1. qsort - 1 Lin / Devi Comp 122 d(n)d(n) Today’s puzzle  How far can you reach with a stack of n blocks, each 2 units long?

2. qsort - 2 Lin / Devi Comp 122 Today’s puzzle  How far can you reach with a stack of n blocks, each 2 units long? d(n) = 1+ 1/2 + 1/3 + 1/4 + 1/5 + 1/6 … d(n)d(n) 1 1/2 1/3 1/4 1/5 1/6 nth harmonic number, Hn =  (lg n)

3. Comp 122, Spring 2004 Quicksort - Randomized

4. qsort - 4 Lin / Devi Comp 122 Quicksort: review Quicksort(A, p, r) if p < r then q := Partition(A, p, r); Quicksort(A, p, q – 1); Quicksort(A, q + 1, r) fi Quicksort(A, p, r) if p < r then q := Partition(A, p, r); Quicksort(A, p, q – 1); Quicksort(A, q + 1, r) fi Partition(A, p, r) x, i := A[r], p – 1; for j := p to r – 1 do if A[j]  x then i := i + 1; A[i]  A[j] fi od; A[i + 1]  A[r]; return i + 1 Partition(A, p, r) x, i := A[r], p – 1; for j := p to r – 1 do if A[j]  x then i := i + 1; A[i]  A[j] fi od; A[i + 1]  A[r]; return i + 1 5 A[p..r] A[p..q – 1] A[q+1..r]  5  5 Partition 5

5. qsort - 5 Lin / Devi Comp 122 Worst-case Partition Analysis Split off a single element at each level: T(n) = T(n – 1) + T(0) + PartitionTime(n) = T(n – 1) +  (n) =  k=1 to n  (k) =  (  k=1 to n k ) =  (n 2 ) n n – 1 n – 2 n – 3 2 1 n Recursion tree for worst-case partition

6. qsort - 6 Lin / Devi Comp 122 Best-case Partitioning  Each subproblem size  n/2.  Recurrence for running time »T(n)  2T(n/2) + PartitionTime(n) = 2T(n/2) +  (n)  T(n) =  (n lg n) cn cn/2 cn/4 cccccc lg n

7. qsort - 7 Lin / Devi Comp 122 Variations  Quicksort is not very efficient on small lists.  This is a problem because Quicksort will be called on lots of small lists.  Fix 1: Use Insertion Sort on small problems.  Fix 2: Leave small problems unsorted. Fix with one final Insertion Sort at end. »Note: Insertion Sort is very fast on almost-sorted lists.

8. qsort - 8 Lin / Devi Comp 122 Unbalanced Partition Analysis What happens if we get poorly-balanced partitions, e.g., something like: T(n)  T(9n/10) + T(n/10) +  (n)? Still get  (n lg n)!! (As long as the split is of constant proportionality.) Intuition: Can divide n by c > 1 only  (lg n) times before getting 1. n  n/c  n/c 2    1= n/c log c n Roughly log c n levels; Cost per level is O(n). (Remember: Different base logs are related by a constant.)  n n  n n  n n

9. qsort - 9 Lin / Devi Comp 122 Intuition for the Average Case  Partitioning is unlikely to happen in the same way at every level. »Split ratio is different for different levels. (Contrary to our assumption in the previous slide.)  Partition produces a mix of “good” and “bad” splits, distributed randomly in the recursion tree.  What is the running time likely to be in such a case?

10. qsort - 10 Lin / Devi Comp 122 Intuition for the Average Case n 0 n – 1 (n – 1)/2 – 1 (n – 1)/2 (n)(n) Bad split followed by a good split: Produces subarrays of sizes 0, (n – 1)/2 – 1, and (n – 1)/2. Cost of partitioning :  (n) +  (n-1) =  (n). n (n – 1)/2 (n)(n) Good split at the first level: Produces two subarrays of size (n – 1)/2. Cost of partitioning :  (n). Situation at the end of case 1 is not worse than that at the end of case 2. When splits alternate between good and bad, the cost of bad split can be absorbed into the cost of good split. Thus, running time is O(n lg n), though with larger hidden constants.

11. qsort - 11 Lin / Devi Comp 122 Randomized Quicksort  Want to make running time independent of input ordering.  How can we do that? »Make the algorithm randomized. »Make every possible input equally likely. Can randomly shuffle to permute the entire array. For quicksort, it is sufficient if we can ensure that every element is equally likely to be the pivot. So, we choose an element in A[p..r] and exchange it with A[r]. Because the pivot is randomly chosen, we expect the partitioning to be well balanced on average.

