Generalized Gradient Flows

• Generalized gradient flows are defined as variational evolutions in abstract state spaces driven by an energy functional and non-quadratic, convex dissipation potentials.
• They enable robust energy-dissipation formulations and structure-preserving time-discrete schemes, ensuring stability even in non-smooth or degenerate cases.
• They find applications in mathematical physics, stochastic processes, and optimization, underpinning analyses in irreversible thermodynamics and variational coarse-graining.

A generalized gradient flow is a variational evolution framework that extends classical gradient flow theory to include non-quadratic, degenerate, or non-metric dissipation, non-Hilbertian state spaces, multi-component energetics, nonlocal interactions, or mixed conservative–dissipative dynamics. The theory underpins the modern analysis of dissipative partial differential equations (PDEs), stochastic processes, variational numerical schemes, and numerous applications in mathematical physics, probability, and optimization.

1. Abstract Definition and Duality Structures

In the most general context, an evolution is viewed as a curve $u:[0,T]\to\mathcal{X}$ in a state space $\mathcal{X}$ (e.g., Banach, metric, or probability spaces), driven by an energy functional $\mathcal{E}:\mathcal{X}\to\mathbb{R}\cup\{\infty\}$ and a family of convex dissipation potentials $\Psi(u,\cdot)$ and their convex duals $\Psi^*(u,\cdot)$. The evolution is governed by the inclusion

$$\partial_t u + \partial_j \Psi^*\big(u, -\nabla \mathcal{E}(u)\big) = 0$$

or, in "flux-form",

$$\partial_t u + j = 0,\quad j\in \partial\Psi^*\big(u, -\nabla \mathcal{E}(u)\big).$$

Here, $\Psi^*(u,\xi) = \sup_{j}\, [j\cdot\xi - \Psi(u,j)]$, and the notion of "gradient" may need to be interpreted in a weak sense in abstract spaces. Dissipative effects arise from $\Psi$, while conservative or Hamiltonian parts can be encoded as anti-symmetric, null-cost directions in $\Psi^*$ (Duong et al., 2015).

This duality structure is closely linked to large-deviation rate functionals of particle systems and underpins the variational formulation of the associated deterministic evolution.

2. Energy-Dissipation Principles and Variational Formulations

Generalized gradient flows are characterized by an energy-dissipation balance of the form

$\mathcal{X}$0

or, in cases with exact solutions,

$\mathcal{X}$1

This balance generalizes the classical gradient flow's energy-dissipation and provides a foundation for solution concepts, compactness tools, well-posedness, and numerical schemes (Jüngel et al., 2018, Ruf, 2024).

In abstract metric spaces, the corresponding notion is the curve of maximal slope, satisfying

$\mathcal{X}$2

Here, $\mathcal{X}$3 is the metric derivative and $\mathcal{X}$4 the local slope, leading to the celebrated Evolution Variational Inequality (EVI) theory (Santambrogio, 2016, Fleißner, 2017).

3. Generalized Dissipations Beyond Quadratic Norms

Unlike classical Hilbert gradient flows—which are defined via $\mathcal{X}$5—generalized gradient flows admit arbitrary, typically convex but possibly degenerate, $\mathcal{X}$6-dependent dissipation potentials. This includes:

• Nonlinear and degenerate mobilities (e.g., porous medium and fast-diffusion equations (Fu et al., 2023, Li et al., 2019))
• Orlicz-growth dissipation (arbitrary $\mathcal{X}$7-growth or even non-reflexive Banach setting (Ruf, 2024))
• Non-quadratic, non-metric jump processes and nonlocal interactions (Peletier et al., 2020, Hoeksema et al., 23 Sep 2025)

The Fenchel duals $\mathcal{X}$8 may encode entropic, large-deviation, or rate-independent (infinite-value or nonsmooth) dissipation.

In the framework of jump processes, generalized gradient flows are formulated for spaces of nonnegative measures, equipped with dissipation functionals built from dual "cosh-type" or other strictly convex cost structures. These do not admit an underlying metric structure but are characterized by energy-dissipation functionals on measure-flux pairs (Peletier et al., 2020, Hoeksema et al., 23 Sep 2025).

