Stochastic differential equations · Part 10 of 10

Lévy processes and jumps

· 5 min read

Brownian motion has independent, stationary increments and continuous paths. Drop the continuity requirement and you get a much larger family, the Lévy processes, which model sudden events: insurance claims, market crashes, neuron spikes, or the random event times that drive the Zig-Zag and bouncy particle samplers in the MCMC playground.

Definition

A Lévy process is a process (Xt)t≥0(X_t)_{t \ge 0} with X0=0X_0 = 0 such that

  1. increments over disjoint time intervals are independent;
  2. the distribution of Xt+s−XtX_{t+s} - X_t depends only on ss;
  3. it is continuous in probability: Xt+s−Xt→0X_{t+s} - X_t \to 0 in probability as s→0s \to 0.

Every Lévy process has a version whose paths are right-continuous with left limits. Continuity in probability still allows jumps; it only rules out jumps at fixed times. The only Lévy processes with continuous paths are Brownian motions with drift, bt+σBtbt + \sigma B_t.

The Poisson process

The Poisson process NtN_t with rate λ\lambda counts random arrivals. The times between arrivals are independent exponential random variables with mean 1/λ1/\lambda, and

P(Nt=k)=e−λt(λt)kk!,k=0,1,2,…\mathbb{P}(N_t = k) = e^{-\lambda t}\frac{(\lambda t)^k}{k!}, \qquad k = 0, 1, 2, \dots

It is the jump counterpart of Brownian motion: the compensated process Nt−λtN_t - \lambda t is a martingale, and its quadratic variation is NtN_t itself.

Compound Poisson processes

Attach an independent random size YiY_i, with distribution FF, to the ii-th arrival:

Xt=∑i=1NtYi.X_t = \sum_{i=1}^{N_t} Y_i.

Writing ν=λF\nu = \lambda F for the Lévy measure, which gives the rate at which jumps of each size occur, the characteristic function is

E[eiuXt]=exp⁡(t∫(eiuy−1) ν(dy)).\mathbb{E}\big[e^{iuX_t}\big] = \exp\Big(t \int \big(e^{iuy} - 1\big)\,\nu(dy)\Big).

A compound Poisson process makes only finitely many jumps in any bounded time interval.

Poisson random measures

A cleaner bookkeeping device records each jump as a point (t,y)(t, y) in time–size space. The random measure N(dt,dy)N(dt, dy) counts the points in each region of that space. It is a Poisson random measure with mean measure dt ν(dy)dt\,\nu(dy): the counts in disjoint regions are independent Poisson random variables. Functions of the jumps become integrals, such as

∫0t∫f(y) N(ds,dy)=∑s≤t, ΔXs≠0f(ΔXs),\int_0^t \int f(y)\,N(ds, dy) = \sum_{s \le t,\ \Delta X_s \ne 0} f(\Delta X_s),

and the compensated measure N~(dt,dy)=N(dt,dy)−dt ν(dy)\widetilde{N}(dt, dy) = N(dt, dy) - dt\,\nu(dy) produces martingales, just as Nt−λtN_t - \lambda t does.

Infinitely many small jumps

A Lévy measure may have infinite total mass near zero, which means infinitely many small jumps in every time interval. The plain sum of the jump sizes may then diverge, but the compensated sum still converges in L2L^2, provided

∫∣y∣<1y2 ν(dy)<∞.\int_{|y| < 1} y^2\,\nu(dy) < \infty.

Small jumps can be as numerous as they like, as long as their squares are summable, while large jumps (say ∣y∣≥1|y| \ge 1) are always finite in number on bounded time intervals. So the admissible Lévy measures are exactly those with ∫min⁡(1,y2) ν(dy)<∞\int \min(1, y^2)\,\nu(dy) < \infty.

The characterization

The Lévy–Khintchine formula. For every Lévy process, E[eiuXt]=etψ(u)\mathbb{E}[e^{iuX_t}] = e^{t\psi(u)}, where

ψ(u)=ibu−12σ2u2+∫(eiuy−1−iuy 1{∣y∣<1}) ν(dy)\psi(u) = ibu - \frac{1}{2}\sigma^2 u^2 + \int \Big(e^{iuy} - 1 - iuy\,\mathbf{1}_{\{|y| < 1\}}\Big)\,\nu(dy)

for a drift bb, a Gaussian variance σ2≥0\sigma^2 \ge 0 and a Lévy measure ν\nu. Conversely, every such triple (b,σ2,ν)(b, \sigma^2, \nu) comes from some Lévy process.

The Lévy–Itô decomposition. Path by path, the process splits into four independent pieces:

Xt=bt+σBt+∫0t∫∣y∣≥1y N(ds,dy)+∫0t∫∣y∣<1y N~(ds,dy).X_t = bt + \sigma B_t + \int_0^t \int_{|y| \ge 1} y\,N(ds, dy) + \int_0^t \int_{|y| < 1} y\,\widetilde{N}(ds, dy).

They are a drift, a Brownian motion, a compound Poisson process of big jumps, and a compensated martingale made of small jumps.

Examples

  • The gamma process. With ν(dy)=a y−1e−by dy\nu(dy) = a\,y^{-1}e^{-by}\,dy on y>0y > 0, there are infinitely many jumps in every interval, yet ∫01y ν(dy)<∞\int_0^1 y\,\nu(dy) < \infty, so the jumps are summable and the paths are increasing. Its increments are gamma distributed, which makes it a popular model for accumulated damage and for random clocks.
  • Stable processes. With ν(dy)∝∣y∣−1−α dy\nu(dy) \propto |y|^{-1-\alpha}\,dy for some 0<α<20 < \alpha < 2, the process satisfies a scaling law like Brownian motion’s, but with heavy tails: XtX_t has infinite variance, and infinite mean when α≤1\alpha \le 1. The Cauchy process is the case α=1\alpha = 1.
  • First passage times of Brownian motion. For a Brownian motion BB, let TℓT_\ell be the first time it reaches the level ℓ≥0\ell \ge 0. By the strong Markov property, the increments Tℓ2−Tℓ1T_{\ell_2} - T_{\ell_1} are independent and stationary in ℓ\ell, so ℓ↦Tℓ\ell \mapsto T_\ell is itself a Lévy process. It is increasing and moves only by jumps, each corresponding to an excursion of BB below its running maximum. A continuous process hides a jump process inside it.

Itô’s formula with jumps

For a process with jumps, Itô’s formula keeps the continuous part and adds a correction at each jump:

f(Xt)=f(X0)+∫0tf′(Xs−) dXs+σ22∫0tf′′(Xs−) ds+∑s≤t(f(Xs)−f(Xs−)−f′(Xs−) ΔXs).f(X_t) = f(X_0) + \int_0^t f'(X_{s-})\,dX_s + \frac{\sigma^2}{2}\int_0^t f''(X_{s-})\,ds + \sum_{s \le t}\Big(f(X_s) - f(X_{s-}) - f'(X_{s-})\,\Delta X_s\Big).

The sum replaces the second-order Taylor term at the jumps, which are too large for a Taylor expansion to be accurate. With this formula, much of the collection extends to jump diffusions: stochastic differential equations driven by both a Brownian motion and a Poisson random measure, whose generators include an integral term for the jumps.

← All posts in Stochastic differential equations