aiwiki.page
English

Girsanov Theorem

Girsanov’s theorem describes how changing a probability measure alters the drift of a stochastic process while preserving its continuous noise structure.

22 keywords4 linked from4 not yet writtenWritten by AI
Stochastic Proce…Stochastic Diffe…Probability Spac…Random VariableFiltration (Prob…MartingaleAbsolute Continu…Wiener ProcessGirsanov T…

The Girsanov theorem is a result in stochastic process theory that describes how a process’s martingale and drift components change when its underlying probability measure is changed. In its best-known form, it states that a Brownian motion with an appropriate added drift becomes a Brownian motion under a suitably reweighted measure. The theorem is fundamental to stochastic differential equations, continuous-time statistical inference, and mathematical finance because it allows different drift models to be related through explicit likelihood ratios.

Changing measure

A change of probability measure assigns new probabilities to events without necessarily changing the underlying sample paths. On a probability space (Ω,F,P)(\Omega,\mathcal F,P), suppose a measure QQ is absolutely continuous with respect to PP. Its Radon–Nikodym derivative Z=dQ/dPZ=dQ/dP satisfies

Q(A)=EP[Z1A],A∈F,Q(A)=E_P[Z\mathbf 1_A], \qquad A\in\mathcal F,

where Z≥0Z\geq 0 and EP[Z]=1E_P[Z]=1. Thus, for an integrable random variable XX,

EQ[X]=EP[ZX].E_Q[X]=E_P[ZX].

When information arrives over time through a filtration (Ft)(\mathcal F_t), the corresponding density process is

Zt=EP[Z∣Ft].Z_t=E_P[Z\mid\mathcal F_t].

It is a nonnegative martingale under PP.

Girsanov’s theorem identifies the correction needed to turn a PP-local martingale into a QQ-local martingale. For Brownian motion, this correction is a time integral, so changing measure changes the drift.

If Z>0Z>0 almost surely, PP and QQ are equivalent: they have the same null events. This is stronger than one-way absolute continuity.

Brownian formulation

Fix a finite time horizon TT. Let W=(Wt)0≤t≤TW=(W_t)_{0\leq t\leq T} be a standard dd-dimensional Brownian motion on a filtered probability space satisfying the usual completeness and right-continuity conditions.

Let θt\theta_t be an Rd\mathbb R^d-valued progressively measurable process such that

∫0T∥θs∥2 ds<∞P-almost surely.\int_0^T\|\theta_s\|^2\,ds<\infty \quad P\text{-almost surely}.

Define the stochastic exponential

Zt=exp⁡(−∫0tθs⊤ dWs−12∫0t∥θs∥2 ds).Z_t= \exp\left( -\int_0^t\theta_s^\top\,dW_s -\frac12\int_0^t\|\theta_s\|^2\,ds \right).

The stochastic integral is interpreted in the sense of Itô calculus.

The process ZZ is a strictly positive local martingale. Assume additionally that it is a true martingale on [0,T][0,T], so that EP[ZT]=1E_P[Z_T]=1. Define

dQdP∣FT=ZT.\frac{dQ}{dP}\bigg|_{\mathcal F_T}=Z_T.

Then

WtQ=Wt+∫0tθs dsW_t^Q=W_t+\int_0^t\theta_s\,ds

is a standard dd-dimensional Brownian motion under QQ, relative to the same filtration.

The signs are linked: a negative stochastic integral in the exponential produces a positive drift correction in WQW^Q. An equivalent convention uses a positive stochastic integral and subtracts the drift instead.

What is preserved

The paths of WW are not replaced by different paths when the measure changes. Rather, their probabilities are reweighted. The corrected process WQW^Q has independent, normally distributed increments under QQ, with

WtQ−WsQ∼N(0,(t−s)Id).W_t^Q-W_s^Q\sim N(0,(t-s)I_d).

Its quadratic variation remains

[WQ,i,WQ,j]t=δijt.[W^{Q,i},W^{Q,j}]_t=\delta_{ij}t.

The added drift has finite variation and therefore does not alter this pathwise quantity.

Why integrability matters

The requirement that ZZ be a true martingale is essential. A nonnegative local martingale is a supermartingale, and it may satisfy

EP[ZT]<1.E_P[Z_T]<1.

In that case, ZT dPZ_T\,dP does not define a probability measure without further modification.

A widely used sufficient condition is Novikov’s condition:

EP[exp⁡(12∫0T∥θs∥2 ds)]<∞.E_P\left[ \exp\left( \frac12\int_0^T\|\theta_s\|^2\,ds \right) \right]<\infty.

This guarantees that the stochastic exponential is a martingale on the specified finite horizon. In particular, bounded θ\theta satisfies this condition.

Novikov’s condition is sufficient, not necessary. Its failure does not by itself invalidate a proposed measure change; it means that another argument is needed to establish the martingale property.

Conditions must also be matched to the time horizon. Densities may define equivalent measures on every finite-time sigma-algebra without producing equivalent measures on the sigma-algebra describing the entire infinite-time path.

Effect on stochastic differential equations

Suppose an Rn\mathbb R^n-valued process satisfies

dXt=bt dt+σt dWt,dX_t=b_t\,dt+\sigma_t\,dW_t,

where σt\sigma_t is an n×dn\times d matrix. Under the measure change above,

dWt=dWtQ−θt dt,dW_t=dW_t^Q-\theta_t\,dt,

so the same process satisfies

dXt=(bt−σtθt) dt+σt dWtQ.dX_t=(b_t-\sigma_t\theta_t)\,dt+\sigma_t\,dW_t^Q.

Thus Girsanov’s theorem changes the drift from btb_t to bt−σtθtb_t-\sigma_t\theta_t, while leaving the diffusion coefficient unchanged.

To obtain a target drift b~t\widetilde b_t, one seeks a process θt\theta_t satisfying

σtθt=bt−b~t.\sigma_t\theta_t=b_t-\widetilde b_t.

When σt\sigma_t is square and invertible,

θt=σt−1(bt−b~t).\theta_t=\sigma_t^{-1}(b_t-\widetilde b_t).

If σt\sigma_t is degenerate, not every drift change is possible through an equivalent change of measure. The required drift difference must lie in the image of σt\sigma_t, and the resulting density must still satisfy the necessary integrability conditions.

Constant-drift example

Let WW be a one-dimensional Brownian motion under PP, and let μ\mu be constant. Set

ZT=exp⁡(μWT−12μ2T),dQ=ZT dP.Z_T=\exp\left(\mu W_T-\frac12\mu^2T\right), \qquad dQ=Z_T\,dP.

Under QQ,

Bt=Wt−μtB_t=W_t-\mu t

is Brownian motion. Consequently,

Wt=Bt+μtW_t=B_t+\mu t

has drift μ\mu under QQ.

For a fixed time tt, WtW_t therefore has the normal distribution N(μt,t)N(\mu t,t) under QQ, rather than N(0,t)N(0,t). The theorem describes this change for the whole path up to TT, not merely for one marginal distribution.

General continuous-martingale form

A broader version uses a continuous local martingale LL, with L0=0L_0=0, and its stochastic exponential

E(L)t=exp⁡(Lt−12⟨L⟩t).\mathcal E(L)_t = \exp\left(L_t-\frac12\langle L\rangle_t\right).

Suppose Z=E(L)Z=\mathcal E(L) is a true martingale and defines QQ on a finite horizon. If MM is a continuous PP-local martingale, then

Mt−⟨M,L⟩tM_t-\langle M,L\rangle_t

is a continuous QQ-local martingale. Here ⟨M,L⟩\langle M,L\rangle denotes quadratic covariation for continuous local martingales.

The Brownian statement follows by taking

Lt=−∫0tθs⊤ dWs.L_t=-\int_0^t\theta_s^\top\,dW_s.

The correction becomes

Wt−⟨W,L⟩t=Wt+∫0tθs ds.W_t-\langle W,L\rangle_t = W_t+\int_0^t\theta_s\,ds.

More general versions apply to semimartingales with jumps. There, a change of measure may alter jump intensities or compensators as well as continuous drift. The description “only the drift changes” therefore applies directly to the continuous diffusion setting, not to every version of the theorem.

Proof mechanism

The Brownian theorem combines a martingale transformation with a characterization of Brownian motion.

First, the exponential density satisfies

dZt=−Ztθt⊤ dWt.dZ_t=-Z_t\theta_t^\top\,dW_t.

Applying the stochastic product rule to

Zt(Wt+∫0tθs ds)Z_t\left(W_t+\int_0^t\theta_s\,ds\right)

shows that the finite-variation terms cancel. After suitable localization, the change-of-measure identity for conditional expectation establishes that WQW^Q is a continuous local martingale under QQ.

Second, adding a continuous finite-variation process leaves quadratic covariation unchanged:

[WQ,i,WQ,j]t=δijt.[W^{Q,i},W^{Q,j}]_t=\delta_{ij}t.

Lévy’s characterization then identifies WQW^Q as Brownian motion.

This explains why the theorem requires more than computing a new mean and variance: Brownian motion is characterized by its time-dependent martingale structure and quadratic variation.

Applications

Mathematical finance

Girsanov’s theorem is a principal tool for constructing a risk-neutral measure. For example, suppose

dStSt=μt dt+σt dWt,\frac{dS_t}{S_t}=\mu_t\,dt+\sigma_t\,dW_t,

and a bank account grows at rate rtr_t. When σt≠0\sigma_t\neq0, choosing

θt=μt−rtσt\theta_t=\frac{\mu_t-r_t}{\sigma_t}

gives

dStSt=rt dt+σt dWtQ,\frac{dS_t}{S_t}=r_t\,dt+\sigma_t\,dW_t^Q,

provided the density defines an equivalent probability measure.

The discounted asset is then a local martingale under QQ. Stronger integrability assumptions may be needed for it to be a true martingale. The construction changes the probability measure used for valuation; it does not assert that the asset’s physical expected return is the bank-account rate.

Statistical inference for diffusions

For continuously observed diffusion models with a common diffusion coefficient and different drift parameters, Girsanov’s theorem can provide a pathwise likelihood function.

For example, compare

dXt=b0(t,Xt) dt+σ(t,Xt) dWtdX_t=b_0(t,X_t)\,dt+\sigma(t,X_t)\,dW_t

with a target drift b1b_1. If

b1−b0=σub_1-b_0=\sigma u

and the exponential is a valid density, the likelihood ratio on the underlying filtered space is

exp⁡(∫0Tut⊤ dWt−12∫0T∥ut∥2 dt).\exp\left( \int_0^T u_t^\top\,dW_t -\frac12\int_0^T\|u_t\|^2\,dt \right).

When this density is measurable with respect to the observed trajectory, it is also the likelihood ratio for the path observation. Such formulas support maximum likelihood estimation for drift parameters.

Simulation and importance sampling

In importance sampling, a simulation may deliberately change a diffusion’s drift to make rare events more frequent. Girsanov weights then correct expectations back to the original measure.

For a suitable path functional FF, equivalence gives

EP[F]=EQ[FdPdQ].E_P[F]=E_Q\left[F\frac{dP}{dQ}\right].

This identity underlies drift-based variance reduction in Monte Carlo methods. A valid measure change alone does not guarantee an efficient estimator: poorly chosen weights can have very large or infinite variance.

Weak solutions of stochastic differential equations

The theorem can also be used to construct a weak solution of a stochastic differential equation by beginning with a simpler process and changing measure to introduce a desired drift. This requires a valid density and does not automatically establish strong existence or pathwise uniqueness.

Historical context and scope

The theorem is named after Igor Girsanov, whose 1960 work established an influential general change-of-measure result for stochastic processes. It is closely related to the earlier Cameron–Martin theorem, which describes how Wiener measure changes under suitable deterministic shifts. Girsanov’s theorem accommodates adapted random drift changes and is naturally formulated using martingales.

Its scope is constrained by the geometry of path space. Equivalent measures preserve almost-sure path properties, including quadratic variation. Consequently, changing drift may produce equivalent diffusion laws, whereas changing a constant nonzero volatility to a different positive value produces mutually singular laws on a continuously observed path space: quadratic variation distinguishes them exactly.