Lecture Notes Combinatorics

Lecture Notes
Combinatorics
Lecture by Maria Axenovich and Torsten Ueckerdt (KIT)
Problem Classes by Jonathan Rollin (KIT)
Lecture Notes by Stefan Walzer (TU Ilmenau)

Last updated: May 30, 2017
1
Contents
0 What is Combinatorics? 4
1 Permutations and Combinations 9

1.1 Basic Counting Principles . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 Addition Principle . . . . . . . . . . . . . . . . . . . . . . 10
1.1.2 Multiplication Principle . . . . . . . . . . . . . . . . . . . 10
1.1.3 Subtraction Principle . . . . . . . . . . . . . . . . . . . . 11
1.1.4 Bijection Principle . . . . . . . . . . . . . . . . . . . . . . 11
1.1.5 Pigeonhole Principle . . . . . . . . . . . . . . . . . . . . . 11
1.1.6 Double counting . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Ordered Arrangements – Strings, Maps and Products . . . . . . 12
1.2.1 Permutations . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Unordered Arrangements –
Combinations, Subsets and Multisets . . . . . . . . . . . . . . . . 15
1.4 Multinomial Coefficients . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 The Twelvefold Way – Balls in Boxes . . . . . . . . . . . . . . . 22
1.5.1 U → L: n Unlabeled Balls in k Labeled Boxes . . . . . . 22
1.5.2 L → U: n Labeled Balls in k Unlabeled Boxes . . . . . . 25
1.5.3 L → L: n Labeled Balls in k Labeled Boxes . . . . . . . 27
1.5.4 U → U: n Unlabeled Balls in k Unlabeled Boxes . . . . 28
1.5.5 Summary: The Twelvefold Way . . . . . . . . . . . . . . . 30
1.6 Binomial Coefficients – Examples and Identities . . . . . . . . . . 30
1.7 Permutations of Sets . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.7.1 Generating Permutations . . . . . . . . . . . . . . . . . . 36
1.7.2 Cycle Decompositions . . . . . . . . . . . . . . . . . . . . 37
1.7.3 Transpositions . . . . . . . . . . . . . . . . . . . . . . . . 39
1.7.4 Derangements . . . . . . . . . . . . . . . . . . . . . . . . . 42
2 Inclusion-Exclusion-Principle
and Möbius Inversion 45
2.1 The Inclusion-Exclusion Principle . . . . . . . . . . . . . . . . . . 45
2.1.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.1.2 Stronger Version of PIE . . . . . . . . . . . . . . . . . . . 52
2.2 Möbius Inversion Formula . . . . . . . . . . . . . . . . . . . . . . 53
3 Generating Functions 58
3.1 Newton’s Binomial Theorem . . . . . . . . . . . . . . . . . . . . . 62
3.2 Exponential Generating Functions . . . . . . . . . . . . . . . . . 64
3.3 Recurrence Relations . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.3.1 Special Solution an = xn and the Characteristic Polynomial 70
3.3.2 Advancement Operator . . . . . . . . . . . . . . . . . . . 76
3.3.3 Non-homogeneous Recurrences . . . . . . . . . . . . . . . 80
3.3.4 Solving Recurrences using Generating Functions . . . . . 82
2
4 Partitions 85
4.1 Partitioning [n] – the set on n elements . . . . . . . . . . . . . . . 85
4.1.1 Non-Crossing Partitions . . . . . . . . . . . . . . . . . . . 86
4.2 Partitioning n – the natural number . . . . . . . . . . . . . . . . 87
4.3 Young Tableau . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3.1 Counting Tableaux . . . . . . . . . . . . . . . . . . . . . . 99
4.3.2 Counting Tableaux of the Same Shape . . . . . . . . . . . 100
5 Partially Ordered Sets 107

5.1 Subposets, Extensions and Dimension . . . . . . . . . . . . . . . 111
5.2 Capturing Posets between two Lines . . . . . . . . . . . . . . . . 117
5.3 Sets of Sets and Multisets – Lattices . . . . . . . . . . . . . . . . 123
5.3.1 Symmetric Chain Partition . . . . . . . . . . . . . . . . . 126
5.4 General Lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6 Designs 132
6.1 (Non-)Existence of Designs . . . . . . . . . . . . . . . . . . . . . 133
6.2 Construction of Designs . . . . . . . . . . . . . . . . . . . . . . . 135
6.3 Projective Planes . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6.4 Steiner Triple Systems . . . . . . . . . . . . . . . . . . . . . . . . 138
6.5 Resolvable Designs . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.6 Latin Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
3
What is Combinatorics?
Combinatorics is a young field of mathematics, starting to be an independent
branch only in the 20th century. However, combinatorial methods and problems
have been around ever since. Many combinatorial problems look entertaining
or aesthetically pleasing and indeed one can say that roots of combinatorics lie
in mathematical recreations and games. Nonetheless, this field has grown to be
of great importance in today’s world, not only because of its use for other fields
like physical sciences, social sciences, biological sciences, information theory and
computer science.
Combinatorics is concerned with:

• Arrangements of elements in a set into patterns satisfying specific rules,
generally referred to as discrete structures. Here “discrete” (as opposed
to continuous) typically also means finite, although we will consider some
infinite structures as well.
• The existence, enumeration, analysis and optimization of discrete struc-

tures.
• Interconnections, generalizations- and specialization-relations between sev-
eral discrete structures.
Existence: We want to arrange elements in a set into patterns satisfying

certain rules. Is this possible? Under which conditions is it possible?
What are necessary, what sufficient conditions? How do we find such
an arrangement?
Enumeration: Assume certain arrangements are possible. How many

such arrangements exist? Can we say “there are at least this many”,
“at most this many” or “exactly this many”? How do we generate all
arrangements efficiently?
Classification: Assume there are many arrangements. Do some of these

arrangements differ from others in a particular way? Is there a natural
partition of all arrangements into specific classes?
Meta-Structure: Do the arrangements even carry a natural underlying
structure, e.g., some ordering? When are two arrangements closer to
each other or more similar than some other pair of arrangements? Are
different classes of arrangements in a particular relation?
Optimization: Assume some arrangements differ from others according
to some measurement. Can we find or characterize the arrangements
with maximum or minimum measure, i.e. the “best” or “worst” ar-
rangements?
4
Interconnections: Assume a discrete structure has some properties (num-
ber of arrangements, . . . ) that match with another discrete structure.
Can we specify a concrete connection between these structures? If
this other structure is well-known, can we draw conclusions about our
structure at hand?
We will give some life to this abstract list of tasks in the context of the
following example.
Example (Dimer Problem). Consider a generalized chessboard of size m×n (m
rows and n columns). We want to cover it perfectly with dominoes of size 2 × 1
or with generalized dominoes – called polyominoes – of size k × 1. That means
we want to put dominoes (or polyominoes) horizontally or vertically onto the
board such that every square of the board is covered and no two dominoes (or
polyominoes) overlap. A perfect covering is also called tiling. Consider Figure
1 for an example.
Figure 1: The 6 × 8 board can be tiled with 24 dominoes. The 5 × 5 board

cannot be tiled with dominoes.
Existence
If you look at Figure 1, you may notice that whenever m and n are both odd
(in the Figure they were both 5), then the board has an odd number of squares
and a tiling with dominoes is not possible. If, on the other hand, m is even or
n is even, a tiling can easily be found. We will generalize this observation for
polyominoes:
Claim. An m × n board can be tiled with polyominoes of size 1 × k if and only
if k divides m or n.
Proof. “⇐” If k divides m, it is easy to construct a tiling: Just cover every
column with m/k vertical polyominoes. Similarly, if k divides n, cover
every row using n/k horizontal polyominoes.
“⇒” Assume k divides neither m nor n (but note that k could still divide
the product m · n). We need to show that no tiling is possible. We
write m = s1 k + r1 , n = s2 k + r2 for appropriate s1 , s2 , r1 , r2 ∈ N and
0 < r1 , r2 < k. Without loss of generality, assume r1 ≤ r2 (the argument
is similar if r2 < r1 ). Consider the colouring of the m × n board with k
colours as shown in Figure 2.
5
1 2 3 4 5 6 7 8 9
1
2
3
4
k 5
k 6
7
8
Figure 2: Our polyominoes have size k × 1. We use k colours (1 = white, k

= black) to colour the m × n board (here: k = 6, m = 8, n = 9). Cutting
the board at coordinates that are multiples of k divides the board into several
chunks. All chunks have the same number of squares of each color, except for
the bottom right chunk where there are more squares of color 1 (here: white)
than of color 2 (here: light gray).
Formally, the colour of the square (i, j) is defined to be ((i−j) mod k)+1.
Any polyomino of size k × 1 that is placed on the board will cover exactly
one square of each colour. However, there are more squares of colour 1
than of colour 2, which shows that no tiling with k×1 dominoes is possible.
Indeed, for the number of squares coloured with 1 and 2 we have:
# squares coloured with 1 = ks1 s2 + s1 r2 + s2 r1 + r2

# squares coloured with 2 = ks1 s2 + s1 r2 + s2 r1 + r2 − 1
Now that the existence of tilings is answered for rectangular boards, we may
be inclined to consider other types of boards as well:
Claim (Mutilated Chessboard). The n×n board with bottom-left and top-right
square removed (see Figure 3) cannot be tiled with (regular) dominoes.
Figure 3: A “mutilated” 6 × 6 board. The missing corners have the same colour.
Proof. If n is odd, then the total number of squares is odd and clearly no tiling
can exist. If n is even, consider the usual chessboard-colouring: In it, the missing
squares are of the same colour, say black. Since there was an equal number of
black and white squares in the non-mutilated board, there are now two more
white squares than black squares. Since dominoes always cover exactly one
black and one white square, no tiling can exist.
6
Other ways of pruning the board have been studied, but we will not consider
them here.
Enumeration
A general formula to determine the number of ways an m × n board can be tiled
with dominoes is known. For an 2m × 2n board the following formula is due to
Temperly and Fisher [TF61] and independently Kasteleyn [Kas61]
n
m Y
Y
jπ
4mn cos2 iπ
2m+1 + cos2 2n+1 .
i=1 j=1
The special case of an 8 × 8 board is already non-trivial:
Theorem (Fischer 1961). There are 24 · 172 · 532 = 12, 988, 816 ways to tile the
8 × 8 board with dominoes.
Classification
Consider tilings of the 4 × 4 board with dominoes. For some of these tilings
there is a vertical line through the board that does not cut through any domino.
Call such a line a vertical cut. In the same way we define horizontal cuts.
As it turns out, for every tiling of the 4 × 4 board at least one cut exists,
possibly several (try this for yourself!).
Hence the set T of all tilings can be partitioned into
T1 = {T | T allows a horizontal cut but no vertical cut},

T2 = {T | T allows a vertical cut but no horizontal cut},
T3 = {T | T allows both a horizontal and a vertical cut}.
Figure 4 shows one tiling for each of these three classes.
Figure 4: Some tilings have horizontal cuts, some have vertical cuts and some
have both.
As another example consider the board B consisting of two 4 × 4 boards

joint by two extra squares as shown in Figure 5.
In a tiling of B it is impossible that one extra square is covered by a domino
containing a square from the left side, while the other extra square is covered by
a domino containing a square from the right side: If we place the two dominoes
as shown on the right in Figure 5, we are left with two disconnected parts of
size 15. And since 15 is odd, this cannot work out. Hence the set T of all tilings
7
Figure 5: The board B consisting of two 4 × 4 boards and two extra squares
connecting them as shown. The partial covering on the right cannot be extended
to a tiling.
.
of B can be partitioned into
T1 = {T | T matches both extra squares to the left}

T2 = {T | T matches both extra squares to the right}.
Meta Structure
We say two tilings are adjacent, if one can be transformed into the other by
taking two dominoes lying like this and turning them by 90 degrees so they
are lying like this (or vice versa). Call this a turn operation. If we draw all
tilings of the 4 × 4 board and then draw a line between tilings that are adjacent,
we get the picture on the left of Figure 6.
With this in mind, we can speak about the distance of two tilings, the number
of turn operations required to transform one into the other. We could also look
for a pair of tilings with maximum distance (this would be an optimization
problem).
We can even discover a deeper structure, but for this we need to identify
different types of turn operations. We can turn into , call this an up-turn,
or into , call this a side-turn. The turn can happen on the background
or the background (white squares form a falling or rising diagonal). We call
an operation a flip if it is an up-turn on or a side-turn on . Call it a flop
otherwise.
Quite surprisingly, walking upward in the graph in Figure 6 always corre-
sponds to flips and walking downwards corresponds to flops. This means that
there is a natural partial ordering on the tilings: We can say a tiling B is greater
than a tiling A if we can get from A to B by a sequence of flips. As it turns out,
the order this gives has special properties. It is a so-called distributive lattice,
in particular, there is a greatest tiling.
Optimization
Different tilings have a different set of decreasing free paths. Such a path (see
Figure 7) proceeds monotonously from the top left corner of the board along
the borders of squares to the bottom right, and does not “cut through” any
domino.
Some tilings admit more such paths then others. It is conjectured that in an
m × n board the maximum number is always attained by one of the two tilings
where all dominoes are oriented the same way (see right of Figure 7). But no
proof has been found yet.
8
Figure 6: On the left: The set of all tilings of the 4 × 4 board. Two tilings
are connected by an edge if they can be transformed into one another by a
single flip. On the right: The same picture, but with the chessboard still drawn
beneath the tilings, so you can check that upward edges correspond to flips (as
defined in the text).
1 Permutations and Combinations

Some people mockingly say that combinatorics is merely about counting things.
In the section at hand we will do little to dispel this prejudice.
We assume the reader is familiar with basic set theory and notions such as
unions, intersections, Cartesian products and differences of two finite sets.
1.1 Basic Counting Principles

We list a few very simple and intuitive observations. Many of them where
already used in the introduction. You may think that they are so easy that
they do not even deserve a name. Still: If we take counting seriously, we should
be aware of simple facts we implicitly use all the time.
9
Figure 7: The green path does not cut through dominoes. We believe that the
boring tiling on the right admits the maximum number of such paths, but this
has not been proved for large boards.
1.1.1 Addition Principle

We say a finite set S is partitioned into parts S1 , . . . , Sk if the parts are disjoint
and their union is S. In other words Si ∩Sj = ∅ for i 6= j and S1 ∪S2 ∪· · ·∪Sk = S.
The addition principle claims that in this case
|S| = |S1 | + |S2 | + · · · + |Sk |.
Example. Let S be the set of students attending the combinatorics lecture. It
can be partitioned into parts S1 and S2 where
S1 = set of students that like easy examples.
S2 = set of students that don’t like easy examples.
If |S1 | = 22 and |S2 | = 8 then we can conclude |S| = 30.
1.1.2 Multiplication Principle

If S is a finite set S that is the product of S1 , . . . , Sk , i.e. S = S1 × S2 . . . × Sk ,
then the multiplication principle claims that
|S| = |S1 | × |S2 | × · · · × |Sm |.
Example. Consider the license plates in Karlsruhe. They have the form
KA - l1 l2 n1 n2 n3
where l1 , l2 are letters from {A, . . . , Z} and n1 , n2 , n3 are digits from {0, . . . , 9}.
However, l2 and n3 may be omitted.
We count the number of possibilities for each position:
K A - l1 l2 n1 n2 n3
↑ ↑ ↑ ↑ ↑ ↑ ↑ ↑
1 1 1 26 27 10 10 11
Here we modeled the case that l2 is omitted by allowing an additional value ⊥
for l2 that stands for “omitted”. The same holds for n3 .
Formally we could say that the set of all valid license plates is given as the
product:
{K}×{A}×{-}×{A, . . . , Z}×{A, . . . , Z, ⊥}×{0, . . . , 9}×{0, . . . , 9}×{0, . . . , 9, ⊥}.
By the Multiplication Principle the size of this set is given as 1 · 1 · 1 · 26 · 27 ·
10 · 10 · 11 = 772200.
10
1.1.3 Subtraction Principle
Let S be a subset of a finite set T . We define S := T \ S, the complement of S
in T . Then the substraction principle claims that
|S| = |T | − |S|.
Example. If T is the set of students studying at KIT and S the set of students
studying neither math nor computer science. If we know |T | = 23905 and
|S| = 20178, then we can compute the number |S| of students studying either
math or computer science:
|S| = |T | − |S| = 23905 − 20178 = 3727.
1.1.4 Bijection Principle

If S and T are finite sets, then the bijection principle claims that
|S| = |T | ⇔ there exists a bijection between S and T .
Example. Let S be the set of students attending the lecture and T the set of
homework submissions for the first problem sheet.
If the number of students and the number of submissions coincide, then there
is a a bijection between students and submissions1 and vice versa.
We now consider two other principles that are similarly intuitive and natural
but will frequently and explicitly occur as patterns in proofs.
1.1.5 Pigeonhole Principle

Let S1 , . . . , Sm be finite sets that are pairwise disjoint and |S1 | + |S2 | + . . . +
|Sm | = n. Then the pigeonhole principle claims that
lnm jnk
∃i ∈ {1, . . . , m} : |Si | ≥ and ∃j ∈ {1, . . . , m} : |Sj | ≤ .
m m
Example. Assume there are 5 holes in the wall where pigeons nest. Say there
is a set Si of pigeons nesting in hole i. Assume there are n = 17 pigeons in
total. Then we know:
• There is some hole with at least d17/5e = 4 pigeons.
• There is some hole with at most b17/5c = 3 pigeons.
1.1.6 Double counting

If we count the same quantity in two different ways, then this gives us a (perhaps
non-trivial) identity. This method often helps get control on overcounting. Let
A and B be finite sets and consider a set S ⊆ A×B. Then the method of double
counting claims that
X X
|{b ∈ B | (a, b) ∈ S}| = |S| = |{a ∈ A | (a, b) ∈ S}|.
a∈A b∈B
1 Whether a “natural” correspondence can easily be determined is an entirely different
matter: That would require that you cleanly write your name onto your submission.
11
Example (Handshaking Lemma). Assume there are n people at a party and
everybody will shake hands with everybody else. What is the total number N of
handshakes that occur? If we just sum up for each guest each of its handshakes,
then we overcount each handshake. In order to control this overcount we shall
instead count the number M of pairs (P, S) where P is a guest and S is a
handshake involving P . We count this number in two ways:
First way: There are n guests and everybody shakes n − 1 hands. So
X X
M= |{(P, S) | S handshake with P }| = (n − 1) = n · (n − 1).
P guest P guest
Second way: Every handshake involves two guests. So

X X
M= |{(P, S) | P involved in S}| = 2 = 2 · N.
S handshake S handshake
Together, we see that n(n − 1) = M = 2N and hence we obtain the identity
n · (n − 1)
N= .
2
Example. In the example above we have seen a first way how two count the
number N (already utilizing double counting). Here we count N in a second
way. Label the guests from 1 to n. To avoid counting a handshake twice, we
count for guest i only the handshakes with guests of smaller numbers. Then the
total number of handshakes is
n
X n−1
X n−1
X
(i − 1) = i= i.
i=1 i=0 i=1
Hence we obtain another identity

n−1
X n · (n − 1)
i=N = .
i=1
2
1.2 Ordered Arrangements – Strings, Maps and Products

Define [n] := {1, . . . , n} to be the set of the first n natural number and let X be
a finite set.
An ordered arrangement of n elements of X is a map s : [n] → X. Ordered
arrangements are so common that they occur in many situations and are known
under different names. We might be inclined to say that “Banana” is a string (or
word) and that “0815422372” is a sequence even though they are both ordered
arrangements of characters. Depending on the circumstances different notation
may be in order as well.
We will introduce the most common perspectives on what an ordered ar-
rangement is and the accompanying names and notation. Throughout, we use
the example X = { , , , , } and the ordered arrangement s of n = 7 ele-
ments given as the table:
12
i 1 2 3 4 5 6 7
s(i)
function: We call [n] the domain of s and s(i) the image of i under s (i ∈ [n]).
The set {x ∈ X | s(i) = x for some i ∈ [n]} is the range of s.
In the example, the domain of s is {1, 2, 3, 4, 5, 6, 7}, the image of 3 is
and the range of s is { , , , }. Note that is not in the range of s.
string: We call s an X-string of length n and write s = s(1)s(2)s(3) · · · s(n).
The i-th position (or character ) in s is denoted by si = s(i). The set X
is an alphabet and its elements are letters. Often s is called a word.
In the example, we would say that s = is a string (or word)
of length n over the five-letter alphabet X. The fourth character of s is
s4 = .
tuple: We can view s as an element of the n-fold Cartesian product X1 ×
. . . × Xn , where Xi = X for i ∈ [n]. We call s a tuple and write it
as (s1 , s2 , . . . , sn ). The element si is called the i-th coordinate (i ∈ [n]).
Viewing arrangements as elements of products makes it easy to restrict the
number of allowed values for a particular coordinate (just choose Xi ( X).
In the example we would write s = ( , , , , , , ). Its first coordinate
is .
sequence: If X is a set of numbers, then s is often called a sequence and

si is the i-th position, term or element in the sequence. We would not
use this terminology in the example above but could have, for instance,
s(i) := 3i + 2/5. In that case the fourth term is equal to 12.4.
In the following we will mostly view arrangements as functions but will freely
switch perspective when appropriate.
1.2.1 Permutations
The most important ordered arrangements are those in which the mapping is
injective, i.e., s(i) 6= s(j) for i 6= j.
Definition 1.1 (Permutation). Let X be a finite set. We define
permutation: A permutation of X is a bijective map π : [n] → X. Usually we
choose X = [n] and denote the set of all permutations of [n] by Sn .
We tend to write permutations as strings if n < 10, take for example
π = 2713546 by which we mean the function:
i 1 2 3 4 5 6 7
π(i) 2 7 1 3 5 4 6
k-Permutation: For 1 ≤ k ≤ |X| a k-permutation of X is an ordered arrange-

ment of k distinct elements of X, i.e. an injective map π : [k] → X. The
set of all k-permutations of X = [n] is denoted by P (n, k). In particular
we have Sn = P (n, n).
13
Circular k-Permutation: We say that two k-permutations π1 , π2 ∈ P (n, k)
are circular equivalents if there exists a shift s ∈ [k] such that the following
implication holds:
i+s≡j (mod k) ⇒ π1 (i) = π2 (j)
This equivalence relation partitions P (n, k) into equivalence classes. A

class is called a circular k-permutation. The set of all circular k-permutations
is denoted by Pc (n, k).
Take for example π1 = 76123, π2 = 12376 and π3 = 32167. Then π1 and π2
6 2 2
1 3 1
π1 7 π2 1 π3 3
2 7 6
3 6 7
Figure 8: Visualization of the 5-permutations π1 , π2 and π3 where values are

arranged in a circular fashion (the first value being on the right of the cycle).
Note that π1 and π2 can be transformed into one another by “turning”. They
are therefore circular equivalents. Flipping is not allowed however, so π3 is not
equivalent to π1 and π2 .
are circular equivalents as witnessed by the shift s = 3. They are therefore

representatives of the same circular 5-permutation (with elements from
[7]). π3 belongs to a different circular 5-permutation. Consider Figure 8
for a visualization.
We now count the number of (circular) k-permutations. For this we need
the notation of factorials: n! := 1 · 2 · · · · · n.
Theorem 1.2. For any natural numbers 1 ≤ k ≤ n we have
n!
(i) |P (n, k)| = n · (n − 1) · . . . · (n − k + 1) = (n−k)! .
n!
(ii) |Pc (n, k)| = k·(n−k)! .
Proof. (i) A permutation is an injective map π : [k] → [n]. We count the

number of ways to pick such a map, picking the images one after the
other. There are n ways to choose π(1). Given a value for π(1), there are
n − 1 ways to choose π(2) (we may not choose π(1) again). Continuing
like this, there are n − i + 1 ways to pick π(i) and the last value we pick is
π(k) with n − k + 1 possibilities. Every k-permutation can be constructed
like this in exactly one way.
The total number of k-permutations is therefore given as the product:
n!
|P (n, k)| = n · (n − 1) · . . . · (n − k + 1) = .
(n − k)!
(ii) We count P (n, k) in two ways:
14
n!
First way: |P (n, k)| = (n−k)! which we proved in (i).
Second way: |P (n, k)| = |Pc (n, k)| · k because every equivalence class in
Pc (n, k) contains k permutations from P (n, k) (since there are k ways
to rotate a k-permutation).
n!
From this we get (n−k)! = |Pc (n, k)| · k which implies the claim.
1.3 Unordered Arrangements –

Combinations, Subsets and Multisets
Let X be a finite set and k ∈ N. An unordered arrangement of k elements of X
is a multiset S of size k with elements from X.
Take for example X = { , , , , }, then an unordered arrangement of 7
elements could be S = { , , , , , , }. Order in sets and multisets does not
matter, so we could write the same multiset as S = { , , , , , , }.
However, we prefer the following notation, where every type x ∈ X occurring
in S is given only once and accompanied by its repetition number rx that cap-
tures how often the type occurs in the multiset. The example above is written
as: S = {2 · , 1 · , 3 · , 1 · } and we would say the type has repetition
number 2, the type has repetition number 1 and so on. We write ∈ S,
r = 2, ∈ S, r = 1 and so on.
The difference between ordered and unordered arrangements is that ordered

arrangements are selections of elements of X that are done one at a time, while
unordered arrangements are selections of objects done all at the same time.
A typical example for unordered arrangements would be shopping lists: You
may need three bananas and two pears, but this is the same as needing a pear,
three bananas and another pear. In a sense, you need everything at once with
no temporal ordering. This is in contrast to something like a telephone number
(which is an ordered arrangement) where dialing a 5 first and a 9 later is differ-
ent from dialing 9 first and 5 later.
As with ordered arrangements, the most important case for unordered ar-
rangements is that all repetition numbers are 1, i.e. rx = 1 for all x ∈ S. Then
S is simply a subset of X, denoted by S ⊆ X.
Definition 1.3. k-Combination: Let X be a finite set. A k-combination of
X is an unordered arrangement of k distinct elements from X. We prefer
the more standard term subset and use “combination” only when we want
to emphasize
the selection process. The set of all k-subsets of X is denoted
by Xk and if |X| = n then we denote

n X
:= .
k k
k-Combination of a Multiset: Let X be a finite set of types and let M be

a finite multiset with types in X and repetition numbers r1 , . . . , r|X| . A
k-combination of M is a multiset with types in X and repetition numbers
P|X|
s1 , . . . , s|X| such that si ≤ ri , 1 ≤ i ≤ |X|, and i=1 si = k.
If for example M = {2 · , 1 · , 3 · , 1 · }, then T = {1 · , 2 · } is a
3-combination of the multiset M , but T 0 = {3 · , 0 · } is not.
15
k-Permutation of a Multiset: Let M be a finite multiset with set of types X.
A k-permutation of M is an ordered arrangement of k elements of M where
different orderings of elements of the same type are not distinguished.
This is a an ordered multiset with types in X and repetition numbers
P|X|
s1 , . . . , s|X| such that si ≤ ri , 1 ≤ i ≤ |X|, and i=1 si = k.
Note that there might be several elements of the same type compared
to a permutation of a set (where each repetition number equals 1). If for
example M = {2· , 1· , 3· , 1· }, then T = ( , , , ) is a 4-permutation
of the multiset M .
Theorem 1.4. For 0 ≤ k ≤ n, we have

n n n!
P (n, k) = · k! and therefore = .
k k k!(n − k)!
Proof. To build an element

from P (n, k), wefirst choose k elements from [n],
by our definition of nk , there are exactly nk ways to do so. Then choose an
order of the k elements, there are |P (k, k)| = k!
0! = k! ways to do so.
Every element from P (n, k) can be constructed like this in exactly one way
so |P (n, k)| = nk · k! which proves the claim.
Using Theorem 1.2(i) now gives the identity on the right as well.

The numbers nk for 0 ≤ k ≤ n are called binomial coefficients because these
are the coefficients when expanding the n-th power of a sum of two variables:
n
X n k n−k
(x + y)n = x y .
k
k=0
1.4 Multinomial Coefficients

The binomial formula can be generalized for more than two variables.
Definition 1.5. For non-negative integers k1 , . . . , kr with k1 + . . . + kr = n the
n
multinomial coefficients k1 ,...,kr
are the coefficients when expanding the n-th
power of a sum of r variables. In other words, define them to be the unique
numbers satisfying the following identity:
X
n
(x1 + . . . + xr )n = xk1 xk2 . . . xkr r .
k1 , . . . , k r 1 2
k1 ,...,kr
k1 +...+kr =n
Example. Verify by multiplying out:
(x1 + x2 + x3 )4 = 1·(x41 + x42 + x43 )

+ 4·(x31 x2 + x31 x3 + x1 x32 + x32 x3 + x1 x33 + x2 x33 )
+ 6·(x21 x22 + x21 x23 + x22 x23 )
+ 12·(x21 x2 x3 + x1 x22 x3 + x1 x2 x23 ).
The coefficient in front of x21 x2 x3 is 12, the coefficient in front of x21 x23 is 6 and
the coefficient in front of x43 is 1.
16
According to the last definition this means:

4 4 4
= 12, = 6, = 1.
2, 1, 1 2, 0, 2 0, 0, 4
Theorem 1.6 (Multinomial Theorem).

For non-negative integers k1 , . . . , kr with k1 + . . . + kr = n we have:

n n n − k1 n − k1 − . . . − kr−1 n!
= · ·. . .· = .
k1 , . . . , kr k1 k2 kr k1 ! · k2 ! · . . . · kr !
Proof. The naive way to multiply out (x1 + . . . + xr )n would be:

r X
X r r
X
(x1 + . . . + xr )n = ... xi1 xi2 . . . xin .
i1 =1 i2 =1 in =1
| {z }
n summation signs
The monomial xi1 xi2 . . . xin is equal to xk11 . . . xkr r if from the indices {i1 , . . . , in }
exactly k1 are equal to 1, k2 equal to 2 and so on.
We count the number of assignments of values to the indices {i1 , . . . , in }
satisfying this.

Choose k1 indices to be equal to 1: There are kn1 ways to do so.
Choose k2 indices to be equal to 2: There are n − k1 indices left to choose
from, so there are n−k
k2
1
ways to choose k2 of them.
Choose kj indices to be equal to j (j ∈ [r]): There n − k1 − . . . − kj−1

indices left to choose from, so there are n−k1 −...−k
kj
j−1
ways to choose kj
of them.
Hence the multinomial coefficient (i.e. the coefficient of xk11 xk22 . . . xkr r ) is

n n n − k1 n − k1 − k2 − . . . − kr−1
= · · ... ·
k1 , . . . , kr k1 k2 kr
and the first identity is proved. Now use Theorem 1.4 to rewrite the binomial
coefficients and obtain
n! (n − k1 )! (n − k1 − . . . − kr−1 )!
· · ... ·
k1 !(n − k1 )! k2 !(n − k1 − k2 )! kr !(n − k1 − . . . − kr )!
Conveniently, many of the factorial terms cancel out, like this:
n! (n−k1
)! (n( −(k1(−(. .(
.−(k(((
r−1 )!
·
( · ... · ( .
k1 !
(n− k1 )! k2 !(
(n(−(k1 − k2 )! kr !(n − k1 − . . . − kr )!
( ( (
Also, (n − k1 − . . . − kr )! = (n − n)! = 0! = 1 so we end up with:

n!
.
k1 ! · k2 ! · . . . · kr !
5
5! 5·4·3·
Example. 3,0,2 = 3!0!2! = 2·
1 = 10.
3·
2·
1·2·1

17
Note that the last Theorem establishes that binomial coefficients are special
cases of multinomial coefficients. We have for 0 ≤ k ≤ n:

n n! n
= = .
k k!(n − k)! k, n − k
We extend
the definition of binomial andmultinomial
coefficients by setting
n n n
k1 ,...,kr = 0 if k i = −1 for some i, and −1 = n+1 = 0. This makes stating
the following lemma more convenient.
Lemma 1.7 (Pascal’s Formula).
If n ≥ 1 and 0 ≤ k ≤ n, we have

n n−1 n−1
= + .
k k k−1
More generally, for n ≥ 1 and k1 , . . . , kr ≥ 0 with k1 + . . . + kr = n, we have

r
X
n n−1
= .
k1 , . . . , kr i=1
k1 , . . . , ki − 1, . . . , kr
Proof. Note first, that in the case of binomial coefficients, the claim can be
rewritten as:

n n−1 n−1
= +
k, n − k k, n − k − 1 k − 1, n − k
so it really is a special case of the second claim, which we prove now. By

definition of multinomial coefficients we have the identity:
X
n
n
(x1 + . . . + xr ) = xk1 . . . xkr r .
k1 , . . . , k r 1
k1 ,...,kr ≥0
k1 +...+kr =n
Exploring a different way to expand (x1 + . . . + xr )n , we obtain:
(x1 + . . . + xr )n = (x1 + . . . + xr ) · (x1 + . . . + xr )n−1

X
n−1 k0 k0
= (x1 + . . . + xr ) · 0 0
x11 . . . xr r
0 0
k1 , . . . , k r
k1 ,...,kr ≥0
k10 +...+kr0 =n−1
r
X X
n−1 k0 k0 +1 k0
= x 1 . . . xi i . . . xr r .
k10 , . . . , kr0 1
i=1 k10 ,...,kr0 ≥0
k10 +...+kr0 =n−1
By substituting indices as ki := ki0 + 1 and kj := kj0 (for j 6= i) this is equal to:

r
X X
n−1
(x1 + . . . + xr )n = xk1 . . . xkr r .
i=1 k1 ,...,kr ≥0
k1 , . . . , ki − 1 . . . , kr 1
ki −1≥0
k1 +...+kr =n
18
Note that the condition ki − 1 ≥ 0 is not needed, because for the summands
with ki − 1 = −1 we defined the coefficient under the sum to be 0. We remove
it and swap the summation signs, ending up with:
X r
X
n−1
(x1 + . . . + xr ) =n
xk1 . . . xkr r
k1 , . . . , ki − 1 . . . , kr 1
k1 ,...,kr ≥0 i=1
k1 +...+kr =n
Now we have two ways to write the coefficients of (x1 + . . . + xr )n and

therefore the identity:
r
X
n n−1
=
k1 , . . . , k r i=1
k1 , . . . , ki − 1 . . . , kr
as claimed.
You may already know Pascal’s Triangle. It is a way to arrange binomial
coefficients in the plane as shown in Figure 9. It is possible to do something
similar in the case of multinomial
coefficients, however, when drawing all coef-
n
ficients of the form k1 ...,kr
, the drawing will be r-dimensional. For r = 3, a
“Pascal Pyramid” is given in Figure 10.
Now that we know what multinomial coefficients are and how to compute
them, it is time to see how they can help us count things.
Example. How many 6-permutations are there of the multiset { , , , , , }?
Trying to list them all (( , , , , , ), ( , , , , , ), ), ( , , , , , ), . . .)
would be tedious. Well, there are 6 possible positions for . Then there are 5
positions left, 2 of which need to contain . The three need to go to the three
remaining positions, so there is just one choice left for them. This
means
there

are 6 · 52 · 1 = 60 arrangements in total. This is equal to 61 52 33 = 1,2,36
,
which is no coincidence:
Theorem 1.8. Let S be a finite multiset with k different types and repetition
numbers r1 , r2 , . . . rk . Let the size of S be n = r1 + r2 + · · · + rk . Then the
number of n-permutations of S equals

n
.
r1 , . . . , rk
Proof. Label the k types as a1 , . . . , ak . In an n-permutation there are n positions

that need to be assigned a type. First choose the r1 positions for the first type,
there are rn1 ways to do so. Then assign r2 positions for the second type, out

of the n − r1 positions that are still left to choose. That amounts to n−r r2
1
choices. Continuing like this, the total number of choices will be:

n n − r1 n − r1 − r2 − . . . − rk−1 Thm 1.6 n
· · ... · = .
r1 r2 rk r1 , r2 , . . . , rk

We have seen before how the coefficient nk = k,n−k n
counts the number
of k-combinations of [n] (i.e. number of subsets of [n]). Now we learned, that
19
0
k=
n=0 1
1
k=
n=1 1 1
2
k=
n=2 1 2 1
3
k=
n=3 1 3 3 1
4
k=
n=4 1 4 6 4 1
5
k=
n=5 1 5 10 10 5 1
6
k=
n=6 1 6 15 20 15 6 1

Figure 9: Arrangement of the binomial coefficients nk . Lemma 1.7 shows that

the number nk is obtained as the sum of the two numbers n−1 k−1 and
n−1
k
directly above it.
1
1
1
1 1
2
1 2 1
3 2
1 3 1 3
4 6 1
1 4 3 6 3
5 12 3 4
10 12 1
5 6 1 10 4
20 12 5
10 4 30 6 1
30 4 20
30 5
10 1 10
20
5 10
5
1

Figure 10: Arrangement of the numbers k1 ,kn2 ,k3 with k1 + k2 + k3 = n in a
pyramid. To make the picture less messy, the numbers in the back (with large
k3 ) are faded out. The three flanks of the pyramid are pascal triangles, one of
which is the black triangle in the front with numbers k1 ,kn2 ,0 . The first number
3

not on one of the flanks is 1,1,1 = 6.
20
it also counts the number of n-permutations of a multiset with two types and
repetition number k and n − k. How come ordered and unordered things are
counted by the same number? We demystify this by finding a natural bijection:
Consider the multiset M := {k · X, (n − k) · X} with types X and X (the
“chosen” type and the “unchosen” type). Now associate with an n-permutation
of M the set of positions that contain X, for instance with n = 5, k = 2 and
the permutation (X, X, X, X, X), the corresponding set would be {1, 4} ⊂ [5]
(since the first and fourth position received X). It is easy to see that every
n-permutation of M corresponds to a unique k element subset of [n] and vice
versa.
Note that so far we only considered n-permutations of multisets of size n.
What about general r-permutations of multisets of size n? For example, the
number of 2 permutations of the multiset { , , , , } is 7, since there is:
, , , , , , .
Note that and is not possible, since we have only of one copy of
and at our disposal. The weird number 7 already suggests that general r-
permutations of n element multisets may not be as easy to count. Indeed, there
is no simple formula as in Theorem 1.8 but we will see an answer using the
principle of inclusion and exclusion later.
There is a special case other than r = n that we can handle, though: If
all repetition numbers ri of a multiset with k types are bigger or equal than
r, for instance when considering 2-permutations of M := { , , , , , , },
then those repetition numbers do not actually impose a restriction, since we
will never run out of copies of any type. We would sometimes sloppily write
M = {∞ · , ∞ · , ∞ · } where the infinity sign indicates that there are “many”
copies of the corresponding elements. The number of r-permutations of M is
then equal to k r , just choose one type of the k types for each of the r positions.
After permutations of multisets, we now consider combinations.
Example. Say you are told to bring two pieces of fruit from the supermarket
and they got , and (large quantities of each). How many choices do you
have? Well, there is: { , }, { , }, { , }, { , }, { , }, { , }, so six
combinations. Note that bringing a and an is the same as bringing an
and a (your selection is not ordered), so this option is counted only once.
We now determine the number of combinations for arbitrary number of types
and number of elements to choose.
Theorem 1.9. Let r, k ∈ N and let S be a multiset with k types and large repe-
tition numbers (each r1 , . . . , rk is at least r), then the number of r-combinations
of S equals
k+r−1
.
r
Proof. For clarity, we do the proof alongside an example. Let the types be
a1 , a2 , . . . , ak , for instance k = 4 and a1 = , a2 = , a3 = , a4 = . Then
imagine the r-combinations laid-out linearly, first all elements of type a1 then
all of type a2 and so on. In our example this could be
for the combination {2 · ,2 · ,1 · }.
21
Now for each i ∈ [k − 1], draw a delimiter between types ai and ai+1 , in our
example:
.
Note that, since we have no elements of type a2 = , there are two delimiters
directly after one another. Given these delimiters, drawing the elements is
actually redundant, just replacing every element with “•” yields:
• • • • •.
The original k-combination can be reconstructed from this sequence of and •:

Just replace every • with a1 until the first occurrence of , then use a2 until the
next , then a3 and so on. So every (r+k −1)-permutation of T := {(k −1)· , r·•}
corresponds to an r-combination of S and vice versa.
We know how to count the former, Theorem 1.8 the number of (r +k −1)-
by
permutations of T is r+k−1
r,k−1 = r+k−1
r .
Counting r-combinations of multisets where repetition numbers r1 , . . . , rk
may be smaller than r is more difficult. We will see later how the inclusion-
exclusion principle provides an answer in this case.
1.5 The Twelvefold Way – Balls in Boxes

The most classic combinatorial problem concerns counting arrangements of n
balls in k boxes. There are four cases: U → L, L → U, L → L, U → U,
for arrangements of Labeled or Unlabeled balls −→in Labeled or Unlabeled
boxes. Here “labeled” means distinguishable and “unlabeled” means indistin-
guishable.
These four cases become twelve cases if we consider the following sub-
variants:
(i) No box may contain more than one ball. Example: When assigning stu-
dents to topics in a seminar, there may be at most one student per topic.
(ii) Each box must contain at least one ball. Example: When you want to get
10 people and 5 cars to Berlin, you have some flexibility in distributing
the people to cars, but a car cannot drive on its own.
(iii) No restriction.
A summary of the results of this section is given in the end in Table 1.
Instead of counting the ways balls can be arranged in boxes, some people
count ways that balls can being put into the boxes or being picked from the
boxes, but this does not actually make a difference. We now systematically
examine all 12 cases.
1.5.1 U → L: n Unlabeled Balls in k Labeled Boxes

Example 1.10. Torsten decided there should always be k = 30 points to a
worksheet and n = 5 problems per worksheet. Stefan believes some problems are
easier than others and deserve more points. He wonders how many ways there
are to distribute the points to the problems. In this setting, balls correspond
22
to points (points are not labeled, they are just points) and boxes correspond to
problems (they are labeled: There is problem 1,2,3,4 and 5).
In other words, we search for solutions to the equation
30 = x1 + x2 + x3 + x4 + x5
with non-negative integers x1 , . . . , x5 , for example:
30 = 8 + 4 + 5 + 5 + 7 or 30 = 10 + 5 + 8 + 4 + 3.
The students say that they would like it if some of the problems are “bonus
problems” worth zero points, so the partition 30 = 10 + 10 + 10 + 0 + 0 should
be permissible. Torsten remains unconvinced.
We come back to this later and examine three cases for general n and k in
the balls-and-boxes formulation:
≤ 1 ball per box Of course, this is only possible if there are at most as many
balls as boxes (n ≤ k). For n = 2 and k = 5 one arrangement would be:
1 2 3 4 5
Each of the k boxes can have two states: Occupied (one ball in it) or
empty (no ball in it) and exactly n boxes are occupied. The number of
ways to choose these occupied boxes is nk .
≥ 1 ball per box Of course, this is only possible if there are at least as many
balls as boxes (n ≥ k). For example for n = 9 and k = 5:
1 2 3 4 5
To count the number of ways to do this, arrange the balls linearly, like
this:
and choose k − 1 out of n − 1 gaps between the balls that correspond to

the beginning of a new box. In the example above this would look like
this:
There is a bijection between the arrangements with no empty box and the
choices of n − 1 gaps for delimiters out of k − 1 gaps in total. We know
how to count the latter: There are n−1
k−1 possibilities.

Going back to the example, we now know there are 29 5 solutions to the
equation:
x1 + x2 + x3 + x4 + x5 = 30
where x1 , . . . , x5 are positive integers.
23
arbitrary number of balls per box Now boxes are allowed to contain any
number of elements, including 0. One example for n = 7 and k = 5 would
be:
1 2 3 4 5
We count the number of such arrangements in three different ways:
1. From an arrangement we obtain a string by “reading” it from left to

right, writing when we see a ball and when a new box starts.
For the example above this yields the string
We see: The permutations of the multiset {n · , (k − 1) · } corre-

spond directly to the arrangements of the balls. Note the difference
to the case of non-empty boxes we discussed before.
We know how to count the permutations of the multiset, by Theorem
1.8 there are n+k−1
k−1 of them.
2. There is a bijection between:
(i) Arrangements of n balls in k boxes
(ii) Arrangements of n + k balls in k boxes with no empty box.
For a map from (i) to (ii), just add one ball to each box. For the
inverse map from (ii) to (i) just remove one ball from each box.

We already know how to count (ii), there are n+k−1
k−1 such arrange-
ments.
3. Assume n, k ≥ 1.To count the arrangements, first choose the number
i of boxes that should be empty (0 ≤ i ≤ k − 1), then choose which
boxes that should be (there are ki choices) and then distribute all n
balls to the remaining boxes k − i boxes such that non of those boxes
is empty (we already know how to count this). This yields:
X
k−1
k

n−1

.
i=0
i k−i−1
This looks different from the other two results, which means we “ac-
cidentally” proved the non-trivial identity:
k−1
X k n − 1
n+k−1
= (n, k ≥ 1).
k−1 i=0
i k−i−1

30+6−1
Going back to the example, we now know there are 6−1 solutions to the
equation:
30 = x1 + x2 + x3 + x4 + x5
with non-negative integers, i.e. that many assignments of points to the five
exercises on exercise sheets such that the total number of points is 30.
24
Torsten thinks that no exercise should be worth more than 10 points. In
the balls and boxes setting this limits the capacity ri of a box i. This makes
counting much more difficult but we will see a way to address this in Chapter 2
using the principle of inclusion and exclusion.
(As for homework problems, it turns out that Jonathan put \def\points{6}
into his latex preamble, which settles the issue.)
1.5.2 L → U: n Labeled Balls in k Unlabeled Boxes

Example 1.11. There are n = 25 kids on a soccer field that want to form k = 4
teams. The kids are on different skill levels (e.g. a team of five bad players is
quite a different thing than a team of five good players) but the teams are
unlabeled (it wouldn’t make a difference if you swap all players of two teams).
In how many ways can the n kids form k teams? In other words: In how many
ways can the set of kids be partitioned into k parts?
Here, kids correspond to balls and teams correspond to boxes. Again, we con-
sider three subcases for general k and n.
balls as boxes (n ≤ k). Each ball will be in its own box. If the balls are
labeled with numbers from [n], there will be one box with the ball with
label 1, one box with the ball with label 2 and so on. In other words, there
is only 1 possibility, for n = 3 and k = 5 the only arrangement looks like
this:
1 2 3
balls as boxes (n ≥ k).
This is the same as the number of partitions of [n] into k non-empty parts
which we also call Stirling Numbers of the second kind and write as sII
k (n).
Some values are easy to determine:
• sII
0 (0) = 1: There S is one way to partition the empty set into non-
empty parts: ∅ = X∈∅ X. Each X ∈ ∅ is non-empty (because no
such X exists).
• sII
0 (n) = 0 (for n ≥ 1): There is no way to partition non-empty sets
into zero parts.
• sII
1 (n) = 1 (for n ≥ 1): Every non-empty set X can be partitioned
into one non-empty set in exactly one way: X = X.
2n −2
• sII
2 (n) = 2 = 2n−1 − 1 (for n ≥ 1): We want to partition [n] into
two non-empty parts. If we consider the parts labeled (there is a first
part and a second part), then choosing the first part fully determines
the second and vice versa. Every subset of [n] is allowed to be the
first part – except for ∅ and [n]. This amounts to 2n − 2 possibilities,
however, since the parts are actually unlabeled (there is no “first”
or “second”) every possibility is counted twice so we need to divide
by 2.
25
• sII
n (n) = 1: There is only one way to partition [n] into n parts: Every
number gets its own part.
A recursion formula for sII

k (n) is given as:
sII II II
k (n) = ksk (n − 1) + sk−1 (n − 1).
To see this, count the arrangements of the n labeled balls in unlabeled

boxes with no empty box as follows:
• The ball of label n may have its own box (with no other ball in
it). The number of such arrangements is equal to the number of
arrangements of the remaining n − 1 balls in k − 1 boxes such that
none of those k − 1 boxes is empty. There are sII
k−1 (n − 1) of those.
• The box with the ball of label n contains another ball. Then, when
removing ball n, there is still at least one ball per box. So removing
ball n gets us to an arrangements of n−1 balls in k non-empty boxes.
There are sIIk (n − 1) of those and for each there are k possibilities
where ball n could have been before removal (note that the boxes
are distinguished by the balls that are already in it). So there are
k · sII
k (n − 1) arrangements where ball n is not alone in a box.
The recursion formula follows from summing up the values for the two
cases.
Note that the Stirling Numbers fulfill a recursion similar to the recursion of
binomial coefficients, so there is something similar to the Pascal Triangle.
See Figure 11.
0
k=
n=0 1
1
k=
n=1 0 1
2
k=
n=2 0 1 1
3
k=
n=3 0 1 3 1
4
k=
n=4 0 1 7 6 1
5
k=
n=5 0 1 15 25 10 1
6
k=
n=6 0 1 31 90 65 15 1
Figure 11: “Stirling Triangle”. A number sIIk (n) is obtained as the sum of the
number sII II
k−1 (n − 1) toward the top left and k times the number sk (n − 1)
towards the top right. E.g. for n = 6, k = 3: 90 = 15 + 3 · 25.
26
We will later see a closed form of sII
k (n) in Chapter 2 using the principle
of inclusion and exclusion.
arbitrary number of balls per box Empty boxes are allowed. To count all
arrangements, first choose the number i of boxes that should be non-empty
(0 ≤ i ≤ k), then count the number of arrangements of n balls into the i
boxes such that non of them is empty. This gives a total of:
k
X
sII
i (n).
i=0
1.5.3 L → L: n Labeled Balls in k Labeled Boxes

Example 1.12. At the end of the semester, each of the n = 70 students of
combinatorics will be assigned one of the k = 11 grades from {1.0, 1.3, . . . , 5.0}.
Both Torsten and the students think that the outcome
C AB H FG I JK
1.0 1.3 1.7 5.0
...
would be preferable to the outcome
H FG I JK C AB
1.0 1.3 1.7 5.0
...
so the boxes (grades) are clearly labeled (we could not draw the boxes 2.0 till
4.0 due to lack of space) . Furthermore, Alice (A) and Bob (B) insist that they
would notice the difference between the arrangement
A E CD FG B
1.0 1.3 1.7 3.0
... ...
and the arrangement
B E CD FG A
1.0 1.3 1.7 3.0
... ...
so the balls (students) are labeled as well. Such arrangements correspond di-
rectly to functions f : [n] → [k], mapping each of the n balls to one of the k
boxes, or here: Mapping every student to their grade.
As before, we consider three subcases:
balls as boxes (n ≤ k).
Such arrangements correspond to injective functions f : [n] → [k]. We
first choose the image of 1 ∈ [n] (there are k possibilities), then the image
27
of 2 ∈ [n] (there are k − 1 possibilities left) and so on. Therefore, the
number of injective functions (and therefore arrangements with at most
one ball per box) is:

k! k
k · (k − 1) · · · · · (k − n + 1) = = n!.
(k − n)! n
These arrangements correspond to surjective functions from [n] to [k].
They can also be thought of as partitions of [n] into k non-empty distin-
guishable (!) parts. So we count the number of ways to partition [n] into
k non-empty indistinguishable parts (there are sII
k (n)) and multiply this
by the number of ways k! to assign labels to the parts afterwards. So in
total, there are
k!sII
k (n)
ways to assign n labeled balls to k non-empty labeled boxes.
arbitrary number of balls per box There are k choices for each of the n
balls, so k n arrangements in total.
1.5.4 U → U: n Unlabeled Balls in k Unlabeled Boxes

Example 1.13. Adria used to play tennis but now wants to sell her old tennis
balls (n in total) on the local flee market. She wants to sell them in batches
since selling them individually is not worth the trouble. So she packs them into
k bags. Some people may want to buy bigger batches than others so she figures
it might be good to have batches of varying sizes ready. She wonders how many
ways there are to pack the balls into bags. One such arrangement (n = 9, k = 4)
could be:
Balls and boxes are unlabeled. Adria cannot distinguish any two balls and can
also not distinguish boxes with the same number of balls.
Even though boxes have no intrinsic ordering, we need to somehow arrange
them on this two-dimensional paper. In order to not accidentally think that
and are different, we use the convention of drawing boxes
in decreasing order of balls. With this convention, an arrangement will look
different on paper if and only if it is actually different in our sense.
With this in mind we see the number of arrangements of n unlabeled balls
in k unlabeled boxes is equal to the number of ways to partition the integer n
into k non-negative summands. For example:
9 = 4 + 2 + 2 + 1.
Partitions where merely the order of the summands differ are considered the
same, so again we use the convention of writing summands in decreasing order.
28
balls as boxes (n ≤ k).
In that case, there is only one way to do it: Put every ball in its own box.
Then there are n boxes with a ball and k − n empty boxes.
As discussed before, we count the number of ways in which the integer n
can be partitioned into exactly k positive parts, i.e.
n = a1 + a2 + . . . + ak , where a1 ≥ a2 ≥ . . . ≥ ak ≥ 1.
This is counted by the partition function, denoted by pk (n). A few values

are easy to determine:
• p0 (n) = 0 (for n ≥ 1): No positive number can be partitioned into

zero numbers.
• pn (n) = 1: To write n as the sum of n positive numbers, there is
exactly one choice:
n = 1 + 1 + ...1.
| {z }
ntimes
To compute other values, observe the following recursion:
pk (n) = pk (n − k) + pk−1 (n − 1).
To see this, observe the two cases:
• Either ak = 1. Then a1 + . . . + ak−1 is a partition of n − 1 and there

are pk−1 (n − 1) ways for this.
• Or ak ≥ 2. This means each ai is at least 2 (for 0 ≤ i ≤ k). There is
a bijection between:
(i) Partitions of n into k parts of size at least 2.
(ii) Partitions of n − k into k parts of size at least 1.
To go from (i) to (ii), just remove 1 from each part and to go back,
just add 1 to each part. We already know how to count (ii), there
are pk (n − k) such partitions.
arbitrary number of balls per box Similar to what we did in the L → U
case, first decide how many boxes i should be non-empty and then count
how many arrangements with exactly i non-empty boxes exist:
k
X
pi (n).
i=1
This can be thought of as the number of integer partitions of n into at

most k non-zero parts.
29
n balls k boxes ≤ 1 per box ≥ 1 per box arbitrary
k
n−1
n+k−1

U L n k−1 k−1
Pk
L U 1 sII
k (n)
II
i=1 si (n)
k

L L n n! sII
k (n)k! kn
Pk
U U 1 pk (n) qk (n) = i=1 pi (n)
Table 1: The twelvefold way.
1.5.5 Summary: The Twelvefold Way

The twelve variations of counting arrangements of balls in boxes are called the
Twelvefold Way. A summary of our results is given in Table 1.
We also summarize the interpretations of n Labeled or Unlabeled balls in k
Labeled or Unlabeled boxes:
Arrangements Correspond to
U→L non-negati ve Integer solutions of x1 + . . . + xk = n.
L→U Partitions of the set [n] into k parts.
L→L Functions from [n] to [k]
U→U Partitions of the number n into k non-negative integers.
1.6 Binomial Coefficients – Examples and Identities

Example 1.14 (Lattice Paths). A monotone lattice paths in the grid L :=
{0, 1, . . . , m}×{0, 1, . . . , n} is a path starting in the bottom left corner (0, 0) and
ending in the top right corner (m, n), taking single steps upwards or rightwards
see Figure 12.
(7,4)
(0,0)
Figure 12: Here m = 7 and n = 4. The lattice path →↑→↑→→→↑→→↑ is

shown.
In other words, a lattice path is a sequence of “↑” (upward steps) and “→”
(rightward steps) with a total of n times ↑ and m times →. In yet other words,
the lattice paths correspond to permutations of the multiset {m · →, n · ↑} and
by Theorem 1.8 their number is

m+n
.
m
30
Example (Cake Number). For positive integers m and n the cake number
c(m, n) is the maximum number of pieces obtained when cutting an m-dimensional
m
Sn by n cuts, i.e. the maximum number of connected components of R \
cake
H
i=1 i , where H i is an (m − 1)-dimensional affine hyperplane. For m = 2, this
simply means: Put n lines into the plane and observe into how many pieces the
plane is cut. See Figure 13 for an example.
2
3
1
10 4
9 11
8
6 5
7
Figure 13: The plane (m = 2) is cut by n = 4 lines. In this particular configu-

ration, this gives 11 pieces. No configuration with this n and m can achieve a
larger number of pieces.
It turns out that

Xm
n
c(m, n) = .
i=0
i
n

Here i is defined to be zero for i > n. For instance

4 4 4
c(2, 4) = + + = 1 + 4 + 6 = 11,
0 1 2
so the example given in Figure 13 is tight.
The cake numbers arise in a seemingly unrelated setting: For m ≥ 3, consider

any finite set X of n points in Rm−1 in general position (no m points on a
hyperplane, no m + 1 on a (m − 2)-dimensional sphere). Then the number of
subsets Y of X that can be separated from X \ Y by a (m − 2)-dimensional
sphere is equal to c(m, n).
For m = 3 this means: Consider n points in the plane where no 3 points
are on a line and no 4 points on a circle2 . Then the number of subsets Y of X
that can be captured by a circle is equal to c(3, n). Consider Figure 14 for an
example. We count all sets that can be captured there.
size 0: The empty set.
size 1: All 5 single element sets.

5
size 2: All except for {b, e} and {a, c}, so 2 − 2 = 8.
size 3: {a, b, c}, {a, b, d}, {a, b, e}, {a, d, e}, {b, c, d}, {b, d, e}, {c, d, e}, so 7 of them.
size 4: All except for {a, b, c, e}, so 4 of them.
2 These assumptions are very week: If you pick random points, say by throwing a bunch of
darts onto the plane, the probability that the points will be in general position is 1.
31
e
d
a
c
b
Figure 14: There are n = 5 points in the plane (m = 3). Some subsets can be
captured by a circle (the pictures shows: {a, d}, {b, d, e}, {c, d, e}, {c}, ∅), some
cannot, for example {b, c, e} (such a circle would also contain d).
size 5: The set {a, b, c, d, e}.

That is 1 + 5 + 8 + 7+ 4 +
1 = 26 sets in total, exactly the cake number
c(3, 5) = 50 + 51 + 52 + 53 = 1 + 5 + 10 + 10 = 26.
There is a seemingly infinite number of useful identities involving binomial co-
efficients, we will show a few of them.
Theorem 1.15 (Proofs by Picture). We have:
P
n 2
(i) k = (n+1) 2−(n+1) = n+1
2 .
k=1
P
n
(ii) (2k − 1) = n2 .
k=1
Proof. (i) For the first identity, arrange 1+2+. . .+n dots in a triangle, mirror
it, and observe that this fills all positions of a square of size (n+1)×(n+1)
except for the diagonal. Here is a picture for n = 6:
• • •••• • •••••
• • ••• • •••• •
• • •• • ••• ••
• • • −→ • •• •••
• • • • ••••
• • •••••
••••••
P
n
Therefore 2 · k = (n + 1)2 − (n + 1) = n(n + 1). Dividing by 2 proves
1
the claim.
(ii) Note how tiles of with sizes of the first n odd numbers can be arranged to
form a square of size n × n, here a picture for n = 5:
1
3
5
7
9
32
Theorem 1.16 (Proofs by Double Counting). We have
P
n
n

(i) 2k k = 3n .
k=0
m
X
n+i n+m+1
(ii) = .
i=0
i m
Proof. (i) We double count strings over the alphabet {0, 1, 2} of length n, in
other words, the n permutations of the multiset {∞ · 0, ∞ · 1, ∞ · 2} where
there is infinite supply of the types 0, 1, 2.
First way: There are three possibilities per character, so 3n possibilities
in total.
Second way: First choose the number of times k that the letter “0”
should be used ( 0 ≤ k ≤ n). Then choose the positions for those
characters, there are nk possibilities. Finally, choose for each of the
remaining (n − k) positions if they should be 1 or 2, there are 2n−k
choices. So in total, there is this number of possibilities:
Xn
n−k n
2 ·
k
k=0
By reversing the order of summation, this is equal to

Xn X n
k n k n
2 · = 2 · .
n−k k
k=0 k=0
This proves the claim.

(ii) We double count the lattice paths in the lattice {0, . . . , m} × {0, . . . , n} of
width m and height n.

First way: We already counted them in Example 1.14. There are n+m m
of them.
Second way: Any path must reach the last row (row n) eventually, using
exactly one of the “↑”-steps from (k, n−1) to (k, n) where 0 ≤ k ≤ m.
See Figure 15 for a sketch.
There is only one way a lattice path can continue
from (k, n), namely
go rightwards until (m, n). There are k+n−1
k ways to get from (0, 0)
to (k, n − 1) though, so the number of lattice paths is:
m
X
k+n−1
.
k
k=0
n+m
Pm k+n−1

This gives the equality m = k=0 k . In the claim, we merely
replaced n by n + 1.
A few other identities can be derived by using the connection of multinomial
coefficients to polynomials.
33
(7,4)
(0,0)
Figure 15: At some point, a lattice path must cross the dashed line, i.e. use
one of the edges going to the last row. We count thenumber of paths using the
highlighted edge from (4, 3) to (4, 4): There are 4+3 ways to get from (0, 0) to
4
(4, 3) and one way to get from (4, 4) to (7, 4), so 4+3
4 · 1 paths in total.
Theorem 1.17 (Proofs by Analysis). We show three equations (i), (ii) and
(iii), see below.
Proof. Start with the binomial formula:
X n
n i n−i
(x + y)n = xy .
i=0
i
Setting y = 1 yields:
Xn
n i
(i) (x + 1)n = x.
i=0
i
Deriving by x gives:
Xn
n i−1
n(x + 1)n−1 = i x .
i=1
i
Now set x = 1 and get:
X n
n
(ii) n · 2n−1 = i .
i=0
i
Taking the second derivative of (i) yields:
X n
n i−2
n(n − 1)(x + 1)n−2 = i(i − 1) x .
i=2
i
Again, we set x = 1
Xn
n
n(n − 1)2n−2 = i(i − 1) .
i=0
i
Adding (ii) to this gives:
Xn
n−2 2 n
(iii) n(n + 1)2 = i .
i=0
i
1.7 Permutations of Sets

Remember that n-permutations of [n] are bijections π : [n] → [n]. From now on
we will just say “permutation of [n]” and drop the “n-”.
34
For such a permutation, the pair (i, j) of two numbers from [n] is called an
inversion if the numbers are ordered, but swapped by the permutation, i.e.
(i, j) is inversion ⇔ i < j ∧ π −1 (i) > π −1 (j).
Take for example the permutation π = 31524. It has the inversions (1, 3), (2, 3),
(2, 5), (4, 5).
For i ∈ [n] define αi := |{j ∈ [n] | (i, j) is inversion}|. The inversion sequence
of π is the sequence α1 α2 . . . αn .
The inversion number or disorder of a permutation π is the number of its
inversions, so α1 +· · ·+αn . For our example the inversion sequence is 1, 2, 0, 1, 0
and its disorder is 4.
Since for any i ∈ [n], there are only n−i numbers bigger than i, any inversion
sequence α1 , . . . , αn satisfies:
0 ≤ αi ≤ n − i (i ∈ [n]). (?)
There are exactly n! integer sequences satisfying (?) (n choices for α1 , n − 1

choices for α2 . . . ), just as many as there are permutations. This is no coin-
cidence: We now show that each permutation has a distinct such sequence as
an inversion sequence. By cardinality we then also know that each such se-
quence is the inversion sequence of some permutation, i.e. there is a one-to-one
correspondence.
Theorem 1.18. Any permutation can be recovered from its inversion sequence.
Proof. Let α = (α1 , . . . , αn ) be integers satisfying (?). We will construct a
permutation π with this inversion sequence and it will be clear that we will not
have any freedom in the construction of π, i.e. π will be the unique permutation
with inversion sequence α.
In the beginning, π has n free positions still to be determined:
π=
| {z }
n positions
We determine the position containing 1, i.e. we determine π −1 (1): Remember

that α1 is the number of elements that are bigger than 1 but arranged to the
left of 1 by π. Since every element is bigger than 1, α1 is just the number of
elements arranged to the left of 1, in other words π −1 (1) = α1 + 1.
Take for instance the inversion sequence α = (5, 3, 4, 0, 2, 1, 0, 0). Then from
α1 = 5 we know that 1 is in the sixth position like this:
π= 1
Now recall that α2 is the number of elements bigger than 2 that are sorted to
the left of 2, this means, α2 is exactly the number of (currently) unoccupied
positions to the left of 2. In the example we have α2 = 3, so π must look like
this:
π= 2 1
For general: i ∈ [n], if we have already placed all number from 1 to i − 1, we
can derive that π −1 (i) must be the position of the free slot that has exactly αi
free slots to the left of it.
35
In the example, since α3 = 4 we put 3 to the unoccupied position that has
four unoccupied positions to the left of it:
π= 2 1 3
Continuing this way, we can reconstruct π completely:
π= 4 7 6 2 5 1 3 8.
1.7.1 Generating Permutations

To generate all permutations of [n] use the following algorithm.
Start: Write 1, 2, . . . , n with left arrows over them.

Step: – Locate the largest mobile integer m, i.e., an integer such that
it arrows to a smaller neighbouring number.
– Swap m with that neighbour.
– Change the direction of the arrows on all integers p, p > m.
Let us illustrate this method with the following example.
←
− ←
− ←
− ←
− ←
−←−←−
1 2 3 4 1 2 3
←
− ←
− ←
− ←
−
1 2 4 3
←
− ←
− ←
− ←
−
1 4 2 3
←
− ←
− ←
− ←
−
4 1 2 3
→
− ←
− ←
− ←
− ←
−←−←−
4 1 3 2 1 3 2
←
− →
− ←
− ←
−
1 4 3 2
←
− ←
− →
− ←
−
1 3 4 2
←
− ←
− ←
− →
−
1 3 2 4
←
− ←
− ←
− ←
− ←
−←−←−
3 1 2 4 3 1 2
←
− ←
− ←
− ←
−
3 1 4 2
←
− ←
− ←
− ←
−
3 4 1 2
←
− ←
− ←
− ←
−
4 3 1 2
→
− →
− ←
− ←
− →
− ←
−←−
4 3 2 1 3 2 1
→
− →
− ←
− ←
−
3 4 2 1
→
− ←
− →
− ←
−
3 2 4 1
→
− ←
− ←
− →
−
3 2 1 4
←
− →
− ←
− ←
− ←
−→− ←
−
2 3 1 4 2 3 1
←
− →
− ←
− ←
−
2 3 4 1
36
←
− ←
− →
− ←
−
2 4 3 1
←
− ←
− →
− ←
−
4 2 3 1
→
− ←
− ←
− →
− ←
−←−→−
4 2 1 3 2 1 3
←
− →
− ←
− →
−
2 4 1 3
←
− ←
− →
− →
−
2 1 4 3
←
− ←
− →
− →
−
2 1 3 4
1.7.2 Cycle Decompositions

Remark. The lecture mentioned a puzzle presented on the Youtube Channel
MinutePhysics. It may have something to do with cycles (but we do not want
to spoil the puzzle for you), feel free to check it out on
http://youtube.com/watch?v=eivGlBKlK6M
We already know several ways to specify a permutation.
• Write it as a string: π = 78546132.
• Write it explicitly as a map:
i 1 2 3 4 5 6 7 8
π(i) 7 8 5 4 6 1 3 2
• Another way to specify the map would be to connect i with π(i), maybe
like this:
1 2 3 4 5 6 7 8
1 2 3 4 5 6 7 8
• In the last drawing every number occurs twice (once as a preimage and
once as an image), now we only take one copy of every number and give
a direction to the edges (from preimage to image):
1
7
6 2 8 4
3
5
In other words, we draw an edge from i → j if π(i) = j. Since π is a map,
every node gets one outgoing edge and since π is bijective, every node gets
exactly one incoming edge. From this it is easy to see that the image we
get is a collection of cycles: When starting at any node i and following the
arrows, i.e. walking along the path i → π(i) → π(π(i)) → π(π(π(i))) →
. . ., we must at some point – since there are only finitely many elements
– come for the first time to an element where we already were. This must
be i, since all other nodes already have an incoming edge. So i was indeed
in a cycle.
37
Instead of drawing the picture we prefer to just list the cycles like this:
π = (17356)(28)(4)
where cycles are group by “()” and elements within a cycle are given in
order. The representation at hand is called a disjoint cycle decomposition.
Note that every element occurs in exactly one cycle (some cycles may have
length 1). Note that this composition is, strictly speaking, not unique, we
could also write π = (4)(73561)(28), where we changed the order of two
cycles and “rotated” 7 to the beginning of its cycle. In the following, we
will not distinguish disjoint cycle decompositions that differ only in these
ways.
Before we generalize disjoint cycle decompositions to arbitrary cycle decom-
positions, we define the product of two permutations:
If π1 , π2 ∈ Sn are permutations, the product π1 · π2 (or just π1 π2 for short)
is the permutation π with π(i) = π2 (π1 (i)) (i.e. first apply π1 , then π2 ). Note,
that when viewed as maps: π1 · π2 = π2 ◦ π1 .
Now if we write a single cycle, e.g. π ∈ S7 , π = (137), we mean the permu-
tation that permutes the elements in the cycle along the cycle (here π(1) = 3,
π(3) = 7, π(7) = 1) and leaves all other elements in place, in our example π is
the permutation with the disjoint cycle decomposition (137)(2)(4)(5)(6).
We can now talk about cycle decompositions where elements may occur more
than once, e.g. π = (136)(435)(7452)(6)(23). It is the product of the involved
cycles ( (136), (435), . . . ).
If π ∈ Sn is a permutation given as a cycle decomposition, then π can be
evaluated by parsing, where parsing an element i ∈ [n] in a cycle decomposition
means
• Have a current element c, initially c = i.
• Go through cycles from left to right.
– If the current element c is part of a cycle, replace c with the next
element in this cycle.
• In the end of this process π(i) is the current element c.
Take for instance π = (136)(435)(7452)(6)(23) and i = 3. Then we start with
the current element c = 3. The first cycle (136) contains c = 3, so we change
c to 6 and go on. The next cycle (435) does not contain c = 6, and neither
does (7452) so we move past these cycles without changing c. The next cycle
(6) contains c = 6 but does not alter it. We therefore end with π(3) = 6. As
another example, consider how π(1) is evaluated:
c=1 c=3 c=5 c=2 c=2 c=3
↓ ↓ ↓ ↓ ↓ ↓
(136) (435) (7452) (6) (23)
So π(1) = 3.
Theorem 1.19. If π1 , π2 ∈ Sn are permutations given in cycle decomposition,
then a cycle decomposition of π1 · π2 is given by concatenating the cycle decom-
positions of π1 and π2 .
38
Proof. Let the cycle decompositions be π1 = C1 C2 . . . Ck , π2 = C10 C20 . . . Cl0 .
When parsing C1 C2 . . . Ck C10 C20 . . . Cl0 with i ∈ [n] then the current element
after C1 C2 . . . Ck will be π1 (i) and therefore, after going through the remaining
cycles C10 . . . Cl0 , the current element will be π2 (π1 (i)) = (π1 · π2 )(i). So the
concatenation is indeed a cycle decomposition of π1 · π2 .
Note that non-disjoint cycle decompositions do not always commute, e.g.
(435)(321) = (13542) 6= (15432) = (321)(435)
where the identities can be verified by parsing. The claims of the following
theorem are easy and we will skip the proofs:
Theorem 1.20. Let π ∈ Sn be given in cycle decomposition.
• Swapping adjacent cycles has no effect if they are disjoint (i.e. no number
occurs in both cycles), e.g. (13)(524) = (524)(13).
• A cycle decomposition of π −1 is obtained by swapping the order of all
−1
cycles and reversing the elements in each cycle, e.g. ((321)(435)) =
(534)(123).
• Cyclic shifts in a cycle have no effect. (534) = (345) = (453).
• Up to cyclic shifts and order of cycles, there is a unique decomposition
into disjoint cycles.
Let id ∈ Sn be the identity permutation ( id(i) = i for i ∈ [n]). Define the order
of π ∈ Sn to be the smallest k such that
πk = π
| · π {z
· . . . · π} = id.
k times
Theorem 1.21. The order of π is the least common multiple of the lengths of
the cycles in the disjoint cycle decomposition of π.
Proof. Assume π is composed of the disjoint cycles C1 C2 . . . Cm and an element
i ∈ [n] is contained in a cycle Cj of length l. Then for any k ∈ N, the value
π k (i) is obtained by parsing i through i
C1 C2 . . . Cm C1 C2 . . . Cm . . . C1 C2 . . . Cm
| {z }
k copies of C1 . . . Cm
Only the copies of Cj can affect the current value if we start with i, so the result
is the same as when parsing i through k copies of just Cj . From this we see
that π k (i) = i if and only if k is a multiple of l. This shows that π k = id if and
only if k is a common multiple of the cycle lengths. So, the order of π is the
least common multiple of them.
1.7.3 Transpositions
A cycle of length 2 is a transposition.
Define discriminant of π as N (π) := n − #C where #C is the number of
cycles in a disjoint cycle decomposition of π. Note that we may not omit single
element cycles now, they count towards #C:
In S5 we have for example N (id) = N ((1)(2)(3)(4)(5)) = n − 5 = 0 and
N ((134)(25)) = 5 − 2 = 3.
39
Theorem 1.22.
(i) Any permutation can be written as the product of transpositions.
(ii) If π is the product of k transpositions (k ∈ N), then N (π) and k have the
same parity (i.e. N (π) ≡ k (mod 2)).
Proof. (i) Let π be any permutation. We already know that we can write π
as a product of cycles. So to show that π can be written as a product
of transpositions, it suffices to show that any cycle can be written as the
product of transpositions.
Let C = (a1 a2 . . . al ) be a cycle of length l (assume l ≥ 2, since cycles of
length 1 correspond to identity permutations and can be omitted from any
cycle decomposition).
Then we can write C as the product of l − 1 transpositions:
C = (a1 a2 . . . al ) = (al−1 al )(al−2 al−1 ) . . . (a2 a3 )(a1 a2 )
Verify this by parsing: If i 6= l, then parsing ai yields:
c=ai c=ai c=ai c=ai c=ai+1 c=ai+1 c=ai+1 c=ai+1 c=ai+1
↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓
(al−1 al ) (al−2 al−1 ) . . . (ai ai+1 ) (ai−1 ai ) ... (a2 a3 ) (a1 a2 )
So parsing sends ai to ai+1 as desired. When parsing al we have:

c=al c=al−1 c=al−2 c=a3 c=a2 c=a1
↓ ↓ ↓ ↓ ↓ ↓
(al−1 al ) (al−2 al−1 ) ... (a2 a3 ) (a1 a2 )
so we get a1 as desired.
(ii) Let π be any permutation with k cycles in a disjoint cycle decomposition.
Let (ab) be any transposition. We determine the number of cycles in a
disjoint cycle decomposition of (ab) · π depending on a and b.
Case 1: a and b are part of different cycles in π. Then there is a dis-
joint cycle decomposition of π looking like this:
π = (aX)(bY )C3 . . . Ck
where X and Y are sequences of elements and C3 , . . . , Ck are cycles.
By parsing we verify:
(ab) · π = (ab)(aX)(bY )C3 . . . Ck = (aY bX)C3 . . . Ck .
The last term is a disjoint cycle decomposition of (ab) · π into k − 1
cycles.
Case 2: a and b are part of the same cycle in π. Then there is a dis-
joint cycle decomposition of π looking like this:
π = (aXbY )C2 . . . Ck
where X and Y are sequences of elements and C2 , . . . , Ck are cycles.
By parsing we verify:
(ab) · π = (ab)(aXbY )C2 . . . Ck = (aY )(bX)C2 . . . Ck .
The last term is a disjoint cycle decomposition of (ab) · π into k + 1
cycles.
40
In both cases, adding a transposition changed the number of cycles by one,
which means that the parity of the number of cycles changed. Therefore,
the claim follows by induction.
We call the number sIk (n) of n-permutations of [n] with exactly k cycles in
a disjoint cycle decomposition the unsigned Stirling number of first kind. More
precisely sIk (n) = |{π ∈ Sn | N (π) = n − k}|.
Theorem 1.23. For all n, k ≥ 1
(i) sI0 (0) = 1 and sIk (0) = sI0 (n) = 0,

(ii) sIk (n) = (n − 1)sIk (n − 1) + sIk−1 (n − 1).
Proof. (i) The empty permutation is the only permutation of 0 elements.
Thus sI0 (0) = 1 and sIk (0) = 0, since the empty permutation has no cycle
and k ≥ 1.
On the other hand every permutation of n ≥ 1 elements has at least one
cycle and hence sI0 (n) = 0.
(ii) Let Snk ⊂ Sn denote the set of permutations in Sn with exactly k cycles
in a disjoint cycle decomposition. Then |Snk | = sIk (n). We define a map
k−1
φ : Snk → Sn−1
k
∪ Sn−1 as follows. Consider a permutation π ∈ Snk and
let (AnB) denote the cycle containing n (A and B may be empty). If
(AnB) = (n), remove the entire cycle (n). Otherwise replace the cycle
(AnB) with (AB). In the first case there are k − 1 cycles in a disjoint
cycle decomposition of φ(π). In the latter case φ(π) still has k cycles in a
disjoint cycle decomposition.
k k−1
Consider a disjoint cycle decomposition of some σ ∈ Sn−1 ∪ Sn−1 . We
k
count how many permutations π ∈ Sn are mapped to σ.
k−1
Case 1: σ ∈ Sn−1 . Then the only permutation in Snk mapped to σ is
obtained by adding the cycle (n) to σ.
k
Case 1: σ ∈ Sn−1 . In this case we may place n between any element
x ∈ [n − 1] and its successor in σ (i.e., put it right behind x in the
cycle containing x). Then n has different positions for distinct choices
of x ∈ [n − 1]. Hence n − 1 distinct elements from Snk are mapped
to σ.
k−1
k
Combining these two cases we see that |Snk | = (n − 1)|Sn−1 | + |Sn−1 |.
Recall that the Stirling numbers of second kind sII

k (n) count the number of
partitions of [n] into k non-empty parts. For these numbers we had the formula
sII II II
k (n) = ksk (n − 1) + sk−1 (n − 1). There is also a similar picture of the
recursion, see Figure 16. From the picture we easily see that sI1 (n) = (n − 1)!
(for n ≥ 1). This is the number of different cycles of length n formed by n
elements.
41
0
k=
n=0 1
1
k=
n=1 0 1
2
k=
n=2 0 1 1
3
k=
n=3 0 2 3 1
4
k=
n=4 0 6 11 6 1
5
k=
n=5 0 24 50 35 10 1
6
k=
n=6 0 120 274 225 85 15 1
Figure 16: A number sIk (n) is obtained as the sum of the number sIk−1 (n − 1)
toward the top left and (n − 1) times the number sIk (n − 1) towards the top
right. E.g. for n = 6, k = 3: 225 = 50 + 5 · 35.
1.7.4 Derangements
A permutation π ∈ Sn is a derangement if it is fixpoint-free, i.e., ∀i ∈ [n] :
π(i) 6= i. Note that this means that each cycle has length at least 2. The set of
all derangements of [n] is denoted by Dn .
We close this chapter by “counting” derangements in Theorem 1.24. This
is also meant to demonstrate that there are several different ways to count.
Depending on the purpose, different counts may be more or less helpful.
Remark. You might find the notation |Dn | =!n in the literature (sic!). Since we
think this notation is too confusing we won’t use it here.
Before we start observe that |D1 | = 0 and |D2 | = 1. We may also agree
on |D0 | = 1, since the empty permutation has no fixpoint, but we try to avoid
using D0 .
42
Theorem 1.24. For a natural number n ≥ 1 we have
|Dn | = (n − 1)(|Dn−1 | + |Dn−2 |), if n ≥ 2, (Recursion (i))
n
|Dn | = n|Dn−1 | + (−1) , (Recursion (ii))
n
X n n−1
X n
n! = |Dk |, |Dn | = n! − |Dk |, (Recursion (iii))
k k
k=0 k=0
Xn Xn
n (−1)k
|Dn | = (−1)n+k k! = n! , (Summation)
k k!
k=0 k=0

n! 1
|Dn | = + , (Explicit)
e 2
√ nn
|Dn | ∼ 2πn n+1 (∼ cn log n for some c ∈ R), (Asymptotic)
e
X |Dn | e−x
xn = ∀x ∈ R \ {1}. (Generating Function)
n! 1−x
n≥0
Proof.
Recursion (i): We prove this statement using a map φ : Dn → Dn−1 ∪ Dn−2
as follows. Consider a derangement π ∈ Dn and let x, y ∈ [n] be the
unique numbers such that π(x) = n and π(n) = y. Written as a string, π
is of the form π = AnBy where A is a sequence of x − 1 elements and B
is a sequence of n − x − 1 elements.
If (case 1) x 6= y, then define φ(π) := AyB. Note that σ = φ(π) ∈ Dn−1 ,
since σ(x) = y 6= x. For example, if π = 45123, then x = 2, y = 3 and
φ(π) = 4312.
If (case 2) x = y, then define φ(π) := A∗ B ∗ where A∗ and B ∗ are
identical to A and B except that all numbers bigger than x are decreased
by 1. We claim that σ = φ(π) ∈ Dn−2 . If i < x, then σ(i) = φ(i) 6= i if
φ(i) < x, and σ(i) = φ(i) − 1 ≥ x > i if φ(i) > x. Otherwise i ≥ x and
σ(i) = φ(i+1) < x < i if φ(i+1) < x, and σ(i) = φ(i+1)−1 6= i+1−1 = i
if φ(i + 1) > x. For example, if π = 45132, then x = y = 2, AB = 413 and
φ(π) = 312.
Claim. Each σ ∈ Dn−1 ∪Dn−2 is the image of exactly (n−1) permutations
π ∈ Dn .
Case 1: σ ∈ Dn−1 . Then there are n − 1 choices to pick a position
x ∈ [n − 1]. Let σ(x) = y, write σ = AyB and define π = AnBy.
Then φ(π) = σ. Note that π ∈ Dn , since π(x) = n 6= x.
Case 2: σ ∈ Dn−2 . Then there are n − 1 choices to put n on position
x (as the first element, into a gap or after the last element). Then
there is a unique way to increase σ(i) by 1 for each i with σ(i) ≥ x.
Finally put x on position n. We obtain π = AnBx, which is mapped
to σ. Note that π ∈ Dn due to similar arguments as above.
Recursion (ii): We will apply induction on n. An induction basis is given by
|D2 | = 1 and |D1 | = 0 . Suppose n ≥ 3. We rewrite Recursion (i) as
|Dn | − n|Dn−1 | = −(|Dn−1 | − (n − 1)|Dn−2 |). By induction hypothesis,
the right side is equal to −(−1)n−1 = (−1)n which proves the claim.
43
Remark. For the number |Sn | = P (n) = n! of permutations of [n] we have
a similar recursion |Sn | = (n − 1)(|Sn−1 | + |Sn−2 |), since n! = (n − 1)((n −
1)! + (n − 2)! and |Sn | = n|Sn−1 |.
Recursion (iii): We count all permutations in Sn . On the one hand |Sn | = n!.
On the other hand each π ∈ Sn induces a derangement on[n]\F (π), where
F (π) is the set of all fixpoints of π. Thus there are nk |Dn−k | different
permutations
Pn in Sn with exactly k fixpoints each. Hence n! = |Sn | =
n
k=0 k |Dn−k |.
P1 (−1)k
summation: Proof by induction on n. If n = 1, then |D1 | = 0 = 1 k=0 k! =
1(1 − 1). This gives an induction basis.
Consider n ≥ 2. Then, using the induction hypothesis (IH):
n−1
X
IH (−1)k
|Dn | = n|Dn−1 | + (−1)n = n(n − 1)! + (−1)n
k!
k=0
n−1
X k n
X
(−1) n! (−1)k
= n! + (−1)n = n! .
k! n! k!
k=0 k=0
P zk
explicit: Recall that ez = k≥0 k! . Then
Pn (−1)k n
|Dn | n! X (−1)k 1
k=0
lim = lim k!
= lim = e−1 = .
n→∞ |Sn | n→∞ n! n→∞ k! e
k=0
So for sufficiently large n we have |Dn | ∼ n!

e . Applying standard argu-
n! 1
ments for the speed of convergence of series we obtain |Dn | = e + 2 .
√ n
asymptotic: This follows immediately from Stirling’s formula n! ≈ 2πn ne
and the explicit result.
P
generating function: Let F (x) := n≥0 |Dn!n | xn . Then the derivative of F
P P
is given by F 0 (x) = n≥1 |Dn!n | nxn−1 = n≥1 (n−1)!
|Dn |
xn−1 . From Recur-
|Dn+1 | |Dn | |Dn−1 |
sion (i) we have n! − (n−1)! = (n−1)! and thus
X |Dn |
(1 − x)F 0 (x) = (xn−1 − xn )
(n − 1)!
n≥1
X |Dn+1 | |Dn |

= − xn
n! (n − 1)!
n≥1
X |Dn−1 |
= xn
(n − 1)!
n≥1
= xF (x).
e−x
This differential equation with F (0) = 1 is solved by F (x) = 1−x .
44
2 Inclusion-Exclusion-Principle
and Möbius Inversion
In the last theorem of the previous chapter, and in several other places, we have
seen summations with alternating signs. This chapter will deal with such kind
of results. Consider a finite set X and some properties P1 , . . . , Pm such that
for each x ∈ X and each i ∈ [m] it is known whether x has property Pi or
not. We are interested in the number of elements from X satisfying none of the
properties Pi .
Example. Let X be the set of students in the room and let P1 and P2 be the
properties “being male” and “studies math”, respectively. Suppose there are 36
students, 26 of which are male and 32 of which study math. Among the male
students 23 study math. We are interested in the number of students which are
neither male nor study math. Let X1 , X2 denote the set of male students and
the set of math students, respectively. Then
|X \ (X1 ∪ X2 )| = |X| − |X1 | − |X2 | + |X1 ∩ X2 | = 36 − 26 − 32 + 23 = 1.
See Figure 17.
36 students
male math
26 23 32
Figure 17: When there are 26 male and 32 math students among 36 students
in the class, and 23 of the male students study math, then there is exactly 1
female student who does not study math.
2.1 The Inclusion-Exclusion Principle

Let P1 , . . . , Pm be some m properties and X be an n-element set, where for
each x ∈ X and each i ∈ {1, . . . , m} we have that x has property Pi or x does
not have property Pi . For a subset S of properties, let N (S) denote the set
of elements in X that have properties P for all P ∈ S. Note that N (∅) = X.
Clearly, a property Pi just corresponds
T to the subset Xi of elements of X having
this property and N (S) = i∈S Xi . However, taking properties instead of sets
has the advantage that we can take the same subset more than once.
Theorem 2.1 (Inclusion-Exclusion Principle).
Let X be a finite set and P1 , . . . , Pm properties. Further define for S ⊆ [m] the
set N (S) = {x ∈ X | ∀i ∈ S : x has Pi }, i.e. the set of all elements from X
having property Pi for all i ∈ S. The number of elements of X that satisfy none
of the properties P1 , . . . , Pm is given by
X
(−1)|S| |N (S)|. (1)
S⊆[m]
45
Proof. Consider any x ∈ X. If x ∈ X has none of the properties, then x ∈ N (∅)
and x 6∈ N (S) for any S 6= ∅. Hence x contributes 1 to the sum (1).
If x ∈ X has exactly k ≥ 1 of the properties, call this set of properties
T ∈ [m]
k . Then x ∈ N (S) iff S ⊆ T . P Pk
The contribution of x to the sum (1) is S⊆T (−1)|S| = i=0 ki (−1)i = 0.
Pk
In the last step we used that for any y ∈ R we have (1−y)k = i=0 ki (−y)i
Pk
which implies (for y = 1) that 0 = i=0 ki (−1)i .
The previous theorem can also be proved inductively.
In some settings we might be interested in a set of elements having a certain

set of properties A and none of the properties from a set of properties B. We
can handle this by considering the opposites Ā of the properties from A and
searching for the elements having none of the properties from Ā and B. In other
settings we might be interested in the number of elements satisfying at least one
of the properties. The following corollary answers this question.
Corollary 2.2. The number of elements of X that have at least one of the
properties P1 , . . . , Pm is given by
X X
|X| − (−1)|S| N (S) = (−1)|S|−1 N (S).
S⊆[m] ∅6=S⊆[m]
Given a set X and three properties P1 , P2 , P3 , the Inclusion-Exclusion-Principle

states that the number of elements not in P1 ∪ P2 ∪ P3 is:
|X| − |P1 | − |P2 | − |P3 | + |P1 ∩ P2 | + |P1 ∩ P3 | + |P2 ∩ P3 | − |P1 ∩ P2 ∩ P3 | .
Figure 18 illustrates this: For example, the number of elements in P1 \ (P2 ∪ P3 )
P1 P2
−−
− + −
−−−
+++
−− − −−
+ +
−
P3
Figure 18: An illustration of the inclusion-exclusion principle for three proper-

ties.
is subtracted once from the total number of elements. The number of elements
in (P1 ∩ P2 ) \ P3 is subtracted twice in the beginning (since the elements are
in P1 as well as in P2 ) and then added back once. The number of elements
46
in P1 ∩ P2 ∩ P3 is subtracted three times in the beginning, then added back
three times and finally subtracted yet again. The same holds for the other
intersections. Altogether one can see that each element in P1 ∪P2 ∪P3 contributes
exactly −1 to the total sum.
2.1.1 Applications
In the following we apply the principle of inclusion and exclusion (PIE) to count
things. The arguments have a fairly rigid pattern:
(i) Define “bad” properties: We identity the things to count as the elements
of some universe X except for those having at least one of a set of prop-
erties P1 , . . . , Pm . The corresponding sets are denoted by X1 , . . . Xm (i.e.
Xi is the set of elements having property Pi ). Given this reformulation of
our problem we want to count X \ (X1 ∪ . . . ∪ Xm ).
(ii) Count N(S): For each S ⊆ [m], determine N (S), the number of elements
of X having all bad properties Pi for i ∈ S.
(iii) Apply PIE: Apply Theorem 2.1, i.e. the principle of inclusion and ex-
clusion. This yields a closed formula for |X \ (X1 ∪ . . . ∪ Xm )|, typically
with one unresolved summation sign.
Theorem 2.3 (Surjections). The number of surjections from [k] to [n] is:
Xn
n
(−1)i (n − i)k .
i=0
i
Proof. Define bad properties: Let X be the set of all maps from [k] to [n].
Define the “bad” property Pi for i ∈ [n] as “i is not in the image of f ”,
i.e.
f : [k] → [n] has property Pi :⇐⇒ ∀j ∈ [k] : f (j) 6= i.
With this definition, the surjective functions are exactly those functions
that have no bad property, i.e. we need to count X \ (X1 ∪ . . . ∪ Xn ).
Count N(S): We claim N (S) = (n−|S|)k , for any S ⊆ [n]. To see this, observe
that f has all properties with indices from S if and only if f (i) ∈
/ S for all
i ∈ [k]. In other words, f must be a function from [k] to [n] \ S and there
are (n − |S|)k of those.
Apply PIE: Using Theorem 2.1, the number of surjections is therefore:
X
(−1)|S| N (S)
pie
X \ (X1 ∪ . . . ∪ Xn ) =
S⊆[n]
X
= (−1)|S| (n − |S|)k
S⊆[n]
n
X
n i
= (−1) (n − i)k .
i=0
i
In the last step we used that (−1)|S| (n − |S|)k only depends on the size of
S and there are ni sets S ⊆ [n] of size i.
47
Corollary 2.4. (i) Consider the case n = k. A function from [n] to [n] is a
surjection if and only if it is a bijection. Since there are n! bijections on
[n] (all permutations) we obtained the identity:
X n
i n
n! = (−1) (n − i)n .
i=0
i
(ii) A surjection from [k] to [n] can be seen as a partition of [k] into n non-
empty distinguishable parts (the map assigns a part to each i ∈ [k]). Since
the partitions of [k] into n non-empty indistinguishable parts is counted by
sII
n (k) and there are n! ways to assign labels to the n parts, we obtain that
the number of surjections is equal to n!sII n (k). This proves the identity:
Xn
i n
n!sII
n (k) = (−1) (n − i)k .
i=0
i
Theorem 2.5 (Derangements revisited). Recall that for n ∈ N, the derange-

ments Dn on n elements are permutations of [n] without fixed points. We claim:
Xn
n
|Dn | = (−1)i (n − i)!
i=0
i
Proof. Define bad properties: Let X be the set of all permutations of [n].
We define the “bad property” Pi that means “π has a fixpoint i”:
π ∈ X has property Pi :⇐⇒ π(i) = i, (i ∈ [n]).
Derangements are exactly permutations that have none of these properties.
Count N(S): We claim N (S) = (n − |S|)! for any S ⊆ [n].
Indeed, π ∈ X has all properties with indices from S if and only if all
i ∈ S are fixed points of π. On the other elements, i.e. on [n] \ S, π may
be an arbitrary bijection so there are (n − |S|)! choices for π.
Apply PIE: Using Theorem 2.1, the number of derangements is therefore:
X
(−1)|S| N (S)
pie
X \ (X1 ∪ . . . ∪ Xn ) =
S⊆[n]
X
= (−1)|S| (n − |S|)!
S⊆[n]
n
X
n i
= (−1) (n − i)!
i=0
i
In the last step we used that (−1)|S| (n − |S|)! only depends on the size of
S and there are ni sets S ⊆ [n] of size i.
Theorem 2.6 (Combinations of multisets). Consider a multiset M with types
1, . . . , m and repetition numbers r1 , . . . , rm . Then the number of k-combinations
of M is: P
X
|S| m − 1 + k − i∈S (ri + 1)
(−1)
m−1
S⊆[m]

where we define ab := 0 for a < b.
48
Proof. Define bad properties: Let X be the set of k-combinations where we
disregard the restrictions the repetition numbers impose, in other words
X is the set of k-combinations of M f, where Mf is the multiset with the
same m types as in M but infinite supply of each type (i.e. “ri = ∞” for
each i ∈ [m]).

Recall that by Theorem 1.9, |X| = m−1+k m−1 . Define the bad property Pi
as:
A ∈ X has property Pi :⇐⇒ type i is repeated in A at least ri + 1 times.
Then the k-combinations of M are exactly those k-combinations in X that

have none of the bad properties.
Count N(S): We claim that for any S ⊆ [m]:
P
m − 1 + k − i∈S (ri + 1)
N (S) = .
m−1
To see this, note that the repetition number of each type i ∈ P S is at least
ri +1. This fixes ri +1 elements of the k-combination. In total
P i∈S (ri +1)
are fixed this way and what remains to be chosen is a (k − i∈S (r + 1))-
P i
combination of M f. Again by Theorem 1.9 there are m−1+k− i∈S (ri +1)
P m−1
such combinations if k − i∈S (ri + 1) ≥ 0 and no such combinations
otherwise.
Apply PIE: Using Theorem 2.1, the number of k-combinations of M is there-
fore:
X
(−1)|S| N (S)
pie
X \ (X1 ∪ . . . ∪ Xm ) =
S⊆[m]
X P
|S| m − 1 + k − i∈S (ri + 1)
= (−1)
m−1
S⊆[m]
In the special case where all the repetition numbers are equal, i.e. r1 = r2 =
. . . = rm = r, this can be simplified to:
Xm
i m m − 1 + k − (r + 1) · i
(−1) .
i=0
i m−1
Before we study a more advanced example, we prove a small result needed there:
2n 2n−r

Lemma. There are 2n−r r binary sequences (with letters 0 and 1) such
that:
• The sequence has length 2n and exactly r copies of 1 and 2n − r copies of
0.
• No two copies of 1 are adjacent. Here the first and the last position of the
sequence count as adjacent (the sequence is cyclic).
For example, if n = 3 and r = 2 there are 9 such sequences, namely:
101000, 100100, 100010, 010100, 010010, 010001, 001010, 001001, 000101.
49
Proof. Since no two copies of 1 may be adjacent, we know that after each 1
there must be a 0. So we can just imagine that one 0 is already “glued” to
every 1 and we are actually building a sequence with 2n − 2r copies of 0 and r
copies of 10.
However, one copy of 10 might “spill” across the border, i.e. the 1 could be
in position 2n and the 0 in position 1. We need to handle this case separately.
Case 1: The last position of the sequence is 1. Then the first position is 0 and
the remaining positions contain r − 1 copies of 10 and 2n − 2r copies of
0, now without any further pitfalls. There are 2n−r−1
r−1 ways to arrange
them.
Case 2: The last position of the sequence is 0. Then our special character
10 does not spill across the border and the sequence is any ordered ar-
rangement of r copies of 10 and 2n − 2r copies of 0. There are 2n−r
r of
them.
In total we have:

2n − r − 1 2n − r r 2n − r 2n − r 2n 2n − r
+ = + = .
r−1 r 2n − r r r 2n − r r
Example 2.7 (Problème des ménages). There are n ≥ 2 married couples

at a dinner party, a husband and a wife each, denoted by H1 , . . . , Hn and
W1 , . . . , Wn . They should be seated on a round table with 2n (distinguishable)
seats such that:
• Men and women alternate, i.e. no two men and no two women sit next
to each other. Equivalently, the seats with even labels are used either
exclusively by women or exclusively by men.
• All couples are separated, i.e. no husband sits next to his wife.
We want to count how many such assignments exist.
First count the number of ways to seat the women. They may be seated in
the even or odd seats and once this is fixed, there are n! ways to seat them, so
2 · n! ways in total.
Assuming the women are already seated, we now count the number of ways
the men can be added. It is easy to see that this number does not depend on
how the women are seated so we can assume without loss of generality that
wife Wi sits in seat 2i and the odd-numbered seats are still free. We count the
number of ways the men can join using PIE.
Define bad properties: Let X be the set of all ways to seat the men without
paying attention to the rule that they must not sit next to their wives.
There are |X| = n! ways to do it.
We define the bad property Pi that captures that the husband Hi sits next
to his wife Wi , i.e. on seat 2i − 1 or on seat 2i + 1 (all seat numbers are
taken modulo 2n, in particular, 2n and 1 are adjacent).
The permitted arrangements are exactly those with none of the bad prop-
erties.
50
Count N (S): This time, calculating N (S) for arbitrary S ⊆ [n] is tricky. In-
stead we calculate
X
N ∗ (r) := N (S), for 0 ≤ r ≤ n.
S⊆[n]
|S|=r
which will be just as helpful. The meaning of this number is a bit subtle:
N ∗ (r) is the number of pairs (π, S) where S ⊆ [n] is of size r and π is a
seating plan such that (wife Wi sits at 2i and) the couples with indices in
S are not separated. To better count these pairs, define a map f :
Under f , the pair (π, S) is mapped to a binary sequence with 2n characters
and exactly |S| copies of 1: For i ∈ S (remember that Hi and Wi will be
assigned adjacent places in π) a one should be put in the position of the
husband Hi – if he sits left of his wife – or in the position of Wi in π, if
she sits left of her husband.
It is time for an example. Consider the pair (π, S) with S = {2, 5} and
the seating
1 2 3 4 5 6 7 8 9 10
π = (H5 W1 H2 W2 H1 W3 H4 W4 H3 W5 ).
Note that the second and fifth couple are indeed not separated (W5 and
H5 are adjacent because the table is circular). Also note that the fourth
couple is also not separated, this is allowed. The mapping f discussed
above would map this pair to the sequence 0010000001 since H2 (in seat
3) and W5 (in seat 10) sit left of their spouse. It is clear that f will
never produce sequences with two adjacent ones: That would mean two
adjacent people are both sitting left of their spouse: Which is impossible
(assuming n ≥ 2). However, any sequence with r copies of 1 and no two
adjacent copies of 1 is the image of (n − r)! pairs (π, S): From the r copes
of 1, the set S (of size r) can be reconstructed as can be the position
of the husbands with indices in S. The (n − r) other husbands can be
distributed arbitrarily onto the (n − r) remaining odd-numbered seats,
there are (n − r)! ways to do so. This proves, using the Lemma above to
count the number of binary sequences of length n with r copies of 1 and
no two adjacent copies of 1:

∗ 2n 2n − r
N (r) = (n − r)! · .
2n − r r
Apply PIE: Using Theorem 2.1, the number of valid ways to arrange the hus-
51
bands between the wives is:
X
(−1)|S| N (S)
pie
X \ (X1 ∪ . . . ∪ Xn ) =
S⊆[n]
n X
X
= (−1)|S| N (S)
r=0 S⊆[n]
|S|=r
n
X X n
X
= (−1)r N (S) = (−1)r N ∗ (r)
r=0 S⊆[n] r=0
|S|=r
n
X
r 2n 2n − r
= (−1) (n − r)! · .
r=0
2n − r r
Multiplying this with 2n!, the number of ways to place the wives the total
number of seatings is (for n ≥ 2):
n
X
r 2n 2n − r
2n! (−1) (n − r)! · .
r=0
2n − r r
2.1.2 Stronger Version of PIE

Theorem 2.1 can be strengthened as follows:
Theorem 2.8 (Stronger PIE). Assume f, g : 2[n] → R are functions mapping
subsets of [n] to real numbers and g can be written in terms of f as:
X
g(A) = f (S) (for A ⊆ [n]).
S⊆A
Then f can also be written in terms of g like this:

X
f (A) = (−1)|A|−|S| g(S) (for A ⊆ [n]).
S⊆A
Before we proof Theorem 2.8, we first observe how it is indeed a generalization

of Theorem 2.1. So assume we have the setting of Theorem 2.1, i.e. properties
P1 , . . . , Pm with corresponding sets X1 , . . . , Xm .
Then for S ⊆ [m] define f (S) to be the number of elements having all
properties with indices not in S and none of the properties with indices in S,
i.e. \ [

f (S) := Xi \ Xi
i∈[m]\S i∈S
S
Note that f ([m]) = |X| − | i∈[m] Xi | is precisely the number we want to
count in Theorem 2.1.
The function g fulfilling the requirement of Theorem 2.8 is given as:
X \
g(A) = f (S) = Xi = N ([m] \ A)
S⊆A i∈[m]\A
52
P
To see the second “=”, note that S⊆A f (S) counts an element x if and only
if the set S of all properties that x does not have is a subset of A. This means
x is counted if and only if it has all properties from [m] \ A.
We now apply Theorem 2.8 to obtain:
X X
f ([m]) = (−1)m−|S| g(S) = (−1)m−|S| N ([m] \ S)
S⊆[m] S⊆[m]
X
= (−1)|S| N (S).
S⊆[m]
where in the last step we changed the order of summation, i.e. summed over
[m] \ S instead of S. This concludes the proof of (Thm. 2.8 ⇒ Thm. 2.1).
Theorem 2.8 establishes a similar result as Möbius Inversion which we con-

sider later. In both cases there are two “linked” functions f and g where g is
given in terms of f . The Theorems assert that f can then also be given in terms
of g. The subset relation “⊆” will be replaced by the notion of divisibility “|”.
Both are order relations and, in fact, there are results even more general than
both Theorem 2.8 and Möbius Inversions, dealing with general partial orders
(but we will not consider them here).
We now proof Theorem 2.8.
Proof. Consider (for A ⊆ [n]) the term that is claimed to be equal to f (A):
X X X X
(−1)|A|−|S| g(S) = (−1)|A|−|S| f (T ) = cT f (T )
S⊆A S⊆A T ⊆S T ⊆A
for appropriate cT (to be determined!), that captures how often f (T ) is en-

countered (for T ⊆ A). Observe cA = 1, since f (A) is only encountered for
T = S = A. For a proper subset T ⊂ A (and k := |A| − |T |) we have:
X k
X
|A|−|S| k−i k
cT = (−1) = (−1) = 0.
i
T ⊆S⊆A i=0
where in the second step we observed that picking a set between T and A is
equivalent to picking a subset of A \ T . The last step is an identity we already
saw.
2.2 Möbius Inversion Formula

Any positive number n ∈ N has a unique decomposition into primes, i.e. n can
be written as n = pa1 1 · pa2 2 · . . . · pakk for primes p1 , . . . , pk with multiplicities
a1 , . . . , ak . In other words, the factors of n are given as a multiset M (n) with
types p1 , . . . , pk and repetition numbers a1 , . . . , ak .
For example 360 = 23 · 32 · 5 and M (360) = {2, 2, 2, 3, 3, 5} = {3 · 2, 2 · 3, 1 · 5}
(which admittedly looks a bit quaint written that way).
We write n|m if n divides m, meaning there is a number a ∈ N such that
a · n = m. In terms of multisets this means M (n) ⊆ M (m).
53
The greatest common divisor (gcd) of m and n, is the number k = gcd(m, n),
with M (k) = M (m)∩M (n). For example gcd(12, 90) = 6 since M (6) = {2, 3} =
{2, 2, 3} ∩ {2, 3, 3, 5} = M (12) ∩ M (90).
Note that M (1) = ∅ (1 is the result of the empty product). If gcd(m, n) = 1,
then m and n are called relatively prime (or coprime).
We define the Euler’s φ-function for n ≥ 2 as:
φ(n) = #{k ∈ N | 1 ≤ k ≤ n, gcd(n, k) = 1}
For example, φ(12) = #{1, 5, 7, 11} = 4, φ(9) = #{1, 2, 4, 5, 7, 8} = 6. A

formula to compute Euler’s φ-function is given in the following Theorem.
Theorem 2.9. If n = pa1 1 pa2 2 . . . pakk then
k
Y 1
φ(n) = n (1 − )
i=1
pi
Consider for instance n = 12 = 22 · 3. Then φ(12) = 12 · (1 − 12 ) · (1 − 13 ) =

12 · 12 · 32 = 4 confirming what we counted by hand before.
Proof. In the proof we use the principle of inclusion and exclusion.
Define bad properties: As ground set we use X = [n]. We say a number

x ∈ X has the bad property Pi if x is divisible by pi , i.e.:
x ∈ [n] has property Pi :⇐⇒ pi |x. (i ∈ [k]).
Then the numbers in [n] that are relatively prime to n are exactly those
that have no bad property.
Count N(S): For any S ⊆ [k] we have N (S) = Q n
i∈S pi .
This is because, if x ∈ X is a multiple of all primes

Q with indices from S,
then x must be a multiple of their product i∈S pi (which is the least
common multiple of those primes). Since this product divides n, there are
no rounding issues
Q and the number of numbers x between 1 and n that
are multiples of i∈S pi is just given as the quotient.
Apply PIE: Using Theorem 2.1, we can write φ(n) as:
X X (−1)|S| Y
(−1)|S| N (S) = n ·
pie 1
φ(n) = Q =n (1 − pi ).
S⊆[k] S⊆[k]i∈S pi i∈[k]
The last identity is best seen by multiplying out the right side: For each
of the k factors we can either choose 1 or we can choose − p1i . The indices
i where − p1i was chosen are captured in the set S. For each S ⊆ [k] we
get exactly the term under the sum.
We now show a few more number theoretic results leading up to Möbius Inver-
sions.
P
Theorem 2.10. n = φ(d).
d|n
54
Proof. We claim that for a divisor d of n the sets {x ∈ [n] | gcd(x, n) = d} and
{y ∈ [ nd ] | gcd(y, nd ) = 1} have the same cardinality. Indeed, it is easy to see
that x 7→ y := xd is a bijection.
This means #{x ∈ [n] | gcd(x, n) = d} = φ( nd ). Summing these identities
for all d|n yields: X
n= φ( nd )
d|n
This is almost identical to the claim, the only thing left to do is to change the
order of summation: Substitute d with d0 := nd and note that d0 divides n iff d
divides n.
We now define the Möbius Function for d ≥ 1 as:


 1 d is the product of an even number of distinct primes
µ(d) := −1 d is the product of an odd number of distinct primes


0 otherwise
For example, µ(15) = 1, µ(7) = µ(30) = −1, µ(12) = 0, since 15 = 3 · 5 is the
product of two distinct primes, 7 = 7 and 30 = 2 · 3 · 5 are the product of an
odd number of primes and 12 = 2 · 2 · 3 is not the product of distinct primes:
We need 2 twice. The numbers n with µ(n) 6= 0 are also called square free since
they do not have a square as a factor (12 is not square-free since it has 4 as a
factor).
Note that µ(1) = 1 since 1 is the product of 0 primes and 0 is even.
Lemma 2.11. (
X 1 if n = 1
µ(d) =
d|n
0 6 1
if n =
Proof. If n = 1 this is obvious since d = 1 is the only divisor of 1 and µ(1) = 1.

For n ≥ 2 we write n = pa1 1 . . . pakk . Then we have:
X X X k
X
|D| k
µ(d) = µ(d) = (−1) = (−1)i = 0.
i=0
i
d|n d|(p1 p2 ...pk ) D⊆[k]
In the first step remember that µ(d) = 0 if d is not square-free. This means if
d contains a prime factor more than once, it does not contribute to the sum.
For the second step, realize that a divisor d of p1 . . . pk is just given by choosing
a subset D of those k primes and multiplying them. Then µ(d) will be 1 if an
even number of primes where chosen and −1 otherwise. The last two steps are
identities we already encountered earlier.
Corollary 2.12.
φ(n) X µ(d)
= .
n d
d|n
Proof. We use an identity that came up in the proof of Theorem 2.9 and argu-
ments similar to those from the last Theorem.
k
Y X X X µ(d)
1 −1|S| µ(d)
φ(n) = n (1 − pi ) =n Q =n =n .
i=1 S⊆[k] i∈S pi d|p1 ...pk
d d
d|n
55
We now have all ingredients to prove the Möbius Inversion Formula:
Theorem 2.13 (Möbius Inversion). P
Let f, g : N → R be functions satisfying g(n) = f (d). Then:
d|n
X
f (n) = µ(d)g( nd ).
d|n
n
Proof. We start by changing the order of summation (d → d) and using the
definition of g.
X X n
µ(d)g( nd ) = µ( )g(d)
d
d|n d|n
X X
= µ( nd ) f (d0 )
d|n d0 |d
X
= cd0 · f (d0 )
d0 |n
Where cd0 are numbers (to be determined!) that count how often f (d0 ) occurs
as a summand. Note that cn = µ(1) = 1 since f (n) occurs only for d0 = d = n.
For d0 |n, d0 6= n we have:
X X
cd0 = µ( nd ) = µ(m) = 0.
d0 |d|n m| dn0
n
Where we substituted the summation index d 7→ m := d and used Lemma 2.11
in the last step. This proves the claim.
Example 2.14 (Circular Sequences of 0’s and 1’s). Circular sequences are for
example:
0 1 1 0 1 0
0 1 0
1 0 0
A= 0, B= 1, C= 1
1 0 0
0 0 0
0 1 0 1 1 0
Circular sequences are considered equal if they can by transformed into one
another by rotation. We would write
A = 001011010 = 110100010 = B 6= C = 100100100.
Let Nn be the number of circular 0/1-sequences of length n and M (d) the

number of aperiodic circular sequences of length d where a sequence is called
aperiodic if it cannot be written as several times a shorter sequence. For exam-
ple, C is not aperiodic since it consists of three copies of “100” while A (and
therefore B) are aperiodic.
For every circular sequence S of length n there is a unique way to describe
it as “repetitions of S 0 ” where S 0 is an aperiodic circular sequence. This is not
56
trivial, but also not hard to see. Clearly, the length of S 0 must divide n. So we
found:
X
Nn = M (d). (?)
d|n
Consider Table 2 for an example with n = 6.
d M(d) corresponding sequences

1 0,1 000000,111111
2 01 010101
3 001,011 001001,011011
6 000001,000011,000101, 000001,000011,000101,
000111,001011,010011, 000111,001011,010011,
001111,010111,011111 001111,010111,011111
Table 2: Aperiodic sequences of all lengths that divide 6. There is a bijection

between them and the circular sequences of length 6.
Another crucial observation is that the number of aperiodic linear sequence

of length d is just d · M (d) (every rotation of an aperiodic circular sequence)
and therefore the number of all linear sequences of length n can be written as
“repetition of some aperiodic linear sequence” so as:
X
2n = d · M (d). (∆)
d|n
Now define g(n) = 2n and f (d) = d · M (d). These choices of f and g fulfill
the requirement of the Möbius Inversion Formula so we obtain:
2.13
X
f (n) = µ(d)g( nd )
d|n
X n
⇔ nM (n) = µ(d)2 d
d|n
(?) X X X d
X X
so Nn = M (d) = 1
d µ(l)2 l = 1
d µ( dl )2l
d|n d|n l|d d|n l|d
XX X X X X
1 d l 1 d l 1 l
= d µ( l )2 = d µ( l )2 = k·l µ(k)2
l|n l|d|n l|n 1| dl | nl l|n k| nl
(d:=k·l)
X 2l X µ(k) 2.12 X 2l φ( n ) X 2l
= = l
= φ( n ).
l n k l nl n l
l|n k| l l|n l|n
While not overwhelmingly pretty, at least our result is an explicit formula only
involving one sum and Euler’s φ-function. This is as good as it gets.
57
3 Generating Functions
In the following we consider sequences (an )n∈N = a0 , a1 , a2 , . . . of non-negative
numbers. Typically, ak is the number of discrete structures of a certain type
and “size” k. P∞
The generating function for (an )n∈N is given as F (x) = n=0 an xn . Despite
the name, a generating function should not be thought of as a thing you plug
values into: It is not meaningful to compute something like F (5): In fact, these
values are often not well-defined because the sum would be divergent. We care
about the thing as a whole: You can either think of generating functions as
functions that are well-defined within their radius of convergence (some area
close to zero) or you just think of the “x” in F (x) as an abstract thing (not a
placeholder for a number) which makes F (x) an element of the ring of formal
power series. In the following, we ignore all technicalities and boldly apply an-
alytic methods as though a generating function were just a simple well-behaved
function. And it just works. If you think this is all a bit arcane, try to look at
it this way:
At first, a generating function is just a silly way to write a sequence (instead
of a0 = 1, a1 = 42, a2 = 23, . . . you would write A(x) = 1 + 42x + 23x2 + . . .).
Some operations on the sequence have a natural correspondence: For example,
shifting the sequence (a0 = 0, a1 = 1, a2 = 42, a3 = 23 . . .) corresponds to
multiplying A(x) by x. Some complicated operations on sequences suddenly
become simple and natural in the world of generating functions where analytic
tools are readily available. P∞ n
P∞When ndealing with generating functions F (x) = n=0 an x and G(x) =
n=0 bn x , we freely use the following operations:
• Differentiate F term-wise
∞
X ∞
X
F 0 (x) = nan xn−1 = (n + 1)an+1 xn
n=1 n=0
• Multiply F (x) by a scalar λ ∈ R term-wise

∞
X
λF (x) = λan xn .
n=0
• Add F (x) and G(x)

∞
X
F (x) + G(x) = (an + bn )xn .
n=0
• Multiply F (x) and G(x)

∞ n
!
X X
F (x) · G(x) = ak bn−k xn .
n=0 k=0
We will later see that the coefficients of the product sometimes count
meaningful things if an and bn did (see Example 3.2 and 3.3).
58
Example 3.1 (Maclaurin series).
(i) Consider (an )n∈N with an = 1 for all n ∈ N. The corresponding generating
function F (x) = 1 + x + x2 + x3 + . . . is called the Maclaurin series.
1
We claim F (x) = 1−x , which looks a lot like the identity for an infinite
geometric series. To verify this, just observe that the product of F (x) and
(1 − x) is one:
(1 − x)F (x) = (1 + x + x2 + x3 + . . .) − (x + x2 + x3 + . . .) = 1.
1
(ii) Differentiating F (x) = 1−x with respect to x on both sides yields:
∞
X 1
F 0 (x) = (n + 1)xn = .
n=0
(1 − x)2
So we found the generating function for (bn )n∈N with bn = n + 1.

