COMPLEX POLYNOMIALS

T. SHEIL-SMALL
University of York
published by the press syndicate of the university of cambridge
The Pitt Building, Trumpington Street, Cambridge, United Kingdom
cambridge university press
The Edinburgh Building, Cambridge CB2 2RU, UK
40 West 20th Street, New York, NY 10011-4211, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
Ruiz de Alarcón 13, 28014 Madrid, Spain
Dock House, The Waterfront, Cape Town 8001, South Africa
http://www.cambridge.org

C Cambridge University Press 2002

This book is in copyright. Subject to statutory exception

and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.

First published 2002

Printed in the United Kingdom at the University Press, Cambridge

Typeface Times 10/13 pt System LATEX 2ε [TB]

A catalogue record for this book is available from the British Library

Library of Congress Cataloguing in Publication data

Sheil-Small, T. (Terence), 1937–
Complex polynomials / T. Sheil-Small.
p. cm. – (Cambridge studies in advanced mathematics ; 75)
Includes bibliographical references and index.
ISBN 0 521 40068 6
1. Polynomials. 2. Functions of complex variables. I. Title. II. Series.
QA161.P59 S45 2002
512.9 42 – dc21 2001052690

ISBN 0 521 40068 6 hardback

 Contents

Preface page xi
List of notation xix
1 The algebra of polynomials 1
1.1 Complex polynomials 1
1.2 The number of zeros of a real analytic polynomial 4
1.3 Real analytic polynomials at infinity 13
2 The degree principle and the fundamental theorem
of algebra 22
2.1 The fundamental theorem of algebra 22
2.2 Continuous functions in the plane 26
2.3 The degree principle 31
2.4 The degree principle and homotopy 40
2.5 The topological argument principle 43
2.6 The coincidence theorem 47
2.7 Locally 1–1 functions 56
2.8 The Borsuk–Ulam theorem 79
3 The Jacobian problem 81
3.1 The Jacobian conjecture 81
3.2 Pinchuk’s example 90
3.3 Polynomials with a constant Jacobian 105
3.4 A topological approach 118
3.5 The resultant and the Jacobian 124
4 Analytic and harmonic functions in the unit disc 125
4.1 Series representations 125
4.2 Positive and bounded operators 138
4.3 Positive trigonometric polynomials 144
4.4 Some inequalities for analytic and trigonometric
polynomials 151

vii
viii Contents

4.5 Cesàro means 156

4.6 De la Vallée Poussin means 160
4.7 Integral representations 167
4.8 Generalised convolution operators 168
5 Circular regions and Grace’s theorem 172
5.1 Convolutions and duality 172
5.2 Circular regions 178
5.3 The polar derivative 184
5.4 Locating critical points 186
5.5 Critical points of rational functions 190
5.6 The Borwein–Erdélyi inequality 193
5.7 Univalence properties of polynomials 196
5.8 Linear operators 203
6 The Ilieff–Sendov conjecture 206
6.1 Introduction 206
6.2 Proof of the conjecture for those zeros on the unit circle 207
6.3 The direct application of Grace’s theorem 208
6.4 A global upper bound 213
6.5 Inequalities relating the nearest critical point to the
nearest second zero 216
6.6 The extremal distance 221
6.7 Further remarks on the conjecture 223
7 Self-inversive polynomials 228
7.1 Introduction 228
7.2 Polynomials with interspersed zeros on the unit circle 232
7.3 Relations with the maximum modulus 238
7.4 Univalent polynomials 241
7.5 A second necessary and sufficient condition for angular
separation of zeros 249
7.6 Suffridge’s extremal polynomials 251
8 Duality and an extension of Grace’s theorem to rational
functions 263
8.1 Linear operators and rational functions 263
8.2 Interpretations of the convolution conditions 270
8.3 The duality theorem for T (1, β) 275
8.4 The duality theorem for T (m, β) 282
8.5 The duality principle 286
8.6 Duality and the class T (α, β) 289
8.7 Properties of the Kaplan classes 293
8.8 The class S(α, β) 296
 Contents ix

8.9 The classes T0 (α, β) 300

8.10 The class T (2, 2) 302
9 Real polynomials 304
9.1 Real polynomials 304
9.2 Descartes’ rule of signs 317
9.3 Strongly real rational functions 319
9.4 Critical points of real rational functions 323
9.5 Rational functions with real critical points 325
9.6 Real entire and meromorphic functions 326
10 Level curves 350
10.1 Level regions for polynomials 350
10.2 Level regions of rational functions 353
10.3 Partial fraction decomposition 355
10.4 Smale’s conjecture 358
11 Miscellaneous topics 370
11.1 The abc theorem 370
11.2 Cohn’s reduction method 372
11.3 Blaschke products 373
11.4 Blaschke products and harmonic mappings 377
11.5 Blaschke products and convex curves 382
11.6 Blaschke products and convex polygons 392
11.7 The mapping problem for Jordan polygons 402
11.8 Sudbery’s theorem on zeros of successive derivatives 407
11.9 Extensions of Sudbery’s theorem 413
References 416
Index 421
 1
The algebra of polynomials

1.1 Complex polynomials

1.1.1 Definitions
A complex polynomial is a function of the form


n
P(z) = ak z k , (1.1)
k=0

where the ak are complex numbers not all zero and where z is a complex
variable. We also use the terms analytic polynomial (reflecting the fact that
the polynomial is an analytic function) and algebraic polynomial (since the
polynomial contains only algebraic operations on the variable z). If an = 0 the
polynomial is said to have degree n. In particular, a polynomial of degree 0
is, by definition, a non-zero constant. The function which is identically zero is
often regarded as being a polynomial of degree −∞. When the ak are all real
numbers, the polynomial P(z) is called a real polynomial. Observe that P(z)
is a real polynomial iff

P(z) = P(z) (1.2)

for all z ∈ C. From this it follows that, if P(a) = 0 then P(a ) = 0; therefore
either a is real or P has the conjugate pair of zeros a and a.

1.1.2 Number of zeros

Lemma A complex polynomial of degree n has at most n zeros.

