Itô's formula is the fundamental change-of-variables rule of stochastic calculus. It describes the evolution of a smooth function evaluated along a random path, particularly one driven by a Wiener process, or Brownian motion. Unlike the ordinary chain rule, it contains an additional second-order term. This correction reflects the accumulation of squared random increments, a phenomenon called quadratic variation.
In its simplest form, if is standard Brownian motion and has two continuous derivatives, then
The stochastic integral is interpreted in the Itô sense. Although Brownian paths are continuous, they are almost surely nowhere differentiable, so this identity cannot be obtained by treating as an ordinary differentiable function of time.
Developed by Kiyosi Itô in the 1940s and early 1950s, the formula provides a working calculus for stochastic differential equations. It connects random motion with differential equations, martingale theory, geometry, numerical simulation, and financial mathematics. Its extensions apply to processes with jumps, functions with limited smoothness, and random evolution in infinite-dimensional spaces.
Historical origins and mathematical setting
The classical calculus of Newton and Leibniz handles change along smooth paths and, more generally, along many paths of finite variation. Brownian motion introduced a fundamentally different kind of trajectory. Its increments over an interval of length have variance , so their typical size is proportional to , rather than . Consequently, squared increments can accumulate to a nonzero quantity even when each individual increment tends to zero.
The construction of Wiener measure by Norbert Wiener in the 1920s gave Brownian motion a rigorous mathematical foundation. During the 1930s, Andrey Kolmogorov developed differential equations governing transition probabilities of Markov processes. These developments established two complementary descriptions of diffusion: random trajectories on one side and deterministic equations for probability distributions on the other. A calculus capable of operating directly on the trajectories was still needed.
Kiyosi Itô's 1944 paper Stochastic Integral introduced the essential integration method underlying his approach. His subsequent work on stochastic differential equations, including On Stochastic Differential Equations in 1951, developed the associated transformation calculus. A decisive feature was the use of integrands evaluated before the corresponding random increment. This preserves the nonanticipating character of the integral: the integrand cannot exploit information from the future increment against which it is integrated.
The resulting theory made it possible to construct diffusions as solutions of equations driven by Brownian motion and to transform those solutions systematically. Instead of differentiating an irregular random trajectory, one works with integral identities whose existence and convergence are established probabilistically. Itô's formula then identifies how an ordinary smooth function interacts with those integrals.
Later developments broadened the framework. Semimartingale theory supplied a class of integrators containing Brownian diffusions and many jump processes. The Stratonovich integral provided an alternative convention with an ordinary-looking chain rule, while local-time formulas extended change of variables to nonsmooth functions. These are not replacements for Itô's formula but related formulations that make different aspects of stochastic change explicit.
The scalar formula and its hypotheses
Let be a filtered probability space, where the filtration represents information available through time . Suppose is Brownian motion relative to this filtration. A scalar Itô process has the form
Here is the drift and is the diffusion coefficient. Standard sufficient assumptions are that the coefficients are progressively measurable and, for every finite ,
The second integral is an Itô integral, defined locally when only the displayed almost-sure integrability is available.
If belongs to , meaning that it has one continuous time derivative and two continuous spatial derivatives, then
The subscripts denote partial derivatives. The identity holds almost surely, simultaneously for all times on which the process and function are defined.
Its differential notation is
with derivatives evaluated at . This is shorthand for the integral equation, not a statement that Brownian motion possesses an ordinary derivative.
The coefficients may themselves be random. In particular, they need not be functions only of the current state. When they have the form and , the process solves a state-dependent stochastic differential equation. Conditions ensuring existence and uniqueness of that equation are separate from the change-of-variables theorem: Itô's formula applies once an appropriate process has been given.
Global boundedness of or its derivatives is not required. A stopping time can restrict the process to bounded regions, where estimates become manageable, and these restrictions can then be removed by localization. If is defined only on an open spatial domain, the formula applies up to the time when exits that domain. This qualification matters for transformations such as the logarithm, which is smooth only on the positive half-line.
Quadratic variation and the origin of the correction
The second-order term arises from the small-scale structure of Brownian motion. For a deterministic partition
whose mesh tends to zero,
in probability. This limit is written . In contrast, a continuous path of finite variation has zero quadratic variation.
The Brownian result can be seen directly from independent increments. Each increment is centered normal with variance , so the expectation of the sum of squared increments is . Its variance is
which tends to zero as the mesh decreases. Thus convergence also holds in mean square for such partitions. The limiting total is not caused by any single large increment; it is the combined effect of increasingly numerous small fluctuations.
For a time-independent smooth function, a Taylor expansion gives the heuristic
The first-order terms converge to drift and stochastic integrals. The second-order terms do not disappear: for an Itô process,
Terms involving only the finite-variation drift, including its squared increments and its products with Brownian increments, vanish in the relevant limits.
This explains the familiar mnemonic
These are rules for retaining terms according to their accumulated variation, not literal equalities between ordinary infinitesimal numbers. Their legitimate use is backed by the corresponding integral convergence theorems.
A rigorous proof first handles bounded coefficients and sufficiently controlled derivatives. The first-order stochastic sums converge by the construction and continuity properties of the Itô integral. Weighted quadratic-increment sums converge to integrals against quadratic variation. Uniform continuity of the second derivative on compact sets controls the Taylor remainder. Localization then removes the boundedness restrictions.
The same reasoning explains why ordinary calculus survives for finite-variation paths: their squared increments accumulate to zero, eliminating the correction. Brownian motion lies at the threshold where second-order increments remain visible. Higher-order Taylor terms disappear under the usual Brownian scaling, leaving a calculus that requires two spatial derivatives but only one time derivative.
Multidimensional form and stochastic product rules
Let be an -dimensional Brownian motion with independent standard components, and let take values in :
Writing
the entries describe instantaneous covariance between the coordinates of . For ,
The second-order term pairs the covariance matrix with the Hessian matrix of . In matrix notation it is
where denotes the matrix trace. Off-diagonal entries matter: correlated coordinate fluctuations interact with mixed second derivatives. Even when the driving Brownian components are independent, coordinates of can be correlated because they share those drivers.
The relevant cross-variation is called quadratic covariation:
For continuous semimartingales it may also be recovered by polarization,
It measures accumulated simultaneous fluctuation, rather than the ordinary covariance of two random variables at a fixed time.
Taking gives stochastic integration by parts:
For example, if and , then the product has the additional drift . Omitting this term produces incorrect formulas for products, ratios, and exponential transformations.
A useful geometric example is the squared distance of -dimensional Brownian motion from the origin. With ,
Each independent coordinate contributes one unit of drift. Away from the origin, applying the formula to gives
This calculation leads to Bessel processes. The restriction away from zero is essential because the norm is not smooth there; any treatment at the origin requires additional boundary analysis or a generalized formula.
Transformations, exponential processes, and martingales
The elementary choice already distinguishes stochastic from ordinary calculus:
Hence
not . The process is a martingale: its future conditional expectation equals its current value. The deterministic subtraction precisely compensates for the average growth of the squared Brownian displacement.
For a general continuous local martingale , the corresponding identity is
Thus is a local martingale, apart from the unchanged initial value. A local martingale becomes a true martingale only under suitable integrability conditions. This distinction is important whenever expectations are taken: the stochastic-integral term cannot automatically be discarded merely because it has no explicit drift.
Exponentiation gives another central example. For a constant ,
satisfies
The negative term in the exponent cancels the second-order Itô correction. The resulting process is a positive martingale with expectation one.
More generally, if is a continuous local martingale, its stochastic exponential is
and Itô's formula gives
It is always a positive local martingale, but need not be a true martingale. Conditions such as Novikov's criterion ensure the latter property. These exponentials are central to Girsanov's theorem, which describes changes of probability measure and the associated changes in drift.
A particularly common state transformation concerns geometric Brownian motion:
For constant coefficients, applying Itô's formula to the natural logarithm yields
Integration gives
The difference between and separates the drift of the level from the drift of its logarithm. It is a direct consequence of curvature, not an additional modeling assumption.
Diffusion generators and partial differential equations
For a time-homogeneous diffusion
define the differential operator
This is the diffusion's infinitesimal generator. Itô's formula implies that
is a local martingale. The operator therefore records the infinitesimal average effect of the random evolution on smooth test functions.
If the diffusion starts at , write its expectation as . Under sufficient integrability, taking expectations gives Dynkin's formula:
Suitable stopped versions apply before exit from a domain. They connect expected exit times and exit locations to boundary-value problems. For example, if a sufficiently regular function satisfies
inside a domain and vanishes on its boundary, then, subject to the required stopping and integrability conditions, represents the expected exit time.
For a terminal payoff , the function
satisfies the backward equation
when the Markov and regularity assumptions justify a classical solution. Applying Itô's formula to explains the equation: its drift must vanish. For standard Brownian motion, , so the corresponding equation is a backward heat equation.
The complementary forward equation concerns the probability density . When such a density and sufficient regularity exist,
This is the Fokker–Planck equation, also called the Kolmogorov forward equation. Its operator is the formal adjoint of . The backward equation evolves functions of the initial state; the forward equation evolves distributions of the state.
Adding a multiplication term leads to the Feynman–Kac formula. For example, solutions of
can be represented, under appropriate hypotheses, by
The proof applies Itô's formula to the product of and the exponential factor. This turns a partial differential equation into an expectation over random paths.
Stratonovich calculus and coordinate changes
The Itô integral uses left-endpoint information in its defining approximations. The Stratonovich integral uses a different convention, related for continuous semimartingales by
The circle distinguishes Stratonovich integration from Itô integration. The covariation term means that the two integrals generally differ even when they use the same integrand and integrator.
For a continuous semimartingale , the Stratonovich change-of-variables rule takes the classical form
The second-order effect has not disappeared. It has been incorporated into the definition of the integral. Converting the right-hand side back to Itô form recovers
Consider the scalar Stratonovich equation
with sufficiently smooth . Its equivalent Itô equation is
Thus specifying the integration convention is part of specifying the model. Writing the same coefficient functions under both conventions generally produces different processes.
For several noise directions, write the diffusion columns as vector fields . The conversion adds
to the Stratonovich drift. This correction expresses how a noise direction changes as the process moves through state space. It is especially important for multiplicative noise, where the diffusion coefficient depends on the current position.
Stratonovich notation is useful on a manifold because its chain rule naturally respects smooth coordinate transformations. An equation expressed using vector fields can be transported between coordinate charts in the same formal way as an ordinary differential equation. An Itô equation instead acquires explicit second-derivative corrections under coordinate changes. The two descriptions are equivalent when these corrections are included.
Smooth approximations to Brownian driving signals often converge to Stratonovich equations, a phenomenon described by Wong–Zakai approximation results under appropriate assumptions. Itô integration, by contrast, is closely aligned with nonanticipating integration and martingale methods. Neither convention is universally preferable: they emphasize different structures, and translating correctly between them preserves the underlying process.
Semimartingales, jumps, and nonsmooth functions
A semimartingale is a process that can be decomposed into a local martingale and a finite-variation process. This class includes continuous Itô diffusions, Poisson processes, and many models combining continuous noise with jumps. It is the standard setting in which stochastic integration and change-of-variables formulas coexist.
For a scalar càdlàg semimartingale —one whose paths are right-continuous with left limits—and , the formula becomes
Here , and is the continuous part of quadratic variation. The left-limit integrand ensures the predictable, nonanticipating interpretation of the integral.
The jump sum is an exact correction, not a quadratic approximation. A jump need not be small, so replacing its effect by a second-order Taylor term would generally be incorrect. For a Poisson process with unit jumps, the continuous quadratic-variation term vanishes, and the formula reduces to
Each event changes the function by its full finite difference.
A different extension relaxes smoothness. The absolute-value function is not differentiable at zero, so ordinary Itô calculus does not directly apply there. For a continuous semimartingale, the Tanaka formula states
using the symmetric local-time convention and .
The local time measures concentration of quadratic variation near level . Its normalization is characterized by the occupation-density identity
for suitable functions . Under appropriate limiting conditions, it can be understood through
For convex functions, generalized Itô formulas integrate local time against the distributional second derivative. This replaces the classical second-derivative term by an object capable of detecting corners and other nonsmooth curvature.
Financial modeling and numerical approximation
A major application is the transformation of asset-price models and derivative values. Suppose an asset follows geometric Brownian motion,
and a derivative has a sufficiently smooth value . Itô's formula gives
The same Brownian increment drives both the asset and the derivative.
In the idealized Black–Scholes model, a self-financing portfolio holding units of the asset and placing the remaining value in a risk-free account eliminates that Brownian exposure. Comparing the resulting drift with the risk-free rate yields the Black–Scholes equation,
The asset's physical drift disappears from the equation. This derivation depends on model assumptions, including frictionless trading and continuous rebalancing; Itô's formula supplies the change-of-variables step, not those economic assumptions. For nonsmooth terminal payoffs, the classical calculation is usually carried out before maturity, where suitable regularity is available, with the payoff imposed as terminal data.
The formula also determines how numerical schemes must account for random fluctuations. The Euler–Maruyama method for
uses
where independent increments satisfy . A higher-order Itô–Taylor expansion gives the scalar Milstein method:
The final term comes from an iterated stochastic integral, evaluated using the identity for the square of a Brownian increment.
Under suitable smoothness, growth, and stability assumptions, Euler–Maruyama has strong convergence order , whereas the scalar Milstein method has strong order . Strong error measures how closely the approximation tracks a solution driven by the same Brownian path. Weak error instead concerns quantities such as differences in expected values of test functions, and can have a different convergence order.
In multidimensional systems, higher-order schemes may require cross-iterated stochastic integrals, including Lévy areas. These encode interactions between distinct noise directions that cannot generally be reconstructed from the endpoint Brownian increments alone. Their appearance is the numerical counterpart of the mixed-derivative and covariation terms in multidimensional Itô calculus.