I need to know that if the following holds for complex vectors $x=[a\cdot A \mid b\cdot B]u$, and $y= [A \mid B]u$ $$\det(I+d \frac{yy^*}{rI})\leq\det(I+\frac{xx^*}{rI})\leq \det(I+c \frac{yy^*}{rI})$$ where $d= \min \{a,b\}$$, c = \max \{a,b\}$ and $r$ are positive real values. $u$ is a complex valued vector and $^*$ means conjugate transform (Hermitian).

By $[A \mid B]$ I mean the composite matrix made from the two matrices $A, B$. Let $A$ and $B$ be complex valued matrices.

It works in simulations. I think it has something to do with eigenvalues of positive definite matrices. I know that $xx^*$ and $yy^*$ are positive definite and so is $I$. Also I think the division by $I$ is not a problem here. But I don't know how to prove it.

Thank you.


(Presumably both $A$ and $B$ are nonzero matrices and $a\neq b$, i.e. $0<d<c$.) Contrary to your claim, your inequality never holds universally, i.e. we can always find a vector $u$ that violates the inequality.

For any complex scalar $\lambda$ and any complex vector $v$, we have $\det(I+\lambda vv^\ast)=1+\lambda v^\ast v$. Now, suppose $A$ is $n\times p$, $B$ is $n\times q$ and $M=[A|B]\neq0$. Then $x=M(aI_p\oplus bI_q)u$ and your inequality is equivalent to $$ 1+\frac{u^\ast (dM^\ast M)u}{r} \leq 1+\frac{u^\ast(aI_p\oplus bI_q)M^\ast M(aI_p\oplus bI_q)u}{r} \leq 1+\frac{u^\ast(cM^\ast M)u}{r}.\tag{1} $$ If it holds for all $u$, we must have $$ dM^\ast M \preceq (aI_p\oplus bI_q)M^\ast M(aI_p\oplus bI_q) \preceq cM^\ast M,\tag{2} $$ where $X\preceq Y$ means $Y-X$ is positive semidefinite. In particular, the analogous inequality must hold for the leading principal $p\times p$ and trailing principal $q\times q$ submatrices. That is, \begin{cases} dA^\ast A \preceq a^2A^\ast A \preceq cA^\ast A,\\ dB^\ast B \preceq b^2B^\ast B \preceq cB^\ast B.\tag{3} \end{cases} Since both $A$ and $B$ are nonzero, the above implies that $d\le a^2\le c$ and $d\le b^2\le c$, i.e. $d\le d^2\le c^2\le c$. Yet, by assumption, $0<d<c$. So, the statement that $d\le d^2\le c^2\le c$ is always false and at least one case in $(3)$ is bound to fail. That is, there always exists a vector in either $\mathbb{R}^p\times 0^q$ or $0^p\times\mathbb{R}^q$ such that $(1)$ does not hold.

For an illustrative example, consider $r=1,\ a=d=\frac12,\ b=c=1$ and $$ M=[A|B]=(1,1),\ u=\pmatrix{1\\ 0},\ y=1,\ x=\frac12. $$ Then the first inequality sign in $\det(I+\frac{dyy^*}{r})=\frac32\color{red}{\le}\det(I+\frac{xx^*}{r})=\frac54\le \det(I+\frac{cyy^*}{r})=2$ is violated.

In the uninteresting cases, a similar analysis shows that your inequality holds for all $u$ iff one of the following four conditions is satisfied:

  1. $A=0$ and $B=0$.
  2. $A\ne0,\ B\ne0$ and $a=b=1$,
  3. $A=0,\ B\ne0$ and $d\le b^2\le c$,
  4. $A\ne0,\ B=0$ and $d\le a^2\le c$.
| cite | improve this answer | |
  • $\begingroup$ Thank you. I have made a mistake in writing the question. I it should be $dydy^*$ not $dyy^*$, same goes with $c$. I want to replace $a,b$ by their min and max. I think now My inequality is correct if I understood your proof? If you wish I can correct my question and then your can change the answer so other readers may know the correct inequality (assuming it is correct when changed). Also a small question about the notation does $(aI_p\oplus bI_q)[A\mid B]$ do same as $[aA\mid bB]$? because $A$ is not a square matrix so I don't know how $aI_pA$ can give $aA$. Once again thank you very much. $\endgroup$ – triomphe Sep 13 '13 at 13:26
  • $\begingroup$ @MLT (1) $aI_p\oplus bI_q$ means $\pmatrix{aI_p\\ &bI_q}$, so the matrix product is $[A|B](aI_p\oplus bI_q)$, not $(aI_p\oplus bI_q)[A|B]$. (2) If the $c$ and $d$ in your question are replaced by $c^2$ and $d^2$, then we would expect the inequality to hold in more cases, but it still does not hold in general. For a counterexample, consider $r=1,\ a=d=1,\ b=c=2,\ [A|B]=(2,1),\ [aA|bB]=(2,2)$. With $u=(1,-1)^T$, we get $y=1,\ x=0$ and the first inequality sign in $1+d^2y^2 = 2 \color{red}{\le} 1+x^2 = 1 \le 1+c^2y^2 = 5$ is still wrong. $\endgroup$ – user1551 Sep 13 '13 at 15:08
  • $\begingroup$ But what about your claim ...implies that $d\le a^2\le c$ and $d\le b^2\le c$,... I thought it will now change to ...implies that $d^2\le a^2\le c^2$ and $d^2\le b^2\le c^2$,... which is true in my case. $\endgroup$ – triomphe Sep 13 '13 at 15:25
  • 1
    $\begingroup$ @MLT If a matrix is positive semidefinite, all principal submatrices should be PSD too. It just happens that, in your first version of the problem (using $c,d$ in the inequality), the inspection of two sets of principal submatrices are enough to rule out all interesting cases. Yet in the corrected version (using $c,d$ in the inequality), $(3)$ only constitute two necessary conditions and they are far from being sufficient. $\endgroup$ – user1551 Sep 13 '13 at 16:24
  • $\begingroup$ In the counterexample mentioned my previous comment, we have $(aI_p\oplus bI_q)M^\ast M(aI_p\oplus bI_q)-d^2M^\ast M=\pmatrix{0&2\\ 2&3}$. While the leading and trailing principal $1\times1$ submatrices (i.e. the two diagonal entries) are PSD, the matrix difference is not. $\endgroup$ – user1551 Sep 13 '13 at 16:24

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.