Shrink the time step and most functions flatten; Brownian motion refuses to.

What it is

Brownian motion is a random walk with variance dt\dt for a walk with time step dt\dt. It's known as a Wiener process and often represented using WW. There are six key properties of the walk:

  1. It starts at zero, W0=0W_0 = 0.
  2. Its increments are independent of each other — what WW does over the interval [s,t][s, t] is independent of anything it did before ss. This also makes it a Markov process (the future is only dependent on the present, not the past — it retains no memory).
  3. It is a martingale (the best prediction of the future is the present — E[Xt∣Fs]=Xs\E[X_t \mid \mathcal{F}_s] = X_s for all s<ts < t).
  4. Step increments are Gaussian with variance equal to elapsed time dt\dt.
  5. It is continuous everywhere with no jumps or discontinuities.
  6. It is not differentiable — the slope grows towards infinity, with 1/dt1/\sqrt{\dt} as dt→0\dt \to 0.

Variances subtract, standard deviations don't

W(0.5)−W(0.25)W(0.5) - W(0.25) is normally distributed with variance 0.5−0.25=0.250.5 - 0.25 = 0.25 and standard deviation 0.25=0.5\sqrt{0.25} = 0.5. Variances subtract, standard deviations don't.

Simulating it

Build the path with time steps dt\dt, then ΔW∼N(0,dt)\Delta W \sim N(0, \dt) — Gaussian because that is the distribution the continuous limit produces from any finite-variance step, according to the central limit theorem.

So the path becomes ∑ΔW\sum \Delta W, which is the cumulative sum of the standard deviation of WW multiplied by a Gaussian variable ZZ, or — ∑dt⋅Z\sum \sqrt{\dt} \cdot Z.

Excess kurtosis by lag, two walks

Excess kurtosis of the increment over kk steps, 40 000 paths of 2 000 steps at dt=10−3\dt = 10^{-3}, seed 20260903 — from the committed experiments/2_2_brownian_motion.json. Zero is Gaussian. The coin column is the central limit theorem arriving.

lag kk (steps)coin walkGaussian walk
1−2.000−0.0005
10−0.198−0.0005
100−0.023−0.0045
250−0.001−0.0018

Continuous everywhere, differentiable nowhere

W(t+dt)−W(t)dt∼W(t)+W(dt)−W(t)dtW(t+dt)−W(t)dt∼W(dt)dt\begin{aligned} \frac{W(t+\dt) - W(t)}{\dt} &\sim \frac{W(t) + W(\dt) - W(t)}{\dt} \\ \frac{W(t+\dt) - W(t)}{\dt} &\sim \frac{W(\dt)}{\dt} \end{aligned}

The typical size of the slope is therefore dt/dt=1/dt\sqrt{\dt} / \dt = 1/\sqrt{\dt}, so as dt→0\dt \to 0 the slope →\to infinity — a smooth function's slope tends to its derivative.

There are no jumps, no discontinuities, since W(t+dt)−W(t)→0W(t+\dt) - W(t) \to 0 as dt→0\dt \to 0 — the data shows that with dt\sqrt{\dt} sampling: the largest move was 0.0056 at dt=10−6\dt = 10^{-6}.

Move and slope against dt

100 paths on a fine grid of step 10−610^{-6} over [0,1][0, 1], measured over every non-overlapping window of width dt\dt — the experiment's h, and the row labels below — seed 20260903, from the committed experiments/2_2_brownian_motion.json. Read down: the move shrinks, the slope grows.

dt\dtmean ∣ΔW∣\lvert \Delta W \rvertmean ∣ΔW/dt∣\lvert \Delta W / \dt \rvertlargest ∣ΔW∣\lvert \Delta W \rvert
10−110^{-1}0.24782.4781.0521
10−210^{-2}0.079237.9230.3492
10−310^{-3}0.0251625.160.1414
10−410^{-4}0.00797779.770.04842
10−510^{-5}0.002523252.30.01589
10−610^{-6}0.0007979797.90.005551

dW\dW is a dt\sqrt{\dt}-sized random increment.

The same at every magnification

D1Random walk → Brownian motion
 

Zoom keeps the first 1/z of the path and stretches space by √z — the scaling Brownian motion is invariant under. The zoomed picture is as rough as the whole one; that roughness is what stops the path having a derivative.

Zooming into the first 1 % of a Brownian path only looks "small" because its variance is 1/c1/c; multiplying the vertical axis by c\sqrt{c} (1010 for a 1 % zoom) restores the variance to 11, showing that the zoomed segment has the same roughness as the full path. Because every zoomed-in piece, once scaled by c\sqrt{c}, statistically reproduces the whole path, Brownian motion never becomes smooth at any magnification.

The best forecast is where you are

E[W(2)∣W(1)=1.5]=1.5\E[W(2) \mid W(1) = 1.5] = 1.5 because this is a martingale — the expected value of W(2)W(2), given the known value of W(1)W(1), is the known value of W(1)W(1). The spread is 1 because WW is a Markov process — no memory: the variance at W(2)W(2) is for the time step 2−1=12 - 1 = 1, therefore the spread is 1=1\sqrt{1} = 1.

Into 2.3

WW is continuous but infinitely rough, so the usual idea "differentiate first, then integrate" is impossible. With no derivative to supply a term like W′(t) dtW'(t)\,\dt, calculus on WW must instead be built from its increments — the dt\sqrt{\dt}-sized kicks whose variance is dt\dt — and from how those increments accumulate. Section 2.3 begins by replacing classical derivatives with this increment-based geometry, and begins to build the rules of stochastic calculus from that starting point.

Check

W is a Brownian motion. What is the distribution of W(0.5) − W(0.25)?

At a fixed time t, you measure the secant slope (W(t+dt) − W(t))/dt for smaller and smaller dt. As dt → 0 the typical size of the slope:

A Brownian path has reached W(1) = 1.5. Seen from now, W(2) has expected value and standard deviation:

You zoom into the first 1 % of a Brownian path and stretch time by 100 to fill the screen. For the picture to look statistically like the whole path, the vertical axis must be stretched by:

Build a walk from ±1 coin flips with √dt scaling, and look at its increments over 250 steps. Their distribution is: