Martingale-Compensator Structure

• Martingale-compensator structure is the canonical decomposition of stochastic processes into a local martingale and a predictable finite-variation compensator.
• It underpins modern methods in filtering, optimal stopping, and risk modeling by isolating random fluctuations from deterministic trends.
• Explicit constructions via dyadic approximations facilitate extensions to non-monotone information settings and multidimensional process analysis.

A martingale-compensator structure refers to the canonical decomposition of stochastic processes—particularly submartingales and semimartingales—into a sum of a local martingale and a finite-variation, predictable process called the compensator. This structure underpins both the analysis of stochastic integrals and the description of jump and point-process dynamics, including modern applications in stochastic calculus, filtering theory, optimal stopping, risk modeling, and stochastic thermodynamics.

1. Fundamental Definitions and General Framework

Let $(\Omega, \mathcal{F}, (\mathcal{F}_t)_{t\ge0}, P)$ be a filtered probability space satisfying the usual right-continuity and completeness assumptions. For an adapted, càdlàg process $A \in \mathcal{V}_{\rm loc}$ (locally integrable finite variation), its compensator (dual predictable projection) $A^p$ is the unique predictable process of finite variation, null at zero, such that $M := A - A^p$ is a local martingale. This definition extends to semimartingale $X$: any such $X$ admits a unique decomposition $X = M + A$ where $M$ is a local martingale and $A$ is a predictable, locally-integrable process of finite variation, i.e., the compensator of the finite-variation part of $X$ (Sokol, 2012).

The Doob–Meyer decomposition theorem formally establishes this: for any integrable submartingale or adapted finite-variation process, there exists a unique predictable, increasing compensator such that the difference is a local martingale (Janson et al., 2010).

2. Structural Results and Absolute Continuity

The absolute continuity properties of compensators are crucial in identifying intensity processes, controlling pathwise representations, and applying stochastic calculus. In the minimal filtration generated by a semimartingale strong Markov process that solves an SDE driven by Wiener and Poisson random measures, every totally inaccessible stopping time admits an absolutely continuous compensator: $A \in \mathcal{V}_{\rm loc}$0 for some nonnegative predictable $A \in \mathcal{V}_{\rm loc}$1 (“hazard rate”) (Janson et al., 2010).

A central result is that all such compensators are absolutely continuous if and only if the Lévy system additive functional is absolutely continuous with respect to Lebesgue measure; this is stable under space-time transformations (e.g., continuous time-changes) and under certain types of measure changes. However, in filtration enlargements (such as initial or progressive expansions), the absolute continuity property for new stopping times may fail unless the compensator in the original filtration is itself absolutely continuous (Janson et al., 2010).

In point-process theory, the compensator structure uniquely encodes the process's law in the one-dimensional setting and generalizes to multidimensional or set-indexed processes via the notion of *–compensators, under suitable conditional independence axioms (Ivanoff, 2013).

3. Explicit Construction and Characterization

The existence of the compensator can be constructively demonstrated by dyadic approximation and a functional-analytic subsequence principle. Specifically, for any finite-variation, càdlàg, adapted process $A \in \mathcal{V}_{\rm loc}$2, one forms stepwise approximations $A \in \mathcal{V}_{\rm loc}$3 and computes discrete compensators $A \in \mathcal{V}_{\rm loc}$4, then constructs convex combinations that converge in $A \in \mathcal{V}_{\rm loc}$5 to a limiting process yielding the desired compensator (Sokol, 2012). This approach provides constructive insight, distinct from the Doob–Meyer abstract argument.

The quadratic variation process $A \in \mathcal{V}_{\rm loc}$6 of a local martingale is similarly constructed as the compensator of $A \in \mathcal{V}_{\rm loc}$7, and an explicit formula via discrete sums converges to $A \in \mathcal{V}_{\rm loc}$8 (Sokol, 2012).

In the context of semimartingales with independent increments—including Lévy processes and their stochastic exponentials—both additive and multiplicative (in the sense of stochastic exponentials) compensators can be explicitly determined from the process's characteristics, and give rise to generalized Lévy–Khintchin formulas for transforms and measure changes (Černý et al., 2020).

