Weighted SDEs: Theory & Applications

• Weighted SDEs are systems that incorporate state-dependent or auxiliary weights to model nonlinear, degenerate, and interacting stochastic processes.
• They enable advanced particle algorithms and high-order numerical schemes for efficient simulation and sampling in high-dimensional and complex settings.
• Weighted SDEs provide explicit control over integrability and regularity using Fisher information and norm estimates, linking SDE theory with nonlinear PDE analysis.

Weighted stochastic differential equations (SDEs) extend classical stochastic analysis by incorporating auxiliary weight processes or position-dependent weights into the dynamics, enabling the accurate description of nonlinear, degenerate, and interacting random systems. Weighted SDEs form the basis for advanced particle algorithms, high-order numerical schemes, information geometry-motivated samplers, and precise norm and entropy bounds in both infinite and finite-dimensional stochastic models.

1. Definitions and General Framework

Weighted SDEs are systems where the evolution of a random process is informed or corrected by an explicitly evolving weight variable or by incorporating state-dependent weighting into the measure of interest. Two pervasive forms appear in the recent literature:

• Explicit weight SDEs: Particle systems $(X_t, w_t)$ with

$$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$

such that the weighted empirical measure is used to approximate a non-conservative or nonlinear PDE, e.g., evolved via a Feynman–Kac representation or as part of a Wasserstein–Fisher–Rao (WFR) gradient flow (Rahimi, 19 Dec 2025).

• Interacting weights: Systems of weighted particles $\{(X_i(t), A_i(t))\}$, where each $A_i(t)$ evolves by an SDE whose coefficients depend on the empirical measure generated by the (possibly infinite) ensemble. The formal limit of the weighted empirical measure solves a nonlinear SPDE (Crisan et al., 2016).

In parallel, weighted SDE analyses introduce auxiliary weights (e.g., Fisher informations, norms) to capture convergence, regularity, or integrability properties of solutions, especially in degenerate or non-reversible settings (Feng et al., 2021, Yamazaki, 13 Jan 2026).

2. Particle Representations and Infinite-Dimensional Weighted SDEs

A canonical framework is the weighted particle system for SPDEs with boundary conditions: - Particles $\{X_i\}$ solve reflecting SDEs on a domain $D\subset\mathbb R^d$:

$$X_i(t) = X_i(0) + \int_0^t \sigma(X_i(s))\,dB_i(s) + \int_0^t c(X_i(s))\,ds + \int_0^t \eta(X_i(s))\,dL_i(s)$$

where $L_i$ is the boundary local time and $\eta$ is inward reflection (Crisan et al., 2016). - Individual weights $A_i(t)$ satisfy:

$$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$0

where $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$1 is the empirical density (Crisan et al., 2016).

A key result is that, as $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$2, the empirical measure $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$3 converges to $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$4, whose density solves a nonlinear SPDE of the form

$$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$5

with boundary conditions driven by the prescribed $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$6.

This framework enables pathwise construction, uniqueness, and verification of solutions to such SPDEs, especially for models like the stochastic Allen–Cahn equation with Dirichlet boundaries (Crisan et al., 2016).

3. Weighted SDEs in High-Order Weak Approximation and Machine Learning

Weighted SDEs play a crucial role in developing dimension-robust numerical solvers for high-dimensional integrals and Kolmogorov PDEs. The central principle is the use of high-order weak approximation schemes for SDEs, with corrections by explicitly computable Malliavin weights:

• For the SDE solution $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$7, the weak approximation

$$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$8

achieves $$\begin{cases} dX_t = b(t, X_t)\,dt + \sigma(t)\,dW_t, \ d\log w_t = \psi(t, X_t)\,dt, \end{cases}$$9, with $\{(X_i(t), A_i(t))\}$0 a polynomial "weight" function of the Brownian increment (Naito et al., 2020).

• High-order schemes (e.g., $\{(X_i(t), A_i(t))\}$1) are realized through explicit formulas for $\{(X_i(t), A_i(t))\}$2 and $\{(X_i(t), A_i(t))\}$3 involving finitely many terms in $\{(X_i(t), A_i(t))\}$4 and the SDE coefficients.

These weights permit the design of stochastic minimization algorithms (e.g., SGD with weighted samples), where only $\{(X_i(t), A_i(t))\}$5 cost per path is needed (independent of the exponential in $\{(X_i(t), A_i(t))\}$6). Self-normalized weighted samples with these Malliavin weights yield estimates that avoid the curse of dimensionality up to $\{(X_i(t), A_i(t))\}$7 in experimental benchmarks (Naito et al., 2020).

