Stochastic differential equations · Part 7 of 10

When diffusions explode: Feller's test

· 4 min read

The existence theorem for stochastic differential equations assumes coefficients that grow at most linearly. Many natural models violate that assumption, and their solutions can run off to infinity in finite time. In one dimension there is a complete answer to whether this happens, due to William Feller.

Solutions up to explosion

Suppose bb and σ\sigma are locally Lipschitz on an interval I=(ℓ,r)I = (\ell, r), whose endpoints may be infinite, and that σ>0\sigma > 0 on II. Applying the existence theorem on larger and larger closed subintervals gives a unique solution up to the explosion time

S=lim⁡n→∞τn,S = \lim_{n \to \infty} \tau_n,

where τn\tau_n is the first time the solution leaves the nn-th subinterval. The solution explodes if S<∞S < \infty, which means XtX_t reaches an endpoint of II in finite time. When I=RI = \mathbb{R}, this is a genuine blow-up to +∞+\infty or −∞-\infty.

Two key functions

Fix a reference point c∈Ic \in I. The scale function is

p(x)=∫cxexp⁡(−∫cy2b(z)σ2(z) dz) dy.p(x) = \int_c^x \exp\Big(-\int_c^y \frac{2b(z)}{\sigma^2(z)}\,dz\Big)\,dy.

It is built so that Lp=bp′+12σ2p′′=0\mathcal{L}p = bp' + \frac{1}{2}\sigma^2 p'' = 0, which makes p(Xt)p(X_t) a local martingale: in the coordinate pp, the process has no drift. Hitting probabilities are ratios of differences of the scale function, as the interval example in the previous post showed for Brownian motion with drift.

Scale describes where the process goes, but not how fast. Timing enters through the speed measure m(dx)=2 dxp′(x) σ2(x)m(dx) = \frac{2\,dx}{p'(x)\,\sigma^2(x)}, and the two combine into the function

v(x)=∫cxp′(y)∫cy2 dzp′(z) σ2(z) dy.v(x) = \int_c^x p'(y) \int_c^y \frac{2\,dz}{p'(z)\,\sigma^2(z)}\,dy.

This function is nonnegative on both sides of cc, and it measures how long the process takes to travel toward each endpoint.

Feller’s test

Theorem. The solution never explodes, that is P(S=∞)=1\mathbb{P}(S = \infty) = 1 for every starting point in II, if and only if

v(ℓ+)=∞andv(r−)=∞.v(\ell+) = \infty \quad \text{and} \quad v(r-) = \infty.

If vv is finite at an endpoint, explosion through that endpoint has positive probability, and the scale function decides how likely it is. For example, if p(r−)<∞p(r-) < \infty and p(ℓ+)=−∞p(\ell+) = -\infty, the process drifts toward rr, and it explodes with probability one exactly when v(r−)<∞v(r-) < \infty.

The intuition is that a finite v(r−)v(r-) means the expected time to travel all the way to rr is finite. A process can be transient, drifting toward an endpoint forever, without ever reaching it in finite time.

Examples

Brownian motion. With b=0b = 0 and constant σ\sigma, the scale function is p(x)=xp(x) = x and v(x)=x2/σ2v(x) = x^2/\sigma^2, which is infinite at both ends. There is no explosion, as expected.

Brownian motion with drift. With constants μ>0\mu > 0 and σ\sigma, write β=2μ/σ2\beta = 2\mu/\sigma^2. Then p(x)=(1−e−βx)/βp(x) = (1 - e^{-\beta x})/\beta is bounded above, so the process drifts to +∞+\infty, but v(±∞)=∞v(\pm\infty) = \infty, so it never gets there in finite time. This is transience without explosion.

A superlinear drift, with and without noise. The deterministic equation x′=x2x' = x^2 blows up. Adding noise does not help. For dXt=Xt2 dt+dBtdX_t = X_t^2\,dt + dB_t, the scale density is p′(x)=e−2x3/3p'(x) = e^{-2x^3/3}, which is integrable at +∞+\infty but not at −∞-\infty, so the process is pushed toward +∞+\infty. For large yy, the inner integral in vv behaves like e2y3/3/y2e^{2y^3/3}/y^2, so the integrand of vv decays like 1/y21/y^2 and v(+∞)<∞v(+\infty) < \infty. The process explodes with probability one. Noise can prevent explosion only when it is strong enough to compete with the drift, for example when σ\sigma also grows quickly, and Feller’s test quantifies exactly how strong it needs to be.

Reaching a boundary at zero. The same functions decide whether a process on (0,∞)(0, \infty) can hit zero. Geometric Brownian motion never does. The squared Bessel process dXt=δ dt+2Xt dBtdX_t = \delta\,dt + 2\sqrt{X_t}\,dB_t hits zero when its dimension δ\delta is below 2 and never hits it when δ≥2\delta \ge 2. The threshold matches the fact that a two-dimensional Brownian motion never hits a given point.

Why it matters

Explosion is not just a pathology. Models in finance, population biology and physics often have superlinear coefficients, and Feller’s test tells you whether such a model makes sense for all time. It is also a good example of a general theme: in one dimension, the scale function and speed measure determine essentially everything about a diffusion, which is why one-dimensional diffusions are so much better understood than higher-dimensional ones.

← All posts in Stochastic differential equations