ETNA Kent State University

Similar documents
On rational approximation of algebraic functions. Julius Borcea. Rikard Bøgvad & Boris Shapiro

Simultaneous Gaussian quadrature for Angelesco systems

Notes on Complex Analysis

INTRODUCTION TO PADÉ APPROXIMANTS. E. B. Saff Center for Constructive Approximation

5 Measure theory II. (or. lim. Prove the proposition. 5. For fixed F A and φ M define the restriction of φ on F by writing.

PADÉ INTERPOLATION TABLE AND BIORTHOGONAL RATIONAL FUNCTIONS

HERMITE-PADÉ APPROXIMANTS FOR A PAIR OF CAUCHY TRANSFORMS WITH OVERLAPPING SYMMETRIC SUPPORTS

Vectors. January 13, 2013

Spurious Poles in Padé Approximation of Algebraic Functions

arxiv: v2 [math.ca] 22 Apr 2016

Orthogonal polynomials with respect to generalized Jacobi measures. Tivadar Danka

Considering our result for the sum and product of analytic functions, this means that for (a 0, a 1,..., a N ) C N+1, the polynomial.

Algorithmic Approach to Counting of Certain Types m-ary Partitions

Large Deviations, Linear Statistics, and Scaling Limits for Mahler Ensemble of Complex Random Polynomials

MATH 722, COMPLEX ANALYSIS, SPRING 2009 PART 5

On probability measures with algebraic Cauchy transform

Zeros of Polynomials: Beware of Predictions from Plots

Solutions to Complex Analysis Prelims Ben Strasser

THE ENVELOPE OF LINES MEETING A FIXED LINE AND TANGENT TO TWO SPHERES

(x 1, y 1 ) = (x 2, y 2 ) if and only if x 1 = x 2 and y 1 = y 2.

Operations On Networks Of Discrete And Generalized Conductors

1 Holomorphic functions

02. Measure and integral. 1. Borel-measurable functions and pointwise limits

Analysis-3 lecture schemes

ULTRASPHERICAL TYPE GENERATING FUNCTIONS FOR ORTHOGONAL POLYNOMIALS

Continued fractions for complex numbers and values of binary quadratic forms

Chebyshev coordinates and Salem numbers

Weighted Sums of Orthogonal Polynomials Related to Birth-Death Processes with Killing

Introduction to Arithmetic Geometry Fall 2013 Lecture #23 11/26/2013

ON SPECTRAL POLYNOMIALS OF THE HEUN EQUATION. II.

INDEX. Bolzano-Weierstrass theorem, for sequences, boundary points, bounded functions, 142 bounded sets, 42 43

Complex Analysis MATH 6300 Fall 2013 Homework 4

1. If 1, ω, ω 2, -----, ω 9 are the 10 th roots of unity, then (1 + ω) (1 + ω 2 ) (1 + ω 9 ) is A) 1 B) 1 C) 10 D) 0

LINEAR EQUATIONS WITH UNKNOWNS FROM A MULTIPLICATIVE GROUP IN A FUNCTION FIELD. To Professor Wolfgang Schmidt on his 75th birthday

I. ANALYSIS; PROBABILITY

Stability of Feedback Solutions for Infinite Horizon Noncooperative Differential Games

Functional equations for Feynman integrals

METRIC HEIGHTS ON AN ABELIAN GROUP

+ 2x sin x. f(b i ) f(a i ) < ɛ. i=1. i=1

We simply compute: for v = x i e i, bilinearity of B implies that Q B (v) = B(v, v) is given by xi x j B(e i, e j ) =

9. Integral Ring Extensions

Complex Analysis review notes for weeks 1-6

2. Intersection Multiplicities

PARTIAL FRACTIONS: AN INTEGRATIONIST PERSPECTIVE

Dirichlet s Theorem. Calvin Lin Zhiwei. August 18, 2007

Part II. Riemann Surfaces. Year

MATH 326: RINGS AND MODULES STEFAN GILLE

(U) =, if 0 U, 1 U, (U) = X, if 0 U, and 1 U. (U) = E, if 0 U, but 1 U. (U) = X \ E if 0 U, but 1 U. n=1 A n, then A M.

A VERY BRIEF REVIEW OF MEASURE THEORY

Notation. General. Notation Description See. Sets, Functions, and Spaces. a b & a b The minimum and the maximum of a and b

Estimates for Bergman polynomials in domains with corners

R1: Sets A set is a collection of objects sets are written using set brackets each object in onset is called an element or member

S-adic sequences A bridge between dynamics, arithmetic, and geometry

Constructions with ruler and compass.

Multiple Orthogonal Polynomials

ELEMENTARY LINEAR ALGEBRA

MATH 56A: STOCHASTIC PROCESSES CHAPTER 1

Matrix Functions and their Approximation by. Polynomial methods

arxiv: v2 [math.ds] 13 Sep 2017

EXACT INTERPOLATION, SPURIOUS POLES, AND UNIFORM CONVERGENCE OF MULTIPOINT PADÉ APPROXIMANTS

1 Structures 2. 2 Framework of Riemann surfaces Basic configuration Holomorphic functions... 3

12. Hilbert Polynomials and Bézout s Theorem

PICARD S THEOREM STEFAN FRIEDL

Here are brief notes about topics covered in class on complex numbers, focusing on what is not covered in the textbook.

Chapter 1. Measure Spaces. 1.1 Algebras and σ algebras of sets Notation and preliminaries

APPROXIMABILITY OF DYNAMICAL SYSTEMS BETWEEN TREES OF SPHERES

CONTENTS. Preface Preliminaries 1

Dirichlet s Theorem. Martin Orr. August 21, The aim of this article is to prove Dirichlet s theorem on primes in arithmetic progressions:

MATH 205C: STATIONARY PHASE LEMMA

Chapter 4: Open mapping theorem, removable singularities

Course 311: Michaelmas Term 2005 Part III: Topics in Commutative Algebra

Module 2: Reflecting on One s Problems

F (z) =f(z). f(z) = a n (z z 0 ) n. F (z) = a n (z z 0 ) n

PROBLEMS. (b) (Polarization Identity) Show that in any inner product space

The Geometry of Cubic Maps

Page 404. Lecture 22: Simple Harmonic Oscillator: Energy Basis Date Given: 2008/11/19 Date Revised: 2008/11/19

Multiple Orthogonal Polynomials

1 Roots of polynomials

SOME PROPERTIES OF THE CANONICAL DIVISOR IN THE BERGMAN SPACE

Complex Analysis Qualifying Exam Solutions

ETNA Kent State University

ASYMPTOTIC MAXIMUM PRINCIPLE

Part V. 17 Introduction: What are measures and why measurable sets. Lebesgue Integration Theory

1 Introduction. or equivalently f(z) =

JORDAN AND RATIONAL CANONICAL FORMS

MATH41011/MATH61011: FOURIER SERIES AND LEBESGUE INTEGRATION. Extra Reading Material for Level 4 and Level 6

POLYNOMIALS WITH COEFFICIENTS FROM A FINITE SET

Mathematical Methods for Engineers and Scientists 1

Möbius Transformation

Chapter 2 Formulas and Definitions:

Math Homework 2

Katholieke Universiteit Leuven Department of Computer Science

MATH 320, WEEK 11: Eigenvalues and Eigenvectors

Canonical lossless state-space systems: staircase forms and the Schur algorithm

SUMMATION TECHNIQUES

Contents Ordered Fields... 2 Ordered sets and fields... 2 Construction of the Reals 1: Dedekind Cuts... 2 Metric Spaces... 3

Infinite Series. 1 Introduction. 2 General discussion on convergence

MCPS Algebra 2 and Precalculus Standards, Categories, and Indicators*

The concentration of the chromatic number of random graphs

Herbert Stahl s proof of the BMV conjecture

Transcription:

