conditional verdict: acceptance should require a bias analysis or at least an independence-control experiment. No change to the verdict is needed, since the reader already made it conditional.","tokens_in":12295,"tokens_out":6981,"duration_ms":77450,"concrete_test":"On synthetic data with analytically known SMI, fix n=20 paired samples and nx=ny=500 unpaired samples, and run LSMI-Sinkhorn with the paper's default settings (b=200, ϵ=0.3, λ,β chosen by their cross-validation) over 50 random trials for (a) independent Gaussians X,Y (true SMI=0) and (b) a Gaussian copula with correlation ρ=0.5 (true SMI computable in closed form). Report mean ± std of \\hat{SMI}. If for case (a) the mean exceeds, say, 0.05 or 2 standard errors above 0, or for case (b) the mean exceeds the true SMI by more than 20%, the coupling step is demonstrably injecting spurious dependence and the central claim fails as an unbiased estimator. This directly tests whether Eq. (5)'s Π-optimization estimates or creates dependence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the coupling matrix Π in Eq. (5) is an adequate surrogate for p(x,y) in the cross-term of Eq. (4). This premise is not supported and is the paper's weakest point. For fixed α, the Π-subproblem in Eq. (6) is an entropy-regularized OT problem whose cost C_α = K^T diag(α)L is exactly the current model's cross-term; Sinkhorn will concentrate mass on pairs with large r_α(x_i,y_j). Updating α in Eq. (7) then fits r_α to this self-selected mixture. Consequently the joint optimization is a positive feedback loop that can inflate the estimated density ratio on unpaired pairs, and \\hat{SMI}, which averages (r_α−1)^2 over all unpaired combinations, inherits this inflation. Proposition 1 only proves monotone descent of the surrogate J; it says nothing about consistency of \\hat{SMI} or about convergence of Π to the true coupling. No bias or consistency analysis is provided. The unregularized limit (ϵ→0) drives Π toward a hard assignment, the LSOM special case, which is a matching method, not a density estimator. Thus the method's ability to 'estimate the SMI without a large number of paired samples' is not established; the empirical curves in Figs. 3–6 cannot rule out that the gap to LSMI(full) is small because both are biased in the same direction, and the independence ('Random') panels show small but apparently positive values with no error bars.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript addresses the problem of estimating squared-loss mutual information (SMI) when only a small number of paired samples and many unpaired samples are available. It introduces a density-ratio model r_α(x,y)=α^T φ(x,y), approximates the cross term ∫ r_α p(x,y) by a weighted combination of paired samples and unpaired samples aligned through a coupling matrix Π, and formulates a joint optimization problem (Eq. 5) regularized by entropy on Π and L2 on α. The proposed LSMI-Sinkhorn algorithm alternates between a closed-form α update (Eq. 8) and a Sinkhorn update for Π (Eq. 6); Proposition 1 proves monotone decrease of the surrogate objective. Experiments on synthetic data, UCI datasets, deep image matching, and photo album summarization are reported to support the claim that SMI can be estimated from tens of pairs plus unpaired marginals.","tokens_in":12651,"tokens_out":6281,"duration_ms":60932,"significance":"If the estimation claim holds, the paper makes a useful contribution: it is, to my knowledge, the first semi-supervised SMI estimator, it avoids the O(n^3) Hungarian matching by using Sinkhorn (O(nx ny)), it comes with code, and its alternating scheme is simple and monotonically decreasing. The applications to image matching and photo album summarization broaden the potential impact beyond MI estimation. However, the statistical validity of the estimator is the load-bearing issue: because Π is learned from the same density-ratio cross-term it is meant to replace, the manuscript needs a bias/consistency argument or a convincing repeated-run bias experiment before the 'can estimate SMI from few pairs' claim can be accepted. The current single-run plots do not provide such evidence.","major_comments":[{"comment":"The central statistical claim is not supported because the coupling matrix Π is learned by maximizing the same density-ratio cross-term it is introduced to estimate. For fixed α, the Sinkhorn subproblem in Eq. (6) is equivalent to maximizing Σ_ij π_ij α^T φ(x'_i,y'_j), so mass is concentrated on pairs with large current density-ratio values; the subsequent α update in Eq. (7) then fits r_α to this self-selected mixture. Proposition 1 proves only monotone decrease of the surrogate J, not convergence of Π to the true joint density, nor unbiasedness or consistency of \\hat{SMI}. The 'ideal' discussion in Eq. (4) also presumes a latent true pairing of the unpaired samples. The manuscript should add either a consistency/bias analysis or, at minimum, a bias experiment on synthetic data with known SMI (including the independent case) with repeated trials before claiming that SMI can be estimated from tens of paired samples.","section":"§3.1, Eqs. (4)–(5); §3.2, Proposition 