{"id":"68208895-2ee5-4ec9-97e5-c77d61a9ec8b","arxiv_id":"2411.13969","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A fibered Wasserstein gradient flow for infinitely many incompressible phases is shown to admit weak solutions, to be stable as the phase measure is discretized, and to converge to the unique entropic optimal transport plan.","lead":"This paper proves existence, stability, and long-time convergence for a continuum of incompressible phases moving under a pointwise volume constraint, viewed as a joint Wasserstein gradient flow. It also connects the discrete-to-continuous limit of the phase distribution to entropic optimal transport.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stability Lemma 13's varying-marginal compactness step is asserted but not proven; since Lemma 16 inherits it, the central convergence claim rests on an unverified estimate.","rationale":"I read the paper in good faith: the central program is coherent, the minimizing-movement construction is detailed, and the claimed entropy-regularized relaxation of AHT is well motivated. The strongest claim, Lemma 16, is that every weak solution tends to the unique minimizer of E, and its proof is structurally reasonable: dissipation is summable, shifted trajectories converge by Lemma 13, and the limiting stationary state is identified via Lemma 15. The weak point I find most load-bearing is not the finite-Fisher-information assumption, which is a legitimate scope restriction even if it excludes piecewise-constant porosity; rather it is the unproved compactness step inside Lemma 13, the stability theorem under varying marginals. That theorem is the bridge used both for the discrete-to-continuous limit and for the long-time asymptotics. The proof states that the varying-marginal case can be handled as in Lemma 10 without carrying out the argument. Since the vertical averages are defined against measures ν_n that only converge weak*, the compactness and the identification of the limit are genuinely non-trivial. A failure here would break the product convergence that is the technical heart of the paper. The reader's conditional verdict already reflects general caution; my concern is a different, more central point, so agreement is partial. The proposed check—filling in the missing Aubin-Lions estimate or constructing a counterexample—would settle whether the concern lands. Until then, I do not recommend changing the conditional verdict.","tokens_in":31993,"tokens_out":27187,"duration_ms":252636,"concrete_test":"Complete the missing step in Lemma 13: prove that the Aubin-Lions time-equicontinuity bound for ω_n is uniform in n when the averages are taken against the varying measure ν_n, and that the limit in (95) is correct. If the estimate fails, exhibit a sequence (µ_n, ν_n, ρ_n, Π_n) satisfying the stated uniform horizontal Fisher bounds and weakly converging marginals for which the sequence ω_n has no strongly convergent subsequence in L2([0,T]×X); such a counterexample would invalidate Eq. (96) and hence Lemma 16.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing link is Lemma 13, the stability of weak solutions under joint variation of the marginals. Its proof reduces the crucial strong convergence of the vertical averages ω_n → ω in L2([0,T]×X) to a single sentence: \"The main difference is that now the marginals µn and νn vary along the converging sequence, but this can be factored into the proof without major issues.\" The averages are taken against ν_n, which only converges weak*; the Aubin-Lions compactness used in Lemma 10 for a fixed ν does not automatically transfer to this setting. If this compactness fails, the product convergence ω_n∇Π_n ⇀ ω∇Π in Eq. (96) fails, and with it both the discrete-to-continuous limit and Lemma 16's asymptotic convergence. This is an omitted proof in a central lemma, not merely a restrictive hypothesis. Independent minor issues include an incorrect inequality in Lemma 12 (WF ≤ E0(t2−t1) should involve the square) and forward references in Lemma 11, but these do not invalidate the argument by themselves.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the stratified drift-diffusion system (1)-(3), in which a continuum of phases y ∈ Y evolves under a joint pointwise volume constraint determined by a fixed first marginal μ. The dynamics are interpreted as a gradient flow in the fibered Wasserstein metric for the entropic optimal transport energy (4). The main results are: existence of global weak solutions via the JKO/minimizing-movement scheme (Section 4, Lemma 10); stability of weak solutions under joint variation of the marginals, including a discrete-to-continuous limit (Section 5, Lemma 13); and long-time convergence of every weak solution to the unique minimizer of the energy, i.e. the entropic optimal transport plan between μ and ν (Section 6, Lemma 16). The paper also shows that in the unregularized limit κ = 0, stationary suboptimal configurations