{"id":"d1343901-d5f4-42b0-9783-2355498029ba","arxiv_id":"2411.09848","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Adding a Sobolev (H1) regularization to quantile-function gradient flows guarantees existence of generalized minimizing movements for positive distance-kernel MMD and rectifies a mass-dissipation defect for the negative kernel.","lead":"This paper adds a mathematical smoothing term to a type of probability-flow equation, making a previously ill-behaved version well-defined. The regularization fixes a known 'mass dissipation' flaw and may extend to higher dimensions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3.6's monotonicity-preservation proof contains an unjustified strict inequality; the identification of the regularized negative-kernel flow with the unconstrained Neumann problem, and hence the numerical dissipation-of-mass demonstration, is not fully established.","rationale":"The reader's weakest_assumption identifies the H1 support restriction as the main limitation, but that is an explicitly stated condition of the theorems and does not affect their internal validity. My stress-test instead examined the proof machinery behind the paper's applied claim that Sobolev regularization rectifies mass dissipation. The negative-kernel existence theorem (3.4) is credible because it uses the full convex functional with the indicator; the positive-kernel theorem (3.10) is credible because the compact Sobolev embedding and the ψ1−ψ2 chain-rule condition are verified. The vulnerable point is Proposition 3.6, which is used to justify dropping the indicator and solving the unconstrained Neumann problem (16). Its proof compares the upper endpoint of the multivalued inclusion at s1 with the lower endpoint at s2, but the multivalued term 2τ[−R^+_ν(y),−R^-_ν(y)] moves in the opposite direction when y decreases. Without an additional small-τ or bounded-density condition, the strict inequality is not guaranteed, so the monotonicity preservation g(t)∈C(0,1) is unproven. Since Algorithm 1 and all Section 4 numerical demonstrations are built on (25), the central practical conclusion is not fully supported as written. I therefore recommend conditional acceptance: the existence theorems may stand, but Proposition 3.6 must be repaired (for example by adding a hypothesis such as τ·sup ρ_ν < 1/2 in the discrete step) or the numerical claims must be reframed as conjectural. My concrete test directly checks whether the invariance actually fails in a simple case, which would settle whether the concern lands.","tokens_in":20755,"tokens_out":20209,"duration_ms":205272,"concrete_test":"Perform one step of the implicit Euler scheme (25) with ν=U[0,1], Qν(s)=s, λ=1, τ=10, and initial gn(s)=0.5, which lies in C(0,1)∩H1(0,1). Solve the boundary-value problem (26) numerically and check whether the computed gn+1 is nondecreasing on (0,1). If gn+1 is non-monotone, Proposition 3.6 is false as stated. As an analytic cross-check, choose a smooth decreasing test function on a subinterval satisfying the Neumann data g'(0)=g'(1)=1 and directly verify whether the middle strict inequality in the proof holds for s1=0.4, s2=0.6; a failure for any admissible test function settles the gap.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central existence theorems 3.4 and 3.10 appear sound: they rely on convex lsc semigroup theory and the Rossi–Savaré compactness/chain-rule machinery. However, the paper's main applied claim—that Sobolev regularization removes the dissipation-of-mass defect—depends on Proposition 3.6, which drops the indicator IC(0,1) and reduces the flow to the Neumann problem (16) with Algorithm 1 solving (25). The proof of g(t)∈C(0,1) in Proposition 3.6 hinges on a strict inequality between s1<s2 in an interval where g'_{n+1}<0: it asserts that the upper endpoint at s1 is strictly below the lower endpoint at s2, i.e. that gn(s1)-gn+1(s1)-2τR^-_ν(gn+1(s1)) < gn(s2)-gn+1(s2)-2τR^+_ν(gn+1(s2)). The increase of the first two terms is essentially gn+1(s1)-gn+1(s2), while the R-terms can decrease by 2τ times the target mass or density in the interval (gn+1(s2), gn+1(s1)]. For ν=U[0,1] with density 1, the net comparison is (1-2τρ)(gn+1(s1)-gn+1(s2)), which is negative once 2τρ>1. Thus for sufficiently large τ or high target density the asserted strict increase can fail, so the claimed contradiction that -g'' is simultaneously increasing and decreasing is not established. Consequently, the invariance g(t)∈C(0,1), the equivalence of (14) and (16), and the validity of the numerical scheme (25) for the regularized flow rest on an unproven smallness condition. Theorems 3.4 and 3.10 themselves are not undermined