Week 9-10: Recurrence Relations and Generating

Functions

April 15, 2019

1 Some number sequences

An infinite sequence (or just a sequence for short) is an ordered array

a0 , a 1 , a 2 , . . . , a n , . . .
of countably many real or complex numbers, and is usually abbreviated as
(an; n ≥ 0) or just (an). A sequence (an) can be viewed as a function f from
the set of nonnegative integers to the set of real or complex numbers, i.e.,
f : Z≥0 → C, f (n) = an, n = 0, 1, 2, . . .
We call a sequence (an) an arithmetic sequence if it is of the form
a0, a0 + q, a0 + 2q, ..., a0 + nq, ...
The general term of such sequences satisfies the recurrence relation
an = an−1 + q, n ≥ 1.
A sequence (an) is called a geometric sequence if it is of the form
a0 , a0q, a0 q 2 , ..., a0 q n , ...
The general term of such sequences satisfies the recurrence relation
an = qan−1, n ≥ 1.

1
The partial sums of a sequence (an) are the summations:
s0 = a0,
s1 = a0 + a1,
s2 = a0 + a1 + a2,
...
sn = a0 + a1 + · · · + an,
...

The partial sums form a new sequence (sn; n ≥ 0).

For an arithmetic sequence an = a0 + nq (n ≥ 0), we have the partial sum
n
X qn(n + 1)
sn = (a0 + kq) = (n + 1)a0 + .
2
k=0

For a geometric sequence an = a0q n (n ≥ 1), we have

n
( qn+1−1
X q−1 a0 if q =
6 1
sn = a0 q n =
k=0 (n + 1)a0 if q = 1.
Example 1.1. Determine the sequence an, defined as the number of regions
which are created by n mutually overlapping circles in general position on the
plane. (By mutually overlapping we mean that each pair of two circles
intersect in two distinct points. Thus non-intersecting or tangent circles are not
allowed. By general position we mean that there are no three circles through
a common point.)
We easily see that the first few numbers are given as
a0 = 1, a1 = 2, a2 = 4, a3 = 8.
It seems that we might have a4 = 16. However, by try-and-error we quickly see
that a4 = 14.
Assume that there are n circles in general position on a plane. When we take
one circle away, say the nth circle, there are n − 1 circles in general position
on the same plane. By induction hypothesis the n − 1 circles divide the plane
2
into an−1 regions. Note that the nth circle intersects each of the n − 1 circles in
2(n − 1) distinct points. Let the 2(n − 1) points on the nth circle be ordered
clockwise as P1, P2, . . . , P2(n−1). Then each of the 2(n − 1) arcs
P1 P2 , P 2 P3 , P 3 P4 , ..., P2(n−2)+1P2(n−1), P2(n−1)P1
intersects one and only one region in the case n − 1 circles and separate the
region into two regions. There are 2(n − 1) more regions produced when the nth
circle is drawn. We thus obtain the recurrence relation
an = an−1 + 2(n − 1), n ≥ 2.
Repeating the recurrence relation we have
an = an−1 + 2(n − 1)
= an−2 + 2(n − 1) + 2(n − 2)
= an−3 + 2(n − 1) + 2(n − 2) + 2(n − 3)
...
= a1 + 2(n − 1) + 2(n − 2) + 2(n − 3) + · · · + 2
(n − 1)n
= a1 + 2 ·
2
= 2 + n(n − 1)
= n2 − n + 2, n ≥ 2.
This formula is also valid for n = 1 (since a1 = 2), although it doesn’t hold for
n = 0 (since a0 = 1).
Example 1.2 (Fibonacci Sequence). A pair of newly born rabbits of opposite
sexes is placed in an enclosure at the beginning of a year. Baby rabbits need
one moth to grow mature; they become an adult pair on the first day of the
second month. Beginning with the second month the female is pregnant, and
gives exactly one birth of one pair of rabbits of opposite sexes on the first day of
the third month, and gives exactly one such birth on the first day of each next
month. Each new pair also gives such birth to a pair of rabbits on the first day
of each month starting from the third month (from its birth). Find the number
of pairs of rabbits in the enclosure after one year?
3
 Let fn denote the number of pairs of rabbits on the first day of the nth month.
Some of these pairs are adult and some are babies. Let an and bn denote the
number of pairs of adult and baby rabbits respectively on the first day of the
nth month. Then the total number of pairs of rabbits on the first day of the nth
month is fn = an + bn, n ≥ 1.
n 1 2 3 4 5 6 7 8 9 10 11 12 13
an 0 1 1 2 3 5 8 13 21 34 55 89 144
bn 1 0 1 1 2 3 5 8 13 21 34 55 89
fn 1 1 2 3 5 8 13 21 34 55 89 144 233
If a pair is adult on the first day of the nth month, then it gives one birth of one
pair on the first day of the next month. So bn+1 = an, n ≥ 1. Note that each
adult pair on the first day of the nth month is still an adult pair on first day of
the next month and each baby pair on the first day of the nth month becomes
an adult pair on the first day of the next month, we have an+1 = an + bn = fn,
n ≥ 1. Thus
fn = an + bn = fn−1 + an−1 = fn−1 + fn−2, n ≥ 3.
Let us define f0 = 0. The sequence f0, f1, f2, f3, . . . satisfying the recurrence
relation 
 fn = fn−1 + fn−2, n ≥ 2
f = 0 (1)
 0
f1 = 1
is known as the Fibonacci sequence, and its terms are known as Fibonacci
numbers.
Example 1.3. The partial sum of Fibonacci sequence is
sn = f0 + f1 + f2 + · · · + fn = fn+2 − 1. (2)
This can be verified by induction on n. For n = 0, we have s0 = f2 − 1 = 0.
Now for n ≥ 1, we assume that it is true for n − 1, i.e., sn−1 = fn+1 − 1. Then
sn = f0 + f1 + · · · + fn
= sn−1 + fn
= fn+1 − 1 + fn (by the induction hypothesis)
= fn+2 − 1. (by the Fibonacci recurrence)
4
Example 1.4. The Fibonacci number fn is even if and only if n is a multiple
of 3.
Note that f1 = f2 = 1 is odd and f3 = 2 is even. Assume that f3k is even,
f3k−2 and f3k−1 are odd. Then f3k+1 = f3k + f3k−1 is odd (even + odd = odd),
and subsequently, f3k+2 = f3k+1 + f3k is also odd (odd + even = odd). It follows
that f3(k+1) = f3k+2 + f3k+1 is even (odd + odd = even).
Theorem 1.1. The general term of the Fibonacci sequence (fn) is given by
à √ !n à √ !n
1 1+ 5 1 1− 5
fn = √ −√ , n ≥ 0. (3)
5 2 5 2

Example 1.5. Determine the number hn of ways to perfectly cover a 2-by-n

board with dominoes. (Symmetries are not counted in counting the number of
coverings.)
We assume h0 = 1 since a 2-by-0 board is empty and it has exactly one
perfect cover, namely, the empty cover. Note that the first few terms can be
easily obtained such as
h0 = 1, h1 = 1, h2 = 2, h3 = 3, h4 = 5.
Now for n ≥ 3, the 2-by-n board can be covered by dominoes in two types:

1 2 n−1 n 1 2 n−2 n

There are hn−1 ways in the first type and hn−2 ways in the second type. Thus
hn = hn−1 + hn−2, n ≥ 2.
Therefore the sequence (hn; n ≥ 0) is the Fibonacci sequence (fn; n ≥ 0) with
f0 = 0 deleted, i.e.,
hn = fn+1, n ≥ 0.
Example 1.6. Determine the number bn of ways to perfectly cover a 1-by-n
board by dominoes and monominoes.

