Some Regularity Results for the Pseudospectral Abscissa and Pseudospectral Radius of a Matrix

Mert Gürbüzbalaban, Michael L. OvertonSIAM J. Optim 2012 · DOI

TL;DR

This paper delves into the mathematical properties of a matrix's "pseudospectrum," a concept crucial for understanding the stability of dynamical systems, especially when matrices are subject to small perturbations. The authors prove that key measures derived from the pseudospectrum, namely the pseudospectral abscissa and radius, are well-behaved (Lipschitz continuous) and that the pseudospectrum itself doesn't have "sharp points" at these critical locations. This resolves a long-standing conjecture and provides stronger theoretical guarantees for analyzing perturbed systems.

What Problem Does This Paper Address?

When analyzing the stability of linear dynamical systems described by $\dot{x} = Ax$, the eigenvalues of the matrix $A$ are paramount. If all eigenvalues have negative real parts, the system is stable. However, real-world systems are rarely perfectly known; they are subject to uncertainties or perturbations. The standard eigenvalue analysis can be misleading if the matrix $A$ is "non-normal," meaning its eigenvectors don't form a complete basis. In such cases, small perturbations can lead to large changes in eigenvalues, making stability predictions unreliable.

The pseudospectrum $\Lambda_\varepsilon(A)$ offers a more robust way to assess stability. It's defined as the union of the spectra of all matrices $A+E$ that are "close" to $A$ (specifically, $\|E\| \le \varepsilon$ for some small $\varepsilon > 0$). This captures how eigenvalues might shift under uncertainty. Two important quantities derived from the pseudospectrum are:

  • Pseudospectral Abscissa ($\alpha_\varepsilon(A)$): The maximum real part of any point in $\Lambda_\varepsilon(A)$. This relates to the decay rate of solutions to $\dot{x} = Ax$ under perturbations.
  • Pseudospectral Radius ($\rho_\varepsilon(A)$): The maximum modulus of any point in $\Lambda_\varepsilon(A)$. This relates to the overall magnitude of solutions under perturbations.

While it's known that the pseudospectrum and these associated measures are continuous with respect to $A$ and $\varepsilon$, their regularity (how smoothly they change) is a more subtle issue. Specifically, Lewis and Pang [LP08] showed that $\alpha_\varepsilon$ and $\rho_\varepsilon$ are Lipschitz continuous (a strong form of smoothness) everywhere except at certain "resolvent-critical" points. They conjectured that the points where $\alpha_\varepsilon$ and $\rho_\varepsilon$ are attained are precisely not resolvent-critical, which would imply Lipschitz continuity everywhere. This paper aims to prove this conjecture and explore the implications for the geometric shape of the pseudospectrum.

Methods

The paper builds upon previous work by Lewis and Pang [LP08] and Alam et al. [ABBO10]. The core of their argument relies on:

  1. Resolvent-Critical Points: They first rigorously define what a "resolvent-critical" point is. A point $z$ is resolvent-critical for $A$ if it's either an eigenvalue of $A$ or a critical point of the norm of the resolvent $(A-zI)^{-1}$ in a generalized sense (using subdifferentials).
  2. Geometric Characterization of the Boundary: A key result (Theorem 3.1) shows that if a point $z$ on the boundary of the pseudospectrum $\partial\Lambda_\varepsilon(A)$ is resolvent-critical, then the pseudospectrum must contain $m \ge 2$ equally spaced circular sectors centered at $z$. This implies a "pointed" or "cusp-like" structure.
  3. Proving the Conjecture: The authors then argue that if the pseudospectrum has such $m \ge 2$ sectors, then the points where $\alpha_\varepsilon$ or $\rho_\varepsilon$ are attained cannot be resolvent-critical. This is because if they were, the pseudospectrum would have this problematic pointed structure, which they later show is impossible. This directly proves the conjecture.
  4. Aubin Property: They leverage their findings to establish the "Aubin property" for the pseudospectrum with respect to both the matrix $A$ and the perturbation parameter $\varepsilon$. This property is a formal way of saying that the pseudospectrum and its critical points change smoothly as $A$ or $\varepsilon$ vary.
  5. "Not Pointed Outwards": Finally, they provide a new proof for the result that the pseudospectrum cannot be "pointed outwards" at nondegenerate points, extending it to the resolvent-critical case using their main theorem.

Datasets

Not discussed in the paper. The paper focuses on theoretical analysis of matrices and their properties, not empirical evaluation on specific datasets.

Key Insights and Ideas

  • Resolvent-Criticality is the Key: The paper establishes that the "trouble spots" for the regularity of $\alpha_\varepsilon$ and $\rho_\varepsilon$ are precisely the resolvent-critical points. By showing that the points where these measures are attained are never resolvent-critical, they resolve the conjecture.
  • Geometric Implication of Resolvent-Criticality: The theorem linking resolvent-critical points to the presence of multiple circular sectors is a powerful geometric insight. It suggests that such points create a specific, potentially problematic, shape in the pseudospectrum.
  • Pseudospectrum is "Smooth" at Critical Points: The main result implies that the pseudospectrum, while potentially complex, doesn't have sharp "points" at the locations that determine $\alpha_\varepsilon$ and $\rho_\varepsilon$. This is crucial for reliable numerical analysis and stability predictions.
  • Aubin Property for Robustness: Establishing the Aubin property confirms that the pseudospectral abscissa and radius, and the critical points themselves, behave predictably and smoothly under small changes in the matrix or the perturbation level $\varepsilon$.

Key Results

  • Conjecture Proven: The paper proves the conjecture that the points where the pseudospectral abscissa $\alpha_\varepsilon(A)$ and pseudospectral radius $\rho_\varepsilon(A)$ are attained are not resolvent-critical.
  • Lipschitz Continuity: As a consequence, $\alpha_\varepsilon$ and $\rho_\varepsilon$ are locally Lipschitz continuous for any matrix $A$ and any $\varepsilon > 0$.
  • Aubin Property: The pseudospectrum $\Lambda_\varepsilon(A)$ exhibits the Aubin property with respect to both the matrix $A$ and the perturbation parameter $\varepsilon$ at points where $\alpha_\varepsilon$ or $\rho_\varepsilon$ are attained.
  • No "Pointed Outwards" Pseudospectra: The paper provides a new proof that the pseudospectrum cannot be "pointed outwards" at critical points.

Limitations and Future Work

  • Conjecture on $m=2$: The authors conjecture that the integer $m$ in Theorem 3.1 (related to the number of sectors around a resolvent-critical point) is always 2. This is supported by examples but not formally proven in this paper. Further investigation into this specific geometric configuration could be fruitful.
  • Numerical Implications: While the theoretical results are strong, exploring how these regularity properties translate into more robust numerical algorithms for computing pseudospectra and related quantities would be a valuable next step.
  • Generalization to Other Norms: The paper focuses on the spectral norm ($\ell_2$ norm). Investigating similar regularity properties for pseudospectra defined using other matrix norms could extend the applicability of these results.

Summary updated

BibTeX
@article{gurbuzbalaban2012some,
  author = {Mert Gürbüzbalaban and Michael L. Overton},
  title = {{Some Regularity Results for the Pseudospectral Abscissa and Pseudospectral Radius of a Matrix}},
  journal = {SIAM J. Optim.},
  year = {2012},
  doi = {10.1137/110822840},
}

Loading related content...

Figure