Downloaded 202 times
![is always right dominant, as is the last occurrence of the maximum element of the array. For example, consider the following list. L = 10, 9, 5, 13, 2, 7, 1, 8, 4, 6, 3 The sequence of right dominant elements are 13, 8, 6, 3 . In order to make this more concrete, we should think about how L is represented. It will make a difference whether L is represented as an array (allowing for random access), a doubly linked list (allowing for sequential access in both directions), or a singly linked list (allowing for sequential access in only one direction). Among the three possible representations, the array representation seems to yield the simplest and clearest algorithm. However, we will design the algorithm in such a way that it only performs sequential scans, so it could also be implemented using a singly linked or doubly linked list. (This is common in algorithms. Chose your rep- resentation to make the algorithm as simple and clear as possible, but give thought to how it may actually be implemented. Remember that algorithms are read by humans, not compilers.) We will assume here that the array L of size n is indexed from 1 to n. Think for a moment how you would solve this problem. Can you see an O(n) time algorithm? (If not, think a little harder.) To illustrate summations, we will first present a naive O(n2 ) time algorithm, which operates by simply checking for each element of the array whether all the subsequent elements are strictly smaller. (Although this example is pretty stupid, it will also serve to illustrate the sort of style that we will use in presenting algorithms.) Right Dominant Elements (Naive Solution) // Input: List L of numbers given as an array L[1..n] // Returns: List D containing the right dominant elements of L RightDominant(L) { D = empty list for (i = 1 to n) isDominant = true for (j = i+1 to n) if (A[i] <= A[j]) isDominant = false if (isDominant) append A[i] to D } return D } If I were programming this, I would rewrite the inner (j) loop as a while loop, since we can terminate the loop as soon as we find that A[i] is not dominant. Again, this sort of optimization is good to keep in mind in programming, but will be omitted since it will not affect the worst-case running time. The time spent in this algorithm is dominated (no pun intended) by the time spent in the inner (j) loop. On the ith iteration of the outer loop, the inner loop is executed from i + 1 to n, for a total of n − (i + 1) + 1 = n − i times. (Recall the rule for the constant series above.) Each iteration of the inner loop takes constant time. Thus, up to a constant factor, the running time, as a function of n, is given by the following summation: n T (n) = (n − i). i=1 To solve this summation, let us expand it, and put it into a form such that the above formulas can be used. T (n) = (n − 1) + (n − 2) + . . . + 2 + 1 + 0 = 0 + 1 + 2 + . . . + (n − 2) + (n − 1) n−1 (n − 1)n = i = . i=0 2 Lecture Notes 6 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-6-2048.jpg)
![is, up to constant factors, roughly (n/e)n . Plugging this in and simplifying yields the Ω(n log n) lower bound. This can also be generalized to show that the average-case time to sort is also Ω(n log n). Linear Time Sorting: The Ω(n log n) lower bound implies that if we hope to sort numbers faster than in O(n log n) time, we cannot do it by making comparisons alone. In some special cases, it is possible to sort without the use of comparisons. This leads to the possibility of sorting in linear (that is, O(n)) time. Here are three such algorithms. Counting Sort: Counting sort assumes that each input is an integer in the range from 1 to k. The algorithm sorts in Θ(n + k) time. Thus, if k is O(n), this implies that the resulting sorting algorithm runs in Θ(n) time. The algorithm requires an additional Θ(n + k) working storage but has the nice feature that it is stable. The algorithm is remarkably simple, but deceptively clever. You are referred to CLRS for the details. Radix Sort: The main shortcoming of CountingSort is that (due to space requirements) it is only practical for a very small ranges of integers. If the integers are in the range from say, 1 to a million, we may not want to allocate an array of a million elements. RadixSort provides a nice way around this by sorting numbers one digit, or one byte, or generally, some groups of bits, at a time. As the number of bits in each group increases, the algorithm is faster, but the space requirements go up. The idea is very simple. Let’s think of our list as being composed of n integers, each having d decimal digits (or digits in any base). To sort these integers we simply sort repeatedly, starting at the lowest order digit, and finishing with the highest order digit. Since the sorting algorithm is stable, we know that if the numbers are already sorted with respect to low order digits, and then later we sort with respect to high order digits, numbers having the same high order digit will remain sorted with respect to their low order digit. An example is shown in Figure 2. Input Output 576 49[4] 9[5]4 [1]76 176 494 19[4] 5[7]6 [1]94 194 194 95[4] 1[7]6 [2]78 278 296 =⇒ 57[6] =⇒ 2[7]8 =⇒ [2]96 =⇒ 296 278 29[6] 4[9]4 [4]94 494 176 17[6] 1[9]4 [5]76 576 954 27[8] 2[9]6 [9]54 954 Fig. 2: Example of RadixSort. The running time is Θ(d(n + k)) where d is the number of digits in each value, n is the length of the list, and k is the number of distinct values each digit may have. The space needed is Θ(n + k). A common application of this algorithm is for sorting integers over some range that is larger than n, but still polynomial in n. For example, suppose that you wanted to sort a list of integers in the range from 1 to n2 . First, you could subtract 1 so that they are now in the range from 0 to n2 − 1. Observe that any number in this range can be expressed as 2-digit number, where each digit is over the range from 0 to n − 1. In particular, given any integer L in this range, we can write L = an + b, where a = L/n and b = L mod n. Now, we can think of L as the 2-digit number (a, b). So, we can radix sort these numbers in time Θ(2(n + n)) = Θ(n). In general this works to sort any n numbers over the range from 1 to nd , in Θ(dn) time. BucketSort: CountingSort and RadixSort are only good for sorting small integers, or at least objects (like characters) that can be encoded as small integers. What if you want to sort a set of floating-point numbers? In the worst-case you are pretty much stuck with using one of the comparison-based sorting algorithms, such as QuickSort, MergeSort, or HeapSort. However, in special cases where you have reason to believe that your numbers are roughly uniformly distributed over some range, then it is possible to do better. (Note Lecture Notes 10 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-10-2048.jpg)
![The most important question in designing a DP solution to a problem is how to set up the subproblem structure. This is called the formulation of the problem. Dynamic programming is not applicable to all optimization problems. There are two important elements that a problem must have in order for DP to be applicable. Optimal substructure: (Sometimes called the principle of optimality.) It states that for the global problem to be solved optimally, each subproblem should be solved optimally. (Not all optimization problems satisfy this. Sometimes it is better to lose a little on one subproblem in order to make a big gain on another.) Polynomially many subproblems: An important aspect to the efficiency of DP is that the total number of subproblems to be solved should be at most a polynomial number. Strings: One important area of algorithm design is the study of algorithms for character strings. There are a number of important problems here. Among the most important has to do with efficiently searching for a substring or generally a pattern in large piece of text. (This is what text editors and programs like “grep” do when you perform a search.) In many instances you do not want to find a piece of text exactly, but rather something that is similar. This arises for example in genetics research and in document retrieval on the web. One common method of measuring the degree of similarity between two strings is to compute their longest common subsequence. Longest Common Subsequence: Let us think of character strings as sequences of characters. Given two sequences X = x1 , x2 , . . . , xm and Z = z1 , z2 , . . . , zk , we say that Z is a subsequence of X if there is a strictly in- creasing sequence of k indices i1 , i2 , . . . , ik (1 ≤ i1 < i2 < . . . < ik ≤ n) such that Z = Xi1 , Xi2 , . . . , Xik . For example, let X = ABRACADABRA and let Z = AADAA , then Z is a subsequence of X. Given two strings X and Y , the longest common subsequence of X and Y is a longest sequence Z that is a subsequence of both X and Y . For example, let X = ABRACADABRA and let Y = YABBADABBADOO . Then the longest common subsequence is Z = ABADABA . See Fig. 4 X= A B R A C A D A B R A LCS = A B A D A B A Y= Y A B B A D A B B A D O O Fig. 4: An example of the LCS of two strings X and Y . The Longest Common Subsequence Problem (LCS) is the following. Given two sequences X = x1 , . . . , xm and Y = y1 , . . . , yn determine a longest common subsequence. Note that it is not always unique. For example the LCS of ABC and BAC is either AC or BC . DP Formulation for LCS: The simple brute-force solution to the problem would be to try all possible subsequences from one string, and search for matches in the other string, but this is hopelessly inefficient, since there are an exponential number of possible subsequences. Instead, we will derive a dynamic programming solution. In typical DP fashion, we need to break the prob- lem into smaller pieces. There are many ways to do this for strings, but it turns out for this problem that considering all pairs of prefixes will suffice for us. A prefix of a sequence is just an initial string of values, Xi = x1 , x2 , . . . , xi . X0 is the empty sequence. The idea will be to compute the longest common subsequence for every possible pair of prefixes. Let c[i, j] denote the length of the longest common subsequence of Xi and Yj . For example, in the above case we have X5 = ABRAC and Y6 = YABBAD . Their longest common subsequence is ABA . Thus, c[5, 6] = 3. Which of the c[i, j] values do we compute? Since we don’t know which will lead to the final optimum, we compute all of them. Eventually we are interested in c[m, n] since this will be the LCS of the two entire strings. The idea is to compute c[i, j] assuming that we already know the values of c[i , j ], for i ≤ i and j ≤ j (but not both equal). Here are the possible cases. Lecture Notes 12 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-12-2048.jpg)
![Basis: c[i, 0] = c[j, 0] = 0. If either sequence is empty, then the longest common subsequence is empty. Last characters match: Suppose xi = yj . For example: Let Xi = ABCA and let Yj = DACA . Since both end in A, we claim that the LCS must also end in A. (We will leave the proof as an exercise.) Since the A is part of the LCS we may find the overall LCS by removing A from both sequences and taking the LCS of Xi−1 = ABC and Yj−1 = DAC which is AC and then adding A to the end, giving ACA as the answer. (At first you might object: But how did you know that these two A’s matched with each other. The answer is that we don’t, but it will not make the LCS any smaller if we do.) This is illustrated at the top of Fig. 5. if xi = yj then c[i, j] = c[i − 1, j − 1] + 1 xi Xi A Xi−1 A Last chars match: yj LCS add to LCS Yj A Yj−1 A Xi−1 A skip xi LCS xi Yj B Last chars do not Xi A max match yj Yj B Xi A LCS Yj−1 B skip yj Fig. 5: The possibe cases in the DP formulation of LCS. Last characters do not match: Suppose that xi = yj . In this case xi and yj cannot both be in the LCS (since they would have to be the last character of the LCS). Thus either xi is not part of the LCS, or yj is not part of the LCS (and possibly both are not part of the LCS). At this point it may be tempting to try to make a “smart” choice. By analyzing the last few characters of Xi and Yj , perhaps we can figure out which character is best to discard. However, this approach is doomed to failure (and you are strongly encouraged to think about this, since it is a common point of confusion.) Instead, our approach is to take advantage of the fact that we have already precomputed smaller subproblems, and use these results to guide us. In the first case (xi is not in the LCS) the LCS of Xi and Yj is the LCS of Xi−1 and Yj , which is c[i − 1, j]. In the second case (yj is not in the LCS) the LCS is the LCS of Xi and Yj−1 which is c[i, j − 1]. We do not know which is the case, so we try both and take the one that gives us the longer LCS. This is illustrated at the bottom half of Fig. 5. if xi = yj then c[i, j] = max(c[i − 1, j], c[i, j − 1]) Combining these observations we have the following formulation: 0 if i = 0 or j = 0, c[i, j] = c[i − 1, j − 1] + 1 if i, j > 0 and xi = yj , max(c[i, j − 1], c[i − 1, j]) if i, j > 0 and xi = yj . Implementing the Formulation: The task now is to simply implement this formulation. We concentrate only on computing the maximum length of the LCS. Later we will see how to extract the actual sequence. We will store some helpful pointers in a parallel array, b[0..m, 0..n]. The code is shown below, and an example is illustrated in Fig. 6 Lecture Notes 13 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-13-2048.jpg)
![Y: 0 1 2 3 4 =n Y: 0 1 2 3 4 =n B D C B B D C B 0 0 0 0 0 0 X = BACDB 0 0 0 0 0 0 1 B 0 1 1 1 1 Y = BDCB 1 B 0 1 1 1 1 X: 2 A 0 1 1 1 1 X: 2 A 0 1 1 1 1 3 C 0 1 1 2 2 3 C 0 1 1 2 2 4 D 0 1 2 2 2 4 D 0 1 2 2 2 m= 5 B 0 1 2 2 3 LCS = BCB m= 5 B 0 1 2 2 3 start here LCS Length Table with back pointers included Fig. 6: Longest common subsequence example for the sequences X = BACDB and Y = BCDB . The numeric table entries are the values of c[i, j] and the arrow entries are used in the extraction of the sequence. Build LCS Table LCS(x[1..m], y[1..n]) { // compute LCS table int c[0..m, 0..n] for i = 0 to m // init column 0 c[i,0] = 0; b[i,0] = SKIPX for j = 0 to n // init row 0 c[0,j] = 0; b[0,j] = SKIPY for i = 1 to m // fill rest of table for j = 1 to n if (x[i] == y[j]) // take X[i] (Y[j]) for LCS c[i,j] = c[i-1,j-1]+1; b[i,j] = addXY else if (c[i-1,j] >= c[i,j-1]) // X[i] not in LCS c[i,j] = c[i-1,j]; b[i,j] = skipX else // Y[j] not in LCS c[i,j] = c[i,j-1]; b[i,j] = skipY return c[m,n] // return length of LCS } Extracting the LCS getLCS(x[1..m], y[1..n], b[0..m,0..n]) { LCSstring = empty string i = m; j = n // start at lower right while(i != 0 && j != 0) // go until upper left switch b[i,j] case addXY: // add X[i] (=Y[j]) add x[i] (or equivalently y[j]) to front of LCSstring i--; j--; break case skipX: i--; break // skip X[i] case skipY: j--; break // skip Y[j] return LCSstring } Lecture Notes 14 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-14-2048.jpg)
![The running time of the algorithm is clearly O(mn) since there are two nested loops with m and n iterations, respectively. The algorithm also uses O(mn) space. Extracting the Actual Sequence: Extracting the final LCS is done by using the back pointers stored in b[0..m, 0..n]. Intuitively b[i, j] = addXY means that X[i] and Y [j] together form the last character of the LCS. So we take this common character, and continue with entry b[i − 1, j − 1] to the northwest ( ). If b[i, j] = skipX , then we know that X[i] is not in the LCS, and so we skip it and go to b[i − 1, j] above us (↑). Similarly, if b[i, j] = skipY , then we know that Y [j] is not in the LCS, and so we skip it and go to b[i, j − 1] to the left (←). Following these back pointers, and outputting a character with each diagonal move gives the final subsequence. Lecture 5: Dynamic Programming: Chain Matrix Multiplication Read: Chapter 15 of CLRS, and Section 15.2 in particular. Chain Matrix Multiplication: This problem involves the question of determining the optimal sequence for perform- ing a series of operations. This general class of problem is important in compiler design for code optimization and in databases for query optimization. We will study the problem in a very restricted instance, where the dynamic programming issues are easiest to see. Suppose that we wish to multiply a series of matrices A1 A2 . . . An Matrix multiplication is an associative but not a commutative operation. This means that we are free to paren- thesize the above multiplication however we like, but we are not free to rearrange the order of the matrices. Also recall that when two (nonsquare) matrices are being multiplied, there are restrictions on the dimensions. A p × q matrix has p rows and q columns. You can multiply a p × q matrix A times a q × r matrix B, and the result will be a p × r matrix C. (The number of columns of A must equal the number of rows of B.) In particular for 1 ≤ i ≤ p and 1 ≤ j ≤ r, q C[i, j] = A[i, k]B[k, j]. k=1 This corresponds to the (hopefully familiar) rule that the [i, j] entry of C is the dot product of the ith (horizontal) row of A and the jth (vertical) column of B. Observe that there are pr total entries in C and each takes O(q) time to compute, thus the total time to multiply these two matrices is proportional to the product of the dimensions, pqr. A * B = C q = Multiplication p p time = pqr r q r Fig. 7: Matrix Multiplication. Note that although any legal parenthesization will lead to a valid result, not all involve the same number of operations. Consider the case of 3 matrices: A1 be 5 × 4, A2 be 4 × 6 and A3 be 6 × 2. multCost[((A1 A2 )A3 )] = (5 · 4 · 6) + (5 · 6 · 2) = 180, multCost[(A1 (A2 A3 ))] = (4 · 6 · 2) + (5 · 4 · 2) = 88. Even for this small example, considerable savings can be achieved by reordering the evaluation sequence. Lecture Notes 15 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-15-2048.jpg)
![number of total possibilities. What saves us here is that there are only O(n2 ) different sequences of matrices. (There are n = n(n − 1)/2 ways of choosing i and j to form Ai..j to be precise.) Thus, we do not encounter 2 the exponential growth. Notice that our chain matrix multiplication problem satisfies the principle of optimality, because once we decide to break the sequence into the product A1..k · Ak+1..n , we should compute each subsequence optimally. That is, for the global problem to be solved optimally, the subproblems must be solved optimally as well. Dynamic Programming Formulation: We will store the solutions to the subproblems in a table, and build the table in a bottom-up manner. For 1 ≤ i ≤ j ≤ n, let m[i, j] denote the minimum number of multiplications needed to compute Ai..j . The optimum cost can be described by the following recursive formulation. Basis: Observe that if i = j then the sequence contains only one matrix, and so the cost is 0. (There is nothing to multiply.) Thus, m[i, i] = 0. Step: If i < j, then we are asking about the product Ai..j . This can be split by considering each k, i ≤ k < j, as Ai..k times Ak+1..j . The optimum times to compute Ai..k and Ak+1..j are, by definition, m[i, k] and m[k + 1, j], respectively. We may assume that these values have been computed previously and are already stored in our array. Since Ai..k is a pi−1 × pk matrix, and Ak+1..j is a pk × pj matrix, the time to multiply them is pi−1 pk pj . This suggests the following recursive rule for computing m[i, j]. m[i, i] = 0 m[i, j] = min (m[i, k] + m[k + 1, j] + pi−1 pk pj ) for i < j. i≤k<j A i..j A i..k A k+1..j A i A i+1 ... A k A k+1 ... A j ? Fig. 8: Dynamic Programming Formulation. It is not hard to convert this rule into a procedure, which is given below. The only tricky part is arranging the order in which to compute the values. In the process of computing m[i, j] we need to access values m[i, k] and m[k + 1, j] for k lying between i and j. This suggests that we should organize our computation according to the number of matrices in the subsequence. Let L = j−i+1 denote the length of the subchain being multiplied. The subchains of length 1 (m[i, i]) are trivial to compute. Then we build up by computing the subchains of lengths 2, 3, . . . , n. The final answer is m[1, n]. We need to be a little careful in setting up the loops. If a subchain of length L starts at position i, then j = i + L − 1. Since we want j ≤ n, this means that i + L − 1 ≤ n, or in other words, i ≤ n − L + 1. So our loop for i runs from 1 to n − L + 1 (in order to keep j in bounds). The code is presented below. The array s[i, j] will be explained later. It is used to extract the actual sequence. The running time of the procedure is Θ(n3 ). We’ll leave this as an exercise in solving sums, but the key is that there are three nested loops, and each can iterate at most n times. Extracting the final Sequence: Extracting the actual multiplication sequence is a fairly easy extension. The basic idea is to leave a split marker indicating what the best split is, that is, the value of k that leads to the minimum Lecture Notes 17 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-17-2048.jpg)
![Chain Matrix Multiplication Matrix-Chain(array p[1..n]) { array s[1..n-1,2..n] for i = 1 to n do m[i,i] = 0; // initialize for L = 2 to n do { // L = length of subchain for i = 1 to n-L+1 do { j = i + L - 1; m[i,j] = INFINITY; for k = i to j-1 do { // check all splits q = m[i, k] + m[k+1, j] + p[i-1]*p[k]*p[j] if (q < m[i, j]) { m[i,j] = q; s[i,j] = k; } } } } return m[1,n] (final cost) and s (splitting markers); } value of m[i, j]. We can maintain a parallel array s[i, j] in which we will store the value of k providing the optimal split. For example, suppose that s[i, j] = k. This tells us that the best way to multiply the subchain Ai..j is to first multiply the subchain Ai..k and then multiply the subchain Ak+1..j , and finally multiply these together. Intuitively, s[i, j] tells us what multiplication to perform last. Note that we only need to store s[i, j] when we have at least two matrices, that is, if j > i. The actual multiplication algorithm uses the s[i, j] value to determine how to split the current sequence. Assume that the matrices are stored in an array of matrices A[1..n], and that s[i, j] is global to this recursive procedure. The recursive procedure Mult does this computation and below returns a matrix. Extracting Optimum Sequence Mult(i, j) { if (i == j) // basis case return A[i]; else { k = s[i,j] X = Mult(i, k) // X = A[i]...A[k] Y = Mult(k+1, j) // Y = A[k+1]...A[j] return X*Y; // multiply matrices X and Y } } In the figure below we show an example. This algorithm is tricky, so it would be a good idea to trace through this example (and the one given in the text). The initial set of dimensions are 5, 4, 6, 2, 7 meaning that we are multiplying A1 (5 × 4) times A2 (4 × 6) times A3 (6 × 2) times A4 (2 × 7). The optimal sequence is ((A1 (A2 A3 ))A4 ). Lecture 6: Dynamic Programming: Minimum Weight Triangulation Read: This is not covered in CLRS. Lecture Notes 18 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-18-2048.jpg)
![4 1 s[i,j] 4 1 m[i,j] 3 2 158 2 3 3 j 1 3 i j 88 104 i 2 3 2 3 1 2 3 120 48 84 4 1 0 0 0 0 5 4 6 2 7 3 Final order A1 A2 A3 A4 1 2 p0 p1 p2 p3 p4 A1 A2 A3 A4 Fig. 9: Chain Matrix Multiplication Example. Polygons and Triangulations: Let’s consider a geometric problem that outwardly appears to be quite different from chain-matrix multiplication, but actually has remarkable similarities. We begin with a number of definitions. Define a polygon to be a piecewise linear closed curve in the plane. In other words, we form a cycle by joining line segments end to end. The line segments are called the sides of the polygon and the endpoints are called the vertices. A polygon is simple if it does not cross itself, that is, if the sides do not intersect one another except for two consecutive sides sharing a common vertex. A simple polygon subdivides the plane into its interior, its boundary and its exterior. A simple polygon is said to be convex if every interior angle is at most 180 degrees. Vertices with interior angle equal to 180 degrees are normally allowed, but for this problem we will assume that no such vertices exist. Polygon Simple polygon Convex polygon Fig. 10: Polygons. Given a convex polygon, we assume that its vertices are labeled in counterclockwise order P = v1 , . . . , vn . We will assume that indexing of vertices is done modulo n, so v0 = vn . This polygon has n sides, vi−1 vi . Given two nonadjacent sides vi and vj , where i < j−1, the line segment vi vj is a chord. (If the polygon is simple but not convex, we include the additional requirement that the interior of the segment must lie entirely in the interior of P .) Any chord subdivides the polygon into two polygons: vi , vi+1 , . . . , vj , and vj , vj+1 , . . . , vi . A triangulation of a convex polygon P is a subdivision of the interior of P into a collection of triangles with disjoint interiors, whose vertices are drawn from the vertices of P . Equivalently, we can define a triangulation as a maximal set T of nonintersecting chords. (In other words, every chord that is not in T intersects the interior of some chord in T .) It is easy to see that such a set of chords subdivides the interior of the polygon into a collection of triangles with pairwise disjoint interiors (and hence the name triangulation). It is not hard to prove (by induction) that every triangulation of an n-sided polygon consists of n − 3 chords and n − 2 triangles. Triangulations are of interest for a number of reasons. Many geometric algorithm operate by first decomposing a complex polygonal shape into triangles. In general, given a convex polygon, there are many possible triangulations. In fact, the number is exponential in n, the number of sides. Which triangulation is the “best”? There are many criteria that are used depending on the application. One criterion is to imagine that you must “pay” for the ink you use in drawing the triangulation, and you want to minimize the amount of ink you use. (This may sound fanciful, but minimizing wire length is an Lecture Notes 19 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-19-2048.jpg)
![important condition in chip design. Further, this is one of many properties which we could choose to optimize.) This suggests the following optimization problem: Minimum-weight convex polygon triangulation: Given a convex polygon determine the triangulation that minimizes the sum of the perimeters of its triangles. (See Fig. 11.) A triangulation Lower weight triangulation Fig. 11: Triangulations of convex polygons, and the minimum weight triangulation. Given three distinct vertices vi , vj , vk , we define the weight of the associated triangle by the weight function w(vi , vj , vk ) = |vi vj | + |vj vk | + |vk vi |, where |vi vj | denotes the length of the line segment vi vj . Dynamic Programming Solution: Let us consider an (n + 1)-sided polygon P = v0 , v1 , . . . , vn . Let us assume that these vertices have been numbered in counterclockwise order. To derive a DP formulation we need to define a set of subproblems from which we can derive the optimum solution. For 0 ≤ i < j ≤ n, define t[i, j] to be the weight of the minimum weight triangulation for the subpolygon that lies to the right of directed chord vi vj , that is, the polygon with the counterclockwise vertex sequence vi , vi+1 , . . . , vj . Observe that if we can compute this quantity for all such i and j, then the weight of the minimum weight triangulation of the entire polygon can be extracted as t[0, n]. (As usual, we only compute the minimum weight. But, it is easy to modify the procedure to extract the actual triangulation.) As a basis case, we define the weight of the trivial “2-sided polygon” to be zero, implying that t[i, i + 1] = 0. In general, to compute t[i, j], consider the subpolygon vi , vi+1 , . . . , vj , where j > i + 1. One of the chords of this polygon is the side vi vj . We may split this subpolygon by introducing a triangle whose base is this chord, and whose third vertex is any vertex vk , where i < k < j. This subdivides the polygon into the subpolygons vi , vi+1 , . . . vk and vk , vk+1 , . . . vj whose minimum weights are already known to us as t[i, k] and t[k, j]. In addition we should consider the weight of the newly added triangle vi vk vj . Thus, we have the following recursive rule: 0 if j = i + 1 t[i, j] = mini<k<j (t[i, k] + t[k, j] + w(vi vk vj )) if j > i + 1. The final output is the overall minimum weight, which is, t[0, n]. This is illustrated in Fig. 12 Note that this has almost exactly the same structure as the recursive definition used in the chain matrix multipli- cation algorithm (except that some indices are different by 1.) The same Θ(n3 ) algorithm can be applied with only minor changes. Relationship to Binary Trees: One explanation behind the similarity of triangulations and the chain matrix multipli- cation algorithm is to observe that both are fundamentally related to binary trees. In the case of the chain matrix multiplication, the associated binary tree is the evaluation tree for the multiplication, where the leaves of the tree correspond to the matrices, and each node of the tree is associated with a product of a sequence of two or more matrices. To see that there is a similar correspondence here, consider an (n + 1)-sided convex polygon P = v0 , v1 , . . . , vn , and fix one side of the polygon (say v0 vn ). Now consider a rooted binary tree whose root node is the triangle containing side v0 vn , whose internal nodes are the nodes of the dual tree, and whose leaves Lecture Notes 20 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-20-2048.jpg)
![vn v0 vj Triangulate at cost t[k,j] vi cost=w(v i ,vk, vj ) Triangulate vk at cost t[i,k] Fig. 12: Triangulations and tree structure. correspond to the remaining sides of the tree. Observe that partitioning the polygon into triangles is equivalent to a binary tree with n leaves, and vice versa. This is illustrated in Fig. 13. Note that every triangle is associated with an internal node of the tree and every edge of the original polygon, except for the distinguished starting side v0 vn , is associated with a leaf node of the tree. root A1 v0 v11 root A11 v1 A2 v10 A10 v2 A3 v9 v3 A9 v8 A4 v4 v7 A8 A5 v A7 5 v6 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A6 Fig. 13: Triangulations and tree structure. Once you see this connection. Then the following two observations follow easily. Observe that the associated binary tree has n leaves, and hence (by standard results on binary trees) n − 1 internal nodes. Since each internal node other than the root has one edge entering it, there are n − 2 edges between the internal nodes. Each internal node corresponds to one triangle, and each edge between internal nodes corresponds to one chord of the triangulation. Lecture 7: Greedy Algorithms: Activity Selection and Fractional Knapack Read: Sections 16.1 and 16.2 in CLRS. Greedy Algorithms: In many optimization algorithms a series of selections need to be made. In dynamic program- ming we saw one way to make these selections. Namely, the optimal solution is described in a recursive manner, and then is computed “bottom-up”. Dynamic programming is a powerful technique, but it often leads to algo- rithms with higher than desired running times. Today we will consider an alternative design technique, called greedy algorithms. This method typically leads to simpler and faster algorithms, but it is not as powerful or as widely applicable as dynamic programming. We will give some examples of problems that can be solved by greedy algorithms. (Later in the semester, we will see that this technique can be applied to a number of graph problems as well.) Even when greedy algorithms do not produce the optimal solution, they often provide fast heuristics (nonoptimal solution strategies), are often used in finding good approximations. Lecture Notes 21 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-21-2048.jpg)
![Activity Scheduling: Activity scheduling and it is a very simple scheduling problem. We are given a set S = {1, 2, . . . , n} of n activities that are to be scheduled to use some resource, where each activity must be started at a given start time si and ends at a given finish time fi . For example, these might be lectures that are to be given in a lecture hall, where the lecture times have been set up in advance, or requests for boats to use a repair facility while they are in port. Because there is only one resource, and some start and finish times may overlap (and two lectures cannot be given in the same room at the same time), not all the requests can be honored. We say that two activities i and j are noninterfering if their start-finish intervals do not overlap, more formally, [si , fi ) ∩ [sj , fj ) = ∅. (Note that making the intervals half open, two consecutive activities are not considered to interfere.) The activity scheduling problem is to select a maximum-size set of mutually noninterfering activities for use of the resource. (Notice that goal here is maximum number of activities, not maximum utilization. Of course different criteria could be considered, but the greedy approach may not be optimal in general.) How do we schedule the largest number of activities on the resource? Intuitively, we do not like long activities, because they occupy the resource and keep us from honoring other requests. This suggests the following greedy strategy: repeatedly select the activity with the smallest duration (fi − si ) and schedule it, provided that it does not interfere with any previously scheduled activities. Although this seems like a reasonable strategy, this turns out to be nonoptimal. (See Problem 17.1-4 in CLRS). Sometimes the design of a correct greedy algorithm requires trying a few different strategies, until hitting on one that works. Here is a greedy strategy that does work. The intuition is the same. Since we do not like activities that take a long time, let us select the activity that finishes first and schedule it. Then, we skip all activities that interfere with this one, and schedule the next one that has the earliest finish time, and so on. To make the selection process faster, we assume that the activities have been sorted by their finish times, that is, f1 ≤ f2 ≤ . . . ≤ fn , Assuming this sorting, the pseudocode for the rest of the algorithm is presented below. The output is the list A of scheduled activities. The variable prev holds the index of the most recently scheduled activity at any time, in order to determine interferences. Greedy Activity Scheduler schedule(s[1..n], f[1..n]) { // given start and finish times // we assume f[1..n] already sorted List A = <1> // schedule activity 1 first prev = 1 for i = 2 to n if (s[i] >= f[prev]) { // no interference? append i to A; prev = i // schedule i next } return A } It is clear that the algorithm is quite simple and efficient. The most costly activity is that of sorting the activities by finish time, so the total running time is Θ(n log n). Fig. 14 shows an example. Each activity is represented by its start-finish time interval. Observe that the intervals are sorted by finish time. Event 1 is scheduled first. It interferes with activity 2 and 3. Then Event 4 is scheduled. It interferes with activity 5 and 6. Finally, activity 7 is scheduled, and it intereferes with the remaining activity. The final output is {1, 4, 7}. Note that this is not the only optimal schedule. {2, 4, 7} is also optimal. Proof of Optimality: Our proof of optimality is based on showing that the first choice made by the algorithm is the best possible, and then using induction to show that the rest of the choices result in an optimal schedule. Proofs of optimality for greedy algorithms follow a similar structure. Suppose that you have any nongreedy solution. Lecture Notes 22 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-22-2048.jpg)
![Huffman’s Algorithm Huffman(int n, character C[1..n]) { Q = C; // priority queue for i = 1 to n-1 { z = new internal tree node; z.left = x = Q.extractMin(); // extract smallest probabilities z.right = y = Q.extractMin(); z.prob = x.prob + y.prob; // z’s probability is their sum Q.insert(z); // insert z into queue } return the last element left in Q as the root; } T T’ T’’ x b b y y c c b c x y x Cost change = Cost change = −(p(b)−p(x))(d(b)−d(x)) −(p(c)−p(y))(d(c)−d(y)) <0 <0 Fig. 19: Correctness of Huffman’s Algorithm. nodes and adding in the new contributions we have B(T ) = B(T ) − p(x)d(x) + p(x)d(b) − p(b)d(b) + p(b)d(x) = B(T ) + p(x)(d(b) − d(x)) − p(b)(d(b) − d(x)) = B(T ) − (p(b) − p(x))(d(b) − d(x)) ≤ B(T ) because (p(b) − p(x))(d(b) − d(x)) ≥ 0. Thus the cost does not increase, implying that T is an optimal tree. By switching y with c we get a new tree T , which by a similar argument is also optimal. The final tree T satisfies the statement of the claim. The above theorem asserts that the first step of Huffman’s algorithm is essentially the proper one to perform. The complete proof of correctness for Huffman’s algorithm follows by induction on n (since with each step, we eliminate exactly one character). Claim: Huffman’s algorithm produces the optimal prefix code tree. Proof: The proof is by induction on n, the number of characters. For the basis case, n = 1, the tree consists of a single leaf node, which is obviously optimal. Assume inductively that when strictly fewer than n characters, Huffman’s algorithm is guaranteed to pro- duce the optimal tree. We want to show it is true with exactly n characters. Suppose we have exactly n characters. The previous claim states that we may assume that in the optimal tree, the two characters of lowest probability x and y will be siblings at the lowest level of the tree. Remove x and y, replacing them with a new character z whose probability is p(z) = p(x) + p(y). Thus n − 1 characters remain. Consider any prefix code tree T made with this new set of n − 1 characters. We can convert it into a prefix code tree T for the original set of characters by undoing the previous operation and replacing z with x Lecture Notes 29 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-29-2048.jpg)
![Representations of Graphs and Digraphs: There are two common ways of representing graphs and digraphs. First we show how to represent digraphs. Let G = (V, E) be a digraph with n = |V | and let e = |E|. We will assume that the vertices of G are indexed {1, 2, . . . , n}. Adjacency Matrix: An n × n matrix defined for 1 ≤ v, w ≤ n. 1 if (v, w) ∈ E A[v, w] = 0 otherwise. If the digraph has weights we can store the weights in the matrix. For example if (v, w) ∈ E then A[v, w] = W (v, w) (the weight on edge (v, w)). If (v, w) ∈ E then generally W (v, w) need not be / defined, but often we set it to some “special” value, e.g. A(v, w) = −1, or ∞. (By ∞ we mean (in practice) some number which is larger than any allowable weight. In practice, this might be some machine dependent constant like MAXINT.) Adjacency List: An array Adj[1 . . . n] of pointers where for 1 ≤ v ≤ n, Adj[v] points to a linked list contain- ing the vertices which are adjacent to v (i.e. the vertices that can be reached from v by a single edge). If the edges have weights then these weights may also be stored in the linked list elements. 1 2 3 Adj 1 1 1 1 1 1 2 3 1 2 0 0 1 2 3 3 0 1 0 3 2 2 3 Adjacency matrix Adjacency list Fig. 22: Adjacency matrix and adjacency list for digraphs. We can represent undirected graphs using exactly the same representation, but we will store each edge twice. In particular, we representing the undirected edge {v, w} by the two oppositely directed edges (v, w) and (w, v). Notice that even though we represent undirected graphs in the same way that we represent digraphs, it is impor- tant to remember that these two classes of objects are mathematically distinct from one another. This can cause some complications. For example, suppose you write an algorithm that operates by marking edges of a graph. You need to be careful when you mark edge (v, w) in the representation that you also mark (w, v), since they are both the same edge in reality. When dealing with adjacency lists, it may not be convenient to walk down the entire linked list, so it is common to include cross links between corresponding edges. Adj 1 2 3 4 1 2 3 4 1 0 1 1 1 1 4 2 1 3 2 1 0 1 0 3 1 1 0 1 3 1 2 4 2 3 4 1 0 1 0 4 1 3 Adjacency matrix Adjacency list (with crosslinks) Fig. 23: Adjacency matrix and adjacency list for graphs. An adjacency matrix requires Θ(V 2 ) storage and an adjacency list requires Θ(V + E) storage. The V arises because there is one entry for each vertex in Adj . Since each list has out-deg(v) entries, when this is summed over all vertices, the total number of adjacency list records is Θ(E). For sparse graphs the adjacency list representation is more space efficient. Lecture Notes 32 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-32-2048.jpg)
![Graph Traversals: There are a number of approaches used for solving problems on graphs. One of the most impor- tant approaches is based on the notion of systematically visiting all the vertices and edge of a graph. The reason for this is that these traversals impose a type of tree structure (or generally a forest) on the graph, and trees are usually much easier to reason about than general graphs. Breadth-first search: Given an graph G = (V, E), breadth-first search starts at some source vertex s and “discovers” which vertices are reachable from s. Define the distance between a vertex v and s to be the minimum number of edges on a path from s to v. Breadth-first search discovers vertices in increasing order of distance, and hence can be used as an algorithm for computing shortest paths. At any given time there is a “frontier” of vertices that have been discovered, but not yet processed. Breadth-first search is named because it visits vertices across the entire “breadth” of this frontier. Initially all vertices (except the source) are colored white, meaning that they are undiscovered. When a vertex has first been discovered, it is colored gray (and is part of the frontier). When a gray vertex is processed, then it becomes black. The search makes use of a queue, a first-in first-out list, where elements are removed in the same order they are inserted. The first item in the queue (the next to be removed) is called the head of the queue. We will also maintain arrays color[u] which holds the color of vertex u (either white, gray or black), pred[u] which points to the predecessor of u (i.e. the vertex who first discovered u, and d[u], the distance from s to u. Only the color is really needed for the search (in fact it is only necessary to know whether a node is nonwhite). We include all this information, because some applications of BFS use this additional information. Breadth-First Search BFS(G,s) { for each u in V { // initialization color[u] = white d[u] = infinity pred[u] = null } color[s] = gray // initialize source s d[s] = 0 Q = {s} // put s in the queue while (Q is nonempty) { u = Q.Dequeue() // u is the next to visit for each v in Adj[u] { if (color[v] == white) { // if neighbor v undiscovered color[v] = gray // ...mark it discovered d[v] = d[u]+1 // ...set its distance pred[v] = u // ...and its predecessor Q.Enqueue(v) // ...put it in the queue } } color[u] = black // we are done with u } } Observe that the predecessor pointers of the BFS search define an inverted tree (an acyclic directed graph in which the source is the root, and every other node has a unique path to the root). If we reverse these edges we get a rooted unordered tree called a BFS tree for G. (Note that there are many potential BFS trees for a given graph, depending on where the search starts, and in what order vertices are placed on the queue.) These edges of G are called tree edges and the remaining edges of G are called cross edges. It is not hard to prove that if G is an undirected graph, then cross edges always go between two nodes that are at most one level apart in the BFS tree. (Can you see why this must be true?) Below is a sketch of a proof that on Lecture Notes 33 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-33-2048.jpg)
![a s e d f c b g s s0 a s0 c s0 a 1 1 c 1 d a 1 1 c 1 d a 1 1 c 1 d Q: a, c, d e 2 Q: c, d, e e 2 b2 Q: d, e, b d s0 b, f, g s0 e s0 a 1 1 c 1 d a 1 1 c 1 d a 1 1 c 1 d e 2 b2 e 2 b2 e 2 b2 f 3 3 g Q: (empty) f 3 3 g Q: b, f, g Q: e, b Fig. 24: Breadth-first search: Example. termination, d[v] is equal to the distance from s to v. (See the CLRS for a detailed proof.) Theorem: Let δ(s, v) denote the length (number of edges) on the shortest path from s to v. Then, on termination of the BFS procedure, d[v] = δ(s, v). Proof: (Sketch) The proof is by induction on the length of the shortest path. Let u be the predecessor of v on some shortest path from s to v, and among all such vertices the first to be processed by the BFS. Thus, δ(s, v) = δ(s, u) + 1. When u is processed, we have (by induction) d[u] = δ(s, u). Since v is a neighbor of u, we set d[v] = d[u] + 1. Thus we have d[v] = d[u] + 1 = δ(s, u) + 1 = δ(s, v), as desired. Analysis: The running time analysis of BFS is similar to the running time analysis of many graph traversal algorithms. As done in CLR V = |V | and E = |E|. Observe that the initialization portion requires Θ(V ) time. The real meat is in the traversal loop. Since we never visit a vertex twice, the number of times we go through the while loop is at most V (exactly V assuming each vertex is reachable from the source). The number of iterations through the inner for loop is proportional to deg(u) + 1. (The +1 is because even if deg(u) = 0, we need to spend a constant amount of time to set up the loop.) Summing up over all vertices we have the running time T (V ) = V + (deg(u) + 1) = V + deg(u) + V = 2V + 2E ∈ Θ(V + E). u∈V u∈V The analysis is essentially the same for directed graphs. Lecture Notes 34 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-34-2048.jpg)
![Lecture 10: Depth-First Search Read: Sections 23.2 and 23.3 in CLR. Depth-First Search: The next traversal algorithm that we will study is called depth-first search, and it has the nice property that nontree edges have a good deal of mathematical structure. Consider the problem of searching a castle for treasure. To solve it you might use the following strategy. As you enter a room of the castle, paint some graffiti on the wall to remind yourself that you were already there. Successively travel from room to room as long as you come to a place you haven’t already been. When you return to the same room, try a different door leaving the room (assuming it goes somewhere you haven’t already been). When all doors have been tried in a given room, then backtrack. Notice that this algorithm is described recursively. In particular, when you enter a new room, you are beginning a new search. This is the general idea behind depth-first search. Depth-First Search Algorithm: We assume we are given an directed graph G = (V, E). The same algorithm works for undirected graphs (but the resulting structure imposed on the graph is different). We use four auxiliary arrays. As before we maintain a color for each vertex: white means undiscovered, gray means discovered but not finished processing, and black means finished. As before we also store predecessor pointers, pointing back to the vertex that discovered a given vertex. We will also associate two numbers with each vertex. These are time stamps. When we first discover a vertex u store a counter in d[u] and when we are finished processing a vertex we store a counter in f [u]. The purpose of the time stamps will be explained later. (Note: Do not confuse the discovery time d[v] with the distance d[v] from BFS.) The algorithm is shown in code block below, and illustrated in Fig. 25. As with BFS, DFS induces a tree structure. We will discuss this tree structure further below. Depth-First Search DFS(G) { // main program for each u in V { // initialization color[u] = white; pred[u] = null; } time = 0; for each u in V if (color[u] == white) // found an undiscovered vertex DFSVisit(u); // start a new search here } DFSVisit(u) { // start a search at u color[u] = gray; // mark u visited d[u] = ++time; for each v in Adj(u) do if (color[v] == white) { // if neighbor v undiscovered pred[v] = u; // ...set predecessor pointer DFSVisit(v); // ...visit v } color[u] = black; // we’re done with u f[u] = ++time; } Analysis: The running time of DFS is Θ(V + E). This is somewhat harder to see than the BFS analysis, because the recursive nature of the algorithm obscures things. Normally, recurrences are good ways to analyze recursively Lecture Notes 35 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-35-2048.jpg)
![Time-stamp structure: There is also a nice structure to the time stamps. In CLR this is referred to as the parenthesis structure. In particular, the following are easy to observe. Lemma: (Parenthesis Lemma) Given a digraph G = (V, E), and any DFS tree for G and any two vertices u, v ∈ V . • u is a descendent of v if and only if [d[u], f [u]] ⊆ [d[v], f [v]]. • u is an ancestor of v if and only if [d[u], f [u]] ⊇ [d[v], f [v]]. • u is unrelated to v if and only if [d[u], f [u]] and [d[v], f [v]] are disjoint. a d 1/10 C 11/14 a d b f C e 2/5 6/9 12/13 F B b f e c C g c g 3/4 7/8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Fig. 26: Parenthesis Lemma. Cycles: The time stamps given by DFS allow us to determine a number of things about a graph or digraph. For example, suppose you are given a graph or digraph. You run DFS. You can determine whether the graph contains any cycles very easily. We do this with the help of the following two lemmas. Lemma: Given a digraph G = (V, E), consider any DFS forest of G, and consider any edge (u, v) ∈ E. If this edge is a tree, forward, or cross edge, then f [u] > f [v]. If the edge is a back edge then f [u] ≤ f [v]. Proof: For tree, forward, and back edges, the proof follows directly from the parenthesis lemma. (E.g. for a forward edge (u, v), v is a descendent of u, and so v’s start-finish interval is contained within u’s, implying that v has an earlier finish time.) For a cross edge (u, v) we know that the two time intervals are disjoint. When we were processing u, v was not white (otherwise (u, v) would be a tree edge), implying that v was started before u. Because the intervals are disjoint, v must have also finished before u. Lemma: Consider a digraph G = (V, E) and any DFS forest for G. G has a cycle if and only the DFS forest has a back edge. Proof: (⇐) If there is a back edge (u, v), then v is an ancestor of u, and by following tree edges from v to u we get a cycle. (⇒) We show the contrapositive. Suppose there are no back edges. By the lemma above, each of the remaining types of edges, tree, forward, and cross all have the property that they go from vertices with higher finishing time to vertices with lower finishing time. Thus along any path, finish times decrease monotonically, implying there can be no cycle. Beware: No back edges means no cycles. But you should not infer that there is some simple relationship between the number of back edges and the number of cycles. For example, a DFS tree may only have a single back edge, and there may anywhere from one up to an exponential number of simple cycles in the graph. A similar theorem applies to undirected graphs, and is not hard to prove. Lecture Notes 37 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-37-2048.jpg)
![Lecture 11: Topological Sort and Strong Components Read: Sects. 22.3–22.5 in CLRS. Directed Acyclic Graph: A directed acyclic graph is often called a DAG for short DAG’s arise in many applications where there are precedence or ordering constraints. For example, if there are a series of tasks to be performed, and certain tasks must precede other tasks (e.g. in construction you have to build the first floor before you build the second floor, but you can do the electrical wiring while you install the windows). In general a precedence constraint graph is a DAG in which vertices are tasks and the edge (u, v) means that task u must be completed before task v begins. A topological sort of a DAG is a linear ordering of the vertices of the DAG such that for each edge (u, v), u appears before v in the ordering. Note that in general, there may be many legal topological orders for a given DAG. To compute a topological ordering is actually very easy, given DFS. By the previous lemma, for every edge (u, v) in a DAG, the finish time of u is greater than the finish time of v. Thus, it suffices to output the vertices in reverse order of finishing time. To do this we run a (stripped down) DFS, and when each vertex is finished we add it to the front of a linked list. The final linked list order will be the final topological order. This is given below. Topological Sort TopSort(G) { for each (u in V) color[u] = white; // initialize L = new linked_list; // L is an empty linked list for each (u in V) if (color[u] == white) TopVisit(u); return L; // L gives final order } TopVisit(u) { // start a search at u color[u] = gray; // mark u visited for each (v in Adj(u)) if (color[v] == white) TopVisit(v); Append u to the front of L; // on finishing u add to list } This is typical example of DFS is used in applications. Observe that the structure is essentially the same as the basic DFS procedure, but we only include the elements of DFS that are needed for this application. As an example we consider the DAG presented in CLRS for Professor Bumstead’s order of dressing. Bumstead lists the precedences in the order in which he puts on his clothes in the morning. We do our depth-first search in a different order from the one given in CLRS, and so we get a different final ordering. However both orderings are legitimate, given the precedence constraints. As with depth-first search, the running time of topological sort is Θ(V + E). Strong Components: Next we consider a very important connectivity problem with digraphs. When digraphs are used in communication and transportation networks, people want to know that there networks are complete in the sense that from any location it is possible to reach any other location in the digraph. A digraph is strongly connected if for every pair of vertices, u, v ∈ V , u can reach v and vice versa. We would like to write an algorithm that determines whether a digraph is strongly connected. In fact we will solve a generalization of this problem, of computing the strongly connected components (or strong components for short) of a digraph. In particular, we partition the vertices of the digraph into subsets such that the induced subgraph of each subset is strongly connected. (These subsets should be as large as possible, and still have this Lecture Notes 38 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-38-2048.jpg)
![Strong Components StrongComp(G) { Run DFS(G), computing finish times f[u] for each vertex u; Compute R = Reverse(G), reversing all edges of G; Sort the vertices of R (by CountingSort) in decreasing order of f[u]; Run DFS(R) using this order; Each DFS tree is a strong component; } a 1/18 b d 2/13 14/17 9 c a d f c e b f a 3/4 5/12 15/16 1 4 c e i g 6/11 d 2 f 5 h 8 b g h i e 7/10 3 g 6 7 i h 8/9 Initial DFS Reversal with new vertex order Final DFS with components Fig. 30: Strong Components Algorithm strong components in a reverse topological order. The question is how to order the vertices so that this is true. Recall from the topological sorting algorithm, that in a DAG, finish times occur in reverse topological order (i.e., the first vertex in the topological order is the one with the highest finish time). So, if we wanted to visit the components in reverse topological order, this suggests that we should visit the vertices in increasing order of finish time, starting with the lowest finishing time. This is a good starting idea, but it turns out that it doesn’t work. The reason is that there are many vertices in each strong component, and they all have different finish times. For example, in the figure above observe that in the first DFS (on the left) the lowest finish time (of 4) is achieved by vertex c, and its strong component is first, not last, in topological order. It is tempting to give up in frustration at this point. But there is something to notice about the finish times. If we consider the maximum finish time in each component, then these are related to the topological order of the component DAG. In particular, given any strong component C, define f (C) to be the maximum finish time among all vertices in this component. f (C) = max f [u]. u∈C Lemma: Consider a digraph G = (V, E) and let C and C be two distinct strong components. If there is an (u, v) of G such that u ∈ C and v ∈ C , then f (C) > f (C ). See the book for a complete proof. Here is a quick sketch. If the DFS visits C first, then the DFS will leak into C (along edge (u, v) or some other edge), and then will visit everything in C before finally returning to C. Thus, some vertex of C will finish later than every vertex of C . On the other hand, suppose that C is visited first. Because there is an edge from C to C , we know from the definition of the component DAG that there cannot be a path from C to C. So C will completely finish before we even start C. Thus all the finish times of C will be larger than the finish times of C . For example, in the previous figure, the maximum finish times for each component are 18 (for {a, b, c}), 17 (for {d, e}), and 12 (for {f, g, h, i}). The order 18, 17, 12 is a valid topological order for the component digraph. Lecture Notes 41 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-41-2048.jpg)
![decreaseKey(u, new key): Decrease the value of u’s key value to new key. A priority queue can be implemented using the same heap data structure used in heapsort. All of the above operations can be performed in O(log n) time, where n is the number of items in the heap. What do we store in the priority queue? At first you might think that we should store the edges that cross the cut, since this is what we are removing with each step of the algorithm. The problem is that when a vertex is moved from one side of the cut to the other, this results in a complicated sequence of updates. There is a much more elegant solution, and this is what makes Prim’s algorithm so nice. For each vertex in u ∈ V − S (not part of the current spanning tree) we associate u with a key value key[u], which is the weight of the lightest edge going from u to any vertex in S. We also store in pred[u] the end vertex of this edge in S. If there is not edge from u to a vertex in V − S, then we set its key value to +∞. We will also need to know which vertices are in S and which are not. We do this by coloring the vertices in S black. Here is Prim’s algorithm. The root vertex r can be any vertex in V . Prim’s Algorithm Prim(G,w,r) { for each (u in V) { // initialization key[u] = +infinity; color[u] = white; } key[r] = 0; // start at root pred[r] = nil; Q = new PriQueue(V); // put vertices in Q while (Q.nonEmpty()) { // until all vertices in MST u = Q.extractMin(); // vertex with lightest edge for each (v in Adj[u]) { if ((color[v] == white) && (w(u,v) < key[v])) { key[v] = w(u,v); // new lighter edge out of v Q.decreaseKey(v, key[v]); pred[v] = u; } } color[u] = black; } [The pred pointers define the MST as an inverted tree rooted at r] } The following figure illustrates Prim’s algorithm. The arrows on edges indicate the predecessor pointers, and the numeric label in each vertex is the key value. To analyze Prim’s algorithm, we account for the time spent on each vertex as it is extracted from the priority queue. It takes O(log V ) to extract this vertex from the queue. For each incident edge, we spend potentially O(log V ) time decreasing the key of the neighboring vertex. Thus the time is O(log V + deg(u) log V ) time. The other steps of the update are constant time. So the overall running time is T (V, E) = (log V + deg(u) log V ) = (1 + deg(u)) log V u∈V u∈V = log V (1 + deg(u)) = (log V )(V + 2E) = Θ((V + E) log V ). u∈V Since G is connected, V is asymptotically no greater than E, so this is Θ(E log V ). This is exactly the same as Kruskal’s algorithm. Lecture Notes 47 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-47-2048.jpg)
![path to be the sum of edge weights along the path. Define the distance between two vertices, u and v, δ(u, v) to be the length of the minimum length path from u to v. (δ(u, u) = 0 by considering path of 0 edges from u to itself.) Single Source Shortest Paths: The single source shortest path problem is as follows. We are given a directed graph with nonnegative edge weights G = (V, E) and a distinguished source vertex, s ∈ V . The problem is to determine the distance from the source vertex to every vertex in the graph. It is possible to have graphs with negative edges, but in order for the shortest path to be well defined, we need to add the requirement that there be no cycles whose total cost is negative (otherwise you make the path infinitely short by cycling forever through such a cycle). The text discusses the Bellman-Ford algorithm for finding shortest paths assuming negative weight edges but no negative-weight cycles are present. We will discuss a simple greedy algorithm, called Dijkstra’s algorithm, which assumes there are no negative edge weights. We will stress the task of computing the minimum distance from the source to each vertex. Computing the actual path will be a fairly simple extension. As in breadth-first search, for each vertex we will have a pointer pred[v] which points back to the source. By following the predecessor pointers backwards from any vertex, we will construct the reversal of the shortest path to v. Shortest Paths and Relaxation: The basic structure of Dijkstra’s algorithm is to maintain an estimate of the shortest path for each vertex, call this d[v]. (NOTE: Don’t confuse d[v] with the d[v] in the DFS algorithm. They are completely different.) Intuitively d[v] will be the length of the shortest path that the algorithm knows of from s to v. This, value will always greater than or equal to the true shortest path distance from s to v. Initially, we know of no paths, so d[v] = ∞. Initially d[s] = 0 and all the other d[v] values are set to ∞. As the algorithm goes on, and sees more and more vertices, it attempts to update d[v] for each vertex in the graph, until all the d[v] values converge to the true shortest distances. The process by which an estimate is updated is called relaxation. Here is how relaxation works. Intuitively, if you can see that your solution is not yet reached an optimum value, then push it a little closer to the optimum. In particular, if you discover a path from s to v shorter than d[v], then you need to update d[v]. This notion is common to many optimization algorithms. Consider an edge from a vertex u to v whose weight is w(u, v). Suppose that we have already computed current estimates on d[u] and d[v]. We know that there is a path from s to u of weight d[u]. By taking this path and following it with the edge (u, v) we get a path to v of length d[u]+ w(u, v). If this path is better than the existing path of length d[v] to v, we should update d[v] to the value d[u] + w(u, v). This is illustrated in Fig. 38. We should also remember that the shortest path to v passes through u, which we do by updating v’s predecessor pointer. u u 3 5 3 5 v relax(u,v) v s 11 s 8 0 0 Fig. 38: Relaxation. Relaxing an edge Relax(u,v) { if (d[u] + w(u,v) < d[v]) { // is the path through u shorter? d[v] = d[u] + w(u,v) // yes, then take it pred[v] = u // record that we go through u } } Lecture Notes 50 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-50-2048.jpg)
![Observe that whenever we set d[v] to a finite value, there is always evidence of a path of that length. Therefore d[v] ≥ δ(s, v). If d[v] = δ(s, v), then further relaxations cannot change its value. It is not hard to see that if we perform Relax(u, v) repeatedly over all edges of the graph, the d[v] values will eventually converge to the final true distance value from s. The cleverness of any shortest path algorithm is to perform the updates in a judicious manner, so the convergence is as fast as possible. In particular, the best possible would be to order relaxation operations in such a way that each edge is relaxed exactly once. Dijkstra’s algorithm does exactly this. Dijkstra’s Algorithm: Dijkstra’s algorithm is based on the notion of performing repeated relaxations. Dijkstra’s algorithm operates by maintaining a subset of vertices, S ⊆ V , for which we claim we “know” the true distance, that is d[v] = δ(s, v). Initially S = ∅, the empty set, and we set d[s] = 0 and all others to +∞. One by one we select vertices from V − S to add to S. The set S can be implemented using an array of vertex colors. Initially all vertices are white, and we set color[v] = black to indicate that v ∈ S. How do we select which vertex among the vertices of V − S to add next to S? Here is where greedy selection comes in. Dijkstra recognized that the best way in which to perform relaxations is by increasing order of distance from the source. This way, whenever a relaxation is being performed, it is possible to infer that result of the relaxation yields the final distance value. To implement this, for each vertex in u ∈ V − S, we maintain a distance estimate d[u]. The greedy thing to do is to take the vertex of V − S for which d[u] is minimum, that is, take the unprocessed vertex that is closest (by our estimate) to s. Later we will justify why this is the proper choice. In order to perform this selection efficiently, we store the vertices of V − S in a priority queue (e.g. a heap), where the key value of each vertex u is d[u]. Note the similarity with Prim’s algorithm, although a different key value is used there. Also recall that if we implement the priority queue using a heap, we can perform the operations Insert(), Extract Min(), and Decrease Key(), on a priority queue of size n each in O(log n) time. Each vertex “knows” its location in the priority queue (e.g. has a cross reference link to the priority queue entry), and each entry in the priority queue “knows” which vertex it represents. It is important when implementing the priority queue that this cross reference information is updated. Here is Dijkstra’s algorithm. (Note the remarkable similarity to Prim’s algorithm.) An example is presented in Fig. 39. Notice that the coloring is not really used by the algorithm, but it has been included to make the connection with the correctness proof a little clearer. Because of the similarity between this and Prim’s algorithm, the running time is the same, namely Θ(E log V ). Correctness: Recall that d[v] is the distance value assigned to vertex v by Dijkstra’s algorithm, and let δ(s, v) denote the length of the true shortest path from s to v. To see that Dijkstra’s algorithm correctly gives the final true distances, we need to show that d[v] = δ(s, v) when the algorithm terminates. This is a consequence of the following lemma, which states that once a vertex u has been added to S (i.e. colored black), d[u] is the true shortest distance from s to u. Since at the end of the algorithm, all vertices are in S, then all distance estimates are correct. Lemma: When a vertex u is added to S, d[u] = δ(s, u). Proof: It will simplify the proof conceptually if we assume that all the edge weights are strictly positive (the general case of nonnegative edges is presented in the text). Suppose to the contrary that at some point Dijkstra’s algorithm first attempts to add a vertex u to S for which d[u] = δ(s, u). By our observations about relaxation, d[u] is never less than δ(s, u), thus we have d[u] > δ(s, u). Consider the situation just prior to the insertion of u. Consider the true shortest path from s to u. Because s ∈ S and u ∈ V − S, at some point this path must first jump out of S. Let (x, y) be the edge taken by the path, where x ∈ S and y ∈ V − S. (Note that it may be that x = s and/or y = u). Lecture Notes 51 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-51-2048.jpg)
![Dijkstra’s Algorithm Dijkstra(G,w,s) { for each (u in V) { // initialization d[u] = +infinity color[u] = white pred[u] = null } d[s] = 0 // dist to source is 0 Q = new PriQueue(V) // put all vertices in Q while (Q.nonEmpty()) { // until all vertices processed u = Q.extractMin() // select u closest to s for each (v in Adj[u]) { if (d[u] + w(u,v) < d[v]) { // Relax(u,v) d[v] = d[u] + w(u,v) Q.decreaseKey(v, d[v]) pred[v] = u } } color[u] = black } [The pred pointers define an ‘‘inverted’’ shortest path tree] } ? 1 ? 0 1 2 1 10 7 ? 5 7 7 7 s 8 s 8 s 8 0 3 2 4 5 0 3 2 4 5 0 3 2 4 5 2 2 2 ? ? 2 ? 2 7 5 5 5 5 1 6 5 1 7 5 1 5 6 6 6 7 7 7 s 8 s 8 s 8 0 3 2 4 5 0 3 2 4 5 0 3 2 4 5 2 2 2 2 7 2 7 2 7 5 5 5 Fig. 39: Dijkstra’s Algorithm example. S pred[u] u s shorter path from s to u? y x d[y] > d[u] Fig. 40: Correctness of Dijkstra’s Algorithm. Lecture Notes 52 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-52-2048.jpg)
![We argue that y = u. Why? Since x ∈ S we have d[x] = δ(s, x). (Since u was the first vertex added to S which violated this, all prior vertices satisfy this.) Since we applied relaxation to x when it was added, we would have set d[y] = d[x] + w(x, y) = δ(s, y). Thus d[y] is correct, and by hypothesis, d[u] is not correct, so they cannot be the same. Now observe that since y appears somewhere along the shortest path from s to u (but not at u) and all subsequent edges following y are of positive weight, we have δ(s, y) < δ(s, u), and thus d[y] = δ(s, y) < δ(s, u) < d[u]. Thus y would have been added to S before u, in contradiction to our assumption that u is the next vertex to be added to S. Lecture 15: All-Pairs Shortest Paths Read: Section 25.2 in CLRS. All-Pairs Shortest Paths: We consider the generalization of the shortest path problem, to computing shortest paths between all pairs of vertices. Let G = (V, E) be a directed graph with edge weights. If (u, v) E, is an edge of G, then the weight of this edge is denoted w(u, v). Recall that the cost of a path is the sum of edge weights along the path. The distance between two vertices δ(u, v) is the cost of the minimum cost path between them. We will allow G to have negative cost edges, but we will not allow G to have any negative cost cycles. We consider the problem of determining the cost of the shortest path between all pairs of vertices in a weighted directed graph. We will present a Θ(n3 ) algorithm, called the Floyd-Warshall algorithm. This algorithm is based on dynamic programming. For this algorithm, we will assume that the digraph is represented as an adjacency matrix, rather than the more common adjacency list. Although adjacency lists are generally more efficient for sparse graphs, storing all the inter-vertex distances will require Ω(n2 ) storage, so the savings is not justified here. Because the algorithm is matrix-based, we will employ common matrix notation, using i, j and k to denote vertices rather than u, v, and w as we usually do. Input Format: The input is an n × n matrix w of edge weights, which are based on the edge weights in the digraph. We let wij denote the entry in row i and column j of w. 0 if i = j, wij = w(i, j) if i = j and (i, j) ∈ E, +∞ if i = j and (i, j) ∈ E. / Setting wij = ∞ if there is no edge, intuitively means that there is no direct link between these two nodes, and hence the direct cost is infinite. The reason for setting wii = 0 is that there is always a trivial path of length 0 (using no edges) from any vertex to itself. (Note that in digraphs it is possible to have self-loop edges, and so w(i, i) may generally be nonzero. It cannot be negative, since we assume that there are no negative cost cycles, and if it is positive, there is no point in using it as part of any shortest path.) The output will be an n × n distance matrix D = dij where dij = δ(i, j), the shortest path cost from vertex i to j. Recovering the shortest paths will also be an issue. To help us do this, we will also compute an auxiliary matrix mid[i, j]. The value of mid[i, j] will be a vertex that is somewhere along the shortest path from i to j. If the shortest path travels directly from i to j without passing through any other vertices, then mid[i, j] will be set to null. These intermediate values behave somewhat like the predecessor pointers in Dijkstra’s algorithm, in order to reconstruct the final shortest path in Θ(n) time. Lecture Notes 53 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-53-2048.jpg)
![This suggests the following recursive rule (the DP formulation) for computing d(k) , which is illustrated in Fig. 41(b). (0) dij = wij , (k) (k−1) (k−1) (k−1) dij = min dij , dik + dkj for k ≥ 1. (n) The final answer is dij because this allows all possible vertices as intermediate vertices. We could write a (k) recursive program to compute dij , but this will be prohibitively slow because the same value may be reevaluated many times. Instead, we compute it by storing the values in a table, and looking the values up as we need them. Here is the complete algorithm. We have also included mid-vertex pointers, mid[i, j] for extracting the final shortest paths. We will leave the extraction of the shortest path as an exercise. Floyd-Warshall Algorithm Floyd_Warshall(int n, int w[1..n, 1..n]) { array d[1..n, 1..n] for i = 1 to n do { // initialize for j = 1 to n do { d[i,j] = W[i,j] mid[i,j] = null } } for k = 1 to n do // use intermediates {1..k} for i = 1 to n do // ...from i for j = 1 to n do // ...to j if (d[i,k] + d[k,j]) < d[i,j]) { d[i,j] = d[i,k] + d[k,j] // new shorter path length mid[i,j] = k // new path is through k } return d // matrix of distances } An example of the algorithm’s execution is shown in Fig. 42. Clearly the algorithm’s running time is Θ(n3 ). The space used by the algorithm is Θ(n2 ). Observe that we deleted all references to the superscript (k) in the code. It is left as an exercise that this does not affect the correctness of the algorithm. (Hint: The danger is that values may be overwritten and then used later in the same phase. Consider which entries might be overwritten and then reused, they occur in row k and column k. It can be shown that the overwritten values are equal to their original values.) Lecture 16: NP-Completeness: Languages and NP Read: Chapt 34 in CLRS, up through section 34.2. Complexity Theory: At this point of the semester we have been building up your “bag of tricks” for solving algorith- mic problems. Hopefully when presented with a problem you now have a little better idea of how to go about solving the problem. What sort of design paradigm should be used (divide-and-conquer, DFS, greedy, dynamic programming), what sort of data structures might be relevant (trees, heaps, graphs) and what representations would be best (adjacency list, adjacency matrices), what is the running time of your algorithm. All of this is fine if it helps you discover an acceptably efficient algorithm to solve your problem. The question that often arises in practice is that you have tried every trick in the book, and nothing seems to work. Although Lecture Notes 55 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-55-2048.jpg)
![Lecture 20: Subset Sum Read: Sections 34.5.5 in CLR. Subset Sum: The Subset Sum problem (SS) is the following. Given a finite set S of positive integers S = {w1 , w2 , . . . , wn } and a target value, t, we want to know whether there exists a subset S ⊆ S that sums exactly to t. This problem is a simplified version of the 0-1 Knapsack problem, presented as a decision problem. Recall that in the 0-1 Knapsack problem, we are given a collection of objects, each with an associated weight wi and associated value vi . We are given a knapsack of capacity W . The objective is to take as many objects as can fit in the knapsack’s capacity so as to maximize the value. (In the fractional knapsack we could take a portion of an object. In the 0-1 Knapsack we either take an object entirely or leave it.) In the simplest version, suppose that the value is the same as the weight, vi = wi . (This would occur for example if all the objects were made of the same material, say, gold.) Then, the best we could hope to achieve would be to fill the knapsack entirely. By setting t = W , we see that the subset sum problem is equivalent to this simplified version of the 0-1 Knapsack problem. It follows that if we can show that this simpler version is NP-complete, then certainly the more general 0-1 Knapsack problem (stated as a decision problem) is also NP-complete. Consider the following example. S = {3, 6, 9, 12, 15, 23, 32} and t = 33. The subset S = {6, 12, 15} sums to t = 33, so the answer in this case is yes. If t = 34 the answer would be no. Dynamic Programming Solution: There is a dynamic programming algorithm which solves the Subset Sum prob- lem in O(n · t) time.2 The quantity n · t is a polynomial function of n. This would seem to imply that the Subset Sum problem is in P. But there is a important catch. Recall that in all NP-complete problems we assume (1) running time is measured as a function of input size (number of bits) and (2) inputs must be encoded in a reasonable succinct manner. Let us assume that the numbers wi and t are all b-bit numbers represented in base 2, using the fewest number of bits possible. Then the input size is O(nb). The value of t may be as large as 2b . So the resulting algorithm has a running time of O(n2b ). This is polynomial in n, but exponential in b. Thus, this running time is not polynomial as a function of the input size. Note that an important consequence of this observation is that the SS problem is not hard when the numbers involved are small. If the numbers involved are of a fixed number of bits (a constant independent of n), then the problem is solvable in polynomial time. However, we will show that in the general case, this problem is NP-complete. SS is NP-complete: The proof that Subset Sum (SS) is NP-complete involves the usual two elements. (i) SS ∈ NP. (ii) Some known NP-complete problem is reducible to SS. In particular, we will show that Vertex Cover (VC) is reducible to SS, that is, VC ≤P SS. To show that SS is in NP, we need to give a verification procedure. Given S and t, the certificate is just the indices of the numbers that form the subset S . We can add two b-bit numbers together in O(b) time. So, in polynomial time we can compute the sum of elements in S , and verify that this sum equals t. For the remainder of the proof we show how to reduce vertex cover to subset sum. We want a polynomial time computable function f that maps an instance of the vertex cover (a graph G and integer k) to an instance of the subset sum problem (a set of integers S and target integer t) such that G has a vertex cover of size k if and only if S has a subset summing to t. Thus, if subset sum were solvable in polynomial time, so would vertex cover. 2 We will leave this as an exercise, but the formulation is, for 0 ≤ i ≤ n and 0 ≤ t ≤ t, S[i, t ] = 1 if there is a subset of {w , w , . . . , w } 1 2 i that sums to t , and 0 otherwise. The ith row of this table can be computed in O(t) time, given the contents of the (i − 1)-st row. Lecture Notes 73 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-73-2048.jpg)
![V(c2) 5 6 5 6 5 c2 4 9 7 4 9 7 4 8 9 8 9 5 7 c3 V(c3) 5 6 5 5 6 5 V(c1) c1 6 5 6 8 6 8 Input graph (k=3) Optimumum Cost = 7 Fig. 61: The k-center problem with optimum centers ci and neighborhood sets V (ci ). Finally, we define the overall bottleneck distance to be D(C) = max D(ci ). ci ∈C This is the maximum distance of any vertex from its nearest center. This distance is critical because it represents the customer that must travel farthest to get to the nearest facility, the bottleneck vertex. Given this notation, we can now formally define the problem. k-center problem: Given a weighted undirected graph G = (V, E), and an integer k ≤ |V |, find a subset C ⊆ V of size k such that D(C) is minimized. The decision-problem formulation of the k-center problem is NP-complete (reduction from dominating set). A brute force solution to this problem would involve enumerating all k-element of subsets of V , and computing D(C) for each one. However, letting n = |V | and k, the number of possible subsets is n = Θ(nk ). If k is k a function of n (which is reasonable), then this an exponential number of subsets. Given that the problem is NP-complete, it is highly unlikely that a significantly more efficient exact algorithm exists in the worst-case. We will show that there does exist an efficient approximation algorithm for the problem. Greedy Approximation Algorithm: Our approximation algorithm is based on a simple greedy algorithm that pro- duces a bottleneck distance D(C) that is not more than twice the optimum bottleneck distance. We begin by letting the first center c1 be any vertex in the graph (the lower left vertex, say, in the figure below). Compute the distances between this vertex and all the other vertices in the graph (Fig. 62(b)). Consider the vertex that is farthest from this center (the upper right vertex at distance 23 in the figure). This the bottleneck vertex for {c1 }. We would like to select the next center so as to reduce this distance. So let us just make it the next center, called c2 . Then again we compute the distances from each vertex in the graph to the closer of c1 and c2 . (See Fig. 62(c) where dashed lines indicate which vertices are closer to which center). Again we consider the bottleneck vertex for the current centers {c1 , c2 }. We place the next center at this vertex (see Fig. 62(d)). Again we compute the distances from each vertex to its nearest center. Repeat this until all k centers have been selected. In Fig. 62(d), the final three greedy centers are shaded, and the final bottleneck distance is 11. Although the greedy approach has a certain intuitive appeal (because it attempts to find the vertex that gives the bottleneck distance, and then puts a center right on this vertex), it is not optimal. In the example shown in the figure, the optimum solution (shown on the right) has a bottleneck cost of 9, which beats the 11 that the greedy algorithm gave. Here is a summary of the algorithm. For each vertex u, let d[u] denote the distance to the nearest center. We know from Dijkstra’s algorithm how to compute the shortest path from a single source to all other vertices in the graph. One way to solve the distance computation step above would be to invoke Dijkstra’s algorithm i times. But there is an easier way. We can modify Dijkstra’s algorithm to operate as a multiple source algorithm. In particular, in the initialization of Dijkstra’s single source algorithm, it sets d[s] = 0 and pred[s] = null. In Lecture Notes 82 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-82-2048.jpg)
![5 6 5 6 5 6 5 6 14 19 23 11 6 c2 11 6 c2 9 7 4 9 7 4 9 7 4 9 7 4 8 9 8 9 8 9 8 9 5 12 19 5 12 4 5 c3 4 5 6 5 5 6 5 5 6 5 5 6 5 c1 6 14 c1 6 9 c1 6 9 6 8 6 8 6 8 6 8 (a) (b) (c) (d) Greedy Cost = 11 Fig. 62: Greedy approximation to k-center. Greedy Approximation for k-center KCenterApprox(G, k) { C = empty_set for each u in V do // initialize distances d[u] = INFINITY for i = 1 to k do { // main loop Find the vertex u such that d[u] is maximum Add u to C // u is the current bottleneck vertex // update distances Compute the distance from each vertex v to its closest vertex in C, denoted d[v] } return C // final centers } Lecture Notes 83 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-83-2048.jpg)
![A Major Improvement: The problem with the previous algorithm is that, even though we have cut the number of comparisons roughly in half, each point is still making lots of comparisons. Can we save time by making only one comparison for each point? The inner while loop is testing to see whether any point that follows P [i] in the sorted list has a larger y-coordinate. This suggests, that if we knew which point among P [i + 1, . . . , n] had the maximum y-coordinate, we could just test against that point. How can we do this? Here is a simple observation. For any set of points, the point with the maximum y- coordinate is the maximal point with the smallest x-coordiante. This suggests that we can sweep the points backwards, from right to left. We keep track of the index j of the most recently seen maximal point. (Initially the rightmost point is maximal.) When we encounter the point P [i], it is maximal if and only if P [i].y ≥ P [j].y. This suggests the following algorithm. Max Dominance: Sort and Reverse Scan MaxDom3(P, n) { Sort P in ascending order by x-coordinate; output P[n]; // last point is always maximal j = n; for i = n-1 downto 1 { if (P[i].y >= P[j].y) { // is P[i] maximal? output P[i]; // yes..output it j = i; // P[i] has the largest y so far } } } The running time of the for-loop is obviously O(n), because there is just a single loop that is executed n − 1 times, and the code inside takes constant time. The total running time is dominated by the O(n log n) sorting time, for a total of O(n log n) time. How much of an improvement is this? Probably the most accurate way to find out would be to code the two up, and compare their running times. But just to get a feeling, let’s look at the ratio of the running times, ignoring constant factors: n2 n = . n lg n lg n (I use the notation lg n to denote the logarithm base 2, ln n to denote the natural logarithm (base e) and log n when I do not care about the base. Note that a change in base only affects the value of a logarithm function by a constant amount, so inside of O-notation, we will usually just write log n.) For relatively small values of n (e.g. less than 100), both algorithms are probably running fast enough that the difference will be practically negligible. (Rule 1 of algorithm optimization: Don’t optimize code that is already fast enough.) On larger inputs, say, n = 1, 000, the ratio of n to log n is about 1000/10 = 100, so there is a 100- to-1 ratio in running times. Of course, we would need to factor in constant factors, but since we are not using any really complex data structures, it is hard to imagine that the constant factors will differ by more than, say, 10. For even larger inputs, say, n = 1, 000, 000, we are looking at a ratio of roughly 1, 000, 000/20 = 50, 000. This is quite a significant difference, irrespective of the constant factors. Divide and Conquer Approach: One problem with the previous algorithm is that it relies on sorting. This is nice and clean (since it is usually easy to get good code for sorting without troubling yourself to write your own). However, if you really wanted to squeeze the most efficiency out of your code, you might consider whether you can solve this problem without invoking a sorting algorithm. One of the basic maxims of algorithm design is to first approach any problem using one of the standard algorithm design paradigms, e.g. divide and conquer, dynamic programming, greedy algorithms, depth-first search. We will talk more about these methods as the semester continues. For this problem, divide-and-conquer is a natural method to choose. What is this paradigm? Lecture Notes 96 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-96-2048.jpg)
![Divide: Divide the problem into two subproblems (ideally of approximately equal sizes), Conquer: Solve each subproblem recursively, and Combine: Combine the solutions to the two subproblems into a global solution. How shall we divide the problem? I can think of a couple of ways. One is similar to how MergeSort operates. Just take the array of points P [1..n], and split into two subarrays of equal size P [1..n/2] and P [n/2 + 1..n]. Because we do not sort the points, there is no particular relationship between the points in one side of the list from the other. Another approach, which is more reminiscent of QuickSort is to select a random element from the list, called a pivot, x = P [r], where r is a random integer in the range from 1 to n, and then partition the list into two sublists, those elements whose x-coordinates are less than or equal to x and those that greater than x. This will not be guaranteed to split the list into two equal parts, but on average it can be shown that it does a pretty good job. Let’s consider the first method. (The quicksort method will also work, but leads to a tougher analysis.) Here is more concrete outline. We will describe the algorithm at a very high level. The input will be a point array, and a point array will be returned. The key ingredient is a function that takes the maxima of two sets, and merges them into an overall set of maxima. Max Dominance: Divide-and-Conquer MaxDom4(P, n) { if (n == 1) return {P[1]}; // one point is trivially maximal m = n/2; // midpoint of list M1 = MaxDom4(P[1..m], m); // solve for first half M2 = MaxDom4(P[m+1..n], n-m); // solve for second half return MaxMerge(M1, M2); // merge the results } The general process is illustrated below. The main question is how the procedure Max Merge() is implemented, because it does all the work. Let us assume that it returns a list of points in sorted order according to x-coordinates of the maximal points. Observe that if a point is to be maximal overall, then it must be maximal in one of the two sublists. However, just because a point is maximal in some list, does not imply that it is globally maximal. (Consider point (7, 10) in the example.) However, if it dominates all the points of the other sublist, then we can assert that it is maximal. I will describe the procedure at a very high level. It operates by walking through each of the two sorted lists of maximal points. It maintains two pointers, one pointing to the next unprocessed item in each list. Think of these as fingers. Take the finger pointing to the point with the smaller x-coordinate. If its y-coordinate is larger than the y-coordinate of the point under the other finger, then this point is maximal, and is copied to the next position of the result list. Otherwise it is not copied. In either case, we move to the next point in the same list, and repeat the process. The result list is returned. The details will be left as an exercise. Observe that because we spend a constant amount of time processing each point (either copying it to the result list or skipping over it) the total execution time of this procedure is O(n). Recurrences: How do we analyze recursive procedures like this one? If there is a simple pattern to the sizes of the recursive calls, then the best way is usually by setting up a recurrence, that is, a function which is defined recursively in terms of itself. We break the problem into two subproblems of size roughly n/2 (we will say exactly n/2 for simplicity), and the additional overhead of merging the solutions is O(n). We will ignore constant factors, writing O(n) just as n, giving: T (n) = 1 if n = 1, T (n) = 2T (n/2) + n if n > 1. Lecture Notes 97 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-97-2048.jpg)
![statistics it is common to return the average of these two elements. We will define the median to be either of these elements. Medians are useful as measures of the central tendency of a set, especially when the distribution of values is highly skewed. For example, the median income in a community is likely to be more meaningful measure of the central tendency than the average is, since if Bill Gates lives in your community then his gigantic income may significantly bias the average, whereas it cannot have a significant influence on the median. They are also useful, since in divide-and-conquer applications, it is often desirable to partition a set about its median value, into two sets of roughly equal size. Today we will focus on the following generalization, called the selection problem. Selection: Given a set A of n distinct numbers and an integer k, 1 ≤ k ≤ n, output the element of A of rank k. The selection problem can easily be solved in Θ(n log n) time, simply by sorting the numbers of A, and then returning A[k]. The question is whether it is possible to do better. In particular, is it possible to solve this problem in Θ(n) time? We will see that the answer is yes, and the solution is far from obvious. The Sieve Technique: The reason for introducing this algorithm is that it illustrates a very important special case of divide-and-conquer, which I call the sieve technique. We think of divide-and-conquer as breaking the problem into a small number of smaller subproblems, which are then solved recursively. The sieve technique is a special case, where the number of subproblems is just 1. The sieve technique works in phases as follows. It applies to problems where we are interested in finding a single item from a larger set of n items. We do not know which item is of interest, however after doing some amount of analysis of the data, taking say Θ(nk ) time, for some constant k, we find that we do not know what the desired item is, but we can identify a large enough number of elements that cannot be the desired value, and can be eliminated from further consideration. In particular “large enough” means that the number of items is at least some fixed constant fraction of n (e.g. n/2, n/3, 0.0001n). Then we solve the problem recursively on whatever items remain. Each of the resulting recursive solutions then do the same thing, eliminating a constant fraction of the remaining set. Applying the Sieve to Selection: To see more concretely how the sieve technique works, let us apply it to the selec- tion problem. Recall that we are given an array A[1..n] and an integer k, and want to find the k-th smallest element of A. Since the algorithm will be applied inductively, we will assume that we are given a subarray A[p..r] as we did in MergeSort, and we want to find the kth smallest item (where k ≤ r − p + 1). The initial call will be to the entire array A[1..n]. There are two principal algorithms for solving the selection problem, but they differ only in one step, which involves judiciously choosing an item from the array, called the pivot element, which we will denote by x. Later we will see how to choose x, but for now just think of it as a random element of A. We then partition A into three parts. A[q] contains the element x, subarray A[p..q − 1] will contain all the elements that are less than x, and A[q + 1..r], will contain all the element that are greater than x. (Recall that we assumed that all the elements are distinct.) Within each subarray, the items may appear in any order. This is illustrated below. It is easy to see that the rank of the pivot x is q − p + 1 in A[p..r]. Let xRank = q − p + 1. If k = xRank , then the pivot is the kth smallest, and we may just return it. If k < xRank , then we know that we need to recursively search in A[p..q −1] and if k > xRank then we need to recursively search A[q +1..r]. In this latter case we have eliminated q smaller elements, so we want to find the element of rank k − q. Here is the complete pseudocode. Notice that this algorithm satisfies the basic form of a sieve algorithm. It analyzes the data (by choosing the pivot element and partitioning) and it eliminates some part of the data set, and recurses on the rest. When k = xRank then we get lucky and eliminate everything. Otherwise we either eliminate the pivot and the right subarray or the pivot and the left subarray. We will discuss the details of choosing the pivot and partitioning later, but assume for now that they both take Θ(n) time. The question that remains is how many elements did we succeed in eliminating? If x is the largest Lecture Notes 104 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-104-2048.jpg)
![pivot p r 5 9 2 6 4 1 3 7 Before partitioing p q r 3 1 2 4 6 9 5 7 After partitioing x A[p..q−1] < x A[q+1..r] > x 5 3 9 1 2 2 6 4 x_rnk=4 4 6 6 6 6 5 1 9 9 5 5 6 3 5 5 7 x_rnk=3 x_rnk=2 (DONE!) 7 7 7 9 Initial Partition Recurse Partition Recurse Partition (k=6) (pivot = 4) (k=6−4=2) (pivot = 7) (k=2) (pivot = 6) Fig. 70: Selection Algorithm. Selection by the Sieve Technique Select(array A, int p, int r, int k) { // return kth smallest of A[p..r] if (p == r) return A[p] // only 1 item left, return it else { x = ChoosePivot(A, p, r) // choose the pivot element q = Partition(A, p, r, x) // partition <A[p..q-1], x, A[q+1..r]> xRank = q - p + 1 // rank of the pivot if (k == xRank) return x // the pivot is the kth smallest else if (k < xRank) return Select(A, p, q-1, k) // select from left subarray else return Select(A, q+1, r, k-xRank)// select from right subarray } } Lecture Notes 105 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-105-2048.jpg)
![or smallest element in the array, then we may only succeed in eliminating one element with each phase. In fact, if x is one of the smallest elements of A or one of the largest, then we get into trouble, because we may only eliminate it and the few smaller or larger elements of A. Ideally x should have a rank that is neither too large nor too small. Let us suppose for now (optimistically) that we are able to design the procedure Choose Pivot in such a way that is eliminates exactly half the array with each phase, meaning that we recurse on the remaining n/2 elements. This would lead to the following recurrence. 1 if n = 1, T (n) = T (n/2) + n otherwise. We can solve this either by expansion (iteration) or the Master Theorem. If we expand this recurrence level by level we see that we get the summation ∞ ∞ n n n 1 T (n) = n + + + · · · ≤ = n . 2 4 i=0 2i i=0 2i ∞ Recall the formula for the infinite geometric series. For any c such that |c| < 1, i=0 c i = 1/(1 − c). Using this we have T (n) ≤ 2n ∈ O(n). (This only proves the upper bound on the running time, but it is easy to see that it takes at least Ω(n) time, so the total running time is Θ(n).) This is a bit counterintuitive. Normally you would think that in order to design a Θ(n) time algorithm you could only make a single, or perhaps a constant number of passes over the data set. In this algorithm we make many passes (it could be as many as lg n). However, because we eliminate a constant fraction of elements with each phase, we get this convergent geometric series in the analysis, which shows that the total running time is indeed linear in n. This lesson is well worth remembering. It is often possible to achieve running times in ways that you would not expect. Note that the assumption of eliminating half was not critical. If we eliminated even one per cent, then the recurrence would have been T (n) = T (99n/100) + n, and we would have gotten a geometric series involving 99/100, which is still less than 1, implying a convergent series. Eliminating any constant fraction would have been good enough. Choosing the Pivot: There are two issues that we have left unresolved. The first is how to choose the pivot element, and the second is how to partition the array. Both need to be solved in Θ(n) time. The second problem is a rather easy programming exercise. Later, when we discuss QuickSort, we will discuss partitioning in detail. For the rest of the lecture, let’s concentrate on how to choose the pivot. Recall that before we said that we might think of the pivot as a random element of A. Actually this is not such a bad idea. Let’s see why. The key is that we want the procedure to eliminate at least some constant fraction of the array after each parti- tioning step. Let’s consider the top of the recurrence, when we are given A[1..n]. Suppose that the pivot x turns out to be of rank q in the array. The partitioning algorithm will split the array into A[1..q − 1] < x, A[q] = x and A[q + 1..n] > x. If k = q, then we are done. Otherwise, we need to search one of the two subarrays. They are of sizes q − 1 and n − q, respectively. The subarray that contains the kth smallest element will generally depend on what k is, so in the worst case, k will be chosen so that we have to recurse on the larger of the two subarrays. Thus if q > n/2, then we may have to recurse on the left subarray of size q − 1, and if q < n/2, then we may have to recurse on the right subarray of size n − q. In either case, we are in trouble if q is very small, or if q is very large. If we could select q so that it is roughly of middle rank, then we will be in good shape. For example, if n/4 ≤ q ≤ 3n/4, then the larger subarray will never be larger than 3n/4. Earlier we said that we might think of the pivot as a random element of the array A. Actually this works pretty well in practice. The reason is that Lecture Notes 106 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-106-2048.jpg)
![roughly half of the elements lie between ranks n/4 and 3n/4, so picking a random element as the pivot will succeed about half the time to eliminate at least n/4. Of course, we might be continuously unlucky, but a careful analysis will show that the expected running time is still Θ(n). We will return to this later. Instead, we will describe a rather complicated method for computing a pivot element that achieves the desired properties. Recall that we are given an array A[1..n], and we want to compute an element x whose rank is (roughly) between n/4 and 3n/4. We will have to describe this algorithm at a very high level, since the details are rather involved. Here is the description for Select Pivot: Groups of 5: Partition A into groups of 5 elements, e.g. A[1..5], A[6..10], A[11..15], etc. There will be exactly m = n/5 such groups (the last one might have fewer than 5 elements). This can easily be done in Θ(n) time. Group medians: Compute the median of each group of 5. There will be m group medians. We do not need an intelligent algorithm to do this, since each group has only a constant number of elements. For example, we could just BubbleSort each group and take the middle element. Each will take Θ(1) time, and repeating this n/5 times will give a total running time of Θ(n). Copy the group medians to a new array B. Median of medians: Compute the median of the group medians. For this, we will have to call the selection algorithm recursively on B, e.g. Select(B, 1, m, k), where m = n/5 , and k = (m + 1)/2 . Let x be this median of medians. Return x as the desired pivot. The algorithm is illustrated in the figure below. To establish the correctness of this procedure, we need to argue that x satisfies the desired rank properties. 14 32 23 5 10 60 29 6 2 3 5 8 1 11 57 2 52 44 27 21 11 14 25 12 17 10 21 29 24 43 12 17 48 1 58 24 30 23 34 19 41 39 6 30 63 34 8 55 39 37 32 52 44 27 55 58 37 25 3 64 19 41 57 43 63 64 48 60 Group Get group medians 8 3 6 2 5 11 1 10 12 14 25 17 29 21 19 23 24 30 34 39 41 27 52 37 32 44 58 55 48 63 57 43 64 60 Get median of medians (Sorting of group medians is not really performed) Fig. 71: Choosing the Pivot. 30 is the final pivot. Lemma: The element x is of rank at least n/4 and at most 3n/4 in A. Proof: We will show that x is of rank at least n/4. The other part of the proof is essentially symmetrical. To do this, we need to show that there are at least n/4 elements that are less than or equal to x. This is a bit complicated, due to the floor and ceiling arithmetic, so to simplify things we will assume that n is evenly divisible by 5. Consider the groups shown in the tabular form above. Observe that at least half of the group medians are less than or equal to x. (Because x is their median.) And for each group median, there are three elements that are less than or equal to this median within its group (because it is the median of its Lecture Notes 107 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-107-2048.jpg)
![set of discrete values with certain probabilities. More formally, it is a function that maps some some discrete sample space (the set of possible values) onto the reals (the probabilities). For 0 ≤ i ≤ n − 1, let Xi denote the random variable that indicates the number of elements assigned to the i-th bucket. Since the distribution is uniform, all of the random variables Xi have the same probability distribution, so we may as well talk about a single random variable X, which will work for any bucket. Since we are using a quadratic time algorithm to sort the elements of each bucket, we are interested in the expected sorting time, which is Θ(X 2 ). So this leads to the key question, what is the expected value of X 2 , denoted E[X 2 ]. Because the elements are assumed to be uniformly distributed, each element has an equal probability of going into any bucket, or in particular, it has a probability of p = 1/n of going into the ith bucket. So how many items do we expect will wind up in bucket i? We can analyze this by thinking of each element of A as being represented by a coin flip (with a biased coin, which has a different probability of heads and tails). With probability p = 1/n the number goes into bucket i, which we will interpret as the coin coming up heads. With probability 1 − 1/n the item goes into some other bucket, which we will interpret as the coin coming up tails. Since we assume that the elements of A are independent of each other, X is just the total number of heads we see after making n tosses with this (biased) coin. The number of times that a heads event occurs, given n independent trials in which each trial has two possible outcomes is a well-studied problem in probability theory. Such trials are called Bernoulli trials (named after the Swiss mathematician James Bernoulli). If p is the probability of getting a head, then the probability of getting k heads in n tosses is given by the following important formula n k n n! P (X = k) = p (1 − p)n−k where = . k k k!(n − k)! Although this looks messy, it is not too hard to see where it comes from. Basically pk is the probability of tossing k heads, (1 − p)n−k is the probability of tossing n − k tails, and n is the total number of different k ways that the k heads could be distributed among the n tosses. This probability distribution (as a function of k, for a given n and p) is called the binomial distribution, and is denoted b(k; n, p). If you consult a standard textbook on probability and statistics, then you will see the two important facts that we need to know about the binomial distribution. Namely, that its mean value E[X] and its variance Var[X] are E[X] = np and Var[X] = E[X 2 ] − E 2 [X] = np(1 − p). We want to determine E[X 2 ]. By the above formulas and the fact that p = 1/n we can derive this as n 1 n 2 1 E[X 2 ] = Var[X] + E 2 [X] = np(1 − p) + (np)2 = 1− + = 2− . n n n n Thus, for large n the time to insert the items into any one of the linked lists is a just shade less than 2. Summing up over all n buckets, gives a total running time of Θ(2n) = Θ(n). This is exactly what our quick-and-dirty analysis gave us, but now we know it is true with confidence. Supplemental Lecture 6: Long Integer Multiplication Read: This material on integer multiplication is not covered in CLRS. Long Integer Multiplication: The following little algorithm shows a bit more about the surprising applications of divide-and-conquer. The problem that we want to consider is how to perform arithmetic on long integers, and multiplication in particular. The reason for doing arithmetic on long numbers stems from cryptography. Most techniques for encryption are based on number-theoretic techniques. For example, the character string to be encrypted is converted into a sequence of numbers, and encryption keys are stored as long integers. Efficient Lecture Notes 109 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-109-2048.jpg)
![encryption and decryption depends on being able to perform arithmetic on long numbers, typically containing hundreds of digits. Addition and subtraction on large numbers is relatively easy. If n is the number of digits, then these algorithms run in Θ(n) time. (Go back and analyze your solution to the problem on Homework 1). But the standard algorithm for multiplication runs in Θ(n2 ) time, which can be quite costly when lots of long multiplications are needed. This raises the question of whether there is a more efficient way to multiply two very large numbers. It would seem surprising if there were, since for centuries people have used the same algorithm that we all learn in grade school. In fact, we will see that it is possible. Divide-and-Conquer Algorithm: We know the basic grade-school algorithm for multiplication. We normally think of this algorithm as applying on a digit-by-digit basis, but if we partition an n digit number into two “super digits” with roughly n/2 each into longer sequences, the same multiplication rule still applies. n n/2 n/2 w x A y z B wz xz wy xy wy wz + xy xz Product Fig. 72: Long integer multiplication. To avoid complicating things with floors and ceilings, let’s just assume that the number of digits n is a power of 2. Let A and B be the two numbers to multiply. Let A[0] denote the least significant digit and let A[n−1] denote the most significant digit of A. Because of the way we write numbers, it is more natural to think of the elements of A as being indexed in decreasing order from left to right as A[n − 1..0] rather than the usual A[0..n − 1]. Let m = n/2. Let w = A[n − 1..m] x = A[m − 1..0] and y = B[n − 1..m] z = B[m − 1..0]. If we think of w, x, y and z as n/2 digit numbers, we can express A and B as A = w · 10m + x B = y · 10m + z, and their product is mult(A, B) = mult(w, y)102m + (mult(w, z) + mult(x, y))10m + mult(x, z). The operation of multiplying by 10m should be thought of as simply shifting the number over by m positions to the right, and so is not really a multiplication. Observe that all the additions involve numbers involving roughly n/2 digits, and so they take Θ(n) time each. Thus, we can express the multiplication of two long integers as the result of four products on integers of roughly half the length of the original, and a constant number of additions and shifts, each taking Θ(n) time. This suggests that if we were to implement this algorithm, its running time would be given by the following recurrence 1 if n = 1, T (n) = 4T (n/2) + n otherwise. Lecture Notes 110 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-110-2048.jpg)
![problem where the burglar can take a fraction of an object for a fraction of the value and weight. This is much easier to solve.) More formally, given v1 , v2 , . . . , vn and w1 , w2 . . . , wn , and W > 0, we wish to determine the subset T ⊆ {1, 2, . . . , n} (of objects to “take”) that maximizes vi , i∈T subject to wi ≤ W. i∈T Let us assume that the vi ’s, wi ’s and W are all positive integers. It turns out that this problem is NP-complete, and so we cannot really hope to find an efficient solution. However if we make the same sort of assumption that we made in counting sort, we can come up with an efficient solution. We assume that the wi ’s are small integers, and that W itself is a small integer. We show that this problem can be solved in O(nW ) time. (Note that this is not very good if W is a large integer. But if we truncate our numbers to lower precision, this gives a reasonable approximation algorithm.) Here is how we solve the problem. We construct an array V [0..n, 0..W ]. For 1 ≤ i ≤ n, and 0 ≤ j ≤ W , the entry V [i, j] we will store the maximum value of any subset of objects {1, 2, . . . , i} that can fit into a knapsack of weight j. If we can compute all the entries of this array, then the array entry V [n, W ] will contain the maximum value of all n objects that can fit into the entire knapsack of weight W . To compute the entries of the array V we will imply an inductive approach. As a basis, observe that V [0, j] = 0 for 0 ≤ j ≤ W since if we have no items then we have no value. We consider two cases: Leave object i: If we choose to not take object i, then the optimal value will come about by considering how to fill a knapsack of size j with the remaining objects {1, 2, . . . , i − 1}. This is just V [i − 1, j]. Take object i: If we take object i, then we gain a value of vi but have used up wi of our capacity. With the remaining j −wi capacity in the knapsack, we can fill it in the best possible way with objects {1, 2, . . . , i− 1}. This is vi + V [i − 1, j − wi ]. This is only possible if wi ≤ j. Since these are the only two possibilities, we can see that we have the following rule for constructing the array V . The ranges on i and j are i ∈ [0..n] and j ∈ [0..W ]. V [0, j] = 0 V [i − 1, j] if wi > j V [i, j] = max(V [i − 1, j], vi + V [i − 1, j − wi ]) if wi ≤ j The first line states that if there are no objects, then there is no value, irrespective of j. The second line implements the rule above. It is very easy to take these rules an produce an algorithm that computes the maximum value for the knapsack in time proportional to the size of the array, which is O((n + 1)(W + 1)) = O(nW ). The algorithm is given below. An example is shown in the figure below. The final output is V [n, W ] = V [4, 10] = 90. This reflects the selection of items 2 and 4, of values $40 and $50, respectively and weights 4 + 3 ≤ 10. The only missing detail is what items should we select to achieve the maximum. We will leave this as an exercise. They key is to record for each entry V [i, j] in the matrix whether we got this entry by taking the ith item or leaving it. With this information, it is possible to reconstruct the optimum knapsack contents. Lecture Notes 112 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-112-2048.jpg)
![0-1 Knapsack Problem KnapSack(v[1..n], w[1..n], n, W) { allocate V[0..n][0..W]; for j = 0 to W do V[0, j] = 0; // initialization for i = 1 to n do { for j = 0 to W do { leave_val = V[i-1, j]; // total value if we leave i if (j >= w[i]) // enough capacity to take i take_val = v[i] + V[i-1, j - w[i]]; // total value if we take i else take_val = -INFINITY; // cannot take i V[i,j] = max(leave_val, take_val); // final value is max } } return V[n, W]; } Values of the objects are 10, 40, 30, 50 . Weights of the objects are 5, 4, 6, 3 . Capacity → j=0 1 2 3 4 5 6 7 8 9 10 Item Value Weight 0 0 0 0 0 0 0 0 0 0 0 1 10 5 0 0 0 0 0 10 10 10 10 10 10 2 40 4 0 0 0 0 40 40 40 40 40 50 50 3 30 6 0 0 0 0 40 40 40 40 40 50 70 4 50 3 0 0 0 50 50 50 50 90 90 90 90 Final result is V [4, 10] = 90 (for taking items 2 and 4). Fig. 73: 0–1 Knapsack Example. Lecture Notes 113 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-113-2048.jpg)
![Supplemental Lecture 8: Dynamic Programming: Memoization Read: Section 15.3 of CLRS. Recursive Implementation: We have described dynamic programming as a method that involves the “bottom-up” computation of a table. However, the recursive formulations that we have derived have been set up in a “top- down” manner. Must the computation proceed bottom-up? Consider the following recursive implementation of the chain-matrix multiplication algorithm. The call Rec-Matrix-Chain(p, i, j) computes and returns the value of m[i, j]. The initial call is Rec-Matrix-Chain(p, 1, n). We only consider the cost here. Recursive Chain Matrix Multiplication Rec-Matrix-Chain(array p, int i, int j) { if (i == j) m[i,j] = 0; // basis case else { m[i,j] = INFINITY; // initialize for k = i to j-1 do { // try all splits cost = Rec-Matrix-Chain(p, i, k) + Rec-Matrix-Chain(p, k+1, j) + p[i-1]*p[k]*p[j]; if (cost < m[i,j]) m[i,j] = cost; // update if better } } return m[i,j]; // return final cost } (Note that the table m[1..n, 1..n] is not really needed. We show it just to make the connection with the earlier version clearer.) This version of the procedure certainly looks much simpler, and more closely resembles the recursive formulation that we gave previously for this problem. So, what is wrong with this? The answer is the running time is much higher than the Θ(n3 ) algorithm that we gave before. In fact, we will see that its running time is exponential in n. This is unacceptably slow. Let T (n) denote the running time of this algorithm on a sequence of matrices of length n. (That is, n = j −i+1.) If i = j then we have a sequence of length 1, and the time is Θ(1). Otherwise, we do Θ(1) work and then consider all possible ways of splitting the sequence of length n into two sequences, one of length k and the other of length n − k, and invoke the procedure recursively on each one. So we get the following recurrence, defined for n ≥ 1. (We have replaced the Θ(1)’s with the constant 1.) 1 if n = 1, T (n) = n−1 1+ k=1 (T (k) + T (n − k)) if n ≥ 2. Claim: T (n) ≥ 2n−1 . Proof: The proof is by induction on n. Clearly this is true for n = 1, since T (1) = 1 = 20 . In general, for n ≥ 2, the induction hypothesis is that T (m) ≥ 2m−1 for all m < n. Using this we have n−1 n−1 T (n) = 1+ (T (k) + T (n − k)) ≥ 1 + T (k) k=1 k=1 n−1 n−2 ≥ 1+ 2k−1 = 1 + 2k k=1 k=0 = 1 + (2 n−1 − 1) = 2 n−1 . In the first line we simply ignored the T (n−k) term, in the second line we applied the induction hypothesis, and in the last line we applied the formula for the geometric series. Lecture Notes 114 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-114-2048.jpg)
![Why is this so much worse than the dynamic programming version? If you “unravel” the recursive calls on a reasonably long example, you will see that the procedure is called repeatedly with the same arguments. The bottom-up version evaluates each entry exactly once. Memoization: Is it possible to retain the nice top-down structure of the recursive solution, while keeping the same O(n3 ) efficiency of the bottom-up version? The answer is yes, through a technique called memoization. Here is the idea. Let’s reconsider the function Rec-Matrix-Chain() given above. It’s job is to compute m[i, j], and return its value. As noted above, the main problem with the procedure is that it recomputes the same entries over and over. So, we will fix this by allowing the procedure to compute each entry exactly once. One way to do this is to initialize every entry to some special value (e.g. UNDEFINED). Once an entries value has been computed, it is never recomputed. Memoized Chain Matrix Multiplication Mem-Matrix-Chain(array p, int i, int j) { if (m[i,j] != UNDEFINED) return m[i,j]; // already defined else if (i == j) m[i,j] = 0; // basis case else { m[i,j] = INFINITY; // initialize for k = i to j-1 do { // try all splits cost = Mem-Matrix-Chain(p, i, k) + Mem-Matrix-Chain(p, k+1, j) + p[i-1]*p[k]*p[j]; if (cost < m[i,j]) m[i,j] = cost; // update if better } } return m[i,j]; // return final cost } This version runs in O(n3 ) time. Intuitively, this is because each of the O(n2 ) table entries is only computed once, and the work needed to compute one table entry (most of it in the for-loop) is at most O(n). Memoization is not usually used in practice, since it is generally slower than the bottom-up method. However, in some DP problems, many of the table entries are simply not needed, and so bottom-up computation may compute entries that are never needed. In these cases memoization may be a good idea. If you have know that most of the table will not be needed, here is a way to save space. Rather than storing the whole table explicitly as an array, you can store the “defined” entries of the table in a hash table, using the index pair (i, j) as the hash key. (See Chapter 11 in CLRS for more information on hashing.) Supplemental Lecture 9: Articulation Points and Biconnectivity Read: This material is not covered in CLR (except as Problem 23–2). Articulation Points and Biconnected Graphs: Today we discuss another application of DFS, this time to a problem on undirected graphs. Let G = (V, E) be a connected undirected graph. Consider the following definitions. Articulation Point (or Cut Vertex): Is any vertex whose removal (together with the removal of any incident edges) results in a disconnected graph. Bridge: Is an edge whose removal results in a disconnected graph. Biconnected: A graph is biconnected if it contains no articulation points. (In general a graph is k-connected, if k vertices must be removed to disconnect the graph.) Biconnected graphs and articulation points are of great interest in the design of network algorithms, because these are the “critical” points, whose failure will result in the network becoming disconnected. Lecture Notes 115 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-115-2048.jpg)
![v Low[u]=d[v] u Fig. 75: Articulation Points and DFS Observation 2: A leaf of the DFS tree is never an articulation point. Note that this is completely consistent with Observation 1, since a leaf will not have any subtrees in the DFS tree, so we can delete the word “internal” from Observation 1. What about the root? Since there are no cross edges between the subtrees of the root if the root has two or more children then it is an articulation point (since its removal separates these two subtrees). On the other hand, if the root has only a single child, then (as in the case of leaves) its removal does not disconnect the DFS tree, and hence cannot disconnect the graph in general. Observation 3: The root of the DFS is an articulation point if and only if it has two or more children. Articulation Points by DFS: Observations 1, 2, and 3 provide us with a structural characterization of which vertices in the DFS tree are articulation points. How can we design an algorithm which tests these conditions? Checking that the root has multiple children is an easy exercise. Checking Observation 1 is the hardest, but we will exploit the structure of the DFS tree to help us. The basic thing we need to check for is whether there is a back edge from some subtree to an ancestor of a given vertex. How can we do this? It would be too expensive to keep track of all the back edges from each subtree (because there may be Θ(e) back edges. A simpler scheme is to keep track of back edge that goes highest in the tree (in the sense of going closest to the root). If any back edge goes to an ancestor of u, this one will. How do we know how close a back edge goes to the root? As we travel from u towards the root, observe that the discovery times of these ancestors of u get smaller and smaller (the root having the smallest discovery time of 1). So we keep track of the back edge (v, w) that has the smallest value of d[w]. Low: Define Low[u] to be the minimum of d[u] and {d[w] | where (v, w) is a back edge and v is a descendent of u}. The term “descendent” is used in the nonstrict sense, that is, v may be equal to u. Intuitively, Low[u] is the highest (closest to the root) that you can get in the tree by taking any one backedge from either u or any of its descendents. (Beware of this notation: “Low” means low discovery time, not low in the tree. In fact Low[u] tends to be “high” in the tree, in the sense of being close to the root.) To compute Low[u] we use the following simple rules: Suppose that we are performing DFS on the vertex u. Initialization: Low[u] = d[u]. Back edge (u, v): Low[u] = min(Low[u], d[v]). Explanation: We have detected a new back edge coming out of u. If this goes to a lower d value than the previous back edge then make this the new low. Tree edge (u, v): Low[u] = min(Low[u], Low[v]). Explanation: Since v is in the subtree rooted at u any single back edge leaving the tree rooted at v is a single back edge for the tree rooted at u. Lecture Notes 117 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-117-2048.jpg)
![Observe that once Low[u] is computed for all vertices u, we can test whether a given nonroot vertex u is an articulation point by Observation 1 as follows: u is an articulation point if and only if it has a child v in the DFS tree for which Low[v] ≥ d[u] (since if there were a back edge from either v or one of its descendents to an ancestor of v then we would have Low[v] < d[u]). The Final Algorithm: There is one subtlety that we must watch for in designing the algorithm (in particular this is true for any DFS on undirected graphs). When processing a vertex u, we need to know when a given edge (u, v) is a back edge. How do we do this? An almost correct answer is to test whether v is colored gray (since all gray vertices are ancestors of the current vertex). This is not quite correct because v may be the parent of v in the DFS tree and we are just seeing the “other side” of the tree edge between v and u (recalling that in constructing the adjacency list of an undirected graph we create two directed edges for each undirected edge). To test correctly for a back edge we use the predecessor pointer to check that v is not the parent of u in the DFS tree. The complete algorithm for computing articulation points is given below. The main procedure for DFS is the same as before, except that it calls the following routine rather than DFSvisit(). Articulation Points ArtPt(u) { color[u] = gray Low[u] = d[u] = ++time for each (v in Adj(u)) { if (color[v] == white) { // (u,v) is a tree edge pred[v] = u ArtPt(v) Low[u] = min(Low[u], Low[v]) // update Low[u] if (pred[u] == NULL) { // root: apply Observation 3 if (this is u’s second child) Add u to set of articulation points } else if (Low[v] >= d[u]) { // internal node: apply Observation 1 Add u to set of articulation points } } else if (v != pred[u]) { // (u,v) is a back edge Low[u] = min(Low[u], d[v]) // update L[u] } } } An example is shown in the following figure. As with all DFS-based algorithms, the running time is Θ(n + e). There are some interesting problems that we still have not discussed. We did not discuss how to compute the bridges of a graph. This can be done by a small modification of the algorithm above. We’ll leave it as an exercise. (Notice that if {u, v} is a bridge then it does not follow that u and v are both articulation points.) Another question is how to determine which edges are in the biconnected components. A hint here is to store the edges in a stack as you go through the DFS search. When you come to an articulation point, you can show that all the edges in the biconnected component will be consecutive in the stack. Supplemental Lecture 10: Bellman-Ford Shortest Paths Read: Section 24.1 in CLRS. Bellman-Ford Algorithm: We saw that Dijkstra’s algorithm can solve the single-source shortest path problem, under the assumption that the edge weights are nonnegative. We also saw that shortest paths are undefined if you Lecture Notes 118 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-118-2048.jpg)
![d=1 Low=1 a e i a 2 1 8 8 b f j b e 3 1 9 8 c g c i 4 3 10 d h d 7 f 1 j 8 5 3 h = articulation pt. 6 g 3 Fig. 76: Articulation Points. have cycles of total negative cost. What if you have negative edge weights, but no negative cost cycles? We shall present the Bellman-Ford algorithm, which solves this problem. This algorithm is slower that Dijkstra’s algorithm, running in Θ(V E) time. In our version we will assume that there are no negative cost cycles. The one presented in CLRS actually contains a bit of code that checks for this. (Check it out.) Recall that we are given a graph G = (V, E) with numeric edge weights, w(u, v). Like Dijkstra’s algorithm, the Bellman-Ford algorithm is based on performing repeated relaxations. (Recall that relaxation updates shortest path information along a single edge. It was described in our discussion of Dijkstra’s algorithm.) Dijkstra’s algorithm was based on the idea of organizing the relaxations in the best possible manner, namely in increasing order of distance. Once relaxation is applied to an edge, it need never be relaxed again. This trick doesn’t seem to work when dealing with graphs with negative edge weights. Instead, the Bellman-Ford algorithm simply applies a relaxation to every edge in the graph, and repeats this V − 1 times. Bellman-Ford Algorithm BellmanFord(G,w,s) { for each (u in V) { // standard initialization d[u] = +infinity pred[u] = null } d[s] = 0 for i = 1 to V-1 { // repeat V-1 times for each (u,v) in E { // relax along each edge Relax(u,v) } } } The Θ(V E) running time is pretty obvious, since there are two main nested loops, one iterated V − 1 times and the other iterated E times. The interesting question is how and why it works. Correctness of Bellman-Ford: I like to think of the Bellman-Ford as a sort of “BubbleSort analogue” for shortest paths, in the sense that shortest path information is propagated sequentially along each shortest path in the graph. Consider any shortest path from s to some other vertex u: v0 , v1 , . . . , vk where v0 = s and vk = u. Since a shortest path will never visit the same vertex twice, we know that k ≤ V − 1, and hence the path consists of at most V − 1 edges. Since this is a shortest path we have δ(s, vi ) (the true shortest path cost from s to vi ) satisfies δ(s, vi ) = δ(s, vi−1 ) + w(vi−1 , vi ). Lecture Notes 119 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-119-2048.jpg)
![? 8 8 8 8 8 8 8 0 −6 ? 0 −6 ? 0 −6 9 0 −6 7 4 5 4 5 4 5 4 5 ? 4 2 2 Initial configuration After 1st relaxation After 2nd relaxation After 3rd relaxation phase phase phase Fig. 77: Bellman-Ford Algorithm. We assert that after the ith pass of the “for-i” loop that d[vi ] = δ(s, vi ). The proof is by induction on i. Observe that after the initialization (pass 0) we have d[v1 ] = d[s] = 0. In general, prior to the ith pass through the loop, the induction hypothesis tells us that d[vi−1 ] = δ(s, vi−1 ). After the ith pass through the loop, we have done a relaxation on the edge (vi−1 , vi ) (since we do relaxations along all the edges). Thus after the ith pass we have d[vi ] ≤ d[vi−1 ] + w(vi−1 , vi ) = δ(s, vi−1 ) + w(vi−1 , vi ) = δ(s, vi ). Recall from Dijkstra’s algorithm that d[vi ] is never less than δ(s, vi ) (since each time we do a relaxation there exists a path that witnesses its value). Thus, d[vi ] is in fact equal to δ(s, vi ), completing the induction proof. In summary, after i passes through the for loop, all vertices that are i edges away (along the shortest path tree) from the source have the correct distance values stored in d[u]. Thus, after the (V − 1)st iteration of the for loop, all vertices u have the correct distance values stored in d[u]. Supplemental Lecture 11: Network Flows and Matching Read: Chapt 27 in CLR. Maximum Flow: The Max Flow problem is one of the basic problems of algorithm design. Intuitively we can think of a flow network as a directed graph in which fluid is flowing along the edges of the graph. Each edge has certain maximum capacity that it can carry. The idea is to find out how much flow we can push from one point to another. The max flow problem has applications in areas like transportation, routing in networks. It is the simplest problem in a line of many important problems having to do with the movement of commodities through a network. These are often studied in business schools, and operations research. Flow Networks: A flow network G = (V, E) is a directed graph in which each edge (u, v) ∈ E has a nonegative capacity c(u, v) ≥ 0. If (u, v) ∈ E we model this by setting c(u, v) = 0. There are two special vertices: a source s, and a sink t. We assume that every vertex lies on some path from the source to the sink (for otherwise the vertex is of no use to us). (This implies that the digraph is connected, and hence e ≥ n − 1.) A flow is a real valued function on pairs of vertices, f : V × V → R which satisfies the following three properties: Capacity Constraint: For all u, v ∈ V , f (u, v) ≤ c(u, v). Skew Symmetry: For all u, v ∈ V , f (u, v) = −f (v, u). (In other words, we can think of backwards flow as negative flow. This is primarily for making algebraic analysis easier.) Flow conservation: For all u ∈ V − {s, t}, we have f (u, v) = 0. v∈V Lecture Notes 120 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-120-2048.jpg)
![1 + 4), 6, 7(= 1 + 6), 10(= 4 + 6), 11(= 1 + 4 + 6) . Note that Li can have as many as 2i elements, but may have fewer, since some subsets may have the same sum. There are two things we will want to do for efficiency. (1) Remove any duplicates from Li , and (2) only keep sums that are less than or equal to t. Let us suppose that we a procedure MergeLists(L1, L2) which merges two sorted lists, and returns a sorted lists with all duplicates removed. This is essentially the procedure used in MergeSort but with the added duplicate element test. As a bit of notation, let L + x denote the list resulting by adding the number x to every element of list L. Thus 1, 4, 6 + 3 = 4, 7, 9 . This gives the following procedure for the subset sum problem. Exact Subset Sum Exact_SS(x[1..n], t) { L = <0>; for i = 1 to n do { L = MergeLists(L, L+x[i]); remove for L all elements greater than t; } return largest element in L; } For example, if S = {1, 4, 6} and t = 8 then the successive lists would be L0 = 0 L1 = 0 ∪ 0 + 1 = 0, 1 L2 = 0, 1 ∪ 0 + 4, 1 + 4 = 0, 1, 4, 5 L3 = 0, 1, 4, 5 ∪ 0 + 6, 1 + 6, 4 + 6, 5 + 6 = 0, 1, 4, 5, 6, 7, 10, 11 . The last list would have the elements 10 and 11 removed, and the final answer would be 7. The algorithm runs in Ω(2n ) time in the worst case, because this is the number of sums that are generated if there are no duplicates, and no items are removed. Approximation Algorithm: To convert this into an approximation algorithm, we will introduce a “trim” the lists to decrease their sizes. The idea is that if the list L contains two numbers that are very close to one another, e.g. 91, 048 and 91, 050, then we should not need to keep both of these numbers in the list. One of them is good enough for future approximations. This will reduce the size of the lists that the algorithm needs to maintain. But, how much trimming can we allow and still keep our approximation bound? Furthermore, will we be able to reduce the list sizes from exponential to polynomial? The answer to both these questions is yes, provided you apply a proper way of trimming the lists. We will trim elements whose values are sufficiently close to each other. But we should define close in manner that is relative to the sizes of the numbers involved. The trimming must also depend on . We select δ = /n. (Why? We will see later that this is the value that makes everything work out in the end.) Note that 0 < δ < 1. Assume that the elements of L are sorted. We walk through the list. Let z denote the last untrimmed element in L, and let y ≥ z be the next element to be considered. If y−z ≤δ y then we trim y from the list. Equivalently, this means that the final trimmed list cannot contain two value y and z such that (1 − δ)y ≤ z ≤ y. We can think of z as representing y in the list. For example, given δ = 0.1 and given the list L = 10, 11, 12, 15, 20, 21, 22, 23, 24, 29 , Lecture Notes 131 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-131-2048.jpg)
![the trimmed list L will consist of L = 10, 12, 15, 20, 23, 29 . Another way to visualize trimming is to break the interval from [1, t] into a set of buckets of exponentially increasing size. Let d = 1/(1 − δ). Note that d > 1. Consider the intervals [1, d], [d, d2 ], [d2 , d3 ], . . . , [dk−1 , dk ] where dk ≥ t. If z ≤ y are in the same interval [di−1 , di ] then y−z di − di−1 1 ≤ i = 1 − = δ. y d d Thus, we cannot have more than one item within each bucket. We can think of trimming as a way of enforcing the condition that items in our lists are not relatively too close to one another, by enforcing the condition that no bucket has more than one item. 1 2 4 8 16 L L’ Fig. 82: Trimming Lists for Approximate Subset Sum. Claim: The number of distinct items in a trimmed list is O((n log t)/ ), which is polynomial in input size and 1/ . Proof: We know that each pair of consecutive elements in a trimmed list differ by a ratio of at least d = 1/(1 − δ) > 1. Let k denote the number of elements in the trimmed list, ignoring the element of value 0. Thus, the smallest nonzero value and maximum value in the the trimmed list differ by a ratio of at least dk−1 . Since the smallest (nonzero) element is at least as large as 1, and the largest is no larger than t, then it follows that dk−1 ≤ t/1 = t. Taking the natural log of both sides we have (k − 1) ln d ≤ ln t. Using the facts that δ = /n and the log identity that ln(1 + x) ≤ x, we have ln t ln t k−1 ≤ = ln d − ln(1 − δ) ln t n ln t ≤ = δ n log t k = O . Observe that the input size is at least as large as n (since there are n numbers) and at least as large as log t (since it takes log t digits to write down t on the input). Thus, this function is polynomial in the input size and 1/ . The approximation algorithm operates as before, but in addition we call the procedure Trim given below. For example, consider the set S = {104, 102, 201, 101} and t = 308 and = 0.20. We have δ = /4 = 0.05. Here is a summary of the algorithm’s execution. Lecture Notes 132 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-132-2048.jpg)
![Approximate Subset Sum Trim(L, delta) { let the elements of L be denoted y[1..m]; L’ = <y[1]>; // start with first item last = y[1]; // last item to be added for i = 2 to m do { if (last < (1-delta) y[i]) { // different enough? append y[i] to end of L’; last = y[i]; } } } Approx_SS(x[1..n], t, eps) { delta = eps/n; // approx factor L = <0>; // empty sum = 0 for i = 1 to n do { L = MergeLists(L, L+x[i]); // add in next item L = Trim(L, delta); // trim away "near" duplicates remove for L all elements greater than t; } return largest element in L; } init: L0 = 0 merge: L1 = 0, 104 trim: L1 = 0, 104 remove: L1 = 0, 104 merge: L2 = 0, 102, 104, 206 trim: L2 = 0, 102, 206 remove: L2 = 0, 102, 206 merge: L3 = 0, 102, 201, 206, 303, 407 trim: L3 = 0, 102, 201, 303, 407 remove: L3 = 0, 102, 201, 303 merge: L4 = 0, 101, 102, 201, 203, 302, 303, 404 trim: L4 = 0, 101, 201, 302, 404 remove: L4 = 0, 101, 201, 302 The final output is 302. The optimum is 307 = 104 + 102 + 101. So our actual relative error in this case is within 2%. The running time of the procedure is O(n|L|) which is O(n2 ln t/ ) by the earlier claim. Lecture Notes 133 CMSC 451](https://image.slidesharecdn.com/designandanalysisofcomputeralgorithms-130328235508-phpapp02/75/Design-and-analysis-of-computer-algorithms-133-2048.jpg)
Uploaded by Krishna Chaytaniah
Design and analysis of computer algorithms
This document provides an overview of a course on the design and analysis of computer algorithms taught by Professor David Mount at the University of Maryland in Fall 2003. The course will cover algorithm design techniques like dynamic programming and greedy algorithms. Major topics will include graph algorithms, minimum spanning trees, shortest paths, and computational geometry. Later sections will discuss intractable problems and approximation algorithms. When designing algorithms, students are expected to provide a description, proof of correctness, and analysis of time and space efficiency. Mathematical background on algorithm analysis, including asymptotic notation and recurrences, will be reviewed.
