{"id":"39550f69-1bc3-4cc7-a7f6-a0183f077d2b","arxiv_id":"2607.19553","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An online proximal-linear algorithm for difference-of-convex-composite objectives and constraints attains O(T/w^2) local regret, with a proximal residual that certifies first-order stationarity.","lead":"This paper develops an online optimization method for a broad class of nonconvex, nonsmooth problems where costs and constraints are differences of convex functions composed with smooth maps. It proves that a proximal-residual based local regret bound holds and connects the residual to first-order stationarity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CDCS Slater condition is genuinely load-bearing: a simple 1D constraint satisfying the paper's assumptions fails it, and the proximal residual then certifies a non-stationary point.","rationale":"The reader's weakest-assumption diagnosis is correct and can be sharpened into a concrete counterexample. The central claim that the proximal residual is a proper stationarity measure is formally conditional on the CDCS Slater condition, and the paper presents no example for which the condition is verified. The counterexample with ζ(x)=x^3/(1+x^2) shows the condition is not merely sufficient but genuinely needed: when it fails, the tangent-cone equality T_{\\tilde L}=T_X can fail and a zero residual can occur at a non-stationary point. This does not invalidate the theorems as stated, but it means the paper's advertised applicability to 'CDCS constraints' is narrower than presented. The reader's CONDITIONAL verdict already captures this; my stress-test confirms the concern and makes it concrete, so no verdict change is needed.","tokens_in":25896,"tokens_out":32040,"duration_ms":264648,"concrete_test":"Implement the 1D problem with X=[0,∞) encoded as c(x)=|ζ(x)|−ζ(x), ζ(x)=x^3/(1+x^2), Φ(x)=−x. At x=0, compute \\tilde L(0) and verify it is {0}; check that the proximal residual is 0 but the Bouligand stationarity condition (24) fails because 0∉{−1}+(−∞,0]. Separately, evaluate the Slater condition (28) for Example 4's Huber constraint at an active boundary point with exactly b entries above δ and the rest zero, and confirm that c'(x;d)≥0 for all d, so Slater fails. If both checks pass, the paper must either add Slater as a standing assumption with verification for all examples or weaken the claimed stationarity certificate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formal stationarity certificate (Theorem 1, Proposition 2) depends on the CDCS Slater condition (Definition 2, eq. (28)). This is not a harmless technicality. Consider X=[0,∞) represented as c(x)=|ζ(x)|−ζ(x) with ζ(x)=x^3/(1+x^2), i.e., ϕ(t)=|t|, ψ(t)=t. This satisfies the favorable-condition assumptions: ϕ,ψ are 1-Lipschitz convex, ζ is smooth with bounded Jacobian, and |c|≤2 on R. At x=0, ζ(0)=0 and Jζ(0)=0, so c'(0;d)=0 for every d; hence the Slater condition (28) fails. Yet the cushioned linearization set is \\tilde L(0)={z: βζ∥z∥²≤0}={0}, because the quadratic cushion terms are positive, whereas T_X(0)=[0,∞). Thus T_{\\tilde L}(0)≠T_X(0), and the key inclusion C(0)⊆T_{\\tilde L}(0) in Theorem 1's proof breaks. For Φ(x)=−x, the proximal subproblem over {0} has S(0)=0, so the residual G(0)=0; the fixed-point condition holds, but 0∉b∂Φ(0)+T_X(0)^* = {−1}+(−∞,0] = (−∞,−1]. So the residual is not a proper stationarity measure without Slater. The paper does not verify Slater for its examples; e.g., the Huber-ℓ0 constraint in Example 4 also has no strict descent direction at active boundary points (c'(x;d)≥0 for all d). This gap is the main barrier to the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online optimization framework for nonconvex nonsmooth losses of CDCS form f_t(θ_t(x))−g_t(θ_t(x)) over a nonconvex feasible set X={x: φ(ζ(x))−ψ(ζ(x))≤0}, where φ,ψ are convex, ψ,g are differentiable, and θ,ζ are smooth. The core algorithmic idea is a time-smoothed proximal-linear method whose inner subproblems are convex over a \"cushioned linearization\" set L̃(·). The paper defines a proximal residual G_t^μ(x)=μ(S_t^μ(x)−x) and a local-regret measure Σ∥G_t^μ(x_t)∥². Its main theoretical claims are: (i) Theorem 1, a tangent-cone characterization T_{L̃(x̄)}=T_X(x̄)=C(x̄) under a CDCS Slater CQ; (ii) Proposition 2, stating that a fixed point of S_t^μ implies first-order stationarity 0∈∂̂Φ_t(x̄)+T_X(x̄)*; (iii) Theorem 2, giving a local-regret bound of order T(δ/w+(L_f+L_g)L_θ/w)² and an inner-iteration bound; and (iv) Proposition 4, an error bound relating ∥G_t^μ(x)∥ to dist(0,∂̂Φ_t(x⁺)+T_X(x⁺)*). The proofs of the regret and iteration bounds are largely self-contained and appear coherent.","tokens_in":26350,"tokens_out":13177,"duration_ms":110417,"significance":"If the stated results