{"id":"a7b315d4-bef7-43ff-bec4-a225a2869a58","arxiv_id":"2602.10691","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For Gaussian distributions, slice-matching to an isotropic target with decaying step sizes converges at rate O(k^{-(2α-1)}) in expectation.","lead":"This paper proves explicit convergence rates for the sliced optimal transport 'slice-matching' algorithm when the source and target are Gaussians. It relies on new Polyak-Łojasiewicz inequalities and shows that sampling full random orthonormal bases stabilizes the eigenvalue spectrum.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop. D.2's spectral-control proof treats the extremal direction θ_i as Haar-uniform; for Λ=I, θ_i maximizes θ_i^TΣ_kθ_i, so E[θ_i^TΣ_kθ_i] > Tr(Σ_k)/d, invalidating the moment bounds that Theorem 4.1 relies on.","rationale":"The reader's verdict identified the non-isotropic target as the weakest assumption, but the more fundamental issue is that the proof of the main theorem for the isotropic case contains a concrete error: the direction θ_i in the eigenvalue recursion is not a uniformly random direction but the extremal one, chosen by maximizing θ^TΣθ. The proof's replacement of its conditional expectation by the isotropic average Tr(Σ)/d is invalid. This breaks the inverse-moment bound (49) and its p-th-order extension (50), which are needed to verify Assumption A and to conclude Theorem 4.1. Consequently, the central claim is unproven as written. The error is not a minor off-by-one or a missing assumption; it is a false step in the core argument. A simple numerical check with d=2 and a covariance satisfying the paper's trace constraint demonstrates the flaw. Even if the theorem might be true by a different argument, the present proof does not establish it. For a paper whose main contribution is this convergence rate, the proof gap warrants rejection rather than conditional acceptance. The reader's concern about non-isotropic targets is legitimate but secondary; the isotropic case itself is not yet proven.","tokens_in":32985,"tokens_out":15320,"duration_ms":124092,"concrete_test":"Implement the exact covariance recursion (18) for d=2, Λ=I, Σ_0 = diag(1.8, 0.1), γ_k = 0.5. For a fixed covariance Σ (e.g., after one step), sample many Haar-random orthonormal bases P, define θ_i(P) = argmax over columns of θ^TΣθ, and compute the empirical mean of θ_i^TΣθ_i. If this mean exceeds Tr(Σ)/d (it will be ≈1.49 vs 0.95 for the stated Σ), the identity used in Proposition D.2 is refuted. A stronger check: compute E[1/λ_min(Σ_{k+1})] − E[1/λ_min(Σ_k)] under the exact recursion over several steps; if it is positive for any k, the claimed contraction (49) fails, breaking the moment bound that feeds Assumption A and Theorem 4.1.","verdict_should_be":"REJECT","load_bearing_attack":"Proposition D.1's lower bound is √λ_min(Σ_{k+1}) ≥ √λ_min(Σ_k)(1−γ_k+γ_k τ_{θ_i}) with τ_θ = sqrt(θ^TΛθ / θ^TΣθ). For the isotropic target Λ=I, τ_θ = 1/√(θ^TΣ_kθ), so the smallest τ — hence the valid θ_i — is the column of P_{k+1} maximizing θ^TΣ_kθ. In Proposition D.2's proof, the conditional expectation of θ_i^TΣ_kθ_i is computed as E_θ[θ^TΣ_kθ] = Tr(Σ_k)/d, which is only correct for a direction drawn uniformly from the sphere independent of Σ_k. θ_i is the argmax over d orthonormal columns and is therefore not uniform and not independent of Σ_k. For example, with d=2 and Σ_k = diag(1.8, 0.1) (so Tr(Σ_k)=1.9 ≤ d), E[max_{columns} θ^TΣ_kθ] = 0.95 + 0.85·E|cos 2φ| = 0.95 + 0.85·(2/π) ≈ 1.49, which exceeds Tr(Σ_k)/d = 0.95. Thus the claimed inequality E[(λ_min(Σ_{k+1}))^{-1} | A_k] ≤ (λ_min(Σ_k))^{-1} does not follow, nor do the p-th moment bounds (49)–(50). Since these moment bounds are the sole mechanism for verifying Assumption A in the Gaussian isotropic case, the proof of Theorem 4.1 is not valid as written. The auxiliary Lemma D.3 is also false (for A = 0.5I, d=2, its RHS for p=2 is λ_max(A)−1 = −0.5 while the LHS is 0.25), compounding the issue.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the slice-matching scheme (IDT with random orthonormal bases) for distribution matching in Wasserstein space. It derives a decomposition of the Sliced-Wasserstein gradient into variance and norm terms, establishes static Łojasiewicz/PL-type inequalities for bounded-density measures and for Gaussian measures with co-diagonalizable covariances, and proves a polynomial-dimension comparison W_2^2 ≤ C_{d,m,M} SW_2^2 in that Gaussian case. The main advertised result is Theorem 4.1: for a centered Gaussian source and an isotropic Gaussian target, with step sizes γ_k=(k+1)^{-α}, 1/2<α<1, the expected SW loss decays as O(k^{-(2α-1)}). The proof strategy is to verify a random PL condition (Assumption A) by controlling inverse moments of λ_min(Σ_k) along the covariance recursion, then applying a Chung-type lemma.","tokens_in":33439,"tokens_out":16065,"duration_ms":136788,"significance":"If the proof were valid, the result would be a useful quantitative non-asymptotic convergence rate for a practical sliced-OT algorithm, with explicit dependence on dimension and step-size, and the W2–SW2 comparison for Gaussians is of independent interest. The paper is also transparent about the isotropic-target restriction and provides reproducible numerical experiments. However, the moment control of λ_min is the linchpin of the proof, and the present argument contains a selection-bias error and a false lemma. Since the advertised rate is not established as written, the contribution is conditional on a correct spectral-control proof.","major_comments":[{"comment":"The conditional-expectation step is invalid because the direction θ_i is selected, not independent. In (45), θ_i is the column of P_{k+1} minimizing θ^TΛθ/θ^TΣ_kθ; when Λ=I it maximizes θ^TΣ_kθ among the d columns. After (51), the proof replaces E[θ_i^TΣ_kθ_i|A_k] by E_θ[θ^TΣ_kθ]=Tr(Σ_k)/d, which holds only for a fixed Haar-uniform θ independent of Σ_k. For d=2, Σ_k=diag(1.8,0.1), E[max_j θ_j^TΣ_kθ_j]≈1.49>Tr(Σ_k)/d=0.95, and can exceed 1. Thus E[(λ_min(Σ_{k+1}))^{-1}|A_k]≤(λ_min(Σ_k))^{-1} does not follow; neither do (49)–(50). Since these moment bounds are the sole verification of Assumption A for the isotropic Gaussian case, Theorem 4.1 is not proved.","section":"Proposition D.2 (Eq. (51))"},{"comment":"The second inequality in Lemma D.3 is false as stated. For A=0.5I and d=2, the left-hand side at p=2 is E[(θ^TAθ−1)^2]=0.25, while the right-hand side is 1+∑_{r=1}^2 C(2,r)λ_max(A)^{r−1}(−1)^r = 1−2+0.5 = −0.5, a negative upper bound. The lemma is used in the proof of Proposition D.2 to bound E[(θ_i^TΣ_kθ_i−1)^p]; the statement needs correction even though a crude finite bound for fixed A is trivial.","section":"Lemma D.3"}],"minor_comments":[{"comment":"The constant d^2(d+2)/λ_min(Σ_k)^2 does not appear to follow directly from Proposition 4.4 with the stated bounds; the derived constant seems to involve an extra factor of order d. This does not affect the rate, but the calculation should be checked.","section":"Appendix E.1, Eq. (54)"},{"comment":"Please state explicitly that θ_i and θ_j are the extremal columns arising in the Ostrowski bound. The current wording 'there exist directions' obscures the selection dependence that is central to the flaw in Proposition D.2.","section":"Proposition D.1"},{"comment":"The inequality F(σ) ≤ 2C_ν(M/m)∥∇W2F(σ)∥_σ is a Łojasiewicz inequality with exponent 1/2, not a PL inequality of the form used in Assumption A. Calling it 'PL-like' may confuse the two notions.","section":"Proposition 4.3"},{"comment":"Minor typos and reference issues: 'weigths' in Proposition 3.3; Li and Moosmüller (2023) should give the arXiv identifier; the Sisouk et al. entry lacks full publication data.","section":"Throughout"},{"comment":"The abstract says convergence for 'Gaussian distributions', but Theorem 4.1 is specifically for an isotropic target. The conclusion is transparent about this, but the abstract should be aligned to avoid overclaiming.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The central issue is in the spectral-control argument. I do not question the paper's novelty or the authors' fairness; the isotropic limitation is disclosed. However, Proposition D.2 is load-bearing and, as written, wrong. I would ask the authors to either supply a correct proof of the inverse-moment control or explicitly downgrade Theorem 4.1 to a conditional result. The paper may be salvageable, but the advertised rate is currently unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the rate may be true, but the proof as written does not establish it. The paper's framework — random PL constants, the covariance recursion, the polynomial W2/SW2 comparison for co-diagonalizable Gaussians — is genuinely useful. But the eigenvalue-control step that makes Theorem 4.1 work has a load-bearing flaw.\n\nThe issue is in Proposition D.2. The direction θ_i that attains the lower bound in D.1 is the column of P that minimizes τ, which for Λ=I means it maximizes θᵀΣ_kθ. That direction is not Haar-uniform and is not independent of Σ_k. So E[θ_iᵀΣ_kθ_i] is not Tr(Σ_k)/d; it is larger. The stress-test example is correct: for d=2, Σ=diag(1.8,0.1), the expected maximum is about 1.49, not 0.95. Consequently the recursive bound on 1/λ_min does not follow, and neither do the moment bounds that verify Assumption A. Lemma D.3 is also simply false for p=2 when λmax<1, since the claimed upper bound is negative. This is not cosmetic; it is the mechanism that controls the PL constants along the trajectory.