During my research I have come across matrices this type

$$C=B\left(B^T B\right)^{-1}B^T\ ,$$

where $B$ is an $m\times n$ real matrix. If $B^TB$ is not invertible, then $\left(B^T B\right)^{-1}$ should be interpreted as the Moore-Penrose pseudoinverse. the I am interested in the diagonal elements of $C$.

Clearly, if $m=n$ and $B$ is invertible, then $C$ is the identity matrix $I_{m\times m}$. It is pretty easy to show that this is still the case when $m<n$ and the rows of $B$ are linearly independent.

However, when the rows of $B$ are not independent, there is not much I can say. From the physical context where this came up, I have a very strong suspicion (supported by numerical experiments) that the diagonal elements of $C$ are bounded from above by $1$ (and clearly from below by $0$). My question is whether someone could give me a hint as to show this, or where to look for in the literature.

  • 2
    $\begingroup$ en.wikipedia.org/wiki/Projection_%28linear_algebra%29 $\endgroup$ – Federico Poloni Feb 26 '15 at 12:51
  • 1
    $\begingroup$ (in particular, you'll need the formula and the part where it says that $\|Pv\|\leq \|v\|$, which you can use for the elements of the canonical basis $v=e_i$ to get what you need). $\endgroup$ – Federico Poloni Feb 26 '15 at 13:06
  • $\begingroup$ After years of experience, I'm still shocked from my own ignorance. Thanks. If you write it as an answer I will accept it. $\endgroup$ – yohbs Feb 26 '15 at 17:32
  • $\begingroup$ Don't feel that way -- This is a direction of linear algebra which, although elementary, is not usually explored in the "classical" courses; it is not shocking that a mathematician working in another field has never seen this in detail. $\endgroup$ – Federico Poloni Feb 26 '15 at 17:38

$C$ is an idempotent matrix, so its eigenvalues are either 1 or zero. $C$ is also Hermitian so Schur's Theorem says the sum of the $k$ largest eigenvalues is greater then the sum of the $k$ largest diagonal elements of $C$ for $k=1,2,...,m$. No diagonal element of $C$ can therefore exceed unity.

  • $\begingroup$ I accepted your answer, but I think @FedericoPoloni's answer is somewhat simpler. The crux is $C$ is idempotent $\Rightarrow$ its eigenvalues are $\in [0,1]$ $\Rightarrow$ foe any $v$ $||Pv||\le||v||$, and specifically for elemets of the cannonical basis $\endgroup$ – yohbs Mar 1 '15 at 9:05

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.