We seek to perform a information theoretic analysis of the RG procedure itself. In order to lay the platform for such a analysis we fix the nomenclature first.

1>Let us label every pair of degrees of freedom in the entangled space as (i,j).

2>Associated with every pair is the feature(F(i,j))-mutual information value F(i,j)=I(i:j) that lies between 0 and 2log2. More the values I(i:j) stronger is the "entangledness" of the pair.

3>Based on this feature we define a classifier(C(i,j)) if I(i:j)<log2 then the pair is "weakly" entangled and C(i,j)=0 else the pair is "strongly" entangled C(i,j)=1.
4>We define a target classification data with respect to the RG fixed point and call that $C_{0}$. In our setup we have 6 RG steps. So we have compute two quantities:
1>Mutual Information content between the target classification $C_{0}$ and Feature set $F_{i}$ at every RG setup, let us denote that as I(F{i}:C{0}).
2>Mutual Information content between the Feature set $F_{6}$ at initial step and Feature at every later RG transformation step $F_{i}$, let us denote that as $I(F_{6}:F_{i})$.
Here $I(F_{i}:C_{0})$ quantifies the amount of essential information of the target classification $C_{0}$ the bottlenecked feature representation $F_{i}$ carries .
Here $I(F_{6}:F_{i})$ quantifies the amount of compression between the initial feature representation and the bottlenecked representation.

Given the demonstration that our RG methodology follows the Information Bottleneck principle we wish to test its efficiency in predicting that a given (i,j) pair is a strongly entangled pair or not. Predictive abilities of this RG can be checked in the following way: given a pair(i,j) at each RG step we compute the conditional probability P(S|(i,j)) of it being strong entangled pair finally? Below will layout the steps neccessary to answer this question.
1>Take as input the pair in question and (Total-N(No of RG transformation steps performed))
2>Check if the ordered pair entities belong to entangled subspace after N transformation layers:
if answer is No
then return P(S|(i,j))=0(as it has been disentangled at that or earlier layer)
else
compute the Joint probability distribution between Feature set at Nth RG step and target classifier, This tell us the association probability between the target class and bottleneck representation of features $F_{N}$ after N transformations of initial layer.