{"id":"e340e6cf-f59a-460f-96f1-30632f222a09","arxiv_id":"2507.12246","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A new Phi-match framework generalizes Sinkhorn and semi-dual gradient ascent for entropic OT, with O(1/N) and O(1/N^2) rates for several variants, plus a path-space Schrodinger bridge extension.","lead":"This paper proposes a family of algorithms for entropic optimal transport, generalizing Sinkhorn with a flexible 'Phi-match' update and offering convergence rate proofs for several variants, including a momentum-accelerated method and a Schrodinger bridge analogue. The rates are the main draw, but one key continuous-space claim rests on a flawed kernel identity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The O(1/N) guarantee for plain SGA rests on an invalid identity-kernel identification in §4.1: the stated kernel gives m_k(ξ)=0 for every atomless ξ, so k-SGA with it is inert and Theorem 1 does not cover SGA.","rationale":"The reader's weakest assumption identifies the same load-bearing flaw. The strongest advertised application of Theorem 1 is the plain SGA update of Section 3.2, and the chain 'SGA = k-SGA with the identity kernel' is the only bridge from the MMD objective to SGA. That bridge is broken: the specified kernel is the Kronecker delta on an uncountable space, whose mean embedding kills all atomless measures, so the kernel is not characteristic and the corresponding k-SGA update is inert. The displayed equality m_k(ξ)(y)=ξ(y) in Section 4.1 is contradicted by the immediate calculation m_k(ξ)(y)=ξ({y})=0 for Lebesgue densities. This is not a stylistic gap but a false mathematical identification at the centre of the SGA claim. I do not see the rest of the paper as irreparably wrong: k-SGA with Gaussian or Laplace kernels, proj-SGA, and proj-SGA++ are separate contributions and Theorem 1 may hold for genuine bounded characteristic kernels. However, the paper's abstract and Section 4.1 explicitly promote the SGA consequence, and the current text does not support that advertised guarantee. There are also unresolved regularity questions such as the finiteness of diam(T_{ϕ0,yanc};L∞(Y)) in Theorem 2, but the identity-kernel defect is the more direct blocker for the headline rate. I therefore keep the reader's REJECT verdict; a revised version that removes the SGA-as-identity-kernel claim and states Theorem 1 only for genuine bounded kernels could be conditionally acceptable.","tokens_in":32308,"tokens_out":9489,"duration_ms":126852,"concrete_test":"Take Y=R, ν=N(0,1), and ξ=ν. For k_Id(y,y')=1{y=y'}, compute m_k(ξ)(y)=∫1{y=y'}ν(y')dy'=ν({y})=0 for every y, while ξ(y)=ν(y)>0. This directly refutes the equality asserted in §4.1. Then evaluate one update of k-SGA with this kernel: since m_k(ν)=0 and m_k(π(ϕ,ϕ+)_Y)=0 for any atomless ν and πY, the update gives ϕ_1=ϕ_0, whereas SGA would move by η(ν−πY). Hence Theorem 1 cannot cover SGA; to retain the SGA rate the authors must either prove it separately or exhibit a genuine bounded positive-definite kernel whose mean embedding equals the identity on all Lebesgue densities, which no such kernel can do.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 claims that for k_Id(y,y')=1 iff y=y', m_k_Id(ξ)(y)=ξ(y), so plain SGA is k-SGA and hence Theorem 1 gives a 1/N MMD rate. For continuous Y with densities with respect to Lebesgue measure this equality is false: m_k_Id(ξ)(y)=ξ({y})=0 for Lebesgue-a.e. y, since a singleton has measure zero. The kernel as written is the Kronecker-delta kernel on an uncountable set, not the Dirac-delta distribution needed for point evaluation of densities. It is not characteristic on P(Y), and the corresponding k-SGA update satisfies ϕ_{n+1}=ϕ_n whenever ν and π(ϕ,ϕ+)_Y are atomless, whereas the SGA update would be ϕ+η(ν−πY). Thus the assertion that SGA is the special case of k-SGA with the identity kernel fails exactly where the paper needs it. Theorem 1, whose hypotheses require a bounded