The Lie of the Liar - Resolving the Liar Paradox with Subjective Numbers

Abstract

The liar sentence, "this sentence is false," produces a familiar cycle: if it is true, it is false; if it is false, it is true. The classical derivation of the contradiction uses bivalence, the truth schema, self-reference, and classical logic, together with one premise it does not state: that the two occurrences of the sentence, the one that evaluates and the one that is evaluated, carry one value at one position. This article represents that premise as a judgment of an evaluating perspective, using the machinery of subjective numbers, and proves that the contradiction depends on it alone.

In that framework a value carries an intrinsic perspective and a judgment carries an evaluating perspective, and the two indices are distinct. The content of the liar is the pair of directed judgments x(Tgt) ⊑_e ¬x(Ptr) and x(Ptr) ⊑_e ¬x(Tgt) at an evaluating perspective e; the unstated premise is the identification x(Ptr) ⊑_e x(Tgt). The closure of e derives a contradiction if and only if it contains such an identification (Observer Theorem). Every derivation of the contradiction carries, in its proof term, the identity entry responsible and the perspective it was adopted from (Provenance Theorem). Identifications pass between perspectives only through the guard CPA, so the paradox is transmissible and blockable (Transmission Theorem). The two-role structure is the coarsest coherent quotient of the dependency unfolding, the classical structure is the next coarser one, and a reference cycle is coherent exactly when its number of negating links is even (Parity Theorem). An identification of two positions is licensed when every coherent valuation of the dependency already agrees on them, forbidden when none does, and open otherwise; licensed identifications change nothing and forbidden ones destroy coherence (Licensing Theorem).

The liar is the case in which the identification is forbidden: every coherent valuation of its content assigns its two occurrences different values. That is its semantic status, it is the same at every evaluator, and it is determinate. The classical error is adding as an axiom what is a theorem where licensed, an unforced choice where open, and refuted where forbidden. Bivalence, disquotation, self-reference, and classical inference hold at every position. The same form, a coherent dependency together with a forbidden identification, is then shown to underlie Grelling's paradox, Curry's paradox, Russell's paradox, Berry's paradox, Yablo's paradox, and the revenge liar, each with its own dependency map and each decided by the same criterion. In the reading ring of the framework article the liar is a unit σ with σ²=1 that reads true at the pointer and false at the target; re-reading after the loop sends σ to −σ, and the parity theorem is its descent criterion. The liar stands to the classical values as √−1 stands to the reals.

The Resolution in Plain Terms

The sentence "this sentence is false" is doing two jobs at once. It is the one saying something, and it is the thing being talked about. Those are two positions, in the way that "I" in a sentence written about oneself is both the author and the subject.

Read that way, the sentence is in order. It says: whatever value the thing talked about has, the thing saying it has the opposite. Two positions, opposite values. There are exactly two ways to satisfy that, they are mirror images of each other, and neither breaks anything.

The contradiction appears only when a reader insists that the two positions must hold the same value, on the ground that it is the same sentence. That insistence comes from the reader, since the sentence does not contain it. And the sentence's own content forbids it: every consistent reading gives the two positions different values. A reader who identifies them contradicts the sentence and then attributes the contradiction to the sentence.

Readers do this because for every ordinary sentence, treating the thing said and the thing talked about as one value is harmless and correct; it is how "'snow is white' is true" works. The habit never fails until the single case in which the thing said is the thing talked about. There it turns a two-position sentence into a one-position knot.

What the machinery of subjective numbers adds is the ability to say who did it and how. A reader is a perspective with its own list of judgments. The identification of the two positions is an entry on that list. Whether the sentence allows the entry can be checked before it is added: allowed for ordinary sentences, forced for "this sentence is true," forbidden for the liar. If a reader adds it anyway, the derivation of the contradiction contains that entry, so it can be pointed to. If the reader took it from another reader, the record shows that as well. The paradox has an address and a chain of custody.

There is also a positive way to say what the liar's value is. There is no real number whose square is −1, and the complex numbers supply one, i, with a conjugation that swaps i and −i. There is no ordinary value that is the negation of itself, and the subjective numbers supply one, σ: it reads true from the pointer and false from the target, it squares to one, and going once around the loop swaps σ and −σ. From either position it adds and multiplies like an ordinary number. What it lacks is a reading that is the same from both, and that lack is the paradox.

The liar has no single classical value, and it is not defective. It is a sentence whose two halves cannot share a value, and the paradox is what happens when a reader makes them. The same check settles the other classical paradoxes of self-reference: each is a consistent dependency plus an identification that the dependency forbids.

Notation Guide

1. Introduction

1.1 The Liar Paradox

The liar paradox, instantiated by the self-referential statement "This statement is false," challenges the assumption that every meaningful statement can be assigned a determinate unindexed truth value. If the statement is assigned true, its content forces it to be false; if it is assigned false, its content is satisfied and it is true. The consequences of this small cycle reach into the theory of meaning, the foundations of mathematics, the definability of truth, and the limits of formal systems.

The paradox uses no exotic machinery. It arises inside ordinary language with ordinary inference. Whatever it reveals must already be present in the ordinary interaction between reference, predication, and the standpoint from which a sentence is evaluated.

1.2 Where the Missing Parameter Lives

Accounts of the liar that keep the language unrestricted and keep classical logic share one diagnosis: the classical derivation omits a parameter, and once the parameter is restored the contradiction dissolves into a dependency that is locally consistent. They differ on where the parameter is placed.

• Between the sentence and the world. In situation semantics (Barwise and Etchemendy 1987) an Austinian proposition is a claim that a particular situation is of a particular type, and the situation a statement is about is a contextual parameter between the sentence and the proposition it expresses. The liar proposition is false, and the fact of its falsity lies in a situation larger than the one the proposition was about.
• Between sentences. Reference-graph analyses (Gaifman 1992; Cook 2004; Walicki 2009; Rabern, Rabern, and Macauley 2013; Beringer and Schindler 2017) assign one truth-bearer to each sentence and record which sentences depend on which. Paradoxicality is a property of the reference pattern. When the graph is unfolded into a tree, a consistent decoration always exists; paradoxicality enters when the decorated tree is re-collapsed onto the graph.
• Between measurement contexts. The sheaf-theoretic account of contextuality (Abramsky and Brandenburger 2011; Abramsky, Barbosa, Kishida, Lal, and Mansfield 2015) describes families of data that are consistent on every context and admit no global section. Liar cycles, encoded as systems of Boolean equations, have the same structure.

