Itô's Lemma: The Chain Rule for Random Motion

The precise statement

Let (Ω, ℱ, (ℱₜ), ℙ) be a filtered probability space satisfying the usual conditions, and let W be a standard (ℱₜ)-Brownian motion. Call X an Itô process if

Xₜ = X₀ + ∫₀ᵗ bₛ ds + ∫₀ᵗ σₛ dWₛ,

where b and σ are (ℱₜ)-progressively measurable with ∫₀ᵗ|bₛ|ds < ∞ and ∫₀ᵗσₛ² ds < ∞ almost surely (so the stochastic integral is well defined). Itô's Lemma. If f ∈ C¹,²([0,∞)×ℝ), then for all t ≥ 0, almost surely,

f(t, Xₜ) = f(0, X₀) + ∫₀ᵗ (∂ₜf + b ∂ₓf + ½ σ² ∂ₓₓf)(s, Xₛ) ds + ∫₀ᵗ (σ ∂ₓf)(s, Xₛ) dWₛ.

In differential shorthand: df = ∂ₜf dt + ∂ₓf dX + ½ ∂ₓₓf σ² dt. The last term — the Itô correction — equals ½ ∂ₓₓf d⟨X⟩, where ⟨X⟩ₜ = ∫₀ᵗσₛ² ds is the quadratic variation. This is the only structural difference from the classical chain rule.

The picture: why (dW)² = dt

Brownian motion is continuous but of infinite total variation and nowhere differentiable, so 'dW' is not a small number you can Taylor-expand away. The right scaling is √: over an interval of length Δt, an increment ΔW is order √Δt, hence (ΔW)² is order Δt — the same order as the drift term, not negligible.

Make this rigorous with quadratic variation: for a partition of [0,t] with mesh → 0, ∑(Wₜᵢ₊₁ − Wₜᵢ)² → t in L² (and a.s. along dyadic refinements). By contrast, for a C¹ path the same sum → 0. So Brownian motion accumulates a genuine 'second-order' budget that a smooth curve does not. When you Taylor-expand f(X) to second order, the term ½f″(ΔX)² normally vanishes in the limit; here it survives as ½f″σ²dt. That surviving term is the entire novelty. Everything downstream — option prices, the Fokker–Planck equation, filtering — is bookkeeping of this one nonzero contribution.

Key idea of the proof

Fix a partition 0 = t₀ < ⋯ < tₙ = t and telescope: f(Xₜ) − f(X₀) = ∑ᵢ [f(Xₜᵢ₊₁) − f(Xₜᵢ)]. Apply a second-order Taylor expansion to each increment:

f(Xₜᵢ₊₁) − f(Xₜᵢ) = f′(Xₜᵢ) ΔXᵢ + ½ f″(Xₜᵢ)(ΔXᵢ)² + o((ΔXᵢ)²).

Summing the first-order terms gives, in the mesh → 0 limit, the Itô integral ∫ f′ dX (this is where progressive measurability and the Itô isometry do the work). The crux is the quadratic term: (ΔXᵢ)² = σ²(ΔWᵢ)² + cross terms + drift², and one shows ∑ σ²(ΔWᵢ)² → ∫₀ᵗ σ² ds in probability — precisely because ∑(ΔWᵢ)² → ⟨W⟩ = t. The drift-squared and cross terms are order (Δt)² and Δt·√Δt, so they vanish. Continuity of f″ (the C² hypothesis) lets you replace f″(Xₜᵢ) by its limit on the integrator ⟨X⟩. Collecting: the quadratic term converges to ½∫ f″ σ² ds. That is Itô's Lemma.

Canonical example: geometric Brownian motion

Take f(x) = eˣ and let Xₜ = (μ − ½σ²)t + σWₜ, so dX = (μ − ½σ²)dt + σ dW with constant coefficients. Set Sₜ = f(Xₜ) = exp((μ − ½σ²)t + σWₜ). Apply Itô with ∂ₓf = eˣ = S and ∂ₓₓf = eˣ = S:

dS = S dX + ½ S σ² dt = S[(μ − ½σ²)dt + σ dW] + ½ S σ² dt = μ S dt + σ S dW.

The −½σ² inside the exponent is exactly cancelled by the Itô correction, leaving the clean SDE dS = μS dt + σS dW — geometric Brownian motion, the Black–Scholes stock model. Run it the other way: solving dS = μS dt + σS dW naïvely as if it were an ODE would give S = S₀exp(μt + σW), which is wrong; the true solution carries the −½σ²t Itô drift. Note E[Sₜ] = S₀eᵘᵗ, so the median (√) lags the mean — a direct footprint of the correction term.

Why the hypotheses matter — and what breaks

C² is not decoration. Take f(x) = |x| and Xₜ = Wₜ. Then f is not differentiable at 0, and the naïve formula fails; the correct statement is Tanaka's formula, d|Wₜ| = sgn(Wₜ) dWₜ + dL⁰ₜ, where L⁰ is the local time at 0 — a new nondecreasing process that replaces the missing ½f″ (a Dirac mass). Convex-but-not-C² functions land you in the Itô–Tanaka theory with a measure ½f″(dx) as second derivative. Continuity of the paths matters too: for processes with jumps you need the Itô–Meyer/semimartingale formula, which adds ∑[f(X) − f(X⁻) − f′(X⁻)ΔX] over jumps. And the convention matters: Itô's formula holds because Itô integrals evaluate σ at the left endpoint (making them martingales); the Stratonovich integral, evaluating at the midpoint, restores the ordinary chain rule d∘f = f′ ∘ dX with no correction — at the cost of losing the martingale property.

Why it matters: what Itô's Lemma unlocks

Itô's Lemma is the computational engine of stochastic analysis. Finance: applying it to an option value V(t, Sₜ) and hedging away the dW term yields the Black–Scholes PDE ∂ₜV + rS∂ₛV + ½σ²S²∂ₛₛV − rV = 0 (Black, Scholes, Merton, 1973). PDE ↔ probability: combined with the martingale property it gives the Feynman–Kac formula, representing solutions of parabolic PDEs as expectations E[φ(Xₜ)] over diffusions — the bridge between the heat equation and Brownian motion. SDE theory: it verifies candidate solutions of stochastic differential equations and underlies existence/uniqueness via the generator ℒ = b∂ₓ + ½σ²∂ₓₓ. Statistics & control: it produces the Fokker–Planck/Kolmogorov forward equation for the density, the Kalman–Bucy filter, and the Hamilton–Jacobi–Bellman equation of stochastic control. The multidimensional version df = ∇f·dX + ½∑ᵢⱼ ∂ᵢⱼf d⟨Xⁱ,Xʲ⟩ carries all of this to interacting systems.

Ordinary calculus vs. Itô calculus: where the extra term comes from

Object	Classical (smooth path)	Itô (Brownian path)
Quadratic variation of the driver over [0,t]	0	⟨W⟩ₜ = t (nonzero!)
Chain rule for f(X)	df = f′ dX	df = f′ dX + ½ f″ d⟨X⟩
Product rule d(XY)	X dY + Y dX	X dY + Y dX + d⟨X,Y⟩
Symbolic multiplication table	(dt)²=dt·dW=(dW)²=0	(dW)² = dt, dt·dW = 0, (dt)² = 0
Regularity needed on f	C¹	C² (resp. C¹,² in (t,x))
Substitution / integration by parts	no correction	correction ½∫f″d⟨X⟩ always present