Given some arbitrary Toeplitz matrix, if I swap two rows, one of the eigenvalues change its sign. For example,
$$X =
\begin{bmatrix}
A & B & C \\
D & A & B \\
E & D & A
\end{bmatrix}$$
and
$$Y = \begin{bmatrix}
D & A & B \\
A & B & C \\
E & D & A
\end{bmatrix}$$
have the same eigenvalues up to sign. I can see this for small examples, but how would I go about proving this?
In particular, if I flip the matrix upside down or left-side-right, half (rounding down) of the eigenvalues flip sign, and the singular values are the same. (It becomes a Hankel matrix.)
Numerical demonstration with MATLAB code:
d = 5;
X = toeplitz(randn(d,1));
-0.8655 -0.1765 0.7914 -1.3320 -2.3299
-0.1765 -0.8655 -0.1765 0.7914 -1.3320
0.7914 -0.1765 -0.8655 -0.1765 0.7914
-1.3320 0.7914 -0.1765 -0.8655 -0.1765
-2.3299 -1.3320 0.7914 -0.1765 -0.8655
J = flipud(eye(d));
0 0 0 0 1
0 0 0 1 0
0 0 1 0 0
0 1 0 0 0
1 0 0 0 0
svd(X)'
4.0897 2.0381 1.8456 0.8649 0.8198
svd(X*J)'
4.0897 2.0381 1.8456 0.8649 0.8198
eig(X)'
-4.0897 -2.0381 -0.8649 0.8198 1.8456
eig(X*J)'
-4.0897 -1.8456 -0.8649 0.8198 2.0381
eig(J*X)'
-4.0897 -1.8456 -0.8649 0.8198 2.0381
EDIT: non-symmetric toeplitz example
Best Answer
What you claim is not necessarily true if you swap two rows arbitrarily. It is easy to generate a random counterexample by computer. However, if $J$ is the backward identity matrix (as shown in your MATLAB code), then the eigenvalues of $JA$ (or $AJ$, which is similar to $JA$) and $A$ are identical except that $\lfloor \frac n2\rfloor$ pairs of them have different signs.
The essential reason behind this phenomenon is that a symmetric Toeplitz matrix has an eigenbasis in which $\lceil \frac n2\rceil$ eigenvectors $v$ are "symmetric" (i.e. $Jv=v$) and the remaining $\lfloor \frac n2\rfloor$ are "skew-symmetric" ($Jv=-v$).
So, if $v$ is a symmetric eigenvector corresponding to an eigenvalue $\lambda$, we have $JAv=J(\lambda v)=\lambda Jv=\lambda v$, i.e. $\lambda$ is also an eigenvalue of $JA$.
In contrast, if $v$ is a skew-symmetric eigenvector corresponding to some eigenvalue $\lambda$, then $JAv=J(\lambda v)=\lambda Jv=-\lambda v$, i.e. the sign of the eigenvalue is flipped in $JA$.