next up previous
Next: Vector norms Up: algebra Previous: Singular value decomposition

Pseudo-inverse

Let ${\bf A}$ be an $m\times n$ matrix. If $m\ne n$, then ${\bf A}$ does not have an inverse. However, we can find its pseudo-inverse ${\bf A}^-$, an $n\times m$ matrix, as shown below.

Note that the two pseudo-inverses defined above for $m>n$ and $m<n$ are essentially the same. To see this, we first assume ${\bf A}$ has more rows than columns ($m>n$), then another matrix defined as ${\bf B}={\bf A}^T$ has more columns than rows. Taking the transpose on both sides of the first pseudo inverse above, we get

\begin{displaymath}({\bf A}^-)^T=[({\bf A}^T{\bf A})^{-1}{\bf A}^T]^T
={\bf A}({\bf A}^T{\bf A})^{-1}=({\bf A}^T)^{-};
\end{displaymath}

i.e.,

\begin{displaymath}{\bf B}^-={\bf B}^T({\bf B}{\bf B}^T)^{-1} \end{displaymath}

which is the same as the second definition of the pseudo inverse.

We can also show that $({\bf A}^-)^-={\bf A}$. If $m>n$, then we have

$\displaystyle ({\bf A}^-)^-$ $\textstyle =$ $\displaystyle [({\bf A}^T{\bf A})^{-1}{\bf A}^T]^-
=[({\bf A}^T{\bf A})^{-1}{\...
...{\bf A}^T{\bf A})^{-1}{\bf A}^T[({\bf A}^T{\bf A})^{-1}{\bf A}^T]^T\right]^{-1}$  
  $\textstyle =$ $\displaystyle {\bf A}({\bf A}^T{\bf A})^{-1}\left[ ({\bf A}^T{\bf A})^{-1}{\bf A}^T{\bf A}({\bf A}^T{\bf A})^{-1}\right]^{-1}$  
  $\textstyle =$ $\displaystyle {\bf A}({\bf A}^T{\bf A})^{-1}({\bf A}^T{\bf A})={\bf A}.$  

Similarly. we can show the same is true if $m<n$. In particular, when $m=n$, ${\bf A}$ is invertible and both versions of the pseudo-inverse defined above become the regular inverse ${\bf A}^-={\bf A}^{-1}$.


next up previous
Next: Vector norms Up: algebra Previous: Singular value decomposition
Ruye Wang 2015-04-27