3
$\begingroup$

In my studies of matrix analysis, particularly in positive semidefinite matrices, I have come across the following question:

Let $ I $ be the $ n \times n $ identity matrix, and let $ v $ be a length n real column vector, we are asked to prove there exists a positive $ \delta > 0 $ such that the matrix $ I - \delta vv^T $ is positive semi definite.

Now, I know for any positive $ \delta $ we know that $ I - \delta vv^T $ is a real symmetric matrix and thus its positive semi definiteness is equivalent to non-negativity of its eigenvalues, but how do I prove there exists a positive $ \delta > 0 $ such that the eigenvalues of $ I - \delta vv^T $ are non negative and I have no real idea on how to do this, or even if this is the right approach, I would certainly appreciate any help on this, I thank all helpers.

$\endgroup$
4
$\begingroup$

Let $A:=I-\delta vv^T$. Complete $v_1:=v$ to an orthonormal basis $v_1,\ldots,v_n$. Note that for $k>1$, $$ A v_k = v_k - \delta v_0v_0^Tv_k = v_k $$ because $v_0^Tv_k=0$. So $A$ has $(n-1)$ eigenvectors with eigenvalue $1$. You know that $\chi_A(t) = (t-1)^{n-1}(t-\lambda)$ is the characteristic polynomial of $A$, where $\lambda$ is some real number. You know that $\operatorname{trace}(vv^T)=\sum_i v_i^2 = v^Tv$ and because the trace of a matrix is the sum of its eigenvalues, $$ \lambda+n-1=\operatorname{trace}(A)=\operatorname{trace}(I-\delta vv^T)=\operatorname{trace}(I)+\operatorname{trace}(\delta vv^T)=n-\delta v^Tv. $$ This implies $\lambda = 1- \delta v^T v = 1 - \delta \cdot \|v\|^2$. If you chose $$\delta<\frac1{\|v\|^2},$$ then $\lambda > 0$.

$\endgroup$
4
$\begingroup$

@Jesko Hüttenhain There is a slightly more direct explanation with a precise bound for $\delta$.

Rank-one matrix $M:=vv^t$ has the following eigenvalues (see explanation below):

  • $\lambda_1=\|v\|^2$ (multiplicity 1)

  • $\lambda_k=0$ ($k=2,3,\cdots n$, i.e., multiplicity $(n-1)$).

Then the eigenvalues of $f(M)=I-\delta M=I-\delta vv^T$ are the $f(\lambda_k)$, i.e.,

  • $1-\delta \|v\|^2$ (multiplicity 1).

  • $1 -\delta 0 = 1$ (multiplicity $(n-1)$).

It suffices thus to ensure that $1-\delta \|v\|^2>0$, i.e.,

$$\delta<\dfrac{1}{\|v\|^2}$$

Explanation about eigenvalues :

$Mv=(vv^t)v=v(v^tv)=v\|v\|^2=\|v\|^2v$ proving that $\|v\|^2$ is an eigenvalue associated with eigenvector $v$. As $M$ is rank 1, all the other eigenvalues are $0$.

$\endgroup$
  • $\begingroup$ You mean "are equal to $1-\delta \lambda_i$ ? Just because we can exhibate an associated eigenvector. I take an example : if $A$ has an eigenvalue/eigenvector pair $(\lambda,V)$, i.e. $AV=\lambda V$, then $I-4A$ has the folllowing eigenpair: $(1-4\lambda,V)$ because $(A-4I)V=(\lambda-4)V$. $\endgroup$ – Jean Marie Aug 31 '16 at 15:37
  • $\begingroup$ Alright, I got it now ;) $\endgroup$ – Jesko Hüttenhain Aug 31 '16 at 15:41

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.