1"},{"comment":"The empirical evidence does not rule out a shared upward bias. The SMI estimation plots report single runs without error bars or repeated trials, and the 'Random' panels show small but apparently positive SMI values. Moreover, 'LSMI (full)' is itself an estimate from 10,000 paired samples and is treated as ground truth in both synthetic and UCI experiments. With single runs, the closeness of LSMI-Sinkhorn to LSMI (full) could simply mean that both estimators are biased in the same direction. Please report mean and standard deviation over multiple random data splits, and where possible compare to an analytic or high-accuracy numerical SMI value.","section":"§5.3, Figures 3, 5, and 6"},{"comment":"The role of the entropic regularization ϵ in the statistical behavior of the estimator is not analyzed. The paper fixes ϵ=0.3 without a sensitivity study, and the ϵ→0 limit of Eq. (6) drives Π toward a hard assignment, reproducing LSOM, which the authors themselves describe as a matching method rather than a density estimator. Since no consistency result is provided for the regularized estimator, the influence of ϵ on the bias-variance trade-off of \\hat{SMI} should be investigated analytically or empirically; otherwise the choice of ϵ appears arbitrary for the estimation claim.","section":"§3.3; §5.1"}],"minor_comments":[{"comment":"The photo album summarization description and Figures 8 and 9 are duplicated verbatim in the manuscript; please remove the duplicate passages.","section":"§5.5"},{"comment":"The text states that H is a positive definite matrix, but as a sum of outer products H is positive semidefinite; only H + λI is guaranteed to be positive definite. Please correct this wording.","section":"§3.2, after Eq. (8)"},{"comment":"The matching-accuracy curves are shown without error bars, and the text does not specify how many random shuffles were averaged or how exactly top-1 and top-2 accuracies are defined; please clarify.","section":"§5.4, Figure 7"},{"comment":"There are small typos such as 'Gromove-Wasserstein' in §3.3, and the related-work discussion of LSOM would benefit from consistently citing the same references ([22] vs [24]) for the original algorithm.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses an interesting and useful problem, and the algorithmic machinery is correct as far as it goes (monotone decrease, closed-form α, Sinkhorn update, O(nx ny) complexity). My main hesitation is that the paper's headline claim—that SMI can be reliably estimated from a few pairs plus unpaired marginals—rests on an unexamined self-training assumption about Π. I would want the revision to provide either a consistency/bias analysis or a carefully designed repeated-run bias experiment with known ground truth, in addition to error bars in the main figures. The duplicated section in the full text should also be fixed before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper generalizes least-squares mutual information estimation to the semi-supervised setting by replacing the permutation coupling in LSOM with an entropy-regularized optimal transport coupling (Sinkhorn). That is a genuine contribution: it handles unequal sample sizes, soft matching, and runs in O(nx ny), versus the Hungarian or Gromov-Wasserstein alternatives. The optimization derivations are correct — alpha has a closed form, the Pi step is a standard Sinkhorn projection, and Proposition 1's monotone decrease of the surrogate objective is sound. The experiments show the method landing closer to LSMI(full) than the baselines on synthetic and UCI data, and the image matching / album summarization demos illustrate the rectangular matching capability. Code is provided.\n\nThe soft spot is the one the stress-test note flags, and it is real. The coupling Pi is optimized against the same cross-term used to fit the density ratio: the cost matrix in Eq. (6) is exactly r_alpha on unpaired pairs. Sinkhorn will concentrate mass where r_alpha is already high, and the alpha update then fits that self-selected mixture. That is a positive feedback loop that can inflate the estimated SMI. Proposition 1 says nothing about consistency or about Pi converging to the true joint, and the paper offers no bias analysis. The 'Random' panels show small positive SMI but without error bars they cannot rule out a systematic upward bias. This does not sink the paper — the method may well work — but the central claim is not yet established.\n\nMinor issues: no error bars or repeated runs in Figures 3, 5, 6, 7; 'LSMI (full)' is a reasonable baseline but not an analytic ground truth; and the manuscript has duplicated Section 5.4/5.5 blocks plus typos that suggest a rushed final pass.\n\nWho this is for: people working on semi-supervised dependence estimation, OT-based matching, or applications needing rectangular matching. It deserves a serious referee. I would send it to peer review and ask for a bias/consistency check — analytic SMI on Gaussian data, repeated runs with variance, and an ablation with Pi fixed to a uniform coupling — before accepting.","headline":"A clean semi-supervised extension of LSMI via Sinkhorn coupling, with a real self-training bias concern that needs analysis before the estimates are trusted.","tokens_in":13176,"tokens_out":4279,"would_cite":true,"duration_ms":47801,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H20","62B10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Squared-loss mutual information can be estimated from a few data pairs plus many unpaired samples, by learning an optimal-transport alignment and fitting a density-ratio model.","keywords":["mutual information estimation","squared-loss mutual information","density ratio estimation","optimal transport","Sinkhorn algorithm","semi-supervised learning","image matching","photo album summarization"],"falsifier":"On synthetic data with a known SMI value (for example $Y=0.5X+\\mathcal{N}(0,0.01)$ with $X$ standard normal), compute the ground-truth SMI from a very large paired sample, then run LSMI-Sinkhorn with $n=10$ true pairs and $n_x=n_y=1000$ unpaired marginal samples; if the estimate remains far from ground truth while the learned $\\Pi$ is sharply concentrated on a visibly wrong matching, the surrogate-coupling assumption is refuted.","tokens_in":12103,"feed_emoji":"📊","tokens_out":9302,"duration_ms":79056,"temperature":0.7,"pith_summary":"This paper sets out to establish that mutual information need not be estimated from a large paired sample: a small set of true pairs plus many unpaired draws from the two marginals can suffice. Its target is Squared-loss Mutual Information (SMI), the chi-squared divergence between the joint density $p(x,y)$ and the product $p(x)p(y)$. The authors reformulate SMI estimation as a joint density-ratio fitting and optimal-transport problem, in which an entropic-regularized coupling matrix $\\Pi$ stands in for the unknown alignment of unpaired samples. They show experimentally that the resulting LSMI-Sinkhorn algorithm tracks the full-sample estimate with as few as tens of pairs, and they apply the same machinery to image matching and photo album summarization.","feed_headline":"Tens of paired samples suffice for mutual information estimation","feed_subtitle":"Unpaired marginal samples fill the joint gap through optimal-transport alignment.","key_machinery":"The load-bearing object is the coupling matrix $\\Pi$ with $\\Pi\\mathbf{1}_{n_y}=n_x^{-1}\\mathbf{1}_{n_x}$ and $\\Pi^\\top\\mathbf{1}_{n_x}=n_y^{-1}\\mathbf{1}_{n_y}$, learned by Sinkhorn iterations; it converts unpaired marginal samples into a soft surrogate for the joint density. The objective couples $\\Pi$ with the density-ratio parameter $\\alpha$ through the vector $h_{\\Pi,\\beta}=\\frac{\\beta}{n}\\sum_i \\phi(x_i,y_i)+(1-\\beta)\\sum_{i,j}\\pi_{ij}\\phi(x'_i,y'_j)$, so that fixing either variable makes the other subproblem convex. This alternating structure, together with the rank-$b$ cost matrix $C_\\alpha=K^\\top\\mathrm{diag}(\\alpha)L$, gives the monotone-decrease guarantee and the $O(n_xn_y)$ iteration cost.","core_discovery":"The central claim is that estimating SMI from limited paired data can be cast as minimizing the squared error between the true density ratio $r(x,y)=p(x,y)/(p(x)p(y))$ and a kernel model $r_\\alpha(x,y)=\\alpha^\\top \\phi(x,y)$, with the joint term in the loss approximated by a weighted mix of the few true pairs and a coupling $\\Pi$ over unpaired samples. The authors solve $\\min_{\\Pi,\\alpha} J(\\Pi,\\alpha)$ under marginal constraints on $\\Pi$, alternating between a closed-form least-squares update $\\hat\\alpha=(H+\\lambda I)^{-1}h_{\\Pi,\\beta}$ and a Sinkhorn update for $\\Pi$. This yields a monotonically decreasing objective, $O(n_xn_y)$ per-iteration cost, and, in the experiments, SMI values close to those of a full-sample estimator while pair-only and Gromov-Wasserstein-based baselines deviate substantially. The paper presents this as the first semi-supervised SMI estimator, and demonstrates downstream use in deep image matching and rectangular photo album summarization.","pith_inferences":["Editorial extension: if the coupling surrogate is reliable, the same plug-in trick should extend to other $f$-divergences or Wasserstein dependency measures wherever unpaired marginals are abundant.","Editorial extension: because the convergence proof only shows monotone decrease of $J$, not convergence of $\\Pi$ to the true joint, a stress test with deliberately misspecified kernels or badly chosen basis vectors would reveal whether the estimator's accuracy comes from the coupling or mainly from the paired-sample term.","Editorial extension: the rank constraint on $C_\\alpha$ suggests that random-feature or Nystr\\\"om approximations of the kernel could scale the method well beyond the 10,000-sample demonstrations reported."],"forward_implications":["With only tens of paired samples and abundant unpaired marginals, SMI values close to a full-sample estimate can be obtained on synthetic and UCI benchmarks.","The alternating scheme decreases the objective monotonically and converges in a few iterations, making the estimator stable enough to be used as a training objective.","Per-iteration cost $O(n_xn_y)$ makes the method feasible for large unpaired sets, with GPU-accelerated