1

The Two Envelopes Problem

Rich Turner and Tom Quilter
The Two Envelopes Problem, like its better known cousin,
the Monty Hall problem, is seemingly paradoxical if you are not
careful with your analysis. In this note we present analyses of
both the careless and careful kind, providing pointers to common
pitfalls for authors on this topic.
As with the Monty Hall problem, the key is to condition on
all the facts you have when you crank the handle of Bayesian
inference. Sometimes facts unexpectedly have to provide you
with information: and you have to take them into account.
This point is well known, and most explanations of the para-
dox point this out. However authors then tend to make a num-
ber of mistakes. Firstly they claim the conventional analysis
corresponds to using a uniform-improper distribution. This is
wrong: it corresponds to a log-uniform improper distribution.
Secondly they claim the root of the paradox is in the use of an
un-normalisable distribution. This is merely a conjecture and a
toy example provided here suggests that improper distributions
exist which do not result in the paradox.
The Two Envelopes Problems ultimate lesson is that infer-
ence always involves making assumptions and any attempt
to blindly use uniformative priors as a method for avoiding
making assumptions can run into trouble. After all the innocu-
ous uniform (or log-uniform) prior corresponds to a very strong
assumption about your data.
0.1 The problem
You are playing a game for money. There are two envelopes on a table.
You know that one contains $X and the other $2X, [but you do not
know which envelope is which or what the number X is]. Initially you
are allowed to pick one of the envelopes, to open it, and see that it
contains $Y . You then have a choice: walk away with the $Y or return
the envelope to the table and walk away with whatever is in the other
envelope. What should you do?
0.2 The Paradox
Clearly, to decide whether we should switch or stick we want to com-
pute the expected return from switching (R) and compare it to
2
the return from sticking (trivially, $Y ). Initially, the chances of choos-
ing the envelope with the highest contents P(C=H) are equal to the
chances of choosing the envelope containing the smaller sum P(C=L),
and they are both
1
2
. A careless analysis analysis might then reason
that the expected return from switching is therefore:
R =
1
2
Y P(C=H) + 2Y P(C=L) (1)
=
1
2
Y
1
2
+ 2Y
1
2
(2)
=
5
4
Y (3)
This would mean we expect to gain
1
4
Y from switching on average.
This is paradoxical as it says It doesnt matter which envelope you
choose initially - you should always switch.
0.3 A thought experiment
So goes the usual explanation of the paradox. Wheres the aw in the
analysis?
In particular, a key question we have to resolve is does observing
the contents of the rst envelope provides us with useful in-
formation. If it does not, then it will not matter whether we stick or
switch [just as it didnt matter which envelope we chose to start with].
First Im going to pursuade you that observing the contents does
provide you with useful information. Then Im going to show you how
to use Bayes theorem to correctly use this information to compute the
optimal decision.
Imagine you make your rst choice, and open up the envelope only
to discover that it contains a bill - the envelope is charging you money
- Y is negative. Not only is this annoying, it is also very unexpected.
Who said we could lose money playing this game? If we use the above
analysis for the new situation, we should now stick and not switch. So
the paradox - that we should switch regardless of Y - (partly) melts
away if the assumption that Y is always positive is broken. Put another
way: if Y can be negative or positive then wed have to observe Y before
we decided whether switch. The moral of the story is that we should
pay closer attention to our assumptions and, in particular, to what Y
is telling us.
3
0.4 Condition on all the infomation
So we made a mistake: When computing the expected reward above
we should have taken into account [technically, condition on] all the
information we have available to us at the time. Equation 1 should
therefore have read:
R =
1
2
Y P(C=H|Y ) + 2Y P(C=L|Y ) (4)
We didnt do this because we naively we thought the amount of
money that the envelope contained didnt provide any useful informa-
tion (ie. P(C=H|Y ) = P(C=H), conditional independence). However
the thought experiment above shows us that this is not necessarily true,
and moreover whether it is or not depends on our assumptions. The
cure is that we should work through the analysis stating the assump-
tions explicitly. The way to incorporate these assumptions explcitly is
to use Bayes theorem:
P(C=i|Y ) =
P(Y |C=i)P(C=i)
P(Y )
(5)
P(C=i|Y ) =
P(Y |C=i)P(C=i)
P(Y |C=H)P(C=H) +P(Y |C=L)P(C=L)
(6)
We have already noted that the probability that P(C=H) = P(C=L) =
1
2
and so the factors of
1
2
in the denominator and numerator all cancel
and we can plug the result into the formula for the expected reward:
R =
1
2
Y P(Y |C=H) + 2Y P(Y |C=L)
P(Y |C=H) +P(Y |C=L)
(7)
=
1
2
Y
1 + 4(X)
1 +(X)
(8)
The expected reward therefore depends on =
P(Y |C=L)
P(Y |C=H)
and we are
now forced to state our beliefs about the two distributions:P(Y |C=L)
and P(Y |C=H) to calculate the expected reward.
0.5 Uniform beliefs
At rst sight a sensible distribution - in the sense that it is non-
commital - is a uniform prior between two limits, Y
min
and Y
max
say:
4
p(Y |C=L) =
1
Y
max
Y
min
(9)
=
1
Z
(10)
p(Y |C=H) =
1
2Y
max
2Y
min
(11)
=
1
2Z
(12)
For the sake of clarity lets think about the situation where Y is
always greater than zero (Y
min
= 0). What are the possible values for
?
Well, if 0 < Y < Y
min
then =
p(Y |C=L)dY
p(Y |C=H)dY
= 1/2, so: R = 3Y/2
and we should always switch. This makes sense - it is twice as likely
that the envelope contains the lower amount in this case, so switching
should yield
2
3
2 +
1
3

