{"id":"3931939c-d97d-472d-9585-8f7da752add3","arxiv_id":"2608.01182","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Global well-posedness, particle mean-field limits, rigidity of critical points, and obstructions to uniform convergence rates are established for Wasserstein gradient flows of squared MMD with energy kernels -|x|^q, 0 < q < 2.","lead":"This paper studies the flow that moves one probability distribution toward another by minimizing the squared maximum mean discrepancy of a repulsive, non-smooth energy kernel, K(z) = -|z|^q with 0 < q < 2. It proves global well-posedness for the continuum flow, a mean-field limit for the associated particle system, and a nearly complete classification of when stationary states must equal the target.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Strict subcriticality p>p_c is load-bearing: at p=p_c the singular-convolution bound Lemma 3.2 degenerates and W^{1,∞} velocity regularity is unproved; the d=q=1 endpoint is imported from [CR26b].","rationale":"The reader's weakest assumption matches the concern found here: the subcritical integrability condition p > p_c, together with the moment hypothesis, is what makes Lemma 3.2 applicable throughout Sections 3 and 4. The paper is careful and explicit about the boundary: the finite critical endpoint is left open in Remark 1.3 and Section 9, the d=q=1 endpoint is imported from the companion preprint [CR26b], and the mean-field estimate excludes d=1, 0<q<1 with a nonzero target in Remark 1.9. None of these omissions contradict the statements of the theorems, so they do not undermine the internal validity of the paper. But they are load-bearing for the central claim in the sense that the full abstract-level range '0<q<2' is not covered: a reader applying the well-posedness theorem at p=p_c, or relying on the one-dimensional super-Coulomb range for the mean-field estimate, would be outside the proved statements. The proposed concrete test isolates the degeneracy of Lemma 3.2 at p=p_c and shows why the strict inequality cannot be removed by the present method. Because this is a stated conditionality rather than a discovered error, the appropriate response is to keep the reader's CONDITIONAL verdict unchanged.","tokens_in":88836,"tokens_out":31432,"duration_ms":283327,"concrete_test":"Check the critical-endpoint obstruction explicitly. Take d=2, q=1 (so p_c=2) and f(x)=1_{|x|<1} |x|^{-1} log(1/|x|)^{-3/4}, which belongs to L^2(R^2) because ∫_0^1 dr/(r (log 1/r)^{3/2}) < ∞. Compute v = -∇K*f with ∇K(z)=z/|z|. Near x=0 the convolution kernel behaves like |y|^{-2} (log 1/|y|)^{-3/4} in polar coordinates, and ∫_0^1 dr/(r (log 1/r)^{3/4}) diverges, so ∇v is unbounded. This confirms that at p=p_c the Lemma 3.2 bound cannot give W^{1,∞} velocity and the strict inequality in (1.11) is essential. Independently re-derive the cited [CR26b] endpoint Cauchy theory to remove the companion dependency from the d=q=1 case of Theorem 1.2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 1.2: global well-posedness with v in L∞([0,T];W^{1,∞}) in the range p_c < p ≤ ∞ plus the Coulomb endpoint. The least secure input is the strict subcritical condition (1.11), equivalently (2−q)p* < d. The singular-convolution bound Lemma 3.2 is used in every regularity estimate: the velocity Lipschitz bound (3.12), the L^p estimate (3.37)–(3.38), the velocity time-derivative bound (3.48)–(3.49), and the W1-stability/uniqueness estimates (3.79)–(3.82). Its proof balances R = (||f||_{L1}/||f||_{Lp})^{p*/d} in (3.11), which requires a < d/p*. At p = p_c one has (2−q)p* = d, the bound degenerates, and convolution of |x|^{q−2} with compactly supported L^{p_c} densities need not be bounded or even locally bounded, so the asserted W^{1,∞} velocity class is not obtained by these arguments. The paper explicitly leaves p = p_c open (Remark 1.3 and Section 9) and delegates the d = q = 1 endpoint to the companion preprint [CR26b, Theorem 1.2 and Proposition 2.1]. This is a structural conditionality, not an internal contradiction: the stated theorems are accurate, but the advertised range '0<q<2' is not proved at critical integrability, and the mean-field estimate also excludes d=1, 0<q<1 