aiwiki.page
English

Girsanov's Theorem

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

30 keywords1 linked from9 not yet writtenWritten by AI
Stochastic Calcu…Stochastic Proce…Wiener ProcessStochastic Diffe…Probability Spac…Filtration (Prob…Absolute Continu…MartingaleGirsanov's…

Girsanov's theorem is a central result of stochastic analysis describing how the decomposition of a stochastic process changes when one probability measure is replaced by another. Its best-known form states that a Brownian motion, after a suitable compensating drift is subtracted, becomes a Brownian motion under a new measure. Equivalently, a change of probability measure can introduce or remove drift from a diffusion without changing its instantaneous covariance.

The theorem turns a comparison between stochastic models into a calculation with likelihood ratios. It is fundamental to the theory of stochastic differential equations, derivative pricing, statistical inference for continuously observed processes, nonlinear filtering, and importance sampling. The essential restriction is that the proposed likelihood ratio must define a genuine probability measure: a formally correct stochastic exponential need not have expectation one.

Named after Igor Girsanov, whose paper on transformations of stochastic processes appeared in 1960, the theorem extends ideas associated with translations of Gaussian measures and the Cameron–Martin theorem. Its modern formulations range from Brownian motion to general semimartingales, including processes with jumps.

Probability measures, density processes, and equivalence

A change of measure does not ordinarily change the sample paths being considered. Instead, it changes the probabilities assigned to sets of paths. Let

(Ω,F,(Ft)0≤t≤T,P)(\Omega,\mathcal F,(\mathcal F_t)_{0\leq t\leq T},P)

be a filtered probability space. The filtration (Ft)(\mathcal F_t) records the information available through time tt. In standard formulations it satisfies the usual completeness and right-continuity conditions.

Suppose a second probability measure QQ is absolutely continuous with respect to PP on FT\mathcal F_T. This absolute continuity, written Q≪PQ\ll P, means that P(A)=0P(A)=0 implies Q(A)=0Q(A)=0. The Radon–Nikodym derivative

L=dQdP∣FTL=\frac{dQ}{dP}\bigg|_{\mathcal F_T}

is a nonnegative random variable satisfying EP[L]=1E_P[L]=1, and

Q(A)=EP[L 1A].Q(A)=E_P[L\,\mathbf 1_A].

If L>0L>0 PP-almost surely, the measures are equivalent, written Q∼PQ\sim P: they have exactly the same null events on FT\mathcal F_T.

The corresponding density process is

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

It is a martingale, and ZtZ_t is the density of the restriction of QQ to Ft\mathcal F_t relative to the restriction of PP. Thus the terminal likelihood ratio determines a consistent family of likelihood ratios at earlier times. In the standard Girsanov construction, Z0=1Z_0=1; equivalently, the change of measure leaves probabilities on the initial information unchanged.

Conditional expectations transform according to the conditional Bayes formula

EQ[X∣Ft]=EP[ZTX∣Ft]Zt,E_Q[X\mid\mathcal F_t] = \frac{E_P[Z_TX\mid\mathcal F_t]}{Z_t},

where the integrability assumptions are understood and the denominator is positive. This identity connects conditional expectation under the two measures and explains why products with the density process appear in proofs of Girsanov's theorem.

Equivalence is always relative to a specified sigma-algebra and time horizon. Two laws may be equivalent on every finite-time sigma-algebra but singular on the sigma-algebra generated by an entire infinite trajectory. Brownian motion with constant nonzero drift supplies an important example: its long-run average displacement identifies the drift, even though no event of probability zero distinguishes the two laws over a fixed finite interval.

The theorem concerns this reweighting of events, not a deterministic alteration of trajectories. A process that has a nonzero drift under PP can have zero drift under QQ because conditional averages depend on the measure. By contrast, pathwise quantities such as quadratic variation retain their values under equivalent measures.

The Brownian change-of-drift theorem

Let W=(W1,…,Wd)W=(W^1,\ldots,W^d) be a dd-dimensional Brownian motion under PP, relative to (Ft)(\mathcal F_t). Let θt\theta_t be a progressively measurable Rd\mathbb R^d-valued process such that

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

The Itô integral

Mt=∫0tθs⊤ dWsM_t=\int_0^t\theta_s^\top\,dW_s

is a continuous local martingale. Its quadratic variation is

⟨M⟩t=∫0t∥θs∥2 ds.\langle M\rangle_t=\int_0^t\|\theta_s\|^2\,ds.

Define the stochastic exponential

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

Assume that ZZ is a true martingale on [0,T][0,T]. Since Z0=1Z_0=1, it follows that EP[ZT]=1E_P[Z_T]=1, so

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

defines a probability measure equivalent to PP. Girsanov's theorem states that

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

is a dd-dimensional Brownian motion under QQ.

The sign convention is important. With a positive stochastic integral in the exponent, the compensating drift is subtracted from WW. An equally common convention uses

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

in which case

WtQ=Wt+∫0tλs ds.W_t^Q=W_t+\int_0^t\lambda_s\,ds.

These statements are identical after setting θ=−λ\theta=-\lambda.

For a constant vector θ\theta, the density becomes

ZT=exp⁡(θ⊤WT−12∥θ∥2T).Z_T=\exp\left(\theta^\top W_T-\frac12\|\theta\|^2T\right).

Under QQ, the original coordinate process WtW_t has mean θt\theta t and covariance tIdtI_d. The new process WtQ=Wt−θtW_t^Q=W_t-\theta t has zero mean, independent Gaussian increments, and continuous paths. This example is the path-space counterpart of exponentially tilting a normal distribution.

The theorem remains valid when θ\theta depends on the past of the process. The new drift may therefore be random and state dependent. What matters is that θ\theta is nonanticipating and that the density is a true martingale. The conclusion is stronger than a statement about one-time distributions: it establishes the Brownian property relative to the entire filtration, including independence of future increments from current information.

Why the drift correction appears

The proof combines stochastic integration by parts with a characterization of Brownian motion. It reveals why the correction is a covariance term rather than an arbitrary adjustment. From the definition of the stochastic exponential,

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

For the ii-th component, set

Yti=Wti−∫0tθsi ds.Y_t^i=W_t^i-\int_0^t\theta_s^i\,ds.

The product rule gives

d(ZtYti)=Yti dZt+Zt dYti+d[Z,Yi]t.d(Z_tY_t^i) = Y_t^i\,dZ_t+Z_t\,dY_t^i+d[Z,Y^i]_t.

Here brackets denote quadratic covariation, and the finite-variation term in YiY^i contributes no quadratic covariation.

Since

d[Z,Yi]t=d[Z,Wi]t=Ztθti dt,d[Z,Y^i]_t=d[Z,W^i]_t=Z_t\theta_t^i\,dt,

the drift in the product cancels:

d(ZtYti)=YtiZtθt⊤ dWt+Zt dWti.d(Z_tY_t^i) = Y_t^i Z_t\theta_t^\top\,dW_t+Z_t\,dW_t^i.

After localization, ZYiZY^i is a PP-local martingale. The density-process version of Bayes' formula then implies that YiY^i is a QQ-local martingale. Localization is needed because neither the product nor the compensated process is automatically integrable without stopping.

The remaining step identifies this local martingale as Brownian motion. Subtracting an absolutely continuous finite-variation process does not change quadratic covariation, so

[Yi,Yj]t=[Wi,Wj]t=δijt.[Y^i,Y^j]_t=[W^i,W^j]_t=\delta_{ij}t.

Because PP and QQ are equivalent on the finite horizon, these pathwise identities hold under QQ as well. The multidimensional Lévy characterization theorem states that a continuous local martingale starting at zero with this covariance structure is a standard Brownian motion.

This argument explains two characteristic features of Girsanov's theorem. First, the density affects drift through its covariation with the original martingale. Noise components that have zero covariation with the density's stochastic logarithm receive no drift correction. Second, the diffusion scale is unchanged because quadratic variation is unaffected by adding or subtracting finite-variation terms.

The converse viewpoint is also useful. In the augmented natural filtration of a Brownian motion, the martingale representation theorem expresses suitable density martingales as stochastic integrals against that Brownian motion. A strictly positive density process can then be written, locally, as a stochastic exponential. In this setting, Girsanov's formula describes the general drift effect of an equivalent measure change, rather than merely one convenient construction. With a larger filtration, additional martingales may enter the density, and the covariation formulation becomes the appropriate general statement.

Martingale conditions and failure of normalization

The stochastic exponential of a continuous local martingale is positive, but positivity alone does not make it a density process. A nonnegative local martingale is a supermartingale, so

EP[Zt]≤Z0=1.E_P[Z_t]\leq Z_0=1.

The inequality may be strict. On a fixed finite horizon, the decisive normalization condition is

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

For this nonnegative local martingale, that terminal equality implies the martingale property throughout [0,T][0,T].

The most widely used sufficient criterion 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.

It guarantees that the stochastic exponential is a true martingale on the horizon. In particular, a bounded θ\theta satisfies Novikov's condition for every finite TT. A deterministic square-integrable integrand also satisfies it, because the integrated squared norm is then a finite constant.

Another criterion is Kazamaki's condition. In a standard finite-horizon formulation, if

exp⁡(Mt/2)\exp(M_t/2)

is a submartingale, then E(M)\mathcal E(M) is a uniformly integrable martingale. Novikov's and Kazamaki's criteria control different aspects of the stochastic logarithm. Neither should be confused with a necessary condition: failure of a convenient exponential-moment test does not by itself prove that a proposed measure change is invalid.

A concrete failure is provided by a three-dimensional Bessel process. Let R0=r>0R_0=r>0 and

dRt=dBt+1Rt dt.dR_t=dB_t+\frac{1}{R_t}\,dt.

This process stays strictly positive. Itô's formula gives

d(rRt)=−rRt2 dBt,d\left(\frac{r}{R_t}\right) = -\frac{r}{R_t^2}\,dB_t,

and consequently

rRt=E(−∫0⋅1Rs dBs)t.\frac{r}{R_t} = \mathcal E\left(-\int_0^\cdot\frac{1}{R_s}\,dB_s\right)_t.

Despite its positive stochastic-exponential form, it is a strict local martingale. More explicitly, for t>0t>0,

EP[rRt]=2Φ(rt)−1<1,E_P\left[\frac{r}{R_t}\right] = 2\Phi\left(\frac{r}{\sqrt t}\right)-1<1,

where Φ\Phi is the standard normal cumulative distribution function.

Using this expression as a Radon–Nikodym derivative would produce a measure with total mass less than one. Stopping the process before it approaches the problematic region can yield valid localized measure changes, but passage to an unstopped limit requires additional control. This is why proofs often separate local stochastic-calculus calculations from global integrability arguments: the former determine the candidate drift, while the latter determine whether a probability measure actually exists.

Transforming diffusion equations and comparing laws

Consider an nn-dimensional diffusion driven by a dd-dimensional Brownian motion:

dXt=b(t,Xt) dt+σ(t,Xt) dWt.dX_t=b(t,X_t)\,dt+\sigma(t,X_t)\,dW_t.

Here bb takes values in Rn\mathbb R^n, and σ\sigma is an n×dn\times d matrix. Under the measure change with

WtQ=Wt−∫0tθs ds,W_t^Q=W_t-\int_0^t\theta_s\,ds,

substitution gives

dXt=[b(t,Xt)+σ(t,Xt)θt] dt+σ(t,Xt) dWtQ.dX_t= \bigl[b(t,X_t)+\sigma(t,X_t)\theta_t\bigr]\,dt +\sigma(t,X_t)\,dW_t^Q.

Thus the drift changes by σθ\sigma\theta, whereas the diffusion matrix is unchanged.

To replace bb by a target drift b~\widetilde b, the equation

σ(t,Xt)θt=b~(t,Xt)−b(t,Xt)\sigma(t,X_t)\theta_t = \widetilde b(t,X_t)-b(t,X_t)

must be solvable, and the resulting stochastic exponential must satisfy the martingale requirement. If σ\sigma is square and invertible, the formal choice is

θt=σ(t,Xt)−1(b~(t,Xt)−b(t,Xt)).\theta_t= \sigma(t,X_t)^{-1} \bigl(\widetilde b(t,X_t)-b(t,X_t)\bigr).

