Logic: Foundations & Applications

• Logic is a mathematically precise study of formal languages, semantics, and proof systems that underpins mathematics, computer science, and AI.
• Different perspectives define logic through internal syntactic structures, representational constraints, and model-theoretic analysis to capture logical consequence.
• Logic informs diverse applications including logic programming, formal verification, and AI integration by providing frameworks for reasoning and computational design.

Searching arXiv for recent and relevant papers on logic across foundations, programming, and AI. arXiv_search(query="all:logic mathematics computer science ai programming representation survey", max_results=10, sort_by="relevance") Refining the search to target broad treatments of logic in mathematics, computer science, and AI. arXiv_search(query="ti:\"Logic\" OR ti:\"logic\" AND (abs:\"mathematics\" OR abs:\"computer science\" OR abs:\"AI\" OR abs:\"programming\")", max_results=10, sort_by="relevance") Logic is the mathematically precise study of formal languages, their semantics and proof systems, and the ways these structures are used to analyze mathematics and underwrite core parts of computer science (Zach, 2024). Contemporary work, however, does not treat logic as a single doctrine. One line characterizes it as the “internal organization of language,” with logical constants, logical truths, and logical consequence grounded in internal syntactic and semantic structure rather than in an ontology of “all possible interpretations” (Čulina, 2021). Another presents logic as the “abstract theory of representation,” centered on the formal constraints that any adequate representational system must satisfy (Plagnol, 2023). In computation, logic programming has been recast as programming in a logic of inductive definitions rather than in the Horn clause fragment of classical first-order logic (Denecker et al., 2023). In artificial intelligence, logic functions both as a computational substrate and as a framework for learning and for reasoning about learned systems (Darwiche, 2020).

1. Historical formation and conceptual scope

Modern symbolic logic was developed in part to provide a formal framework for mathematics. Frege, Peano, Whitehead and Russell, and Hilbert developed logical systems either as foundations of mathematics or as formal analogs of mathematical reasoning amenable to mathematical study, including Hilbert’s consistency program (Zach, 2024). From this emerged the now-standard distinction between syntax, semantics, and proof; the metatheoretic study of soundness, completeness, compactness, and undecidability; and later the mechanization of inference in theorem proving, verification, and proof assistants (Zach, 2024).

Within that broad field, contemporary research presents competing descriptions of what logic fundamentally is. On one account, logic is “always the logic of a language”: a normative, language-relative discipline governing how truths may properly determine other truths through the internal semantic structure of connectives and quantifiers (Čulina, 2021). On another, valid inference is secondary to representation itself: logic studies the structural conditions under which systems of representations can adequately present a universe, and laws of logic are accordingly construed as laws of representation (Plagnol, 2023). These conceptions do not eliminate more familiar model-theoretic and proof-theoretic approaches; rather, they relocate them within broader debates about language, meaning, and representation.

A further controversy concerns paradox and self-reference. A revisionary proposal argues that liar-like paradoxes, Russell’s paradox, Gödel’s incompleteness theorems, and the halting problem all rely on violations of a sharpened law of identity, and it reformulates logic around four set-theoretic principles: “Don’t talk about empty,” “Elements of set should have identity,” “Sets should have definitions,” and “A total set should be defined” (Yang, 2023). This is a distinctly revisionary stance, but it illustrates that even the boundaries of meaningful logical discourse remain contested.

2. Meaning, logical constants, and consequence

A central issue in logic is how to characterize logical constants and logical consequence. The “received view,” associated with Tarski and later invariance accounts, identifies logicality through invariance under all interpretations or permutations of domains. An internal alternative instead defines a symbol as logical when its semantic rule “does not refer to the reality the language speaks of, except possibly referring to external assumptions of the language use” (Čulina, 2021). On this basis, logical truth is truth “determined by the internal semantic structure of the language,” and logical consequence is truth preservation determined by that internal structure rather than by the subject matter described (Čulina, 2021).

For first-order language, this internal approach yields a sharp classification. Propositional connectives are logical constants because they are given by Boolean functions on truth values. Quantifiers are logical when they are determined by functions from non-empty sets of truth values to a truth value, yielding exactly $2^3 = 8$ logical quantifiers of type $\langle 1 \rangle$; among them, $\forall$ and $\exists$ form a functionally complete set, and with negation either one alone is already functionally complete (Čulina, 2021). By contrast, neither equality nor cardinal quantifiers count as logical constants on this criterion: equality requires object-level identity facts, and quantifiers such as “there is exactly one” or “there are at least $n$” require counting or comparing objects in the domain (Čulina, 2021).