Each of these accounts identifies the operation that manufactures the contradiction: the demand for one situation, one decoration, one global section. None of them represents the party that makes the demand. The re-collapse in a reference-graph analysis has no agent performing it. The global section in a sheaf model has no observer requiring it. The situation in Austinian semantics is a portion of the world. The question this article asks is therefore who collapses the parameter, and what follows from representing that party explicitly. Section 11 compares the accounts in detail.

1.3 The Evaluating Perspective

Subjective numbers give that party a formal home. A subjective number a(s) is a value a together with the intrinsic perspective s at which it is held. A closed judgment a(s) ⊑_p b(t) is validated by an evaluating perspective p, which is a separate index from s and t. The framework carries two kinds of perspective at once: the perspective a value belongs to, and the perspective from which a relation between values is judged.

For the liar, the two occurrences of the sentence are two intrinsic perspectives, Ptr and Tgt. The falsity predicate, read from either occurrence toward the other, contributes the directed judgments x(Tgt) ⊑_e ¬x(Ptr) and x(Ptr) ⊑_e ¬x(Tgt). These are validated by an evaluating perspective e that is neither occurrence. The identification of the two occurrences, x(Ptr) ⊑_e x(Tgt), is contributed by neither the falsity predicate nor the demonstrative. It is an additional judgment that e may hold, adopt from another perspective, or refuse. The contradiction is derivable at e exactly when that judgment is in the closure of e.

• Two indices. The occurrence that holds a value and the standpoint that judges a relation are distinct parameters.
• Collapse as judgment. The identification of pointer and target is a kernel entry of an evaluator.
• Guarded passage. One perspective may use another's judgments only where the guard CPA permits, and the guard is relative to where the adopted judgment lands.
• Local classicality. Each fixed position uses ordinary Boolean values and ordinary inference.

1.4 What a Resolution Requires

A resolution of the liar is held to five requirements. It must assign the sentence a determinate semantic status. It must identify the step at which the classical derivation fails and show that the step is unjustified. It must preserve the classical principles as far as they can be preserved and state exactly what is given up. It must treat revenge sentences by the same mechanism it uses for the liar. And it must not relocate the contradiction into the meta-theory. Section 8 states the resolution as a single theorem and checks it against each requirement.

Universe 00110000

2. The Classical Derivation and Its Hidden Premise

2.1 Formal Derivation of the Paradox

Let L denote the liar statement "This statement is false." Introduce an unindexed truth evaluation function τ mapping statements to values in {true, false}. The content of L requires

The equation determines the truth value of L by the content of L, which refers to that same truth value. Both cases fail:

• If τ(L) = true, then the claim made by L holds only if τ(L) = false.
• If τ(L) = false, then the claim made by L is correct, so τ(L) = true.

No Boolean value satisfies τ(L) = ¬τ(L).

2.2 The Hidden Premise

The derivation appears to use only familiar principles:

• Bivalence: Every statement has a definite truth value, either true or false.
• Truth Schema: The statement "X is true" is true if and only if X is true.
• Self-Reference: Statements may refer to themselves.
• Classical Logic: Excluded middle, non-contradiction, and ordinary rules of inference.

It uses one more. The two occurrences of the sentence, the one that says and the one that is said about, are held to a single value at a single position. Writing the paradox as an equation in this way is standard, and Wen (2001) treats the semantic paradoxes as equations throughout. The equation τ(L) = ¬τ(L) is written only after this identification has been made: τ has one argument, and the argument is the expression rather than an occurrence of it.

The identification is performed by whoever evaluates the sentence. The derivation is an act of an evaluator who requires a single value for the expression across both of its positions. This demand is the Universality Assumption: the requirement that an evaluating standpoint validate one value for the expression irrespective of the position at which the expression occurs. In the vocabulary of sheaf-theoretic contextuality, it is the demand for a global section. The demand hides inside the truth schema. The schema True(⌜S⌝) ↔ S connects S as evaluated, the argument of the truth predicate, with S as asserted, and so bundles two principles. One is disquotation: an attribution of a value to a sentence is correct exactly when the sentence has that value. The other is occurrence identity: the sentence evaluated and the sentence asserted carry one value. For a grounded sentence the second principle is a theorem of the sentence's dependency, since the same non-semantic fact fixes the value at both occurrences and every coherent valuation agrees on them (Theorem 8.3). The classical derivation adds it as an axiom for every sentence, and for the liar the axiom is refuted by every coherent valuation of the sentence's own content. The term "universality" is used here for a property of an evaluator; the expressive universality of natural language discussed by Simmons (1993) is a different notion.

2.3 Meta-Theoretical Setting

The framework is formulated inside ordinary set-theoretic mathematics, using sets, functions, relations, categories, and morphisms. It introduces no axiom asserting a universal truth predicate over its own total language. If ZFC is consistent, the external mathematics used to define the role-indexed models is consistent relative to ZFC. The liar analysis takes place inside object-level indexed structures. The meta-theory describes the family of perspectives; it is not a member of the perspective set and validates no judgment of the form ⊑_p. It adds no truth value, no inference rule, and no hidden evaluating perspective to the object-level system.

2.4 Static Evaluation Convention

Unless an update parameter is explicitly introduced, evaluative positions are fixed. Repeated traversal of the liar relation is represented as a directed cycle of indexed judgments, and a time-dependent sequence of changing truth values is a different representation of the same traversal; Section 11 compares it with revision sequences.

3. Role-Indexed Truth Bearers

Role-indexing distinguishes two semantic positions occupied by one expression. The phrase "this sentence" preserves referential identity: the same sentence L is referred to. It does not settle role identity, whether the evaluating occurrence and the evaluated occurrence are one semantic position, and it does not settle value identity, whether the value assigned at one position must equal the value assigned at the other.

Three claims are therefore separated that the classical derivation compresses into one:

• Referential identity: the pointer and the target concern the same expression L.
• Role identity: the pointer position and the target position are one semantic position.
• Value identity: the values assigned at those positions are required to be equal.

The liar sentence supplies the first claim together with a directed attribution of falsity. It supplies neither the second nor the third.

3.1 Role-Indexed Bearers

Let Roles = {Ptr, Tgt}. A role-indexed truth bearer is a pair (L, r) with r ∈ Roles. A role valuation is a function

Write v_Ptr(L) for v(Ptr) and v_Tgt(L) for v(Tgt). The reference map sends both bearers to the same expression:

Self-reference is preserved. What is represented is one expression at two positions of one dependency.