persist, so regularization is essential for unconditional convergence, and it presents numerical experiments illustrating convergence, stability, and a bottleneck test with discontinuous μ.","tokens_in":32237,"tokens_out":5457,"duration_ms":60628,"significance":"If the main claims hold, this is a valuable contribution: it gives the first unconditional convergence guarantee for an entropic relaxation of the Angenent-Haker-Tannenbaum scheme and provides a rigorous variational framework for multiphase porous media flows with a continuum of phases. The paper is largely self-contained, the chain of estimates is detailed, and the energy functional is not fitted to the conclusions: the minimizer structure is cited from an independent source and the entropy parameter κ, time step τ, and discretization parameters are natural model parameters. The numerical section is qualitative and uses small academic examples, but it does support the theoretical statements and even probes regimes outside the stated assumptions. The main weakness is that Lemma 13, which is load-bearing for the stability and asymptotic-convergence results, contains an asserted but not proved compactness step for the varying-marginal case; this needs to be repaired before the central claims can be regarded as fully established.","major_comments":[{"comment":"The proof of the strong convergence of the vertical averages in (95) and (96) is not supplied: after stating that the argument is 'similar' to Lemma 10, the text says that the variation of the marginals μn and νn 'can be factored into the proof without major issues', but no details are given. This is a load-bearing point: Lemma 16 uses Lemma 13 to pass to the limit in the cluster-point argument for long-time convergence, and Lemma 11 and Lemma 14 rely on the product convergence (96) for lower semicontinuity of the dissipation. In particular, the averages ωn are taken against νn, which is only known to converge weak*, so the Aubin-Lions compactness used for fixed ν in Lemma 10 does not transfer automatically. Please provide a complete proof of (95) and (96), or state and prove the needed variant of Aubin-Lions for varying reference measures.","section":"Section 5, Lemma 13"},{"comment":"The statement (83) reads W2(ρt1,ρt2) ≤ W2_F(ρt1,ρt2) ≤ E(ρ0)(t2-t1), but the proof at the end of Lemma 12 derives W_F^2 ≤ (t2-t1) E(ρ0), so the correct inequality is W_F ≤ sqrt(E(ρ0)(t2-t1)). The printed bound is false in general and also has the wrong scaling in both E(ρ0) and time. Please correct the statement and check all places that use Lemma 12 for equicontinuity; the corrected square-root bound is sufficient for that purpose.","section":"Lemma 12, Eq. (83)"},{"comment":"The standing assumption (15) that μ is bounded away from 0 and ∞ and has finite Fisher information excludes piecewise-constant porosities, which are a standard case in porous media modeling. The numerical bottleneck experiment in Section 8 uses exactly such a discontinuous μ and says that the observed convergence 'may indicate that the Fisher information bound hypothesis could be further relaxed', but no proof is offered. This is a limitation of scope rather than an internal inconsistency, but it should be stated explicitly: the numerical experiment is outside the hypotheses of the theorems, and if a relaxation is intended, it should be proved or explicitly labeled as a conjecture.","section":"Section 1.5 and Section 8, Eq. (116)"}],"minor_comments":[{"comment":"The sentence after (72) says 'weak* convergence of ωτ dx dt to ωτ dx dt'; the second factor should be ω dx dt. Also verify the displayed equation (72) itself, which appears to have a typo in the final equality.","section":"Lemma 10, Eq. (72)"},{"comment":"The proof refers to 'the regularity properties outlined in Lemma 12' and to Lemma 13 before those lemmas are stated. Please reorder the presentation or add explicit forward references so the logical dependence is clear.","section":"Lemma 11"},{"comment":"There are several typographical issues, e.g. 'Radon-Nykodym' for 'Radon-Nikodym', 'defing' for 'defining', and 'sufficiently' for 'sufficiently'. These do not affect the mathematics but should be corrected.","section":"Throughout"},{"comment":"The flipped-initialization experiment with N → ∞ is described as having initial entropy tending to infinity, so it is outside the assumptions of the existence/stability theorems; this should be stated clearly in the text, not only implicitly.","section":"Section 8"}],"recommendation":"major_revision","confidential_remarks":"The core ideas are interesting and the results, if correct, would merit publication. My recommendation rests on the incomplete proof of Lemma 13, which is used in the central stability and asymptotic-convergence theorems. If the authors can supply a complete proof of (95)-(96) for varying marginals (or adjust the statements to cover only the cases in which the current argument is complete), I would be willing to support acceptance after a further round. The paper is within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does something genuinely new: it pushes the multiphase porous-media gradient flow from finitely many phases to an arbitrary continuum ν, and proves existence, stability under ν-convergence, and convergence as t → ∞ to the unique entropic optimal transport plan. The conceptual frame—fibered Wasserstein metric plus JKO scheme—is clean, and the paper honestly connects the model to the AHT scheme, even showing that the unregularized plan relaxation still has suboptimal stationary points. That is real content, not packaging.