The Infinitesimal Generator of a Diffusion

Itô's lemma tells you how a function of a diffusion changes over an instant — but written out in full it's a formula, not an object you can manipulate on its own. Distilling exactly the deterministic part of that formula into a single operator turns out to be one of the most useful moves in stochastic calculus: it's what converts "solve this stochastic process" into "solve this ordinary partial differential equation," which is how Black-Scholes actually gets derived.

The analogy: the instantaneous rate-of-change machine

Think of a machine that takes any function $g$ you feed it — "the option payoff," "the bond price," "the squared distance from some level" — and hands back a new function describing exactly how fast $E[g(X_t)]$ is changing, right now, given the diffusion $X_t$ is currently at some point $x$. That machine is the infinitesimal generator. It doesn't care what $g$ is; feed it any reasonable function and it returns the instantaneous expected rate of change, built entirely from the diffusion's drift and volatility.

Writing it down

For a diffusion $dX_t = \mu(X_t)\,dt + \sigma(X_t)\,dW_t$, the infinitesimal generator $\mathcal{L}$ acting on a twice-differentiable function $g$ is

$$\mathcal{L}g(x) = \mu(x) g'(x) + \tfrac{1}{2}\sigma(x)^2 g''(x).$$

In words: take the drift $\mu(x)$ times the function's slope $g'(x)$ (how much of the change comes from the average direction of motion), add half the variance rate $\sigma(x)^2$ times the function's curvature $g''(x)$ (how much extra drift the randomness itself contributes to the expected value, thanks to Jensen's-inequality-style convexity effects). This is exactly the deterministic part of Itô's lemma, stripped of the martingale (noise) term — formally,

$$\mathcal{L}g(x) = \lim_{h \to 0} \frac{E[g(X_{t+h}) \mid X_t = x] - g(x)}{h}.$$

In words: $\mathcal{L}g(x)$ is the instantaneous rate at which the expected value of $g$ is changing, given you're currently sitting at $x$ — a derivative with respect to time, taken through the lens of the diffusion's own dynamics rather than a fixed path.

The Ornstein-Uhlenbeck paths above have generator $\mathcal{L}g(x) = \theta(\mu - x)g'(x) + \tfrac12\sigma^2 g''(x)$. Watch how a path far from the mean-reversion level snaps back faster — that snap-back speed is precisely $\mu(x) = \theta(\mu-x)$, the drift term feeding directly into $\mathcal{L}$.

Worked example 1: the generator of Brownian motion, checked against a known answer

For plain Brownian motion, $\mu(x)=0$ and $\sigma(x)=1$, so $\mathcal{L}g(x) = \tfrac12 g''(x)$. Apply it to $g(x)=x^2$: $g'(x)=2x$, $g''(x)=2$, so $\mathcal{L}g(x) = \tfrac12 \times 2 = 1$. This says $E[W_t^2]$ grows at rate exactly 1 per unit time from any starting point — and indeed $E[W_t^2] = t$ for standard Brownian motion starting at 0, whose derivative with respect to $t$ is exactly 1. The generator's abstract formula reproduces a fact you can check directly.

Worked example 2: the generator as a PDE solver, for a bond price

A short rate follows $dr_t = a(b - r_t)\,dt + \sigma\,dW_t$ (a Vasicek-style mean-reverting process). The price of a bond, $P(r,t)$, satisfies the PDE $\frac{\partial P}{\partial t} + \mathcal{L}P - rP = 0$ — this is the Feynman-Kac connection (see The Feynman-Kac Theorem): the generator turns the expected-discounted-payoff pricing problem into a solvable PDE. Concretely, with $a=0.5$, $b=0.03$, $\sigma=0.01$, at $r=0.04$, $\mathcal{L}P = 0.5(0.03-0.04)P_r + \tfrac12(0.01)^2 P_{rr} = -0.005\,P_r + 0.00005\,P_{rr}$. Plugging in derivatives of a candidate solution and requiring the PDE to balance is literally how the closed-form Vasicek bond price formula is derived — the generator is the bridge between the SDE you wrote down and the PDE you can actually solve.

What this means in practice

• Every closed-form diffusion-based pricing formula (Black-Scholes, Vasicek, CIR) is derived by writing down the generator, forming the associated PDE via Feynman-Kac, and solving it — the generator is the mechanical link between "here is the SDE" and "here is the pricing equation."
• Stationary distributions of a diffusion satisfy $\mathcal{L}^* p = 0$ where $\mathcal{L}^*$ is the generator's adjoint (see The Fokker-Planck and Kolmogorov Equations) — the generator and its adjoint together describe both "how a function's expectation evolves" and "how the probability density itself evolves."
• Numerical schemes for SDEs (finite-difference PDE solvers for options) are, under the hood, discretizing the generator operator directly.