You are correct-ish.

Left and right inverses are still inverses, just not commutative.

Moreover, since V is full rank and orthonormal, the SVD is trivially:

   svd(V) = V I I

This means that while V' is the left inverse, it is also the right Penrose
inverse in the sense that the least squares solution of argmin_X norm(V X -
I) is V'.

On Mon, Sep 10, 2012 at 2:05 PM, Dmitriy Lyubimov <[email protected]> wrote:

> Yep that's my understanding too. V cannot have inverse since it is not
> square (in common case).
>
> But since it is orthonormal, transpose produces a pseudoinverse which
> we can use just the same.
>
> On Mon, Sep 10, 2012 at 2:03 PM, Sean Owen <[email protected]> wrote:
> > Yes, doesn't the V V' != I bit mean it's not an inverse? I thought
> > only square matrices had a real inverse. This is a one-sided inverse,
> > which is (IIRC?) slightly stronger than being a pseudo-inverse.
> >
> > On Mon, Sep 10, 2012 at 9:59 PM, Ted Dunning <[email protected]>
> wrote:
> >> No.  It is the inverse.  V' V = I
> >>
> >> On the other hand, V V' != I.  We do know that norm(A V V' - A) is
> small.
>

Reply via email to