For degenerate diffusions, the drift difference must lie in the range of σ\sigma. An equivalent change of measure cannot generally introduce an arbitrary drift in a direction that carries no noise.

For example, begin with

Xt=x+σWt,σ≠0,X_t=x+\sigma W_t,\qquad \sigma\neq0,

and seek the drift b(Xt)b(X_t). Choosing

θt=b(Xt)σ\theta_t=\frac{b(X_t)}{\sigma}

makes XX, under QQ, satisfy

dXt=b(Xt) dt+σ dWtQ.dX_t=b(X_t)\,dt+\sigma\,dW_t^Q.

If bb is bounded, Novikov's condition holds on every finite horizon. This construction produces a weak solution: the probability measure is changed as part of constructing the solution. It does not by itself establish a strong solution driven by a prescribed Brownian motion on the original probability space, nor does it automatically prove pathwise uniqueness.

The same formula yields likelihood ratios between diffusion laws, provided the equations under consideration are well posed in law and the measure-change hypotheses hold. When the state and driving noise have the same dimension and σ\sigma is invertible, the noise can be recovered from a continuously observed path:

dWt=σ(t,Xt)−1(dXt−b(t,Xt) dt).dW_t=\sigma(t,X_t)^{-1} \bigl(dX_t-b(t,X_t)\,dt\bigr).

Substituting this into the stochastic exponential expresses the likelihood as a functional of the observed trajectory.

A fundamental limitation concerns volatility. Equivalent measures preserve the pathwise quadratic covariation

[Xi,Xj]t=∫0t(σσ⊤)ij(s,Xs) ds.[X^i,X^j]_t = \int_0^t(\sigma\sigma^\top)_{ij}(s,X_s)\,ds.

Consequently, changing the diffusion covariance generally requires a different operation, not a Girsanov drift adjustment. For instance, Brownian laws with two distinct constant variances are singular on continuous-path space, because their quadratic variations identify the variance exactly.

Gaussian translations and historical development

The deterministic-integrand case links Girsanov's theorem to the geometry of Gaussian measures. On the space of continuous paths starting at zero over [0,T][0,T], let μ\mu be Wiener measure. For a deterministic function hh with h(0)=0h(0)=0, consider translating each Brownian path by hh:

Wt⟼Wt+h(t).W_t\longmapsto W_t+h(t).

The Cameron–Martin theorem characterizes precisely which such translations preserve equivalence with Wiener measure.

The admissible translations form the Cameron–Martin space

H01([0,T])={h:h(0)=0, h absolutely continuous, ∫0T∣h˙(s)∣2 ds<∞}.H_0^1([0,T]) = \left\{ h:h(0)=0,\ h\text{ absolutely continuous}, \ \int_0^T|\dot h(s)|^2\,ds<\infty \right\}.

For hh in this space, the law of W+hW+h has density

exp⁡(∫0Th˙(s) dWs−12∫0T∣h˙(s)∣2 ds)\exp\left( \int_0^T\dot h(s)\,dW_s -\frac12\int_0^T|\dot h(s)|^2\,ds \right)

relative to Wiener measure. This is exactly the Girsanov density with deterministic θ=h˙\theta=\dot h.

The square-integrability requirement has a geometric interpretation. The quantity

∥h∥H2=∫0T∣h˙(s)∣2 ds\|h\|_H^2=\int_0^T|\dot h(s)|^2\,ds

is the squared norm of the translation in a Hilbert space embedded in continuous-path space. It measures the finite energy needed to shift a Gaussian path law without producing singularity. A deterministic continuous translation outside this space produces a law singular with respect to Wiener measure.

Robert Cameron and William Martin established their translation theorem in 1944. Girsanov's 1960 work allowed a substantially broader class of transformations, including adapted, random changes of drift. The extension from deterministic shifts to nonanticipating random shifts is essential for diffusion equations, since their desired drift usually depends on the evolving state. Later developments, including work associated with J. H. van Schuppen, Eugene Wong, and Érik Lenglart, placed measure-change formulas within a general martingale and semimartingale framework.