\n\nWhat the paper does well: the proof strategy is modular and clear — gradient-variance decomposition, static PL inequality for Gaussians, then trajectory control. Proposition C.2, the polynomial W2/SW2 comparison for diagonal Gaussians, looks correct up to constants and is of independent interest. The experiments illustrate dimension dependence and the benefit of orthonormal basis sampling, though they lack error bars and a pinned code commit. The authors honestly state the isotropic-target limitation, and the general-Λ condition (21) is left unverified.\n\nMinor issues: the γ1/γ0 indexing typo in Proposition B.8 is annoying but fixable; the non-isotropic case is honestly flagged as open.\n\nMy verdict: not acceptable as-is, but worth a serious referee. The gap is technical rather than conceptual — a repaired spectral control argument might well salvage the theorem. I'd send it to reviewers with a clear request to check the extremal-direction dependence in D.2, and I would not cite the main rate until that is fixed.","headline":"The main theorem is not proven as written: Prop. D.2 treats a covariance-dependent extremal direction as if it were Haar-uniform, and Lemma D.3 is plainly false; the framework is still worth engaging.","tokens_in":33976,"tokens_out":5749,"would_cite":false,"duration_ms":55338,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","65K10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The slice-matching scheme provably converges to an isotropic Gaussian target with expected loss O(k^{1−2α}).","keywords":["Sliced-Wasserstein distance","distribution matching","slice-matching scheme","Polyak–Łojasiewicz inequality","stochastic gradient descent","Gaussian measures","non-asymptotic convergence rates"],"falsifier":"Run the explicit covariance recursion (18) for a non-isotropic Gaussian target with a fixed α and track E[λ_min(Σ_k)^{−p}] over many random seeds; if this quantity diverges with k for any p ≥ 1, then condition (21) fails and Theorem 4.1 cannot be extended to that target. Equivalently, for the isotropic case, a mismatch between the empirical slope of log E[F(σ_k)] and 1 − 2α over k up to about 10^4 would signal that the claimed rate, or its dimension constant, does not hold as stated.","tokens_in":32857,"feed_emoji":"🎯","tokens_out":6590,"duration_ms":65969,"temperature":0.7,"pith_summary":"This paper derives quantitative non-asymptotic convergence rates for the slice-matching scheme, an iterative method that matches distributions by repeatedly applying one-dimensional optimal transport maps along a random orthonormal basis. The central result is that when both the source and the target are Gaussian and the target is isotropic, the expected Sliced-Wasserstein objective decays as C/k^{2α−1} for step sizes γ_k = (k+1)^{−α} with 1/2 < α < 1. The proof works by showing that the Sliced-Wasserstein objective satisfies a Polyak–Łojasiewicz inequality along the trajectory, with random constants controlled through the smallest eigenvalue of the current covariance. A key ingredient is the choice of an orthonormal basis at each step: it matches first and second moments and keeps inverse eigenvalues bounded in expectation. The rate matters because it turns an empirically successful but theoretically poorly understood heuristic into a provably convergent stochastic optimization method.","feed_headline":"Slice matching hits a Gaussian target at polynomial speed","feed_subtitle":"Expected error decays as C/k^(2α−1); orthonormal-basis sampling is what keeps the proof's constants under control.","key_machinery":"The load-bearing objects are the slice-matching map T_{σ,P}(x) = Σ_i t_{θ_i}(θ_iᵀx)θ_i acting along the columns of a Haar-random orthonormal basis P, and the covariance recursion Σ_{k+1} = ((1−γ_k)I + γ_k P D_k Pᵀ) Σ_k ((1−γ_k)I + γ_k P D_k Pᵀ)ᵀ with D_k = diag(√(θ_iᵀΛθ_i / θ_iᵀΣ_kθ_i)). The proof combines three ingredients: the gradient-variance decomposition 2F = ‖∇F‖²_σ + E‖T̄ − T_P‖²_σ; a Polyak–Łojasiewicz inequality F(σ) ≤ d²(d+2) / λ_min(Σ)² · ‖∇F‖²_σ along the iterates; and a recursion bounding λ_min(Σ_{k+1}) from below by λ_min(Σ_k) times a factor depending on one direction's ratio τ_θ. Orthonormal-basis sampling is what makes the drift term E[θᵀΣ_kθ − 1] ≤ 0, giving uniform-in-k mo","core_discovery":"Writing the slice-matching update as a stochastic gradient step in Wasserstein space, the authors prove Theorem 4.1: if σ = N(0, Σ) and μ = N(0, Id), then for γ_k = (k+1)^{−α}, 1/2 < α < 1, there is C > 0 with E[F(σ_k)] ≤ C / k^{2α−1}, where F(σ) = (d/2) SW²(σ, μ). The argument decomposes the