Stochastic Integral Framework

• Stochastic Integral Framework is a rigorously defined system that integrates random processes such as semimartingales, Lévy processes, and rough paths.
• It unifies multiple approaches including classical Itô theory, pathwise and model-free constructions, ensuring robust change-of-measure and isometric properties.
• The framework extends to infinite-dimensional settings, jump processes, Volterra-type integrals, and stochastic control, offering broad applicability in uncertainty quantification.

A stochastic integral framework is a rigorously defined system for integrating non-deterministic processes—typically semimartingales, Lévy processes, or stochastic kernels—against random integrators, accommodating a wide spectrum of settings ranging from finite-dimensional diffusions and jump processes to infinite-dimensional path spaces, rough paths, and stochastic Volterra equations. Central to these frameworks are constructions that respect underlying filtrations, support pathwise and quasi-sure definitions, allow for robust change-of-measure results, and are extendable to model-free or non-dominated families. Modern developments unify classical Itô theory, pathwise constructions, rough integration, stochastic control via path integrals, and stochastic calculus on random time-horizons.

1. Foundational Frameworks: Semimartingale and Itô Integration

The canonical construction of the stochastic integral arises in the context of semimartingales. Given a filtered probability space $(\Omega,\mathcal F,(\mathcal F_t),P)$ and a real or complex semimartingale $X$, the stochastic integral $\int_0^t H_s\,dX_s$ is defined first for simple, predictable integrands $H$ and then extended via $L^2$ convergence or ucp convergence for adapted processes whose integrability is determined by quadratic variation or the semimartingale characteristics. The Itô isometry

$\E\left[ \left(\int_0^T H_s\,dW_s\right)^2 \right] = \E\left[ \int_0^T H_s^2\,ds \right]$

characterizes the isometric extension for Wiener integrators and underlies the martingale property of the integral (Nielsen, 2018).

For multidimensional and complex-valued processes, integration is constructed via real and imaginary projections, and for infinite-dimensional settings, cylindrical Wiener processes and Malliavin calculus are employed to define integrals via expansion in appropriate bases and weak convergence arguments (Benth et al., 2013, Olivera, 2012).

Key to the uniqueness and existence are properties that:

• The integral of a simple, $\mathcal{F}_t$-predictable process coincides with the sum of increments;
• The integral extends continuously in probability when the $L^2$ norm of the integrand converges.

Extensions to arbitrary filtrations refine the theory, showing that the value of the stochastic integral can depend on the underlying filtration unless integrands are left-continuous (LCRL), bounded, or if the semimartingale is regular; various theorems formalize when different filtrations give coincident integrals (Karandikar et al., 2020).

2. Pathwise, Model-Free, and Infinite-Dimensional Integration

Pathwise constructions aim to define stochastic integrals as pointwise limits along the path, independent of any fixed probability measure. A core result is the medial limit approach: given a family $\mathcal{P}$ of (possibly singular) probability measures under which the integrator is a semimartingale, one can approximate any predictable integrand by finite-variation integrands and define the integral as the medial limit of the corresponding Lebesgue–Stieltjes sums. This construction yields a process that, under every $P\in\mathcal{P}$, coincides almost surely with the traditional Itô integral. This aggregation theorem thus provides a universal definition across non-dominated models and is especially significant under uncertainty or in robust finance (Nutz, 2011).

When stochastic integration is sought without probabilistic structure (i.e., on typical paths or under outer measures), the framework employs an outer measure $\mu$ defined in terms of minimal superhedging for claims nonnegative on prediction sets. Integrals are constructed via uniform BDG-type estimates and pathwise limits of simple strategies or step integrators. For paths with absolutely continuous quadratic variation, $L^2$ approaches offer further regularity and enable the solution of SDEs in a purely pathwise, model-independent sense (Bartl et al., 2018).

In rough path theory, the integration extends to $p$-variation paths, allowing definition of rough integrals using the limit of left-point Riemann sums along carefully chosen partitions (Property (RIE)), and connecting with classical quadratic variation in the sense of Föllmer. The rough integral construction ensures pathwise existence and convergence for a broad class of integrators, including non-gradient and path-dependent integrands, thus extending the range of admissible trading strategies or control laws well beyond semimartingale classes (Allan et al., 2021).

3. Stochastic Integrals with Jumps, Lévy Processes, and Discrete Diffusions

In settings with jump processes or discrete state spaces, stochastic integrals are constructed against Poisson random measures or Lévy-type integrators. For a finite state space $\mathcal{X}$ and a predictable intensity $\lambda_t(y)$, the random measure $N[\lambda](dt, dy)$ encodes a point process of jumps, and stochastic integrals are written as sums over $K_t(y)\,N[\lambda](dt, dy)$. The compensated random measure $\tilde N[\lambda] = N[\lambda] - \lambda_t(y)\,dt$ yields martingale properties and isometry relations analogous to the continuous case (Ren et al., 2024).

Essential structural features include:

• The explicit Itô isometry for Poisson integrals;
• Change-of-measure results (discrete-time Girsanov theorem) giving the Radon–Nikodym derivative as an exponential of the integral with respect to the jump measure;
• Decomposition of KL-error in discrete schemes (τ-leaping) into truncation, approximation, and discretization, permitting sharp control of the error for algorithmic implementations in machine learning, sampling, or stochastic simulation.

Change of measure (Girsanov) results for Poisson random measures (as with Brownian motion) permit reweighting of intensities in the likelihood ratio process, essential for backward sampling and for discrete-diffusion error bounds.

4. Volterra-Type Integrals, SVIEs, and Fractional Kernels

A rich class of integrals arises in systems with memory, where dynamics are expressed in terms of stochastic Volterra integral equations (SVIEs) or their backward analogues (BSVIEs). The pathwise integral is defined using products of stochastic Volterra kernels—deterministic and stochastic—organized as Banach algebras under *-products. For instance, the forward variation-of-constants formula for linear SVIEs writes the solution as

$X = \psi + Q * \psi + R * \psi,$

where $(Q,R)$ are the (,)-resolvents for the Volterra kernels, and * denotes the product of kernels acting on the underlying $L^2$-process. Singular kernels, such as those appearing in fractional SDEs, can be accommodated, and explicit representations can be given in terms of resolvent series.

BSVIEs admit analogous backward variation-of-constants formulae, with integration duality and "integration by parts" principles mapping directly between adjoint/primal processes, and these dualities underpin stochastic maximum-principle structure in stochastic control and differential games (Hamaguchi, 2021, Wang et al., 2010).

5. Stochastic Integration for Path Integral Control and Optimal Control

Path integral stochastic control frameworks recast stochastic optimal control (SOC) problems as expectations over trajectory distributions, using pathwise Feynman–Kac representations of the value function: $\Psi(x,t) = \E_{x,t}\left[\exp\left(-\frac{1}{\lambda} \left( q(x_T) + \int_t^T q(x_s, s)\,ds \right)\right)\right].$ Optimal controls are synthesized from gradients of desirability functions (exponentially transformed value functions), and the corresponding stochastic integral expressions provide closed-form feedback laws and enable efficient Monte Carlo sampling for high-dimensional control (Pan et al., 2014, Patil, 24 Apr 2025).

The iterative data-driven path integral control approach leverages Gaussian process learned dynamics with moment-matching to propagate distributions and analytically compute one-step integrals in the control update recursion, ensuring sample efficiency and analytic tractability even under uncertain, nonlinear dynamics.

6. Infinite-Dimensional and Generalized Frameworks

In infinite-dimensional settings—e.g., when integrators and integrands take values in Hilbert or Banach spaces—stochastic integration is constructed via Skorohod integrals, Malliavin calculus, and expansions in basis elements. For volatility-modulated Volterra processes, the stochastic integral decomposes as a sum of a Skorohod integral with respect to a cylindrical Wiener process and a correction involving the Malliavin derivative, yielding a two-term representation and enabling anticipating Itô formulas (Benth et al., 2013).

Frameworks for infinite collections of continuous semimartingales define integrals via semimartingale-topology closure, with isomorphic representations in terms of reproducing kernel Hilbert spaces built from the covariation structure. This is pivotal in modeling large or infinite-asset financial markets and establishes exact analogues of hedging, optional decomposition, and completeness as in finite-dimensional markets (Kardaras, 2019).

7. Stochastic Integration on Random Time-Horizons and Intervals

Recent developments formalize stochastic integration on random or stochastic sets of interval type, crucial in contexts such as sudden-stop markets or uncertain time-horizons in finance. Given a stochastic set $B$ of interval type (e.g., defined by a stopping time $T$), all processes, integrals, and strategies are constructed on $B$, and their properties (linearity, isometry, Itô formula) are established componentwise via fundamental coupled sequences and restriction to intervals bounded by increasing stopping times. This localization enables the definition of self-financing and admissible strategies that are robust to sudden termination of market activity (Yue et al., 18 Jun 2025).

Table: Principal Frameworks and Core Objects

Framework	Integrator Class	Main Construction Principle
Classical Itô/semimartingale (Nielsen, 2018)	Semimartingales	L²-predictable limit, isometry, filtration
Pathwise/medial limit (Nutz, 2011)	Càdlàg paths, quasisure	Finite-variation approximation, medial limit
Lévy/Poisson (Ren et al., 2024)	Jump/point processes	Compensated random measure, Itô isometry
Rough paths (Allan et al., 2021)	Càdlàg p-variation paths	RIE, left-point sums, p-rough path lift
Volterra/SVIE (Hamaguchi, 2021, Wang et al., 2010)	Integral operators, memory	Resolvent kernels, *-algebra, duality
Infinite-dimensional (Benth et al., 2013, Kardaras, 2019)	Hilbert- or Banach-valued	Skorohod/Malliavin, RKHS, semimartingale topology
Path integral SOC (Pan et al., 2014, Patil, 24 Apr 2025)	Controlled SDEs	Feynman-Kac expansion, desirability function
Interval-type sets (Yue et al., 18 Jun 2025)	Processes on random horizon	Localization by FCS, restriction, series representation

Conclusion

The stochastic integral framework encompasses a constellation of interrelated methodologies that enable rigorous integration in stochastic environments beyond the deterministic or classical measure-theoretic settings. By accommodating pathwise, filtration-sensitive, jump-modulated, rough, infinite-dimensional, and interval-restricted objects, these frameworks form the backbone of stochastic analysis, robust finance, optimal control, and the rigorous mathematical modeling of systems under uncertainty.