Martingale-Coboundary Decomposition

• Martingale-coboundary decomposition is a technique that represents a stochastic process as the sum of a martingale component and a telescoping coboundary term.
• It transfers complex asymptotic behavior to the martingale part, thereby facilitating the application of limit theorems like the CLT, invariance principles, and large deviations.
• The framework extends to non-stationary, multidimensional, and functional-analytic settings, with applications in ergodic theory, random fields, and dynamical systems.

Searching arXiv for recent and foundational papers on martingale-coboundary decomposition. Martingale–coboundary decomposition is a representation of a stochastic process, random field, or observable as the sum of a martingale component and a telescoping correction term, usually called a coboundary. In discrete time, the basic form is

$X_i = Y_i + U_i - U_{i+1},$

where $(Y_i)$ is a martingale difference sequence and $(U_i)$ is an adapted sequence. Equivalently, partial sums satisfy

$S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$

so the asymptotic behavior of $(X_i)$ can be reduced to that of the martingale part provided the telescoping term is negligible under the relevant normalization. In stationary ergodic theory this is the representation

$f=m+g-g\circ T,$

with $(m\circ T^i)$ a martingale difference sequence. The modern theory extends this scheme to non-stationary processes, random fields, nonuniformly hyperbolic dynamics, Orlicz spaces, Hardy martingales, and continuous-time Markov processes generated by semi-Dirichlet forms (Volny, 2023).

1. Classical formulation and core mechanism

The classical one-parameter formulation starts from a stationary sequence $X_i=f\circ T^i$ on a probability space with a measure-preserving transformation $T$. One seeks

$f=m+g-g\circ T,$

where $(Y_i)$0 is a martingale difference sequence. In non-stationary discrete time, the corresponding representation is

$(Y_i)$1

with $(Y_i)$2 adapted and satisfying $(Y_i)$3 (Volny, 2023).

The functional significance of the decomposition lies at the level of partial sums. Because

$(Y_i)$4

the process inherits CLT, invariance principle, LIL, and large-deviation behavior from the martingale term once the coboundary contribution is negligible after normalization. In the one-dimensional stationary setting this is the basic mechanism behind Gordin-type approximations; in the non-stationary setting the same logic persists, but the relevant projective conditions are indexed by time rather than generated by a single shift (Volny, 2023).

A common misconception is that martingale approximation is only an $(Y_i)$5, stationary, or ergodic-theoretic device. The recent general formulation treats arbitrary adapted sequences relative to a two-sided filtration and works in general Banach function spaces $(Y_i)$6, including Orlicz spaces, provided conditional expectation is continuous in the $(Y_i)$7-norm (Volny, 2023). This shows that stationarity is a special case rather than the conceptual core.

2. Non-stationary representation and projective characterizations

A central modern result gives a necessary and sufficient criterion for non-stationary martingale–coboundary representation. Let $(Y_i)$8 be adapted to $(Y_i)$9, with $(U_i)$0 $(U_i)$1-measurable and

$(U_i)$2

Define

$(U_i)$3

Then the existence of

$(U_i)$4

with $(U_i)$5 a martingale difference sequence and with the conditional vanishing properties

$(U_i)$6

is equivalent to convergence in $(U_i)$7 of the projective series $(U_i)$8 and $(U_i)$9 for every $S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$0. Moreover,

$S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$1

where

$S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$2

is the martingale projection (Volny, 2023).

This characterization is sharp in the sense that the decomposition exists if and only if the forward and backward projective series converge. The conditions are genuinely two-sided. In the stationary case they collapse to the familiar pair of series

$S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$3

recovering the traditional stationary theory as a corollary (Volny, 2023).

For stationary random fields indexed by $S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$4, the representation becomes multidirectional. One obtains decompositions of the form

$S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$5

with the top-level term generating an orthomartingale difference field and the remaining terms acting as higher-order coboundaries. Necessary and sufficient $S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$6 conditions are formulated through multi-parameter projection operators $S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$7, while sufficient conditions arise from multidimensional Hannan and Maxwell–Woodroofe criteria (Giraudo, 2017). A related projective approach yields an orthomartingale–coboundary decomposition for stationary random fields under coordinatewise Gordin-type conditions and transfers orthomartingale limit theorems to general fields (Machkouri et al., 2014). Volný’s refinement gives $S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$8 and $S_n=\sum_{i=1}^n X_i=\sum_{i=1}^n Y_i+U_1-U_{n+1},$9 necessary-and-sufficient criteria for stationary random fields with completely commuting filtrations, extending Heyde’s and Volný’s one-dimensional results and providing new WIP and large-deviation estimates (Volny, 2017).