The proof of this is an entirely elementary fact from algebra; this is to be

contrasted with the stronger fundamental theorem of algebra (see chapter 2)
which states that a polynomial of degree n ≥ 0 has exactly n zeros; the usual

1
2 1 The algebra of polynomials
proofs of this use methods of analysis or topology (it is not a result which
follows purely from the algebraic field property of the complex numbers). The
proof of the weaker statement is by induction. The result is trivial when n = 0.
Assume the statement proved for polynomials of degree n −1, where n ≥ 1, and
let P be a polynomial of degree n given say by (1.1). Either P has no zeros and
there is nothing further to prove, or ∃a ∈ C such that P(a) = 0. We then have

P(z) P(z) − P(a)
n

zk − ak
n

k−1
= = ak = ak z k−1− j a j , (1.3)
z−a z−a k=0
z − a k=1 j=0

and the last expression is clearly a polynomial of degree n − 1. Therefore by

the induction hypothesis P(z)/(z − a) has at most n − 1 zeros and so P(z) has
at most n zeros. The result follows by induction.

This result has a number of important consequences.

1.1.3
Uniqueness theorem If P(z) and Q(z) are polynomials of degree not exceeding
n and if the equation
P(z) = Q(z) (1.4)

is satisfied at n + 1 distinct points, then P = Q.

For otherwise, P − Q is a polynomial of degree not exceeding n with n + 1

zeros. We deduce the next result.

1.1.4
Theorem: Lagrange’s interpolation formula Let z 1 , z 2 , . . . , z n + 1 be n + 1
distinct points and let w1 , w2 , . . . , wn+1 be arbitrary complex numbers (not
necessarily distinct but not all zero). Among all polynomials of degree not
exceeding n there is a unique polynomial P(z) such that

P(z k ) = wk (1 ≤ k ≤ n + 1). (1.5)

This has the representation

n+1
Q(z)
P(z) = wk , (1.6)
k=1
Q
(z k )(z − z k )
n+1
where Q(z) = k=1 (z − z k ).
 1.1 Complex polynomials 3
By the previous result there is at most one such polynomial; on the other hand
the polynomial so represented is immediately seen to have the desired property.

1.1.5
An alternative formulation of Lagrange’s formula is the following.
n+1
Corollary Let z 1 , z 2 , . . . , z n+1 be n+1 distinct points and let Q(z) = k=1 (z−
z k ). If P(z) is a polynomial of degree not exceeding n, then


n+1
Q(z)
P(z) = P(z k ) . (1.7)
k=1
Q
(z k )(z − z k )

In other words we have an explicit method for determining the values of a

polynomial of degree n in terms of its values at n + 1 known points.

1.1.6
n
Example Let Q(z) = z −1 = k=0 (z −ωk ), where ωk = e2πik/(n+1) are the
n+1

(n + 1)th roots of unity. Then, if P(z) is a polynomial of degree not exceeding

n, we have
1
n
P(z) = P(ωk )en (ωk z), (1.8)
n + 1 k=0
where en (z) = 1 + z + · · · + z n .

1.1.7 Representation for harmonic polynomials

A harmonic polynomial T (z) is a function of the form T (z) = Q(z) + P(z),
where Q and P are analytic polynomials, and so is a complex-valued harmonic
function in C (the complex plane). T can be represented in the form


n
T (z) = ak r |k| eikθ (z = r eiθ ). (1.9)
k=−n

T has degree n if either an or a−n is non-zero. Note that a harmonic polynomial

is the sum of a polynomial in the variable z and a polynomial in the variable
z. The polynomial is uniquely determined by the 2n + 1 coefficients ak . These
coefficients also determine uniquely the trigonometric polynomial


n
T (eiθ ) = ak eikθ . (1.10)
k=−n
4 1 The algebra of polynomials
Indeed, given this latter expression the coefficients can be recovered from the
formula
 2π
1
ak = e−ikθ T (eiθ )dθ (−n ≤ k ≤ n). (1.11)
2π 0
We can apply Lagrange’s interpolation formula to obtain a representation for a
harmonic polynomial as follows. Let ζk = e2πik/(2n+1) (0 ≤ k ≤ 2n) and set
Dn (z) = z n + · · · + z + 1 + z + · · · + z n . (1.12)
Then, if T (z) is a harmonic polynomial of degree not exceeding n, we have

1
2n
T (z) = T (ζk )Dn (ζ k z). (1.13)
2n + 1 k=0

This follows easily by applying example 1.1.6 to einθ T (eiθ ), which is a poly-
nomial of degree at most 2n in the variable eiθ . We obtain the above formula
on the unit circle. Since both sides of the equation are harmonic polynomials,
it follows that the equality holds for all z ∈ C.

1.2 The number of zeros of a real analytic polynomial

1.2.1
Definition A real analytic polynomial is an expression of the form


m

n
P(x, y) = a j,k x jy k (1.14)
j=0 k=0

where the coefficients a j,k are real or complex numbers and where x and y are
real variables. The degree of the term a j,k x j y k is j + k provided that a j,k = 0,
and the degree of P is the largest of the degrees of the individual terms. Every
algebraic polynomial and every harmonic polynomial is real analytic and their
degrees earlier defined agree with the above definition.

1.2.2
Bézout’s theorem Let P(x, y) = u(x, y) + iv(x, y) (with u and v real) be
a complex-valued real analytic polynomial, where u has degree m and v has
degree n, and suppose that u and v are relatively prime (i.e. contain no non-
trivial common factors). Then P has at most mn zeros in C.

Proof This is an algebraic result which is proved using linear techniques.