5
 bn = bn−1 + bn−2, b0 = b1 = 1, b2 = 2.
Theorem 1.2. The Fibonacci number fn can be written as
bX2 cµ
n−1
¶
n−k−1
fn = , n ≥ 0.
k
k=0

Proof. Let g0 = 0 and

bX2 cµ
n−1
¶
n−k−1
gn = , n ≥ 1.
k
k=0
¥ n−1 ¦ ³ ´
m
Note that k > 2 is equivalent to k > n − k − 1. Since p = 0 for any
integers m and p such that p > m, we may write gn as
n−1 µ
X ¶
n−k−1
gn = , n ≥ 1.
k
k=0

To prove the theorem, it suffices to show that the sequence (gn) satisfies the
Fibonacci recurrence relation with the same initial values. In fact, g0 = 0,

6
 ¡0¢
g1 = 0 = 1, and for n ≥ 0,
Xn µ ¶ X n−1 µ ¶
n−k n−k−1
gn+1 + gn = +
k k
k=0 k=0
³n´ X n µ ¶ X n µ ¶
n−k n−k
= + +
0 k k−1
k=1 k=1
³n´ X n ·µ ¶ µ ¶¸
n−k n−k
= + +
0 k k−1
k=1
³n´ X n µ ¶
n−k+1
= + (By the Pascal formula)
0 k
k=1
µ ¶ X n µ ¶ µ ¶
n+1 n−k+1 0
= + +
0 k n+1
k=1
n+1 µ
X ¶
(n + 2) − k − 1
= = gn+2.
k
k=0

We conclude that the sequence (gn) is the Fibonacci sequence (fn).

2 Linear recurrence relations

Definition 2.1. A sequence (xn; n ≥ 0) of numbers is said to satisfy a linear

recurrence relation of order k if
xn = α1(n)xn−1 + α2(n)xn−2 + · · · + αk (n)xn−k + βn, (4)
where αk (n) 6= 0, n ≥ k, and the coefficients α1(n), α2(n), . . ., αk (n) and βn are
functions of n. The linear recurrence relation (4) is said to be homogeneous
if βn = 0 for all n ≥ k, and is said to have constant coefficients if α1(n),
α2(n), . . ., αk (n) are constants. The recurrence relation
xn = α1(n)xn−1 + α2(n)xn−2 + · · · + αk (n)xn−k , (5)
where αk (n) 6= 0, n ≥ k, is called the corresponding homogeneous linear
recurrence relation of (4).
7
 A solution of the recurrence relation (4) is a sequence (un) satisfying (4).
The general solution of (4) is a solution
xn = un(c1, c2, . . . , ck ) (6)
with some parameters c1, c2, . . . , ck , provided that for arbitrary initial values
u0, u1, . . . , uk−1 there exist constants c1, c2, . . . , ck such that (6) is the unique
sequence satisfying both the recurrence relation (4) and the initial conditions.
Let S∞ denote the set of all sequences (an; n ≥ 0). Then S∞ is an infinite-
dimensional vector space under the ordinary addition and scalar multiplication
of sequences. Let Nk be the set all solutions of the nonhomogeneous linear
recurrence relation (4), and Hk the set of all solutions of the homogeneous linear
recurrence relation (5). We shall see that Hk is a k-dimensional subspace of S∞,
and Nk a k-dimensional affine subspace of S∞.
Theorem 2.2. (Structure Theorem for Linear Recurrence Relations)
(a) The solution space Hk is a k-dimensional subspace of the vector space
S∞ of sequences. Thus, if (an,1), (an,2), . . ., (an,k ) are linearly independent
solutions of the homogeneous linear recurrence relation (5), then the general
solution of (5) is
xn = c1an,1 + c2an,2 + · · · + ck an,k , n ≥ 0,
where c1, c2, . . . , ck are arbitrary constants.
(b) Let (an) be a particular solution of the nonhomogeneous linear recur-
rence relation (4). Then the general solution of (4) is
xn = an + hn, n ≥ 0,
where (hn) is the general solution of the corresponding homogeneous linear
recurrence relation (5). In other words, Nk is a translate of Hk in S∞, i.e.,
Nk = (an) + Hk .
Proof. (a) To show that Hk is a vector subspace of S∞, we need to show that
Hk is closed under the addition and scalar multiplication of sequences. Let (an)

8
and (bn) be solutions of (5). Then
an + bn = [α1(n)an−1 + α2(n)an−2 + · · · + αk (n)an−k ]
+[α1(n)bn−1 + α2(n)bn−2 + · · · + αk (n)bn−k ]
= α1(n)(an−1 + bn−1) + α2(n)(an−2 + bn−2) +
· · · + αk (n)(an−k + bn−k ), n ≥ k.
For scalars c,
can = c[α1(n)an−1 + α2(n)an−2 + · · · + αk (n)an−k ]
= α1(n)can−1 + α2(n)can−2 + · · · + αk (n)can−k , n ≥ k.
This means that Hk is closed under the addition and scalar multiplication of
sequences.
To show that Hk is k-dimensional, consider the projection π : S∞ → Rk ,
defined by
π(x0, x1, x2, . . .) = (x0, x1, . . . , xk−1).
We shall see that the restriction π|Hk : Hk → Rk is a linear isomorphism. For
each (a0, a1, . . . , ak−1) ∈ Rk , define an inductively by
an = α1(n)an−1 + α2(n)an−2 + · · · + αk (n)an−k , n ≥ k.
Obviously, we have π(a0, a1, a2, . . .) = (a0, a1, . . . , ak−1). This means that π|Hk
is surjective. Now let (un) ∈ Hk be such that π(u0, u1, u2, . . .) = (0, 0, . . . , 0),
i.e.,
u0 = u1 = · · · = uk−1 = 0.
Applying the recurrence relation (5) for n = k, we have uk = 0. Again applying
(5) for n = k + 1, we obtain uk+1 = 0. Continuing to apply (5), we have un = 0
for n ≥ k. Thus (un) is the zero sequence. This means that π is injective. We
have finished the proof that π|Hk is a linear isomorphism from Hk onto Rk . So
dim Hk = k.
(b) For each solution (vn) of (4), we claim that the sequence hn := vn − un
(n ≥ 0) is a solution of (5). So
vn = un + hn, n ≥ 0.

9
In fact, since (un) and (vn) are solutions of (4), applying the recurrence relation
(4), we have
hn = [α1(n)vn−1 + α2(n)vn−2 + · · · + αk (n)vn−k + βn]
−[α1(n)un−1 + α2(n)un−2 + · · · + αk (n)un−k + βn]
= α1(n)(vn−1 − un−1) + α2(n)(vn−2 − un−2) + · · · + αk (n)(vn−k − un−k )
= α1(n)hn−1 + α2(n)hn−2 + · · · + αk (n)hn−k , n ≥ k.
This means that (hn) is a solution of (5). Conversely, for any solution (hn) of
(5), we have
un + hn = [α1(n)un−1 + α2(n)un−2 + · · · + αk (n)un−k + βn]
+[α1(n)hn−1 + α2(n)hn−2 + · · · + αk (n)hn−k ]
= α1(n)(un−1 + hn−1) + α2(n)(un−2 + hn−2)
+ · · · + αk (n)(un−k + hn−k ) + βn
for n ≥ k. This means that the sequence (un + hn) is a solution of (4).

Definition 2.3. The Wronskian Wk (n) of k solutions (un,1), (un,2), . . .,

(un,k ) of the homogeneous linear recurrence relation (5) is the determinant
 
un,1 un,2 · · · un,k
 
 un+1,1 un+1,2 · · · un+1,k 
Wk (n) = det  .. ... ...  , n ≥ 0.
 . 
un+k−1,1 un+k−1,2 · · · un+j−1,k
The Wronskian of (un,1), (un,2), . . ., (un,k ) is also an infinite sequence
(Wk (n); n ≥ 0).
Proposition 2.4. If Wk (m) = 0, then Wk (n) = 0 for all n ≥ m.
Proof. Since Wk (m) = 0, there are constants c1, c2, . . . , ck , not all zero, such
that for i = 0, 1, . . . , k − 1,
k
X
cj um+i,j = c1um+i,1 + c2um+i,2 + · · · + ck um+i,k = 0.
j=1

10
We claim that Wk (m+1) = 0. It is enough to show that for the same coefficients
c1, c2, . . . , ck above, the linear relation
k
X
cj um+k,j = c1um+k,1 + c2ui,2 + · · · + ck um+k,k = 0
j=1

holds. In fact, applying the recurrence relation (5) to (um+k,1), (um+k,2), . . .,

(um+k,k ) respectively, we have
k
X k
X k−1
X
cj um+k,j = cj αi(m + k)um+i,j
j=1 j=1 i=0
k−1
X k
X
= αi(m + k) cj um+i,j = 0.
i=0 j=1

This means that the vectors (um+i,1), (um+i,2), . . ., (um+i,k ), with 1 ≤ i ≤ k,

are linearly dependent. So Wk (m + 1) = 0. Continue the procedure, we see that
Wk (n) = 0 for n ≥ m.
Theorem 2.5. The solutions (un,1), (un,2), . . ., (un,k ) of the homogeneous
linear recurrence relation (5) are linearly independent if and only if there is
a nonnegative integer n0 such that the Wronskian
Wk (n0) 6= 0.
In other words, the solutions (un,1), (un,2), . . ., (un,k ) are linearly dependent
if and only if Wk (n) = 0 for all n ≥ 0.
Proof. We show that the sequences (un,1), (un,2), . . ., (un,k ) are linearly depen-
dent if and only if Wk (n) = 0 for all n ≥ 0. If (un,1), (un,2), . . ., (un,k ) are
linearly dependent, then for n ≥ 0 the columns of the matrix
 
un,1 un,2 ··· un,k
 
 un+1,1 un+1,2 · · · un+1,k 
 ... ... ... 
 
un+k−1,1 un+k−1,2 · · · un+j−1,k

11
are linearly dependent because the columns are part of the sequences (un,1),
(un,2), . . ., (un,k ) respectively. It follows from linear algebra that the determinant
of the matrix is zero, i.e., the Wronskian Wk (n) = 0 for all n ≥ 0.
Conversely, if Wk (n) = 0 for all n ≥ 0, in particular, Wk (0) = 0, then there
are constants c1, c2, . . . , ck , not all zero, such that
k
X
cj ui,j = c1ui,1 + c2ui,2 + · · · + ck ui,k = 0, i = 0, 1, . . . , k − 1.
j=1
P
We claim that kj=1 cj un,j = 0 for all n by induction. Assume it is true for
n ≤ m − 1. Applying the recurrence relation (5) to the sequences (um,1), (um,2),
. . ., (um,k ), we have
k
X k
X k
X
cj um,j = cj αi(k)um−i,j
j=1 j=1 i=1
k
X k
X
= αi(k) cj um−i,j = 0.
i=1 j=1

This means that the sequences (un,1), (un,2), . . ., (un,k ) are linearly dependent.

3 Homogeneous linear recurrence relations with constant coeffi-

cients

In this section we only consider linear recurrence relations of the form

xn = α1xn−1 + α2xn−2 + · · · + αk xn−k , αk 6= 0, n ≥ k, (7)
where α1, α2, . . ., αk are constants. We call this kinds of recurrence relations as
homogeneous linear recurrence relations of order k with constant
coefficients. Sometimes it is convenient to write (7) as of the form
α0xn + α1xn−1 + · · · + αk xn−k = 0, n≥k (8)
where α0 6= 0 and αk 6= 0. The following polynomial equation
α0tk + α1tk−1 + · · · + αk−1t + αk = 0, (9)
12
is called the characteristic equation associated with the recurrence relation
(8), and the polynomial on the left side of (9) is called the characteristic
polynomial.
Example 3.1. The Fibonacci sequence (fn; n ≥ 0) satisfies the linear recur-
rence relation
fn = fn−1 + fn−2, n ≥ 2
of order 2 with α1 = α2 = 1 in (7).
Example 3.2. The geometric sequence (xn; n ≥ 0), where xn = q n, satisfies
the linear recurrence relation
xn = qxn−1, n≥1
of order 1 with α1 = q in (7).
It is quite heuristic that solutions of the first order homogeneous linear recur-
rence relations are geometric sequences. This hints that the recurrence relation
(7) may have solutions of the form xn = q n. The following theorem confirms the
speculation.
Theorem 3.1. (a) For any number q 6= 0, the geometric sequence
xn = q n
is a solution of the kth order homogeneous linear recurrence relation (8) with
constant coefficients if and only if the number q is a root of the characteristic
equation (9).
(b) If the characteristic equation (9) has k distinct roots q1, q2, . . ., qk ,
then the general solution of (8) is
xn = c1q1n + c2q2n + · · · + ck qkn, n ≥ 0. (10)
Proof. (a) Put xn = q n into the recurrence relation (8). We obtain
α0q n + α1q n−1 + · · · + αk q n−k = 0. (11)
Since q 6= 0, dividing both sides of (11) by q n−k , we have
α0q k + α1q k−1 + · · · + αk−1q + αk = 0 (12)
13
This means that (11) and (12) are equivalent. This finishes the proof of Part
(a).
(b) Since q1, q2, . . . , qk are roots of the characteristic equation (9), xn = qin are
solutions of the homogeneous linear recurrence relation (8) for all i (1 ≤ i ≤ k).
Since the solution space of (8) is a vector space, the linear combination
xn = c1q1n + c2q2n + · · · + ck qkn, n≥0
are also solutions (8). Now given arbitrary values for x0, x1, . . . , xk−1, the se-
quence (xn) is uniquely determined by the recurrence relation (8). Set
c1q1i + c2q2i + · · · + ck qki = xi, 0 ≤ i ≤ k − 1.
The coefficients c1, c2, . . . , ck are uniquely determined by Cramer’s rule as fol-
lows:
det Ai
ci = , 1≤i≤k
det A
where A is the Vandermonde matrix
 
1 1 ··· 1
 q q2 · · · qk 
 1 
 2 2 2 
A =  q1 q2 · · · q k  ,
 .. ... ... 
 . 
q1k−1 q2k−1 · · · qkk−1
and Ai is the matrix obtained from A by replacing its ith column by the column
[x0, x1, . . . , xk−1]T . The determinant of A is given by
Y
det A = (qj − qi) 6= 0.
1≤i<j≤k

This finishes the proof of Part (b).

Example 3.3. Find the sequence (xn) satisfying the recurrence relation
xn = 2xn−1 + xn−2 − 2xn−3, n≥3
and the initial conditions x0 = 1, x1 = 2, and x2 = 0.

14
Solution. The characteristic equation of the recurrence relation is
x3 − 2x2 − x + 2 = 0.
Factorizing the equation, we have
(x − 2)(x + 1)(x − 1) = 0.
There are three roots x = 1, −1, 2. By Theorem 3.1, we have the general solution
xn = c1(−1)n + c2 + c32n.
Applying the initial conditions,

 c1 +c2 +c3 = 1
c −c2 +2c3 = 2
 1
c1 +c2 +4c3 = 0
Solving the linear system we have c1 = 2, c2 = −2/3, c3 = −1/3. Thus
2 1
xn = 2 − (−1)n − 2n.
3 3
Theorem 3.2. (a) Let q be a root with multiplicity m of the characteristic
equation (9) associated with the kth order homogeneous linear recurrence
relation (8) with constant coefficients. Then the m sequences
xn = q n , nq n, ..., nm−1q n
are linearly independent solutions of the recurrence relation (8).
(b) Let q1, q2, . . . , qs be distinct nonzero roots with the multiplicities
m1 , m2 , ..., ms
respectively for the characteristic equation (9). Then the sequences
xn = q1n, nq1n, ..., nm1−1q1n;
q2n, nq2n, ..., nm2−1q2n;
... ... ...
qsn, nqsn, ..., nms−1qsn; n≥0
are linearly independent solutions of the homogeneous linear recurrence re-
lation (8). Their linear combinations form the general solution of the re-
currence relation (8).
15
Proof. The falling factorial polynomial of degree i is the polynomial
[t]i := t(t − 1) · · · (t − i + 1).
Let q be a root of the characteristic polynomial
P (t) = α0tk + α1tk−1 + · · · + αk−1t + αk = (t − q)mQ(t)
with multiplicity m. Then q is a root of the ith derivative P (i)(t) of P (t), where
0 ≤ i ≤ m−1. We claim that xn = [n]iq n is a solution of the recurrence relation
for each i = 0, 1, . . . , m − 1. In fact,
LHS of (8) = α0[n]iq n + α1[n − 1]iq n−1 + · · · + αk [n − k]iq n−k
¯
di ¯¯ i
¡ n n−1 n−k
¢
= q α t + α t + · · · + α t
dti ¯t=q
0 1 k
¯
i¯
i d ¯
¡ k k−1
¢ n−k
= q · i¯ α0t + α1t + · · · + αk−1t + αk t
dt t=q
¯
i¯
d
= q i · i ¯¯ (t − q)mR(t), where R(t) = Q(t)tn−k .
dt t=q
Applying the Leibniz rule,
i µ ¶µ j ¯
X ¶µ ¯ ¶
i i d ¯¯ d ¯¯
i−j
LHS of (8) = q (t − q)m R(t)
j dtj ¯
j=0 t=q dti−j ¯ t=q
i µ ¶ ¯ µ ¯ ¶
X i ¯ di−j ¯¯
= qi [m]j (t − q)m−j ¯¯ R(t) = 0.
j=0
j t=q dti−j ¯t=q

