{"id":"8aa749a6-d141-4896-a19a-b450ebed7b30","arxiv_id":"2506.02712","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The authors prove partial-domain-adaptation bounds based on partial Wasserstein distances and derive source-sample weights from the optimal transport coupling.","lead":"This paper derives new generalization bounds for partial domain adaptation using partial optimal transport, and uses those bounds to justify a weighting scheme for source training examples. The work matters because it gives a theoretical basis for a widely used heuristic approach and shows small accuracy gains when combined with a state-of-the-art method.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"WARMPOT optimizes cross-entropy, but Theorems 3.2/3.3 require a bounded metric loss; the proofs' triangle inequalities fail for cross-entropy, so the bounds do not cover the deployed objective.","rationale":"I agree with the reader that the metric-loss assumption is the weakest load-bearing point. The theoretical results are internally consistent under their stated assumptions, and the partial-Wasserstein construction with coupling-derived weights is a genuine contribution. The gap is load-bearing because the algorithm and all experiments use cross-entropy: it fails the triangle inequality, is unbounded, and is not ζ-Lipschitz, so Lemma A.4, the reverse-triangle step in Appendix C, and the Hoeffding step in Lemma 3.4 cannot hold for WARMPOT's objective. Thus Eq. (19) is not a minimization of a proven upper bound, and the empirical weight ablations do not directly validate the theoretical claim. The paper's limitation paragraph acknowledges approximation of the bounds but omits this more fundamental loss mismatch. The CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":23905,"tokens_out":10328,"duration_ms":113003,"concrete_test":"Verify the proof gap analytically: set K=2, a=(0.01,0.99), b=(0.99,0.01), e0=(1,0), and compute CE(a,e0)=4.605 versus CE(a,b)+CE(b,e0)=4.570. The strict violation of the triangle inequality used in Lemma A.4 and Appendix C shows that the proofs of (5) and (9) cannot be instantiated with WARMPOT's cross-entropy loss; to close the gap, the paper would need either bounds for a bounded surrogate of cross-entropy or a metric loss in (19).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Both Theorem 3.2 and Theorem 3.3 assume ℓ is a metric on Y and bounded in [0,1]; Theorem 3.3 additionally assumes ζ-Lipschitzness. The proofs rely on the triangle and reverse-triangle inequalities of ℓ (Lemma A.4; Appendix C, Eq. (59)) and on the Hoeffding step in Lemma 3.4, which requires ℓ ∈ [0,1]. Section 5.2 sets ℓ to cross-entropy in the WARMPOT objective (19), both in the weighted source term and inside the partial-transport cost. Cross-entropy is not a metric: for K=2, a=(0.01,0.99), b=(0.99,0.01), e0=(1,0), CE(a,e0)=4.605 > CE(a,b)+CE(b,e0)=4.570. It is also unbounded above and not Lipschitz in either argument near zero probabilities. Consequently, none of the theorem statements upper-bounds the quantity WARMPOT minimizes; the claimed theoretical motivation for the weighted cross-entropy loss and the exact PW cost in (19) is not established. This does not invalidate the theorems under their stated assumptions, but it leaves a gap between theory and the deployed algorithm, one that the paper's own limitation paragraph (which only discusses the cost of exact bound minimization) does not acknowledge.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives two families of generalization bounds for partial domain adaptation (PDA) based on partial optimal transport. For a fixed hypothesis w = g∘f, Theorem 3.2 bounds the empirical target loss by a weighted empirical source loss, a partial Wasserstein distance between empirical source and target feature distributions, a total-variation term, and a non-computable difficulty term. Theorem 3.3 gives an analogous bound using joint feature-label distributions, with the target labels replaced by predictions under w. In both bounds, the source weights pi and target weights qj are the row and column sums of the optimal partial-transport coupling matrix. Lemma 3.4 converts these fixed-hypothesis bounds into PAC-Bayes generalization bounds (Corollaries 3.5 and 3.6). Based on these results, the authors propose the WARMPOT algorithm that minimizes a weighted cross-entropy source loss plus a partial Wasserstein alignment term computed with cross-entropy in the transport cost, and they report experiments on Office-Home and ImageNet-to-Caltech showing that WARMPOT is competitive with recent methods and that replacing the ARPM weights with the WARMPOT weights improves ARPM's average accuracy.","tokens_in":24149,"tokens_out":6803,"duration_ms":68514,"significance":"If the bounds are valid, they provide a genuinely new theoretical foundation for