Henrici Departure-from-Normality

• Henrici’s departure-from-normality is a scalar index that measures the geometric distortion in directed Laplacians by comparing the Frobenius norm with the eigenvalue sum.
• It is computed using key metrics that reveal non-orthogonality in eigenvectors, directly impacting spectral stability and the reliability of associated algorithms.
• Empirical evidence shows higher Henrici values correlate with increased instability, making the index essential for assessing robustness in directed graph signal processing.

Henrici’s departure-from-normality is a scalar index that quantifies the extent to which a square matrix, particularly a directed Laplacian, fails to be normal. For Laplacians of directed graphs, non-normality yields non-orthogonal eigenvectors, challenging the classical spectral tools and stability analyses applicable to undirected (normal) graph Laplacians. Henrici’s departure is constructed from the deficit between the Frobenius norm of a matrix and the $\ell_2$ sum of its eigenvalues, providing a concise, interpretable measure of geometric distortion induced by non-normality. The index serves as a critical parameter for assessing robustness and reliability in spectral analysis, filtering, reconstruction, and sampling on non-normal digraphs (Gokavarapu, 1 Jan 2026).

1. Formal Definitions and Related Indices

Henrici’s departure-from-normality is one of three quantitative measures introduced for expressing non-normality in a matrix $M \in \mathbb{C}^{n \times n}$:

• Asymmetry index:

$$\alpha(M) = \frac{\| M - M^\top \|_F}{\|M\|_F}, \quad \alpha(0) = 0$$

• Commutator-based departure:

$$\nu(M) = \frac{\| M\,M^* - M^*\,M \|_F}{\|M\|_F^2}, \quad \nu(0) = 0$$

• Henrici’s Frobenius-norm departure:

$$\Delta(M) = \sqrt{\,\|M\|_F^2 - \sum_{k=1}^n |\lambda_k|^2\,}$$

where $\lambda_k$ are the eigenvalues of $M$. For normal matrices, $\Delta(M) = 0$. Thus $\Delta(M)$ provides a rigorous, basis-invariant measure that vanishes precisely for normal matrices and grows with the degree of non-normality.

2. Calculation for Directed Laplacians

For a weighted digraph $G$ with out-degree matrix $D_{\mathrm{out}}$ and adjacency matrix $A$, the combinatorial directed Laplacian is given by

$L = D_{\mathrm{out}} - A.$

The computation of $\Delta(L)$ involves the following steps:

1. Compute the Frobenius norm: $\|L\|_F = \sqrt{ \sum_{i,j} |L_{ij}|^2 }$.
2. Compute all eigenvalues $\lambda_1, ..., \lambda_n$ of $L$.
3. Form the sum $S = \sum_{k=1}^n |\lambda_k|^2$.
4. Compute the departure:

$\Delta(L) = \sqrt{\, \|L\|_F^2 - S \,}$

Equivalently, using the singular values $\sigma_i(L)$ of $L$,

$$\Delta(L) = \sqrt{ \sum_i \sigma_i^2 - \sum_k | \lambda_k |^2 }$$

3. Theoretical Role and Connections to Stability

While the majority of analytic stability bounds in signal processing on directed graphs are formulated in terms of the condition number $\kappa(V) = \sigma_{\max}(V)/\sigma_{\min}(V)$ of the right-eigenvector matrix $V$, Henrici’s departure $\Delta(L)$ is recognized as a complementary, informative index of non-normality.

Specifically, both $\kappa(V)$ and $\Delta(L)$ act as intrinsic difficulty indices: large values indicate increased instability in spectral filtering, amplification of coefficient-to-signal perturbations, and heightened geometric distortion. For example, the coefficient-to-signal amplification theorem quantifies:

$$\frac{ \| V(\widehat{x} + \eta) - V\widehat{x} \|_2 }{ \|x\|_2 } \leq \kappa(V) \frac{ \|\eta\|_2 }{ \| \widehat{x}\|_2 }$$

Experimental evidence shows that $\Delta(L)$ and $\kappa(V)$ typically increase together under non-normal perturbations, so $\Delta(L)$ also reliably predicts instability in practice (Gokavarapu, 1 Jan 2026).

Moreover, energy and variation distortion bounds rely on the extremal singular values of $V$, which are empirically correlated with $\Delta(L)$. Specifically,

$$\sigma_{\min}^2(V)\|\widehat{x}\|_2^2 \le \|x\|_2^2 \le \sigma_{\max}^2(V)\|\widehat{x}\|_2^2,$$

and,

$$\sigma_{\min}^2(V)\sum_k|\lambda_k|^2|\widehat{x}_k|^2 \le \|Lx\|_2^2 \le \sigma_{\max}^2(V)\sum_k|\lambda_k|^2|\widehat{x}_k|^2.$$

Thus, large $\kappa(V)$ and, correspondingly, large $\Delta(L)$ amplify energy and smoothness distortions.

4. Empirical Illustration: Normal Versus Non-Normal Digraphs

The performance of $\Delta(L)$ as a non-normality indicator is illustrated by experiments on $n=20$ node digraphs:

Graph	$\kappa(V)$	$\Delta(L)$	$\alpha(L)$	$\nu(L)$
Directed cycle	1	0	1	0
Perturbed cycle	16.80	5.98	0.49	0.065

A directed cycle, which is normal, yields $\kappa(V) = 1$, $\Delta(L) = 0$. A randomly perturbed cycle (with additional edges), which is non-normal, yields $\kappa(V) \approx 16.80$, $\Delta(L) \approx 5.98$. Reconstruction error versus input noise for a 5-bandlimited signal shows a pronounced amplification of noise in the perturbed (non-normal) case, in line with the increase in both indices.

5. Impact on Directed Graph Signal Processing

Henrici’s departure $\Delta(L)$ emerges as a practical “trust metric” in spectral methods on digraphs. Its vanishing characterizes cases where spectral tools and diagonal filtering in the Laplacian eigenbasis remain reliable. As $\Delta(L)$ increases, non-normal effects—such as the emergence of pseudospectra and operator ill-conditioning—dominate. In these cases, the standard spectral framework becomes less reliable, and more robust operator choices or regularization may be required.

Large $\Delta(L)$ correlates with:

• Enhanced distortion in energy and smoothness in the BGFT domain.
• Increased sampling-set instability and noise sensitivity.
• Amplified coefficient-to-signal ratios in filtering and reconstruction.

The index is thus essential for a priori assessment of algorithmic stability and for the design of stable directed spectral methods.

6. Summary and Implications

Henrici’s departure-from-normality, $\Delta(L)$, provides a rigorous, efficiently computable measure of the geometric distortion induced by the non-normality of directed Laplacians. It complements the condition number of the eigenvector basis in diagnosing the intrinsic difficulty of spectral analysis and filtering for digraphs. Its role as a stability and trust metric is empirically validated and theoretically motivated in the context of signal energy preservation, reconstruction, and resilience to noise, making it an indispensable diagnostic in harmonic analysis on directed networks (Gokavarapu, 1 Jan 2026).