Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

Thus the expected value is

0 1
1 @ 1
E[S] = (n 1)! + (n 1)! + · · · + (n 1)!A = · n · (n 1)! = 1.
n! | {z } n!
n times

Cute, right? Now let’s bring out the artillery.

§2.2 Linearity of Expectation

The crux result of this section is the following theorem.

Theorem 2.2 (Linearity of Expectation)

Given any random variables X1 , X2 , . . . , Xn , we always have

E[X1 + X2 + · · · + Xn ] = E[X1 ] + E[X2 ] + · · · + E[Xn ].

This theorem is obvious if the X1 , X2 , . . . , Xn are independent of each other – if I roll

100 dice, I expect an average of 350. Duh. The wonderful thing is that this holds even if
the variables are not independent. And the basic idea is just the double-counting we did
in the earlier example: even if the variables depend on each other, if you look only at the
expected value, you can still add just by columns. The proof of the theorem is just a
bunch of sigma signs which say exactly the same thing, so I won’t bother including it.
Anyways, that means we can now nuke our original problem. The trick is to define
indicator variables as follows: for each i = 1, 2, . . . , n let
(
def 1 if person i gets his own name tag
Si =
0 otherwise.

Obviously,
S = S 1 + S 2 + · · · + Sn .
Moreover, it is easy to see that E[Si ] = P(Si = 1) = n1 for each i: if we look at any
particular person, the probability they get their own name tag is simply n1 . Therefore,

1 1 1
E[S] = E[S1 ] + E[S2 ] + · · · + E[Sn ] = + + · · · + = 1.
n
| n {z n
}
n times

Now that was a lot easier! By working in the context of expected value, we get a
framework where the “double-counting” idea is basically automatic. In other words,
linearity of expectation lets us only focus on small, local components when computing an
expected value, without having to think about why it works.

§2.3 More Examples

Example 2.3 (HMMT 2006)

At a nursery, 2006 babies sit in a circle. Suddenly, each baby randomly pokes either
the baby to its left or to its right. What is the expected value of the number of
unpoked babies?

3
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

Solution. Number the babies 1, 2, . . . , 2006. Define

(
def 1 if baby i is unpoked
Xi =
0 otherwise.

We seek E[X1 + X2 + · · · + X2006 ]. Note that any particular baby has probability 1 2
2 = 1
4
of being unpoked (if both its neighbors miss). Hence E[Xi ] = 14 for each i, and

1 1003
E[X1 + X2 + · · · + X2006 ] = E[X1 ] + E[X2 ] + · · · + E[X2006 ] = 2006 · = .
4 2
Seriously, this should feel like cheating.

§2.4 Practice Problems

The first two problems are somewhat straightforward applications of the methods de-
scribed above.
Problem 2.4 (AHSME 1989). Suppose that 7 boys and 13 girls line up in a row. Let
S be the number of places in the row where a boy and a girl are standing next to each
other. For example, for the row GBBGGGBGBGGGBGBGGBGG we have S = 12.
Find the expected value of S.
Problem 2.5 (AIME 2006 #6). Let S be the set of real numbers that can be represented
as repeating decimals of the form 0.abc where a, b, c are distinct digits. Find the sum of
the elements of S.
The next three problems are harder; in these problems linearity of expectation is not
the main idea of the solution. All problems below were written by Lewis Chen.
Problem 2.6 (NIMO 4.3). One day, a bishop and a knight were on squares in the same
row of an infinite chessboard, when a huge meteor storm occurred, placing a meteor in
each square on the chessboard independently and randomly with probability p. Neither
the bishop nor the knight were hit, but their movement may have been obstructed by
the meteors. For what value of p is the expected number of valid squares that the bishop
can move to (in one move) equal to the expected number of squares that the knight can
move to (in one move)?
Problem 2.7 (NIMO 7.3). Richard has a four infinitely large piles of coins: a pile of
pennies, a pile of nickels, a pile of dimes, and a pile of quarters. He chooses one pile at
random and takes one coin from that pile. Richard then repeats this process until the
sum of the values of the coins he has taken is an integer number of dollars. What is the
expected value of this final sum of money, in cents?
Problem 2.8 (NIMO 5.6). Tom has a scientific calculator. Unfortunately, all keys
are broken except for one row: 1, 2, 3, + and -. Tom presses a sequence of 5 random
keystrokes; at each stroke, each key is equally likely to be pressed. The calculator then
evaluates the entire expression, yielding a result of E. Find the expected value of E.
(Note: Negative numbers are permitted, so 13-22 gives E = 9. Any excess operators
are parsed as signs, so -2-+3 gives E = 5 and -+-31 gives E = 31. Trailing operators
are discarded, so 2++-+ gives E = 2. A string consisting only of operators, such as -++-+,
gives E = 0.)