(iii) Substituting x = −y in the Maclaurin series gives the generating function
of the alternating sequence (i.e. an = (−1)n ):
∞
X
1
= (−1)n y n .
1 + y n=0
How do we count using Generating Functions?

(k)
Example 3.2. Let an be the number of arrangements of n unlabeled balls in
k labeled boxes and at least 1 ball per box. For each k, this gives a generating
function:
∞
X
B (k) (x) = a(k) n
n x .
n=0
Considering the case with only one box, we have

(
0 for n = 0
a(1)
n =
1 for n ≥ 1.
So B (1) (x) is just the Maclaurin Series from 3.1(i) shifted by one position (i.e.
multiplied by x), meaning:
∞
X x
B (1) (x) = 0 + x + x2 + x3 + . . . = x · xn = .
n=0
1−x
The key observation we make now is that multiplying two generating functions
has a meaningful correspondence in our balls-in-boxes setting:
For two numbers of boxes s and t, we have the identity:
n
X (s) (t)
an(s+t) = al an−l
l=0
This is merely the observation that arrangements of n balls in s + t boxes are

given by arranging some l balls in the first s boxes and the remaining n − l balls
59
in the remaining t boxes. Looking at the right side of the equation, note that
these numbers are exactly the coefficients of the product B (s) (x) · B (t) (x) (check
this!). So we obtained:
B (s+t) (x) = B (s) (x) · B (t) (x)
and therefore
k x k
(k) (1)
B (x) = B (x) = .
1−x
P (k)
However, we want to write B (k) (x) in the form n an xn to see the coefficients
(k)
an . To do this, note that deriving k − 1 times the term (1 − x)−1 yields
(k − 1)!(1 − x)−k . Using this we obtain:
k−1 ∞
!
(k) xk xk d 1 xk dk−1 X n
B (x) = = · = x
(1 − x)k (k − 1)! dxk−1 1 − x (k − 1)! dxk−1 n=0
∞
X
xk
= n · (n − 1) · . . . · (n − k + 2)xn−k+1
(k − 1)!
n=k−1
∞
X n!
= xn+1
(n − k + 1)!(k − 1)!
n=k−1
X∞ X∞
n n−1 n
= xn+1 = x
k−1 n=0
k−1
n=k−1

where in the last step use that ab = 0 if a < b and additionally make an index
shift of 1.
So we found a new proof that the number of arrangements of n unlabeled
(k)
balls in k labeled boxes and at least one ball per box is an = n−1
k−1 .
Example 3.3. We want to count integer solutions for a + b + c = n with a non-

negative even integer a, a non-negative integer b and c ∈ {0, 1, 2}. Equivalently,
we can think of this as counting arrangements of n unlabeled balls in boxes
labeled a, b and c where box a should receive an even number of balls and c at
most 2 balls.
We first consider the more simple situations where only one of the vari-
ables/boxes exists:
• Solutions to a = n with even a. Clearly, there is a unique solution for

even n and no solution for odd n. The corresponding generating function
is:
X∞ X∞
1
A(x) = x2n = (x2 )n = .
n=0 n=0
1 − x2
• Solutions to b = n where b is an integer. Clearly, there is exactly one so-

lution for each n. The corresponding generating function is the Maclaurin
series:
X∞
1
B(x) = xn = .
n=0
1−x
60
• Solutions to c = n where c ∈ {0, 1, 2}. Clearly, there is exactly one solution
if n ∈ {0, 1, 2} and no solution otherwise. The corresponding generating
function is:
C(x) = 1 + x + x2 .
With the same argument as in the previous example, multiplying the generating
functions yields the generating function of the sequence we are interested in, so
we calculate:
(1 + x + x2 ) (1 + x + x2 )
F (x) = A(x) · B(x) · C(x) = =
(1 − x2 )(1 − x) (1 + x)(1 − x)2 .
We would like to write this as a linear combination of generating function we
understand well, so we search for real numbers R, S, T with:
(1 + x + x2 ) R S T
F (x) = = + +
(1 + x)(1 − x)2 1 + x 1 − x (1 − x)2
Multiplying by the denominators yields:
(1 + x + x2 ) = R(1 − x)2 + S(1 − x2 ) + T (1 + x)
We compare the coefficients in front of 1, x and x2 and get the equations 1 =

R + S + T , 1 = −2R + T , 1 = R − S. This system has the unique solution
R = 41 , S = − 34 , T = 32 .
Using this and the identities we obtained in Example 3.1 yields:
∞
X ∞
X ∞
X
1 1 1 n n n
F (x) = 1
4 −3 +3 = 1
(−1) x − 43 x + 23 (n+1)xn .
1 + x 4 1 − x 2 (1 − x)2 4
n=0 n=0 n=0
So the coefficient of xn , and therefore the number of solutions to a + b + c = n

we want to count is:
(−1)n 3 3(n + 1)
− + .
4 4 2
Example 3.5. We want to count the number an of well-formed parenthesis
expressions with n pairs of parenthesis. For example (()())() is a well-formed
expression with 4 pairs of parenthesis but )())(( is not. Formally, a permutation
of the multiset {n · “(”, n · “)”} is well-formed if reading it from left to right and
counting “+1” for every “(” and “-1” for every “)” will never yield a negative
number at any time (no prefix contains more closing than opening parenthesis).
Every well-formed expression with n ≥ 1 pairs of parenthesis starts with “(”
and there is a unique matching “)” such that the sequence in between and the
sequence after is a well-formed (possibly empty) expression. For example:
(())()() (()(())) ()(()(()))

In other words, a well-formed expression with n pairs of parenthesis is obtained
by putting a well-formed expression with k pairs in between “(” and “)” and
then append a well-formed expression with n − k − 1 pairs of parenthesis. This
gives the equation:
n−1
X
an = ak an−k−1
k=0
61
So if F (x) is the generating function belonging to an , then we know:
∞
X ∞ n−1
X X ∞ X
X n
F (x) = an xn = 1 + ( ak an−k−1 )xn = 1 + ( ak an−k )xn+1
n=0 n=1 k=0 n=0 k=0
∞ X
X n
=1+x· ( ak an−k )xn = 1 + x · F (x)2 .
n=0 k=0
And with analytic methods we can find a solution to this as:

√
1 − 1 − 4x
F (x) = .
2x
So far it is unclear what this means for the coefficients of F (x) (and therefore
for the number of well-formed expressions).
√ It seems we need additional tools
to understand the power series of 1 − 4x.
3.1 Newton’s Binomial Theorem

Recall the following special case of the Multinomial Theorem (Theorem 1.6).
Xn ∞
n n k X n k
(1 + x) = x = x , ∀n ∈ N.
k k
k=0 k=0

Where, as usual, nk = 0 for k > n. This shows that (1 + x)n is the generating
n
function for the series (ak )k∈N with ak = k .
We extend this result from natural numbers n ∈ N to any real number n ∈ R.
To this end we first extend the definition of binomial coefficients.
Definition 3.7 (Binomial Coefficients for Real Numbers). Recall that for in-
tegers n, k ∈ N we have:

n |P (n, k)| n(n − 1) · · · · · (n − k + 1) n!
= = =
k k! k! k!(n − k)!
It does not make sense to talk about permutations of sets of size n ∈ R and it is
unclear what n! should be, but the formula in between is well-defined for general
n ∈ R. With this in mind, we define p(n, k) := n · (n − 1) · . . . · (n − k + 1) and

n
k
:= p(n,k)
k! . Note that the new definition matches the old one if n is integer.
We can now talk about numbers such as “−7/2 choose 5” by which me mean:
−7 −9 −11 −13 −15
−7/2 2 · 2 · 2 · 2 · 2 9009
= =− .
5 5! 256
Note that for n ∈ R we have p(n, 0) = 1 and for k ≥ 1 the recursions:
p(n, k) = (n − k + 1) · p(n, k − 1) (?)

= n · p(n − 1, k − 1).
Given our extended definition of binomial coefficients, we can state the following
Theorem (but will omit the proof).
62
Theorem 3.8 (Newton’s Binomial Theorem). For all non-zero n ∈ R we have:
∞
X
n n
(1 + x) = xk .
k
k=0
√
Setting n = 1/2 yields an identity for 1 + x that we require to proceed
in Example 3.5. But before we can use it, we need to better understand the
coefficients of the form 1/2
· .
Lemma 3.9. For any integer n ≥ 1 we have

1/2 n+1 2n − 2 1 1
= (−1) 2n−1
· .
n n−1 2 n
Proof. We do induction on n. For n = 1:

1
1/2 1 1
= 2 = =1·1· ·1
1 1! 2 2
For the induction step (n → n + 1) we use recursion (?) from above:

1/2 p(1/2, n + 1) (1/2 − (n + 1) + 1)p(1/2, n) n − 1/2 1/2
= = =−
n+1 (n + 1)! (n + 1) · n! n+1 n

IH n − 1/2 2n − 2 1 1
=− (−1)n+1 ·
n+1 n − 1 22n−1 n

2n 2n − 1 2n − 2 1 1
= · (−1)n+2 ·
2n 2n n − 1 22n−1 n + 1
(2n − 2)!(2n − 1)(2n) 1 1
= (−1)n+2 · .
(n − 1)!(n − 1) · n · n 22n+1 n + 1
| {z }
(2n
n)
Proposition 3.10. Using the last Theorem and Lemma we obtain:

∞
X ∞
X
√ 3.8 1/2 n 3.9 2n − 2 1 1
1+x = x = 1+ −2 (−1)n 2n · · xn .
n=0
n n=1
n − 1 2 n
Example 3.5 (Continued). Using the proposition we are now able to find the
coefficients of the generating function from Example 3.5 above, i.e. the number
of well-formed parenthesis expressions.
√ ∞
1 − 1 − 4x 3.10 1 X 2n − 2 1 1
F (x) = = 2 (−1)n 2n (−4x)n
2x 2x n=1 n−1 2 n
∞ ∞
1 X 2n − 2 1 n X 2n 1
= x = xn .
x n=1 n − 1 n n=0
n n + 1
1
The numbers Cn := 2n n n+1 are called Catalan Numbers. They do not only
count well-formed parenthesis expressions but occur in other situations as well.
63
3.2 Exponential Generating Functions
Until now we mapped sequences to functions like this:
∞
X
(an )n∈N 7→ an xn ∈ R[x]
n=0
In other words, we used the coefficients an to obtain a linear combination of the

basis {xn }n∈N of R[x].
Our choice of a basis was useful in the cases we considered, but we can
consider other bases as well that will be useful in other situations. The following
sets are all bases of R[x]:
n n

x −x x 1
p(x, n) e
n∈N n! n∈N n! n∈N nx n∈N
In the
P∞last case we get from any sequence (an )n∈N a corresponding Dirichlet
series n=0 anns . Such series are important in algebraic number theory (in that
setting the variable is typically called s instead of x). As an example, we state
the following result (without proof).
Theorem 3.11 (Euler Product). The Dirichlet series for (µ(n))n∈N satisfies
X∞ Y
µ(n) 1
= = (1 − p−s )
n=0
ns ζ(s) p prime
P∞ 1
where ζ(s) is the Riemann zeta function ζ(s) := s
.
n=0 n
We will not examine Dirichlet series further, dealing with exponential gener-
ating functions instead. In contrast to ordinary generating functions (the kind
n
we considered before) the basis is not {xn }n∈N but { xn! }n∈N and instead of
counting arrangements of unlabeled objects (unlabeled balls in labeled boxes,
solutions to a1 + a2 + . . . + al = n with constraints, indistinguishable parenthe-
sis in an ordered string), exponential generating functions are useful to count
arrangements of labeled objects (permutations, derangements, partitions, . . . )
as we will see shortly.
Before we get started, note the exponential generating functions A(x) and
B(x) of (an )n∈N and (bn )n∈N with an = 1 and bn = n! (n ∈ N) are
X∞
xn
A(x) = = ex
n=0
n!
∞
X X∞
xn 1
B(x) = n! = xn =
n=0
n! n=0
1 − x.
The following two observations show the connection of exponential generating

functions to arrangement of labeled objects.
Observation 3.12. Say the numbers (an )n∈N and (bn )n∈N count arrangements
of type A and B, respectively, using n labeled objects.
64
Assume arrangements of type C with n objects are obtained by a unique
split of the n objects into two sets and then forming an arrangement of type A
with the first set and an arrangement of type B with the second.
Then cn , the number of arrangements of type C and size n, is given as
n
X n
cn = ak · bn−k .
k
k=0
Crucially, the exponential generating functions A(x), B(x), C(x) of the three
sequences reflect this relationship as C(x) = A(x) · B(x), which is easy to verify:
∞
! ∞
! ∞ n
!
X xn X xn X X ak bn−k
A(x) · B(x) = an · bn = xn
n=0
n! n=0
n! n=0 k=0
k! (n − k)!
∞ n !
X X n x n
= ak · bn−k = C(x).
n=0
k n!
k=0
Observation 3.13. An index shift by one in the sequence corresponds to a

derivation, meaning
∞
X ∞
X ∞
X
xn derive xn−1 xn
F (x) = an ⇒ F 0 (x) = an = an+1 .
n=0
n! n=1
(n − 1)! n=0 n!
We already encountered Stirling numbers before. We distinguish three kinds:
• The unsigned Stirling numbers of the first kind sIk (n) fulfill the recursion
sIk (n) = (n − 1)sIk (n − 1) + sIk−1 (n − 1)
and count the number of n-permutations of [n] with k cycles.

• The signed Stirling numbers of the first kind are simply
s̄Ik (n) = (−1)n−k sIk (n).
• The Stirling numbers of the second kind sII

k (n) fulfill
sII II II
k (n) = ksk (n − 1) + sk−1 (n − 1)
and count the number of partitions of [n] into k non-empty parts.

Theorem 3.14. For k ∈ N, we have the following identity for the exponential
generating function Fk (x) of the Stirling numbers (sII
k (n))n∈N :
∞
X xn 1
Fk (x) = sII
k (n) = (ex − 1)k .
n=0
n! k!
Proof. We proceed by induction on k. For k = 1 we have:

(
II 1 for n ≥ 1
sk (n) =
0 for n = 0
65
P
∞
xn
and the claim follows from the identity n! = ex − 1.
n=1
If k ≥ 2, we use the recursion of Stirling numbers from above:
sII II II
k (n + 1) = ksk (n) + sk−1 (n).
Taking the generating functions for each term (remember that derivations cor-
respond to index shifts!) and then applying the induction hypothesis yields:
IH 1
Fk0 (x) = k · Fk (x) + Fk−1 (x) = kFk (x) + (ex − 1)k−1 .
(k − 1)!
This differential equation has the solution

1 x
Fk (x) = (e − 1)k
k!
as claimed, where we used Fk (0) = 0.
Note that solutions to differential equations may be hard to find but are easy
to verify. Finding the solutions is not the topic of this lecture. It is perfectly
alright if you use computers (or techniques from other lectures).
To later obtain the exponential generating function for Stirling numbers of

the first kind, we first prove two identities:
Lemma 3.15.
n
X
(i) sIk (n)xk = p(x + n − 1, n).
k=0
n
X
(ii) s̄Ik (n)xk = p(x, n).
k=0
Proof. Note that p(x + n − 1, n) is a polynomial of degree

Pn n with variable x so
there are constants bn,k such that p(x + n − 1, n) = k=0 bn,k xk . They fulfill
the following equation:
n
X
bn,k xk = p(x + n − 1, n) = (x + n − 1) · p(x + n − 2, n − 1)
k=0
n−1
X n
X n−1
X
= (x + n − 1) bn−1,k xk = bn−1,k−1 xk + (n − 1) bn−1,k xk .
k=0 k=1 k=0
Comparing the coefficients of xk we get:
bn,k = bn−1,k−1 + (n − 1)bn−1,k
where we define bn,k = 0 for k > n or k < 0.

So the numbers bn,k fulfill the same recursion as sIk (n) and since the starting
values sI0 (0) = b0,0 = 1 match as well, this proves bn,k = sIk (n) and thus (i).
66
For (ii), plug −x into the equation above:
n
X n
Y
(−1)k sIk (n)xk = p(−x + n − 1, n) = (−x + n − k)
k=0 k=1
n
Y
= (−1)n (x − n + k) = (−1)n p(x, n).
k=1
n
Multiplying by (−1) gives the desired result:
n
X
(−1)n−k sIk (n) xk = p(x, n).
| {z }
k=0
s̄Ik (n)
Theorem 3.16. The exponential generating function of (s̄Ik (n))n∈N fulfills:

∞
X xn 1
s̄Ik (n) = (log(1 + x))k .
n=0
n! k!
Proof. One way to expand (1 + x)z is:

∞
X 1
(1 + x)z = ez log(1+x) = (log(1 + x))k z k
k!
k=0
On the other hand, we can also expand (1 + x)z using Newton’s Binomial The-
orem and the last Lemma:
X∞ ∞ ∞ n
z 3.8 z n X p(z, n) n 3.15 X xn X I
(1 + x) = x = x = s̄k (n)z k
n=0
n n=0
n! n=0
n!
k=0
∞ X ∞ ∞ ∞
!
X n
x k X X xn
I I
= s̄k (n) z = s̄k (n) zk .
↑ n=0 n! n=0
n!
k=0 k=0
sIk (n)=0
for k > n
We have written (1 + x)z in two ways and the coefficients of z k in both repre-
sentations must match. This proves the claim.
3.3 Recurrence Relations

We have already seen (in the case of Stirling numbers) that relations between
numbers in a sequence may fully characterize the sequence (if some starting
value is given). A sequence (an )n∈N is defined recursively if
an = f (an−1 , an−2 , . . . , an−k )
for some function f , a positive integer k, and all n ≥ k. This identity an =

f (an−1 , an−2 , . . . , an−k ) is called a recurrence or a recurrence relation or a re-
cursion of order k. If a0 , a1 , . . . , ak−1 are known, they are called initial values
or initial conditions. The sequence (an )n∈N is then called a solution of the
recursion f with initial values a0 , . . . , ak−1 .
In the following we examine special cases of such recurrence relations and
show how the sequence can be derived from them.
67
Example 3.17 (Fibonacci Numbers). Rabbits were first brought to Australia
by settlers in 1788. Assume for simplicity, the first pair of young rabbits (male
and female) arrives in January (at time n = 1). A month later (n = 2) this
pair of rabbits reaches adulthood. Another month later (n = 3) they produced
a new young pair of rabbits as offspring. In general, assume that in the course
of a month every pair of young rabbits grows into adulthood and every pair of
adult rabbits produces one pair of young rabbits as offspring. If Fn denotes the
number of rabbits at time n, then we have Fn = Fn−1 + Fn−2 , since exactly
the rabbits that already existed at time n − 2 will be adults at time n − 1 and
produce offspring between time n − 1 and n. This gives sequence:
F0 = 0, F1 = 1, F2 = 1, F3 = 2, F4 = 3, F5 = 5, F6 = 8, F7 = 13, F8 = 21, . . .
The sequence (Fn )n∈N is called the Fibonacci sequence and its elements are
the famous Fibonacci numbers. Rabbits in Australia are now a serious prob-
lem, causing substantial damage to crops and have been combated with ferrets,
fences, poison, firearms and the myxoma virus.
Photos taken by Daniel Oines and Esdras Calderan
Figure 19: Spiraling patterns can be seen in many plants. The pine cone on the
left has 8 spiral arms spiraling out counterclockwise and 13 spiral arms spiraling
out clockwise. In the sunflower, different spiral patterns stand out depending
on where you look. Close to the center, a counterclockwise spiral with 21 arms
and a clockwise spiral with 34 arms can be seen. Closer to the border another
counterclockwise spiral with 55 arms emerges. Fibonacci numbers everywhere!
We’re not making this up, zoom in and count for yourselves (or better yet, go
outside and look at nature directly), the patterns are really there!
Fibonacci numbers also frequently occur in plants (see Figure 19). In case
you do not know Vi Hart (seriously?!) you should definitely check out her
Youtube Channel and watch her videos on this topic.
Another example where Fibonacci numbers occur is tilings of checkerboards

of size n × 2 with dominoes as shown in Figure 20. Let cn be the number of
such tilings. Note that there are two possibilities to cover the rightmost column
(see Figure 21: Either with a vertical domino, then a board of size (n − 1) × 2
68
Figure 20: A tiling of the 9 × 2 board with dominoes.
remains to be tiled, or with horizontal dominoes, then a board of size (n − 2) × 2

remains to be tiled. This means cn = cn−1 + cn−2 and together with c1 = 1 and
c2 = 2 (and setting c0 = 1) this implies cn = Fn+1 .
Figure 21: The last column can either be covered by a vertical domino or by
horizontal dominoes.
Example 3.18 (Ternary n-strings without substring 20). Consider strings

over the alphabet {0, 1, 2} that do not contain the substring 20, for instance
0010221021 or 1111011102. Let tn be the number of such strings of length n.
We have t0 = 1 (the empty string), t1 = 3 (each string of length 1), and t2 = 8
(all strings of length 2 except for 20).
In general, an n-string without 20 consists of an (n − 1)-string without 20
and an additional digit, which gives 3 · tn−1 possibilities. However, this may
produce strings that end in 20 (but have no 20 in other places): There are tn−2
of them.
This gives tn = 3 · tn−1 − tn−2 .
The Fibonacci sequence and the sequence (tn )n∈N from the last example have
in common that they are solutions for some recurrence relation of order 2, i.e.,
their elements can be described in terms of the previous two elements. Moreover,
each element could be expressed as the sum of multiples of previous elements.
We now develop a general theory dealing with such recurrence relations.
Definition 3.19. A recurrence relation of the form
an = c1 (n)an−1 + c2 (n)an−2 + · · · + ck (n)an−k + g(n)
for functions g : N → R and ci : N → R, i = 1, . . . , k, is called a linear recurrence.

In case g ≡ 0 (that is, g(n) = 0 for all n ∈ N), the recurrence is called linear
homogeneous, otherwise linear non-homogeneous.
In this chapter we shall treat the case when the coefficients ci ’s are constants,
so
an = c1 an−1 + c2 an−2 + · · · + ck an−k .
We assume that ck 6= 0, otherwise we treat the recursion as one of a smaller
order.
69
The Fibonacci sequence fulfills a homogeneous linear recurrence relation with
constant coefficients
Fn = Fn−1 + Fn−2
where k = 2, c1 = 1, c2 = 1, g ≡ 0.
The same is true for tn from Example 3.18
tn = 3tn−1 − tn−2
where this time k = 2, c1 = 3, c2 = −1, g ≡ 0.
3.3.1 Special Solution an = xn and the Characteristic Polynomial

Throughout this section assume we are given a linear homogeneous recurrence
relation f of the form
an = c1 an−1 + c2 an−2 + . . . + ck an−k
of order k. If one replaces ai with xi and plugs it in the recursion, the following
is obtained:
xn = c1 xn−1 + c2 xn−2 + · · · + ck xn−k .
Dividing by xn−k and bringing terms to one side of the equation gives:
xk − c1 xk−1 − c2 xk−2 − · · · − ck = 0.
Solving this equation allows to find some solutions. The characteristic polyno-
mial of f is given by
p(x) = xk − c1 xk−1 − c2 xk−2 − · · · − ck−1 x − ck .
We see that a homogeneous linear recursion could be reconstructed from its
characteristic polynomial by replacing xi with an−k+i :
an − c1 an−1 − c2 an−2 − · · · − ck an−k = 0,
k k−1 k−2
x − c1 x − c2 x − · · · − ck−1 x − ck = 0.
Note further, that we can assume that all the roots of the characteristic poly-
nomial are non-zero because ck 6= 0.
In this section we shall present some methods to solve linear homogeneous
recurrences and some specific non-homogeneous recurrences.
Lemma 3.20 (Linearity Property). Let f be a linear homogeneous recurrence
with constant coefficients. If (bn )n∈N is a solution of f with some initial values
and (dn )n∈N is a solution of f with some initial values, then for any constants
β, δ ∈ R, (βbn + δdn )n∈N is a solution of f with some initial values.
Proof. Let an = f (an−1 , . . . , an−k ) = c1 an−1 +· · ·+ck an−k . Thus bn = c1 bn−1 +
· · · + ck bn−k and dn = c1 dn−1 + · · · + ck dn−k for all n ≥ k. Multiplying the first
equation by β, the second by δ and adding them gives βbn + δdn = c1 (βbn−1 +
δdn−1 ) + · · · + ck (βbn−k + δdn−k ).
Lemma 3.21 (Divisibility of Polynomials). Let p(x) be a polynomial. Then p(x)
is divisible by (x − q)m , for a positive integer m and a nonzero constant q if and
only if the following polynomials are divisible by (x − q): p0 (x) = p(x), p1 (x) =
xp00 (x), p2 (x) = xp01 (x), . . . , pk (x) = xp0k−1 (x), . . . , pm−1 (x) = xp0m−2 (x).
70
Proof. Assume first that p(x) is divisible by (x−q)m . Then p(x) = (x−q)m t(x),
where t(x) is a polynomial. We use induction on k to show that (x − q)m−k
divides pk , k = 0, . . . , m − 1. When k = 0, (x − q)(m−0) divides p0 (x) = p(x).
Assume that (x − q)m−k+1 divides pk−1 (x). Let us show that (x − q)m−k divides
pk (x). We have then that pk−1 = (x−q)m−k+1 s(x) for some polynomial s. Then
pk (x) = xp0k−1 (x) = x (m − k + 1)(x − q)m−k s(x) + (x − q)m−k+1 s0 (x) , so pk
is divisible by (x − q)m−k .
Now, assume that each of p0 , . . . , pm−1 is divisible by x − q. We shall show
that pm−k (x) is divisible by (x − q)k , k = 0, . . . , m − 1, so in particular that
p0 (x) is divisible by (x − q)m . We do this by induction on k. When k = 1,
the statement holds by assumption. Now, assume that pm−k (x) is divisible
by (x − q)k , let us prove that pm−k−1 (x) is divisible by (x − q)k+1 . Assume
not, i.e., that pm−k−1 (x) = (x − q)` t(x), t(q) 6= 0, ` ≤ k. Then pm−k (x) =
x(x − q)`−1 (`t(x) + t0 (x)(x − q)), so the highest power of (x − q) dividing pm−k
is ` − 1 < k. We see that pm−k (x) is not divisible by (x − q)k .
Lemma 3.22 (Fundamental Solution). Let f be a linear homogeneous recur-
rence with constant coefficients. If q is a root of the characteristic polynomial
of f , then (q n )n∈N is a solution of f for some initial values. Moreover, if q is
a root of the characteristic polynomial of multiplicity s, s > 1, then (ni q n )n∈N ,
0 ≤ i ≤ s − 1 is a solution of f .
Proof. Let an = f (an−1 , . . . , an−k ) = c1 an−1 + · · · + ck an−k and bn = q n , where
q is a root of the characteristic polynomial, so q k − c1 q k−1 − · · · − ck q 0 = 0.
Multiply the last equality by q n−k to obtain q n − c1 q n−1 − · · · − ck q n−k = 0. So
(q n )n∈N is a solution of f .
Now, assume that q has multiplicity s, i.e., the characteristic polynomial
p1 (x) has a form p1 (x) = (x − q)s p(x), where p(x) is a polynomial. Observe
that q is a root of each of the following polynomials: p2 (x) = p1 (x)xn−k , p02 (x),
p3 (x) = xp02 (x), p03 (x), p4 (x) = xp03 (x), etc. till ps+2 (x). Indeed, the (x − q)
term remains after performing the product rule. So, in particular, for i ≤ s
pi+2 (q) = 0. (2)
We have
p2 (x) = xn − c1 xn−1 − · · · − ck xn−k ,

p02 (x) = nxn−1 − c1 (n − 1)xn−2 − · · · − ck (n − k)xn−k−1 ,
p3 (x) = nxn − c1 (n − 1)xn−1 − · · · − ck (n − k)xn−k ,
p03 (x) = n2 xn−1 − c1 (n − 1)2 xn−2 − · · · − ck (n − k)2 xn−k−1 ,
p4 (x) = n2 xn − c1 (n − 1)2 xn−1 − · · · − ck (n − k)2 xn−k ,
..
.
pi+2 (x) = ni xn − c1 (n − 1)i xn−1 − · · · − ck (n − k)i xn−k .
We need to verify that for bn = ni q n , (bn )n∈N gives a solution of the recurrence
an − c1 an−1 − · · · − ck an−k = 0. Let us plug bn for an in the left hand side of
71
the equation:
bn − c1 bn−1 − · · · − ck bn−k =
i n i n−1 i n−i
n q − c1 (n − 1) q − · · · − ck (n − k) q =
pi+2 (q) = 0.
Here the last equality holds by (2).
Theorem 3.23 (Existence and Uniqueness). Let f be a linear homogeneous
recursion with constant coefficients and p(x) be its characteristic polynomial
with roots q1 , . . . , qr having multiplicities s1 , . . . , sr , respectively. Then for any
initial values, there is a solution (an )n∈N of f written in the form
an = (?q1n + ?nq1n + · · · + ?ns1 −1 q1n ) + · · · + (?qrn + ?nqrn + · · · + ?nsr −1 qrn ), (3)
where ?’s are some constants. In particular, if all roots of the characteristic
polynomial are distinct, an = ?q1n + ?q2n + · · · + ?qrn .
Proof. From Lemma 3.22, all the listed terms are solutions of the recursion
with some initial values. From Lemma 3.20, any of their linear combinations is
a solution of the recursion with some initial values. It remains to verify that
there are coefficients (?’s) of the linear combination so that it is a solution of the
recursion with the given initial values a0 , a1 , . . . , ak−1 . Denote the consecutive
coefficient in (3) in front of nj qin by γi,j , i = 1, . . . , r, j = 0, . . . , si − 1. That is,
the γi,j ’s are the unknowns solving equation (3) for all n ∈ N, which can hence
be written as
 
Xr sXi −1 Xr sXi −1
an = γi,j nj qin =  γi,j nj  qin .

i=1 j=0 i=1 j=0
Recall that a0 , a1 , . . . , ak−1 are the known initial values. Plugging n = 0, n =

1, . . . , n = k − 1 we get:
a0 = (γ1,0 + 0 + · · · + 0) + · · · + (γr,0 + 0 + · · · + 0)
a1 = (γ1,0 + γ1,1 + · · · + γ1,s1 −1 )q1 +
· · · + (γr,0 + · · · + γr,sr −1 )qr
a2 = (γ1,0 + 2γ1,1 + · · · + 2s1 −1 γ1,s1 −1 )q12 +
· · · + (γr,0 + · · · + 2sr −1 γr,sr −1 )qr2
..
. (4)

ak−1 = γ1,0 + (k − 1)γ1,1 + · · · + (k − 1)s1 −1 γ1,s1 −1 q1k−1 +

· · · + γr,0 + · · · + (k − 1)sr −1 γr,sr −1 qrk−1
The matrix of this system of linear equations with unknowns γi,j , i = 1, . . . , r,
j = 1, . . . , si − 1 is:
 
1 0 0 0 ··· 1 0 0 0 ···
 q1 q q q1 ··· q2 q2 q2 q2 ··· 
 2 1 1 
 q1 2q12 22 q12 23 q12 ··· q22 2q22 22 q22 23 q22 ··· 
 3 
 q1 3q13 32 q13 33 q13 ··· q23 3q23 32 q23 33 q23 ···  (5)
 4 
 q1 4q14 42 q14 43 q14 ··· q24 4q24 42 q24 43 q24 ··· 
 
.. .. .. .. .. .. .. .. ..
. . . . ··· . . . . .
72
This is a square k × k matrix. We shall show that it is non-singular by proving
that its rows r1 , r2 , . . . , rk are linearly independent. Assume that there are
coefficients α1 , . . . , αk , not all equal to zero, so that α1 r1 +α2 r2 +· · ·+αk r1 = 0.
The components of this vector correspond to respective columns in the matrix
in the following form:
α1 + α2 q1 + α3 q12 + · · · + αk q1k−1 = 0
α1 0 + α2 q1 + α3 2q12 + · · · + αk (k − 1)q1k−1 = 0
α1 0 + α2 q1 + α3 22 q12 + · · · + αk (k − 1)2 q1k−1 = 0
..
.
α1 + α2 qr + α3 qr2 + · · · + αk qrk−1 = 0
α1 0 + α2 qr + α3 2qr2 + · · · + αk (k − 1)qrk−1 = 0
α1 0 + α2 qr + α3 22 qr2 + · · · + αk (k − 1)2 qrk−1 = 0
..
.
Let h(x) = α1 +α2 x+α3 x2 . . .+αk xk−1 . Then we see that q1 is a root of h(x), as
well as a root of h2 (x) = xh01 (x), h3 (x) = xh02 (x), . . ., hs1 . By Lemma 3.21 h(x)
is divisible by (x − q1 )s1 . Similarly,
Qr h(x) is divisible by (x − q2 )s2 , . . . , (x − qr )sr .
ThusPh(x) is divisible by i=1 (x − qi )si . Thus the degree of h(x) is at least
r
k = i=1 si , a contradiction to the definition of h(x). Thus there are no such
coefficients αi ’s and thus our matrix is non-singular. Therefore the system (4)
has a solution for any given right-hand side (a0 , a1 , . . . , ak−1 ).
Theorem 3.23 gives us a tool to solve any linear homogeneous recursion with
constant coefficients. The main steps are the following.
1. Derive the characteristic polynomial p(x).
2. Compute the roots q1 , . . . , qr of p(x) with respective multiplicities s1 , . . . , sr .
3. Solve the system of linear equations (4) (for example by Gaussian Elimi-
nation).
Pr Psi −1
4. Write down the explicit solution an = i=1 j=0 γi,j nj qin .
Let us illustrate this approach with a few examples.
Example 3.24. We solve the recursion
an = −5an−1 + 6an−2 , a0 = 3, a1 = 10.
The characteristic polynomial is x2 + 5x − 6, its roots are q1 = 1 and q2 =

−6, both with multiplicity 1, i.e., s1 = s2 = 1. Thus the general solution is
γ1,0 1n + γ2,0 (−6)n . Plugging initial values, we obtain
a0 = γ1,0 + γ2,0 = 3,
a1 = γ1,0 − 6γ2,0 = 10.
Thus γ2,0 = −1, γ1,0 = 4, and the recursion is solved by (an )n∈N with
an = 4 − (−6)n .
73
an = −4an−1 − 4an−2 , a0 = 3, a1 = 10.
The characteristic polynomial is x2 + 4x + 4, its roots are q1 = −2, with mul-

tiplicity 2, i.e., s1 = 2. Thus the general solution is γ1,0 (−2)n + γ1,1 n(−2)n .
Plugging initial values, we obtain
a0 = γ1,0 = 3,
a1 = γ1,0 (−2) + γ1,1 (−2) = 10.
Thus γ1,0 = 3, γ1,1 = −8, and the recursion is solved by (an )n∈N with
an = 3(−2)n − 8n(−2)n .
an = −4an−2 , a0 = 6, a1 = 20.
The characteristic polynomial is x2 + 4, its roots are q1 = 2i and q2 = −2i, both

with multiplicity 1, i.e., s1 = s2 = 1. Thus the general solution is γ1,0 (2i)n +
γ2,0 (−2i)n . Plugging initial values, we obtain
a0 = γ1,0 + γ2,0 = 6,
a1 = γ1,0 (2i) + γ2,0 (−2i) = 20.
Thus γ1,0 = 3 − 5i, γ2,0 = 3 + 5i. Therefore an = (3 − 5i)(2i)n + (3 + 5i)(−2i)n =

(2i)n (3 − 5i + (−1)n 3 + (−1)n 5i). So, if n is odd, an = (2i)n (−2 · 5i), if n is
even an = (2i)n 6. Note that in = en ln i = eπni/2 = cos(πn/2) + i sin(πn/2), so
for n = 4k, in = 1, n = 4k + 2, in = −1, for n = 4k + 1, in = i, for n = 4k + 3,
in = −i.
Hence the recursion is solved by (an )n∈N with


 5 · 2n+1 , if n = 4k + 1

6 · 2n , if n = 4k + 2
an =


 −5 · 2n+1 , if n = 4k + 3

−6 · 2n , if n = 4k + 4.
Recall from Example 3.17 that the Fibonacci sequence (Fn )n∈N solves the
recursion
Fn = Fn−1 + Fn−2 , F0 = 1, F1 = 1.
We shall now derive an explicit formula for Fn .
Theorem 3.27 (Binet’s Formula). For the Fibonacci numbers we have:
√ !n √ !n !
1 1+ 5 1− 5
Fn = √ −
5 2 2
√
1+ 5
where Φ := 2 is called the Golden Ratio.
74
Figure 22: In a drawing like this one, obtained by squares “spiraling” out of the
center, the ratio of width and height will approach the golden ratio. Of course,
you already know all about spirals, since you watched all Vi Hart Videos, right?
Proof. The recurrence relation is Fn+2 = Fn+1 + Fn with starting values F0 =

characteristic polynomial is p(x) = x2 − x − 1. The roots of
0, F1 = 1. So the √
1± 5
p(x) are q1,2 = 2 , i.e.
√ ! √ !
1+ 5 1− 5
p(x) = x − x− .
2 2
The general solution for Fn is therefore (by Theorem 3.23)

√ !n √ !n
1+ 5 1− 5
Fn = γ1,0 + γ2,0 .
2 2
We need to find γ1,0 , γ2,0 from initial values3 .

√ !0 √ !0
1+ 5 1− 5
0 = F0 = γ1,0 + γ2,0 = γ1,0 + γ2,0 .
2 2
√ !1 √ !1
1+ 5 1− 5
1 = F1 = γ1,0 + γ2,0
2 2
The unique solution is γ1,0 = √1 , γ2,0 = − √15 .

5
√
In the explicit formula we have just shown, the term ( 1−2 5 )n ≈ (−0.62)n
quickly goes to zero. This allows us to ignore it when computing Fibonacci
numbers:
j n k
Corollary 3.28. Fn is the integer closest to √15 Φn , i.e. Fn = Φ
√ + 1 .
5 2
√ n
√1
1− 5
Proof. It suffices to show that 5
2 < 12 .
3 Any two starting values will do, we could determine γ
1,0 , γ2,0 from F4 = 3, F5 = 5 if we
wanted to. But using F0 and F1 yields the simplest system of equations.
75
√ n
1 1 − 5 |1 − Φ|n |1 − Φ| 1 2 2 1
√ = √ ≤ √ = √ <√ √ = < .
5 2 5 5 Φ 5 5 5 5 2
Example 3.29 (Ternary n-string without 20). In Example 3.18 we found start-
ing values and the recurrence but we have yet to determine the values. Recall:
t1 = 3, t2 = 8, tn+2 = 3tn+1 − tn .
√
3± 5
So p(x) = (x2 −3x+1). The characteristic polynomial has the roots q1,2 = 2 .
So by Theorem 3.23 the general solution is
√ !n √ !n
3+ 5 3− 5
t(n) = γ1,0 + γ2,0 .
2 2
Again we use the initial values to determine γ1,0 and γ2,0 . For simplicity,
use t0 = 1 instead of t2 = 8:
1 = t(0) = γ1,0 + γ2,0

√ √
3+ 5 3− 5
3 = t(1) = γ1,0 · + γ2,0
2 2
√ √
5+3 5 5−3 5
Solving yields γ1,0 = 10 , γ2,0 = 10 and therefore:
√ ! √ !n √ ! √ !n
5+3 5 3+ 5 5−3 5 3− 5
tn = − .
10 2 10 2
3.3.2 Advancement Operator

Remark. This section contains an alternative treatment of recurrence relations
in terms of the so-called advancement operator. We present this material, even
though it is repetitive and everything is already covered in Section 3.3.1 with
the approach of the characteristic polynomial.
We can consider any sequence (fn )n∈N as a function f : N → R where
f (n) = fn .
If f fulfills a linear recurrence relation with constant coefficients, then there
is a unique way to extend f to Z such that the recurrence relation is still fulfilled.
Take for instance the Fibonacci sequence: Given the starting values there is a
unique way to calculate backwards using F (n − 2) = F (n) − F (n − 1):
n ... −5 −4 −3 −2 −1 0 1 2 3 4 5 ...
F(n) ... 5 −3 2 −1 1 0 1 1 2 3 5 ...
This makes f an element of the vector space of all functions from Z to R.
We define the advancement operator A acting on this vector space. It maps a
function f to the function Af with:
(Af )(n) = f (n + 1).
76
An advancement operator polynomial is a polynomial in A, for example 3A2 −
A + 6. This is also an operator which maps f to the function (3A2 − A + 6)f
with:
((3A2 − A + 6)f )(n) = 3f (n + 2) − f (n + 1) + 6f (n).
This notation allows us to write linear recurrence equations with constant coef-
ficients as:
p(A)f = g
where p is some polynomial. For the Fibonacci numbers we would write (A2 −
A − 1)F = 0.
In the following we try to solve equations of this kind. We start with the eas-
iest case of a homogeneous equation and an advancement operator polynomial
of degree one.
Lemma 3.30. For a real number r 6= 0 the solution to (A − r)f = 0 with initial
value f (0) = c is given by f (n) = crn .
Proof.
(A − r)f (n) = 0 ⇔ f (n + 1) − rf (n) = 0
⇔ f (n + 1) = rf (n)
⇔ f (n) = rn · f (0) = rn · c
Raising the difficulty a bit, now consider the advancement operator polyno-
mial p(A) = A2 + A − 6 = (A + 3) · (A − 2) and the corresponding homogeneous
recurrence p(A)f = (A + 3)(A − 2)f = 0. Note that solutions of (A − 2)f = 0
or (A + 3)f are also solutions of (A + 3)(A − 2)f = 0. 4
Actually, all solutions are sums of such solutions, which means that, using
the last Lemma, we know that f is of the form
f (n) = c1 (−3)n + c2 2n .
Here, c1 and c2 are constants that can be derived from two initial values for f .
It is no coincidence that we found a two dimensional space of functions:
Lemma 3.31. If p(A) has degree k and its constant term is non-zero, then the
set of all solutions f to p(A)f = 0 is a k-dimensional subspace of functions from
Z → C, parametrized by c1 , . . . , ck which can be determined by k initial values
for f .
We do not give a formal proof. However, you should not be surprised by
the statement. It is clear that set of all solutions form a subspace: If f, g are
solutions and α, β ∈ C\{0} then αf +βg is also a solution since p(A)(αf +βg) =
αp(A)f + βp(A)g = α · 0 + β · 0 = 0. It is also intuitive that there are k degrees
of freedom: For every choice for values of f on {1, 2, . . . , k} the recurrence gives
a unique way to extend f (upwards and downwards) to other values.
Generalizing our observation for the advancement operator polynomial (A +
3)(A − 2), we state (also without proof):
4 There is some low-level notational magic involved here: We need ((A + 3) · (A − 2))f =
(A + 3)((A − 2)f ) = (A − 2)((A + 3)f ) where · is multiplication of polynomials and “nothing”

is the application of an operator to a function.
77
Proposition 3.32. If p(A) = (A − r1 ) · (A − r2 ) · . . . · (A − rk ) for distinct
r1 , . . . , rk , then all solutions of p(A)f = 0 are of the form
f (n) = c1 r1n + c2 r2n + . . . + ck rkn .
Every polynomial of degree k has k complex roots, but those roots are not
necessarily distinct. The following Theorem handles the case of roots with
multiplicity at least two and is therefore the last piece in the puzzle.
From now on, instead of “all solutions are of the form. . . ” we say the “general
solution is . . . ”.
Theorem 3.33. If r 6= 0 and p(A) = (A − r)k , then the general solution of the
homogeneous system p(A)f = 0 is
f (n) = c1 rn + n · c2 rn + n2 c3 rn + . . . + nk−1 ck · rn .
If p(A) = q1 (A) · q2 (A) and q1 (A), q2 (A) have no root in common, then the
general solution of p(A)f = 0 is the sum of the general solutions of q1 (A)f = 0
and q2 (A)f = 0.
Example. Let p(A) = (A − 1)5 · (A + 1)2 · (A − 3). Then the general solution
to p(A)f = 0 is
f (n) = c1 + n · c2 + n2 · c3 + n3 · c4 + n4 · c5 + c6 (−1)n + n · c7 (−1)n + c8 3n .
Note that the condition that r (the constant term of p(A)) may not be zero
is no real restriction: An advancement operator polynomial of the form p(A) · A
describes the same recurrence as p(A), just at a different index.
Theorem 3.34 (Binet’s Formula). For the Fibonacci numbers we have:
√ !n √ !n !
1 1+ 5 1− 5
Fn = √ −
5 2 2
√
1+ 5
where Φ := 2 is called the Golden Ratio.
Proof. The recurrence relation is Fn+2 = Fn+1 + Fn with starting values F0 =

0, F1 = 1. Rewrite this in terms of advancement operator polynomial:
f (n + 2) − f (n + 1) − f (n) = 0 ⇐⇒ (A2 − A − 1) f (n) = 0.

| {z }
p(A)
√
1± 5
The roots of p(A) are 2 , i.e.
√ ! √ !
1+ 5 1− 5
p(A) = A− A− .
2 2
The general solution for f (n) is therefore (by Proposition 3.32)

√ !n √ !n
1+ 5 1− 5
f (n) = c1 + c2 .
2 2
78
Figure 23: In a drawing like this one, obtained by squares “spiraling” out of the
center, the ratio of width and height will approach the golden ratio. Of course,
you already know all about spirals, since you watched all Vi Hart Videos, right?
We need to find c1 , c2 from initial values5 .

√ !0 √ !0
1+ 5 1− 5
0 = f (0) = c1 + c2 = c1 + c2 .
2 2
√ !1 √ !1
1+ 5 1− 5
1 = f (1) = c1 + c2
2 2
The unique solution is c1 = √1 , c2 = − √15 .

5
√
In the explicit formula we have just shown, the term ( 1−2 5 )n ≈ (−0.62)n
quickly goes to zero. This allows us to ignore it when computing Fibonacci
numbers:
j n k
Corollary 3.35. Fn is the integer closest to √15 Φn , i.e. Fn = Φ
√ + 1 .
5 2
√ n

Proof. It suffices to show that √15 1−2 5 < 12 .
√ n
1 1 − 5 |1 − Φ|n |1 − Φ| 1 2 2 1
√ = √ ≤ √ = √ <√ √ = < .
5 2 5 5 Φ 5 5 5 5 2
Example 3.36 (Ternary n-string without 20). In Example 3.18 we found start-
ing values and the recurrence but we have yet to determine the values. Recall:
t1 = 3, t2 = 8, tn+2 = 3tn+1 − tn .
So p(A)t√= (A2 − 3A + 1)t = 0. The advancement operator polynomial has the

roots 3±2 5 . So by Proposition 3.32 the general solution is
√ !n √ !n
3+ 5 3− 5
t(n) = c1 + c2 .
2 2
5 Any two starting values will do, we could determine c , c from F = 3, F = 5 if we
1 2 4 5
wanted to. But using F0 and F1 yields the simplest system of equations.
79
Again we use the initial values to determine c1 and c2 . For simplicity, use
t0 = 1 instead of t2 = 8:
1 = t(0) = c1 + c2
√ √
3+ 5 3− 5
3 = t(1) = c1 · + c2
2 2
√ √
5+3 5 5−3 5
Solving yields c1 = 10 , c2 = 10 and therefore:
√ ! √ !n √ ! √ !n
5+3 5 3+ 5 5−3 5 3− 5
tn = − .
10 2 10 2
3.3.3 Non-homogeneous Recurrences

So far we ignored non-homogeneous linear recurrence relations, i.e. those of the
form
an = c1 (n)an−1 + · · · + ck (n)an−k + g(n)
for some non-trivial function g : N → R.
Lemma 3.37. Any two solutions (an )n∈N , (a0n )n∈N of a non-homogeneous linear
recurrence an = f (an−1 , . . . , an−k ) + g(n) “differ by” a solution of the corre-
sponding homogeneous recurrence an = f (an−1 , . . . , an−k ), i.e. there is (a∗n )n∈N
such that an = a∗n + a0n for all n ∈ N with a∗n = f (a∗n−1 , . . . , a∗n−k ) for all n ≥ k.
Proof. Just set a∗n = an − a0n for all n ∈ N. Then for all n ≥ k we have
k
X k
X
f (a∗n−1 , . . . , a∗n−k ) = ci (n)a∗n−i = ci (n) an−i − a0n−i
i=1 i=1
k
X k
X
= ci (n)an−i − ci (n)a0n−i = (an − g(n)) − (a0n − g(n))
i=1 i=1
= an − a0n = a∗n ,
as desired.
This means that to find all solutions of an = f (an−1 , . . . , an−k ) + g(n) it suffices
to find a single such solution (a0n )n∈N , the particular solution, and all solutions
(a∗n )n∈N of an = f (an−1 , . . . , an−k ), the general solution.
Then the general solution of the non-homogeneous recurrence is given by
an = a∗n + a0n for n ∈ N.
Since there is no framework that always works to find particular solutions,
this is actually the difficult part.
Example 3.38. Recall from Example 1.6 (and the corresponding problem on
the exercise sheet) that when cutting a two dimensional cake into the maximum
number of pieces, the n-th cut yields n additional pieces, i.e.
sn = sn−1 + n, s0 = 1 and therefore sn = c1 (n)sn−1 + g(n)
for c1 (n) = 1 and g(n) = n for n ∈ N.
80
1. Find general solution for the homogeneous recurrence:
sn = sn−1 ⇒ p(x) = x − 1 ⇒ q1 = 1, s1 = 1
⇒ a∗n = γ1,0 q1n = γ1,0 .
2. Find particular solution for non-homogeneous recurrence: sn = sn−1 + n.

We guess that the solution is “similar” to the right hand side. Since the
right hand side (i.e. g(n) = n) is a polynomial of degree 1 over n, we guess
that the solution could be a polynomial of degree 2 over n meaning
a0n = d1 n2 + d2 n + d3
where d1 , d2 , d3 are to be determined. We plug this guess into the desired

recurrence:
!
d1 n2 + d2 n + d3 = a0n = a0n−1 + n = d1 (n − 1)2 + d2 (n − 1) + d3 + n
= d1 n2 + d2 n + d3 + (−2d1 n + d1 − d2 + n)
which has the solution d1 = d2 = 21 (note that we “lost” d3 on the way,

so it is not needed and we just set it to zero). So we found one particular
solution:
0 1 2 1 n(n + 1) n+1
an = n + n = = .
2 2 2 2
3. The general solution for the non-homogeneous recurrence is the sum

n+1
sn = a∗n + a0n = γ1,0 +
2
We still need to determine γ1,0 using the initial value:
1 = s0 = γ1,0 + 0 ⇒ γ1,0 = 1

n+1 n n n
⇒ sn = +1= + + .
2 2 1 0
Example 3.39. We solve the recurrence:
an = 4an−1 − 4an−2 + 3n + 2n
1. General solution for homogeneous recurrence:
an = 4an−1 − 4an−2 ⇒ p(x) = x2 − 4x + 4 = (x − 2)2 ⇒ q1 = 2, s1 = 2

⇒ a∗n = γ1,0 q1n + γ1,1 nq1n = γ1,0 2n + γ1,1 n2n .
2. Particular solution of the non-homogeneous recurrence:
an = 4an−1 − 4an−2 + 3n + 2n
Guess something similar to the right hand side. Our attempt is6 :
a0n = d1 · 3n + d2 n + d3 .
6 We could have guessed another summand of d4 · n2 , but as it turns out, it is not needed.
81
Which means we need to find d1 , d2 , d3 such that:
!
d1 · 3n + d2 n + d3 = a0n = 4an−1 − 4an−2 + 3n + 2n
= 4d1 · 3n−1 + 4d2 (n − 1) + 4d3
− 4d1 · 3n−2 − 4d2 (n − 2) − 4d3 + 3n + 2n
= 8d1 · 3n−2 + 4d2 + 3n + 2n
= d1 · 3n + d2 n + d3
+ (−d1 · 3n−2 − d2 n − d3 + 4d2 + 3n + 2n)
= d1 · 3n + d2 n + d3
+ ((9 − d1 )3n−2 + (2 − d2 )n − d3 + 4d2 )
and we get d1 = 9, d2 = 2, d3 = 8, i.e.
a0n = 3n+2 + 2n + 8.
3. The general solution of non-homogeneous recurrence is therefore

an = a∗n + a0n = γ1,0 2n + γ1,1 n2n + 3n+2 + 2n + 8.
Starting values would now allow us to determine γ1,0 and γ1,1 but we omit
this here.
3.3.4 Solving Recurrences using Generating Functions

We now examine one example of a recurrence relation with non-constant coef-
ficients.
Example 3.40 (Derangements yet again). Recall from Theorem 1.24 that for
dn = |Dn |, the number of derangements on [n], we have
d0 = 1, d1 = 0, d2 = 1, dn = (n − 1)(dn−1 + dn−2 ) (for n ≥ 2).
This constitutes a homogeneous linear recurrence relation with k = 2 and coef-
ficients c1 (n) = c2 (n) = n − 1, g ≡ 0.
Let D(x) be the exponential generating function for (dn )n∈N , i.e.
∞
X xn
D(x) = dn
n=0
n!
Then using Observation 3.13 and the recurrence we find:

∞
X X∞ ∞
xn xn X xn
D0 (x) = dn+1 = ndn + ndn−1
n=0
n! n=0
n! n=1 n!
X∞ X∞
xn−1 xn−1
=x dn +x dn−1
n=1
(n − 1)! n=1
(n − 1)!
= xD0 (x) + xD(x).
This differential equation has (given d1 = 0, d2 = 1) the unique solution:
e−x
D(x) =
1 − x.
82
Luckily, we already know the exponential generation functions for e−x and for
1
1−x so we can write:
∞
! ∞
!
1 X n X xn
−x nx
D(x) = e · = (−1) · n! .
1−x n=0
n! n=0
n!
∞ n
!
3.12
X X n xn
n−k
= k!(−1) .
n=0
k n!
k=0
P
n
n

So we found yet another proof that dn = k k!(−1)n−k .
k=0
Let us give two more examples showing how we can solve recurrences using
generating functions.
Example 3.41. Consider the recurrence
an = −an−1 + 6n−2 (for n ≥ 2), a0 = 1, a1 = 3.
With the characteristic polynomial p(x) = x2 + x −P 6 in mind we find the

∞
following equation for the generating function F (x) = n=0 an xn of (an )n∈N .
∞
X ∞
X ∞
X
F (x)(1 + x − 6x2 ) = an xn + an−1 xn − 6 an−2 xn
n=0 n=1 n=2
= a0 x0 + a1 x1 + a0 x1 = 1 + 4x.
Thanks to our approach and using the recurrence for n ≥ 2, all but a finite
number of terms canceled out. In the last step we used the initial values.
Using partial fraction decomposition (not handled here) we can simplify the
identity and obtain:
1 + 4x 1 + 4x 6 1 1 1
F (x) = 2
= = 5 · − 5 ·
1 + x − 6x (1 − 2x)(1 + 3x) 1 − 2x 1 + 3x
∞ ∞
6X 1X
= (2x)n − (−3x)n
5 n=0 5 n=0
1
P∞
where in the last step we applied 1−y = n=0 y n for y = 2x and y = −3x. We
can now see the coefficients:
an = 6
5 · 2n − 1
5 · (−3)n .
Example 3.42. Consider the following non-homogeneous recurrence:
bn = bn−1 + 2bn−2 + 2n (for n ≥ 2), b0 = 2, b1 = 1.
Multiplying the equivalent statement bn = bn−1 − 2bn−2 = 2n with xn for each

n ≥ 2 and adding yields a sum in which we can rewrite the terms:
∞
X ∞
X ∞
X ∞
X
bn xn − bn−1 xn −2 bn−2 xn = 2n xn .
n=2 n=2 n=2 n=2
| {z } | {z } | {z } | {z }
F (x)−b0 −xb1 x·(F (x)−b0 ) x2 F (x) 1
1−2x −1−2x
83
Solving for F (x) yields:
1
F (x)(1 − x − 2x2 ) = + 1 − 3x
1 − 2x
2 − 5x + 6x2
⇒ F (x) =
(1 − 2x)(1 − x − 2x2 )
From here, apply partial fraction decomposition with a method of your choice,
to get
1 1 1
F (x) = − 91 +2 + 13
9 1 + x.
1 − 2x 3 (1 − 2x)2
From this we can obtain the coefficients again using identities we know.
∞ ∞
!2 ∞
X X X
n n n n
F (x) = − 19 2 x + 2
3 2 x + 13
9 (−1)n xn .
n=0 n=0 n=0
X∞ ∞
X X∞
= − 19 2 n xn + 2
3 (n + 1)2n xn + 13
9 (−1)n xn .
n=0 n=0 n=0
The coefficients are therefore:
bn = − 91 2n + 32 (n + 1)2n + 13 n
9 (−1) .
84
4 Partitions
4.1 Partitioning [n] – the set on n elements
A partition of [n] into given by its parts A1 , . . . , Ak with
k
[
Ai 6= ∅ (for i = 1, . . . k), Ai ∩ Aj = ∅ (for i 6= j), Ai = [n].
i=1
7
1
5 3
6 4
9
2 8
Figure 24: A partition of [9] into 4 sets.
The parts are unlabeled, i.e. if two partitions use the same parts Ai only
in a different order, we consider them to be identical. If we fix the number of
parts, i.e. if we want to count partitions of [n] into exactly k non-empty parts,
then their number is given by sIIk (n) as we have already seen in Section 1.5.2
where we considered arrangements of n labeled balls in k unlabeled boxes with
at least one ball per box.
We define the Bell Number as
n
X
Bn = sII
k (n).
k=0
It counts the total number of partitions of [n] (into an arbitrary number of sets).
Don’t get confused over the specialS case of n = 0: There is exactly one partition
of ∅ into non-empty parts: ∅ = A∈∅ A. Every A ∈ ∅ is non-empty, since no
such A exists. So we also have B0 = sII 0 (0) = 1.
A different way to define the Bell numbers is to consider a square free number
k ∈ N, meaning
k = p1 · p2 · . . . · pn for distinct primes p1 , . . . , pn .
The number of ways to write k as product of integers bigger than one is exactly
Bn . Take for instance k = 2 · 3 · 5 · 7, then some ways of writing k would be:
k = 6 · 35, k = 2 · 5 · 21, k = 210
which directly corresponds to the partitions:
{2, 3, 5, 7} = {2, 3}∪{5, 7}, {2, 3, 5, 7} = {2}∪{5}∪{3, 7}, {2, 3, 5, 7} = {2, 3, 5, 7}.
In the following we try to find an explicit formula for Bn . We start by finding

a recursion:
85
Theorem 4.1.
X
n−1
n−1

Bn = Bk (n ≥ 1).
k
k=0
Proof. Every partition of [n] has one part that contains the number n. In ad-
dition to n this part contains k other numbers (for some 0 ≤ k ≤ n − 1). The
remaining n − 1 − k elements are partitioned arbitrarily. From this correspon-
dence we obtain the desired identity:
X
n−1
n−1
X n−1
n−1 X n − 1
n−1
Bn = Bn−1−k = Bn−1−k = Bk .
k n−1−k k
k=0 k=0 k=0
Theorem 4.2 (Dobinski’s Formula).

∞
1 X kn
Bn = .
e k!
k=0
Proof. Consider the exponential generating function B(x) of (Bn )n∈N .

∞
X xn
B(x) = Bn
n=0
n!
Using the recursion from Theorem 4.1 we see

∞ ∞ n
!
X xn X X n xn
0
B (x) = Bn+1 = Bk
n=0
n! n=0 k=0
k n!
∞
! ∞
!
3.12
X xn X xn
= Bn · 1 = B(x) · ex .
n=0
n! n=0
n!
x
−1
⇒ B(x) = ee (using B0 = B(0) = 1).
∞ ∞ ∞
x 1 X (ex )k 1 X 1 X (xk)n
= 1e ee = =
e k! e k! n=0 n!
k=0 k=0
∞ ∞
!
X 1 X k n xn
=
n=0
e k! n!
k=0
4.1.1 Non-Crossing Partitions

Imagine the elements of [n] to be laid out in a circular way.
Two disjoint sets A, B ⊆ [n] are crossing if there are numbers i < j < k <
l ∈ [n] such that {i, k} ⊆ A, {j, l} ⊆ B.
A non-crossing partition is a partition in which the parts are pairwise non-
crossing. In cyclic drawings the notion is very intuitive. See Figure 25 for
examples.
We denote the number of non-crossing partitions of [n] by NCn . We prove
now that NCn is equal to Cn , the n-th Catalan number. We already came
across these numbers in Example 3.5, where we counted well formed parenthesis
expressions.
86
4 3 4 3
2 2
5 5
1 1
6 6
9 9
7 8 7 8
Figure 25: A non-crossing partition of [9] on the left and a crossing partition of
[9] on the right.
Theorem 4.3.
1 2n
NCn = Cn = .
n+1 n
Proof. Recall the values C0 = 1, C1 = 1, C2 = 2 (corresponding to the paren-
thesis expressions “”, “()”, “()()”, “(())”) and the recursion
n
X
Cn+1 = Ck Cn−k .
k=0
It suffices to prove that the sequence (NCn )n∈N has the same starting values
and satisfies the same recursion.
It is easy to check that NC0 = 1, NC1 = 1, NC2 = 2. Actually those numbers
are just the Bell numbers since partitions of [n] can only be crossing for n ≥ 4.
We now have to prove the recursion
n
X
NCn+1 = NCk · NCn−k .
k=0
To this end, consider any non-crossing partition P of [n + 1]. The last element
n + 1 is in some part S ⊆ [n + 1]. Let k be the biggest number in S other than
n + 1 if such an element exists and k = 0 if S = {n + 1}. Now observe that in
the partition P, every part contains either only numbers that are bigger than k
or only numbers that are at most k: Otherwise, such a part would cross S.
This means that P decomposes into a non-crossing partition of [k] and a
non-crossing partition of {k + 1, . . . , n}. Here we ignored n + 1: It must be in
the same part as k and will never produce a crossing if there has not already
been one. Such a decomposition is unique: Every non-crossing partition of
[n + 1] uniquely decomposes and every pair of non-crossing partitions of [k] and
{k + 1, . . . , n} corresponds to a non-crossing partition of [n + 1].
This proves the claimed recursion and therefore the Theorem.
4.2 Partitioning n – the natural number

We are interested in ways to write n as the sum of positive natural numbers.
We would say, for instance, that
n = 17 = 5 + 5 + 4 + 3.
87
is a partition of n = 17 into the (unlabeled) parts 5, 5, 4 and 3.
Alternatively we write n = λ = (5, 5, 4, 3) and say λ is the partition of n
(even though λ is a sorted tuple) hoping this will not be confusing.
We already considered this in Section 1.5.4 were we counted arrangements
of n unlabeled balls in k unlabeled boxes and at least one ball per box. The
number of such arrangements is given by pk (n). We established a recursive
formula 

 0 if k > n,

0 if n ≥ 1, k = 0,
pk (n)

 1 if n = k = 0,


pk (n − k) + pk−1 (n − 1) if 1 ≤ k ≤ n.
We define the total number of partitions of n (into an arbitrary number of parts)
as the partition function
Xn
p(n) = pk (n).
k=0
We illustrate a partition by its Ferrer diagram (also called Young diagram).

It consists of rows of squares (left aligned) corresponding to the parts (largest
part in the topmost row). Consider for instance the partition λ = (5, 5, 4, 3) of
n = 17 which has the Ferrer diagram
Observation 4.4. The number of partitions of n into at most k parts is equal

to the number of partitions of n with parts of size at most k.
Proof. The Ferrer diagrams for partitions with at most k parts are those with
at most k rows. The Ferrer diagrams for partitions with parts of size at most
k are those with at most k columns. Clearly there is a bijection between those
sets of diagrams: Flip them along the diagonal!
For instance, flipping yields .
Which means the diagram for the partition λ = (7, 5, 3, 3, 3, 2) is mapped to the
diagram for the partition λ = (6, 6, 5, 2, 2, 1, 1).
Formal Proof. The bijection on the diagrams corresponds to a bijection on the
partitions that maps the partition n = (λ1 , . . . , λk ) with biggest part l to the
conjugate partition n = (λ∗1 , . . . , λ∗l ) defined as
λ∗i = |{j | λj ≥ i}| .
Theorem 4.5 (Euler). ∞ ∞

X Y 1
p(n)xn = .
n=0
1 − xk
k=1
88
Proof. We boldly rewrite the infinite
P∞ product as an infinite product of infinite
1
sums using the identity: 1−y = n=0 y n .
∞
Y 1 1 1 1
= · · · ...
1 − xk 1 − x 1 − x2 1 − x3
k=1
∞
! ∞
! ∞
!
X X X
n1 2n2 3n3
= x x x ...
n1 =0 n2 =0 n3 =0
In an expansion of this, the coefficient P in front of xn is just the number of ways

we can choose n1 , n2 , n3 , . . . such that i≥0 i · ni = n. But such choices exactly
correspond to partitions of n, i.e. the coefficient is p(n).
Take for instance our favorite partition n = 17 = 5+5+4+3. It corresponds
to choosing n5 = 2, n4 = 1, n3 = 1 and all other indices ni as 0. This proves
the claim.
Define podd (n) as the number of partitions of n into odd parts and pdist (n)
as the number of partitions of n into distinct parts.
Take for instance n = 7. We have podd (7) = 5 since
7 = 1+1+1+1+1+1+1, 7 = 1+1+1+1+3, 7 = 1+3+3, 7 = 1+1+5, 7 = 7,
and pdist (7) = 5 since
7 = 3 + 4, 7 = 1 + 2 + 4, 7 = 2 + 5, 7 = 1 + 6, 7 = 7.
This is no coincidence, in fact, in the problem class you already saw:

∞
X ∞
Y ∞
Y X∞
1
podd (n)xn = = (1 + xk
) = pdist (n)xn .
n=0
1 − x2k+1 n=0
k=0 k=0
This proves podd = pdist , but we will prove it yet again, this time by constructing
a bijection.
Theorem 4.6. podd = pdist .
Proof. Let n = λ1 + λ2 + . . . + λk be a partition of n into distinct parts. We
separate the powers of 2 from the λi , i.e. we write:
n = u1 2a1 + u2 2a2 + . . . + uk 2ak
for odd numbers ui and λi = ui 2ai . Note that the ui need no longer be distinct,
for example if λ1 = 5, λ2 = 10 we would have u1 = u2 = 5. We sort the
summands according to the ui which gives
n = µ1 · (2α1 + 2α2 + . . . + 2αl1 ) + µ2 · (2β1 + 2β2 + . . . + 2βl2 ) + . . . (?)

| {z } | {z }
r1 r2
where the values µi are all distinct and take the roles of ui , in particular, the
following multisets coincide
{u1 , u2 , . . . , uk } = {µ1 , µ1 , . . . , µ1 , µ2 , µ2 , . . . , µ2 , . . .}.

| {z } | {z }
l1 times l2 times
89
Note that the values ri are sums of distinct powers of 2.
For the bijection, we map the original partition into distinct parts to the
following partition into odd parts
n = µ1 + µ1 + . . . + µ1 + µ2 + µ2 + . . . + µ2 + . . . .
| {z } | {z }
r1 times r2 times
We still need to argue that this constitutes a bijection. To do so, we explain

the inverse mapping.
Given a partition into odd numbers with repetition numbers ri , there is a
unique way to write these ri as sums of distinct powers of 2, the binary expansion
of ri . This gets us to the situation (?) and from there we get back to a partition
into distinct parts by multiplying out.
Example. To illustrate the last Theorem, take the partition into this distinct
parts:
26 = 12 + 6 + 4 + 3 + 1.
separate the powers of two and sort the terms according to the odd numbers
that remain:
26 = 3 · 22 + 3 · 21 + 1 · 22 + 3 · 20 + 1 · 20
= 3(20 + 21 + 22 ) + 1(20 + 22 )
=3·7+1·5
which yields the partition
26 = 3 + 3 + 3 + 3 + 3 + 3 + 3 + 1 + 1 + 1 + 1 + 1.
All steps are reversible, as there is only one way to write 5 and 7 as sums of
distinct powers of two. In terms of Ferrer diagrams we have mapped:
7→
Now define peven

d (n) and podd
d (n) as the numbers of partitions of n into an
even number of distinct parts and an odd number of distinct parts, respectively.
We have, for instance
peven
d (7) = #{“1+6”, “2+5”, “3+4”} = 3,
podd
d (7) = #{“1+2+4”, “7”} = 2.
Apparently, these numbers can differ, but we now show that they can differ
by at most 1, and characterize when this is the case.
To this end, define for k ∈ Z the pentagonal number wk := (3k−1)k
2 . Note
(3k+1)k
that with this definition w−k = 2 . Some values are:
90
k ... -3 -2 -1 0 1 2 3 ...
wk ... 15 7 2 0 1 5 12 ...
Lemma 4.7. (
(−1)k if n = wk for some k ∈ Z,
peven
d (n) − podd
d (n) =
0 otherwise.
Proof. Consider the Ferrer Diagrams for a partition into distinct parts. In it,
two rows may never have the same length. Define the slope S of such a diagram
to be a maximal staircase going diagonally down starting from the top-right
square and define the bottom B as the last row of the diagram. The following
diagram has a slope of length 3 (highlighted in red) and a bottom of size 2
(highlighted in blue):
Note that, in a few special diagrams, B and S may have a single square in
common. We define the set ∆ := B ∩ S, it contains the single common square
if it exists, and is empty otherwise. We will be sloppy in notation using B, S
and ∆ to simultaneously denote the set and the size of the set.
Now distinguish three types partitions corresponding to three types of dia-
grams where:
• Type 1: S ≥ B + ∆
• Type 2: S < B − ∆
• Type 3: B − ∆ ≤ S < B + ∆
Note that the latter case can only occur for ∆ = 1 and S ∈ {B, B − 1}.
Claim. (i) Type 3 partitions can only occur if n = wk for some k ∈ Z.
(ii) Conversely, if n = wk for some k ∈ Z, then there is exactly one Type 3
partition of n.
Proof of (i). Consider the case k := S = B − 1 first, where the diagram looks
like this (for k = 4):
which gives
k
X (k + 1)k (3k + 1)k
n = k2 + i = k2 + = = w−k .
i=1
2 2
The other case is k := S = B meaning the diagram looks like this (for k = 4)
91
which gives
(3k + 1)k 2k (3k − 1)k

n = w−k − k = − = = wk .
2 2 2
Proof of (ii). The existence is already clear from our proof of (i): We found
for arbitrary k ∈ N a type 3 partition of w−k and wk . For the uniqueness, note
that no two diagrams of type 3 have the same size: We can iterate through all
of them by alternatingly adding a column and a row.
Next we prove the following claim:
Claim (iii). For every k ∈ N there is a bijection between Type 1 partitions on
k rows and Type 2 partitions on k − 1 rows.
Note that from this the Theorem follows, since it guarantees a bijection that
maps every partition to a partition with one row more or one row less, so every
partition into an even number of parts is mapped to a partition into an odd
number of parts and vice versa. This shows that the number of even and odd
partitions must coincide. The only disturbance can be the single type 3 partition
that exists if n = wk and is even if k is even and odd if k is odd.
Proof of (iii). Consider a Type 1 partition, its slope is at least as large as its
bottom, maybe like this:
We take away the bottom and distribute the squares among the first |B| rows
of the diagram, like this:
There is enough room to do this (since the slope was at least as big as the
bottom) and the resulting partition is a partition into distinct parts. The size
of the bottom has increased and the size of the new slope is the size of the old
bottom. Therefore the new diagram is of type 2 or type 3 and, looking more
closely, type 3 can actually not occur as result of our operation, so it really is of
type 2. The inverse operation is to take the current slope and create a new row
from it. After checking that this maps type 2 partitions to type 1 partitions we
are done.
Theorem 4.8 (Euler’s Pentagonal Number Theorem).
∞
Y ∞
X ∞
X
(1 − xk ) = (−1)k xwk = 1 + (−1)k (xwk + xw−k ).
k=1 k=−∞ k=1
92
Proof. Multiply out the left hand side. The coefficient for xn counts the number
of partitions of n into distinct parts, however, partitions into an odd number of
parts are counted with negative sign. Therefore
∞
Y ∞
X ∞
X
Thm 4.7
(1 − xk ) = (peven
d (n) − podd
d (n))x
n
= (−1)k xwk .
k=1 n=0 k=−∞
We have yet to explain why wk are called pentagonal numbers. It is basically

analogous to square numbers and triangle numbers: All occur naturally when
filling a two-dimensional region with dots.
We start with one dot, and say it is “layer 0”, then add layer by layer filling
an area of the respective kind. For pentagonal numbers, there are four dots in
layer 1, seven dots in layer 2 and so on. Generally, in layer i there are 3i + 1
dots since there are three sides with i + 1 dots each, but two dots are shared by
two sides. In a drawing with k layers there are:
k−1
X k−1
X
k (3k − 1)k
3i + 1 = k + 3( i) = k + 3 = = wk .
i=0 i=0
2 2
4.3 Young Tableau

A Young tableau T is a Ferrer diagram of a partition λ of n that is filled with
all numbers 1, . . . , n. For instance if λ = (7, 6, 5, 3, 2, 2, 1) is a partition of 26:
11 26 23 13 19 14 24
8 18 9 15 22 21
5 1 2 3 6
17 4 7
10 12
16 25
20
We say T is a tableau of λ or T has shape λ. There are n! tableaux of shape λ.

Definition 4.9. A standard Young tableau is a Young tableau where the num-
bers in the rows are increasing from left-to-right and the numbers in the columns
are increasing from top-to-bottom, take for instance the following standard
93
Young tableau:
1 2 3 6 12 16 25
4 7 8 17 18 21
5 11 13 19 26
9 15 22
10 23
14 24
20
From now on we only consider standard Young tableaux.