Firstly, we may note that P(0, y) is not identically zero, for otherwise x is
 1.2 The number of zeros of a real analytic polynomial 5
a common factor of u and v. We then write u and v in the form

m
n
u(x, y) = u k (x)y k , v(x, y) = vk (x)y k , (1.15)
k=0 k=0

where u k are real polynomials in x of degree at most m−k and vk are real
polynomials in x of degree at most n−k. Also, after an affine transformation of
the variables, we may assume that u m and vn = 0. (Affine transformations are
first degree real analytic polynomial mappings, which are bijective mappings
of C; for an account of these see chapter 2, section 2.7.21.) Now the equation
P = 0 holds iff u = 0 and v = 0 simultaneously. Let us assume this is the case
at a particular pair (x, y). We use the method of Sylvester to eliminate the
quantity y from the two equations u(x, y) = 0, v(x, y) = 0. The method is to
observe that the following m + n equations hold:
y µ u(x, y) = 0 (0 ≤ µ ≤ n − 1); y ν v(x, y) = 0 (0 ≤ ν ≤ m − 1).
(1.16)
Here y 0 = 1 by definition. Each of these equations can be interpreted as the
vanishing of a linear combination of the m + n quantities 1, y, y 2 , . . . , y m+n−1 .
This implies the vanishing of the determinant of coefficients:
 
u 0 (x) u 1 (x) u 2 (x) 0 

 0 u 0 (x) u 1 (x) u 2 (x)
D(x) =  = 0. (1.17)
v
 0 (x) v1 (x) v2 (x) 0 
 0 v0 (x) v1 (x) v2 (x) 
This example is the case m = n = 2. The determinant D(x) is a polynomial
in x. We will show that (a) D(x) does not vanish identically and (b) D(x) has
degree at most mn. When these facts are established, it will follow that there
are at most mn values of x for which the determinant can vanish. For each
such x there are only finitely many values of y such that P(x, y) = 0. Thus we
obtain at most finitely many points (x, y) for which P(x, y) = 0. Because of
this we may now perform an affine transformation on (x, y) so that, after the
transformation to new variables x
, y
, the resulting polynomial has for each
fixed x
at most one zero in y
. (This is easily seen geometrically; e.g. a small
rotation will do.) It follows immediately from the above applied to the new
coordinates that P has at most mn zeros altogether, proving the theorem.
To prove (b) we let wi, j denote the terms of the (m + n) × (m + n) matrix
of coefficients and observe that each term is a polynomial in x whose degree
satisfies

m − j +i (1 ≤ i ≤ n),
degree of wi, j ≤ (1.18)
−j + i (n + 1 ≤ i ≤ m + n).
6 1 The algebra of polynomials
Here we have taken the degree of the identically zero polynomial to be −∞.
Now the value of the determinant is given by the formula


m+n
|wi, j | = (−1)σ wi,φ(i) (1.19)
i=1

where the sum ranges over all permutations φ(i) of the numbers 1, 2, . . . , m + n,
and where σ = 0 for even permutations and σ = 1 for odd permutations. It
follows that the degree of the determinant cannot exceed


n

m+n

m+n
(m − φ(i) + i) + (−φ(i) + i) = mn + (−φ(i) + i) = mn,
i=1 i=n+1 i=1
(1.20)
 
since φ(i) = i, as φ(i) runs through the numbers i in some order. This
proves (b).
To prove (a) we assume that the determinant vanishes identically and show
that this implies that u and v have a common factor, contradicting the hypothe-
ses. Firstly, we observe that the vanishing of the determinant is equivalent to
the vanishing of the determinant of the transpose matrix. Secondly, we observe
that the elements of the matrix are polynomials in x and therefore belong to the
field of rational functions in x. Because the determinant vanishes for all x, we
may interpret this as the statement that the determinant of this matrix consisting
of elements of this field vanishes, as all the operations required to evaluate the
determinant are algebraically defined over any field. Therefore from the alge-
braic theory of determinants we can assert the existence of elements αk , βk of
the field not all zero and satisfying the set of equations
u 0 α0 + v0 β0 = 0, u 1 α0 + u 0 α1 + v1 β0 + v0 β1 = 0, . . . etc.
(1.21)
or generally


r

r
u r −k αk + vr −k βk = 0 (0 ≤ r ≤ m + n − 1), (1.22)
k=r −m k=r −n

where αk = 0 for k ≥ n and βk = 0 for k ≥ m. We now set

n−1

m−1
A(y) = αk y k , B(y) = − βk y k (1.23)
k=0 k=0

and note that by multiplying out and equating coefficients to zero the above set
of equations is equivalent to the single equation
u A − v B = 0. (1.24)
 1.2 The number of zeros of a real analytic polynomial 7
Since not both A and B are zero, it follows from this equation that neither A
nor B is the zero element, as otherwise one of u and v is zero, contradicting the
hypothesis. Now the coefficients of A and B are rational functions in x; therefore
multiplying through by a suitable polynomial in x, we obtain an equation of the
same form with A and B polynomials in both x and y. Furthermore, we may
divide out the equation by any common factor of A and B. Thus we obtain an
equation of the above form with A and B relatively prime polynomials, where
the degree of A in y is smaller than the degree of v in y and the degree of B
in y is smaller than the degree of u in y. We will show that f = v/A = u/B
is a polynomial in x and y, which is then the required common factor of u and
v. Consider first the case when B is constant in the variable y, and so a pure
polynomial in x. B can be factorised as a product of irreducible polynomials;
hence it is enough to show that each irreducible factor R of B is a factor of u.
Expanding u in powers of y, this is true iff every coefficient of u is divisible by
R. Write u = u 1 + u 2 , where u 1 consists of those terms in the expansion of u
whose coefficients are divisible by R, and u 2 consists of the remaining terms.
Similarly, write A = A1 + A2 . Since R is not a factor of A, A2 = 0. Thus we
have v B = Au = A1 u 1 + A1 u 2 + A2 u 1 + A2 u 2 , from which we see that A2 u 2
is divisible by R. This implies that u 2 = 0, as otherwise the lowest coefficient
of this product in y is divisible by R, but also is the product of a coefficient of
A2 and a coefficient of u 2 , which is not divisible by R. Thus B divides u in this
case.
Returning to the general case, from the usual division algorithm for polyno-
mials we can write
v ρ u σ
= f + , =g+ (1.25)
A A B B
where f, g, ρ and σ are polynomials in y with rational coefficients in x, and
where ρ has y-degree smaller than A and σ has y-degree smaller than B.
Hence as y → ∞, ρ/A and σ/B → 0 and so f − g → 0. This clearly implies
that f = g. We therefore obtain ρ/A = σ/B and so either (i) ρ = σ = 0 or
(ii) A/ρ = B/σ . In case (ii) we see that we are back to the same problem for A
and B that we had for u and v, but A and B have smaller degrees. Therefore we
may use an induction argument to deduce that A and B have a common factor,
contradicting our legitimate assumption. It follows that case (i) holds and so f
is a common factor of u and v. The coefficients of f are rational functions in x,
but by multiplying through by the smallest polynomial p in x to cancel out the
denominators of the coefficients, we find as above that p divides both A and
B. Since A and B are relatively prime, f is a polynomial in both x and y. This
proves (a) and so completes the proof of Bézout’s theorem.
8 1 The algebra of polynomials
1.2.3
Bézout’s theorem can be used to put a bound on the number of zeros of a
real analytic polynomial even in the absence of algebraic information on the
polynomial. For example it may often be the case that on analytic or topological
grounds one can establish the finiteness of the number of zeros. In this case we
have the following result.

