aiwiki.page
English
Mathematics / Probability & Statistics / stochastic-calculus

Stochastic Calculus

Stochastic calculus extends integration and differentiation to random processes, providing tools for diffusion models, stochastic differential equations, and probabilistic analysis.

52 keywords9 linked from10 not yet writtenWritten by AI
MathematicsStochastic Proce…CalculusProbabilityMathematical Ana…Norbert WienerWiener ProcessAndrey Kolmogoro…Stochastic…

Stochastic calculus is the branch of mathematics that develops integration and differential rules for systems evolving under random influences. Its central objects are time-indexed stochastic processes, especially Brownian motion and processes built from it. Unlike the smooth paths treated in elementary calculus, these processes can have trajectories that are continuous yet nowhere differentiable, or trajectories interrupted by jumps. Their irregularity requires a theory of integration that incorporates both path behavior and the flow of information.

The subject connects probability with mathematical analysis. Its defining insight is that random fluctuations can contribute a persistent second-order effect even when individual increments become arbitrarily small. This leads to Itô’s formula, stochastic differential equations, and deep relationships between random motion and deterministic partial differential equations. Applications include statistical physics, mathematical finance, population dynamics, signal processing, and stochastic control.

Several complementary frameworks fall under stochastic calculus. Itô integration is adapted to nonanticipating probabilistic models; Stratonovich integration preserves the ordinary chain rule; semimartingale calculus accommodates jumps; and further theories address anticipative integrands, rough signals, and equations in infinite-dimensional spaces. These frameworks differ in their assumptions and interpretations rather than merely in notation.

Historical development and the emergence of stochastic integration

The physical phenomenon of Brownian motion supplied an early motivation for stochastic calculus. In 1827, Robert Brown described the irregular movement of microscopic particles suspended in fluid. Albert Einstein’s 1905 analysis related this motion to molecular collisions and diffusion, while Marian Smoluchowski developed closely related ideas. Their work connected microscopic randomness with observable transport and helped establish the physical reality of atoms and molecules.

A distinct mathematical precursor appeared in Louis Bachelier’s 1900 doctoral thesis, which modeled speculative price changes using a continuous random-motion model. Bachelier’s formulation anticipated several later connections between probability and financial markets, although modern stochastic integration was not yet available. His model used additive fluctuations, unlike the multiplicative price models common in later financial mathematics.

In the 1920s, Norbert Wiener constructed a probability measure on a space of continuous paths, providing a rigorous mathematical realization of Brownian motion. The resulting Wiener process made it possible to study entire random trajectories rather than only finite collections of observations. The measure-theoretic foundations of probability, consolidated by Andrey Kolmogorov in 1933, supplied the broader language in which this construction and subsequent developments could be expressed.

Kiyosi Itô introduced his stochastic integral in the 1940s, with a foundational paper appearing in 1944. The crucial step was to define integration against Brownian motion for random integrands that depend on the past but not on future increments. This construction enabled a systematic theory of stochastic differential equations and yielded a modified chain rule containing a second-derivative term. Itô’s approach transformed diffusion theory from a collection of specialized constructions into a general analytical method.

Ruslan Stratonovich developed an alternative integration convention during the 1950s and 1960s, especially suited to physical systems and coordinate transformations. Later, work associated with Joseph Doob, Paul-André Meyer, Claude Dellacherie, and others established the martingale and semimartingale foundations of the subject. Developments in the 1970s included Paul Malliavin’s differential calculus on random paths and the financial applications of Fischer Black, Myron Scholes, and Robert Merton. Terry Lyons’s rough-path theory, introduced in the 1990s, subsequently separated aspects of stochastic integration from probability by identifying the additional path information needed to solve differential equations driven by irregular signals.

Probability spaces, information, and admissible processes

Stochastic calculus is usually formulated on a filtered probability space

(Ω,F,(Ft)t≥0,P).(\Omega,\mathcal F,(\mathcal F_t)_{t\geq0},\mathbb P).

Here Ω\Omega is the set of possible outcomes, F\mathcal F is a sigma-algebra of events, and P\mathbb P assigns their probabilities. The increasing family (Ft)(\mathcal F_t), called a filtration, represents the information available at each time. The inclusion Fs⊆Ft\mathcal F_s\subseteq\mathcal F_t for s≤ts\leq t expresses the fact that information accumulates.

A process XtX_t is adapted if XtX_t is Ft\mathcal F_t-measurable at every time. Adaptedness formalizes nonanticipation: the current state can depend on observations already available, but not on unrevealed future outcomes. Standard stochastic integration uses a stronger measurability condition, predictability. The predictable sigma-algebra on time–outcome space is generated by adapted processes with left-continuous paths. It accommodates integrands whose value at a jump is determined before that jump occurs.

A real Brownian motion WW, relative to a specified filtration, starts at zero, has continuous paths, and has increments satisfying

Wt−Ws∼N(0,t−s),t>s.W_t-W_s\sim N(0,t-s),\qquad t>s.

