See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/233691648

Rational Functions

Article in The American Mathematical Monthly · November 2009

DOI: 10.4169/000298909X474873

CITATIONS READS
3 3,078

1 author:

Luis Verde-Star
Metropolitan Autonomous University
65 PUBLICATIONS 598 CITATIONS

SEE PROFILE

All content following this page was uploaded by Luis Verde-Star on 10 October 2014.

The user has requested enhancement of the downloaded file.

 Rational Functions
Luis Verde-Star

1. INTRODUCTION. Rational functions are often seen as pairs of polynomials that

can be added and multiplied. Their vector space properties are rarely mentioned in
elementary textbooks.
In this paper we use a linear algebraic approach to study the vector space of rational
functions over the field of complex numbers. We find a basis for the space of proper
rational functions and prove the general partial fractions decomposition formula and
some other basic results.
Duality in the linear algebra sense plays a central role in our development. We use
pairs of dual bases and also the duality associated with a bilinear form on the space
of rational functions, which allows us to consider the proper rational functions as a
distinguished subspace of the dual space of the polynomials.
We also consider some spaces that are isomorphic to the proper rational functions.
The exponential polynomials and the linearly recurrent sequences are examples of such
spaces. Each of them has a natural multiplication that can be transferred to any other
isomorphic space. This explains the origin of the Hadamard and Hurwitz convolutions
of rational functions, and the Duhamel convolution of exponential polynomials. The
isomorphisms among those spaces provide an algebraic foundation for the transform
methods used to find solutions of linear functional equations.
Duality can also be used to induce an algebraic structure on the vector dual of
an algebra. This leads to the concepts of coalgebra, bialgebra, and Hopf algebra. In
Section 6 we describe briefly the Hopf algebra of the polynomials and its Hopf algebra
dual, which turns out to be isomorphic to the proper rational functions with the Hurwitz
multiplication and some additional structure. For this purpose we use some abstract
algebra ideas, such as tensor products and ideals. In the rest of the paper we use only
elementary linear algebra tools, dual bases, Taylor functionals, and the Leibniz rule.
2. POLYNOMIALS. In this section we recall some basic facts about polynomials
that we will use to obtain properties of rational functions. We also introduce some
basic definitions and notation.
If V is a complex vector space and L : V → C is a linear map, we say that L is
a linear functional on V . We will write L , u to denote the value of L applied to an
element u of V . The set of all linear functionals on V is a vector space called the dual
of V and denoted by V ∗ .
Let P denote the complex vector space of polynomials in one variable, that is, the
set of all finite linear combinations of nonnegative powers of z.
By repeated differentiation we get the formula
 
Dk m
(z − a) =
m
(z − a)m−k , a ∈ C, k, m ∈ N. (2.1)
k! k
Evaluation of the right-hand side at z = a gives us zero if k  = m, and one if k = m.
We define the Taylor functionals Ta,k on the space P by
 
Ta,k , p(z) =
p (k) (a)
k!
, a ∈ C, k ∈ N, p ∈ P. (2.2)

doi:10.4169/000298909X474873

804 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
Then we have
 
Ta,k , (z − a)n = δk,n , a ∈ C, k, n ∈ N, (2.3)

where δk,n is Kronecker’s function, whose value is one if k = n and zero otherwise.
Note that Ta,k can be applied to any function f for which f (k) (a) exists, not only to
polynomials. Thus Ta,k can be considered as a linear functional on many vector spaces
whose elements are sufficiently differentiable functions at a. We will use Ta,k to denote
functionals with different domains, indicating what the domain is whenever it is not
clear from the context. We will apply Taylor functionals to rational functions in the
next section.
Using (2.3) it is easy to show that for each a the set {(z − a)n : n ∈ N} is linearly
independent. In particular, the nonnegative powers of z are linearly independent and
therefore they form a basis of . P
The binomial formula gives

n  
n
(z − b)n = (z − a + a − b)n = (a − b)n−k (z − a)k . (2.4)
k=0
k

This means that every finite linear combination of the powers (z − b)k , for some num-
ber b, is also a finite linear combination of powers (z − a)k for any a. Therefore, for
each complex number a the set {(z − a)n : n ∈ N} is a basis for . P
P
For n ≥ 0 we denote by n the subspace of P
of all the polynomials with degree
P
at most equal to n. Note that, since the dimension of n is n + 1, for each a the set
P
{(z − a)k : 0 ≤ k ≤ n} is a basis for n . From (2.3) we see that, for each complex
P
number a, every polynomial p(z) in n has a unique representation of the form


n
 
p(z) = Ta,k , p (z − a)k .
k=0

This result is usually called Taylor’s theorem for polynomials.

P
Note that (2.4) implies that n is invariant under the translation operators that send
p(z) to p(z − b) for some complex number b.
The proof of the linear independence of the powers of z − a, using relation (2.3),
appears quite often in elementary textbooks. The usual method of setting up a system
of linear equations by evaluation at distinct points leads to a Vandermonde matrix, and
showing that such a matrix is nonsingular requires some additional work.
The following general duality result gives us a convenient way to prove linear inde-
pendence, especially in vector spaces whose elements are functions, as we will see in
the next section. The proof is an easy and straightforward exercise.

E
Theorem 2.1. Let {v0 , v1 , . . . , vn } be a subset of a vector space that generates a
subspace F E
and let {L 0 , L 1 , . . . , L n } be a subset of the dual space ∗ that generates
G
a subspace . Suppose that

L k , vm  = δk,m , 0 ≤ k ≤ n, 0 ≤ m ≤ n.

Then we have:
F
(i) {v0 , v1 , . . . , vn } is a basis for .
G
(ii) {L 0 , L 1 , . . . , L n } is a basis for .

November 2009] RATIONAL FUNCTIONS 805

 (iii)G is isomorphic to the dual space of F.
(iv) For any f in Fwe have


n
f = L k , f  vk .
k=0

(v) For any φ in G we have


n
φ= φ, vk  L k .
k=0

In part (iii) the isomorphism is the function that maps each element of G to its
F
restriction to the subspace .
The relations of the form L k , vm  = δk,m are called biorthogonality relations. Note
that the biorthogonality relation (2.3) that gives us Taylor’s theorem for polynomials
also gives the following results about the dual space. For any a the set of the restrictions
P P
to n of the Taylor functionals Ta,k , for 0 ≤ k ≤ n, is a basis for n∗ , and for any φ
P
in n∗ we have


n
 
φ= φ, (z − a)k Ta,k .
k=0

It follows that the dual space of every finite-dimensional subspace of the polynomials
is generated by a set of restrictions of Taylor functionals. We can see that not every
P
element of ∗ is a finite linear combination of Taylor functionals as follows. First we
P P
note that, since {z n : n ∈ N} is a basis for , every φ in ∗ is completely determined
by the sequence φ, z n , for n ≥ 0, and conversely, any given sequence of numbers
P
f (n) determines an element μ of ∗ , defined by μ, z n  = f (n), for n ≥ 0. Consider
now a linear functional of the form

s
φ= α j Ta j ,k j ,
j =1

where the a j are complex numbers, the k j are nonnegative integers, and the pairs
(a j , k j ) are distinct. Then


s  
n n−k j
φ, z  =
n
αj a , (2.5)
j =1
kj j

and, as a function of n, each term of the sum in (2.5) is the product of a polynomial
and an exponential function. Therefore, any sequence f (n) that is not given by a sum
like the one in (2.5) can be used to define a functional μ that is not a finite linear
combination of Taylor functionals. For example, it is not difficult to show that the
functional μ defined by μ, z n  = 1/n! for n ≥ 0 is not a linear combination of Taylor
functionals.
One of the main tools for the computation of derivatives is the Leibniz rule

D m ( f g)  m
D k f D m−k g
= . (2.6)
m! k=0
k! (m − k)!