Theorem Let P(x, y) = u(x, y) + iv(x, y) be a complex-valued real analytic

polynomial, where u has degree m ≥ 0 and v has degree n ≥ 0. Then P has at
most mn isolated zeros in C. In particular, if P has at most finitely many zeros,
then P has at most mn zeros in C.

Proof The following argument is an adaptation of a method given by

Wilmshurst [72]. If m = 0, then u is a non-zero constant and so P has no
zeros; similarly if n = 0. If m = 1, then either (i) u divides v or (ii) u and v
are relatively prime. In case (ii) Bézout’s theorem shows that P has at most n
zeros; in case (i) P vanishes exactly when u vanishes, which is on a line. Hence
P has no isolated zeros. Thus the theorem follows in this case. We proceed by
induction and assume the theorem is true for polynomials U + i V , where U has
degree smaller than m and V has degree smaller than n. If u and v are relatively
prime, then Bézout’s theorem gives the result. Therefore we may assume that
u and v have highest common factor p say, where p is a real polynomial in
(x, y) of degree r ≥ 1. Then u/ p and v/ p are relatively prime polynomials of
degrees m − r and n − r respectively, and so P/ p has at most (m − r )(n − r )
zeros. The remaining zeros of P are the zeros of p. Now p is a real-valued
continuously differentiable function in the plane. It follows from the implicit
function theorem (see chapter 2) that at a zero c of p either (i) p vanishes on
a curve passing through c or (ii) both partial derivatives px and p y are zero
at c. In case (i) c is a non-isolated zero of P. Thus if c is an isolated zero of
P, which is also a zero of p, then c is an isolated zero of p, and so from case
(ii) the polynomial Q = p + i px has an isolated zero at c. Since p has degree r
and px has degree r −1, it follows by the induction hypothesis that Q has at most
r (r − 1) isolated zeros. Hence P has at most (m − r )(n − r ) + r (r − 1) < mn
isolated zeros. The result follows by induction.

1.2.4 A complex approach

It is clear that Bézout’s theorem will be equally valid for a pair of complex
polynomials u(z, ζ ), v(z, ζ ) of complex variables z and ζ. From the two equa-
tions u(z, ζ ) = 0 and v(z, ζ ) = 0 we can use Sylvester’s method to eliminate
 1.2 The number of zeros of a real analytic polynomial 9
the variable ζ , obtaining a determinant D(z), which is an algebraic polynomial
in z whose zeros contain those z for which there exists a pair (z, ζ ) satisfying
the two equations. The maximum degree of D is the product of the degrees of
u and v. We can apply this result to our original problem of finding the zeros
of a complex-valued real analytic polynomial P. P is a polynomial in the real
variables x and y; by making the substitutions x = 12 (z + z ), y = 2i1 (z − z ), P
becomes a polynomial P(z, z) in the ‘variables’ z and z; in other words a real
analytic polynomial can be written in the form of a finite sum


P(z, z) = a j,k z j z k , (1.26)
j,k

where the a j,k are complex numbers and the degree of P is the maximum of
j + k over terms where a j,k = 0. This is clearly the restriction of P(z, ζ ) =

j,k a j,k ζ z to ζ = z. To apply the Sylvester method of elimination we require
j k

two equations. However, P(z, z) = 0 iff P(z, z) = 0; thus we eliminate ζ from

the two equations

P(z, ζ ) = 0, P(ζ , z) = 0 (1.27)

which take the form

a j,k ζ j z k = 0, a j,k z j ζ k = 0. (1.28)
j,k j,k

We obtain the determinant equation D(z) = 0, where D is an algebraic poly-

nomial in z whose degree does not exceed n 2 , where n is the degree of P. D
is identically zero iff the above pair of polynomials in z and ζ have a common
factor. We leave it as an exercise to show that this holds iff, writing P = u + iv
(u, v real), u and v have a common factor. Thus in this formulation we have lost
the distinction between the possible different degrees of u and v, as n is clearly
the larger of these two degrees. On the other hand we have constructed an alge-
braic polynomial D(z) whose set of zeros contains all the zeros of the original
real analytic polynomial P. In fact, if (z, ζ ) is any pair satisfying the above two
equations, then D(z) = 0. In general it need not be the case that the ζ satisfies
ζ = z, and therefore D is likely to have more zeros than P. However, there are
certainly cases where D and P have the same zeros. Indeed this is the case if
P is itself an algebraic polynomial: then D = c P n , where c is a constant = 0.
Of course, if we count multiplicities, then D has n 2 zeros, whereas P has only
n zeros. We will give an example in chapter 2 (subsection 2.6.11) where D has
n 2 simple zeros all of which are zeros of P.
10 1 The algebra of polynomials
It is worth recording here the form of the 2n × 2n matrix of which D is the
determinant. We have






P(z, z) = a j,k z j z k = a j,k z k z j = a j,k z j z k (1.29)
j,k j k k j

and so


n

n
P(z, z) = b j (z)z j , P(z, z) = c j (z)z j , (1.30)
j=0 j=0

where


n− j

n− j
b j (z) = a j,k z k , c j (z) = ak, j z k . (1.31)
k=0 k=0

The terms of the matrix are given by

wi, j = b j−i for i ≤ j ≤ n + i and 1 ≤ i ≤ n,

wi, j = c j−i+n for i − n ≤ j ≤ i and n + 1 ≤ i ≤ 2n,
wi, j = 0 otherwise. (1.32)

As recorded earlier these relations imply that the maximum degree of wi, j is
given by

n− j +i (i ≤ j ≤ n + i and 1 ≤ i ≤ n),
degree of wi, j ≤
−j + i (i − n ≤ j ≤ i and n + 1 ≤ i ≤ 2n).
(1.33)

Otherwise degree of wi, j = −∞ (so trivially satisfies the above inequalities).

