{"id":"8ee39e97-2c09-4154-8464-8e212e501936","arxiv_id":"2606.11773","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"OMWU has last-iterate asymptotic convergence to a saddle point for smooth convex-concave saddle-point problems with small constant step size, via a new boundary argument on cluster points.","lead":"The paper proves last-iterate asymptotic convergence of Optimistic Multiplicative-Weights Update (OMWU) to a saddle point for smooth convex-concave problems using a small constant learning rate. This fills a known gap for the entropic version of a popular saddle-point solver used in optimization and machine learning.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"Validity of the boundary argument establishing inactive-coordinate KKT conditions for OMWU cluster points","rationale":"The reader's weakest_assumption correctly isolates the single step whose failure would invalidate the last-iterate result. Because the paper supplies an explicit (if appendix-only) argument for this step, the appropriate response is to test that argument directly rather than adjust the overall verdict without verification.","tokens_in":1653,"tokens_out":285,"duration_ms":13010,"concrete_test":"Extract the precise OMWU update rule and the statement of the boundary lemma from the appendix; independently re-derive the limit inequality for inactive coordinates by taking the limit along a convergent subsequence without invoking the ChatGPT-assisted steps, and check whether the resulting variational inequality holds for the optimistic term.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that every cluster point of the OMWU trajectory satisfies the KKT conditions on inactive coordinates (those that approach the boundary). This is presented as the main new technical ingredient, documented in the appendix. Standard last-iterate arguments for OGDA rely on interior-point behavior or variational inequalities that do not automatically extend when coordinates hit zero under the multiplicative (entropic) update; if the boundary argument contains a gap specific to the optimistic correction term in the OMWU recursion, the asymptotic convergence conclusion fails to hold for general convex-concave problems.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims to establish last-iterate asymptotic convergence of the Optimistic Multiplicative Weights Update (OMWU) algorithm to a saddle point for smooth convex-concave saddle-point problems, using a sufficiently small constant learning rate. The result requires no uniqueness, strict complementarity, error bounds, or special initialization. The central technical contribution is a new boundary argument (placed in the appendix) showing that every cluster point of the trajectory satisfies the KKT conditions on coordinates that approach the boundary.","tokens_in":1760,"tokens_out":310,"duration_ms":7979,"significance":"If the boundary argument is valid, the result would be significant: it extends the classical last-iterate convergence known for OGDA since the 1980s to the entropic (multiplicative) setting of OMWU, which is widely used in game theory and online learning. The proof is presented as self-contained and free of auxiliary assumptions on the solution set.","major_comments":[{"comment":"Appendix (boundary argument): the claim that every cluster point satisfies the inactive-coordinate KKT inequalities must be verified specifically for the OMWU recursion, including the interaction between the optimistic correction term and coordinates that reach zero under the multiplicative update. Standard interior-point or variational-inequality arguments used for OGDA do not automatically carry over; a gap here would render the asymptotic convergence statement false for general convex-concave problems.","section":"Appendix"}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their careful review and for highlighting the need for explicit verification of the boundary argument. We address the concern below and will revise the appendix accordingly.","responses":[{"response":"We agree that a self-contained verification specific to the OMWU recursion is required and that standard OGDA arguments do not transfer directly. The appendix already contains a direct analysis of the OMWU update rule in the limit: when a coordinate x_i approaches the boundary, the multiplicative factor forces the product term to zero while the optimistic correction is shown to remain bounded, yielding the required KKT inequality for inactive coordinates. To strengthen clarity, we will expand the appendix with an additional lemma that isolates the interaction between the optimistic term and the multiplicative update near zero, including explicit limit calculations for both the primal and dual variables. This revision will make the argument fully explicit without relying on interior-point analogies.","revision_made":"yes","referee_comment":"[Appendix] Appendix (boundary argument): the claim that every cluster point satisfies the inactive-coordinate KKT inequalities must be verified specifically for the OMWU recursion, including the interaction between the optimistic correction term and coordinates that reach zero under the multiplicative update. Standard interior-point or variational-inequality arguments used for OGDA do not automatically carry over; a gap here would render the asymptotic convergence statement false for general convex-concave problems."}],"tokens_in":1250,"tokens_out":299,"duration_ms":10987,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that this paper gives the first proof that the last iterate of OMWU converges asymptotically to a saddle point for smooth convex-concave problems with a small fixed learning rate. It avoids the usual extra assumptions like uniqueness or strict complementarity.\n\nThe result itself is useful because OMWU is the entropic counterpart to OGDA, and the Euclidean case had this guarantee since the 1980s while the multiplicative version did not. The paper states the claim cleanly and isolates the new technical step: a boundary argument showing that every cluster point satisfies the inactive-coordinate KKT inequalities. That step is placed in the appendix and is presented as the ingredient that lets the argument go through when coordinates approach the boundary under the multiplicative update.\n\nThe soft spot is exactly that boundary argument. The stress-test note flags whether it correctly accounts for the optimistic correction term once coordinates near zero; standard interior or variational-inequality arguments do not automatically carry over. Because the argument was found with ChatGPT assistance and is not visible in the abstract, its correctness is the load-bearing piece. If it holds, the conclusion follows; if it has a gap specific to OMWU, the asymptotic claim does not.\n\nThe work is aimed at researchers who need last-iterate guarantees for non-Euclidean saddle-point algorithms in optimization and machine learning. A reader who cares about algorithm choice in min-max settings would find the result relevant once the appendix is checked. It deserves a serious referee because it addresses a documented open question with a direct mathematical argument rather than fitting or simulation.\n\nRecommendation: send it to review, but instruct the referees to examine the boundary argument in detail.","headline":"OMWU last-iterate convergence is now claimed via a boundary KKT argument, but that step is the part that needs direct verification.","tokens_in":2216,"tokens_out":406,"would_cite":true,"duration_ms":18184,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"OMWU converges asymptotically to saddle points in smooth convex-concave problems with small constant learning rate.","keywords":["optimistic multiplicative weights update","last-iterate convergence","saddle-point problems","convex-concave optimization","asymptotic convergence","KKT conditions","boundary argument"],"falsifier":"A concrete counterexample, either numerical or analytical, in which the sequence generated by OMWU with a small constant learning rate fails to approach any saddle point in a smooth convex-concave problem.","tokens_in":2546,"feed_emoji":"","tokens_out":578,"duration_ms":16638,"temperature":0.7,"pith_summary":"The paper proves that the last iterate of Optimistic Multiplicative Weights Update converges to a saddle point for smooth convex-concave saddle-point problems. Convergence holds with any sufficiently small fixed learning rate and requires none of the usual extra conditions such as uniqueness of the solution, strict complementarity, or initialization near an equilibrium. The result closes the gap with the known last-iterate behavior of its Euclidean counterpart, Optimistic Gradient Descent Ascent. Establishing this property matters because many applications rely on the final output of the algorithm rather than on averages over the entire trajectory.","feed_headline":"OMWU last iterate converges to saddle points","feed_subtitle":"Small constant learning rate suffices for asymptotic convergence in smooth convex-concave problems without uniqueness or initialization assu","key_machinery":"A boundary argument showing that every cluster point satisfies the inactive-coordinate KKT inequalities.","core_discovery":"OMWU converges asymptotically for smooth convex-concave saddle-point problems, with a small enough constant learning rate. The result does not require uniqueness, strict complementarity, an error bound, or initialization near a solution. The main new ingredient is a boundary argument showing that every cluster point satisfies the inactive-coordinate KKT inequalities.","pith_inferences":["Similar boundary arguments could be checked for other non-Euclidean optimistic methods.","The result suggests that last-iterate convergence may hold more broadly for multiplicative-weights dynamics than previously known.","Applications that rely on single-point outputs, such as certain game-solving or equilibrium-finding tasks, gain a stronger justification for using OMWU."],"forward_implications":["OMWU can be used without averaging the iterates or requiring special initialization.","Convergence holds even when the set of saddle points is not a singleton.","No error-bound or strict-complementarity condition is needed for the guarantee.","The algorithm remains reliable for arbitrary starting points inside the domain."],"fun_headline_variants":["OMWU last iterate converges in smooth convex-concave problems","Asymptotic OMWU convergence holds with small constant rate","OMWU reaches saddle points without uniqueness assumptions","Boundary argument proves OMWU KKT satisfaction at clusters","OMWU converges last iterate for saddle-point problems"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The boundary argument that every cluster point satisfies the inactive-coordinate KKT inequalities holds for the specific dynamics of OMWU.","fun_headline_variants_meta":{"raw":{"variants":["OMWU last iterate converges in smooth convex-concave problems","Asymptotic OMWU convergence holds with small constant rate","OMWU reaches saddle points without uniqueness assumptions","Boundary argument proves OMWU KKT satisfaction at clusters","OMWU converges last iterate for saddle-point problems"]},"model":"grok-4.3","cost_usd":0.003819,"raw_usage":{"total_tokens":1928,"prompt_tokens":586,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":38187000,"prompt_tokens_details":{"text_tokens":586,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1270,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":586,"tokens_out":72,"duration_ms":9664,"temperature":1.0,"reasoning_tokens":1270,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T09:01:07.480604+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A concrete counterexample, either numerical or analytical, in which the sequence generated by OMWU with a small constant learning rate fails to approach any saddle point in a smooth convex-concave problem.","supporting_citations":[],"review_version":1}