A stochastic differential equation (SDE) is an equation describing the time evolution of a stochastic process under the combined influence of systematic forces and random fluctuations. A common form is
where specifies the drift, specifies how noise enters the system, and is a Wiener process, also called Brownian motion. Unlike an ordinary differential equation, an SDE generally cannot be interpreted through ordinary derivatives: Brownian paths are almost surely nowhere differentiable, so the equation requires a theory of stochastic integration.
SDEs provide a mathematical language for diffusion, thermal fluctuations, randomly varying populations, financial prices, and many other evolving systems. They connect individual random trajectories with equations for probability distributions, and connect probabilistic models with analysis, geometry, computation, and statistical inference. Their development grew from studies of Brownian motion, including Albert Einstein’s diffusion theory in 1905 and Paul Langevin’s dynamical formulation in 1908. Kiyosi Itô’s work beginning in the 1940s established the stochastic calculus that underlies much of the modern theory.
Mathematical formulation and information structure
The standard finite-dimensional Brownian SDE takes values in :
Here is an -dimensional Brownian motion, takes values in , and is a matrix. The number of noise sources need not equal the number of state variables. The associated instantaneous covariance matrix is
Different matrices can produce the same , a distinction that matters when comparing trajectory-level constructions with distributional descriptions.
The differential notation abbreviates an integral equation:
The first integral is an ordinary time integral, usually understood as a Lebesgue integral. The second is a stochastic integral. Consequently, specifying the symbols , , and is not sufficient: the interpretation of that stochastic integral is part of the equation. Unless otherwise stated, mathematical treatments commonly use the Itô interpretation.
The probabilistic setting is a probability space equipped with a filtration , an increasing family of collections of events representing information available by each time. The solution must be adapted: is measurable with respect to . Brownian increments after time must be independent of . These requirements prevent a model from using future noise to determine its present state. Initial data may be deterministic or random, provided they are compatible with this information structure.
For small , sufficiently regular coefficients give the local approximations
and
Thus drift increments have size proportional to , while typical noise increments have size proportional to . This difference in scaling explains why squared noise increments contribute to the calculus rather than disappearing.
An equation with random coefficients but no stochastic integral can be a random ordinary differential equation rather than an SDE in this narrower sense. For example, , with a continuous random coefficient , can be solved path by path using ordinary integration. Brownian-driven SDEs require additional machinery because the driving paths have unbounded variation on every nontrivial time interval, almost surely.
Stochastic integration and the change-of-variables rule
The Itô integral is built first for simple adapted processes. If on each interval , its integral is defined by
The integrand is evaluated before the corresponding Brownian increment occurs. Completion in a suitable mean-square space extends this definition to progressively measurable processes satisfying
A central identity, the Itô isometry, is
Such integrals are martingales with zero mean when started at zero. Local versions permit integrands without global square integrability.
Brownian motion has quadratic variation equal to elapsed time. Along increasingly fine deterministic partitions of , the sums of squared Brownian increments converge in probability to . For an Itô process, this becomes
in the scalar case. The mnemonic records this second-order contribution; it is not ordinary multiplication of infinitesimals.
The resulting change-of-variables formula, Itô’s formula, is fundamental. If
and has one continuous time derivative and two continuous space derivatives, then
The second-derivative term distinguishes this formula from the ordinary chain rule. For and , it gives
The Stratonovich integral, denoted by , uses a different limiting convention and obeys the ordinary chain rule for smooth changes of variables. In one dimension,
is equivalent, under sufficient smoothness, to
In several dimensions, if is the th noise vector field, the correction is .
Neither convention is universally preferable. Itô integration aligns naturally with nonanticipating information and martingale methods. Stratonovich integration is convenient for coordinate changes and often appears when rapidly fluctuating smooth forcing is replaced by idealized white noise. Smooth approximations can converge to Stratonovich equations through Wong–Zakai-type results, but the limiting interpretation depends on the approximation and scaling. Changing conventions without changing the drift generally changes the model.
Existence, uniqueness, and meanings of solution
A strong solution is constructed on a prescribed probability space using a specified Brownian motion and initial condition. Apart from the initial randomness, it is determined by the driving noise. A weak solution permits the probability space, Brownian motion, and process to be constructed together. Here “weak” refers to the probabilistic formulation, not to weak spatial derivatives or the distributional interpretation used for partial differential equations.
There are corresponding notions of uniqueness. Pathwise uniqueness means that two solutions on the same space, driven by the same Brownian motion and having the same initial condition, are indistinguishable: with probability one, they agree at every time. Uniqueness in law means that all solutions have the same probability distribution on path space. These are different assertions. The Yamada–Watanabe theorem links them: in the classical Brownian SDE setting, weak existence together with pathwise uniqueness yields strong existence and uniqueness in law.
A standard sufficient condition for a global strong solution is global Lipschitz continuity in the state variable, together with a linear growth bound. On each finite time interval, suppose
and
With suitable measurability and a square-integrable initial condition, these conditions give a unique adapted continuous solution with finite second moments. The proof uses successive approximation, the Itô isometry, estimates for martingale maxima, and integral inequalities.
Local Lipschitz continuity generally gives a solution only up to a possible explosion time. The solution is constructed until it exits larger and larger bounded regions; explosion occurs if these exit times approach a finite limit while the state becomes unbounded. A growth condition or a suitable Lyapunov estimate can rule this out. Noise does not automatically prevent explosion: even equations with no noise can diverge in finite time, as does for positive initial data.
The standard sufficient conditions are not necessary. The square-root coefficient in
is not Lipschitz at zero, yet the nonnegative solution is pathwise unique for the usual nonnegative parameters. Conversely, nonuniqueness can already occur without noise: the initial-value problem
allows solutions that remain at zero for an arbitrary time before increasing.
Boundary behavior and the permitted state space are therefore part of a well-posed problem. A differential expression valid only inside a domain may not determine what happens at its boundary. Killing, absorption, reflection, or other boundary mechanisms can define different stochastic models with identical interior coefficients.
Canonical equations and explicit solutions
A basic SDE is Brownian motion with constant drift and volatility:
For deterministic , its distribution at time is normal with mean and variance . Its increments are stationary and independent. It models unconstrained diffusion with systematic transport, but its state can take any real value, which makes it unsuitable for quantities that must remain positive.
The Ornstein–Uhlenbeck process introduces mean reversion:
Multiplication by an integrating factor gives
For deterministic , this is Gaussian, with mean
and variance
Its invariant distribution is normal with mean and variance . When initialized in that distribution, it is a stationary process, with covariance decaying exponentially in the time separation. The model originated in the study of Brownian velocity and remains a standard example of linear stochastic dynamics.
The geometric Brownian motion equation is
For , Itô’s formula applied to the logarithm yields
The correction is essential. Although the expectation is , the almost-sure long-run rate of is . Expected growth and typical trajectory growth can therefore differ substantially. The process remains positive and has lognormal one-time distributions.
The square-root, or Cox–Ingersoll–Ross, diffusion uses
For , it has a gamma invariant distribution with shape and scale . Starting from a positive state, zero is unattainable when ; when , it can be reached. Its transition law is a scaled noncentral chi-squared law.
These models illustrate additive noise, multiplicative noise, mean reversion, and state-dependent boundary behavior. Explicit formulas are exceptional rather than typical, but they supply benchmarks for theoretical results, simulation algorithms, and statistical procedures.
Generators, transition laws, and partial differential equations
Under conditions ensuring a unique solution law, an SDE with deterministic coefficients defines a process with the Markov property: conditional on its present state, its future law does not depend on its earlier history. Time-dependent coefficients generally produce a time-inhomogeneous Markov process. In the time-homogeneous case, the transition operators
form a semigroup, satisfying . This identity is the operator form of the Chapman–Kolmogorov equations.
The infinitesimal generator describes the short-time action of this semigroup. For smooth ,
Itô’s formula implies that
is a local martingale, and a martingale when suitable integrability holds. This property can itself define the process through a martingale problem. That approach is particularly useful when coefficients are irregular or when convergence of approximate models is easier to establish through their generators than through their trajectories.
For a terminal payoff , let
When regularity permits a classical interpretation, solves the backward Kolmogorov equation
The evolution of a probability density is described instead by the Fokker–Planck equation:
This is the forward equation, formally involving the adjoint of the generator. These equations connect SDEs with partial differential equations, including the heat equation as the special case of Brownian diffusion.
The Feynman–Kac formula extends the correspondence to equations containing a potential or discount term. For example, under suitable assumptions, the solution of
has the representation
Source terms can be incorporated through additional time integrals.
Smooth densities do not follow automatically from an SDE. Uniformly nondegenerate diffusion is a standard route to smoothing, while degenerate systems require more delicate conditions. Hörmander-type results show that noise acting in only some directions can still produce smooth densities when the drift and noise vector fields spread randomness through their commutators. The Langevin equation for position and velocity is a central example.
Long-time behavior, boundaries, and rare events
An invariant probability measure satisfies
for suitable test functions. Starting the process with law makes its one-time distributions unchanged by time evolution. Existence, uniqueness, and convergence toward a stationary distribution are separate questions. A process can have several invariant measures, or none, and invariance alone does not ensure that every initial state approaches the same equilibrium.
A common route to long-time control is a Lyapunov function , typically increasing toward infinity, for which
with positive constants and . This inequality expresses an average restoring tendency outside a bounded region. Combined with appropriate accessibility and regularity properties, it can establish recurrence, existence of an invariant law, and quantitative convergence to equilibrium. Ergodic results then justify time-average limits such as
under the theorem’s assumptions.
For the overdamped gradient diffusion
a candidate equilibrium density is
provided the normalizing constant is finite and the dynamics are well defined. The drift drives the state toward low-potential regions, while noise disperses it. With the usual conditions, this diffusion is reversible and satisfies detailed balance. Other diffusions possess stationary probability currents and are not reversible.
Boundary questions concern whether a process reaches or leaves the edge of its state space. In one dimension, scale functions and speed measures organize the classification of boundaries. An absorbing boundary retains the process after arrival; a killed process is sent to a cemetery state; a reflecting boundary pushes it back into the domain. Reflection can be represented schematically by
where the finite-variation process acts only at the boundary. In one-dimensional reflecting models, it is closely related to boundary local time.
Small noise introduces a further timescale. For
trajectories can remain near one local minimum of for long periods before crossing a potential barrier. Under appropriate nondegeneracy assumptions, mean escape times have leading exponential dependence , where is the barrier height. Large deviation theory describes the exponentially small probabilities of atypical paths and supplies the mathematical basis for this metastability analysis. Such rare transitions may dominate system behavior even when short simulations show only local fluctuations.
Numerical approximation and simulation
Most SDEs do not have explicit solutions, so numerical approximation is essential. For a step size , the Euler–Maruyama method updates
where independent Brownian increments satisfy
The simulation must preserve the intended covariance between noise components. Replacing Brownian increments by deterministic perturbations or by random variables with incorrect scaling changes the model.
Two principal error criteria serve different purposes. Strong error compares an approximation with the exact solution driven by the same Brownian path, for example
Weak error compares expectations,
for a suitable class of test functions. Under standard Lipschitz assumptions, Euler–Maruyama has strong order ; with additional smoothness and growth assumptions, it has weak order . These rates concern discretization error, not statistical sampling error.
For a scalar autonomous equation, the Milstein method adds a correction:
Under suitable regularity it has strong order . In multidimensional systems with noncommuting noise vector fields, analogous higher-order schemes require iterated stochastic integrals, including Lévy areas. Their simulation can be substantially more expensive than generating ordinary Gaussian increments.
Standard convergence claims can fail outside their assumptions. Explicit Euler approximations may have divergent moments for equations with superlinearly growing coefficients, even when the exact solution has finite moments. Tamed, truncated, and implicit methods modify the update to control this behavior. Numerical schemes also need to respect structural features: an unconstrained Euler step can make a nonnegative square-root diffusion negative, while a generic method may fail to preserve a manifold constraint or an equilibrium law. Numerical stability and preservation of the state space can therefore matter as much as formal order.
Simulation-based expectation estimates combine time discretization with Monte Carlo sampling. For independent paths with finite payoff variance, sampling error decreases proportionally to the inverse square root of the number of paths. Reducing the step size does not eliminate that error. Multilevel Monte Carlo couples simulations on coarse and fine grids so that most samples are computed cheaply, while fewer fine-grid samples estimate the smaller correction between levels.
In sensitivity calculations and inverse problems, the coupling of random increments is particularly important. Using common Brownian paths across parameter values can greatly reduce the variance of differences, whereas independently resimulating each parameter value can obscure small effects behind simulation noise.
Physical modeling, finance, control, and inference
In statistical physics, the underdamped Langevin equation separates position and momentum :
Here is the mass matrix, the potential energy, a scalar friction coefficient, Boltzmann’s constant, and temperature. The noise amplitude and friction are linked by the fluctuation–dissipation relation. Under appropriate confinement, the invariant density is proportional to the exponential of minus total energy divided by . This model supports molecular simulation and coarse-grained descriptions of thermal motion.
In population biology and chemical kinetics, diffusion models often approximate discrete birth, death, and reaction events. For a reaction channel with state-change vector and propensity , a formal diffusion approximation contributes
The approximation is most defensible when many reaction events occur over the relevant timescale. Near extinction or with small molecule counts, the underlying jump process may be necessary because a continuous approximation can misrepresent discreteness and boundary behavior.
In mathematical finance, geometric Brownian motion underlies the Black–Scholes model introduced in 1973. Under its pricing assumptions, a suitable change of probability measure replaces a traded asset’s physical drift by the risk-free rate, adjusted for payouts when present. Discounted prices then have a martingale structure. The Girsanov theorem formalizes changes of drift under appropriate absolute-continuity and integrability conditions; it does not arbitrarily change the diffusion covariance. Stochastic volatility and jump models extend the framework beyond constant volatility and continuous price paths.
Controlled SDEs allow the drift or diffusion to depend on an intervention :
Optimizing an expected running and terminal cost connects the model with mathematical optimization and dynamic programming. The resulting Hamilton–Jacobi–Bellman equation contains the same second-order diffusion operator as the generator, together with optimization over admissible controls. Partial observation adds a filtering problem: the state must be estimated from noisy measurements before or alongside control decisions.
Statistical inference typically observes the process at discrete times rather than continuously. Exact transition densities, when available, support maximum likelihood estimation. Otherwise, methods use approximate transitions, estimating equations, simulation, or Bayesian inference. Drift and diffusion parameters carry different information: high-frequency observations over a fixed interval can reveal quadratic variation accurately, but generally do not provide unlimited information about drift. Long-span observation is often needed for consistent estimation of drift parameters. Measurement noise, hidden variables, and model misspecification further distinguish inference from merely fitting a numerical trajectory.
Jump noise, infinite-dimensional equations, and geometric extensions
Brownian noise describes continuous fluctuations, but many systems undergo abrupt changes. An SDE with jumps can be written
when the compensated integral is well defined. Here denotes the state just before time , is a Poisson random measure of jumps with marks , and subtracts its intensity measure. More general formulations separate compensated small jumps from uncompensated large jumps. Solutions typically have paths that are right-continuous with left limits.
Jump equations have a nonlocal calculus. Itô’s formula includes the finite difference at each jump rather than only a local second-derivative correction. Their generators therefore contain integral terms, and the corresponding distributional equations are often integro-differential equations. Lévy processes provide a broad class of stationary independent-increment drivers containing Brownian motion and Poisson processes as special cases.
When the state is a spatial field, the model becomes a stochastic partial differential equation. A symbolic example is
The state may live in a Hilbert space or Banach space, and may be an infinite-dimensional noise. Spatial dimension and noise regularity become decisive: a solution may exist only as a generalized function, making nonlinear operations such as powers ill-defined without additional construction. Mild solutions combine an evolution semigroup with a stochastic convolution; variational solutions use appropriate function spaces and duality.
SDEs can also evolve on a manifold. In Stratonovich form,
tangent vector fields give an intrinsically geometric description. Ordinary coordinate transformations preserve this form. Brownian motion on a Riemannian manifold has generator one-half of the Laplace–Beltrami operator. Local Itô coordinates generally introduce additional drift terms reflecting the geometry, so merely projecting Euclidean noise into a coordinate chart is not sufficient.
Finally, rough path theory studies differential equations driven by irregular signals by augmenting a path with iterated-integral information. This makes it possible to define a continuous solution map for appropriately enhanced drivers. Brownian motion admits such enhancements, but different second-order enhancements correspond to different integration conventions. Ordinary uniform convergence of driving paths alone does not capture this information, explaining why apparently similar smooth approximations can produce different limiting equations. Fractional Brownian drivers and path-dependent equations extend the framework further, often beyond the semimartingale or finite-dimensional Markov setting.