12. qsort - 12 Lin / Devi Comp 122 Variations (Continued)  Input distribution may not be uniformly random.  Fix 1: Use “randomly” selected pivot. »We’ll analyze this in detail.  Fix 2: Median-of-three Quicksort. »Use median of three fixed elements (say, the first, middle, and last) as the pivot. »To get O(n 2 ) behavior, we must continually be unlucky to see that two out of the three elements examined are among the largest or smallest of their sets.

13. qsort - 13 Lin / Devi Comp 122 Randomized Version Randomized-Partition(A, p, r) i := Random(p, r); A[r]  A[i]; Partition(A, p, r) Randomized-Partition(A, p, r) i := Random(p, r); A[r]  A[i]; Partition(A, p, r) Randomized-Quicksort(A, p, r) if p < r then q := Randomized-Partition(A, p, r); Randomized-Quicksort(A, p, q – 1); Randomized-Quicksort(A, q + 1, r) fi Randomized-Quicksort(A, p, r) if p < r then q := Randomized-Partition(A, p, r); Randomized-Quicksort(A, p, q – 1); Randomized-Quicksort(A, q + 1, r) fi Want to make running time independent of input ordering.

14. Comp 122, Spring 2004 Probabilistic Analysis and Randomized Algorithms

15. qsort - 15 Lin / Devi Comp 122 The Hiring Problem  You are using an employment agency to hire a new assistant.  The agency sends you one candidate each day.  You interview the candidate and must immediately decide whether or not to hire that person. But if you hire, you must also fire your current office assistant—even if it’s someone you have recently hired.  Cost to interview is c i per candidate.  Cost to hire is c h per candidate.  You want to have, at all times, the best candidate seen so far.  When you interview a candidate who is better than your current assistant, you fire the current assistant and hire the candidate.  You will always hire the first candidate that you interview.  Problem: What is the cost of this strategy?

16. qsort - 16 Lin / Devi Comp 122 Pseudo-code to Model the Scenario Hire-Assistant (n) best  0 ;;Candidate 0 is a least qualified sentinel candidate for i  1 to n do interview candidate i if candidate i is better than candidate best then best  i hire candidate i Hire-Assistant (n) best  0 ;;Candidate 0 is a least qualified sentinel candidate for i  1 to n do interview candidate i if candidate i is better than candidate best then best  i hire candidate i Cost Model: Slightly different from the model considered so far. However, analytical techniques are the same. Want to determine the total cost of hiring the best candidate. If n candidates interviewed and m hired, then cost is nc i +mc h. Have to pay nc i to interview, no matter how many we hire. So, focus on analyzing the hiring cost mc h. m varies with order of candidates.

17. qsort - 17 Lin / Devi Comp 122 Worst-case Analysis  In the worst case, we hire all n candidates.  This happens if each candidate is better than all those who came before. Candidates come in increasing order of quality.  Cost is nc i +nc h.  If this happens, we fire the agency. What should happen in the typical or average case?

18. qsort - 18 Lin / Devi Comp 122 Probabilistic Analysis  We need a probability distribution of inputs to determine average-case behavior over all possible inputs.  For the hiring problem, we can assume that candidates come in random order. »Assign a rank rank(i), a unique integer in the range 1 to n to each candidate. »The ordered list  rank(1), rank(2), …, rank(n)  is a permutation of the candidate numbers  1, 2, …, n . »Let’s assume that the list of ranks is equally likely to be any one of the n! permutations. »The ranks form a uniform random permutation. »Determine the number of candidates hired on an average, assuming the ranks form a uniform random permutation.

19. qsort - 19 Lin / Devi Comp 122 Randomized Algorithm  Impose a distribution on the inputs by using randomization within the algorithm.  Used when input distribution is not known, or cannot be modeled computationally.  For the hiring problem: »We are unsure if the candidates are coming in a random order. »To make sure that we see the candidates in a random order, we make the following change. The agency sends us a list of n candidates in advance. Each day, we randomly choose a candidate to interview. »Thus, instead of relying on the candidates being presented in a random order, we enforce it.

20. Comp 122, Spring 2004 Discrete Probability See Appendix C & Chapter 5.

21. qsort - 21 Lin / Devi Comp 122 Discrete probability = counting  The language of probability helps count all possible outcomes.  Definitions: »Random Experiment (or Process) Result (outcome) is not fixed. Multiple outcomes are possible. Ex: Throwing a fair die. »Sample Space S Set of all possible outcomes of a random experiment. Ex: {1, 2, 3, 4, 5, 6} when a die is thrown. »Elementary Event A possible outcome, element of S, x  S; Ex: 2 – Throw of fair die resulting in 2. »Event E Subset of S, E  S; Ex: Throw of die resulting in {x > 3} = {4, 5, 6} »Certain event : S »Null event :  »Mutual Exclusion Events A and B are mutually exclusive if A  B= .

