A stochastic integral is an integral in which the integrator, and usually also the integrand, is a stochastic process. Its characteristic form is
where describes a randomly evolving signal and specifies how its increments are weighted. When has highly irregular paths, ordinary integration methods do not generally apply. Stochastic integration supplies alternative constructions that use probabilistic information, pathwise structure, or both.
The most widely used construction is the Itô integral, developed for integration against Brownian motion and subsequently extended to much broader classes of processes. Another major construction, the Stratonovich integral, obeys the familiar change-of-variables rule of ordinary calculus, but differs from the Itô integral by a correction determined by the fluctuations of the processes involved. These are not competing notations for an identical operation: they encode different limiting procedures.
Stochastic integrals are fundamental to stochastic differential equations, martingale theory, diffusion processes, mathematical finance, and models of fluctuating physical systems. Their theory explains why random fluctuations produce additional terms in differentiation formulas, how continuous and discontinuous noise can be treated within one framework, and which approximations lead to which limiting equations.
Origins and the obstruction to ordinary integration
For a sufficiently regular function , an integral of the form can be constructed from weighted increments,
If has finite total variation and is continuous, the resulting Riemann–Stieltjes integral is insensitive to the choice of evaluation points within the subintervals. Brownian motion behaves differently: almost every path is continuous but has infinite total variation on every nontrivial compact interval. Ordinary finite-variation integration therefore cannot supply a general integral against Brownian paths.
The decisive feature is quadratic variation. For Brownian motion , along deterministic partitions of whose mesh tends to zero,
in probability. In contrast, a continuous finite-variation function has zero quadratic variation. Brownian increments over an interval of length are typically of size , so their squares accumulate at order one even as the individual increments vanish.
This behavior makes evaluation conventions consequential. The algebraic identity
shows that left-endpoint sums converge to
By comparison, sums using the arithmetic mean of the two endpoint values satisfy
The discrepancy is a structural consequence of quadratic variation, not numerical error.
The subject emerged from several strands of twentieth-century probability. Norbert Wiener constructed a mathematical measure on continuous Brownian paths in the 1920s and developed integration for deterministic integrands. Kiyosi Itô’s work beginning in the 1940s made it possible to integrate random, nonanticipating integrands and to formulate differential equations driven by Brownian motion. Ruslan Stratonovich introduced a different convention in the 1950s that retained the ordinary chain rule.
Later developments connected stochastic integration to the general theory of martingales and to processes with jumps. This development changed the central question from whether random paths possess ordinary derivatives to which classes of integrators support a stable integration operation. The modern answer depends both on the path properties of the integrator and on the information available when the integrand is chosen.
Construction of the Itô integral
The Brownian Itô integral is built on a filtered probability space
The filtration records the information available through time . Brownian motion is adapted to it, and is independent of for . Thus the information structure is part of the definition: a process that is Brownian relative to one filtration need not remain Brownian after future information is added.
For an elementary predictable integrand,
where is -measurable, define
The coefficient multiplying an increment is chosen before that increment occurs. This nonanticipation condition is essential: it makes the coefficient independent of the forthcoming Brownian increment in the conditional sense required by the construction.
For square-integrable elementary processes, independence and zero conditional means eliminate cross terms when the square of the sum is averaged. The result is the Itô isometry:
This identity links the expected value of a squared random integral to an ordinary integral. It also gives the continuity estimate needed to extend the definition.
Predictable processes satisfying
can be approximated by elementary processes in the norm on the right-hand side. Their integrals then converge in the space of square-integrable random variables, and the limit defines the Itô integral. The Hilbert space structure ensures that the result does not depend on the approximating sequence. Progressively measurable square-integrable integrands are also admitted through this construction, with the appropriate equivalence of integrands under the product measure .
The integral process
has a continuous version and is a square-integrable martingale under the stated integrability condition. In particular, and . Its increments need not be independent or normally distributed: those stronger properties generally fail when is random.
A localization procedure extends the definition to predictable satisfying only
for every finite . One stops the process before its accumulated squared integrand becomes too large, constructs the stopped integrals, and joins them consistently. The resulting integral is a continuous local martingale, but it need not be a true martingale or have finite second moment.