4
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

§3 Direct Existence Proofs

In its simplest form, we can use expected value to show existence as follows: suppose
we know that the average score of the USAMO 2014 was 12.51. Then there exists a
contestant who got at least 13 points, and a contestant who got at most 12 points. This
is similar in spirit to the pigeonhole principle, but the probabilistic phrasing is far more
robust.

§3.1 A First Example

Let’s look at a very simple example, taken from the midterm of a class at the San Jose
State University.2

Example 3.1 (SJSU M179 Midterm)

Prove that any subgraph of Kn,n with at least n2 n+1 edges has a perfect matching.

We illustrate the case n = 4 in the figure.

Figure 1: The case n = 4. There are n2 n + 1 = 13 edges, and the matching is

highlighted in green.

This problem doesn’t “feel” like it should be very hard. After all, there’s only a total
of n2 possible edges, so having n2 n + 1 edges means we have practically all edges
present.3
So let’s be really careless and just randomly pair off one set of points with the other,
regardless of whether there is actually an edge present. We call the score of such a pairing
the number of pairs which are actually connected by an edge. We wish to show that
some pairing has score n, as this will be the desired perfect matching.
So what’s the expected value of a random pairing? Let v1 , . . . , vn be the n vertices on
the left. For each i, let4 ,
(
def 1 if the pair with vi has an edge
Xi =
0 otherwise.
deg vi
Then the score of the configuration is X = X1 +X2 +· · ·+Xn . Now we have E[Xi ] = n ,
so
E[X] = E[X1 ] + · · · + E[Xn ]
2
For a phrasing of the problem without graph theory: given n red points and n blue points, suppose we
connect at least n2 n + 1 pairs of opposite colors. Prove that we can select n segments, no two of
which share an endpoint.
3
On the other hand, n2 n + 1 is actually the best bound possible. Can you construct a counterexample
with n2 n?
4
Thanks to D. Grozev for a correction here

5
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

deg v1 deg v2 deg vn

= + + ··· +
n n n
n2 n + 1 1
= =n 1+ .
n n
Since X takes only integer values, there must be some configuration which achieves
X = n. Thus, we’re done.

§3.2 Ramsey Numbers

Let’s do another simple example. Before we begin, I will quickly introduce a silly algebraic
lemma, taken from [5, page 30].

Lemma 3.2
For any positive integers n and k,
✓ ◆
n 1 ⇣ en ⌘k
< .
k e k

Here e ⇡ 2.718 . . . is Euler’s constant.

Proof. Do n
k < nk
k! and then use calculus to prove that k! e(k/e)k . Specifically,
Z k
ln 1 + ln 2 + · · · + ln k ln x dx = k ln k k + 1
x=1

whence exponentiating works.

Algebra isn’t much fun, but at least it’s easy. Let’s get back to the combinatorics.

Example 3.3 (Ramsey Numbers)

Let n and k be integers with n  2k/2 and k 3. Then it is possible to color
the edges of the complete graph on n vertices each either red or blue with the
following property: one cannot find k vertices for which the k2 edges among them
are monochromatic.

Remark. In the language of Ramsey numbers, prove that R(k, k) > 2k/2 .

