Stochastic differential equations · Part 1 of 10

Probability, the measure-theoretic way

· 5 min read

Stochastic calculus is about random functions of continuous time, and that is where elementary probability runs out of road. “The probability that a Brownian path stays below 1 for all t≤1t \le 1” involves uncountably many random variables at once, and making sense of it requires the measure-theoretic framework. This post collects the parts of that framework the rest of the collection relies on.

Probability spaces

A probability space is a triple (Ω,F,P)(\Omega, \mathcal{F}, \mathbb{P}). The sample space Ω\Omega lists all possible outcomes, F\mathcal{F} is a σ-algebra of subsets of Ω\Omega called events, and P\mathbb{P} assigns each event a probability, with P(Ω)=1\mathbb{P}(\Omega) = 1 and countable additivity: for disjoint events A1,A2,…A_1, A_2, \dots,

P(⋃nAn)=∑nP(An).\mathbb{P}\Big(\bigcup_{n} A_n\Big) = \sum_n \mathbb{P}(A_n).

Why not assign a probability to every subset? On Ω=[0,1]\Omega = [0, 1] with the uniform distribution, there are sets to which no consistent length can be assigned, so probabilities are defined only on a σ-algebra: a collection of sets that contains Ω\Omega and is closed under complements and countable unions. That restriction is more than a technicality. It also carries meaning, as the next section explains.

σ-algebras as information

A useful way to think about a σ-algebra G⊆F\mathcal{G} \subseteq \mathcal{F} is as a state of knowledge: G\mathcal{G} contains exactly the events whose occurrence you can decide with the information it represents. For a coin tossed three times, the σ-algebra generated by the first toss contains only four events: the empty event, the whole space, “the first toss is heads” and “the first toss is tails”. A finer σ-algebra means more information.

Information that grows over time is described by a filtration: an increasing family of σ-algebras, F0⊆F1⊆⋯\mathcal{F}_0 \subseteq \mathcal{F}_1 \subseteq \cdots in discrete time or Fs⊆Ft\mathcal{F}_s \subseteq \mathcal{F}_t for s≤ts \le t in continuous time. A process is adapted if its value at each time tt is known at time tt. Adaptedness is exactly the condition that will make the Itô integral work: integrands may use the past, but never the future.

Random variables and their laws

A random variable is a function X:Ω→RX : \Omega \to \mathbb{R} that is measurable, meaning {X∈B}∈F\{X \in B\} \in \mathcal{F} for every Borel set BB, so that all of its probabilities are defined. Its law is the distribution μX(B)=P(X∈B)\mu_X(B) = \mathbb{P}(X \in B) that it induces on the real line, and σ(X)\sigma(X), the σ-algebra generated by XX, is the information obtained by observing it.

Two examples show how much a simple probability space can hold.

  • Uniform and Gaussian variables. Take Ω=[0,1]\Omega = [0, 1] with length as probability. Then U(ω)=ωU(\omega) = \omega is uniform, and X=Φ−1(U)X = \Phi^{-1}(U), where Φ\Phi is the standard normal distribution function, is a standard Gaussian. In the same way, any distribution on the real line can be produced from a single uniform variable.
  • Infinitely many coin tosses from one number. Write U=∑n≥1ξn2−nU = \sum_{n \ge 1} \xi_n 2^{-n} in binary. The digits ξ1,ξ2,…\xi_1, \xi_2, \dots are independent fair coin flips, so the unit interval already contains an infinite sequence of coin tosses. The partial sums Sn=∑k≤n(2ξk−1)S_n = \sum_{k \le n} (2\xi_k - 1) then form a simple random walk, the discrete ancestor of Brownian motion.

Expectation and convergence

The expectation E[X]=∫ΩX dP\mathbb{E}[X] = \int_\Omega X\,d\mathbb{P} is a Lebesgue integral. Its great advantage over the Riemann integral is how well it behaves under limits: the monotone and dominated convergence theorems give simple conditions under which E[lim⁡nXn]=lim⁡nE[Xn]\mathbb{E}[\lim_n X_n] = \lim_n \mathbb{E}[X_n].

Random variables can converge in several inequivalent senses:

  • almost surely, if Xn(ω)→X(ω)X_n(\omega) \to X(\omega) for every ω\omega outside an event of probability zero;
  • in probability, if P(∣Xn−X∣>ε)→0\mathbb{P}(|X_n - X| > \varepsilon) \to 0 for every ε>0\varepsilon > 0;
  • in LpL^p, if E∣Xn−X∣p→0\mathbb{E}|X_n - X|^p \to 0;
  • in distribution, if E[f(Xn)]→E[f(X)]\mathbb{E}[f(X_n)] \to \mathbb{E}[f(X)] for every bounded continuous function ff.

Almost sure convergence and LpL^p convergence each imply convergence in probability, which in turn implies convergence in distribution, and none of the reverse implications holds in general. The random walk illustrates two of them: Sn/n→0S_n / n \to 0 almost surely, by the strong law of large numbers, while Sn/nS_n / \sqrt{n} converges only in distribution, to N(0,1)N(0, 1), by the central limit theorem. In stochastic calculus, L2L^2 convergence does most of the work, because the Itô integral is constructed as an L2L^2 limit.

Conditional expectation

Given a σ-algebra G\mathcal{G}, the conditional expectation E[X∣G]\mathbb{E}[X \mid \mathcal{G}] is the best guess of XX that uses only the information in G\mathcal{G}. Formally, it is the G\mathcal{G}-measurable random variable YY that satisfies

E[X1A]=E[Y1A]for every A∈G.\mathbb{E}[X \mathbf{1}_A] = \mathbb{E}[Y \mathbf{1}_A] \quad \text{for every } A \in \mathcal{G}.

For square-integrable XX, it is the orthogonal projection of XX onto the subspace of G\mathcal{G}-measurable random variables, so it minimizes the mean squared error E[(X−Y)2]\mathbb{E}[(X - Y)^2] among all such YY. Three rules cover almost every calculation:

  • the tower property: E[E[X∣G]∣H]=E[X∣H]\mathbb{E}\big[\mathbb{E}[X \mid \mathcal{G}] \mid \mathcal{H}\big] = \mathbb{E}[X \mid \mathcal{H}] when H⊆G\mathcal{H} \subseteq \mathcal{G};
  • taking out what is known: E[ZX∣G]=Z E[X∣G]\mathbb{E}[ZX \mid \mathcal{G}] = Z\,\mathbb{E}[X \mid \mathcal{G}] when ZZ is G\mathcal{G}-measurable;
  • independence: E[X∣G]=E[X]\mathbb{E}[X \mid \mathcal{G}] = \mathbb{E}[X] when XX is independent of G\mathcal{G}.

A first glimpse of martingales

Apply these rules to the random walk with Fn=σ(ξ1,…,ξn)\mathcal{F}_n = \sigma(\xi_1, \dots, \xi_n). The next step is independent of the past and has mean zero and variance one, so

E[Sn+1∣Fn]=Sn,E[Sn+12−(n+1)∣Fn]=Sn2−n.\mathbb{E}[S_{n+1} \mid \mathcal{F}_n] = S_n, \qquad \mathbb{E}\big[S_{n+1}^2 - (n+1) \mid \mathcal{F}_n\big] = S_n^2 - n.

Processes with this “the best prediction of the future is the present” property are martingales. Both identities have exact analogues for Brownian motion, where BtB_t and Bt2−tB_t^2 - t are martingales, and the second one is the first sign of the quadratic variation that makes stochastic calculus different from ordinary calculus. The next post constructs Brownian motion itself.

← All posts in Stochastic differential equations