{"id":"e5e89f02-664c-48dd-9708-2364aedd4275","arxiv_id":"2501.06804","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A smoothed consensus-based optimization algorithm is shown to reach almost-sure consensus and, under a sufficient condition on initial data and parameters, to land near the global minimum of nonsmooth nonconvex functions.","lead":"This paper modifies consensus-based optimization (CBO), a swarm algorithm, by replacing the objective function with a smooth approximation, and proves that the particle system reaches consensus and can land near the global minimum of nonsmooth, nonconvex objectives. The appeal is a finite-particle convergence proof that avoids the usual mean-field limit.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.2 relies on an unproved deterministic bounded set X in equation (39); the smoothing constants κ and η are path-dependent, so the error estimates do not follow as written.","rationale":"The paper's central novel contribution is the finite-particle error estimate in Theorem 4.2, and that theorem rests on equation (39) providing a deterministic bounded set X on which the smoothing constants κ, η and the lower bound \\tilde f_min are controlled. The proof of Theorem 4.1 establishes only almost-sure convergence to a random limit, which does not imply such a deterministic set; and the explicit formula (23) shows that even with bounded initial data the paths have positive probability of escaping any fixed ball, due to the Brownian exponent. Thus (39) is not a minor technicality: without it, the inequalities (40)-(43) do not follow because κ, η and \\tilde f_min become path-dependent random variables that cannot be factored out of the expectations. The reader's weakest assumption identifies exactly this gap, and I agree. The consensus results (Theorem 3.1 and Theorem 4.1) are independent and appear correct. Corollary 4.1 has the further issue that its proof chooses the initial distribution xin to satisfy (59), whereas the corollary is stated for any given initial data under Assumptions 1-3; this makes the corollary over-stated. Both issues are likely repairable by adding bounded-support or appropriate finite-moment assumptions and by restating Corollary 4.1 existentially. The verdict CONDITIONAL remains appropriate: the central error estimate should be revised, but the paper contains substantial correct material. Hence verdict_should_be UNCHANGED.","tokens_in":25889,"tokens_out":9656,"duration_ms":99222,"concrete_test":"Set d=1, N=2, λ=1, σ=1, x1(0)=0, x2(0)=1. Equation (23) gives x1(t)-x2(t) = -exp(-(3/2)t + W(t)). Since P(sup_{t≥0}(W(t)-1.5t) > log 10) = 10^{-3} > 0, with positive probability the particles are not contained in any fixed ball of radius 5, so no deterministic bounded set satisfies (39). Then independently re-derive (40)-(43) with X = X(ω) path-dependent: the constants κ(ω), η(ω), and \\tilde f_min(ω) cannot be moved outside the expectations, so condition (45) as stated is insufficient. If this re-derivation cannot be completed without an expectation version of (45), Theorem 4.2 is unproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2, equation (39) asserts that there exists a bounded set X containing xi(t) and xbar*(t) almost surely for all t. This is not established in Theorem 4.1, which only proves almost-sure convergence to a random limit; a.s. convergence does not yield a deterministic bound uniform in ω. Moreover, the claim is false even for bounded initial data: from (23) the coordinate differences are geometric Brownian motions, and for any R > 0 there is positive probability that exp(σW_l(t) - (λ+σ^2/2)t) exceeds R for some t, so paths can leave any fixed ball on a set of positive probability. Since Definition 2.2 supplies κ, η, and \\tilde f_min only on a bounded set, the estimates (40)-(43) require κ, η, and \\tilde f_min to be deterministic constants. If X is path-dependent, these quantities become random and cannot be pulled out of the expectations in (40)-(43). Theorem 4.2 is therefore not proved as stated, and Corollary 4.1 inherits the gap. A repair requires either a uniform almost-sure bound on paths (e.g., bounded-support initial data with a pathwise estimate) or a reformulation of (45) using expectations of random κ and η. The consensus results (Theorems 3.1 and 4.1) are not affected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SCBO, a finite-particle consensus-based optimization algorithm (Algorithm 1, Eqs. (14)-(15)) that replaces a nonsmooth, possibly non-Lipschitz objective f in (1) by a smoothing function tilde