Solution. Again we just randomly color the edges and hope for the best. We use a coin
flip to determine the color of each of the n2 edges. Let’s call a collection of k vertices
bad if all k2 edges are the same color. The probability that any collection is bad is
✓ ◆( k ) 1
1 2
.
2
The number of collections in n
k , so the expected number of bad collections is
n
E[number of bad collections] = k
.
2( 2 )
k
1

We just want to show this is less than 1. You can check this fairly easily using Lemma 3.2;
in fact, we have a lot of room to spare.

6
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

§3.3 A Tricky Application

To cap off this section, we give a tricky proof (communicated to me via [3]) of the
following result.

Theorem 3.4 (Ajtai-Komlós-Szemerédi)

Given a triangle-free graph G with average degree d and N vertices, we can find an
independent set with size at least 0.01 Nd log d.

Here, triangle-free just means there are no three vertices which are all adjacent to each
other.5 Another phrase for this is locally sparse.
Our first move is to try and replace the “average degree” d with “maximum degree”
. Here’s the trick: notice that at most half of the vertices have degree greater than 2d.
So if we throw away these vertices, we still have half the vertices and left, and now the
maximum degree is  2d. If we let n = N /2, then we just need an independent set of
size 0.04 n log in our new graph.
So now we have n vertices with maximum degree . Here’s the trick: consider all
possible independent sets, and pick one set S uniformly at random (!). For this set S, we
define a score X as follows:

• For each vertex u in S, we write a + at that u.

• For each vertex v adjacent to something in S, we write +1 at that vertex. A vertex

can receive +1 multiple times. However, note that since S is independent, this
means v 2/ S.

• Define the score X to be the sum of all numbers written.

+1

+1

+
+ +2

+1

Figure 2: Assigning scores. The elements of S are the large red vertices.

Obviously, X  2 |S|, since each vertex in S bestows to itself and at most

among its neighbors. Now, we will place a bound on E[X], which will give us the result.
Consider any vertex v, and consider its set of neighbors. Note that by the triangle-free
condition, no neighbors arePadjacent to each other. Let Xv denote the sum of the scores
given to vertex v (so X = v Xv ). We are going to show that E[Xv ] 0.08 log . This
is enough, because then E[X] 0.08n log , and for a good choice of X, we then have
|S| 0.04n log .

5
If you’re familiar with the notation R(m, n), here’s some food for thought: what’s the connection between
this and R(3, t)?

7
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

Neighbors

v ...

Figure 3: Ignoring things.

Suppose we’re selecting an independent set, and we’re done selecting everything aside
from v and its neighbors. We’ll prove that regardless of how the stuff outside is chosen,
E[Xv ] 0.08 log still holds. Assume that, not including v, there are m other vertices in
the neighborhood which we can still pick (i.e. they are not adjacent to anything outside
that has been selected).
There are a few ways we can pick the remaining set:

• We can pick v, but then we can no longer pick any of its neighbors.

• We can pick any nonempty subset of the m remaining vertices, but then we can no
longer pick v.

• We can pick no vertices.

There are a total of 1 + (2m 1) + 1 possibilities. In the first scenario, the Xv = + . In

the second and third scenario, Xv = E[# neighbors chosen] = 12 m. So,
⇢
1· + 2m · 12 m 1 m 1
E[Xv ] = = m + · > max ,m .
2m + 1 2 +1 1+2 m 2 4 2m

It remains to prove this is at least 0.08 log . You can check this, because if m 1
log2 ,
p 2
then 14 m is enough; otherwise, 2m which is certainly sufficient.

§3.4 Practice Problems

The first two problems are from [2]; the last one is from [4].
Problem 3.5. Show that one can construct a (round-robin) tournament with more than
1000 people such that in any set of 1000 people, some contestant beats all of them.
Problem 3.6 (BAMO 2004). Consider n real numbers, not all zero, with sum zero.
Prove that one can label the numbers as a1 , a2 , . . . , an such that

a1 a2 + a2 a3 + · · · + an a1 < 0.

Problem 3.7 (Russia 1996). In the Duma there are 1600 delegates, who have formed
16000 committees of 80 people each. Prove that one can find two committees having no
fewer than four common members.

8
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

§4 Heavy Machinery
Here are some really nice ideas used in modern theory. Unfortunately I couldn’t find
many olympiad problems that used them. If you know of any, please let me know!

