Optimal-Transport Kinetic Loss

• The paper introduces twisted quadratic cost functions that couple spatial and velocity components to improve convergence in kinetic systems.
• It details a methodology combining optimal transport with kinetic dissipation, yielding exponential contraction rates and robust hypocoercivity.
• Practical implementations leverage sample-based computations and matrix regularization to facilitate deep learning and PDE analysis.

An optimal-transport-inspired kinetic loss refers to a family of loss functionals designed for systems with kinetic (second-order, phase-space or velocity-dependent) structure, in which optimal transport (OT) theory dictates the geometry of probability densities, and the associated transport cost is sensitive to dynamics such as velocity, acceleration, or dissipation. These losses capture hypocoercive effects, interpolation between distributions, and convergence rates in settings relevant to kinetic equations, generative models, and machine learning. This entry surveys canonical constructions, analytic properties, and practical algorithms centered on kinetic OT losses.

1. Twisted Quadratic Cost and Kinetic Wasserstein Metrics

The prototypical kinetic loss arises from replacing the standard quadratic ground cost in $\mathbb{R}^d$ with a phase-space cost adapted to kinetic equations. For points $z = (x, v) \in \mathbb{R}^2$, fix $$A = \begin{pmatrix} a & b \ b & c \end{pmatrix} \succ 0$$ and define

$$c\big((x, v), (x', v')\big) = \left\|(x, v) - (x', v')\right\|^2_A = (x - x', v - v')^\top A (x - x', v - v')\,.$$

The corresponding Wasserstein-type metric is

$$W_A^2(\mu, \nu) = \inf_{\gamma \in \Pi(\mu, \nu)} \int_{\mathbb{R}^2 \times \mathbb{R}^2} \left\|z_1 - z_2\right\|_A^2 \, \gamma(dz_1, dz_2)\,,$$

where $\Pi(\mu, \nu)$ is the set of couplings of $\mu$ and $\nu$. $W_A$ is metrically equivalent to the classical $W_2$ via

$$\nu_{\min} W^2_2(\mu, \nu) \leq W_A^2(\mu, \nu) \leq \nu_{\max} W^2_2(\mu, \nu)\,,$$

with $\nu_{\min}, \nu_{\max}$ as the minimum and maximum eigenvalues of $A$ (Salem, 2021).

This "twisted" quadratic cost naturally encodes both spatial and velocity scales, and is optimally adapted to the kinetic Fokker–Planck flow.

2. Kinetic Dissipation and Loss Formulation

In kinetic models, notably the 1D kinetic Fokker–Planck equation,

$$\partial_t f_t + v\,\partial_x f_t - \partial_v\left(U'(x) f_t\right) = \partial_v\left(v f_t + \partial_v f_t\right)\,,$$

the relevant energy dissipation is not the entropy alone but a functional that couples transport and diffusion. Consider the convex potential $\varphi_t$ generating the optimal map from equilibrium $f_\infty$ to $f_t$ in $W_A$, and define the kinetic dissipation functional

$$J_A(f_t \, | \, f_\infty) = -\int_{\mathbb{R}^2}\langle B(T(z)) - B(z), T(z) - z \rangle_A f_\infty(z)\,dz + \ldots,$$

where $$B(x,v) = \begin{pmatrix} v \ -U'(x) - v \end{pmatrix}$$ and $T = A^{-1}\nabla\varphi_t$ (Salem, 2021).

The kinetic loss functional is then: $$\mathcal{L}_{\mathrm{kinetic}}(f) = W_A^2(f, f_\infty)\,,$$ or, in its dissipation variant,

$$\mathcal{L}_{\mathrm{diss}}(f) = J_A(f \,|\, f_\infty)\,.$$

Minimizing either functional enforces hypocoercive convergence to equilibrium, with the rate governed by spectral properties of $A$ and the coercivity constant $\kappa$ (see Section 5).

3. Analytic Properties and Coercivity

The core analytic property distinguishing kinetic OT losses is the direct, sharp contraction estimate along the kinetic flow. Precisely, for solutions $f_t$,

$$\frac{d}{dt} W_A^2(f_t, f_\infty) \le -2\kappa W_A^2(f_t, f_\infty)\,,$$

so that

$$W_A(f_t, f_\infty) \le e^{-\kappa t} W_A(f_0, f_\infty)\,.$$

This exponential decay holds even when the confinement potential $U$ has only a "mostly convex" structure (compactly supported perturbation of quadratic with small $C^2$ norm). It contrasts with standard $W_2$-dissipation, which fails in the absence of global convexity of $U$.

The loss is algorithmically robust: both $W_A$ and $J_A$ are computable directly from samples without simulating stochastic trajectories, provided samples from $f_\infty$ and the transport map $T$ are accessible. This enables their use as training losses in variational and deep learning frameworks targeting kinetic equilibrium (Salem, 2021).

4. Comparison to Classical and Alternative Kinetic OT Losses

The classical $W_2$ Wasserstein loss is inadequate for capturing convergence in systems with hypocoercive kinetic structure unless the drift is strictly convex. In contrast, the twisted kinetic loss $W_A$ fully incorporates the interplay of velocity dissipation and spatial transport, yielding faster and more robust trends to equilibrium, and often sharper rates compared to entropy-based approaches such as Villani's entropy-hypocoercivity method.

Alternative kinetic losses include:

• Second-order (acceleration-based) kinetic OT discrepancies, such as the OTIKIN cost, involving optimal transport with acceleration-based costs and Benamou–Brenier–Vlasov PDEs (Brigati et al., 21 Feb 2025).
• Kinetic energy minimization functionals in continuous normalizing flows and diffusion paths, important in generative modeling (Shaul et al., 2023).
• OT-inspired regularized losses, such as the Sinkhorn divergence, that act as kinetic proxies in neural network architectures but focus on first-order (velocity-based) dynamics (Khamlich et al., 2023).

Each generalizes the principle that the transport cost—whether defined via $A$-twisted distances, acceleration, or entropy regularization—can serve as a loss function directly encoding desired dynamical or geometric properties.

5. Practical Considerations and Implementation

Implementing a kinetic OT-inspired loss requires:

• Selection of the matrix $A$ to capture the coupling between position and velocity; explicit constructions depend on the kinetic equation's structure and the confinement potential's convexity properties.
• Computation of the optimal transport map, typically via a potential $\varphi(z)$ (twisted Brenier map) satisfying $\nabla^2 \varphi \succeq 0$ and $T(z) = A^{-1}\nabla\varphi(z)$.
• In the learning context, parametrizing the potential $\varphi$ via deep neural networks and optimizing $\mathcal{L}_{\mathrm{kinetic}}$ or $\mathcal{L}_{\mathrm{diss}}$ using stochastic sampling from $f_\infty$.

For non-convex or high-dimensional settings, regularization and preconditioning via $A$ are critical for numerical stability and effective convergence.

6. Broader Context and Influence

The optimal-transport-inspired kinetic loss formalism extends OT theory to dynamic settings, merging geometric, analytical, and algorithmic advances. Its adoption enables explicit interpretation of loss landscapes in kinetic learning, dimension-free hypocoercivity rates, and robust handling of local non-convexities in modeling complex dynamical systems (Salem, 2021).

In summary, OT-inspired kinetic losses represent a principled methodology to enforce and exploit phase-space geometric structure in the optimization of kinetic systems, with wide applicability in kinetic theory, PDE analysis, and modern machine learning.