\n\nWhat the paper does well: the proofs are mostly detailed and self-contained, the chain from discrete minimizing movements to weak solutions is careful, and the stability result is the right tool for the discrete-to-continuous limit. No fitted parameters, no circularity beyond citing an established result for E-minimizers. The numerics are qualitative but they genuinely illustrate the claims, including the bottleneck case that lies outside the stated assumptions—the authors flag that themselves, which is honest.\n\nThe soft spots are real but not fatal. The one that matters is Lemma 13. The proof of strong convergence of the vertical averages ω_n → ω when the marginals ν_n vary is hand-waved: \"this can be factored into the proof without major issues.\" That is not a minor line-item; Lemma 16 and the product-convergence step (96) depend on it. The Aubin-Lions framework used in Lemma 10 is set up for a fixed reference measure, and transporting it to varying ν_n requires an argument, not a sentence. I suspect it is fixable using the uniform Fisher-information bounds and the strong convergence of µ_n, but as written it is a genuine gap in a central lemma. The referee should demand the missing details.\n\nSmaller issues: Lemma 12 states WF(ρ_t1, ρ_t2) ≤ E(ρ0)(t2−t1), but the proof gives WF² ≤ (t2−t1)E(ρ0); dimensionally it should have a square root. Not load-bearing, but it should be corrected. The assumption that µ has finite Fisher information excludes piecewise-constant porosity, which is standard in porous media and which the authors themselves simulate; they are upfront about it, but the numerical evidence suggests the theorem might extend further, and that extension is left open.\n\nWho this is for: anyone working on Wasserstein gradient flows, multiphase porous media, or entropic optimal transport algorithms. It deserves a serious referee. I would send it out, with instructions to focus on Lemma 13 and to check whether the compactness claim can be made rigorous without extra assumptions.","headline":"A substantial and mostly rigorous paper that proves existence, stability, and long-time convergence for a continuum of coupled Wasserstein gradient flows; the main caveat is a skipped compactness argument in the stability lemma that deserves a careful referee.","tokens_in":32756,"tokens_out":1979,"would_cite":true,"duration_ms":23213,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35K51","35A15","35A35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A continuum of coupled Wasserstein gradient flows gives the entropic relaxation of the AHT scheme an unconditional convergence guarantee that the original algorithm lacks.","keywords":["multiphase porous media flow","fibered Wasserstein gradient flow","minimizing movement scheme","Angenent–Haker–Tannenbaum scheme","entropic optimal transport","weak solutions","long-time asymptotics","discrete-to-continuous limit"],"falsifier":"Take $X=Y=[0,1]$, $\\mu=\\nu$ uniform, $V(x,y)=|x-y|^2/2$, and $\\kappa>0$; run the minimizing-movement scheme at decreasing $\\tau$ from the flipped initial plan $\\rho_0=(\\mathrm{id},1-\\mathrm{id})_\\sharp\\mu$. The theorem predicts $E(\\rho_t)-E(\\rho^*) \\to 0$ along the whole trajectory; observing that the trajectory approaches any stationary point other than the entropic plan $\\rho^*$, or that the energy gap plateaus above the discretization error as $\\tau \\to 0$, would falsify Lemma 16.","tokens_in":31808,"feed_emoji":"♾️","tokens_out":8139,"duration_ms":73177,"temperature":0.7,"pith_summary":"The paper studies a system of drift-diffusion equations for a potentially infinite number of incompressible phases occupying a porous domain, each phase labelled by $y \\in Y$ and all sharing a fixed volume-occupancy profile $\\mu(x)$. It interprets this system as a gradient flow in the space of couplings between $\\mu$ and $\\nu$, measured with a fibered Wasserstein distance that transports mass separately inside every $y$-fiber. The main results are existence of weak solutions by a minimizing-movement construction, stability as the phase distribution $\\nu$ approaches a continuum, and long-time convergence: every weak solution converges to the unique minimizer of the energy, which is exactly the objective of entropy-regularized optimal transport. Since that