Girsanov Transform Martingales

• Girsanov transform martingales are exponential densities that enable an absolutely continuous change of measure for stochastic differential equations and Markov processes.
• They establish a link between generator structures and nonlinear PDEs, elucidating conditions for path-independence and applications in stochastic control and statistical mechanics.
• Ensuring the true martingale property involves precise integrability and boundary criteria, with extensions to infinite-dimensional, sublinear, and manifold frameworks.

A Girsanov transform martingale is a fundamental object in stochastic analysis, representing the exponential martingale density that facilitates an absolutely continuous change of measure between solutions of stochastic differential equations (SDEs), and more generally between Markov processes or martingale problems. The deep connection between transform martingales, generator structure, and (non-linear) partial differential equations is central to the theory, with critical implications for the structure of conditioned processes, the path-dependence (or independence) of likelihood ratios, and the analytic criteria for such measures to be well-defined.

1. Girsanov Transform Martingales: Foundations and Classical Formulation

Given a $d$-dimensional Itô SDE of the form

$dX_t = b(t, X_t) dt + \sigma(t, X_t) dB_t,$

with smooth and invertible $\sigma$, the classical Girsanov theorem produces the exponential local martingale

$$M_t \equiv \exp\left\{ \int_0^t \langle \sigma^{-1}b(s, X_s), dB_s \rangle + \frac{1}{2}\int_0^t |\sigma^{-1}b|^2(s, X_s) ds \right\},$$

which—under Novikov-type integrability conditions—is a true martingale and serves as the Radon–Nikodym density for a new measure $Q$ on $\mathcal{F}_t$: $dQ|_{\mathcal{F}_t} = M_t dP|_{\mathcal{F}_t}.$ This change of measure transforms $X$ into a process with drift $b$, possibly modifying the probabilistic structure such as the martingale property or generator of $X$.

2. Path-Independence and Martingale–PDE Correspondence

A central question is when the Girsanov density is "path-independent"—that is, when it can be written as

$M_t = \exp \{ v(0, X_0) - v(t, X_t) \}$

for a deterministic function $v (t, x)$. Truman, Wang, Wu, and Yang show that this is equivalent to the drift being of gradient type,

$$b(t, x) = \sigma(t, x)\sigma(t, x)^\top \nabla_x v(t, x),$$

where $v$ satisfies the nonlinear, time-reversed KPZ (Burgers–KPZ type) PDE: $$\partial_t v(t, x) = -\frac{1}{2} \Big\{ \operatorname{Tr}[\sigma \sigma^\top \nabla_x^2 v] (t, x) + \|\sigma^\top \nabla_x v(t, x)\|^2 \Big\}.$$ This characterizes exactly when the likelihood ratio between two measures depends only on the endpoints in state space and time, not on the sample path traversed—leading to significant simplifications and deeper connections to stochastic control, statistical mechanics, and integrable systems (Truman et al., 2010).

Table: Characterization of Path-Independence

Criterion	Condition on $b$	PDE for $v$
Path-independence holds	$b = \sigma \sigma^\top \nabla v$	$$\partial_t v = -\frac{1}{2}(\operatorname{Tr} + \|\cdot\|^2)$$

The assertion extends to $X_t$ evolving on a connected complete manifold, exchanging Euclidean gradient and Laplacian for their Riemannian analogues.

3. Martingale Criteria and Uniqueness

The validity of a Girsanov change hinges on the density being a true martingale, not just a local martingale. In one-dimensional diffusions

$dX_t = b(X_t)dt + \sigma(X_t)dW_t,$

if a new drift $\mu$ is targeted, the density

$$Z_t = \exp\left( \int_0^t \theta(X_s)dW_s - \frac{1}{2} \int_0^t \theta^2(X_s)ds \right), \quad \theta(x) = \frac{\mu(x) - b(x)}{ \sigma(x) }$$

is a martingale if and only if the associated boundary tests fail to trigger explosion: the integral criteria, formulated in scale-speed language, furnish necessary and sufficient conditions on the endpoints (Desmettre et al., 2019). Uniform integrability or $L^1$-martingale property can be tested by boundary integrals in the tilted scale and speed measures:

• For every finite endpoint $\xi$, if $X$ can reach $\xi$ in finite time and an associated additive functional explodes, $Z$ fails to be a martingale.

Extensions to Markov jump processes admit similar exponential martingale densities, with explicit formulas for the Radon–Nikodym derivative in terms of the compensated jump measure and predictable intensities. Uniqueness of the (generalized) martingale problem identifies the density as the only possible candidate—cf.\ the quadratic family for finite-state continuous-time Markov processes (Wang, 7 Sep 2025).

