## Problem 1

For any $\alpha\ge 1$, a cut $C$ in an undirected (multi)graph $G(V,E)$ is called an α-min-cut if $|C|\le\alpha|C^*|$ where $C^*$ is a min-cut in $G$.

1. Give a lower bound to the probability that Karger’s Random Contraction algorithm returns an α-min-cut in a graph $G(V,E)$ of n vertices.

2. Use the above bound to estimate the number of distinct α-min cuts in $G$.

### My Solution

1.From Karger’s algorithm we know that the maximum size of α-min-cut C is α|C*|, the degree of multi-graph $G_i:degree(G_i) = 2|E_i| \geq |V_i| |C^*| \ge |V_i|\cdot|C|/ \alpha$ ,so the probability $p_i =\Pr[e_i \notin C \vert \forall j < i,e_i \notin C]$ can be computed as:

$$\begin{array} {lcl} p_i & = & 1 - \frac{\vert C \vert}{\vert E_i \vert} \\ & \geq & 1 - \frac{2\alpha}{|V_i|} \\ & = & 1- \frac{2\alpha}{n-i+1} \end{array}$$

and further, apply the Karger’s algorithm until vertices remain:

$$\begin{array} {lcl} p_{\text{survives until 2α vertices remain}} & \geq & \Pr[\,C_{2α}\mbox{ is returned by }{RandomContract}\,]\\ & = & \Pr[\,e_i\not\in C\mbox{ for all }i=1,2,\ldots,n-2α\,]\\ & = & \prod_{i=1}^{n-2α}\Pr[e_i\not\in C\mid \forall j<i, e_j\not\in C]\\ & \geq & \prod_{i=1}^{n-2α}\left(1-\frac{2α}{n-i+1}\right)\\ & = & \frac{(n-2α)!}{n!/(2α)!} \\ & = & \frac{(n-2α)!(2α)!}{n!} \\ & = & \frac{1}{n \choose 2α} \end{array}$$

then, choose a random cut in the remaining multi-graph:G

$$\begin{array} {lcl} p_{\text{correct}} & = & \Pr[\text{C survives until 2α vertices remain}] \times \Pr[\text{C survives random cut}]\\ & \geq & \frac{1}{n \choose 2α} \times \frac{2}{2^{2α}}\\ & = & \frac{(2α)!}{2^{2α-1}} \times \frac{(n-2α)!}{n!}\\ & \geq & \frac{(2α)!}{2^{2α-1}} \times \frac{1}{n^{2α}}\\ & \geq & \frac{1}{n^{2α}} \end{array}$$

So, a lower bound to the probability that Karger’s Random Contraction algorithm returns an α-min-cut in a graph G(V,E) of n vertices can be $n^{-2α}$.

2.By the analysis of Karger’s algorithm, we know $p_C\ge\frac{1}{n^{2α}}$. And since $p_{correct}$ is a well defined probability, due to the unitarity of probability, it must hold that $p_{\text{correct}}\le 1$. Therefore:

$$1 \ge p_{correct} = \sum_{C \in \sf{C}}p_C \ge |\sf{C}| \frac{1}{n^{2α}}$$

and this means that $|\sf{C}| \le n^{2α}$, So the up bound of the number of distinct α-min cuts in G is $n^{2α}$.

## Problem 2

Let $G(V,E)$ be an undirected graph with positive edge weights $w:E\to\mathbb{Z}^+$. Given a partition of $V$ into $k$ disjoint subsets $S_1,S_2,\ldots,S_k$, we define $w(S_1,S_2,\ldots,S_k)=\sum_{uv\in E\atop \exists i\neq j: u\in S_i,v\in S_j}w(uv)$ as the cost of the k-cut ${S_1,S_2,\ldots,S_k}$. Our goal is to find a k-cut with maximum cost.

1. Give a poly-time greedy algorithm for finding the weighted max $k-cut$. Prove that the approximation ratio is $(1 − 1 / k)$.

