Stochastic differential equations · Part 4 of 10

Itô's formula

· 4 min read

Itô’s formula is the chain rule of stochastic calculus, and it is the most used tool in the subject: almost every calculation in the rest of this collection is an application of it.

The idea

For a smooth function ff, Taylor’s theorem gives

f(Bt+h)−f(Bt)=f′(Bt) ΔB+12f′′(Bt) (ΔB)2+⋯ ,ΔB=Bt+h−Bt.f(B_{t + h}) - f(B_t) = f'(B_t)\,\Delta B + \tfrac{1}{2} f''(B_t)\,(\Delta B)^2 + \cdots, \qquad \Delta B = B_{t+h} - B_t.

In ordinary calculus the second-order term would vanish in the limit. Here (ΔB)2(\Delta B)^2 has expectation hh, the same order as the time step, and it does not average away. Summed over a partition, the second-order terms add up to 12∫f′′(Bs) ds\tfrac{1}{2}\int f''(B_s)\,ds, because quadratic variation turns (dB)2(dB)^2 into dtdt. The third- and higher-order terms do vanish, because ∣ΔB∣3|\Delta B|^3 is of order h3/2h^{3/2}.

The basic formula

For a twice continuously differentiable ff,

f(Bt)=f(B0)+∫0tf′(Bs) dBs+12∫0tf′′(Bs) ds,f(B_t) = f(B_0) + \int_0^t f'(B_s)\,dB_s + \frac{1}{2}\int_0^t f''(B_s)\,ds,

or, in differential form, df(Bt)=f′(Bt) dBt+12f′′(Bt) dtdf(B_t) = f'(B_t)\,dB_t + \tfrac{1}{2}f''(B_t)\,dt. With f(x)=x2f(x) = x^2, this recovers Bt2=2∫0tBs dBs+tB_t^2 = 2\int_0^t B_s\,dB_s + t from the previous post. When ff also depends on time,

df(t,Bt)=(∂tf+12∂xxf)(t,Bt) dt+∂xf(t,Bt) dBt.df(t, B_t) = \Big(\partial_t f + \frac{1}{2}\partial_{xx} f\Big)(t, B_t)\,dt + \partial_x f(t, B_t)\,dB_t.

General Itô processes

For an Itô process dXt=bt dt+σt dBtdX_t = b_t\,dt + \sigma_t\,dB_t, the same expansion, now with d⟨X⟩t=σt2 dtd\langle X \rangle_t = \sigma_t^2\,dt, gives

df(t,Xt)=(∂tf+bt ∂xf+12σt2 ∂xxf)(t,Xt) dt+σt ∂xf(t,Xt) dBt.df(t, X_t) = \Big(\partial_t f + b_t\,\partial_x f + \frac{1}{2}\sigma_t^2\,\partial_{xx} f\Big)(t, X_t)\,dt + \sigma_t\,\partial_x f(t, X_t)\,dB_t.

The dtdt term is the drift of f(t,Xt)f(t, X_t), and the dBtdB_t term is its noise. A trick that recurs throughout this collection is to choose ff so that the drift vanishes, which makes f(t,Xt)f(t, X_t) a (local) martingale.

Several dimensions

For a vector of Itô processes X=(X1,…,Xd)X = (X^1, \dots, X^d) and a twice continuously differentiable ff on Rd\mathbb{R}^d,

df(Xt)=∑i∂if(Xt) dXti+12∑i,j∂ijf(Xt) d⟨Xi,Xj⟩t.df(X_t) = \sum_{i} \partial_i f(X_t)\,dX_t^i + \frac{1}{2}\sum_{i,j} \partial_{ij} f(X_t)\,d\langle X^i, X^j \rangle_t.

For a dd-dimensional Brownian motion, d⟨Bi,Bj⟩t=δij dtd\langle B^i, B^j \rangle_t = \delta_{ij}\,dt, so the second-order term becomes 12Δf(Bt) dt\tfrac{1}{2}\Delta f(B_t)\,dt. Two special cases are worth memorizing:

  • The product rule: d(XtYt)=Xt dYt+Yt dXt+d⟨X,Y⟩td(X_t Y_t) = X_t\,dY_t + Y_t\,dX_t + d\langle X, Y \rangle_t.
  • Harmonic functions: if Δf=0\Delta f = 0, then f(Bt)f(B_t) is a local martingale. This is where the connection between Brownian motion and potential theory begins.

Example: geometric Brownian motion

The Black–Scholes model of a stock price is

dSt=μSt dt+σSt dBt.dS_t = \mu S_t\,dt + \sigma S_t\,dB_t.

Apply Itô’s formula to log⁡St\log S_t, using f(x)=log⁡xf(x) = \log x, f′(x)=1/xf'(x) = 1/x and f′′(x)=−1/x2f''(x) = -1/x^2:

dlog⁡St=dStSt−12d⟨S⟩tSt2=(μ−σ22)dt+σ dBt.d\log S_t = \frac{dS_t}{S_t} - \frac{1}{2}\frac{d\langle S \rangle_t}{S_t^2} = \Big(\mu - \frac{\sigma^2}{2}\Big)dt + \sigma\,dB_t.

Integrating both sides,

St=S0exp⁡((μ−σ22)t+σBt).S_t = S_0 \exp\Big(\Big(\mu - \frac{\sigma^2}{2}\Big)t + \sigma B_t\Big).

The mean grows like eμte^{\mu t}, but a typical path grows at the slower rate μ−σ2/2\mu - \sigma^2/2. If σ2>2μ\sigma^2 > 2\mu, almost every path tends to zero even though the mean grows without bound: the mean is carried by rare, enormous paths.

Example: exponential martingales

Taking μ=0\mu = 0 and σ=θ\sigma = \theta in the computation above shows that

Zt=exp⁡(θBt−θ22t)satisfiesdZt=θZt dBt,Z_t = \exp\Big(\theta B_t - \frac{\theta^2}{2}t\Big) \quad \text{satisfies} \quad dZ_t = \theta Z_t\,dB_t,

so ZZ is a martingale with mean one. More generally, for an adapted process θt\theta_t, the stochastic exponential exp⁡(∫0tθs dBs−12∫0tθs2 ds)\exp\big(\int_0^t \theta_s\,dB_s - \frac{1}{2}\int_0^t \theta_s^2\,ds\big) solves dZ=θZ dBdZ = \theta Z\,dB. These are the densities that change the drift of a Brownian motion in Girsanov’s theorem, later in the collection.

Example: computing moments

Itô’s formula turns moment calculations into ordinary integrals. For f(x)=x4f(x) = x^4,

d(Bt4)=4Bt3 dBt+6Bt2 dt.d(B_t^4) = 4B_t^3\,dB_t + 6B_t^2\,dt.

The stochastic integral has mean zero, so E[Bt4]=6∫0tE[Bs2] ds=6∫0ts ds=3t2\mathbb{E}[B_t^4] = 6\int_0^t \mathbb{E}[B_s^2]\,ds = 6\int_0^t s\,ds = 3t^2, the familiar fourth moment of a Gaussian. The same idea gives a recursion for every moment.

Example: the radial part of Brownian motion

Let BB be a dd-dimensional Brownian motion with d≥2d \ge 2, started away from the origin, and let Rt=∣Bt∣R_t = |B_t|. Away from the origin, f(x)=∣x∣f(x) = |x| has gradient x/∣x∣x/|x| and Laplacian (d−1)/∣x∣(d - 1)/|x|, so Itô’s formula gives

dRt=d−12Rt dt+dWt,Wt=∫0tBs∣Bs∣⋅dBs.dR_t = \frac{d - 1}{2R_t}\,dt + dW_t, \qquad W_t = \int_0^t \frac{B_s}{|B_s|} \cdot dB_s.

The process WW is a one-dimensional Brownian motion, because it is a continuous local martingale with quadratic variation tt (this uses Lévy’s characterization, from the post on martingales). So RR is the Bessel process of dimension dd. Its drift pushes the radius outward, more strongly in higher dimensions, which is the analytic reason Brownian motion in three or more dimensions eventually wanders off to infinity.

← All posts in Stochastic differential equations