2
$\begingroup$

Suppose we have positive-definite matrices $A$, $B$, if $A>B>0$, can we always find a positive real number $\alpha$, $0<\alpha < 1$ such that $ \alpha A \geq B $? If it has, then what is it?

$\endgroup$
5
  • $\begingroup$ $A>B$ means $A-B$ is positive-definite matrix. $\endgroup$
    – wayne
    Jun 7, 2015 at 3:07
  • $\begingroup$ Naturally. Thank you. Can't you prove a little sharper result: $\ \alpha\cdot A\ >\ B\ $ ? $\endgroup$ Jun 7, 2015 at 3:08
  • $\begingroup$ If you mean my proposition is always true? $\endgroup$
    – wayne
    Jun 7, 2015 at 3:10
  • 2
    $\begingroup$ The real part the relevant expression related to $\ A - B\ $ is greater than certain $\ \epsilon > 0\ $ when evaluated on the unit sphere. If you change $\ A-B\ $ a little, i.e. if you change $\ A\ $ a little then it'll be fine. $\endgroup$ Jun 7, 2015 at 3:29
  • $\begingroup$ Thanks for you suggestion, after following your suggestion, I arrange it as follows: $A-B>\epsilon I$ which is equivalent to $A-\epsilon I> B$, also $(I-\epsilon A^{-1}) A> B$, but this still can't show $\alpha A>B$ where $0< \alpha <1$. $\endgroup$
    – wayne
    Jun 7, 2015 at 3:45

2 Answers 2

5
$\begingroup$

It seems the following.

Let $A$ and $B$ be $n\times n$ matrices and $S\subset\Bbb R^n$ be the unit sphere. Then positive definiteness of the matrix $A$ is equivalent to

$$m(A)=\min\{(Ax,x):x\in S\}>0.$$

Similarly, we have

$$m(A-B)=\min\{(A-B)x,x):x\in S\}>0.$$

Put

$$M(B)=\max\{(Bx,x):x\in S\}>0.$$

Let $\alpha>\frac{M(B)}{m(A-B)+M(B)}<1$ be any number and $x\in S$ be any vector. Then

$$(\alpha Ax,x)-(Bx,x)=$$ $$\alpha((A-B)x,x)-(1-\alpha)(Bx,x)\ge$$ $$\alpha m(A-B)-(1-\alpha)M(B)=$$ $$\alpha(m(A-B)+M(B))-M(B)>$$ $$\frac{M(B)}{m(A-B)+M(B)} (m(A-B)+M(B))-M(B)>0.$$

Then

$$m(\alpha A-B)= \min\{(\alpha A-B)x,x):x\in S\}>0.$$

So the matrix $\alpha A−B$ is positive-definite, that is $\alpha A>B$

$\endgroup$
6
  • $\begingroup$ Thanks for your efforts, I have 2 questions regarding your answer: 1) why let $A$ and $B$ be $n×n$ matrices and $S\subset R^n$ be the unit sphere? 2) $(\alpha Ax, x)-(Bx,x)>0$ may not imply $m(\alpha A-B)>0$. $\endgroup$
    – wayne
    Jun 7, 2015 at 7:49
  • $\begingroup$ The general definition considers the complex field, not real. $\endgroup$ Jun 7, 2015 at 8:14
  • $\begingroup$ Can $\alpha =\frac{M(B)}{M(B)+m(A-B)}$ such that $\alpha A\geq B$? $\endgroup$
    – wayne
    Jun 7, 2015 at 12:44
  • $\begingroup$ 1) I just fixed $n$. $S$ represents all non-zero vectors of the space $\Bbb R^n$. I considered only a real case, but I hope that the proof for the complex case is similar. $\endgroup$ Jun 8, 2015 at 5:25
  • 1
    $\begingroup$ 2) The sphere $S$ is compact, so a continuous positive-valued function on $S$ attains its (positive) infimum (that is why I wrote $”\min”$ instead of $”\inf”$.) $\endgroup$ Jun 8, 2015 at 5:25
3
$\begingroup$

The @Alex Ravsky's answer is good, however, I still want to share my answer:

Proof: To prove $\alpha A\geq B$ where $0<\alpha<1$, we introduce an extra parameter $\lambda$, $0<\lambda<1$, $\alpha=1-\lambda$, such that $(1-\lambda)A\geq B$, which is equivalent to find a $\lambda$, $0<\lambda<1$, such that $(1-\lambda)A\geq B$.

Arrange it, we get, $\lambda A\leq A-B$, to make it hold, we shall find $\lambda$ such that $\lambda A \leq \lambda_{\min}(A-B)I$ holds, which is equivalent to $\lambda I \leq \lambda_{\min}(A-B) A^{-1}$. Further, let $\lambda = \frac{\lambda_{\min}(A-B)}{\lambda_{\max}(A)}$, the above equations are automatically satisfied.

$\endgroup$
9
  • $\begingroup$ I expect that, in fact, these $\lambda_{m\dots}$’s are the respective (minimal of maximal) eigenvalues of the matrices. $\endgroup$ Jun 8, 2015 at 5:30
  • 1
    $\begingroup$ 1) I expect the following. Let $C$ be a positive definite symmetric $n\times n$ real-value matrix. Then $$\min\{(Cx,x):x\in S\}=\lambda_{min}(C)$$ and $$\max\{(Cx,x):x\in S\}=\lambda_{max}(C),$$ where $\lambda_{min}$ and $\lambda_{max}$ are, respectively, minimal and maximal eigenvalues of the matrix $C$. $\endgroup$ Jun 8, 2015 at 10:09
  • 1
    $\begingroup$ Thanks for your comments, it's clear now. Regarding your reply 1): I don't know $\lambda_{\min}$ or $\lambda_{\max}$ can be interpreted in this form. 3): I tried and arrived at $\frac{\lambda_{\max}(B)}{\lambda_{\max}(B)+\lambda_{\min}(A-B)} - \frac{\lambda_{\max}(A)-\lambda_{\min}(A-B)}{\lambda_{\max}(A)}=\frac{\lambda_{\min}(A-B)(\lambda_{\min}(A-B)-\lambda_{\max}(A)+\lambda_{\max}(B))}{\lambda_{\max}(A)\lambda_{\max}(B)+\lambda_{\max}(A)\lambda_{\min}(A-B)}$. Since it's hard to decide whether $(\lambda_{\min}(A-B)-\lambda_{\max}(A)+\lambda_{\max}(B))>0$ or not, I stop here. $\endgroup$
    – wayne
    Jun 8, 2015 at 14:46
  • 1
    $\begingroup$ 3) It seems the following. Yes, I arrived to this inequality too. From here we can proceed as follows. Let $x\in S$ be an arbitrary vector. Then $$(Ax, x)-(Bx, x)=((A-B)x , x)\ge\lambda_{min}(A-B)$$ and $$(Ax , x)\le\lambda_{max}(A).$$ So $$(Bx, x)\le \lambda_{max}(A)- \lambda_{min}(A-B).$$ Since the vector $x$ was chosen arbitrarily, we may claim that $$\lambda_{max} (B)\le \lambda_{max}(A)- \lambda_{min}(A-B).$$ $\endgroup$ Jun 8, 2015 at 20:12
  • 1
    $\begingroup$ Thanks for your efforts, everything is clear now. It seems that your bound is smaller than mine. $\endgroup$
    – wayne
    Jun 9, 2015 at 1:00

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.