The Kolmogorov forward equation for a distributed model of regime-switching diffusions

1. Problem Statement

The paper addresses the mathematical modeling of regime-switching diffusion processes, which are stochastic processes combining continuous dynamics (diffusion) with discrete events (switching between states).

• Context: Traditional models use a finite number of discrete hidden states ($M$ states), leading to a coupled system of $M+1$ partial differential equations (PDEs), specifically Kolmogorov forward (Fokker-Planck) equations. As $M$ increases, the system becomes computationally complex and difficult to solve analytically.
• Objective: The authors propose transitioning from a discrete set of states to a continuum of states distributed over a segment (e.g., $s \in [0, 1]$). The goal is to derive a single integro-differential equation that describes the density of states, allowing for explicit analytical solutions under specific conditions and providing a framework to approximate discrete models.
• Scope: The study covers both standard switching diffusions and those with an advection term (generalized Ornstein-Uhlenbeck processes), which exhibit mean reversion.

2. Methodology

The authors employ a constructive algorithm based on spectral methods and Fourier transforms to solve the derived integro-differential equation.

A. Mathematical Formulation

The continuous model is defined by the following integro-differential equation for the density $p(t, x, s)$, where $x$ is the spatial variable and $s$ is the continuous state variable:
$$p_t = \int_0^1 K(x, s, \xi)p(t, x, \xi) d\xi - (\beta(x, s)p)_x + \frac{1}{2}(R^2(x, s)p)_{xx}$$

• Kernel $K$: Represents the switching rates between continuous states. It satisfies an analog of the $q$-property (non-negative off-diagonal, zero row sum).
• Coefficients: $\beta(x, s)$ is the drift (advection), and $R(x, s)$ is the diffusion coefficient.
• Discrete Approximation: The paper demonstrates how a step-wise kernel $K$ and coefficients can approximate a discrete $M$-state model by partitioning the interval $[0, 1]$.

B. Solution Construction

The solution is constructed via the following steps:

1. Fourier Transform: The equation is transformed with respect to the spatial variable $x$ (variable $\mu$).
2. Variable Transformation:
   • If advection is zero ($b=0$), a simple exponential transformation removes the diffusion term.
   • If advection is non-zero ($b \neq 0$), a more complex transformation involving time-dependent scaling is used to handle the drift and mean reversion.
3. Spectral Expansion: The transformed solution is expanded in an orthonormal basis $X_l(s)$ (e.g., trigonometric functions or Haar wavelets) over the state domain $s \in [0, 1]$.
   $$V(t, \mu, s) = \sum_{l=0}^{\infty} a_l(t, \mu) X_l(s)$$
4. Reduction to ODE/PDE Systems: Substituting the expansion into the equation reduces the problem to a system of linear ordinary differential equations (for $b=0$) or first-order PDEs (for $b \neq 0$) for the coefficients $a_l(t, \mu)$.
   • The kernel $K(s, \xi)$ is expanded as $K(s, \xi) = \sum A_{nm} X_n(s) X_m(\xi)$.
   • The system becomes $\dot{a}_k = \sum A_{kl} a_l$.
5. Inverse Transform: The coefficients are solved, and the inverse Fourier transform is applied to recover the density $p(t, x, s)$.

3. Key Contributions

• Continuous State Formulation: The paper successfully formulates the Kolmogorov forward equation for a regime-switching process with a continuously distributed state variable, unifying the description of discrete switching models.
• Constructive Analytical Algorithm: The authors provide a rigorous algorithm to find explicit solutions for the Cauchy problem. This is a significant departure from standard numerical approaches required for high-dimensional discrete systems.
• Haar Wavelet Basis: The paper highlights the utility of the Haar wavelet basis for approximating discrete models. Because Haar functions are piecewise constant, they align perfectly with step-wise kernels used to model discrete states, allowing for exact reconstruction of discrete dynamics from the continuous model.
• Explicit Solutions for Specific Initial Data: The authors derive closed-form solutions for:
  • Uniform Gaussian initial distributions.
  • Step-wise Gaussian distributions.
  • Delta-function (point mass) initial distributions.
• Steady-State Analysis: For models with mean reversion ($b < 0$), the paper analyzes the long-term behavior ($t \to \infty$), showing convergence to non-trivial steady states that can be multi-peaked.

4. Results

The paper presents specific case studies to validate the methodology:

• Case $M=1$ (One Hidden State):

  • The continuous model approximates a 2-state system.
  • Using the Haar basis, the authors derive explicit formulas for the density evolution.
  • Result: For $b < 0$ (mean reversion), the system converges to a steady state. If the initial data is a delta function, the steady state is explicitly calculable. If the initial data is Gaussian, the steady state involves Whittaker functions.
  • Observation: The steady-state structure depends on the initial conditions (specifically the projection onto the first basis mode), which is an unusual property for diffusion equations.

• Case $M=3$ (Four-State System):

  • The model is applied to a 4-state hierarchical system (multifractal model).
  • Using the first four Haar basis functions, the authors solve the system explicitly.
  • Result: The solution is a sum of four terms involving decaying exponentials and oscillatory terms (due to complex eigenvalues of the switching matrix).
  • Stability: The eigenvalues of the switching matrix have negative real parts, ensuring stability. For $b < 0$, the steady state consists of four distinct peaks.

• Numerical Illustrations:

  • Figures in the paper demonstrate the time evolution of densities for main and hidden states, showing how initial Gaussian or delta distributions spread and settle into steady states with multiple peaks.

5. Significance and Implications

• Analytical Insight: The work provides a rare example where high-dimensional regime-switching problems can be solved analytically, offering deep insight into the interplay between diffusion, drift, and switching rates.
• Approximation of Discrete Models: It establishes a rigorous link between discrete and continuous regime-switching models. The continuous model acts as a "master equation" that can be discretized to recover specific finite-state models with high accuracy using the Haar basis.
• Computational Efficiency: While the paper focuses on analytical solutions, the spectral method suggests efficient numerical schemes for complex switching systems where traditional PDE solvers struggle with dimensionality.
• Generalization: The methodology is extensible to multivariate spatial variables ($x \in \mathbb{R}^n$) and can be applied to generalized Ornstein-Uhlenbeck processes, which are crucial in finance (interest rate modeling) and physics.
• Theoretical Contribution: It proves the existence and uniqueness of classical non-negative solutions for the proposed integro-differential equation under Lipschitz continuity assumptions, bridging stochastic process theory with functional analysis.

In summary, Bratus and Rozanova demonstrate that by treating the regime-switching states as a continuous distribution, one can transform a complex coupled system of PDEs into a solvable integro-differential equation, yielding explicit solutions that reveal the structural dynamics of mean-reverting switching diffusions.