3.2 The Directed Liar Relation

The demonstrative fixes the target of evaluation; the predicate "is false" performs the evaluation from the position of the whole sentence. Read from the pointer toward the target, the predicate contributes

The target occurrence is itself the sentence, and so is itself a pointer whose target is the pointer occurrence. Read from the target toward the pointer, the predicate contributes

As constraints on a valuation the two equations are one equation. As judgments they are two directed edges, and Section 4 uses the directed form.

The proposition is elementary. Its content lies in the adequacy of the representation: the two positions are positions of one dependency. Section 3.3 gives the semantic ground and Section 7 gives the structural ground, by showing that the two roles are the classes of the coarsest coherent quotient of the dependency unfolding.

3.3 Why There Are Two Positions

Two distinct operations are performed by the sentence. Reference determines what is evaluated; predication determines how the evaluation is made. The demonstrative contributes a target; the predicate contributes a transformation applied to the value found at that target. An expression can be the target of its own predicate only by occurring twice in one act of evaluation: once as the thing referred to, once as the thing doing the referring. These are two occurrences of one type. Discourse-level analyses reach the same conclusion from the other side: the carriers of truth values are sentence tokens, and saying the same sentence type at two points of a network of cross-references produces two statements, each a node of the discourse graph (Gaifman 1992; Dyrkolbotn and Walicki 2014). The role split records the same distinction inside a single sentence, where the network has exactly two nodes and one type.

The split decides neither the value at either position nor whether the positions carry one value. Those are the second and third claims of the list above, and the second is the one the classical derivation makes without stating it.

4. The Evaluating Perspective

4.1 Two Indices

A closed judgment a(s) ⊑_p b(t) carries three perspective indices: the intrinsic perspective s of the source value, the intrinsic perspective t of the target value, and the evaluating perspective p that validates the judgment. For the liar, s and t range over Roles, and p is an evaluating perspective e that is neither role. The reading of the judgment throughout is determination, as in Section 5 of the framework article: from the standpoint of e, the value a at position s fixes the value b at position t.

Let P ⊇ {Ptr, Tgt, e} and V = {true, false}. The roles are positions only: their kernels are empty and their guards are identically 0. The base kernel of e records the directed liar relation as judgments of e:

These four entries are the liar edges at e. They are negating edges between the roles. An identity edge between the roles is an entry R^∘_e(Ptr, Tgt, x, x) = 1 or R^∘_e(Tgt, Ptr, x, x) = 1. The collapse C_e is the set of all four identity edges. Neither the demonstrative nor the predicate contributes an identity edge. An identity edge is an additional judgment of e: the judgment that the value at one occurrence is the value at the other.

A contradiction at e is a closed judgment of the form x(r) ⊑_e ¬x(r), a same-role negating judgment. Under the determination reading it says that the value x at r fixes the value ¬x at r, which no valuation with v(r) = x can respect. It is a positive judgment, never a derived negation; the framework's Theorem 3.4 is untouched. When both values are forbidden at a role, no valuation survives, and e is incoherent in the sense of the framework's Section 5.

Throughout this section the base kernel of e contains role edges only and the guard CPA_e is identically 0, so that the closure of e is generated by its own kernel. Adoption from other perspectives is treated in Section 5.

4.2 Coherent Role-Valuations

This is the framework's notion of coherence with the positions of e equal to the two roles. By Lemma 5.1 of the framework article, a valuation is coherent exactly when it respects every generator of e, that is, every kernel entry and every imported judgment: a reflexive judgment is respected by every valuation, and respect is preserved under transitive composition. A same-role negating judgment x(r) ⊑_e ¬x(r) is respected by a valuation only if v(r) ≠ x; it forbids one value at one role. The full collapse forbids both values at both roles.

4.3 The Observer Theorem

Every closed judgment of e between role positions is witnessed by a path of base edges. Each base edge is a step between the two roles and is either negating, sending x to ¬x, or identity, sending x to x. The value carried around a path is therefore determined by the number of steps of each kind.

A partial collapse, in which I contains identity edges for one value only, forbids that value at the affected role and leaves one alternating valuation coherent. The full collapse forbids every value at every role. In each case the presence and extent of the contradiction at e are read off the identity edges that e validates.

Universe 00110000

4.4 Global View Without Global Section

Two demands on an evaluating perspective must be kept apart. The first is that it represent the whole dependency: that it hold the liar edges and thereby see that the value at each occurrence determines the value at the other. The second is that it hold one value for the expression: that some coherent valuation assign the same value to both occurrences.

The Universality Assumption at e is the demand for a global section. Proposition 4.6 says that no perspective with a global view can meet the demand, and Theorem 4.3 says what happens when the demand is enforced anyway by adding the collapse: the perspective loses its local sections as well. Under Theorem 4.3 (i), e sees everything and asserts one value for nothing; under (ii), e has demanded one value and can coherently hold none.

The meta-theory is the limiting case of a global view. It defines R^⋆_p for every p ∈ P, so it sees every perspective's judgments, including every collapse and every refusal to collapse. It belongs to no perspective, validates no judgment of the form ⊑_p, and assigns L no value. Surveying the multiplicity of positions and adjudicating between them are different acts, and only the second is the source of the paradox.

4.5 Provenance of the Contradiction

The closure of a perspective is generated by derivations, and the framework records each derivation as a proof term: base_p(s,t,a,b) for a kernel entry, refl_p(x) for reflexivity, trans(π,ρ) for transitive composition, and cpa_p,q,u(π,ρ) for a guarded composition whose right edge is the q-derivation ρ. Substitution is vacuous over {true, false}. The provenance functor H_p sends a derivation to its history, recording each base entry and each adoption event together with the perspective adopted from. A contradiction at a perspective is therefore a derivation with a record.

Traceability is what distinguishes a resolution from a prohibition. A hierarchy forbids the identification by making it unformulable. The present account permits it, records it when it occurs, and exhibits it as the sole cause of the contradiction in every derivation that reaches one.

5. Transmission of Collapse Between Perspectives

Perspectives do not evaluate in isolation. The closure of e may compose a judgment validated by e with a judgment validated by another perspective q, provided the guard permits it. The rule is

The guard is relative to the intrinsic perspective u at which the adopted edge lands. Since the left premise can be reflexive, the rule lets e take on every judgment of q that lands at u (framework article, Theorem 3.6). A perspective may accept what q concludes about one position and refuse what q concludes about another.

