Scribd
Upload
199 views

Strassen

Strassen matrix multiplication

Uploaded by

amukhopadhyay
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
199 views

Strassen

Strassen matrix multiplication

Uploaded by

amukhopadhyay
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 8
42 Strassen's algorithm for matrix multiplication 7s ray is 0. How would you change any of the algorithms that do not allow empty subarrays to permit an empty subarray to be the result? 415 Use the following ideas to develop a nonrecursive, linear-time algorithm for the maximum-subarray problem. Start at the left end of the array, and progress toward the right, keeping track of the maximum subarray seen so far. Knowing a maximum, subarray of A[.. j]}, extend the answer to find a maximum subarray ending at in- dex j +1 by using the following observation: a maximum subarray of A[I.. 7 + 1] is either a maximum subarray of 4[I.. j] ora subarray Ali. j 4. 1), for some 1 =i integer. ‘Suppose that we partition each of A, B, and C into four n/2 x-n/2 matrices An An By By Cr Cn a(4 an): 8 ( ) c= ) (49) Bu Baa Cu Coa So that we rewrite the equation C = A- Bas (2 &)-Ce Cet) a0 Equation (4.10) corresponds to the four equations, and by assuming that n is an 2, the dimension /2 is an Cu = An+Bu | Ai+ Ba, (a1 C2 = Au Bis + Arr+ Bro. (4.12) Ca = Anis Bu + Are + Bar. (4.13) Cx = Anis Biz | Aza + Bo (4.14) Each of these four equations specifies two multiplications of n/2 x n/2 matrices and the addition of their n/2 x n/2 products. We can use these equations to create a straightforward, recursive, divide-and-conquer algorithm: 42 Strassen's algorithm for matrix multiplication ” ‘SQUARE-MATRIX-MULTIPLY-RECURSIVE(A, B) 1 n= A.rows 2. let C be anew m xn matrix 3 ifm 4 en =aurbn 5 else partition A, B, and C as in equations (4.9) 6 Ci = SQUARE-MATRIX-MULTIPLY-RECURSIVE(A}1, Bui) | SQUARE-MATRIX-MULTIPLY-RECURSIVE (Ay2, Boi) 7 C12 = SQUARE-MATRIX-MULTIPLY-RECURSIVE(A 1, By2) 4 SQUARE-MATRIX-MULTIPLY-RECURSIVE (Ay2, Bz2) 8 Cx, = SQUARE-MATRIX-MULTIPLY-RECURSIVE(Ap1, Bi) ++ SQUARE-MATRIX-MULTIPLY-RECURSIVE(Aza, Boy) 9 Co. = SQUARE-MATRIX-MULTIPLY-RECURSIVE(Ap1, Biz) ++ SQUARE-MATRIX-MULTIPLY-RECURSIVE( Aza, Boo) 10 return € This pseudocode glosses over one subtle but important implementation detail How do we partition the matrices in line 5? If we were to create 12 new n/2xn/2 matrices, we would spend @(n?) time copying entries. In fact, we can partition the matrices without copying entries. The trick is to use index calculations. We identify a submatrix by a range of row indices and a range of column indices of the original matrix. We end up representing a submatrix a little differently from how we represent the original matrix, which is the subtlety we are glossing over. ‘The advantage is that, since we can specify submatrices by index calculations, executing line 5 takes only @(1) time (although we shall see that it makes no difference asymptotically to the overall running time whether we copy or partition in place). Now, we derive a recurrence to characterize the running time of SQUARE- MareiX-MULTIPLY-RECURSIVE. Let T(n) be the time to multiply two m x matrices using this procedure. In the base case, when = 1, we perform just the ‘one scalar multiplication in line 4, and so TQ) = 04) (4.15) ‘The recursive case occurs when 1 > 1. As discussed, partitioning the matrices in line 5 takes @(1) time, using index calculations. In lines 6-9, we recursively call SQUARE-MATRIX-MULTIPLY-RECURSIVE @ total of eight times. Because each recursive call multiplies two n/2 x n/2 matrices, thereby contributing T(1r/2) to the overall running time, the time taken by all eight recursive calls is 87(1/2). We also must account for the four matrix additions in lines 6-9. Each of these matrices contains n?/4 entries, and so each of the four matrix additions takes @(n?) time. Since the number of matrix additions is a constant, the total time spent adding ma- 78 Chaprer 4 Divide-and-Conguer trices in lines 6-9 is ©(n?). (Again, we use index calculations to place the results of the matrix additions into the correct positions of matrix C, with an overhead of @(1) time per entry.) The total time for the recursive case, therefore, is the sum of the partitioning time, the time for all the recursive calls, and the time to add the matrices resulting from the recursive calls: Tin) = O01). 