Let $A_1$ be a $n \times n$ real symmetric matrix whose eigenvalues are $\lambda_1$, $\lambda_2, \lambda_3, \ldots , \lambda_n$.

Let's assume that if $i < j$, then $|\lambda_i| > |\lambda_j|$ (so, $\lambda_1$ is the dominant eigenvalue...)

Also, for any of these eigenvalues $\lambda_i$, define $v_i$ as being the unit associated eigenvector (So, $A_1v_i = \lambda_i v_i$ and $\|v_i\| = 1$)

Is it true that the matrix

$$A_2 = A_1 - A_1v_1v_1^T$$

has $\lambda_2$ as its dominant eigenvalue?

(So, we somehow eliminate the original dominant eigenvalue $\lambda_1$ with this subtraction).

And, in a more general form, is it true that

$$A_k = A_1 - \sum_{i=1}^{k-1}A_1v_iv_i^T$$

has $\lambda_k$ as its dominant eigenvalue?

I read this proposition on a paper about NIPALS and also on this Wikipedia's section, but both places don't present any proof.

I made a computer program to test it, and it seems to be true. Also, I discovered that when I use eigenvectors whose norm is not equal to 1, the test program fails, so, this hypothesis is needed.


As a consequence of the spectral theorem, any symmetric matrix of rank $r$ can be expressed in the form $$ A = \sum_{i=1}^r \lambda_i\,v_iv_i^T $$ where each $\lambda_i$ is a non-zero eigenvalue, and the vectors $v_i$ are chosen in any way such that each $v_i$ is an eigenvector of length $1$, and any pair $v_i,v_j$ are mutually orthogonal.

Starting from here, I think you'll find it easy to show that your statement is true.

  • $\begingroup$ The eigenvectors of a symmetric matrix are linear independent, right? But they are also orthogonal? $\endgroup$ Aug 13 '15 at 17:49
  • 1
    $\begingroup$ We can always choose the eigenvectors to be orthogonal for a symmetric matrix. If all of the eigenvalues are distinct, then no matter how you choose the eigenvectors, they will end up orthogonal. $\endgroup$ Aug 13 '15 at 17:50

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.