hold, this would be a meaningful extension of the local-regret framework to a broad structured nonconvex nonsmooth class with nonconvex constraints, while keeping all inner updates convex. The tangent-cone characterization in Theorem 1 is of independent interest for offline CDCS-constrained optimization. The main regret bound has the expected scaling and the iteration-complexity analysis is nontrivial. However, the significance is currently weakened by two substantial gaps: the stationarity certificate depends on a CDCS Slater condition that is not verified for the paper's own examples and is not automatic even for simple constraints; and Proposition 4 relies on smoothness assumptions on ∇g and ∇ψ that are not stated in the paper's assumptions. These gaps affect the central claim that the residual is a proper stationarity measure, and one of the paper's advertised contributions.","major_comments":[{"comment":"The CDCS Slater CQ (28) is load-bearing for the equality T_{L̃(x̄)}=T_X(x̄) and hence for Proposition 2, but it is not automatic and is not verified for the examples. Counterexample: let X=[0,∞) be represented by φ(t)=|t|, ψ(t)=t, ζ(x)=x³/(1+x²). All favorable-condition assumptions hold. At x̄=0, Jζ(0)=0, so no d0 satisfies φ'(0;0)<0; Slater fails. Then L̃(0)={0}, so T_{L̃}(0)={0}, but T_X(0)=[0,∞). With Φ(x)=−x, S(0)=0 and G(0)=0, yet 0∉∂̂Φ(0)+T_X(0)*=(−∞,−1]. Thus the residual is not a proper stationarity measure without Slater. Moreover, Example 4 (the Huber ℓ0 budget constraint, eq. (17)) fails (28) at every active point: at such a point the zero coordinates are exactly zero and the large coordinates have zero derivative, so φ'(x;d)−∇ψ(x)ᵀd = Σ_{j:x_j=0}|d_j| ≥ 0 for all d. The abstract's unqualified claim that the residual is a proper stationarity measure is therefore too broad; the","section":"§3.1, Definition 2 and Theorem 1, eq. (28)"},{"comment":"The proof of Proposition 4 invokes Lipschitz constants β_g and β_ψ for ∇g_s and ∇ψ, respectively. Neither Assumption 1 nor the favorable condition (Definition 1) includes smoothness of ∇g or ∇ψ; only L-Lipschitzness and differentiability of g and ψ are assumed. In bounding ∥∇g_s(θ_s(x⁺))−∇g_s(θ_s(x))∥ and the analogous term for ∇ψ, the proof uses exactly these unstated constants. Consequently, the error bound (81) is not established under the assumptions stated in the proposition. The proposition should be restated with explicit smoothness assumptions on ∇g_s and ∇ψ (and then these should be verified for the examples, e.g., the Huber function does have a Lipschitz gradient), or the claim should be modified to avoid these terms.","section":"§3.5, Proposition 4, eqs. (95) and (99)"}],"minor_comments":[{"comment":"The displayed bound (81) has malformed typesetting with stray '+ +' and missing parentheses, making it hard to parse; it should be corrected.","section":"§3.5, eq. (81)"},{"comment":"The phrase 'proper stationarity measure for the original problem' should be qualified by the CDCS Slater condition, since Proposition 2 explicitly depends on it.","section":"Abstract and §1.1"},{"comment":"The constraint (15) is only sufficient, not equivalent, to the ℓ0 budget (14); the paper says 'it suffices,' but the subsequent discussion of feasible regions should not imply that (15) exactly characterizes the sparsity constraint.","section":"Example 4, eqs. (14)–(17)"},{"comment":"The reference 'Zhu and Li (2026)' cites arXiv:2607.08954 with a July 2026 date; please verify that this reference is complete and accurately described.","section":"§1.3"}],"recommendation":"major_revision","confidential_remarks":"The main theorems are technically sophisticated and the regret bound appears sound, but the stress-test concern is legitimate: the CDCS Slater condition is not a benign technical assumption, it fails for a simple one-dimensional constraint satisfying the paper's regularity assumptions, and it is not verified for the paper's own Example 4. The authors should either add the Slater condition as an explicit standing assumption throughout, prove it for the examples, or substantially revise the claims about the residual being a proper stationarity measure. The unstated smoothness in Proposition 4 is also a real correctness gap, though it does not affect Theorem 2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real extension, not a repackaging. It is the first to put CDCS objectives together with CDCS constraints in the online local-regret framework, and the tangent-cone characterization (Theorem 1) is a new technical tool. The proximal residual is a sensible stationarity measure when the CDCS Slater condition holds: the proof of Theorem 1 is coherent, and Proposition 2 follows from it. The regret and iteration-complexity bounds in Theorem 2 are standard but correct, given the Lipschitz/smoothness assumptions. I would send this to referees.