Multiplication algorithms : bilinear complexity and fast asymptotic methods

Svyatoslav Covanov 1, 2
Abstract : Since 1960 and the result of Karatsuba, we know that the complexity of the multiplication (of integers or polynomials) is sub-quadratic: given a ring R, the product in R[X] of polynomials a_0 + a_1 X and b_0 + b_1 X, for any a_0, a_1, b_0 and b_1 in R, can be computed with three and not four multiplications over R: (a_0 + a_1X)(b_0 + b_1X) = m_0 + (m_2 - m_0 - m_1)X + m_1X^2, with the three multiplications m_0 = a_0b_0, m_1 = a_1b_1 et m_2 = (a_0 + a_1)(b_0 + b_1). In the same manner, Strassen's algorithm allows one to multiply two matrices 2nx2n with only seven products of matrices nxn. The two previous examples fall in the category of bilinear maps: these are functions of the form Phi : K^m x K^n -> K^l, given a field K, linear in each variable. Among the most classical bilinear maps, we have the multiplication of polynomials, matrices, or even elements of algebraic extension of finite fields. Given a bilinear map Phi, computing the minimal number of multiplications necessary to the evaluation of this map is a NP-hard problem. The purpose of this thesis is to propose algorithms minimizing this number of multiplications. Two angles of attack have been studied. The first aspect of this thesis is to study the problem of the computation of the bilinear complexity under the angle of the reformulation of this problem in terms of research of matrix subspaces of a given rank. This work led to an algorithm taking into account intrinsic properties of the considered products such as matrix or polynomial products over finite fields. This algorithm allows one to find all the possible decompositions, over F_2, for the product of polynomials modulo X^5 and the product of matrices 3x2 by 2x3. Another aspect of this thesis was the development of fast asymptotic methods for the integer multiplication. There is a particular family of algorithms that has been proposed after an article by Fürer published in 2007. This article proposed a first algorithm, relying on fast Fourier transform (FFT), allowing one to multiply n-bit integers in O(n log n 2^{O(log^* n)}), where log^* is the iterated logarithm function. In this thesis, an algorithm, relying on a number theoretical conjecture, has been proposed, involving the use of FFT and generalized Fermat primes. With a careful complexity analysis of this algorithm, we obtain a complexity in O(nlog n 4^{log^* n})
Liste complète des métadonnées

Cited literature [51 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01825744
Contributor : Abes Star <>
Submitted on : Thursday, June 28, 2018 - 3:37:25 PM
Last modification on : Tuesday, December 18, 2018 - 4:18:26 PM
Document(s) archivé(s) le : Thursday, September 27, 2018 - 7:55:57 AM

File

DDOC_T_2018_0057_COVANOV.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01825744, version 1

Citation

Svyatoslav Covanov. Multiplication algorithms : bilinear complexity and fast asymptotic methods. Symbolic Computation [cs.SC]. Université de Lorraine, 2018. English. ⟨NNT : 2018LORR0057⟩. ⟨tel-01825744⟩

Share

Metrics

Record views

367

Files downloads

198