3. Functional-analytic settings: $(X_i)$0, Orlicz, and beyond

The abstract non-stationary theory is formulated in a Banach space $(X_i)$1 of integrable random variables whose norm depends only on the distribution, is closed under conditional expectation, and for which conditional expectation is continuous in the norm. This includes $(X_i)$2 spaces for any $(X_i)$3 and Orlicz spaces $(X_i)$4, but excludes $(X_i)$5 and some Orlicz spaces where martingale convergence fails (Volny, 2023).

A key example is the Orlicz space with

$(X_i)$6

equipped with the Luxemburg norm

$(X_i)$7

In such spaces forward martingale convergence may hold while backward martingale convergence can fail; explicit counterexamples exist both in $(X_i)$8 and in $(X_i)$9 (Volny, 2023). This matters because many transfer arguments in martingale approximation use backward conditional convergence implicitly. The general projective-series characterization avoids assuming it.

This broader functional-analytic perspective also appears in adjacent decomposition theories. In vector-lattice stochastic analysis, an $f=m+g-g\circ T,$0-bounded martingale in a Dedekind complete Riesz space with conditional expectation can be decomposed into three martingales

$f=m+g-g\circ T,$1

where $f=m+g-g\circ T,$2 is $f=m+g-g\circ T,$3-bounded with controlled increments, $f=m+g-g\circ T,$4 has absolutely convergent difference series, and $f=m+g-g\circ T,$5 is uniformly bounded in $f=m+g-g\circ T,$6. The paper does not formulate an ergodic coboundary $f=m+g-g\circ T,$7, but the absolutely summable component is structurally analogous to the bounded or summable remainder that often becomes coboundary-like in stationary settings (Niouar et al., 2024).

A plausible implication is that the conceptual reach of martingale–coboundary methods is larger than the literal formula $f=m+g-g\circ T,$8: what matters is the extraction of a martingale core plus a residual term whose cumulative effect is boundary-dominated, summable, or otherwise negligible under scaling.

4. Dynamical systems, random fields, and geometric variants

For nonuniformly hyperbolic dynamical systems modeled by Young towers, the decomposition is constructed directly on the tower. If $f=m+g-g\circ T,$9 is Hölder with $(m\circ T^i)$0, one lifts to

$(m\circ T^i)$1

and obtains

$(m\circ T^i)$2

where $(m\circ T^i)$3 is the tower map and $(m\circ T^i)$4 its transfer operator. The martingale component $(m\circ T^i)$5 belongs to $(m\circ T^i)$6, the coboundary $(m\circ T^i)$7 belongs to $(m\circ T^i)$8, and $(m\circ T^i)$9 is negligible under $X_i=f\circ T^i$0-scaling in the limit theorems derived in the paper (Korepanov et al., 2016).

The same framework extends uniformly to families $X_i=f\circ T^i$1 of nonuniformly expanding or hyperbolic maps and observables $X_i=f\circ T^i$2, allowing WIP, covariance convergence, and homogenization for discrete fast–slow systems with varying fast dynamics. This is one of the main settings where a single explicit martingale–coboundary decomposition, rather than a sequence of approximations, becomes technically decisive (Korepanov et al., 2016).

In the random-field setting, orthomartingale analogues replace one-parameter martingale differences. For stationary $X_i=f\circ T^i$3-actions with completely commuting filtrations, the decomposition separates a field into an orthomartingale difference component and directional coboundaries. The multidimensional structure is not a mere product version of the one-dimensional case: additional mixed terms $X_i=f\circ T^i$4 arise, representing higher-order coboundaries across several directions (Giraudo, 2017).

There are also geometric and categorical generalizations. In a cohomological framework where time is modeled by a small category and filtrations are contravariant functors to probability spaces, martingales arise as $X_i=f\circ T^i$5-cocycles after a normalization called the $X_i=f\circ T^i$6-gauge, and exact $X_i=f\circ T^i$7-cochains represent coboundary gains generated by price systems. The first cohomology group measures “homological arbitrage,” interpreted as a global obstruction to representing consistent gains as coboundaries (Adachi, 2 May 2026). This is not a classical limit-theorem setting, but it retains the same structural division between cocycles, coboundaries, and residual cohomological obstructions.

5. Transfer of limit theorems and deviation bounds

The primary use of martingale–coboundary decomposition is reduction of asymptotic questions for dependent processes to corresponding martingale theorems. In the non-stationary discrete-time setting, if

$X_i=f\circ T^i$8

with $X_i=f\circ T^i$9 and $T$0 in $T$1, and

$T$2

then the CLT for normalized sums of $T$3 holds if and only if it holds for $T$4. With the additional condition

$T$5

the same equivalence holds for the invariance principle (Volny, 2023).

