Some Extensions of the Crouzeix-Palencia Result

Trevor C. Caldwell, Anne Greenbaum, Kenan LiSIAM J. Matrix Anal. Appl 2018 · DOI

TL;DR

This paper extends a recent result showing that the numerical range of a matrix is a $(1+\sqrt{2})$-spectral set. The authors demonstrate that other regions in the complex plane can also be spectral sets, sometimes with constants close to $1+\sqrt{2}$, and identify specific cases where the constant can be improved to 2. This work deepens our understanding of the relationship between a matrix's spectrum, its numerical range, and the behavior of analytic functions applied to matrices.

What Problem Does This Paper Address?

The paper builds upon the work of Crouzeix and Palencia, who proved that the closure of the numerical range of a matrix $A$, denoted $W(A)$, is a $(1+\sqrt{2})$-spectral set. This means that for any function $f$ analytic inside $W(A)$ and continuous on its boundary, the inequality $||f(A)|| \le (1+\sqrt{2})||f||_{W(A)}$ holds, where $|| \cdot ||$ is the operator 2-norm and $|| \cdot ||_{W(A)}$ is the supremum of $|f(z)|$ over $z \in W(A)$. Crouzeix conjectured that the constant could be improved to 2.

This paper addresses two main questions:

  1. Can this spectral set property be extended to regions $\Omega$ in the complex plane that do not necessarily contain $W(A)$ but do contain the spectrum of $A$?
  2. Are there specific cases where the constant $(1+\sqrt{2})$ can be replaced by 2, as conjectured by Crouzeix?

The authors aim to generalize the techniques used by Crouzeix and Palencia to explore these questions, providing both theoretical extensions and numerical investigations.

Methods

The core of the paper's methodology involves extending the arguments from Crouzeix and Palencia's work [5]. This primarily relies on:

  • Cauchy Integral Formula: Functions of matrices are represented using the Cauchy integral formula over the boundary of a region $\Omega$.
  • Operator Analysis: The analysis involves studying properties of operators derived from the resolvent of $A$ and functions of $A$. Specifically, they define a Hermitian operator $\mu(\sigma, A)$ and its modified version $M(\sigma, A)$ to analyze spectral properties.
  • Lemmas and Theorems: The paper establishes several key lemmas and theorems. Lemma 2.1 provides an upper bound for an operator $S = f(A) + g(A)^* + \gamma I$ related to the spectral properties of $M(\sigma, A)$. Theorem 2.2 provides an upper bound for the spectral set constant $c_\Omega$ for convex regions $\Omega$ containing the spectrum of $A$.
  • Numerical Studies: The authors conduct numerical experiments to investigate the behavior of $A_{min}(\mu(\sigma, A))$ for points $\sigma$ inside and outside $W(A)$, particularly for circular regions around the spectrum. This helps to estimate the spectral set constants for various regions.
  • Optimal Blaschke Products: The paper utilizes properties of optimal Blaschke products (Theorem 5.1) which are crucial for constructing functions that achieve the maximum norm ratio $||f(A)||/||f||_\Omega$.

The approach is largely analytical, building upon existing spectral theory and functional calculus for operators, augmented by numerical exploration to gain insights and test hypotheses.

Datasets

Not discussed in the paper. The paper focuses on theoretical results and numerical experiments involving matrices, rather than datasets in the conventional sense. The "data" consists of matrices generated for numerical studies (e.g., random complex upper triangular matrices, perturbed Jordan blocks).

Key Insights and Ideas

  • Generalization of Spectral Sets: The paper shows that regions $\Omega$ other than the numerical range $W(A)$ can also be K-spectral sets for $A$, even if they don't contain $W(A)$, as long as they contain the spectrum of $A$. This broadens the applicability of spectral set theory.
  • Constant Near $1+\sqrt{2}$: For these generalized regions, the constant $K$ can be close to $1+\sqrt{2}$, suggesting that the original result is not an isolated phenomenon.
  • Cases for Constant 2: The authors identify specific scenarios where the spectral set constant can be proven to be 2, aligning with Crouzeix's conjecture. This is a significant step towards proving the conjecture in its full generality.
  • Sharpness of Bounds: The paper demonstrates a case (a 3x3 Jordan block) where the derived bound for spectral sets is sharp, validating the theoretical framework and providing a benchmark for future research.
  • Behavior of $A_{min}(\mu(\sigma, A))$: Numerical studies reveal that $A_{min}(\mu(\sigma, A))$ decreases rapidly as $\sigma$ moves towards the spectrum of $A$ and increases slowly as $\sigma$ moves away from it. This provides intuition about how spectral set constants change with region geometry.

Key Results

  • Theorem 2.2: For a convex domain $\Omega$ with a smooth boundary containing the spectrum of $A$, the spectral set constant $c_\Omega$ is bounded by $1 + \frac{1}{2} + \sqrt{2} + \delta + \delta^2/4 + \gamma$, where $\delta$ and $\gamma$ are defined based on the minimum eigenvalue of $\mu(\sigma, A)$ on the boundary of $\Omega$.
  • Disk as a Spectral Set: If $\Omega$ is a closed disk containing the spectrum of $A$, then $\Omega$ is a 2-spectral set for $A$ (i.e., $c_\Omega \le 2$). This is a new proof of a known result.
  • Sharpness Example: For a 3x3 Jordan block with eigenvalue 0 and a disk $\Omega$ of radius $r < 1$ centered at the origin, the bound $\max_{f \in A(\Omega)} ||f(A)||/||f||_\Omega \le 2 + \delta$ is shown to be sharp, with the value being $1/r^2$.
  • Numerical Findings: Table 1 and Table 2 present numerical results for various matrices and disks, showing that the derived bounds ($K_\delta$) are often better than those obtained from the Cauchy integral formula ($K_{Cauchy}$). For example, for a 12x12 random upper triangular matrix, $W(A)$ is shown to be a 2.0113-spectral set.

Limitations and Future Work

  • General Proof of Crouzeix's Conjecture: While the paper provides cases where the constant is 2 and numerical evidence supporting the conjecture, a general proof for all matrices remains an open problem. The authors pose the question of whether $||f(A)|| \le ||f(A) + g(A)^*||$ holds for the optimal function $f$, which would prove the conjecture.
  • Analytic Formula for Optimal Blaschke Product Roots: The paper notes that an analytic formula for the roots of the optimal Blaschke product is not known, and determining other properties of this optimal product is an interesting open problem.
  • Theoretical Bounds for Numerical Results: The paper suggests that it is an open question whether the numerically obtained bounds for spectral sets (like the 2.0113 for the 12x12 matrix) can be proven theoretically.
  • Negative Values of $\delta$: The paper mentions that $\delta$ can take negative values, which is relevant for regions that contain $W(A)$ but are not equal to it. Further exploration of these cases could be fruitful.

Summary updated

BibTeX
@article{caldwell2018some,
  author = {Trevor C. Caldwell and Anne Greenbaum and Kenan Li},
  title = {{Some Extensions of the Crouzeix-Palencia Result}},
  journal = {SIAM J. Matrix Anal. Appl},
  year = {2018},
  doi = {10.1137/17m1140832},
}

Loading related content...

Figure