Bregman level proximal subdifferentials and new characterizations of Bregman proximal operators

Ziyuan Wang, Andreas ThemelisarXiv 2025 · DOI · arXiv

Abstract

Classic subdifferentials in variational analysis may fail to fully represent the Bregman proximal operator in the absence of convexity. In this paper, we fill this gap by introducing the left and right Bregman level proximal subdifferentials and investigate them systematically. Every Bregman proximal operator turns out to be the resolvent of a Bregman level proximal subdifferential under a standard range assumption, even without convexity. Aided by this pleasant feature, we establish new correspondences among useful properties of the Bregman proximal operator, the underlying function, and the (left) Bregman level proximal subdifferential, generalizing classical equivalences in the Euclidean case. Unlike the classical setting, asymmetry and duality gap emerge as natural consequences of the Bregman distance. Along the way, we improve results by Kan and Song and by Wang and Bauschke on Bregman proximal operators. We also characterize the existence and single-valuedness of the Bregman level proximal subdifferential, investigate coincidence results, and make an interesting connection to relative smoothness. Abundant examples are provided to justify the necessity of our assumptions.

Introduction

In optimization, the proximal operator is a fundamental tool for decomposing complex problems into simpler sub-problems. In the standard Euclidean setting with convex functions, a well-established one-to-one relationship exists: the proximal operator is exactly the resolvent of the function’s subdifferential. However, this elegant correspondence falters in two significant scenarios:

  1. Non-convexity: Standard subdifferentials do not fully capture the behavior of the proximal operator for non-convex functions.
  2. Bregman Geometry: When employing Bregman distances—which adapt to problem-specific geometry (e.g., in Mirror Descent)—the inherent asymmetry (where the "left" distance differs from the "right") complicates the mathematical framework.

While previous work by Rockafellar addressed non-convexity in Euclidean space using "level proximal subdifferentials," this paper extends these concepts to the Bregman setting, filling a critical gap in variational analysis.

Methodology

The authors utilize a rigorous theoretical framework within variational analysis to bridge these gaps. Key methodological contributions include:

  • Asymmetric Subdifferentials: To account for Bregman asymmetry, the authors define distinct "left" ($\overleftarrow{\partial}_{\phi/\lambda}^p f$) and "right" ($\overrightarrow{\partial}_{\phi/\lambda}^p g$) versions of the level proximal subdifferential.
  • Warped Resolvent Representation: A central proof demonstrates that the Bregman proximal mapping is a "warped resolvent" of these new subdifferentials, an identity that holds regardless of function convexity.
  • Property Characterization: The framework establishes rigorous correspondences between function properties (e.g., $B_\phi$-weak convexity), operator properties (e.g., monotonicity), and subdifferential characteristics.

Key Insights and Results

The most significant insight is that every Bregman proximal operator can be viewed as a resolvent of a Bregman level proximal subdifferential. This provides a unified lens for studying proximal operators without the restrictive assumption of convexity.

Generalizing Classical Equivalences

The paper successfully generalizes several classical Euclidean results to the Bregman setting. For instance, Theorem 4.4 establishes equivalent conditions for $B_\phi$-weak convexity, linking it to the maximal monotonicity of the proximal operator.

Comparison of Euclidean and Bregman settings for weak convexity characterizations.

Furthermore, the authors explore the relationship between function convexity and operator nonexpansiveness. Theorem 4.9 establishes the equivalence between function convexity and the "$\nabla\phi$-firm nonexpansiveness" of its proximal operator. The study also extends to the Bregman-Moreau envelope, highlighting how the "duality gap" in Bregman space affects envelope convexity.

Equivalences for convexity, nonexpansiveness, and envelope properties.

Refinement of Existing Theory

Beyond new characterizations, the authors refine previous results by Kan, Song, Wang, and Bauschke. By removing unnecessary assumptions and providing more general proofs, they offer a more robust foundation for Bregman-based optimization.

Limitations and Future Work

Despite its contributions, the theory relies on certain standard but restrictive assumptions:

  • Range Assumption: Key results often require the range of the proximal operator to remain within the interior of the distance-generating function's domain.
  • 1-Coercivity: The distance-generating function $\phi$ is assumed to be 1-coercive, which may exclude certain edge cases.

Future research directions include investigating "variational convexity" (a local form of convexity) within the Bregman framework and further exploring the interplay between the convexity of the Bregman-Moreau envelope and the underlying function.

Summary updated

Abstract

Classic subdifferentials in variational analysis may fail to fully represent the Bregman proximal operator in the absence of convexity. In this paper, we fill this gap by introducing the left and right \emph{Bregman level proximal subdifferentials} and investigate them systematically. Every Bregman proximal operator turns out to be the resolvent of a Bregman level proximal subdifferential under a standard range assumption, even without convexity. Aided by this pleasant feature, we establish new correspondences among useful properties of the Bregman proximal operator, the underlying function, and the (left) Bregman level proximal subdifferential, generalizing classical equivalences in the Euclidean case. Unlike the classical setting, asymmetry and duality gap emerge as natural consequences of the Bregman distance. Along the way, we improve results by Kan and Song and by Wang and Bauschke on Bregman proximal operators. We also characterize the existence and single-valuedness of the Bregman level proximal subdifferential, investigate coincidence results, and make an interesting connection to relative smoothness. Abundant examples are provided to justify the necessity of our assumptions.
BibTeX
@misc{wang2025bregman,
  author = {Ziyuan Wang and Andreas Themelis},
  title = {{Bregman level proximal subdifferentials and new characterizations of Bregman proximal operators}},
  year = {2025},
  eprint = {2506.07333},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.07333},
  howpublished = {arXiv},
}
Figure