Let $\mathbf{X}$ be nonnegative irreducible matrix such that $\mathbf{X} \ne \mathbf{X}^T$. Let $\mathbf{p}(\mathbf{A})$ denote a right eigenvector corresponding to the Perron root of $\mathbf{A}$, $\rho(\mathbf{A})$. Similarly, $\mathbf{q}(\mathbf{A})$ denotes a left eigenvector corresponding to $\rho(\mathbf{A})$. (In the textbook I'm reading, left eigenvectors are column vectors corresponding to the right eigenvector of $\mathbf{A}^T$.)

A remark in the textbook seems to imply that as long as $\mathbf{X} \ne \mathbf{X} ^ T$, $\mathbf{p}(\mathbf{X}) \ne \alpha \mathbf{p}(\mathbf{X}^T)$ for any constant $\alpha$.

EDIT: As requested by @user1551, here is the exact remark from Fundamentals of Resource Allocation in Wireless Networks:

Finally, as $\mathbf{X}^T\in W_K(\mathbf{X})$ for any $\mathbf{X} \in X_K$, it follows from Theorem 1.14 that $\rho((1-\mu) \mathbf{X} + \mu\mathbf{X}^T)$ is a concave function on $\mu \in [0,1]$. If, in addition, $\mathbf{X}\ne \mathbf{X}^T$, we have $\mathbf{p}(\mathbf{X}) \ne \alpha \mathbf{p}(\mathbf{X^T})$ for any constant $\alpha$.

$X_K$ is the set of nonnegative irreducible $K$-by-$K$ matrices. For $\mathbf{X}\in X_K$, $W_K(\mathbf{X}) = \{\mathbf{Y}\in X_K : \mathbf{q}(Y)\circ \mathbf{p}(\mathbf{Y}) = \mathbf{q}(X)\circ \mathbf{p}(\mathbf{X}) \in \Pi_K^{+}\}$, where $\Pi_K^+$ is the standard simplex restricted to positive values.

In an effort to prove the remark, I started a proof by contradiction. Suppose that $\mathbf{p}(\mathbf{X}^T)\equiv \mathbf{q}(\mathbf{X})$ is also a right Perron eigenvector of $\mathbf{X}$. By definition of left and right Perron eigenvectors, $\mathbf{X}^T \mathbf{q}(\mathbf{X}) = \rho(\mathbf{X})\mathbf{q}(\mathbf{X})$ and $\mathbf{X}\mathbf{q}(\mathbf{X}) = \rho(\mathbf{X})\mathbf{q}(\mathbf{X})$.

Subtracting the two equations, I get $(\mathbf{X}-\mathbf{X}^T)\mathbf{q}(\mathbf{X}) = \mathbf{0}$, and I'm not sure how to proceed. The only conclusion I get from this equation is that $\mathbf{q}(\mathbf{X})$ is in the null space of $\mathbf{X} - \mathbf{X}^T$.

  • 1
    $\begingroup$ "A remark in the textbook seems to imply that...". What exactly does the remark say? $\endgroup$
    – user1551
    May 11, 2020 at 9:20
  • $\begingroup$ I included an edit to include the remark. @user8675309 provided a good counterexample. For doubly stochastic matrix, the $\rho$ function is linear on $\mu \in [0,1]$. $\endgroup$
    – owovrokfop
    May 11, 2020 at 9:33

2 Answers 2


Your contradiction approach will fail because the claim is not false. In terms of simple examples: consider any doubly stochastic matrix, e.g. a permutation matrix and in particular one that is not symmetric (I’d suggest the cyclic shift matrix).


Another argument for why the stated claim is false.

The space of $n \times n$ irreducible Perron matrices $X$ has dimension $n^2$ (if you want, restrict yourself to positive matrices).

Fix $\rho >0$ and $p$ a positive vector. The space of $n \times n$ irreducible Perron matrices $X$ such that $Xp = \rho p$ has dimension $n^2-n$ (we have n linear equations to check).

The space of $n \times n$ irreducible Perron matrices $X$ such that $Xp = \rho p$ and $p^t X = \rho p^t$ has dimension $n^2-2n+1 = (n-1)^2$ (we have $n$ more linear equations to check, the last one being redundant).

The space $n \times n$ irreducible symmetric Perron matrices $X$ such that $Xp = \rho p$ is included in the later, and has dimension $n(n+1)/2-n = n(n-1)/2$. If $n>2$, then $n(n-1)/2 < (n-1)^2$, so that there are non-symmetric Perron matrices with main eigenvector $p$ and main eigen-covector $p^t$.


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.