Source-linked AI summary

An inertial forward-backward algorithm for the minimization of the sum of two nonconvex functions

Radu Ioan Bot, Ernö Robert Csetnek, Szilárd László

arXiv:1410.0641v1math.OCmath.NA

TL;DR

The paper tackles minimization of a smooth plus nonsmooth objective when both components may be nonconvex. It proposes an inertial forward-backward algorithm and proves convergence to critical points under a KL condition on an appropriate regularization, illustrating the analysis with optimization and image-restoration experiments.

  • Problem

    Forward-backward methods are less understood in the full nonconvex setting, where the proximal point operator may be multivalued.

  • Method

    The paper introduces an inertial forward-backward algorithm and analyzes it using sufficient decrease, subgradient bounds, and KL properties.

  • Results

    The iterates converge to critical points when an appropriate regularization satisfies the KL property, with experiments on local-optimum detection and noisy blurred-image restoration.

  • Takeaways & Limitations

    Inertial effects can detect optimal solutions not found by non-inertial versions, while the framework also addresses nonconvex image restoration.

  • Takeaways & Limitations

    The stated convergence results require either coercivity or a bounded generated sequence when the objective is bounded below.

Abstract

from arXiv · show

We propose a forward-backward proximal-type algorithm with inertial/memory effects for minimizing the sum of a nonsmooth function with a smooth one in the nonconvex setting. The sequence of iterates generated by the algorithm converges to a critical point of the objective function provided an appropriate regularization of the objective satisfies the Kurdyka-Łojasiewicz inequality, which is for instance fulfilled for semi-algebraic functions. We illustrate the theoretical results by considering two numerical experiments: the first one concerns the ability of recovering the local optimal solutions of nonconvex optimization problems, while the second one refers to the restoration of a noisy blurred image.

1 Introduction

The paper addresses the less-understood full nonconvex setting by introducing an inertial forward-backward algorithm and analyzing its convergence under a KL-based condition. Two experiments examine inertial methods for detecting local optima and restoring noisy blurred images.

  • The introduction situates proximal-gradient splitting as a framework that applies gradient steps to smooth terms and proximal steps to nonsmooth terms.
  • Inertial schemes define each new iterate using the previous two iterates and originate from second-order differential inclusions.
  • The paper extends inertial forward-backward methods to objectives whose nonsmooth and smooth components may both be nonconvex.Earlier work allowed a nonconvex smooth part but assumed the nonsmooth part was convex.
  • Convergence is guaranteed when an appropriate regularization of the objective satisfies the Kurdyka-Lojasiewicz property.The proof combines sufficient decrease, a subgradient lower bound for iterate gaps, and KL analytic properties.
  • The numerical experiments test local-optimum recovery that non-inertial methods miss and restoration of a noisy blurred image using nonconvex terms.

2 Preliminaries

The preliminaries establish the KL framework, limiting subdifferential tools, and sequence lemmas used to prove convergence. They also connect KL functions to broad analytic classes, including semi-algebraic functions.

  • The limiting subdifferential is defined through limits of Fréchet subgradients along sequences whose points and function values converge.
  • For convex functions, the Fréchet and limiting subdifferentials coincide with the convex subdifferential.
  • The Fermat rule states that a local minimizer has zero in its limiting subdifferential, while sums with smooth functions obey a subdifferential sum rule.
  • The KL property is defined for proper lower-semicontinuous functions through a desingularizing function satisfying a gradient-subgradient inequality near a point.
  • KL functions include semi-algebraic, real sub-analytic, semiconvex, uniformly convex, and suitable convex functions.
  • The analysis uses uniformized KL properties on compact sets, subdifferential closedness, and sequence convergence lemmas.

3 A forward-backward algorithm

The paper introduces an inertial forward-backward scheme for fully nonconvex objectives and establishes convergence under coercivity, boundedness, and KL-type assumptions. Its iterates approach critical points, with full convergence when the regularized objective is a KL function such as a semi-algebraic function.

  • Algorithm: The proposed algorithm minimizes a nonsmooth f plus smooth g in the fully nonconvex setting using inertial and memory effects.It is formulated through a Bregman-distance-based iterative scheme with gradient, inertial, and proximal terms.
  • Algorithm: The iteration is well-defined when f is proper, lower semicontinuous, bounded below and the Bregman distance is coercive in its first argument.Alternatively, coercivity of each iteration’s minimized objective can replace boundedness of f.
  • Relations to existing methods: The scheme contains known methods as special cases, including inertial proximal-point iterations when g = 0 and classical forward-backward forms in the convex setting.Setting the inertial parameter to zero recovers non-inertial variants discussed in the paper.
  • Convergence analysis: Under the stated parameter conditions, a Lyapunov-type sequence is decreasing and the objective values (f + g)(x_n) converge.The descent inequality combines objective values with squared successive-step terms.
  • Convergence analysis: If f + g is coercive, every cluster point of the generated sequence is a critical point of f + g.Coercivity also yields bounded iterates, enabling subsequential convergence and criticality of cluster points.
  • KL convergence: When the regularized objective is a KL function, the iterates have finite total step length and converge to a critical point.The paper states the result for semi-algebraic f + g as a direct corollary; boundedness below plus bounded iterates can replace coercivity.

4 Numerical experiments

The experiments assess inertial methods on nonconvex minimization and noisy blurred-image restoration. They compare inertial and non-inertial variants across optimization settings and reconstruction conditions.

  • 4.1 Detecting minimizers of nonconvex optimization problems: The first experiment examines whether inertial or memory effects recover optimal solutions missed by non-inertial methods.The study uses a nonconvex nonsmooth function and compares trajectories from four starting points.
  • 4.1 Detecting minimizers of nonconvex optimization problems: Figure 2 compares 100-iteration trajectories from four starting points using non-inertial β = 0 and inertial β = 1.99 or β = 2.99.The first, second, and third columns correspond to these parameter choices, respectively.
  • 4.1 Detecting minimizers of nonconvex optimization problems: The test problem has two optimal solutions, (0, 1/2) and (0, −1/2), obtained from its first-order optimality conditions.The objective combines the nonconvex function f(x1, x2) = |x1| − |x2| with a smooth term whose gradient has Lipschitz constant L∇g = 9/4.
  • 4.2 Restoration of noisy blurred images: The image experiment restores a 256 × 256 boat image from Gaussian-blurred and noisy observations using a nonconvex misfit and nonconvex regularization.The blur operator has size 9 × 9 and standard deviation 4, while the added zero-mean white Gaussian noise has standard deviation 10^-6.
  • 4.2 Restoration of noisy blurred images: After 300 iterations, very small inertial parameters produce ISNR values competitive with, and slightly better than, the non-inertial version.The comparison uses β = 0 for the non-inertial algorithm and evaluates ISNR for different constant β values.
  • 4.2 Restoration of noisy blurred images: Figure 3 presents the original and blurred noisy images in the first row and reconstructed images after 300 iterations in the second row.The reconstructions compare the non-inertial version with the inertial version using β = 10^-7.
Loading 1410.0641v1…