I’ve been thinking about the Sleeping Beauty problem for years, and I still haven’t got a hill I’m prepared to die on. As soon as one answer starts to look secure, the other pops up and reminds you why you found the problem compelling in the first place.
That’s odd for a problem whose arithmetic is barely beyond fractions. You’ve got a fair coin, a woman who can’t remember her previous awakening, and a fairly straightforward question: what’s the probability that the coin came down heads?
My problem starts when I try to work out precisely what she’s being asked to count.
A fair coin and an unreliable calendar
On Sunday, Sleeping Beauty learns the rules of the experiment. A fair coin will be tossed when she goes to sleep. If it’s heads, she will be awakened and interviewed only on Monday. If it’s tails, she will be awakened and interviewed on both Monday and Tuesday.
After each interview, she goes back to sleep, and her memory of the awakening is wiped. The experiment ends on Wednesday, and she can tell the difference between that and the experimental interviews. During an interview, she knows the protocol but has no idea whether it’s Monday or Tuesday.
Each time, the researchers ask for her degree of belief that the coin came up heads. The two familiar answers are 1/2 and 1/3. The 1/2 camp are the halfers and the 1/3 camp are the thirders. Adam Elga defended the thirder position in his 2000 paper; David Lewis defended the halfer position in his reply the following year. [1, 2]
The case for leaving the coin alone
The argument for 1/2 starts with something very hard to dispute. Beauty knows the coin is fair and she knows she’ll be woken up either way. Heads leads to an interview. Tails leads to an interview too.
If the event was certain under either outcome, then why should its occurrence have changed her assessment of those outcomes? Nothing in the room tells you anything about the coin, and nothing in her memory tells you which day it was. The coin hasn’t become biased whilst she’s been asleep.
This is the appeal of the halfing. It appears to respect the difference between uncertainty about the coin and uncertainty about her position on the schedule.
But there’s a question lurking beneath that reassurance. Does the fact that she’ll wake up at least once in the course of the day actually encapsulate what she knows when she’s actually sitting in this interview?
The case for counting interviews
Imagine doing the experiment 1,000 times. You’d expect about 500 heads and 500 tails. The 500 heads would give you about 500 interviews, and the 500 tails would give you about 1,000.
If every interview generates a transcript, then we’re looking at around 1,500 transcripts in total. Around 500 of those are head transcripts. If you draw uniformly from that whole collection, the probability of drawing a head transcript is about 1 in 3.
The corresponding long-run fraction is:
\displaystyle \frac{\text{heads interviews}}{\text{all interviews}}=\frac{500}{500+1{,}000}=\frac{1}{3}
Now imagine a different procedure. Pick an experiment uniformly at random, then pick one of its interviews uniformly at random. The probability of heads is 1/2 because the first step is an equal selection over the experiments. A tails experiment doesn’t get any extra weighting simply because it has two interviews.
Both calculations are correct. They represent different selection procedures.
The thirder’s substantive claim is that Beauty should reason as if she were at an interview in the first collection. The halfer asks why her current experience should be treated as a uniform draw from all the interviews. The lab never in fact makes that draw.
That final step is where counting becomes philosophy. The transcript calculation gives you a frequency. Establishing that frequency in terms of Beauty’s current degree of belief requires an argument.
Three possibilities do not mean three equal chances
At an interview, Beauty considers three options: heads on Monday, tails on Monday, and tails on Tuesday. Both sides can, of course, represent her ambivalence in terms of those same three options. The disagreement is about the weights.
| Beauty’s possible situation | Standard thirder weight | Lewis’s halfer weight |
|---|---|---|
| Heads, Monday | 1/3 | 1/2 |
| Tails, Monday | 1/3 | 1/4 |
| Tails, Tuesday | 1/3 | 1/4 |
| Total | 1 | 1 |
The thirder assigns equal weight to each interview situation. Lewis maintains a total probability of 1/2 for tails and splits it equally between the two indistinguishable tails awakenings. [1, 2]
It’s not sufficient to simply list three possibilities to establish that each has probability 1/3. A list that includes “heads,” “tails before lunch,” and “tails after lunch” wouldn’t make an ordinary fair coin any more likely to land on tails. We need to defend the significance of Beauty’s temporal uncertainty.
This is also why describing the dispute as a choice of sample space is incomplete. The same three rows can support either answer. The difficulty lies in deciding how to assign probability across them.
What happens when someone says it’s Monday?
Suppose the researchers reliably tell Beauty the day at each interview. At this interview she is told it’s Monday. No further selection is involved in the decision to reveal the day.
Under the standard thirder account, the Tuesday row goes away. The two surviving rows were of equal weight, so her probability of heads is:
P(H\mid\text{Monday}) =\frac{1/3}{1/3+1/3} =\frac{1}{2}Under Lewis’s halfer account, the same conditioning gives:
P(H\mid\text{Monday}) =\frac{1/2}{1/2+1/4} =\frac{2}{3}So the standard Thirder goes from 1/3 to 1/2, the Lewisian Halfer from 1/2 to 2/3. That’s the opposite of what happens in the original exchange. [1, 2]
The thirder result is the one that appeals to me. Every experiment has exactly one Monday interview regardless of what the coin does. Once you exclude Tuesday, the reason for giving tails the extra weight vanishes. Lewis’s result is perfectly consistent with his initial weights, but it forces you to accept that learning that it’s Monday makes heads more likely than tails.
Neither side ends up with all the attractive intuitions. That’s part of what makes the problem so persistent.
When the argument puts money on the table
Betting makes the consequences tangible. Suppose Beauty adopts a policy of taking an even money bet on heads at each interview (paying her $1 if heads and losing $1 if tails). All payments are made at the end of the experiment, so her balance at that point gives her no information during the experiment.
Heads: one winning bet, total gain $1. Tails: two losing bets, total loss $2. That policy has an expected return of minus 50 cents prior to the experiment.
For the policy of buying a ticket at every interview (a ticket that pays $1 if it’s heads and nothing if it’s tails), the break-even price is 1/3 of a dollar. The policy is to buy one ticket if the coin toss is heads and two if it’s tails. The multiplicity is important even though we’re talking about the same toss.
We can go even further. Suppose Beauty also accepts a Sunday wager that pays $1.50 on tails and costs $1.50 on heads. Combine that with the $1 even money heads wager at every interview:
| Coin result | Sunday wager | Interview wagers combined | Total |
|---|---|---|---|
| Heads | −$1.50 | +$1.00 | −$0.50 |
| Tails | +$1.50 | −$2.00 | −$0.50 |
The loss is now assured. This is a Dutch book (a set of bets which must lose under every possible outcome of the coin toss). Hitchcock developed a Dutch book argument in favor of thirding. [3]
The table presents a problem for a policy that accepts those wagers. Whether it also rules out a degree of belief of 1/2 depends on the connection between belief and betting. A sophisticated halfer can, of course, reject the interview wagers by appealing to the fact that the same decision will face them in the tails branch. The literature is very much aware of the distinction between credences and acceptable betting odds. [4]
Calling the bets correlated makes no difference to the loss. Likewise, showing the loss in no way establishes that every halfer must accept the package.
Why I cannot simply declare both sides right
There’s a very tempting way to finish the discussion: halfers do the experiments, thirders do the interviews, and we’re all happy. It accounts for why both 1/2 and 1/3 appear in the correct calculations. It leaves Beauty’s question unanswered.
She’s asked to assess one proposition only (that the coin in her experiment came up heads). The philosophical issue is which of the two lines of reasoning she should adopt given all that she knows at that point. Describing the two sampling procedures doesn’t establish that either is a complete model of her situation.
I’m therefore not inclined to say the problem is simply underspecified (as if we’d all been there all along and just needed one more instruction to clarify it). A sampling rule makes a particular sampling question precise. Whether Beauty in fact requires such a rule or whether rational belief alone is sufficient to determine the appropriate way to deal with her ambiguity is itself in dispute.
Where I remain
I keep coming back to the difference between the experiment as a whole and Beauty’s experience within it. The external observer can record one coin outcome per experiment. Beauty has to make her assessment at a moment when she cannot tell where she is in the sequence.
The extra interview changes the distribution of the experiences. It leaves the mechanism that produces the coin outcome completely untouched. I can understand both statements and calculate the consequences of each. My uncertainty is about the bridge between them.
That’s why I’m still in the lurch. The halfer defends a very plausible principle about a fair coin and about guaranteed evidence. The thirder offers a very plausible account of the interviews that Beauty actually undergoes. Each attacks something I would normally be prepared to accept.
After all these years of thinking about Sleeping Beauty, I’m more certain than ever about where the uncertainty lies. I know perfectly well why half the experiments produce heads and why only a third of the interviews are from those experiments. What I haven’t settled is which fact should determine the answer from inside the room.
The coin is fair. I’m still not sure how to be fair to Beauty.
References
1. Elga, A. (2000). Self-locating belief and the Sleeping Beauty problem. Analysis, 60(2), 143–147.
2. Lewis, D. (2001). Sleeping Beauty: Reply to Elga. Analysis, 61(3), 171–176.
3. Hitchcock, C. (2004). Beauty and the bets. Synthese, 139, 405–420.
4. Milano, S. (2022; first published online 2020). Bayesian Beauty. Erkenntnis, 87, 657–676.
