5
$\begingroup$

Let $A \in \mathbb{C}^{n\times n}$ be a complex matrix. We let $a_{i,j}$ be the $\left(i,j\right)$-th entry of $A$ for all $i, j \in \left[n\right]$ (where $\left[n\right]$ denotes $\left\{1,2,\ldots,n\right\}$). For each $i \in \left[n\right]$, we define the $i$-th Gershgorin disk of $A$ to be the closed disk \begin{align} D_i\left(A\right) := \left\{z \in \mathbb{C} : \left|z-a_{i,i}\right| \leq \sum_{j \neq i} \left|a_{i,j}\right| \right\}. \end{align} (Here, the summation sign $\sum\limits_{j \neq i}$ means a sum over all $j \in \left[n\right]$ satisfying $j \neq i$.)

The following is well-known (see the Wikipedia for an easy proof):

Theorem 1 (Gershgorin's first theorem). Each eigenvalue of $A$ belongs to at least one of the $n$ disks $D_1\left(A\right), D_2\left(A\right), \ldots, D_n\left(A\right)$.

As has already been observed here on MO, this does not mean that there each of these $n$ disks contains at least one eigenvalue of $A$; there are counterexamples even for $n = 2$.

However, something in this direction does hold when the $n$ disks can be split into two sets that have no overlap:

Theorem 2 (Gershgorin's second theorem). Let $k$ be an integer with $0 \leq k \leq n$. Assume that $D_i\left(A\right) \cap D_j\left(A\right) = \varnothing$ for all $i$ and $j$ with $i \leq k < j$. Then, there are exactly $k$ eigenvalues of $A$ that belong to the first $k$ Gershgorin disks $D_1\left(A\right), D_2\left(A\right), \ldots, D_k\left(A\right)$, and there are exactly $n-k$ eigenvalues of $A$ that belong to the remaining $n-k$ Gershgorin disks $D_{k+1}\left(A\right), D_{k+2}\left(A\right), \ldots, D_n\left(A\right)$. (Here, eigenvalues are counted with their algebraic multiplicities, so there are always $n$ of them in total.)

Unlike Theorem 1, this is not trivial at all. Gershgorin's original proof (Theorem 2 is Satz III in his 1931 paper) uses a not-very-rigorous continuity argument. The idea is nice: We let $B$ be the diagonal $n\times n$-matrix whose diagonal entries are those of $A$. Consider the eigenvalues of the matrix $\left(1-t\right)B + tA$ for each $t \in \left[0,1\right]$. For $t = 0$, these eigenvalues are the $a_{i,i}$, whereas for $t = 1$ they are the eigenvalues of $A$. Now, it is intuitively clear but nontrivial to formalize and prove that the eigenvalues of a matrix depend continuously on the entries of the matrix; thus, when $t$ progresses from $0$ to $1$, the eigenvalues cannot "jump". However, by Theorem 1, these eigenvalues are always in the union $D_1\left(\left(1-t\right)B + tA\right) \cup D_2\left(\left(1-t\right)B + tA\right) \cup \cdots \cup D_n\left(\left(1-t\right)B + tA\right)$, which in turn is easily seen to be a subset of the union $D_1\left(A\right) \cup D_2\left(A\right) \cup \cdots \cup D_n\left(A\right)$. Since the latter union decomposes into the two disjoint closed subsets $D_1\left(A\right) \cup D_2\left(A\right) \cup \cdots \cup D_k\left(A\right)$ and $D_{k+1}\left(A\right) \cup D_{k+2}\left(A\right) \cup \cdots \cup D_n\left(A\right)$, and since we know how many eigenvalues of $\left(1-t\right)B + tA$ lie in which of these two subsets for $t = 0$ (namely, $k$ eigenvalues lie in the former, and $n-k$ eigenvalues in the latter), we thus conclude that the same numbers of eigenvalues are in these subsets for $t = 1$, that is, for the matrix $A$.

Horn and Johnson, in their Matrix analysis (2nd edition 2003) (proof of Theorem 6.1.1), suggest a variant of this argument. They, too, consider the matrix $\left(1-t\right)B + tA$, but instead of looking at specific eigenvalues, they use the argument principle from complex analysis to represent the number of eigenvalues of $\left(1-t\right)B + tA$ that lie within the union $D_1\left(A\right) \cup D_2\left(A\right) \cup \cdots \cup D_k\left(A\right)$ (or, rather, within a rectifiable curve that surrounds this union) as a meromorphic integral. The continuity of this integral, I assume, follows from some basic complex analysis. However, this too leaves something to be desired from a formalistic point of view: Why do we know that we can surround $D_1\left(A\right) \cup D_2\left(A\right) \cup \cdots \cup D_k\left(A\right)$ by a rectifiable curve (that does not contain any of the other disks in its interior)?

Thus, my...

Question: Is there a proof of Theorem 2 that avoids any analytic technicalities, or perhaps even any analysis whatsoever?

Of course, the very fact that $A$ has $n$ eigenvalues requires analysis (as it depends on the Fundamental Theorem of Algebra). However, we can take it for an assumption. Even better, Theorem 2 could be restated as "no more than $k+1$ eigenvalues of $A$ can lie within $D_1\left(A\right) \cup D_2\left(A\right) \cup \cdots \cup D_k\left(A\right)$", which looks like a perfectly elementary claim susceptible to attack using (e.g.) the triangle inequality just like Theorem 1 is commonly proved. It seems natural to start by observing that if you have $k+1$ eigenvalues of $A$, you can find an $A$-invariant subspace $W$ of $\mathbb{C}^n$ such that these $k+1$ eigenvalues are precisely the eigenvalues of the restriction $A\mid_W : W \to W$. Unfortunately, it is not obvious to me that what can be said about the Gershgorin disks of $A\mid_W$; this clearly depends on the right choice of basis for $W$. Maybe an echelon basis?

$\endgroup$
13
  • $\begingroup$ Does the following attempt to rigorize the continuity argument work? You have a "set of eigenvalues" map of algebraic varieties $\mathbb{A}^1_{\mathbb{C}}\rightarrow\operatorname{Sym}^n(\mathbb{A}^1_{\mathbb{C}})$ which sends $t$ to the set of eigenvalues of $(1-t)B+tA$. Restricting to $[0,1]$ gives you a continuous map of manifolds $[0,1]\rightarrow\operatorname{Sym}^n(\mathbb{A}^1_{\mathbb{C}}).$ As you note, in this case the image lies in $\operatorname{Sym}^n$ of the union of disks, which decomposes into connected components indexed by how many points lie in each part. $\endgroup$
    – dhy
    Sep 10 at 17:25
  • $\begingroup$ @dhy I'm not quite sure how the topology on the Sym is defined (is it just the quotient topology?) and why the eigenvalues map is continuous. But the impetus behind the question is to have a proof accessible to 1st year grad students in a course I'm teaching... so is rather avoid any topology that could not be explained in 20 minutes. $\endgroup$ Sep 10 at 17:40
  • $\begingroup$ It is the quotient topology. Alternatively, there is an isomorphism $\operatorname{Sym}^n(\mathbb{A}^1)\cong\mathbb{A}^n$ (this is the statement that elementary symmetric polynomials generate all symmetric polynomials) and so the map $\mathbb{A}^1\rightarrow\mathbb{A}^n$ comes from taking coefficients of the characteristic polynomial and so is easily seen to be continuous. $\endgroup$
    – dhy
    Sep 10 at 18:32
  • $\begingroup$ Phrased like this the crux is to show that the loci in $\mathbb{A}^n$ corresponding to having eigenvalues in $D_1\cup D_2\cdots\cup D_n$ is disconnected and has the correct connected components. This can be seen by looking at the preimage of this loci under the map $(\mathbb{A}^1)^n\rightarrow\mathbb{A}^n.$ $\endgroup$
    – dhy
    Sep 10 at 18:34
  • $\begingroup$ * to have a proof accessible to 1st year grad students* Hmmm... You can easily make the original Gershgorin proof completely rigorous if the students know that a) A finite union of disks is compact, b) Every continuous positive function on a compact set has positive minimum, c) Rouche theorem for the disk (no need for fancy curves) and d) that a locally constant function on $[0,1]$ is constant. Which of these facts is unavailable to you? The algebraic proof, even if it exists, will most likely be quite contrived and the homotopy argument is too useful not to expose your students to it :-) $\endgroup$
    – fedja
    Sep 12 at 2:39

Your Answer

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

Browse other questions tagged or ask your own question.