because they work with the full convex functional including IC(0,1).","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies Wasserstein gradient flows on P2(R) of MMD functionals with distance kernel K(x,y)=±|x-y|. Using the isometric embedding of P2(R) into the cone C(0,1)⊂L2(0,1) via quantile functions, it reformulates the flow as an L2 gradient flow for an associated functional, and adds a Sobolev H1 regularization λ/2∫|u'|^2 with Neumann boundary conditions (shifted by the target quantile for the negative kernel). Main results: Theorem 3.4 proves existence and uniqueness of a strong solution of the Cauchy problem for the regularized negative-kernel functional, yielding a unique Wasserstein gradient flow; Theorem 3.10 proves existence of generalized minimizing movements for the regularized positive-kernel functional, solving a Cauchy problem with the limiting subdifferential. Proposition 3.6 gives conditions under which the cone constraint is automatically preserved, reducing the flow to the unconstrained Neumann problem. The paper also derives explicit Dirac-to-Dirac examples, a Stefan-type free boundary problem for the negative kernel after contact, convergence as λ↓0 and t↑∞ remarks, and numerical experiments showing that the regularization removes a 'dissipation-of-mass' defect.","tokens_in":21133,"tokens_out":29500,"duration_ms":286919,"significance":"The existence theory for the positive distance kernel appears to be new; the verification of the Rossi–Savaré compactness and chain-rule hypotheses via the compact Sobolev embedding is natural and sound. The negative-kernel result is a regularized extension of the existing convex-flow theory, and the explicit formulas and numerics illustrate a practically relevant improvement. The paper is generally careful: the central theorems are proved from stated assumptions via standard semigroup theory, the regularization parameter λ is explicit and not fitted, and the text is candid about parts that remain numerical (the Stefan problem after contact) or sketched (fractional Sobolev relaxation). The referee checked the disputed monotonicity argument in Proposition 3.6; it is correct, because the term gn(s2)-gn(s1)≥0, the strict decrease of gn+1, and the inequality R^-_ν(gn+1(s1))≥R^+_ν(gn+1(s2)) make the central strict inequality true independently of τ and of the target density. Overall this is a solid contribution; the issues below are local and presentation-level.","major_comments":[],"minor_comments":[{"comment":"The statement that for a proper convex and lsc functional F one has dom(∂F)=dom(F) is false in general; the paper's own FH in Lemma 3.2 is a counterexample, since dom(FH)=H1(0,1) while dom(∂FH) is a proper dense subset. Please replace this with the correct statement that dom(∂F) is dense in dom(F).","section":"Section 2.2 (after Eq. (5))"},{"comment":"The claim that the convergence in (3) is 'globally uniform in t∈[0,∞)' is stronger than the standard compact-interval convergence result for convex lsc semigroups; please either supply a proof or weaken the statement to uniform convergence on compact intervals.","section":"Theorem 3.4"},{"comment":"The proof chooses an interval (a,b)⊂(0,1) with g'_{n+1}(a)=g'_{n+1}(b)=0 around a point where g'_{n+1}<0; if the negative region touches the boundary, one of the endpoints may be 0 or 1. The monotonicity contradiction is unaffected, but the proof should allow one-sided intervals or explicitly justify that boundary endpoints can be handled by approximation.","section":"Proposition 3.6 (proof)"},{"comment":"The Stefan problem for t≥t* is introduced formally, and the paper does not prove well-posedness for it or the claimed emergence of a Dirac point at 0. Since this is presented as a numerical observation, this is acceptable, but the formal nature of the Stefan formulation should be stated more explicitly.","section":"Example 3.7"},{"comment":"The equivalence of (25) and (26) is stated only for continuous CDFs, yet the Dirac-to-Dirac example has a discontinuous Rν; the text should make explicit that in that case the BVP solver is applied only to the absolutely continuous part, with the singular part treated by the separate procedure described in Section 4.2.","section":"Section 4.1 / Algorithm 1"},{"comment":"The claim that Theorems 3.4 and 3.10 remain valid with fractional Sobolev spaces is only a sketch; please label it as an outlook/conjecture rather than a statement of established results.","section":"Remark 3.13"},{"comment":"There are minor typographical issues: 'Figues' should be 'Figures', 'T echnical' should be 'Technical', and 'absolute continuous' should be 'absolutely