### The situation

Some researchers would like to put you to sleep. Depending on the secret toss of a fair coin, they will briefly awaken you either once (Heads) or twice (Tails). After each waking, they will put you back to sleep with a drug that makes you forget that awakening. When you are awakened, to what degree should you believe that the outcome of the coin toss was Heads?

(OK, maybe you don’t want to be the subject of this experiment! Suppose instead that Sleeping Beauty (SB) agrees to it (with the full approval of the Magic Kingdom’s Institutional Review Board, of course). She’s about to go to sleep for one hundred years, so what are one or two more days, anyway?)

[Detail of a Maxfield Parrish illustration.]

### Are you a Halfer or a Thirder?

The Halfer position. Simple! The coin is fair--and SB knows it--so she should believe there's a one-half chance of heads.

The Thirder position. Were this experiment to be repeated many times, then the coin will be heads only one third of the time SB is awakened. Her probability for heads will be one third.

### Thirders have a problem

• On Sunday evening, just before SB falls asleep, she must believe the chance of heads is one-half: that’s what it means to be a fair coin.

• Whenever SB awakens, she has learned absolutely nothing she did not know Sunday night. What rational argument can she give, then, for stating that her belief in heads is now one-third and not one-half?

### Some attempted explanations

• SB would necessarily lose money if she were to bet on heads with any odds other than 1/3. (Vineberg, inter alios)

• One-half really is correct: just use the Everettian “many-worlds” interpretation of Quantum Mechanics! (Lewis).

• SB updates her belief based on self-perception of her “temporal location” in the world. (Elga, i.a.)

• SB is confused: “[It] seems more plausible to say that her epistemic state upon waking up should not include a definite degree of belief in heads. … The real issue is how one deals with known, unavoidable, cognitive malfunction.” [Arntzenius]

### The question

Accounting for what has already been written on this subject (see the references as well as a previous post), how can this paradox be resolved in a statistically rigorous way? Is this even possible?

### References

Arntzenius, Frank (2002). Reflections on Sleeping Beauty Analysis 62.1 pp 53-62.

Bradley, DJ (2010). Confirmation in a Branching World: The Everett Interpretation and Sleeping Beauty. Brit. J. Phil. Sci. 0 (2010), 1–21.

Elga, Adam (2000). Self-locating belief and the Sleeping Beauty Problem. Analysis 60 pp 143-7.

Franceschi, Paul (2005). Sleeping Beauty and the Problem of World Reduction. Preprint.

Groisman, Berry (2007). The end of Sleeping Beauty’s nightmare. Preprint.

Lewis, D (2001). Sleeping Beauty: reply to Elga. Analysis 61.3 pp 171-6.

Papineau, David and Victor Dura-Vila (2008). A Thirder and an Everettian: a reply to Lewis’s ‘Quantum Sleeping Beauty’.

Pust, Joel (2008). Horgan on Sleeping Beauty. Synthese 160 pp 97-101.

Vineberg, Susan (undated, perhaps 2003). Beauty’s Cautionary Tale.

-
I was moved to post this as a separate question based on comments at stats.stackexchange.com/questions/23779. – whuber Oct 25 '12 at 20:10
It would be good if you could describe the experiment a bit clearer. Without reading the original post, it is really hard to understand what the paradox is about. – sebhofer Oct 25 '12 at 22:51
My comment wasn't meant to be rude btw. I realized later it might have come across a bit harsh. Hope you didn't take it the wrong way. – sebhofer Oct 28 '12 at 0:42
You might be interested in the (now large) literature in philosophy on this paradox. Here is a fairly complete bibliography (with links): philpapers.org/browse/sleeping-beauty – user16414 Oct 31 '12 at 14:13
Thank you, @jpust! That is a wonderful resource. It's going to take a while to get through it. :-) – whuber Jan 17 '13 at 1:53

### Strategy

I would like to apply rational decision theory to the analysis, because that is one well-established way to attain rigor in solving a statistical decision problem. In trying to do so, one difficulty emerges as special: the alteration of SB’s consciousness.

• Rational decision theory has no mechanism to handle altered mental states.

• In asking SB for her credence in the coin flip, we are simultaneously treating her in a somewhat self-referential manner both as subject (of the SB experiment) and experimenter (concerning the coin flip).

Let’s alter the experiment in an inessential way: instead of administering the memory-erasure drug, prepare a stable of Sleeping Beauty clones just before the experiment begins. (This is the key idea, because it helps us resist distracting--but ultimately irrelevant and misleading--philosophical issues.)

• The clones are like her in all respects, including memory and thought.

• SB is fully aware this will happen.

We can clone, in principle. E. T. Jaynes replaces the question "how can we build a mathematical model of human common sense"--something we need in order to think through the Sleeping Beauty problem--by "How could we build a machine which would carry out useful plausible reasoning, following clearly defined principles expressing an idealized common sense?" Thus, if you like, replace SB by Jaynes' thinking robot, and clone that.

(There have been, and still are, controversies about "thinking" machines.

"They will never make a machine to replace the human mind—it does many things which no machine could ever do."

You insist that there is something a machine cannot do. If you will tell me precisely what it is that a machine cannot do, then I can always make a machine which will do just that!”

--J. von Neumann, 1948. Quoted by E. T. Jaynes in Probability Theory: The Logic of Science, p. 4.)

--Rube Goldberg

### The Sleeping Beauty experiment restated

Prepare $n \ge 2$ identical copies of SB (including SB herself) on Sunday evening. They all go to sleep at the same time, potentially for 100 years. Whenever you need to awaken SB during the experiment, randomly select a clone who has not yet been awakened. Any awakenings will occur on Monday and, if needed, on Tuesday.

I claim that this version of the experiment creates exactly the same set of possible results, right down to SB's mental states and awareness, with exactly the same probabilities. This potentially is one key point where philosophers might choose to attack my solution. I claim it's the last point at which they can attack it, because the remaining analysis is routine and rigorous.

Now we apply the usual statistical machinery. Let's begin with the sample space (of possible experimental outcomes). Let $M$ mean "awakens Monday" and $T$ mean "awakens Tuesday." Similarly, let $h$ mean "heads" and "t" mean tails. Subscript the clones with integers $1, 2, \ldots, n$. Then the possible experimental outcomes can be written (in what I hope is a transparent, self-evident notation) as the set

\eqalign{ \{&hM_1, hM_2, \ldots, hM_n, \\ &(tM_1, tT_2), (tM_1, tT_3), \ldots, (tM_1, tT_n), \\ &(tM_2, tT_2), (tM_2, tT_3), \ldots, (tM_2, tT_n), \\ &\cdots, \\ &(tM_{n-1}, tT_2), (tM_{n-1}, tT_3), \ldots, (tM_{n-1}, tT_n) & \}. }

### Monday probabilities

As one of the SB clones, you figure your chance of being awakened on Monday during a heads-up experiment is ($1/2$ chance of heads) times ($1/n$ chance I’m picked to be the clone who is awakened). In more technical terms:

• The set of heads outcomes is $h = \{hM_j, j=1,2, \ldots,n\}$. There are $n$ of them.

• The event where you are awakened with heads is $h(i) = \{hM_i\}$.

• The chance of any particular SB clone $i$ being awakened with the coin showing heads equals $$\Pr[h(i)] = \Pr[h] \times \Pr[h(i)|h] = \frac{1}{2} \times \frac{1}{n} = \frac{1}{2n}.$$

### Tuesday probabilities

• The set of tails outcomes is $t = \{(tM_j, tT_k): j \ne k\}$. There are $n(n-1)$ of them. All are equally likely, by design.

• You, clone $i$, are awakened in $(n-1) + (n-1) = 2(n-1)$ of these cases; namely, the $n-1$ ways you can be awakened on Monday (there are $n-1$ remaining clones to be awakened Tuesday) plus the $n-1$ ways you can be awakened on Tuesday (there are $n-1$ possible Monday clones). Call this event $t(i)$.

• Your chance of being awakened during a tails-up experiment equals $$\Pr[t(i)] = \Pr[t] \times P[t(i)|t] = \frac{1}{2} \times \frac{2(n-1}{n(n-1)} = \frac{1}{n}.$$

### Bayes' Theorem

Now that we have come this far, Bayes' Theorem--a mathematical tautology beyond dispute--finishes the work. Any clone's chance of heads is therefore $$\Pr[h | t(i) \cup h(i)] = \frac{\Pr[h]\Pr[h(i)|h]}{\Pr[h]\Pr[h(i)|h] + \Pr[t]\Pr[t(i)|t]} = \frac{1/(2n)}{1/n + 1/(2n)} = \frac{1}{3}.$$

Because SB is indistinguishable from her clones--even to herself!--this is the answer she should give when asked for her degree of belief in heads.

### Interpretations

The question "what is the probability of heads" has two reasonable interpretations for this experiment: it can ask for the chance a fair coin lands heads, which is $\Pr[h] = 1/2$ (the Halfer answer), or it can ask for the chance the coin lands heads, conditioned on the fact that you were the clone awakened. This is $\Pr[h|t(i) \cup h(i)] = 1/3$ (the Thirder answer).

In the situation in which SB (or rather any one of a set of identically prepared Jaynes thinking machines) finds herself, this analysis--which many others have performed (but I think less convincingly, because they did not so clearly remove the philosophical distractions in the experimental descriptions)--supports the Thirder answer.

The Halfer answer is correct, but uninteresting, because it is not relevant to the situation in which SB finds herself. This resolves the paradox.

This solution is developed within the context of a single well-defined experimental setup. Clarifying the experiment clarifies the question. A clear question leads to a clear answer.

I guess that, following Elga (2000), you could legitimately characterize our conditional answer as "count[ing] your own temporal location as relevant to the truth of h," but that characterization adds no insight to the problem: it only detracts from the mathematical facts in evidence. To me it appears to be just an obscure way of asserting that the "clones" interpretation of the probability question is the correct one.

This analysis suggests that the underlying philosophical issue is one of identity: What happens to the clones who are not awakened? What cognitive and noetic relationships hold among the clones?--but that discussion is not a matter of statistical analysis; it belongs on a different forum.

-
This answer summarizes a talk I prepared in December 2008 and posted on the Web at that time in PowerPoint format. Its conclusion appears to be substantially similar to Groisman's, even though the justification may be different: "If we mean ‘This awakening is a Head-awakening under setup of wakening’, then her answer should be 1/3, but if we mean ‘The coin landed Heads under setup of coin tossing’, her answer should be 1/2." See philsci-archive.pitt.edu/3382/1/SB_PhilSci.pdf. – whuber Oct 25 '12 at 20:11
I attack it at exactly your unscaled underbelly. See my detailed analysis below. – Dax Fohl Sep 1 at 5:29

"Whenever SB awakens, she has learned absolutely nothing she did not know Sunday night." This is wrong, as wrong as saying "Either I win the lottery or I don't, so the probability is $50\%$." She has learned that she has woken up. This is information. Now she should believe each possible awakening is equally likely, not each coin flip.

If you are a doctor and a patient walks into your office, you have learned that the patient has walked into a doctor's office, which should change your assessment from the prior. If everyone goes to the doctor, but the sick half of the population goes $100$ times as often as the healthy half, then when the patient walks in you know the patient is probably sick.

