# Proof of Congruence Transformation

## Main Question or Discussion Point

Consider the following statement

Let T (size: nxm) be a complex matrix. Then if A of dimension nxn is positive semidefinite then T*AT >= 0.

Now I was wondering if the converse is true aswel? In my math book they used the converse statement to proof something, but is it possible to say that if T*AT >= 0 (positive semidefinite) then A>= 0?

Note: I used the symbol * to indicate the Hermittian.

Someone got some tips for me?

Related Linear and Abstract Algebra News on Phys.org
jbunniii
Homework Helper
Gold Member
It's not true in general. Suppose, for example, that T is n x 1. So it's a particular column vector; let's call it $$x_0$$ to emphasize that fact.

Then in that case, T*AT >= 0 means simply that

$$x_0^* A x_0 \geq 0$$ (scalar inequality)

for that specific $$x_0$$ (or any scalar multiple of $$x_0$$). But in order to conclude that A is positive semidefinite, you would need

$$x^* A x \geq 0$$

for EVERY column vector x.

: A sufficient condition for the conclusion you desire would be that T must be surjective. Do you see why?

Thanks for the fast replay. Let me see if i got.

A>=0 implies T*AT>=0 if T has full column rank. That means that for every column (i) of T we have Ti*ATi>=0.

If T did not have full column rank then there are linear dependent columns for which it does not hold that Ti*ATi>=0.

jbunniii
Homework Helper
Gold Member
Thanks for the fast replay. Let me see if i got.

A>=0 implies T*AT>=0 if T has full column rank. That means that for every column (i) of T we have Ti*ATi>=0.

If T did not have full column rank then there are linear dependent columns for which it does not hold that Ti*ATi>=0.
No, that's not right. A >= 0 implies T*AT >= 0 for ANY complex n x m matrix T.

Why is this? Because

x*(T*AT)x = (Tx)*A(Tx) = y*Ay >= 0 because A is positive semidefinite

(here I have defined y = Tx for added clarity)

You asked about the converse: when does T*AT >= 0 imply A >= 0?

A sufficient condition is for T to be surjective, i.e., for T to have full ROW rank. (Not full column rank! The n x 1 example I gave in the last post has full column rank, but the desired result does not hold.)

Why is it sufficient for T to be surjective? Let y be any n x 1 vector. We want

y*Ay >= 0.

Since T is surjective, there exists a (m x 1) vector x such that y = Tx.

Then:

y*Ay = (Tx)*A(Tx) = x*(T*AT)x* >= 0 because T*AT is positive semidefinite. QED.

The next natural question is: is it also NECESSARY for T to be surjective? I'll let you think about that one.