We then have


2n
D(z) = (−1)σ wi,φ(i) (1.34)
i=1

where the sum ranges over all permutations φ(i) of the numbers 1, 2, . . . , 2n,
and where σ = 0 for even permutations and σ = 1 for odd permutations.
2n n
Each product i=1 wi,φ(i) has degree not exceeding i=1 (n − φ(i) + i) +
2n
i=n+1 (−φ(i) + i) = n 2
, with equality only if every term wi,φ(i) in the product
attains its maximum degree n − φ(i) + i for 1 ≤ i ≤ n or −φ(i) + i for n + 1 ≤
i ≤ 2n.
As an application of these remarks, suppose that P has coefficients a j,k where
a j,n− j = 0 for 1 ≤ j ≤ n and a0,n = 0; in other words the only term of highest
degree n in the expansion of P is the term in z n . Then b0 (z) has full degree n,
and c0 (z) has degree 0, but the remaining polynomials b j (z) (1 ≤ j ≤ n) and
c j (z) (0 ≤ j ≤ n − 1) have degree strictly smaller than n − j. It follows that the
 1.2 The number of zeros of a real analytic polynomial 11

only permutation giving a product of degree n 2 is the case φ(i) = i (1 ≤ i ≤ 2n).

Thus D(z) has exact degree n 2 , and therefore P has at most n 2 zeros.
Collecting up the results of this subsection we have the next theorem.

1.2.5
Theorem Let P(z, z) = u +iv (u and v real) be a complex-valued real analytic
polynomial of degree n. Then the Sylvester resultant D(z) of P(z, ζ ) and P(ζ , z)
is a polynomial in z of degree at most n 2 . Every zero of P(z, z) is a zero of D(z);
indeed, if P(z, ζ ) = P(ζ , z) = 0, then D(z) = 0. D(z) vanishes identically if,
and only if, u and v have a non-trivial common factor. Finally, if
P(z, z) = Q(z, z) + Az n , (1.35)
where Q is a real analytic polynomial of degree < n and A is a non-zero
constant, then D(z) has exact degree n 2 and therefore P has at most n 2 zeros.
Indeed, in this case, P attains every value w at most n 2 times and is therefore
an n 2 -valent mapping of the plane.

The final remark comes from applying the theorem to P − w. We shall see in
chapter 2 that P of this form does actually attain every value w, and so is a
surjective mapping of the plane.

1.2.6 The Sylvester resultant of P − w

In studying the valence of a mapping P(z, z) it is natural to consider the
Sylvester resultant for P − w, where w is an arbitrary complex number. The
resultant will take the form D(z, w, w), where this is an algebraic polynomial
in z and a real analytic polynomial in w. Furthermore, if P has degree n, then
as a real analytic polynomial in w, D has at most degree n. This follows from
the fact that the only terms involving w in the matrix are the terms b0 − w and
c0 − w, so w only appears with degree 1 in wi, j if j = i for 1 ≤ i ≤ n and if
j = i − n for n + 1 ≤ i ≤ 2n. All other terms have degree 0 or −∞ in w. For
any permutation φ of the numbers 1, 2, . . . , 2n, there are at most n values i such
that (i) φ(i) = i where 1 ≤ i ≤ n and (ii) φ(i) = i − n where n + 1 ≤ i ≤ 2n.

1.2.7 Evaluating the Sylvester resultant

Let b0 , b1 , . . . , bn and c0 , c1 , . . . , cn be the generating elements of the 2n × 2n
Sylvester matrix. Then the resultant can be written in the form


n
(−1)σ bφ(i)−i cφ(i+n)−i (1.36)
φ i=1
12 1 The algebra of polynomials
summed over all permutations φ of the numbers 1, 2, . . . , 2n such that for
1≤i ≤n
0 ≤ φ(i) − i ≤ n, 0 ≤ φ(i + n) − i ≤ n. (1.37)
Note that, for each permutation φ, the sum of the subscripts of the bs and the
cs is exactly n 2 .

1.2.8 Further properties of the resultant

Let u and v be real analytic polynomials of degrees m and n respectively written
in the form

m
n
u(x, y) = u k (x)y k , v(x, y) = vk (x)y k , (1.38)
k=0 k=0

where u m = 0 and vn = 0. If the resultant is not identically zero, then the set
of equations


r

r
u r −k αk + vr −k βk = λr (0 ≤ r ≤ m + n − 1) (1.39)
k=r −m k=r −n

can be uniquely solved for αk and βk , where αk = 0 for k ≥ n and βk = 0 for

k ≥ m. Here the λr are given arbitrary elements in the field of rational functions
in x, and the solutions αk and βk are then also elements in this field. Writing


m+n−1

n−1

m−1
(x, y) = λr (x)y r , A(x, y) = αk (x)y k , B(x, y) = βk (x)y k ,
r =0 k=0 k=0
(1.40)
the above equations are equivalent to the single equation
u A + v B = . (1.41)
In other words, if the resultant of u and v does not vanish identically, then given
 of degree in y at most m + n − 1 we can find a unique A of degree at most
n − 1 in y and a unique B of degree at most m − 1 in y to satisfy this equation.
In particular, taking  = 1, we can find a unique p of degree at most n − 1
and a unique q of degree at most m − 1 such that
up + vq = 1. (1.42)
Here p and q are polynomials in y with coefficients which are rational functions
in x. If R(x) is the lowest common denominator of these rational coefficients,
then we obtain the relation
uP + vQ = R (1.43)
 1.3 Real analytic polynomials at infinity 13
where P = p R, Q = q R are now polynomials in both x and y. The polynomial
R(x) is a factor of the resultant: if we use the Sylvester matrix to solve the
equation up + vq = 1 for p and q, then, applying Cramer’s rule, the solutions
for each of the coefficients of p and q (expanded as polynomials in y) will be
the ratio of two determinants; the numerator is ± the determinant of one of
the minors of the matrix corresponding to the first column; the denominator
is the determinant of the matrix, i.e. the Sylvester resultant. Since u and v have
polynomial coefficients, it follows that the Sylvester resultant is a common
denominator of all the rational coefficients of p and q. Hence the lowest common
denominator R is a factor of the resultant; if R is a proper factor, then there is
a factor of the resultant which is a common factor of the determinants of each
of the minors corresponding to the first column.

