Reynolds Paradox in System F Semantics

• Reynolds Paradox is a demonstration that impredicative universal quantification in System F yields no nontrivial set-theoretic model due to Cantor-style diagonal self-reference.
• It is formalized in Coq using a Church-encoded preinitial algebra and Lawvere’s fixed-point theorem to reconstruct the contradiction within the impredicative Prop universe.
• The paradox informs type-theoretic kernel design by illustrating that preserving extensionality requires stratified universes or weakened impredicativity to avoid self-referential collapses.

Reynolds paradox is the diagonal obstruction showing that System F, understood with impredicative universal quantification interpreted set-theoretically by intersection, has no non-trivial set-theoretic model. In the standard semantic aspiration, one assigns to each type a set, interprets function types as ordinary function sets, and interprets $\forall X.\,A$ by ranging over all subsets of a universe of sets. Reynolds’s result is that this combination of impredicativity, functoriality, and extensionality collapses through a Cantor-style self-reference. In the mechanised presentation in Reynolds.v from the coq-paradoxes library, the contradiction is reconstructed inside impredicative Prop by means of the [PHI](https://www.emergentmind.com/topics/preference-hijacking-phi) construction, a Church-encoded preinitial algebra, and Lawvere’s fixed-point theorem, yielding a proof of False from exactly the hypotheses that Rocq’s kernel is designed to reject (Alonso, 26 May 2026).

1. Semantic content of the paradox

System F, the second-order polymorphic $\lambda$-calculus, extends the simply-typed $\lambda$-calculus by impredicative universal quantification over types. Its type grammar is

$A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$

A set-theoretic model assigns to each closed type $A$ a set $\llbracket A\rrbracket$ and to each term $t:A$ an element of $\llbracket A\rrbracket$, with the clauses

$$\llbracket A\to B\rrbracket = \llbracket A\rrbracket \to \llbracket B\rrbracket, \qquad \llbracket \forall X.\,A\rrbracket = \bigcap_{S\subseteq U}\llbracket A\rrbracket_{[X:=S]}.$$

Here $U$ is a universe of sets, and impredicativity means that in forming $\lambda$0 one quantifies over all subsets $\lambda$1, including those that themselves involve $\lambda$2-types.

The paradox arises because these semantic requirements force more than a naive interpretation of polymorphism. The function-space clause is ordinary set-theoretic semantics. The universal-quantifier clause is impredicative by construction. The semantic account also requires type-substitution to be functorial: interpreting a type-level function $\lambda$3 should give a set-theoretic map on each interpretation $\lambda$4. Reynolds’s theorem, as stated in the later exposition, is that there is no non-trivial set-theoretic model of System F that validates the impredicative intersection clause; equivalently, System F has no model in the usual category of sets that interprets $\lambda$5 by intersection (Coquand, 2023).

The key intuition is diagonal. If polymorphism is interpreted by unrestricted intersection over a universe of sets, then the interpretation recreates a Cantor-style self-reference. The mechanised exposition sharpens this intuition into the claim that one cannot have a set Heyt into which all propositions inject in a way that respects extensionality (Alonso, 26 May 2026).

2. The impredicative PHI construction

In Reynolds.v, the argument is conducted in Coq’s impredicative Prop universe. One fixes a carrier Heyt : Prop together with a purported injection $t:A$0 This hypothesis packages the extensionality condition needed for the diagonal collapse: equality in Heyt must reflect logical equivalence of propositions.

The central endofunctor analogue is the double-power construction $t:A$1 The exposition renders the idealised set-theoretic notation as

$\lambda$6

From this one builds the Church-encoded preinitial algebra $t:A$2 An element $\lambda$7 is a polymorphic fold: for every algebra structure $\lambda$8, it produces an element of $\lambda$9. Because Prop is impredicative, A0 can itself be defined in Prop. However, the construction is only preinitial. Its universal property holds up to provable equality rather than definitional equality (Alonso, 26 May 2026).

This distinction is structurally important. The paradox does not begin from a strong initial algebra already granted by the ambient theory. Instead, it extracts from impredicativity a weak enough algebraic object to support diagonalisation, and then repairs the gap between preinitiality and strict inverse laws by quotienting with a partial-equivalence relation in the next stage.

3. Lawvere diagonalisation and the contradiction

The categorical core is Lawvere’s fixed-point theorem: in a cartesian-closed setting, a surjection $\lambda$0 yields a fixed point for every endomorphism $\lambda$1. In the simple set-theoretic form used in the exposition,

$\lambda$2

In the mechanisation, the statement is specialised to $\lambda$3 and $\lambda$4, using a particular surjection khi : A0 → (A0 → Heyt) supplied by the Reynolds construction (Alonso, 26 May 2026).

To obtain that surjection, Reynolds.v defines fold and unfold maps for the Church-encoded algebra: $t:A$3 These are the algebra and coalgebra structures

$\lambda$5

Because A0 is only preinitial, the mechanisation quotients by a partial-equivalence relation E0 so that cons and match become strict inverses modulo E0. This yields the bona fide isomorphism

$\lambda$6

The crucial map is then taken to be $t:A$4 and is shown surjective onto the reflexive elements of the relevant setoid of subsets of $\lambda$7.

