Infinitely Divisible Laws & Lévy Processes

the inverse-Gaussian process

The inverse-Gaussian process is the subordinator whose increments are inverse-Gaussian distributed — and the name is literal: it is the process of first-passage times of a Brownian motion with drift to successive levels. It is a natural increasing random clock and, used to subordinate Brownian motion, it produces the normal-inverse-Gaussian (NIG) model, a flexible finance model competing with variance-gamma.

Concretely, if W_t = mu t + B_t is Brownian motion with positive drift mu, then T_a = inf{t : W_t = a}, the first time it reaches level a, defines an increasing process in a; the increment T_{a+da} - T_a is inverse-Gaussian. As a Lévy process T_a has Laplace exponent Phi(u) = delta(sqrt(2u + gamma^2) - gamma) (with parameters delta, gamma > 0), Lévy measure nu(dx) = (delta / sqrt(2 pi)) x^(-3/2) e^(-gamma^2 x / 2) dx on x > 0, no drift, no Gaussian part. The x^(-3/2) singularity at zero makes nu infinite (infinitely many small jumps) but integrable against x, so the process is finite variation, like the gamma process. The inverse-Gaussian law is itself the canonical first-passage distribution, infinitely divisible, with explicit density and moment generating function.

Subordinating Brownian motion by an IG process gives the normal-inverse-Gaussian (NIG) Lévy process of Barndorff-Nielsen, which has semi-heavy (exponentially-tilted) tails and is heavily used to model financial returns and turbulence. The honest point: 'inverse-Gaussian' does not mean the reciprocal of a Gaussian — the name comes from the inverse relationship between the cumulant generating functions of the first-passage time and the Gaussian, a historical naming that regularly confuses newcomers.

Let W_t = t + B_t hit level a for the first time at T_a. Then T_a is inverse-Gaussian with mean a (the time to drift distance a) and the process a -> T_a is the IG subordinator. Subordinating an independent Brownian motion B'_{T_a} yields the NIG process, whose tails decay exponentially-times-polynomially — lighter than stable, heavier than Gaussian.

First-passage times of drifting Brownian motion form the IG subordinator.

Inverse-Gaussian does NOT mean 1/Gaussian; the name reflects an inverse relationship between cumulant generating functions. The IG process has infinite Lévy measure but finite variation, and like all subordinators has no Gaussian part.

Also called
IG processfirst-passage subordinatorIG 過程