next up previous
Next: Practical Implementation Up: The BioCluct Clustering Algorithm Previous: Algorithm Running Time

Algorithm Correctness

Here we will give an outline of the proof. For the details of the proof please refer to [1].
\begin{theorem}Chernoff 1952 \cite{Chernoff52} \\
Let $X \sim \text{Binomial}(n...
...by $D(a\Vert p)=p\log(\frac{p}{a}) + (1-p)\log(\frac{1-p}{1-a})$ .
\end{theorem}


\begin{theorem}Let H be a clique graph with $\gamma$ -cluster structure. For the...
...\in \Omega(H,p)$\space the output $A(G)$\space is $H$\space w.h.p.
\end{theorem}


\begin{proof}(outline) \\
Consider the partition $\overline{C}^0 = \{\overline{...
...elta E(\overline{C})\vert < \vert E \Delta E(C)\vert$ ;
\end{itemize}\end{proof}

Below we will show in some detail a result, which proves the last statement of the above outline.


\begin{theorem}Let $H$\space be a clique graph with a $\gamma$ -cluster structur...
...h.p. $\vert E(H) \Delta E(G)\vert < \vert E(C) \Delta E(G)\vert$ .
\end{theorem}


\begin{proof}% latex2html id marker 130
(outline) \\
First we will show that gi...
... E(H) \Delta E(G)\vert < \vert E(C) \Delta E(G)\vert$ ;
\end{itemize}\end{proof}


  
Figure: Schematic proof of the key observation. This figure depicts a graph G (blank), original clique graph H (green) and some other clique graph C (green with pattern). Consider $E(H) \Delta E(G)$ and $E(C) \Delta E(G)$. Each edge (u,v) which belongs to neither C nor to H, but belongs to G increments both $E(H) \Delta E(G)$ and $E(C) \Delta E(G)$. Each edge (u,v) which belongs to both C and H, but does not belong to G also increments both $E(H) \Delta E(G)$ and $E(C) \Delta E(G)$. Thus, the difference between $E(H) \Delta E(G)$ and $E(C) \Delta E(G)$ is due to patches that correspond to $E(H) \Delta E(C)$. Let us denote by $\Delta ^{*}$ the symmetric difference constrained to edges in $E(C) \Delta E(H)$. But $\vert E(C) \Delta ^{*} E(G)\vert + \vert E(H) \Delta ^{*} E(G)\vert = \vert E(H) \Delta E(C)\vert$. Thus $\vert E(C) \Delta ^{*} E(G)\vert < \vert E(H) \Delta ^{*} E(G)\vert$ iff $\vert E(H) \Delta ^{*} E(G)\vert > \frac {\vert E(H) \Delta E(C)\vert}{2}$. Hence, $\vert E(C) \Delta E(G)\vert < \vert E(H) \Delta E(G)\vert$ iff $\vert E(H) \Delta ^{*} E(G)\vert > \frac {\vert E(H) \Delta E(C)\vert}{2}$.
\includegraphics{lec12_fig/keyObservation.eps}



Peer Itsik
2001-02-01