* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Lecture No. 10(A) : Method of Conditional Probabilities 1 - CSE-IITM
Birthday problem wikipedia , lookup
Computational electromagnetics wikipedia , lookup
Computational complexity theory wikipedia , lookup
Hardware random number generator wikipedia , lookup
Simplex algorithm wikipedia , lookup
Factorization of polynomials over finite fields wikipedia , lookup
Dijkstra's algorithm wikipedia , lookup
Simulated annealing wikipedia , lookup
Clique problem wikipedia , lookup
Drift plus penalty wikipedia , lookup
IITM-CS6845: Theory Toolkit March 9, 2011 Lecture No. 10(A) : Method of Conditional Probabilities Lecturer: N. S. Narayanaswamy Scribe: Sajin Koroth Theme: Explicit Constructions Lecture Plan: Today we will be seeing the method of conditional probabilities. It is more of an application than an elaborate theory like d-wise independent sample spaces which we saw in the last lecture. We will motivate the method by applying it to a non-trivial problem, i.e to obtain 2-coloring of a complete graph to get at most nr 2r monochromatic 2(2) r-cliques. We will see one more application, which is a maximization problem(unlike the earlier one which tries to minimize cliqes), k-SAT. We will end the lecture by generalizing the method of conditional probabilities to a framework which incorporates all the above mentioned applications. 1 Explicit Construction of 2-coloring of a complete graph with bounded monochromatic r-cliques using method of Conditional Probabilities Recall that in the very first lecture to obtain a lowerbound on the Ramsey number R(r, r) we used the following probabilistic method argument. Let us consider a random 2-coloring of the edge set of complete graph, Kn . The expected number of monochromatic r-cliques are r n 2 E [#r − cliques] = ∵ Two of the 2(2) colorings are monochromatic r r 2(2) r Hence if nr 21−(2) < 1, then there is at least one coloring where there are no monochromatic r-clique, hence for all such n, r, R(r, r) > n. The above analysis also guarantees that for r any complete graph Kn , there is a 2-coloring of its edge set such there are at most nr 21−(2) monochromatic r-cliques. We are interested in constructing such a 2-coloring. Note that this task is not obvious, no naive approaches like color everything red or color everything blue would not work, as you need to control both red r-cliques and blue r-cliques. But note that if we analyze the coloring step by step, then we can write expression for conditional probability of introducing an r-monochromatic clique, by coloring the ei+1 th edge given a coloring of first i-edges in some fixed ordering of the edges. Let Wi denote the number of No. 10(A)-1 monochromatic r-cliques in the partially colored graph Gi where only the edges, e1 , . . . , ei are colored. Hence the we can write E[Wi+1 ] as E[Wi+1 ] = E[Wi+1 |ei+1 = red]Pr[ei+1 = red] + E[Wi+1 |ei+1 = blue]Pr[ei+1 = blue] E[Wi+1 |ei+1 = red] + E[Wi+1 |ei+1 = blue] = 2 Hence we get that, E[Wi+1 ] ≥ min {E[Wi+1 |ei+1 = red], E[Wi+1 |ei+1 = blue]} So if we color the ei+1 ’th edge to the one color which achieves the minimum in the above expression, we are guaranteeing that the number of monochromatic cliques in our graph is at most the expected number of monochromatic cliques in a random partial coloring. Hence after coloring all the edges we get to a coloring where the number of monochromatic r r-cliques is at most nr 21−(2) . The costliest step in our explicit construction is checking the number of monochromatic r-cliques in the partially colored graph after fixing the color of ith edge to a specific color, for both the colors. The naive algorithm for checking r-cliques takes time nr and hence if r is a constant our explicit construction algorithm runs in time polynomial. 2 Explicit construction for a maximization problem Consider the following problem, MAX-SAT, which given a 3 − CN F formula on n variables and m clauses which is such that in any clause the literals appearing are distinct, asks for an assignment to the variables x1 , . . . , xn which maximizes the number of satisfied clauses. Let Xj , 1 ≤ j ≤ m be the indicator random variable denoting whether the jth clause is satisfied by the random assignment or not. Since the three literals appearing in the jth clause are distinct, out of the 23 = 8 possible values of these literals only one (0, 0, 0) is bad. P Hence E[Xj ] = 87 . Let X denote the total number of P satisfied clauses, by definition X = j Xj . Hence using linearity of expectation E[X] = j E[Xj ] = 7m 8 . We are interested in, given a MAX-SAT instance φ construct an assignment which satisfies at least 7m 8 clauses. The construction is very similar to the construction of above. But the subtle difference is in the fact that unlike the earlier problem this is a maximization problem. Let us analyze a random assignment of truth values to x1 , . . . , xn step by step. Let Wi denote the expected number of satisfied clauses in a partial assignment which fixes only x1 , . . . , xi . Similar to the above problem we can write Wi+1 as E[Wi+1 ] = E[Wi+1 |ei+1 = 0]Pr[ei+1 = 0] + E[Wi+1 |ei+1 = 1]Pr[ei+1 = 1] E[Wi+1 |ei+1 = 0] + E[Wi+1 |ei+1 = 1] = 2 No. 10(A)-2 Hence, E[Wi+1 ] ≤ max {E[Wi+1 |ei+1 = 0], E[Wi+1 |ei+1 = 1]} So if we assign xi+1 such that it maximizes the above expression we guarantee that the number of satisfied clauses are at least that of the expected number of satisfied clauses. Note that to find the number of satisfied clauses is easy as you just have to go through each clause and see if at least one of the three literals is set to 1, which a naive approach solves in time O(nm). Hence the construction is a polynomial time construction. 3 Generalized Conditional Probability Method Consider the uniform distribution on {0, 1}l . Let A1 , . . . , As be a set of events such that s X Pr [Ai ] = k i=1 If we define indicator random variable Xi for the event Ai then the above condition tells us that " s # X E [X] = E Xi i=1 = = s X i=1 s X E [Xi ] (∵ Linearity of Expectation) Pr [Ai ] i=1 = k That is the expected number of events which happen from {A1 , . . . , As } is k. Hence this also tells us that there is at least one point in {0, 1}l where the number of events which happen from {A1 , . . . , As } is at most k. To obtain that point algorithmically we can use the method of conditional probabilities as follows. Consider the fixation of xi , 1 ≤ i ≤ l to either 1 or 0 based on whether Pr[Ai |x1 = b1 , . . . , xi−1 = bi−1 , xi = 0] is the minimum or Pr[Ai |x1 = b1 , . . . , xi−1 = bi−1 , xi = 1] is the minimum (where b1 , . . . , bi−2 are the values fixed by the algorithm in stages 1, . . . , i − 1). This will give us a point in {0, 1}l in which No. 10(A)-3 at most k events from A1 , . . . , As will happen because, 1 Pr[Ai |x1 · · · xi−1 = b1 · · · bi−1 , xi = 0] + 2 1 Pr[Ai |x1 · · · xi−1 = b1 · · · bi−1 , xi = 1] 2 ≥ min{Pr[Ai |x1 · · · xi−1 = b1 · · · bi−1 , xi = 0], Pr[Ai |x1 · · · xi−1 = b1 · · · bi−1 ] = Pr[Ai |x1 · · · xi−1 = b1 · · · bi−1 , xi = 1]} Also note that the algorithm maintains that ∀i, 1 ≤ i ≤ l, Pr[Ai |x1 · · · xi−1 = b1 · · · bi−1 ] ≥ Pr[Ai |x1 · · · xi = b1 · · · bi ] Hence, Pr[Ai ] ≥ Pr[Ai |x1 · · · xl = b1 · · · bl ] So we are guaranteed that at x1 · · · xl = b1 · · · bl at most k events occur. Also note that we could have ensured at least k events occur by choosing max instead of min in our algorithm. No. 10(A)-4