Neural Stochastic Flows

• Neural stochastic flows are trainable models that combine deterministic bijections with stochastic layers to represent complex transport and transition measures.
• They interleave invertible mappings with stochastic processes, such as Brownian motion or Markov kernels, to achieve tractable densities and efficient sampling.
• These models power applications in Bayesian inference, multiscale reduction, and accelerated simulations by leveraging path-space optimization and solver-free SDE transitions.

Neural stochastic flows are trainable stochastic-flow models that use neural parameterizations to represent transport, transition, or path measures in settings where deterministic flows alone are insufficient. In the literature, the term spans several related constructions: stochastic normalizing flows that interleave invertible maps with stochastic layers; neural SDE-based flows that treat Brownian motion as a latent variable and reduce training to random neural ordinary differential equations; finite-time Schrödinger–Föllmer processes for Bayesian inference; and conditional normalizing-flow models that directly learn SDE transition laws in a solver-free manner (Hodgkinson et al., 2020). Across these formulations, the common objective is to retain the structural advantages of flow-based models—tractable densities, reparameterized sampling, or explicit transition kernels—while incorporating stochasticity through Markov kernels, Brownian forcing, or latent diffusion mechanisms (Hagemann et al., 2021).

1. Conceptual scope and formal definitions

In the broadest usage, a stochastic flow is a family of mappings $$\Phi_{s,t}(\cdot,\omega):\mathbb{R}^d\to\mathbb{R}^d$$, indexed by time pairs $(s,t)$, such that for each realization $\omega$ of the driving Brownian motion,

$X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$

and the Chapman–Kolmogorov or semi-group property holds:

$$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$

Marginalizing out $\omega$ gives the usual transition density $p(X_t\mid X_s=x_s)$ (Kiyohara et al., 29 Oct 2025).

A normalizing flow is a bijection $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$ such that $X\approx \mathcal{T}(Z)$ in distribution, with density given by the change-of-variables formula

$$p_{\mathcal{T}_*P_Z}(x)=p_Z(\mathcal{T}^{-1}(x))\cdot |\det\nabla \mathcal{T}^{-1}(x)|.$$

To allow richer transformations, one interleaves deterministic invertible maps with stochastic “layers.” The resulting object is called a Stochastic Normalizing Flow and is most naturally described as a pair of Markov chains, one forward and one reverse (Hagemann et al., 2021).

One neural-SDE formulation models the generative flow as the solution $(s,t)$0 of

$(s,t)$1

where $(s,t)$2 is an $(s,t)$3-dimensional standard Brownian motion and $(s,t)$4 are learnable parameters. Using the theory of rough paths, the underlying Brownian motion is treated as a latent variable and approximated, enabling efficient training of neural SDEs as random neural ordinary differential equations (Hodgkinson et al., 2020).

A distinct but closely related solver-free line defines Neural Stochastic Flows as conditional normalizing flows that learn to approximate the transition laws of a stochastic differential equation in a single-shot manner. These models introduce a learnable mapping

$(s,t)$5

so that $(s,t)$6 induces a conditional density $(s,t)$7 by change of variables (Kiyohara et al., 29 Oct 2025).

2. Stochastic normalizing flows and path-space formulations

A central mathematical formalization treats a stochastic normalizing flow as a pair of Markov chains: $(s,t)$8

$(s,t)$9

with positive marginal densities, factorized path measures, and mutual absolute continuity of the reverse-and-forward conditional measures (Hagemann et al., 2021). In this formulation, deterministic invertible layers, overdamped Langevin layers, Metropolis–Hastings layers, Metropolis-adjusted Langevin layers, variational autoencoder layers, and diffusion normalizing flow layers are all admissible transition operators (Hagemann et al., 2021).

The fundamental training objective is the path-space Kullback–Leibler divergence

$\omega$0

which upper-bounds $\omega$1 because the KL of joints is greater than or equal to the KL of marginals (Hagemann et al., 2021). This path-space viewpoint clarifies why stochastic layers can increase expressivity: purely deterministic normalizing flows are continuous bijections of $\omega$2, so the image of a connected support remains connected, whereas a stochastic layer can split and re-join mass and thereby map a unimodal base to a multimodal target without requiring Jacobian blow-up along connecting paths (Hagemann et al., 2021).

In lattice field theory, the same hybrid construction is interpreted as a non-equilibrium transformation tied to Jarzynski’s equality. An SNF alternates deterministic flow layers with stochastic Monte Carlo updates, and each stage contributes a “work” increment from the action change, Jacobian term, or MCMC heat term. The total work satisfies Jarzynski’s equality, so unbiased estimates of partition functions and observables are obtained by reweighting with $\omega$3 (Caselle et al., 2022). A related formulation writes the variational loss as

$\omega$4

linking SNF training directly to path-space variational inference (Caselle et al., 2022).

This suggests that “neural stochastic flow” is not a single model class but a family of constructions unified by path-space transport with stochastic transitions. In some papers the emphasis is expressive density modeling; in others it is unbiased non-equilibrium estimation, Markov-chain design, or surrogate modeling of stochastic dynamics (Caselle et al., 2022).

3. Flow-based learning of stochastic differential equation laws

A major recent development uses conditional normalizing flows to learn SDE transition laws directly, rather than learning an underlying drift-diffusion pair and then solving numerically. In this formulation, the induced conditional density is

$\omega$5

with

$\omega$6

In practice $\omega$7 is built from $\omega$8 coupling layers whose scales and shifts depend on $\omega$9, and each scale or shift network is multiplied by a time-scaling factor $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$0 satisfying $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$1, ensuring that when $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$2 the entire flow reduces to the identity map (Kiyohara et al., 29 Oct 2025).

To recover key stochastic-flow structure, these models impose identity at zero interval, semi-group regularization through a bidirectional KL penalty comparing $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$3 to the composition of two shorter transitions, and stationarity for autonomous SDEs by omitting explicit conditioning on $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$4 when the drift and diffusion are time-homogeneous (Kiyohara et al., 29 Oct 2025). For fully observed trajectories, training maximizes the exact conditional log-likelihood

$X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$5

while latent variants combine a latent transition flow $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$6, an emission model $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$7, and an approximate posterior $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$8 within an ELBO (Kiyohara et al., 29 Oct 2025).

The computational motivation is explicit: once $X_t(\omega)=\Phi_{s,t}(X_s(\omega),\omega),$9 is trained, one-shot sampling from $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$0 to $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$1 requires only a single pass through $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$2 coupling layers, yielding complexity $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$3 independent of $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$4, whereas Euler–Maruyama requires $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$5 small steps and costs $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$6 (Kiyohara et al., 29 Oct 2025). The reported empirical summary states that on Ornstein–Uhlenbeck, Lorenz, and 2D nonlinear drift examples, NSFs match the true transition densities under MMD and KL metrics while sampling 10–100× faster than Euler–Maruyama for $$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$7, and that on CMU motion-capture trajectories and Stochastic Moving MNIST, predictive log-likelihoods are nearly identical to conventional SDE solvers (Kiyohara et al., 29 Oct 2025).

A PDE-oriented alternative learns transition probability density functions by solving the corresponding Fokker–Planck equation with an atomic initial distribution, parametrically with respect to the location of the initial mass. The Neural Galerkin Normalizing Flow framework searches for the solution as a transformation of the transition probability density function of a reference stochastic process, which automatically preserves positivity and mass conservation (Saporiti et al., 19 Mar 2026). Here the density ansatz

$$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$8

is coupled to a Dirac–Frenkel condition that yields an ODE system for the time evolution of the normalizing flow parameters (Saporiti et al., 19 Mar 2026). The two NSF lines differ in mechanism—data-driven transition-law fitting versus PDE residual minimization—but both target transition densities and both use normalizing flows as structure-preserving surrogates.

4. Bayesian inference, multiscale reduction, and posterior uncertainty

Neural stochastic flows also appear as inference engines. Neural Schrödinger–Föllmer Flows cast approximate Bayesian inference as a stochastic-control problem. The uncontrolled prior process is

$$\Phi_{u,t}(\Phi_{s,u}(x,\omega),\omega)=\Phi_{s,t}(x,\omega),\qquad 0\le s\le u\le t.$$9

and the controlled process is

$\omega$0

with the terminal law constrained to equal the posterior $\omega$1 (Vargas et al., 2021). The optimal control minimizes a relative-entropy functional against Wiener measure; after Euler–Maruyama discretization, gradients are computed by backpropagating through the time-unfolded SDE simulation, and the framework admits a variance-reduction strategy called the STL estimator whose directional derivative at the optimum vanishes almost surely (Vargas et al., 2021). The empirical summary reports that N-SFS attains 89.28% accuracy and ECE $\omega$2 on the Banana dataset, 98.89% accuracy and ECE $\omega$3 on MNIST, and 61.56% accuracy and ECE $\omega$4 on CIFAR10 (Vargas et al., 2021).

In multiscale stochastic modeling, normalizing flows are used to parameterize an invariant measure that is otherwise defined implicitly by an intractable PDE. The slow–fast system is modeled as coupled SDEs

$\omega$5

and stochastic averaging yields an effective SDE for the slow component involving the invariant measure $\omega$6 of the fast process (Saha et al., 10 May 2026). The invariant distribution is modeled as $\omega$7 with exact density

$\omega$8

where $\omega$9 is built from RealNVP affine-coupling layers (Saha et al., 10 May 2026). Learning proceeds by a penalized negative log-likelihood induced by the reduced SDE, with drift estimated by Monte Carlo integration through the flow, and uncertainty quantification is obtained via a second normalizing flow $p(X_t\mid X_s=x_s)$0 that approximates the posterior over model parameters through an ELBO (Saha et al., 10 May 2026). In the tagged-particle in solvent benchmark, the flow-based estimator achieves mean-squared error an order of magnitude smaller than an unstructured neural network baseline and yields credible bands that cover the true drift (Saha et al., 10 May 2026).

These constructions illustrate two distinct roles for stochastic flows in inference. In Schrödinger–Föllmer models, the stochastic flow itself is the approximate posterior transport. In multiscale reduction, the flow parameterizes a latent invariant law or posterior over parameters that enters an effective stochastic model.

5. Architectural variants and domain-specific instantiations

The literature includes several specialized neural stochastic-flow architectures. One class concerns stochastic neural networks with structured posterior parameterizations. Kronecker Flow transforms a base random matrix $p(X_t\mid X_s=x_s)$1 with i.i.d. standard normal entries into a random weight matrix $p(X_t\mid X_s=x_s)$2 by applying a row-wise flow $p(X_t\mid X_s=x_s)$3 independently to each column and a column-wise flow $p(X_t\mid X_s=x_s)$4 independently to each row. If $p(X_t\mid X_s=x_s)$5 and $p(X_t\mid X_s=x_s)$6 are linear, one recovers the matrix-normal reparameterization $p(X_t\mid X_s=x_s)$7 with $p(X_t\mid X_s=x_s)$8 parameters instead of $p(X_t\mid X_s=x_s)$9; allowing $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$0 and $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$1 to be arbitrary invertible maps generalizes this to a more expressive family (Huang et al., 2019). The resulting variational posterior is used for ELBO optimization, PAC-Bayes bound estimation, and approximate Thompson sampling in contextual bandits (Huang et al., 2019).

Another sequential construction is Recurrent Autoregressive Flows, which model a stochastic process by conditioning a time-varying normalizing flow on a recurrent hidden state. At time $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$2, the conditional density is

$\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$3

where $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$4 is produced by a GRU-based recurrent core (Mern et al., 2020). In the reported experiments, RAF attains average test log-probability density $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$5 on the stochastic maze-navigation task, compared with $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$6 for an RNN–Gaussian baseline (Mern et al., 2020).

A hybrid continuous-and-discrete formulation is Neural Jump Stochastic Differential Equations, which model a hidden state with deterministic drift, optional diffusion, and event-driven jumps:

$\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$7

In the original paper $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$8, so the continuous part is an ODE plus jumps, and event times are governed by a conditional intensity $\mathcal{T}_\theta:\mathbb{R}^d\to\mathbb{R}^d$9 (Jia et al., 2019). This construction was introduced for marked point processes with piecewise-continuous latent trajectories, including Stack Overflow badges, medical records, and earthquake monitoring (Jia et al., 2019).

