{"id":"56538327-ed8c-487b-8cde-2dfe6acb3c6d","arxiv_id":"1908.09043","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Continuous-time proximal gradient and Douglas-Rachford splitting flows are shown to be globally exponentially stable using integral quadratic constraints, with explicit rates.","lead":"This paper rewrites two optimization algorithms, proximal gradient and Douglas-Rachford splitting, as continuous-time flows and proves that their solutions converge exponentially fast to the optimum when the smooth part of the objective is strongly convex. It also proves exponential convergence of a smooth surrogate, the forward-backward envelope, under a generalized Polyak-Lojasiewicz condition.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the IQC stability proofs are sound; remaining gaps are secondary (Section 4.2 unproved, Appendix A sign error).","rationale":"The stress-test confirms the mathematical core. The reader's weakest assumption—that Lemma 1's sector bound with σ<1 is load-bearing—is actually proven, so the force of that concern does not land. However, the paper's secondary claims have gaps: Section 4.2's unproved dual extension and the sign error in Appendix A. These justify keeping a CONDITIONAL verdict (addressable before publication) but do not elevate to REJECT. The strongest_claim holds under the stated assumptions.","tokens_in":12195,"tokens_out":41511,"duration_ms":356138,"concrete_test":"Re-derive Appendix A: substitute f = F_μ − g_+ − (μ/2)||G||² + μ⟨∇f,G⟩ into (A.3); the resulting inequality is (μκ−1)/2||G||² ≥ κ(F_μ−F⋆) + (μκ−1)/μ(g−g_+) + (μκ−1)⟨∇f,G⟩. With this corrected (A.4), combine with (A.5) to verify (8). If the sign of the inner-product term is not corrected, the claimed implication fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I checked the central claim—global exponential stability of the proximal gradient flow (15) and DR splitting flow (26) under Assumption 1. Lemma 1's sector inequality (17a) follows correctly from firm nonexpansiveness of prox, Lipschitz smoothness, and Nesterov's inequality (6); the coefficient calculation yields σ²=max{(1−μL_f)²,(1−μm_f)²}, and σ<1 for μ∈(0,2/L_f). The LMI in Theorem 2 (22)-(23) is solved correctly, giving rate ρ≤1−σ. The same IQC multiplier applies to the DR nonlinearity R_g∘R_f because R_f is σ-contractive and R_g is non-expansive. Theorem 3 is a direct Lyapunov argument under Assumption 2 and is valid. The weaknesses I found are not load-bearing for these theorems: (i) Section 4.2 asserts dual DR stability without proof, (ii) Appendix A's derivation of the proximal PL lower bound contains a sign error—the term multiplying ⟨∇f,G⟩ in (A.4) should be +(μκ−1), not −(μκ−1). With that sign fixed, (A.4) and (A.5) imply (8). These are addressable issues that do not affect the central stability claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces continuous-time models for proximal gradient and Douglas-Rachford splitting applied to nonsmooth composite problems and analyzes their global stability. For problems of the form min_x f(x)+g(x) with f strongly convex and ∇f Lipschitz, Theorem 2 proves that the equilibrium of the proximal gradient flow (15) is globally exponentially stable for μ∈(0,2/L_f) with rate ρ≤1−σ, where σ=max{|1−μm_f|,|1−μL_f|}. Lemma 1 and Lemma 4 establish the underlying contraction and sector properties, and Theorem 6 carries the same rate to the DR splitting flow (26). Under a proximal Polyak-Lojasiewicz condition (Assumption 2), Theorem 3 establishes exponential decay of the forward-backward envelope. The paper also discusses extensions to dual DR dynamics and connects the proximal PL condition to a known PL formulation in Appendix A.","tokens_in":12428,"tokens_out":20875,"duration_ms":182602,"significance":"If the results hold, the paper provides a clean, control-theoretic treatment of continuous-time proximal algorithms with explicit exponential rates. The central proofs are complete: Lemma 1's sector bound (17a) follows from firm nonexpansiveness of the proximal operator, strong convexity, and Nesterov's inequality; the LMI (22)–(23) in Theorem 2 is verified with P=pI; and Theorem 3 is a standard Lyapunov argument. The paper is also honest about the limitation of the PL branch in Remark 4. The rates are explicit and the assumptions are standard, and the proofs do not rely on fitted constants. These are useful tools for the optimization-as-dynamical-systems literature and the paper is likely to be of interest to readers of Automatica.","major_comments":[],"minor_comments":[{"comment":"The global exponential stability claim for the dual DR dynamics (30) is asserted as 'readily established' without proof; because the conjugates f1 and g1 require verifying the relevant strong-convexity, smoothness, and IQC conditions, please either provide the proof or explicitly mark this as a conjecture.","section":"Section 4.2"},{"comment":"There is a sign error in the term multiplying ⟨∇f(x), G_μ(x)⟩: it should be +(μκ−1) rather than −(μκ−1). As written, the transition from (A.4) to the displayed bound with γ=2κ/|μκ−1| is not valid. Because Theorem 3 assumes (8) directly, this error does not undermine the main theorem, but the appendix should be corrected and the derivation checked.","section":"Appendix A, Eq. (A.4)"},{"comment":"In the proof of Lemma 5, the expression 'prox_{μf}' should read 'prox_{μg}'; otherwise the displayed identity concerns the wrong resolvent.","section":"Lemma 5"},{"comment":"In the proof of Theorem 6, 'systems (11)' should be 'systems (16)'; the feedback interconnection used there is (16a) with nonlinearity (27).","section":"Theorem 6, proof"},{"comment":"The sentence 'this is the best achievable convergence rate for system (15)' is a strong optimality claim with no proof or citation; please add a derivation or soften the statement.","section":"Remark 2"}],"recommendation":"minor_revision","confidential_remarks":"The central theorems are sound and the presentation is generally careful. The missing proof in Section 4.2 and the Appendix A sign error are local and do not affect the main stability claims, so I recommend minor revision. The paper fits the journal's scope well."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: the main theorems check out, and the continuous-time Douglas-Rachford splitting flow is a real new contribution. But the paper has a couple of warts—an unproved claim in Section 4.2 and a sign error in Appendix A—that need a pass before I'd sign off.\n\nWhat's actually new: the DR splitting flow (26) and its global exponential stability theorem (Theorem 6). The proof is neat: R_μg is firmly nonexpansive, R_μf is σ-contractive with the same σ as the proximal gradient nonlinearity, so the same IQC multiplier applies and the LMI (22) gives the rate. That is a clean observation. The PL-based envelope convergence theorem (Theorem 3) is also new, and it's a straightforward Lyapunov argument that works. The main proofs are complete and the rates are explicit; the optimal μ = 2/(L_f+m_f) recovers the expected 2/(κ+1) rate. No fitted parameters, no circularity.\n\nSoft spots, in proportion. The novelty claim in the introduction is a bit strong: the proximal gradient flow already appeared in the authors' own CDC 2018 paper [27], so only the DR flow is genuinely first here. Section 4.2 says stability of DR on the dual problem is 'readily established' and provides no proof; that's an assertion, not a result, and it should be either proven or trimmed. Appendix A has a sign error in (A.4): the term multiplying ⟨∇f, G⟩ should be +(μκ−1), not −(μκ−1). It's a typo—the reader and the stress-test both caught it, and with the sign fixed the derivation works—but it needs flagging. These are all minor issues.\n\nBottom line: the central stability claims hold up. The paper is honest about its limitations, including the absence of an x(t) rate under PL without strong convexity. It is written for people working on continuous-time optimization dynamics and IQC-based convergence. I'd send it to review; the correct verdict is minor revision, not accept as is.\n\nBest,","headline":"A sound IQC-based analysis with a genuinely new continuous-time DR flow result; fix the appendix sign error and Section 4.2 before publication.","tokens_in":12979,"tokens_out":8641,"would_cite":true,"duration_ms":66468,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","65K10","93D23"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves global exponential stability of the continuous-time proximal gradient and Douglas-Rachford splitting flows, with explicit rates determined by the strong-convexity and smoothness constants.","keywords":["proximal gradient flow","Douglas-Rachford splitting","global exponential stability","integral quadratic constraints","forward-backward envelope","proximal PL condition","composite optimization","nonsmooth optimization"],"falsifier":"Take the scalar problem $f(x)=\\frac{L_f}{2}x^2$ with $g=0$; the proximal gradient flow becomes the linear ODE $\\dot{x}=-\\mu L_f x$, whose exact decay exponent is $\\mu L_f$, while the theorem certifies the rate $1-\\sigma$ with $\\sigma=|1-\\mu L_f|$, so checking $\\mu$ across $(0,2/L_f)$ settles whether the claimed rate ever overstates the true contraction.","tokens_in":11976,"feed_emoji":"📉","tokens_out":13608,"duration_ms":120269,"temperature":0.7,"pith_summary":"This paper treats proximal gradient and Douglas-Rachford splitting as ordinary differential equations and asks whether their equilibria are globally exponentially stable. For composite objectives in which the smooth part is strongly convex with Lipschitz gradient and the nonsmooth part is convex, the answer is yes for every step size $\\mu \\in (0, 2/L_f)$, with an explicit rate $\\rho \\le 1 - \\max\\{|1-\\mu m_f|, |1-\\mu L_f|\\}$. The proof rewrites each algorithm as a feedback loop between a stable linear system and a nonlinear proximal step that is a strict contraction, then uses integral quadratic constraints, a control-theoretic stability certificate, to certify global exponential stability. Under a proximal gradient-dominance condition, the same ideas show the forward-backward envelope, a smooth surrogate for the objective, converges exponentially even without strong convexity, although the trajectory itself need not. The payoff is a unified explanation of convergence whose rates are tied directly to the conditioning of the problem.","feed_headline":"Proximal and Douglas-Rachford flows converge exponentially","feed_subtitle":"Control-theoretic proof yields explicit rates from the problem's strong-convexity and smoothness constants.","key_machinery":"The load-bearing object is a pointwise quadratic inequality for the nonlinear map $u(\\xi) = \\mathrm{prox}_{\\mu g}(\\xi - \\mu\\nabla f(\\xi))$: for any two points, $\\|u(\\xi)-u(\\hat{\\xi})\\|^2 \\le \\sigma^2 \\|\\xi-\\hat{\\xi}\\|^2$ with $\\sigma = \\max\\{|1-\\mu m_f|, |1-\\mu L_f|\\}$. This sector bound is what integral quadratic constraints require, and it makes $u$ a strict contraction exactly when $\\mu < 2/L_f$. It is derived by combining firm nonexpansiveness of the proximal operator with a two-sided inequality that controls the inner product of two gradient differences in terms of both the distance between points and the norm of the gradient difference. The same bound, with the reflected proximal operator $R_{\\mu f} = 2\\,\\mathrm{prox}_{\\mu f} - I$ in place of $u$, covers the Douglas-Rachford nonlinearity $R_{\\mu g}R_{\\mu f}$. Feeding this bound into the exponential-stability test turns the global stability question into the solvability of a small matrix inequality, which holds precisely when $\\rho \\le 1-\\sigma$.","core_discovery":"The central discovery is that the continuous-time proximal gradient flow, $\\dot{x} = -(x - \\mathrm{prox}_{\\mu g}(x - \\mu \\nabla f(x)))$, and the Douglas-Rachford splitting flow, $\\dot{z} = -z + R_{\\mu g}R_{\\mu f}(z)$, are globally exponentially stable when $f$ is $m_f$-strongly convex with $L_f$-Lipschitz gradient, $g$ is a convex nonsmooth function, and $\\mu \\in (0, 2/L_f)$. The exponential rate is $\\rho \\le 1 - \\sigma$ with $\\sigma = \\max\\{|1-\\mu m_f|, |1-\\mu L_f|\\}$, and the choice $\\mu = 2/(L_f+m_f)$ gives $\\rho \\le 2/(\\kappa+1)$, where $\\kappa = L_f/m_f$ is the condition number. The proof is uniform: both nonlinear maps satisfy the same pointwise quadratic inequality, so both fit the same stability test. When strong convexity is absent but the proximal gradient-dominance condition holds, the paper shows the forward-backward envelope $F_\\mu(x(t))$ decays as $F_\\mu(x(t)) - F_\\mu^\\star \\le e^{-\\gamma\\mu(1-\\mu L_f)t}(F_\\mu(x(0)) - F_\\mu^\\star)$.","pith_inferences":["The rate bound is designed for the continuous-time dynamics; the discrete algorithms' actual rates should differ from $1-\\sigma$ by an extra discretization factor, and setting up a matched stability test for the discrete iteration could quantify that gap.","The proof uses only the strong-convexity and smoothness constants of $f$, so any nonlinear map with the same contraction parameters would satisfy the same stability certificate; this suggests immediate extensions to preconditioned or inexact proximal steps.","Inside the proximal gradient-dominance branch, the rate $\\gamma\\mu(1-\\mu L_f)$ is maximized at $\\mu = 1/(2L_f)$, giving $\\gamma/(4L_f)$; the paper does not discuss this optimization, but it follows directly from the displayed bound.","Since the Douglas-Rachford flow is stable for a merely convex nonsmooth part, the continuous-time viewpoint may help analyze alternating schemes that lack an obvious Lyapunov function."],"forward_implications":["The standard proximal gradient and Douglas-Rachford iterations are explicit forward-Euler discretizations of these flows, so the continuous-time theorem gives an idealized convergence rate for the algorithms they discretize.","For a fixed condition number $\\kappa = L_f/m_f$, the best step size in this analysis is $\\mu = 2/(L_f+m_f)$, yielding the rate $2/(\\kappa+1)$, the same condition-number dependence as classical gradient descent in the smooth case.","The same IQC proof applies to the Douglas-Rachford flow formulated on the dual problem, covering the ADMM-equivalent algorithm whenever the constraint matrix has full row rank.","Without strong convexity, the proximal gradient-dominance condition still forces the forward-backward envelope to converge exponentially at rate $\\gamma\\mu(1-\\mu L_f)$, but it does not give an exponential rate for the distance to the optimizer.","Because both algorithms share one quadratic characterization, any tightening of the sector bound would automatically improve the certified rate for both flows."],"supporting_citations":[{"why":"Supplies firm nonexpansiveness of the proximal operator and the Moreau envelope identities used