§4.1 Alteration
In previous arguments we often proved a result by showing E[bad] < 1. A second method
is to select some things, find the expected value of the number of “bad” situations, and
subtract that off. An example will make this clear.

Example 4.1 (Weak Turán)

A graph G has n vertices and average degree d. Prove that it is possible to select an
independent set of size at least 2d
n
.

Proof. Rather than selecting 2d n

vertices randomly and hoping the number of edges is 1,
we’ll instead select each vertex with probability p. (We will pick a good choice of p later.)
That means the expected number of vertices we will take is np. Now there are 12 nd
edges, so the expected number of “bad” situations (i.e. an edge in which both vertices
are taken) is 12 nd · p2 .
Now we can just get rid of all the bad situations. For each bad edge, delete one of its
endpoints arbitrarily (possibly with overlap). This costs us at most 12 nd · p2 vertices, so
the expected value of the number of vertices left is

1 2 1
np ndp = np 1 dp .
2 2
It seems like a good choice of p is d,
1
which now gives us an expected value of 2d ,
n
as
desired.
A stronger result is Problem 6.5.

§4.2 Union Bounds and Markov’s Inequality

A second way to establish existence is to establish a nonzero probability. One way to do
this is using a union bound.

Proposition 4.2 (Union Bound)

Consider several events A1 , A2 , . . . , Ak . If

P(A1 ) + P(A2 ) + · · · + P(Ak ) < 1

then there is a nonzero probability that none of the events occur.

The following assertion is sometimes useful for this purpose.

Theorem 4.3 (Markov’s Inequality)

Let X be a random variable taking only nonnegative values. Suppose E[X] = c.
Then
1
P(X rc)  .
r

9
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

§5.2 Partial Solution Using Alteration

p
We’ll show the bound c n for any c < 23 .
First, we need to bound the number of triangles.

Claim — There are at most 13 n2 triangles.

Proof. Consider each of the n2 intersection of two lines. One can check it is the vertex
of at most two triangles. Since each triangle has three vertices, this implies there are at
most 23 n2 < 13 n2 triangles.

It is also not hard to show there are at most 12 n2 finite regions6 .

Now color each line blue with probability p. The expected value of the number of lines
chosen is
E[lines] = np.
The expected number of completely blue triangles is less than
1
E[bad triangles] < n2 · p3 .
3
For the other finite regions, of which there are at most 12 n2 , the probability they are
completely blue is at most p4 . So the expected number of completely blue regions here is
at most
1
E[bad polygons with 4+ sides] < n2 · p4 .
2
Note that the expected number of quadrilaterals (and higher) is really small compared
to any of the preceding quantities; we didn’t even bother subtracting off the triangles
that we already counted earlier. It’s just here for completeness, but we expect that it’s
going to die out pretty soon.
Now we do our alteration – for each bad, completely blue region, we un-blue one line.
Hence the expected number of lines which are blue afterwards is
✓ 2 ◆ ✓ 2 ◆ ✓ ◆
n 3 n 4 np2 np3
np ·p · p = np 1 .
3 2 3 2
p
Ignore the rightmost np2 for now, since it’s really small. We want p = k/ n for some k;
3

the value is roughly k · (1 k 2 /3) at this point, so an optimal value of p is p = n 1/2

(that is, k = 1); this gives
✓ ◆
p 2 27 1 2p 81
n· p = n .
3 16 n 3 32
p
For n sufficiently large, this exceeds c n, as desired.

§5.3 Interlude – Sketch of Official Solution Obtaining c = 1

This is not probabilistic, but we include it for completeness anyways. It is in fact just a
greedy algorithm.
Suppose we have colored k of the lines blue, and that it is not possible to color any
additional lines. That means any of the n k non-blue lines is the side of some finite
region with an otherwise entirely blue perimeter. For each such line `, select one such

12
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

Problem 6.2 (Romania 2004). Prove that for any complex numbers z1 , z2 , . . . , zn ,
satisfying |z1 |2 + |z2 |2 + · · · + |zn |2 = 1, one can select "1 , "2 , . . . , "n 2 { 1, 1} such that
n
X
"k zk  1.
k=1

Problem 6.3 (Shortlist 1999 C4). Let A be a set of N residues (mod N 2 ). Prove that
there exists a set B of of N residues (mod N 2 ) such that A + B = {a + b|a 2 A, b 2 B}
contains at least half of all the residues (mod N 2 ).
Problem 6.4 (Iran TST 2008/6). Suppose 799 teams participate in a round-robin
tournament. Prove that one can find two disjoint groups A and B of seven teams each
such that all teams in A defeated all teams in B.
Problem 6.5 (Caro-Wei Theorem). Consider a graph G with vertex set V . Prove that
one can find an independent set with size at least
X 1
.
deg v + 1
v2V

Remark. Note that, by applying Jensen’s inequality, our independent set has size at least
d+1 , where d is the average degree. This result is called Turán’s Theorem (or the
n

complement thereof).

Problem 6.6 (USAMO 2012/6). For integer n 2, let x1 , x2 , . . . , xn be real numbers

satisfying x1 +x2 +. . .+xn = 0 and x21 +x22 +. . .+x2n = 1. For each subset A ✓ {1, 2, . . . , n},
define X
SA = xi .
i2A

(If A is the empty set, then SA = 0.) Prove that for any positive number , the number
of sets A satisfying SA is at most 2n 3 / 2 . For which choices of x1 , x2 , . . . , xn ,
does equality hold?
Problem 6.7 (Online Math Open, Ray Li). Kevin has 2n 1 cookies, each labeled with
a unique nonempty subset of {1, 2, . . . , n}. Each day, he chooses one cookie uniformly at
random out of the cookies not yet eaten. Then, he eats that cookie, and all remaining
cookies that are labeled with a subset of that cookie. Compute the expected value of the
number of days that Kevin eats a cookie before all cookies are gone.
Problem 6.8. Let n be a positive integer. Let ak denote the number of permutations of
n elements with k fixed points. Compute

a1 + 4a2 + 9a3 + · · · + n2 an .

Problem 6.9 (Russia 1999). In a certain school, every boy likes at least one girl. Prove
that we can find a set S of at least half the students in the school such that each boy in
S likes an odd number of girls in S.
Problem 6.10 (Sperner). Consider N distinct subsets S1 , S2 , . . . , SN of {1, 2, . . . , n}
such that no Si is a subset of any Sj . Prove that
✓ ◆
n
N  ⌅1 ⇧ .
2n

15
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

Problem 6.11. Let n be a positive integer. Suppose 11n points are arranged in a circle,
colored with one of n colors, so that each color appears exactly 11 times. Prove that one
can select a point of every color such that no two are adjacent.
Problem 6.12 (Sweden 2010, adapted). In a town with n people, any two people either
know each other,por they both know someone in common. Prove that one can find a
group of at most n log n + 1 people, such that anyone else knows at least one person in
the group.

Remark. In graph theoretic language

p – given a graph with diameter 2, prove that a
dominating set of size at most n log n + 1 exists.

Problem 6.13 (Erdös). Prove that in any set S of n distinct positive integers we can
always find a subset T with 13 n or more elements with the property that a + b 6= c for
any a, b, c 2 T (not necessarily distinct).

Remark. Such sets are called sum-free.

Problem 6.14 (Korea 2016). Let U be a set of m triangles. Prove that there exists a
subset W ✓ U with at least 0.45m0.8 triangles, with the following property: there are no
points A, B, C, D, E, F for which ABC, BCD, CDE, DEF , EF A, F AB are all in W .

§7 Solution Sketches
2.4 Answer: 9.1. Make an indicator variable for each adjacent pair.

2.5 Answer: 360. Pick a, b, c randomly and compute E[0.abc]. Then multiply by |S|.

2.6 8(1 p) = 4 · (1 p) + (1 p)2 + (1 p)3 + . . . .

2.7 Let xn be the EV at a state with n (mod 100). Then x0 = 0 and

1
xn = ((xn+1 + 1) + (xn+5 + 5) + (xn+10 + 10) + (xn+25 + 25)) .
4
Do algebra.

