Reformulating Stochastic Differential Equations as Ordinary Differential Equations

22 concepts12 papers1 proposal

Stochastic Differential Equations (SDEs) are the natural mathematical language for systems evolving under uncertainty, appearing everywhere from computational biology to financial markets. However, SDEs are notoriously difficult to optimize, simulate, and infer. Because Brownian motion is nowhere differentiable, standard deterministic calculus breaks down. Practitioners cannot easily compute gradients with respect to parameters, utilize adaptive high-order numerical solvers, or track probability densities without resorting to computationally expensive Monte Carlo path simulations.

A major research direction addresses this challenge by transforming stochastic systems into deterministic or pathwise-differentiable ordinary differential equations (ODEs). By decoupling the source of randomness from the time-dependent dynamics, these methods unlock the entire toolkit of deterministic numerical analysis—such as the adjoint method, gradient matching, and adaptive step-size integration—for stochastic inference and generative modeling.

Pathwise Smooth Approximations: SDEs to RODEs

The primary reason SDEs resist standard calculus is the "roughness" of continuous noise. If we look at a single realization of a Brownian motion path, it is jagged at every scale. However, if we can approximate this rough path with a smooth, differentiable trajectory, the SDE transforms into a Random Ordinary Differential Equation (RODE). Along any single sampled path, the system behaves completely deterministically.

Simo Särkkä et al. in Series Expansion Approximations of Brownian Motion for Non-Linear Kalman Filtering of Diffusion Processes demonstrated that representing continuous stochastic noise as a truncated series expansion converts SDEs into randomized ODEs. This conversion allows the use of high-order numerical integration techniques and adaptive solvers that are typically reserved for deterministic systems.

Building on this pathwise view, Sanmitra Ghosh et al. in Differentiable Bayesian inference of SDE parameters using a pathwise series expansion of Brownian motion utilized a finite series expansion to create a fully differentiable framework for parameter estimation. Because the resulting system is pathwise differentiable, gradients can be computed efficiently via adjoint sensitivity analysis, enabling advanced, gradient-based Bayesian inference methods like the No-U-Turn Sampler (NUTS).

Standard SDE:  dX_t = f(X_t)dt + g(X_t)dW_t   (dW_t is rough and non-differentiable)
                                    │
                       [Pathwise Approximation]
                                    ▼
Random RODE:   dX_t/dt = f(X_t) + g(X_t)ψ_t    (ψ_t is a smooth, differentiable sample path)

This perspective is highly versatile:

Absorbing Noise into the Drift Term

Instead of approximating individual noise paths, another strategy alters the system's coordinates or optimization constraints so that the stochasticity is mathematically absorbed into a modified, deterministic drift function.

For instance, when modeling complex environmental noise like fractional white noise—which exhibits long-range dependencies—Anh Tong et al. in Learning Fractional White Noises in Neural Stochastic Differential Equations showed that the system can be reformulated as an equivalent standard SDE by absorbing the fractional characteristics into modified drift and diffusion functions, enabling traditional backpropagation.

Taking this concept to its logical conclusion, one can eliminate the diffusion term entirely to form a probability flow ODE. Zhenyi Zhang et al. in Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport and Zhenyi Zhang et al. in Modeling Cell Dynamics and Interactions with Unbalanced Mean Field Schrödinger Bridge utilized Fisher information regularization to reformulate complex SDE constraints into computationally tractable ODE constraints. This transformation captures the intrinsic system noise and diffusion while leveraging the stability and speed of standard neural ODE solvers.

In sensitivity analysis, changing the diffusion (noise) coefficient usually creates mathematical pathologies because paths generated with different diffusion rates are mutually singular. Jérémie Klinger in Computing Nonequilibrium Responses with Score-Shifted Stochastic Differential Equations resolved this through a diffusion-to-drift transformation. By incorporating the score function of the system's probability distribution, a perturbation in the diffusion term is mathematically mapped into an equivalent perturbation in the drift term, allowing standard path-reweighting techniques to succeed.

Confining Randomness to Initial Conditions

A radical variation of the stochastic-to-deterministic transformation completely removes time-dependent randomness from the process altogether. Hossein Gorji in Logarithmic Gradient Transformation and Chaos Expansion of Itô Processes introduced a logarithmic gradient transformation for Itô processes driven by white noise.

This method reformulates the SDE into an equivalent system where the time-evolving noise is entirely eliminated. Instead, the randomness is confined strictly to an uncertain initial condition:

$$\frac{dY_t}{dt} = \mathcal{F}(Y_t, t), \quad Y_0 = \xi$$

Once the system is recast as an ODE with a random starting point $\xi$, established deterministic solution techniques like polynomial chaos expansion can be applied. The coefficients of this expansion evolve deterministically over time, bridging the gap between analyzing inherently stochastic processes and deterministic dynamical systems.

(Conversely, adding noise can sometimes be a useful mathematical trick for deterministic systems; Bo Thiesson et al. in ARMA Time-Series Modeling with Graphical Models demonstrated that transforming deterministic time-series models into stochastic ones by injecting a small Gaussian noise component enables the application of standard Expectation-Maximization algorithms for parameter estimation).

Why It Works: The Intuition of Decoupling

The core reason these approaches work is that they decouple what is random from how the system evolves.

In a standard SDE, randomness is injected continuously at every infinitesimal time step. This forces the use of specialized stochastic calculus (such as Itô or Stratonovich calculus) and prevents the use of standard derivative-based optimization because the next step's noise is always unpredictable and non-differentiable.

By converting the SDE into an ODE or RODE, we shift the perspective in one of two ways:

  1. Pathwise Conditioning: We sample a global, smooth approximation of the noise path before solving the differential equation. Conditioned on this concrete sample path, the trajectory is completely deterministic and differentiable, allowing us to compute exact parameter gradients using standard backpropagation and the adjoint method.
  2. Distributional Equivalence: Instead of tracking individual jittering particles, we track the evolution of the overall probability density. The aggregate effect of random diffusion can be perfectly described by a smooth, deterministic vector field (the score function of the distribution). This turns a chaotic stochastic particle system into a deterministic fluid flow.

Ultimately, these transformations allow machine learning models and numerical frameworks to treat stochasticity not as an intractable mathematical barrier, but as a structured, differentiable input.

Key Foundational Work

Series Expansion Approximations of Brownian Motion for Non-Linear Kalman Filtering of Diffusion Processes (2014) — contributed 2 concept elements to this direction.
ARMA Time-Series Modeling with Graphical Models (2004) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_stochastic_to_deterministic_transformati_02a4a718,
  title = {stochastic-to-deterministic transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stochastic-to-deterministic-transformation/rdr_02a4a718a2eb455ba939fa2e8e46e2b8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_stochastic_to_deterministic_transformati_02a4a718_v1,
  title = {stochastic-to-deterministic transformation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/stochastic-to-deterministic-transformation/rdr_02a4a718a2eb455ba939fa2e8e46e2b8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake