Logarithmic residue based methods for

computing zeros of analytic functions

and related problems



P Kravanja M Van Barel and A Haegemans

Department of Computer Science Katholieke Universiteit Leuven

Celestijnenlaan A B Heverlee Belgium

Abstract

Given an f and a p ositively oriented Jordan curve we consider the

problem of computing all the zeros of f that lie inside together with their resp ective mul

tiplicities By exploiting the connection b etween logarithmic residue integrals and the theory

of formal orthogonal p olynomials we obtain an accurate algorithm that pro ceeds by solving

generalized eigenvalue problems and a Vandermonde system We show how similar techniques

can b e applied to the following problems computing zeros and p oles of meromorphic functions

and lo cating clusters of zeros of analytic functions

Intro duction

Let W b e a simply connected region in C f W C analytic in W and a p ositively oriented

Jordan curve in W that do es not pass through any zero of f We consider the problem of computing

al l the zeros of f that lie in the interior of together with their resp ective multiplicities

Let N denote the total numb er of zeros of f that lie in the interior of ie the numb er of zeros

where each zero is counted according to its multiplicity Supp ose that N Let the sequence

Z Z consist of all the zeros of f that lie inside Each zero is rep eated according to its

N

0

f has a simple p ole at multiplicity An easy calculation shows that the f

each zero of f with residue equal to the multiplicity of the zero It follows that

Z

0

f z

dz N

i f z

and thus N can b e calculated via numerical integration Integrals of this typ e are called logarithmic

residue integrals Metho ds for the determination of zeros of analytic functions that are based on the

numerical evaluation of integrals are called quadrature methods A review of such metho ds is given

in A classical approach is to consider the integrals

Z

0

f z

p

s dz p z

p

i f z



Corresp onding author Email PeterKravanjananetornlgov

The implies that the s s are equal to the Newton sums of the unknown zeros

p

p p

p Z s Z

p

N

Delves and Lyness considered the monic p olynomial P z of degree N that has zeros Z Z

N N

and calculated its co ecients via Newtons identities In this way they reduced the problem to the

easier problem of computing the zeros of a p olynomial Unfortunately the map from the Newton

sums to the co ecients in the standard monomial basis of P z is usually illconditioned Also

N

the p olynomials that arise in practice may b e such that small changes in the co ecients pro duce

much larger changes in some of the zeros The lo cation of the zeros determines their sensitivity to

p erturbations of the co ecients Multiple zeros and very close zeros are extremely sensitive but

even a succession of mo derately close zeros can result in severe illconditioning The comp osite map

from the Newton sums to the zeros and their resp ective multiplicities is also usually illconditioned

A similar approach was taken by Li who considered as a system of p olynomial equations

Li used a homotopy continuation metho d to solve this system

What is wrong with these approaches in our opinion is that they consider the wrong set of

unknowns One should consider the mutually distinct zeros and their resp ective multiplicities sep

arately Let n denote the numb er of mutually distinct zeros of f that lie in the interior of Let

z z b e these zeros and their resp ective multiplicities In Section we will show how

n n

these unknowns can b e calculated by solving generalized eigenvalue problems and a Vandermonde

system Our approach exploits the connection b etween logarithmic residue integrals and the theory

of formal orthogonal p olynomials In Section we show how similar techniques can b e applied to

the following problems computing zeros and p oles of meromorphic functions and lo cating clusters

of zeros of analytic functions

Formal orthogonal p olynomials

Let P b e the linear space of p olynomials with complex co ecients We dene a symmetric bilinear

form h i P P C by setting

Z

n

0

X

f z

z z dz z z h i

k k k

i f z

k

for any two p olynomials P Note that h i can b e evaluated via numerical integration

A monic p olynomial of degree t that satises

t

k

hz z i k t

t

is called a formal orthogonal FOP of degree t Observe that condition is void for

t The adjective formal emphasizes the fact that in general the form h i do es not dene a

true inner pro duct An imp ortant consequence of this fact is that in contrast to p olynomials that

are orthogonal with resp ect to a true inner pro duct FOPs need not exist or need not b e unique

t

for every degree t For details see for example and the references cited therein If is satised

and is unique then is called a regular FOP and t a regular index If we set

t t

t t

z u u z u z z

t t t tt

then condition translates into the YuleWalker system

s s s

s u

t

t t

s u

s

t t

s u

t tt s s

t t

t

Hence the regular FOP of degree t exists if and only if the Hankel matrix H s is

t k l

k l

regular Thus the rank prole of H s determines which regular FOPs exist

k l k l

Theorem n rank H for every integer p In particular n rank H

np N

Proof Let p b e a nonnegative integer As

n

X

p

p

s h z i z

p k

k

k

for p the matrix H can b e written as

np

n

h i

X

T

np

T

H z z where z

z z

np k k k

k

k

k

k

This implies that rank H n However H is regular Indeed one can easily verify that H can

