Stochastic differential equations · Part 2 of 10

Brownian motion

· 6 min read

Brownian motion is the central object of stochastic calculus. It is the continuous-time limit of a random walk, the simplest continuous martingale, and the source of randomness in every stochastic differential equation in this collection.

Definition

A standard Brownian motion is a process (Bt)t≥0(B_t)_{t \ge 0} such that

  1. B0=0B_0 = 0;
  2. for any times 0≤t0<t1<⋯<tn0 \le t_0 < t_1 < \cdots < t_n, the increments Bt1−Bt0,…,Btn−Btn−1B_{t_1} - B_{t_0}, \dots, B_{t_n} - B_{t_{n-1}} are independent;
  3. each increment is Gaussian, with Bt−Bs∼N(0,t−s)B_t - B_s \sim N(0, t - s) for s<ts < t;
  4. the path t↦Btt \mapsto B_t is continuous, almost surely.

Equivalently, BB is a continuous Gaussian process with mean zero and covariance E[BsBt]=min⁡(s,t)\mathbb{E}[B_s B_t] = \min(s, t). All of its finite-dimensional distributions follow from the transition density

pt(x,y)=12πt e−(y−x)2/(2t),p_t(x, y) = \frac{1}{\sqrt{2\pi t}}\, e^{-(y - x)^2 / (2t)},

because the joint density of (Bt1,…,Btn)(B_{t_1}, \dots, B_{t_n}) is the product pt1(0,x1) pt2−t1(x1,x2)⋯ptn−tn−1(xn−1,xn)p_{t_1}(0, x_1)\, p_{t_2 - t_1}(x_1, x_2) \cdots p_{t_n - t_{n-1}}(x_{n-1}, x_n). As a function of tt and yy, the transition density solves the heat equation ∂tp=12∂y2p\partial_t p = \tfrac{1}{2} \partial_y^2 p, a first hint of the connection between diffusions and PDEs that a later post develops.

Symmetries

Several transformations turn a Brownian motion into another Brownian motion:

  • Scaling: for any c>0c > 0, the process (c−1/2Bct)t≥0\big(c^{-1/2} B_{ct}\big)_{t \ge 0} is a Brownian motion.
  • Reflection: (−Bt)t≥0(-B_t)_{t \ge 0} is a Brownian motion.
  • Restarting: for a fixed time ss, the process (Bs+t−Bs)t≥0(B_{s+t} - B_s)_{t \ge 0} is a Brownian motion, independent of the path up to time ss.
  • Time inversion: the process (tB1/t)t>0(t B_{1/t})_{t > 0}, set to 00 at t=0t = 0, is a Brownian motion.

Scaling is the most important of these. It says that zooming in on a Brownian path by a factor cc in time and c\sqrt{c} in space gives a statistically identical picture. That already explains why the path cannot be differentiable: over a short time hh, its increments have size about h\sqrt{h}, which is much larger than hh.

Does it exist?

The definition asks for a lot, and it is not obvious that any process satisfies all four conditions. Three constructions are worth knowing.

Kolmogorov’s route. The Kolmogorov extension theorem builds a process with the right finite-dimensional distributions on the space of all functions from [0,∞)[0, \infty) to R\mathbb{R}. That space is too large: continuity depends on uncountably many coordinates and is not even an event there. The Kolmogorov–Chentsov criterion fixes this. Because E∣Bt−Bs∣4=3∣t−s∣2\mathbb{E}|B_t - B_s|^4 = 3|t - s|^2, and similar bounds hold for higher moments, the process has a version whose paths are continuous, and in fact Hölder continuous of every order below 1/21/2. The resulting probability measure on the space of continuous functions is called Wiener measure.

Lévy’s midpoint construction. Build the path on the dyadic times by successive refinement. Start with B0=0B_0 = 0 and B1∼N(0,1)B_1 \sim N(0, 1). Given the values at the two ends of an interval of length hh, set the value at its midpoint to the average of the end values plus an independent N(0,h/4)N(0, h/4) correction, which is exactly the conditional distribution of a Brownian bridge at its midpoint. Interpolating linearly at each stage, the approximations can be written as a sum of tent-shaped functions (Schauder functions) with independent Gaussian coefficients. The largest of 2n2^n independent standard Gaussians grows only like 2nlog⁡2\sqrt{2n\log 2}, so the corrections shrink fast enough for the approximations to converge uniformly, almost surely, to a continuous limit.

Random Fourier series. On the interval [0,π][0, \pi], with independent standard Gaussians ξ0,ξ1,…\xi_0, \xi_1, \dots,

Bt=ξ0 tπ+2π∑n=1∞ξnsin⁡(nt)nB_t = \frac{\xi_0\, t}{\sqrt{\pi}} + \sqrt{\frac{2}{\pi}} \sum_{n=1}^{\infty} \xi_n \frac{\sin(nt)}{n}

defines a Brownian motion. You can check the covariance directly from the Fourier series identity

stπ+2π∑n=1∞sin⁡(ns)sin⁡(nt)n2=min⁡(s,t),s,t∈[0,π].\frac{st}{\pi} + \frac{2}{\pi}\sum_{n=1}^{\infty} \frac{\sin(ns)\sin(nt)}{n^2} = \min(s, t), \qquad s, t \in [0, \pi].

This is essentially Wiener’s original construction, and it pictures Brownian motion as random noise in which each frequency nn is weighted by 1/n1/n.

Infinite length, finite quadratic variation

Take a partition 0=t0<t1<⋯<tn=t0 = t_0 < t_1 < \cdots < t_n = t with mesh Δ=max⁡i(ti+1−ti)\Delta = \max_i (t_{i+1} - t_i), and consider the sum of squared increments

Qn=∑i(Bti+1−Bti)2.Q_n = \sum_{i} \big(B_{t_{i+1}} - B_{t_i}\big)^2.

Each term has mean ti+1−tit_{i+1} - t_i and variance 2(ti+1−ti)22(t_{i+1} - t_i)^2, and the terms are independent, so

E[Qn]=t,Var⁡(Qn)=2∑i(ti+1−ti)2≤2tΔ→0.\mathbb{E}[Q_n] = t, \qquad \operatorname{Var}(Q_n) = 2\sum_i (t_{i+1} - t_i)^2 \le 2t\Delta \to 0.

So the quadratic variation of Brownian motion on [0,t][0, t] is exactly tt: it is not random at all. This has a striking consequence. If a path had finite total variation VV, then Qn≤max⁡i∣Bti+1−Bti∣⋅VQ_n \le \max_i |B_{t_{i+1}} - B_{t_i}| \cdot V, which tends to zero by continuity, contradicting Qn→t>0Q_n \to t > 0. So Brownian paths have infinite length on every interval. In fact, by a theorem of Paley, Wiener and Zygmund, they are nowhere differentiable, almost surely.

Infinite variation is the reason the integral ∫H dB\int H\,dB cannot be defined path by path as a Stieltjes integral, and nonzero quadratic variation is the reason Itô’s formula has an extra term. The next two posts deal with both.

Martingale and Markov properties

Let Ft\mathcal{F}_t be the information generated by the path up to time tt. Then each of the following is a martingale:

Bt,Bt2−t,exp⁡(λBt−λ2t2)(λ∈R).B_t, \qquad B_t^2 - t, \qquad \exp\Big(\lambda B_t - \frac{\lambda^2 t}{2}\Big) \quad (\lambda \in \mathbb{R}).

Each claim follows from the fact that Bt−BsB_t - B_s is independent of Fs\mathcal{F}_s. Brownian motion is also a Markov process: given its present position, its future is independent of its past. The strong Markov property extends this from fixed times to stopping times, and it leads to the reflection principle: for a>0a > 0,

P(max⁡s≤tBs≥a)=2 P(Bt≥a).\mathbb{P}\Big(\max_{s \le t} B_s \ge a\Big) = 2\,\mathbb{P}(B_t \ge a).

The reason is a symmetry argument. Each path that reaches aa and ends below it can be matched with the path that is reflected about aa after its first visit there, which ends above it, and the two are equally likely.

A brief history

Robert Brown watched pollen particles jitter in water in 1827. Louis Bachelier used the same kind of random motion to model stock prices in his 1900 thesis, five years before Albert Einstein explained the physical phenomenon as the result of molecular collisions and used it to estimate the size of molecules. Norbert Wiener gave the first rigorous mathematical construction in 1923, which is why Brownian motion is also called the Wiener process.

← All posts in Stochastic differential equations