Three features of the theorem deserve emphasis. First, the collapse does not propagate by decree; it propagates only through an explicit composition that the receiving perspective has permitted, and the composition is recorded in the provenance of the resulting judgment. Second, although the guard is relative to the landing position, permission at either role transmits the full contradiction, because the liar's own edges connect the roles; immunity to a collapsed perspective requires refusing its judgments at both roles. Third, a collapse can be distributed: if q₁ validates x(Ptr) ⊑ x(u) and q₂ validates x(u) ⊑ x(Tgt) for a position u outside the roles, neither holds an identity edge between the roles, yet an evaluator that adopts both assembles one. The general condition is the first clause of Corollary 5.2: what matters is whether a net-identity path between the roles is assembled in the evaluator's closure, from whatever sources. The parity of Lemma 4.2 is a property of the net transformation along a path, so Theorems 4.3 and 4.7 hold with "identity edge" read as "net-identity path," and the provenance of such a path records every perspective that contributed a segment. The framework's Theorem 5.3 (4) states the general fact: an identification enters a perspective's closure only through its own kernel or through a permitted import, and the history names the route.

Universe 00110000

6. Collapse and Fixed Points

The liar is one member of a family. A reflexive construction carries a transformation F: V → V across the reference: the value found at the target is transformed into the value asserted at the pointer. For the liar, F = ¬. For the truth-teller, "this sentence is true," F is the identity. In general the edges at e are

and the collapse C_e is as before.

For the liar, ¬ has no fixed point, and the collapsed evaluator has no coherent valuation. For the truth-teller, every value is a fixed point, and the collapsed evaluator has two coherent constant valuations; the collapse is coherent and leaves the value undetermined. Self-reference alone is therefore never the source of contradiction. What matters is the transformation carried across the reference and whether the evaluator has identified the two occurrences.

On the collapsed object, Theorem 6.1 is the fixed-point criterion of Kripke (1975): the liar is paradoxical because v = ¬v has no solution, and the truth-teller is undetermined because v = v has too many. Beringer and Schindler (2017) show that the reference-graph notion of paradoxicality coincides with Kripke's, and Theorem 6.1 is that coincidence restated at a single evaluator. What the present account adds is the step before the test: whether the collapse belongs to the closure of a given evaluating perspective, by what path it arrived there, and whether the evaluator was entitled to make it. The last question is decided by the licensing criterion of Section 8, which reads the entitlement off the coherent valuations of the content before any collapse is made.

7. The Unfolding and the Parity Theorem

Two roles were introduced in Section 3 by observing that the sentence occurs twice in one act of evaluation. But the target occurrence is the sentence, and so contains a demonstrative that refers to the sentence, which occurs again. Following the reference produces a sequence of positions. The two-role structure is a quotient of that sequence, and the choice of quotient is where the paradox is decided.

7.1 The Dependency Unfolding

Let the positions of the unfolding be r_n for n ∈ ω, with r₀ the outermost occurrence. Each occurrence r_n refers to r_n+1 and attributes falsity to it. Each such link contributes the constraint v(r_n) = ¬v(r_n+1), recorded at an evaluating perspective e as the pair of negating edges

The unfolding is the ray obtained from the single-node reference graph of the liar by opening its self-loop, and a valuation of the unfolding is a decoration of that ray. Beringer and Schindler (2017) prove that the unfolded structure always admits a decoration and that paradoxicality enters only when the decorated structure is re-collapsed. A different construction, the unwinding of Cook (2004), replaces the loop by an ω-sequence each of whose members refers to all later members; it preserves paradoxicality and turns the liar into Yablo's paradox. The unfolding used here removes paradoxicality until a quotient is imposed. The two constructions answer different questions.

7.2 Quotients of the Unfolding

A quotient of the unfolding is a partition of its positions into classes. At e, the quotient is the addition of the identity edges x(r_m) ⊑_e x(r_n) for all x and all positions r_m, r_n in a common class. The quotient of period k ≥ 1 identifies r_n with r_n+k for every n; equivalently, it is the cycle of k classes [r₀], …, [r_k−1] joined by k negating links.

The period-2 quotient has two classes, the even positions and the odd positions. Its links become the edges x(Tgt) ⊑_e ¬x(Ptr) and x(Ptr) ⊑_e ¬x(Tgt) with Ptr = [r₀] and Tgt = [r₁]. This is exactly the two-role model of Section 4, and under the quotient the link from r₁ to r₂ becomes a link from the target back to the pointer, which is the reading of Section 3.2 that the target occurrence is itself a pointer. The period-1 quotient has one class and one self-loop, x(r) ⊑_e ¬x(r): it is the classical unindexed liar of Section 2.1.

This answers the question of why there are two roles and not three or thirty. There are infinitely many positions, two is the smallest number of classes into which they can be gathered without contradiction, and every coherent way of gathering them refines the split into pointer and target. The role split of Section 3 is the coarsest representation of the liar that an evaluating perspective can hold coherently, and the classical derivation is the attempt to hold one class fewer.

The period-2 quotient has a name. Two positions, each attributing falsity to the other, is the open pair: two sentences of which each asserts that the other is false (Sorensen 2001; Armour-Garb and Woodbridge 2006). Its dependency has exactly two coherent valuations, exchanged by the symmetry that swaps the positions, and nothing in its content selects one. The liar, evaluated without the collapse, is the open pair, and the symmetry of the open pair is total for the liar because the two positions are occurrences of one sentence.

The status has two components. The constraint v(Ptr) ≠ v(Tgt) is determinate: it holds in every coherent valuation and is the content of the sentence. The value at either position is not determinate: the symmetry exchanges the two valuations and nothing selects between them. In the vocabulary of fixed-point semantics the liar is ungrounded, and the accounts that treat the open pair as pathological because of this symmetry are describing the same fact. Section 8 turns the constraint into the verdict: the identification of the two occurrences is forbidden, and that is what distinguishes the liar from every ungrounded sentence whose identification is merely open.

7.3 Mixed Cycles and the Odd-Cycle Criterion

A reference cycle need not attribute falsity at every link. A link at which one sentence asserts that the next is true contributes identity edges rather than negating edges. Let a cycle have k positions and m negating links, the remaining k − m links being identity links.