Design and analysis of computer algorithms
- 1. CMSC 451 Design and Analysis of Computer Algorithms1 David M. Mount Department of Computer Science University of Maryland Fall 2003 1 Copyright, David M. Mount, 2004, Dept. of Computer Science, University of Maryland, College Park, MD, 20742. These lecture notes were prepared by David Mount for the course CMSC 451, Design and Analysis of Computer Algorithms, at the University of Maryland. Permission to use, copy, modify, and distribute these notes for educational purposes and without fee is hereby granted, provided that this copyright notice appear in all copies. Lecture Notes 1 CMSC 451
- 2. Lecture 1: CourseIntroduction Read: (All readings are from Cormen, Leiserson, Rivest and Stein, Introduction to Algorithms, 2nd Edition). Review Chapts. 1–5 in CLRS. What is an algorithm? Our text defines an algorithm to be any well-defined computational procedure that takes some values as input and produces some values as output. Like a cooking recipe, an algorithm provides a step-by-step method for solving a computational problem. Unlike programs, algorithms are not dependent on a particular programming language, machine, system, or compiler. They are mathematical entities, which can be thought of as running on some sort of idealized computer with an infinite random access memory and an unlimited word size. Algorithm design is all about the mathematical theory behind the design of good programs. Why study algorithm design? Programming is a very complex task, and there are a number of aspects of program- ming that make it so complex. The first is that most programming projects are very large, requiring the coor- dinated efforts of many people. (This is the topic a course like software engineering.) The next is that many programming projects involve storing and accessing large quantities of data efficiently. (This is the topic of courses on data structures and databases.) The last is that many programming projects involve solving complex computational problems, for which simplistic or naive solutions may not be efficient enough. The complex problems may involve numerical data (the subject of courses on numerical analysis), but often they involve discrete data. This is where the topic of algorithm design and analysis is important. Although the algorithms discussed in this course will often represent only a tiny fraction of the code that is generated in a large software system, this small fraction may be very important for the success of the overall project. An unfortunately common approach to this problem is to first design an inefficient algorithm and data structure to solve the problem, and then take this poor design and attempt to fine-tune its performance. The problem is that if the underlying design is bad, then often no amount of fine-tuning is going to make a substantial difference. The focus of this course is on how to design good algorithms, and how to analyze their efficiency. This is among the most basic aspects of good programming. Course Overview: This course will consist of a number of major sections. The first will be a short review of some preliminary material, including asymptotics, summations, and recurrences and sorting. These have been covered in earlier courses, and so we will breeze through them pretty quickly. We will then discuss approaches to designing optimization algorithms, including dynamic programming and greedy algorithms. The next major focus will be on graph algorithms. This will include a review of breadth-first and depth-first search and their application in various problems related to connectivity in graphs. Next we will discuss minimum spanning trees, shortest paths, and network flows. We will briefly discuss algorithmic problems arising from geometric settings, that is, computational geometry. Most of the emphasis of the first portion of the course will be on problems that can be solved efficiently, in the latter portion we will discuss intractability and NP-hard problems. These are problems for which no efficient solution is known. Finally, we will discuss methods to approximate NP-hard problems, and how to prove how close these approximations are to the optimal solutions. Issues in Algorithm Design: Algorithms are mathematical objects (in contrast to the must more concrete notion of a computer program implemented in some programming language and executing on some machine). As such, we can reason about the properties of algorithms mathematically. When designing an algorithm there are two fundamental issues to be considered: correctness and efficiency. It is important to justify an algorithm’s correctness mathematically. For very complex algorithms, this typically requires a careful mathematical proof, which may require the proof of many lemmas and properties of the solution, upon which the algorithm relies. For simple algorithms (BubbleSort, for example) a short intuitive explanation of the algorithm’s basic invariants is sufficient. (For example, in BubbleSort, the principal invariant is that on completion of the ith iteration, the last i elements are in their proper sorted positions.) Lecture Notes 2 CMSC 451
- 3. Establishing efficiency isa much more complex endeavor. Intuitively, an algorithm’s efficiency is a function of the amount of computational resources it requires, measured typically as execution time and the amount of space, or memory, that the algorithm uses. The amount of computational resources can be a complex function of the size and structure of the input set. In order to reduce matters to their simplest form, it is common to consider efficiency as a function of input size. Among all inputs of the same size, we consider the maximum possible running time. This is called worst-case analysis. It is also possible, and often more meaningful, to measure average-case analysis. Average-case analyses tend to be more complex, and may require that some probability distribution be defined on the set of inputs. To keep matters simple, we will usually focus on worst-case analysis in this course. Throughout out this course, when you are asked to present an algorithm, this means that you need to do three things: • Present a clear, simple and unambiguous description of the algorithm (in pseudo-code, for example). They key here is “keep it simple.” Uninteresting details should be kept to a minimum, so that the key compu- tational issues stand out. (For example, it is not necessary to declare variables whose purpose is obvious, and it is often simpler and clearer to simply say, “Add X to the end of list L” than to present code to do this or use some arcane syntax, such as “L.insertAtEnd(X).”) • Present a justification or proof of the algorithm’s correctness. Your justification should assume that the reader is someone of similar background as yourself, say another student in this class, and should be con- vincing enough make a skeptic believe that your algorithm does indeed solve the problem correctly. Avoid rambling about obvious or trivial elements. A good proof provides an overview of what the algorithm does, and then focuses on any tricky elements that may not be obvious. • Present a worst-case analysis of the algorithms efficiency, typically it running time (but also its space, if space is an issue). Sometimes this is straightforward, but if not, concentrate on the parts of the analysis that are not obvious. Note that the presentation does not need to be in this order. Often it is good to begin with an explanation of how you derived the algorithm, emphasizing particular elements of the design that establish its correctness and efficiency. Then, once this groundwork has been laid down, present the algorithm itself. If this seems to be a bit abstract now, don’t worry. We will see many examples of this process throughout the semester. Lecture 2: Mathematical Background Read: Review Chapters 1–5 in CLRS. Algorithm Analysis: Today we will review some of the basic elements of algorithm analysis, which were covered in previous courses. These include asymptotics, summations, and recurrences. Asymptotics: Asymptotics involves O-notation (“big-Oh”) and its many relatives, Ω, Θ, o (“little-Oh”), ω. Asymp- totic notation provides us with a way to simplify the functions that arise in analyzing algorithm running times by ignoring constant factors and concentrating on the trends for large values of n. For example, it allows us to reason that for three algorithms with the respective running times n3 log n + 4n2 + 52n log n ∈ Θ(n3 log n) 15n2 + 7n log3 n ∈ Θ(n2 ) 3n + 4 log5 n + 19n2 ∈ Θ(n2 ). Thus, the first algorithm is significantly slower for large n, while the other two are comparable, up to a constant factor. Since asymptotics were covered in earlier courses, I will assume that this is familiar to you. Nonetheless, here are a few facts to remember about asymptotic notation: Lecture Notes 3 CMSC 451
- 4. Ignore constant factors:Multiplicative constant factors are ignored. For example, 347n is Θ(n). Constant factors appearing exponents cannot be ignored. For example, 23n is not O(2n ). Focus on large n: Asymptotic analysis means that we consider trends for large values of n. Thus, the fastest growing function of n is the only one that needs to be considered. For example, 3n2 log n + 25n log n + (log n)7 is Θ(n2 log n). Polylog, polynomial, and exponential: These are the most common functions that arise in analyzing algo- rithms: Polylogarithmic: Powers of log n, such as (log n)7 . We will usually write this as log7 n. √ Polynomial: Powers of n, such as n4 and n = n1/2 . Exponential: A constant (not 1) raised to the power n, such as 3n . An important fact is that polylogarithmic functions are strictly asymptotically smaller than polynomial function, which are strictly asymptotically smaller than exponential functions (assuming the base of the exponent is bigger than 1). For example, if we let mean “asymptotically smaller” then loga n nb cn for any a, b, and c, provided that b > 0 and c > 1. Logarithm Simplification: It is a good idea to first simplify terms involving logarithms. For example, the following formulas are useful. Here a, b, c are constants: loga n logb n = = Θ(loga n) loga b loga (nc ) = c loga n = Θ(loga n) bloga n = nloga b . Avoid using log n in exponents. The last rule above can be used to achieve this. For example, rather than saying 3log2 n , express this as nlog2 3 ≈ n1.585 . Following the conventional sloppiness, I will often say O(n2 ), when in fact the stronger statement Θ(n2 ) holds. (This is just because it is easier to say “oh” than “theta”.) Summations: Summations naturally arise in the analysis of iterative algorithms. Also, more complex forms of analy- sis, such as recurrences, are often solved by reducing them to summations. Solving a summation means reducing it to a closed form formula, that is, one having no summations, recurrences, integrals, or other complex operators. In algorithm design it is often not necessary to solve a summation exactly, since an asymptotic approximation or close upper bound is usually good enough. Here are some common summations and some tips to use in solving summations. Constant Series: For integers a and b, b 1 = max(b − a + 1, 0). i=a Notice that when b = a − 1, there are no terms in the summation (since the index is assumed to count upwards only), and the result is 0. Be careful to check that b ≥ a − 1 before applying this formula blindly. Arithmetic Series: For n ≥ 0, n n(n + 1) i = 1 + 2 + ··· + n = . i=0 2 2 This is Θ(n ). (The starting bound could have just as easily been set to 1 as 0.) Lecture Notes 4 CMSC 451
- 5. Geometric Series: Letx = 1 be any constant (independent of n), then for n ≥ 0, n xn+1 − 1 xi = 1 + x + x2 + · · · + xn = . i=0 x−1 If 0 < x < 1 then this is Θ(1). If x > 1, then this is Θ(xn ), that is, the entire sum is proportional to the last element of the series. Quadratic Series: For n ≥ 0, n 2n3 + 3n2 + n i2 = 12 + 22 + · · · + n2 = . i=0 6 Linear-geometric Series: This arises in some algorithms based on trees and recursion. Let x = 1 be any constant, then for n ≥ 0, n−1 (n − 1)x(n+1) − nxn + x ixi = x + 2x2 + 3x3 · · · + nxn = . i=0 (x − 1)2 As n becomes large, this is asymptotically dominated by the term (n − 1)x(n+1) /(x − 1)2 . The multi- plicative term n − 1 is very nearly equal to n for large n, and, since x is a constant, we may multiply this times the constant (x − 1)2 /x without changing the asymptotics. What remains is Θ(nxn ). Harmonic Series: This arises often in probabilistic analyses of algorithms. It does not have an exact closed form solution, but it can be closely approximated. For n ≥ 0, n 1 1 1 1 Hn = = 1 + + + ··· + = (ln n) + O(1). i=1 i 2 3 n There are also a few tips to learn about solving summations. Summations with general bounds: When a summation does not start at the 1 or 0, as most of the above for- mulas assume, you can just split it up into the difference of two summations. For example, for 1 ≤ a ≤ b b b a−1 f (i) = f (i) − f (i). i=a i=0 i=0 Linearity of Summation: Constant factors and added terms can be split out to make summations simpler. (4 + 3i(i − 2)) = 4 + 3i2 − 6i = 4+3 i2 − 6 i. Now the formulas can be to each summation individually. Approximate using integrals: Integration and summation are closely related. (Integration is in some sense a continuous form of summation.) Here is a handy formula. Let f (x) be any monotonically increasing function (the function increases as x increases). n n n+1 f (x)dx ≤ f (i) ≤ f (x)dx. 0 i=1 1 Example: Right Dominant Elements As an example of the use of summations in algorithm analysis, consider the following simple problem. We are given a list L of numeric values. We say that an element of L is right dominant if it is strictly larger than all the elements that follow it in the list. Note that the last element of the list Lecture Notes 5 CMSC 451
- 6. is always rightdominant, as is the last occurrence of the maximum element of the array. For example, consider the following list. L = 10, 9, 5, 13, 2, 7, 1, 8, 4, 6, 3 The sequence of right dominant elements are 13, 8, 6, 3 . In order to make this more concrete, we should think about how L is represented. It will make a difference whether L is represented as an array (allowing for random access), a doubly linked list (allowing for sequential access in both directions), or a singly linked list (allowing for sequential access in only one direction). Among the three possible representations, the array representation seems to yield the simplest and clearest algorithm. However, we will design the algorithm in such a way that it only performs sequential scans, so it could also be implemented using a singly linked or doubly linked list. (This is common in algorithms. Chose your rep- resentation to make the algorithm as simple and clear as possible, but give thought to how it may actually be implemented. Remember that algorithms are read by humans, not compilers.) We will assume here that the array L of size n is indexed from 1 to n. Think for a moment how you would solve this problem. Can you see an O(n) time algorithm? (If not, think a little harder.) To illustrate summations, we will first present a naive O(n2 ) time algorithm, which operates by simply checking for each element of the array whether all the subsequent elements are strictly smaller. (Although this example is pretty stupid, it will also serve to illustrate the sort of style that we will use in presenting algorithms.) Right Dominant Elements (Naive Solution) // Input: List L of numbers given as an array L[1..n] // Returns: List D containing the right dominant elements of L RightDominant(L) { D = empty list for (i = 1 to n) isDominant = true for (j = i+1 to n) if (A[i] <= A[j]) isDominant = false if (isDominant) append A[i] to D } return D } If I were programming this, I would rewrite the inner (j) loop as a while loop, since we can terminate the loop as soon as we find that A[i] is not dominant. Again, this sort of optimization is good to keep in mind in programming, but will be omitted since it will not affect the worst-case running time. The time spent in this algorithm is dominated (no pun intended) by the time spent in the inner (j) loop. On the ith iteration of the outer loop, the inner loop is executed from i + 1 to n, for a total of n − (i + 1) + 1 = n − i times. (Recall the rule for the constant series above.) Each iteration of the inner loop takes constant time. Thus, up to a constant factor, the running time, as a function of n, is given by the following summation: n T (n) = (n − i). i=1 To solve this summation, let us expand it, and put it into a form such that the above formulas can be used. T (n) = (n − 1) + (n − 2) + . . . + 2 + 1 + 0 = 0 + 1 + 2 + . . . + (n − 2) + (n − 1) n−1 (n − 1)n = i = . i=0 2 Lecture Notes 6 CMSC 451
- 7. The last stepcomes from applying the formula for the linear series (using n − 1 in place of n in the formula). As mentioned above, there is a simple O(n) time algorithm for this problem. As an exercise, see if you can find it. As an additional challenge, see if you can design your algorithm so it only performs a single left-to-right scan of the list L. (You are allowed to use up to O(n) working storage to do this.) Recurrences: Another useful mathematical tool in algorithm analysis will be recurrences. They arise naturally in the analysis of divide-and-conquer algorithms. Recall that these algorithms have the following general structure. Divide: Divide the problem into two or more subproblems (ideally of roughly equal sizes), Conquer: Solve each subproblem recursively, and Combine: Combine the solutions to the subproblems into a single global solution. How do we analyze recursive procedures like this one? If there is a simple pattern to the sizes of the recursive calls, then the best way is usually by setting up a recurrence, that is, a function which is defined recursively in terms of itself. Here is a typical example. Suppose that we break the problem into two subproblems, each of size roughly n/2. (We will assume exactly n/2 for simplicity.). The additional overhead of splitting and merging the solutions is O(n). When the subproblems are reduced to size 1, we can solve them in O(1) time. We will ignore constant factors, writing O(n) just as n, yielding the following recurrence: T (n) = 1 if n = 1, T (n) = 2T (n/2) + n if n > 1. Note that, since we assume that n is an integer, this recurrence is not well defined unless n is a power of 2 (since otherwise n/2 will at some point be a fraction). To be formally correct, I should either write n/2 or restrict the domain of n, but I will often be sloppy in this way. There are a number of methods for solving the sort of recurrences that show up in divide-and-conquer algo- rithms. The easiest method is to apply the Master Theorem, given in CLRS. Here is a slightly more restrictive version, but adequate for a lot of instances. See CLRS for the more complete version of the Master Theorem and its proof. Theorem: (Simplified Master Theorem) Let a ≥ 1, b > 1 be constants and let T (n) be the recurrence T (n) = aT (n/b) + cnk , defined for n ≥ 0. Case 1: a > bk then T (n) is Θ(nlogb a ). Case 2: a = bk then T (n) is Θ(nk log n). Case 3: a < bk then T (n) is Θ(nk ). Using this version of the Master Theorem we can see that in our recurrence a = 2, b = 2, and k = 1, so a = bk and Case 2 applies. Thus T (n) is Θ(n log n). There many recurrences that cannot be put into this form. For example, the following recurrence is quite common: T (n) = 2T (n/2) + n log n. This solves to T (n) = Θ(n log2 n), but the Master Theorem (either this form or the one in CLRS will not tell you this.) For such recurrences, other methods are needed. Lecture 3: Review of Sorting and Selection Read: Review Chapts. 6–9 in CLRS. Lecture Notes 7 CMSC 451
- 8. Review of Sorting:Sorting is among the most basic problems in algorithm design. We are given a sequence of items, each associated with a given key value. The problem is to permute the items so that they are in increasing (or decreasing) order by key. Sorting is important because it is often the first step in more complex algorithms. Sorting algorithms are usually divided into two classes, internal sorting algorithms, which assume that data is stored in an array in main memory, and external sorting algorithm, which assume that data is stored on disk or some other device that is best accessed sequentially. We will only consider internal sorting. You are probably familiar with one or more of the standard simple Θ(n2 ) sorting algorithms, such as Insertion- Sort, SelectionSort and BubbleSort. (By the way, these algorithms are quite acceptable for small lists of, say, fewer than 20 elements.) BubbleSort is the easiest one to remember, but it widely considered to be the worst of the three. The three canonical efficient comparison-based sorting algorithms are MergeSort, QuickSort, and HeapSort. All run in Θ(n log n) time. Sorting algorithms often have additional properties that are of interest, depending on the application. Here are two important properties. In-place: The algorithm uses no additional array storage, and hence (other than perhaps the system’s recursion stack) it is possible to sort very large lists without the need to allocate additional working storage. Stable: A sorting algorithm is stable if two elements that are equal remain in the same relative position after sorting is completed. This is of interest, since in some sorting applications you sort first on one key and then on another. It is nice to know that two items that are equal on the second key, remain sorted on the first key. Here is a quick summary of the fast sorting algorithms. If you are not familiar with any of these, check out the descriptions in CLRS. They are shown schematically in Fig. 1 QuickSort: It works recursively, by first selecting a random “pivot value” from the array. Then it partitions the array into elements that are less than and greater than the pivot. Then it recursively sorts each part. QuickSort is widely regarded as the fastest of the fast sorting algorithms (on modern machines). One explanation is that its inner loop compares elements against a single pivot value, which can be stored in a register for fast access. The other algorithms compare two elements in the array. This is considered an in-place sorting algorithm, since it uses no other array storage. (It does implicitly use the system’s recursion stack, but this is usually not counted.) It is not stable. There is a stable version of QuickSort, but it is not in-place. This algorithm is Θ(n log n) in the expected case, and Θ(n2 ) in the worst case. If properly implemented, the probability that the algorithm takes asymptotically longer (assuming that the pivot is chosen randomly) is extremely small for large n. x QuickSort: x partition <x x >x sort sort split merge MergeSort: sort HeapSort: buildHeap extractMax Heap Fig. 1: Common O(n log n) comparison-based sorting algorithms. Lecture Notes 8 CMSC 451
- 9. MergeSort: MergeSort alsoworks recursively. It is a classical divide-and-conquer algorithm. The array is split into two subarrays of roughly equal size. They are sorted recursively. Then the two sorted subarrays are merged together in Θ(n) time. MergeSort is the only stable sorting algorithm of these three. The downside is the MergeSort is the only algorithm of the three that requires additional array storage (ignoring the recursion stack), and thus it is not in-place. This is because the merging process merges the two arrays into a third array. Although it is possible to merge arrays in-place, it cannot be done in Θ(n) time. HeapSort: HeapSort is based on a nice data structure, called a heap, which is an efficient implementation of a priority queue data structure. A priority queue supports the operations of inserting a key, and deleting the element with the smallest key value. A heap can be built for n keys in Θ(n) time, and the minimum key can be extracted in Θ(log n) time. HeapSort is an in-place sorting algorithm, but it is not stable. HeapSort works by building the heap (ordered in reverse order so that the maximum can be extracted efficiently) and then repeatedly extracting the largest element. (Why it extracts the maximum rather than the minimum is an implementation detail, but this is the key to making this work as an in-place sorting algorithm.) If you only want to extract the k smallest values, a heap can allow you to do this is Θ(n + k log n) time. A heap has the additional advantage of being used in contexts where the priority of elements changes. Each change of priority (key value) can be processed in Θ(log n) time. Which sorting algorithm should you implement when implementing your programs? The correct answer is probably “none of them”. Unless you know that your input has some special properties that suggest a much faster alternative, it is best to rely on the library sorting procedure supplied on your system. Presumably, it has been engineered to produce the best performance for your system, and saves you from debugging time. Nonetheless, it is important to learn about sorting algorithms, since the fundamental concepts covered there apply to much more complex algorithms. Selection: A simpler, related problem to sorting is selection. The selection problem is, given an array A of n numbers (not sorted), and an integer k, where 1 ≤ k ≤ n, return the kth smallest value of A. Although selection can be solved in O(n log n) time, by first sorting A and then returning the kth element of the sorted list, it is possible to select the kth smallest element in O(n) time. The algorithm is a variant of QuickSort. Lower Bounds for Comparison-Based Sorting: The fact that O(n log n) sorting algorithms are the fastest around for many years, suggests that this may be the best that we can do. Can we sort faster? The claim is no, pro- vided that the algorithm is comparison-based. A comparison-based sorting algorithm is one in which algorithm permutes the elements based solely on the results of the comparisons that the algorithm makes between pairs of elements. All of the algorithms we have discussed so far are comparison-based. We will see that exceptions exist in special cases. This does not preclude the possibility of sorting algorithms whose actions are determined by other operations, as we shall see below. The following theorem gives the lower bound on comparison-based sorting. Theorem: Any comparison-based sorting algorithm has worst-case running time Ω(n log n). We will not present a proof of this theorem, but the basic argument follows from a simple analysis of the number of possibilities and the time it takes to distinguish among them. There are n! ways to permute a given set of n numbers. Any sorting algorithm must be able to distinguish between each of these different possibilities, since two different permutations need to treated differently. Since each comparison leads to only two possible outcomes, the execution of the algorithm can be viewed as a binary tree. (This is a bit abstract, but given a sorting algorithm it is not hard, but quite tedious, to trace its execution, and set up a new node each time a decision is made.) This binary tree, called a decision tree, must have at least n! leaves, one for each of the possible input permutations. Such a tree, even if perfectly balanced, must height at least lg(n!). By Stirling’s approximation, n! Lecture Notes 9 CMSC 451
- 10. is, up toconstant factors, roughly (n/e)n . Plugging this in and simplifying yields the Ω(n log n) lower bound. This can also be generalized to show that the average-case time to sort is also Ω(n log n). Linear Time Sorting: The Ω(n log n) lower bound implies that if we hope to sort numbers faster than in O(n log n) time, we cannot do it by making comparisons alone. In some special cases, it is possible to sort without the use of comparisons. This leads to the possibility of sorting in linear (that is, O(n)) time. Here are three such algorithms. Counting Sort: Counting sort assumes that each input is an integer in the range from 1 to k. The algorithm sorts in Θ(n + k) time. Thus, if k is O(n), this implies that the resulting sorting algorithm runs in Θ(n) time. The algorithm requires an additional Θ(n + k) working storage but has the nice feature that it is stable. The algorithm is remarkably simple, but deceptively clever. You are referred to CLRS for the details. Radix Sort: The main shortcoming of CountingSort is that (due to space requirements) it is only practical for a very small ranges of integers. If the integers are in the range from say, 1 to a million, we may not want to allocate an array of a million elements. RadixSort provides a nice way around this by sorting numbers one digit, or one byte, or generally, some groups of bits, at a time. As the number of bits in each group increases, the algorithm is faster, but the space requirements go up. The idea is very simple. Let’s think of our list as being composed of n integers, each having d decimal digits (or digits in any base). To sort these integers we simply sort repeatedly, starting at the lowest order digit, and finishing with the highest order digit. Since the sorting algorithm is stable, we know that if the numbers are already sorted with respect to low order digits, and then later we sort with respect to high order digits, numbers having the same high order digit will remain sorted with respect to their low order digit. An example is shown in Figure 2. Input Output 576 49[4] 9[5]4 [1]76 176 494 19[4] 5[7]6 [1]94 194 194 95[4] 1[7]6 [2]78 278 296 =⇒ 57[6] =⇒ 2[7]8 =⇒ [2]96 =⇒ 296 278 29[6] 4[9]4 [4]94 494 176 17[6] 1[9]4 [5]76 576 954 27[8] 2[9]6 [9]54 954 Fig. 2: Example of RadixSort. The running time is Θ(d(n + k)) where d is the number of digits in each value, n is the length of the list, and k is the number of distinct values each digit may have. The space needed is Θ(n + k). A common application of this algorithm is for sorting integers over some range that is larger than n, but still polynomial in n. For example, suppose that you wanted to sort a list of integers in the range from 1 to n2 . First, you could subtract 1 so that they are now in the range from 0 to n2 − 1. Observe that any number in this range can be expressed as 2-digit number, where each digit is over the range from 0 to n − 1. In particular, given any integer L in this range, we can write L = an + b, where a = L/n and b = L mod n. Now, we can think of L as the 2-digit number (a, b). So, we can radix sort these numbers in time Θ(2(n + n)) = Θ(n). In general this works to sort any n numbers over the range from 1 to nd , in Θ(dn) time. BucketSort: CountingSort and RadixSort are only good for sorting small integers, or at least objects (like characters) that can be encoded as small integers. What if you want to sort a set of floating-point numbers? In the worst-case you are pretty much stuck with using one of the comparison-based sorting algorithms, such as QuickSort, MergeSort, or HeapSort. However, in special cases where you have reason to believe that your numbers are roughly uniformly distributed over some range, then it is possible to do better. (Note Lecture Notes 10 CMSC 451
- 11. that this isa strong assumption. This algorithm should not be applied unless you have good reason to believe that this is the case.) Suppose that the numbers to be sorted range over some interval, say [0, 1). (It is possible in O(n) time to find the maximum and minimum values, and scale the numbers to fit into this range.) The idea is the subdivide this interval into n subintervals. For example, if n = 100, the subintervals would be [0, 0.01), [0.01, 0.02), [0.02, 0.03), and so on. We create n different buckets, one for each interval. Then we make a pass through the list to be sorted, and using the floor function, we can map each value to its bucket index. (In this case, the index of element x would be 100x .) We then sort each bucket in as- cending order. The number of points per bucket should be fairly small, so even a quadratic time sorting algorithm (e.g. BubbleSort or InsertionSort) should work. Finally, all the sorted buckets are concatenated together. The analysis relies on the fact that, assuming that the numbers are uniformly distributed, the number of elements lying within each bucket on average is a constant. Thus, the expected time needed to sort each bucket is O(1). Since there are n buckets, the total sorting time is Θ(n). An example illustrating this idea is given in Fig. 3. B 0 1 .10 .14 .17 2 A 3 .38 .42 .71 .10 .14 .86 .38 .59 .17 .81 .56 4 .42 5 .56 .59 6 7 .71 8 .81 .86 9 Fig. 3: BucketSort. Lecture 4: Dynamic Programming: Longest Common Subsequence Read: Introduction to Chapt 15, and Section 15.4 in CLRS. Dynamic Programming: We begin discussion of an important algorithm design technique, called dynamic program- ming (or DP for short). The technique is among the most powerful for designing algorithms for optimization problems. (This is true for two reasons. Dynamic programming solutions are based on a few common elements. Dynamic programming problems are typically optimization problems (find the minimum or maximum cost so- lution, subject to various constraints). The technique is related to divide-and-conquer, in the sense that it breaks problems down into smaller problems that it solves recursively. However, because of the somewhat different nature of dynamic programming problems, standard divide-and-conquer solutions are not usually efficient. The basic elements that characterize a dynamic programming algorithm are: Substructure: Decompose your problem into smaller (and hopefully simpler) subproblems. Express the solu- tion of the original problem in terms of solutions for smaller problems. Table-structure: Store the answers to the subproblems in a table. This is done because subproblem solutions are reused many times. Bottom-up computation: Combine solutions on smaller subproblems to solve larger subproblems. (Our text also discusses a top-down alternative, called memoization.) Lecture Notes 11 CMSC 451
- 12. The most importantquestion in designing a DP solution to a problem is how to set up the subproblem structure. This is called the formulation of the problem. Dynamic programming is not applicable to all optimization problems. There are two important elements that a problem must have in order for DP to be applicable. Optimal substructure: (Sometimes called the principle of optimality.) It states that for the global problem to be solved optimally, each subproblem should be solved optimally. (Not all optimization problems satisfy this. Sometimes it is better to lose a little on one subproblem in order to make a big gain on another.) Polynomially many subproblems: An important aspect to the efficiency of DP is that the total number of subproblems to be solved should be at most a polynomial number. Strings: One important area of algorithm design is the study of algorithms for character strings. There are a number of important problems here. Among the most important has to do with efficiently searching for a substring or generally a pattern in large piece of text. (This is what text editors and programs like “grep” do when you perform a search.) In many instances you do not want to find a piece of text exactly, but rather something that is similar. This arises for example in genetics research and in document retrieval on the web. One common method of measuring the degree of similarity between two strings is to compute their longest common subsequence. Longest Common Subsequence: Let us think of character strings as sequences of characters. Given two sequences X = x1 , x2 , . . . , xm and Z = z1 , z2 , . . . , zk , we say that Z is a subsequence of X if there is a strictly in- creasing sequence of k indices i1 , i2 , . . . , ik (1 ≤ i1 < i2 < . . . < ik ≤ n) such that Z = Xi1 , Xi2 , . . . , Xik . For example, let X = ABRACADABRA and let Z = AADAA , then Z is a subsequence of X. Given two strings X and Y , the longest common subsequence of X and Y is a longest sequence Z that is a subsequence of both X and Y . For example, let X = ABRACADABRA and let Y = YABBADABBADOO . Then the longest common subsequence is Z = ABADABA . See Fig. 4 X= A B R A C A D A B R A LCS = A B A D A B A Y= Y A B B A D A B B A D O O Fig. 4: An example of the LCS of two strings X and Y . The Longest Common Subsequence Problem (LCS) is the following. Given two sequences X = x1 , . . . , xm and Y = y1 , . . . , yn determine a longest common subsequence. Note that it is not always unique. For example the LCS of ABC and BAC is either AC or BC . DP Formulation for LCS: The simple brute-force solution to the problem would be to try all possible subsequences from one string, and search for matches in the other string, but this is hopelessly inefficient, since there are an exponential number of possible subsequences. Instead, we will derive a dynamic programming solution. In typical DP fashion, we need to break the prob- lem into smaller pieces. There are many ways to do this for strings, but it turns out for this problem that considering all pairs of prefixes will suffice for us. A prefix of a sequence is just an initial string of values, Xi = x1 , x2 , . . . , xi . X0 is the empty sequence. The idea will be to compute the longest common subsequence for every possible pair of prefixes. Let c[i, j] denote the length of the longest common subsequence of Xi and Yj . For example, in the above case we have X5 = ABRAC and Y6 = YABBAD . Their longest common subsequence is ABA . Thus, c[5, 6] = 3. Which of the c[i, j] values do we compute? Since we don’t know which will lead to the final optimum, we compute all of them. Eventually we are interested in c[m, n] since this will be the LCS of the two entire strings. The idea is to compute c[i, j] assuming that we already know the values of c[i , j ], for i ≤ i and j ≤ j (but not both equal). Here are the possible cases. Lecture Notes 12 CMSC 451
- 13. Basis: c[i, 0]= c[j, 0] = 0. If either sequence is empty, then the longest common subsequence is empty. Last characters match: Suppose xi = yj . For example: Let Xi = ABCA and let Yj = DACA . Since both end in A, we claim that the LCS must also end in A. (We will leave the proof as an exercise.) Since the A is part of the LCS we may find the overall LCS by removing A from both sequences and taking the LCS of Xi−1 = ABC and Yj−1 = DAC which is AC and then adding A to the end, giving ACA as the answer. (At first you might object: But how did you know that these two A’s matched with each other. The answer is that we don’t, but it will not make the LCS any smaller if we do.) This is illustrated at the top of Fig. 5. if xi = yj then c[i, j] = c[i − 1, j − 1] + 1 xi Xi A Xi−1 A Last chars match: yj LCS add to LCS Yj A Yj−1 A Xi−1 A skip xi LCS xi Yj B Last chars do not Xi A max match yj Yj B Xi A LCS Yj−1 B skip yj Fig. 5: The possibe cases in the DP formulation of LCS. Last characters do not match: Suppose that xi = yj . In this case xi and yj cannot both be in the LCS (since they would have to be the last character of the LCS). Thus either xi is not part of the LCS, or yj is not part of the LCS (and possibly both are not part of the LCS). At this point it may be tempting to try to make a “smart” choice. By analyzing the last few characters of Xi and Yj , perhaps we can figure out which character is best to discard. However, this approach is doomed to failure (and you are strongly encouraged to think about this, since it is a common point of confusion.) Instead, our approach is to take advantage of the fact that we have already precomputed smaller subproblems, and use these results to guide us. In the first case (xi is not in the LCS) the LCS of Xi and Yj is the LCS of Xi−1 and Yj , which is c[i − 1, j]. In the second case (yj is not in the LCS) the LCS is the LCS of Xi and Yj−1 which is c[i, j − 1]. We do not know which is the case, so we try both and take the one that gives us the longer LCS. This is illustrated at the bottom half of Fig. 5. if xi = yj then c[i, j] = max(c[i − 1, j], c[i, j − 1]) Combining these observations we have the following formulation: 0 if i = 0 or j = 0, c[i, j] = c[i − 1, j − 1] + 1 if i, j > 0 and xi = yj , max(c[i, j − 1], c[i − 1, j]) if i, j > 0 and xi = yj . Implementing the Formulation: The task now is to simply implement this formulation. We concentrate only on computing the maximum length of the LCS. Later we will see how to extract the actual sequence. We will store some helpful pointers in a parallel array, b[0..m, 0..n]. The code is shown below, and an example is illustrated in Fig. 6 Lecture Notes 13 CMSC 451
- 14. Y: 0 1 2 3 4 =n Y: 0 1 2 3 4 =n B D C B B D C B 0 0 0 0 0 0 X = BACDB 0 0 0 0 0 0 1 B 0 1 1 1 1 Y = BDCB 1 B 0 1 1 1 1 X: 2 A 0 1 1 1 1 X: 2 A 0 1 1 1 1 3 C 0 1 1 2 2 3 C 0 1 1 2 2 4 D 0 1 2 2 2 4 D 0 1 2 2 2 m= 5 B 0 1 2 2 3 LCS = BCB m= 5 B 0 1 2 2 3 start here LCS Length Table with back pointers included Fig. 6: Longest common subsequence example for the sequences X = BACDB and Y = BCDB . The numeric table entries are the values of c[i, j] and the arrow entries are used in the extraction of the sequence. Build LCS Table LCS(x[1..m], y[1..n]) { // compute LCS table int c[0..m, 0..n] for i = 0 to m // init column 0 c[i,0] = 0; b[i,0] = SKIPX for j = 0 to n // init row 0 c[0,j] = 0; b[0,j] = SKIPY for i = 1 to m // fill rest of table for j = 1 to n if (x[i] == y[j]) // take X[i] (Y[j]) for LCS c[i,j] = c[i-1,j-1]+1; b[i,j] = addXY else if (c[i-1,j] >= c[i,j-1]) // X[i] not in LCS c[i,j] = c[i-1,j]; b[i,j] = skipX else // Y[j] not in LCS c[i,j] = c[i,j-1]; b[i,j] = skipY return c[m,n] // return length of LCS } Extracting the LCS getLCS(x[1..m], y[1..n], b[0..m,0..n]) { LCSstring = empty string i = m; j = n // start at lower right while(i != 0 && j != 0) // go until upper left switch b[i,j] case addXY: // add X[i] (=Y[j]) add x[i] (or equivalently y[j]) to front of LCSstring i--; j--; break case skipX: i--; break // skip X[i] case skipY: j--; break // skip Y[j] return LCSstring } Lecture Notes 14 CMSC 451
- 15. The running timeof the algorithm is clearly O(mn) since there are two nested loops with m and n iterations, respectively. The algorithm also uses O(mn) space. Extracting the Actual Sequence: Extracting the final LCS is done by using the back pointers stored in b[0..m, 0..n]. Intuitively b[i, j] = addXY means that X[i] and Y [j] together form the last character of the LCS. So we take this common character, and continue with entry b[i − 1, j − 1] to the northwest ( ). If b[i, j] = skipX , then we know that X[i] is not in the LCS, and so we skip it and go to b[i − 1, j] above us (↑). Similarly, if b[i, j] = skipY , then we know that Y [j] is not in the LCS, and so we skip it and go to b[i, j − 1] to the left (←). Following these back pointers, and outputting a character with each diagonal move gives the final subsequence. Lecture 5: Dynamic Programming: Chain Matrix Multiplication Read: Chapter 15 of CLRS, and Section 15.2 in particular. Chain Matrix Multiplication: This problem involves the question of determining the optimal sequence for perform- ing a series of operations. This general class of problem is important in compiler design for code optimization and in databases for query optimization. We will study the problem in a very restricted instance, where the dynamic programming issues are easiest to see. Suppose that we wish to multiply a series of matrices A1 A2 . . . An Matrix multiplication is an associative but not a commutative operation. This means that we are free to paren- thesize the above multiplication however we like, but we are not free to rearrange the order of the matrices. Also recall that when two (nonsquare) matrices are being multiplied, there are restrictions on the dimensions. A p × q matrix has p rows and q columns. You can multiply a p × q matrix A times a q × r matrix B, and the result will be a p × r matrix C. (The number of columns of A must equal the number of rows of B.) In particular for 1 ≤ i ≤ p and 1 ≤ j ≤ r, q C[i, j] = A[i, k]B[k, j]. k=1 This corresponds to the (hopefully familiar) rule that the [i, j] entry of C is the dot product of the ith (horizontal) row of A and the jth (vertical) column of B. Observe that there are pr total entries in C and each takes O(q) time to compute, thus the total time to multiply these two matrices is proportional to the product of the dimensions, pqr. A * B = C q = Multiplication p p time = pqr r q r Fig. 7: Matrix Multiplication. Note that although any legal parenthesization will lead to a valid result, not all involve the same number of operations. Consider the case of 3 matrices: A1 be 5 × 4, A2 be 4 × 6 and A3 be 6 × 2. multCost[((A1 A2 )A3 )] = (5 · 4 · 6) + (5 · 6 · 2) = 180, multCost[(A1 (A2 A3 ))] = (4 · 6 · 2) + (5 · 4 · 2) = 88. Even for this small example, considerable savings can be achieved by reordering the evaluation sequence. Lecture Notes 15 CMSC 451
- 16. Chain Matrix MultiplicationProblem: Given a sequence of matrices A1 , A2 , . . . , An and dimensions p0 , p1 , . . . , pn where Ai is of dimension pi−1 × pi , determine the order of multiplication (represented, say, as a binary tree) that minimizes the number of operations. Important Note: This algorithm does not perform the multiplications, it just determines the best order in which to perform the multiplications. Naive Algorithm: We could write a procedure which tries all possible parenthesizations. Unfortunately, the number of ways of parenthesizing an expression is very large. If you have just one or two matrices, then there is only one way to parenthesize. If you have n items, then there are n − 1 places where you could break the list with the outermost pair of parentheses, namely just after the 1st item, just after the 2nd item, etc., and just after the (n − 1)st item. When we split just after the kth item, we create two sublists to be parenthesized, one with k items, and the other with n − k items. Then we could consider all the ways of parenthesizing these. Since these are independent choices, if there are L ways to parenthesize the left sublist and R ways to parenthesize the right sublist, then the total is L · R. This suggests the following recurrence for P (n), the number of different ways of parenthesizing n items: 1 if n = 1, P (n) = n−1 k=1 P (k)P (n − k) if n ≥ 2. This is related to a famous function in combinatorics called the Catalan numbers (which in turn is related to the number of different binary trees on n nodes). In particular P (n) = C(n − 1), where C(n) is the nth Catalan number: 1 2n C(n) = . n+1 n Applying Stirling’s formula (which is given in our text), we find that C(n) ∈ Ω(4n /n3/2 ). Since 4n is exponen- tial and n3/2 is just polynomial, the exponential will dominate, implying that function grows very fast. Thus, this will not be practical except for very small n. In summary, brute force is not an option. Dynamic Programming Approach: This problem, like other dynamic programming problems involves determining a structure (in this case, a parenthesization). We want to break the problem into subproblems, whose solutions can be combined to solve the global problem. As is common to any DP solution, we need to find some way to break the problem into smaller subproblems, and we need to determine a recursive formulation, which represents the optimum solution to each problem in terms of solutions to the subproblems. Let us think of how we can do this. Since matrices cannot be reordered, it makes sense to think about sequences of matrices. Let Ai..j denote the result of multiplying matrices i through j. It is easy to see that Ai..j is a pi−1 × pj matrix. (Think about this for a second to be sure you see why.) Now, in order to determine how to perform this multiplication optimally, we need to make many decisions. What we want to do is to break the problem into problems of a similar structure. In parenthesizing the expression, we can consider the highest level of parenthesization. At this level we are simply multiplying two matrices together. That is, for any k, 1 ≤ k ≤ n − 1, A1..n = A1..k · Ak+1..n . Thus the problem of determining the optimal sequence of multiplications is broken up into two questions: how do we decide where to split the chain (what is k?) and how do we parenthesize the subchains A1..k and Ak+1..n ? The subchain problems can be solved recursively, by applying the same scheme. So, let us think about the problem of determining the best value of k. At this point, you may be tempted to consider some clever ideas. For example, since we want matrices with small dimensions, pick the value of k that minimizes pk . Although this is not a bad idea, in principle. (After all it might work. It just turns out that it doesn’t in this case. This takes a bit of thinking, which you should try.) Instead, as is true in almost all dynamic programming solutions, we will do the dumbest thing of simply considering all possible choices of k, and taking the best of them. Usually trying all possible choices is bad, since it quickly leads to an exponential Lecture Notes 16 CMSC 451
- 17. number of totalpossibilities. What saves us here is that there are only O(n2 ) different sequences of matrices. (There are n = n(n − 1)/2 ways of choosing i and j to form Ai..j to be precise.) Thus, we do not encounter 2 the exponential growth. Notice that our chain matrix multiplication problem satisfies the principle of optimality, because once we decide to break the sequence into the product A1..k · Ak+1..n , we should compute each subsequence optimally. That is, for the global problem to be solved optimally, the subproblems must be solved optimally as well. Dynamic Programming Formulation: We will store the solutions to the subproblems in a table, and build the table in a bottom-up manner. For 1 ≤ i ≤ j ≤ n, let m[i, j] denote the minimum number of multiplications needed to compute Ai..j . The optimum cost can be described by the following recursive formulation. Basis: Observe that if i = j then the sequence contains only one matrix, and so the cost is 0. (There is nothing to multiply.) Thus, m[i, i] = 0. Step: If i < j, then we are asking about the product Ai..j . This can be split by considering each k, i ≤ k < j, as Ai..k times Ak+1..j . The optimum times to compute Ai..k and Ak+1..j are, by definition, m[i, k] and m[k + 1, j], respectively. We may assume that these values have been computed previously and are already stored in our array. Since Ai..k is a pi−1 × pk matrix, and Ak+1..j is a pk × pj matrix, the time to multiply them is pi−1 pk pj . This suggests the following recursive rule for computing m[i, j]. m[i, i] = 0 m[i, j] = min (m[i, k] + m[k + 1, j] + pi−1 pk pj ) for i < j. i≤k<j A i..j A i..k A k+1..j A i A i+1 ... A k A k+1 ... A j ? Fig. 8: Dynamic Programming Formulation. It is not hard to convert this rule into a procedure, which is given below. The only tricky part is arranging the order in which to compute the values. In the process of computing m[i, j] we need to access values m[i, k] and m[k + 1, j] for k lying between i and j. This suggests that we should organize our computation according to the number of matrices in the subsequence. Let L = j−i+1 denote the length of the subchain being multiplied. The subchains of length 1 (m[i, i]) are trivial to compute. Then we build up by computing the subchains of lengths 2, 3, . . . , n. The final answer is m[1, n]. We need to be a little careful in setting up the loops. If a subchain of length L starts at position i, then j = i + L − 1. Since we want j ≤ n, this means that i + L − 1 ≤ n, or in other words, i ≤ n − L + 1. So our loop for i runs from 1 to n − L + 1 (in order to keep j in bounds). The code is presented below. The array s[i, j] will be explained later. It is used to extract the actual sequence. The running time of the procedure is Θ(n3 ). We’ll leave this as an exercise in solving sums, but the key is that there are three nested loops, and each can iterate at most n times. Extracting the final Sequence: Extracting the actual multiplication sequence is a fairly easy extension. The basic idea is to leave a split marker indicating what the best split is, that is, the value of k that leads to the minimum Lecture Notes 17 CMSC 451
- 18. Chain Matrix Multiplication Matrix-Chain(array p[1..n]) { array s[1..n-1,2..n] for i = 1 to n do m[i,i] = 0; // initialize for L = 2 to n do { // L = length of subchain for i = 1 to n-L+1 do { j = i + L - 1; m[i,j] = INFINITY; for k = i to j-1 do { // check all splits q = m[i, k] + m[k+1, j] + p[i-1]*p[k]*p[j] if (q < m[i, j]) { m[i,j] = q; s[i,j] = k; } } } } return m[1,n] (final cost) and s (splitting markers); } value of m[i, j]. We can maintain a parallel array s[i, j] in which we will store the value of k providing the optimal split. For example, suppose that s[i, j] = k. This tells us that the best way to multiply the subchain Ai..j is to first multiply the subchain Ai..k and then multiply the subchain Ak+1..j , and finally multiply these together. Intuitively, s[i, j] tells us what multiplication to perform last. Note that we only need to store s[i, j] when we have at least two matrices, that is, if j > i. The actual multiplication algorithm uses the s[i, j] value to determine how to split the current sequence. Assume that the matrices are stored in an array of matrices A[1..n], and that s[i, j] is global to this recursive procedure. The recursive procedure Mult does this computation and below returns a matrix. Extracting Optimum Sequence Mult(i, j) { if (i == j) // basis case return A[i]; else { k = s[i,j] X = Mult(i, k) // X = A[i]...A[k] Y = Mult(k+1, j) // Y = A[k+1]...A[j] return X*Y; // multiply matrices X and Y } } In the figure below we show an example. This algorithm is tricky, so it would be a good idea to trace through this example (and the one given in the text). The initial set of dimensions are 5, 4, 6, 2, 7 meaning that we are multiplying A1 (5 × 4) times A2 (4 × 6) times A3 (6 × 2) times A4 (2 × 7). The optimal sequence is ((A1 (A2 A3 ))A4 ). Lecture 6: Dynamic Programming: Minimum Weight Triangulation Read: This is not covered in CLRS. Lecture Notes 18 CMSC 451
- 19. 4 1 s[i,j] 4 1 m[i,j] 3 2 158 2 3 3 j 1 3 i j 88 104 i 2 3 2 3 1 2 3 120 48 84 4 1 0 0 0 0 5 4 6 2 7 3 Final order A1 A2 A3 A4 1 2 p0 p1 p2 p3 p4 A1 A2 A3 A4 Fig. 9: Chain Matrix Multiplication Example. Polygons and Triangulations: Let’s consider a geometric problem that outwardly appears to be quite different from chain-matrix multiplication, but actually has remarkable similarities. We begin with a number of definitions. Define a polygon to be a piecewise linear closed curve in the plane. In other words, we form a cycle by joining line segments end to end. The line segments are called the sides of the polygon and the endpoints are called the vertices. A polygon is simple if it does not cross itself, that is, if the sides do not intersect one another except for two consecutive sides sharing a common vertex. A simple polygon subdivides the plane into its interior, its boundary and its exterior. A simple polygon is said to be convex if every interior angle is at most 180 degrees. Vertices with interior angle equal to 180 degrees are normally allowed, but for this problem we will assume that no such vertices exist. Polygon Simple polygon Convex polygon Fig. 10: Polygons. Given a convex polygon, we assume that its vertices are labeled in counterclockwise order P = v1 , . . . , vn . We will assume that indexing of vertices is done modulo n, so v0 = vn . This polygon has n sides, vi−1 vi . Given two nonadjacent sides vi and vj , where i < j−1, the line segment vi vj is a chord. (If the polygon is simple but not convex, we include the additional requirement that the interior of the segment must lie entirely in the interior of P .) Any chord subdivides the polygon into two polygons: vi , vi+1 , . . . , vj , and vj , vj+1 , . . . , vi . A triangulation of a convex polygon P is a subdivision of the interior of P into a collection of triangles with disjoint interiors, whose vertices are drawn from the vertices of P . Equivalently, we can define a triangulation as a maximal set T of nonintersecting chords. (In other words, every chord that is not in T intersects the interior of some chord in T .) It is easy to see that such a set of chords subdivides the interior of the polygon into a collection of triangles with pairwise disjoint interiors (and hence the name triangulation). It is not hard to prove (by induction) that every triangulation of an n-sided polygon consists of n − 3 chords and n − 2 triangles. Triangulations are of interest for a number of reasons. Many geometric algorithm operate by first decomposing a complex polygonal shape into triangles. In general, given a convex polygon, there are many possible triangulations. In fact, the number is exponential in n, the number of sides. Which triangulation is the “best”? There are many criteria that are used depending on the application. One criterion is to imagine that you must “pay” for the ink you use in drawing the triangulation, and you want to minimize the amount of ink you use. (This may sound fanciful, but minimizing wire length is an Lecture Notes 19 CMSC 451
- 20. important condition inchip design. Further, this is one of many properties which we could choose to optimize.) This suggests the following optimization problem: Minimum-weight convex polygon triangulation: Given a convex polygon determine the triangulation that minimizes the sum of the perimeters of its triangles. (See Fig. 11.) A triangulation Lower weight triangulation Fig. 11: Triangulations of convex polygons, and the minimum weight triangulation. Given three distinct vertices vi , vj , vk , we define the weight of the associated triangle by the weight function w(vi , vj , vk ) = |vi vj | + |vj vk | + |vk vi |, where |vi vj | denotes the length of the line segment vi vj . Dynamic Programming Solution: Let us consider an (n + 1)-sided polygon P = v0 , v1 , . . . , vn . Let us assume that these vertices have been numbered in counterclockwise order. To derive a DP formulation we need to define a set of subproblems from which we can derive the optimum solution. For 0 ≤ i < j ≤ n, define t[i, j] to be the weight of the minimum weight triangulation for the subpolygon that lies to the right of directed chord vi vj , that is, the polygon with the counterclockwise vertex sequence vi , vi+1 , . . . , vj . Observe that if we can compute this quantity for all such i and j, then the weight of the minimum weight triangulation of the entire polygon can be extracted as t[0, n]. (As usual, we only compute the minimum weight. But, it is easy to modify the procedure to extract the actual triangulation.) As a basis case, we define the weight of the trivial “2-sided polygon” to be zero, implying that t[i, i + 1] = 0. In general, to compute t[i, j], consider the subpolygon vi , vi+1 , . . . , vj , where j > i + 1. One of the chords of this polygon is the side vi vj . We may split this subpolygon by introducing a triangle whose base is this chord, and whose third vertex is any vertex vk , where i < k < j. This subdivides the polygon into the subpolygons vi , vi+1 , . . . vk and vk , vk+1 , . . . vj whose minimum weights are already known to us as t[i, k] and t[k, j]. In addition we should consider the weight of the newly added triangle vi vk vj . Thus, we have the following recursive rule: 0 if j = i + 1 t[i, j] = mini<k<j (t[i, k] + t[k, j] + w(vi vk vj )) if j > i + 1. The final output is the overall minimum weight, which is, t[0, n]. This is illustrated in Fig. 12 Note that this has almost exactly the same structure as the recursive definition used in the chain matrix multipli- cation algorithm (except that some indices are different by 1.) The same Θ(n3 ) algorithm can be applied with only minor changes. Relationship to Binary Trees: One explanation behind the similarity of triangulations and the chain matrix multipli- cation algorithm is to observe that both are fundamentally related to binary trees. In the case of the chain matrix multiplication, the associated binary tree is the evaluation tree for the multiplication, where the leaves of the tree correspond to the matrices, and each node of the tree is associated with a product of a sequence of two or more matrices. To see that there is a similar correspondence here, consider an (n + 1)-sided convex polygon P = v0 , v1 , . . . , vn , and fix one side of the polygon (say v0 vn ). Now consider a rooted binary tree whose root node is the triangle containing side v0 vn , whose internal nodes are the nodes of the dual tree, and whose leaves Lecture Notes 20 CMSC 451
- 21. vn v0 vj Triangulate at cost t[k,j] vi cost=w(v i ,vk, vj ) Triangulate vk at cost t[i,k] Fig. 12: Triangulations and tree structure. correspond to the remaining sides of the tree. Observe that partitioning the polygon into triangles is equivalent to a binary tree with n leaves, and vice versa. This is illustrated in Fig. 13. Note that every triangle is associated with an internal node of the tree and every edge of the original polygon, except for the distinguished starting side v0 vn , is associated with a leaf node of the tree. root A1 v0 v11 root A11 v1 A2 v10 A10 v2 A3 v9 v3 A9 v8 A4 v4 v7 A8 A5 v A7 5 v6 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A6 Fig. 13: Triangulations and tree structure. Once you see this connection. Then the following two observations follow easily. Observe that the associated binary tree has n leaves, and hence (by standard results on binary trees) n − 1 internal nodes. Since each internal node other than the root has one edge entering it, there are n − 2 edges between the internal nodes. Each internal node corresponds to one triangle, and each edge between internal nodes corresponds to one chord of the triangulation. Lecture 7: Greedy Algorithms: Activity Selection and Fractional Knapack Read: Sections 16.1 and 16.2 in CLRS. Greedy Algorithms: In many optimization algorithms a series of selections need to be made. In dynamic program- ming we saw one way to make these selections. Namely, the optimal solution is described in a recursive manner, and then is computed “bottom-up”. Dynamic programming is a powerful technique, but it often leads to algo- rithms with higher than desired running times. Today we will consider an alternative design technique, called greedy algorithms. This method typically leads to simpler and faster algorithms, but it is not as powerful or as widely applicable as dynamic programming. We will give some examples of problems that can be solved by greedy algorithms. (Later in the semester, we will see that this technique can be applied to a number of graph problems as well.) Even when greedy algorithms do not produce the optimal solution, they often provide fast heuristics (nonoptimal solution strategies), are often used in finding good approximations. Lecture Notes 21 CMSC 451
- 22. Activity Scheduling: Activityscheduling and it is a very simple scheduling problem. We are given a set S = {1, 2, . . . , n} of n activities that are to be scheduled to use some resource, where each activity must be started at a given start time si and ends at a given finish time fi . For example, these might be lectures that are to be given in a lecture hall, where the lecture times have been set up in advance, or requests for boats to use a repair facility while they are in port. Because there is only one resource, and some start and finish times may overlap (and two lectures cannot be given in the same room at the same time), not all the requests can be honored. We say that two activities i and j are noninterfering if their start-finish intervals do not overlap, more formally, [si , fi ) ∩ [sj , fj ) = ∅. (Note that making the intervals half open, two consecutive activities are not considered to interfere.) The activity scheduling problem is to select a maximum-size set of mutually noninterfering activities for use of the resource. (Notice that goal here is maximum number of activities, not maximum utilization. Of course different criteria could be considered, but the greedy approach may not be optimal in general.) How do we schedule the largest number of activities on the resource? Intuitively, we do not like long activities, because they occupy the resource and keep us from honoring other requests. This suggests the following greedy strategy: repeatedly select the activity with the smallest duration (fi − si ) and schedule it, provided that it does not interfere with any previously scheduled activities. Although this seems like a reasonable strategy, this turns out to be nonoptimal. (See Problem 17.1-4 in CLRS). Sometimes the design of a correct greedy algorithm requires trying a few different strategies, until hitting on one that works. Here is a greedy strategy that does work. The intuition is the same. Since we do not like activities that take a long time, let us select the activity that finishes first and schedule it. Then, we skip all activities that interfere with this one, and schedule the next one that has the earliest finish time, and so on. To make the selection process faster, we assume that the activities have been sorted by their finish times, that is, f1 ≤ f2 ≤ . . . ≤ fn , Assuming this sorting, the pseudocode for the rest of the algorithm is presented below. The output is the list A of scheduled activities. The variable prev holds the index of the most recently scheduled activity at any time, in order to determine interferences. Greedy Activity Scheduler schedule(s[1..n], f[1..n]) { // given start and finish times // we assume f[1..n] already sorted List A = <1> // schedule activity 1 first prev = 1 for i = 2 to n if (s[i] >= f[prev]) { // no interference? append i to A; prev = i // schedule i next } return A } It is clear that the algorithm is quite simple and efficient. The most costly activity is that of sorting the activities by finish time, so the total running time is Θ(n log n). Fig. 14 shows an example. Each activity is represented by its start-finish time interval. Observe that the intervals are sorted by finish time. Event 1 is scheduled first. It interferes with activity 2 and 3. Then Event 4 is scheduled. It interferes with activity 5 and 6. Finally, activity 7 is scheduled, and it intereferes with the remaining activity. The final output is {1, 4, 7}. Note that this is not the only optimal schedule. {2, 4, 7} is also optimal. Proof of Optimality: Our proof of optimality is based on showing that the first choice made by the algorithm is the best possible, and then using induction to show that the rest of the choices result in an optimal schedule. Proofs of optimality for greedy algorithms follow a similar structure. Suppose that you have any nongreedy solution. Lecture Notes 22 CMSC 451
- 23. Input: Add 1: 1 1 2 2 3 3 4 Sched 1; Skip 2,3 4 5 5 6 6 7 7 8 8 Add 7: Add 4: 1 1 2 2 Sched 4; Skip 5,6 3 3 Sched 7; Skip 8 4 4 5 5 6 6 7 7 8 8 Fig. 14: An example of the greedy algorithm for activity scheduling. The final schedule is {1, 4, 7}. Show that its cost can be reduced by being “greedier” at some point in the solution. This proof is complicated a bit by the fact that there may be multiple solutions. Our approach is to show that any schedule that is not greedy can be made more greedy, without decreasing the number of activities. Claim: The greedy algorithm gives an optimal solution to the activity scheduling problem. Proof: Consider any optimal schedule A that is not the greedy schedule. We will construct a new optimal schedule A that is in some sense “greedier” than A. Order the activities in increasing order of finish time. Let A = x1 , x2 , . . . , xk be the activities of A. Since A is not the same as the greedy schedule, consider the first activity xj where these two schedules differ. That is, the greedy schedule is of the form G = x1 , x2 , . . . , xj−1 , gj , . . . where gj = xj . (Note that k ≥ j, since otherwise G would have more activities than the optimal schedule, which would be a contradiction.) The greedy algorithm selects the activity with the earliest finish time that does not conflict with any earlier activity. Thus, we know that gj does not conflict with any earlier activity, and it finishes before xj . Consider the modified “greedier” schedule A that results by replacing xj with gj in the schedule A. (See Fig. 15.) That is, A = x1 , x2 , . . . , xj−1 , gj , xj+1 , . . . , xk . A: x1 x2 x3 x4 x5 A’: x1 x2 g3 x4 x5 G: x1 x2 g3 Fig. 15: Proof of optimality for the greedy schedule (j = 3). This is a feasible schedule. (Since gj cannot conflict with the earlier activities, and it does not conflict with later activities, because it finishes before xj .) It has the same number of activities as A, and therefore A Lecture Notes 23 CMSC 451
- 24. is also optimal.By repeating this process, we will eventually convert A into G, without decreasing the number of activities. Therefore, G is also optimal. Fractional Knapsack Problem: The classical (0-1) knapsack problem is a famous optimization problem. A thief is robbing a store, and finds n items which can be taken. The ith item is worth vi dollars and weighs wi pounds, where vi and wi are integers. He wants to take as valuable a load as possible, but has a knapsack that can only carry W total pounds. Which items should he take? (The reason that this is called 0-1 knapsack is that each item must be left (0) or taken entirely (1). It is not possible to take a fraction of an item or multiple copies of an item.) This optimization problem arises in industrial packing applications. For example, you may want to ship some subset of items on a truck of limited capacity. In contrast, in the fractional knapsack problem the setup is exactly the same, but the thief is allowed to take any fraction of an item for a fraction of the weight and a fraction of the value. So, you might think of each object as being a sack of gold, which you can partially empty out before taking. The 0-1 knapsack problem is hard to solve, and in fact it is an NP-complete problem (meaning that there probably doesn’t exist an efficient solution). However, there is a very simple and efficient greedy algorithm for the fractional knapsack problem. As in the case of the other greedy algorithms we have seen, the idea is to find the right order in which to process items. Intuitively, it is good to have high value and bad to have high weight. This suggests that we first sort the items according to some function that is an decreases with value and increases with weight. There are a few choices that you might try here, but only one works. Let ρi = vi /wi denote the value-per-pound ratio. We sort the items in decreasing order of ρi , and add them in this order. If the item fits, we take it all. At some point there is an item that does not fit in the remaining space. We take as much of this item as possible, thus filling the knapsack entirely. This is illustrated in Fig. 16 35 40 $140 30 $90 40 $160 60 + + + 40 30 20 $100 20 $100 20 + + 20 $100 10 5 5 $30 5 $30 knapsack $30 $20 $100 $90 $160 $270 $220 $260 ρ= 6.0 2.0 5.0 3.0 4.0 Greedy solution to Greedy solution Optimal solution Input fractional problem. to 0−1 problem. to 0−1 problem. Fig. 16: Example for the fractional knapsack problem. Correctness: It is intuitively easy to see that the greedy algorithm is optimal for the fractional problem. Given a room with sacks of gold, silver, and bronze, you would obviously take as much gold as possible, then take as much silver as possible, and then as much bronze as possible. But it would never benefit you to take a little less gold so that you could replace it with an equal volume of bronze. More formally, suppose to the contrary that the greedy algorithm is not optimal. This would mean that there is an alternate selection that is optimal. Sort the items of the alternate selection in decreasing order by ρ values. Consider the first item i on which the two selections differ. By definition, greedy takes a greater amount of item i than the alternate (because the greedy always takes as much as it can). Let us say that greedy takes x more Lecture Notes 24 CMSC 451
- 25. units of objecti than the alternate does. All the subsequent elements of the alternate selection are of lesser value than vi . By replacing x units of any such items with x units of item i, we would increase the overall value of the alternate selection. However, this implies that the alternate selection is not optimal, a contradiction. Nonoptimality for the 0-1 Knapsack: Next we show that the greedy algorithm is not generally optimal in the 0-1 knapsack problem. Consider the example shown in Fig. 16. If you were to sort the items by ρi , then you would first take the items of weight 5, then 20, and then (since the item of weight 40 does not fit) you would settle for the item of weight 30, for a total value of $30 + $100 + $90 = $220. On the other hand, if you had been less greedy, and ignored the item of weight 5, then you could take the items of weights 20 and 40 for a total value of $100 + $160 = $260. This feature of “delaying gratification” in order to come up with a better overall solution is your indication that the greedy solution is not optimal. Lecture 8: Greedy Algorithms: Huffman Coding Read: Section 16.3 in CLRS. Huffman Codes: Huffman codes provide a method of encoding data efficiently. Normally when characters are coded using standard codes like ASCII, each character is represented by a fixed-length codeword of bits (e.g. 8 bits per character). Fixed-length codes are popular, because its is very easy to break a string up into its individual characters, and to access individual characters and substrings by direct indexing. However, fixed-length codes may not be the most efficient from the perspective of minimizing the total quantity of data. Consider the following example. Suppose that we want to encode strings over the (rather limited) 4-character alphabet C = {a, b, c, d}. We could use the following fixed-length code: Character a b c d Fixed-Length Codeword 00 01 10 11 A string such as “abacdaacac” would be encoded by replacing each of its characters by the corresponding binary codeword. a b a c d a a c a c 00 01 00 10 11 00 00 10 00 10 The final 20-character binary string would be “00010010110000100010”. Now, suppose that you knew the relative probabilities of characters in advance. (This might happen by analyzing many strings over a long period of time. In applications like data compression, where you want to encode one file, you can just scan the file and determine the exact frequencies of all the characters.) You can use this knowledge to encode strings differently. Frequently occurring characters are encoded using fewer bits and less frequent characters are encoded using more bits. For example, suppose that characters are expected to occur with the following probabilities. We could design a variable-length code which would do a better job. Character a b c d Probability 0.60 0.05 0.30 0.05 Variable-Length Codeword 0 110 10 111 Notice that there is no requirement that the alphabetical order of character correspond to any sort of ordering applied to the codewords. Now, the same string would be encoded as follows. a b a c d a a c a c 0 110 0 10 111 0 0 10 0 10 Lecture Notes 25 CMSC 451
- 26. Thus, the resulting17-character string would be “01100101110010010”. Thus, we have achieved a savings of 3 characters, by using this alternative code. More generally, what would be the expected savings for a string of length n? For the 2-bit fixed-length code, the length of the encoded string is just 2n bits. For the variable-length code, the expected length of a single encoded character is equal to the sum of code lengths times the respective probabilities of their occurrences. The expected encoded string length is just n times the expected encoded character length. n(0.60 · 1 + 0.05 · 3 + 0.30 · 2 + 0.05 · 3) = n(0.60 + 0.15 + 0.60 + 0.15) = 1.5n. Thus, this would represent a 25% savings in expected encoding length. The question that we will consider today is how to form the best code, assuming that the probabilities of character occurrences are known. Prefix Codes: One issue that we didn’t consider in the example above is whether we will be able to decode the string, once encoded. In fact, this code was chosen quite carefully. Suppose that instead of coding the character ‘a’ as 0, we had encoded it as 1. Now, the encoded string “111” is ambiguous. It might be “d” and it might be “aaa”. How can we avoid this sort of ambiguity? You might suggest that we add separation markers between the encoded characters, but this will tend to lengthen the encoding, which is undesirable. Instead, we would like the code to have the property that it can be uniquely decoded. Note that in both the variable-length codes given in the example above no codeword is a prefix of another. This turns out to be the key property. Observe that if two codewords did share a common prefix, e.g. a → 001 and b → 00101, then when we see 00101 . . . how do we know whether the first character of the encoded message is a or b. Conversely, if no codeword is a prefix of any other, then as soon as we see a codeword appearing as a prefix in the encoded text, then we know that we may decode this without fear of it matching some longer codeword. Thus we have the following definition. Prefix Code: An assignment of codewords to characters so that no codeword is a prefix of any other. Observe that any binary prefix coding can be described by a binary tree in which the codewords are the leaves of the tree, and where a left branch means “0” and a right branch means “1”. The code given earlier is shown in the following figure. The length of a codeword is just its depth in the tree. The code given earlier is a prefix code, and its corresponding tree is shown in the following figure. 0 1 a 0 1 0 c 0 1 10 b d 110 111 Fig. 17: Prefix codes. Decoding a prefix code is simple. We just traverse the tree from root to leaf, letting the input character tell us which branch to take. On reaching a leaf, we output the corresponding character, and return to the root to continue the process. Expected encoding length: Once we know the probabilities of the various characters, we can determine the total length of the encoded text. Let p(x) denote the probability of seeing character x, and let dT (x) denote the length of the codeword (depth in the tree) relative to some prefix tree T . The expected number of bits needed to encode a text with n characters is given in the following formula: B(T ) = n p(x)dT (x). x∈C Lecture Notes 26 CMSC 451
- 27. This suggests thefollowing problem: Optimal Code Generation: Given an alphabet C and the probabilities p(x) of occurrence for each character x ∈ C, compute a prefix code T that minimizes the expected length of the encoded bit-string, B(T ). Note that the optimal code is not unique. For example, we could have complemented all of the bits in our earlier code without altering the expected encoded string length. There is a very simple algorithm for finding such a code. It was invented in the mid 1950’s by David Huffman, and is called a Huffman code.. By the way, this code is used by the Unix utility pack for file compression. (There are better compression methods however. For example, compress, gzip and many others are based on a more sophisticated method called the Lempel-Ziv coding.) Huffman’s Algorithm: Here is the intuition behind the algorithm. Recall that we are given the occurrence probabil- ities for the characters. We are going to build the tree up from the leaf level. We will take two characters x and y, and “merge” them into a single super-character called z, which then replaces x and y in the alphabet. The character z will have a probability equal to the sum of x and y’s probabilities. Then we continue recursively building the code on the new alphabet, which has one fewer character. When the process is completed, we know the code for z, say 010. Then, we append a 0 and 1 to this codeword, given 0100 for x and 0101 for y. Another way to think of this, is that we merge x and y as the left and right children of a root node called z. Then the subtree for z replaces x and y in the list of characters. We repeat this process until only one super-character remains. The resulting tree is the final prefix tree. Since x and y will appear at the bottom of the tree, it seem most logical to select the two characters with the smallest probabilities to perform the operation on. The result is Huffman’s algorithm. It is illustrated in the following figure. The pseudocode for Huffman’s algorithm is given below. Let C denote the set of characters. Each character x ∈ C is associated with an occurrence probability x.prob. Initially, the characters are all stored in a priority queue Q. Recall that this data structure can be built initially in O(n) time, and we can extract the element with the smallest key in O(log n) time and insert a new element in O(log n) time. The objects in Q are sorted by probability. Note that with each execution of the for-loop, the number of items in the queue decreases by one. So, after n − 1 iterations, there is exactly one element left in the queue, and this is the root of the final prefix code tree. Correctness: The big question that remains is why is this algorithm correct? Recall that the cost of any encoding tree T is B(T ) = x p(x)dT (x). Our approach will be to show that any tree that differs from the one constructed by Huffman’s algorithm can be converted into one that is equal to Huffman’s tree without increasing its cost. First, observe that the Huffman tree is a full binary tree, meaning that every internal node has exactly two children. It would never pay to have an internal node with only one child (since such a node could be deleted), so we may limit consideration to full binary trees. Claim: Consider the two characters, x and y with the smallest probabilities. Then there is an optimal code tree in which these two characters are siblings at the maximum depth in the tree. Proof: Let T be any optimal prefix code tree, and let b and c be two siblings at the maximum depth of the tree. Assume without loss of generality that p(b) ≤ p(c) and p(x) ≤ p(y) (if this is not true, then rename these characters). Now, since x and y have the two smallest probabilities it follows that p(x) ≤ p(b) and p(y) ≤ p(c). (In both cases they may be equal.) Because b and c are at the deepest level of the tree we know that d(b) ≥ d(x) and d(c) ≥ d(y). (Again, they may be equal.) Thus, we have p(b) − p(x) ≥ 0 and d(b) − d(x) ≥ 0, and hence their product is nonnegative. Now switch the positions of x and b in the tree, resulting in a new tree T . This is illustrated in the following figure. Next let us see how the cost changes as we go from T to T . Almost all the nodes contribute the same to the expected cost. The only exception are nodes x and b. By subtracting the old contributions of these Lecture Notes 27 CMSC 451
- 28. smallest a: 05 b: 48 c: 07 d: 17 e: 10 f: 13 smallest 12 b: 48 d: 17 e: 10 f: 13 a: 05 c: 07 smallest 22 b: 48 d: 17 f: 13 12 e: 10 a: 05 c: 07 smallest 22 b: 48 30 12 e: 10 d: 17 f: 13 a: 05 c: 07 smallest 52 b: 48 22 30 12 e: 10 d: 17 f: 13 a: 05 c: 07 Final Tree 0 1 b: 48 0 1 1 0 1 0 1 e: 10 d: 17 f: 13 0 1 001 010 011 a: 05 c: 07 0000 0001 Fig. 18: Huffman’s Algorithm. Lecture Notes 28 CMSC 451
- 29. Huffman’s Algorithm Huffman(int n, character C[1..n]) { Q = C; // priority queue for i = 1 to n-1 { z = new internal tree node; z.left = x = Q.extractMin(); // extract smallest probabilities z.right = y = Q.extractMin(); z.prob = x.prob + y.prob; // z’s probability is their sum Q.insert(z); // insert z into queue } return the last element left in Q as the root; } T T’ T’’ x b b y y c c b c x y x Cost change = Cost change = −(p(b)−p(x))(d(b)−d(x)) −(p(c)−p(y))(d(c)−d(y)) <0 <0 Fig. 19: Correctness of Huffman’s Algorithm. nodes and adding in the new contributions we have B(T ) = B(T ) − p(x)d(x) + p(x)d(b) − p(b)d(b) + p(b)d(x) = B(T ) + p(x)(d(b) − d(x)) − p(b)(d(b) − d(x)) = B(T ) − (p(b) − p(x))(d(b) − d(x)) ≤ B(T ) because (p(b) − p(x))(d(b) − d(x)) ≥ 0. Thus the cost does not increase, implying that T is an optimal tree. By switching y with c we get a new tree T , which by a similar argument is also optimal. The final tree T satisfies the statement of the claim. The above theorem asserts that the first step of Huffman’s algorithm is essentially the proper one to perform. The complete proof of correctness for Huffman’s algorithm follows by induction on n (since with each step, we eliminate exactly one character). Claim: Huffman’s algorithm produces the optimal prefix code tree. Proof: The proof is by induction on n, the number of characters. For the basis case, n = 1, the tree consists of a single leaf node, which is obviously optimal. Assume inductively that when strictly fewer than n characters, Huffman’s algorithm is guaranteed to pro- duce the optimal tree. We want to show it is true with exactly n characters. Suppose we have exactly n characters. The previous claim states that we may assume that in the optimal tree, the two characters of lowest probability x and y will be siblings at the lowest level of the tree. Remove x and y, replacing them with a new character z whose probability is p(z) = p(x) + p(y). Thus n − 1 characters remain. Consider any prefix code tree T made with this new set of n − 1 characters. We can convert it into a prefix code tree T for the original set of characters by undoing the previous operation and replacing z with x Lecture Notes 29 CMSC 451
- 30. and y (addinga “0” bit for x and a “1” bit for y). The cost of the new tree is B(T ) = B(T ) − p(z)d(z) + p(x)(d(z) + 1) + p(y)(d(z) + 1) = B(T ) − (p(x) + p(y))d(z) + (p(x) + p(y))(d(z) + 1) = B(T ) + (p(x) + p(y))(d(z) + 1 − d(z)) = B(T ) + p(x) + p(y). Since the change in cost depends in no way on the structure of the tree T , to minimize the cost of the final tree T , we need to build the tree T on n − 1 characters optimally. By induction, this exactly what Huffman’s algorithm does. Thus the final tree is optimal. Lecture 9: Graphs: Background and Breadth First Search Read: Review Sections 22.1 and 22.2 CLR. Graph Algorithms: We are now beginning a major new section of the course. We will be discussing algorithms for both directed and undirected graphs. Intuitively, a graph is a collection of vertices or nodes, connected by a collection of edges. Graphs are extremely important because they are a very flexible mathematical model for many application problems. Basically, any time you have a set of objects, and there is some “connection” or “re- lationship” or “interaction” between pairs of objects, a graph is a good way to model this. Examples of graphs in application include communication and transportation networks, VLSI and other sorts of logic circuits, surface meshes used for shape description in computer-aided design and geographic information systems, precedence constraints in scheduling systems. The list of application is almost too long to even consider enumerating it. Most of the problems in computational graph theory that we will consider arise because they are of importance to one or more of these application areas. Furthermore, many of these problems form the basic building blocks from which more complex algorithms are then built. Graphs and Digraphs: Most of you have encountered the notions of directed and undirected graphs in other courses, so we will give a quick overview here. Definition: A directed graph (or digraph) G = (V, E) consists of a finite set V , called the vertices or nodes, and E, a set of ordered pairs, called the edges of G. (Another way of saying this is that E is a binary relation on V .) Observe that self-loops are allowed by this definition. Some definitions of graphs disallow this. Multiple edges are not permitted (although the edges (v, w) and (w, v) are distinct). 1 4 1 3 2 3 2 4 Digraph Graph Fig. 20: Digraph and graph example. Definition: An undirected graph (or graph) G = (V, E) consists of a finite set V of vertices, and a set E of unordered pairs of distinct vertices, called the edges. (Note that self-loops are not allowed). Lecture Notes 30 CMSC 451
- 31. Note that directedgraphs and undirected graphs are different (but similar) objects mathematically. Certain notions (such as path) are defined for both, but other notions (such as connectivity) may only be defined for one, or may be defined differently. We say that vertex v is adjacent to vertex u if there is an edge (u, v). In a directed graph, given the edge e = (u, v), we say that u is the origin of e and v is the destination of e. In undirected graphs u and v are the endpoints of the edge. The edge e is incident (meaning that it touches) both u and v. In a digraph, the number of edges coming out of a vertex is called the out-degree of that vertex, and the number of edges coming in is called the in-degree. In an undirected graph we just talk about the degree of a vertex as the number of incident edges. By the degree of a graph, we usually mean the maximum degree of its vertices. When discussing the size of a graph, we typically consider both the number of vertices and the number of edges. The number of vertices is typically written as n or V , and the number of edges is written as m or E or e. Here are some basic combinatorial facts about graphs and digraphs. We will leave the proofs to you. Given a graph with V vertices and E edges then: In a graph: Number of edges: 0 ≤ E ≤ n = n(n − 1)/2 ∈ O(n2 ). 2 Sum of degrees: v∈V deg(v) = 2E. In a digraph: Number of edges: 0 ≤ E ≤ n2 . Sum of degrees: v∈V in-deg(v) = v∈V out-deg(v) = E. Notice that generally the number of edges in a graph may be as large as quadratic in the number of vertices. However, the large graphs that arise in practice typically have much fewer edges. A graph is said to be sparse if E ∈ Θ(V ), and dense, otherwise. When giving the running times of algorithms, we will usually express it as a function of both V and E, so that the performance on sparse and dense graphs will be apparent. Paths and Cycles: A path in a graph or digraph is a sequence of vertices v0 , v1 , . . . , vk such that (vi−1 , vi ) is an edge for i = 1, 2, . . . , k. The length of the path is the number of edges, k. A path is simple if all vertices and all the edges are distinct. A cycle is a path containing at least one edge and for which v0 = vk . A cycle is simple if its vertices (except v0 and vk ) are distinct, and all its edges are distinct. A graph or digraph is said to be acyclic if it contains no simple cycles. An acyclic connected graph is called a free tree or simply tree for short. (The term “free” is intended to emphasize the fact that the tree has no root, in contrast to a rooted tree, as is usually seen in data structures.) An acyclic undirected graph (which need not be connected) is a collection of free trees, and is (naturally) called a forest. An acyclic digraph is called a directed acyclic graph, or DAG for short. Simple Nonsimple Free Tree Forest DAG cycle cycle Fig. 21: Illustration of some graph terms. We say that w is reachable from u if there is a path from u to w. Note that every vertex is reachable from itself by a trivial path that uses zero edges. An undirected graph is connected if every vertex can reach every other vertex. (Connectivity is a bit messier for digraphs, and we will define it later.) The subsets of mutually reachable vertices partition the vertices of the graph into disjoint subsets, called the connected components of the graph. Lecture Notes 31 CMSC 451
- 32. Representations of Graphsand Digraphs: There are two common ways of representing graphs and digraphs. First we show how to represent digraphs. Let G = (V, E) be a digraph with n = |V | and let e = |E|. We will assume that the vertices of G are indexed {1, 2, . . . , n}. Adjacency Matrix: An n × n matrix defined for 1 ≤ v, w ≤ n. 1 if (v, w) ∈ E A[v, w] = 0 otherwise. If the digraph has weights we can store the weights in the matrix. For example if (v, w) ∈ E then A[v, w] = W (v, w) (the weight on edge (v, w)). If (v, w) ∈ E then generally W (v, w) need not be / defined, but often we set it to some “special” value, e.g. A(v, w) = −1, or ∞. (By ∞ we mean (in practice) some number which is larger than any allowable weight. In practice, this might be some machine dependent constant like MAXINT.) Adjacency List: An array Adj[1 . . . n] of pointers where for 1 ≤ v ≤ n, Adj[v] points to a linked list contain- ing the vertices which are adjacent to v (i.e. the vertices that can be reached from v by a single edge). If the edges have weights then these weights may also be stored in the linked list elements. 1 2 3 Adj 1 1 1 1 1 1 2 3 1 2 0 0 1 2 3 3 0 1 0 3 2 2 3 Adjacency matrix Adjacency list Fig. 22: Adjacency matrix and adjacency list for digraphs. We can represent undirected graphs using exactly the same representation, but we will store each edge twice. In particular, we representing the undirected edge {v, w} by the two oppositely directed edges (v, w) and (w, v). Notice that even though we represent undirected graphs in the same way that we represent digraphs, it is impor- tant to remember that these two classes of objects are mathematically distinct from one another. This can cause some complications. For example, suppose you write an algorithm that operates by marking edges of a graph. You need to be careful when you mark edge (v, w) in the representation that you also mark (w, v), since they are both the same edge in reality. When dealing with adjacency lists, it may not be convenient to walk down the entire linked list, so it is common to include cross links between corresponding edges. Adj 1 2 3 4 1 2 3 4 1 0 1 1 1 1 4 2 1 3 2 1 0 1 0 3 1 1 0 1 3 1 2 4 2 3 4 1 0 1 0 4 1 3 Adjacency matrix Adjacency list (with crosslinks) Fig. 23: Adjacency matrix and adjacency list for graphs. An adjacency matrix requires Θ(V 2 ) storage and an adjacency list requires Θ(V + E) storage. The V arises because there is one entry for each vertex in Adj . Since each list has out-deg(v) entries, when this is summed over all vertices, the total number of adjacency list records is Θ(E). For sparse graphs the adjacency list representation is more space efficient. Lecture Notes 32 CMSC 451
- 33. Graph Traversals: Thereare a number of approaches used for solving problems on graphs. One of the most impor- tant approaches is based on the notion of systematically visiting all the vertices and edge of a graph. The reason for this is that these traversals impose a type of tree structure (or generally a forest) on the graph, and trees are usually much easier to reason about than general graphs. Breadth-first search: Given an graph G = (V, E), breadth-first search starts at some source vertex s and “discovers” which vertices are reachable from s. Define the distance between a vertex v and s to be the minimum number of edges on a path from s to v. Breadth-first search discovers vertices in increasing order of distance, and hence can be used as an algorithm for computing shortest paths. At any given time there is a “frontier” of vertices that have been discovered, but not yet processed. Breadth-first search is named because it visits vertices across the entire “breadth” of this frontier. Initially all vertices (except the source) are colored white, meaning that they are undiscovered. When a vertex has first been discovered, it is colored gray (and is part of the frontier). When a gray vertex is processed, then it becomes black. The search makes use of a queue, a first-in first-out list, where elements are removed in the same order they are inserted. The first item in the queue (the next to be removed) is called the head of the queue. We will also maintain arrays color[u] which holds the color of vertex u (either white, gray or black), pred[u] which points to the predecessor of u (i.e. the vertex who first discovered u, and d[u], the distance from s to u. Only the color is really needed for the search (in fact it is only necessary to know whether a node is nonwhite). We include all this information, because some applications of BFS use this additional information. Breadth-First Search BFS(G,s) { for each u in V { // initialization color[u] = white d[u] = infinity pred[u] = null } color[s] = gray // initialize source s d[s] = 0 Q = {s} // put s in the queue while (Q is nonempty) { u = Q.Dequeue() // u is the next to visit for each v in Adj[u] { if (color[v] == white) { // if neighbor v undiscovered color[v] = gray // ...mark it discovered d[v] = d[u]+1 // ...set its distance pred[v] = u // ...and its predecessor Q.Enqueue(v) // ...put it in the queue } } color[u] = black // we are done with u } } Observe that the predecessor pointers of the BFS search define an inverted tree (an acyclic directed graph in which the source is the root, and every other node has a unique path to the root). If we reverse these edges we get a rooted unordered tree called a BFS tree for G. (Note that there are many potential BFS trees for a given graph, depending on where the search starts, and in what order vertices are placed on the queue.) These edges of G are called tree edges and the remaining edges of G are called cross edges. It is not hard to prove that if G is an undirected graph, then cross edges always go between two nodes that are at most one level apart in the BFS tree. (Can you see why this must be true?) Below is a sketch of a proof that on Lecture Notes 33 CMSC 451
- 34. a s e d f c b g s s0 a s0 c s0 a 1 1 c 1 d a 1 1 c 1 d a 1 1 c 1 d Q: a, c, d e 2 Q: c, d, e e 2 b2 Q: d, e, b d s0 b, f, g s0 e s0 a 1 1 c 1 d a 1 1 c 1 d a 1 1 c 1 d e 2 b2 e 2 b2 e 2 b2 f 3 3 g Q: (empty) f 3 3 g Q: b, f, g Q: e, b Fig. 24: Breadth-first search: Example. termination, d[v] is equal to the distance from s to v. (See the CLRS for a detailed proof.) Theorem: Let δ(s, v) denote the length (number of edges) on the shortest path from s to v. Then, on termination of the BFS procedure, d[v] = δ(s, v). Proof: (Sketch) The proof is by induction on the length of the shortest path. Let u be the predecessor of v on some shortest path from s to v, and among all such vertices the first to be processed by the BFS. Thus, δ(s, v) = δ(s, u) + 1. When u is processed, we have (by induction) d[u] = δ(s, u). Since v is a neighbor of u, we set d[v] = d[u] + 1. Thus we have d[v] = d[u] + 1 = δ(s, u) + 1 = δ(s, v), as desired. Analysis: The running time analysis of BFS is similar to the running time analysis of many graph traversal algorithms. As done in CLR V = |V | and E = |E|. Observe that the initialization portion requires Θ(V ) time. The real meat is in the traversal loop. Since we never visit a vertex twice, the number of times we go through the while loop is at most V (exactly V assuming each vertex is reachable from the source). The number of iterations through the inner for loop is proportional to deg(u) + 1. (The +1 is because even if deg(u) = 0, we need to spend a constant amount of time to set up the loop.) Summing up over all vertices we have the running time T (V ) = V + (deg(u) + 1) = V + deg(u) + V = 2V + 2E ∈ Θ(V + E). u∈V u∈V The analysis is essentially the same for directed graphs. Lecture Notes 34 CMSC 451
- 35. Lecture 10: Depth-FirstSearch Read: Sections 23.2 and 23.3 in CLR. Depth-First Search: The next traversal algorithm that we will study is called depth-first search, and it has the nice property that nontree edges have a good deal of mathematical structure. Consider the problem of searching a castle for treasure. To solve it you might use the following strategy. As you enter a room of the castle, paint some graffiti on the wall to remind yourself that you were already there. Successively travel from room to room as long as you come to a place you haven’t already been. When you return to the same room, try a different door leaving the room (assuming it goes somewhere you haven’t already been). When all doors have been tried in a given room, then backtrack. Notice that this algorithm is described recursively. In particular, when you enter a new room, you are beginning a new search. This is the general idea behind depth-first search. Depth-First Search Algorithm: We assume we are given an directed graph G = (V, E). The same algorithm works for undirected graphs (but the resulting structure imposed on the graph is different). We use four auxiliary arrays. As before we maintain a color for each vertex: white means undiscovered, gray means discovered but not finished processing, and black means finished. As before we also store predecessor pointers, pointing back to the vertex that discovered a given vertex. We will also associate two numbers with each vertex. These are time stamps. When we first discover a vertex u store a counter in d[u] and when we are finished processing a vertex we store a counter in f [u]. The purpose of the time stamps will be explained later. (Note: Do not confuse the discovery time d[v] with the distance d[v] from BFS.) The algorithm is shown in code block below, and illustrated in Fig. 25. As with BFS, DFS induces a tree structure. We will discuss this tree structure further below. Depth-First Search DFS(G) { // main program for each u in V { // initialization color[u] = white; pred[u] = null; } time = 0; for each u in V if (color[u] == white) // found an undiscovered vertex DFSVisit(u); // start a new search here } DFSVisit(u) { // start a search at u color[u] = gray; // mark u visited d[u] = ++time; for each v in Adj(u) do if (color[v] == white) { // if neighbor v undiscovered pred[v] = u; // ...set predecessor pointer DFSVisit(v); // ...visit v } color[u] = black; // we’re done with u f[u] = ++time; } Analysis: The running time of DFS is Θ(V + E). This is somewhat harder to see than the BFS analysis, because the recursive nature of the algorithm obscures things. Normally, recurrences are good ways to analyze recursively Lecture Notes 35 CMSC 451
- 36. DFS(a) DFS(b) return c DFS(f) d e DFS(c) return b DFS(g) a a a 1/.. 1/.. 1/.. b a f b b f 2/.. b 2/5 2/5 6/.. c g c c g c 3/.. 3/4 3/4 7/.. return g a d a 1/10 11/14 1/10 return f DFS(d) return a b f e DFS(e) b f 2/5 6/9 12/13 return e 2/5 6/9 return f c g c g 3/4 7/8 3/4 7/8 Fig. 25: Depth-First search tree. defined algorithms, but it is not true here, because there is no good notion of “size” that we can attach to each recursive call. First observe that if we ignore the time spent in the recursive calls, the main DFS procedure runs in O(V ) time. Observe that each vertex is visited exactly once in the search, and hence the call DFSVisit() is made exactly once for each vertex. We can just analyze each one individually and add up their running times. Ignoring the time spent in the recursive calls, we can see that each vertex u can be processed in O(1 + outdeg(u)) time. Thus the total time used in the procedure is T (V ) = V + (outdeg(u) + 1) = V + outdeg(u) + V = 2V + E ∈ Θ(V + E). u∈V u∈V A similar analysis holds if we consider DFS for undirected graphs. Tree structure: DFS naturally imposes a tree structure (actually a collection of trees, or a forest) on the structure of the graph. This is just the recursion tree, where the edge (u, v) arises when processing vertex u we call DFSVisit(v) for some neighbor v. For directed graphs the other edges of the graph can be classified as follows: Back edges: (u, v) where v is a (not necessarily proper) ancestor of u in the tree. (Thus, a self-loop is consid- ered to be a back edge). Forward edges: (u, v) where v is a proper descendent of u in the tree. Cross edges: (u, v) where u and v are not ancestors or descendents of one another (in fact, the edge may go between different trees of the forest). It is not difficult to classify the edges of a DFS tree by analyzing the values of colors of the vertices and/or considering the time stamps. This is left as an exercise. With undirected graphs, there are some important differences in the structure of the DFS tree. First, there is really no distinction between forward and back edges. So, by convention, they are all called back edges by convention. Furthermore, it can be shown that there can be no cross edges. (Can you see why not?) Lecture Notes 36 CMSC 451
- 37. Time-stamp structure: Thereis also a nice structure to the time stamps. In CLR this is referred to as the parenthesis structure. In particular, the following are easy to observe. Lemma: (Parenthesis Lemma) Given a digraph G = (V, E), and any DFS tree for G and any two vertices u, v ∈ V . • u is a descendent of v if and only if [d[u], f [u]] ⊆ [d[v], f [v]]. • u is an ancestor of v if and only if [d[u], f [u]] ⊇ [d[v], f [v]]. • u is unrelated to v if and only if [d[u], f [u]] and [d[v], f [v]] are disjoint. a d 1/10 C 11/14 a d b f C e 2/5 6/9 12/13 F B b f e c C g c g 3/4 7/8 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Fig. 26: Parenthesis Lemma. Cycles: The time stamps given by DFS allow us to determine a number of things about a graph or digraph. For example, suppose you are given a graph or digraph. You run DFS. You can determine whether the graph contains any cycles very easily. We do this with the help of the following two lemmas. Lemma: Given a digraph G = (V, E), consider any DFS forest of G, and consider any edge (u, v) ∈ E. If this edge is a tree, forward, or cross edge, then f [u] > f [v]. If the edge is a back edge then f [u] ≤ f [v]. Proof: For tree, forward, and back edges, the proof follows directly from the parenthesis lemma. (E.g. for a forward edge (u, v), v is a descendent of u, and so v’s start-finish interval is contained within u’s, implying that v has an earlier finish time.) For a cross edge (u, v) we know that the two time intervals are disjoint. When we were processing u, v was not white (otherwise (u, v) would be a tree edge), implying that v was started before u. Because the intervals are disjoint, v must have also finished before u. Lemma: Consider a digraph G = (V, E) and any DFS forest for G. G has a cycle if and only the DFS forest has a back edge. Proof: (⇐) If there is a back edge (u, v), then v is an ancestor of u, and by following tree edges from v to u we get a cycle. (⇒) We show the contrapositive. Suppose there are no back edges. By the lemma above, each of the remaining types of edges, tree, forward, and cross all have the property that they go from vertices with higher finishing time to vertices with lower finishing time. Thus along any path, finish times decrease monotonically, implying there can be no cycle. Beware: No back edges means no cycles. But you should not infer that there is some simple relationship between the number of back edges and the number of cycles. For example, a DFS tree may only have a single back edge, and there may anywhere from one up to an exponential number of simple cycles in the graph. A similar theorem applies to undirected graphs, and is not hard to prove. Lecture Notes 37 CMSC 451
- 38. Lecture 11: TopologicalSort and Strong Components Read: Sects. 22.3–22.5 in CLRS. Directed Acyclic Graph: A directed acyclic graph is often called a DAG for short DAG’s arise in many applications where there are precedence or ordering constraints. For example, if there are a series of tasks to be performed, and certain tasks must precede other tasks (e.g. in construction you have to build the first floor before you build the second floor, but you can do the electrical wiring while you install the windows). In general a precedence constraint graph is a DAG in which vertices are tasks and the edge (u, v) means that task u must be completed before task v begins. A topological sort of a DAG is a linear ordering of the vertices of the DAG such that for each edge (u, v), u appears before v in the ordering. Note that in general, there may be many legal topological orders for a given DAG. To compute a topological ordering is actually very easy, given DFS. By the previous lemma, for every edge (u, v) in a DAG, the finish time of u is greater than the finish time of v. Thus, it suffices to output the vertices in reverse order of finishing time. To do this we run a (stripped down) DFS, and when each vertex is finished we add it to the front of a linked list. The final linked list order will be the final topological order. This is given below. Topological Sort TopSort(G) { for each (u in V) color[u] = white; // initialize L = new linked_list; // L is an empty linked list for each (u in V) if (color[u] == white) TopVisit(u); return L; // L gives final order } TopVisit(u) { // start a search at u color[u] = gray; // mark u visited for each (v in Adj(u)) if (color[v] == white) TopVisit(v); Append u to the front of L; // on finishing u add to list } This is typical example of DFS is used in applications. Observe that the structure is essentially the same as the basic DFS procedure, but we only include the elements of DFS that are needed for this application. As an example we consider the DAG presented in CLRS for Professor Bumstead’s order of dressing. Bumstead lists the precedences in the order in which he puts on his clothes in the morning. We do our depth-first search in a different order from the one given in CLRS, and so we get a different final ordering. However both orderings are legitimate, given the precedence constraints. As with depth-first search, the running time of topological sort is Θ(V + E). Strong Components: Next we consider a very important connectivity problem with digraphs. When digraphs are used in communication and transportation networks, people want to know that there networks are complete in the sense that from any location it is possible to reach any other location in the digraph. A digraph is strongly connected if for every pair of vertices, u, v ∈ V , u can reach v and vice versa. We would like to write an algorithm that determines whether a digraph is strongly connected. In fact we will solve a generalization of this problem, of computing the strongly connected components (or strong components for short) of a digraph. In particular, we partition the vertices of the digraph into subsets such that the induced subgraph of each subset is strongly connected. (These subsets should be as large as possible, and still have this Lecture Notes 38 CMSC 451
- 39. shorts socks 1/10 shorts shirt 11/14 socks 15/16 pants shoes 2/9 pants tie 12/13 shirt belt tie 3/6 belt 7/8 shoes jacket 4/5 jacket Final order: socks, shirt, tie, shorts, pants, shoes, belt, jacket Fig. 27: Topological sort. property.) More formally, we say that two vertices u and v are mutually reachable if u and reach v and vice versa. It is easy to see that mutual reachability is an equivalence relation. This equivalence relation partitions the vertices into equivalence classes of mutually reachable vertices, and these are the strong components. Observe that if we merge the vertices in each strong component into a single super vertex, and joint two su- pervertices (A, B) if and only if there are vertices u ∈ A and v ∈ B such that (u, v) ∈ E, then the resulting digraph, called the component digraph, is necessarily acyclic. (Can you see why?) Thus, we may be accurately refer to it as the component DAG. c a b a,b,c d f h d,e e g f,g,h,i i Digraph and Strong Components Component DAG Fig. 28: Strong Components. The algorithm that we will present is an algorithm designer’s “dream” (and an algorithm student’s nightmare). It is amazingly simple and efficient, but it is so clever that it is very difficult to even see how it works. We will give some of the intuition that leads to the algorithm, but will not prove the algorithm’s correctness formally. See CLRS for a formal proof. Strong Components and DFS: By way of motivation, consider the DFS of the digraph shown in the following figure (left). By definition of DFS, when you enter a strong component, every vertex in the component is reachable, so the DFS does not terminate until all the vertices in the component have been visited. Thus all the vertices in a strong component must appear in the same tree of the DFS forest. Observe that in the figure each strong component is just a subtree of the DFS forest. Is it always true for any DFS? Unfortunately the answer is no. In general, many strong components may appear in the same DFS tree. (See the DFS on the right for a counterexample.) Does there always exist a way to order the DFS such that it is true? Fortunately, the answer is yes. Suppose that you knew the component DAG in advance. (This is ridiculous, because you would need to know the strong components, and that is the problem we are trying to solve. But humor me for a moment.) Further Lecture Notes 39 CMSC 451
- 40. suppose that youcomputed a reversed topological order on the component digraph. That is, (u, v) is an edge in the component digraph, then v comes before u in this reversed order (not after as it would in a normal topological ordering). Now, run DFS, but every time you need a new vertex to start the search from, select the next available vertex according to this reverse topological order of the component digraph. Here is an informal justification. Clearly once the DFS starts within a given strong component, it must visit every vertex within the component (and possibly some others) before finishing. If we do not start in reverse topological, then the search may “leak out” into other strong components, and put them in the same DFS tree. For example, in the figure below right, when the search is started at vertex a, not only does it visit its component with b and c, but the it also visits the other components as well. However, by visiting components in reverse topological order of the component tree, each search cannot “leak out” into other components, because other components would have already have been visited earlier in the search. i e c a 1/8 9/12 13/18 1/18 b d d a 2/13 14/17 h f 2/3 4/7 10/11 14/17 c e f 3/4 5/12 15/16 g b 5/6 15/16 g 6/11 i 7/10 h 8/9 Fig. 29: Two depth-first searches. This leaves us with the intuition that if we could somehow order the DFS, so that it hits the strong components according to a reverse topological order, then we would have an easy algorithm for computing strong compo- nents. However, we do not know what the component DAG looks like. (After all, we are trying to solve the strong component problem in the first place). The “trick” behind the strong component algorithm is that we can find an ordering of the vertices that has essentially the necessary property, without actually computing the component DAG. The Plumber’s Algorithm: I call this algorithm the plumber’s algorithm (because it avoids leaks). Unfortunately it is quite difficult to understand why this algorithm works. I will present the algorithm, and refer you to CLRS for the complete proof. First recall that GR (what CLRS calls GT ) is the digraph with the same vertex set as G but in which all edges have been reversed in direction. Given an adjacency list for G, it is possible to compute GR in Θ(V + E) time. (I’ll leave this as an exercise.) Observe that the strongly connected components are not affected by reversing all the digraph’s edges. If u and v are mutually reachable in G, then certainly this is still true in GR . All that changes is that the component DAG is completely reversed. The ordering trick is to order the vertices of G according to their finish times in a DFS. Then visit the nodes of GR in decreasing order of finish times. All the steps of the algorithm are quite easy to implement, and all operate in Θ(V + E) time. Here is the algorithm. Correctness: Why visit vertices in decreasing order of finish times? Why use the reversal digraph? It is difficult to justify these elements formally. Here is some intuition, though. Recall that the main intent is to visit the Lecture Notes 40 CMSC 451
- 41. Strong Components StrongComp(G) { Run DFS(G), computing finish times f[u] for each vertex u; Compute R = Reverse(G), reversing all edges of G; Sort the vertices of R (by CountingSort) in decreasing order of f[u]; Run DFS(R) using this order; Each DFS tree is a strong component; } a 1/18 b d 2/13 14/17 9 c a d f c e b f a 3/4 5/12 15/16 1 4 c e i g 6/11 d 2 f 5 h 8 b g h i e 7/10 3 g 6 7 i h 8/9 Initial DFS Reversal with new vertex order Final DFS with components Fig. 30: Strong Components Algorithm strong components in a reverse topological order. The question is how to order the vertices so that this is true. Recall from the topological sorting algorithm, that in a DAG, finish times occur in reverse topological order (i.e., the first vertex in the topological order is the one with the highest finish time). So, if we wanted to visit the components in reverse topological order, this suggests that we should visit the vertices in increasing order of finish time, starting with the lowest finishing time. This is a good starting idea, but it turns out that it doesn’t work. The reason is that there are many vertices in each strong component, and they all have different finish times. For example, in the figure above observe that in the first DFS (on the left) the lowest finish time (of 4) is achieved by vertex c, and its strong component is first, not last, in topological order. It is tempting to give up in frustration at this point. But there is something to notice about the finish times. If we consider the maximum finish time in each component, then these are related to the topological order of the component DAG. In particular, given any strong component C, define f (C) to be the maximum finish time among all vertices in this component. f (C) = max f [u]. u∈C Lemma: Consider a digraph G = (V, E) and let C and C be two distinct strong components. If there is an (u, v) of G such that u ∈ C and v ∈ C , then f (C) > f (C ). See the book for a complete proof. Here is a quick sketch. If the DFS visits C first, then the DFS will leak into C (along edge (u, v) or some other edge), and then will visit everything in C before finally returning to C. Thus, some vertex of C will finish later than every vertex of C . On the other hand, suppose that C is visited first. Because there is an edge from C to C , we know from the definition of the component DAG that there cannot be a path from C to C. So C will completely finish before we even start C. Thus all the finish times of C will be larger than the finish times of C . For example, in the previous figure, the maximum finish times for each component are 18 (for {a, b, c}), 17 (for {d, e}), and 12 (for {f, g, h, i}). The order 18, 17, 12 is a valid topological order for the component digraph. Lecture Notes 41 CMSC 451
- 42. This is abig help. It tells us that if we run DFS and compute finish times, and then run a new DFS in decreasing order of finish times, we will visit the components in topological order. The problem is that this is not what we wanted. We wanted a reverse topological order for the component DAG. So, the final trick is to reverse the digraph, by forming GR . This does not change the strong components, but it reverses the edges of the component graph, and so reverses the topological order, which is exactly what we wanted. In conclusion we have: Theorem: Consider a digraph G on which DFS has been run. Sort the vertices by decreasing order of finish time. Then a DFS of the reversed digraph GR , visits the strong components according to a reversed topological order of the component DAG of GR . Lecture 12: Minimum Spanning Trees and Kruskal’s Algorithm Read: Chapt 23 in CLRS, up through 23.2. Minimum Spanning Trees: A common problem in communications networks and circuit design is that of connect- ing together a set of nodes (communication sites or circuit components) by a network of minimal total length (where length is the sum of the lengths of connecting wires). We assume that the network is undirected. To minimize the length of the connecting network, it never pays to have any cycles (since we could break any cycle without destroying connectivity and decrease the total length). Since the resulting connection graph is connected, undirected, and acyclic, it is a free tree. The computational problem is called the minimum spanning tree problem (MST for short). More formally, given a connected, undirected graph G = (V, E), a spanning tree is an acyclic subset of edges T ⊆ E that connects all the vertices together. Assuming that each edge (u, v) of G has a numeric weight or cost, w(u, v), (may be zero or negative) we define the cost of a spanning tree T to be the sum of edges in the spanning tree w(T ) = w(u, v). (u,v)∈T A minimum spanning tree (MST) is a spanning tree of minimum weight. Note that the minimum spanning tree may not be unique, but it is true that if all the edge weights are distinct, then the MST will be distinct (this is a rather subtle fact, which we will not prove). Fig. 31 shows three spanning trees for the same graph, where the shaded rectangles indicate the edges in the spanning tree. The one on the left is not a minimum spanning tree, and the other two are. (An interesting observation is that not only do the edges sum to the same value, but in fact the same set of edge weights appear in the two MST’s. Is this a coincidence? We’ll see later.) 10 10 10 b e b e b e 4 8 7 6 4 8 7 6 4 8 7 6 a 9 d 5 g a 9 d 5 g a 9 d 5 g 9 9 9 8 2 2 8 2 2 8 2 2 c f c f c f 1 1 1 Cost = 33 Cost = 22 Cost = 22 Fig. 31: Spanning trees (the middle and right are minimum spanning trees. Steiner Minimum Trees: Minimum spanning trees are actually mentioned in the U.S. legal code. The reason is that AT&T was a government supported monopoly at one time, and was responsible for handling all telephone connections. If a company wanted to connect a collection of installations by an private internal phone system, Lecture Notes 42 CMSC 451
- 43. AT&T was required(by law) to connect them in the minimum cost manner, which is clearly a spanning tree . . . or is it? Some companies discovered that they could actually reduce their connection costs by opening a new bogus installation. Such an installation served no purpose other than to act as an intermediate point for connections. An example is shown in Fig. 32. On the left, consider four installations that lie at the corners of a 1 × 1 square. Assume that all edge lengths are just Euclidean distances. It is easy to see that the cost of any MST for this configuration is 3 (as shown on the left). However, if you introduce a new installation at the center, whose √ distance to each of the other four points is 1/ 2. It is now possible to connect these five points with a total cost √ √ of 4/ 2 = 2 2 ≈ 2.83. This is better than the MST. MST SMT Steiner point 1 Cost = 3 Cost = 2 sqrt(2) = 2.83 Fig. 32: Steiner Minimum tree. In general, the problem of determining the lowest cost interconnection tree between a given set of nodes, assum- ing that you are allowed additional nodes (called Steiner points) is called the Steiner minimum tree (or SMT for short). An interesting fact is that although there is a simple greedy algorithm for MST’s (as we will see below), the SMT problem is much harder, and in fact is NP-hard. (Luckily for AT&T, the US Legal code is rather ambiguous on the point as to whether the phone company was required to use MST’s or SMT’s in making connections.) Generic approach: We will present two greedy algorithms (Kruskal’s and Prim’s algorithms) for computing a min- imum spanning tree. Recall that a greedy algorithm is one that builds a solution by repeated selecting the cheapest (or generally locally optimal choice) among all options at each stage. An important characteristic of greedy algorithms is that once they make a choice, they never “unmake” this choice. Before presenting these algorithms, let us review some basic facts about free trees. They are all quite easy to prove. Lemma: • A free tree with n vertices has exactly n − 1 edges. • There exists a unique path between any two vertices of a free tree. • Adding any edge to a free tree creates a unique cycle. Breaking any edge on this cycle restores a free tree. Let G = (V, E) be an undirected, connected graph whose edges have numeric edge weights (which may be positive, negative or zero). The intuition behind the greedy MST algorithms is simple, we maintain a subset of edges A, which will initially be empty, and we will add edges one at a time, until A equals the MST. We say that a subset A ⊆ E is viable if A is a subset of edges in some MST. (We cannot say “the” MST, since it is not necessarily unique.) We say that an edge (u, v) ∈ E − A is safe if A ∪ {(u, v)} is viable. In other words, the choice (u, v) is a safe choice to add so that A can still be extended to form an MST. Note that if A is viable it cannot contain a cycle. A generic greedy algorithm operates by repeatedly adding any safe edge to the current spanning tree. (Note that viability is a property of subsets of edges and safety is a property of a single edge.) When is an edge safe? We consider the theoretical issues behind determining whether an edge is safe or not. Let S be a subset of the vertices S ⊆ V . A cut (S, V − S) is just a partition of the vertices into two disjoint subsets. An edge (u, v) crosses the cut if one endpoint is in S and the other is in V − S. Given a subset of edges A, we Lecture Notes 43 CMSC 451
- 44. say that acut respects A if no edge in A crosses the cut. It is not hard to see why respecting cuts are important to this problem. If we have computed a partial MST, and we wish to know which edges can be added that do not induce a cycle in the current MST, any edge that crosses a respecting cut is a possible candidate. An edge of E is a light edge crossing a cut, if among all edges crossing the cut, it has the minimum weight (the light edge may not be unique if there are duplicate edge weights). Intuition says that since all the edges that cross a respecting cut do not induce a cycle, then the lightest edge crossing a cut is a natural choice. The main theorem which drives both algorithms is the following. It essentially says that we can always augment A by adding the minimum weight edge that crosses a cut which respects A. (It is stated in complete generality, so that it can be applied to both algorithms.) MST Lemma: Let G = (V, E) be a connected, undirected graph with real-valued weights on the edges. Let A be a viable subset of E (i.e. a subset of some MST), let (S, V − S) be any cut that respects A, and let (u, v) be a light edge crossing this cut. Then the edge (u, v) is safe for A. Proof: It will simplify the proof to assume that all the edge weights are distinct. Let T be any MST for G (see Fig. ). If T contains (u, v) then we are done. Suppose that no MST contains (u, v). We will derive a contradiction. 9 u 4 v u 4 v u 4 v 7 x 8 x x 8 y y y 6 A T + (u,v) T’ = T − (x,y) + (u,v) Fig. 33: Proof of the MST Lemma. Edge (u, v) is the light edge crossing cut (S, V − S). Add the edge (u, v) to T , thus creating a cycle. Since u and v are on opposite sides of the cut, and since any cycle must cross the cut an even number of times, there must be at least one other edge (x, y) in T that crosses the cut. The edge (x, y) is not in A (because the cut respects A). By removing (x, y) we restore a spanning tree, call it T . We have w(T ) = w(T ) − w(x, y) + w(u, v). Since (u, v) is lightest edge crossing the cut, we have w(u, v) < w(x, y). Thus w(T ) < w(T ). This contradicts the assumption that T was an MST. Kruskal’s Algorithm: Kruskal’s algorithm works by attempting to add edges to the A in increasing order of weight (lightest edges first). If the next edge does not induce a cycle among the current set of edges, then it is added to A. If it does, then this edge is passed over, and we consider the next edge in order. Note that as this algorithm runs, the edges of A will induce a forest on the vertices. As the algorithm continues, the trees of this forest are merged together, until we have a single tree containing all the vertices. Observe that this strategy leads to a correct algorithm. Why? Consider the edge (u, v) that Kruskal’s algorithm seeks to add next, and suppose that this edge does not induce a cycle in A. Let A denote the tree of the forest A that contains vertex u. Consider the cut (A , V − A ). Every edge crossing the cut is not in A, and so this cut respects A, and (u, v) is the light edge across the cut (because any lighter edge would have been considered earlier by the algorithm). Thus, by the MST Lemma, (u, v) is safe. Lecture Notes 44 CMSC 451
- 45. The only trickypart of the algorithm is how to detect efficiently whether the addition of an edge will create a cycle in A. We could perform a DFS on subgraph induced by the edges of A, but this will take too much time. We want a fast test that tells us whether u and v are in the same tree of A. This can be done by a data structure (which we have not studied) called the disjoint set Union-Find data structure. This data structure supports three operations: Create-Set(u): Create a set containing a single item v. Find-Set(u): Find the set that contains a given item u. Union(u, v): Merge the set containing u and the set containing v into a common set. You are not responsible for knowing how this data structure works (which is described in CLRS). You may use it as a “black-box”. For our purposes it suffices to know that each of these operations can be performed in O(log n) time, on a set of size n. (The Union-Find data structure is quite interesting, because it can actually perform a sequence of n operations much faster than O(n log n) time. However we will not go into this here. O(log n) time is fast enough for its use in Kruskal’s algorithm.) In Kruskal’s algorithm, the vertices of the graph will be the elements to be stored in the sets, and the sets will be vertices in each tree of A. The set A can be stored as a simple list of edges. The algorithm is shown below, and an example is shown in Fig. 34. Kruskal’s Algorithm Kruskal(G=(V,E),w) { A = {} // initially A is empty for each (u in V) Create_Set(u) // create set for each vertex Sort E in increasing order by weight w for each ((u,v) from the sorted list) { if (Find_Set(u) != Find_Set(v)) { // u and v in different trees Add (u,v) to A Union(u, v) } } return A } 1 10 2 10 2 10 b e b e b e 4 8 7 6 4 8 7 6 4 8 7 6 a 9 d 5 g a 9 c 5 g a 9 c 5 c 9 9 9 8 2 2 8 2 2 8 2 2 c c c c c c 1 1 1 4 10 9 9 10 8 7 6 10 5 10 c c a c a e 4 8 7 6 4 8 7 6 4 8 6 7 c 9 c 5 c a 9 c 5 c a 9 c 5 c 9 9 9 8 2 2 8 2 2 8 2 2 c c c c c c 1 1 1 Fig. 34: Kruskal’s Algorithm. Each vertex is labeled according to the set that contains it. Analysis: How long does Kruskal’s algorithm take? As usual, let V be the number of vertices and E be the number of edges. Since the graph is connected, we may assume that E ≥ V − 1. Observe that it takes Θ(E log E) time to Lecture Notes 45 CMSC 451
- 46. sort the edges.The for-loop is iterated E times, and each iteration involves a constant number of accesses to the Union-Find data structure on a collection of V items. Thus each access is Θ(V ) time, for a total of Θ(E log V ). Thus the total running time is the sum of these, which is Θ((V + E) log V ). Since V is asymptotically no larger than E, we could write this more simply as Θ(E log V ). Lecture 13: Prim’s and Baruvka’s Algorithms for MSTs Read: Chapt 23 in CLRS. Baruvka’s algorithm is not described in CLRS. Prim’s Algorithm: Prim’s algorithm is another greedy algorithm for minimum spanning trees. It differs from Kruskal’s algorithm only in how it selects the next safe edge to add at each step. Its running time is essentially the same as Kruskal’s algorithm, O((V + E) log V ). There are two reasons for studying Prim’s algorithm. The first is to show that there is more than one way to solve a problem (an important lesson to learn in algorithm design), and the second is that Prim’s algorithm looks very much like another greedy algorithm, called Dijkstra’s algorithm, that we will study for a completely different problem, shortest paths. Thus, not only is Prim’s a different way to solve the same MST problem, it is also the same way to solve a different problem. (Whatever that means!) Different ways to grow a tree: Kruskal’s algorithm worked by ordering the edges, and inserting them one by one into the spanning tree, taking care never to introduce a cycle. Intuitively Kruskal’s works by merging or splicing two trees together, until all the vertices are in the same tree. In contrast, Prim’s algorithm builds the tree up by adding leaves one at a time to the current tree. We start with a root vertex r (it can be any vertex). At any time, the subset of edges A forms a single tree (in Kruskal’s it formed a forest). We look to add a single vertex as a leaf to the tree. The process is illustrated in the following figure. 12 12 10 10 6 6 r 7 r 7 11 3 4 u u 9 5 5 Fig. 35: Prim’s Algorithm. Observe that if we consider the set of vertices S currently part of the tree, and its complement (V − S), we have a cut of the graph and the current set of tree edges A respects this cut. Which edge should we add next? The MST Lemma from the previous lecture tells us that it is safe to add the light edge. In the figure, this is the edge of weight 4 going to vertex u. Then u is added to the vertices of S, and the cut changes. Note that some edges that crossed the cut before are no longer crossing it, and others that were not crossing the cut are. It is easy to see, that the key questions in the efficient implementation of Prim’s algorithm is how to update the cut efficiently, and how to determine the light edge quickly. To do this, we will make use of a priority queue data structure. Recall that this is the data structure used in HeapSort. This is a data structure that stores a set of items, where each item is associated with a key value. The priority queue supports three operations. insert(u, key): Insert u with the key value key in Q. extractMin(): Extract the item with the minimum key value in Q. Lecture Notes 46 CMSC 451
- 47. decreaseKey(u, new key):Decrease the value of u’s key value to new key. A priority queue can be implemented using the same heap data structure used in heapsort. All of the above operations can be performed in O(log n) time, where n is the number of items in the heap. What do we store in the priority queue? At first you might think that we should store the edges that cross the cut, since this is what we are removing with each step of the algorithm. The problem is that when a vertex is moved from one side of the cut to the other, this results in a complicated sequence of updates. There is a much more elegant solution, and this is what makes Prim’s algorithm so nice. For each vertex in u ∈ V − S (not part of the current spanning tree) we associate u with a key value key[u], which is the weight of the lightest edge going from u to any vertex in S. We also store in pred[u] the end vertex of this edge in S. If there is not edge from u to a vertex in V − S, then we set its key value to +∞. We will also need to know which vertices are in S and which are not. We do this by coloring the vertices in S black. Here is Prim’s algorithm. The root vertex r can be any vertex in V . Prim’s Algorithm Prim(G,w,r) { for each (u in V) { // initialization key[u] = +infinity; color[u] = white; } key[r] = 0; // start at root pred[r] = nil; Q = new PriQueue(V); // put vertices in Q while (Q.nonEmpty()) { // until all vertices in MST u = Q.extractMin(); // vertex with lightest edge for each (v in Adj[u]) { if ((color[v] == white) && (w(u,v) < key[v])) { key[v] = w(u,v); // new lighter edge out of v Q.decreaseKey(v, key[v]); pred[v] = u; } } color[u] = black; } [The pred pointers define the MST as an inverted tree rooted at r] } The following figure illustrates Prim’s algorithm. The arrows on edges indicate the predecessor pointers, and the numeric label in each vertex is the key value. To analyze Prim’s algorithm, we account for the time spent on each vertex as it is extracted from the priority queue. It takes O(log V ) to extract this vertex from the queue. For each incident edge, we spend potentially O(log V ) time decreasing the key of the neighboring vertex. Thus the time is O(log V + deg(u) log V ) time. The other steps of the update are constant time. So the overall running time is T (V, E) = (log V + deg(u) log V ) = (1 + deg(u)) log V u∈V u∈V = log V (1 + deg(u)) = (log V )(V + 2E) = Θ((V + E) log V ). u∈V Since G is connected, V is asymptotically no greater than E, so this is Θ(E log V ). This is exactly the same as Kruskal’s algorithm. Lecture Notes 47 CMSC 451
- 48. Q: 4,8,?,?,?,? Q: 8,8,10,?,? Q: 1,2,10,? 10 4 10 8 10 4 ? 10 10 4 8 7 6 4 8 7 6 4 8 7 6 9 ? 5 ? 9 8 5 ? 9 2 5 ? 9 9 9 8 2 2 8 2 2 8 2 2 8 ? 8 ? 1 1 1 1 1 Q: <empty> Q: 2,5 Q: 2,2,5 10 5 2 10 2 10 5 5 4 8 7 6 4 8 7 6 4 8 7 6 9 5 9 5 2 9 2 5 2 9 9 9 8 2 2 8 2 2 8 2 2 1 1 1 Fig. 36: Prim’s Algorithm. Baruvka’s Algorithm: We have seen two ways (Kruskal’s and Prim’s algorithms) for solving the MST problem. So, it may seem like complete overkill to consider yet another algorithm. This one is called Baruvka’s algorithm. It is actually the oldest of the three algorithms (invented in 1926, well before the first computers). The reason for studying this algorithm is that of the three algorithms, it is the easiest to implement on a parallel computer. Unlike Kruskal’s and Prim’s algorithms, which add edges one at a time, Baruvka’s algorithm adds a whole set of edges all at once to the MST. Baruvka’s algorithm is similar to Kruskal’s algorithm, in the sense that it works by maintaining a collection of disconnected trees. Let us call each subtree a component. Initially, each vertex is by itself in a one-vertex component. Recall that with each stage of Kruskal’s algorithm, we add the lightest-weight edge that connects two different components together. To prove Kruskal’s algorithm correct, we argued (from the MST Lemma) that the lightest such edge will be safe to add to the MST. In fact, a closer inspection of the proof reveals that the cheapest edge leaving any component is always safe. This suggests a more parallel way to grow the MST. Each component determines the lightest edge that goes from inside the component to outside the component (we don’t care where). We say that such an edge leaves the component. Note that two components might select the same edge by this process. By the above observation, all of these edges are safe, so we may add them all at once to the set A of edges in the MST. As a result, many components will be merged together into a single component. We then apply DFS to the edges of A, to identify the new components. This process is repeated until only one component remains. A fairly high-level description of Baruvka’s algorithm is given below. Baruvka’s Algorithm Baruvka(G=(V,E), w) { initialize each vertex to be its own component; A = {}; // A holds edges of the MST do { for (each component C) { find the lightest edge (u,v) with u in C and v not in C; add {u,v} to A (unless it is already there); } apply DFS to graph H=(V,A), to compute the new components; } while (there are 2 or more components); return A; // return final MST edges Lecture Notes 48 CMSC 451
- 49. There are anumber of unspecified details in Baruvka’s algorithm, which we will not spell out in detail, except to note that they can be solved in Θ(V + E) time through DFS. First, we may apply DFS, but only traversing the edges of A to compute the components. Each DFS tree will correspond to a separate component. We label each vertex with its component number as part of this process. With these labels it is easy to determine which edges go between components (since their endpoints have different labels). Then we can traverse each component again to determine the lightest edge that leaves the component. (In fact, with a little more cleverness, we can do all this without having to perform two separate DFS’s.) The algorithm is illustrated in the figure below. 13 10 13 10 b f h a h h 1 8 3 1 8 3 a 9 d 6 i a 9 c 6 h 12 11 12 11 14 2 7 14 2 7 c e g c e e 15 4 15 4 13 10 13 10 h h h a h h 1 8 3 1 8 3 h 9 h 6 h a 9 a 6 h 12 11 12 11 14 2 7 14 2 7 h h h a h h 15 4 15 4 Fig. 37: Baruvka’s Algorithm. Analysis: How long does Baruvka’s algorithm take? Observe that because each iteration involves doing a DFS, each iteration (of the outer do-while loop) can be performed in Θ(V + E) time. The question is how many iterations are required in general? We claim that there are never more than O(log n) iterations needed. To see why, let m denote the number of components at some stage. Each of the m components, will merge with at least one other component. Afterwards the number of remaining components could be a low as 1 (if they all merge together), but never higher than m/2 (if they merge in pairs). Thus, the number of components decreases by at least half with each iteration. Since we start with V components, this can happen at most lg V time, until only one component remains. Thus, the total running time is Θ((V + E) log V ) time. Again, since G is connected, V is asymptotically no larger than E, so we can write this more succinctly as Θ(E log V ). Thus all three algorithms have the same asymptotic running time. Lecture 14: Dijkstra’s Algorithm for Shortest Paths Read: Chapt 24 in CLRS. Shortest Paths: Consider the problem of computing shortest paths in a directed graph. We have already seen that breadth-first search is an O(V + E) algorithm for finding shortest paths from a single source vertex to all other vertices, assuming that the graph has no edge weights. Suppose that the graph has edge weights, and we wish to compute the shortest paths from a single source vertex to all other vertices in the graph. By the way, there are other formulations of the shortest path problem. One may want just the shortest path between a single pair of vertices. Most algorithms for this problem are variants of the single-source algorithm that we will present. There is also a single sink problem, which can be solved in the transpose digraph (that is, by reversing the edges). Computing all-pairs shortest paths can be solved by iterating a single-source algorithm over all vertices, but there are other global methods that are faster. Think of the vertices as cities, and the weights represent the cost of traveling from one city to another (nonex- istent edges can be thought of a having infinite cost). When edge weights are present, we define the length of a Lecture Notes 49 CMSC 451
- 50. path to bethe sum of edge weights along the path. Define the distance between two vertices, u and v, δ(u, v) to be the length of the minimum length path from u to v. (δ(u, u) = 0 by considering path of 0 edges from u to itself.) Single Source Shortest Paths: The single source shortest path problem is as follows. We are given a directed graph with nonnegative edge weights G = (V, E) and a distinguished source vertex, s ∈ V . The problem is to determine the distance from the source vertex to every vertex in the graph. It is possible to have graphs with negative edges, but in order for the shortest path to be well defined, we need to add the requirement that there be no cycles whose total cost is negative (otherwise you make the path infinitely short by cycling forever through such a cycle). The text discusses the Bellman-Ford algorithm for finding shortest paths assuming negative weight edges but no negative-weight cycles are present. We will discuss a simple greedy algorithm, called Dijkstra’s algorithm, which assumes there are no negative edge weights. We will stress the task of computing the minimum distance from the source to each vertex. Computing the actual path will be a fairly simple extension. As in breadth-first search, for each vertex we will have a pointer pred[v] which points back to the source. By following the predecessor pointers backwards from any vertex, we will construct the reversal of the shortest path to v. Shortest Paths and Relaxation: The basic structure of Dijkstra’s algorithm is to maintain an estimate of the shortest path for each vertex, call this d[v]. (NOTE: Don’t confuse d[v] with the d[v] in the DFS algorithm. They are completely different.) Intuitively d[v] will be the length of the shortest path that the algorithm knows of from s to v. This, value will always greater than or equal to the true shortest path distance from s to v. Initially, we know of no paths, so d[v] = ∞. Initially d[s] = 0 and all the other d[v] values are set to ∞. As the algorithm goes on, and sees more and more vertices, it attempts to update d[v] for each vertex in the graph, until all the d[v] values converge to the true shortest distances. The process by which an estimate is updated is called relaxation. Here is how relaxation works. Intuitively, if you can see that your solution is not yet reached an optimum value, then push it a little closer to the optimum. In particular, if you discover a path from s to v shorter than d[v], then you need to update d[v]. This notion is common to many optimization algorithms. Consider an edge from a vertex u to v whose weight is w(u, v). Suppose that we have already computed current estimates on d[u] and d[v]. We know that there is a path from s to u of weight d[u]. By taking this path and following it with the edge (u, v) we get a path to v of length d[u]+ w(u, v). If this path is better than the existing path of length d[v] to v, we should update d[v] to the value d[u] + w(u, v). This is illustrated in Fig. 38. We should also remember that the shortest path to v passes through u, which we do by updating v’s predecessor pointer. u u 3 5 3 5 v relax(u,v) v s 11 s 8 0 0 Fig. 38: Relaxation. Relaxing an edge Relax(u,v) { if (d[u] + w(u,v) < d[v]) { // is the path through u shorter? d[v] = d[u] + w(u,v) // yes, then take it pred[v] = u // record that we go through u } } Lecture Notes 50 CMSC 451
- 51. Observe that wheneverwe set d[v] to a finite value, there is always evidence of a path of that length. Therefore d[v] ≥ δ(s, v). If d[v] = δ(s, v), then further relaxations cannot change its value. It is not hard to see that if we perform Relax(u, v) repeatedly over all edges of the graph, the d[v] values will eventually converge to the final true distance value from s. The cleverness of any shortest path algorithm is to perform the updates in a judicious manner, so the convergence is as fast as possible. In particular, the best possible would be to order relaxation operations in such a way that each edge is relaxed exactly once. Dijkstra’s algorithm does exactly this. Dijkstra’s Algorithm: Dijkstra’s algorithm is based on the notion of performing repeated relaxations. Dijkstra’s algorithm operates by maintaining a subset of vertices, S ⊆ V , for which we claim we “know” the true distance, that is d[v] = δ(s, v). Initially S = ∅, the empty set, and we set d[s] = 0 and all others to +∞. One by one we select vertices from V − S to add to S. The set S can be implemented using an array of vertex colors. Initially all vertices are white, and we set color[v] = black to indicate that v ∈ S. How do we select which vertex among the vertices of V − S to add next to S? Here is where greedy selection comes in. Dijkstra recognized that the best way in which to perform relaxations is by increasing order of distance from the source. This way, whenever a relaxation is being performed, it is possible to infer that result of the relaxation yields the final distance value. To implement this, for each vertex in u ∈ V − S, we maintain a distance estimate d[u]. The greedy thing to do is to take the vertex of V − S for which d[u] is minimum, that is, take the unprocessed vertex that is closest (by our estimate) to s. Later we will justify why this is the proper choice. In order to perform this selection efficiently, we store the vertices of V − S in a priority queue (e.g. a heap), where the key value of each vertex u is d[u]. Note the similarity with Prim’s algorithm, although a different key value is used there. Also recall that if we implement the priority queue using a heap, we can perform the operations Insert(), Extract Min(), and Decrease Key(), on a priority queue of size n each in O(log n) time. Each vertex “knows” its location in the priority queue (e.g. has a cross reference link to the priority queue entry), and each entry in the priority queue “knows” which vertex it represents. It is important when implementing the priority queue that this cross reference information is updated. Here is Dijkstra’s algorithm. (Note the remarkable similarity to Prim’s algorithm.) An example is presented in Fig. 39. Notice that the coloring is not really used by the algorithm, but it has been included to make the connection with the correctness proof a little clearer. Because of the similarity between this and Prim’s algorithm, the running time is the same, namely Θ(E log V ). Correctness: Recall that d[v] is the distance value assigned to vertex v by Dijkstra’s algorithm, and let δ(s, v) denote the length of the true shortest path from s to v. To see that Dijkstra’s algorithm correctly gives the final true distances, we need to show that d[v] = δ(s, v) when the algorithm terminates. This is a consequence of the following lemma, which states that once a vertex u has been added to S (i.e. colored black), d[u] is the true shortest distance from s to u. Since at the end of the algorithm, all vertices are in S, then all distance estimates are correct. Lemma: When a vertex u is added to S, d[u] = δ(s, u). Proof: It will simplify the proof conceptually if we assume that all the edge weights are strictly positive (the general case of nonnegative edges is presented in the text). Suppose to the contrary that at some point Dijkstra’s algorithm first attempts to add a vertex u to S for which d[u] = δ(s, u). By our observations about relaxation, d[u] is never less than δ(s, u), thus we have d[u] > δ(s, u). Consider the situation just prior to the insertion of u. Consider the true shortest path from s to u. Because s ∈ S and u ∈ V − S, at some point this path must first jump out of S. Let (x, y) be the edge taken by the path, where x ∈ S and y ∈ V − S. (Note that it may be that x = s and/or y = u). Lecture Notes 51 CMSC 451
- 52. Dijkstra’s Algorithm Dijkstra(G,w,s) { for each (u in V) { // initialization d[u] = +infinity color[u] = white pred[u] = null } d[s] = 0 // dist to source is 0 Q = new PriQueue(V) // put all vertices in Q while (Q.nonEmpty()) { // until all vertices processed u = Q.extractMin() // select u closest to s for each (v in Adj[u]) { if (d[u] + w(u,v) < d[v]) { // Relax(u,v) d[v] = d[u] + w(u,v) Q.decreaseKey(v, d[v]) pred[v] = u } } color[u] = black } [The pred pointers define an ‘‘inverted’’ shortest path tree] } ? 1 ? 0 1 2 1 10 7 ? 5 7 7 7 s 8 s 8 s 8 0 3 2 4 5 0 3 2 4 5 0 3 2 4 5 2 2 2 ? ? 2 ? 2 7 5 5 5 5 1 6 5 1 7 5 1 5 6 6 6 7 7 7 s 8 s 8 s 8 0 3 2 4 5 0 3 2 4 5 0 3 2 4 5 2 2 2 2 7 2 7 2 7 5 5 5 Fig. 39: Dijkstra’s Algorithm example. S pred[u] u s shorter path from s to u? y x d[y] > d[u] Fig. 40: Correctness of Dijkstra’s Algorithm. Lecture Notes 52 CMSC 451
- 53. We argue thaty = u. Why? Since x ∈ S we have d[x] = δ(s, x). (Since u was the first vertex added to S which violated this, all prior vertices satisfy this.) Since we applied relaxation to x when it was added, we would have set d[y] = d[x] + w(x, y) = δ(s, y). Thus d[y] is correct, and by hypothesis, d[u] is not correct, so they cannot be the same. Now observe that since y appears somewhere along the shortest path from s to u (but not at u) and all subsequent edges following y are of positive weight, we have δ(s, y) < δ(s, u), and thus d[y] = δ(s, y) < δ(s, u) < d[u]. Thus y would have been added to S before u, in contradiction to our assumption that u is the next vertex to be added to S. Lecture 15: All-Pairs Shortest Paths Read: Section 25.2 in CLRS. All-Pairs Shortest Paths: We consider the generalization of the shortest path problem, to computing shortest paths between all pairs of vertices. Let G = (V, E) be a directed graph with edge weights. If (u, v) E, is an edge of G, then the weight of this edge is denoted w(u, v). Recall that the cost of a path is the sum of edge weights along the path. The distance between two vertices δ(u, v) is the cost of the minimum cost path between them. We will allow G to have negative cost edges, but we will not allow G to have any negative cost cycles. We consider the problem of determining the cost of the shortest path between all pairs of vertices in a weighted directed graph. We will present a Θ(n3 ) algorithm, called the Floyd-Warshall algorithm. This algorithm is based on dynamic programming. For this algorithm, we will assume that the digraph is represented as an adjacency matrix, rather than the more common adjacency list. Although adjacency lists are generally more efficient for sparse graphs, storing all the inter-vertex distances will require Ω(n2 ) storage, so the savings is not justified here. Because the algorithm is matrix-based, we will employ common matrix notation, using i, j and k to denote vertices rather than u, v, and w as we usually do. Input Format: The input is an n × n matrix w of edge weights, which are based on the edge weights in the digraph. We let wij denote the entry in row i and column j of w. 0 if i = j, wij = w(i, j) if i = j and (i, j) ∈ E, +∞ if i = j and (i, j) ∈ E. / Setting wij = ∞ if there is no edge, intuitively means that there is no direct link between these two nodes, and hence the direct cost is infinite. The reason for setting wii = 0 is that there is always a trivial path of length 0 (using no edges) from any vertex to itself. (Note that in digraphs it is possible to have self-loop edges, and so w(i, i) may generally be nonzero. It cannot be negative, since we assume that there are no negative cost cycles, and if it is positive, there is no point in using it as part of any shortest path.) The output will be an n × n distance matrix D = dij where dij = δ(i, j), the shortest path cost from vertex i to j. Recovering the shortest paths will also be an issue. To help us do this, we will also compute an auxiliary matrix mid[i, j]. The value of mid[i, j] will be a vertex that is somewhere along the shortest path from i to j. If the shortest path travels directly from i to j without passing through any other vertices, then mid[i, j] will be set to null. These intermediate values behave somewhat like the predecessor pointers in Dijkstra’s algorithm, in order to reconstruct the final shortest path in Θ(n) time. Lecture Notes 53 CMSC 451
- 54. Floyd-Warshall Algorithm: TheFloyd-Warshall algorithm dates back to the early 60’s. Warshall was interested in the weaker question of reachability: determine for each pair of vertices u and v, whether u can reach v. Floyd realized that the same technique could be used to compute shortest paths with only minor variations. The Floyd-Warshall algorithm runs in Θ(n3 ) time. As with any DP algorithm, the key is reducing a large problem to smaller problems. A natural way of doing this is by limiting the number of edges of the path, but it turns out that this does not lead to the fastest algorithm (but is an approach worthy of consideration). The main feature of the Floyd-Warshall algorithm is in finding a the best formulation for the shortest path subproblem. Rather than limiting the number of edges on the path, they instead limit the set of vertices through which the path is allowed to pass. In particular, for a path p = v1 , v2 , . . . , v we say that the vertices v2 , v3 , . . . , v −1 are the intermediate vertices of this path. Note that a path consisting of a single edge has no intermediate vertices. (k) Formulation: Define dij to be the shortest path from i to j such that any intermediate vertices on the path are chosen from the set {1, 2, . . . , k}. In other words, we consider a path from i to j which either consists of the single edge (i, j), or it visits some intermediate vertices along the way, but these intermediate can only be chosen from among {1, 2, . . . , k}. The path is free to visit any subset of these vertices, and to do so in any order. For example, in the digraph shown in (k) the Fig. 41(a), notice how the value of d5,6 changes as k varies. (k−1) i dij j (0) d5,6 = INF (no path) 5 1 3 (1) d5,6 = 13 (5,1,6) 1 6 Vertices 1,2,...,k−1 4 (2) 4 9 2 d5,6 = 9 (5,2,6) 3 (3) (k−1) (k−1) 1 d5,6 = 8 (5,3,2,6) dik dkj 4 1 6 (4) d5,6 = 6 (5,4,1,6) k (a) (b) (3) Fig. 41: Limiting intermediate vertices. For example d5,6 can go through any combination of the intermediate vertices {1, 2, 3}, of which 5, 3, 2, 6 has the lowest cost of 8. (k) Floyd-Warshall Update Rule: How do we compute dij assuming that we have already computed the previous ma- trix d(k−1) ? There are two basic cases, depending on the ways that we might get from vertex i to vertex j, assuming that the intermediate vertices are chosen from {1, 2, . . . , k}: Don’t go through k at all: Then the shortest path from i to j uses only intermediate vertices {1, . . . , k − 1} (k−1) and hence the length of the shortest path is dij . Do go through k: First observe that a shortest path does not pass through the same vertex twice, so we can assume that we pass through k exactly once. (The assumption that there are no negative cost cycles is being used here.) That is, we go from i to k, and then from k to j. In order for the overall path to be as short as possible we should take the shortest path from i to k, and the shortest path from k to j. Since of (k−1) (k−1) these paths uses intermediate vertices only in {1, 2, . . . , k − 1}, the length of the path is dik + dkj . Lecture Notes 54 CMSC 451
- 55. This suggests thefollowing recursive rule (the DP formulation) for computing d(k) , which is illustrated in Fig. 41(b). (0) dij = wij , (k) (k−1) (k−1) (k−1) dij = min dij , dik + dkj for k ≥ 1. (n) The final answer is dij because this allows all possible vertices as intermediate vertices. We could write a (k) recursive program to compute dij , but this will be prohibitively slow because the same value may be reevaluated many times. Instead, we compute it by storing the values in a table, and looking the values up as we need them. Here is the complete algorithm. We have also included mid-vertex pointers, mid[i, j] for extracting the final shortest paths. We will leave the extraction of the shortest path as an exercise. Floyd-Warshall Algorithm Floyd_Warshall(int n, int w[1..n, 1..n]) { array d[1..n, 1..n] for i = 1 to n do { // initialize for j = 1 to n do { d[i,j] = W[i,j] mid[i,j] = null } } for k = 1 to n do // use intermediates {1..k} for i = 1 to n do // ...from i for j = 1 to n do // ...to j if (d[i,k] + d[k,j]) < d[i,j]) { d[i,j] = d[i,k] + d[k,j] // new shorter path length mid[i,j] = k // new path is through k } return d // matrix of distances } An example of the algorithm’s execution is shown in Fig. 42. Clearly the algorithm’s running time is Θ(n3 ). The space used by the algorithm is Θ(n2 ). Observe that we deleted all references to the superscript (k) in the code. It is left as an exercise that this does not affect the correctness of the algorithm. (Hint: The danger is that values may be overwritten and then used later in the same phase. Consider which entries might be overwritten and then reused, they occur in row k and column k. It can be shown that the overwritten values are equal to their original values.) Lecture 16: NP-Completeness: Languages and NP Read: Chapt 34 in CLRS, up through section 34.2. Complexity Theory: At this point of the semester we have been building up your “bag of tricks” for solving algorith- mic problems. Hopefully when presented with a problem you now have a little better idea of how to go about solving the problem. What sort of design paradigm should be used (divide-and-conquer, DFS, greedy, dynamic programming), what sort of data structures might be relevant (trees, heaps, graphs) and what representations would be best (adjacency list, adjacency matrices), what is the running time of your algorithm. All of this is fine if it helps you discover an acceptably efficient algorithm to solve your problem. The question that often arises in practice is that you have tried every trick in the book, and nothing seems to work. Although Lecture Notes 55 CMSC 451
- 56. 1 1 0 8 ? 1 0 8 ? 1 4 (0) 4 (1) 1 8 ? 0 1 ? 1 8 ? 0 1 ? d = d = 4 2 2 4 ? 0 ? 4 2 2 4 12 0 5 ? 2 9 0 ? 2 9 0 5 12 9 1 9 1 ? = infinity 3 3 1 1 5 4 0 8 9 1 4 0 8 9 1 1 8 (2) ? 0 1 ? 1 7 (3) 5 0 1 6 2 d = 2 8 d = 4 2 4 12 0 5 4 2 4 12 0 5 ? 2 3 0 6 7 2 3 0 5 9 12 5 9 12 3 1 3 1 3 3 1 5 4 0 3 4 1 1 7 3 (4) 5 0 1 6 2 d = 4 2 4 7 0 5 6 5 4 7 7 2 3 0 3 1 3 Fig. 42: Floyd-Warshall Example. Newly updates entries are circled. Lecture Notes 56 CMSC 451
- 57. your algorithm cansolve small problems reasonably efficiently (e.g. n ≤ 20) the really large applications that you want to solve (e.g. n = 1, 000 or n = 10, 000) your algorithm never terminates. When you analyze its √ n running time, you realize that it is running in exponential time, perhaps n n , or 2n , or 2(2 ) , or n!, or worse! Near the end of the 60’s where there was great success in finding efficient solutions to many combinatorial prob- lems, but there was also a growing list of problems for which there seemed to be no known efficient algorithmic solutions. People began to wonder whether there was some unknown paradigm that would lead to a solution to these problems, or perhaps some proof that these problems are inherently hard to solve and no algorithmic solutions exist that run under exponential time. Near the end of the 60’s a remarkable discovery was made. Many of these hard problems were interrelated in the sense that if you could solve any one of them in polynomial time, then you could solve all of them in polynomial time. This discovery gave rise to the notion of NP-completeness, and created possibly the biggest open problems in computer science: is P = NP? We will be studying this concept over the next few lectures. This area is a radical departure from what we have been doing because the emphasis will change. The goal is no longer to prove that a problem can be solved efficiently by presenting an algorithm for it. Instead we will be trying to show that a problem cannot be solved efficiently. The question is how to do this? Laying down the rules: We need some way to separate the class of efficiently solvable problems from inefficiently solvable problems. We will do this by considering problems that can be solved in polynomial time. When designing algorithms it has been possible for us to be rather informal with various concepts. We have made use of the fact that an intelligent programmer could fill in any missing details. However, the task of proving that something cannot be done efficiently must be handled much more carefully, since we do not want leave any “loopholes” that would allow someone to subvert the rules in an unreasonable way and claim to have an efficient solution when one does not really exist. We have measured the running time of algorithms using worst-case complexity, as a function of n, the size of the input. We have defined input size variously for different problems, but the bottom line is the number of bits (or bytes) that it takes to represent the input using any reasonably efficient encoding. By a reasonably efficient encoding, we assume that there is not some significantly shorter way of providing the same information. For example, you could write numbers in unary notation 111111111 = 1002 = 8 rather than binary, but that would be unacceptably inefficient. You could describe graphs in some highly inefficient way, such as by listing all of its cycles, but this would also be unacceptable. We will assume that numbers are expressed in binary or some higher base and graphs are expressed using either adjacency matrices or adjacency lists. We will usually restrict numeric inputs to be integers (as opposed to calling them “reals”), so that it is clear that arithmetic can be performed efficiently. We have also assumed that operations on numbers can be performed in constant time. From now on, we should be more careful and assume that arithmetic operations require at least as much time as there are bits of precision in the numbers being stored. Up until now all the algorithms we have seen have had the property that their worst-case running times are bounded above by some polynomial in the input size, n. A polynomial time algorithm is any algorithm that runs in time O(nk ) where k is some constant that is independent of n. A problem is said to be solvable in polynomial time if there is a polynomial time algorithm that solves it. Some functions that do not “look” like polynomials (such as O(n log n)) are bounded above by polynomials (such as O(n2 )). Some functions that do “look” like polynomials are not. For example, suppose you have an algorithm which inputs a graph of size n and an integer k and runs in O(nk ) time. Is this a polynomial? No, because k is an input to the problem, so the user is allowed to choose k = n, implying that the running time would be O(nn ) which is not a polynomial in n. The important thing is that the exponent must be a constant independent of n. Of course, saying that all polynomial time algorithms are “efficient” is untrue. An algorithm whose running time is O(n1000 ) is certainly pretty inefficient. Nonetheless, if an algorithm runs in worse than polynomial time (e.g. 2n ), then it is certainly not efficient, except for very small values of n. Lecture Notes 57 CMSC 451
- 58. Decision Problems: Manyof the problems that we have discussed involve optimization of one form or another: find the shortest path, find the minimum cost spanning tree, find the minimum weight triangulation. For rather tech- nical reasons, most NP-complete problems that we will discuss will be phrased as decision problems. A problem is called a decision problem if its output is a simple “yes” or “no” (or you may think of this as True/False, 0/1, accept/reject). We will phrase many optimization problems in terms of decision problems. For example, the minimum spanning tree decision problem might be: Given a weighted graph G and an integer k, does G have a spanning tree whose weight is at most k? This may seem like a less interesting formulation of the problem. It does not ask for the weight of the minimum spanning tree, and it does not even ask for the edges of the spanning tree that achieves this weight. However, our job will be to show that certain problems cannot be solved efficiently. If we show that the simple decision problem cannot be solved efficiently, then the more general optimization problem certainly cannot be solved efficiently either. Language Recognition Problems: Observe that a decision problem can also be thought of as a language recognition problem. We could define a language L L = {(G, k) | G has a MST of weight at most k}. This set consists of pairs, the first element is a graph (e.g. the adjacency matrix encoded as a string) followed by an integer k encoded as a binary number. At first it may seem strange expressing a graph as a string, but obviously anything that is represented in a computer is broken down somehow into a string of bits. When presented with an input string (G, k), the algorithm would answer “yes” if (G, k) ∈ L implying that G has a spanning tree of weight at most k, and “no” otherwise. In the first case we say that the algorithm “accepts” the input and otherwise it “rejects” the input. Given any language, we can ask the question of how hard it is to determine whether a given string is in the language. For example, in the case of the MST language L, we can determine membership easily in polynomial time. We just store the graph internally, run Kruskal’s algorithm, and see whether the final optimal weight is at most k. If so we accept, and otherwise we reject. Definition: Define P to be the set of all languages for which membership can be tested in polynomial time. (Intuitively, this corresponds to the set of all decisions problems that can be solved in polynomial time.) Note that languages are sets of strings, and P is a set of languages. P is defined in terms of how hard it is computationally to recognized membership in the language. A set of languages that is defined in terms of how hard it is to determine membership is called a complexity class. Since we can compute minimum spanning trees in polynomial time, we have L ∈ P. Here is a harder one, though. M = {(G, k) | G has a simple path of length at least k}. Given a graph G and integer k how would you “recognize” whether it is in the language M ? You might try searching the graph for a simple paths, until finding one of length at least k. If you find one then you can accept and terminate. However, if not then you may spend a lot of time searching (especially if k is large, like n − 1, and no such path exists). So is M ∈ P? No one knows the answer. In fact, we will show that M is NP-complete. In what follows, we will be introducing a number of classes. We will jump back and forth between the terms “language” and “decision problems”, but for our purposes they mean the same things. Before giving all the technical definitions, let us say a bit about what the general classes look like at an intuitive level. P: This is the set of all decision problems that can be solved in polynomial time. We will generally refer to these problems as being “easy” or “efficiently solvable”. (Although this may be an exaggeration in many cases.) Lecture Notes 58 CMSC 451
- 59. NP: This isthe set of all decision problems that can be verified in polynomial time. (We will give a definition of this below.) This class contains P as a subset. Thus, it contains a number of easy problems, but it also contains a number of problems that are believed to be very hard to solve. The term NP does not mean “not polynomial”. Originally the term meant “nondeterministic polynomial time”. But it is bit more intuitive to explain the concept from the perspective of verification. NP-hard: In spite of its name, to say that problem is NP-hard does not mean that it is hard to solve. Rather it means that if we could solve this problem in polynomial time, then we could solve all NP problems in polynomial time. Note that for a problem to be NP hard, it does not have to be in the class NP. Since it is widely believed that all NP problems are not solvable in polynomial time, it is widely believed that no NP-hard problem is solvable in polynomial time. NP-complete: A problem is NP-complete if (1) it is in NP, and (2) it is NP-hard. That is, NPC = NP∩NP-hard. The figure below illustrates one way that the sets P, NP, NP-hard, and NP-complete (NPC) might look. We say might because we do not know whether all of these complexity classes are distinct or whether they are all solvable in polynomial time. There are some problems in the figure that we will not discuss. One is Graph Isomorphism, which asks whether two graphs are identical up to a renaming of their vertices. It is known that this problem is in NP, but it is not known to be in P. The other is QBF, which stands for Quantified Boolean Formulas. In this problem you are given a boolean formula with quantifiers (∃ and ∀) and you want to know whether the formula is true or false. This problem is beyond the scope of this course, but may be discussed in an advanced course on complexity theory. QBF NP−Hard Harder No Ham. Cycle Knapsack NPC Hamiltonian Cycle Satisfiability NP Graph Isomorphism? P MST Strong connectivity Easy One way that things ‘might’ be. Fig. 43: The (possible) structure of P, NP, and related complexity classes. Polynomial Time Verification and Certificates: Before talking about the class of NP-complete problems, it is im- portant to introduce the notion of a verification algorithm. Many language recognition problems that may be very hard to solve, but they have the property that it is easy to verify whether a string is in the language. Consider the following problem, called the Hamiltonian cycle problem. Given an undirected graph G, does G have a cycle that visits every vertex exactly once. (There is a similar problem on directed graphs, and there is also a version which asks whether there is a path that visits all vertices.) We can describe this problem as a language recognition problem, where the language is HC = {(G) | G has a Hamiltonian cycle}, where (G) denotes an encoding of a graph G as a string. The Hamiltonian cycle problem seems to be much harder, and there is no known polynomial time algorithm for this problem. For example, the figure below shows two graphs, one which is Hamiltonian and one which is not. However, suppose that a graph did have a Hamiltonian cycle. Then it would be a very easy matter for someone to convince us of this. They would simply say “the cycle is v3 , v7 , v1 , . . . , v13 ”. We could then inspect the Lecture Notes 59 CMSC 451
- 60. Nonhamiltonian Hamiltonian Fig. 44: Hamiltonian cycle. graph, and check that this is indeed a legal cycle and that it visits all the vertices of the graph exactly once. Thus, even though we know of no efficient way to solve the Hamiltonian cycle problem, there is a very efficient way to verify that a given graph is in HC. The given cycle is called a certificate. This is some piece of information which allows us to verify that a given string is in a language. More formally, given a language L, and given x ∈ L, a verification algorithm is an algorithm which given x and a string y called the certificate, can verify that x is in the language L using this certificate as help. If x is not in L then there is nothing to verify. Note that not all languages have the property that they are easy to verify. For example, consider the following languages: UHC = {(G) | G has a unique Hamiltonian cycle} HC = {(G) | G has no Hamiltonian cycle}. Suppose that a graph G is in the language UHC. What information would someone give us that would allow us to verify that G is indeed in the language? They could give us an example of the unique Hamiltonian cycle, and we could verify that it is a Hamiltonian cycle, but what sort of certificate could they give us to convince us that this is the only one? They could give another cycle that is NOT Hamiltonian, but this does not mean that there is not another cycle somewhere that is Hamiltonian. They could try to list every other cycle of length n, but this would not be at all efficient, since there are n! possible cycles in general. Thus, it is hard to imagine that someone could give us some information that would allow us to efficiently convince ourselves that a given graph is in the language. The class NP: Definition: Define NP to be the set of all languages that can be verified by a polynomial time algorithm. Why is the set called “NP” rather than “VP”? The original term NP stood for “nondeterministic polynomial time”. This referred to a program running on a nondeterministic computer that can make guesses. Basically, such a computer could nondeterministically guess the value of certificate, and then verify that the string is in the language in polynomial time. We have avoided introducing nondeterminism here. It would be covered in a course on complexity theory or formal language theory. Like P, NP is a set of languages based on some complexity measure (the complexity of verification). Observe that P ⊆ NP. In other words, if we can solve a problem in polynomial time, then we can certainly verify membership in polynomial time. (More formally, we do not even need to see a certificate to solve the problem, we can solve it in polynomial time anyway). However it is not known whether P = NP. It seems unreasonable to think that this should be so. In other words, just being able to verify that you have a correct solution does not help you in finding the actual solution very much. Most experts believe that P = NP, but no one has a proof of this. Next time we will define the notions of NP-hard and NP-complete. Lecture Notes 60 CMSC 451
- 61. Lecture 17: NP-Completeness:Reductions Read: Chapt 34, through Section 34.4. Summary: Last time we introduced a number of concepts, on the way to defining NP-completeness. In particular, the following concepts are important. Decision Problems: are problems for which the answer is either yes or no. NP-complete problems are ex- pressed as decision problems, and hence can be thought of as language recognition problems, assuming that the input has been encoded as a string. We encode inputs as strings. For example: HC= {G | G has a Hamiltonian cycle} MST = {(G, x) | G has a MST of cost at most x}. P: is the class of all decision problems which can be solved in polynomial time, O(nk ) for some constant k. For example MST ∈ P but HC is not known (and suspected not) to be in P. Certificate: is a piece of evidence that allows us to verify in polynomial time that a string is in a given language. For example, suppose that the language is the set of Hamiltonian graphs. To convince someone that a graph is in this language, we could supply the certificate consisting of a sequence of vertices along the cycle. It is easy to access the adjacency matrix to determine that this is a legitimate cycle in G. Therefore HC ∈ NP. NP: is defined to be the class of all languages that can be verified in polynomial time. Note that since all languages in P can be solved in polynomial time, they can certainly be verified in polynomial time, so we have P ⊆ NP. However, NP also seems to have some pretty hard problems to solve, such as HC. Reductions: The class of NP-complete problems consists of a set of decision problems (languages) (a subset of the class NP) that no one knows how to solve efficiently, but if there were a polynomial time solution for even a single NP-complete problem, then every problem in NP would be solvable in polynomial time. To establish this, we need to introduce the concept of a reduction. Before discussing reductions, let us just consider the following question. Suppose that there are two problems, H and U . We know (or you strongly believe at least) that H is hard, that is it cannot be solved in polynomial time. On the other hand, the complexity of U is unknown, but we suspect that it too is hard. We want to prove that U cannot be solved in polynomial time. How would we do this? We want to show that (H ∈ P) ⇒ (U ∈ P). / / To do this, we could prove the contrapositive, (U ∈ P) ⇒ (H ∈ P). In other words, to show that U is not solvable in polynomial time, we will suppose that there is an algorithm that solves U in polynomial time, and then derive a contradiction by showing that H can be solved in polynomial time. How do we do this? Suppose that we have a subroutine that can solve any instance of problem U in polynomial time. Then all we need to do is to show that we can use this subroutine to solve problem H in polynomial time. Thus we have “reduced” problem H to problem U . It is important to note here that this supposed subroutine is really a fantasy. We know (or strongly believe) that H cannot be solved in polynomial time, thus we are essentially proving that the subroutine cannot exist, implying that U cannot be solved in polynomial time. (Be sure that you understand this, this the basis behind all reductions.) Example: 3-Colorability and Clique Cover: Let us consider an example to make this clearer. The following prob- lem is well-known to be NP-complete, and hence it is strongly believed that the problem cannot be solved in polynomial time. Lecture Notes 61 CMSC 451
- 62. 3-coloring (3Col): Givena graph G, can each of its vertices be labeled with one of 3 different “colors”, such that no two adjacent vertices have the same label. Coloring arises in various partitioning problems, where there is a constraint that two objects cannot be assigned to the same set of the partition. The term “coloring” comes from the original application which was in map drawing. Two countries that share a common border should be colored with different colors. It is well known that planar graphs can be colored with 4 colors, and there exists a polynomial time algorithm for this. But determining whether 3 colors are possible (even for planar graphs) seems to be hard and there is no known polynomial time algorithm. In the figure below we give two graphs, one is 3-colorable and one is not. 3−colorable Not 3−colorable Clique cover (size = 3) Fig. 45: 3-coloring and Clique Cover. The 3Col problem will play the role of the hard problem H, which we strongly suspect to not be solvable in polynomial time. For our unknown problem U , consider the following problem. Given a graph G = (V, E), we say that a subset of vertices V ⊆ V forms a clique if for every pair of vertices u, v ∈ V (u, v) ∈ E. That is, the subgraph induced by V is a complete graph. Clique Cover (CCov): Given a graph G = (V, E) and an integer k, can we partition the vertex set into k subsets of vertices V1 , V2 , . . . , Vk , such that i Vi = V , and that each Vi is a clique of G. The clique cover problem arises in applications of clustering. We put an edge between two nodes if they are similar enough to be clustered in the same group. We want to know whether it is possible to cluster all the vertices into k groups. Suppose that you want to solve the CCov problem, but after a while of fruitless effort, you still cannot find a polynomial time algorithm for the CCov problem. How can you prove that CCov is likely to not have a polynomial time solution? You know that 3Col is NP-complete, and hence experts believe that 3Col ∈ P. You / feel that there is some connection between the CCov problem and the 3Col problem. Thus, you want to show that (3Col ∈ P) ⇒ (CCov ∈ P), / / which you will show by proving the contrapositive (CCov ∈ P) ⇒ (3Col ∈ P). To do this, you assume that you have access to a subroutine CCov(G, k). Given a graph G and an integer k, this subroutine returns true if G has a clique cover of size k and false otherwise, and furthermore, this subroutine runs in polynomial time. How can we use this “alleged” subroutine to solve the well-known hard 3Col problem? We want to write a polynomial time subroutine for 3Col, and this subroutine is allowed to call the subroutine CCov(G, k) for any graph G and any integer k. Both problems involve partitioning the vertices up into groups. The only difference here is that in one problem the number of cliques is specified as part of the input and in the other the number of color classes is fixed at 3. In the clique cover problem, for two vertices to be in the same group they must be adjacent to each other. In the 3-coloring problem, for two vertices to be in the same color group, they must not be adjacent. In some sense, the problems are almost the same, but the requirement adjacent/non-adjacent is exactly reversed. Lecture Notes 62 CMSC 451
- 63. We claim thatwe can reduce the 3-coloring problem to the clique cover problem as follows. Given a graph G for which we want to determine its 3-colorability, output the pair (G, 3) where G denotes the complement of G. (That is, G is a graph on the same vertices, but (u, v) is an edge of G if and only if it is not an edge of G.) We can then feed the pair (G, 3) into a subroutine for clique cover. This is illustrated in the figure below. _ _ G G H H 3−colorable Coverable by 3 cliques Not 3−colorable Not coverable Fig. 46: Clique covers in the complement. Claim: A graph G is 3-colorable if and only if its complement G has a clique-cover of size 3. In other words, G ∈ 3Col iff (G, 3) ∈ CCov. Proof: (⇒) If G 3-colorable, then let V1 , V2 , V3 be the three color classes. We claim that this is a clique cover of size 3 for G, since if u and v are distinct vertices in Vi , then {u, v} ∈ E(G) (since adjacent vertices / cannot have the same color) which implies that {u, v} ∈ E(G). Thus every pair of distinct vertices in Vi are adjacent in G. (⇐) Suppose G has a clique cover of size 3, denoted V1 , V2 , V3 . For i ∈ {1, 2, 3} give the vertices of Vi color i. We assert that this is a legal coloring for G, since if distinct vertices u and v are both in Vi , then {u, v} ∈ E(G) (since they are in a common clique), implying that {u, v} ∈ E((G). Hence, two vertices / with the same color are not adjacent. Polynomial-time reduction: We now take this intuition of reducing one problem to another through the use of a subroutine call, and place it on more formal footing. Notice that in the example above, we converted an instance of the 3-coloring problem (G) into an equivalent instance of the Clique Cover problem (G, 3). Definition: We say that a language (i.e. decision problem) L1 is polynomial-time reducible to language L2 (written L1 ≤P L2 ) if there is a polynomial time computable function f , such that for all x, x ∈ L1 if and only if f (x) ∈ L2 . In the previous example we showed that 3Col ≤P CCov. In particular we have f (G) = (G, 3). Note that it is easy to complement a graph in O(n2 ) (i.e. polynomial) time (e.g. flip 0’s and 1’s in the adjacency matrix). Thus f is computable in polynomial time. Intuitively, saying that L1 ≤P L2 means that “if L2 is solvable in polynomial time, then so is L1 .” This is because a polynomial time subroutine for L2 could be applied to f (x) to determine whether f (x) ∈ L2 , or equivalently whether x ∈ L1 . Thus, in sense of polynomial time computability, L1 is “no harder” than L2 . The way in which this is used in NP-completeness is exactly the converse. We usually have strong evidence that L1 is not solvable in polynomial time, and hence the reduction is effectively equivalent to saying “since L1 is not likely to be solvable in polynomial time, then L2 is also not likely to be solvable in polynomial time.” Thus, this is how polynomial time reductions can be used to show that problems are as hard to solve as known difficult problems. Lemma: If L1 ≤P L2 and L2 ∈ P then L1 ∈ P . Lecture Notes 63 CMSC 451
- 64. Lemma: If L1≤P L2 and L1 ∈ P then L2 ∈ P . / / One important fact about reducibility is that it is transitive. In other words Lemma: If L1 ≤P L2 and L2 ≤P L3 then L1 ≤P L3 . The reason is that if two functions f (x) and g(x) are computable in polynomial time, then their composition f (g(x)) is computable in polynomial time as well. It should be noted that our text uses the term “reduction” where most other books use the term “transformation”. The distinction is subtle, but people taking other courses in complexity theory should be aware of this. NP-completeness: The set of NP-complete problems are all problems in the complexity class NP, for which it is known that if any one is solvable in polynomial time, then they all are, and conversely, if any one is not solvable in polynomial time, then none are. This is made mathematically formal using the notion of polynomial time reductions. Definition: A language L is NP-hard if: L ≤P L for all L ∈ NP. (Note that L does not need to be in NP.) Definition: A language L is NP-complete if: (1) L ∈ N P and (2) L is NP-hard. An alternative (and usually easier way) to show that a problem is NP-complete is to use transitivity. Lemma: L is NP-complete if (1) L ∈ N P and (2) L ≤P L for some known NP-complete language L . The reason is that all L ∈ N P are reducible to L (since L is NP-complete and hence NP-hard) and hence by transitivity L is reducible to L, implying that L is NP-hard. This gives us a way to prove that problems are NP-complete, once we know that one problem is NP-complete. Unfortunately, it appears to be almost impossible to prove that one problem is NP-complete, because the defini- tion says that we have to be able to reduce every problem in NP to this problem. There are infinitely many such problems, so how can we ever hope to do this? We will talk about this next time with Cook’s theorem. Cook showed that there is one problem called SAT (short for boolean satisfiability) that is NP-complete. To prove a second problem is NP-complete, all we need to do is to show that our problem is in NP (and hence it is reducible to SAT), and then to show that we can reduce SAT (or generally some known NPC problem) to our problem. It follows that our problem is equivalent to SAT (with respect to solvability in polynomial time). This is illustrated in the figure below. Lecture 18: Cook’s Theorem, 3SAT, and Independent Set Read: Chapter 34, through 34.5. The reduction given here is similar, but not the same as the reduction given in the text. Recap: So far we introduced the definitions of NP-completeness. Recall that we mentioned the following topics: P: is the set of decision problems (or languages) that are solvable in polynomial time. NP: is the set of decision problems (or languages) that can be verified in polynomial time, Lecture Notes 64 CMSC 451
- 65. Your problem Your reduction Known NPC SAT NPC SAT NPC NPC NP NP NP P P P Proving a problem is in NP Proving a problem is NP−hard Resulting structure Fig. 47: Structure of NPC and reductions. Polynomial reduction: L1 ≤P L2 means that there is a polynomial time computable function f such that x ∈ L1 if and only if f (x) ∈ L2 . A more intuitive to think about this, is that if we had a subroutine to solve L2 in polynomial time, then we could use it to solve L1 in polynomial time. Polynomial reductions are transitive, that is, if L1 ≤P L2 and L2 ≤P L3 , then L1 ≤P L3 . NP-Hard: L is NP-hard if for all L ∈ NP, L ≤P L. Thus, if we could solve L in polynomial time, we could solve all NP problems in polynomial time. NP-Complete: L is NP-complete if (1) L ∈ NP and (2) L is NP-hard. The importance of NP-complete problems should now be clear. If any NP-complete problems (and generally any NP-hard problem) is solvable in polynomial time, then every NP-complete problem (and in fact every problem in NP) is also solvable in polynomial time. Conversely, if we can prove that any NP-complete problem (and generally any problem in NP) cannot be solved in polynomial time, then every NP-complete problem (and generally every NP-hard problem) cannot be solved in polynomial time. Thus all NP-complete problems are equivalent to one another (in that they are either all solvable in polynomial time, or none are). An alternative way to show that a problem is NP-complete is to use transitivity of ≤P . Lemma: L is NP-complete if (1) L ∈ N P and (2) L ≤P L for some NP-complete language L . Note: The known NP-complete problem L is reduced to the candidate NP-complete problem L. Keep this order in mind. Cook’s Theorem: Unfortunately, to use this lemma, we need to have at least one NP-complete problem to start the ball rolling. Stephen Cook showed that such a problem existed. Cook’s theorem is quite complicated to prove, but we’ll try to give a brief intuitive argument as to why such a problem might exist. For a problem to be in NP, it must have an efficient verification procedure. Thus virtually all NP problems can be stated in the form, “does there exists X such that P (X)”, where X is some structure (e.g. a set, a path, a partition, an assignment, etc.) and P (X) is some property that X must satisfy (e.g. the set of objects must fill the knapsack, or the path must visit every vertex, or you may use at most k colors and no two adjacent vertices can have the same color). In showing that such a problem is in NP, the certificate consists of giving X, and the verification involves testing that P (X) holds. In general, any set X can be described by choosing a set of objects, which in turn can be described as choosing the values of some boolean variables. Similarly, the property P (X) that you need to satisfy, can be described as a boolean formula. Stephen Cook was looking for the most general possible property he could, since this should represent the hardest problem in NP to solve. He reasoned that computers (which represent the most general Lecture Notes 65 CMSC 451
- 66. type of computationaldevices known) could be described entirely in terms of boolean circuits, and hence in terms of boolean formulas. If any problem were hard to solve, it would be one in which X is an assignment of boolean values (true/false, 0/1) and P (X) could be any boolean formula. This suggests the following problem, called the boolean satisfiability problem. SAT: Given a boolean formula, is there some way to assign truth values (0/1, true/false) to the variables of the formula, so that the formula evaluates to true? A boolean formula is a logical formula which consists of variables xi , and the logical operations x meaning the negation of x, boolean-or (x ∨ y) and boolean-and (x ∧ y). Given a boolean formula, we say that it is satisfiable if there is a way to assign truth values (0 or 1) to the variables such that the final result is 1. (As opposed to the case where no matter how you assign truth values the result is always 0.) For example, (x1 ∧ (x2 ∨ x3 )) ∧ ((x2 ∧ x3 ) ∨ x1 ) is satisfiable, by the assignment x1 = 1, x2 = 0, x3 = 0 On the other hand, (x1 ∨ (x2 ∧ x3 )) ∧ (x1 ∨ (x2 ∧ x3 )) ∧ (x2 ∨ x3 ) ∧ (x2 ∨ x3 ) is not satisfiable. (Observe that the last two clauses imply that one of x2 and x3 must be true and the other must be false. This implies that neither of the subclauses involving x2 and x3 in the first two clauses can be satisfied, but x1 cannot be set to satisfy them either.) Cook’s Theorem: SAT is NP complete. We will not prove this theorem. The proof would take about a full lecture (not counting the week or so of background on Turing machines). In fact, it turns out that a even more restricted version of the satisfiability problem is NP-complete. A literal is a variable or its negation x or x. A formula is in 3-conjunctive normal form (3-CNF) if it is the boolean-and of clauses where each clause is the boolean-or of exactly 3 literals. For example (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x3 ∨ x4 ) ∧ (x2 ∨ x3 ∨ x4 ) is in 3-CNF form. 3SAT is the problem of determining whether a formula in 3-CNF is satisfiable. It turns out that it is possible to modify the proof of Cook’s theorem to show that the more restricted 3SAT is also NP-complete. As an aside, note that if we replace the 3 in 3SAT with a 2, then everything changes. If a boolean formula is given in 2SAT, then it is possible to determine its satisfiability in polynomial time. Thus, even a seemingly small change can be the difference between an efficient algorithm and none. NP-completeness proofs: Now that we know that 3SAT is NP-complete, we can use this fact to prove that other problems are NP-complete. We will start with the independent set problem. Independent Set (IS): Given an undirected graph G = (V, E) and an integer k does G contain a subset V of k vertices such that no two vertices in V are adjacent to one another. For example, the graph shown in the figure below has an independent set (shown with shaded nodes) of size 4. The independent set problem arises when there is some sort of selection problem, but there are mutual restrictions pairs that cannot both be selected. (For example, you want to invite as many of your friends to your party, but many pairs do not get along, represented by edges between them, and you do not want to invite two enemies.) Note that if a graph has an independent set of size k, then it has an independent set of all smaller sizes. So the corresponding optimization problem would be to find an independent set of the largest size in a graph. Often the vertices have weights, so we might talk about the problem of computing the independent set with the largest total weight. However, since we want to show that the problem is hard to solve, we will consider the simplest version of the problem. Lecture Notes 66 CMSC 451
- 67. Fig. 48: IndependentSet. Claim: IS is NP-complete. The proof involves two parts. First, we need to show that IS ∈ NP. The certificate consists of the k vertices of V . We simply verify that for each pair of vertex u, v ∈ V , there is no edge between them. Clearly this can be done in polynomial time, by an inspection of the adjacency matrix. boolean formula (in 3−CNF) graph and integer 3SAT F IS yes (G,k) f no polynomial time computable Fig. 49: Reduction of 3-SAT to IS. Secondly, we need to establish that IS is NP-hard, which can be done by showing that some known NP-complete problem (3SAT) is polynomial-time reducible to IS, that is, 3SAT ≤P IS. Let F be a boolean formula in 3-CNF form (the boolean-and of clauses, each of which is the boolean-or of 3 literals). We wish to find a polynomial time computable function f that maps F into a input for the IS problem, a graph G and integer k. That is, f (F ) = (G, k), such that F is satisfiable if and only if G has an independent set of size k. This will mean that if we can solve the independent set problem for G and k in polynomial time, then we would be able to solve 3SAT in polynomial time. An important aspect to reductions is that we do not attempt to solve the satisfiability problem. (Remember: It is NP-complete, and there is not likely to be any polynomial time solution.) So the function f must operate without knowledge of whether F is satisfiable. The idea is to translate the similar elements of the satisfiable problem to corresponding elements of the independent set problem. What is to be selected? 3SAT: Which variables are assigned to be true. Equivalently, which literals are assigned true. IS: Which vertices are to be placed in V . Requirements: 3SAT: Each clause must contain at least one literal whose value it true. IS: V must contain at least k vertices. Restrictions: 3SAT: If xi is assigned true, then xi must be false, and vice versa. Lecture Notes 67 CMSC 451
- 68. IS: If uis selected to be in V , and v is a neighbor of u, then v cannot be in V . We want a function f , which given any 3-CNF boolean formula F , converts it into a pair (G, k) such that the above elements are translated properly. Our strategy will be to create one vertex for each literal that appears within each clause. (Thus, if there are m clauses in F , there will be 3m vertices in G.) The vertices are grouped into clause clusters, one for each clause. Selecting a true literal from some clause corresponds to selecting a vertex to add to V . We set k to the number of clauses. This forces the independent set to pick one vertex from each clause, thus, one literal from each clause is true. In order to keep the IS subroutine from selecting two literals from some clause (and hence none from some other), we will connect all the vertices in each clause cluster to each other. To keep the IS subroutine from selecting both a literal and its complement, we will put an edge between each literal and its complement. This enforces the condition that if a literal is put in the IS (set to true) then its complement literal cannot also be true. A formal description of the reduction is given below. The input is a boolean formula F in 3-CNF, and the output is a graph G and integer k. 3SAT to IS Reduction k ← number of clauses in F ; for each clause C in F create a clause cluster of 3 vertices from the literals of C; for each clause cluster (x1 , x2 , x3 ) create an edge (xi , xj ) between all pairs of vertices in the cluster; for each vertex xi create edges between xi and all its complement vertices xi ; return (G, k); Given any reasonable encoding of F , it is an easy programming exercise to create G (say as an adjacency matrix) in polynomial time. We claim that F is satisfiable if and only if G has an independent set of size k. Example: Suppose that we are given the 3-CNF formula: (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ). The reduction produces the graph shown in the following figure and sets k = 4. x1 x2 x3 x1 x2 x3 x1 x1 x1 x1 x2 x2 x2 x2 x3 x3 x3 x3 x1 x2 x3 x1 x2 x3 The reduction Correctness (x1=x2=1, x3=0) Fig. 50: 3SAT to IS Reduction for (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ). In our example, the formula is satisfied by the assignment x1 = 1, x2 = 1, and x3 = 0. Note that the literal x1 satisfies the first and last clauses, x2 satisfies the second, and x3 satifies the third. Observe that by selecting the corresponding vertices from the clusters, we get an independent set of size k = 4. Lecture Notes 68 CMSC 451
- 69. Correctness Proof: Weclaim that F is satisfiable if and only if G has an independent set of size k. If F is satisfiable, then each of the k clauses of F must have at least one true literal. Let V denote the corresponding vertices from each of the clause clusters (one from each cluster). Because we take vertices from each cluster, there are no inter-cluster edges between them, and because we cannot set a variable and its complement to both be true, there can be no edge of the form (xi , xi ) between the vertices of V . Thus, V is an independent set of size k. Conversely, if G has an independent set V of size k. First observe that we must select a vertex from each clause cluster, because there are k clusters, and we cannot take two vertices from the same cluster (because they are all interconnected). Consider the assignment in which we set all of these literals to 1. This assignment is logically consistent, because we cannot have two vertices labeled xi and xi in the same cluster. Finally the transformation clearly runs in polynomial time. This completes the NP-completeness proof. Observe that our reduction did not attempt to solve the IS problem nor to solve the 3SAT. Also observe that the reduction had no knowledge of the solution to either problem. (We did not assume that the formula was satisfiable, nor did we assume we knew which variables to set to 1.) This is because computing these things would require exponential time (by the best known algorithms). Instead the reduction simply translated the input from one problem into an equivalent input to the other problem, while preserving the critical elements to each problem. Lecture 19: Clique, Vertex Cover, and Dominating Set Read: Chapt 34 (up through 34.5). The dominating set proof is not given in our text. Recap: Last time we gave a reduction from 3SAT (satisfiability of boolean formulas in 3-CNF form) to IS (indepen- dent set in graphs). Today we give a few more examples of reductions. Recall that to show that a problem is NP-complete we need to show (1) that the problem is in NP (i.e. we can verify when an input is in the language), and (2) that the problem is NP-hard, by showing that some known NP-complete problem can be reduced to this problem (there is a polynomial time function that transforms an input for one problem into an equivalent input for the other problem). Some Easy Reductions: We consider some closely related NP-complete problems next. Clique (CLIQUE): The clique problem is: given an undirected graph G = (V, E) and an integer k, does G have a subset V of k vertices such that for each distinct u, v ∈ V , {u, v} ∈ E. In other words, does G have a k vertex subset whose induced subgraph is complete. Vertex Cover (VC): A vertex cover in an undirected graph G = (V, E) is a subset of vertices V ⊆ V such that every edge in G has at least one endpoint in V . The vertex cover problem (VC) is: given an undirected graph G and an integer k, does G have a vertex cover of size k? Dominating Set (DS): A dominating set in a graph G = (V, E) is a subset of vertices V such that every vertex in the graph is either in V or is adjacent to some vertex in V . The dominating set problem (DS) is: given a graph G = (V, E) and an integer k, does G have a dominating set of size k? Don’t confuse the clique (CLIQUE) problem with the clique-cover (CC) problem that we discussed in an earlier lecture. The clique problem seeks to find a single clique of size k, and the clique-cover problem seeks to partition the vertices into k groups, each of which is a clique. We have discussed the facts that cliques are of interest in applications dealing with clustering. The vertex cover problem arises in various servicing applications. For example, you have a compute network and a program that checks the integrity of the communication links. To save the space of installing the program on every computer in the network, it suffices to install it on all the computers forming a vertex cover. From these nodes all the links can be tested. Dominating set is useful in facility location problems. For example, suppose we want to select where to place a set of fire stations such that every house in the city is within 2 minutes of the nearest Lecture Notes 69 CMSC 451
- 70. fire station. Wecreate a graph in which two locations are adjacent if they are within 2 minutes of each other. A minimum sized dominating set will be a minimum set of locations such that every other location is reachable within 2 minutes from one of these sites. The CLIQUE problem is obviously closely related to the independent set problem (IS): Given a graph G does it have a k vertex subset that is completely disconnected. It is not quite as clear that the vertex cover problem is related. However, the following lemma makes this connection clear as well. G G G V’ is CLIQUE of iff V’ is an IS of iff V−V’ is a VC of size k in G size k in G size n−k in G Fig. 51: Clique, Independent set, and Vertex Cover. Lemma: Given an undirected graph G = (V, E) with n vertices and a subset V ⊆ V of size k. The following are equivalent: (i) V is a clique of size k for the complement, G. (ii) V is an independent set of size k for G. (iii) V − V is a vertex cover of size n − k for G. Proof: (i) ⇒ (ii): If V is a clique for G, then for each u, v ∈ V , {u, v} is an edge of G implying that {u, v} is not an edge of G, implying that V is an independent set for G. (ii) ⇒ (iii): If V is an independent set for G, then for each u, v ∈ V , {u, v} is not an edge of G, implying that every edge in G is incident to a vertex in V − V , implying that V − V is a VC for G. (iii) ⇒ (i): If V − V is a vertex cover for G, then for any u, v ∈ V there is no edge {u, v} in G, implying that there is an edge {u, v} in G, implying that V is a clique in G. V is an independent set for G. Thus, if we had an algorithm for solving any one of these problems, we could easily translate it into an algorithm for the others. In particular, we have the following. Theorem: CLIQUE is NP-complete. CLIQUE ∈ NP: The certificate consists of the k vertices in the clique. Given such a certificate we can easily verify in polynomial time that all pairs of vertices in the set are adjacent. IS ≤P CLIQUE: We want to show that given an instance of the IS problem (G, k), we can produce an equiv- alent instance of the CLIQUE problem in polynomial time. The reduction function f inputs G and k, and outputs the pair (G, k). Clearly this can be done in polynomial time. By the above lemma, this instance is equivalent. Theorem: VC is NP-complete. VC ∈ NP: The certificate consists of the k vertices in the vertex cover. Given such a certificate we can easily verify in polynomial time that every edge is incident to one of these vertices. Lecture Notes 70 CMSC 451
- 71. IS ≤P VC:We want to show that given an instance of the IS problem (G, k), we can produce an equivalent instance of the VC problem in polynomial time. The reduction function f inputs G and k, computes the number of vertices, n, and then outputs (G, n − k). Clearly this can be done in polynomial time. By the lemma above, these instances are equivalent. Note: Note that in each of the above reductions, the reduction function did not know whether G has an inde- pendent set or not. It must run in polynomial time, and IS is an NP-complete problem. So it does not have time to determine whether G has an independent set or which vertices are in the set. Dominating Set: As with vertex cover, dominating set is an example of a graph covering problem. Here the condition is a little different, each vertex is adjacent to the members of the dominating set, as opposed to each edge being incident to each member of the dominating set. Obviously, if G is connected and has a vertex cover of size k, then it has a dominating set of size k (the same set of vertices), but the converse is not necessarily true. However the similarity suggests that if VC in NP-complete, then DS is likely to be NP-complete as well. The main result of this section is just this. Theorem: DS is NP-complete. As usual the proof has two parts. First we show that DS ∈ NP. The certificate just consists of the subset V in the dominating set. In polynomial time we can determine whether every vertex is in V or is adjacent to a vertex in V . Reducing Vertex Cover to Dominating Set: Next we show that an existing NP-complete problem is reducible to dominating set. We choose vertex cover and show that VC ≤P DS. We want a polynomial time function, which given an instance of the vertex cover problem (G, k), produces an instance (G , k ) of the dominating set problem, such that G has a vertex cover of size k if and only if G has a dominating set of size k . How to we translate between these problems? The key difference is the condition. In VC: “every edge is incident to a vertex in V ”. In DS: “every vertex is either in V or is adjacent to a vertex in V ”. Thus the translation must somehow map the notion of “incident” to “adjacent”. Because incidence is a property of edges, and adjacency is a property of vertices, this suggests that the reduction function maps edges of G into vertices in G , such that an incident edge in G is mapped to an adjacent vertex in G . This suggests the following idea (which does not quite work). We will insert a vertex into the middle of each edge of the graph. In other words, for each edge {u, v}, we will create a new special vertex, called wuv , and replace the edge {u, v} with the two edges {u, wuv } and {v, wuv }. The fact that u was incident to edge {u, v} has now been replaced with the fact that u is adjacent to the corresponding vertex wuv . We still need to dominate the neighbor v. To do this, we will leave the edge {u, v} in the graph as well. Let G be the resulting graph. This is still not quite correct though. Define an isolated vertex to be one that is incident to no edges. If u is isolated it can only be dominated if it is included in the dominating set. Since it is not incident to any edges, it does not need to be in the vertex cover. Let VI denote the isolated vertices in G, and let I denote the number of isolated vertices. The number of vertices to request for the dominating set will be k = k + I. Now we can give the complete reduction. Given the pair (G, k) for the VC problem, we create a graph G as follows. Initially G = G. For each edge {u, v} in G we create a new vertex wuv in G and add edges {u, wuv } and {v, wuv } in G . Let I denote the number of isolated vertices and set k = k + I. Output (G , k ). This reduction illustrated in the following figure. Note that every step can be performed in polynomial time. Correctness of the Reduction: To establish the correctness of the reduction, we need to show that G has a vertex cover of size k if and only if G has a dominating set of size k . First we argue that if V is a vertex cover for G, then V = V ∪ VI is a dominating set for G . Observe that |V | = |V ∪ VI | ≤ k + I = k . Note that |V ∪ VI | might be of size less than k + I, if there are any isolated vertices in V . If so, we can add any vertices we like to make the size equal to k . Lecture Notes 71 CMSC 451
- 72. f k=3 k’=3+1=4 G G’ Fig. 52: Dominating set reduction. To see that V is a dominating set, first observe that all the isolated vertices are in V and so they are dominated. Second, each of the special vertices wuv in G corresponds to an edge {u, v} in G implying that either u or v is in the vertex cover V . Thus wuv is dominated by the same vertex in V Finally, each of the nonisolated original vertices v is incident to at least one edge in G, and hence either it is in V or else all of its neighbors are in V . In either case, v is either in V or adjacent to a vertex in V . This is shown in the top part of the following figure. vertex cover for G dominating set for G’ dominating set for G’ using original vertices vertex cover for G Fig. 53: Correctness of the VC to DS reduction (where k = 3 and I = 1). Conversely, we claim that if G has a dominating set V of size k = k + I then G has a vertex cover V of size k. Note that all I isolated vertices of G must be in the dominating set. First, let V = V − VI be the remaining k vertices. We might try to claim something like: V is a vertex cover for G. But this will not necessarily work, because V may have vertices that are not part of the original graph G. However, we claim that we never need to use any of the newly created special vertices in V . In particular, if some vertex wuv ∈ V , then modify V by replacing wuv with u. (We could have just as easily replaced it with v.) Observe that the vertex wuv is adjacent to only u and v, so it dominates itself and these other two vertices. By using u instead, we still dominate u, v, and wuv (because u has edges going to v and wuv ). Thus by replacing wu,v with u we dominate the same vertices (and potentially more). Let V denote the resulting set after this modification. (This is shown in the lower middle part of the figure.) We claim that V is a vertex cover for G. If, to the contrary there were an edge {u, v} of G that was not covered (neither u nor v was in V ) then the special vertex wuv would not be adjacent to any vertex of V in G , contradicting the hypothesis that V was a dominating set for G . Lecture Notes 72 CMSC 451
- 73. Lecture 20: SubsetSum Read: Sections 34.5.5 in CLR. Subset Sum: The Subset Sum problem (SS) is the following. Given a finite set S of positive integers S = {w1 , w2 , . . . , wn } and a target value, t, we want to know whether there exists a subset S ⊆ S that sums exactly to t. This problem is a simplified version of the 0-1 Knapsack problem, presented as a decision problem. Recall that in the 0-1 Knapsack problem, we are given a collection of objects, each with an associated weight wi and associated value vi . We are given a knapsack of capacity W . The objective is to take as many objects as can fit in the knapsack’s capacity so as to maximize the value. (In the fractional knapsack we could take a portion of an object. In the 0-1 Knapsack we either take an object entirely or leave it.) In the simplest version, suppose that the value is the same as the weight, vi = wi . (This would occur for example if all the objects were made of the same material, say, gold.) Then, the best we could hope to achieve would be to fill the knapsack entirely. By setting t = W , we see that the subset sum problem is equivalent to this simplified version of the 0-1 Knapsack problem. It follows that if we can show that this simpler version is NP-complete, then certainly the more general 0-1 Knapsack problem (stated as a decision problem) is also NP-complete. Consider the following example. S = {3, 6, 9, 12, 15, 23, 32} and t = 33. The subset S = {6, 12, 15} sums to t = 33, so the answer in this case is yes. If t = 34 the answer would be no. Dynamic Programming Solution: There is a dynamic programming algorithm which solves the Subset Sum prob- lem in O(n · t) time.2 The quantity n · t is a polynomial function of n. This would seem to imply that the Subset Sum problem is in P. But there is a important catch. Recall that in all NP-complete problems we assume (1) running time is measured as a function of input size (number of bits) and (2) inputs must be encoded in a reasonable succinct manner. Let us assume that the numbers wi and t are all b-bit numbers represented in base 2, using the fewest number of bits possible. Then the input size is O(nb). The value of t may be as large as 2b . So the resulting algorithm has a running time of O(n2b ). This is polynomial in n, but exponential in b. Thus, this running time is not polynomial as a function of the input size. Note that an important consequence of this observation is that the SS problem is not hard when the numbers involved are small. If the numbers involved are of a fixed number of bits (a constant independent of n), then the problem is solvable in polynomial time. However, we will show that in the general case, this problem is NP-complete. SS is NP-complete: The proof that Subset Sum (SS) is NP-complete involves the usual two elements. (i) SS ∈ NP. (ii) Some known NP-complete problem is reducible to SS. In particular, we will show that Vertex Cover (VC) is reducible to SS, that is, VC ≤P SS. To show that SS is in NP, we need to give a verification procedure. Given S and t, the certificate is just the indices of the numbers that form the subset S . We can add two b-bit numbers together in O(b) time. So, in polynomial time we can compute the sum of elements in S , and verify that this sum equals t. For the remainder of the proof we show how to reduce vertex cover to subset sum. We want a polynomial time computable function f that maps an instance of the vertex cover (a graph G and integer k) to an instance of the subset sum problem (a set of integers S and target integer t) such that G has a vertex cover of size k if and only if S has a subset summing to t. Thus, if subset sum were solvable in polynomial time, so would vertex cover. 2 We will leave this as an exercise, but the formulation is, for 0 ≤ i ≤ n and 0 ≤ t ≤ t, S[i, t ] = 1 if there is a subset of {w , w , . . . , w } 1 2 i that sums to t , and 0 otherwise. The ith row of this table can be computed in O(t) time, given the contents of the (i − 1)-st row. Lecture Notes 73 CMSC 451
- 74. How can weencode the notion of selecting a subset of vertices that cover all the edges to that of selecting a subset of numbers that sums to t? In the vertex cover problem we are selecting vertices, and in the subset sum problem we are selecting numbers, so it seems logical that the reduction should map vertices into numbers. The constraint that these vertices should cover all the edges must be mapped to the constraint that the sum of the numbers should equal the target value. An Initial Approach: Here is an idea, which does not work, but gives a sense of how to proceed. Let E denote the number of edges in the graph. First number the edges of the graph from 1 through E. Then represent each vertex vi as an E-element bit vector, where the j-th bit from the left is set to 1 if and only if the edge ej is incident to vertex vi . (Another way to think of this is that these bit vectors form the rows of an incidence matrix for the graph.) An example is shown below, in which k = 3. e1 e2 e3 e4 e5 e6 e7 e8 v7 v6 v1 1 0 0 0 0 0 0 0 v2 1 1 1 0 0 0 0 0 e6 v3 0 0 1 1 0 0 0 1 v1 e5 v5 v4 0 0 0 0 1 1 1 1 e1 v5 0 0 0 0 0 0 1 0 e7 e2 e4 v6 0 1 0 0 0 1 0 0 v2 v4 v7 0 0 0 1 1 0 0 0 e3 e8 v3 Fig. 54: Encoding a graph as a collection of bit vectors. Now, suppose we take any subset of vertices and form the logical-or of the corresponding bit vectors. If the subset is a vertex cover, then every edge will be covered by at least one of these vertices, and so the logical-or will be a bit vector of all 1’s, 1111 . . . 1. Conversely, if the logical-or is a bit vector of 1’s, then each edge has been covered by some vertex, implying that the vertices form a vertex cover. (Later we will consider how to encode the fact that there only allowed k vertices in the cover.) e1 e2 e3 e4 e5 e6 e7 e8 v7 v6 v1 1 0 0 0 0 0 0 0 v2 1 1 1 0 0 0 0 0 e6 v3 0 0 1 1 0 0 0 1 v1 e5 v5 v4 0 0 0 0 1 1 1 1 e1 v5 0 0 0 0 0 0 1 0 e7 e2 e4 v6 0 1 0 0 0 1 0 0 v2 v4 v7 0 0 0 1 1 0 0 0 e3 e8 t= 1 1 1 1 1 1 1 1 v2 v v3 v v4 v3 Fig. 55: The logical-or of a vertex cover equals 1111 . . . 1. Since bit vectors can be thought of as just a way of representing numbers in binary, this is starting to feel more like the subset sum problem. The target would be the number whose bit vector is all 1’s. There are a number of problems, however. First, logical-or is not the same as addition. For example, if both of the endpoints of some edge are in the vertex cover, then its value in the corresponding column would be 2, not 1. Second, we have no way of controlling how many vertices go into the vertex cover. (We could just take the logical-or of all the vertices, and then the logical-or would certainly be a bit vectors of 1’s.) Lecture Notes 74 CMSC 451
- 75. There are twoways in which addition differs significantly from logical-or. The first is the issue of carries. For example, the 1101 ∨ 0011 = 1111, but in binary 1101 + 0011 = 1000. To fix this, we recognize that we do not have to use a binary (base-2) representation. In fact, we can assume any base system we want. Observe that each column of the incidence matrix has at most two 1’s in any column, because each edge is incident to at most two vertices. Thus, if use any base that is at least as large as base 3, we will never generate a carry to the next position. In fact we will use base 4 (for reasons to be seen below). Note that the base of the number system is just for own convenience of notation. Once the numbers have been formed, they will be converted into whatever form our machine assumes for its input representation, e.g. decimal or binary. The second difference between logical-or and addition is that an edge may generally be covered either once or twice in the vertex cover. So, the final sum of these numbers will be a number consisting of 1 and 2 digits, e.g. 1211 . . . 112. This does not provide us with a unique target value t. We know that no digit of our sum can be a zero. To fix this problem, we will create a set of E additional slack values. For 1 ≤ i ≤ E, the ith slack value will consist of all 0’s, except for a single 1-digit in the ith position, e.g., 00000100000. Our target will be the number 2222 . . . 222 (all 2’s). To see why this works, observe that from the numbers of our vertex cover, we will get a sum consisting of 1’s and 2’s. For each position where there is a 1, we can supplement this value by adding in the corresponding slack value. Thus we can boost any value consisting of 1’s and 2’s to all 2’s. On the other hand, note that if there are any 0 values in the final sum, we will not have enough slack values to convert this into a 2. There is one last issue. We are only allowed to place only k vertices in the vertex cover. We will handle this by adding an additional column. For each number arising from a vertex, we will put a 1 in this additional column. For each slack variable we will put a 0. In the target, we will require that this column sum to the value k, the size of the vertex cover. Thus, to form the desired sum, we must select exactly k of the vertex values. Note that since we only have a base-4 representation, there might be carries out of this last column (if k ≥ 4). But since this is the last column, it will not affect any of the other aspects of the construction. The Final Reduction: Here is the final reduction, given the graph G = (V, E) and integer k for the vertex cover problem. (1) Create a set of n vertex values, x1 , x2 , . . . , xn using base-4 notation. The value xi is equal a 1 followed by a sequence of E base-4 digits. The j-th digit is a 1 if edge ej is incident to vertex vi and 0 otherwise. (2) Create E slack values y1 , y2 , . . . , yE , where yi is a 0 followed by E base-4 digits. The i-th digit of yi is 1 and all others are 0. (3) Let t be the base-4 number whose first digit is k (this may actually span multiple base-4 digits), and whose remaining E digits are all 2. (4) Convert the xi ’s, the yj ’s, and t into whatever base notation is used for the subset sum problem (e.g. base 10). Output the set S = {x1 , . . . , xn , y1 , . . . , yE } and t. Observe that this can be done in polynomial time, in O(E 2 ), in fact. The construction is illustrated in Fig. 56. Correctness: We claim that G has a vertex cover of size k if and only if S has a subset that sums to t. If G has a vertex cover V of size k, then we take the vertex values xi corresponding to the vertices of V , and for each edge that is covered only once in V , we take the corresponding slack variable. It follows from the comments made earlier that the lower-order E digits of the resulting sum will be of the form 222 . . . 2 and because there are k elements in V , the leftmost digit of the sum will be k. Thus, the resulting subset sums to t. Conversely, if S has a subset S that sums to t then we assert that it must select exactly k values from among the vertex values, since the first digit must sum to k. We claim that these vertices V form a vertex cover. In particular, no edge can be left uncovered by V , since (because there are no carries) the corresponding column would be 0 in the sum of vertex values. Thus, no matter what slack values we add, the resulting digit position could not be equal to 2, and so this cannot be a solution to the subset sum problem. Lecture Notes 75 CMSC 451
- 76. e1 e2 e3e4 e5 e6 e7 e8 x1 1 1 0 0 0 0 0 0 0 x2 1 1 1 1 0 0 0 0 0 x3 1 0 0 1 1 0 0 0 1 x4 1 0 0 0 0 1 1 1 1 Vertex values x5 1 0 0 0 0 0 0 1 0 x6 1 0 1 0 0 0 1 0 0 x7 1 0 0 0 1 1 0 0 0 y1 0 1 0 0 0 0 0 0 0 y2 0 0 1 0 0 0 0 0 0 y3 0 0 0 1 0 0 0 0 0 y4 0 0 0 0 1 0 0 0 0 y5 Slack values 0 0 0 0 0 1 0 0 0 y6 0 0 0 0 0 0 1 0 0 y7 0 0 0 0 0 0 0 1 0 y8 0 0 0 0 0 0 0 0 1 t 3 2 2 2 2 2 2 2 2 vertex cover size (k=3) Fig. 56: Vertex cover to subset sum reduction. e1 e2 e3 e4 e5 e6 e7 e8 x1 1 1 0 0 0 0 0 0 0 x2 1 1 1 1 0 0 0 0 0 x3 1 0 0 1 1 0 0 0 1 x4 1 0 0 0 0 1 1 1 1 Vertex values x5 1 0 0 0 0 0 0 1 0 (take those in vertex cover) x6 1 0 1 0 0 0 1 0 0 x7 1 0 0 0 1 1 0 0 0 y1 0 1 0 0 0 0 0 0 0 y2 0 0 1 0 0 0 0 0 0 y3 0 0 0 1 0 0 0 0 0 y4 0 0 0 0 1 0 0 0 0 y5 Slack values 0 0 0 0 0 1 0 0 0 y6 (take one for each edge that has 0 0 0 0 0 0 1 0 0 y7 only one endpoint in the cover) 0 0 0 0 0 0 0 1 0 y8 0 0 0 0 0 0 0 0 1 t 3 2 2 2 2 2 2 2 2 vertex cover size Fig. 57: Correctness of the reduction. Lecture Notes 76 CMSC 451
- 77. It is worthnoting again that in this reduction, we needed to have large numbers. For example, the target value t is at least as large as 4E ≥ 4n (where n is the number of vertices in G). In our dynamic programming solution W = t, so the DP algorithm would run in Ω(n4n ) time, which is not polynomial time. Lecture 21: Approximation Algorithms: VC and TSP Read: Chapt 35 (up through 35.2) in CLRS. Coping with NP-completeness: With NP-completeness we have seen that there are many important optimization problems that are likely to be quite hard to solve exactly. Since these are important problems, we cannot simply give up at this point, since people do need solutions to these problems. How do we cope with NP-completeness: Use brute-force search: Even on the fastest parallel computers this approach is viable only for the smallest instances of these problems. Heuristics: A heuristic is a strategy for producing a valid solution, but there are no guarantees how close it is to optimal. This is worthwhile if all else fails, or if lack of optimality is not really an issue. General Search Methods: There are a number of very powerful techniques for solving general combinatorial optimization problems that have been developed in the areas of AI and operations research. These go under names such as branch-and-bound, A∗ -search, simulated annealing, and genetic algorithms. The perfor- mance of these approaches varies considerably from one problem to problem and instance to instance. But in some cases they can perform quite well. Approximation Algorithms: This is an algorithm that runs in polynomial time (ideally), and produces a solu- tion that is within a guaranteed factor of the optimum solution. Performance Bounds: Most NP-complete problems have been stated as decision problems for theoretical reasons. However underlying most of these problems is a natural optimization problem. For example, the TSP optimiza- tion problem is to find the simple cycle of minimum cost in a digraph, the VC optimization problem is to find the vertex cover of minimum size, the clique optimization problem is to find the clique of maximum size. Note that sometimes we are minimizing and sometimes we are maximizing. An approximation algorithm is one that returns a legitimate answer, but not necessarily one of the smallest size. How do we measure how good an approximation algorithm is? We define the ratio bound of an approximation algorithm as follows. Given an instance I of our problem, let C(I) be the cost of the solution produced by our approximation algorithm, and let C ∗ (I) be the optimal solution. We will assume that costs are strictly positive values. For a minimization problem we want C(I)/C ∗ (I) to be small, and for a maximization problem we want C ∗ (I)/C(I) to be small. For any input size n, we say that the approximation algorithm achieves ratio bound ρ(n), if for all I, |I| = n we have C(I) C ∗ (I) max , ≤ ρ(n). I C ∗ (I) C(I) Observe that ρ(n) is always greater than or equal to 1, and it is equal to 1 if and only if the approximate solution is the true optimum solution. Some NP-complete problems can be approximated arbitrarily closely. Such an algorithm is given both the input, and a real value > 0, and returns an answer whose ratio bound is at most (1 + ). Such an algorithm is called a polynomial time approximation scheme (or PTAS for short). The running time is a function of both n and . As approaches 0, the running time increases beyond polynomial time. For example, the running time might be O(n 1/ ). If the running time depends only on a polynomial function of 1/ then it is called a fully polynomial- time approximation scheme. For example, a running time like O((1/ )2 n3 ) would be such an example, whereas O(n1/ ) and O(2(1/ ) n) are not. Although NP-complete problems are equivalent with respect to whether they can be solved exactly in polynomial time in the worst case, their approximability varies considerably. Lecture Notes 77 CMSC 451
- 78. • For someNP-complete problems, it is very unlikely that any approximation algorithm exists. For example, if the graph TSP problem had an approximation algorithm with a ratio bound of any value less than ∞, then P = NP. • Many NP-complete can be approximated, but the ratio bound is a (slow growing) function of n. For example, the set cover problem (a generalization of the vertex cover problem), can be approximated to within a factor of ln n. We will not discuss this algorithm, but it is covered in CLRS. • Some NP-complete problems can be approximated to within a fixed constant factor. We will discuss two examples below. • Some NP-complete problems have PTAS’s. One example is the subset problem (which we haven’t dis- cussed, but is described in CLRS) and the Euclidean TSP problem. In fact, much like NP-complete problems, there are collections of problems which are “believed” to be hard to approximate and are equivalent in the sense that if any one can be approximated in polynomial time then they all can be. This class is called Max-SNP complete. We will not discuss this further. Suffice it to say that the topic of approximation algorithms would fill another course. Vertex Cover: We begin by showing that there is an approximation algorithm for vertex cover with a ratio bound of 2, that is, this algorithm will be guaranteed to find a vertex cover whose size is at most twice that of the optimum. Recall that a vertex cover is a subset of vertices such that every edge in the graph is incident to at least one of these vertices. The vertex cover optimization problem is to find a vertex cover of minimum size. How does one go about finding an approximation algorithm. The first approach is to try something that seems like a “reasonably” good strategy, a heuristic. It turns out that many simple heuristics, when not optimal, can often be proved to be close to optimal. Here is an very simple algorithm, that guarantees an approximation within a factor of 2 for the vertex cover problem. It is based on the following observation. Consider an arbitrary edge (u, v) in the graph. One of its two vertices must be in the cover, but we do not know which one. The idea of this heuristic is to simply put both vertices into the vertex cover. (You cannot get much stupider than this!) Then we remove all edges that are incident to u and v (since they are now all covered), and recurse on the remaining edges. For every one vertex that must be in the cover, we put two into our cover, so it is easy to see that the cover we generate is at most twice the size of the optimum cover. The approximation is given in the figure below. Here is a more formal proof of its approximation bound. G and opt VC The 2−for−1 Heuristic Fig. 58: The 2-for-1 heuristic for vertex cover. Claim: ApproxVC yields a factor-2 approximation for Vertex Cover. Proof: Consider the set C output by ApproxVC. Let C ∗ be the optimum VC. Let A be the set of edges selected by the line marked with “(*)” in the figure. Observe that the size of C is exactly 2|A| because we add two vertices for each such edge. However note that in the optimum VC one of these two vertices must have been added to the VC, and thus the size of C ∗ is at least |A|. Thus we have: |C| |C| = |A| ≤ |C ∗ | ⇒ ≤ 2. 2 |C ∗ | Lecture Notes 78 CMSC 451
- 79. 2-for-1 Approximation forVC ApproxVC { C = empty-set while (E is nonempty) do { (*) let (u,v) be any edge of E add both u and v to C remove from E all edges incident to either u or v } return C; } This proof illustrates one of the main features of the analysis of any approximation algorithm. Namely, that we need some way of finding a bound on the optimal solution. (For minimization problems we want a lower bound, for maximization problems an upper bound.) The bound should be related to something that we can compute in polynomial time. In this case, the bound is related to the set of edges A, which form a maximal independent set of edges. The Greedy Heuristic: It seems that there is a very simple way to improve the 2-for-1 heuristic. This algorithm simply selects any edge, and adds both vertices to the cover. Instead, why not concentrate instead on vertices of high degree, since a vertex of high degree covers the maximum number of edges. This is greedy strategy. We saw in the minimum spanning tree and shortest path problems that greedy strategies were optimal. Here is the greedy heuristic. Select the vertex with the maximum degree. Put this vertex in the cover. Then delete all the edges that are incident to this vertex (since they have been covered). Repeat the algorithm on the remaining graph, until no more edges remain. This algorithm is illustrated in the figure below. G and opt VC The Greedy Heuristic Fig. 59: The greedy heuristic for vertex cover. Greedy Approximation for VC GreedyVC(G=(V,E)) { C = empty-set; while (E is nonempty) do { let u be the vertex of maximum degree in G; add u to C; remove from E all edges incident to u; } return C; } It is interesting to note that on the example shown in the figure, the greedy heuristic actually succeeds in find- ing the optimum vertex cover. Can we prove that the greedy heuristic always outperforms the stupid 2-for-1 heuristic? The surprising answer is an emphatic “no”. In fact, it can be shown that the greedy heuristic does not even have a constant performance bound. That is, it can perform arbitrarily poorly compared to the optimal algorithm. It can be shown that the ratio bound grows as Θ(log n), where n is the number of vertices. (We leave Lecture Notes 79 CMSC 451
- 80. this as amoderately difficult exercise.) However, it should also be pointed out that the vertex cover constructed by the greedy heuristic is (for typical graphs) smaller than that one computed by the 2-for-1 heuristic, so it would probably be wise to run both algorithms and take the better of the two. Traveling Salesman Problem: In the Traveling Salesperson Problem (TSP) we are given a complete undirected graph with nonnegative edge weights, and we want to find a cycle that visits all vertices and is of minimum cost. Let c(u, v) denote the weight on edge (u, v). Given a set of edges A forming a tour we define c(A) to be the sum of edge weights in A. Last time we mentioned that TSP (posed as a decision problem) is NP-complete. For many of the applications of TSP, the problem satisfies something called the triangle inequality. Intuitively, this says that the direct path from u to w, is never longer than an indirect path. More formally, for all u, v, w ∈ V c(u, w) ≤ c(u, v) + c(v, w). There are many examples of graphs that satisfy the triangle inequality. For example, given any weighted graph, if we define c(u, v) to be the shortest path length between u and v (computed, say by the Floyd-Warshall algorithm), then it will satisfy the triangle inequality. Another example is if we are given a set of points in the plane, and define a complete graph on these points, where c(u, v) is defined to be the Euclidean distance between these points, then the triangle inequality is also satisfied. When the underlying cost function satisfies the triangle inequality there is an approximation algorithm for TSP with a ratio-bound of 2. (In fact, there is a slightly more complex version of this algorithm that has a ratio bound of 1.5, but we will not discuss it.) Thus, although this algorithm does not produce an optimal tour, the tour that it produces cannot be worse than twice the cost of the optimal tour. The key insight is to observe that a TSP with one edge removed is a spanning tree. However it is not necessarily a minimum spanning tree. Therefore, the cost of the minimum TSP tour is at least as large as the cost of the MST. We can compute MST’s efficiently, using, for example, either Kruskal’s or Prim’s algorithm. If we can find some way to convert the MST into a TSP tour while increasing its cost by at most a constant factor, then we will have an approximation for TSP. We shall see that if the edge weights satisfy the triangle inequality, then this is possible. Here is how the algorithm works. Given any free tree there is a tour of the tree called a twice around tour that traverses the edges of the tree twice, once in each direction. The figure below shows an example of this. start start MST Twice−around tour Shortcut tour Optimum tour Fig. 60: TSP Approximation. This path is not simple because it revisits vertices, but we can make it simple by short-cutting, that is, we skip over previously visited vertices. Notice that the final order in which vertices are visited using the short-cuts is exactly the same as a preorder traversal of the MST. (In fact, any subsequence of the twice-around tour which visits each vertex exactly once will suffice.) The triangle inequality assures us that the path length will not increase when we take short-cuts. Claim: Approx-TSP has a ratio bound of 2. Proof: Let H denote the tour produced by this algorithm and let H ∗ be the optimum tour. Let T be the minimum spanning tree. As we said before, since we can remove any edge of H ∗ resulting in a spanning tree, and Lecture Notes 80 CMSC 451
- 81. TSP Approximation ApproxTSP(G=(V,E)) { T = minimum spanning tree for G r = any vertex L = list of vertices visited by a preorder walk ot T starting with r return L } since T is the minimum cost spanning tree we have c(T ) ≤ c(H ∗ ). Now observe that the twice around tour of T has cost 2c(T ), since every edge in T is hit twice. By the triangle inequality, when we short-cut an edge of T to form H we do not increase the cost of the tour, and so we have c(H) ≤ 2c(T ). Combining these we have c(H) c(H) ≤ c(T ) ≤ c(H ∗ ) ⇒ ≤ 2. 2 c(H ∗ ) Lecture 22: The k-Center Approximation Read: Today’s material is not covered in CLR. Facility Location: Imagine that Blockbuster Video wants to open a 50 stores in some city. The company asks you to determine the best locations for these stores. The condition is that you are to minimize the maximum distance that any resident of the city must drive in order to arrive at the nearest store. If we model the road network of the city as an undirected graph whose edge weights are the distances between intersections, then this is an instance of the k-center problem. In the k-center problem we are given an undirected graph G = (V, E) with nonnegative edge weights, and we are given an integer k. The problem is to compute a subset of k vertices C ⊆ V , called centers, such that the maximum distance between any vertex in V and its nearest center in C is minimized. (The optimization problem seeks to minimize the maximum distance and the decision problem just asks whether there exists a set of centers that are within a given distance.) More formally, let G = (V, E) denote the graph, and let w(u, v) denote the weight of edge (u, v). (w(u, v) = w(v, u) because G is undirected.) We assume that all edge weights are nonnegative. For each pair of vertices, u, v ∈ V , let d(u, v) = d(u, v) denote the distance between u to v, that is, the length of the shortest path from u to v. (Note that the shortest path distance satisfies the triangle inequality. This will be used in our proof.) Consider a subset C ⊆ V of vertices, the centers. For each vertex v ∈ V we can associate it with its nearest center in C. (This is the nearest Blockbuster store to your house). For each center ci ∈ C we define its neighborhood to be the subset of vertices for which ci is the closest center. (These are the houses that are closest to this center. See Fig. 61.) More formally, define: V (ci ) = {v ∈ V | d(v, ci ) ≤ d(v, cj ), for i = j}. Let us assume for simplicity that there are no ties for the distances to the closest center (or that any such ties have been broken arbitrarily). Then V (c1 ), V (c2 ), . . . , V (ck ) forms a partition of the vertex set of G. The bottleneck distance associated with each center is the distance to its farthest vertex in V (ci ), that is, D(ci ) = max d(v, ci ). v∈V (ci ) Lecture Notes 81 CMSC 451
- 82. V(c2) 5 6 5 6 5 c2 4 9 7 4 9 7 4 8 9 8 9 5 7 c3 V(c3) 5 6 5 5 6 5 V(c1) c1 6 5 6 8 6 8 Input graph (k=3) Optimumum Cost = 7 Fig. 61: The k-center problem with optimum centers ci and neighborhood sets V (ci ). Finally, we define the overall bottleneck distance to be D(C) = max D(ci ). ci ∈C This is the maximum distance of any vertex from its nearest center. This distance is critical because it represents the customer that must travel farthest to get to the nearest facility, the bottleneck vertex. Given this notation, we can now formally define the problem. k-center problem: Given a weighted undirected graph G = (V, E), and an integer k ≤ |V |, find a subset C ⊆ V of size k such that D(C) is minimized. The decision-problem formulation of the k-center problem is NP-complete (reduction from dominating set). A brute force solution to this problem would involve enumerating all k-element of subsets of V , and computing D(C) for each one. However, letting n = |V | and k, the number of possible subsets is n = Θ(nk ). If k is k a function of n (which is reasonable), then this an exponential number of subsets. Given that the problem is NP-complete, it is highly unlikely that a significantly more efficient exact algorithm exists in the worst-case. We will show that there does exist an efficient approximation algorithm for the problem. Greedy Approximation Algorithm: Our approximation algorithm is based on a simple greedy algorithm that pro- duces a bottleneck distance D(C) that is not more than twice the optimum bottleneck distance. We begin by letting the first center c1 be any vertex in the graph (the lower left vertex, say, in the figure below). Compute the distances between this vertex and all the other vertices in the graph (Fig. 62(b)). Consider the vertex that is farthest from this center (the upper right vertex at distance 23 in the figure). This the bottleneck vertex for {c1 }. We would like to select the next center so as to reduce this distance. So let us just make it the next center, called c2 . Then again we compute the distances from each vertex in the graph to the closer of c1 and c2 . (See Fig. 62(c) where dashed lines indicate which vertices are closer to which center). Again we consider the bottleneck vertex for the current centers {c1 , c2 }. We place the next center at this vertex (see Fig. 62(d)). Again we compute the distances from each vertex to its nearest center. Repeat this until all k centers have been selected. In Fig. 62(d), the final three greedy centers are shaded, and the final bottleneck distance is 11. Although the greedy approach has a certain intuitive appeal (because it attempts to find the vertex that gives the bottleneck distance, and then puts a center right on this vertex), it is not optimal. In the example shown in the figure, the optimum solution (shown on the right) has a bottleneck cost of 9, which beats the 11 that the greedy algorithm gave. Here is a summary of the algorithm. For each vertex u, let d[u] denote the distance to the nearest center. We know from Dijkstra’s algorithm how to compute the shortest path from a single source to all other vertices in the graph. One way to solve the distance computation step above would be to invoke Dijkstra’s algorithm i times. But there is an easier way. We can modify Dijkstra’s algorithm to operate as a multiple source algorithm. In particular, in the initialization of Dijkstra’s single source algorithm, it sets d[s] = 0 and pred[s] = null. In Lecture Notes 82 CMSC 451
- 83. 5 6 5 6 5 6 5 6 14 19 23 11 6 c2 11 6 c2 9 7 4 9 7 4 9 7 4 9 7 4 8 9 8 9 8 9 8 9 5 12 19 5 12 4 5 c3 4 5 6 5 5 6 5 5 6 5 5 6 5 c1 6 14 c1 6 9 c1 6 9 6 8 6 8 6 8 6 8 (a) (b) (c) (d) Greedy Cost = 11 Fig. 62: Greedy approximation to k-center. Greedy Approximation for k-center KCenterApprox(G, k) { C = empty_set for each u in V do // initialize distances d[u] = INFINITY for i = 1 to k do { // main loop Find the vertex u such that d[u] is maximum Add u to C // u is the current bottleneck vertex // update distances Compute the distance from each vertex v to its closest vertex in C, denoted d[v] } return C // final centers } Lecture Notes 83 CMSC 451
- 84. the modified multiplesource version, we do this for all the vertices of C. The final greedy algorithm involves running Dijkstra’s algorithm k times (once for each time through the for-loop). Recall that the running time of Dijkstra’s algorithm is O((V + E) log V ). Under the reasonable assumption that E ≥ V , this is O(E log V ). Thus, the overall running time is O(kE log V ). Approximation Bound: How bad could greedy be? We will argue that it has a ratio bound of 2. To see that we can get a factor of 2, consider a set of n + 1 vertices arranged in a linear graph, in which all edges are of weight 1. The greedy algorithm might pick any initial vertex that it likes. Suppose it picks the leftmost vertex. Then the maximum (bottleneck) distance is the distance to the rightmost vertex which is n. If we had instead chosen the vertex in the middle, then the maximum distance would only be n/2, which is better by a factor of 2. Greedy Cost = n Opt Cost =n/2 Fig. 63: Worst-case for greedy. We want to show that this approximation algorithm always produces a final distance D(C) that is within a factor of 2 of the distance of the optimal solution. Let O = {o1 , o2 , . . . , ok } denote the centers of the optimal solution (shown as black dots in Fig. 64, and the lines show the partition into the neighborhoods for each of these points). Let D∗ = D(O) be the optimal bottleneck distance. Let G = {g1 , g2 , . . . , gk } be the centers found by the greedy approximation (shown as white dots in the figure below). Also, let gk+1 denote the next center that would have been added next, that is, the bottleneck vertex for G. Let D(G) denote the bottleneck distance for G. Notice that the distance from gk+1 to its nearest center is equal D(G). The proof involves a simple application of the pigeon-hole principal. g3 o1 o4 g1 o3 g5 g2 g6 <Dopt >D* o2 o5 <Dopt g4 Fig. 64: Analysis of the greedy heuristic for k = 5. The greedy centers are given as white dots and the optimal centers as black dots. The regions represent the neighborhood sets V (oi ) for the optimal centers. Theorem: The greedy approximation has a ratio bound of 2, that is D(G)/D∗ ≤ 2. Proof: Let G = {g1 , g2 , . . . , gk , gk+1 } be the (k + 1)-element set consisting of the greedy centers together with the next greedy center gk+1 First observe that for i = j, d(gi , gj ) ≥ D(G). This follows as a result of our greedy selection strategy. As each center is selected, it is selected to be at the maximum (bottleneck) distance from all the previous centers. As we add more centers, the maximum distance between any pair of centers decreases. Since the final bottleneck distance is D(G), all the centers are at least this far apart from one another. Lecture Notes 84 CMSC 451
- 85. Each gi ∈G is associated with its closest center in the optimal solution, that is, each belongs to V (om ) for some m. Because there are k centers in O, and k + 1 elements in G , it follows from the pigeon-hole principal, that at least two centers of G are in the same set V (om ) for some m. (In the figure, the greedy centers g4 and g5 are both in V (o2 )). Let these be denoted gi and gj . Since D∗ is the bottleneck distance for O, we know that the distance from gi to ok is of length at most D∗ and similarly the distance from ok to gj is at most D∗ . By concatenating these two paths, it follows that there exists a path of length 2D∗ from gi to gj , and hence we have d(gi , gj ) ≤ 2D∗ . But from the comments above we have d(gi , gj ) ≥ D(G). Therefore, D(G) ≤ d(gi , gj ) ≤ 2D∗ , from which the desired ratio follows. Lecture 23: Approximations: Set Cover and Bin Packing Read: Set cover is covered in Chapt 35.3. Bin packing is covered as an exercise in CLRS. Set Cover: The set cover problem is a very important optimization problem. You are given a pair (X, F ) where X = {x1 , x2 , . . . , xm } is a finite set (a domain of elements) and F = {S1 , S2 , . . . , Sn } is a family of subsets of X, such that every element of X belongs to at least one set of F . Consider a subset C ⊆ F . (This is a collection of sets over X.) We say that C covers the domain if every element of X is in some set of C, that is X= Si . Si ∈C The problem is to find the minimum-sized subset C of F that covers X. Consider the example shown below. The optimum set cover consists of the three sets {S3 , S4 , S5 }. S3 S4 S5 S1 S2 S6 Fig. 65: Set cover. Set cover can be applied to a number of applications. For example, suppose you want to set up security cameras to cover a large art gallery. From each possible camera position, you can see a certain subset of the paintings. Each such subset of paintings is a set in your system. You want to put up the fewest cameras to see all the paintings. Complexity of Set Cover: We have seen special cases of the set cover problems that are NP-complete. For example, vertex cover is a type of set cover problem. The domain to be covered are the edges, and each vertex covers the subset of incident edges. Thus, the decision-problem formulation of set cover (“does there exist a set cover of size at most k?”) is NP-complete as well. There is a factor-2 approximation for the vertex cover problem, but it cannot be applied to generate a factor- 2 approximation for set cover. In particular, the VC approximation relies on the fact that each element of the domain (an edge) is in exactly 2 sets (one for each of its endpoints). Unfortunately, this is not true for the general Lecture Notes 85 CMSC 451
- 86. set cover problem.In fact, it is known that there is no constant factor approximation to the set cover problem, unless P = NP. This is unfortunate, because set cover is one of the most powerful NP-complete problems. Today we will show that there is a reasonable approximation algorithm, the greedy heuristic, which achieves an approximation bound of ln m, where m = |X|, the size of the underlying domain. (The book proves a somewhat stronger result, that the approximation factor of ln m where m ≤ m is the size of the largest set in F . However, their proof is more complicated.) Greedy Set Cover: A simple greedy approach to set cover works by at each stage selecting the set that covers the greatest number of “uncovered” elements. Greedy Set Cover Greedy-Set-Cover(X, F) { U = X // U are the items to be covered C = empty // C will be the sets in the cover while (U is nonempty) { // there is someone left to cover select S in F that covers the most elements of U add S to C U = U - S } return C } For the example given earlier the greedy-set cover algorithm would select S1 (since it covers 6 out of 12 ele- ments), then S6 (since it covers 3 out of the remaining 6), then S2 (since it covers 2 of the remaining 3) and finally S3 . Thus, it would return a set cover of size 4, whereas the optimal set cover has size 3. What is the approximation factor? The problem with the greedy set cover algorithm is that it can be “fooled” into picking the wrong set, over and over again. Consider the following example. The optimal set cover consists of sets S5 and S6 , each of size 16. Initially all three sets S1 , S5 , and S6 have 16 elements. If ties are broken in the worst possible way, the greedy algorithm will first select sets S1 . We remove all the covered elements. Now S2 , S5 and S6 all cover 8 of the remaining elements. Again, if we choose poorly, S2 is chosen. The pattern repeats, choosing S3 (size 4), S4 (size 2) and finally S5 and S6 (each of size 1). Thus, the optimum cover consisted of two sets, but we picked roughly lg m, where m = |X|, for a ratio bound of (lg m)/2. (Recall the lg denotes logarithm base 2.) There were many cases where ties were broken badly here, but it is possible to redesign the example such that there are no ties and yet the algorithm has essentially the same ratio bound. S4S3 S2 S1 S5 Optimum: {S5, S6} S6 Greedy: {S1, S2, S3, S4, S5, S6} Fig. 66: An example in which the Greedy Set cover performs poorly. However we will show that the greedy set cover heuristic nevers performs worse than a factor of ln m. (Note that this is natural log, not base 2.) Before giving the proof, we need one important mathematical inequality. Lemma: For all c > 0, c 1 1 1− ≤ . c e where e is the base of the natural logarithm. Lecture Notes 86 CMSC 451
- 87. Proof: We usethe fact that for all x, 1 + x ≤ ex . (The two functions are equal when x = 0.) Now, if we substitute −1/c for x we have (1 − 1/c) ≤ e−1/c , and if we raise both sides to the cth power, we have the desired result. The theorem of the approximation bound for bin packing proven here is a bit weaker from the one in CLRS, but I think it is easier to understand. Theorem: Greedy set cover has the ratio bound of at most ln m where m = |X|. Proof: Let c denote the size of the optimum set cover, and let g denote the size of the greedy set cover minus 1. We will show that g/c ≤ ln m. (This is not quite what we wanted, but we are correct to within 1 set.) Initially, there are m0 = m elements left to be covered. We know that there is a cover of size c (the optimal cover) and therefore by the pigeonhole principle, there must be at least one set that covers at least m0 /c elements. (Since otherwise, if every set covered less than m0 /c elements, then no collection of c sets could cover all m0 elements.) Since the greedy algorithm selects the largest set, it will select a set that covers at least this many elements. The number of elements that remain to be covered is at most m1 = m0 − m0 /c = m0 (1 − 1/c). Applying the argument again, we know that we can cover these m1 elements with a cover of size c (the optimal cover), and hence there exists a subset that covers at least m1 /c elements, leaving at most m2 = m1 (1 − 1/c) = m0 (1 − 1/c)2 elements remaining. If we apply this argument g times, each time we succeed in covering at least a fraction of (1 − 1/c) of the remaining elements. Then the number of elements that remain is uncovered after g sets have been chosen by the greedy algorithm is at most mg = m0 (1 − 1/c)g . How long can this go on? Consider the largest value of g such that after removing all but the last set of the greedy cover, we still have some element remaining to be covered. Thus, we are interested in the largest value of g such that g 1 1≤m 1− . c We can rewrite this as c g/c 1 1≤m 1− . c By the inequality above we have g/c 1 1≤m . e Now, if we multiply by eg/c and take natural logs we get that g satisfies: g eg/c ≤ m ⇒ ≤ ln m. c This completes the proof. Even though the greedy set cover has this relatively bad ratio bound, it seems to perform reasonably well in practice. Thus, the example shown above in which the approximation bound is (lg m)/2 is not “typical” of set cover instances. Bin Packing: Bin packing is another well-known NP-complete problem, which is a variant of the knapsack problem. We are given a set of n objects, where si denotes the size of the ith object. It will simplify the presentation to assume that 0 < si < 1. We want to put these objects into a set of bins. Each bin can hold a subset of objects whose total size is at most 1. The problem is to partition the objects among the bins so as to use the fewest possible bins. (Note that if your bin size is not 1, then you can reduce the problem into this form by simply dividing all sizes by the size of the bin.) Lecture Notes 87 CMSC 451
- 88. Bin packing arisesin many applications. Many of these applications involve not only the size of the object but their geometric shape as well. For example, these include packing boxes into a truck, or cutting the maximum number of pieces of certain shapes out of a piece of sheet metal. However, even if we ignore the geometry, and just consider the sizes of the objects, the decision problem is still NP-complete. (The reduction is from the knapsack problem.) Here is a simple heuristic algorithm for the bin packing problem, called the first-fit heuristic. We start with an unlimited number of empty bins. We take each object in turn, and find the first bin that has space to hold this object. We put this object in this bin. The algorithm is illustrated in the figure below. We claim that first-fit uses at most twice as many bins as the optimum, that is, if the optimal solution uses b∗ bins, and first-fit uses bff bins, then bff ≤ 2. b∗ s6 s3 s5 s7 s2 s1 s4 Fig. 67: First-fit Heuristic. Theorem: The first-fit heuristic achieves a ratio bound of 2. Proof: Consider an instance {s1 , . . . , sn } of the bin packing problem. Let S = i si denote the sum of all the object sizes. Let b∗ denote the optimal number of bins, and bff denote the number of bins used by first-fit. First observe that b∗ ≥ S. This is true, since no bin can hold a total capacity of more than 1 unit, and even if we were to fill each bin exactly to its capacity, we would need at least S bins. (In fact, since the number of bins is an integer, we would need at least S bins.) Next, we claim that bff ≤ 2S. To see this, let ti denote the total size of the objects that first-fit puts into bin i. Consider bins i and i + 1 filled by first-fit. Assume that indexing is cyclical, so if i is the last index (i = bff ) then i + 1 = 1. We claim that ti + ti+1 ≥ 1. If not, then the contents of bins i and i + 1 could both be put into the same bin, and hence first-fit would never have started to fill the second bin, preferring to keep everything in the first bin. Thus we have: bff (ti + ti+1 ) ≥ bff . i=1 But this sum adds up all the elements twice, so it has a total value of 2S. Thus we have 2S ≥ bff . Combining this with the fact that b∗ ≥ S we have bff ≤ 2S ≤ 2b∗ , implying that bff /b∗ ≤ 2, as desired. There are in fact a number of other heuristics for bin packing. Another example is best fit, which attempts to put the object into the bin in which it fits most closely with the available space (assuming that there is sufficient available space). There is also a variant of first-fit, called first fit decreasing, in which the objects are first sorted in decreasing order of size. (This makes intuitive sense, because it is best to first load the big items, and then try to squeeze the smaller objects into the remaining space.) Lecture Notes 88 CMSC 451
- 89. A more carefulproof establishes that first fit has a approximation ratio that is a bit smaller than 2, and in fact 17/10 is possible. Best fit has a very similar bound. First fit decreasing has a significantly better bound of 11/9 = 1.222 . . .. Lecture 24: Final Review Overview: This semester we have discussed general approaches to algorithm design. The intent has been to investi- gate basic algorithm design paradigms: dynamic programming, greedy algorithms, depth-first search, etc. And to consider how these techniques can be applied on a number of well-defined problems. We have also discussed the class NP-completeness, of problems that believed to be very hard to solve, and finally some examples of approximation algorithms. How to use this information: In some sense, the algorithms you have learned here are rarely immediately applicable to your later work (unless you go on to be an algorithm designer) because real world problems are always messier than these simple abstract problems. However, there are some important lessons to take out of this class. Develop a clean mathematical model: Most real-world problems are messy. An important first step in solving any problem is to produce a simple and clean mathematical formulation. For example, this might involve describing the problem as an optimization problem on graphs, sets, or strings. If you cannot clearly describe what your algorithm is supposed to do, it is very difficult to know when you have succeeded. Create good rough designs: Before jumping in and starting coding, it is important to begin with a good rough design. If your rough design is based on a bad paradigm (e.g. exhaustive enumeration, when depth-first search could have been applied) then no amount of additional tuning and refining will save this bad design. Prove your algorithm correct: Many times you come up with an idea that seems promising, only to find out later (after a lot of coding and testing) that it does not work. Prove that your algorithm is correct before coding. Writing proofs is not always easy, but it may save you a few weeks of wasted programming time. If you cannot see why it is correct, chances are that it is not correct at all. Can it be improved?: Once you have a solution, try to come up with a better one. Is there some reason why a better algorithm does not exist? (That is, can you establish a lower bound?) If your solution is exponential time, then maybe your problem is NP-hard. Prototype to generate better designs: We have attempted to analyze algorithms from an asymptotic perspec- tive, which hides many of details of the running time (e.g. constant factors), but give a general perspective for separating good designs from bad ones. After you have isolated the good designs, then it is time to start prototyping and doing empirical tests to establish the real constant factors. A good profiling tool can tell you which subroutines are taking the most time, and those are the ones you should work on improving. Still too slow?: If your problem has an unacceptably high execution time, you might consider an approximation algorithm. The world is full of heuristics, both good and bad. You should develop a good heuristic, and if possible, prove a ratio bound for your algorithm. If you cannot prove a ratio bound, run many experiments to see how good the actual performance is. There is still much more to be learned about algorithm design, but we have covered a great deal of the basic material. One direction is to specialize in some particular area, e.g. string pattern matching, computational geometry, parallel algorithms, randomized algorithms, or approximation algorithms. It would be easy to devote an entire semester to any one of these topics. Another direction is to gain a better understanding of average-case analysis, which we have largely ignored. Still another direction might be to study numerical algorithms (as covered in a course on numerical analysis), or to consider general search strategies such as simulated annealing. Finally, an emerging area is the study of algorithm engineering, which considers how to design algorithms that are both efficient in a practical sense, as well as a theoretical sense. Lecture Notes 89 CMSC 451
