Let $A$ be an $m\times n$ complex matrix of rank $r$ and let $$\lambda_1 (AA^*)\geq \lambda _2(AA^*)\geq \ldots \lambda_r(AA^*)$$ denote the non-zero eigen values of $AA^*$. If the real scalar $\alpha$ satisfies $$0<\alpha<\frac 2{\lambda_1(AA^*)}$$ then prove that the eigen values of $\alpha AA^*-AA^\dagger$ are $$\begin{cases} \alpha \lambda_i(AA^*)-1 & i=1,2,3,\ldots ,r\\ 0 & i =r+1,\ldots,m \end{cases}$$

Here the notation $A^*$ is complex conjugate transpose of $A$ and $A^\dagger$ is moore penrose inverse of $A$. That is, $A^\dagger$ satisfies, $$AA^\dagger A=A\quad A^\dagger AA^\dagger=A^\dagger\quad (AA^\dagger)^*=AA^\dagger\quad (A^\dagger A)^*=A^\dagger A$$


Let $A=USV^*$ be the reduced SVD of $A$ ($U\in\mathbb{C}^{m\times r}$, $S:=\mathrm{diag}(\sigma_1,\ldots,\sigma_r)$ is $r\times r$ a diagonal positive definite matrix, $V\in\mathbb{C}^{n\times r}$). In your notation, you can choose $\sigma_i$ such that $\sigma_i^2=\lambda_i(AA^*)$, $i=1,\ldots,r$. Then $A^\dagger=VS^{-1}U^*$ and $$ \alpha AA^*-AA^\dagger=\alpha USV^*VSU^*-USV^*VS^{-1}U^* =U\left(\alpha S^2-I_r\right)U^*. $$ Now let $U^\perp\in\mathbb{C}^{m\times (m-r)}$ be such that $\tilde{U}:=[U,U^\perp]$ is a square unitary matrix. Then $$ \alpha AA^*-AA^\dagger=U\left(\alpha S^2-I_r\right)U^*=\tilde{U}\pmatrix{\alpha S^2-I_r\\&0_{m-r}}\tilde{U}^* $$ is the spectral decomposition of the Hermitian matrix $\alpha AA^*-AA^\dagger$.

Note that the assumption on the bounds for $\alpha$ is irrelevant.

Another way without SVD by using the definition of the pseudoinverse

It is sufficient to show, that:

  • if $AA^*v=\lambda v$, $\lambda\neq 0$, $v\neq 0$, then $AA^\dagger v=v$,
  • if $AA^*w=0$, then $AA^\dagger w=0$.

Take this as a fact for now. The first item implies that $$ (\alpha AA^*v-AA^\dagger) v=(\alpha\lambda-1)v $$ and the second item gives $$ (\alpha AA^*-AA^\dagger)w=0-0=0. $$ This shows what you already wanted.

Now to show the first item is true, assume $AA^*v=\lambda v$ and $\lambda\neq 0$. Then $$ AA^\dagger v=AA^\dagger(\lambda^{-1}AA^*v)=\lambda^{-1}AA^\dagger AA^*v =\lambda^{-1}AA^*v=\lambda^{-1}\lambda v=v. $$ For the second item, note that $AA^*w=0$ if and only if $A^*w=0$. We show that the kernels of $A^\dagger$ and $A^*$ are identical. Since $A^\dagger=A^\dagger AA^\dagger=A^\dagger (AA^\dagger)^*=A^\dagger (A^\dagger)^* A^*$, we have $\mathrm{ker}(A^*)\subseteq\ker(A^\dagger)$. On the other hand, $A^*=(AA^\dagger A)^*=A^*(AA^\dagger)^*=A^*AA^\dagger$ and hence $\mathrm{ker}(A^\dagger)\subseteq\mathrm{ker}(A^*)$.

  • $\begingroup$ thank you for the answer.. its awesome.. I am thinking no need to take $U_{m\times r}$ rather we can take $U_{m\times m}$ itself. That is also possible!!(similarly for $V$). and $S=diag(\sigma_1,\ldots,\sigma_r,0,\ldots,0)$. So that computations are becoming simple. $\endgroup$ – David Nov 28 '15 at 15:56

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.