1
2
, which it does.
However, if Y
min
< Y < Y
max
then = 0 and R =
1
2
- we must
have chosen the envelope containing the larger amount, the expected
reward from switching is therefore $
1
2
and so we should stick.
There is no paradox.
0.6 Improper uniform beliefs
We might now say, well, Im very unsure about what value Y would
take and therefore Ill make Y
max
extremely large, and take the limit
Y
max

1
. This turns out to be a terrible assumption for any
practical situation. As we take this limit p(Y |C=L) and p(Y |C=H)
become very ill matched to any possible real data in such a way
that this causes problems. All data we observe will lie beneath Y
max
(as its tending to innity). However, in this regime we believe it
is twice as likely that the envelope contains the lower amount, and
so we switch whatever we see. Put another way, in order to stick we
would have to observe a Y which is larger than Y
max
, but well never
encounter an envelope containing more than innity dollars
2
so well
always switch.
Although we tried to make the most non-commital assumptions
possible - we ended up making very denite, absurd assumptions which
1
In this limit the uniform distribution is un-normalisable and such distributions are
called improper
2
although there are numbers greater than innity, its tough to write legal cheques for
these amounts
5
makes sensible inference impossible.
3
Before we move on, its worth mentioning that it might have been
tempting to set equal to one in the above case. After all, in the limit
both p(Y |C=L) and p(Y |C=H) are (improper) uniform distributions
between zero and innity. We would then have = 1 and recover the
old result R = 5Y/4. However, setting the two densities equal is not
consistent with the assumption that one envelope contains an amount
twice that in the other. This assumption demands one density be half
of the other (and to have twice the range). So the original analysis
does not correspond to assuming a uniform improper distribution (as
some authors claim).
To conclude, its impossible to use this particular improper distribu-
tions without being careful (or else we introduce a contradiction with
our other assumptions) and even if we are careful they correspond to
absurd assumptions (where we switch unless Y is greater than innity).
0.7 The log-uniform beliefs
If the = 1 case does not correspond to pair of uniform distributions
p(Y |C=H) and p(Y |C=L), does it correspond to another choice? The
answer, somewhat suprisingly, is yes. To see this we need to do some
more work to simplify the form of . We can do this by realising
that specifying p(Y |C=L) immediately species p(Y |C=H) (due to the
doubling constraint). We can derive a relation which relates the two via
manipulation which is easy to skrew up (as, for example, many authors
on this paradox do - even in publications). Heres one way to make sure
you get the right result: We want to sum up all the probability at Y

in the old distribution which gets mapped to Y = 2Y

in the new
distribution. So we write down:
p(Y |C=H) =

0
p(Y

|C=L)(2Y

Y )dY

(13)
=

p(U/2|C=L)(U Y )dU/2 (14)

= p(Y/2|C=L)/2 (15)
This makes intuitive sense: A chunk of probability in the old dis-
tribution p(Y