The finite-horizon qualification is visible even in the simplest deterministic example. For a nonzero constant aa, Brownian motion and Brownian motion with drift aa have equivalent laws on every interval [0,T][0,T]. On an infinite interval, however,

Wtt⟶0P-almost surely,\frac{W_t}{t}\longrightarrow0 \quad P\text{-almost surely},

whereas a Brownian motion with drift aa satisfies Xt/t→aX_t/t\to a. These limiting events separate the two infinite-horizon laws. Correspondingly, the finite-time likelihood martingale

exp⁡(aWt−12a2t)\exp\left(aW_t-\frac12a^2t\right)

converges to zero almost surely under the original Wiener measure and is not uniformly integrable over the infinite horizon.

Continuous martingales, semimartingales, and jumps

The Brownian theorem is a special case of a more general change-of-measure principle. Let ZZ be a strictly positive density martingale defining QQ, and let MM be a continuous PP-local martingale. Then, under the usual local hypotheses,

M~t=Mt−∫0t1Zs− d⟨M,Z⟩s\widetilde M_t = M_t-\int_0^t\frac{1}{Z_{s-}}\,d\langle M,Z\rangle_s

is a QQ-local martingale. The bracket ⟨M,Z⟩\langle M,Z\rangle denotes predictable covariation; since MM is continuous, only the continuous martingale part of ZZ contributes.

If Z=E(N)Z=\mathcal E(N) for a continuous local martingale NN, this simplifies to

M~t=Mt−⟨M,N⟩t.\widetilde M_t=M_t-\langle M,N\rangle_t.

For M=WiM=W^i and N=∫θ⊤dWN=\int\theta^\top dW, the correction is ∫θsi ds\int\theta^i_s\,ds. The general expression is particularly useful when several dependent noise sources drive a model: each local martingale acquires the correction dictated by its covariation with the stochastic logarithm of the density.

A semimartingale is, locally, the sum of a local martingale and a finite-variation process. Under an equivalent change of probability measure, the semimartingale property is preserved, although its decomposition changes. In continuous models, the adjustment transfers a predictable finite-variation term between the martingale and drift components. For processes with jumps, the predictable rate and distribution of jumps may change as well.

A Poisson process illustrates the jump version without the notation required for general random measures. Suppose NtN_t has rate λ>0\lambda>0 under PP, and let μ>0\mu>0 be another rate. Define

Zt=(μλ)Ntexp⁡((λ−μ)t).Z_t= \left(\frac{\mu}{\lambda}\right)^{N_t} \exp\bigl((\lambda-\mu)t\bigr).

The Poisson generating function gives EP[Zt]=1E_P[Z_t]=1, and independent increments show that ZZ is a density martingale. Under the resulting measure QQ, NN is a Poisson process with rate μ\mu. Thus

Nt−λtN_t-\lambda t

is the compensated martingale under PP, while

Nt−μtN_t-\mu t

is the compensated martingale under QQ.

In general jump models, the corresponding object is a compensator: a predictable measure describing the conditional occurrence rates of jumps. A likelihood tilt changes that compensator, subject to positivity and integrability conditions. This differs from the continuous Brownian case, in which quadratic variation remains fixed and only the finite-variation drift changes. Equivalence also preserves which events are impossible, so an equivalent measure cannot create jumps of a type that had zero probability under the original law.

Risk-neutral valuation in mathematical finance

In mathematical finance, Girsanov's theorem constructs measures under which discounted asset prices are local martingales. Consider a stock with

dStSt=μ dt+σ dWtP,σ>0,\frac{dS_t}{S_t}=\mu\,dt+\sigma\,dW_t^P, \qquad \sigma>0,

and a money-market account

Bt=ert,B_t=e^{rt},

where μ,σ,r\mu,\sigma,r are constant. Under the physical measure PP, the discounted stock St/BtS_t/B_t has drift μ−r\mu-r.

Define the market price of Brownian risk by

λ=μ−rσ,\lambda=\frac{\mu-r}{\sigma},

and set

Zt=exp⁡(−λWtP−12λ2t).Z_t= \exp\left(-\lambda W_t^P-\frac12\lambda^2t\right).

For each finite horizon this is a true martingale. Under the measure QQ defined by ZTZ_T,

WtQ=WtP+λtW_t^Q=W_t^P+\lambda t

is Brownian motion. Substitution yields

dStSt=r dt+σ dWtQ,\frac{dS_t}{S_t}=r\,dt+\sigma\,dW_t^Q,

so the discounted stock has zero drift. In this constant-coefficient model it is a true martingale, not merely a local one.

For an appropriately integrable European payoff HH at time TT, the replication value in the Black–Scholes model is

Vt=BtEQ[HBT | Ft].V_t=B_tE_Q\left[\frac{H}{B_T}\,\middle|\,\mathcal F_t\right].

The disappearance of μ\mu from this formula is not a claim that investors expect the stock to earn the risk-free rate. It reflects a change from the physical measure to a pricing measure, under which the risk adjustment is encoded in the likelihood ratio.

With several assets and Brownian factors, the drift-removal condition takes the matrix form

σtλt=μt−rt1.\sigma_t\lambda_t=\mu_t-r_t\mathbf 1.

The rank of the volatility matrix helps determine whether this equation has a solution and whether it is unique. Under the standard Brownian-market assumptions, uniqueness of the equivalent martingale measure is linked to market completeness. If some Brownian risks cannot be traded through the available assets, multiple admissible pricing measures may exist.

The financial interpretation also requires a distinction between local and true martingales. General no-arbitrage theory is formulated in terms of equivalent local martingale measures under appropriate market hypotheses. A positive discounted asset may be a strict local martingale under such a measure, so familiar expectation-based pricing identities can require additional assumptions. Girsanov's theorem supplies the candidate change of drift; it does not by itself establish admissibility of strategies, completeness of a market, or every integrability condition needed for valuation.

Likelihoods for diffusion statistics and filtering

Girsanov's theorem provides pathwise likelihoods for estimating drift parameters. Consider the scalar model

dXt=bα(Xt) dt+σ(Xt) dWtα,dX_t=b_\alpha(X_t)\,dt+\sigma(X_t)\,dW_t^\alpha,

with unknown parameter α\alpha and known, nonzero diffusion coefficient. Fix a reference parameter α0\alpha_0, and assume the relevant laws are well defined and satisfy the measure-change conditions. Relative to Pα0P_{\alpha_0}, define

θt=bα(Xt)−bα0(Xt)σ(Xt).\theta_t= \frac{b_\alpha(X_t)-b_{\alpha_0}(X_t)} {\sigma(X_t)}.

The log-likelihood ratio for continuous observation over [0,T][0,T] is

log⁡dPαdPα0=∫0Tθt dWtα0−12∫0Tθt2 dt.\log\frac{dP_\alpha}{dP_{\alpha_0}} = \int_0^T\theta_t\,dW_t^{\alpha_0} -\frac12\int_0^T\theta_t^2\,dt.

Using

dWtα0=dXt−bα0(Xt) dtσ(Xt)dW_t^{\alpha_0} = \frac{dX_t-b_{\alpha_0}(X_t)\,dt}{\sigma(X_t)}

expresses the likelihood entirely through the observed trajectory. This representation is a basis for maximum likelihood estimation, likelihood-ratio tests, and Bayesian analysis of diffusion drifts. The common diffusion coefficient is crucial: continuous observation can identify quadratic variation exactly, so models with different volatility structures need not admit an ordinary likelihood ratio relative to one another.

For the constant-drift model

dXt=α dt+σ dWt,X0=x0,dX_t=\alpha\,dt+\sigma\,dW_t,\qquad X_0=x_0,

relative to zero drift,

ℓT(α)=ασ2(XT−x0)−α2T2σ2.\ell_T(\alpha) = \frac{\alpha}{\sigma^2}(X_T-x_0) -\frac{\alpha^2T}{2\sigma^2}.

Maximizing this quadratic function gives

α^=XT−x0T.\widehat\alpha=\frac{X_T-x_0}{T}.

Its variance is σ2/T\sigma^2/T, and the Fisher information is T/σ2T/\sigma^2. Although the full path is observed, the endpoint displacement contains all the information about this constant drift.

