{"id":"7a13c26c-ebbe-4c73-9f37-ce71a60dab93","arxiv_id":"2505.12097","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper defines and analyzes D^c_epsilon(P||Q) = inf_R { T_c(P,R) + epsilon D(R||Q) }, a general infimal-convolution divergence with duality, dynamic mean-field-game formulation, and explicit Gaussian examples.","lead":"Researchers introduce the proximal optimal transport divergence, a way to compare probability distributions that blends the cost of transporting one distribution into another with a statistical divergence penalty. The framework unifies several known divergence families and connects to mean-field games and modern generative modeling algorithms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix C's proof of dual maximizer existence truncates in the wrong direction: f_n^{(l)}=max{f_n,l} makes the truncated objective smaller, so inequality (C.43) has the wrong sign and the monotone-convergence argument for the central optimizer is not valid as written.","rationale":"The reader's weakest assumption (Assumption A-D) is actually less fragile than stated: for KL, the level sets {KL(·||Q)≤α} are tight by a tail-entropy estimate, and KL(·||Q) is strictly convex; so the existence/uniqueness Theorem 2.4 has a sound basis for the KL case. The load-bearing issue is instead the proof of the dual maximizer used in the first-variation theorem. Theorem C.6 contains a sign error in the truncation: using max{f,l} decreases the truncated potentials, so inequality (C.43) is reversed. This makes the existence assertion unproven as written. Since Theorem 4.1 and the ICNN-discriminator algorithms rely on φ*, this is the single most load-bearing concern. The repair (replace max by min) appears straightforward and the theorem is likely correct, so I do not move the verdict: conditional acceptance with a request to repair the appendix is appropriate.","tokens_in":35223,"tokens_out":27777,"duration_ms":301162,"concrete_test":"Re-derive Theorem C.6 using truncation f^{(l)} = min{f_n,l} and verify the inequalities: f^{(l)} ≤ f_n, φ_n^{(l)} ≥ φ_n, ψ_n^{(l)} ≥ ψ_n, so the truncated objective is ≥ the original; then check that the decreasing-in-l pointwise limits (or a diagonal subsequence) satisfy the constraint φ⊕ψ≤c and attain the supremum. If this repair succeeds, the proof gap is a typographical error; if it fails (e.g., because the truncated ψ_n^{(l)} lose the constraint φ+ψ≤c), then Theorems 3.3(c) and 4.1 need a different proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Reader's weakest assumption (Assumption A-D) is not the main issue: for KL the level sets {KL(·||Q)≤α} are tight by a tail-entropy estimate and KL(·||Q) is strictly convex, so Theorem 2.4 has a sound basis in the KL case. The central risk sits in Theorem 3.3(c) and Theorem 4.1, whose proof in Appendix C (Theorem C.6) is internally inconsistent. With f_n = \\tilde a - φ_n ≥ 0 and g_n = \\tilde b - ψ_n ≥ 0, the manuscript defines f_n^{(l)} = max{f_n,l} and g_n^{(l)} = max{g_n,l}. Since max ≥ f_n, the truncated potentials φ_n^{(l)} = \\tilde a - f_n^{(l)} and ψ_n^{(l)} = \\tilde b - g_n^{(l)} are pointwise ≤ φ_n and ≤ ψ_n. The dual objective is monotone increasing in (φ,ψ), so (C.43) should be ≤, not ≥. Thus the claimed upper bound that allows taking weak limits is unjustified. The intended construction is almost certainly f^{(l)} = min{f,l}; with that replacement the monotone-convergence step would go through, so the theorem is likely repairable. But as written, the existence of a maximizer (φ*,ψ*) — required for the first-variation formula and for the ICNN-discriminator algorithms — is not established.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a family of discrepancies between probability measures, called proximal optimal transport divergences, defined by infimal convolution of an optimal transport cost T_c(P,R) with an information divergence D(R||Q), controlled by a parameter ε. The main theoretical contributions are: elementary properties (finiteness, convexity, lower semicontinuity, uniqueness of the proximal minimizer), interpolation limits as ε→0 and ε→∞, a dual variational representation for the KL case, existence and characterization of dual optimizers, a data-processing inequality, additivity for product measures, a first-variation formula for the divergence, a dynamic Benamou–Brenier / mean-field-game formulation, and closed-form Gaussian examples. The paper also connects the construction to generative modeling algorithms, gradient flows, and input-convex neural network parameterizations.","tokens_in":35428,"tokens_out":6545,"duration_ms":68521,"significance":"If the main results stand, the paper provides a useful unifying framework that interpolates between information