2.8 Answer: 1866. Show that one can replace + or - buttons with STOP. Show that one
can replace 1 and 3 buttons with 2. Let p = 35 . Compute 2(p + 10p2 + · · · + 104 p5 ).

3.5 Suppose there are n people, and decide each edge with a coin flip. Compute the
expected number of 1000-subsets for which there is no one better than all. Check that
this is less than 1 for very large n.

3.6 Show that a random permutations has expected value at most 0. Why are the
inequalities strict?

3.7 Let ni be the number of committees which the ith delegate is in. Pick two committees
randomly and find the expected value
P ofnithe number of common members. Use Jensen’s
inequality to get a lower bound on 2 .

16
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

6.1 Pick the contestants randomly. Find the expected number of problems both miss.

6.2 Select each of the "i randomly with a coin flip. Square the left-hand side and use
the fact that |z|2 = zz for any z.

6.3 Randomly selecting B works; you can even permit repeated elements in B. You may
need the inequality 1 n1  1e .
n

6.4 Let dk be the number of teams which defeat the kth team (here 1  k  799). Select
A randomly and compute the expected number of teams dominated by everyone in A.
You need Jensen on the function x7 .

6.5 Use the following greedy algorithm – pick a random vertex, then delete it and all its
neighbors. Repeat until everything is gone.

6.6 Compute E[SA

2 ] for a random choice of A. Markov Inequality.

6.7 The number of days equals the number of times a cookie is chosen. Compute the
probability any particular cookie is chosen; i.e. the expected value of the number of times
the cookie is chosen. Sum up.

6.8 For a random permutation let X be the number of fixed points. We already know
E[X] = 1. Compute E[ X2 ]. Use this to obtain E[X 2 ].

6.9 Use a coin flip to decide whether to select each girl, then take as many boys as
possible. Show that any person, girl or boy, has exactly a 50% chance of being chosen.

6.10 First prove that

N
X 1
n  1.
k=1 |Sk |

To do this, consider a random maximal chain of subsets

; = T0 ⇢ T1 ⇢ T2 ⇢ · · · ⇢ Tn = {1, 2, . . . , n}.

Compute the expected number of intersections of this chain with {S1 , S2 , . . . , SN }.

6.11 LLL. Here p = 11 2 and d = 42.

6.12 If any vertex has small degree, p then its neighbors are already the desired set. So
assume all degrees are greater than n log n. Pick each person with probability p for
some well-chosen p; then we expect to pick np people. Show that the probability someone
fails is less than n1 and use a union bound. The inequality 1 p  e p is helpful.

6.13 Work modulo a huge prime p = 3k + 2. Find a nice sum-free (mod p) set U of size
k + 1 first, and then consider Un = {nx | x 2 U } for a random choice of n. Compute
E[|S \ Un |].

6.14 Fix U and use alteration. Add a triangle to W with probability p, then for every
bad 6-tuple contained in W , delete one of the triangles from W .

17
Evan Chen《陳誼廷》 — 11 August 2014 Expected Uses of Probability

References
[1] pythag011 at http://www.aops.com/Forum/viewtopic.php?f=133&t=481300

[2] Ravi B’s collection of problems, available at:

http://www.aops.com/Forum/viewtopic.php?p=1943887#p1943887.

[3] Problem 6 talk (c > 1) by Po-Shen Loh, USA leader, at the IMO 2014.

[4] Also MOP lecture notes: http://math.cmu.edu/~ploh/olympiad.shtml.

[5] Lecture notes by Holden Lee from an MIT course:

http://web.mit.edu/~holden1/www/coursework/math/18997/notes.pdf

Thanks to all the sources above. Other nice reads that I went through while preparing
this, but eventually did not use:

1. Alon and Spencer’s The Probabilistic Method. The first four chapters are here:
http://cs.nyu.edu/cs/faculty/spencer/nogabook/.
2. A MathCamp lecture that gets the girth-chromatic number result:
http://math.ucsb.edu/~padraic/mathcamp_2010/class_graph_theory_probabilistic/
lecture2_girth_chromatic.pdf

18