runs on 10,000 unpaired samples taking seconds.","Because it allows $n_x\\neq n_y$, the method extends matching and summarization to rectangular layouts such as 320 images into a $16\\times 20$ grid, where square-only methods fail.","The learned coupling is concentrated along the diagonal for dependent data and uniform for independent data, providing a direct visual diagnostic of dependence."],"supporting_citations":[{"why":"Defines Squared-loss Mutual Information and the least-squares density-ratio estimation approach that the paper extends.","marker":"[20]"},{"why":"Introduces the Sinkhorn algorithm for entropic-regularized optimal transport used to optimize the coupling matrix.","marker":"[5]"},{"why":"Supplies the log-stabilized Sinkhorn variant that makes the alternating updates numerically robust.","marker":"[16]"},{"why":"Presents Least-Squares Object Matching, which the paper shows is a special case of its optimization framework.","marker":"[22]"},{"why":"Provides the cross-domain object matching formulation and model-selection ideas the paper builds on and contrasts with.","marker":"[24]"},{"why":"Defines Gromov-Wasserstein distances, the main alignment baseline and runtime comparison in the experiments.","marker":"[11]"},{"why":"Gives the computational complexity and entropic-regularization baselines for Gromov-Wasserstein that the paper compares against.","marker":"[14]"},{"why":"Introduces kernelized sorting, the prior approach for photo album summarization that LSMI-Sinkhorn extends to rectangular layouts.","marker":"[15]"}],"fun_headline_variants":["LSMI-Sinkhorn: MI from few pairs via OT","Sinkhorn fills the pair gap for mutual information","Optimal transport leverages unpaired data for MI","Semi-supervised MI estimation via optimal transport"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned coupling matrix $\\Pi$ is a valid stand-in for the true joint density $p(x,y)$ over the unpaired samples; if Sinkhorn converges to a coupling that is well matched to the density-ratio model but not to the real correspondence, the estimated SMI can be systematically wrong, and the paper offers no consistency or bias analysis to rule this out.","fun_headline_variants_meta":{"raw":{"variants":["LSMI-Sinkhorn: MI from few pairs via OT","Sinkhorn fills the pair gap for mutual information","Optimal transport leverages unpaired data for MI","Semi-supervised MI estimation via optimal transport"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000671,"raw_usage":{"total_tokens":3081,"prompt_tokens":995,"completion_tokens":2086,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":611,"completion_tokens_details":{"reasoning_tokens":2023}},"tokens_in":611,"tokens_out":2086,"duration_ms":17258,"temperature":1.0,"reasoning_tokens":2023,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:51:50.885620+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On synthetic data with a known SMI value (for example $Y=0.5X+\\mathcal{N}(0,0.01)$ with $X$ standard normal), compute the ground-truth SMI from a very large paired sample, then run LSMI-Sinkhorn with $n=10$ true pairs and $n_x=n_y=1000$ unpaired marginal samples; if the estimate remains far from ground truth while the learned $\\Pi$ is sharply concentrated on a visibly wrong matching, the surrogate-coupling assumption is refuted.","supporting_citations":[{"cited_title":"BMC Bioin- formatics 10(S52) (2009)","cited_arxiv_id":null,"evidence_quote":"Defines Squared-loss Mutual Information and the least-squares density-ratio estimation approach that the paper extends."},{"cited_title":"In: NIPS (2013)","cited_arxiv_id":null,"evidence_quote":"Introduces the Sinkhorn algorithm for entropic-regularized optimal transport used to optimize the coupling matrix."},{"cited_title":"SIAM Journal on Scientific Computing 41(3), A1443–A1481 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the log-stabilized Sinkhorn variant that makes the alternating updates numerically robust."},{"cited_title":"IEEE transactions on Pat- tern Analysis and Machine Intelligence 37(9), 1764–1776 (2015)","cited_arxiv_id":null,"evidence_quote":"Presents Least-Squares Object Matching, which the paper shows is a special case of its optimization framework."},{"cited_title":"In: AISTATS (2011)","cited_arxiv_id":null,"evidence_quote":"Provides the cross-domain object matching formulation and model-selection ideas the paper builds on and contrasts with."},{"cited_title":"Foundations of Computational Mathematics 11(4), 417–487 (2011)","cited_arxiv_id":null,"evidence_quote":"Defines Gromov-Wasserstein distances, the main alignment baseline and runtime comparison in the experiments."},{"cited_title":"In: ICML (2016)","cited_arxiv_id":null,"evidence_quote":"Gives the computational complexity and entropic-regularization baselines for Gromov-Wasserstein that the paper compares against."},{"cited_title":"IEEE transactions on pattern analysis and machine intelligence 32, 1809–1821 (2010)","cited_arxiv_id":null,"evidence_quote":"Introduces kernelized sorting, the prior approach for photo album summarization that LSMI-Sinkhorn extends to rectangular layouts."}],"review_version":1}