{"id":"b3f6fa4d-b28c-400f-8f0d-7f7e474eb18b","arxiv_id":"2412.06691","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Speed restart of the Hessian-damped inertial system (WIN) yields linear decay of objective values under a Polyak-Lojasiewicz condition, with no strong convexity required.","lead":"This paper proves a linear convergence guarantee for a speed restart procedure applied to a damped inertial optimization dynamics with Hessian-driven damping. The result gives theoretical support to a restart heuristic used to accelerate inertial optimization algorithms on convex problems that need not be strongly convex.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algebraic inversion in Proposition 3.8: solving the displayed inequality yields T(z) ≤ τ + α/(2µγ)/[(1−e^{−ατ})^2 Ψ(τ)], not the stated expression with the factor in the numerator; the stated bound collapses as τ→0 and breaks Theorem 2.2.","rationale":"The reader identified the Polyak-Łojasiewicz inequality as the weakest assumption, but a direct check of the proof reveals a concrete algebraic error in Proposition 3.8 that is more immediately load-bearing. The error is internal to the manuscript, not a matter of external consensus: the displayed inequality in the proof, when solved correctly, produces exactly the reciprocal of the factor that appears in the proposition's statement. As written, the claimed upper bound on the restart time tends to τ as τ→0, which cannot be true because Corollary 3.7 gives a positive lower bound τ3 on T(z). This also makes the definition of τ* (used in Theorem 2.2) ill-posed: the minimization over (0,τ2) has infimum 0, not a positive attained value. The proof of Theorem 2.2 depends on a positive uniform upper bound on restart times, so this error invalidates the central claim as presented. The error appears correctable by inverting the fraction, in which case the proof strategy may be salvageable with a redefinition of τ*; therefore a conditional acceptance is appropriate rather than outright rejection. The numerical check described is straightforward: solving the scalar linear ODE yields a T(z) that contradicts the paper's bound for τ small, confirming the algebraic inversion. This does not impugn the authors' honesty; it is a technical error in the argument that must be fixed before the theorem can be accepted.","tokens_in":14049,"tokens_out":13537,"duration_ms":124379,"concrete_test":"Re-derive the bound in Proposition 3.8 by solving R(T(z)−τ) ≤ 1/(2µ)‖∇f(z)‖² for T(z)−τ, where R = γ/α (1−e^{−ατ})^2 Ψ(τ)‖∇f(z)‖². The correct upper bound has the reciprocal factor: T(z) ≤ τ + α/(2µγ)/[(1−e^{−ατ})^2 Ψ(τ)]. Then test numerically on f(x)=x²/2 with α=γ=1, β=0, z=1: integrate x''+x'+x=0 and compute T(z) as the first positive time with d/dt‖ẋ‖²≤0. Use τ=0.1. The paper's bound gives ≈0.105, which is violated by the numerically observed T(z)≈1.8; the corrected bound gives ≈52.7 and holds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Proposition 3.8 claims T(z) ≤ τ + α/(2µγ)(1−e^{−ατ})^2 Ψ(τ). But the proof's own displayed inequality is R(T(z)−τ) ≤ f(x(τ))−f(x(T(z))) ≤ f(z)−f* ≤ (1/(2µ))‖∇f(z)‖², with R = γ/α (1−e^{−ατ})^2 Ψ(τ)‖∇f(z)‖². Isolating T(z)−τ gives T(z)−τ ≤ α/(2µγ) / [(1−e^{−ατ})^2 Ψ(τ)], i.e., the squared term and Ψ(τ) belong in the denominator, not the numerator. The paper's version sends the upper bound to τ as τ→0, contradicting Corollary 3.7 (τ3 ≤ T(z) with τ3 > 0) and making the defined τ* = argmin{τ + ...} over (0,τ2) have infimum 0 without a positive minimizer. Since Theorem 2.2's proof uses the existence of a positive uniform restart-time upper bound τ*, the presented argument is invalid. The subsequent rate proof cannot proceed as written; the central claim is not established.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the weighted inertial Newton-like system (2.2) with constant coefficients and proposes a speed-restart scheme in which the dynamics are re-initialized whenever the speed stops increasing. The main claim, Theorem 2.2, is a linear convergence rate for the objective values along the restarted trajectory under convexity, C^2 regularity, L-smoothness, and the Polyak-Łojasiewicz inequality (2.1), with constants independent of the starting point. The proof is organized through estimates on the restart time and on the per-restart function decrease (Lemmas 3.1–3.5, Propositions 3.6–3.9), followed by numerical illustrations.","tokens_in":14333,"tokens_out":45524,"duration_ms":427026,"significance":"If the main theorem is correct, the paper extends speed-restart linear-convergence results, previously known for strongly convex objectives in the setting of Su-Boyd-Candès and for Hessian-driven damping variants, to the wider Polyak-Łojasiewicz class. The paper is self-contained, gives explicit constants, and covers both β=0 and β>0. The numerical experiments support the qualitative benefit of restarting, although they are not the main contribution. The main issues are in the technical proof of the restart-time lower bound and in the uniform definition of the restart-time upper bound.","major_comments":[{"comment":"The lower bound for ⟨x',x''⟩ is not valid for all t∈(0,τ1). In the expansion of ⟨x',A(t)⟩, the cross term is γ(2−e^{αt})⟨I_z(t)+J_z(t),∇f(z)⟩/e^{2αt}. When e^{αt}<2 the coefficient 2−e^{αt} is positive, so the correct lower bound is −γ(2−e^{αt})‖I_z(t)+J_z(t)‖‖∇f(z)‖/e^{2αt}, which is negative, whereas the manuscript replaces it by the positive quantity −γ²(e^{αt}−2)(e^{αt}−1)(1−H)/H/(αe^{2αt})‖∇f(z)‖². A direct check with f(x)=x²/2, α=3, β=0, γ=1, z=1, and t=0.1 gives ⟨x'(t),x''(t)⟩≈0.06353, while the right-hand side of Lemma 3.5 evaluates to approximately 0.06366, so the lemma is false as stated. Since Corollary 3.7 and the lower restart-time bound in (2.4) depend on Lemma 3.5, the proof of Theorem 2.2 is incomplete as written.","section":"Lemma 3.5 (Section 3.2)"},{"comment":"The quantity used as an upper restart-time bound in the proof of Theorem 2.2 is not defined independently of the starting point. After Proposition 3.8 the text defines τ* as an argmin over (0,τ2)∩(0,T(z)], a set that depends on z, while the proof of Theorem 2.2 counts restarts using a single τ* valid for all z. The authors should define τ* = argmin over (0,τ2) of φ(τ)=τ+α/(2µγ(1−e^{−ατ})²Ψ(τ)) and then show uniform T(z)≤τ* by the two-case argument: if τ*≤T(z) use Proposition 3.8, and if τ*>T(z) the inequality is trivial. The same issue affects the use of Proposition 3.9: the per-restart factor Q must be chosen uniformly and shown to lie in (0,1), for instance by taking a fixed sufficiently small τ in (0,min(τ2,τ3)).","section":"Section 2 and Proposition 3.8"}],"minor_comments":[{"comment":"The displayed bound T(z) ≤ τ + α/(2µγ(1−e^{−ατ})²Ψ(τ)) is ambiguous in the supplied text; the proof shows that (1−e^{−ατ})²Ψ(τ) should be in the denominator. Please ensure the typeset formula is unambiguous.","section":"Proposition 3.8"},{"comment":"The lower and upper restart-time bounds are both denoted by symbols that look identical in the plain text (τ∗ and τ∗). Please distinguish τ_* and τ^* clearly.","section":"Section 2"},{"comment":"Table II has no caption, and one entry reads “3.39-e2”, which appears to be a typo for 3.39e−2.","section":"Table II"},{"comment":"In Table III the regression is presented as f(x_k)∼Ae^{−Bt}, but the data are discrete in k; the notation should be f(x_k)∼Ae^{−Bk}.","section":"Section 4"},{"comment":"In the speed-restart branch the pseudocode sets xk = xk−1 when the norm of the step decreases. Please clarify whether this is intentional or whether the intended reset is to the preceding velocity/iterate state.","section":"Algorithm 1"},{"comment":"There is a typo “Propositoin” in the sentence before Theorem 2.2, and the phrase “Polyak-/suppress Lojasiewicz” should read “Polyak–Łojasiewicz”.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's overall strategy is plausible and the main theorem may be correct, but the false estimate in Lemma 3.5 is load-bearing because it underlies the lower bound on restart times. This error appears localized and likely correctable, but the proof needs to be repaired before the central claim can be accepted. The τ* definition also needs to be made uniform in z."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper extends speed-restart analysis from strong convexity to PL and from two to three coefficients, and the proof skeleton is sound except for one load-bearing algebraic inversion in Proposition 3.8. That inversion is wrong, and it breaks the stated theorem as written.\n\nWhat's genuinely new: nobody had a linear convergence guarantee for speed-restarted (WIN) under PL. The authors correctly acknowledge WIN is a time reparameterization of DIN, so the novelty is modest, but the PL assumption is a real extension over the strong-convexity results in [23,36]. The decomposition into restart-time bounds (Prop 3.8) and per-cycle function decrease (Prop 3.9) is the right structure, and the lemmas building Mz(t) and H(t) are careful and check out.\n\nThe problem: Proposition 3.8 claims T(z) ≤ τ + (α/(2μγ))(1−e^{−ατ})²Ψ(τ). The proof's own inequality is R(T−τ) ≤ (1/(2μ))||∇f(z)||² with R = (γ/α)(1−e^{−ατ})²Ψ(τ)||∇f(z)||². Solving gives the reciprocal: T−τ ≤ α/(2μγ)/[(1−e^{−ατ})²Ψ(τ)]. With the printed bound, the upper bound tends to τ as τ→0, contradicting Corollary 3.7's positive lower bound τ3, and the infimum of τ + C(1−e^{−ατ})²Ψ(τ) on (0,τ2) is 0, so the claimed τ* minimizer doesn't exist. Theorem 2.2's proof needs that positive τ*, so the central claim is not established as written. The good news: the correct reciprocal has a positive minimizer, so the theorem is probably true with the same proof structure after fixing this slip.\n\nMinor: the choice of uniform τ* and Q in Theorem 2.2 is implicit rather than explicit; the numerical section is illustrative only; there are some typos. \n\nVerdict: worth a serious referee, but not in this form. If the authors fix the inversion, re-check the minimizer, and verify Q<1, I'd expect it to go through. As is, the main theorem lacks a valid proof.","headline":"The PL-extension is real and the proof structure is sound, but Proposition 3.8 contains a load-bearing algebraic inversion that breaks the stated theorem as written.","tokens_in":14886,"tokens_out":4042,"would_cite":false,"duration_ms":37520,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["37N40","90C25","65K10","34A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"A speed-restart rule makes a Hessian-damped inertial system converge linearly in objective value for any convex, smooth function satisfying the Polyak-Lojasiewicz inequality, with no strong convexity assumption.","keywords":["convex optimization","inertial methods","Hessian-driven damping","speed restart","Polyak-Lojasiewicz inequality","linear convergence","continuous-time dynamical system","heavy ball"],"falsifier":"Integrate the system for $f(x)=\\frac{1}{2}\\|x\\|^2$, $\\alpha=3$, $\\beta=0$, $\\gamma=1$, from $z=(1,0,\\ldots,0)$, and locate the first time $T(z)$ at which $\\frac{d}{dt}\\|x'(t)\\|^2$ changes sign; Theorem 2.2 requires $T(z)$ to lie between the paper's $\\tau_*$ and $\\tau^*$ and the first restart to shrink the objective gap by at most the factor $Q$ from Proposition 3.9. A violation of either check, or an analogous check on any convex $C^2$, $L$-smooth PL function, would refute the claimed uniform linear convergence.","tokens_in":13846,"feed_emoji":"⚡","tokens_out":17066,"duration_ms":155266,"temperature":0.7,"pith_summary":"This paper studies a continuous-time inertial optimization dynamics with Hessian-driven damping, $x'' + \\alpha x' + \\beta \\nabla^2 f(x) x' + \\gamma \\nabla f(x) = 0$, under a speed-restart rule: once the squared velocity stops increasing, the trajectory is cut and restarted from its current position with zero velocity. The authors' goal is to prove that this restarted dynamics has linear convergence of objective values, $f(x(t))-f^* \\leq C e^{-Kt} (f(z)-f^*)$, for every convex $C^2$ function with Lipschitz gradient that satisfies the Polyak-Lojasiewicz inequality $2\\mu(f(z)-f^*) \\leq \\|\\nabla f(z)\\|^2$. This matters because the linear rate is obtained without assuming strong convexity, which is the usual hypothesis in the prior speed-restart analyses. The proof