Incidence matrices

CJ Quines
May 14, 2022

Warmup
1. In a certain committee, each member belongs to exactly three subcommittees, and each
subcommittee has exactly three members. Prove that the number of members equals the
number of subcommittees.
Sketch: Count the number of (subcommittee, member) pairs such that the member is in
the subcommittee. There are n subcommittees with 3 members each, and m members in 3
subcommittees each. Thus 3n = 3m so n = m.

2. Let A1 , A2 , . . . , A6 be 4-element subsets of S = {1, 2, . . . , 8}. Suppose each element in S is in

exactly r of the Ai s. Find r.
Sketch: Count the number of (subset, element) pairs such that the element is in the subset.
There are six subsets with four elements each, and eight elements in r subsets each. Thus
6 × 4 = 8r so r = 3.

Definitions
Here’s the setup. Let A1 , A2 , . . . , Ar be subsets of {1, 2, . . . , c}. (A set of subsets of S is called a
family of subsets of S.) Consider an r × c matrix, where the ith row and jth column has 1 if j ∈ Ai
and 0 otherwise. This is the incidence matrix of the subsets.
For example: {1, 2, 3, 4, 5, 6} has subsets {1, 2, 3}, {2, 4, 6}, {3, 4, 5}. Its incidence matrix is
 
1 1 1 0 0 0
0 1 0 1 0 1 .
0 0 1 1 1 0

A family of subsets is also known as an incidence structure. This is because we can interpret this
geometrically, where the elements of the base set are points, and the subsets are “lines” that contain
these points. Here’s the above incidence matrix, represented geometrically:

Incidence structures can also be interpreted as edges in a bipartite graph. On one side, your
vertices are the elements. On the other side, your vertices are the subsets. An edge joins an element
to each subset it belongs to.
The incidence matrix is not a method, just a tool. The method we’ll explore is double counting:
first counting by rows, then counting by columns.
2 CJ Quines

Examples
1. (Baltic Way 2001) A set of 8 problems was prepared for an exam. Each student was given 3
of them. No two students received more than one common problem. What is the maximum
number of students?
Sketch: Construct an incidence matrix where columns are problems and rows are students.
We want to count the total number of ones.
Count by rows: If there are s students, there are s rows with 3 each, giving 3s in total.
Count by columns: We want the number of ones in each column, but we don’t know that,
so instead we bound. Suppose a column had 4 ones. Then the rows that had those ones must
have 2 other ones, so there are 8 more ones among those rows. But no two of these ones can
be in the same column, otherwise there would be two students who got two or more common
problems. There are only 7 remaining columns, so this is impossible. Thus, there are at most
3 ones in each column. There are 8 columns with at most 3 each, giving at most 24 in total.
Thus 3s ≤ 24, so s ≤ 8. It remains to find a construction.
Remark 1: Please remember how we bounded the number of ones in each column. We won’t
have another example with this trick, but it will appear several times in the problems.
Remark 2: Here’s a cool idea that almost works. Construct an incidence matrix where
columns are pairs of problems and rows are students. An entry would be 1 if the student was
given that pair of problems, and 0 otherwise. Then each pair of problems could’ve been given
to at most one student, as otherwise two students would’ve had two common problems. This
gives us 3s ≤ 82 , or s ≤ 9, which isn’t a strong enough bound.

2. (China 1996) Eight singers participate in an art festival where several songs are performed.
Each song is performed by 4 singers, and each pair of singers performs together in λ songs.
Determine the minimum value of λ.
Sketch 1: Construct an incidence matrix where columns are singers and rows are songs. The
trick is to count the number of “row pairs” of ones: that is, the number of pairs of ones that
lie in the same row. In this case, rows represent songs, so pairs of ones in the same row are
pairs of singers performing a song.
Count by rows: Suppose there are b rows. Each row has 4 ones, giving 6 row pairs, and 6b
pairs in total.
Count by columns: Pick a pair of columns; there are 28 of them. Each pair of columns
gives λ row pairs. This gives 28λ pairs in total.
Thus 6b = 28λ, so λ ≥ 3. It remains to find a construction; this is actually the painful part.
Sketch 2: We still count row pairs of ones, but this time, we fix a column and count only the
pairs that have a one in this column. In other words, we’re counting the pairs of singers who
sing in the same song, with one of the singers fixed. Suppose this given singer sang r songs.
Count by rows: Suppose the given singer sang r songs. Each of the r songs gives 3 row
pairs, and 3r pairs in total.
Count by columns: Each of the other 7 singers sang λ songs with this singer, giving 7λ
pairs in total.
Thus 3r = 7λ and λ ≥ 3; it remains to find a construction.
3 CJ Quines

Remark: For these problems, the write-ups are easier to understand if we don’t mention
incidence matrices (or rows, or columns) at all. They’re just a tool to help us find what to
double count.

