On effective analytic continuation

Joris van der Hoeven

Mathématiques, Bâtiment 425

CNRS, Université Paris-Sud

91405 Orsay Cedex

France

Email: joris@texmacs.org

Web: http://www.math.u-psud.fr/~vdhoeven

Until now, the area of symbolic computation has mainly focused on the manipulation of algebraic expressions. It would be interesting to apply a similar spirit of “exact computations” to the field of mathematical analysis.

One important step for such a project is the ability to compute with computable complex numbers and computable analytic functions. Such computations include effective analytic continuation, the exploration of Riemann surfaces and the study of singularities. This paper aims at providing some first contributions in this direction, both from a theoretical point of view (such as precise definitions of computable Riemann surfaces and computable analytic functions) and a practical one (how to compute bounds and analytic continuations in a reasonably efficient way).

We started to implement some of the algorithms in the Mmxlib library. However, during the implementation, it became apparent that further study was necessary, giving rise to the present paper.

Keywords: analytic continuation, Riemann surface, algorithm, differential equation, convolution equation, relaxed power series, error bound

A.M.S. subject classification: 03F60, 30-04, 30B40, 30F99

1.Introduction

Although the field of symbolic computation has given rise to several softwares for mathematically correct computations with algebraic expressions, similar tools for analytic computations are still somewhat inexistent.

Of course, a large amount of software for numerical analysis does exist, but the user generally has to make several error estimates by hand in order to guarantee the applicability of the method being used. There are also several systems for interval arithmetic, but the vast majority of them works only for fixed precisions. Finally, several systems have been developed for certified arbitrary precision computations with polynomial systems. However, such systems cannot cope with transcendental functions or differential equations.

The first central concept of a systematic theory for certified computational analysis is the notion of a computable real number. Such a number is given by an approximation algorithm which takes with on input and which produces an -approximation for with . One defines computable complex numbers in a similar way.

The theory of computable real numbers and functions goes back to Turing [Tur36] and has been developed further from a theoretical point of view [Grz55, Alb80, BB85, Wei00]. It should be noticed that computable real and complex numbers are a bit tricky to manipulate: although they easily be added, multiplied, etc., there exists no test for deciding whether a computable real number is identically zero. Nevertheless, possibly incomplete zero-tests do exist for interesting subclasses of the real numbers [Ric97, MP00, vdH01b]. In section 2.5, we will also introduce the concept of semi-computable real numbers, which may be useful if a zero-test is really needed.

The subject of computable real numbers also raises several practical and complexity issues. At the ground level, one usually implements a library for the evaluation of basic operations , , , etc. and special functions , , , etc. Using fast multiplication methods like the FFT [KO63, CT65, SS71], this raises the question of how to do this in an asymptotically efficient way [Bre76a, Bre76b, CC90, Kar91, vdH99a, vdH01a, vdH05b]. At an intermediate level, one needs a software interface for certified operations with arbitrary precision numbers. Several implementations exist [FHL+05, GPR03, Mül00, vdH99b, vdH06b], which are mostly based on correct rounding or interval arithmetic [Moo66, AH83, Neu90, JKDW01, BBH01, Bla02]. At the top level, one may finally provide a data type for real numbers [MM96, Mül00, Lam06, O'C05, vdH06a, vdH06b]. Given the real number result of a complex computation, an interesting question is to globally optimize the cost of determining a given number of digits of the result, by automatically adjusting the precisions for intermediate computations [vdH06a, vdH06b].

The next major challenge for computational analysis is the efficient resolution of more complicated problems, like differential or functional equations. In our opinion, it is important to consider this problem in the complex domain. There are several reasons for this:

This paper aims at providing a basic theoretical framework for computations with computable analytic functions and effective analytic continuation. When possible, our study is oriented to efficiency and concrete implementability.

The history of analytic continuation of solutions to complex dynamical systems goes back to the 19-th century [BB56]. Although interval arithmetic and Taylor models have widely been used for certified numeric integration of dynamical systems [Moo66, Loh88, MB96, Loh01, MB04], most implementations currently use a fixed precision [Ber98]. Some early work on effective analytic continuation in the multiple precision context was done in [CC90, vdH99a, vdH01a, vdH05b]; see also [vdH07] for some applications. Of course, fast arithmetic on formal power series [BK75, BK78, vdH02b] is an important ingredient from the practical point of view. Again, the manipulation of computable analytic functions is very tricky. For instance, even for convergent local solutions to algebraic differential equations with rational coefficients and initial conditions, there exists no general algorithm for determining the radius of convergence [DL89]. Of course, one also inherits the zero-test problem from computable complex numbers.

Let us detail the structure and the main results of this paper. In section 2, we start by recalling some basic definitions and results from the theory of computable real numbers. In particular, we recall the concepts of left computable and right computable real numbers, which correspond to computable lower resp. upper bounds of real numbers.

In section 3, we introduce the concept of a computable Riemann surface. In a similar way as computable real numbers are approximated by “digital numbers” in , we will approximate computable Riemann surfaces by so called “digital Riemann surfaces”, which are easier to manipulate from an effective point of view. For instance, in section 3.2, we will see how to identify two branches in a digital Riemann surface. However, from a conceptual point of view, it is not always convenient to see Riemann surfaces as limits of sequences of digital approximations. In sections 3.4 and 3.5, we will therefore discuss two equivalent ways to represent computable Riemann surfaces. Notice that all Riemann surfaces in this paper are above .

The next section 4 deals with constructions of several kinds of computable Riemann surfaces. We start with the definition of computable coverings (which can be thought of as morphisms of computable Riemann surfaces) and the construction of the limit of a sequence of coverings. We proceed with the definition of disjoint unions, covering products, quotients and joins at a point. For instance, if and are the Riemann surfaces of two analytic functions resp. , then and are defined on the covering product of and . In section 4.4, we consider Riemann surfaces which admit a distinguished point, the root. This allows for the definition of a smallest “organic” Riemann surface which contains a prescribed set of “broken line paths”. Universal covering spaces and so called convolution products of rooted Riemann surfaces are special cases of organic Riemann surfaces.

In section 5, we come to the main subject of computable analytic functions. In [vdH05a], a first definition was proposed. Roughly speaking, the idea was to see a computable analytic function as an instance of an abstract data type , with methods for computing

This point of view is very natural from a computational point of view if we want to solve a differential or more general functional equation, since it is often possible to locally solve such equations. However, the computed bounds are usually not sharp, so we need some additional global conditions in order to ensure that analytic continuation can be carried out effectively at all points where the solutions are defined.

Now the more systematic theory of computable Riemann surfaces of this paper makes it possible to directly define the concept of a computable analytic function on a given computable Riemann surface. Although this makes definitions easier, one still has to show how to construct the Riemann surface of a computable analytic function. Using the results from section 4, we will do this for many classical operations, like , , , , , , , , algebraic and differential equations, convolution products, etc. Especially in the case of convolution products, the global knowledge of an underlying Riemann surface is very important. What is more, we will show that it is possible to construct the Riemann surfaces incrementally, on explicit demand by the user. Also, whereas all underlying Riemann surfaces from [vdH05a] were simply connected, the present theory enables us to identify certain branches where the function takes identical values. Nevertheless, the local approach from [vdH05a] remains useful, because any “locally computable analytic function” induces a natural “globally computable analytic function” (see theorem 5.7).

During the implementation of some of the algorithms from [vdH05a] in our Mmxlib library, it turned out that bad bounds and could lead to extremely inefficient algorithms. Therefore, it is essential to have algorithms for the efficient computation of accurate bounds. In section 6, we will study this problem in a systematic way. Our leitmotiv is to work with truncated power series expansions at an order with a bound for the remainder. On the one hand, we will study how such expansions and bounds can be computed efficiently and accurately (sections 6.3 and 6.4). On the other hand, we will show how to use them for computing the absolute value of the smallest zero of an analytic function (section 6.1) and for computing extremal values on a compact disk (section 6.2). Several of the ideas behind our algorithms already occur in the literature about Taylor models and polynomial root finding. However, the context is a bit different, so our exposition may have some interest for its own sake.

For the sake of simplicity, we have limited ourselves to the study of univariate analytic functions. It should be possible to generalize to the multivariate case along the same lines. The main extra difficulty we foresee is integration, because it requires an automatic algorithm for the deformation of paths. Nevertheless, in sections 4.8 and 5.5, we study convolution products, and a similar approach might be used for integration. Some of the algorithms in this paper have been implemented in the Mmxlib library. However, our implementation is still quite unstable and work is in progress to include the ideas from the present paper.

2.Computable real and complex numbers

2.1.Computable functions and relations on effective sets

We assume that the reader is familiar with basic notions of the theory of Turing machines. We recall that a Turing machine computes a function , where if the Turing machine does not halt on the input . A function is said to be computable if for some Turing machine . A subset of is said to be recursively enumerable, or shortly enumerable, if or if there exists a Turing machine with . We say that is computable if both and are enumerable. Denoting by the set of Turing machines, there exists a bijection , whose inverse encodes each Turing machine by a unique natural number.

More generally, an encoding (or effective representation) of a set is a partial surjective function , which is not necessarily injective. In that case, we call (or more precisely the pair ) an effective set. If is computable or enumerable, then we call an abstract computable resp. enumerable set. For instance, the set of Turing machines which halt on all inputs is an effective set, but not an abstract computable set, because of the halting problem. If and are effective sets, then so is , for the encoding , where . By induction, is an effective set for each . Many other classical sets, like finite sequences or trees over an effective set admit straightforward encodings, which will not be detailed in what follows.

A function between two effective sets and is said to be computable if there exists a Turing machine such that for all . In that case, each with provides an encoding for , and we denote by the effective set of all computable functions from to . A partial function is said to be computable if there exists a Turing machine with for all .

Sometimes, it is convenient to allow for generalized encodings , where is another encoded set. Indeed, in that case, the composition yields an encoding in the usual sense. For instance, , where encodes each function by the Turing machine which computes it. Given , we will write and . To each object , given by its encoding with , we may naturally associate its representation in . However, this association does not lead to a mapping , since we do not necessarily have . In particular, in order to implement a computable function via a computable function , using , one has to make sure that whenever .

An -ary relation on an effective set is said to be computable, if there exists a computable subset of , with . Equivalently, we may require the existence of a computable function with for all . Similarly, is enumerable, if there exists an enumerable subset of , with . This is equivalent to the existence of a computable function with for all . Here denotes the set of increasing computable functions , divided by the equivalence relation with . Notice that and are equal as sets, but not as effective sets. A computable function will be called an ultimate test. Notice that the equality relation on an effective set is not necessarily computable or enumerable, even if is an abstract computable set.

Since a subset is also a unary relation on , the above definition in particular yields the notions of computable and enumerable subsets of . We also define to be a sequentially enumerable subset of if or if there exists a computable function with . Similarly, we say that is sequentially computable if both and are sequentially enumerable. If is sequentially enumerable and admits a computable equality test, then is enumerable. If is enumerable and is an abstract enumerable set, then is sequentially enumerable. If is sequentially computable, then is an abstract enumerable set.

There are several other interesting notions which deserve further study, but which will not be used in what follows. For instance, we may define a subset of an effective set to be pseudo-computable, if there exists a computable function with , where is defined similarly as . For instance, given a Turing machine , the set is a pseudo-computable subset of .

