Note on generalized Archimedes–Borchardt algorithm

We investigate convergence and invariance properties of the generalized Archimedes–Borchardt algorithm. The main tool is reducing the problem to an appropriate Gauss iteration process.


Introduction
Iterative procedures, among others those approximating various irrational numbers using some means, have been well-known for a long time. One of the classical iterative algorithms is the Newton process being a formalization of the Babylonian method of extracting the square root of a positive number a. Starting with an arbitrary positive x 0 the sequence (x k ) k∈N0 is (strictly) decreasing and bounded, so convergent: it approximates the number √ a (see, for instance, [4] by Carlson). Putting y k := a/x k we see that (1.1) is equivalent to x k+1 = 1 2 (x k + y k ) and 1/y k+1 = 1 2 (1/x k + 1/y k ) , k ∈ N 0 , or x k+1 = A (x k , y k ) and y k+1 = H (x k , y k ) , k ∈ N 0 , (1.2) where A and H denote, respectively, the arithmetic and harmonic means (cf. [8, p. 190] by Foster and Phillips, also [6] by Daróczy). It follows from the definition of (y k ) k∈N0 that it strictly increases to √ a. Consequently, Clearly (1.2) is a particular case of the recurrent process x k+1 = M (x k , y k ) and y k+1 = N (x k , y k ) , where M and N are means on an interval I. Here M : I 2 → I is called a (bivariate) mean on I provided it satisfies the inequalities min{x, y} ≤ M (x, y) ≤ max{x, y}, x,y∈ I.
Another important example of a recurrent algorithm (1.3) is that with M = A and N = G, i.e. the arithmetic and geometric means, respectively: Both these sequences have the common limit called the arithmetic-geometric mean (medium arithmeticum-geometricum) of x 0 and y 0 , denoted by A ⊗ G (x 0 , y 0 ). The algorithm (1.4) occured first in 1784 in the work [17] by Lagrange in connection with reduction and evaluation of elliptic integrals (see also [18, pp. In general, the value of the arithmetic-geometric mean at an arbitrary point (x 0 , y 0 ) ∈ (0, +∞) 2 was determined by Gauss in 1818: [10]), also [11, pp. 352-355]). For a systematic description of Gauss' theory we refer to the comprehensive article [5] by Cox. The reader interested in other compound means like the arithmetic-geometric mean is referred to the book [2] by Borweins.
In 1800 Gauss suggested studying the process which is superficially similar to (1.4). Apparently he realized that both (x k ) k∈N and (y k ) k∈N approach a common limit. Unexpectedly it is expressed not by elliptic functions like in the case of (1.4) but by trigonometric or hyperbolic functions (cf. [4]). As it follows from [12, pp. 234, 284] the same was known for Pfaff at the very beginning of the 19th century. In 1880 algorithm (1.5) and its fundamental properties were rediscovered by Borchardt [1] (see also [24] and [4]). Since then it occasionally bears his name. If we change the Borchardt algorithm by replacing the arithmetic mean in the first equality by the harmonic one, then we will come to the process Also here both sequences (x k ) k∈N and (y k ) k∈N tend to a common limit. In particular, starting with x 0 = 2 √ 3 and y 0 = 3 we obtain the algorithm attributed to Archimedes (see [13], also [9,22]) for estimating the number π. For a longer and comprehensive story about the Archimedean approximations to π the reader is referred to the book [2] by Borweins.
Observe that sequences (x k ) k∈N , (y k ) k∈N satisfy the Borchardt process if and only if the sequences (1/x k ) k∈N , (1/y k ) k∈N satisfy the Archimedean one. Thus we need consider only one of these two algorithms, for instance algorithm (1.5).
Both processes: Borchardt's (1.5) as well as Archimedes' (1.6) are particular cases of the algorithm where M and N are bivariate means on an interval I and x 0 , y 0 ∈ I. Observe that if M is one of the means A, G, H on I = (0, +∞), then M is strict: For a discussion about some other examples of the Archimedes-Borchardt process (1.7) see, for instance, [8].

Reduction to generalized Gauss algorithm
In the present paper we study the generalized Archimedes-Borchardt algorithm and prove a convergence result extending Theorem FP. Given an interval I of reals and a positive integer p a function M : I p → I is called a mean (in p variables) on I when [3]). It is said to be strict when for all x 1 , . . . , x p ∈ I and every permutation σ of the set {1, . . . , p}, then the mean M is called symmetric.
Given means M 1 , . . . , M p : I p → I and points x 0,1 , . . . , x 0,p ∈ I consider the recurrent process which extends both algorithms (1.5) and (1.7). Our aim is to prove, among others, the following result. As we see Theorem 2.1 considerably generalizes the convergence part of Theorem FP. The generalization comes in three respects: (i) two bivariate means have been replaced here by p means in p variables; (ii) the convergence is uniform with respect to x 0,1 , . . . , x 0,p on each compact subset of the cube I p ; (iii) the assumption of symmetricity turned out to be superfluous.
However, Theorem 2.1 says nothing about the monotonicity of the sequences considered there, unlike Theorem FP. So the problem below seems to be of interest. In what follows, given an interval I and mappings M 1 , . . . , M p : I p → I, we recurrently define functions N 1 , . . . , N p : I p → I by for each x = (x 1 , . . . , x p ) ∈ I p .
The simple observation described in Proposition 2.3 (iv) shows that the investigation of the Archimedes-Borchardt process (2.1) can be reduced to the generalized Gauss algorithm (2.3) which is relatively well-studied (see [15, Sect. 2 and the references therein], also [7]). Notice that equalities (2.3) are equivalent to the condition Here, as usual, The description of the generalized Gauss algorithm and its limit behaviour, presented in Theorem M, originates in the research of Matkowski in his paper [21] (see also [19,20] and, for some further details, [15,Section 2]). In some earlier publications Matkowski assumed that at most one of the means N 1 , . . . , N p is not strict (see [19] for the case p = 2, also [20] where a gap from [19] was filled). The remark below shows how to easily comply with the assumptions imposed on the means N 1 , . . . , N p in Theorem M. Proof. The only mean in one variable is the identity function, so the assertion clearly holds when p = 1. So assume that p ≥ 2. We may also assume that the means N 1 , . . . , N p−1 are strict. Take any x ∈ I p such that equalities (2.5) and (2.6) hold and suppose that Thus ( Theorem M has been extended in a number of directions. Notice that, among others, its versions for parametrized means and random means were proved in [14] and [16], respectively. Now we are in a position to prove Theorem 2.1.

Invariants
When passing from geometry to analysis we often make use of calculations based on algorithms. Then some invariants associated with them turn out to be important and useful. Notice the following result which is another consequence of Proposition 2.3, Remark 2.4 and Theorem M. It is usually difficult to find the form of the invariant mean L (cf., for instance, [4,22][8]). Remember that even in such a seemingly simple case of the classical Gauss algorithm (1.4) the unique continuous (A, G)-invariant mean is defined using an elliptic integral. In the case of the generalized Archimedes-Borchardt algorithm the situation seems to be even more complicated as the mean L existing on account of Theorem 2.1 is invariant with respect to the auxiliary sequence (N 1 , . . . , N p ), not to the original sequence (M 1 , . . . , M p ). In such a way we come to the next problem ending this note. Problem 3.2. Find a class of processes (2.1) for which the common limit L of the sequences (x k,1 ) k∈N , . . . , (x k,p ) k∈N can be determined using invariant equation (3.1).
Open Access. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creativecommons.org/licenses/by/4.0/.
Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.