VO Combinatorics

(Module: “Combinatorics” (MALK))

Markus Fulmek
Summer Term 2018
The present (draft of) lecture notes is based on lectures “Combinatorics” held
since summer term 2012. It contains selected material from the following text
books:
‚ Species: Combinatorial Species and tree–like Structures by F. Bergeron,
G. Labelle and P. Leroux [1],
‚ Partially ordered sets: Enumerative Combinatorics I by R. Stanley [7],
‚ Asymptotics: Analytic Combinatorics by P. Flajolet and R. Sedgewick
[3].
The reader is assumed to be familiar with basic concepts from Discrete Math-
ematics (see, for instance, lecture notes [4]). Moreover, some knowledge from
Analysis (power series), Linear Algebra, Algebra (basic group theory) and Com-
plex Analysis (contour integrals) is a necessary precondition for the material
presented here.
This draft was translated to English by Ilse Fischer, Markus Fulmek, Christian
Krattenthaler and Michael Schlosser. Many thanks in advance for any hints
regarding typos and other flaws: most likely, this draft contains many of them!
Vienna, summer term 2019
 Contents

Chapter 1. Generating functions and Lagrange’s Inversion Formula 1

1.1. A short recap: formal power series 1
1.2. Lagrange’s Inversion formula 1

Chapter 2. Species: Enumeration of combinatorial objects 7

2.1. Motivating Examples 7
2.2. Species, labelled and unlabelled 9
2.3. Unlabelled species and the enumeration of trees 10
2.3.1. Enumeration of binary trees 15
2.4. Bijective combinatorics on rooted trees 17
2.4.1. Dyck paths of length 2n 17
2.4.1.1. Planar trees Ø Dyck paths 18
2.4.1.2. Ordered binary trees Ø Planar trees 18
2.5. Unordered rooted trees 18
2.6. labelled species 19
2.6.1. Combinatorial proof of Cayley’s theorem 24
2.7. Pólya’s theorem 27
2.8. A generalisation 31
2.9. Cycle index series 32

Chapter 3. Partially ordered sets 43

3.1. Definition and Examples 43
3.2. Construction of posets 48
3.3. Lattices 50
3.3.1. Distributive lattices 54
3.4. Incidence algebra and Möbius inversion 57
3.4.1. The incidence algebra of locally finite posets 58
3.4.2. The zeta function of locally finite posets 59
3.4.3. Möbius inversion 61
3.4.4. Calculation of Möbius functions of several concrete posets 63
3.4.5. Möbius algebra of a locally finite lattice 67

Chapter 4. Asymptotic enumeration 73

4.1. Landau’s notation 73
4.2. Recapitulation: Elements of complex analysis 76
4.3. Singularity analysis 80
4.3.1. The Exponential Growth Formula 81
4.3.2. Rational functions 83
4.3.3. Meromorphic functions 84
4.3.4. Several dominant singularities 89
iii
iv CONTENTS

4.3.5. “Standardized” singularity analysis 89

4.3.5.1. The Gamma–function 90
4.3.5.2. Asymptotics for “standard functions” 93
4.3.5.3. Transfer theorems 100
4.4. Saddle point method 105
4.4.1. Heuristics: contour through saddle point 105
4.4.2. Hayman’s theorem 114
4.4.3. A saddle point theorem for large powers 121
4.5. Asymptotics of combinatorial sums 125
4.6. Exp–log scheme 130
Bibliography 133
 CHAPTER 1

Generating functions and Lagrange’s Inversion Formula

1.1. A short recap: formal power series

Recall that the set Crrzss of formal power series (in some variable z) with com-
plex coefficients
# +
ÿ
Crrzss “ c n ¨ zn : c n P C
ně0
is a commutative algebra with unity over C.
The notion “formal” refers to the fact that we do not actually “compute the
infinite sum c pαq” for some concrete number α, but view c pzq as a convenient
notation for the sequence of coefficients pcn qně0. For some formal power series
c pzq “ c0 ` c1 ¨ z ` c2 ¨ z2 ` ¨ ¨ ¨ , however, we adopt the intuitive notation c p0q
for the constant term c0 , i.e.: c p0q :“ c0 .
T HEOREM 1.1.1. The formal power series a pzq “ a0 ` a1 ¨ z ` a2 ¨ z2 ` ¨ ¨ ¨ possesses
a multiplicative inverse if and only if a p0q “ a0 ‰ 0.
This inverse (if existing) is unique. 
D EFINITION 1.1.2. Let a pzq and b pzq be two formal power series, let b pzq “ 0 (i.e.,
b pzq “ b1 ¨ z ` b2 ¨ z2 ` . . . , b0 “ 0). Then the composition pa ˝ bq pzq of a and b is
defined by
ÿ
pa ˝ bq pzq :“ ai pb pzqqi .
iě0

Note that the condition b p0q “ 0 is necessary, since otherwise we could en-
counter infinite sums for the coefficients of the composition: we do not deal
with infinite sums in the calculus of formal power series.
T HEOREM 1.1.3. Let a pzq “ a1 ¨ z ` a2 ¨ z2 ` ¨ ¨ ¨ be a formal power series with a p0q “
0. Then there is a unique compositional inverse b pzq “ b1 ¨ z ` b2 ¨ z2 ` ¨ ¨ ¨ , i.e.
pa ˝ bq pzq “ pb ˝ aq pzq “ z,
if and only if a1 ‰ 0. 

1.2. Lagrange’s Inversion formula

Let f pzq be a formal power series with vanishing constant term (i.e., f p0q “ 0):
By Theorem 1.1.3 we know that there is a (unique) compositional inverse F pzq,
i.e.,
F p f pzqq “ f pF pzqq “ z.
1
2 1. GENERATING FUNCTIONS AND LAGRANGE’S INVERSION FORMULA

But how can we find this inverse? In fact, there is a quite general and very useful
formula which we shall derive in the following: For that, we consider a slightly
more general problem, which involves an extension of the calculus of formal
power series.
D EFINITION 1.2.1. A formal Laurent series (in some variable z) is a (formal) sum
ÿ
an ¨ zn
něN

with coefficients an P C, for some N P Z: N may be smaller than zero, i.e., there may
be negative powers of z.
We denote the set of all formal Laurent series by Cppzqq

Addition, multiplication and composition for formal Laurent series are defined
exactly as for formal power series.
T HEOREM 1.2.2. A formal Laurent series a pzq possesses a (unique) multiplicative
inverse if and only if a pzq ‰ 0. 

This implies that the algebra of formal Laurent series is, in fact, a field.
In the following we shall always assume that the formal power series f pzq (with
f pzq “ 0) under consideration “starts with z1 , i.e., is of the form
f pzq “ f 1 ¨ z1 ` f 2 ¨ z2 ` ¨ ¨ ¨
f pzq
where f 1 ‰ 0 (we might express this as z ‰ 0): The general case

g pzq “ gm ¨ zm ` gm`1 ¨ zm`1 ` ¨ ¨ ¨ for m ą 1

can be easily reduced to this case.
L EMMA 1.2.3. Consider some formal power series f pzq “ f 1 ¨ z ` f 2 ¨ z2 ` ¨ ¨ ¨ where
f 1 “ f p0q ‰ 0. Then we have for all n P Z
r z ´ ¯
z´1 f n´1 pzq ¨ f 1 pzq “ rn “ 0s .

(Here, we made use of Iverson’s notation: r“some assertion”s equals 1 if “some asser-
tion” is true, otherwise it equals 01.)
` ˘
P ROOF. Observe that for n ‰ 0 we have n ¨ f n´1 pzq ¨ f 1 pzq “ p f n pzqq1 : For
n ą 0, this is a formal power series; and since the coefficient of z´1 equals 0 for
every formal power series, the assertion is true for all n ą 0.
If n “ ´m ă 0, then f n pzq “ c´m ¨ z´m ` ¨ ¨ ¨ ` c´1 ¨ z´1 ` c0 ` c1 ¨ z1 ` . . . and
p f n pzqq1 “ p´mq ¨ c´m ¨ z´m´1 ` ¨ ¨ ¨ ` p´1q ¨ c´1 ¨ z´2 ` c1 ` . . . ,
so the assertion is also true for all n ă 0.
1Iverson’s notation is a generalization of Kronecker’s delta: δ “ ri “ js.
i,j
 1.2. LAGRANGE’S INVERSION FORMULA 3

For n “ 0, we simply compute:

r z r z f `2¨ f ¨z`¨¨¨
1 2
z´1 f ´1 pzq ¨ f 1 pzq “ z´1
f 1 ¨ z ` f 2 ¨ z2 ` ¨ ¨ ¨
r z 1 f `2¨ f ¨z`¨¨¨
1 2
“ z´1
z f1 ` f2 ¨ z ` ¨ ¨ ¨
r z 1 1 ` 2 ¨ f2 ¨ z ` ¨ ¨ ¨
´1 f
“ z ˆ 1 ˙
z f2
1´z ´ ´¨¨¨
f1
loooooooomoooooooon
“:hpzq
r z1 ˆ f2
˙´ ¯
“ z´1 1`2 z`¨¨¨ 1 ` h pzq ` h pzq2 ` ¨ ¨ ¨
z f1
“ 1.

T HEOREM 1.2.4. Let g pzq be a formal Laurent series and f pzq “ f 1 ¨ z ` f 2 ¨ z2 ` ¨ ¨ ¨
be a formal power series with f p0q “ 0, f 1 ‰ 0. Suppose there is an expansion of g pzq
in powers f pzq, i.e.,
ÿ
g pzq “ ck ¨ f k pzq . (1.1)
kěN
Then the coefficients cn are given by
1 r ´1 z 1
cn “ z g pzq ¨ f ´n pzq for n ‰ 0. (1.2)
n
An alternative expression for these coefficients is
r z
cn “ z´1 g pzq ¨ f 1 pzq ¨ f ´n´1 pzq for n P Z. (1.3)

P ROOF. For the first claim (1.2) we take the derivative of (1.1),which gives
ÿ
g1 pzq “ k ¨ ck ¨ f 1 pzq ¨ f k´1 pzq .
k

Now we multiply both sides by f ´n pzq:

ÿ
g1 pzq ¨ f ´n pzq “ k ¨ ck ¨ f 1 pzq ¨ f k´n´1 pzq .
k

Taking the coefficients of z´1 on both sides, by Lemma 1.2.3 we obtain n ¨ cn on

the right–hand side and the claim follows immediately.
For the second claim we multiply (1.1) by f 1 pzq ¨ f ´n´1 pzq, which gives
ÿ
g pzq ¨ f ´n´1 pzq ¨ f 1 pzq “ ck ¨ f k´n´1 pzq ¨ f 1 pzq ,
kěN

and the claim follows, again, by taking the coefficients of z´1 on both sides. 

Now Lagrange’s inversion formula is the following special case of Theorem 1.2.4:
4 1. GENERATING FUNCTIONS AND LAGRANGE’S INVERSION FORMULA

C OROLLARY 1.2.5 (Lagrange’s inversion formula). Let f pzq “ f 1 ¨ z ` f 2 ¨ z2 ` ¨ ¨ ¨

be a formal power series with f 1 ‰ 0, let F pzq be the compositional inverse of f . Then
we have for every k P N:

k r ´k z ´n
Jzn K Fk pzq “ z f pzq for n ‰ 0. (1.4)
n
An alternative expression for these coefficients is
r z
´k´1
n k
Jz K F pzq “ z f 1 pzq f ´n´1 pzq . (1.5)

P ROOF. Since F pzq is the compositional inverse of f pzq, we have Fk p f pzqq “ zk ,

pkq
i.e. (writing Fn :“ Jzn K Fk pzq):
ÿ pkq
Fn ¨ f pzqn “ zk
n

By setting g pzq “ zk in Theorem 1.2.4 and applying (1.2), we obtain

pkq 1 r ´1 z k r ´k z ´n
Fn “ z k ¨ zk´1 ¨ f ´n pzq “ z f pzq ,
n n
which is equivalent to (1.4). In the same way, applying (1.3) gives (1.5). 
E XAMPLE 1.2.6 (Catalan numbers). Let Fn be the number of triangulations of an
pn ` 1q–gon (with vertices labelled 1, 2, . . . , n successively): Setting F1 “ 1 (i.e., as-
suming that there is a single triangulation of a 2–gon), the sequence pFn qně1 starts like
this:
1, 1, 2, 5, 14, 42, 132, 429, 1430, 4862, . . .
For n ą 1 and any such triangulation, consider the triangle containing the edge which
connects the vertices 1 and 2: This triangle “cuts” the pn ` 1q–gon in two parts, namely
a pk ` 1q–gon and an pn ´ k ` 1q–gon, where 1 ď k ď n ´ 1. This leads to the recur-
sion
n´1
ÿ
Fn “ Fk ¨ Fn´k for n ą 1,
k“1

which together with the initial condition F1 “ 1 implies the following

ř functional equa-
tion for the generating function F pzq “ F1 ¨ z ` F2 ¨ z ` ¨ ¨ ¨ “ ně1 Fn ¨ zn :
2

pF pzqq2 “ F pzq ´ z.

This may be rewritten as

z “ F pzq ´ F2 pzq .
Stated otherwise: F pzq is the compositional inverse of the (quite short) formal power
series z ´ z2 . Applying Lagrange’s Inversion formula (1.4) (for k “ 1) to f pzq “ z ´ z2
 1.2. LAGRANGE’S INVERSION FORMULA 5

we obtain:
1 r ´1 z
Jzn K F pzq “ z pz ´ z2 q´n
n
1 r ´1 z ´n
“ z z p1 ´ zq´n
n
1 r n´1 z
“ z p1 ´ zq´n
nˆ ˙
1 ´n
“ p´1qn´1
n n´1
ˆ ˙
1 2n ´ 2
“ .
n n´1
Clearly, Fn “ Cn´1 , i.e., we get (once again) the well–known Catalan–numbers.
 CHAPTER 2

Species: Enumeration of combinatorial objects

The theory of species provides a general framework for labelled and unlabelled
combinatorial objects and their generating functions. One can tackle this theory
very abstractly; however here we would like to approach the theme by means
of a couple of illustrative examples. (Much more material than treated here can
be found in the textbook [1].)

2.1. Motivating Examples

E XAMPLE 2.1.1. The numbers
Bn :“ # ppartitions of rnsq
are called Bell numbers. By considering the block which contains the largest element
n, one immediately obtains the recursion
ÿn ˆ ˙
n
Bn`1 “ ¨ Bk for n ě 0, (2.1)
n´k
k“0
with initial condition B0 “ 1. The sequence pBn qně0 starts like this:
1, 1, 2, 5, 15, 52, 203, 877, 4140, 21147, 115975, . . .
We consider the exponential generating function (see e.g. [4, Section 2.2]) of these
numbers
ÿ zn
B pzq :“ Bn ¨ .
ně0
n!
n
If we multiply (2.1) with zn! and (formally1) sum over all n ě 0 summieren, then we
obtain
˜ ˆ ˙ ¸ n
ÿ ÿ ÿ n ÿ ÿ n
zn n z Bk 1
Bn`1 ¨ “ Bk “ zn .
ně0
n! ně0 k“0
n ´ k n! ně0 k“0
k! pn ´ kq!
loooooomoooooon loooooooooooomoooooooooooon
“
dBpzq “Bpzq¨ez
dz

We can solve the differential equation

dB d log B
“ “ ez
B dz dz
very easily:
log pBq “ ez ` C ùñ B pzq “ ee
z `C
,
1D.h.: The infinite sum is for the time being only a notation for the equations (2.1)! The
equation for the series simply says that the coefficients of zn on the left- and right-hand sides
are equal for all n ě 0.
7
8 2. SPECIES

and from B p0q “ B0 “ 1 we immediately get C “ ´1, thus altogether

B pzq “ ee
z ´1
. (2.2)
At this point one can make the following observation: The exponential function ez
emerges as the exponential generating function egf sets of the number of all subsets of
rns with n elements. Admittedly, this number is combinatorially not very interesting
(it is namly constantly equal to 1 for all n ě 0), but of course we can then also interpret
ez ´ 1 as the exponential generating function egfsets1 of the number of all nonempty
subsets of rns with n elements (this number is merely negligibly more complicated: 1 for
n ą 0 und 0 für n “ 0, oder shorter mit Iverson’s notation: rn ą 0s). The exponential
generating function egfpartitions thus satisfies
` ˘
egf partitions “ egf sets egfsets1 .
And partitions are indeed “sets of nonempty sets”! This certainly might only be an
accidental analogy — but we will see that there is a “system” behind this.
In a partition of rns the ordering of the blocks does not matter — we can thus
assume that the blocks
‚ have their largest elements “marked”,
‚ and that they are ordered corresponding to their “markings”.
For instance, the blocks of the following partition appear in their “canonical”
order like this:
t1,3u Y t2,5,6u Y t9uYt7,10u Y t4,8,11,12u

3 1 6 5 2 9 10 7 12 11 8 4

Here one could have the idea to “forget the numbering”. One could imagine
this graphically for the above example like this:
2 ` 3 `1` 2 ` 4

Without the numbering these “blocks of indistinguishable elements” only de-

pend on the following:
‚ the cardinality of the single blocks
‚ and the ordering of the blocks of different size
Very obviously we can interpret this as compositions of the natural number n.
E XAMPLE 2.1.2. The same idea as in Example 2.1.1 readily leads to the recursion for
the number Kn of compositions of n:
n
ÿ
Kn`1 “ 1 ¨ Kk für n ě 0, (2.3)
k“0
again with the initial condition K0 “ 1. We consider now the (ordinary) generating
function of these numbers ÿ
K pzq :“ Kn ¨ zn .
ně0
 2.2. SPECIES, LABELLED AND UNLABELLED 9

If we multiply (2.3) by zn and (formally) sum over all n ě 0, then we obtain

˜ ¸
ÿ ÿ ÿ n
1
Kn`1 ¨ zn “ 1 ¨ Kk zn “ K pzq ¨ ,
ně0 ně0 k“0
1 ´ z
loooooomoooooon
Kpzq´K0
“ z

which apparently leads to the equation

z
K´1 “ K
1´z
for K pzq; with the solution
1´z z
K pzq “ “ 1` .
1 ´ 2z 1 ´ 2z
The last equation leads immediately to the explicit (and well-known) representation
Kn “ 2n´1 for n ą 0.
Also here one can make the analogue observation to Example 2.1.1: The geometric series
1
1´z emerges as the (ordinary) generating function gfsets of the number of all multisets
with n elements, which are all equal to ˝ (this number is again constantly equal to 1 for
1
all n ě 0), and 1´z ´ 1 emerges as the (ordinary) generating function gfsets1 of the
number of all nonempty sets with n elements, which are all equal to ˝ (this number is
again rn ą 0s).
Since the elements are not labelled here (and are therefore “identical”), we are actu-
ally not dealing with sets but with (very simple — after all there is only one element
which can occur arbitrarily often) multisets: To make the notation not too complicated,
we keep the term sets. Since the elements are all identical, their order is irrelevant:
(Multi-)sets of indistinguishable elements are thus the same “combinatorial objects”
as sequences! And for the (ordinary) generating function gfcompositions indeed again
` ˘
gfcompositions “ gfsequences gfsets1
holds. And indeed, also compositions are “sequences of nonempty sets” (in the sense
that there is only one (multi-)set mit k indistinguishable elements, which we can con-
fidently identify with the number k). We thus see again the same “analogy” as in
Example 2.1.1.

2.2. Species, labelled and unlabelled

It is commonly known that generating functions can be thought of in the follow-
ing way: Given a family A of combinatorial objects, specify a weight function
ω : A Ñ R (R is a ring, ω pOq typically assumes the values zn , where n is a
“characteristic value” of O, such as size, length, number of parts, etc.) and form
the formal sum ÿ
ω pOq ,
OPA
which we (also formally) can write out as follows:
ÿ
c x ¨ x,
xPR
10 2. SPECIES

where c x is the number of objects in A with assigned weight x: Of course, the

number c x should be defined for any particular (finite) x, in order that the sum
makes sense. This observation immediately leads to the following definition:
D EFINITION 2.2.1. We denote a family A of combinatorial objects, of which every
single object consists of a finite number n of “atoms” (and which moreover may possess
a more or less complicated “structure”), where the atoms
‚ either may be labelled by the first n natural numbers [see Example 2.1.1],
‚ or may be completely indistinguishable (“unlabelled”) [see Example 2.1.2],
as labelled species or as unlabelled species (generally: species). The number n of
atoms, an object A of a species consists of, is called the size of A: }A} “ n.
This can be expressed a bit more abstractly as follows: To a species A belongs a size
function #A : A Ñ N0 , which is given by #A pAq “ }A}.
For a species we always assume that the subfamilies
An :“ tA P A : }A} “ nu
for all n P N are finite, i.e.,
an :“ |An | ă 8 @n P N.
In particular is a species A always countable.
Of course, the size function delivers a weight function
ωA pAq :“ z}A} for unlabelled species,
z}A}
ωA pAq :“ for labelled species.
}A}!
The generating function of a (unlabelled or labelled) species A is then the formal
power series
ÿ ÿ ÿ
gf A pzq “ ωA pAq “ z}A} “ an zn ,
APA APA ně0
ÿ ÿ z}A} ÿ zn
egf A pzq “ ωA pAq “ “ an
}A}! ně0 n!
APA APA

2.3. Unlabelled species and the enumeration of trees

E XAMPLE 2.3.1. A totally simple species consists only of a single object, which itself
consists of a single atom:
Atom “ t˝u
the generating function is
gfatom pzq “ z.
E XAMPLE 2.3.2. Another simple species is (multi-)sets (of atoms, which are indistin-
guishable as unlabelled objects):
sets “ tH, t˝u , t˝, ˝u , . . . u
the generating function is
1
gfsets pzq “ 1 ` z ` z2 ` ¨ ¨ ¨ “ .
1´z
 2.3. UNLABELLED SPECIES AND THE ENUMERATION OF TREES 11

A simple “modification” would be the species sets1 of nonempty sets (of atoms); the
generating function is then
z 1
gfsets1 pzq “ z ` z2 ` z3 ` ¨ ¨ ¨ “ “ ´ 1.
1´z 1´z
R EMARK 2.3.3. Since the atoms here are unlabelled (i.e., not distinguishable), their
ordering does not matter: Therefore the species sets (of atoms) is equal to the species
sequences (of atoms). This however depends on the indistinguishability of the el-
ements (atoms): sequences of distinguishable elements are of course not the same
combinatorial objects as (multi-)sets of these elements!
D EFINITION 2.3.4. A tree is a simply connected graph without circles.
A rooted tree is a tree with a special vertex, which is referred to as root. A rooted
tree always has a “natural orientation” (“from the root away”), it thus emerges as a
directed graph: Vertices with outgoing degree 0 are called leaves, all other vertices
are called inner vertices.
Every tree is of course planar (is embeddable in the plane such that two arbitrary em-
bedded edges intersect in at most one vertex): A rooted tree in a fixed embedding in
the plane, where the order of the subtrees is relevant, is called planar oder ordered
rooted tree.
The size of a tree (in the sense of Definition 2.2.1) can be given by the number of its
vertices: Trees are thus a further example of species.
For instance, there are the following 2 planar rooted trees with 3 vertices:

With 4 vertices there are the following 5 planar rooted trees, although the sec-
ond and the third of them would be viewed as the same “normal” (i.e., “not-
planar”) rooted tree:

R EMARK 2.3.5 (Here always: rooted trees!). Especially for computer sciences algo-
rithms (e.g., search algorithms) and data structures which can be “composed” of trees
are interesting objects. Since these trees often have a “distinguished vertex” (a root), we
will in the following always consider rooted trees (if not explicitly speficied otherwise)
but for simplicity talk about trees.
As we have seen in the introductory Example 2.1.2, the combinatorial construc-
tion “form finite sequences of objects” transfers to the algebraic construction
“insert the generating function of the objects into the geometric series”. We
shall examine this “transfer”
combinatorial construction Ñ algebraic construction
12 2. SPECIES

in further examples:
D EFINITION 2.3.6 (Disjoint union). Given two species A and B , we can consider
their disjoint union C “ A Y
9 B ; the corresponding size function is
#
# pxq x P A,
#C pxq :“ A
#B pxq x P B .

Apparently this construction transfers to generating functions as follows:

gf C “ gf A ` gf B .
D EFINITION 2.3.7 (Cartesian Product). Given two species A and B , we can consider
their cartesisian product C “ A ˆ B ; the corresponding size function is
#C ppx, yqq :“ #A pxq ` #B pyq .
Apparently this construction transfers to generating functions as follows:
gf C “ gf A ¨ gf B .
We visualize this graphically by means of the objects of size 3 in the product sets ˆ
planarrootedtrees:

řn
` ˘ k“0 m k ¨ wn´k
H, (Here: n “ 3.)

m 0 ¨ w3 “ 1 ¨ 2 “ 2
` ˘
H,

` ˘
t˝u, m 1 ¨ w2 “ 1 ¨ 1 “ 1
` ˘
t˝, ˝u, m 2 ¨ w1 “ 1 ¨ 1 “ 1
` ˘
t˝, ˝, ˝u, m 3 ¨ w0 “ 1 ¨ 1 “ 1

A special case is the k–fold cartesian product of a species A whose objects are sequences
of length k of objects from A:

A ˆ A ˆ ¨ ˆ A : gf Ak “ pgf A qk .
Ak “ loooooooomoooooooon
k times

If k “ 0, we obtain the single-element species A0 “ tǫu which only contains the empty
sequence ǫ “ pq with }ǫ} “ 0.
For instance, we can view the combinatorial objects “compositions of a number n with
exactly k parts” as species K (with size function n): This apparently corresponds to
ˆ ˙k
k
` ˘k z
K “ sets1 , thus gf K “ gfsets1 “ .
1´z
 2.3. UNLABELLED SPECIES AND THE ENUMERATION OF TREES 13

D EFINITION 2.3.8 (Sequences). Given a species A which does not contain any object
of size 0, we can consider arbitrarily long finite sequences of objects from A, i.e., the
species A‹ ď
A‹ :“ Ak
kě0
According to the previous examples this transfers to generating functions as follows:
ÿ 1
gf A‹ “ pgf A qk “ .
1 ´ gf A
kě0
We have already seen a concrete examples for this in Example 2.1.2:
compositions “ sets‹1 .
R EMARK 2.3.9. Why did we have to assume here that A contains no objects of size 0?
Because otherwise A‹ would contain infinitely many objects of size 0, in contradiction
to our assumption for species:
} pq } “ } pǫq } “ } pǫ, ǫq } “ } pǫ, ǫ, ǫq } “ ¨ ¨ ¨ “ 0.
A PPLICATION 2.3.10. If one removes the root from a nonempty rooted tree W, a
(possibly empty) “list” of connected components arises which themselves consist of
nonempty rooten trees: If the rooted tree W is
‚ ordered, then one should also view the “list” as ordered (i.e., as sequence) —
then from the (ordered) sequence one can of course again uniquely reconstruct
the original (ordered) rooted tree W,
‚ unordered, then one should also view the “list” as unordered (i.e., as multi-
set) — then from the (unordered) multiset one can of course again uniquely
reconstruct the original (unordered) rooted tree.
The following graphic illustrates this simple thought:
´ ¯
n
1
, 2 , 3 ,...,
n
planar rooted tree
1 2 3

! )
2 n
, , 1
,..., 3
W
“normal” rooted tree

If we denote the number of planar rooted trees with n vertices with f n , then we
obtain by definition the generating function of the species of planar rooted trees
with at least one vertex:
ÿ
f pzq “ f n ¨ zn “ z ` z2 ` 2z3 ` 5z4 ` ¨ ¨ ¨
ně1
From the Application 2.3.10 we first get
planarrootedtrees 1 “ atom ˆ planarrootedtrees ‹1
and hence then the following functional equation:
´ ¯ z
f “ z ¨ f0 ` f1 ` f2 ` ¨¨¨ “ ðñ f 2 ´ f ` z “ 0.
1´ f
14 2. SPECIES

This quadratic equation for f has two solutions of which the following has the
“right” (positive) coefficients:
?
1 ´ 1 ´ 4z ÿ 1 ˆ2n˙
f pzq “ “ zn`1 . (2.4)
2 ně0
n ` 1 n

Before we start the next venturesome calculation, we remind ourselves of the

...
Transfer principle for formal power series

If an identity for analytic functions also makes sense for the corresponding for-
mal power series, then it is automatically also an identity for formal power
seriesn.
Conversely: If an identity for formal power series also makes sense for the cor-
responding analytic functions (this means that there exists a nontrivial common
radius of convergence for all involved series), then it is automatically also an
identity for analytic functions.

E XAMPLE 2.3.11 (Sets). Let A be a species which does not contain any object of size 0.
If we in sequences of A “forget about the order”, we obtain the species C “ sets pAq
(as said: actually multisets — elements can occur repeatedly): the size function for an
object
tA1 , A2 , . . . , Ak u P C
is given (just as in sequences) by
k
ÿ
#C ptA1 , A2 , . . . , Ak uq “ #A pAi q .
i“1

Thus we obtain for the generating function

ÿ ! )
#C pCq
gf C pzq “ z A1 , . . . , A1 ,looooomooooon
Ð C“ looooomooooon A2 , . . . , A2 ,...
CPC k1 k2
¨ ˛ ¨ ˛
ÿ ´ ¯k1 ÿ ´ ¯k 2
“˝ z}A1 } ‚¨ ˝ z}A2 } ‚¨ ¨ ¨
k1 ě0 k2 ě0
ź ÿ´ ¯k
“ z}A}
APA kě0
źˆ 1
˙ ai
“ Ð ai :“|tAPA:}A}“iu|ă8
iě1
1 ´ zi
´ 2 3
¯ ´ 4 6
¯
a1 z` z2 ` z3 `¨¨¨ `a2 z2 ` z2 ` z3 `¨¨¨ `¨¨¨ ř zn
“e Ðlog 1
1´z “ n
˜ ` ˘ ` ˘ ¸
gf A z2 gf A z3
“ exp gf A pzq ` ` `¨¨¨ .
2 3
 2.3. UNLABELLED SPECIES AND THE ENUMERATION OF TREES 15

For instance we obtain the species numberpartitions (i.e., “unordered” compositions)

as sets psets1 q.
E XAMPLE 2.3.12 (Composition). Given two species A and B such that B contains
no empty object (of size zero; compare with Remark 2.3.9):
EB P B : #B pBq “ 0.
Then we construct the composition of species A pB q as follows: We imagine all the
atoms of an object from A “blown up” and insert in each such atom an object from
B . The following graphic visualizes the construction for A “ rootedtrees und B “
sets1 :

{ } { }
{ }
B:
{ }
{ } { }
A pB q:

A:
{ }

However, unfortunately this construction does not translate to the composition of the
respective generating functions! Foir instance, we have:
sets psets1 q “ numberpartitions,
but
ź 18
1
z ‰ .
1 ´ 1´z 1 ´ zi
i“1
(The problem is apparently: the construction as such is not unique — in the above visu-
alization we could also have distributed the objects from B differently among the atoms
of the objects from A. We would need to, say, first construct all possible “insertions
of objects in atoms”, and then figure out which of those are actually different!)

2.3.1. Enumeration of binary trees. We now want to show the usefulness

of the viewpoint of species on typical enumeration problems.
D EFINITION 2.3.13. A rooted tree with the property that every vertex has outdegree
ď2 is called binary tree. With other words: On each vertex of a binary tree two subtrees
are hanging which may also be empty. (More generally, a rooted tree with the property
that evey vertex has outdegree ď m is called an m–ary tree.) An ordered binary tree
is a binary tree where at every inner vertex the order of its two subtrees matters (i.e.,
there is always a left and a right subtree; even if one of the subtrees is empty! Note
that a binary ordered tree is not the same concept as a planar binary tree!)
A binary tree with the property that each inner vertex has outdegree “ 2 is called
complete binary tree. (More generally, a rooted tree with the property that each inner
vertex has outdegree m is called a complete m–ary tree.)
16 2. SPECIES

We would like to determine the number of complete binary ordered trees with
n inner vertices: This is the same as the number of ordered binary trees with n
vertices. The following graphic shows the objects of this species B with size
(number of vertices) n “ 3:

The “combinatorial decomposition” of this species is simple: An ordered binary

tree is
‚ either empty
‚ or consists of a root (an atom) and a pair of subtrees.
Formally thus:
´ ¯
2
B “ tǫu Y9 atom ˆ B .
This translates to the following equation for the generating function:
gf B “ 1 ` z ¨ pgf B q2 .
This quadratic equation for gf B has again two solutions of which
?
1 ´ 1 ´ 4z ÿ 1 ˆ2n˙
gf B “ “ zn (2.5)
2z ně0
n ` 1 n

is the “right” one.

This simple idea of course also works generally for the species of unordered
m–ary trees Tm :
´ ¯m
gf Tm “ 1 ` z ¨ gf Tm . (2.6)
For m ą 2 it is however simpler to determine the coefficients of the generating
function with the Lagrange inversion formula; see Corollary 1.2.5, which we
shall repeat here:
C OROLLARY 2.3.14 (Lagrange–inversion formula). Let f pzq be a formal power se-
ries with vanishing constant term and F pzq the corresponding compositional inverse
series (i.e., F p f pzqq “ f pF pzqq “ z). Then one has
n k k r ´k z ´n
Jz K F pzq “ z f pzq for n ‰ 0, (2.7)
n
or r z
´k´1
n k
Jz K F pzq “ z f 1 pzq f ´n´1 pzq. (2.8)
´ ¯
After all, (2.6) just means that gf Tm ´ 1 is the compositional inverse of
z
.
p1 ` zqm
 2.4. BIJECTIVE COMBINATORICS ON ROOTED TREES 17

According to (2.7) one thus has for n ą 0

ˆ ˙
n 1 r ´1 z p1 ` zqm¨n
Jz K gf Tm “ z Ðbinomial theorem!
n zn
ˆ ˙
1 m¨n
“ Г
pm¨nqpm¨n´1q¨¨¨pm¨n´n`2q
n n´1 n¨pn´1q!
ˆ ˙
1 m¨n
“ .
n ¨ pm ´ 1q ` 1 n
Exercise 1: Show that the number of all unlabelled ordered nonempty rooted trees with n vertices,
where every inner vertex has 2 or 3 branches, equals
ˆ ˙ˆ ˙
1ÿ n j
.
n j 3j ´ n ` 1
j

Hint: Find an equation for the generating function and use Langrange’s inversion formula.

Exercise 2: How many ways are there to (properly) parenthesize n pairwise non–commuting elements
of a monoid? And how does this number change if the n elements are pairwise commuting?
For example, consider 6 non–commuting elements x1 , x2 , . . . , x6 . Two different ways to parenthesize
them properly would be
ppx2 x5 qppx1 px4 x6 qqx3 qq and ppx3 px1 px4 x6 qqqpx5 x2 qq.
However, these would be equivalent for commuting elements.

Hint: Translate parentheses to labelled binary trees: The outermost pair of parentheses corresponds
to the root, and the elements of the monoid correspond to the leaves.

2.4. Bijective combinatorics on rooted trees

It is noticeable that the generating functions for nonempty planar rooted trees
(2.4) and ordered binary trees (2.5) are equal up to factor z, and the numbers of
these objects with fixed size (number of vertices) is essentially a Catalan number
` ˘
1 2k
Ck “ k`1 k : This can also be shown using bijections.

2.4.1. Dyck paths of length 2n.

D EFINITION 2.4.1. The integer lattice is the infinite directed graph with vertex set
Z ˆ Z and edge set (step set)

tppx, yq , px ` 1, y ` 1qq : px, yq P Z ˆ Zu

Y tppx, yq , px ` 1, y ´ 1qq : px, yq P Z ˆ Zu .
(The directed edges hence always go either “one step right–up”: u or “one step right–
down”: d.)
A path in the integer lattice which starts in p0, 0q and ends in p2n, 0q and which never
comes below the x–axis, is called Dyck path of length 2n.
18 2. SPECIES

4
u
3 d
u u u u
2 d d d d
u u u
1 d d d
u
0 d
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

It is a well known fact

` ˘that the number of Dyck paths of length 2n is equal to the Catalan
1 2n
number Cn “ n`1 n (see e.g. [4, Anhang A.2.2]).

2.4.1.1. Planar trees Ø Dyck paths. From the graphic it becomes apparent
that a Dyck path can be coded as as word of length 2n over the alphabet tu, du,
where
‚ the number of u’s is equal to the number of d’s, i.e. n,
‚ and in each “initial part” of the word at least as many u’s occur as d’s.
For the above graphic this “code word” is:
uududuududuudduddd.
This immediately translates to a coding for ordered trees (plant louse coding).
Each tree on n ` 1 vertices has exactly n edges: We imagine a “plant louse”
which sequentially walks along the edges of the ordered rooted tree (always
starting with the left subtree).

ud
u

ud
d
udu

ud
ud

ud

ud

2.4.1.2. Ordered binary trees Ø Planar trees. The corresponding rotation corre-
spondence is best made clear in a graphic:
rotate by 45˝ : new root: new edges:

2.5. Unordered rooted trees

There are the following 9 unordered rooted trees with 5 vertices:

According to Application 2.3.10, by “tearing out the root” from an unordered

nonempty rooted tree, a (multi–)set of rooted trees arises, i.e., for the species of
unordered rooted trees T one has:
T “ t˝u ˆ sets pT q . (2.9)
 2.6. LABELLED SPECIES 19

This transfers as follows to the generating function (see Example 2.3.11):

ˆ ˙t 1 ˆ ˙ t2 ˆ ˙ t3
1 1 1
gf T pzq “ z ¨ ¨ ¨ ¨ ¨ (Cayley)
1´z 1 ´ z2 1 ´ z3
ˆ ´ ¯ 1 ´ ¯ ˙
1 2 3
“ z ¨ exp gf T pzq ` gf T z ` gf T z ` ¨ ¨ ¨ (Polya).
2 3
From the formula of Cayley the coefficients ti can be recursively computed:
gf T pzq “ t1 z ` t2 z2 ` t3 z3 ` ¨ ¨ ¨ Ðp´t
n q“p´1q p n q
n t`n´1
ˆ ˆ ˙ ˙ ˆ ˆ ˙ ˙
t1 ` 1 2 2 t2 ` 1 4
“ z 1 ` t1 z ` z ` ¨ ¨ ¨ ¨ 1 ` t2 z ` z `¨¨¨ ¨¨¨
2 2
ˆˆ ˙ ˙ ˆˆ ˙ ˙
2 t1 ` 1 3 t1 ` 2
“ z ` t1 z ` ` t2 z ` ` t1 t2 ` t3 z4 ` ¨ ¨ ¨
2 3
Concretely one obtains:

gf T pzq “ z ` z2 ` 2z3 ` 4z4 ` 9z5 ` 20z6 ` 48z7 ` 115z8

` 286z9 ` 719z10 ` 1842z11 ` ¨ ¨ ¨

2.6. labelled species

E XAMPLE 2.6.1. The labelled species atom again only consists of a single element, for
which there is clearly just one labeling: “˝ is labelled by 1”. The generating function is
hence again
z
egf atom pzq “ “ z.
1!
E XAMPLE 2.6.2. A further species is sets (of atoms): In the labelled objects the atoms
are distinguishable but, since in a set the order does not matter, there is exactly one
labelled set with n elements:
sets “ tH, t1u , t1, 2u , . . . u ,
the (exponential) generating function thus is
z z2
egf sets pzq “ 1 ` ` ` ¨ ¨ ¨ “ exp pzq .
1! 2!
The (exponential) generating function of sets1 then is
z z2 z3
` ` ` ¨ ¨ ¨ “ exp pzq ´ 1.
1! 2! 3!
E XAMPLE 2.6.3. Since the atoms here are labelled, the (labelled) species sets (of
atoms) is not anymore equal to the (labelled) species sequences (of atoms): The lat-
ter is simply einfach equal to the (labelled) species permutations with the generating
function
1! 2! 1
egfpermutations pzq “ 1 ` z ` z2 ` ¨ ¨ ¨ “ .
1! 2! 1´z
20 2. SPECIES

The species cycles of the cyclic permutations (i.e. of the permutations whose unique
cycle decomposition consists of a single cycle) apparently has the generating function
0! 1! 2! 1
egfcycles pzq “ z ` z2 ` z3 ` ¨ ¨ ¨ “ log .
1! 2! 3! 1´z
Just as with the unlabelled species there is also here a “transfer”
combinatorial construction Ñ algebraic construction.
E XAMPLE 2.6.4 (Disjoint union). Given two labelled species A and B , we can (just
as with the unlabelled species) consider their disjoint union C “ A Y
9 B , and also here
one has for the generating functions:
egf C “ egf A ` egf B .
E XAMPLE 2.6.5 (Product). Given two labelled species A and B , we define their prod-
uct C “ A ‹ B as follows: We first consider the cartesian product A ˆ B . From
`m`n ˘ a pair of
objects pA, Bq P A ˆ B (where }A} “ m and }B} “ n) we construct n ddifferent
labelled objects according to the following rule:
‚ Consider all decompositions of the set rm ` ns in two disjoint parts
rm ` ns “ X Y
9 Y with |X| “ m, |Y| “ n,
‚ For every such decomposition replace the numbers from A resp. B according
to the order-preserving bijections2
rms Ñ X
rns Ñ Y
We illustrate this by an example:
2 3 1
pA, Bq “ 1 2
}A}`}B}“3`2“5:
2 3 4
tt1,2,3u ,t4,5uu Ñ 1 5
2 4 3
tt1,2,4u ,t3,5uu Ñ 1 5

4 5 1
tt3,4,5u ,t1,2uu Ñ 3 2

The size function is of course again given by

#C ppA, Bqq :“ #A pAq ` #B pBq .
This construction indeed transfers again to generating functions:
˜ ¸
ÿ ÿ n ÿ cn
n! ak bn´k
egf A ¨ egf B “ zn “ zn “ egf C ,
ně0
n! k! pn ´ kq! ně0
n!
k“0
2Between any two finite ordered sets X and Y with |X| “ |Y| there of course exists exactly
one order-preserving bijection.
 2.6. LABELLED SPECIES 21
řn ` n˘
since cn “ k“0 k ¨ ak ¨ bn´k by construction.
For instance:
‚ sets ‹ sets “ subsets of a given set, which one could also interpret as
“functions to t0, 1u”. That the cardinality of the power set of an n–element set
equals 2n is now also readily seen from the generating function:
ÿ 2n
pegfsets pzqq2 “ pez q2 “ e2z “ zn .
ně0
n!

‚ sets1 ‹ sets1 are thus to be viewed as “surjective functions to t0, 1u”.

A special case is again the k–fold product of a (labelled) species A:

A ‹ A ‹ ¨ ¨ ¨ ‹ A : egf Ak “ pegf A qk .
Ak :“ loooooooomoooooooon
k times

For instance:
‚ setsk “ ordered decomposition in k blocks, which one could also in-
terpret as “functions to rks”.
‚ sets1 k are thus to be viewed as “surjective functions to rks”.
E XAMPLE 2.6.6 (Sequences). Also for a labelled species A which contains no object
of size 0 we can consider arbitrary finite sequences of objects from A, i.e. the species A‹
ď
A‹ :“ Ak .
kě0

Again this transfers to generating functions as follows:

ÿ 1
egf A‹ “ pegf A qk “ .
1 ´ egf A
kě0

For instance, we obtain for the species of surjections τ : rns Ñ rks (with }τ} “ n)
surjections “ sets1 ‹ .
It immediately follows: The exponential generating function of the numbers3
n
ÿ
Sn,k ¨ k!
k“0

(where Sn,k denotes the Stirling numbers of the second kind) thus is
1 1
“ .
1 ´ pe z ´ 1q 2 ´ ez
E XAMPLE 2.6.7 (Sets). If in an object from A‹ we “forget about the order”, we obtain
the species
C “ sets pAq .
3These are the numbers of all surjections from rns Ñ rks, where k “ 0, 1, . . . , n is arbitrary.
22 2. SPECIES

By the labeling the k A–objects of an object from Ak is anyway distinguishable (i.e.: a

multiset of labelled objects is always a set); for each set with k A–objects there are k!
ordered k–tuples of these objects in Ak . Thus we obtain for the generating function
ÿ pegf pzqqk
egf C pzq “ A
“ eegfA pzq .
k!
kě0

For instance we obtain the species

setpartitions “ sets psets1 q ;
the exponential generating function of the Bell numbers hence is
ee ´1
z

(see also the introductory Example 2.1.1).

R EMARK 2.6.8. As we have seen, in many cases there is a clear difference between la-
belled and unlabelled species. But for ordered trees nothing new arises by the labeling,
because the vertices of ordered trees anyway have a “natural ordering” (root, successor
of the root from left to right, successor of the successor of the root from left to right, . . . ),
such that therefore holds:
# punlabelled ordered treesq ¨ n! “ # plabelled ordered treesq
(where n denotes the number of the vertices), i.e., the generating functions for unla-
belled and labelled ordered trees are equal.
E XAMPLE 2.6.9 (Composition). Given two species A and B , such that B contains no
“empty object” (with size zero):
EB P B : #B pBq “ 0.
Then we can construct the composition of species A pB q as follows:
‚ Take a k–tuple pB1 , B2 , . . . , Bk q of objects from B k .
‚ Order the k–tuple by the smallest number in each component Bi : De facto we
thus consider k–element sets (compare with Example 2.6.7)
(
B j1 ă B j2 ă ¨ ¨ ¨ ă B jk ,
i.e. a species with the generating function

pegf B pzqqk
.
k!
‚ Take an A–object A of size #A pAq “ k and imagine the atoms of A “blown
up”: In the labelled atoms now insert in the “canonical order” the objects
B j1 , ¨ ¨ ¨ , B jk (i.e., in the A–atom with number 1 one inserts B j1 , in the A–atom
with number 2 one inserts B j2 , etc.).
‚ The size of the thus produced object from C P C :“ A pB q is then of course
given by
#C pCq “ #B pB1 q ` ¨ ¨ ¨ ` #B pBk q .
 2.6. LABELLED SPECIES 23

Here the construction indeed translates to the composition of the generating functions:
ÿ ÿ 1 ÿ z}Bj1 }`¨¨¨`}Bjk }
egf C pzq “ ` ˘
k! !
kě0 APA
) }B j1 } ` ¨ ¨ ¨ ` }B jk } !
}A}“k B j1 ,...,B j PB k
k

ÿ pegf B pzqqk
“ ak “ egf A pegf B pzqq .
k!
kě0

Exercise 3: Develop a theory for weighted generating functions (for labelled and unlabelled species).
I.e., let A be some species with weight function ω which assigns to every object A P A some element
in a ring R (for instance, R “ Z rys, the ring of polynomials in y with coefficients in Z). So the
generating function to be considered is
ÿ
z} A} ¨ ω pAq .
A PA
How should we define the weight function for sums, products and composition of species, so that the
corresponding assertions for generating functions remain valid?

Exercise 4: Let f pm, nq be the number of all paths from p0, 0q to pm, nq in N ˆ N, where each single
steps is either p1, 0q (step to the right) or p0, 1q (step upwards) or p1, 1q (diagonal step upwards).
Use the language of species to show that
ÿ 1
f pm, nq x m yn “ .
1´x´y´x¨y
m,ně0

E XAMPLE 2.6.10 (Cayley’s theorem). Let T denote the species of nonempty unorderd
labelled rooted trees. With the same consideration as in Section 2.5 we obtain the same
“species equation” (2.9) also for the labelled rooted trees, thus
T “ t˝u ˆ sets pT q .
This equation here translates according to Example 2.6.7 resp. 2.6.9 to
egf T pzq “ z ¨ eegfT pzq ,
hence egf T pzq is the compositional inverse of z ¨ e´z . We can very easily determine
their coefficients using the Lagrange inversion formula (Corollary 1.2.5: (2.7) for k “
1):
1 r ´1 z ` ˘´n
Jzn K gf T pzq “ z z ¨ e´z
n
1 r ´1 z ´n nz
“ z z e
n
1 r n´1 z nz
“ z e
n
1 nn´1
“
n pn ´ 1q!
for n ą 0. This means, the number of labelled unordered rooted trees on n ą 0 vertices
is
tn “ nn´1 .
24 2. SPECIES

If we “forget” about the root, we then obtain Cayley’s formula for the number of all
labelled trees (not rooted trees!):
# plabelled treeson n verticesq “ nn´2 . (2.10)
Exercise 5: Determine the number of all unlabelled ordered binary rooted trees with n vertices and
k leaves.
Hint: Consider the generating function in 2 variables z and y, where every rooted tree W with
n vertices and k leaves is assigned ω pWq :“ zn yk . The following picture shows these trees for
n “ 1, 2, 3, 4:

n“1:

n“2:

n“3:

n“4:

. . . and these 7 trees, vertically reflected.

I.e., the first terms of the generating function are:

ÿ ´ ¯ ´ ¯
T pz, yq :“ ω pWq “ z ¨ y ` z2 ¨ 2y ` z3 y2 ` 4y ` z4 6y2 ` 8y ` ¨ ¨ ¨
W
Find an equation for this generating function T, from which the series expansion can be derived.

Exercise 6: Determine the number of all labelled unordered rooted trees with n vertices and k leaves.
Hint: Consider the exponential generating function in 2 variables z and y (as in Exercise 5) and use
Lagrange’s inversion formula.

2.6.1. Combinatorial proof of Cayley’s theorem. The Prüfer correspondence

delivers a bijection between
‚ unordered, labelled rooted trees with n vertices,
‚ and sequences of numbers of length n ´ 1 from rns.
For a given rooted tree this works as follows: Look for the leaf with the smallest
number (labeling), note down the number of the inner vertex, on which this leaf
is hanging, and remove the leaf: Repeat this until only the root is left. In a small
example:
7 1 11 6
4 10 3 5
9 8 12
10, 9, 3, 3, 10, 2, 8, 2, 3, 12, 2
2
 2.6. LABELLED SPECIES 25

Conversely one recovers from the Prüfer code

pp pv1 q , p pv2 q , . . . , p pvn´1 qq
the corresponding rooted tree again by proceeding as follows: Let v1 be the
smallest number from rns, which does not appear in the Prüfer code. The num-
bers which do not appear in the Prüfer code correspond to the leaves in the
original rooted tree; v1 is thus the number of the first removed leaf. By con-
struction this leaf was hanging on the vertex with the number p pv1 q, we thus
note down
pv1 , p pv1 qq (v1 is hanging on p pv1 q)
and continue with the remaining Prüfer code
pp pv2 q , p pv3 q , . . . , p pvn´1 qq
recursively: Let v2 be the smallest number in rns z tv1 u which does not appear
in the remaining Prüfer code; we thus note down
pv2 , p pv2 qq (v2 is hanging on p pv2 q),
etc. For the above example the following sequence of notations arises:
pp1,10q,p4,9q,p5,3q,p6,3q,p7,10q,p9,2q,p10,8q,p8,2q,p11,3q,p3,12q,p12,2qq,

and it is clear that the root has number 2: With this the rooted tree can be
uniquely reconstructed.

Exercise 7: Prove Cayley’s formula (the number of labelled trees on n vertices equals nn´2 ) as
follows: Take a labelled tree on n vertices and tag two vertices S and E. View S and E as the
starting point and ending point of the unique path p connecting S and E in the tree. Now orient
all edges belonging to p “from S to E”, and all edges not belonging to p “towards p”. Now travel
along p from S to E and write down the labels of the vertices: Whenever a new maximal label is
encountered, close a cycle (by inserting an oriented edge from the vertex before this new maximum
to the start of the “current cycle”) and start a new cycle. Interpret the resulting directed graph as a
function rns Ñ rns (i.e., a directed edge from a to b indicates that the function maps a to b).

Exercise 8: Show that the number of all graphs on n vertices, m edges and k components equals
the coefficient of un αm βk {n! in
˜ ¸β
ÿ n
pn2 q u
p1 ` αq .
n!
n ě0

Hint: Find a connection between the generating function of all labelled graphs (weight ω pGq :“
u|V pGq| α|EpGq| ) and the generating function of all connected labelled graphs.

Exercise 9: Show that the number of labelled unicyclic graphs (i.e., connected graphs with exactly
one cycle) on n vertices equals
n ˆ ˙
1ÿ n
j! nn´1´ j .
2 j
j “3

Hint: Find a representation as a composition of species.

26 2. SPECIES
ř n ´1
Exercise 10: Let T puq “ ně0 pn ` 1q un {n!. Prove the identity
T j puq ÿ ul
“ pl ` jql .
1 ´ uT puq l!
l ě0

Hint: For a bijective proof consider the species W of labelled trees, where the vertex with the largest
label (i.e.: n, if the tree has n vertices) is tagged as the root, but this label is erased, and the root
does not contribute to the size of the tree (i.e: if t has n vertices (including the root), then we have
}t}W “ n ´ 1): Obviously, T “ gfW .
Now consider functions f : rls Ñ rl ` js. Visualize such function f as a directed graph with vertex
set rl ` js and directed edges px, f pxqq. The following graphic illlustrates this for the case l “ 20,
j “ 2 and
p f pnqqln“1 “ p2, 6, 2, 5, 2, 7, 8, 5, 4, 4, 7, 21, 21, 18, 16, 18, 22, 15, 15, 15q :

9
1 3 12
4 10
2 13 21
6 5 14

7 8 16 18

19
11
20 15 17 22

The components of this graph are rooted trees or unicyclic graphs; all vertices in rl ` js z rls appear
as roots of corresponding trees.

Exercise 11: The derivative A1 of a labelled species A is defined as follows: Objects of species A1
with size n ´ 1 are objects of A with size n, whose atoms are numbered from 1 to n ´ 1 (not from
1 to n), such that there is one atom without a label. A typical element of Sequences1 is
p3, 1, 2, 5, ˝, 4q ,
where ˝ indicates the unlabelled atom.
Show: The generating function of A1 is precisely the derivative of the generating function of A.
Moreover, show:
(1) pA ` B q1 “ A1 ` B 1.
(2) pA ‹ B q1 “ `A1 ‹ B `
˘ B 1 ‹ A.
1
(3) pA ˝ B q “ A1 ˝ B ‹ B 1.
These equations are to be understood as size–preserving bijections.

Exercise 12: Show the following identities for labelled species:

(1) oPar1 “ oPar2 ‹ Sets, where oPar denotes the species of ordered set partitions (i.e., the
order of the blocks of the partitions matters).
(2) Polyp1 “ Sequences pAtomq ‹ Sequences p2Atomq , where Polyp denotes the species
` ˘
Cycles Sequencesě1 ;
i.e., an object of Polyp is a cycle, where there is a nonempty sequence attached to each
atom of the cycle.
 2.7. PÓLYA’S THEOREM 27

Exercise 13: Let A be the (labelled) species of (unordered) rooted trees, U the (labelled) species of
trees (without root) and F the (labelled) species of rooted forests. Show the following equations:
(1) A1 “ F ‹ Sequences pAq,
(2) U 2 “ F ‹ A1,
` ˘2 ` ˘2
(3) A2 “ A1 ` A1 ‹ Sequences pAq.

Exercise 14: Compute all derivatives of Sets2 and of Sequences.

2.7. Pólya’s theorem

We start with two motivating examples.
E XAMPLE 2.7.1. Given a cube, how many essentially different possibilities exist to
colour its sides with the colours red and blue? (“Essentially different” shall mean here:
We view two colourings, which turn into each other by a rotation, as equal.)
There are in total 10 possibilities: all sides red, all sides blue, one side red, one side blue,
2 opposite sides red, 2 opposite sides blue, 2 adjacent sides blue, 2 adjacent sides red, 3
in a corner colliding sides blue, 3 in a row connected sides blue.
E XAMPLE 2.7.2. Given pearles in k colours. How many possibilities exist to form a
necklace of length n, where 2 necklaces are viewed as equal ansehen, if they can be
turned into each other by a rotation (but not by a reflection!)?

n“4
k“2
n“5
k“2

If we want to describe Example 2.7.2 “abstractly”, then we can do this as fol-

lows: Suppose an object on n (labelled) atoms is given. These atoms are mapped
into a set of values R “ rks of k colours (in the example: k “ 2). We ask about the
number of such mappings, where we consider 2 mappings f 1 , f 2 as equivalent
if there exists a cyclic permutation σ P Sn of labelled atoms, such that
f1 ˝ σ “ f2 .
For instance:
4 3 4 3
f1 f2 f 1 ˝ p1 2 3 4q “ f 2 .
1 2 1 2

This problem can be generalised as follows:

D EFINITION 2.7.3. Let a finite set rns “ t1, 2, . . . , nu and a group G Ď Sn of per-
mutations be given. The corresponding idea is: We have a (labelled) object O on n
atoms which we identify with the elements from rns, and G is the group of symmetry
mappings, i.e., the group of permutations of atomes of O, which leave O invariant; for
instance:
28 2. SPECIES

3 2 4
S3 xǫ,p1 2q,p3 4q,p2 4qp1 3qp5 6qy
5 6
1 2 1 3

Further, let R “ rks be a given set of k colours. We are interested in the number of all
nonequivalent mappings f : rns Ñ R, where f 1 „ f 2 if f 1 ˝ g “ f 2 for any g P G.
With other words: We search for the number of all equivalence classes of mappings
under the equivalence relation
f 1 „ f 2 :ðñ Dg P G : f 1 ˝ g “ f 2 .
We call these equivalence classes of “colourings” patterns.
Let each element r P R be assigned a weight śnω prq (usually a monomial, e.g. simply
r
z ), the weight of a function be ω p f q :“ i“1 ω p f piqq. It is clear: f 1 „ f 2 ùñ
ω p f 1 q “ ω p f 2 q; the weight of a pattern shall be ω p f q, where f is any representative
of the pattern (i.e. of the equivalence class).
As always we are interested in the generating function of all patterns, i.e. in the
sum ÿ
ω pfq.
pattern

E XAMPLE 2.7.4. We consider the example of the cube with ω preadq “ x, ω pblueq “ 1:
The weight of a pattern is hence x#pred facesq . The generating function is then:
x6 ` x5 ` 2x4 ` 2x3 ` 2x2 ` x ` 1.
L EMMA 2.7.5 (Burnside’s Lemma, Theorem of Cauchy–Frobenius). Consider the
problem from Definition 2.7.3. Let ψα pgq be the number of all functions f with weight
ω p f q “ α and f ˝ g “ f . Then the number of all patterns with weight α is equal to
1 ÿ
ψα pgq .
|G|
gPG

P ROOF. Consider all pairs pg, f q with f ˝ g “ f , ω p f q “ α: By double counting

one immediately has ÿ ÿ
ψα pgq “ η pfq,
gPG f :ωp f q“α
where η p f q denotes the number of all g with f ˝ g “ f .
For fixed f , G f :“ tg P G : f ˝ g “ f u is a subgroup of G.
Now consider all f ˝ h with h P G: Considering this as set, this is simply the
equivalence class f of f : How often does each element occur?
´ ¯
´1
f ˝ h1 “ f ˝ h2 ðñ f ˝ h1 ˝ h2 “ f
ðñ h1 ˝ h´1
2 P G f ðñ h1 P G f ˝ h2 .

I.e., for the elements of a fixed right coset of G f one always gets the same func-
tions. Hence we have:
ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ
ˇ ˇ ˇ ˇ ˇ ˇ ˇ ˇ
|G| “ ˇ f ˇ ¨ G f ùñ η p f q “ G f “ |G| { ˇ f ˇ .
 2.7. PÓLYA’S THEOREM 29

Wir thus obtain:

ÿ ÿ |G| ÿ 1
η pfq “ ˇ ˇ “ |G| ¨
ˇ ˇ
ˇ ˇ
ˇ ˇ
f :ωp f q“α f :ωp f q“α ˇ f ˇ f :ωp f q“α ˇ f ˇ
ÿ ÿ 1 ÿ
“ |G| ¨ ˇ ˇ “ |G| ¨
ˇ ˇ 1
f :ω p f q“α f P f ˇ f ˇ f :ωp f q“α

“ # ppattern with weight αq ¨ |G| .


D EFINITION 2.7.6. The cycle index of a group G Ď Sn is
1 ÿ ` c1 c2 ˘ 1 ÿ` ˘
PG px1 , x2 , . . . , xn q :“ x1 ¨ x2 ¨ ¨ ¨ xncn “ x|z1 | ¨ x|z2 | ¨ ¨ ¨ ,
|G| |G|
gPG gPG

where z1 ¨ z2 ¨ ¨ ¨ “ g denotes the disjoint cylce decomposition of the permutation g and

|zi | denotes the length of the cycle zi , and wehre pc1 , c2 , . . . q denotes the cycle type of
g (i.e., ci is equal to the number of cycles of length i in g).
E XAMPLE 2.7.7. The rotation group of the cube can be described by permutations of
its faces:
‚ the identity delivers x16 ,
‚ rotation around the axis, which is determined by the midpoints of two opposite
faces, delivers 3 ¨ x22 ¨ x12 (for 180˝) and 6 ¨ x4 ¨ x12 (for ˘90˝ ),
‚ rotation around the axis, which is determined by the midpoints of two diago-
nally apposite (parallel) edges, delivers (for 180˝ ) 6 ¨ x23 ,
‚ rotation around the spatial diagonals (for ˘120˝ ) delivers 8 ¨ x32 .
In total thus:
1 ´ 6 ¯
PG “ x1 ` 3x12 x22 ` 6x12 x4 ` 6x23 ` 8x32 .
24
E XAMPLE 2.7.8. A rotation of a chain of length n corresponds to a permutation π with
π piq “ i ` m pmod nq.
A cycle of length d of π is a (minimal) sequence
l “ pi, i ` m, . . . , i ` pd ´ 1q mq ,
such that i ` d ¨ m ” i pmod mq, i.e., d ¨ m ” 0 pmod nq and because of the mini-
mality of l, we have d “ gcdnm,n .
n
In π there are therefore only cycles of length d, and in fact exactly d many: The weight
n{d
is ω pπq “ xd . The only question left is: How many m’s are there with gcd m, n “
n{d, 0 ď m ă n? From the numbers i ¨ nd
´ n n n¯
0, 1 ¨ , 2 ¨ , . . . , pd ´ 1q ¨
d d d
these are those where i and d are relative prime:
n n
gcd n, i ¨ “ ðñ gcd d, i “ 1.
d d
30 2. SPECIES

Their number is ϕ pdq, the Eulerian ϕ–function:

ˆ ˙ ˆ ˙
k1 kl 1 1
d “ p1 ¨ ¨ ¨ pl : ϕ pdq “ d 1 ´ ¨¨¨ 1´ .
p1 pl
Thus:
1ÿ n
PG “ ϕ pdq xdd . (2.11)
n
d|n

T HEOREM 2.7.9 (Pólya’s theorem). Let G Ď Sn . The generating function for all
patterns f , f : rns Ñ R, is
˜ ¸
ÿ ÿ ÿ ÿ
ω pmq “ PG ω prq , ω prq2 , ω prq3 , . . .
m pattern rPR rPR rPR

P ROOF. Consider α “ ω p f q for a function f : rns Ñ R. By Lemma 2.7.5 the

number of all patterns with weight α is equal to
1 ÿ
ψα pgq ,
|G|
gPG

where ψα pgq “ # pth : ω phq “ α and h ˝ g “ huq.

We multiply this with α and sum over all α:
ÿ ÿ 1 ÿ
ω pmq “ ψα pgq α
m pattern α
|G|
gPG
1 ÿÿ
“ ψα pgq α
|G| α gPG
1 ÿ ÿ
“ ω pfq
|G|
gPG f : f ˝g“ f
“ ...
Think: If g “ z1 ˝ z2 ˝ ¨ ¨ ¨ ˝ zl denotes the disjoint cycle decomposition of g, then
f ˝ g “ f if and only if the function f is constant on all cycles z1 , z2 , . . . zl . Thus:
1 ÿ ÿ
... “ ω p f p1qq|z1 | ¨ ω p f p2qq|z2 | ¨ ¨ ¨ ω p f plqq|zl |
|G|
gPG f :rlsÑR
˜ ¸ ˜ ¸ ˜ ¸
1 ÿ ÿ ÿ ÿ
“ ω prq|z1 | ¨ ω prq|z2 | ¨ ¨ ¨ ω prq|zl |
|G|
gPG rPR rPR rPR
˜ ¸
ÿ ÿ ´ř ¯
“ PG ω prq1 , ω prq2 , . . . . Ðx|zi | ÞÑ rPR ωprq|zi |
rPR rPR


R EMARK 2.7.10. If the set R of colours contains only one element which is simply
assigned the weight z, then there is of courses only one pattern, and the generating
 2.8. A GENERALISATION 31

function of this single pattern is simply zn , where n is the size of the object — and
indeed also ´ ¯
PG z, z2 , . . . “ zn
holds.
E XAMPLE 2.7.11. For the cube we indeed obtain:
ˆ ´ ¯ ´ ¯2 ´ ¯2
1
6 x4 ` 1 px ` 1q2 ` 8 x3 ` 1 ` 3 x2 ` 1 px ` 1q2 `
24
´ ¯3 ˙
6 x ` 1 ` px ` 1q “ x6 ` x5 ` 2x4 ` 2x3 ` 2x2 ` x ` 1.
2 6

E XAMPLE 2.7.12. For a chain of length n which is coloured with k colours, we assign
the weight xi to the colour i and obtain:
1ÿ ´ ¯n{d
ϕ pdq x1d ` x2d ` ¨ ¨ ¨ ` xkd .
n
d|n

For the special case k “ 2 this yields:

1ÿ ´ ¯n{d 1ÿ ÿ ˆn{d ˙ pn{d´lq¨d
d d
ϕ pdq x1 ` x2 “ ϕ pdq x1d¨l x2 Ðm:“d¨l
n n l
d|n d|n l
ÿ ÿ ˆ ˙
1 m n´m n{d
“ x x ϕ pdq .
n m 1 2 m{d
d|gcd n,m

For k colours one more generally has:

ˆ ˙
1ÿ ÿ n{d d¨l
ϕ pdq x1d¨l1 ¨ ¨ ¨ xk k “ Ðmi :“d¨li
n l1 , . . . , l k
d|n l1 `...lk “n{d
ÿ ÿ ˆ ˙
1 m n{d
x1m1 ¨ ¨ ¨ xk k ϕ pdq .
n m `¨¨¨m m1 {d, . . . , mk {d
1 k “n d|gcd n,m1 ,...,mk

Exercise 15: How many different necklaces of n pearls in k colours are there? (This should be
understood “as in real life”, where rotations and reflections of necklaces are considered equal; in
contrast to the presentation in the lecture course.)

2.8. A generalisation
We can apparently generalise Definition 2.7.3, by considering also on the set of
colours R a permutation group H which acts on R: How many patterns are there
which under the action of H are yet different? I.e, we now consider two patterns
f 1 , f 2 as equivalent if there is a g P G (G is the perrmutation group on the atoms)
and a h P H (H is permutation group on the colours) such that
f 1 „ f 2 :ðñ f 1 ˝ g “ h ˝ f 2
E XAMPLE 2.8.1. In Example 2.7.1 (colourings of the faces of the cube with colours
R “ tblue, redu), let the group be H “ SR : I.e., there are now only 6 possibilities for
“essentially different” patterns:
32 2. SPECIES

‚ 6 faces have the same colour,

‚ exactly 5 faces have the same colour,
‚ exactly 2 opposite faces have the same colour,
‚ exactly 2 adjacent faces have the same colour,
‚ exacly 3 in a corner colliding faces have the same colour,
‚ eactly 3 faces in a row have the same colour.
If we define the weight of a pattern by xmaxp#predq,#pblueqq , then the generating function
is:
x6 ` x5 ` 2x4 ` 2x3 .
T HEOREM 2.8.2.
ÿ 1 ÿ ÿ
ω pmq “ ω pfq. (2.12)
m pattern
|G| ¨ |H|
gPG f : f ˝g“h˝ f
hPH

P ROOF. Obviously there holds:

f ˝ g1 “ f ˝ g2 ùñ g1 P G f ¨ g2 ,
h1 ˝ f “ h2 ˝ f ùñ h1 P h2 ¨ H f .
ÿ ÿ |G| ¨ |H|
ω pfq “ ω p f q ˇˇ ˇ ˇ ˇ
f ,g,h f
G f ˇ ¨ ˇH f ˇ
f ˝g“h˝ f
ÿ
“ |G| ¨ |H| ¨ ω pmq .
w pattern

2.9. Cycle index series

We now want to apply these considerations to labelled species: For that we
(essentially) “sum” the cycle indices for all objects.
D EFINITION 2.9.1. Let F be a labelled species: Then the symmetric group Sn acts on
the subfamilies
Fn :“ tF P F : }F} “ nu
“by permutation of the labeling of the atoms”. We denote the action of such a permu-
tation σ on F with σ pFq; if such a permutation σ fixes (leaves invariant) an object F,
then we write for it: σ pFq “ F. It is clear: For each object F P Fn
symF pFq :“ tσ P Sn : σ pFq “ Fu
is a subgroup4 of Sn : We refer to it as symmetry group of the object F.
E XAMPLE 2.9.2. The permutation σ “ p12q p3q (in cycle notation) obviously fixes the
following unordered rooted tree:
1 2
3
4algebraically: the stabilizer–subgroup of F.
 2.9. CYCLE INDEX SERIES 33

The same object, viewed as ordered rooted tree, would however aber only be fixed by
the identity (compare also with Remark 2.6.8).
D EFINITION 2.9.3 (Cycle index series). Let F be a labelled species. For a permuta-
tion σ P Sn we define:
fixF pσq :“ tF P Fn : σ pFq “ Fu .
|fixF pσq| is hence the number of objects F in F for which σ belongs to the symmetry
group of F.
The cycle type of a permutation σ we denote by
c pσq :“ pc1 , c2 , . . . , cn q .
(I.e., ci is the number of cycles of length i in σ.)
Further, for a sequence x :“ px1 , x2 , . . . q of infinitely many variables we define:
xcpσq :“ x1c1 ¨ x2c2 ¨ ¨ ¨ xncn .
The cycle index series of a species F is then the formal power series (in infinitely
many variables)
ÿ xcpσq
ZF pxq “
}F}!
σ,F
σpFq“F
ÿ 1 ÿ
“ |fixF pσq| ¨ xcpσq (2.13)
ně0
n! σPS n
ÿ 1 ÿ ÿ
“ xcpσq
n!
ně0 σPSn FPFn :
σpFq“F
ÿ 1 ÿ ÿ
“ xcpσq . (2.14)
n!
ně0 FPFn σPSn :
σpFq“F

R EMARK 2.9.4. Unordered labelled rooted trees differ from ordered labelled rooted trees
“only” by the property that the first have a nontrivial symmetry group (i.e..: not only
consisting of the identity), compare with Example 2.9.2. With other words: For the
species of ordered labelled rooted trees
W1 „ W2 :ðñ “W1 and W2 are identical as unordered objects”
is clearly an equivalence relation whose equivalence classes can be naturally identi-
fied with the unordered labelled rooted trees. “From the view of the unordered trees”
this equivalence relation is given by its symmetry group G (which consists of “rear-
rangements” of subtrees):
W1 „ W2 :ðñ D σ P G : σ pW1 q “ W2 .
This can be generalized as follows: Let Fp be a labelled specied which contains only
the trivial symmetry group (i.e., all n! relabelings of an object Fp P Fp are viewed as
different): ´ ¯
sym p Fp “ tidu for all Fp P Fp .
F
34 2. SPECIES

Let G be a “larger” (casually spoken) symmetry group for the objects from Fp ; i.e.,
tidu Ă G “ GFp Ă Sn for all Fp P Fpn Ď Fp .
Then one can consider the species F of the equivalence classes of Fp with respect to
this symmetry group, whose objects F P F are determined by the prpoerty that for any
two representatives F1 , F2 P F Ď Fp
F1 „ F2 :ðñ D σ P G : σ pF1 q “ F2 (2.15)
holds. For a Fp P Fpn let
´ ¯ ! ´ ¯ )
Sn Fp :“ σ Fp : σ P Sn (all “relabelings” of F),
p

Sn pFq :“ tσ pFq : σ P Sn u (all equivalence classes w.r.t. (2.15).