A different meaning of stochasticity appears in “Stochastic Sampling from Deterministic Flow Models,” where a deterministic ODE sampler is converted into a family of Itô SDEs with the same marginals. For deterministic flows $X\approx \mathcal{T}(Z)$0, choosing $X\approx \mathcal{T}(Z)$1 yields

$X\approx \mathcal{T}(Z)$2

which preserves all intermediate marginals $X\approx \mathcal{T}(Z)$3 (Singh et al., 2024). The paper emphasizes that this provides a family of stochastic samplers for fixed deterministic flow models rather than a new training objective (Singh et al., 2024).

Family	Core mechanism	Representative papers
Stochastic normalizing flows	Invertible maps interleaved with stochastic kernels or MCMC updates	(Hagemann et al., 2021, Caselle et al., 2022, Caselle et al., 2022)
Solver-free SDE transition models	Conditional normalizing flows for $X\approx \mathcal{T}(Z)$4 with flow-consistency regularization	(Kiyohara et al., 29 Oct 2025)
Bayesian and multiscale inference flows	Controlled SDEs or normalizing-flow parameterizations of latent invariant/posterior laws	(Vargas et al., 2021, Saha et al., 10 May 2026)

6. Applications, limitations, and recurrent misconceptions

Applications span sampling, inference, surrogate modeling, and scientific computing. In lattice field theory, stochastic normalizing flows are used to sample Boltzmann distributions and to estimate free-energy differences in the two-dimensional $X\approx \mathcal{T}(Z)$5 theory (Caselle et al., 2022). In Bayesian learning, Neural Schrödinger–Föllmer Flows provide a finite-time and low-variance framework for posterior approximation (Vargas et al., 2021). In multiscale systems, normalizing flows are used to learn effective stochastic dynamics from a single observed slow trajectory while the fast variables remain unobserved (Saha et al., 10 May 2026). In transition-density approximation, Neural Galerkin Normalizing Flow serves as a surrogate model for many-query problems associated with stochastic differential equations, including Bayesian inference, simulation, and diffusion bridge generation (Saporiti et al., 19 Mar 2026).

A common misconception is that stochasticity is introduced only to improve sample diversity. The cited literature assigns several more technical roles to stochasticity: breaking topological constraints of deterministic bijections (Hagemann et al., 2021); restoring exact target marginals when converting an ODE sampler into an SDE sampler (Singh et al., 2024); matching multi-point statistics of stochastic gradient descent through cylindrical noise in stochastic modified flows (Gess et al., 2023); or encoding epistemic uncertainty through variational posteriors over flow parameters (Saha et al., 10 May 2026).

Another misconception is that neural stochastic flows are synonymous with one architecture. The papers instead describe multiple non-equivalent paradigms. Some models are path-space generative models built from forward and reverse Markov chains (Hagemann et al., 2021). Some are neural SDEs trained as random ODEs (Hodgkinson et al., 2020). Some are conditional normalizing flows for arbitrary-time SDE transition kernels (Kiyohara et al., 29 Oct 2025). Some are stochastic-control processes for Bayesian inference (Vargas et al., 2021). The unifying feature is not a single parameterization but the use of neural transport or transition maps in explicitly stochastic settings.

Limitations are also formulation-specific. In lattice applications, the number of deterministic and stochastic blocks is a key hyperparameter: too few flow blocks give limited expressivity, too few stochastic blocks give poor ergodicity, and deterministic dominance may cause mode collapse (Caselle et al., 2022). In jump-SDE models, adaptive ODE solves can be slow when events are frequent, and storing jump states and Jacobians increases memory and compute cost (Jia et al., 2019). In solver-free SDE transition models, the semi-group property is only approximately enforced through regularization rather than imposed exactly (Kiyohara et al., 29 Oct 2025). In Schrödinger–Föllmer inference, multimodal toy posteriors may miss distant modes unless exploration noise is large (Vargas et al., 2021).

Taken together, the literature indicates that neural stochastic flows are best understood as a methodological umbrella for neural models of stochastic transport, stochastic dynamics, and path-space inference. The precise object being learned may be a density push-forward, a Markov kernel, a transition law, a controlled diffusion, or a posterior over latent parameters; the choice depends on whether the primary problem is generative modeling, Bayesian inference, multiscale reduction, numerical acceleration, or stochastic process learning.