Itô's lemma is the change-of-variables theorem of stochastic calculus. It describes how a sufficiently smooth function of a random process evolves when that process has continuous fluctuations of the kind generated by Brownian motion. Unlike the ordinary chain rule, it contains a second-order term involving the function's curvature. This additional term is necessary because the accumulated squared increments of Brownian motion do not vanish as the observation intervals become arbitrarily short.
Named after the Japanese mathematician Kiyosi Itô, the formula is a central tool for studying stochastic differential equations, diffusion processes, and continuous-time random models. It connects stochastic integration with differential equations, underlies important methods in mathematical finance, and explains how noise changes under nonlinear transformations. Its extensions cover multidimensional processes, jumps, nonsmooth functions, and processes taking values on geometric spaces.
Historical development and mathematical setting
Classical calculus was developed for functions whose local changes can be approximated by a linear expression, with higher-order errors becoming negligible under integration. Random paths presented a different problem. Brownian motion, introduced into mathematical probability as a model of irregular motion, has continuous sample paths but almost surely no ordinary derivative. Its paths also have infinite total variation on every nontrivial finite interval. Consequently, the usual interpretation of an integral against a smooth or bounded-variation path does not provide a general calculus for Brownian motion.
Norbert Wiener gave a rigorous mathematical construction of Brownian motion in the 1920s. During the 1930s, Andrey Kolmogorov developed analytic equations governing transition probabilities of diffusion processes. These developments established both a pathwise probabilistic object and differential equations describing its distributions. A remaining task was to formulate equations driven directly by individual random trajectories and to manipulate those equations systematically.
Itô's work in the 1940s supplied the decisive construction. His 1944 paper on stochastic integrals introduced an integration theory suitable for random integrands depending on the information already available. The resulting integral was not simply an extension of ordinary integration: its left-endpoint sampling convention encoded the distinction between present information and future random increments. The change-of-variables formula associated with this theory made it possible to calculate with stochastic differential equations without pretending that Brownian paths were differentiable.
The name “lemma” understates the formula's importance. It is a theorem serving much the same organizing role in stochastic analysis that the chain rule serves in ordinary calculus. It identifies the correct differential operator associated with a diffusion, transforms equations into more tractable forms, and converts local descriptions of random motion into identities for functions of that motion.
Later developments broadened the formula beyond its original Brownian setting. Semimartingale theory provided a common framework for continuous noise and jumps, while local-time formulas extended change of variables to functions with corners. Alternative stochastic integrals, especially the Stratonovich integral, clarified the relation between stochastic equations and geometry. These extensions preserve the underlying insight: the correct transformation rule depends not only on first-order increments but also on their accumulated second-order behavior.
Stochastic integration and quadratic variation
A standard setting is a probability space equipped with a filtration , an increasing family of information sets. Brownian motion is adapted to this filtration, and its future increments are independent of the information available at the present time. For , the increment has a normal distribution with mean zero and variance .
For a simple adapted integrand, the Itô integral is constructed from sums of the form
The coefficient uses information available at the left endpoint . Under square-integrability assumptions, completion of these simple integrands gives an integral satisfying the Itô isometry:
This identity controls the stochastic integral through an ordinary time integral. Such integrals are martingales when the relevant square-integrability conditions hold.
The crucial second-order quantity is quadratic variation. For Brownian motion, along deterministic partitions whose mesh tends to zero,
in probability. This is written . For a uniform partition with intervals, the sum has mean and variance , directly exhibiting the convergence. Special partition sequences also give almost sure convergence, but that stronger conclusion should not be confused with the basic convergence-in-probability statement.
A continuous semimartingale can be decomposed as
where has finite variation and is a continuous local martingale. The finite-variation part contributes no quadratic variation, so . In particular, if
then
This separates drift from fluctuation. Drift accumulates through ordinary integration in time; the random martingale part contributes both a first-order stochastic integral and a second-order quadratic-variation term. Itô's lemma is the precise rule combining those contributions when a nonlinear function is applied to .
The one-dimensional formula and its hypotheses
Let be a real-valued Itô process:
This differential notation abbreviates the integral equation
A usual local assumption is that the coefficients are progressively measurable and, for every finite ,
Stronger assumptions may be imposed when expectations or global moment estimates are needed.
Suppose belongs to : it has a continuous first partial derivative in time and continuous first and second partial derivatives in space. Itô's lemma states that
The equality holds as an identity of continuous processes, up to indistinguishability under the standard formulation.
Its compact differential form is
with derivatives evaluated at . The correction
is the feature distinguishing the formula from the ordinary chain rule. A positive second derivative produces a positive local drift contribution from noise, whereas a negative second derivative produces a negative contribution.
For a general continuous semimartingale, the same theorem takes the coordinate-free-in-time form
This version does not require quadratic variation to have a density with respect to time.
The hypotheses concern both smoothness and the region in which the function is defined. For example, can be used only while . A standard method is to stop the process before it exits compact subsets of the function's domain and then remove the stopping boundaries when justified. Such localization also handles unbounded derivatives. The formula itself does not require globally bounded derivatives, nor does it assert existence or uniqueness of the stochastic equation to which it is applied.
Why the second-order term survives
The mechanism can be seen by applying a second-order Taylor expansion along a partition
For Brownian motion and a time-independent smooth function,
where . Summing the left-hand side telescopes to . The first sum on the right converges to an Itô integral. The second converges to
Brownian increments over an interval of length have typical size , not . Their squares therefore have size , and summing approximately such contributions leaves a finite result. By contrast, the squared increments of a smooth path have size , whose sum vanishes. This difference explains why ordinary calculus stops at first order while Brownian calculus retains a second-order term.
The familiar mnemonic
summarizes these limiting rules. It is not an algebra of ordinary infinitesimal numbers. Each expression represents a statement about limits of accumulated products of increments. For example, mixed time–Brownian sums vanish because their factors have incompatible scales, while squared Brownian increments accumulate to elapsed time.
A rigorous proof also controls the Taylor remainders. After stopping Brownian motion in a bounded region, continuity of gives a modulus-of-continuity bound
where as . Continuity of the Brownian path makes the largest increment tend to zero, while the sum of squared increments remains bounded in probability. Their product therefore vanishes in probability.
For a general Itô process, drift increments have finite variation and the martingale part supplies the quadratic variation. Approximation of the stochastic integral and localization then extend the argument to admissible random coefficients. This proof structure explains why the ordinary theorem needs only two continuous spatial derivatives, rather than an indefinitely differentiable function: the higher-order terms are controlled as a remainder, not retained as a full infinite expansion.
Several variables, covariance, and the product rule
Let take values in and be driven by an -dimensional standard Brownian motion:
Write
This matrix is the instantaneous covariance matrix of the increments. Its entries determine the quadratic covariations:
For a scalar function , the multidimensional formula is
Using the gradient and Hessian matrix, the second-order term is
Thus the correction is determined by the interaction between covariance and curvature, not by the separate variances alone.
If two Brownian motions and have constant correlation , then
For , the second-order contribution is
The mixed derivative appears twice in the double sum, and the prefactor leaves coefficient . Treating correlated drivers as independent would omit this contribution and generally produce an incorrect drift.
An important special case follows by taking . For continuous semimartingales and ,
Consequently, stochastic integration by parts contains a covariation term absent from the bounded-variation version. For Brownian motion,
which gives
The formula applies component by component to a vector-valued transformation . Each transformed coordinate has its own Hessian correction. This matters when changing coordinates in systems of stochastic equations: a nonlinear transformation changes not just the diffusion coefficient through its Jacobian, but also the drift through second derivatives. Even a process with zero drift in its original coordinates can acquire nonzero drift after a nonlinear coordinate change.
Explicit transformations and representative examples
Powers of Brownian motion give a direct family of examples. For an integer ,
For , the process is a martingale. For ,
Taking expectations and using yields
The stochastic integral has zero expectation here because the required integrability holds.
The geometric Brownian motion equation
illustrates the logarithmic correction. Applying the formula to gives
Therefore,
The subtraction of distinguishes the drift of the logarithm from the proportional drift of . For constant coefficients, , although the mean logarithmic increment is .
A related construction is the stochastic exponential. If is a continuous local martingale with , then
satisfies
The compensating quadratic-variation term cancels the curvature term introduced by exponentiation. The resulting process is a positive local martingale. It need not be a true martingale without additional conditions; positivity alone does not guarantee preservation of expectation.
A geometric example comes from the distance of a -dimensional Brownian motion from the origin. Set . Away from , the derivatives of the Euclidean norm give
The stochastic integral has quadratic variation , and can be represented locally by a one-dimensional Brownian driver. The resulting radial process is a Bessel process. Its drift is generated entirely by the curvature of the distance function. The singularity at the origin requires separate boundary analysis, particularly in low dimensions; the smooth formula cannot simply be applied there.
Diffusion generators and partial differential equations
For a time-homogeneous diffusion
Itô's lemma identifies the infinitesimal generator
Under suitable regularity, this operator describes the first-order change in conditional expectations:
The connection is exact at the process level: Itô's formula says that
is a local martingale.
After suitable integrability or stopping, taking expectations gives Dynkin's formula,
This converts a differential identity along random paths into an identity for their distributions. If is sufficiently smooth, it satisfies the backward evolution equation
For standard Brownian motion, , so the equation is the heat equation with diffusivity .
The corresponding forward equation acts on a probability density :
This is the Fokker–Planck equation. The backward and forward equations encode the same diffusion from different directions: the former evolves functions of the initial state, while the latter evolves distributions of the state.
The Feynman–Kac formula adds a potential or discount term. Suppose a sufficiently regular function satisfies
Applying Itô's lemma to
cancels its drift. If the resulting local martingale is a true martingale, then
Terminal or boundary conditions, coefficient regularity, and integrability determine the precise scope of this representation. The formula supplies a bridge between partial differential equations and random-path methods, including probabilistic solutions of boundary problems and simulation-based evaluation of expectations.
Financial modeling and numerical computation
In mathematical finance, Itô's lemma converts an assumed asset-price equation into the dynamics of a contingent claim. In the classical constant-coefficient model,
and a smooth claim value satisfies
The first and second spatial derivatives measure different effects: controls exposure to small price changes, while determines the curvature correction.
The Black–Scholes model, developed in the work of Fischer Black, Myron Scholes, and Robert Merton in 1973, uses this identity in a frictionless continuous-trading setting. A self-financing portfolio containing shares and a cash position has dynamics
Matching the stochastic term gives . Matching the drift then yields
The physical drift cancels. This calculation depends on the specified market assumptions and the self-financing condition; it is not a statement that any dynamically adjusted portfolio is riskless or that actual market prices obey the model exactly.
In numerical analysis, the same second-order structure determines approximation schemes for stochastic equations. For
the Euler–Maruyama method uses
where independent increments have distribution . Under standard global Lipschitz, growth, and moment assumptions, this method generally has strong convergence order . Strong error concerns the approximation of paths driven by the same Brownian motion; weak error concerns expectations of suitable test functions.
A higher-order scalar approximation, the Milstein method, retains an iterated-integral contribution:
The term comes from the identity for , not from an ordinary deterministic Taylor expansion. With appropriate regularity, the scalar method has strong order . In multidimensional systems, higher-order schemes may require cross iterated integrals and stochastic areas, making the covariance structure of the driving noise computationally significant.
Stratonovich calculus and coordinate changes
The Stratonovich integral uses a different convention for stochastic integration. For continuous semimartingales,
Under suitable conditions, symmetric or midpoint-type approximations lead to this integral, whereas left-endpoint approximations lead to the Itô integral. The two constructions assign different values to the same formal integrand and integrator.
Stratonovich notation restores the familiar form of the chain rule:
This does not eliminate the quadratic-variation phenomenon. Instead, it incorporates the correction into the definition of the integral. For example,
whereas the Itô value is . Both identities are correct within their respective conventions.
For a scalar equation,
the equivalent Stratonovich equation is
Thus using the same printed drift in both conventions generally defines different processes. For a vector equation with noise columns , the correction is
Each derivative here acts on the spatial coordinates.
This distinction is particularly useful in differential geometry. On a smooth manifold, Stratonovich equations can be written in terms of vector fields and transformed between coordinate charts using the ordinary chain rule. An Itô equation instead acquires a second-order drift contribution under a nonlinear coordinate transformation. That contribution is necessary for different coordinate descriptions to represent the same underlying random motion.
Stratonovich equations also arise in approximation problems. Under standard Wong–Zakai conditions, ordinary differential equations driven by smooth approximations to Brownian motion converge to a Stratonovich equation. The conclusion depends on the approximation scheme and hypotheses, so it is not a universal rule for every model with rapidly varying noise. Itô integration remains especially natural when nonanticipation, martingale properties, and conditional expectations are central. The choice of convention is therefore part of the mathematical specification of a stochastic model, rather than a purely notational preference.
Jumps, nonsmooth functions, and extended formulas
For a semimartingale with jumps, second-order differentiation alone cannot describe the effect of every increment. A jump can be finite, so the full nonlinear change in the function must be retained. If is a real-valued càdlàg semimartingale, meaning that its paths are right-continuous with left limits, and , then
Here , and is the continuous part of quadratic variation. The left-limit integrand is predictable, so its value is determined before the jump occurs.
The jump sum is an exact correction, not a second-order approximation. For a Poisson process with unit jumps,
If the intensity is a constant , the associated generator is
This finite-difference operator contrasts with the second-order differential operator of a diffusion.
A separate extension weakens smoothness of the function. The absolute-value function has a corner, so the ordinary formula does not apply at zero. For a continuous semimartingale, Tanaka's formula gives
where is symmetric local time, with the corresponding symmetric sign convention. Local time records the accumulation of quadratic variation near the level . For Brownian motion, its normalization is characterized by the occupation-density identity
for suitable measurable .
More generally, a convex function has a second derivative in the sense of a measure, even if it has no ordinary second derivative everywhere. The smooth correction can then be replaced by
with the appropriate one-sided derivative used in the stochastic integral. This is a generalized Itô–Tanaka formula and is useful for positive-part functions, barriers, and reflected processes.
Further extensions treat random functions, infinite-dimensional processes, and path-dependent functionals. In the Itô–Wentzell formula, the transformed function is itself a random field, so additional covariation terms arise between its randomness and the process being substituted into it. Infinite-dimensional versions require an appropriate second derivative and trace condition. Each extension must specify which second-order variation exists; continuity or randomness alone is insufficient to justify an Itô-type formula.