energy is the entropic version of the Angenent–Haker–Tannenbaum (AHT) scheme's cost, the paper establishes that this relaxation is unconditionally convergent, in contrast to the classical AHT scheme, which can stall at suboptimal stationary maps.","feed_headline":"Entropic variant of optimal transport flow now converges globally","feed_subtitle":"The relaxed scheme reaches the unique optimal plan where the classical algorithm gets stuck","key_machinery":"The carrying object is the fibered Wasserstein distance $W_F$, which for couplings with the same second marginal $\\nu$ disintegrates each coupling into fiber measures on $X$ and integrates their squared $W_2$ distances over $Y$. The construction proceeds by a minimizing-movement (JKO-type) scheme with steps $E_\\tau(\\rho\\mid\\bar\\rho)=\\frac{1}{2\\tau}W_F(\\rho,\\bar\\rho)^2+E(\\rho)$; each step is equivalent to a convex program whose optimality conditions yield the discrete pressure $\\Pi^n$ and the transport relation $\\rho^{n-1}_y = (\\mathrm{id}+\\tau\\nabla_X(V-\\Pi^n+\\kappa\\log r^n))_\\sharp\\,\\rho^n_y$. The pressure $\\Pi$ is determined by the elliptic equation $\\nabla\\cdot(\\mu\\nabla\\Pi)=\\kappa\\Delta\\mu+\\nabla\\cdot(\\int_Y \\nabla_X V\\, r\\,d\\nu)$, acting as the Lagrange multiplier for the volume constraint. The technical bottleneck is that both density and pressure converge only weakly, so the product $\\rho\\nabla\\Pi$ is handled through strong $L^2$ convergence of vertical averages of the density, which gives the joint convergence of the product needed in the weak formulation. Long-time convergence is closed by showing that vanishing dissipation $I(\\rho,\\Pi)=\\int|\\nabla_X(V-\\Pi+\\kappa\\log r)|^2\\,d\\rho$ forces $\\rho$ to be the unique diagonal-scaling minimizer of $E$.","core_discovery":"The central claim is that the coupled system (1)–(3) is a well-posed gradient flow whose longtime limit solves entropic optimal transport. More precisely, for fixed marginals $\\mu$ and $\\nu$, positive regularization $\\kappa$, and a finite-energy initial coupling, the minimizing-movement sequence built with the fibered Wasserstein distance $W_F$ yields a weak solution $(\\rho_t, \\Pi_t)$ of the coupled PDEs; a key step is showing that vertical averages of the density converge strongly enough to pass the product of pressure gradient and density to the limit, jointly over infinitely many phases. Any such weak solution converges as $t \\to \\infty$ to the unique minimizer of $E(\\rho)=\\int V\\,d\\rho+\\kappa H(\\rho)$, namely the entropic optimal transport plan between $\\mu$ and $\\nu$. The paper also shows that this convergence property is stable under approximation of the second marginal by finitely many phases, recovering multiphase porous-media models in the discrete-to-continuous limit. In the unregularized limit $\\kappa=0$, by contrast, suboptimal stationary configurations persist, so relaxation alone does not fix the AHT scheme; some regularization is necessary.","pith_inferences":["The numerically observed linear decay of $E(\\rho_t)-E(\\rho^*)$ suggests a quantitative dissipation inequality; proving one would turn the asymptotic guarantee into a complexity bound for entropy-regularized optimal transport computed by minimizing movements.","The bottleneck experiment with a discontinuous porosity $\\mu$ indicates that the Fisher-information hypothesis may be relaxable to piecewise-smooth densities; a proof would extend the results to layered or fractured porous media.","Because the invariant measure of the flow is the entropic optimal transport plan, the dynamics can be read as an annealing of a Schrödinger bridge; this may connect the convergence proof to quantitative rates for iterative proportional fitting methods such as Sinkhorn's algorithm.","The same fibered-gradient-flow mechanism may apply to multi-marginal or dynamic transport problems where AHT-style projection algorithms stall, as long as a fiber metric with strong vertical-average compactness can be defined."],"forward_implications":["Every weak solution of (1)–(3) converges as $t \\to \\infty$ to the unique entropic optimal transport plan between $\\mu$ and $\\nu$, so the entropic relaxation of the AHT scheme is unconditionally convergent.","The continuum model is the discrete-to-continuous limit of finite multiphase porous-media flows: trajectories for $\\nu$ made of finitely many Dirac masses converge to those for a general probability measure $\\nu$.","The minimizing-movement scheme gives a constructive existence proof and a numerical algorithm; the experiments indicate linear convergence of the relative energy gap, opening a route to novel entropic optimal transport solvers on geometric domains.","For $\\kappa=0$, suboptimal stationary configurations exist, so merely relaxing AHT to the space of plans without regularization does not guarantee convergence."],"supporting_citations":[{"why":"Introduces the AHT scheme whose lack of convergence