Theorem 7.5 is the balance theorem for signed graphs (Harary 1953) applied to reference cycles: a signed graph is balanced exactly when every cycle carries an even number of negative edges, and a reference cycle is coherent exactly when it is balanced. It recovers the classical criterion for finite reference paradoxes. A cycle of k sentences each asserting that the next is false is paradoxical exactly when k is odd: the open pair is coherent, the liar triple is not. A cycle of N sentences in which each asserts that the next is true and the last asserts that the first is false has m = 1 and is paradoxical for every N; any N − 1 of its links are jointly satisfiable and all N are not, and for N = 4 the structure is the one Abramsky, Barbosa, Kishida, Lal, and Mansfield (2015) identify with the Popescu–Rohrlich box. A cycle with m = 0 is a truth-teller cycle, coherent and undetermined. In digraph terms the criterion is that a finite reference graph without an odd cycle has a kernel, and the dangerous finite patterns are the odd cycles (Rabern, Rabern, and Macauley 2013).

The parity invariant is coarser than the cohomological obstruction used to witness contextuality, which distinguishes structures that parity does not. It is finer than the two-role model taken alone, which sees every cycle as one transformation and one collapse. Cycles of different lengths and signatures are different quotients of different unfoldings, and the framework distinguishes them by the parity of the net transformation each quotient forces around its loop.

7.4 The Liar as a Subjective Number

The framework article makes the unfolding into a ring. Over a field K of characteristic other than 2, a subjective number over the loop graph is a function from the positions r_n of the unfolding to K, its reading at each position, with n now ranging over all integers so that the ray of Section 7.1 is extended backward by the occurrences that refer to r₀; readings add and multiply pointwise, so every reading is a ring homomorphism, and the deck transformation, which moves every position one step along the reference, acts by a ring automorphism. Encode true as 1 and false as −1, so that the liar's transformation F=¬ is multiplication by −1.

This is the liar's value stated positively. The sentence has a value, σ; the value is an ordinary element of an ordinary ring; from either occurrence it behaves like any other number; and its one peculiarity, that it reads true from one occurrence and false from the other, is the content of the sentence. The classical evaluator's demand for one value is the demand that σ be a constant, which it is not, and the demand yields nothing because there is nothing to yield. There is no real number whose square is −1, and the complex numbers supply one together with a conjugation that swaps it with its negative; there is no classical value that is its own negation, and the subjective numbers supply one together with a re-reading that swaps it with its negative. The parity theorem, the coarsest-quotient theorem, and the Observer Theorem are the statements that σ descends to even quotients, to the two-role quotient first, and to the classical quotient never. Spencer-Brown (1969) and Kauffman (2002) proposed an imaginary truth value for the liar, modeled on √−1 and on the oscillation between true and false; the unit σ differs in satisfying σ²=1, in being derived from the unfolding rather than assigned, and in carrying the descent criterion that ties it to the licensing of the collapse.

Universe 00110000

8. Licensing and the Resolution

The results of Sections 3 through 7 assemble into a criterion and a theorem. The criterion says when an evaluator may identify two positions. The theorem says what the liar is.

8.1 Licensing

Two of the constructions below, Curry's sentence and the revenge liar, relate more than two positions. For them the kernel is allowed to contain a conjunctive edge a₁(s₁), …, a_k(s_k) ⊑_e b(t), which a valuation respects when v(s_i) = a_i for every i implies v(t) = b. Coherence is respect of every kernel entry, identity edges are as before, and Theorem 8.2 holds word for word, since its proof uses only that coherence is respect of every entry. The parity results of Sections 4 and 7 concern binary edges and are not used for conjunctive ones.

8.2 Classification

The classification separates the liar from every other case by a single feature. For the liar, one identification, of one sentence's two occurrences, is forbidden by the content of that sentence alone. The truth-teller's identification is licensed, the grounded sentence's is licensed and determining, the open pair's are open, and the cycles and Yablo need the classical evaluator to identify several sentences' occurrences at once before anything is forbidden. Section 9 shows that Grelling, Russell, and Berry share the liar's feature, and that Curry is the open case with a consequence.

8.3 The Resolution Theorem

8.4 The Requirements Met

8.5 What Each Component of the Machinery Does

8.6 The Classical Error

The classical evaluator adds the identification of every sentence's occurrences to its kernel as an axiom. Theorem 8.3 says what that axiom is in each case. For a grounded sentence it is a theorem: the same fact fixes both occurrences, and this is why the habit exists and why it is harmless almost everywhere. For the truth-teller it is a theorem of the content and adds nothing. For the open pair it is an unforced choice among coherent valuations. For an odd cycle or for Yablo it is a set of individually coherent choices that is jointly refuted. For the liar it is a single judgment refuted by the sentence's own content. The classical evaluator treats these cases alike, and the paradox is the case in which the axiom contradicts the content. The error is one of adoption: what the dependency licenses in some cases, leaves open in others, and forbids in one is taken as an axiom in all. The machinery makes each case visible: the identification is a kernel entry, its status is read off the coherent valuations, and its consequences and provenance are theorems.

9. The Identification Schema

Sections 4 through 8 treat one sentence. The form of the result is general. A construction supplies a dependency between positions, the classical reading identifies some of those positions, and the identification has a status that the dependency fixes in advance. This section states the form as a definition and a theorem, and then works through the classical paradoxes of self-reference with it.

9.1 The Schema

The schema says that a paradox of this family has three parts and that only one of them is defective. The dependency is coherent. The evaluator is a perspective with a kernel. The identification is an entry of that kernel whose status the dependency already fixes. A construction is paradoxical when the entry is forbidden and the evaluator adds it anyway, and the theorem says that the resulting contradiction is located at the entry, is traceable to it, and travels only along guards.

9.2 Instances

Each instance below is given by its positions, its content, the identification the classical reading makes, and the verdict of Theorem 8.2 on that identification.

(a) Grelling. A word is heterological when it does not describe itself (Grelling and Nelson 1908). Is "heterological" heterological? The predicate applied to the word is one occurrence of the question, H^Ptr, and the word applying to itself, which the predicate asks about, is the other, H^Tgt. The content is that the predicate applies exactly when the word does not apply to itself: v(H^Ptr) = ¬v(H^Tgt), recorded as the four liar edges between H^Ptr and H^Tgt. By Theorem 8.3 (i) the identification is forbidden. Grelling's paradox is the liar with a predicate in place of a demonstrative.

