5
$\begingroup$

The following problem is exercise 1.4.5(a) in One Thousand Exercises in Probability:

5. The Monty Hall problem: goats and cars. (a) Cruel fate has made you a contestant in a game show; you have to choose one of three doors. One conceals a new car, [sic] two conceal old goats. You choose, but your choosen door is not opened immediately. Instead, the presenter opens another door to reveal a goat, and he offers you the opportunity to change your choice to the third door (unopened and so far unchosen). Let $p$ be the (conditional) probability that the third door conceals the car. The value of $p$ depends on the presenter's protocol. Devise protocols to yield the values $p = \frac{1}{2}$, $p =\frac{2}{3}$. Show that, for $\alpha \in [\frac{1}{2}, \frac{2}{3}]$, there exists a protocol such that $p = \alpha$. Are you well advised to change your choice to the third door?

This is the solution given later in the book:

5. (a) One cannot compute probabilities without knowing the rules governing the conditional probabilities. If the first door chosen conceals a goat, then the presenter has no choice in the door to be opened, since exactly one of the remaining doors conceals a goat. If the first door conceals the car, then a choice is necessary, and this is governed by the protocol of the presenter. Consider two 'extremal' protocols for this latter situation.

(i) The presenter opens a door chosen at random from the two available.

(ii) There is some ordering of the doors (left to right, perhaps) and the presenter opens the earlier door in this ordering which conceals a goat.

Analysis of the two situations yields $p = \frac{2}{3}$ under (i), and $p = \frac{1}{2}$ under (ii).

Let $\alpha \in [\frac{1}{2}, \frac{2}{3}]$, and suppose the presenter possesses a coin which falls with heads upwards with probability $\beta = 6\alpha - 3$. He flips the coin before the show, and adopts strategy (i) if and only if the coin shows heads. The probability in question is now $\frac{2}{3} \beta + \frac{1}{2}(1 - \beta) = \alpha$.

You never lose by swapping, but whether you gain depends on the presenter's protocol.

This solution seems completely wrong to me for multiple reasons:

  1. The deterministic protocol adds information, and the probability shouldn't be reduced with more information.

  2. By switching, the contestant is guaranteed the car as long as he does not select the car first, and this outcome is independent of the chosen protocol, since the protocol only applies when the contestant selects the car first.

  3. This question appears to have already been answered on StackExchange, and the answer there is not in agreement with this given solution.

As far as I can tell, the authors are confusing "the conditional probability that the third door conceals the car" with the conditional probability of winning by switching when the location of the car is unknown.

To see what I mean, suppose we label the doors 1, 2, and 3, and we let $C^{(d)}$, $G^{(d)}$, and $S^{(d)}$ be the events that the $d$th door contains a car, contains a goat, or is first selected by the contestant, respectively. Let $d_1$, $d_2$, and $d_3$ be the doors first selected by the contestant, revealed by the host, and then leftover, respectively.

With this notation, the conditional probability $p$ is

$P(C^{(d_3)} | G^{(d_2)}, S^{(d_1)}) = \dfrac{P(C^{(d_3)}, G^{(d_2)} | S^{(d_1)})}{P(G^{(d_2)} | S^{(d_1)})} = \dfrac{\frac{2}{3}}{1} = \dfrac{2}{3}$

However, we can construct a specific situation where the conditional probability is $\frac{1}{2}$: Suppose the host always reveals the smallest numbered door containing a goat. If we take $d_1 = 3$ and $d_2 = 1$, then we have

$P(C^{(2)} | G^{(1)}, S^{(3)}) = 1 - P(G^{(2)} | G^{(1)}, S^{(3)}) = 1 - \dfrac{P(G^{(2)}, G^{(1)} | S^{(3)})}{P(G^{(1)} | S^{(3)})} = 1 - \dfrac{\frac{1}{3}}{\frac{2}{3}} = \dfrac{1}{2}$

But this is a specific situation that relies on more than just the protocol of the host, since there is a $\frac{1}{3}$ chance that the car could be behind the first door, so that, by picking the third door and having the second door revealed, we would know with certainty that the car is behind door one before we switch. Thus, the conditional probability of $\frac{1}{2}$ is only attained when the location of the car is unknown after the reveal. (And our sanity check of $\frac{1}{3} \cdot 1 + \frac{2}{3} \cdot \frac{1}{2} = \frac{2}{3}$ agrees with our conditional probability $p$ that the third door contains a car, not knowing which specific doors are opened).

That the authors made a mistake is further cemented in the last sentence of their solution: "whether you gain depends on the presenter's protocol." In fact, regardless of the presenter's protocol, there is always an event of non-zero probability where you benefit by swapping. It should read, "whether you have the possibility of not gaining by switching after conditioning depends on the presenter's protocol."

Their solution seems wrong to me, but the authors have fancy credentials (Ph.D.s from Oxford) and positions at Oxford and Cambridge, so I'm hesitant to write it off as an error. It is possible I could have misread the problem or completely miscalculated the conditional probabilities. The details do, after all, seem rather nuanced.

Is this problem/solution an error? Am I just nitpicking wording?

  • 1
    Actually, I think the error in the answer is the claim that $\frac12$ is an extremal point. Indeed, there's a protocol where your chance of winning the car after changing is *zero.* That protocol is: If the show master has the option to open the door with a car, he does so. In that case, if he opens an empty door, the only possible reason that he does so is that you have chosen the door with the car (so he couldn't open that without violating the rules), and therefore if you change the doors, you are guaranteed not to get the car.2017-02-20
  • 0
    I think you are misreading their "protocol ii". I believe they mean to say that the presenter invariably opens the leftmost unchosen door regardless of what it contains. that is to say, if you choose door $1$, he opens door $2$. If you choose $2$ or $3$, he opens $1$. In this case, of course, it turns out to hide a goat but that is an accident.2017-02-20
  • 0
    No, the presenter always reveals a goat after the first selection. The protocol is how the presenter decides which goat to reveal in the case that there are two choices (when the contestant has initially picked the car). This is clearly stated in both the question and solution. "If the first door chosen conceals a goat, then the presenter has no choice in the door to be opened, since exactly one of the remaining doors conceals a goat. If the first door conceals the car, then a choice is necessary, and this is governed by the protocol of the presenter."2017-02-20
  • 0
    In that case I have trouble seeing how they get to $\frac 12$. I get $\frac 12$ when the presenter has no idea where the car is and opens randomly (risking the accidental reveal of the car). If the presenter does know where the car is and carefully avoids revealing it, then it seems to me that the symmetry should be broken.2017-02-20
  • 0
    I believe the authors are referring not to the probability of winning the car but rather to the probability that the last door holds the car after conditioning on the specific doors opened, which (given a fixed protocol and by selecting the right doors) can be $\frac{1}{2}$, as computed in my question.2017-02-20
  • 0
    @user144527: You're right, I misread the protocol. With the correct reading, while I didn't calculate, it seems at least plausible: With the ordered door protocol, if the show master had opened the last door, the probability that the car is in the second door is $1$ (as otherwise the show master would have opened that instead; note that this is exactly the same type of argumentation as with my misunderstood problem). Therefore in return, the probability for the car to be behind the third door when the second door is opened obviously might be reduced.2017-02-20
  • 0
    @user144527: The probability that the last door holds the car *is* the probability that you will win if you change to that door. [That is, assuming that "winning" means "getting the car", of course.](https://xkcd.com/1282/) :-)2017-02-20
  • 0
    Anyway, in the *problem statement* it wasn't said what exactly counts as "strategy".2017-02-20
  • 0
    Yes, the probability that the last door holds the car is the probability that you will win if you switch; that is not contested. The difference is whether or not you condition on specific doors and a fixed protocol.2017-02-20
  • 0
    From the exercise text (emphasis by me): "Let $p$ be the **(conditional)** probability that **the third** door conceals the car. The value of $p$ **depends on** the presenter's protocol." I think that leaves no room for interpretation whether or not you should condition on specific doors and protocols.2017-02-20
  • 0
    The doors are not numbered in the problem. The "third" door means the third door selected, i.e., not the first door (the one the contestant selected first) and not the second door (the one revealed by the presenter). Even if they were labeled, this would mean nothing because you could still alter the presenter's protocol or the first selection to get any result wanted.2017-02-21

2 Answers 2

3

As far as I can tell, the authors are confusing "the conditional probability that the third door conceals the car" with the conditional probability of winning by switching when the location of the car is unknown.

They are the same thing.   If the remaining door conceals a car, then you will win by switching.

To see what I mean, suppose we label the doors 1, 2, and 3, and we let C(d), G (d) be the events that the dth door contains a car, contains a goat, or is first selected by the contestant, respectively.

Simplify it.   You only need to label two doors.   The events we are interesting in are: What the contestants first choice hides, which of the other two doors the host reveals, what the remaining door hides.

(Notice, though, what the contestant's first choice hides determines what the remaining door hides; the car can only be behind one of them.)

Without loss of generality, the contestant chooses whichever door and it hides either (1/3) a car or (2/3) a goat.   There are two doors remaining; one on the left of the other; now label those doors $1$ and $2$.   After the host opens one of these such that it will reveal a goat, then evaluate the probability the other hides a car.

  • Case 1: The contestant's choice hides a goat.   The host reveals whichever door also hides a goat.   Protocol is irrelevant; the host has no choice if a goat is to be revealed.   There is $50\%$ chance this will be either door and the other will certainly hide a car.

$$\begin{align} \mathsf P_\star(F{=}g, H{=}1, R{=}c) &= \tfrac 23\tfrac 12=\tfrac 13 \\ \mathsf P_\star(F{=}g, H{=}2, R{=}c) &= \tfrac 23\tfrac 12 = \tfrac 13 \\ \mathsf P_\star(F{=}g, H{=}1, R{=}g) &= 0\\ \mathsf P_\star(F{=}g, H{=}2, R{=}g) &= 0 \end{align}$$

  • Case 2: The contestant's choice hides a car.   The host may therefore reveal either of the two doors, as they both hide a goat.   Which door he then reveals depends on the protocol.

    • Protocol I: Random Protocol: The host will then open either door with equal probability. $$\begin{align}\mathsf P_1(F{=}c, H{=}1, R{=}g) &= \tfrac 13\tfrac 12=\tfrac 16 \\ \mathsf P_1(F{=}c, H{=}2, R{=}g) &= \tfrac 13\tfrac 12 = \tfrac 16 \\ \mathsf P_1(F{=}c, H{=}1, R{=}c) &= 0\\ \mathsf P_1(F{=}c, H{=}2, R{=}c) &= 0 \end{align}$$

    • Protocol II: Leftmost Protcol: The host will open door 1. $$\begin{align}\mathsf P_2(F{=}c, H{=}1, R{=}g) &= \tfrac 13\\ \mathsf P_2(F{=}c, H{=}2, R{=}g) &= 0 \\ \mathsf P_2(F{=}c, H{=}1, R{=}c) &= 0\\ \mathsf P_2(F{=}c, H{=}2, R{=}c) &= 0 \end{align}$$

Now evaluate $\mathsf P_1(R{=}c\mid H{=}1)$, $\mathsf P_1(R{=}c\mid H{=}2)$ , $\mathsf P_2(R{=}c\mid H{=}1)$ , $\mathsf P_2(R{=}c\mid H{=}2)$ .

$$\begin{align}\mathsf P_1(R{=}c\mid H{=}1) & =\dfrac{\mathsf P_1(R{=}c\cap H{=}1)}{\mathsf P_1(R{=}c\cap H{=}1)+\mathsf P_1(R{=}g\cap H{=}1)} \\ & = \dfrac{\tfrac 13}{\tfrac 13+\tfrac 16} \\ & =\dfrac 23 \\[2ex] \mathsf P_1(R{=}c\mid H{=}2) & = \dfrac{2}{3} \\[2ex] \mathsf P_2(R{=}c\mid H{=}1) & = \dfrac{\tfrac 13}{\tfrac 13+\tfrac 13} \\ & = \dfrac 12 \\[2ex] \mathsf P_2(R{=}c\mid H{=}2) & = \dfrac{\tfrac 13}{\tfrac 13+0} \\ &= 1 \end{align}$$

Conclusion:

If the host uses Protocol I (never reveal a car and choose randomly when the contestant chooses the car), then the car will be behind the remaining door with probability $2/3$ regardless of which door the host opens.

If the host is known to use Protocol II (opening the leftmost door unless there is a car behind it) then the car will be behind the remaining door $1/2$ of the times when the host opens the leftmost door, and always when the host opens the rightmost door.


This solutions seems completely wrong to me for multiple reasons:

1.The deterministic protocol adds information, and the probability shouldn't be reduced with more information.

The deterministic probability adds information and depending on what information is added (open leftmost or open rightmost), the conditional probability that the car is behind the remaining door either decreases to $1/2$ or increases to $1$.

2.By switching, the contestant is guaranteed the car as long as he does not select the car first, and this outcome is independent of the chosen protocol, since the protocol only applies when the contestant selects the car first.

Yes.   However the contestant does not know whether he selected the car first, except that protocol II may add some clues.   Because a protocol II host never chooses the rightmost door unless there is a car behind the leftmost door.

