Wasserstein Lagrangian Mechanics

• WLM is a variational framework that casts the evolution of probability measures as action-minimizing paths in Wasserstein space, generalizing both classical mechanics and gradient flows.
• The framework employs a Langevin deformation with a damping parameter to rigorously interpolate between geodesic (Hamiltonian) and dissipative (gradient) dynamics.
• WLM underpins advanced computational methods and machine learning algorithms for trajectory inference, physical system identification, and population dynamics.

Wasserstein Lagrangian Mechanics (WLM) is the variational and geometric framework that describes dynamical evolution of probability measures as action-minimizing (critical) paths in optimal transport spaces, primarily the $L^2$-Wasserstein space $\mathcal{P}_2(M)$ over a Riemannian manifold $(M,g)$. By formulating kinetic energy and action principles in Wasserstein geometry, WLM generalizes both classical mechanics and gradient-flow formulations, allowing interpolation between geodesic (Hamiltonian) and dissipative (gradient) dynamics on spaces of densities. The recent extension via Langevin deformation introduces a damping parameter, enabling rigorous interpolation between the geodesic flow and gradient flow, and admits Perelman-type $W$-entropy monotonicity and rigidity results. WLM serves as a unifying foundation for analysis, numerical computation, and machine learning in transport-driven population dynamics, and underpins a class of learning algorithms for trajectory inference and physical system identification.

1. Action Principles and Geodesic Flows in Wasserstein Space

The basis of WLM is the identification of probability measure evolution as the critical paths of a Lagrangian (action) functional defined over the Wasserstein space $\mathcal{P}_2(M)$. For smooth densities $\rho(t,x)\,\mathrm{d}\mu$ ($\mathrm{d}\mu = e^{-f}\,\mathrm{dvol}_g$) and potential $\phi \in C^\infty(M)$, the tangent space structure induced by Otto's formal Riemannian metric takes

$s = -\nabla \cdot(\rho \nabla\phi)$

with kinetic energy

$$\langle s_1, s_2 \rangle_\rho = \int_M \langle \nabla\phi_1, \nabla\phi_2 \rangle \rho\,\mathrm{d}\mu.$$

The corresponding Lagrangian and action are: $\mathcal{P}_2(M)$0 The variational principle with continuity constraint $\mathcal{P}_2(M)$1 leads to the Euler–Lagrange system: $\mathcal{P}_2(M)$2 where the second line is the Hamilton–Jacobi equation. These equations characterize geodesics (displacement interpolations) in $\mathcal{P}_2(M)$3, and provide the optimal transport map between given endpoints $\mathcal{P}_2(M)$4, $\mathcal{P}_2(M)$5 (Li et al., 2016).

2. Langevin Deformation: Damped Eulerian Dynamics and Interpolation

To rigorously interpolate between geodesic flow (purely inertial, Hamiltonian regime) and gradient flow (overdamped, entropy-minimizing regime), the Langevin deformation introduces a friction parameter $\mathcal{P}_2(M)$6: $\mathcal{P}_2(M)$7 equivalently expressed as the damped compressible Euler system for $\mathcal{P}_2(M)$8: $\mathcal{P}_2(M)$9 As $(M,g)$0, one recovers the Wasserstein gradient flow (heat equation), while $(M,g)$1 recovers the geodesic/Hamiltonian evolution. Well-posedness is established in Sobolev spaces for each fixed $(M,g)$2 (Li et al., 2016).

This construction generalizes to $(M,g)$3-Wasserstein spaces on $(M,g)$4, with $(M,g)$5, via suitable change in kinetic energy: $(M,g)$6 admitting analogous Euler–Lagrange equations, Langevin deformations, and PDE well-posedness for $(M,g)$7 (Lei et al., 2 Jun 2025).

3. Entropy Formulas, Monotonicity, and Rigidity: Perelman-type $(M,g)$8-Entropy

WLM admits a variational entropy structure reflecting the Perelman $(M,g)$9-entropy formalism. For $W$0 equipped with Bakry–Émery curvature, the Boltzmann–Shannon entropy is

$W$1

The geodesic flow satisfies a $W$2-entropy monotonicity, with

$W$3

and dissipation inequality

$W$4

provided non-negative Bakry–Émery curvature. The Langevin deformation admits corresponding time-dependent $W$5 functionals, establishing comparison results to the explicit time-dependent Gaussian model and ensuring monotonicity (Li et al., 2016). In $W$6-Wasserstein space, the $W$7-entropy formula generalizes with similar monotonicity and rigidity properties (Lei et al., 2 Jun 2025).

Rigidity theorems specify that equalities in entropy-dissipation force the underlying manifold to be isometric to Euclidean space, $W$8, $W$9, and the solution path coincides with the Gaussian model (or its time-dependent analogue in the Langevin deformation) (Li et al., 2016, Lei et al., 2 Jun 2025).

