Algebra view: matrix, column space Geometric view: operators, image

From least square

Penrose-I

Projection

Consider an operator $A:\RR^n\to\RR^m$.

The question is: what is the orthogonal projection onto subspace $\operatorname{im}(A)$, i.e. what is $P_A$?

For any vector $u\in\RR^m$,

\[P_A u =A v\]

for some $v\in\RR^n$ by the definition of $\operatorname{im}(A)$. By orthogonality, for any vector $w\in\RR^n$,

\[\left\langle Aw, u-P_A u \right\rangle =0 .\]

This is equivalent to

\[\left\langle w, A^* u-A^*A v \right\rangle =0\]

for any $w\in\RR^n$, so that

\[A^* u-A^*A v=0.\]

If $A^A$ is invertible, we have $v=(A^A)^{-1}A^u$, and thus $P_Au=A(A^A)^{-1}A^*u$.

If $A^*A$ is not invertible, we can consider any inverse, for example the Penrose inverse and obtain the representation

\[P_A=A(A^*A)^\dagger A^*.\] \[A^*P_Au=A^*u\]

In the world of matrices, $A\in M$ is a $m\times n$ matrix, $A^*$ is a $n\times m$ matrix.

Nystrom