Mirror Flow & Bregman Potentials

• Mirror Flow and Bregman Potentials are techniques that leverage convex geometry to define continuous optimization dynamics and convergence properties.
• They underpin both discrete and continuous algorithms in areas like policy learning, variational inference, and generative modeling by adapting to specific constraints.
• The approach ensures stability and convergence by dynamically encoding geometric structure through Bregman divergences and tailored potential functions.

Mirror flow refers to the continuous-time limit of mirror descent—a fundamental optimization paradigm wherein first-order updates are preconditioned by the geometry of a convex potential, known as a Bregman potential. The concept and machinery of Bregman potentials and divergences underpin a broad class of discrete and continuous methods in convex optimization, variational inference, stochastic control, policy learning, sampling, and generative modeling. The choice of Bregman potential directly encodes the geometric structure, induces the flow dynamics, and determines both convergence and implicit bias of these algorithms.

1. Bregman Potentials, Divergences, and Mirror Flows

A Bregman potential (also called a mirror or Legendre potential) is a strictly convex, twice continuously differentiable function $\psi: \mathcal{X} \to \mathbb{R}$ defined on a convex set $\mathcal{X}\subseteq \mathbb{R}^d$. The associated Bregman divergence is

$$D_\psi(u, v) = \psi(u) - \psi(v) - \langle \nabla\psi(v),\,u - v \rangle,$$

which, due to convexity, is non-negative and vanishes if and only if $u = v$ (Sethi et al., 3 Jun 2025, CHA et al., 23 Oct 2025). The Bregman divergence is generally asymmetric and induces a non-Euclidean geometric structure on $\mathcal{X}$.

Mirror flow is the continuous-time dynamical system arising as the limit of mirror descent when step sizes vanish: $$\frac{dx(t)}{dt} = -\left[\nabla^2\psi(x(t))\right]^{-1}\nabla f(x(t))$$ with objective $f$, or equivalently $d/dt\, \nabla\psi(x(t)) = -\nabla f(x(t))$ (Tzen et al., 2023, CHA et al., 23 Oct 2025, Liang et al., 2024). The mapping $\nabla\psi$ is the mirror (or primal-dual) map, and $\psi^*$ is its Legendre dual, satisfying $\nabla\psi^* = (\nabla\psi)^{-1}$.

Choice of potential $\psi$ influences the flow's geometry: quadratic potentials generate Euclidean flows, entropic potentials produce simplex-respecting dynamics, and more general Bregman potentials adapt to specific constraints or anisotropies.

2. Discrete and Continuous Mirror-Based Algorithms

Discrete-time mirror descent updates are formulated by solving, at each step,

$$x^{(k+1)} = \arg\min_{x\in\mathcal{X}} \left\{\eta_k\langle \nabla f(x^{(k)}), x\rangle + D_\psi(x, x^{(k)})\right\}.$$

The induced first-order optimality condition is

$$\nabla\psi(x^{(k)}) - \nabla\psi(x^{(k+1)}) = \eta_k\nabla f(x^{(k)}) + n^{(k+1)}$$

with $n^{(k+1)}$ a normal vector to the feasible set at $x^{(k+1)}$ (Lin et al., 2022). In policy mirror descent (PMD) for Markov decision processes, the policy update is characterized analogously, with step sizes and Bregman divergence regularization governing the convergence regime.

In the continuous-time limit, these updates converge to mirror flow ODEs, providing a unifying geometric perspective (Tzen et al., 2023, CHA et al., 23 Oct 2025). The evolution variational inequality (EVI) formalism yields, for any reference $y$,

$$\frac{d}{dt} D_\psi(x(t) \| y) + f(x(t)) - f(y) \leq 0,$$

ensuring Lyapunov-type monotonicity of Bregman divergence along optimally designed mirror flows (CHA et al., 23 Oct 2025).

3. Convergence Theory: Rates and Stability

Convergence rates of mirror flow are fundamentally determined by the strong convexity of $\psi$ and the (possibly relative) convexity of the objective $f$ (CHA et al., 23 Oct 2025, Sethi et al., 3 Jun 2025, Lin et al., 2022):

• For merely convex objectives, objective suboptimality and Bregman divergence decay as $O(1/t)$ in continuous time, $O(1/k)$ in discrete time (Tzen et al., 2023, CHA et al., 23 Oct 2025, Lin et al., 2022).
• When $f$ is strongly convex relative to $\psi$, exponential (linear) convergence in Bregman divergence is obtained:

$$D_\psi(x(t)\|x^*) \leq e^{-\kappa t} D_\psi(x(0) \| x^*)$$

with contractivity factor $\kappa = \mu/(\mu + L)$ and $\mu$ as the convexity modulus of $\psi$ (CHA et al., 23 Oct 2025).