22. qsort - 22 Lin / Devi Comp 122 Axioms of Probability & Conclusions  A probability distribution Pr{} on a sample space S is a mapping from events of S to real numbers such that the following are satisfied: »Pr{A}  0 for any event A. »Pr{S} = 1. (Certain event) »For any two mutually exclusive events A and B, Pr(A  B ) = Pr(A)+Pr(B).  Conclusions from these axioms: »Pr{  } = 0. »If A  B, then Pr{A}  Pr{B}. »Pr(A  B ) = Pr(A)+Pr(B)-Pr(A  B)  Pr(A)+Pr(B) »

23. qsort - 23 Lin / Devi Comp 122 Independent Events  Events A and B are independent if Pr{A|B} = Pr{A}, if Pr{B} != 0 i.e., if Pr{A  B} = Pr{A}Pr{B}  Example: Experiment: Rolling two independent dice. Event A: Die 1 < 3 Event B: Die 2 > 3 A and B are independent.

24. qsort - 24 Lin / Devi Comp 122 Conditional Probability  Example: On the roll of two independent dice, what is the probability of a total of 8? »S = {(1,1), (1,2), …, (6,6)} »|S| = 36 »A = {(2,6), (3,5), (4,4), (5,3), (6,2)} »Pr{A} = 5/36

25. qsort - 25 Lin / Devi Comp 122 Conditional Probability  Example: On the roll of two independent dice, if at least one face is known to be an even number, what is the probability of a total of 8? »One die is even, sum is 8  Second die is also even. »No. of elementary events in the original sample space: 36 »The fact that one face is even, precludes outcomes where both faces are odd. »Hence, sample space size is reduced to 27 (9 elementary events have both odd faces). »Number of elementary events in the reduced sample space that are successes : 3 --{(2,6), (4,4), (6,2)} »Hence, Probability is 3/27 = 1/9.

26. qsort - 26 Lin / Devi Comp 122 Conditional Probability  Formalizes the notion of having prior partial knowledge of the outcome of an experiment.  The conditional probability of an event A given that another event B occurs is defined to be  In the previous example »A: Event that sum on the faces is 8. »B: Event that one of them is even. »Pr{B} = 27/36 »Pr{A  B} = 3/36 »Pr{A|B} = Pr {A  B}/Pr{B} = 3/27 = 1/9.

27. qsort - 27 Lin / Devi Comp 122 Discrete Random Variables  A random variable X is a function from a sample space S to the real numbers.  If the space is finite or countably infinite, a random variable X is called a discrete random variable.  Maps each possible outcome of an experiment to a real number.  For a random variable X and a real number x, the event X=x is {s  S : X(s)=x }.  Pr{X=x} =  { s  S:X{s}=x} Pr{s}  f(x) = Pr{X=x} is the probability density function of the random variable X.

28. qsort - 28 Lin / Devi Comp 122 Discrete Random Variables  Example: »Rolling 2 dice. »X: Sum of the values on the two dice. »Pr{X=7} = Pr{(1,6),(2,5),(3,4),(4,3),(5,2),(6,1)} = 6/36 = 1/6.

29. qsort - 29 Lin / Devi Comp 122 Expectation  Average or mean  The expected value of a discrete random variable X is E[X] =  x x Pr{X=x}  Linearity of Expectation »E[X+Y] = E[X]+E[Y], for all X, Y »E[aX+Y]=aE[X]+E[Y], for constant a and all X, Y  For mutually independent random variables X 1, X 2, …, X n »E[X 1 X 2 … X n ] = E[X 1 ]E[X 2 ]…E[X n ]

30. qsort - 30 Lin / Devi Comp 122 Expectation – Example  Let X be the RV denoting the value obtained when a fair die is thrown. What will be the mean of X, when the die is thrown n times. »Let X 1, X 2, …, X n denote the values obtained during the n throws. »The mean of the values is (X 1 +X 2 +…+X n )/n. »Since the probability of getting values 1 thru 6 is (1/6), on an average we can expect each of the 6 values to show up (1/6)n times. »So, the numerator in the expression for mean can be written as (1/6)n·1+(1/6)n·2+…+(1/6)n·6 »The mean, hence, reduces to (1/6)·1+(1/6)·2+…(1/6)·6, which is what we get if we apply the definition of expectation.

