Proof of KΓΆnig's Theorem
KΓΆnig's theorem establishes the fundamental duality between matchings and vertex covers in bipartite graphs. It is a cornerstone of combinatorial optimization and linear programming duality.
Statement
In a bipartite graph , the maximum matching size equals the minimum vertex cover size: .
Proof
. For any matching and vertex cover : each edge of has at least one endpoint in , and these endpoints are distinct (since is a matching). So , hence .
. We construct a vertex cover of size from a maximum matching .
Let be a maximum matching. Define (unmatched vertices on the left). Let be the set of all vertices reachable from by alternating paths (paths that alternate between non-matching and matching edges).
Define: (reachable left vertices), (reachable right vertices).
Claim: is a minimum vertex cover.
is a vertex cover. Consider any edge with :
- If (i.e., ): , so the edge is covered.
- If : is reachable from by an alternating path. If : the path to extended by reaches , so . If : is matched to . Since is reachable, and (matched vertices are not in ), was reached via some with , then ... Actually, the alternating path to ends with a matching edge into , meaning the previous vertex is in , and . But is matched to by , so , giving .
. Each matching edge contributes exactly one vertex to : if with : (and ). If : (and ). We cannot have : since is matched and reachable, the alternating path goes (with , so ), meaning , contradiction. We cannot have : reachable means an alternating path reaches via a non-matching edge from some , then if the path continues to , so . Since is matched to in , , contradiction.
So each matching edge has exactly one endpoint in , and no vertex of is unmatched (unmatched vertices in are in , hence not in ; unmatched vertices in are not reachable, hence not in ). Thus .
KΓΆnig's theorem is a special case of LP duality. The maximum matching LP: subject to , . The dual: subject to for , . For bipartite graphs, the LP polytope is integral (totally unimodular constraint matrix), so LP optimum = IP optimum = .