np n n

n r T

and D is the where V is the Vandermonde matrix V z b e factorized as H V D V

n n n n n n

rs s n

diagonal matrix D diag Therefore rank H n It follows that rank H n 2

n n np np

By assumption s N and thus the regular FOP of degree exists It is given by z z

where s s Theorem implies that the regular FOP of degree n exists and tells us also

n

that regular FOPs of degree larger than n do not exist The p olynomial is easily seen to b e

n

n

Y

z z z

k n

k

It is the monic p olynomial of degree n that has z z as simple zeros

n

If H is strongly regular ie if all its leading principal submatrices are regular then we have a

n

full set f g of regular FOPs

n

What happ ens if H is not strongly regular By lling up the gaps in the sequence of existing

n

1

with a monic p olynomial of degree t regular FOPs it is p ossible to dene a sequence f g

t t

t

such that if these p olynomials are group ed into blo cks according to the sequence of regular indices

then p olynomials b elonging to dierent blo cks are orthogonal with resp ect to h i More precisely

1

dene f g as follows If t is a regular index then let b e the regular FOP of degree t Else

t t

t

dene as where r is the largest regular index less than t and is an arbitrary monic

t r tr tr

tr

p olynomial of degree t r In the latter case is called an inner polynomial If z z then

t tr

1

can b e we say that is dened by using the standard monomial basis These p olynomials f g

t t

t

group ed into blo cks Each blo ck starts with a regular FOP and the remaining p olynomials are inner

p olynomials Note that the last blo ck has innite length

The socalled block orthogonality prop erty is expressed by the fact that the Gram matrix G

n

n

h i is blo ck diagonal The diagonal blo cks are regular symmetric and zero ab ove the

r s

rs

main antidiagonal If all the inner p olynomials in a certain blo ck are dened by using the standard

monomial basis then the corresp onding diagonal blo ck has Hankel structure The matrix G

n

n

h i is blo ck tridiagonal The diagonal blo cks are symmetric and lower antiHessenb erg

r s

rs

ie its entries are equal to zero along all the antidiagonals that lie ab ove the main antidiagonal

except for the antidiagonal that precedes the main antidiagonal Again if all the inner p olynomials

in a certain blo ck are dened by using the standard monomial basis then the corresp onding diagonal

blo ck is a Hankel matrix The entries of the odiagonal blo cks are all equal to zero except for the

entry in the southeast corner For pro ofs and further details we refer to

Theorem The eigenvalues of the pencil G G are z z In other words they

n n

n

are given by z z where s s

n

nn

Proof Dene V as the Vandermondelike matrix V z and let D and D b e

n n r s n

r s

n

the diagonal matrices D diag and D diag z z Then G

n n n n n

n

T T

and G can b e factorized as G V D V and G V D V Let b e an eigenvalue of the

n n n n

n n n n n

p encil G G and x a corresp onding eigenvector Then

n

n

G x G x

n

n

T T

V D V x V D V x

n n n

n n n

T

x y D y if y V D

n

n n

diag z z y y

n

This proves the theorem 2

If n and are known then we can apply the previous theorem to obtain z z

n n

By Theorem the value of n can b e computed as the rank of H but this is not a very practical

N

1

approach Instead we will start to compute the p olynomials f g one by one and determine the

t

t

value of n as the degree of the last existing regular FOP The upp er b ound N will play a crucial role

in the computations

Theorem can b e interpreted as follows the zeros of the nth degree regular FOP can b e calculated

by solving a generalized eigenvalue problem This prop erty holds for al l regular FOPs This will

enable us to compute regular FOPs in their pro duct representation which is numerically very stable

as Dene the matrices G and G

k

k

k k

and G h i G h i

r s k r s

rs k rs

for k

Theorem Let t be a regular index and let z z be the zeros of the regular FOP

t tt t

Then the eigenvalues of the pencil G G are given by z z In other words they

t t tt

t

are given by z z where s s

t tt

t

We will rst show that the zeros of s as H Proof Dene the Hankel matrix H

t k l

k l t t

are given by the eigenvalues of the p encil H H The zeros of are given by the eigenvalues of

t t

t

its companion matrix C Let b e an eigenvalue of C and x a corresp onding eigenvector As H is

t t t

regular we may conclude that C x x H C x H x Using one can easily verify that

t t t t

H C H

t t t

Let A b e the unit upp er triangular matrix that contains the co ecients of

t t

T

Then G can b e factorized as G A H A As z z where s s the matrix G

t t t t

t

t

t t

T

is given by h z i G The matrix h z i can b e written as A H A and thus

r s t r s t

rs rs

t t

T

G A H H A Now let b e an eigenvalue of the p encil H H and x a corresp onding

t t t

t

t t t

eigenvector Then

H x H x

t

t

H H x H x

t t

t

T T

A H H A y A H A y if y A x

t t t t

t

t t t

G y G y

t

t

This proves the theorem 2

Regular FOPs are characterized by the fact that the determinant of a Hankel matrix is dierent

from zero while inner p olynomials corresp ond to singular Hankel matrices By using an explicit

determinant expression for regular FOPs one can show that h i det H det H Therefore

t t t t

if t is a regular index then t is a regular index if and only if h i However from a

t t

numerical p oint of view a test is equal to zero do es not make sense Because of rounding errors in

the evaluation of h i we would encounter only regular FOPs Strictly sp eaking one could say that

inner p olynomials are not needed in numerical calculations However the opp osite is true Let us

call a regular FOP wel lconditioned if its corresp onding YuleWalker system is wellconditioned

and il lconditioned otherwise To obtain a numerically stable algorithm it is crucial to generate only

wellconditioned regular FOPs and to replace illconditioned regular FOPs by inner p olynomials

Stable lo okahead solvers for linear systems of equations that have Hankel structure are based on

this principle In this approach the diagonal blo cks in G are taken slightly larger than

n

strictly necessary to avoid illconditioned blo cks A disadvantage is that part of the structure of G

n

and G gets lost and that there will b e some additional llin

n

We will ask the user for two thresholds and with to decide whether

stop cond stop cond

the algorithm may stop or not and to determine the size of a blo ck Supp ose that the algorithm

has just generated a wellconditioned regular FOP If r N then we may stop Else we

r

pro ceed to calculate h i If jh ij then we generate as a regular FOP ie

r r r r cond r

we solve a generalized eigenvalue problem to obtain the zeros of Else we scan the sequence

r

N r

t t

jhz ij If jhz ij for t N r then we conclude that n r

r r r r stop

t

and stop Else we search for the rst element that is larger than The corresp onding value of t

cond

then determines the size of the blo ck of p olynomials If all the elements are less than then we

cond

use the value of t that corresp onds to the maximum to determine the blo ck size and warn the user

that we could not obtain the level of wellconditioning that he or she requested

Once approximations for the zeros have b een calculated the multiplicities can b e found by solving

the Vandermonde system

s

s z z

n

n

n

s z z

n n

n

A Fortran implementation of our algorithm is available Numerical tests show that it gives very

accurate results Note that it do es not require initial guesses for the zeros Also as not only

approximations for the zeros are computed but also the corresp onding multiplicities one can use the

quadratically convergent mo died Newtons iteration to rene the approximations for the zeros

Related problems

Computing zeros and p oles of meromorphic functions

Let W and b e dened as ab ove and let f W C b e meromorphic in W Supp ose that f has

neither zeros nor p oles on We consider the problem of computing al l the zeros and p oles of f that

lie in the interior of together with their resp ective multiplicities and orders Let N denote the

total numb er of zeros of f that lie inside and let P denote the total numb er of p oles of f that lie

inside Supp ose that N P Let n denote the numb er of mutually distinct zeros of f that lie

inside Let z z b e these zeros and their resp ective multiplicities Let p denote the

n n

numb er of mutually distinct p oles of f that lie inside Let y y b e these p oles and

p p

0

their resp ective orders The logarithmic derivative f z f z has a simple p ole at z with residue

k k

for k n and a simple p ole at y with residue for l p It follows that

l l

Z

p

n

0

X X

f z

dz z z z z y y h i

k k k l l l

i f z

k l

for any two p olynomials P The FOP of degree n p with resp ect to this form is given by

p

n

Y Y

z z z y

k l

k l

Our algorithm for computing zeros of analytic functions is essentially an algorithm for computing

zeros of FOPs Provided that an upp er b ound for n p is known one can easily verify that the

algorithm can b e mo died to compute the zeros of ie the mutually distinct zeros and p oles

of f that lie inside Once these have b een calculated their corresp onding multiplicities and orders

can b e calculated by solving a Vandermonde system The sign of a comp onent of the solution of this

systems tells us whether the corresp onding zero of the FOP of degree n p is a zero of f or a p ole

