{"id":"d32ea69f-5c46-4d71-85e2-823934a02b96","arxiv_id":"2506.15632","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The authors prove exponential convergence for a Hessian-damped heavy-ball ODE and linear convergence for its Heavy Ball and Nesterov-type discretizations on strongly quasiconvex functions.","lead":"This paper shows that adding a curvature-dependent damping term to a momentum optimization method gives provably fast convergence for a broad class of nonconvex functions with a single minimum. The result matters because it extends accelerated methods beyond convex optimization while reducing the oscillation problems of classical momentum schemes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Numerical validation in Examples 20–21 lies outside the hypotheses of Theorems 13 and 17, so the experiments do not support the claimed linear-convergence results; the theorems' proven regime is also much narrower than the experiments suggest.","rationale":"The paper's internal proofs for Theorems 5, 13, and 17 appear algebraically coherent; I found no arithmetic error in the Lyapunov derivations under the stated hypotheses. The central weakness is the gap between those hypotheses and the paper's own validation. The reader's conditional verdict is justified: the numerical section is the only place where the methods are shown to do the advertised oscillation suppression, and it deliberately uses parameters that the theorems exclude. If the theorems are the central claim, the violation is not a refutation, but it prevents the experiments from serving as support and suggests that the practically relevant momentum regime is not covered. The continuous-time assumption (23) is also not implied by strong quasiconvexity, but the paper acknowledges this with Example 9 and supplies sufficient conditions, so I do not treat that as the primary blocker. A rerun inside the legal parameter regime would settle whether the empirical phenomenon is real; in any event the manuscript needs either to restrict its numerical claims or to widen the parameter ranges so that the experiments are covered.","tokens_in":19995,"tokens_out":29828,"duration_ms":279241,"concrete_test":"Substitute the Example 20 parameters into (40) and (59) with L=6, γ=1/2: verify that α=0.8 violates α<√2/2, that θ+β=0.0917 cannot satisfy θ+β≤(1−2α²)/L because the upper bound is negative, and that β=1/24>γ/(ηL²) for every η>1. Then rerun Figure 1 with parameters inside the theorems' regimes, e.g. for (39) take α=0.2, θ=0.01, β=0.05, and for (55) take β=0.005, η=2 and choose the largest α, θ satisfying (59). Compare the Hessian-corrected and vanilla variants. If the advantage disappears, the experimental support claim fails; if it persists, the paper must still state explicitly that the displayed behavior is outside the proven parameter range.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Example 20 (h(x)=x²+2sin²x, L=6, γ=1/2) uses α=0.8, θ=0.05, β=1/(4L)=1/24 for method (39), but Theorem 13 condition (40) requires α<√2/2≈0.707 and θ+β≤(1−2α²)/L. With α=0.8, (1−2α²)/L=(1−1.28)/6<0, so no positive θ+β can satisfy the condition. For method (55), Theorem 17 requires β<γ/(ηL²) for some η>1, i.e. β<0.5/(36η)≤1/72≈0.0139, while β=1/24≈0.0417. Example 21 similarly selects α=0.8 and 0.9 with β=0.0025 and tuned θ without checking (40) or (59). Thus the oscillation-reduction and linear-rate behavior shown in Figures 1 and 3 is not an instantiation of the theorems; the claim that the numerics support the obtained results is unsupported. The restriction is not a minor technicality: for (55), combining β<γ/(ηL²) with the small right-hand side of (59) forces α+θL to be at most O(γ²/L²) in the relevant regime, so the proven Nesterov-type update is close to gradient descent, far from the α=0.6–0.9 experiments. The theorems may be correct as conditional statements, but the paper's main advertised demonstrations lie outside their scope.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies unconstrained minimization of strongly quasiconvex functions through a second-order ODE with Hessian-driven damping, and derives two discrete-time algorithms from it: a Heavy Ball method with Hessian correction (39) and a Nesterov-type method with adaptive momentum (55). For the ODE, Theorem 5 proves exponential convergence of function values and iterates under an additional quasar-convexity-type assumption (23). For the discrete methods, Theorem 13 proves linear convergence of the Heavy Ball iteration under parameter restriction (40), and Theorem 17 proves linear convergence of the Nesterov-type iteration under restriction (59). Section 3.2 analyzes sufficient conditions for assumption (23), including Propositions 10 and 11 and counterexamples. Section 5 gives numerical experiments on two nonconvex strongly quasiconvex functions. The paper concludes that Hessian-driven damping accelerates convergence and reduces oscillations relative to classical momentum methods.","tokens_in":20418,"tokens_out":7089,"duration_ms":71254,"significance":"If the results hold in the stated regimes, the paper would extend the known exponential/linear convergence guarantees for strongly convex functions and for the heavy ball ODE in [29] to a class of nonconvex strongly quasiconvex functions with Hessian damping, and it would provide two explicit discretizations with linear convergence. The Lyapunov arguments in Theorems 5, 13, and 17 are carefully executed, the algebra appears coherent under the stated parameter conditions, and the limiting cases θ=0 and β=0 connect cleanly to existing results in [29]. The counterexamples in Section 3.2 are valuable and show the authors are aware that assumption (23) is not automatic. However, the numerical experiments do not satisfy the hypotheses of the theorems, and the parameter regime allowed by Theorem 17 is narrow enough to undermine the practical claim of acceleration. The paper is a meaningful contribution to the theory of Hessian-damped momentum methods for generalized convex functions, but its advertised experimental support and its abstract-level claims need to be reconciled with the proven conditional statements.","major_comments":[{"comment":"The numerical experiments do not fall within the hypotheses of the theorems they are claimed to illustrate. In Example 20, the Heavy Ball run uses α=0.8, but Theorem 13 condition (40) requires α<√2/2≈0.707; moreover, with α=0.8 the quantity (1−2α²)/L is negative, so no positive θ+β can satisfy θ+β≤(1−2α²)/L. For the Nesterov-type run, Theorem 17 requires β<γ/(ηL²) for some η>1, which with γ=1/2 and L=6 gives β<1/72, while the experiment uses β=1/24≈0.0417. Example 21 similarly chooses α=0.8 and 0.9 with β=0.0025 and tuned θ without checking (40) or (59). Therefore the oscillation-reduction and linear-rate behavior shown in Figures 1 and 3 is not an instantiation of the proven theorems, and the abstract's statement that numerical experiments support the obtained results is not justified. The experiments should be rerun inside the proven parameter regimes or explicitly labeled as heuristic extensions outside the scope of the theorems.","section":"Section 5, Examples 20–21"},{"comment":"The proven parameter regime for the Nesterov-type method is so restrictive that the method is effectively gradient descent rather than an accelerated method. For the allowed range β<γ/(ηL²), the quantity µ1 is close to 1 and µ2 is at most 1−1/η, so condition (59) forces α+θL to be very small. For example, with γ=1/2, L=6, η=2 and β=1/150, the admissible ǫ is at most about 1.3×10⁻⁴ and (59) forces α+θL ≲ 7×10⁻⁵. In this regime the momentum coefficient α is negligible and the method is essentially gradient descent. The authors should state this limitation explicitly or find a less restrictive condition if the advertised acceleration is to be meaningful.","section":"Section 4.2, Theorem 17 and condition (59)"},{"comment":"Theorem 5 is stated for strongly quasiconvex functions but its proof relies crucially on assumption (23), which is not implied by strong quasiconvexity. Example 9 (and also Example 12) explicitly constructs a strongly quasiconvex function that is not quasar-convex and hence need not satisfy (23). Since the theorem is conditional on (23), the abstract and introduction should state more carefully that the exponential-convergence result applies to strongly quasiconvex functions satisfying assumption (23), with Section 3.2 providing sufficient conditions. As written, the claim that the paper studies the dynamical system 'tailored for a class of nonconvex functions called strongly quasiconvex' overstates the domain of validity.","section":"Section 3, Theorem 5 and Examples 9, 12"},{"comment":"The proof divides by α²−3θ²L² when deriving inequality (50). This quantity is positive under condition (40) because θ<α/(L√3), but the division is done without explicitly noting the positivity. In the boundary case α=0 the condition θ∈[0, α/(L√3)) is empty, so the theorem statement should restrict α>0 or treat α=0 separately. This is a local clarity issue, but it should be fixed because an unnamed division by a potentially zero quantity is a correctness hazard in an otherwise sound proof.","section":"Section 4.1, proof of Theorem 13 around Eq. (50)"}],"minor_comments":[{"comment":"The text defines h(x)=x²+2sin²x, but the caption of Figure 1 states h(x)=x²+3sin²(x); these should be made consistent.","section":"Section 5, Example 20 and Figure 1"},{"comment":"The sentence 'it can be verified that h it is strongly quasiconvex with modulus' is grammatically incomplete and omits the value of the modulus; 'parameters turning' should read 'parameter tuning'.","section":"Section 5, Example 21"},{"comment":"The display in the proof of Corollary 7 is malformed: the fragment '˜C κ ∈ (2, +∞)' and the duplicated exponential-integral inequalities make the statement very hard to parse. The corollary should be restated with explicit constants or with a cleaner piecewise formulation.","section":"Corollary 7"},{"comment":"The concluding paragraph says 'future research lines are discussed in Section 5', but the conclusions and future directions are in Section 6.","section":"Section 6"},{"comment":"The remark says Theorem 5 allows 'a free damping term', but Theorem 5 restricts α to the interval (0, √(γ(κ+2)²/(8κ))], so the damping parameter is not literally free; the wording should be softened.","section":"Remark 8"}],"recommendation":"major_revision","confidential_remarks":"The central proofs are largely sound as conditional statements, but the paper's own experiments lie outside the proven regimes and Theorem 17's condition appears to force the method close to gradient descent. I would ask the authors to either prove convergence for parameter choices covering their experiments or to clearly relabel the experiments as heuristic demonstrations outside the theorem scope, and to adjust the abstract accordingly. The paper could become acceptable after a major revision that addresses the scope of the numerical claims and the restrictiveness of the parameter conditions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper proves exponential convergence for a Hessian-damped inertial ODE and linear convergence for two discretizations on strongly quasiconvex functions. That is a genuine extension of the authors' earlier heavy-ball result [29] and of the convex Hessian-damping work [3,4]. The Lyapunov arguments are coherent within the stated parameter regimes, and the sufficient conditions in Section 3.2 — especially the counterexample showing strong quasiconvexity does not imply quasar-convexity — are worth reading.\n\nThe soft spots are real, and the stress-test note is correct. The numerical experiments in Examples 20 and 21 use parameters outside the hypotheses of Theorems 13 and 17. Example 20 takes α=0.8 for the heavy-ball method, while Theorem 13 requires α<√2/2 and θ+β≤(1−2α²)/L, which is negative at α=0.8; no positive θ+β can satisfy it. For the Nesterov method, β=1/24 with γ=1/2, L=6 violates β<γ/(ηL²) for any η>1. So Figure 1 is not an instantiation of the theorems. Example 21 similarly tunes parameters without checking the conditions. The authors could fix this by running within the proven regimes or explicitly labeling the experiments as heuristic, but as written the numerical section overclaims.\n\nA more structural concern is that Theorem 17's condition (59) is so restrictive that the allowable α+θL is tiny — for small ε it is on the order of µ1µ2/(µ1/ε), which is very small unless ε is near its upper bound. In practice, the proven Nesterov update is close to gradient descent. The experiments, with α=0.6–0.9, are far outside that regime. So the paper's advertised acceleration is not actually covered by the proof.\n\nTwo minor technical issues: Theorem 13's proof divides by α²−3θ²L², which is zero at α=θ=0, a point included in the theorem's range; treating that case separately fixes it. The ODE well-posedness for (22) is never stated. There is also a small inconsistency in Example 20 (x²+2sin²x in text, x²+3sin²x in the figure caption).\n\nOverall, the conditional theoretical results are legitimate and will interest people working on generalized convexity and inertial dynamics. The paper should go to peer review, but it needs major revision: either move the experiments inside the proven regimes or label them as heuristic, and address the narrowness of the Nesterov regime. The core theorems are not fatally flawed, but the presentation overstates what is proven.","headline":"Mostly sound conditional theory for Hessian-damped momentum on strongly quasiconvex functions, but the numerical validation lies outside the proven parameter regimes and the proven Nesterov regime is too narrow to support the acceleration claims.","tokens_in":20884,"tokens_out":12595,"would_cite":false,"duration_ms":109481,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","65K10","34A34"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves exponential convergence for a Hessian-damped ODE and linear convergence for its two discretized momentum algorithms on strongly quasiconvex nonconvex objectives.","keywords":["strongly quasiconvex","Hessian-driven damping","heavy ball method","Nesterov acceleration","linear convergence","exponential convergence","nonconvex optimization","second-order dynamical system"],"falsifier":"Simulate the damped ODE (22) on the Example 9 function $h(x)=h_1(x)+x^2$, which is strongly quasiconvex but not quasar-convex: if the trajectory still decays at the claimed rate with the claimed constants, Assumption (23) is not necessary; if the decay is slower or fails, the theorem's scope is exactly as stated. For the discrete side, run the Heavy Ball method (39) with the Experiment 20 parameters $\\alpha=0.8$, $\\theta=0.05$, and $\\beta=1/24$, and check whether the proof's contraction factor $1-\\rho/\\sigma$ is actually below one, since those parameters lie outside the box (40).","tokens_in":19820,"feed_emoji":"📉","tokens_out":8987,"duration_ms":87677,"temperature":0.7,"pith_summary":"The paper studies minimization of differentiable strongly quasiconvex functions, a family that includes strongly convex functions along with genuinely nonconvex examples such as sub-unit norms and certain fractional ratios. It proposes a second-order dynamical system with both constant viscous damping and Hessian-driven damping, and proves that under a kappa-quasar-convexity inequality every trajectory converges exponentially to the unique minimizer. Discretizing that system produces a Heavy Ball method and a Nesterov-type method, each carrying a finite-difference Hessian-correction term; the paper proves linear convergence of both iterates and function values for each method. The point of the package is to bring accelerated momentum guarantees to a nonconvex class while using curvature information to suppress the oscillations that plain momentum produces.","feed_headline":"Hessian damping yields exponential decay for nonconvex momentum methods","feed_subtitle":"A damped dynamical system and two momentum algorithms converge linearly on strongly quasiconvex objectives.","key_machinery":"The load-bearing object is a Lyapunov energy function. For the continuous system the energy is $E(t)=h(x(t))-h_*+\\frac{1}{2}\\|\\lambda(x(t)-\\bar{x})+\\dot{x}(t)+\\beta\\nabla h(x(t))\\|^2$ with $\\lambda=2\\alpha/(\\kappa+2)$, and its derivative is shown to satisfy $\\dot{E}(t)+\\frac{\\lambda\\kappa}{2}E(t)\\le 0$. For the discrete methods the energy is $E_k=h(x_k)-h_*+\\frac{\\alpha^2}{\\theta+\\beta}\\|x_k-x_{k-1}\\|^2+\\frac{\\theta^2}{\\theta+\\beta}\\|\\nabla h(x_{k-1})\\|^2$ for the Heavy Ball variant and a position-momentum energy for the Nesterov variant. Each proof combines the descent lemma for $L$-smooth functions with the differential characterization of strong quasiconvexity, gradient dominance, and Assumption (23). The Hessian-correction term enters only through gradient differences, so neither algorithm requires an actual Hessian evaluation.","core_discovery":"The central discovery is that Hessian-driven damping, already known to improve heavy-ball dynamics for convex problems, supplies the same Lyapunov contraction for strongly quasiconvex objectives. Theorem 5 states that if $h$ is twice differentiable, strongly quasiconvex with modulus $\\gamma>0$, and satisfies Assumption (23), then for $\\alpha\\in(0,\\sqrt{\\gamma(\\kappa+2)^2/(8\\kappa)}]$ and $\\beta\\in(0,(\\kappa+2)/(\\kappa\\alpha))$ the trajectory of (22) satisfies $h(x(t))-h_*\\le C e^{-\\alpha\\kappa t/(\\kappa+2)}$. The discrete counterpart replaces the Hessian term $\\nabla^2 h(x)\\dot{x}$ by the gradient difference $\\theta(\\nabla h(x_k)-\\nabla h(x_{k-1}))$; Theorem 13 gives linear convergence for the Heavy Ball form (39) under parameter restrictions (40), and Theorem 17 gives linear convergence for the Nesterov-type form (55) under condition (59). In both discrete theorems a Lyapunov energy contracts by a factor in $(0,1)$ at every step, so iterates and function values converge linearly to the unique minimizer.","pith_inferences":["If the linear rates survive outside the parameter box (40), the same gradient-difference correction could be added to stochastic or proximal heavy-ball variants, giving curvature awareness without Hessian computations.","The sufficient conditions (35)-(36) turn strong quasiconvexity into quasar-convexity under quadratic-growth control; a natural next step is to relax Assumption (23) to a local or tail inequality and ask how much of the exponential rate is lost.","The continuous-time analysis suggests that Hessian-driven damping should also suppress oscillations for other nonconvex classes satisfying gradient dominance, such as Polyak-Łojasiewicz-type functions, a transfer the paper does not make."],"forward_implications":["For every twice differentiable strongly quasiconvex function satisfying Assumption (23), the damped trajectory (22) reaches its unique minimizer with function values obeying $h(x(t))-h_*\\le C e^{-\\alpha\\kappa t/(\\kappa+2)}$ from any initial condition.","The Heavy Ball method with Hessian correction (39) converges linearly for $L$-smooth strongly quasiconvex functions whenever the parameters satisfy condition (40).","The Nesterov-type method with adaptive momentum (55) converges linearly under condition (59), with contraction factor $\\mu_1(1+\\varepsilon)$.","Both discrete methods inherit oscillation suppression from the continuous Hessian damping because the term $\\theta(\\nabla h(x_k)-\\nabla h(x_{k-1}))$ acts as a finite-difference surrogate for $\\nabla^2 h(x_k)(x_k-x_{k-1})$.","Exponential and linear convergence are established without convexity; the classical strongly convex behavior is included as a special case of the strongly quasiconvex class."],"supporting_citations":[{"why":"Supplies the Hessian-driven damping system whose behavior the paper extends to strongly quasiconvex functions.","marker":"[3]"},{"why":"Supplies inertial systems with Hessian-driven damping whose discretization and rate analysis the paper adapts.","marker":"[4]"},{"why":"Supplies the definition of quasar-convexity that underlies Assumption (23).","marker":"[22]"},{"why":"Provides the strongly quasiconvex heavy-ball and gradient-method baselines to which the new rates are compared.","marker":"[29]"},{"why":"Introduces strongly quasiconvex functions, the class studied throughout the paper.","marker":"[41]"},{"why":"Provides the differential characterization of strongly quasiconvex functions used in the Lyapunov proofs.","marker":"[44]"},{"why":"Supplies existence of a unique minimizer and the quadratic growth bound used for the trajectory error.","marker":"[28]"},{"why":"Analyzes Nesterov acceleration for strongly quasar-convex functions, the setting the Nesterov part extends.","marker":"[21]"}],"fun_headline_variants":["Hessian-damped momentum converges linearly on quasiconvex objectives","Curvature-aware momentum achieves linear convergence for quasiconvex problems","Damped momentum methods converge linearly without convexity","Nonconvex momentum gets exponential decay from Hessian damping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The continuous-time exponential rate rests on Assumption (23), a $\\kappa$-quasar-convexity inequality that strong quasiconvexity alone does not imply; the paper itself gives an unbounded strongly quasiconvex function that violates it.","fun_headline_variants_meta":{"raw":{"variants":["Hessian-damped momentum converges linearly on quasiconvex objectives","Curvature-aware momentum achieves linear convergence for quasiconvex problems","Damped momentum methods converge linearly without convexity","Nonconvex momentum gets exponential decay from Hessian damping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00138,"raw_usage":{"total_tokens":5607,"prompt_tokens":983,"completion_tokens":4624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":4555}},"tokens_in":599,"tokens_out":4624,"duration_ms":36252,"temperature":1.0,"reasoning_tokens":4555,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:33:03.818386+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the damped ODE (22) on the Example 9 function $h(x)=h_1(x)+x^2$, which is strongly quasiconvex but not quasar-convex: if the trajectory still decays at the claimed rate with the claimed constants, Assumption (23) is not necessary; if the decay is slower or fails, the theorem's scope is exactly as stated. For the discrete side, run the Heavy Ball method (39) with the Experiment 20 parameters $\\alpha=0.8$, $\\theta=0.05$, and $\\beta=1/24$, and check whether the proof's contraction factor $1-\\rho/\\sigma$ is actually below one, since those parameters lie outside the box (40).","supporting_citations":[{"cited_title":"Alvarez, H","cited_arxiv_id":null,"evidence_quote":"Supplies the Hessian-driven damping system whose behavior the paper extends to strongly quasiconvex functions."},{"cited_title":"Attouch, Z","cited_arxiv_id":null,"evidence_quote":"Supplies inertial systems with Hessian-driven damping whose discretization and rate analysis the paper adapts."},{"cited_title":"Hinder, A","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of quasar-convexity that underlies Assumption (23)."},{"cited_title":"Vladimirov, Ju.E","cited_arxiv_id":null,"evidence_quote":"Provides the differential characterization of strongly quasiconvex functions used in the Lyapunov proofs."},{"cited_title":"Hermant, J.-F","cited_arxiv_id":null,"evidence_quote":"Analyzes Nesterov acceleration for strongly quasar-convex functions, the setting the Nesterov part extends."}],"review_version":2}