Doob–Meyer Decomposition in Stochastic Processes

• Doob–Meyer decomposition is the canonical representation of a class D submartingale as the sum of a martingale and a predictable, integrable increasing process.
• The method leverages discrete decompositions, the Komlós lemma, and limit procedures to rigorously establish the uniqueness and predictability of the compensator.
• Extensions to nonlinear, fractional, and multiplicative frameworks highlight its pivotal role in optimal stopping, stochastic control, and mathematical finance.

The Doob–Meyer decomposition is the canonical representation of a class D submartingale as the sum of a martingale and a predictable, integrable, increasing process. This decomposition is both unique and foundational in the theory of stochastic processes, underpinning much of the modern development of martingale theory, stochastic integration, stochastic control, and optimal stopping. Variants and extensions exist in nonlinear, sublinear, and noncommutative frameworks, as well as for specific classes of processes relevant in mathematical finance, probability, and analysis.

1. Formal Statement and Uniqueness

Let $$(\Omega, \mathcal{F}, \{\mathcal{F}_t\}_{t\ge0}, P)$$ be a filtered probability space with right-continuous and complete filtration. A real-valued, adapted process $S=(S_t)_{t\in[0,1]}$ with càdlàg paths is a submartingale of class D if $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$ is uniformly integrable and for all $0\leq s\leq t\leq1$, $E[|S_t|]<\infty$, $E[S_t|\mathcal{F}_s]\ge S_s$ a.s.

Doob–Meyer Theorem: For every such $S$, there exist unique processes $(M, A)$ such that:

• $M$ is a càdlàg martingale with $M_0 = S_0$,
• $S=(S_t)_{t\in[0,1]}$0 is integrable, predictable, increasing, $S=(S_t)_{t\in[0,1]}$1,
• $S=(S_t)_{t\in[0,1]}$2 for every $S=(S_t)_{t\in[0,1]}$3 (Beiglboeck et al., 2010).

Uniqueness holds in the sense that if $S=(S_t)_{t\in[0,1]}$4, then $S=(S_t)_{t\in[0,1]}$5 (as predictable, increasing processes) and $S=(S_t)_{t\in[0,1]}$6.

2. Key Ingredients and Proof Structure

The standard proof leverages several elements:

• Discrete Doob decomposition: For each fixed grid, the process is represented as $S=(S_t)_{t\in[0,1]}$7 with explicit formulas for the compensator $S=(S_t)_{t\in[0,1]}$8 given by $S=(S_t)_{t\in[0,1]}$9 (Beiglboeck et al., 2010).
• Komlós lemma: Uniform integrability ensures existence of convex combinations of martingale increments converging strongly in $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$0 (Beiglboeck et al., 2010).
• Limit procedures: Piecewise-constant or left-continuous approximations of $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$1 (along dyadics) are shown to converge to a predictable, increasing process.
• Predictability: The predictability of $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$2 is ensured by the fact that the limit of left-continuous, adapted processes is still predictable (using limiting arguments and the structure of the predictable $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$3-algebra).
• Nonstandard analysis approach: An alternative proof uses hyperfinite time grids and transfer principles to construct the decomposition, bypassing classical subsequence and compactness arguments (Matsunaga, 23 Aug 2025).

3. Extensions and Generalizations

G-Expectation Framework

In nonlinear settings—such as sublinear expectation spaces (G-expectation, initiated by Peng)—a version of the Doob–Meyer decomposition holds for G-supermartingales, taking the form $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$4, where $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$5 is a G-martingale and $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$6 is an increasing, right-continuous, adapted process (Chen, 2013). The "naturalness" of $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$7 (orthogonality to G-martingales) replaces the classical predictability.

$\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$8-Supermartingale Systems

For $\{S_\tau:\,\tau\text{ stopping in }[0,1]\}$9-supermartingales associated with backward stochastic differential equations (BSDEs) driven by a driver $0\leq s\leq t\leq1$0, one obtains a Doob–Meyer–Mertens decomposition: for any $0\leq s\leq t\leq1$1-supermartingale system $0\leq s\leq t\leq1$2, there exist processes $0\leq s\leq t\leq1$3, $0\leq s\leq t\leq1$4, $0\leq s\leq t\leq1$5, and an orthogonal local martingale $0\leq s\leq t\leq1$6 such that

$0\leq s\leq t\leq1$7

with $0\leq s\leq t\leq1$8 predictable, increasing, $0\leq s\leq t\leq1$9 (Bouchard et al., 2015).

