# Sum over products of weighted subset of certain size

Consider a commutative semiring (R,+,\cdot). \mathbb{0} is the identity for (R,+), and \mathbb{1} is the identity for (R,\cdot). Let f,g:V\to R, w:V\to \mathbb{N} and Z\subset \mathbb{N}. It is common that we are interested in computing expressions of the following form.

\displaystyle \sum_{S\subset V, \sum_{x\in S} w(x) \in Z} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x)

Examples:

1. If w(x)=1 for all x, and f(x) be the probability that event x occurs, g=1-f, we find the probability that the number of event occurs t times, where t\in Z. In probability, this is computing the Poisson distribution.

2. If (R,+,\cdot) = (\N,+,\cdot), f=g=1, for all x and w(x)=x and V\subset \N and Z=\{t\}, then we find the number of subsets that have element sum t.

3. If (R,+,\cdot) = (\N,\max,+), V\subset \N, g=0 and Z=\{0,\ldots,W\}, then this solves the knapsack problem with knapsack size W, value f and cost w.

4. An actual application inspired this post: An automated test suite that runs n subtests, and it is allowed to rerun a subtest if it fails the first time. A subtest passes if first run passes or the rerun passes. The test is successful if all the subtests passes and the number of total reruns is at most k. Assume probability of passing is independent for each subtest. One want to estimate the probability of a successful test given the probability a run passes for a specific subtest.

Let \max Z = k and |V| = n. The naive algorithm runs in O(n2^n) time (assuming semiring operation takes O(1) time). There is a common transformation that turns this problem that sum over all subsets to a problem that sums over Z. So it runs in O(nk) time.

Let V=\{v_1,\ldots,v_n\} and V_j = \{v_1,\ldots,v_j\}. Define \displaystyle D(i,j) = \sum_{S\subset V_j, \sum_{x\in S} w(x) = i} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x) .

Certainly, \displaystyle \sum_{S\subset V, \sum_{x\in S} w(x) \in Z} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x) = \sum_{i\in Z} D(i,n)

We only incur a O(k) number of semiring operations once we compute all D(i,n) for 0\leq i\leq k.

Let [P] be the Iverson bracket notation, namely

\displaystyle [P] = \begin{cases} \mathbb{1} & \text{if } P \text{ is true;}\\ \mathbb{0} & \text{otherwise.} \end{cases}

Theorem1
1. D(i,0) = [i \neq 0]

2. For j\geq 1, D(i,j) = [i\geq w(v_j)] f(v_j)D(i-w(v_j),j-1) + g(v_j) D(i,j-1).

Proof

The base case can be verified easily, we show part of a inductive step.

\displaystyle \begin{aligned} f(v_j)D(i-w(v_j),j-1) + g(v_j)D(i,j-1) &= f(v_j) \sum_{S\subset V_{j-1}, \sum_{x\in S} w(x) = i-w(v_j)} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x) + g(v_j) \sum_{S\subset V_{j-1}, \sum_{x\in S} w(x) = i)} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x) \\ &= \sum_{v_j\in S\subset V_j, \sum_{x\in S} w(x) = i} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x) + \sum_{v_j\not\in S\subset V_j, \sum_{x\in S} w(x) = i} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x)\\ &= \sum_{S\subset V_j, \sum_{x\in S} w(x) = i} \prod_{x\in S} f(x) \prod_{x\in V\backslash S} g(x) \end{aligned}

Posted by Chao Xu on 2014-08-11.
Tags: algorithm.