Let $A$ be an $n \times n$ real symmetric matrix with row and column sums zero. For example, $$ A=\begin{bmatrix}1 & -2 & 1\\ -2 & 1 & 1\\ 1 & 1 & -2 \end{bmatrix}. $$ I have the following interesting observation about $A$ in general.

Claim: Suppose $\mathrm{rank}(A)=n-1$, and let $v_1, v_2,\dots ,v_{n-1}$ be the $n-1$ normalized eigenvectors (with unit length) corresponding to the $n-1$ nonzero eigenvalues. Let $\mathbf{V}=[v_1,\dots,v_{n-1}]$ be an $n\times(n-1)$ matrix of which each column $i$ is the eigenvector $v_i$. We have $$ I-\mathbf{V}\mathbf{V}^{T}=\frac{1}{n}\begin{bmatrix}1 & \dots & 1\\ \vdots & \vdots & \vdots\\ 1 & \dots & 1 \end{bmatrix}, $$ where $I$ is the identity matrix, and $n$ is the number of columns.

As for our particular $A$ in the display, we have $$ \mathbf{V}=\begin{bmatrix}\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{6}}\\ -\frac{1}{\sqrt{2}} & \frac{1}{\sqrt{6}}\\ 0 & -\frac{2}{\sqrt{6}} \end{bmatrix}. $$ One can easily verify the above claim for this example. I have randomly generated many such matrices, and the claim holds. So it might be correct.

My question is how to prove it. After spending many hours, I have made little progress so far. The only thing meaningful I have found is that any eigenvector of $A$ must sum to be zero, because $0=1^TAv=\lambda 1^Tv$. Here $\{\lambda, v\}$ denotes a generic pair of eigenvalue and eigenvector. An additional observation is that all cofactors of $A$ are identical. But these observations are far from enough to understand this claim. Any thought is welcomed. Thanks.

  • $\begingroup$ If A has eigenspaces of dimension greater than 1 we'd better use an orthonormal basis. $\endgroup$
    – blue
    Aug 27, 2014 at 0:24

1 Answer 1


The key point is that row and column sums being zero (equivalent for a symmetric matrix), just means the final eigenvector is $$\frac{1}{\sqrt{n}}\begin{pmatrix} 1 \\ \vdots \\ 1 \end{pmatrix}$$, call this vector $w$ Let $W$ be the $n \times n$ matrix where we have adjoined this eigenvector $w$ as a final column. Your claim is then that in this case

$$ WW^T = I = VV^T + ww^T$$

Which follows just from the rules of matrix multiplication, as pointed out in the comments.

  • $\begingroup$ At this point what you're claiming is that W is orthogonal because $W W^T = I$ iff $W^T W = I$. But this is obvious since we normalized the eigenvectors and eigenvectors of symmetric matrices with respect to different eigenvalues are orthogonal. $\endgroup$
    – blue
    Aug 27, 2014 at 0:19
  • $\begingroup$ @blue of course, I was invoking the finite dimensional real spectral theorem to get $W$ orthogonal, the real meat of the claim is that $VV^T + ww^T$ is the whole product. $\endgroup$
    – jxnh
    Aug 27, 2014 at 1:57
  • $\begingroup$ @JHance I think you have already proved my claim. Given $WW^T = VV^T + ww^T$, we only need to argue $WW^T=W^TW$ for $W^TW=I$. Since $W$ is orthogonal, W commutes with its transpose, that is $WW^T=W^TW$. $\endgroup$
    – semibruin
    Aug 27, 2014 at 5:13
  • 1
    $\begingroup$ @JHance The equation $WW^T = VV^T + ww^T$ follows from elementary matrix multiplication: Every entry of LHS is a dot product of entries of $v_1$ to $v_{n-1}$ plus $1/n$. That the claim is equivalent to $VV^T+ww^T = I$ is also elementary. I don't see any other gaps. $\endgroup$
    – blue
    Aug 28, 2014 at 11:41

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.