two widely used ingredients in PDA algorithms: the partial Wasserstein distance as a domain-alignment term and non-uniform weighting of source samples. A notable strength is that the Wasserstein term in the bounds is expressed directly through empirical measures, avoiding the additional concentration-of-measure step present in earlier OT-based domain-adaptation bounds. The weights arise constructively from the optimal coupling rather than being imposed heuristically, and the authors ship an open-source implementation. However, the theoretical results are stated under a metric loss bounded in [0,1], while the deployed objective uses cross-entropy, which is neither a metric nor bounded. This gap means that the theorems, as stated, do not cover the actual quantity minimized by WARMPOT. The paper also lacks a clear validation protocol for its tuned hyperparameters. For these reasons the significance of the results is real but conditional on closing the theory-practice gap.","major_comments":[{"comment":"There is a fundamental mismatch between the loss covered by the theory and the loss used in the algorithm. Theorems 3.2 and 3.3 assume that the loss function ℓ is a metric on Y and bounded in [0,1]; Theorem 3.3 further requires ζ-Lipschitzness in each argument. Section 5.2 states that WARMPOT sets ℓ to the cross-entropy loss in the objective (19), both in the weighted source term and inside the partial-transport cost. Cross-entropy is not a metric (it is not symmetric and does not satisfy the triangle inequality; it is also unbounded above on the probability simplex). Consequently, the proofs in Appendix B and C, which rely on the triangle and reverse-triangle inequalities (Lemma A.4, Eq. (26); Eq. (59)), and the Hoeffding step in Lemma 3.4 (which requires ℓ∈[0,1]), do not apply to the objective actually minimized. The paper's limitation paragraph in Section 6 acknowledges only the computational difficulty of exact bound minimization, not this loss mismatch. The authors should either extend the theorems to a class of losses that includes cross-entropy (with boundedness and smoothness assumptions appropriate for softmax outputs) or modify WARMPOT to employ a loss satisfying the metric and boundedness assumptions, and then discuss the implications for the theoretical motivation of the algorithm.","section":"Section 5.2; Appendix E"},{"comment":"The hyperparameters (αmax, β, η1, η2, ε) are said to be obtained 'through a parameter search', but the manuscript does not specify whether this search was performed on a separate validation split or directly on the test tasks. Without a clear validation protocol, the reader cannot assess whether the reported competitive results are affected by test-set feedback. Please state explicitly the split used for hyperparameter selection and, if any tuning was performed on the test tasks, provide an evaluation in which hyperparameters are selected without access to test labels, or report the sensitivity of the conclusions to this choice.","section":"Appendix E"}],"minor_comments":[{"comment":"The comparison of weighting schemes is potentially confounded by different weight-update frequencies: WARMPOT updates weights every mini-batch, while BA3US and ARPM update them only every 500 iterations on the full dataset. A control experiment with matched update schedules would strengthen the claim that the weighting scheme itself, rather than the update frequency, drives the observed improvement.","section":"Table 1"},{"comment":"The claim that ARPM+our-weights achieves state-of-the-art performance rests on a 0.3 percentage-point average improvement over ARPM (82.0 vs 81.7 with standard deviations of 0.4 and 0.7). A paired significance test or confidence intervals over the common seeds would make the superiority claim more robust.","section":"Table 2"},{"comment":"The definition of L̂f in (13) with Ξ in (63) is difficult to parse because it mixes minimizations over g′ with the coupling-dependent weights. A short intuitive explanation of why this term is non-computable and what it represents in terms of the difficulty of the PDA problem would improve readability.","section":"Theorem 3.3"},{"comment":"There are several typographical errors, e.g., 'Wassserstein' in Section 1 ('Our proof techniques ... Wassserstein distance') and the inconsistent spelling of 'WARMPOT' in Table captions. A careful proofreading pass is recommended.