{"id":"20ad1390-c015-49b5-bb55-bf95651246b0","arxiv_id":"2509.22879","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A moment-SOS semidefinite hierarchy computes best W2 and TV mixture approximations over semi-algebraic parameter sets and can recover the mixture order from a rank condition.","lead":"This paper builds semidefinite programming relaxations to approximate a probability distribution, known only through finitely many moments, by a mixture of standard parametric distributions. The method gives convergence guarantees and can estimate the number of mixture components and their parameters, with clustering as an application.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Flatness of M_d(phi) alone does not certify finite convergence; a d=1 Bernoulli-vs-Gaussian example falsifies the recovery theorem.","rationale":"The reader's weakest assumption identified the same gap: flatness of M_d(phi*) alone is asserted to certify finite convergence and recovery of the full solution, but no flatness or extension condition is proved for the coupling lambda* or for the TV measures psi+*, psi-*. My stress test makes the concern concrete and stronger: the gap is not merely a missing proof detail; the stated rank condition can hold at a non-optimal pseudo-solution. The Bernoulli-vs-Gaussian example satisfies all hypotheses of Theorem 3.2 at d=1 while producing a false zero transport cost; the true optimal mixture has a larger component count and strictly smaller W2 distance. This directly attacks the central algorithmic claim (Algorithm 1 and the finite-recovery guarantee), not just the asymptotic convergence part, which appears standard. Therefore the paper's main theorem is false as stated, and a conditional acceptance based on 'repairable proof gap' is insufficient; the finite-recovery claim needs substantial reformulation or additional assumptions. I recommend REJECT of the current version, while acknowledging that the asymptotic hierarchy and the experiments may still be salvageable in a revised version.","tokens_in":23369,"tokens_out":17988,"duration_ms":193905,"concrete_test":"Implement the one-dimensional instance in the paper's TSSOS/Mosek pipeline: mu=1/2(delta_0+delta_1), S_theta=[0,1] x [0.05,1], family N(m,sigma^2), R=1, epsilon=1e-3, and solve (12) at d=1. Verify that the SDP returns tau_1 = epsilon (up to solver tolerance) and that an optimal phi* has rank M_1(phi*)=1 with extracted parameters (m,sigma)=(0.5,0.5). Then compute the true W2^2 between mu and the extracted N(0.5,0.25) using a fine-grid discrete optimal transport solver, and compare it with W2^2(mu, 1/2 N(0,0.05^2)+1/2 N(1,0.05^2)). If, as expected, the former is about 0.10 and the latter at most 0.0025, then the rank condition has certified a non-optimal mixture, refuting Theorem 3.2. For the TV case, run the same comparison with (17) at d=1.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing concern is not merely a missing proof but a false finite-convergence certificate. Theorem 3.2/3.3 conclude from rank M_d(phi*)=rank M_{d-dmin}(phi*) that the flat-extended phi* solves the original infinite-dimensional problem. The proof in Appendix A.3 invokes Theorem A.1, but that theorem applies only to accumulation points of optimal solutions as d grows; a flat extension at one fixed d is not shown to be such an accumulation point, and no flatness/extension condition is imposed on the coupling lambda*(d) (or on psi+*, psi-* in the TV case). Flatness of M_d(phi) only yields a K-atomic mixing measure; it does not produce a full coupling with marginals mu and nu_{phi*}.\n\nThis gap is fatal. Take n=1, mu=1/2(delta_0+delta_1), family N(m,sigma^2) with (m,sigma) in [0,1] x [0.05,1], R=1 (constant), and any epsilon>0. At d=1, phi*=delta_{(0.5,0.5)} has M_1(phi*) of rank 1, so (13) holds with K=1. The truncated SDP (12) is feasible with lambda_{1,0}=lambda_{0,1}=1/2, lambda_{2,0}=lambda_{0,2}=1/2, lambda_{1,1}=1/2; the corresponding 3x3 moment matrix is PSD and the cost is lambda_{2,0}+lambda_{0,2}-2lambda_{1,1}=0. Hence tau_1 = epsilon (since R=1) and the rank condition certifies finite convergence. But no true coupling between mu and N(0.5,0.25) has zero transport cost: W2^2(mu, N(0.5,0.25)) is about 0.10. Meanwhile the two-component mixture 1/2 N(0,0.05^2)+1/2 N(1,0.05^2) has W2^2 <= 0.0025 < 0.10. Thus the extracted K=1 mixture is not optimal, and (13) does not imply tau_d = tau_infty. The same defect invalidates Theorem 3.3.