Stochastic differential equations · Part 9 of 10

Girsanov's theorem

· 4 min read

Girsanov’s theorem says that adding a drift to a Brownian motion is the same as reweighting its paths. It turns hard problems about processes with drift into easier problems about Brownian motion, and it is the foundation of risk-neutral pricing.

Changing measure

If Z≥0Z \ge 0 is a random variable with E[Z]=1\mathbb{E}[Z] = 1, then Q(A)=E[Z1A]\mathbb{Q}(A) = \mathbb{E}[Z \mathbf{1}_A] defines a new probability measure, under which EQ[X]=E[ZX]\mathbb{E}_{\mathbb{Q}}[X] = \mathbb{E}[ZX]. We write Z=dQ/dPZ = d\mathbb{Q}/d\mathbb{P}. When information arrives over time, the density process Zt=E[Z∣Ft]Z_t = \mathbb{E}[Z \mid \mathcal{F}_t] is a martingale, and conditional expectations under Q\mathbb{Q} follow a version of Bayes’ rule: for s≤ts \le t and an Ft\mathcal{F}_t-measurable XX,

EQ[X∣Fs]=E[ZtX∣Fs]Zs.\mathbb{E}_{\mathbb{Q}}[X \mid \mathcal{F}_s] = \frac{\mathbb{E}[Z_t X \mid \mathcal{F}_s]}{Z_s}.

A discrete warm-up

Let ξ∼N(0,1)\xi \sim N(0, 1) under P\mathbb{P}, and reweight by Z=eμξ−μ2/2Z = e^{\mu\xi - \mu^2/2}. Completing the square,

E[Zf(ξ)]=∫f(x) eμx−μ2/2 e−x2/22π dx=∫f(x) e−(x−μ)2/22π dx,\mathbb{E}\big[Z f(\xi)\big] = \int f(x)\, \frac{e^{\mu x - \mu^2/2}\,e^{-x^2/2}}{\sqrt{2\pi}}\,dx = \int f(x)\, \frac{e^{-(x - \mu)^2/2}}{\sqrt{2\pi}}\,dx,

so under Q\mathbb{Q}, ξ∼N(μ,1)\xi \sim N(\mu, 1). An exponential tilt shifts the mean of a Gaussian without changing its variance. Do this independently for each step of a Gaussian random walk, with a shift that may depend on the past, and the walk acquires a drift. Girsanov’s theorem is the continuous-time limit of this observation.

The theorem

Let θt\theta_t be an adapted process, and define the stochastic exponential

Zt=exp⁡(∫0tθs dBs−12∫0tθs2 ds),Z_t = \exp\Big(\int_0^t \theta_s\,dB_s - \frac{1}{2}\int_0^t \theta_s^2\,ds\Big),

which solves dZt=θtZt dBtdZ_t = \theta_t Z_t\,dB_t by Itô’s formula.

Girsanov’s theorem. If ZZ is a martingale on [0,T][0, T], define Q\mathbb{Q} by dQ/dP=ZTd\mathbb{Q}/d\mathbb{P} = Z_T. Then under Q\mathbb{Q}, the process

B~t=Bt−∫0tθs ds,0≤t≤T,\widetilde{B}_t = B_t - \int_0^t \theta_s\,ds, \qquad 0 \le t \le T,

is a Brownian motion.

A convenient sufficient condition for ZZ to be a true martingale, and not just a local one, is Novikov’s condition, Eexp⁡(12∫0Tθs2 ds)<∞\mathbb{E}\exp\big(\frac{1}{2}\int_0^T \theta_s^2\,ds\big) < \infty. It certainly holds when θ\theta is bounded.

Why it works. By Lévy’s characterization, it is enough to show that B~\widetilde{B} is a local martingale under Q\mathbb{Q}, because its quadratic variation is tt under any equivalent measure. By Bayes’ rule, that is the same as showing that ZtB~tZ_t \widetilde{B}_t is a local martingale under P\mathbb{P}. The product rule gives

d(ZtB~t)=Zt (dBt−θt dt)+B~t θtZt dBt+θtZt dt=Zt (1+θtB~t) dBt,d(Z_t \widetilde{B}_t) = Z_t\,(dB_t - \theta_t\,dt) + \widetilde{B}_t\,\theta_t Z_t\,dB_t + \theta_t Z_t\,dt = Z_t\,(1 + \theta_t \widetilde{B}_t)\,dB_t,

where the term θtZt dt\theta_t Z_t\,dt in the middle is the covariation d⟨Z,B~⟩td\langle Z, \widetilde{B} \rangle_t. The dtdt terms cancel exactly, leaving a stochastic integral.

What can and cannot be changed

Girsanov’s theorem changes drifts, never diffusion coefficients. Quadratic variation can be read off a single path, because sums of squared increments converge to ⟨X⟩t\langle X \rangle_t almost surely along suitable partitions, and events of probability one are the same under equivalent measures. So the laws of BB and 2B2B on [0,T][0, T] are mutually singular: no reweighting turns one into the other.

Even drifts have limits. Shifting Brownian motion by a deterministic path hh with h(0)=0h(0) = 0 gives an equivalent measure exactly when ∫0Th′(t)2 dt<∞\int_0^T h'(t)^2\,dt < \infty, which is the Cameron–Martin theorem; otherwise the shifted law is singular.

Applications

Hitting times with drift. The law of Xt=Bt+μtX_t = B_t + \mu t on [0,T][0, T] has density eμBT−μ2T/2e^{\mu B_T - \mu^2 T/2} with respect to the law of Brownian motion. Any question about the drifted process becomes an expectation over Brownian paths with an extra weight. Combined with the reflection principle, this gives, for example, the distribution of the first time a Brownian motion with drift reaches a level.

Weak solutions without Lipschitz conditions. To solve dXt=b(Xt) dt+dBtdX_t = b(X_t)\,dt + dB_t for a drift bb that is only bounded and measurable, start with a Brownian motion XX and change measure with θt=b(Xt)\theta_t = b(X_t). Under the new measure, Xt−∫0tb(Xs) dsX_t - \int_0^t b(X_s)\,ds is a Brownian motion, which says exactly that XX is a weak solution.

Likelihoods for diffusions. Girsanov densities are likelihood ratios between path measures. For dXt=β b(Xt) dt+dBtdX_t = \beta\,b(X_t)\,dt + dB_t observed continuously on [0,T][0, T], the log-likelihood of β\beta relative to β=0\beta = 0 is

β∫0Tb(Xt) dXt−β22∫0Tb(Xt)2 dt,\beta\int_0^T b(X_t)\,dX_t - \frac{\beta^2}{2}\int_0^T b(X_t)^2\,dt,

a quadratic function of β\beta that is maximized in closed form.

Risk-neutral pricing. In the Black–Scholes model, dSt=μSt dt+σSt dBtdS_t = \mu S_t\,dt + \sigma S_t\,dB_t. Changing measure with θ=(r−μ)/σ\theta = (r - \mu)/\sigma turns the drift μ\mu into the interest rate rr, so that discounted prices e−rtSte^{-rt}S_t become martingales. Prices of derivatives are then discounted expectations under this risk-neutral measure, and investors’ attitudes to risk drop out entirely.

Rare-event simulation. To estimate a small probability such as P(max⁡t≤TBt>a)\mathbb{P}(\max_{t \le T} B_t > a) for a large level aa, simulate under a measure that pushes paths toward the rare event, and correct each sample with its Girsanov weight. This is importance sampling on path space, and a good choice of drift can reduce the variance by orders of magnitude.

← All posts in Stochastic differential equations