guarantees motivates the entropic relaxation studied in this paper.","marker":"[2]"},{"why":"Supplies the fibered Wasserstein metric and its topology, used to define the minimizing-movement scheme and to prove convergence of trajectories.","marker":"[32]"},{"why":"Provides the gradient-flow and compactness machinery (refined Ascoli–Arzelà and lower semicontinuity) used to extract limit trajectories.","marker":"[1]"},{"why":"Source of the JKO and dual-potential method used to derive first-order optimality conditions for each discrete step.","marker":"[35]"},{"why":"Gives the structure and uniqueness of the minimizers of the entropic transport energy, used in Lemmas 4 and 15.","marker":"[6]"},{"why":"Supplies the cluster-point argument used in Lemma 16 to prove long-time convergence to the minimizer.","marker":"[28]"},{"why":"Primal-dual algorithm used in the numerical simulations of the minimizing-movement scheme.","marker":"[16]"},{"why":"Multiphase porous-media model whose discrete-to-continuous limit is identified with the continuum dynamics in Remark 5.","marker":"[11]"}],"fun_headline_variants":["Coupled Wasserstein flows converge to entropic plan","Entropic relaxation yields unconditionally convergent flow","Infinite-phase flows relax to unique entropic optimum","Relaxed AHT scheme converges unconditionally"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The porosity profile $\\mu$ must be bounded away from zero and infinity and have finite Fisher information $\\int_X |\\nabla\\sqrt{\\mu}|^2\\,dx < \\infty$; this regularity is what gives the horizontal $H^1$ control on the density and $L^2$ control on the pressure gradient, so the proof does not cover discontinuous porosities such as the piecewise-constant bottleneck tested numerically.","fun_headline_variants_meta":{"raw":{"variants":["Coupled Wasserstein flows converge to entropic plan","Entropic relaxation yields unconditionally convergent flow","Infinite-phase flows relax to unique entropic optimum","Relaxed AHT scheme converges unconditionally"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001145,"raw_usage":{"total_tokens":4752,"prompt_tokens":952,"completion_tokens":3800,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":3740}},"tokens_in":568,"tokens_out":3800,"duration_ms":27930,"temperature":1.0,"reasoning_tokens":3740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:41:02.382519+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $X=Y=[0,1]$, $\\mu=\\nu$ uniform, $V(x,y)=|x-y|^2/2$, and $\\kappa>0$; run the minimizing-movement scheme at decreasing $\\tau$ from the flipped initial plan $\\rho_0=(\\mathrm{id},1-\\mathrm{id})_\\sharp\\mu$. The theorem predicts $E(\\rho_t)-E(\\rho^*) \\to 0$ along the whole trajectory; observing that the trajectory approaches any stationary point other than the entropic plan $\\rho^*$, or that the energy gap plateaus above the discretization error as $\\tau \\to 0$, would falsify Lemma 16.","supporting_citations":[{"cited_title":"Minimizing flows for the Monge-Kantorovich problem","cited_arxiv_id":null,"evidence_quote":"Introduces the AHT scheme whose lack of convergence guarantees motivates the entropic relaxation studied in this paper."},{"cited_title":"Heterogeneous gradient flows in the topology of fibered optimal transport","cited_arxiv_id":null,"evidence_quote":"Supplies the fibered Wasserstein metric and its topology, used to define the minimizing-movement scheme and to prove convergence of trajectories."},{"cited_title":"Lectures in Mathematics ETH Z¨ urich","cited_arxiv_id":null,"evidence_quote":"Provides the gradient-flow and compactness machinery (refined Ascoli–Arzelà and lower semicontinuity) used to extract limit trajectories."},{"cited_title":"Optimal Transport for Applied Mathematicians , volume 87 of Progress in Nonlinear Differential Equations and Their Applications","cited_arxiv_id":null,"evidence_quote":"Source of the JKO and dual-potential method used to derive first-order optimality conditions for each discrete step."},{"cited_title":"Domain decomposition for entropy regularized optimal transport","cited_arxiv_id":null,"evidence_quote":"Gives the structure and uniqueness of the minimizers of the entropic transport energy, used in Lemmas 4 and 15."},{"cited_title":"Self-similarity in a thin film Muskat problem.SIAM Journal on Mathematical Analysis, 49(4):2790–2842, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the cluster-point argument used in Lemma 16 to prove long-time convergence to the minimizer."},{"cited_title":"A first-order primal-dual algorithm for convex problems with applications to imaging","cited_arxiv_id":null,"evidence_quote":"Primal-dual algorithm used in the numerical simulations of the minimizing-movement scheme."},{"cited_title":"Canc` es, T","cited_arxiv_id":null,"evidence_quote":"Multiphase porous-media model whose discrete-to-continuous limit is identified with the continuum dynamics in Remark 5."}],"review_version":1}