acceptodds
Under review as a conference paper at ICLR 2027

ZOFOSGDA: A Hybrid Method for Nonsmooth Nonconvex-Nonconcave Minimax Optimization

Abstract

Minimax optimization arises in many machine-learning problems, including robust learning, adversarial training, and AUC maximization. Some of these formulations are nonsmooth and may exhibit nonconvex–nonconcave geometry.However, most existing optimization methods are designed primarily for smooth objectives, and their convergence guarantees typically apply only to weakly convex–(strongly) concave or nonconvex–(strongly) concave settings. To address this gap, we study stochastic noncomposite minimax optimization with a nonsmooth, nonconvex outer variable and a smooth, possibly nonconcave inner variable that satisfies a uniform projected Polyak-ojasiewicz condition. We propose ZOFOSGDA, a hybrid method that combines randomized two-point zeroth-order estimates for the outer variable, stochastic first-order gradients for the inner variable, and SPIDER variance reduction. Our stationarity criterion pairs a projected Goldstein residual in the outer block with a projected-gradient residual in the inner block. Under a uniformly bounded initial potential gap and the canonical smoothing radius , ZOFOSGDA returns both an expected blockwise generalized Goldstein stationary point and an expected projected Goldstein stationary point of the value function. Its oracle costs are stochastic function queries and stochastic inner-gradient queries, where is the outer dimension. An outer-gradient tracking bound enables the value-function guarantee. A controlled classification benchmark verifies the nonconcave projected-P geometry, while a supplementary AUC-maximization task provides an additional finite-sample stress test.

Primary area:
optimization
Keywords:
minimax optimization, zeroth-order optimization, nonsmooth optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ZLHQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies stochastic nonsmooth nonconvex–nonconcave minimax optimization under a uniform projected-PL condition on the inner problem. The authors propose ZOFOSGDA, combining zeroth-order randomized smoothing, first-order gradient ascent, and SPIDER variance reduction. It achieves expected blockwise Goldstein stationarity and projected Goldstein stationarity of the value function, with oracle complexities of (~d11.5ϵ−4)\widetilde(d_1^1.5\epsilon^-4) for stochastic function evaluations and (~d10.5ϵ−4)\widetilde(d_1^0.5\epsilon^-4) for stochastic inner-gradient evaluations. Numerical experiments on classification and AUC maximization are provided. I have not identified a clear mathematical error in the main arguments examined.

My main concern is the limited novelty. W. Shi et al. (2026) have already studied a closely related hybrid zeroth- and first-order method for nonsmooth minimax optimization. The present work uses SPIDER instead of momentum, considers projected-PL inner geometry, and provides an additional Goldstein stationarity guarantee for the original value function. While these are meaningful technical extensions, the main ingredients build on established techniques (e.g., Chen et al., 2023; Xu et al., 2023). Overall, I find the conceptual contribution largely incremental.

Shi, Wanli, et al. "Hybrid Zeroth-and First-Order Method for Stochastic Nonsmooth Minimax Optimization." (2026). Chen, Lesi, Jing Xu, and Luo Luo. "Faster gradient-free algorithms for nonsmooth nonconvex stochastic optimization." International Conference on Machine Learning. PMLR, 2023. Xu, Zi, et al. "Zeroth-order alternating gradient descent ascent algorithms for a class of nonconvex-nonconcave minimax problems." Journal of Machine Learning Research 24.313 (2023): 1-25.

Critical strengths

The main strength is the extension of convergence guarantees to nonsmooth minimax problems with a potentially nonconcave inner objective satisfying a projected-PL condition. In particular, the analysis also establishes Goldstein stationarity of the original value function.

Critical weaknesses

My primary concern is the limited novelty. W. Shi et al. (2026) have already proposed a closely related hybrid zeroth- and first-order method for nonsmooth minimax optimization with a similar blockwise Goldstein stationarity guarantee. The use of SPIDER variance reduction and PL-based analysis also builds on established techniques (Chen et al., 2023; Xu et al., 2023). Although the extensions require some technical effort, I find the conceptual and algorithmic contributions largely incremental.

The experimental validation is also not fully convincing. The paper lacks a direct comparison with the closely related hybrid method of W. Shi et al. (2026), making it difficult to assess the practical benefit of SPIDER relative to this existing alternative.

Questions for authors

Assumption 3.2 can be viewed as a projected-gradient analogue of the KL inequality with exponent 1/21/2. Is this exponent essential to the analysis, or could the convergence guarantees be extended to a more general projected-gradient inequality with exponent (1/2,1)(1/2,1), possibly with different convergence rates? Clarifying this would help assess the generality of the theoretical results.

10 comments
AuthorsOpenReview ↗

Reponse (Part 1)

Thank you for your careful review and for raising these concerns about the relationship to prior work, the technical contribution, and the generality of our results. Below, we respond to each point in turn, explain the supporting analysis, and describe the corresponding revisions to the manuscript.


W1: Limited novelty relative to Shi et al. (2026) and lack of a direct comparison\color{blue}{\text{W1: Limited novelty relative to Shi et al. (2026) and lack of a direct comparison}}


Thank you for pointing us to the related hybrid method of Shi et al. We have carefully examined the cited manuscript. It is inappropriate to treat an unpublished manuscript containing demonstrable theoretical errors as an established benchmark for judging our contribution to be incremental. We also disagree with using the missing comparison as a basis for rejection. Our response has three parts.

  1. Policy. Using the missing preprint comparison as a rejection ground violates ICLR's reviewer guidelines.

  2. Errors in the cited analysis. Two counterexamples expose theoretical errors that undermine its use as an established baseline for assessing our contribution.

  3. Our contribution. We establish a finite-query theory for stationarity of the original nonsmooth minimax value function under nonconcave projected-PL inner geometry, a guarantee that Shi et al.'s analysis does not provide.

We explain each point in detail below.

1. The policy violation and the correction we request.

(a) ICLR explicitly excludes this missing comparison as a rejection ground.

The [ICLR 2027 Reviewer Guidelines], under the question about very recent and non-peer-reviewed work, state

While authors are not required to compare to contemporaneous work or unpublished arxiv papers, they are strongly encouraged to cite such related work if they are aware of it. Reviewers can make authors aware of related contemporaneous work or arxiv papers, but the lack of such comparisons cannot be a basis for rejection.

The guideline makes three distinct points. Authors are not required to compare with the specified contemporaneous or unpublished work. Authors who know of relevant preprints are strongly encouraged to cite them. Reviewers may suggest such work, but “the lack of such comparisons cannot be a basis for rejection.” The final clause explicitly restricts which grounds may support a rejection recommendation. A suggestion to add a comparison must therefore remain consistent with that restriction.

