An inexact golden ratio primal-dual algorithm with linesearch step for a saddle point problem

Changjie Fang,Jinxiu Liu,Jingtao Qiu,Shenglan Chen

2024-08-16

Abstract:In this paper, we propose an inexact golden ratio primal-dual algorithm with linesearch step(IP-GRPDAL) for solving the saddle point problems, where two subproblems can be approximately solved by applying the notations of inexact extended proximal operators with matrix norm. Our proposed IP-GRPDAL method allows for larger stepsizes by replacing the extrapolation step with a convex combination step. Each iteration of the linesearch requires to update only the dual variable, and hence it is quite cheap. In addition, we prove convergence of the proposed algorithm and show an O(1/N) ergodic convergence rate for our algorithm, where N represents the number of iterations. When one of the component functions is strongly convex, the accelerated O(1/N2) convergence rate results are established by choosing adaptively some algorithmic parameters. Furthermore, when both component functions are strongy convex, the linear convergence rate results are achieved. Numerical simulation results on the sparse recovery and image deblurring problems illustrate the feasibility and efficiency of our inexact algorithms.

Optimization and Control

What problem does this paper attempt to address?

### Problems Addressed by the Paper

The paper primarily aims to address the saddle point problem, particularly by proposing an approximate golden ratio primal-dual algorithm (IP-GRPDAL) when subproblems are difficult to solve exactly. Specifically, the paper addresses the following key issues:

1. **Approximate Solution of Subproblems**:
- When the proximal operators of the objective functions $f$ and $g$ are difficult to compute, a method involving an extended proximal operator with matrix norms is proposed.
- The proposed IP-GRPDAL method can approximately solve the two subproblems under the type 2 approximate criterion.

2. **Convergence and Convergence Rate**:
- Under the step size condition $\tau\sigma\|A\|^2_T < \phi$, the global convergence and asymptotic convergence rate of $O(1/N)$ are proven, where $N$ denotes the number of iterations.
- When one or both component functions are strongly convex, accelerated convergence rates of $O(1/N^2)$ and linear convergence rates are established, respectively.

3. **Adaptive Step Size Strategy**:
- A line search strategy is used to update the dual variable, allowing for larger step sizes and thus reducing the computational cost of algorithm iterations.
- The explicit computation for the next iteration is improved compared to previous work.

4. **Numerical Experiment Validation**:
- Numerical experiment results on sparse recovery and image deblurring problems show that the proposed algorithm outperforms several existing methods.

In summary, the paper aims to propose a new approximate golden ratio primal-dual algorithm and validate its effectiveness and superiority through theoretical analysis and numerical experiments.