0
$\begingroup$

I'm not sure how to prove any of this. The only thing that seems to be related to this is: How does the SVD solve the least squares problem? but it does not answer any of the questions below

3.32. Prove that the pseudoinverse $\boldsymbol{A}^+$ of an $m\times n$ matrix $\boldsymbol A$, as defined using the SVD in Section 3.6.1, satisfies the following four properties, known as the Moore-Penrose conditions.

(a) $\boldsymbol{A}\boldsymbol{A}^+\boldsymbol{A}=\boldsymbol{A}$.

(b) $\boldsymbol{A}^+\boldsymbol{A}\boldsymbol{A}^+=\boldsymbol{A}^+$.

(c) $(\boldsymbol{A}\boldsymbol{A}^+)^T=\boldsymbol{A}\boldsymbol{A}^+$.

(d) $(\boldsymbol{A}^+\boldsymbol{A})^T=\boldsymbol{A}^+\boldsymbol{A}$.

$\endgroup$
  • $\begingroup$ A little bit difficult to answer without having section 3.6.1. I suppose $A^+$ and $A$ are defined in terms of the matrices commonly called $U,\Sigma,V$ in the SVD and that you are supposed to substitute in left and right hand sides and see if you can get them to be equal. $\endgroup$ – mathreadler Mar 31 '17 at 2:18
  • 1
    $\begingroup$ Please try and use less images of text in questions and more typeset content even if there probably was a good thought behind this time. $\endgroup$ – mathreadler Mar 31 '17 at 2:20
  • $\begingroup$ Please consider updating the title to use 'Moore-Penrose conditions'. This helps others to find your question. $\endgroup$ – dantopa Mar 31 '17 at 3:30
0
$\begingroup$

$$ \mathbf{A} = \mathbf{U} \, \Sigma \, \mathbf{V}^{*}, \qquad \mathbf{A}^{+} = \mathbf{U} \, \Sigma^{+} \, \mathbf{V}^{*} $$

$\Sigma$ gymnastics

The problem distills down to an exercise in $\Sigma$ gymnastics. The diagonal matrix $\mathbf{S}$ contains the ordered singular values has size $\rho \times \rho.$ Note $\mathbf{S}^{T} = \mathbf{S}.$ The inverse matrix has reciprocal values on the diagonal.

First step: $$ \begin{align} \Sigma = \left[ \begin{array}{cc} \mathbf{S} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{m\times n}, \qquad % \Sigma^{T} = \left[ \begin{array}{cc} \mathbf{S} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{n\times m}, \qquad % \Sigma^{+} = \left[ \begin{array}{cc} \mathbf{S}^{-1} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{n\times m} % \end{align} $$ Second step: $$ \begin{align} \Sigma \, \Sigma^{T} = \left[ \begin{array}{cc} \mathbf{S}^{2} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{m\times m}, \qquad % \Sigma^{T} \, \Sigma &= \left[ \begin{array}{cc} \mathbf{S}^{2} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{n\times n} % \end{align} $$ Third step: $$ \begin{align} \Sigma \, \Sigma^{+} = \left[ \begin{array}{cc} \mathbf{I}_{\rho} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{m\times m}, \qquad % \Sigma^{+} \, \Sigma &= \left[ \begin{array}{cc} \mathbf{I}_{\rho} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{n\times n} % \end{align} $$ Fourth step: $$ \begin{align} \Sigma \, \Sigma^{+} \, \Sigma &= \left[ \begin{array}{cc} \mathbf{S} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{m\times n} = \Sigma, \qquad % \Sigma^{+} \, \Sigma \, \Sigma^{+} = \left[ \begin{array}{cc} \mathbf{S} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right]_{n\times m} = \Sigma^{+} % \end{align} $$


a

$$ \begin{align} \mathbf{A} \, \mathbf{A}^{+} \, \mathbf{A} &= % A \left( \mathbf{U} \, \Sigma \, \mathbf{V}^{*} \right) % A \left( \mathbf{V} \, \Sigma^{+} \, \mathbf{U}^{*} \right) % A+ \left( \mathbf{U} \, \Sigma \, \mathbf{V}^{*} \right) \\ % & = \mathbf{U} \, \Sigma \, \Sigma^{+} \, \Sigma \, \mathbf{V}^{*} \\ % & = \mathbf{U} \, \Sigma \, \mathbf{V}^{*} \\ % & = \mathbf{A} % \end{align} $$

b

$$ \begin{align} \mathbf{A}^{+} \mathbf{A} \mathbf{A}^{+} & = \mathbf{V} \, \Sigma^{+} \, \Sigma \, \Sigma^{+} \, \mathbf{U}^{*} \\ & = \mathbf{V} \, \Sigma^{+} \, \mathbf{U}^{*} \\ % & = \mathbf{A}^{+} % \end{align} $$

c

$$ \begin{align} \left( \mathbf{A} \, \mathbf{A}^{+} \right)^{*} &= % A \left( \mathbf{U} \, \Sigma \, \Sigma^{+} \, \mathbf{U}^{*} \right)^{*} \\ % A &= \left( \mathbf{U} \, \left[ \begin{array}{cc} \mathbf{I}_{\rho} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right] \, \mathbf{U}^{*} \right)^{*} \\ &= \mathbf{U} \, \left[ \begin{array}{cc} \mathbf{I}_{\rho} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right] \, \mathbf{U}^{*} \\ &= \mathbf{A} \, \mathbf{A}^{\dagger} % \end{align} $$

d

$$ \begin{align} \left( \mathbf{A}^{+} \, \mathbf{A} \right)^{*} &= % A \left( \mathbf{V} \, \Sigma^{+} \, \Sigma \, \mathbf{V}^{*} \right)^{*} \\ % A &= \mathbf{V} \, \left[ \begin{array}{cc} \mathbf{I}_{\rho} & \mathbf{0} \\ \mathbf{0} & \mathbf{0} \end{array} \right] \, \mathbf{V}^{*} \\ &= \mathbf{A}^{+} \, \mathbf{A} % \end{align} $$

| cite | improve this answer | |
$\endgroup$
0
$\begingroup$

I'm going to assume that the SVD definition of the pseudo-inverse is $$ A^+ = V\Sigma^+U^*$$ where the SVD of $A$ is $A=U\Sigma V^*$ and $\Sigma^+$ is a diagonal matrix with diagonal entries $$[\Sigma^+]_{ii}=\left\{ \begin{array} 00 & \text{if} [\Sigma]_{ii}=0\\ ([\Sigma]_{ii})^{-1}&\text{otherwise}\end{array}\right.$$

It's then just a matter of plugging in the SVD of $A$ and $A^{+}$ into the equations (a)-(d). For example for (a) you have $$ AA^{+}A=(U\Sigma V^*)(V\Sigma^{+}U^*)(U\Sigma V^*)=U\Sigma\Sigma^+\Sigma V^*=U\Sigma V^*=A $$ $$

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ How would the transpose be used in this? $\endgroup$ – steve Mar 31 '17 at 2:39
  • $\begingroup$ According to wikipedia the transposes should be conjugate transposes (Hermitian transpose), but when the matrices are real these can be changed to transposes. $\endgroup$ – Peder Mar 31 '17 at 2:53

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.