Let $A \in \mathbb{R}^{n \times p}$, let $D$ be a diagonal matrix with positive entries. $\dagger$ denotes the Moore-Penrose pseudoinverse.

Is it true in general that: $$(A^\top D A)^\dagger A^\top D = (A^\top D^2 A)^\dagger A^\top D^2 \enspace ? $$ (and if not, is it easy to determine when this holds?)

If $D$ is proportional to identity, ok. Multiplying by $A$ on the right yields an identity which looks reasonable when $A^\top D A$ is invertible.

Numerical simulations with $p > n$ give frequent equality, but not always, and I'm unable to determine in which case the equality holds.


Using the identity $(X^\top X)^\dagger X^\top=X^\dagger$, the equality in question can be rewritten as $(\sqrt{D}A)^\dagger\sqrt{D}=(DA)^\dagger D$.

It is true when $A$ has full row rank (but numerical verification may fail because of rounding errors). In this case, $(BA)^\dagger=A^\dagger B^\dagger$ for any $B$ with full column rank (see Wikipedia). Hence $$ (\sqrt{D}A)^\dagger\sqrt{D}=A^\dagger \sqrt{D}^\dagger\sqrt{D}=A^\dagger=A^\dagger D^\dagger D=(DA)^\dagger D. $$ Alternatively, when $A$ has full row rank, $(\sqrt{D}A)^\top$ and $(DA)^\top$ have the same column spaces. So, using the fact that $X^\dagger X$ is the orthogonal projection onto the column space of $X^\top$, we see that $(\sqrt{D}A)^\dagger\sqrt{D}A=(DA)^\dagger DA$. Right-multiply both sides by the right inverse of $A$, we obtain $(\sqrt{D}A)^\dagger\sqrt{D}=(DA)^\dagger D$.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.