• In policy mirror descent, sufficient step size and advantage gap guarantee finite-step support recovery: suboptimal actions are pruned in finite iterations depending on the threshold imposed by $\nabla\psi$ (Lin et al., 2022).
• For constrained stochastic control, exponential convergence follows if the stagewise Hamiltonian is strongly convex in actions relative to $D_\psi$ (Sethi et al., 3 Jun 2025).

Table: Potentials and Resulting Geometry in Mirror Flows

Bregman Potential	Geometry/Flow	Special Property
$\psi(x) = \frac{1}{2}\|x\|^2$	Euclidean gradient flow	Reduces to vanilla SGD/grad descent
$\psi(x) = \sum x_i\log x_i$	Simplex, entropic mirror flow	Yields exponentiated gradient steps
$\psi(x) = \|x\|_p^p/p$	$\ell_p$ balls/normed space	Induces sparsity ($p=1$)

4. Implicit Bias and Asymptotic Behavior

In separable classification, the direction of convergence for mirror flow is determined not by the local behavior of $\psi$, but by its "horizon function" $\phi_\infty$, which captures the linear asymptotic growth rate: $$\phi_\infty(u) = \lim_{\alpha\to\infty} \psi(\alpha u)/\alpha.$$ As $t \to \infty$, $w(t)$ converges in direction to the unique $\phi_\infty$-max-margin vector: $$w^* = \arg\min \{\phi_\infty(w) : \min_i y_i\langle x_i, w \rangle \geq 1\}$$ which generalizes classical $L_2$- or $L_1$-maximum margins to any Bregman geometry. This maximizer characterizes the implicit bias of the mirror flow dynamics, explicitly connecting geometry to margin (Pesme et al., 2024).

In neural network optimization, mirror flows with "unscaled" separable potentials remain in the kernel regime and exhibit lazy training, yielding the same RKHS-biased solution as standard gradient flow in the infinite-width limit (Liang et al., 2024). Scaled potentials permit richer, non-RKHS bias described by Bregman-divergence penalties on network curvature.

5. Extensions: Nonsmoothness, Constraints, and Stochastic Flows

Bregman potentials enable generalizations beyond smooth unconstrained optimization:

• Constrained Optimization: Bregman mirror flows operate efficiently on convex sets, handle constraints through the geometry of $\psi$, and can incorporate Bregman damping for additional stability in distributed settings (Chen et al., 2021).
• Nonsmooth Objectives: The Bregman–Moreau envelope constructs smooth surrogates of nonsmooth $g$, using generalized proximity operators:

$$g^\lambda_\psi(x) = \inf_{u} \{g(u) + (1/\lambda) D_\psi(u, x)\}$$

enabling mirror-proximal updates and Bregman-proximal Langevin sampling (Lau et al., 2022).

• Stochastic and Langevin Dynamics: Mirror-Langevin diffusions exploit the local geometry of $\psi$, yielding preconditioned stochastic samplers, with gradient and noise scaled by $(\nabla^2\psi)^{-1}$ (Tzen et al., 2023, Lau et al., 2022).

6. Generative Modeling, Flow Matching, and Constrained Synthesis

Mirror flow concepts underpin recent advances in generative modeling on constrained domains. Regularized mirror maps guarantee well-posed primal–dual flows, controlling the geometry and tail behavior of dual representations. Mirror flow matching (MFM) uses a dual coordinate linear interpolation and learns optimal velocity fields in the Bregman-induced geometry, providing convergence guarantees and enjoying stable training even under heavy-tailed targets via Student-t priors (Guan et al., 10 Oct 2025). These frameworks yield state-of-the-art sample efficiency and rigorous feasibility when synthesizing data subject to convex constraints.

7. Synthesis: Geometric Control, Design, and Outlook

The choice and analysis of Bregman potentials provide a unifying framework for designing and analyzing optimization and learning algorithms suited to complex geometries, constraints, and task-specific biases. Mirror flows interpolate between gradient flows and natural/geometric-descents, with discrete analogues (mirror descent, PMD, etc.) inheriting the stability, convergence, and implicit regularization properties. Systematic selection of $\psi$ enables adaptation to domain constraints, promotes desired structures (e.g., sparsity vs. uniformity), and facilitates both analytical and empirical performance guarantees (Lin et al., 2022, CHA et al., 23 Oct 2025, Sethi et al., 3 Jun 2025, Guan et al., 10 Oct 2025, Liang et al., 2024).

A plausible implication is that continued exploration of non-Euclidean, data-adaptive Bregman potentials will enable further advances in robustness, implicit bias control, and geometric adaptation in both optimization and generative modeling.