(b) Curry. Let K be the sentence "if K is true then C," for an arbitrary sentence C (Curry 1942). The positions are K^Ptr, K^Tgt, and C. The content is the equation v(K^Ptr) = (v(K^Tgt) → v(C)), recorded as the four conjunctive edges x(K^Tgt), y(C) ⊑_e (x → y)(K^Ptr) for x, y ∈ V. The coherent valuations are the four choices of (v(K^Tgt), v(C)), each with v(K^Ptr) determined. The classical reading identifies K^Ptr with K^Tgt. By Theorem 8.2 the surviving valuations are those with v(K^Tgt) = (v(K^Tgt) → v(C)). With v(K^Tgt) = true this requires v(C) = true; with v(K^Tgt) = false the right side is true, so no valuation survives. Exactly one valuation survives, v(K^Ptr) = v(K^Tgt) = v(C) = true. The identification is therefore open, since one of the four coherent valuations agrees on the pair and three do not, and it is determining, and what it determines is v(C) = true. This is Curry's paradox stated exactly: the classical derivation passes between K as asserted and K as evaluated through the truth schema, that passage is the identification, the identification is an unforced choice, and the arbitrary conclusion C is the value that the choice determines. When C is fixed false on independent grounds, recorded as the entry true(C) ⊑_e false(C), no valuation survives the identification and it is forbidden: Curry's sentence with a false consequent is the liar.

(c) Russell. Let R be the class of all sets that are not members of themselves (Russell 1903). The comprehension principle says that x ∈ R exactly when x ∉ x, and the instance x := R reads: R ∈ R exactly when R ∉ R. The two sides are two occurrences of the membership fact. On the left, R is the container, the class that comprehension defines; on the right, R is the argument, a set asked whether it belongs to itself. Let m^Ptr be the fact with R as container and m^Tgt the fact with R as argument. The content is v(m^Ptr) = ¬v(m^Tgt), the four liar edges, and by Theorem 8.3 (i) the identification of the two occurrences is forbidden. The evaluator that makes it is naive comprehension read with one membership relation regardless of the role a set plays. The theory of types (Russell 1908) and stratified comprehension (Quine 1937) keep the container occurrence and the argument occurrence at different indices, so that the identification is never made; in the present terms they refuse a forbidden identification.

(d) Berry. Consider the phrase "the least integer not nameable in fewer than nineteen syllables," which has eighteen syllables (Russell 1908). Let n be the integer the phrase describes. Two stocks of names are in play: the stock the phrase quantifies over, and that stock enlarged by the phrase itself. Let N^Tgt be the statement that n is nameable in fewer than nineteen syllables, evaluated in the first stock, and N^Ptr the same statement evaluated in the second. The content has three parts. By the definition of n, N^Tgt is false, recorded as true(N^Tgt) ⊑_e false(N^Tgt). Every name in the first stock is a name in the second, recorded as true(N^Tgt) ⊑_e true(N^Ptr). And the phrase names n in eighteen syllables, so N^Ptr is true, recorded as false(N^Ptr) ⊑_e true(N^Ptr). The only coherent valuation has v(N^Tgt) = false and v(N^Ptr) = true, so the identification of the two stocks is forbidden. The classical evaluator treats the stock of names as one stock regardless of whether the phrase is counted in it, and that is the identification.

(e) Yablo. Let sentences Y_n, for n ∈ ω, each say that every later sentence is false (Yablo 1993; Cook 2014). The positions are Y_n^Ptr and Y_n^Tgt for every n, and the content at n is the conjunctive constraint that Y_n^Ptr is true exactly when Y_m^Tgt is false for every m > n. The valuation with every target false and every pointer true is coherent. A single identification Y_n^Ptr ~ Y_n^Tgt is open: the valuation with every target false except Y_n^Tgt, which is true, with Y_m^Ptr false for m < n and true for m ≥ n, is coherent and agrees at n, while the all-targets-false valuation is coherent and disagrees. The full set of identifications is forbidden: with every pointer identified with its target, if some Y_n is true then every later sentence is false, so Y_n+1 is false and some sentence later than n + 1 is true, against the falsity of every sentence later than n; and if every sentence is false then every sentence later than 0 is false, which makes Y₀ true. This is Theorem 8.3 (vi).

(f) Revenge. The revenge liar, "this sentence is not true at any perspective," has positions R^Ptr(e) and R^Tgt(q) for every accessible perspective q. Its single self-identification is forbidden when the evaluator's only accessible target is its own and open otherwise, and its joint self-identification across every self-accessing perspective is forbidden. Section 10 proves this.

9.3 Uniform Accounts of Paradox

Two uniform accounts of the paradoxes of self-reference are the closest relatives of the schema. Yanofsky (2003), following Lawvere (1969), shows that they share a diagonal argument: a transformation without a fixed point is applied along a self-referential map, and the absence of the fixed point is the contradiction. In the present account that argument appears as Theorem 6.1, on the collapsed object, where F has no fixed point. Priest (2002) presents the paradoxes through the inclosure schema, as limits that are transcended and closed at once, and takes the resulting contradictions to be true. The identification schema locates the same constructions one step earlier. The content of each is coherent, the diagonal argument runs only after an identification has been added, and the identification has a status that can be read off the content before it is added. What the schema supplies is a classification the other accounts do not attempt: which identifications are safe, which are choices with consequences, and which are refuted.

10. Revenge

10.1 Revenge Sentences

A revenge sentence quantifies over evaluating perspectives from within one: "this sentence is not true at any perspective." In the role-indexed representation the sentence R has a pointer occurrence R^Ptr(e) at the evaluating perspective e and a target occurrence R^Tgt(q) at every perspective q in the domain U ⊆ P quantified over. Which targets e may use is fixed by the guard: the effective domain is

The content of the sentence at e is a conjunction over the effective domain: R^Ptr(e) is true exactly when every accessible target is false. It is recorded as conjunctive edges in the sense of Section 8.1, and a valuation is coherent for the revenge dependency when it satisfies the biconditional

As with the liar, the content contributes no identity edge: nothing in the sentence says that R^Ptr(e) and R^Tgt(e) carry one value. The phrase "any perspective" can denote all modeled perspectives, all perspectives accessible from e, or an unrestricted totality, and these choices give different effective domains and different dependencies. A revenge sentence therefore has a determinate dependency only once its evaluating perspective, its domain, and the guard values are fixed; the results below hold for every such choice.

10.2 Self-Access

Self-access is the liar's identity edge at the level of the guard. When the evaluator's only accessible target is its own, the revenge sentence is the liar, and its identification is forbidden by Theorem 8.3 (i). When other targets are accessible, the single identification is open, and every valuation that makes it also makes the sentence false at e and true at some other target. What the revenge argument needs is more than the single identification. It runs: by the account, R has the value true at no perspective; so what R says is the case; so R is true; so R is true at e, against the first step. The first step identifies R^Ptr(q) with R^Tgt(q) at every perspective, and the third identifies R^Ptr(e) with them. The argument makes the joint self-identification, and by Theorem 10.1 (c) that identification is forbidden. Revenge is decided by the same criterion as the liar, and it is the joint case of Theorem 8.3, alongside the odd cycles and Yablo.