4. Applications: Filtering, Optimal Stopping, Nonlinear PDEs

The martingale-compensator structure is essential in modern stochastic filtering. In incomplete information models such as defaultable asset pricing—where the default is the first hitting time of a hidden process—the compensator for the default indicator $A \in \mathcal{V}_{\rm loc}$9 in the observed filtration can be explicitly identified, with the intensity process expressed as a ratio of conditional expectations. Filtering theory (e.g., the Kushner–Stratonovich or Zakai SPDE) provides explicit expressions for these conditional laws, leading to effective computation of compensators for filtering-based asset pricing and risk management (Çetin, 2012).

In optimal stopping and control, the Snell envelope's Doob–Meyer decomposition reveals that the finite-variation part of the envelope is absolutely continuous with respect to the negative variation of the finite-variation part of the gains process; in Markovian settings, this leads to explicit membership criteria for the value function in the extended generator's domain, duality principles, and smooth-fit conditions in classical boundary problems (Jacka et al., 2017).

For path-dependent processes and associated PDEs, evolutionary semigroup frameworks identify martingale–compensator decompositions analytically: a process of the form $A^p$0 is a martingale if and only if a time-shifted version satisfies a mild final-value problem involving a path-dependent generator, with the compensator $A^p$1 corresponding to an $A^p$2–derivative plus the generator action. This extends the classical link between Kolmogorov’s backward equation and Feynman–Kac martingale representations to general path-dependent and non-Markovian stopping problems (Denk et al., 2 Jul 2025).

5. Generalizations: Non-Monotone Information, Infinitesimal Compensators, and Multidimensional Settings

In non-monotone information settings—where information may be lost as well as gained, e.g., in marked point processes with deletion—classical compensator theory is replaced by an “infinitesimal” approach. Here, forward and backward (in the sense of time partition) compensators—respectively quantifying conditional innovation and information loss—are central to martingale representations. The overall structure involves the sum of martingale increments for both innovation and loss, generalizing the Doob–Meyer theory to dynamic information environments (Christiansen, 2018).

For planar or multidimensional counting processes, the *–compensator structure incorporates set-indexed versions of avoidance probability and cumulative hazard, with laws completely characterized in the presence of a strong conditional independence property. The regenerative decomposition for the compensator is analogous to the one-dimensional hazard-based formula, but now indexed by random “layers” defined by the process history (Ivanoff, 2013).

6. Central Limit and Weak Limit Theorems: Compensator-Driven Limit Processes

The structure of compensators underlies functional limit theorems for martingales and stochastic processes. Under continuity of the compensator in the limit, weak convergence of sequences of martingales yields limiting processes of the form $A^p$3, where $A^p$4 is a Brownian motion independent of the compensator process $A^p$5. Applications include central limit theorems for occupation time processes, realized volatility estimation, and other functionals driven by compensator behavior (Rémillard et al., 2023).

In Markov chain summations, solutions to the Poisson equation provide explicit martingale-compensator decompositions, enabling analysis of scaling limits and variance parameters in classical probabilistic limit theorems (Glynn et al., 2022).

---

References

• (Janson et al., 2010) Absolutely Continuous Compensators
• (Sokol, 2012) Proving existence results in martingale theory using a subsequence principle
• (Černý et al., 2020) Simplified calculus for semimartingales: Multiplicative compensators and changes of measure
• (Çetin, 2012) On absolutely continuous compensators and nonlinear filtering equations in default risk models
• (Jacka et al., 2017) On the compensator in the Doob-Meyer decomposition of the Snell envelope
• (Rémillard et al., 2023) Central limit theorems for martingales-I : continuous limits
• (Ivanoff, 2013) A Compensator Characterization of Planar Point Processes
• (Christiansen, 2018) A martingale concept for non-monotone information in a jump process framework
• (Denk et al., 2 Jul 2025) Martingales and Path-Dependent PDEs via Evolutionary Semigroups
• (Glynn et al., 2022) Solving Poisson's Equation: Existence, Uniqueness, Martingale Structure, and CLT