Projection-Free Methods for Solving Nonconvex-Concave Saddle Point Problems

In this paper, we investigate a class of constrained saddle point (SP) problems where the objective function is nonconvex-concave and smooth. This class of problems has wide applicability in machine learning, including robust multi-class classification and dictionary learning. Several projection-based primal-dual methods have been developed for tackling this problem; however, the availability of methods with projection-free oracles remains limited. To address this gap, we propose efficient single-loop projection-free methods reliant on first-order information. In particular, using regularization and nested approximation techniques, we propose a primal-dual conditional gradient method that solely employs linear minimization oracles to handle constraints. Assuming that the constraint set in the maximization is strongly convex, our method achieves an $\epsilon$-stationary solution within $\mathcal{O}(\epsilon^{-6})$ iterations. When the projection onto the constraint set of maximization is easy to compute, we propose a one-sided projection-free method that achieves an $\epsilon$-stationary solution within $\mathcal{O}(\epsilon^{-4})$ iterations. Moreover, we present improved iteration complexities of our methods under a strong concavity assumption. To the best of our knowledge, our proposed algorithms are among the first projection-free methods with convergence guarantees for solving nonconvex-concave SP problems.

Paper

Similar papers

Peer review

Reviewer cP1u5/10 · confidence 3/52023-06-29

Summary

This paper studies the constrained nonconvex-concave minimax problem. This problem has been studied in several papers in the literature, but this paper proposes a projection free (single loop) algorithm to solve this problem.

Strengths

The proposed algorithms are interesting and extend Frank-Wolf type methods to the nonconvex setting.

Weaknesses

1. The authors have made some effort into explaining why the FMO oracle might be much more computationally efficient as compared to a projection onto a set. The motivating example seems to be that of the nuclear norm constraint. Can the authors describe in a little more detail as to why this is the case for this constraint. It would make the paper more complete and provide motivation to study projection free methods. 2. Once again, the motivating example of the nuclear constraint does not seem to be addressed in the assumption of $\alpha$-strongly-convex sets. Are these constraints not strongly convex? If this is the case, then there seems to be a mismatch between the motivating examples and the assumptions. 3. What is the relation between the strong convexity of the set and the strong convexity of the objective function? For example, if the inner problem is strongly concave (like in the setting of Theorem 4.4), do we still need to assume the strong convexity of the constraint set?

Questions

See above

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

See above

Reviewer LwGk5/10 · confidence 3/52023-07-02

Summary

This paper investigates algorithms for constrained saddle point (SP) problems where the objective function is nonconvex-concave and smooth. Existing methods are usually projection-based and this paper focuses on developing single-loop projection-free algorithms which only use linear minimization oracles. In particular, this paper provides convergence guarantees for nonconvex-concave SP problems and nonconvex-strongly concave SP problems. This paper also investigates one-sided projection-free methods which can achieve an improved convergence matching the SOTA results for projection-based methods.

Strengths

This paper is well-written and easy to follow. The contributions of this paper are also straightforward: developing projection-free algorithms for saddle point problems and providing convergence guarantees for the proposed methods. By considering the LMO-PO oracle, the rate obtained matches the SOTA convergence of projection-based algorithms.

Weaknesses

1) When analyzing the convergence guarantees for the R-PDCG method, the author also assumes that Y is S-Convex set. Is this condition inevitable? Also, the rate for R-PDCG is slightly worse than projection-based methods, is this rate improvable, or it is already optimal for projection-free algorithms? 2) It seems a bit strange that the experiment results are shown in the Introduction part. The paper would be more convincing if the authors can present more extensive experiments and additional numerical results. 3) It seems that parameter $\mu$ and $\tau$ are related to parameter $K$; how to set $K$ in practice?

Questions

See the "Weaknesses" part.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The authors addressed the limitations adequately.

Reviewer dDgV6/10 · confidence 4/52023-07-04

Summary

This paper proposes projection-free optimization algorithms for constrained nonconvex-(strongly) concave saddle point problem. Solution concept: $\epsilon$-stationarity To this end, they propose R-PDCG and CG-RPGA algorithms. 1. Without projection, iteration complexity for R-PDCG is $O(\epsilon^{-6})$ for nonconvex-concave and $O(\epsilon^{-4})$ nonconvex- strongly concave objective. 2. With projection for the maximization step, iteration complexity for CG-RPGA is $O(\epsilon^{-4})$ for nonconvex-concave and $O(\epsilon^{-2})$ nonconvex- strongly concave objective. They illustrate the results through experiments.

