The techniques of trilinear aggregating and the recent progress in the asymptotic acceleration of matrix operations (Q800727)

From MaRDI portal
scientific article
Language Label Description Also known as
English
The techniques of trilinear aggregating and the recent progress in the asymptotic acceleration of matrix operations
scientific article

    Statements

    The techniques of trilinear aggregating and the recent progress in the asymptotic acceleration of matrix operations (English)
    0 references
    1984
    0 references
    N\(\times N\) matrix multiplication (and consequently many other computations) can be performed using \(2N^ 3-N^ 2\) arithmetic operations in the straightforward algorithm. This bound can be reduced to \(cN^ a\) for \(a<3\) via recursive application of a basis algorithm for multiplication of matrices of a specific size if an efficient basis algorithm is available. The ingenious ad hoc basis algorithm for the \(2\times 2\) case (found by \textit{V. Strassen} [Numer. Math. 13, 354-356 (1969; Zbl 0185.401)]), gave \(a=2.807..\). In 10 years general techniques of trilinear aggregating for the design of the basis algorithms [the author, Usp. Mat. Nauk 27, No.5, 249-250 (1972; Zbl 0261.65025)] were reintroduced [the author, Proc. 19th Ann. Symp. Found. Comput. Sci., 166- 176 (1978)]. \textit{D. Bini, M. Capovani, G. Lotti} and \textit{F. Romani} [Inf. Process. Lett. 8, 234-235 (1979; Zbl 0395.68048)] and then \textit{A. Schönhage} [SIAM J. Comput. 10, 434-455 (1981; Zbl 0462.68018)] extended the classes of algorithms that can be used as the bases of the recursive constructions, \textit{D. Coppersmith} and \textit{S. Winograd} [SIAM J. Comput. 11, 472-492 (1982; Zbl 0486.68030)] found a suitable modification of the recursive construction itself. All this accentuated the power of trilinear aggregating, so Strassen's exponent was eventually reduced to \(a<2.5\) in a series of several papers by several authors. This paper presents a concise but complete proof of the possibility to reduce the exponent below 2.5 with the inclusion of all the major techniques of algorithmic design known for matrix multiplication and of their application to the direct sum conjecture for tensor ranks. The author published a more elementary and less advanced survey in SIAM Rev. 26, 393-415 (1984) and a more detailed presentation of the subject in the monograph ''How to Multiply Matrices Faster'' [Lect. Notes Comput. Sci. 179 (1984)].
    0 references
    0 references
    0 references
    0 references
    0 references
    arithmetic complexity
    0 references
    matrix multiplication
    0 references
    trilinear aggregating
    0 references
    direct sum
    0 references
    tensor ranks
    0 references
    0 references
    0 references