divergences and optimal transport, with potential applications in generative modeling, distributionally robust optimization, and gradient flows. The explicit dual representation (Theorem 3.1) and the transport-and-reweighting interpretation of the proximal minimizer are attractive and likely to be used by practitioners. The paper is also commendable for including detailed proofs of the central duality theorem, explicit Gaussian computations, and concrete algorithmic reformulations. However, the proof of dual-maximizer existence contains a sign/truncation error that currently undermines Theorem 3.3(c) and Theorem 4.1; this is repairable but requires a substantive revision of Appendix C.","major_comments":[{"comment":"The truncation argument proving existence of a dual maximizer is invalid as written. With f_n = \\tilde a - φ_n ≥ 0 and g_n = \\tilde b - ψ_n ≥ 0, the manuscript defines f_n^{(l)} = max{f_n,l} and g_n^{(l)} = max{g_n,l}. Since max{f_n,l} ≥ f_n, the truncated potentials φ_n^{(l)} = \\tilde a - f_n^{(l)} and ψ_n^{(l)} = \\tilde b - g_n^{(l)} are pointwise ≤ φ_n and ≤ ψ_n. The dual objective is monotone increasing in (φ,ψ), so inequality (C.43) has the wrong sign: the truncated objective is a lower bound, not an upper bound, for the original objective. Consequently the weak-convergence and monotone-convergence steps leading to (C.47) do not establish the existence of (φ*,ψ*). The intended construction is almost certainly f_n^{(l)} = min{f_n,l} (and similarly for g_n), and with that replacement the argument likely goes through; but as written, Theorem 3.3(c) and the first-variation formula in Theorem 4.1, which both rely on this maximizer, are not proven.","section":"Appendix C, Theorem C.6, Eqs. (C.39)–(C.43)"},{"comment":"The proof of Theorem 4.1 uses the boundedness of the cost c at a load-bearing point, saying “Since c is bounded, arguing as in Theorem C.4,” but Theorem 4.1 assumes only that c is uniformly continuous. Uniformly continuous costs on noncompact Polish spaces need not be bounded (for example c(x,y)=|x−y| on R), and in that case the asserted equiboundedness of the extended potentials \\hat φ_α and ψ_α is not justified. The theorem should either add a boundedness hypothesis or provide a localization argument that controls the potentials without global boundedness.","section":"Appendix E, proof of Theorem 4.1"},{"comment":"The data-processing inequality is stated as a theorem but its proof is omitted: Appendix D says only that it “follows by a similar argument and is therefore omitted.” This is not a proof of the stated result. The inequality involves the averaged cost Kc and the kernel-induced measures K#P and K#Q, and it is not immediately evident that the dual representation interacts with the kernel in the claimed way. A complete proof, or an explicit statement that this is a conjecture/observation, is needed before this result can be used as a theorem of the paper.","section":"Section 3.3, Theorem 3.6"},{"comment":"The general results in Section 2 are formulated for arbitrary information divergences under Assumption A-D, which requires the level sets {R : D(R||Q) ≤ α} to be precompact and strictly convex. The paper states that these assumptions are satisfied by KL, but for common f-divergences (for example α-divergences with α>1, or total variation) strict convexity or weak precompactness of level sets can fail without additional moment or compactness control. Since Theorem 2.4, the uniqueness of the proximal minimizer R*, and Definition 2.5 are stated in this claimed generality, the scope of the theory is broader than is justified. The main KL-based claims are unaffected, but the framework should either restrict the generality or verify Assumption A-D for each divergence family claimed.","section":"Section 2.2, Assumption A-D and Theorem 2.4"}],"minor_comments":[{"comment":"In the display defining ψ_n^{(l)}, the right-hand side reads \\tilde b - f_n^{(l)}; it should be \\tilde b - g_n^{(l)}.","section":"Appendix C, Eq. (C.42)"},{"comment":"The formula for m_R contains the fragment “µ_2+” which appears to be a typographical artifact; please clarify the intended expression.","section":"Appendix F, Theorem F.2, Eq. (F.11)"},{"comment":"The notation \\tilde φ and \\hat φ is introduced in the text but the displayed equations are not fully consistent about which objects are convex and which are Legendre transforms; please standardize the notation.","section":"Section 5.2, Eq. (5.5)–(5.6)"},{"comment":"The dual class Φ_c is defined with bounded continuous functions, while Theorem 3.3(c) states that the maximizer lies in L1(P)×L1(Q). The relationship between the two classes should be stated explicitly, since bounded and L1 potentials are not the same set for unbounded costs.","section":"Section 3.1, Eq. (3.1) and Theorem 3.3(c)"}],"recommendation":"major_revision","confidential_remarks":"The central duality formula (Theorem 3.1) appears to be proven through Theorem C.5, and the Gaussian examples are self-consistent, so the paper is not beyond repair. However, the truncation error in Theorem C.6 is exactly in the argument that supplies the dual maximizer used by Theorem 3.3(d) and Theorem 4.1; this is a load-bearing gap, not a typo. I would send back for major revision with the clear request to fix the truncation direction, clarify the boundedness hypotheses in Theorem 4.1, and provide a real proof of the data-processing inequality. If the authors repair these points, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nTwo things you should know about arXiv:2505.12097. The core object — the proximal OT divergence D^c_ε(P||Q) = inf_R {T_c(P,R) + ε D(R||Q)} — is a genuine extension of the (f,Γ)-divergence line of work, and the duality theorem (Thm 3.1) is the right kind of result: it covers general lower semicontinuous costs and the proof goes through standard approximation arguments. The paper also supplies interpolation limits, proximal minimizer existence, a first-variation formula, a dynamic mean-field-game interpretation, and explicit Gaussian computations. Those are solid contributions and the framework genuinely unifies a lot of existing material (1-Wasserstein divergences, Wasserstein-regularized KL, JKO steps, OT-flows).\n\nThe soft spot is in the proof of general dual-maximizer existence (Thm 3.3(c), Appendix C, Thm C.6). The truncation f_n^(l) = max{f_n,l} makes the truncated potentials pointwise smaller than the original ones, so the objective inequality (C.43) goes the wrong way and the monotone-convergence argument fails. The intended construction is min{f_n,l}; with that replacement the step likely goes through, but as written the existence of (ϕ*,ψ*) in the L1 setting is not established. This matters because Theorem 4.1 and the ICNN-based algorithms rely on having an optimizer. The main duality equality does not depend on this proof, so the central claim survives.\n\nThe reader's other concerns are legitimate but minor: the DPI (Thm 3.6) is asserted without proof, the rescaled dynamic formula (2.20) has a scaling mismatch between the continuity equation and the 1/T^2 factor — a presentational bug rather than a mathematical error — and the relation to Moreau-Yosida f-divergences [53] is left vague. None of these contradict the main results. The assumption A-D is fine for the KL case by standard tightness of KL sublevel sets, so the proximal minimizer results are on solid ground.\n\nThis paper deserves a serious referee. The framework is worth taking seriously; the existence proof needs repair and the minor issues need cleanup. I'd recommend major revision, not desk reject.","headline":"Proximal OT divergences is a useful unification with a correct duality theorem, but the proof of general dual-maximizer existence has a truncation error that is repairable.","tokens_in":36079,"tokens_out":13376,"would_cite":true,"duration_ms":118059,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","49J40","62B10","68T07","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The proximal optimal transport divergence interpolates between optimal transport distances and information divergences through an infimal convolution, and the paper proves its dual variational form and a transport proximal operator.","keywords":["proximal optimal transport divergence","infimal convolution","Kullback-Leibler divergence","Wasserstein distance","variational duality","mean-field games","gradient flows","generative modeling"],"falsifier":"Take $P=\\delta_0$, $Q=\\mathcal{N}(0,1)$, $\\varepsilon=1$, and $c(x,y)=|x-y|^2$, and let $\\rho=\\delta_1-\\delta_0$; compute the one-sided derivative $\\lim_{\\alpha\\downarrow 0}[D^c_{\\mathrm{KL},1}(P+\\alpha\\rho\\parallel Q)-D^c_{\\mathrm{KL},1}(P\\parallel Q)]/\\alpha$ and compare it with $\\int\\hat{\\phi}^*\\,d\\rho$ using the dual optimizer $\\hat{\\phi}^*$. A disagreement would refute Theorem 4.1; a simpler boundary check is that with total variation in place of KL and $P=\\delta_0,Q=\\delta_1$, the proximal minimizer is non-unique, so Theorem 2.4 cannot hold without the strict-convexity assumption.","tokens_in":34945,"feed_emoji":"📐","tokens_out":8215,"duration_ms":84812,"temperature":0.7,"pith_summary":"The paper introduces the proximal optimal transport divergence, a single discrepancy measure built as an infimal convolution: to compare $P$ with $Q$, you first pay an optimal transport cost to move $P$ onto an intermediate measure $R$, then pay an information divergence to reweight $R$ toward $Q$. It claims this construction is a genuine divergence that stays finite even when $P$ and $Q$ have disjoint supports, interpolates between the transport distance and the divergence as $\\varepsilon$ varies, admits a variational dual usable with convex neural networks, and has a dynamic formulation as a first-order mean-field game. If the paper is right, transport proximals already used in generative modeling, distributionally robust optimization, and gradient flows are all instances of one unified object, and the new dual forms give them a common algorithmic backbone.","feed_headline":"One divergence bridges optimal transport and information divergence","feed_subtitle":"Proximal OT divergences stay finite even when supports are disjoint and give generative flows one shared foundation.","key_machinery":"The load-bearing object is the transport proximal operator $\\operatorname{prox}^c_{\\varepsilon D(\\cdot\\parallel Q)}(P)=\\arg\\min_R\\{T_c(P,R)+\\varepsilon D(R\\parallel Q)\\}$, whose unique output $R^*$ is the intermediate measure. The proof machinery is convex duality: the paper swaps the infimum over $R$ and the supremum in the Kantorovich dual of $T_c$, then applies the Gibbs variational principle $-\\varepsilon\\log\\mathbb{E}_Q[e^{-\\psi/\\varepsilon}]=\\inf_R\\{\\mathbb{E}_R[\\psi]+\\varepsilon D_{\\mathrm{KL}}(R\\parallel Q)\\}$, producing the dual representation (3.8). Everything downstream, including the neural network algorithm, the first variation, and the particle gradient flow, runs on that representation.","core_discovery":"The central claim is that $D^c_\\varepsilon(P\\parallel Q)=\\inf_{R}\\{T_c(P,R)+\\varepsilon D(R\\parallel Q)\\}$ is itself a divergence, and that for KL divergence it satisfies the duality $D^c_{\\mathrm{KL},\\varepsilon}(P\\parallel Q)=\\sup_{(\\phi,\\psi)\\in\\Phi_c}\\{\\mathbb{E}_P[\\phi]-\\varepsilon\\log\\mathbb{E}_Q[e^{-\\psi/\\varepsilon}]\\}$ with $\\Phi_c=\\{(\\phi,\\psi):\\phi(x)+\\psi(y)\\le c(x,y)\\}$. From this dual identity the paper derives the existence and uniqueness of the intermediate minimizer $R^*$, the Gibbs form $dR^*/dQ=e^{-\\psi^*/\\varepsilon}/\\mathbb{E}_Q[e^{-\\psi^*/\\varepsilon}]$, and the first-variation formula $\\delta D^c_{\\mathrm{KL},\\varepsilon}/\\delta P=\\hat{\\phi}^*$, the $c$-transform of the optimal dual potential. These results are what turn the divergence into a computable objective and a gradient flow.","pith_inferences":["Because the dual proof uses only the Gibbs variational principle and Kantorovich duality, the same construction should extend to any divergence with a Donsker-Varadhan-style representation, such as R\\'enyi or $\\alpha$-divergences; the paper states the extension to $f$-divergences but proves the KL case in full.","The additivity of the divergence for product measures suggests a tensorized computational strategy in high dimensions: evaluate one-dimensional proximal divergences per coordinate block and sum them, which would make the objective practical for factorized variational inference.","Interpreting $R^*$ as the output of a transport map followed by an exact Gibbs reweighting suggests a testable regularization for generative flows: enforce that the learned pushforward ends at a Gibbs density relative to the target, rather than only matching samples."],"forward_implications":["Because $D^c_\\varepsilon(P\\parallel Q)\\le \\min\\{T_c(P,Q),\\varepsilon D(P\\parallel Q)\\}$, the divergence is finite in cases where KL diverges, so it can compare singular or manifold-supported distributions without a density ratio.","The dual form makes the divergence a concave maximization over transport potentials, so it can be evaluated by convex programming for discrete measures and by input-convex neural networks in continuous spaces.","As $\\varepsilon\\to 0$ the rescaled divergence $(1/\\varepsilon)D^c_\\varepsilon$ increases to $D(P\\parallel Q)$, while as $\\varepsilon\\to\\infty$ it increases to $T_c(P,Q)$; thus $\\varepsilon$ is a genuine dial between reweighting and transport.","The dynamic formulation is a first-order mean-field game with a backward Hamilton-Jacobi equation coupled to a forward continuity equation, giving a flow-based picture of the proximal minimizer as the terminal distribution of an optimal control problem.","The first variation exists along singular perturbations and equals the $c$-transform of the optimal potential, which yields gradient descent algorithms in probability space and particle generative schemes."],"supporting_citations":[{"why":"supplies the $(f,\\Gamma)$-divergence construction that the proximal divergence generalizes to arbitrary transport costs.","marker":"[7]"},{"why":"provides a divergence that compares measures without absolute continuity and establishes the infimal-convolution pattern extended here.","marker":"[20]"},{"why":"gives the Kantorovich duality and $c$-transform results used in the proof of the variational representation.","marker":"[47]"},{"why":"supplies the Donsker-Varadhan variational formula for KL used to pass from the primal to the dual.","marker":"[19]"},{"why":"gives the input-convex neural network parameterization used to compute the dual divergence.","marker":"[33]"},{"why":"defines the OT-flow generative model that the paper re-derives as a proximal OT divergence.","marker":"[40]"},{"why":"supplies the Benamou-Brenier dynamic formulation extended to the mean-field-game representation.","marker":"[5]"},{"why":"defines mean-field games, the framework used to interpret the dynamic formulation.","marker":"[29]"},{"why":"introduces the JKO scheme, which the paper identifies as an instance of the transport proximal operator.","marker":"[28]"}],"fun_headline_variants":["OT and info divergences now one framework","Proximal OT divergence unifies transport and information","Infimal convolution bridges OT and info divergences","New OT divergence flows from a mean-field game","Interpolate smoothly between OT and info divergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the information divergence $D$ has weakly precompact, strictly convex sublevel sets (Assumption A-D), which secures existence and uniqueness of the intermediate measure $R^*$; for common divergences beyond KL this can fail.","fun_headline_variants_meta":{"raw":{"variants":["OT and info divergences now one framework","Proximal OT divergence unifies transport and information","Infimal convolution bridges OT and info divergences","New OT divergence flows from a mean-field game","Interpolate smoothly between OT and info divergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000621,"raw_usage":{"total_tokens":2878,"prompt_tokens":947,"completion_tokens":1931,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":1860}},"tokens_in":563,"tokens_out":1931,"duration_ms":14042,"temperature":1.0,"reasoning_tokens":1860,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:42:29.676238+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $P=\\delta_0$, $Q=\\mathcal{N}(0,1)$, $\\varepsilon=1$, and $c(x,y)=|x-y|^2$, and let $\\rho=\\delta_1-\\delta_0$; compute the one-sided derivative $\\lim_{\\alpha\\downarrow 0}[D^c_{\\mathrm{KL},1}(P+\\alpha\\rho\\parallel Q)-D^c_{\\mathrm{KL},1}(P\\parallel Q)]/\\alpha$ and compare it with $\\int\\hat{\\phi}^*\\,d\\rho$ using the dual optimizer $\\hat{\\phi}^*$. A disagreement would refute Theorem 4.1; a simpler boundary check is that with total variation in place of KL and $P=\\delta_0,Q=\\delta_1$, the proximal minimizer is non-unique, so Theorem 2.4 cannot hold without the strict-convexity assumption.","supporting_citations":[{"cited_title":"(f-Γ)-divergences: Interpolating between f-divergences and integral probability metrics","cited_arxiv_id":null,"evidence_quote":"supplies the $(f,\\Gamma)$-divergence construction that the proximal divergence generalizes to arbitrary transport costs."},{"cited_title":"Formulation and properties of a divergence used to com- pare probability measures without absolute continuity","cited_arxiv_id":null,"evidence_quote":"provides a divergence that compares measures without absolute continuity and establishes the infimal-convolution pattern extended here."},{"cited_title":"Optimal transport for applied mathematicians , volume 87 of Progress in Nonlinear Differential Equations and their Applications","cited_arxiv_id":null,"evidence_quote":"gives the Kantorovich duality and $c$-transform results used in the proof of the variational representation."},{"cited_title":"Dupuis and R.S","cited_arxiv_id":null,"evidence_quote":"supplies the Donsker-Varadhan variational formula for KL used to pass from the primal to the dual."},{"cited_title":"Optimal transport map- ping via input convex neural networks","cited_arxiv_id":null,"evidence_quote":"gives the input-convex neural network parameterization used to compute the dual divergence."},{"cited_title":"OT-flow: Fast and accurate continuous normalizing flows via optimal transport","cited_arxiv_id":null,"evidence_quote":"defines the OT-flow generative model that the paper re-derives as a proximal OT divergence."},{"cited_title":"Mean field games","cited_arxiv_id":null,"evidence_quote":"defines mean-field games, the framework used to interpret the dynamic formulation."},{"cited_title":"The variational formulation of the fokker–planck equation","cited_arxiv_id":null,"evidence_quote":"introduces the JKO scheme, which the paper identifies as an instance of the transport proximal operator."}],"review_version":1}