Show $AA^T-(X^T X)^{-1}$ is positive semidefinite where $AX=I$

101 Views Asked by At

Show $AA^T-(X^T X)^{-1}$ is positive semidefinite where $AX=I$. $A, X$ may not be invertible or square.

Attempt: One way is to check $$AA^T-(X^T X)^{-1}=$$$$\bigg(A-(X^TX)^{-1}X^T\bigg)\bigg(A-(X^TX)^{-1}X^T\bigg)^T$$ which is not hard if come up with this equality first. I feel this is like complete square with respect numbers but without a general procedure to follow. How do one complete square with matrices in general? Are there any more intuitive approach to the proof. Thank you.

1

There are 1 best solutions below

0
On

Here is an alternative approach. To prove that $$ v^TAA^Tv\ge v^T(X^TX)^{-1}v $$ for every vector $v$, write $v=X^Tu$ for some vector $u$ (in fact, we may take $u=A^Tv$). Thus the problem boils down to showing that $$ I\ge X(X^TX)^{-1}X^T, $$ but this latter inequality is evident because $X(X^TX)^{-1}X^T$ is an orthogonal projection.

You may also rewrite the above solution in the form of completion of square (well, it actually isn't a square, but a Gram matrix): \begin{aligned} AA^T-(X^TX)^{-1} &=A\left[I-X(X^TX)^{-1}X^T\right]A^T\\ &=A\left[I-X(X^TX)^{-1}X^T\right]\left[I-X(X^TX)^{-1}X^T\right]A^T\\ &=\left[A-(X^TX)^{-1}X^T\right]\left[A-(X^TX)^{-1}X^T\right]^T. \end{aligned} The moral: try to exploit the use of orthogonal projections.