(b) The cited manuscript is unpublished, so the above policy applies.

The [manuscript by Shi et al. you cite] is unpublished and therefore falls within the scope of the above policy. The absence of a direct experimental comparison with this manuscript cannot be a basis for rejection. Following the guideline's encouragement to cite relevant preprints, we have added the citation in Related Work and a dedicated discussion in Appendix I of the revised manuscript, explaining the differences in assumptions, analysis, and guarantees.

(c) Your review uses the missing comparison in the way the guideline prohibits.

In your review, under “Critical Weaknesses,” you write

The experimental validation is also not fully convincing. The paper lacks a direct comparison with the closely related hybrid method of W. Shi et al. (2026), making it difficult to assess the practical benefit of SPIDER relative to this existing alternative.

Your review explicitly uses the missing comparison to conclude that our experimental validation is “not fully convincing” and lists this under “Critical Weaknesses.” This makes the absence of the preprint comparison part of the justification for rejection, contrary to the quoted guideline. We therefore ask you to exclude this point as a basis for rejection.

We next examine whether the cited guarantees provide a sound theoretical baseline for your novelty assessment, then explain the additional guarantees established by our work.

2. Errors in the cited paper's theoretical analysis.

Your review cites “a similar blockwise Goldstein stationarity guarantee” in support of the assessment that our contribution is incremental. Using the cited manuscript as an established theoretical baseline is unjustified because its analysis contains two explicit errors in the derivation of those guarantees. We give counterexamples to the maximizer identity and estimator unbiasedness, then show how these errors affect the supplied proofs.

(a) A false maximizer identity is used repeatedly in the theoretical analysis.

Lemma B.1 of the cited version (p. 16, lines 848–850) identifies the maximizer of the averaged objective with the maximizer at an individual perturbed point,

=\arg\max_{y\in\mathcal Y}\mathbb E_u f(x+\delta u,y).$$ To see why this identity fails under their assumptions, consider $$f(x,y)=-y^2+2y\sin x,\qquad x\in\mathbb R,\quad y\in[-2,2],\quad u\sim\operatorname{Unif}[-1,1].$$

Averaging over uu gives fδ(x,y)=−y2+2y(sin⁡δ/δ)sin⁡xf_\delta(x,y)=-y^2+2y(\sin\delta/\delta)\sin x. Therefore, for 0<δ<10<\delta<1,

y_\delta^ *(x)=\frac{\sin\delta}{\delta}\sin x.$$ For example, at $x=0$, the smoothed maximizer is $y_ \delta^ *(0)=0$, whereas $y^ *(\delta u)=\sin(\delta u)$ is nonzero for almost every sampled $u$. **The asserted identity therefore fails even under strong concavity.** The distinction is consequential because smoothing and maximization generally do not commute. The example satisfies the cited assumptions. On the stated domain, valid constants are $L=2\sqrt{13}$, $L_{fy}=2\sqrt2$, $L_{fxy}=2$, and strong-concavity parameter $\tau=2$. A deterministic oracle has zero variance, and $\max_y f_\delta(x,y)=(\sin\delta/\delta)^2\sin^2x\ge0$, so the oracle and lower-boundedness assumptions also hold. **This false identity is used repeatedly across the paper's theoretical analysis, and its consequences propagate to the main convergence and complexity proofs.** It appears directly in the smoothness argument, Eq. (43) (p. 17), and the outer-gradient tracking argument, Eq. (51) (p. 18). The resulting estimates enter Eq. (86), the Lyapunov recursion Eq. (89), and the final convergence bound Eq. (100) in the proof of Theorem 5.10. Remark 5.11 derives the query complexity from that bound. Thus, the theoretical guarantees used in your comparison depend on this invalid identification of two different maximizing responses. **(b) Shi et al.'s specified estimator is biased, contrary to their Lemma 5.4.** Shi et al. specify a two-point estimator with coefficient $d_1/(2\delta)$ and directions uniform inside the unit ball (p. 5, lines 270–274 and the sampling description on that page). Their Lemma 5.4 claims that this estimator is unbiased for $\nabla_x f_\delta$, with the claimed identity derived in their Eq. (34). **That unbiasedness claim is false for the stated sampling distribution.** To check this claim under the specified distribution, consider $$f(x,y)=\psi(a^\top x)-y^2,\qquad \psi(t)=\max\\{-1,\min\\{t,1\\}\\},\quad a\ne0,\quad y\in[-1,1].$$ This objective is globally Lipschitz on the stated domain and $2$-strongly concave in $y$, with inner gradient $-2y$ and an outer directional derivative independent of $y$. Its deterministic oracle has zero variance, and $\max_y f_\delta(x,y)\ge-1$. At $x=0$, choose $\delta\|a\|<1$. All sampled perturbations then lie in the linear region of $\psi$, and $\nabla_x f_\delta(0,y)=a$. Shi et al.'s stated estimator therefore gives $$\mathbb E\widehat g_x(0,y) =d_1\mathbb E[uu^\top]a =\frac{d_1}{d_1+2}a \ne a=\nabla_x f_\delta(0,y).$$ **The factor $d_1/(d_1+2)$ proves that the specified estimator is biased.** Averaging multiple samples does not remove this bias. The proof additionally assumes $\|u\|=1$ (p. 16), which contradicts the stated ball-sampling distribution. Sphere sampling would give the standard unbiased estimator for ball smoothing, but would change the distribution specified for the method. The stated estimator and its unbiasedness proof are therefore inconsistent. These counterexamples establish errors in two key ingredients of the cited theory, including an identity used repeatedly in the convergence analysis and an unbiasedness claim that fails for the specified estimator. **It is therefore unjustified to use a manuscript with these theoretical errors as an established baseline for concluding that our contribution is incremental.** That comparison credits the cited work with guarantees its analysis does not establish. **3. A finite-query theory for the original nonsmooth minimax objective.** **Our contribution is a framework that provably attains projected Goldstein stationarity of the original nonsmooth value function $\Phi(x)=\max_y f(x,y)$ under nonconcave projected-PL inner geometry, with explicit SZO and SFO query guarantees.** The method obtains this certificate using stochastic function values and inner gradients at approximate inner responses, allowing the maximizing response to be nonunique. The comparison with Shi et al. explains the problem it covers, the objective it certifies, and the analysis needed to attain it.

(a) A theory for nonconcave inner optimization with nonunique maximizing responses. Our framework controls inner-response error under the uniform projected-PL condition of the original problem. Lemma 4.5 bounds distance to the full solution set Y∗(x)Y^*(x), and Lemma 4.6 establishes averaged-gap control with an explicit smoothing correction. These estimates make it possible to analyze nonsmooth minimax optimization with nonconcave inner objectives and nonunique maximizers without assuming that smoothing preserves PL. Shi et al. impose strong concavity and build their analysis around a unique maximizing response. That analysis does not cover the inner geometry addressed by our theory.

