Bregman Proximal Point Algorithm

• Bregman Proximal Point Algorithm is a family of regularization methods that replaces the quadratic proximity term with a Bregman distance generated by a convex differentiable kernel.
• It underpins various optimization techniques including convex minimization, equilibrium problems, and primal-dual splitting, offering rigorous convergence guarantees.
• The algorithm’s flexibility supports inexact, accelerated, and manifold-based implementations, making it vital for robust, geometry-adapted optimization in diverse applications.

The Bregman proximal point algorithm (BPPA) is a family of proximal regularization methods in which the quadratic proximity term of the classical proximal point algorithm is replaced by a Bregman distance generated by a convex differentiable kernel. A standard form in the cited literature is

$$x_{k+1}\in \arg\min_{x\in C}\left\{r(x)+\frac{1}{\lambda_k}D_h(x,x_k)\right\},$$

while, for monotone inclusions, the same idea is expressed as a resolvent relation

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$

This non-Euclidean reformulation appears in convex minimization, equilibrium problems, primal-dual splitting, augmented Lagrangian methods, stochastic and distributed optimization, and manifold-valued problems (Zhou et al., 2015, Jiang et al., 2022). At the same time, several recent works emphasize that many algorithms described as “Bregman proximal” are not full proximal point methods, but rather proximal-gradient, prox-linear, or model-based variants derived from the BPPA viewpoint (Chraibi et al., 2024, Ochs et al., 2017). The arXiv record for “Proximal algorithms with Bregman distances for bilevel equilibrium problems with application to the problem of ‘how routines form and change’ in Economics and Management Sciences” gives only a placeholder without PDF or source, so its exact formulas and convergence statements are not available from the record itself (Bento et al., 2014).

1. Kernel geometry and the Bregman distance

The Bregman distance used throughout this literature is

$D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$

or, in manifold notation,

$$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$

It is nonnegative under convexity of the generator, but it is generally neither symmetric nor a metric. In the Euclidean quadratic case $h(x)=\tfrac12\|x\|^2$, it reduces to $\tfrac12\|x-y\|^2$; Mahalanobis and entropy geometries furnish other canonical special cases (Zhou et al., 2015, Sharma et al., 20 Jan 2026).

The kernel assumptions are typically phrased in Legendre language. Across the cited work, $h$ or $\phi$ is taken to be proper, closed, convex, differentiable on the interior of its domain, and often essentially smooth and essentially strictly convex. Some papers strengthen this further to 1-coercivity, while others formulate a “Bregman function” with bounded level sets and continuity properties that convert vanishing Bregman distance into actual convergence (Ochs et al., 2017, Wang et al., 2021). In finite-dimensional convex settings, one common device is to choose $\operatorname{dom}\phi=\overline{\mathcal C}$, so that the kernel absorbs feasibility and the iterates remain in $\mathcal C$ automatically (Yang et al., 2021).

Several identities organize BPPA analysis. A central one is the three-point identity,

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$0

which is repeatedly used to derive Fejér-type monotonicity, descent inequalities, and telescoping bounds (Zhou et al., 2015). Closely related four-point identities appear in modern inexact analyses and primal-dual derivations (Yang et al., 2021).

2. Canonical formulations across problem classes

In convex minimization, BPPA appears in its most direct form: proximal regularization of the entire objective by a Bregman term. This is the formulation from which many later variants inherit their structure. In the finite-sum, constrained, or composite settings, the same idea can be lifted to primal-dual inclusions, dual problems, or regularized saddle systems rather than applied to the primal objective alone (Yang et al., 2021, Yan et al., 2020).

In monotone inclusion form, the Bregman proximal step is a non-Euclidean resolvent. This viewpoint is especially important in splitting theory, where the proximal point operator is applied to the KKT system or to a maximal monotone map on a product space. The update

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$1

is the conceptual backbone from which several Bregman primal-dual algorithms are derived (Jiang et al., 2022).

In equilibrium problems on Hadamard manifolds, the same regularization principle acts on a bifunction $$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$2 rather than a scalar objective. One formulation solves

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$3

where

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$4

The target problem is to find $$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$5 such that $$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$6 for all $$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$7, under monotonicity and geodesic convexity assumptions (Sharma et al., 20 Jan 2026).

For constrained variational problems, the latent-variable proximal point framework rewrites BPPA at the continuous level. Starting from

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$8

it introduces the latent variable

$$\nabla \phi(u^{(k)})-\nabla \phi(u^{(k+1)})\in F(u^{(k+1)}).$$9

so that the proximal subproblem becomes the saddle system

$D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$0

This equivalence is exact at the continuous level and is motivated by discretization convenience for pointwise inequality constraints (Dokken et al., 7 Mar 2025).

3. Descent, inexactness, and acceleration

The classical convergence picture is Bregman-Fejér monotonicity. In the convex setting, one obtains monotone decrease of the objective together with monotone decrease of the Bregman distance to the solution set. A particularly clean instance arises from the equivalence between Bregman proximal gradient and BPPA with modified generator $D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$1, which yields

$D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$2

$D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$3

and

$D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$4

Thus the standard convex rate is $D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$5 in function value (Zhou et al., 2015).

Modern work has refined the inexact theory. A notable development is the inexact BPPA condition

$D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$6

Its distinctive feature is the split between the point where the $D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$7-subdifferential of $D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$8 is evaluated and the point where the Bregman gradient term is evaluated. This circumvents the feasibility difficulty that arises when $D_h(x,y)=h(x)-h(y)-\langle \nabla h(y),x-y\rangle,$9 and $$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$0 are awkward to satisfy simultaneously at one approximate iterate (Yang et al., 2021).

Acceleration has been pursued along several lines. The inertial V-iBPPA uses extrapolated points, estimate sequences, and a quadrangle scaling exponent $$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$1, achieving

$$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$2

when the proximal parameter is bounded above and below, and $$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$3 when the proximal parameter is constant and the kernel is strongly convex with Lipschitz continuous gradient (Yang et al., 2021). A distinct accelerated Bregman proximal point construction improves the generic rate from

$$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$4

to

$$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$5

under a triangle-scaling assumption on the Bregman divergence (Yan et al., 2020). In unbalanced optimal transport, an estimate-sequence acceleration yields $$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$6 under the triangle scaling property, although for the pure entropy kernel the theoretical exponent is $$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$7 (Chen et al., 2024).

4. Relationship to proximal-gradient, splitting, and model-based methods

A central structural fact is that Bregman proximal gradient is exactly a Bregman proximal point method with a different generator. For the composite convex problem $$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$8, the update

$$D_\phi(x,y)=\phi(x)-\phi(y)-\langle \operatorname{grad}\phi(y),\exp_y^{-1}x\rangle.$$9

can be rewritten as

$h(x)=\tfrac12\|x\|^2$0

This observation collapses the analysis of Bregman proximal gradient into the analysis of BPPA itself (Zhou et al., 2015).

Primal-dual splitting methods exploit the same viewpoint at the level of KKT inclusions. The Bregman Condat–Vu algorithms are derived by applying BPPA to the primal-dual monotone inclusion for

$h(x)=\tfrac12\|x\|^2$1

with carefully chosen product-space kernels $h(x)=\tfrac12\|x\|^2$2 and $h(x)=\tfrac12\|x\|^2$3. The resulting primal and dual updates are then understood not as ad hoc extensions, but as explicit splittings of a single Bregman proximal point step (Jiang et al., 2022).

The Bregman augmented Lagrangian method is likewise a proximal point method on the dual. For the dual function $h(x)=\tfrac12\|x\|^2$4, the Bregman proximal point step is

$h(x)=\tfrac12\|x\|^2$5

and the corresponding saddle form yields BALM. This proximal-point interpretation extends to accelerated BALM, where improved dual rates translate into improved primal objective and feasibility rates for linearly constrained convex programs (Yan et al., 2020).

A recurrent misconception is that every method with a Bregman proximal subproblem is a BPPA in the strict sense. Several papers explicitly reject that identification. Delay-tolerant distributed Bregman proximal algorithms are “not presented as a pure Bregman proximal point method for $h(x)=\tfrac12\|x\|^2$6”; they are Bregman proximal-gradient methods with a linearization of the smooth part (Chraibi et al., 2024). Likewise, non-smooth non-convex Bregman minimization builds a convex model of the objective, computes an approximate Bregman proximal point of that model, and then performs an Armijo-like line search, so the proximal step acts on a surrogate rather than on the original objective (Ochs et al., 2017).

