Is the following a valid identity?
$$ (I - XX^T)^{1/2} X = X (I - X^T X)^{1/2} $$
Without the square root, it is quite easy to see that
$$ (I - XX^T) X = X (I - X^T X) $$
holds but I have trouble verifying that the upper equation holds.
Is the following a valid identity?
$$ (I - XX^T)^{1/2} X = X (I - X^T X)^{1/2} $$
Without the square root, it is quite easy to see that
$$ (I - XX^T) X = X (I - X^T X) $$
holds but I have trouble verifying that the upper equation holds.
On
Making use of the SVD $X = U \Sigma V^T$, one arrives at
$$ (I - X X^T)^{1/2} X = U (I - \Sigma^2)^{1/2} \Sigma V $$
and
$$ X (I - X^T X)^{1/2} = U \Sigma (I - \Sigma^2)^{1/2} V $$
where the diagonal entries may be swapped, $\Sigma (I - \Sigma^2)^{1/2} = (I - \Sigma^2)^{1/2} \Sigma$. Hence the identity holds!
A very general rule exists for these things.
Let $A,B\,$ be any two matrices such that the product $AB$ is a square matrix, and let $f(z)$ be an analytic function of a scalar argument.
Then the following is true $$f(AB)\cdot A = A\cdot f(BA)$$
The proof consists of a series expansion of the function coupled with the associativity of matrix multiplication.
In this particular case, $\,A=X,\,B=X^T\,$ and $\,f(z) = \sqrt{1-z}$