(b) A stationarity certificate for the original worst-case objective. We prove projected Goldstein stationarity of Φ(x)\Phi(x), certifying the outer decision against its maximizing response while accounting for the error in the auxiliary inner iterate. Shi et al. claim blockwise generalized Goldstein stationarity at a pair (x,y)(x,y), and their Eq. (16) analyzes max⁡yfδ(x,y)\max_y f_\delta(x,y). Our value-function certificate requires control of Φδ(x)=Eumax⁡yf(x+δu,y)\Phi_\delta(x)=\mathbb E_u\max_y f(x+\delta u,y). Smoothing and maximization generally do not commute, so the two smoothed objectives differ even when the inner problems are solved exactly. Lemmas D.4 and D.6 control this discrepancy through gradient tracking and establish the transfer to the original Φ\Phi. Our theory therefore certifies the original minimax objective itself, a result that the cited blockwise guarantee does not establish. The framework also proves the blockwise stationarity guarantee.

(c) A stochastic algorithm proved to attain this certificate at a quantified query cost. Our coupled SPIDER analysis controls both stochastic estimator errors together with the inner-response tracking error as both variables move. This joint control establishes that the available hybrid queries suffice to attain the original value-function certificate. Corollaries 5.7–5.8 prove both expected stationarity guarantees, including E[rΦ(xR)]≤ϵ\mathbb E[\mathfrak r_\Phi(x_R)]\le\epsilon. Under the stated assumptions, including the uniform initial-potential-gap bound Δpot\Delta_{\mathrm{pot}}, the query costs are

Q_{\mathrm{SZO}}=\widetilde O\left(\Delta_{\mathrm{pot}}d_1^{3/2}\delta^{-1}\epsilon^{-3}\right),\\ Q_{\mathrm{SFO}} =\widetilde O\left(\Delta_{\mathrm{pot}}d_1^{1/2}\delta^{-1}\epsilon^{-3}\right). \end{aligned}$$ Both costs include the work needed to obtain sufficiently accurate inner responses. For $\delta=c_\delta\epsilon$ with sufficiently small fixed $c_\delta$, both bounds have accuracy dependence $\epsilon^{-4}$. In comparison, Shi et al.'s momentum method reports $\widetilde O(d_1^{3/2}\delta^{-1}\epsilon^{-4})$ SZO queries in Remark 5.11, giving $\epsilon^{-5}$ after the same radius scaling, under its stated assumptions. As shown in point 2, that reported bound is not established by the supplied proof. **Our algorithm has a proved finite-query guarantee for the original value-function certificate under nonconcave inner geometry, together with a lower SZO accuracy exponent than the one claimed in Shi et al.** The averaged-gap, gradient-tracking, and coupled stochastic estimates are essential to establishing this guarantee. **The contribution is a proved finite-query theory for the original nonsmooth minimax objective under nonconcave inner geometry.** Shi et al.'s claimed blockwise result under strong concavity does not establish this guarantee. It is unjustified to dismiss our contribution by treating an unpublished manuscript with the demonstrated mathematical errors as an established theoretical benchmark. The cited work cannot be credited with guarantees that its analysis does not establish, and ICLR's guidelines explicitly prohibit using the missing preprint comparison as a basis for rejection. We therefore ask you to reassess this concern based on the guarantees actually established by the respective analyses. The revised Appendix I provides the detailed comparison.
Reviewer_5uVjOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies stochastic noncomposite minimax optimization with a nonsmooth, nonconvex outer variable and a smooth, possibly nonconcave inner variable that satisfies a uniform projected Polyak-Łojasiewicz condition. We propose ZOFOSGDA, a hybrid method that combines randomized two-point zeroth-order estimates for the outer variable, stochastic first-order gradients for the inner variable, and SPIDER variance reduction, and establishes projected Goldstein stationarity guarantees at both the block level and the value-function level. The theoretical results show that the proposed algorithm have a better dimension dependence than existing nonsmooth minimax methods with O(d13/2ε−4)O(d^{3/2}_1 \varepsilon^-4) and O(d11/2ε−4)O(d^{1/2}_1 \varepsilon^-4), respectively.

Critical strengths
  1. The authors develop ZOFOSGDA for nonsmooth nonconvex minimax problems with a smooth, potentially nonconcave inner objective satisfying a uniform projected-PŁ condition.
  2. This paper is written clearly. The authors propose ZOFOSGDA, ahybrid method that combines randomized two-point zeroth-order estimates for the outer variable, stochastic first-order gradients for the inner variable, and SPIDER variance reduction, which bringing two heterogeneous query types into a single convergence framework.
  3. The theoretical contribution is solid: by establishing an outer gradient tracking bound, it lifts the block-wise residual to projected Goldstein stationarity of the value function, and provides guarantees on both sides along with an explicit radius and a high-probability confidence-dependent version.
  4. The theoretical results show that the proposed algorithm have a better dimension dependence than existing nonsmooth minimax methods with O(d13/2ε−4)O(d^{3/2}_1 \varepsilon^-4) and O(d11/2ε−4)O(d^{1/2}_1 \varepsilon^-4), respectively
Critical weaknesses
  1. In this paper, it would be good if the author could provide practical examples that meet the projected PL assumption, or if he could extend the analysis to weaker conditions and add appropriate discussion in the paper.
  2. In the experimental part, the problem of AUC maximization of binary classification is solved. If some practical examples of large-scale languages can be added, it will be more practical
  3. It seems that under the query complexity of ε−4\varepsilon^ {-4}, please add some discussion in the paper to show that the proposed zero order algorithm theory can effectively alleviate the dimension cost or give actual performance.
  4. Numerical comparison against comparable nonsmooth minimax baselines is insufficient, so the theoretical advantage in dimension dependence is not directly reflected in the experiments.
Questions for authors
  1. In this paper, it would be good if the author could provide practical examples that meet the projected PL assumption, or if he could extend the analysis to weaker conditions and add appropriate discussion in the paper.
  2. Whether the author can add a baseline comparison of nonsmooth minimax, it is better to test a large model task, such as confrontation training, to illustrate the advantages of this dimension.
  3. It seems that under the query complexity of ε−4\varepsilon^ {-4}, please add some discussion in the paper to show that the proposed zero order algorithm theory can effectively alleviate the dimension cost or give actual performance.
8 comments
AuthorsOpenReview ↗

Respones (Part 1)

Thank you for your careful review and for recognizing our algorithm and theoretical guarantees. Following your suggestions, we have added a projected-KL extension in Appendix H, robust-control experiments in Appendix J, and OPT-1.3B adaptation on eight language tasks in Appendix K. We address your questions about the assumptions, applications, and computational cost below.


W1 / Q1(a): Applications and Verification of Projected PL.\color{blue}{\text{W1 / Q1(a): Applications and Verification of Projected PL.}}


Thank you for raising this question. We verify uniform projected PL for a physical-design model over an entire design region, and we extend our analysis to every fixed projected-KL exponent θ∈(1/2,1)\theta\in(1/2,1) in Appendix H. The application examples and calculations below show where the condition holds and how to check it.

1. Why we require projected PL.

For an inner projected-gradient mapping Gy,λG_{y,\lambda}, our condition requires

Φ(x)−f(x,y)≤12μ∥Gy,λ(x,y)∥2\Phi(x)-f(x,y)\le\frac{1}{2\mu}\|G_{y,\lambda}(x,y)\|^2

with the same μ>0\mu>0 and admissible interval 0<λ≤λˉ0<\lambda\le\bar\lambda throughout Xδˉ×YX_{\bar\delta}\times Y. A small residual in a general smooth nonconcave problem can occur far below the global maximum. This inequality turns a computable inner residual into a bound on the global maximization gap. Its uniformity keeps that bound valid as the outer decision and its smoothing perturbations change.

Our algorithm evaluates its outer update at an approximate inner response, while our target is Φ(x)=max⁡yf(x,y)\Phi(x)=\max_y f(x,y). We use the gap bound in our tracking and coupled stochastic estimates to control the resulting outer-direction error. This lets us prove stationarity of the original value function while accounting for the finite queries used by the approximate inner updates.

2. Application models with nonconcave PL geometry.

We present two examples from real applications—robust feedback control and mechanical vibration design—whose inner objectives are nonconcave yet satisfy a PL inequality under explicit model conditions. For the vibration model, we further verify the uniform projected-PL condition required by our analysis.

(a) Robust feedback control.

Zhang et al. [3] study finite-horizon robust control with dynamics and feedback policies

\qquad u_t=-K_ts_t,\quad w_t=-L_ts_t.$$ The outer variable $K=(K_t)$ designs the controller, and the inner variable $L=(L_t)$ selects the disturbance-feedback policy. The minimax trajectory cost is $$\min_{K\in\mathcal K}\max_L J(K,L), \qquad J(K,L)=\mathbb E\left[ \sum_{t=0}^{T-1} \left(s_t^\top Q_ts_t+u_t^\top R_t^u u_t-w_t^\top R_t^w w_t\right) +s_T^\top Q_Ts_T\right].$$ Here the state weights are positive semidefinite, the control and disturbance weights are positive definite, and $\mathcal K$ is the robust-feasible controller set in [3, Lemma 3.3]. The disturbance gains enter products of the closed-loop matrices $A_t-B_tK_t-D_tL_t$ along the state trajectory, so quadratic stage costs do not make $J(K,\cdot)$ concave in the policy parameters. Under the robust-feasibility and nondegenerate-noise conditions in [3], Lemma 3.5 establishes, for every fixed admissible $K$, $$\|\nabla_LJ(K,L)\|_F^2 \ge c_K\bigl[J(K,L^*(K))-J(K,L)\bigr], \qquad c_K>0.$$ **This PL inequality lets us control the global quality of a disturbance policy while retaining its nonconcave feedback-policy formulation.** We add the outer sparsity cost $\tau\sum_t\|K_t\|_1$, motivated by sparse feedback design [4]. This gives a nonsmooth controller-design objective while preserving the inner maximizers, gradients, gaps, and PL inequality. We evaluate this extension in Appendix J. **(b) Mechanical vibration design.** Consider a damped oscillator, $m\ddot z+c\dot z+kz=F\cos(\omega t)$, with positive mass $m$, stiffness $k$, damping $c$, and forcing amplitude $F$. Its squared steady-state displacement amplitude at excitation frequency $\omega$ [1] is $$H(m,k,\omega)=\frac{F^2}{(k-m\omega^2)^2+c^2\omega^2}.$$ Minimizing peak response over a frequency band is an established design objective [2]. With absolute-value adjustment costs, the problem is $$\min_ {(m,k)\in X}\max_ {\omega\in[a,b]} \\{H(m,k,\omega)+\tau_m|m-m_0|+\tau_k|k-k_0|\\}.$$ The outer variables select mass and stiffness, while the inner variable selects the most adverse frequency in $[a,b]$, with $0<a<b$. Define $z_*=k/m-c^2/(2m^2)$. On a compact convex design set with positive mass and stiffness throughout $X_{\bar\delta}$, the condition $3a^2<z_*<b^2$ places the unique positive-frequency maximum $\omega_*=\sqrt{z_*}$ inside the band and gives $H_{\omega\omega}(m,k,a)>0$. By continuity, positive curvature also occurs inside the band. **The inner response is nonconcave for every design in the region.**

Writing D=(k−mω2)2+c2ω2D=(k-m\omega^2)^2+c^2\omega^2 and d∗=c2k/m−c4/(4m2)>0d_*=c^2k/m-c^4/(4m^2)>0, direct calculation gives