f(x, mu_t) with mu_t decaying to 0. It proves: (i) almost-sure global consensus for arbitrary initial data via an explicit geometric Brownian motion formula (23); (ii) existence of a common almost-sure limit x_infty (Theorem 4.1) under 2*lambda > sigma^2; (iii) an error estimate (Theorem 4.2) asserting ess inf f(x_infty) <= fmin + E(beta) with E(beta) -> 0 as beta -> infinity, provided condition (45) holds; and (iv) Corollary 4.1 claiming that for every delta > 0 parameters can be chosen to make E(beta) <= delta. Numerical experiments compare SCBO with a deterministic smoothing gradient method and with the CBO algorithm of [21].","tokens_in":26181,"tokens_out":15994,"duration_ms":150399,"significance":"If Theorems 4.2 and Corollary 4.1 were fully proved, the paper would make a meaningful contribution: it extends finite-particle CBO analysis to nonsmooth, non-Lipschitz objectives without passing to the mean-field limit, and it provides a quantitative error bound as beta -> infinity. The explicit GBM representation (23) is a clear strength, and the a.s. consensus theorem and the common-consensus-state theorem are supported by a direct, self-contained argument. The manuscript also gives reproducible numerical evidence that SCBO is competitive with existing methods. However, the central error estimate currently rests on an unproved uniform boundedness assertion, so the advertised error bound is not established as written.","major_comments":[{"comment":"Equation (39) asserts that there exists a bounded set X such that xi(t) and xbar*(t) lie in X almost surely for all t. This does not follow from Theorem 4.1, which only proves almost-sure convergence of xi(t) to a random limit x_infty. Almost-sure convergence gives a sample-path-dependent eventual bound, not a deterministic bounded set. Moreover, for any nonzero initial difference, Eq. (23) shows that xi_l(t) - xj_l(t) is a geometric Brownian motion with negative drift; the supremum of such a process over [0, infinity) exceeds any fixed level R with positive probability. Hence no deterministic bounded set can contain the paths for almost all omega, even if the initial data are bounded. Because Definition 2.2(iii) provides kappa and eta only on a given bounded set, the constants kappa and eta used in (40)-(43) become omega-dependent if X is path-dependent, and they cannot be factored out of the expectations in (42)-(43) as written. Consequently, the derivation of (46)-(52), and hence Theorem 4.2 and Corollary 4.1, is not justified. A repair requires either a genuine uniform-in-omega bound on the paths (for example by adding a projection or truncation to Algorithm 1), or a reformulation of condition (45) using expectations of the random quantities kappa, eta, and tilde f_min.","section":"Section 4.2, Eq. (39)"},{"comment":"The proof of Theorem 4.1 uses (33) to conclude that E[integral_0^t (xi_l(s) - xbar*_l(s))^2 ds] is finite, and then applies the martingale convergence theorem to the stochastic integral. For (33) to be finite, one needs E[max_{1<=i<=N, 1<=l<=d} (xi_l(0) - xbar_l(0))^2] < infinity. Assumption 2 only says that the initial data are i.i.d. with common law xin and does not impose any moment condition. Thus the proof of the common-consensus-state theorem does not cover initial laws with infinite second moment, although the abstract and Theorem 4.1 claim convergence 'with any initial data'. Please either add a finite second-moment hypothesis to Assumption 2 and Theorem 4.1, or provide a localization argument that avoids the expectation bound.","section":"Section 4.1, proof of Theorem 4.1 and Eq. (33)"}],"minor_comments":[{"comment":"The proof applies Ito's formula to ln x_ij_l(t) without treating the case x_ij_l(0)=0. If the initial difference is zero, the logarithm is undefined, but Eq. (23) remains valid by continuity (the difference is identically zero). Please handle this case separately before applying the log-transform.","section":"Section 3.2, proof of Theorem 3.1"},{"comment":"In the estimate for E[Q3], the displayed intermediate inequality omits the factor 1/2 that appears in the definition of Q3 in Eq. (37). The final bound is still valid because dropping the factor 1/2 yields a weaker (more negative) lower bound, but the displayed inequality should be derived explicitly to avoid confusion.","section":"Section 4.2, Eq. (43)"},{"comment":"The phrase 'a suitable xin' in the proof is misleading: the initial law should be part of the problem data, not a parameter to be selected. In fact, since e^{beta(tilde f_min - f(xin))} <= 1 and converges to 1 as beta -> 0, a sufficiently small beta works for any fixed xin; please reword the proof so that the quantifier over the given initial law is clear.","section":"Corollary 4.1, proof"},{"comment":"There are several typographical errors, including 'dose not' (should be 'does not') and broken spacing in 'focus es'. Please proofread the text.