806 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
Since evaluation of polynomials at a point a is a linear and multiplicative map, from
the Leibniz rule we get

 
   
P,
m
Ta,m , pq = Ta,k , p Ta,m−k , q , p, q ∈ (2.7)
k=0

which we call the Leibniz rule for Taylor functionals. Note that it can be seen as a
formula for the computation of the product of p and q, when we represent p, q, and
pq as linear combinations of powers of z − a.  
Suppose that f (z) = (z − a)m g(z), where g is a function such that Ta,k , g is well
defined for 0 ≤ k ≤ m and g(a)  = 0, that is, a is a root of f with multiplicity m. Then,
by (2.3) and the Leibniz rule we have

  n
   n
 
Ta,n , f (z) = Ta,k , (z − a)m Ta,n−k , g(z) = δk,m Ta,n−k , g(z) .
k=0 k=0
     
 Ta,n , f (z) = 0 if n < m and Ta,n , f (z) = Ta,n−m , g(z) if n ≥ m. In par-
Therefore
ticular Ta,m , f (z) = Ta,0 , g(z) = g(a)  = 0. This result will be used in the next sec-
tion.

3. RATIONAL FUNCTIONS. We consider first a very simple rational function. Let

t be a complex number and let f (z) = 1/(t − z). Note that f is not defined at z = t
and it cannot be made continuous at t because the limit of f (z) when z → t does not
exist. This is in constrast with the behaviour of polynomials, which are continuous
everywhere. In fact, they are infinitely differentiable everywhere and their derivatives
of sufficiently high order are identically zero. For the rational function f we have

f (k) (z) 1
= f k+1 (z) = . (3.1)
k! (t − z)k+1
Therefore f is infinitely differentiable at any z  = t and its succesive derivatives never
become zero.
Equation (3.1) allows us to extend the definition of the Taylor functionals to obtain
 
1 1
Ta,k , = , a  = t. (3.2)
t−z (t − a)k+1
Therefore f has the Taylor series

1 ∞
(z − a)m
= , (3.3)
t −z m=0
(t − a)m+1

which is clearly a geometric series that converges for all z such that |z − a| < |t − a|.
We apply the operator D k /k! with respect to z to each term in the series and use
(3.1) to get
∞   ∞  
1 m (z − a)m−k n+k (z − a)n
= = . (3.4)
(t − z)k+1 m=k
k (t − a)m+1 n=0
k (t − a)n+k+1

Since this series is obtained from (3.3) by repeated differentiation, it has the same disk
of convergence. Since (3.4) is the expansion of a negative power of t − z, it is called

November 2009] RATIONAL FUNCTIONS 807

the binomial series. Compare it with the binomial formula (2.4), which has a finite
number of summands.
Q
We introduce next some definitions and notation. Let denote the set of all rational
functions, that is, the quotients of the form p/q, where p, q are polynomials and q is
not identically zero. It is clear that Q is a complex vector space that has additional
algebraic structure because its elements can be multiplied. It is in fact a field, and we
will see that it has an interesting additional structure.
The division algorithm for polynomials says that for any pair of polynomials p and
q, with q not identically zero, there exists a unique pair of polynomials r and s such
that p = sq + r and either r = 0 or the degree of r is strictly less than the degree of
q. Dividing by q we obtain p/q = s + r/q.
R
Let be the set of rational functions r/q with r = 0 or degree of r strictly less than
the degree of q. We will assume that q is monic and that r and q are relatively prime.
We call R the space of proper rational functions. By the division algorithm every
rational function has a unique representation as the sum of a polynomial and a proper
rational function. Since P R Q
and are vector subspaces of we have =Q P R⊕ .
R
Note also that is closed under the usual multiplication of rational functions. Let us
try to find a basis for the vector space .R
Let a be a complex number. Since the set of nonnegative powers of z − a is a basis
P
for , the set Ba = {(z − a)m : m < 0} seems to be a good candidate for a basis of
R . We will show that it is linearly independent. But Ba is not a basis because, by
reduction to a common denominator, we can see that all the elements generated by Ba
have the form p(z)/(z − a)n for some positive n, and this clearly shows that Ba does
not generate .R
Let us show that (z − a)−1 and (z − b)−1 are linearly independent if a  = b. Suppose
that α(z − a)−1 + β(z − b)−1 is identically zero. Then so is its numerator α(z − b) +
β(z − a), and by the linear independence of {1, z} we get the equations α + β = 0 and
αb + βa = 0, which imply α = β = 0, since a  = b.
In a similar way we can prove that Ba ∪ {(z − b)−m } is linearly independent for
R
every positive m, whenever a  = b. This suggests that a basis for should contain all
the sets Ba , for a in C.
We prove next that Ba is linearly independent. Given a finite subset C of Ba , there
exists a nonnegative integer m such that C is contained in the set Ba,m = {(z − a)−k−1 :
0 ≤ k ≤ m}. We will try to find a proof that uses Theorem 2.1 and the biorthogonality
relation (2.3). For this purpose we must pass from negative powers to nonnegative
powers of z − a. This is done by multiplying each element of Ba,m by (z − a)m+1 to
obtain the powers (z − a)k for 0 ≤ k ≤ m. Then we can use (2.3) as follows. For each
R
j such that 0 ≤ j ≤ m, define the linear functional φ j on the subspace of generated
by Ba,m as follows:
   
φ j , f (z) = Ta,m− j , (z − a)m+1 f (z) .

Then by (2.3) we have

   
φ j , (z − a)−k−1 = Ta,m− j , (z − a)m−k = δ j,k , 0 ≤ j ≤ m, 0 ≤ k ≤ m.

Therefore, by Theorem 2.1 we see that Ba,m is linearly independent, and consequently,
Ba is linearly independent.
The idea used in the construction of the functionals φ j can be modified to show that
the union of a finite collection of sets of the form Bai ,m i , with distinct ai ’s, is linearly
independent. This is done in the proof of our next theorem. Let us introduce some
convenient notation first.

808 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 For each complex number a and each nonnegative integer k define

1
ra,k (z) = . (3.5)
(z − a)1+k

Theorem 3.1. The set B = {ra,k : a ∈ C, k ∈ N} is a basis for the space R of proper
rational functions.

Proof. Every finite subset of B is contained in a set of the form

A = Ba1 ,m 1 ∪ Ba2 ,m 2 ∪ · · · ∪ Bas ,m s ,

where the ai are distinct complex numbers, the m i are nonnegative integers, and
Ba j ,m j = {ra j ,k : 0 ≤ k ≤ m j }. Define the index set

J = {(i, j) : 1 ≤ i ≤ s, 0 ≤ j ≤ m i }.

For each (i, j) in J define a linear functional φi, j on the subspace generated by A by
   
φi, j , f (z) = Tai ,m i − j , (z − ai )m i +1 f (z) .

