Assume the following:
• $L\leq K$ .
• $\Gamma\in M_{K,L}$ is a $L$ rank ${ 0,1} $ matrix, without identical rows or the zeros row.
• $N\in M_{K,K}$ is a diagonal matrix, whose diagonal is a distribution: all entries are non-negative and $trace\left(N\right)=1$.
• $\Lambda\in M_{K,K}$ is a diagonal matrix, satisfying $\Lambda_{i,i}=\sum_{j=1}^{L}\Gamma_{i,j}$, or equivalently: $diag\left(\Lambda\right)=diag\left(\Gamma\Gamma^{T}\right)$.
Denote:
$X=\Gamma^{T}N\Lambda^{-1}\Gamma\qquad=\sum_{i=1}^{K}N_{i,i}\Lambda_{i,i}^{-1}\{ (\Gamma[i,:])^{T}\Gamma[i,:]} $
$Y=\Gamma^{T}N\Gamma\;\;\;\;\;\;\;\;=\sum_{i=1}^{K}N_{i,i}\{ (\Gamma[i,:])^{T}\Gamma[i,:]}$
$Z=\Gamma^{T}N\Lambda\Gamma\;\;\;\;\;\;=\sum_{i=1}^{K}N_{i,i}\Lambda_{i,i}\{ (\Gamma[i,:])^{T}\Gamma[i,:]\} $
I'm interested in showing that the expression $M=X-YZ^{-1}Y$ is small in some sense (under non-singularity assumption of all matrices noted).
Observations so far:
If $K=L$ then we get that $\Gamma$ is invertible and $M=0$.
$M$ is positive semidefinite.
Experiments showed so far that M greatly tends component wise to positive small values.
Has anyone seen something similar?, I'd like to find a tight bound on each entry of $M$, or even understand why it almost always ends up double non-negative.