throughout the stability proofs.","marker":"[6]"},{"why":"Supplies the integral-quadratic-constraint framework that converts sector bounds on nonlinearities into stability certificates.","marker":"[34]"},{"why":"Provides the exponential-decay matrix inequality used to derive the rate rho <= 1 - sigma.","marker":"[42]"},{"why":"Provides the two-sided inequality combining strong convexity and Lipschitz smoothness, which produces the contraction constant.","marker":"[40]"},{"why":"Supplies the forward-backward envelope properties and its gradient formula used in the non-strongly-convex convergence proof.","marker":"[36]"},{"why":"Supplies the proximal gradient-dominance condition and its link to linear convergence, which the paper generalizes to the nonsmooth setting.","marker":"[41]"},{"why":"Supplies the proximal augmented Lagrangian and the primal-dual flow whose restriction yields the proximal gradient flow.","marker":"[11]"},{"why":"Supplies the reflected proximal operators used to write the Douglas-Rachford optimality condition and dynamics.","marker":"[44]"}],"fun_headline_variants":["Exponential stability proven for proximal and DR splitting flows","Control theory locks exponential rates for optimization flows","Strong convexity prescribes exponential settling for two flows","IQCs certify exponential stability for proximal and DR dynamics","Two flows, one IQC: exponential stability proven"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof needs the proximal step's nonlinear map to be a strict contraction, which forces the smooth part to be strongly convex with Lipschitz gradient and the step size to satisfy $\\mu < 2/L_f$; once $\\sigma = \\max\\{|1-\\mu m_f|, |1-\\mu L_f|\\}$ reaches 1, the algebra behind the stability certificate stops working.","fun_headline_variants_meta":{"raw":{"variants":["Exponential stability proven for proximal and DR splitting flows","Control theory locks exponential rates for optimization flows","Strong convexity prescribes exponential settling for two flows","IQCs certify exponential stability for proximal and DR dynamics","Two flows, one IQC: exponential stability proven"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3339,"prompt_tokens":1034,"completion_tokens":2305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":2241}},"tokens_in":650,"tokens_out":2305,"duration_ms":17543,"temperature":1.0,"reasoning_tokens":2241,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:25:22.481291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the scalar problem $f(x)=\\frac{L_f}{2}x^2$ with $g=0$; the proximal gradient flow becomes the linear ODE $\\dot{x}=-\\mu L_f x$, whose exact decay exponent is $\\mu L_f$, while the theorem certifies the rate $1-\\sigma$ with $\\sigma=|1-\\mu L_f|$, so checking $\\mu$ across $(0,2/L_f)$ settles whether the claimed rate ever overstates the true contraction.","supporting_citations":[{"cited_title":"Proximal algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies firm nonexpansiveness of the proximal operator and the Moreau envelope identities used throughout the stability proofs."},{"cited_title":"System analysis via integ ral quadratic constraints,","cited_arxiv_id":null,"evidence_quote":"Supplies the integral-quadratic-constraint framework that converts sector bounds on nonlinearities into stability certificates."},{"cited_title":"Exponential decay rate conditions for uncertain linear systems using integral quadratic constraints,","cited_arxiv_id":null,"evidence_quote":"Provides the exponential-decay matrix inequality used to derive the rate rho <= 1 - sigma."},{"cited_title":"Nesterov, Introductory lectures on convex optimization: A basic course, 2013, vol","cited_arxiv_id":null,"evidence_quote":"Provides the two-sided inequality combining strong convexity and Lipschitz smoothness, which produces the contraction constant."},{"cited_title":"Forward-backward truncated Newton methods for convex composite optimization","cited_arxiv_id":"1402.6655","evidence_quote":"Supplies the forward-backward envelope properties and its gradient formula used in the non-strongly-convex convergence proof."},{"cited_title":"Linear convergen ce of gradient and proximal-gradient methods under the Polyak - Lojasiewicz condition,","cited_arxiv_id":null,"evidence_quote":"Supplies the proximal gradient-dominance condition and its link to linear convergence, which the paper generalizes to the nonsmooth setting."},{"cited_title":"The proximal augmented Lagrangian method for nonsmooth composite optimization,","cited_arxiv_id":null,"evidence_quote":"Supplies the proximal augmented Lagrangian and the primal-dual flow whose restriction yields the proximal gradient flow."},{"cited_title":"Linear convergence and metri c selection for Douglas-Rachford splitting and ADMM,","cited_arxiv_id":null,"evidence_quote":"Supplies the reflected proximal operators used to write the Douglas-Rachford optimality condition and dynamics."}],"review_version":1}