continuous'.","section":"Throughout (Section 4.2)"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is within the scope of math.AP and the central theorems appear sound. The referee's concerns are local: a false general claim about dom(∂F), an overstrong convergence statement, and some presentation issues around the Stefan problem and the numerical equivalence for discontinuous CDFs. None of these affects the main existence results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers the first existence result for generalized minimizing movements of the positive-kernel MMD in one dimension, via a Sobolev regularization on quantile functions, and a clean existence theorem for the regularized negative-kernel flow. The two main theorems (3.4 and 3.10) are proven with standard semigroup theory and the Rossi–Savaré machinery, and the explicit Dirac-to-Dirac examples are genuinely illuminating. The numerics also make a convincing case that the regularization cures the dissipation-of-mass defect, even without shipped code.\n\nThe main soft spot is Proposition 3.6. The stress-test note is right: the strict inequality in the middle of the proof does not follow from the stated assumptions. For a uniform target with density ρ, the comparison reduces roughly to (1−2τρ)(g_{n+1}(s1)−g_{n+1}(s2)) plus lower-order terms, which can be negative once 2τρ > 1. So the claim that the implicit Euler step preserves the cone C(0,1) for arbitrary τ is not established. This matters because the numerical scheme (25) and the equivalence between (14) and (16) rest on that proposition. For the small τ used in the experiments this is probably harmless, but the proposition as written is too strong. The fix is likely a smallness condition on τ relative to the target density, or a different argument. The central existence theorems 3.4 and 3.10 are not affected, since they work with the full convex functional including the indicator.\n\nThe other limitation, that H1 assumptions force compact convex support, is honestly stated in Remark 3.13, and the fractional Sobolev relaxation is sketched but not developed. That is a real restriction, but it is not hidden. The Stefan problem in Example 3.7 is left open; that is fine, it is flagged as more involved.\n\nOverall, the paper is a solid contribution to the theory of MMD gradient flows. The existence results are new, the techniques are appropriate, and the applied motivation is clear. The citation pattern is fine; the prior self-cited results are the natural foundation. The paper deserves a serious referee. I would send it to review and ask for a repair of Proposition 3.6 before acceptance; the rest of the analysis holds up.","headline":"Solid existence theory for regularized 1D MMD flows, with a repairable gap in the negative-kernel reduction that does not touch the main theorems.","tokens_in":21689,"tokens_out":2851,"would_cite":true,"duration_ms":29108,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35K55","49Q22","46E35"],"pacs":[],"model":"deepseek-v4-flash","headline":"By adding a Sobolev H1 regularizer to the quantile formulation, this paper proves that one-dimensional MMD gradient flows with both positive and negative distance kernels exist—uniquely for the negative kernel and as generalized…","keywords":["Wasserstein gradient flow","maximum mean discrepancy","distance kernel","quantile functions","Sobolev regularization","generalized minimizing movement","Cauchy problem","nonconvex functional"],"falsifier":"Run the negative-kernel Dirac-to-Dirac example (initial $\\delta_{-1}$, target $\\delta_0$) with the implicit Euler scheme (25) and compare the pre-contact phase with the explicit heat-equation solution (19); then track the free boundary $\\beta(t)$ in the Stefan phase. If, for some $\\lambda>0$ and small step size $\\tau$, the numerical quantile $g(t,\\cdot)$ leaves the cone $C(0,1)$ or the free boundary $\\beta(t)$ decreases, the H1 regularization would fail to preserve the Wasserstein/quantile structure, contradicting Proposition 3.6 and the Wasserstein interpretation of Theorem 3.4.","tokens_in":20536,"feed_emoji":"🌊","tokens_out":11569,"duration_ms":104104,"temperature":0.7,"pith_summary":"The paper shows that adding a Sobolev term to the quantile formulation removes two obstructions in one-dimensional Wasserstein gradient flows of maximum mean discrepancy (MMD) functionals with distance kernels $K(x,y)=\\pm|x-y|$. Since measures on $\\mathbb{R}$ are represented by quantile functions in the cone $C(0,1)\\subset L^2(0,1)$, these flows become Cauchy problems on $L^2(0,1)$; the negative-kernel functional is convex, while the positive-kernel functional is not. Adding $\\frac{\\lambda}{2}\\int_0^1 |u'|^2\\,ds$ yields a unique strong solution—hence a unique Wasserstein gradient flow—for the regularized negative kernel, and a generalized minimizing movement for the regularized positive kernel. The same regularization moves the support toward the target instead of leaving the dissipating mass tail seen in the unregularized flow, which the paper demonstrates on explicit Dirac-to-Dirac examples and in numerics.","feed_headline":"Regularized MMD flows exist for both distance kernels","feed_subtitle":"Adding an H1 term to the quantile formulation guarantees flows and removes a mass tail.","key_machinery":"The carrying device is the isometric embedding of $P_2(\\mathbb{R})$ into the cone $C(0,1)\\subset L^2(0,1)$ of quantile functions, combined with the H1 Sobolev regularizer. The regularizer is $F_H(u)=\\frac{\\lambda}{2}\\int_0^1 |u'(s)|^2\\,ds$ (infinite outside $H^1(0,1)$), and for the negative kernel it is shifted to $F_{H,\\nu}(u)=F_H(u-Q_\\nu)$ so that the Neumann boundary conditions are prescribed by the target quantile; its subdifferential is $-\\lambda u''$ with those boundary conditions. Adding $I_{C(0,1)}$ forces the flow to stay in the quantile cone. For the negative kernel the regularized functional is convex, so maximal monotone operator theory applies; for the positive kernel the compact Sobolev embedding $H^1(0,1)\\hookrightarrow L^2(0,1)$ gives compact sublevels in the minimizing-movement scheme, and convexity of the pieces supplies the chain rule needed for the limiting-subdifferential formulation.","core_discovery":"The paper's central claim is that the Sobolev-regularized functionals $\\widetilde F^\\pm_\\nu = F^\\pm_\\nu + F_H^{(\\nu)} + I_{C(0,1)}$ have well-defined Wasserstein-type gradient flows in one dimension. For the negative kernel, $\\widetilde F^-_\\nu$ is proper, convex, and lower semicontinuous, and Theorem 3.4 asserts a unique strong solution $g\\in H^1_{\\mathrm{loc}}([0,\\infty);L^2(0,1))$ of $\\partial_t g\\in -\\partial \\widetilde F^-_\\nu(g)$, whose push-forward is the unique Wasserstein gradient flow of $\\widetilde F^-_\\nu$. For the positive kernel, $\\widetilde F^+_\\nu$ is not $\\lambda$-convex, but the H1 term supplies compact sublevels and the required chain rule, so Theorem 3.10 asserts the existence of generalized minimizing movements $g\\in \\mathrm{GMM}(\\widetilde F^+_\\nu,g_0)$ that are strong $H^1(0,T;L^2(0,1))$ solutions of $\\partial_t g\\in -\\partial_l \\widetilde F^+_\\nu(g)$ with the limiting subdifferential, and the associated curves in $P_2(\\mathbb{R})$ are Wasserstein absolutely continuous flows. The positive-kernel Dirac example gives $g(t,s)=-1-t$, so the regularized positive flow is not the time reversal of the negative one.","pith_inferences":["The paper does not state this, but the 'horns' at the support boundary, whose height is set by the target density, suggest that in higher dimensions a Laplacian-type penalty could similarly prevent mass from spreading to infinity; a testable extension is to run the same construction with a radial Laplacian term on quantile-like radial maps.","The fractional Sobolev relaxation in Remark 3.13 is only sketched; making it precise for $\\sigma<1/2$ would likely extend the theorems to initial and target measures with unbounded or disconnected support, and the natural first check is whether the compactness argument survives without the absolute continuity of $g_0$.","Theorem 3.10 asserts existence but not uniqueness of the generalized minimizing movement, so different subsequences of step sizes or different $\\lambda$ values may select different positive-kernel flows; this is a concrete numerical question for non-constant initial data.","The Stefan problem in Example 3.7 is a well-defined free-boundary problem; a rigorous analysis of $\\beta(t)$ would turn the numerically observed formation of a Dirac at the target into a theorem and give a benchmark for Wasserstein numerical schemes."],"forward_implications":["For the negative distance kernel, the regularized flow exists uniquely, can be approximated by an implicit Euler scheme that solves a Neumann boundary-value problem at each step, and converges to the unregularized flow as $\\lambda\\downarrow 0$.","For the positive distance kernel, existence of a flow is restored despite the lack of $\\lambda$-convexity: generalized minimizing movements exist on every finite time horizon and yield Wasserstein absolutely continuous curves.","The regularization gives a stronger long-time bound, $\\mathrm{MMD}_K^2(\\mu(t),\\nu)+|Q_{\\mu(t)}-Q_\\nu|^2_{H^1} \\le W_2^2(\\mu_0,\\nu)/(2t)$, so the flow approaches the target at least as fast as the