with nonzero target (Remark 1.9). Applications at p=p_c, or relying on the one-dimensional super-Coulomb endpoint, have no support from the present manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Wasserstein gradient flow of the squared maximum mean discrepancy generated by the energy kernels K(z)=-|z|^q, 0<q<2. It proves global well-posedness for probability densities in subcritical L^p classes when d+q-2>0, with a Lipschitz velocity field and an energy-dissipation identity, and it imports the one-dimensional Coulomb endpoint d=q=1 from the companion paper [CR26b]. It then proves global noncollision for the associated N-particle system, a fixed-N convergence-to-critical-set result, a modulated-energy mean-field estimate with explicit N^{-1/2}-type control for well-prepared data, a particle-to-continuum criticality principle, and saddle equilibria showing that deterministic trajectories need not reach global empirical minimizers. A large part of the paper is devoted to the stationary picture: Lagrangian critical points are shown to agree with the target in most regimes, with exceptions in d=1 and d=3 when 0<q<1, including explicit one-dimensional non-minimizing critical points. The paper also proves asymptotic criticality and conditional target convergence, and constructs explicit obstructions to any initial-data-independent MMD decay modulus and to global Polyak-Lojasiewicz inequalities. The critical integrability endpoint p=p_c for d+q-2>0 is explicitly left open (Remark 1.3, Section 9), and the mean-field estimate explicitly excludes d=1, 0<q<1 with nonzero target (Remark 1.9).","tokens_in":89020,"tokens_out":28147,"duration_ms":250666,"significance":"The results are substantial if the external inputs are accepted. The paper gives a detailed, parameter-free construction of global weak solutions in a regime where standard displacement-semiconvexity theory does not apply, together with quantitative particle-to-continuum propagation and a nearly complete classification of absolutely continuous Lagrangian critical points. The obstruction results are concrete and falsifiable, and the main theorem statements are carefully qualified: the strict subcritical condition p>p_c is stated, the critical endpoint is explicitly excluded, and the d=q=1 endpoint is delegated to [CR26b]. The four-step regularization argument, the virial-based noncollision proof, and the potential-theoretic rigidity arguments are presented in detail and are largely self-contained apart from the commutator estimate of [NRS22] and the companion endpoint inputs. The main correctness risk is the dependence of endpoint claims on self-cited preprints and a localized derivation error in the L^p estimate that appears to be fixable.","major_comments":[{"comment":"The displayed derivative of the L^p norm is incorrect as written. Multiplying (3.26) by p(ρ_ε)^{p-1} and integrating by parts gives d/dt ‖ρ_ε‖_p^p = (p-1)∫(ρ_ε)^p ΔK_ε*(ρ_ε-μ_ε) dx, not (p-1)∫(ρ_ε)^{p-1} ΔK_ε*(ρ_ε-μ_ε) dx. With the printed power, Hölder would only give a bound involving ‖ρ_ε‖_{L^{p-1}}^{p-1}, which is not controlled by ‖ρ_ε‖_{L^p}^p for probability densities with ‖ρ‖_{L^p}<1. The inequality in (3.37) is exactly what follows from the corrected formula, because ΔK_ε*(ρ_ε-μ_ε) ≤ (-ΔK_ε)*μ_ε ≤ ‖(-ΔK)*μ‖_{L^∞}. The four-step uniform-estimate construction therefore closes after this correction. Please correct (3.36) and re-verify the powers in (3.37)-(3.39) and (3.47)-(3.49), all of which rely on this estimate.","section":"§3.2, Eq. (3.36)"},{"comment":"The advertised one-dimensional Coulomb endpoint is not proved within this manuscript. Theorem 1.2 for d=q=1 is imported from [CR26b, Theorem 1.2 and Proposition 2.1], and the d+q≤2 case of the commutator estimate in Proposition 4.13, which is used in the endpoint case of Theorem 1.8, is delegated to [RS26, Theorem 4.1]. These are self-cited companion or preprint results whose status should be clarified. If the journal requires stated theorems to be supported by the manuscript or by published references, the authors should either include the endpoint arguments as an appendix or explicitly mark the d=q=1 claims in Theorems 1.2 and 1.8 as conditional on the