The Sleeping Beauty Problem

What each side is actually defending

It helps to see that this is not really an argument about a coin. Each side is protecting a different principle, and the two principles collide only here.

Thirders protect the connection between credence and the structure of possibilities. If there are three ways things could be from your point of view, and two of them are tails-ways, then that has to show up somewhere in your beliefs. Beauty is not asking "how do coins behave?" — she knows how coins behave. She is asking "where am I?", and the answer to that question depends on how many places there are to be.

Halfers protect the connection between credence and evidence. Belief is not supposed to drift. If your confidence in a proposition changes, there must be something you learned, and you must be able to say what it was. "I woke up, as I knew I would" is not an answer. Allow credences to shift on the mere fact of being somewhere, and it becomes very hard to say what stops them shifting elsewhere.

Both principles are ones we would ordinarily insist on. That is the mark of a genuine open problem rather than a puzzle with a hidden trick.

The Dutch books

The classic way to break a tie in formal epistemology is to show that one side's credences would let a bookmaker take money from them with certainty. Both sides have produced such an argument. Both arguments work. That should have been a warning.

Against the halfer (Christopher Hitchcock, 2004). Suppose Beauty's credence in heads on waking is 1/2. Then two transactions are, by her own lights, exactly fair, and she must be willing to take both.

• On Sunday, she sells a bet: she takes £1.50 now and pays out £3 if heads. Everyone agrees her Sunday credence in heads is 1/2, so this is fair.
• At each awakening, she buys a bet: she pays £1 and receives £2 if heads. Fair at credence 1/2, which is what the halfer says she holds.

	Sunday bet (sold)	Awakening bets (bought)	Total
Heads	+£1.50 − £3 = −£1.50	1 bet: −£1 + £2 = +£1.00	−£0.50
Tails	+£1.50 − £0 = +£1.50	2 bets: −£2 + £0 = −£2.00	−£0.50

She loses fifty pence whatever the coin does. The reason is visible in the table: her awakening bets are billed twice on tails and once on heads, so pricing them at even money is a mistake. Repeat the calculation with a thirder, who will pay at most £2/3 for the awakening bet, and the sure loss disappears — she comes out 16.7 pence down on heads and 16.7 pence up on tails, which is a gamble, not a fleecing.

Against the thirder (sketched by Lewis in 2001, developed by others). Thirding requires Beauty's credence in heads to be 1/2 on Sunday, 1/3 on waking, and 1/2 again on Wednesday when the experiment is over and she has still not been told the result. Two shifts, and neither is triggered by any evidence. A bookie who transacts on either side of a predictable, unmotivated credence shift can arrange a sure loss — this is the standard diachronic Dutch book, the same argument used to justify conditionalisation in the first place. If credences may swing back and forth without evidence, the whole apparatus that made Dutch books compelling comes apart.

Why neither wins. Rachael Briggs argued in 2010 that these two arguments are not really competing on the same ground. The first assumes bets are settled per awakening; the second assumes an agent whose transactions are counted per run. Once Beauty's decisions can be duplicated without her knowledge, the usual correspondence between credence and fair betting price breaks, and different decision theories pull the two apart. The betting arguments turn out to be tracking the protocol, not the belief. This is the best current reason to think the problem cannot be settled by pragmatic means, and it is one of the few points of broad agreement.

The double halfer

A third position deserves naming. Double halfers — including Nick Bostrom's "hybrid" model and Christopher Meacham's 2008 treatment — hold that Beauty's credence in heads is 1/2 on waking and remains 1/2 when she learns it is Monday. This dodges the halfer's embarrassment about the future fair coin.

The price is steep: 1/2 is not the conditional probability that the halfer's own distribution assigns to heads given Monday, so double halfers must reject conditionalisation on self-locating information outright. They accept that cheerfully, on the grounds that no satisfactory rule for updating de se belief exists anyway. Critics reply that abandoning conditionalisation is a very large concession to avoid a small oddity.

There is also a school — represented by parts of Piccione and Rubinstein's original discussion — holding that the question is defective: that credence is not well defined for an agent who may be asked the same question twice with no memory of the first, and that we should specify a decision problem instead. On this reading the paradox is a signal that "degree of belief" has been pushed past the domain where it has clear content.

Where the field stands

There is no consensus and no trend towards one. Surveys of professional philosophers — including the large PhilPapers survey of 2020, which included the question directly — find thirders somewhat ahead of halfers, with a substantial fraction declining to take a side and neither position anywhere near a majority. Papers defending each view continue to appear.

This entry therefore does not have a resolution section, because there is no resolution. What can be said is what the disagreement is about: how to extend Bayesian updating to beliefs about where and when you are. That is a genuine gap in the theory, not a failure of anyone's arithmetic.

1990. Arnold Zuboff published "One Self: The Logic of Experience" in Inquiry, containing a version of the problem in the course of a broader argument about personal identity. It attracted little attention at the time.

1997. Michele Piccione and Ariel Rubinstein published "On the Interpretation of Decision Problems with Imperfect Recall" in Games and Economic Behavior. Their absent-minded driver problem — a driver who cannot remember whether he has already passed one junction — has the same formal structure, and they discuss a scenario equivalent to Sleeping Beauty. They arrived at it from decision theory rather than epistemology, and the two literatures took some years to notice each other.

2000. Adam Elga, then completing a doctorate at MIT, published a three-page note in Analysis. He gave the problem its name, credited it to Robert Stalnaker, who had it from Zuboff, and argued for 1/3. The paper is remarkable for how much subsequent literature it generated per page.

2001. David Lewis replied in the same journal with "Sleeping Beauty: reply to Elga", defending 1/2 and accepting the consequence about Monday. It was among the last things he published; he died in October of that year. That the two canonical statements of the two positions are both under six pages, in the same journal, one year apart, is part of why the problem became a standard teaching case.

2004–2010. The betting arguments were worked out — Hitchcock's Dutch book against the halfer in Synthese in 2004, various diachronic books against the thirder — and then largely defused by Rachael Briggs's demonstration in 2010 that they encode decision-theoretic commitments rather than epistemic ones.

2008 onwards. Christopher Meacham and others developed the double-halfer position; Michael Titelbaum introduced Technicolor Beauty and pressed Lewis's "no new evidence" premise; the connection to anthropic reasoning, already implicit in Bostrom's 2002 book, became the main reason the problem is cited outside philosophy.

Now. It remains open. It appears in graduate epistemology syllabuses as the standard illustration that Bayesianism, which is otherwise in excellent order, has no agreed account of self-locating belief.

Most paradoxes on this site are surprising and then explained. This one is surprising and then not explained, and that is what it contributes.

It found a real gap in Bayesian epistemology. Bayesian updating is the most successful formal theory of rational belief we have. The Sleeping Beauty problem shows that it is silent on a specific and unavoidable class of cases: beliefs about which time it is, which person you are, which of several identical situations you are in. Before 2000 this was not widely recognised as a gap. Nobody has closed it since.

It disciplined anthropic reasoning. Arguments from "the fact that we exist" were common and loose — the Doomsday argument, fine-tuning arguments, the simulation argument, Boltzmann brains. Sleeping Beauty turns the loose intuitions into two named principles, SSA and SIA, with sharply different consequences and a documented cost each. That is a substantial improvement in a field that badly needed one, and it means that arguments in that space can now be located rather than merely asserted.

It separated belief from betting. The debate established, more clearly than any other case, that "the odds at which you should bet" and "how confident you should be" are not always the same quantity. Whenever an agent's decisions can be duplicated, or its memory manipulated, or its existence made contingent on the outcome, the usual bridge between them fails. That matters wherever agents are copied — which is now an engineering question and not only a philosophical one.

It is an unusually clean example of an unresolved problem. The setup takes a paragraph, the two arguments take a paragraph each, and there is no consensus. For anyone learning how philosophical disagreement actually works, that combination is rare and valuable.

Cosmology and the measure problem. In eternal inflation, essentially every outcome occurs somewhere, and predictions require a rule for weighting observers across an infinite multiverse. That is the Sleeping Beauty question with the numbers taken to infinity, and cosmologists working on the measure problem cite the halfer–thirder debate directly. Whether typical observers are ordinary or are Boltzmann brains depends on which weighting you adopt.

