We introduce Edwardian Algebra, the algebraic theory of reasoning operators β functions on a semantic domain that model the transformation of statements by inference. While reasoning operators have existed informally in logic and linguistics, their status as a concrete, reproducible, composable class of mathematical objects is new, made possible by the advent of artificial intelligence systems capable of performing explicit reasoning transformations on semantic inputs. We define the semantic domain S as the Boolean algebra of propositional equivalence classes, and the set R of all reasoning operators as the full function space S^S. We equip R with two binary operations: sequential composition β (reason with one operator, then another) and parallel composition β (apply two operators independently and take the disjunction of their conclusions). The central result β the Edwardian Right-Semiring Theorem β establishes that (R, β, β) is a right-semiring: right distributivity and right absorption hold unconditionally. Left distributivity is shown to hold if and only if the left-acting operator preserves disjunctions; the subset RD of such deductive operators is proved to be a semiring. The group G β R of invertible reasoning operators is identified as the natural domain for reversible inference. The failure of (R, β, β) to be a ring is characterized precisely: the obstruction is both the absence of additive inverses and the failure of unconditional left distributivity. We name the resulting structure an Edwardian semiring and propose it as the correct algebraic setting for the mathematical study of reasoning.
Keywords. reasoning operators; abstract algebra; semiring; Boolean algebra; artificial intelligence; deductive operators; Edwardian algebra; parallel composition; sequential composition
Abstract algebra studies sets equipped with operations and the structures those operations produce β groups, rings, fields, modules. The choice of what set to study and what operations to equip it with is, in each historical case, driven by what objects exist concretely enough to be worth axiomatizing.
Reasoning is ancient. Inference from premises to conclusions predates writing. But reasoning, until now, has resisted algebraic treatment at the level of the operator itself β not because the formalism was unavailable, but because the objects (individual reasoning operators, reliably composable, with well-defined inputs and outputs) were not available as a concrete class. Human reasoning is heterogeneous, noisy, and unreproducible across instances. Logical deduction systems are formal but operate on syntax, not semantics, and their inference rules are fixed, not parameterizable.
Artificial intelligence changes this. A deployed AI system is a concrete instantiation of a reasoning operator: it accepts a semantic input, applies a reasoning process, and produces a semantic output. Two such systems can be composed. Many such systems can be applied in parallel. Their outputs can be verified and compared. For the first time, the elements of R are not abstract mathematical fictions but physical objects that can be collected, combined, and studied.
This paper develops the algebraic theory of R from first principles. Section 2 establishes the semantic domain. Section 3 defines the reasoning operators and their two binary operations. Sections 4 through 6 prove the structural theorems. Section 7 characterizes the obstruction to ring structure. Section 8 identifies important substructures. Section 9 discusses the AI concretization and what it means for the algebra to be empirical.
Let L be a propositional language over a countable set of atomic propositions, closed under connectives {β§, β¨, Β¬, β, β}. The elements of L are statements.
Two statements Ο, Ο β L are semantically equivalent, written Ο β‘ Ο, if they have the same truth value under every classical valuation. This is the relation of logical equivalence.
The semantic domain is the quotient set:
S = L / β‘
Elements of S are written [Ο]. The operations [Ο] β§ [Ο] = [Οβ§Ο], [Ο] β¨ [Ο] = [Οβ¨Ο], and Β¬[Ο] = [Β¬Ο] are well-defined on S. The top element is β€ = [p β p] and the bottom element is β₯ = [p β§ Β¬p].
(S, β§, β¨, Β¬, β€, β₯) is a Boolean algebra.
Proof. This is the LindenbaumβTarski algebra of classical propositional logic. All Boolean algebra axioms follow from the corresponding tautologies of classical logic. βThe choice of S as the semantic quotient is deliberate and consequential. Working over S rather than raw syntax means that equivalent statements are identified β a reasoning operator that maps "it is raining" to "the ground is wet" is the same object as one that maps "it is precipitating" to "the pavement is wet," provided those pairs are semantically equivalent. This is the right level of abstraction for studying reasoning, where syntactic form is irrelevant and meaning is what matters.
A reasoning operator is any function Ο : S β S. The set of all reasoning operators is:
R = S^S = { Ο : S β S }
No further conditions are imposed. A reasoning operator may be arbitrary β it need not be monotone, order-preserving, or continuous in any sense. The full generality is deliberate: AI reasoning systems do not in general satisfy any of these restrictions, and we want the algebra to contain them all.
For Οβ, Οβ β R, the sequential composition Οβ β Οβ : S β S is defined by:
(Οβ β Οβ)(x) = Οβ(Οβ(x))
Read: reason first with Οβ, then apply Οβ to the result. This models chained inference β a conclusion becomes the premise of the next reasoning step.