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"The central derivation is internally consistent: the proofs in Appendices B–D appear correct under the stated metric-loss assumptions, and the empirical-measure formulation of the Wasserstein term is a genuine improvement over earlier bounds. The load-bearing issue is the theory-practice gap: the theorems require a bounded metric loss, while the algorithm minimizes cross-entropy, which is outside the scope of the theorems. This is fixable in principle—either by restricting the algorithm to a loss covered by the theory (e.g., a bounded metric on the label simplex with appropriate normalization) or by extending the proofs to cross-entropy under specific assumptions on the softmax outputs—but it is not a cosmetic issue. I also recommend asking the authors to disclose the hyperparameter validation split, because the reported state-of-the-art claim depends on it. If the authors can bridge the loss gap and clarify the validation protocol, the paper would make a solid contribution to the PDA literature; in its current form, the theoretical support for the deployed objective is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The theorem side of this paper is the real contribution, and the proofs hold up under their stated assumptions. Unlike earlier OT-based DA bounds (Shen et al. 2018, Courty et al. 2017a), the partial Wasserstein term here is defined on empirical feature distributions, so it directly motivates the feature-alignment practice used in most modern algorithms. The source weights derived from the optimal coupling are new, and the ablation gives them credible support: replacing ARPM's weights with these buys a small but consistent gain on Office-Home and ImageNet→Caltech, and the code is public.\n\nThe soft spots are in the bridge from theory to the deployed algorithm. Theorems 3.2 and 3.3 require ℓ to be a bounded metric, and 3.3 also ζ-Lipschitz in each argument. The WARMPOT objective (19) uses cross-entropy, which is neither bounded nor a metric. The stress-test note gives a concrete counterexample to the triangle inequality for CE, and the proof of Lemma A.4 relies on the triangle and reverse-triangle inequalities throughout. So the bounds do not upper-bound the actual quantity WARMPOT minimizes. The paper's limitation paragraph acknowledges the approximation cost of exact bound minimization but not this mismatch. That is an overclaim, and it needs to be stated rather than implied.\n\nTwo smaller issues: the reported gains of the proposed weights over BA3US-style weights are within one standard deviation on several tasks, and the SOTA gain from ARPM+our-weights is an average improvement of 0.3 points. The hyperparameters α,β,η1,η2,ε are selected by search, but there is no described validation protocol. The sensitivity analysis suggests performance is not brittle, yet the tuning procedure is underspecified.\n\nNone of this invalidates the theorems. The theory is a genuine step forward for the PDA subfield, and the coupling-derived weights are a useful idea that could outlive this particular algorithm. The paper deserves a serious referee. I would recommend acceptance after a major revision that either restricts the claims to metric losses or extends the analysis to the surrogate actually used (for instance, bounding CE by a metric loss with label dimension scaling, or using a bounded Lipschitz surrogate such as the hinge loss). The referee should also ask for a clear validation protocol for the hyperparameters.","headline":"Real theoretical contribution on partial Wasserstein alignment, but the paper overclaims the link to its cross-entropy objective.","tokens_in":24745,"tokens_out":2879,"would_cite":true,"duration_ms":27717,"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":"Partial optimal transport gives partial domain adaptation a theorem-backed way to weight source samples and align features.","keywords":["partial domain adaptation","partial optimal transport","partial Wasserstein distance","generalization bounds","negative transfer","source reweighting","PAC-Bayes","unsupervised domain adaptation"],"falsifier":"On synthetic data with known target labels and a metric label loss, compute the empirical target loss and the right-hand side of the bound in Theorem 3.2 for many fixed $\\gamma$-Lipschitz classifiers and feature maps; any violation of the inequality would refute the theorem. For the algorithmic claim, train the same model with cross-entropy and with a metric label loss; if the metric-loss version does not match or improve on cross-entropy, the experimental success is not explained by the paper's assumptions.","tokens_in":23672,"feed_emoji":"🎯","tokens_out":13990,"duration_ms":123140,"temperature":0.7,"pith_summary":"The paper sets out to show that partial domain adaptation—learning from a source domain whose label set contains the target label set—has a first-principles justification in partial optimal transport. It proves upper bounds on the target loss in which the alignment term is a partial Wasserstein distance and the source losses are weighted by the row sums of the optimal transport coupling, so the weights are derived quantities rather than heuristics. A second, joint feature-label version of the bound, followed by a PAC-Bayes argument that adds a KL-divergence term, lifts the result to high-probability generalization bounds. If the argument is right, two common practices, aligning feature distributions and reweighting source samples to reduce negative transfer, become components of a provable bound with an explicit recipe for the weights. The paper also presents WARMPOT as a practical minimizer of these terms, noting that exact minimization is prohibitively expensive and that its implementation relies on entropic mini-batch partial transport.","feed_headline":"Source weights for partial domain adaptation come from transport plan","feed_subtitle":"New bounds make the weighted source loss and the alignment term consequences of one optimal transport plan.","key_machinery":"The engine is the partial Wasserstein distance $\\mathrm{PW}_\\alpha(1/\\beta P_s^f, Q_t^f)$: an optimal transport problem that moves only $\\alpha$ units of mass under the marginal constraints $\\Pi 1_{n_t} \\le P_s^f/\\beta$ and $\\Pi^T 1_{n_s} \\le Q_t^f$. Its optimal coupling $\\Pi^\\star$ simultaneously supplies the alignment cost and, through its row and column sums, the source and target weights $p_i, q_j$ that reweight the empirical losses. Two auxiliary devices carry the proof: a change-of-measure inequality that converts the unevenly weighted target loss back to the uniform empirical target loss at the price of a total-variation term, and a difficulty term $L_f$ (or $\\hat L_f$) that measures the best achievable loss on the combined labeled data. A PAC-Bayes lemma then lifts the fixed-sample inequality to a high-probability bound on the population target loss, adding a KL-divergence term and a variance term of order $1/n_t$.","core_discovery":"The paper's central claim is that, with a fixed feature extractor $f$ and any classifier $g$ that is $\\gamma$-Lipschitz with respect to a metric label loss $\\ell$, the empirical target loss obeys $$L_{\\tilde z}(w) \\le \\sum_{i=1}^{n_s} \\frac{p_i}{\\$\\alpha$} \\ell(w(x_i), y_i) + \\frac{2}{\\$\\alpha$} \\mathrm{PW}_\\$\\alpha$\\left(\\frac{1}{\\$\\beta$} P_s^f, Q_t^f\\right) + \\frac{1}{2}\\sum_{j=1}^{n_t} \\left|\\frac{1}{n_t} - \\frac{q_j}{\\$\\alpha$}\\right| + 2L_f,$$ where $p_i$ and $q_j$ are the row and column sums of the optimal coupling matrix $\\Pi^\\star$ of the partial Wasserstein problem and $L_f$ is a non-computable term measuring problem difficulty. A second bound replaces the feature-only distributions by joint feature-label distributions and adds a label cost to the transport metric. The paper proves both bounds and converts them into PAC-Bayes generalization bounds, emphasizing that the partial Wasserstein term is computed directly on empirical feature distributions, so the bound explicitly depends on the learned feature map and motivates feature alignment.","pith_inferences":["The theorems assume a metric label loss, while the experiments train with cross-entropy, which is neither a metric nor Lipschitz in the label argument; whether the bounds extend to such losses is unsettled, and training with a metric label cost would test how much theory covers the deployed objective.","The per-sample weights double as outlier scores: a source point with near-zero $p_i$ is one the transport plan is willing to discard, so the same coupling could drive sample selection, open-set adaptation, or robust training.","Because the bound depends explicitly on the feature map $f$, it could serve as a model-selection or training criterion for feature extractors, not only as a justification of alignment; the paper does not develop this use.","Appendix H shows that with $\\alpha=1$ the WARMPOT weights solve a Wasserstein-1 problem with the box constraint $\\hat p_i \\le 1/(\\beta n_s)$, in contrast to ARPM's variance constraint, suggesting box-constrained transport as a general template for reweighting schemes in other domain adaptation settings."],"forward_implications":["The weights in a weighted source loss are no longer free: the optimal partial transport coupling prescribes them, so any PDA method using a weighted source term has a principled default.","The partial Wasserstein distance with its two budget parameters is supported as a domain alignment term, with $\\beta$ controlling the inflated source mass and $\\alpha$ controlling how much target mass participates.","Because the bound is evaluated on empirical feature distributions, it explicitly depends on the learned feature map, motivating training the feature extractor and not just the classifier.","The coupling automatically suppresses negative transfer: in the paper's P→A experiment, outlier source classes receive only about 6% of the source weight mass.","When the theoretically derived weights are used inside the ARPM algorithm, the paper reports the highest average Office-Home accuracy among the compared methods (82.0%), suggesting