Then
´ “permutation
¯ of the labelings from rns” delivers a group action from Sn to
Sn Fp which induces an action on the equivalence classes Sn pFq. Clearly, GF is
the stabilizer5 of F:
GF “ symF pFq ,
those there holds (see also the proof of Lemma 2.7.5):
n! n!
|Sn pFq| “ “ .
symrF s pFq |GF |
´ ¯
For F1 , F2 P Sn Fp there holds (by definition) F2 “ τ pF1 q for a τ P Sn . Then there
holds6 for any such τ P Sn :
` ˘
τ symF pF1 q τ ´1 “ symF pF2 q .
Algebraically this means: For two elements in the same orbit the corresponding stabi-
lizers are conjugated subgroups; in particular they are isomorphic. Combinatorially
this means that the symmetries do not depend on the concrete labeling of the atoms (for
our rooted trees this is obviously clear).
Then we can also consider the species Fr of the corresponding unlabelled structures
whose objects are the equivalence classes of Fp under the equivalence relation “is
obtained by an EMarbitrary relabeling of”, i.e. with the symmetry group equal to the
full symmetric group.
Graphical illustration of the general situation (GF :“ symF pFq):
τ2 pFq“pτ2 GF τ2´1 qpτ2 p Fpqq τm pFq“pτm GF τm
´1
qpτm p Fpqq

´ ¯ ´ ¯ ´ ¯
´ ¯ τ2 Fp τ3 Fp τm Fp
Fp
r
F “ Sn Fp “

F“GF p F
pq τ3 pFq“pτ3 GF τ3´1 qpτ3 p Fpqq

5Because G fixes by definition the equivalence class F.

F
6By definition of a group action.
 2.9. CYCLE INDEX SERIES 35

Casually spoken: Generally labelled species F appear “in the middle between the cor-
responding unlabelled species Fr and the corresponding totally unsymmetric species
Fp ”; expressed by the corresponding symmetry groups:

´ ¯ ´ ¯
symFp Fp “ tidu Ă GF “ symF pFq Ă Sn “ symFr Fr .

E XAMPLE 2.9.5. To make this completely concrete, we consider again the unordered
labelled rooted tree from Example 2.9.2:

p
“Fix numbering of F”:
1 2
Fp 1 Fp 2
3
3
GF “ tid, p3q p12qu » Z2 .

´ ¯
Consider S6 Fp , i.e., all reorderings of F:
p
σ“p123q τ“p1qp23q id
2 3 1 3 1 2
1 2 1 2 1 2
´ ¯ 1 2 3
S6 Fp “ 3 3 2 3 3 1 2 3 1
1 2 1 2 1 2
1 2 3
3 3 3

σ pFq τ pFq F

With this considerations in the background we can now formulate:

L EMMA 2.9.6. With the notation introduced in Remark 2.9.4 there holds: The cycle
indikator series of a labelled species F is the sum over the cycle indices of all objects of
the corresponding unlabelled species Fr . This means:

ÿ
PGF pxq “ ZF pxq . (2.16)
r Fr
FP
36 2. SPECIES

P ROOF. The proof consists only of a longer transformation:

ÿ ÿ 1 ÿ
PGF pxq “ xcpγq ÐDefinition 2.7.6
|GF |
r Fr
FP r Fr
FP γPGF
ÿ ÿ ÿ 1
“ ¨ xcpγq Гregrouping of terms”
ně0 γPSn FP
|GF |
r Frn
γPGF
ÿ 1 ÿ ÿ n!
“ ¨ xcpγq ÐGF is stabilizer of F
ně0
n! |GF |
r Frn
γPSn FP
γPGF
ÿ 1 ÿ ÿ
“ xcpγq Ð n!
|GF |
“|Sn pFq|
n! γPS
ně0 n FPFn :
γpFq“F
ÿ 1 ÿ
“ |fixF pγq| ¨ xcpγq
ně0
n!
γPSn
“ ZF pxq . 
T HEOREM 2.9.7. There holds:
ZF pz, 0, 0, . . . q “ egf F pzq , (2.17)
´ ¯
2 3
ZF z, z , z , . . . “ gf Fr pzq , (2.18)

where Fr is the unlabelled species associated to F (with the same symmetries).

P ROOF. Equation (2.17) is obtained by a simple calculation from Definition (2.13):
ÿ 1 ÿ
ZF pz, 0, 0, . . . q “ |fixF pσq| ¨ zc1 ¨ 0c2 ¨ 0c3 ¨ ¨ ¨
ně0
n! σPS
n
ÿ 1
“ |fixF pidn q| ¨ zn
ně0
n!
ÿ 1
“ ¨ f n ¨ zn “ egf F pzq .
ně0
n!
` ˘
Equation (2.18) is a direct corollary from Lemma 2.9.6, since PGF z, z2 , . . . is
simply equal to zn fory any F P Fn .

E XAMPLE 2.9.8. For the species atom we obtain Zatom “ x1 .
For the species sets we obtain according to (2.13)
ÿ 1 ÿ
Zsets “ xcpσq ,
ně0
n!
σPSn
because there is exactly one set on n elements which (of course) is fixed by any permu-
tation σ of its atoms:
fixsets pσq “ 1.
 2.9. CYCLE INDEX SERIES 37

It is well-known that the number of all permutations of Sn of cycle type pc1 , c2 , . . . , cn q

is exactly7
n!
. (2.19)
1 c1 ¨ c 1 ! ¨ 2 c2 ¨ c 2 ! ¨ ¨ ¨ n c n ¨ c n !
Therefore we further obtain:
ÿ ÿ x1c1 ¨ x2c2 ¨ ¨ ¨ xncn
Zsets “
ně0 1¨c1 `2¨c2 `¨¨¨`n¨cn “n
1 c1 ¨ c 1 ! ¨ 2 c2 ¨ c 2 ! ¨ ¨ ¨ n c n ¨ c n !
` x ˘c 1 ` x ˘ c 2
ÿ 1 2
1 2
“ ¨ ¨¨¨
c1 ! c2 !
1¨c1 `2¨c2 `¨¨¨ă8
x2 x3
“ ex1 ` 2 ` 3 `¨¨¨ .
Combined with (2.18) thsi yields again the (ordinary) generating function of the (un-
labelled) species sets:
´ ¯ 1
Zsets z, z2 , . . . “ “ gfsets pzq .
1´x
For the species cycles we obtain (compare with Observation 2.9.4):
ÿ 1 ÿ ÿ
Zcycles “ xcpσq Ð(2.14)
n!
ně1 APcyclesn σPSn :
σpAq“A
ÿ 1 ÿ
“ pn ´ 1q! xcpσq ÐDef. cycle index
n!
ně1 σPSn :
σpAq“A
ÿ 1ÿ n{d
“ ϕ pdq xd Ð(2.11)
n
ně1 d|n
ÿ ÿ 1
“ ϕ pdq xdk Ðn“k¨d
k¨d
dě1 kě1
ÿ ϕ pdq 1
“ log .
d 1 ´ xd
dě1

For the species permutations we think: Let σ P Sn be of cycle type pc1 , c2 , . . . , cn q.

How does a permutation π look like which is fixed by σ?
σ
i ÝÝÝÑ σ piq
§ §
§
πđ
§
πđ
σ
π piq ÝÝÝÑ σ pπ piqq “ π pσ piqq
7Simple counting argument: Consider an arbitrary permutation τ as a cycle decomposition
(the first c1 elemente are the fixpoints, the following 2 ¨ c2 elements are the 2-cycles, etc.): In
how many ways is one and the same permutation σ of cycle type pc1 , . . . , ck q obtained in this
manner?
38 2. SPECIES

This means: π is fixed by σ, if and only if

π “ σ´1 ˝ π ˝ σ ðñ π ´1 “ σ´1 ˝ π ´1 ˝ σ ðñ π ˝ σ “ σ ˝ π.

Suppose i and π piq “ j belong to cycles of different length in σ: W.l.o.g. let the σ–cycle
of i be the shorter one (otherwise consider π ´1 instaed of π); let k be its length:
σ σ σ
i ÝÝÝÑ σ piq ÝÝÝÑ ¨ ¨ ¨ ÝÝÝÑ σk piq “ i
§ § § §
§
πđ
§
πđ πđ
§ §
πđ
σ σ σ
j ÝÝÝÑ σ pjq ÝÝÝÑ ¨ ¨ ¨ ÝÝÝÑ σk pjq ‰ j
Hence in π always only those numbers are permuted which in σ belong to cycles of
equal length! Hoe many possibilities are there? We consider the ck cycles of length k
of σ:
` ˘
pi . . . q pi . . . q
loomoon loomoon looomooon
1 2 . . . i c k
. . . .
k k k
Let i be the smallest number in these ck k–cycles: The permutation π can map i to ck ¨ k
numbers: Thereby, at the same time, also the images
´ ¯
k´1
π piq , π pσ piqq , . . . , π σ piq

are already determined! For the next smaller number j there are thus only pck ´ 1q ¨ k
possibilities leftn, etc.: It is clear that in this manner all permutations π, which are
fixed by σ, are correctly counted; the number of these permutations thus is in total
c 1 ! ¨ 1 c1 ¨ c 2 ! ¨ 2 c2 ¨ ¨ ¨
On the other side there are exactly
n!
1 c1 ¨ c 1 ! ¨ 2 c2 ¨ c 2 ! ¨ ¨ ¨
permutations σ P Sn of cycle type pc1 , c2 , . . . q (see (2.19)). Thus we obtain:
ÿ 1 ÿ ˇ ˇ
Zpermutations “ ˇfixpermutations pσqˇ ¨ xcpσq
ně0
n!
σPSn
ÿ 1 ÿ n! ¨ pc1 ! ¨ 1c1 ¨ ¨ ¨ cn ! ¨ ncn q c1
“ ¨ x1 ¨ ¨ ¨ xncn
ně0
n! 1 ¨ c1 ! ¨ ¨ ¨ c n ! ¨ n
c 1 c n
pc1 ,...,cn q
ÿ c
ź 1
“ x11 ¨ x2c2 ¨ ¨ ¨ “ .
1 ´ xi
pc1 ,c2 ,... q iě1

The following theorems are concerned with product and composition of species
with symmetries: We actually would first need to define the symmetries for
these constructions — but how these symmetries are to be meant8, will become
clear in the proofs.

8The interpretations being dealt with are anyway the most obvious ones.
 2.9. CYCLE INDEX SERIES 39

T HEOREM 2.9.9. Let A and B be two species. Then there hold:

ZA Y9 B “ ZA ` ZB , (2.20)
ZA ‹B “ ZA ¨ ZB . (2.21)
IfA contains no object of size zero, then there also holds:
1
ZA ‹ “ . (2.22)
1 ´ ZA
P ROOF. Equation (2.20) immediately follows from Definition (2.13).
` ˘
For (2.21) we consider an element C “ pA, Bq P nk ¨ pA ˆ B q, where n “
}A} ` }B} and k “ }A}. A permutation σ which leaves such a pair invariant,
decomposes9 in σ “ pσ1 , σ2 q with
σ ppA, Bqq “ pσ1 pAq , σ2 pBqq .
Therefore ˆ ˙
n
|fixC pσq| “ ¨ |fixA pσ1 q| ¨ |fixB pσ2 q| ,
k
and the cycle type of σ of course satisfies
c pσq “ c pσ1 q ` c pσ2 q .
For (2.22) we only need to remember
ď
9
A‹ “ kě0 A
k
.
The claim then follows from (2.20) and (2.21). 
T HEOREM 2.9.10. Let A, B be two species, let B contain no object of size 0. With the
short notation
xm :“ pxm , x2m , x3m , . . . q
we have:
ZApBq pxq “ ZA pZB px1 q , ZB px2 q , ZB px3 q , . . . q .
P ROOF. By Definition (2.14) there holds:
ÿ 1 ÿ
ZApBq px1 , x2 , x3 , . . . q “ xcpσq .
ně0
n!
pC,σq:
CPApB q: }C}“n
σPSn : σpCq“C

Indeed, each object C P A pB q by construction “consists”

‚ of an A–object A of size k
‚ and of k B –objects B1 , . . . , Bk ;
is thus, so to say, “of the form”
C “ pA; B1 , B2 , . . . , Bk q .
A permutation σ, which fixes this object, can be described as follows:
‚ σ corresponds “macroscopically” to a permutation ρ P Sk , which fixes
A fixiert (notation: rρ | σs),
9I.e.: This is how the action on the product is meant!
40 2. SPECIES

‚ Each cycle pi1 , . . . , il q of ρ induces a cycle of σ which maps the atoms of

Bim to the ataoms of Bim`1 (the subindices are of course considered mod-
` ˘
ulo the cykle length lt): In particular the objects Bi1 , . . . , Bil , which
correspond to this cycle, are identical as unlabelled objects (“structures”,
compare with the considerations in Observation 2.9.4).
Each pair pC, σq P A pB q ˆ Sn from the region of smmation defines an ordered
set partition of rns “ B1 Y9 B2 ¨ ¨ ¨ Y
9 Bk : We can achieve by a relabeling τ of the
atoms of C Ñ C1 another arbitrary ordered set partition (with the same sequence
of block sizes, of course) and we obtaine another pair
´ ¯
C1 , σ1 “ τ ˝ σ ˝ τ ´1
from the region of summation, which delivers the same cycle type:
` ˘
c pσq “ c σ1 .
There are (independent from the order of the blocks Bi )
n!
}B1 }! ¨ ¨ ¨ }Bk }!
such ordered set partitions. Of these — up to relabeling — identical objects we
of course always only need to consider one; e.g. the one with the “canonical”
set partition
tt1, 2, . . . , }B1 }u , t}B1 } ` 1, . . . , }B1} ` }B2 }u , . . . u .
On the other hand, each relabeling ν of the Atome from A Ñ A1 first induces a
permutation ρ1 “ ν ˝ ρ ˝ ν´1 for which also ρ1 pA1 q “ A1 holds, and if we reorder
the blocks Bi accordingly
Bi ÞÑ Bνpiq ,
we obtain in this way again simply the same object C (with the same permutation
σ which fixes it): Clearly there are k! such relabelings.
Therefor we first obtain:
ÿ 1 ÿ ÿ n! 1
ZApBq pxq “ ¨ ¨ xcpσq
ně0
n! }B1 }! ¨ ¨ ¨ }Bk }! k!
pC,σq: ρ:rρ|σs
C“pA;B1 ,B2 ,...,Bk q
pB1 ,...,Bk q “canonically”
σPSn : σpCq“C
ÿ 1 ÿ ÿ ÿ 1 ÿ 1
“ ¨ ¨ ¨ ¨ xcpσq .
k!
kě0 APAk σ:rρ|σs B1 :ζ 1
}B1 }! ℓ pζ 1 q
B :ζ
}Bm }! ℓ pζ m q
m m
ρpAq“A
ρ“ζ 1 ˝ζ 2 ...ζ m

Two things are to be made clear:

‚ How many “completed” permutations σ belong to a fixed ρ?
‚ How does the cycle type c pσq of the “completed” permutation σ arise
from the permutation ρ or its cycles?
For this we consider a fixed cycle ζ of ρ with ℓ pζq “ l which “sends around in
a circle” the B –object B:
 2.9. CYCLE INDEX SERIES 41

τ1 τ2
B
τl τ3

This cycle is “completed” by the maps (“relabeling of the original object B”)
τ1 , τ2 , . . . , τl to cycles of the “full” permutation σ. Here it should be observed
that
τ :“ τ1 ˝ τ2 ¨ ¨ ¨ τl (2.23)
gives a permutation which fixes B. For instance, in the following schematic
graphic the cycle length is l “ 4 and the permutation (in cycle notation) is
τ “ p12q p3q:

τ1
1 4
2 5
3 6

τ4 τ2
12 9
11 8
10 7

τ3
From the graphic it is immediately clear: If the cycle decomposition of τ is given
by
τ “ ξ1 ˝ ξ2 ¨ ¨ ¨ ξs ,
then the corresponding cycles of the “composed permutation” σ have the lengths
pℓ pξ 1 q ¨ l, ℓ pξ 2 q ¨ l, . . . q.
Further is follows: For a fixed permutation τ which fixes B, there are

p}B}!ql´1
“completions” pτ1 , τ2 , . . . , τl q. For the first pl ´ 1q τi ’s can be chosen arbitrarily;
τl is then uniquely determined by (2.23).
In total we can thus continue the calculation in the following way:
` ˘ ` ˘
ÿ 1 ÿ ÿ xl , x2l , . . . cpτ1 q ÿ xl , x2l , . . . cpτm q
ZA p B q “ 1 1
¨¨¨ m m
k! }B1 }! }Bm }!
kě0 APAk B1 Pζ 1 Bm Pζ m
ρpAq“A ℓpζ 1 q“l1 ℓpζ m q“lm
ρ“ζ 1 ˝ζ 2 ...ζ m τ1 pB1 q“B1 τm pBm q“Bm
“ ZA pZB px1 , x2 , . . . q , ZB px2 , x4 , . . . q , . . . q .

42 2. SPECIES

E XAMPLE 2.9.11. We have

` ˘
Zpermutations “ Zsets Zcycles px1 , x2 , . . . q , Zcycles px2 , x4 , . . . q ,
ř ϕpnq 1 1 ř ϕpnq
log 1´x n ` 2 log 1´x1 `¨¨¨
“e ně1 n ně1 n 2n
ř 1 ř ϕpdq
ř
“ e ně1 log 1´xn d|n n Ð d|n
ϕpdq
n “1
ź 1
“ .
1 ´ xn
ně1

Exercise 16: Determine the cycle index series of the species Fixfree of permutations without fixed
points.
Hint: Show the relation Sets ¨ Fixfree “ Permutations.

Exercise 17: Determine the cycle index series for the species “set partitions”.

Exercise 18: Given some arbitrary species A, show the formula

ˆ ˙
B
ZA1 px1 , x2 , . . . q “ Z px1 , x2 , . . . q .
Bx1 A

Exercise 19: Show: ˜ ¸

8
ź ÿ 1 xn
1
“ exp .
1 ´ xk n 1 ´ xn
k “1 n ě1

Hint: Permutations “ Sets pCyclesq .

 CHAPTER 3

Partially ordered sets

3.1. Definition and Examples

D EFINITION 3.1.1. A partially ordered set (or short poset) is a set P together with
a relation ď (“less than or equal to”) that satisfies the following conditions:
(1) @x P P : x ď x (reflexivity)
(2) @x, y P P : x ď y and y ď x ùñ x “ y (antisymmetry)
(3) @x, y, z P P : x ď y and y ď z ùñ x ď z (transitivity)
The following notation is commonly used (“strictly less than”):
x ă y :ðñ x ď y and x ‰ y.
In addition, we use the following notation:
x Í y :ðñ x ą yandEz : x ą z ą y
We say: “x covers y”.
Two elements x and y are said to be comparable if x ď y or y ď x, otherwise they are
said to be incomparable.
In case two posets P, Q are considered, we denote the
‚ order relation on P by ďP ,
‚ while the order relation on Q is denoted by ďQ ,
in order to avoid confusions. (However, we simply use ď for the usual order on N.)
For finite posets P (to be more precise: if P does not have too many elements ;-),
Hasse–diagrams are a useful description of the poset, where the order relations
are represented as the edges of a graph with vertex set P. The following fig-
ure illustrates this for P “ ta, b, c, d, e, f u with the following covering relations
(which of course uniquely determine a finite poset!)
b Ì a, c Ì b, e Ì a, e Ì d, f Ì c, f Ì e.
a d

b
e
c

E XAMPLE 3.1.2. Examples of posets:

(1) P “ rns linearly ordered: An order relation is said to be linear, if any two
elements x, y are comparable (the Hasse–diagram looks like a line then):
43
44 3. POSETS

(2) P “ 2rns (Power set of rns) ordered by set inclusion: This poset is said to be
the Boolean algebra Bn . E.g., for n “ 3:
t1,2,3u

t1,2u t1,3u t2,3u

t1u t2u t3u

(3) P “ Sn with weak order: σ Ì π if and only if the permutation π is ob-

tained from the permutation σ by interchanging two adjacent elements and
the number of inversions is increased by this transposition.
σ “ ¨ ¨ ¨ πi ă πi`1 ¨ ¨ ¨
π “ ¨ ¨ ¨ πi`1 ą πi ¨ ¨ ¨

321

231 312

213 132

123

(4) P “ Tn :“ td P N : d | nu, ordered by a ď b :ðñ a | b: This poset is said to

be the divisor lattice. E.g., n “ 30:
30

6 10 15

2 3 5

(5) P “ P n , the family of all (set–)partitions of rns, ordered by “refinement”: For

two partitions π, τ we have π ď τ, if every block of π is fully contained in
some block of τ (i.e., “π is constructed from τ by subdividing blocks”). E.g.,
for n “ 3:
 3.1. DEFINITION AND EXAMPLES 45
tt1,2,3uu

tt1u,t2,3uu tt2u,t1,3uu tt3u,t1,2uu

tt1u,t2u,t3uu

(6) P “ Vn,q , the family of subspaces of a finite vectorspace pGF pqqqn over a finite
field GF pqq, ordered by “subspace–relation” (that is set inclusion).
D EFINITION 3.1.3. Given two posets P, Q with order relation ďP (for P) and ďQ (for
Q). A map φ : P Ñ Q is said to be order-preserving, if, for all x, y P P, we have:
x ďP y ùñ φ pxq ďQ φ pyq .
The posets P and Q are said to be isomorphic if there exists a bijective map φ : P Ñ Q
such that φ and φ´1 both are order-preserving (such a map φ is said to be an order
isomorphism), that is
x ďP y ðñ φ pxq ďQ φ pyq .
E XAMPLE 3.1.4. Let P “ t0, 1un be the set of all t0, 1u–vectors of length n, ordered by
pv1 , . . . , vn q ď pw1 , . . . , wn q :ðñ vi ď wi for i “ 1, 2, . . . , n.
The interpretation of these vectors as characteristic functions of subsets induces an
order-preserving bijection P Ñ Bn on the Boolean algebra Bn (see example 3.1.2).
Exercise 20: Let P be some finite poset and f : P Ñ P an order–preserving bijection. Show that
f ´1 is also order–preserving.
Show that this is not true in general for infinite posets.

D EFINITION 3.1.5. Let P be a poset: A partially ordered subset Q Ď P is said to be

‚ a weak subposet of P if x ďQ y ùñ x ďP y for all x, y P Q,
‚ an induced subposet of P if x ďQ y ðñ x ďP y for all x, y P Q.
There is an analogy to the difference between subgraphs and induced subgraphs;
the following figure illustrates the difference:
Q “ ta, e, f u Ă P “ ta, b, c, d, e, f u:
a d a a

b
Q induziert
P e e
c
Q schwach
f f e f

D EFINITION 3.1.6. Let P be a poset and x, y P P with x ď y: An induced subset

rx, ys :“ tz : x ď z ď yu
is said to be an interval (in particular: rx, xs “ txu).
A poset that has only intervals of finite cardinality is said to be locally finite.
46 3. POSETS

D EFINITION 3.1.7 (Chains and Antichains). Let P be a poset: An element x P P

with
EyPP:yăx
is said to be a minimal element of P, while an element x P P with
EyPP:yąx
is said to be a maximal element of P.
An element x P P with x ď y (resp. x ě y) for all y P P is said to be the minimum
(resp. maximum). If P has a minimum (maximum), then it is of course uniquely
determined: We denote it by 0̂ (resp. 1̂).
If P contains a minimum 0̂, then we refer to an element that covers 0̂ as atom of P.
If P contains a maximum 1̂, then we refer to an element that is covered by 1̂ as coatom
of P.
A subset C of P that is linearly ordered as induced subposet is said to be a chain in
P. A subset A of P where any two elements are incomparable is said to be an antichain
(or Sperner family) of P.
If a chain C “ tx1 ă x2 ă ¨ ¨ ¨ ă xm u is finite, then ℓ pCq :“ |C| ´ 1 is the length of
C.
A (finite) chain is said to be saturated, if “its elements cover each other (in P)”, that is
if
C “ tx1 Ì x2 Ì ¨ ¨ ¨ Ì xm u .
A saturated chain C is said to be maximal, if there exists no saturated chain D with
C Ď D and ℓ pDq ą ℓ pCq.
Exercise 21: (a) Find a (finite) poset P where
‚ the length of a longest chain is l,
‚ every element of P belongs to a chain of length l,
which nevertheless has a maximal chain of length ă l.
(b) Let P be a (finite) poset with connected Hasse–diagram, where the longest chain has length l
(there might be several longest chains). Moreover, assume that for all x, y P P such that y Í x (y
covers x), x and y belong to a chain of length l: Show that under this assumption all maximal chains
have length l.

D EFINITION 3.1.8 (Order ideal). A subset I Ď P with the property

x P I and y P P with y ď x ùñ y P I
is said to be an order ideal of P. For an arbitrary subset S Ď P, the set
xSy :“ ty P P : Dx P S : y ď xu
is always an order ideal; it is the order ideal generated by S. If S consists of a single
element x, then xSy “ xtxuy is the principal order ideal generated by x.
A subset I Ď P with the property
x P I and y P P with y ě x ùñ y P I
is said to be a dual order ideal of P (dual principal order ideal is defined in a similar
manner).
 3.1. DEFINITION AND EXAMPLES 47

The set of all order ideals of P, ordered by inclusion, is itself a poset, which is denoted
by J pPq in this text: Such a poset always has a minimum 0̂ (the empty set) and always
has a maximum 1̂ (P itself).
P ROPOSITION 3.1.9. Let P be a poset, I P J pPq be a order ideal of P, and x a maximal
element in I.
Then also Iz txu is an order ideal of P.

In case that P is finite, there is a bijection between order ideals I and antichains
A:
I ÞÑ A :“ tx : x maximal in Iu
A ÞÑ I “ xAy :“ ty : Dx P A : y ď xu
D EFINITION 3.1.10 (Linear extension). An order-preserving bijection from a finite
poset P with |P| “ n to rns is said to be a linear extension of P.
P ROPOSITION 3.1.11. The number of linear extensions of a finite poset P equals the
number of maximal chains in J pPq.

P ROOF. Let σ : P Ñ rns be a linear extension. Set I0 :“ H and Ik :“ σ´1 prksq

for k “ 1, 2, . . . , n. Then the subset Ik Ď P is always an order ideal of P (since σ is
order-preserving) with |Ik | “ k (as σ is bijective); this implies in J pPq:
0̂ “ I0 Ì I1 Ì ¨ ¨ ¨ Ì In “ P “ 1̂.
This means the following: Every linear extension σ of P is associated with a
maximal chain in J pPq.
Conversely, let C be a maximal chain in J pPq: The length of such a chain is
obviously n; indeed, one can obtain each such maximal chain algorithmically
as follows: Start with I0 “ H; if Ik´1 ‰ P has already been constructed, choose
a minimal element xk from PzIk´1 and set Ik :“ Ik´1 Y txk u. For this chain C, we
define σ : P Ñ rns:
σ pxk q “ k;
The map σ is of course bijective; in addition, the map is order-preserving, since
xi ďP x j ùñ i ď j.
(As i ą j ùñ xi ęP x j .)
In summary, we observe the following: There exists a bijection between the
linear extensions of P and the maximal chains in J pPq. 
D EFINITION 3.1.12. A (finite) poset is said to be graded of rank n, if each maximal
chain has the same length n. For such a poset P, there exists a uniquely determined
rank function ρ : P Ñ t0, 1, . . . , nu such that

‚ ρ pxq “ 0, if x is a minimal element of P,

‚ ρ pyq “ ρ pxq ` 1, if y Í x.
48 3. POSETS

In case ρ pxq “ i, we say: x has rank i. For such a poset P, let pi be the number of
elements of rank i: Then the polynomial
n
ÿ
F pP, qq :“ p i qi
i“0
is said to be the rank generating function of P.
Exercise 22: Consider the “zigzag–poset” Zn with elements x1 , x2 , . . . , x n and cover relations
x2i´1 Ì x2i for i ě 1, 2i ď n and x2i Í x2i`1 for i ě 1, 2i ` 1 ď n
a) How many order ideals are there in Zn ?
b) Let Wn pqq be the rank generating function of the lattice of order ideals J pZn q of Zn . For instance,
W0 pqq “ 1, W1 pqq “ 1 ` q, W2 pqq “ 1 ` q ` q2 , W3 pqq “ 1 ` 2q ` q2 ` q3 . Show:
8
ÿ 1 ` p1 ` qqz ´ q2 z3
W pq, zq :“ Wn pqq zn “ .
n “0
1 ´ p1 ` q ` q2 qz2 ` q2 z4

c) Let en be the number of all linear extensions of Zn . Show:

8
ÿ zn 1
en “ tan z ` .
n “0
n! cos z

3.2. Construction of posets

Given two posets, there are several possibilities to construct new posets from
them. The simplest poset (disregarding the empty poset), which consists of a
single element, is denoted by ˝ in the following.
D EFINITION 3.2.1. Let P and Q be posets, and suppose that P and Q are disjoint as
sets. The direct sum direct sum of P and Q is denoted by P ` Q: It is the poset on the
union P Y Q such that x ď y if (and only if)
‚ x, y P P and x ďP y
‚ or x, y P Q and x ďQ y.
Moreover, we denote the ordinal sum of P and Q by P ‘ Q: It is the poset on the union
P Y Q such that x ď y if (and only if)
‚ x, y P P and x ďP y
‚ or x, y P Q and x ďQ y
‚ r x P P and y P Q.

Using Hasse diagrams, the direct sum and the ordinal sum can be illustrated as
follows:

P
P`Q P‘Q
Q
 3.2. CONSTRUCTION OF POSETS 49

We obviously have
P ` Q » Q ` P,
however, in general,
P ‘ Q fi Q ‘ P.
Both, the direct sum and the ordinal sum, are associative.
An antichain with n elements is obviously isomorphic to the n-fold direct sum
n ¨ ˝ :“ ˝loooooomoooooon
` ˝ ` ¨ ¨ ¨ ˝,
n times
and a chain with n elements is isomorphic to the n-fold ordinal sum
˝loooooomoooooon
‘˝‘¨¨¨˝.
n times

D EFINITION 3.2.2. Let P and Q be posets.

We denote the direct product of P and Q by P ˆ Q, which is defined as the poset on
the cartesian product P ˆ Q such that px, yq ď px1 , y1 q if (and only if)
x ďP x1 and y ďQ y1 .
Moreover, we denote the ordinal product of P and Q by P b Q, which is defined as
the poset on the cartesian product P ˆ Q such that px, yq ď px1 , y1 q if (and only if)
x “ x1 and y ďQ y1 , oder x ăP x1 .

Using Hasse diagrams, the direct sum and the ordinal sum can be illustrated as
follows:

P
PˆQ PbQ
Q

Again, we have
P ˆ Q » Q ˆ P,
but, in general,
P b Q fi Q b P.
The direct product is associative.
Exercise 23: Let P, Q be graded posets, let r and s be the maximal ranks of P and Q, respectively,
and let F pP, qq and F pQ, qq be the corresponding rank generating functions. Show:
a) If r “ s (otherwise maximal chains would be of different lengths), then F pP ` Q, qq “ F pP, qq `
F pQ, qq.
b) F pP ‘ Q, qq “ F pP, qq ` qr`1 F pQ, qq.
c) F pP ˆ Q, qq “ F pP, qq ¨ F pQ, qq.
` ˘
d) F pP b Q, qq “ F P, qs`1 ¨ F pQ, qq.
50 3. POSETS

D EFINITION 3.2.3. Let P and Q be posets.

Then Q P denotes the set of all order-preserving maps f : P Ñ Q with order relation
f ď g, if f pxq ďQ g pxq for all x P P.
E XAMPLE 3.2.4. The Boolean lattice can also be obtained as follows, see Example 3.1.4:
Bn » t0, 1un¨˝ .
Exercise 24: Let P, Q, R be posets. Find order isomorphisms for the following relations:
a) P ˆ pQ ` Rq » pP ˆ Qq ` pP ˆ Rq.
b) R P`Q » R P ˆ R Q .
` ˘P
c) R Q » R QˆP .

ř
Exercise 25: Let P be a finite poset and define GP pq, tq :“ I q| I | tmp I q , where the summation range
is the set of all order ideals I of P, and where m pIq denotes the number of maximal elements of I.
(For instance: GP pq, 1q is the rank generating function of J pPq.)
a) Let Q be a poset with n elements. Show:
` ` ˘˘
GPbQ pq, tq “ GP qn , q´n ¨ GQ pq, tq ´ 1 ,
where P b Q denotes the ordinal product.
b) Let P be a poset with p elements. Show:
ˆ ˙
q´1
GP q, “ qp.
q

3.3. Lattices
D EFINITION 3.3.1 (Lattice). Let P be a poset, and let x, y P P: An element z with
z ě x and z ě y
is said to be an upper bound of x and y; z is said to be the least upper bound or
supremum of x and y, if, for every upper bound w of x and y, we have
w ě z.
If there exists a least upper bound z of x and y, then it is of course unique; we write
x _ y and say “x sup y”.
Analogously, we define the greatest lower bound or infimum (if it exists): We write
x ^ y and say “x inf y”.
A poset in which each pair px, yq of elements has a least upper bound as well as a
greatest lower bound is said to be a lattice.
A finite lattice has a (unique) minimum 0̂ and a (unique) maximum 1̂.
An element x in a lattice that covers 0̂ is said to be an atom.
An element x in a lattice that is covered by 1̂ is said to be a coatom.
E XAMPLE 3.3.2. Examples of finite lattices are as follows:
(1) rns with the usual linear order.
(2) the power set 2rns , ordered by set inclusion.
(3) Sn with the weak order.
 3.3. LATTICES 51

(4) The set Tn “ td P N : d | nu of (positive) divisors of a number n P N, ordered

by divisibility (d1 ď d2 :ðñ d1 | d2 ).
(5) The set Πn of (set–)partitions of rns, ordered by “refinement of partitions”.
(6) The set Vn,q of subspaces of an n–dimensional vector space over a finite field,
ordered by the subset relation (which is again set inclusion).
C OROLLARY 3.3.3. In each lattice, we have the following basic rules:
(1) ^ and _ are commutative and associative operations in L.
(2) @x P L : x _ x “ x ^ x “ x (^ and _ are idempotent operations in L).
(3) x ^ y “ x ðñ x _ y “ y ðñ x ď y.
(4) x ^ px _ yq “ x “ x _ px ^ yq (rule of absorption).

P ROOF. (1): Commutativity is of course obvious.

Associativity for _ can be shown as follows: x _ py _ zq is greater than or equal
‚ to x and y _ z,
‚ and therefore also to x, y and z,
‚ and therefore also to x _ y and z,
‚ and therefore also to px _ yq _ z;
this implies x _ py _ zq ě px _ yq _ z, and the same is true also for the other
direction of the inequality sign.
The argument is the same for ^.
(2) This is trivial.
(3) Using x ď y, we conclude x _ y “ y, and x _ y “ y implies y ě x. The
argument is the same for ^.
(4) This follows immediately from (3). 
C OROLLARY 3.3.4. Let L be a lattice, and let a1 ď a in L. Then we have following for
all b P L
a1 ^ b ď a ^ b (3.1)
a1 _ b ď a _ b. (3.2)

P ROOF. The statement is an immediate consequence of

# #
b b
a1 ^ b ď 1 resp. a _ b ě .
a ďa a ě a1

C OROLLARY 3.3.5 (Modular inequality). Let L be a lattice, and let x, y, z P L. Then
we have
x ď z ùñ x _ py ^ zq ď px _ yq ^ z. (3.3)

P ROOF. Using the definition, we can immediately conclude:

x ď x _ y and x ď z ùñ x ď px _ yq ^ z,
y ^ z ď y ď x _ y and y ^ z ď z ùñ y ^ z ď px _ yq ^ z.
The two inequalities yield (3.3). 
52 3. POSETS

It is possible to give an axiomatic characterization of lattices that is based on the

operations ^ and _.
T HEOREM 3.3.6. Let P be a set with binary operations _ and ^ that satisfy items 1, 2
and 4 from Corollary 3.3.3. Then P is a lattice, where ď is defined by
x ď y :ðñ x ^ y “ x.

P ROOF. First we show the following: The relation defined above is indeed a
partial order.

‚ Reflexivity follows from item 2 in Corollary 3.3.3:

x ^ x “ x ùñ x ď x.
‚ Antisymmetry follows directly from the definition (and from item 1 in
Corollary 3.3.3: commutativity):
x ď y and y ď x ðñ x ^ y “ x and y ^ x “ y ùñ x “ y.
‚ Transitivity: Using
x ď y and y ď z ðñ x ^ y “ x and y ^ z “ y,
we can conclude (employ item 1 in Corollary 3.3.3: associativity)
x ^ z “ px ^ yq ^ z “ x ^ py ^ zq “ x ^ y “ x,
that is x ď z.
Next we show: For each x, y P P, there exists an infimum. The notation already
suggests the right claim: x ^ y is the infimum, since
‚ x ^ y ď x: This follows directly from Corollary 3.3.3, item 1+2:
x ^ px ^ yq “ px ^ xq ^ y “ x ^ y.
‚ For z with z ď x and z ď y, we have (Corollary 3.3.3: associativity):
z ^ px ^ yq “ pz ^ xq ^ y “ z ^ y “ z ùñ z ď x ^ y.
Finally, we show the symmetry relation
x ^ y “ x ðñ x _ y “ y.
Both directions follow from item 4 in Corollary 3.3.3:
‚ x _ y “ px ^ yq _ y “ y,
‚ x ^ y “ x ^ px _ yq “ x.
Using this symmetry relation it becomes evident that (as expected) also x _ y is
the supremum. This concludes the proof. 
P ROPOSITION 3.3.7. Let P be a finite poset with 1̂ such that there exists the infimum
x ^ y for all x, y P P. Then P is already a lattice.
 3.3. LATTICES 53

P ROOF. It suffices to show that supremum x _ y exists for all x, y P P. For this
purpose, it makes sense to define
ľ
x _ y :“ z.
zěx,y

The set over which the infimum is taken is not empty (as 1̂ ě x, y) and also
finite, which guarantees that x _ y is well-defined. It also has the property of
a supremum: Ź As a, b ě x, y ùñ x, y ď a ^ b, it can be proven (induction)
that x, y ď zěx,y z, and, for every upper bound a ě x, y, we obviously have
Ź
a ě zěx,y z. 
R EMARK 3.3.8. Let P and Q be lattices. Then the direct product P ˆ Q is also a lattice,
and we have
px1 , y1 q _ px2 , y2 q “ px1 _ x2 , y1 _ y2 q ,
px1 , y1 q ^ px2 , y2 q “ px1 ^ x2 , y1 ^ y2 q .
Similarly, Q P is a lattice with
p f _ gq pxq “ f pxq _ g pxq ,
p f ^ gq pxq “ f pxq ^ g pxq .

Exercise 26: Let L be a finite lattice. Show that the following three conditions are equivalent for all
x, y P L:
(a) L is graded (i.e.: all maximal chains have in L the same length), and for the rank function rk of
L there holds
rk pxq ` rk pyq ě rk px ^ yq ` rk px _ yq .
(b) If y covers the element x ^ y, then x _ y covers the element x.
(c) If x and y both cover element x ^ y, then x _ y covers both elements x and y.
(A lattice L obeying one of these conditions is called semimodular.)
Hint: Employ an indirect proof for (c) ùñ (a): For the first assertion in (a), if there are intervals
which are not graded, then we may choose an interval ru, vs among them which is minimal with respect
to set–inclusion (i.e., every sub–interval is graded). Then there are two elements x1 , x2 P ru, vs, which
both cover u, and the length of all maximal chains in rx i , vs is ℓi , such that ℓ1 ‰ ℓ2 . Now apply (b)
or (c) to x1 , x2 .
For the second assertion in (a), take a pair x, y P L with
rk pxq ` rk pyq ă rk px ^ yq ` rk px _ yq , (3.4)
such that the length of the interval rx ^ y, x _ ys is minimal, and under all such pairs, also rk pxq `
rk pyq is minimal. Since it is impossible that both x and y cover x ^ y (why?), w.l.o.g. there is an
element x1 with x ^ y ă x1 ă x. Show that X “ x, Y “ x1 _ y is a pair such that rk pXq ` rk pYq ă
rk pX ^ Yq ` rk pX _ Yq, but where the length of the interval rX ^ Y, X _ Ys is less than the length
of rx ^ y, x _ ys.

Exercise 27: Let L be a finite semimodular lattice. Show that the following two conditions are
equivalent:
a) For all elements x, y, z P L with z P rx, ys (i.e., x ď y) there is an element u P rx, ys, such that
z ^ u “ x and z _ u “ y (u is a “complement” of z in the interval rx, ys).
b) L is atomic, i.e.: Every element can be represented as the supremum of atoms.
(A finite semimodular lattice obeying one of these conditions is called geometric.)
54 3. POSETS

Exercise 28: Let G be a (labelled) graph on n vertices. A partition of the vertex–set V pGq is called
connected if every block of the partition corresponds to a connected induced subgraph of G. The set
of all connected partitions is a subposet of the poset of partitions of V pGq, and thus a poset itself.
(If G is the complete graph, the poset of connected partitions of V pGq is the same as the poset of
all partitions of V pGq.)
Show that the poset of connected partitions of G is a geometric lattice.

Exercise 29: A lattice L is called modular if it is graded and for all x, y P L there holds:
rk pxq ` rk pyq “ rk px ^ yq ` rk px _ yq . (3.5)
(In particular, the lattice L pVq of subspaces of a finite vector space is modular.)
Show: A finite lattice L is modular if and only if for all x, y, z P L with x ď z there holds:
x _ py ^ zq “ px _ yq ^ z. (3.6)

Hint: Show that (3.6) implies the Diamond Property: The mappings
ψ : ry, y _ zs Ñ ry ^ z, zs , ψ pxq “ x ^ z
ϕ : ry ^ z, zs Ñ ry, y _ zs , ϕ pxq “ x _ y
are order preserving bijections with φ ˝ ψ “ id, see the following picture:

y_z

ϕ
y z

y^z

Exercise 30: Show: The lattice Πn of all partitions of an n–element set is not modular.

3.3.1. Distributive lattices. A particularly important class of lattices are the

distributive lattices.
D EFINITION 3.3.9. A lattice L is said to be distributive, if, for all a, b, c P L, we have
the following laws of distributivity:
a ^ pb _ cq “ pa ^ bq _ pa ^ cq ,
a _ pb ^ cq “ pa _ bq ^ pa _ cq .
R EMARK 3.3.10. It suffices to require either of these laws as the other then follows. We
show how to deduce the second law from the first law:
pa _ bq ^ pa _ cq “ ppa _ bq ^ aq _ ppa _ bq ^ cq Ð1. dist.l.

“ a _ ppa _ bq ^ cq Ðabsorption

“ a _ ppa ^ cq _ pb ^ cqq Ð1. dist.l.

“ a _ pb ^ cq Ðass.l. and absorption

 3.3. LATTICES 55

R EMARK 3.3.11. Let L1 and L2 be two distributive lattices. Then also the direct prod-
uct L1 ˆ L2 is a distributive lattice.
E XAMPLE 3.3.12. Examples for distributive lattices are:
(1) Linear order
(2) Boolean lattice Bn : Here we have A _ B “ A Y B and A ^ B “ A X B; the
laws of distributivity are the set theoretic identities
A X pB Y Cq “ pA X Bq Y pA X Cq
A Y pB X Cq “ pA Y Bq X pA Y Cq
(3) Divisor lattice Tn “ td P N : d | nu: Indeed, let n “ p1k1 ¨ ¨ ¨ prkr be the unique
prime factor decomposition of n, then
Tn » r0, k1 s ˆ ¨ ¨ ¨ r0, kr s .
(4) For an arbitrary poset P, the lattice of order ideals J pPq is always distributive.
On the other hand, simple examples of lattices that are not distributive are the following
(given by their Hasse diagram):

1̂ 1̂
c
M5 a b c N5 b
a
0̂ 0̂
a^pb_cq‰pa^bq_pa^cq a_pb^cq‰pa_bq^pa_cq

T HEOREM 3.3.13 (Structure theorem for finite distributive lattices, Birkhoff’s

Theorem). Let L be a finite distributive lattice. Then there exists a poset P such that
L » J pPq .
D EFINITION 3.3.14. Let L be lattice. An element x ‰ 0̂ P L is said to be supremum-
irreducible, if x cannot be represented as follows
x “ y _ z,
where y ă x and z ă x.
L EMMA 3.3.15. Let L be distributive lattice, and let a P L be supremum–irreducible.
Then we have
a ď x _ y ùñ pa ď x oder a ď yq .
P ROOF. As a “ a ^ px _ yq “ pa ^ xq _ pa ^ yq, it follows from the assumption
that
a ^ x “ a oder a ^ y “ a,
which implies a ď x or a ď y. 
P ROOF OF 3.3.13. Let P be the set of supremum–irreducible elements of L, in-
terpreted as a poset with the order relation “inherited” from L.
We define a map ϕ : L Ñ J pPq by
` ˘ (
ϕ 0̂ “ H; ϕ pxq :“ I pxq :“ x1 P P : x1 ď x for x ‰ 0̂. (3.7)
56 3. POSETS

Claim: ϕ is bijective with inverse map

ł
ψ : I ÞÑ y
yPI

for I ‰ H P J pPq (Special case: ψ pHq “ 0̂).

For the special case x “ 0̂ resp. y “ H we have ϕ pxq “ y and ψ pyq “ x. It
remains to show:
@x P L, x ‰ 0̂ : pψ ˝ ϕq pxq “ x, (3.8)
@y P J pPq , y ‰ H : pϕ ˝ ψq pyq “ y. (3.9)
We first show (3.8). For this purpose, we observe that x ‰ 0̂ is always repre-
sentable as supremum of supremum–irreducible elements: In case x is itself
supremum–irreducible, this is of course trivial:
ł
x“ y.
y“x

Therefore suppose that x is not supremum–irreducible, then, by definition, there

exist elements x1 ă x and x2 ă x with x “ x1 _ x2 . The same considera-
tions can also be applied to x1 and x2 , etc.: Since L is finite, the algorithm must
stop at some point, and we obtain the required decomposition into supremum–
irreducible elements
x “ x1 _ x2 _ ¨ ¨ ¨ _ xk with xi ă x for i “ 1, . . . , k.
This implies
tx1 , . . . , xk u Ď ϕ pxq “ I pxq ,
that is ł
x “ x1 _ x2 _ ¨ ¨ ¨ _ x k ď y ď x,
yPIpxq
Ž
and we can conclude yPIpxq y “ x, and so (3.8) is proven.
Now we prove (3.9): Suppose y “ tx1 , x2 . . . , xk u P J pPq. Let x “ ψ pyq “
x1 _ x2 _ ¨ ¨ ¨ _ xk . It is obvious that y Ď ϕ pxq; it remains to show ϕ pxq Ď y, that
is that each supremum–irreducible element x1 with x1 ď x is included in y:
` ˘ ` ˘
x 1 “ x 1 ^ x “ x 1 ^ x1 _ ¨ ¨ ¨ _ x 1 ^ x k ,
and, since x1 is supremum–irreducible, there is an i with x1 “ x1 ^ xi : However,
this means that x1 ď xi , and thus (3.9) is proven.
Claim: ϕ is order preserving, i.e.:
x1 ď x2 ðñ I px1 q Ď I px2 q .
Consider for x1 ď x2 an arbitrary element x1 P I px1 q: x1 ď x1 ùñ x1 ď x2 ùñ
x1 P I px2 q, therefore I px1 q Ď I px2 q. Conversely, I px1 q Ď I px2 q implies
¨ ˛ ¨ ˛
ł ł ł ł
yď˝ y‚_ ˝ y‚ “ y.
yPIpx1 q yPIpx1 q yPIpx2 qzIpx1 q yPIpx2 q

The two claims imply Theorem 3.3.13. 

E XAMPLE 3.3.16. The following example illustrates the situation:
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 57

ta,b,cu

c L c P tc,bu J pPq ta,bu

b a b a tbu tau

T HEOREM 3.3.17. Let L be a finite distributive lattice. Then L has a rank function rk;
it is given by
rk pxq “ |I pxq| ,
where I pxq is the ideal defined in (3.7) (of the poset of supremum-irreducible elements
of L). In particular, it follows that the rank rk pLq of L is equal to the number of
supremum-irreducible elements of L.
P ROOF. By Birkhoff’s Theorem (Theorem 3.3.13), we have
L » J pPq ,
where P is the poset of supremum-irreducible elements in L. Consequently, it
suffices to show that all maximal chains in J pPq have the same length.
Let
0̂ “ H “ I0 Ì I1 Ì ¨ ¨ ¨ Ì Ir “ 1̂ “ P
be a maximal chain in P. We claim: |Ik | “ |Ik´1 | ` 1. Indeed, if S :“ Ik zIk´1 “
tx1 , . . . , xm u, then, using S ‰ H, there exists a maximal element xi (with respect
to the order of P) in S. But then Ik z txi u is also an order ideal of P (see Corol-
lary 3.1.9), and, since Ik Í Ik´1 , we can conclude Ik´1 “ Ik z txi u. Hence all
maximal chains in J pPq have the same length |P|. 

3.4. Incidence algebra and Möbius inversion

Let A, B and C be three sets with
A X B “ A X C “ B X C “ A X B X C “: D.

AzD CzD
D

BzD

Obviously (following the inclusion-exclusion principle) we have:

|A Y B Y C| “ 1 ¨ |A| ` 1 ¨ |B| ` 1 ¨ |C| ` p´2q ¨ |D| .
Evidently, the coefficients that can occur (in our case 1, 1, 1, ´2) only depend on
the order relation (set inclusion) of the involved sets A Y B Y C, A, B, C and
D: Indeed, it is possible to compute these coefficients quite elegantly from the
order relation as we will see in the following.
58 3. POSETS

3.4.1. The incidence algebra of locally finite posets.

D EFINITION 3.4.1. Let P be a locally finite poset (recall that this means that for all
x, y P P with x ď y the interval rx, ys is finite). We consider the set I pPq of all
functions
f : PˆP Ñ C
with the property
f px, yq ” 0 if x ę y (3.10)
(that is to say that the function f is only “defined on intervals”) with the ordinary
addition and scalar multiplication of functions
f , g P I pPq : p f ` gq px, yq :“ f px, yq ` g px, yq ,
f P I pPq , λ P C : pλ ¨ f q px, yq :“ λ ¨ p f px, yqq
and the following (non-commutative) multiplication (convolution):
ÿ
p f ‹ gq px, yq :“ f px, zq g pz, yq .
zPP
The range of summation is finite (!), as P is assumed to be locally finite, and, by (3.10),
the sum can also be written as:
ÿ ÿ
f px, zq g pz, yq “ f px, zq g pz, yq .
zPP xďzďy

Evidently, I pPq is a vector space over C under the addition and scalar multiplication.
As for the multiplication (convolution), it satisfies the law of associativity:
ÿ
pp f ‹ gq ‹ hq px, yq “ p f ‹ gq px, zq h pz, yq
zPP
˜ ¸
ÿ ÿ
“ f px, uq g pu, zq h pz, yq
zPP uPP
ÿÿ
“ f px, uq g pu, zq h pz, yq
uPP zPP
ÿ
“ f px, uq pg ‹ hq pu, yq
uPP
“ p f ‹ pg ‹ hqq px, yq .
Furthermore, we have the laws of distributivity
f ‹ pg ` hq “ f ‹ g ` f ‹ h,
p f ` gq ‹ h “ f ‹ h ` g ‹ h
(as can be shown easily). Finally, with
#
1 : x “ y,
δ P I pPq : δ px, yq “ ,
0 otherwise
we obviously have
δ ‹ f “ f ‹ δ “ f,
that is, I pPq is a (non-commutative) algebra with unit δ: We refer to it as the inci-
dence algebra of P.
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 59

E XAMPLE 3.4.2. For P “ rns, I pPq is equal to the algebra of n ˆ n upper triangular
matrices over C: Indeed, each value f px, yq of an element f P I pPq can be interpreted
as the entry in row x and columns y of a corresponding triangular matrix M, and,
conversely, each such matrix defines uniquely a function f P I pPq. It can also be seen
easily that the multiplication in I pPq corresponds to matrix multiplication.
T HEOREM 3.4.3. Let P be a locally finite poset. An element f P I pPq has an inverse
element f ´1 (i.e.: D f ´1 P I pPq with f ‹ f ´1 “ f ´1 ‹ f “ δ) if and only if f px, xq ‰
0 @x P P. If an inverse exists, then it is uniquely determined.

P ROOF. One implication ( ùñ ) is obvious:

´ ¯ ÿ
´1
1 “ f ‹ f px, xq “ f px, zq f ´1 pz, xq ùñ f px, xq ‰ 0.
xďzďx

As for the other implication ( ðù ), we consider an f with f px, xq ‰ 0 @x P

1
P. We define g px, xq :“ f px,xq , and, for x ă y, recursively (“inductively with
respect to the length of the intervals rx, ys”):
ÿ
f px, zq g pz, yq “ 0,
xďzďy
ř
i.e., g px, yq f px, xq “ ´ xăzďy f px, zq g pz, yq and, therefore,
1 ÿ
g px, yq “ ´ f px, zq g pz, yq .
f px, xq xăzďy

The so constructed g satisfies f ‹ g “ δ; analogously, we can also find an h with

g ‹ h “ δ: Then we have
f “ f ‹ δ “ f ‹ pg ‹ hq “ p f ‹ gq ‹ h “ δ ‹ h “ h,
and so we also have g ‹ f “ δ, and f ´1 :“ g is the inverse. 

3.4.2. The zeta function of locally finite posets.

D EFINITION 3.4.4. Let P be a locally finite poset. The function ζ P I pPq:
#
1 for x ď y
ζ px, yq “
0 sonst

is said to be the zeta function of P.

R EMARK 3.4.5. From an abstract point of view, the zeta function of a poset “is” es-
sentially the order relation of the poset: Indeed, each relation „ on P “is” a subset of
the cartesian product P ˆ P (namely, the set of all pairs px, yq with x „ y), and the
zeta function is from this point of view simply the characteristic function of the order
relation.
For P “ rns (see Example 3.4.2), the zeta function corresponds to the upper triangular
matrix that has only 1’s above the main diagonal.
60 3. POSETS

The following functions are useful: For instance,

ÿ
ζ 2 px, yq “ ζ px, zq ζ pz, yq “ # pelements in rx, ysq ,
looooooomooooooon
xďzďy
1
and, more general,
ÿ
ζ k px, yq “ ζ px, x1 q ζ px1 , x2 q ¨ ¨ ¨ ζ pxk´1 , yq,
looooooooooooooooooomooooooooooooooooooon
x0 “xďx1 﨨¨ďxk´1 ďxk “y
1
this is the number of “multichains” of length k from x to y. In order to count
“ordinary” chains of length k from x to y, it is useful to consider pζ ´ δq:
#
1 für x ă y
pζ ´ δq px, yq “
0 otherwise

The number of interest in then pζ ´ δqk px, yq.

R EMARK 3.4.6. Although the algebra is not commutative, each element of course com-
mutes with the unit δ. This implies the following special case of the Binomial Theo-
rem:
ÿk ˆ ˙
k k
pζ ´ δq px, yq “ p´1q j ζ k´j px, yq .
j
j“0
This identity is also a special case of the inclusion-exclusion principle: Indeed, if we
delete precisely j of those elements in a “multichain”
px “ x0 , x1 , . . . , xk “ yq
of length k from x to y that are repeated after their first appearance, then we are
`˘
left with a “multichain” of length pk ´ jq, and, conversely, there are kj possibilities
to construct a “multichain” of length k from a “multichain” of length pk ´ jq: Each
“multichain” of length pk ´ jq consists of k ` 1 ´ j elements, and each of these elements
can be repeated (also multiple times!): Since we need to add a total of j “repetitions”,
this corresponds to a multiset with j elements selected from a set of with pk ` 1 ´ jq
elements, and therefore we have
ˆ ˙ ˆ ˙
pk ` 1 ´ jq ` j ´ 1 k
“ .
j j
An additional example is p2δ ´ ζq,
$
’
&1 if x “ y,
p2δ ´ ζq “ ´1 if x ă y,
’
%0 otherwise.
According to Theorem 3.4.3, this function is invertible.
Now we aim at counting all chains (of any length) from x to y. Let l be the length
of the interval rx, ys, then, according to considerations above, the number is
l
ÿ
pζ ´ δqk px, yq ,
k“0
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 61

and, using a well-known trick (telescoping series) for the geometric series, we
have
l
ÿ
pδ ´ pζ ´ δqq ‹ pζ ´ δqk “ δ ´ loooomoooon
pζ ´ δql`1 “ δ,
k“0
“0

and so the number is given by p2δ ´ ζq´1 px, yq.

3.4.3. Möbius inversion.

D EFINITION 3.4.7. The zeta function ζ of a locally finite poset P is according to The-
orem 3.4.3 invertible: µ :“ ζ ´1 is said to be the Möbiusfunction of P. By definition,
we have:
ÿ ÿ
ζ px, zq ¨µ pz, yq “
loomoon µ pz, yq “ 0 for all x ă y, (3.11)
xďzďy xďzďy
“1
ÿ ÿ
µ px, zq ¨ loomoon
ζ pz, yq “ µ px, zq “ 0 for all x ă y (3.12)
xďzďy xďzďy
“1

and µ px, xq ” 1.

T HEOREM 3.4.8 (Möbius inversion). Let P be a locally finite poset such that every
principal order ideal is finite. Let f , g : P Ñ C be functions on P with values in C.
Then the following assertions are equivalent:

ÿ
g pxq “ f pyq @x P P (3.13)
yďx
ÿ
f pxq “ g pyq µ py, xq @x P P. (3.14)
yďx

R EMARK 3.4.9. This is a generalisation of Möbius inversion in number theory: The

natural numbers, ordered by the divisor relation, is a locally finite poset such that every
principal ideal (the principal ideal of n is the set of divisors of n) is finite:

ÿ ÿ ´n¯
g pnq “ f pdq ðñ f pnq “ µ g pdq
d
d|n d|n

(the interval rd, ns in this poset is isomorphic to r1, n{ds, that is µ pd, nq “ µ p1, n{dq “:
µ pn{dq).
62 3. POSETS

P ROOF OF THE M ÖBIUS INVERSION . For the proof, we only need to perform
two simple calculations:
ÿ ÿ ÿ
g pyq ¨ µ py, xq “ f pzq ¨ µ py, xq Ðby (3.13)
yďx yďx zďy
ÿ ÿ
“ µ py, xq f pzq ¨ ζ pz, yq Ðdefinition ζ
yďx zďy
ÿ ÿ
“ f pzq ζ pz, yq ¨ µ py, xq
zPP zďyďx
ÿ
“ f pzq δ pz, xq
zPP
“ f pxq ,
and, conversely:
ÿ ÿ ÿ
f pxq “ g pyq ¨ µ py, xq Ðby (3.14)
xďs xďs yďx
ÿ ÿ
“ ζ px, sq g pyq ¨ µ py, xq Ðdefinition ζ
xďs yďx
ÿ ÿ
“ g pyq µ py, xq ¨ ζ px, sq
yPP yďxďs
ÿ
“ g pyq δ py, sq
yPP
“ g psq .

R EMARK 3.4.10. From an abstract point of view, the key observation is that the equa-
tions (3.13) and (3.14) are equivalent to the “symmetric” equations
ÿ
g pxq “ f pyq ¨ ζ py, xq @x P P,
yPP
ÿ
f pxq “ g pyq ¨ µ py, xq @x P P,
yPP

and that I pPq acts on the vector space C P of all functions P Ñ C as algebra of linear
mappings (from the right) with
ÿ
p f ξq pxq :“ f pyq ξ py, xq für f P C P , ξ P I pPq .
yPP

Möbius inversion is then nothing else but

f ζ “ g ðñ f “ gµ.

We have the following dual version of Theorem 3.4.8:

T HEOREM 3.4.11. Let P be a locally finite poset such that every dual principal order
ideal is finite. Let f , g : P Ñ C be functions on P with values in C. Then the following
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 63

assertions are equivalent:

ÿ
g pxq “ f pyq @x P P (3.15)
yěx
ÿ
f pxq “ µ px, yq g pyq @x P P. (3.16)
yěx

E XAMPLE 3.4.12. Let S1 , S2 , . . . , Sn be sets. Consider the poset that consists of all in-
tersections of selections of these sets, ordered by inclusion. Here the empty intersection
is defined as č
Si :“ S1 Y S2 Y ¨ ¨ ¨ Y Sn “ 1̂.
iPH
We aim at deriving a formula for |S1 Y S2 Y ¨ ¨ ¨ Y Sn | and
Ť define for this purpose func-
tions g, f : P Ñ C as g pTq :“ |T| and f pTq :“ |Tz p T1 ăT T 1 q| (in words: f pTq is
the number of elements in T that are not contained in any T 1 P P with T 1 Ř T).
First we observe that in general we have the following (think for a moment!):
ÿ ` ˘
g pTq “ f T1 .
T 1 ĎT
Using Möbius inversion, we obtain
ÿ ` ˘ ` ˘
f pTq “ g T1 ¨ µ T1, T .
T 1 ĎT
In particular, we have
` ˘ ÿ ` ˘ ` ˘ ÿ ` ˘ˇ ˇ
0 “ f 1̂ “ g T 1 ¨ µ T 1 , 1̂ “ µ T 1 , 1̂ ˇT 1 ˇ ,
T 1 Ď1̂ T 1 Ď1̂
we leads us to the following formulation of the inclusion-exclusion principle:
ˇ ˇ ÿ ` ˘ˇ ˇ
|S1 Y S2 Y ¨ ¨ ¨ Y Sn | “ ˇ1̂ˇ “ ´ µ T 1 , 1̂ ˇT 1 ˇ .
T 1 Ř1̂

In the introductory example (see the beginning of Abschnitt 3.4) we then have
the following:
` ˘
1̂ µ 1̂, 1̂ “ 1
` ˘
A B C µ ., 1̂ “ ´1
` ˘
0̂ µ 0̂, 1̂ “ 2

3.4.4. Calculation of Möbius functions of several concrete posets. Möbius

inversion is of course only useful if we can compute the Möbius function.
E XAMPLE 3.4.13. Let P “ N (with the natural order). It follows immediately from
the defining equation (3.11) that
$
’
&1 x “ y,
µ px, yq “ ´1 y “ x ` 1,
’
%0 otherwise.
64 3. POSETS

The Möbius inversion implies the “discrete analogue of the fundamental theorem of
calculus”
#
ÿn
g p1q for n “ 1,
g pnq “ f piq @n ě 1 ðñ f pnq “
i“1
g pnq ´ g pn ´ 1q for n ą 1.
From an abstract point of view, this means that the summation operator
n
ÿ
pΣ cq pnq :“ c piq
i“1
and the difference operator
#
c p1q n“1
p∆ cq pnq :“
c pnq ´ c pn ´ 1q ną1
are inverse operators on the vector space CN of all functions c : N Ñ C.
T HEOREM 3.4.14. Let P and Q be locally finite posets. For the direct product P ˆ Q,
we have
µPˆQ ppx1 , y1 q , px2 , y2 qq “ µP px1 , y1 q ¨ µQ px2 , y2 q .
P ROOF. Recall that in the direct product P ˆ Q we have
px1 , y1 q ď px2 , y2 q :ðñ x1 ď x2 and y1 ď y2 .
We check whether the given function has the defining property of the Möbius
function:
ÿ
µP px3 , x2 q ¨ µQ py3 , y2 q “
px1 ,y1 qďpx3 ,y3 qďpx2 ,y2 q
ÿ ÿ
µP px3 , x2 q ¨ µQ py3 , y2 q “ δP px1 , x2 q ¨ δQ py1 , y2 q
x1 ďx3 ďx2 y1 ďy3 ďy2
“ δPˆQ ppx1 , y1 q , px2 , y2 qq .

E XAMPLE 3.4.15 (Classical Möbius function). The divisor lattice Tn :“ td : d | nu
(order relation “ divisor relation) for n with prime factor decomposition n “ p1k1 ¨ p2k2 ¨
pkmm is isomorphic to the product of linear posets
Tn » r0, k1 s ˆ ¨ ¨ ¨ ˆ r0, km s .
By Theorem 3.4.14, we obtain:
´ ¯
β β
µTn px, yq “ µTn p1α1 ¨ ¨ ¨ pαmm , p1 1 ¨ ¨ ¨ pmm
“ µr0,k1 s pα1 , β 1 q ¨ ¨ ¨ µr0,km s pαm , β m q
# řm
p´1q i“1 pβi ´αi q if pβ i ´ αi q P t0, 1u @i,
“
0 otherwise.
Since we have the following isomorphy
”yı
rx, ys “ 1,
x
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 65

in the divisor lattice, the Möbius function satisfies

µ pd, d ¨ kq “ µ p1, kq “: µ pkq ,
and we obtain the classical Möbius function appearing in number theory
#
p´1qm k “ p1 ¨ ¨ ¨ pm pi ‰ j ùñ pi ‰ p j q,
µ pkq “
0 sonst.
E XAMPLE 3.4.16 (Euler’s ϕ–function). Euler’s ϕ–function, which also appears in
number theory, is defined as
φ pnq :“ |tx P rns : ggT px, nq “ 1u| .
(In words: φ pnq is the number of positive x no greater than n that are coprime to n.
For instance, if n “ 12 these numbers are 1, 5, 7, 11, and so we have ϕ p12q “ 4.)
Now the following holds ÿ
n“ φ pdq .
d|n
This is because rns is the disjoint union
ď ! n )
9
rns “ d|n y ¨ : y ď d and ggT py, dq “ 1 .
d
looooooooooooooooooooomooooooooooooooooooooon
Number: ϕpdq

Using Möbius inversion (set g pnq “ n and f pnq “ φ pnq), we obtain

ÿ
φ pnq “ µ pd, nq d
d|n
ÿ ´n ¯ n
“ µ ,n
d d
d|n
ÿ n
“ µ p1, dq
d
d|n
ÿ 1
“n µ pdq
d
d|n
źˆ 1
˙
“n 1´ .
p
p|n
k
The last step can be justified as follows: Let n “ p11 ¨ ¨ ¨ pkmm , then the penultimate sum
ranges over all d that are products of distinct primes, that is d “ pi1 ¨ ¨ ¨ pi j (the Möbius
function is 0 for all other divisors) with µ pdq “ p´1q j .
´ ¯ ´ ¯
For example, we have ϕ p12q “ 12 ¨ 1 ´ 12 ¨ 1 ´ 31 “ 4.

E XAMPLE 3.4.17 (Boolean lattice). The power set 2rns of rns can be interpreted as
family of characteristic functions (see Example 3.1.4):
A Ď rns Ø f A : rns Ñ t0, 1u , f A pxq “ rx P As 1.
1Iversons notation.
66 3. POSETS

This implies that the following posets are isomorphic:

2rns » t0, 1un .

The Möbius function of the (linearly order) poset t0, 1u is obviously
µ p0, 0q “ µ p1, 1q “ 1; µ p0, 1q “ ´1.
This implies for T Ď S Ď rns immediately
n
ź
µ pT, Sq “ µ pri P Ts , ri P Ssq “ p´1q|SzT| .
i“1

The inclusion-exclusion principle can also be formulated as follows: Let E be a

set of properties that might or might not be fulfilled for individual elements
of a ground set A (e.g., A could be the set of permutations of rns, and E “
tE1 , . . . , En u could consist of the following n properties
Ei :“ “has i as fixpoint”, i “ 1, 2, . . . , n.q
For an arbitrary subset S Ď E, let
f “ pSq :“ # px P A, that have precisely the properties in Sq
(that is these elements have no properties “outside” of S), and, for an arbitrary
subset T Ď E, let
f ě pTq :“ # px P A, that have at least the properties in Tq
(that is these elements can also have properties “outside” of T; e.g., the number
of permutations in Sn , that have at least the k different fixpoints i1 , i2 , . . . , ik , is
equal to pn ´ kq!). Evidently, we have
ÿ
f ě pTq “ f “ pSq .
SĚT

Using Möbius inversion, we have:

ÿ
f “ pTq “ p´1q|SzT| f ě pSq .
SĚT

For the special case T “ H, we have:

ÿ
f “ pHq “ p´1q|S| f ě pSq . (3.17)
S

E.g., the number of permutations in Sn without fixpoint (i.e., the well-known

number of derangements) is:
n ˆ ˙
ÿ n
ÿ
n p´1qk
p´1qk pn ´ kq! “ n! .
k k!
k“0 k“0

From this point of view, (3.17) is indeed the inclusion-exclusion principle:

 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 67

C OROLLARY 3.4.18 (Inklusion–exclusion). Let A1 , A2 , . . . , An be n (not necessarily

disjoint!) sets. Then we have:
|A1 Y ¨ ¨ ¨ Y An | “ |A1 | ` ¨ ¨ ¨ ` |An | ´ |A1 X A2 | ´ ¨ ¨ ¨
´ |An´1 X An | ` |A1 X A2 X A3 | ` ¨ ¨ ¨ ´ ¨ ¨ ¨
` p´1qn´1 |A1 X A2 X ¨ ¨ ¨ X An |
P ROOF. In the context of Example 3.4.17: Let A “ A1 Y ¨ ¨ ¨ Y An be sets, and let
E “ tE1 , . . . , En u be the following n properties
Ei :“ “is contained in Ai ”, i “ 1, 2, . . . , n.
Then we obviously have f “ pHq “ 0, and the assertion follows immediately
from (3.17). 
C OROLLARY 3.4.19. Let L be a finite distributive lattice. The rank function on L
satisfies
rk px1 _ ¨ ¨ ¨ _ xn q “ rk px1 q ` ¨ ¨ ¨ ` rk pxn q ´ rk px1 ^ x2 q ´ ¨ ¨ ¨
´ rk pxn´1 ^ xn q ` rk px1 ^ x2 ^ x3 q ` ¨ ¨ ¨ ´ ¨ ¨ ¨
` p´1qn´1 rk px1 ^ ¨ ¨ ¨ ^ xn q .
P ROOF. According to Birkhoff’s Theorem (Theorem 3.3.13), there exists a poset
P such that
L » J pPq .
In the isomorphy, each x P L corresponds to an order ideal Ix of P, and rk pxq “
|Ix |. In addition Ix^y “ Ix X Iy : This follows immediately from Corollary 3.4.18.

3.4.5. Möbius algebra of a locally finite lattice.
D EFINITION 3.4.20. Let L be a locally finite
ř poset. We consider the complex vector
space of the formal linear combinations xPL c x x with coefficients c x P C. With the
bilinear (that is: “extended distributively”) multiplication
x ¨ y :“ x ^ y @x, y P L,
this vector space becomes a commutative algebra A pLq. The algebra is frequently called
Möbiusalgebra. If L is finite, then L is obviously a basis of the vector space of idempo-
tent elements (x ¨ x “ x ^ x “ x).
For each x P L, we define the vector δx P A pLq as linear combination:
ÿ
δx :“ µ py, xq y.
0̂ďyďx

For the special vectors δx P A pLq, Möbius inversion gives

ÿ ÿ ÿ ÿ ÿ
δy “ µ ps, yq s “ s µ ps, yq “ x. (3.18)
yďx yďx sďy sďx sďyďx
(
In other words: The elements δy : y P L span the vector space A pLq; under the as-
sumption that L is finite, they are also a basis.
68 3. POSETS

L EMMA 3.4.21. Let L be a locally finite lattice. In A pLq we have:

δx ¨ δy “ rx “ ys δx . (3.19)
P ROOF. First we note that
˜ ¸ ¨ ˛
ÿ ÿ
δx ¨ δy “ µ pz1 , xq z1 ¨˝ µ pz2 , yq z1 ‚
z1 ďx z2 ďy
ÿ ÿ
“ µ pz1 , xq µ pz2 , yq z1 ^ z2 .
z1 ďx z2 ďy
ř
By (3.18), z1 ^ z2 “ sďz1 ^z2 δs , this is further equal to
¨ ˛¨ ˛
˚ ‹˚ ‹
ÿ ˚ ÿ ‹˚ ÿ ‹
˚ ‹˚ ‹
δs ˚ µ pz1 , xq‹ ˚ µ pz2 , yq‹ “ rx “ ys δx .
˚ ‹ ˚ ‹
sďx^y ˝looooooomooooooon
sďz1 ďx ‚˝looooooomooooooon
sďz2 ďy ‚
“rs“xs “rs“ys


C OROLLARY 3.4.22. Let L be a finite lattice with at least 2 elements, and let 1̂ ‰ a P L.
Then we have ÿ ` ˘
µ x, 1̂ “ 0. (3.20)
x:x^a“0̂

P ROOF. According to (3.18) and (3.19), we have on the one hand:

˜ ¸
ÿ
a ¨ δ1̂ “ δb ¨ δ1̂ “ 0 P A pLq if a ‰ 1̂.
bďa
On the other hand, by definition of δx , we have
ÿ ` ˘ ÿ ` ˘
a ¨ δ1̂ “ a ¨ µ x, 1̂ x “ µ x, 1̂ x ^ a.
xPL xPL
Now a ¨ δ1̂ can be expanded (as any other element) in the basis L:
ÿ
a ¨ δ1̂ “ c x x,
xPL
and, as a ¨ δ1̂ “ 0 (zero vector in A pLq), we have in particular c0̂ “ 0:
ÿ ` ˘
0 “ c0̂ “ µ x, 1̂ .
x:x^a“0̂
ř ` ˘
(This sum is similar to the defining relation x µ x, 1̂ “ 0, however, it contains
less terms in general.) 
E XAMPLE 3.4.23 (Möbius function of the partition lattice). For two partitions σ,
π of rns, we have σ ď π : ðñ “all blocks of σ are contained in blocks of π”. For
instance:
π0 “ 1 4 5 6 7|2 8 9 10|3
σ0 “ 1|4 5|6 7|2|8 9 10|3
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 69

In this example, it can be seen immediately that: rσ0 , π0 s » Π3 ˆ Π2 ˆ Π1 .