Here is another slight variation. Suppose whatever the outcome of the coin toss was, Sleeping Beauty will be woken up twice. However, if it is tails, she will be woken up nicely twice. If it is heads, she will be woken up nicely once, and will have a bucket of ice dumped on her once. If she wakes up in a pile of ice, she has information that the coin came up heads. If she wakes up nicely, she has information that the coin probably didn't come up heads. She can't have a nondegenerate test whose positive result (ice) tells her heads is more likely without the negative result (nice) indicating that heads is less likely.

-

This post has been locked while disputes about its content are being resolved. For more info visit meta.

Intriguing (+1). But I can't help thinking that a Halfer might come back with something like "but SB knew in advance that she would be awakened, so the experience of awakening provides no new information." It seems akin to the more prosaic example of an ordinary coin flip. After the coin is flipped--but before you learn the outcome--you know the coin has been flipped. But it's either nonsensical or useless to then assert the probability of heads is either 1 or 0. Your credence for heads remains exactly the same it was before the flip. Some kinds of information do not change probabilities. – whuber Oct 27 '12 at 20:26
In the ice/nicely variation, would the Halfer say Sleeping Beauty gains some information from finding out that she is woken up nicely? The original puzzle is equivalent to this case, so the update to the probabilities should be the same. – Douglas Zare Oct 28 '12 at 0:42
The ice/nice variation is interesting indeed--well worth careful consideration. Because even its sample space is different, how do you convincingly demonstrate that the original problem is equivalent to it? Your final statement makes sense, but what is the proof of it? – whuber Oct 28 '12 at 16:07
I think you need to represent this as some sort of filtered probability space, and then there should be an isomorphism between the two. I haven't done this yet. – Douglas Zare Oct 30 '12 at 15:09
@DouglasZare I initially agreed with you, but changed my opinion (see my updated answer). – gui11aume Jan 5 '13 at 11:46

I just thought of the following situation which does not require fairies, miracles nor magic potions. Flip a fair coin on Monday noon. Upon 'Tails' send a mail to Alice and Bob (in a way that they don't know that the other has received a mail from you, and that they cannot communicate). Upon 'Heads', send a mail to one of them at random (with probability $1/2$).

When Alice receives a mail, what is the probability that the coin landed on 'Heads'? The probability that she receives a letter is $1/2 \times 1/2 + 1/2 = 3/4$, and the probability that the coin landed on 'Heads' is $1/3$.

Here there is no paradox because Alice does not receive a letter with probability $1/4$, in which case she knows the coin landed on 'Heads'. The fact that we don't ask her opinion in that case, does make this probability equal to 0.

So, what is the difference? Why would Alice gain information by receiving a mail, and SB would learn nothing being awakened?

Moving on to a more miraculous situation, we put 2 different SB to sleep. If the coin lands on 'Tails' we wake up both, if it lands on 'Heads' we wake up one of them at random. Here again, each of the SB should say that the probability of the coin landing on 'Heads' is $1/3$ and again there is no paradox because there is a $1/4$ chance that this SB would not be awakened.

But this situation is very close to the original paradox because erasing the memory (or cloning) is equivalent to having two different SB. So, I am with @Douglas Zare here (+1). SB has learned something by being awakened. The fact that she cannot express her opinion on Tuesday when the coin is 'Heads' up because she is sleeping does not erase the information she has by being awakened.

In my opinion the paradox lies in "she has learned absolutely nothing she did not know Sunday night" which is stated without justification. We have this impression because the situations when she is awakened are identical, but this is just like Alice receiving a mail: it is the fact that she is asked her opinion that gives her information.

MAJOR EDIT: After giving it a deep thought, I change my opinion: Sleeping Beauty has learned nothing and the example I give above is not a good analogue of her situation.