2.2.Computable real numbers

Let be the set of digital or dyadic numbers. Given an ordered ring , we denote , , etc. Given and , we say that is an -approximation of if . An approximator for is a computable function which sends to an -approximation of . If admits such an approximator, then we call a computable real number and encode by . We denote by the set of approximators and by the effective set of computable real numbers. Given , both the problems of testing whether resp. are undecidable.

The usual topologies on and naturally induce topologies on and . Given an open subset of , an element of is called a computable real function. Notice that such a function admits a natural encoding by an element , where . Many classical functions like , , , , , , are easily seen to be computable. It can be shown (see [Grz55, Grz57, Wei00] and theorem 2.3 below) that a computable real function is necessarily continuous. Consequently, the step and stair functions are not computable. Intuitively speaking, this stems from the fact that the sign function cannot be computed effectively for computable real numbers.

It is convenient to express part of the semantics of computations with computable real numbers by providing a signature for the available operations. For instance, the class comes with two main functions

Similarly, the class provides operations

However, we take care not to provide functions for comparisons.

2.3.Left and right computable real numbers

There exist many equivalent definitions for computable real numbers and several alternative encodings [Wei00, Chapter 4]. A particularly interesting alternative encoding is to define an approximator (or two-sided approximator) of to be a computable function with

and . This definition admits two variants: a left approximator (resp. right approximator) of is a computable increasing (resp. decreasing) function , with . A real number is said to be left computable (resp. right computable) if it admits a left (resp. right) approximator.

Intuitively speaking, a left (resp. right) computable real number corresponds to a computable lower (resp. upper) bound. Indeed, in what follows, it will frequently occur that we can compute sharper and sharper lower or upper bounds for certain real numbers, without being able to compute an optimal bound. We denote by , , and the left and right analogues of and .

Remark 2.1. The above definitions of left, right and two-sided approximators naturally extend to the case of sequences in the set of extended digital numbers. This leads to natural counterparts , , , etc.

Remark 2.2. For actual implementations, it is a good idea to let the index of approximators correspond to the estimated cost of the computation of (see also [vdH06b]). We also notice that left, right and two-sided approximators can be implemented by a common class real with a method approximate, which returns a bounding interval as a function of . In the case of left (resp. right) approximators, we would have (resp. ).

Let be an open subset of or . A function is said to be lower continuous (resp. upper continuous), if for every and every (resp. ), there exists a neighbourhood of , such that (resp. ) for all . We have [Grz55, Grz57, Wei00]:

Theorem 2.3. Let be an open subset of . Then

  1. Any is continuous.

  2. Any is lower continuous.

  3. Any is upper continuous.

Proof. We will prove (b); the other two assertions are proved in a similar way. The function admits an encoding . Let with approximator

Let be a left approximator for . Given , there exists a with . Now the computation of by only depends on for some finite . Increasing if necessary, we may assume without loss of generality that

Let , with approximator . For a certain , we have . Now consider the alternative approximator of with for and otherwise. Then, by construction, satisfies . We conclude that .

The “lower step function” , defined by if and otherwise, is lower computable in the sense that . Indeed, given , we may take . Similarly, the function is lower computable, while is upper computable. In particular, this shows that . Besides the projections

typical lower computable functions on are:

Here the dot in indicates the argument of the function . Left computable numbers are turned into right computable numbers and vice versa by the following operations:

More generally, increasing computable real functions induce both increasing lower and upper computable real functions, while decreasing computable real functions turn left computable real numbers into right computable real numbers and vice versa.

2.4.Computable complex numbers

The complexification of provides a natural definition for the set of computable complex numbers. Typical operations on include

The complexification of also provides a natural encoding for and, setting , the approximation function for numbers in extends to

Clearly, functions like , , , etc. can only be defined on simply connected subsets of . On the other hand, is effectively algebraically closed in the sense that there exists an algorithm which takes a polynomial of degree on input and which returns its set of roots in .

2.5.Semi-computable numbers

For many applications, the absence of computable comparisons for computable real or complex numbers can be a big problem. One solution to this problem is to systematically consider all possible answers of zero tests or sign computations and to use these answers as hypotheses during subsequent tests. For instance, if we assume that , then a subsequent test should return .

The above approach can be formalized as follows. A system of real constraints is a pair with and for . We say that is satisfied if for . We denote by the set of systems of real constraints. A semi-computable real number is encoded by a computable function , where is a finite subset of such that at least one element of is satisfied and whenever both and are satisfied. We denote by the set of semi-computable real numbers. A semi-computable function is a function . Such a function naturally induces a function . Indeed, given , encoded by , we may take and , whenever .

Example 2.4. The step function is semi-computable. Indeed, given , we first compute an -approximation of with (e.g. ) and . If , then we let

and take . Otherwise, we let

and take with and .

From a practical point of view, computations with semi-computable numbers can be implemented using non-deterministic evaluation and we point to the similarity with the computation with parameterized expressions [vdH97, Chapter 8]. Each branch of the non-deterministic computation process comes with a system of real constraints in . A constraint checker is used in order to eliminate branches for which is contradictory.

In many applications, the numbers belong to a polynomial algebra and one may use classical algorithms from real algebraic geometry to check the consistency of [BPR03]. Modulo further progress in automatic proofs of identities [Ric92, Zei90, vdH02a], we hope that more and more powerful constraint checkers will be constructed for increasingly general classes of constants (like algebraic exp-log expressions in ). This would allow for the automatic elimination of a large number of inconsistent branches. Notice also that it is recommended to spend a roughly equivalent time in trying to prove and disprove constraints. Of course, proving is easy, since it suffices to find a non zero digit of .

As in the case of computations with parameterized expressions, many algorithms for computable real numbers naturally generalize to semi-computable real numbers. This is due to the fact that all numbers involved often belong to a fixed polynomial algebra , in which the Noetherianity of this algebra may be used in termination proofs. We refer to [vdH97] for examples.

Remark 2.5. In our definition of systems of real constraints, we have considered sign conditions on computable real numbers. The same construction may be applied to more general types of constraints, like , for a certain number of fixed subsets of the real numbers. However, we have not yet found any practical use for such a generalization.

3.Computable Riemann surfaces

A classical Riemann surface (above ) is a topological space , together with a projection , so that every admits a neighbourhood for which is a homeomorphism of on an open ball of . A Riemann surface with border is defined similarly, except that each now admits a neighbourhood for which is a homeomorphism of on a subset of which is homeomorphic to . A classical covering is a local homeomorphism between two Riemann surfaces, which commutes with the projections, i.e. . Throughout this paper, coverings are not required to be surjective.

3.1.Digital Riemann surfaces

An encoding of a digital Riemann surface is a tuple , where is a finite set of nodes, a scale, a projection and a symmetric adjacency relation, such that

DR1

If , then .

DR2

If and are such that , then .

DR3

Let be such that for and such that three relations among , , and hold. Then the fourth relation holds as well.

The conditions DR2 and DR3 are illustrated in figure 3.1 below. In the case when with pairwise distinct projections and satisfy , then we will also write . Notice that .

Figure 3.1. Illustration of the axioms DR2 (top) and DR3 (bottom) for digital Riemann surfaces. When regarding the left hand sides as digital Riemann pastings, the right hand sides also correspond to their normalizations.

Let us show how to associate a Riemann surface in the classical sense to an encoding as above. To each , we may associate a compact square by

We now consider the topological space

where is a copy of for each . Whenever , the squares and admit a common edge in . Gluing the corresponding copies and together in according to this edge determines a new topological space

The space is a Riemann surface with a border, whose projection on is naturally determined by the projections of the on . Indeed, DR2 (resp. DR3) implies that points on the edges (resp. vertices) of the are either in the interior of or on its border. The interior of , endowed with its natural projection on , is a Riemann surface in the classical sense; we call it the digital Riemann surface associated to . It will be convenient to write and denote the interior of by . More generally, if , then we write and for its interior.

Example 3.1. If the mapping is injective, then the induced projection is a homeomorphism on its image. In that case, we will identify with the subset of , and call a digital subset of . Conversely, any together with a finite subset determines a natural digital subset , encoded by with .

Example 3.2. One of the simplest examples of a digital Riemann surface for which is not injective is shown in figure 3.2. Formally speaking, this surface is encoded by

Figure 3.2. Example of a digital Riemann surfaces with non-trivial fibers.

Consider an encoding of a digital Riemann surface at scale . This encoding induces a natural “doubled encoding” , by associating four nodes with to each . Given , we set if and only if and , or and . The doubled encoding encodes the same digital Riemann surface , but at the smaller scale . By induction, it is possible to obtain encodings at any scale with .

Given a digital Riemann surface , the above argument shows that there exists a maximal scale , such that admits an encoding at scale for every . Inversely, the encoding of at a given scale is essentially unique (up to bijections ). Indeed, given , the center of each () corresponds to a unique point in . Furthermore, given with , we have if and only if the segment lifts to a segment on . If the scale is clear from the context, then it will be convenient to denote “the” encoding of by . If is the result of some computation, then we will denote by the scale of the corresponding representation.

Remark 3.3. In practice, it is more efficient to work with a set of scaled nodes instead of . Each element of is a pair with , and . A scaled node corresponds to nodes in with and for all . For simplicity, we will directly work with nodes in in what follows. Nevertheless, with some additional effort, our algorithms can be adapted to work with scaled nodes.

Let be a digital Riemann surface, with a fixed encoding . We write

for the set of digital points on . Such a point can be encoded by a pair with and such that . This encoding is unique, except when lies on the border of two squares. Notice that is an abstract computable set. Similarly, we write

for the set of computable points on . Such a point can be encoded by a pair with and , such that the distance between and is bounded by . Hence, we have , or with , or with . In particular, admits a computable open neighbourhood , such that is a homeomorphism onto a rectangle with corners in . Notice that there exists no algorithm for testing whether for given and .

3.2.Digital Riemann pastings

During actual computations with digital Riemann surfaces, the conditions DR2 and DR3 are not always met a priori. In that case, we need an automatic procedure to identify nodes when necessary. This will be the main objective of this section.

Consider a tuple as in the previous section which only satisfies DR1. Then the construction of the previous section still yields a topological space with a projection , even though may now contain points which are not locally homeomorphic to open subsets of . We will call a digital Riemann pasting with encoding . For instance, taking , , and , we obtain the digital Riemann pasting shown at the upper left hand side of figure 3.1.

A quotient structure on is a pair , where is an equivalence relation on and an adjacency relation on , such that

In that case, when setting for all , the tuple again encodes a digital Riemann pasting.

The intersection of two quotient structures and is again a quotient structure. Moreover, if both and encode digital Riemann surfaces, then so does . Consequently, generates a smallest quotient structure for which encodes a digital Riemann surface. We call the normalization of and the corresponding digital Riemann surface the normalization of . It can be checked that normalization commutes with the doubling operator , so that that this definition indeed does not depend on the chosen scale. Two examples of normalizations are shown in figure 3.1.

Example 3.4. Let be a digital Riemann pasting and let be an equivalence relation on with . Given an encoding , we may define an equivalence relation on by . Then encodes a digital Riemann surface, which we will denote by .