Quantum mechanics. In Everettian interpretations, a measurement branches the observer, and the observer's uncertainty just after branching is self-locating — "which branch am I in?" — rather than uncertainty about the world. Attempts to derive the Born rule from decision theory, following David Deutsch and David Wallace, need an account of rational belief under branching. Sleeping Beauty is the sandbox in which the required principles are argued about.

AI systems that are copied. A model that can be instantiated many times, with identical context and no memory across instances, is in Beauty's position by construction. If such a system is to reason about its own situation — how many copies are running, whether this is a test or deployment, what its observations imply about the world — it needs exactly the rule this problem shows we lack. This has moved from a curiosity to a design question, and papers on AI self-location cite Elga and Lewis directly.

Simulation and existential-risk arguments. Nick Bostrom's simulation argument counts observers across simulated and unsimulated worlds; forecasts of catastrophic risk that appeal to our position in history do the same. Both are sensitive to the SSA/SIA choice in ways that can flip their conclusions entirely, so any confident use of them is implicitly taking a side in a debate that is not settled.

Experimental design under memory manipulation. The problem also has a modest practical descendant. When respondents may be sampled more than once, or when a trial's stopping rule depends on outcomes, the frequency of an event among observations diverges from its frequency among runs. Deciding which denominator a quantity refers to is the ordinary, unglamorous version of the question Beauty is being asked.

Setting it up formally

Let $\mathcal{W} = {H, T}$ be the uncentered worlds and let the centered possibilities available to Beauty on waking be

$$\Omega = {,H_1,; T_1,; T_2,}$$

where the subscript is the day: $H_1$ is "heads, and it is Monday", $T_1$ is "tails, and it is Monday", $T_2$ is "tails, and it is Tuesday". Write $\mathrm{Mon} = H_1 \cup T_1$ and $H = {H_1}$, $T = {T_1, T_2}$.

Beauty's Sunday credence function $P_0$ is defined over $\mathcal{W}$ with $P_0(H) = 1/2$. Her waking credence $P$ is defined over $\Omega$. The problem is that no uncontroversial rule takes you from $P_0$ to $P$: they are not even functions on the same space. Everything below is an argument about what that rule should be.

Elga's derivation

Two premises.

(E1) Restricted indifference. $T_1$ and $T_2$ are subjectively indistinguishable and correspond to the same uncentered world, so

$$P(T_1 \mid T) = P(T_2 \mid T) = \tfrac{1}{2} \quad\Longrightarrow\quad P(T_1) = P(T_2)$$

(E2) The Principal Principle, applied on Monday. In the variant where the coin is tossed on Monday night, conditioning on $\mathrm{Mon}$ leaves a fair future toss with known chance $1/2$ and no inadmissible evidence, so

$$P(H_1 \mid \mathrm{Mon}) = \tfrac{1}{2} \quad\Longrightarrow\quad P(H_1) = P(T_1)$$

Together, $P(H_1) = P(T_1) = P(T_2)$, and since they exhaust $\Omega$,

$$P(H) = P(H_1) = \tfrac{1}{3}$$

Elga's own presentation runs the argument in the other direction, deriving the Sunday-to-Monday shift as a consequence rather than a cost, and he is explicit that the conclusion requires abandoning the assumption that credence in uncentered propositions can only change by conditionalisation.

Lewis's derivation

One premise.

(L1) No new evidence. Let $E$ be Beauty's total evidence on waking. Lewis argues $P_0(E) = 1$ — on Sunday she was already certain she would be awake and questioned. Conditionalisation on a probability-one proposition is the identity map, so

$$P(H) = P_0(H) = \tfrac{1}{2}$$

Lewis accepts (E1), giving

$$P(H_1) = \tfrac{1}{2}, \qquad P(T_1) = P(T_2) = \tfrac{1}{4}$$

and therefore

$$P(H_1 \mid \mathrm{Mon}) = \frac{1/2}{1/2 + 1/4} = \frac{2}{3}$$

Lewis rejects (E2) explicitly. His defence is the admissibility clause built into the Principal Principle: chance constrains credence only for agents whose evidence is admissible with respect to the chance event, and Beauty's self-locating evidence — that it is Monday — is, in this contrived setup, correlated with the toss outcome and therefore inadmissible. Whether that use of admissibility is legitimate or ad hoc is one of the live sub-debates.

The three positions in one table

	$P(H_1)$	$P(T_1)$	$P(T_2)$	$P(H)$	$P(H \mid \mathrm{Mon})$	Gives up
Thirder (Elga)	1/3	1/3	1/3	1/3	1/2	Credence changes only on evidence
Halfer (Lewis)	1/2	1/4	1/4	1/2	2/3	Principal Principle for a future fair toss
Double halfer	1/2	1/4	1/4	1/2	1/2	Conditionalisation itself

Every row gives up something a Bayesian would rather keep. That is the content of the paradox, stated as compactly as it can be.

Generalising

Let tails produce $n$ indistinguishable awakenings and heads produce one, with a coin of chance $p$ for heads. The thirder's answer is obtained by weighting each world by its number of awakenings:

$$P_{1/3}(H) ;=; \frac{p \cdot 1}{p \cdot 1 + (1-p), n}$$

The halfer's answer is $P_{1/2}(H) = p$, independent of $n$. With $p = 1/2$ these are $1/(1+n)$ and $1/2$. At $n = 2$ they are 1/3 and 1/2; at $n = 10^6$ they are about $10^{-6}$ and 1/2.

The corresponding conditional credences on learning it is the first day are $p$ for the thirder — always, for every $n$ — and

$$P_{1/2}(H \mid \text{day } 1) ;=; \frac{p}{p + (1-p)/n} ;=; \frac{np}{np + 1 - p}$$

for the halfer, which tends to 1 as $n$ grows. The two costs scale in opposite directions and neither is bounded.

Anthropic reasoning

The reason this matters far outside the puzzle is that the two positions are instances of two general principles for reasoning about your own existence, catalogued by Nick Bostrom in Anthropic Bias (2002).

The Self-Sampling Assumption (SSA). Reason as if you are a random sample from the observers in your reference class within your world. Applied to Beauty: the worlds keep their chance-given weights $1/2$ and $1/2$, and within the tails world she is a random sample from two awakenings. That is exactly Lewis's distribution. SSA yields halfing.

The Self-Indication Assumption (SIA). Reason as if you are a random sample from all possible observers, which weights each world in proportion to how many observers it contains. The tails world contains twice as many awakenings, so it gets twice the weight. That is exactly Elga's distribution. SIA yields thirding.

Now the Doomsday argument, in the Carter–Leslie form. Compare a hypothesis on which humanity ends soon, with perhaps $10^{11}$ humans in total, against one on which it spreads and lasts, with perhaps $10^{17}$. Under SSA, you should treat your birth rank — roughly the 100-billionth human — as a random draw from the humans in your world. A rank that low is enormously more probable under the small total than under the large one, so the likelihood ratio shifts your credence sharply towards doom soon, whatever your prior was.

Under SIA, the long-lived hypothesis is up-weighted at the start by its far greater number of observers — by a factor of about $10^6$ in this example. That prior boost is precisely the same factor as the likelihood penalty SSA imposes. Ken Olum showed in 2002 that the two cancel exactly, leaving your original prior untouched. If you are a thirder, the Doomsday argument does not go through. If you are a halfer, it does.

This is not an incidental connection. It is the same computation. A three-page note about a fictional coin toss turns out to determine whether an argument about the extinction of the species is valid.

SIA is not free either. Bostrom's Presumptuous Philosopher objection: suppose physics has narrowed cosmology to two theories, equally supported by all the evidence, one predicting a universe with $10^{12}$ observers and one predicting $10^{18}$. SIA instructs you to be a trillion times more confident in the second before any experiment is run, and to tell the experimentalists not to bother. Many find that as unacceptable as the Doomsday conclusion. Whichever way Beauty goes, somebody's cosmology gets strange.

Related territory

The same machinery is doing work in several live debates: whether the Everettian branching of quantum mechanics generates self-locating uncertainty that can ground the Born rule; how to reason about being a Boltzmann brain in an eternally inflating cosmology; and how to assess the simulation argument, which is an observer-counting argument of exactly this shape. In each case the answer depends on a rule for self-locating belief that the Sleeping Beauty problem shows we do not have.