But here is an equivalent problem that is not paradoxical. I could play the following game with Alice and Bob: I toss a coin secretly and independently bet them 1\$that they cannot guess it. But if the coin landed on 'Tails', the bet of either Alice of Bob is cancelled (money does not change hand). Given that they know the rules, what should they bet? 'Heads' obviously. If the coin lands on 'Heads', they gain 1\$, otherwise, they lose 0.5\$on average. Does it mean that they believe that the coin has a 2/3 chance of landing on 'Heads'? Sure not. Simply the protocol is such that they do not gain the same amount of money for each answer. I believe that Sleeping Beauty is in the same situation as Alice or Bob. The events give her no information about the toss, but if she is asked to bet, her odds are not 1:1 because of asymmetries in the gain. I believe that this is what @whuber means by The Halfer answer is correct, but uninteresting, because it is not relevant to the situation in which SB finds herself. This resolves the paradox. - +1. As explained in my comment to Zare's answer, I'm struggling to understand the distinction you are making between knowing in advance you will be awakened and knowing you have been awakened. What specifically is learned upon awakening, when you were 100% sure that the awakening would occur? – whuber Oct 27 '12 at 20:31 @whuber your comment led me think about it again. See the updated answer. – gui11aume Jan 5 '13 at 11:45 I just re-tripped across this. I've refined some of my thoughts since that last post, and thought I might find a receptive audience for them here. First off, on the philosophy of how to address such a controversy: Say arguments A and B exist. Each has a premise, a sequence of deductions, and a result; and the results differ. The best way way to prove one argument is incorrect is to invalidate one of its deductions. If that were possible here, there wouldn't be a controversy. Another is to disprove the premise, but you can't do that directly. You can argue for why you don’t believe one, but that won't resolve anything unless you can convince others to stop believing it. To prove a premise wrong indirectly, you have to form an alternate sequence of deductions from it that leads to an absurdity or to a contradiction of the premise. The fallacious way is to argue that the opposing result violates your premise. That means that one is wrong, but it doesn't indicate which. +++++ The halfer's premise is "no new information." Their sequence of deductions is empty - none are needed. Pr(Heads|Awake) = Pr(Heads)=1/2. The thirders (specifically, Elga) have two premises - that Pr(H1|Awake and Monday) = Pr(T1|Awake and Monday), and Pr(T1|Awake and Tails) = Pr(T2|Awake and Tails). An incontrovertible sequence of deductions then leads to Pr(Heads|Awake) = 1/3. Note that the thirders don't ever assume there is new information - their premises are based on whatever information exists - "new" or not - when SB is awake. And I've never seen anyone argue for why a thirder premise is wrong, except that it violates the halfer result. So the halfers have provided none of the valid arguments I've listed. Just the fallacious one. But there are other deductions possible from "no new information," with a sequence of deductions that start with Pr(Heads|Awake) = 1/2. One is that Pr(Heads|Awake and Monday) = 2/3 and Pr(Tails|Awake and Monday) = 1/3. This does contradict the thirder premise, but like I said, that doesn’t help the halfer cause since it still could be their premise that is wrong. Ironically, this result does prove something - that the halfer premise contradicts itself. On Sunday, SB says Pr(Heads|Monday) = Pr(Tails|Monday), so adding the information "Awake" has allowed her to update these probabilities. It is new information. So I have proven the halfer premise can't be right. That doesn't mean the thirders are right, but it does mean that halfers have not provided any contrary evidence. +++++ There is another argument I find more convincing. It isn't completely original, but I'm not sure if the proper viewpoint has been emphasized enough. Consider a variation of the experiment: SB is always wakened on both days; usually it is in a room that is painted blue, but on Tuesday after Heads it is in a room that is painted red. What should she say the probability of Heads is, if she finds herself awake in a blue room? I don’t think anybody would seriously argue that it is anything but 1/3. There are three situations that could correspond to her current one, all are equally likely, and only one includes Heads. The salient point is that there is no difference between this version, and the original. What she "knows" - her "new information" - is that it is not H2. It does not matter how, or IF, she would know it could be H2 if it could. Her capability to observe situations that she knows do not apply is irrelevant if she knows they do not apply. I can not believe the halfer premise. It is based on a fact - that she can't observe H2 - that cannot matter since she can, and does, observe that it isn't H2. So I hope that I have provided a convincing argument for why the halfer premise is invalid. Along the way, I know I have demonstrated that the thirder result must be correct. - Pr(Heads, Monday, Blue)=50%, Pr(Heads, Monday, Red)=0%, Pr(Heads, Tuesday, Blue)=0%, Pr(Heads, Tuesday, Red)=0%, Pr(Tails, Monday, Blue)=25%, Pr(Tails, Monday, Red)=0%, Pr(Tails, Tuesday, Blue)=25%, Pr(Tails, Tuesday, Red)=0%. Thus, halfer. – Dax Fohl Sep 1 at 4:36 "Whenever SB awakens, she has learned absolutely nothing she did not know Sunday night." This isn't correct, which is the error in the halfer argument. One thing that makes it hard to argue with,tho, is that the halfer argument which is based on this statement is seldom expressed with any more rigor than what I quoted. There are three problems. First, the argument does not define what "new information" means. It seems to mean "An event that originally had a non-zero probability cannot have occurred based on the evidence." Second, it never enumerates what is known on Sunday to see if it fits this definition; and it can, if you look at it properly. Finally, there is no theorem that says "if you have no new information of this kind, you can't update." If you do have it, Bayes Theorem will produce an update. But it is a fallacy to conclude, if you don't have this new information, that you can't update. Being a fallacy doesn't mean it isn't true, it means you can't make this conclusion based on this evidence alone. On Sunday Night, say SB rolls an imaginary six-sided die of her own. Since it is imaginary, she can't look at the result. But the purpose is to see if it matches the day she is awake: an even number means it matches Monday, and an odd number means Tuesday. But it can't match both, which effectively distinguishes the two days. SB can now (that is, on Sunday) calculate the probability for the eight possible combinations of {Heads/Tails, Monday/Tuesday, Match/No Match}. Each will be 1/8. But when she is awake, she knows that {Heads, Tuesday, Match} and {Heads, Tuesday, No Match} did not happen. This constitutes "new information" of the form the halfers argument says doesn’t exist, and it allows SB to update the probability that the researcher's coin landed on heads. It is 1/3 whether or not her imaginary coin matches the actual day. Since it is the same either way, it is 1/3 whether or not she knows if there is a match; and in fact, whether or not she rolls, or imagines rolling, the die. This extra die seems like a lot to go through to get a result. In fact, it isn’t necessary, but you need a different definition of "new information" to see why. Updating can occur anytime the significant (i.e., independent and not zero-probability) events in the prior sample space differ from the significant events in the posterior sample space. That way, the denominator of the ratio in Bayes Theorem is not 1. While this usually occurs when the evidence makes some of the events have zero probability, it can also occur when the evidence changes whether events are independent. This is a very unorthodox interpretation, but it works because Beauty is given more than one opportunity observe an outcome. And the point of my imaginary die, which distinguished the days, was to render the system into one where the total probability was 1. On Sunday, SB knows P(Awake,Monday,Heads) = P(Awake,Monday,Tails) = P(Awake,Tuesday,Tails)=1/2. These add up to more than 1/2 because the events are not independent based on the information SB has on Sunday. But they are independent when she is awake. The answer, according to Bayes Theorem, is (1/2)/(1/2+1/2+1/2)=1/3. There is nothing wrong with a denominator that is greater that 1; but the imaginary coin argument was designed to accomplish the same things without such a denominator. - Welcome to CV, @JeffJo. This is an interesting argument, but the tone comes across as somewhat testy. You should be cautious about that, lest people misinterpret it as rudeness. – gung Jan 15 '13 at 0:08 Sorry about that tone - it really wasn't intended that way. The problem with probability paradoxes is that there are undefinable terms, multiple paths to solution, and simple shortcuts that are often taken without proper justification. The upshot is that, to convince a proponent of the "wrong" answer that yours is "rigorous," you have to both demonstrate