Stochastic Calculus: Itô versus Stratonovich

The Itô–Stratonovich distinction is a working concern wherever multiplicative noise appears. Simulating overdamped Langevin dynamics with position-dependent diffusion — colloids near a wall, ions in a channel, molecular motors — demands that the code declare its convention: the Euler–Maruyama scheme is intrinsically Itô, so a physically derived Stratonovich model must have the spurious drift $\tfrac12 bb'$ added by hand, or the simulation silently converges to the wrong Boltzmann distribution and reports incorrect escape rates. Enforcing detailed balance and the fluctuation–dissipation theorem for such systems is, at bottom, a convention-bookkeeping exercise. The same machinery governs quantitative finance, where the non-anticipating Itô convention is mandatory because trading strategies cannot use future prices, and the martingale property underlies risk-neutral pricing; the celebrated $-\tfrac12\sigma^2$ term in the Black–Scholes log-price is nothing but the Itô correction of Example 1. Noise-induced phenomena — stochastic resonance, Brownian ratchets that rectify fluctuations into directed motion, noise-induced transitions that create or destroy peaks in a stationary distribution — all hinge on the spurious drift and therefore on getting the convention right before any analysis begins.

Imagine measuring the area under a wildly fluctuating curve with a sum of rectangles. For a smooth curve it does not matter whether you take each rectangle's height at the left edge, the middle, or the right edge of its strip — refine the mesh and all three answers converge to the same integral. For white noise the curve is so rough that the strip width $\mathrm dt$ and the height fluctuation $\sqrt{\mathrm dt}$ conspire: the sampling choices differ by a finite amount that survives every refinement. That irreducible discrepancy is the entire content of the Itô–Stratonovich question. It exists only because $(\mathrm dW)^2=\mathrm dt$ rather than the ordinary-calculus $(\mathrm dW)^2\to 0$. Track a particle taking a step $b(X)\,\mathrm dW$. In the Itô prescription you compute $b$ before the step, at the left endpoint; in Stratonovich you compute $b$ at the half-step, the midpoint. Because the step and the change in $b$ are both driven by the same increment $\mathrm dW$, they are correlated, and the midpoint feels an extra push $\tfrac12 b\,b'\,(\mathrm dW)^2 = \tfrac12 b b'\,\mathrm dt$. This is the noise-induced or spurious drift — a systematic, directed motion produced by state-dependent noise even when the deterministic force averages to zero. It is not an artifact of sloppy bookkeeping: it is a real physical effect whenever the noise strength varies in space, and getting its sign and magnitude right is the difference between a simulation that respects detailed balance and one that quietly drifts to the wrong steady state.

The two figures carry the thesis. The first shows why the dilemma exists: as the mesh refines, $\sum(\Delta W_i)^2$ converges to the elapsed time $t$ while $\sum(\Delta t_i)^2=t\,\Delta t$ vanishes — so $(\mathrm dW)^2$ is the one second-order term you cannot discard, exactly the term ordinary calculus throws away. The second shows what it costs: for the multiplicative SDE $\mathrm dX=\sigma X\,\mathrm dW$, the same physical noise gives $\langle X\rangle=X_0$ in the Itô reading (a martingale, flat) but $\langle X\rangle=X_0 e^{\sigma^2 t/2}$ in the Stratonovich reading (growing). The gap is precisely the spurious drift $\tfrac12 bb'=\tfrac12\sigma^2 X$ — not a rounding error, but the signature of multiplicative noise. A physical noise always has a small but nonzero correlation time $\tau$; over $\tau$ the path is smooth and ordinary calculus applies exactly. Sending $\tau\to0$ therefore keeps the smooth-calculus chain rule and lands on Stratonovich — this is the Wong–Zakai theorem, and it is why physicists derive their models in the Stratonovich convention. Itô, by contrast, is natural when the future is genuinely unknowable at each instant, as for a trader who can react only to past prices; its left-point, non-anticipating structure is what makes it a martingale. For additive noise the whole issue disappears. If the noise amplitude does not depend on the state, $b'=0$, so the spurious drift $\tfrac12 bb'$ is zero: both prescriptions give identical trajectories and identical Fokker–Planck dynamics. All the ambiguity is a multiplicative-noise phenomenon, which is why the Ornstein–Uhlenbeck process and the ordinary damped Langevin equation never force a choice.

The starting point is the multiplication table for stochastic differentials, dictated by the quadratic variation of the Wiener process. It is the one rule from which Itô's lemma and the drift conversion both follow.

where: $\mathrm dW$ — the Wiener increment over $\mathrm dt$, Gaussian with variance $\mathrm dt$; $\mathrm dt$ — the time step. The first identity says the squared increment is deterministic at leading order; the others say cross- and higher-order terms are negligible. A stochastic differential equation with drift $a$ and noise amplitude $b$ is written as follows; it is incomplete until the interpretation of the $b\,\mathrm dW$ term is fixed.

where: $X$ — the stochastic state; $a$ — the drift (deterministic force); $b$ — the noise amplitude, multiplicative when it depends on $X$. When $b'=\partial_X b\neq0$ the convention matters. Itô's lemma is the stochastic chain rule: because $(\mathrm dX)^2=b^2\,\mathrm dt$ is first order, the Taylor expansion of $f(X)$ keeps the second-derivative term.

where: $f$ — a smooth function of the state; $\tfrac12 b^2 f''$ — the Itô correction, absent in ordinary calculus. Setting $f=X^2$ with $a=0,\ b=$ const recovers $\mathrm d\langle X^2\rangle=b^2\,\mathrm dt$, i.e. ordinary diffusion — a check we use below. The Stratonovich integral samples the integrand at the midpoint of each subinterval, which is what preserves the ordinary chain rule.

where: $(S)$ — the Stratonovich symbol; the integrand is evaluated at the midpoint $\tfrac12(X_i+X_{i+1})$, correlating it with the increment it multiplies. The two conventions describe the same process provided the drift is shifted by the spurious drift $\tfrac12 bb'$. This single conversion rule is the bridge between the calculi.

where: $a_{\text{Itô}},a_{\text{Strat}}$ — the drifts in the two representations of the same physical process; $\tfrac12 bb'$ — the noise-induced drift; $b'=\mathrm db/\mathrm dx$. Equivalently $a_{\text{Strat}}=a_{\text{Itô}}-\tfrac12 bb'$. Each convention has its own Fokker–Planck equation for the density $P(x,t)$. In the Itô form the diffusion coefficient $b^2$ sits inside both spatial derivatives.

where: $P(x,t)$ — probability density; $a_{\text{Itô}}$ — Itô drift; $b^2$ — diffusion coefficient inside two derivatives. The Stratonovich Fokker–Planck equation instead reads $\partial_t P=-\partial_x(a_{\text{Strat}}P)+\tfrac12\,\partial_x\!\big(b\,\partial_x(bP)\big)$, with one factor of $b$ pulled outside; the two are equivalent under $(5)$, as shown below.

How the two interpretations relate across the important regimes:

Example 1 — Itô's lemma recovers the Wiener variance. Show that the Itô correction is exactly what makes $\langle W^2(t)\rangle=t$, and contrast with the wrong answer the naive chain rule gives.

Step 1. Take $f=W^2$ with $\mathrm dW$ the standard increment ($a=0,\ b=1$). Itô's lemma gives $\mathrm d(W^2)=2W\,\mathrm dW+\tfrac12\cdot 2\cdot 1\cdot\mathrm dt=2W\,\mathrm dW+\mathrm dt$.

Step 2. Average over the noise. The martingale term $2W\,\mathrm dW$ has zero mean (the left-point $W$ is independent of the future increment), leaving $\mathrm d\langle W^2\rangle=\mathrm dt$, so $\langle W^2(t)\rangle=t$.

Step 3. The naive chain rule would give $\mathrm d(W^2)=2W\,\mathrm dW$ with no $\mathrm dt$, hence $\langle W^2\rangle=0$ for all $t$ — a flat, obviously wrong prediction. Numerically integrating $\mathrm d(W^2)$ over $N$ Wiener steps confirms the accumulated $\int\mathrm dt=t$ term.

Result: Itô's lemma gives $\langle W^2(t)\rangle=\mathbf{t}$, matching the Wiener variance; dropping the $\tfrac12 b^2 f''$ term gives the wrong answer $\langle W^2\rangle=0$. The $+\mathrm dt$ is the whole content of stochastic calculus.

Example 2 — Converting a physical Stratonovich model for an Euler–Maruyama simulator. A diffusing particle with position-dependent noise obeys the Stratonovich SDE $\mathrm dX=-\gamma X\,\mathrm dt+\sqrt{2D(X)}\circ\mathrm dW$. Find the Itô drift a left-point (Euler–Maruyama) integrator must use.

Step 1. Identify $b(X)=\sqrt{2D(X)}$, so $b'=\dfrac{D'(X)}{\sqrt{2D(X)}}$.

Step 2. Compute the spurious drift: $\tfrac12 b\,b'=\tfrac12\sqrt{2D}\cdot\dfrac{D'}{\sqrt{2D}}=\tfrac12 D'(X)$.

Step 3. The Itô drift is $a_{\text{Itô}}=a_{\text{Strat}}+\tfrac12 bb'=-\gamma X+\tfrac12 D'(X)$. The simulator integrates $\mathrm dX=(-\gamma X+\tfrac12 D')\,\mathrm dt+\sqrt{2D}\,\Delta W$ with $\Delta W=\sqrt{\Delta t}\,\mathcal N(0,1)$.

Result: the Euler–Maruyama scheme must use the Itô drift $\mathbf{a_{\text{Itô}}=-\gamma X+\tfrac12 D'(X)}$; omitting the $\tfrac12 D'$ term would drive the particle to the wrong steady state and break detailed balance.

References. Gardiner, Handbook of Stochastic Methods, Ch. 4 · Risken, The Fokker–Planck Equation, 2nd ed., Ch. 3 · van Kampen, Stochastic Processes in Physics and Chemistry, Ch. IX · Øksendal, Stochastic Differential Equations · Kardar, Statistical Physics of Particles, Ch. 9 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 10.

Provenance. The multiplication table, Itô's lemma, and the drift conversion are standard and cross-checked against Gardiner Ch. 4, Risken Ch. 3, van Kampen Ch. IX, and Øksendal. A verifier recomputes: (i) Itô's lemma from the second-order Taylor expansion with $(\mathrm dW)^2=\mathrm dt$; (ii) the geometric-Brownian-motion solution $X=X_0\exp[(\mu-\tfrac12\sigma^2)t+\sigma W]$ with $\langle X\rangle=X_0 e^{\mu t}$; (iii) the drift conversion $a_{\text{Itô}}=a_{\text{Strat}}+\tfrac12 bb'$ by expanding the midpoint rule (checked for $b=\sigma x$ and $b=\sqrt{2D(x)}$); (iv) the equivalence of the two Fokker–Planck forms via the product rule; (v) $\mathrm d(W^2)=2W\mathrm dW+\mathrm dt\Rightarrow\langle W^2\rangle=t$. The interpolating stationary distribution with convention parameter $\alpha$ (0 = Itô, ½ = Stratonovich, 1 = Hänggi–Klimontovich) is given in Risken §3.3.

Builds on: Wiener Process and White Noise · The Langevin Equation · Brownian Motion · The Fokker–Planck Equation. Leads to: The Fokker–Planck Equation · Kramers' Escape Rate Theory · The Kramers–Moyal Expansion · The Ornstein–Uhlenbeck Process.