More generally: Let π “ tB1 , B2 , . . . , Bk u, and let precisely λi blocks of σ be con-
¯ we have´rσ, πs » Πλ1 ¯ˆ Πλ2 ˆ ¨ ¨ ¨ ˆ Πλk ; here we have σ „
tained in Bi , then
´
0̂Πλ , . . . , 0̂Πλ and π „ 1̂Πλ , . . . , 1̂Πλ . We conclude that
1 k 1 k
´ ¯ ´ ¯ ´ ¯
µ pσ, πq “ µ 0̂Πλ , 1̂Πλ ¨ µ 0̂Πλ , 1̂Πλ ¨ ¨ ¨ µ 0̂Πλ , 1̂Πλ .
1 1 2 2 k k

It follows
` that we ˘can compute the Möbius function of Πn in general, if we can deter-
mine µ 0̂Πm , 1̂Πm for all m P N. We use Corollary 3.4.22. Let a “ 1 2 . . . pm ´ 1q |m.
If x ^ a “ 0̂Πm , then we either have x “ 0̂Πm or x “ m i| premainder in blocks with 1 elementq
for an i “ 1, 2, . . . , m ´ 1. According to (3.20), we have
` ˘ ÿ ` ˘
µm :“ µ 0̂Πm , 1̂Πm “ ´ µ x, 1̂Πm “ ´ pm ´ 1q µm´1 .
x“m i|premainderq

With the initial condition µ1 “ 1, we have

` ˘
µm “ µ 0̂Πm , 1̂Πm “ p´1qm´1 pm ´ 1q! (3.21)
For our concrete example, we obtain
µ pσ0 , π0 q “ p´1q2 2! p´1q 1!1 “ ´2.
C OROLLARY 3.4.24. Let L be finite lattice, and let X Ď L be such that
‚ 1̂ L R X,
‚ y P L and y ‰ 1̂ L ùñ Dx P X with y ď x.
(That is, X contains all coatoms of L.) Then we have
` ˘ ÿ
µ 0̂ L , 1̂ L “ p´1qk Nk ,
k

where Nk the number of all k–subsets of X with infimum 0̂ L .

P ROOF. For x P L, we have in A pLq

ÿ ÿ ÿ
1̂ L ´ x “ δy ´ δy “ δy .
yď1̂ L yďx yęx

On the one hand, this implies

ź` ˘ ÿ
1̂ L ´ x “ δy Ð by (3.19)
xPX y:yęx @xPX
“ δ1̂L . Ð by assumption

On the other hand, we can also simply expand the product in order to obtain
an expansion in terms of the basis L:
ÿ ÿ ÿ ` ˘
1̂ L ´ x` x^y´`¨¨¨ “ µ x, 1̂ L x .
xPX x,yPX looooooomooooooon
xPL
δ1̂
L

As comparison of the coefficients of x “ 0̂ L yields the assertion. 

70 3. POSETS

R EMARK 3.4.25. In Corollary 3.4.24, it is of course natural to let X be the set of

coatoms of L. In particular, we have
` the following:
˘ If 0̂ L is not representable as the
infimum of coatoms, then we have µ 0̂ L , 1̂l “ 0.

C OROLLARY 3.4.26. Let L be a finite lattice. Suppose 0̂ L is not representable as the

infimum of coatoms or 1̂ L is not representable as the supremum of atoms, then we have
` ˘
µ 0̂ L , 1̂ L “ 0.
E XAMPLE 3.4.27 (Möbiusfunction of a finite distributive lattice). For a finite dis-
tributive lattice L, Birkhoff’s Theorem (Theorem 3.3.13) implies:
L » J pPq .
Let I Ď I 1 P J pPq, then we have
#
p´1q| I zI |
1
` 1˘ if rI, I 1s » Boolean lattice,
µ I, I “
0 otherwise.
Indeed, let I 1 “ I Y tx1 , . . . , xn , . . . , xm u, where I Y tx1 u , . . . , I Y txn u are the atoms
in rI, I 1 s. Then there are only two possibilities:
Case 1: I Y tx1 , . . . , xn u “ I 1 , then we have rI, I 1s » 2rns .
Case 2: I Y tx1 , . . . , xn u ‰ I 1 , then we have rI, I 1s fi 2rns , because in the Boolean lattice
1̂ is always representable as the supremum of atoms.

(See also the paper of Rota and Stanley [6] regarding Möbiusfunctions.)
E XAMPLE 3.4.28 (Möbiusfunction of Vn pqq “ GF pqqn ). Let V,”!W)ď Vn pqqı be two
subspaces with V ď W. Then the interval rV, Ws is isomorphic to ~0 , W{V , where
´! ) dim pV{Wq
the latter interval only depends on the dimension ¯ “ dim pVq ´ dim pWq.
Therefore, it suffices to compute µn :“ µ ~0 , Vn pqq . For this purpose, we use the
dualised version of Corollary 3.4.22:
ÿ ` ˘
µ 0̂, x “ 0. (3.22)
x:x_a“1̂

Let a be an atom: From x _ a “ 1̂, we conclude that

‚ either a ď x ùñ x “ 1̂,
‚ or a ę x ùñ a ^ x “ 0̂.
In Vn pqq we have in general
rk pxq ` rk pcq “ rk px ^ cq ` rk px _ cq
Therefore, a ^ x “ 0̂ implies immediately rk pxq ` 1 “ 0 ` n, and so: If x is a coatom,
then (3.22) implies
` ˘ ÿ ` ˘
µ 0̂, 1̂ “ ´ µ 0̂, x .
aęx,x coatom
 3.4. INCIDENCE ALGEBRA AND MÖBIUS INVERSION 71
“ n ‰
Let a be a fixed atom in Vn pqq, i.e., a subspace of dimension 1. Vn pqq has n´1 q “
rnsq “ qn´1 ` qn´2 ` ¨ ¨ ¨ ` q ` 1 coatoms, that is subspaces 2 of dimension n ´ 1,
“n´1‰
among these n´2 q
“ rn ´ 1sq “ qn´2 ` ¨ ¨ ¨ ` q ` 1 include the subspace a. Moreover,
we have
` ˘
µn :“ µVn pqq 0̂, 1̂ “ ´qn´1 µn´1 .
This implies, using the initial condition µ1 “ ´1, that

µn “ p´1qn qp 2q .
n

As a small application, we compute the number of spanning sets in Vn pqq. Let f pWq
be the number of subsets that span W, and let g pWq be the number of non-empty
subspaces of W. Evidently, we have:
ÿ
g pWq “ f pTq .
TďW

Using Möbius inversion, we obtain:

ÿ
f pWq “ µ pT, Wq g pTq ,
TďW

in particular
ÿ n „ 
ÿ ´ ¯
n
p´1qn´k qp q 2q k ´ 1 .
n´k
f pVn pqqq “ µ pT, Vn pqqq g pTq “ 2
k q
T k“0

Exercise 31: Prove the “NBC–Theorem” (“Non–broken circuit theorem”) of G.–C. Rota: Let L be
geometric lattice. We assume that the atoms Žof L are labelled (with natural numbers 1, 2, . . . ). A
set B of atoms is called independent, if rk p Bq “ |B|, otherwise it is called dependent. A set C
of atoms is called a circuit, if C is a minimal dependent set. A broken circuit is a set corresponding
to a circuit from which its largest atom (with respect to the labeling of atoms) was removed. A
non–broken circuit is a set B of atoms which does not contain a broken circuit. Then Rota’s Theorem
states: ´ ¯
` ˘ ł
µ 0̂, x “ p´1qrkp xq ¨ # non–broken circuits B with B“x .

Exercise 32: Show: The Möbius function µ px, yq of a semimodular lattice is alternating, i.e.

p´1qlength of r x,ys µ px, yq ě 0.

Moreover, show that the Möbius function of a geometric lattice is strictly alternating, i.e.
p´1qlength of r x,ys µ px, yq ą 0.

2Direct counting argument: There are rns “ qn ´1 subspaces of dimension 1 in V pqq.

q q ´1 n
A basis of a k-dimensional subspace W can be extended to a basis of Vn pqq by adding n ´ k
vectors; the span of these n ´ k vectors is isomorphic to Vn´k pqq. An arbitrary k-dimensional
subspace of Vn pqq can be written in rnsq ¨ rn ´ 1sq ¨ ¨ ¨ rn ´ k ` 1sq ways as an (ordered) direct sum
of 1-dimensional subspaces; this observation leads directly to the number of k–dimensional
“ ‰ rns ¨rn´1s ¨¨¨rn´k`1s
subspaces: nk q “ qrks ¨rk´q1s ¨¨¨r1s q .
q q q
72 3. POSETS

Hint: Use the following formula for the Möbius function of a lattice:
ÿ ` ˘
µ 0̂, x “ 0 for all a P L. (3.23)
x:x _ a“1̂
Show that if a is an atom, then there follows:
` ˘ ÿ ` ˘
µ 0̂, 1̂ “ ´ µ 0̂, x . (3.24)
x coatom
xğa
 CHAPTER 4

Asymptotic enumeration

So far, we used formal power series in an entirely algebraic way, retrieving

informations about their coefficients — either explicit formulae or recursions
— by elementary computations.
In this chapter we will see, that we can obtain even more information by ana-
lytical methods (a lot more about these methods can be found in [3].)
E XAMPLE 4.0.1. We already computed the exponential generating function of the
Bell–numbers Bn , which enumerate all (set–)partitions of an n–element set (2.2):
ÿ Bn ÿ 1 ÿ 1 ˆm ˙
B pzq “ z “e
n e z ´1
“ z m
pe ´ 1q “ p´1qm´j e jz ,
n! m! m! j
m,j

which immediately yields the formula:

ÿ 1 ˆm ˙
Bn “ p´1qm´j jn .
m! j
m,j

Now we may ask: What is the “order of magnitude” of the number Bn , for
n large? The analogous question “what is the order of magnitude of n!?” is
answered by Stirling’s Formula
? ´ n ¯n
n! „ 2πn for n Ñ 8. (4.1)
e
We want to find similar “asymptotic” results also in other cases.

4.1. Landau’s notation

D EFINITION 4.1.1. Let S be an arbitrary set and let f , g be real– or complex–valued
functions S Ñ R (or S Ñ C). Then the meaning of the O–notation “ f is a Big–O of
g on S” is:
f pzq “ O pg pzqq :ðñ D C ą 0 : @z P S : | f pzq| ď C ¨ |g pzq| . (4.2)
In many cases, the set of interest S is an “appropriately chosen” (but not necessarily
fixed) neighbourhood of some point ζ (ζ “ 8 is also possible): If, for instance, f , g are
actually sequences (i.e., functions N Ñ C), then
f pnq “ O pg pnqq for n Ñ 8
means that there is a neighbourhood S of 8 (i.e., S “ tn P N : n ě au for some a), for
which (4.2) holds.
Similarly, the meaning of the o–notation “ f is a Small–O of g for z Ñ ζ” is:
f pzq
f pzq “ o pg pzqq for z Ñ ζ :ðñ lim “ 0. (4.3)
zÑζ g pzq
73
74 4. ASYMPTOTIC ENUMERATION

Moreover, we shall use the following notation for asymptotic equivalence:

f pzq
f pzq „ g pzq for z Ñ ζ :ðñ lim “ 1. (4.4)
zÑζ g pzq
R EMARK 4.1.2. The O–notation and o–notation must not be misunderstood as “equa-
tion”: For instance, the meaning of
´ ¯
n3 ` 2n2 ´ 1 “ O n3

is “n3 ` 2n2 ´ 1 belongs to the class

ˇ ˇof all functions f : N Ñ C, for which there is
some C ą 0 such that | f pnq| ď C ˇn3 ˇ. In spoken language this misunderstanding is
avoided by saying “. . . is a Big–O of . . . ” and not “. . . is equal to Big–O of . . . ”. But
since O–notation and o–notation are commonly used, we stick to it.
R EMARK 4.1.3. The “„”–notation for asymptotic equivalence basically is superfluous,
since it can be replaced by the o–notation:
f pxq „ g pxq ðñ f pxq “ g pxq p1 ` o p1qq ðñ f pxq “ eop1q ¨ g pxq .
But, again, this notation is commonly used, so we stick to it.
E XAMPLE 4.1.4. O–Notation:
´ ¯
n3 ` 2n2 ´ 1 “ O n3 pn Ñ 8q
z2 “ O pzq pz Ñ 0q
´ ¯
z “ O z2 pz Ñ 8q
`? ˘
plog zq5 “ O z pz Ñ 8q
E XAMPLE 4.1.5. o–Notation:
´ ¯
n ` 2n ´ 1 “ o n4 pn Ñ 8q
3 2

z2 “ o pzq pz Ñ 0q
1
“ o p1q pn Ñ 8q
n ´ n ¯n
?
n! “ 2πn p1 ` o p1qq pn Ñ 8q
e
o p f pzq ¨ g pzqq “ f pzq ¨ o pg pzqq pz Ñ ζq
The meaning of the last “equation” is (cf. Remark 4.1.2): The class of functions φ, for
which
φ pzq
lim “0
zÑζ f pzq ¨ g pzq
holds, is contained in the class of functions f ¨ ψ, for which
ψ pzq
lim “0
zÑζ g pzq
φpzq
holds (simply set ψ pzq :“ f pzq ).
 4.1. LANDAU’S NOTATION 75

E XAMPLE 4.1.6. Asymptotic equivalence:

z ` 1 „ z pz Ñ 8q
1
sinh z „ ez pz Ñ 8q
2 ?
n! „ e´n nn 2πn pn Ñ 8q
z
π pzq „ pz Ñ 8q
log pzq
The idea behind these notations is, of course, to replace a complicated function by a
simpler one, which is asymptotically equivalent.

Exercise 33: Let f , g : N Ñ R` be two functions (from the natural numbers to the nonnegative
reals). Which of the following rules is valid for n Ñ 8 (under which preconditions)?
O p f pnqq ` O pg pnqq “ O p f pnq ` g pnqq O p f pnqq ´ O pg pnqq “ O p f pnq ´ g pnqq
O p f pnqq
O p f pnqq ¨ O pg pnqq “ O p f pnq ¨ g pnqq “ O p f pnq {g pnqq
O pg pnqq
´ ¯
O p f pnqqOpgpnqq “ O f pnq gpnq exp pO p f pnqqq “ O pexp p f pnqqq
b ˆb ˙
O p f pnqq “ O f pnq g pO p f pnqqq “ O pg f pnqq

e f pnq`Opgpnqq “ e f pnq p1 ` O pg pnqqq log p f pnq ` g pnqq “ log p f pnqq

` O pg pnq { f pnqq .
(The “equations” should be interpreted as follows: O p f pnqq ` O pg pnqq is the class of all functions
of the form f ‹ pnq ` g‹ pnq, where f ‹ pnq “ O p f pnqq and g‹ pnq “ O pg pnqq; the first “equation”
means, that this class is contained in the class O p f pnq ` g pnqq.)

Exercise 34: Same question as in the preceding exercise, where O p.q is replaced by o p.q.

Exercise 35: Let f 1 , f 2 , g1 , g2 be functions N Ñ C, such that f 1 pnq „ f 2 pnq and g1 pnq „ g2 pnq
for n Ñ 8. Which of the following rules are valid for n Ñ 8 (under which preconditions)?
f 1 pnq ` g1 pnq „ f 2 pnq ` g2 pnq f 1 pnq ´ g1 pnq „ f 2 pnq ´ g2 pnq
f 1 pnq f pnq
f 1 pnq ¨ g1 pnq „ f 2 pnq ¨ g2 pnq „ 2
g1 pnq g2 pnq
f 1 pnq g 1 pnq „ f 2 pnq g2 pnq exp p f 1 pnqq „ exp p f 2 pnqqq
b b
f 1 pnq „ f 2 pnq g1 p f 1 pnqq „ g2 p f 2 pnqq
log p f 1 pnqq „ log p f 2 pnqqq .

E XAMPLE 4.1.7. Here are some typical generating functions that we want to examine
in the following:
ř n
‚ Derangements: Dn zn! “ e´z ¨ 1´z 1
,
ř ?
n 1´ 1´4z
‚ Catalan numbers: Cn z “ 2z ,
ř zn e z ´1
‚ Bell–numbers: Bn n! “ e .
76 4. ASYMPTOTIC ENUMERATION

For the Derangements (i.e., the number of fixed–point–free permutations) we obtained

an exact formula by the principle of inclusion–exclusion (cf. the presentation preceding
Corollary 3.4.18), from which we may directly derive an asmyptotic formula:
ˆ ˙
1 1 1 n 1 n!
Dn “ n! ´ ` ´ ¨ ¨ ¨ ` p´1q „ .
0! 1! 2! n! e
For the Catalan numbers, we may easily derive an asmyptotic formula by Stirling’s
formula (4.1):
ˆ ˙2n ´ ¯
p2nq! 2n e 2n 1 1
Cn “ „ ? “ 4n ¨ ? 3{2 .
pn ` 1q n! 2 e n n πn πn

The asymptotics for the Catalan numbers shows the typical feature: If a gen-
erating function f has singularities, let ρ be a singularity of minimal absolute
value. Then the coefficients of f show the following asymptotic behaviour:
ˆ ˙n
n 1
Jz K f pzq „ ¨ Θ pnq
|ρ|
a
with limnÑ8 n Θ pnq “ 1; the type of the singularity determines Θ pnq. This
kind of asymptotic analysis is called singularity analysis (cf. Abschnitt 4.3). If
there are no (or no “simple”) singularities, then the saddle point method proves to
be useful in many cases (cf. Abschnitt 4.4).

4.2. Recapitulation: Elements of complex analysis

We will quickly recapitulate some basic concepts from complex analysis, which
we shall need in the following:
D EFINITION 4.2.1. A domain G is an open connected subset of the complex num-
bers C: G Ď C. A function f : G Ñ C is called analytic at z0 P G if there is a
neighbourhood U Ď G of z0 in which the series expansion
ÿ
f pzq “ cn pz ´ z0 qn
ně0

holds (for certain coefficients cn P C). A function f : G Ñ C is called analytic (on G)

if f is analytic for all z P G.

For a function f , which is analytic at z0 , we have:

‚ There is a positive radius of convergence R, such that the series con-
verges for every point in the interior of the disk
Dpz0 , Rq :“ tz P C : |z ´ z0 | ď Ru
with center z0 and radius R.
‚ f pzq is analytic in the interior tz P C : |z ´ z0 | ă Ru of this disk of conver-
gence.
‚ The series diverges for all z with |z ´ z0 | ą R.
 4.2. RECAPITULATION: ELEMENTS OF COMPLEX ANALYSIS 77

D EFINITION 4.2.2. A function f : G Ñ C is called (complex!) differentiable at z0 , if

the limit
f pz0 ` δq ´ f pz0 q
lim (δ is a complex number!)
δÑ0 δ
exists; f is called differentiable on G if f is differentiable for all z P G.
T HEOREM 4.2.3. A function f : G Ñ C is analytic on G if and only if f is differen-
tiable on G.
Let f and g be two analytic functions, let λ P C. Then f ` g, λ ¨ f and f pgq are also
f
analytic; and g is analytic at all points ζ with g pζq ‰ 0. 

D EFINITION 4.2.4. A function h : G Ñ C is called meromorphic at z0 if h pzq can

be written as a quotient of two analytic functions f , g in a neighbourhood U of z0 :
f pzq
@z P Uz tz0 u : g pzq ‰ 0 and h pzq “ .
g pzq
In this case we also have
ÿ
h pzq “ cn pz ´ z0 qn
ně´m
for all z ‰ z0 in the disk centered at z0 . If m is the largest natural number, for which
c´m ‰ 0 in this series expansion, then z0 is called a pole of order m. The coefficient
c´1 of pz ´ z0 q´1 in this series expansion is called the residue of h at the point z0 , we
also denote it as Res ph, z0 q:
r z
Res ph, z0 q “ pz ´ z0 q´1 h pzq .

D EFINITION 4.2.5. Let φ : r0, 1s Ñ G be a differentiable function which determines

a contour Γ in the domain G (with starting pointş φ p0q and end point φ p1q); let f :
G Ñ C be a function. Then the contour integral Γ f pzq d pzq is defined as
ż1
f pφ ptqq φ1 ptq dt.
0

The contour Γ is called closed if φ p0q “ φ p1q.

R EMARK 4.2.6. The contour integral only depends on the contour Γ, not on the con-
crete parametrization φ ptq.
T HEOREM 4.2.7 (Cauchy’s Theorem). Let f : G Ñ C be analytic on G, and let Γ be
an arbitray closed contour in G. Then we have
ż
f pzq dz “ 0.
Γ

T HEOREM 4.2.8 (Residue Theorem). Let h pzq be meromorphic on G, let Γ be a closed

contour in G. Then we have
ż ÿ
1
h pzq dz “ ns ¨ Res ph, sq ,
2π i Γ s
78 4. ASYMPTOTIC ENUMERATION

where the sum runs over all poles s of h, and ns denotes the winding number1 of Γ
with respect to s.
ř
C OROLLARY 4.2.9 (Cauchy’s integral formula). Let f pzq “ ně0 f n zn be analytic
in a disk centered at 0, and let Γ be a contour in the interior of this disk, which winds
around 0 exactly once (in positive orientation) (i.e., the winding number with respect
to 0 is 1), then we have
ż
n 1 f pzq
Jz K f pzq “ f n “ dz. (4.5)
2π i Γ zn`1
ř
D EFINITION 4.2.10. Let f pzq “ ně0 cn pz ´ z0 qn be an analytic function with ra-
dius of convergence R; let ζ be a point in the boundary
BDpz0 , Rq “ tz P C : |z ´ z0 | “ Ru
of the disk of convergence. If there is a neighbourhood U of ζ and an analytic function
fˆ on U, such that ˇ
ˆf ˇˇ “ f |UXDpz0 ,Rq ,
UXDpz0 ,Rq

then this fˆ is called an analytic continuation of f in ζ. If there is such analytic

continuation, then it is unique.
ζ is called a singularity of f , if there is no analytic continuation of f in ζ f .
We denote the set of singularities of f by Sing p f q. Sing p f q X BDpz0 , Rq is always
closed; the case
Sing p f q “ BDpz0 , Rq
is possible: Then BDpz0 , Rq is called the natural boundary of f .
E XAMPLE 4.2.11. The function
? 1 ÿ ˆ1˙
1 ` z “ p1 ` zq 2 “ 2 zk
k
kě0

has no analytic continuation at ζ “ ´1, so ´1 is a singularity of this function.

The functions 1 e´z both have a singularity (more precisely: a pole) at ζ “ 1.
and 1´z
1´z

Exercise 36: Let f , g be complex functions which are analytic on some given domain. Which of the
following rules are valid (under which preconditions)?
Sing p f ˘ gq Ď Sing p f q Y Sing pgq Sing p f ¨ gq Ď Sing p f q Y Sing pgq
Sing p f {gq Ď Sing p f q Y Sing pgq Y Null pgq Sing p f ˝ gq Ď Sing pgq Y gp´1q pSing p f qq
´a ¯
Sing f Ď Sing p f q Y Null p f q Sing plog f q Ď Sing p f q Y Null p f q
´ ¯ ` ` ˘˘
Sing f p´1q Ď f pSing p f qq Y f Null f 1

Here, Sing p f q denotes the set of singular points of f , and Null p f q denotes the set of zeroes of f .

1Loosely speaking, the winding number enumerates how often the contour Γ “winds
around” s: In the examples we consider here, the winding number will always be equal to
1.
 4.2. RECAPITULATION: ELEMENTS OF COMPLEX ANALYSIS 79
ř
T HEOREM 4.2.12. Let f pzq “ f n ¨ zn be analytic at 0 with radius of convergence
R ă 8. Then there exists a singularity ζ in the boundary of the disk of convergence
(i.e., |ζ| “ R).

P ROOF. Indirect: Suppose not; then there exists an analytic continuation of f at

all points with |z| “ R. Therefore, for all z with |z| “ R there
Ť is a neighbourhood
Uz , such that f has an analytic continuation on U. Now z:|z|“R Uz is an open
cover of BDp0, Rq: since the boundary of the disk is compact, there exists a finite
subcover. But this implies that f is analytic on Dp0, R1 q with R1 ą R, and for
some R0 with R ă R0 ă R1 we have, according to (4.5):
ż
n 1 f pzq
f n “ Jz K f pzq “ dz.
2π i |z|“R0 zn`1

We may parametrize the circle with center 0 and radius R0 as follows: z pθq “
R0 ¨ eiθ and dθ
d
z pθq “ iz pθq. Then we obtain for this contour:
ż 2π ` ˘
1 f R 0 ¨ ei θ
fn “ dθ.
2π 0 R0n ¨ en¨iθ

From this we get immediately the upper bound

1
| fn| ď ¨ max | f pzq| , (4.6)
R0n |z|“R0
` ˘ ř
i.e., f n “ O R´n 0 pn Ñ 8q. Let q with R ă q ă R 0 , then f pzq “ f n zn
2
converges for |z| ă q according to the well–known root test , a contradiction
(since q ą R). 
ř
T HEOREM 4.2.13 (Pringsheim). Let f pzq “ f n zn be analytic at 0 with radius of
convergence R. Assume that f n ě 0 for all n ě 0. Then the (real) number R is a
singularity of f .

P ROOF. Indirect: Suppose, f pzq is analytic at ζ “ R. Then there is a radius of

convergence r such that f pzq is analytic for all z with |z ´ R| ă r. Set h “ 3r and
z0 “ R ´ h. Then we have
ÿ
f pzq “ gm pz ´ z0 qm
m

for certain coefficients gm and all z with |z ´ z0 | ď 2h.

2From | f | C
a q ?
n
n ď R0n we get for |z| ă q ă R0 : n
| f n zn | ă R0 C, in particular
a q
lim supnÑ8 n | f n zn | ă 1 (since R0 ă 1).
80 4. ASYMPTOTIC ENUMERATION

2h r

z0

Equivalently, we have:
ÿ 2r
f pz0 ` zq “ gm ¨ zm for |z| ď 2h “
m
3
and therefore
ˆ ˙m
m 1 d
gm “ Jz K f pz ` z0 q “ f pz ` z0 q |z“0
m! dz
1 pmq
“ f pz0 q
m!
1 ÿ
“ f n ¨ n ¨ pn ´ 1q ¨ ¨ ¨ pn ´ m ` 1q z0n´m
m! n
ÿ ˆn˙
“ f n ¨ z0n´m .
něm
m
So we have ˜ ¸
ÿ ÿ ˆn˙
pz0 ` 2hq “
f looomooon f n ¨ z0n´m p2hqm .
m něm
m
“R`h
The right–hand side is a converging double sum (since R ` h lies in the inte-
rior of DpR, rq) and has only non–negative summands: Therefore it is absolutely
convergent, and we may reorder the summands arbitrarily.
ÿ ÿ n ˆ ˙
n n´m
f pR ` hq “ fn z0 p2hqm
n m“0
m
ÿ ÿ
“ f n pz0 ` 2hqn “ f n pR ` hqn ,
n n
a contradiction, since R ` h ą R. 

4.3. Singularity analysis

E XAMPLE 4.3.1. Most of the power series we are interested in are generating func-
tions stemming from some enumeration problem and thus cleary having non–negative
coefficients: So Pringsheim’s Theorem is applicable in these cases.
 4.3. SINGULARITY ANALYSIS 81

(1) For the Derangement–numbers Dn we know

ÿ Dn e´z
zn “
n! 1´z
and Dn!n „ e1 ; therefore the radius of convergence is 1 — and ζ “ 1 is, indeed,
a singularity.
(2) Let Sn be the number of surjections rns Ñ rks (k ď n arbitrary): From the
combinatoriel point of view it is clear that we encounter the species sequences psetn1 q
(cf. Example 2.6.6), therefore we have the generating function
ÿ Sn 1 1
zn “ “ .
n! 1 ´ pe ´ 1q
z 2 ´ ez
The singularities of this function obviously are log 2 ` 2π ik, k P Z; the only
real singularity is log 2, which is, indeed, the radius of convergence.
(3) The generating function for the Catalan numbers is
?
1 ´ 1 ´ 4z
2z
with singularity ζ “ 41 ; hence the radius of convergence is 41 , too.
(4) The generating function for the species cycles ist
1 z2 z3
log “ z` ` `¨¨¨
1´z 2 3
with singularity ζ “ 1; hence the radius of convergence is 1, too.

4.3.1. The Exponential Growth Formula.

D EFINITION 4.3.2. A singularity ζ of f of minimal absolute value is called domi-
nant singularity.
D EFINITION 4.3.3. We say that a sequence of numbers pan q8n“0 is of exponential
order K n and introduce the short notation an ’ K n :
a
an ’ K n :ðñ lim sup n |an | “ K, (4.7)
nÑ8
i.e., @ǫ ą 0 we have
|an | ą pK ´ ǫqn for infinitely many n, (4.8)
n
|an | ă pK ` ǫq for fast all n. (4.9)

Stated otherwise:
an “ K n ¨ Θ pnq ,
where Θ is subexponential, i.e.
a
lim sup n
|Θ pnq| “ 1.
nÑ8
Typical examples for subexponential sequences are
? 2 π
1, n3 , n´3{2 , log n, p´1qn , e n
, elog n , n2 sin n , . . .
2
82 4. ASYMPTOTIC ENUMERATION
ř
T HEOREM 4.3.4. Let f pzq “ n f n ¨ zn be analytic at 0, let R be the absolute value of
a dominant singularity. Then we have:
ˆ ˙n
1
fn ’ .
R

P ROOF. This follows immediately from řCauchy–Hadamard’s formula for the ra-
dius of convergence R of a power series n f n zn :
1
R“ a . (4.10)
lim supnÑ8 n
| fn|
(For the absolute value of a dominant singularity clearly equals the radius of
convergence.) 
Exercise 37: Let p pnq be the number of (integer) partitions of n. We know that
8
ÿ 8
ź 1
p pnq zn “ .
n “0 i “1
1 ´ zi

What are the (dominant) singular points of this generating function? What does this imply for the
asymptotic behaviour of p pnq for n Ñ 8?

We sum up these results:

P RINCIPLE 4.3.5 (First řprinciple of singularity analysis). The (absolute values of)
singularities of f pzq “ f n ¨ zn determine the exponential growth of the coefficients
fn.

As examples illustrating this principle, we already considered the Derangement–

numbers Dn , the number of surjections Sn and the Catalan numbers Cn .
D EFINITION 4.3.6. Let G be a domain, let z0 P G. Let f : Gzz0 Ñ C be an analytic
function. Then the point z0 is called an isolated singularity of f . There are precisely
three possibilities: The singularity z0 may be:
‚ a removable singularity, if there is an analytic continuation of f on the whole
of G3.
‚ a pole, if z0 is not a removable singularity is and there is some natural number
m, such that pz ´ z0 qm ¨ f pzq has a removable singularity at z0 . If m is the
minimal such number, then z0 is called a pole of m–th order.
‚ an essential singularity, if z0 is neither removable nor a pole.
E XAMPLE 4.3.7. The following functions Cz t0u Ñ C have an isolated singularity at
0:
‚ For sinz z , 0 is a removable singularity,
‚ for z1m , 0 is a pole of m–th order,
1
‚ for e z , 0 is an essential singularity.
3Accoriding to Riemann’s Theorem this is possible, if f is bounded in a neighbourhood of z .
0
Analytic continuations are always unique.
 4.3. SINGULARITY ANALYSIS 83

R EMARK 4.3.8. The logarithmus is a covering map which is not analytic on any
domain containing 0, so 0 is not an isolated singularity of the logarithm (and thus of
?
the root function n z “ z1{n “ exp n1 log z).
P ROPOSITION 4.3.9. Let f pzq be analytic on |z| ă R with 0 ď R ă 8. Then we have
for arbitrary r P p0, Rq:
max|z|“r | f pzq|
| fn| ď ,
rn
i.e.
max|z|“r | f pzq|
| f n | ď inf .
0ărăR rn
If f pzq only has non–negative coefficients, then we have
f prq f prq
fn ď n
ùñ f n ď inf .
r 0ărăR r n

P ROOF. The fundamental bound is derived from Cauchy’s integral formula

(4.5) (see the proof of Theorem 4.2.12): Setting z pθq “ r ¨ eiθ and dθ
d
z pθq “ iz pθq,
we obtain ` ˘
ż ż
1 f pzq 1 2π f r ¨ eiθ
fn “ dz “ dθ
2π i |z|“r zn`1 2π 0 r n en¨iθ

R EMARK 4.3.10. If f pzq has only non–negative coefficients, then the inf is assumed
at ˆ ˙
f prq 1 f 1 prq f prq f 1 prq
“ 0 ðñ ´ n n`1 “ 0 ðñ r “ n.
rn rn r f prq
We shall make use of this simple observation when we will deal with the saddle point
method.

Now we want to explain the following principle:

P RINCIPLE 4.3.11 (Second principle of singularity analysis). The nature of the
dominant singularities determines the subexponential part Θ pnq in
f n „ K n ¨ Θ pnq .

4.3.2. Rational functions. From the well–known partial fraction decomposi-

tion of a rational function f (expressed as the quotient of two polynomials p
and q, which do not have a common root)
ÿk ÿ e
p pzq p pzq i
γi,l
f pzq “ “ e1 ek “ R
lo
o pzq
moo n `
q pzq C pz ´ z1 q ¨ ¨ ¨ pz ´ zk q l
i“1 l“1 pz ´ zi q
polynomial

together with the binomial Theorem

´l ´l
ÿ ˆ´l ˙ ˆ z ˙n l
ÿ ˆ n ` l ´ 1˙
pz ´ zi q “ p´zi q ´ “ p´1q zn zi´n´l
ně0
n zi ně0
l´1
84 4. ASYMPTOTIC ENUMERATION

we immediately obtain the following exact formula for the coefficients f n “

Jzn K f pzq, n ą deg pRq:
ei
k ÿ
ÿ ˆ ˙
l n ` l ´ 1 ´n´l
fn “ p´1q γi,l zi
l´1
i“1 l
ÿk ÿei ˆ ˙
´n l n ` l ´ 1 ´l
“ zi p´1q γi,l zi
l´1
i“1 l
loooooooooooooooomoooooooooooooooon
polynomial in n

From this, we derive immediately the asymptotics:

T HEOREM 4.3.12. Let 1, . . . , am be the dominant singularities (poles) of the rational
function f pzq; let e1 , . . . , em be the corresponding orders of these poles. Then we have:
m
ÿ ` ´n ˘
fn “ a´n polynomial pnq
i loooooooomoooooooon
i `O b
i“1
of degree ei ´1

where b is a non–dominant singularity of minimal absolute value.

E XAMPLE 4.3.13. The generating function of the Fibonacci–numbers Fn is
ÿ 1
Fn ¨ zn “ .
ně0
1 ´ z ´ z2
?
The singularities of this rational function are ´1˘2 5 . The partial fraction decomposi-
tion is
? ?
1 1` 5 1 1´ 5 1
“ ? ? ´ ? ? ,
1 ´ z ´ z2 2 5 1 ´ 1` 5 z 2 5 1 ´ 1´ 5 z
2 2
from which we immediately obtain the asymptotics:
¨¨ ˛n`1 ˛
ˆ ? ˙n`1 ˚˚ 1 ´ ?5 ‹ ‹
1 1` 5 ˚ ‹ ‹
Fn “ ? ` O ˚˚˝ ‹.
5 2 2 ‚ ‚
˝ loomoon
»0.618034

Further simple examples:

1 ÿ
2
“ pn ` 1q 2n zn (one pole of second order),
p1 ´ 2zq ně0
1 ÿ 1` n n˘ n
“ p1q ` p´1q z (two dominant poles).
1 ´ z2 ně0 2

4.3.3. Meromorphic functions. Meromorphic functions can be treated in

much the same way as rational functions:
 4.3. SINGULARITY ANALYSIS 85

T HEOREM 4.3.14. Let f pzq be meromorphic for |z| ď R with poles at a1 , . . . , am ; let
e1 , . . . , em be the orders of these poles. Let f be analytic at 0 and for all z with |z| “ R.
Then we have:
ÿm
` ´n ˘
fn “ a´n
i polynomial i pnq `O R
loooooooomoooooooon
i“1
of degree ei ´1

P ROOF. The proof proceeds by “subtraction of singularities”.

Consider the pole a1 . In a neighbourhood of a1 we may expand f pzq int a Lau-
rent series:
ÿ
c´e1 pz ´ a1 q´e1 ` ¨ ¨ ¨ c´1 pz ´ a1 q´1 `
f pzq “ loooooooooooooooooooooomoooooooooooooooooooooon cn pz ´ a1 qn .
ně0
loooooooomoooooooon
“:S a pzq
1
“:Ha1 pzq

Here Sa1 pzq is the “singular part” at a1 , while Ha1 pzq is analytic at a1 . This
“subtraction of singular parts” can, of course, be repeated for all the remaining
poles, whence we obtain
˜ ¸
ÿm ÿm
f pzq “ Sai pzq ` f pzq ´ Sai pzq ,
i“1 i“1
loooomoooon looooooooooomooooooooooon
“:Spzq “:Hpzq

where S pzq is a rational function with poles a1 , . . . , am and H pzq is an analytic

function in the disk |z| ď R ` ǫ for some ǫ ą 0 (since f is analytic for all z with
|z| “ R, the “usual” compactness argument yields the claim). Since we have for
the coefficients hn of H pzq ` ˘
hn “ O R´n
(cf. (4.6)), the assertion follows. 
E XAMPLE 4.3.15. We already computed the exponential generating function of the
numbers Sn of all surjections rns Ñ rks (arbitrary k):
1
f pzq “ .
2 ´ ez
This function has poles at log 2 ` k ¨ 2π i, k P Z; by “subtracting” the singular part at
log 2 we obtain
ˆ ˙ ˜ ¸
1 1 1 1
f pzq “ ´ ` H pzq “ ` H pzq
2 z ´ log 2 2 log 2 1 ´ log1 2 z
z´log 2
(since limzÑlog 2 2´ez “ ´ 12 , by de L’Hospital’s rule), where H pzq is analytic for all
z with |z| ď 2π. Hence we get from Theorem 4.3.14
ˆ ˙n`1
1 1 ` ˘
fn “ ` O p2πq´n ,
2 log 2
which implies for the numbers Sn :
ˆ ˙n`1
n! 1 ` ˘
Sn “ ` O n! p2πq´n (4.11)
2 log 2
86 4. ASYMPTOTIC ENUMERATION

E XAMPLE 4.3.16. The (exponential) generating function of cycles‹ (i.e., sequences pcyclesq)
is
1
f pzq “ 1
.
1 ´ log 1´z
Obviously there is one pole at z “ 1 and another at
1 1 e´1 1
log “ 1 ðñ “ e ðñ z “ “ 1 ´ » 0.632121,
1´z 1´z e e
hence the dominant singularity is ζ “ 1 ´ e´1 . Again, by de L’Hospital’s rule, we get
z´ζ 1
lim 1
“ lim ´1 “ ζ ´ 1 “ ´e´1 ,
zÑζ 1 ´ log 1´z zÑζ
1´z

i.e.,
e´1
f pzq “ ´ ` ˘ ` H pzq .
z ´ 1 ´ e´1
H pzq is analytic in the (closed!) disk with radius R “ 1 ´ ǫ (ǫ ą 0), therefore we
obtain the asympotics
ˆ ˙n`1
1 1 ` ´n ˘ 1
fn “ ` O p1 ´ ǫq » 1.58198n`1 ` O p1.0001n q .
e 1´e ´1 e
E XAMPLE 4.3.17. The number Cn,k of surjections rns Ñ rks can be expressed in terms
of the Stirling–numbers of the second kind:
Cn,k “ k! ¨ Sn,k .
We want to investigate the following question: If we assume uniform probability
distribution on the set of all surjections rns Ñ rks (n fixed, k arbitray), what is the ex-
pected value of the cardinality of the surjection’s image? Our approach is to consider
the generating function in two variables
ÿ ÿ zn k ÿ k z 1
f pt, zq :“ Cn,k ¨ ¨t “ t pe ´ 1qk “ .
n! 1 ´ t pez ´ 1q
kě0 ně0 kě0

Then we have ˜ ¸
ˇ n
d f ˇˇ ÿ ÿ zn ez ´ 1
“ k ¨ C ¨ “ .
dt ˇt“1 ně0 n,k
n! p2 ´ e z q2
k“0
The poles (order: 2) of this function are log 2 ` k ¨ 2π i, k P Z, and the Laurent–series
expansion is
ez ´ 1 1 1
“ ` ` H pzq ,
p2 ´ ez q2
4 pz ´ log 2q 2 4 pz ´ log 2q
´ ¯
where H pzq “ ´ 487 1
` 48 pz ´ log 2q ` O pz ´ log 2q2 is analytic in the (closed!)
b
disk of radius R “ 4π 2 ` plog 2q2 ´ ǫ (ǫ ą 0; we may choose R » 6.3213). From
 4.3. SINGULARITY ANALYSIS 87

this we obtain immediately

ˆ ˙n
ˆ ˙n
ez ´ 1 n`1 1 1 1 ` ˘
n
Jz K “ ´ ` O R´n
p2 ´ e q
z 2
4 plog 2q2 log 2
4 log 2 log 2
ˆ ˙n`2
1 1 ` ˘
“ pn ` 1 ´ log 2q ` O R´n .
4 log 2
We must combine this result with (4.11). . .
ˆ ˙n`1
n 1 1 1 ` ´n ˘
Jz K “ ` O R ,
2 ´ ez 2 log 2
since the expected value we are interested is the quotient of coefficients:
´ ¯n`2 ` ´n ˘
1 1
4 pn ` 1 ´ log 2q log 2 ` O R n`1 1 ` ˘
´ ¯n`1 “ ´ ` O R´n .
1 1 2 log 2 2
2 log 2 ` O pR´n q
(Since log 2 „ 0.693147 ă 1; cf (4.16) in the following Lemma.)
The following Lemma states some “typical manipulations” concerning the O–
notation; in particular the transformation from the preceding example:
L EMMA 4.3.18. Let f pnq and h pnq be arbitratry sequences (i.e., functions N Ñ C).
Then we have
O p f pnqq ¨ h pnq “ O p f pnq ¨ h pnqq . (4.12)
If in particular h pnq “ O p1q, then we simply have
O p f pnqq ¨ h pnq “ O p f pnqq . (4.13)
Let r : N Ñ C with limnÑ8 r pnq “ 0 (i.e., r “ o p1q). Then we have:
1
“ 1 ` O pr pnqq for n Ñ 8. (4.14)
1 ` O pr pnqq
1
Moreover, let g : N Ñ C, such that gpnq “ O p1q. Then we have:
1 O pr pnqq
“ 1` for n Ñ 8. (4.15)
1` Oprpnqq g pnq
gpnq
f pnq
Finally, let f : N Ñ C with “ O p1q. Then we have
pgpnqq2
f pnq ` O pr pnqq f pnq
“ ` O pr pnqq (4.16)
g pnq ` O pr pnqq g pnq
P ROOF. (4.12): Let ζ pnq P O p f pnqq be some fixed (but arbitrarily chosen) ele-
ment from O p f pnqq. Then we have by definition
|ζ pnq ¨ h pnq| “ |ζ pnq| ¨ |h pnq|
ď K ¨ | f pnq| ¨ |h pnq|
“ K ¨ | f pnq ¨ h pnq| .
(4.13): Simply continue the above chain of inequalities with “ď K ¨ K1 ¨ | f pnq|”,
which holds by definition.
88 4. ASYMPTOTIC ENUMERATION

(4.14): By definition of the limit and of O, respectively, there is some C ą 0

and some n0 P N, such that for a fixed (but arbitrarily chosen) element ζ pnq P
O pr pnqq in the denominator of (4.14) there holds:
1
|ζ pnq| ď C ¨ |r pnq| ď for all n ě n0 .
2
But this implies that for all n ě n0
ˇ ˇ
ˇ 1 ˇ 1 1
ˇ ˇ
ˇ 1 ` ζ pnq ˇ “ |1 ` ζ pnq| ď 1 ´ |ζ pnq|
1
ď Ð1´C|rpnq|ď1´ζpnq
1 ´ C ¨ |r pnq|
C ¨ |r pnq|
“ 1` Ð1´C¨|rpnq|ě 12
1 ´ C ¨ |r pnq|
ď 1 ` p2 ¨ Cq ¨ |r pnq| ,
which proves (4.14).
(4.15): By assumption there is some ǫ ą 0, such that |g pnq| ą ǫ for n ą N. Let
ζ pnq be as above, but now choose n0 such that
ǫ
|ζ pnq| ď C ¨ |r pnq| ď for all n ě n0 .
2
Then we have |C ¨ |r pnq| {g pnq| ď 1{2 for all n ě max pn0 , Nq, and
ˇ ˇ
ˇ ˇ
ˇ 1 ˇ 1 1
ˇ ˇď ď
ˇ ζpnq ˇ |ζpnq| C¨|rpnq|
ˇ 1 ` gpnq ˇ 1 ´ |gpnq| 1 ´ |gpnq|
C ¨ |r pnq| { |g pnq|
“ 1` ÐC¨|rpnq|{|gpnq|ď 21
1 ´ C ¨ |r pnq| { |g pnq|
|r pnq|
ď 1 ` p2 ¨ Cq ¨ ,
|g pnq|
which proves (4.15).
(4.16): Finally, we have
f pnq ` O pr pnqq f pnq O pr pnqq
“ ` .
g pnq ` O pr pnqq g pnq ` O pr pnqq g pnq ` O pr pnqq
For the first term on the right–hand side we have
¨ ˛
f pnq f pnq ˝ 1 ‚
“
g pnq ` O pr pnqq g pnq 1 ` Oprpnqq
gpnq
ˆ ˙
f pnq O pr pnqq
“ 1` Ð by (4.15)
g pnq g pnq
f pnq f pnq O pr pnqq
“ `
g pnq pg pnqq2
f pnq
“ ` O pr pnqq Ð by assumption and (4.13).
g pnq
 4.3. SINGULARITY ANALYSIS 89

For the second term on the right–hand side we have according to (4.13)
O pr pnqq
“ O pr pnqq ,
g pnq ` O pr pnqq
1
since also |g pnq ` O pr pnqq| ą ǫ{2 for n “large enough”, so gpnq`Oprpnqq “ O p1q.
Of course we have O pr pnqq ` O pr pnqq “ O pr pnqq, so also (4.16) is proved. 
4.3.4. Several dominant singularities.
1
E XAMPLE 4.3.19 (Cancellation). Consider the sum of rational functions 1`z2
and
1
1´z3
:
1 1 2 ` z2 ´ z3 2 ´ z2 ` z3 ` z4 ` z8 ` z9 ´ z10
` “ “ .
1 ` z2 1 ´ z3 p1 ` z2 q p1 ´ z3 q 1 ´ z12
The coefficients of zn vanish for n ” 1, 5, 6, 7, 11 pmod 12q.
E XAMPLE 4.3.20 (Nonperiodic fluctuations). Consider the rational function
1 6 11 84 3 779 4 2574 5
6
“ 1 ` z ` z2 ´ z ´ z ´ z `¨¨¨
1´ 5z ` z
2 5 25 125 625 3125
with partial fraction decomposition
1 3i 1 3i
1 2` 8 ¯ 2´ 8 ¯
“ ´ ` ´ .
1 ´ 56 z ` z2 1 ´ 53 ´ 45i z 1 ´ 53 ` 45i z
The roots of the denominators have absolute value 1 and are inverses of one another, i.e.
3 4i 3 4i
` “ eiθ and ´ “ e´iθ .
5 5 5 5
So the coefficient of zn may be written as
einθ ` e´inθ 3 einθ ´ e´inθ 3
` ¨ “ cos pn ¨ θq ` sin pn ¨ θq
2 4 2i 4
cos θ
“ cos pn ¨ θq ` sin pn ¨ θq
sin θ
sin ppn ` 1q ¨ θq
“ .
sin θ
4.3.5. “Standardized” singularity analysis. We shall now present the “(quasi–
) mechanical algorithm” of (standard) singularity analysis for determining the
asymptotics of the coefficients of some generating function f pzq, as developped
by Philippe Flajolet and Andrew Odlyzko. It consists of the following steps:
Normalization of the singularity: If ρ is the “relevant singularity” of f pzq,
consider fˆ pzq :“ f pρ ¨ zq: then 1 is the “relevant singularity” of fˆ pzq.
Asymptotics for “standard functions”: For “standard functions” of the
form ˆ ˙β
´α 1 1
p1 ´ zq log
z 1´z
there is a general result describing the asymptotics of the coefficients.
90 4. ASYMPTOTIC ENUMERATION

Transfer–Theorems: If we know some asymptotic bound for a function,

we may immediately derive an asymptotic bound for its coefficients:
˜ ˆ ˙β¸
1 ´ ¯
f pzq “ O p1 ´ zq´α log ùñ Jzn K f pzq “ O nα´1 plog nq β
1´z

4.3.5.1. The Gamma–function.

D EFINITION 4.3.21 (Gamma–function). The Gamma–function Γ psq is defined by
Euler’s Integral Representation
ż8
Γ psq “ e´t ¨ ts´1 dt. (4.17)
0

for ℜ psq ą 0. The Gamma–function fulfils the functional equation

Γ ps ` 1q “ s ¨ Γ psq . (4.18)
It follows that Γ psq has an analytic continuation to Cz t0, ´1, ´2, . . . u: We have for
all m P N0
p´1qm 1
Γ psq „ for s Ñ ´m. (4.19)
m! s ` m
For the Gamma–function we also have Euler’s Reflection Formula:
π
Γ psq Γ p1 ´ sq “ . (4.20)
sin pπsq

From the functional equation (4.18) we immediately obtain

Γ ps ` nq “ sn ¨ Γ psq “ s ¨ ps ` 1q ¨ ¨ ¨ ps ` n ´ 1q ¨ Γ psq . (4.21)

In particular, from (4.17) we see immediately Γ p1q “ 1, so we have for n ě 0 P Z
Γ pn ` 1q “ n!
´ ¯ ?
1
and likewise with Γ 2 “ π
ˆ ˙ ˆ ˙n ?
1 1 ? p2nq! ¨ π
Γ n` “ ¨ π“ . (4.22)
2 2 n! ¨ 4n
For the Gamma–function, Stirling’s formula gives an asymptotic approximation:
ˆ ˙
? 1 1 139
Γ ps ` 1q „ s e
s ´s
2πs 1 ` ` ´ `¨¨¨ . (4.23)
12s 288s2 51840s3
R EMARK 4.3.22 (Ψ–function). The Ψ–function is defined as the logarithmic deriva-
tive of the Gamma–function:
d Γ1 pzq
Ψ pzq “ log Γ pzq “ .
dz Γ pzq
 4.3. SINGULARITY ANALYSIS 91

p
The following formula gives evaluations of the Ψ–function at rational numbers z “ q ;
cf. [2, section 1.7.3, equation (29)]:
ˆ ˙ ˆ ˙
p π πp
Ψ “ ´γ ´ log q ´ cot
q 2 q
q
21
ÿ ˆ ˙ ˆ ˆ ˙˙
pn n
` cos 2π log 2 ´ 2 cos 2π , (4.24)
q q
n“1
q
where the prime ( 2 1)
in the upper bound of the summation means that for evev q the
q
last summand (with index n “ 2 ) is multiplied by 21 . (Here, γ denotes the Euler–
Mascheroni constant.) For example:
ˆ ˙ ´π ¯ 1
1 π
Ψ “ ´γ ´ log 2 ´ cot ` cos π log p2 ´ 2 cos pπqq
2 2 2 2
“ ´γ ´ 2 log 2, (4.25)
`π˘
since cot 2 “ 0, cos π “ ´1 and log 4 “ 2 log 2.
R EMARK 4.3.23. For the following Theorem it is useful is to recall the polar coordi-
nates for complex numbers: Let z P C with ℜ pzq “ x and ℑ pzq “ y, i.e., z “ x ` iy.
Then we may express z in polar coordinates pr, φq, i.e. z “ r ¨ pcos φ ` i sin φq with
r “ |z| and φ “ φ0 ` 2πk for k P Z, where φ0 denotes the main value (in the interval
´π ă arg pzq ď π)of the argument arg pzq. (for r “ 0 we may choose the angle φ
arbitrarily.) We have:
ex`iy “ ex ¨ pcos y ` i sin yq , (4.26)
log px ` iyq “ log |x ` iy| ` i arg px ` iyq . (4.27)
In particular, this means that the (complex) logarithmus (the inverse function of the ex-
ponential function) is a multi–valued function, which might be “forced” to be single–
valued by constraining its range (for instance, by considering only the main values of
the argument).
Moreover, we have the following connection of the trigonometricn functions with the
exponential function:
eiθ ` e´iθ
cos θ “ ,
2
eiθ ´ e´iθ
sin θ “ .
2i
T HEOREM 4.3.24 (Presentation with Hankel–contour integral). For the Gamma–
function we have ż
1 1
“ p´tq´s ¨ e´t dt, (4.28)
Γ psq 2π i H
where H “ H´ ` H˝ ` H` is a contour which cosists of the following parts
‚ H´ “ tw ´ i : w ě 0u, (
‚ H˝ “ ´eiφ : ´ π2 ď φ ď π2 ,
‚ H` “ tw ` i : w ě 0u
See the following graphical illustration:
92 4. ASYMPTOTIC ENUMERATION

P ROOF. Let D be a contour, starting at a point ρ ą 0 on the real axis, winding

once around zero in the anti–clockwise orientation and returning zu ρ.

D
ρ

For z P CzN with ℜ pzq ą 0, consider the integral

ż
p´tqz´1 e´t dt. (4.29)
D
The factor
p´tqz´1 “ epz´1q¨logp´tq “ epz´1q¨logp|´t|q`argp´tq
in the integrand is unique if we require log p´tq P R for t ă 0: I.e., for all points
t P D we have ´π ď arg pzq ď π.
The integrand is not analytic in the interior of D (since the logarithm is analytic
only in the slitted complex plane, i.e., in no neighbourhood of 0), but according to
Cauchy’s Theorem 4.2.7 we may “deform” the contour D to another contour D1
without changing the value of the integral:

D1 ρ

In the first (straight) part of this new contour D1 we have arg p´tq “ ´π, i.e.,
p´tqz´1 “ epz´1qplog t´iπq ,
and in the last (straight) part we have arg p´tq “ π, i.e.,
p´tqz´1 “ epz´1qplog t`iπq .
(For both parts we have log t P R.) Parametrizing the small circle as ´t pθq “
δeiθ we obtain:
ż żδ żρ
´iπpz´1q z´1 ´t
p´tq z´1 ´t
e dt “ e t e ` eiπpz´1q tz´1 e´t
D ρ δ
żπ ´ ¯z´1 iθ
`i δ ei θ eδe δeiθ dθ “
´π
żρ żπ
iθ
´ 2i sin pπzq t e dt ` iδ
z´1 ´t z
eizθ`δe dθ.
δ ´π

(Since sin pπ pz ´ 1qq “ sin pπz ´ πq “ ´ sin pπzq.)

 4.3. SINGULARITY ANALYSIS 93

This equality holds for all positive real numbers δ ď ρ: for δ Ñ 0 we have
şπ iθ şπ
δz Ñ 0 (since ℜ pzq ą 0) and ´π eizθ`δe dθ Ñ ´π eizθ dθ (since the integrand
converges uniformly); so we obtain:
ż żρ
p´tq e dt “ ´2i sin pπzq tz´1 e´t dt.
z´1 ´t
D 0
This holds for all positive real numbers ρ: H is obviously the “limiting case of
the contour” D for ρ Ñ 8, so there holds
ż ż8
z´1 ´t
p´tq e dt “ ´2i sin pπzq tz´1 e´t dt,
H 0
i.e., ż
1
Γ pzq “ ´ p´tqz´1 e´t dt. (4.30)
2i sin pπzq H
The contour H is not close to 0, therefore we do not need the assumption ℜ pzq ą
0 any more4: The integral in (4.30) is a (single–valued) analytic function, which
equals Γ pzq for ℜ pzq ą 0; so Hankel’s formula (4.30) for the Gamma–function
holds for all z P CzZ. Replacing z by 1 ´ z and applying Euler’s Reflection
Formula (4.20), we obtain the assertion. 
4.3.5.2. Asymptotics for “standard functions”. According to the binomial the-
orem, for the “standard function” p1 ´ zq´α (where α P C is arbitrary) we obtain
ˆ ˙ ˆ ˙
n ´α n ´α n`α´1
Jz K p1 ´ zq “ p´1q “ .
n n
If α P t0, ´1, ´2, . . . u, then the asymptotics of these coefficients is trivial (since
in that case p1 ´ zq´α is a polynomial, and the coefficients are all 0 for n ą ´α).
Otherwise, we can write
ˆ ˙
n`α´1 α ¨ pα ` 1q ¨ ¨ ¨ pα ` n ´ 1q Γ pn ` αq
“ “ .
n n! Γ pαq Γ pn ` 1q
Hence, we could determine the asymptotics by means of Stirling’s formula for
the gamma function (4.23)
ˆ ˆ ˙˙
? 1
Γ ps ` 1q “ s e
s ´s
2πs 1 ` O .
s
However, it is more convenient to apply Cauchy’s integral formula (4.5):
T HEOREM 4.3.25 (Asymptotics for coefficients of standard functions, 1). Let α P
Cz t0, ´1, ´2, . . . u. We consider the “standard function”
f pzq “ p1 ´ zq´α .
Then, asymptotically as n Ñ 8, we have
˜ ¸
ÿ8
nα´1 ek pαq
Jzn K f pzq „ 1` , (4.31)
Γ pαq nk
k“1

4An analytic continuation for the logarithm exists only in the slitted complex plane!
94 4. ASYMPTOTIC ENUMERATION

where ek pαq is a polynomial of degree 2k in α, which is divisible by αk`1 “ α ¨

pα ´ 1q ¨ ¨ ¨ pα ´ kq. The first terms read:
˜
n α´1 α ¨ pα ´ 1q α ¨ pα ´ 1q ¨ pα ´ 2q ¨ p3α ´ 1q
Jzn K p1 ´ zq´α „ 1` `
Γ pαq 2n 24n2
ˆ ˙¸
α2 ¨ pα ´ 1q2 ¨ pα ´ 2q ¨ pα ´ 3q 1
` 3
`O . (4.32)
48n n4

P ROOF. Due to Cauchy’s integral formula (4.5), we have

ż
n ´α 1 p1 ´ zq´α
Jz K p1 ´ zq “ dz, (4.33)
2iπ C zn`1
where C is an arbitrary closed contour (with winding number 1) around 0, which
avoids the ray of real numbers ě 1. Concretely, we choose the contour C R,n that
is described in the following figure:

Kurve CR,n
R

1{n
0 1 1{n

For all n ą ℜ p´αq, the integral becomes arbitrarily small along the “circle
piece” of C R,n if we let R Ñ 85, so that “only” the “Hankel-like piece” Hn “
H´ ˝ `
n ` H n ` H n of C R,n plays a role, which consists of the subpieces
i (
‚ H´n “!w ´ n : w ě 1 , )
iφ
‚ H˝n “ 1 ´ en : ´ π2 ď φ ď π2 ,
‚ H` “ w ` i : w ě 1(.
n n
Performing the substitution,
t 1
z “ 1 ` , dz “ dt
n n
we obtain ż ˆ ˙
n nα´1
´α ´α t ´n´1
Jz K p1 ´ zq “ p´tq 1` dt, (4.34)
2i π H n
where H is the Hankel contour from Theorem 4.3.24.
ˇ ˇ
5The reason is that the modulus of the integrand can be estimated from above as ˇˇ p1´zq ˇ
´α
ˇď
ˇ z n`1 ˇ
Rℜp´αq´n´1 , and hence the integral is less than or equal to 2πRℜp´αq´n .
 4.3. SINGULARITY ANALYSIS 95
` t
˘
For |t| ă n the power series expansion of log 1 ` n converges, hence we ob-
tain
ˆ ˙
t ´n´1
1` “ e´pn`1q logp1`t{nq
n
ˆ ˆ ˙˙
t t2 t3
“ exp ´ pn ` 1q ´ ` ´`¨¨¨
n 2n2 3n3
ˆ ˙
t2 ´ 2t 3t2 ´ 2t3
“ exp ´t ` ` ` ¨ ¨ ¨ Ðsort n´k
2n 6n2
ˆ 2 ˙
t ´ 2t 3t2 ´ 2t3
“ e ¨ exp
´t
` `¨¨¨
2n 6n2
ˆ ˙
t2 ´ 2t 3t4 ´ 20t3 ` 24t2
“e ´t
1` ` `¨¨¨ . (4.35)
2n 24n2
That is, for |t| ď R (R ă 8 arbitrary but fixed) the integrand of (4.34) converges
for n Ñ 8 uniformly to p´tq´α e´t . “Asymptotically”, we therefore have
ˆ ˙ ˆ ˆ ˙˙
t ´n´1 1
1` “e ´t
1`O for n Ñ 8.
n n
“Formally”, by substitution in (4.34), we obtain
ż ˆ ˆ ˙˙
´α nα´1 ´α ´t 1
n
Jz K p1 ´ zq “ p´tq e 1`O dt
2i π H n
ˆ ˆ ˙˙
nα´1 1
“ 1`O .
Γ pαq n
That was admittedly somewhat bold, and it needs justification. To this end, we
split the Hankel contour H
‚ into the part with ℜ ptq ě log2 n: we shall show that this part is “negli-
gible”;
‚ and into the part with ℜ ptq ă log2 n: inbthat case, for n large enough
2
log n ?
we have |t|n „ n ă 1 (because |t| ď ℜ ptq 2
` 1 and log z ă z for
all z ą 0) and the expansion (4.35) “works”.
To see that the first part is negligible, some finer work is required for the esti-
mation of the integral. The subparts of the integrals over H` and H´ can be
dealt with “in once” because, to start with, we have
ˇż ˇ ż
ˇ 8 p´t ˘ i q α ˇ 8
|t ` i|α
ˇ ˇ
ˇ dt ˇ ď ˇ ˇn`1 dt.
ˇ log2 n `1 ` t˘i ˘n`1 ˇ log2 n ˇ1 ` t`i ˇ
n n
From
zα “ eα log z “ epℜpαq`iℑpαqqplog|z|`i arg zq
“ eℜpαq log|z|´ℑpαq arg z ¨ eipℑpαq log|z|`ℜpαq arg zq
it follows that
|zα | “ eℜpαq log|z|´ℑpαq arg z ,
96 4. ASYMPTOTIC ENUMERATION

thus here concretely

ˇ ˇ
ˇpt ` iqα ˇ “ |t ` i|ℜpαq ¨ e´ℑpαq argpt`iq ď C ¨ |t|ℜpαq
since arg pt ` iq Ñ 0 for t Ñ 8 and
? d
t `1
2 t2 ` 1
lim “ lim “ 1.
tÑ8 t tÑ8 t2

Furthermore, we have ˇ ˇ ˆ ˙
ˇ1 ` t ` i ˇ ě 1 ` t ,
ˇ ˇ
ˇ n ˇ n
hence we may continue the estimation of the integral (we put α1 “ ℜ pαq):
ˇż ˇ ż8
ˇ 8 p´t ˘ iqα ˇ tα
1
ˇ ˇ
ˇ dtˇ ď C ¨ ˇ ˇn`1 dt.
ˇ log2 n `1 ` t˘i ˘n`1 ˇ log2 n ˇ1 ` t ˇ
n n

Here, the integration variable t ě log2 n is real, which we assume always in

what follows. Then we have
ˇ ˇ 2
ˇ ˇ
ˇ1 ` ˇ ě 1 ` log n ,
t
ˇ nˇ n
from which we may infer
ˇ ˇ ˜ ¸´n
ˇ
ˇ1 ` t ˇˇ´n log2 n
ˇ ď 1`
nˇ n
for all n P N.
From real analysis, we know (see for example [5, Satz 26.1]): for a sequence
pxn qn with limnÑ8 xn “ 0, and with xn ‰ 0 and xn ą ´1 for all n, we have
lim p1 ` xn q1{xn “ e.
nÑ8
log2 n
We choose the special sequence6 n and obtain
˜ ¸n{ log2 n
2
log n
lim 1` “ e,
nÑ8 n
or, equivalently (the real logarithm is continuous):
´ ¯
log2 n
n log 1 ` n
lim “1
nÑ8 log2 n
This implies that, at least, for ǫ ą 0 we have
˜ ¸
log2 n
n log 1 ` ě p1 ´ ǫq log2 n, (4.36)
n
if n is large enough.
6Apply de l’Hospital’s rule twice: lim log2 n 2 log n
nÑ8 n “ limnÑ8 n “ limnÑ8 n2 “ 0.
 4.3. SINGULARITY ANALYSIS 97

Now we claim: for fixed k, we may choose ann such that for t ě log2 n we have
1
tα 1
` ˘ ď . (4.37)
t n`1 n k t2
1` n
For, if we take the logarithm on both sides of this claim, then we obtain the
equivalent inequality
ˆ ˙
` 1 ˘ t
α ` 2 log t ´ pn ` 1q log 1 ` ď ´k log n (4.38)
n
(since the real logarithm is a monotone increasing function), and if we substi-
tute for t Ñ log2 n, then it is obvious that the following inequality holds for n
large: ˜ ¸
` 1 ˘ 2
log n
α ` 2 log log2 n ´ pn ` 1q log 1 ` ď ´k log n
n
looooooooooooooomooooooooooooooon
ěp1´ǫq log2 n by (4.36)

(since log2 n grows faster than log log2 n and log n). If we consider the deriva-
tive of (4.38), then we see that
2 ` α1 n ` 1 1 2 ` α1 n ` 1
´ “ ´ ď 0,
t n 1 ` t{n t n`t
np2`α1 q
for t ě n´α1 ´1 , and since the right-hand side of this inequality converges to
np2`α1 q
2 ` α1 , for n large enough we have also log2 n ě n´α1 ´1 . Thus, the inequality
(4.37) is established, and we may now complete our estimation:
ˇż ˇ ż8
ˇ 8 p´t ˘ i q α ˇ 1 C
ˇ ˇ
ˇ ` ˘ dt ˇ ď C ¨ dt “ .
ˇ log2 n 1 ` t˘i n`1 ˇ k 2
log2 n n t nk log2 n
n
By (4.35), for large n the integral (4.34) equals a sum of terms of the form
ż
ck p´tq´α e´t tk dt,
H :ℜptqďlog2 n
“except for the above error term”. We would of course like to argue that these
integrals “essentially” are equal to
ż
2π i 2π i ¨ pα ´ kq ¨ pα ´ k ` 1q ¨ ¨ ¨ pα ´ 1q
p´tq´α`k e´t dt “ “
H Γ pα ´ kq Γ pαq
(by Hankel’s formula (4.28) and the functional equation (4.21), respectively). In
order to see this, we argue that the “cut-off remaining integral” (using the same
manipulations as in the proof of Theorem 4.3.24) leads to the real integral
ż8
2i sin pπαq e´t p´tq´α dt,
log2 n
whose integrand clearly becomes very small if t becomes large:
ˇ ´t ˇ
ˇe p´tq´α ˇ ď e´t tℜpαq ď 1
n k t2
98 4. ASYMPTOTIC ENUMERATION

(since limtÑ8 e´t tℜpαq`2 “ 0). Hence, we can bound the error as before.
If we put everything together, we arrive at the claim. 
E XAMPLE 4.3.26 (Catalan numbers). Applying (4.32) to the generating function of
the Catalan numbers, ?
ÿ 1´ 1´4¨z
n
Cn´1 ¨ z “ ,
2
ně1
we have α “ ´ 12 . Setting y “ 4 ¨ z, we obtain:
4n n
Cn´1 “ ´ Jy K p1 ´ yq´α
2
´3{2
ˆ ´ ¯˙
n´1 n? 3 25 105 ´4
»4 ¨ ¨ 1` ` ` `O n .
π 8n 128n2 1024n3
R EMARK 4.3.27. For the coefficients ek in (4.31) we have
2k
ÿ
ek “ λk,l pα ´ 1q ¨ pα ´ 2q ¨ ¨ ¨ pα ´ lq
l“k
q y
with λk,l “ vk t l et ¨ p1 ` vtq´1´1{v .
T HEOREM 4.3.28 (Asymptotics for coefficients of standard functions, 2). Let α P
Cz t0, ´1, ´2, . . . u. We consider the “standard function”
ˆ ˙β
´α 1 1
f pzq “ p1 ´ zq log . (4.39)
z 1´z
Then, asymptotically as n Ñ 8, we have
˜ ¸
8
ÿ
nα´1 Ck pαq
Jzn K f pzq „ plog nq β 1` , (4.40)
Γ pαq logk n
k“1
` β˘ ˇ
dk 1 ˇ
where Ck “ k
Γ pαq ˇ .
dsk Γpsq s“α

R EMARK 4.3.29. The factor 1z in (4.39) is “just” there, in order that f pzq is a power
series, also in the case where β R Z (the logarithm is analytic in a neighbourhood of
z “ 1):
ˆ ˙β ˆ ˆ ˙˙
1 1 1 1
log “ exp β log log
z 1´z z 1´z
ˆ ˆ ˙˙
z z2 z3
“ exp β log 1 ` ` ` ` ¨ ¨ ¨
2 3 4
ˆ ˆ ˙˙
z 5z2 z3
“ exp β ` ` `¨¨¨
2 24 8
R EMARK 4.3.30. The formulae in Theorems 4.3.25 and 4.3.28 actually remain valid
also in the case α P t0, ´1, ´2, . . . u if we interpret them as limit cases, i.e., if we set
1 1
“ lim “ 0.
Γ pαq zÑα Γ pzq
 4.3. SINGULARITY ANALYSIS 99

In (4.40), the limit should be taken after cancelling the Gamma–factors Γ pαq, so only
the first term vanishes: For instance, we have
˜ ¸
z 1 2 ¨ γ 1
Jzn K “´ ` `O .
log p1 ´ zq´1 n ¨ log2 n n ¨ log3 n n ¨ log4 n
S KETCH OF PROOF. In analogy to the proof of Theorem 4.3.25, we choose a
“Hankel-type” contour in Cauchy’s integral formula. After the substitution
z Ñ 1 ` nt , we obtain an integral over the “ordinary” Hankel contour H with
integrand
ˆ ˙ˆ ˙ ˆ ˙ ´ n ¯¯β ˆ ˙
1 t t ´n´1 1 t ´α ´ t ´β´n´1
f 1` 1` “ ´ log ´ 1`
n n n n n t n
¨ ˛
logp´tq β
1 ´ log n
„ nα´1 p´tq´α plog nq β ˝ ‚ e´t .
1 ´ ´tn

For |t| ď log n, we have

ˆ ˙ ˆ ˙ ˆ ˙
t ´β ÿ ´β tk log n
1` “ “ 1`O .
n k nk n
kě0
1
Hence, these terms are negligible in our -expansion. Consequently, we
logk n
may simplify further:
ˆ ˙ˆ ˙ ˆ ˙
1 t t ´n´1 ´α log p´tq β
f 1` 1` „n α´1
plog nq e p´tq
β ´t
1´ ,
n n n log n
and expand the part which is independent of t by the binomial theorem:
˜ ˆ ˙2 ¸
log p´tq β pβ ´ 1q log p´tq
e´t p´tq´α 1 ´ β ` ´ ¨ ¨ ¨ ..
log n 2! log n
Analogously to the proof of Theorem 4.3.25, we must show that this expansion
is correct “up to an asymptotically negligible rest”; this allows us to integrate
termwise, where we interchange differentiation
d d
p´tq´s “ exp p´s log p´tqq “ ´ log p´tq exp p´s log p´tqq
ds ds
and integration:
ż k
ˆ ż ˙
1 ´s ´t k d 1 ´s ´t
p´tq e log p´tq dt “ p´1q
k
p´tq e dt
2π i H dsk 2π i H
dk 1
“ p´1qk . Ðby (4.28)
dsk Γ psq
This establishes the claim. 
E XAMPLE 4.3.31. A typical application of Theorem 4.3.28 is (α “ 21 , β “ ´1):
˜ ¸
1 1 1 γ ` 2 log 2 1
Jzn K ? “? 1´ `O ,
1 ´ z 1z log 1´z
1 nπ log n log n log2 n
100 4. ASYMPTOTIC ENUMERATION
´ ¯ ? ´ ¯1 ´ ¯
since Γ 1
2 “ π, Γ pzq´1 “ ´ Ψpzq
Γpzq and Ψ 1
2 “ ´γ ´ 2 log 2 (see (4.25)).