A representational analysis reframes these same notions. In that setting, conjunction expresses co-presence, disjunction organizes alternatives across possible worlds, negation marks incompatibility with a reference situation, and implication functions as a symbolic shortcut from one situation to another (Plagnol, 2023). The adequacy of a representational system is then assessed by three logical properties: completeness, faithfulness, and coherence (Plagnol, 2023). This suggests that logical constants can be understood not only as semantically invariant expressions, but also as content-insensitive operators governing how representations may be combined, extended, and contrasted.

3. Proof theory, model theory, and infinitary development

Classical logic is distinguished by the interaction of proof theory and model theory. Proof theory studies calculi such as Hilbert systems, natural deduction, sequent calculus, and tableaux; model theory studies interpretations, truth, and consequence, formalized as $\Gamma \vdash A$ and $\Gamma \vDash A$, with soundness and completeness relating them (Zach, 2024). Completeness yields model existence theorems and compactness; compactness, in turn, yields non-standard models of arithmetic and the non-categoricity of first-order theories of infinite structures (Zach, 2024). Quantifier elimination provides decidability results for theories such as real closed fields, while proof-theoretic transformations support ordinal analysis, reverse mathematics, and proof mining (Zach, 2024).

Infinitary logic extends this landscape. Standard $L_{\lambda,\theta}$ logics allow conjunctions and disjunctions of length $<\lambda$ and quantifier blocks of length $<\theta$. Shelah’s $\langle 1 \rangle$0, defined not by an ordinary syntax but via equivalence classes of models under the borrowing game $\langle 1 \rangle$1, is characterized for $\langle 1 \rangle$2 as the maximal logic above $\langle 1 \rangle$3 satisfying Strong Undefinability of Well Order (Džamonja et al., 2019). Karp’s chain logic takes a different route: formulas are ordinary $\langle 1 \rangle$4-formulas, but models are replaced by weak or proper chain models $\langle 1 \rangle$5 with modified witness conditions for quantifiers. For $\langle 1 \rangle$6 singular of countable cofinality, chain logic lies above $\langle 1 \rangle$7, satisfies Interpolation and $\langle 1 \rangle$8, has a Union Lemma and a Completeness Theorem, yet is not $\langle 1 \rangle$9-compact (Džamonja et al., 2019). These results show that major logical properties can depend as much on the notion of model as on the syntax of formulas.

4. Nonclassical, graded, evidential, and algebraic logics

The field of logic includes many systems that weaken, refine, or reinterpret classical consequence.

Family	Characteristic feature	Paper
Paradefinite logic	Belnap-Dunn expansion with falsity and implication; drops LNC and LEM while preserving the standard deduction theorem	(Middelburg, 7 Jan 2026)
Rational fuzzy attribute logic	Graded entailment for rational truth-degree if-then formulas; Pavelka complete; decidable fragment of RPL or $\forall$0	(Vychodil, 2015)
Affine and Łukasiewicz logics	$\forall$1 gives commutativity of weak conjunction and restricted contraction; $\forall$2 yields classical Łukasiewicz logic	(Arthan et al., 2014)
Logic of evidence	Language with $\forall$3, $\forall$4, and $\forall$5; sound and complete axiomatization; doubly-exponential full decision problem and NP-complete propositional fragment	(Halpern et al., 2014)
Eigenlogic	Logical operators as commuting projection operators; truth values as eigenvalues $\forall$6	(Toffano, 2015)

Paradefinite logic is designed for theories that may be inconsistent or incomplete. A 2026 proposal argues that the most natural paradefinite logic relative to classical logic is the expansion of Belnap-Dunn logic with a falsity connective and an implication connective for which the standard deduction theorem holds; the resulting system is obtained from a classical sequent-style presentation by retaining the ordinary clauses for $\forall$7 and omitting exactly the law of non-contradiction and the law of excluded middle (Middelburg, 7 Jan 2026). This preserves maximal agreement with classical logic except where paradefiniteness requires otherwise.

Graded and many-valued logics extend consequence to degrees. Rational fuzzy attribute logic studies implications $\forall$8 between finite rational fuzzy sets of attributes, defines graded semantic and syntactic entailment, proves Pavelka-style completeness, and characterizes entailment via least models and closure operators (Vychodil, 2015). A related substructural line analyzes affine logic and Łukasiewicz logic: intuitionistic Łukasiewicz logic is obtained by adding the schema $\forall$9 to intuitionistic affine logic, and classical Łukasiewicz logic is then the extension of that system by $\exists$0 (Arthan et al., 2014). The same paper shows that the usual negative translations remain correct for intuitionistic Łukasiewicz logic even though full contraction is absent (Arthan et al., 2014).

Logic also has quantitative evidential forms. A logic for reasoning about evidence treats evidence as a function from priors to posteriors, with operators $\exists$1, $\exists$2, and $\exists$3, a sound and complete axiomatization, and complexity bounds ranging from doubly-exponential time for the full quantified language to NP-completeness for a propositional fragment (Halpern et al., 2014). At the opposite end of the spectrum, Eigenlogic reconstructs propositional logic in finite-dimensional linear algebra: logical operators are commuting projection operators, truth values are the eigenvalues $\exists$4 and $\exists$5, and Boole’s multilinear elective decomposition becomes a spectral decomposition over tensored elementary projectors (Toffano, 2015).

5. Logic in programming, verification, and type theory

In programming languages and formal methods, logic serves both as a semantics of computation and as an executable proof theory. A major reappraisal of logic programming argues that logic programs are not Horn-clause theories of classical first-order logic but theories in a logic of inductive definitions, together with a Herbrand Axiom stating that function symbols are constructors of the universe (Denecker et al., 2023). On this view, prototypical Prolog programs such as member/2 and append/3 are formal inductive definitions, the least Herbrand model is recovered as the unique model of a definitional theory $\exists$6, and well-founded semantics reflects a broader theory of non-monotone inductive definitions (Denecker et al., 2023).

This interaction between logic and programming becomes especially explicit in constraint-logic object-oriented programming. Work on Muli extends logic variables from primitive types to reference types, introducing set-based applicable type constraints $\exists$7, structural equality constraints on objects, and special choice points for method invocation, type tests, casts, and structural equality (Dageförde, 2018). The result is a logically disciplined account of symbolic execution over Java-like class hierarchies, where objects themselves become logical unknowns and the constraint store integrates arithmetic, type, and structural constraints (Dageförde, 2018).

Logic also structures verification and typed programming. Automated reasoning uses Herbrand reductions, unification, SAT, and resolution; program verification uses Floyd–Hoare logic and temporal logics such as LTL and CTL; and type systems realize the Curry–Howard correspondence, under which propositions correspond to types, proofs to programs, and proof normalization to program evaluation (Zach, 2024). In current practice, these ideas support proof assistants based on higher-order or dependent type theory, as well as computer verification at the level of programs, protocols, and mathematical proofs (Zach, 2024).

6. Logic in AI, machine learning, and learned reasoning

In contemporary AI, logic is not confined to symbolic theorem proving. One influential account identifies three modern roles for logic in AI: logic as a basis for computation, logic for learning from a combination of data and knowledge, and logic for reasoning about the behavior of machine learning systems (Darwiche, 2020). The common technical substrate is tractable Boolean circuits—NNF circuits refined into DNNF, d-DNNF, SDDs, and OBDDs—together with knowledge compilation. Under decomposability, determinism, smoothness, and structured decomposability, tasks such as satisfiability, weighted model counting, marginals, MAP-style queries, explanation, robustness analysis, and bias analysis become polynomial in circuit size (Darwiche, 2020).

This computational role now extends inside learned architectures. LogicCBMs enhance concept bottleneck models by inserting a differentiable propositional logic module between concepts and labels, using fuzzy logical operators such as AND, OR, XOR, NOT, and implication over learned concept activations (Vemuri et al., 8 Dec 2025). The model remains end-to-end trainable, but the concept-to-label map is no longer merely linear: each class logit becomes a weighted combination of learned predicates. On CUB, LogicCBM reports 81.13% validation accuracy versus 75.20% for a vanilla CBM, and it also reports higher Concept Correction Gain and more effective interventions (Vemuri et al., 8 Dec 2025). This indicates a current tendency to use logic not only for post hoc explanation but as part of the predictive architecture itself.

A complementary line asks whether logical regularities can be learned directly from perceptual data without preset reasoning patterns. “Logic could be learned from images” formulates LiLi tasks in which two input images encode numbers and the output image is related by Bitwise And, Bitwise Or, Bitwise Xor, Addition, Subtraction, or Multiplication, while the model is given neither digit semantics nor the underlying relation (Guo et al., 2019). Standard neural models learn the easier bitwise tasks and, with enough data, addition and subtraction, but perform poorly on multiplication; a divide-and-conquer model that adds label information and separates carry and non-carry subtasks achieves high testing accuracy on the difficult logic task (Guo et al., 2019). The broader implication is that logic can appear both as an explicit symbolic formalism and as a latent rule-like structure recoverable from data, though the harder cases still benefit from architectural decomposition aligned with the logical structure of the task.