Axioms and assumptions

18.600 Problem Set 2, due February 21

Welcome to your second 18.600 problem set! We will continue to explore some combinatorics
along with probability and the axioms of probability.
Before we get to work, let’s indulge in a bit of reflection. When we say “The probability
that A will happen is p” where does p come from? Sometimes the evidence convinces pretty
much everyone that A will or will not happen. Informally, the probability that a predicted lunar
eclipse will happen on schedule is pretty much 1, and the probability that Mars and Jupiter
will collide this month is pretty much 0. In other simple situations (die rolls, coin tosses, etc.)
experience may lead us to agree on probabilities that aren’t 0 or 1. The assumption that all
outcomes are equally likely (for random permutations or die rolls or coin tosses) is sometimes a
natural starting point. This assumption is implicitly made in a few of the problems here.
In more complicated real world settings, one can sometimes define the risk neutral proba-
bility, a probability measure derived from the market prices of contracts whose values depend
on future events. If we want to know the risk neutral probability that a given candidate will
win an election, or that an athletic team will win a game, we can look at betting markets.
(Check out [Link], [Link], [Link], [Link], and similar
websites.) As we will see later in the course, if we want to know the risk neutral probability
that the price of a share of Apple stock will exceed some value by the end of the year, we can
work this out by looking at current prices of derivatives (contracts whose future value depends
on future share prices). The total amount of money at stake in derivative markets is estimated
at over a quadrillion dollars per year (try googling derivatives quadrillion).
Some argue that betting markets set up perverse incentives. If I buy a contract that gives
me $500,000 if my house burns down, that’s useful insurance. But if I buy a contract that gives
me $500,000 if your house burns down, that gives me an unhealthy incentive to burn your house
down. People similarly worry about a world in which hedge funds can bet that a company will
collapse and then actively cause it to collapse. Rules are required to prevent such things, but
foolproof (and evil-genius-proof) rules are hard to design and enforce. Preventing bettors from
deliberately spreading misinformation (in order to manipulate prices) is also a challenge.
On the other hand, one might argue that the absence of betting markets is part of the reason
that some questions in politics and law are divisive. It is hard to place a bet on the proposition
that “my candidate would do more to advance long term happiness and prosperity than yours”
or “my client is innocent,” so there is no market mechanism for producing a commonly accepted
probability. Different groups can claim to have different probability estimates, the expression
of which may advance their own agendas, but without a market we cannot tell which parties
would actually be willing to bet money at the corresponding rates. Some studies claim that
people answering questions about the economy are both more accurate and less partisan when
they are paid (even a very small amount) for correct answers. Maybe there is something to be
said for having money on the line.

1
A. FROM ROSS 8TH EDITION CHAPTER TWO:

1. Card Problem: A 5-card hand is dealt from a well-shuffled deck of 52 playing cards.
What is the probability that exactly three of the four suits are represented in the hand?

2. Theoretical Exercise 6: Let E, F , and G be three events. Find expressions for the
events so that, of E, F , and G,
(a) all three events occur;
(b) none of the events occur;
(c) only E occurs;
(d) both E and G, but not F , occur
(e) at least one of the events occur;
(f) at most one of the events occurs;
3. Theoretical Exercise 20: Consider an experiment whose sample space consists of a
countably infinite number of points. Show that not all points can be equally likely. Can
all points have a positive probability of occurring?
B. In Essay Admission Country, each applicant submits one essay to each of 12 universities.
There are 12 distinct types of essays (personal vignette, accomplishment overview, political
manifesto, absurdist ramble, etc.) and each applicant is capable of writing a good essay of any
type. All universities invite students to submit any type they’d like. However, each year the
national zeitgeist selects a single type (uniformly at random) to be preferred by all the
universities that year. Each university (independently) accepts an applicant with probability
p = .5 if their type is preferred and 0 otherwise. Applicants don’t know which type is preferred.
1. Compute (give numeric percentages) the probability of being accepted to at least one
school if one submits (i) the same essay to all 12 schools or (ii) a different type to each.
Explain intuitively (in a sentence or two) why one answer is larger for any p ∈ (0, 1].
2. Now suppose each university has a different preferred type (with all 12! permutations
equally likely). So now, instead of all looking for the same thing, every school has a
different preference. As before, each university (independently) accepts an applicant
with probability p if the essay type is preferred and 0 otherwise. Compute the
probability of being accepted to at least one school under (i) and (ii)Pfrom above. [Hint:
(ii) requires inclusion-exclusion.] Simplify by pretending that e−x ≈ 12 k
k=0 (−x) /k! holds
exactly. Is one answer larger for all p ∈ (0, 1]? Explain why in a sentence or two.
REMARK: What do we do in real life when we have many ways of presenting ourselves and
we don’t know for sure what will or won’t impress a reviewer? The stories above are not
accurate models, but they illustrate that different assumptions lead to different conclusions.
Are preferences clustered (common zeitgeist) or diversified (product differentiation)? Are
preferences knowable? Later in the course, as we study expectation, we’ll see that in all four
scenarios above the expected number of schools accepting the applicant is p.

2
C. Alice and Bob are playing a game of tennis and have reached the game state called
“deuce.” From here the players keep playing points until one player’s point-win total exceeds
the other player’s total by 2, at which point the player ahead by 2 is declared winner of the
game. Suppose that Alice wins each point with probability p (independently of all previous
points) and Bob wins each point with probability q = (1 − p) (independently of all previous
points). Find the probability that Alice wins the game, as a function of p. (Hint: consider
what happens over the course of the next two points. Either Alice wins both and the game is
over, or Bob wins both and the game is over, or each player wins a point and the players are
back where they started. Compute the probabilities of these three outcomes and work from
there. What is the conditional probability that Alice wins given that the game ends on the kth
of these outcomes?) Based on your answer, do you agree or disagree with the following
statement? If Alice is k times as likely as Bob to win a point, then Alice is k 2 times as likely
as Bob to win the game if the current score is deuce.
D. The online comic strip [Link] has a “random” button one can click to choose one of the
previous n ≈ 3050 strips. Assume there are exactly 3050 numbered strips and that clicking the
“random” button yields the kth strip, where k is chosen uniformly from {1, 2, . . . , n}. If one
observes m strips this way, what is the probability that one sees at least one strip more than
once? (This is a variant of the birthday problem.) Find the values of m that make this
probability roughly equal to 1/4 and 1/2 and 3/4. (Hint: go to [Link] and enter
something like Prod[ (1-k/3050), {k,0,39}]. Then try different numbers in place of 39.)
E. A “gender reveal” party is held to announce the gender of an expected newborn. 15 cups
are filled in advance with colored beads and covered: if the baby is a boy, 8 are filled with blue
beads and 7 with pink. If the baby is a girl, 7 are filled with blue and 8 with pink. When the
audience arrives the cups are knocked over (revealing bead colors) in a uniformly random order
until the audience has seen 8 cups of the same color (and thereby knows the gender). Let N
be the number of cups turned over before the gender is known. Compute the probability that
N = k for k ∈ {8, 9, 10, . . . , 15}. Is the probability that N = 15 (and one has to wait to the
very end) more or less than 1/2? (This problem was communicated to me by a friend who
found N = 15 in a real life implementation and wanted to know how surprising that was.)
F. This problem concerns arbitrage opportunites in betting markets. I’ll use data from a time
when such an opportunity was available. On Sep. 12, 2019, I looked up Democratic
presidential nomination contracts on two trading platforms: predictit and betfair. According
the table below, one could purchase a predictit YES contract on Elizabeth Warren for 35
cents. This is worth $1 if Warren wins the nomination and $0 otherwise. Similarly, one could
purchase a predictit NO contract on Elizabeth Warren for 66 cents. This is worth $1 if Warren
is not elected and $0 otherwise. (When you purchase a NO contract for 66 cents you are
technically selling a YES contract for 34 cents to somebody else offering to buy it for that
price. The gap between the 34 offer-to-buy and the 35 offer-to-sell is called the bid-ask spread.

3
Because predictit only trades in integers, the spread is typically 1 cent — which is why YES
and NO prices sum to 101 instead of 100.)

predictit YES predictit NO betfair YES betfair NO

Elizabeth Warren 35 66 34.5 66.2
Joe Biden 27 74 21.7 79.2
Bernie Sanders 17 84 13.5 87.2
Andrew Yang 12 89 6.1 94.6
Kamala Harris 11 90 10.9 89.6
Pete Buttigieg 9 92 5.0 95.8
Hillary Clinton 5 96 3.6 96.7
Cory Booker 4 97 2.5 98.3
Beto O’Rourke 3 98 0.8 99.6
Julian Castro 2 99 1.0 99.1
Tulsi Gabbard 3 98 1.0 99.3
Tom Steyer 3 98 0.8 99.9
Amy Klobuchar 2 99 0.7 99.5