4. Computational and Algorithmic Frameworks

The action-based WLM formulation enables a unified computational approach for solving variational transport problems, including geodesic flow, Schrödinger bridge, unbalanced transport, and physically augmented flows. The primal problem seeks action-minimizing density paths under velocity and potential terms; the dual problem exploits the Legendre transform with respect to the time-dependent potential, resulting in a minimax (saddle-point) optimization over densities and potentials.

Recent deep learning algorithms parameterize the potential as a neural network, sample interpolated trajectories, and optimize a loss derived from the dual variational objective. Marginal constraints are enforced via pointwise coupling, and Monte Carlo integration replaces analytical computations. This framework accommodates a wide range of transport and physics-driven scenarios, including unbalanced transport and interaction potentials, with empirical performance validated in single-cell trajectory inference and other domains (Neklyudov et al., 2023).

A recent class of learning algorithms for WLM (e.g., population mechanics) directly parameterizes the potential or action via neural networks, integrates N-particle ODEs under the Hamiltonian/Lagrangian evolution, and minimizes a metric (e.g., Sinkhorn divergence) between simulated and observed marginals. These methods employ symplectic integrators for stability and utilize gradient-based training via backpropagation through the trajectory integration steps (Guan et al., 8 May 2026).

5. Extensions: Population Learning and Second-Order Population Dynamics

WLM offers a structured model for collective phenomena, unifying classical, quantum, and gradient-flow mechanics on measure spaces. The WLM approach departs from first-order gradient flows by explicitly representing inertial effects and dissipation/interpolation via the parameter $\mathcal{P}_2(M)$0 or $\mathcal{P}_2(M)$1. The Euler–Lagrange system in WLM for density $\mathcal{P}_2(M)$2 and Kantorovich potential $\mathcal{P}_2(M)$3 takes the form: $\mathcal{P}_2(M)$4 where $\mathcal{P}_2(M)$5 can encode external potentials, interaction energies, or entropy (Guan et al., 8 May 2026). The flexibility to express both conservative (Hamiltonian, periodic) and dissipative (diffusive) regimes enables faithful modeling of phenomena ranging from fluid vortices and flocking to embryonic development and chemical reaction networks.

Empirical benchmarks confirm that WLM-based learning methods outperform classical gradient-flow and flow-matching baselines in interpolation and forecasting tasks where oscillatory or conservative features predominate. Notably, WLM recovers agent-level interaction structure (e.g., repulsion-alignment-cohesion in boid flocking) directly from population marginals (Guan et al., 8 May 2026).

6. Discretization Schemes and Numerical Methods

Practical computation in WLM is facilitated by time-space discretizations formulated in Lagrangian coordinates. The Benamou–Brenier action provides a continuous variational principle for Wasserstein geodesics, and regularized Lagrangian flow maps $\mathcal{P}_2(M)$6 allow for unconstrained minimization in $\mathcal{P}_2(M)$7 of particle positions. Discrete schemes preserve mass, positivity, and energy dissipation, and are robust across a range of nonlinear flows, including porous medium, aggregation, and chemotaxis models (Cheng et al., 2024).

Lagrangian schemes, by operating directly on flow maps, outperform or complement Eulerian (grid-based) methods in capturing sharp interfaces and enforcing conservative properties by construction. These schemes can be adapted to arbitrary internal energies and external potentials, making them well-suited for efficient time stepping in WLM-driven PDEs and for use in learning-based simulation pipelines.

7. Generalizations and Theoretical Developments

The WLM paradigm extends to $\mathcal{P}_2(M)$8-Wasserstein spaces ($\mathcal{P}_2(M)$9), expanding the flexibility of variational population mechanics to encompass $\rho(t,x)\,\mathrm{d}\mu$0-Laplacian flows and higher-order transport regimes. The associated entropy formulas, damped Eulerian systems, and rigidity results all have precise analogues, forming a broad unifying framework for geometric evolution equations and collective dynamics on measure spaces (Lei et al., 2 Jun 2025).

Ongoing developments address existence, uniqueness, and regularity of the Langevin deformation, as well as convergence to gradient-flow or geodesic regimes under parameter limits. The WLM framework thus bridges optimal transport, geometric analysis, and emerging machine learning approaches to population-level mechanics.

---

Key References:

• "W-entropy formulas and Langevin deformation of flows on Wasserstein space over Riemannian manifolds" (Li et al., 2016)
• "W-entropy formulas and Langevin deformation on the $\rho(t,x)\,\mathrm{d}\mu$1-Wasserstein space over Riemannian manifolds" (Lei et al., 2 Jun 2025)
• "A Computational Framework for Solving Wasserstein Lagrangian Flows" (Neklyudov et al., 2023)
• "A Call to Lagrangian Action: Learning Population Mechanics from Temporal Snapshots" (Guan et al., 8 May 2026)
• "A new flow dynamic approach for Wasserstein gradient flows" (Cheng et al., 2024)