Theorem 4.10 (Robinson-Schensted-Correspondence). There is a bijection be-
tween the permutations of [n] and the set
[
{(T1 , T2 ) | T1 and T2 are standard Young tableaux of shape λ}.
λ is partition of n
i.e. the set of ordered pairs (T1 , T2 ) where T1 and T2 are standard Young tableaux
of the same shape.
Proof. We will see a geometric construction that builds, given a permutation π,
a corresponding pair of standard Young tableaux.
Let π be a permutation of [n] and X(π) = {(i, π(i)) | i = 1, . . . , n} the corre-
sponding point set in the plane. For instance, for the permutation π = 3271546
of [7] the point set X(π) is:
π(x)
7
6
5
4
3
2
1
x
1 2 3 4 5 6 7
A point p is minimal if there is no other point that is both to the left and below
of p. The set of all minimal points in X is therefore
min(X) = {(x, y) ∈ X | ∀(x0 , y 0 ) ∈ X \ {x, y} : x0 > x or y 0 > y}.
The shadowline S(X) for a point set X is the weakly decreasing rectilinear
line through all points in min(X) and convex bends exactly in min(X). Here,
by convex we mean and by concave we mean . In the following drawing,
94
min(X) (red) and S(X) (black) are shown.
π(x)
7
6
5
4
3
2
1
x
1 2 3 4 5 6 7
In case you like pseudo-code, a concise algorithmic description of our con-

struction is given below. It does not yet provide the pair of tableaux we want,
but all the auxiliary objects we require to define them. A detailed description
in words follows.
Algorithm 1 Geometric Variant of Robinson-Schensted-Correspondence

X1 ← X(π)
i←0
while Xi+1 6= ∅ do
i←i+1
Xi0 ← Xi
j←0
while Xi0 6= ∅ do
j ←j+1
Sij ← S(Xi0 )
Xi0 ← Xi0 \ min(Xi0 )
end while
ni ← j
Xi+1 ← convex bends of Si1 , . . . , Sini
end while
m←i
The algorithm proceeds in phases (counted by the variable i), the total
number of phases m is not known beforehand (but bounded by n). We start
every phase i with a non-empty point set Xi . For Xi we construct a sequence
of shadowlines Si1 , . . . , Sini where the j-th shadowline is taken for the point set
consisting of those points from Xi that were not used for previous shadowlines.
Sj−1
Si1 = S(Xi ), Si2 = S(Xi \ Si1 ). In general: Sij = S(Xi \ k=1 Sik ).
The i-th phase ends as soon as all points from Xi were contained in one of the
shadowlines Si1 , Si2 , . . . , Sini .
The shadowlines of phase i determine the i-th row of the tableaux T1 and
T2 we want to construct. Let xji be the x-coordinate of first segment of the
shadowline Sij , i.e. the x-coordinate at which Sij leaves the picture on the top
and yij the y-coordinate of last segment of Sij , i.e. the y-coordinate at which Sij
leaves the picture on the right. Then the i-th row of T1 and T2 consists of the
numbers x1i , . . . , xni i and yi1 , . . . , yini , respectively.
95
Shadowlines may contain concave bends (they do iff they have two or more
convex bends). The set Xi+1 is defined as the set of all concave bends occurring
in the shadowlines Si1 , . . . , Sini . If Xi+1 is empty, we are done, otherwise, it
serves as point set for phase i + 1. (proof continues later)
Before we verify that the construction yields the bijection we desire, we give
an example.
Example. Consider again π = 3271546. The first phase of the algorithm will
find three shadowlines S11 , S12 , S13 . They leave the diagram on x positions 1, 3
and 7 and on y positions 1, 4 and 6, giving rise to the partial tableaux as shown.
π(x)
7
(1)
6 T1 = 1 3 7
5
4
3 (1)
2 T2 = 1 4 6
1
x
1 2 3 4 5 6 7
There are four concave bends on these shadowlines which give the point set
for the next phase: X2 = {(2, 3), (4, 2), (5, 7), (6, 5)}. In phase 2 we get two
shadowlines and add corresponding second lines to the tableaux.
π(x)
7 (2) 1 3 7
6 T1 =
2 5
5
4
3 1 4 6
(2)
2 T2 =
1 2 5
x
1 2 3 4 5 6 7
There are still convex bends, so we proceed with phase 3
π(x)
1 3 7
7 T1 = 2 5
6
4 6
5
4
3 1 4 6
2 T2 = 2 5
1 3 7
x
1 2 3 4 5 6 7
No concave bends were made this time, our construction is done. The pair
(T1 , T2 ) is the result.
96
We now establish, in a series of claims, that the construction constitutes
a bijection between permutations and pairs of Young tableaux of the same
shape as desired. We use the notion of a chain, which is a subset Y of a
point set X such that Y is increasing, i.e. Y = {(x1 , y1 ), (x2 , y2 ), . . . , (xk , yk )}
with y1 < y2 < . . . < yk and x1 < x2 < . . . < xk .
Claim. The number ni of shadows lines in phase i is the length of the longest
chain in Xi .
Proof of Claim: Let Y be a longest chain in Xi . No shadowline can contain more

than one point of Y , since shadowlines are decreasing while Y is increasing. This
shows ni ≥ |Y |.
Now observe that for the element µ ∈ Y with smallest x- and y-coordinate
we have µ ∈ min(Xi ), otherwise there would be a point to the left and below
of µ enabling us to extend Y . Therefore, the first shadowline Si1 will contain µ
(and, indeed, an element of every longest chain) so the longest chain in Xi \ Si1
will be of size |Y | − 1. The next shadowline will contain another point of Y and,
by induction, we conclude that ni ≤ |Y | since X1+|Y | will only contain chains
of size 0, meaning X1+|Y | is empty. (claim)
Claim. The tableaux have valid shape, i.e. ni+1 ≤ ni for all i = 1, . . . , m − 1.
Proof of Claim: Let Y be a longest chain in Xi+1 , consisting of some positions

of concave bends of the shadowlines from the previous phase i. Since the concave
bends of every shadowline are decreasing, Y can only contain one concave bend
from each shadowline of phase i. This means |Y | ≤ ni . We already know
|Y | = ni+1 from the previous claim, so we are done. (claim)
Claim. The rows of T1 and T2 are increasing.

Proof of Claim: This is equivalent to saying: Shadowlines constructed later in
the phase will leave the picture further to the right and further to the top. This
is obvious by construction: Every shadowline goes through the unused point
with smallest x coordinate and the unused point with smallest y coordinate,
making that point unavailable for later shadowlines. (claim)
Claim. The columns of T1 and T2 are increasing.
Proof of Claim: We consider two consecutive phases i and i + 1 and show that
in the step from row i to row i + 1 every column is increasing.
In phase i the shadowlines Si1 , . . . Sini where constructed with corresponding
sets of concave bends Bi1 , . . . , Bini (some of which may be empty) that together
form the set Xi+1 . In the following we implicitly use that shadow lines of
the same phase are non-crossing. We have Bi1 ⊆ min(Xi+1 ), meaning that
the points Bi1 will be “consumed” by the first shadowline Si+1 1
of phase i + 1.
Therefore Si+1 will be constructed from a subset of Bi , . . . , Bini , and will use up
2 2
j
all remaining points from Bi2 (if any remain). Generally, Si+1 will be constructed
j ni j
from a subset of Bi , . . . , Bi . The x-coordinate of the leftmost point of Si+1
j
is therefore at least the x-coordinate of the leftmost concave bend of Si and
therefore to the right of the leftmost point of Sij . This proves that columns of
T1 are increasing, for T2 do the same argument for y coordinates. (claim)
97
So far we proved that our map is well-defined, i.e. it gives rise to a pair of
standard Young tableaux for any permutation π ∈ Sn . The final step is to show
that the map constitutes a bijection.
Claim. There is an inverse map, i.e. from any pair (T1 , T2 ) of standard Young
tableaux of the same shape we can recover a corresponding permutation π ∈ Sn .
Proof. We demonstrate the inverse procedure with an example, hoping the gen-
eral case will be apparent from this. Consider the following pair of Young
tableaux:
1 2 4 8 1 3 6 7
3 6 2 4
T1 := , T2 :=
5 5
7 8
These tableaux contain eight numbers in four rows, so we try to recover a
four-phased construction and annotate the x and y coordinates of an 8 × 8 grid
with the index of the phases at which we wish a corresponding shadowlines to
leave the picture.
π(x)1 1 2 1 3 2 4 1
8 4
7 1
6 1
5 3
4 2
3 1
2 2
1 1
x
1 2 3 4 5 6 7 8
We claim there is a unique way to draw the shadowlines of phases 4, 3, 2 and 1

(one after the other) that correspond to a forward construction (see Figure 26).
First connect the indices of the last phase with L-shaped shadowlines (the
lines of the last phase do not have convex bends). In our example, there is only
one such line (drawn in red). Now draw the shadowlines for phase 3, again
there is only one in our example. It must connect the corresponding numbers
at the border of the picture and have a concave bend at the convex bend of the
shadowline of phase 4 on the way, there is only one way to do it. The lines for
phase 2 (shown in blue) are added one after the other (right-most first). The
first line must visit the convex bend at (7, 5) if it didn’t, then this point would
not be reachable for the next blue shadowline (since they must not cross). In
the last step the shadowlines of phase 1 are added (yellow). Convince yourself
that there is, again, only one way to draw them: Every yellow shadowline must
visit exactly those convex bends of the blue lines that where not visited by a
previous yellow shadowline and that would otherwise become unreachable for
subsequent yellow shadowlines.
Now the permutation can be obtained by taking the convex bends of the
shadowlines of phase 1. It is: π = 28561437.
With techniques similar to what we have already seen one can show that
this backwards procedure works for arbitrary tableaux.
98
π(x)1 1 2 1 3 2 4 1 π(x)1 1 2 1 3 2 4 1
8 4 8 4
7 1 7 1
6 1 6 1
5 3 5 3
4 2 4 2
3 1 3 1
2 2 2 2
1 1 1 1
x x
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
π(x)1 1 2 1 3 2 4 1 π(x)1 1 2 1 3 2 4 1
8 4 8 4
7 1 7 1
6 1 6 1
5 3 5 3
4 2 4 2
3 1 3 1
2 2 2 2
1 1 1 1
x x
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Figure 26: Reverse construction of the Robinson-Schensted-Correspondence.
4.3.1 Counting Tableaux

Given the geometric variant of the Robinson-Schensted-Correspondence (the
original construction, although equivalent, was formulated very differently), it
is not hard to count all Young tableaux with n elements.
We say a permutation π ∈ Sn is an involution if π 2 = id, i.e. applying
π twice yields the identity permutation. This is equivalent to saying that the
disjoint cycle decomposition of π contains only cycles of size 2 (transpositions)
and cycles of size 1 (fixed points). For instance:
i: 1 2 3 4 5 6 7 8
π(i): 1 2 3 4 5 6 7 8
π 2 (i): 1 2 3 4 5 6 7 8
We would write this involution as π = (1 4)(2 6)(7 5).

RSC
Lemma 4.11. Let π 7→ (T1 , T2 ), i.e. π is mapped to (T1 , T2 ) via the Robinson-
Schensted-Correspondence. Then
RSC
(i) π −1 7→ (T2 , T1 ).
(ii) π is an involution iff T1 = T2 .
99
(iii) The length of the longest increasing subsequence in π is the length of the
first row in T1 .
Proof. (i) Recall that the construction starts with
X(π) = {(i, π(i)) | i ∈ [n]} = {(π −1 (j), j) | j ∈ [n]}.
The point set X(π −1 ) is therefore obtained by flipping the point set X(π)
along the diagonal x = y, in other words, changing the roles of x and y.
The shadowlines we obtain will also be flipped, so every x-coordinate we
would have written into T1 is now a y-coordinate and therefore written
into T2 and vice versa. This just means that the roles of T1 and T2 are
swapped.
(i)
(ii) π 2 = id ⇔ π = π −1 ⇔ (T1 , T2 ) = (T2 , T1 ) ⇔ T1 = T2 .
(iii) Consider an increasing subsequence, i.e.
i1 < i2 < . . . ik with π(i1 ) < π(i2 ) < . . . π(ik ).
The corresponding set Y = {(ij , π(ij )) | j ∈ [k]} is a chain in X(π). Now

we can use a sub claim from the last theorem which asserted that the
length of the i-th row of the tableaux is the length of a longest chain in
Xi , using it for i = 1.
Note that by Lemma 4.11(ii) we have that the number in of involutions of

[n] and the number of standard Young tableaux with n squares coincide. The
following Lemma therefore gives a recurrence for the number of Young tableaux.
Lemma 4.12. For the number in of involutions of [n] we have
i1 = 1, i2 = 2, in = in−1 + (n − 1)in−2 .
Proof. Consider involutions in their disjoint cycle decomposition. The cycles

contain one or two elements, so think of an involution as an arrangement of n
labeled balls in unlabeled boxes and one or two balls per box. To count them,
we distinguish two cases:
Case 1: The ball with label n is in its own box. The rest is an arrangement
with n − 1 balls.
Case 2: The ball with label n is in a box together with another ball x. There
are n − 1 choices for x and the rest is an arrangement with n − 2 balls.
4.3.2 Counting Tableaux of the Same Shape

Let T (λ) be the set of all standard Young tableaux with shape λ. Take for
instance

λ= , T (λ) = 1 2 3 , 1 2 4 , 1 2 5 , 1 3 4 , 1 3 5 .
4 5 3 5 3 4 2 5 2 4
100
In general, let λ = (n1 , n2 , . . . , nm ) where n1 ≥ n2 ≥ . . . ≥ nm are numbers
that sum up to n. With this in mind we define the function t : Zm → N as
(
|T ((n1 , . . . , nm ))| if n1 ≥ n2 ≥ . . . ≥ nm ≥ 0,
t(n1 , . . . , nm ) =
0 otherwise.
For technical reasons, we allow trailing zero-length rows, for instance we could
write T ((3, 2, 0, 0)) = T ((3, 2, 0)) = T ((3, 2)) = 5. We claim the function is
fully characterized by the following identities:
(1) If the numbers ni fail to be weakly decreasing then
t(n1 , . . . , nm ) = 0.
(2) If the last number is zero, then it can be removed
t(n1 , . . . , nm , 0) = t(n1 , . . . , nm ).
(3) If the numbers are weakly decreasing and none is zero then
m
X
t(n1 , . . . , nm ) = t(n1 , n2 , . . . , ni−1 , ni − 1, ni+1 , . . . , nm )
i=1
= t(n1 − 1, . . . , nm ) + t(n1 , n2 − 1, . . . , nm ) + . . . .
(4) If only one number n ≥ 0 is left then
t(n) = 1.
It is obvious that t fulfills (1), (2) and (4). And (3) is just the observation that
the biggest number n must be the last number in one of the m rows. Consider
for instance λ = (5, 4, 4, 3) then in any Young tableau the number n = 16 will
be in one of the following positions:
n
n
n
n
For each case, we count the number of Young diagrams of the shape where the
corresponding square was removed, i.e. we consider the shapes
, , ,
So we compute t(4, 4, 4, 3)+t(5, 3, 4, 3)+t(5, 4, 3, 3)+t(5, 4, 4, 2). You may object

that the second shape is not valid, but that’s no problem since t(5, 3, 4, 3) was
defined to be zero.
Note that (1)-(4) gives a complete recursion, i.e. for each input the value is
either defined explicitly or given in terms of values for smaller inputs (in (2) the
number of terms is decreased, in (3) the sum of the inputs is decreased). This
101
means there is a unique solution. While we do not try the long and arduous
journey of discovering the solution ourselves, we can, given the solution, verify
that it is indeed correct.
To this end, we need to first examine the Vandermonde determinant which
is defined7 as Y
∆(x1 , . . . , xm ) := (xi − xj ).
1≤i<j≤m
It has the curious property that swapping two input values changes the sign
of the output. To see this, observe firstly what happens if the adjacent inputs
xi and xi+1 in the argument list of ∆ are swapped for some i ∈ [m − 1]. All
factors remain unchanged with the exception of (xi − xi+1 ) which is replaced
by (xi−1 − xi ) = −(xi − xi−1 ) as claimed. If two non-adjacent values xi , xi+k
are swapped, then this can be simulated by an odd number of swaps of adjacent
elements. Think of a race with n runners where Alice is in place i and Bob in
place i + k. We can make them change places as follows: First Bob overtakes
k people, putting him in front of Alice and then Alice (who is now in position
i+1) must be fall behind k −1 positions. This gives 2k −1 overtaking operations
in total, an odd number.
We remark (without proof), that the Vandermonde is indeed a determinant,
namely
 
1 x1 x21 ··· xm−1
1
 1 x2 x2 ··· xm−1 
(m
)  2 2 

∆(x1 , . . . , xm ) = (−1) 2 det  .. .. .. .. .

 . . . .
··

·
2 m−1
1 xm xm ... xm
Before we can get back to counting tableaux, we need to prove a technical
Lemma:
Lemma 4.13. We have the following identity on polynomials over x1 , x2 , . . . , xm , y:
m
X
m

xk ∆(x1 , . . . , xk + y, . . . , xm ) = x1 + x2 + . . . + xm + 2 y ∆(x1 , . . . , xm ).
k=1
Proof. Let g be the polynomial given as the sum on the left hand side. Observe
what happens if we swap the roles of xi and xj (with i < j) in g. The summands
for k ∈
/ {i, j} will just change sign (by our previous observation). The summand
with k = i turns into:
xj ∆(x1 , . . . , xj + y , . . . , xi , . . . , xm ) = −xj ∆(x1 , . . . , xi , . . . , xj + y , . . . , xm ).

↑ ↑ ↑ ↑
position i position j position i position j
So the term for k = i became the negated term with k = j, and this works vice
versa as well — altogether, the value of g changes sign if xi and xj swap roles.
Now consider the case of xi = xj , then swapping xi and xj obviously doesn’t
change anything. The only number that doesn’t change if its sign changes is
zero, so g = 0 whenever two xi and xj coincide (for i 6= j).
Now think of all the variables x2 , . . . , xm , y as some (distinct) integer con-
stants and of x1 as the only actual variable, then g is a polynomial of degree
7 Our definition differs from the usual one in that it may have a different sign.
102
n over x1 . It has several zeroes, one of which is x1 = x2 . This allows us to
divide g by the degree 1 polynomial x1 − x2 (using polynomial long division)
and we obtain g = p · (x1 − x2 ) for some polynomial p. In the same way we
separate the other zeroes getting g = p0 · (x1 − x2 )(x1 − x3 ) . . . (x1 − xm ) for
some polynomial p0 . Then, looking at p0 , we switch perspective thinking of x2
as the variable and of the other values as (suitable distinct) constants. Looking
at it that way we find the zeroes x2 = x3 , x2 = x4 , . . . , x2 = xm of p0 and can
separate corresponding factors from p0 . We repeat this for the remaining m − 2
variables as well. With this we eventually obtain:
Y
g = p̂ · (xi − xj ). (?)
1≤i<j≤m
for some polynomial p̂. Note that this was the important step. Since this is not
an algebra lecture, we allowed ourselves to be a bit sketchy: Formally we would
have to argue that no funny business is going on when doing the polynomial
long division and switching perspective between thinking of the xi as constants
or variables.
m
Given (?), everything falls into place: The degreem
of the polynomial g is
2 + 1 while the degree of the right hand side is 2 plus the degree of p̂. So
the degree of p̂ is one!
This means that g is of the form:
Y
g = (a1 x1 + a2 x2 + . . . + am xm + by) · (xi − xj )
1≤i<j≤m
= (a1 x1 + a2 x2 + . . . + am xm + by)∆(x1 , . . . , xm ).
for some constants a1 , . . . , am , b still to be determined8 . But recall that g was

originally defined to be:
m
X
g= xk ∆(x1 , . . . , xk + y, . . . , xm ).
k=1
Since the equation must hold for the special case of y = 0, we quickly see that
a1 = a2 = . . . = am = 1.
Now to determine b, we multiply out g and collect the sum of all monomials
containing y with multiplicity 1. We start by analyzing the k-th summand of g:
xk · ∆(x1 , . . . , xk + y, . . . , xm )
Y
= xk · (xi − xj ) (x1 − xk − y)(x2 − xk − y) . . . (xk + y − xm ).
1≤i<j≤m
i,j6=k
To get monomials with a single y, choose y from one of the factors and choose
the part without y everywhere else. This allows to reassemble ∆ with the
exception of a single missing factor (we write it into the denominator). The
8 You may object that p̂ may have a constant term. We call a polynomial homogeneous, if
all monomials have full degree. Since g and ∆(x1 , . . . , xm ) are homogeneous, we can conclude
that p̂ must be homogeneous as well.
103
m − 1 summands (one for each occurrence of y) add up to
 
Y
−y −y y
xk ·  (xi − xj ) + + ... +
x1 − xk x2 − xk xk − xm
1≤i<j≤m
k−1 m
!
X xk X xk
= y∆(x1 , . . . , xm ) − +
i=1 i
x − x k xk − xi
i=k+1
X −xk
= y∆(x1 , . . . , xm ) .
xi − xk
i6=k
If we sum up these these terms for all k we get

Xm X
−xk
y∆(x1 , . . . , xm )
xi − xk
k=1 i6=k
X −xk −xi

= y∆(x1 , . . . , xm ) +
xi − xk xk − xi
1≤i<k≤m
X
xk − xi m
= y∆(x1 , . . . , xm ) = y∆(x1 , . . . , xm ) .
xk − xi 2
1≤i<k≤m
m

This shows b = 2 , completing the proof.
Theorem 4.14.
∆(x1 , . . . , xm )n!
t(n1 , . . . , nm ) =
x1 !x2 ! . . . xm !
Pm
where n = i=1 ni , xi = ni + m − i for i = 1, . . . , m and x1 ≥ . . . ≥ xm ≥ 0.
Proof. We show that the right hand side is a solution to the recurrence we found
for t, i.e. if we define t by the right hand side then (1) − (4) from page 101 hold.
First note what happens if xi are not strictly decreasing, i.e. xi = xi+1 for some
i. This gives a value of 0 since the factor 0 = xi − xi+1 occurs in ∆(x1 , . . . , xm ).
We also know:
ni + m − i = ni+1 + m − (i + 1) ⇔ ni = ni+1 − 1,
so the ni are not weakly decreasing which means these values do not correspond
to a valid shape for Young tableaux. Some other invalid shapes were already
excluded by restricting ourselves to weakly decreasing xi . All in all, we are
consistent with:
(1) t(n1 , . . . , nm ) = 0 unless n1 ≥ . . . ≥ nm ≥ 0.
The second thing we need to show is
(2) t(n1 , . . . , nm−1 , 0) = t(n1 , . . . , nm−1 ).
Since nm = 0 means xm = 0, we calculate:
Y Y m−1
Y
∆(x1 , . . . , xm−1 , 0) = (xi − xj ) = (xi − xj ) xi
1≤i<j≤m 1≤i<j≤m−1 i=1
m−1
Y m−1
Y
= ∆(x1 , . . . , xm−1 ) xi = ∆(x1 − 1, . . . , xm−1 − 1) xi .
i=1 i=1
104
With this it is easy to verify:
∆(x1 , . . . , xm−1 , 0)n!

t(n1 , . . . , nm−1 , 0) =
x1 ! . . . xm−1 !0!
∆(x1 − 1, . . . , xm−1 − 1)n!
= = t(n1 , . . . , nm−1 ).
(x1 − 1)! . . . (xm − 1)!
The hard part is (3), but we did most of the work already in the last Lemma.
Pm
(3) t(n1 , . . . , nm ) = k=1 t(n1 , . . . , nk − 1, . . . , nm ).
First note:
m
X m
X m−1
X
m
xi = (ni + m − i) = n + j =n+ .
i=1 i=1 j=0
2
Now Lemma 4.13 gives for y = −1

m
X
m
xk ∆(x1 , . . . , xk − 1, . . . , xm ) = (x1 + . . . + xm − ) ∆(x1 , . . . , xm )
2
k=1 | {z }
n
Hence:
Pm
∆(x1 , . . . , xm )n! xk ∆(x1 , . . . , xk − 1, . . . , xm )(n − 1)!
k=1
t(n1 , . . . , nm ) = =
x1 ! . . . xm ! x1 ! . . . xm !
m
X Xm
∆(x1 , . . . , xk − 1, . . . , xm )(n − 1)!
= = t(n1 , . . . , nk − 1, . . . , nm )
x1 ! . . . (xk − 1)! . . . xm !
k=1 k=1
The last property is again easy to verify. For m = 1 and n = n1 , we have:

∆(x1 )n! ∆(n1 )n1 !
(4) t(n1 ) = = = ∆(n1 ) = 1.
x1 ! n1 !
where ∆(n1 ) = 1 because it is an empty product.
With the last Theorem it would already be possible to count Young tableaux
of any shape fairly conveniently. However, with a geometric interpretation, it
gets even easier. This geometric interpretation uses hooks.
Consider a partition λ = (n1 , . . . , nm ) and a square (i, j) ∈ λ, by which we
mean the square in row i and column j. The hook hi,j rooted in (i, j) is the
union of all squares to the right of the square in the same row and below of the
square in the same column. Here is a picture, a Ferrer diagram where the hook
h2,3 is highlighted, it has length 7.
105
Theorem 4.15 (Hook length formula). Let λ be a partition of n. Then
n!
t(λ) = Q
|hi,j |.
(i,j)∈λ
Proof. Let λ = (n1 , . . . , nm ). We multiply the lengths of all the hooks, going
through them row by row. For row i, the first hook hi,1 starts at square (m, 1)
then goes upwards and rightwards and ends at square (i, ni ). It has length
(ni + m − i) = xi . The other hooks in row i also end in (i, ni ) but they start
in other positions. We go through these positions from left to right as shown in
the following figure (for i = 3).
•
•
•
•
Now if there was a hook starting in each integer position this line passes through,
the product of the hook lengths would just be xi !. However, the positions
marked with a dot do not correspond to starts of hooks so for each dot we
have to divide by the length that a hook starting there would have. The dot
in line j (j > i) is in position (j, nj + 1) so a hook going to (i, ni ) has length
(j − i + ni − nj ) = (xi − xj ). So the product of all valid hooks for line i is
xi !
(xi −xj ) . Now for the product of all hook lengths of all rows we get:
Q
j>i
Y Y xi ! x1 ! . . . xm ! x1 ! . . . xm ! 4.14 n!
|hi,j | = Q = Q = =
(xi − xj ) (xi − xj ) ∆(x1 , . . . , xm ) t(λ).
(i,j)∈λ 1≤i≤m
j>i i<j
Take for instance the following Ferrer diagram, that we annotate with the
lengths of the hooks rooted at the respective position:
7 6 3 1
→ 5 4 1
3 2
2 1
The number of Young tableaux of this shape is by the Hook length formula 11!
divided by all the hook lengths, so:
11! 11 · 10 · 9 · 8
= = 11 · 10 · 3 · 4 = 1320.
7·6·3·1·5·4·1·3·2·2·1 3·2
106
5 Partially Ordered Sets
Example 5.1. As was accurately observed by Randall Munroe, creator of xkcd,
different fruit not only differ in their tastiness, but also in the difficulty of
preparing them. Some types fruit are clearly superior to others, for instance
Figure 27: https://xkcd.com/388/: Coconuts are so far down to the left they
couldn’t be fit on the chart. Ever spent half an hour trying to open a coconut
with a rock? – Randall Munroe
seedless grapes are both more tasty and easier to prepare than oranges. However,
if you compare pineapples with bananas then pineapples are more tasty but
harder to prepare, so there is no clear winner.
To model these situations where some things are bigger/better/above/dominating
others things but some pairs of things may also be incomparable/on the same
level/of equal rank, we introduce the concept of partial orders.
Definition 5.2. A partially ordered set or poset or partial order is a pair

P = (X, ≤) where X is a (for us usually finite) set and ≤ is a binary re-
lation that is:
(i) reflexive: x ≤ x, ∀x ∈ X.
(ii) transitive: (x ≤ y and y ≤ z) ⇒ x ≤ z, ∀x, y, z ∈ X.
(iii) antisymmetric: (x ≤ y and y ≤ x) ⇒ x = y, ∀x, y ∈ X.
You probably already know some (partial) orders, and recognize a few ex-
amples from Table 3. Note that a total order (where for any x, y ∈ X we have
x ≤ y or y ≤ x) is a special case of a partial order. We now introduce some
notation.
• If x ≤ y and x 6= y then we write x < y.
• If x ≤ y we also write y ≥ x.
107
X ≤ Example Relations
natural numbers order by value 23 ≤ 42, 42 23
natural numbers order by divisibility 6 7, 7 6, 13 ≤ 91
polynomials order by divisibility X − 7 ≤ X 2 − 10X + 21
words over {A, . . . , Z} order lexicographically ELEPHANT ≤ MOUSE
real intervals order by inclusion [3, 5] ≤ [2, 7], [2, 3] [4, 6]
real intervals “completely left of” [3, 5] [2, 7], [2, 3] ≤ [4, 6]
real valued functions point-wise domination 1 + x x2 , sin(x) ≤ 2
subsets of [N ] inclusion {1, 3} ≤ {1, 2, 3, 5}
Table 3: Some posets P = (X, ≤) with informal definitions of their order rela-
tions and examples.
• If neither x ≤ y nor y ≤ x then x, y are incomparable denoted by x k y.
• If x < y and there is no z with x < z < y, then x < y is a cover relation
(or cover for short), denoted by x l y.
We depict a poset by its Hasse diagram. In it, every x ∈ X corresponds to a
point in the plane and for every x, y ∈ X with xly the corresponding points are
connected by a y-monotone line where x is the lower (with smaller y-coordinate)
endpoint. In particular “transitive edges” are omitted meaning x, y ∈ P are not
directly connected if there is a z such that x < z < y. The poset from Example
5.1 has the Hasse diagram shown in Figure 28.
Figure 28: Hasse Diagram for the poset originating from Figure 5.1. For instance
lemons and bananas are connected by a line since bananas are both tastier and
easier to prepare. Lemons and cherries are not connected by a direct line since
they are already connected via bananas. Bananas and watermelons are not
connected at all since they are incomparable.
108
Sometimes it may be easier to define the cover relations than to define the
entire relation.
Example. Let X be the set of states of a Rubik’s Cube. For x ∈ X define r(x)
to be the minimum number of moves needed to solve the cube from state x.
We want x l y if and only if r(x) < r(y) and x and y can be transformed
into one another by one move. This is the cover relation of a poset (without
proof).
We now define some key concepts and corresponding notation.

Definition 5.3. Let P = (X, ≤) poset.
• A chain is a sequence of elements in increasing order, i.e. y1 < y2 < . . . <
yk for yi ∈ X. It has length k.
• The height of P is the length of a longest chain in P .

• A set Y ⊆ X where elements are pairwise incomparable (y1 k y2 for all
y1 6= y2 ∈ Y ) is an antichain.
• The width of P is the size of a largest antichain in P .
• The sets of minimal and maximal elements of P are given as
min(P ) := {x ∈ X | ∀y ∈ X : x ≤ y or x k y},
max(P ) := {x ∈ X | ∀y ∈ X : x ≥ y or x k y}.
Looking at the poset P from Figure 28 again, an example for a chain would
be {Watermelons, Pears, Strawberries}. There are several longest chains, one
of which is {Grapefruit, Oranges, Bananas, Plums, Pears, Blueberries, Seed-
less Grapes}. The height of P is therefore 7. An example for an antichain
is {Pineapple, Cherries, Plums, Red Apples}. No other antichain is longer so
the width of P is 4. The maximal elements are max(P ) = {Seedless Grapes,
Peaches} and the minimal elements are min(P ) = {Pineapples, Pomegranates,
Grapefruit, Lemons}. We now study partitions of posets into chains and an-
tichains. We start with the easier case.
Theorem 5.4 (Antichain Partitioning). The elements of every poset P =
(X, ≤) can be partitioned into h(P ) antichains (and not less).
Proof. Note first that we cannot partition P into fewer antichains: No antichain
can contain two elements of a chain, since elements of chains are pairwise com-
parable and elements of antichains are pairwise incomparable. Since P contains
a chain Y of size h(P ), at least h(P ) antichains are needed.
To see that h(P ) antichains suffice, first note that min(X) is an antichain:
Two minimal elements are always unrelated, otherwise the “bigger” minimal
element would not be minimal at all. Also, every maximal chain in P contains an
element from min(X): If Y = {y1 , . . . , yk } is a maximal chain with y1 < . . . < yk
and y1 is not minimal, then we would find y0 < y1 and therefore a longer chain.
So we take the first antichain to be min(P ), then we still need to partition
X \ min(P ) into h(P ) − 1 antichains. Since the maximal chains in X \ min(P )
have size at most h(P ) − 1 we can do this by induction.
109
Figure 29: Partition of the poset into 7 antichains obtained by iteratively putting
the minimal remaining elements into a new antichain.
To get back to our fruit example, Figure 29 shows how the poset can be
partitioned into 7 antichains.
Theorem 5.5 (Dilworth’s Theorem). Every poset P = (X, ≤) can be parti-
tioned into w(P ) chains (and not less).
Proof. Clearly, w(P ) chains are necessary: P contains an antichain of size w(P )
and no two of its elements can be contained in the same chain.
To show w(P ) chains suffice, we do induction on |X|. Consider a maximum
antichain A = {x1 , . . . , xw(P ) }.
The idea is to split P along A into two parts, take for instance our fruit
poset and A = {Seeded Grapes, Cherries, Plums, Red Apples}, then the two
parts are shown in Figure 30.
Formally we define P1 = (X1 , ≤) and P2 = (X2 , ≤) with elements
X1 := {y ∈ X | ∃x ∈ A : x ≤ y}, X2 := {y ∈ X | ∃x ∈ A : y ≤ x},
and the same relations as before. Note that with this definition X1 ∩ X2 = A,
since if y ∈ (X1 ∩ X2 ) then there is x1 ∈ A and x2 ∈ A with x1 ≤ y ≤ x2 . Since
A is an antichain this implies x1 = x2 = y so y ∈ A.
Now if we can partition P1 into chains C11 , . . . , C|A|
1
and P2 into chains
2 2 1 2
C1 , . . . , C|A| where Ci and Ci both contains xi then we can attach the chains
to one another obtaining a partition of P into chains C1 , . . . , C|A| . We can find
these partitions by induction unless P1 or P2 fail to be smaller than P .
Convince yourself that P1 = P ⇔ A = min(P ) and P2 = P ⇔ A = max(P ).
So our only problem is the case where there is no largest antichain except for
min(P ), max(P ) or both.
In that case, let C be any maximal chain. In the same way as in the previous
Theorem we argue that min(P ) ∩ C 6= ∅ and max(P ) ∩ C 6= ∅. Then P \ C has
110
Figure 30: We split the poset P along an antichain into P1 (the upper half) and
P2 (the lower half). The elements of the antichain are contained in both posets
after the split.
width w(P ) − 1 (since we removed an element from each largest antichain) so

it can be partitioned into w(P ) − 1 chains by induction and we are done.
5.1 Subposets, Extensions and Dimension