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the consensus half of this paper is a real step beyond Ha–Jin–Kim; the error theorem is not proved as stated, and the fix is doable.\n\nWhat's new: they prove a.s. consensus and existence of a common consensus state for a CBO variant where f is merely continuous and possibly non-Lipschitz, using a smoothing function and a finite-particle SDE with common noise. The explicit formula for coordinate differences (geometric Brownian motion) does the heavy lifting, and Theorems 3.1 and 4.1 are clean and correct. That is genuinely new: the only prior finite-particle proof required C_b^2 and fixed beta. The idea of letting the smoothing parameter decay in time and tracking the weighted mean is natural, but the analysis is nontrivial.\n\nThe soft spot is Theorem 4.2, equation (39). They assert there is a bounded set X containing all paths and xbar*(t) a.s. for all t, so the smoothing constants kappa and eta are uniform. But Theorem 4.1 only gives a.s. convergence to a random limit; it does not give a deterministic bounded set. The stress-test note is right: even with bounded initial data, the geometric Brownian factors make P(sup_t |x_i_l(t) - xbar_l(t)| > R) > 0 for every R, so paths can leave any fixed ball on a set of positive probability. The set X would have to be random, and then kappa and eta are random and cannot be pulled out of the expectations in (40)-(43). So the error estimate does not follow as written. This is the load-bearing part of the paper's main claim, so it has to be fixed before the result is usable. The repair is probably adding bounded-support or strong moment assumptions on the initial data and proving a uniform-in-omega bound, or alternatively carrying kappa and eta inside the expectations and reworking condition (45). Corollary 4.1 is also over-stated: the proof chooses the initial data distribution as well as the parameters, so it is an existential statement over initial data too, and the text should say so.\n\nMinor issue: the proof assumes finite second moments of the initial data (e.g., E[max_l (x_i_l(0)-bar_x_l(0))^2]) without stating it in Assumption 2.\n\nNet: the consensus theorems are solid and worth keeping; the error analysis is conditional and needs revision. I would send this to a serious referee—the gap is identifiable and probably fixable, and the finite-particle nonsmooth CBO result is worth having in the literature.\n\nFor peer review: major revision, with a referee who knows SDE estimates.","headline":"New finite-particle CBO convergence for nonsmooth objectives; consensus results are solid but the main error theorem has a genuinely flawed uniformity step.","tokens_in":26688,"tokens_out":3354,"would_cite":true,"duration_ms":33273,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","37N40","65K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a smoothing-based consensus optimization scheme converges almost surely to a common state for continuous nonconvex objectives and, under explicit parameter conditions, drives the objective value at that state…","keywords":["consensus-based optimization","smoothing method","nonsmooth nonconvex optimization","global consensus","error estimation","non-Lipschitz optimization","finite particle system","stochastic differential equations"],"falsifier":"Take a continuous, non-Lipschitz objective such as f(x) = |x|^p on R with p in (0,1), choose initial data and parameters satisfying Assumptions 1-3 and inequality (45), and simulate the SCBO scheme; if the empirical essential infimum of f at the consensus state exceeds fmin + E(β), or if a positive fraction of paths leaves every bounded set, the central guarantee fails. A direct check of whether condition (39) holds for this example would settle whether the proof's uniformity assumption is valid.","tokens_in":25674,"feed_emoji":"🎯","tokens_out":9767,"duration_ms":84699,"temperature":0.7,"pith_summary":"Consensus-based optimization (CBO) is a swarm method whose finite-particle convergence theory previously required smooth objectives or a mean-field limit. This paper proposes a smoothing variant, SCBO, that replaces the objective f with a smoothing function whose smoothing parameter decays to zero over time. It