\input{template}
\input{macros}

\usepackage{color, graphicx}
\usepackage{amssymb, amsmath}
\usepackage{epsfig}


\begin{document}
\lecture{26}{Primal-Dual algorithm for shortest path problem
}{Pragya Goyal}
%% -----------------------------------------------------------------------------
%%                          Beginning of Section 1.
%% -----------------------------------------------------------------------------

\section{Recap of last lecture}
In last class we had formed the ILP for the Shortest Path problem
as follows:
\begin{equation}
\text{min}\sum_e w_{uv} x_{uv}
\end{equation}
\begin{equation}
\sum_u x_{su} = 1
\end{equation}
\begin{equation}
\sum_v x_{vt} = 1
\end{equation}
\begin{equation}
\forall p \sum_q x_{pq} - \sum_r x_{rp} = 0
\end{equation}
\begin{equation}
1 \geq x_{uv} \geq 0
\end{equation}
\begin{equation}
x_{uv} \text{is an integer}
\end{equation}
Expression(1)is the summation of weight of all the edges selected,
where $x_{i}$ takes value $1$ if edge $(u,v)$ is selected otherwise
remains $0$. Equations (2) and (3), represent the constraints that
in the shortest path the source s must have exactly one edge
leaving it and destination t must have exactly one edge entering
into it, respectively. Equation (4) is
written to satisfy the constraint that all the intermediate
vertices (i.e. apart from s and t) in the path must have one edge
entering and one edge leaving them. Inequality(5) is to ensure
that $x_{uv}$ takes values in $[0,1]$. To reduce the problem to LP we
drop the $x_{uv}\leq 1$ constraint. This can be done because equalities (2) and (3) along with $x_{uv}$ constraint ensure that each $x_{uv}$ does not take value greater than 1.\newline\newline
 Now let us look at the dual.
We will be solving the dual and will observe that it is Dijkstra's
algorithm.\newline\newline

\section{Solving the Dual}
In the dual all the constraints of Primal are represented by
variables. Here we have one variable for each vertex. Hence dual
can be represented as:
\begin{equation}
\text{max } y_s - y_t
\end{equation}
\begin{equation}
 \forall u,v \hspace{10mm} y_v - y_u \leq w_{uv}
\end{equation}

Now let us try solving the dual. Here $y_i$ can be thought of variable corresponding to ith vertex. To maximize $y_s - y_t$ let us
begin with assuming $y_t = 0$. Keeping $y_t$ fixed at 0 we raise all other variables by 1 at a time until for atleast one of them, say $y_1$, $y_1 - y_t = w_{1t}$. at this point stop increasing the value of $y_1$ and fix it to the last reached value. Repeat the procedure keeping $y_t$ and $y_1$ fixed now. Include any new vertex $y_2$ as soon as $y_2-y_1$ equals $w_{12}$. We iterate over all vertex variables until we reach $y_s$. We stop the iteration when $y_s$ is reached. \newline\newline


\textbf{\emph{Marble-string analogy:}}\newline
 The procedure is similar to fixing a marble tied to
many strings and then pulling the strings apart. Constraint being
limited length of the strings. Fixed marble is a vertex in this
case, strings are the edges entering the vertex and string lengths
correspond to edge weights. We can observe here that the string
with least length (which corresponds to least weight edge entering
fixed vertex) will become taut first. Becoming taut means equality
is satisfied in the constraint. Select this edge. Now fix the
vertex from which this edge emanates. Apply similar raising method
for this vertex. We iterate over the vertices until we reach the
source vertex. The edges selected at each iteration form the
shortest path from s to t. This algorithm is \textbf{Dijkstra's
algorithm}.\newline\newline

\subsection{Reverse Delete}
The edges we selected in the above procedure contain the shortest path along with some additional edges. To remove those edges we apply the procedure of \emph{Reverse Delete}. This is done by starting deletion of maximum weight edges, each time assuring that a path exists from s to t. As soon as we reach an edge, deleting which leaves no path from s to t, we stop the procedure. The path left (including this last egde seen) is the required shortest path.


\end{document}