1.3 Real analytic polynomials at infinity

1.3.1 Resolving the singularity: the blow-up method
Let P(x, y) be a real analytic polynomial of degree N . The limiting behaviour
at infinity of P can be determined by performing a sequence of transformations,
which we now describe. Firstly, after a suitable affine transformation we may
assume that both the real and imaginary parts of P(0, y) have full degree N .
We then make the transformation
x → 1/x, y → y/x, (1.44)
obtaining

1 y Q(x, y)
R(x, y) = P , = , (1.45)
x x xN
where Q(x, y) is a polynomial of degree N such that Q(0, y) has degree N .
Any finite limiting value of P at infinity is then a limiting value of R as x → 0
and y remains bounded. We expand Q in powers of x, obtaining


N
Q k (y)x k
k=0
R(x, y) = . (1.46)
xN
It is clear that R(x, y) → ∞ as x → 0 and y remains bounded, unless y tends to a
real zero of Q 0 (y); for otherwise the term Q 0 (y)/x N dominates the expression.
Thus to determine the possible finite limiting values of R we will consider
x → 0 through positive values and y → a, where a is a real zero of Q 0 . We
write
Q k (y) = (y − a)rk Sk (y) (0 ≤ k ≤ N ), (1.47)
14 1 The algebra of polynomials
where rk is the multiplicity of the zero a of Q k (if Q k (a) = 0, then rk = 0); if
Q k ≡ 0, for the purposes of this discussion we may take rk = +∞. Next, let
us consider the substitution
y = a + t x p, (1.48)
where p is a positive number to be determined. Note that the mapping (x, y) →
(x, t) is a 1–1 mapping of the half-plane x > 0 onto itself. We obtain



N
−N rk prk +k
R(x, y) = x Sk (y)t x . (1.49)
k=0

Now observe that, for small positive p, pr0 < prk + k (1 ≤ k ≤ N ) and the
dominating term is S0 (a)t r0 /x N − pr0 . If pr0 = N and pr0 < prk + k (1 ≤ k ≤ N ),
then with the choice p = N /r0 we obtain the limiting values S0 (a)t r0 (t real). If
r0 is odd, this is a simply described straight line through the origin; if r0 is even,
it is a doubly described ray with endpoint at the origin. Otherwise, we choose
p > 0 to satisfy pr0 = pr j + j for some j and pr0 ≤ prk + k (1 ≤ k ≤ N );
in other words
j
p = min , (1.50)
r0 − r j
where the minimum is taken over those j (1 ≤ j ≤ N − 1) for which r0 > r j .
We obtain
F(x, t)
R(x, a + t x p ) = N − pr , (1.51)
x 0

where F(x, t) is polynomial in t of degree at most N , and contains fractional

powers of x – in fact is polynomial in x and x p . Note also that F(0, t) is a
polynomial in t of degree r0 , which contains only those powers t r j for which
pr0 = pr j + j. Suppose now that p = µ/ν, where µ and ν are relatively prime
natural numbers. Then with the substitution x = s ν we obtain
A(s, t)
R(s ν , a + ts µ ) = ν N −µr , (1.52)
s 0

where A(s, t) is a polynomial in s and t and A(0, t) = F(0, t) has degree r0 .

Furthermore, since the powers t r j occurring in A(0, t) satisfy pr0 = pr j + j,
we obtain
νj
r0 − r j = , (1.53)
µ
and, since µ and ν are relatively prime and r0 − r j is an integer, µ divides j
and so r0 − r j is an integral multiple of ν. It follows that A(0, t) has the form
A(0, t) = t α B(t ν ), (1.54)
where α is a non-negative integer and B(t) is a polynomial in t of degree ≥ 1.
 1.3 Real analytic polynomials at infinity 15
1.3.2
Now the form of R after this transformation is of the same general type in the
variables s and t as it was in the variables x and y, namely a polynomial in s
and t divided by a power of s. Therefore it is open to us to repeat this reducing
process. Indeed, it follows from our comparison principle (see 1.3.6 below) that
any finite limiting value of R will be attained along some curve y = a + t x p ,
with x → 0 and t remaining bounded, where p is a positive and indeed rational
number. On the other hand R → ∞ along such curves, if p is smaller than the
above minimum choice. Therefore, finite limiting values can only be attained
for either this or larger values of p. This implies that with our chosen minimum
value of p and the above transformation, limiting values of R (associated with
the zero a) will be attained with s → 0 and t remaining bounded (in fact t → 0
if the limiting value corresponds to a larger value of p). As before, each limiting
value will correspond to a real zero, b say, of the polynomial A(0, t). If the
multiplicity of the zero b is s0 , then we make a transformation t = b + us q ,
where q is chosen according to the same minimum process with which we
chose p. Now clearly s0 ≤ r0 ≤ N ; if s0 = r0 , then A(0, t) = c(t − b)r0 , where
c is a non-zero constant, and therefore A(0, t) contains all powers of t from
0 to r0 . It follows that ν = 1 and so the power of s in the denominator is at
most N − r0 . We see, therefore, that, if we continue repeating the process, at
each step either the degree of the leading term strictly decreases or the power
of the denominator strictly decreases by an integer amount. Thus eventually
the process will terminate in the following way. Using the original notation
Q(x, y)/x N as a standard form, the final transformation will take the form
y = a + t x p , where p = N /r0 . As earlier described either this will lead to
a line or ray of asymptotic values or this choice of p will coincide with the
minimum choice p = min( j/(r0 − r j )). Following the transformation as before
we obtain a curve of asymptotic values A(0, t), a polynomial in the real variable
t of degree r0 ≥ 1. Indeed, as the denominator is eliminated at the final step,
the sequence of transformations will send our original polynomial P to a new
polynomial A(s, t). It is clear that for the existence of asymptotic values it is both
necessary and sufficient that there exist such a finite sequence of transformations
from P to another polynomial A. If we perform our construction for every real
zero of our leading term at every stage, we will obtain all possible asymptotic
values with x > 0 tending to ∞. We can obtain in this way a maximum of
N = degree of P asymptotic value curves, which are polynomial images of
the real axis. Of course, if at one or other stage the leading term has no real
zeros, then the polynomial tends to ∞ along the resulting tract. In a similar
way, to obtain the asymptotic values of P with x < 0 tending to −∞, we
simply apply the same reasoning to the function (−1) N Q(−x, y)/x N , again
16 1 The algebra of polynomials
taking x > 0 and tending to 0. This gives another N possible asymptotic value
curves.