10.3 Worked Revenge Model

Let P ⊇ {e, a, b}, let U = {a, b}, and let CPA_e(a, R^Tgt(a)) = CPA_e(b, R^Tgt(b)) = 1. In a coherent valuation with both accessible targets false, R^Ptr(e) is true. No self-identification is in play, since e ∉ U_e, and no contradiction follows.

Now take U' = {e, a, b} with CPA_e(e, R^Tgt(e)) = 1 as well. The dependency is still coherent: R^Tgt(e) false and R^Ptr(e) true satisfies it. The self-identification at e is open: the valuation with R^Ptr(e) false, R^Tgt(e) false, and R^Tgt(a) true is coherent and agrees. Let a and b evaluate R as well, each with effective domain U'. The joint identification of the three self-pairs is forbidden by Theorem 10.1 (c). An evaluator that adds it, natively or by adopting judgments that assemble it, has no coherent valuation, and by Theorem 9.2 (3) the record of its derivation names the entries and the sources.

10.4 Revenge on the Unfolding

The sharpest revenge sentence for the present account quantifies over the positions of the unfolding itself: "this sentence is false at every position." Its content at each position r_n is that r_n^Ptr is true exactly when every target r_m^Tgt in the quantified range is false; with the range restricted to later positions, this is Yablo's paradox with its occurrences split, and Section 9.2 (e) applies. Without identity edges the dependency is coherent, each single identification is open, and the full set is forbidden. By Theorem 9.2 (2), any derivation of the contradiction contains identity entries. The revenge sentence behaves exactly as the liar does under the licensing criterion: its content is coherent, and what is forbidden is the classical evaluator's identification of its occurrences.

10.5 The Diagonal Reading

Theorem 10.1 is a diagonal fact, and it should be read as one. For every evaluating perspective in P there is a revenge sentence whose self-identification is forbidden at that perspective, and for every community of perspectives with self-access the joint self-identification is forbidden. The account stores this parameter in a different place from a hierarchy of languages. There the restriction is imposed uniformly on the language, and every speaker inherits it. Here access is a guard value of a perspective: two perspectives in one model may have different guards, one may include itself and face a forbidden identification while another excludes itself and does not, and the consequences of each choice are theorems about that perspective. The meta-theory surveys all of them. It is not a member of P, has no guards, and is the subject of no revenge sentence formulated over P. The meta-theory can itself be made a perspective and subjected to a further revenge sentence, at the cost of a larger perspective set, in which the same theorem holds again.

11. Comparison with Prior Accounts

For each approach the table records where the parameter omitted by the classical derivation is stored, whether the party that performs the collapse is represented, what the approach says about the liar, and how it relates to the present account.

11.1 Neighbors of the Machinery

The formal apparatus has relatives outside the theory of truth. Nested epistemic operators (Fagin, Halpern, Moses, and Vardi 1995) can express that one agent holds a relation between what two other agents hold; the subjective-number judgment a(s) ⊑_p b(t) expresses the same content with the perspective of each value carried by the value itself, so that the value is an object and the judgment a relation between objects. Trust logics in which one agent trusts another about a topic (Liau 2003) have a guard of the same arity as CPA_p(q, u); here the topic is the position at which an adopted judgment lands, and Theorem 5.1 shows that for the liar the topic-relativity of trust does not isolate the roles. Justification logic (Artemov 2001) annotates formulas with the evidence that supports them, and provenance semirings (Green, Karvounarakis, and Tannen 2007) annotate data with its derivation. The provenance functor belongs to this family, and it is what allows Example 5.1 to distinguish a perspective that collapsed on its own from one that inherited a collapse.

None of these neighbors supplies the combination used above: a value with an intrinsic perspective, a judgment with a separate evaluating perspective, and a guard on composition indexed by the landing position. The Observer Theorem and the Transmission Theorem depend on that combination.

11.2 Local Classicality

Inside a fixed position, Boolean reasoning proceeds normally. Explosion is avoided because the liar edges alone never place both x and ¬x at one position under one evaluator. A contradiction appears at an evaluator only when that evaluator has identified the positions, and it appears then as a same-position negating judgment whose derivation is recorded. No fixed indexed position validates both a value and its negation unless its evaluator has collapsed, and an evaluator that has collapsed is one whose kernel or adopted judgments say so.

Slide Deck

↗ 12 slides

12. Conclusion

The liar sentence occupies two positions of one dependency, and its content is a pair of negating edges between them. An evaluating perspective that holds those edges holds the whole dependency and admits exactly two coherent valuations: at that perspective the liar is the open pair. A contradiction appears when, and only when, the perspective's closure contains an identity edge between the positions, the judgment that one occurrence has the value of the other. That judgment is contributed by neither the demonstrative nor the predicate. It is contributed by the evaluator, natively or by adoption through a guard, and every derivation of the contradiction names it and its source.

The two positions are themselves a quotient. Following the reference of the sentence produces an infinite ray, every quotient of that ray that keeps the two parities apart is coherent, every quotient that merges them is not, and the two-role structure is the coarsest coherent quotient. The classical derivation is the passage from two classes to one. Mixed reference cycles obey the same parity, which is the classical criterion that the dangerous finite patterns are the odd cycles. Made into a number, the liar is the unit σ with σ²=1 whose two readings are the two values the classical evaluator tried to merge, and re-reading it around the loop is the conjugation that records the refusal.

Bivalence, disquotation, self-reference, and classical inference are all retained at every position. The one principle given up is the identification of a sentence's value across its two occurrences where that identification is forbidden, and for the liar it is forbidden by the sentence's own content. To resolve the paradox with this machinery is to represent the evaluator, locate the identification in its kernel, decide by the coherent valuations whether it is licensed, forbidden, or open, and trace it in its provenance when it is made anyway. The same procedure decides Grelling, Curry, Russell, Berry, Yablo, and the revenge liar, and it says in each case which identification the classical reading made and what status it had.

The meta-theory sees every perspective and belongs to none. It has a global view of the liar and no global section for it, and the distinction between those two is the distinction between representing a multiplicity of positions and demanding one value across them. The paradox lives in the second demand, and the demand has an address: the kernel of whoever makes it.

