Abstract:We investigate the role of constraints in the computational complexity of min-max optimization. First, we show that when the constraints are jointly convex (i.e., the min player and max player share the same constraints), computing a local min-max equilibrium with a nonconvex-concave objective is PPAD-hard. This improves the result of Daskalakis, Skoulakis, and Zampetakis [2021] along multiple directions: it applies to nonconvex-concave objectives (instead of nonconvex-nonconcave ones) that are degree-two polynomials, and it's essentially tight in the parameters. Second, we show that with general constraints (i.e., the min player and max player have different constraints), even convex-concave min-max optimization becomes PPAD-hard. Conversely, local min-max equilibria for nonconvex-concave and convex-concave objectives can be computed in polynomial time under simpler classes of constraints. Therefore, our results show that constraints are a key driver of the complexity of min-max optimization problems. Along the way, we also provide PPAD-membership of a general problem related to quasi-variational inequalities, which has applications beyond our problem.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is to explore how different types of constraints affect the computational complexity of min - max optimization problems. Specifically, the author has studied the computational complexity of local min - max equilibrium points in min - max optimization problems under constraints of different structures (such as jointly convex constraints, bilinear constraints, and product constraints). Through this research, the author hopes to reveal the crucial role of constraints in the complexity of min - max optimization problems. ### Main Contributions 1. **Non - convex - concave objective functions under jointly convex constraints**: - The author has proven that when the constraint is jointly convex, even if the objective function is non - convex - concave, computing the local min - max equilibrium point is still a PPAD - hard problem. This result improves the work of Daskalakis et al. [DSZ21], is applicable to a wider range of quadratic polynomial - form objective functions, and is almost optimal in terms of parameters. 2. **Convex - concave objective functions under general constraints**: - The author has further shown that under general constraints (i.e., the minimizing player and the maximizing player have different constraints), even convex - concave min - max optimization problems become PPAD - hard. This indicates that the complexity of the constraints significantly increases the difficulty of the problem. 3. **Solvability under simplified constraints**: - Conversely, under certain simplified categories of constraints, local min - max equilibrium points can be computed in polynomial time. This shows that the structure of the constraints has an important impact on the solvability of the problem. 4. **PPAD membership of quasi - variational inequalities**: - As a by - product, the author has also proven that a class of problems related to quasi - variational inequalities belongs to the PPAD class, which also has wide applications in other fields. ### Key Techniques - **Linear variational inequalities**: The author first proved that the linear variational inequality problem on a hypercube is PPAD - hard, which is obtained by reduction from the Nash equilibrium problem of polymatrix games. - **Pseudo - gradient fields**: The author distinguishes between pseudo - gradient fields and standard gradient fields and uses the complexity of pseudo - gradient fields to establish hardness results. - **Perfect imitation and bilinear constraints**: Use bilinear constraints to achieve simple imitation of linear operators, so that the behavior of one player can exactly imitate that of another player. - **Approximate imitation and jointly convex constraints**: Under jointly convex constraints, the author has designed an approximate imitation mechanism, so that the behaviors of the two players are consistent to a certain extent, thus solving the challenges in previous methods. ### Conclusion Through these results, the author has revealed the important role of constraints in the complexity of min - max optimization problems, especially in the cases of non - convex - concave and convex - concave objective functions. In addition, the author has also provided some new PPAD membership results, further enriching the theoretical foundation in this field.

On the Role of Constraints in the Complexity of Min-Max Optimization

The Complexity of Constrained Min-Max Optimization

On the Complexity of a Practical Primal-Dual Coordinate Method

Optimality-based domain reduction for inequality-constrained NLP and MINLP problems

Constraint Qualifications for Optimality Conditions and Total Lagrange Dualities in Convex Infinite Programming

A Weak Maximum Principle for Discrete Optimal Control Problems with Mixed Constraints

Complexity Lower Bounds for Nonconvex-Strongly-Concave Min-Max Optimization.

A Maximal Tractable Class of Soft Constraints

Constrained Phi-Equilibria

A Minimization Approach for Minimax Optimization with Coupled Constraints

Statistical Mechanics of Min-Max Problems

Beyond first-order methods for non-convex non-concave min-max optimization

Faster Accelerated First-order Methods for Convex Optimization with Strongly Convex Function Constraints

Lower Bounds for Matroid Optimization Problems with a Linear Constraint

Chasing Convex Functions with Long-term Constraints

Bridging the Gap Between General and Down-Closed Convex Sets in Submodular Maximization

On the equivalence between the minimax theorem and strong duality of conic linear programming

Extreme Point Pursuit -- Part I: A Framework for Constant Modulus Optimization

Constrained Non-Monotone Submodular Maximization: Offline and Secretary Algorithms

Expected Maximin Fairness in Max-Cut and other Combinatorial Optimization Problems

Near-Optimal Solutions of Constrained Learning Problems