The Markov Property

A process is Markov when the future depends on the past only through the present. Knowing the entire history up to time t tells you nothing more about what happens next than simply knowing where you are right now, X_t. For every horizon s > 0 and every event A,

\mathbb{P}\big(X_{t+s} \in A \mid \mathcal{F}_t\big) = \mathbb{P}\big(X_{t+s} \in A \mid X_t\big).

This is the memoryless property: the current state X_t screens off the past. Once you condition on it, the route you took to get there is irrelevant to the law of the future.

Past, present, fan of futures

Step through the picture. First a sample path is drawn up to a present time t; then that past is greyed out; then a fan of possible continuations springs from the present point. The fan would look identical for any path that ended at the same X_t — that is the Markov property made visible.

Brownian motion is Markov

Brownian motion is the archetypal Markov process, and the reason is its independent increments. Let us prove it carefully, one line at a time, so that nothing is smuggled in. Fix a present time t, a horizon s > 0, and write \mathcal{F}_t for the history of the path up to t (the filtration). We want the future value W_{t+s}.

Step 1 — split the future into present plus increment. Add and subtract the present value. This is just algebra and is always true:

W_{t+s} = \underbrace{W_t}_{\text{present}} + \underbrace{(W_{t+s} - W_t)}_{\text{the increment}}.

Step 2 — name what each piece knows. The present value W_t is \mathcal{F}_t-measurable — it is part of the history, so conditioning on \mathcal{F}_t treats it as a known constant. The increment W_{t+s} - W_t, by the independent-increments property of Brownian motion, is independent of \mathcal{F}_t:

(W_{t+s} - W_t) \perp \mathcal{F}_t.

Step 3 — condition on the whole history. Take the conditional law of W_{t+s} given \mathcal{F}_t. The first term is a known constant W_t; the second is independent of \mathcal{F}_t, so conditioning on \mathcal{F}_t does nothing to it — its conditional law equals its unconditional law, which is N(0, s) (a Brownian increment over a gap of length s is centred at 0 with variance s). A constant plus an independent N(0,s) is Gaussian, recentred at the constant:

\mathcal{L}\big(W_{t+s} \mid \mathcal{F}_t\big) = N\big(W_t,\; s\big).

Step 4 — read off the only input that survived. The right-hand side mentions the past only through the single number W_t; the rest of the history \mathcal{F}_t has dropped out entirely. So conditioning on the full history gives the same answer as conditioning on the present value alone:

\mathbb{P}\big(W_{t+s} \in A \mid \mathcal{F}_t\big) = \mathbb{P}\big(W_{t+s} \in A \mid W_t\big).

That is exactly the Markov property. Independent increments did all the work: they are precisely what makes the route taken irrelevant once you know where you are now.

Brownian motion (W_t)_{t\ge 0} is a Markov process. For every t \ge 0 and horizon s > 0, the conditional law of W_{t+s} given the history \mathcal{F}_t depends on that history only through the present value, and is Gaussian: \mathcal{L}\big(W_{t+s} \mid \mathcal{F}_t\big) = N\big(W_t,\, s\big), \qquad\text{hence}\qquad \mathbb{P}\big(W_{t+s} \in A \mid \mathcal{F}_t\big) = \mathbb{P}\big(W_{t+s} \in A \mid W_t\big).

For a Markov process the future law is captured by a transition probability p(s, x, t, y) — the probability density (or kernel) of being at y at time t given you were at x at the earlier time s. Because the present screens off the past, these kernels are all you need; they chain together through the Chapman–Kolmogorov relation p(s, x, u, z) = \int p(s, x, t, y)\, p(t, y, u, z)\, dy, which is just the Markov property written for densities. For Brownian motion the kernel is the Gaussian p(s,x,t,y) = \tfrac{1}{\sqrt{2\pi(t-s)}}\, e^{-(y-x)^2 / 2(t-s)}.

The property we proved fixes the present time t in advance. The strong Markov property upgrades this: it lets you restart the clock not at a deterministic instant but at a stopping time \tau — a random time whose arrival you can recognise without peeking into the future. The claim is that, conditioned on \mathcal{F}_\tau (everything known up to \tau), the shifted process

\widetilde W_s := W_{\tau + s} - W_\tau, \qquad s \ge 0,

is again a Brownian motion, started afresh at the origin and independent of the past \mathcal{F}_\tau. In other words the process forgets its history not only at fixed clock times but at any rule-based moment — the instant it first hits a level, say. This is not automatic: it holds because Brownian paths are continuous and a stopping time can be approximated from above by discrete times, at each of which the ordinary (fixed-time) Markov property applies. The strong version is what makes hitting-time and reflection arguments work, and it is the engine behind much of the pricing of path-dependent options.