positive-definite kernel, therefore cannot be specialised to SGA, and the sentence following Theorem 1 claiming that SGA inherits the 1/N rate is unsupported. The MMD-descent interpretation of SGA itself is also invalidated, because no bounded kernel on a continuous space has mean embedding equal to the identity on densities; mean embeddings are RKHS-valued and cannot reproduce arbitrary Lebesgue densities pointwise. k-SGA with Gaussian or Laplace kernels may remain valid, but the paper's advertised plain-SGA guarantee and its example of the identity kernel as a characteristic kernel do not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a unified optimisation perspective on entropic optimal transport (eOT), introducing a class of methods called Phi-match that interpolates between SGA (semi-dual gradient ascent) and Sinkhorn. It then derives several interpretations: alternating projections on the set Q of joint distributions with prescribed X-marginal, a local greedy update, and a mirror-descent viewpoint. The main theoretical results are a non-asymptotic O(1/N) guarantee for a kernelised variant k-SGA in squared maximum mean discrepancy (Theorem 1), O(1/N) and O(1/N^2) guarantees for a projected semi-dual gradient ascent and its accelerated version (Theorems 3 and 4), and a O(1/N) rate for a signed semi-dual ascent (Theorem 2). The framework is extended to the dynamical Schroedinger bridge problem, with a path-space analogue of Phi-match and an SDE-drift implementation. The paper is primarily theoretical, with no numerical experiments, and relies on imported lemmas from prior work on mirror descent and Sinkhorn.","tokens_in":32686,"tokens_out":10554,"duration_ms":126642,"significance":"If the main results were fully correct, the paper would make a useful contribution by providing non-asymptotic convergence guarantees for a family of eOT algorithms that do not require strict tail or log-concavity assumptions on the marginals. The k-SGA theorem and the proj-SGA/acceleration proofs follow standard templates and appear internally consistent. The framework productively unifies primal projections, local greedy updates, and mirror-descent views of Sinkhorn, and the extension to path-space Schr\\\"odinger bridges is conceptually attractive. The paper also gives clear attribution to prior work and does not rely on fitted constants or circular reasoning. However, a central advertised claim---that plain SGA inherits the O(1/N) MMD rate via an identity kernel---is based on a false identification in Section 4.1, and Theorem 2 contains an unproved finiteness assumption. These issues materially weaken the paper's headline contributions as currently stated.","major_comments":[{"comment":"The claim that the identity kernel k_Id(y,y') = 1 iff y = y' satisfies m_k(\\xi)(y) = \\xi(y) is false for atomless probability measures on a continuous space. For any Lebesgue-absolutely-continuous measure \\xi, the integral \\int k_Id(y,y') d\\xi(y') equals \\xi({y}) = 0 for Lebesgue-almost every y, not the density \\xi(y). The kernel as written is the Kronecker-delta kernel on an uncountable set, not the Dirac-delta distribution needed for point evaluation of densities. Consequently, the identity kernel is not characteristic on P(Y), the map V_Phi for SGA does not coincide with the first variation of L_{k_Id}, and plain SGA is not a special case of k-SGA. The sentence following Theorem 1 claiming that SGA inherits the 1/N rate is therefore unsupported. This is load-bearing because the advertised connection between SGA and MMD is the motivation for introducing k-SGA.","section":"Section 4.1"},{"comment":"The theorem states a rate depending on diam(T_{\\phi_0,y_{\\rm anc}}; L^\\infty(Y)), but it never proves that this diameter is finite. The set T is a superlevel set of the semi-dual J intersected with the anchoring condition \\phi(y_{\\rm anc})=0, and without additional assumptions nothing prevents J(\\phi) \\ge J(\\phi_0) from holding for functions of arbitrarily large L^\\infty norm. If the diameter is infinite, the displayed bound is vacuous and the claimed non-asymptotic rate for sign-SGA is not meaningful. The proof in Appendix A.3.1 uses diam in a Young-inequality step without verifying its finiteness. The theorem needs either a proof that T has finite L^\\infty diameter under the stated assumptions or an explicit assumption to that effect.","section":"Theorem 2"},{"comment":"The continuous-time acceleration result in Appendix B again uses the same invalid identity kernel k(y,y') = 1 iff y = y'. For atomless marginals, L_k(\\hat\\pi_t^Y, \\nu) is identically zero under this kernel, so the claimed O(1/t^2) decay does not establish convergence of the Y-marginal in any meaningful metric. This appendix should be revised to use a bounded characteristic kernel or be clearly marked as heuristic.","section":"Appendix B, Lemma 13"}],"minor_comments":[{"comment":"The two observations labelled 'Fact 1' and 'Fact 2' are unnumbered; numbering them would make cross-referencing in the proofs easier.","section":"Section 3.1.1"},{"comment":"The sentence listing 'the identity, Gaussian, and Laplace kernels' as examples of characteristic kernels should be corrected: the identity kernel is not characteristic on uncountable spaces, and its mean embedding does not act on densities in the claimed way.","section":"Section 4.1"},{"comment":"The phrase 'under minimal assumptions' is used prominently, but Theorem 1 requires a bounded kernel, Theorem 3 requires bounded cost with \\lambda(B)<\\infty, and Theorem 2 requires (at least implicitly) a finite L^\\infty diameter. The paper should qualify these statements carefully.","section":"Abstract and Introduction"},{"comment":"After Theorem 1, the paper says that for characteristic kernels '\\pi_n^Y approaches \\nu'. Strictly, the theorem gives convergence of L_k(\\pi_N^Y,\\nu) to zero; if the kernel is characteristic this implies convergence in the MMD metric, and the paper should spell out that this is the intended sense.","section":"Theorem 1"},{"comment":"The path-space extension is described with high-level statements; Proposition 3 and Lemma 7 rely on external results (Girsanov, h-transforms, time reversal) without a full derivation of all integrability conditions. A short remark on the required regularity would improve readability.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The identity-kernel error is serious and is repeated in Appendix B, but it is localisable: the k-SGA theorem with a genuine bounded characteristic kernel appears sound, and the proj-SGA/acceleration results are also standard conditional on their stated assumptions. A revision that removes the unsupported plain-SGA claim and either proves or explicitly assumes finiteness of the diameter in Theorem 2 could make the paper acceptable. The authors should also reconsider the 'minimal assumptions' framing, since several theorems now require explicit boundedness or finiteness conditions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: the paper has a sensible core idea and a serious flaw in its headline continuous-space claim. The Phi-match abstraction is a genuinely nice unifying view, with Sinkhorn and SGA as two instantiations, and the kernelised, projected, and accelerated variants are new and potentially useful. The k-SGA proof follows the standard relative-smoothness template and looks internally consistent; the projected and accelerated methods are standard adaptations and likely correct. The path-space extension of IPF is a real generalization and is clearly explained.