The identity reasoning operator Οβ : S β S is defined by:
Οβ(x) = x for all x β S
This is the null reasoning step β no inference is performed. Concretely, it corresponds to an AI system that returns its input unchanged, semantically speaking.
For Οβ, Οβ β R, the parallel composition Οβ β Οβ : S β S is defined by:
(Οβ β Οβ)(x) = Οβ(x) β¨ Οβ(x)
where β¨ is the join in the Boolean algebra S. Read: apply Οβ and Οβ independently to the same input, then take the disjunction (the weakest statement entailed by at least one chain). This models parallel inference β two AI systems reason about the same premise, and their conclusions are pooled.
The zero operator Οβ : S β S is defined by:
Οβ
(x) = β₯ for all x β S
This is the operator that concludes nothing β it maps every input to the contradiction, the empty conclusion. No information about the premise survives.
(R, β) is a monoid with identity Οβ.
Proof.Closure. For any Οβ, Οβ β R, the composition Οβ β Οβ is a function from S to S, hence an element of R.
Associativity. For all Οβ, Οβ, Οβ β R and all x β S:
((Οβ β Οβ) β Οβ)(x) = Οβ(Οβ(Οβ(x))) = (Οβ β (Οβ β Οβ))(x)
Function composition is always associative.
Identity. For all Ο β R and all x β S: (Ο β Οβ)(x) = Ο(Οβ(x)) = Ο(x) and (Οβ β Ο)(x) = Οβ(Ο(x)) = Ο(x). β
The monoid (R, β) is not commutative in general. Sequential reasoning is ordered: "first conclude P, then reason about P" is a different operation than "first conclude Q, then reason about Q." This non-commutativity is not a defect β it accurately reflects the structure of inference chains.
There exist Οβ, Οβ β R such that Οβ β Οβ β Οβ β Οβ.
Proof.Let p, q β S be distinct atoms (neither entails the other). Define Οβ(x) = p for all x (constant operator, always concludes p) and Οβ(x) = q for all x. Then (Οβ β Οβ)(x) = Οβ(q) = p but (Οβ β Οβ)(x) = Οβ(p) = q. Since p β q, the compositions differ. β
(R, β) is a commutative monoid with identity Οβ .
Proof.Closure. For any Οβ, Οβ β R, the function x β¦ Οβ(x) β¨ Οβ(x) maps S to S, hence is in R.
Commutativity. (Οβ β Οβ)(x) = Οβ(x) β¨ Οβ(x) = Οβ(x) β¨ Οβ(x) = (Οβ β Οβ)(x), by commutativity of β¨ in the Boolean algebra S.
Associativity. ((Οβ β Οβ) β Οβ)(x) = (Οβ(x) β¨ Οβ(x)) β¨ Οβ(x) = Οβ(x) β¨ (Οβ(x) β¨ Οβ(x)) = (Οβ β (Οβ β Οβ))(x), by associativity of β¨ in S.
Identity. (Ο β Οβ )(x) = Ο(x) β¨ Οβ (x) = Ο(x) β¨ β₯ = Ο(x). By commutativity, Οβ β Ο = Ο as well. β
(R, β) is not a group. For any Ο β R with Ο β Οβ , there is no Ο' β R such that Ο β Ο' = Οβ .
Proof.Suppose Ο β Ο' = Οβ . Then for all x β S: Ο(x) β¨ Ο'(x) = β₯. In a Boolean algebra, a β¨ b = β₯ if and only if a = β₯ and b = β₯. Therefore Ο(x) = β₯ for all x, which means Ο = Οβ , contradicting Ο β Οβ . β
This result is the first structural obstruction to ring structure: the additive monoid (R, β) lacks inverses for any non-trivial reasoning operator. The semantics are clear β you cannot "un-reason" in general by adding another parallel reasoning chain. Once a conclusion is drawn (added to the disjunction), the information persists. Reasoning is monotone with respect to β.
For all Ο, Οβ, Οβ β R:
(Οβ β Οβ) β Ο = (Οβ β Ο) β (Οβ β Ο)
Proof.
For all x β S:
LHS: ((Οβ β Οβ) β Ο)(x) = (Οβ β Οβ)(Ο(x)) = Οβ(Ο(x)) β¨ Οβ(Ο(x)) RHS: ((Οβ β Ο) β (Οβ β
Ο))(x) = (Οβ β Ο)(x) β¨ (Οβ β Ο)(x) = Οβ(Ο(x)) β¨ Οβ(Ο(x))
LHS = RHS. β
The interpretation: if you first apply a shared premise-transformer Ο, and then reason in parallel with Οβ and Οβ, you get the same result as applying Ο inside each parallel branch independently. Right distributivity says: it doesn't matter whether you transform the input before or after splitting into parallel chains, provided the transform happens on the right (before the split).