=\frac{8F^2m^2\omega^2d_*}{D^3} \left[H(m,k,\omega_*)-H(m,k,\omega)\right].$$ The coefficient has a positive minimum on the compact domain. We can therefore choose $$\mu=\min_{(m,k,\omega)\in X_{\bar\delta}\times[a,b]} \frac{8F^2m^2\omega^2d_*}{D^3}>0.$$ To verify projection, let $L_\omega=\max|H_{\omega\omega}|$ on this domain. The gradient points toward $\omega_*$ and satisfies $|H_\omega(\omega)|\le L_\omega|\omega-\omega_*|$. For **every $0<\lambda\le1/L_\omega$**, the ascent point stays between the current frequency and $\omega_*$. Projection is inactive, so the same $\mu$ applies to the projected-gradient mapping throughout the common stepsize interval. The adjustment costs cancel from the inner gap. For example, $F=c=1$, $X=[0.9,1.1]\times[1.9,2.1]$, $[a,b]=[0.5,2]$, and $\bar\delta=0.05$ give $1.2306<z_*<1.8374$ throughout an envelope containing the augmented domain. Hence $3a^2=0.75<z_*<4=b^2$ uniformly. The outer value is $V(m,k)=F^2m^2/[c^2(km-c^2/4)]$, whose Hessian has negative determinant when $km>5c^2/4$, as holds throughout this region. Positive adjustment costs preserve nonconvex curvature away from their kinks and introduce nonsmoothness when $(m_0,k_0)$ lies in the interior. **We have therefore verified the required uniform projected-PL condition for a nonsmooth nonconvex outer problem with a smooth nonconcave inner objective throughout the augmented design region.** The strict parameter margins make our certificate persist under sufficiently small perturbations of the design bounds and physical parameters. **3. Checking uniformity and projection.** The calculation above gives four checks for a structured model: identify the global inner maximizer and gap; derive the residual-to-gap coefficient; bound it below over the augmented outer domain; and verify a common projected stepsize interval. In the oscillator, the response formula and strict parameter bounds complete all four checks. Domain-wide bounds establish uniformity, while empirical residuals along a training trajectory can supplement this verification. The remaining regularity and stochastic-oracle conditions are separate assumptions of our theorem. **4. Extension to projected KL.** In Appendix H, we replace projected PL with the more general residual-to-gap condition $$\\|G_{y,\lambda}(x,y)\\|\ge c\,[\Phi(x)-f(x,y)]^\theta, \qquad\theta\in(1/2,1),$$ for every fixed $\theta$, uniformly over the augmented domain and a common stepsize interval. We retain the remaining assumptions. We first bound distance to the inner solution set by the gap raised to $1-\theta$ (Lemma H.2), then derive a Hölder tracking bound with exponent $p=(1-\theta)/\theta$ on the squared residual and smoothing error (Lemma H.3). Using accuracy-dependent weights, stepsizes, and batch sizes, we close the coupled stochastic recursion in Theorem H.4. **We obtain both expected blockwise generalized Goldstein stationarity and expected projected Goldstein stationarity of the original value function under this broader geometry.** Under the stated uniform initial-potential-gap condition, choosing $\delta=c_\delta\epsilon^{\theta/(1-\theta)}$ with sufficiently small $c_\delta>0$ yields $$\begin{aligned} Q_{\mathrm{SZO}} &=\widetilde O\!\left(\Delta_{\mathrm{pot}}d_1^{3/2} \epsilon^{-4\theta/(1-\theta)}\right),\\ Q_{\mathrm{SFO}} &=\widetilde O\!\left(\Delta_{\mathrm{pot}}d_1^{1/2} \epsilon^{-4\theta/(1-\theta)}\right). \end{aligned}$$ We state these bounds in Corollary H.5 for fixed $\theta$, with constants that may depend on $\theta$ and the problem parameters. The accuracy exponent approaches $4$ as $\theta$ approaches $1/2$ from above. **We retain both stationarity targets and make the additional accuracy cost of the weaker residual-to-gap exponent explicit.** The new analysis is marked in blue. --- ## $\color{blue}{\text{W2 / Q1(b): Large-Scale and Application Experiments}}$ --- Thank you for this suggestion. **We have expanded the evaluation with OPT-1.3B adaptation on eight language tasks in Appendix K and nonsmooth robust-control experiments in Appendix J, both marked in blue.** The language tasks test performance on a large pretrained model; the control tasks additionally allow exact evaluation against the global inner optimum. **1. OPT-1.3B adaptation.** We use OPT-1.3B on BoolQ, CB, COPA, RTE, SST-2, WiC, WinoGrande, and WSC. The objective follows the minimax meta-training formulation of Shirkavand et al. [5], $$\min_{x,\hat y}\max_y\; F(x,\hat y;\mathcal D_f) +\theta\bigl[F(x,\hat y;\mathcal D_p)-F(x,y;\mathcal D_p)\bigr], \qquad\theta>0.$$

Here xx denotes the shared backbone parameters, y^\hat y and yy are task-specific adaptation blocks, and Dp\mathcal D_p and Df\mathcal D_f are support and query data. The outer variables minimize query loss and the support-loss penalty. Maximizing over yy minimizes the reference adapter's support loss, so the penalty encourages y^\hat y to attain a similarly low support loss. The outer optimization therefore depends on the quality of the inner adaptation.

We parameterize both adaptation blocks with LoRA and evaluate single-task adaptation using separate training and test splits. We use up to 500 validation examples and 1,000 test examples per task, subject to available labeled data. We compare against first-order LoRA, full-model and LoRA-based MeZO [6], Bilevel-ZOFO [5], PGFDA, and ZOFO. We run our method for 2,000 steps and report mean test accuracy over three random seeds. We specify the step-size search in Appendix K.

2. Accuracy across eight language tasks.

In Table 12, we report the following unweighted averages of the mean accuracies across the eight tasks.

MethodAdaptation modeAverage test accuracy
First-orderLoRA73.45%
MeZOFull model65.72%
MeZOLoRA67.06%
Bilevel-ZOFOLoRA74.39%
PGFDALoRA73.49%
ZOFOLoRA73.32%
ZOFOSGDALoRA74.69%

Our method achieves 74.69% average accuracy, improving on Bilevel-ZOFO by 0.30 percentage points and ZOFO by 1.37 percentage points. We also improve on PGFDA and first-order LoRA by 1.20 and 1.24 percentage points, respectively. We attain the highest mean accuracy on COPA and RTE and tie with PGFDA on WinoGrande; we give the full taskwise comparison in Table 12. Our method attains these accuracies on real language tasks using a 1.3-billion-parameter backbone.

3. Worst-case control performance.

We adapt the finite-horizon robust-control model of Zhang et al. [3], where controller gains KK and disturbance-feedback gains LL interact through a state trajectory. The tested objectives include both τ=0\tau=0 and the nonsmooth design cost