In order to compute the normalization of a digital Riemann pasting , it is convenient to maintain

Given a subset of nodes such that consists of a singleton , we may then glue all nodes in together using

In order to normalize , we now keep on doing the following operations until both DR2 and DR3 are satisfied:

The normalization procedure finishes, because the size of strictly decreases for the first operation and the number of strictly increases for the second operation.

3.3.Digital coverings and computable Riemann surfaces

A digital covering is a covering in the classical sense between two digital Riemann surfaces. Let be a digital covering and let and be encodings of and at the same scale . Then induces a mapping , which sends to the unique with , where stands for the center of . This mapping satisfies

(3.1)
(3.2)

Inversely, given a mapping which satisfies (3.1) and (3.2), we obtain a covering in the classical sense by sending to the unique point with . In other words, the digital covering may be encoded by the triple . We will denote by the set of all digital coverings.

Example 3.5. Let be a digital Riemann surface encoded by . Consider the equivalence relation on and the projection . Then the tuple with , and encodes the digital complex subset of and is a digital covering.

Example 3.6. The definition of digital coverings naturally extends to digital Riemann pastings. The normalization of a digital Riemann pasting comes with a natural digital covering . In particular, given an equivalence relation on with , we obtain a natural covering . This generalizes the previous example, by taking . Moreover, any digital covering induces a digital covering which commutes with .

Given digital Riemann surfaces and coverings , we call

(3.3)

a digital covering sequence. Such a sequence admits a natural limit

(3.4)

where is the smallest equivalence relation with for each , and has the natural structure of a Riemann surface. We will write for the composed covering and for the covering which sends to . We say that the covering sequence (3.3) is computable, if the mapping is computable. In that case, we call a computable Riemann surface. Notice that coverings are not necessarily injective. This corresponds to the idea that better knowledge of the Riemann surface of a function may lead to the detection of identical leaves.

Example 3.7. Let be an open rectangle with corners in or an open disk with center in and radius in . For each , let

By example 3.1, and determine a digital subset of . The limit of the sequence of embeddings is a computable digital Riemann surface, which is homeomorphic to . More generally, the limit of a computable sequence of embeddings of digital subsets of is called a computable open subset of .

Example 3.8. The example 3.2 can be adapted to turn infinitely many times around the hole in the middle. Indeed, consider the “infinite digital Riemann surface” encoded by:

Given , the restriction of to those with determines a digital Riemann surface in the usual sense. The natural inclusions determine a digital covering sequence whose limit corresponds to . Notice that is isomorphic to the universal covering space of ; see also section 4.7.

Let be a fixed computable Riemann surface. We denote by

the sets of digital and computable points on . A digital point (and similarly for a computable point ) may be encoded by a partial sequence such that and for all . We notice that is an abstract enumerable set. We have natural computable mappings

As in the case of digital Riemann surfaces, each admits a computable open neighbourhood , such that is a homeomorphism onto a rectangle with corners in .

3.4.Atlas representation of a computable Riemann surface

Instead of using the digital representation of computable Riemann surfaces (i.e. as limits of digital covering sequences), we may also try to mimic more classical representations of Riemann surfaces. For instance, a computable atlas representation of a Riemann surface with projection is a tuple , where

Proposition 3.9. Any computable Riemann surface admits an atlas representation.

Proof. Let be the limit of a digital covering sequence of digital Riemann surfaces and define

Given , let be an encoding of . We have already noticed that for , with or with . We may thus take . Conversely, given , the composition of and the restriction of to determines an immersion of into . Finally, given pairs , we may ultimately check whether : given , we check whether and .

Proposition 3.10. Any Riemann surface with a computable atlas representation can be given the structure of a computable Riemann surface.

Proof. Let be an enumeration of and an enumeration of all pairs with .

Let us first assume that each is a digital subset of . Consider the disjoint union , together with the smallest equivalence relation for which corresponding squares in and are equivalent if and only if . Setting , we obtain a natural computable digital covering sequence . We claim that is isomorphic to the limit of this sequence.

Indeed, the construction implies natural coverings which pass to the limit . Inversely, naturally immerses into , with inverse . Gluing these immersions together for all , we obtain a covering with (since every is contained in ), proving that .

In the general case, each is the computable limit of a sequence of immersions. We may now construct another computable atlas representation of , by taking , , etc. We conclude by applying the above argument to this new computable atlas representation.

Remark 3.11. From the proofs of the above propositions, it becomes clear that the class of Riemann surfaces with a computable atlas representation does not change if we require the computable open sets to be of a prescribed type, like open rectangles with corners in or open balls with centers in and radii in .

3.5.Intrinsic representation of a computable Riemann surface

Let be a classical Riemann surface above and denote

Given and , we denote

(3.5)

Given a point , let be the largest radius such that there exists an open disk for which admits an open neighbourhood so that the restriction of to is a homeomorphism between and . Given with , we denote by or the unique point in with . In particular, the notations (3.5) naturally generalize to the case of balls and in , for (resp. ).

A computable intrinsic representation of is a tuple such that

Simplifying notations , and , we thus have the following signature:

Proposition 3.12. Any computable Riemann surface admits a computable intrinsic representation.

Proof. Let be the limit of a digital covering sequence . For , we take the set of encodings of points and we already have a computable mapping .

Let be the encoding of a point . The distance of to the border is easily computed for each . Since and , the sequence encodes . Similarly, given with , it is easy to compute in for each sufficiently large . Then the sequence encodes .

Finally, yields an enumeration of . Given with encodings and , we may ultimately check whether holds: given , we test whether and .

Proposition 3.13. Let be a Riemann surface with a computable intrinsic representation. Then is a computable Riemann surface.

Proof. Let be the enumeration of and an enumeration of all pairs such that holds.

For each , we may compute a square with corners in , for some such that . Now let , where is the smallest equivalence relation induced by identifying matching squares in and for pairs . We claim that the limit of the induced digital covering sequence is isomorphic to .

Indeed, we have natural coverings for each , which pass to the limit . Inversely, for each , the set can be immersed in some , where is large enough such that all pairs with are among . Gluing these immersions together, we this obtain an immersion with , proving that .

3.6.Optional features of computable Riemann surfaces

Let be a computable Riemann surface. In certain cases, we may design an algorithm to compute the distance of a point to the border. In that case, we say that is a delimited computable Riemann surface.

Remark 3.14. One might prefer to call computable Riemann surfaces in our sense lower computable Riemann surfaces and delimited computable Riemann surfaces simply computable Riemann surfaces (and similarly for computable open sets). However, in what follows, we will mainly have to deal with computable Riemann surfaces for which we do not have a computable distance function . Therefore, we will stick to the original definition.

Assume that and consider two points . Even under the assumption that , we notice that there exists no test in order to decide whether . Indeed, given encodings and of resp. , we do not know whether there exists an index with . Nevertheless, we naturally do have such a test in the case when the coverings are embeddings. In this case, we say that has computable branches. Conversely, assume that we have a conditional equality test

where returns the result of the test , provided that we are given the answer to the test . Equivalently, one may assume a predicate

such that holds if and only if , provided that . Then we may associate a new digital Riemann surface to each , by identifying all squares with whose centers are equal (using the normalization algorithm from section 3.2). This leads to a new representation of , for which the induced coverings are embeddings. When using the atlas representation, has computable branches if and only if we have a computable test for deciding whether .

4.Constructions of computable Riemann surfaces

4.1.Computable coverings

Consider two computable Riemann surfaces and . A covering is said to be computable if its restriction to is a computable mapping . A digital representation of such a covering is a triple , such that represents , represents and is a computable sequence of digital coverings , such that

(4.1)

commutes and for any . If each is an immersion, then we call a computable immersion (of representations). If is also surjective, then we call a computable subdivision (of representations), and is said to be a subdivision of .

Lemma 4.1. Let be the representation of a computable Riemann surface. Then we may compute a computable subdivision of , such that there exist with for all and .

Proof. Without loss of generality, we may assume that the are encoded at scales . Given a digital Riemann surface encoded by , let stand for its restriction to the subset of inner nodes which admit four distinct neighbours . Taking , and , the inclusion mappings determine a computable immersion of the Riemann surface represented by into . Since , this immersion is actually a subdivision and we have for all .

Lemma 4.2. Let be the limit of a computable covering sequence and a digital Riemann surface such that is compact. Then we may compute an and a digital Riemann surface with .

Proof. The set forms an open covering of . Since is compact, it follows that there exists an with . Since and are both digital Riemann surfaces, we may actually check whether , and therefore compute the first for which this holds.

Proposition 4.3. Let be a computable covering. Let and be representations for resp. . Modulo subdividing and reindexing , the covering admits a computable digital representation of the form .

Proof. By lemma 4.1, we may assume without loss of generality that there exist with for all and . In particular, is a compact subset of for all . By lemma 4.2, we may compute a digital Riemann surface with . We next increase further until there exists a digital covering which commutes with . On the one hand, the digital coverings , whose incarnations at a suitable scale are finite in number, can easily be computed. Using the predicate , we also have an ultimate test for checking whether . Trying all values of in parallel, we know that one of these tests will ultimately succeed. Increasing still further so as to ensure that , this completes the construction of the digital representation of .

Remark 4.4. A representation of a computable Riemann surface is said to be proper if there exist with for all and . From the proof of proposition 4.3, it follows that it is not necessary to subdivide , provided that is proper.

A computable covering sequence is a computable sequence

(4.2)

where each is a computable Riemann surface and each a computable covering. Let be a proper representation of for each . By induction over , and modulo reindexation of , we may construct a digital representation for , such that we have the following commutative diagram:

In particular, we obtain a new computable Riemann surface

We call the limit of the computable covering sequence (4.2). This limit satisfies the following universal property:

Proposition 4.5.

For every Riemann surface and coverings , there exists a unique covering with for all . Moreover, if is computable and the are given by a computable mapping, then is also computable and we may compute it as a function of and the .

4.2.Disjoint unions and covering products

Let and be two digital Riemann surfaces which are encoded at the same scale . We define their disjoint union by

It is not hard to verify that this construction does not depend on and that is indeed a digital Riemann surface. We have natural inclusions and . The disjoint union satisfies the following universal property:

Proposition 4.6.

Given any digital Riemann surface with digital coverings and , there exists a unique covering with and . Moreover, is a digital covering which can be computed as a function of , and .

Similarly, we define the covering product of and by taking

We have natural digital coverings and which are not necessarily surjective. The covering product does satisfy the following universal property:

Proposition 4.7.

Given any digital Riemann surface with digital coverings and , then there exists a unique covering with and . Moreover, is a digital covering which can be computed as a function of , and .

Let and be computable Riemann surfaces represented by resp. . The disjoint union of and is the computable Riemann surface represented by the sequence . The sequences and determine computable immersions and and the universal properties for pass to the limit. Similarly, the covering product of and is the computable Riemann surfaces represented by the sequence . Again, we have natural computable coverings and which satisfy the universal property for products.

Proposition 4.8. Let and be computable Riemann surfaces.