We now define when a poset is part of another poset, that is, when P = (X, ≤P )
is a subposet of Q = (Y, ≤Q ). This should be the case if X ⊆ Y and ∀x1 , x2 ∈
X : (x1 ≤P x2 ⇔ x1 ≤Q x2 ). In that case we say P is the subposet of Q induced
by X. But since we do not care about relabeling of elements, the notion of
subposet is actually a bit more generous:
P is a subposet of Q if there is an injective function f : X → Y such that
∀x1 , x2 ∈ X : (x1 ≤P x2 ⇔ f (x1 ) ≤Q f (x2 )). The function f is an embedding
and the image of f is a copy of P in Q.
We say Q = (X, ≤Q ) is an extension of P = (X, ≤P ) (note that they use the
same ground set X) if ∀x1 , x2 ∈ X : x1 ≤P x2 ⇒ x1 ≤Q x2 . When interpreting
relations as subsets of X × X we could write ≤P ⊆ ≤Q .
We say L = (X, ≤L ) is a linear extension of P if it is an extension of P and
a total order (meaning no two elements are incomparable in ≤L ).
Lemma 5.6. Let P1 = (X, ≤1 ), P2 = (X, ≤2 ) be two posets on the same ground
set X. Then we define P = P1 ∩ P2 := (X, ≤) with the relation ≤ := ≤1 ∩ ≤2 ,
meaning
x ≤ y ⇔ (x ≤1 y and x ≤2 y).
We claim P is a poset and P1 and P2 are extensions of P .
111
e
z
c d
Q := P := x y P 0 :=
w
a b
Figure 31: With the posets as shown (via their Hasse diagrams) P is a subposet
of Q as witnessed by the map w 7→ a, x 7→ c, y 7→ d, z 7→ e. Actually, there are
four different embeddings, since w could also be mapped to b and x and y can
also be mapped to c and d the other way round. These two maps correspond
to two copies of P in Q, namely {a, c, d, e} and {b, c, d, e}. The poset P 0 is not
a subposet of Q. However, Q is an extension of P 0 .
Proof. It is easy to verify that ≤ is an order relation. It is also clear that all pairs
that are related via ≤ are related via ≤1 and ≤2 , so ≤1 and ≤2 are extensions
of ≤.
Two-dimensional posets. Recall the poset P from the fruit example. It has
two natural linear extensions: The first is the total order LTaste in which the
fruit are arranged in order of increasing tastiness:
Lemons ≤LTaste Grapefruit ≤LTaste Tomatoes ≤LTaste . . . ≤LTaste Peaches.
In other words, LTaste is the order obtained when projecting all elements to the
tastiness-axis (note how this is an extension of P and a total order, assuming no
two fruit have exactly the same tastiness) The second is the total order LEase ,
in which the fruits are arranged in order of increasing ease of preparing them:
Pineapples ≤LEase Pomegranates ≤LEase . . . ≤LEase Seedless Grapes.
Fruit are ordered if and only if they are ordered according to both linear orders
so P = L1 ∩ L2 .
We want to capture this property in a new notion and define: A poset is 2-
dimensional if it has a two dimensional picture9 , i.e. an assignment f : P → R2
of positions in the plane to each element such that x < y in P if and only if
f (y) is above and right of f (x).
Not every poset is 2-dimensional. Consider the spider poset with three
legs, the Hasse Diagram is shown on the left of Figure 32. This spider has a
head H three knees K1 , K2 , K3 considered bigger than the head and three feet
F1 , F2 , F3 considered smaller than the corresponding knee. There are no further
relations.
We try to find a two dimensional embedding into the plane, i.e. assign
positions in the plane to each element (see right side of Figure 32).
The head H has to go somewhere which partitions the plane into four quad-
rants (above and right of H, below and right of H, above and left of H, below
and left of H). The knees must all go above and right of H since they are bigger
than H. Since knees are incomparable, the ones with bigger x coordinate must
9 Strictly speaking we would have to say: It has a two-dimensional picture but no one-
dimensional picture, for details refer to the general definition later.
112
K1
K2 K1 K3 K2
K3
F2 F1 H F3
H
Figure 32: On the left: The spider with three legs. On the right: Sketch for the
proof that the spider is not two-dimensional.
have smaller y coordinate so the knees must lie on a decreasing curve as shown.
Without loss of generality, K2 is the second point on this decreasing curve. Now
observe that there is no suitable point to put F2 : It cannot be in the quadrant
below and left of H nor in the quadrant above and right of H since H k F2 .
Since it must also be below and left of K2 , it must be in the shaded area. But
points in that area are below and left of either K1 or K3 (or both), which is not
allowed since F2 k K1 and F2 k K3 . This completes the proof.
Sums of Chains. We define an additional bit of notation. By i⊕k for i, k ∈ N

we mean the poset that is the union of a chain of length i and a chain of length
k and no comparabilities between the chains. For instance 1 ⊕ 1 is an antichain
of size 2. The poset important in the following is 1 ⊕ 2 = , let’s label with
c
a
b
. It consists of three elements a, b, c with b < c and a k b, a k c. Note that
has three linear extensions, a < b < c, b < c < a, b < a < c, the last
of which will have a special standing in the following: Given a linear extension
L of some poset P and a copy of in P then we say L orders this copy of
if the lonely element (corresponding to a above) does not come in between
the other two elements (corresponding to b and c above). Take for instance the
linear ordering L : a < d < b < c < e < f of following poset (shown on the left).
It contains 8 copies of listed on the right and almost all of them are ordered
by L.
f e f f
%
a d c c
d e b X b a X d e X
c
c e c
a b d d e e d
b X b X a X a X
f
The only exception is c d . It is not obvious whether or not a different linear
extension would have ordered every copy of (you will have to find that out
on the exercise sheet). What we show now is the significance of the existence of
such a linear extension.
Theorem 5.7. A poset P is 2-dimensional if and only if there is a linear

extension L of P ordering all copies of 1 ⊕ 2 in P .
Proof. “⇒” Assume P is 2-dimensional. Then P = L1 ∩ L2 for the two linear
extension L1 , L2 of P corresponding to an embedding into R2 . We claim
that any of the linear extensions, say L1 , orders every copy of .
113
Assume not, then we find elements a, b, c in P that form an unordered
c
copy of , i.e. we have the situation a b with b <L1 a <L1 c. Since
b ≤P c we also have b ≤L2 c so we are in this situation:
L2
c
b
L1
We know from L1 that, horizontally, a is between b and c, but vertically it

must be below b (since a k b) and it must be above c (since a k c). Clearly
this is not possible (it would have to be in both shaded areas at once)
which contradicts the assumption that L1 does not order a copy of .
“⇐” Let L be a linear extension of P ordering all copies of . From this
we construct an embedding into R2 . First label the elements of P in
accordance with L, i.e. x1 <L . . . <L xn . The x-coordinate of xi will
just be i. We define the y-coordinates from left to right, i.e. we define xi
assuming those of x1 , . . . , xi−1 are already given.
Considering the y-coordinates of {xj | j < i, xj ≤P xi } and place xi barely
above the maximum such y-coordinate.
Clearly, this procedure ensures that if we have xj ≤P xi then xi is put
above and right of xj . But we also have to ensure that incomparable
elements are not placed this way. Assume there is a smallest index i
where xj k xi for some j < i and the above procedure will still put xi
above and right of xj .
y
xi
xj
x
We must have put xi so high for a reason: There must be some xk (k < i)
with xk <P xi and xi was therefore assigned a y-coordinate barely above
xk . In the drawing above, xk is therefore on one of the horizontal lines,
either to the left or to the right of xj . It cannot be on the right (i.e. on
the blue line and j < k < i) since that would imply xj < xk (otherwise we
would have made a mistake already earlier, but i was minimal) and because
of xk < xi and transitivity we would also have xj < xi contradicting our
assumption of xj k xi .
So we have that xk is to the left of xj (i.e. on the red line and k < j).
x
This means xk k xj and therefore xj xik is a copy of in P . Since in
L we have xk < xj < xi it is not ordered by L, a contradiction.
Higher dimensional posets. There is a natural way to generalize the notion

from the previous few pages to arbitrary dimension. First note that for d ∈ N
the set Rd becomes an (infinite) poset via the dominance order which simply
means that for x = (x1 , . . . , xd ) ∈ Rd , y = (y1 , . . . , yd ) ∈ Rd we have
x ≤dom y :⇔ ∀i ∈ [d] : xi ≤ yi .
114
We say a poset P is d-dimensional if it is a subposet of (Rd , ≤dom ) and not a
subposet of (Rd−1 , ≤dom ).
Theorem 5.8. A poset P is d-dimensional if and only if d is smallest number

of linear extensions of P whose intersection is P .
Td
Proof. It suffices to show that P is a subposet of Rd if and only if P = i=1 Li
for some linear extensions L1 , . . . , Ld of P .
“⇒” We can assume, without loss of generality, that no two points share a
coordinate (we always break ties without changing the relations in the
poset). Then define Li as the order of the elements on the projection of P
to the i-th coordinate. This is a set of linear extensions with intersection
P.
“⇐” Given linear extensions L1 , . . . , Ld we can take these to assign coordinates.
The coordinate of x ∈ P will be (r1 , . . . , rd ) ∈ Rd where ri is the rank of
x in the i-th linear extension, i.e. ri = |{y ∈ P | y ≤Li x}|. It is easy to
see that this gives an embedding of P into Rd .
Next we observe that every poset has a well-defined dimension, in other

words, every poset can be written as the intersection of d linear orders for some
large enough d. We call such a set of linear extensions a realizer of P .
Ttaking all linear extensions L = {L | L is linear extension of P }
We claim that
works, i.e. P = L∈L L. It is clear that “⊆” holds. We leave the other direction
as an exercise.
Our next result shows that the dimension does not only exist but is actually
bounded by the width of the poset. We require some preparation though.
Definition 5.9. Let P = (X, ≤) be a poset.
• For S ⊆ X define the open downset D(S) = {y ∈ X | ∃x ∈ S : y < x} and
the closed downset D[S] = D(S) ∪ S.
• For single elements sets S = {s} we will simply write D(s) and D[s]
instead of D({s}) and D[{s}].
• In the same way define open and closed upsets:
U (S) = {y ∈ X | ∃x ∈ S : x < y}, U [S] = U (S) ∪ S,

U (s) = U ({s}), U [s] = U [{s}].
Consider the left of Figure 33 for an example.

Theorem 5.10. For any poset P its dimension is at most its width, i.e.
dim(P ) ≤ w(P ).
Proof. Our goal is to find a realizer on w = w(P ) dimensions.

By Dilworth’s Theorem we can partition P into w(P ) chains C1 , . . . , Cw .
Claim. For any chain C there is a linear extension L of P such that for any
x ∈ C and x k y we have y <L x.
115
X5
X4
∅ X3
X2
X1
X0
Figure 33: On the left: The open upset of {Cherries, Red Apples, Blueberries}
is shown in red and the closed downset of {Pineapples Bananas} is shown in
dashed blue.
On the right: Sketch for the claim from Theorem 5.10. Consider the chain
C = {Oranges, Watermelons, Cherries, Pears, Strawberries}. Then L is a linear
extension that puts the elements from C as late as possible, for instance like this
(elements from C highlighted): L : Grapefruit, Lemons, Tomatoes, Pineapples,
Pomegranates, Oranges, Bananas, Red Apples, Watermelons, Plums, Green
Apples, Seeded Grapes, Cherries, Pears, Peaches, Blueberries, Strawberries,
Seedless Grapes
Note first that once we have proved this claim, we have proved the theorem
since the linear extensions TL1 , . . . , Lw we get for the chains C1 , . . . , Cw are a
w
realizer of P , meaning P = i=0 Li . It is clear that the right side is an extension
of the left side. Now consider if x k y. Then x ∈ Ci for some i ∈ [w], and y ∈ Cj
for some j 6= i. Then we have y <Li x and x <Lj y so neither relation occurs
in Li ∩ Lj .
Think of Li as a linear extension of P that puts the elements from Ci as late
as possible. A sketch is given on the right of Figure 33.
Proof of claim. Denote the elements of C by x1 < x2 < . . . < xk and consider
their upsets. Note that clearly U (x1 ) ⊂ U (x2 ) ⊂ . . . ⊂ U (xk ).
Now define X0 := X \ U [x1 ], Xj := U (xj ) \ U [xj+1 ] (for 1 ≤ j < k) and
Xk := U (xk ) and let Pj := (Xj , ≤) be the poset induced by Xj (0 ≤ j ≤ k).
Given a linear extension Lj for each Pj (0 ≤ j ≤ k) we now define the linear
extension L of P as:
L : L0 x1 L1 x2 L2 . . . xk Lk .
First, convince yourself that L really is a linear extension of P , i.e. every element
occurs exactly once and if x <P y then x occurs before y.
116
Now assume x ∈ C and x k y. Say x = xl . Then y ∈/ U (xl ) so y is not part
of any Xj for j ≥ l. This means y <L xl = x as claimed.
Example 5.11 (Standard Examples). For a positive integer n define the poset
Sn = ({a1 , . . . , an , b1 , . . . , bn }, ≤) where ai ≤ bj ⇔ i 6= j and no (non-reflexive)
relations within {a1 , . . . , an } and {b1 , . . . , bn } respectively. For instance:
b1 b1 b2 b1 b2 b3 b1 b2 b3 b4
S1 = , S2 = S3 = S4 =
a1 a1 a2 a1 a2 a3 a1 a2 a3 a4
We call these posets the standard examples. We claim that dim(Sn ) = n =

w(Sn ) (for n ≥ 2), so the standard examples make the bound from the last
Theorem tight.
To see that dim(Sn ) ≥ n, consider a realizer of Sn (consisting of linear
extensions). Since ai k bi for i ∈ [n], some linear extension L in the realizer
must have bi <L ai . Since every remaining bj is greater than ai they must all
go right of ai and every remaining aj is smaller than bi , so they must all go left
of bi . This means that bj <L aj for no j 6= i. To reverse all pairs (ai , bi )i∈[n]
therefore requires n linear extensions.
5.2 Capturing Posets between two Lines
l2
T
S E
l1 P O
Figure 34: On the left we drew five shapes between two horizontal lines. From
this we obtain a poset (shown on the right) by considering a shape less than
another shape if it is left of it.
Imagine l1 and l2 are two horizontal lines in the plane. An object O ⊆ R2

is spanned between l1 , l2 if it is contained in the strip in between l1 and l2 , is
connected and touches both l1 and l2 .
With a set of such objects we associate a poset where the strict relation is
given as “left of”, meaning that O1 ≤ O2 if and only if O1 ∩ O2 = ∅ and the
leftmost point of O1 is left of the leftmost point of O2 . This clearly gives an
order relation (once we add the reflexive relations). Note that since the objects
touch the top and bottom line we avoid situations like this
l2
O1 O2
l1 O3
where one might be inclined to say that O1 is left of O3 is left of O2 but O1 is

not left of O2 . Since this is forbidden our “left of” is really consistent with all
reasonable intuitions one might have about leftness.
117
In the following we ask: What kind of posets can be represented in such a
way and what shapes are required?
Observation 5.12 (Straight Lines). If P is 2-dimensional then P can be rep-
resented by straight segments spanned between the two lines.
To see this, assume P = L1 ∩ L2 for two linear orders L1 , L2 . Then place
the elements of P on l1 in the order given by L1 , also place them onto l2 in the
order given by L2 and connect the points corresponding to the same element.
This give a line segment sx for each x ∈ X.
Now observe:
x <P y ⇔ (x <L1 y and x <L2 y) ⇔ sx left of sy .
The reverse holds as well, i.e. any poset represented by straight segments
spanned between l1 and l2 is at most two-dimensional and a realizer is given by
sorting the segments according to their endpoints on l1 and l2 .
The second type of object we consider are axis aligned rectangles. Note that
since those rectangles must be tangential to l1 and l2 , they are already uniquely
determined by their leftmost and rightmost x-coordinate. In fact, the setting is
not “really” two-dimensional and is easily seen to be equivalent to the setting
of interval orders where:
An interval order P = (X, ≤) is given by a set X of open bounded intervals
in R with (a, b) <P (c, d) iff b ≤R c.
We remark that not all interval orders are 2-dimensional but we will not
prove this until later. We first characterize them in terms of a forbidden sub-
poset:
Theorem 5.13 (Fishburn). P = (X, ≤) is an interval order if and only if
2 ⊕ 2 * P , i.e. there is no copy of the poset contained in P .
Proof. “⇒” Assume there is a copy of 2 ⊕ 2 labeled like this.
b d
a c
If P could be written as an interval order, then a, b, c, d would be rep-

resented by intervals (al , ar ), (bl , br ), (cl , cr ), (dl , dr ) with the following
properties:
• ar <R bl (since a <P b)
• bl <R cr (since c kP b)
• cr <R dl (since c <P d)
• dl <R ar (since a kP d)
This requires ar <R bl <R cr <R dl <R ar , which is clearly not possible.
“⇐” Let P be (2⊕2)-free. The crucial observation is that for any two elements
b, d ∈ X we have D(b) ⊆ D(d) or D(b) ⊇ D(d). Assume not: Then
D(b) * D(d) and D(b) + D(d). This implies b k d and the existence of
a ∈ D(b) \ D(d) and c ∈ D(d) \ D(b). Convince yourself that this means
that a, b, c, d forms a copy of (check a < b, c < d, a k c, a k d, b k c)
contradicting our assumption.
118
Given the observation, we can order the downsets D = {D(x) | x ∈ X}
by strict inclusion, i.e. ∅ = D0 ⊂ D1 ⊂ D2 ⊂ . . . ⊂ Dµ .
For x ∈ X we choose the interval (ax , bx ) such that
(
min{β | x ∈ Dβ } if x ∈ / max(P )
D(x) = Dax , bx =
µ+1 if x ∈ max(P ).
Observe that ax < bx . Also, if x <P y then x ∈ D(y) so bx ≤ ay and

conversely, if x k y then x ∈ / D(y) so bx ≥ ay and also y ∈ / D(x) so
by ≥ ax . This means (ax , bx ) ∩ (ay , by ) 6= ∅ and neither interval is left of
the other.
h
h g
f
e
f g d
c
d e b
a
a b c 0 1 2 3 4 5
D={ D0 = ∅ , D1 = ab, D2 = abc, D3 = abcde, D4 = abcdeg }

| {z } | {z } | {z } | {z } | {z }
D(a)=D(b)=D(c) D(d) D(e) D(f )=D(g) D(h)
Figure 35: Example for the construction in Fishburn’s Theorem. The poset on
the left contains no 2 ⊕ 2. We determine all downsets D0 ⊆ . . . ⊆ D4 . We pick
the interval orders as the Theorem suggests, for instance, the interval for d is
(1, 3) since D1 is the downset of d and D3 is the first downset containing d.
We now consider the case where the objects are (open) triangles spanned
between l1 and l2 with the base on l1 and tip on l2 (see pictures below).
Theorem 5.14. P is a triangle order if and only if there is a linear extension
of P that orders all copies of 2 ⊕ 2.
Here we say a 2 ⊕ 2 is ordered by a linear order L, if L puts the element of one
chain both before the elements of the other chain. In other words, if 2 ⊕ 2 is
labeled like this:
b d
a c
then we want either a <L b <L c <L d or c <L d <L a <L b.
Proof. “⇒” If P is a triangle order then we claim that taking L as the order
of the tips of the triangles from left to right works. Note that L is a
total order (obviously) and a linear extension of P : If a triangle is left of
another triangle (T1 <P T2 ) , then in particular its tip is left of the other
triangle (T1 <L T2 ).
Now consider a copy of 2 ⊕ 2 with the same labeling as above.
119
Then a, b, c, d correspond to triangles Ta , Tb , Tc , Td and Ta is left of Tb .
Ta Tb
Assume for contradiction that the tip of Tc is in between the tips of Ta

and Tb . Since c k b we need Tc to intersect Tb , maybe like this:
Ta Tc Tb
Now there is no way to add Td : It must be right of Tc but not right of Ta

which is clearly impossible.
“⇐” Assume L : x1 < x2 < . . . < xn is a linear extension of P that orders
all copies of 2 ⊕ 2. We argue that P is a triangle order. We construct a
triangle Tx for every x ∈ P , given by the horizontal position of the three
points, namely the tip tx and base (ax , bx ).
The tips should simply be ordered according to L, so txi = i for each
i ∈ [n]. We pick the bases of the triangles one after the other, going
through the elements of P from left to right according to L.
For x1 we define its base arbitrarily. Having defined triangles for x1 , . . . , xk ,
k ≥ 1, we now show how to add Tx for x = xk+1 with the desired rela-
tionships to the already existing triangles.
Consider the downset D(x). We choose ax = max{by | y ∈ D(x)}, so as
far left as possible while still right of every triangle that Tx should be right
of. The right end point bx is chosen somewhere far right.
y1 u y2 y3 s x
↑ ↑
ax far right
In the picture we assume D(x) = {y1 , y2 , y3 }, so ax = max{by1 , by2 , by3 } =

by3 . This ensures that x >P y implies that Tx is right of Ty (for any y).
Also (by choice of the tips) Tx is not left of an existing triangle. However,
the pictures shows that there might be u ∈ / D(x) such that Tx is right of
Tu . We need to fix this.
We do this by shifting ax to the left. But remember that ax needs to be
right of by for each y ∈ D(x), so as soon as ax would overtake such a by ,
we just take it with us, shifting it to the left as well. Now we need to argue
that this does not cause Ty to be suddenly left of some Tu for y u. So
assume ax and by are about to overtake the left endpoint au of Tu , maybe
like this:
t y u x
↑
shifting by and ax left,
almost overtaking au
120
All Orders
Convex Orders
Triangle Orders
dim(P ) ≤ 3 dim(P ) ≤ 2 Interval Orders
Figure 36: Venn diagram showing the relationships between some types of orders
we consider here and in the following.
Now remember: The reason we are shifting is that there is some t x

with Tt still left of Tx . Since we almost overtook au already, this triangle
Tt is also left of Tu so t < u (otherwise we would have made a mistake
earlier). Combining what we know (y < x, t < u, y u, t x) we see that
y, x, t, u forms a copy of 2 ⊕ 2 like this:
x u
y t
Since x is the most recent element, it comes last in L, and since L orders
every copy of 2 ⊕ 2 we have t <L u <L y <L x and the picture above was
actually misleading. In truth we are in this situation:
t u y x
Here, shifting ax and by past au is no problem as Tu and Ty will still

intersect. So we can shift as far as we need which proves the claim.
As an overview over the results of this section consider the Venn diagram
in Figure 36. The class of convex and triangle orders are already defined geo-
metrically as orders arising from convex sets or triangles spanned between lines,
respectively. But remember that interval orders and orders of dimension at most
two also arise from shapes in that way: From axis aligned boxes and segments.
In Theorem 5.15 we will furthermore see that the set of all orders arises from
y-monotone curves.
The containments are easy to see: Any straight line is a (degenerated) tri-
angle and every triangle is convex. To see the inclusion of interval orders in
triangle orders, observe that if we only allow “boring” triangles that have the
tip above their base, then triangles intersect iff their bases intersect, so these
triangles behave like intervals. Alternatively, this relation immediately follows
from Theorem 5.13 and Theorem 5.14.
In Lemma 5.16 we show that there is indeed a non-convex three dimensional
order. We remark without proof that the “small” circle of interval orders al-
ready contains posets of arbitrary dimension: The interval order defined by all
121
closed intervals with end points in [n] has dimension at least log log n + ( 21 +
o(1)) log log log n.
Theorem 5.15. Every poset P can be represented by y-monotone curves spanned
between l1 and l2 .
Proof. Take any realizer of P , i.e. P = L1 ∩ L2 ∩ . . . ∩ Lk for linear orders
L1 , . . . , Lk . Assume without loss of generality k ≥ 2 and introduce k − 2 lines in
between l1 and l2 . This gives k lines in total. On the i-th line we distribute the
element of P in increasing order according to Li . We then connect all points
belonging to the same element with a straight segments.
ordered by L5 : l2
ordered by L4 :
ordered by L3 :
ordered by L2 :
ordered by L1 : l1
x w y z
Now it is obvious that x <Li y for each i if and only if the line for x is left of
the line for y.
Lemma 5.16. There is a 3-dimensional order that is not a convex order.
Proof. Here is one such poset. It is a modified standard example:
d1 d2 d3
c1 c2 c3
b1 b2 b3
a1 a2 a3
Assume it can be represented by convex shapes Ca1 , Ca2 , . . . , Cd3 that are spanned
between l1 and l2 . Note that this implies that each such shape Cx contains a
straight line Sx that is spanned between l1 and l2 .
Since ai k di for i = 1, 2, 3 the shapes Cai and Cdi intersect. So take a
point xi ∈ Cai ∩ Cdi . Assume without loss of generality that x1 has the middle
y-coordinate. With respect to the straight line from x2 to x3 , we have that x1
is either left of it (red area) or right of it (blue) area:
l2
x2
x3
l1
Assume x1 is on the left. Then since c1 > a2 and c1 > a3 , the shape Cc1 must
be right of x2 and x3 (since x2 ∈ Ca2 , x3 ∈ Ca3 ), but since c1 < d1 it must also
be left of x1 (since x1 ∈ Cd1 ). The same holds for the straight line Sc1 . Clearly
such a straight line does not exist.
If x1 is on the right instead, we run into the same problem with the line Sb1 :
It has to be right of x1 but left of x2 and x3 .
122
5.3 Sets of Sets and Multisets – Lattices
We now consider posets whose elements are sets ordered by inclusion meaning
A ≤ B iff A ⊆ B (we will prefer the latter notation). Usually the sets we consider
are subsets of [n] and the most important example is the Boolean lattice Bn ,
the family of all subsets of [n]. Take for instance B3 :
{1, 2, 3}
{1, 2} {2, 3} {1, 3}
{2} {1} {3}
Two sets are incomparable if they are pairwise not included in one another. For
instance {1} k {2, 3} and {1, 2} k {1, 3}. Note also that {{1, 2}, {2, 3}, {1, 3}} is
a largest antichain in B3 .
The next Theorem generalizes this observation.

Theorem 5.17 (Sperner’s Theorem). w(Bn ) = d nn e .
2
Proof. Note that two different sets of the same size are never included in one
another. So taking all k-subsets of [n] (i.e. all subsets of size k) yields an
antichain of size nk . Choosing k = d n2 e maximizes its size and proves the lower
bound.
To prove the upper bound, we introduce a new notion: For a permutation π
of [n] we say that π meets a set A ⊆ [n] if the elements of A form a prefix of π,
meaning:
A = {π(1), π(2), . . . , π(|A|)}.
For instance, the permutation π = 24513 meets {2, 4} and {1, 2, 4, 5} but does
not meet {4, 5}.
Now consider a antichain F. We count the number of permutations that
meet an element of F. Note that the sets met by a single permutation π form
a chain, so no π can meet several elements of F.
So clearly the sets {π | π meets A} are disjoint for different A ∈ F and we
have: X
|{π | π meets A}| ≤ n!
A∈F
For any given A ⊆ [n] there are |A|!(n − |A|)! permutations that meet A: We
know that the first |A| elements of such a π are given by A and can be arranged
in |A|! ways, and the remaining n − |A| elements can be arranged in (n − |A|)!
123
ways. So we have:
X
|A|!(n − |A|)! ≤ n!
A∈F
X 1
⇔ n
≤1
A∈F |A|
X 1
⇒ n
≤1
A∈F dn
2e

n
⇔ |F| ≤ .
d n2 e
n
step we divided both sides by n!, then we used that k = d 2 e maxi-
In the first
n
mizes k and thus made the term under the sum independent of A.
Remark. Note that if F actually is a largest antichain in Bn , then all inequalities
n
used in the above proof must be tight. In particular, each |A| must actually
n
n

have been equal to b n c = d n e . So only sets of size b 2 c or d n2 e are used in
n
2 2
F. From this it is easy to see that:

• If n is even, then F = [n]n is the unique largest antichain.
2
[n]
[n]

• If n is odd, then all largest antichains are contained in bn
∪ dn
.
2c 2e
Next we consider so-called intersecting families. We say F ⊆ 2n is intersect-

ing if ∀A, B ∈ F : A ∩ B 6= ∅.
Observation 5.18. If F ⊆ 2[n] is intersecting, then |F| ≤ 2n−1 and this bound
can be attained.
Proof. For any element A ∈ F, the complement [n] \ A is disjoint from A and
cannot be in F. So taking complements is an injection that maps elements from
F to elements not in F. Therefore |F| ≤ 2n − F and thus |F| ≤ 2n−1 .
To attain this bound, consider Fx = {A ⊆ [n] | x ∈ A} for some fixed
x ∈ [n]. Clearly this is an intersecting family of size 2n−1 .
So this problem has a fairly straightforward and boring solution. But what
about the maximum cardinality of an intersecting k-family, i.e. we only allow
sets of size k? If k = 1, then we can clearly only pick one set (any two different
sets of size 1 do not intersect). If k = 2, then it is easy to see that n − 1 is best
possible using the sets {1, 2}, {1, 3}, {1, 4}, . . . , {1, n}. Is the best strategy still
the “obvious” one in general? It turns out the answer is “yes”, but the proof is
a bit more involved.
n
Theorem 5.19 (Erdős-Ko-Rado). For two integers n and 0 < k ≤ the
2
maximum size of an intersecting k-family of [n] is n−1
k−1 .
Note that we have the requirement of 2k ≤ n since the problem becomes

trivial otherwise: Two sets of size bigger than n/2 always intersect.
124
Proof. The lower bound is easy, just take Fxk = {A ⊆ [n] | x ∈ A, |A| = k}, for
some fixed x ∈ [n]. Clearly, Fxk is an intersecting family of the desired size.
For the upper bound we use a similar approach as in Sperner’s Theorem.
Recall that a circular permutation σ of [n] is an arrangement of the numbers
1, . . . , n on a circle where we do not distinguish between circular shifts, e.g.
53142 ≡ 31425. The number of circular permutations of [n] is (n − 1)!.
We say a circular permutation σ meets a set A ⊆ [n] if the elements of A
appear consecutively on σ. For instance A = {1, 2, 4} is met by σ = 25341 but
B = {2, 3, 4} is not.
Now fix F to be an intersecting k-family.
Claim. For any circular permutation σ, the number of sets A ∈ F met by σ is
at most k.
Proof by Picture. Fix one k-set A0 met by σ, we draw it in red (here k = 5,
n = 12):
Consider a different k-set B ∈ F that is also met by σ (here drawn in blue),

it must intersect A0 . There are 2(k − 1) choices for such a B, (k − 1) for
intersecting A0 at the “clockwise” side and (k − 1) for intersecting A0 at the
“counterclockwise” side. But note that if B ∈ F intersects some part of A0
until some gap between two consecutive elements on σ, then the set B 0 that
intersects the opposite part of A0 starting from that gap cannot be in F (shown
dashed in the picture). This is because B 0 ∩ B = ∅. Here we use that 2k ≤ n
so B and B 0 do not intersect on the opposite side either. This shows that there
are at most 2(k − 1)/2 = k − 1 sets met by σ other than A0 , so at most k in
total. (end claim)
Just like in the proof of Sperner’s Theorem, the number of circular permu-
tations meeting A ∈ F is |A|!(n − |A|)! = k!(n − k)!. We now double count the
set
S := {(A, σ) | A ∈ F and σ is circular permutation meeting A}.
First Way:
X X
|S| = |{A ∈ F | σ meets A}| ≤ k = (n − 1)! · k.
σ σ
Second Way:
X X
|S| = {σ | σ meets A} = k!(n − k)! = |F| · k!(n − k)!
A∈F A∈F
Putting this together we obtain:

(n − 1)!k (n − 1)! n−1
|F | ≤ = = .
k!(n − k)! (k − 1)!(n − k)! k−1
125
5.3.1 Symmetric Chain Partition
We have already
seen in Sperner’s Theorem that the width of the Boolean lattice
Bn is d nn e . Using Dilworth’s Theorem, this means Bn can be partitioned into
n
2
d 2 e chains. We are going to prove a version of this that is stronger in two ways.
n
Firstly, we are going to look at a generalization of Boolean lattices. Secondly, we

find a particularly interesting partition into chains. The next definition captures
what we mean by “interesting”.
Definition 5.20.
• Let P = (X, ≤) be a poset. Then the rank of x ∈ X is the size of a largest
chain beneath x, i.e.
rank(x) = max{|C| | C is a chain with max(C) = x}, x ∈ X.
Consider for instance the following poset in which we annotated each

element with its rank:
3 4
3
2 2 1
• Define rank(P ) to be the maximum rank of an element of P (this is just

a new name for the height of the poset). The above poset has rank 4.
• A poset P is ranked if all maximal chains ending in an element x have the
same size. The poset shown above is not ranked since there are maximal
chains of size 2, 3 and 4 all ending in the same element. Another way to
think about it, is that a poset is ranked iff for any cover relation x l y the
elements x and y have adjacent ranks.
• A chain C is unrefinable if there is no z ∈ X − C with min(C) < z <
max(C) such that C ∪ {z} is a chain. Note that if P is ranked, then C is
unrefinable if and only if it does not skip any rank.
• A chain C is symmetric if it is unrefinable and
rank(min(C)) + rank(max(C)) = rank(P ) + 1.
In Boolean lattices this is a very intuitive concept.

• A symmetric chain decomposition is a partition of the elements of P into
symmetric chains. Figure 37 shows decompositions of some Boolean lat-
tices.
We define B(m1 , m2 , . . . , mk ) to be the poset on all submultisets of M =
{m1 ·1, m2 ·2, . . . , mk ·k} ordered by inclusion. Here, mi is the repetition number
of the element i. The elements 1,2,. . . ,k do not really need to be numbers, we
might as well have chosen instead of 1 and instead of 2. The elements
just need to be distinguishable and numbers happen to be convenient. Recall
that if A, B are multisets with types 1, . . . , k and repetition numbers a1 , . . . , ak
for A and b1 , . . . , bk for B then by A ⊆ B we mean ai ≤ bi for all i.
126
Note that B(1, . . . , 1) = Bk since in that case M = {1, . . . , k}.
| {z }
k
Figure 38 depicts B(1, 2, 1), a three dimensional poset that is similar to B3
except it has an additional “plane” since we can have the element 2 not only
zero or one times, but also twice.
There are many other ways to explain what these posets are, maybe you
prefer one of the following perspectives:
• The poset B(m1 , . . . , mk ) is isomorphic to the poset of all divisors of
mk
pm m2
1 · p2 · . . . · pk
1
ordered by divisibility, where p1 , . . . , pk are distinct
primes. This is illustrated for B(1, 2, 1) and p1 = 2, p2 = 3, p3 = 5 on the
right of Figure 38.
• B(m1 , . . . , mk ) is the product of chains:
B(m1 , . . . , mk ) = C1 × . . . × Ck
where each Ci is a chain on mi + 1 elements. By the product P × Q of two

posets P and Q we mean the posets with elements {(x, y) | x ∈ P, y ∈ Q}
and dominance order, so
(x, y) ≤P ×Q (x0 , y 0 ) :⇔ x ≤P x0 and y ≤Q y 0 .
• Using particularly easy chains we can view B(m1 , . . . , mk ) as the set [m1 +
1] × [m2 + 1] × . . . × [mk + 1] with dominance order ≤Rk .
Note that B(m1 , . . . , mk ) is ranked. A set A = {r1 · 1, . . . , rk · k} has rank
rank(A) = r1 +. . .+rk +1 and the rank of the entire poset is rank(B(m1 , . . . , mk )) =
m1 + . . . + mk + 1.
Theorem 5.21. B(m1 , . . . , mk ) has a symmetric chain decomposition.
Proof. We do induction on k. If k = 1, then B(m1 ) is a (symmetric) chain of
length m1 + 1. So the trivial partition works.
For k ≥ 2, consider the subposet P = B(m1 , . . . , mk−1 , 0) of B(m1 , . . . , mk ),
consisting of those multisets with repetition number 0 for type k. Clearly P is
isomorphic to B(m1 , . . . , mk−1 ) and has, by induction hypothesis, a symmetric
chain decomposition.
1234
123
12 123 124 134 234
1 12 13 23
1 2 12 13 14 23 24 34
∅ 1 2 3
∅ 1 2 3 4
∅
∅
Figure 37: Symmetric chain decompositions of B1 , B2 , B3 and B4 .
127
{1, 2 · 2, 3}
90
{2 · 2, 3} {1, 2, 3}
{1, 2 · 2} 45 30
18
{2, 3} {1, 3} 15 10
{2 · 2} {1, 2} 9 6
{3} 5
3 2
{2} {1}
1
{}
Figure 38: Left: The Hasse Diagram of the poset B(1, 2, 1).
Right: Divisors of 90 = 21 · 32 · 51 ordered by divisibility.
Now the idea is really simple, see Figure 39. Given a symmetric chain
composition of P , we first partition B(m1 , . . . , mk ) into “curtains” that run
along the chains in P and then we partition the curtains (which are essentially
two dimensional grids) into symmetric chains. Now for the formal argument.
If P = C1 ∪ . . . ∪ CR is a partition into symmetric chains, then B(m1 , . . . , mk ) =
Cur1 ∪ . . . ∪ CurR is a partition into R sets where
Curi := {A ∪ {j · k} | A ∈ Ci , 0 ≤ j ≤ mk }.
Now focus on one of the chains C = Ci with elements A1 ⊂ A2 ⊂ . . . ⊂ Al . The

corresponding curtain forms a grid of size mk × l standing on its tip as shown
in Figure 40. Note that it is not necessarily square. The grid can easily be
partitioned into unrefinable chains that look like hooks as shown. If mk + 1 = l
the last hook will be a chain of size 1, otherwise, the last hook will like straight
in the picture, either pointing to the top left if l > mk + 1 or to the top right if
l < mk + 1. We only need to check that all these “hook-chains” are symmetric.
Consider the first hook
A1 ⊂ A2 ⊂ . . . ⊂ Al ⊂ Al ∪ {k} ⊂ . . . ⊂ Al ∪ {mk · k}.
Its minimum has rank |A1 | + 1 and its maximum has rank |Al | + mk + 1. Re-
member that C was symmetric in P so we had |A1 | + 1 + |Al | + 1 = rank(P ) + 1
and get:
|A1 | + 1 + |Al | + mk + 1 = rank(P ) + 1 + mk = rank(B(m1 , . . . , mk )) + 1.
This proves that the first hook is a symmetric chain. Subsequent hooks have
their minima at higher ranks but the ranks of the maxima are correspondingly
lower so it is easy to see that they are symmetric as well.
Figure 37 shows the results of this construction for the Boolean lattices B1 ,
B2 , B3 and B4 . You may want to verify your understanding by constructing
these symmetric chain partitions yourself. Note that in the case of Boolean
lattices all curtains have height 2 and will therefore be partitioned into only one
or two hooks.
128
Figure 39: On the left: Partition of B(2, 3) into three symmetric chains. On
the right: The corresponding partition of B(2, 3, 5) into three “curtains”.
Since Bn grows exponentially in n, it may be impractical to compute the

symmetric chain partition of Bn explicitly for large n. However, there is a simple
implicit characterization that allows us to compute for any set A ⊆ [n] what the
other sets in the symmetric chain containing A are. We state it here without
proof.
The first step is to associate with A ⊆ [n] the characteristic zero-one string
cA that encodes for each i ∈ {1, . . . , 12} in increasing order whether or not i is
in A. For A = {1, 2, 7, 8, 10, 11} ⊆ [12] this would be cA = 110000110110.
Now, roughly speaking, think of the 0s as opening parenthesis and 1s as
closing parenthesis, then cA is a (not necessarily well-formed) parenthesis ex-
pression where some matching pairs can be found. In our example this would
be:
1 1 0 0 0 0 1 1 0 1 1 0
1 2 3 4 5 6 7 8 9 10 11 12
The matched positions are those elements that do not vary within the chain.
The matched 1s form the minimum m of the chain, here m = {7, 8, 10, 11}, the
matched 0s are the elements missing from the maximum of the chain, so here
M = [12] \ {4, 5, 6, 9} = {1, 2, 3, 7, 8, 10, 11, 12}. The unmatched positions, here
{1, 2, 3, 12} will vary within the chain and are added from left to right. So in
the symmetric chain partition of Bn the set A will be part of the symmetric
129
Al ∪ {mk · k}
Al−1 ∪ {mk · k}
·
··
Al ∪ {2 · k} Al−2 ∪ {mk · k}
·
··
Al ∪ {k} Al−1 ∪ {2 · k}
··
·
··
·
Al Al−1 ∪ {k} Al−2 ∪ {2 · k} ... A3 ∪ {mk · k}
Al−1 Al−2 ∪ {k} A2 ∪ {mk · k}
··
·
··
Al−2 · A3 ∪ {2 · k} A1 ∪ {mk · k}
··
·
··
·
A3 ∪ {k} A2 ∪ {2 · k}
··
·
··
·
A3 A2 ∪ {k} A1 ∪ {2 · k}
A2 A1 ∪ {k}
A1
Figure 40: The elements of a curtain can be partitioned into symmetric chains
as shown.
chain:
{7, 8,10, 11} ⊂ {1, 7, 8, 10, 11} ⊂ {1, 2, 7, 8, 10, 11}

⊂{1, 2, 3, 7, 8, 10, 11} ⊂ {1, 2, 3, 7, 8, 10, 11, 12}.
5.4 General Lattices

A poset is a lattice if any two elements x, y have a least upper bound, i.e.
∀x, y ∈ P : ∃z : (z ≥ x and z ≥ y and (∀z 0 : (z 0 ≥ x and z 0 ≥ y) ⇒ z ≤ z 0 )).
We write z = x ∨ y and call z the join of x and y (it is necessarily unique). We

also require that any two elements have a largest lower bound, i.e.
∀x, y ∈ P : ∃z : (z ≤ x and z ≤ y and (∀z 0 : (z 0 ≤ x and z 0 ≤ y) ⇒ z ≥ z 0 )).
We write z = x ∧ y and call z the meet of x and y (it is necessarily unique).

