the law of the iterated logarithm

The law of the iterated logarithm pins down the EXACT envelope of the fluctuations of Brownian motion — not just the typical size sqrt(t) of B_t (that is the central limit scaling), but the precise extreme size of the wildest oscillations, almost surely. It sits between the law of large numbers (which says B_t / t -> 0) and the central limit theorem (which says B_t / sqrt(t) is normal), giving the sharp almost-sure boundary the path keeps brushing against but does not cross.

The statement at infinity is: almost surely, limsup_(t -> infinity) B_t / sqrt(2 t log log t) = +1 and liminf_(t -> infinity) B_t / sqrt(2 t log log t) = -1. The function phi(t) = sqrt(2 t log log t) is the exact upper envelope: the path B_t exceeds (1 - epsilon) phi(t) infinitely often but exceeds (1 + epsilon) phi(t) only finitely often, for every epsilon > 0. By Brownian time-inversion (t B_(1/t) is again a Brownian motion) there is a mirror statement at t = 0: almost surely limsup_(t -> 0+) B_t / sqrt(2 t log log (1/t)) = +1, controlling the fine local roughness near a point. The proof combines an exponential (Gaussian tail) upper bound summed along a geometric subsequence (Borel-Cantelli) with an independent-increments lower bound, and the constant 2 inside the square root is exactly what makes the upper and lower halves match.

Why it matters: the LIL is the gold standard for an almost-sure rate, used to calibrate strong approximation (KMT) theorems, iterated random walks, and the precision of Monte Carlo. The double logarithm is genuinely delicate — replacing it by log t or by a constant gives a false boundary. The honest content is that the bound is achieved infinitely often but the constant 1 is never beaten: the path returns arbitrarily close to its envelope again and again, forever, yet stays inside (1 + epsilon) of it eventually. The LIL is about the limsup of fluctuations, not the typical value, and it says nothing about any single time t.