Bregman Geometry: Foundations & Applications

• Bregman geometry is a differential-geometric framework that uses asymmetric Bregman divergences to establish dual flat coordinate systems for efficient computation.
• It generalizes traditional metric structures by extending Euclidean and KL divergences, providing a unified approach for statistical, optimization, and machine learning applications.
• Its algorithmic constructs, including Bregman projections and mirror descent, underpin scalable methods in clustering, inference, and manifold learning.

Bregman geometry is the differential-geometric framework arising from the structure induced by Bregman divergences, a rich class of asymmetric distance-like functionals that extend squared Euclidean distance, Kullback–Leibler divergence, and numerous others. This geometry underpins dually flat statistical manifolds, forms the core of information geometry, generalizes metric and Riemannian structures, and supports a wide range of computational and inferential methodologies in optimization, statistics, and machine learning.

1. Foundations: Bregman Divergence and Dually Flat Geometry

A Bregman divergence is specified by a strictly convex, differentiable generator function $F: \Omega \to \mathbb{R}$ (with $\Omega \subset \mathbb{R}^d$ open, convex):

$$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$

Key properties include non-negativity ($D_F(p \| q) \geq 0$ with equality iff $p = q$), convexity in the first argument, and in general absence of symmetry or the triangle inequality (Pham et al., 9 Apr 2025, 0709.2196). The most prominent examples are:

Generator $F(x)$	Bregman Divergence $D_F(x \| y)$	Manifold/Application
$\frac{1}{2}\|x\|^2$	$\frac{1}{2}\|x-y\|^2$	Euclidean geometry
$\sum_{i} x_i\log x_i$	$\Omega \subset \mathbb{R}^d$0	KL-divergence, probability simplex
$\Omega \subset \mathbb{R}^d$1	$\Omega \subset \mathbb{R}^d$2 (Itakura–Saito)	Positive orthant, inverse problems

Bregman divergences naturally induce a pair of dual coordinate systems via Legendre–Fenchel transformation:

$\Omega \subset \mathbb{R}^d$3

so that $\Omega \subset \mathbb{R}^d$4 (primal/dual coordinate) and $\Omega \subset \mathbb{R}^d$5 (Nielsen et al., 2024). The primal chart is flat for the affine connection $\Omega \subset \mathbb{R}^d$6, while the dual chart is flat for $\Omega \subset \mathbb{R}^d$7. This “dually flat” property is the hallmark of Bregman geometry and lends itself to efficient computations, projection theorems, and tractable statistical modeling (Cho et al., 19 Jun 2026).

2. Metric, Connections, and the Pythagorean Theorem

The Hessian metric on $\Omega \subset \mathbb{R}^d$8 is defined as $\Omega \subset \mathbb{R}^d$9, i.e., the local squared length of a tangent vector is measured by the Hessian. Geodesics are straight lines in the $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$0 (primal) coordinates, while dual geodesics are straight lines in the $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$1 (dual) coordinates. In matrix settings (e.g., PD$$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$2), these generalize to affine lines in the space of positive-definite matrices and their Legendre images (Kanamori et al., 2010).

Bregman projections onto affine submanifolds minimize divergence and satisfy the generalized Pythagorean theorem:

$$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$3

where $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$4 is the Bregman projection of $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$5 onto the convex set $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$6 (Pham et al., 9 Apr 2025). Such decompositions are foundational for statistical inference (e.g., maximum likelihood via e-projection), mirror descent, and information projection methods (Cho et al., 19 Jun 2026).

3. Geometric and Algorithmic Structures

Canonical geometric objects in Bregman geometry include:

• Bregman balls: For $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$7, $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$8; convex in $$D_F(p \, \|\, q) = F(p) - F(q) - \langle \nabla F(q),\, p - q \rangle.$$9 (Pham et al., 9 Apr 2025).
• Voronoi diagrams: Cells defined as $D_F(p \| q) \geq 0$0 partition space into convex polytopes in the dual chart (0709.2196). Dual and symmetrized variants further extend this structure.
• Bregman triangulations: Dual to Voronoi diagrams, with both straight-edge (lifted) and geodesic (curved-edge) versions; simplices are circumscribed by unique Bregman spheres (0709.2196).

Algorithms that leverage these constructs:

• Clustering: Lloyd's algorithm generalizes to Bregman divergences; clusters update via arithmetic means in the relevant chart (Pham et al., 9 Apr 2025, Gomes-Gonçalves et al., 2018).
• Bregman Hausdorff divergence: Extends set-to-set comparison using Bregman rather than metric balls; asymmetric, with left/right variants and Chernoff symmetrization strategies (Pham et al., 9 Apr 2025).

The overall computational tractability of Bregman geometric constructs follows from their affine or projective geometric realization, leveraging isometries, embedding techniques, and reduction to half-space or power-diagram computations (0709.2196, Gomes-Gonçalves et al., 2018).