objective into squared gradient plus variance, obtains a Polyak–Łojasiewicz inequality whose constant is proportional to 1/λ_min(Σ_k)², and then controls inverse moments through an eigenvalue recursion derived from the explicit covariance update. The crucial step is that isotropy of the target makes the conditional drift E[θᵀΣ_kθ − 1 | A_k] = Tr(Σ_k)/d − 1 ≤ 0, so λ_min(Σ_k)^{−p} has bounded expectation","pith_inferences":["If the claimed rate extends to general Gaussian targets, the sufficient condition (21) must hold; a natural test is to run the explicit covariance recursion with a non-isotropic Λ and check whether E[λ_min(Σ_k)^{−p}] stays bounded for all p. The paper's own numerical experiments hint that it does, but no proof is given.","The theorem is silent on the empirically best regime α ≤ 1/2, including α = 0 (the original constant-step algorithm). A plausible reading, which the authors flag as open, is that the polynomial bound captures only the tail of the convergence and that small-α behavior is governed by a different, perhaps dimension-dependent constant.","The event-based handling of random PL constants may transfer to other distribution-matching schemes (for instance, ones with diffusive regularization) as long as one can control the analogous random constants in expectation; if a similar spectral drift inequality holds, the same rate template would apply.","The single-direction variant's lack of inverse-moment control suggests that orthogonality is not merely an implementation detail: without it, worst-case drift can accumulate in low-eigenvalue directions, so one would expect strictly worse worst-case rates. This is an inference from the paper's analysis, not a proved statement."],"forward_implications":["For an isotropic Gaussian target and a Gaussian source, with step sizes 1/(k+1)^α and 1/2 < α < 1, the expected Sliced-Wasserstein loss satisfies E[F(σ_k)] ≤ C/k^{2α−1}; convergence to the target is guaranteed at a polynomial rate.","The constants grow polynomially (roughly d²(d+2)) in the ambient dimension, so the bound predicts slower convergence in higher dimensions, which the reported experiments reproduce.","Sampling an orthonormal basis of directions at each step, rather than a single direction, is what allows inverse-eigenvalue moments to be controlled from the first iteration; the experiments show that this choice also suppresses fluctuations in the extreme eigenvalues.","Even without the Polyak–Łojasiewicz inequality, the stochastic-optimization viewpoint gives a convergence-to-critical-point guarantee at rate K^{α−1} for weighted average gradient norms, complementing earlier asymptotic results.","For co-diagonalizable Gaussian covariances with eigenvalues in [m, M], the new comparison W2² ≤ d(d+2)(M/m) SW2² replaces the usual exponential dimension dependence with a polynomial one."],"fun_headline_variants":["Polynomial rates for sliced OT matching on Gaussians","Orthonormal sampling secures slice-matching convergence rates","Gaussian slice-matching error decays as C/k^(2α−1)","Slice-matching gets provable polynomial speed on Gaussians","Sliced OT matches Gaussians with polynomial decay"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The quantitative rate rests on the target being the isotropic Gaussian Λ = Id; for Λ ≠ Id the paper only provides an unverified sufficient condition (Eq. 21), so the rate is not established beyond the isotropic case.","fun_headline_variants_meta":{"raw":{"variants":["Polynomial rates for sliced OT matching on Gaussians","Orthonormal sampling secures slice-matching convergence rates","Gaussian slice-matching error decays as C/k^(2α−1)","Slice-matching gets provable polynomial speed on Gaussians","Sliced OT matches Gaussians with polynomial decay"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001536,"raw_usage":{"total_tokens":5967,"prompt_tokens":712,"completion_tokens":5255,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":5182}},"tokens_in":456,"tokens_out":5255,"duration_ms":36379,"temperature":1.0,"reasoning_tokens":5182,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T01:00:55.706294+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the explicit covariance recursion (18) for a non-isotropic Gaussian target with a fixed α and track E[λ_min(Σ_k)^{−p}] over many random seeds; if this quantity diverges with k for any p ≥ 1, then condition (21) fails and Theorem 4.1 cannot be extended to that target. Equivalently, for the isotropic case, a mismatch between the empirical slope of log E[F(σ_k)] and 1 − 2α over k up to about 10^4 would signal that the claimed rate, or its dimension constant, does not hold as stated.","supporting_citations":[],"review_version":1}