5. Generalizations beyond Euclidean convex analysis

On Hadamard manifolds, BPPA inherits genuinely geometric difficulties. The Bregman regularization term induced by a Bregman function is, in general, nonconvex on Hadamard manifolds unless the curvature is zero. To recover existence and convergence, the 2026 manifold equilibrium framework imposes the strong assumption that, for each $h(x)=\tfrac12\|x\|^2$7, the set

$h(x)=\tfrac12\|x\|^2$8

is geodesically convex. Under this assumption, together with monotonicity of the bifunction and a weaker coercivity condition than those typically assumed in the earlier literature on Bregman regularization, the generated sequence is well defined, every cluster point solves the equilibrium problem, and the whole sequence converges (Sharma et al., 20 Jan 2026).

Abstract convexity pushes the construction even further. There, one replaces ordinary linear functions by an abstract class $h(x)=\tfrac12\|x\|^2$9, ordinary subgradients by $\tfrac12\|x-y\|^2$0, and the classical Bregman divergence by

$\tfrac12\|x-y\|^2$1

The proximal-point-like iteration becomes

$\tfrac12\|x-y\|^2$2

and the principal convergence statement is

$\tfrac12\|x-y\|^2$3

with cluster points optimal when minimizers exist (Millán et al., 2024).

The latent-variable proximal point algorithm offers a different kind of extension: not a broader notion of convexity, but a continuous-level reparameterization of BPPA for inequality-constrained variational problems. By moving from the feasible image to latent coordinates through $\tfrac12\|x-y\|^2$4 and $\tfrac12\|x-y\|^2$5, it allows discretization in standard linear finite element spaces, preserves feasibility after reconstruction, and exhibits observed mesh-independence across a range of examples (Dokken et al., 7 Mar 2025). A plausible implication is that BPPA is not only a convergence device, but also a discretization interface.

The earlier arXiv notice on bilevel equilibrium problems already located Bregman distances on Hadamard manifolds and under pseudomonotonicity, but the absence of PDF/source prevents a paper-faithful comparison of its update rules with the later 2026 manifold framework (Bento et al., 2014).

6. Applications and empirical roles

Optimal transport has become one of the most visible application areas. In KL-based unbalanced optimal transport, the problem

$\tfrac12\|x-y\|^2$6

is solved by an inexact Bregman proximal point method with entropy kernel $\tfrac12\|x-y\|^2$7. Each outer proximal subproblem is an entropic UOT problem with shifted cost matrix, and the scaling algorithm is used only as an approximate proximal-operator solver. The reported benefits are convergence to the true UOT solution, robust regularization parameter selection, mitigation of numerical stability issues, and practical computational complexity comparable to the scaling algorithm in some regimes (Chen et al., 2024).

The standard discrete OT problem has also served as a testbed for inexact and inertial BPPA. With quadratic kernels, the subproblems become Euclidean projections onto the transport polytope; with entropic kernels, they reduce to Sinkhorn-type subproblems. The principal empirical message is that the split-point inexact criterion is practical for structured feasible sets, and that inertial V-iBPPA can reduce outer iterations and total runtime when the inner tolerances are sufficiently tight (Yang et al., 2021).

In machine learning, BPPA has been analyzed as an implicit regularizer on separable classification problems. For constant stepsize and a distance-generating function $\tfrac12\|x-y\|^2$8 that is both strongly convex and smooth with respect to a chosen norm, BPPA yields an asymptotic margin lower bound

$\tfrac12\|x-y\|^2$9

In the Mahalanobis case $h$0, the condition number is $h$1, and BPPA converges to the maximum $h$2-margin classifier exactly. The dependence on the condition number is shown to be tight, so the divergence choice materially affects classifier quality (Li et al., 2021).

BPPA-type updates also appear as subroutines inside larger alternating frameworks. In robust nonnegative matrix factorization with missing values and outliers, Bregman-proximal point steps with $h$3 are embedded in an augmented Lagrangian / ADMM-like scheme for the $h$4- and $h$5-updates, while $h$6 and $h$7 are updated by quadratic minimization and soft-thresholding. The application to bladder-cancer gene expression analysis uses these Bregman-proximal components to stabilize multiplicative, positivity-preserving factor updates (Chrétien et al., 2015).

Taken together, these developments portray BPPA as a geometric principle rather than a single algorithmic template: exact or inexact, primal or dual, Euclidean or entropy-like, explicit or latent, finite-dimensional or manifold-valued. What remains invariant is the use of a Bregman-generated proximity term to regularize a difficult step in a geometry adapted to the structure of the problem.