Abstract:We provide a general method to convert a "primal" black-box algorithm for solving regularized convex-concave minimax optimization problems into an algorithm for solving the associated dual maximin optimization problem. Our method adds recursive regularization over a logarithmic number of rounds where each round consists of an approximate regularized primal optimization followed by the computation of a dual best response. We apply this result to obtain new state-of-the-art runtimes for solving matrix games in specific parameter regimes, obtain improved query complexity for solving the dual of the CVaR distributionally robust optimization (DRO) problem, and recover the optimal query complexity for finding a stationary point of a convex function.
What problem does this paper attempt to address?
The core problem that this paper attempts to solve is: **How to efficiently obtain an approximately optimal solution to the dual optimization problem by only using a black - box algorithm for solving the regularized primal optimization problem**. Specifically, the author proposes a general method that can transform a "primal" black - box algorithm for solving the regularized convex - concave minimax optimization problem into an algorithm for solving the corresponding dual maximin optimization problem.
### Specific description of the problem
The paper mainly focuses on the following two types of optimization problems:
1. **Primal minimax optimization problem (P)**:
\[
\min_{x \in X} f(x) \quad \text{where} \quad f(x) := \max_{y \in Y} \psi(x, y),
\]
2. **Dual maximin optimization problem (D)**:
\[
\max_{y \in Y} \phi(y) \quad \text{where} \quad \phi(y) := \min_{x \in X} \psi(x, y).
\]
Here, \(\psi: \mathbb{R}^d\times\mathbb{R}^n\rightarrow\mathbb{R}\) is a differentiable convex - concave objective function, that is, for a fixed \(y\), \(\psi(\cdot, y)\) is a convex function; for a fixed \(x\), \(\psi(x,\cdot)\) is a concave function. \(X\subseteq\mathbb{R}^d\) and \(Y\subseteq\mathbb{R}^n\) are non - empty, closed and convex constraint sets.
### Main contributions of the paper
1. **General dual extraction framework**: The author provides a general framework that can transform a black - box algorithm for solving the regularized primal optimization problem into an algorithm for solving the dual optimization problem. This framework gradually approximates the dual optimal solution within a logarithmic number of rounds through the recursive regularization method.
2. **New running - time complexity**: By applying this framework, the author obtains new optimal running - time complexity for solving matrix games and distributionally robust optimization (DRO) problems within certain parameter ranges.
3. **Recovery of the optimal query complexity for smooth convex functions**: In addition, the author also shows that this framework can recover the optimal query complexity for finding the stable points of smooth convex functions.
### Application examples
1. **Bilinear matrix games**: In the case of the bilinear objective function \(\psi(x, y)=x^{\top}Ay\), the author solves the dual problem - solving difficulty by achieving the same expected running - time complexity as the existing optimal algorithms through this framework.
2. **Distributionally robust optimization (DRO) at CVaR level \(\alpha\)**: For the DRO problem, the author obtains a new optimal query complexity through this framework, which significantly reduces the number of queries required to solve the dual problem.
3. **Calculation of stable points of smooth convex functions**: The author shows that this framework can be used to calculate the approximate stable points of smooth convex functions and achieves the optimal gradient query complexity.
### Summary
This paper solves the problem of efficient transformation from the primal optimization problem to the dual optimization problem by proposing a general dual extraction framework, and has made significant progress especially in areas such as bilinear matrix games and distributionally robust optimization.