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript develops semidefinite programming hierarchies for the following problem: given a target probability measure μ known only through moments up to order 2d, find a mixture ν_φ from a parametric family (Gaussian, Poisson, exponential, etc.) over a compact basic semi-algebraic parameter set that minimizes a regularized squared 2-Wasserstein or total variation distance. The main theoretical claims are (i) asymptotic convergence of the relaxation values to the true optimum, and (ii) finite convergence and recovery of the mixture order and parameters whenever the optimal pseudo-moment matrix M_d(φ*) satisfies the flatness condition rank M_d(φ*) = rank M_{d-dmin}(φ*). The paper also presents an algorithm for extracting mixture parameters and numerical experiments on synthetic GMMs and MNIST, with an application to initializing k-means and EM.","tokens_in":23967,"tokens_out":12502,"duration_ms":99684,"significance":"The problem is timely, and the general approach—reformulating mixture approximation as a generalized moment problem and applying moment-SOS relaxations—is potentially valuable. The TV reformulation via the Hahn-Jordan decomposition is a nice device, and the paper ships reproducible code with experiments. The asymptotic convergence arguments follow standard moment-SOS patterns, and the Carleman-type bounds for Gaussian, Poisson, and exponential mixtures are useful ingredients. However, the central advertised finite-convergence/recovery theorem is not correct as stated, and the numerical claims that rely on it are not supported.","major_comments":[{"comment":"The flatness condition (13) on M_d(φ*) alone does not certify finite convergence or optimality. Counterexample: n=1, μ=(δ_0+δ_1)/2, family N(m,σ^2) with S_θ=[0,1]×[0.05,1], R≡1, ε>0 small, d=1. Let φ*=δ_(0.5,0.5). Then rank M_1(φ*)=rank M_0(φ*)=1, and (12) is feasible with λ_{1,0}=λ_{0,1}=λ_{2,0}=λ_{0,2}=λ_{1,1}=1/2, giving objective 0. Yet W2^2(μ,N(0.5,0.5^2))=0.5, while 1/2 N(0,0.05^2)+1/2 N(1,0.05^2) has W2^2≤0.0025. Thus (13) holds but φ* is not optimal for (11) and K=1 is not the optimal mixture order. The proof's appeal to Theorem A.1 is inapplicable: that theorem concerns accumulation points as d→∞; a flat extension at fixed d need not be such a point, and no flatness is imposed on λ*. This invalidates the finite-convergence claim.","section":"Theorem 3.2 / Appendix A.3"},{"comment":"The same gap occurs in the TV case. The rank condition is only on M_d(φ*); no flatness or extension condition is stated for M_d(ψ_+*) or M_d(ψ_-*). The proof says 'proceed as in proof of Theorem 3.2' and inherits the defect. Moreover, (17c)-(17d) are truncated moment inequalities; even if φ* is K-atomic, the pseudo-moment pair (ψ_+*,ψ_-*) need not arise from measures satisfying ψ_+*≤μ and ψ_-*≤ν_φ*. The d=1 example above can be adapted to show K-atomic recovery can fail.","section":"Theorem 3.3 / Appendix A.4"},{"comment":"Lemma A.1 is stated for any mixture family satisfying Assumption 2.2, but its proof treats only Gaussian, Poisson, and exponential families; the final sentence asserts other families 'can also be shown' without proof. Theorems A.1 and A.2 invoke Lemma A.1 to obtain moment determinacy of ν_φ* and of the y-marginal of λ*. As a result, the asymptotic convergence theorems are not established at the stated level of generality. Please restrict the claims to the named families or provide a general proof.","section":"Lemma A.1 / Theorems A.1-A.2"}],"minor_comments":[{"comment":"The summation on the right-hand side should run over s(p,d), not s(n,p).","section":"Eq. (18)"},{"comment":"Line 1 has an extra closing parenthesis in M_d(φ*(d))); line 7 repeats 'global global'.","section":"Algorithm 2"},{"comment":"'Hanh-Jordan' should be 'Hahn-Jordan'.","section":"§3.2"},{"comment":"The value τ^{W2}_{4,0,R}=0 is the truncated relaxation value, not the true W2 distance. The sentence claiming the recovered mixtures are 'best' in the W2/TV sense relies on the invalid finite-convergence theorem and should be rephrased.","section":"Figure 1 caption"}],"recommendation":"reject","confidential_remarks":"The false finite-convergence theorem is a load-bearing error: it is the basis for the abstract, Algorithm 1, and the interpretation of the experiments. The simple d=1 counterexample shows the claimed rank condition on M_d(φ*) alone cannot certify optimality of the extracted mixture. The asymptotic convergence part may be salvageable, but the present manuscript's central claim is not."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jan,\n\nThe paper is a serious extension of Lasserre's moment-SOS machinery to mixture approximation, and the asymptotic part is solid: the W2 and TV formulations, the extension to multivariate and general polynomial-moment families, the Hahn-Jordan device for TV, and the Carleman bounds for Gaussian, Poisson, and exponential mixtures are all genuinely useful. The authors deserve credit for these.\n\nThe load-bearing problem is the finite-convergence claim. Theorem 3.2 (and 3.3) says that rank-flatness of M_d(phi*) alone certifies finite convergence and lets you extract the optimal K-component mixture. That is not just missing a proof; it is false. The proof shows phi* is K-atomic, but never shows the full pseudo-coupling lambda* has a representing measure with the correct marginals. Flatness of the mixing measure does not produce a true coupling.\n\nA concrete counterexample: mu = 1/2(delta_0 + delta_1), Gaussian family with m in [0,1], sigma in [0.05,1], R=1. At d=1, phi* = delta_{(0.5,0.5)} gives rank M_1 = rank M_0 = 1, and the SDP has a feasible lambda with zero transport cost, so tau_1 = epsilon. But the true optimal mixture is two Gaussians near 0 and 1 with small variance, cost roughly 0.0025 + epsilon, far better than any single Gaussian. The rank condition holds, yet tau_1 is not tau_infty. Same defect applies to Theorem 3.3. The proof invokes Theorem A.1, but that theorem concerns accumulation points as d grows; a flat solution at one fixed d is not automatically an accumulation point.\n\nMinor issues: Lemma A.1 is stated for any polynomial-moment family but proven for the three named families; likely repairable. The experiments are preliminary and lack baselines, and the regularization parameters are heuristics. Those are secondary.\n\nWho should read it: people working on moment-SOS approaches to mixtures, particularly for W2/TV distances. The asymptotic hierarchy may well be correct and useful, but the finite-recovery result is the main selling point and it does not hold as stated. I would send this to a serious referee rather than desk-reject, because the core idea is worth fixing and the error is pinpointable; adding flatness/rank conditions on the coupling lambda* might repair the theorem.\n\nRecommendation: engage, but expect major revision.\n\nBest,","headline":"A useful asymptotic moment-SOS hierarchy for mixture approximation, but the advertised finite-convergence and recovery theorem is false as stated; the flatness condition on the mixing measure alone does not certify a true coupling.","tokens_in":24415,"tokens_out":4928,"would_cite":false,"duration_ms":39193,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C22","90C26","44A60","62H30","60E99"],"pacs":[],"model":"deepseek-v4-flash","headline":"A measure known only through finitely many moments can be optimally approximated by a parametric mixture, with a semidefinite hierarchy that converges — and sometimes outright recovers the mixture order and parameters.","keywords":["mixture models","semidefinite programming","moment relaxations","Wasserstein distance","total variation distance","flat extension","clustering","moment determinacy"],"falsifier":"Construct a target measure μ that is itself a known K-component mixture, solve the SDP relaxation at order d, and check whether the flatness condition on M_d(ϕ*) triggers while the extracted K parameters yield a distance strictly larger than the SDP lower bound. That outcome would falsify the recovery claim. A more direct search: find a case where flatness holds on M_d(ϕ*) but M_d(λ*) (the coupling pseudo-moment matrix) admits no representing measure — then the claimed best approximation is not realized.","tokens_in":23277,"feed_emoji":"🧮","tokens_out":8472,"duration_ms":52653,"temperature":0.7,"pith_summary":"This paper studies the problem of approximating a probability measure, observed only through its moments up to some finite order, by a mixture of distributions from a parametric family such as Gaussians, Poissons, or exponentials. The approximation quality is measured by the squared 2-Wasserstein distance or the total variation distance, and the parameter set is not assumed finite; instead it is a compact semialgebraic set. The authors construct a hierarchy of semidefinite relaxations whose optimal values increase to the true optimal distance, and prove that when a rank-flatness condition holds on the optimal mixing pseudo-moment matrix, the convergence is finite, the mixing measure is K-atomic, and the K component parameters can be extracted by linear algebra. They also show that these estimates can initialize k-means or EM to reduce iteration counts and variance in clustering tasks.","feed_headline":"Semidefinite relaxations provably find the best mixture from moments","feed_subtitle":"Guaranteed convergence, and under a flatness condition, exact recovery of the mixture order and parameters.","key_machinery":"The load-bearing object is the pseudo-moment matrix M_d(ϕ) of the mixing measure ϕ, together with the flat-extension criterion rank M_d(ϕ*) = rank M_{d−d_min}(ϕ*) = K, which certifies that the truncated moment sequence comes from a K-atomic measure on the parameter set. The polynomial moment assumption — that every moment of the mixture is a polynomial in the parameters — converts mixture moments into linear functionals of ϕ and makes the problem a generalized moment problem amenable to a semidefinite hierarchy. For the total variation case, the additional device is the decomposition of the signed measure μ−ν into two positive measures bounded respectively by μ and ν; those domination constr","core_discovery":"The central claim is that the infinite-dimensional problem of minimizing a distance between a target measure and a mixture from a parametric family becomes a tractable semidefinite program once the target is represented by its moments up to order 2d, provided the family's moments are polynomials in the mixture parameters. The mixture moments are then linear functionals of the mixing measure, making the whole problem an instance of the generalized moment problem. The paper proves that as the relaxation order d increases, the optimal value converges to the true distance; if the optimal pseudo-moment matrix of the mixing measure satisfies the flat-extension condition (its rank at order d equals","pith_inferences":["The same moment-relaxation logic could be applied to other distances that admit moment-based lower bounds, such as energy distances or maximum mean discrepancy, as long as the distance can be certified by polynomial constraints.","The flatness condition used here is only checked on the mixing measure; testing whether joint flatness on the coupling (W2 case) or on the TV components (ψ+*, ψ−*) is necessary would clarify the actual robustness of the extraction step under numerical perturbation.","The paper's univariate projection experiments suggest that a systematic study of random projections could turn this method into a practical dimension-agnostic preprocessing routine for large-scale clustering."],"forward_implications":["When the flatness condition holds, the mixture order K and the component parameters can be read off the optimal solution without any a priori bound on the number of components.","The framework applies to any parametric family with polynomial moments and compact semialgebraic parameter sets, including Gaussian, Poisson, and exponential mixtures.","The recovered estimates can serve as high-quality initializations for k-means and EM, reducing both iteration counts and variance, as demonstrated on random Gaussian mixtures and MNIST.","For the total variation distance, asymptotic convergence holds even without flatness, and any accumulation point of the optimal mixing measures is an optimal (not necessarily atomic) mixture."],"fun_headline_variants":["Semidefinite relaxations find optimal mixtures from moments","Optimal mixture via semidefinite hierarchy from moment data","Guaranteed convergence to best mixture fit from moments","Exact recovery under flatness: SDP for mixtures","Mixture fitting solved by moment-sum-of-squares"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proofs of finite convergence rely on the flatness condition on the mixing measure alone (rank M_d(ϕ*) = rank M_{d−d_min}(ϕ*)) being enough to extend the entire truncated solution to a true coupling or to true dominating measures; no flatness on those auxiliary variables is established, so if that extension fails, the recovery guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["Semidefinite relaxations find optimal mixtures from moments","Optimal mixture via semidefinite hierarchy from moment data","Guaranteed convergence to best mixture fit from moments","Exact recovery under flatness: SDP for mixtures","Mixture fitting solved by moment-sum-of-squares"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1298,"prompt_tokens":709,"completion_tokens":589,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":525}},"tokens_in":453,"tokens_out":589,"duration_ms":20361,"temperature":1.0,"reasoning_tokens":525,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T14:47:03.119477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a target measure μ that is itself a known K-component mixture, solve the SDP relaxation at order d, and check whether the flatness condition on M_d(ϕ*) triggers while the extracted K parameters yield a distance strictly larger than the SDP lower bound. That outcome would falsify the recovery claim. A more direct search: find a case where flatness holds on M_d(ϕ*) but M_d(λ*) (the coupling pseudo-moment matrix) admits no representing measure — then the claimed best approximation is not realized.","supporting_citations":[],"review_version":1}