Abstract
We consider the Koopman operator semigroup $(K^t)_{t\ge 0}$ associated with stochastic differential equations (SDEs) of the form $dX_t = AX_t\,dt + B\,dW_t$, where $A$ and $B$ are constant matrices and $W_t$ represents Brownian motion. This paper establishes that the reproducing kernel Hilbert space (RKHS) $\mathbb{H}_C$ generated by a Gaussian kernel with a positive definite covariance matrix $C$ is invariant under each Koopman operator $K^t$ if the matrices $A$, $B$, and $C$ satisfy a specific Lyapunov-like matrix inequality: $AC^2 + C^2A^\top \le 2BB^\top$. Additionally, we provide a characterization for the inclusion of Gaussian RKHSs, proving that $\mathbb{H}_{C_1} \subset \mathbb{H}_{C_2}$ if and only if $C_1^2 \ge C_2^2$.
Introduction
The Koopman operator has emerged as a cornerstone of modern dynamical systems theory, offering a way to analyze nonlinear dynamics through a linear lens by shifting focus from state-space trajectories to the evolution of observable functions. In the context of stochastic systems, the Koopman operator $K^t$ is defined as a conditional expectation: $K^t f(x) = \mathbb{E}[f(X_t) | X_0 = x]$.
For practitioners using kernel-based methods like Kernel Extended Dynamic Mode Decomposition (kEDMD), a critical theoretical requirement is the invariance of the chosen function space under the Koopman operator. If the RKHS is invariant, the operator maps functions within the space back into the same space, allowing for rigorous error bounds and stable numerical approximations. However, identifying which kernels are compatible with specific stochastic dynamics remains a significant challenge. This study addresses this gap for linear SDEs (Ornstein-Uhlenbeck processes) and Gaussian kernels.
Characterizing Gaussian RKHS Inclusion
A fundamental contribution of this work is clarifying the relationship between the kernel's covariance matrix and the resulting Hilbert space's "size." It is a known technical hurdle in functional analysis to determine when one RKHS is a subset of another.
The authors prove that for two positive definite matrices $C_1$ and $C_2$, the inclusion $\mathbb{H}_{C_1} \subset \mathbb{H}_{C_2}$ holds if and only if $C_1^2 \ge C_2^2$ in the sense of the Loewner partial order. This result provides the necessary mathematical machinery to track how the Koopman operator transforms the "effective" covariance of a function over time and whether that transformed function remains within the original RKHS.
The Invariance Condition
The core result of the paper (Theorem 2.1) identifies a sufficient condition for the invariance of $\mathbb{H}_C$ under the Koopman semigroup. For a system defined by drift matrix $A$ and diffusion matrix $B$, the RKHS defined by covariance $C$ is invariant if:
$$\frac{1}{2}(AC^2 + C^2A^\top) \le BB^\top$$
This inequality represents a balancing act between three forces:
- Drift ($A$): The deterministic component of the system.
- Diffusion ($B$): The stochastic noise or "chaos" injected into the system.
- Kernel Geometry ($C$): The scale and orientation of the Gaussian basis functions.
In the simplified case of isotropic radial basis function (RBF) kernels, where $C = \sigma I$, the condition reduces to an intuitive form: $\frac{1}{2}(A + A^\top) \le \frac{1}{\sigma^2} BB^\top$. This suggests that higher levels of noise ($B$) or smaller kernel bandwidths (larger $\sigma$) make it easier to achieve invariance.
Operator Norms and Stability
Beyond proving invariance, the paper provides bounds on the norm of the Koopman operator within these spaces. Under the identified Lyapunov condition, the operator satisfies:
$$\|K^t f\|_C \le e^{t \cdot \text{Tr}(-A)/2} \|f\|_C$$
This bound is significant for numerical stability. For systems where the drift matrix $A$ is Hurwitz (stable), the trace of $-A$ is positive, but the specific scaling suggests that the operator's growth is strictly controlled by the system's internal dissipation. This ensures that errors in data-driven approximations do not explode over time, provided the invariance condition is met.
Conclusion and Future Directions
This research provides a rigorous framework for selecting Gaussian kernels that are mathematically compatible with linear stochastic dynamics. By satisfying the Lyapunov-like inequality, researchers can guarantee that their Koopman-based models are theoretically sound.
Several avenues for future work remain. While the condition $AC^2 + C^2A^\top \le 2BB^\top$ is proven to be sufficient, its necessity remains an open question. Furthermore, extending these results to nonlinear SDEs and non-Gaussian kernels (such as Matérn or polynomial kernels) would broaden the applicability of these stability guarantees to a wider range of real-world dynamical systems.
Summary updated