1.3.3 Repetition of asymptotic values

However, as we shall see, each asymptotic value of P at ∞ is repeated along
quite separate tracts, and in general each asymptotic value curve appears twice
(except in one circumstance), either repeated or reversed. This is a conse-
quence of the fact that, according to the above reasoning, the asymptotic values
are attained as limits along rational algebraic curves of a particularly simple
type. To be specific, if we put together the sequence of transformations from the
initial R(x, y) to the final polynomial A(s, t), we obtain a single transformation
of the form

x = sq , y = φ(s) + ts β, (1.55)

where q and β are natural numbers and where φ(s) is a real polynomial of
degree < β; furthermore, the highest common factor of the powers of s appear-
ing in the expression φ(s) + ts β is 1. In short, we have

R(s q , φ(s) + ts β ) = A(s, t). (1.56)

m
If φ(s) = k=0 ck s k , then the coefficients ck are zeros of successive leading
terms in our sequence of transformations. Note that the mapping (s, t) → (x, y)
is a 1–1 mapping of the half-plane {s > 0} onto the half-plane {x > 0}. Indeed,
this is clearly the case at each stage of the resolution process, which consists
of blowing up a chosen zero into an entire line, the remainder of the line being
pushed off to ∞. Furthermore, the mapping is 1–1 from {s < 0} onto {x < 0}, if
q is odd, and onto {x > 0}, if q is even. The curve A(0, t) is a limiting curve of
asymptotic values as s → 0 through either positive values of s or negative values
of s. If q is odd, the approach through negative values of s will correspond in
the (x, y)-plane to a tract in {x < 0} diametrically opposite the constructed tract
in {x > 0}, and therefore the asymptotic value curve is repeated in a completely
different portion of the plane. If q is even, the curve will again appear twice as
limits along separate tracts in {x > 0}, unless φ(−s) = φ(s).
This last possibility will occur if φ(s) is a polynomial in s 2 and β is odd. We
will show that repetition of asymptotic values still occurs, though in a slightly
strange way. Let d be the highest power of 2 which is a factor of q and the
powers of s appearing in the expansion of φ. Then x = s σ d , y = ψ(s d ) + ts β ,
where ψ(u) is a real polynomial in u and σ d = q. Now suppose that
ωd = 1 and γ = ωβ ; (1.57)
 1.3 Real analytic polynomials at infinity 17
then
x = (ωs)σ d , y = ψ((ωs)d ) + (γ t)(ωs)β (1.58)
and so
A(ωs, γ t) = A(s, t). (1.59)
In particular, since d is even and β is odd, this holds when ω = −1, γ = −1;
i.e.
A(−s, −t) = A(s, t), (1.60)
and in particular A(0, −t) = A(0, t). It follows that our polynomial curve is
an even function and so has the form A(0, t) = B(t 2 ). The curve traced out is
thus B(t) from +∞ to 0, followed by the reverse curve B(t) from 0 to +∞;
i.e. we obtain half of B(t) traversed twice. We show that in a separate tract
the other half of B(t) with t < 0 is similarly traversed. Firstly, note that either
(i) σ = q/d is odd, or (ii) σ is even. In case (i) we consider the tract given by
x = −s σ d = (λs)σ d , y = ψ(−s d ) + ts β = ψ((λs)d ) + ηt(λs)β , (1.61)
β
where λd = −1 and η = λ . Thus
R(x, y) = A(λs, ηt) → A(0, ηt) as s → 0. (1.62)
Now A(0, t) = A(0, γ t) and so the non-vanishing terms in the expansion of
A(0, t) involve only those powers t k for which γ k = 1; i.e. ωβk = 1. If ω =
e2πi/d , this implies that d divides βk and so, as d is a power of 2 and β is odd,
d divides k. Thus A(0, t) has the form
A(0, t) = F(t d ), (1.63)
where F(t) is a real polynomial. Thus A(0, ηt) = F(ηd t d ) = F(−t d ). Thus with
s → 0 through positive values, x → 0 through negative values; i.e. our tract lies
in the negative half-plane and has limiting values the half-curve described by
F(t) with t < 0, as asserted. In case (ii) we consider the tract
x = s σ d = (λs)σ d , y = ψ(−s d ) + ts β = ψ((λs)d ) + ηt(λs)β , (1.64)
and note that exactly the same reasoning applies, but our tract will now lie in
the positive half-plane. However, in this case the polynomial ψ(s) contains at
least one odd power of s and so we obtain a distinct tract.

1.3.4
As we have seen, apart from the above case, the asymptotic value curve A(0, t)
appears twice corresponding to the distinct tract obtained by changing the
18 1 The algebra of polynomials
variable s to −s. The parameter β determines whether the curve is repeated
or reversed as the tract is traversed in the positive direction: the positive direc-
tion corresponds to y increasing. If β is odd, s → −s gives the tract

x = (−s)q , y = φ(−s) − ts β (1.65)

and for s > 0, y is clearly a decreasing function of t. Therefore, as the tract

is traversed positively, we obtain the limiting curve A(0, −t); i.e. the curve is
reversed. On the other hand, when β is even, s → −s gives y = φ(−s) + ts β
is an increasing function of t, and so the limiting curve is A(0, t); i.e. the curve
is repeated in the same direction.