  1. If and are delimited, then so are and .

  2. If and have computable branches, then so have and .

Proof. All properties are easy. For instance, given , we have

4.3.Quotient spaces and gluing at a point

Let be a computable Riemann surface and a sequentially enumerable relation with . In particular, we may compute a computable sequence , where each is a pair such that is the -th pair in the enumeration of .

For each , let be the smallest equivalence relation on generated by the relations for and . Setting , we have natural computable coverings and . Let be the limit of . The mappings induce a computable surjective covering . For every we have . It is not hard to verify the following universal property of :

Proposition 4.9.

Given a Riemann surface and a covering with , there exists a unique covering with . Moreover, if and are computable, then so is and we may compute it as a function of and .

Let us now consider two computable Riemann surfaces and . Given and with , consider the relation on which is reduced to the singleton . We call the join of and at . If and are not important, or clear from the context, then we also write for . We will denote the natural coverings and by resp. .

Proposition 4.10. Assume that and are connected. Then is connected.

Proof. Assume for contradiction that is not connected and let , where is the connected component of . Then we may define an equivalence relation on by . The quotient set has a natural structure of a Riemann surface and there exists a natural covering . By the universal property of , it follows that , which is impossible.

The proposition ensures in particular that we may apply the following classical theorem:

Theorem 4.11. (van Kampen) Let and be path-connected topological spaces, such that is non-empty and path connected. Denote by and the natural inclusions of in resp. . Then the homotopy group of is given by

where is the normal subgroup of the free product of and generated by elements with .

Corollary 4.12. If and are simply connected computable Riemann surfaces, then so is .

4.4.Computable rooted Riemann surfaces

A broken line path is a finite sequence and we write

Intuitively speaking, corresponds to a path . We write for the set of broken line paths and denote by

the subsets of digital and computable paths. The empty path is denoted by . We say that is a truncation of and write if for some . Given two paths , we write . When no confusion is possible, paths of length will be identified with numbers. If , then we will also write for the path .

A Riemann surface is said to be rooted if it admits a special element called the root of . If is also computable and , then we call a computable rooted Riemann surface. Unless explicitly stated otherwise, we will always assume that rooted Riemann surfaces are connected. A root-preserving covering between two rooted Riemann surfaces will be called a rooted covering. We denote by the class of computable rooted Riemann surfaces. Given , we have an additional method in the signature of .

Let be a computable rooted Riemann surface. We define the path domain of to be the set of , so that

are all well-defined. We will also write . The digital and computable path domains of are defined by

We notice that is an abstract computable set with a computable equality test, whereas is only an effective set. A broken line path naturally induces a continuous path by setting

for and . This path is rooted in the sense that .

Proposition 4.13. Let and be computable rooted Riemann surfaces. Then there exists at most one rooted covering . Such a covering is necessarily computable and computable as a function of and .

Proof. Assume that there exists a covering . By continuity, it suffices to show how to compute for all . Since is connected, there exists a path with . Given , we claim that we may compute such a path . Indeed, the set is enumerable and, given , we may ultimately test whether . We perform these ultimate tests in parallel, for all , until one of them succeeds. Since is connected, we have , so our claim implies .

Proposition 4.14. Let be a computable rooted Riemann surface and assume that is given the natural topology of . Then

  1. , and are open subsets of , resp. .

  2. is a dense subset of and is a dense subset of both and .

Proof. Let us prove the proposition by induction over for each of the subspaces , , etc. The assertions are clear for . Assume that is open, with as a dense subset. We have

where . Now the restriction is computable, so is lower continuous, by theorem 2.3. Assume that and let . Then admits an open neighbourhood with for all . Consequently, is an open neighbourhood of . This proves that , and are open subsets of , resp. . In order to show that is a dense subset of , it suffices to prove that any open ball intersects . Now is an open ball of , which intersects , say in . Furthermore, is a disk with radius . Taking with , we thus have . The other density statements are proved similarly.

Proposition 4.15. Let be the limit of a computable covering sequence .

  1. If are all connected, then so is .

  2. If are all simply connected, then so is .

Proof. Assume that where and are non-empty open sets. Then both intersects and for sufficiently large . Consequently, is not connected. This proves (a). As to (b), assume that are simply connected and consider a loop with . Then is compact, so for a sufficiently large . In a similar way as in lemma 4.2, we may find a such that the restriction of to is a homeomorphism. But then is a loop in which may be contracted to a point. Composing with , we obtain a contraction of into a point.

Proposition 4.16. Given a not necessary connected computable rooted Riemann surface , we may compute the connected component of the root.

Proof. Let . Modulo taking a subsequence, we may assume without loss of generality that contains a point with . It is easy to compute the connected component of in for each . By proposition 4.15(a), the limit of the sequence yields .

4.5.Organic Riemann surfaces

Assume now that we are given an enumerable set of paths and a computable mapping such that, given and , we have if and only if . Reordering terms when necessary, we may assume that is presented as an enumeration such that for all . Assume that we are also given a number ; we call an organic triple.

Let us define a computable rooted covering sequence , such that for all and with . We proceed by induction over . Denote by the computable ball with center and radius . We start with and . Assume that has been constructed. Then the path necessarily occurs before in our enumeration, whence , so that and are well-defined. Now we take

with root and . By construction, for all and with . Indeed, if , then . If , then . This completes the construction of our covering sequence. Its limit is called the organic Riemann surface associated to . Organic Riemann surfaces are always simply connected, by corollary 4.12 and proposition 4.15. They satisfy two universal properties:

Proposition 4.17. Given a rooted Riemann surface with and , there exists a unique rooted covering . Moreover, if is computable, then is computable and computable as a function of .

Proof. Let us show by induction over that there exists a unique rooted covering , where

This is clear for . Assume that the assertion holds for a given . There exists a covering

By the universal property of joins, it follows that there exists a rooted covering with and . We obtain by proposition 4.5 and we conclude by proposition 4.13.

Proposition 4.18. Let and be organic triples with . Then there exists a unique rooted covering , which is computable and computable as a function of and .

Proof. Notice that for all . Denote the counterparts of , , etc. in the construction of by , , etc. For each , there exists a computable such that . By a similar induction as in the proof of proposition 4.17, one shows that there exists a rooted covering for every . Passing to the limit, we obtain .

Remark 4.19. If we only have a mapping such that for any and with , then we may still define , where

is an enumerable set, which fulfills the stronger requirement that if and only if .

4.6.Universal computable covering spaces

Let be a computable rooted Riemann surface. The construction of organic Riemann surfaces may in particular be applied for , and . In that case, we denote and it can be proved that . In the construction of , each is naturally isomorphic to the ball . By induction over , each therefore comes with a natural rooted covering . Taking limits, we obtain a natural rooted covering and it is readily verified that for all . The universal computable covering space admits the following universal properties:

Proposition 4.20. Given a rooted covering with , there exists a unique rooted covering and satisfies . If is computable, then is computable and computable as a function of .

Proof. With as in the proof of proposition 4.17, the universal property of joins implies that for all . Taking limits for , we conclude that .

Proposition 4.21. Given a computable rooted covering , there exists a unique rooted covering and we have . Moreover, is computable and computable as a function of .

Proof. The existence, uniqueness and computability properties of follow from proposition 4.18. The rooted coverings and are identical by proposition 4.13.

Proposition 4.22. Let be a root-preserving computable covering between two rooted computable Riemann surfaces and with . Then any path with can be lifted uniquely to a path with .

Proof. Let . Since is uniformly continuous, we may approximate by a broken line path with

Since , we have . Consequently, lifts to a path on . Since , we also have for all . Consequently, , so that lifts to the path .

Corollary 4.23.

is isomorphic to the universal covering space of .

4.7.Digital covering spaces

Let be a rooted digital Riemann surface, encoded by . Assume that is such that . In this section, we will then show that the universal covering space can be constructed in a more explicit way.

A digital path is a tuple with . We denote by the set of digital paths on , for which . Given , we write . The set comes with a natural projection and a natural adjacency relation: if and only if or for some .

Let be the subset of of paths of lengths . Then is a Riemann pasting and we denote by its associated digital Riemann surface. The root can be lifted to a root of for and the natural inclusions induce natural rooted coverings for .

Proposition 4.24. With the above notations the limit of is isomorphic to the universal covering space of .

Proof. In view of proposition 4.13, it suffices to prove that there exist rooted coverings and . Since , we have natural rooted coverings . This yields a rooted covering when passing to the limit. Conversely, any path can naturally be approximated by a digital path , in the sense that , after possible reparameterization of . Setting , we then have , which shows the existence of a rooted covering .

Proposition 4.25. The mappings are injective.

Proof. The theoretical definition of the normalization of a Riemann pasting also applies in the case of when is infinite and one has resp. for if and only if these relations hold for a sufficiently large with . For each there exists a digital path of smallest length with and we denote this length by . Let for each , so that . For every , the path induces an element of , which shows that the natural rooted covering is surjective. Since is obtained by gluing a finite number of squares to , corollary 4.12 implies that is simply connected, by induction over . Consequently, is isomorphic to for each , and , as desired.

Corollary 4.26. Let be a rooted digital Riemann surface and let be such that . Then there exists an algorithm to decide whether and are homotopic.

Proof. Since has computable branches by proposition 4.25, we have a test for deciding whether . Now this is the case if and only if and are homotopic.

Remark 4.27. Several other algorithms can be developed in order to obtain topological information about digital Riemann surfaces. For instance, let us sketch an algorithm to compute generators of the homotopy group :

  1. Let , , , let be the restriction of to .

  2. Let .

  3. If then return .

  4. Pick an element of minimal length and set .

  5. If , then go to step 3.

  6. Let be obtained by gluing a new square above to .

  7. If there exists a with , then set and identify with inside .

  8. Replace by and go to step 2.

The above algorithm returns a set of digital paths each of which elements corresponds to a generator in .

4.8.Convolution products

Let and be two computable Riemann surfaces with roots above . Organic Riemann surfaces are also useful for the construction of a new Riemann surface such that the convolution product of analytic functions and on resp. will be defined on .

A digital folding on is a computable mapping such that for all and with and . We denote by the enumerable set of digital foldings on . We also write for the subset of rooted digital foldings with . Given , we define by

We define to be the set of all foldings with , such that and lift to rooted foldings on resp. . We notice that is enumerable.

Now any induces a path by , where . By construction, we have and . Let be the enumerable set of all paths which are induced by foldings in . Given , we let

(4.3)

Given with , we claim that . Indeed, let be such that and define with by

By construction, we have and . In view of remark 4.19, we may now define the convolution product of and by .

Proposition 4.28. Let be a continuous function with , such that and its mirror lift into functions and on resp. with and . Then the path can be lifted to a path on . In particular, given and on resp. , the convolution product can be analytically continued along :

where , whenever .

Proof. We first observe that a digital folding induces a natural continuous mapping by

Let

Since is uniformly continuous, we may approximate it by a digital folding with

Moreover, we may take such that and . By our choice of , the foldings and lift to resp. , so that . Moreover, the broken line path satisfies for all , again by the choice of . Consequently, and its associated continuous path lifts to a path on with the same endpoints as . Once more by the choice of , we have for all and . Consequently, lifts to the path on .

The convolution product comes with natural computable rooted coverings and , since any in particular induces a path with . The following universal property follows from proposition 4.18:

Proposition 4.29.

Let and be two computable rooted coverings. Then there exists a unique rooted covering . This covering is computable and can be computed as a function of and . Moreover, and .

5.Computable analytic functions

In [vdH05a], a computable analytic function was defined locally as a “computable germ” with a computable method for analytic continuation. In section 5.1, we recall an improved version of this definition. In section 5.3, we define the new concepts of globally and incrementally computable analytic functions. These concepts allow for computations with analytic functions on computable Riemann surfaces as studied in the previous sections. A locally computable analytic function in the sense of section 5.1 will naturally give rise to a globally computable analytic function on an organic Riemann surface. However, common operations on globally analytic functions, as studied in sections 5.4 and 5.5, may give rise to computable Riemann surfaces which are not necessarily simply connected. Our new definition therefore has the advantage that identical branches may be detected effectively in many cases.

5.1.Locally computable analytic functions

Let be a convergent power series at the origin. We will write for its radius of convergence. Given with , we also define

Finally, given , we will denote by the analytic continuation of along the straightline segment , so that for small .

A locally computable analytic function is an object encoded by a quadruple

where

We denote by the set of locally computable analytic functions. Given , we call its computable radius of convergence. Usually, is smaller than the genuine radius of convergence of .

Remark 5.1. We notice that the definition of the encoding is recursive, because of the method for analytic continuation. Such recursive quadruples can in their turn be encoded by terms in a suitable -calculus, and thereby make the definition fit into the setting introduced in section 2.1.

Example 5.2. One important example of a locally computable analytic function is the identity function centered at a given . We may implement it as a function with

The constructor can be implemented in a similar way.

Example 5.3. Basic operations on can easily be implemented in a recursive manner. For instance, the addition of may be computed by taking

In sections 3 and 4 of [vdH05a], algorithms were given for several other basic operations and for the resolution of differential equations. Modulo minor modifications, these algorithms remain valid. In particular, we have implementations for the following operations:

(5.1)

In the cases of and , the second argument specifies the value of the function at .

It is instructive to rethink the definition of in terms of signatures. First of all, we have a class of computable power series with a method for the extraction of coefficients

Then the class of locally computable analytic functions is determined by the methods

(5.2)

For the last two methods, we understand that and are defined if and only if resp. .

The recursive definition of raises the question when two elements should be considered identical. In what follows, we will use the criterion that if and only if the signature (5.2) does not allow for the distinction of and . In other words, whenever and are such that and are both defined and , we require that , and . We warn that we may have for two different elements .

Remark 5.4. There are a few changes between the present definition and the definition of computable analytic functions in [vdH05a]. First of all, we have loosened the requirements for bound computations by allowing the results of and to be only left resp. right computable. In [vdH05a], we also required a few additional global consistency conditions in our definition of computable analytic functions. The homotopy condition will no longer be needed because of theorem 5.7 below, even though its satisfaction may speed up certain algorithms. The continuity condition also becomes superfluous because of theorem 2.3.

5.2.Improved bounds and default analytic continuation

Given , we have already noticed that the computable radius of convergence of does not necessarily coincide with its theoretical radius of convergence . This raises a problem when we want to analytically continue , because we are not always able to effectively continue at all points where is theoretically defined. By contrast, bad upper bounds for on compact disks only raise an efficiency problem. Indeed, we will show now how to improve bad bounds into exact bounds.

Let us first introduce some new concepts. The path domain of is the set of such that for every . Given , we denote and . The digital path domain of , which is defined by , is enumerable. Given , we say that improves , and we write , if , and for all and .

Assume now that we are given with and let us show how to compute an -approximation for . Approximating sufficiently far, we first compute an with . Now let and choose

(5.3)

sufficiently large such that

(5.4)

Using an algorithm which will be specified in section 6.2, we next compute an -approximation for , where . Then is the desired -approximation of . We have proved:

Proposition 5.5.

Given , we may compute an improvement of , such that for all and .

Another situation which frequently occurs is that the radius of convergence can be improved via the process of analytic continuation and that we want to compute bounds on larger disks, corresponding to the new radii of convergence. This problem may be reformulated by introducing the class of weak locally computable analytic functions. The signatures of and are the same except that comes with a second radius function with ; given , we only require computable bounds for . We have a natural inclusion and the notions of path domain and improvement naturally extend to .

Assume now that and that we want to compute a bound for on for a given . We have an algorithm for computing a with from . Consider any computable sequence with

Since is compact and the function is continuous (by theorem 2.3), it follows that there exists an with for all . In particular, the balls form an open covering of , whence the sequence is necessarily finite. Let be its last term. Then

is a computable upper bound for . We have proved:

Proposition 5.6.

Given a weak locally computable analytic function , we may compute an improvement of .

The bound (5.4) may also be used in order to provide a default analytic continuation method of a computable power series inside a given computable radius of convergence , assuming that we have an algorithm for the computation of upper bounds , . Indeed, let , and be such that and assume that we want to compute an -approximation of !. Now choose with and let . Then the majoration [vdH05a, vdH03]

yields the majoration

so that

Taking in a similar way as in (5.3), we thus have

Let be an -approximation of . Then is also an -approximation of .

5.3.Globally and incrementally computable analytic functions

Let us now consider an analytic function on a computable rooted Riemann surface . We say that is a globally computable analytic function on , if there exists a computable function , which maps to a locally computable analytic function , such that

(5.5)
(5.6)
(5.7)

for all , and . We denote by the set of such functions. We also denote by the set of all computable analytic functions on some connected computable (and computable as a function of ) rooted Riemann surface . In other words, the signature of is given by

Here the projection is required to satisfy . The signature for can also be given in a more intrinsic way:

Notice that the method has been replaced by an exact method with codomain , in view of proposition 5.5. For the intrinsic representation, the compatibility conditions become and , where denotes the Riemann surface with its root moved by .

Using analytic continuation, let us now show how to improve a locally computable analytic function into a globally computable analytic function .

Theorem 5.7. There exists an algorithm which takes on input and produces an improvement of on output. Given any other improvement of , we have .

Proof. For the underlying Riemann surface of , we take , with for all . By the construction of , we may compute a sequence of open balls with , and , such that

Given , we may therefore compute an with (which may depend on the encoding ) and take

Given , we may also compute

Given with , we may finally compute

By propositions 5.6 and 5.5, we may therefore compute for every with . Since , proposition 4.14 and its adaptation to imply , whence . The universal property of (proposition 4.17) implies that for any other improvement of .

In practice, it is not very convenient to compute with global computable analytic functions , because we have no control over the regions where we wish to investigate first. An alternative formalization relies on the incremental extension of the Riemann surface on which is known. Consider the class with the following signature:

Given and , where , the method returns an extension of on a Riemann surface with (in particular, there exists a computable rooted covering ). For simplicity, it will be convenient to assume that . For consistency, we also assume that successive calls of for paths and with yield extended surfaces and for which there exists a rooted covering . This ensures the following:

Proposition 5.8. Consider an enumeration of . Then the limit of the computable rooted covering sequence does not depend on the particular ordering of the enumeration (up to isomorphism).

Corollary 5.9. There exists an algorithm which takes on input and produces an improvement of on output. Given any other improvement of , we have .

Any locally computable analytic function naturally determines an incrementally computable analytic function : starting with , each successive call of joins a ball with radius to at the end of , just like in the construction of organic Riemann surfaces. However, as we will see in the next section, the method may also be used to identify identical branches in the Riemann surface of a function.

5.4.Operations on computable analytic functions

In this section, we improve the implementations of the operations (5.1) so as to identify branches of the underlying computable Riemann surface of an analytic function , whenever we know that takes the same values on both branches. We will also consider several other operations on computable analytic functions.

Constructors
The inclusion and identity are easy to implement, since it suffices to take for the Riemann surface and return for all .

Entire functions
Let us now consider the case of addition for . We take

where stands for the rooted covering product, i.e. is the connected component of the root of . This root is computed by applying the universal property of computable covering products to the immersions of a small ball into neighbourhoods of the roots of and . As to the method , we may simply take

The consistency condition for successive applications of is naturally met, because of the universal property of covering products. The cases of subtraction, multiplication, exponentiation and precomposition with any other computable entire functions in several variables can be dealt with in a similar way.

Multiplicative inverses
Assume now that we want to compute for with . Clearly, we may take

It remains to be shown that is a computable rooted Riemann surface. It suffices to show this in the case when is a digital Riemann surface. Indeed,

Now for every point above , we will show in section 6.1 how to compute a maximal disk on which does not vanish. For the -th approximation of , it suffices to take the union of all with (starting with an for which contains the root of ).

Differentiation
Given , we may take

Integration
Given and , let

Let be the limit of a covering sequence of digital Riemann surfaces. Given , we have sketched in remark 4.27 how to compute generators for the homotopy group of . The relations induce a computable equivalence relation on . Setting , the covering gives rise to a natural covering . We take

Logarithm
Given and with , we may take

However, in this particular case, the integrals of over the above generators are always multiples of , so they can be computed exactly. More precisely, let be the limit of . We now replace by , where is the equivalence relation defined by

Given , we may check whether , by computing -approximations and of resp. and testing whether . The covering induces a natural covering and we take

Solving algebraic equations
Let be such that the polynomial is square-free. Let

Let be the digital Riemann surface with

and with an adjacency relation defined as follows. Solving the equation at the center of yields solutions which we attach arbitrarily to the with . We set if and if the analytic continuation of from to coincides with . This can be tested effectively, since there are no multiple roots, whence all branches are bounded away from each other when computing with a sufficient precision. By a similar argument, the root of may be lifted to , if has a prescribed value , and the rooted covering may be lifted to a rooted covering . We now take

Denoting , we also take

Integral equations
Consider an equation of the form

(5.8)

where is a vector of indeterminates, a polynomial in and . Any algebraic differential equation can be rewritten in this form. In section 6.3 below, we will discuss techniques for computing the power series solution to (5.8) at the origin, as well as bounds and . Given with for all , we have

(5.9)

By what has been said at the end of section 5.2, we may compute . Consequently, the equations (5.8) and (5.9) have the same form, and the analytic continuation process may be used recursively, so as to yield a solution . Since we do not have any a priori knowledge about identical branches in the Riemann surfaces of the , we simply solve (5.8) in by taking . Notice that the decomposition of the integral in (5.9) may be used for more general implicit equations involving integration, even if they are not given in normal form (5.8).

Composition
Let us first show how to compute for given with . Assuming by induction over that , we denote

and set

In section 6.2, we will show how to compute , so that and .

Assume now that are such that , so that is well-defined by what precedes. Let be the canonical Riemann surface of , as in theorem 5.7, and let be the subspace induced by paths for which . A digital folding on is said to be a digital homotopy between the paths associated to and if and is constant. The set of pairs such that determine digitally homotopic paths on is enumerable. We take , where stands for digital homotopy on . We also take

Remark 5.10. With a bit more effort, the computation of the Riemann surface of can be done more efficiently, by directly working with digital approximations and using corollary 4.26.