Stable Processes and Fractional Calculus

For $E[|S_t|]<\infty$0 with $E[|S_t|]<\infty$1 a strictly $E[|S_t|]<\infty$2-stable Lévy process and $E[|S_t|]<\infty$3, the Doob–Meyer decomposition has a martingale term given by compensated Poisson integrals and an explicit finite variation compensator involving local times and fractional powers (Cano et al., 2022).

4. Applications and Structural Roles

• Optimal Stopping and the Snell Envelope: The canonical decomposition of the Snell envelope $E[|S_t|]<\infty$4 is crucial in optimal stopping. The compensator $E[|S_t|]<\infty$5 is absolutely continuous with respect to the decreasing part of the drift in the gains process, and the value function inherits regularity properties from the decomposition (e.g., smooth pasting in free boundary problems) (Jacka et al., 2017).
• Stochastic Control and Duality: The Doob–Meyer martingale part features in duality formulations of stochastic control, notably in the Davis–Karatzas–Rogers dual for optimal stopping, where the optimal control is characterized through the generator domain via the decomposition (Jacka et al., 2017).
• Convex Function Analysis: By decomposing $E[|S_t|]<\infty$6 for convex $E[|S_t|]<\infty$7 and Brownian motion $E[|S_t|]<\infty$8, one relates the increasing process to the Hessian measure of $E[|S_t|]<\infty$9, leading to stochastic proofs of Alexandrov's theorem on second-order differentiability almost everywhere (Nguyen, 23 Mar 2026).
• Explicit Solution of Multi-dimensional Stopping Problems: In explicit optimal stopping for sum- and product-type payoffs (e.g., Poisson disorder, investment, house-selling), the decomposition allows explicit determination of optimal stopping boundaries through identification of the drift density in the additive or multiplicative Doob–Meyer form (Christensen et al., 2017).

5. Multiplicative and Nonlinear Decompositions

When dealing with positive special semimartingales, a multiplicative Doob–Meyer decomposition is available. For example, for $E[S_t|\mathcal{F}_s]\ge S_s$0 with $E[S_t|\mathcal{F}_s]\ge S_s$1, one can write $E[S_t|\mathcal{F}_s]\ge S_s$2 with $E[S_t|\mathcal{F}_s]\ge S_s$3, $E[S_t|\mathcal{F}_s]\ge S_s$4, splitting the process into a positive local martingale factor and a "growth" factor (Christensen et al., 2017).

Nonlinear and infinite-dimensional settings, as well as frameworks lacking right-continuity or possessing only sublinear structure, support their own analogues of the decomposition, often with compensator uniqueness characterized by naturalness or regularity conditions rather than simple predictability (Chen, 2013, Bouchard et al., 2015).

6. Predictability, Compensators, and Doleans–Dade Characterization

The predictable, increasing compensator in the Doob–Meyer decomposition is characterized as the unique predictable process such that the difference with the original submartingale is a martingale. Equivalently, Doléans–Dade's theorem relates predictability and naturalness (orthogonality to bounded martingales) (Matsunaga, 23 Aug 2025):

$E[S_t|\mathcal{F}_s]\ge S_s$5

In G-martingale frameworks, predictability is substituted by right-continuity and naturality as the defining property of the compensator (Chen, 2013).

7. Influence and Generalizations

The Doob–Meyer decomposition has had wide influence:

• It underlies the theory of semimartingales and general stochastic integration,
• Forms the backbone of stochastic control, filtering, and dynamic risk measure theory,
• Admits short, elementary, or nonstandard proofs via hyperfinite grids and standardization (Matsunaga, 23 Aug 2025),
• Extends through regularization to systems without right-continuity, to monotone stopping problems in multidimensional settings, and to various forms of nonlinear expectations,
• Facilitates solution of certain PDEs with obstacles via connection to potential theory and local time formulae (Jacka et al., 2017, Nguyen, 23 Mar 2026),
• Provides explicit calculational tools in optimal stopping for both classical and non-classical processes (Christensen et al., 2017).

The decomposition also serves as a primary example of the interaction between path properties, filtration regularity, integrability, and the interplay of martingale and drift terms in advanced stochastic analysis.