Quadratic variation and Itô calculus
The Itô integral has an associated calculus in which quadratic fluctuations contribute to differentiation. If
then its quadratic variation is
For two such integrals,
their quadratic covariation is
These identities quantify the accumulated size and joint fluctuation of random increments, rather than their ordinary total variation.
Suppose
under the usual local integrability conditions. For a function with one continuous time derivative and two continuous space derivatives, Itô’s formula states
The second-derivative term has no counterpart in ordinary differentiation along a smooth path. A second-order Taylor expansion explains its origin: squared Brownian increments survive after summation, whereas the relevant higher-order remainder terms vanish under suitable localization.
For example, applying the formula to gives
Applying it instead to yields
The correction exactly cancels the additional second-order term. This exponential is an important example of a positive martingale and a building block for changes of probability measure.
The corresponding product rule is
for continuous semimartingales. Thus stochastic integration by parts contains a covariation term. The notation , often used in informal calculations, summarizes this behavior; it is not literal multiplication of infinitesimal real numbers.
In several dimensions, let be a standard -dimensional Brownian motion and let , where is an matrix. The second-order contribution becomes
The matrix is the instantaneous covariance matrix of the noise in . It determines the second-order differential operator associated with the diffusion and creates a direct connection between stochastic integration and partial differential equations.
Semimartingales, localization, and jumps
Brownian motion is not the most general useful integrator. The standard general framework uses a semimartingale, a process that can be decomposed as
where is a local martingale and is an adapted process of locally finite variation. Processes are ordinarily taken to have right-continuous paths with left limits, a regularity condition called càdlàg. For a predictable integrand , integration separates into
The second term is a pathwise Lebesgue–Stieltjes integral; the first requires the martingale construction and localization.
Predictability is especially important when the integrator jumps. If is adapted and càdlàg, its left-limit process is predictable and represents the value immediately before time . At a jump,
For an equation written , the multiplier is therefore the pre-jump state. Replacing it with the post-jump value generally changes the equation and can introduce anticipation.
A basic example is the Poisson process with intensity . It has finitely many jumps on bounded intervals almost surely, so
Its compensated version, , is a martingale. For suitable predictable ,
Compensation subtracts the predictable average jump accumulation while retaining the random fluctuations.
Itô’s formula must also account for jumps that are not infinitesimally small. For a real semimartingale and a twice continuously differentiable ,
Here is the continuous part of quadratic variation. The jump sum retains the exact nonlinear change at each jump rather than approximating it solely through a quadratic term.
The importance of semimartingales is also structural. The Bichteler–Dellacherie characterization identifies them, subject to the standard process assumptions, as the “good integrators” for which integration of bounded elementary predictable processes has the appropriate continuity under uniform smallness of the integrands. Processes outside this class may still admit meaningful integrals, but generally require a different topology, additional path information, or restrictions on the integrands.
The Stratonovich convention and smooth approximations
For continuous semimartingales and , the Stratonovich integral is defined by
The circle distinguishes it from the Itô integral. Under standard conditions, it is also obtained from symmetric endpoint sums. The correction vanishes when has continuous finite-variation paths, but it is usually nonzero when itself responds to the same noise as .
Its defining advantage is the ordinary-looking chain rule. For a continuous semimartingale and sufficiently smooth ,
Consequently,
whereas the corresponding Itô integral is . The second-order contribution has not disappeared; it is incorporated into the definition of the integral instead of appearing explicitly in the change-of-variables formula.
A scalar Stratonovich equation
is equivalent, under suitable smoothness, to the Itô equation
The two equations describe the same process only after this drift adjustment. Writing identical drift and diffusion coefficients with different integral conventions generally specifies different models.
The distinction is particularly visible for multiplicative noise. The equation
has solution . Its Itô form has drift . By contrast, the Itô equation has solution . The latter has constant expectation when is a fixed constant; the former generally does not.