Note that this implies that lattices have a unique minimum which we call 0 and
a unique maximum which we call 1.
130
The following poset P is not a lattice although it has a maximum and mini-
mum. Note that a and b have no join, the upper bounds of a and b are {c, d, 1}
but there is no least upper bound (since c and d are incomparable). For similar
reasons, c and d have no meet. If we modify P by adding an element x as shown
we obtain a lattice L. In it, we have for instance a ∨ b = x, c ∧ d = x.
1 1
c d c d
P := L := x
a b a b
0 0
Note that for two elements x, y of a lattice with x ≤ y we always have x ∧ y = x
and x ∨ y = y. This also implies that 1 is the neutral element of the ∧-operation
and 0 is the neutral element of the ∨ operation. The Boolean lattices Bn really
are lattices where ∧ = ∩ and ∨ = ∪, since A∪B really is the least set containing
A and B and A ∩ B is the largest set contained in A and B.
We now consider Young lattices Y (m, n). Its elements are Ferrer diagrams
with at most m rows and at most n columns ordered by inclusion. Figure 41
shows Y (2, 3).
Figure 41: The Young lattice Y (2, 3).
Young lattices are ranked. Diagrams with s squares have rank s + 1.
131
6 Designs
Assume you are the leader of a local brewery and have just invented seven
kinds of new beers. You are pretty sure all of them are awesome but are curious
whether other people share your opinion. There are some experts, each of which
can judge 3 beers (beyond that point they become tipsy and you do not trust
their judgment).
You want to make sure that each beer is evaluated in contrast to each other
beer, i.e. for each pair of beers there should be one expert that tries both beers.
Actually, make sure that there is exactly one expert for each pair, since experts
are expensive and you don’t have any money to waste.
So can you assign beers to the experts and meet this requirement? It turns
out your problem has a solution with seven experts as shown in Figure 42.
b7
b4 b6
b5
b1 b2 b3
Figure 42: The beers b1 , . . . , b7 are represented by points. Each set of beers that
one expert tries is represented by a straight line or circle. Note that any pair of
points uniquely determines a line or circle containing that pair.
In terms of the following definition we have just found a 2-(7, 3, 1)-design:

v = 7 beers, k = 3 beers per expert, and each pair (t = 2) of beers tasted by
λ = 1 of the experts. This particular design is known as the Fano plane.
Definition 6.1. For numbers t, v, k, λ ∈ N, a t-(v, k, λ) design with point set V
is a multiset B of sets (“blocks”) of points with
(i) |V | = v,
(ii) |B| = k for each B ∈ B,
(iii) each set of t points is a subset of exactly λ blocks.
A t-(v, k, λ) design is also denoted by Sλ (t, k, v) and sometimes called bal-
anced block design.
Example 6.2.
• Designs with k = v, i.e. where each block contains all points, are called
trivial. This gives a t-(v, v, |B|)-design for any t ∈ [v]. From now on we
assume k < v.

• Taking all point sets of size k, i.e. B = [v] v−t
k , yields a t-(v, k, k−t )-design
for any t ∈ [k] (verify this!).
• Consider the point set V = F42 \ {~0}, so all bit-strings of length four
with addition modulo 2, except for ~0 = (0, 0, 0, 0). We have for in-
stance (1, 1, 0, 0) + (1, 0, 1, 0) = (0, 1, 1, 0). Now consider the blocks B =
{{x, y, z} | x, y, z ∈ F42 , x + y + z = ~0}.
132
Note that any pair of points x 6= y uniquely determines a third point z
such that their sum is zero, namely z = x + y (note that x = −x in F42 so
x 6= −y and z = x + y 6= 0). So B is a 2-(v = 15, k = 3, λ = 1)-design.
6.1 (Non-)Existence of Designs

As a first step we want to understand under which conditions on λ, t, k and v
designs with those parameters may exist. In other words we are looking for
necessary conditions on the parameters λ, t, k and v.
Theorem 6.3. If B is a t-(v, k, λ)-design then

(i) the number of blocks in B is |B| = λ · vt / kt ,
(ii) if I ⊆ V is a set of
size i ≤ t, then the number of blocks containing I is
ri = λ · v−i
t−i / k−i
t−i .
Note that for i = t we get rt = λ as it should be. Also note that the
fact that every point is contained in r1 blocks means that B is an r1 -
regular hypergraph (you can safely ignore this remark if don’t know what
hypergraphs are).
Proof. (i) We double count the set
S := {(T, B) | B ∈ B, T ⊆ B, |T | = t}.

Firstly, each set T ⊆ V of size t (and there are vt of those) is contained
in exactly λ blocks since that is what a design requires. So |S| = vt λ.

Secondly, each blocks contains kt subsets of size t, so |S| = |B| · kt .

Together we get: |B| · kt = vt λ from which the claim follows.
(ii) Fix an i-set I and double count the set
S := {(T, B) | B ∈ B, I ⊆ T ⊆ B, |T | = t}.

Firstly, there are v−it−i sets T of size t containing I and for each such T

there are λ blocks B containing T . This gives |S| = v−i t−i · λ.

Secondly, for each of the rI blocks B with I ⊆ B there are k−i t−i sets T

of size t with I ⊆ T ⊆ B. This gives |S| = rI · k−i
t−i .

Together we obtain v−i k−i
t−i · λ = rI · t−i . From this we see that rI actually
only depends on |I| = i and the claim follows.
Remark. We consider t = 2 to be the default case. Sometimes a 2-(v, k, λ) design
is just called a (v, k, λ)-design. In the other notation the parameter λ = 1 can
be omitted so an S1 (t, k, v)-design is simply a S(t, k, v)-design. In that case we
also call it Steiner System (hence the “S”).
Corollary 6.4. If B is a (v, k, λ)-design (so t = 2) the results of the last
Theorem become
v(v−1)
(i) |B| = λ k(k−1) = r kv where r is:
133
v−1
(ii) r = r1 = λ k−1 .
Remark 6.5. For any t-(v, k, λ)-design all the numbers we derived above, such
as |B|, ri (1 ≤ i ≤ t) are integers. In particular, if some choice of parameters
t, k, v, λ does not yield integers, no design with those parameters exists.
For example, in any (v, k = 3, λ = 1)-design, we have r1 = v−1 2 , which is
(v−1)v
integer only if v is odd, and |B| = 6 which is integer only if v ∈ {0, 1, 3, 4}
(mod 6), so together we need v ∈ {1, 3} (mod 6). Note that for v = 3 we get
the trivial design and for v = 7 we get the Fano plane, so things seem to fit so
far.
However, the necessary conditions for the existence of designs are not yet
sufficient. In the following Theorem we derive another necessary condition, this
time a lower bound on v.
Theorem 6.6. In any non-trivial t-(v, k, λ = 1)-design we have v ≥ (t + 1)(k −
t + 1).
Proof. First observe that for A, B ∈ B with A 6= B we have |A ∩ B| ≤ t − 1.

Otherwise some set T ⊆ |A ∩ B| of size t would be contained in both A and B,
contradicting λ = 1.
Now choose some B0 ∈ B and a set S of size t+1 such that |S ∩B0 | = t. This
implies that S is not contained in any B ∈ B (otherwise the t-set S ∩B0 would be
contained in both B0 and B). Still each of the t-sets T ⊆ S must be contained
in some block BT . Consider two such blocks BT and BT 0 . Each contains k
elements, t of which are in S and k − t are outside of S. The intersection
BT ∩ BT 0 certainly contains T ∩ T 0 which is already of size t − 1. This means
(by the observation in the beginning of the proof) that BT and BT 0 must not
have any more further intersections outside of S.
B{c,d,a} B{b,c,d}
d c
a b
B{d,a,b} B0 = B{a,b,c}
Figure 43: Consider the case t = 3 and k = 9. After choosing some B0 ∈ B

we find a set S = {a, b, c, d} of size t + 1 that intersects B0 in exactly t = 3
points a, b, c. Now no block may contain S (as argued above) but each set
{a, b, c}, {b, c, d}, {c, d, a}, {d, a, b} must be contained in some block. These four
blocks must be disjoint outside of S (they already intersect in t − 1 elements
within S). So we count k − t = 6 elements for each of them plus the elements
from S, so v ≥ 24 + 4 = 28.
This means for each of the blocks BT (and there are t + 1 of them) that
there are k − t elements outside of BT that are not contained in any other BT 0 .
134
Together with the elements from S this gives
v ≥ (t + 1) + (t + 1)(k − t) = (t + 1)(k − t + 1).

| {z } | {z }
∈S k − t for each BT
So we can conclude, for instance, that no design with parameters λ = 1,

t = 10, v = 72 and k = 16 exists (even though the divisibility conditions of
Theorem 6.3 are satisfied).
Remark. In a recent, so far unpublished paper10 Peter Keevash shows that for
any parameters λ, t, k there exist t-(v, k, λ) designs for sufficiently large v. The
proof is very sophisticated so we will not include it here.
In the next Theorem, we find a lower bound to the number of blocks of a
design.
Theorem 6.7 (Fisher’s inequality). Let B be a non-trivial (v, k, λ)-design (so
k < v). Then |B| ≥ v.
Proof. Consider the incidence matrix A with v rows, |B| columns and values
(ap,B )p∈V,B∈B defined as
(
1, p ∈ B,
ap,B :=
0, otherwise.
If ap , aq are rows in A, then ap · aTq is the number of blocks containing both p

and q. This value is λ for p 6= q and r = r1 for p = q. So we have
 
r λ ··· ··· λ
 
 λ r λ 
.
 
..
 
T  .. .. 
A·A = . . .
.
 
..
..
..
 . 
 . 
 . r λ 
.
..
λ ··· ··· λ r
The determinant is det(A · AT ) = rk(r − λ)v−1 (verify this!).

v−1
By Corollary 6.4(i) we have r = λ k−1 > λ. So the determinant is not zero
and thus the matrix has full rank, i.e. rank v. This means that the matrix
A must also have rank at least v. In particular A has at least v columns and
therefore |B| ≥ v as claimed.
6.2 Construction of Designs

So far we talked about necessary conditions for the existence of designs without
actually constructing any – which is our goal in the following. We have to start
with a few definitions, though.
Definition 6.8.
• A (v, k, λ)-design with |B| = v is called symmetric. For those designs the
matrix A we discussed above is square.
10 http://arxiv.org/abs/1401.3665 submitted on 15 Jan 2014.
135
• A symmetric (v, k, 1)-design is called projective plane. The Fano plane
from Figure 42 is such a projective plane with 7 points and 7 blocks
(“lines”).
• A (v, k = 3, λ = 1)-design or, equivalently, a Steiner System with param-

eters S(t = 2, k = 3, v) is called a Steiner Triple System.
In the following we write Zn for the cyclic group with n elements, in other
words, the numbers modulo n.
Definition 6.9. Let 2 ≤ k < v and λ ≥ 1. A (v, k, λ)-difference set is a set

D = {d1 , . . . , dk } ⊆ Zv such that each element in Zv \ {0} has multiplicity λ in
the multiset {di − dj | i, j ∈ [k], i 6= j}. In other words, each non-zero number
can be written as the difference of two numbers from D in exactly λ ways.
Example 6.10. The set {1, 3, 4, 5, 9} ⊆ Z11 is a (11,5,2)-difference set, since
each of the numbers can be written as a difference in two ways, as follows:
Number First Way Second Way

1 4−3 5−4
2 3−1 5−3
3 1−9 4−1
4 5−1 9−5
5 9−4 3−9
6 9−3 4−9
7 1−5 5−9
8 1−4 9−1
9 1−3 3−5
10 3−4 4−5
Theorem 6.11. If D = {d1 , . . . , dk } is a (v, k, λ)-difference set then
B := {D, 1 + D, 2 + D, . . . , v − 1 + D}
is a symmetric (v, k, λ)-design. Here, by a + D we mean the set {a + d | d ∈ D}.

Proof. Note that there are k · (k − 1) ways to form terms i − j with i 6= j and
i, j ∈ D. On the other hand, each non-zero number can be written in exactly λ
ways like this, so λ(v − 1) = k · (k − 1) which implies λ = k(k−1)
v−1 < k.
From this we conclude that the blocks we proposed above are all distinct:
Assume a + D = b + D for a 6= b. Then we have a permutation π ∈ Sk such
that a + di = b + dπ(i) for all i ∈ [k]. Then a − b = dπ(i) − di for each i ∈ [k] so
we have k > λ ways to write a − b as a difference, contradicting the property of
the difference set.
By construction we have v points in total and blocks of size k. The only thing
left to check is that each pair of two distinct elements x, y ∈ Zv is contained in
exactly λ of the blocks. Let d = x − y be the difference of those two elements.
Then there are exactly λ ways to pick two elements xi , yi ∈ D with difference d
and for each such pair there is a unique shift a such that x = xi + a, y = yi + a.
This implies that x, y ∈ a + D for exactly λ choices of a.
136
Example. If v is a prime power with v ≡ 3 (mod 4) then
{a2 | a ∈ Zv , a2 6= 0}
v−1 v−3
is a (v, k, λ)-difference set with k = 2 and λ = 4 . We do not prove this.
6.3 Projective Planes

We should mention that projective planes actually come from geometry. There,
a projective plane of order q is defined as a set of q 2 + q + 1 points with a family
of subsets of points (lines) such that (i) each line is of size q + 1 and (ii) each
pair of points is contained in a unique line.
In our setting, projective planes of order q are symmetric (v, k, λ = 1)-designs
where k = q + 1. In that case we can indeed conclude
sym. Cor.6.4 v
v = |B| = r ⇒ r =k =q+1
k
so each point is contained in q + 1 lines and
Cor.6.4 v−1
r = ⇒ v = (q + 1)q + 1 = q 2 + q + 1,
k−1
so the number of points matches as well. Before we continue, note the following
property:
Claim. Any two lines (blocks) of a projective plane intersect in a unique point.
Proof of Claim. Consider two distinct lines L1 , L2 and some x ∈ L1 \ L2 . Then
consider the family of lines (Ly )y∈L2 where Ly is the unique line containing x
and y. These lines are all distinct (otherwise we would find some y1 , y2 contained
in two lines: L2 and Ly1 ), so we found q + 1 lines that all contain x. This means
we must have counted L1 in the process, otherwise there would be q + 2 lines
containing x (contradicting r = q + 1). We conclude L1 ∩ L2 6= ∅. Clearly
|L1 ∩ L2 | ≤ 1, otherwise we would have a pair of points contained in L1 and L2
(contradicting λ = 1).
Construction of Projective Planes

In the following construction of a projective plane of order q we use Fq to denote
a field of size q where q is a prime power.
The points of the plane are equivalence classes of the set X = {(x0 , x1 , x2 ) |
xi ∈ Fq , (x0 , x1 , x2 ) 6= 0} where two vectors are equivalent if they are scalar
multiples of one another. Formally, a point [x0 , x1 , x2 ] is the set:
[x0 , x1 , x2 ] = {c · (x0 , x1 , x2 ) | c ∈ Fq \ {0}}.
Note the same point can be described in different ways, for instance [x0 , x1 , x2 ] =
[2x0 , 2x1 , 2x2 ] (for q 6= 2). Since |X| = q 3 − 1 and q − 1 elements represent the
3
−1
same point (are in the same class), we have qq−1 = q 2 + q + 1 points in total,
as desired.
For (a0 , a1 , a2 ) ∈ F3q \ {~0} we define the line L([a0 , a1 , a2 ]) as
L([a0 , a1 , a2 ]) = {[x0 , x1 , x2 ] | a0 x0 + a1 x1 + a2 x2 = 0}.
137
Note that lines are well defined, that is the definition respects the equivalence
classes (either all elements representing a point satisfy the equation of the line
or neither of them).
The number of solutions to a0 x0 + a1 x1 + a2 x2 is q 2 since, without loss of
generality, a2 6= 0 and thus x0 and x1 can be chosen arbitrarily and x2 is then
uniquely determined. Disregarding the solution x0 = x1 = x2 = 0 (which does
not represent any point) we have q 2 − 1 solutions in X that are part of lines,
2
−1
meaning qq−1 = q + 1 points are part of the line, as desired.
Now consider two points [x0 , x1 , x2 ] 6= [y0 , y1 , y2 ]. We show that they are
contained in a unique line. Indeed, if L[a0 , a1 , a2 ] contains both, then we have
a0 x0 + a1 x1 + a2 x2 = 0, a0 y0 + a1 y1 + a2 y2 = 0. This is a homogeneous sys-
tem with two equations and three variables a0 , a1 , a2 , so there exists a solution
(a0 , a1 , a2 ) 6= 0 and all solutions are of the form c · (a0 , a1 , a2 ). All those solution
triples define the same line, so L[a0 , a1 , a2 ] is uniquely determined.
This concludes the construction (and the verification thereof).
Remark. It is conjectured that the order of every projective plane is a prime
power, but no proof is known.
Theorem 6.12 (Bruck-Ryser-Chowla-Theorem). If v, k, λ ∈ N with λ(v − 1) =

k(k − 1) and a symmetric (v, k, λ)-design exists then
• if v is even, then k − λ is a square,
v−1
• if v is odd, then z 2 = (k − λ)x2 + (−1) 2 λy 2 has a solution (x, y, z) 6= 0.
6.4 Steiner Triple Systems

We leave projective planes for now and come back to Steiner Triple Systems, i.e.
(v, k = 3, λ = 1)-designs. From Remark 6.5 we know that v ∈ {1, 3} (mod 6).
The following Theorem shows that this necessary condition on the existence is
also sufficient.
Theorem 6.13. For each v ∈ N with v ∈ {1, 3} (mod 6) a Steiner Triple
System with v points exists.
Proof. The construction is different for v ≡ 1 and v ≡ 3 (mod 6). We start

with the latter case.
The points are V = Z2n+1 × Z3 so v = (2n + 1) · 3 = 6n + 3 ≡ 3 (mod 6).
There are two types of blocks:
type 1: {(x, 0), (x, 1), (x, 2)} for each x ∈ Z2n+1 ,
type 2: {(x, i), (y, i), ( x+y
2 , i + 1)} for all x 6= y and each i ∈ Z3 .
We need to show that each pair of elements is contained in a unique block. So

consider a pair (x, i) 6= (y, j).
Case 1: x = y: Then the pair is contained in the block {(x, 0), (x, 1), (x, 2)}
(of type 1) and in no other block.
Case 2: x 6= y, i = j: The pair is contained in the block {(x, i), (y, i), ( x+y
2 , i+
1)} (of type 2) and in no other block.
138
Case 3: x 6= y, i 6= j: Assume without loss of generality that j = i + 1 (since
i, j ∈ Z3 we always have this or j = i − 1). Then the pair is contained
0
in the block {(x, i), (y 0 , i), ( x+y 0
2 , i + 1)} where y = 2y − x. The pair is
contained in no other block.
This concludes the case of v ≡ 3 (mod 6), we proceed with v ≡ 1 (mod 6).
The construction and verification is more complicated.
As point set, choose V = (Z2n × Z3 ) ∪ {∞}. To simplify notation we write xi
to denote the pairs (x, i) ∈ Z2n × Z3 . The element ∞ is special and we assume
x + ∞ = ∞ for any x ∈ V . Before we define the blocks of the Triple System,
we define four types of base blocks first:
• {00 , 01 , 02 },
• {∞, 00 , n1 }, {∞, 01 , n2 }, {∞, 02 , n0 },
• {00 , x1 , (−x)1 }, {01 , x2 , (−x)2 }, {02 , x0 , (−x)0 }, for each x ∈ [n − 1],

• {n0 , x1 , (1 − x)1 }, {n1 , x2 , (1 − x)2 }, {n2 , x0 , (1 − x)0 }, for each x ∈ [n].
The blocks of the Steiner Triple System are
B = {a0 + B | a ∈ {0, 1, . . . , n − 1}, B is base block}.
where the addition of elements is defined coordinate wise, in particular a0 + xi = (a+

x)i .
We need to check that there is a unique block containing u, v with u 6= v.
We do not consider all cases here, just two important ones.
Case 1: u = ∞, v = (x, i): If x ≤ n − 1, then the block is x0 + {∞, 0i , ni+1 }.

If x ≥ n then we find (x − n)0 + {∞, 0i−1 , ni }.
Case 2: u = xi , v = yi : Without loss of generality x < y. Either y − x = 2s
(the difference is even), then xi , yi is contained in the block (y − s)0 +
{0i−1 , si , (−s)i }. Or y − x = 2s − 1 then xi , yi is contained in the block
(y − s)0 + {ni−1 , si , (1 − s)i }. ...
6.5 Resolvable Designs

Consider a game played by three players (like skat). Assume there is a tourna-
ment for this game, with n players. Each set of three players should play exactly
λ times against each other. There are several time slots during which the games
are scheduled. Of course every player can only play one game at a time. In an
optimal schedule all players have a game in each time slot. In the sense of the
following definition, such a schedule corresponds to a resolvable (n, 3, λ)-design,
where parallel classes correspond to the time slots.
Definition 6.14. A parallel class of a (v, k, λ)-design is a set of disjoint blocks

forming a partition of the point set V .
A partition of B into parallel classes is called resolution. A design is resolvable
if it has a resolution.
139
Note that projective planes are examples for non-resolvable designs since no
disjoint blocks exist. An example for a resolvable (v = 4, k = 2, λ = 1)-design
with the corresponding parallel classes is shown in the next picture, where the
points are A, B, C, D and the blocks are depicted by edges.
D C D C D C D C
−→ and and
A B A B A B A B
This is a special case of a (v = q 2 , k = q, λ = 1)-design. Such designs are called
affine planes and are, as we show now, always resolvable.
Theorem 6.15. Any (v = q 2 , k = q, λ = 1)-design is resolvable.
Proof. The main observation is the following:
/ B there is a unique block B 0 with x ∈ B 0
Claim. For each block B and x ∈
0
and B ∩ B = ∅.
Proof of Claim. For each y ∈ B there is a unique block By such that {x, y} ∈
By . These blocks are distinct (since |By ∩ B| ≤ 1) and all contain x. Because
Cor.6.4 2
−1
of r = λ k−1 v−1
= qq−1 = q + 1 there is exactly one block left that contains x
that is different from each By . It is disjoint from B.
With the claim proved, everything falls into place. Start with a maximal set
of pairwise disjoint blocks. This set forms a parallel class (this is not obvious, but
easy to prove). Then, in the next phase, take some other set of disjoint blocks
not yet considered. This forms a parallel class in the same way. Continue like
this until all blocks are handled (it is easy to verify that this works).
6.6 Latin Squares

In the following we consider a structure that seems unrelated to designs at first,
but the final Theorem will establish an interesting connection to affine planes.
A Latin square of order n ≥ 1 is an n × n array filled with numbers from the
cyclic group Zn = {0, . . . , n − 1} such that each row and each column contains
each number from Zn exactly once. Consider for instance the following two
Latin squares
1 0 2 3 1 0 3 2
0 1 3 2 3 2 1 0
A= B=
2 3 1 0 0 1 2 3
3 2 0 1 2 3 0 1.
We denote the set of positions containing the number i in a Latin square L by
L(i), in the example above we have for instance B(3) = {(1, 3), (2, 1), (3, 4), (4, 2)}.
Another way to think about Latin squares is to say that they are a partition
L(0) ∪ . . . ∪ L(n − 1) of [n] × [n] such that each set L(i) are a set of positions
such that rooks (from the game chess) placed onto those positions cannot attack
each other. Every solution to a Sudoku puzzle is a Latin square with n = 9,
although the reverse is not true: Latin squares do not respect the constraint for
the 3 × 3 boxes.
There are many ways to transform a Latin square to similar Latin squares:
You could permute the rows or columns or “rename” the numbers (swap A(1)
140
and A(0) for instance). At least the latter operation seems uninteresting as
renaming numbers yields the same partition, just with different labels. To
capture the idea of “entirely different Latin squares, we propose the notion
of orthogonality defined in the following.
Let A, B be Latin squares of order n with entries A = (aij )i,j , B = (bij )i,j .
The juxtaposition of A and B is the n × n array where each position simply
contains the corresponding numbers of A and B, i.e.
A ⊗ B = ((aij , bij ))i,j with entries in Zn × Zn .
For instance, with A and B from the example above we get
(1,1) (0,0) (2,3) (3,2)

(0,3) (1,2) (3,1) (2,0)
A⊗B =
(2,0) (3,1) (1,2) (0,3)
(3,2) (2,3) (0,0) (1,1) .
We say A, B are orthogonal if A ⊗ B contains each pair from Zn × Zn exactly

once. The two Latin squares above are not orthogonal because some pairs occur
twice (e.g. (2, 0)) and, necessarily, some pairs therefore occur not at all (e.g.
(1, 3)).
An example for two Latin squares A and C with A orthogonal to C is
1 0 2 3 1 0 2 3 (1,1) (0,0) (2,2) (3,3)

0 1 3 2 3 2 0 1 (0,3) (1,2) (3,0) (2,1)
A= , C= , A⊗C =
2 3 1 0 0 1 3 2 (2,0) (3,1) (1,3) (0,2)
3 2 0 1 2 3 1 0 (3,2) (2,3) (0,1) (1,0) .
Note that this notion of orthogonality has little to do with geometric notions
orthogonality you may be familiar with. Observe the following:
• If A is orthogonal to B, then B is orthogonal to A.
• If A is orthogonal to B, then “renaming” numbers in A or B (for instance
replace every 0 with a 1 and vice versa) preserves orthogonality.
Remark. For n ∈ {2, 6} there are no two orthogonal Latin squares of order n.
For n = 2 this is easy to see since the only Latin squares with n = 2 are
0 1 1 0
L1 = and L2 =
1 0 0 1.
They are not orthogonal to each other nor to themselves (only for n = 1 can a
Latin square be orthogonal to itself). For n = 6 the argument is non-trivial.
For n ∈ N \ {2, 6}, a pair of orthogonal Latin squares of order n exists. The
proof of this is highly non-trivial, in fact it took over a hundred years to show
that there is a pair of orthogonal Latin squares of order 10.
Our goal in the following is to construct a large set A1 , . . . , Ak of Latin
squares of order n that are MOLS (mutually orthogonal Latin squares), meaning
Ai is orthogonal to Aj for i 6= j.
Theorem 6.16. Let n be a positive integer, r ∈ [n − 1] non-zero and co-prime
to n, i.e. gcd(n, r) = 1. Then Lrn = (r · i + j (mod n))i,j is a Latin square.
141
To clarify how Lrn looks, consider the example n = 5 and r = 2. “Going
right” corresponds to +1 and going down corresponds to +2 which gives
3 4 0 1 2
0 1 2 3 4
L25 = 2 3 4 0 1
4 0 1 2 3
1 2 3 4 0.
Proof. We have to show that each row and column contains each number exactly
once. For rows this is clear, since the i-th row contains r·i+1, r·i+2, . . . , r·i+n,
which traverses all numbers modulo n. The column j contains r + j, 2r +
j, . . . , n · r + j, all modulo n. Assume two of those numbers are identical, say
i1 · r + j ≡ i2 · r + j we conclude (i1 − i2 ) · r ≡ 0 (mod n). Since gcd(n, r) = 1
(meaning r has an inverse modulo n) we actually had i1 = i2 . So no column
contains a number twice and Lrn really is a Latin square of order n.
Theorem 6.17. If n is prime, then L1n , . . . , Ln−1
n are n − 1 MOLS of order n.
Proof. Since n is prime, gcd(n, i) = 1 for all i ∈ [n − 1] so by Theorem 6.16
L1n , . . . , Ln−1
n are Latin squares of order n.
Now consider two of those squares Lrn and Lsn with r 6= s. We need to show
that they are orthogonal, so suppose some pair of numbers from Zn ×Zn appears
in Lrn ⊗ Lsn in positions (i, j) and (k, l). By definition of Lrn and Lsn this gives
the identity:
(r · i + j, s · i + j) = (r · k + l, s · k + l).
So r · (i − k) = l − j = s · (i − k) which implies (r − s)(i − k) = 0. The numbers
modulo n form a field and a product can only be zero if one of the factors is
zero. Since r 6= s we obtain i = k and therefore also l = j. In particular we
showed that the same pair cannot appear in distinct positions and of Lrn ⊗ Lsn ,
so Lrn is orthogonal to Lsn as desired.
Remark. For a prime power n = pk there is a field Fn = {α0 = 0, α1 , . . . , αn−1 }
of order n. For it we can define Lα
n = (αr · αi + αj )i,j and generalize Theorem
r
6.16 and Theorem 6.17 accordingly.

Lemma 6.18. For any n ≥ 2 there are at most n − 1 MOLS of order n.
Proof. Assume we have a set A1 , . . . , Ak of MOLS. As orthogonality is not
affected by “renaming” the numbers in a Latin square, we may assume without
loss of generality that each Ai has 0 1 2 . . . n − 1 as its first row. Now
consider the entry at position (2, 1). If A is orthogonal to B then a21 6= b21
because all pairs (0, 0), . . . , (n − 1, n − 1) already appear in row 1 of A ⊗ B. So
the entries in (2, 1) must be mutually distinct. They must also be non-zero since
the first column of each Ai already contains zero in position (1, 1). Therefore,
we started with at most k ≤ n − 1 Latin squares.
And now for the theorem that establishes the connection to designs. We will
only prove the equivalence of (i) and (iv), but we will do so constructively.
Theorem 6.19. For n ≥ 2 each of the following is equivalent:
(i) There exist n − 1 mutually orthogonal Latin squares of order n.
142
(ii) There exists a finite field of order n.
(iii) n is a prime power, i.e. n = pk .
(iv) There exists a (v = n2 , k = n, λ = 1)-design (an affine plane).

Proof. (i) ⇒ (iv). Let A1 , . . . , An−1 be n − 1 MOLS of order n. We need to
construct a (v = n2 , k = n, λ = 1)-design. Recall (from Corollary 6.4)
that the number of blocks in such a design is necessarily |B| = n(n + 1)
and from Theorem 6.15 that the design is necessarily resolvable, so it splits
into n + 1 parallel classes of n disjoint blocks each (this resolution will be
apparent).
The points of the design are [n] × [n], so the set of positions in Latin
squares of order n.
Recall how each Latin square A is a partition of the positions [n] × [n] =
A(0)∪. . .∪A(n−1) where A(i) are the positions containing number i ∈ Zn .
These will be blocks of the design. In addition, each row R(i) = {(i, j) |
j ∈ [n]} and each column C(j) = {i, j) | i ∈ [n]} is a block. So in total
the blocks are:
B = {Ar (s) | r ∈ [n − 1], s ∈ Zn } ∪ {R(i) | i ∈ [n]} ∪ {C(j) | j ∈ [n]}.
We need to show that each pair of distinct points (i, j) and (k, l) appears
in exactly one block. We first show that they are contained in at most one
block.
Case 1: i = k. Both points are contained in the i-th row so both are
in R(i). They cannot be contained in the same column (otherwise
they would not be distinct) and they are not both contained in any
Ar (s) since that would mean that the Latin square Ar contains the
number s twice in the i-th row. In particular, no block other than
R(i) contains both points.
Case 2: j = l. Similar to Case 1, the points are contained in C(j) and
in no other block.
Case 3: i 6= k, j 6= l. The points are not in the same row or column so
in no R(i) or C(j). But assume (for contradiction) that we have
(i, j), (k, l) ∈ Ar (s) ∩ At (u) for (r, s) 6= (t, u). Since the blocks orig-
inating from the same Latin square are disjoint (since those blocks
form a partition), we have r 6= t. So in Ar there is the number s
in both positions (i.e. at (i, j) and (k, l)) and in At there is u in
both positions. This means Ar ⊗ At has (s, u) in both positions, con-
tradicting the fact that Ar and At are orthogonal. So each pair of
positions is contained in at most one block.
To see that each pair of positions is contained in at least one block we

double count the set:
S = {((i, j), (k, l), B) | B ∈ B, (i, j) 6= (k, l), B contains (i, j) and (k, l)}.
143
X
Firstly, |S| = {B | B contains (i, j) and (k, l)}
(i,j)6=(k,l)
X
n2
≤ 1= . (uses “at most one”)
2
(i,j)6=(k,l)
X
Secondly, |S| = {((i, j), (k, l)) | (i, j) 6= (k, l), B contains (i, j) and (k, l)}
B
X |B| 2
n n
= = (n2 + n) = .
2 2 2
B
So we realize that the “≤” is actually an equality so in particular each

pair of points is contained in exactly one block.
(iv) ⇒ (i) Assume we have a (v = n2 , k = n, λ = 1)-design. As argued before

(Corollary 6.4), there are |B| = (n + 1) · n blocks in total. Since affine
planes are resolvable by Theorem 6.15 we get n + 1 parallel classes B =
B1 ∪ . . . ∪ Bn+1 consisting of n blocks each. We use two of these parallel
classes B1 and B2 to find “coordinates” for the n2 points and the remaining
n − 1 parallel classes to define MOLS.
We label the blocks from B1 and B2 as B1 = {R(1), R(2), . . . , R(n)} and
B2 = {C(1), C(2), . . . , C(n)}. For i, j ∈ [n], we know that every pair
of blocks R(i) and C(j) intersects in exactly one point pij : Because of
t = 2 and λ = 1, they cannot intersect in two points and because of
the claim from the proof of Theorem 6.15 they cannot be disjoint. This
means that the blocks from B1 and B2 intersect as shown in Figure 44 so
C(1) C(2) C(3) C(4) C(5)
R(1)
R(2)
R(3)
R(4)
R(5)
Figure 44: The two different parallel classes B1 and B2 intersect like this when
arranging the points accordingly.
we can interpret the blocks R(i) as rows and C(j) as columns (i, j ∈ [n]).
To simplify notation, identify the point point pij of the design with the
position (i, j) ∈ [n] × [n] in Latin squares. Then for each l ∈ {3, . . . , n + 1}
we interpret the n blocks Bl (0), Bl (2), . . . , Bl (n − 1) of the parallel class
Bl as a Latin square where the positions containing a number r ∈ Zn
are given by Bl (r). This really is a valid Latin square since each Bl (r)
intersects each R(i) and C(j) in exactly one point so each number occurs
in exactly one row and column.
144
It is also easy to verify that the n − 1 Latin squares we get from B3 to
Bn+1 are mutually orthogonal, since for any distinct s, t ∈ {3, . . . , n + 1}
and r1 , r2 ∈ Zn we know that |Bs (r1 ) ∩ Bt (r2 )| = 1 which just means that
the pair (r1 , r2 ) ∈ Zn × Zn occurs exactly once in the juxtaposition of the
Latin square for Bs and Bt .
References
[Kas61] Pieter W. Kasteleyn. The statistics of dimers on a lattice: I. the number
of dimer arrangements on a quadratic lattice. Physica, 27(12):1209–
1225, 1961.
[TF61] H. N. V. Temperley and Michael E. Fisher. Dimer problem in statistical
mechanics—an exact result. Philos. Mag. (8), 6:1061–1063, 1961.
145

Lecture Notes Combinatorics

Uploaded by

Copyright:

Available Formats

Lecture Notes Combinatorics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture Notes Combinatorics

Uploaded by

Copyright:

Available Formats

Lecture Notes

Lecture Notes by Stefan Walzer (TU Ilmenau)

1 Permutations and Combinations 9

5 Partially Ordered Sets 107

Combinatorics is concerned with:

• The existence, enumeration, analysis and optimization of discrete struc-

Existence: We want to arrange elements in a set into patterns satisfying

Enumeration: Assume certain arrangements are possible. How many

Classification: Assume there are many arrangements. Do some of these

Figure 1: The 6 × 8 board can be tiled with 24 dominoes. The 5 × 5 board

Figure 2: Our polyominoes have size k × 1. We use k colours (1 = white, k

# squares coloured with 1 = ks1 s2 + s1 r2 + s2 r1 + r2

The special case of an 8 × 8 board is already non-trivial:

T1 = {T | T allows a horizontal cut but no vertical cut},

Figure 4 shows one tiling for each of these three classes.

As another example consider the board B consisting of two 4 × 4 boards

of B can be partitioned into

T1 = {T | T matches both extra squares to the left}

1 Permutations and Combinations

1.1 Basic Counting Principles

1.1.1 Addition Principle

1.1.2 Multiplication Principle

|S| = |T | − |S| = 23905 − 20178 = 3727.

1.1.4 Bijection Principle

|S| = |T | ⇔ there exists a bijection between S and T .

1.1.5 Pigeonhole Principle

1.1.6 Double counting

Second way: Every handshake involves two guests. So

Together, we see that n(n − 1) = M = 2N and hence we obtain the identity

Hence we obtain another identity

1.2 Ordered Arrangements – Strings, Maps and Products

sequence: If X is a set of numbers, then s is often called a sequence and

k-Permutation: For 1 ≤ k ≤ |X| a k-permutation of X is an ordered arrange-

i+s≡j (mod k) ⇒ π1 (i) = π2 (j)

This equivalence relation partitions P (n, k) into equivalence classes. A

Figure 8: Visualization of the 5-permutations π1 , π2 and π3 where values are

are circular equivalents as witnessed by the shift s = 3. They are therefore

Proof. (i) A permutation is an injective map π : [k] → [n]. We count the

(ii) We count P (n, k) in two ways:

1.3 Unordered Arrangements –

The difference between ordered and unordered arrangements is that ordered

k-Combination of a Multiset: Let X be a finite set of types and let M be

Theorem 1.4. For 0 ≤ k ≤ n, we have

Proof. To build an element

1.4 Multinomial Coefficients

Example. Verify by multiplying out:

(x1 + x2 + x3 )4 = 1·(x41 + x42 + x43 )

Theorem 1.6 (Multinomial Theorem).

Proof. The naive way to multiply out (x1 + . . . + xr )n would be:

Also, (n − k1 − . . . − kr )! = (n − n)! = 0! = 1 so we end up with:

More generally, for n ≥ 1 and k1 , . . . , kr ≥ 0 with k1 + . . . + kr = n, we have

so it really is a special case of the second claim, which we prove now. By

Exploring a different way to expand (x1 + . . . + xr )n , we obtain:

(x1 + . . . + xr )n = (x1 + . . . + xr ) · (x1 + . . . + xr )n−1

By substituting indices as ki := ki0 + 1 and kj := kj0 (for j 6= i) this is equal to:

Now we have two ways to write the coefficients of (x1 + . . . + xr )n and

Proof. Label the k types as a1 , . . . , ak . In an n-permutation there are n positions

for the combination {2 · ,2 · ,1 · }.

The original k-combination can be reconstructed from this sequence of and •: