• Home
  • Textbooks
  • The Design of Approximation Algorithms
  • Further Uses of Randomized Rounding of Semidefinite Programs

The Design of Approximation Algorithms

David P. Williamson, David B. Shmoys

Chapter 13

Further Uses of Randomized Rounding of Semidefinite Programs - all with Video Answers

Educators


Chapter Questions

01:37

Problem 1

Show that the graph coloring algorithm of Section 13.2 can be used to color a 3 -colorable graph with $\bar{O}\left(n^{1 / 4}\right)$ colors.

Nick Johnson
Nick Johnson
Numerade Educator

Problem 2

In this problem, we consider an unweighted version of the minimum multicut problem from Section 8.3. We are given a graph $G=(V, E)$ and $k$ pairs of source-sink vertices, $s_i, t_i \in V$ for $i=1, \ldots, k$. We wish to find a subset of edges $F \subseteq E$ that minimizes $|F|$ such that for each $i=1, \ldots, k$, there is no $s_i-t_i$ path in $(V, E-F)$.
Consider the following vector program:

$$
\begin{array}{lrl}
\operatorname{minimize} & \sum_{(i, j) \in E}\left(1-v_i-v_j\right) & \\
\text { subject to } & \begin{array}{rll}
v_{s_i}-v_{\tau_i} & =0, & \\
v_j-v_j & =1, & \\
v_j & \in \Re^a, & \\
& \forall j \in V,
\end{array}, \\
& & \forall j \in V .
\end{array}
$$

Consider the demand graph $H=\left(V, E^{\prime}\right)$, where $E^{\prime}=\left\{\left(s_i, t_i\right): i=1, \ldots, k\right\}$. Let $\Delta$ be the maximum degree of a vertex in the demand graph. Suppose that that the optimal value of the vector programming relaxation is $\epsilon|E|$. Consider the following algorithm which is similar to the algorithm in Section 6.5 for coloring 3 -colorable graphs. We draw $\ell=\left\lceil\log _2(\Delta / t)\right\rceil$ random vectors $\tau_1, \ldots, \tau_t$. The $\ell$ random vectors define $2^t$ different regions into which the vectors $v_i$ can fall: one region for each distinct possibility of whether $r_j \cdot v_i \geq 0$ or $r_j \cdot v_i<0$ for all $j=1, \ldots, t$. Remove all edges ( $i, j$ ) from the graph such that $v_i$ and $v_j$ are in different regions. If for any $s_i-t_i$ pair, there still exists an $s_i-t_i$ path, remove all edges incident on $s_i$. We now analyze this algorithm.
(a) Prove that the vector program is a relaxation of the unweighted minimum multicut problem.
(b) For any $(i, j) \in E$, prove that the probability that $i$ and $j$ are in different regions is at most $t \cdot \sqrt{1-v_i-v_j}$.
(c) Prove that for any $i=1, \ldots, k$, the probability that we end up removing all the edges incident on $i$ is at most $\Delta 2^{-t}$.
(d) Show that the expected number of edges removed is at most $O(\sqrt{\epsilon} \log (\Delta / \epsilon))|E|$.

For the final item, it may be useful to use Jensen's inequality, which states that for any convex function $f$ (that is, $f^{\prime \prime}(x) \geq 0$ ) and any positive $p_i$,

$$
f\left(\frac{\sum_i p_i x_i}{\sum_i p_i}\right) \leq \frac{1}{\sum_i p_i} \sum_i p_i f\left(x_i\right)
$$

The arithmetic-geometric mean inequality given in Fact 5.8 is a special case of Jensen's inequality with $f(x)=-\log x$ and all the $p_i=1$.

Check back soon!

Problem 3

In this problem, we consider another algorithm for the unweighted minimum multicut problem from the previous exercise, using the same vector programming relaxation as above. As before, assume that the optimal value of the vector programming relaxation is $\epsilon|E|$, and that $\Delta$ is the maximum degree of a vertex in the demand graph. We set a threshold $\alpha=C \sqrt{\ln (\Delta / \epsilon)}$ for some constant $C$. Suppose we draw a random vector $r$. Consider the set $S(\alpha)=\left\{i \in V: \nu_i-r \geq \alpha\right\}$, and $S^{\prime}(\alpha)=S(\alpha)-\bigcup_{i=1}^k\left\{s_i, t_i: s_i, t_i \in S(\alpha)\right\}$; that is, $S^{\prime}(\alpha)$ is all elements in $S(\alpha)$ except for the $s_i, t_i$ pairs that both end up in $S(\alpha)$.
The following inequality will be useful for the analysis of the algorithm. Let $v_i$ and $v_j$ be unit vectors, and let $r$ be a random vector. Let $\alpha$ be a quantity such that $\alpha>1$ and $\bar{\Phi}(\alpha)<1 / 3$. Then

$$
\operatorname{Pr}\left[v_i \cdot r \geq \alpha \text { and } v_j \cdot r<\alpha\right]=O\left(\sqrt{v_i \cdot v_j} \Phi(\alpha) \sqrt{\log (1 / \Phi(\alpha))}\right)
$$
(a) Give a randomized algorithm that works by repeatedly drawing random vectors $r$ and considering the set $S^{\prime}(\alpha)$ as defined above. Show that the expected size of the multicut returned by the algorithm is at most

$$
\sum_{(i, j) \in E} \frac{\operatorname{Pr}\left[(i, j) \in \delta\left(S^{\prime}(\alpha)\right)\right]}{\operatorname{Pr}\left[i \in S^{\prime}(\alpha) \text { or } j \in S^{\prime}(\alpha)\right]}
$$

(b) Prove that for any $(i, j) \in E$,

$$
\frac{\operatorname{Pr}\left[(i, j) \in \delta\left(S^{\prime}(\alpha)\right)\right]}{\operatorname{Pr}\left[i \in S^{\prime}(\alpha) \text { or } j \in S^{\prime}(\alpha)\right]} \leq O\left(\sqrt{v_i \cdot v_j}\right) \alpha+\epsilon .
$$

(c) Show that the randomized algorithm returns a multicut of size $O(\sqrt{\epsilon \log (\Delta / \epsilon)})|E|$ edges in expectation.

Check back soon!