Abstract
The Bregman proximal mapping and Bregman-Moreau envelope are traditionally studied for functions defined on the entire space $\mathbb{R}^n$, even though these constructions depend only on the values of the function within (the interior of) the domain of the distance-generating function (dgf). While this convention is largely harmless in the convex setting, it leads to substantial limitations in the nonconvex case, as it fails to embrace important classes of functions such as relatively weakly convex ones. In this work, we revisit foundational aspects of Bregman analysis by adopting a domain-aware perspective: we define functions on the natural domain induced by the dgf and impose properties only relative to this set. This framework not only generalizes existing results but also rectifies and simplifies their statements and proofs. Several examples illustrate both the necessity of our assumptions and the advantages of this refined approach.
Introduction
In modern optimization and variational analysis, the proximal mapping and the Moreau envelope are indispensable tools for algorithm design and convergence analysis. When moving beyond standard Euclidean geometry, researchers often employ a Bregman divergence—a non-symmetric pseudo-distance derived from a convex distance-generating function (dgf), denoted as $h$.
Traditionally, the literature assumes that the functions $f$ being optimized are defined over the entire ambient space $\mathbb{R}^n$. However, this convention creates significant theoretical hurdles in the nonconvex setting. If the domain of the dgf is not closed, many useful nonconvex functions—specifically "relatively weakly convex" functions—cannot be extended to the whole space without losing essential properties like lower semicontinuity. This paper addresses these limitations by rebuilding the foundational mathematics of Bregman operators through a domain-aware perspective.
The Domain-Aware Framework
Instead of forcing functions to exist globally on $\mathbb{R}^n$, this research defines them strictly on the subsets where they naturally operate: the domain of the dgf ($\mathcal{X}$) and its interior ($\mathcal{Y}$).
Relative Topology
By treating these subsets as topological spaces in their own right, the authors demonstrate that properties such as continuity or compactness can be established without reference to the ambient space. This shift avoids the "extended arithmetics" often required to resolve $\infty - \infty$ cases when functions are evaluated outside their valid domains.
Direct Operator Analysis
Because Bregman divergence is asymmetric, it generates distinct "left" and "right" proximal mappings. Standard practice often involves converting right operators into left ones via complex transformations. This work instead analyzes right operators directly using epi-composite functions. Furthermore, the authors frame these operators within the theory of generalized $\Phi$-conjugacy, showing how left and right envelopes act as natural duals in an optimal transport-like framework.
Key Insights and Theoretical Corrections
Intrinsic Topology and Literature Rectification
A central insight of this work is that the intrinsic topology of the dgf domain is sufficient for all relevant analysis. The paper identifies and corrects several imprecise claims in existing literature. For instance, previous studies often claimed that certain operator graphs were closed in the ambient space $\mathbb{R}^n \times \mathbb{R}^n$. The authors prove this is generally false; such graphs are only closed relative to the restricted domain $\mathcal{Y} \times \mathbb{R}^n$.
Embracing Nonconvexity
By abandoning the full-space requirement, the Bregman framework can now accommodate relatively weakly convex functions. These are functions that become convex only when a multiple of the dgf is added to them. Such functions are prevalent in modern machine learning and signal processing, and this new framework allows them to be analyzed with the same rigor as their convex counterparts.
Euclidean Reformulations
The authors establish that both left and right Bregman operators can be represented globally as standard Euclidean operators applied to appropriately transformed functions. This provides a vital theoretical bridge between non-Euclidean and Euclidean optimization theories.
Mathematical Results
As a theoretical advancement, the results are presented as foundational theorems:
- Expanded Validity: The authors prove that Bregman proximal mappings remain continuous and Moreau envelopes remain smooth for a much broader class of functions than previously established.
- Relative Smoothness and Cocoercivity: The paper establishes a novel mathematical equivalence linking "Bregman cocoercivity" and "anisotropic strong convexity." This result serves as a Bregman-space equivalent to the Baillon-Haddad theorem, a cornerstone of standard convex analysis.
- Proof Simplification: By restricting the domain, the authors eliminate the need for complex workarounds, leading to more elegant and shorter proofs for established theorems.
Limitations and Future Directions
The current framework assumes that the distance-generating function is differentiable on the interior of its domain. Additionally, the focus remains on the mathematical foundations rather than the implementation of specific algorithms.
Future research may extend this domain-aware perspective to Klee envelopes or explore the relaxation of the differentiability assumption. Developing a rigorous framework for non-differentiable or merely convex generators could lead to an even more comprehensive theory of Bregman-type optimization methods.
Summary updated