2. Consider the following local search algorithm for the weighted max cut (max 2-cut).

Fill in the blank parenthesis. Give an analysis of the running time of the algorithm. And prove that the approximation ratio is 0.5.

start with an arbitrary bipartition of $V$ into disjoint $S_0,S_1$;

while (true) do

if $\exists i\in{0,1}$ and $v\in S_i$ such that (__)

then $v$ leaves $S_i$ and joins $S_{1 − i}$;

continue;

end if

break;

end

### My Solution

1.(1) algorithm

$S_1={v_1},S_2={v_2},\ldots,S_k={v_k}$,$V =V - {v_1,v_2,\ldots,v_k}$ ;

while ( V is not empty) do:

Randomly select a vertic $v$ in V;

$V = V- v$

put $v$ to $S_j$ for the biggest increased value

(2) proof:

$$\begin{array} {lcl} E(w) & = & \sum_{(i,j) \in E} w_{i,j} \Pr[i \in V_m,j\in V_n, \forall m \neq n] \\ & = & \sum_{(i,j) \in E} w_{i,j} (1 - \Pr[i \in V_m,j\in V_m, \forall m ]) \\ & = & \sum_{(i,j) \in E} w_{i,j} (1 - \sum_{m=1}^k\Pr[i \in V_m,j\in V_m]) \\ & = & \sum_{(i,j) \in E} w_{i,j} (1 - \sum_{m=1}^k\frac{1}{k^2}) \\ & = & (1-\frac{1}{k})\sum_{(i,j) \in E} w_{i,j} \\ & \ge & (1-\frac{1}{k})OPT \end{array}$$

2.(1) the blank: $\underline{v \in S_{1-i} \ has\ bigger\ cut \ result}$

(2) analysis: calculate the cut between $S_0,S_1$ is $O(n^2)$, and in each loop the value is increaed at least 1, thus the maximum possible cut value is $\sum_{e \in E}w_e$,then the loop size is $\sum_{e \in E}w_e$ at most. finally the total running time is $O(n^2 \cdot \sum_{e \in E}w_e)$.

(3) proof:

• the OPT can’t be larger than the sum of all egde weights, then : $\sum_{e \in E} w_e\geq OPT$

• if we move $v$ from $S_i \ to \ S_{1-i}$, then $\sum_{v \in S_{1-i}, (u,v) \in E}w(u,v) \geq \sum_{v \in S_{i}, (u,v) \in E}w(u,v)$, according to the mean value theorem, we have $\sum_{v \in S_{1-i}, (u,v) \in E}w(u,v) \geq \frac{1}{2}\sum_{v :(u,v) \in E}w(u,v)$, and for any $u’ \in S_{1-i}$ , we can get $\sum_{v \in S_i, (u’,v) \in E}w(u’,v) \geq \frac{1}{2}\sum_{v :(u’,v) \in E}w(u’,v)$ similarly.

Finally, for all vertices in V, the max-cut is :

$$\begin{array} {lcl} V(S) &= & \sum_{u \in S_i,v \in S_{1-i},(u,v) \in E } \\ & = & \sum_{v \in S_i, (u,v) \in E}w(u,v) + \sum_{v \in S_{1-i}, (u,v) \in E}w(u,v) \\ & \ge & \frac{1}{2}\sum_{v :(u’,v) \in E}w(u,v) + \frac{1}{2}\sum_{v :(u,v) \in E}w(u,v) \\ & = & \frac{1}{2} \sum_{e \in E}w_e \\ & \ge &\frac{1}{2}OPT \end{array}$$

## Problem 3

Given $m$ subsets $S_1,S_2,\ldots, S_m\subseteq U$of a universe $U$ of size $n$, we want to find a $C\subseteq{1,2,\ldots, n}$ of fixed size $k = | C |$ with the maximum coverage $\left|\bigcup_{i\in C}S_i\right|$.

• Give a poly-time greedy algorithm for the problem. Prove that the approximation ratio is $1 − (1 − 1 / k)k > 1 − 1 / e$.

### My Solution

algorithm:

GreedyCover

Input: sets $S_1,S_2,\ldots,S_m$;

initially, $U=\bigcup_{i=1}^mS_i$, and $C=\emptyset$, count = 0;

while $U\neq\emptyset$ and count < k do

find $i\in{1,2,\ldots, m}$ with the largest $|S_i\cap U|$;

let $C=C\cup{i}$and $U=U- S_i$;

count = count + 1

return C;

proof: let $c_i$ denote the elements of selected $S_j$ in the i-th round , and $C_i=\sum_{i=1}^ic_i$, the optimum solution is $OPT$, the remaining elements in $OPT$ is $R_i$, we have: $R_i = OPT - C_i,R_0 = OPT, C_0 =0$.

At each round , the Greedy algorithm selects the subset $S_j$ with maximum size of uncoverd elements yet, there exists one set that cover at least $\frac{1}{k}$ fraction of the remaning uncovered elements $R_i$, so we have $c_{i+1} \ge \frac{R_i}{k}$.

