Minimum Cost Flow

In this section I will briefly walk over my understanding of min cost flow. I would highly recommend reading through CP-Algorithm's Min Cost Flow to understand the solution idea first. Additionally, check the TopCoder tutorial for a more detailed explanation.

Triangle Inequality

In graph theory, the Triangle Inequality states that if there is an edge $u\to v$ with weight $w$, and $d_k$ be the shortest path to node $k$ (for some reasonable definition of shortest path), then $d_v-d_u\le w$.

Johnson's Algorithm

The main idea of Johnson's Algorithm is that if all edge weights are positive, then running Dijkstra's from each node would result in a $\mathcal{O}(VE\log E)$ algorithm. If there are any negative edges, Johnson's Algorithm defines a potential function $\pi$, such that for every edge $u,v,w$, the following holds: $w>=\pi (u)-\pi (v)$. Then, each edge weight can be transformed into $w\to w+\pi (v)-\pi (u)$, resulting in positive weight. This condition coincides with the Triangle Inequality, so we can arbitrarily pick a node and run a $\mathcal{O}(VE)$ shortest path algorithm to determine this function.

Minimum Cost Flow

The general idea of Min Cost Flow is to repeatedly push flow along the shortest path. Since flow graphs have negative edges, each step naively would take $\mathcal{O}(VE)$ time. To speed it up, we can use the same potential function from Johnson's Algorithm to employ Dijkstra for this process. In this case we must use distance from $S$, the source node, as the $\pi$ function. At each step, run Dijkstra's using the $\pi$ function, update the $\pi$ function to match the current distances, and then push flow along the shortest path, reversing edges as needed. Once flow is met or the sink is unreachable, terminate.

Implementation

With all this being said, here is my implementation.

1template<int MN, int MM>
2struct MCF//MN = nodes, MM = edges [assume edges one-directional]
3{
4public:
5    int N, M, S, T;
6    int flow[MM*2], cap[MM*2], hd[MN], nx[MM*2], to[MM*2], cost[MM*2];
7    int pi[MN], p[MN], d[MN];
8    int vis[MN];
9    void init(int n, int s, int t)
10    {

Also check out Benq's Implementation and KACTL's Implementation (which are a lot better than mine).

Problems

Applications

Assignment Problem