8T(H/2) | OW) = 87 (n/2) + O(n?) (4.16) Notice that if we implemented partitioning by copying matrices, which would cost (1?) time, the recurrence would not change, and hence the overall running time ‘would increase by only a constant factor. Combining equations (4.15) and (4.16) gives us the recurre time of SQUARE-MATRIX-MULTIPLY-RECURSIVE: for the running fou ifn =1, Tn) = o [87 (1/2)+ 0?) ifn>1 4.17 As we shall see from the master method in Section 4.5, recurrence (4.17) has the solution T(1) = @(n*). Thus, this simple divide-and-conquer approach is no faster than the straightforward SQUARE-MATRIX-MULTIPLY procedure. Before we continue on to examining Strassen’s algorithm, let us review where the components of equation (4.16) came from. Partitioning each n x matrix by index calculation takes ©(1) time, but we have two matrices to partition. Although you could say that partitioning the two matrices takes (2) time, the constant of 2 is subsumed by the @-notation. Adding two matrices, each with, say, k entries, takes @(k) time. Since the matrices we add each have n?/4 entries, you could say that adding each pair takes @(n?/4) time. Again, however, the @-notation subsumes the constant factor of 1/4, and we say that adding two n?/4 x n?/4 matrices takes @(n?) time. We have four such matrix additions, and once again, instead of saying that they take ©(4n) time, we say that they take @(n?) time. (OF course, you might observe that we could say that the four matrix additions take ©(4n7/4) time, and that 4n?/4 = n2, but the point here is that @-notation subsumes constant factors, whatever they are.) Thus, we end up with two terms of ©(n2), which we can combine into one. When we account for the eight recursive calls, however, we cannot just sub- sume the constant factor of 8. In other words, we must say that together they take 87(n/2) time, rather than just T(n/2) time. You can get a feel for why by looking back at the recursion tree in Figure 2.5, for recurrence (2.1) (which is identical to recurrence (4.7)), with the recursive case T(11) = 27 (n/2) +O (1). The factor of 2 determined how many children each tree node had, which in turn determined how ‘many terms contributed to the sum at each level of the tree. If we were to ignore 42 Strassen's algorithm for matrix multiplication 7 the factor of 8 in equation (4.16) or the factor of 2 in recurrence (4.1), the recursion tree would just be linear, rather than “bushy.” and each level would contribute only ‘one term to the sum. Bear in mind, therefore, that although asymptotic notation subsumes constant multiplicative factors, recursive notation such as T(n/2) does not. Strassen’s method ‘The key to Strassen’s method is to make the recursion tree slightly less bushy. ‘That is, instead of performing eight recursive multiplications of n/2 x n/2 matrices, it performs only seven. The cost of eliminating one matrix multiplication will be several new additions of n/2 x n/2 matrices, but still only a constant number of additions. As before, the constant number of matrix additions will be subsumed by @-notation when Wwe set up the recurrence equation to characterize the running time Strassen’s method is not at all obvious. (This might be the biggest understate ment in this book.) It has four steps: 1. Divide the input matrices A and B and output matrix C into n/2 xn/2 subma- trices, as in equation (4.9). This step takes ©(1) time by index calculation, just as in SQUARE-MATRIX-MULTIPLY-RECURSIVE. Create 10 matrices 5,,83,.... Sig, each of which is n/2 x n/2 and is the sum or difference of two matrices created in step 1. We can create all 10 matrices in O(n?) time. 3. Using the submatrices created in step 1 and the 10 matrices created in step 2, recursively compute seven matrix products P1, P2,..., Ps. Each matrix P; is nf2xn/2. 