So if a protocol II host selects the rightmost door you surely know where the car hides.

Though if a protocol II host chooses the leftmost door it is even odds whether switching will win a car or lose it.

Of course, knowing which protocol the host uses is the trick; that's the business about the host selecting protocol before the show using a biased coin.

However, the situations where a protocol II host chooses the rightmost door happen $1/3$ of the time while those where the host chooses the leftmost door happen $2/3$ of the time, so before playing the game you have a total probability of $\tfrac 13\cdotp1+\tfrac 23\tfrac 12=\tfrac 23$ for winning if you switch.

3.This question appears to have already been answered on StackExchange, and the answer there is not in agreement with this given solution.

They are in fact in agreement.   Protocol doesn't matter until the host chooses a door.

  • 0
    This doesn't answer the question.2017-02-21
  • 0
    This answers the question.2017-02-21
  • 0
    It does not answer the question; it simply restates what was posted in the question.2017-02-21
  • 1
    The question: Did the authors make a mistake, since there is no protocol that ensures a conditional probability on the third door of $p=\frac{1}{2}$ but rather the protocol in the solution gives $\mathbb{P}(p=\frac{1}{2})=\frac{2}{3}$ and $\mathbb{P}(p=1)=\frac{1}{3}$. You have restated the "but rather" part.2017-02-21
  • 0
    Also, the two probabilities you point out at the start of the question are not the same. The difference is "when the location of the car is unknown."2017-02-21
0

This does not address your particular beef with the book's analysis, but I would like to point out that in this particular description, it is not even clear whether the presenter knows where the prize is, or whether he will always open a door (maybe some shows he does not; indeed, maybe he only opens a door whenever you did initially pick the door with the prize ... And if you know that, then you have a $0$ chance of winning by switching). So the analysis from the book is already assuming all kinds of things that are not at all clear from the book's description of the problem.

  • 0
    "Instead, the presenter opens another door to reveal a goat." There is always a reveal, and it is always a goat.2017-02-20
  • 0
    The way I read it: I am on this particular show, and this particular one time the presenter behaves a certain way.2017-02-20
  • 0
    That is irrelevant. "Always" applies to the probability space.2017-02-20
  • 0
    I agree ... If only they had used the word "always" in their description. Too bad they didn't.2017-02-20
  • 0
    It doesn't matter. They stated that it is true independent of the choices of doors. This means that it is true in every event, which means it has probability one, which means "always."2017-02-20
  • 0
    Well, I guess we just read the description differently. I see a description of a one-time event, and you see it differently. But knowing all the variants on the Monty Hall, and how important it is to phrase it exactly right, surely the authors could have been a lot more careful in their phrasing so that my reading would be ruled out. As it is stated, I don't think it can be ruled out.2017-02-20
  • 0
    No, it's not reading it differently. Whether it's a "one-time event" or not does not affect the probability space.2017-02-20
  • 0
    If you didn't get to choose which door you opened, it wouldn't be your choice. The question is not telling you that you chose a door without allowing you to actually choose the door. This is unambiguous because of the tense of the verb ("choose" instead of "chose"). I will concede, however, that "read the description differently" could mean "misread the description."2017-02-20
  • 0
    Also, any lingering doubt should be immediately resolved by the solution, so it absolutely can be ruled out.2017-02-20
  • 0
    Just to be clear: I have no trouble coming up with scenarios (or protocols) where the $p$ ranges from $\frac{1}{2}$ to $\frac{2}{3}$, so I can answer that question no problem. But when at the end they asked whether I should stay or switch, I really have no idea how to answer that, because the description of what I am faced with as a contestant is unclear ... and that lingering doubt about the *question* is of course not going to be resolved by the authors' *solution*. I mean, if I were to get that question on a test, I can't really look at the solution to figure out what is meant.2017-02-21
  • 0
    You should have problems coming up with protocols where $p$ is anything other than $\frac{2}{3}$ because it's not possible the way $p$ is defined. But even if you could, that you should switch would be trivial: if you have at least a $\frac{1}{2}$ probability of getting the car when switching, then you should obviously switch to try to get the car.2017-02-21
  • 0
    Yes, technically, you could want the goat, and technically the next number in the sequence 1, 2, 3, 4, _ could be $\pi$ instead of 5, but that's clearly not the case. My issue with the problem is not an unlikely alternative possibility but rather that what the author's intention is not consistent and cannot be interpreting correctly even after filling in the blanks, which all math texts require one to do.2017-02-21