\n\nThe problem is Section 4.1. The paper claims the identity kernel k_Id(y,y')=1 iff y=y' is characteristic and gives m_k(ξ)(y)=ξ(y), so SGA is a special case of k-SGA and inherits the 1/N MMD rate. For continuous Y this is false: the integral is zero for Lebesgue-a.e. y because singletons have measure zero. That kernel is not characteristic on atomless measures, and the mean embedding is not the identity on densities. So Theorem 1, which requires a bounded positive-definite kernel, cannot be specialised to SGA, and the sentence after Theorem 1 claiming SGA gets the 1/N rate is unsupported. This is not a cosmetic issue; it removes the advertised continuous-space guarantee for plain SGA.\n\nA second, softer issue: Theorem 2's rate depends on diam(T, L∞), and finiteness is never proved. With shift-invariance removed by anchoring, the superlevel set could still be unbounded; the proof needs an argument that the diameter is finite.\n\nWhat stands: k-SGA with Gaussian or Laplace kernels may well be correct, and the proj-SGA/proj-SGA++ rates likely hold. The framework is a genuine derivation against an external benchmark, with no fitted constants or self-citations. The paper is salvageable, but the central claim as stated is not supported.\n\nThis paper is for researchers working on entropic OT algorithms and mirror-descent interpretations. It deserves a serious referee, but it needs major revision before publication. I would send it out and ask the authors to fix the kernel identification, prove the diameter bound, and restate what is actually proved for SGA.","headline":"Interesting Phi-match framework undermined by a false identity-kernel identification in Section 4.1; the 1/N rate for plain SGA is unsupported, though k-SGA and the projected methods may survive.","tokens_in":33171,"tokens_out":2089,"would_cite":false,"duration_ms":24694,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper derives a unified family of Sinkhorn-like algorithms for entropic optimal transport from a semi-dual optimisation viewpoint, with non-asymptotic $1/N$ and $1/N^2$ rates under minimal assumptions.","keywords":["entropic optimal transport","Sinkhorn algorithm","semi-dual","maximum mean discrepancy","mirror descent","Schrödinger bridge","non-asymptotic convergence","accelerated optimization"],"falsifier":"On $\\mathbb{R}$ with $\\nu$ Lebesgue-absolutely continuous, compute $m_{k_{\\mathrm{Id}}}(\\nu)(y)=\\int \\mathbf{1}\\{y=y'\\}\\nu(y')\\,dy'$: the integrand is nonzero only on a set of measure zero, so the kernel mean is $0$ almost everywhere and the claimed equality $m_k(\\xi)=\\xi$ fails. A valid bounded positive-definite kernel returns a smoothed version of $\\xi$, so the $1/N$ bound for plain SGA depends on a missing evaluation kernel.","tokens_in":32124,"feed_emoji":"🧮","tokens_out":10931,"duration_ms":116921,"temperature":0.7,"pith_summary":"Entropic optimal transport is usually solved with a matrix-scaling algorithm whose classical non-asymptotic guarantees degrade exponentially as the regularisation parameter shrinks. This paper tries to establish that the problem can be redesigned from scratch as infinite-dimensional optimisation, maximising a semi-dual objective whose gradient is the mismatch between the Y-marginal of a structured coupling and the target distribution. From this viewpoint it defines a general update, $\\Phi$-match, that contains the Sinkhorn iteration and plain semi-dual gradient ascent as two natural instantiations, and it shows every iterate remains in the family of couplings that can be completed to the entropic plan. This yields non-asymptotic rates for the Y-marginal, and hence for the coupling, under no assumptions on the marginals: $1/N$ for MMD-based, signed, and projected variants, and $1/N^2$ for an accelerated projected variant. The same machinery transfers to path measures, giving drift-update algorithms for the dynamical Schr\\\"odinger bridge problem.","feed_headline":"New entropic-OT methods hit 1/N and 1/N^2 rates","feed_subtitle":"A unified optimisation view of Sinkhorn-style updates gives assumption-free convergence and extends to Schrödinger bridges.","key_machinery":"The central object is the $\\Phi$-match update together with the semi-dual $J$ of entropic optimal transport. The semi-dual is built from the map $\\phi_+(x)=\\log\\int_Y \\exp((\\phi(y)-c(x,y))/\\varepsilon)\\,d\\nu(y)$, and the joint density $\\pi(\\phi,\\phi_+)\\propto \\exp((\\phi(y)-\\phi_+(x)-c(x,y))/\\varepsilon)\\,\\mu\\otimes\\nu$; the operator $\\Phi$ selects which discrepancy between $\\pi_Y$ and $\\nu$ the method minimises. The machinery shows that $\\Phi$-match is simultaneously an alternating projection, a local greedy root-finding step, and a mirror-descent step, which is what makes non-asymptotic convergence proofs possible.","core_discovery":"The central discovery is that entropic-optimal-transport algorithms are best understood not as scaling procedures but as iterates of an optimisation method on the semi-dual $J(\\phi)=\\int_Y \\phi\\,d\\nu - \\int_X \\phi_+\\,d\\mu$, whose first variation is $\\delta J(\\phi)=\\nu - \\pi(\\phi,\\phi_+)_Y$. The proposed class $M_{\\Phi\\text{-match}}(\\phi;\\eta)=\\phi-\\eta(\\log\\Phi(\\pi(\\phi,\\phi_+)_Y)-\\log\\Phi(\\nu))$ keeps each joint distribution inside the structured set $Q$, so any method that brings the Y-marginal to $\\nu$ automatically produces the optimal coupling. For a bounded positive-definite kernel $k$, kernelised SGA satisfies $L_k(\\pi_N^Y;\\nu)\\le \\max\\{2c_k,1\\}\\,d_{\\mathrm{KL}}(\\pi^*\\|\\pi_0)/N$; under a cost-dependent smoothness condition, projected SGA and its accelerated version attain $O(1/N)$ and $O(1/N^2)$ semi-dual gaps. The same recursive update, lifted to path measures, gives a Schr\\\"odinger-bridge solver whose identity-operator special case is iterative proportional fitting.","pith_inferences":["Editorial extension: The cleanest domain for the $1/N$ theorem is kernelised SGA with a bounded positive-definite kernel; the paper's remarks on Gaussian and Laplace kernels suggest a particle implementation, and one could test whether finite-sample kernelised SGA beats Sinkhorn when $\\varepsilon$ is small and costs are ill-conditioned.","Editorial extension: Accelerated projected SGA demonstrates the benefit of staying in the dual; an implicit corollary is that primal momentum on the non-convex set $Q$ may control only marginals, not couplings, because convex combinations of iterates can leave $Q$.","Editorial extension: Replacing $\\Phi$ in path-$\\Phi$-match yields a one-parameter family of Schr\\\"odinger-bridge samplers; a natural experiment is to compare $\\Phi=\\exp$ against $\\Phi=\\mathrm{id}$ on Brownian bridges to see whether the MMD-style drift correction converges faster in practice."],"forward_implications":["For any bounded positive-definite kernel, kernelised SGA converges in squared MMD at rate $O(1/N)$: $L_k(\\pi_N^Y;\\nu)\\le \\max\\{2c_k,1\\}d_{\\mathrm{KL}}(\\pi^*\\|\\pi_0)/N$, and since each iterate lies in $Q$, this drives the full coupling toward the entropic optimal plan.","Signed SGA and projected SGA maximise the semi-dual at rate $O(1/N)$, while accelerated projected SGA reaches $O(1/N^2)$ using only a cost-dependent constant rather than assumptions on the marginals.","The rate's dependence on the regularisation parameter is polynomial rather than the exponential contraction factor $e^{-\\|c\\|_\\infty/\\varepsilon}$ seen in classical Sinkhorn analyses, so the methods target the small-regularisation regime.","Path-space $\\Phi$-match turns the same recursion into SDE drift updates, making the dynamical Schr\\\"odinger bridge problem solvable by the same optimisation template, with iterative proportional fitting as one parameter choice."],"supporting_citations":[{"why":"Introduces the semi-dual formulation and the discrete-space semi-dual gradient ascent update that this paper generalises.","marker":"Genevay et al. (2016)"},{"why":"Supplies the first-variation formula for the semi-dual and the initialisation bound $d_{\\mathrm{KL}}(\\pi^*\\|\\pi_0)\\le d_{\\mathrm{KL}}(\\pi^*\\|\\pi_{\\mathrm{ref}})$.","marker":"Léger (2021)"},{"why":"Provides the relative-smoothness inequality used as Proposition 1 and the primal root/projection