4. Information Geometry, Gradient Flows, and Feynman–Kac Representation

Weighted SDEs generalize standard diffusion-based sampling and variational schemes. In the context of Wasserstein–Fisher–Rao (WFR) metric flows:

• The marginal law $\{(X_i(t), A_i(t))\}$8 evolves by a reaction–diffusion–transport PDE

$\{(X_i(t), A_i(t))\}$9

(Rahimi, 19 Dec 2025).

• The associated Feynman–Kac representation shows that $A_i(t)$0, with $A_i(t)$1 and $A_i(t)$2, provides particle paths with evolving weights such that the normalized weighted empirical measure approximates $A_i(t)$3.

This formulation realizes non-reversible samplers and WFR gradient flows, as the reaction (weight) term can geometrically and operator-theoretically modify spectral gaps and improve mixing, especially in nonconvex settings. Weighted SDEs thus provide a bridge between SDE-based sampling, optimal transport, and nonlinear PDEs (Rahimi, 19 Dec 2025).

5. Weighted Norms, Fisher Information, and Integrability Estimates

Weighting not only applies to particle-based or explicit auxiliary SDEs but also to analytic functionals such as norms and Fisher information. Two key directions are:

• Weighted integrability (norm equivalence): For SDEs on $A_i(t)$4, if $A_i(t)$5 solves an SDE with locally Lipschitz coefficients, and $A_i(t)$6 is a $A_i(t)$7 weight satisfying compatibility bounds with the growth of $A_i(t)$8 and $A_i(t)$9, then (Yamazaki, 13 Jan 2026)

$\{X_i\}$0

with fully quantitative $\{X_i\}$1 expressed in terms of $\{X_i\}$2 from the weight conditions.

• Weighted Fisher information: For hypoelliptic SDEs, a weighted Fisher information of the form

$\{X_i\}$3

serves as a Lyapunov functional, enabling explicit decay inequalities and exponential convergence in $\{X_i\}$4 and Kullback–Leibler divergence under explicit structural conditions filling in degenerate directions (Feng et al., 2021).

Weighted estimates are fundamental for rigorous control of integrability, regularity, and convergence rates in both particle and functional analytic settings.

6. Connections to Nonlinear and Degenerate Models

Weighted SDE techniques are indispensable for handling degeneracy, nonlinearity, and interactions in stochastic dynamics.

• In particle systems approximating nonlinear SPDEs (e.g., the stochastic Allen–Cahn equation), weighted SDEs on the weights with path-space interaction (empirical measure feedback) yield the correct macroscopic nonlinear evolution via propagation of chaos (Crisan et al., 2016).
• For non-reversible or degenerate diffusions (e.g., underdamped Langevin, interacting oscillator chains), weighted Fisher informations with suitable auxiliary directions enable the derivation of exponential ergodicity estimates otherwise unavailable for degenerate SDEs (Feng et al., 2021).

A plausible implication is that weighted SDEs, both at the trajectory and at the measure level, provide a unified formalism for lifting classical Markov formulations to settings where reaction terms, nonlinearities, or degeneracies preclude direct application of standard theory.

7. Algorithmic and Computational Aspects

Weighted SDEs are amenable to efficient computation due to explicit weight updates and fully parallelizable particle methods:

• In sampling and machine learning, high-order weighted schemes using Malliavin weights (as in WA2.0 and WA3.0) scale linearly in the ambient dimension and require only polynomials in Brownian increments and Jacobian-vector products (Naito et al., 2020).
• WFR-based weighted SDE algorithms (simulation loop detailed in (Rahimi, 19 Dec 2025)) involve standard Euler–Maruyama steps for particle locations and additive ODEs for weights, with optional resampling for variance reduction.
• Particle systems for SPDEs require simulation of an infinite (mean-field) or large number of interacting weight-driven trajectories, but admit a construction via fixed-point arguments and moment bounds (Crisan et al., 2016).

The convergence guarantees, variance properties, and avoidance of curse of dimensionality depend critically on the correct design of weight updates (either probabilistically justified as in Feynman–Kac or algorithmically as in SGD for weighted minimization).

---

References: (Crisan et al., 2016, Naito et al., 2020, Feng et al., 2021, Rahimi, 19 Dec 2025, Yamazaki, 13 Jan 2026).