The Accelerated Euclidean Algorithm

Transcription

1 The Accelerated Euclidean Algorithm Sidi Mohamed Sedjelmaci To cite this version: Sidi Mohamed Sedjelmaci The Accelerated Euclidean Algorithm Laureano Gonzales-Vega and Thomas Recio Eds 2004, University of Cantabria, Santander, Spain, pp , 2004 <hal > HAL Id: hal Submitted on 2 Dec 2004 HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not The documents may come from teaching and research institutions in France or abroad, or from public or private research centers L archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d enseignement et de recherche français ou étrangers, des laboratoires publics ou privés

2 The Accelerated Euclidean Algorithm Sidi Mohamed Sedjelmaci Abstract We propose a new GCD algorithm called Accelerated Euclidean Algorithm, or AEA for short, which matches the On log 2 n loglog n time complexity of Schönhage algorithm for n-bit inputs This new GCD algorithm is designed for both integers and polynomials We only focus our study to the integer case, the polynomial case is currently addressed [3] ccsd , version 1-2 Dec Introduction The algorithm is based on a half-gcd like procedure, but unlike Schönhage s algorithm, it is iterative and therefore avoids the penalizing calls of the repetitive recursive procedures The new half-gcd procedure reduces the size the integers at least a half word-memory bits per iteration only in single precision By a dynamic updating process, we obtain the same recurrence and the same time performance as in the Schönhage approach Throughout, the following notation is used W is a word memory, ie: W 16, 32 or 64 Let u v > 2 be positive integers where u has a n bits with n 32 Given a non-negative integer x N, lx represents the number of its significant bits, not counting leading zeros, ie: lx log 2 x+1 For the sake of readability integers U and V will be represented as a concatenation of l sets of W bits integers except for the last set which may be shorter, with l n/w, ie: if U l 1 i0 2iW U l i with U 1 0 and V l 1 i0 2iW V l i, then we write symbollically U U 1 U 2 U l and V V 1 V l We use specific vectors which represent interval subsets from of U and V, by: Ui U X[ij] i+1 U j Ui ; X[i] ; for 1 i < j l V i V i+1 V j V i Let Mx be the cost of a multiplication of two x-bit integers The function Mx depends on the algorithm used to carry out the multiplications The fast Schönhage-Strassen algorithm [6] performs these multiplications in Mx On log n log log n 2 AEA: The Accelerated Euclidean Algorithm The following algorithm is based on two main ideas: The computations are done in a Most Signifivant digit First MSF way Update by multiplying, with the current matrix, ONLY twice the number of leading bits we have already chopped, NOT the others leading bits 1

3 Algorithm AEA Input : u v > 2 with u 8W; n lu; Output : a 2 2 matrix M and U,V such that M U,V U,V and lv lu 2 log 2 n/w 1 W Begin 1 U,V u,v; s log 2 n/w ; 2 if lv lu/2 + 1 return I; 3 if lv > lu/ For i 1 to 2 s 1 5 if U i 0 or V i 0 Regular case 6 h 0; 7 if i odd L 0 ILEX[ii + 1]; update L i,h; 8 else 9 R 0 ILEX[ii + 1]; update R i,h; 10 x i/2; h h + 1; 11 While x even 12 R h R h 1 L h 1 ; update R i,h; 13 x x/2; h h + 1; 14 Enwhile 15 L h R h 1 L h 1 ; update L i,h; 16 Endelse 17 else Irregulari U i 0 and V i 0 ; 18 EndFor 19 Return L h and U,V ; End The algorithm ILE borrowed from [7] runs the extended Euclidean algorithm and stops when the remainder has roughly the half size of the inputs Algorithm ILE Input : u 0 u 1 0 Output : a 2 2 matrix M and u i 1,u i such that M u 0,u 1 u i 1,u i and lu i 1 2 lu 0 Begin 1 n lu 0 ; p lu 1 ; 2 if p < n/2 + 1 return M if p n/ m p n/2 1; 5 Apply Extended Euclid Algorithm until a i 2 m < a i+1 ; ai 1 b 6 return M i 1 and u i 1,u i End a i b i ; 2

4 The functions update L or update R update not all the bits of the operands, but only the next useful small vectors, in order to get the next needed matrix The irregular case is when U i 0 and V i 0, ie: one or many components are all equal to zero Roughly speaking, we perform an euclidean division in order to make an efficient reduction and continue our process The aim is to preserve, at most, the general scheme of our process The basic idea is to use full updated vectors, ie: vectors updated with all the previous matrices L h 3 An Example In order to carry out single-precision computations we take W 20 Recall that the notation means 2 A a c A a c B b d B b W +, where A, B, a, b, c and d d are integers cf notation in Section 1 u Let, then, with our notation, we obtain v U u V v We have U and V then lu 1 lv 1 20, n p 20 and m 1 p 1 n 2 9 Thus, in order to compute ILEU 1,V 1, we must stop the Extended Euclid Algorithm at a i 2 9 We obtain the matrix a i < N 1 ILEU 1,V 1 then : U U1 2 N 1 N 20 + U 2 U1 V 1 V N + V N 1 1 V 2 U1 2 So N U V , hence U and V Now we can apply ILE to the new integers less than 32 bits U and V We have lu 31 and lv 29 We obtain m 8 and the second matrix in 32-bits single precision and M N 2 N 1 N 2 ILEU,V Moreover, at this level, we may consider that U 1 and V 1 are eliminated chopped, even if U 2 has some extra bits, and that U 2 and are updated as follows: 3

