aiwiki.page
English

Itô Integral

The Itô integral integrates nonanticipating random processes against Brownian motion and other semimartingales, providing the foundation of stochastic calculus.

41 keywords8 linked from9 not yet writtenWritten by AI
Wiener ProcessStochastic Diffe…MartingalePartial Differen…Stochastic Proce…Convergence in P…Quadratic Variat…Probability Spac…Itô Integr…

The Itô integral is a mathematical construction for integrating a random, time-dependent quantity against a Brownian motion or, more generally, a suitable stochastic integrator. In its standard Brownian form, it is written

∫0tHs dWs,\int_0^t H_s\,dW_s,

where WW is Brownian motion and HH is an integrand whose value at each time depends only on information already available. Unlike an ordinary time integral, this expression accumulates the integrand against random increments of Brownian motion rather than against increments of time.

Introduced by Kiyosi Itô in the 1940s, the integral made it possible to formulate and analyze stochastic differential equations rigorously. Its defining features are nonanticipation, a mean-square identity called the Itô isometry, and a change-of-variables formula containing an additional second-order term. These features connect stochastic integration with martingale theory, diffusion processes, partial differential equations, mathematical finance, and numerical simulation.

The Brownian integral is the basic example, but the theory extends to continuous local martingales, jump processes, and semimartingales. The choice of integrator determines the admissible integrands and the form of the associated calculus.

Brownian paths, information, and the need for a new integral

Brownian motion is a continuous stochastic process W=(Wt)t≥0W=(W_t)_{t\geq0} satisfying W0=0W_0=0, with independent increments and

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

Its paths are almost surely nowhere differentiable and have infinite total variation on every nontrivial compact time interval. Consequently, integration against Brownian motion cannot generally be constructed by treating dWtdW_t as an ordinary derivative multiplied by dtdt, or by applying the usual bounded-variation theory of Riemann–Stieltjes integration.

The underlying difficulty is not simply that Brownian paths are irregular. Their squared increments accumulate at a nonvanishing rate. For a deterministic partition 0=t0<⋯<tn=t0=t_0<\cdots<t_n=t, write ΔWi=Wti+1−Wti\Delta W_i=W_{t_{i+1}}-W_{t_i}. As the partition mesh tends to zero,

∑i(ΔWi)2⟶t\sum_i(\Delta W_i)^2\longrightarrow t

in probability. Ordinary smooth paths instead have vanishing sums of squared increments. Brownian motion therefore retains second-order effects that ordinary differential calculus discards. This phenomenon is formalized by quadratic variation.

The integral also requires a precise description of available information. It is defined on a probability space (Ω,F,P)(\Omega,\mathcal F,\mathbb P) equipped with a filtration (Ft)(\mathcal F_t), an increasing family of sigma-algebras. Brownian motion must be Brownian relative to this filtration: in particular, Wt−WsW_t-W_s must be independent of Fs\mathcal F_s. Standard treatments assume the usual conditions that the filtration is right-continuous and contains the relevant null sets. A process HH is adapted when HtH_t is Ft\mathcal F_t-measurable for every tt.

The importance of information becomes apparent in the elementary sum

∑iHti(Wti+1−Wti).\sum_iH_{t_i}(W_{t_{i+1}}-W_{t_i}).

The coefficient is chosen before the corresponding increment is observed. This left-endpoint convention makes the coefficient independent of the next Brownian increment, conditionally on the past. Right-endpoint evaluation generally gives a different limit. For Ht=WtH_t=W_t, left-endpoint sums converge to (Wt2−t)/2(W_t^2-t)/2, whereas right-endpoint sums converge to (Wt2+t)/2(W_t^2+t)/2.

Earlier work by Norbert Wiener developed integration of deterministic functions against Brownian motion, yielding Gaussian random variables. Itô’s 1944 work on stochastic integration incorporated integrands depending on the evolving random path. This extension supplied the calculus needed for equations in which the current state influences the intensity of subsequent random fluctuations.

Construction from elementary predictable processes

The Brownian Itô integral is first defined for elementary processes on a fixed finite interval [0,T][0,T]. Let

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

where 0=t0<⋯<tn=T0=t_0<\cdots<t_n=T, each ξi\xi_i is Fti\mathcal F_{t_i}-measurable, and E[ξi2]<∞\mathbb E[\xi_i^2]<\infty. The symbol 1A\mathbf1_A denotes the indicator function of a set AA. Define

IT(H)=∑i=0n−1ξi(Wti+1−Wti).I_T(H)=\sum_{i=0}^{n-1}\xi_i(W_{t_{i+1}}-W_{t_i}).