4. Compute the desired submatrices C,,, Ci2-C2y.Caz of the result matrix C by adding and subtracting Various combinations of the P; matrices. We can com- pute all four submatrices in @(n?) time. We shall see the details of steps 2-4 in a moment, but we already have enough information to set up a recurrence for the running time of Strassen’s method. Let us assume that once the matrix size n gets down to 1, we perform a simple scalar mul- tiplication, just as in line 4 of SQUARE-MATRIX-MULTIPLY-RECURSIVE. When n> 1, steps 1, 2, and 4 take a total of @(n?) time, and step 3 requires us to per- form seven multiplications of n/2.xn/2 matrices. Hence, we obtain the following recurrence for the running time T(x) of Strassen’s algorithm: _jew ifm=1 Twn J 77/2) +O?) ifn >1 (4.18) 80 Chaprer 4 Divide-and-Conguer We have traded off one matrix multiplication for a constant number of matrix ad- ditions. Once we understand recurrences and their solutions, we shall see that this tradeoff actually leads to a lower asymptotic running time. By the master method in Section 4.5, recurrence (4.18) has the solution T(n) = @(n=?) We now proceed to describe the details. In step 2, we create the following 10 matrices: Sy Bi — Bn. S: = Ant+Ap. Ss An + An. Sy = Bu-Bu, Ss = Ant+An, So = ButBr Sy; = Ap— An Ss = Bu +B So = An—Ao So = Bu} Bo Since we must add or subtract n/2 x n/2 matrices 10 tim take @(n?) time, Instep 3, we recursively multiply /2n/2 matrices seven times to compute the following n/2 x n/2 matrices, each of which is the sum or difference of products of A and B submatrices: this step does indeed Pr= Ane Si = Ane Bis — Ane Ba Po = Spe Ba = Ai Br + Arz+ Bar Ps = Sy+By = An Bu | An Bu Py = An Sp = Ana Ba — Ana Bur Ps = Ss-Se = Aus Bur + Atr+ Bor + Anas Bu ~ Azz Ba. Po = Sp+8s = Aiz+ Bay | Atz+ Box — Ara + Bor — Ana» Boo. Pr = SoS = Ans Butane Bu An: Bi Note that the only multiplications we need to perform are those in the middle col- umn of the above equations. The right-hand column just shows what these products equal in terms of the original submatrices created in step 1 Step 4 adds and subtracts the P; matrices created in step 3 to construct the four n/2.X n/2 submatrices of the product C. We start with 11 = Ps + Py— Pa = Py 42. Stwassen’s algorithm for matris multiplication 8) Expanding out the right-hand side, with the expansion of each P; on its own line and vertically aligning terms that cancel out, we see that Cy, equals Aur Bu + Ani: Bas + Aza Bui + Aza: Baz = An By An-Ba = An Bas = Ain Bas ~ Ang Bra Azz Bay + Aa Baa + Ara* Bas Au Bu | Ara Bar which corresponds to equation (4.11). Similarly, we set Ca = Pit and so Ci2 equals 1-Bis— ABs 4+ Ais+Baa ~ Aia+Baa By ~ Ata Bor comresponding to equation (4.12). Setting Cu = Py) Ps makes Co equal An Bu + Ane Bu Aza By ~ Anas Boy comresponding to equation (4.13). Finally, we set Cop = Ps} P= PsP, so that Caz equals Aur Bur + Ari Baz + Aza Bui + Are Bar Aur+Bo2 An Bi 2+ Bi = An Bu Au-Bu Ayy+Biz + Ay By | Aa Biz An B: + Aa Bi 2 Chaprer 4 Divide-and-Conguer which corresponds to equation (4.14). Altogether, we add or subtract n/2 x n/2 ‘matrices eight times in step 4, and so this step indeed takes @(n”) time Thus, we see that Strassen’s algorithm, comprising steps 1-4, produces the cor- rect matrix product and that recurrence (4.18) characterizes its running time. Since ‘we shall see in Section 4.5 that this recurrence has the solution T(n) = @(n'*7), Swrassen's method is asymptotically faster than the straightforward SQUARE- MATRIX-MULTIPLY procedure. The notes at the end of this chapter discuss some of the practical aspects of Strassen’ algorithm, Exer Note: Although Exercises 4.2-3, 4.2-4, and 4.2-5 are about variants on Strassen’s algorithm, you should read Section 4.5 before trying to solve them. 42-1 Use Strassen’s algorithm to compute the matrix product, (3 s)(E 2) 4.22 Write pseudocode for Strassen’s algorithm, 423 How would you modify Strassen’s algorithm to multiply » x1 matrices in which n is not an exact power of 2? Show thatthe resulting algorithm runs in time @(n'*”). 424 What is the largest & such that if you can multiply 3 x3 matrices using k multi- plications (not assuming commutativity of multiplication), then you can multiply nxn matrices in time o(n'*7)? What would the running time of this algorithm be? 4.25 V. Pan has discovered a way of multiplying 68 x 68 matrices using 132,464 mul- tiplications, a way of multiplying 70 > 70 matrices using 143,640 multiplications, and a way of multiplying 72 x 72 matrices using 155,424 multiplications. Which ‘method yields the best asymptotic running time when used in a divide-and-conquer ‘matrix-multiplication algorithm? How does it compare to Strassen’s algorithm?

You might also like