proves that, for any initial data, the particles converge almost surely to a common random state, and that under explicit drift, noise, and smoothing-schedule conditions the objective value at that state is bounded by fmin + E(β), with E(β) tending to zero as β tends to infinity. The result matters because it extends rigorous CBO convergence to continuous, nonconvex, possibly non-Lipschitz objective functions while working directly with the finite particle system.","feed_headline":"Smoothed swarm optimizer solves nonsmooth global problems","feed_subtitle":"Finite-particle consensus reaches any desired error to the global minimum, without a mean-field limit.","key_machinery":"The load-bearing object is the smoothing function f̃(x, µ) of Definition 2.2: twice continuously differentiable in x, converging pointwise to f as µ decreases to 0, and satisfying the bounds |∇_µ f̃(x, µ)| ≤ κ $µ^{{-q}}$ and ‖∇²_xx f̃(x, µ)‖ ≤ η $µ^{{-q-1}}$ on bounded sets. In the algorithm, f̃ enters the weights exp(-β f̃(x_i(t), µ_t)) that define the consensus target x̄*(t), making those weights smooth enough for Ito's formula even though f itself is nonsmooth. The mechanism that carries the proof is the explicit separation formula for particle coordinates, which gives exponential decay; the Ito-formula inequality for the empirical average of the smoothed weights, which turns smoothing-parameter decay into a lower bound; and the Laplace principle, which converts that lower bound into the error estimate for f(x∞).","core_discovery":"The central claim is stated as Theorem 4.2 and Corollary 4.1. For a continuous f mapping R^d to R_+ with a unique global minimizer, and for a smoothing function satisfying Definition 2.2, the SCBO dynamics produce a consensus state x∞ such that, whenever Assumptions 1-3 and inequality (45) hold, the essential infimum of f at x∞ is at most fmin + E(β), where E(β) tends to zero as β tends to infinity. Corollary 4.1 strengthens this to a guarantee: for any prescribed tolerance δ there exist parameters β, µ0, λ, and σ satisfying (45) with E(β) ≤ δ. The proof proceeds through an explicit log-difference formula that gives almost-sure and L² consensus, a martingale argument that yields the common limit x∞, and Ito's formula applied to the smoothed exponential weights, which produces a differential inequality that the Laplace principle converts into the final objective-error estimate.","pith_inferences":["The proof's equation (39) assumes one fixed bounded set contains all particle paths and x̄*(t) almost surely for all time. Since the paper proves only almost-sure convergence to a random limit and does not assume bounded initial data, establishing a pathwise or high-probability bound on sup_t ‖x_i(t)‖ under moment conditions would make the expectation estimates unconditional.","Because the algorithm is gradient-free and the conditions are dimension-independent, the SCBO template is an attractive candidate for high-dimensional nonsmooth machine-learning objectives; the numerical experiments support good success rates, but the theory does not yet provide a complexity rate in dimension.","The same smoothing-by-convolution device could extend other stochastic swarm or metaheuristic algorithms that rely on Ito calculus to nonsmooth objectives, provided the smoothing parameter decays slowly enough relative to the consensus rate."],"forward_implications":["For any initial data, the SCBO particles reach a common consensus state almost surely whenever 2λ is greater than σ², with an explicit L² decay rate for pairwise disagreement.","For any target accuracy δ greater than zero, Corollary 4.1 guarantees parameters under which the objective value at the consensus state is within δ of the global minimum.","All sufficient conditions are dimension-independent, so the theoretical guarantees do not degrade as the search dimension grows.","An exponentially decaying smoothing schedule µ_t = µ_0 e^{-αt} with (q+1)α less than 2λ - σ² satisfies the key integrability assumption, giving an explicit tuning rule."],"supporting_citations":[{"why":"supplies the original consensus-based optimization dynamics and the Laplace principle, stated as Proposition 2.5, that converts exponential estimates into the objective-error bound.","marker":"[31]"},{"why":"provides the component-wise diffusion CBO model in (4) that the SCBO dynamics modifies by replacing f with its smoothing function.","marker":"[7]"},{"why":"gives the finite-particle CBO convergence proof with shared environmental noise that SCBO extends to nonsmooth objectives.","marker":"[21]"},{"why":"supplies the smoothing-method theory for nonsmooth nonconvex minimization that motivates the construction of f̃.","marker":"[9]"},{"why":"provides the convolution construction used to define smoothing functions satisfying Definition 2.2.","marker":"[24]"},{"why":"supplies the variational analysis theorem used for the continuity of the convolution smoothing functions.","marker":"[32]"},{"why":"gives the large-deviations background behind Proposition 2.5, used in the limit as β tends to infinity.","marker":"[14]"},{"why":"supplies the supermartingale convergence theorem used to prove existence of the almost-sure consensus state.","marker":"[19]"},{"why":"supplies Ito's formula, the starting point for both the consensus and error estimates.","marker":"[1]"},{"why":"supplies the law of the iterated logarithm used to show that the drift term dominates Brownian noise, giving almost-sure consensus.","marker":"[28]"}],"fun_headline_variants":["Smoothing unlocks swarm consensus for nonsmooth objectives","Consensus-based optimizer tackles nonsmooth nonconvex problems","Swarm algorithm bypasses mean-field for nonsmooth global search","Smoothed particle swarm hits global min without smoothness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The error bound requires a single fixed bounded region that contains every particle path and the consensus target for all time, giving uniform smoothing constants; the paper proves only almost-sure convergence to a random limit, so the existence of that fixed region is not established.","fun_headline_variants_meta":{"raw":{"variants":["Smoothing unlocks swarm consensus for nonsmooth objectives","Consensus-based optimizer tackles nonsmooth nonconvex problems","Swarm algorithm bypasses mean-field for nonsmooth global search","Smoothed particle swarm hits global min without smoothness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1291,"prompt_tokens":912,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":528,"tokens_out":379,"duration_ms":100783,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:52:23.322113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a continuous, non-Lipschitz objective such as f(x) = |x|^p on R with p in (0,1), choose initial data and parameters satisfying Assumptions 1-3 and inequality (45), and simulate the SCBO scheme; if the empirical essential infimum of f at the consensus state exceeds fmin + E(β), or if a positive fraction of paths leaves every bounded set, the central guarantee fails. A direct check of whether condition (39) holds for this example would settle whether the proof's uniformity assumption is valid.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the original consensus-based optimization dynamics and the Laplace principle, stated as Proposition 2.5, that converts exponential estimates into the objective-error bound."},{"cited_title":"ESAIM Cont rol Optim","cited_arxiv_id":null,"evidence_quote":"provides the component-wise diffusion CBO model in (4) that the SCBO dynamics modifies by replacing f with its smoothing function."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the finite-particle CBO convergence proof with shared environmental noise that SCBO extends to nonsmooth objectives."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the smoothing-method theory for nonsmooth nonconvex minimization that motivates the construction of f̃."},{"cited_title":"Springer-Verlag, Berlin (1993)","cited_arxiv_id":null,"evidence_quote":"provides the convolution construction used to define smoothing functions satisfying Definition 2.2."},{"cited_title":"Spr inger, New York (1998)","cited_arxiv_id":null,"evidence_quote":"supplies the variational analysis theorem used for the continuity of the convolution smoothing functions."},{"cited_title":"Springer, Berlin, Heidelberg (1998)","cited_arxiv_id":null,"evidence_quote":"gives the large-deviations background behind Proposition 2.5, used in the limit as β tends to infinity."},{"cited_title":"Springer, Berlin (2013)","cited_arxiv_id":null,"evidence_quote":"supplies the supermartingale convergence theorem used to prove existence of the almost-sure consensus state."},{"cited_title":"Springer, Berlin, Heridelberg (1985)","cited_arxiv_id":null,"evidence_quote":"supplies Ito's formula, the starting point for both the consensus and error estimates."},{"cited_title":"Springer, London (2008 )","cited_arxiv_id":null,"evidence_quote":"supplies the law of the iterated logarithm used to show that the drift term dominates Brownian noise, giving almost-sure consensus."}],"review_version":1}