08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Pro<strong>of</strong>: Say that t is a “busy time” if there exists at least one 2-clause or 1-clause at time<br />

t, and define a time-window [r + 1, s] to be a “busy window” if time r is not busy but<br />

then each t ∈ [r + 1, s] is a busy time. We will prove that for some constant c 2 , with<br />

probability 1 − o(1), all busy windows have length at most c 2 ln n.<br />

Fix some r, s and consider the event that [r + 1, s] is a busy window. Since SC always<br />

decreases the total number <strong>of</strong> 1 and 2-clauses by one whenever it is positive, we must<br />

have generated at least s − r new 2-clauses between r and s. Now, define an indicator<br />

random variable for each 3-clause which has value one if the clause turns into a 2-clause<br />

between r and s. By Lemma 4.21 these variables are independent and the probability that<br />

a particular 3-clause turns into a 2-clause at a time t is at most 3/(2(n − 2)). Summing<br />

over t between r and s,<br />

Prob ( a 3-clause turns into a 2-clause during [r, s] ) ≤<br />

3(s − r)<br />

2(n − 2) .<br />

Since there are cn clauses in all, the expected sum <strong>of</strong> the indicator random variables is<br />

cn 3(s−r) ≈ 3c(s−r) . Note that 3c/2 < 1, which implies the arrival rate into the queue <strong>of</strong> 2<br />

2(n−2) 2<br />

and 1-clauses is a constant strictly less than one. Using Chern<strong>of</strong>f bounds, if s−r ≥ c 2 ln n<br />

for appropriate constant c 2 , the probability that more than s−r clauses turn into 2-clauses<br />

between r and s is at most 1/n 3 . Applying the union bound over all O(n 2 ) possible choices<br />

<strong>of</strong> r and s, we get that the probability that any clause remains a 2 or 1-clause for more<br />

than c 2 ln n steps is o(1).<br />

Now, assume the 1 − o(1) probability event <strong>of</strong> Lemma 4.22 that no clause remains a<br />

2 or 1-clause for more than c 2 ln n steps. We will show that this implies it is unlikely the<br />

SC algorithm terminates in failure.<br />

Suppose SC terminates in failure. This means that at some time t, the algorithm<br />

generates a 0-clause. At time t − 1, this clause must have been a 1-clause. Suppose the<br />

clause consists <strong>of</strong> the literal w. Since at time t − 1, there is at least one 1-clause, the<br />

shortest clause rule <strong>of</strong> SC selects a 1-clause and sets the literal in that clause to true.<br />

This other clause must have been ¯w. Let t 1 be the first time either <strong>of</strong> these two clauses,<br />

w or ¯w, became a 2-clause. We have t − t 1 ≤ c 2 ln n. Clearly, until time t, neither <strong>of</strong> these<br />

two clauses is picked by SC. So, the literals which are set to true during this period are<br />

chosen independent <strong>of</strong> these clauses. Say the two clauses were w + x + y and ¯w + u + v<br />

at the start. x, y, u, and v must all be negations <strong>of</strong> literals set to true during steps t 1 to<br />

t. So, there are only O ( (ln n) 4) choices for x, y, u, and v for a given value <strong>of</strong> t. There are<br />

O(n) choices <strong>of</strong> w, O(n 2 ) choices <strong>of</strong> which two clauses i, j <strong>of</strong> the input become these w<br />

and ¯w, and n choices for t. Thus, there are O ( n 4 (ln n) 4) choices for what these clauses<br />

contain and which clauses they are in the input. On the other hand, for any given i, j,<br />

the probability that clauses i, j both match a given set <strong>of</strong> literals is O(1/n 6 ). Thus the<br />

probability that these choices are actually realized is therefore O ( n 4 (ln n) 4 /n 6) = o(1),<br />

as required.<br />

113

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

Saved successfully!

Ooh no, something went wrong!