The Sleeping Beauty Problem

Imagine you are put to sleep on Sunday. A fair coin is flipped.

• Heads: you are woken once, on Monday, then the experiment ends.
• Tails: you are woken on Monday, given a memory-erasing drug that puts you back to sleep, then woken again on Tuesday.

Each time you wake up, you have no idea which day it is or whether you have been woken before. You are asked: "What is your credence — your degree of belief — that the coin landed heads?"

The coin is fair, so before the experiment you'd say 1/2. But now you are inside the experiment, awake, and you know that awakenings are not equally distributed between heads and tails. Does that change anything?

Two camps have argued fiercely since philosopher David Elga posed the problem in 2000:

• Thirders say the answer is $\frac{1}{3}$, because only one of three equally likely awakenings is a heads-waking.
• Halfers say the answer remains $\frac{1}{2}$, because waking up gives you no new information about the coin.

Both positions are defensible. Neither is universally accepted. The disagreement is not about arithmetic — it is about what probability means for an agent who does not know where in time she is.

Each trial flips a fair coin. Heads produces one awakening; tails produces two awakenings (with memory wiped between them). The bar chart tracks the cumulative fraction of awakenings that are heads-awakenings.

<!-- {{c_layout_comment}} -->
<div class="controls">
  <button id="run1" type="button">{{btn_run1}}</button>
  <button id="run100" type="button">{{btn_run100}}</button>
  <button id="reset" type="button" class="ghost">{{btn_reset}}</button>
</div>
<div class="stats-row">
  <div class="stat-box">
    <div class="stat-label">{{label_trials}}</div>
    <div class="stat-value" id="trials">0</div>
  </div>
  <div class="stat-box">
    <div class="stat-label">{{label_awakenings}}</div>
    <div class="stat-value" id="awakenings">0</div>
  </div>
  <div class="stat-box">
    <div class="stat-label">{{label_heads_wakes}}</div>
    <div class="stat-value" id="heads-wakes">0</div>
  </div>
  <div class="stat-box highlight">
    <div class="stat-label">{{label_freq}}</div>
    <div class="stat-value" id="freq">—</div>
  </div>
</div>
<!-- {{c_bar_comment}} -->
<div class="bar-area">
  <div class="bar-track">
    <div id="bar-fill" class="bar-fill" style="width:0%"></div>
    <div class="bar-marker half" title="1/2">½</div>
    <div class="bar-marker third" title="1/3">⅓</div>
  </div>
  <div class="bar-labels">
    <span>0</span>
    <span>{{label_one_third}}</span>
    <span>{{label_one_half}}</span>
    <span>1</span>
  </div>
</div>
<div class="narrative" id="narrative">{{msg_start}}</div>

/* {{c_css_comment}} */
* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; padding: .5rem; }
.controls { display: flex; gap: .5rem; flex-wrap: wrap; margin-bottom: .8rem; }
button { font: 600 14px system-ui, sans-serif; padding: .4rem .85rem;
         border: 1px solid #1d3557; background: #1d3557; color: #fff;
         border-radius: 8px; cursor: pointer; }
button.ghost { background: #fff; color: #1d3557; }
.stats-row { display: flex; gap: .5rem; flex-wrap: wrap; margin-bottom: .9rem; }
.stat-box { flex: 1 1 70px; background: #f0f4f8; border: 1px solid #d0dbe6;
            border-radius: 8px; padding: .4rem .6rem; text-align: center; }
.stat-box.highlight { background: #e8f4ea; border-color: #a3c9a8; }
.stat-label { font-size: .72rem; color: #555; text-transform: uppercase; letter-spacing: .04em; }
.stat-value { font: 700 1.3rem ui-monospace, monospace; color: #1d3557; margin-top: .1rem; }
.stat-box.highlight .stat-value { color: #1a7d33; }
/* {{c_bar_css_comment}} */
.bar-area { margin-bottom: .7rem; }
.bar-track { position: relative; height: 32px; background: #dde5ed; border-radius: 16px; overflow: hidden; }
.bar-fill { height: 100%; background: linear-gradient(90deg, #1d3557 0%, #457b9d 100%);
            border-radius: 16px; transition: width .25s ease; }
.bar-marker { position: absolute; top: 0; height: 100%; width: 2px; background: #e63946; opacity: .7; }
.bar-marker::after { content: attr(title); position: absolute; top: -1.4rem; left: -1ch;
                      font-size: .7rem; color: #e63946; font-weight: 700; }
.bar-marker.half { left: 50%; }
.bar-marker.third { left: 33.33%; }
.bar-labels { display: flex; justify-content: space-between;
              font-size: .7rem; color: #666; margin-top: .2rem; padding: 0 2px; }
.narrative { font-size: .9rem; color: #444; min-height: 1.5em; line-height: 1.5; }
.narrative.heads { color: #1a7d33; font-weight: 600; }
.narrative.tails { color: #c92f3c; font-weight: 600; }

