08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

we decrease the residual length <strong>of</strong> the clause by one.<br />

At a general stage, suppose the fates <strong>of</strong> i variables have already been decided and n−i<br />

remain. The residual length <strong>of</strong> each clause is known. Among the clauses that are not yet<br />

satisfied, choose a random shortest length clause. Among the n − i variables remaining,<br />

pick one uniformly at random, then pick it or its negation as the new literal. Include<br />

this literal in the clause thereby satisfying it. Since the clause is satisfied, the algorithm<br />

deletes it. For each other clause, do the following. If its residual length is l, decide with<br />

probability l/(n − i) to include the new variable in the clause and if so with probability<br />

1/2 each, include it or its negation. If literal that was set to true is included in a clause,<br />

delete the clause as it is now satisfied. If its negation is included in a clause, then just<br />

delete the literal and decrease the residual length <strong>of</strong> the clause by one.<br />

Why does this yield the same distribution as the original one? First, observe that the<br />

order in which the variables are picked by the method <strong>of</strong> deferred decisions is independent<br />

<strong>of</strong> the clauses; it is just a random permutation <strong>of</strong> the n variables. Look at any one clause.<br />

For a clause, we decide in order whether each variable or its negation is in the clause. So<br />

for a particular clause and a particular triple i, j, and k with i < j < k, the probability<br />

that the clause contains the i th , the j th , and k th literal (or their negations) in the order<br />

determined by deferred decisions is:<br />

( ) ( ) ( )<br />

1 −<br />

3 1 −<br />

3 · · · 1 −<br />

3 3<br />

n<br />

(<br />

1 −<br />

2<br />

n−i<br />

(<br />

1 − 1<br />

n−j<br />

n−1<br />

) (<br />

1 −<br />

2<br />

n−i−1<br />

) (<br />

1 − 1<br />

n−j−1<br />

)<br />

· · ·<br />

(<br />

n−i+2<br />

n−i+1<br />

1 − 2<br />

n−j+2<br />

)<br />

)<br />

· · · (1<br />

− 1<br />

n−k+2<br />

2<br />

n−j+1<br />

)<br />

1<br />

n−k+1 = 3<br />

n(n−1)(n−2) ,<br />

where the (1 − · · · ) factors are for not picking the current variable or negation to be included<br />

and the others are for including the current variable or its negation. Independence<br />

among clauses follows from the fact that we have never let the occurrence or nonoccurrence<br />

<strong>of</strong> any variable in any clause influence our decisions on other clauses.<br />

Now, we prove the lemma by appealing to the method <strong>of</strong> deferred decisions to generate<br />

the formula. T i = t if and only if the method <strong>of</strong> deferred decisions does not put the current<br />

literal at steps 1, 2, . . . , t − 1 into the i th clause, but puts the negation <strong>of</strong> the literal at<br />

step t into it. Thus, the probability is precisely<br />

1<br />

2<br />

(<br />

1 −<br />

3<br />

n<br />

) (<br />

1 −<br />

3<br />

n−1<br />

)<br />

· · ·<br />

(<br />

1 −<br />

3<br />

n−t+2<br />

)<br />

3<br />

n−t+1 ≤ 3<br />

2(n−2) ,<br />

as claimed. Clearly the T i are independent since again deferred decisions deal with different<br />

clauses independently.<br />

Lemma 4.22 There exists a constant c 2 such that with probability 1 − o(1), no clause<br />

remains a 2 or 1-clause for more than c 2 ln n steps. I.e., once a 3-clause becomes a<br />

2-clause, it is either satisfied or reduced to a 0-clause in O(ln n) steps.<br />

112

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!