3. (IMC 2002) Two hundred students participated in a mathematical contest. They had six
problems to solve. It is known that each problem was correctly solved by at least 120
participants. Prove that there must be two participants such that every problem was solved
by at least one of these two students.
Sketch: Construct an incidence matrix where columns are problems and rows are students.
Assume the opposite. That means each pair of students didn’t solve at least one problem. In
other words, each pair of rows has at least one column where they’re both zero! That means
we want to count column pairs of zeroes.
Count by rows: Pick a pair of rows; there are 200


2 = 19 900 of them. Each pair of rows
gives at least one column pair, so at least 19 900 in total.
Count by columns: Each column has at least 120
ones, so there are at most 80 zeroes in
each column. There are six columns, so at most 6 80
2 = 18 960 in total.
Combining the two counts gives a contradiction.
Remark: The way I’d write this solution would be as follows. Suppose the contrary. Let T
be the number of tuples (student, student, problem), such that there are two distinct students
and a problem that neither one solved. Then [. . .] so T ≥ 19 900. Also [. . .] so T ≤ 18 960.
Thus 19 900 ≤ T ≤ 18 960, contradiction.

4. In a chess club, n people gathered to play chess against each other, as they pleased. No two
people played against each other more than once. At the end of the day, it was observed that
a total of 3n games had been played. Moreover, if we choose any two players, say A and B,
there would be at most one other player who had played with both A and B. Prove that
n > 30.
Sketch: Construct an incidence matrix where columns and rows are people. We count the
column pairs of ones.
Count by rows: Pick a pair of rows; there are n2 of them. Each pair of rows gives at most

one column pair, as at
most one person played with the people represented by these two rows.
There are at most n2 column pairs in total.
Count Pby columns: Suppose the ith column had ci ones. Because there are a total of 3n
games, ci = 6n. Applying Jensen’s, we get
n   P   
X ci ( ci ) /n 6
≥n =n = 15n.
2 2 2
i=1

n


Combining the two counts, 15n ≤ 2 , thus n ≥ 31 as desired.
Remark: When we can’t get good column bounds, we can sometimes assign variables and use
Jensen’s. Using Jensen’s like this is too weak, however, because equality isn’t always possible.
a+b

For example, it’s true that a2 + 2b ≥ 2 22 . But if a + b is odd, you can get a better lower



a+b−1
a+b+1

bound of 2
2
+ 2
2
.
4 CJ Quines

Problems
1. (Balanced block designs) Let A1 , A2 , . . . , Ab be k-element subsets of S = {1, 2, . . . , v}. Suppose
each pair of elements in S is in exactly λ of the Ai s. Prove that each element in S is contained
in exactly r of the Ai s, for r = λ(v − 1)/(k − 1) = bk/v. Hint: 4
2. (USA TST 2005) Let n be an integer greater than 1. For a positive integer m, let Sm =
{1, 2, . . . , mn}. Suppose there exists a 2n-element set T such that
a) each element of T is an m-element subset of Sm ,
b) each pair of elements of T shares at most one common element, and
c) each element of Sm is contained in exactly two elements of T .
Determine the maximum possible value of m in terms of n. Hint: 9

3. Let S1 , S2 , . . . , S5n be subsets of A = {1, 2, . . . , 100} such that every element of A appears in
exactly 4n of the Si s. Prove there exists i, j, k such that Si ∪ Sj ∪ Sk = A.
4. (China 1994?) Let A1 , A2 , . . . , Ak be 5-element subsets of S = {1, 2, . . . , 10} such that
|Ai ∩ Aj | ≤ 2 for 1 ≤ i < j ≤ k. Find the maximum value of k. Hint: 5
5. (China TST 1992) Sixteen students took part in a math competition where every problem
was a multiple choice question with four choices. After the contest, it is found that any two
students had at most one answer in common. Determine the maximum number of questions.
Hint: 5

6. (IMO 1998) In a competition, there are a contestants and b judges, where b ≥ 3 is an odd
integer. Each judge rates each contestant as either “pass” or “fail”. Suppose k is a number
such that for any two judges, their ratings coincide for at most k contestants. Prove that
k b−1
≥ .
a 2b
7. (USA 2011) Let X be a set with |X| = 225. Suppose further there are eleven subsets
A1 , . . . , A11 of X such that |Ai | = 45 for 1 ≤ i ≤ 11 and |Ai ∩ Aj | = 9 for 1 ≤ i < j ≤ 11.
Prove that |A1 ∪ · · · ∪ A11 | ≥ 165, and give an example for which equality holds. Hint: 9

Harder problems
8. (Singapore 2010) Let n be a positive integer. Find the smallest positive integer k with the
property that for any coloring of the squares of a 2n × k chessboard with n colors, there are 2
columns and 2 rows such that the 4 squares in their intersections have the same color. Hint: 3
9. (USA 1979) An organization has n members, and it has n+1 distinct three-member committees.
Prove that there are two committees that share exactly one member. Hint: 6
10. (Singapore 2006) Let n be a positive integer. Let S1 , S2 , . . . , Sk be a collection of 2n-element
subsets of {1, 2, . . . , 4n − 1, 4n} so that Si ∩Sj contains at most n elements for all 1 ≤ i < j ≤ k.
Show that k ≤ 6(n+1)/2 . Hint: 1
11. (USA 1984) A math exam has two papers, where each paper consists of at least one question
and both papers have 28 questions altogether. Each pupil attempted 7 questions. Each pair
of questions was attempted by exactly two pupils. Show that one pupil attempted either zero
or at least four questions in the first paper. Hint: 2
5 CJ Quines

12. (China 1993) A group of 10 people went to a bookstore. It is known that everyone bought
exactly 3 books, and for every two people, there is at least one book that both of them bought.
What is the least number of people that could have bought the book purchased by the greatest
number of people? Hint: 7

13. (IMO 2001) Twenty-one girls and twenty-one boys took part in a mathematical competition.
It turned out that each contestant solved at most six problems, and for each pair of a girl and
a boy, there was at least one problem that was solved by both the girl and the boy. Show that
there is a problem that was solved by at least three girls and at least three boys. Hint: 8

14. (IMO 2005) In a mathematical competition 6 problems were posed to the contestants. Each
pair of problems was solved by more than 25 of the contestants. Nobody solved all 6 problems.
Show that there were at least 2 contestants who solved exactly 5 problems.

Hints
1. Use Remark 2 from the first example.

2. How many students are there? How many problems did each solve?

3. The case n = 4 is similar to the China TST 1992 problem.

4. This generalizes the second example.

5. For the construction, try to put things into sets greedily, and that’ll probably work.

6. Induct on n.

7. First show the answer is at least 4. Then show it can’t be 4.

8. Bound triplets of matching problems in rows and columns.

9. For the construction, represent something geometrically, like the figure on page 1.

Sketches
1. We count row pairs of ones, where one of the ones is in a given column. Count by rows: If the
given column has r ones, each gives k − 1 pairs along its row. Count by columns: If we pick
one of the other v − 1 columns, there are λ rows that contain a pair. Thus r(k − 1) = λ(v − 1).
The bk = vr follows from counting all the ones.
Remark: Note that v, k, and λ determine b and r. We call the family of subsets a 2-balanced
design, because the subsets are balanced with respect to the 2-element subsets of S. We
want to prove that every 2-balanced design is also 1-balanced. In fact, for every s < t, every
t-balanced design is also s-balanced; the proof is similar.

2. We count column pairs of ones. Count by rows: There are 2n

2 pairs of rows, and each pair
of row has at most one
2n

pair. Count by columns: Each of the mn columns has exactly two
ones. Thus mn ≤ 2 , so m ≤ 2n − 1. It remains to find a construction; consider 2n lines in
general position.

3. Assume otherwise; we count column triplets of zeroes. Count by rows: There are 5n


3
triplets of rows, each giving at least one column pair. Count by columns: Each column has
n zeroes, giving 100 n3 triplets. Thus 5n n




3 ≤ 100 3 , contradiction after some algebra.
6 CJ Quines

4. We count column pairs of ones. Count by rows: There are k2 pairs of rows, each giving at

most two Pcolumn

ci

pairs.k/2Count

by columns: The total
k/2

number
k

of ones is 5k, so by Jensen,
there are 2 ≥ 10 2 column pairs. Thus 10 2 ≤ 2 2 , so k ≤ 6. It remains to find a
construction, e.g. 01234, 01567, 02589, 13689, 24679, 34578.

5. Construct the incidence matrix where columns are problems, rows are students, and entries
are the students’ answers. We count column pairs of matching answers. Count by rows:
16


There are 2 pairs of rows, each giving at most one pair. Count by columns: Suppose
there are p columns. By Jensen, each column gives at least 4 42 pairs. Thus 4p 42 ≤ 16




2 , so
p ≤ 5. It remains to find a construction; consider families of parallel lines in F4 × F4 .

6. Construct the incidence matrix where columns are contestants and
rows are judges. We count
column pairs of matching ratings. Count by rows: There are 2b pairs of rows, each giving
at most
k pairs.
Count by columns: By Jensen, each of the a columns
gives at least
(b−1)/2 (b+1)/2 2 2 b
2 + 2 = (b − 1) /4 pairs. Combining, we get a(b − 1) /4 ≤ k 2 , as desired.

7. We count column pairs of ones. Count by rows: There are 11

2 pairs of rows, each giving 9
column pairs. Count by columns: Suppose there are n columns, P and the ith

that 495/n
column had
ci
ci ones. There are a total of 45 × 11 ones. By Jensen, we get 2 ≥ n 2 . Combining,
11 495/n



we get 9 2 ≥ n 2 , or n ≥ 165. It remains to find a construction; consider 11 planes in
general position.

References
Problems were mostly lifted from Loh’s Combinatorics of Sets External-Link-Alt handout, Shi-Jie’s Incidence Matrix
External-Link-Alt handout, and Zhao’s Double Counting External-Link-Alt handout.