Abstract
This paper extends the "invariant imbedding" method from deterministic calculus to the domain of stochastic differential equations (SDEs). By transforming complex boundary value problems (BVPs) into a family of simpler initial value problems (IVPs), the author provides a robust framework for proving the existence and uniqueness of solutions for "anticipating" SDEs—equations where the state depends on both past and future noise. The study utilizes Malliavin calculus and generalized Stratonovich integrals to handle non-adapted integrands, establishing critical conditions for solution stability in both one-dimensional and multi-dimensional systems.
Introduction
In standard stochastic calculus, systems are typically modeled as Initial Value Problems (IVPs), where the state evolves from a known initial condition $x_0$ based on drift and random diffusion. A key property of these systems is that they are "adapted" to the natural filtration of the Brownian motion; they do not require knowledge of future noise.
However, many physical and economic systems are better described as Boundary Value Problems (BVPs), where constraints are imposed at both the start and the end of a time interval (e.g., $h_0 x_0 + h_1 x_T = v_0$). These problems are "anticipating" because the solution at any time $t$ must account for the boundary condition at the future time $T$. Standard Itô calculus fails in this context because the integrands are no longer adapted. This paper introduces a systematic approach to prove the existence and uniqueness of solutions for such equations using the principle of invariant imbedding.
Methodology
The author employs Stochastic Invariant Imbedding, a technique that treats the parameters of the boundary condition—specifically the interval length and the boundary value—as variables.
Transformation to Initial Value Problems
The core strategy involves transforming the BVP into a system of coupled partial differential equations (PDEs) with initial conditions. This "imbedding" philosophy suggests that rather than solving one difficult BVP, one should solve a family of easier IVPs. By varying the boundary "target," the entire space of possible solutions is mapped.
Malliavin Calculus and Generalized Integrals
To address the anticipating nature of the equations, the paper utilizes the framework of Malliavin calculus, specifically the work of Nualart and Pardoux regarding generalized Stratonovich integrals. This allows for integration against Brownian motion even when the integrand "sees" the future. The author introduces auxiliary stochastic processes, $\tilde{R}$ and $\tilde{V}$, which are adapted. By solving these standard SDEs and inverting the resulting relationship, the solution to the original non-adapted BVP is recovered.
Key Results
The research establishes several foundational theorems for the study of anticipating SDEs:
- Existence and Uniqueness: Theorems 1 and 2 provide general conditions for the existence and uniqueness of solutions to SDEs with affine boundary conditions.
- Condition $\mathcal{H}2$ (The Invertibility Key): The paper identifies a crucial condition—the invertibility of a specific random matrix $Z(\tau, v)$. If this matrix remains invertible, a unique solution to the BVP is guaranteed.
- Dimensional Analysis:
- For one-dimensional problems, the condition for uniqueness is simple: a solution exists if the boundary coefficients $h_0$ and $h_1$ have the same sign (Theorem 3).
- For multi-dimensional systems, the author provides concrete bounds on the "steepness" of the drift function relative to the boundary matrices (Theorem 4).
- Limit Theorem: The author proves that the solution to a BVP driven by a rapidly oscillating Markov process converges weakly to the solution of a stochastic BVP as the oscillation frequency approaches infinity, demonstrating a bridge between deterministic and stochastic methods.
Discussion and Future Work
While the framework is powerful, it relies on significant regularity constraints. The uniqueness result requires the solution to reside in a specific Malliavin-Sobolev space ($\mathbb{L}_{C,loc}^{1,8}$), implying a high degree of smoothness. Furthermore, the current work is limited to affine boundary conditions. Extending this methodology to fully non-linear boundary conditions $F(X(0), X(1)) = 0$ remains a significant challenge for future research. The author also suggests that regularity hypotheses on the drift function could be weakened in future iterations, particularly when the diffusion coefficient depends only on time.
Summary updated