{"id":"47f60483-c105-45a8-b808-ccabcaacff03","arxiv_id":"2607.07078","paper_version":1,"verdict":"ACCEPT","confidence":"UNKNOWN","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A forgetting-factor regret metric with exponentially decaying weights is introduced for online zero-sum games, with tracking bounds proven for gradient, Frank-Wolfe, and gradient-free algorithms under time-varying payoffs.","lead":"The paper introduces a forgetting-factor regret metric for online zero-sum games that weights recent saddle gaps exponentially, enabling real-time tracking of time-varying Nash equilibria. It provides regret bounds for three algorithms (gradient descent-ascent, Frank-Wolfe, and zeroth-order) with explicit dependence on equilibrium variation, payoff variation, and gradient-estimation error.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No significant objection identified. The contraction analysis and regret bounds are correct as verified; the interior-NE limitation is real but standard and acknowledged.","rationale":"The reader correctly identifies the interior-NE condition as the main limitation. I verified the core analytical steps: (i) the cross-gradient cancellation identity (81), (ii) the contraction factor derivation leading to (28), (iii) the Frank-Wolfe gap contraction in Theorem 2, and (iv) the zeroth-order error bound in Theorem 3. All check out. The convergence argument (vanishing sequence convolved with summable geometric kernel → 0) is standard and correctly applied. The limitations — strong convexity-strong concavity, interiority, vanishing NE variation, and the restrictive ν > 0 condition for Frank-Wolfe — are all explicitly stated as assumptions or acknowledged as future work. The extension from OCO [42] to games involves genuinely new analysis of cross-gradient coupling (Assumption 4 and the decomposition in (79)–(86)), which is non-trivial. The motivating example (equations 5–7) correctly demonstrates that sublinear standard regret does not imply tracking. The numerical example uses parameters satisfying the stated conditions (α=1 < 2/(1+0.01) ≈ 1.98). No circularity, no internal inconsistency, no unsupported claims detected. The verdict of ACCEPT is appropriate.","tokens_in":24228,"tokens_out":9107,"duration_ms":228982,"concrete_test":"Independently re-derive the contraction factor β(α) in (28) from the squared-norm recursion (95) by expanding [√(κ² + α²L_ײ) + τ]² with κ = 1−α(L+μ)/2 and τ = α(L−μ)/2, and verify that the minimizer α* in (31) satisfies the admissibility condition (29). If the expanded form does not match (28) term-by-term, the sharp contraction claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I carefully traced the three main proofs. For Theorem 1, the key identity (81) — that cross-gradient perturbations satisfy ⟨d_x^t, a_t⟩ + ⟨d_y^t, b_t⟩ = 0 — checks out via the fundamental theorem of calculus applied along the line segments z_x^t(η) and z_y^t(η). The contraction factor β(α) = [√(κ² + α²L_ײ) + τ]² with κ = 1 − α(L+μ)/2 and τ = α(L−μ)/2 correctly expands to (28), and the admissible range (29) correctly ensures β ∈ [0,1). The geometric convolution argument for G-Regret_F_T → 0 when V_t → 0 is standard and valid: both the initial-error term Σ ρ^{T−t}β^t and the variation term Σ_i Σ_{t≥i} ρ^{T−t}β^{t−i}·(variation_i) vanish as T→∞ when the variation sequence is o(1), since both ρ,β ∈ (0,1) produce summable kernels. For Theorem 2, the Frank-Wolfe contraction on the saddle gap (rather than on distance to NE) is correctly derived through (99)–(111), with the step size η_t = νg_t/C minimizing the per-step upper bound. The condition ν > 0 is restrictive (requires NE well-interior and small cross-gradient coupling) but is stated as a hypothesis. For Theorem 3, the zeroth-order gradient error bound ||h_x^t − r_x^t|| ≤ c_t via (119) is correct by the mean value theorem and the choice of δ_t in (42). The reader's identified weakness — the interior NE condition used at (82) to set ∇f_t(x*_t, y*_t) = 0 — is a genuine limitation: without interiority, stationarity becomes a variational inequality and the decomposition (84) fails. However, this is explicitly assumed in all three theorems and acknowledged in Section VI. It is standard in the saddle-point optimization literature and does not constitute an internal inconsistency or correctness error. I do not find a more fundamental concern.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This paper introduces a forgetting-factor regret for online two-player zero-sum games with time-varying convex-concave payoff functions. The metric G-Regret_F_T = Σ ρ^{T-t} (f_t(x_t, y*_t) - f_t(x*_t, y_t)) assigns exponentially decaying weights to past saddle gaps, emphasizing recent performance and directly linking regret minimization to terminal Nash equilibrium (NE) tracking. Three algorithms are analyzed: projected online gradient descent-ascent (OGDA), a projection-free online Frank-Wolfe method, and a deterministic finite-difference zeroth-order method. For each, the authors establish regret bounds that explicitly decompose the effects of NE variation, payoff variation, and gradient-estimation error, and provide sufficient conditions under which G-Regret_F_T → 0, certifying asymptotic tracking of time-varying NEs.","tokens_in":25146,"tokens_out":1340,"duration_ms":107069,"significance":"The paper addresses a genuine gap in the online saddle-point literature: existing regret notions (individual regret, saddle-point regret, NE-regret, duality gap) aggregate historical performance with uniform weights and do not guarantee terminal tracking of a moving NE. The forgetting-factor regret metric is well-motivated and the connection to terminal tracking (Eqs. 12–13, 16) is clean and correct. The technical contributions are substantial: Lemma 2 provides a refined gradient estimate for smooth strongly convex functions (proved from first principles in Appendix B) that enables a sharp contraction factor β(α) for OGDA (Eq. 28), explicitly capturing the interplay between strong convexity, smoothness, and cross-gradient coupling. The Frank-Wolfe analysis (Theorem 2) is notable for deriving contraction on the saddle gap rather than on distance to NE, with an adaptive step size η_t = νg_t/C. The zeroth-order extension (Theorem 3) cleanly separates equilibrium variation from gradient-estimation error. The proofs are detailed and technically careful; the key identity (81) and the contraction derivation (79–95) were verified and are correct.","major_comments":[{"comment":"The interior NE condition (used at Eq. 82 to set ∇f_t(x*_t, y*_t) = 0) is required in all three theorems but is stated only in the theorem hypotheses, not as a standalone assumption. This is a load-bearing condition: without it, the gradient decomposition (84) fails and the contraction analysis breaks down. The authors acknowledge in Section VI that weakening strong convexity-strong concavity is future work, but the interiority limitation itself is not discussed. A brief remark (e.g., after Assumption 4 or in Section II) explaining that interiority is needed for the first-order stationarity conditions, and noting that it excludes games with active constraints at the NE, would improve transparency and help readers assess applicability.","section":null}],"minor_comments":[{"comment":"The role of the forgetting factor ρ deserves more discussion. The regret bounds (Eqs. 34, 40, 46) hold for any ρ ∈ (0,1), but the choice of ρ affects the tightness of the bound and the tracking timescale. A remark on how ρ trades off historical smoothing against tracking responsiveness would help practitioners. Additionally, the simulation uses ρ = 0.5 without justification.","section":null},{"comment":"Section V (Simulations): The numerical example uses a payoff function (Eq. 47) with L_× = 0.1, which is a weakly coupled game. It would strengthen the paper to include a scenario with larger L_× to demonstrate that the algorithms remain effective when cross-gradient coupling is significant, since this is a central feature of the analysis. Also, the Frank-Wolfe algorithm (Algorithm 2) is included in the simulation but the payoff variation F_sup_t does not appear in the example since the payoff is piecewise constant; a case with continuous payoff variation would better illustrate the regime where F_sup_t enters the bound.","section":null},{"comment":"Theorem 2 requires the condition ν > 0 (Eq. 38), which involves the interior distances δ_X, δ_Y and the cross-gradient constants. This condition is somewhat restrictive and its verifiability is not discussed. A remark on when ν > 0 is satisfied in practice, or at least noting that it requires the NE to be well-interior and cross-gradient coupling to be moderate, would help readers assess applicability.","section":null},{"comment":"In Definition 1 (Eq. 15), the regret is defined as a sum from t=1 to T, but the algorithms start from t=0. The indexing convention should be clarified (e.g., whether the t=0 gap is excluded by design or folded into initialization).","section":null},{"comment":"The convergence argument 'a vanishing sequence convolved with a stable geometric kernel remains vanishing' is invoked at the end of each proof (Appendices C–E) without explicit justification. While standard, a one-line lemma or reference would improve rigor and