4.3.5.3. Transfer theorems.

π
D EFINITION 4.3.32. For two real numbers φ, R with R ą 1 and 0 ă φ ă 2, we define
the open domain
∆ pφ, Rq :“ tz P C : |z| ă R, z ‰ 1, |arg pz ´ 1q| ą φu .
We call such a domain a ∆-domain.7 The left part in the following figure illustrates
this.

R r γ2
φ γ1 1 θ
n
0 1 0 1
γ4
γ3

T HEOREM 4.3.33. Let α, β P R (not C!) be arbitrary, and let f pzq be a function that
is analytic on a ∆-domain. If f in the intersection of a neighbourhood of 1 and the
∆-domain satisfies
‚ the property that, if
˜ ˆ ˙β ¸
´α 1
f pzq “ O p1 ´ zq log ,
1´z

then ´ ¯
n α´1 β
Jz K f pzq “ O n plog nq ; (4.41)
‚ the property that, if
˜ ˆ ˙β ¸
´α 1
f pzq “ o p1 ´ zq log ,
1´z

then ´ ¯
Jzn K f pzq “ o nα´1 plog nq β . (4.42)

P ROOF. As usual, we use Cauchy’s integral formula

ż
n 1 f pzq
f n “ Jz K f pzq “ dz
2π i γ zn`1

7Philippe Flajolet, in line with his cultural background, used to call these domains
“Camembert domains”.
 4.3. SINGULARITY ANALYSIS 101

for a contour γ in the ∆-domain, which encircles the origin once in positive
direction. Concretely, we choose (see the right part of the above figure) the
contour γ “ γ1 Y γ2 Y γ3 Y γ4 which consists of the following parts:
" *
1
γ1 “ z P C : |z ´ 1| “ , |arg pz ´ 1q| ě θ ,
n
" *
1
γ2 “ z P C : ď |z ´ 1| , |z| ď r, |arg pz ´ 1q| “ θ ,
n
γ3 “ tz P C : |z| “ r, |arg pz ´ 1q| ě θu ,
" *
1
γ4 “ z P C : ď |z ´ 1| , |z| ď r, |arg pz ´ 1q| “ ´θ ;
n
where 1 ă r ă R and φ ă θ ă π2 . We treat each of the individual corresponding
parts of the integral separately:
ż
j 1 f pzq
fn “ dz,
2π i γ j zn`1
j “ 1, 2, 3, 4. Then we have:
ˆ ˙ ˜ˆ ˙ ¸
1 1 ´α ´ ´ ¯¯
f n1 “ O ¨O plog nq β “ O nα´1 log n β ,
n n
´ ¯
since f “ O nα plog nq β along γ1 by assumption, the length of the integration
´ ¯
path is O n1 , and z´n´1 “ O p1q along γ1 .
Furthermore, for r1 :“ n ¨ r we have
ˇ ˇ ż r1 ˇ ˇ´α ˇ ˇ β ˇˇ iθ t ˇˇ´n´1
ˇ 2ˇ 1 ˇtˇ ˇ n
ˇ ˇ ˇlog ˇ ˇ1 `ˇ e ˇ
ˇ fn ˇ ď dt Ðlog nt ďlog n
2πn 1 ˇ n ˇ t ˇ n ˇ
ż8 ˇ ˇ´n´1
1 α´1 β ´α ˇ
ˇ eiθ t ˇˇ ˇ ˇ
ˇ eiθ t ˇ
´ iθ ¯
e t
ď n plog nq t ˇ1 ` ˇ dt Ð ˇ1` n ˇě1` ℜ n
2π 1 n
ż8 ˇ ˇ
1 α´1 β ´α ˇ
ˇ t cos θ ˇˇ´n´1
ď n plog nq t ˇ1 ` dt.
2π 1 n ˇ
Since
ż8 ˇ ˇ ż8
ˇ t cos θ ˇˇ´n´1
lim t ´α ˇ
ˇ1 ` n ˇ dt “ t´α e´t cos θ dt ă 8, Ð because 0ăθă π2
nÑ8 1 1
it follows that ´ ¯
f n2 “ O nα´1 plog nq β ,
and completely analogously this also holds for f n4 .
` ˘
Along γ3 , the function f pzq is bounded, but z´n “ O r´n ; consequently f n3 is
“exponentially small”.
The estimation with o (instead of O) works in an analogous fashion. (It is how-
ever slightly more complicated: the straight pieces must be “decomposed” in a
similar way as in the proof of Theorem 4.3.25. . . .) 
102 4. ASYMPTOTIC ENUMERATION

In the following examples, the well-known Gaußian integral

ż8 ?
s2
e´ 2 ds “ 2π. (4.43)
´8

appears again and again.

E XAMPLE 4.3.34 (Unary-binary trees). We consider the species T of (unlabelled non-

empty) planar unary–binary trees (i.e., each internal node has either one or two “de-
scendants”). The following figure illustrates schematically the decomposition of this
species:

T “ ` `
T T T

This decomposition can be directly translated into the following functional equation for
the (ordinary) generating function T pzq:
´ ¯
T pzq “ z 1 ` T pzq ` T 2 pzq .

Its solution is
a
1´z´ p1 ` zq p1 ´ 3zq
T pzq “ .
2z

The dominant singularity is visibly z “ 13 , and the function is analytic in a ∆-domain,

cf. the following figure:

´1 0 1
3

With the substitution

1´u
1 ´ 3z “ u ðñ z “ ,
3
 4.3. SINGULARITY ANALYSIS 103

we obtain
c´ ¯
1 ´ 1´u
3 ´ 1´u
1` 3 u
T puq “
2 1´u
b 3
2 ` u ´ 3 13 u p4 ´ uq
“ Ð2`u“2p1´uq`3u
2 p1 ´ uq
? ? b` ˘
3u ´ 3 u 2 1 ´ u4
“ 1` Ð2 binom. th. “trivial”
2 p1 ´ uq
? ?
3u ´ 2 3 u p1 ` O puqq
“ 1` p1 ` O puqq
2
? ? 3u ´ ¯
“ 1´ 3 u` ` O u3{2 .
2
Altogether, this leads to the expansion (now again in z):
? ? 3 ´ ¯
T pzq “ 1 ´ 3 1 ´ 3z ` p1 ´ 3zq ` O p1 ´ 3zq3{2 .
2

By (4.31) (with α “ ´ 12 ), we have

¨´ ¯˛
1
e k ´2
83
? ÿ
n´ 2
Js K 1 ´ s „ ´ ¯ ˝1 `
n
k
‚.
Γ ´1 k“1
n
2

By (4.22) in combination with (4.20), we have

ˆ ˙
1 ?
Γ ´ “ ´2 π,
2
3
´ by (4.41)¯ (with α “ ´ 2 and β “ 0), we see that the coefficient of the rest
and
O p1 ´ sq3{2 is of the order
´ 5¯
O n´ 2 .

Altogether, we obtain
c ´ ¯
n n 3 n ´ 52
Tn “ Jz K T pzq “ 3 `O 3 n
4πn3
for the coefficient Tn of zn in T pzq.

E XAMPLE 4.3.35 (2-regular graphs). The labelled species of 2-regular graphs can be
interpreted as sets of cycles of length at least 3 (“ cycles in the sense of graph theory),
where we identify cycles if they differ only by a reflection. Hence, the corresponding
104 4. ASYMPTOTIC ENUMERATION

exponential generating function equals

ˆ ˆ 3 ˙˙
1 z z4
f pzq “ exp ` `¨¨¨
2 3 4
ˆ ˙
1 1 z z2
“ exp log ´ ´
2 1´z 2 4
z z2
e´ 2 ´ 4
“ ?
1´z
We do the substitution u “ 1 ´ z:
1´u 1´2u`u2 3 u2
e´ 2 ´ 4 e´ 4 `u´ 4
f puq “ ? “ ?
u u
´ 34
e ´ ´ ¯¯
“ ? 1 ` u ` O u2
u
3 1 3 1
´ 3
¯
“ e´ 4 p1 ´ zq´ 2 ` e´ 4 p1 ´ zq 2 ` O p1 ´ zq 2 .

We basically proceed as in the previous example and apply Theorem 4.3.25 to the stan-
dard functions

‚ p1 ´ zq´1{2 , where we use two terms of the asymptotic expansion in (4.31),

‚ and to p1 ´ zq1{2 , where we use only the dominant term in (4.31).

This gives
3 1 3 5 ´ 5¯
f n “ e´ 4 ? ´ e´ 4 ? ` O n´ 2 .
πn 8 πn3
Hence, the number of 2-regular graphs is asymptotically equal to n! ¨ f n .

E XAMPLE 4.3.36 (Children’s rounds). The labelled species “children’s rounds” is a

partition of the set of children into several “individual rounds”, in each of which one
child sits in the centre; see the following illustration:

3 7

6 1 4

2 5

n!
There are exactly n´1 different such “rounds” with n children (since cyclic permuta-
tions of the children that sit around in a circle are identified); the exponential generating
function of such “single rounds” is therefore

z3 z4
z2 ` ` `¨¨¨ ,
2 3
 4.4. SADDLE POINT METHOD 105

and “children’s rounds” are simply sets of such “single rounds”. Consequently, the
corresponding generating function equals
ˆ ˙
2 z3 z4
f pzq “ exp z ` ` ` ¨ ¨ ¨
2 3
ˆ ˙
1
“ exp z log
1´z
1
“ p1 ´ zq´z “ p1 ´ zq´z`1
1´z
1
“ exp pp1 ´ zq log p1 ´ zqq .
1´z
This function is analytic in CzRě1 , and it has a singularity at ζ “ 1. We expand the
exponential function about z0 “ 0:
plogp1´zqp1´zqq2
1 ` log p1 ´ zq p1 ´ zq ` 2! `¨¨¨
f pzq “
1´z
1 ÿ plog p1 ´ zqqn p1 ´ zqn´1
“ ` log p1 ´ zq `
1´z n!
ně2
ˆ ˙
1 1 1 2 1
“ ´ log ` ¨ p1 ´ zq ¨ log
1´z 1´z 2 1´z
ˆ ˆ ˙˙
1
` O p1 ´ zq2 ¨ log3 (see [3, p. 397]).
1´z
1 1
The part 1´z ´ log 1´z contributes 1 ´ n1 to the coefficient. After application of Re-
mark 4.3.30, we obtain an estimation for the rest from (4.41) (with α “ ´1 and β “ 2,
respectively α “ ´2 and β “ 3). Thus, altogether we have
ˆ ˆ ˙˙ ´ ¯
1 ´2 1 ´3 2
f n “ 1 ´ ´ n ¨ plog nq ¨ 1 ` O ` O n ¨ log n
n log n
1 ´ ¯
´2 ´2
“ 1 ´ ´ n ¨ log n ` O n .
n
4.4. Saddle point method
Singularity analysis turned out to be extremely useful for the determination of
asymptotics. However, what do we do if the generating function
z2
‚ either has no singularities at all, such as e.g. ez (sets), ez` 2 (involutions)
or ee ´1 (Bell numbers),
z

‚ or cannot be approximated by our “standard functions” about the dom-

z
inant singularities, such as e.g. e 1´z (sets of non-empty permutations)?

4.4.1. Heuristics: contour through saddle point. In such cases, the follow-
ing heuristic argument may help. Starting point is again Cauchy’s integral for-
mula: ż
n 1 f pzq
Jz K f pzq “ dz. (4.44)
2π i C zn`1
106 4. ASYMPTOTIC ENUMERATION

The idea is to deform the contour C around 0 with winding number 1 inside the
domain G where the function f is analytic such that
f pzq
‚ the integrand zn`1 is so small on a part C1 of the contour C such that the
ş f pzq
part 2π1 i C1 zn`1 dz of the integral is negligible in the asymptotic consid-
erations,
‚ so that the asymptotically relevant part of the integral is the one coming
from the “rest” C2 of the contour C (where moreover this rest should
be easily computable, at least asymptotically).
This sounds somewhat unclear and vague, so that one cannot imagine how
this is going to work in practice. As a matter of fact, this vague idea can be
successfully carried through in so many cases such that it acquired a name:
saddle point method.8
So, to begin with, we have to estimate the modulus of the integrand in (4.44),
ˇ ˇ
ˇ f pzq ˇ
ˇ ˇ
ˇ zn`1 ˇ .
This real valued function appears graphically as a “mountain landscape” over
the complex plane; according to the above vague requirement, the path de-
scribed by the contour C should “almost always be close to the plane, and only
once lead through a mountain pass”. (See the following figure.)

f pzq
So, let us consider the function g pzq :“ zn`1
. By Taylor’s theorem, we have

1 ´ ¯
1 2 2 3
g pzq “ g pz0 q ` pz ´ z0 q g pzo q ` pz ´ z0 q g pz0 q ` O pz ´ z0 q for z Ñ z0 .
2
We write |g| in a neighbourhood of z0 :
Case 1: g pz0 q “ 0. In this case, the function |g| has a global minimum in z0 , see
the following figure:
8Recall that a trick that works at least three times becomes a method . . .
 4.4. SADDLE POINT METHOD 107

g1 pz q
Case 2: g pz0 q ‰ 0 and g1 pz0 q ‰ 0. Then we write gpz 0q “ λeiφ in polar coordi-
0
nates, and we see that
ˇ ´ ¯ˇ
ˇ ˇ
|g pzq| “ |g pz0 q| ¨ ˇ1 ` r ¨ λ ¨ eipφ`θq ` O r2 ˇ .

For small r along a line segment through z0 with angle θ, we have in first ap-
proximation:
‚ for θ “ ´φ ˘ π, the function |g pzq| grows, from a minimum „ |g pz0 q| p1 ´ rλq
to a maximum „ |g pz0 q| p1 ` rλq; ? ` ˘
‚ for θ “ ´φ ˘ π2 , we have |1 ˘ rλi| “ 1 ` r2 λ2 “ 1 ` O r2 “ 1 ` o prq,
so that the function g pzq is “essentially constant”.
g2 pz q
Case 3: g pz0 q ‰ 0 and g1 pz0 q “ 0. Let 21 gpz 0q “ λeiφ in polar coordinates. Then
0
ˇ ´ ¯ˇ
ˇ 2 ip2θ`φq ˇ
|g pzq| “ |g pz0 q| ˇ1 ` λr e ` O r3 ˇ .

For angle θ equal to

‚ θ “ ´ 2 , we have eip2θ`φq “ 1; the function |g pzq| behaves along this

φ

direction “essentially” like a Y-shaped parabola;

‚ θ “ ´ 2 ` π2 , we have eip2θ`φq “ ´1; the function |g pzq| behaves along
φ

this direction “essentially” like a X-shaped parabola.

In other words, the point z0 is a saddle point of |g pzq|. The “strategy” of the
saddle point method consists is to deform the the contour Γ for Cauchy’s integral
formula such that
‚ it passes through the saddle point
φ
‚ in the direction of “steepest descent”9 (i.e., in direction ´ 2 ` π2 , with
the above notation in polar coordinates);
hoping that the contour integral can be split into two parts:
‚ into one (small) part about the saddle point, which provides the “dom-
inant” contribution to the integral;
‚ and in a remaining part, which is “negligible”.
9This explains that the saddle point method is alternatively also called “method of steepest
descent”.
108 4. ASYMPTOTIC ENUMERATION

R EMARK 4.4.1 (Fundamental theorem of algebra). Each polynomial p pzq of degree

1
n ą 0 over the field C has a root in C. For, suppose not, then f pzq :“ ppzq would be
bounded,
| f pzq| ď C for all z P C,
and then, by Cauchy’s integral formula, for all Taylor coefficients of f we would get
C 1
|Jzn K f pzq| ď for all R ą 0,
2π Rn
that is, |Jzn K f pzq| “ 0 for n ą 0, a contradiction.
Let us see how the saddle point method works in a familiar example.
ř
E XAMPLE 4.4.2 (Stirling’s formula). Let f pzq “ ez “ ně0 n! 1 n
z . We want to find
a “good” contour of integration Γ for Cauchy’s integral formula
ż
1 1 ez
“ dz.
n! 2π i Γ zn`1
So we look for a saddle point: the equation
ˆ z ˙1
e ez ez
“ 0 ðñ ´ pn ` 1q “0
zn`1 zn`1 zn`2
has clearly the solution z “ n ` 1 — however, we do not need to stick strictly to our
“saddle point heuristic”, we may equally well try z “ n ´(n „ ¯n ` 1 for n Ñ 8; n is
so-to-speak an “asymptotic saddle point”). For g pzq :“ ez , the values g pnq and
zn`1
g2 pnq are real, thus the angle φ (see the preceding considerations: “Case 3”) is 0, and
as contour that cuts the real axis in the point n orthogonally we simply choose the circle
with radius r “ n:
z pθq “ n ¨ eiθ with derivative z1 pθq “ i ¨ n ¨ eiθ “ i ¨ z. (4.45)
We consider the contour integral
ż 2π ż 2π iθ
1 ezpθq 1 en¨e
1
¨ z pθq dθ “ ¨ i ¨ n ¨ eiθ dθ
2π i 0 pz pθqqn`1 2π i 0 nn`1 eipn`1qθ
ż
n´n 2π n¨eiθ ´i¨nθ
“ e dθ
2π 0
´ n ¯´n 1 ż 2π iθ
“ enpe ´1´iθq dθ.
e 2π 0
We split this integral now “skilfully”:
ż ż ż
1 2π 1 δ 1 2π´δ
“ ` ,
2π 0 2π
loomoon
´δ 2π
loooomoooon
δ

“:In0 “:In1

where we choose δ “ ˇn´2{5 .ˇ

ˇ iθ ˇ iθ
We first consider In1 : ˇen¨e ˇ “ eℜpn¨e q “ en cos θ is
‚ decreasing on the interval pδ, πq, the maximum is attained at δ,
‚ increasing on the interval pπ, 2π ´ δq, the maximum is attained at 2π ´ δ.
 4.4. SADDLE POINT METHOD 109

Consequently, we may bound the integrand In1 from above as follows:

ˇ ˇ
ˇ npeiθ ´1´iθq ˇ iθ
ˇe ˇ “ eℜpnpe ´1´iθqq
“ enpcos θ´1q
ř
ď enpcos δ´1q Ðcos z“ m z2m p´1qm p2mq!
1

2 ?
5n
´n δ2
„e “e . ´ 2

´ ´ 1
¯¯
Altogether we obtain In1 “ O exp ´ 21 ¨ n 5 .
Now we consider In0 :
żδ żδ ´ 2 ¯
npeiθ ´1´iθ q n ´ θ2 `Opθ 3 q ř
e dθ “ e dθ Ðe z “ zn
n n!
´δ ´δ
ˆ ˙
żδ 2 ´ 51
´n θ2 `O n
´ 65
“ e dθ Ð|θ 3 |ďδ3 “n
´δ
żδ ´ ´ ¯¯
θ2 1
“ e´n 2 ¨ 1 ` O n´ 5 dθ
´δ
´ ´ ¯¯ ż δ θ2
´ 15
“ 1`O n e´n 2 ¨ dθ.
´δ
We make small a computation on the side:
ż8 2
ż8
e´ 2 ¨pδ `2δt`t q dt Ðθ“t`δ
n 2 2
´n¨ θ2
e dθ “
δ 0
2
ż8
´ nδ2
ďe e´nδt dt
ˆ0 ˙ˇ
nδ2 1 ´nδt ˇˇ8
“e ´ 2
´ e ˇ
nδ 0
ˆ 1
˙
1 ´ nδ 2
2
“ e 2 “O e ´n 5
. Ðδ“n´ 5
nδ
Hence, we have
żδ ż8 ´ ¯
2
´n¨ θ2 ´n¨ θ2
2
´n1{5 ?n
e dθ “ e dθ ` O e Ðs“θ 2
´δ ´8
c ż ´ ¯
2 8 ´s2 ´n1{5
“ e ds `O e
n looooomooooon
´8
?
π
c ´ 1¯
2π
“ ` O n´ 5 . ÐOpe´x q“Op 1x q
n
If we put all this together, then we obtain the following variant of Stirling’s formula:
1 ´ n ¯´n 1 ´ ´ ¯¯
“ ? 1 ` O n´1{5 .
n! e 2πn
110 4. ASYMPTOTIC ENUMERATION

E XAMPLE 4.4.3. Let un be the number of ordered set partitions; these are like ordinary
(set) partitions except that the order of the elements in each block is relevant. In the
language of species, these are
sets ppermutationens 1 q ,
from which we directly read off the (exponential) generating function:
ÿ zn ˆ ˙
z
un “ exp .
ně0
n! 1 ´ z

For Cauchy’s integral formula, we have to deal with the integrand

` z ˘
exp 1´z
f pzq “ .
zn`1
Its derivative is
` z ˘ ˜ ¸
exp 1´z z
f 1 pzq “ n`2
¨ 2
´ pn ` 1q . Ðp 1´z
z 1
q “ p1´zq
1
` z 2“ 1 2
z p1 ´ zq p1´zq p1´zq

If we equate f 1 pzq to zero, we are led to the equation

z2 pn ` 1q ´ z p2 pn ` 1q ` 1q ` n ` 1 “ 0.
Its roots are
?
2n ` 3 ˘ 4n2 ` 12n ` 9 ´ 4n2 ´ 8n ´ 4 ?
b2 ´4ac
z1,2 “ Ðz1,2 “ ´b˘
2 pn ` 1q 2a
?
2n ` 3 ˘ 4n ` 5
“ .
2 pn ` 1q
We choose the smaller root
a
1 4 pn ` 1q ` 1
z1 “ 1 ` ´ ,
2 pn ` 1q 2 pn ` 1q
and, from the series expansion
? z z2 z3 5z4 ´ ¯
1`z “ 1` ´ ` ´ ` O z5 p|z| ă 1q
2 8 16 128
of the square root about z “ 0, we obtain
b
? 1
4z ` 1 1 ` 4z 1 ´ ´
´1
¯¯
“ ? “ ? 1 ` O p4zq ; pz Ñ 8q ,
2z z z
and, for z “ pn ` 1q, we get
ˆ ˆ ˙˙
1 1 1
z1 “ 1 ` ´? 1`O pn Ñ 8q .
2 pn ` 1q n`1 4 pn ` 1q
By the series expansion
1 z 3z2 5z3 35z4 ´ ¯
? “ 1´ ` ´ ` ` O z5 p|z| ă 1q
1`z 2 8 16 128
 4.4. SADDLE POINT METHOD 111

about z “ 0, we now obtain

ˆ ˆ ˙˙
1 1 1 1 1
? “? b “? 1`O pn Ñ 8q .
n`1 n 1` 1 n n
n

Altogether, we have
1 1 ´ ¯
z1 “ 1 ´ ? ` ` O n´3{2 .
n 2 pn ` 1q
Again, we approach the situation “generously”. We choose the “asymptotic saddle
point”
1
r “ 1´ ?
n
on the real axis, which is at the same time the radius of the circle about 0 that we choose
as integration contour (see the preceding example, in particular (4.45)). Then Cauchy’s
integral formula yields
ż ˆ ˙
un 1 2π r ei θ
“ exp ´ niθ r´n dθ.
n! 2π 0 1 ´ r ei θ
As in the preceding example, we split this integral “skilfully”:
ż 2π ż 2π´δ żδ ż 2π´δ
“ “ ` ,
0 ´δ ´δon loδomoon
loomo
“:In0 “:In1

where we choose δ “ n´7{10 .

Since r “ 1 ´ ?1n , we have

r ei θ
“ r ¨ eiθ ` r2 ¨ e2iθ ` r3 ¨ e3iθ ` ¨ ¨ ¨ Ðeiθ “1`iθ´ θ2 `Opθ3 q
2

1 ´ re i θ
¨ ˛ ¨ ˛

˚ ‹ ˚ ‹
r ` r2 ` r3 ` ¨ ¨ ¨‚` iθ ˚
“ 1 ˝loooooooomoooooooon ˝ r ` 2r 2
` 3r 3
` ¨ ¨
loooooooooomoooooooooon ¨‹
‚
r ? r ?
1´r “ n´1 p1´rq2
“n´ n
¨ ˛

θ2 ˚ ‹
˚ 2 2 2 3 ‹
´ r ` 2 r ` 3 r ` ¨ ¨ ¨‹
˚looooooooooomooooooooooon
2 ˝ ‚
rp1`rq ?
“ n´3n`2n3{2
p1´rq3
¨ ˛
˚ ‹
´ ¯˚ ‹
3 ˚ 3 2 3 3 ‹
r ` 2 r ` 3 r ` ¨ ¨ ¨ ‹ pn Ñ 8q .
` O θ ˚ looooooooooomooooooooooon
˚ ‹
˝ rp1`4r`r2q ˆ ˙ ‚
p1´rq4
“O 1
p1´rq4
“O p n 2 q
112 4. ASYMPTOTIC ENUMERATION
` 3˘ ` 2˘ ´ ¯
3
For the part In0 ,
we have |θ| ď δ3 “ n´21{10 , hence O θ O n “ O n ´1{10 .
Consequently, we get
ˆ ˙ ˆ ˙
żδ rp1`rq θ3
1´r `iθ
r r
1 ´n ´θ 2 `O
In0 r n “ e p1´rq2 2p1´rq3 p1´rq4
dθ
2π ´δ
żδ ´? ´ ¯ ´ ¯¯
1 ?
“ exp n ´ 1 ´ iθ n ´ θ 2 n3{2 ` O pnq ` O n´1{10 dθ
2π ´δ
żδ ´? ´ ¯¯
1
“ exp n ´ 1 ´ θ 2 n3{2 ` O n´1{10 dθ,
2π ´δ
where the last simplification follows from
? ´ ¯ ´ ¯
´2{10 2 ´4{10
θ n“O n , θ O pnq “ O n .

As in the preceding example, we have

ż8 ż8
e´pδ `2δt`t qn dt Ðθ“δ`t
´θ 2 n3{2 2 2 3{2
e dθ “
δ 0
ż8
´n3{2 δ2 3{2
ďe e´2δtn dt
ˆ0 ˙ˇ8
3{2 2 1 ´2δtn3{2 ˇ
ˇ
“e ´n δ
e ˇ
´2δn3{2 0
3{2
e´n δ
2 ´ 1{10
¯
“ 3{2
“ O e´n , Ðδ“n´7{10 ,
2δn
and thus
¨ ˛
˚ż 8 ´ ¯ ‹
1 ?n´1 ˚ ‹´ ´ ¯¯
In0 r n “ e ˚
˚ exp ´θ 2 3{2
n dθ ‹ 1 ` O n´1{10 .
‹
2π ˝looooooooooooomooooooooooooon
´8 ‚
?
πn´3{4

Together with the expansion

ˆ ˆ ˙˙
´n 1
r “ exp ´n log 1 ´ ?
n
ˆ ˆ ´ ¯˙˙
1 1 ´3{2
“ exp ´n ´ ? ´ `O n
n 2n
ˆ ´ ¯˙
? 1 ´1{2
“ exp n` `O n
2
? ? ´ ´ ¯¯
n` 21 Opn´1{2 q n` 12
“e e “e 1`O n ´1{2
,

we eventually arrive at
?
´
e2 ´n´ 12 ¯¯
In0 “ ? 3{4 1 ` O n´1{10 .
2 πn
 4.4. SADDLE POINT METHOD 113

For the remaining part of the integral, which we denoted by In1 , we collect the following
z 1
facts: we have (of course) e 1´z “ e 1´z ´1 and (equally obviously) 1´z 1 1´z
“ p1´zqp1´zq ;
from this, it follows immediately that
ˇ z ˇ 1 ˇˇ ˆ
1 ´ r e ´i θ
˙ˇ
ˇ 1
ˆ
1 ´ r cos θ
˙
ˇ 1´z ˇ
ˇe ˇ “ ˇˇexp ˇ “ exp .
e 1 ´ 2r cos θ ` r2 ˇ e 1 ´ 2r cos θ ` r2
1´rz
The derivative of the function 1´2rz`r 2 (z P r´1, 1s, 0 ă r ă 1), with respect to z
rp1´r q
2
is 2 ą 0. Consequently, it is (strictly) monotone increasing. However, this
p1´2rz`r2 q
implies that, for z “ reiθ and δ ď θ ď 2π ´ δ, we have
ˇ z ˇ ˆ ˙
ˇ 1´z ˇ 1 ´ r cos δ
ˇe ˇ ď e exp
´1
,
1 ´ 2r cos δ ` r2
and ˜ ¸
1 ´ r cos n´7{10 ´? ¯
1{10
exp “ exp n´n ` O p1q .
1 ´ 2r cos n´7{10 ` r2
(This is seen — after the substitutions r “ 1 ´ ?1n and n “ z´10 — from the Laurent
series expansion
` 5 ˘ ` ˘ ´ ¯
z ´ 1 cos z7 ` 1 1 1
“ 5 ´ ` O z3 ,
z ´ 2z ` 2 pz ´ 1q cos pz q ` 2
10 5 5 7 z z
and the “backward substitution” z “ n´1{10 .) Furthermore,
2π ´n ?n´n1{10 Op1q ´ ? 1{10 ¯
In1 ď r e e “ O e2 n´n ,
2π
with the expansion of r´n as before. This may not look very spectacular: however,
asymptotically, In1 is “much smaller” than In0 :
ˇ 1ˇ ˇ ˇ
ˇ In ˇ
ˇ ˇ ď C ˇˇn3{4 e´n1{10 ˇˇ ,
ˇ I0 ˇ
n
hence ´ ¯
In1 “ o In0 pn Ñ 8q .

R EMARK 4.4.4 (Longer asymptotic expansion). We may take more terms in the
Taylor series expansion in θ — for ez (see Example 4.4.2) we get for instance
ż8 ż8 ´ 2 3 4 θ 5 `O θ 6
¯
npeiθ ´1´iθ q n ´ θ2 ´i θ6 ` θ24 `i 120 p q
e dθ “ e dθ Ðez “1`z`Opz2 q!
´8 ´8
ż8 ˆ ´ ¯˙
2
´n θ2 θ3 θ4 θ5
“ e 1 ´ ni ` n ` ni ` nO θ 6 dθ
´8 6 24 120
? ż
2π n 8 4 ´θ2 n
niO p. . . q `
“ ? ´ loooomoooon θ e 2 dθ
n 24 ´8
“0, see below.
ˆ ż8 ˙
6 ´θ 2 n2
`O n θ e dθ .
´8
114 4. ASYMPTOTIC ENUMERATION

Here, the imaginary part is zero since θ 2k´1 is an odd function.

With the integrals
ż8 ?
2 n 3 2π
θ 4 e´θ 2 dθ “ 5{2
´8 n
ż8 ?
2 n 15 2π
θ 6 e´θ 2 dθ “ 7{2
,
´8 n
we obtain ? ?
2π 1 2π ´ ¯
? ` ´5{2
`O n .
n 8 n3{2

ř 4.4.2. nHayman’s theorem. Our typical situation is the following: let f pzq “
ně0 f n z be analytic about z “ 0; we want to express the coefficient f n by using
Cauchy’s integral formula, where we integrate along a circle z pθq “ reiθ :
ż ` ˘
r´n π f reiθ
fn “ dθ.
2π ´π eniθ
The subsequent steps follow the general pattern:
‚ show that the “tails” of the integral (i.e., θ R p´δ, δq) are “asymptotically
negligible in comparison to the main part”;
‚ show that, for the main part of the integral (i.e., θ P p´δ, δq), there exists
2
an approximation of the integrand of the type e´θ ;
‚ show that the tails of these “new, approximative integrals” (i.e., the
integration domains θ ă ´δ and θ ą δ) are again asymptotically negli-
gible;
see the following schematic illustration.

ignore tails
I:

´π ´δ δ π
2
approximate by e´t
II: in the centre

´δ δ
add tails
III:
´8 `8

We now make situations in which this “general pattern” works more precise.
For r ě 0 in the circle of convergence with f prq ą 0 and small θ, we have the
series expansion in θ
´ ´ ¯¯ 8
ÿ piθq p
log f reiθ “ log p f prqq ` α p prq . (4.46)
p!
p“1
 4.4. SADDLE POINT METHOD 115

Of particular interest for our purposes are the first two terms, which, using the
notation h prq :“ log p f prqq, we write
f 1 prq
a prq :“ α1 prq “ rh1 prq “ r (4.47)
f prq
b prq :“ α2 prq “ r2 h2 prq ` rh1 prq
˜ ˆ 1 ˙2 ¸
f 1 prq f 2 prq f prq
“r ` r2 ´ . (4.48)
f prq f prq f prq

D EFINITION 4.4.5. Let f pzq be analytic about 0 with radius of convergence 0 ă R ď

8, and assume furthermore that f pzq ą 0 for z P pR0 , Rq for some 0 ă R0 ă R.
The function f pzq is called Hayman-admissible, if it satisfies the following three
conditions:
‚ H1: limrÑR a prq “ `8 and limrÑR b prq “ `8;
‚ H2: there is a function δ prq on pR0 , Rq with 0 ă δ prq ă π such that we have
uniformly for |θ| ď δ prq:
´ ¯
f reiθ “ f prq eiθaprq´pθ bprqq{2`op1q pr Ñ Rq ;
2

‚ H3: we have uniformly for δ prq ď |θ| ă π:

˜ ¸
´ ¯ f prq
f r ei θ “ o a pr Ñ Rq .
b prq

In practice, the first thing to do is to find the function δ prq: with the notation
from (4.46), by H2 and H3 we should for r Ñ R on the one hand have

α2 prq δ prq2 Ñ 8, Ð (follows from the combination of H2 and H3),

and on the other hand

α3 prq δ prq3 Ñ 0. Ð (follows from H2 since the error term in the exponent is of the order o p1q)

Altogether, this leads to the necessary condition

α3 prq2
lim “ 0.
rÑR α2 prq3
For the “saddle point Ansatz”, we need
´ ¯
´1{2 ´1{3
α2 prq “ o pδ prqq and δ prq “ o α3 prq , (4.49)

that is, e.g.

δ prq “ α2 prq´1{4 ¨ α3 prq´1{6 .
z2
R EMARK 4.4.6. For example, ez , ee ´1 , and ez` 2 (with R “ 8) or e 1´z (with R “
z z

2
1) are Hayman-admissible functions. Not Hayman-admissible is for example ez
(becomes “too large” about π and violates H3).
116 4. ASYMPTOTIC ENUMERATION
ř
T HEOREM 4.4.7 (Hayman). Let f pzq “ něo f n zn be Hayman-admissible. For n P
N, let ζ “ ζ pnq be the largest (and, thus, unique) solution of the equation

f 1 pzq
a pzq “ z “n
f pzq

in the interval pR0 , Rq. Then, for n Ñ 8, we have

f pζq
fn „ a ,
ζ n 2πb pζq

d 2 d
where b pzq “ z2 dz 2 log f pzq ` z dz log f pzq.

