Abstract
This paper investigates the stochastic recursive control problem where the aggregator (or generator) of the backward stochastic differential equation (BSDE) describing the running cost is continuous and monotonic, but not necessarily Lipschitz continuous with respect to the first unknown variable and the control. We establish the Dynamic Programming Principle (DPP) and characterize the value function as the unique viscosity solution of the associated Hamilton-Jacobi-Bellman (HJB) equation. These results are achieved through a generalized comparison theorem for BSDEs and the stability properties of viscosity solutions. Finally, we apply this framework to the continuous-time Epstein-Zin utility model, providing a rigorous theoretical foundation for its use in portfolio optimization.
Introduction
In stochastic control theory, the recursive utility framework—modeled via Backward Stochastic Differential Equations (BSDEs)—is a powerful tool for capturing complex risk preferences. Traditionally, proving the Dynamic Programming Principle (DPP) and deriving the Hamilton-Jacobi-Bellman (HJB) equation required the "aggregator" (the function $f$ driving the cost) to satisfy a Lipschitz continuity condition. This condition ensures that the cost function changes predictably relative to its inputs.
However, many sophisticated models in mathematical finance, such as the Epstein-Zin utility, involve aggregators that are continuous but fail to meet the strict Lipschitz requirement. This creates a theoretical disconnect: while these models are widely used in practice, the mathematical justification for solving them via standard dynamic programming or PDE methods was previously incomplete. This research bridges that gap by relaxing the Lipschitz constraint in favor of monotonicity and continuity.
Theoretical Framework
The Problem of Non-Lipschitz Aggregators
The core of the stochastic recursive control problem is the value function defined by the solution to a BSDE. When the aggregator $f$ is non-Lipschitz, standard existence and uniqueness results for BSDEs, as well as the comparison theorems necessary for the DPP, become more difficult to establish. The authors address this by assuming the aggregator is continuous and monotonic with respect to the first unknown variable ($y$).
Establishing the Dynamic Programming Principle (DPP)
To prove the DPP (Theorem 3.5), the authors utilize a Generalized Comparison Theorem. This theorem ensures that the order of terminal costs is preserved in the value function even without Lipschitz smoothness. By employing the framework of backward semigroups, they formally define the value function in a non-Markovian setting, allowing the control problem to be decomposed into smaller, nested sub-problems.
The HJB Equation and Viscosity Solutions
The second major contribution is connecting the value function to a partial differential equation (PDE). Because the value function may not be smooth enough to solve the HJB equation in a classical sense, the authors turn to viscosity solutions.
To prove this connection (Theorem 4.8), they use a "smootherization" technique:
- Mollification: The rough, non-Lipschitz aggregator is approximated by a sequence of smooth (Lipschitz) functions.
- Convergence: They prove that as these smooth approximations converge to the original aggregator, their corresponding value functions converge to the original value function.
- Stability: Using the stability property of viscosity solutions, they demonstrate that the limit of these solutions is indeed the unique viscosity solution of the original HJB equation.
Application: Epstein-Zin Utility
The practical utility of this theory is demonstrated through the continuous-time Epstein-Zin utility model. This model is favored in macroeconomics and finance because it allows for the separation of risk aversion and the intertemporal elasticity of substitution. However, its aggregator is typically non-Lipschitz.
The authors apply their results to a portfolio optimization problem involving wealth and consumption. By applying the generalized DPP and HJB framework, they derive the specific HJB equation for this model, providing the rigorous mathematical grounding necessary for its application in complex financial environments.
Key Insights
- Monotonicity over Smoothness: The study reveals that monotonicity is a more fundamental requirement than Lipschitz continuity for maintaining the structure of stochastic control problems.
- Approximation as a Solution Path: The use of mollifiers proves that "rough" problems can be solved by treating them as the limit of a series of well-behaved problems, provided the transition is stable.
- Bridging Theory and Practice: By justifying the HJB approach for Epstein-Zin utilities, the paper provides a solid foundation for researchers using these models to study risk and consumption.
Limitations and Future Directions
While the paper significantly advances the field, it notes specific limitations:
- Independence of Diffusion: In the derivation of the HJB equation, the aggregator $f$ is assumed to be independent of the second unknown variable ($z$), which represents the diffusion or risk component. Extending these results to $z$-dependent, non-Lipschitz aggregators remains a significant open challenge.
- Stochastic Environment: The current work focuses on standard diffusion cases. Future research could explore integrating these non-Lipschitz results into models involving jumps, delays, or other complex stochastic processes.
Summary updated