2105.03607
Mean Subtraction and Mode Selection in Dynamic Mode Decomposition
Gowtham S Seenivasaharagavan, Milan Korda, Hassan Arbabi, Igor Mezić
correctmedium confidence
- Category
- Not specified
- Journal tier
- Strong Field
- Processed
- Sep 28, 2025, 12:56 AM
- arXiv Links
- Abstract ↗PDF ↗
Audit review
The paper proves that, when Z = ČΘ with Č full column rank and n ≥ r, the DMD eigenvalues with nontrivial modes coincide exactly with the Koopman eigenvalues: σ_nontriv[ČΘ, c*(ČΘ)] = σ(Λ) (Theorem 5.1), with the key mechanism supplied by Appendix D (Lemma D.2) showing that true Koopman eigenvalues yield nonzero DMD modes while spurious eigenvalues yield trivial modes . The candidate’s counterexample hinges on treating the Vandermonde power vector [1, μ, …, μ^{n−1}]^T as a right eigenvector of the companion T[c*]; in this paper’s convention, those Vandermonde rows are left eigenvectors (since (V*)^{-1} is Vandermonde), and Lemma D.2 then forbids a nonzero Koopman eigenvalue from having a trivial DMD mode, contradicting the counterexample . Thus the paper’s result stands; the model’s claimed counterexample arises from an eigenvector orientation mix-up.
Referee report (LaTeX)
\textbf{Recommendation:} minor revisions
\textbf{Journal Tier:} strong field
\textbf{Justification:}
The manuscript gives a rigorous, practically relevant account of mean subtraction and mode selection in DMD, and proves that, under sufficient sampling and Koopman-invariant observables, nontrivial DMD eigenvalues equal the Koopman spectrum. The arguments are clear and correct, though notation is dense. Clarifying the left/right eigenvector conventions for the companion operator would help prevent misunderstandings.