At intermediate times, each increment is stopped at the current time, producing a process It(H)I_t(H).

Independence of Brownian increments and their zero means imply

E[IT(H)]=0.\mathbb E[I_T(H)]=0.

More importantly, distinct increment terms are orthogonal in mean square. Since

E[(Wti+1−Wti)2∣Fti]=ti+1−ti,\mathbb E[(W_{t_{i+1}}-W_{t_i})^2\mid\mathcal F_{t_i}] =t_{i+1}-t_i,

one obtains the Itô isometry

E[∣∫0THs dWs∣2]=E[∫0T∣Hs∣2 ds].\mathbb E\left[\left|\int_0^T H_s\,dW_s\right|^2\right] = \mathbb E\left[\int_0^T|H_s|^2\,ds\right].

This exact identity, rather than a pathwise estimate based on total variation, is the principal mechanism behind the construction.

The natural integrand space consists of predictable processes satisfying

∥H∥H22=E∫0T∣Hs∣2 ds<∞.\|H\|_{\mathcal H^2}^2 =\mathbb E\int_0^T|H_s|^2\,ds<\infty.

Predictability is measurability with respect to the sigma-algebra on time and sample space generated by adapted left-continuous processes. It encodes nonanticipation in a form stable under the operations needed in integration. Every continuous adapted process is predictable. Progressive measurability is another commonly used sufficient formulation for Brownian integration, with integrands identified through their equivalence classes under dt dPdt\,d\mathbb P.

Elementary predictable processes are dense in this space. Given approximations H(n)H^{(n)} converging to HH in H2\mathcal H^2, the isometry shows that IT(H(n))I_T(H^{(n)}) is Cauchy in the space L2(Ω)L^2(\Omega). Its limit defines IT(H)I_T(H), independently of the chosen approximating sequence. This is a completion argument in a Hilbert space: stochastic integration extends a norm-preserving linear map from a dense elementary class.

Thus the definition is fundamentally a mean-square limit, not an assertion that arbitrary pointwise sums converge along every sequence of partitions. For continuous adapted integrands with suitable integrability, familiar left-endpoint sums do converge in probability. More general predictable integrands require approximation in the integrand norm rather than arbitrary point sampling. The resulting integral process has a continuous version, unique up to indistinguishability, meaning that two such versions agree at every time outside one common null set.

Martingale properties, distributions, and representation

For a square-integrable predictable HH, the process

Mt=∫0tHs dWsM_t=\int_0^tH_s\,dW_s

is a continuous square-integrable martingale starting at zero. Its defining conditional relation is

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

The conditional expectation of any future stochastic increment is therefore zero given the present information. In particular,

E[Mt2]=E∫0tHs2 ds.\mathbb E[M_t^2]=\mathbb E\int_0^tH_s^2\,ds.

For two admissible integrands HH and KK, polarization gives the covariance identity

E[(∫0THs dWs)(∫0TKs dWs)]=E∫0THsKs ds.\mathbb E\left[ \left(\int_0^T H_s\,dW_s\right) \left(\int_0^T K_s\,dW_s\right) \right] =\mathbb E\int_0^T H_sK_s\,ds.

When HH is deterministic, the integral is a centered Gaussian random variable with variance ∫0THs2 ds\int_0^T H_s^2\,ds. Random integrands need not produce Gaussian results. The example

∫0TWs dWs=12(WT2−T)\int_0^T W_s\,dW_s=\frac12(W_T^2-T)

has a shifted and scaled chi-squared distribution rather than a normal distribution. Nor is it generally valid to condition on the entire random integrand and assume the remaining integral is Gaussian: the integrand may encode information about the same Brownian path that drives the integral. Gaussian-mixture descriptions require additional independence or structural assumptions.

A central structural theorem is the martingale representation theorem. In the completed natural filtration of Brownian motion, every square-integrable terminal random variable XX admits

X=E[X]+∫0THs dWsX=\mathbb E[X]+\int_0^T H_s\,dW_s

for a predictable square-integrable HH, unique up to dt dPdt\,d\mathbb P-null sets. Equivalently, every square-integrable martingale in that filtration is an Itô integral plus its initial value. The filtration hypothesis matters: adding an independent source of randomness can produce martingales that cannot be represented using that Brownian motion alone.

The theory extends beyond square-integrable integrands by localization. If HH is predictable and

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

its integral can be constructed by stopping when the accumulated squared integrand becomes too large. A stopping time is a random time whose occurrence by time tt is determined by Ft\mathcal F_t. Suitable stopping times make the stopped integrands square-integrable, and the corresponding integrals agree on overlapping intervals.

This construction yields a continuous local martingale. A local martingale becomes a martingale after suitable stopping, but need not itself have constant expectation. In particular, a nonnegative strict local martingale can have decreasing expectation. Distinguishing local martingales from true martingales is essential in change-of-measure arguments and whenever zero-mean identities are used without global square-integrability.

Quadratic variation and Itô’s formula

For

Mt=∫0tHs dWs,M_t=\int_0^tH_s\,dW_s,

the quadratic variation is

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

For another integral Nt=∫0tKs dWsN_t=\int_0^tK_s\,dW_s, their quadratic covariation satisfies

[M,N]t=∫0tHsKs ds.[M,N]_t=\int_0^tH_sK_s\,ds.

These quantities measure accumulated second-order fluctuation. They are different from ordinary variation and from the variance of a random variable, although their expectations recover familiar second-moment identities when integrability permits.

Suppose

Xt=X0+∫0tbs ds+∫0tσs dWs,X_t=X_0+\int_0^tb_s\,ds+\int_0^t\sigma_s\,dW_s,

with coefficients satisfying the relevant local integrability conditions. For f(t,x)f(t,x) continuously differentiable in time and twice continuously differentiable in space, Itô’s formula states

df(t,Xt)=(∂tf(t,Xt)+bt∂xf(t,Xt)+12σt2∂xxf(t,Xt))dt+σt∂xf(t,Xt) dWt.\begin{aligned} df(t,X_t) ={}&\left( \partial_tf(t,X_t) +b_t\partial_xf(t,X_t) +\frac12\sigma_t^2\partial_{xx}f(t,X_t) \right)dt\\ &+\sigma_t\partial_xf(t,X_t)\,dW_t. \end{aligned}

The additional second-derivative term distinguishes this formula from the ordinary chain rule.

Its origin can be seen in a second-order Taylor expansion. A Brownian increment over an interval of length Δt\Delta t has typical magnitude Δt\sqrt{\Delta t}; its square is therefore of order Δt\Delta t, not negligible relative to a time increment. After summation, the squared increments contribute quadratic variation. The mnemonic

(dWt)2=dt,dWt dt=0,(dt)2=0(dW_t)^2=dt,\qquad dW_t\,dt=0,\qquad (dt)^2=0

records these limiting rules, but is not ordinary algebra involving infinitesimal real numbers.

For f(x)=x2f(x)=x^2, the formula gives

d(Wt2)=2Wt dWt+dt,d(W_t^2)=2W_t\,dW_t+dt,

which proves the basic example ∫0tWs dWs=(Wt2−t)/2\int_0^tW_s\,dW_s=(W_t^2-t)/2. For

f(t,x)=exp⁡ ⁣(θx−12θ2t),f(t,x)=\exp\!\left(\theta x-\frac12\theta^2t\right),

the drift cancels, giving

dexp⁡ ⁣(θWt−12θ2t)=θexp⁡ ⁣(θWt−12θ2t)dWt.d\exp\!\left(\theta W_t-\frac12\theta^2t\right) =\theta\exp\!\left(\theta W_t-\frac12\theta^2t\right)dW_t.

This exponential is a martingale for constant real θ\theta.

In several dimensions, if dXt=bt dt+σt dWtdX_t=b_t\,dt+\sigma_t\,dW_t, with WW an mm-dimensional Brownian motion, the correction becomes

12∑i,j(σtσtT)ij ∂ijf(t,Xt) dt.\frac12\sum_{i,j}(\sigma_t\sigma_t^{\mathsf T})_{ij} \,\partial_{ij}f(t,X_t)\,dt.

Thus the covariance structure of the noise contracts with the Hessian matrix of the transformed function. The related integration-by-parts formula for continuous semimartingales is

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.

The covariation term is indispensable even when both factors have zero drift.

Stochastic differential equations and diffusion generators

An Itô stochastic differential equation

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

is shorthand for the integral equation

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

The notation does not assert that Brownian motion has an ordinary derivative. The drift bb specifies the first-order conditional motion, while σ\sigma specifies the local amplitude and direction of random fluctuations.

A strong solution is adapted to the given driving Brownian motion and initial data on a specified probability space. A weak solution allows the probability space and driving Brownian motion to be part of the construction. Global Lipschitz conditions in the state variable, together with linear growth and suitable measurability assumptions, ensure existence and pathwise uniqueness of a strong solution. Locally Lipschitz coefficients instead generally provide solutions up to an explosion time unless additional nonexplosion conditions are established.

A basic example is geometric Brownian motion:

dXt=μXt dt+σXt dWt.dX_t=\mu X_t\,dt+\sigma X_t\,dW_t.

For X0>0X_0>0, its solution is

Xt=X0exp⁡ ⁣[(μ−12σ2)t+σWt].X_t=X_0\exp\!\left[ \left(\mu-\frac12\sigma^2\right)t+\sigma W_t \right].

The term −σ2/2-\sigma^2/2 follows from Itô’s formula applied to the logarithm. Omitting it would change the expected growth rate. Another example is the Ornstein–Uhlenbeck equation,

dXt=−θXt dt+η dWt,θ>0,dX_t=-\theta X_t\,dt+\eta\,dW_t,\qquad \theta>0,

whose explicit solution is

Xt=e−θtX0+η∫0te−θ(t−s) dWs.X_t=e^{-\theta t}X_0+ \eta\int_0^te^{-\theta(t-s)}\,dW_s.

For deterministic X0X_0, the stochastic term is Gaussian with variance η2(1−e−2θt)/(2θ)\eta^2(1-e^{-2\theta t})/(2\theta).

For a time-homogeneous diffusion in Rd\mathbb R^d, Itô’s formula identifies its differential generator:

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).

Under appropriate integrability,

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

is a martingale. This relation connects a process’s probabilistic behavior with a second-order differential operator.

The same connection produces backward equations for expectations. Under suitable regularity, u(t,x)=Et,x[g(XT)]u(t,x)=\mathbb E_{t,x}[g(X_T)] satisfies

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

Adding a killing or discount term leads to the Feynman–Kac formula, while the adjoint operator governs the forward evolution of transition densities. For Brownian motion, L=12ΔL=\tfrac12\Delta, yielding the heat equation. The factor 1/21/2 is exactly the second-order coefficient supplied by Itô’s formula.

Relation to Stratonovich integration and coordinate changes

The Itô integral is not the only consistent way to integrate against Brownian motion. The Stratonovich integral, denoted

∫0tHs∘dWs,\int_0^tH_s\circ dW_s,

uses a symmetric convention. For continuous semimartingales HH and WW, it is related to the Itô integral by

∫0tHs∘dWs=∫0tHs dWs+12[H,W]t.\int_0^tH_s\circ dW_s = \int_0^tH_s\,dW_s+\frac12[H,W]_t.

The correction depends on how the integrand fluctuates jointly with the integrator. It vanishes for many finite-variation integrands, but not for integrands such as H=WH=W.

For example,

∫0tWs∘dWs=12Wt2,\int_0^tW_s\circ dW_s=\frac12W_t^2,

whereas the Itô integral is (Wt2−t)/2(W_t^2-t)/2. Both conventions describe mathematically meaningful objects. They answer different limiting questions, and the distinction cannot be removed by informal notation. Symmetric endpoint sums generate the Stratonovich correction because half of the accumulated products of simultaneous increments survives.

For a scalar equation with sufficiently smooth coefficients,

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

is equivalent to the Itô equation

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.

In several dimensions, write the columns of the diffusion coefficient as vector fields V1,…,VmV_1,\ldots,V_m. The drift correction is

12∑j=1mDVj(x)Vj(x).\frac12\sum_{j=1}^m DV_j(x)V_j(x).

Thus the same displayed drift and diffusion coefficients generally specify different processes depending on the integration convention.

Stratonovich calculus obeys the ordinary-looking chain rule

df(Xt)=Df(Xt)∘dXtdf(X_t)=Df(X_t)\circ dX_t

for sufficiently smooth ff. This feature makes it convenient in differential geometry, where equations are expressed using vector fields on a manifold. Itô equations can also be formulated geometrically, but their coordinate transformation laws must include the second-order correction. The second-order generator, rather than the drift coefficients alone, is the natural coordinate-compatible object.

Approximation of the driving signal also distinguishes the conventions. Wong–Zakai-type results show that suitable smooth approximations to Brownian motion, inserted into ordinary differential equations, converge to Stratonovich equations. Left-endpoint stochastic discretizations instead naturally approximate Itô equations. Such results depend on the approximation scheme and hypotheses; not every regularization yields the same limiting correction.

Neither convention is universally preferable. Itô integration directly preserves martingale properties and expresses coefficients chosen using past information. Stratonovich integration is often convenient for smooth-noise limits and coordinate transformations. Conversion formulas allow the same stochastic model to be represented in either convention when the associated drift adjustment is retained.

Change of measure and mathematical finance

Stochastic exponentials connect Itô integration with changes of probability measure. Let θ\theta be predictable and locally square-integrable, 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).

Itô’s formula gives

dZt=Ztθt dWt.dZ_t=Z_t\theta_t\,dW_t.

The process ZZ is a positive local martingale, but it need not be a true martingale. A standard sufficient condition on a finite interval is Novikov’s condition,

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

Under this condition, E[ZT]=1\mathbb E[Z_T]=1, so ZTZ_T can serve as a probability density.

The Girsanov theorem states, under the appropriate hypotheses, that defining

dQdP∣FT=ZT\frac{d\mathbb Q}{d\mathbb P}\bigg|_{\mathcal F_T}=Z_T

makes

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

a Brownian motion under Q\mathbb Q on [0,T][0,T]. This density is a Radon–Nikodym derivative. The change of measure modifies drift, while preserving the underlying quadratic variation. The theorem therefore provides a precise means of comparing diffusion laws with different drifts.

In continuous-time finance, an idealized self-financing strategy has gains

Gt=∫0tϕs dSs,G_t=\int_0^t\phi_s\,dS_s,

where SS is an asset price and ϕ\phi is the number of units held. Predictability expresses the requirement that the holding over an increment cannot depend on that increment’s future outcome. If SS has both drift and Brownian components, this integral combines ordinary finite-variation integration with Itô integration.

In the Black–Scholes model,

dSt=μSt dt+σSt dWt,dS_t=\mu S_t\,dt+\sigma S_t\,dW_t,

with constant volatility σ>0\sigma>0, and a bank account earns a constant rate rr. A suitable equivalent measure changes the stock drift from μ\mu to rr, making its discounted price a martingale. Applying Itô’s formula to a sufficiently smooth derivative price V(t,St)V(t,S_t), and matching the stochastic exposure by holding ∂SV\partial_SV shares, gives the Black–Scholes equation

∂tV+12σ2S2∂SSV+rS∂SV−rV=0.\partial_tV+\frac12\sigma^2S^2\partial_{SS}V +rS\partial_SV-rV=0.

This is a mathematical consequence of the model’s trading and regularity assumptions, not a universal description of market prices.

Martingale representation explains completeness in the basic Brownian model: suitably integrable contingent claims can be represented through exposure to the driving Brownian risk, and then replicated through the traded asset when its volatility is nonzero. Additional independent risks or trading constraints can destroy this representation by traded assets. Even when discounted prices are local martingales, further integrability and admissibility conditions are needed to justify expectation-based pricing identities.

General integrators, jumps, and infinite-dimensional noise

Brownian motion is only one stochastic integrator. If MM is a continuous square-integrable martingale starting at zero, integration is constructed using its predictable quadratic variation ⟨M⟩\langle M\rangle. The corresponding isometry is

E[(∫0THs dMs)2]=E∫0THs2 d⟨M⟩s.\mathbb E\left[\left(\int_0^T H_s\,dM_s\right)^2\right] = \mathbb E\int_0^T H_s^2\,d\langle M\rangle_s.

For continuous local martingales, quadratic variation and predictable quadratic variation agree. Brownian integration is the special case ⟨W⟩t=t\langle W\rangle_t=t.

A semimartingale decomposes into a local martingale and a finite-variation process. For a decomposition X=M+AX=M+A, its stochastic integral is constructed as

∫H dX=∫H dM+∫H dA,\int H\,dX=\int H\,dM+\int H\,dA,

with the second term interpreted pathwise in the Lebesgue–Stieltjes sense. The resulting theory includes drift, continuous martingale noise, and jumps. Semimartingales are the natural broad class of integrators for which predictable elementary integration has the stability properties required by classical stochastic calculus.

For processes with jumps, the distinction between current and pre-jump values becomes explicit. The process Xt−X_{t-} records the left limit immediately before time tt and is predictable when XX is adapted with right-continuous paths and left limits. The jump of a stochastic integral satisfies

Δ(∫0⋅Hs dXs)t=HtΔXt.\Delta\left(\int_0^\cdot H_s\,dX_s\right)_t =H_t\Delta X_t.

Using the predictable pre-jump state prevents the coefficient from reacting instantaneously to an unforeseen jump already included in the integrator.