unregularized one.","In the Dirac-to-Dirac example, the negative-kernel flow is explicit before it touches the target (a heat equation with forcing $2s$) and then becomes a Stefan free-boundary problem, showing the regularization turns a dissipating mass tail into a moving support.","The regularized positive-kernel flow from $\\delta_{-1}$ away from $\\delta_0$ is the translation $\\delta_{-1-t}$, so positive and negative kernel flows are not time reversals of one another."],"supporting_citations":[{"why":"Supplies the quantile-functional form $F_\\nu$ in Lemma 3.1, the prior existence analysis of the unregularized negative-kernel MMD flow, and the proof template for Theorem 2.4.","marker":"[13]"},{"why":"Provides the Wasserstein gradient flow and minimizing-movement theory, including $\\lambda$-convexity and convergence of discrete minimizers, used in Theorem 3.4 and Remark 3.9.","marker":"[1]"},{"why":"Supplies the generalized minimizing movement theory for nonconvex Hilbert-space functionals, including the compactness lemma and chain-rule conditions used in Theorem 3.10.","marker":"[26]"},{"why":"Provides the maximal monotone operator and semigroup results used to obtain unique strong solutions of the convex regularized Cauchy problem.","marker":"[10]"},{"why":"Contains the explicit unregularized Dirac-to-Dirac MMD flow whose dissipation-of-mass defect motivates the regularization.","marker":"[18]"},{"why":"Gives the subdifferential sum rule used in Proposition 3.6 to identify $\\partial(F_\\nu+F_{H,\\nu})$ with $\\partial F_\\nu+\\partial F_{H,\\nu}$.","marker":"[24]"}],"fun_headline_variants":["Sobolev term tames MMD flows for both kernels","H1 regularization guarantees MMD gradient flows","Distance-kernel MMD flows exist with Sobolev fix","Regularized MMD flows: existence for plus and minus kernels","Sobolev-regularized MMD flows now rigorously defined"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Both main theorems assume the initial quantile $g_0$ (and, for the negative kernel, the target quantile $Q_\\nu$) lies in $C(0,1)\\cap H^1(0,1)$, which forces the initial and target measures to have compact and convex support; the paper's fractional-Sobolev relaxation of this restriction is only sketched.","fun_headline_variants_meta":{"raw":{"variants":["Sobolev term tames MMD flows for both kernels","H1 regularization guarantees MMD gradient flows","Distance-kernel MMD flows exist with Sobolev fix","Regularized MMD flows: existence for plus and minus kernels","Sobolev-regularized MMD flows now rigorously defined"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1421,"prompt_tokens":1089,"completion_tokens":332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":705,"completion_tokens_details":{"reasoning_tokens":248}},"tokens_in":705,"tokens_out":332,"duration_ms":3368,"temperature":1.0,"reasoning_tokens":248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:14:57.812041+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the negative-kernel Dirac-to-Dirac example (initial $\\delta_{-1}$, target $\\delta_0$) with the implicit Euler scheme (25) and compare the pre-contact phase with the explicit heat-equation solution (19); then track the free boundary $\\beta(t)$ in the Stefan phase. If, for some $\\lambda>0$ and small step size $\\tau$, the numerical quantile $g(t,\\cdot)$ leaves the cone $C(0,1)$ or the free boundary $\\beta(t)$ decreases, the H1 regularization would fail to preserve the Wasserstein/quantile structure, contradicting Proposition 3.6 and the Wasserstein interpretation of Theorem 3.4.","supporting_citations":[{"cited_title":"Ambrosio, N","cited_arxiv_id":null,"evidence_quote":"Provides the Wasserstein gradient flow and minimizing-movement theory, including $\\lambda$-convexity and convergence of discrete minimizers, used in Theorem 3.4 and Remark 3.9."},{"cited_title":"Rossi and G","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized minimizing movement theory for nonconvex Hilbert-space functionals, including the compactness lemma and chain-rule conditions used in Theorem 3.10."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the maximal monotone operator and semigroup results used to obtain unique strong solutions of the convex regularized Cauchy problem."},{"cited_title":"Hertrich, R","cited_arxiv_id":null,"evidence_quote":"Contains the explicit unregularized Dirac-to-Dirac MMD flow whose dissipation-of-mass defect motivates the regularization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the subdifferential sum rule used in Proposition 3.6 to identify $\\partial(F_\\nu+F_{H,\\nu})$ with $\\partial F_\\nu+\\partial F_{H,\\nu}$."}],"review_version":1}