interpretation of Sinkhorn that underlies the k-SGA rate.","marker":"Aubin-Frankowski et al. (2022)"},{"why":"Gives the $\\eta$-Sinkhorn mirror-descent form, the mirror map $\\delta\\varphi^*$, and the path-space IPF flow that $\\Phi$-match extends.","marker":"Reza Karimi et al. (2024)"},{"why":"Bounds the Schr\\\"odinger potentials in $L^\\infty$, fixing the ball radius $B$ used by projected SGA and its accelerated version.","marker":"Di Marino and Gerolin (2020)"},{"why":"Supplies the FISTA/ISTA accelerated proximal template that accelerated projected SGA adapts to the semi-dual.","marker":"Beck and Teboulle (2009)"},{"why":"Establishes the disintegration of the dynamical Schr\\\"odinger bridge into a static entropic plan plus a reference bridge, which motivates path-$\\Phi$-match.","marker":"Föllmer (1988)"},{"why":"Provides the non-asymptotic $O(1/N)$ rate for iterative proportional fitting that the paper's path-space results generalise.","marker":"Bernton et al. (2019)"}],"fun_headline_variants":["Entropic OT: optimization view gives 1/N and 1/N^2 rates","Optimization perspective yields faster entropic OT solvers","Entropic OT rates 1/N and 1/N^2 from optimization view","New entropic OT algorithms with optimal convergence rates","Unified optimization framework for entropic OT and bridges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the identity kernel acts as an evaluation map on densities ($m_{k_{\\mathrm{Id}}}(\\xi)=\\xi$), a property that holds for atomic measures but not for the continuous Lebesgue-density setting in which the paper states its guarantees.","fun_headline_variants_meta":{"raw":{"variants":["Entropic OT: optimization view gives 1/N and 1/N^2 rates","Optimization perspective yields faster entropic OT solvers","Entropic OT rates 1/N and 1/N^2 from optimization view","New entropic OT algorithms with optimal convergence rates","Unified optimization framework for entropic OT and bridges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2438,"prompt_tokens":939,"completion_tokens":1499,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":1411}},"tokens_in":555,"tokens_out":1499,"duration_ms":11600,"temperature":1.0,"reasoning_tokens":1411,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:51:05.712076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On $\\mathbb{R}$ with $\\nu$ Lebesgue-absolutely continuous, compute $m_{k_{\\mathrm{Id}}}(\\nu)(y)=\\int \\mathbf{1}\\{y=y'\\}\\nu(y')\\,dy'$: the integrand is nonzero only on a set of measure zero, so the kernel mean is $0$ almost everywhere and the claimed equality $m_k(\\xi)=\\xi$ fails. A valid bounded positive-definite kernel returns a smoothed version of $\\xi$, so the $1/N$ bound for plain SGA depends on a missing evaluation kernel.","supporting_citations":[{"cited_title":"Stochastic O ptimization for L arge-scale O ptimal T ransport","cited_arxiv_id":null,"evidence_quote":"Introduces the semi-dual formulation and the discrete-space semi-dual gradient ascent update that this paper generalises."},{"cited_title":"Mirror D escent with R elative S moothness in M easure S paces, with application to S inkhorn and EM","cited_arxiv_id":null,"evidence_quote":"Provides the relative-smoothness inequality used as Proposition 1 and the primal root/projection interpretation of Sinkhorn that underlies the k-SGA rate."},{"cited_title":"Sinkhorn Flow as Mirror Flow: A Continuous-Time Framework for Generalizing the S inkhorn Algorithm","cited_arxiv_id":null,"evidence_quote":"Gives the $\\eta$-Sinkhorn mirror-descent form, the mirror map $\\delta\\varphi^*$, and the path-space IPF flow that $\\Phi$-match extends."},{"cited_title":"An optimal transport approach for the S chr\\\"odinger bridge problem and convergence of S inkhorn algorithm","cited_arxiv_id":null,"evidence_quote":"Bounds the Schr\\\"odinger potentials in $L^\\infty$, fixing the ball radius $B$ used by projected SGA and its accelerated version."}],"review_version":1}