Given $ w,x,y \in \mathbf{R^{n}}$
which one has better numerical properties?
(1) $ w = z - (x \cdot y^{T}) \cdot z$
(2) $ w = z - x \cdot (y^{T} \cdot z)$
First one needs more FLOPS (thus more possible floating point arithmetic errors) and storage to be solved. Which is more robust to perturbations?