4. Generalized Gradient Flows in Metric and Probability Spaces

The metric theory, foundationally developed for Wasserstein spaces and optimal transport, generalizes further to:

• Gradient flows in spaces of probability laws endowed with generalized transport distances (e.g., $\mathcal{X}$9-Wasserstein, entropic, or weighted distances) (Fu et al., 2023, Zeng et al., 18 Sep 2025, Li et al., 2019)
• Flows of $\mathcal{E}:\mathcal{X}\to\mathbb{R}\cup\{\infty\}$0-divergences under non-standard geodesic metrics derived from the Hessian transport of entropy functionals (Li et al., 2019)
• Variational "diffusive limits" for random walks, tessellations, and large interacting particle systems, which reveal continuum PDEs as macroscopic generalized gradient flows (Hraivoronska et al., 2022)

The generalized gradient-flow structure is central to the variational coarse-graining of stochastic and deterministic systems, where a duality-based $\mathcal{E}:\mathcal{X}\to\mathbb{R}\cup\{\infty\}$1-convergence approach passes from microscopic or more detailed dynamics to reduced or effective equations in a fully unified manner (Duong et al., 2015).

The minimization movement or JKO (Jordan-Kinderlehrer-Otto) scheme,

$\mathcal{E}:\mathcal{X}\to\mathbb{R}\cup\{\infty\}$2

captures the time-discrete incremental principle for nonlinear, nonquadratic, or spatially inhomogeneous dissipation (Santambrogio, 2016, Fu et al., 2023, Zeng et al., 18 Sep 2025).

5. Analytical and Numerical Methodologies

The abstract theory provides the justification for robust time-discretizations:

• Structure-preserving time-discrete schemes such as minimizing movements, De Giorgi and Gonzalez-type discrete gradients, and alternating minimization for combined dissipation mechanisms (Jüngel et al., 2018, Mielke et al., 2023, 1908.10246)
• High-order, unconditionally stable, variational extrapolation algorithms for general gradient flows, directly generalizing minimizing-movement methods to higher temporal accuracy (1908.10246)
• Augmented Lagrangian and primal-dual splitting methods for generalized Wasserstein and reaction-diffusion flows, including effective numerical schemes for high-dimensional or multi-species systems (Fu et al., 2023, Zeng et al., 18 Sep 2025)

This structure allows consistent, stable approximations even for nonquadratic or degenerate flows, as in reaction-diffusion, aggregation-diffusion, or jump process models.

Numerical analysis and convergence rely crucially on the preservation of the underlying energy-dissipation and variational framework, ensuring stability and robust performance even in non-smooth or nonconvex cases.

6. Applications Across Mathematical Physics and Analysis

Generalized gradient flows are pervasive in:

• Stochastic processes (diffusions and jump processes), in particular in large-deviation rate-functional derivations of hydrodynamic and scaling limits (Duong et al., 2015, Hoeksema et al., 23 Sep 2025, Peletier et al., 2020)
• Irreversible thermodynamics and Onsager–Machlup–GENERIC systems, encoding both dissipative and conservative evolution (Jüngel et al., 2018)
• Quantum field theory, supersymmetry, and nonlinear sigma models (including generalized gradient flows for nonlinearly realized symmetries and supersymmetric gauge invariants) (Aoki et al., 2015)
• Metric measure geometry, heat flow on non-smooth spaces, and the synthetic analysis of Ricci curvature lower bounds (RCD theory) (Santambrogio, 2016, Fleißner, 2017, Fleißner, 2017)
• Convex and geometric flows on domains/shapes, e.g. gradient flows for Laplacian eigenvalue optimization (Holle, 2020)

The framework is robust to degenerate, non-coercive, or infinity-valued dissipation, permits non-monotone subdifferential structures, and applies in non-reflexive Banach and Orlicz-type function spaces, thus encompassing a vast range of PDEs and evolution problems (Ruf, 2024).

7. Recent Developments and Directions

Ongoing advances include:

• The extension of generalized gradient-flow theory to singular nonlocal jump kernels, with new classes of reflecting solutions and refined compactness-stability results (Hoeksema et al., 23 Sep 2025)
• Acceleration and optimization: fixed-time convergent generalized gradient flows with element-wise normalization, including momentum extensions and saddle-escape guarantees for machine learning and convex optimization (Baranwal et al., 2022)
• Coupled or multi-dissipation gradient flows using time-splitting methods and inf-convolution of dual dissipations, with applications to multi-physics systems and visco-elasto-plastic evolution (Mielke et al., 2023)
• Variational coarse-graining methods and $\mathcal{E}:\mathcal{X}\to\mathbb{R}\cup\{\infty\}$3-limit techniques for model reduction in kinetic and stochastic systems (Duong et al., 2015)

This synthesis of variational, geometric, probabilistic, and metric approaches continues to inform the design of algorithms, the analysis of complex evolutions, and the mathematical understanding of dissipation in broad scientific contexts.