TL;DR
This paper provides the first rigorous proof that Hermitian Dynamic Mode Decomposition (DMD) converges to the true spectral properties of self-adjoint Koopman operators. By establishing a general theorem for the convergence of spectral measures—even for unbounded operators—the authors bridge a critical gap between data-driven dynamical systems modeling and functional analysis, demonstrating their results on 2D Schrödinger equations.
What Problem Does This Paper Address?
In the study of nonlinear dynamical systems, the Koopman operator is a powerful tool that "lifts" nonlinear dynamics into an infinite-dimensional linear space. To make this practical, researchers use data-driven methods like Extended Dynamic Mode Decomposition (EDMD) to find finite-dimensional matrix approximations of these operators.
However, there is a major theoretical "gotcha": standard EDMD often fails to converge to the correct spectral properties (eigenvalues and spectral measures). It can suffer from "spectral pollution" (spurious results) or "spectral invisibility" (missing parts of the spectrum). While a variant called Hermitian DMD—which forces the approximation to be self-adjoint to reflect physical constraints—has shown great empirical success, it lacked a formal proof of convergence. This paper addresses that missing link.
Methods
The authors focus on Hermitian DMD, a "physics-informed" approach that ensures the resulting matrix approximation preserves the self-adjoint (Hermitian) property of the underlying operator.
- Algorithmic Optimization: The authors derive a simplified, more efficient formula for the Hermitian DMD matrix: $K = G^{-1} \left( \frac{A + A^*}{2} \right)$. This avoids the computationally expensive and potentially unstable Singular Value Decompositions (SVD) used in previous iterations of the algorithm.
- General Spectral Theorem: The core of the paper is a new, general convergence theorem (Theorem 4.5). Unlike previous proofs that only applied to "bounded" operators (those that don't "blow up"), this theorem applies to unbounded operators. This is crucial for physics, as many fundamental operators (like the Hamiltonian in quantum mechanics) are unbounded.
- Weak Convergence: They prove that as you add more data (snapshots) and more basis functions (the dictionary), the "spectral measures" of the approximation converge weakly to the true measures of the system.
Datasets
As a theoretical and numerical analysis paper, it uses synthetic data generated from a physical system rather than a standard machine learning benchmark:
- 2D Schrödinger Equation: Specifically, a system with an external harmonic potential.
- Dictionary: A set of 400 Gaussian "bumps" distributed across the state space.
- Snapshots: $300 \times 300$ (90,000) snapshot points used to discretize the integrals required for the DMD process.
Key Insights and Ideas
- Self-Adjointness as a Safeguard: Forcing the model to be Hermitian isn't just a physical constraint; it acts as a mathematical "anchor" that prevents the approximation from drifting into the numerical instabilities (like spectral pollution) that plague standard DMD.
- Handling the "Unbounded" Reality: Most real-world physical operators are unbounded. By using rational functions (the "resolvent") instead of simple polynomials in their proof, the authors created a framework that finally matches the reality of physical modeling.
- Efficiency Gains: The discovery that the symmetric Procrustes problem (the heart of Hermitian DMD) simplifies significantly when the basis is orthonormalized means that these "guaranteed" models are actually faster to compute than previous versions.
Key Results
- Theoretical Guarantee: Theorem 4.6 formally proves that Hermitian DMD spectral measures converge to the true Koopman spectral measures as the number of data points ($M$) and dictionary size ($N$) go to infinity.
- Numerical Accuracy: In the Schrödinger example, the algorithm accurately captured the first 50 eigenvalues of the system with high precision.
- Spectral Measure Alignment: The paper visualizes the "spikes" of the spectral measure (representing energy distributions). As the dictionary size increases, these spikes align almost perfectly with the exact analytical solution (see Figure 4 in the paper).
Limitations and Future Work
- The "Core" Assumption: The proof relies on "Assumption 4.3," which requires the existence of a "core"—a subset of the domain where the operator behaves nicely. While common in physics, identifying this core for an entirely unknown black-box system remains a challenge.
- Dictionary Selection: The convergence still depends on the choice of the dictionary (basis functions). If the dictionary doesn't "span" the relevant parts of the Hilbert space, convergence will be slow or incomplete.
- Extension to Non-Self-Adjoint Systems: While this paper masters the Hermitian case, many systems (like those with friction or dissipation) are not self-adjoint. Extending these rigorous spectral guarantees to general non-Hermitian systems remains a "holy grail" in the field.
Summary updated