References

• Abramsky, Samson, Rui Soares Barbosa, Kohei Kishida, Raymond Lal, and Shane Mansfield. "Contextuality, Cohomology and Paradox." 24th EACSL Annual Conference on Computer Science Logic (CSL 2015), Leibniz International Proceedings in Informatics 41, 2015, 211–228.
• Abramsky, Samson, and Adam Brandenburger. "The Sheaf-Theoretic Structure of Non-Locality and Contextuality." New Journal of Physics 13, no. 11, 2011, 113036.
• Aczel, Peter. Non-Well-Founded Sets. CSLI Lecture Notes 14, 1988.
• Armour-Garb, Bradley, and James A. Woodbridge. "Dialetheism, Semantic Pathology, and the Open Pair." Australasian Journal of Philosophy 84, no. 3, 2006, 395–416.
• Artemov, Sergei. "Explicit Provability and Constructive Semantics." Bulletin of Symbolic Logic 7, no. 1, 2001, 1–36.
• Aspvall, Bengt, Michael F. Plass, and Robert Endre Tarjan. "A Linear-Time Algorithm for Testing the Truth of Certain Quantified Boolean Formulas." Information Processing Letters 8, no. 3, 1979, 121–123.
• Barwise, Jon, and John Etchemendy. The Liar: An Essay on Truth and Circularity. Oxford University Press, 1987.
• Beringer, Timo, and Thomas Schindler. "A Graph-Theoretic Analysis of the Semantic Paradoxes." Bulletin of Symbolic Logic 23, no. 4, 2017, 442–492.
• Burge, Tyler. "Semantical Paradox." The Journal of Philosophy 76, no. 4, 1979, 169–198.
• Cook, Roy T. "Patterns of Paradox." The Journal of Symbolic Logic 69, no. 3, 2004, 767–774.
• Cook, Roy T. The Yablo Paradox: An Essay on Circularity. Oxford University Press, 2014.
• Curry, Haskell B. "The Inconsistency of Certain Formal Logics." The Journal of Symbolic Logic 7, no. 3, 1942, 115–117.
• Dyrkolbotn, Sjur, and Michał Walicki. "Propositional Discourse Logic." Synthese 191, no. 5, 2014, 863–899.
• Fagin, Ronald, Joseph Y. Halpern, Yoram Moses, and Moshe Y. Vardi. Reasoning About Knowledge. MIT Press, 1995.
• Field, Hartry. Saving Truth from Paradox. Oxford University Press, 2008.
• Gaifman, Haim. "Pointers to Truth." The Journal of Philosophy 89, no. 5, 1992, 223–261.
• Glanzberg, Michael. "A Contextual-Hierarchical Approach to Truth and the Liar Paradox." Journal of Philosophical Logic 33, no. 1, 2004, 27–88.
• Green, Todd J., Grigoris Karvounarakis, and Val Tannen. "Provenance Semirings." Proceedings of the 26th ACM Symposium on Principles of Database Systems (PODS 2007), 2007, 31–40.
• Grelling, Kurt, and Leonard Nelson. "Bemerkungen zu den Paradoxieen von Russell und Burali-Forti." Abhandlungen der Fries'schen Schule, Neue Folge 2, 1908, 301–334.
• Gupta, Anil. "Critical Notice: Jon Barwise and John Etchemendy, The Liar: An Essay on Truth and Circularity." Philosophy of Science 56, no. 4, 1989, 697–709.
• Gupta, Anil, and Nuel Belnap. The Revision Theory of Truth. MIT Press, 1993.
• Harary, Frank. "On the Notion of Balance of a Signed Graph." Michigan Mathematical Journal 2, no. 2, 1953, 143–146.
• Kauffman, Louis H. "Time, Imaginary Value, Paradox, Sign and Space." In Computing Anticipatory Systems: CASYS 2001, AIP Conference Proceedings 627, American Institute of Physics, 2002.
• Kripke, Saul. "Outline of a Theory of Truth." The Journal of Philosophy 72, no. 19, 1975, 690–716.
• Lawvere, F. William. "Diagonal Arguments and Cartesian Closed Categories." In Category Theory, Homology Theory and Their Applications II, Lecture Notes in Mathematics 92, Springer, 1969, 134–145.
• Leitgeb, Hannes. "What Truth Depends On." Journal of Philosophical Logic 34, no. 2, 2005, 155–192.
• Liau, Churn-Jung. "Belief, Information Acquisition, and Trust in Multi-Agent Systems: A Modal Logic Formulation." Artificial Intelligence 149, no. 1, 2003, 31–60.
• MacFarlane, John. Assessment Sensitivity: Relative Truth and Its Applications. Oxford University Press, 2014.
• Penrose, Roger. "On the Cohomology of Impossible Figures." Leonardo 25, no. 3/4, 1992, 245–247. Originally published in Structural Topology 17, 1991, 11–16.
• Priest, Graham. Beyond the Limits of Thought. 2nd ed., Oxford University Press, 2002.
• Priest, Graham. In Contradiction: A Study of the Transconsistent. 2nd ed., Oxford University Press, 2006.
• Quine, W. V. "New Foundations for Mathematical Logic." The American Mathematical Monthly 44, no. 2, 1937, 70–80.
• Rabern, Landon, Brian Rabern, and Matthew Macauley. "Dangerous Reference Graphs and Semantic Paradoxes." Journal of Philosophical Logic 42, no. 5, 2013, 727–765.
• Russell, Bertrand. The Principles of Mathematics. Cambridge University Press, 1903.
• Russell, Bertrand. "Mathematical Logic as Based on the Theory of Types." American Journal of Mathematics 30, no. 3, 1908, 222–262.
• Simmons, Keith. Universality and the Liar: An Essay on Truth and the Diagonal Argument. Cambridge University Press, 1993.
• Sorensen, Roy. Vagueness and Contradiction. Oxford University Press, 2001.
• Spencer-Brown, George. Laws of Form. Allen and Unwin, 1969.
• Tarski, Alfred. "The Concept of Truth in Formalized Languages." In Logic, Semantics, Metamathematics, Oxford University Press, 1956, 152–278. Originally published 1935.
• Walicki, Michał. "Reference, Paradoxes and Truth." Synthese 171, no. 1, 2009, 195–226.
• Wen, Lan. "Semantic Paradoxes as Equations." The Mathematical Intelligencer 23, no. 1, 2001, 43–48.
• Yablo, Stephen. "Paradox without Self-Reference." Analysis 53, no. 4, 1993, 251–252.
• Yanofsky, Noson S. "A Universal Approach to Self-Referential Paradoxes, Incompleteness and Fixed Points." Bulletin of Symbolic Logic 9, no. 3, 2003, 362–386.