// Code not found

Watch where the bar settles. The long-run frequency of "I'm awake and it was heads" converges to 1/3, not 1/2 — because two out of every three awakenings (on average) happen on tails trials. Thirders say that frequency is the correct credence. Halfers say frequency across awakenings is the wrong thing to measure: the coin is still fair, so the probability of heads for this flip is still 1/2.

The disagreement runs deeper than a classroom coin-flip.

The thirder argument (Elga, 2000) uses a Principal Principle-style reasoning: if you are equally likely to be in any of the three awakenings (Mon-Heads, Mon-Tails, Tue-Tails), then by symmetry each gets probability $\frac{1}{3}$. Heads corresponds to exactly one of them, so $P(\text{Heads}) = \frac{1}{3}$.

The halfer argument (Lewis, 2001) insists that waking up is not new information — you already knew you would wake up. Your prior was $\frac{1}{2}$, and nothing you observe upon waking is evidence that should update it. So $P(\text{Heads}) = \frac{1}{2}$.

The real fault line is about self-locating belief: can a rational agent assign different probabilities to the same objective event depending on how many times she is woken? Thirders say yes — probability can depend on which awakening you are. Halfers say no — probability tracks facts about the world, not about your position in a schedule.

A deeper version of the puzzle involves Bayesian inference: thirders update on the fact of being awake; halfers argue that "being awake" has probability 1 in every scenario and carries no Bayesian weight. Both are applying Bayes' rule correctly — to different sample spaces.

Neither camp has a knock-down proof. The problem remains open in philosophy of probability: no consensus exists as of 2025.

"I'm awake — what should I believe about the world?" sounds like an idle puzzle. It isn't:

• Cosmology and the anthropic principle: when reasoning about how many observers exist in a universe, do you weight by observer-moments or by universe histories? Thirder-style reasoning says your beliefs should shift because you exist at all; halfer-style says existence gives no new evidence.
• AI and decision theory: an agent that is copied, paused, or run on multiple instances faces the same question. Should a copied AI think it is twice as likely to be the original? The Nash equilibrium of multi-instance agents depends sensitively on which stance they take.
• Doomsday argument: if you reason about how likely you are to be "early" in humanity's history, you are doing self-locating probability. The same halfer/thirder split applies.
• Simulation hypotheses: reasoning about whether you are in a simulation — where simulated instances might vastly outnumber originals — is structurally identical to the Sleeping Beauty setup.

Whether you are a halfer or a thirder, the puzzle forces you to confront what rational belief means when the subject of the belief is also part of the experimental design.

The Sleeping Beauty problem is deceptively simple: one fair coin, one forgetful princess, and a question that has occupied philosophers for a quarter century. The arithmetic is trivial; the disagreement is genuine.

Thirders count awakenings and find $\frac{1}{3}$. Halfers count coin flips and find $\frac{1}{2}$. Both are right given their assumptions about what probability means. The puzzle is a mirror held up to those assumptions.

Long after the princess wakes for the last time, the debate continues — not because anyone made an error, but because the concept of rational credence for a self-locating agent is not yet fully settled. That is what makes it one of the most entertaining open problems in philosophy.