- 90. Material for thefinal exam: Old Material: Know general results, but I will not ask too many detailed questions. Do not forget DFS and DP. You will likely an algorithm design problem that will involve one of these two techniques. All-Pairs Shortest paths: (Chapt 25.2.) Floyd-Warshall Algorithm: All-pairs shortest paths, arbitrary edge weights (no negative cost cycles). Running time O(V 3 ). NP-completeness: (Chapt 34.) Basic concepts: Decision problems, polynomial time, the class P, certificates and the class NP, polynomial time reductions. NP-completeness reductions: You are responsible for knowing the following reductions. • 3-coloring to clique cover. • 3SAT to Independent Set (IS). • Independent Set to Vertex Cover and Clique. • Vertex Cover to Dominating Set. • Vertex Cover to Subset Sum. It is also a good idea to understand all the reductions that were used in the homework solutions, since modifications of these will likely appear on the final. NP-complete reductions can be challenging. If you cannot see how to solve the problem, here are some suggestions for maximizing partial credit. All NP-complete proofs have a very specific form. Explain that you know the template, and try to fill in as many aspects as possible. Suppose that you want to prove that some problem B is NP-complete. • B ∈ NP. This almost always easy, so don’t blow it. This basically involves specifying the certificate. The certificate is almost always the thing that the problem is asking you to find. • For some known NP-complete problem A, A ≤P B. This means that you want to find a polynomial time function f that maps an instance of A to an instance of B. (Make sure to get the direction correct!) • Show the correctness of your reduction, by showing that x ∈ A if and only if f (x) ∈ B. First suppose that you have a solution to x and show how to map this to a solution for f (x). Then suppose that you have a solution to f (x) and show how to map this to a solution for x. If you cannot figure out what f is, at least tell me what you would like f to do. Explain which elements of problem A will likely map to which elements of problem B. Remember that you are trying to translate the elements of one problem into the common elements of the other problem. I try to make at least one reduction on the exam similar to one that you have seen before, so make sure that understand the ones that we have done either in class or on homework problems. Approximation Algorithms: (Chapt. 35, up through 35.2.) Vertex cover: Ratio bound of 2. TSP with triangle inequality: Ratio bound of 2. Set Cover: Ratio bound of ln m, where m = |X|. Bin packing: Ratio bound of 2. k-center: Ratio bound of 2. Many approximation algorithms are simple. (Most are based on simple greedy heuristics.) The key to proving many ratio bounds is first coming up with a lower bound on the optimal solution (e.g., TSPopt ≥ MST). Next, provide an upper bound on the cost of your heuristic relative to this same quantity (e.g., the shortcut twice-around tour for the MST is at most twice the MST cost). Lecture Notes 90 CMSC 451
- 91. Supplemental Lecture 1:Asymptotics Read: Chapters 2–3 in CLRS. Asymptotics: The formulas that are derived for the running times of program may often be quite complex. When designing algorithms, the main purpose of the analysis is to get a sense for the trend in the algorithm’s running time. (An exact analysis is probably best done by implementing the algorithm and measuring CPU seconds.) We would like a simple way of representing complex functions, which captures the essential growth rate properties. This is the purpose of asymptotics. Asymptotic analysis is based on two simplifying assumptions, which hold in most (but not all) cases. But it is important to understand these assumptions and the limitations of asymptotic analysis. Large input sizes: We are most interested in how the running time grows for large values of n. Ignore constant factors: The actual running time of the program depends on various constant factors in the im- plementation (coding tricks, optimizations in compilation, speed of the underlying hardware, etc). There- fore, we will ignore constant factors. The justification for considering large n is that if n is small, then almost any algorithm is fast enough. People are most concerned about running times for large inputs. For the most part, these assumptions are reasonable when making comparisons between functions that have significantly different behaviors. For example, suppose we have two programs, one whose running time is T1 (n) = n3 and another whose running time is T2 (n) = 100n. (The latter algorithm may be faster because it uses a more sophisticated and complex algorithm, and the added sophistication results in a larger constant factor.) For small n (e.g., n ≤ 10) the first algorithm is the faster of the two. But as n becomes larger the relative differences in running time become much greater. Assuming one million operations per second. n T1 (n) T2 (n) T1 (n)/T2 (n) 10 0.001 sec 0.001 sec 1 100 1 sec 0.01 sec 100 1000 17 min 0.1 sec 10,000 10,000 11.6 days 1 sec 1,000,000 The clear lesson is that as input sizes grow, the performance of the asymptotically poorer algorithm degrades much more rapidly. These assumptions are not always reasonable. For example, in any particular application, n is a fixed value. It may be the case that one function is smaller than another asymptotically, but for your value of n, the asymptot- ically larger value is fine. Most of the algorithms that we will study this semester will have both low constants and low asymptotic running times, so we will not need to worry about these issues. Asymptotic Notation: To represent the running times of algorithms in a simpler form, we use asymptotic notation, which essentially represents a function by its fastest growing term and ignores constant factors. For example, suppose we have an algorithm whose (exact) worst-case running time is given by the following formula: T (n) = 13n3 + 5n2 − 17n + 16. As n becomes large, the 13n3 term dominates the others. By ignoring constant factors, we might say that the running time grows “on the order of” n3 , which will will express mathematically as T (n) ∈ Θ(n3 ). This intuitive definition is fine for informal use. Let us consider how to make this idea mathematically formal. Definition: Given any function g(n), we define Θ(g(n)) to be a set of functions: Θ(g(n)) = {f (n) | there exist strictly positive constants c1 , c2 , and n0 such that 0 ≤ c1 g(n) ≤ f (n) ≤ c2 g(n) for all n ≥ n0 }. Lecture Notes 91 CMSC 451
- 92. Let’s dissect thisdefinition. Intuitively, what we want to say with “f (n) ∈ Θ(g(n))” is that f (n) and g(n) are asymptotically equivalent. This means that they have essentially the same growth rates for large n. For example, functions such as √ 4n2 , (8n2 + 2n − 3), (n2 /5 + n − 10 log n), and n(n − 3) are all intuitively asymptotically equivalent, since as n becomes large, the dominant (fastest growing) term is some constant times n2 . In other words, they all grow quadratically in n. The portion of the definition that allows us to select c1 and c2 is essentially saying “the constants do not matter because you may pick c1 and c2 however you like to satisfy these conditions.” The portion of the definition that allows us to select n0 is essentially saying “we are only interested in large n, since you only have to satisfy the condition for all n bigger than n0 , and you may make n0 as big a constant as you like.” An example: Consider the function f (n) = 8n2 + 2n − 3. Our informal rule of keeping the largest term and throwing away the constants suggests that f (n) ∈ Θ(n2 ) (since f grows quadratically). Let’s see why the formal definition bears out this informal observation. We need to show two things: first, that f (n) does grows asymptotically at least as fast as n2 , and second, that f (n) grows no faster asymptotically than n2 . We’ll do both very carefully. Lower bound: f (n) grows asymptotically at least as fast as n2 : This is established by the portion of the definition that reads: (paraphrasing): “there exist positive constants c1 and n0 , such that f (n) ≥ c1 n2 for all n ≥ n0 .” Consider the following (almost correct) reasoning: f (n) = 8n2 + 2n − 3 ≥ 8n2 − 3 = 7n2 + (n2 − 3) ≥ 7n2 = 7n2 . Thus, if we set c1 = 7, then we are done. But in the above reasoning we have implicitly made the assumptions that 2n ≥ 0 and n2 √ 3 ≥ 0. These are not true for all n, but they are√ for all sufficiently − true large n. In particular, if n ≥ 3, then both are true. So let us select n0 = 3, and now we have f (n) ≥ c1 n2 , for all n ≥ n0 , which is what we need. Upper bound: f (n) grows asymptotically no faster than n2 : This is established by the portion of the definition that reads “there exist positive constants c2 and n0 , such that f (n) ≤ c2 n2 for all n ≥ n0 .” Consider the following reasoning (which is almost correct): f (n) = 8n2 + 2n − 3 ≤ 8n2 + 2n ≤ 8n2 + 2n2 = 10n2 . This means that if we let c2 = 10, then we are done. We have implicitly made the assumption that 2n ≤ 2n2 . This is not true for all n, but it is true for all n ≥ 1. So, let us select n0 = 1, and now we have f (n) ≤ c2 n2 for all n ≥ n0 , which is what we need. √ From the lower bound, we have n0 ≥ √ and from the upper bound we have n0 ≥ 1, and so combining these 3 √ we let n0 be the larger of the two: n0 = 3. Thus, in conclusion, if we let c1 = 7, c2 = 10, and n0 = 3, then we have 0 ≤ c1 g(n) ≤ f (n) ≤ c2 g(n) for all n ≥ n0 , and this is exactly what the definition requires. Since we have shown (by construction) the existence of con- stants c1 , c2 , and n0 , we have established that f (n) ∈ n2 . (Whew! That was a lot more work than just the informal notion of throwing away constants and keeping the largest term, but it shows how this informal notion is implemented formally in the definition.) Now let’s show why f (n) is not in some other asymptotic class. First, let’s show that f (n) ∈ Θ(n). If this were / true, then we would have to satisfy both the upper and lower bounds. It turns out that the lower bound is satisfied (because f (n) grows at least as fast asymptotically as n). But the upper bound is false. In particular, the upper bound requires us to show “there exist positive constants c2 and n0 , such that f (n) ≤ c2 n for all n ≥ n0 .” Informally, we know that as n becomes large enough f (n) = 8n2 + 2n − 3 will eventually exceed c2 n no matter Lecture Notes 92 CMSC 451
- 93. how large wemake c2 (since f (n) is growing quadratically and c2 n is only growing linearly). To show this formally, suppose towards a contradiction that constants c2 and n0 did exist, such that 8n2 + 2n − 3 ≤ c2 n for all n ≥ n0 . Since this is true for all sufficiently large n then it must be true in the limit as n tends to infinity. If we divide both side by n we have: 3 lim 8n + 2 − ≤ c2 . n→∞ n It is easy to see that in the limit the left side tends to ∞, and so no matter how large c2 is, this statement is violated. This means that f (n) ∈ Θ(n). / Let’s show that f (n) ∈ Θ(n3 ). Here the idea will be to violate the lower bound: “there exist positive constants / c1 and n0 , such that f (n) ≥ c1 n3 for all n ≥ n0 .” Informally this is true because f (n) is growing quadratically, and eventually any cubic function will exceed it. To show this formally, suppose towards a contradiction that constants c1 and n0 did exist, such that 8n2 + 2n − 3 ≥ c1 n3 for all n ≥ n0 . Since this is true for all sufficiently large n then it must be true in the limit as n tends to infinity. If we divide both side by n3 we have: 8 2 3 lim + − 3 ≥ c1 . n→∞ n n2 n It is easy to see that in the limit the left side tends to 0, and so the only way to satisfy this requirement is to set c1 = 0, but by hypothesis c1 is positive. This means that f (n) ∈ Θ(n3 ). / O-notation and Ω-notation: We have seen that the definition of Θ-notation relies on proving both a lower and upper asymptotic bound. Sometimes we are only interested in proving one bound or the other. The O-notation allows us to state asymptotic upper bounds and the Ω-notation allows us to state asymptotic lower bounds. Definition: Given any function g(n), O(g(n)) = {f (n) | there exist positive constants c and n0 such that 0 ≤ f (n) ≤ cg(n) for all n ≥ n0 }. Definition: Given any function g(n), Ω(g(n)) = {f (n) | there exist positive constants c and n0 such that 0 ≤ cg(n) ≤ f (n) for all n ≥ n0 }. Compare this with the definition of Θ. You will see that O-notation only enforces the upper bound of the Θ definition, and Ω-notation only enforces the lower bound. Also observe that f (n) ∈ Θ(g(n)) if and only if f (n) ∈ O(g(n)) and f (n) ∈ Ω(g(n)). Intuitively, f (n) ∈ O(g(n)) means that f (n) grows asymptotically at the same rate or slower than g(n). Whereas, f (n) ∈ O(g(n)) means that f (n) grows asymptotically at the same rate or faster than g(n). For example f (n) = 3n2 + 4n ∈ Θ(n2 ) but it is not in Θ(n) or Θ(n3 ). But f (n) ∈ O(n2 ) and in O(n3 ) but not in O(n). Finally, f (n) ∈ Ω(n2 ) and in Ω(n) but not in Ω(n3 ). The Limit Rule for Θ: The previous examples which used limits suggest alternative way of showing that f (n) ∈ Θ(g(n)). Limit Rule for Θ-notation: Given positive functions f (n) and g(n), if f (n) lim = c, n→∞ g(n) for some constant c > 0 (strictly positive but not infinity), then f (n) ∈ Θ(g(n)). Lecture Notes 93 CMSC 451
- 94. Limit Rule forO-notation: Given positive functions f (n) and g(n), if f (n) lim = c, n→∞ g(n) for some constant c ≥ 0 (nonnegative but not infinite), then f (n) ∈ O(g(n)). Limit Rule for Ω-notation: Given positive functions f (n) and g(n), if f (n) lim =0 n→∞ g(n) (either a strictly positive constant or infinity) then f (n) ∈ Ω(g(n)). This limit rule can be applied in almost every instance (that I know of) where the formal definition can be used, and it is almost always easier to apply than the formal definition. The only exceptions that I know of are strange instances where the limit does not exist (e.g. f (n) = n(1+sin n) ). But since most running times are fairly well-behaved functions this is rarely a problem. For example, recall the function f (n) = 8n2 + 2n − 3. To show that f (n) ∈ Θ(n2 ) we let g(n) = n2 and compute the limit. We have 8n2 + 2n − 3 2 3 lim = lim 8 + − 2 = 8, n→∞ n2 n→∞ n n (since the two fractional terms tend to 0 in the limit). Since 8 is a nonzero constant, it follows that f (n) ∈ Θ(g(n)). You may recall the important rules from calculus for evaluating limits. (If not, dredge out your calculus book to remember.) Most of the rules are pretty self evident (e.g., the limit of a finite sum is the sum of the individual limits). One important rule to remember is the following: L’Hˆ pital’s rule: If f (n) and g(n) both approach 0 or both approach ∞ in the limit, then o f (n) f (n) lim = lim , n→∞ g(n) n→∞ g (n) where f (n) and g (n) denote the derivatives of f and g relative to n. Exponentials and Logarithms: Exponentials and logarithms are very important in analyzing algorithms. The fol- lowing are nice to keep in mind. The terminology lgb n means (lg n)b . Lemma: Given any positive constants a > 1, b, and c: nb lgb n lim =0 lim = 0. n→∞ an n→∞ nc We won’t prove these, but they can be shown by taking appropriate powers, and then applying L’Hˆ pital’s rule. o The important bottom line is that polynomials always grow more slowly than exponentials whose base is greater than 1. For example: n500 ∈ O(2n ). For this reason, we will try to avoid exponential running times at all costs. Conversely, logarithmic powers (sometimes called polylogarithmic functions) grow more slowly than any polynomial. For example: lg500 n ∈ O(n). Lecture Notes 94 CMSC 451
- 95. For this reason,we will usually be happy to allow any number of additional logarithmic factors, if it means avoiding any additional powers of n. At this point, it should be mentioned that these last observations are really asymptotic results. They are true in the limit for large n, but you should be careful just how high the crossover point is. For example, by my calculations, lg500 n ≤ n only for n > 26000 (which is much larger than input size you’ll ever see). Thus, you should take this with a grain of salt. But, for small powers of logarithms, this applies to all reasonably large input sizes. For example lg2 n ≤ n for all n ≥ 16. Asymptotic Intuition: To get a intuitive feeling for what common asymptotic running times map into in terms of practical usage, here is a little list. • Θ(1): Constant time; you can’t beat it! • Θ(log n): This is typically the speed that most efficient data structures operate in for a single access. (E.g., inserting a key into a balanced binary tree.) Also it is the time to find an object in a sorted list of length n by binary search. • Θ(n): This is about the fastest that an algorithm can run, given that you need Θ(n) time just to read in all the data. • Θ(n log n): This is the running time of the best sorting algorithms. Since many problems require sorting the inputs, this is still considered quite efficient. • Θ(n2 ), Θ(n3 ), . . ..: Polynomial time. These running times are acceptable either when the exponent is small or when the data size is not too large (e.g. n ≤ 1, 000). • Θ(2n ), Θ(3n ): Exponential time. This is only acceptable when either (1) your know that you inputs will be of very small size (e.g. n ≤ 50), or (2) you know that this is a worst-case running time that will rarely occur in practical instances. In case (2), it would be a good idea to try to get a more accurate average case analysis. • Θ(n!), Θ(nn ): Acceptable only for really small inputs (e.g. n ≤ 20). Are their even bigger functions? Definitely! For example, if you want to see a function that grows inconceivably fast, look up the definition of Ackerman’s function in our text. Max Dominance Revisited: Returning to our Max Dominance algorithms, recall that one had a running time of T1 (n) = n2 and the other had a running time of T2 (n) = n log n + n(n − 1)/2. Expanding the latter function and grouping terms in order of their growth rate we have n2 n T2 (n) = + n log n − . 2 2 We will leave it as an easy exercise to show that both T1 (n) and T2 (n) are Θ(n2 ). Although the second algorithm is twice as fast for large n (because of the 1/2 factor multiplying the n2 term), this does not represent a significant improvement. Supplemental Lecture 2: Max Dominance Read: Review Chapters 1–4 in CLRS. Faster Algorithm for Max-Dominance: Recall the max-dominance problem from the last two lectures. So far we have introduced a simple brute-force algorithm that ran in O(n2 ) time, which operated by comparing all pairs of points. Last time we considered a slight improvement, which sorted the points by their x-coordinate, and then compared each point against the subsequent points in the sorted order. However, this improvement, only improved matters by a constant factor. The question we consider today is whether there is an approach that is significantly better. Lecture Notes 95 CMSC 451
- 96. A Major Improvement:The problem with the previous algorithm is that, even though we have cut the number of comparisons roughly in half, each point is still making lots of comparisons. Can we save time by making only one comparison for each point? The inner while loop is testing to see whether any point that follows P [i] in the sorted list has a larger y-coordinate. This suggests, that if we knew which point among P [i + 1, . . . , n] had the maximum y-coordinate, we could just test against that point. How can we do this? Here is a simple observation. For any set of points, the point with the maximum y- coordinate is the maximal point with the smallest x-coordiante. This suggests that we can sweep the points backwards, from right to left. We keep track of the index j of the most recently seen maximal point. (Initially the rightmost point is maximal.) When we encounter the point P [i], it is maximal if and only if P [i].y ≥ P [j].y. This suggests the following algorithm. Max Dominance: Sort and Reverse Scan MaxDom3(P, n) { Sort P in ascending order by x-coordinate; output P[n]; // last point is always maximal j = n; for i = n-1 downto 1 { if (P[i].y >= P[j].y) { // is P[i] maximal? output P[i]; // yes..output it j = i; // P[i] has the largest y so far } } } The running time of the for-loop is obviously O(n), because there is just a single loop that is executed n − 1 times, and the code inside takes constant time. The total running time is dominated by the O(n log n) sorting time, for a total of O(n log n) time. How much of an improvement is this? Probably the most accurate way to find out would be to code the two up, and compare their running times. But just to get a feeling, let’s look at the ratio of the running times, ignoring constant factors: n2 n = . n lg n lg n (I use the notation lg n to denote the logarithm base 2, ln n to denote the natural logarithm (base e) and log n when I do not care about the base. Note that a change in base only affects the value of a logarithm function by a constant amount, so inside of O-notation, we will usually just write log n.) For relatively small values of n (e.g. less than 100), both algorithms are probably running fast enough that the difference will be practically negligible. (Rule 1 of algorithm optimization: Don’t optimize code that is already fast enough.) On larger inputs, say, n = 1, 000, the ratio of n to log n is about 1000/10 = 100, so there is a 100- to-1 ratio in running times. Of course, we would need to factor in constant factors, but since we are not using any really complex data structures, it is hard to imagine that the constant factors will differ by more than, say, 10. For even larger inputs, say, n = 1, 000, 000, we are looking at a ratio of roughly 1, 000, 000/20 = 50, 000. This is quite a significant difference, irrespective of the constant factors. Divide and Conquer Approach: One problem with the previous algorithm is that it relies on sorting. This is nice and clean (since it is usually easy to get good code for sorting without troubling yourself to write your own). However, if you really wanted to squeeze the most efficiency out of your code, you might consider whether you can solve this problem without invoking a sorting algorithm. One of the basic maxims of algorithm design is to first approach any problem using one of the standard algorithm design paradigms, e.g. divide and conquer, dynamic programming, greedy algorithms, depth-first search. We will talk more about these methods as the semester continues. For this problem, divide-and-conquer is a natural method to choose. What is this paradigm? Lecture Notes 96 CMSC 451
- 97. Divide: Divide theproblem into two subproblems (ideally of approximately equal sizes), Conquer: Solve each subproblem recursively, and Combine: Combine the solutions to the two subproblems into a global solution. How shall we divide the problem? I can think of a couple of ways. One is similar to how MergeSort operates. Just take the array of points P [1..n], and split into two subarrays of equal size P [1..n/2] and P [n/2 + 1..n]. Because we do not sort the points, there is no particular relationship between the points in one side of the list from the other. Another approach, which is more reminiscent of QuickSort is to select a random element from the list, called a pivot, x = P [r], where r is a random integer in the range from 1 to n, and then partition the list into two sublists, those elements whose x-coordinates are less than or equal to x and those that greater than x. This will not be guaranteed to split the list into two equal parts, but on average it can be shown that it does a pretty good job. Let’s consider the first method. (The quicksort method will also work, but leads to a tougher analysis.) Here is more concrete outline. We will describe the algorithm at a very high level. The input will be a point array, and a point array will be returned. The key ingredient is a function that takes the maxima of two sets, and merges them into an overall set of maxima. Max Dominance: Divide-and-Conquer MaxDom4(P, n) { if (n == 1) return {P[1]}; // one point is trivially maximal m = n/2; // midpoint of list M1 = MaxDom4(P[1..m], m); // solve for first half M2 = MaxDom4(P[m+1..n], n-m); // solve for second half return MaxMerge(M1, M2); // merge the results } The general process is illustrated below. The main question is how the procedure Max Merge() is implemented, because it does all the work. Let us assume that it returns a list of points in sorted order according to x-coordinates of the maximal points. Observe that if a point is to be maximal overall, then it must be maximal in one of the two sublists. However, just because a point is maximal in some list, does not imply that it is globally maximal. (Consider point (7, 10) in the example.) However, if it dominates all the points of the other sublist, then we can assert that it is maximal. I will describe the procedure at a very high level. It operates by walking through each of the two sorted lists of maximal points. It maintains two pointers, one pointing to the next unprocessed item in each list. Think of these as fingers. Take the finger pointing to the point with the smaller x-coordinate. If its y-coordinate is larger than the y-coordinate of the point under the other finger, then this point is maximal, and is copied to the next position of the result list. Otherwise it is not copied. In either case, we move to the next point in the same list, and repeat the process. The result list is returned. The details will be left as an exercise. Observe that because we spend a constant amount of time processing each point (either copying it to the result list or skipping over it) the total execution time of this procedure is O(n). Recurrences: How do we analyze recursive procedures like this one? If there is a simple pattern to the sizes of the recursive calls, then the best way is usually by setting up a recurrence, that is, a function which is defined recursively in terms of itself. We break the problem into two subproblems of size roughly n/2 (we will say exactly n/2 for simplicity), and the additional overhead of merging the solutions is O(n). We will ignore constant factors, writing O(n) just as n, giving: T (n) = 1 if n = 1, T (n) = 2T (n/2) + n if n > 1. Lecture Notes 97 CMSC 451
- 98. (2,14) (2,14) 14 14 (7,13) (7,13) 12 (12,12) 12 (12,12) (9,10) (9,10) 10 (4,11) (14,10) 10 (4,11) (14,10) 8 8 (15,7) (15,7) 6 (7,7) 6 (7,7) 4 (11,5) (16,4) 4 (11,5) (16,4) 2 (13,3) 2 (13,3) (5,1) (5,1) 2 4 6 8 10 12 14 16 2 4 6 8 10 12 14 16 Input and initial partition. Solutions to subproblems. (2,14) 14 (7,13) 12 (12,12) (9,10) 10 (4,11) (14,10) 8 (15,7) 6 (7,7) 4 (11,5) (16,4) 2 (13,3) (5,1) 2 4 6 8 10 12 14 16 Merged solution. Fig. 68: Divide and conquer approach. Lecture Notes 98 CMSC 451
- 99. Solving Recurrences byThe Master Theorem: There are a number of methods for solving the sort of recurrences that show up in divide-and-conquer algorithms. The easiest method is to apply the Master Theorem that is given in CLRS. Here is a slightly more restrictive version, but adequate for a lot of instances. See CLRS for the more complete version of the Master Theorem and its proof. Theorem: (Simplified Master Theorem) Let a ≥ 1, b > 1 be constants and let T (n) be the recurrence T (n) = aT (n/b) + cnk , defined for n ≥ 0. Case (1): a > bk then T (n) is Θ(nlogb a ). Case (2): a = bk then T (n) is Θ(nk log n). Case (3): a < bk then T (n) is Θ(nk ). Using this version of the Master Theorem we can see that in our recurrence a = 2, b = 2, and k = 1, so a = bk and case (2) applies. Thus T (n) is Θ(n log n). There many recurrences that cannot be put into this form. For example, the following recurrence is quite common: T (n) = 2T (n/2) + n log n. This solves to T (n) = Θ(n log2 n), but the Master Theorem (either this form or the one in CLRS will not tell you this.) For such recurrences, other methods are needed. Expansion: A more basic method for solving recurrences is that of expansion (which CLRS calls iteration). This is a rather painstaking process of repeatedly applying the definition of the recurrence until (hopefully) a simple pattern emerges. This pattern usually results in a summation that is easy to solve. If you look at the proof in CLRS for the Master Theorem, it is actually based on expansion. Let us consider applying this to the following recurrence. We assume that n is a power of 3. T (1) = 1 n T (n) = 2T +n if n > 1 3 First we expand the recurrence into a summation, until seeing the general pattern emerge. n T (n) = 2T +n 3 n n n 2n = 2 2T + + n = 4T + n+ 9 3 9 3 n n 2n n 2n 4n = 4 2T + + n+ = 8T + n+ + 27 9 3 27 3 9 . . . k−1 k−1 n 2i n n = 2k T + = 2k T +n (2/3)i . 3k i=0 3i 3k i=0 The parameter k is the number of expansions (not to be confused with the value of k we introduced earlier on the overhead). We want to know how many expansions are needed to arrive at the basis case. To do this we set n/(3k ) = 1, meaning that k = log3 n. Substituting this in and using the identity alog b = blog a we have: log3 n−1 log3 n−1 T (n) = 2log3 n T (1) + n (2/3)i = nlog3 2 + n (2/3)i . i=0 i=0 Lecture Notes 99 CMSC 451
- 100. Next, we canapply the formula for the geometric series and simplify to get: 1 − (2/3)log3 n T (n) = nlog3 2 + n 1 − (2/3) = nlog3 2 + 3n(1 − (2/3)log3 n ) = nlog3 2 + 3n(1 − nlog3 (2/3) ) = nlog3 2 + 3n(1 − n(log3 2)−1 ) = nlog3 2 + 3n − 3nlog3 2 = 3n − 2nlog3 2 . Since log3 2 ≈ 0.631 < 1, T (n) is dominated by the 3n term asymptotically, and so it is Θ(n). Induction and Constructive Induction: Another technique for solving recurrences (and this works for summations as well) is to guess the solution, or the general form of the solution, and then attempt to verify its correctness through induction. Sometimes there are parameters whose values you do not know. This is fine. In the course of the induction proof, you will usually find out what these values must be. We will consider a famous example, that of the Fibonacci numbers. F0 = 0 F1 = 1 Fn = Fn−1 + Fn−2 for n ≥ 2. The Fibonacci numbers arise in data structure design. If you study AVL (height balanced) trees in data structures, you will learn that the minimum-sized AVL trees are produced by the recursive construction given below. Let L(i) denote the number of leaves in the minimum-sized AVL tree of height i. To construct a minimum-sized AVL tree of height i, you create a root node whose children consist of a minimum-sized AVL tree of heights i − 1 and i − 2. Thus the number of leaves obeys L(0) = L(1) = 1, L(i) = L(i − 1) + L(i − 2). It is easy to see that L(i) = Fi+1 . L(0) = 1 L(1)=1 L(2)=2 L(3)=3 L(4)=5 Fig. 69: Minimum-sized AVL trees. If you expand the Fibonacci series for a number of terms, you will observe that Fn appears to grow exponentially, but not as fast as 2n . It is tempting to conjecture that Fn ≤ φn−1 , for some real parameter φ, where 1 < φ < 2. We can use induction to prove this and derive a bound on φ. Lemma: For all integers n ≥ 1, Fn ≤ φn−1 for some constant φ, 1 < φ < 2. Proof: We will try to derive the tightest bound we can on the value of φ. Basis: For the basis cases we consider n = 1. Observe that F1 = 1 ≤ φ0 , as desired. Induction step: For the induction step, let us assume that Fm ≤ φm−1 whenever 1 ≤ m < n. Using this induction hypothesis we will show that the lemma holds for n itself, whenever n ≥ 2. Since n ≥ 2, we have Fn = Fn−1 + Fn−2 . Now, since n − 1 and n − 2 are both strictly less than n, we can apply the induction hypothesis, from which we have Fn ≤ φn−2 + φn−3 = φn−3 (1 + φ). Lecture Notes 100 CMSC 451
- 101. We want toshow that this is at most φn−1 (for a suitable choice of φ). Clearly this will be true if and only if (1 + φ) ≤ φ2 . This is not true for all values of φ (for example it is not true when φ = 1 but it is true when φ = 2.) At the critical value of φ this inequality will be an equality, implying that we want to find the roots of the equation φ2 − φ − 1 = 0. By the quadratic formula we have √ √ 1± 1+4 1± 5 φ = = . 2 2 √ Since 5 ≈ 2.24, observe that one of the roots is negative, and hence would not be a possible candidate for φ. The positive root is √ 1+ 5 φ = ≈ 1.618. 2 There is a very subtle bug in the preceding proof. Can you spot it? The error occurs in the case n = 2. Here we claim that F2 = F1 + F0 and then we apply the induction hypothesis to both F1 and F0 . But the induction hypothesis only applies for m ≥ 1, and hence cannot be applied to F0 ! To fix it we could include F2 as part of the basis case as well. Notice not only did we prove the lemma by induction, but we actually determined the value of φ which makes the lemma true. This is why this method is called constructive induction. √ By the way, the value φ = 1 (1 + 5) is a famous constant in mathematics, architecture and art. It is the golden 2 ratio. Two numbers A and B satisfy the golden ratio if A A+B = . B A It is easy to verify that A = φ and B = 1 satisfies this condition. This proportion occurs throughout the world of art and architecture. Supplemental Lecture 3: Recurrences and Generating Functions Read: This material is not covered in CLR. There a good description of generating functions in D. E. Knuth, The Art of Computer Programming, Vol 1. Generating Functions: The method of constructive induction provided a way to get a bound on Fn , but we did not get an exact answer, and we had to generate a good guess before we were even able to start. Let us consider an approach to determine an exact representation of Fn , which requires no guesswork. This method is based on a very elegant concept, called a generating function. Consider any infinite sequence: a0 , a1 , a2 , a3 , . . . If we would like to “encode” this sequence succinctly, we could define a polynomial function such that these are the coefficients of the function: G(z) = a0 + a1 z + a2 z 2 + a3 z 3 + . . . This is called the generating function of the sequence. What is z? It is just a symbolic variable. We will (almost) never assign it a specific value. Thus, every infinite sequence of numbers has a corresponding generating func- tion, and vice versa. What is the advantage of this representation? It turns out that we can perform arithmetic Lecture Notes 101 CMSC 451
- 102. transformations on thesefunctions (e.g., adding them, multiplying them, differentiating them) and this has a corresponding effect on the underlying transformations. It turns out that some nicely-structured sequences (like the Fibonacci numbers, and many sequences arising from linear recurrences) have generating functions that are easy to write down and manipulate. Let’s consider the generating function for the Fibonacci numbers: G(z) = F0 + F1 z + F2 z 2 + F3 z 3 + . . . = z + z 2 + 2z 3 + 3z 4 + 5z 5 + . . . The trick in dealing with generating functions is to figure out how various manipulations of the generating function to generate algebraically equivalent forms. For example, notice that if we multiply the generating function by a factor of z, this has the effect of shifting the sequence to the right: G(z) = F0 + F1 z + F2 z 2 + F3 z 3 + F4 z 4 + ... zG(z) = F0 z + F1 z 2 + F2 z 3 + F3 z 4 + ... z 2 G(z) = F0 z 2 + F1 z 3 + F2 z 4 + ... Now, let’s try the following manipulation. Compute G(z) − zG(z) − z 2 G(z), and see what we get (1 − z − z 2 )G(z) = F0 + (F1 − F0 )z + (F2 − F1 − F0 )z 2 + (F3 − F2 − F1 )z 3 + . . . + (Fi − Fi−1 − Fi−2 )z i + . . . = z. Observe that every term except the second is equal to zero by the definition of Fi . (The particular manipulation we picked was chosen to cause this cancellation to occur.) From this we may conclude that z G(z) = . 1 − z − z2 So, now we have an alternative representation for the Fibonacci numbers, as the coefficients of this function if expanded as a power series. So what good is this? The main goal is to get at the coefficients of its power series expansion. There are certain common tricks that people use to manipulate generating functions. The first is to observe that there are some functions for which it is very easy to get an power series expansion. For example, the following is a simple consequence of the formula for the geometric series. If 0 < c < 1 then ∞ 1 ci = . i=0 1−c Setting z = c, we have 1 = 1 + z + z2 + z3 + . . . 1−z (In other words, 1/(1 − z) is the generating function for the sequence (1, 1, 1, . . .). In general, given an constant a we have 1 = 1 + az + a2 z 2 + a3 z 3 + . . . 1 − az is the generating function for (1, a, a2 , a3 , . . .). It would be great if we could modify our generating function to be in the form of 1/(1 − az) for some constant a, since then we could then extract the coefficients of the power series easily. In order to do this, we would like to rewrite the generating function in the following form: z A B G(z) = = + , 1 − z − z2 1 − az 1 − bz Lecture Notes 102 CMSC 451
- 103. for some A,B, a, b. We will skip the steps in doing this, but it is not hard to verify the roots of (1 − az)(1 − bz) (which are 1/a and 1/b) must be equal to the roots of 1 − z − z 2 . We can then solve for a and b by taking the reciprocals of the roots of this quadratic. Then by some simple algebra we can plug these values in and solve for A and B yielding: √ √ z 1/ 5 −1/ 5 1 1 1 G(z) = = + = √ − , 1 − z − z2 1 − φz 1−φˆ 5 1 − φz 1−φˆ √ √ ˆ where φ = (1 + 5)/2 and φ = (1 − 5)/2. (In particular, to determine A, multiply the equation by 1 − φz, and then consider what happens when z = 1/φ. A similar trick can be applied to get B. In general, this is called the method of partial fractions.) Now we are in good shape, because we can extract the coefficients for these two fractions from the above function. From this we have the following: 1 G(z) = √ 5 ( 1 + φz + φ2 z 2 + ... −1 ˆ + −φz + −φ2 z 2 ˆ + ... ) Combining terms we have ∞ 1 G(z) = √ ˆ (φi − φi )z i . 5 i=0 We can now read off the coefficients easily. In particular it follows that 1 Fn = √ (φn − φn ). ˆ 5 This is an exact result, and no guesswork was needed. The only parts that involved some cleverness (beyond the invention of generating functions) was (1) coming up with the simple closed form formula for G(z) by taking appropriate differences and applying the rule for the recurrence, and (2) applying the method of partial fractions to get the generating function into one for which we could easily read off the final coefficients. This is a rather remarkable, because it says that we can express the integer Fn as the sum of two powers of to ˆ irrational numbers φ and φ. You might try this for a few specific values of n to see why this is true. By the way, ˆ when you observe that φ < 1, it is clear that the first term is the dominant one. Thus we have, for large enough √ n, Fn = φn / 5, rounded to the nearest integer. Supplemental Lecture 4: Medians and Selection Read: Chapter 9 of CLRS. Selection: We have discussed recurrences and the divide-and-conquer method of solving problems. Today we will give a rather surprising (and very tricky) algorithm which shows the power of these techniques. The problem that we will consider is very easy to state, but surprisingly difficult to solve optimally. Suppose that you are given a set of n numbers. Define the rank of an element to be one plus the number of elements that are smaller than this element. Since duplicate elements make our life more complex (by creating multiple elements of the same rank), we will make the simplifying assumption that all the elements are distinct for now. It will be easy to get around this assumption later. Thus, the rank of an element is its final position if the set is sorted. The minimum is of rank 1 and the maximum is of rank n. Of particular interest in statistics is the median. If n is odd then the median is defined to be the element of rank (n + 1)/2. When n is even there are two natural choices, namely the elements of ranks n/2 and (n/2) + 1. In Lecture Notes 103 CMSC 451
- 104. statistics it iscommon to return the average of these two elements. We will define the median to be either of these elements. Medians are useful as measures of the central tendency of a set, especially when the distribution of values is highly skewed. For example, the median income in a community is likely to be more meaningful measure of the central tendency than the average is, since if Bill Gates lives in your community then his gigantic income may significantly bias the average, whereas it cannot have a significant influence on the median. They are also useful, since in divide-and-conquer applications, it is often desirable to partition a set about its median value, into two sets of roughly equal size. Today we will focus on the following generalization, called the selection problem. Selection: Given a set A of n distinct numbers and an integer k, 1 ≤ k ≤ n, output the element of A of rank k. The selection problem can easily be solved in Θ(n log n) time, simply by sorting the numbers of A, and then returning A[k]. The question is whether it is possible to do better. In particular, is it possible to solve this problem in Θ(n) time? We will see that the answer is yes, and the solution is far from obvious. The Sieve Technique: The reason for introducing this algorithm is that it illustrates a very important special case of divide-and-conquer, which I call the sieve technique. We think of divide-and-conquer as breaking the problem into a small number of smaller subproblems, which are then solved recursively. The sieve technique is a special case, where the number of subproblems is just 1. The sieve technique works in phases as follows. It applies to problems where we are interested in finding a single item from a larger set of n items. We do not know which item is of interest, however after doing some amount of analysis of the data, taking say Θ(nk ) time, for some constant k, we find that we do not know what the desired item is, but we can identify a large enough number of elements that cannot be the desired value, and can be eliminated from further consideration. In particular “large enough” means that the number of items is at least some fixed constant fraction of n (e.g. n/2, n/3, 0.0001n). Then we solve the problem recursively on whatever items remain. Each of the resulting recursive solutions then do the same thing, eliminating a constant fraction of the remaining set. Applying the Sieve to Selection: To see more concretely how the sieve technique works, let us apply it to the selec- tion problem. Recall that we are given an array A[1..n] and an integer k, and want to find the k-th smallest element of A. Since the algorithm will be applied inductively, we will assume that we are given a subarray A[p..r] as we did in MergeSort, and we want to find the kth smallest item (where k ≤ r − p + 1). The initial call will be to the entire array A[1..n]. There are two principal algorithms for solving the selection problem, but they differ only in one step, which involves judiciously choosing an item from the array, called the pivot element, which we will denote by x. Later we will see how to choose x, but for now just think of it as a random element of A. We then partition A into three parts. A[q] contains the element x, subarray A[p..q − 1] will contain all the elements that are less than x, and A[q + 1..r], will contain all the element that are greater than x. (Recall that we assumed that all the elements are distinct.) Within each subarray, the items may appear in any order. This is illustrated below. It is easy to see that the rank of the pivot x is q − p + 1 in A[p..r]. Let xRank = q − p + 1. If k = xRank , then the pivot is the kth smallest, and we may just return it. If k < xRank , then we know that we need to recursively search in A[p..q −1] and if k > xRank then we need to recursively search A[q +1..r]. In this latter case we have eliminated q smaller elements, so we want to find the element of rank k − q. Here is the complete pseudocode. Notice that this algorithm satisfies the basic form of a sieve algorithm. It analyzes the data (by choosing the pivot element and partitioning) and it eliminates some part of the data set, and recurses on the rest. When k = xRank then we get lucky and eliminate everything. Otherwise we either eliminate the pivot and the right subarray or the pivot and the left subarray. We will discuss the details of choosing the pivot and partitioning later, but assume for now that they both take Θ(n) time. The question that remains is how many elements did we succeed in eliminating? If x is the largest Lecture Notes 104 CMSC 451
- 105. pivot p r 5 9 2 6 4 1 3 7 Before partitioing p q r 3 1 2 4 6 9 5 7 After partitioing x A[p..q−1] < x A[q+1..r] > x 5 3 9 1 2 2 6 4 x_rnk=4 4 6 6 6 6 5 1 9 9 5 5 6 3 5 5 7 x_rnk=3 x_rnk=2 (DONE!) 7 7 7 9 Initial Partition Recurse Partition Recurse Partition (k=6) (pivot = 4) (k=6−4=2) (pivot = 7) (k=2) (pivot = 6) Fig. 70: Selection Algorithm. Selection by the Sieve Technique Select(array A, int p, int r, int k) { // return kth smallest of A[p..r] if (p == r) return A[p] // only 1 item left, return it else { x = ChoosePivot(A, p, r) // choose the pivot element q = Partition(A, p, r, x) // partition <A[p..q-1], x, A[q+1..r]> xRank = q - p + 1 // rank of the pivot if (k == xRank) return x // the pivot is the kth smallest else if (k < xRank) return Select(A, p, q-1, k) // select from left subarray else return Select(A, q+1, r, k-xRank)// select from right subarray } } Lecture Notes 105 CMSC 451
- 106. or smallest elementin the array, then we may only succeed in eliminating one element with each phase. In fact, if x is one of the smallest elements of A or one of the largest, then we get into trouble, because we may only eliminate it and the few smaller or larger elements of A. Ideally x should have a rank that is neither too large nor too small. Let us suppose for now (optimistically) that we are able to design the procedure Choose Pivot in such a way that is eliminates exactly half the array with each phase, meaning that we recurse on the remaining n/2 elements. This would lead to the following recurrence. 1 if n = 1, T (n) = T (n/2) + n otherwise. We can solve this either by expansion (iteration) or the Master Theorem. If we expand this recurrence level by level we see that we get the summation ∞ ∞ n n n 1 T (n) = n + + + · · · ≤ = n . 2 4 i=0 2i i=0 2i ∞ Recall the formula for the infinite geometric series. For any c such that |c| < 1, i=0 c i = 1/(1 − c). Using this we have T (n) ≤ 2n ∈ O(n). (This only proves the upper bound on the running time, but it is easy to see that it takes at least Ω(n) time, so the total running time is Θ(n).) This is a bit counterintuitive. Normally you would think that in order to design a Θ(n) time algorithm you could only make a single, or perhaps a constant number of passes over the data set. In this algorithm we make many passes (it could be as many as lg n). However, because we eliminate a constant fraction of elements with each phase, we get this convergent geometric series in the analysis, which shows that the total running time is indeed linear in n. This lesson is well worth remembering. It is often possible to achieve running times in ways that you would not expect. Note that the assumption of eliminating half was not critical. If we eliminated even one per cent, then the recurrence would have been T (n) = T (99n/100) + n, and we would have gotten a geometric series involving 99/100, which is still less than 1, implying a convergent series. Eliminating any constant fraction would have been good enough. Choosing the Pivot: There are two issues that we have left unresolved. The first is how to choose the pivot element, and the second is how to partition the array. Both need to be solved in Θ(n) time. The second problem is a rather easy programming exercise. Later, when we discuss QuickSort, we will discuss partitioning in detail. For the rest of the lecture, let’s concentrate on how to choose the pivot. Recall that before we said that we might think of the pivot as a random element of A. Actually this is not such a bad idea. Let’s see why. The key is that we want the procedure to eliminate at least some constant fraction of the array after each parti- tioning step. Let’s consider the top of the recurrence, when we are given A[1..n]. Suppose that the pivot x turns out to be of rank q in the array. The partitioning algorithm will split the array into A[1..q − 1] < x, A[q] = x and A[q + 1..n] > x. If k = q, then we are done. Otherwise, we need to search one of the two subarrays. They are of sizes q − 1 and n − q, respectively. The subarray that contains the kth smallest element will generally depend on what k is, so in the worst case, k will be chosen so that we have to recurse on the larger of the two subarrays. Thus if q > n/2, then we may have to recurse on the left subarray of size q − 1, and if q < n/2, then we may have to recurse on the right subarray of size n − q. In either case, we are in trouble if q is very small, or if q is very large. If we could select q so that it is roughly of middle rank, then we will be in good shape. For example, if n/4 ≤ q ≤ 3n/4, then the larger subarray will never be larger than 3n/4. Earlier we said that we might think of the pivot as a random element of the array A. Actually this works pretty well in practice. The reason is that Lecture Notes 106 CMSC 451
- 107. roughly half ofthe elements lie between ranks n/4 and 3n/4, so picking a random element as the pivot will succeed about half the time to eliminate at least n/4. Of course, we might be continuously unlucky, but a careful analysis will show that the expected running time is still Θ(n). We will return to this later. Instead, we will describe a rather complicated method for computing a pivot element that achieves the desired properties. Recall that we are given an array A[1..n], and we want to compute an element x whose rank is (roughly) between n/4 and 3n/4. We will have to describe this algorithm at a very high level, since the details are rather involved. Here is the description for Select Pivot: Groups of 5: Partition A into groups of 5 elements, e.g. A[1..5], A[6..10], A[11..15], etc. There will be exactly m = n/5 such groups (the last one might have fewer than 5 elements). This can easily be done in Θ(n) time. Group medians: Compute the median of each group of 5. There will be m group medians. We do not need an intelligent algorithm to do this, since each group has only a constant number of elements. For example, we could just BubbleSort each group and take the middle element. Each will take Θ(1) time, and repeating this n/5 times will give a total running time of Θ(n). Copy the group medians to a new array B. Median of medians: Compute the median of the group medians. For this, we will have to call the selection algorithm recursively on B, e.g. Select(B, 1, m, k), where m = n/5 , and k = (m + 1)/2 . Let x be this median of medians. Return x as the desired pivot. The algorithm is illustrated in the figure below. To establish the correctness of this procedure, we need to argue that x satisfies the desired rank properties. 14 32 23 5 10 60 29 6 2 3 5 8 1 11 57 2 52 44 27 21 11 14 25 12 17 10 21 29 24 43 12 17 48 1 58 24 30 23 34 19 41 39 6 30 63 34 8 55 39 37 32 52 44 27 55 58 37 25 3 64 19 41 57 43 63 64 48 60 Group Get group medians 8 3 6 2 5 11 1 10 12 14 25 17 29 21 19 23 24 30 34 39 41 27 52 37 32 44 58 55 48 63 57 43 64 60 Get median of medians (Sorting of group medians is not really performed) Fig. 71: Choosing the Pivot. 30 is the final pivot. Lemma: The element x is of rank at least n/4 and at most 3n/4 in A. Proof: We will show that x is of rank at least n/4. The other part of the proof is essentially symmetrical. To do this, we need to show that there are at least n/4 elements that are less than or equal to x. This is a bit complicated, due to the floor and ceiling arithmetic, so to simplify things we will assume that n is evenly divisible by 5. Consider the groups shown in the tabular form above. Observe that at least half of the group medians are less than or equal to x. (Because x is their median.) And for each group median, there are three elements that are less than or equal to this median within its group (because it is the median of its Lecture Notes 107 CMSC 451
- 108. group). Therefore, thereare at least 3((n/5)/2 = 3n/10 ≥ n/4 elements that are less than or equal to x in the entire array. Analysis: The last order of business is to analyze the running time of the overall algorithm. We achieved the main goal, namely that of eliminating a constant fraction (at least 1/4) of the remaining list at each stage of the algorithm. The recursive call in Select() will be made to list no larger than 3n/4. However, in order to achieve this, within Select Pivot() we needed to make a recursive call to Select() on an array B consisting of n/5 elements. Everything else took only Θ(n) time. As usual, we will ignore floors and ceilings, and write the Θ(n) as n for concreteness. The running time is 1 if n = 1, T (n) ≤ T (n/5) + T (3n/4) + n otherwise. This is a very strange recurrence because it involves a mixture of different fractions (n/5 and 3n/4). This mixture will make it impossible to use the Master Theorem, and difficult to apply iteration. However, this is a good place to apply constructive induction. We know we want an algorithm that runs in Θ(n) time. Theorem: There is a constant c, such that T (n) ≤ cn. Proof: (by strong induction on n) Basis: (n = 1) In this case we have T (n) = 1, and so T (n) ≤ cn as long as c ≥ 1. Step: We assume that T (n ) ≤ cn for all n < n. We will then show that T (n) ≤ cn. By definition we have T (n) = T (n/5) + T (3n/4) + n. Since n/5 and 3n/4 are both less than n, we can apply the induction hypothesis, giving n 3n 1 3 T (n) ≤ c +c + n = cn + +n 5 4 5 4 19 19c = cn + n = n +1 . 20 20 This last expression will be ≤ cn, provided that we select c such that c ≥ (19c/20) + 1. Solving for c we see that this is true provided that c ≥ 20. Combining the constraints that c ≥ 1, and c ≥ 20, we see that by letting c = 20, we are done. A natural question is why did we pick groups of 5? If you look at the proof above, you will see that it works for any value that is strictly greater than 4. (You might try it replacing the 5 with 3, 4, or 6 and see what happens.) Supplemental Lecture 5: Analysis of BucketSort Probabilistic Analysis of BucketSort: We begin with a quick-and-dirty analysis of bucketsort. Since there are n buckets, and the items fall uniformly between them, we would expect a constant number of items per bucket. Thus, the expected insertion time for each bucket is only a constant. Therefore the expected running time of the algorithm is Θ(n). This quick-and-dirty analysis is probably good enough to convince yourself of this algorithm’s basic efficiency. A careful analysis involves understanding a bit about probabilistic analyses of algorithms. Since we haven’t done any probabilistic analyses yet, let’s try doing this one. (This one is rather typical.) The first thing to do in a probabilistic analysis is to define a random variable that describes the essential quantity that determines the execution time. A discrete random variable can be thought of as variable that takes on some Lecture Notes 108 CMSC 451
- 109. set of discretevalues with certain probabilities. More formally, it is a function that maps some some discrete sample space (the set of possible values) onto the reals (the probabilities). For 0 ≤ i ≤ n − 1, let Xi denote the random variable that indicates the number of elements assigned to the i-th bucket. Since the distribution is uniform, all of the random variables Xi have the same probability distribution, so we may as well talk about a single random variable X, which will work for any bucket. Since we are using a quadratic time algorithm to sort the elements of each bucket, we are interested in the expected sorting time, which is Θ(X 2 ). So this leads to the key question, what is the expected value of X 2 , denoted E[X 2 ]. Because the elements are assumed to be uniformly distributed, each element has an equal probability of going into any bucket, or in particular, it has a probability of p = 1/n of going into the ith bucket. So how many items do we expect will wind up in bucket i? We can analyze this by thinking of each element of A as being represented by a coin flip (with a biased coin, which has a different probability of heads and tails). With probability p = 1/n the number goes into bucket i, which we will interpret as the coin coming up heads. With probability 1 − 1/n the item goes into some other bucket, which we will interpret as the coin coming up tails. Since we assume that the elements of A are independent of each other, X is just the total number of heads we see after making n tosses with this (biased) coin. The number of times that a heads event occurs, given n independent trials in which each trial has two possible outcomes is a well-studied problem in probability theory. Such trials are called Bernoulli trials (named after the Swiss mathematician James Bernoulli). If p is the probability of getting a head, then the probability of getting k heads in n tosses is given by the following important formula n k n n! P (X = k) = p (1 − p)n−k where = . k k k!(n − k)! Although this looks messy, it is not too hard to see where it comes from. Basically pk is the probability of tossing k heads, (1 − p)n−k is the probability of tossing n − k tails, and n is the total number of different k ways that the k heads could be distributed among the n tosses. This probability distribution (as a function of k, for a given n and p) is called the binomial distribution, and is denoted b(k; n, p). If you consult a standard textbook on probability and statistics, then you will see the two important facts that we need to know about the binomial distribution. Namely, that its mean value E[X] and its variance Var[X] are E[X] = np and Var[X] = E[X 2 ] − E 2 [X] = np(1 − p). We want to determine E[X 2 ]. By the above formulas and the fact that p = 1/n we can derive this as n 1 n 2 1 E[X 2 ] = Var[X] + E 2 [X] = np(1 − p) + (np)2 = 1− + = 2− . n n n n Thus, for large n the time to insert the items into any one of the linked lists is a just shade less than 2. Summing up over all n buckets, gives a total running time of Θ(2n) = Θ(n). This is exactly what our quick-and-dirty analysis gave us, but now we know it is true with confidence. Supplemental Lecture 6: Long Integer Multiplication Read: This material on integer multiplication is not covered in CLRS. Long Integer Multiplication: The following little algorithm shows a bit more about the surprising applications of divide-and-conquer. The problem that we want to consider is how to perform arithmetic on long integers, and multiplication in particular. The reason for doing arithmetic on long numbers stems from cryptography. Most techniques for encryption are based on number-theoretic techniques. For example, the character string to be encrypted is converted into a sequence of numbers, and encryption keys are stored as long integers. Efficient Lecture Notes 109 CMSC 451
- 110. encryption and decryptiondepends on being able to perform arithmetic on long numbers, typically containing hundreds of digits. Addition and subtraction on large numbers is relatively easy. If n is the number of digits, then these algorithms run in Θ(n) time. (Go back and analyze your solution to the problem on Homework 1). But the standard algorithm for multiplication runs in Θ(n2 ) time, which can be quite costly when lots of long multiplications are needed. This raises the question of whether there is a more efficient way to multiply two very large numbers. It would seem surprising if there were, since for centuries people have used the same algorithm that we all learn in grade school. In fact, we will see that it is possible. Divide-and-Conquer Algorithm: We know the basic grade-school algorithm for multiplication. We normally think of this algorithm as applying on a digit-by-digit basis, but if we partition an n digit number into two “super digits” with roughly n/2 each into longer sequences, the same multiplication rule still applies. n n/2 n/2 w x A y z B wz xz wy xy wy wz + xy xz Product Fig. 72: Long integer multiplication. To avoid complicating things with floors and ceilings, let’s just assume that the number of digits n is a power of 2. Let A and B be the two numbers to multiply. Let A[0] denote the least significant digit and let A[n−1] denote the most significant digit of A. Because of the way we write numbers, it is more natural to think of the elements of A as being indexed in decreasing order from left to right as A[n − 1..0] rather than the usual A[0..n − 1]. Let m = n/2. Let w = A[n − 1..m] x = A[m − 1..0] and y = B[n − 1..m] z = B[m − 1..0]. If we think of w, x, y and z as n/2 digit numbers, we can express A and B as A = w · 10m + x B = y · 10m + z, and their product is mult(A, B) = mult(w, y)102m + (mult(w, z) + mult(x, y))10m + mult(x, z). The operation of multiplying by 10m should be thought of as simply shifting the number over by m positions to the right, and so is not really a multiplication. Observe that all the additions involve numbers involving roughly n/2 digits, and so they take Θ(n) time each. Thus, we can express the multiplication of two long integers as the result of four products on integers of roughly half the length of the original, and a constant number of additions and shifts, each taking Θ(n) time. This suggests that if we were to implement this algorithm, its running time would be given by the following recurrence 1 if n = 1, T (n) = 4T (n/2) + n otherwise. Lecture Notes 110 CMSC 451
- 111. If we applythe Master Theorem, we see that a = 4, b = 2, k = 1, and a > bk , implying that Case 1 holds and the running time is Θ(nlg 4 ) = Θ(n2 ). Unfortunately, this is no better than the standard algorithm. Faster Divide-and-Conquer Algorithm: Even though the above exercise appears to have gotten us nowhere, it ac- tually has given us an important insight. It shows that the critical element is the number of multiplications on numbers of size n/2. The number of additions (as long as it is a constant) does not affect the running time. So, if we could find a way to arrive at the same result algebraically, but by trading off multiplications in favor of additions, then we would have a more efficient algorithm. (Of course, we cannot simulate multiplication through repeated additions, since the number of additions must be a constant, independent of n.) The key turns out to be a algebraic “trick”. The quantities that we need to compute are C = wy, D = xz, and E = (wz + xy). Above, it took us four multiplications to compute these. However, observe that if instead we compute the following quantities, we can get everything we want, using only three multiplications (but with more additions and subtractions). C = mult(w, y) D = mult(x, z) E = mult((w + x), (y + z)) − C − D = (wy + wz + xy + xz) − wy − xz = (wz + xy). Finally we have mult(A, B) = C · 102m + E · 10m + D. Altogether we perform 3 multiplications, 4 additions, and 2 subtractions all of numbers with n/2 digitis. We still need to shift the terms into their proper final positions. The additions, subtractions, and shifts take Θ(n) time in total. So the total running time is given by the recurrence: 1 if n = 1, T (n) = 3T (n/2) + n otherwise. Now when we apply the Master Theorem, we have a = 3, b = 2 and k = 1, yielding T (n) ∈ Θ(nlg 3 ) ≈ Θ(n1.585 ). Is this really an improvement? This algorithm carries a larger constant factor because of the overhead of recur- sion and the additional arithmetic operations. But asymptotics says that if n is large enough, then this algorithm will be superior. For example, if we assume that the clever algorithm has overheads that are 5 times greater than the simple algorithm (e.g. 5n1.585 versus n2 ) then this algorithm beats the simple algorithm for n ≥ 50. If the overhead was 10 times larger, then the crossover would occur for n ≥ 260. Although this may seem like a very large number, recall that in cryptogrphy applications, encryption keys of this length and longer are quite reasonable. Supplemental Lecture 7: Dynamic Programming: 0–1 Knapsack Problem Read: The introduction to Chapter 16 in CLR. The material on the Knapsack Problem is not presented in our text, but is briefly discussed in Section 17.2. 0-1 Knapsack Problem: Imagine that a burglar breaks into a museum and finds n items. Let vi denote the value of the i-th item, and let wi denote the weight of the i-th item. The burglar carries a knapsack capable of holding total weight W . The burglar wishes to carry away the most valuable subset items subject to the weight constraint. For example, a burglar would rather steal diamonds before gold because the value per pound is better. But he would rather steal gold before lead for the same reason. We assume that the burglar cannot take a fraction of an object, so he/she must make a decision to take the object entirely or leave it behind. (There is a version of the Lecture Notes 111 CMSC 451
