Take a matrix $A \in \mathbb{R}^{n \times n}$, and define the spectral radius $\rho(A) = |\lambda_{\text{max}}(A)|$. $A$ is stable if $\rho(A) < 1$, or since it is relevant for my particular application, I'll call it $\gamma$-stable if for $\gamma \in [0, 1)$, $\rho(A) \le \gamma$.

The set of all $\gamma$-stable matrices $S = \{A\ |\ \rho(A) \le \gamma\}$, is not convex (unless $A$ is symmetric). I am interested in the convex hull of $S$, and have conjectured that $\textbf{conv}\ S = \{A\ |\ ||A|| \le \gamma\} \overset{\Delta}{=} C$. Since $\rho(A) \le ||A||$, and $C$ is convex, it is certainly true that $\textbf{conv}\ S \subseteq C$, but I'm unsure of the other direction.

EDIT: I confused the inequality, $C \subseteq S$, not the other way around. The question still stands, is there a known characterization of $\mathbf{conv}\ S$?

Is there a known characterization of $\textbf{conv}\ S$? I have not been able to find one myself, and surprisingly, my searches are coming up empty handed.

  • $\begingroup$ it is certainly true that $\textbf{conv}\ S \subseteq C$: hmmm... $\endgroup$ – Omnomnomnom Jul 8 '17 at 17:42
  • $\begingroup$ I suspect that the convex hull of the nilpotent matrices alone yields all trace-zero matrices $\endgroup$ – Omnomnomnom Jul 8 '17 at 17:44
  • $\begingroup$ I would conjecture that $\operatorname{co} S = \{ A | \, |\operatorname{tr} A | \le n \gamma \}$. $\endgroup$ – copper.hat Jul 9 '17 at 21:30
  • $\begingroup$ @copper.hat It does not seem possible to write $\begin{bmatrix}2 & 0\\0 & 0\end{bmatrix}$ as a convex combination of $1$-stable matrices. Such matrices will necessarily have trace $2$, both eigenvalues $1$ and, hence, be triangular with identities on the main diagonal. $\endgroup$ – A.Γ. Jul 11 '17 at 20:44
  • 2
    $\begingroup$ @A.Γ.: The matrix you wrote is in the (real) closure of $\operatorname{co} S$, see the aside in the answer below. $\endgroup$ – copper.hat Jul 11 '17 at 23:59

The matrix $T_n = \begin{bmatrix} \gamma & n \\ 0 & \gamma \end{bmatrix}$ is $\gamma$-stable for all $n$ but $\|T_n\| \ge n$ for all $n$.

Hence $S \not\subset C$.

Aside: (Too cumbersome for a comment.)

If $M(t) = (1-2t)\begin{bmatrix} 2 &1 \\ -1 & 0 \end{bmatrix}+ t\begin{bmatrix} 0 &-{1 \over t} \\ 0 & 0 \end{bmatrix} + t\begin{bmatrix} 0 &0 \\ {1 \over t} & 0 \end{bmatrix} = \begin{bmatrix} (1-2t)2 &-2t \\ 2t & 0 \end{bmatrix}$, then $M(t) \in \operatorname{co} S$ for all $t \in (0,1)$, hence $\begin{bmatrix} 2 &0 \\ 0 & 0 \end{bmatrix} \in \overline{\operatorname{co}} S$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.