Mathematical Methods Spring 2019 Jeremy’s solutions

Solutions to problems from the final:

8.

a/ You and a friend are playing with a standard deck of cards. You deal three cards to yourself,
and then three cards to your friend.
1/ What is the probability that you have exactly two aces between you?
Because we are considering all six cards together, this question is of the form “what is
the probability of k successes in n trials”, with the additional constraint that we are
sampling without replacement. We assume that there are 4 aces and N − 4 non-aces
in a deck of N cards. (I’m keeping the number of cards general because the math is
the same regardless of N , but different people might assume different values for N ).
Now, the probability of getting six cards with two aces amongst them is
4 · 3 · N · (N − 4) · (N − 5) · (N − 6) 4! (N − 4)! (N − 6)!
= .
N · (N − 1) · (N − 2) · (N − 3) · (N − 4) · (N − 5) 2! (N − 8)! N !
Then, we must consider the number of different ways we could arrange two aces and
four non-aces. This is 6!/(2!4!), and so the final probability is
   
4 N −4 N
.
2 4 6
This is easily understood as “the ways to choose 2 things from 4, times the ways to
choose 4 things from N − 4, divided by the ways to choose 6 things from N ”. Some
common values for N would be 40 (in Spain), 48 (in Portugal), or 52 (in France and
the U.S.A.). For those values, the probabilities would be ≈ 0.092, ≈ 0.066, or ≈ 0.057,
respectively.
2/ Your friend complains that how you are dealing the cards is unfair. For example, it’s
more likely that your first card will be an ace than it is that your friend’s first card will
be an ace. Calculate the ratio of these probabilities to see if your friend’s objection is
a good one.

The probability that your first card is an ace is 4/N . The probability that your friend’s
first card is an ace is found as follows. We must account for every way that you could
be dealt three cards followed by your friend being dealt an ace. But because between
zero or three of your cards could be aces, the probability of your friend being dealt an
ace depends on what you have been dealt.
We first note that, from the previous part, we have a general expression for “the
probability of drawing k1 things from n1 and k2 things from n2 ”,
   
n1 n2 n1 + n2
.
k1 k2 k1 + k2
Let n1 = 4, n2 = N − 4. Let k1 be the number of aces you are dealt, k2 be the number
of non-aces, and let k1 + k2 = 3. Then, the probability that your friend is dealt an
ace is
" 3 #
X 4 − i 4N − 4 N 
,
i=0
N −3 i 3−i 3

where the factor of (4 − i)/(N − 3) is the probability that your friend’s first card is
an ace if you’ve been dealt i aces in 3 cards. Carrying out this sum reveals that the

1
Mathematical Methods Spring 2019 Jeremy’s solutions

probability of your friend’s first card being an ace is 4/N . So, the ratio of probabilities
is 1—exactly the same! Despite your friend’s concern, it seems that how you’re dealing
the cards is fair.
b/ On a usual night, a pizza restaurant receives an order once every six minutes on average.
1/ Last night, the restaurant received two orders within three minutes of each other. What
is the chance of this happening?
We assume that orders to the restaurant are Poisson-distributed, with λ = 1/6 in
units of orders/minute. Because we are concerned with the time interval between
two calls, we should use the exponential distribution. (Equivalently, we can use the
Poisson distribution, with k = 1, which also finds the probability of the interval
between two events.) The probability of there being 3 minutes between orders is
3 −3/6
Pr(T = 3) = e ≈ 0.303 .
6
2/ Later that same night, the cashier decided to take a ten-minute break right after taking
an order. What is the chance that the restaurant will receive no orders during this
break?
The chance that the restaurant will receive no orders during this break is the proba-
bility that the next order will come in ten minutes or more. That is, Pr(T ≥ 10) =
1 − Pr(T < 10). So, we must use the cumulative distribution function for the expo-
nential distribution, F (t) = 1 − e−λt , and so find

Pr(T ≥ 10) = e−10/6 ≈ 0.189 .

Note that this is identical to using the Poisson distribution to find the probability of
zero events in 10 minutes.
c/ In order to measure the average height of the people of a certain country, we take the
following sample (in centimeters): {171, 174, 189, 164, 171, 180, 174, 177}. Estimate the
average of the population, with its standard error. In order to improve the estimate, we
will take a larger sample. What is the minimum size of the new sample in order to state
that the average height of the people of this country is lower than 176 cm with a 90%
confidence? (Assume that the average and the variance of the new sample are the same
as the ones computed before, and that the sampling distribution is Gaussian.) In both
cases, we will use the sample statistic to estimate the population statistic. To estimate
the average (mean) of the population, we must first find the average of the sample. To do
this, we sum the eight measurements and divide by eight to find x = 175 cm. To find the
standard error, which is the standard deviation, we first calculate the sample variance:
8
1X
n2 = (xi − x)2 = 47.5 .
8 i=1
√
Then, the standard deviation is just the square root of the variance: s = n2 = 6.89.
Note that if we had applied Bessel’s correction, which constructs a better estimator
p of the
population standard deviation from the sample standard deviation via σ̂ = pN/(N − 1)s,
where N is the number of data items, we would have instead found σ̂ = 8/7s = 7.37.
Either answer was accepted on the exam. So, we have found the estimate of the population
average as
µ̂ = 175 ± 6.89 cm .

Now, we want to take a larger sample, such that we can be 90% confident that the average
height is less than 176 cm. Because we are assuming that the sampling distribution is

2
Mathematical Methods Spring 2019 Jeremy’s solutions

Gaussian, we can invoke the Central Limit Theorem to say √ that the standard deviation of
the new sample, which has N data items, will be s0 = s/ N , where s = 6.89 from above.
Because we just want to know that the average is less than 176 cm, we will take a one-sided
interval, i.e., [0, µ+ ] only. This means that α = 0.1, and α and µ+ are related by

µ+ = µ̂obs + σµ̂ Φ−1 (1 − α) .

√
Here, note that µ̂obs = 175, σµ̂ = s/ N , and Φ−1 indicates the inverse CDF of the
standard Gaussian distribution. Since we want to have µ+ = 176, we can plug in our
known values and solve for N :
 −1 2
sΦ (1 − α)
N= ≈ 78 ,
µ+ − µ̂obs

where we rounded up to the nearest whole number. So, we need at least 78 people in the
new sample to be 90% sure that the population mean is less than 176 cm.
d/ We wish to estimate the inverse rate parameter, β = 1/λ, of a population which follows
the exponential distribution f (x) = λe−λx , for x ∈ (0, ∞). (This is common, for instance,
when measuring lifetimes of radioactive particles.) Given a sample {x1 , x2 , . . . , xk }, we
choose x1 as our estimator.
1/ Find the bias and the variance of this estimator.
We have β̂ = x1 . Because x1 is an element of our sample, it is a random variable
which follows the population distribution. Thus,
Z ∞
x −x/β
E[β̂] = E[x1 ] = e dx = β .
0 β

Thus this estimator is unbiased. This same reasoning applies to finding the variance
of the estimator, and so
Z ∞
(x − β)2 −x/β
V [β̂] = V [x1 ] = e dx = β 2 .
0 β

2/ Instead of estimating β, we wish to estimate λ. Since x1 worked out so well for β, we

try 1/x1 as the estimator for λ. Is this estimator unbiased?
Now trying to apply the same approach as before, we calculate the expectation value
of λ̂ = 1/x1 as
Z ∞
λ −λx
E[λ̂] = E[1/x1 ] = e dx .
0 x

But here we have a problem—this integral does not converge on the range given. So
while we can’t calculate the exact value of λ̂’s bias, we can be quite sure that it isn’t
unbiased!