Electronic Transactions on Numerical Analysis. Volume 4, pp. 95-222, 2002. Copyright 2002,. ISSN 068-963. ETNA ASYMPTOTICS FOR QUADRATIC HERMITE-PADÉ POLYNOMIALS ASSOCIATED WITH THE EXPONENTIAL FUNCTION HERBERT STAHL Abstract. The asymptotic behavior of quadratic Hermite-Padé polynomials p n, q n, r n P n of type I and p n, q n, r n P 2n of type II associated with the exponential function are studied. In the introduction the background of the definition of Hermite-Padé polynomials is reviewed. The quadratic Hermite-Padé polynomials p n, q n, r n P n of type I are defined by the relation p n(z) + q n(z)e z + r n(z)e 2z = O(z 3n+2 ) as z 0, and the polynomials p n, q n, r n P 2n of type II by the two relations p n(z)e z q n(z) = O(z 3n+ ) as z 0, p n(z)e 2z r n(z) = O(z 3n+ ) as z 0. Analytic descriptions are given for the arcs, on which the contracted zeros of both sets of the polynomials {p n, q n, r n} and {p n, q n, r n} cluster as n. Analytic expressions are also given for the density functions of the asymptotic distributions of these zeros. The description is based on an algebraic function of third degree and a harmonic function defined on the Riemann surface, which is associated with the algebraic function. The existence and basic properties of the asymptotic distributions of the zeros and the arcs on which these distributions live are proved, the asymptotic relations themselves are only conjectured. Numerical calculations are presented, which demonstrate the plausibility of these conjectures. Key words. Quadratic Hermite-Padé polynomials of type I and type II, the exponential function, German and Latin polynomials, Hermite-Padé approximants. AMS subject classifications. 4A2, 30E0.. Introduction and Review. The first part of the talk has review character, the polynomials considered in the second part of the talk are special cases of general Hermite-Padé polynomials, which can be seen as a generalization of Padé polynomials, and therefore also as generalizations of orthogonal polynomials. The approximants associated with the Hermite- Padé polynomials are generalizations of the concepts of Padé approximants and continued fractions. In order to motivate the definition of Hermite-Padé polynomials we start by repeating the definition of diagonal Padé approximants and their associated orthogonal polynomials... Padé Approximants. Let the function f be analytic at the origin. Then there exist polynomials p n, q n P n such that (.) p n (z) q n (z)f(z) = O(z 2n+ ) as z 0 with O( ) denoting Landau s big oh. The rational function [n/n] f := p n /q n is uniquely determined by (.), and it is called the diagonal Padé approximant of degree (n, n) to the function f. The denominator polynomial q n of [n/n] f can be characterized by orthogonality. Let Q n (z) := z n q n (/z) denote the reversed polynomial of q n. Then the polynomial q n satisfies relation (.) (together with an appropriate choice of p n ) if and only if the reversed polynomial Q n satisfies the orthogonality relation (.2) ζ l Q n (ζ)f(/ζ)dζ = 0, l = 0,..., n, C Received December 2, 2000. Accepted for publication May 5, 200. Communicated by Sven Ehrich. TFH-Berlin/FB II; Luxemburger Strasse 0; 3353 Berlin; Germany (stahl@tfh-berlin.de). The research has been done while the author was visiting the INRIA, Sophia-Antipolis. 95

96 Herbert Stahl where C is an integration path around infinity. In case of Markov, Stieltjes, or Hamburger functions, relation (.2) transforms into an orthogonality relation defined by a positive measure on R (cf. [3], Chapter 5.3, [7], Chapter 2, or [8], Kapitel 9 & 0). We have f [n/n] f for n sufficiently large if and only if f is a rational function. This assertion is known as Kronecker s Theorem. In a letter to Jacobi, Hermite raised the question whether something similar, but more general than continued fractions (more general than Padé approximants, we would say today) could be defined so that the algebraic character of a function with degree m > could be detected in the same way as rational functions can be detected by continued fractions (or Padé approximants) via Kronecker s Theorem. (Note that rational functions are algebraic of degree.) A discussion of the correspondence between Hermite and Jacobi can be found in [5], page -, where one also finds further references about this topic..2. Hermite-Padé Polynomials. The question posed by Hermite led to the introduction of what is now called the Jacobi-Perron algorithm (cf. [], [9], [7], Chapter 4.5) and further to the definition of Hermite-Padé polynomials and their associated approximants. Here, we will not discuss the Jacobi-Perron algorithm, which is a generalization of the continued fraction approach, and instead concentrate on the Hermite-Padé polynomials and their associated approximants. Let f = (f 0,..., f m ), m, be a system of functions, which are analytic in a neighborhood of the origin. DEFINITION.. Hermite-Padé Polynomials of Type I (Latin polynomials in K. Mahler s terminology [6]): For any multi-index n = (n 0,..., n m ) N m+ there exists a vector of polynomials (p 0,..., p m ) P n 0 P n... P nm such that (.3) m p j (z)f j (z) = O(z n ) as z 0, j=0 where n := n 0 +... + n m and P k := { p P k p monic, p 0}. The vector (p 0,..., p m ) is called Hermite-Padé form of type I, and its elements are the Hermite-Padé polynomials of type I. DEFINITION.2. Hermite-Padé Polynomials of Type II (German polynomials in K. Mahler s terminology [6]): For any multi-index n = (n 0,..., n m ) N m+ there exists a vector of polynomials (p 0,..., p m ) P N 0 P N... P Nm with N j := n n j, j = 0,..., m, such that (.4) p i (z)f j (z) p j (z)f i (z) = O(z n + ) as z 0, for i, j = 0,..., m, i j. The vector (p 0,..., p m ) is called Hermite-Padé form of type II, and its elements are the Hermite-Padé polynomials of type II. Remarks: () The assumption p 0 P n 0 and p 0 P N 0 implies a normalization of the whole form (p 0,..., p m ) and (p 0,..., p m ), respectively. There may exist situations in which a normalization by the first component is not possible; however, there always exists one of the m + components by which a normalization is possible. (2) In general the Hermite-Padé polynomials are not unique. Their existence is immediate since both relations (.3) and (.4) can be rewritten as homogeneous linear systems of equations for the coefficients of the polynomials. (3) In (.4) there are (m + )m/2 formally different relations. However, at most m of these relations are linearly independent.

Hermite-Padé Polynomials 97.3. Hermite-Padé Approximants. If one takes m =, f 0, and f = f, then relation (.3) reduces to (.), and by taking m =, f 0, and f = f, in relation (.4) one again gets relation (.). Hence, the concept of Padé polynomials p n, q n P n defined by (.) is a special case of both types of Hermite-Padé polynomials. If f 0 (0) 0, then one can assume without loss of generality that f 0, and under this assumption we deduce from (.4) that (.5) p 0 (z)f j (z) p j (z) = O(z n + ) as z 0 for j =,..., m. DEFINITION.3. Hermite-Padé Simultaneous Rational Approximants: For a given multi-index n N m+ let p 0,..., p m be the Hermite-Padé polynomials of type II define by (.5). Then the vector of rational functions (.6) ( p (z),..., p ) m (z) p 0 p 0,with common denominator polynomial p 0 is called the (Hermite-Padé) simultaneous rational approximant to the (reduced) system of functions f red = (f,..., f m ). For m = we have the Padé approximants to f with numerator and denominator degrees (n 0, n ) as special case of (.6). Besides the simultaneous rational approximants there exists a second type of approximants: the algebraic Hermite-Padé approximants. They are defined with the help of Hermite- Padé polynomials of type I. Let f be a function analytic at the origin, and define the system of functions f as (.7) f = (f 0,..., f m ) := (, f,..., f m ). DEFINITION.4. Algebraic Hermite-Padé Approximants: For a given multi-index n N m+ let p 0,..., p m Pn... P 0 n m be the Hermite-Padé polynomials of type I defined by (.3) with the special choice of (.7). Let the algebraic function y = y(z) be defined by (.8) m p j (z)y(z) j 0. j=0 From the m branches of y we select the branch y = y n which has the highest contact with f at the origin. This branch y n is the algebraic Hermite-Padé approximant to f associated with the multi-index n. Again, it is immediate that, for m = Definition.4 leads to the Padé approximants [n 0 /n ] f introduced after (.). For m > the two types of Hermite-Padé approximants introduced in Definition.3 and.4 split up in two different directions: In the first case we continue to use rational functions as approximants, but we now approximate simultaneously several functions; in the second case only one function f is approximated, but now by an algebraic function of m th degree so that in principle again m branches of the function f can be approximated simultaneously by the m branches of the approximant. A survey over asymptotics of Hermite-Padé polynomials together with related results about the convergence of Hermite-Padé approximants can be found in [2].

