{"id":"ee1868ca-e96a-417a-b317-cc5607b9d430","arxiv_id":"2505.10800","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A new contractive difference-of-convex algorithm solves each subproblem by Picard iteration of a contraction with an adaptive termination rule, and provably converges to critical points.","lead":"This paper rewrites the inner step of a standard difference-of-convex optimization method as a fixed point of a contraction, then solves it with simple Picard iterations and an adaptive stopping rule. The result is an algorithm, cDCA, with a proof of convergence to critical points and numerical tests suggesting it runs faster than two recent accelerated variants.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.2 needs ∇h Lipschitz on co({x_k}); the ℓ1−2 experiments use h=γ‖x‖, which is not differentiable at 0, so the whole-sequence convergence theorem does not cover the paper's main test case as stated.","rationale":"After checking Lemma 3.1, the descent estimate (17), and the boundedness/summability arguments, the core contraction-based analysis appears sound; I do not see an algebraic error in the main proof. The reader's other worries—the tolerance-based stopping in Step 2 versus Lemma 3.2 and the x_k=x_{k-1} edge case—are less central: if the algorithm does not stop, every accepted x_{k+1} differs from x_k (the stopping test would otherwise have triggered), so stepsize thresholds in (16) are positive and the inner loop is finite. The genuinely load-bearing soft spot is the condition on ∇h in Lemma 4.1/Theorem 4.2. It is not a standing assumption on h, it is not verified for the ℓ1−2 test problem, and h=γ‖x‖ is in fact nondifferentiable at 0. The paper's citation of [34] does not supply the missing lower bound. Hence the advertised whole-sequence convergence is, for a flagship example, conditional on an unproven and possibly false regularity property. The appropriate verdict remains conditional: the theory is likely repairable by stating Theorem 4.2 for smooth h (or adding a verified nondegeneracy condition), but as written the claim overreaches.","tokens_in":16598,"tokens_out":22461,"duration_ms":214655,"concrete_test":"Log the ℓ1−2 cDCA runs from Table 3 with the paper's parameters (λ=0.1Lf, δ=1.99λ/Lf, tol=1e-6) and record min_k ‖x_k‖ and the quantity sup_{u,v in co({x_i})} ‖∇h(u)-∇h(v)‖/‖u-v‖ along the trajectory. If any run has ‖x_k‖=0 or the sup-Lipschitz constant over the convex hull grows without bound as k increases, Theorem 4.2's hypothesis fails for (32). An analytic companion is to solve the LPM subproblem (10) for h=γ‖x‖ at x_k=0 with η_k=0 and f≡0, g=γ‖x‖_1: the minimizer is 0, showing the nonsmooth point is reachable; the numerical logging on the actual Table 3 instances settles whether it is reached.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing point is the extra smoothness hypothesis in Theorem 4.2/Lemma 4.1: ∇h must exist and be L_h-Lipschitz continuous on co({x_k}). The proof of Lemma 4.1 literally writes ∇h(x_{k+1}) and ∇h(x_k), so h must be differentiable at every generated point. In the ℓ1−2 experiments (Section 5, Eq. (32)), h(x)=γ‖x‖. This h is nondifferentiable at x=0, and ∇h(x)=γx/‖x‖ is not Lipschitz on any convex set that contains, or accumulates at, 0. The paper says (after Eq. (32)) that γ<1/(2‖A^T b‖_∞) makes the assumptions of Theorem 4.2 satisfied, citing [34, Subsection 5.1], but the cited passage does not establish a positive lower bound on ‖x_k‖ or differentiability of h along the cDCA sequence. Since the LPM subproblem has a strong convexity term but no term preventing a zero minimizer, the generated sequence can in principle hit 0; if it does, Lemma 4.1 and Theorem 4.2 do not apply. Thus the abstract's 'global convergence of the whole sequence' is not actually established for one of the two numerical test problems; for that example only the subsequential Theorem 4.1 is guaranteed. This is a scope/application gap rather than an internal contradiction in the proof, but it is the most load-bearing limitation of the advertised claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the difference-of-convex minimization problem min F(x)=f(x)+g(x)-h(x), where f is smooth convex with Lipschitz gradient, g is prox-friendly convex, and h is convex. Its first contribution is Lemma 3.1, which reformulates the linearized proximal method (LPM) subproblem as a fixed-point problem of a contraction, Prox_{μg}[T_k^{λ,μ}], with contraction coefficient 1-μλ when μ∈(0,2/(2λ+L_f)]. Building on this, the authors propose cDCA, which solves each LPM subproblem approximately by Picard iteration with the adaptive inner stopping rule (16), and an outer stopping rule in Step 2. The main theoretical results are Theorem 4.1, establishing subsequential convergence to a critical point and summability of squared successive displacements, and Theorem 4.2, establishing whole-sequence convergence under a Kurdyka-Łojasiewicz condition on an auxiliary function E and Lipschitz differentiability of ∇h on the convex hull of the generated sequence. Numerical experiments compare cDCA with ADCA and pDCAe on ℓ1-2 and logarithmic regularized least-squares problems.","tokens_in":16883,"tokens_out":15244,"duration_ms":138900,"significance":"The contraction reformulation of the LPM subproblem is clean, and the adaptive inner termination rule is a useful idea: if the theory holds, each outer iteration is computed by finitely many inner steps whose error is controlled by the previous outer displacement, avoiding a fixed inner tolerance schedule. The descent estimate (17) in Theorem 4.1 is carefully derived, and the subsequential convergence proof is largely sound. The main limitation is that the whole-sequence convergence theorem's smoothness assumption is not verified for the two numerical test problems, so the advertised global convergence is broader than what is actually established for those examples.","major_comments":[{"comment":"Theorem 4.2 requires ∇h to be L_h-Lipschitz continuous on co({x_k}). For the ℓ1−2 test problem (32), h(x)=γ‖x‖, which is not differentiable at 0; for the logarithmic problem (33), the chosen h(x)=∑γ[|x_i|/ε − log(|x_i|+ε)+log ε] is also nondifferentiable on coordinate hyperplanes. The sentence after (32) asserting that γ<1/(2‖A^T b‖∞) makes the assumptions of Theorem 4.2 satisfied is not justified: the cited passage in [34] concerns a different algorithm, and no argument is given that the cDCA iterates, or their convex hull, avoid the nondifferentiability set. Consequently the abstract's whole-sequence convergence claim is not established for either numerical test case; only the subsequential Theorem 4.1 is guaranteed there. Please either prove the needed avoidance/differentiability property for cDCA or explicitly state that Theorem 4.2 applies only to problems whose h is differentiable on the generated convex hull.","section":"Section 5, Eqs. (32)–(33) and Theorem 4.2"},{"comment":"In Step 2 of Algorithm 3.2, when the first inner tolerance test fails, the algorithm sets x_{k+1}=x_{m_k+1}^{k+1}; the stationarity equation (19) and its limit version (27) are written only for the iterate x_{m_k}^{k+1}. The sentence following (24) says the extra case follows 'by a similar derivation', but the analogous equation for x_{m_k+1}^{k+1} is not displayed, and the proof of (ii) invokes (19) directly for the final outer iterate. The gap is easily repaired because the contraction property gives ||x_{m_k+1}^{k+1}−x_{m_k}^{k+1}|| ≤ (1−μλ)||x_{m_k}^{k+1}−x_{m_k−1}^{k+1}|| ≤ δ||x_k−x_{k−1}||, so the same bound (16) holds for the extra iterate; please add this argument.","section":"Theorem 4.1(ii), around Eq. (27)"}],"minor_comments":[{"comment":"The stopping criterion is only shown to detect a critical point in the exact equality case x_{m_k}^{k+1}=x_k=x_{m_k+1}^{k+1}. In finite precision the algorithm stops when the relative distance is below tol, but no quantitative relation between tol and a subgradient residual of F is given; the text after the lemma calls the output an 'approximate critical point', which is stronger than what is proven. A bound along the lines of Lemma 4.1 would make the termination rule more defensible.","section":"Section 3.2, Lemma 3.2"},{"comment":"The parameter λ is tuned in Section 5.1 on the same distribution of random instances used for the comparisons in Table 3 and Table 4, while ADCA and pDCAe use literature defaults; the reported advantages in iterations and CPU time should be interpreted with this selection bias in mind.","section":"Sections 5.1 and 5.2"},{"comment":"Typographical issues include 'ffrstly' and 'ffxed' in the abstract, the corrupted string 'Kurdyka—/suppress Lojasiewicz' in Definition 2.1 and Remark 4.2, and minor grammar issues in the abstract; these should be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The theoretical core is sound, but the advertised whole-sequence convergence is broader than what is verified for the numerical examples. If the authors can close the gap (or explicitly restrict the claim), I would support acceptance; otherwise the paper should be revised to make the scope clear. The λ-tuning issue also weakens the empirical section but is secondary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, two things you should know about arXiv:2505.10800. First, the core idea is genuinely nice: the LPM subproblem is equivalent to the fixed point of a contraction, so the inner iteration can be solved by Picard iteration with an adaptive stopping criterion (16) that depends on the previous outer displacement. This is a real improvement over prespecified summable error tolerances, and the contraction reformulation in Lemma 3.1 is correct. Second, the paper overclaims whole-sequence convergence for one of its two test problems: Theorem 4.2 requires ∇h to be Lipschitz on the convex hull of the sequence, and for the ℓ1−2 regularizer h(x)=γ||x|| that condition is not verified and is plausibly false. The citation to [34] does not fill the gap, since that paper does not establish a positive lower bound on ||x_k|| or differentiability of h along the generated sequence.\n\nThe convergence analysis itself is mostly solid. Theorem 4.1 is careful: the descent inequality with the auxiliary function E is coherent, and it yields subsequential convergence plus finite inner termination under the adaptive rule. The KL-based whole-sequence argument in Theorem 4.2 follows a standard template and is presumably correct under its hypotheses. The numerical experiments show cDCA consistently beating ADCA and pDCAe in CPU time, which is plausible but not independently checkable without code.\n\nThe soft spots are proportionate. Lemma 3.2 analyzes a termination condition slightly different from the algorithm's actual test, but that is minor. The edge case x_k=x_{k-1} can prevent finite inner termination and is dismissed quickly. The parameter λ is tuned on the same problem distribution used for the comparisons, and no error bars or code are provided, so the reported speedups are suggestive rather than definitive. The major issue is the mismatch between the abstract's global whole-sequence claim and what is actually proven for the ℓ1−2 test case. The subproblem has no term keeping iterates away from zero, and nothing in the cited reference guarantees that the generated sequence avoids zero on its convex hull. Without that, Theorem 4.2 simply does not apply to one of the two advertised numerical examples.\n\nWho gets value from this paper? People working on DC optimization and fixed-point algorithms will find the contraction reformulation and the adaptive termination rule worth studying. The paper deserves a serious referee, but it needs revision: either prove the sequence stays bounded away from zero for this class of problems or soften the claim to subsequential convergence for the ℓ1−2 example. The core mathematics is sound; the gap is in the scope of the application claim, not in the main proof machinery.\n\nMy recommendation: send it to peer review, with the expectation of major revision. The idea is strong enough to be worth the refereeing effort.","headline":"A clean contraction-based subproblem solver with a genuinely adaptive stopping rule; the whole-sequence convergence claim overreaches for the ℓ1−2 experiments.","tokens_in":17478,"tokens_out":4130,"would_cite":true,"duration_ms":40115,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C30","65K05","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that each linearized proximal subproblem in a difference-of-convex method is a fixed point problem for a contraction, and uses that reformulation to build cDCA, an algorithm with an adaptive inner stopping rule and global…","keywords":["difference-of-convex optimization","linearized proximal method","contractive mapping","Picard iteration","adaptive termination rule","Kurdyka-Lojasiewicz property","fixed point","convergence analysis"],"falsifier":"Run cDCA on a one-dimensional problem with $f(x)=x^2/2$, $g=0$, and $h(x)=|x|^{3/2}$, so $\\nabla h$ is convex but not Lipschitz on the convex hull of any sequence passing through $0$, while the other hypotheses of Theorem 4.1 hold. If the outer sequence still converges to a critical point, the Lipschitz-gradient assumption in Theorem 4.2 is not needed; if it fails or stalls, that assumption is load-bearing. Also record whether the inner termination rule (16) is ever satisfied when two consecutive outer iterates coincide exactly.","tokens_in":16342,"feed_emoji":"🧮","tokens_out":14192,"duration_ms":127593,"temperature":0.7,"pith_summary":"The paper targets difference-of-convex problems minimizing $F = f + g - h$, where standard DCA iterations reduce to convex subproblems that usually lack closed-form solutions and are solved by slow iterative schemes. The core claim is that the linearized proximal method subproblem (10) is equivalent to the fixed point problem of a contraction, $\\operatorname{Prox}_{\\mu g}[T_k^{\\lambda,\\mu}]$, for $\\mu\\in(0,2/(2\\lambda+L_f)]$. Based on this, the paper proposes cDCA, which approximates each subproblem by Picard iteration and stops the inner loop adaptively when the inner displacement drops below a fixed fraction of the previous outer displacement. The convergence results say every cluster point is critical, the objective values converge, and under a standard asymptotic regularity condition the whole sequence converges to a critical point without any pre-tuned decreasing inner tolerance schedule.","feed_headline":"Difference-of-convex subproblems become contraction fixed points","feed_subtitle":"A new algorithm solves each inner loop by Picard iteration, stopping adaptively from the last outer step.","key_machinery":"The central object is the composite map $\\operatorname{Prox}_{\\mu g}[T_k^{\\lambda,\\mu}] = \\operatorname{Prox}_{\\mu g}[(1-\\mu\\lambda)(I-\\frac{\\mu}{1-\\mu\\lambda}\\nabla f)(\\cdot)+\\mu\\lambda x_k+\\mu\\eta_k]$. It is a contraction because $\\nabla f$ is $1/L_f$-cocoercive, so $I-\\nu\\nabla f$ is nonexpansive for $\\nu\\le2/L_f$, and $\\operatorname{Prox}_{\\mu g}$ is firmly nonexpansive; the affine shift preserves the contraction constant $1-\\mu\\lambda$. This object carries the argument: it turns the convex subproblem into a Banach fixed point problem, justifies Picard iteration, and makes the adaptive stopping rule (16) natural, because inner contraction error can be controlled by outer displacement. The auxiliary function $E$ then supplies the descent framework that drives both subsequential and whole-sequence convergence.","core_discovery":"Lemma 3.1 is the hinge: for $x_k$ and $\\eta_k\\in\\partial h(x_k)$, the unique minimizer of (10) is exactly the unique fixed point of $\\operatorname{Prox}_{\\mu g}[T_k^{\\lambda,\\mu}]$, where $T_k^{\\lambda,\\mu}(x)=(1-\\mu\\lambda)x-\\mu\\nabla f(x)+\\mu\\lambda x_k+\\mu\\eta_k$. For $\\mu\\in(0,2/(2\\lambda+L_f)]$ the composed operator is a contraction with coefficient $1-\\mu\\lambda$, and the best coefficient $L_f/(2\\lambda+L_f)$ is attained at $\\mu=2/(2\\lambda+L_f)$. The paper therefore replaces the usual inner proximal-gradient loop by Picard iteration of this contraction, terminated by (16): $\\|x_{m_k}^{k+1}-x_{m_k-1}^{k+1}\\|\\le\\delta\\|x_{k-1}-x_k\\|$ with $\\delta\\in(0,2\\lambda/L_f)$. The convergence analysis shows the auxiliary function $E(x,y)=F(x)+\\tau\\|x-y\\|^2$ decreases along the pairs $(x_k,x_{k-1})$, that outer step lengths are square-summable, and that all cluster points satisfy $0\\in\\nabla f(x)+\\partial g(x)-\\partial h(x)$. Theorem 4.2 then upgrades subsequential convergence to whole-sequence convergence when $E$ has the KL property at the limit and $\\nabla h$ is Lipschitz on the convex hull of the iterates.","pith_inferences":["Beyond the paper, the contraction constant $L_f/(2\\lambda+L_f)$ gives a direct bound on the inner iteration count: aiming for inner error $\\varepsilon$ costs about $\\log(\\varepsilon)/\\log(L_f/(2\\lambda+L_f))$ Picard steps, so the adaptive rule can be converted into an explicit worst-case complexity estimate.","Beyond the paper, the fixed-point equivalence depends on $g$ only through $\\operatorname{Prox}_{\\mu g}$, so the same contraction argument should apply to any convex prox-friendly regularizer without changing the inner-loop analysis.","Beyond the paper, a natural test of Theorem 4.2's necessity is to run cDCA with $h(x)=|x|^{3/2}$ in one dimension; if the whole sequence converges anyway, the $L_h$-Lipschitz assumption could be removed or weakened.","Beyond the paper, the adaptive stopping criterion (16) is portable: any outer method that produces a stabilizing quadratic term could adopt a relative inner-error rule keyed to the previous outer displacement instead of an absolute tolerance schedule."],"forward_implications":["Every outer iteration of cDCA needs only finitely many inner Picard steps when consecutive outer iterates differ, so no user-supplied inner tolerance sequence is required.","All cluster points of the cDCA sequence are critical points of $F$, and the sequence of objective values converges.","Adding the KL property and Lipschitz differentiability of $h$ along the iterates upgrades this to convergence of the whole sequence to a critical point, with rates when the KL exponent is known.","In the reported experiments on $\\ell_1\\text{-}2$ and logarithmic regularized least squares, cDCA uses fewer total iterations and less CPU time than ADCA and pDCAe at comparable final objective values."],"supporting_citations":[{"why":"introduces the linearized proximal method whose subproblem (10) is the object cDCA solves as a contraction fixed point.","marker":"[31]"},{"why":"supplies the cocoercivity-to-nonexpansiveness lemma used to prove the composed operator is a contraction.","marker":"[35]"},{"why":"provides the firm nonexpansiveness of proximal mappings and the gradient cocoercivity facts behind Lemma 3.1.","marker":"[5]"},{"why":"provides the KL-based descent convergence framework invoked after Lemma 4.1 for whole-sequence convergence.","marker":"[22]"},{"why":"provides the pDCAe baseline and the regularizer examples whose assumptions are cited in the numerical experiments.","marker":"[34]"},{"why":"provides the ADCA baseline used in the numerical comparisons.","marker":"[28]"},{"why":"supplies the closedness result for subdifferentials used to show cluster points are critical points.","marker":"[30]"}],"fun_headline_variants":["Contraction fixed points speed up difference-of-convex algorithms","Turning DCA subproblems into contraction fixed points","Adaptive stopping for contraction-based DCA","Picard iteration replaces inner proximal-gradient loops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"For the whole sequence to converge (not just along subsequences), the proof needs the slope of the convex function $h$ to change at a bounded rate all along the path the iterates travel, and the auxiliary function $E$ to satisfy a standard asymptotic flatness condition; neither follows from the basic assumptions and neither is verified in the experiments.","fun_headline_variants_meta":{"raw":{"variants":["Contraction fixed points speed up difference-of-convex algorithms","Turning DCA subproblems into contraction fixed points","Adaptive stopping for contraction-based DCA","Picard iteration replaces inner proximal-gradient loops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000563,"raw_usage":{"total_tokens":2723,"prompt_tokens":1047,"completion_tokens":1676,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1617}},"tokens_in":663,"tokens_out":1676,"duration_ms":11490,"temperature":1.0,"reasoning_tokens":1617,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:03:34.177010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run cDCA on a one-dimensional problem with $f(x)=x^2/2$, $g=0$, and $h(x)=|x|^{3/2}$, so $\\nabla h$ is convex but not Lipschitz on the convex hull of any sequence passing through $0$, while the other hypotheses of Theorem 4.1 hold. If the outer sequence still converges to a critical point, the Lipschitz-gradient assumption in Theorem 4.2 is not needed; if it fails or stalls, that assumption is load-bearing. Also record whether the inner termination rule (16) is ever satisfied when two consecutive outer iterates coincide exactly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces the linearized proximal method whose subproblem (10) is the object cDCA solves as a contraction fixed point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the cocoercivity-to-nonexpansiveness lemma used to prove the composed operator is a contraction."},{"cited_title":"Springer, New York, 2017","cited_arxiv_id":null,"evidence_quote":"provides the firm nonexpansiveness of proximal mappings and the gradient cocoercivity facts behind Lemma 3.1."},{"cited_title":"Imaging Sci","cited_arxiv_id":null,"evidence_quote":"provides the KL-based descent convergence framework invoked after Lemma 4.1 for whole-sequence convergence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the pDCAe baseline and the regularizer examples whose assumptions are cited in the numerical experiments."},{"cited_title":"In: Pro- ceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI-18, pp","cited_arxiv_id":null,"evidence_quote":"provides the ADCA baseline used in the numerical comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the closedness result for subdifferentials used to show cluster points are critical points."}],"review_version":1}