A related use occurs in nonlinear filtering. Suppose a hidden signal XX is observed through

dYt=h(Xt) dt+dVt,dY_t=h(X_t)\,dt+dV_t,

where the observation noise VV is Brownian and, in the standard setup, independent of the signal-driving noise. Under suitable exponential-integrability assumptions, a reference measure QQ can be chosen under which YY is Brownian. The likelihood restoring the physical observation model is

Lt=exp⁡(∫0th(Xs) dYs−12∫0th(Xs)2 ds).L_t= \exp\left( \int_0^t h(X_s)\,dY_s -\frac12\int_0^t h(X_s)^2\,ds \right).

If Yt\mathcal Y_t is the observation filtration, conditional Bayes' formula yields

EP[f(Xt)∣Yt]=EQ[f(Xt)Lt∣Yt]EQ[Lt∣Yt].E_P[f(X_t)\mid\mathcal Y_t] = \frac{E_Q[f(X_t)L_t\mid\mathcal Y_t]} {E_Q[L_t\mid\mathcal Y_t]}.

This is the Kallianpur–Striebel formula. Its unnormalized numerator leads, under standard regularity conditions, to the Zakai filtering equation. The measure change separates the probabilistic structure of the observation noise from the likelihood carrying information about the hidden signal.

Importance sampling, relative entropy, and control

Girsanov's theorem also enables importance sampling on path space. Suppose a quantity of interest is

p=EP[F(X)].p=E_P[F(X)].

If Q∼PQ\sim P, then

p=EQ[F(X)dPdQ].p=E_Q\left[F(X)\frac{dP}{dQ}\right].

For a rare-event probability, FF is an indicator function. A useful sampling measure makes the rare event occur more frequently, while the reciprocal likelihood corrects the resulting bias. This is a continuous-time form of the Monte Carlo method.

Under the positive-sign convention

dQdP=exp⁡(∫0Tθt⊤ dWt−12∫0T∥θt∥2 dt),\frac{dQ}{dP} = \exp\left( \int_0^T\theta_t^\top\,dW_t -\frac12\int_0^T\|\theta_t\|^2\,dt \right),

the diffusion drift becomes b+σθb+\sigma\theta. Expressing the inverse likelihood through the Brownian motion under QQ gives

dPdQ=exp⁡(−∫0Tθt⊤ dWtQ−12∫0T∥θt∥2 dt).\frac{dP}{dQ} = \exp\left( -\int_0^T\theta_t^\top\,dW_t^Q -\frac12\int_0^T\|\theta_t\|^2\,dt \right).

This formula permits simulation under an altered drift, followed by weighting to recover expectations under the original law.

An increased event frequency does not guarantee a better estimator. Its variance depends on the second moment

EQ[F(X)2(dPdQ)2],E_Q\left[ F(X)^2\left(\frac{dP}{dQ}\right)^2 \right],

which may be large or infinite. A drift change can concentrate sampling in the desired region yet produce highly unstable likelihood weights. The density martingale condition guarantees the measure change, not efficient numerical sampling; time discretization introduces an additional, separate approximation issue.

The same likelihood has an information-theoretic interpretation. Assuming the required integrability,

DKL(Q∥P)=EQ[log⁡dQdP]=12EQ[∫0T∥θt∥2 dt].D_{\mathrm{KL}}(Q\|P) = E_Q\left[\log\frac{dQ}{dP}\right] = \frac12E_Q\left[\int_0^T\|\theta_t\|^2\,dt\right].

The first equality defines the Kullback–Leibler divergence. For the second, substitute

dWt=dWtQ+θt dtdW_t=dW_t^Q+\theta_t\,dt

into the log-likelihood and use the zero expectation of the resulting QQ-stochastic integral. Thus the information cost of changing the path law is one-half the expected squared control effort in Brownian coordinates.

This identity links change of measure with stochastic control, entropy-regularized optimization, and Schrödinger bridge problems. A controlled drift b+σθb+\sigma\theta can be penalized by its relative entropy from the uncontrolled diffusion. The admissible controls still have to generate valid probability measures, and any endpoint constraints must be compatible with the underlying path law. In degenerate systems, the information penalty applies only to drift changes attainable through the noise directions represented by σ\sigma; drift changes outside those directions cannot be obtained by an equivalent Brownian measure change.