5 ; lu 2 21; l 20 at 1 b t 1 On the other hand, it is easy to check that the final matrix M U M V Now, if u and v were larger in size, namely: u u3 u 4 u k v v 3 v 4 v k, a t b t satisfies then we have to continue the half-gcd process Since W bits have been already chopped, we have to update only the double, ie: multiply the next 2W leading bits of U and V by M, namely perform: U3 U 4 V 3 V 4 U3 U M 4 V 3 V 4, and disregard all the other bits of U and V U 3 U1 U 2 Then do the same process as before with instead of to chop V 3 V 1 V 2 the vector, and so on, repeating the process till we reach the middle of the size of U 4 An Example Let us consider two consecutive Fibonacci numbers u,v F 59,F 58, ie: F F First, we must compute mmax which gives 2 mmax, the maximum size of the output matrix L 1 mmax lv lu Let U 1,V , then lu 1 lv 1 20, n p 20 and m p 1 n 2 9 We run the Extended Euclid algorithm and stops at a i 2 9 We obtain the matrix L 0 and the remainders r 1,r 2 : L 0 ILEU 1,V 1 and r ,r ,1404 U U1 10 Hence L 0 L 9 + U 2 U1 V 0 V L + V V 9 + L 0 1 V 2, so By update1,0 we compute L 0 4

6 U1 10 L 9 + U V where FIX occurs in the last equality, hence the new values of U 1 and V 1 : U and V We apply ILE to the new integers U and V We have lu 21 and lv 21 Repeating the same process as before, we obtain m 9, the second matrix R 0 and remainders r 1,r R 0 ILEU,V and r ,r ,1583 Since L 1 R 0 L 0 and using update2,0 we obtain U L 1 R V R U F32 Thus L 1 V F and L 1 R 0 L We can apply one more Euclid step because the matrix Q L 1 still satifies a i 2 mmax So after one Euclid step we finally obtain: U F31 Q L 1, V F 30 and after L 1 Q L 1, we have: L Moreover, at this level, we may consider that U 1 and V 1 are eliminated chopped, even if U 2 has some extra bits, and that U 2 and are updated as follows: ; lu 2 21; l 20 at 1 b t 1 On the other hand, it is easy to check that the final matrix M U M V Now, if u and v were larger in size, namely: u u3 u 4 u k v v 3 v 4 v k, a t b t, satisfies then we have to continue the half-gcd process Since W bits have been already chopped, we have to update only the double, ie: multiply the next 2W leading bits of U and V by M, namely perform: 5

7 U3 U 4 V 3 V 4 U3 U M 4 V 3 V 4, and disregard all the other bits of U and V U 3 U1 U 2 Then do the same process as before with instead of to chop V 3 V 1 V 2 the vector, and so on, repeating the process till we reach the middle of the size of U Here we stress that this is the main difference between our approach and the other Sorenson s like algorithms, where all the bits of U and V are updated by multiplying them with the matrix M In our approach, we only update the double of what we have already chopped 5 Remarks Unlike the recursive versions of GCD algorithms [1],[4], our aim is not to balance the computations on each leaf of the binary tree computations, but to make full of single precision everytime, computing therefore the maximum of quotients in single precision This lead to different computations each step in the algorithm AEA and the other recursive GCD algorithms Moreover the fundamental difference between AEA and Schönhage s approach is that AEA deals straightforward with the most significant leading bits first MSF computation Consequently, we can stop the algorithm AEA at any time and still obtain the leading bits of the result Thus AEA is a strong MSF algorithm and can be considered for on line arithmetics, where all the basic operations can be carried out simultaneously as soon as only the needed bits are available This new algorithm should be an alternative to the Schönhage GCD algorithm On the other hand, the derived GCD algorithm deals with many applications where long euclidean divisions are needed We have identified and started to study many applications such as subresultants and Cauchy index computation, Pade-approximates or LLL-algorithms 6 References [1] Gathen, J von zur, Gerhard, G Modern Computer Algebra In Cambridge University Press 1999 [2] DH Lehmer Euclid s algorithm for large numbers, American Math Monthly, 45, 1938, [3] MF Roy and SM Sedjelmaci The Polynomial Accelerated Euclidean Algorithm, Subresultants and Cauchy index, work in progress, 2004 [4] A Schönhage Schnelle Berechnung von Kettenbruchentwicklugen, Acta Informatica, 1, 1971, [5] A Schönhage, A,F,W Grotefeld and E Vetter Fast Algorithms, A Multitape Turing Machine Implementation, BI-Wissenschaftsverlag, Mannheim, Leipzig, Zürich,

8 [6] A Schönhage, V Strassen Schnelle Multiplication grosser Zalen Computing 7,1971, [7] SM, Sedjelmaci On A Parallel Lehmer-Euclid GCD Algorithm, in Proc of the International Symposium on Symbolic and Algebraic Computation ISSAC 2001, 2001, Sidi Mohamed Sedjelmaci LIPN CNRS UMR 7030, Université Paris-Nord Av J-B Clément, Villetaneuse, France sms@lipnuniv-paris13fr 7