On the other hand, consider the conclusion that we want to prove:$C_k = OPT - R_k \ge (1- (1-\frac{1}{k})^k)\cdot OPT$, that equals to: $R_k \leq (1- \frac{1}{k})^k \cdot OPT$,it is hard to prove it directly, we use the technique of mathematical induction: $R_i \leq (1- \frac{1}{k})^i \cdot OPT$

• for i =0, $R_0 \le OPT$, it is true
• suppose i, it is true: $R_i \le (1 - \frac{1}{k})^i\cdot OPT$
• when it comes to i+1: $R_{i+1} \leq R_{i} - c_{i+1} \leq R_i - \frac{R_i}{k} = R_i(1- \frac{1}{k}) \le (1- \frac{1}{k})^{i+1}\cdot OPT$
• we have proved it!

## Problem 4

We consider minimum makespan scheduling on parallel identical machines when jobs are subject to precedence constraints.

We still want to schedule $n$ jobs $j=1,2,\ldots, n$ on $m$ identical machines, where job $j$ has processing time $p_j$. But now a partial order $\preceq$ is defined on jobs, so that if $j\prec k$ then job $j$ must be completely finished before job $k$ begins. The following is a variant of the List algorithm for this problem: we still assume that the input is a list of $n$ jobs with processing times $p_1,p_2,\ldots, p_n$.

whenever a machine becomes idle

assign the next available job on the list to the machine;

Here a job $k$ is available if all jobs $j\prec k$ have already been completely processed.

• Prove that the approximation ratio is 2.

### My Solution

proof:

• with precedence constraints,it is known that: $$OPT \ge \max_{\mathcal{A}:i \preceq j}p_{\mathcal{A}}, OPT\ge\frac{1}{m}\sum_{j=1}^np_j$$

• with greedy stratege,it is known that the max completion time is less than the sum of the max processing time $p_{max}$ in precedence constraints and the average processing time $p_{mean}$ among m identical machines, which is: $$C_{max} \le \max_{\mathcal{A}:i \preceq j}p(\mathcal{A}) + \frac{1}{m}\sum_{j=1}^np_j = 2\cdot OPT$$

## Problem 5

For a hypergraph $H(V,E)$ with vertex set $V$, every hyperedge $e\in E$ is a subset $e\subset V$ of vertices, not necessarily of size 2. A hypergraph $H(V,E)$ is k-uniform if every hyperedge $e\in V$ is of size $k = | e |$.

A hypergraph $H(V,E)$ is said to have property B (named after Bernstein) if $H$ is 2-coloable; that is, if there is a proper 2-coloring $f:V \to \{ {\color{red}{R},\color{blue}{B}} \}$ which assigns each vertex one of the two colors $\color{red}{Red}$ or $\color{blue}{Blue}$, such that none of the hyperedge is monochromatic.

1. Let $H(V,E)$ be a k-uniform hypergraph in which every hyperedge $e\in E$ shares vertices with at most d other hyperedges.

• Prove that if $2\mathrm{e}\cdot (d+1)\le 2^{k}$, then $H$ has property B.

• Describe how to use Moser’s recursive Fix algorithm to find a proper 2-coloring of $H$. Give the pseudocode. Prove the condition in in terms of $d$ and $k$ under which the algorithm can find a 2-coloring of $H$ with high probability.

• Describe how to use Moser-Tardos random solver to find a proper 2-coloring of H. Give the pseudocode. Prove the condition in in terms of d and k under which the algorithm can find a 2-coloring of $H$ within bounded expected time. Give an upper bound on the expected running time.

2. Let $H(V,E)$ be a hypergraph (not necessarily uniform) with at least $n\ge 2$ vertices satisfying that $\forall v\in V, \sum_{e\ni v}(1-1/n)^{-|e|}2^{-|e|+1}\le \frac{1}{n}$.

• Prove that $H$ has property B.

• Describe how to use Moser-Tardos random solver to find a proper 2-coloring of $H$. Give an upper bound on the expected running time.

### My Solution

1.(1)proof:

• if $2\mathrm{e}\cdot (d+1)\le 2^{k}$, it equal to $\mathrm{e}\cdot \frac{1}{2^{k-1}} (d+1)\le 1 \qquad(1)$, on the other hand, the size of each edge $|e| = k$, let the probability of one edge cannot be 2-coloable in edge $e_i$ is $Pr[e_i]= 2 / 2^k = 1/2^{k-1} \qquad (2)$, it is a bad thing. From the two formulas above, we know that it meets Lovász Local Lemma condition . So the probability that not all bad things happen is greater than 0. In our problem, it means H has property B.

(2)algorithm:

• $\phi$: k-CNF of max degree d with m clauses on n variables, $|e| = m$, for clause $C_i$, bad event $A_i: C_i$ is not satisfied, which is the vertices in edge is colored the same color.

• $n$ variables: $x_1,x_2,\ldots,x_n, \ n$ is the total number of vertices, $x_i=\{0,1\}$

• For each clause C in φ, we denote by $\mathsf{vbl}(C)\subseteq\mathcal{X}$ the set of variables on which C is defined.

• We also abuse the notation and denote by $\Gamma(C)=\{\text{clause }D\text{ in }\phi\mid D\neq C, \mathsf{vbl}(C)\cap\mathsf{vbl}(D)\neq\phi\}$ the neighborhood of C, i.e. the set of other clauses in φ that shares variables with C, and $\Gamma^+(C)=\Gamma(C)\cup\{C\}$ the inclusive neighborhood of C,

Solve($\phi$)

1. Pick values of $x_1,x_2\ldots,x_n$ uniformly and independently at random;

2. while ∃ unsatisfied clause C in φ :

Fix ($C$);

Fix(C):

1. Replace the values of variables in $\mathsf{vbl}(C)$ with uniform and independent random values;

2. while ∃ unsatisfied clause D overlapping with C:

Fix(D);

• proof:
• first, for the edges number m, we have $m \le 2^{k-1}$, that is $k \ge \log_2^m +1$, the proof is as follows: the probability of one edge cannot be 2-coloable in edge $e_i$ is $Pr[e_i]= 2 / 2^k = 1/2^{k-1}$, therefore the probability of the bad 2-coloring is:$Pr(\bigwedge_{i=1}^m e_i)\le \sum_{i=1}^mPr(e_i)=\frac{m}{2^{k-1}} \le 1$, so $k \ge \log_2^m +1 \qquad(1)$.
• second, using Moser’s recursive Fix algorithm, there are m recursion trees, t total nodes, total of random bits is: $n+tk$, the sequence of random bits can be recoverd from final assignment+ reciursion trees, for each recursion tree, the root uses $\left \lceil \log_2m\right \rceil$ bits, each internal node uses at most : $\log_2d + c$ bits, which is $m\left \lceil \log_2m\right \rceil + t(\log_2d + c)$, with the Incompressibility Theorem(N uniform random bits cannot be encoded to substantially less than N bits.), we have $n+tk \le m\left \lceil \log_2m\right \rceil + t(\log_2d + c)$, and it is equal to $t(k-c-\log_2d \le m\left \lceil \log_2m\right \rceil + \log_2n)$, and $k-c-\log_2d >0$, we get : $d\le 2^{k-c} \qquad(2)$, with Lovász Local Lemma, we know that $ep(d+1) \leq 1$, and here is $p = 1/2^{k-1}$, so $d \le \frac{2^{k-1}}{e} - 1\qquad(3)$ ,and the probability the algorithm can find a 2-coloring of H is $n+tk=O(n+km\log_2m)$ .
• using (1) ,(2) and (3), we can get the conditions of d and k: $k \ge \log_2^m +1$,$d \le \frac{2^{k-1}}{e} - 1$

(3)algorithm:

• $\mathcal{X}$ is a set of mutually independent random variables.

RandomSolver:

1. sample all $X \in \mathcal{X}$;

2. while there is non-violated bad event $A \in \mathcal{A}$:

resample all $X \in vbl(A)$;

• proof:
• from above problem, we know that $k \ge \log_2^m +1 \qquad(1)$.
• with Lovász Local Lemma, we know that $d \le \frac{2^{k-1}}{e} - 1\qquad(2)$
1. (1) proof(unsolved):
• let $k= |e|$, m edges, n vertices, $n \ge 2$, max degree of dependency graph: d, then we have $2(1-1/n) \ge 1$

• $\forall v\in V, \frac{1}{n} \geq \sum_{e \ni v}(1-\frac{1}{n})^{-|e|}2^{-|e|+1} = \sum_{e \ni v}2(2(1-1/n))^{-|e|}= 2 \sum_{i=1}^n \frac{k_i}{2(2-1/n)^{k_i}}$

• $\forall i, p = \Pr[A_i] \le \frac{2}{2^{\min(k)}} \le \frac{1}{2}, d \le \max(k_i)$

(2) algorithm(unsolved)