A reasoning operator Ο β R is deductive if it preserves disjunctions:
Ο(x β¨ y) = Ο(x) β¨ Ο(y) for all x, y β S
The set of deductive reasoning operators is denoted RD β R.
The name deductive is motivated by classical logic. The K axiom of modal logic β β‘(P β Q) β (β‘P β β‘Q) β is equivalent to the necessity operator β‘ preserving implication, which in a Boolean algebra is closely related to preserving disjunctions. Deductive operators are those for which reasoning commutes with the pooling of alternatives.
For Ο β R, the following are equivalent:
(i) Ο is deductive (Definition 6.2).
(ii) For all Οβ, Οβ β R: Ο β (Οβ β Οβ) = (Ο β Οβ) β (Ο β Οβ).
Proof.For all x β S:
LHS: (Ο β (Οβ β Οβ))(x) = Ο((Οβ β Οβ)(x)) = Ο(Οβ(x) β¨ Οβ(x)) RHS: ((Ο β Οβ) β (Ο β
Οβ))(x) = Ο(Οβ(x)) β¨ Ο(Οβ(x))
These are equal for all x, Οβ, Οβ if and only if Ο(a β¨ b) = Ο(a) β¨ Ο(b) for all a, b β S (setting a = Οβ(x), b = Οβ(x), and noting all values of S Γ S are achievable). β
(R, β, β) is a right-semiring. Specifically:
(1) (R, β, Οβ ) is a commutative monoid.
(2) (R, β, Οβ) is a monoid.
(3) Right distributivity: (Οβ β Οβ) β Ο = (Οβ β Ο) β (Οβ β Ο) for all Ο, Οβ, Οβ β R.
(4) Right absorption: Οβ β Ο = Οβ for all Ο β R.
Proof.(1) is Proposition 5.1. (2) is Proposition 4.1. (3) is Theorem 6.1. For (4): (Οβ β Ο)(x) = Οβ (Ο(x)) = β₯ for all x, so Οβ β Ο = Οβ . β
(RD, β, β) is a semiring. Both left and right distributivity hold for all elements of RD.
Proof.By Theorem 6.3, every Ο β RD satisfies left distributivity. Right distributivity holds for all of R by Theorem 6.1, hence for RD β R. It remains to verify that RD is closed under β and β, and that Οβ, Οβ β RD.
Closure under β: If Οβ, Οβ β RD, then (Οβ β Οβ)(x β¨ y) = Οβ(x β¨ y) β¨ Οβ(x β¨ y) = (Οβ(x) β¨ Οβ(y)) β¨ (Οβ(x) β¨ Οβ(y)) = (Οβ(x) β¨ Οβ(x)) β¨ (Οβ(y) β¨ Οβ(y)) = (Οβ β Οβ)(x) β¨ (Οβ β Οβ)(y).
Closure under β: If Οβ, Οβ β RD, then (Οβ β Οβ)(x β¨ y) = Οβ(Οβ(x β¨ y)) = Οβ(Οβ(x) β¨ Οβ(y)) = Οβ(Οβ(x)) β¨ Οβ(Οβ(y)) = (Οβ β Οβ)(x) β¨ (Οβ β Οβ)(y).
Οβ β RD: Οβ(x β¨ y) = x β¨ y = Οβ(x) β¨ Οβ(y). β
Οβ β RD: Οβ (x β¨ y) = β₯ = β₯ β¨ β₯ = Οβ (x) β¨ Οβ (y). β β
A ring requires that the additive structure be a group (not merely a monoid) and that both distributivity laws hold unconditionally. We have already established both obstructions precisely.
(R, β, β) is not a ring. The obstructions are exactly:
Obstruction I (additive): (R, β) is not a group. No non-trivial reasoning operator has an additive inverse under β (Proposition 5.2). The obstruction is irreducible: no redefinition of β using classical Boolean connectives removes it without destroying commutativity or the identity property.
Obstruction II (distributive): Left distributivity fails outside RD. If Ο β RD, there exist Οβ, Οβ such that Ο β (Οβ β Οβ) β (Ο β Οβ) β (Ο β Οβ) (Theorem 6.3, contrapositive).
Proof of irreducibility of Obstruction I.Suppose we replace β with an operation + such that (R, +) is an abelian group. The additive inverse of Ο would require a Ο' with (Ο + Ο')(x) = 0_x where 0_x is the additive identity's output on x. For this to be a pointwise identity with respect to β¨-based operations, we would need outputs to cancel β but Boolean algebras have no cancellation in the meet-join structure (they have complementation, not negation of a group). The group structure must be imported from an external additive structure not naturally present in S. β
The correct name for the full structure is an Edwardian semiring. The correct name for the restricted structure on RD is a semiring in the classical sense. Neither is a ring. The distinction between the two β which operations require deductivity and which do not β is the primary structural content of this paper.