As s h i N P can b e computed via numerical integration and n p N P s P

it follows that an upp er b ound for P immediately leads to the required upp er b ound for n p In

case is the unit circle such a b ound for P can b e obtained by using the heuristic approach of

Gleyse and Kaliaguine

Lo cating clusters of zeros of analytic functions

Supp ose that f is analytic and supp ose that the zeros of f that lie inside can b e group ed into m

clusters In this case the problem of computing all the zeros of f that lie inside is notoriously

illconditioned Let I I b e index sets that dene these clusters and let

m

X X

z and c

k k j k j

j

k 2I k 2I

j j

for j m In other words is equal to the total numb er of zeros that form cluster j

j

its weight whereas c is equal to the arithmetic mean of the zeros in cluster j its centre of

j

gravity We will show how our algorithm for computing zeros of analytic functions can b e used

to obtain approximations for the centres c c of the clusters together with the corresp onding

m

weights This information enables one to zo om into a certain cluster its zeros can b e

m

calculated separately from the other zeros of f By shifting the origin in the to

the centre of a certain cluster its zeros b ecome b etter relatively separated which is appropriate in

oating p oint arithmetic and reduces the illconditioning

For k n we dene z c if k I From the denition of and c it follows that

k k j j j j

X

j m

k k

k 2I

j

Dene the symmetric bilinear form h i by

m

m

X

c c h i

j j j m

j

for any two p olynomials P This form is related to the form h i in an obvious way instead

of the zeros z z and their multiplicities we now use the centres of gravity c c

n n m

and the weights of the clusters Let

m

max j j

k

k n

The following theorem tells us that h i approximates h i and vice versa

m

Theorem Let P Then h i h i O

m

Proof This follows immediately by expanding the terms in

n m

X X X

h i z z c c

k k k k j k j k

j

k k 2I

j

into a Taylor and by taking into account equation 2

In other words if is suciently small then we can exp ect our algorithm for computing zeros of

analytic functions to b ehave as if the underlying symmetric bilinear form is h i instead of h i

m

The FOP of degree m will b e a regular FOP and its zeros will b e go o d approximations for the centres

c c The FOPs of degree larger than m will b e illconditioned The following result conrms

m

this intuition

Corollary The matrix H is regular if Let t m Then det H O

m t

Let z be the FOP of degree m with respect to the form h i Then c O

m m j

p

for j m Also hz z i O for al l p m

m

Once approximations for the centres have b een found the corresp onding weights can b e calculated

by solving a Vandermonde system For more details including a pro of of the previous corollary

numerical examples as well as a dierent approach based on rational interp olation at ro ots of unity

we refer to

Acknowledgements

This research was supp orted by the Fund for Scientic ResearchFlanders FWOV pro jects Or

thogonal Systems and their Applications grant G and Counting and computing all

isolated solutions of systems of nonlinear equations grant G and by the Flemish Insti

tute for the Promotion of Scientic and Technological Research in the Industry IWT

References

A W Bo janczyk and G Heinig A multistep algorithm for Hankel matrices J Complexity

no

A Bultheel and M Van Barel Linear algebra rational approximation and orthogonal polyno

mials Studies in Computational Mathematics vol Elsevier

S Cabay and R Meleshko A weakly stable algorithm for Pade approximants and the inversion

of Hankel matrices SIAM J Matrix Anal Appl no

L M Delves and J N Lyness A numerical method for locating the zeros of an analytic function

Math Comput

R W Freund and H Zha A lookahead algorithm for the solution of general Hankel systems

Numer Math

B Gleyse and V Kaliaguine On algebraic computation of number of poles of meromorphic func

tions in the unit disk Nonlinear Numerical Metho ds and Rational Approximation I I A Cuyt

ed Kluwer Academic Publishers pp

W B Gragg and M H Gutknecht Stable lookahead versions of the Euclidean and Cheby

shev algorithms Approximation and Computation A Festschrift in Honor of Walter Gautschi

R V M Zahar ed Birkhauser pp

N I Ioakimidis Quadrature methods for the determination of zeros of transcendental functions

a review Numerical Integration Recent Developments Software and Applications P Keast and

G Fairweather eds Reidel Dordrecht The Netherlands pp

P Kravanja R Co ols and A Haegemans Computing zeros of analytic mappings A logarithmic

residue approach BIT no To app ear

P Kravanja T Sakurai and M Van Barel On locating clusters of zeros of analytic functions

Technical Rep ort KULeuven Dept Computer Science Celestijnenlaan A B Hever

lee Belgium

TY Li On locating al l zeros of an analytic function within a bounded domain by a revised

DelvesLyness method SIAM J Numer Anal no