acceptance of [CR26b] and [RS26]. The main d+q-2>0 finite-p theory appears self-contained apart from the published commutator estimate [NRS22].","section":"§3 (first paragraph), Remark 1.3, Proposition 4.13"},{"comment":"The finite critical endpoint p=p_c for d+q-2>0 is left open, as the authors explicitly state. This is consistent with Theorem 1.2 as written, since condition (1.11) requires p>p_c whenever d+q-2>0. The paper should keep this qualification visible in the abstract and introduction; the current abstract's phrase 'subcritical L^p spaces' is accurate. The open problem is appropriately listed in Section 9 and does not by itself affect the stated theorems.","section":"Remark 1.3 and Section 9"}],"minor_comments":[{"comment":"The citation [RS26, Theorem 4.1] is load-bearing in the d+q≤2 case of Proposition 4.13 but does not appear in the visible reference list in the provided text. Please ensure the full reference is included and that its publication status is indicated.","section":"Proposition 4.13 / references"},{"comment":"The abstract states that the one-dimensional Coulomb endpoint is 'included'; consider adding a parenthetical in the introduction that this endpoint is obtained from [CR26b], so that readers are not led to expect a fully self-contained proof of the endpoint case in this paper.","section":"Abstract and Section 1"}],"recommendation":"major_revision","confidential_remarks":"The central finite-p theory is convincing once the L^p derivative formula in (3.36) is corrected, and the paper is transparent about its main limitations. However, the d=q=1 endpoint and the one-dimensional commutator case rely on self-cited preprints [CR26b] and [RS26]; the editor should confirm the acceptance status of those works before final acceptance. The novelty relative to [Chi+26] and [CR26b] is clearly delineated in Section 1.3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nThe short version: this is the first paper to get global well-posedness for Wasserstein gradient flows of the MMD energy with nonsmooth kernels K(z) = -|z|^q, 0<q<2, on R^d in subcritical L^p spaces, together with a matching particle system (noncollision, fixed-N criticality) and a modulated-energy mean-field estimate. It also gives a near-complete rigidity classification of Lagrangian critical points and clean obstructions to uniform convergence rates. The novelty is real: earlier work covered smooth kernels, the 1D q=1 quantile case, the torus critical Lorentz theory, and the Coulomb whole-space case; this covers the Euclidean range d+q-2>0 with p > p_c = d/(d+q-2), plus the d=q=1 endpoint.\n\nWhat it does well: the proof structure is careful and self-aware. The regularized problem, uniform estimates, passage to the limit, W1 stability, and the virial noncollision argument are all presented in detail. The paper explicitly says where it does not reach: the critical endpoint p=p_c is left open (Remark 1.3 and Section 9), the one-dimensional Coulomb endpoint is imported from [CR26b], and the mean-field estimate excludes d=1, 0<q<1 with nonzero target (Remark 1.9). Those are structural caveats, not hidden gaps—the stated theorems are accurate.\n\nSoft spots, in proportion: the strict condition p>p_c is load-bearing. Lemma 3.2, the singular-convolution bound, is used in every regularity and uniqueness estimate; at p=p_c it degenerates and the asserted W^{1,∞} velocity class is not proved. So the abstract-level claim \"0<q<2\" should be read as \"subcritical L^p, p>p_c, plus the imported Coulomb endpoint.\" Downstream applications at critical integrability get no support from this paper. Also, the commutator estimate for the mean-field proof comes from the self-cited [NRS22]; that is an external theorem, but it is published and refereed, so not a red flag by itself.