4. Infinite-Dimensional and SPDE Girsanov Transform Martingales

In infinite-dimensional settings such as SPDEs on Hilbert space $H$,

$dZ = (AZ + b(Z)) dt + dW(t), \qquad Z_0 = x,$

the law of the solution admits a Girsanov-type density on path space. The "mild" Girsanov formula (Prato et al., 2023) expresses the Radon–Nikodym derivative between the laws of the linear and the nonlinear equation as

$$\frac{d\mathbb{P} \circ Z_x^{-1} }{d\mathcal{N}_{Q_T} } ( h + e^{\cdot A} x ) = \exp \left\{ - \frac{1}{2} \| \gamma_x(h) \|_{H_{Q_T}}^2 + I(\gamma_x)(h) \right\},$$

where $\mathcal{N}_{Q_T}$ is the Gaussian law of the stochastic convolution. The proof utilizes the Ramer change-of-variables formula; existence of such Radon–Nikodym derivatives hinges on quasi-nilpotency and Malliavin calculus. This density transforms the Gaussian measure into the law of the nonlinear process, with the exponential martingale $M_t$ serving as the density.

5. Structural and Analytical Conditions for True Exponential Martingales

The exponential martingale property (true martingale vs.\ merely local martingale) is subtle and depends on fine integrability and boundary behaviour. For one-dimensional diffusions (Desmettre et al., 2019):

• Explicit scale–speed–integral tests determine martingale property.
• Violation yields strict local martingales (e.g., in certain generalized Heston models the risk-neutral measure fails to exist if the Feller condition is violated).

In the Markov process context (Chen, 2011), criteria for uniform integrability of $M_t$ can be easier to check than Novikov's condition; for continuous MAFs, finiteness of expected total variation suffices, and in pure-jump processes, the jump-size lower bound and quadratic moment integrability suffice. The Girsanov transform is also tightly connected with enlargement properties for Kato classes and Dirichlet forms following measure change.

Furthermore, under a finite entropy condition (Léonard, 2011), all classical results about Girsanov transform martingales remain valid: if $P$ is absolutely continuous with respect to $R$ and $H(P\mid R)<\infty$, the Radon–Nikodym process is a true martingale and all process characteristics are transformed as per Riesz representation in Hilbert or Orlicz spaces.

6. Generalizations: Martingales in Sublinear Expectation and Probability-Free Frameworks

Girsanov transform martingales generalize to non-classical probabilistic settings:

• Sublinear expectation/G-Brownian motion: The exponentials constructed as $$M_t = \exp( \int_0^t \langle \theta_s, dB_s \rangle - \frac{1}{2} \int_0^t \langle \theta_s, d \langle B \rangle_s \theta_s \rangle )$$ are symmetric G-martingales, and the Girsanov transform reconstructs a G-Brownian motion with shifted drift (Osuka, 2011, Hu et al., 2012, Liu, 2018).
• Probability-free martingale theory: In game-theoretic capital process models, the Girsanov transform is interpreted as a change of numéraire, and the resulting process $M^I_t := M_t - \int_0^t d[I,M]_s/I_s$ is a martingale under the new numéraire, matching the classical structure absent a probability measure (Vovk et al., 2017).

7. Extensions: Manifold SDEs, Volterra Processes, and Banach-Valued Measures

The structure of Girsanov transform martingales extends to SDEs on manifolds (with drift given by Riemannian gradients and corresponding Burgers–KPZ PDEs), to stochastic processes driven by Volterra kernels (fundamental martingales for Gaussian Volterra/fBM processes, with clique-factorized exponential martingales governing graph-indexed collections (Hu et al., 2024)), and even to settings with Banach-valued measures (vector-valued Birkhoff integrals (Candeloro et al., 2019)).

These advances unify martingale transforms, measure-change density processes, and underlying geometric or functional structures. Transform martingales are intimately connected to the analytic solution properties of associated PDEs, the smoothness and invertibility of noise coefficients, and the fine stochastic calculus of the underlying random processes.

---

References:

• "A Burgers-KPZ Type Parabolic Equation for the Path-Independence of the Density of the Girsanov Transformation" (Truman et al., 2010)
• "Change of drift in one-dimensional diffusions" (Desmettre et al., 2019)
• "A mild Girsanov formula" (Prato et al., 2023)
• "Uniform integrability of exponential martingales and spectral bounds..." (Chen, 2011)
• "Girsanov theory under a finite entropy condition" (Léonard, 2011)
• "Girsanov's formula for G-Brownian motion" (Osuka, 2011)
• "Towards a probability-free theory of continuous martingales" (Vovk et al., 2017)
• "The fundamental martingale with applications to Markov Random Fields" (Hu et al., 2024)
• "A Girsanov Result through Birkhoff Integral" (Candeloro et al., 2019)
• "Martingale Problem and Quadratic Family" (Wang, 7 Sep 2025)