0
$\begingroup$

Let $A$ be real square matrix.

Let $\mathcal{F}(A)$ be the set of real matrices $A'$ of the same size such that $A'_{ii}=A_{ii}$ for all $i$, and for all $i,j$, $A_{ij}=0\Rightarrow A'_{ij}=0 \land A_{i,j} \ne 0 \implies A'_{i,j} \ne 0$: in other words, obtained from $A$ by modifying nonzero nondiagonal entries.

Q1 What are sufficient conditions for $A$ s.t. exists a positive definite $A'$ in $\mathcal{F}(A)$?

Since there are several definitions of positive definite, we don't require neither $A$ or $A'$ to be symmetric.

In practice this is possible for some $A$.

Partial answers are welcome. The definition of positive definite from mathworld

$M$ is positive definite if for all nonzero real vectors $x$, $x^T M x >0$.

$\endgroup$
2
  • 1
    $\begingroup$ Please indicate what you mean by positive definite (I understand it means $X^tAX>0$ for every nonzero column vector $X$?) $\endgroup$
    – YCor
    Dec 31, 2015 at 11:02
  • $\begingroup$ @YCor Exactly, will edit soon, citing mathworld (wikipedia requires symmetric). $\endgroup$
    – joro
    Dec 31, 2015 at 11:20

1 Answer 1

3
$\begingroup$

The answer is quite boring... Necessary and sufficient conditions are $A_{ii}>0$ for each $i$. In this case, let $B_\varepsilon$ the matrix with entries $$ (B_\varepsilon)_{ij} = \begin{cases} A_{ij} & i=j,\\ \varepsilon & i\neq j, A_{ij}\neq 0,\\ 0 & i\neq j, A_{ij}=0. \end{cases} $$ The matrix $B_0$ is diagonal and posdef, so for any sufficiently small value of $\varepsilon$ the matrix $B_\varepsilon$ (which belongs to $\mathcal{F}(A)$) is posdef, too, since positive definiteness is an open condition.

On the other hand, if $A_{ii}\leq 0$ for some $i$ there is no hope, since for each $A'\in\mathcal{F}(A)$ we have $x^\top A'x \leq 0$ when $x$ is the vector $e_i$ of the canonical basis.

$\endgroup$
5
  • $\begingroup$ Thanks. Ycor misquoted my definition. I edited: $A_{i,j} \ne 0 \implies A'_{i,j} \ne 0$ (I had this in wording). Does your answer still applies? $\endgroup$
    – joro
    Dec 31, 2015 at 11:22
  • 1
    $\begingroup$ @Joro Yes. Just put epsilons instead of zeros in the entries where you need nonzeros. Any sufficiently small modification of $D$ is posdef, since posdefness is an open condition. $\endgroup$ Dec 31, 2015 at 11:26
  • $\begingroup$ Thank you. I suggest to edit, so I accept your answer (The title doesn't accept zeros and I had explicitly nonzero, but Ycor's edit missed that). $\endgroup$
    – joro
    Dec 31, 2015 at 11:37
  • $\begingroup$ I discarded the condition $A_{ij}\neq 0\Rightarrow A'_{ij}\neq 0$ because it does not affect the existence of a positive-definite matrix in $F(A)$, by the openness condition Federico mentioned. $\endgroup$
    – YCor
    Dec 31, 2015 at 11:41
  • $\begingroup$ Happy New year celebration! $\endgroup$
    – joro
    Dec 31, 2015 at 11:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.