{"id":"cc426375-ca95-413d-a861-40718677d22d","arxiv_id":"2607.10205","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Anisotropic CBO's computational complexity depends exponentially only on the intrinsic dimension of an additively separable objective, not the ambient dimension, under aligned anisotropic noise.","lead":"Anisotropic consensus-based optimization can break the curse of dimensionality when the objective splits into lower-dimensional additive pieces. Complexity then scales with the largest piece size, not the full ambient dimension, if the noise axes match the split.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified; the independence-plus-alignment premise is already the paper's own explicit caveat and is correctly scoped by the theorems.","rationale":"The strongest claim is a conditional complexity statement: under additive separability (Assumption 1.1), component-wise tractability (Assumption 2.10), regularity (Assumption 2.5), and the independence of the projected initials, anisotropic CBO attains an ℓ^∞-error ≤ ε_total with probability at least (1-δ_{1})^L - δ_{2}, where the exponential dependence sits only on the intrinsic dimension d. The paper supplies a complete chain (separation \to mean-field convergence with α_{0} independent of D \to high-probability finite-particle bound) and openly flags the alignment requirement both in the abstract and in the numerical section. The reader's identification of that requirement as the weakest assumption is therefore accurate, yet it does not constitute an internal inconsistency or an unstated gap; it is the natural domain of the result. Consequently the ACCEPT verdict stands and no adjustment is warranted.","tokens_in":48095,"tokens_out":563,"duration_ms":7774,"concrete_test":"Independently re-derive the product structure of the success probability in the proof of Theorem 3.8 (display (55)) starting from the decoupled SDEs (38) without invoking the path-wise factorization (42); if the product form fails to appear, the high-probability bound is not justified by the mean-field separation alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption (statistical independence of the projected initial laws ρ^ℓ_0 together with coordinate-wise alignment of the anisotropic noise to the partition {I_ℓ}) is precisely the hypothesis required for the mean-field decoupling in Theorem 3.3 and for the subsequent complexity claims in Theorems 3.5/3.8 and Corollary 3.9. The paper states the hypothesis explicitly, proves that the path-wise laws factorize only under it (Eqs. (41)–(42)), and shows numerically that success collapses when the rotation misaligns the structure (Figs. 1, 5, 7). Because the central claim is conditioned on these hypotheses, the premise does not undermine the theorems; it merely delimits their applicability. No hidden gap in the derivation of α_0 ~ O(d^{2} log d) or of the (1-δ_{1})^L success probability was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies anisotropic consensus-based optimization (CBO) for nonsmooth nonconvex objectives that are additively separable into lower-dimensional components (Assumption 1.1). It proves that the mean-field anisotropic dynamics separate into independent lower-dimensional McKean processes under statistical independence of the projected initial measures (Theorem 3.3), yielding mean-field global convergence of the Lyapunov functional V_\\infty with hyperparameters \\alpha_0 = max_\\ell \\alpha_{0,\\ell} that depend only on the intrinsic dimension d = max d_\\ell and the component-wise inverse-continuity constants of Assumption 2.10 (Theorem 3.5). Combining this with the strong mean-square approximation of [16] produces a high-probability \\ell_\\infty bound for the practical discrete finite-particle scheme whose particle number N and consensus parameter scale with d (not ambient D) while the time-step contribution depends at most linearly on D (Theorem 3.8, Corollary 3.9). Supporting numerics on (block-rotated) Rastrigin functions illustrate the predicted degradation with increasing d and with increasing non-convexity inside the blocks.","tokens_in":48382,"tokens_out":957,"duration_ms":17864,"significance":"The result is a genuine advance in the rigorous analysis of multi-agent derivative-free methods: it supplies the first explicit complexity statement showing that anisotropic CBO can break the curse of dimensionality when the objective admits an aligned additive structure. The component-wise quantitative Laplace principle (Proposition 4.7), the refined tractability assumptions, the explicit scaling \\alpha_0 ~ O(d^{2} log d), and the product success probability (1-\\delta_1)^L are technically clean and useful. The derivation is self-contained once the independence/alignment hypotheses are granted, builds carefully on the existing mean-field and strong-convergence literature, and is corroborated by systematic heat-map experiments. The open problem of learning the noise axes on the fly is clearly motivated by the same analysis.","major_comments":[{"comment":"Theorem 3.3 establishes path-wise factorization only for the mean-field laws under the independence hypothesis on {\\rho_0^\\ell}. The subsequent high-probability bound of Theorem 3.8 (and the product structure (1-\\delta_1)^L) therefore inherits this mean-field factorization via independent copies of the McKean process. While the paper correctly flags that finite-N trajectories do not separate, a quantitative estimate of the rate at which the empirical measure approaches the product structure (or a fully microscopic argument under the same hypotheses) would make the complexity claim for the implementable algorithm more self-contained.","section":"Theorem 3.3, proof of Theorem 3.8 (display (53)–(55))"},{"comment":"The linear-in-D factor inside C_NA (arising from the anisotropic diffusion) appears in the admissible step-size of Corollary 3.9. Although this is only polynomial and does not restore an exponential curse, the manuscript never isolates its precise dependence; a short remark quantifying the constant would clarify that the ambient dimension truly enters only mildly.","section":"Corollary 3.9 and the citation to [16, Theorem 2.4]"}],"minor_comments":[{"comment":"The informal statement of Theorem 1.2 and the rigorous Corollary 3.9 use slightly different normalizations of \\epsilon; a single consistent expression would improve readability.","section":"Theorem 1.2 / Corollary 3.9"},{"comment":"Figures 1, 5, 7 and 9–10 are informative but the color-scale legends and exact success criterion (|E(x_\\alpha)-E*| < 0.25) are stated only in the text; placing them in the captions would help.","section":"Section 5"},{"comment":"Notation switches between d and bold d for the intrinsic dimension; a uniform choice would avoid momentary confusion.","section":"Throughout"}],"recommendation":"accept","confidential_remarks":"The manuscript is a solid, carefully written contribution that fits well in a top optimization or applied-analysis journal. The independence-plus-alignment hypothesis is stated with complete honesty and is already the paper’s own caveat; I see no hidden circularity. I would accept essentially as is."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The new result is clean: under additive separability (Assumption 1.1) and independent projected initials, the anisotropic mean-field dynamics fully decouples (Thm 3.3), so the high-probability finite-particle bound (Thm 3.8 / Cor 3.9) has exponential cost only in the intrinsic dimension d = max d_ℓ and the component-wise inverse-continuity constants, not ambient D. α0 is max of the per-block α0,ℓ; N is polynomial in 1/ε²δ with the usual exp(α) factor; Δt/H pick up at most linear D through C_NA. That is a genuine improvement over the general-objective analyses (Fornasier–Klock–Riedl, Riedl thesis, the strong-convergence companion) that paid ambient-dimension cost in α0.\n\nThe math is coherent. Well-posedness, pathwise separation via Grönwall + consensus stability, Lyapunov on V_∞, quantitative Laplace per block (Prop 4.7), and the high-probability lift from the earlier strong-convergence paper all fit together without circular fitting. The numerics (rotated Rastrigin heatmaps) do exactly what they should: success collapses when the rotation misaligns the blocks with the coordinate noise, and worsens with larger component non-convexity—matching the theory’s dependence on d and the local constants.\n\nThe soft spot is real but already scoped by the authors: everything hinges on statistical independence of the projected initials and on the anisotropic noise axes lining up with the partition {I_ℓ}. If the true separability is rotated relative to the coordinates, the decoupling fails and you are back to ambient-D cost. They state this, prove the factorization only under it, and show the collapse numerically. No hidden gap; just a clear applicability boundary. The “learn the alignment on the fly” remark is left for future work, which is honest.