Uniqueness of solution to least squares gradient descent using normal equations

71 Views Asked by At

It is known that any solution to the normal equations will solve the least squares minimization problem.

So finding a $w$ that satisfies $X^TXw=X^Ty$ will show that $w$ is optimal. However, it may be the case $X^TX$ is not invertible, so this would imply that the solution to the least squares may not be unique. However, we can use QR decomposition by setting $X=QR$ where $Q$ has orthogonal columns and $R$ is square and upper triangular with positive entries on the diagonal. This decomposition can always be done and is unique. Given $R$'s properties, it is always invertible. Plugging in this decomposition into the normal equations, after some simplifications we get $$Rw=Q^Ty$$ Since $R$ is invertible, we have $w=R^{-1}Q^Ty$ which I believe is unique. So would this not imply that the solution to the least squares gradient descent is always unique, regardless of the invertibility of $X^TX$?

1

There are 1 best solutions below

0
On BEST ANSWER

The entries of $R$ can be chosen to be positive only if $X$ has full column rank. If $X$ does not have independent columns, at least one diagonal entry of $R$ will be zero. Thus $R$ might not be invertible.