So xn = [n]iq n is a solution. Note that

[t]i = a0 + a1t + · · · + aiti, (i = 0, 1, . . . , m − 1)
for some integers a0, a−1, . . . , ai. It follows that there some integers b0, b1, . . . , bi
such that
ti = b0[t]0 + b1[t]1 + · · · + bi[t]i, i = 0, 1, . . . , m − 1.
We see that
xn = niq n = b0[n]0q n + b1[n]1q n + · · · + bi[n]iq n
16
is a solution, where 0 ≤ i ≤ m − 1. Next we show that q n, nq n, . . . , nm−1q n are
linearly independent. Set
c0q n + c1nq n + · · · + cm−1nm−1q n = 0, ∀n ≥ 0.
Since q 6= 0, we have
c0 + c1n + · · · + cm−1nm−1 = 0 ∀n ≥ 0.
Consider c0, c1, . . . , cm−1 as variables and let n = 1, 2, . . . , m. We have a system
of linear equations about c0, c1, . . . , cm−1. The coefficient matrix
 
2 m−1
1 1 1 ··· 1
 1 2 22 · · · 2m−1 
 
 
A =  1 3 32 · · · 3m−1 
 .. .. .. ... 
. . . 
1 m m2 · · · mm−1
Q
and det A = 1≤i<j≤m(j − i) 6= 0. Hence c0 = c1 = · · · = cm−1 = 0.
Proof of General Linear Independence of Solutions: Consider the linear com-
bination s
X
(cj,1qjn + cj,2nqjn + · · · + cj,mj nmj −1qjn) = 0 (13)
j=1
with coefficients cj,1, cj,2, . . . , cj,mj , where j = 1, . . . , s. We assume the roots
q1, . . . , qs are already arranged in increasing order of their modulus as |q1| ≤
|q2| ≤ · · · ≤ |qs|, and the roots of equal modulus are are already arranged
in increasing order of their multiplicities. Let qr , . . . , qs be all roots of largest
modulus with the same largest multiplicity m = mr = · · · = ms. We may have
r = 1. Assume r > 1. Then
|q1| ≤ · · · ≤ |qr−1| ≤ |qr | = · · · = |qs|.
We have either |qr−1| < |qr | or |qr−1| = |qr |. If |qr−1| = |qr |, we must have
mr−1 < mr . Let ωj = qj /qs for j = 1, . . . , s. Then ωr , . . . , ωs are distinct roots
of unity. Divide both sides of (13) by nm−1qsn and move the lower order terms

17
to the right, we obtain
s
X
cj,mωjn = ε(n) → 0 (n → ∞).
j=r

Let A(n) be the (s − r + 1) × (s − r + 1) matrix

 
ωrn n
ωr+1 ··· ωsn
 n+1 n+1 
 ωr ωr+1 · · · ωsn+1 
A(n) =  ... ... ... ,
 
n+s−r
ωrn+s−r ωr+1 · · · ωsn+s−r
and Aj (n) the matrix obtained from A(n) by replacing its jth column with
[ε(n), ε(n + 1), . . . , ε(n + s − r)]T .
Q
Then det A(n) = ωrnωr+1 n
· · · ωsn r≤i<j≤s(ωj − ωi) 6= 0. By Cramer’s rule, we
have
det Aj (n)
cj,m = → 0 (n → ∞), j = r, . . . , s.
det A(n)
It follows that cj,m = 0 for j = r, . . . , s. Likewise, applying the same method to
other nonzero coefficients with highest order, we see that all coefficients in (13)
are zero. ¤

4 Matric representation of linear recurrence relations

Let us write x0 = [x0, x1, . . . , xk−1]T , x1 = [xk , xk+1, . . . , x2k−1]T , . . . ,

xn = [xnk , xnk+1, . . . , xnk+k−1]T , n ≥ 0.
Assume xm = a1xm−1 + a2xm−2 + · · · + ak xm−k , m ≥ k. We see that
xn = Axn−1, n ≥ 1.
Then
xn = A n x 0 , n ≥ 0.
Example 4.1. Find the sequence (xn) satisfying the recurrence relation
xn = 2xn−1 + xn−2 − 2xn−3, n≥3
and the initial conditions x0 = 1, x1 = 2, and x2 = 0.
18
The recurrence relation can be repeated as

 x3 = 2x2 + x1 − 2x0 = −2x0 + x1 + 2x2
x = 2x3 + x2 − 2x1 = −4x0 + 5x2
 4
x5 = 2x4 + x3 − 2x2 = −10x0 + x1 + 10x2
Thus     
x3 −2 1 2 x0
 x4  =  −4 0 5   x1  , i.e., x1 = Ax0
x5 −10 1 10 x2
It follows that xn = Axn−1 = A2xn−2 = · · · = Anx0.
   n  
x3n −2 1 2 x0
xn =  x3n+1  =  −4 0 5   x1  , n ≥ 0.
x3n+2 −10 1 10 x2

5 Nonhomogeneous linear recurrence relations with constant co-

efficients

Theorem 5.1. Given a nonhomogeneous linear recurrence relation of the

first order
xn = αxn−1 + βn, n ≥ 1. (14)
(a) Let βn = cq n be an exponential function of n. Then (14) has a particular
solution of the following form.

• If q 6= α, then xn = Aq n.
• If q = α, then xn = Anq n.
P
(b) Let βn = ki=0 bini be a polynomial function of n with degree k.
• If α 6= 1, then (14) has a particular solution of the form
x n = A 0 + A 1 n + A2 n 2 + · · · + A k n k ,

19
 where the coefficients A0, A1, . . ., Ak are recursively determined as
bk
Ak = ,
1−α 
Xk µ ¶
1  j
Ai = bi + α (−1)j−i Aj  , 0 ≤ i ≤ k − 1.
1−α j=i+1
i

• If α = 1, then the solution of (14) is given by

n
X
xn = x0 + βi.
i=1

Proof. (a) We may assume q 6= 0; otherwise the recurrence (14) is homogeneous.

For the case q 6= α, put xn = Aq n in (14); we have
Aq n = αAq n−1 + cq n.
The coefficient A is determined as A = cq/(q − α).
For the case q = α, put xn = Anq n in (14); we have
Anq n = αA(n − 1)q n−1 + cq n.
Since q = α, then αAq n−1 = cq n. The coefficient A is determined as A = cq/α.
Pk
(b) For the case α 6= 1, put xn = j=0 Aj nj in (14); we obtain
k
X k
X k
X
j j
Aj n = α Aj (n − 1) + bj n j .
j=0 j=0 j=0

Then
k
X k
X j µ ¶
X k
X
j j i j−i
Aj n = α Aj n (−1) + bj n j .
j=0 j=0 i=0
i j=0
k
X k
X k
X µ ¶ Xk
j
Ai n i = α ni (−1)j−i Aj + bi n i .
i=0 i=0 j=i
i i=0
 
X k Xk µ ¶
Ai − bi − α j
(−1)j−i Aj  ni = 0.
i=0 j=i
i

20
The coefficients A0, A1, . . . , Ak are determined recursively as
bk
Ak = ,
1−α 
k
X µ ¶
1  j
Ai = bi + α (−1)j−i Aj  , 0 ≤ i ≤ k − 1.
1−α j=i+1
i

As for the case α = 1, iterate the recurrence relation (14); we have

xn = xn−1 + βn = xn−2 + βn−1 + βn
= xn−1 + βn−2 + βn−1 + βn = · · ·
= x0 + β1 + β2 + · · · + βn.

Example 5.1. Solve the difference equation

½
xn = xn−1 + 3n2 − 5n3, n ≥ 1
x0 = 2.
Solution.
n
X n
X ¡ 2 3
¢
xn = x0 + bi = 2 + 3i − 5i
i=1 i=1
X n n
X
= 2+3 i2 − 5 3
i
i=1 i=1
µ ¶2
n(n + 1)(2n + 1) n(n + 1)
= 2+3× −5× .
6 2
We have applied the following identities
n
X n(n + 1)(2n + 1)
i2 = ,
i=1
6
n
X µ ¶2
n(n + 1)
i3 = .
i=1
2

21
Example 5.2. Solve the equation
½
xn = 3xn−1 − 4n, n ≥ 1
x0 = 2.
Solution. Note that xn = 3nc is the general solution of the corresponding
homogeneous linear recurrence relation. Let xn = An + B be a particular
solution. Then
An + B = 3[A(n − 1) + B] − 4n
Comparing the coefficients of n0 and n, it follows that A = 2 and B = 3. Thus
the general solution is given by
xn = 2n + 3 + 3nc.
The initial condition x0 = 2 implies that c = −1. Therefore the solution is
xn = −3n + 2n + 3.
Theorem 5.2. Given a nonhomogeneous linear recurrence relation of the
second order
xn = α1xn−1 + α2xn−2 + cq n. (15)
Let q1 and q2 be solutions of its characteristic equation
x2 − α1x − α2 = 0.
Then (15) has a particular solution of the following forms, where A is a
constant to be determined.
(a) If q 6= q1, q 6= q2, then xn = Aq n.
(b) If q = q1, q1 6= q2, then xn = Anq n.
(c) If q = q1 = q2, then xn = An2q n.
Proof. The homogeneous linear recurrence relation corresponding to (15) is
xn = α1xn−1 + α2xn−2, n ≥ 2. (16)
We may assume q 6= 0. Otherwise (15) is homogeneous.

22
 (a) Put xn = Aq n into (15); we have
Aq n = α1Aq n−1 + α2Aq n−2 + cq n.
Then
A(q 2 − a1q − a2) = cq 2.
Since q is not a root of the characteristic equation x2 = α1x + α2, that is,
q 2 − α1q − α2 6= 0, the coefficient A is determined as
cq 2
A= 2 .
q − α1q − α2
(b) Since q = q1 6= q2, then xn = q n is a solution of (16) but xn = nq n is not,
that is,
q 2 − α1q − α2 = 0 and nq n 6= α1(n − 1)q n−1 + α2(n − 2)q n−2.
It follows that
nq 2 − α1(n − 1)q − α2(n − 2) = n(q 2 − α1q − α2) + α1q + 2α2
= α1q + 2α2 6= 0.
Put xn = Anq n into (15); we have
Anq n = α1A(n − 1)q n−1 + α2A(n − 2)q n−2 + cq n.
Then £ ¤
A nq 2 − α1(n − 1)q − α2(n − 2) = cq 2.
Since α1q + 2α2 6= 0, the coefficient A is determined as
cq 2
A= .
α1q + 2α2
(c) Since q = q1 = q2, then both xn = q n and xn = nq n are solutions of (16),
but xn = n2q n is not. It then follows that
q 2 − α1q − α2 = 0, α1q + 2α2 = 0, and

n2q 2 − α1(n − 1)2q − α2(n − 2)2 = n2(q 2 − α1q − α2) + 2n(α1q + 2α2) − α1q −
= −α1q − 4α2 6= 0.
23
Put xn = An2q n into (15); we have
£ ¤
Aq n−2 n2q 2 − α1(n − 1)2q − α2(n − 2)2 = cq n.
The coefficient A is determined as
cq 2
A=− .
α1q + 4α2

Example 5.3. Solve the equation


 xn = 10xn−1 − 25xn−2 + 5n+1, n ≥ 2
x = 5
 0
x1 = 15.
Put xn = An2 × 5n into the recurrence relation. We have
An2 × 5n = 10A(n − 1)2 × 5n−1 − 25A(n − 2)2 × 5n−2 + 5n+1.
Dividing both sides we further have
An2 = 2A(n − 1)2 − A(n − 2)2 + 5.
Thus A = 5/2. The general solution is given by
5
xn = n25n + c15n + c2n5n.
2
Applying the initial conditions x0 = 5 and x1 = 15, we have c1 = 5 and
c2 = −9/2. Hence µ ¶
5 2 9
xn = n − n + 5 5n.
2 2
Theorem 5.3. Given a nonhomogeneous linear recurrence relation of the
second order
xn = α1xn−1 + α2xn−2 + βn, n ≥ 2, (17)
where βn is a polynomial function of n with degree k.

24
(a) If α1 + α2 6= 1, then (17) has a particular solution of the form
x n = A0 + A 1 n + · · · + Ak n k ,
where A0, A1, . . ., Ak are constants to be determined. If k ≤ 2, then a
particular solution has the form
x n = A0 + A1 n + A 2 n 2 .

(b) If α1 + α2 = 1, then (17) can be reduced to a first order recurrence

relation
yn = (α1 − 1)yn−1 + βn, n ≥ 2,
where yn = xn − xn−1 for n ≥ 1.
Pk j
Pk j
Proof. (a) Let βn = b
j=0 j n . Put x n = j=0 Aj n into the recurrence
relation (17); we obtain
k
X k
X k
X k
X
j j j
Aj n = α1 Aj (n − 1) + α2 Aj (n − 2) + bj n j
j=0 j=0 j=0 j=0
k
X j
X µ ¶
j
= α1 Aj (−1)j−i ni
j=0 i=0
i
X k X j µ ¶ Xk
j−i j i
+α2 Aj (−2) n + bj n j
j=0 i=0
i j=0
Xk X k µ ¶
j
= α1 ni (−1)j−i Aj
i=0 j=i
i
X k X k µ ¶ Xk
j
+α2 ni (−2)j−i Aj + bj n j .
i=0 j=i
i j=0

Collecting the coefficients of ni, we have

 
X k X k µ ¶ Xk µ ¶ Xk
Ai − α1 j j
(−1) j−i
Aj − α2 (−2)j−i
Aj − bi ni = 0.
i=0 j=i
i j=i
i i=0

25
Since α1 + α2 6= 1, the coefficients A0, A1, . . ., Ak are determined as
bk
Ak = ,
1 − α1 − α2  
k
X µ ¶³ ´
1 bi + j
Ai = (−1)j−i
α1 + 2 α2 Aj  ,
j−i
1 − α1 − α2 j=i+1
i

where 0 ≤ i ≤ k − 1.
(b) The recurrence relation (17) becomes
xn = α1xn−1 + (1 − α1)xn−2 + βn, n ≥ 2.
Set yn = xn − xn−1 for n ≥ 1; the recurrence (17) reduces to a first order
recurrence relation.
Example 5.4. Solve the following recurrence relation

 xn = 6xn−1 − 9xn−2 + 8n2 − 24n
x = 5
 0
x1 = 5.
Solution. Put xn = A0 + A1n + A2n2 into the recurrence relation; we obtain
A0+A1n+A2n2 = 6[A0+A1(n−1)+A2(n−1)2]−9[A0+A1(n−2)+A2(n−2)2]+8n2−
Collecting the coefficients of n2, n, and the constant, we have
(4A2 − 8)n2 + (4A1 − 24A2 + 24)n + (4A0 − 12A1 + 30A2) = 0.
We conclude that A2 = 2, A1 = 6, and A0 = 3. So xn = 2n2 + 6n + 3 is a
particular solution. Then the general solution of the recurrence is
xn = 2n2 + 6n + 3 + 3nc1 + 3nnc2.
Applying the initial condition x0 = x1 = 5, we have c1 = 2, c2 = −4. The
sequence is finally obtained as
xn = 2n2 + 6n + 3 + 2 × 3n − 4n × 3n.

