On the proximal point algorithm for strongly quasiconvex functions in Hadamard spaces

Nicholas Pischke

2024-11-11

Abstract:We prove the convergence of the proximal point algorithm for finding the unique minimizer of a strongly quasiconvex function in general nonlinear Hadamard spaces, generalizing a recent result due to F. Lara. Our argument is rather elementary and brief and relies only on a few properties of strongly quasiconvex functions and their proximal operators which are established here for the first time over these nonlinear spaces. In particular, our convergence proof is fully effective and actually yields fast (ranging up to linear) rates of convergence for the iterates towards the solution and for the function values towards the minimum. These rates are novel even in the context of Euclidean spaces.

Optimization and Control

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to solve the **convergence problem of the proximal point algorithm for strongly quasiconvex functions in Hadamard spaces**. Specifically, the author attempts to prove that the proximal point algorithm can find the unique minimum of strongly quasiconvex functions in Hadamard spaces and provides an estimate of the convergence rate of this algorithm.

#### Main problem background

1. **General form of optimization problems**:
The paper considers the problem of minimizing a function \( f: X \to \mathbb{R} \) on a given space \( X \) and a constraint set \( C \subseteq X \):
\[
\min_{x \in C} f(x)
\]

2. **Classical applications of the proximal point algorithm**:
When \( f \) is convex and lower semi - continuous (lsc), the most well - known solution method is the proximal point algorithm (Proximal Point Algorithm), which constructs the next iteration point by solving the following sub - problem:
\[
\arg\min_{y \in C} \left( f(y) + \frac{1}{2c_k} \| y - x_k \|^2 \right)
\]
where \( x_k \) is the current iteration point and \( c_k>0 \) is a regularization parameter.

3. **Challenges in non - convex cases**:
When the convexity assumption of \( f \) is relaxed, the proximal point algorithm becomes complicated. In particular, for quasiconvex functions, the proximal sub - problem may no longer be uniquely solvable, and the fixed points of the proximal mapping are not necessarily minimum points.

4. **Advantages of strongly quasiconvex functions**:
Strongly quasiconvex functions are a special class of quasiconvex functions that avoid some of the above problems. For example, strongly quasiconvex functions have a unique minimum point, and the proximal point algorithm is still effective in this case.

#### Main contributions of the research

1. **Extension to Hadamard spaces**:
The author extends the previous results of Lara et al. in Euclidean spaces to Hadamard spaces (i.e., complete non - positively curved geodesic metric spaces). This is an important class of nonlinear hyperbolic spaces.

2. **Provision of quantitative convergence results**:
The author not only proves the convergence of the proximal point algorithm but also gives a very fast convergence rate (from sub - linear to linear). These convergence rates are new in Hadamard spaces and are also proposed for the first time even in Euclidean spaces.

3. **Concise proof methods**:
The author uses the generalization of some key properties and short argument methods, making the proof process more intuitive and easier to understand. This method avoids relying on complex tools such as sequential compactness and can thus be applied to a wider range of nonlinear spaces.

In summary, this paper solves the problems of convergence and convergence rate of the proximal point algorithm for strongly quasiconvex functions in Hadamard spaces and provides new theoretical support for this field.