\qquad\tau=0.01.$$ We compare our method with TTSGDA, PGFDA, and ZOFO on a time-varying system with $T=10$ and fixed systems with $T=10$ and $T=20$, giving 90 or 180 parameters per policy block. We run each configuration for 5,000 main-loop iterations with four seeds and two penalty values, giving 96 runs. We select parameters from the four-seed confirmation runs on the time-varying instance and reuse them on both fixed systems without retuning. The methods have different oracle budgets, including our refresh batches. We specify these budgets, parameter selection, and the unconstrained policy implementation in Appendix J. For each controller, we compute its exact worst-case cost by backward Riccati recursion and report $$\Phi_\tau(K)=\sup_LJ_\tau(K,L),\qquad \Delta_{\mathrm{in}}(K,L)=\Phi_\tau(K)-J_\tau(K,L).$$ **We use these metrics to evaluate the controller against its globally optimal disturbance and to check the quality of the disturbance learned by our algorithm.** An inaccurate learned disturbance therefore cannot artificially improve the controller cost. For $\tau=0.01$, we obtain the following final worst-case costs in Table 11, expressed as means $\pm$ sample standard deviations over four seeds. | Method | Time-varying, $T=10$ | Fixed, $T=10$ | Fixed, $T=20$ | |---|---|---|---| | ZOFOSGDA | $7.0057\pm0.0277$ | $7.0202\pm0.0186$ | $13.9531\pm0.0380$ | | TTSGDA | $6.9526\pm0.1409$ | $7.0569\pm0.1971$ | $13.9897\pm0.3772$ | | PGFDA | $7.9620\pm0.0261$ | $7.9586\pm0.0303$ | $15.6696\pm0.0558$ | | ZOFO | $8.9769\pm0.1135$ | $8.9637\pm0.1372$ | $18.9890\pm0.2586$ | **We obtain lower mean worst-case costs than PGFDA and ZOFO in all three systems for both penalty values.** On the nonsmooth instances, our mean cost is approximately $0.76\\%$ higher than TTSGDA on the time-varying system and slightly lower on both fixed systems. TTSGDA uses first-order information in both blocks, while our outer updates use function values. Our mean inner gaps are $0.00160$, $0.00149$, and $0.00460$, compared with $0.05306$, $0.05984$, and $0.14006$ for TTSGDA, approximately **30 to 40 times smaller**. Our final controller costs also have smaller sample standard deviations than TTSGDA. We plot both trajectories in Figures 7–9; all 96 runs remain robust-feasible at the recorded checkpoints. **The small inner gaps show that our learned disturbances approach the global worst-case values used to evaluate the controllers.** We retain lower controller costs than PGFDA and ZOFO on the fixed systems without retuning the selected parameters.
Reviewer_PgBWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies stochastic nonsmooth nonconvex–nonconcave minimax optimization with a smooth inner objective satisfying a uniform projected-PL condition. The authors propose ZOFOSGDA, combining randomized smoothing, hybrid zeroth- and first-order updates, and SPIDER variance reduction. They establish convergence guarantees for blockwise generalized Goldstein stationarity and projected Goldstein stationarity of the value function.

Critical strengths

The paper establishes a value-function Goldstein stationarity guarantee for nonsmooth nonconvex–nonconcave minimax optimization under a projected-PL condition.

The numerical experiments demonstrate promising performance compared with several existing methods.

Critical weaknesses

Insufficient motivation and empirical justification: Given that Shi and Luo (2026) already study nonsmooth nonconvex–concave minimax optimization, the paper does not sufficiently explain the significance of allowing nonconcave inner objectives while additionally requiring inner smoothness and a uniform projected-PL condition. The main example in Section 6.1 is specifically constructed to satisfy this condition, while the supplementary AUC problem has a strongly concave inner objective. Thus, the significance of this extension is not sufficiently demonstrated.

Incomplete comparison with existing work: Table 1 does not clearly distinguish the smoothness assumptions on the inner objective. As a result, the comparison does not fully reflect the trade-off between relaxing inner concavity and imposing additional smoothness assumptions. Clarifying this distinction would help readers better assess the scope and significance of the theoretical contributions.

Questions for authors

.

9 comments
AuthorsOpenReview ↗

Response (Part 1)

Thank you for your review and for recognizing our value-function stationarity guarantee and numerical results. Your main concerns are the significance of allowing nonconcave inner objectives and the assumptions needed to obtain that guarantee. We address these below with two complete application examples, new robust-control experiments in Appendix J, an explicit comparison of assumptions.


W1. motivation and empirical justification.\color{blue}{\text{W1. motivation and empirical justification.}}


Thank you for raising this question. We explain below why nonconcave inner problems matter, which application models exhibit PL geometry, what the new experiments show, and how our assumptions enable the value-function guarantee

1. Why nonconcave inner problems matter in applications.

Nonconcave inner objectives arise naturally in many minimax applications. The underlying model determines how the loss or cost depends on the inner variable, and this dependence can be smooth without being concave. Requiring inner concavity therefore restricts the application models covered by the theory.

For example, the inner variable in adversarial learning is an input perturbation evaluated through a nonlinear prediction model. In robust control, it consists of disturbance-feedback gains that affect subsequent states and their accumulated cost. In mechanical design, it can be an excitation frequency entering the physical frequency-response formula. These dependencies can produce nonconcavity in the perturbation, feedback gains, or frequency, placing such instances outside the scope of concavity-based guarantees.

Outer nonsmoothness serves a separate modeling purpose. An entrywise ℓ1\ell_1 controller cost promotes sparse feedback connections, while absolute-value adjustment costs penalize changes to a physical design. Combining these costs with a nonconcave inner maximization gives the nonsmooth minimax structure we study. The research question is how to retain this structure while making approximate inner optimization sufficient for a guarantee on the original worst-case objective.

2. PL geometry supplies global inner-error control and occurs in concrete models.

The role of PL is to connect an inner gradient residual to the gap from the globally worst response. For a general smooth nonconcave problem, a small gradient can occur at a response whose value is far below the global maximum. Smoothness alone does not control this gap. A PL inequality does, so an approximately stationary inner response is also approximately worst-case in objective value. Our uniform projected-PL condition requires this control to hold consistently as the outer design changes, using the projected-gradient residual appropriate to the inner constraint.

This identifies a structured class of nonconcave problems for which the inner accuracy needed by minimax optimization can be quantified. The two complete examples below show how the structure arises from application models. Zhang et al. [3] establish PL for the disturbance-policy objective at each fixed admissible controller. For mechanical vibration design, we verify uniform projected PL over an augmented design region. In both cases, the outer nonsmooth design cost leaves the inner gradient and maximization gap unchanged.

(a) Robust feedback control provides a model-derived nonconcave PL objective.

Zhang et al. [3] study finite-horizon robust control, where a controller must perform well against disturbances that respond to the evolving state. Their dynamics and feedback policies are

\qquad u_t=-K_ts_t,\quad w_t=-L_ts_t, \qquad t=0,\ldots,T-1.$$ Here $s_t$ is the state, $u_t$ the control, $w_t$ the disturbance, and $\xi_t$ process noise. **The outer variable $K=(K_t)$ designs the controller, while the inner variable $L=(L_t)$ selects the most adverse disturbance-feedback policy.** The minimax objective is $$\min_{K\in\mathcal K}\max_L J(K,L),$$ $$J(K,L)=\mathbb E\!\left[ \sum_{t=0}^{T-1} \left(s_t^\top Q_ts_t+u_t^\top R_t^u u_t-w_t^\top R_t^w w_t\right) +s_T^\top Q_Ts_T\right],$$ with positive semidefinite state weights, positive definite control and disturbance weights, and the robust-feasible controller set $\mathcal K$ specified in [3, Lemma 3.3]. The cost balances state deviations and control effort against a penalty on disturbance energy; robust feasibility ensures a finite worst-case cost. **The inner nonconcavity comes from optimizing a feedback policy.** Substituting the policies gives the closed-loop matrix $A_t-B_tK_t-D_tL_t$. Products of these matrices determine subsequent states, so linear state dynamics and quadratic stage costs do not make the trajectory cost concave in $L$. Under the robust-feasibility and nondegenerate-noise conditions in [3], Lemma 3.5 establishes, for each fixed admissible $K$, $$\\|\nabla_LJ(K,L)\\|_F^2 \ge c_K\bigl[J(K,L^*(K))-J(K,L)\bigr], \qquad c_K>0,$$ where $L^*(K)$ is a globally optimal disturbance policy.