The Wong–Zakai approximation results explain why Stratonovich integration often appears when smooth signals approximate Brownian motion. Under appropriate hypotheses, solutions of ordinary differential equations driven by piecewise smooth Brownian approximations converge to a Stratonovich equation. The conclusion depends on the approximation procedure; specially constructed multidimensional approximations can produce additional correction terms.
Stratonovich notation is also useful on a manifold. Equations expressed through smooth vector fields transform naturally under coordinate changes because the ordinary chain rule remains valid. Itô equations can equally well describe manifold-valued processes, but their coordinate expressions require second-order transformation terms. For jump processes, several distinct extensions of the Stratonovich idea exist; the simple continuous-process conversion formula does not by itself determine which jump convention is intended.
Stochastic differential equations and diffusion models
A stochastic differential equation is interpreted through stochastic integrals, not through an ordinary time derivative of Brownian motion. The expression
means
The drift integral describes accumulated finite-variation motion, while the stochastic integral describes fluctuations whose magnitude and direction depend on the current state.
Under global Lipschitz continuity and linear-growth conditions on and , with an appropriate initial condition, this equation has a unique nonexplosive strong solution. Here “strong” means that the solution is constructed using a specified Brownian motion and initial data; it does not refer to classical differentiability. Local Lipschitz assumptions often yield existence and uniqueness up to a possible explosion time. A weak solution instead allows the probability space and driving Brownian motion to be part of the construction.
The Ornstein–Uhlenbeck equation,
illustrates how ordinary variation-of-constants arguments combine with stochastic integration:
For deterministic , the solution is Gaussian with mean and variance
The limiting variance reflects a balance between deterministic relaxation and continual random forcing.
For a time-homogeneous diffusion, Itô’s formula identifies its infinitesimal generator:
When a sufficiently regular transition density exists, it satisfies the forward equation
This equation describes the evolution of the entire probability distribution rather than an individual sample path.
The Feynman–Kac formula gives a complementary link to differential equations. Under suitable regularity and integrability assumptions,
solves
Applying Itô’s formula to the discounted process makes the connection explicit: the partial differential equation cancels its drift, leaving a local martingale. Such representations support probabilistic methods for heat flow, diffusion with killing, boundary-value problems, and numerical evaluation through random sampling.
Martingale representation, changes of measure, and finance
Stochastic integration can reconstruct martingales from a basic source of noise. In the completed natural filtration of Brownian motion, every square-integrable terminal random variable measurable at time has a representation
for a predictable square-integrable integrand , unique up to equality under . Equivalently,
This is a form of the martingale representation theorem.
The result depends on the filtration, not just on the presence of Brownian motion. If the filtration contains additional independent random sources, Brownian stochastic integrals alone need not represent all its martingales. In jump-driven filtrations, compensated jump integrals play a corresponding role. Representation theorems identify which random fluctuations can be expressed through the available driving processes.
Stochastic exponentials provide a related mechanism for changing probability measures. For a predictable , consider
This process is a positive local martingale. A sufficient condition for it to be a true martingale on is Novikov’s condition,
When defines a probability measure by , Girsanov’s theorem states that
is Brownian motion under . The transformation changes drift, while quadratic variation remains unchanged.
In continuous-time finance, a predictable trading strategy specifies holdings chosen using information available before the next price change. If is a semimartingale price process, its accumulated trading gain is
Predictability rules out choosing a position after observing the price increment to which it is applied. The finite-variation part of contributes ordinary accumulated gains, while its local-martingale part contributes stochastic gains.
In the idealized Black–Scholes model,
an appropriate change of measure replaces the drift with the risk-free rate . Discounted prices then become martingales under the transformed measure. Martingale representation supplies the mathematical mechanism for replicating suitable contingent claims using the traded asset and the bank account.
These conclusions depend on explicit assumptions, including the filtration, allowed trading strategies, integrability, and the absence of specified market frictions. In models with untraded sources of randomness, martingale representation relative to traded assets can fail, and a claim may not have a unique replication price. The stochastic integral formalizes gains; it does not by itself establish market completeness.
Approximation and numerical computation
A stochastic integral is usually simulated through increments rather than by constructing an entire continuous sample path. For Brownian motion on a grid , the increments are independent with
For an adapted step approximation , the sum
is itself an exact Itô integral of the corresponding piecewise constant process. The numerical error comes from approximating the intended integrand or the solution that determines it.
The Euler–Maruyama method for an Itô equation is
Its left-endpoint coefficient evaluation reflects the nonanticipating convention. Under standard global regularity and moment assumptions, its strong convergence order is : a typical root-mean-square error at a fixed terminal time is . With additional smoothness, its weak order is typically , meaning an error in expectations of sufficiently smooth test functions.
Strong and weak accuracy answer different questions. Strong approximation compares a numerical path with the exact solution driven by the same Brownian path. Weak approximation compares distributions through expected test functions. Computing an average terminal payoff may require weak accuracy, while pathwise coupling, filtering algorithms, or approximations of stopping events can require more detailed control.
For a scalar equation with sufficiently smooth coefficients, the Milstein method adds a term that accounts for an iterated stochastic integral:
The correction follows from
Under suitable assumptions, this raises strong order to . In several dimensions, higher-order schemes can require cross-iterated integrals, including Lévy areas, which are not determined by the Brownian endpoint increments alone.
Numerical stability and consistency with the integration convention are separate concerns. Applying an ordinary smooth-path solver to interpolated Brownian data can approximate a Stratonovich equation rather than the intended Itô equation. Conversely, a left-point stochastic scheme can introduce the wrong drift if a model was specified in Stratonovich form without conversion. For some equations with superlinearly growing coefficients, the explicit Euler–Maruyama method can exhibit divergent moments even when the exact solution is well behaved.
The Monte Carlo method adds sampling error to discretization error. With independent simulated outcomes of finite variance, the standard error of an average decreases proportionally to the inverse square root of the sample size. Reducing the time step therefore does not eliminate statistical error, and increasing the number of paths does not remove bias from an inaccurate time discretization.
Beyond classical semimartingale integration
Classical Itô theory is broad, but it does not cover every random signal. A major example is fractional Brownian motion , a centered Gaussian process with covariance
At it is ordinary Brownian motion. For other values it is not a semimartingale in its natural filtration. Its paths are locally Hölder continuous of every order less than , and the parameter also controls the dependence of its increments.
One pathwise alternative is Young integration. If and are Hölder continuous with exponents and , and
then can be defined as a limit of ordinary increment sums without a probabilistic correction. This is especially useful for fractional Brownian motion with and sufficiently regular integrands. It generally does not handle integration of Brownian motion against itself, because Brownian Hölder exponents lie below .
Rough path theory, developed by Terry Lyons in the 1990s, extends pathwise integration by enhancing a signal with iterated integrals. For paths of roughness corresponding to , first-level increments and second-level iterated integrals provide the principal data. The latter record information such as signed areas that cannot be recovered continuously from uniform convergence of the original paths alone. Different enhancements can encode Itô-type or Stratonovich-type behavior. This approach separates deterministic continuity results for differential equations from the probabilistic construction of their driving signals.
Another extension uses Malliavin calculus. Its divergence operator defines the Skorokhod integral for certain integrands that may depend on future noise. When an integrand is adapted and satisfies the relevant square-integrability conditions, this integral agrees with the Itô integral. For anticipative integrands, however, it is a different operation, characterized through a duality relation with differentiation on Gaussian probability space. It should not be interpreted simply as an Itô sum with the nonanticipation restriction omitted.
Infinite-dimensional stochastic integration is needed for random evolution in function spaces. If is a Hilbert-space-valued Wiener process with covariance operator , integrability of an operator-valued process is governed by a condition such as
where is the Hilbert–Schmidt norm. Cylindrical Wiener noise requires related constructions because it need not define a random variable in the underlying Hilbert space.
These integrals enter stochastic partial differential equations through expressions such as the stochastic convolution
where is the evolution semigroup of a deterministic equation. The smoothing supplied by , the spatial covariance of the noise, and the dimension of the domain jointly determine whether the integral exists as a function-valued process or only in a space of generalized functions.