Arnold Schönhage

Comments on Selected Topics

Citations [..] in this text are referring to our List of Publications,
there furnished with corresponding return pointers `/b'
- that are always placed at the end of the preceding line -
like [26] related to median finding,  or [32] about UNION-FIND.


1. Fast multiplication of integers and polynomials

After [9] and similar complexity bounds for integer multiplication obtained by Toom and Cook, the improved estimate  O(N · log N · loglog N)  from [16]  has been the best we knew for many years, as far as circuit complexity or the speed of multitape Turing machines are concerned -- until M. Fürer could now prove the better bound  O(N · log N · 2O(log*N))  for these models of computation, cf. Proc. 39th ACM STOC (2007), 57-66.
    Our studies of linking automata, however, (about SMM's, see [13], [23], and section 6 of the final paper [35], in particular)  have revealed that such "pointer machines" are capable of performing integer multiplication in linear time !
    In 1981/82 we reconsidered these issues under more practical aspects, especially towards numerical computations with complex polynomials. Section 1 of [39] generalizes the Fermat number method of [16] to fast multiplication  mod (2N + 1)  with so-called "suitable" numbers N, and [44] reports our first experiences with a multitape Turing machine implementation of that approach, which then became the starting point of our TP Project.

In algebraic complexity (counting adds, subs and mults at unit cost), analogous methods apply for multiplying polynomials over rings, with [31] settling the particular case of characteristic 2.  For (approximate) numerical multiplication of polynomials in R[x] or C[x], however, bit complexity models are more adequate. Then methods of binary segmentation as in [39] are preferable, saving a logarithmic factor. For the related case of numerical multiplication  mod xn, we can save another 25% of the running time by the methods of [53].

2. Gcd computations

In 1987, developing our fast routines for computing integer gcd's, we have replaced the original divide-and-conquer method of [17] by a much better suited version called  "Controlled Euclidean Descent", as briefly outlined in Section 7.2.1 of [51].  In [49] one finds the same ideas carried out for the analogous case of binary quadratic forms.

Polynomial gcd's in Z[x] are the theme of [47], and how to reduce them to integer gcd computations by means of binary segmentation, while the case C[x] treated in [43] requires quite different methods, deeply connected with the infinitesimal features of complex numbers and polynomials.

3. Matrix computations

Algorithmic issues of linear algebra have always been among our favorite topics of research, starting with [4] and [7] on the quadratic convergence of Jacobi's method.
    Then, after Strassen's seminal paper Gaussian elimination is not optimal,  the challenge was, of course, to exploit his new matrix multiplication method for other matrix computations as well, like our blockwise orthogonalization methods in [19] and [21] - as tutorial in [22] - in particular yielding a fast (and stable) method for unitary transformation of complex matrices to upper Hessenberg form. Methods and theorems from [36] have been used for deriving smaller bounds for the exponent(s) of matrix multiplication.

To compute characteristic polynomials, we should distinguish two cases. Over fields of finite characteristic, [50] offers a purely algebraic approach via computing sufficiently many power sums of the eigenvalues.
    For the numerical case (working approximately over C), one can apply Keller-Gehrig's algebraic method to matrices in upper Hessenberg form, as we have briefly outlined at the end of the preliminary report [40] (now also available for downloads). - Meanwhile P. Kirrinnis has considerably extended and developed these initial ideas to an impressive work on fast computation of invariant subspaces.

4. Computational complexity of analytic functions

In discussing algorithmic problems dealing with complex numbers, we prefer a strictly constructive approach susceptible to actual computing, thus very different from alternate views like the perspective propagated with the BSS-model  and in the book reviewed in [55].
    Algorithms are to be specified for some realistic machine model (like multitape Turing machines, or simple RAM's)  with some fair cost measure of bit complexity. Any input number z shall be given by some oracle answering any `queries' r  by providing binary approximations x, y with   |(x + iy) - z| < 2-r.
    Then it is fairly obvious how to compose two oracles for inputs u and v  to obtain a machine program serving as a new oracle for simple functions like  w = u+v,  or  w = u×v,  the latter yielding N-bit precision in running time  O(µ(N)),  where µ(N) shall denote a time bound for N-bit integer multiplication (such that  µ(N)/N  is monotone), like  µ(N) = cN  for pointer machines.

Evaluation complexity of holomorphic functions f should be considered locally. Restricted to any (closed) disk D in C with convergent power series expansion,  fD is said to be computable in time  t(s),  if there exists a machine M such that on input of integer  s > 0  and any z in D  given by an oracle,  M  is stopping after at most  t(s) steps with output w such that  |w - f(z)| < 2-s.
    There is the beautiful result due to R. Brent (see references in [48] )  that the elementary functions  exp, ln, cos, sin, arctan,  etc. (on any regular disk) are all computable in time  O(µ(s)·log s)  - but what about higher transcendental functions ?
    Interesting upper bounds have been found for the Gamma, zeta, and error function, to name a few, like evaluation of  Gamma(z) in time sß for any  ß > 1.5  (see Borwein & Borwein, Pi and the AGM, page 332). The methods in [46] have a different focus, aiming at efficient mass production for multiple evaluations of Riemann's zeta function. The last section of [48] contains an excursus how to use these techniques for computing a single value  zeta(½ + it)  for large t  up to moderate precision t -c within time  O(tß),  so for any  ß > 0.375 .

With regard to analytic continuation, we have the following general

Theorem.  If  fD admits analytic continuation along curve C to some other germ  fEE -> C  on disk E,  then there exists  c > 0  such that computability of  fD in time t(s) implies computability of  fE in time  O(st(cs).
    With  t(cs) = O(sß),  this implies computability of  fE in time  O(sß+1).
Conjecture: In this polynomially bounded case, much sharper  fE should then also be computable in time  O(t(s)).

5. Computational complexity of algebraic functions

The evaluation complexity of any algebraic function is simply bounded by a constant multiple of integer multiplication time, which holds in the following strongly uniform sense, cf. Prop. (6.1) of [48]:

Consider  A0, . . . ,An  in  Z[z1, . . . ,zk]  with complex inputs z1,...,zk varying in some compact subset K of Ck, there satisfying  maxj|Aj(z)| > ½,  say. Then the branches w of the algebraic functions defined by the equation
F(z,w) = A0(z) + A1(zw + · · · + An(zwn = 0   are computable up to relative precison 2-s in time  c·µ(s),  with an estimation constant c merely depending on F and K.

This general result is implied by our splitting circle method, see [45] and the report [40], in particular. Beyond that there remains, of course, an abundance of more special problems to find small estimation factors c of this sort, depending on the degree n of the defining equation and many other circumstances. - Just think of little examples like
  F(z,w) = z1 - z2·w,   or   z - w3 = 0,  leading to the discussion of efficient algorithms for complex division, or for complex cube roots.

After the integer and gcd routines had been developed and implemented, we started in 1988/89  to carry out all those detailed studies needed for an efficient implementation of the splitting circle method, by and by also implementing all the corresponding routines. - Our "TP-book" [51] from 1993/94  describes an intermediate stage of these efforts. - The routines for root finding as the main application could eventually be completed by the end of 1998.
    Meanwhile many enhancements have been added, while implementation of some items is still pending, e.g. the improved polynomial division method from [56] by means of Graeffe steps. Turing Processing tells more about the present state of all that software.


Last modified: Sun Jul 1 11:36:31 CEST 2018