the weight recipe transfers beyond the paper's own objective."],"supporting_citations":[{"why":"It defines the optimal partial transport problem whose optimal coupling supplies both the alignment cost and the weights.","marker":"Figalli (2010)"},{"why":"It provides the partial Wasserstein distance definition used in the paper's bounds.","marker":"Caffarelli & McCann (2010)"},{"why":"It is the feature-based Wasserstein bound that Theorem 3.2 extends to the partial and weighted setting.","marker":"Shen et al. (2018)"},{"why":"It contributes the joint distribution optimal transport bound and the feature-label cost used in Theorem 3.3.","marker":"Courty et al. (2017a)"},{"why":"It gives the change-of-measure lemma that produces the total-variation term in the bounds.","marker":"Ohnishi & Honorio (2021)"},{"why":"It supplies the max-min inequality used to define the non-computable difficulty term in the loss-change lemma.","marker":"Boyd & Vandenberghe (2004)"},{"why":"It provides the PAC-Bayes template that Lemma 3.4 follows and that yields the generalization bounds.","marker":"Alquier (2024)"},{"why":"It supplies the Hoeffding and Chernoff inequalities used inside the PAC-Bayes proof.","marker":"Wainwright (2019)"},{"why":"It supplies the mini-batch partial transport solver that WARMPOT uses to make the alignment term computable.","marker":"Nguyen et al. (2022)"},{"why":"It provides the ARPM baseline whose weighting scheme is replaced by the paper's weights in the improved comparison.","marker":"Gu et al. (2024)"}],"fun_headline_variants":["Optimal transport plan sets source weights for domain adaptation","Partial transport plan yields source weights and theoretical bounds","Source weights come from an optimal transport plan","How to weight source data: follow the transport plan"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the label loss is a true distance between labels (symmetric and satisfying the triangle inequality) and, for the joint version, Lipschitz in each argument; the algorithm and experiments instead use cross-entropy, which is neither, and the paper does not close that gap.","fun_headline_variants_meta":{"raw":{"variants":["Optimal transport plan sets source weights for domain adaptation","Partial transport plan yields source weights and theoretical bounds","Source weights come from an optimal transport plan","How to weight source data: follow the transport plan"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1574,"prompt_tokens":989,"completion_tokens":585,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":525}},"tokens_in":605,"tokens_out":585,"duration_ms":5610,"temperature":1.0,"reasoning_tokens":525,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:18:44.796997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On synthetic data with known target labels and a metric label loss, compute the empirical target loss and the right-hand side of the bound in Theorem 3.2 for many fixed $\\gamma$-Lipschitz classifiers and feature maps; any violation of the inequality would refute the theorem. For the algorithmic claim, train the same model with cross-entropy and with a metric label loss; if the metric-loss version does not match or improve on cross-entropy, the experimental success is not explained by the paper's assumptions.","supporting_citations":[{"cited_title":"The optimal partial transport problem","cited_arxiv_id":null,"evidence_quote":"It defines the optimal partial transport problem whose optimal coupling supplies both the alignment cost and the weights."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It provides the partial Wasserstein distance definition used in the paper's bounds."},{"cited_title":"Wasserstein distance guided representation learning for domain adaptation","cited_arxiv_id":null,"evidence_quote":"It is the feature-based Wasserstein bound that Theorem 3.2 extends to the partial and weighted setting."},{"cited_title":"and Honorio, J","cited_arxiv_id":null,"evidence_quote":"It gives the change-of-measure lemma that produces the total-variation term in the bounds."},{"cited_title":"and Vandenberghe, L","cited_arxiv_id":null,"evidence_quote":"It supplies the max-min inequality used to define the non-computable difficulty term in the loss-change lemma."},{"cited_title":"User-friendly introduction to PAC-Bayes bounds","cited_arxiv_id":null,"evidence_quote":"It provides the PAC-Bayes template that Lemma 3.4 follows and that yields the generalization bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the Hoeffding and Chernoff inequalities used inside the PAC-Bayes proof."},{"cited_title":"Improving mini-batch optimal transport via partial transportation","cited_arxiv_id":null,"evidence_quote":"It supplies the mini-batch partial transport solver that WARMPOT uses to make the alignment term computable."}],"review_version":1}