\n\nThis is for people who already work on CBO / mean-field particle methods or who care about derivative-free high-dimensional optimization with modular structure. The proofs are careful, the citations are the right ones, and the claim is supported under the stated hypotheses. I would send it to referees without hesitation; it is a proper incremental advance, not a re-packaging.","headline":"Solid, correctly scoped theory paper: anisotropic CBO complexity tracks intrinsic d (not ambient D) under additive separability + coordinate alignment; the independence caveat is already the authors' own.","tokens_in":48995,"tokens_out":581,"would_cite":true,"duration_ms":7596,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65K10","90C26","90C56","35Q90","35Q84"],"pacs":[],"model":"grok-4.5","headline":"Anisotropic CBO finds global minima of additively separable high-dimensional objectives with complexity that depends only on the intrinsic dimension, not the ambient one.","keywords":["consensus-based optimization","anisotropic diffusion","additive separability","curse of dimensionality","mean-field limit","global nonconvex optimization","derivative-free methods"],"falsifier":"Take a fully separable Rastrigin function, apply a dense orthogonal rotation that mixes all coordinates, and check whether the empirical success probability of anisotropic CBO still scales only with the original one-dimensional blocks or collapses to the ambient-dimension scaling.","tokens_in":48991,"feed_emoji":"📐","tokens_out":656,"duration_ms":6555,"temperature":0.7,"pith_summary":"High-dimensional global optimization is usually cursed by dimension: cost grows exponentially with the number of variables. This paper shows that anisotropic consensus-based optimization (CBO)—a swarm of particles that mix a drift toward a consensus point with coordinate-wise random exploration—automatically exploits additive separability. When the objective is a sum of lower-dimensional pieces living on disjoint coordinate blocks, the particle system decouples into independent lower-dimensional problems. Consequently the number of particles and the temperature parameter needed for high-probability global convergence grow only with the largest block size (the intrinsic dimension), not with the full ambient dimension. The same analysis also shows that only the local geometry of each block matters, so a single poorly conditioned coordinate does not force the whole algorithm to pay the worst-case cost. Numerical heat-maps confirm that success rates collapse once the blocks are rotated away from the coordinate axes, underscoring that the noise must be aligned with the separability structure.","feed_headline":"CBO complexity tracks intrinsic dimension, not ambient size","feed_subtitle":"When an objective splits into low-dim blocks, anisotropic noise lets the swarm ignore the rest of the variables","key_machinery":"Mean-field decoupling: under additive separability and independent initial projections, the anisotropic Fokker–Planck equation factors into L independent lower-dimensional McKean–Vlasov equations; the quantitative Laplace principle then applies block-wise, producing an α0 that is the maximum of the block-wise thresholds.","core_discovery":"For an objective that is additively separable into L lower-dimensional components of dimensions at most d, anisotropic CBO converges in high probability to the unique global minimizer with particle number and inverse-temperature that scale only with d and with the tractability constants of the individual components, independent of ambient dimension D.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Anisotropic CBO complexity scales only with intrinsic dimension d","Additive separability lets anisotropic CBO ignore ambient dimension D","Particle count and inverse temperature depend solely on component dim d","Anisotropic noise exploits additive structure to beat high-D curse","CBO cost set by tractability of low-dim blocks not full landscape"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The coordinate-wise noise of the algorithm must line up with the unknown partition of the variables; if the true separable blocks are rotated relative to the axes, the decoupling fails and the dimension-reduction guarantee disappears.","fun_headline_variants_meta":{"raw":{"variants":["Anisotropic CBO complexity scales only with intrinsic dimension d","Additive separability lets anisotropic CBO ignore ambient dimension D","Particle count and inverse temperature depend solely on component dim d","Anisotropic noise exploits additive structure to beat high-D curse","CBO cost set by tractability of low-dim blocks not full landscape"]},"model":"grok-4.5","effort":"low","cost_usd":0.00609,"raw_usage":{"total_tokens":1625,"prompt_tokens":819,"num_sources_used":0,"completion_tokens":88,"cost_in_usd_ticks":60900000,"prompt_tokens_details":{"text_tokens":819,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":718,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":819,"tokens_out":88,"duration_ms":6030,"temperature":1.0,"reasoning_tokens":718,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T13:30:30.309159+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Take a fully separable Rastrigin function, apply a dense orthogonal rotation that mixes all coordinates, and check whether the empirical success probability of anisotropic CBO still scales only with the original one-dimensional blocks or collapses to the ambient-dimension scaling.","supporting_citations":[],"review_version":1}