Strengths

1. The problem is well-motivated and the applications are clear. 2. This is the first projection-free method for constrained nonconvex-concave saddle point problem.

Weaknesses

I am judging the paper by it's theoretical contribution because the applications are supportive of the theory but that's not the main message. The main weakness of the work is the novelty of the theoretical tools used in the proofs. **Lemma 4.1 + Lemma B.1 are the keys to the proofs. The proof techniques, specifically (12)-(14) (rest of the proof follows by assumptions) are standard steps in any FW-based method for smooth functions (see [14] for example).**

Questions

1. Could you comment on the optimality of the bounds, i.e., how tight are the bounds? 2. Could you highlight the novelties required for the proof beyond Lemma 4.1, and Lemma B.1? 3. In the experiments, the algorithm does converge to a stationary point, but does stationarity guarantee a good solution for these applications? 4. It seems like if the function is smooth nonconvex-nonconcave, the algorithm should guarantee stationarity. If yes, you should highlight that in the paper. If not, could you explain why? In other words, how important is the convexity?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

2 fair

Limitations

See "Weakness" section.

Reviewer dDgV2023-08-19

Thanks for the response

I am happy with the response and keep my score.

Reviewer WRhg6/10 · confidence 3/52023-07-06

Summary

This paper proposed two projection-free algorithms for solving smooth nonconvex- (strongly) concave saddle point problems. The authors showed that the convergence rates of the proposed algorithms matches the state-of-the-art convergence rate of projection-based methods. Experimental results on dictionary learning verify that the proposed algorithms show great advantage in terms of training time compared to existing projection-based methods.

Strengths

The motivation and contribution of this paper are clear. Projections in algorithms could be problematic in practice and potentially slow down the training. This work fills this gap in nonconvex-(strongly) concave saddle point problems.

Weaknesses

I do not see any major weakness, but I do have some questions. Please see the Questions section.

Questions

1. In the related work section, I would recommend the authors to discuss the works on projection-free methods for solving bilevel optimization. Since saddle point problem can be viewed as a special case of bilevel problem, methods proposed for solving bilevel problems should be applicable for SP problems. 2. The result in line 247 needs a reference. I believe it is related to Danskin's Lemma. 3. In the experiment part, I noticed that the one-sided projection-free method CG-RPGA has better performance in terms of training time compared with the fully projection-based method R-PDCG. The authors argue that this matches the convergence rates. I'm not sure about this argument, because the convergence rates are in terms of number of iterations and time per-iteration is obviously different for each methods. I think the training time plots are trying to verify that even AGP has the same convergence rate as CG-RPGA, AGP is slower mainly due to the projection oracle. If the authors would like to verify the different convergence rates of CG-RPGA and R-PDCG, a plot of gap function and iteration number is more appropriate.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

Limitations are well discussed in the last section of the paper.

Reviewer RF277/10 · confidence 4/52023-07-26

Summary

This paper proposes two Frank-Wolfe (FW) based algorithms for solving a class of nonconvex-(strongly) concave saddle point problems. The proposed algorithms are among the first projection-free methods with convergence guarantees for such problems as authors claimed. The paper uses regularization and nested approximation techniques to deal with the nonsmooth component, and apply it to the primal-dual scheme. Shortly, it approximates the nonsmooth function $f(x)$ by $\mu$ in convex-concave setting. If the objective function is strongly concave in $y$, the regularization can be avoided by setting $\mu=0$. In terms of novelty, the techniques used in this paper are common, and the analysis seems quite classic to me. However, this simple combination still brings interesting results. I think this is a very good and well written paper, and it has made a great contribution. Therefore, I suggest accepting this paper, but I may change my perspective based on other comments.

Strengths

Originality. This paper is a good combination of the regularization technique and Frank-Wolfe method. This allows to obtain a single-loop projection-free method with cheaper computational cost for the nonconvex-concave problem. Quality. As far as I see, the proofs are correct. Experiments show the advantage of the projection-free methods. It would be better to add another simulation example for situations where projecting on constraints $X$ and $Y$ are difficult. Clarity. The paper is easy to understand and the results are clearly stated and well-organized. I would like to suggest the authors to double check language, symbols, and definitions. For example, "problem 1" should be changed to "problem (1)", and $\mathcal{G}_X(\bar z)$ should be changed to $\mathcal{G}_X(\bar x,\bar y)$ in Definition 2.1. Significance. This paper considers a class of nonconvex-concave saddle-point problems, which widely exists in robust optimization, reinforcement learning and adversarial learning. Given existing results, the main contribution of this paper are about solving such problems via projection-free schemes, which reduce the computational complexity in dealing with the problem with structured complicated constraint set, such as nuclear norm ball. The proposed methods can be useful in practice because of its cheaper computational cost and ability to solve the problem with complicated constraint sets.

Weaknesses

The convergence requirement of the fully projection-free method R-PDCG is that the set $Y$ is strongly convex, which is very limited in practical applications. If this assumption can be removed while achieving faster convergence performance (comparable to projection based methods), it would be a better result. In addition, the value of step size $\tau_k$ and the parameter $\mu_k$ is related to the total iteration $K$. If the total number of iterations is large, this will result in a small step size of the algorithms and slow convergence. It would be better to improve the step size $\tau_k$ to a constant that is independent of the total number of iterations.

Questions

The four theorems proposed in this paper said that "there exists $t\in\{\cdots\}$ such that ... satisfy the following bounds". Does this mean that only a limited amount of iterations satisfies the boundary? Is this measure reasonable and what is its practical significance?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

None

Reviewer RF272023-08-15

In the plots of the algorithms in terms of iteration counters in Example 2 (Dictionary Learning), I wonder why the fully projection-based algorithm AGP has a slower convergence performance than FW-based algorithms. It seems that AGP has a comparable convergence result to the result of CG-RPGA and is even faster than R-PDCG in theory.

Authorsrebuttal2023-08-15

We appreciate the reviewer for the follow-up question. We would like to highlight that in the examples we considered in the paper, the projection onto the constraint set $X$ requires full SVD decomposition, therefore, it leads to a higher computational cost for AGP algorithm. As we fixed the running time of algorithms in all the experiments, AGP will take fewer iterations compared to other methods. Therefore, it is very important to acknowledge the distinct oracles these algorithms employ when comparing their complexity results as stated in Table 1 of our paper. Moreover, it should be noted that the complexity results available for these methods are only upper bounds for the gap functions and the algorithms may have better performances on specific examples. The examples provided in the numerical experiments support the motivation behind the development of projection-free methods for saddle point problems, particularly when an LMO is available.

Reviewer RF272023-08-15

It would be better to provide the step sizes of different algorithms used in Simulation. In addition, please state the reason for the choice of $\tau_k$ and $\mu_k$ , which are not discussed in the whole paper.

Authorsrebuttal2023-08-15

Thank you for pointing this out. (i) Due to space limitations, we have relegated the details of our experiment to section F of the Appendix. In the final version of the paper, having one additional page will allow us to relocate this information to the main body of the paper. As it is mentioned in Appendix section F, ``For all the algorithms, the step-sizes are selected as suggested by the papers and scaled to have the best performance. For AGP we let the primal step-size $\frac{1}{\sqrt{k}}$, dual step-size as 0.2, and the dual regularization parameter as $\frac{10^{-1}}{k^{1/4}}$; for SPFW both primal and dual step-sizes are selected to be diminishing as $\frac{2}{k+2}$". (ii) The selection of the step size $\tau_k$ and parameter $\mu_k$ is discussed in the proof of Corollaries 4.3 and 5.2. These parameters are selected to minimize the upper bound derived in Theorems 4.2 and 5.1, respectively. For instance, in Theorem 4.2, we have provided an explicit upper bound on the primal and dual gap functions. Considering the dominant terms in the aggregation of these two bounds in terms of $\tau,\mu$, and $K$ we observe that $\mathcal{G}_Z(x_t,y_t)\leq \mathcal O(\frac{1}{\tau K}+\frac{\tau^{1/3}}{\mu^{2/3}}+\mu)$. Therefore, selecting $K=\mathcal O(\epsilon^{-6})$, $\tau=\mathcal O(\epsilon^5)$ and $\mu=\mathcal O(\epsilon)$ implies that $\mathcal{G}_Z(x_t,y_t)\leq \epsilon$ after $K=\mathcal O(\epsilon^{-6})$ iterations.

Reviewer RF272023-08-16

Many thanks for the authors' response. I strongly agree with the author's statement that projection-free methods may exhibit better convergence performance than projection-based methods in terms of the running time under some special set constraints. But I think the projected method should exhibit comparable or better performance than the projection-free method in terms of the iteration in theory even in hard-to-project cases, which is inconsistent with the first plot (left figure) of the second example. Could the author explain the reason why the FW-based algorithms exhibit faster convergence performance than AGP in terms of iteration in the second example? Also, I have one more question. In the simulation, what method was used to solve the projection operation of the projected-based algorithm AGP, especially in the case of the constraint set being a nuclear norm ball, whether there are differences in the algorithm performance caused by different solution methods.

Authorsrebuttal2023-08-16

(i) Thank you for your follow-up questions. One important property of FW-type methods is that they usually add just one point (often a vertex) in each iteration. This is an important property in many applications For example, when the optimization has a nuclear norm ball constraint each step adds only one rank-1 matrix and it leads to a sparse iterates. After $t$ iterations, the resulting solution has a rank of at most $t+1$. A reason that our methods outperform AGP in these examples can stem from this important property of FW-type methods as the optimal (stationary) solutions are potentially sparse. See section 1.1. of [R1] for more details. To have a better understanding of AGP performance, we re-executed the method with tuned stepsizes, resulting in a comparable outcome for AGP when compared to R-PDCG. In the final version of the paper, we will incorporate new plots showcasing the performance of AGP. Additionally, for the Robust Multiclass Classification problem, we intend to conduct further AGP iterations to provide enhanced clarity regarding its performance. (ii) The primal projection step in AGP involves projecting onto nuclear-ball constraint for variable $C'$ and projecting onto column-wise unit $l2$-norm ball for variable $D'$. For projection onto nuclear-ball constraint, we invoked svd function in Matlab to decompose the gradient at each iteration, and subsequently the vector of singular values is projected onto the simplex set. [R1] Braun, G., Carderera, A., Combettes, C. W., Hassani, H., Karbasi, A., Mokhtari, A., \& Pokutta, S. (2022). Conditional gradient methods. arXiv preprint arXiv:2211.14103.

Reviewer RF272023-08-18

Thanks for the authors' response. I think it is a solid paper. I will keep my score unchanged.

Reviewer WRhg2023-08-15

Thanks the authors for the response and additional plots. For the Robust Multiclass Classification experiment, due to the few iterations of AGP shown in the iteration plots in Figure 1 and Figure 2, it is still not clear how AGP performs compared with the proposed methods in terms of iterations. Moreover, as Reviewer RF27 mentioned in the comment, in the Dictionary Learning experiment, AGP converges much slower in terms of iterations. This is does not match the theory. I wonder if the authors have any insights on this observation. For now, I will keep my score unchanged.

Authorsrebuttal2023-08-15

This aligns precisely with the core motivation of our paper. As detailed in our response to Reviewer cP1u, projecting onto the nuclear-norm constraint incurs a higher computational cost compared to the corresponding LMO. The reason is that the projection operation requires a full SVD decomposition while LMO requires finding the left and right singular vectors corresponding to the largest singular value of $\nabla_x\mathcal L(x_k,y_k)$ (Please also see our response to reviewer RF27). In the Robust Multiclass Classification example, we observe that the high per-iteration computational cost of the projection operator significantly impacts AGP, with more than three iterations taking over 300 seconds reflecting the benefit of projection-free algorithms for a certain class of problems. Therefore, it is very important to acknowledge the distinct oracles these algorithms employ when comparing their complexity results as stated in Table 1 of our paper. For problems with easy-to-project constraints, projection-based algorithms such as AGP may have a better performance, however, the examples provided in the numerical experiments support the motivation behind the development of projection-free methods for saddle point problems with hard-to-project constraints, particularly when an LMO is available. For the final version of the paper, we will run AGP for additional iterations to enhance the clarity of its performance. We appreciate the reviewer's question and would be more than happy to address any other concerns they may have.

Reviewer cP1u2023-08-15

Response to Rebuttal

I'd like to thank the authors for their response. I have increased my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC