Moore-Penrose pseudoinverse are limits: $A^+ = \lim_{\alpha \searrow 0} (A^T A + \alpha I)^{-1} A^T$.

I have noticed that: $A^+ A = \lim_{\alpha \searrow 0}(A^T A + \alpha I)^{-1} A^T A = I$, even $A$ is not invertible. But I wonder why does calculating the pseudoinverse require to add $\alpha I$ even $\alpha$ converged zero.

  • 1
    $\begingroup$ because $A^TA$ may not be invertible? - Also note that for some $A$, $\lim_{ \alpha\to 0}\bigl((A^TA+\alpha I)^{-1}A^TA\bigr)\ne I$. Just try $A=0$. $\endgroup$ – Hagen von Eitzen Apr 1 '17 at 10:42
  • 2
    $\begingroup$ An analogy with the scalar case: $\lim_{\alpha\to0}(a^2+\alpha)^{-1}a$ is defined even when $a=0$. $\endgroup$ – Yves Daoust Apr 1 '17 at 10:49
  • $\begingroup$ $A^TA$ is not sure be full rank. Any 0 eigen/singular value of $A$ will be 0 for $A^TA$ $\endgroup$ – mathreadler Apr 1 '17 at 11:09

$A^TA$ is not sure be full rank. Any $0$ eigen/singular value of $A$ will be $0$ for $A^TA$, but any non-zero eigenvalue of $A$ will be $>0$ for $A^TA$ and $+\alpha I$ moves all eigenvalues by $+\alpha$ which is positively so any previous $0$ will be $>0$ and any previous non-zero will be even bigger so sure to not become a $0$.


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.