4. Statistical and Optimization Applications

Bregman geometry is natural for:

• Statistical learning and inference: Flatness translates to tractable projections (e.g., e-projection = maximum likelihood for exponential families, m-projection = moment matching (Cho et al., 19 Jun 2026)). Expectation-Maximization, Variational Inference, and Expectation Propagation correspond to alternating Bregman projections along e- and m-flat submanifolds (Hayashi, 2022, Cho et al., 19 Jun 2026).
• Quasi-Newton methods: Bregman divergence on PD$D_F(p \| q) \geq 0$1 generalizes Hessian update formulas (BFGS/DFP) as m- and e-projections, with extension to self-scaling updates and invariance under affine transformations (Kanamori et al., 2010).
• Proximal and mirror descent: Each step is a Bregman projection in a chosen geometry; convergence rates, stability, and adaptation to non-Euclidean and Banach settings are governed by properties of the Bregman generator (CHA et al., 23 Oct 2025, Zhang et al., 17 Sep 2025, Azizian et al., 2022).
• Matrix and manifold information geometry: Total Bregman divergences define robust, closed-form means and statistical detectors for positive-definite matrices, with algorithmic advantages over affine-invariant Riemannian geometry (Hua et al., 2020).

The explicit Pythagorean decomposition and generalized orthogonality drive estimation, calibration, and robust optimization methods throughout modern statistics and machine learning (Cho et al., 19 Jun 2026, Hayashi, 2022).

5. Extensions: Beyond Riemannian Flatness and Generalized Frameworks

Several lines extend classical Bregman geometry:

• Curved generalizations: Logarithmic $D_F(p \| q) \geq 0$2-divergence interpolates between Bregman (flat) and constant-negative-curvature statistical manifolds, yielding conformal deformation and curvature-controlled departures from dual flatness (Wong et al., 2019).
• Banach–Bregman geometry: Generalizes to infinite-dimensional and non-Hilbert spaces, replacing inner products by duality pairings. Enables mirror descent, stochastic approximation, and adaptive methods on simplices, sparse domains, or natural parameter spaces with super-relaxation and variance reduction (Zhang et al., 17 Sep 2025).
• Symplectic Bregman divergences: Generalization to vector spaces equipped with a symplectic form, unifying convex and symplectic geometry, with implications for dissipative mechanics, contrast functions, and learning dynamics (Nielsen, 2024).
• Optimal transport and statistical manifolds: The Bregman–Wasserstein divergence defines geodesics, barycenters, and displacement interpolation in probability spaces by transporting using Bregman geometry rather than squared Euclidean cost, leading to new statistical manifold structures and optimization methods for gradient flows (Kainth et al., 2023).
• Contact geometry in optimization: Continuous-time accelerated optimization flows and contact Hamiltonian discretizations leverage Bregman geometry for invariance and structure-preserving integrators (Bravetti et al., 2019).

These extensions preserve the algebraic skeleton of duality and projection while enabling new geometric, algorithmic, and inferential tools.

6. Statistical Centroids, Barycenters, and Implementation

The theory of centroids and means in Bregman geometry exhibits a clear separation due to asymmetry:

• The right-sided centroid minimizes average divergence from data to a point (arithmetic mean in primal coordinates); the left-sided centroid (arithmetic mean in dual coordinates, pulled back via the Legendre map) (0711.3242).
• The symmetrized centroid (Jensen–Shannon or Burbea–Rao) is obtained as the point equidistant (in a symmetrized sense) between the right- and left-sided centroids, computed efficiently via geodesic-walk/bisection algorithms (0711.3242).
• In separable settings, explicit isometries map the Bregman geometry to Euclidean space; clustering and quantization algorithms exploit this for efficient partitioning, codebook update, and accuracy (Gomes-Gonçalves et al., 2018).
• The pyBregMan Python library implements Bregman manifold operations, dual potentials, geometric objects (balls, bisectors), and algorithms (mirror descent, barycenters, statistics), facilitating rapid prototyping and deployment in both statistical and machine learning domains (Nielsen et al., 2024).

7. Synthesis and Ongoing Directions

Bregman geometry provides the canonical analytic and algorithmic infrastructure for non-Euclidean, information-geometric, and optimization-theoretic settings, accommodating divergence, dual coordinate systems, efficient projections, and tractable structure even on high-dimensional manifolds. Its generalization to nonflat, Banach, symplectic, and manifold settings continues to yield new foundational results and scalable algorithms across fields, including matrix learning, optimal transport, variational and Bayesian inference, and large-scale learning systems.

The dichotomy between flat Bregman (dually affine) structure and modifications incorporating curvature, symplecticity, or non-quadratic geometry determines both computational tractability and statistical/convergence behavior, as seen in the classification of convergence rates by Legendre exponent (Azizian et al., 2022), the deployment of new split Gibbs sampling for Poisson inverse problems (Faye et al., 15 Nov 2025), and the emergence of robust matrix detectors and Bregman–Hausdorff set distances in machine learning (Hua et al., 2020, Pham et al., 9 Apr 2025). The Bregman geometric paradigm thus remains central to theoretical, computational, and applied developments in modern data sciences.