98 Herbert Stahl.4. Orthogonality. As in (.2), and also in case of Hermite-Padé polynomials, one can express the defining relations (.3), (.4) or (.5) by orthogonality in an equivalent way. Here, we discuss the orthogonality relations only for the diagonal case, i.e., we assume that all multi-indices n N m+ are of the form n = (k,..., k) with k N. The reversed polynomials of p 0,..., p m and p 0,..., p m are defined by (.9) P j (z) := z k p j (/z) and P j (z) := z mk p j (/z), j =,..., m. Further, we assume that f 0. Under these assumptions a vector of polynomials (p,..., p m ) P k... P k satisfies relation (.3) together with an appropriately chosen polynomial p 0 Pk if and only if (.0) m ζ l C j= P j (ζ)f j (/ζ)dζ = 0 for l = 0,..., mk. A polynomial p 0 Pmk satisfies relation (.4) together with an appropriately chosen set of polynomials p,..., p m P mk if and only if (.) ζj l P 0(ζ)f j (/ζ)dζ = 0 for l = 0,..., k, j =,..., m. C In both integrals C is an integration path encircling infinity. There also exist orthogonality relations analogous to (.) for the other polynomials P j, j =,..., m, of type II. If the functions f,..., f m are of Markov, Stieltjes, or Hamburger type, then the m orthogonality relations in (.) take a more conventional form (cf. [7], Chapter 4). The polynomial P 0 in (.) is called multiple orthogonal since it satisfies simultaneously m different orthogonality relations, each one with a partial degree n j. The total number of restrictions provided by the multiple orthogonality relation (.) is n n 0 = n +...+n m, which is the maximally possible degree of P 0. The theory of multiple orthogonality is still in an early stage of development. Nevertheless, there exists already a considerable amount of knowledge and special results. Surveys about the state of art in this field can be found in [7], Chapter 4, [], and [23]..5. Mahler Relations. Up to this point one may have the impression that the Definitions. and.2 of both types of Hermite-Padé polynomials are rather independent if m >. But this is not so. Under certain normality assumptions one can calculate one type of polynomials from the other one. The situation is especially simple if the system of functions f is perfect. DEFINITION.5. Perfect Systems: A System of functions f = (f 0,..., f m ) is called perfect if for all multi-indices n N m+ the left-hand side of (.3) starts with the term Az n and A 0. The system is called weakly perfect if this property holds true only for close-to-diagonal multi-indices n N m+. More precisely, it is weakly perfect if it holds true for each multiindex n = (n 0,..., n m ) N m+ with the property that there exists k N with 0 n j k for all j = 0,..., m. The perfectness of special systems has been studied in [6], [7], and [2]. Weak perfectness has been established for Angelesco and Nikishin systems (cf. [7], Chapter 4). In order to define the (diagonal) Mahler relations, we introduce for each k N two sets of multi-indices n(k) 0,..., n(k) m N m+ and n(k) 0,..., n(k) m N m+ with n(k) i =

Hermite-Padé Polynomials 99 (n i0,..., n im ), n(k) i = (n i0,..., n im ), i = 0,..., m. The n ij = n ij (k) and n ij = n ij (k) are defined by (.2) n ij (k) := k + δ ij, n ij (k) := k δ ij, i, j = 0,..., m, k N, where δ ij denotes Kronecker s symbol. We have n(k) i = (m + )k + and n(k) i = (m + )k for i = 0,..., m. For the m+ multi-indices n(k) i N m+, i = 0,..., m, and the system f = (f 0,..., f m ) the m + vectors of Hermite-Padé polynomials of type I are denoted by (.3) (p i0,..., p im ) P k... P k... P k, i = 0,..., m. The assumptions in (.3) imply a normalization different from that used in Definition.. Here, the i th component is used for normalizing the vector (p i0,..., p im ). If the system f is weakly perfect, then this normalization is always possible. For the multi-indices n(k) i N m+, i = 0,..., m, and the system f, the vectors of Hermite-Padé polynomials of type II are denoted by (.4) (p i0,..., p im ) P mk... P mk... P mk, i = 0,..., m. Again, the normalization implied by (.4) is different from that used in Definition.2. The vector of polynomials in (.3) and (.4) are put together in two matrices: p 00 p 0m p 00 p 0m (.5) P :=.., P :=... p m0 p mm p m0 p mm THEOREM.6. Let f = (f 0,..., f m ) be a weakly perfect system. If the matrices P and P of polynomials are defined as in (.2) - (.5), then we have (.6) P P t = z (m+)k I m+ with I m+ denoting the identity matrix with m + rows and columns. Remark. Identity (.6) is called a Mahler relation. It shows that at least in case of weakly perfect systems there exists a functional relationship between the Hermite-Padé polynomials of both types. We shall see in the more detailed investigation of Hermite-Padé polynomials associated with the exponential function that despite of the functional relationship the properties of both types of polynomials can be rather different..6. Hermite-Padé Polynomials to the Exponential Function. In the second part of the talk we are concerned with the asymptotic behavior of quadratic Hermite-Padé polynomials associated with the exponential function. In a general situation a system of exponential functions is of the form (.7) f(z) = (f 0 (z),..., f m (z)) := (, e z,..., e mz ), j =,..., m. From Definition. we know that for a multi-index n = (n 0,..., n m ) N m+ the corresponding Hermite-Padé polynomials p j,n P nj, j =,..., m, of type I are now defined by the relation (.8) p 0,n (z) + p,n (z)e z +... + p m,n (z)e mz = O(z n ) as z 0,

200 Herbert Stahl FIG... The zeros of the polynomials p 30 (stars), q 30 (boxes), r 30 (diamonds), and some of the zeros of the error term e 30 (triangles). (Notice that the axes have different scales.) and the Hermite-Padé polynomials p j,n P n nj, j =,..., m, of type II are defined by the m relations (.9) p j,n (z) p 0,n (z)e jz = O(z n + ) as z 0, j =,..., m. The polynomials of both types have been introduced in [0], and they have been used and intensively studied in number theory and approximation theory (cf. [4]-[6], [7], [24])..7. Quadratic Hermite-Padé Polynomials. After moving from the general problem to the more special situation of a system of exponential functions, we continue to specialize, and restrict our interest now to quadratic Hermite-Padé polynomials, i.e., to the case m = 2, which is the simplest situation that does not coincide with the much studied and well understood case of Padé approximants to the exponential function. The (diagonal, quadratic) Hermite-Padé polynomials of type I associated with the exponential function e z are denoted by p n, q n, r n P n, and they are defined by the relation (.20) e n (z) := p n (z) + q n (z)e z + r n (z)e 2z = O(z 3n+2 ) as z 0. The (diagonal, quadratic) Hermite-Padé polynomials of type II of degree 2n are denoted by p 2n, q 2n, r 2n P 2n, and corresponding to Definition.2, they are defined by the two relations (.2) (.22) e,2n (z) := p 2n (z)e z q 2n (z) = O(z 3n+ ) as z 0, e 2,2n (z) := p 2n (z)e 2z r 2n (z) = O(z 3n+ ) as z 0. Note that in (.20) the polynomials p n, q n, r n are of degree n, and not of degree n, as assumed in Definition..

Hermite-Padé Polynomials 20 FIG..2. The zeros of the polynomials p 60 (stars), q 60 (boxes), r 60 (diamonds). (Notice that the axis have different scales.) The polynomials p n, q n, r n are basic for the definition of quadratic approximants (.23) α n (z) := ( q n (z) ± ) q n (z) 2r n (z) 2 p n (z)r n (z) to e z developed at z = 0, and the polynomials p 2n, q 2n, r 2n lead to simultaneous rational approximants to the system (e z, e 2z ) defined by (.24) r,2n (z) := q 2n /p 2n (z) and r 2,2n (z) := r 2n /p 2n (z). These are the two types of approximants that have been introduced in the Definitions.4 and.3. In the remainder of the talk the asymptotic behavior of the polynomials of both types will be investigated for n. Detailed studies of the polynomials p n, q n, r n can be found in [6], [8], and [9]. In [6] among other things a 4-term recurrence relation and very precise asymptotic estimates for the polynomials p n, q n, r n and for the error term e n have been derived. While in [6], like in (.20), only the diagonal case has been studied, the investigations have been extended to the non-diagonal case in [8] and [9]. Interesting connections with special functions have been established in [9], and the paper contains results about the location of the zeros of the polynomials p n, q n, and r n. Results achieved in [6], [8], and [9] have been extended to the general case (.8) in [24]. In Figure. the zeros of the Hermite-Padé polynomials p n, q n, r n, together with some of the zeros of the error term e n in (.20) are plotted for n = 30. In Figure 2. the zeros of the Hermite-Padé polynomials p 2n, q 2n, r 2n of type II are plotted again for n = 30. The regularity displayed in these plots certainly suggests that there should exist analytic expressions

202 Herbert Stahl that allow to describe the asymptotic behavior of the zeros, and also the asymptotic behavior of the polynomials p n, q n, r n, p 2n, q 2n, r 2n and the error term e n themselves. Results in this direction are the main topic of the next two sections. 2. Asymptotics of Quadratic Hermite-Padé Polynomials. The zeros of the polynomials p n, r n, p 2n, q 2n, r 2n, and nearly all zeros of the polynomial q n tend to infinity as n. Because of this convergence to infinity, many specific aspects of the asymptotic behavior can not be detected in the complex plane. The asymptotic zero distributions and the asymptotics for the polynomials themselves become more informative if the independent variable z is rescaled in such a way that the zeros of the transformed polynomials have finite cluster points as n. This concept has been used successfully by Szegö in [22] for the study of the asymptotic behavior of Taylor polynomials associated with the exponential function, and by Saff and Varga in [2] for the study of zeros and poles of Padé polynomials associated with the exponential function. In the same spirit as in these investigations we introduce as a new independent variable (2.) w := z, n =, 2,..., 3n for the study of the quadratic Hermite-Padé polynomials that will be presented in this talk. The (transformed) polynomials P n, Q n, R n, P n, Q n, and R n are then defined by (2.2) (2.3) P n (w) := p n (3nw), Q n (w) := q n (3nw), R n (w) := r n (3nw), P n (w) := p n (3nw), Q n (w) := q n (3nw), R n (w) := r n (3nw). These new polynomials satisfy the relations (2.4) (2.5) (2.6) E n (w) :=P n (w) ( e 3w) n + Qn (w) + R n (w) ( e 3w) n = O(z 3n+2 ), Q 2n (w)(e 3w ) n P 2n (w) = O(w 3n+ ), Q 2n (w)(e 3w ) n R 2n (w) = O(w 3n+ ) as w 0. These relations are equivalent to the relations (.20) - (.22) in the last section. The error term E n in (.20) is related to e n by E n (w) = e n (3nw)e 3nw. Note that in (.20) not only the variable z has been substituted by 3nw, but the relation has also been multiplied by e w in order to make the symmetry more evident, which is intrinsic to the problem. In a similar way the relations (.2) and (.22) have been transformed in order to get (2.5) and (2.6). The polynomials P n, Q n, R n, and P 2n, Q 2n, R 2n are normalized by assuming that (2.7) P n (w) = w n +... P n and P 2n (w) = w 2n +... P 2n. Since the system (e w,, e w ) is perfect (cf. [6]) the polynomials P n, Q n, R n, and P 2n, Q 2n, R 2n are uniquely determined by (2.4) - (2.7). It then follows from (2.4) that P n (w) = R n ( w) and Q n (w) = Q n ( w), and from (2.5) and (2.6) it follows that P n (w) = R n ( w) and Q n (w) = Q n ( w). 2.. Asymptotic Distributions of Zeros. By Z(p) we denote the (multi) set of zeros of a polynomial p P n (multiplicities of zeros are represented by repetition), by ν p the counting measure (2.8) ν p := x Z(p) δ x, p P n,

Hermite-Padé Polynomials 203 and by the weak convergence of measures in C, i.e., µ n µ means that fdµ n fdµ holds for every real function f continuous on C, as n. The aim in the present investigation is to give an analytic interpretation of the regular configurations of zeros that can be observed in the Figures. and.2. Only parts of the result will be proved in the talk. These are mainly the statements connected with the definition of the asymptotic expressions. On the other hand the asymptotic relations themselves will not be proved. Thus, for instance, the existence of asymptotic distributions for the zeros in the Figures. and.2 is only conjectured. CONJECTURE 2.. There exist six probability measures ω P, ω Q, ω R, ω P, ω Q, ω R on C such that the limits (2.9) (2.0) hold true as n. n ν P n 2n ν P n ω P, ω P, n ν Q n 2n ν Q n ω Q, ω Q, n ν R n ω R, 2n ν R n ω R The supports of the measures ω i, i {P, Q, R, P, Q, R}, are analytic arcs or the union of analytic arcs. The definition of theses arcs is one of the topics in the next subsection. 2.2. The Riemann Surface R. We start by defining a Riemann surface R with three sheets and genus zero together with an algebraic function ψ : R C, which maps R bijectively on C. It will turn out that this Riemann surface is fundamental for all definitions relevant for the description of the asymptotics of the polynomials P n, Q n, R n, P 2n, Q 2n, R 2n. The Riemann surface R is introduced in order to make the function (2.) v w = w(v) := v2 /3 v(v 2 ), v C, bijective. Hence, R is the Riemann surface with canonical projection π : R C and the property that there exists a bijection ψ : R C such that π ψ (v) = w(v) for all v C. The last requirement fully determines the surface R and the mapping ψ. The function ψ is algebraic of third degree. The surface R has three sheets and four simple branch points ζ j, j =,..., 4, over the four base points (2.2) w j := 4 /3e iϕj with ϕ j = 5 2 π, 7 2 Indeed, the derivative (2.3) w(v) = v4 + /3 v 2 (v 2 ) 2 π, 7 2 9 π, π, j =,..., 4. 2 has simple zeros at the four roots v j = 4 /3, j =,..., 4, and it is easy to check that the four points in (2.2) are defined by w j = π ψ (v j ) for j =,..., 4 if the v j s are ordered appropriately. We note that for a given ζ R the value v = ψ(ζ) can be calculated very efficiently by solving the cubic equation (2.4) π(ζ) v (v 2 ) v 2 + 3 = 0. The value v = ψ(ζ) then is one of the three solutions of (2.4). The selection is determined by the sheet on which the point ζ lies.

204 Herbert Stahl The three sheets of R will be denoted by B, B 0, B. For a point ζ R we write ζ (j) if ζ lies on the sheet B j. The canonical projection π : R C is bijective on each sheet. By π j : C B j, j =, 0,, we denote the three branches of the inverse function π. Points on R will generally be denoted by ζ, and by ζ (j) if it is clear on which sheet B j the point is lying. Base points will generally be denoted by w. For brevity we call the four points ζ j R, j =,..., 4, and also their four base points (2.2) branch points. The information given so far leaves the three sheets rather arbitrary, but their definition will become more concrete as the analysis advances. In a first step we assume that the two sheets B and B 0 are pasted together cross-wise in the usual way along a closed curve C on R, which is lying over an arc Γ C that connects the two branch points w and w 4. It is assumed that Γ intersects R between 0 and. Analogously, the two sheets B and B 0 are pasted together cross-wise along a curve C lying over an arc Γ C. The arc Γ connects the two branch points w 2 with w 3 and intersects R between and 0. Except for the assumptions about the intersections with R and the specific connections of the branch points, the arcs Γ and Γ are still fully arbitrary. (They will be determined in Lemma 2.4, below.) It is not difficult to deduce from (2.) that the numbering of the sheets can be done in such a way that (2.5) ψ(0 (0) ) =. It then follows from (2.) that ψ(0 (j) ) = j /3 for j =,, and ψ( (j) ) = j for j =, 0,. 2.3. Definition of the Functions h j. In the next step we shall show that there exists a function u such that the three branches of the function h = Re(u ψ) have developments near infinity that model the asymptotic behavior of the three terms n log P n(w) e 3nw, n log Q n(w), n log R n(w) e 3nw as n with P n, Q n, R n defined by relation (2.2). In the next lemma it is shown that the function h and thereby also the function u is uniquely determined by properties that follow immediately from (2.4) and (2.7). LEMMA 2.2. Let h be a function harmonic in R\{ ( ), (0), (), 0 (0) } and assume that (2.6) (2.7) (2.8) (2.9) h(ζ) = 3 Re π(ζ) + log π(ζ) + O( π(ζ) ) as ζ ( ), h(ζ) = log π(ζ) + O() as ζ (0), h(ζ) = 3 Re(π(ζ)) + log π(ζ) + O() as ζ (), h(ζ) = 3 log π(ζ) + O() as ζ 0 (0). Then the function h is uniquely determined, and it is given by h = Re(u ψ) with (2.20) u(v) := 2 v2 v 2 + log 2 3 v (v 2 ). DEFINITION 2.3. By (2.2) (2.22) h j (w) = h πj (w) for j =, 0,, and w near, h (w) = h π0 (w) for w near 0

Hermite-Padé Polynomials 205 we define four harmonic function elements in neighborhoods of and 0. In (2.2) and (2.22) the π j, j =, 0,, are the three branches of the inverse π of the canonical projection π : R C associated with the sheets B, B 0, and B. Remarks. () Up to now the sheets B, B 0, and B are still rather arbitrary; however, the assumptions made about the arcs Γ and Γ guarantee that the functions πj, j =, 0,, are well defined in neighborhoods of and 0, and therefore the function elements (2.2) and (2.22) are well defined. Their global definition, however, depends on the specification of the sheets B 0, j =, 0,, which will be done in Lemma 2.4 below. (2) The normalization (2.7) implies that n log P n(w) = log w +O(/w) as w, and consequently we have n log P n(w)e 3nw = 3 Re(w)+log w +O(/w) as w, which corresponds to (2.6). In the same way we see that the two other terms in the middle part of relation (2.4) correspond to the expressions given in (2.7) and (2.8). In these last two relations we have an error term O() instead of O(/w) since the normalization (2.7) is assumed only for the polynomial P n. Since the term E n in (2.4) has a zero of order at least 3n + 2 at 0, it follows that n log E n(w) = 3 log w + O() as w, and consequently also (2.9) follows directly from (2.5). (3) Formula (2.20) allows one to derive as many terms in the developments of the function elements h, h 0, h, and h as one wants. (4) The function h is defined via the function u, which has been defined in (2.20) as a function in the v plane. It is interesting and also important for the investigations below that the derivatives u and w = (π ψ ) have zeros at the same places in the v plane. Indeed, we have (2.23) u (v) = 3 (v4 + /3) v (v 2 ) 2, and a comparison with (2.3) shows that both functions have the same set of zeros. No proofs will be given in the present section; all results that are not stated as conjectures will be proved in Section 3. This is also the case for Lemma 2.2. In the next lemma arcs will be fixed with the help of assumptions about the global structure of the harmonic continuations of the four function elements h j, j =, 0,,. Among these arcs are the two arcs Γ and Γ, which determine the sheets B j, j =, 0,, of R. LEMMA 2.4. (i) There exist uniquely two analytic Jordan arcs Γ and Γ such that the two function elements h and h defined in (2.2) have harmonic continuations throughout the domains C \ Γ and C \ Γ, respectively, and the extended functions are continuous throughout C. The arc Γ connects the two branch points w 2 and w 3 in {Re(w) < 0}, and the arc Γ connects the two branch points w and w 4 in {Re(w) > 0}. The arc Γ is the image of Γ under reflection on the imaginary axis. (ii) There uniquely exists a continuum K 0 C such that the function element h 0 defined in (2.2) has an harmonic continuation throughout the domain C \ K 0, the extended function is continuous throughout C, and the continuum K 0 has no subarcs in common with Γ or Γ. The continuum K 0 is the union of five analytic Jordan arcs Γ 00,..., Γ 04, and it connects all four points w,..., w 4. The subarc Γ 00 is the interval [ i y, i y ] with y a positive number. The subarcs Γ 0 and Γ 02 connect the two branch points w and w 2 with i y, the subarcs Γ 03 and Γ 04 connect the branch points w 3 and w 4 with i y. For y we have the numerical value (2.24) y. = 0.6239

206 Herbert Stahl FIG. 2.. The arcs Γ, Γ, the set K 0 = Γ 00... Γ 04, and parts of the set K = Γ... Γ 4. (iii) There uniquely exists a continuum K C such that the function element h defined in (2.22) has a harmonic continuation throughout the domain C \ ({0} K ), the extended function is continuous throughout C \ {0}, and the continuum K has no subarcs in common with Γ, Γ, or K 0. The continuum K is the union of four analytic Jordan arcs Γ,..., Γ 4 C, each Γ j, j =,..., 4, connects the branch point w j with. The Γ,..., Γ 4 are disjoint in C. DEFINITION 2.5. By h j, j =, 0,,, we denote the harmonic continuations of the function elements h j into the domains C \ Γ, C \ K 0, C \ Γ, and C \ K, respectively. Because of the continuity assumption, these continuations extend to the whole C. Remarks. () With the determination of the two Jordan arcs Γ and Γ in part (i) of the Lemma 2.4 the shape of the three sheets B, B 0, B of the surface R is finally fixed. The definition of the sheets is unique up to the attribution of the boundary curves C and C to each of the neighboring sheets. This can always be done in a satisfactory way. (2) The arcs Γ, Γ, the set K 0, and parts of the set K are shown in Figure 2.. (3) Below, in Theorem 2.6, tools will be introduced which allow to calculate all arcs mentioned in Lemma 2.4 in a very efficient way. The existence of these tools allows us to say that the arcs Γ, Γ, and the sets K 0, K are defined in a constructive fashion. (4) The surface R has three sheets, but in Lemma 2.4 we have considered four branches h j, j =, 0,,, of the function h. Consequently, everywhere in C \ (Γ K 0 Γ K ) two of the four branches have to be identical. The harmonicity implies that the identical pair has to be the same in each component of the set C \ (Γ K 0 Γ K ). The pairing can only change if w crosses one of the subarcs of Γ K 0 Γ K. 2.4. Definition of the Measures ν j. For the functions h j, j =, 0,, of Definition 2.5 there exist representations which involve logarithmic potentials, and these potentials de-

