This work considers the nonconvex, nonsmooth problem of minimizing a composite objective of the form $f(g(x))+h(x)$ where the inner mapping $g$ is a smooth finite summation or expectation amenable to variance reduction. In such settings, prox-linear methods can enjoy variance-reduced speed-ups despite the existence of nonsmoothness. We provide a unified convergence theory applicable to a … Read more
Many of the primal ingredients of convex optimization extend naturally from Euclidean to Hadamard spaces — nonpositively curved metric spaces like Euclidean, Hilbert, and hyperbolic spaces, metric trees, and more general CAT(0) cubical complexes. Linear structure, however, and the duality theory it supports are absent. Nonetheless, we introduce a new type of subgradient for convex … Read more
Accelerated proximal gradient methods have become a useful tool in large-scale convex optimization, specially for variational regularization with non-smooth priors. Prevailing convergence analysis considers that users can perform the proximal and the gradient steps exactly. Still, in some practical applications, the proximal or the gradient steps must be computed inexactly, which can harm convergence speed … Read more
It is well known that the performance of the classical Markowitz model for portfolio optimization is extremely sensitive to estimation errors on the expected asset returns. Robust optimization mitigates this issue. We focus on ellipsoidal uncertainty sets around a point estimate of the expected asset returns. An important issue is the choice of the parameters … Read more
The Alternating Direction Method of Multipliers (ADMM) is widely recognized for its efficiency in solving separable optimization problems. However, its application to optimization on Riemannian manifolds remains a significant challenge. In this paper, we propose a novel inertial Riemannian gradient ADMM (iRG-ADMM) to solve Riemannian optimization problems with nonlinear constraints. Our key contributions are as … Read more
Abs-smooth functions are given by compositions of smooth functions and the evaluation of the absolute value. The linear independence kink qualification (LIKQ) is a fundamental assumption in optimization problems governed by these abs-smooth functions, generalizing the well-known LICQ from smooth optimization. In particular, provided that LIKQ holds it is possible to derive optimality conditions for … Read more
Building on [J. Glob. Optim. 89 (2024) 899–926], we continue to focus on solving a nonconvex and nonsmooth structured optimization problem with linear and closed convex set constraints, where its objective function is the sum of a convex (possibly nonsmooth) function and a smooth (possibly nonconvex) function. Based on the traditional augmented Lagrangian construction, we … Read more
Surprisingly, recent work has shown that gradient descent can be accelerated without using momentum–just by judiciously choosing stepsizes. An open question raised by several papers is whether this phenomenon of stepsize-based acceleration holds more generally for constrained and/or composite convex optimization via projected and/or proximal versions of gradient descent. We answer this in the affirmative … Read more
We show that for separable convex optimization, random stepsizes fully accelerate Gradient Descent. Specifically, using inverse stepsizes i.i.d. from the Arcsine distribution improves the iteration complexity from $O(k)$ to $O(k^{1/2})$, where $k$ is the condition number. No momentum or other algorithmic modifications are required. This result is incomparable to the (deterministic) Silver Stepsize Schedule which … Read more
In this paper, we focus on a linearly constrained composite minimization problem whose objective function is possibly nonsmooth and nonconvex. Unlike the traditional construction of augmented Lagrangian function, we provide a proximal-perturbed augmented Lagrangian and then develop a new Bregman Alternating Direction Method of Multipliers (ADMM). Under mild assumptions, we show that the novel augmented … Read more