Lecture 10. Paradigm #8: Randomized Algorithms Back to the “majority problem” (finding the majority element in an array A). FIND-MAJORITY(A, n) while (true) do randomly choose 1 ≤ I ≤ n; if A[i] is the majority then return (A[i]); If there is a majority element, we get it with probability > ½ each round. The expected number of rounds is 2

Computing ∑ s in S Pr[s] s Each time a new index of the array is guessed, we find the majority element with probability p > 1/2. So the expected number of guesses is p*1 + (1-p)p*2 + (1-p) 2 p* = Σ i ≥ 1 (1-p) i-1 ⋅ p ⋅ i = x. To evaluate this, multiply x by (1-p) and add y = ∑ i ≥ 1 (1-p) i p. We get ∑ i ≥ 1 (1-p) i p (i+1), which is just x-p. Now, using the sum of a geometric series, Σ i ≥ 0 ar i = a (1-r ∞ )/ (1-r) y evaluates to 1-p. So (1-p)x + (1-p) = x-p. So x = 1/p < 2

Another example: order statistics Problem: Given a list of numbers of length n, and an integer i (with 1 ≤ i ≤ n), determine the i-th smallest member in the list. This problem is also called "computing order statistics". Special cases: computing the minimum corresponds to i = 1, computing the maximum corresponds to i = n, computing the 2nd largest corresponds to i = n-1, and computing the median corresponds to i = ceil(n/2). Obviously you could compute the i'th smallest by sorting, but that takes O(n log n) time, which is not optimal. We know how to solve it when i = 1, n-1, n in linear time. We now give a randomized algorithm to do selection in expected linear time. By expected (or average) linear time, we mean the average over all random choices used by the algorithm, and not depending on the particular input.

Randomized selection Remember the PARTITION algorithm for Quicksort in CS240. RANDOMIZED-PARTITION chooses a random index j to partition. RANDOMIZED-SELECT(A,p,r,i) /* Find the i'th smallest element from the array A[p..r] */ if p = r then return A[p]; q := RANDOMIZED-PARTITION(A,p,r); k := q-p+1; /* number of elements in left side of partition */ if i = k then return A[q] ; else if i < k then return RANDOMIZED-SELECT(A,p,q-1,i) ; else return RANDOMIZED-SELECT(A,q+1,r,i-k);

Expected running time: O(n) Expected running time: T(n) ≤ (1/n)*Σ 1 ≤ k ≤ n T(max(k-1,n-k)) + O(n) Prove T(n) ≤ cn by induction. Assume it is true for n ≥ 3. Now assume it is true for n' < n. Then T(n) ≤ ( (1/n)*Σ 1 ≤ k ≤ n T(max(k-1,n-k)) ) + O(n) ≤ ((2/n)*Σ floor(n/2) ≤ k ≤ n-1 T(k)) + O(n) ≤ ((2/n)*Σ floor(n/2) ≤ k ≤ n-1 ck) + an (for some a ≥ 1) = ( (2c/n)*Σ floor(n/2) ≤ k ≤ n-1 k) + an = ( (2c/n)*(Σ 1 ≤ k ≤ n-1 k - Σ 1 ≤ k ≤ floor(n/2)-1 k ) + an ≤ (2c/n)* ( n(n-1)/2 - (n/2 - 1)(n/2-2)/2) + an = (2c/n)*(n 2 - n - n 2 /4 + n + n/2 - 2)/2 + an = (c/n)*(3n 2 /4 + n/2 - 2) + an ≤ 3cn/4 + c/2 + an < cn for c large enough.

Example: Matrix multiplication Given nxn matrices, A,B,C, test: C=AxB Trivial O(n 3 ) Strassen O(n log7 ) Best deterministic alg O(n ) Consider the following randomized alg: repeat k times generate a random nx1 vector x in {-1,1} n ; if A(Bx) ≠ Cx, then return AB ≠ C; return AB = C;

Error bound. Theorem. This algorithm errs with prob ≤ 2 -k. Proof. If C’=AB ≠ C, then for some i, j, c ij ≠ ∑ l=1..n a il b lj =c’ ij Then C’x ≠ Cx for either x j = 1 or x j = -1. Thus with probability ½, one round will tell if AB≠C. After k rounds, error probability is ≤ 2 -k.

Some comments about Randomness You might object: in practice, we do not use truly random numbers, but rather "pseudo- random" numbers. The difference may be crucial. For example, this is one way a pseudorandom number is generated X n+1 = a X n + c (mod m) As von Neumann said in 1951: “Anyone who considers arithmetical methods of producing random digits is, of course, in a state of sin."

Which sequence is random? If you go to a casino, you flip a coin they provided, and you get 100 heads in a row. You complain to the casino: “Your coin is biased, H 100 is not a random sequence!” The casino owner would demand: “You give me a random sequence.” You confidently flipped your own coin 100 times and obtained a sequence S, and tell the casino that you would trust S to be random. But casino can object: the two sequences have the same probability! Why is S more random than H 100

How to define Randomness Laplace once said: a sequence is “extraordinary” (non-random) when it contains regularity. Solomonoff, Kolmogorov, Chatin (1960’s): a random sequence is a sequence that is not (algorithmically) compressible (by a computer / Turing machine). Remember, we have used such sequences to obtain the average case complexity of algorithms. In your assignment 2, Problem 2, you have shown: random sequences exist (in fact, most sequences are random sequences).