Stochastic Differential Equations (SDEs)
- Stochastic differential equations (SDEs) are integral equations governing continuous-time dynamics driven by random processes, with or without Brownian motion, They describe physical, biological, financial, biomedical, ecological, control, optimization, and machine learning models.
- SDEs are defined on a filtered probability space, including Wiener processes (Brownian motion) whose increments follow a Gaussian distribution and stochastic integral defined through limiting procedures for adapted processes. Key mathematical expressions include Itô and Stratonovich interpretations, each having distinct implications and applications.
- Numerical methods for SDEs, such as Euler- Maruyama and Improved Euler-Heun, discretize and approximate solutions. These methods have varying orders of accuracy and are used in applications like pharmacokinetic, biomedical, growth, and ecological modeling. The robust for these methods is only grow as the order number appear at the exponent
Stochastic differential equations (SDEs) are differential or integral equations describing continuous-time dynamics driven by random processes, conventionally Brownian motion or a Wiener process. A general Itô SDE has the form
where the drift specifies the conditional systematic tendency and the diffusion coefficient specifies the scale and geometry of stochastic variation. The differential notation is shorthand for an integral identity; in particular, does not denote an ordinary derivative. SDEs provide pathwise models, while their probability laws evolve according to Fokker–Planck equations. They are used in physical, biological, financial, biomedical, ecological, control, optimization, and machine-learning models.
1. Mathematical formulation and stochastic calculus
An SDE is defined on a filtered probability space carrying a Brownian motion. For ,
and the increment is independent of the information available up to time . The rigorous meaning of
is the integral equation
The first integral is an ordinary Lebesgue integral; the second is a stochastic integral defined through limiting procedures for adapted processes. The distinction between a stochastic process and one of its sample paths is fundamental: is a random variable, whereas a realization is an ordinary function of time.
In the Itô interpretation, the integrand is evaluated using information available at the beginning of each time increment. The basic formal rules are
For a sufficiently smooth function , Itô’s formula is
The second-derivative term is the Itô correction and reflects the nonzero quadratic variation of Brownian motion.
A Stratonovich SDE is written
Its integral obeys the ordinary chain rule of deterministic calculus. In one dimension, the Stratonovich equation is equivalent to the Itô equation
The choice between Itô and Stratonovich interpretations is therefore part of the model specification, not merely a notational convention (Roberts, 2012).
The quadratic variation of Brownian motion is
It is obtained from sums of squared increments,
rather than from an ordinary Stieltjes integral. Consequently,
for Brownian motion, and
This distinguishes stochastic calculus from deterministic calculus. Brownian paths are continuous but almost surely nowhere differentiable and of unbounded variation, so ordinary pathwise Riemann–Stieltjes integration is generally unavailable. Henstock–Kurzweil or Stieltjes-complete integration provides a broader cancellation-based framework, but Itô integration remains a distinct probabilistic construction (Muldowney, 2014).
2. Drift, diffusion, and probability laws
For a scalar Itô diffusion,
the drift satisfies
while the conditional variance satisfies
In dimensions, the diffusion covariance is
The drift and diffusion have distinct inferential roles: systematic local motion should not be conflated with intrinsic stochasticity.
If denotes the density of , then formally
This is the Fokker–Planck equation. The SDE represents individual trajectories, whereas the Fokker–Planck equation represents the evolution of their ensemble density. The generator acting on smooth test functions is
For the inward-pointing discontinuous drift
the classical global Lipschitz assumptions fail. A formal stationary Fokker–Planck calculation gives
a Laplace density centered at zero with scale $1/(2k)$. Euler–Maruyama density recursions and smooth approximations of the sign function approach this candidate numerically, although the existence, uniqueness, Euler–Maruyama convergence, and convergence of stationary densities are not established in that analysis (Simonsen et al., 2013).
SDEs can also be coupled to nonlinear evolution of probability laws. A distribution-dependent SDE, or McKean–Vlasov SDE, has the form
Its law solves a nonlinear Fokker–Planck equation. Weak existence can be formulated through a superposition principle: a solution of the nonlinear Fokker–Planck equation corresponds to a weak solution of the distribution-dependent SDE. Well-posedness is commonly obtained by fixed-point arguments on curves of probability measures, with Wasserstein-distance estimates controlling dependence on the law (Huang et al., 2020).