Heuristic continuation
Assume that we are given a computable convergent power series at the origin. Sometimes, it is interesting to associate a function to by determining and in a heuristic way. For instance, given the first coefficients of , the radius of convergence may be determined heuristically, by looking at the convex hull of in and considering the edge from to with . Then

In order to determine , it suffices to compute until several successive values are small with respect to and approximate . A similar approximation may be used for the analytic continuation to a point with . Finally, one may determine by heuristically identifying branches in where the germs of above the same point coincide up to a given order and at a given accuracy.

Remark 5.11. Even if one may not want to crucially depend on heuristic computations, so as to obtain only certified answers, one may still use them as a complement to theoretically correct computations, in order to obtain an idea about the quality of a bound or to guide other computations. For instance, given , assume that we want to obtain a lower bound for with “expected relative error” . Then we may keep producing better and better lower bounds and heuristic bounds (at expansion order ), until .

5.5.Convolution products

Let . The convolution product of and is locally defined by

(5.10)

If we want to evaluate up to many digits at a small , then we may simply compute the Taylor series expansions of and at and evaluate the primitive of . Assuming that the series expansions of and are given, this algorithm requires a time for the computation of a -approximation of . More generally, if is a broken-line path with , then

Modulo the replacement of each by for a sufficiently large , we may thus compute a -approximation of using the above method, in time .

In order to obtain a complete power series expansion of at or , it is convenient to consider the Borel and Laplace transforms

Then we have

(5.11)
(5.12)

These formula allow for the efficient (and possibly relaxed) computation of the coefficients , since the Borel and Laplace transforms can be computed in essentially linear time.

More precisely, let , and . Assume that and for all and that we are given -approximations of and -approximations of . Then the naive evaluation of the formula (5.12) using interval arithmetic yields -approximations of . In order to use (5.11) and fast multiplication methods based on the FFT, one may subdivide the multiplication of and into squares like in relaxed multiplication method from [vdH02b, Figure 3]. For each square, one may then apply the scaling technique from [vdH02b, Section 6.2.2], so as to allow for FFT-multiplication without precision loss. This yields an algorithm for the computation of -approximations for . Notice that this algorithm is relaxed.

If we want the power series expansion of at a path with , then consider the formula

(5.13)

Assuming that the and are sufficiently small, we also have

(5.14)

for all . Now if is sufficiently small, we may compute the series expansion of at as a function of the series expansion of the same function at the origin, using a variant of [vdH02b, Section 3.4.1]. This yields -digit expansions for coefficients of at in time .

Let us now define and by induction over , in such a way that implies and . Assuming that , we take

(5.15)

Clearly, for with , we have and . Given , we take

(5.16)

This completes the induction and the construction of . If , then we have , since (5.15) reduces to (4.3). If , then we suspect that , but we have not tried to check this in detail.

In practice, if we want to analytically continue along a path which is known to belong to , it can be quite expensive to “randomly” compute a part of which contains . During the analytic continuation of along , it is therefore recommended to progressively compute equivalent paths for which avoid singularities as well as possible. These paths may then be used for the computation of better bounds and in order to accelerate the computation of a part of which contains .

More precisely, let and assume that is fixed. Let

By construction, we have and is the limit of a sequence with . Let be fixed and consider the set of all paths with . Given , let

Here denotes the distance between and the border of and we recall that

stands for the continuous path on associated to . Using Dijkstra's shortest path algorithm, we may enumerate such that . As soon as we find an with

then we stop (this condition can be checked ultimately by computing a sufficiently precise digital approximation of using the techniques from section 4.7). If is small enough, this yields a path which is homotopic to on and for which is small. The idea is now to replace by in the right-hand side of (5.15) resp. (5.16), if this yields a better bound.

The above approach raises several subtle problems. First of all, the computed path depends on the number . When computing a -th approximation for , one possibility is to take . A second problem is that the choice of depends on and , so we no longer have . Nevertheless, it should be possible to adapt the theory to the weaker condition that whenever , where we notice that our change can only lead to improved bounds. Finally, if becomes small, then the shortest path algorithm may become inefficient. One approach to this problem would be to use the shortest path at a larger scale for an accelerated computation of the shortest path at a smaller scale. As a first approximation, one may also try to continuously deform as a function of . We wish to come back to these issues in a forthcoming paper.

6.Bound computations

For actual implementations of computable analytic functions it is very important that bound computations (i.e. lower bounds for convergence radii and upper bounds for the norm on compact disks) can be carried out both accurately and efficiently.

A first problem is to find a good balance between efficiency and accuracy: when bounds are needed during intermediate computations, rough bounds are often sufficient and faster to obtain. However, bad bounds may lead to pessimistic estimates and the computation of more terms in power series expansions in order to achieve a given precision for the end-result. Therefore, it is important that cheap bounds are also reasonably accurate.

Another point is that it is usually a good idea to use different algorithms for rough and high precision bound computations. Indeed, only when sufficient knowledge is gathered about the function using rough bound computations, it is usually possible to fulfill the conditions for applying a high precision method, such as Newton's method. Furthermore, such asymptotically fast methods may only be more efficient when large precisions are required, which requires the study of the trade-off between different methods.

In this section, we will present several techniques for efficient and/or accurate bound computations. Some of the algorithms have been implemented in Mmxlib. However, the topic of bound computations deserves a lot of further study.

6.1.Lower bounds for the smallest zero of an analytic function

Let with and . The problem of computing a lower bound for the radius of convergence of reduces to the computation of a such that has no zeros on . We may start with the simpler problem of computing a lower bound for

where with has been fixed. A natural approach is to approximate the problem by a root finding problem of complex polynomials.

More precisely, we may approximate real and complex numbers by elements of the sets and of real intervals with endpoints in resp. complex balls with centers in and radii in [vdH06b]. Let for some with . We start by picking , and the computation of complex ball approximations for , as well as a bound for the remainder

The bound may be integrated into the constant coefficient by setting . Now we compute a lower bound for the norm of the smallest root of the polynomial

using some classical numerical method and interval/ball arithmetic. The result will then be presented as an interval and yields the desired lower bound for .

We have implemented two experimental versions of the above method for the two numerical methods from [Car96] and a variant of [Pan96, Appendix A]. The first method is based on repeated squaring in the ring . However, it is cumbersome to adapt to the case when there exist almost multiple roots. Also, we observed a lot of precision loss in our context of certified computations with complex balls. This might be due to the divisions. The second method is based on Graeffe transforms and rapidly provided us with rough lower bounds for of an acceptable quality. Let us quickly explain this method.

First of all, we recall that Graeffe's transform sends a polynomial of degree with roots to another polynomial with roots . Such a polynomial can be computed efficiently using FFT-squaring:

Given a monic polynomial with , we also observe that the norm of the largest root of lies in the interval . Indeed, if , then , whence . Similarly, if is such that for all , then for all .

Now let be a polynomial of degree and assume that we want an upper bound for the largest root of with a relative accuracy . If we rather want a lower bound, then we replace by . We start by making monic by setting . We next let be smallest such that . Starting with and , we now repeat the following:

  1. Compute .

  2. Scale .

  3. Replace .

  4. If , then return .

  5. Set and .

Consider the factorizations and , where denotes the original. Then we observe that , each time when we arrive at step 4. When the approximations were sufficiently precise, it follows that we obtain an -approximation of the largest root of on exit.

Remark 6.1. Notice that we simplified the method from [Pan96, Appendix A], since we do not need Turan's proximity test. Instead, we use a variant of bound (B.7) mentioned in Appendix B, by rescaling at each step. Notice that FFT-multiplication leads to huge precision loss when applied to polynomials which have not been scaled.

Remark 6.2. If there exists a unique and simple root of maximal modulus, then after a few steps, we have , with , whence a good approximation of can be read off from . Now if , then either or . Going the way back up, we may thus compute a good approximation of . At a second stage, this approximation may be further improved using Newton's method.

Remark 6.3. The worst case for the above algorithm is when admits a single root of multiplicity . In that case, each iteration typically gives rise to a precision loss of binary digits, when using a fast algorithm for multiplication.

Let us now come back to the original problem of computing a lower bound for the radius of convergence of . Given , we thus have to find an -th lower approximation for with and . We start by computing the -th lower approximation of . For , we may now take if and otherwise (alternatively, one may choose as a function of a heuristic approximation of the radius of convergence of ; see remark 5.11). Using the above algorithm, we may now compute a lower bound for , using an expansion of at order (or an order like which makes the total computation time more or less proportional to ) and . We may then take if and otherwise.

6.2.Computing extremal values on compact disks

Let and be such that . By definition, we have a method for computing an upper bound for . Since this bound may be pessimistic, we will now show how to compute better approximations for .

We start by computing an with and picking an expansion order . If we want an -approximation of , then we may take as in (5.3) and (5.4). We next compute approximations for the first coefficients of the series and set . We now have to approximate

Let be a power of two larger with and . We may efficiently approximate the vector using the FFT and compute

More generally, we may efficiently approximate using the FFT for small values of . Let . Then

for , and where for polynomials of degree . In other words,

(6.1)

We also have

where . We may thus compute an approximation using one FFT of order . More generally, for a fixed , and modulo choosing a larger , we may compute an approximation using one FFT of order .

In practice, the above method is more powerful. Indeed, if is a truncated power series, then the right-hand side of (6.1) is usually of the order for a small . Also, in the favorable but frequent case when the maximal value of is obtained near a unit which “clearly dominates the others” (this case typically occurs when we approach an isolated singularity), one may consider the shifted polynomial and apply Newton's method near in order to efficiently find high precision approximations of . If the upper bound for was pessimistic, one may also directly recompute the Taylor expansion of at order and apply Newton's method for this series. This allows us to use a much sharper bound for the tail of the expansion of on than (5.4). Alternatively, one may investigate the use of a steepest descent method. Notice that the method may still be applied in the slightly less favorable case of a small number of units which dominate the others.

Remark 6.4. One feature of the above method is that it can easily be applied to the computation of approximations of

Indeed, it suffices to replace and by the corresponding , and , . The efficient computation of and is interesting in order to compute upper bounds for resp. on compact disks. In the case of , one needs to require that has no roots on , so that .

Remark 6.5. The previous remark actually generalizes to extrema of the form

where is a more general continuous and real-valued function which can be evaluated efficiently. However, suitable analogues of (6.1) are harder to obtain in that case.

6.3.Relaxed Taylor series and bounds for the remainders

In sections 6.1 and 6.2, an important ingredient of the algorithms is the computation of a bound for the tail of the power series expansion of on a compact disk . Until now, sharp bounds for the tail were obtained by computing a rough bound on a slightly larger disk and using Cauchy's formula. However, if is pessimistic, then we will have to choose quite large in order to reduce the bound for . This raises the questing of finding more direct ways for bounding on . In this section, we will see how to adapt the strategies of lazy and relaxed computations with formal power series in order to directly take into account error bounds for the tails.

Notations
Given a power series and , we will denote:

Assuming algorithms for the computation of bounds and for resp. on , we will also denote by the resulting bound for on . Finally, in the case when , then we will abbreviate , , etc. by , and so on.

Relaxed power series
We recall that the technique of lazy computations with formal power series relies on the observation that solutions to implicit equations usually can be put into a form which expresses the -th coefficient of a solution in terms of the previous ones. For instance, if with , then the formula yields a way to compute the coefficients of using

In the case of relaxed computation [vdH02b], additional tricks are used so as to accelerate these computations using FFT-multiplication. This enables us to compute coefficients in time , where corresponds to the complexity of multiplication of polynomials of degree . The lazy and relaxed strategies have the big advantage that the resolution of a functional equation can be done in approximately the same time as the evaluation of the defining implicit equation.

One disadvantage of FFT-multiplication is that it increases numerical instability in the case when the coefficients do not have the same orders of magnitude. Using transformations of the kind , where is the “numerical” radius of convergence of , it has been shown in [vdH02b, Section 6.2] how to reduce this numerical instability. In our case, we are rather interested in the computation of -approximations of for . Assume that is the solution of some implicit equation using the operations , , , , , and . Using the rules

we may then construct an implicit equation for which can be evaluated as efficiently as itself. Without loss of generality, we may thus assume that and compute -approximations for the coefficients for an which does not depend on . If we need coefficients, usually suffices. This trick therefore reduces the general case to fixed point arithmetic and FFT-multiplication of degree polynomials only accounts for a precision loss of digits.

Bounds for the remainders
Having computed , we have seen in the previous section how to compute a bound for . The next question is to compute a bound for . Clearly, we may take

(6.2)
(6.3)
(6.4)

where

can be computed in time . One may also compute a bound for using automatic differentiation. For especially nice postcompositions, one may take:

(6.5)
(6.6)

For more general postcompositions with , with , and , one may use

The case of convolution products will be discussed below.

Implicit equations
Let us now show how to deal with implicit equations. We start with the case when for some expression which involves operations for which we can compute bounds of the type (6.26.6). When making the hypothesis that for some , we may formally compute the bound . If , then we claim that the hypothesis was correct and that we may indeed take . Indeed, since the formulas (6.26.6) are positive and real analytic, the function is real analytic with a power series expansion which is positive at the origin. Therefore, forms a sequence of analytic functions on which converges uniformly to and such that for all . By continuity, it follows that .

In order to find the smallest fixed-point of , we may use the secant method:

If for some or if exceeds a given threshold, then the method fails and we set . Otherwise, converges quadratically to . As soon as , for some given , we check whether for , in which case we stop. The resulting is an approximation of with relative accuracy .

The above technique generalizes to systems of implicit equations. In this case, the hypothesis and the bound are vectors and the secant method becomes:

where

We may also consider systems such that is recursively built up using the standard operations , , , , etc., together with extra operations like and which involve the recursive resolution of other systems of implicit equations. Indeed, theoretically speaking, such a system may be rewritten as one big system of the above kind. In practice however, we also want to preserve the lazy computation paradigm, which can be achieved by storing the hypotheses and the corresponding bounds in a hash table, which is passed as a reference to the bound computation method.

Lower bounds for the radius of convergence
Let be arbitrary. Modulo a transformation of the type , the above algorithms can be used in order to compute a possibly infinite upper bound for . In particular, when applying this method for different values of , we obtain an algorithm for computing a lower bound for . Indeed, we keep decreasing or increasing depending on whether resp. . More precisely, assuming that for a starting approximation and , we keep setting and at each iteration, until we obtain an adequate precision. When a starting approximation is not beforehand, one may use a second iteration resp. in order to obtain a reasonable value for , while taking .

Let us now consider the dependence of the computation of for a solution to as a function of (assuming that we perform the necessary scalings for each ). When the implicit equation was constructed using , , , and recursive solutions to implicit equations of the same kind, then it can be checked that

(6.7)

for . Consequently, the function indeed does have a fixed point for sufficiently small , and our algorithm yields a computable lower bound for . In particular, our technique can be used as an alternative for the classical majorant method [vK75, vdH03]. Moreover, it easily adapts to slightly more general functional equations, which involve composition or other operations: it suffices to check that (6.7) holds for .

Assuming that lower bounds for radii of convergence are computed as above, we claim that coincides with the largest theoretical simply connected Riemann surface on which and are defined. In order to see this, we first observe that the algorithm for computing may theoretically be applied to arbitrary paths and with . Since was constructed using the common operations , , , , etc., we have whenever and depends continuously on and . Consequently, the supremum

is lower continuous in . Now assume for contradiction that and take

Setting , there exists a neighbourhood of with for all . Taking with , we thus obtain . This contradiction completes the proof of our claim. Notice the analogy with [vdH05a, Theorem 3].

Composition equations
The case of implicit equations which involve compositions has to be treated with additional care. For instance, consider an equation of the type

(6.8)

Assuming that the equation admits a solution at the origin, its analytic continuation to requires the prior analytic continuation of to for any and . Naive implementations may therefore lead to infinite loops.

One solution to this problem is to introduce a “freezing” operator . Given , the function is the restriction of to its current Riemann surface . In particular, for all . Then we may replace (6.8) by

This approach avoids infinite loops, by handing over to the user the responsibility of ensuring that all values with are already defined. Of course, this may be automatized by trying brutal continuations in all directions. One may also consider delayed freezing operators , which only freeze after postcompositions.

In the very particular case when the generate a finite group for the composition operator, we notice that (6.8) may be rewritten as a system of equations in the unknowns with . After a local resolution at the origin, these equations do no longer involve composition. A particularly important special case of this situation is when and with .

Convolution equations
The power series expansion of the analytic continuation of a convolution product may be computed using (5.13) and (5.14). Unfortunately, the translation of a power series by a small is not very convenient for relaxed computations, which naturally occur if and are unknowns in a convolution equation [É85], such as

Nevertheless, in the equation (5.14), the functions and are known except when resp. . Modulo one subdivision of the path, we may also assume without loss of generality that . This reduces the resolution of the convolution equation to the problem of determining the coefficients of at a small as a function of the coefficients of at in a relaxed manner, assuming that the coefficients of at are already known. Now we may again write

(6.9)

The coefficients of may be computed in a relaxed manner by what precedes. The second member may be expanded in using

(6.10)

However, the evaluation of each ! at a precision of digits still requires a time , which is not very convenient if we want to evaluate up to order . On the other hand, if the power series expansion of has convergence radius , then the translated expansion of still has convergence radius . The idea is now to use (6.9) and (6.10) for the computation of good bounds and not for the expansion of itself, using the formulas

If is close to , then may typically remain finite even for . In that case, we have a method to analytically continue beyond .

Remark 6.6. With the above method, in order to obtain an order expansion of the solution to a convolution equation at a path , one generally needs an order expansion of at the origin, where is more or less proportional to (it also depends on the positions of the singularities of ). It remains an interesting question whether the order can be reduced.

6.4.Improved bounds for remainders of Taylor series

Division
The error bounds computed in section 6.3 are not optimal in the case of division

(6.11)

Indeed, the fixed-point method yields

The denominator is unnecessarily pessimistic: even if exceeds , the function itself might be bounded away from . This is particularly annoying in the case when for large values of . Indeed, when using the fixed-point method in a direct way on this example, the computable radius of convergence of would be instead of .

For this reason, it is good to treat the case of division (6.11) in an ad hoc manner. When rewriting (6.11) in terms of , we obtain the solution

Now we may compute a lower bound for on using the technique from section 6.2. Consequently, we may take

Exponentiation
Similarly, when applying the technique from the previous section to the case of exponentiation

(6.12)

we obtain a bound

Although this bound is a bit better than in the bound for division (roughly speaking, we effectively “see” the part of with ), we again obtain a better ad hoc bound by solving (6.12) in terms of :

Section 6.2 again yields an efficient algorithm for computing order bounds and for and on . We may then take

Implicit equations
Let us now return to the case of a general implicit equation and again consider the decomposition . We may rewrite each subexpression of as , where and are new expressions in , such that corresponds to the “coefficient of in :

Composition is treated in a similar way as integration. Applying the above rules to , we obtain

We now replace the equation by

and compute bounds and as in the previous section with the above improvement for the final division by . In the case of possibly nested systems of implicit equations , subexpressions are decomposed as

where is a vector and stands for the vector product.

Example 6.7. Consider the implicit equation

(6.13)

For , we have

and

for the polynomial with . Then (6.13) is equivalent to

Dynamical systems
Instead of taking in the above case of implicit equations, it would be nice to rather extract the linear part of in . Unfortunately, the resulting linear equation in is often not so easy to solve. Nevertheless, for implicit equations of a particular shape, such a resolution may be feasible. For instance, consider the case of an ordinary differential equation

(6.14)

where is an expression which is also a power series in . We may then rewrite (6.14) as

(6.15)

We next set

for polynomials of degree and numbers and which are approximated at successive stages using the secant method. Then (6.15) admits an explicit solution

Now order upper bounds for and can be computed using the method from section 6.2. Then we may take

With some more work, this method can be adapted to the case of systems of ordinary differential equations (6.14), with and . The case when is polynomial can also be treated with the majorant technique [vdH03, Section 5].

6.5.Approaches for limiting the precision loss

Computing in the jet space
Consider the solution to some ordinary differential equation with given initial conditions at the origin. Assuming that are given by complex ball representations , we may in principle compute coefficients of using complex ball arithmetic. However, this may lead to overestimation of the error due to the fact that we do not keep track of possible cancellations between the errors in during the computation.

One approach to this problem is to use Taylor models [MB96, Ber98] in which we consider as formal parameters, with . Instead of computing with coefficients in , we now compute with coefficients in the jet space

For and with , we take . Given , the constant coefficient is stored at a precision which is one or a few words higher than the precision of the .

Taylor models can be used in many variants. For instance, each of the coefficients with may be taken to be finite precision floating point numbers instead of balls, in which case rounding errors are incorporated into the error of . If , then one may also take (), which allows for the computations of bounds for the derivatives in the parameters . If is continued analytically from to , then we also notice that the initial conditions at may again be taken in the jet space for the errors at . This is useful for the computation of return maps and limit cycles. When the constant coefficients of such jets become less precise than , it may sometimes be useful to unjettify and replace each by . We next rejettify the vector by replacing each by .

Remark 6.8. The jet-space technique can also be used for studying the dependence of the analytic continuation of on initial conditions. For instance, return maps for limit cycles may be computed in such a way.

Remark 6.9. Clearly, the technique of jettification is not limited to differential equations: it applies to more general functional equations whose local expansions are determined by the equation in terms of a finite number of initial conditions.

The wrapping effect
A well known problem with certified integration of dynamical systems using interval methods is the wrapping effect [Moo66, Loh01, MB04]. Consider a simple equation like

Given an initial condition

at , integration of the equation from to yields with

Now if is given by an enclosing rectangle, then left multiplication by turns this rectangle by , so that we lose bit of precision when enclosing the result by a new rectangle.

Now a similar problem is encountered when using complex interval arithmetic in order to compute the -th power of using . Therefore, one possible remedy is to adapt ball arithmetic to matrices and represent transition matrices such as by balls , where is an exact matrix and denotes the space of matrices with norm (i.e. for all vectors ). In this representation, taking the (naive) -th power of the matrix only gives rise to a precision loss of bits. Although the above idea applies well to matrices whose eigenvalues are of approximately the same order of magnitude, the error bounds may again be pessimistic in other cases. In addition, one may therefore adapt the norms in an incremental manner using numerical preconditioning. Equivalently, one may adapt the coordinate system as a function of [Loh88, MB04].

We notice that the divide and conquer technique may also be used to the wrapping effect. Indeed, in the case of linear equations, the transition matrices verify the relation

Instead of computing in the naive way, using

one may use binary splitting:

Even in the case when ordinary interval arithmetic is used in order to represent the matrices , the computation of using this technique gives rise to a precision loss of only bits. With some more work, this technique also applies to non-linear equations.

Preservation laws
In the case when a dynamical system is subject to preservation laws or symmetries, then one may project the bounding region for the numerical solution on the variety of actual solutions, after each numerical integration step. In lucky cases, this may help to further reduce overestimation and precision loss.

Acknowledgment
The author would like to thank the anonymous referees for their careful reading an corrections.

Glossary

Encoding for an effective set 4

Effective set of computable functions from into 4

Encoding of 5

Set of digital numbers 5

Set of strictly positive elements in 5

Set of positive or zero elements in 5

Set of approximators of real numbers 5

Set of computable real numbers 5

Set of left computable real numbers 6

Set of right computable real numbers 6

Dot notation for arguments 7

Scale of the encoding of a digital Riemann surface 9

Nodes of the encoding of a digital Riemann surface 9

Projection of on 9

Adjacency relation on 9

Circular adjacency relation 10

Square associated to a node 10

Set of digital points on 11

Set of computable points on 12

Normalization of a digital Riemann pasting 12

Set of digital coverings 13

Open ball of radius and center resp. 16

Closed ball of radius and center resp. 16

Disjoint union of and 20

Covering product of and 20

Join of at and at 21

Set of broken line paths 22

Root of a Riemann surface 22

Set of rooted Riemann surfaces 22

Path domain of 22

Organic Riemann surface associated to 24

Covering space of 25

Convolution product of and 28

Radius of convergence of 28

Maximum of on disk of radius 28

Analytic continuation of along the straightline segment 28

Effective lower bound for 28

Effective upper bound for 28

Set of locally computable analytic functions 29

Path domain of 30

improves 30

Set of weak locally computable functions 31

Set of computable analytic functions 32

Projection on 32

Set of incrementally computable Riemann surfaces 33

Extension mapping for incrementally computable Riemann surfaces 33

The head 42

The tail 42

The restriction 42

Bibliography

[AH83]

G. Alefeld and J. Herzberger. Introduction to interval analysis. Academic Press, 1983.

[Alb80]

O. Alberth. Computable analysis. McGraw-Hill, 1980.

[BB56]

C.A. Briot and J.C. Bouquet. Propriétés des fonctions définies par des équations différentielles. Journal de l'École Polytechnique, 36:133–198, 1856.

[BB85]

E. Bishop and D. Bridges. Foundations of constructive analysis. Die Grundlehren der mathematische Wissenschaften. Springer, Berlin, 1985.

[BBH01]

J. Blanck, V. Brattka, and P. Hertling, editors. Computability and complexity in analysis, volume 2064 of Lect. Notes in Comp. Sc. Springer, 2001.

[Ber98]

M. Berz. Cosy infinity version 8 reference manual. Technical Report MSUCL-1088, Michigan State University, 1998.

[BK75]

R.P. Brent and H.T. Kung. algorithms for composition and reversion of power series. In J.F. Traub, editor, Analytic Computational Complexity, 1975. Proc. of a symposium on analytic computational complexity held by Carnegie-Mellon University.

[BK78]

R.P. Brent and H.T. Kung. Fast algorithms for manipulating formal power series. Journal of the ACM, 25:581–595, 1978.

[Bla02]

J. Blanck. General purpose exact real arithmetic. Technical Report CSR 21-200, Luleå University of Technology, Sweden, 2002. http://www.sm.luth.se/~jens/.

[BPR03]

S. Basu, R. Pollack, and M.F. Roy. Algorithms in real algebraic geometry, volume 10 of Algorithms and computation in mathematics. Springer-Verlag, 2003.

[Bre76a]

R.P. Brent. The complexity of multiprecision arithmetic. In Complexity of Computational problem solving, 1976.

[Bre76b]

R.P. Brent. Fast multiple-precision evaluation of elementary functions. Journal of the ACM, 23(2):242–251, 1976.

[Car96]

J. P. Cardinal. On two iterative methods for approximating the roots of a polynomial. In J. Renegar, M. Shub, and S. Smale, editors, Proc. AMS-SIAM Summer Seminar on Math. of Numerical Analysis, volume 32 of Lectures in Applied Math., pages 165–188, Providence, 1996. American Mathematical Society Press. Park City, Utah.

[CC90]

D.V. Chudnovsky and G.V. Chudnovsky. Computer algebra in the service of mathematical physics and number theory (computers in mathematics, stanford, ca, 1986). In Lect. Notes in Pure and Applied Math., volume 125, pages 109–232, New-York, 1990. Dekker.

[CT65]

J.W. Cooley and J.W. Tukey. An algorithm for the machine calculation of complex Fourier series. Math. Computat., 19:297–301, 1965.

[DL89]

J. Denef and L. Lipshitz. Decision problems for differential equations. The Journ. of Symb. Logic, 54(3):941–950, 1989.

[É85]

J. Écalle. Les fonctions résurgentes I–III. Publ. Math. d'Orsay 1981 and 1985, 1985.

[FHL+05]

L. Fousse, G. Hanrot, V. Lefèvre, P. Pélissier, and P. Zimmermann. Mpfr: A multiple-precision binary floating-point library with correct rounding. Technical Report RR-5753, INRIA, 2005.

[GPR03]

M. Grimmer, K. Petras, and N. Revol. Multiple precision interval packages: Comparing different approaches. Technical Report RR 2003-32, LIP, École Normale Supérieure de Lyon, 2003.

[Grz55]

A. Grzegorczyk. Computable functionals. Fund. Math., 42:168–202, 1955.

[Grz57]

A. Grzegorczyk. On the definitions of computable real continuous functions. Fund. Math., 44:61–71, 1957.

[JKDW01]

L. Jaulin, M. Kieffer, O. Didrit, and E. Walter. Applied interval analysis. Springer, London, 2001.

[Kar91]

E.A. Karatsuba. Fast evaluation of transcendental functions. Problems of Information Transmission, 27:339–360, 1991.

[KO63]

A. Karatsuba and J. Ofman. Multiplication of multidigit numbers on automata. Soviet Physics Doklady, 7:595–596, 1963.

[Lam06]

B. Lambov. The RealLib project. http://www.brics.dk/~barnie/RealLib, 2001–2006.

[Loh88]

R. Lohner. Einschließung der Lösung gewöhnlicher Anfangs- und Randwertaufgaben und Anwendugen. PhD thesis, Universität Karlsruhe, 1988.

[Loh01]

R. Lohner. On the ubiquity of the wrapping effect in the computation of error bounds. In U. Kulisch, R. Lohner, and A. Facius, editors, Perspectives of enclosure methods, pages 201–217. Springer, 2001.

[MB96]

K. Makino and M. Berz. Remainder differential algebras and their applications. In M. Berz, C. Bischof, G. Corliss, and A. Griewank, editors, Computational differentiation: techniques, applications and tools, pages 63–74, SIAM, Philadelphia, 1996.

[MB04]

K. Makino and M. Berz. Suppression of the wrapping effect by taylor model-based validated integrators. Technical Report MSU Report MSUHEP 40910, Michigan State University, 2004.

[Mül00]

N. Müller. iRRAM, exact arithmetic in C++. http://www.informatik.uni-trier.de/iRRAM/, 2000.

[MM96]

V. Ménissier-Morain. Arbitrary precision real arithmetic: design and algorithms. http://www-calfor.lip6.fr/~vmm/documents/submission_JSC.ps.gz, 1996.

[Moo66]

R.E. Moore. Interval analysis. Prentice Hall, Englewood Cliffs, N.J., 1966.

[MP00]

M. N. Minh and M. Petitot. Lyndon words, polylogarithms and the Riemann function. Discrete Maths, 217(1–3):273–292, 2000.

[Neu90]

A. Neumaier. Interval methods for systems of equations. Cambridge university press, Cambridge, 1990.

[O'C05]

R. O'Connor. A monadic, functional implementation of real numbers. Technical report, Institute for Computing and Information Science, Radboud University Nijmegen, 2005.

[Pan96]

Victor Y. Pan. On approximating complex polynomial zeros: Modified quadtree (Weyl's) construction and improved Newton's iteration. Technical Report RR-2894, INRIA Sophia, 1996.

[Ric92]

D. Richardson. The elementary constant problem. In Proc. ISSAC '92, pages 108–116, 1992.

[Ric97]

D. Richardson. How to recognise zero. JSC, 24:627–645, 1997.

[SS71]

A. Schönhage and V. Strassen. Schnelle Multiplikation grosser Zahlen. Computing, 7:281–292, 1971.

[Tur36]

A. Turing. On computable numbers, with an application to the Entscheidungsproblem. Proc. London Maths. Soc., 2(42):230–265, 1936.

[vdH97]

J. van der Hoeven. Automatic asymptotics. PhD thesis, École polytechnique, France, 1997.

[vdH99a]

J. van der Hoeven. Fast evaluation of holonomic functions. TCS, 210:199–215, 1999.

[vdH99b]

J. van der Hoeven. GMPX, a C-extension library for gmp. http://www.math.u-psud.fr/~vdhoeven/, 1999. No longer maintained.

[vdH01a]

J. van der Hoeven. Fast evaluation of holonomic functions near and in singularities. JSC, 31:717–743, 2001.

[vdH01b]

J. van der Hoeven. Zero-testing, witness conjectures and differential diophantine approximation. Technical Report 2001-62, Prépublications d'Orsay, 2001.

[vdH02a]

J. van der Hoeven. A new zero-test for formal power series. In Teo Mora, editor, Proc. ISSAC '02, pages 117–122, Lille, France, July 2002.

[vdH02b]

J. van der Hoeven. Relax, but don't be too lazy. JSC, 34:479–542, 2002.

[vdH03]

J. van der Hoeven. Majorants for formal power series. Technical Report 2003-15, Université Paris-Sud, Orsay, France, 2003.

[vdH05a]

J. van der Hoeven. Effective complex analysis. JSC, 39:433–449, 2005.

[vdH05b]

J. van der Hoeven. Efficient accelero-summation of holonomic functions. Technical Report 2005-54, Univ. Paris-Sud, Orsay, France, 2005. Accepted for JSC.

[vdH06a]

J. van der Hoeven. Computations with effective real numbers. TCS, 351:52–60, 2006.

[vdH06b]

J. van der Hoeven. Effective real numbers in Mmxlib. In D. Saunders, editor, Proc. ISSAC '06, Genova, Italy, July 2006.

[vdH07]

J. van der Hoeven. Around the numeric-symbolic computation of differential Galois groups. JSC, 42(1–2):236–264, 2007.

[vK75]

S. von Kowalevsky. Zur Theorie der partiellen Differentialgleichungen. J. Reine und Angew. Math., 80:1–32, 1875.

[Wei00]

K. Weihrauch. Computable analysis. Springer-Verlag, Berlin/Heidelberg, 2000.

[Zei90]

D. Zeilberger. A holonomic systems approach to special functions identities. Journal of Comp. and Appl. Math., 32:321–368, 1990.