Fun Stuff / Logics

The Surprise Exam: What Exactly Did the Teacher Announce?

A teacher promises a surprise exam next week. The students rule out Friday, then Thursday, and eventually every day. On Wednesday, the teacher gives the exam. Where did the reasoning go wrong?

Possible worlds · Public announcements · Fixed points · The topology of surprise

1. The puzzle and its conditional variation

The Surprise Exam Paradox

James Simpson opens A new epistemic solution to the Surprise Exam Paradox (2025, pp. 1–2) with the familiar story:

A trustworthy teacher tells her students that this week’s only exam will be a surprise exam. A surprise exam is an exam such that the students won’t have a justified belief the morning the exam occurs that they’ll be having an exam that day.

A clever student argues that the teacher cannot keep this promise. Suppose no exam has occurred by Friday morning. Since an exam must occur during the week, it must occur on Friday, so the student will expect it. Friday therefore cannot be the date of a surprise exam.

Now suppose no exam has occurred by Thursday morning. Friday has already been ruled out, so the exam must occur on Thursday. But then Thursday’s exam will also be expected. Repeat the same argument for Wednesday, Tuesday, and Monday: no day seems capable of accommodating the promised exam.

Yet a teacher appears perfectly capable of announcing a surprise exam and then giving one on Wednesday, when the students do not expect it. The puzzle is the conflict between the apparently valid backward argument and this apparently possible outcome.

The Conditionally Expected Exam Paradox

The variation changes what the student must expect. Simpson states it as follows (§2.1, p. 6):

The teacher announces that this week’s only exam won’t be a conditionally expected exam. An exam is conditionally expected iff, on the morning of the exam, a student has a justified belief that if there is an exam at all this week, then it is today.

The relevant judgment is conditional: if there is an exam this week, it must be today.

Friday is again ruled out: if no exam has occurred before Friday, any exam that occurs during the week must occur on Friday. This conditional remains true even if there might be no exam. Once Friday is excluded, an exam on Thursday would also be conditionally expected if none had occurred earlier. The same backward argument eliminates Wednesday, Tuesday, and Monday. Yet an exam on Thursday seems capable of being conditionally unexpected when the student still regards Friday as possible.

Simpson also discusses Williamson’s version (p. 7), in which the teacher makes only the conditional promise: if an exam occurs, it will not be conditionally expected. This promise can be fulfilled by giving no exam.

The quotations use Simpson’s language of justified belief. Our formal model below first studies deductive knowledge; Section 6 returns to his treatment of justification and fallible testimony.

Three choices hidden in the stories

Understanding what the backward calculation establishes requires separating three questions:

Is an exam guaranteed, or is no exam also possible?
Does surprise mean not knowing that an exam is today, or not knowing that an exam, if there is one, must be today?
Is the announcement evaluated before it is learned, or must it remain true after its own consequences are understood?

These choices are independent. In particular, a conditionally expected exam changes the definition of surprise; a self-referential announcement changes the information against which surprise is evaluated. One formal model can keep track of all three.

2. Possible worlds and morning information

There is at most one exam during a five-day week. Let

\[\Omega=\{0,1,2,3,4,5\}, \qquad G=\{1,2,3,4,5\}.\]

State \(i\in G\) means the exam occurs on day \(i\); state \(0\) means no exam occurs. The event \(G\) says that an exam occurs sometime during the week. Write \(E_i=\{i\}\) for an exam on day \(i\).

Let \(A\subseteq\Omega\) be the possibilities the students retain before observing the passage of days. On the morning of day \(i\), provided no exam has yet occurred, their information is

\[I_i(A)=A\cap\bigl(\{0\}\cup\{i,i+1,\ldots,5\}\bigr).\]

The no-exam state survives every morning observation.

For any event \(H\subseteq\Omega\), define knowledge at this history by

\[K_i^A H \quad\Longleftrightarrow\quad I_i(A)\subseteq H.\]

We use this definition at possible histories, where \(I_i(A)\) is nonempty. Students know a proposition when it holds at every remaining possibility.

Ordinary surprise

At an actual exam state \(i\in A\cap G\), ordinary surprise means

\[\neg K_i^A E_i.\]

Equivalently,

\[\boxed{\text{Ordinary surprise at }i \iff 0\in A\ \text{or}\ \exists j\in A\cap G:\ j>i.}\]

Either no exam is still possible, or a later exam is still possible. Either alternative prevents certainty that the exam is today.

Conditional surprise

An exam is conditionally expected when the students know that, if an exam occurs at all, it occurs today:

\[K_i^A(G\Rightarrow E_i).\]

Here \(G\Rightarrow E_i\) is the material conditional, the event \(\{0,i\}\). Conditional surprise is its negation:

\[\boxed{\neg K_i^A(G\Rightarrow E_i) \iff \exists j\in A\cap G:\ j>i.}\]

The required alternative is a later exam date, since the conditional is true at the no-exam state.

For example, on Friday morning with \(A=\{0,5\}\), the students do not know that an exam will occur. Yet they do know that, if one occurs this week, it must be today. A Friday exam would be ordinarily surprising but conditionally expected. On Thursday with \(A=\{0,4,5\}\), an exam that morning would satisfy both definitions of surprise because Friday remains possible.

When an exam is known to be guaranteed, \(A\subseteq G\), the two definitions coincide.

3. One elimination framework

First consider a teacher who promises that any exam given will be surprising. The two definitions produce the operators

\[\begin{aligned} T_O(A)&=(A\cap\{0\})\\ &\quad\cup\{i\in A\cap G:0\in A\ \text{or}\ \exists j\in A\cap G,\ j>i\},\\[4pt] T_C(A)&=(A\cap\{0\})\\ &\quad\cup\{i\in A\cap G:\exists j\in A\cap G,\ j>i\}. \end{aligned}\]

The subscripts indicate ordinary and conditional surprise. Both retain state \(0\) whenever it is available: a promise about what happens if an exam occurs is vacuously satisfied when there is none.

If the teacher also guarantees an exam, intersect the output with \(G\):

\[F_O(A)=G\cap T_O(A), \qquad F_C(A)=G\cap T_C(A).\]

Each operator removes possibilities; each is monotone under set inclusion. Adding possible worlds preserves existing alternatives and can supply new ones that make an exam surprising.

These operators describe the content of the announcement relative to \(A\). Whether they should be applied once or iterated is a separate question.

4. Interpretation A: evaluate once, then announce

Suppose an exam is already known to occur, so initially \(A_0=G\). The teacher evaluates surprise using these initial possibilities and announces that the actual date belongs to \(T_O(A_0)\). Since ordinary and conditional surprise coincide here, either operator gives

\[A_1=T_O(G)=T_C(G)=\{1,2,3,4\}.\]

Friday is excluded. The original claim concerned surprise relative to \(A_0\).

If the actual date is Thursday, the original claim was true: without the announcement, Friday would have remained possible on Thursday morning. After the announcement, Thursday becomes predictable if no earlier exam occurs. Learning an epistemic claim can change the epistemic condition it describes.

This gives a precise reading of the teacher’s words: the exam would have been surprising without the information conveyed by this announcement. It is coherent, but weaker than the natural promise that the exam will surprise students who have heard the promise. Reapplying the operator requires an additional premise that the same surprise condition continues to hold in the updated model.

The baseline also matters. If \(A_0=\Omega\) and the guarantee and ordinary-surprise claim arrive together, then

\[F_O(\Omega)=G, \qquad F_O(G)=\{1,2,3,4\}.\]

Before the first announcement, even Friday is ordinarily surprising because giving no exam remains possible. Thus “the first update eliminates Friday” presupposes that the exam guarantee is already incorporated, or uses conditional surprise.

5. Interpretation B: include the announcement’s own consequences

Now take the stronger promise:

After learning this announcement and reasoning through its consequences, you will still be surprised if an exam occurs.

Let \(A^*\) be the possibilities retained after this information is fully incorporated. Every retained state must satisfy the promise when surprise is evaluated using \(A^*\) itself. For the optional-exam promise, this means

\[\boxed{A^*=T_O(A^*)\quad\text{or}\quad A^*=T_C(A^*).}\]

For a guaranteed exam, replace \(T_O,T_C\) by \(F_O,F_C\). Because these operators always return subsets of their inputs, the requirement that every retained state satisfy the promise already yields the fixed-point equality.

There may be several fixed points. If the announcement conveys no additional restriction on the date, select the greatest fixed point: the largest set compatible with the promise. In this finite model, obtain it by starting from all initially available possibilities and repeatedly applying the chosen operator until the set stops changing. The iteration computes one strong announcement’s meaning.

Guaranteed exam: no state survives

Starting from \(G\), both definitions give the familiar calculation:

Round	Remaining exam dates
0	Monday, Tuesday, Wednesday, Thursday, Friday
1	Monday, Tuesday, Wednesday, Thursday
2	Monday, Tuesday, Wednesday
3	Monday, Tuesday
4	Monday
5	None

Every nonempty finite set of dates has a last date. That date is predictable if no earlier exam occurs, so it cannot survive. The only fixed point is \(\varnothing\).

The conclusion is an impossible guarantee. The no-exam state was excluded by the guarantee. An empty information set represents inconsistency among the maintained requirements; \(\{0\}\) represents the information that no exam occurs.

Optional exam: the definitions separate

Under ordinary surprise, retaining \(0\) protects every exam date from certainty:

\[T_O(\Omega)=\Omega.\]

So the greatest fixed point is the full state space. Learning the promise leaves every possibility open.

Under conditional surprise, the last possible exam date is removed at each round:

\[\begin{aligned} \{0,1,2,3,4,5\}&\longrightarrow\{0,1,2,3,4\}\\ &\longrightarrow\{0,1,2,3\}\\ &\longrightarrow\{0,1,2\}\\ &\longrightarrow\{0,1\}\\ &\longrightarrow\{0\}. \end{aligned}\]

The set \(\{0\}\) is the greatest fixed point and the only nonempty fixed point. The empty set is also a fixed point.

Strong announcement	Ordinary surprise	Conditional surprise
Any exam will be surprising	\(\Omega\)	\(\{0\}\)
An exam will occur and be surprising	\(\varnothing\)	\(\varnothing\)

The entries are greatest fixed points starting from \(\Omega\). They distinguish a no-exam outcome, an inconsistent guarantee, and an unchanged information set.

Proof: all fixed points in the finite model

If \(0\in A\), then \(T_O(A)=A\). If \(A\) is nonempty and excludes \(0\), its largest date is removed, so it is not a fixed point of \(T_O\). Thus the ordinary operator’s fixed points are \(\varnothing\) and every subset containing \(0\).

For \(T_C\), any set containing an exam date has a largest such date, which the operator removes. The only fixed points are therefore \(\varnothing\) and \(\{0\}\).

A fixed point of either guaranteed-exam operator must be contained in \(G\). The largest-date argument then excludes every nonempty candidate.

Finally, let \(T\) denote any of the four operators and set \(A_{n+1}=T(A_n)\), starting from \(A_0=\Omega\). The sequence decreases and must stabilize because \(\Omega\) is finite. If \(B=T(B)\), induction and monotonicity give \(B\subseteq A_n\) for every \(n\). The terminal set consequently contains every fixed point and is itself a fixed point. It is the greatest one.

6. Simpson: what happens when trust is defeasible?

James Simpson’s A new epistemic solution to the Surprise Exam Paradox (2025, §§2–2.1) studies justified belief, with a reliable but fallible teacher. His diagnosis is that the student’s reasoning undermines the justification on which it depends. In the ordinary puzzle, the teacher’s testimony may be the student’s only reason to expect any exam. Rejecting that testimony can remove the reason for expecting a Friday exam. For conditional surprise, his Thursday argument restores Friday as a possible exam date once the restriction on which exams the teacher may give loses credibility.

We can express these changes in terms of the student’s remaining possibilities. If ordinary surprise is considered on Friday and the student’s possibilities become \(\{0,5\}\), then

\[\neg K_5^{\{0,5\}}E_5, \qquad K_5^{\{0,5\}}(G\Rightarrow E_5).\]

Losing certainty that an exam occurs restores ordinary surprise. Friday’s exam remains conditionally expected. Conditional surprise requires a later exam alternative, as on Thursday:

\[\neg K_4^{\{0,4,5\}}(G\Rightarrow E_4).\]

The substantive move is reopening possibilities previously excluded using the teacher’s testimony. That is belief revision, whereas our fixed-point calculation keeps the strong announcement as a hard constraint throughout.

Our strong interpretation gives a different conclusion for the promise “if there is an exam, it will not be conditionally expected.” At \(\{0\}\), the promise is vacuously true. Deducing that no exam occurs is therefore consistent with accepting the testimony.

I find it useful to separate the questions. The fixed-point model asks whether a guarantee can be true and accepted under specified information rules. Simpson asks what justification remains when a student’s argument calls the teacher’s testimony into question.

7. Baltag and coauthors: the same operator in a broader language

Alexandru Baltag, Nick Bezhanishvili, and David Fernández-Duque’s The Topology of Surprise (2022, §§4–5) develops a topological treatment. For the guaranteed five-day exam, its elimination operator is exactly ours. Their self-referential formulation is

\[P=\langle P\rangle\Diamond\top.\]

Here \(\langle P\rangle\) means that \(P\) is true and, after publicly learning it, the following condition holds; \(\Diamond\top\) says the actual state cannot be isolated by available observations. They select the greatest fixed point, called the perfect core. An empty core makes the strong announcement impossible; a nonempty core allows it to be truthful.

To see the equivalence directly, return to \(X=G\) and let the observable sets be

\[O_i=\{i,i+1,\ldots,5\}, \qquad \mathcal T=\{\varnothing,O_1,\ldots,O_5\}.\]

These observations record days passing without an exam, up to the exam morning.

For \(A\subseteq X\), define its derivative by

\[d(A)=\{x\in X:\text{every open neighborhood of }x \text{ meets }A\setminus\{x\}\}.\]

At \(i\), the smallest open neighborhood is \(O_i\). Hence, for \(i\in A\),

\[i\in d(A) \iff O_i\cap(A\setminus\{i\})\ne\varnothing \iff\exists j\in A:\ j>i.\]

Therefore

\[\boxed{T_O(A)=T_C(A)=A\cap d(A),\qquad A\subseteq G.}\]

The equivalence concerns this information structure and the strong guaranteed-exam model. Our inclusion of the no-exam state and distinction between the two surprise predicates makes those variants explicit.

8. What changes on a continuum?

The finite definition “there is a later possible date” uses the order of the week. The more general definition “no permitted observation isolates the actual state” works even when observations have no such order. This is where topology becomes useful.

The following examples apply that criterion to identifying a hidden number.

Arbitrarily accurate measurements and exact identification

Suppose a hidden number lies in \(X=[0,1]\). Observations are neighborhoods in the usual relative topology: measurements locate the number within intervals of positive, arbitrarily small width. Every neighborhood of every \(x\in[0,1]\) contains another possible value, including at the endpoints. Thus

\[d(X)=X, \qquad X\cap d(X)=X.\]

The strong announcement that the exact value will remain unidentified through these observations is consistent. Learning it leaves all values possible, preserving the uncertainty it describes.

The available evidence consists of finite-precision measurements and their finite combinations.

An informative announcement with a nonempty survivor set

For a slightly richer example, take

\[X=\{0\}\cup\{1/n:n=1,2,\ldots\}\cup[2,3]\]

with the usual relative topology. Every \(1/n\) is isolated, whereas \(0\) is approached by the sequence and every point in \([2,3]\) has neighboring interval points. The successive survivor sets are

\[\begin{aligned} A_0&=X,\\ A_1&=\{0\}\cup[2,3],\\ A_2&=[2,3],\\ A_3&=[2,3]. \end{aligned}\]

The first round removes the isolated sequence points. Once they disappear, \(0\) becomes isolated too and is removed in the second round. The interval remains forever. The strong announcement is informative: the student learns that the number lies in \([2,3]\), while its exact value remains unidentified.

The role of the observation structure

A continuum equipped with the discrete topology has singleton observations. Then \(d(X)=\varnothing\), and the strong announcement fails immediately despite the uncountably many states. Conversely, even a two-point space with only the observations \(\varnothing\) and \(X\) has no isolated points.

What matters is which states observations can distinguish. The five-day impossibility relies on a last remaining distinguishable date.

Technical note: iteration beyond finitely many rounds

For a general topological space, use the contracting operator \(A\mapsto A\cap d(A)\) and define

\[\begin{aligned} A_0&=X,\\ A_{\alpha+1}&=A_\alpha\cap d(A_\alpha),\\ A_\lambda&=\bigcap_{\alpha<\lambda}A_\alpha \quad\text{at limit ordinals }\lambda. \end{aligned}\]

At a limit stage, retain the states that survived every earlier stage. Continue until the sequence stabilizes, which in general may require more than countably many rounds. The sequence must eventually stabilize because it decreases through subsets of the fixed set \(X\). Monotonicity ensures that every fixed point is contained in every stage, so the terminal set is the greatest fixed point.

One can equivalently iterate \(d\) from \(X\): these iterates decrease, and at a limit stage monotonicity keeps the next derivative inside all preceding stages. The common terminal set satisfies \(A=d(A)\) and is the perfect core. Thus the derivative and contracting formulations yield the same greatest fixed point.

9. What I take the paradox to show

Under the strong interpretation, the backward calculation tests whether the teacher’s promise can survive its own dissemination. With a guaranteed exam in a finite week, it cannot. With only a conditional-surprise restriction, it can survive by implying no exam. With imperfect observations on a continuum, it can survive while leaving many states possible.

Under the one-step interpretation, further elimination requires a renewed guarantee of surprise in the updated model. Under a fallible-testimony interpretation, an argument that undermines trust must also reconsider the exclusions that depended on that trust. These are different diagnoses because they answer different formal questions.

The comparison with the forty cheating husbands puzzle helps locate the difference. There, subsequent publicly observed inaction, combined with the known decision rule, supplies new information each day. Here, the backward rounds can all be performed before Monday: they compute the proposed guarantee’s fixed point. The interpretation of the message determines whether those rounds are licensed.

For me, the most interesting lesson is that a self-referential announcement can be both informative and true. A statement can refer to what will remain unknown after it is learned. Whether it can be true depends on the possibilities and observations that remain.