3. Existence, uniqueness, regularity, and constraints
Classical existence and uniqueness theorems typically require local or global Lipschitz conditions and suitable growth bounds. Under such assumptions, strong solutions are constructed on a prescribed probability space and are pathwise unique. When coefficients are discontinuous, singular, distribution-dependent, or constrained by algebraic relations, these conclusions may fail or require modified notions of solution.
For discontinuous drifts, one approach replaces a discontinuous function by its Filippov regularization
The SDE is then interpreted as an inclusion,
Smooth approximations, uniform Itô estimates, tightness, Skorohod representations, and weak convergence can yield martingale solutions. In finite-dimensional systems with measurable linear-growth drift and Lipschitz diffusion, existence of at least one martingale solution is obtained, but uniqueness is not asserted in general. The same strategy extends to stochastic heat equations and stochastic parabolic systems (Barbu et al., 2015).
A different pathwise theory treats discontinuous diffusion coefficients in one-dimensional equations driven by Hölder processes: If is Hölder continuous with exponent greater than $1/2$, has locally bounded variation and fixed sign, and is locally integrable, the Lamperti transform
gives the explicit solution
Uniqueness holds up to the first time at which and is global when never vanishes. The theory applies, in particular, to fractional Brownian motion with Hurst parameter (Torres et al., 2019).
Constraints introduce further compatibility requirements. An explicit stochastic differential-algebraic equation may be written
together with
If the constraint manifold is smooth, every diffusion direction must be tangent to it: Equivalently, each diffusion direction must satisfy
Failure of this stochastic tangency condition can imply nonexistence of a pathwise Itô solution. For index-one systems, Itô differentiation of the algebraic constraint yields an ordinary SDE for the combined differential and algebraic variables. For completely high-index systems, repeated stochastic differentiation may be required, and some systems are ill-posed. Approximate methods enforce the constraint either with probability one locally or with a prescribed probability bound (Suthar et al., 2021).
The -framework replaces a single probability law by a sublinear expectation representing volatility uncertainty. A -Brownian motion has uncertain quadratic variation,
A -SDE contains both and terms. Its pathwise representation can reduce the equation to a parametrized ordinary differential equation, and comparison theorems hold quasi-surely under conditions involving the nonlinear function
4. Numerical integration and likelihood approximation
The Euler–Maruyama method discretizes
as
where
For sufficiently regular equations, its deterministic order is $1$ and its usual strong order is $1/2$. At low observation frequency, Euler–Maruyama transition likelihoods can be substantially biased, especially for mean-reverting or nonlinear diffusions.
A modified stochastic Improved Euler or Heun method introduces an auxiliary random sign for Itô equations:
For Stratonovich equations, . Under smoothness assumptions, the method has global strong order one; specially structured additive-noise equations can exhibit strong order two (Roberts, 2012).
Higher strong-order stochastic Runge–Kutta methods require iterated Itô integrals. For example,
whereas cross-integrals for require additional approximations, such as Wiktorsson-type random series. Strong methods approximate individual paths and must use consistent Brownian increments across resolutions. Weak methods approximate expectations and distributions; they can replace Gaussian increments by discrete random variables with matching moments and require large Monte Carlo ensembles (Gevorkyan et al., 2016).
For likelihood-based inference, the exact SDE transition density is generally unavailable. A high-order Hermite expansion can approximate it after a Lamperti-type transformation to unit diffusion. For an order- expansion,
is expressed through powers of and recursively computed coefficients. The order-$2$ approximation is used in the cited simulations. It performs better than Euler–Maruyama at sparse or irregular observation frequencies, although deriving the multidimensional coefficients can be algebraically demanding (Picchini et al., 2010).
When random effects describe between-subject variability, the resulting stochastic differential mixed-effects model is
The diffusion represents within-subject stochasticity, while represents systematic between-subject variability. Integrating out multiple random effects is performed with a Laplace approximation. Subject-level modes are found through internal optimization, while population parameters are optimized externally. This permits likelihood-based inference for multidimensional SDEs with multiple random effects in pharmacokinetic, biomedical, growth, and ecological applications (Picchini et al., 2010).
5. Modeling, inference, and learning
SDE parameters may be fixed, random, covariate-dependent, or distribution-dependent. Varying-coefficient SDEs replace constant parameters by predictors such as
where may be a linear effect, factor effect, random effect, or penalized spline. This allows drift and diffusion parameters to vary with time, temperature, location, individual identity, or behavioral phase while retaining a mechanistic SDE form. Approximate transition likelihoods are evaluated using closed-form transitions when available or Euler–Maruyama otherwise. Penalized likelihood, generalized additive-model bases, random effects, automatic differentiation, and Laplace approximation provide computationally efficient inference (Michelot et al., 2020).
