{"id":"b0e3777a-e213-43f0-9118-8aba176bba93","arxiv_id":"2603.04940","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"NSGDA-M is claimed to achieve O(ε⁻⁴) stochastic complexity for nonconvex-strongly concave minimax under generalized smoothness, but the constrained-case proof rests on a false lemma.","lead":"An optimization paper claims a new stochastic algorithm, NSGDA-M, that solves nonconvex min-max problems at O(ε⁻⁴) cost with constant sample sizes under a relaxed smoothness assumption. The proof, however, relies on a lemma that fails for constrained feasible sets, so the main theorems are not established as stated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3 is false when Y is compact convex and y*(x) lies on the boundary; Theorems 1 and 2 rely on it through (6) and (12), so the stated constrained-case guarantees are unsupported.","rationale":"The reader's weakest assumption is exactly the load-bearing flaw. Lemma 3 is not cosmetic; it is the only mechanism that converts the squared-gradient term ηy²∥∇yL∥² in the y-update inequality into a duality gap that can be absorbed by small ηy. In the compact-convex case, a boundary maximizer has zero duality gap but generically a nonzero gradient, so the inequality fails. Because Assumption 2(b) explicitly includes compact convex Y and Theorems 1 and 2 are stated for it, the central complexity claims are not established as stated. The unconstrained R^m case may be salvageable and the NSGDA-M idea may be sound, but the present manuscript overclaims. The concrete counterexample satisfies the assumptions, so this is a genuine correctness gap rather than a stylistic issue.","tokens_in":29616,"tokens_out":5588,"duration_ms":58309,"concrete_test":"Instantiate the counterexample: take L(x,y)=y−(µ/2)y² with µ=0.5, Y=[0,1], and x fixed. Verify that y*(x)=1, ∇yL(x,1)=0.5, and the duality gap is 0. Compute y′=1+0.5/Ly0 for any Ly0≥0.5; confirm y′∉Y, so the step in Lemma 3 that uses the definition of y* is invalid. Then evaluate recursion (6) at y_t=1 and compare the claimed contraction with the actual projected update y_{t+1}=1.","verdict_should_be":"REJECT","load_bearing_attack":"Lemma 3 asserts, for any (x,y), that ∥∇yL(x,y)∥² ≤ 2(Ly0+Ly1∥∇yL(x,y)∥)(L(x,y*(x))−L(x,y)). The proof sets y′ = y + ∇yL(x,y)/(Ly0+Ly1∥∇yL(x,y)∥) and then invokes L(x,y*(x)) ≥ L(x,y′), which requires y′ ∈ Y. Under Assumption 2(b), Y may be a compact convex set, and y′ can be infeasible. Concrete counterexample: L(x,y) = y − (µ/2)y², Y = [0,1], µ<1. Then y*(x)=1, ∇yL(x,1)=1−µ>0, and the gap L(x,1)−L(x,1)=0, so Lemma 3 would give (1−µ)²≤0. This L satisfies Assumptions 1–4 with Ly0=µ, Ly1=0, B=1−µ. The same failure persists for y=1−ε: LHS ≈ (1−µ)² while RHS = O(ε). Lemma 3 is used verbatim to derive the y-tracking contraction inequalities (6) and (12); without it, the term −2ηy(1−ηyL̄y)(gap) that absorbs ηy²∥∇yL∥² is lost, and the tracking bound ∥yt−y*(xt)∥≤1/(8Lx,1) is not established. Since Theorems 1–2 explicitly include Assumption 2(b), and the experiments use a simplex, the main claim is not proven for the stated problem class.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies stochastic nonconvex-strongly concave minimax problems (1) under (L0,L1)-generalized smoothness with Y either R^m or a compact convex set. It proposes NSGDA-M (Algorithm 1), which updates the inner variable by projected stochastic gradient ascent and the outer variable by normalized SGD with momentum. The main theoretical claims are an O(epsilon^-4) expected stochastic-gradient complexity (Theorem 1) and an O(epsilon^-4 (log(1/delta))^{3/2}) high-probability complexity (Theorem 2) with a constant batch size, improving on the O(epsilon^-2) batch size and delta^{-4} factor of Xian et al. An appendix analyzes normalized SGDA. Numerical experiments on simplex-constrained DRO logistic regression are reported.","tokens_in":29965,"tokens_out":16446,"duration_ms":160198,"significance":"If the results were valid, the paper would make a useful contribution to the growing literature on nonconvex minimax optimization under generalized smoothness. The analysis is self-contained, the stepsize choices are explicit, and no fitted parameters or circularly tuned constants are used; the constant-batchsize aspect and the improved delta dependence are meaningful. However, the central lemma used to control the y-tracking error is false for compact convex Y, and the constrained case, including the paper's own experiments, is exactly where the proof fails. The advertised improvement over [34] is therefore not established for the stated problem class.","major_comments":[{"comment":"Lemma 3 asserts that for any (x,y) in R^n x R^m, ||grad_y L(x,y)||^2 <= 2(Ly0 + Ly1||grad_y L(x,y)||)(L(x,y*(x))-L(x,y)). The proof sets y' = y + grad_y L(x,y)/(Ly0 + Ly1||grad_y L(x,y)||) and then invokes L(x,y*(x)) >= L(x,y'), which requires y' to be feasible for Y. Under Assumption 2(b), Y may be a compact convex set and this feasibility can fail. A concrete counterexample is L(x,y) = y - (mu/2)y^2 with Y = [0,1] and 0 < mu < 1. Then y*(x)=1, grad_y L(x,1)=1-mu>0, while the gap L(x,y*(x))-L(x,y*(x)) is zero; Lemma 3 would give (1-mu)^2 <= 0. This L satisfies Assumptions 1-4 with Ly0=mu, Ly1=0, and B=1-mu. The proof's key step is therefore invalid for the constrained case.","section":"Section 3, Lemma 3"},{"comment":"The false Lemma 3 is the only source of the negative gap term that absorbs eta_y^2||grad_y L||^2 in the y-tracking recursion. In deriving (6) and (12), the term -2 eta_y (1 - eta_y Lbar_y)(L(x_t,y*(x_t))-L(x_t,y_t)) is used to obtain contraction. Without Lemma 3, that term is unavailable, and the recursive estimate leading to ||y_t-y*(x_t)|| <= 1/(8 Lx,1) on the stopping-time event is not established. This bound is load-bearing: it is used to define T0, to bound I2 in the momentum-error estimate, and to justify the final stationarity guarantee. The same defect propagates to the constrained-case claims in the Appendix Theorems 3-4. Since Assumption 2(b) explicitly includes compact convex Y and the experiments use a simplex, the main theorems do not support the paper's stated claims for that setting.","section":"Theorems 1 and 2, inequalities (6) and (12)"}],"minor_comments":[{"comment":"The claim of O(epsilon^-4) in expectation should be qualified. The displayed parameter choice contains delta-dependent terms, and the final complexity expression includes 1/(epsilon^3 sqrt(delta)) and 1/(epsilon^2 delta). The O(epsilon^-4) statement only holds when delta is treated as a fixed constant or when epsilon is sufficiently small relative to delta.","section":"Section 3.1, discussion after Theorem 1"},{"comment":"The proof cites Lemma 4 before Lemma 4 is introduced in Section 3.2; renumber or add a forward reference.","section":"Lemma 1 proof"},{"comment":"Assumption 3 uses 1/Lx,1 and 1/Ly,1, which are undefined when those constants are zero. State the convention for zero constants, and make explicit in Lemma 3 that the constructed point y' must lie in Y when Y is a proper subset of R^m.","section":"Assumption 3 / Lemma 3"},{"comment":"The normalized update x_{t+1} = x_t - eta_x m_{t+1}/||m_{t+1}|| is undefined when m_{t+1}=0; specify a convention, e.g., x_{t+1}=x_t in that case.","section":"Algorithm 1, eq. (3)"},{"comment":"The front-matter abstract and the abstract at the start of the full text disagree. The former mentions an expected stationarity bound O(delta^{-3/4}T^{-1/4}) + G_Phi delta and a T^{-1/7} rate that do not appear in the body, and the two abstracts state different logarithmic dependencies for the high-probability result. These should be aligned.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"I agree with the reader's assessment. The false Lemma 3 is not a local presentation issue: the constrained case is an advertised part of the contribution, and the proof breaks exactly there. The unconstrained case may be salvageable, but the current manuscript does not support Theorems 1-4 as stated for compact convex Y. A resubmission with either a correct projected-gradient tracking argument or a restricted scope and appropriately reinterpreted experiments would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper has a real idea, but the main theorems as stated are not proven. The algorithm (normalized momentum on x, projected stochastic ascent on y) is a natural extension of normalized SGD to minimax, and the claim of O(ε⁻⁴) with constant batchsize under generalized smoothness would be a genuine improvement over Xian et al.'s Θ(ε⁻²) batchsize. The high-probability analysis via a direct martingale argument is also a reasonable attempt to shave the δ dependence from δ⁻⁴ to log^{3/2}. Credit where due: the paper is self-contained, the parameter choices follow from the analysis, and the appendix gives a full treatment of normalized SGDA that the prior work only gestured at.\n\nThe soft spot is not cosmetic; it's load-bearing. Lemma 3 asserts a quadratic-growth inequality for the full y-gradient in terms of the duality gap. The proof constructs y′ by stepping along the gradient and uses feasibility of y′ to invoke the definition of y*(x). That only works when Y is the whole space. When Y is a compact convex set, y′ can leave Y. The counterexample is simple: L(x,y)=y−(µ/2)y² on [0,1] with µ<1. At y=y*(x)=1 the gradient is 1−µ>0 while the gap is zero, so Lemma 3 gives (1−µ)²≤0. This L satisfies the assumptions with B=1−µ. Since the paper explicitly includes compact convex Y in Assumption 2(b), and the experiments use a simplex, this is not a side issue. Theorems 1 and 2 both rely on Lemma 3 through inequalities (6) and (12); without it the y-tracking recursion loses the term that absorbs the squared gradient, and the stated tracking bound does not follow.\n\nThere are also presentation problems. The arXiv abstract advertises an O(δ^{-3/4}T^{-1/4})+G_Φδ bound that does not appear in the body, and the high-probability complexity in the abstract (O(ε⁻⁴log(1/(εδ)))) differs from the theorem (O(ε⁻⁴ log^{3/2}(1/δ))). No code or error bars, but that's minor for a theory paper.\n\nThe unconstrained case Y=R^m looks salvageable—there y′ is always feasible and Lemma 3 goes through. So the core idea is worth pursuing, and the bug is specific enough that a careful revision might fix it. If I were an editor, I would send this to referees rather than desk-reject: the question is timely and the mistake is identifiable, not a sign of sloppiness in the whole framework. But as submitted, the constrained-case claims are unsupported. I'd tell the authors to fix Lemma 3, restrict the claims to what they can prove, and clean up the abstract.","headline":"The constant-batchsize idea is sensible, but Lemma 3 is false on compact Y, so the stated main theorems don't cover the constrained case; the unconstrained version may survive.","tokens_in":30524,"tokens_out":2368,"would_cite":false,"duration_ms":26971,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C47","65K05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A normalized stochastic descent-ascent method with momentum reaches ε-stationarity in O(ε⁻⁴) stochastic-gradient evaluations for nonconvex–strongly concave minimax problems under generalized smoothness, with a constant batch size.","keywords":["nonconvex-strongly concave minimax optimization","(L0,L1)-generalized smoothness","normalized stochastic gradient descent ascent","momentum","constant batch size","ε-stationary point","stochastic oracle complexity","distributionally robust optimization"],"falsifier":"Take L(x,y)=f(x)+y−(μ/2)y² on Y=[0,1] with 0<μ<1 and any bounded nonconvex f. The unique y*-maximizer is 1 for every x; at y=1, ∇yL=1−μ>0 while the right-hand side of Lemma 3 is 0, so Lemma 3's conclusion is false. Since Lemma 3 is invoked in inequalities (6) and (12), checking whether NSGDA-M's stated O(ε⁻⁴) tracking bound holds on this instance (or on a simplex-constrained robust regression problem with a boundary maximizer) directly tests the constrained-case claim.","tokens_in":29396,"feed_emoji":"🧮","tokens_out":9311,"duration_ms":86293,"temperature":0.7,"pith_summary":"This paper tries to show that stochastic minimax optimization beyond Lipschitz smoothness does not need large batches. It proposes NSGDA-M, a single-loop method that updates the outer variable by normalized stochastic gradient descent with momentum and the inner variable by projected stochastic gradient ascent, and claims an ε-stationary point of the primal function Φ(x)=max_{y∈Y} L(x,y) in O(ε⁻⁴) stochastic-gradient evaluations in expectation and O(ε⁻⁴(log(1/δ))^{3/2}) with high probability. Under the (L0,L1) generalized smoothness condition, this would match the known ε⁻⁴ oracle lower bound for nonconvex stochastic optimization and remove the Θ(ε⁻²) batch-size requirement of earlier generalized-smooth minimax methods. The theorems cover both unconstrained and compact convex inner feasible sets; the proof's y-tracking recursion, which keeps the inner iterate near the exact maximizer y*(x), is the step on which the rates depend.","feed_headline":"Constant-batch minimax method reaches ε-stationarity in O(ε⁻⁴)","feed_subtitle":"For generalized-smooth nonconvex–strongly concave problems, momentum normalization removes the Θ(ε⁻²) batch requirement.","key_machinery":"The load-bearing object is the tracking error ∥y_t − y*(x_t)∥ and the two inequalities that control it. Lemma 1 gives Lipschitz continuity of the exact maximizer y*(·) with modulus κ=(Ly,0+Ly,1B)/μ; Lemma 3 bounds the squared inner gradient by 2(Ly,0+Ly,1∥∇yL∥)(L(x,y*(x))−L(x,y)), converting the y-ascent gain into a contraction. Plugging these into the y recursion produces inequality (6) in the expectation proof and inequality (12) in the high-probability proof. On the x side, the normalized momentum step makes the noise enter as a geometrically discounted martingale sum, which is what lets the method work with a single sample per iteration.","core_discovery":"The paper's central claim is that NSGDA-M achieves the optimal ε-dependence for nonconvex stochastic optimization in a minimax setting with generalized smoothness: O(ε⁻⁴) stochastic-gradient calls in expectation (Theorem 1) and O(ε⁻⁴(log(1/δ))^{3/2}) with probability at least 1−δ (Theorem 2), with a batch size of one. The x update uses a momentum-accelerated normalized stochastic gradient, and the y update is a projected stochastic ascent step. The only initial-condition requirement is that y0 be within O(√δ/Lx,1) (or O(1/Lx,1) in Theorem 2) of y*(x0). For the constrained case, Assumption 4 bounds the inner gradient at the maximizer, and the analysis bounds the tracking error ∥y_t−y*(x_t)∥ b","pith_inferences":["(Editorial extension) The proof's dependence on Lemma 3 suggests that the constrained-case theorem is only as solid as the feasibility of the trial point; a repair would likely project the trial step onto the tangent or normal cone of Y, and if that repair works, the same rates should carry over.","(Editorial extension) The normalized-momentum mechanism isolates noise handling from sample size; substituting a variance-reduced estimator for the momentum term is a natural next step that could sharpen the log(1/δ) factor in the high-probability bound.","(Editorial extension) If the claimed bounds hold, the practical message for distributionally robust optimization is that single-sample updates are enough, making generalized-smooth minimax solvers usable in streaming settings without growing batches.","(Editorial extension) Testing the algorithm on an exact boundary-maximizer example would separate a proof gap from a genuine algorithmic failure; if the method still converges at the predicted rate despite Lemma 3 failing, the theorem may be repairable without changing the algorithm."],"forward_implications":["If the guarantee holds, generalized-smooth nonconvex–strongly concave minimax problems are solved with the same ε⁻⁴ oracle complexity as standard Lipschitz-smooth nonconvex stochastic optimization.","The constant batch size enables streaming or single-pass use, since one sample per iteration suffices instead of batches of size Θ(ε⁻²).","The failure-probability dependence improves to (log(1/δ))^{3/2}, making high-probability certificates practical at small δ.","The stated theorems require a warm start y0 near y*(x0); absent that, the tracking-error argument is not covered.","Experiments on distributionally robust logistic regression over nine benchmark datasets show the batch-size-one momentum variant converging as reliably as a batch-size-50 normalized ascent baseline."],"fun_headline_variants":["Minimax beyond Lipschitz: O(ε⁻⁴) with constant batch","NSGDA-M: One-sample updates hit optimal ε-stationarity","Stochastic minimax at optimal cost: O(ε⁻⁴) without Lipschitz","Generalized smoothness? Still O(ε⁻⁴) with batch one"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The rate depends on Lemma 3, which assumes that moving from a current inner point along its gradient direction (with step size governed by the local smoothness constants) produces a point still inside the feasible set Y; for compact convex Y with a boundary maximizer this can fail, and then the tracking-error contraction used in Theorems 1 and 2 does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Minimax beyond Lipschitz: O(ε⁻⁴) with constant batch","NSGDA-M: One-sample updates hit optimal ε-stationarity","Stochastic minimax at optimal cost: O(ε⁻⁴) without Lipschitz","Generalized smoothness? Still O(ε⁻⁴) with batch one"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":2962,"prompt_tokens":815,"completion_tokens":2147,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":2057}},"tokens_in":559,"tokens_out":2147,"duration_ms":15992,"temperature":1.0,"reasoning_tokens":2057,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T05:53:53.993622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take L(x,y)=f(x)+y−(μ/2)y² on Y=[0,1] with 0<μ<1 and any bounded nonconvex f. The unique y*-maximizer is 1 for every x; at y=1, ∇yL=1−μ>0 while the right-hand side of Lemma 3 is 0, so Lemma 3's conclusion is false. Since Lemma 3 is invoked in inequalities (6) and (12), checking whether NSGDA-M's stated O(ε⁻⁴) tracking bound holds on this instance (or on a simplex-constrained robust regression problem with a boundary maximizer) directly tests the constrained-case claim.","supporting_citations":[],"review_version":1}