26
6 Generating functions

The (ordinary) generating function of an infinite sequence

a0 , a 1 , a 2 , . . . , a n , . . .
is the infinite series
A(x) = a0 + a1x + a2x2 + · · · + anxn + · · · .
A finite sequence
a0 , a 1 , a 2 , . . . , a n
can be regarded as the infinite sequence
a0, a1, a2, . . . , an, 0, 0, . . .
and its generating function
A(x) = a0 + a1x + a2x2 + · · · + anxn
is a polynomial.
Example 6.1. The generating function of the constant infinite sequence
1, 1, . . . , 1, . . .
is the function
1
A(x) = 1 + x + x2 + · · · + xn + · · · = .
1−x
Example 6.2. For any positive integer n, the generating function for the bi-
nomial coefficients
³n´ ³n´ ³n´ ³n´
, , , ..., , 0, . . .
0 1 2 n
is the function n ³ ´
X n k
x = (1 + x)n.
k
k=0

27
Example 6.3. For any real number α, the generating function for the infinite
sequence of binomial coefficients
³α´ ³α´ ³α´ ³α´
, , , ..., , ...
0 1 2 n
is the function ∞ ³ ´
X α n
x = (1 + x)α .
n=0
n
Example 6.4. Let k be a positive integer and let
a0 , a 1 , a 2 , . . . , a n , . . .
be the infinite sequence whose general term an is the number of nonnegative
integral solutions of the equation
x1 + x2 + · · · + xk = n.
Then the generating function of the sequence (an) is
 
X∞ X X∞ X
A(x) =  
1 x =n
xi1+···+ik
n=0 i1 +···+ik =n n=0 i1 +···+ik =n
   
∞
X ∞
X 1
=  x i1 
··· x ik  =
i1 =0 ik =0
(1 − x)k
∞
X µ ¶ ∞ µ
X ¶
−k n+k−1
= (−1)n xn = xn .
n=0
n n=0
n
Example 6.5. Let an be the number of integral solutions of the equation
x1 + x2 + x3 + x4 = n,
where 0 ≤ x1 ≤ 3, 0 ≤ x2 ≤ 2, x3 ≥ 2, and 3 ≤ x4 ≤ 5. The generating
function of the sequence (an) is
¡ 2 3
¢¡ 2
¢¡ 2 3
¢¡ 3 4 5
¢
A(x) = 1 + x + x + x 1 + x + x x + x + · · · x + x + x
5
¡ 2 3
¢¡ ¢
2 2
x 1+x+x +x 1+x+x
= .
1−x
28
Example 6.6. Determine the generating function for the number of n-combinations
of apples, bananas, oranges, and pears where in each n-combination the num-
ber of apples is even, the number of bananas is odd, the number of oranges is
between 0 and 4, and the number of pears is at least two.
The required generating function is
Ã∞ !à ∞ !à 4 !à ∞ !
X X X X
2i 2i+1 i
A(x) = x x x xi
i=0 i=0 i=0 i=2
3 5
x (1 − x )
= .
(1 − x2)2(1 − x)2
Example 6.7. Determine the number an of bags with n pieces of fruit (apples,
bananas, oranges, and pears) such that the number of apples is even, the number
bananas is a multiple of 5, the number oranges is at most 4, and the number of
pears is either one or zero.
The generating function of the sequence (an) is
Ã∞ !à ∞ !à 4 !à 1 !
X X X X
A(x) = x2i x5i xi xi
i=0 i=0 i=0 i=0
(1 + x + x2 + x + x4)(1 + x)
3
=
(1 − x2)(1 − x5)
(1 + x)(1 − x5)/(1 − x)
=
(1 + x)(1 − x)(1 − x5)
X∞ µ ¶
1 n −2 n
= = (−1) x
(1 − x)2 n=0 n
X∞ µ ¶ X∞
n+1
= xn = (n + 1)xn.
n=0
n n=0

Thus an = n + 1.
Example 6.8. Find a formula for the number an,k of integer solutions (i1, i2, . . . , ik )
of the equation
x1 + x2 + · · · + xk = n
such that i1, i2, . . . , ik are nonnegative odd numbers.
29
 The generating function of the sequence (an) is
̰ ! ̰ !
X X xk
A(x) = x2i+1 · · · x2i+1 =
i=0 i=0
(1 − x2)k
X∞ µ ¶ X∞ µ ¶
i + k − 1 i + k − 1
= xk x2i = x2i+k
i i
 i=0 ³ i=0
 P∞ j+r−1 ´ 2j
 j=r j−r x for k = 2r
=
 P∞ ³ j+r ´ 2j+1
 j=r j−r x for k = 2r + 1.
¡ s+r−1 ¢ ¡ s+r ¢
We then conclude that a2s,2r = s−r , a2s+1,2r+1 = s−r , and an,k = 0
otherwise. We may combine the three cases into
³ n k ´
 b 2 c+d 2 e−1 if n − k = even,
b n2 c−b k2 c
an,k =

0 if n − k = odd.
Example 6.9. Let an denote the number of nonnegative integral solutions of
the equation
2x1 + 3x2 + 4x3 + 5x4 = n.
Then the generating function of the sequence (an) is
 
X∞
 X 
A(x) =  1 xn
n=0 i,j,k,l≥0
2i+3j+4k+5l=0
Ã∞ ! ∞ Ã
∞
!Ã ∞ !
X X X X
2i  3j  4k
= x x x x5l
i=0 j=0 k=0 l=0
1
= .
(1 − x2)(1 − x3)(1 − x4)(1 − x5)
Theorem 6.1. Let sn be the number of nonnegative integral solutions of
the equation
a1x1 + a2x2 + · · · + ak xk = n.

30
Then the generating function of the sequence (sn) is
1
A(x) = .
(1 − xa1 )(1 − xa2 ) · · · (1 − xak )

7 Recurrence and generating functions

Since
∞ µ
X ¶ ∞ µ
X ¶
1 −n k n + k − 1 k
= (−x) = x , |x| < 1;
(1 − x)n k k
k=0 k=0

then
∞ µ
X ¶ ∞ µ
X ¶
1 −n k n + k − 1 k k 1
= (−ax) = a x , |x| < .
(1 − ax)n k k |a|
k=0 k=0

Example 7.1. Determine the generating function of the sequence

0, 1, 22, . . . , n2, . . . .
1
P∞ k
Since 1−x = k=0 x, then
µ ¶ X ∞ ∞
1 d 1 d ¡ k ¢ X k−1
= = x = kx .
(1 − x)2 dx 1 − x dx
k=0 k=0
x
P∞
Thus (1−x)2
= k=0 kxk . Taking the derivative with respect to x we have
X ∞
1+x 2 k−1
= k x .
(1 − x)3
k=0

Therefore the desired generating function is

x(1 + x)
A(x) = .
(1 − x)3
Example 7.2. Solve the recurrence relation

 an = 5an−1 − 6an−2, n≥2
a = 1
 0
a1 = −2

31
 P∞ n
Let A(x) = n=0 an x . Applying the recurrence relation, we have
∞
X
A(x) = a0 + a1x + (5an−1 − 6an−2) xn
n=2
= a0 + a1x − 5xa0 + 5xA(x) − 6x2A(x).
Applying the initial values and collecting the coefficient functions of A(x), we
further have ¡ ¢
2
1 − 5x + 6x A(x) = 1 − 7x.
Thus the function g(x) is solved as
1 − 7x
A(x) = .
1 − 5x + 6x2
Observing that 1 − 5x + 6x2 = (1 − 2x)(1 − 3x) and applying partial fraction,
1 − 7x A B
= + .
1 − 5x + 6x2 1 − 2x 1 − 3x
The constants A and B can be determined by
A(1 − 3x) + B(1 − 2x) = 1 − 7x.
Then ½
A +B = 1
−3A −2B = −7
Thus A = 5, B = −4. Hence
1 − 7x 5 4
A(x) = = − .
1 − 5x + 6x2 1 − 2x 1 − 3x
Since ∞ ∞
1 X 1 X
n n
= 2 x and = 3nxn
1 − 2x n=0 1 − 3x n=0
We obtain the sequence
an = 5 × 2n − 4 × 3n, n ≥ 0.

