Change of Trajectory Measure Lemma

• The Change of Trajectory Measure Lemma is a foundational result that provides explicit Radon–Nikodym derivatives for transforming trajectory measures under time shifts.
• It leverages temporal homogeneity and differentiability in the stochastic process to connect shifted paths with original distributions via Jacobian factors and density ratios.
• This lemma underpins applications in stochastic calculus, large deviations, and statistical mechanics, ensuring rigorous measure transformations in complex systems.

The Change of Trajectory Measure Lemma is a foundational result in probability theory and stochastic analysis, describing how the probability law of a trajectory-valued process transforms under nonlinear changes of time, measure, or sampling regime. It unifies a family of change-of-measure formulas—ranging from Girsanov theorems for diffusions, to exponential tiltings for Markov processes, to geometric measure decompositions in billiard dynamics—by providing explicit Radon–Nikodym derivatives between path-space measures associated to transformed or shifted trajectories. The lemma supports advanced developments in stochastic calculus, statistical mechanics, kinetic theory, and large deviations.

1. Foundational Framework and General Setup

The prototypical context is a stochastic process $X = W + A$, where $W = (W_s)_{s \in \mathbb{R}}$ is a two-sided Brownian motion in $\mathbb{R}^d$, initiated from a random point $W_0$ with density $m(x) > 0$, and $A$ a measurable functional of $W$, typically with $A_0 = 0$ and paths in the Skorokhod space. The associated law $P_\nu$ for $X$ captures both the random initial condition and the pathwise stochasticity.

Critical structure is provided by temporal homogeneity: for all $v \in \mathbb{R}$, shifting $W$ and its functional by $v$ relates the shifted and original process via

$$X_s\bigl(W_{\cdot + v} + A_v(W)\bigr) = X_{s + v}(W),$$

ensuring the process transforms in a manner compatible with time shifts (Löbus, 2013).

2. Precise Statement of the Lemma

Under conditions of differentiability and invertibility for the map $w_0 \mapsto X_{-t}(w_0, W_{\cdot \neq 0})$, the lemma establishes that $P_\nu \circ \{X_\cdot \mapsto X_{\cdot - t}\}^{-1}$ is absolutely continuous relative to $P_\nu$, with explicit Radon–Nikodym derivative: $$\frac{P_\nu\Bigl(dX_{\cdot - t}\Bigr)}{P_\nu\Bigl(dX_\cdot\Bigr)} = \frac{m(X_{-t})}{m(X_0)} \prod_{i=1}^d \left|\frac{\partial X_{-t}^{(i)}}{\partial w_0^{(i)}}\right| .$$ Here, the Jacobian factor reflects the transformation of the initial data, and the ratio of densities compensates for the initial law shift. The conditions required include: almost sure $C^1$-diffeomorphic dependence on $w_0$, nonvanishing determinant, and integrability of the candidate density (Löbus, 2013).

3. Core Mechanisms and Proof Strategy

The proof decomposes as follows:

• Path-space decomposition: Temporal homogeneity allows the shifted trajectory measure to be written in terms of the process applied to a shifted Brownian base path ($X(W^{[-t]})$), exploiting the structure of the stochastic flow.
• Change of variable and density ratio: The Jacobian comes from the transformation $w_0' = w_0 - A_{-t}(W)$, realized as a $C^1$ diffeomorphism. The density picks up contributions from both the initial measure $m$ and the Jacobian determinant.
• Jump process adaptation: When the functional $A$ incorporates jumps, the structure still holds provided jumps are well behaved (finitely many per compacts), and their contributions are absorbed in the Jacobian.
• Synthesis: Combining these ingredients delivers the explicit Radon–Nikodym derivative (Löbus, 2013).

4. Extension to Markov Processes, Diffusions, and Large Deviations

For continuous-time Markov processes (including both jump and diffusion cases), the lemma generalizes to exponential tiltings of path-space measures. Given original and tilted generators ($\mathcal{Q}$, $\mathcal{Q}^f$), the path measure under the tilted generator is absolutely continuous with respect to the original. The Radon–Nikodym derivative is: $$\frac{d\mathbb{P}^f_x}{d\mathbb{P}_x}(X_{[0,T]}) = \exp\biggl\{ f(X_T) - f(X_0) - \int_0^T H\bigl(\delta_{X_t}, f\bigr)\,dt \biggr\},$$ where $H(\mu, f)$ is the non-linear generator (Hamiltonian) and $f$ the tilting function. For empirical measures, the Lagrangian arises as the Legendre dual of the Hamiltonian and encodes the rate function for the associated large deviation principle. The infinitesimal Radon–Nikodym density is always interpretable as a relative entropy rate per unit time (Redig et al., 2013).

5. Connections to Billiards, Geometry, and Deterministic Systems

In deterministic geometric settings, the change of trajectory measure lemma appears as the generalized Santalo formula, which directly relates integration over trajectory space to boundary data. In billiard settings, for an obstacle $K$ in a Riemannian manifold, the formula states

$$\int_{S\Omega \setminus \text{Trap}(\Omega)} f(x)\,d\mu(x) = \int_{S^+(\partial M_0) \setminus \text{Trap}^+(\partial M_0)} \left( \int_0^{\tau(x)} f(\varphi_t(x))\,dt \right) d\mu_\partial(x).$$

This allows measures over phase space to be computed from boundary data and trajectory times, underpinning results in inverse scattering and stability of trapping sets (Stoyanov, 2016).

6. Applications: Stochastic Calculus, Statistical Mechanics, and Beyond

The lemma is essential in a wide range of applications:

• Stochastic calculus: Enables partial integration relative to generators on path space, crucial for rigorous anticipative stochastic calculus (e.g., involving jumps or nonadapted processes).
• Relative compactness: Uniform control of the Radon–Nikodym densities yields relative compactness of path-space laws, facilitating weak convergence arguments (Löbus, 2013).
• Statistical mechanics: Underpins the many-to-one lemma and spinal decompositions in branching random walks, where sampling according to a tilted Gibbs measure translates to a change-of-measure along a distinguished trajectory—e.g., proving convergence toward Brownian excursion limits (Chen et al., 2015).
• Geometrical analysis: Delivers measure-preserving formulas in kinetic theory and enables resolution of weak Harnack and Sobolev estimates via precise Jacobian control for trajectory maps (Dietert et al., 20 Aug 2025).
• Mathematical finance: Corrects and extends multidimensional Girsanov-Girsanov schemas, ensuring measure changes in models with correlation or hard-to-borrow constraints (Liu, 2020).

7. Structural and Technical Prerequisites

Successful application of the Change of Trajectory Measure Lemma requires:

• Differentiability and invertibility of the trajectory map with respect to initial data or control parameters.
• Temporal homogeneity or other invariance properties to facilitate path-space transformations.
• Integrability or tightness conditions ensuring that candidate densities are martingales, especially important in contexts where Novikov-type integrability may fail—here, tightness criteria provide necessary and sufficient conditions (Blanchet et al., 2012).
• Adaptation to jump processes and hybrid processes via careful handling of discontinuities in the underlying functional.

The lemma thus serves as a unifying analytical and probabilistic tool, translating between measures on path space under various transformations, and providing the explicit Radon–Nikodym factors necessary for rigorous analysis in stochastic calculus, large deviations, statistical mechanics, geometrical dynamics, and applied probability.