\n\nThe soft spots are real. First, the CDCS Slater condition (Definition 2) is load-bearing. A bounded variant of the stress-test example makes this concrete: take X=[0,∞) via c(x)=|ζ(x)|−ζ(x) with ζ=x^3/(1+x^4). At x=0, Jζ(0)=0 so no d0 satisfies the strict inequality; the cushioned set collapses to {0}, T_{~L}(0)≠T_X(0), and for Φ(x)=−x the residual is zero at 0 while 0 is not stationary. The stress-test note as written uses ζ=x^3/(1+x^2) and says the range is bounded, which is wrong because c is unbounded, but the bounded variant makes the same point. The paper does not verify Slater for Example 4; the Huber-ℓ_0 constraint appears to fail it at saturated boundary points. Second, Proposition 4's proof uses β_g and β_ψ smoothness of ∇g_s and ∇ψ that are not in Assumption 1 or the favorable condition. That is fixable by adding assumptions, but as written the error-bound claim is not fully supported. Third, the paper does not discuss what happens when Slater fails: the residual may still be a reasonable heuristic, but it is not a certified stationarity measure, and the local-regret bound loses its interpretation.\n\nI do not think these are fatal. The main framework and the tangent-cone geometry are sound, and the missing pieces are identifiable rather than hidden. But the paper should not be accepted as is: the Slater condition needs to be checked for the examples, and Proposition 4's hypotheses need to be stated. A serious editor should send this to peer review, with referees asked to focus on Slater verification and the error-bound assumptions.","headline":"A genuine extension of local-regret to CDCS objectives with CDCS constraints, but the stationarity certificate depends on a Slater condition that is not verified in the examples; worth refereeing after revision.","tokens_in":26823,"tokens_out":7126,"would_cite":false,"duration_ms":56119,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C25","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that a proximal residual computed from convex subproblems certifies first-order stationarity for online nonconvex optimization with difference-of-convex composite losses and constraints.","keywords":["online optimization","difference-of-convex","nonconvex nonsmooth optimization","proximal residual","stationarity measure","tangent cone","local regret","constraint qualification"],"falsifier":"Build a two-constraint CDCS example where both constraints are active at a boundary point so condition (28) has no solution, run the algorithm, and check whether a fixed point x̄=S_t^μ(x̄;v) still satisfies 0∈b∂Φ_t(x̄)+T_X(x̄)^*. If a fixed point occurs without the stationarity inclusion, that example marks exactly where the Slater assumption is load-bearing.","tokens_in":25776,"feed_emoji":"🎯","tokens_out":7913,"duration_ms":75027,"temperature":0.7,"pith_summary":"The paper studies online learning where each round's loss is a difference-of-convex function composed with a smooth map, and where the feasible region is itself defined by a constraint of the same kind. Because generic nonconvex problems are intractable, the paper adopts a stationarity-based benchmark rather than global optimality. Its central claim is that a time-smoothed proximal-linear algorithm, whose inner updates are convex subproblems, has a natural residual that is a proper stationarity measure: whenever the residual's fixed-point condition holds at a point, that point satisfies a first-order stationarity condition for the original nonconvex constrained problem. The key enabling result is a tangent-cone characterization showing that a carefully cushioned linearization of the constraint set preserves the local geometry of the true feasible set. If correct, the paper delivers a local-regret bound, a bound on the number of inner convex solves, and an error bound that turns the residual into a computable certificate of approximate stationarity.","feed_headline":"Proximal residual certifies stationarity in nonconvex online learning","feed_subtitle":"A fixed point of the convex-computed residual means first-order stationarity of the original nonconvex problem.","key_machinery":"The cushioned linearization set L̃(x): a convex set formed by linearizing the inner smooth map ζ at x inside the convex function φ, linearizing the entire composite concave part ψ∘ζ, and adding quadratic cushion terms ±½Lβ‖y−x‖² to both sides. Theorem 1 says that under the CDCS Slater condition, T_{L̃(x)}(x)=T_X(x)=C(x), where C(x) is the cone of directions d satisfying φ'(ζ(x);Jζ(x)d) ≤ ∇ψ(ζ(x))ᵀJζ(x)d. This tangent-cone equality is what allows the proximal-linear mapping S_t^μ(x), a convex problem over L̃(x), to certify stationarity of the original nonconvex problem. The proximal residual G_t^μ(x) is the scaled displacement produced by this mapping.","core_discovery":"On its own terms, the paper's discovery is that the proximal-residual mapping G_t^μ(x)=μ(S_t^μ(x)−x), defined as the difference between the current point and the minimizer of a convex proximal-linear surrogate, is a faithful first-order stationarity measure for the nonconvex constrained problem. Proposition 2 proves that the fixed-point condition x=S_t^μ(x;v) implies 0∈b∂Φ_t(x)+T_X(x)^*, which is Bouligand stationarity of the time-smoothed loss over the original feasible set. The mechanism is Theorem 1: under a CDCS Slater condition, the tangent cone of the convex surrogate set equals the tangent cone of the true feasible set, so moving between the surrogate and the original problem loses no","pith_inferences":["Not in the paper: the tangent-cone characterization is likely reusable for offline optimization with CDCS constraints, where it would justify convex-surrogate algorithms without a separate stationarity analysis.","Not in the paper: when the feasible region is a max of several CDCS constraints, the single-constraint Slater condition must be checked for that max representation; verifying or relaxing it for multiple simultaneously active constraints would extend the framework to common constraint structures.","Not in the paper: replacing the uniform sliding window with adaptive or exponentially weighted smoothing should preserve the proof's logic, since the regret argument only needs the smoothed objectives to change by a bounded amount between rounds.","Not in the paper: the residual could serve as a practical stopping criterion in stochastic or bandit-like online settings where only partial loss information is available at decision time."],"forward_implications":["The local-regret bound Regret_T ≲ T(δ/w+(L_f+L_g)L_θ/w)² says the iterates are approximately stationary on average, with per-round deviation controlled by the stopping tolerance and the smoothing window.","Each inner iteration solves a convex optimization problem, so the method runs with a convex oracle even though the objective and constraints are nonconvex.","The bound on total inner iterations, O(Mμ²w²/(μ−(L_f+L_g)β_θ)δ⁻²(1+T/w)), keeps the computational overhead finite and polynomial in the precision requirement.","Feeding the same loss at every round recovers an offline method: suitable choices of window and tolerance yield an iterate whose expected squared residual is below any ε, i.e., a first-order stationary point.","Under additional regularity, Proposition 4's error bound makes the residual a quantitative certificate: a small residual implies the next iterate is close to the true stationarity set."],"fun_headline_variants":["Proximal residual fixed point is a stationarity certificate in online nonconvex","Convex-computed residual: a faithful stationarity measure for nonconvex online","Fixed point of proximal residual implies first-order stationarity (online nonconvex)","Proximal residual: a convex-oracle check for stationarity in nonconvex online","Residual fixed point = stationarity for nonconvex online optimization"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"At every boundary point of the feasible region where the constraint is active, there must exist a direction along which the convex part grows strictly slower than the linearized concave part; without such a Slater-type direction, the cushioned surrogate set may fail to reproduce the true tangent cone, and the stationarity certificate can break.","fun_headline_variants_meta":{"raw":{"variants":["Proximal residual fixed point is a stationarity certificate in online nonconvex","Convex-computed residual: a faithful stationarity measure for nonconvex online","Fixed point of proximal residual implies first-order stationarity (online nonconvex)","Proximal residual: a convex-oracle check for stationarity in nonconvex online","Residual fixed point = stationarity for nonconvex online optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000717,"raw_usage":{"total_tokens":3048,"prompt_tokens":721,"completion_tokens":2327,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":2223}},"tokens_in":465,"tokens_out":2327,"duration_ms":17167,"temperature":1.0,"reasoning_tokens":2223,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:24:06.055455+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a two-constraint CDCS example where both constraints are active at a boundary point so condition (28) has no solution, run the algorithm, and check whether a fixed point x̄=S_t^μ(x̄;v) still satisfies 0∈b∂Φ_t(x̄)+T_X(x̄)^*. If a fixed point occurs without the stationarity inclusion, that example marks exactly where the Slater assumption is load-bearing.","supporting_citations":[],"review_version":1}