8
$\begingroup$

Let

$$K=\begin{pmatrix} K_{11} & K_{12}\\ K_{21} & K_{22} \end{pmatrix}$$

be a symmetric positive semidefinite real matrix (PSD) with $K_{12}=K_{21}^T$. Then, for $|r| \le 1$,

$$K^*=\begin{pmatrix} K_{11} & rK_{12}\\ rK_{21} & K_{22} \end{pmatrix}$$

is also a PSD matrix. Matrices $K$ and $K^*$ are $2 \times 2$ and $K_{21}^T$ denotes the transpose matrix. How do I prove this?

$\endgroup$
  • 2
    $\begingroup$ I think this question needs the self-study tag. $\endgroup$ – Michael Chernick Jan 9 '18 at 19:34
  • $\begingroup$ Please add the [self-study] tag & read its wiki. Then tell us what you understand thus far, what you've tried & where you're stuck. We'll provide hints to help you get unstuck. $\endgroup$ – gung Jan 9 '18 at 21:54
  • 1
    $\begingroup$ If K is 2x2, does that mean K_21 is a scalar? If so, why are you talking about its transpose? $\endgroup$ – Acccumulation Jan 9 '18 at 22:51
14
$\begingroup$

This is a nice opportunity to apply the definitions: no advanced theorems are needed.

To simplify the notation, for any number $\rho$ let $$\mathbb{A}(\rho)=\pmatrix{A&\rho B\\\rho B^\prime&D}$$ be a symmetric block matrix. (If working with block matrices is unfamiliar to you, just assume at first that $A$, $B$, $D$, $x$, and $y$ are numbers. You will get the general idea from this case.)

For $\mathbb{A}(\rho)$ to be positive semidefinite (PSD) merely means that for all vectors $x$ and $y$ of suitable dimensions

$$\eqalign{ 0 &\le \pmatrix{x^\prime&y^\prime} \mathbb{A}(\rho) \pmatrix{x\\y} \\ &= \pmatrix{x^\prime&y^\prime} \pmatrix{A&\rho B\\\rho B^\prime&D}\pmatrix{x\\y} \\ &=x^\prime A x + 2\rho y^\prime B^\prime x + y^\prime D y.\tag{1} }$$

This is what we have to prove when $|\rho|\le 1$.

We are told that $\mathbb{A}(1)$ is PSD. I claim that $\mathbb{A}(-1)$ also is PSD. This follows by negating $y$ in expression $(1)$: as $\pmatrix{x\\y}$ ranges through all possible vectors, $\pmatrix{x\\-y}$ also ranges through all possible vectors, producing

$$\eqalign{ 0 &\le \pmatrix{x^\prime&-y^\prime}\mathbb{A}(1)\pmatrix{x\\-y} \\ &= x^\prime A x + 2(-y)^\prime B^\prime x + (-y)^\prime D (-y) \\ &= x^\prime A x + 2(-1)y^\prime B^\prime x + y^\prime D y \\ &= \pmatrix{x^\prime&y^\prime}\mathbb{A}(-1)\pmatrix{x\\y}, }$$

showing that $(1)$ holds with $\rho=-1.$

Notice that $\mathbb{A}(\rho)$ can be expressed as a linear interpolant of the extremes $\mathbb{A}(-1)$ and $\mathbb{A}(1)$:

$$\mathbb{A}(\rho) = \frac{1-\rho}{2}\mathbb{A}(-1) + \frac{1+\rho}{2}\mathbb{A}(1).\tag{2}$$

When $|\rho|\le 1$, both coefficients $\color{blue}{(1-\rho)/2}$ and $\color{blue}{(1+\rho)/2}$ are non-negative. Therefore, since both ${\pmatrix{x^\prime&y^\prime}\mathbb{A}(1)\pmatrix{x\\y}}$ and $\pmatrix{x^\prime&y^\prime}\mathbb{A}(-1)\pmatrix{x\\y}$ are nonnegative, so is the right hand side of

$$\eqalign{ &\pmatrix{x^\prime&y^\prime}\mathbb{A}(\rho)\pmatrix{x\\y} \\ &= \color{blue}{\left(\frac{1-\rho}{2}\right)}\pmatrix{x^\prime&y^\prime}\mathbb{A}(-1)\pmatrix{x\\y} + \color{blue}{\left(\frac{1+\rho}{2}\right)}\pmatrix{x^\prime&y^\prime}\mathbb{A}(1)\pmatrix{x\\y} \\ &\ge \color{blue}{0}(0) + \color{blue}{0}(0) = 0. }$$

(I use colors to help you see the four separate non-negative terms that are involved.)

Because $x$ and $y$ are arbitrary, we have proven $(1)$ for all $\rho$ with $|\rho|\le 1$.

$\endgroup$
  • 4
    $\begingroup$ This is quite beautiful in its simplicity :-) $\endgroup$ – TenaliRaman Jan 9 '18 at 17:15
7
$\begingroup$

There is already a great answer by @whuber, so I will try to give an alternative, shorter proof, using a couple theorems.

  1. For any $A$ - PSD and any $Q$ we have $Q^TAQ$ - PSD
  2. For $A$ - PSD and $B$ - PSD also $A + B$ - PSD
  3. For $A$ - PSD and $q > 0$ also $qA$ - PSD

And now:

\begin{align*} K^* &= \begin{pmatrix} K_{1,1} & rK_{1,2} \\ rK_{2,1} & K_{2,2} \\ \end{pmatrix} \\ &= \begin{pmatrix} K_{1,1} & rK_{1,2} \\ rK_{2,1} & r^2K_{2,2} \\ \end{pmatrix} + \begin{pmatrix} 0 & 0 \\ 0 & qK_{2,2} \\ \end{pmatrix}, \text{ where $q = 1 - r^2 > 0$} \\ &= \begin{pmatrix} I & 0 \\ 0 & rI \\ \end{pmatrix}^T \begin{pmatrix} K_{1,1} & K_{1,2} \\ K_{2,1} & K_{2,2} \\ \end{pmatrix} \begin{pmatrix} I & 0 \\ 0 & rI \\ \end{pmatrix} + q\begin{pmatrix} 0 & 0 \\ 0 & K_{2,2} \\ \end{pmatrix} \end{align*}

Matrix $K$ is PSD by definition and so is its submatrix $K_{2, 2}$

$\endgroup$
  • 4
    $\begingroup$ +1 Nice demonstration! It might be made a little clearer by using "$q$" instead of "$r$" in your statement of fact (3). $\endgroup$ – whuber Jan 9 '18 at 16:59

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.