2
$\begingroup$

I am confused about the bi-conditional statement I outlined in the title.

I'll write it again here: (This was part of a solution to an assignment question)

$A$ has full column rank if and only if the symmetric matrix $B=A^TA$ is positive definite.

The definition of column rank that I am aware of states that a $m \times n$ Matrix $A$ has full column rank if each of the columns are linearly independent. So it would be full rank if $rank(A) = n$ in this case.

And I guess, the dimension of $B = A^TA$ is $n \times n$. So it would be a square matrix that is full column rank. Now going back to the statement, why can we say that this matrix $B$ is symmetric, let alone a positive definite?

$\endgroup$
  • 1
    $\begingroup$ Note that $x^TBx=(Ax)^T(Ax)\geq0$, so $B$ is positive semidefinite no matter what $A$ is. When is $(Ax)^T(Ax)=0$? $\endgroup$ – Arthur Dec 20 '16 at 22:04
  • $\begingroup$ @πr8 yes I know that. $\endgroup$ – h94 Dec 20 '16 at 22:05
  • $\begingroup$ @Arthur never thought of it that way. That makes a lot of sense. But still don't understand the symmetry. $\endgroup$ – h94 Dec 20 '16 at 22:07
  • $\begingroup$ @πr8 But where would the column rank come into play? $\endgroup$ – h94 Dec 20 '16 at 22:09
  • $\begingroup$ It doesn't, for the symmetry - the column rank is only relevant (I think) for the positive-definiteness. $\endgroup$ – πr8 Dec 20 '16 at 22:09
2
$\begingroup$

The definition of column rank is the dimension of the image of $A$, so 'full column rank' means that $A$ is surjective. The columns are linearly independent iff $A$ is injective. $B := A^{\mathsf T} A$ is symmetric: $$ B^{\mathsf T} = (A^{\mathsf T} A)^{\mathsf T} = A^{\mathsf T} A = B $$ $A$ is injective iff $B$ is positive definite: Let $A$ be injective. Then $v^{\mathsf T} B v = v^{\mathsf T} A^{\mathsf T} A v = \|Av\|^2 > 0$ for all $v \neq 0$. If $B$ is positive definite, then it is injective. Since $B = A^{\mathsf T} A$ we have $\ker(A) \subset \ker(B) = 0$, which implies that $A$ is injective.

$\endgroup$

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.