P ROOF. As always, we use Cauchy’s integral formula in the “split” Version

ż ż
1 δ ´ iθ ¯ ´niθ 1 2π´δ ´ iθ ¯ ´niθ
fnr “n
f re e dθ ` f re e dθ ,
2π ´δ
loooooooooooooomoooooooooooooon 2π δ
loooooooooooooooomoooooooooooooooon
In0 In1

where we set δ “ δ prq. The part In1 can be immediately approximated by as-
sumption H3:
˜ ¸
ˇ ˇ 1 f prq
ˇ 1ˇ
ˇ In ˇ ď 2π ¨ o a .
2π b prq

For the main part In0 , we want to show:

˜ ˜ ¸ ¸
2
f prq pa prq ´ nq
In0 “ a exp ´ ` o p1q pr Ñ Rq (4.50)
2πb prq 2b prq

(independent of n). To achieve this, we use assumption H2:

żδ ˆ ˙
1 1
In0 “ f prq exp i pa prq ´ nq θ ´ b prq θ ` o p1q dθ
2
2π ´δ 2
żδ
f prq
eiθpaprq´nq´ 2 bprqθ p1 ` o p1qq dθ
1 2
“ Ðeop1q “1`op1q
2π
˜´δ
ż ˜ż ¸¸
δ δ
f prq iθpaprq´nq´ 12 bprqθ 2 ´ 21 bprqθ 2
“ e dθ ` o e dθ Ð|eiθpa´nq |“1
2π ´δ ´δ
˜ż ¸
f prq δ ´ ¯ 2 ?
“ e iθpaprq´nq´ 12 bprqθ 2 dθ ` o b prq´1{2
ş8 ´ t 2b
. ´8 e
Ð dt“ ?2π
2π ´δ
b

The assumptions H2 and H3 both hold for θ “ δ prq. Together, they yield
a
b prqeiδprqaprq´δprq
2
bprq{2
Ñ 0 ùñ δ prq2 b prq Ñ 8 for r Ñ R.
 4.4. SADDLE POINT METHOD 117
b ?
bprq
We perform the substitution t “ 2 θ (i.e., dθ “ ? 2
dt), and we obtain for
bprq
r Ñ R:
¨ b ˛
ż δ bprq d b ´ ¯
f prq ˝ 2 2 ´t `it bprq paprq´nq
2 2
In0 “ ¨ b e dt ` o b prq´1{2 ‚
2π ´δ
bprq
2
b prq
¨ b ˛
ż δ bprq b
f prq ˝ 2 ´t `it bprq paprq´nq
2 2
“ a b e dt ` o p1q‚ Ðform a square
π 2b prq ´δ
bprq
2
˜ż ´ b ¯2 ¸
8 paprq´nq2
f prq ´ t´i 2bprq1
paprq´nq ´ 2bprq ?
“ a e dt ` o p1q Ðδ bprq{2Ñ8
π 2b prq ´8
¨ ˛
˚ b
ż 8´i 1 paprq´nq ‹
f prq ˚ paprq´nq2
˚ ´ 2bprq 2bprq 2
‹
‹
“ a ˚e b e ´s
ds `o p1q ‹.
˚
π 2b prq ˝ ´8´i 2bprq paprq´nq
1 ‹
looooooooooooooomooooooooooooooon ‚
?
π

Thus, we established (4.50), and the claim of the theorem follows immediately:
for, by H1, we have limrÑR a prq “ `8, and hence
lim ζ pnq “ R;
nÑ8
we simply put r “ ζ pnq. 
The convenient feature of the concept of Hayman-admissibility is that there
hold the following “closure properties”: “Large classes of functions are Hayman-
admissible”.
T HEOREM 4.4.8. Let f pzq and g pzq be Hayman-admissible functions, and let p pzq be
a polynomial with real coefficients. Then:
‚ f pzq ¨ g pzq and e f pzq are also Hayman-admissible;
‚ f pzq ` p pzq is Hayman-admissible; if the leading coefficient of p pzq is posi-
tive, then also f pzq p pzq and p p f pzqq are Hayman-admissible;
‚ If almost all Taylor coefficients of e ppzq are positive, then also e ppzq is Hayman-
admissible.

S KETCH . In the first place, one has to find the function δ prq: to this end, one
uses (4.49). 
We give some examples in order to illustrate the concept of Hayman-admissibility.
E XAMPLE 4.4.9. For f pzq “ ez , we have the expansion
´ ¯ θ2
i “ er`riθ´r 2 `Oprθ q
3
f re θ

2
“ f prq ¨ eriθ´r 2 `Oprθ q .
θ 3

Hence, a prq “ b prq “ r, and thus H1 is of course satisfied.

118 4. ASYMPTOTIC ENUMERATION

For H2 we must choose δ “ δ prq such that

lim rδ3 “ 0.
rÑR
´ ¯
Thus, δ “ o r´1{3 , and for H3 to hold we must achieve
ˇ ´ ¯ˇ ˇ iθ ˇ ˇ ˇ ˆ r˙
ˇ i θ ˇ ˇ re ˇ ˇ r cos θ ˇ !?! e
ˇ f r e ˇ “ ˇe ˇ “ ˇe ˇ“o ? .
r
2 ` ˘ ? 1
Since r cos θ “ r ´ r θ2 ` O rθ 4 and r “ e 2 log r , these conditions will be satisfied
for ´ ¯
log r “ o rθ 2 ,
so, for example, for δ “ r´2{5 .
1 1
E XAMPLE 4.4.10. For f pzq “ 1´z , we consider the Taylor series expansion of log 1´reit
about t “ 0:
˜ ¸
1 1 ir ´r r ¨ pr ` 1q
log i “ log ¨ t0 ` ¨ t1 ` ¨ t2 ` O ¨ t3 .
1 ´ re t 1´r 1´r p1 ´ rq 2
p1 ´ rq 3

We obtain the expansion

˜ ˜ ¸¸
´ ¯ 1 ir r r ¨ pr ` 1q 3
f re i θ
“ exp ¨θ´ 2
¨θ `O ¨θ
1´r 1´r p1 ´ rq2 p1 ´ rq3
r r
Clearly, here we have a prq “ and b prq “ , and H1 is satisfied (R “ 1). In
1´r p1´rq2
order to satisfy H2, we should have θ “ o p1 ´ rq. For H3 to hold we must achieve
˜ ¸
ˇ ´ ¯ˇ f prq
ˇ ˇ
ˇ f r ei θ ˇ “ o a ,
b prq
that is, ˜ ¸
ˇ ˇ 1 ˆ ˙
ˇ 1 ˇ !?! 1
ˇ ˇ “ o 1´r
? “o ? .
ˇ 1 ´ r ei θ ˇ r r
1´r
However,
1 1
ˇ ˇ“b
ˇ1 ´ reiθ ˇ
p1 ´ r cos θq2 ` pr sin θq2
1
“? , Ðsin2 θ`cos2 θ“1
1 ´ 2r cos θ ` r2
and for r Ñ R “ 1 we have θ Ñ 0 (since θ must be of order θ “ o p1 ´ rq); more
precisely,
θ2 ´ ¯
cos θ “ 1 ´ ` O θ 4 “ 1 ´ o p1 ´ rq2 .
2
Hence,
1 1
ˇ1 ´ reiθ ˇ “
ˇ ˇ b ,
p1 ´ rq2 ` 2r ¨ o p1 ´ rq
 4.4. SADDLE POINT METHOD 119

and for r Ñ 1 this tends to 8: consequently f pzq is not Hayman-admissible.

2
E XAMPLE 4.4.11. For f pzq “ ez , we have
´ ¯ θ2 ´ ¯
log f reiθ “ r2 e2iθ “ r2 ` 2r2 iθ ´ 4r2 ` O r2 θ 3 ,
2
thus a prq “ 2r2 and b prq “ 4r2 . However, H3 cannot be achieved. For, if we would
assume that condition, then, for δ ď |θ| ď π, we would need
˜ 2¸
ˇ ´ ¯ˇ er
ˇ ˇ
ˇ f reiθ ˇ “ er cos 2θ “ o
2 ???
pr Ñ R “ 8q ,
2r
which is of course wrong for θ “ π.
E XAMPLE 4.4.12. Let f n be the number of permutations of rns, which have only cycle
lengths 2 and 3 in their disjoint
` cycle decomposition.
˘ In the language of species, we
consider the composition sets cycles2,3 :
ÿ fn z2 z3
f pzq “ zn “ e 2 ` 3 .
ně0
n!

We have
´ ¯ 1 1
log f reiθ “ e2iθ r2 ` e3iθ r3
2 3
ˆ 2 3 ˙ ´ ¯ ´ ¯ 2 ´ ¯
r r 3 θ
“ ` ` r ` r iθ ´ 2r ` 3r
2 3 2
` O θ3
2 3 loooomoooon looooomooooon 2
aprq bprq

This function is Hayman-admissible, and hence

ζ2 ζ3
e2`3
f n „ n! a , as n Ñ 8, (4.51)
ζn 2πζ 2 p3ζ ` 2q
where ζ “ ζ pnq is a solution of
a pζq “ ζ 2 p1 ` ζq “ n; (4.52)
i.e., ζ pnq „ n1{3 . However, since ζ appears in the exponential in (4.51), this first
approximation for ζ is not good enough for getting an asymptotic formula for f n in
terms of familiar functions; we need more terms in the asymptotic expansion for ζ. How
this is done is explained below. It illustrates another important concept in asymptotic
analysis, the so-called “bootstrap method”.10
In (4.52), we make the Ansatz ζ “ n1{3 ` ρ, with ρ of smaller asymptotic order than
n1{3 . We expand the product and obtain
´ ¯
n ` n2{3 p1 ` 3ρq ` n1{3 2ρ ` 3ρ2 ` ρ2 ` ρ3 “ n,

10The name comes from the idea to pull oneself out of one’s boots, similar to Münch-
hausen’s famous story where he pulls himself out of a swamp. This is somehow how the
method works.
120 4. ASYMPTOTIC ENUMERATION

or, equivalently,
´ ¯
n2{3 p1 ` 3ρq ` n1{3 2ρ ` 3ρ2 ` ρ2 ` ρ3 “ 0. (4.53)

Now we suppose that ρ is of larger asymptotic order than a constant. In that case, the
term 3n2{3 ρ is asymptotically larger than all other terms on the left-hand side of (4.53).
This is a contradiction because then the left-hand side cannot be zero. Hence, ρ must be
asymptotically constant. In that case, the asymptotically largest terms on the left-hand
side of (4.53) are n2{3 p1 ` 3ρq. They must (asymptotically) “cancel” each other, by
which we mean that ρ must be chosen such that they actually are of smaller order. We
infer that ρ “ ´ 13 ` o p1q.
We need to continue since, by our findings so far, the term ζ 3 (which appears in the
exponential in (4.51)) is still not approximated to an order that would tend to zero. So,
we let ζ “ n1{3 ´ 31 ` ρ, with ρ of smaller (asymptotic) order than a constant.11 If we
substitute this Ansatz in (4.53), then we obtain
1 1 2
3n2{3 ρ ` 3n1{3 ρ2 ´ n1{3 ` ρ3 ´ ρ ` “ 0. (4.54)
3 3 27
By arguments
´ that¯are analogous to the previous ones, here we find that ρ must be of
the order O n´1{3 . In that case, the asymptotically largest terms in (4.54) are 3n2{3 ρ
and ´ 13 n1{3 . They
´ must
¯ (asymptotically) cancel each other, and therefore we see that
ρ “ 91 n´1{3 ` o n´1{3 .

This is still not good enough. Another round of “bootstrap” reveals that
1 1 ´1{3 2 ´ ¯
ζ “ n1{3 ´ ` n ´ n´2{3 ` o n´2{3 .
3 9 81
If this is substituted in (4.51), we obtain
1 1 2{3 ´ 1 n1{3 ` 1
e 3 n` 2 n 6 18
f n „ n! ? n 1
, as n Ñ 8.
6πn 3 ` 2
Here we used that
` ˘n
ζ n “ nn{3 ζn´1{3
ˆ ´ ¯˙n
n{3 1 ´1{3 1 ´2{3 2 ´1 ´1
“n 1´ n ` n ´ n `o n
3 9 81
ˆ ˆ ´ ¯˙˙
n{3 1 ´1{3 1 ´2{3 2 ´1 ´1
“ n exp n log 1 ´ n ` n ´ n `o n
3 9 81
´ ¯
“ nn{3 exp ´ 13 n2{3 ` 18
1 1{3
n ` o p1q , as n Ñ 8.

11We allow ourselves the liberty of using the same symbol again although the ρ here has
nothing to do with the earlier ρ.
 4.4. SADDLE POINT METHOD 121

4.4.3. A saddle point theorem for large powers.

E XAMPLE 4.4.13. If we consider power series F and f which are compositional inverses
of each other, that is,
F p f pzqq “ z,
then the coefficients of F pzq can be determined by the use of Lagrange inversion:
1 r ´1 z
n
Jz K F pzq “ z f pzq´n .
n
We write f as
z
f pzq “ .
ϕ pzq
Then we obtain the following contour integral for the coefficients of F pzq:
ż n
n 1 1 ϕ pzq
Jz K F pzq “ dz.
n 2π i Γ zn
E XAMPLE 4.4.14. Let Wn denote the number of all paths in the lattice Z ˆ Z that
consist of right steps p1, 0q, diagonal up-steps p1, 1q, and diagonal down-steps p1, ´1q,
and which start at p0, 0q and end at pn, 0q. We have
r z ˆ1 ˙n ż ` ˘n
0 1 1 ` z ` z2
Wn “ z `1`z “ dz.
z 2π i Γ zn`1
These two examples motivate to consider a more general problem: determine
r z ż
N n 1 A pzq B pzqn
z A pzq B pzq “ dz
2π i Γ z N`1
for n Ñ 8 and N Ñ 8.
ř ř
T HEOREM 4.4.15. Let A pzq “ jě0 a j z j and B pzq “ jě0 b j z j be two functions
analytic about 0 and satisfying the following properties:
L1: A and B have non-negative coefficients and B p0q ‰ 0.
` (˘
L2: B pzq is aperiodic, i.e., gcd j : b j ą 0 “ 1.
L3: The radius of convergence R of B pzq is ă 8; the radius of convergence of A pzq is
ě R.
Let T be the (left) limit
B1 psq
T :“ lim s .
sÑR´ B psq
N
Then, for λ :“ n, we have: if λ P p0, Tq, then
r z B pζqn
N n
z A pzq B pzq “ A pζq a p1 ` o p1qq , (4.55)
ζ N`1 2πn ¨ ξ
where ζ is the largest solution of the equation
B1 pζq
α1B pζq “ ζ “λ
B pζq
(cf. (4.47)), and
d2
ξ :“ plog B psq ´ λ log sq|s“ζ .
ds2
122 4. ASYMPTOTIC ENUMERATION

This holds uniformly for λ in an arbitrary compact subinterval of p0, Tq.

For the proof we need the following simple lemma.

ř
L EMMA 4.4.16. Let f pzq “ ně0 f n zn be analytic about 0 for |z| ă R with non-
negative coefficients, of which at least two are positive. If there is a non-real z with
|z| ă r, for which
| f pzq| “ f p|z|q ,
then
‚ this z must be of the form
p
2π i q
z “ r¨e ,
where 0 ă p ă q and gcd pp, qq “ 1,
‚ and there must exist an a P Z such that f n ‰ 0 only for n ” a pmod qq.

P ROOF OF THE LEMMA . Of course, we have

ˇ ˇ
ˇÿ ˇ ÿ ÿ
ˇ nˇ
ˇ fn z ˇ ď | f n zn | “ f n |zn | . Ð f n ě0
ˇ ˇ
ně0 ně0 ně0

Equality for a z “ reiθ can only hold if all powers zn “ r n eniθ (interpreted as
vectors in R2 ) point in the same direction. So let two arbitrary coefficients be
non-zero, say f n1 , f n2 ‰ 0. Then we must have
en1iθ “ en2iθ ðñ |n1 ´ n2 | θ “ 2kπ,
k
and thus θ “ 2π |n , where we may of course choose k ă |n1 ´ n2 |. Let
1 ´n2 |
p k
q “ |n1 ´n2 | be the reduced fraction. Then we have

q | n1 ´ n2 ,
or, equivalently, n1 ” n2 pmod qq. 

P ROOF OF THE THEOREM . For fixed r with 0 㡠r `ă R,˘ˇby assumption L2 (ape-

riodicity of B) and Lemma 4.4.16, the function ˇB reiθ ˇ attains its unique maxi-
mum at θ “ 0. Therefore there exists a (small) θ1 P p0, πq such that
ˇ ` ˘ˇ ˇ ` ˘ˇ
‚ ˇˇB reiθ ˇ ă ˇ B reiθ1 ˇ for θ P rθ1 , πs;
` ˘ˇ
‚ ˇB reiθ ˇ is (strictly) monotone decreasing on r0, θ1 s.

Consequently, if we integrate along the contour z “ ζ eiθ , then we obtain the

desired coefficient as J pπq, where
żθ ´ ¯ ´ ¯n
1 i i
J pθq :“ N A ζe θ
B ζe θ
e´Niθ dθ,
ζ 2π ´θ
and where for n Ñ 8 the difference
ˇ ` Jiθpπq
˘ˇ ´ J pθ1 q is exponentially small (be-
cause of the above estimation for B re ˇ). If we expand as in (4.46), we obtain
ˇ
 4.4. SADDLE POINT METHOD 123

for the integrand the approximation (N “ λn)

¨ ¨ ˛˛
˚ ˚ B ‹‹
A pζq B pζqn ¨ exppiθ ˝α1A pζq ` n ˝ loooomoooon
α1 pζq ´ λ ‚‚
“0 by assumpt.

θ2 ´ A ¯ ´ ¯
´ α2 pζq ` nα2B pζq ` O n ¨ θ 3 q,
2
for α3A pζq and α3B pζq are O p1q for n Ñ 8 since ζ does not depend on n. Now let
θ0 :“ n´2{5 . Then, for θ0 ď θ ď θ1 , we have (of course)
2 ´ ¯ ´ ¯
´ θ2 pα2A pζq`nα2B pζqq`Opn¨θ 3 q ´n1{5 Opn´1{5 q ´n1{5
e “O e ¨e “O e ,
´ 1{5
¯
for n Ñ 8. Hence also J pθ1 q ´ J pθ0 q “ O e´n for n Ñ 8, that is, exponen-
tially small. The desired coefficient is therefore´asymptotically ¯ J pθ0 q: for θ ă θ0
2
we have however (of course) θ2 ¨ α2A pζq “ O n´4{5 pn Ñ 8q — that is, the
desired coefficient is (after “appending” the asymptotically negligible integral
ends) asymptotically equal to
ż
A pζq B pζq n 8 ´ θ 2 nαB pζq A pζq B pζq n
e 2 2 dθ “ b .
2πζ N ´8 ζ N B
2πn ¨ α2 pζq
` ˘
The coefficient α2B is the third Taylor coefficient in the expansion of log B reiθ ,
hence
d2 ´ ¯ˇˇ B1 prq 2
ˆ 1 ˙2
i 2 B prq 2 B prq
α2 prq “ ´ 2 log B re ˇ
B θ ˇ
“r `r ´r ,
dθ θ“0 B prq B prq B prq
and we have
ˇ ˆ 1 ˙2
d2 ˇ
ˇ B 2 pζq B pζq λ α2B pζq
ξ“ plog B psq ´ λ log sq ˇ “ ´ ` “ .
ds2 s“ζ B pζq B pζq ζ2 ζ2
This establishes our claim. 
E XAMPLE 4.4.17. Let f pzq be given by the equation
f pzq “ z ¨ e f pzq ðñ f pzq e´ f pzq “ z.
I.e., f is the compositional inverse of z ¨ e´z . By Lagrange inversion, we get
1 r ´1 z ` ´z ˘´n
fn “ z ze
n
1 r ´1 z enz
“ z
n zn
r
1 n´1 z
“ z pez qn
n
1
“ Jzn K pez qn ¨ z.
n
124 4. ASYMPTOTIC ENUMERATION

In the notation of Theorem 4.4.15, we have A pzq “ z and B pzq “ ez , as well as n “ N

(hence λ “ 1). By the theorem, we obtain
nn´1 1 enζ
„ ζ a ,
n! n ζ n`1 2πnξ
where λ “ ξ “ ζ “ 1, thus (not very surprisingly, see (4.1)):
nn´1 1 en
„ ? .
n! n 2πn
E XAMPLE 4.4.18. We consider the trinomial coefficients
´ ¯n
n 2
Tn :“ Jz K 1 ` z ` z .
` ˘
In the notation of Theorem 4.4.15, we have A pzq “ 1 and B pzq “ 1 ` z ` z2 , as well
as n “ N (hence λ “ 1). By the theorem, we get
` ˘n
1 ` ζ ` ζ2
Tn „ a ,
ζ n 2πnξ
where ζ is determined by the equation
1 ` 2ζ
ζ “ λ “ 1,
1 ` ζ ` ζ2
that is, ζ “ 1, and
d2
ξ“ plog B psq ´ λ log sq|s“ζ“1
ds2ˆ ˙ˇ
d 1 ` 2s 1 ˇˇ
“ ´ Ðλ“1
ds 1 ` s ` s2 s ˇs“1
ˇ
1 ´ 2s ´ 2s2 1 ˇˇ 2
“ 2
` 2ˇ “ .
p1 ` s ` s2 q s ˇ 3
s“1
Thus, we obtain
3n`1{2
Tn „ ? .
2 πn
E XAMPLE 4.4.19 (Asymptotics for unary-binary trees). In Example 4.3.34, we had
already considered the number of unary-binary planar rooted trees with n ą 0 vertices,
ř
which we denoted by Tn . The corresponding generating function T pzq “ Tn zn
satisfies the equation ´ ¯
T pzq “ z 1 ` T pzq ` T pzq2 .
z T
In other words, T is the compositional inverse of 1`z`z 2 : 1`T`T 2 “ z.

By Lagrange inversion, we obtain

` ˘n
1 r ´1 z 1 ` z ` z2
Mn “ z
n zn
1 n ´ ¯n
2
“ Jz K z 1 ` z ` z .
n
 4.5. ASYMPTOTICS OF COMBINATORIAL SUMS 125

Completely analogously as in the preceding example, here we obtain

3n`1{2
Tn „ ? ,
2 πn3
in agreement with our finding in Example 4.3.34.

4.5. Asymptotics of combinatorial sums

The idea that we successfully applied in the saddle point method, can also be
applied to sums (instead of to integrals). We illustrate this by considering an
example.
E XAMPLE 4.5.1 (Motzkin numbers). The number of all paths in the lattice Z ˆ Z
which
‚ run from p0, 0q to pn, 0q;
‚ consist of right steps p1, 0q, diagonal up-steps p1, 1q, and diagonal down-steps
p1, ´1q;
‚ never pass below the x-axis;
is called n-th Motzkin number, and is denoted by Mn . If, in such a Motzkin path,
we delete the horizontal steps, then what remains is evidently a Dyck path that consists
of 2k ď n diagonal steps. This leads us straightforwardly to the formula
ÿ 1 ˆ2k˙ˆ n ˙
n{2
Mn “ . (4.56)
k`1 k n ´ 2k
loooooooooooomoooooooooooon
k“0
“:spn,kq

The summands s pn, kq in this sum “optically” show a similar behaviour as the function
2
e´t that strongly decreases away from zero, see the following plot for the summands
s p100, kq:
Summands for the 100-th Motzkin-Number
1.2 × 1044

1.0 × 1044

8.0 × 1043

6.0 × 1043

4.0 × 1043

2.0 × 1043

10 20 30 40 50

For the determination of the place of the “peak” of these summands, we look for the k,
for which s pn, k ` 1q {s pn, kq “ 1. This leads to the equation
k ` 1 pn ´ 2kq pn ´ 2k ´ 1q
“ 1,
k`2 pk ` 1q2
respectively equivalently
pn ´ 2kq pn ´ 2k ´ 1q “ pk ` 1q pk ` 2q .
126 4. ASYMPTOTIC ENUMERATION

Again, we are not necessarily interested in the “exact” location of the peak. With the
“approximate” Ansatz k „ λn, we are “approximatively” led to the quadratic equation
p1 ´ 2λq2 “ λ2 ,
with obvious solutions λ1 “ 1 and λ2 “ 31 . Therefore, we set k0 “ n3 , and approximate
s pn, n{3 ´ k1 q with the help of Stirling’s formula (4.1), in logarithmic form:
ˆ ˙
1 1
log n! „ ´n ` n ` log n ` plog 2πq .
2 2
We have
n!
s pn, kq “
pk ` 1q! pkq! pn ´ 2kq!
˜ˆ ˙
1
“ exp n` log n ` 1 ´ log 2π
2
ˆ ˙
1
´ n ´ 2k ` log pn ´ 2kq
2
ˆ ˙ ˆ ˙ ¸
1 3
´ k` logpkq ´ k ` logpk ` 1q .
2 2
Now we substitute k “ n{3 ´ k1 and use
´ ˆ ˙
n¯ 3z
log ´z ` “ log n ´ log 3 ` log 1 ´
3 n
(for z “ k1 and z “ k1 ´ 1) together with
z2 ´ ¯
log p1 ´ zq “ z ` ` O z3
2
to obtain (after some computation):
˜
9k21 ´ 3k1 ` 3
s pn, kq “ exp ´ 2 logpnq ` n logp3q ` 2
n
˜ ¸¸
5 logp3q k31
´ logp2πq ` `O
2 n2
5
˜ ˜ ¸¸
3n` 2 2´ 1 p9k2 ´3k1 `3q k31
“ e n 1 1 ` O .
2πn2 n2
(The expression
´ ¯ becomes so simple because many terms can be “subsumed” in the error
k31
term O n2
.) computations ok, fix this??? Consequently, we have
ÿ 3n` 25 ˆ ˆ 3 ˙˙
´ n1 p9k2 ´3k`3q k
Mn “ e 2
2
e 1`O . (4.57)
2πn n2
k
We split this sum as follows:
ÿ ÿ
Mn “ p¨ ¨ ¨ q ` p¨ ¨ ¨ q .
|k|ďn3{5 |k|ąn3{5
 4.5. ASYMPTOTICS OF COMBINATORIAL SUMS 127
´Q U ¯
For |k| ą n3{5 (i.e., k “ ˘ n3{5 ` k1 with k1 ą 0), the summand for |k| “ n3{5
dominates the other summands, provided n is large enough:
´ ¯
2 ?
´ n1 9pn3{5 `k1 q ˘3pn3{5 `k1 q`3 2
“ e´9pn nq `Opn´2{5 q
3{5´1{2 `k
e 1{ .
Consequently, we can bound that part of the sum by
ˇ ˇ
ˇ ˇ ´ ¯
ˇ ÿ ˇ
ˇ p¨ ¨ ¨ q ˇ “ O n ¨ 3n n´2 e´9n1{5 .
ˇ ˇ
ˇ|k|ąn3{5 ˇ

This is exponentially small in comparison with 3n {n3{2 . The other part of the sum is
ÿ 3n` 25 2´9´ ?k ¯2 ´ ´ ¯¯
e n ¨ 1`O n ´1{5
.
3{5
2πn2
|k|ďn

If we extend the domain of summation of this sum to all of Z, then we introduce an

error, which however is exponentially small. Thus we arrive at
´ ¯2
3n`5{2 ? ÿ 1 ´9 ?kn
n ? e ,
2πn2 n
kPZ
ş8 2 ?
and the sum can be identified as Riemann sum for ´8 e´9t dt “ 31 π. Hence, we
get
n`5{2 1 ? n`3{2
23 2 3?
Mn „ e π“e .
2πn3{2 3 2 πn3
Exercise 38: Let wn be the number of possibilities of paying an amount of n Euro using 1–Euro–
coins, 2–Euro–coins and 5–Euro–notes (the order of the coins and notes is irrelevant).
ř
(1) Determine the generating function ně0 wn zn .
(2) Determine the asymptotic behaviour of wn for n Ñ 8.

Exercise 39: Let Dn,k be the number of permutations of rns, whose disjoint cycle decomposition
does not contain any cycle of length ď k. (So Dn,1 is the number of fixed–point–free permutations
of rns.)
(1) Show:
z 2 z k
ÿ Dn,k e´z´ 2 ´¨¨¨´ k
zn “ .
n! 1´z
n ě0
(2) For k fixed, what is the asymptotic behaviour of Dn,k for n Ñ 8?

Exercise 40: Let wn,k be the number of possibilities of paying an amount of n Euro using 1–Euro–
coins, 2–Euro–coins and 5–Euro–notes, where exactly k coins or notes are used (again, the order of
the coins and notes is irrelevant).
ř
(1) Determine the generating function n,kě0 wn,k zn tk .
ř
(2) If we assume that all possibilities which are enumerated by wn “ k wn,k have the same
probability: What is the asymptotic behaviour of the expected value for the number of
coins and notes which are used to pay an amount of n Euro, for n Ñ 8?
128 4. ASYMPTOTIC ENUMERATION

Exercise 41: Let Rn be the number of possibilities of (completely) tiling a 2 ˆ n rectangle by 1 ˆ 1

squares and 1 ˆ 2 rectangles (dominoes).
ř
(1) Determine the generating function ně0 Rn zn .
(2) Determine the asymptotic behaviour of Rn for n Ñ 8?

Hint: The fact that the dominant singularity can not be calculated explicitly is not an obstacle. One
has to continue to calculate with the dominant singularity symbolically.

Exercise 42: Let p pn, kq be the number of all (integer) partitions of n with at most k summands.
Show:
ÿ 1
p pn, kq zn “ .
n ě0
p1 ´ zqp1 ´ z2 q ¨ ¨ ¨ p1 ´ zk q
Determine the asymptotic behaviour of ppn, kq for fixed k and n Ñ 8.

Exercise 43: The exponential generating function of the Bernoulli numbers bn is

ÿ z
bn z n “ z .
n ě0
e ´1
Determine the asymptotic behaviour of bn for Ñ 8.

1
Exercise 44: The fraction Γ pz q
is an entire function with zeroes 0, ´1, ´2, . . . . Show Weierstraß’
product representation:
8 ´
1 ź s ¯ ´s{n
“ seγs 1` e ,
Γ psq n
n “1
where γ denotes Eulers constant
γ “ lim pHn ´ log nq
nÑ8
and
n
ÿ 1
Hn “
j
j “1
denotes the n–th harmonic number.

Exercise 45: Show the reflection formula for the gamma function:
π
Γ pzq Γ p1 ´ zq “ . (4.58)
sin πz
Hint: Use the product representation of the sine:
ź8 ˆ ˙
z2
sin z “ z 1´ 2 2 . (4.59)
n “1
n π

Exercise 46: Show the duplication formula for the gamma function:
ˆ ˙
1 1
Γ pzq Γ z ` “ 21´2z π 2 Γ p2zq
2
and its generalisation
źm ˆ ˙
j 1 m´1
Γ z` “ m 2 ´mz p2πq 2 Γ pmzq .
m
j “0
 4.5. ASYMPTOTICS OF COMBINATORIAL SUMS 129

Exercise 47: Show using Stirling’s Formula for the Γ–Function

ˆ ˙ ˆ ˆ ˙˙
n`α´1 n α ´1 1
“ 1`O
n Γ pαq n

Hint: Here is Stirling’s Formula:

ˆ ˆ ˙˙
? 1
Γ ps ` 1q „ s es ´s
2πs 1 ` O .
s

Exercise 48: A Motzkin path is a lattice path, where every step is of the form p1, 0q, p1, 1q, p1, 1q
(horizontal, up– and down–steps), which starts at the origin, returns to the x–axis and never goes
below the x–axis. Let Mn be the number of all Motzkin paths of length (i.e., number of steps) n.
Show that the generating function for Motzkin paths is given by
?
ÿ 1 ´ z ´ 1 ´ 2z ´ 3z2
n
Mn z “ .
n ě0
2z2
Derive an explizit formula for Mn . Use the generating function to determine the asymptotic behaviour
of Mn for n Ñ 8.

Exercise 49: A Schröder path is a lattice path consisting of steps p2, 0q, p1, 1q and p1, ´1q (i.e.,
double horizontal, upward and downward steps) which starts at the origin, returns to the x–axis
but never falls below the x–axis. If we assume that all Schröder paths of length n have the same
probability: What is the asymptotics of the expected value of the number of steps for a Schröder
path of length n for n Ñ 8?

Exercise 50: Consider the number of cycles in the disjoint cycle decomposition of permutations of
rns on average: What is the asymptotics for this average for n Ñ 8?

řn
Exercise 51: Let Hn “ j “1 j´1 be the n–th harmonic number. Show that
ÿ 1 1
Hn z n “ log ,
n ě0
1´z 1´z
and use this result together with singularity analysis to obtain an asymptotic expansion of Hn for
n Ñ 8.

Exercise 52: Let un be the number of permutations of rns, which only have cycles of odd length in
their decomposition into disjoint cycles. Determine the asymptotic behaviour of un for n Ñ 8.
Hint: Observe that the generating function is analytic in a “Double–Delta–Domain” (i.e., in a disk
with two “dents” at the two singularities), so we have two contributions according to the Transfer
Theorem.

Exercise 53: Determine the asymptotic behaviour of the sum

n ˆ ˙ˆ ˙
ÿ n 2k
fn “
k k
k “0
for n Ñ 8.
130 4. ASYMPTOTIC ENUMERATION

Hint: Compute the generating function of these sums, i.e., multiply the above expression by zn and
sum over all n ě 0; apply the binomial theorem for simplifying the double sum thus obtained.

Exercise 54: Denote by In the number of all involutions (an involution is a self–inverse bijection) on
rns.
(1) Show
ÿ zn z2
In “ ez` 2 .
n!
n ě0
(2) Use the saddle point method to determine the asymptotic behaviour of In for n Ñ 8.

Exercise 55: The exponential generating function of the Bell–numbers Bn (Bn is the number of all
partitions of rns) is
ÿ zn
“ ee ´ 1 .
z
Bn
n ě0
n!
Use the saddle point method to determine the asymptotic behaviour of Bn for n Ñ 8.

Exercise 56: Determine the asymptotic behaviour of the sum

n ˆ ˙
ÿ n p´1qk
k k!
k “0
for n Ñ 8.
Hint: Determine the generating function for this sum!

Exercise 57: The saddle point method can also be used for the asymptotics of the Motzkin numbers
Mn (see exercise 48) for n Ñ 8:
Show r z ´ ¯n r z ´ ¯n
Mn “ z 0 z ` 1 ` z ´1 ´ z 2 z ` 1 ` z ´1
and obtain a complex contour integral for Mn , which can be dealt with using the saddle point method.

4.6. Exp–log scheme

As another application, we mention the following (without proof):
D EFINITION 4.6.1. A function G pzq which is analytic at 0, has only non–negative
coefficients and finite radius of convergence ρ, is said to be of logarithmic type with
parameters pκ, λq, where κ, λ P R, κ ‰ 0, if the following conditions hold:
(1) the number ρ is the unique singularity of G pzq on |z| “ ρ,
(2) G pzq is continuable to a ∆–domain at ρ,
(3) G pzq satisfies
˜ ¸
1 1
G pzq “ κ ¨ log `λ`O as z Ñ ρ in ∆. (4.60)
1´z plog p1 ´ z{ρqq2
 4.6. EXP–LOG SCHEME 131

D EFINITION 4.6.2. The labelled construction

F “ Sets pG q
is called a (labelled) exp–log–scheme if the exponential generating function G pzq of G
is of logarithmic type.
The unlabelled construction
F “ Multisets pG q
is called an (unlabelled) exp–log–scheme if the ordinary generating function G pzq of G
is of logarithmic type, with ρ ă 1.
In both cases (labelled and unlabelled), the quantities pκ, λq from (4.60) are called the
parameters of the scheme.
T HEOREM 4.6.3 (Exp–log scheme). Consider an exp–log scheme with parameters
pκ, λq.
Then we have
κ ´ ´ ¯¯
´2
Jzn K G pzq “ ¨ 1 ` O plog nq ,
n ¨ ρn
n eλ`r0 κ´1 ´n ´ ´
´2
¯¯
Jz K F pzq “ ¨n ¨ ρ ¨ 1 ` O plog nq ,
Γ pκq
ř G pρ j q
where r0 “ 0 in the labelled case and r0 “ jě2 j in the unlabelled case.
If we consider the number X of G –components in a (randomly chosen) F –object of size
n, then the expected value of X is
´ ¯ d
κ ¨ plog n ´ Ψ pκqq ` λ ` r1 ` O plog nq´1 (where Ψ psq “ Γ psq ),
ds
ř ` j˘
where r1 “ 0 in the labelled case and r0 “ jě2 G ρ in the unlabelled case. The
variance of X is O plog nq.
Exercise 58: Determine the asymptotic behaviour of expected value and variance of the number of
connected components of 2–regular labelled graphs with n vertices for n Ñ 8.
 Bibliography

[1] F. Bergeron, G. Labelle, and P. Leroux. Combinatorial Species and tree–like Structures, vol-
ume 67 of Encyclopedia of Mathematics and its Applications. Cambridge University Press, 1998.
[2] A. Erdélyi, editor. Higher Transcendental Functions, volume 1. McGraw–Hill, 1953.
[3] P. Flajolet and R. Sedgewick. Analytic Combinatorics. Cambridge University Press, 2009.
[4] M. Fulmek and C. Krattenthaler. Diskrete Mathematik. Vorlesungsskriptum.
[5] H. Heuser. Lehrbuch der Analysis. B.T. Teubner, 12. edition, 1998.
[6] G. C. Rota and R. Stanley. Theory of Möbiusfunction. Z. Wahrsch. Theorie, 2:340–368, 1964.
[7] R. Stanley. Enumerative Combinatorics, volume 1. Wadsworth & Brooks/Cole, 1986.

133