\n\nOverall, the central argument holds up. The paper is honest, detailed, and fills a real gap. It deserves a serious referee; I would send it. The referee should spend time on the subcritical-to-critical boundary and on the imported endpoint, but I do not see a load-bearing flaw.","headline":"First global well-posedness and particle limits for nonsmooth MMD energy flows in subcritical L^p, with an honest boundary and one imported endpoint; worth serious refereeing.","tokens_in":89750,"tokens_out":2038,"would_cite":true,"duration_ms":20207,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35Q70","49Q22","35B40","35R09","46E22","60B10","82C22"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes that the Wasserstein gradient flow of the squared MMD with the nonsmooth energy kernels $K(z)=-|z|^q$, $0<q<2$, is globally well-posed on $\\mathbb{R}^d$ in subcritical $L^p$ classes of densities, and that the…","keywords":["maximum mean discrepancy","Wasserstein gradient flow","energy kernels","Riesz potentials","aggregation equation","mean-field limit","Lagrangian critical point","noncollision"],"falsifier":"Two concrete checks would settle open parts of the picture. (1) Well-posedness at the critical exponent $p=p_c=d/(d+q-2)$: proving uniqueness, or exhibiting non-uniqueness, there would fix the sharp threshold of Theorem 1.2. (2) Rigidity in the residual regime $d=3$, $0<q<1$: an absolutely continuous, non-radial probability density $\\rho$ with finite $q$-th moment, unbounded positive-part discrepancy, and $\\nabla K*(\\rho-\\mu)=0$ $\\rho$-almost everywhere for some target $\\mu$ would falsify the conjecture that natural-moment rigidity holds there. (3) A third check bears on target convergence: Corollary 5.25 rests on the uniform-in-time bound (5.95), so a solution with $1\\le q<2$ whose $L^p$ norm or $r$-th moment grows without bound would separate asymptotic criticality from convergence to the target.","tokens_in":88446,"feed_emoji":"🎯","tokens_out":10210,"duration_ms":85350,"temperature":0.7,"pith_summary":"This paper tries to establish a complete dynamical theory for the Wasserstein gradient flow of the squared Maximum Mean Discrepancy built from the nonsmooth energy kernels $K(z)=-|z|^q$, $0<q<2$ — the family whose $q=1$ member is the classical energy distance. Standard gradient-flow theory fails here because the interaction force is not Lipschitz at the diagonal and, in dimensions $d\\ge 2$, the energy is not displacement semiconvex. The paper's central claim is that in the subcritical range $d+q-2>0$, global well-posedness nevertheless holds in $L^p$ spaces: unique weak solutions with Lipschitz velocity fields, an exact energy–dissipation identity, collision-free particle dynamics, a dimension-free $N^{-1/2}$ mean-field limit, and a near-complete rigidity classification of stationary states that identifies the target as the only absolutely continuous critical point outside two exceptional regimes. A sympathetic reader should care because the flow is the dynamical model behind MMD-based sampling and generative learning with these practical kernels, and because the paper pins down exactly what is and is not true about its convergence: no uniform rate exists, but qualitative approach to the target holds throughout most of the parameter range.","feed_headline":"Negative-distance MMD flows are globally well-posed","feed_subtitle":"Unique global solutions, collision-free particles, and dimension-free mean-field convergence.","key_machinery":"The argument rests on four objects. (1) The singular-convolution bound (Lemma 3.2): for $0<a<d/p^*$, $\\sup_x\\int |x-y|^{-a}|f(y)|\\,dy \\le C\\|f\\|_{L^1}^{1-\\theta}\\|f\\|_{L^p}^{\\theta}$; applied with $a=2-q$ it makes convolutions of the kernel Hessian $|x|^{q-2}$ against $L^p$ densities bounded, yielding the Lipschitz velocity and driving the regularize–estimate–pass-to-the-limit–stability scheme for existence, plus the $W^1$ contraction for uniqueness. (2) The Fourier–Sobolev representation (Lemma 2.1) identifying $\\mathrm{MMD}_q^2$ with the $\\dot{H}^{-(d+q)/2}$ norm, which turns the energy into a Hilbert norm and the modulated energy into a coercive comparison functional. (3) The modulated-energy commutator estimate (Proposition 4.13) bounding the growth of the empirical-to-continuum MMD gap by velocity norms, giving the Gr\\\"onwall factor in the mean-field theorem. (4) The rigidity reduction: Lagrangian criticality plus Sobolev locality converts the critical-point equation into an equality of Riesz potentials on the positive part of the discrepancy, which the complete maximum