Because the betting probabilities are not the same on the two sites, there may be
opportunities for arbitrage, i.e., opportunities to make money without taking risk. (For now,
let us ignore taxes, fees, interest and the risk associated with either of these companies going
bankrupt before the election; of course these things matter in practice.) If I buy Biden NO on
predictit for 74 and Biden YES on betfair for 21.7 then I have only spent 95.7. But I am
guaranteed to win 100. If I purchase all 13 predictit NO contracts then the price comes to
$11.80. But because at least 12 of these candidates are guaranteed to lose, I will be guaranteed
to win at least $12 before fees (and about $11.88 after fees). Note that per predictit rules, I
only have to leave enough money on the site to cover the amount I will owe in the worst case
scenario; this means that I can take my 8 cents of profit now without having to leave any
money on the platform for the duration between now and the election. So: 8 free cents.

1. Using the chart above, find three other arbitrage opportunities, i.e., ways to make a risk
free profit (again ignoring the caveats mentioned above: fees, tax, interest, etc.)

2. Per the efficient market hypothesis, arbitrage opportunities should not exist. Give a
short speculative explanation for why there seem to be arbitrage opportunities here.
(This will not be graded except to check that you thought about it.)

REMARK: Predictit traders are only allowed to spend $850 on each candidate. (This was
part of the deal making predictit legal in the US, unlike betfair.) So one can buy 42500 shares
of Klobuchar YES (2 cents each) but only 858 shares of Klobuchar NO (99 cents each). Some

4
people think these constraints cause unlikely YES contracts to be overpriced. In theory,
arbitrageurs should correct this overpricing by buying lots of NO contracts (thus bidding down
the YES prices); but if each potential arbitrageur can only make a little “free money” this way
(due to spending caps), it may just not be worth the effort.
REMARK: Suppose that you think the “true” value of Warren YES is 34.5 but have a
personal reason for wanting to own a Warren YES share. Then there are two things you can
do. One, you can go straight to the site and buy a share for 35 (basically overpaying by half a
cent). Or two, you can put up an “offer” to buy for 34 and wait until somebody is willing to
take the other side of the bet (by buying NO at 66). If you take the second approach, you
might get a better price. But you will have to wait in line behind everyone else offering to buy
at 34. And there is some risk that you never get to the front of the line (perhaps the price of
Warren YES shares goes up before you manage to buy at 34) — and the scenarios in which
you fail to get to the front of the line might be precisely the scenarios in which you would
most like to own Warren YES. Sometimes on predictit the lines can be long (tens or even
hundreds of thousands of shares) and because trades can only take place at integer values, you
cannot jump to the front of the line by offering 34.1. For example, at one point there were
over 350,000 offers to sell Mark Zuckerberg YES for 1 cent. (Zuckerberg was not running.)
Some people making those offers would be willing to sell for half a cent or a tenth of a cent.
But because of the strict integer rule, they can’t make those offers. They have to wait in line.
G. The following is a popular and rather instructive puzzle. A standard deck of 52 cards (26
red and 26 black) is shuffled so that all orderings are equally likely. We then play the following
game: I place the deck face down and begin turning over the cards from the top of the deck
one at a time so that you can see them. At some point (before I have turned over all 52 cards)
you say “now.” At this point I turn over the next card and if the card is red, you receive one
dollar; otherwise you receive nothing. You would like to design a strategy to maximize the
probability that you will receive the dollar. How should you decide when to say “now”?
Your first observation is that a good time to say “now” is when you know a high fraction of
the cards remaining in the deck are red. On the other hand, if you wait for this fraction to
increase, there is a chance you’ll see more red cards while you wait, so that the fraction
decreases. What’s the right way to balance these concerns, i.e., what is the optimal strategy
for deciding when to say “now”? The next page has a hint but don’t look until you have to.

5
HINT: Imagine a variant of the game in which, after you say “now,” I turn over the bottom
card on the deck. Observe that in this variant, it makes no difference when you say “now”
(since you’re going to see the same bottom card of the deck regardless). Now try to argue that
your probability of winning with a given strategy in the modified game is the same as your
probability of winning with that strategy in the original game. Conclude that in the original
game, it also makes no difference which strategy you choose. Your odds of winning the dollar
are .5 regardless.
REMARK: The fraction of red cards in the deck turns out to be something called a
martingale and the fact that it makes no difference when you bet can be derived from
something called the optional stopping theorem. We’ll have more on this at the end of the
course. But I like this puzzle because it’s something you can appreciate right now. This is a
strategy game that Warren Buffett and a chimpanzee would play equally well. (Unlike the
memory game here [Link] which chimpanzees are
apparently better at than humans.)