|C=L)dY

is mapped to a region centred at Y = 2Y

but it is smeared out to extend over a length 2dY

so the density at
p(Y |C=L) must be half that of p(Y/2|C=L).
3
Note, however, that as we take the limit the paradox (that we should always switch
regardless of Y ) does not rear its ugly head.
6
If equals one then this means: p(Y |C=L) = p(Y |C=H) = p(Y/2|C =
L)/2 for all Y , which holds when p(Y |C=L) = 1/Y . This is a uniform
distribution on log Y and says, I have know idea about the scale of
Y . I think Y is as likely to be between 1 and 10 as it is to be between
10 and 100. This seems more sensible than the uniform prior before -
for one thing, it decreases with Y .
However, this distribution is not normalisable (without introducing
a lower and upper cut-o and then is not 1 for all Y ). Limiting
arguments fail for the same reason they did in the uniform case: in
the region where both the distributions are non-zero, you should al-
ways switch. So taking Y
max
to innity pushes the regime where you
should stick to a place where no data will land. Again, the limit corre-
sponding to the improper distribution corresponds to a terrible set of
assumptions.
0.8 A sensible approach
Is there a sensible approach? In a word yes. However, it will depend
on what assumptions you want to make. In a way the previous dis-
cussion is a red-herring: you might be quite happy to use a uniform
distribution, or a log-uniform distribution - and everything will be ok
so long as you think carefully about the choice for the ranges (am I
playing with a friend for his poket money, or am I through to the -
nal stages of a game-show called Who wants to be a billionaire?).
4
One nal alternative (which does not have hard cut-os) is to place an
exponential distribution over Y :
P(Y |C=L) = exp(Y ) (16)
1/ is the characteristic length scale over which the density decays.
From eqn. 8 we should switch when:
1
2
(Y ) (17)
= 2 exp(Y/2) (18)
(An intuition for this equation is that it is the result of a hypothesis
test: we shuld switch when it is more likely weve chosen the envelope
containing the smaller amount. That is when P(C = L|Y ) > P(C =
H|Y ). Plugging in values well get the same result as above.)
4
One alternative is to place a prior over the ranges and integrate them out:
P(Y |C=L) =
R
P(Y |C=L, Y
min
, Y
max
)P(Y
min
, Y
max
)dY
min
dY
max
7
So we switch when Y
2

ln 4
3

, which is proportional to the

characteristic length scale, as expected.
0.9 I want to specify a marginal, not a conditional!
We have discussed how specifying the conditional distrution P(Y |C =
L) immediately species the other conditional P(Y |C = H) =
1
2
P(Y/2|C =
L). Together these two distributions combine to form the marginal dis-
tribution, or more intuitively, the distribution over values of money we
expect to see in envelopes: P(Y ) = P(Y |C = L)P(C = L) +P(Y |C =
H)P(C = H). This is uniquely determined once we specify P(Y |C =
L) as P(Y ) =
1
2
P(Y |C = L) +
1
4
P(Y/2|C = L). This opens up the
possibility of the reverse: specifying a marginal and then deriving the
two conditionals.
The above equation can be solved by recursion, yielding: P(Y |C =
L) = 2

n=0
(2)
n
P(2
n
Y ) under the proviso that 2
N
P(2
N
Y |C =
L) 0 as N . This equation says: You give me a marginal, Ill
give you the conditional.
0.10 A sensible improper distribution
Does using an improper distribution always lead to the paradox? Some
authors claim so, but it hasnt been proved. More over the intuition
from the above examples is that it wasnt the improper nature of the
distribution which lead to the paradox (as some authors glibly state) -
it was the fact that they corresponded to stupid assumptions.
Lets try and invent a distribution which correspondeds to reason-
able assumptions in the limit where they become un-normalisable.
5
The key intuition is this: well be ne as long as the point we
always switch is not situated in the extreme of the range of allowed Y
values. For instance, if the conditional distributions have some periodic
struture of correctly chosen amplitude then there can be regions when
one distribution is much higher than the other, and regions when the
converse is true. In this case the paradox will not arise. Whatsmore, the
marginal distribution P(Y ) implied by the two periodic conditionals
may not be such a stupid one
6
. Let me give a simple example:
5
Another point noted by authors is that the improper distribution above has an innite
mean value - this may correspond to another undesireable assumption too. For instance
there are proper distributions which have innite expected value and in such cases the
paradox can occur too.
6
Although a counter example might be enough to convince the mathematicians the
paradox doesnt arise from un-normalisabilty, it would be nice if such a counter example
corresponded to a realistic
8
P(Y |C = L) = 4 2n + 1 > x > 2n (19)
P(Y |C = L) = 1 otherwise (20)
Over some range L that we extend to innity (and therefore 0)
to make our improper prior.
This species the other conditional:
P(Y |C = H) = 2 4n + 2 > x > 4n (21)
P(Y |C = H) = 1/2 otherwise (22)
And the marginal:
P(Y ) = 3 4n + 1 > x > 4n (23)
P(Y ) = 3/2 4n + 2 > x > 4n + 1 (24)
P(Y ) = 9/4 4n + 3 > x > 4n + 2 (25)
P(Y ) = 3/4 4n + 4 > x > 4n + 3 (26)
Is this a sensible marginal? Well, imagine that some types of
coins were harder to come by than others. For example, imagine only
1,5,9,13... valued coins existed. Then we might like something like this
(perhaps wed like to add in a 1/Y decay too).
When should we switch? This depends on Y in the following way:
Region P(Y |C = L) P(Y |C = H) switch?
I 4n + 1 > x > 4n 4 2 doesnt matter
II 4n + 2 > x > 4n + 1 1 2 no
III 4n + 3 > x > 4n + 2 4 1/2 yes
IV 4n + 4 > x > 4n + 3 1 1/2 doesnt matter
So we do need to listen to what the data are telling us even though the
condtionals are improper.