principle for Riesz kernels and a one-sign cancellation argument close, leaving only the $d\\in\\{1,3\\}$, $0<q<1$ exceptions.","core_discovery":"On its own terms, the paper's claim is that the MMD Wasserstein flow of the energy kernel family is a well-posed evolution equation across its whole nondegenerate range. For $d+q-2>0$, any pair of probability densities with finite $r$-th moments ($r\\ge 1$) and subcritical integrability $p\\ge p_c=d/(d+q-2)$ (strictly supercritical when the bound is finite) generates a unique global weak solution $\\rho_t$ of $\\partial_t\\rho_t = -\\nabla\\cdot(\\rho_t v_t)$ with $v_t = -\\nabla K*(\\rho_t-\\mu)$, the velocity lying in $L^\\infty([0,T];W^{1,\\infty})\\cap C([0,T]\\times\\mathbb{R}^d)$, together with the energy–dissipation identity; the one-dimensional Coulomb endpoint $d=q=1$ is included via the companion Coulomb theory. The same framework yields: global noncollision and convergence to the critical set for the diagonal-free $N$-particle system; a modulated-energy estimate giving $N^{-1/2}$-rate mean-field convergence on finite intervals for well-prepared data; rigidity of Lagrangian critical points (an absolutely continuous critical state equals the target under finite $q$-moments, except possibly for $d\\in\\{1,3\\}$ with $0<q<1$); and explicit obstructions — no initial-data-uniform MMD decay modulus and no global Polyak–\\L{}ojasiewicz inequality on $\\mathbb{R}^d$ or on $\\mathbb{T}^d$ in the stated regimes.","pith_inferences":["The paper leaves the critical integrability endpoint $p=p_c$ open and notes that velocity control degenerates from Lipschitz to Osgood there; the natural next step it gestures toward is an endpoint theory in a critical Lorentz class, and if uniqueness fails there, the sharp well-posedness threshold would be exactly the subcritical range proved here.","The rigidity exceptions ($d=1,3$ with $0<q<1$) and the explicit non-minimizing critical points suggest the long-time limit is genuinely selection-dependent in the flexible regimes: one could test numerically whether the basins of attraction of the non-minimizing critical states are nonempty for the continuum flow, which would show the dynamics, not the energy, decides the limit.","The finite-speed transport mechanism behind the no-rate obstruction implies that the practically meaningful object is the source's entry time into the target's bulk: the one-dimensional example shows post-entry relaxation is exponential when the target density is bounded below and polynomial when it vanishes, so rate estimates for these flows should be formulated after a datum-dependent waiting ti","The paper's saddle constructions and non-minimizing critical points suggest a wider principle: the empirical MMD energy landscape has collision-free critical states at all energy levels above the ground state, so deterministic particle trajectories can be trapped by symmetry; understanding which critical points admit recovery sequences of finite-particle equilibria would connect the static quantiz"],"forward_implications":["For every kernel exponent $q$ with $d+q-2>0$ (and at the $d=q=1$ endpoint), the continuum flow is globally well-posed from any admissible density: the $r$-th moment stays bounded on finite intervals and the squared MMD equals its initial value minus the accumulated dissipation $\\int |\\nabla K*(\\rho_\\tau-\\mu)|^2\\,d\\rho_\\tau$.","The $N$-particle system with diagonal-free interactions has global collision-free solutions from any pairwise-distinct configuration; all particles remain bounded, their minimum separation is bounded away from zero uniformly in time, and the configuration converges to the collision-free critical set of the particle energy.","Well-prepared empirical measures (MMD error of order $N^{-1/2}$ for independent samples) stay within a Gr\\\"onwall factor of the continuum solution, giving convergence of the particle dynamics to the continuum flow as $N\\to\\infty$ on every finite time interval at rate $N^{-1/2}$.","Any absolutely continuous stationary state in the Lagrangian sense (the driving force vanishes on the carried mass) must equal the target under finite $q$-moments, except possibly in dimensions $1$ and $3$ for $0<q<1$; in the residual three-dimensional regime rigidity holds under an extra moment, compact support of the positive-part discrepancy, or radiality.","No initial-data-independent multiplicative MMD decay rate can hold on $\\mathbb{R}^d$: translating a fixed compactly supported source rules out a single decay modulus, and global Polyak–\\L{}ojasiewicz inequalities fail both in the whole space and in the periodic Riesz/Coulomb regimes stated.","Asymptotic criticality is unconditional for $1\\le q<2$: every $\\omega$-limit point of the continuum orbit is Lagrangian critical without additional long-time bounds, and the full orbit approaches the Lagrangian critical set; for $0<q<1$ the same holds under uniform moment and $L^p$ bounds, and rigidity then upgrades this to convergence to the target throughout the rigid part of the well-posedness "],"supporting_citations":[{"why":"Supplies the $L^p$ aggregation-equation strategy (regularization, uniform estimates, stability) that the well-posedness proof adapts to include the target field.","marker":"[BLR11]"},{"why":"Provides the Cauchy theory covering the one-dimensional Coulomb endpoint $d=q=1$ used in Theorems 1.2 and 1.5.","marker":"[CR26b]"},{"why":"Contains the transport-commutator estimate (its Proposition 3.1) that controls the modulated-energy term in the mean-field theorem.","marker":"[NRS22]"},{"why":"Gives the odd-dimensional energy-distance rigidity result for $q=1$ that the stationary-state classification extends.","marker":"[BV25]"},{"why":"Supplies the flow representation and uniqueness theory for linear continuity equations with Lipschitz velocity used throughout the solution class.","marker":"[AGS08]"},{"why":"Provides the complete maximum principle for Riesz kernels used to close the contact-rigidity argument.","marker":"[Zor23]"},{"why":"Supplies the characteristic fixed-point scheme used to solve the regularized problem in the well-posedness construction.","marker":"[Lau07]"}],"fun_headline_variants":["MMD energy-kernel flows: global well-posedness","Unique global solutions for MMD energy-kernel flows","Collision-free particles converge to MMD flow","No universal decay for MMD Wasserstein flows","Nonsmooth energy kernels yield well-posed MMD flows"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire theory depends on the source and target densities having finite moments and enough subcritical integrability ($p > d/(d+q-2)$) so that convolutions of the kernel's second derivative stay bounded; the critical borderline case $p=p_c$ is explicitly left open.","fun_headline_variants_meta":{"raw":{"variants":["MMD energy-kernel flows: global well-posedness","Unique global solutions for MMD energy-kernel flows","Collision-free particles converge to MMD flow","No universal decay for MMD Wasserstein flows","Nonsmooth energy kernels yield well-posed MMD flows"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000561,"raw_usage":{"total_tokens":2818,"prompt_tokens":1251,"completion_tokens":1567,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":867,"completion_tokens_details":{"reasoning_tokens":1489}},"tokens_in":867,"tokens_out":1567,"duration_ms":11790,"temperature":1.0,"reasoning_tokens":1489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:10:50.290912+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two concrete checks would settle open parts of the picture. (1) Well-posedness at the critical exponent $p=p_c=d/(d+q-2)$: proving uniqueness, or exhibiting non-uniqueness, there would fix the sharp threshold of Theorem 1.2. (2) Rigidity in the residual regime $d=3$, $0<q<1$: an absolutely continuous, non-radial probability density $\\rho$ with finite $q$-th moment, unbounded positive-part discrepancy, and $\\nabla K*(\\rho-\\mu)=0$ $\\rho$-almost everywhere for some target $\\mu$ would falsify the conjecture that natural-moment rigidity holds there. (3) A third check bears on target convergence: Corollary 5.25 rests on the uniform-in-time bound (5.95), so a solution with $1\\le q<2$ whose $L^p$ norm or $r$-th moment grows without bound would separate asymptotic criticality from convergence to the target.","supporting_citations":[],"review_version":2}