gives uniform bounds on the time between restarts and a uniform decrease factor per restart, and the numerical experiments show the rule also improves the observed decay rates for the discrete inertial algorithm they test.","feed_headline":"Speed restarts give linear decay without strong convexity","feed_subtitle":"A speed-stall restart forces exponential convergence for Polyak-Lojasiewicz objectives.","key_machinery":"The argument is carried by three quantitative objects: the restart time $T(z)=\\inf\\{t>0: \\frac{d}{dt}\\|x'(t)\\|^2 \\le 0\\}$; an auxiliary function $H(t)$, defined in (3.5) as $H(t)=1+\\frac{2L\\gamma}{\\alpha^2}-\\frac{L\\beta}{\\alpha}-\\frac{Lt(\\frac{\\gamma}{\\alpha}e^{\\alpha t}+\\frac{\\gamma}{\\alpha}-\\beta)}{e^{\\alpha t}-1}$, which decreases from 1 to $-\\infty$ and whose zeros $\\tau_1$, $\\tau_2$ mark the time scale on which the frozen-gradient bounds are valid; and a companion function $G(t)$ whose zero $\\tau_3$ gives the uniform lower bound on $T(z)$. The proof also uses the supremum $M_z(t)=\\sup_{u\\in(0,t]} \\|x'(u)\\|/(1-e^{-\\alpha u})$ to bound the integral terms $I_z(t)$ and $J_z(t)$ defined in (3.3). Lemma 3.1 bounds those integrals by $L$ times $M_z(t)$ times explicit functions of $t$; Lemma 3.5 shows $\\langle x'(t),x''(t)\\rangle$ is at least $\\gamma^2(e^{\\alpha t}-1)\\|\\nabla f(z)\\|^2/(\\alpha e^{2\\alpha t} H(t)^2) G(t)$, so while $H,G$ stay positive the speed is still increasing and $T(z)$ has not been reached. Proposition 3.8 uses the PL inequality to bound $T(z)$ from above, and Proposition 3.9 integrates a lower bound on the speed over $[0,\\tau]$ to get the per-restart factor $Q$. In short, $H$ and $G$ control the kinematics of the restart, and the PL inequality converts those kinematics into progress on the objective.","core_discovery":"The central discovery is that the speed-restart mechanism, previously known to give linear convergence for strongly convex objectives, works under the strictly weaker Polyak-Lojasiewicz condition. For the system (2.2) with any coefficients $\\alpha>0$, $\\beta\\ge0$, $\\gamma>0$, the paper constructs explicit positive constants $\\tau_*$ and $\\tau^*$ such that every restart interval has length between them, independent of the starting point, and proves that each restart reduces the function-value gap by a uniform factor $Q\\in(0,1)$. Iterating the decrease factor and using the uniform time bounds yields the exponential estimate of Theorem 2.2. The technical route rewrites the ODE as an integral equation, controls the gradient and Hessian terms through a supremum of the normalized speed, and then converts a lower bound on $\\langle x', x''\\rangle$ into the restart-time bounds; the Polyak-Lojasiewicz inequality is what turns a lower speed bound into a quantified gap decrease. A corollary is that the same linear-rate statement covers the Heavy Ball case $\\beta=0$, so the Hessian term is not needed for the rate, only for damping oscillations.","pith_inferences":["If the PL constant $\\mu$ is unknown, the explicit upper bound on the restart time cannot be computed; an adaptive scheme that estimates $\\mu$ online while running the dynamics is a natural extension that the paper does not analyze.","The continuous-time theorem suggests that any inertial algorithm whose continuum limit is (WIN) should inherit linear convergence under PL when restarted at speed-stall events; the paper tests only one discretization, so whether other standard accelerated-gradient discretizations obey the same discrete rate remains open.","A testable prediction of the proof is that conditioning enters the rate only through $L$ and $\\mu$, not through the starting point; one could verify this by fitting $C$ and $K$ on a family of quadratic PL functions with increasing condition numbers and comparing them with the theorem's constants.","The discrete restart condition used in Algorithm 1, $\\|x_{k+1}-x_k\\| < \\|x_k-x_{k-1}\\|$, is a natural proxy for the continuous speed-stall condition; the continuous analysis provides a heuristic justification for it, not a proof for the discrete method."],"forward_implications":["For every convex $C^2$, $L$-smooth $f$ satisfying the PL inequality with some $\\mu>0$, the speed-restarted trajectory of (WIN) satisfies $f(x(t))-f^* \\le C e^{-Kt}(f(z)-f^*)$ with constants $C,K$ independent of the starting point $z$.","Because $\\beta$ is allowed to be zero, the result applies to the classical Heavy Ball dynamics, giving linear convergence under PL rather than only under strong convexity.","The uniform bounds $\\tau_* \\le T(z) \\le \\tau^*$ imply the number of restarts up to time $t$ grows linearly in $t$, so the exponential decay rate is governed by the worst-case per-restart decrease factor $Q$.","Since strong convexity implies the PL inequality, the theorem covers every strongly convex smooth objective and extends the earlier speed-restart analyses to non-strongly-convex PL objectives.","In the numerical experiments, the speed-restart rule improves the fitted exponential decay rate $B$ for both the continuous system (with and without Hessian damping) and for a discretized inertial gradient algorithm; the warm-start variant does not improve the slope."],"supporting_citations":[{"why":"supplies the speed-restart rule and the template linear-rate proof for a strongly convex objective, which this paper extends to the PL case.","marker":"[36]"},{"why":"the direct predecessor that applies speed restart to a Hessian-damped inertial system; this paper generalizes its analysis to three constant coefficients and PL objectives.","marker":"[23]"},{"why":"introduces the Hessian-driven damping term and the (DIN) system from which the model (WIN) is built.","marker":"[5]"}],"fun_headline_variants":["Speed restarts ensure linear decay without strong convexity","Linear convergence via speed restarts under P-L condition","Speed restart scheme cracks non-strongly convex linear rate","Hessian-damped system: linear rate with Polyak-Lojasiewicz","Restart trick yields exponential decay for P-L objectives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the Polyak-Lojasiewicz inequality $2\\mu(f(z)-f^*) \\le \\|\\nabla f(z)\\|^2$ holding with a known positive $\\mu$; if that inequality fails or $\\mu$ is unknown, the upper bound on the restart time and the fixed per-restart decrease both disappear, and the linear-rate proof collapses.","fun_headline_variants_meta":{"raw":{"variants":["Speed restarts ensure linear decay without strong convexity","Linear convergence via speed restarts under P-L condition","Speed restart scheme cracks non-strongly convex linear rate","Hessian-damped system: linear rate with Polyak-Lojasiewicz","Restart trick yields exponential decay for P-L objectives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1276,"prompt_tokens":829,"completion_tokens":447,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":365}},"tokens_in":445,"tokens_out":447,"duration_ms":4554,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:21:59.393474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Integrate the system for $f(x)=\\frac{1}{2}\\|x\\|^2$, $\\alpha=3$, $\\beta=0$, $\\gamma=1$, from $z=(1,0,\\ldots,0)$, and locate the first time $T(z)$ at which $\\frac{d}{dt}\\|x'(t)\\|^2$ changes sign; Theorem 2.2 requires $T(z)$ to lie between the paper's $\\tau_*$ and $\\tau^*$ and the first restart to shrink the objective gap by at most the factor $Q$ from Proposition 3.9. A violation of either check, or an analogous check on any convex $C^2$, $L$-smooth PL function, would refute the claimed uniform linear convergence.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the speed-restart rule and the template linear-rate proof for a strongly convex objective, which this paper extends to the PL case."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"the direct predecessor that applies speed restart to a Hessian-damped inertial system; this paper generalizes its analysis to three constant coefficients and PL objectives."},{"cited_title":"´Alvarez, H","cited_arxiv_id":null,"evidence_quote":"introduces the Hessian-driven damping term and the (DIN) system from which the model (WIN) is built."}],"review_version":1}