- 112. problem where theburglar can take a fraction of an object for a fraction of the value and weight. This is much easier to solve.) More formally, given v1 , v2 , . . . , vn and w1 , w2 . . . , wn , and W > 0, we wish to determine the subset T ⊆ {1, 2, . . . , n} (of objects to “take”) that maximizes vi , i∈T subject to wi ≤ W. i∈T Let us assume that the vi ’s, wi ’s and W are all positive integers. It turns out that this problem is NP-complete, and so we cannot really hope to find an efficient solution. However if we make the same sort of assumption that we made in counting sort, we can come up with an efficient solution. We assume that the wi ’s are small integers, and that W itself is a small integer. We show that this problem can be solved in O(nW ) time. (Note that this is not very good if W is a large integer. But if we truncate our numbers to lower precision, this gives a reasonable approximation algorithm.) Here is how we solve the problem. We construct an array V [0..n, 0..W ]. For 1 ≤ i ≤ n, and 0 ≤ j ≤ W , the entry V [i, j] we will store the maximum value of any subset of objects {1, 2, . . . , i} that can fit into a knapsack of weight j. If we can compute all the entries of this array, then the array entry V [n, W ] will contain the maximum value of all n objects that can fit into the entire knapsack of weight W . To compute the entries of the array V we will imply an inductive approach. As a basis, observe that V [0, j] = 0 for 0 ≤ j ≤ W since if we have no items then we have no value. We consider two cases: Leave object i: If we choose to not take object i, then the optimal value will come about by considering how to fill a knapsack of size j with the remaining objects {1, 2, . . . , i − 1}. This is just V [i − 1, j]. Take object i: If we take object i, then we gain a value of vi but have used up wi of our capacity. With the remaining j −wi capacity in the knapsack, we can fill it in the best possible way with objects {1, 2, . . . , i− 1}. This is vi + V [i − 1, j − wi ]. This is only possible if wi ≤ j. Since these are the only two possibilities, we can see that we have the following rule for constructing the array V . The ranges on i and j are i ∈ [0..n] and j ∈ [0..W ]. V [0, j] = 0 V [i − 1, j] if wi > j V [i, j] = max(V [i − 1, j], vi + V [i − 1, j − wi ]) if wi ≤ j The first line states that if there are no objects, then there is no value, irrespective of j. The second line implements the rule above. It is very easy to take these rules an produce an algorithm that computes the maximum value for the knapsack in time proportional to the size of the array, which is O((n + 1)(W + 1)) = O(nW ). The algorithm is given below. An example is shown in the figure below. The final output is V [n, W ] = V [4, 10] = 90. This reflects the selection of items 2 and 4, of values $40 and $50, respectively and weights 4 + 3 ≤ 10. The only missing detail is what items should we select to achieve the maximum. We will leave this as an exercise. They key is to record for each entry V [i, j] in the matrix whether we got this entry by taking the ith item or leaving it. With this information, it is possible to reconstruct the optimum knapsack contents. Lecture Notes 112 CMSC 451
- 113. 0-1 Knapsack Problem KnapSack(v[1..n], w[1..n], n, W) { allocate V[0..n][0..W]; for j = 0 to W do V[0, j] = 0; // initialization for i = 1 to n do { for j = 0 to W do { leave_val = V[i-1, j]; // total value if we leave i if (j >= w[i]) // enough capacity to take i take_val = v[i] + V[i-1, j - w[i]]; // total value if we take i else take_val = -INFINITY; // cannot take i V[i,j] = max(leave_val, take_val); // final value is max } } return V[n, W]; } Values of the objects are 10, 40, 30, 50 . Weights of the objects are 5, 4, 6, 3 . Capacity → j=0 1 2 3 4 5 6 7 8 9 10 Item Value Weight 0 0 0 0 0 0 0 0 0 0 0 1 10 5 0 0 0 0 0 10 10 10 10 10 10 2 40 4 0 0 0 0 40 40 40 40 40 50 50 3 30 6 0 0 0 0 40 40 40 40 40 50 70 4 50 3 0 0 0 50 50 50 50 90 90 90 90 Final result is V [4, 10] = 90 (for taking items 2 and 4). Fig. 73: 0–1 Knapsack Example. Lecture Notes 113 CMSC 451
- 114. Supplemental Lecture 8:Dynamic Programming: Memoization Read: Section 15.3 of CLRS. Recursive Implementation: We have described dynamic programming as a method that involves the “bottom-up” computation of a table. However, the recursive formulations that we have derived have been set up in a “top- down” manner. Must the computation proceed bottom-up? Consider the following recursive implementation of the chain-matrix multiplication algorithm. The call Rec-Matrix-Chain(p, i, j) computes and returns the value of m[i, j]. The initial call is Rec-Matrix-Chain(p, 1, n). We only consider the cost here. Recursive Chain Matrix Multiplication Rec-Matrix-Chain(array p, int i, int j) { if (i == j) m[i,j] = 0; // basis case else { m[i,j] = INFINITY; // initialize for k = i to j-1 do { // try all splits cost = Rec-Matrix-Chain(p, i, k) + Rec-Matrix-Chain(p, k+1, j) + p[i-1]*p[k]*p[j]; if (cost < m[i,j]) m[i,j] = cost; // update if better } } return m[i,j]; // return final cost } (Note that the table m[1..n, 1..n] is not really needed. We show it just to make the connection with the earlier version clearer.) This version of the procedure certainly looks much simpler, and more closely resembles the recursive formulation that we gave previously for this problem. So, what is wrong with this? The answer is the running time is much higher than the Θ(n3 ) algorithm that we gave before. In fact, we will see that its running time is exponential in n. This is unacceptably slow. Let T (n) denote the running time of this algorithm on a sequence of matrices of length n. (That is, n = j −i+1.) If i = j then we have a sequence of length 1, and the time is Θ(1). Otherwise, we do Θ(1) work and then consider all possible ways of splitting the sequence of length n into two sequences, one of length k and the other of length n − k, and invoke the procedure recursively on each one. So we get the following recurrence, defined for n ≥ 1. (We have replaced the Θ(1)’s with the constant 1.) 1 if n = 1, T (n) = n−1 1+ k=1 (T (k) + T (n − k)) if n ≥ 2. Claim: T (n) ≥ 2n−1 . Proof: The proof is by induction on n. Clearly this is true for n = 1, since T (1) = 1 = 20 . In general, for n ≥ 2, the induction hypothesis is that T (m) ≥ 2m−1 for all m < n. Using this we have n−1 n−1 T (n) = 1+ (T (k) + T (n − k)) ≥ 1 + T (k) k=1 k=1 n−1 n−2 ≥ 1+ 2k−1 = 1 + 2k k=1 k=0 = 1 + (2 n−1 − 1) = 2 n−1 . In the first line we simply ignored the T (n−k) term, in the second line we applied the induction hypothesis, and in the last line we applied the formula for the geometric series. Lecture Notes 114 CMSC 451
- 115. Why is thisso much worse than the dynamic programming version? If you “unravel” the recursive calls on a reasonably long example, you will see that the procedure is called repeatedly with the same arguments. The bottom-up version evaluates each entry exactly once. Memoization: Is it possible to retain the nice top-down structure of the recursive solution, while keeping the same O(n3 ) efficiency of the bottom-up version? The answer is yes, through a technique called memoization. Here is the idea. Let’s reconsider the function Rec-Matrix-Chain() given above. It’s job is to compute m[i, j], and return its value. As noted above, the main problem with the procedure is that it recomputes the same entries over and over. So, we will fix this by allowing the procedure to compute each entry exactly once. One way to do this is to initialize every entry to some special value (e.g. UNDEFINED). Once an entries value has been computed, it is never recomputed. Memoized Chain Matrix Multiplication Mem-Matrix-Chain(array p, int i, int j) { if (m[i,j] != UNDEFINED) return m[i,j]; // already defined else if (i == j) m[i,j] = 0; // basis case else { m[i,j] = INFINITY; // initialize for k = i to j-1 do { // try all splits cost = Mem-Matrix-Chain(p, i, k) + Mem-Matrix-Chain(p, k+1, j) + p[i-1]*p[k]*p[j]; if (cost < m[i,j]) m[i,j] = cost; // update if better } } return m[i,j]; // return final cost } This version runs in O(n3 ) time. Intuitively, this is because each of the O(n2 ) table entries is only computed once, and the work needed to compute one table entry (most of it in the for-loop) is at most O(n). Memoization is not usually used in practice, since it is generally slower than the bottom-up method. However, in some DP problems, many of the table entries are simply not needed, and so bottom-up computation may compute entries that are never needed. In these cases memoization may be a good idea. If you have know that most of the table will not be needed, here is a way to save space. Rather than storing the whole table explicitly as an array, you can store the “defined” entries of the table in a hash table, using the index pair (i, j) as the hash key. (See Chapter 11 in CLRS for more information on hashing.) Supplemental Lecture 9: Articulation Points and Biconnectivity Read: This material is not covered in CLR (except as Problem 23–2). Articulation Points and Biconnected Graphs: Today we discuss another application of DFS, this time to a problem on undirected graphs. Let G = (V, E) be a connected undirected graph. Consider the following definitions. Articulation Point (or Cut Vertex): Is any vertex whose removal (together with the removal of any incident edges) results in a disconnected graph. Bridge: Is an edge whose removal results in a disconnected graph. Biconnected: A graph is biconnected if it contains no articulation points. (In general a graph is k-connected, if k vertices must be removed to disconnect the graph.) Biconnected graphs and articulation points are of great interest in the design of network algorithms, because these are the “critical” points, whose failure will result in the network becoming disconnected. Lecture Notes 115 CMSC 451
- 116. a e i a a e e i b f j b f j c g c c g Biconnected Articulation point components Bridge d h d h Fig. 74: Articulation Points and Bridges Last time we observed that the notion of mutual reachability partitioned the vertices of a digraph into equivalence classes. We would like to do the same thing here. We say that two edges e1 and e2 are cocyclic if either e1 = e2 or if there is a simple cycle that contains both edges. It is not too hard to verify that this defines an equivalence relation on the edges of a graph. Notice that if two edges are cocyclic, then there are essentially two different ways of getting from one edge to the other (by going around the the cycle each way). Biconnected components: The biconnected components of a graph are the equivalence classes of the cocylicity relation. Notice that unlike strongly connected components of a digraph (which form a partition of the vertex set) the biconnected components of a graph form a partition of the edge set. You might think for a while why this is so. We give an algorithm for computing articulation points. An algorithm for computing bridges is simple modifi- cation to this procedure. Articulation Points and DFS: In order to determine the articulation points of an undirected graph, we will call depth- first search, and use the tree structure provided by the search to aid us. In particular, let us ask ourselves if a vertex u is an articulation point, how would we know it by its structure in the DFS tree? We assume that G is connected (if not, we can apply this algorithm to each individual connected component). So we assume is only one tree in the DFS forest. Because G is undirected, the DFS tree has a simpler structure. First off, we cannot distinguish between forward edges and back edges, and we just call them back edges. Also, there are no cross edges. (You should take a moment to convince yourself why this is true.) For now, let us consider the typical case of a vertex u, where u is not a leaf and u is not the root. Let’s let v1 , v2 , . . . , vk be the children of u. For each child there is a subtree of the DFS tree rooted at this child. If for some child, there is no back edge going to a proper ancestor of u, then if we were to remove u, this subtree would become disconnected from the rest of the graph, and hence u is an articulation point. On the other hand, if every one of the subtrees rooted at the children of u have back edges to proper ancestors of u, then if u is removed, the graph remains connected (the backedges hold everything together). This leads to the following. Observation 1: An internal vertex u of the DFS tree (other than the root) is an articulation point if and only there exists a subtree rooted at a child of u such that there is no back edge from any vertex in this subtree to a proper ancestor of u. Please check this condition carefully to see that you understand it. In particular, notice that the condition for whether u is an articulation point depends on a test applied to its children. This is the most common source of confusion for this algorithm. What about the leaves? If u is a leaf, can it be an articulation point? Answer: No, because when you delete a leaf from a tree, the rest of the tree remains connected, thus even ignoring the back edges, the graph is connected after the deletion of a leaf from the DFS tree. Lecture Notes 116 CMSC 451
- 117. v Low[u]=d[v] u Fig. 75: Articulation Points and DFS Observation 2: A leaf of the DFS tree is never an articulation point. Note that this is completely consistent with Observation 1, since a leaf will not have any subtrees in the DFS tree, so we can delete the word “internal” from Observation 1. What about the root? Since there are no cross edges between the subtrees of the root if the root has two or more children then it is an articulation point (since its removal separates these two subtrees). On the other hand, if the root has only a single child, then (as in the case of leaves) its removal does not disconnect the DFS tree, and hence cannot disconnect the graph in general. Observation 3: The root of the DFS is an articulation point if and only if it has two or more children. Articulation Points by DFS: Observations 1, 2, and 3 provide us with a structural characterization of which vertices in the DFS tree are articulation points. How can we design an algorithm which tests these conditions? Checking that the root has multiple children is an easy exercise. Checking Observation 1 is the hardest, but we will exploit the structure of the DFS tree to help us. The basic thing we need to check for is whether there is a back edge from some subtree to an ancestor of a given vertex. How can we do this? It would be too expensive to keep track of all the back edges from each subtree (because there may be Θ(e) back edges. A simpler scheme is to keep track of back edge that goes highest in the tree (in the sense of going closest to the root). If any back edge goes to an ancestor of u, this one will. How do we know how close a back edge goes to the root? As we travel from u towards the root, observe that the discovery times of these ancestors of u get smaller and smaller (the root having the smallest discovery time of 1). So we keep track of the back edge (v, w) that has the smallest value of d[w]. Low: Define Low[u] to be the minimum of d[u] and {d[w] | where (v, w) is a back edge and v is a descendent of u}. The term “descendent” is used in the nonstrict sense, that is, v may be equal to u. Intuitively, Low[u] is the highest (closest to the root) that you can get in the tree by taking any one backedge from either u or any of its descendents. (Beware of this notation: “Low” means low discovery time, not low in the tree. In fact Low[u] tends to be “high” in the tree, in the sense of being close to the root.) To compute Low[u] we use the following simple rules: Suppose that we are performing DFS on the vertex u. Initialization: Low[u] = d[u]. Back edge (u, v): Low[u] = min(Low[u], d[v]). Explanation: We have detected a new back edge coming out of u. If this goes to a lower d value than the previous back edge then make this the new low. Tree edge (u, v): Low[u] = min(Low[u], Low[v]). Explanation: Since v is in the subtree rooted at u any single back edge leaving the tree rooted at v is a single back edge for the tree rooted at u. Lecture Notes 117 CMSC 451
- 118. Observe that onceLow[u] is computed for all vertices u, we can test whether a given nonroot vertex u is an articulation point by Observation 1 as follows: u is an articulation point if and only if it has a child v in the DFS tree for which Low[v] ≥ d[u] (since if there were a back edge from either v or one of its descendents to an ancestor of v then we would have Low[v] < d[u]). The Final Algorithm: There is one subtlety that we must watch for in designing the algorithm (in particular this is true for any DFS on undirected graphs). When processing a vertex u, we need to know when a given edge (u, v) is a back edge. How do we do this? An almost correct answer is to test whether v is colored gray (since all gray vertices are ancestors of the current vertex). This is not quite correct because v may be the parent of v in the DFS tree and we are just seeing the “other side” of the tree edge between v and u (recalling that in constructing the adjacency list of an undirected graph we create two directed edges for each undirected edge). To test correctly for a back edge we use the predecessor pointer to check that v is not the parent of u in the DFS tree. The complete algorithm for computing articulation points is given below. The main procedure for DFS is the same as before, except that it calls the following routine rather than DFSvisit(). Articulation Points ArtPt(u) { color[u] = gray Low[u] = d[u] = ++time for each (v in Adj(u)) { if (color[v] == white) { // (u,v) is a tree edge pred[v] = u ArtPt(v) Low[u] = min(Low[u], Low[v]) // update Low[u] if (pred[u] == NULL) { // root: apply Observation 3 if (this is u’s second child) Add u to set of articulation points } else if (Low[v] >= d[u]) { // internal node: apply Observation 1 Add u to set of articulation points } } else if (v != pred[u]) { // (u,v) is a back edge Low[u] = min(Low[u], d[v]) // update L[u] } } } An example is shown in the following figure. As with all DFS-based algorithms, the running time is Θ(n + e). There are some interesting problems that we still have not discussed. We did not discuss how to compute the bridges of a graph. This can be done by a small modification of the algorithm above. We’ll leave it as an exercise. (Notice that if {u, v} is a bridge then it does not follow that u and v are both articulation points.) Another question is how to determine which edges are in the biconnected components. A hint here is to store the edges in a stack as you go through the DFS search. When you come to an articulation point, you can show that all the edges in the biconnected component will be consecutive in the stack. Supplemental Lecture 10: Bellman-Ford Shortest Paths Read: Section 24.1 in CLRS. Bellman-Ford Algorithm: We saw that Dijkstra’s algorithm can solve the single-source shortest path problem, under the assumption that the edge weights are nonnegative. We also saw that shortest paths are undefined if you Lecture Notes 118 CMSC 451
- 119. d=1 Low=1 a e i a 2 1 8 8 b f j b e 3 1 9 8 c g c i 4 3 10 d h d 7 f 1 j 8 5 3 h = articulation pt. 6 g 3 Fig. 76: Articulation Points. have cycles of total negative cost. What if you have negative edge weights, but no negative cost cycles? We shall present the Bellman-Ford algorithm, which solves this problem. This algorithm is slower that Dijkstra’s algorithm, running in Θ(V E) time. In our version we will assume that there are no negative cost cycles. The one presented in CLRS actually contains a bit of code that checks for this. (Check it out.) Recall that we are given a graph G = (V, E) with numeric edge weights, w(u, v). Like Dijkstra’s algorithm, the Bellman-Ford algorithm is based on performing repeated relaxations. (Recall that relaxation updates shortest path information along a single edge. It was described in our discussion of Dijkstra’s algorithm.) Dijkstra’s algorithm was based on the idea of organizing the relaxations in the best possible manner, namely in increasing order of distance. Once relaxation is applied to an edge, it need never be relaxed again. This trick doesn’t seem to work when dealing with graphs with negative edge weights. Instead, the Bellman-Ford algorithm simply applies a relaxation to every edge in the graph, and repeats this V − 1 times. Bellman-Ford Algorithm BellmanFord(G,w,s) { for each (u in V) { // standard initialization d[u] = +infinity pred[u] = null } d[s] = 0 for i = 1 to V-1 { // repeat V-1 times for each (u,v) in E { // relax along each edge Relax(u,v) } } } The Θ(V E) running time is pretty obvious, since there are two main nested loops, one iterated V − 1 times and the other iterated E times. The interesting question is how and why it works. Correctness of Bellman-Ford: I like to think of the Bellman-Ford as a sort of “BubbleSort analogue” for shortest paths, in the sense that shortest path information is propagated sequentially along each shortest path in the graph. Consider any shortest path from s to some other vertex u: v0 , v1 , . . . , vk where v0 = s and vk = u. Since a shortest path will never visit the same vertex twice, we know that k ≤ V − 1, and hence the path consists of at most V − 1 edges. Since this is a shortest path we have δ(s, vi ) (the true shortest path cost from s to vi ) satisfies δ(s, vi ) = δ(s, vi−1 ) + w(vi−1 , vi ). Lecture Notes 119 CMSC 451
- 120. ? 8 8 8 8 8 8 8 0 −6 ? 0 −6 ? 0 −6 9 0 −6 7 4 5 4 5 4 5 4 5 ? 4 2 2 Initial configuration After 1st relaxation After 2nd relaxation After 3rd relaxation phase phase phase Fig. 77: Bellman-Ford Algorithm. We assert that after the ith pass of the “for-i” loop that d[vi ] = δ(s, vi ). The proof is by induction on i. Observe that after the initialization (pass 0) we have d[v1 ] = d[s] = 0. In general, prior to the ith pass through the loop, the induction hypothesis tells us that d[vi−1 ] = δ(s, vi−1 ). After the ith pass through the loop, we have done a relaxation on the edge (vi−1 , vi ) (since we do relaxations along all the edges). Thus after the ith pass we have d[vi ] ≤ d[vi−1 ] + w(vi−1 , vi ) = δ(s, vi−1 ) + w(vi−1 , vi ) = δ(s, vi ). Recall from Dijkstra’s algorithm that d[vi ] is never less than δ(s, vi ) (since each time we do a relaxation there exists a path that witnesses its value). Thus, d[vi ] is in fact equal to δ(s, vi ), completing the induction proof. In summary, after i passes through the for loop, all vertices that are i edges away (along the shortest path tree) from the source have the correct distance values stored in d[u]. Thus, after the (V − 1)st iteration of the for loop, all vertices u have the correct distance values stored in d[u]. Supplemental Lecture 11: Network Flows and Matching Read: Chapt 27 in CLR. Maximum Flow: The Max Flow problem is one of the basic problems of algorithm design. Intuitively we can think of a flow network as a directed graph in which fluid is flowing along the edges of the graph. Each edge has certain maximum capacity that it can carry. The idea is to find out how much flow we can push from one point to another. The max flow problem has applications in areas like transportation, routing in networks. It is the simplest problem in a line of many important problems having to do with the movement of commodities through a network. These are often studied in business schools, and operations research. Flow Networks: A flow network G = (V, E) is a directed graph in which each edge (u, v) ∈ E has a nonegative capacity c(u, v) ≥ 0. If (u, v) ∈ E we model this by setting c(u, v) = 0. There are two special vertices: a source s, and a sink t. We assume that every vertex lies on some path from the source to the sink (for otherwise the vertex is of no use to us). (This implies that the digraph is connected, and hence e ≥ n − 1.) A flow is a real valued function on pairs of vertices, f : V × V → R which satisfies the following three properties: Capacity Constraint: For all u, v ∈ V , f (u, v) ≤ c(u, v). Skew Symmetry: For all u, v ∈ V , f (u, v) = −f (v, u). (In other words, we can think of backwards flow as negative flow. This is primarily for making algebraic analysis easier.) Flow conservation: For all u ∈ V − {s, t}, we have f (u, v) = 0. v∈V Lecture Notes 120 CMSC 451
- 121. (Given skew symmetry,this is equivalent to saying, flow-in = flow-out.) Note that flow conservation does NOT apply to the source and sink, since we think of ourselves as pumping flow from s to t. Flow conservation means that no flow is lost anywhere else in the network, thus the flow out of s will equal the flow into t. The quantity f (u, v) is called the net flow from u to v. The total value of the flow f is defined as |f | = f (s, v) v∈V i.e. the flow out of s. It turns out that this is also equal to v∈V f (v, t), the flow into t. We will show this later. The maximum-flow problem is, given a flow network, and source and sink vertices s and t, find the flow of maximum value from s to t. Example: Page 581 of CLR. Multi-source, multi-sink flow problems: It may seem overly restrictive to require that there is only a single source and a single sink vertex. Many flow problems have situations in which many source vertices s1 , s2 , . . . , sk and many sink vertices t1 , t2 , . . . , tl . This can easily be modelled by just adding a special supersource s and a supersink t , and attaching s to all the si and attach all the tj to t . We let these edges have infinite capacity. Now by pushing the maximum flow from s to t we are effectively producing the maximum flow from all the si to all the tj ’s. Note that we don’t care which flow from one source goes to another sink. If you require that the flow from source i goes ONLY to sink i, then you have a tougher problem called the multi-commodity flow problem. Set Notation: Sometimes rather than talking about the flow from a vertex u to a vertex v, we want to talk about the flow from a SET of vertices X to another SET of vertices Y . To do this we extend the definition of f to sets by defining f (X, Y ) = y ∈ Y f (x, y). x∈X Using this notation we can define flow balance for a vertex u more succintly by just writing f (u, V ) = 0. One important special case of this concept is when X and Y define a cut (i.e. a partition of the vertex set into two disjoint subsets X ⊆ V and Y = V − X). In this case f (X, Y ) can be thought of as the net amount of flow crossing over the cut. From simple manipulations of the definition of flow we can prove the following facts. Lemma: (i) f (X, X) = 0. (ii) f (X, Y ) = −f (Y, X). (iii) If X ∩ Y = ∅ then f (X ∪ Y, Z) = f (X, Z) + f (Y, Z) and f (Z, X ∪ Y ) = f (Z, X) + f (Z, Y ). Ford-Fulkerson Method: The most basic concept on which all network-flow algorithms work is the notion of aug- menting flows. The idea is to start with a flow of size zero, and then incrementally make the flow larger and larger by finding a path along which we can push more flow. A path in the network from s to t along which more flow can be pushed is called an augmenting path. This idea is given by the most simple method for computing network flows, called the Ford-Fulkerson method. Almost all network flow algorithms are based on this simple idea. They only differ in how they decide which path or paths along which to push flow. We will prove that when it is impossible to “push” any more flow through the network, we have reached the maximum possible flow (i.e. a locally maximum flow is globally maximum). Lecture Notes 121 CMSC 451
- 122. Ford-Fulkerson Network Flow FordFulkerson(G, s, t) { initialize flow f to 0; while (there exists an augmenting path p) { augment the flow along p; } output the final flow f; } Residual Network: To define the notion of an augmenting path, we first define the notion of a residual network. Given a flow network G and a flow f , define the residual capacity of a pair u, v ∈ V to be cf (u, v) = c(u, v)−f (u, v). Because of the capacity constraint, cf (u, v) ≥ 0. Observe that if cf (u, v) > 0 then it is possible to push more flow through the edge (u, v). Otherwise we say that the edge is saturated. The residual network is the directed graph Gf with the same vertex set as G but whose edges are the pairs (u, v) such that cf (u, v) > 0. Each edge in the residual network is weighted with its residual capacity. Example: Page 589 of CLR. Lemma: Let f be a flow in G and let f be a flow in Gf . Then (f + f ) (defined (f + f )(u, v) = f (u, v) + f (u, v)) is a flow in G. The value of the flow is |f | + |f |. Proof: Basically the residual network tells us how much additional flow we can push through G. This implies that f + f never exceeds the overall edge capacities of G. The other rules for flows are easy to verify. Augmenting Paths: An augmenting path is a simple path from s to t in Gf . The residual capacity of the path is the MINIMUM capacity of any edge on the path. It is denoted cf (p). Observe that by pushing cf (p) units of flow along each edge of the path, we get a flow in Gf , and hence we can use this to augment the flow in G. (Remember that when defining this flow that whenever we push cf (p) units of flow along any edge (u, v) of p, we have to push −cf (p) units of flow along the reverse edge (v, u) to maintain skew-symmetry. Since every edge of the residual network has a strictly positive weight, the resulting flow is strictly larger than the current flow for G. In order to determine whether there exists an augmenting path from s to t is an easy problem. First we construct the residual network, and then we run DFS or BFS on the residual network starting at s. If the search reaches t then we know that a path exists (and can follow the predecessor pointers backwards to reconstruct it). Since DFS and BFS take Θ(n + e) time, and it can be shown that the residual network has Θ(n + e) size, the running time of Ford-Fulkerson is basically Θ((n + e)(number of augmenting stages)). Later we will analyze the latter quantity. Correctness: To establish the correctness of the Ford-Fulkerson algorithm we need to delve more deeply into the theory of flows and cuts in networks. A cut, (S, T ), in a flow network is a partition of the vertex set into two disjoint subsets S and T such that s ∈ S and t ∈ T . We define the flow across the cut as f (S, T ), and we define the capcity of the cut as c(S, T ). Note that in computing f (S, T ) flows from T to S are counted negatively (by skew-symmetry), and in computing c(S, T ) we ONLY count constraints on edges leading from S to T ignoring those from T to S). Lemma: The amount of flow across any cut in the network is equal to |f |. Lecture Notes 122 CMSC 451
- 123. Proof: f (S, T ) = f (S, V ) − f (S, S) = f (S, V ) = f (s, V ) + f (S − s, V ) = f (s, V ) = |f | (The fact that f (S − s, V ) = 0 comes from flow conservation. f (u, V ) = 0 for all u other than s and t, and since S − s is formed of such vertices the sum of their flows will be zero also.) Corollary: The value of any flow is bounded from above by the capacity of any cut. (i.e. Maximum flow ≤ Minimum cut). Proof: You cannot push any more flow through a cut than its capacity. The correctness of the Ford-Fulkerson method is based on the following theorem, called the Max-Flow, Min-Cut Theorem. It basically states that in any flow network the minimum capacity cut acts like a bottleneck to limit the maximum amount of flow. Ford-Fulkerson algorithm terminates when it finds this bottleneck, and hence it finds the minimum cut and maximum flow. Max-Flow Min-Cut Theorem: The following three conditions are equivalent. (i) f is a maximum flow in G, (ii) The residual network Gf contains no augmenting paths, (iii) |f | = c(S, T ) for some cut (S, T ) of G. Proof: (i) ⇒ (ii): If f is a max flow and there were an augmenting path in Gf , then by pushing flow along this path we would have a larger flow, a contradiction. (ii) ⇒ (iii): If there are no augmenting paths then s and t are not connected in the residual network. Let S be those vertices reachable from s in the residual network and let T be the rest. (S, T ) forms a cut. Because each edge crossing the cut must be saturated with flow, it follows that the flow across the cut equals the capacity of the cut, thus |f | = c(S, T ). (iii) ⇒ (i): Since the flow is never bigger than the capacity of any cut, if the flow equals the capacity of some cut, then it must be maximum (and this cut must be minimum). Analysis of the Ford-Fulkerson method: The problem with the Ford-Fulkerson algorithm is that depending on how it picks augmenting paths, it may spend an inordinate amount of time arriving a the final maximum flow. Con- sider the following example (from page 596 in CLR). If the algorithm were smart enough to send flow along the edges of weight 1,000,000, the algorithm would terminate in two augmenting steps. However, if the algo- rithm were to try to augment using the middle edge, it will continuously improve the flow by only a single unit. 2,000,000 augmenting will be needed before we get the final flow. In general, Ford-Fulkerson can take time Θ((n + e)|f ∗ |) where f ∗ is the maximum flow. An Improvement: We have shown that if the augmenting path was chosen in a bad way the algorithm could run for a very long time before converging on the final flow. It seems (from the example we showed) that a more logical way to push flow is to select the augmenting path which holds the maximum amount of flow. Computing this path is equivalent to determining the path of maximum capacity from s to t in the residual network. (This is exactly the same as the beer transport problem given on the last exam.) It is not known how fast this method works in the worst case, but there is another simple strategy that is guaranteed to give good bounds (in terms of n and e). Lecture Notes 123 CMSC 451
- 124. Edmonds-Karp Algorithm: TheEdmonds-Karp algorithm is Ford-Fulkerson, with one little change. When finding the augmenting path, we use Breadth-First search in the residual network, starting at the source s, and thus we find the shortest augmenting path (where the length of the path is the number of edges on the path). We claim that this choice is particularly nice in that, if we do so, the number of flow augmentations needed will be at most O(e · n). Since each augmentation takes O(n + e) time to compute using BFS, the overall running time will be O((n + e)e · n) = O(n2 e + e2 n) ∈ O(e2 n) (under the reasonable assumption that e ≥ n). (The best known algorithm is essentially O(e · n log n). The fact that Edmonds-Karp uses O(en) augmentations is based on the following observations. Observation: If the edge (u, v) is an edge on the minimum length augmenting path from s to t in Gf , then δf (s, v) = δf (s, u) + 1. Proof: This is a simple property of shortest paths. Since there is an edge from u to v, δf (s, v) ≤ δf (s, u) + 1, and if δf (s, v) < δf (s, u) + 1 then u would not be on the shortest path from s to v, and hence (u, v) is not on any shortest path. Lemma: For each vertex u ∈ V −{s, t}, let δf (s, u) be the distance function from s to u in the residual network Gf . Then as we peform augmentations by the Edmonds-Karp algorithm the value of δf (s, u) increases monotonically with each flow augmentation. Proof: (Messy, but not too complicated. See the text.) Theorem: The Edmonds-Karp algorithm makes at most O(n · e) augmentations. Proof: An edge in the augmenting path is critical if the residual capacity of the path equals the residual capacity of this edge. In other words, after augmentation the critical edge becomes saturated, and disappears from the residual graph. How many times can an edge become critical before the algorithm terminates? Observe that when the edge (u, v) is critical it lies on the shortest augmenting path, implying that δf (s, v) = δf (s, u) + 1. After this it disappears from the residual graph. In order to reappear, it must be that we reduce flow on this edge, i.e. we push flow along the reverse edge (v, u). For this to be the case we have (at some later flow f ) δf (s, u) = δf (s, v) + 1. Thus we have: δf (s, u) = δf (s, v) + 1 ≥ δf (s, v) + 1 since dists increase with time = (δf (s, u) + 1) + 1 = δf (s, u) + 2. Thus, between the time that an edge becomes critical, its tail vertex increases in distance from the source by two. This can only happen n/2 times, since no vertex can be further than n from the source. Thus, each edge can become critical at most O(n) times, there are O(e) edges, hence after O(ne) augmentations, the algorithm must terminate. In summary, the Edmonds-Karp algorithm makes at most O(ne) augmentations and runs in O(ne2 ) time. Maximum Matching: One of the important elements of network flow is that it is a very general algorithm which is capable of solving many problems. (An example is problem 3 in the homework.) We will give another example here. Consider the following problem, you are running a dating service and there are a set of men L and a set of women R. Using a questionaire you establish which men are compatible which which women. Your task is to pair up as many compatible pairs of men and women as possible, subject to the constraint that each man is paired with at most one woman, and vice versa. (It may be that some men are not paired with any woman.) This problem is modelled by giving an undirected graph whose vertex set is V = L ∪ R and whose edge set consists of pairs (u, v), u ∈ L, v ∈ R such that u and v are compatible. The problem is to find a matching, Lecture Notes 124 CMSC 451
- 125. that is asubset of edges M such that for each v ∈ V , there is at most one edge of M incident to v. The desired matching is the one that has the maximum number of edges, and is called a maximum matching. Example: See page 601 in CLR. The resulting undirected graph has the property that its vertex set can be divided into two groups such that all its edges go from one group to the other (never within a group, unless the dating service is located on Dupont Circle). This problem is called the maximum bipartite matching problem. Reduction to Network Flow: We claim that if you have an algorithm for solving the network flow problem, then you can use this algorithm to solve the maximum bipartite matching problem. (Note that this idea does not work for general undirected graphs.) Construct a flow network G = (V , E ) as follows. Let s and t be two new vertices and let V = V ∪ {s, t}. E = {(s, u)|u ∈ L} ∪ {(v, t)|v ∈ R} ∪ {(u, v)|(u, v) ∈ E}. Set the capacity of all edges in this network to 1. Example: See page 602 in CLR. Now, compute the maximum flow in G . Although in general it can be that flows are real numbers, observe that the Ford-Fulkerson algorithm will only assign integer value flows to the edges (and this is true of all existing network flow algorithms). Since each vertex in L has exactly 1 incoming edge, it can have flow along at most 1 outgoing edge, and since each vertex in R has exactly 1 outgoing edge, it can have flow along at most 1 incoming edge. Thus letting f denote the maximum flow, we can define a matching M = {(u, v)|u ∈ L, v ∈ R, f (u, v) > 0}. We claim that this matching is maximum because for every matching there is a corresponding flow of equal value, and for every (integer) flow there is a matching of equal value. Thus by maximizing one we maximize the other. Supplemental Lecture 12: Hamiltonian Path Read: The reduction we present for Hamiltonian Path is completely different from the one in Chapt 36.5.4 of CLR. Hamiltonian Cycle: Today we consider a collection of problems related to finding paths in graphs and digraphs. Recall that given a graph (or digraph) a Hamiltonian cycle is a simple cycle that visits every vertex in the graph (exactly once). A Hamiltonian path is a simple path that visits every vertex in the graph (exactly once). The Hamiltonian cycle (HC) and Hamiltonian path (HP) problems ask whether a given graph (or digraph) has such a cycle or path, respectively. There are four variations of these problems depending on whether the graph is directed or undirected, and depending on whether you want a path or a cycle, but all of these problems are NP-complete. An important related problem is the traveling salesman problem (TSP). Given a complete graph (or digraph) with integer edge weights, determine the cycle of minimum weight that visits all the vertices. Since the graph is complete, such a cycle will always exist. The decision problem formulation is, given a complete weighted graph G, and integer X, does there exist a Hamiltonian cycle of total weight at most X? Today we will prove that Hamiltonian Cycle is NP-complete. We will leave TSP as an easy exercise. (It is done in Section 36.5.5 in CLR.) Lecture Notes 125 CMSC 451
- 126. Component Design: Upto now, most of the reductions that we have seen (for Clique, VC, and DS in particular) are of a relatively simple variety. They are sometimes called local replacement reductions, because they operate by making some local change throughout the graph. We will present a much more complex style of reduction for the Hamiltonian path problem on directed graphs. This type of reduction is called a component design reduction, because it involves designing special subgraphs, sometimes called components or gadgets (also called widgets). whose job it is to enforce a particular constraint. Very complex reductions may involve the creation of many gadgets. This one involves the construction of only one. (See CLR’s presentation of HP for other examples of gadgets.) The gadget that we will use in the directed Hamiltonian path reduction, called a DHP-gadget, is shown in the figure below. It consists of three incoming edges labeled i1 , i2 , i3 and three outgoing edges, labeled o1 , o2 , o3 . It was designed so it satisfied the following property, which you can verify. Intuitively it says that if you enter the gadget on any subset of 1, 2 or 3 input edges, then there is a way to get through the gadget and hit every vertex exactly once, and in doing so each path must end on the corresponding output edge. Claim: Given the DHP-gadget: • For any subset of input edges, there exists a set of paths which join each input edge i1 , i2 , or i3 to its respective output edge o1 , o2 , or o3 such that together these paths visit every vertex in the gadget exactly once. • Any subset of paths that start on the input edges and end on the output edges, and visit all the vertices of the gadget exactly once, must join corresponding inputs to corresponding outputs. (In other words, a path that starts on input i1 must exit on output o1 .) The proof is not hard, but involves a careful inspection of the gadget. It is probably easiest to see this on your own, by starting with one, two, or three input paths, and attempting to get through the gadget without skipping vertex and without visiting any vertex twice. To see whether you really understand the gadget, answer the question of why there are 6 groups of triples. Would some other number work? DHP is NP-complete: This gadget is an essential part of our proof that the directed Hamiltonian path problem is NP-complete. Theorem: The directed Hamiltonian Path problem is NP-complete. Proof: DHP ∈ NP: The certificate consists of the sequence of vertices (or edges) in the path. It is an easy matter to check that the path visits every vertex exactly once. 3SAT ≤P DHP: This will be the subject of the rest of this section. Let us consider the similar elements between the two problems. In 3SAT we are selecting a truth assignment for the variables of the formula. In DHP, we are deciding which edges will be a part of the path. In 3SAT there must be at least one true literal for each clause. In DHP, each vertex must be visited exactly once. We are given a boolean formula F in 3-CNF form (three literals per clause). We will convert this formula into a digraph. Let x1 , x2 , . . . , xm denote the variables appearing in F . We will construct one DHP-gadget for each clause in the formula. The inputs and outputs of each gadget correspond to the literals appearing in this clause. Thus, the clause (x2 ∨ x5 ∨ x8 ) would generate a clause gadget with inputs labeled x2 , x5 , and x8 , and the same outputs. The general structure of the digraph will consist of a series vertices, one for each variable. Each of these vertices will have two outgoing paths, one taken if xi is set to true and one if xi is set to false. Each of these paths will then pass through some number of DHP-gadgets. The true path for xi will pass through all the clause gadgets for clauses in which xi appears, and the false path will pass through all the gadgets for clauses in which xi appears. (The order in which the path passes through the gadgets is unimportant.) When the paths for xi have passed through their last gadgets, then they are joined to the next variable vertex, xi+1 . This is illustrated in the following figure. (The figure only shows a portion of the construction. There will be paths coming into Lecture Notes 126 CMSC 451
- 127. Gadget What it looks like inside i1 o1 i1 o1 i2 o2 i2 o2 i3 o3 i3 o3 Path with 1 entry i1 o1 i1 o1 i2 o2 i2 o2 i3 o3 i3 o3 Path with 2 entries i1 o1 i1 o1 i2 o2 i2 o2 i3 o3 i3 o3 Path with 3 entries i1 o1 i1 o1 i2 o2 i2 o2 i3 o3 i3 o3 Fig. 78: DHP-Gadget and examples of path traversals. Lecture Notes 127 CMSC 451
- 128. these same gadgetsfrom other variables as well.) We add one final vertex xe , and the last variable’s paths are connected to xe . (If we wanted to reduce to Hamiltonian cycle, rather than Hamiltonian path, we could join xe back to x1 .) xi xi ... xi xi xi xi xi xi xi _ _ _ _ xi+1 xi xi xi xi _ _ _ _ xi xi xi xi ... Fig. 79: General structure of reduction from 3SAT to DHP. Note that for each variable, the Hamiltonian path must either use the true path or the false path, but it cannot use both. If we choose the true path for xi to be in the Hamiltonian path, then we will have at least one path passing through each of the gadgets whose corresponding clause contains xi , and if we chose the false path, then we will have at least one path passing through each gadget for xi . For example, consider the following boolean formula in 3-CNF. The construction yields the digraph shown in the following figure. (x1 ∨ x2 ∨ x3 ) ∧ (x1 ∨ x2 ∨ x3 ) ∧ (x2 ∨ x1 ∨ x3 ) ∧ (x1 ∨ x3 ∨ x2 ). path starts here T x1 _ to x3 F x1 x1 _ x2 _ x1 x2 x3 to x2 x2 T x2 _ x _1 _ xe F x3 x3 x3 x2 to x3 T x3 to x2 F Fig. 80: Example of the 3SAT to DHP reduction. The Reduction: Let us give a more formal description of the reduction. Recall that we are given a boolean formula F in 3-CNF. We create a digraph G as follows. For each variable xi appearing in F , we create a variable vertex, named xi . We also create a vertex named xe (the ending vertex). For each clause c, we create a DHP-gadget whose inputs and outputs are labeled with the three literals of c. (The order is unimportant, as long as each input and its corresponding output are labeled the same.) We join these vertices with the gadgets as follows. For each variable xi , consider all the clauses c1 , c2 , . . . , ck in which xi appears as a literal (uncomplemented). Join xi by an edge to the input labeled with xi in the gadget for c1 , and in general join the the output of gadget cj labeled xi with the input of gadget cj+1 with this same label. Finally, join the output of the last gadget ck to the next vertex variable xi+1 . (If this is the last variable, then join it to xe instead.) The resulting chain of edges is called the true path for variable xi . Form a second chain in exactly the same way, but this time joining the gadgets for the clauses in which xi appears. This is called the false path for xi . The resulting digraph is the output of the reduction. Observe that the entire construction can be performed in polynomial time, by simply inspecting the formula, creating the appropriate vertices, and adding the appropriate edges to the digraph. The following lemma establishes the correctness of this reduction. Lemma: The boolean formula F is satisfiable if and only if the digraph G produced by the above reduction has a Hamiltonian path. Lecture Notes 128 CMSC 451
- 129. Start here T x1 to x3 x1 x2 x1 x2 to x2 x2 T _ _ xe x3 x3 x3 F A satisfying assignment hits all gadgets Start here x1 _ to x3 F x1 x2 _ x2 T x2 _ x _1 xe x3 x3 x3 to x2 F A nonsatisfying assignment misses some gadgets Fig. 81: Correctness of the 3SAT to DHP reduction. The upper figure shows the Hamiltonian path resulting from the satisfying assignment, x1 = 1, x2 = 1, x3 = 0, and the lower figure shows the non-Hamiltonian path resulting from the nonsatisfying assignment x1 = 0, x2 = 1, x3 = 0. Proof: We need to prove both the “only if” and the “if”. ⇒: Suppose that F has a satisfying assignment. We claim that G has a Hamiltonian path. This path will start at the variable vertex x1 , then will travel along either the true path or false path for x1 , depending on whether it is 1 or 0, respectively, in the assignment, and then it will continue with x2 , then x3 , and so on, until reaching xe . Such a path will visit each variable vertex exactly once. Because this is a satisfying assignment, we know that for each clause, either 1, 2, or 3 of its literals will be true. This means that for each clause, either 1, 2, or 3, paths will attempt to travel through the corresponding gadget. However, we have argued in the above claim that in this case it is possible to visit every vertex in the gadget exactly once. Thus every vertex in the graph is visited exactly once, implying that G has a Hamiltonian path. ⇐: Suppose that G has a Hamiltonian path. We assert that the form of the path must be essentially the same as the one described in the previous part of this proof. In particular, the path must visit the variable vertices in increasing order from x1 until xe , because of the way in which these vertices are joined together. Also observe that for each variable vertex, the path will proceed along either the true path or the false path. If it proceeds along the true path, set the corresponding variable to 1 and otherwise set it to 0. We will show that the resulting assignment is a satisfying assignment for F . Any Hamiltonian path must visit all the vertices in every gadget. By the above claim about DHP-gadgets, if a path visits all the vertices and enters along input edge then it must exit along the corresponding output edge. Therefore, once the Hamiltonian path starts along the true or false path for some variable, it must remain on edges with the same label. That is, if the path starts along the true path for xi , it must travel through all the gadgets with the label xi until arriving at the variable vertex for xi+1 . If it starts along the false path, then it must travel through all gadgets with the label xi . Since all the gadgets are visited and the paths must remain true to their initial assignments, it follows that for each corresponding clause, at least one (and possibly 2 or three) of the literals must be true. Therefore, this is a satisfying assignment. Lecture Notes 129 CMSC 451
- 130. Supplemental Lecture 13:Subset Sum Approximation Read: Section 37.4 in CLR. Polynomial Approximation Schemes: Last time we saw that for some NP-complete problems, it is possible to ap- proximate the problem to within a fixed constant ratio bound. For example, the approximation algorithm pro- duces an answer that is within a factor of 2 of the optimal solution. However, in practice, people would like to the control the precision of the approximation. This is done by specifying a parameter > 0 as part of the input to the approximation algorithm, and requiring that the algorithm produce an answer that is within a relative error of of the optimal solution. It is understood that as tends to 0, the running time of the algorithm will increase. Such an algorithm is called a polynomial approximation scheme. For example, the running time of the algorithm might be O(2(1/ ) n2 ). It is easy to see that in such cases the user pays a big penalty in running time as a function of . (For example, to produce a 1% error, the “constant” factor would be 2100 which would be around 4 quadrillion centuries on your 100 Mhz Pentium.) A fully polynomial approximation scheme is one in which the running time is polynomial in both n and 1/ . For example, a running time of O((n/ )2 ) would satisfy this condition. In such cases, reasonably accurate approximations are computationally feasible. Unfortunately, there are very few NP-complete problems with fully polynomial approximation schemes. In fact, recently there has been strong evidence that many NP-complete problems do not have polynomial approximation schemes (fully or otherwise). Today we will study one that does. Subset Sum: Recall that in the subset sum problem we are given a set S of positive integers {x1 , x2 , . . . , xn } and a target value t, and we are asked whether there exists a subset S ⊆ S that sums exactly to t. The optimization problem is to determine the subset whose sum is as large as possible but not larger than t. This problem is basic to many packing problems, and is indirectly related to processor scheduling problems that arise in operating systems as well. Suppose we are also given 0 < < 1. Let z ∗ ≤ t denote the optimum sum. The approximation problem is to return a value z ≤ t such that z ≥ z ∗ (1 − ). If we think of this as a knapsack problem, we want our knapsack to be within a factor of (1 − ) of being as full as possible. So, if = 0.1, then the knapsack should be at least 90% as full as the best possible. What do we mean by polynomial time here? Recall that the running time should be polynomial in the size of the input length. Obviously n is part of the input length. But t and the numbers xi could also be huge binary numbers. Normally we just assume that a binary number can fit into a word of our computer, and do not count their length. In this case we will to be on the safe side. Clearly t requires O(log t) digits to be store in the input. We will take the input size to be n + log t. Intuitively it is not hard to believe that it should be possible to determine whether we can fill the knapsack to within 90% of optimal. After all, we are used to solving similar sorts of packing problems all the time in real life. But the mental heuristics that we apply to these problems are not necessarily easy to convert into efficient algorithms. Our intuition tells us that we can afford to be a little “sloppy” in keeping track of exactly full the knapsack is at any point. The value of tells us just how sloppy we can be. Our approximation will do something similar. First we consider an exponential time algorithm, and then convert it into an approximation algorithm. Exponential Time Algorithm: This algorithm is a variation of the dynamic programming solution we gave for the knapsack problem. Recall that there we used an 2-dimensional array to keep track of whether we could fill a knapsack of a given capacity with the first i objects. We will do something similar here. As before, we will concentrate on the question of which sums are possible, but determining the subsets that give these sums will not be hard. Let Li denote a list of integers that contains the sums of all 2i subsets of {x1 , x2 , . . . , xi } (including the empty set whose sum is 0). For example, for the set {1, 4, 6} the corresponding list of sums contains 0, 1, 4, 5(= Lecture Notes 130 CMSC 451
- 131. 1 + 4),6, 7(= 1 + 6), 10(= 4 + 6), 11(= 1 + 4 + 6) . Note that Li can have as many as 2i elements, but may have fewer, since some subsets may have the same sum. There are two things we will want to do for efficiency. (1) Remove any duplicates from Li , and (2) only keep sums that are less than or equal to t. Let us suppose that we a procedure MergeLists(L1, L2) which merges two sorted lists, and returns a sorted lists with all duplicates removed. This is essentially the procedure used in MergeSort but with the added duplicate element test. As a bit of notation, let L + x denote the list resulting by adding the number x to every element of list L. Thus 1, 4, 6 + 3 = 4, 7, 9 . This gives the following procedure for the subset sum problem. Exact Subset Sum Exact_SS(x[1..n], t) { L = <0>; for i = 1 to n do { L = MergeLists(L, L+x[i]); remove for L all elements greater than t; } return largest element in L; } For example, if S = {1, 4, 6} and t = 8 then the successive lists would be L0 = 0 L1 = 0 ∪ 0 + 1 = 0, 1 L2 = 0, 1 ∪ 0 + 4, 1 + 4 = 0, 1, 4, 5 L3 = 0, 1, 4, 5 ∪ 0 + 6, 1 + 6, 4 + 6, 5 + 6 = 0, 1, 4, 5, 6, 7, 10, 11 . The last list would have the elements 10 and 11 removed, and the final answer would be 7. The algorithm runs in Ω(2n ) time in the worst case, because this is the number of sums that are generated if there are no duplicates, and no items are removed. Approximation Algorithm: To convert this into an approximation algorithm, we will introduce a “trim” the lists to decrease their sizes. The idea is that if the list L contains two numbers that are very close to one another, e.g. 91, 048 and 91, 050, then we should not need to keep both of these numbers in the list. One of them is good enough for future approximations. This will reduce the size of the lists that the algorithm needs to maintain. But, how much trimming can we allow and still keep our approximation bound? Furthermore, will we be able to reduce the list sizes from exponential to polynomial? The answer to both these questions is yes, provided you apply a proper way of trimming the lists. We will trim elements whose values are sufficiently close to each other. But we should define close in manner that is relative to the sizes of the numbers involved. The trimming must also depend on . We select δ = /n. (Why? We will see later that this is the value that makes everything work out in the end.) Note that 0 < δ < 1. Assume that the elements of L are sorted. We walk through the list. Let z denote the last untrimmed element in L, and let y ≥ z be the next element to be considered. If y−z ≤δ y then we trim y from the list. Equivalently, this means that the final trimmed list cannot contain two value y and z such that (1 − δ)y ≤ z ≤ y. We can think of z as representing y in the list. For example, given δ = 0.1 and given the list L = 10, 11, 12, 15, 20, 21, 22, 23, 24, 29 , Lecture Notes 131 CMSC 451
- 132. the trimmed listL will consist of L = 10, 12, 15, 20, 23, 29 . Another way to visualize trimming is to break the interval from [1, t] into a set of buckets of exponentially increasing size. Let d = 1/(1 − δ). Note that d > 1. Consider the intervals [1, d], [d, d2 ], [d2 , d3 ], . . . , [dk−1 , dk ] where dk ≥ t. If z ≤ y are in the same interval [di−1 , di ] then y−z di − di−1 1 ≤ i = 1 − = δ. y d d Thus, we cannot have more than one item within each bucket. We can think of trimming as a way of enforcing the condition that items in our lists are not relatively too close to one another, by enforcing the condition that no bucket has more than one item. 1 2 4 8 16 L L’ Fig. 82: Trimming Lists for Approximate Subset Sum. Claim: The number of distinct items in a trimmed list is O((n log t)/ ), which is polynomial in input size and 1/ . Proof: We know that each pair of consecutive elements in a trimmed list differ by a ratio of at least d = 1/(1 − δ) > 1. Let k denote the number of elements in the trimmed list, ignoring the element of value 0. Thus, the smallest nonzero value and maximum value in the the trimmed list differ by a ratio of at least dk−1 . Since the smallest (nonzero) element is at least as large as 1, and the largest is no larger than t, then it follows that dk−1 ≤ t/1 = t. Taking the natural log of both sides we have (k − 1) ln d ≤ ln t. Using the facts that δ = /n and the log identity that ln(1 + x) ≤ x, we have ln t ln t k−1 ≤ = ln d − ln(1 − δ) ln t n ln t ≤ = δ n log t k = O . Observe that the input size is at least as large as n (since there are n numbers) and at least as large as log t (since it takes log t digits to write down t on the input). Thus, this function is polynomial in the input size and 1/ . The approximation algorithm operates as before, but in addition we call the procedure Trim given below. For example, consider the set S = {104, 102, 201, 101} and t = 308 and = 0.20. We have δ = /4 = 0.05. Here is a summary of the algorithm’s execution. Lecture Notes 132 CMSC 451
- 133. Approximate Subset Sum Trim(L, delta) { let the elements of L be denoted y[1..m]; L’ = <y[1]>; // start with first item last = y[1]; // last item to be added for i = 2 to m do { if (last < (1-delta) y[i]) { // different enough? append y[i] to end of L’; last = y[i]; } } } Approx_SS(x[1..n], t, eps) { delta = eps/n; // approx factor L = <0>; // empty sum = 0 for i = 1 to n do { L = MergeLists(L, L+x[i]); // add in next item L = Trim(L, delta); // trim away "near" duplicates remove for L all elements greater than t; } return largest element in L; } init: L0 = 0 merge: L1 = 0, 104 trim: L1 = 0, 104 remove: L1 = 0, 104 merge: L2 = 0, 102, 104, 206 trim: L2 = 0, 102, 206 remove: L2 = 0, 102, 206 merge: L3 = 0, 102, 201, 206, 303, 407 trim: L3 = 0, 102, 201, 303, 407 remove: L3 = 0, 102, 201, 303 merge: L4 = 0, 101, 102, 201, 203, 302, 303, 404 trim: L4 = 0, 101, 201, 302, 404 remove: L4 = 0, 101, 201, 302 The final output is 302. The optimum is 307 = 104 + 102 + 101. So our actual relative error in this case is within 2%. The running time of the procedure is O(n|L|) which is O(n2 ln t/ ) by the earlier claim. Lecture Notes 133 CMSC 451
- 134. Approximation Analysis: Thefinal question is why the algorithm achieves an relative error of at most over the optimum solution. Let Y ∗ denote the optimum (largest) subset sum and let Y denote the value returned by the algorithm. We want to show that Y is not too much smaller than Y ∗ , that is, Y ≥ Y ∗ (1 − ). Our proof will make use of an important inequality from real analysis. Lemma: For n > 0 and a real numbers, a n (1 + a) ≤ 1 + ≤ ea . n Recall that our intuition was that we would allow a relative error of /n at each stage of the algorithm. Since the algorithm has n stages, then the total relative error should be (obviously?) n( /n) = . The catch is that these are relative, not absolute errors. These errors to not accumulate additively, but rather by multiplication. So we need to be more careful. Let L∗ denote the i-th list in the exponential time (optimal) solution and let Li denote the i-th list in the approx- i imate algorithm. We claim that for each y ∈ L∗ there exists a representative item z ∈ Li whose relative error i from y that satisfies (1 − /n)i y ≤ z ≤ y. The proof of the claim is by induction on i. Initially L0 = L∗ = 0 , and so there is no error. Suppose by 0 induction that the above equation holds for each item in L∗ . Consider an element y ∈ L∗ . We know that i−1 i−1 y will generate two elements in L∗ : y and y + xi . We want to argue that there will be a representative that is i “close” to each of these items. By our induction hypothesis, there is a representative element z in Li−1 such that (1 − /n)i−1 y ≤ z ≤ y. When we apply our algorithm, we will form two new items to add (initially) to Li : z and z + xi . Observe that by adding xi to the inequality above and a little simplification we get (1 − /n)i−1 (y + xi ) ≤ z + xi ≤ y + xi . zy L*i−1 L i−1 L* i Li z’ y z’’ y+xi z z+xi Fig. 83: Subset sum approximation analysis. The items z and z + xi might not appear in Li because they may be trimmed. Let z and z be their respective representatives. Thus, z and z are elements of Li . We have (1 − /n)z ≤z ≤ z (1 − /n)(z + xi ) ≤ z ≤ z + xi . Lecture Notes 134 CMSC 451
- 135. Combining these withthe inequalities above we have (1 − /n)i−1 (1 − /n)y ≤ (1 − /n)i y ≤ z ≤ y (1 − /n)i−1 (1 − /n)(y + xi ) ≤ (1 − /n)i (y + xi ) ≤ z ≤ z + yi . Since z and z are in Li this is the desired result. This ends the proof of the claim. Using our claim, and the fact that Y ∗ (the optimum answer) is the largest element of L∗ and Y (the approximate n answer) is the largest element of Ln we have (1 − /n)n Y ∗ ≤ Y ≤ Y ∗ . This is not quite what we wanted. We wanted to show that (1 − )Y ∗ ≤ Y . To complete the proof, we observe from the lemma above (setting a = − ) that n (1 − ) ≤ 1 − . n This completes the approximate analysis. Lecture Notes 135 CMSC 451