Hermite-Padé Polynomials 207 termine three probability measures that are the asymptotic distributions of the zeros of the polynomials P n, Q n, R n. LEMMA 2.6. There exist three probability measures ν, ν 0, ν such that (2.25) h (w) = 3 Re(w) log w x dν (x), (2.26) h 0 (w) = log(2) log w x dν 0(x), (2.27) h (w) = 3 Re(w) log w x dν (x). We have supp(ν j ) = Γ j for j =,, and supp(ν 0 ) = K 0. The measure ν is the image of ν under reflection on the imaginary axis. Remarks. () The measures ν, ν 0, ν are absolutely continuous with respect to linear Lebesgue measure on supp(ν j ) for j =, 0,, respectively. Below, in Theorem 2.7, tools will be presented that allow an efficient calculation of the density functions of these three measures. (2) The density functions are real-analytic with respect to arc length inside of the arcs that form the supports. Near the branch points w,..., w 4 the density functions are of the form const dist(w, w j ) + O( w w j ) for w supp(ν j ) and w w j, j =,..., 4. 2.5. Asymptotics I. The definitions of the last two subsections allow to formulate the first group of asymptotic results, which are a core piece of the present talk. These statements are still conjectures since parts of the proofs still have not been worked out. CONJECTURE 2.7. Let the functions h j, j =, 0,,, the arcs Γ, Γ, and the continua K 0, K, be defined as in the Lemma 2.4 and Definition 2.5. Then we have (2.28) (2.29) (2.30) (2.3) lim n n log P n(w) = h (w) + 3 Re(w) locally uniformly for w C \ Γ, lim n lim n lim n n log Q n(w) = h 0 (w) locally uniformly for w C \ K 0, n log R n(w) = h (w) 3 Re(w) locally uniformly for w C \ Γ, n log E n(w) = h (w) locally uniformly for w C \ ({0} K ). CONJECTURE 2.8. The three measures ω P, ω Q, ω R in Conjecture 2. are given by (2.32) ω P = ν, ω Q = ν 0, ω R = ν, where ν j, j =, 0,, are the probability measures introduced in Lemma 2.6. In Figure 2.2 the zeros from Figure. are plotted together with the arcs introduced in Lemma 2.4. The zeros in Figure. have been calculated in the z variable. In order to make them comparable to the scales of Figure 2., Figure. has been transformed by the function (2.) with n = 30. The arcs Γ, Γ, and the set K 0 are the supports of the measures ν j, j =, 0,. As Figure 2.2 shows there exists a good accordance between the zeros and the supports of their asymptotic distributions already for n = 30. In Conjecture 2.8 only weak convergence has been considered; in a forthcoming paper also a strong version of the

208 Herbert Stahl FIG. 2.2. An overlay of Figure. with Figure 2. after a shrinking of the scales of Figure. in accordance with (2.). asymptotic zero distributions will be proved. These strong asymptotic relations are precise enough to fix approximate positions of individual zeros. Also for the zeros of the error term E n an asymptotic distribution can be found with a method that is similar to that used to define the three probability measures ν, ν 0, ν. However, this asymptotic distribution has no compact support and its mass is infinite. We shall not address the problem in the present talk. 2.6. Definition of the Functions g j and the Measures ψ j. Our next aim is to present asymptotic relations for the Hermite-Padé polynomials P 2n, Q 2n, R 2n of type II. The approach will be analogous to that applied in the last two subsections for the asymptotic relation of the polynomials P n, Q n, R n of type I. In a first step we define functions g j and measures ψ j, j =, 0,, which will be the building blocks of the asymptotic relations. LEMMA 2.9. There exist uniquely two analytic Jordan arcs Γ,2, Γ 2, and a set K such that (i) the three function elements h, h, and h 0 defined in (2.2) have harmonic continuations g, g, and g 0 throughout the domains C \ (Γ,2 {0}), C \ (Γ 2 {0}), and C \ (K {0}), respectively, (ii) at the origin we have (2.33) g j (w) = 3 log w + O() as w 0 for j =, 0,, and (iii) the functions g, g 0, g extend continuously throughout C \ {0}. The function elements h, h, and h 0 represent the three branches at infinity of the function h defined in Definition 2.3. Comparing Lemma 2.4 with Lemma 2.9 we see that near infinity three functions h, h 0, h and the newly introduced functions g, g 0, g are identical. However, globally these functions are different. The boundaries Γ,2, K, Γ 2 of the domains of definition of the functions g, g 0, g are defined by a similar principle as that

Hermite-Padé Polynomials 209 applied in Lemma 2.4 for Γ, Γ and K 0. Therefore, it is not surprising that there exists a connection between the new arcs Γ,2, Γ 2 and set K on one hand and the arcs Γ, Γ and set K 0 defined in Lemma 2.4 on the other hand. The connections are stated in the next lemma. LEMMA 2.0. (i) We have K = Γ Γ. (ii) The arc Γ,2 connects the branch point w 2 with w 3, and the arc Γ 2 the branch point w with w 4. The subarcs Γ 02 and Γ 03 of K 0 introduced in Lemma 2.4 are subarcs of the arc Γ,2, and the subarcs Γ 0 and Γ 04 of K 0 are subarcs of the arc Γ 2. (iii) The arc Γ,2 intersects R in the interval (0, ), and the arc Γ 2 intersects R in the interval (, 0). Remark. The arcs Γ,2, Γ 2 and the set K are plotted in Figure 2.3. For the functions g j, j =, 0,, there exist representations involving logarithmic potentials. With the help of these potentials we introduce the probability measures ψ j, j =, 0,. LEMMA 2.. There exist three probability measures ψ, ψ 0, ψ such that (2.34) g (w) = 3 Re(w) + 3 log w + 2 log w x dψ (x), (2.35) g 0 (w) = log(2) + 3 log w + 2 log w x dψ 0(x), (2.36) g (w) = 3 Re(w) + 3 log w + 2 log w x dψ (x). We have supp(ψ j ) = Γ j,2 for j =,, and supp(ψ 0 ) = K. The measure ψ is the image of ψ under reflection on the imaginary axis. The measure ψ 0 is symmetric with respect to the imaginary axis, and we have ψ 0 = (ν + ν )/2. Remark. Like the measures ν, ν 0, ν, the measures ψ, ψ 0, ψ are also absolutely continuous with respect to linear Lebesgue measure on the supports supp(ψ j ), j =, 0,. Below, in Theorem 2.7, tools will be presented that allow an efficient calculation of the density functions of the three measures. 2.7. Asymptotics II. With the definitions of the last subsection we are prepared to formulate the asymptotic relations for the Hermite-Padé polynomials P 2n, Q 2n, R 2n of type II. CONJECTURE 2.2. Let the functions g j, j =, 0,, the arcs Γ,2, Γ 2, and the set K, be defined as in the Lemma 2.9. Then locally uniformly we have (2.37) lim n (2.38) lim n (2.39) lim n n log P 2n(w) = g (w) + 3 log w + 3 Re(w) for w C \ (Γ,2 {0}), n log Q 2n(w) = g 0 (w) + 3 log w for w C \ (K {0}), n log R 2n(w) = g (w) + 3 log w 3 Re(w) for w C \ (Γ 2 {0}).

20 Herbert Stahl FIG. 2.3. The arcs Γ,2, Γ 2 and the set K = Γ Γ. The arcs of Figure 2. are included into the figure and represented by dashed lines ( ). CONJECTURE 2.3. The three measures ω P, ω Q, ω R in Conjecture 2. are given by (2.40) ω P = ψ, ω Q = ψ 0, ω R = ψ, where ψ j, j =, 0,, are the probability measures introduced in Lemma 2.. In Figure 2.4 the zeros of the Hermite-Padé polynomials P 60, Q 60, R 60 of type II are plotted together with the arcs introduced in Lemma 2.9 and plotted in Figure 2.3. Note that Figure. has been rescaled by function (2.) with n = 30. As in Figure 2.2, we observe a good approximation of the support supp(ψ j ), j =, 0,, by the zeros of the Hermite-Padé polynomials P 60, Q 60, R 60, respectively. 2.8. A Numerical Method. In the last subsection we presented a numerical method which allows to calculate the functions h j, j =, 0,,, and g j, j =, 0,, in the asymptotic relations of the Conjectures 2.7 and 2.2, the arcs Γ,..., Γ 2 introduced in the Lemmas 2.4 and 2.9, and the density functions of the measures ν, ν 0, ν, ψ, ψ 0, ψ introduced in the Lemmas 2.6 and 2.. The function h defined on the Riemann surface R and introduced in Lemma 2.2 is basic for the definition of the functions h j and g j. Let f denote the function (2.4) f(v) := Re 2 v2 v 2 + log 2 3 log v(v 2 ), which is the real part of function (2.20) in Lemma 2.2. From Lemma 2.2 and the definition of the functions h j and g j in the Lemmas 2.4 and 2.9 it follows that for any given w C the values h j (w) and g j (w) are equal to f(v j ) with v j = ψ π l j (w) if the branch π l j of the inverse projection π is chosen appropriately. The calculation of ψ π l (w) can be done

Hermite-Padé Polynomials 2 FIG. 2.4. An overlay of Figure.2 with Figure 2.3 after a shrinking of the scales of Figure.2 in accordance with (2.). very efficiently, as we have seen in Subsection 2.2. The value of ψ π l (w) is one of the three solutions v l, l =, 0,, of the equation (2.42) w v (v 2 ) v 2 + 3 = 0. However, the selection of the right solution is a problem that is equivalent to choosing the appropriate branch of π in ψ π l j (w). In the next theorem we present a strategy for making these selections by continuation along chains of points in the domains of definition of the functions h j and g j. Let D j, j =, 0,,, denote the domain of definition of the function h j, i.e., by Lemma 2.4 we have D j := C \ Γ j for j =,, and D j := C \ K j for j = 0,. Let G j, j =, 0,, further denote the domain of definition of the function g j, i.e., by Lemma 2.9 we have G j := C\(Γ j2 {0}) for j =,, and G 0 := C\(K {0}). THEOREM 2.4. (i) Let w C lie close to. Then we can choose the index l of the three solutions v l of equation (2.42) in such a way that v l lies close to l for each l =, 0,, and it follows that (2.43) h j (w) = g j (w) = f(v j ) for j =, 0,. (ii) Let w C lie close to 0. Then there exists exactly one solution v of equation (2.42) that lies close to (the other two solutions lie close to ± /3). With this choice of v it follows that (2.44) h (w) = f(v ). (iii) For an arbitrary w D j, j =, 0,, or w G j, j =, 0,, the values h j (w) or g j (w), respectively, can be calculated by continuation of these functions through D j or G j starting from a neighborhood of infinity. If we assume that for a point w = w 0 D j

22 Herbert Stahl the solution v j = v j (w 0 ) of equation (2.42) has been chosen in such a way that h j (w 0 ) = f(v j (w 0 )), and if w D j lies close to w 0, then one has to choose, from the three roots of equation (2.42) with w = w, the solution v j = v j (w ) that lies closest to v j = v j (w 0 ). For this new v j one has (2.45) h j (w ) = f(v j (w )). In an analogous way one can proceed for the functions g j and w G j, j =, 0,. (iv) The function h is calculated by continuation in D using the same strategy as that described in (iii), only that now one has to start from a neighborhood of w = 0. Next we address the problem of calculating the arcs introduced in Lemmas 2.4 and 2.9. DEFINITION 2.5. In order to simplify notation, we define Γ := Γ, Γ, := Γ, Γ 3 := Γ Γ 4, Γ,3 := Γ 2 Γ 3. (The arcs Γ,2 and Γ 2 have already been defined in Lemma 2.9.) It is immediate that all arcs introduced in the Lemmas 2.4 and 2.9 are subarcs of the system Γ ji, j =,, i =, 2, 3, together with Γ 00 = [ iy, iy ]. A numerical value for y has been given in (2.24). THEOREM 2.6. (i) The three arcs Γ, Γ 2, Γ 3 are analytic, they are disjoint in C \ {w, w 4 }, and each one of them connects w with w 4 in C. All three arcs are symmetric with respect to R, and we have Γ 3. At w the arcs Γ, Γ 2, Γ 3 have tangential directions (2.46) ϕ = 65 π/36, ϕ 2 = 4 π/36, ϕ 3 = 7 π/36, respectively. The tangential directions at w 4 follow from (2.46) and the symmetry with respect to R. (ii) For w = w and w = w 4 equation (2.42) has two identical solutions. For each w Γ j, j =, 2, 3, a pair can be selected from the three solutions of equation (2.42) in such a way that the two elements of this pair are a continuation along Γ j of the two elements of the identical pair at w or w 4. The two solutions selected in that way are denoted by v j+ = v j+ (w) and v j = v j (w) for w Γ j. Let t w D denote the tangent on the arc Γ j at the point w Γ j. Then we have (2.47) t w = ±i v j+(w) v j (w) v j+ (w) v j (w) for w Γ j, j =, 2, 3. (iii) The three arcs Γ,j, j =, 2, 3, are the images of the arcs Γ j, j =, 2, 3, under reflection on the imaginary axis. (iv) For the subarcs of K 0 and K introduced in Lemma 2.4 we have (2.48) (2.49) (2.50) (2.5) (2.52) (2.53) Γ 0 = Γ 2 {Re(w) > 0, Im(w) > 0}, Γ 02 = Γ,2 {Re(w) < 0, Im(w) > 0}, Γ 03 = Γ,2 {Re(w) < 0, Im(w) < 0}, Γ 04 = Γ 2 {Re(w) > 0, Im(w) < 0}, Γ = Γ 3 {Im(w) > 0}, Γ 4 = Γ 3 {Im(w) < 0}, Γ 2 = Γ,3 {Im(w) > 0}, Γ 3 = Γ,3 {Im(w) < 0}.

Hermite-Padé Polynomials 23 Remark. The initial directions of the arcs Γ j, j =, 2, 3, at one of the branch points w or w 4 given in part (i) of the theorem together with the formula (2.47) for tangent, form the basis for an efficient calculation of the Jordan arcs Γ j, j =, 2, 3. By symmetry this also allows to calculate the Jordan arcs Γ,j, j =, 2, 3. The information given in part (iv) of the theorem allow to break down the arcs Γ ij into those pieces, which are needed for the construction of the sets K 0, K. The arcs shown in the Figures 2. - 2.4 are calculated in this way. The last theorem in the present section contains a numerical method for calculating the density functions of the measures ν,..., ψ. In a first step we introduce five new measures µ ij, i =,, j =, 2, µ 00 on the arcs Γ ij, Γ 00 = [ iy, iy ], respectively. THEOREM 2.7. (i) Let for w Γ j, j =, 2, the two solutions v j+ = v j+ (w) and v j = v j (w) of equation (2.42) be selected as described in part (ii) of Theorem 2.6. We define measures µ j, j =, 2, on Γ and Γ 2 by (2.54) dµ j (w) := 3 2π v j+(w) v j (w) ds w, w Γ j, j =, 2, where ds w denotes the line element on the arcs Γ j, j =, 2. (ii) Let the two measures µ,j, j =, 2, be the images of the measures µ j, j =, 2, under the reflection on the imaginary axis. (iii) For w Γ 00 = [ iy, iy ] let v 0+ = v 0+ (w) and v 0 = v 0 (w) denote the two solutions of equation (2.42) that are symmetric with respect to the imaginary axis. The measure µ 00 is defined on Γ 00 by (2.55) dµ 00 (w) := 3 2π v 0+(w) v 0 (w) ds w, w Γ 00. With the definitions of part (i), (ii), and (iii) we have (2.56) (2.57) (2.58) (2.59) ν j = µ j, j =,, ν 0 = µ 00 + µ {Re(w)<0},2 + µ {Re(w)>0} 2, ψ j = 2 µ j2, j =,, ψ 0 = 2 (µ, + µ ). 3. Proofs. The lemmas 2.2, 2.4, 2.6, 2.9, 2.0, 2., and the Theorems 2.4, 2.6, and 2.7 are proved in the present section. 3.. Proof of Lemma 2.2. Let the function h be defined as Re(u ψ) on R with u given by (2.20), then this function has the developments (2.6-2.9) at the points ( ), (0), (), and 0 (0). This can be verified by straightforward calculations using (2.) for the definition of π ψ. In order to prove uniqueness, we assume that g is another function that is harmonic in R \ { ( ), (0), (), 0 (0) } and satisfies the assumptions made in (2.6-2.9) for the function h. Then h g is harmonic throughout the compact surface R, and consequently h g is constant. From (2.6) we know that (h g)( ( ) ) = 0, which proves uniqueness.

24 Herbert Stahl 3.2. Auxiliary Lemmas. Next we state and prove two auxiliary lemmas. LEMMA 3.. We have (3.) (u ψ π j ) (w) = 3 (ψ π j )(w) = 3 ψ(ζ (j) ) for w C, j =, 0,, with u defined in (2.20), w = π(ζ (j) ), ζ (j) R. Proof. From the chain rule it follows that (3.2) (u ψ π j ) (w) = u ((ψ π j )(w))(ψ π j ) (w) = u (v) (π ψ ) (v) with v = ψ(ζ (j) ). Using the expressions (2.3) and (2.23), this yields (u ψ π j ) (w) = 3 v, which proves (3.). LEMMA 3.2. Define the set N as (3.3) N := { w C h π (w) = h π 0 (w) }, where the branches πj, j =, 0,, of π are determined by the choice of the sheets B, B 0, B of R. (i) The set N is independent of the choice of the sheets B, B 0, B if the assumptions made in Subsection 2.2 are satisfied and if in addition we have N Γ =. (ii) If the assumptions formulated in part (i) are satisfied, then the set N is the union of three analytic Jordan arcs Γ, Γ 2, Γ 3. Each of the three arcs Γ j, j =, 2, 3, connects the two branch points w and w 4 in C, and at w the arcs Γ, Γ 2, Γ 3 have the tangential directions (3.4) ϕ = 65 36 π, ϕ 2 = 4 36 π, ϕ 3 = 7 36 π, respectively. (iii) We have Γ 3, and (3.5) w = ( ) 3 log 2 + i Im(w) + O w as w, w Γ 3. (iv) For the intersection points of N with R and ir we have the following numerical values (3.6) Γ R = {0.59999}, Γ 2 R = { 0.3793}, Γ 2 i R = { i 0.6239, i 0.6239}. Remark. The arcs Γ, Γ 2, and parts of the arc Γ 3 are plotted in Figure 3.. 3.3. Proof of Lemma 3.2. The proof of the lemma will be rather involved. However, a great part of the investigations will also be used in the subsequent proofs of lemmas and theorems from Section 2. Let Γ, Γ be Jordan arcs that satisfy the assumptions made in Subsection 2.2. In addition to that we assume that Γ is the reflection of Γ on the imaginary axis. The two arcs Γ, Γ determine the three sheets B, B 0, B of the Riemann surface R, and consequently also the functions h π and h π0 used in definition (3.3) of N are determined by the

Hermite-Padé Polynomials 25 FIG. 3.. The arcs Γ,Γ 2, and parts of the arc Γ 3. The three arcs form the set N defined in (3.3). The three domains D () 0, D(), D() 2 are the components of the set C\N. specific choice of the arcs Γ, Γ. Since we have on Γ a change over between the two functions h π and h π0, the function (3.7) d(w) := h π (w) h π 0 (w) changes sign when w crosses Γ. From this we see that the definition of the set N itself is independent of variations of the arc Γ if we have Γ N =. From Lemma 3. and the definitions of h = Re(u ψ) and h j = h πj, j =, 0,, in Lemma 2.2 and Definition 2.3 we deduce that (3.8) (3.9) x h j(w) = Re [ (u ψ π j ) (w) ] = 3 Re(v j ), y h j(w) = Im [ (u ψ π j ) (w) ] = 3 Im(v j ) with v j := ψ π j (w), w = x + iy C, j =, 0,. From the harmonicity of the function d it follows that the set N consists of analytic arcs. From (3.3), (3.8-3.9), and (3.7) it further follows that (3.0) ( ) 2 ( ) 2 x d(w) + y d(w) = 9 v v 0 2. It is a consequence of (3.0) that the arcs of N can have no bifurcations in C \ {w,..., w 4 } since otherwise we would have (3.) x d(w) = d(w) = 0, y w = x + iy,

26 Herbert Stahl at such a point, which, however, is only possible at the branch points w and w 4 because of (3.0). It is immediate that w, w 4 N. In order to understand the structure of the set N in a neighborhood of the branch point w, we consider the development of the function h in the local coordinate ζ = w + η 2. From Lemma 3. and the definitions in Subsection 2.2 we derive the development (3.2) u ψ(ζ) = u(v ) + 3 v η 2 + 2 2 (π ψ ) (v ) η3 +... with v j := ψ(ζ ) = 4 /3. Evaluating (π ψ ) (v ) = w (v ) then yields (3.3) h(ζ) = h(ζ ) + 3 Re ( v η 2) + Re ( ) /4 4 3 2 η3 +... i 33/2 Taking into account the form of the two branches π and π0 in a neighborhood of the branch point w, we deduce from (3.7) and (3.3) that (3.4) d(w + η 2 ) = 2 η 3 Re ( ) /4 4 3 arg(η)/2 3 2 ei + O( η 4 ) as η 0. i 33/2 The condition d(w + η 2 ) = 0 then implies that for η 0 we have (3.5) 5 72 π + 3 2 arg(η2 ) 2 π + mod(π), and this implies that the set N consists of three arcs in a neighborhood of w, which have the tangential directions given in (3.4) at w. Using the mapping function (2.), one can compare the behavior of the values v := ψ π (w) and v 0 := ψ π0 (w) while w runs through R +. It then is possible to verify that (3.6) Re(v (w) v 0 (w)) { > 0 for w > ( Γ R) < 0 for w < ( Γ R). In a similar, but somewhat more involved way, one can show that (3.7) Im(v (w) v 0 (w)) > 0 for all w i R +. From (3.7), (3.8-3.9), (3.6), and (3.7) we deduce that { (3.8) x d(w) < 0 for w > ( Γ R) > 0 for w < ( Γ R) and (3.9) y d(w) > 0 for all w i R +, i.e., the function d is strictly monotonic on R +, i R +, and i R. On R + at the intersection point of Γ and R we have a change over between the two branches h and h 0, and therefore