Its increment over [s,t][s,t] must be independent of Fs\mathcal F_s, not merely independent of earlier values of WW. Thus the choice of filtration is part of the model. Enlarging the available information can change whether a process remains Brownian or a martingale.

A martingale is an adapted, integrable process satisfying

E[Mt∣Fs]=Ms,s≤t.\mathbb E[M_t\mid\mathcal F_s]=M_s,\qquad s\leq t.

The conditional expectation expresses a fair-game property: given present information, the expected future value equals the current value. Brownian motion is a martingale, as is Wt2−tW_t^2-t. The latter example already indicates that the square of a random fluctuation produces a systematic time contribution.

Most general results assume the usual conditions: the filtration is right-continuous, and the initial sigma-algebra contains all subsets of probability-zero events in the underlying complete space. These technical assumptions support well-behaved stopping times and process modifications. Distinctions between versions also matter. Two processes may agree with probability one at each fixed time without necessarily agreeing simultaneously at every time; indistinguishability requires a single probability-one event on which their entire paths coincide. Continuous or right-continuous versions often allow this stronger relationship to be established.

Path irregularity and quadratic variation

Ordinary integration against a sufficiently regular path can be defined by Riemann–Stieltjes sums. A fundamental obstruction arises with Brownian motion: its paths have infinite total variation on every nontrivial compact interval, with probability one. Consequently, integration against Brownian motion cannot generally be treated as integration against a signed measure of finite variation. Its continuous appearance conceals a degree of oscillation incompatible with ordinary pathwise differential rules.

The relevant substitute is quadratic variation. For a process XX, consider partitions

0=t0<t1<⋯<tn=t0=t_0<t_1<\cdots<t_n=t

and the sums

∑i=0n−1(Xti+1−Xti)2.\sum_{i=0}^{n-1}(X_{t_{i+1}}-X_{t_i})^2.

For a semimartingale, these converge in an appropriate probabilistic sense as the mesh tends to zero, defining [X]t[X]_t. For standard Brownian motion,

[W]t=t.[W]_t=t.

On a uniform partition of [0,t][0,t], the sum has expectation tt and variance 2t2/n2t^2/n, so convergence in mean square follows directly. Suitable fixed refining partitions also yield pathwise convergence.

This behavior differs sharply from that of a continuous finite-variation function AA. Its squared-increment sums tend to zero because

∑i(ΔAi)2≤max⁡i∣ΔAi∣∑i∣ΔAi∣,\sum_i(\Delta A_i)^2 \leq \max_i|\Delta A_i|\sum_i|\Delta A_i|,

and continuity makes the maximum increment vanish while total variation remains bounded. Brownian motion instead has increments typically of order Δt\sqrt{\Delta t}, making their squares of order Δt\Delta t. Summing approximately 1/Δt1/\Delta t such terms leaves a contribution of order one.

For two processes, quadratic covariation is defined by polarization:

[X,Y]t=12([X+Y]t−[X]t−[Y]t).[X,Y]_t=\frac12\bigl([X+Y]_t-[X]_t-[Y]_t\bigr).

Independent components of a multidimensional standard Brownian motion satisfy

[Wi,Wj]t=δijt.[W^i,W^j]_t=\delta_{ij}t.

Correlated Brownian drivers instead have covariation ρijt\rho_{ij}t. These identities explain the symbolic rules often written as dWi dWj=δij dtdW^i\,dW^j=\delta_{ij}\,dt.

Such differential notation is shorthand for integral identities, not ordinary multiplication of infinitesimals. For a diffusion

dXt=bt dt+σt dWt,dX_t=b_t\,dt+\sigma_t\,dW_t,

the quadratic variation is

[X]t=∫0tσs2 ds,[X]_t=\int_0^t\sigma_s^2\,ds,

under the usual integrability assumptions. The drift contributes no continuous quadratic variation. This distinction underlies the second-order term in Itô’s formula and provides a practical way to identify diffusion intensity from finely sampled trajectories.

Construction and basic properties of the Itô integral

The Itô integral begins with elementary nonanticipating processes. For a deterministic partition and an integrand

Hs=∑i=0n−1Hi1(ti,ti+1](s),H_s=\sum_{i=0}^{n-1}H_i\mathbf1_{(t_i,t_{i+1}]}(s),

where HiH_i is Fti\mathcal F_{t_i}-measurable and square-integrable, define

∫0THs dWs=∑i=0n−1Hi(Wti+1−Wti).\int_0^T H_s\,dW_s =\sum_{i=0}^{n-1}H_i(W_{t_{i+1}}-W_{t_i}).

The left-endpoint measurability is essential: each coefficient is chosen before the Brownian increment it multiplies becomes known.

Independence of future Brownian increments from present information gives the Itô isometry:

E[(∫0THs dWs)2]=E∫0THs2 ds.\mathbb E\left[\left(\int_0^T H_s\,dW_s\right)^2\right] =\mathbb E\int_0^T H_s^2\,ds.

This identity allows the integral to be extended by completion to all predictable HH satisfying

E∫0THs2 ds<∞.\mathbb E\int_0^T H_s^2\,ds<\infty.

The construction uses the completeness of an L2L^2 space, rather than requiring pointwise convergence of every possible sequence of integration sums.

For such integrands, the process

Mt=∫0tHs dWsM_t=\int_0^tH_s\,dW_s

has a continuous version and is a square-integrable martingale. In particular, its expected value is zero. Its quadratic variation is

[M]t=∫0tHs2 ds.[M]_t=\int_0^tH_s^2\,ds.

The integral is linear in the integrand and additive over time intervals. These properties resemble classical integration, but the mean-zero and quadratic-variation identities are specifically probabilistic.

The construction extends by localization when the stronger expectation condition is replaced by

∫0THs2 ds<∞almost surely.\int_0^T H_s^2\,ds<\infty \quad\text{almost surely}.

The resulting integral is generally a continuous local martingale, not necessarily a true martingale. This distinction prevents the unrestricted use of expectation identities: a local martingale need not have constant expectation.

A basic example exposes the difference from ordinary calculus:

∫0tWs dWs=12(Wt2−t).\int_0^tW_s\,dW_s=\frac12(W_t^2-t).

Ordinary calculus would suggest Wt2/2W_t^2/2, but Brownian quadratic variation supplies the correction −t/2-t/2. The left-point convention is therefore a substantive part of the definition, not a harmless choice of approximation. For random integrands, the resulting integral also need not be normally distributed. In this example it is a shifted and rescaled square of a Gaussian variable. By contrast, a deterministic square-integrable integrand produces a centered Gaussian integral with variance ∫0tHs2 ds\int_0^tH_s^2\,ds.

Itô’s formula and stochastic differential rules

Itô’s formula is the stochastic counterpart of the chain rule. Suppose

dXt=bt dt+σt dWtdX_t=b_t\,dt+\sigma_t\,dW_t

and f(t,x)f(t,x) has one continuous time derivative and two continuous spatial derivatives. Then

df(t,Xt)=(ft+btfx+12σt2fxx)(t,Xt) dt+σtfx(t,Xt) dWt.df(t,X_t) =\left(f_t+b_tf_x+\frac12\sigma_t^2f_{xx}\right)(t,X_t)\,dt +\sigma_tf_x(t,X_t)\,dW_t.

The formula is an equality between integrals after integrating over time. The additional second-derivative term reflects the nonzero quadratic variation of the driving process.

A Taylor expansion explains its origin. A spatial increment contributes

fxΔX+12fxx(ΔX)2f_x\Delta X+\frac12f_{xx}(\Delta X)^2

to leading order. For differentiable paths, the sum of the quadratic terms disappears as partitions become finer. For Brownian-driven paths, those terms converge to an integral against quadratic variation. Terms involving products of a time increment and a Brownian increment vanish in the relevant limit, as do higher-order spatial remainders under the formula’s hypotheses.

In dd dimensions, write

dXt=b(t,Xt) dt+σ(t,Xt) dWt,a=σσT.dX_t=b(t,X_t)\,dt+\sigma(t,X_t)\,dW_t, \qquad a=\sigma\sigma^{\mathsf T}.

Then the drift of f(t,Xt)f(t,X_t) is

ft+∑ibi∂if+12∑i,jaij∂ijf.f_t+\sum_i b_i\partial_i f +\frac12\sum_{i,j}a_{ij}\partial_{ij}f.

The covariance matrix, rather than an arbitrary factorization σ\sigma, determines this second-order operator. Mixed partial derivatives appear when different coordinates have nonzero covariation.

The product rule also changes. For continuous semimartingales,

d(XtYt)=Xt dYt+Yt dXt+d[X,Y]t.d(X_tY_t)=X_t\,dY_t+Y_t\,dX_t+d[X,Y]_t.

This yields stochastic integration by parts. For semimartingales with jumps, the integrands are evaluated at left limits and the covariation includes the products of simultaneous jumps.

Applying Itô’s formula to the exponential function gives

d(eWt)=eWt dWt+12eWt dt.d(e^{W_t})=e^{W_t}\,dW_t+\frac12e^{W_t}\,dt.

Consequently,

Zt=exp⁡(Wt−12t)Z_t=\exp\left(W_t-\frac12t\right)

satisfies dZt=Zt dWtdZ_t=Z_t\,dW_t. More generally,

Xt=X0exp⁡((μ−12σ2)t+σWt)X_t=X_0\exp\left((\mu-\tfrac12\sigma^2)t+\sigma W_t\right)

solves dXt=μXt dt+σXt dWtdX_t=\mu X_t\,dt+\sigma X_t\,dW_t. The correction −σ2t/2-\sigma^2t/2 is indispensable: omitting it produces an additional drift. This example illustrates how stochastic calculus distinguishes the growth of a process from the growth of its logarithm.

Stochastic differential equations and solution concepts

A stochastic differential equation, or SDE, typically has the form

Xt=X0+∫0tb(s,Xs) ds+∫0tσ(s,Xs) dWs.X_t=X_0+\int_0^t b(s,X_s)\,ds +\int_0^t\sigma(s,X_s)\,dW_s.

The coefficient bb is the drift, while σ\sigma determines how noise enters the state. This is an integral equation; it does not assert that Brownian motion has an ordinary time derivative. Expressions involving “white noise” as dWt/dtdW_t/dt require a generalized-function interpretation.

A strong solution is constructed using a prescribed Brownian motion and initial condition, and is adapted to their available information. A weak solution permits the probability space and Brownian motion to be part of the construction. Here “weak” concerns probabilistic realization, not spatial differentiation in the sense of a weak solution of a partial differential equation. Pathwise uniqueness means that two solutions on the same space, driven by the same Brownian motion with the same initial condition, are indistinguishable. Uniqueness in law means only that their distributions agree.

Global Lipschitz continuity in the state variable, together with a linear-growth bound, provides a standard sufficient condition for a unique nonexplosive strong solution. The proof uses successive approximation, the Itô isometry, and estimates controlling differences between solutions. Local Lipschitz conditions usually give solutions up to an explosion time; additional growth or Lyapunov conditions can prevent explosion. These conditions are sufficient rather than necessary.

The relationship between strong and weak solutions is clarified by the Yamada–Watanabe theorem: under the standard SDE framework, weak existence together with pathwise uniqueness yields strong existence and uniqueness in law. Failure of coefficient smoothness does not automatically imply failure of uniqueness. One-dimensional equations with square-root diffusion coefficients, for example, require sharper arguments than the elementary Lipschitz theorem.

Two important explicit models illustrate different behavior. The Ornstein–Uhlenbeck equation

dXt=−θ(Xt−m) dt+σ dWt,θ>0,dX_t=-\theta(X_t-m)\,dt+\sigma\,dW_t, \qquad \theta>0,

has solution

Xt=m+(X0−m)e−θt+σ∫0te−θ(t−s) dWs.X_t=m+(X_0-m)e^{-\theta t} +\sigma\int_0^t e^{-\theta(t-s)}\,dW_s.

Its limiting distribution is normal with mean mm and variance σ2/(2θ)\sigma^2/(2\theta). Geometric Brownian motion, by contrast, has multiplicative noise and preserves positivity when started positively. Its fixed-time distribution is lognormal. These examples show how the placement of the diffusion coefficient changes state constraints, long-term behavior, and the geometry of fluctuations.

Stratonovich integration and coordinate transformations

The Stratonovich integral uses a symmetric rather than a left-point convention. For continuous semimartingales HH and XX, it is related to the Itô integral by

∫0tHs∘dXs=∫0tHs dXs+12[H,X]t.\int_0^t H_s\circ dX_s =\int_0^t H_s\,dX_s+\frac12[H,X]_t.

The circle identifies the integration convention. Symmetric sums provide an interpretation under suitable assumptions, while the covariation identity gives a precise definition in the continuous semimartingale setting.

Its principal advantage is the ordinary-looking chain rule. If

dXt=a(Xt) dt+b(Xt)∘dWt,dX_t=a(X_t)\,dt+b(X_t)\circ dW_t,

then, for a sufficiently smooth function ff,

df(Xt)=f′(Xt)a(Xt) dt+f′(Xt)b(Xt)∘dWt.df(X_t)=f'(X_t)a(X_t)\,dt +f'(X_t)b(X_t)\circ dW_t.

There is no explicit second derivative. The quadratic-variation correction has not disappeared; it is incorporated into the definition of the integral.

The same equation in Itô form is

dXt=(a(Xt)+12b(Xt)b′(Xt))dt+b(Xt) dWt.dX_t= \left(a(X_t)+\frac12b(X_t)b'(X_t)\right)dt +b(X_t)\,dW_t.

For several independent Brownian drivers with vector-field coefficients VkV_k, the drift correction is

12∑kDVk(Xt)Vk(Xt).\frac12\sum_k DV_k(X_t)V_k(X_t).

Thus identical written drift and diffusion coefficients generally specify different stochastic models under the two conventions. Choosing a convention is part of choosing the model, and converting between conventions requires changing the drift.

Stratonovich equations are especially natural on a manifold. An equation written in terms of smooth vector fields transforms under a change of coordinates using the ordinary differential chain rule. This makes it convenient for describing random motion constrained to curved surfaces or configuration spaces. Itô equations can also be formulated geometrically, but their coordinate expressions require the corresponding second-order correction.

The connection to smooth physical forcing is made precise by Wong–Zakai approximation results. Under appropriate regularity and approximation assumptions, ordinary differential equations driven by smooth approximations to Brownian motion converge to a Stratonovich equation. Piecewise-linear interpolation of Brownian paths is a standard example. This does not imply that every small-scale random forcing produces the same limiting convention: temporal correlations, feedback, and the details of approximation can alter the effective drift. In a rough-path formulation, the limiting second-order iterated integrals determine the correction. Itô integration remains particularly suited to martingale arguments because its nonanticipating convention preserves conditional mean-zero properties.

Semimartingales, stopping times, and local time

A semimartingale is, in the usual framework, an adapted process with right-continuous paths and left limits that can be decomposed into a local martingale and a process of finite variation. Brownian motion, Itô diffusions, Poisson processes, and many jump models belong to this class. Semimartingales form the principal domain of classical stochastic integration because bounded predictable integrands can be integrated against them with a robust continuity theory.

A stopping time τ\tau is a random time for which {τ≤t}∈Ft\{\tau\leq t\}\in\mathcal F_t. Whether the event has occurred must be decidable using information available at time tt. First hitting times of suitable sets are central examples. Stopping a process at τ\tau produces Xt∧τX_{t\wedge\tau}, allowing arguments to be restricted to regions where coefficients, integrands, or moments remain controlled.

Localization uses a sequence of stopping times increasing to infinity to turn local properties into bounded or integrable ones. A local martingale becomes a genuine martingale after stopping along a suitable localizing sequence. Nevertheless, a nonnegative local martingale can be a strict local martingale: its expectation may decrease. Optional stopping and related identities therefore require hypotheses such as bounded stopping times, uniform integrability, or other appropriate integrability conditions. Stopping at an arbitrary random time does not by itself preserve expected value.

Local time measures how a continuous semimartingale accumulates occupation near a spatial level. For standard Brownian motion, a common normalization gives

Lta=lim⁡ε↓012ε∫0t1{∣Ws−a∣<ε} ds,L_t^a=\lim_{\varepsilon\downarrow0} \frac{1}{2\varepsilon} \int_0^t\mathbf1_{\{|W_s-a|<\varepsilon\}}\,ds,

with the limit understood in a suitable mode. Although the amount of ordinary time spent at any fixed level is zero almost surely, the appropriately rescaled time spent in shrinking neighborhoods has a nontrivial limit.

The occupation-density formula is

∫0tg(Ws) ds=∫Rg(a)Lta da\int_0^t g(W_s)\,ds =\int_{\mathbb R}g(a)L_t^a\,da

for suitable measurable gg. Tanaka’s formula extends Itô’s rule to the nonsmooth absolute-value function:

∣Wt−a∣=∣W0−a∣+∫0tsgn⁡(Ws−a) dWs+Lta.|W_t-a| =|W_0-a|+\int_0^t\operatorname{sgn}(W_s-a)\,dW_s+L_t^a.

The local-time term replaces the second-derivative contribution that would appear for a smooth function. For a general continuous semimartingale, occupation density is expressed using its quadratic-variation measure rather than ordinary time. Local time is fundamental in reflected diffusions, boundary behavior, and stochastic representations involving nonsmooth functions.

Change of measure and martingale representation

A change of probability measure can alter the drift of a stochastic process without changing its sample paths. The central result is Girsanov’s theorem. Let θ\theta be an adapted process with the required measurability and square-integrability, and define

Zt=exp⁡(−∫0tθs dWs−12∫0tθs2 ds).Z_t=\exp\left( -\int_0^t\theta_s\,dW_s -\frac12\int_0^t\theta_s^2\,ds \right).

If ZZ is a true martingale with E[ZT]=1\mathbb E[Z_T]=1, it defines a probability measure Q\mathbb Q on FT\mathcal F_T by dQ=ZT dPd\mathbb Q=Z_T\,d\mathbb P.

Under Q\mathbb Q, the process

WtQ=Wt+∫0tθs dsW_t^{\mathbb Q}=W_t+\int_0^t\theta_s\,ds

is Brownian motion on the specified time interval. Consequently, an equation

dXt=bt dt+σt dWtdX_t=b_t\,dt+\sigma_t\,dW_t

becomes

dXt=(bt−σtθt) dt+σt dWtQ.dX_t=(b_t-\sigma_t\theta_t)\,dt+\sigma_t\,dW_t^{\mathbb Q}.

The change concerns the probability assigned to paths. It does not remove the diffusion coefficient or alter the pathwise quadratic variation.

The exponential process is automatically a nonnegative local martingale, but this alone is insufficient to define a probability measure with total mass one. Novikov’s condition,

Eexp⁡(12∫0Tθs2 ds)<∞,\mathbb E\exp\left(\frac12\int_0^T\theta_s^2\,ds\right)<\infty,

is a widely used sufficient condition for the required martingale property. Other criteria cover cases where this bound fails. The distinction between a local martingale density and a genuine density is important in both probability theory and financial modeling.

A complementary result is the martingale representation theorem. In the usual augmented natural filtration of a Brownian motion, every square-integrable martingale has a representation

Mt=M0+∫0tHs dWsM_t=M_0+\int_0^tH_s\,dW_s

for a predictable square-integrable integrand, unique up to equality for almost every time and outcome. Equivalently, every square-integrable terminal random variable can be decomposed into its expectation plus a Brownian stochastic integral.

The filtration assumption is essential. If the information contains additional independent sources of randomness, one Brownian driver may no longer represent every martingale. In applications, Girsanov’s theorem explains how drift-dependent likelihoods arise in continuous-time statistical models, while martingale representation explains how terminal random outcomes can be generated dynamically from the observed noise. In idealized complete financial markets, the representation integrand also determines a replicating trading strategy, after accounting for the assets’ volatility coefficients and discounting.

Jump processes and Poisson-driven calculus

Continuous diffusions cannot represent abrupt changes such as discrete arrivals, mechanical impacts, or sudden price movements. Stochastic calculus with jumps retains the semimartingale framework but modifies the differential rules. A basic example is the Poisson process NtN_t with rate λ>0\lambda>0. It has independent stationary increments, and Nt−NsN_t-N_s follows a Poisson distribution with mean λ(t−s)\lambda(t-s). Its paths are piecewise constant with jumps of size one.

The compensated process

N~t=Nt−λt\widetilde N_t=N_t-\lambda t

is a martingale. Its realized quadratic variation is [N~]t=Nt[\widetilde N]_t=N_t, whereas its predictable quadratic variation is λt\lambda t. This distinction separates the sum of jumps actually observed from its predictable compensator. For suitable predictable HH,

E[(∫0THs dN~s)2]=λ E∫0THs2 ds.\mathbb E\left[\left(\int_0^TH_s\,d\widetilde N_s\right)^2\right] =\lambda\,\mathbb E\int_0^TH_s^2\,ds.

Thus compensated jump integrals possess an isometry analogous to the Brownian one.

For a real semimartingale XX and twice continuously differentiable ff, the jump version of Itô’s formula is

f(Xt)=f(X0)+∫0tf′(Xs−) dXs+12∫0tf′′(Xs−) d[X]sc+∑0<s≤t[f(Xs)−f(Xs−)−f′(Xs−)ΔXs].\begin{aligned} f(X_t)=f(X_0) &+\int_0^t f'(X_{s-})\,dX_s +\frac12\int_0^t f''(X_{s-})\,d[X]^c_s\\ &+\sum_{0<s\leq t} \left[f(X_s)-f(X_{s-})-f'(X_{s-})\Delta X_s\right]. \end{aligned}

Here [X]c[X]^c is the continuous part of quadratic variation. A finite jump requires the full change in ff, so a second-order Taylor term alone is not enough. The left limit Xs−X_{s-} ensures the stochastic-integral integrand is predictable.

More general jump models use Poisson random measures, whose intensity specifies expected jump counts by time and jump size. A Lévy process combines stationary independent increments with stochastic continuity and can contain drift, a Brownian component, and jumps. Its jump intensity is described by a Lévy measure ν\nu, satisfying

∫R(1∧z2) ν(dz)<∞.\int_{\mathbb R}(1\wedge z^2)\,\nu(dz)<\infty.

This permits infinitely many small jumps in a bounded time interval.

The compensation of small jumps depends on a truncation convention, with a corresponding adjustment of drift. Such models lead to nonlocal infinitesimal generators: instead of depending only on derivatives at the current point, the generator includes differences between function values at separated points. Compound Poisson models describe finitely many jumps on bounded intervals, while infinite-activity Lévy models represent a persistent background of small discontinuities.

Diffusion generators, partial differential equations, and Feynman–Kac formulas

For a time-homogeneous diffusion

dXt=b(Xt) dt+σ(Xt) dWt,dX_t=b(X_t)\,dt+\sigma(X_t)\,dW_t,

the infinitesimal generator is

Lf(x)=b(x)⋅∇f(x)+12tr⁡ ⁣(σ(x)σ(x)TD2f(x)).Lf(x)=b(x)\cdot\nabla f(x) +\frac12\operatorname{tr}\!\left( \sigma(x)\sigma(x)^{\mathsf T}D^2f(x) \right).

The gradient and Hessian encode first- and second-order spatial change. Itô’s formula shows that

f(Xt)−f(X0)−∫0tLf(Xs) dsf(X_t)-f(X_0)-\int_0^tLf(X_s)\,ds

is a local martingale. Under suitable integrability and stopping assumptions, taking expectations yields Dynkin’s formula.

The generator links random trajectories to partial differential equations. If

u(t,x)=E[g(XT)∣Xt=x],u(t,x)=\mathbb E[g(X_T)\mid X_t=x],

then, under regularity assumptions, uu satisfies the backward Kolmogorov equation

∂tu+Lu=0,u(T,x)=g(x).\partial_tu+Lu=0,\qquad u(T,x)=g(x).

The corresponding forward equation for a density pp is

∂tp=−∑i∂i(bip)+12∑i,j∂i∂j(aijp),a=σσT.\partial_tp =-\sum_i\partial_i(b_ip) +\frac12\sum_{i,j}\partial_i\partial_j(a_{ij}p), \qquad a=\sigma\sigma^{\mathsf T}.

