Advanced Algorithm - Weighted Set Cover

An example of randomized rounding from LP

呈前一講,我們對有限制的 WSC 問題有 $l$-approx. algorithm,但可能並不是太好(比方說 $l=|U|$ 之類),於是我們嘗試引進一些隨機性,雖然犧牲了 deterministic 的 approx. ratio ,有時候甚至會得到更爛的結果(甚至不滿足 constraint @@),但可以證明在大部分時候,都可以得到一個還不錯 (approx. ratio ISN'T too bad)的結果。

Weighted Set Cover Problem

讓我們明確定義這個問題

Given $U = \lbrace e_1,e_2,\cdots,e_n \rbrace, , S_1,S_2,\cdots S_k \subseteq U , \text{with} , \text{cost},(S_i)$

Want: Find a min-cost collection of subsets $C$ such that $\underset{C}{\bigcup} S_i = U$

Formulation

Let $x_i = $𝟙[subset $S_i$ is chosen or not]

\[ \min \sum_i x_i \cdot \text{cost}(S_i) \quad s.t. \begin{cases} \sum_{e_j \in S_i} x_i \geq 1 \;& , \forall j=1 \sim n \;(\because \text{每個element 都至少要被蓋到一次})\\ 0 \leq x_i \leq 1 \; & , \cancel{x_i \in \mathbb{Z}} \;(\text{LP Relaxation}) \end{cases} \]

利用 LP solver 得到的解 $\mathbf{x}^{\star} = (x_1^{\star},x_2^{\star},\cdots,x_k^{\star})$ 當作選擇 $\mathbf{S_i}$ 進 collection 的機率

Lemma

\[ \Pr[C \, \text{is a set cover} \, \land \, \text{cost}(C) \leq \mathbf{OPT} \cdot c_1 \, \log n] \geq \frac{1}{2} \]

cost 很大的機率

At each round, $S_i$ will be chosen with prob. $x_i^{\star}$ independently, and we repeat $c \log n$ rounds

因此,

$$ \mathbb{E}[\, \text{cost in 1 round} \,] = \, \sum_{i=1}^{k} \text{cost}(S_i) \cdot \Pr[ S_i \, \text{is chosen}] = \sum_{i=1}^{k} \, \text{cost}(S_i) \cdot x_i^{\star} \leq \mathbf{OPT} $$

$$ \Rightarrow \mathbb{E}[\, \text{cost of Alg. output} \,] \leq c \log n \cdot \mathbf{OPT} $$

\[ \text{According to \textbf{Markov Ineq.}} \quad \boxed{\Pr[X \geq a] \leq \frac{\mathbb{E}[X]}{a}} \]

$$ \Rightarrow \boxed{\Pr[\, \text{cost of Alg. output} \, \geq \color{blue}{4} \, \mathbf{OPT} \cdot c \log n\,] \leq \frac{1}{\color{blue}{4}}} $$

collection $C$ 不是 set cover 的機率

For each elemetn $e_j$

\[ \Pr[e_j \, \text{is not covered in round 1}] \, = \underset{e_j \in S_i}{\prod} (1 - x_i^{\star}) \underset{\text{見以下}}{\leq} \underset{e_j \in S_i}{\prod} e^{-x_i^{\star}} = \exp(-\underset{e_j \in S_i}{\sum} x_i^{\star}) \underset{\text{constraint}}{\leq} e^{-1} \]
\[ \text{之後很常用到的不等式} \; \boxed{ 1-x \leq e^{-x}} \; \forall x \geq 0 \]

所以在 $c \log n$ 個 round 內, $e_j$ 都不會被選到的機率 $ \leq (\frac{1}{e})^{c \log n}$,選擇適當的 $c$ ,使得該機率 $\leq \frac{1}{4n}$, 再利用 Union Bound,將單一 element 不被選到的機率加總

\[ \Rightarrow \boxed{\Pr[\text{Alg. miss some element}] \leq \frac{1}{4}} \]

再一次利用 Union Bound,將這兩件壞事的機率相加,可以得到其 $\leq \frac{1}{2}$,得證!

※ 如果造得出 $\frac{1}{2}$,對於任意給定的常數 error tolerance $\epsilon$ ,我們都可以藉由調控 $c$ (也就是要 run 多少 round) 得到 (但怪怪的地方是說,因 approx. ratio 並非 constant ,而是 $\mathcal{\Theta}(\log n)$,所以 $c$ 直接被忽略,才可以 在降低非法解的可能時,同時保證 approx. ratio)

 
comments powered by Disqus