32
Theorem 7.1. Let (an; n ≥ 0) be a sequence satisfying the homogeneous
linear recurrence relation of order k with constant coefficients, i.e.,
an = α1an−1 + α2an−2 + · · · + αk an−k , (18)
P∞ n
where αk 6= 0, n ≥ k. Then its generating function A(x) = n=0 an x is a
rational function of the form
P (x)
A(x) = , (19)
Q(x)
where Q(x) is a polynomial of degree k with a nonzero constant term and
P (x) is a polynomial of degree strictly less than k.
Conversely, given such polynomials P (x) and Q(x), there exists a unique
sequence (an) satisfying the linear homogeneous recurrence relation (18),
and its generating function is the rational function in (19).
Proof. The generating function A(x) of the sequence (an) can be written as
k−1 ∞ k−1 ∞
à k !
X X X X X
A(x) = ai x i + an x n = ai x i + αian−i xn
i=0 n=k i=0 n=k i=1
k−1
X Xk ∞
X k−1
X k
X ∞
X
= ai x i + αi an−ixn = ai x i + αi anxn+i
i=0 i=1 n=k i=0 i=1 n=k−i
 
k−1
X ∞
X k−1
X ∞
X k−i−1
X
= aixi + αk xk anxn + αixi  an x n − aj x j 
i=0 n=0 i=1 n=0 j=0
k−1
X k
X k−1
X k−i−1
X
i i i
= aix + A(x) αix − αix aj x j
i=0 i=1 i=1 j=0
k
X k−1
X k−1
X l
X
i i l
= A(x) αix + ai x − x αial−i.
i=1 i=0 l=1 i=1

33
Then
à k
! k−1 k−1 l
X X X X
A(x) 1 − αixi = ai x i − xl αial−i
i=1 i=0 l=1 i=1
k−1
à i
!
X X
= a0 + al − αial−i xl .
l=1 i=1

Thus
k−1
à l
!
X X
P (x) = a0 + al − αial−i xl ,
l=1 i=1
k
X
Q(x) = 1 − αixi.
i=1

Conversely, let (an) be the sequence whose generating function is A(x) =

P (x)/Q(x). Write
∞
X k
X k
X
A(x) = an x n , P (x) = bi x i , Q(x) = 1 − αixi.
n=0 i=0 i=1
P (x)
Then A(x) = Q(x) is equivalent to
à k
!Ã ∞ ! k
X X X
1− αixi anxn = bi x i .
i=1 n=0 i=0

The polynomial Q(x) can be viewed as an infinite series with αi = 0 for i > k.
Thus à n !
X∞ X∞ X X k
n n
anx − αian−i x = bi x i .
n=0 n=0 i=1 i=0
n
Equating the coefficients of x , we have the recurrence relation
k
X
an = αian−i, n ≥ k.
i=1

34
Proposition 7.2 (Partial Fractions). (a) If P (x) is a polynomial of degree
less than k, then
P (x) A1 A2 Ak
= + + · · · + ,
(1 − ax)k 1 − ax (1 − ax)2 (1 − ax)k
where A1, A2, . . . , Ak are constants to be determined.
(b) If P (x) is a polynomial of degree less than p + q + r, then
P (x) A1(x) A2(x) A3(x)
= + + ,
(1 − ax)p(1 − bx)q (1 − cx)r (1 − ax)p (1 − bx)q (1 − cx)r
where A1(x), A2(x), and A3(x) are polynomials of degree q + r, p + r, and
p + q, respectively.

8 A geometry example

A polygon P in R2 is called convex if the segment joining any two points in

P is also contained in P . Let Cn denote the number of ways to divide a labeled
convex polygon with n + 2 sides into triangles. The first a few such numbers are
C1 = 1, C2 = 2, C3 = 5.
We first establish a recurrence relation for Cn+1 in terms of C0, C1, . . ., Cn.
Let Pv1v2...vn+3 denote a convex (n + 3)-polygon with vertices v1, v2, . . . , vn+3. In
each triangular decomposition of Pv1v2...vn+3 into triangles, the segment v1vn+3 is
one side of a triangle ∆ in the decomposition; the third vertex of the triangle ∆ is
one of the vertices v2, v3, . . . , vn+2. Let vk+2 be the third vertex of ∆ other than
v1 and vn+3 (0 ≤ k ≤ n); see Figure 1 below. Then we have a convex (k + 2)-
vk+2

vn+2

vn+3 v3

v1 v2

Figure 1: vk+2 is the third vertex of the triangle with the side v1 vn+3

polygon Pv1v2...vk+2 and another convex (n − k + 2)-polygon Pvk+2vk+3...vn+3 . Then

35
by induction there are Ck ways to divide Pv1v2...vk+2 into triangles, and Cn−k ways
to divide Pvk+2vk+3...vn+3 into triangles. We thus have the recurrence relation
n
X
Cn+1 = Ck Cn−k with C0 = 1.
k=0
P
Consider the generating function F (x) = ∞ n
n=0 Cn x . Then
Ã∞ !à ∞ !
X X
F (x)F (x) = Cnxn Cnxn
n=0 n=0
∞
à n !
X X
= Ck Cn−k xn
n=0 k=0
X∞ ∞
n 1X
= Cn+1x = Cnxn
n=0
x n=1
F (x) 1
= − .
x x
We thus obtain the equation
xF (x)2 − F (x) + 1 = 0.
Solving for F (x), we obtain
√
1± 1 − 4x
F (x) = .
2x
Note that
∞
X µ1¶
√
1 − 4x = 1 + (−1)n 2 4nxn
n=1
n
X∞
= 1+ anxn,
n=1

36
where
· µ ¶ µ ¶. ¸
1 1 1
an = (−1)n − 1 ··· −n+1 n! · 2n · 2n
2 2 2
(−1)(−3)(−5) · · · (−2 n − 1 + 1) n
= (−1)n · ·2
n!
1 · 3 · 5 · · · (2 n − 1 − 1) n
= − ·2
n!
(2(n − 1))!
= −2 · .
n!(n − 1)!
Then
∞
X
√ (2(n − 1))!
1 − 4x = 1 − 2 xn
n=1
n!(n − 1)!
X∞
(2n)!
= 1−2 xn+1.
n=0
n!(n + 1)!

We conclude that
√
1− 1 − 4x
F (x) =
2x
∞
X (2n)!
= xn
n=0
n!(n + 1)!
X∞ µ ¶
1 2n
= xn .
n=0
n+1 n

Hence the sequence (Cn) is given by the binomial coefficients:

µ ¶
1 2n
Cn = , n ≥ 0.
n+1 n
The sequence (Cn) is known as the Catalan sequence and the numbers Cn
as the Catalan numbers.
Example 8.1. Let Cn be the number of ways to evaluate a matrix product
A1A2 · · · An+1, n≥0
37
by adding various parentheses. For instance, C0 = 1, C1 = 1, C2 = 2, and
C3 = 5. In general the formula is given by
µ ¶
1 2n
Cn = .
n+1 n
Note that each way of evaluating the matrix product A1A2 · · · An+2 will be
finished by multiplying of two matrices at the end. There are exactly n + 1 ways
of multiplying the two matrices at the end:
A1A2 · · · An+2 = (A1 · · · Ak+1)(Ak+2 · · · An+2), 0 ≤ k ≤ n.
This yields the recurrence relation
n
X
Cn+1 = Ck Cn−k .
k=0
1
¡ 2n ¢
Thus Cn = n+1 n , n ≥ 0.

9 Exponential generating functions

The ordinary generating function method is a powerful algebraic tool for finding
unknown sequences, especially when the sequences are certain binomial coeffi-
cients or the order is immaterial. However, when the sequences are not binomial
type or the order is material in defining the sequences, we may need to consider
a different type of generating functions. For example, the sequence an = n! is
the counting of the number of permutations of n distinct objects; its ordinary
generating function
X∞ X∞
n
an x = n!xn
n=0 n=0
cannot be easily figure out as a closed known expression. However, the generating
function ∞ ∞
X an n X n 1
x = x =
n=0
n! n=0
1−x
is obvious.

38
 The exponential generating function of a sequence (an; n ≥ 0) is the
infinite series ∞
X an n
E(x) = x .
n=0
n!
Example 9.1. The exponential generating function of the sequence
P (n, 0), P (n, 1), . . . , P (n, n), 0, . . .
is given by
n
X P (n, k)
E(x) = xk
k!
k=0
n ³
X n´
= xk
k
k=0
= (1 + x)n.
Example 9.2. The exponential generating function of the constant sequence
(an = 1; n ≥ 0) is
X∞
xn
E(x) = = ex.
n=0
n!
The exponential generating function of the geometric sequence (an = an; n ≥ 0)
is ∞
X an x n
E(x) = = eax.
n=0
n!
Theorem 9.1. Let M = {n1α1, n2α2, . . . , nk αk } be a multiset over the set
S = {α1, α2, . . . , αk } with n1 many α1’s, n2 many α2’s, . . ., nk many αk ’s.
Let an be the number of n-permutations of the multiset M . Then the expo-
nential generating function of the sequence (an; n ≥ 0) is given by
Ãn !Ã n ! Ãn !
Xx1 i Xx 2 i Xxk i
E(x) = ··· . (20)
i=0
i! i=0
i! i=0
i!