1.3.5 Inverting the transformation

The mapping x = s , y = φ(s) + ts β has the inverse
q

s = x 1/q , t = x −β/q (y − φ(x 1/q )) (1.66)

and so
R(x, y) = A(x 1/q , x −β/q (y − φ(x 1/q ))). (1.67)
This gives

P(x, y) = A(x −1/q , x β/q (y/x − φ(x −1/q ))), (1.68)

and so

P(x q , y) = A(x −1 , −x β φ(x −1 ) + yx β−q ) = A(x −1 , ψ(x) + yx α ), (1.69)

where α = β − q and, since φ has degree β − h, where h ≥ 1, ψ(x) =

−x β φ(1/x) is a polynomial of the form ψ(x) = cx h + · · · , where c = 0.
This may be regarded as the inverse of the formula

A(s, t) = P(s −q, φ(s)s −q + ts α ). (1.70)

From these relations we see that on the curve y = x q φ(x −1 ) we have

P(x q , y) = A(x −1, 0), (1.71)

and, if q ≥ β − h, y remains bounded on this curve as x → 0. Thus A(x −1 , 0)

remains bounded as x → 0; i.e. A(x, 0) is a bounded polynomial as x → ∞.
This implies A(x, 0) is constant, and so P(x q, y) is constant on the curve.
Therefore, unless the polynomial P(x, y) is constant on some curve,

q < β − h < β, (1.72)

 1.3 Real analytic polynomials at infinity 19
and in particular
β ≥q +2 and α > h. (1.73)
To reiterate: knowing the tract equation y = φ(s)s −q + ts α, we obtain the
polynomial A(s, t) from P(x, y); conversely, using the reverse tract equation
t = −s β φ(1/s) + ys α , we can recover P(x q, y) from A(s, t). However, if we
work directly with A(s, t), applying our process, but not assuming any prior
knowledge of the asymptotic tracts, we will not necessarily obtain the above
tract for t. The reason for this is that α need not be greater than the largest
power of s in the expression s β φ(1/s). No term s γ with γ ≥ α is needed in
determining the asymptotic tract appropriate for A(s, t) in the above direction;
we will obtain a tract by deleting all terms involving those powers s γ with
γ ≥ α. On the other hand we can obtain precisely invertible transformations
if φ has the form φ(x) = cr x r + · · · , where cr = 0 and r > q; for then
α > deg(s β φ(1/s)). Then the asymptotic tract is such that y → 0 as x → ∞.
In this case, if the process is actually carried through, at each stage the value
of p is always an integer and furthermore the zero is always unique and is a
multiple zero of exact degree N ; this is because the initial largest power of
t is t N , but also the final leading term is P(0, y), which has exact degree N
(by our initial assumption). Therefore the leading term has exact degree N at
each stage. This phenomenon indicates that the possible polynomials A(s, t)
which can be obtained by applying the process to a polynomial P are of a rather
special type. A(s, t) has the asymptotic value curve P(0, y) along a curve with
s → ∞, t → 0; but all other asymptotic values of A, if any, will occur along
curves where s → 0, t → ∞.

1.3.6 The comparison principle

We have omitted from our discussion of asymptotic values an important general
principle, which establishes the algebraic nature of the asymptotic tracts and
also the fact that, for real analytic polynomials, the only finite limiting values
at ∞ are asymptotic values. The argument is easily sketched as follows.
Suppose we have a finite sum of terms, for which along a sequence the sum
tends to a finite value, though some individual terms tend to ∞. We compare
two terms by considering for a suitable global subsequence the ratio of the
absolute values of the two terms. If the limit of the ratio is finite and non-zero,
we say the terms are comparable. If the limit is zero, then the upper term of the
ratio is smaller than the lower term; and vice versa. In this way we can order the
groups of comparable terms and pick out the largest grouping in this ordering. It
20 1 The algebra of polynomials
is clear that the largest group must contain at least two terms, for if it contained
only one, then this term would dominate the sum, and so the sum would tend
to ∞. For polynomial terms the comparison of two such terms gives a tract of
the type described above in which the sequence must lie.

1.3.7 Solving algebraic equations and algebraic functions

An algebraic function f (x) is a function satisfying an equation of the form

P(x, f (x)) = 0, (1.74)

where P(x, y) is a real analytic polynomial. In other words y = f (x) is a

solution of the equation P(x, y) = 0. Of course, at a given x the equation
may have no solutions or several solutions, though no more than the degree
of P. We are interested in continuous solutions for an interval of values of x.
We can use our procedure to construct an explicit form for a solution. After
a normalisation we may assume that P(0, 0) = 0, and our aim is to construct
a solution f (x) with f (0) = 0. Let us attempt to find a solution for small
x > 0; if such a solution exists then for an arbitrary natural number N the
function R N (x, y) = P(x, y)/x N has 0 as an asymptotic value as x → 0;
namely along the curve y = f (x). Of course, our procedure is designed to find
all limiting values of R N as x → 0; we obtain a finite number of algebraic
curves y = φk (x) + t x βk , which are polynomials in a fractional power of x,
each of which reduces R N (x, y) to a polynomial in a fractional power of x
and t; we then put x = 0 giving a polynomial in t, whose values for t real
are the limiting values; we are concerned with the values of t which yield the
limiting value 0, i.e. the zeros of the final polynomial, which is simply taking
the procedure to the next stage. Thus 0 will be a limiting value of R N along
a finite number of curves y = ψk,N (x). It is possible that for one, or more, of
the ψk,N we have R N (x, ψk,N (x)) = 0; then ψk,N (x) is a required solution and
furthermore we have an explicit formula in powers of x for the solution. On the
other hand any actual continuous solution y = f (x) must satisfy for some k

f (x) = φk (x) + t(x)x βk , (1.75)

where t(x) tends to a zero of the final polynomial as x → 0; in other words one
of the ψk,N is an approximation to f . Since ψk,N is a partial sum of ψk,N +r , we
see that we obtain a series expansion for f by letting N →∞. At each stage the
degree of the leading term does not increase, so eventually becomes a constant
ν. Thus for large N the values of p will be integers and each leading term will
 1.3 Real analytic polynomials at infinity 21
be a pure νth power of a linear polynomial giving just one multiple zero. This
implies that the series expansion for f is a power series in x 1/q for some natural
number q. Our procedure continued indefinitely by increasing N will yield all
the solutions y = f (x) for x > 0 near 0 (i.e. all the branches of the algebraic
function) and each such solution is a convergent power series, and therefore an
analytic function, in a fractional power of x.