readability.","section":null},{"comment":"Assumption 5 (zeroth-order) requires query access on an expanded domain (X + δ̄_x B) × (Y + δ̄_y B). For sets like probability simplices or structured polytopes, this expansion may not be natural. A brief remark on what sets satisfy this, or whether projection onto the original set suffices, would clarify the assumption's scope.","section":null},{"comment":"The paper does not provide a formal comparison table summarizing the three algorithms' regret bounds, required assumptions, and convergence conditions side by side. Such a table would help readers quickly compare the results.","section":null},{"comment":"Reference [42] is cited as motivation but the key differences between the OCO setting (single-agent, no cross-gradient coupling) and the game setting (two-agent, cross-gradient terms L_xy, L_yx) are not explicitly enumerated. A sentence or two highlighting what is genuinely new beyond porting the forgetting-factor idea to games would strengthen the novelty claim.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is technically sound and the proofs check out. The main limitation (interior NE condition) is real but standard for this type of analysis and is acknowledged as future work. The contribution is incremental relative to the authors' own prior work [42] on OCO, but the extension to games is non-trivial due to cross-gradient coupling and the two-player structure. I recommend minor revision with the interiority discussion as the primary item to address."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper extends the authors' own forgetting-factor regret from online convex optimization to two-player zero-sum games with time-varying payoffs. The metric weights recent saddle gaps more heavily than old ones, so vanishing regret directly certifies terminal NE tracking rather than just good historical average performance. That is a genuine gap in the existing online saddle-point literature, where uniform-weight regrets can be sublinear while the current iterate is far from the current NE. The motivating example in the introduction (where Follow-The-Leader has O(log T) regret but fails to track a switching equilibrium) makes this point cleanly.","headline":"Solid regret metric for time-varying zero-sum games; the main limitation (strong convexity + interior NE) is real but standard and acknowledged.","tokens_in":25363,"tokens_out":188,"would_cite":true,"duration_ms":33879,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Exponential forgetting certifies real-time Nash tracking in zero-sum games","keywords":[],"falsifier":"Construct a time-varying zero-sum game where the Nash equilibrium sits on the boundary of the feasible set and moves slowly, run gradient descent-ascent, and check whether the forgetting-factor regret still converges to zero. If the interior-NE condition is essential, the regret should fail to vanish even with arbitrarily slow equilibrium drift.","tokens_in":24529,"feed_emoji":"⚖️","tokens_out":982,"duration_ms":162878,"temperature":0.7,"pith_summary":"The paper introduces a new performance metric for online two-player zero-sum games where the payoff function changes over time. Standard regret measures aggregate all past performance with equal weight, so an algorithm can look good on average while failing to track the current Nash equilibrium. The authors propose a forgetting-factor game regret that weights each past saddle gap by an exponentially decaying factor, so recent rounds dominate. When this weighted regret converges to zero, the terminal saddle gap must vanish, which means the players' strategies at the final time actually approach the equilibrium of the game being played at that moment. The paper proves this for three algorithms: projected gradient descent-ascent, a projection-free Frank-Wolfe method, and a zeroth-order finite-difference method that uses only function values. For each, the regret bound decomposes into explicit terms involving how fast the Nash equilibrium moves, how fast the payoff function itself changes, and the gradient-estimation error. If the equilibrium drift and (for some algorithms) the payoff drift vanish, the regret goes to zero and tracking is certified.","feed_headline":"Exponential forgetting certifies real-time Nash tracking in zero-sum games","feed_subtitle":"Weighting past saddle gaps by a decay factor turns average-regret bounds into terminal tracking guarantees for three online algorithms.","key_machinery":"The contraction factor beta(alpha) for projected gradient descent-ascent, derived from a refined gradient estimate (Lemma 2) that splits the gradient difference into a strong-convexity component and a smoothness residual, combined with a cross-gradient perturbation term bounded by Assumption 4. The same contraction structure propagates to the zeroth-order case with an additional finite-difference error term. For Frank-Wolfe, the key machinery is a duality-gap contraction (beta_FW) that uses the interiority of the Nash equilibrium to lower-bound the Frank-Wolfe gap and upper-bound the saddle gap, yielding a scalar recursion.","core_discovery":"The central object is the forgetting-factor game regret, defined as the sum over time of exponentially discounted two-sided saddle gaps against the current Nash equilibrium. The key mechanism is that because the forgetting factor lies strictly between zero and one, old errors are geometrically suppressed, so convergence of the weighted sum forces the most recent (undiscounted) saddle gap to zero. This converts an aggregate regret bound into a terminal tracking guarantee. The paper then shows that for gradient descent-ascent, Frank-Wolfe, and zeroth-order finite differences, the per-step tracking error contracts geometrically with a factor that depends on the strong convexity, smoothness, and","pith_inferences":["The forgetting-factor approach could extend to general-sum games or potential games, where the relevant benchmark would be a time-varying coarse correlated equilibrium rather than a Nash equilibrium.","If the forgetting factor is adapted online based on observed payoff variation, the metric could interpolate between average regret (slow environments) and tracking regret (fast environments) without prior knowledge of the drift rate.","The cross-gradient coupling condition (Assumption 4) likely has an information-theoretic interpretation: it bounds how much information about one player's strategy leaks through the other player's gradient, and its absence would make decentralized tracking fundamentally harder."],"forward_implications":["Any online algorithm whose tracking error contracts geometrically per step can be certified under the forgetting-factor regret framework, provided the equilibrium variation is summable or vanishing.","The decomposition of regret into NE variation, payoff variation, and gradient error terms suggests that adaptive step sizes or forgetting factors could be tuned to the observed rate of environmental change.","The zeroth-order result implies that tracking can be maintained with only function-value queries, which is relevant for adversarial or black-box settings where gradients are not available.","The interior-NE requirement for the Frank-Wolfe analysis suggests that constrained games with boundary equilibria would need a different gap-based contraction argument."],"fun_headline_variants":["Discounted regret metric turns aggregate bounds into real-time Nash tracking","Forgetting factor suppresses old saddle gaps to certify terminal NE tracking","Exponential decay on past gaps yields per-step Nash equilibrium convergence","Weighted saddle-gap regret links three online algorithms to time-varying NE tracking","Geometric suppression of past errors forces current saddle gap to zero"],"cache_read_input_tokens":0,"weakest_assumption_plain":"Every payoff function must be strongly convex in one player's variable and strongly concave in the other's, and the Nash equilibrium must lie in the interior of the feasible set so that the gradient at the equilibrium is exactly zero. Without interiority, the stationarity conditions become inequalities and the contraction analysis does not go through.","fun_headline_variants_meta":{"raw":{"variants":["Discounted regret metric turns aggregate bounds into real-time Nash tracking","Forgetting factor suppresses old saddle gaps to certify terminal NE tracking","Exponential decay on past gaps yields per-step Nash equilibrium convergence","Weighted saddle-gap regret links three online algorithms to time-varying NE tracking","Geometric suppression of past errors forces current saddle gap to zero"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":675,"prompt_tokens":585,"completion_tokens":90,"prompt_tokens_details":null},"tokens_in":585,"tokens_out":90,"duration_ms":24210,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T20:27:05.557278+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Construct a time-varying zero-sum game where the Nash equilibrium sits on the boundary of the feasible set and moves slowly, run gradient descent-ascent, and check whether the forgetting-factor regret still converges to zero. If the interior-NE condition is essential, the regret should fail to vanish even with arbitrarily slow equilibrium drift.","supporting_citations":[],"review_version":1}