For the LIL, a sufficient tail control is

$T$6

for some $T$7. Then the LIL for $T$8 is equivalent to the LIL for the martingale difference part $T$9 (Volny, 2023). The telescoping structure is essential here, because the normalized coboundary reduces to a boundary term.

Large deviations admit analogous transfers. Under uniform $f=m+g-g\circ T,$0-bounds for $f=m+g-g\circ T,$1 and $f=m+g-g\circ T,$2 with $f=m+g-g\circ T,$3,

$f=m+g-g\circ T,$4

Under uniform boundedness $f=m+g-g\circ T,$5, $f=m+g-g\circ T,$6,

$f=m+g-g\circ T,$7

and under exponential moment bounds on $f=m+g-g\circ T,$8 and $f=m+g-g\circ T,$9,

$(Y_i)$00

for all sufficiently large $(Y_i)$01 (Volny, 2023).

In random fields, the corresponding transfer goes through orthomartingales. Under multidimensional Hannan or Maxwell–Woodroofe conditions, orthomartingale approximation yields functional CLTs to Brownian sheets (Giraudo, 2017). Under completely commuting filtrations, Volný’s decomposition leads to weak invariance principles and large-deviation estimates in $(Y_i)$02 and Orlicz settings (Volny, 2017). El Machkouri and Giraudo use orthomartingale-coboundary decomposition to derive maximal inequalities, WIP over VC-classes, and Hölder-space invariance principles for stationary $(Y_i)$03-fields (Machkouri et al., 2014).

For nonuniformly hyperbolic dynamics, the primary decomposition gives WIP with covariance

$(Y_i)$04

and a secondary martingale–coboundary decomposition of the conditional variance process yields ASIP rates depending on the integrability of the return time $(Y_i)$05 (Korepanov et al., 2016).

6. Related decompositions, analogies, and misconceptions

Several decomposition theories are closely related but not identical to martingale–coboundary decomposition. The Gundy–Stein decomposition writes a terminal martingale $(Y_i)$06 as

$(Y_i)$07

with a localized part $(Y_i)$08, an absolutely summable martingale part $(Y_i)$09, and a bounded part $(Y_i)$10, with explicit constants in the positive closed case and a further four-term refinement splitting $(Y_i)$11 into a stopped martingale component and a predictable compensator (Hormozi et al., 30 Mar 2026). The authors do not use ergodic-theoretic coboundary language explicitly, but the predictable compensator and stopped-martingale splitting are structurally close to martingale–coboundary separation.

Hardy-martingale decompositions provide another analogue. For vector-valued Hardy martingales one has

$(Y_i)$12

where $(Y_i)$13 has increments bounded by the past magnitude of $(Y_i)$14, or even by $(Y_i)$15, while $(Y_i)$16 has absolutely summable increments. This is not written as $(Y_i)$17, but the summable remainder behaves like a boundary correction, and in stationary embeddings it can play a coboundary-like role (Müller, 2015).

In continuous time, Fukushima-type decomposition for quasi-regular semi-Dirichlet forms gives

$(Y_i)$18

where $(Y_i)$19 is a martingale additive functional and $(Y_i)$20 is a continuous additive functional of zero quadratic variation on the predictable set $(Y_i)$21. This is the continuous-time analogue of martingale plus coboundary/remainder separation, with $(Y_i)$22 encoding drift or potential-type effects (Ma et al., 2014).

A recurrent misconception is that coboundaries are always negligible. They are negligible only under a specified normalization and under conditions ensuring that the boundary term does not contribute asymptotically. Invariance principles, LILs, and large deviations each require different control of $(Y_i)$23, $(Y_i)$24, or their analogues (Volny, 2023). Another misconception is that the decomposition is unique in a canonical sense. Some settings admit explicit formulas for the transfer term, but uniqueness generally depends on the ambient filtration, regularity class, and whether one fixes additional normalization conditions; in continuous time, even uniqueness of the Doob–Meyer-type split can fail on optional interval-type sets unless one passes to the predictable set $(Y_i)$25 (Ma et al., 2014).

Taken together, these developments show that martingale–coboundary decomposition is best understood as a structural reduction principle. Its exact algebraic form varies—$(Y_i)$26, $(Y_i)$27, orthomartingale plus mixed directional coboundaries, martingale additive functional plus zero-quadratic-variation term—but the underlying aim is stable: isolate a martingale core that governs fluctuations, and relegate dependence, drift, or geometry to a residual term whose contribution can be quantified, transferred, or annihilated by telescoping (Volny, 2023).