31. qsort - 31 Lin / Devi Comp 122 Indicator Random Variables  A simple yet powerful technique for computing the expected value of a random variable.  Convenient method for converting between probabilities and expectations.  Helpful in situations in which there may be dependence.  Takes only 2 values, 1 and 0.  Indicator Random Variable for an event A of a sample space is defined as:

32. qsort - 32 Lin / Devi Comp 122 Indicator Random Variable Lemma 5.1 Given a sample space S and an event A in the sample space S, let X A = I{A}. Then E[X A ] = Pr{A}. Lemma 5.1 Given a sample space S and an event A in the sample space S, let X A = I{A}. Then E[X A ] = Pr{A}. Proof: Let Ā = S – A (Complement of A) Then, E[X A ] = E[I{A}] = 1·Pr{A} + 0·Pr{Ā} = Pr{A}

33. qsort - 33 Lin / Devi Comp 122 Indicator RV – Example Problem: Determine the expected number of heads in n coin flips. Method 1: Without indicator random variables. Let X be the random variable for the number of heads in n flips. Then, E[X] =  k=0..n k·Pr{X=k} We solved this last class with a lot of math.

34. qsort - 34 Lin / Devi Comp 122 Indicator RV – Example  Method 2 : Use Indicator Random Variables  Define n indicator random variables, X i, 1  i  n.  Let X i be the indicator random variable for the event that the i th flip results in a Head.  X i = I{the i th flip results in H}  Then X = X 1 + X 2 + …+ X n =  i=1..n X i.  By Lemma 5.1, E[X i ] = Pr{H} = ½, 1  i  n.  Expected number of heads is E[X] = E[  i=1..n X i ].  By linearity of expectation, E[  i=1..n X i ] =  i=1..n E[ X i ].  E[X] =  i=1..n E[ X i ] =  i=1..n ½ = n/2.

35. qsort - 35 Lin / Devi Comp 122 Randomized Hire-Assistant Randomized-Hire-Assistant (n) Randomly permute the list of candidates best  0 ;;Candidate 0 is a least qualified dummy candidate for i  1 to n do interview candidate i if candidate i is better than candidate best then best  i hire candidate i Randomized-Hire-Assistant (n) Randomly permute the list of candidates best  0 ;;Candidate 0 is a least qualified dummy candidate for i  1 to n do interview candidate i if candidate i is better than candidate best then best  i hire candidate i How many times do you find a new maximum?

36. qsort - 36 Lin / Devi Comp 122 Analysis of the Hiring Problem (Probabilistic analysis of the deterministic algorithm)  X – RV that denotes the number of times we hire a new office assistant.  Define indicator RV’s X 1, X 2, …, X n.  X i = I{candidate i is hired}.  As in the previous example, »X = X 1 + X 2 + …+ X n »Need to compute Pr{candidate i is hired}.  Pr{candidate i is hired} »i is hired only if i is better than 1, 2,…,i-1. »By assumption, candidates arrive in random order Candidates 1, 2, …, i arrive in random order. Each of the i candidates has an equal chance of being the best so far. Pr{candidate i is the best so far} = 1/i. E[X i ] = 1/i. (By Lemma 5.1)

37. qsort - 37 Lin / Devi Comp 122 Analysis of the Hiring Problem  Compute E[X], the number of candidates we expect to hire. By Equation (A.7) of the sum of a harmonic series. Expected hiring cost = O(c h ln n).

38. qsort - 38 Lin / Devi Comp 122 Analysis of the randomized hiring problem  Permutation of the input array results in a situation that is identical to that of the deterministic version.  Hence, the same analysis applies.  Expected hiring cost is hence O(c h ln n).

39. Comp 122, Spring 2004 Quicksort - Randomized

40. qsort - 40 Lin / Devi Comp 122 Avg. Case Analysis of Randomized Quicksort Let RV X = number of comparisons over all calls to Partition. Suffices to compute E[X]. Why? Notation: Let z 1, z 2, …, z n denote the list items (in sorted order). Let Z ij = {z i, z i+1, …, z j }. Let RV X ij = Thus, 1 if z i is compared to z j 0 otherwise X ij is an indicator random variable. X ij =I{z i is compzred to z j }.

41. qsort - 41 Lin / Devi Comp 122 Analysis (Continued) We have: Note: E[X ij ] = 0·P[X ij =0] + 1·P[X ij =1] = P[X ij =1] This is a nice property of indicator RVs. (Refer to notes on Probabilistic Analysis.) So, all we need to do is to compute P[z i is compared to z j ].

42. qsort - 42 Lin / Devi Comp 122 Analysis (Continued) z i and z j are compared iff the first element to be chosen as a pivot from Z ij is either z i or z j. Exercise: Prove this. So,

43. qsort - 43 Lin / Devi Comp 122 Analysis (Continued) Substitute k = j – i.

44. qsort - 44 Lin / Devi Comp 122 Deterministic vs. Randomized Algorithms  Deterministic Algorithm : Identical behavior for different runs for a given input.  Randomized Algorithm : Behavior is generally different for different runs for a given input. Algorithms DeterministicRandomized Worst-case Analysis Worst-case Running Time Probabilistic Analysis Average Running Time Probabilistic Analysis Average Running Time