Lawvere’s theorem is then applied to the endomorphism $t:A$5 to obtain a fixed point $\lambda$8 with $\lambda$9. The final step is elementary logic combined with the injectivity hypothesis: one shows that there is no $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$0 such that $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$1, so the fixed point is impossible. Coq therefore closes the theorem $t:A$6 The mathematical content is Reynolds’s diagonal or Cantor argument recast in Prop (Alonso, 26 May 2026).

4. Consequences for impredicativity and kernel design

The mechanised result is formulated as a no-model theorem for impredicative polymorphism. Any attempt to interpret the impredicative $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$2-quantifier of System F in a genuine set universe, together with an extensionality or injectivity requirement, leads to diagonal collapse and contradiction. In the formulation given in the exposition, one must give up at least one of the following: stratification of universes, impredicativity built into the function-type, or extensionality/injectivity of the interpretation map (Alonso, 26 May 2026).

Rocq’s kernel resolves the situation in three ways. First, universes are stratified so that one cannot form a type of “all types.” Second, Prop is impredicative, but Set is not impredicative by default. Third, the kernel refuses to admit an injective map I : Prop → Heyt by universe-constraint failures. In this sense, Reynolds.v functions as a parametrized proof of False, parametrized by the very hypotheses that the kernel refuses to discharge.

Within the broader coq-paradoxes package, Reynolds paradox is one of four formalised paradoxes that “draw the boundary of the Calculus of Inductive Constructions from the outside.” The package as a whole is presented as a negative specification of what Rocq’s kernel had to be designed to refuse, and the four examples together establish three boundary conditions on the kernel: the placement of impredicativity, the restriction of large elimination, and the discipline of universe constraints (Alonso, 26 May 2026).

5. Relation to Hurkens and Coquand’s variation

Later work presents Reynolds paradox together with two type-theoretic refinements: Hurkens’ paradox and Coquand’s variation. The comparative structure is explicit.

Construction	Ambient setting	Key obstruction
Reynolds	Set-theoretic semantics of System F	No model interpreting $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$3 by intersection
Hurkens	Pure type system $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$4	Weak $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$5-algebra yields $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$6
Coquand variation	Minimal HOL and $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$7	A single judgmental weak-algebra law yields $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$8

Hurkens internalises the diagonal in a pure type system with two universes $A,B \;::=\; X \mid A\to B \mid \forall X.\,A.$9 and $A$0 and an impredicative low universe. The relevant functor is

$A$1

and a weak $A$2-algebra consists of a type $A$3 with maps

$A$4

satisfying the judgmental equation

$A$5

From this, Hurkens derives a closed term of type $A$6.

Coquand’s variation sharpens the diagnosis. In minimal HOL, one sets

$A$7

assumes maps $A$8 and $A$9, and proves two impossibility results. The simplified theorem states that there is no type $\llbracket A\rrbracket$0 for which $\llbracket A\rrbracket$1 is a judgmental retract of $\llbracket A\rrbracket$2. The refined theorem states that there is no $\llbracket A\rrbracket$3 with maps $\llbracket A\rrbracket$4 and $\llbracket A\rrbracket$5 satisfying

$\llbracket A\rrbracket$6

The same refined paradox is then embedded into $\llbracket A\rrbracket$7 by taking

$\llbracket A\rrbracket$8

and defining intro and match so that the exact definitional law above holds (Coquand, 2023).

The explicit comparative claim is that Reynolds’s original impossibility was couched in set-theoretic semantics, Hurkens internalised the diagonal in a pure type system, and Coquand’s variation isolates a single judgmental weak-algebra equation as the culprit. The variation therefore refines the semantic no-go theorem into a more sharply targeted syntactic obstruction (Coquand, 2023).

6. Misconceptions, scope, and technical interpretation

A common overstatement is that the paradox requires a strong initial algebra, full extensionality, or the machinery of partial-equivalence relations from the outset. The later variation states instead that one does not need full extensionality, nor partial-equivalence relations, nor a strong initial algebra—just a weak initial-algebra law at the judgmental level (Coquand, 2023). In the mechanised Reynolds proof itself, the quotient by the partial-equivalence relation E0 appears only because the Church encoding supplies preinitiality rather than definitional initiality.

Another common misunderstanding is to read the paradox as a curiosity about one proof assistant encoding. The coq-paradoxes article explicitly argues against that reading. The package is presented not as a collection of curiosities but as a negative specification of what Rocq’s kernel had to be designed to refuse, and as evidence that the refusal is being made for the right reasons (Alonso, 26 May 2026). This suggests that the paradox is best understood as a structural boundary result for impredicative type theory and its semantics.

A further technical point concerns what exactly fails in set-theoretic semantics. The contradiction does not say that all semantics of polymorphism are impossible. The statement given is narrower and more precise: there is no non-trivial set-theoretic model of System F that validates the impredicative clause

$\llbracket A\rrbracket$9

In the Coq mechanisation, the corresponding forbidden ingredient is an injective interpretation of propositions into a carrier Heyt compatible with extensionality. A plausible implication is that any consistent semantics must block the diagonal by constraining universes, weakening impredicativity, or refusing the relevant injectivity or functoriality principles.

The paradox therefore occupies a precise place in the landscape of impredicative reasoning. Reynolds’s original theorem rules out the naive set-theoretic dream for System F; the mechanisation in Reynolds.v reconstructs that theorem through a fixed-point argument over an impredicative Church encoding; and later variations show that the decisive obstruction can be isolated in even weaker, more syntactic forms.