Brownian motion (the Wiener process)
/ VEE-ner /
Drop a speck of pollen in still water and watch it under a microscope: it jitters, drifting in tiny random jerks in every direction, never settling, the path a frantic crinkly mess. That jiggle is what the botanist Robert Brown saw in 1827, and Brownian motion is the idealised mathematical model of it. Picture a tireless drunkard taking infinitely many infinitely tiny random steps per second — the smooth-looking but wildly wiggly curve traced out is a Brownian path. It is the single most important continuous-time random process, the bridge from coin-flip randomness to the calculus of finance and physics.
Formally a standard Brownian motion is a random function of time B(t), with t at least 0, fixed by four properties. It starts at the origin: B(0) = 0. Its increments are independent: the change B(t) - B(s) over a future window tells you nothing about the change over an earlier, non-overlapping window. Its increments are stationary and Gaussian: the change over any window of length h is Normal(0, h) — mean zero, variance exactly equal to the elapsed time h, so the typical distance wandered grows like the square root of time. And its paths are continuous: you can draw the whole curve without lifting your pen. These four demands — start at zero, independent increments, Gaussian increments with variance equal to elapsed time, continuous paths — pin the process down completely.
Brownian motion sits at the centre of probability because nearly everything continuous and random is built from it. It is the scaling limit of the simple random walk (Donsker's theorem), so it is the continuous-time CLT: zoom out on any well-behaved random walk and it looks Brownian. It is simultaneously a martingale and a Gaussian process. Its strange paths — continuous everywhere yet differentiable nowhere — force a brand-new calculus (the Ito calculus) on us. And from it spring the models of diffusing heat, fluctuating stock prices, and noisy physical systems. The honest warning: everything in this field is the deep end of probability, sketched here for intuition; the rigorous constructions and proofs need real measure theory.
Track a stock-price log over a year as B(t) with t in years. The change over the second half-year, B(1) - B(0.5), is Normal(0, 0.5): it has mean 0 and standard deviation about sqrt(0.5) = 0.71, and — by independent increments — it is statistically unrelated to what happened in the first half-year, B(0.5) - B(0).
Over a window of length h the displacement is Normal(0, h), independent of the past.
Variance equals the elapsed time, so the standard deviation grows like the square root of time — not linearly. The drunkard wanders, but slowly; doubling the time only multiplies the typical distance by sqrt(2).