Let (i, j) and (k, n) be elements of J . If k  = i then by the Leibniz rule we have
 
  m i +1 1
φi, j , rak ,n = Tai ,m i − j , (z − ai ) = 0,
(z − ak )1+n

since m i − j < m i + 1.
If k = i then the biorthogonality relation (2.3) gives
 
  (z − ai )m i +1
φi, j , rai ,n = Tai ,m i − j , = δ j,n .
(z − ai )1+n

Therefore
 
φi, j , rak ,n = δ(i, j ),(k,n) , (i, j), (k, n) ∈ J,

and by Theorem 2.1 the set A is linearly independent. Therefore B is linearly indepen-
dent.
It is easy to see that a finite linear combination of elements of B can be written
in the form p/w by reducing to a common denominator, and that the degree of the
numerator p is strictly less than the degree of w. Therefore B generates a subspace of
R .
It remains to show that every proper rational function is in the span of B. Let p/w
R
be in . Then, by the fundamental theorem of algebra, there exist distinct complex
numbers a1 , a2 , . . . , as and nonnegative integers m 1 , m 2 , . . . , m s such that
s
w(z) = (z − ai )1+m i ,
i=1

and if we let N + 1 be the degree of w, then the degree of p is at most N .

November 2009] RATIONAL FUNCTIONS 809

 Let A and J be as previously defined. An element in the span of A has the form

g(z) = αi, j rai , j (z).
(i, j )∈J

By multiplying and dividing by w(z) we get

1  w(z)
g(z) = αi, j . (3.6)
w(z) (i, j )∈J (z − ai )1+ j

Define the polynomials

w(z)
qi, j (z) = , (i, j) ∈ J. (3.7)
(z − ai )1+m i − j

From the definition of w we see that qi, j is a polynomial whose degree is equal to N −
m i + j and has ai as a root with multiplicity j. We will prove that C = {qi, j : (i, j) ∈
P
J } is a basis for the vector space N . Since the index set J has N + 1 elements and
P
the dimension of N is N + 1, it is enough to show that C is linearly independent.
Define the linear functionals L k,n on the space by P
 
 
L k,n , u(z) = Tak ,n ,
u(z)
qk,0 (z)
, (k, n) ∈ J, u ∈ P. (3.8)

Let (i, j) and (k, n) be in J . Then by (3.7) we have

 
  (z − ak )1+m k
L k,n , qi, j (z) = Tak ,n , . (3.9)
(z − ai )1+m i − j

If k  = i then the quotient in (3.9) has ak as a root of multiplicity m k + 1, and by the

Leibniz rule the value of the right-hand side is zero.
If k = i then the quotient in (3.9) becomes (z − ak ) j and hence the value of the
right-hand side is δn, j . Therefore (3.9) gives
 
L k,n , qi, j (z) = δ(k,n),(i, j ) , (k, n), (i, j) ∈ J, (3.10)

P
and by Theorem 2.1 we have that C is a basis for N and {L k,n : (k, n) ∈ J } is its
dual basis.
We conclude that the proper rational function p/w can be written in the form (3.6)
and hence is a finite linear combination of elements of B. This completes the proof of
the theorem.

The rational functions ra,k are called basic rational functions because they are ele-
ments of the basis B.
P
Note that the basis {qi, j : (i, j) ∈ J } of N is a generalized Lagrange interpolation
basis. It is usually called the Lagrange-Sylvester basis. If all the m i are zero then
qi,0 (z) = w(z)/(z − ai ) and L k,0 is evaluation at ak divided by w  (ak ).
The proof of the previous theorem gives us an important property of the proper
rational functions, called the partial fractions decomposition.

810 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
Corollary 3.1. Let p/w be a proper rational function with
s
w(z) = (z − ai )1+m i ,
i=1

where the ai are distinct complex numbers and the m i are nonnegative integers. Let
J = {(i, j) : 1 ≤ i ≤ s, 0 ≤ j ≤ m i }. Then there exists a unique representation

p(z)  βi, j
= , (3.11)
w(z) (i, j )∈J (z − ai )1+m i − j

where the coefficients are given by

 
p(z)(z − ai )1+m i
βi, j = Tai , j , . (3.12)
w(z)

Example. Let p(z) = 2(z + 1) and w(z) = (z − i)(z − 1)2 . We take a1 = i and a2 =
1. Then J = {(1, 0), (2, 0), (2, 1)} and (3.12) gives
 
2(z + 1) 2(i + 1)
β1,0 = Ti,0 , = = i − 1,
(z − 1)2 (i − 1)2

 
2(z + 1) 4
β2,0 = T1,0 , = = 2(1 + i),
z−i 1+i

   
2(z + 1) (z − i) − (z + 1)
β2,1 = T1,1 , = 2 T1,0 , = 1 − i.
z−i (z − i)2

Therefore
2(z + 1) i −1 1−i 2(1 + i)
= + + .
(z − i)(z − 1) 2 z −i z − 1 (z − 1)2

If w is a polynomial with two distinct roots then the partial fractions decomposition
of 1/w gives us the multiplication formula for basic rational functions that we describe
in the next corollary.

Corollary 3.2. Let (a, k) and (b, m) be elements of C × N, with a  = b. Then


k
  
m
 
ra,k rb,m = Ta, j , rb,m ra,k− j + Tb, j , ra,k rb,m− j . (3.13)
j =0 j =0

If we take k = m = 0 then (3.13) becomes

1 1 1 1 1
= + . (3.14)
(z − a)(z − b) (a − b) (z − a) (b − a) (z − b)
Note that (3.13) can be obtained from (3.14) by repeated differentiation with respect
to a, and then with respect to b, using the Leibniz rule.

November 2009] RATIONAL FUNCTIONS 811

Theorem 3.2. The set of Taylor functionals {Ta,k : a ∈ C, k ∈ N} is a linearly inde-
pendent subset of ∗ . P
Proof. Let E be a finite set of Taylor functionals. Then there exist distinct complex
numbers a1 , a2 , . . . , as and nonnegative integers m 1 , m 2 , . . . , m s such that E is con-
tained in the set {Tai , j : 1 ≤ i ≤ s, 0 ≤ j ≤ m i }. Let J = {(i, j) : 1 ≤ i ≤ s, 0 ≤
j ≤ m i }, define
s
w(z) = (z − ai )m i +1 ,
i=1

and let N + 1 be the degree of w. Let qi, j be the polynomials defined in (3.7). Since ai
is not a root of qi,0 , the rational function 1/qi,0 (z) has a Taylor series expansion around
z = ai . Define the Taylor approximation of degree d of 1/qi,0 (z) at z = ai by
d 
 
1
pi,d (z) = Tai , j , (z − ai ) j , d ≥ 0.
j =0
qi,0 (z)

Define the polynomials

Hi, j (z) = qi, j (z) pi,m i − j (z), (i, j) ∈ J. (3.15)

Note that qi,0 (z) is a factor of Hi, j (z) for 0 ≤ j ≤ m i . Note also that all the Hi, j have
degree N .
We will prove the biorthogonality relation
 
Tak ,n , Hi, j (z) = δ(k,n),(i, j ), (k, n), (i, j) ∈ J. (3.16)

It is clear that (3.16) holds if k  = i, since ak is a root of Hi, j (z) with multiplicity
m k + 1.
 now the case k = i in (3.16). Since qi, j (z) = qi,0 (z)(z − ai ) , for  < j
j
Consider
we have Tai , , qi, j (z) = 0 and for  ≥ j we have
   
Tai , , qi, j (z) = Tai ,− j , qi,0 (z) .

If  ≥ j then n −  ≤ n − j ≤ m i − j and the definition of pi,m i − j (z) yields

 
  1
Tai ,n− , pi,m i − j (z) = Tai ,n− , .
qi,0 (z)
Suppose that n ≥ j. By the Leibniz rule and the previous remarks we have

  n
  
Tai ,n , Hi, j (z) = Tai , , qi, j (z) Tai ,n− , pi,m i − j (z)
=0

  
n
  1
= Tai ,− j , qi,0 (z) Tai ,n− ,
= j
qi,0 (z)
 
= Tai ,n− j , 1 = δn, j .

Note that if n < j then all the terms in the above sum are equal to zero. This completes
the proof of (3.16).

812 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 By the biorthogonality relation (3.16) we see that {Tai , j : 1 ≤ i ≤ s, 0 ≤ j ≤ m i }
is linearly independent. Therefore the set of all the Taylor functionals is linearly inde-
P
pendent in ∗ .

The polynomials Hi, j are called the Hermite interpolation polynomials associated
with the multiset of roots of w. They are a basis for the space N . From (3.16) we P
obtain the Hermite interpolation formula
 
u(z) =

Tai , j , u Hi, j (z), u ∈ N. P
(3.17)
(i, j )∈J

Let p/w be a proper rational function and let a1 , a2 , . . . , as be the distinct roots of
w. The coefficient of 1/(z − ai ) in the partial fractions decomposition of p/w is called
the residue of p/w at ai . Residues are important because the basic rational functions
of the form rb,0 are the only ones that are not derivatives of rational functions, and this
fact gives them important properties in connection to contour integration of rational
functions.
The coefficients in (3.13) are
 
  j +m (−1) j
Ta, j , rb,m =
m (a − b)1+m+ j

and
 
  j +k (−1) j
Tb, j , ra,k = .
k (b − a)1+k+ j
   
It is easy to see that Ta,k , rb,m + Tb,m , ra,k = 0. This means that the sum of the
residues of 1/w is zero for every polynomial w with two distinct roots. Note that
this is true for the example given after Corollary 3.1. This result is generalized in our
next theorem.

Theorem 3.3. Let p/w be a proper rational function with the degree of w equal to
N + 1. Then the sum of the residues of p/w is equal to the coefficient of z N in the
representation of p(z) as a linear combination of powers of z.

Proof. We use the notation of Theorem 3.1. The residue at ak of p/w is given by
 
p(z)(z − ak )1+m k
Tak ,m k , .
w(z)

We write p in terms of the polynomials qi, j as follows:


p(z) = αi, j qi, j (z). (3.18)
(i, j )∈J

Then by the definition of the qi, j in (3.7) we have

p(z)(z − ak )1+m k  (z − ak )1+m k

= αi, j ,
w(z) (i, j )∈J
(z − ai )1+m i − j

November 2009] RATIONAL FUNCTIONS 813

and using the Leibniz rule we see that
 
p(z)(z − ak )1+m k
Tak ,m k , = αk,m k ,
w(z)
and therefore k αk,m k is the sum of the residues of p/w. From the definition of the
qi, j we see that the monic polynomials qk,m k have degree equal to N for 1 ≤ k ≤ s, and
all the others have smaller degree. Thus by (3.18) the sum of the αk,m k is the coefficient
of z N in p(z).

Note that if the degree of w minus the degree of p is at least two, then the sum of
the residues of p/w is zero.

Q
4. A BILINEAR FORM ON . Let f (z) be a rational function. Then f (z) =
u(z) + p(z)/w(z) where u is a polynomial and p/w is a proper rational function.
By the partial fractions decomposition (Corollary 3.1),
p(z)  βi, j
= , (4.1)
w(z) (i, j )∈J (z − ai )1+m i − j

where the ai are the distinct roots of w. The principal part of f at a root ak of w is
defined as the sum of the terms in (4.1) that have i = k, and it is denoted by Pak ( f )(z),
that is,

mk
βk, j
Pak ( f )(z) = . (4.2)
j =0
(z − ak )1+m k − j

Note that f is equal to u plus the sum of its principal parts.

Taylor’s theorem for polynomials says that, for any given complex number a, ev-
ery polynomial u(z) can be written in a unique way as a finite linear combination of
nonnegative powers of z − a. For rational functions we have the following generalized
Taylor’s theorem.

Theorem 4.1. Let f (z) = u(z) + p(z)/w(z) be as previously defined and let a be a
complex number. Then we have:
(i) If a is not a root of w then f (z) has a unique representation as a power series
of the form

∞
f (z) = cn (z − a)n , (4.3)
n=0

which converges if |z − a| < min{|a − a j | : 1 ≤ j ≤ s}, and the coefficients
are given by cn = Ta,n , f , for n ≥ 0.
(ii) If a = ak for some k such that 1 ≤ k ≤ s, then f (z) has a unique representation
of the form

∞
f (z) = Pak ( f )(z) + dn (z − ak )n , (4.4)
n=0

which converges if 0 < |z − ak | < min{|ak − a j | : 1 ≤ j ≤ s, j  = k}. For 

nonnegative n the coefficient of (z − ak )n is given by dn = Tak ,n , f − Pak ( f ) .

814 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 For the negative powers of (z − ak ) that appear in Pak ( f )(z) the coefficients
are given by (3.12).

Proof. A pair of suitable sign changes in (3.4) gives us

∞   ∞
1 n + k (−1)n (z − a)n  
= = Ta,n , rt,k (z − a)n , t  = a.
(z − t) k+1
n=0
k (a − t) n+k+1
n=0
(4.5)
This series converges if |z − a| < |a − t|.
If a is not a root of w then each summand in (4.1) has a convergent Taylor series
like (4.5). The polynomial part of f has a finite expansion in terms of powers of z − a.
Therefore f has a convergent Taylor series of the form (4.3). This proves part (i).
If a is a root ak of w then ak is not a root of the denominator of f − Pak ( f ).
Applying part (i) to f − Pak ( f ) we obtain (4.4), and this completes the proof of the
theorem.

The previous theorem says that given a number a and a rational function f , either
f has a Taylor series of the form (4.3) or there is a unique series of the form

∞
f (z) = α j (z − a) j , (4.6)
j =−m−1

called a Laurent series, where m is a nonnegative integer and the series converges in a
set {z : 0 < |z − a| < ρ} for some positive ρ.
From our previous discussion it is clear that the Taylor functionals Ta,k and the basic
rational functions ra,k play a central role in the theory of rational functions. They are
related by (3.2) and (3.3), which we can write as

1 ∞
= ra,m (t) (z − a)m . (4.7)
t −z m=0

We say that 1/(t − z) is a generating function for the sequence ra,m because the co-
efficients in the expansion of 1/(t − z) as a series in powers of z − a are the rational
functions ra,m (t).
By Theorem 3.2 the set of Taylor functionals {Ta,k : (a, k) ∈ C × N} is linearly
P
independent in the dual space of . Therefore the bijective map that sends ra,k to Ta,k ,
for (a, k) in C × N, induces an isomorphism of complex vector spaces between the
R
space of proper rational functions and the space T
of all finite linear combinations
of Taylor functionals. This means that R
can be identified with a subspace of ∗ .P
We will use residues to find an interpretation of the bijective correspondence between
Taylor functionals and basic rational functions. Such an interpretation will allow us to
consider rational functions as elements of the dual space of all rational functions.
Suppose that f (z) is a rational function that has a Taylor series of the form (4.3) for
some number a. Then

∞
ra,k (z) f (z) = cn (z − a)n−k−1 ,
n=0

−1
and in this series ck is the coefficient of (z − a) , that is, ck is the residue of
ra,k (z) f (z) at a. But we also have ck = Ta,k , f . Therefore the map that sends f
to the residue at a of ra,k f coincides with the Taylor functional Ta,k .

November 2009] RATIONAL FUNCTIONS 815

 Since ra,k ra,m = ra,k+m+1 , it is clear that the residue at a of ra,k ra,m is equal to zero
for all nonnegative integers k, m. Therefore the map that sends f to the residue at a
of ra,k f is also defined for functions of the form (4.6) and this means that it is well
defined on the space Q of all rational functions. Note that Ta,k is not defined on the
rational functions ra,m for m ≥ 0. Therefore ra,k induces an extension of Ta,k to all of
Q , and consequently every proper rational function induces a linear functional on . Q
We define
 
ra,k (z), f (z) = Residue at a of ra,k (z) f (z), f ∈ Q. (4.8)

Since this map coincides with Ta,k when f is defined at a we have

 
  n n−k
ra,k (z), z =
n
a , a ∈ C, k, n ∈ N, (4.9a)
k

 
  k k +m 1
ra,k , rb,m = (−1) , a  = b, k, n ∈ N, (4.9b)
k (a − b)1+k+m

and by (4.8) we also have

 
ra,k , ra,m = 0, a ∈ C, k, m ∈ N. (4.9c)
   
From (4.9b) it is easy to see that ra,k , rb,m = − rb,m , ra,k .
Q
Recall that the set {z n : n ∈ N} ∪ {ra,k : (a, k) ∈ C × N} is a basis for . We extend
Q Q
the map  ,  to an antisymmetric bilinear map from × to C as follows. Define
Q
first the bilinear map on the basis of by equations (4.9a), (4.9b), (4.9c), and
 
z n , z k = 0, k, n ∈ N. (4.9d)

The antisymmetry condition

 f, g = − g, f  , f, g in the basis of Q, (4.10)

is consistent with the previous equations and completes the definition of the map on
Q
the basic elements of . Extending the map  ,  to Q
by linearity we obtain an
antisymmetric bilinear form on the space of all the rational functions.
The bilinear form is not positive definite and therefore we cannot introduce geomet-
Q
ric properties in the space . But the bilinear form can be used to study properties of
Q
linear operators on . For example, we can define adjoints or duals of operators with
respect to the bilinear form.
We present some basic properties of this bilinear form. The proofs are simple com-
putations and are omitted.
Let p/w be an element of R Q
and let g be in . Then by the partial fractions de-
composition we have
p  p 
,g = Residue of g at a, (4.11)
w a
w

where the sum runs over the set of all the distinct roots of w(z).

816 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 Let f and g be rational functions. Then

 f, pg =  p f, g , p∈ P, (4.12)
 f, Dg = −D f, g , (4.13)

where D is the differentiation operator,

  n
  
ra,n , f g = ra,k , f ra,n−k , g , (4.14)
k=0

and

 f (z), g(z) =  f (z + a), g(z + a) , a ∈ C. (4.15)

Note that (4.12) means that the operators of multiplication by a polynomial are self-
adjoint and by (4.13) the adjoint of D is −D. Equation (4.14) is the Leibniz rule and
(4.15) is the translation invariance property.
We define the reversion map R from to byQ Q
 
1
R f (z) = f
z
1
z
, f ∈ . Q (4.16)

Note that

Rz n = z −1−n = r0,n (z), n ∈ N,

and
zk
Rra,k (z) = , a ∈ C, k ∈ N. (4.17)
(1 − az)1+k

Note that R is an involution, that is, R 2 = I . We also have  f, g = Rg, R f  and

 f, Rg = −R f, g for any rational functions f and g.
Let p/u be a proper rational function and let q and v be polynomials. Then
p pq r
, qv = ,v = ,v , (4.18)
u u u
where pq = wu + r and the degree of r is strictly smaller than the degree of u.
Let u and v be polynomials and let f be a rational function which is defined on the
roots of v, that is, the roots of v are not poles of f . Then
   
1 1
,vf = , f , (4.19)
uv u

and if, in addition, u and v have no common roots and f is defined on the roots of u,
then
     
1 1 f 1 f
,f = , + , . (4.20)
uv u v v u

These properties are called the reduction and decomposition formulas, respectively.

November 2009] RATIONAL FUNCTIONS 817

 Let w be a polynomial of positive degree. The linear functional that sends f to
1/w, f  is called the divided difference of f with respect to the multiset of roots of w.
The partial fractions decomposition shows that this functional is a linear combination
of Taylor functionals at the roots of w. If all the roots ak of w are simple we get
  
1 s
f (ak )
, f (z) = . (4.21)
w(z) k=0
w  (a )
k

Divided differences are important tools for the solution of polynomial and rational
interpolation problems.
R
If p/w is in then the sequence
 
p(z) n
c(n) = ,z , n ∈ N,
w(x)
is the solution of a linear difference equation with constant coefficients and it is called
a linearly recurrent sequence. The bilinear form is a very useful tool to study such
sequences. See [8].
Let t be a complex number and f be a rational function defined at t. The formulas
 
1
, f (z) = f (t)
z−t
and
 
1 f (k) (t)
, f (z) = , k ≥ 0,
(z − t)k+1 k!
are an algebraic version of the Cauchy integral representation formulas of complex
analysis. The bilinear form can be represented as an integral over a closed path in the
complex plane.

R
5. VECTOR SPACES ISOMORPHIC TO . We consider some vector space iso-
morphisms that allow us to solve linear functional equations that include linear non-
homogeneous differential equations and difference equations. The method of solution
can be seen as a unified shortcut to the traditional transform methods, which include
the Laplace, Mellin, and Z transforms.
From (4.7) we see that
 
1
Ta,m , = ra,m (t), (a, m) ∈ C × N.
t −z
Since the Taylor functionals can be applied to functions that need not be rational, we
can generalize this construction. For example, if instead of 1/(t − z) we use e zt we
obtain the functions
  t m at
ea,m (t) = Ta,m , e zt = e , (a, m) ∈ C × N, (5.1)
m!
E
called basic exponential polynomials. Let be the complex vector space generated
by the set {ea,m (t) : (a, m) ∈ C × N}. It is not difficult to show that this set is linearly
independent in the space of differentiable functions of the complex variable t. The el-
E
ements of are called exponential polynomials, or pseudo-polynomials. It is obvious

818 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 R E
that and are isomorphic as complex vector spaces and that the element of that E
R
corresponds to a given f (z) in is obtained by applying f (z), considered as a linear
combination of Taylor functionals, to e zt .
In the next example we use a discrete variable n instead of t. Define the sequences
 
  n n−m
sa,m (n) = ra,m (z), z =
n
a , n ∈ N, (a, m) ∈ C × N, (5.2)
m
called basic linearly recurrent sequences. They generate the vector space S of linearly
R
recurrent sequences, which is also isomorphic to and

S = { f (z), z  : f ∈ R }.
n

We can construct many other spaces isomorphic to R using the following formal
construction. Suppose that G(z, t) is a complex-valued function defined on C × X for
some set X , and such that for each fixed t and each complex a the function G(z, t) has
a formal Taylor series expansion around z = a, that is, the functions of t defined by
 
ga,m (t) = Ta,m , G(z, t)

are well defined for (a, m) in C × N. Suppose also that the ga,m are linearly indepen-
G R
dent as functions of t. Then they generate a vector space isomorphic to . We say
G
that G(z, t) is the generating function of the space .
Consider the following example. Let G(z, t) = e z log t = t z , where z and t are com-
plex variables and we take the principal branch of log t with imaginary part in [0, 2π).
Then
  (log t)m a
ga,m (t) = Ta,m , G(z, t) = t .
m!
A vector space isomorphic to R may have additional algebraic structure, for ex-
ample, a “natural” multiplication of its elements as functions of t. In the case of the
exponential polynomials the multiplication formula is
 
k+m
ea,k (t)eb,m (t) = ea+b,k+m (t). (5.3)
m
For the linearly recurrent sequences the natural multiplication is
   
n n−k n n−m
sa,k (n)sb,m (n) = a b , (5.4)
k m
called Hadamard multiplication.
Since the multiplications are defined first on the basic elements and then extended
by linearity to the whole space, we can transfer a multiplication from one space to any
other isomorphic vector space using the bijective correspondence between the bases.
R
For example, using (5.3) we define a multiplication  in as follows:
 
k+m
ra,k  rb,m = ra+b,k+m . (5.5)
m
R
This is very different from the natural multiplication in , which is described in (3.13).
The multiplication  is called Hurwitz convolution. It is usually defined by means of
contour integrals in the theory of functions of a complex variable.

November 2009] RATIONAL FUNCTIONS 819

 We can also use the natural multiplication in R to define a new multiplication ∗ in
E . This is done by a direct replacement of the basic elements of R by the correspond-
ing basic elements of E in formula (3.13). In this way we get
 k    m  
j +m (−1) j j +k (−1) j
ea,k ∗ eb,m = e a,k− j + eb,m− j ,
j =0
m (a − b)1+m+ j j =0
k (b − a)1+k+ j
(5.6)

for a  = b, and for the remaining case,

ea,k ∗ ea,m = ea,k+m+1 . (5.7)

The multiplication ∗ is called convolution, or, sometimes, Duhamel convolution. Note

that (5.6) and (5.7) can be used to define the convolution product in any space G
isomorphic to . R
E R
Let : → be the vector space isomorphism defined by ea,k = ra,k for (a, k)
in C × N. Then we have
 
ea,k ∗ eb,m = −1 ea,k eb,m

and
 −1 −1

ra,k  rb,m = ra,k rb,m ,

where the multiplications in the right-hand sides are the natural multiplications in R
E
and , respectively.
Note that, to be able to transfer the Hadamard multiplication (5.4) to some other
isomorphic space, we must find an explicit expression for the right-hand side of (5.4)
S
as a linear combination of basic elements of . This is done in [7] using Newton’s
polynomial interpolation formula. The resulting formula is
d 
 
k+m− j
sa,k sb,m = a m− j bk− j sab,k+m− j ,
j =0
j, k − j, m − j

where d is the minimum of k and m.

The bijective correspondence between bases of isomorphic vector spaces can also
be used to transfer an operator on one of the spaces to a corresponding operator on
E
the other space. For example, if A is a linear operator on then A −1 is the corre-
sponding operator on . R
G R
Let be a vector space isomorphic to with generating function G(z, t) and basis
{ga,m : (a, m) ∈ C × N}. We define a linear operator L : → by G G
Lga,m = a ga,m + ga,m−1 , a ∈ C, m ≥ 1, (5.8)

and

Lga,0 = a ga,0 , a ∈ C. (5.9)

G
We call L the distinguished operator on . If we identify G(z, t) with its formal Taylor
series and write

G(z, t) = ga,m (t) (z − a)m ,
m≥0

820 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
then we can apply L to each one of the coefficients of this series and define

LG(z, t) = Lga,m (t) (z − a)m .
m≥0

Using (5.8) and (5.9) we get

LG(z, t) = aG(z, t) + (z − a)G(z, t) = zG(z, t).

This relation can be used to determine the operator L. For example, for the space E
the distinguished operator is differentiation D, since Dt e zt = z e zt . For the linearly
recurrent sequences S the generating function is G(z, n) = z n and the distinguished
operator is the shift E defined by Eg(n) = g(n + 1), since E z n = z z n . It is easy to
verify that D and E satisfy (5.8) and (5.9).
R
Denote by Sz the distinguished linear operator on , defined by

Sz ra,m (z) = a ra,m (z) + ra,m−1 (z), m ≥ 1, (5.10)

and

Sz ra,0 (z) = a ra,0 (z). (5.11)

Let us show that Sz is just multiplication by z followed by reduction modulo the poly-
nomials. Consider first
1 z−a+a a
z ra,0 (z) = z = =1+ .
z−a z−a z−a
Reduction modulo the polynomials yields (5.11). For m ≥ 1 multiplication of ra,m (z)
by z gives
1 a+z−a a 1
z = = + ,
(z − a) 1+m (z − a) 1+m (z − a) 1+m (z − a)m
and this is equivalent to (5.10). Note that in this case the reduction modulo the poly-
nomials has no effect.
Let w be a polynomial of degree N + 1. Let L be the distinguished operator on the
G G
space and let g be a given element of . Consider the following general problem.
G
Find the set of all f in that satisfy the linear equation w(L) f = g.
If G is the space of exponential polynomials then the equation is w(D) f = g,
and it is a linear differential equation with constant coefficients. For the space S the
equation is a linear difference equation. Since the distinguished operator corresponds
to differentiation on the space of exponential polynomials, the equations w(L) f = g
are called differential-like equations. See [9]. Since the equation w(L) f = g is linear,
G
if we know how to solve it in some space then we know how to solve it in any other
isomorphic space.
In R the equation becomes w(Sz ) f (z) = g(z). Since Sz is multiplication by z
followed by reduction modulo the polynomials, it is easy to see that we must have
w(z) f (z) − g(z) = p(z), for some polynomial p(z). Solving for f we get f (z) =
R
p(z)/w(z) + g(z)/w(z), and since f must be in , p/w must be a proper rational
function. Therefore the set of solutions of w(Sz ) f (z) = g(z) is given by
 
V =
p
w
1
+g : p∈ N .
w
P
November 2009] RATIONAL FUNCTIONS 821
Note that g/w is a particular solution of w(Sz ) f (z) = g(z) and for each p in
P N , the proper rational function p/w is a solution of the homogeneous equation
w(Sz ) f (z) = 0.
Suppose that
s
w(z) = (z − ai )1+m i , (5.12)
i=1

and let

J = {(i, j) : 1 ≤ i ≤ s, 0 ≤ j ≤ m i }. (5.13)

Then
1
g = g ra1 ,m 1 ra2 ,m 2 · · · ras ,m s , (5.14)
w
and by the partial fractions decomposition (Corollary 3.1) every solution of the homo-
geneous equation is of the form
p 
= βi, j rai ,m i − j , (5.15)
w (i, j )∈J

for some complex coefficients βi, j , and thus the set {rai ,m i − j : (i, j) ∈ J } is a basis for
the kernel of w(Sz ).
Note that g/w is given in (5.14) in terms of the natural multiplication of proper
rational functions.
Now we can use the isomorphisms and the convolution product to describe the
solution of the general problem.

G
Corollary 5.1. Let be a vector space with basis {ga,k : (a, k) ∈ C × N}. Let L be
the distinguished linear operator on G
and g be a given element of . Let w be a G
polynomial as in (5.12) and J the index set defined in (5.13). Then the set of solutions
of the equation w(L) f = g is

V = {h + u : u ∈ G },
w

where

h = g ∗ ga1 ,m 1 ∗ ga2 ,m 2 ∗ · · · ∗ gas ,m s ,

∗ denotes the convolution product in the space G , and G w is the vector subspace of G
generated by the set {gai ,m i − j : (i, j) ∈ J }.

In the case of the exponential polynomials this corollary describes the solution sets
of equations of the form w(D) f (t) = g(t), where g is an exponential polynomial. An
example of such an equation is: f (4) (t) − 3 f (2) (t) + f  (t) − 2 f (t) = e−t + cos 2t.
Equations of this type are usually solved by the Laplace transform method.
In addition to the factorization of w, the main computational work required to find
the solution set is the convolution product that produces the particular solution h. This
is done by means of formulas (5.6) and (5.7). It is a straightforward computation that
may be quite laborious if the degree of w is large. The computation of the convolution
product of the ga j ,m j is equivalent to computing the partial fractions decomposition of
1/w.

822 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 Using the integral representation of the convolution we can also solve equations
G
where g is not an exponential polynomial. There are spaces for which the distin-
guished operator L is a differential operator with variable coefficients. See [9].

6. ADDITIONAL ALGEBRAIC STRUCTURE. In this section we describe briefly

an important algebraic structure of the space of proper rational functions, obtained by
duality from the algebraic structure of the space of polynomials.
Let V be a complex vector space and let V ∗ be its dual space. Consider the following
example that shows how the duality relationship can be used to associate to a given
map related to V a corresponding map related to V ∗ .
Let A : V → V be a linear map. It is well known that there is a linear map A :
V → V ∗ , called the adjoint of A, such that
∗

  
A f, x =  f, Ax , f ∈ V ∗ , x ∈ V.

If V is finite-dimensional, taking suitable bases we see that the matrix representation

of A is the transpose of the matrix representation of A.
The construction of adjoints can be applied to a commutative diagram and yields an-
other commutative diagram where the arrows are reversed and the spaces are replaced
by their duals.
If V has some additional algebraic structure, for example, an associative multipli-
cation, then the vector space duality can be used to construct a corresponding object
in V ∗ . A multiplication on V is a map μ : V ⊗ V → V . Using duality we can obtain
a map : V ∗ → (V ⊗ V )∗ .
If V is finite-dimensional then so is V ∗ and there exists a natural isomorphism
between (V ⊗ V )∗ and V ∗ ⊗ V ∗ . Thus we can consider as a map from V ∗ to V ∗ ⊗
V ∗ which is related to μ by duality and is called a comultiplication.
If V is infinite-dimensional then (V ⊗ V )∗ may not be isomorphic to V ∗ ⊗ V ∗ and
in this case the comultiplication cannot be defined on the whole dual V ∗ , but it is
defined on some suitable subspace.
Q R T
The bilinear map on allows us to identify with the subspace , generated by
the Taylor functionals, of the vector space dual of the polynomials. It turns out that R
P
is a suitable subspace of ∗ in the sense that we can define an algebraic structure on
R P
induced by duality by the structure of . The space of polynomials is not only a
commutative algebra over C, it is also a bialgebra with an involution that makes it a
Hopf algebra. We will define these terms below.
Let C be a complex vector space and let i denote the identity map on C. A linear
map : C → C ⊗ C is a comultiplication if it satisfies the coassociativity property

(i ⊗ )◦ =( ⊗ i) ◦ . (6.1)

A linear map : C → C is a counit if ( ⊗ i) ◦ is the natural isomorphism from

C to C ⊗ C and (i ⊗ ) ◦ is the natural isomorphism from C to C ⊗ C. A space C
with a coassociative comultiplication and a counit is called a coalgebra.
In the space Pthe substitution map that sends p(z) to p(t + z) and the binomial
formula for (t + z)n motivate the definition of the map on the basic elements of P
by
 n  
n j
zn = z ⊗ z n− j , n ≥ 0. (6.2)
j =0
j

It is easy to see that is a comultiplication on P . The counit is defined by p = p(0).

November 2009] RATIONAL FUNCTIONS 823

 Let C ∗ be the dual space of a coalgebra C. The comultiplication induces a mul-
tiplication in C ∗ as follows:

αβ, f  = α ⊗ β, f, α, β ∈ C ∗ , f ∈ C. (6.3)

The adjoint of is a unit on C ∗ . Therefore C ∗ is an associative algebra with unit

element. In an analogous way, if we start with a finite-dimensional algebra A then,
using duality, we can give A∗ a coalgebra structure. This is not always true if A is
infinite dimensional. See [1, 4].
A space H that has both an algebra and a coalgebra structure is called a bialgebra
if and are algebra morphisms. This means that if μ and η are the multiplication
and the unit of H and i is the identity map on the field then ( f g) = f g, ◦ η =
(η ⊗ η) ◦ i, i ◦ ( ⊗ ) = ◦ μ, and ◦ η = i. These properties also mean that μ and
η are morphisms of coalgebras. A bialgebra that has an endomorphism S compatible
in a certain sense with the bialgebra structure is a Hopf algebra. See [4]. S is called the
antipode map.
The space P is a simple and important example of a Hopf algebra. The multiplica-
tion is the usual one

z j z k = z j +k , j, k ∈ N, (6.4)

and the unit element is z 0 = 1. The comultiplication and the counit were defined pre-
viously. The antipode is

Sz k = (−z)k , k ∈ N. (6.5)

Using the Chu-Vandermonde convolution formula it is easy to show that is an alge-

bra map.
If H is a Hopf algebra we define

H ◦ = {α ∈ H ∗ : μ∗ (α) ∈ H ∗ ⊗ H ∗ }, (6.6)

where μ is the multiplication in H and μ∗ is the comultiplication in H ∗ , the full lin-

ear dual of H . It is easy to see that H ◦ is a subalgebra of H ∗ , μ∗ (H ◦ ) ⊂ H ◦ ⊗ H ◦ ,
∗
(H ◦ ⊗ H ◦ ) ⊂ H ◦ , and S ∗ (H ◦ ) ⊂ H ◦ . Therefore H ◦ is a Hopf algebra and it is
called the dual Hopf algebra of H , or the continuous dual of H . The following propo-
sition is an important characterization of the dual Hopf algebra. For a proof see [1,
Theorem 2.2.12].

Proposition 6.1. H ◦ is the set of elements of H ∗ whose kernel contains an ideal of H

of finite codimension.

Recall that we can identify R P

with a subspace of ∗ . The Hopf algebra structure
ofP R R
induces by duality a structure on that makes isomorphic to the Hopf dual
of the polynomials.
R
Leibniz’s rule (4.14) gives us the comultiplication on , which we denote by ,


n
ra,n = ra,k ⊗ ra,n−k , a ∈ C, n ∈ N. (6.7)
k=0

824 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
The counit, denoted by , is given by
 
ra,n = ra,n , 1 = δ0,n , a ∈ C, n ∈ N. (6.8)

The multiplication on R , denoted by , is determined as follows:

n  
   n   
ra,k  rb,m , z =
n
ra,k , z j rb,m , z n− j
j =0
j
 n     
n j j −k n − j n−m− j
= a b
j =0
j k m
  
k+m n
= (a + b)n−k−m ,
k k+m
and therefore
 
k+m
ra,k  rb,m = ra+b,k+m . (6.9)
k
Notice that this is the Hurwitz convolution of equation (5.5). The rational function r0,0
is the unit for this multiplication.
For the antipode we have
 
    n n−k
ra,k (z), Sz n = ra,k (z), (−1)n z n = (−1)n a ,
k
and thus

S ∗ra,k = (−1)k r−a,k . (6.10)

R
It is easy to see that these definitions give the structure of a Hopf algebra. The next
R
theorem shows that is the Hopf algebra dual of . P
Theorem 6.1. R is the set of elements of P ∗
whose kernel contains an ideal of P of
finite codimension.

Proof. All the ideals of P P

are of the form w , where w is a polynomial. By the
P P
division algorithm for any nonzero w the quotient /w has dimension equal to the
degree of w. Thus any nonzero ideal of P
has finite codimension.
R
Let p/w be an element of . By (4.19) and (4.9d) we have
p
w
, wv =  p, v = 0, v∈ P,
P
and therefore the ideal w is contained in the kernel of p/w.
P
Let α be an element of ∗ such that the kernel of α contains the ideal w , whereP
w is a polynomial of degree N + 1. Let p be a polynomial. By the division algorithm
P
p = qw + u where u is a polynomial in N . Then αp = αu. Therefore, since = P
P P P
w ⊕ N , the functional α is determined by its restriction to N , call it β. From
the biorthogonality relation (3.16) we have that {Tak ,n : (k, n) ∈ J } is a basis for the
P
dual space of N . Since these basis elements are Taylor functionals associated with
the roots of w, β is a linear combination of such functionals and thus corresponds to a
proper rational function of the form p/w.

November 2009] RATIONAL FUNCTIONS 825

 R
If we look again at the definition of the Hopf algebra structure of , we can see
that all the maps that define the structure are defined in terms of the indices a and k of
the basic elements ra,k . This suggests the following definition.
Let B be the free complex vector space generated by the set C × N. Define the
multiplication  by
 
k+m
(a, k)  (b, m) = (a + b, k + m),
k
the comultiplication by

n
(a, n) = (a, k) ⊗ (a, n − k),
k=0

the counit  by

(a, n) = δ0,n ,

and the antipode

S ∗ (a, n) = (−1)n (−a, n).

B R B
It is clear that is a Hopf algebra isomorphic to . We call the binomial Hopf
algebra over the complex numbers. Note that the vector subspace of B
generated by
P
the elements of the form (0, k), for k in N, is a Hopf algebra isomorphic to . See
[7] for a more detailed discussion about the binomial Hopf algebra and its connection
with transform methods.

R
7. FINAL REMARKS. The space can be seen as the direct sum of the subspaces
R a = {r a,k : k ∈ N}, where a runs over the complex numbers. Note that for each a,
R Pa and are isomorphic as vector spaces. This means that, as a vector space, R is
P
much larger than . On the other hand, R
can be identified with a proper subset of
P P × .
The complexity of the natural multiplication of basic rational functions, expressed
in (3.13), and of the convolution of exponential polynomials, given in (5.6), is es-
sentially a consequence of the Leibniz rule. It seems that such complexity cannot be
reduced by using other bases for the spaces.
Note that the Leibniz rule is also involved in the computation of the coefficients of
partial fractions decompositions, and in the comultiplication of the dual of the polyno-
mials.
From the point of view of the notation, the need to use pairs (i, j) as indices is
the main source of complexity in some of the propositions of this paper. This need is a
direct consequence of the fact that the roots of a polynomial must be seen as a multiset,
because each root has a positive multiplicity.
Rational functions are useful for solving combinatorial problems using the method
of generating functions. See [6].
Linear functionals on spaces of polynomials or rational functions are very important
in numerical analysis and related fields, but the algebraic structure of the spaces of
linear functionals has rarely been used in those fields. See [2].
The bilinear form on Q gives a very direct connection between rational functions
and linearly recurrent sequences. This is used in [8] to obtain numerous properties of
linearly recurrent sequences.

826 
c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
 The algebraic structure of the linearly recurrent sequences has been studied in [3]
and [5].

REFERENCES

1. E. Abe, Hopf Algebras, Cambridge University Press, Cambridge, 1980.

2. C. Brezinski and P. Maroni, The algebra of linear functionals on polynomials, with applications to Padé
approximation, Numer. Algorithms 11 (1996) 25–33. doi:10.1007/BF02142486
3. L. Cerlienco, M. Mignotte, and F. Piras, Suites récurrentes linéaires, Enseign. Math. 33 (1987) 67–108.
4. C. Kassel, Quantum Groups, Springer-Verlag, New York, 1995.
5. B. Peterson and E. J. Taft, The Hopf algebra of linearly recursive sequences, Aequationes Math. 20 (1980)
1–17. doi:10.1007/BF02190488
6. R. P. Stanley, Enumerative Combinatorics, vol. I, Wadsworth and Brooks/Cole, Monterey, CA, 1986.
7. L. Verde-Star, An algebraic approach to convolutions and transform methods, Adv. in Appl. Math. 19
(1997) 117–143. doi:10.1006/aama.1997.0530
8. , Sections and lacunary sums of linearly recurrent sequences, Adv. in Appl. Math. 23 (1999) 176–
197. doi:10.1006/aama.1998.0640
9. , Solving linear differential-like equations, this M ONTHLY 107 (2000) 213–226. doi:10.2307/
2589314

LUIS VERDE-STAR received his Ph.D. from the University of Wisconsin-Madison in 1977. His mathe-
matical interests include linear algebra, interpolation and approximation, combinatorics, and linear functional
equations.
Universidad Autónoma Metropolitana-Iztapalapa, Apartado 55–534, México D.F. 09340, México
verde@xanum.uam.mx

From the Letters of Lady Mary Wortley Montagu

“I beleive there are few heads capable of makeing Sir I[saac] Newton’s calcu-
lations, but the result of them is not difficult to be understood by a moderate
capacity.”
Lady Mary Wortley Montagu, letter to her daughter, Lady Bute, Jan. 28, 1753,
in R. Halsband, ed., The Complete Letters of Lady Mary Wortley Montagu,
Vol. III: 1752–1762, Oxford University Press, London, 1967, p. 23.

—Submitted by Robert Haas, Cleveland Heights, OH

November 2009] RATIONAL FUNCTIONS 827

View publication stats