A neural SDE-learning strategy can embed an Euler–Maruyama transition directly into a neural architecture: with . The approximate conditional transition is Gaussian with mean
and covariance
Maximum likelihood jointly estimates drift and diffusion. The covariance term is essential: treating stochastic residuals as deterministic prediction error prevents estimation of the diffusion. Applications to geometric Brownian motion and a stochastic Lorenz-63 system show that explicit diffusion modeling improves recovery of stochastic behavior relative to deterministic neural networks (Dridi et al., 2021).
Neural Laplace methods instead represent a stochastic trajectory through its pathwise Laplace transform,
For geometric Brownian motion,
the mean transform satisfies
when . The approach is exploratory: it studies geometric Brownian motion, does not explicitly decompose drift and diffusion, and does not establish general SDE-identification results. In the reported experiment, deterministic neural ODE baselines outperform Neural Laplace because the selected geometric Brownian motion regime is dominated by exponential drift (Carrel, 2024).
SDEs also arise as continuous-time models of stochastic gradient descent. For least-squares problems, the corresponding SDE has gradient drift and a diffusion covariance matching the covariance of stochastic-gradient fluctuations: In finite-sample overparameterized problems, the diffusion range lies in the span of the training data, and both drift and diffusion vanish on the interpolation set. Consequently, invariant measures can be degenerate and supported on interpolating solutions. In population models with persistent observation noise, the invariant distribution is generally nondegenerate (Schertzer et al., 2024).
6. Extensions, geometry, and applications
SDE theory extends beyond semimartingale Brownian drivers. Rough stochastic differential equations combine an Itô Brownian component with a pathwise rough component: The rough path stores first- and second-level increments and satisfies Chen’s relation. Stochastic controlled rough paths separate the martingale component from the rough controlled component. The resulting solution map is pathwise and locally stable in the rough-path metric while retaining progressive measurability and Itô integration for the Brownian component. This framework supports conditional SDEs, filtering, stochastic control, common-noise McKean–Vlasov systems, and rough Fokker–Planck equations (Friz et al., 2021).
Loop-driven SDEs probe noncommutative geometry. If a Brownian motion is conditioned to close in a free nilpotent group through level , all iterated integrals up to level vanish at the terminal time. The first surviving effects arise from vector-field commutators of length . For a Brownian bridge, corresponding to , the leading operator is
If the relevant commutators vanish, the terminal random variable collapses to its initial value; if they span the state space, the terminal law has a smooth density. Thus loop-driven equations connect SDEs with free nilpotent groups, Hörmander theory, Malliavin calculus, stochastic holonomy, and hypoelliptic density asymptotics (Baudoin, 2014).
Random delays produce a different departure from ordinary Langevin models. In the equation
the stochasticity enters through a randomly selected past argument rather than an additive or multiplicative white-noise term. Under the self-averaging interpretation, the continuous-time equation becomes deterministic: Its solution is
where is the modified Bessel function. A finite-step discretization retains intrinsic fluctuations, and the deterministic continuous limit is recovered only when the step size tends to zero at fixed physical time (Krapivsky et al., 2011).
SDEs may also include random effects, memory, uncertain volatility, singular perturbations, and boundary constraints. For -stochastic delay equations, practical exponential stability in the th moment can be transferred among the exact delayed equation, an auxiliary nondelayed equation, and their Euler–Maruyama approximations when the delay and step size are sufficiently small. The stability estimate has the form
where allows convergence to a nonzero residual neighborhood (Lu, 2024).
Applications include pharmacokinetics and pharmacodynamics, biomedical modeling, animal movement, body-condition dynamics, diving behavior, population ecology, stochastic control, finance, and optimization. In ecological models, varying drift and diffusion can describe temperature-dependent elephant movement, covariate-dependent lipid accumulation in elephant seals, and phase-dependent variability in whale diving. In population models, multiplicative noise can produce extinction events absent from deterministic dynamics. The central methodological distinction throughout these applications is between systematic structure encoded by the drift, intrinsic stochastic fluctuations encoded by the diffusion, heterogeneity encoded by random effects or law dependence, and measurement or observation error, which requires an additional observation model.