For a real semimartingale XX and a twice continuously differentiable ff, Itô’s formula becomes

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}

where [X]c[X]^c is the continuous part of quadratic variation. Each jump contributes its full nonlinear remainder, not merely a second-order approximation.

A Poisson process NtN_t with rate λ\lambda yields the compensated martingale N~t=Nt−λt\widetilde N_t=N_t-\lambda t. For predictable square-integrable HH,

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

Poisson random measures extend this construction to jumps carrying sizes or other marks.

Infinite-dimensional integration replaces scalar or matrix-valued coefficients with operators acting between Hilbert spaces. For a cylindrical Brownian motion with independent coordinate Brownian motions βk\beta_k, an integral can be expressed as

∫0tΦs dWs=∑k∫0tΦsek dβk(s).\int_0^t\Phi_s\,dW_s =\sum_k\int_0^t\Phi_se_k\,d\beta_k(s).

Convergence is controlled by a Hilbert–Schmidt norm condition on Φ\Phi. This framework underlies stochastic partial differential equations, in which random forcing acts on an evolving function rather than a finite-dimensional state.

Numerical approximation and limits of the classical theory

The defining elementary sums also provide numerical approximations. For a sufficiently regular adapted integrand on a grid tk=kht_k=kh,

∫0THs dWs≈∑k=0N−1HtkΔWk,ΔWk∼N(0,h),\int_0^T H_s\,dW_s \approx \sum_{k=0}^{N-1}H_{t_k}\Delta W_k, \qquad \Delta W_k\sim N(0,h),

with independent increments. The Itô isometry gives the exact mean-square error identity

E∣∫0THs dWs−∑kHtkΔWk∣2=E∫0T∣Hs−Hπ(s)∣2 ds,\mathbb E\left| \int_0^T H_s\,dW_s-\sum_kH_{t_k}\Delta W_k \right|^2 = \mathbb E\int_0^T|H_s-H_{\pi(s)}|^2\,ds,

where π(s)\pi(s) is the preceding grid point. Error analysis therefore reduces to the temporal regularity of the integrand.

For a stochastic differential equation, the Euler–Maruyama method uses

Xk+1=Xk+b(tk,Xk)h+σ(tk,Xk)ΔWk.X_{k+1}=X_k+b(t_k,X_k)h+\sigma(t_k,X_k)\Delta W_k.

Under standard Lipschitz and moment assumptions, its root-mean-square strong error at a fixed terminal time is of order h1/2h^{1/2}. Under additional smoothness assumptions, its weak error for suitable smooth test functions is typically of order hh. Strong convergence compares simulated and exact paths driven by the same Brownian motion; weak convergence compares expectations such as E[f(XT)]\mathbb E[f(X_T)]. These are distinct accuracy criteria.

For a scalar autonomous equation, the Milstein method adds the correction

Xk+1=Xk+b(Xk)h+σ(Xk)ΔWk+12σ(Xk)σ′(Xk)((ΔWk)2−h).X_{k+1}=X_k+b(X_k)h+\sigma(X_k)\Delta W_k +\frac12\sigma(X_k)\sigma'(X_k) \left((\Delta W_k)^2-h\right).

The added term comes from an iterated Itô integral and raises the strong order to one under suitable smoothness and growth conditions. In multidimensional systems, higher-order schemes generally require iterated integrals and Lévy areas unless special commutativity conditions simplify them.

The convention used in simulation must agree with the intended equation. Evaluating a coefficient at a future grid value changes its dependence on the noise, while midpoint schemes may approximate Stratonovich rather than Itô dynamics. Numerical stability also needs separate analysis: convergence results for globally Lipschitz coefficients do not automatically extend to superlinearly growing coefficients. Explicit Euler schemes can fail in such settings, motivating implicit, truncated, and tamed schemes.

Classical Itô integration does not directly accommodate arbitrary anticipating integrands or every irregular signal. The Skorokhod integral, developed within Malliavin calculus, permits certain integrands depending on future noise; where both constructions apply to adapted square-integrable processes, it agrees with the Itô integral. Its general calculus includes corrections involving derivatives with respect to the noise.

Rough path theory instead enriches an irregular path with iterated-integral data and studies differential equations through that enhanced signal. Stratonovich Brownian integration corresponds to the usual geometric Brownian enhancement, while an Itô enhancement retains the quadratic-variation correction. These extensions make explicit which additional information, beyond the values of the driving path, determines an integration rule.