yours with no room for objection, and find an inescapable hole in the opposing argument. I think my attempts to point out that hole are what you found "testy." – JeffJo Jan 15 '13 at 22:01 The paradox lies in the perspective change between a single experiment and its limit point. If # of experiments is taken into account, you can understand this even more precisely than the "either/or" of halvers and thirders: # Single Experiment: Halvers are right If there is a single experiment, there are three outcomes and you just have to figure the probabilities from the perspective of the awakened: 1. Heads was tossed: 50% 2. Tails was tossed and this is my first awakening: 25% 3. Tails was tossed and this is my second awakening: 25% So, in a single experiment, at any wakeup event, you should assume 50/50 that you are in a state where heads was tossed # Two experiments: 42%ers are right Now, try two experiments: 1. Heads was tossed twice: 25% (for both awakenings combined) 2. Tails was tossed twice: 25% (for all four awakenings combined) 3. Heads then Tails and this is my first awakening: 25%/3 4. Heads then Tails and this is my 2nd or 3rd awakening: 25%*2/3 5. Tails then Heads and this my 1st or 2nd awakening: 25%*2/3 6. Tails then Heads and this is my 3rd awakening: 25%/3. So here, {1, 3, 6} are your Heads states, with a combined probability of (25 + 25/3 + 25/3)%, 41.66%, which is less than 50%. If two experiments are run, at any wakeup event, you should assume 41.66% chance you are in a state where Heads was thrown # Infinite experiments: Thirders are right I'm not going to do the math here, but if you look at the two-experiments options, you can see #1 and #2 drive it toward halves, and the rest drive it toward thirds. As the number of experiments increases, the options driving toward halves (all heads/all tails) will decrease in probability down to zero, leaving the "thirds" options to take over. If infinite experiments are run, at any wakeup event, you should assume 1/3 chance you are in a state where heads was thrown # Preempting Retorts: ### But, gambling? Yes in the single experiment instance, you should still "gamble" by the thirds. This is not an inconsistency; it's just because you may be placing the same bet multiple times given a certain outcome, and know this in advance. (Or if you don't, the mafia does). ### Okay, how about two single experiments? Discrepancy much? No, because knowledge about whether you're on the first or 2nd experiment adds to your, erm, knowledge. Let's look at the "two experiments" options and filter them by knowledge that you're on the first experiment. 1. Applicable for first awakening (1/2) 2. Applicable for first two awakenings (2/4) 3. Applicable 4. Never applicable 5. Applicable for first awakening (1/2) 6. Not applicable Okay, take the Heads ones (1,3,6) multiply these, odds by applicability: 25/2 + 25/3 + 0 = 125/6. Now take the Tails ones (2,4,5) and do the same: 25*4/2 + 0 + 25*(2/3)/2 = 125/6. Viola, they're the same. The added information about which experiment you're in in fact adjusts the odds of what you know. ### But, the clones!! Simply put, contrary to the OP's answer's postulate, that cloning creates an equivalent experiment: cloning plus random selection does change the knowledge of the experimentee, in the same way "multiple experiments" changes the experiment. If there are two clones, you can see the probabilities of each clone correspond to the Two Experiments probabilities. Infinite clones converges to thirders. But it's not the same experiment, and it's not the same knowledge, as a single experiment with a single non-random subject. ### You say "random one of infinite" and I say Axiom of Choice dependency I don't know, my set theory isn't that great. But given for N less than infinity, you can establish some sequence that converges from half to a third, the infinite case equaling a third will either be true or undecidable at worst, no matter which axioms you invoke. - I feel I disagree with the fact that the probability of heads given I have been awakened is 50%, there is now new information. – rwolst Sep 1 at 12:11 A simple explanation for this would be that there are 3 ways in which sleeping beauty can wake up two of which are from a Tails toss. So the probability has to be 1/3 for a heads every time she wakes up. I've outlined it in a blog post The main argument against the "halfer" point of view is the following: In a bayesian sense, SB is always looking to see what new information she has. In reality, the moment she has decided to take part in the experiment, she has additional information that when she wakes up it could be in of the days. Or put in other words the lack of information (wiping out the memory) is what is providing the evidence here, subtly though. - Yes, this is part of the Thirder argument. But it does not explain why the Halfer argument is incorrect. – whuber Nov 12 '12 at 21:49 I like this, and I think a slight tweak will improve it further: suppose that if the coin is "heads", one will be awoken on Monday, and if it's tails one will be awoken on Tuesday and again on Wednesday. There are three days when one may wake up, and all three are equally likely. The times one wakes up on Monday, the coin will have been heads; on Tuesday or Wednesday, tails. – supercat Feb 21 '14 at 23:41 @supercat This just assumes "three things" are "three equal probabilities", "just because". Assume the coin is weighted a million to one heads, and that argument falls apart. You have to calculate actual probabilities. Monday is 50%. – Dax Fohl Sep 1 at 4:27 For that scenario, put 1,000,001 beauties in separate rooms, and count how many awakening have occurred after a "heads" flip and how many after a "tails" flip. There will be 1,000,002 awakenings, of which 2 will have occurred after tails flips, so the odds would be 500,000:1 heads. – supercat Sep 1 at 14:45 @DaxFohl: See previous comment. For the Monday/Tuesday/Wednesday scenario, the probabilities are equal because the head/tail probabilities are equal. Biasing the coin would mean that 1,000,000 awakenings out of 1,000,002 would be on Monday after heads, one would be Tuesday after tails, and one Wednesday after tails. – supercat Sep 1 at 16:39 One third of possible wakings are Heads wakings, and two thirds of possible wakings are Tails wakings. However, one half of princesses (or whatever) are Heads princesses, and one half are Tails princesses. The Tails princesses, individually and in aggregate, experience twice as many wakings as the Heads princesses. From the perspective of the princess, on waking up, there are three possibilities. She is either a Heads princess awaking for the first (and only) time ($H1$), a Tails princess awaking for the first time ($T1$), or a tails princess awaking for a second time ($T2$). There seems no reason to assume that these three outcomes are equally likely. Rather$P[H1]=0.5$,$P[T1]=0.25$, and$P[T2]=0.25$. I haven't read Vineberg's reasoning, but I think I can see how she arrives at a fair bet of$\$1/3$. Suppose that every time a princess awakens, she makes a bet of $\$x$that she is a Heads princess, receiving \$1 if she is indeed a Heads princess, and \$0 otherwise. Then a Heads princess will receive$\$(1-x)$, and a Tails princess will receive $\$(-x)$each time she plays. Since the Tails princesses must play twice, and since half of princesses are Heads princesses, the expected return is$\$(1-3x)/2$, and the fair price is $\$1/3$. Normally this would be conclusive evidence that the probability is$1/3$, but the usual reasoning does not hold in this case: the princesses who are destined to lose the bet are obliged to play the game twice, whereas those who are destined to win will play only once! This imbalance uncouples the usual relationship between probabilities and fair bets. (On the other hand, a technician who was assigned to help with the waking process really would have only a one third chance of being assigned to a Heads princess.) - As many questions, it depends of the exact meaning of the question: When you are awakened, to what degree should you believe that the outcome of the coin toss was Heads? If you are interpret it as "what are the odds that a tossed coin is Heads", obviously the answer is "half the odds". But what you are asking is not (in my interpretation) that, but "which is the chance that the current awakening was caused by a Heads?". In that case, obviously only a third of the awakenings are caused by a Heads, so the most probable answer is "Tails". - But there is not "third of". There's one or two, not three or six or infinity. So in that respect, the questions are the same, and both have answer "1/2". – Dax Fohl Sep 1 at 4:19 This is a very interesting question. I will give my answer as if I were to be sleeping beauty. I feel a key point to understand is that we 100% trust the experimenter. 1) On Sunday night, if you ask me what the probability the coin is heads, I will tell you$\frac{1}{2}$. 2) Whenever you wake me up and ask me, I will tell you$\frac{1}{3}$. 3) When you tell me that this is the last time you are awakening me I will immediately switch to telling you the probability is$\frac{1}{2}$. Clearly (1) follows from the fact the coin is fair. (2) follows from the fact that when you are are woken, you are in one of 3 equally likely situations from your point of view. Each of them can occur with probability$\frac{1}{2}\$.

Then (3) follows in the same manner except that as soon as you are told this is the last time you are being awakened, the number of situations you can be in collapses to 2 (as now tails and this being the first time you were awakened is impossible).

-
How can 3 equally likely situations occur with probability 1/2? – Dax Fohl Sep 1 at 4:21
@DaxFohl Because they are not mutually exclusive. – isaacg Sep 1 at 8:31

The question is ambiguous and so there only appears to be a paradox. The question is posed this way:

When you are awakened, to what degree should you believe that the outcome of the coin toss was Heads?

Which is confused with this question:

When you are awakened, to what degree should you believe Heads was the reason you were awakened?

In the first question the probability is 1/2. In the second question, 1/3.

The problem is that the first question is stated, but second question is implied in the context of the experiment. Those who subconsciously accept the implication say it's 1/3. Those who read the question literally say it's 1/2.

Those who are confused are not sure which question they're asking!

-
I'm a halver with respect to the 2nd question as well. – Dax Fohl Sep 1 at 4:16

Beauty's credence for Heads is p=1/2.

## Halfers love chocolate frogs

Let's jolly-up the experiment by awarding Beauty with a chocolate frog if the coin is flipped Heads. Upon each wake-up, we'll ask Beauty: how many chocolate frogs do you expect, on average, to receive in this experiment? Of course this number is exactly her credence for Heads, p.

Now we, as outsider, know that Beauty will receive on average half a frog per experiment. The halfer position is that, when asked, she will concur: p=1/2. This, being the correct value of the average number of frogs, is consistent with the premise that Beauty is rational and mathematically competent.

Thirders claim that Beauty will report p=1/3, apparently violating the premise of Beauty's rationality.

## Thirders die young

A more complicated experiment has Beauty being presented at each wake-up with two enormous jars of indistinguishable "beans": a mix of tasty "J"ellybeans and lethal "K"ill-pills. The jars are filled as follows:

• if Heads, the ratio of J:K in jar "A" is 7:0, in "B" it's 0:7.
• if Tails, the ratio of J:K in jar "A" is 1:6, in "B" it's 6:1

She must pick one bean (from either "A" or "B"), set it aside, and will consume her choice(s) at the end of the experiment. She is not told any of this beforehand, but learns it (to her dismay!) upon wake-up.

Picking "A" gives an overall survival probability of p+(1-p)/49; picking "B" gives 36(1-p)/49. Halfers will find "A" the more attractive choice, but thirders will apparently be fooled into picking "B", once again violating the premise of Beauty's rationality.

## Elga got it wrong (but so did Lewis)

Adam Elga is the original thirder. Let me write "Cr" for "credence". Elga's argument is basically that because ...

• Cr(Mon|Tails) = Cr(Tue|Tails)

... then Cr(Heads,Mon) = Cr(Tails,Mon) = Cr(Tue,Tails) and, by virtue of those joint events being exhaustive and exclusive, they must all be equal to one third.

This is a neat sleight of hand. It fails because "Mon" and "Tue" are not probabilistic events, so using Bayes rule to multiply-up by Cr(Mon), as if that were a probability, is inadmissible. The only probabilistic event is Heads/Tails. "Cr(Heads|Mon)" is a genuine probability -- we could equally write Pr(Heads|Mon) -- because Heads is a probabilistic event; but "Cr(Mon|Heads)" is something else.

Incidentally, David Lewis, the original halfer, in reply to Elga, makes the same mistake of treating Mon as a probabilistic event. In that case, it leads to the unwelcome (wrong) additional deduction that Cr(Heads|Mon) = 2/3. Actually that credence should be 1/2 again: the so-called "double halfer" position.

## Huber's incredulous clones

In this thread, William Huber formulates a variant of the Sleeping Beauty experiment in which one (Heads) or two (Tails) SB clones are awoken. He posits:

I claim that this version of the experiment creates exactly the same set of possible results, right down to SB's mental states and awareness, with exactly the same probabilities.

In the ordinary SB experiment, Beauty is unsurprised to be woken for a Monday-or-Tuesday interview. In the variant, let's say with a billion clones, she might be quite incredulous! Since the whole SB puzzle of one of "credence", this red-flags the variant as not equivalent at all to the original puzzle.

The outcome of all the honest-to-goodness random variables (just the coin flip in our case) defines a "world". For this problem, I like to call it a "trajectory" of wake-ups. If you know what trajectory you're on, you should reason about your location on that trajectory (i.e. the day of the week) by the principle of indifference: lacking evidence to the contrary, each is equally likely. None of this should be controversial.

So, for example, speaking about the Tails trajectory, we can say Cr(Mon|Tails) = Cr(Tue|Tails) = 1/2.

When the trajectory is unknown, you have to form your credence as a weighted average over trajectories: Cr(thing) = Pr(Heads).Cr(thing|Heads) + Pr(Tails).Cr(thing|Tails) -- the law of total expectation. Thus, credence, "Cr", is an expectation, not in general a probability.

To put it into symbols:

• Cr(X) = sum_E Pr(E).Cr(X|E) ... summing over exhaustive, exclusive, probabilistic events E
• Cr(E) = Pr(E) ... if E is a probabilistic event
• Cr(X|E) = Cr*(X|E) ... if E fully defines a trajectory, where Cr* indicates reasoning by the principle of indifference

### Example

What's Beauty's credence for Monday? In unnecessarily-tedious detail, let's average over trajectories to find out:

• Cr(Mon) = 1/2.Cr*(Mon|Heads) + 1/2.Cr*(Mon|Tails) ... because the trajectories are fully defined by Heads and Tails
• Cr(Mon) = 1/2.1 + 1/2.1/2 ... by the principle of indifference on those trajectories
• Cr(Mon) = 3/4 ... by adding up

## Self-location assumptions

Of course we can't prove anything about credence, because at its heart we are making some philosophical assumptions. The key distinction for us is that of the Self Sampling Assumption (SSA) vs the Self Indication Assumption (SIA) from the philosophy of anthropic probability. SSA basically says we should reason by taking the average over "worlds" (as we did above); SIA says we should average over outcomes (e.g. the triumvirate of Heads+Mon, Tails+Mon and Tails+Tue).

So readers might now feel it's all been a wild goose chase leading to philosophical hogwash. Maybe, but I hope not. Although SSA and SIA are applied (or rejected!) widely in more general anthropic philosophy settings, they are not needed explicitly for this problem, so the SSA/SIA controversy should not let us just wave Sleeping Beauty away as intractable.

For further reading, admittedly supporting the "double halfer"-style reasoning given in this post, consider starting with Nick Bostrum's "Sleeping beauty and self-location: A hybrid model".

-
Halfers love chocolate frogs - "Of course this number is exactly her credence for Heads, p." - this is an unjustified leap. You can conclude that you will receive 1/2 chocolate frog in the whole experiment, on average, without believing that in a specific awakening the coin went one way or the other. – spookylukey Sep 1 at 12:31
"Thirders die young" - the scenario doesn't work, because it doesn't define what happens in the case of Tails when she will choose 2 beans. Modifications to fix this either change the scenario i.e. give her new information, or involve lying to her, in which case all choices based on trust are off. – spookylukey Sep 1 at 12:36
Re frogs: if Beauty believes with credence p=1/3 that she will receive one chocolate frog (because the coin landed Heads) and 1-p=2/3 that she will receive no chocolate frogs (because the coin landed Tails), then she expects to receive p=1/3 chocolate frogs on average per experiment. Re beans: she sets aside her choice for later consumption (either one bean in total if Heads, else two beans in total if Tails); she eats both on Wednesday. How do thirders, with p=1/3, decide to pick from jar "A", I wonder? – Creosote Sep 1 at 19:02

I really like this example but I would argue that there is one point to make confounded with a couple of nuisance distractions.

To avoid nuisance distractions, one arguable should try to discern an abstract diagrammatic representation of the problem that is clearly beyond reasonable doubt (as an adequate representation) and can be verifiably manipulated (re-manipulated by qualified others) to demonstrate the claims. As a simple example think of an (abstract mathematical) rectangle and the claim that it can be made into two triangles.

Draw a free hand rectangle as a representation of a mathematical rectangle (in your drawing the four angles will not add exactly to 180 degrees and the adjacent lines will not be exactly equal or straight but there will be no real doubt that it represents a true rectangle). Now manipulate it by drawing a line from one opposite corner to another, which anyone else could do and you get a representation of two triangles that no one would reasonably doubt. Any questioning of can this be so seems nonsense, it just is.

The point I am try to make here is that if you get a beyond a reasonable doubt representation of the SB problem as a joint probability distribution and can condition on an event that happens in the experiment in this representation - then claims of whether anything is learned by that event can be demonstrated by verifiable manipulation and require no (philosophical) discussion or questioning.

Now I better present my attempt and readers will need to discern if I have succeeded. I will use a probability tree to represent joint probabilities for day sleeping in the experiments (DSIE), coin flip outcome on Monday (CFOM) and woken given one was sleeping in the experiment (WGSIE). I will draw it out (actually just write it out here) in terms of p(DSIE)*p(CFOM|DSIE)*p(WGSIE|DSIE,CFOM).

I would like to call DSIE and CFOM possible unknowns and WGSIE the possible known, then p(DSIE,CFOM) is a prior and p(WGSIE| DSIE,CFOM) is a data model or likelihood and Bayes theorem applies, without this labelling it’s just conditional probability which is logically the same thing.

Now we know p(DSIE=Mon) + p(DSIE=Tues) = 1 and p(DSIE=Tues) = ½ p(DSIE=Mon)

so p(DSIE=Mon)=2/3 and p(DSIE=Tues)=1/3.

Now P(CFOM=H|DSIE=Mon) = 1/2 , P(CFOM=T|DSIE=Mon) = 1/2 , P(CFOM=T|DSIE=Tues)=1.

P(WGSIE| DSIE=.,CFOM=.) is always equals to one.

Prior equals

P(DSIE=Mon ,CFOM=H) = 2/3 * ½ = 1/3

P(DSIE=Mon ,CFOM=T) = 2/3 * ½ = 1/3

P(DSIE=Tues ,CFOM=T) = 1/3 *1 = 1/3

So marginal prior for CFOM = 1/3 H and 2/3 T, and the posterior given you were woken while sleeping in the experiment – will be the same (as no learning occurs) – so you prior is 2/3 T.

OK – where did I go wrong? Do I need to review my probability theory?

-
I am having a hard time seeing how this helps resolve the paradox. To what prior distribution are you referring? (And please--this is not the place for bringing up the Monty Hall problem. That notorious situation always generates more discussion than insight.) – whuber Oct 26 '12 at 21:13
I have responded to the comment from @whuber . – phaneron Oct 30 '12 at 13:58

I just thought of a new way to explain my point, and what is wrong with the 1/2 answer. Run two versions of the experiment at the same time, using the same coin flip. One version is just like the original. In the other, three (or four - it doesn’t matter) volunteers are needed; each is assigned a different combination of Heads-or-Tails and Monday-or-Tuesday (the Heads+Tuesday combination is omitted if you use only three volunteers). Label them HM, HT, TM, and TT, respectively (possibly omitting HT).

If a volunteer in the second version is woken up this way, she knows she was equally likely to have been labeled HM, TM, or TT. In other words, the probability she was labeled HM, given that she is awake, is 1/3. Since the coin flip and day correspond to this assignment, she can trivially deduce that P(Heads|Awake)=1/3.

The volunteer in the first version could be woken more than once. But since "today" is only one of those two possible days, when she is awake she has exactly the same information as the awake volunteer in the second version. She knows that her current circumstances can correspond to the label applied to one, AND ONLY ONE, of other volunteers. That is, she can say to herself "either the volunteer labeled HM, or HT, or TT is also awake. Since each is equally likely, there is a 1/3 chance it is HM and so a 1/3 chance the coin landed tails."

The reason people make a mistake is that they confuse "is awake sometime during the experiment" with "is awake now." The 1/2 answer comes from the original SB saying to herself "either HM is the only other awake volunteer NOW, or TM and TT are BOTH awake SOMETIME DURING THE EXPERIMENT. Since each situation is equally likely, there is a 1/2 chance it is HM and so a 1/2 chance the coin landed tails." It is a mistake because only one other volunteer is awake now.

-
Each is not equally likely. Why would it be? Say the coin was weighted a million to one toward heads. You can't say three things are equally likely just because there are three of them. – Dax Fohl Sep 1 at 4:10

Rather than giving a statistically rigorous answer, I'd like to modify the question slightly in a way that might convince people whose intuition leads them to be halfers.

Some researchers want to put you to sleep. Depending on the secret toss of a fair coin, they will awaken you either once (Heads) or nine-hundred and ninety-nine times (Tails). After each awakening they will put you back to sleep with a drug that makes you forget that awakening.

When you are awakened, what degree of belief should you have that the outcome of the coin toss was Heads?

Following the same logic as before, there could be two camps -

• Halfers - the coin toss was fair, and SB knows this, so she should believe there is a one-half chance of heads.
• Thousanders - if the experiment was repeated many times, the coin toss would be heads only one in a thousand times, so she should believe that the chance of heads is one in a thousand.

I believe that some of the confusion from the question as originally worded arises simply because there isn't much difference between a half and a third. People naturally think of probabilities as somewhat fuzzy concepts (particularly when the probability is a degree-of-belief rather than a frequency) and it's difficult to intuit the difference between degrees of belief of a half and a third.

However, the difference between a half and one in a thousand is much more visceral. I claim that it will be intuitively obvious to more people that the answer to this problem is one in a thousand, rather than a half. I would be interested to see a "halfer" defend their argument using this version of the problem instead.

-
Chris, the defence is unchanged. For a more intuitive view of the halfer position, consider what you'd do in the following (rather horrible) modification of the experiment. Each time you wake, you must say "A" or "B". At the end of the experiment (1) if Heads and you had said "A" then you score 1 point; (2) if Tails and you always said "B" then you score 0.1 points; (3) otherwise you score nothing. Then a random number in the range 0 to 1 is picked: if it's higher than your points total, you are killed. Are you really so confident in Tails that you'd always say "B" ...? – Creosote Sep 5 at 9:17

Before SB goes to sleep she believes that the chance of the next coin flip being heads is 1/2. After she awakens, she believes that the chance that the most recent coin flip was heads is 1/3. Those events are not the same thing because there is not a one to one correspondence between awakening and coin flips.

-