Continuous, yet broken everywhere
In the previous guide you built Brownian motion as the continuous limit of a random walk: shrink the time step and the jump size together in just the right way, and the jagged staircase melts into a single unbroken curve W(t). That limit hands you two facts that seem to fight each other. First, the path is continuous — there are no jumps, no gaps; you really could trace it with a pen without lifting the tip. Second, the path is nowhere differentiable — at no instant does it have a well-defined slope, velocity, or tangent line. Let that sink in: a curve you can draw in one stroke, yet which has a sharp corner at every single point.
Why must this happen? Recall the increment over a small time gap h has Gaussian size: W(t+h) - W(t) ~ Normal(0, h), so its typical magnitude is the standard deviation, of order sqrt(h). The would-be derivative is the ratio (W(t+h) - W(t)) / h, whose typical size is sqrt(h)/h = 1/sqrt(h). As h shrinks toward zero, 1/sqrt(h) blows up to infinity. The slope over a tiny window does not settle down to a number — it explodes. There is no limit, so there is no derivative. The very scaling that makes the path continuous (steps of size sqrt(h)) is exactly what makes it non-smooth (slopes of size 1/sqrt(h)).
Self-similar: a fractal that looks the same at every zoom
Why is there a corner everywhere and not just somewhere? Because the path looks statistically the same no matter how far you zoom in. This is the self-similarity of Brownian motion, and it has a precise scaling law. If you take the path and rescale time by a factor c, you must rescale space by sqrt(c) to recover the same kind of process: the rescaled path W(c*t)/sqrt(c) is again a standard Brownian motion. Speed up the clock by 100 and shrink the vertical axis by 10, and the new picture is statistically indistinguishable from the original.
Standard Brownian motion is self-similar:
W(c*t) / sqrt(c) has the same law as W(t)
Time stretches by c <--> space stretches by sqrt(c)
(this is the same sqrt-of-time scaling as Var(W(t)) = t)This is exactly what self-similarity means for a fractal: zoom in on any tiny slice of the path, rescale it by these matched factors, and you see a curve with the same wiggly character as the whole. There is no smooth scale lurking underneath — no magnification at which the jaggedness smooths out into a nice line, the way a circle eventually looks straight under a microscope. The roughness is present at every scale at once. That is why every point is a corner: each point, magnified, reveals the same storm of wiggling that the full path shows.
Infinite wiggling: length, variation, and quadratic variation
All this wiggling has a startling consequence for length. Take any time interval, say [0, 1], and try to measure how far the pen actually travels along the path — its total up-and-down distance, what mathematicians call the total variation. For a Brownian path this is infinite. Over [0, 1] the curve travels an infinite total distance even though it never strays far from the origin (its value stays of order 1). The path crams an unbounded amount of motion into a bounded region by wiggling on every scale at once.
But here is the deep twist that the next guide will make its whole subject. If ordinary length (the sum of the absolute increments) is infinite, try instead the sum of the SQUARES of the increments — the quadratic variation. Chop [0, t] into n tiny pieces and add up (change in W)^2 over each piece. For a smooth function this would shrink to zero (squaring tiny things makes them tinier). For Brownian motion it does NOT vanish: it converges to t, the length of the interval, with probability 1. The first power of the increments diverges to infinity; the second power converges to something finite and exact.
This single fact — that the quadratic variation of Brownian motion over [0, t] equals exactly t — is the hinge on which all of stochastic calculus turns. It is the reason ordinary calculus fails for these paths and the reason a new calculus (Ito's) is both necessary and possible. We will not prove it here; the next guide is devoted to it. For now, just hold the strange shape of the result in your mind: a curve so rough that its length is infinite, yet so regular that the sum of its squared wiggles lands precisely on t.
The path's secret life: zeros, records, and the arcsine surprise
The fine structure of a single sample path hides more surprises. Look at the set of times where the path is exactly back at zero. You might guess these zeros are scattered like isolated dots. In fact they form a strange dust: there are infinitely many of them in any interval right next to a zero, yet they contain no solid stretch of time — between any two zeros there is always a gap where the path is strictly positive or strictly negative. The zero set is uncountable, has zero total length, and is itself a fractal. Right after the path touches zero, it immediately crosses back and forth infinitely often, because the same 1/sqrt(h) explosion that kills the derivative also makes the path oscillate furiously near any point it visits.
Now a result so counterintuitive it deserves to be famous: the arcsine law. Over the interval [0, 1], ask what fraction of the time the path spends ABOVE zero (on the positive side). Your intuition shouts "about half — it is symmetric, after all." Wrong. The fraction of time spent positive is most likely to be near 0 or near 1, and least likely to be near 1/2. The typical Brownian path does not split its time evenly between the sides; it tends to commit to one side and linger there for most of the interval. The probability that the positive fraction is below x follows the arcsine distribution, whose density piles up at the two ends 0 and 1.
How big does it get? The law of the iterated logarithm
One last piece of geometry pins down the path's envelope — how high and low it ranges as time grows. We know the typical size at time t is sqrt(t), since Var(W(t)) = t. But the path is random, so it sometimes overshoots. How far can it stray above sqrt(t), again and again, forever? The answer is astonishingly precise. The law of the iterated logarithm says that as t grows, W(t) keeps brushing up against the envelope sqrt(2 * t * ln(ln(t))) — it touches that boundary infinitely often but essentially never exceeds it. The extra factor sqrt(2 * ln(ln(t))) over the typical sqrt(t) is tiny and grows agonizingly slowly, yet it captures the exact ceiling of the path's largest recurring excursions.
It is worth pausing on how sharp this is. The law of the iterated logarithm does not merely bound the path; it gives the exact constant. The value 2 inside the square root is not a loose estimate — the path comes arbitrarily close to sqrt(2 * t * ln(ln(t))) over and over (so any smaller envelope is breached infinitely often), yet for any envelope with a constant bigger than 2, the path eventually stays below it forever. This is the same flavour of result as the strong law of large numbers from earlier rungs: a statement that holds for the individual random path itself, with probability 1, not merely on average.
Step back and take in the whole creature. A Brownian path is continuous yet nowhere differentiable; self-similar and fractal at every scale; of infinite ordinary length but finite quadratic variation equal to t; its zeros a length-zero fractal dust; its time split lopsidedly by the arcsine law; and its growth fenced precisely by the iterated logarithm. None of this fits the smooth curves of ordinary calculus, where dt is negligible compared to dt and squared terms are thrown away. That mismatch is not a flaw — it is the doorway. In the next guide we make the mismatch precise through quadratic variation, and then build the Ito calculus that is custom-made for exactly this geometry.