# Minimum of submodular function over family of subsets

Theorem1

Let L and L' are two lattices. If f:L \to \R is a submodular function and P:L'\to 2^{L} is a function with the property that if X\in P(A) and Y\in P(B), then X\wedge Y\in P(A\wedge B) and X\vee Y\in P(A\vee B). f_P:L'\to \R defined as \displaystyle f_P(X) = \min_{Y\in P(X)} f(Y)\\ is submodular.

Proof

Let X^* = \argmin_{Y\in P(X)} f(Y), note since X^*\in P(X) and Y^*\in P(Y), we have X^*\vee Y^* \in P(X\vee Y) and X^*\vee Y^* \in P(X\wedge Y). \displaystyle \begin{aligned} f_P(X) + f_P(Y) &= f(X^*) + f(Y^*)\\ &\geq f(X^* \vee Y^*) + f(X^*\wedge Y^*)\\ &\geq f((X\vee Y)^*) + f((X\wedge Y)^*)\\ &= f_P(X\vee Y) + f_P(X\wedge Y) \end{aligned}

This is quite useful, for starters, it proves that we can create a monotone submodular function from any submodular function.

Theorem2

Let f:2^V\to \R be a submodular function, then f_*,f^*:2^V\to \R defined as \displaystyle f_*(X) = \min \{f(Y)|Y\subset X\}\\ f^*(X) = \min \{f(Y)|X\subset Y\} are monotone and submodular.

A practical application is to generalize the cut function. Consider for a directed graph graph G=(V,E). We would define \delta^+(A) to be the set of out going edges from A to V\setminus A. f=|\delta^+| is a submodular function. An alternate definition for f is the minimum number of edges to be removed so there is no path from A to V\setminus A.

A simple generalization is when we only care about T\subset V. We can define f_T(A) to be the minimum number of edges to be removed so there is no path from A to T\setminus A. Amazingly(or not not surprisingly, depending on your intuition), f_T is also a submodular function by invoking the next theorem, which is a direct corollary of our first theorem.

Theorem3

Let f:2^V\to \R be a submodular function, then f_T:2^T\to \R defined as \displaystyle f_T(X) = \min \{f(Y)|Y\subset X, T\setminus X\subset V\setminus Y\}\\ is submodular.

Posted by Chao Xu on 2014-08-13.
Tags: submodular.