Algorithms - Lecture 10 - Minimum cost perfect matching

1 Definition

• A matching is a set of edges without common vertices. Or, each vertex has a degree $\le 1$.
• $G=(V,E)$ is a bipartite graph that does not contain any odd-length cycles.
  • Assume $V$ can be partitioned in to $X$ and $Y$, such that $|X|=|Y|=n$, and each edge has one endpoint in $X$ and one endpoint in $Y$.
  • Assume $G$ admits a perfect matching, i.e., a matching of cardinality $n$.
• Want to compute a minimum-cost perfect matching of $G$.

2 Algorithm

2.1 Residual graph

Definition of residual graph $G_M$

$G_M$ is V ∪ { s , t } V \cup \{s,t\} V∪{s,t}, for each vertex $x$ in $X$

• if $x$ is not matched in $M$, add directed edge $(s,x)$ with cost zero
• if $y$ is not matched in $M$, add directed edge $(y,t)$ with cost zero
• for each edge $e=(x,y)\in M$, $G_M$ contains a directed edge from $y$ to $x$ with cost $-c_e$
• for each edge $e=(x,y)\in E/M$, $G_M$ contains a directed edge from $x$ to $y$ with cost $c_e$

Augmenting paths - Simple directed paths in $G_M$

• There is a one-to-one correspondence between (simple) directed $s-t$ paths in $G_M$ and augmenting paths in $G$ w.r.t. $M$.
  • Let $M^{\prime }$ be a perfect matching of $G$.
  • $M\oplus M^{\prime }$ corresponds to a collection of vertex-disjoint alternating paths and cycles in $G$. (it’s easy to see that the symmetric difference induces degree at most 2 on any vertex. And a graph with degree at most 2 has a very simple structure; it can only contain paths and cycles.)
  • Since $|M^{\prime }|>|M|$, this collection includes an augmenting path $P$ in $G$ w.r.t. $M$.

2.2 Alternating cycles

Simple directed cycles in $G_M$ - Alternating cycles in $G$

There is one-to-one correspondence between simple directed cycles in $G_M$ and alternating cycles in $G$ w.r.t. perfect matching $M$.

Characterizing Minimum-Cost perfect matchings

$M$ is a perfect matching of $G$ $\leftrightarrow$ no negative-cost cycle in $G_M$.

Since each vertex is matched in two perfect matchings $M,M^{\prime }$, thus the symmetric difference $M\oplus M^{\prime }$ result in that each vertex has a degree in {0,2}, thus there cannot be paths (since the beginning and end of vertex of a path has a degree=1).

2.3 Vertex prices and Reduced costs in $G_M$

Definition

• Maintain a price $p(v)$ for each vertex $v$ in $V+s$
• For each edge in $G_M$ not incident on $t$
  • Each directed edge $(s,x)$ is assigned a reduced cost $p(s)-p(x)$
  • For each edge $e=(x,y)$
    • if $e=(x,y)\in G_M$, reduced cost $=c_e+p(x)-p(y)$
    • if $e=(y,x)\in G_M$, reduced cost $=-c_e+p(y)-p(x)$

Nonnegative reduced costs

• $G_M$ does not contain any negative-cost directed cycles
• Use Dijkstra’s SSSP algorithm to compute minimum-cost s-t path in $G_M$

2. Iterative framework

• From $M_i$ to $M_{i+1}$, identify an augmenting path $P_i$ and set $M_{i+1}=M_i\oplus P_i$.
  • $P_i$ is a simple path in $G$ of odd length, with one endpoint in $X$ and the other endpoint in $Y$.
  • Edges of $P_i$ alternate between edges in $M_i$ and edges not in $M_i$.
  • Neither endpoint is matched in $M_i$, so $P_i$ begins and ends with an edge that is not in $M_i$.
• The cost of $M_{i+1}$ is equal to the cost of $M_i$ plus the cost of $s\to t$ of $P_i$ in $G_M$.