The inequality certifies that an approximately stationary disturbance policy is approximately worst-case in objective value. This is useful for controller optimization because performance must be assessed against the strongest disturbance. It supplies global inner-error control while preserving the nonconcave feedback-policy objective.

Outer nonsmoothness also has an engineering motivation. Sparse feedback design penalizes controller coefficients to reduce the required communication links [4]. Combining that design objective with the robust-control formulation gives

\\{J(K,L) + \tau \sum_ {t=0}^ {T-1}\\|K_ t\\|_ 1 \\},\tau>0.$$ The entrywise $\ell_1$ cost is independent of $L$, so it preserves the inner maximizers, gradients, gaps, and fixed-controller PL inequality. Thus, a nonsmooth controller-design requirement can be added while retaining the inner structure supplied by the control model. Appendix J evaluates this combination of [3] and the sparse-design motivation in [4]. The constant $c_K$ in [3] is controller-dependent. The next example directly verifies the uniform projected-PL condition over an augmented design region. **(b) Mechanical vibration design admits a uniform projected-PL certificate.** Consider the standard damped oscillator, $$m\ddot z+c\dot z+kz=F\cos(\omega t),\qquad H(m,k,\omega)=\frac{F^2}{(k-m\omega^2)^2+c^2\omega^2}.$$ Here $H$ is the squared steady-state displacement amplitude [1]. The outer player selects the mass and stiffness, and the inner player selects the most adverse excitation frequency. Minimizing peak response over a frequency band is an established vibration-design objective [2]. A corresponding design problem is $$\min_{(m,k)\in X}\max_{\omega\in[a,b]} \\{H(m,k,\omega)+\tau_m|m-m_0|+\tau_k|k-k_0|\\},$$ where $0<a<b$, $F,c>0$, and the absolute-value terms represent design-adjustment costs. We combine the physical response formula in [1] with the peak-response design objective motivated by [2]. **The inner geometry is therefore determined by the oscillator's frequency response.** We next identify a design region where that response is nonconcave and satisfies uniform projected PL. The required geometry can be verified over a design region. Define $$z_*(m,k)=\frac{k}{m}-\frac{c^2}{2m^2}.$$ For a compact convex $X$ with positive mass and stiffness throughout $X_{\bar\delta}$, the sufficient condition $$3a^2<z_*(m,k)<b^2\qquad\text{for every }(m,k)\in X_{\bar\delta}$$ places the unique positive-frequency maximum $\omega_* =\sqrt{z_*}$ inside the band and gives $H_{\omega\omega}(m,k,a)>0$. By continuity, positive curvature also occurs inside the band. Thus, **the inner objective is nonconcave for every design in the region**. At the same time, writing $$D=(k-m\omega^2)^2+c^2\omega^2,\qquad d_* =\frac{c^2k}{m}-\frac{c^4}{4m^2}>0,$$ direct calculation gives $$\frac12|H_\omega(m,k,\omega)|^2 =\frac{8F^2m^2\omega^2d_*}{D^3} \bigl[H(m,k,\omega_*)-H(m,k,\omega)\bigr].$$ The coefficient has a positive uniform lower bound $\mu$ on $X_{\bar\delta}\times[a,b]$. To verify the projected condition as well, let $L_\omega=\max|H_{\omega\omega}|$ on that domain. The gradient points toward $\omega_*$ and satisfies $|H_\omega(\omega)|\le L_\omega|\omega-\omega_*|$. Consequently, for **every $0<\lambda\le1/L_\omega$**, the ascent point lies between $\omega$ and $\omega_*$, projection is inactive, and the same inequality holds for the projected-gradient mapping. The adjustment costs cancel from the inner gap. This verifies exactly the uniformity in the outer variable and common stepsize interval required by our projected-PL assumption. For example, in normalized units, $F=c=1$, $X=[0.9,1.1]\times[1.9,2.1]$, $[a,b]=[0.5,2]$, and $\bar\delta=0.05$ give $$1.2306<z_*<1.8374,\qquad 3a^2=0.75,\qquad b^2=4$$ throughout an envelope containing the augmented domain. The outer value is $$V(m,k)=\max_\omega H(m,k,\omega) =\frac{F^2m^2}{c^2(km-c^2/4)}.$$ Its Hessian has negative determinant when $km>5c^2/4$, which also holds throughout this example. Positive adjustment costs preserve this nonconvex curvature away from their kinks and introduce nonsmoothness when $(m_0,k_0)$ lies in the interior. **The physical response formula thus yields a nonsmooth nonconvex outer problem and a smooth nonconcave inner problem with uniform projected PL throughout the augmented design region.** The strict parameter margins make the certificate persist under sufficiently small perturbations. This establishes the geometry for a family of physical designs across a region, rather than only at a selected iterate.

3. Appendix J evaluates the approach on a robust-control application.

To address your concern about application evidence, we have added Appendix J, “Finite-Horizon Robust Control,” marked in blue. Adapted from the time-varying system in Zhang et al. [3, §5.5], the experiment evaluates the controller and disturbance-feedback policies described above. We test both the original cost (τ=0\tau=0) and the nonsmooth controller-cost extension (τ=0.01\tau=0.01), writing Jτ(K,L)=J(K,L)+τ∑t∣Kt∣1J_\tau(K,L)=J(K,L)+\tau\sum_t\\|K_t\\|_1. The task therefore combines optimization through state trajectories with a nonsmooth outer design cost.

(a) Three systems, four methods, and exact worst-case evaluation. We compare ZOFOSGDA, TTSGDA, PGFDA, and ZOFO on a time-varying system with horizon T=10T=10 and fixed systems with T=10T=10 and T=20T=20. Each policy block has 9T9T parameters, giving 90 at T=10T=10 and 180 at T=20T=20. Every configuration uses 5,000 iterations and four seeds, {10,11,12,13}\{10,11,12,13\}. All 96 reported runs complete and remain robust-feasible at the recorded checkpoints. The two fixed systems reuse the parameters selected on the time-varying T=10T=10 instance without retuning.

We evaluate the controller against its globally optimal disturbance response, computed by backward Riccati recursion, and report

\Delta_{\mathrm{in}}(K,L)=\Phi_\tau(K)-J_\tau(K,L).$$ **The two metrics answer complementary questions.** The Riccati-evaluated cost measures controller quality against the globally optimal disturbance, so an insufficiently optimized learned adversary cannot make the controller appear artificially successful. The inner gap then measures how accurately the algorithm's own disturbance policy attains that worst-case value. Appendix J.2 specifies the oracles, parameter search, and evaluation protocol. The time-varying results are the selected four-seed confirmation runs, and the fixed systems test transfer of those parameters. **(b) Results on the nonsmooth control objective.** The table below reproduces the final worst-case costs for $\tau=0.01$ from Table 11 in Appendix J.3. Entries are means $\pm$ sample standard deviations over four seeds; lower is better within each system. | Method | Time-varying system, $T=10$ | Fixed system, $T=10$ | Fixed system, $T=20$ | |---|---|---|---| | ZOFOSGDA | $7.0057\pm0.0277$ | $7.0202\pm0.0186$ | $13.9531\pm0.0380$ | | TTSGDA | $6.9526\pm0.1409$ | $7.0569\pm0.1971$ | $13.9897\pm0.3772$ | | PGFDA | $7.9620\pm0.0261$ | $7.9586\pm0.0303$ | $15.6696\pm0.0558$ | | ZOFO | $8.9769\pm0.1135$ | $8.9637\pm0.1372$ | $18.9890\pm0.2586$ | **ZOFOSGDA attains lower mean worst-case costs than PGFDA and ZOFO in all three systems, for both tested penalty values.** On the nonsmooth instances shown above, its mean cost is approximately $0.76\\%$ higher than TTSGDA on the time-varying system and slightly lower on both fixed systems. TTSGDA uses first-order information in both blocks, whereas our outer updates use function values. ZOFOSGDA's corresponding mean inner gaps are $0.00160$, $0.00149$, and $0.00460$, compared with $0.05306$, $0.05984$, and $0.14006$ for TTSGDA, approximately **30 to 40 times smaller**. Its final costs also have smaller sample standard deviations across seeds. Table 11 reports both penalty values, and Figures 7 to 9 show the cost and inner-gap trajectories. These results show that ZOFOSGDA obtains the reported worst-case controller costs together with disturbance-policy values close to the global inner maximum. Reusing the selected parameters on the fixed systems also tests the approach across changes in dynamics and horizon. The four-seed comparisons fix the main-loop iteration count and use different oracle budgets, including ZOFOSGDA's refresh batches; they describe performance under those configurations. **(c) Why the original and new experiments are both informative.** The controlled classification model in Section 6.1 isolates the geometry required by our theorem, making its nonconcavity and projected-PL condition analytically checkable. With $$T_\alpha(y)=y+\frac{\alpha}{2\pi}\sin(2\pi y),\qquad\alpha=0.7,$$ the appendix verifies positive inner curvature at $y_i=-1/4$ and the uniform projected-PL inequality with $$\mu=\frac cN,\qquad 0<\lambda\le\frac{N(1-\alpha)^2}{c(1+\alpha)^2}.$$ The construction makes it possible to test the algorithm on an instance whose geometry is verified explicitly. The AUC problem has a strongly concave inner variable and supplies supplementary numerical evidence. **Appendix J addresses the separate application question by evaluating worst-case performance in control policy space, including a nonsmooth controller cost.** The revision therefore adds a dynamical-system application alongside the controlled test of the theoretical assumptions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Results

We thank the reviewers for their careful assessments and constructive suggestions. During the discussion period, we have clarified the theoretical contribution, extended the analysis beyond projected-PL geometry, and expanded the empirical evaluation to robust control and OPT-1.3B adaptation. We summarize the main updates below to facilitate further assessment and discussion.

1. Theoretical contribution, generality, and assumptions

  • Clarified the core theoretical contribution and its relation to prior work (Appendix I and the detailed responses). Our main novelty is the unified analysis connecting approximate inner computations to projected Goldstein stationarity of the original nonsmooth value function, Φ(x)=max⁡yf(x,y)\Phi(x)=\max_y f(x,y). Because smoothing and maximization generally do not commute, the framework controls the full averaged inner gap and the resulting outer-gradient error using the original projected-PL geometry. The joint stochastic convergence analysis then establishes that the required residual accuracy is attainable, with both outer and inner query costs included. The revised comparison explains these guarantees alongside the assumptions and objectives of the related theories.

  • Extended both stationarity guarantees to uniform projected-KL geometry (Appendix H, Theorem H.4 and Corollary H.5). For every fixed exponent θ∈(1/2,1)\theta\in(1/2,1), a Hölder tracking bound and accuracy-dependent parameter choices yield both expected blockwise generalized Goldstein stationarity and expected projected Goldstein stationarity of the original value function. The extension retains the remaining assumptions. With the prescribed smoothing radius and uniform initial-potential-gap condition, the query bounds have accuracy exponent 4θ/(1−θ)4\theta/(1-\theta), approaching the projected-PL exponent of 44 as θ↓1/2\theta\downarrow1/2.

  • Strengthened the application justification and clarified the assumption trade-off (Table 1 and the application responses). We provide robust-feedback-control and mechanical-vibration design examples with nonconcave inner objectives. The vibration example verifies uniform projected PL over an augmented design region and a common projected stepsize interval. Revised Table 1 explicitly separates inner regularity, geometry, oracle access, stationarity targets, and query costs, making clear the additional conditions supporting our nonconcave results.

2. Additional experimental evidence

  • Finite-horizon robust control (Appendix J, Table 11 and Figures 7 to 9). We compare four methods on three systems, with two controller-penalty settings and four seeds. Backward Riccati recursion evaluates each controller against its globally optimal disturbance, and we separately report the learned disturbance's inner gap. ZOFOSGDA attains lower mean worst-case costs than PGFDA and ZOFO in all three systems. For the nonsmooth objective, its mean inner gaps are approximately 30 to 40 times smaller than TTSGDA's; its mean cost is 0.76% higher on the time-varying system and slightly lower on both fixed systems. These comparisons use equal main-loop iteration counts and different oracle budgets.

  • OPT-1.3B adaptation on eight language tasks (Appendix K, Table 12). We add comparisons with first-order LoRA, MeZO, Bilevel-ZOFO, PGFDA, and ZOFO. Across three seeds, ZOFOSGDA achieves the highest reported unweighted average test accuracy, 74.69%, compared with 74.39% for Bilevel-ZOFO and 73.32% for ZOFO. These results extend the empirical evaluation to a 1.3-billion-parameter model; they assess single-task adaptation, and the global projected-PL condition is not verified for this LLM objective.

The corresponding manuscript additions and revisions are marked in blue. We appreciate the reviewers' time and would welcome further questions about these results or any remaining concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.