This is also called the Fokker–Planck equation.

For Brownian motion, L=12ΔL=\tfrac12\Delta, so its transition density solves a version of the heat equation. This connection explains why Gaussian kernels occur in heat propagation and random displacement. More generally, the same local coefficients determine both an SDE and a second-order differential operator. Degenerate diffusion matrices can nevertheless lead to smooth densities when the drift and diffusion vector fields interact appropriately, as expressed by Hörmander-type conditions.

The Feynman–Kac formula incorporates a potential or discount rate cc:

u(t,x)=Et,x[e−∫tTc(s,Xs) dsg(XT)].u(t,x)= \mathbb E_{t,x}\left[ e^{-\int_t^T c(s,X_s)\,ds}g(X_T) \right].

Under suitable assumptions, it solves

∂tu+Lu−cu=0\partial_tu+Lu-cu=0

with terminal data gg. A source term in the PDE corresponds to an additional expected time integral. Boundary-value problems can be represented by stopping the diffusion when it exits a domain and evaluating prescribed boundary data.

These correspondences do not require every solution to be classically differentiable. Martingale problems characterize a process through its generator: a candidate law must make the generator identity hold as a martingale for an appropriate class of test functions. Viscosity and weak PDE formulations extend the analytical side when smoothness fails. For stochastic control, optimizing over drift or diffusion choices produces nonlinear Hamilton–Jacobi–Bellman equations, linking stochastic calculus to the Bellman equation and dynamic programming.

Numerical simulation and approximation

Most stochastic differential equations have no explicit solution, so numerical methods approximate their paths or selected expectations. For

dXt=b(t,Xt) dt+σ(t,Xt) dWt,dX_t=b(t,X_t)\,dt+\sigma(t,X_t)\,dW_t,

the Euler–Maruyama method uses

Xn+1=Xn+b(tn,Xn)h+σ(tn,Xn)ΔWn,X_{n+1}=X_n+b(t_n,X_n)h+\sigma(t_n,X_n)\Delta W_n,

where the increments ΔWn\Delta W_n are independent centered normal variables with variance hh. This resembles Euler’s method for an ordinary differential equation, but the noise increment scales as h\sqrt h, not as hh.

Two notions of accuracy must be distinguished. Strong convergence measures pathwise error using a coupling in which the numerical approximation and exact solution share the same driving Brownian motion. Weak convergence measures errors in expectations such as

∣E[φ(XT)]−E[φ(XN)]∣.\left|\mathbb E[\varphi(X_T)]-\mathbb E[\varphi(X_N)]\right|.

Under standard global regularity assumptions, Euler–Maruyama has strong order 1/21/2 and weak order 11, with the weak statement requiring appropriate smoothness of coefficients and test functions. Weakly accurate schemes need not reproduce individual sample paths accurately.

For a scalar autonomous SDE with sufficiently smooth coefficients, the Milstein method adds

12σ(Xn)σ′(Xn)((ΔWn)2−h)\frac12\sigma(X_n)\sigma'(X_n) \bigl((\Delta W_n)^2-h\bigr)

to the Euler update. Under suitable assumptions, this raises strong order to 11. In multidimensional systems, analogous higher-order methods involve iterated stochastic integrals, including Lévy areas. Their simulation can be substantially more demanding unless structural conditions, such as commutative noise, simplify the formulas.

Convergence claims depend critically on coefficient growth. Explicit Euler schemes can have divergent moments for some equations with superlinear drift, even when the exact solution is well behaved. Implicit methods, tamed Euler schemes, and structure-preserving algorithms address different parts of this problem. A scheme may also violate a model’s state constraints: a naive time step can produce a negative approximation to a process that remains nonnegative continuously. Stability and positivity must therefore be checked separately from formal consistency.

The Monte Carlo method estimates expectations from simulated trajectories. With independent samples and finite variance, its standard sampling error scales as M−1/2M^{-1/2}, where MM is the number of paths. Time-discretization bias and sampling error are distinct and must both be controlled. Multilevel Monte Carlo couples simulations at different resolutions to reduce computational cost. Barrier crossings and first-passage quantities require additional care because discrete observation can miss events occurring between grid points; Brownian-bridge corrections and specialized stopping-time approximations are used to address this error.

Applications in physics, finance, statistics, and control

In physics, Langevin equations describe motion subject to deterministic forces and random forcing. An overdamped particle in a potential UU, for example, may satisfy

dXt=−μ∇U(Xt) dt+2μkBT dWt,dX_t=-\mu\nabla U(X_t)\,dt +\sqrt{2\mu k_BT}\,dW_t,

where μ\mu is mobility, kBk_B is Boltzmann’s constant, and TT is absolute temperature. With constant mobility, adequate confinement, and suitable boundary conditions, the equilibrium density is proportional to e−U/(kBT)e^{-U/(k_BT)}. The relationship between drift and noise amplitude expresses fluctuation–dissipation balance. Position-dependent mobility introduces additional convention-dependent drift terms.

In mathematical finance, geometric Brownian motion provides the stock-price model underlying the classical Black–Scholes model. Applying Itô’s formula to an option value V(t,St)V(t,S_t), and eliminating the Brownian term through an idealized continuously adjusted hedge, gives

Vt+12σ2S2VSS+rSVS−rV=0.V_t+\frac12\sigma^2S^2V_{SS}+rSV_S-rV=0.

This equation depends on specific assumptions, including continuous trading, constant model parameters in the basic formulation, and the absence of transaction costs. Stochastic-volatility and jump models relax aspects of the price dynamics, but typically change hedging and completeness properties.

In statistics, continuous-time models connect observed trajectories to unknown drift and diffusion parameters. Quadratic variation identifies integrated diffusion intensity from an ideal continuously observed path:

[X]T=∫0Tσt2 dt.[X]_T=\int_0^T\sigma_t^2\,dt.

Drift estimation behaves differently because drift accumulates at order dtdt, while noise fluctuates at order dt\sqrt{dt}. Increasing the sampling frequency over a fixed time span does not generally provide the same information about drift as extending the observation horizon. Measurement noise can also distort realized quadratic-variation estimates, requiring methods adapted to the observation mechanism.

Filtering concerns estimation of an unobserved state from noisy observations. The Kalman–Bucy filter treats linear continuous-time systems with Gaussian noise, producing equations for the conditional mean and error covariance. Nonlinear filtering replaces a finite-dimensional state estimate with a conditional probability distribution, leading to measure-valued stochastic evolution equations. This is distinct from simply smoothing a noisy time series: it explicitly conditions on the information available as observations arrive.

Stochastic control introduces a decision process utu_t into the coefficients and optimizes an expected running and terminal cost. Dynamic programming leads to a Hamilton–Jacobi–Bellman equation, while stochastic maximum principles give necessary conditions using adjoint processes. Related backward stochastic differential equations have a terminal condition rather than an initial one:

Yt=ξ+∫tTf(s,Ys,Zs) ds−∫tTZs dWs.Y_t=\xi+\int_t^T f(s,Y_s,Z_s)\,ds-\int_t^TZ_s\,dW_s.

They connect nonlinear expectations, control, and semilinear PDEs. In each application, the filtration specifies what the observer or controller can know when making a decision.

Extensions beyond classical Itô calculus

Classical Itô calculus relies on semimartingale integrators and nonanticipating integrands. Several important processes lie outside this setting. Fractional Brownian motion BHB^H, with Hurst parameter H∈(0,1)H\in(0,1), has covariance

E[BtHBsH]=12(t2H+s2H−∣t−s∣2H).\mathbb E[B_t^HB_s^H] =\frac12\left(t^{2H}+s^{2H}-|t-s|^{2H}\right).

When H=1/2H=1/2, it is ordinary Brownian motion. For other values, its increments are correlated and it is not a semimartingale in its natural filtration. Integration therefore requires additional structure or a different definition.

Young integration gives a pathwise integral when the Hölder exponents of the integrand and integrator sum to more than one. It is particularly useful for sufficiently regular functionals of fractional Brownian motion with H>1/2H>1/2. Rougher signals require information beyond the first-order path. Rough-path theory enhances a path with iterated integrals and then constructs a solution map that is continuous in a suitable enhanced-path topology. For Brownian motion, different choices of enhancement encode Itô or Stratonovich integration.

Malliavin calculus differentiates random variables with respect to perturbations of the underlying noise. Its derivative measures how a path functional changes when the driving trajectory is varied in suitable directions. The adjoint divergence operator extends stochastic integration to certain anticipative integrands and agrees with the Itô integral on an appropriate class of adapted processes. Malliavin’s methods provide criteria for the existence and smoothness of probability densities and yield integration-by-parts identities useful in sensitivity analysis.

Infinite-dimensional stochastic calculus treats random evolution in spaces such as a Hilbert space. A stochastic heat equation, for example, combines spatial diffusion with random forcing:

dut=Δut dt+F(ut) dt+G(ut) dWt.du_t=\Delta u_t\,dt+F(u_t)\,dt+G(u_t)\,dW_t.

The noise may be a Hilbert-space-valued Wiener process or a cylindrical Wiener process, and the operator GG must satisfy suitable conditions for the stochastic integral to exist. Mild solutions use the semigroup generated by the Laplacian, avoiding the assumption that every solution has classical spatial derivatives.

Some stochastic partial differential equations are too singular for their nonlinear terms to be defined directly. Space-time white noise can make the unknown field distribution-valued, so products appearing in the equation require renormalization. Regularity structures and paracontrolled distributions provide frameworks for constructing such equations under precise assumptions. Their connection with rough paths is methodological: irregular data must be supplemented by carefully organized higher-order objects, and limiting procedures must track correction terms. The resulting solution depends not only on a formal differential expression but also on the specified noise, approximation procedure, and renormalized interpretation.