Please note that in the spirit of collaboration I shared with some of my classmates the helper
function \texttt{drawgraph(...)} used in exercise 3 and included in the file \texttt{bisection\_spectral.m}, and the functions \texttt{runtest(...)} and \texttt{recurse(...)} in file \texttt{bench\_rec\_bisection.m} for exercise 4. I am therefore claiming here I am the sole author of these functions.
Please also note this sharing was done with Edoardo Vecchi's supervision.
In figure \ref{fig:run1} there are graph outputs respectively from \textit{Grid9}, \textit{Small}, and \textit{Eppstein}. Please note that these results can be reproduced using script \texttt{ex3\_plots.m}.
Colors to represent the partitions and the eigenvector components are the same as the ones used in the the assignment's Figure 3. It is quite easy to evince from these plots that spectral partitioning indeed partitions the graph verticies based on the sign of their respective Fiedler eigenvector component: the only intersection between the Graph plane and the surface the eigenvector entries lay corresponds to the edgecut produced by spectral partitioning. This is due to the fact that in the figure the graph is positioned at $z =0$.
Again the data shown here can reproduced using a MATLAB script (\texttt{Bench\_rec\_bisection.m}). Please note that the given recursion helper function \texttt{rec\_bisection} was not used. Instead, the function \texttt{runtest} was implemented to allow for partitioning graphs to be generated at will.
\section{Compare recursive bisection to direct $k$-way partitioning\punkte{10}}
I summarize my results in table \ref{table:Compare_Metis}. Additionally, plots for the $n =32$ cases can be found in figure \ref{fig:bikmetis}. Please note that both the table and the plots can be reproduced using script \texttt{Bench\_metis.m}.
I had the suspicion that k-way partitioning would perform better in all cases, since in this the partitioning algorithm has a chance to see the entire problem and to perform decision that could affect all partitions. Instead, for the \texttt{Crack} mesh, recursive partitioning surprisedly performed better than k-way for both $n =16$ and $n =32$, producing slighly smaller edge cuts.
\begin{figure}[h]
\begin{subfigure}{0.5\textwidth}
\centering
\resizebox{0.8\linewidth}{!}{
\input{ex5_air_n}}
\caption{\textit{Airfoil} with recursive partitioning}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\resizebox{0.8\linewidth}{!}{
\input{ex5_air_k}}
\caption{\textit{Airfoil} with k-way partitioning}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\resizebox{0.8\linewidth}{!}{
\input{ex5_crack_n}}
\caption{\textit{Crack} with recursive partitioning}
\end{subfigure}
\begin{subfigure}{0.5\textwidth}
\centering
\resizebox{0.8\linewidth}{!}{
\input{ex5_crack_k}}
\caption{\textit{Crack} with k-way partitioning}
\end{subfigure}
\caption{Graph outputs for \textit{METIS} recursive and k-way partitioning with $n=32$. Edge cuts are displayed in black}
\caption{Comparing the number of cut edges for recursive bisection and direct multiway partitioning in Metis 5.0.2. Results for recursive partitioning shown on the left, k-way on the right.}