39
Proof. Note that an = 0 for n > n1 + · · · + nk . Thus E(x) is a polynomial. The
right side of (20) can be expanded to the form
n1 ,n2 ,...,nk n1 +n2 +···+nk
X xi1+i2+···+ik X xn X n!
= .
i1 ,i2 ,...,ik
i !i
=0 1 2
! · · · i k ! n=0
n! i1 +i2 +···+ik =n
i1!i2! · · · ik !
0≤i1 ≤n1 ,...,0≤ik ≤nk

Note that the number of permutation of M with exactly i1 α1’s, i2 α2’s, . . .,

and ik αk ’s such that
i1 + i 2 + · · · + i k = n
is the multinomial coefficient
µ ¶
n n!
= .
i1 , i 2 , . . . , i k i1!i2! · · · ik !
It turns out that the sequence (an) is given by
X µ ¶
n
an = , n ≥ 0.
i +i +···+i =n
i ,
1 2i , . . . , i k
1 2 k
0≤i1 ≤n1 ,...,0≤ik ≤nk

We proved the closed form of the required exponential generating function.

Example 9.3. Determine the number of ways to color the squares of a 1-by-n
chessboard using the colors, red, white, and blue, if an even number of squares
are colored red.
Let an denote the number of ways of such colorings and set a0 = 1. Each such
coloring can be considered as a permutation of three objects r (for red), w (for
white), and b (for blue) with repetition allowed, and the element r appears even

40
number of times. The exponential generating function of the sequence (an) is
Ã∞ !à ∞ !2
X x 2n Xx n
E(x) =
n=0
(2n)! n=0
n!
ex + e−x 2x 1 ¡ 3x ¢
= e = e + ex
Ã2∞ 2
∞
!
1 X n n
3 x X x n
= +
2 n=0 n! n=0
n!
∞
1X n xn
= (3 + 1) · .
2 n=0 n!
Thus the sequence is given by
3n + 1
an = , n ≥ 0.
2
Example 9.4. Determine the number an of n digit (under base 10) numbers
with each digit odd where the digit 1 and 3 occur an even number of times.
Assume a0 = 1. The number an equals the number of n-permutations of the
multiset M = {∞1, ∞3, ∞5, ∞7, ∞9}, in which 1 and 3 occur an even number
of times. The exponential generating function of the sequence an is
Ã∞ !2 à ∞ !3
X x2n X xn
E(x) =
n=0
(2n)! n=0
n!
µ x ¶2
e + e−x
= e3x
2
1 ¡ 5x ¢
= e + 2e3x + ex
4Ã !
∞ ∞ ∞
1 X5 x n n X3 x n n Xx n
= + +
4 n=0 n! n=0
n! n=0
n!
X ∞ µ n ¶
5 + 2 × 3n + 1 xn
= .
n=0
4 n!
Thus
5n + 2 × 3n + 1
an = , n ≥ 0.
4
41
Example 9.5. Determine the number of ways to color the squares of a 1-by-n
board with the colors, red, blue, and white, where the number of red squares is
even and there is at least one blue square.
The exponential generating function for the sequence is
Ã∞ !à ∞ !à ∞ !
X x2i X xi X xi
E(x) =
i=0
(2i)! i=0
i! i=1
i!
ex + e−x x x
= e (e − 1)
2
1 ¡ 3x ¢
= e − e2x + ex − 1
2
∞
1 X 3n − 2n + 1 xn
= − + ·
2 n=0 2 n!
Thus
3n − 2n + 1
an = , n≥1
2
and
a0 = 0.

10 Combinatorial interpretations

Theorem 10.1. The combinatorial interpretations of ordinary generating

functions:
(a) The number of ways of placing n indistinguishable balls into m distin-
guishable boxes is the coefficient of xn in
̰ !m
¡ ¢m X 1
1 + x + x2 + · · · = xk = .
(1 − x)m
k=0

(b) The number of ways of placing n indistinguishable balls into m distin-

guishable boxes with at most rk balls in the kth box is the coefficient of
xn in the expression
m
Y ¡ 2 rk
¢
1 + x + x + ··· + x .
k=1

42
(c) The number of ways of placing n indistinguishable balls into m distin-
guishable boxes with at least sk balls in the kth box is the coefficient of
xn in the expression
Ym
sk
¡ 2
¢ xs1+···+sm
x 1 + x + x + ··· = .
(1 − x)m
k=1

(d) The number of ways of placing n indistinguishable balls into m distin-

guishable boxes, such that the number of balls held in the kth box is
allowed in a subset Ck ⊆ Z≥0 (1 ≤ k ≤ m), is the coefficient of xn in
the expression
    
X X X
 x k  k
x ···  xk  .
k∈C1 k∈C2 k∈Cm

Proof. Let us consider the situation of placing infinitely many indistinguishable

balls into m distinguishable boxes. We use the symbol x to represent a ball and
xk to represent indistinguishable k balls. In the course of placing the balls into
m boxes, each box contains either none of balls (x0 = 1), or one ball (x), or two
balls (x2), or three balls (x3), etc.; that is, each box contains
1 + x + x2 + x3 + · · ·
balls, where the plus sign “+” means “or”.

Box 1 Box 2 Box 3 · · · Box m

1 1 1 1 1
x x x x x
x2 x2 x2 x2 x2
x3 x3 x3 x3 x3
... ... ... ... ...

If we use multiplication to represent “and”, then all possible distributions of

balls in the m boxes in the course of placing are contained in the expression
(1 + x + x2 + · · · )m. (21)
43
Thus, when there are n distinguishable balls to be placed into the m distinguish-
able boxes, the number of distributions of the n balls in the m boxes will be the
coefficient of xn in the expression (21). Indeed, to compute the coefficient of xn
in (21), we select xn1 from column 1, xn2 from column 2, . . ., xnm from column
m in the above table such that xn1 xn2 · · · xnm = xn, and do this in all possible
ways. The number of such ways is clearly the number of nonnegative integer
solutions of the equation
y1 + y2 + · · · + ym = n,
¡ ¢
and is given by m+n−1
n .
The situation for other cases are similar.
Theorem 10.2. The combinatorial interpretation of exponential generating
functions:
(a) The number of ways of selecting an ordered n objects with repetition
n
allowed from an m-set is the coefficient of xn! in the expression
µ 2 3
¶m ÃX ∞ k
!m
x x x x
1+ + + + ··· = = emx. (22)
1! 2! 3! k!
k=0

(b) The number of ways of selecting an ordered n objects from an m-set

such that the number of kth objects is allowed in a subset Ck ⊆ Z≥0, is
n
the coefficient of xn! in the expression
    
X xk X xk X xk
  ··· . (23)
k! k! k!
k∈C1 k∈C2 k∈Cm

Example 10.1. Find the number of 4-permutations of the multiset

M = {a, a, b, b, b, c}.

44
 There are 5 sub-multisets of M having cardinality 4.

{a, a, b, b} {a, a, b, c} {a, b, b, b} {a, b, b, c} {b, b, b, c}

4!/2!2!0! 4!/2!1!1! 4!/1!3!0! 4!/1!2!1! 4!/0!3!1!

Then the answer is

4! 4! 4! 4! 4!
+ + + + = 38,
2!2!0! 2!1!1! 1!3!0! 1!2!1! 0!3!1!
4
which is the coefficient of x4! in the expansion of
µ ¶µ ¶
x x2 x x2 x3 ³ x´
E(x) = 1 + + 1+ + + 1+ .
1! 2! 1! 2! 3! 1!
In fact, the coefficient of x4/4! is
X 4! 4! 4! 4! 4! 4!
a4 = = + + + + .
i+j+k=4
i!j!k! 2!2!0! 2!1!1! 1!3!0! 1!2!1! 0!3!1!
i≤2,j≤3,k≤1

45