10

We will call a matrix positive matrix if all elements in the matrix are positive, and we will denote the largest eigenvalue with $\lambda_{\max}$, what is exist because of the Perron–Frobenius theorem.

Theorem. Let $A$ be a positive square matrix. Then the minimal row sum is a lower bound and the maximal row sum is an upper bound of $\lambda_{\max}$.

My questions.

  1. Is there a name for this theorem and can anybody say books or papers what refer to it?

  2. How to prove it?

Rodrigo de Azevedo
  • 23,223
user153012
  • 12,890
  • 5
  • 42
  • 115

2 Answers2

8

Let us denote by $r$ the Perron-root of the positive matrix $A \in \mathbb{R}^{n \times n}$. Then by the Collatz-Wielandt formula we have:

$$\max_{x \in S}\min_{\substack{i=1, \ldots,n}} \frac{(Ax)_i}{x_i} = r = \min_{x \in S}\max_{\substack{i=1, \ldots,n}} \frac{(Ax)_i}{x_i}, $$ where $S := \{x \in \mathbb{R}^n\setminus\{0\}: x_i > 0, \forall i=1,\ldots,n\}$. Now it is clear that $A$ and $A^T$ have same eigenvalues, since $\det(M)=\det(M^T)$ and for every $\lambda \in \mathbb{R}$ we have $$\det(A-\lambda I)=\det((A-\lambda I)^T)= \det(A^T-\lambda I).$$ Furthermore $A$ strictly positive implies $A^T$ strictly positive, thus this formula also holds for $A^T$. It follows that we have $$\max_{x \in S}\min_{\substack{i=1, \ldots,n}} \frac{(A^Tx)_i}{x_i} = r = \min_{x \in S}\max_{\substack{i=1, \ldots,n}} \frac{(A^Tx)_i}{x_i}.$$ This clearly implies that for every $y \in S$ we have $$\min_{\substack{i=1, \ldots,n}} \frac{(A^Ty)_i}{y_i} \leq r \leq \max_{\substack{i=1, \ldots,n}} \frac{(A^Ty)_i}{y_i}.$$ Choose $y = (1,1,\ldots,1)$ to get your bounds. Note also that using the same trick on $A$ directly you will get the same upper/lower bound but with the columns instead of the rows.

For reference, I recommend (in increasing order of technicality/generality):

  1. "Matrix analysis" by Horn and Johnson
  2. "Nonnegative Matrices in the Mathematical Sciences" by Bermann and Plemmons
  3. "Nonlinear Perron-Frobenius theory" by Nussbaum and Lemmens

There are even more general versions discussed in recent literature, e.g. in this paper

Surb
  • 57,262
  • 11
  • 68
  • 119
  • Consider $A^T=\begin{bmatrix} 3 & 1 & 1\ 4 & 1 & 1 \ 5 & 1 & 1 \end{bmatrix}$ and $y=(1, 0, 0)$. It can be checked that $r$ is greater than $5$, but $(A^Ty)_i/y_i=3$. The upper bound in the last inequality does not hold. – kswim Oct 27 '21 at 06:51
  • Please see item 7 of "positive matrices" in Section "Statements" in https://en.wikipedia.org/wiki/Perron–Frobenius_theorem – kswim Oct 27 '21 at 06:53
  • In the reference you mentioned (I think you would refer to "Matrix analysis and applied linear algebra" by Carl D. Meyer), as far as I could determine, the upper bounds are not stated. – kswim Oct 27 '21 at 06:56
  • Yes. But, in your answer, $S$ contains non-negative vectors. – kswim Oct 27 '21 at 07:16
  • @kswim looks like, 7 years ago, I did not realize the importance of the infimum (minimum), being taken on vectors in the interior of the cone. I corrected my answer, thanks for noting the mistake. – Surb Oct 27 '21 at 07:26
  • 1
    I found out that this led another questions about whether the formula is correct or not. So, I wanted to make things correct. Thanks for the correction! – kswim Oct 27 '21 at 07:34
2

See propositions 2.3 and 2.4 of the book "Banach Lattices and Positive operators" of H.H. Schaefer.

Daniel
  • 6,057
  • Can you maybe offer a good ebook version of it, or send me an image about the pages what you talking about? It could really help me. – user153012 Aug 26 '14 at 21:05
  • 1
    Or just write it here. – user153012 Aug 26 '14 at 21:21
  • 1
    See this link. http://books.google.com.br/books?id=GQ-oAAAAIAAJ&focus=searchwithinvolume&q=row+sums. I could not find a better look inside this book. I forgot to say that the page is 7. The number $r(A)$ is the spectral radius of $A$, which is the biggest positive eigenvalue for this type of matrix. – Daniel Aug 27 '14 at 01:25