ISSN 0001-4346, Mathematical Notes, 2015, Vol. 97, No. 4, pp. 531–555. © Pleiades Publishing, Ltd., 2015. Original Russian Text © S. B. Gashkov, 2015, published in Matematicheskie Zametki, 2015, Vol. 97, No. 4, pp. 529–555. Arithmetic Complexity of Certain Linear Transformations S. B. Gashkov * Lomonosov Moscow State University, Moscow, Russia Received September 1, 2012; in nal form, September 5, 2014 AbstractWe obtain order-sharp quadratic and slightly higher estimates of the computa- tional complexity of certain linear transformations (binomial, Stirling, Lah, Gauss, Serpi ´ nski, Sylvester) in the basis {x + y}∪{ax : |a|≤ C} consisting of the operations of addition and in- ner multiplication by a bounded constant as well as upper bounds O(n log n) for the compu- tational complexity in the basis {ax + by : a, b R} consisting of all linear functions. Lower bounds of the form Ω(n log n) are obtained for the basis consisting of all monotone linear functions {ax + by : a, b > 0}. For the nite arithmetic basis B +,*,/ = {x ± y, xy, 1/x, 1} and for the bases {x ± y}∪{nx : n N}∪{x/n : n N} and B +,* = {x + y, xy, 1}, estimates O(n log n log log n) are obtained in certain cases. In the case of a eld of characteristic p, com- putations in the basis {x + y mod p} are also considered. DOI: 10.1134/S0001434615030256 Keywords: linear transformation (binomial, Stirling, Lah, Gauss, Serpi ´ nski, Sylvester), arith- metic computational complexity of linear transformations, nite arithmetic basis, eld of characteristic p, 1. INTRODUCTION In the present paper, we consider widely used nite-dimensional linear transformations over elds of real and complex numbers (and over nite elds) and estimated their computational complexity using various collections of elementary arithmetic operations in these elds. These sets of elementary operations are called bases (there is usually no mistake of identifying them with the notion of basis in a linear space). By computational complexity we mean the number of performed operations (in what follows, we shall present the formal denition). In combinatorics [1], [2], we have the binomial transformation R n R n and its inverse dened, respectively, by the formulas y k = k m=0 k m x m , x k = k m=0 (1) km k m y m , k =0,...,n 1, where the ( k m ) = k!/(m!(k m)!) are the binomial coecients. The matrix of this transformation is triangular; its nonzero elements constitute a Pascal triangle of order n. Using Pascal’s identity, we can easily compute all the numbers generating the Pascal triangle by using (n 1)(n 2)/2 operations of addition 1 . This algorithm is optimal in the case of the additive basis B + = {x + y, 1} and it is also optimal in a wider basis 2 . The computational model used in the case of the basis B + is known as addition chains [3]. By an addition chain of length l we mean a sequence of natural numbers a 0 ,...,a l , in which a 0 =1 and each number a i is equal to the sum of two preceding numbers a j , a k , j<i, k<i (not necessarily j = k; in the case j = k, the corresponding step of the chain is called a doubling step). An addition chain can * E-mail: sbgashkov@gmail.com 1 In view of its symmetry, the actual number of additions is almost twice less. 2 Because the number of operations in a computation cannot be less than the number of numbers to be computed. 531