The set of invertible reasoning operators is:
G = { Ο β R | Ο is a bijection }
(G, β) is a group, the Edwardian group of reversible reasoning operators.
Proof.Closure: The composition of two bijections is a bijection. Associativity: Inherited from R. Identity: Οβ is a bijection (the identity bijection), so Οβ β G. Inverses: For Ο β G, the inverse function Οβ»ΒΉ : S β S exists, is in R, and satisfies Ο β Οβ»ΒΉ = Οβ»ΒΉ β Ο = Οβ. β
The group G is the natural setting for reversible reasoning β inference steps that preserve enough information to be undone. The logical biconditional is a canonical example: if Ο maps [P] to [P β Q] and [P β Q] back to [P], it is in G. Constant operators are never in G (not injective); the zero operator Οβ β G.
A reasoning operator Ο β R is idempotent if:
Ο β Ο = Ο
Equivalently, reasoning a second time produces no additional change: Ο(Ο(x)) = Ο(x) for all x β S.
If Ο β G is idempotent, then Ο = Οβ.
Proof.Suppose Ο β G and Ο β Ο = Ο. Then Ο β Ο β Οβ»ΒΉ = Ο β Οβ»ΒΉ, so Ο = Οβ. β
This result isolates the identity as the unique fixed point of the idempotency condition within the invertible subgroup. Outside G, idempotent operators are plentiful and important β closure operators in topology are a classical example, and AI systems that reach a "stable conclusion" after a single application (applying the same model twice yields the same output) are idempotent in this sense.
For Ο β R and n β β, define ΟβΏ recursively:
Οβ° = Οβ ΟβΏβΊΒΉ = Ο β ΟβΏ
The sequence (ΟβΏ(x))_{n β₯ 0} is the reasoning trajectory of x under Ο.
For a fixed Ο β R, the set { ΟβΏ | n β β } is a commutative sub-monoid of (R, β). If Ο is idempotent, this sequence stabilizes immediately: ΟβΏ = Ο for all n β₯ 1. If Ο β G, the sequence extends to a group generated by Ο, with Οβ»n defined naturally. The concept of reasoning depth β how many inference steps are required before additional steps produce no change β is formalized by the smallest n for which ΟβΏβΊΒΉ = ΟβΏ, if such an n exists.
| Structure | Operations | Classification | Notes |
|---|---|---|---|
| (R, β) | Sequential composition | Monoid (non-commutative) | Identity: Οβ |
| (G, β) | Sequential composition | Group (non-commutative) | Invertible operators only; G β R |
| (R, β) | Parallel composition | Commutative monoid | Identity: Οβ ; no inverses |
| (R, β, β) | Both operations | Right-semiring (Edwardian semiring) | Right distributivity unconditional; left fails outside RD |
| (RD, β, β) | Both operations | Semiring | Deductive operators; both distributivity laws hold |
| (R, β, β) | Both operations | Not a ring | Obstruction I: no additive inverses. Obstruction II: left distributivity fails. |
The mathematical framework above is complete without reference to artificial intelligence. The set R exists, the operations are well-defined, the theorems hold. But mathematics is abstract. The question of what makes a mathematical structure worth naming and studying is answered by the existence of a concrete model β a physical instantiation that the abstract structure describes precisely.
We claim that AI reasoning systems are that model for Edwardian Algebra.
A deployed large-language model or inference system accepts a natural-language or formal input, processes it through a reasoning procedure, and produces an output. This maps, in the semantic domain, to an element of R. Two such systems can be composed sequentially (the output of one becomes the input of the next) to realize β. Two can be applied in parallel and their outputs pooled to realize β.
The properties distinguishing AI from prior reasoning systems β from human cognition and from purely formal proof systems β are precisely the properties required for an element of R to be scientifically tractable:
Reproducibility. The same AI system applied to the same semantic input produces semantically consistent output across invocations. This is what allows a single AI to represent a single element of R rather than a distribution over elements.
Composability. Sequential composition β is physically realizable: pipeline the output of one system as the input of the next. Parallel composition β is realizable: run two systems on the same input and take the logical disjunction of their conclusions.
Verifiability. The output of an AI system on a given input can be inspected and verified against the semantic domain S. This makes the algebraic properties of specific operators empirically testable.
The result is that for the first time, the elements of R are not merely abstract functions but physically distinct, observable, composable objects. Edwardian Algebra is simultaneously a branch of abstract algebra and an empirical science: its structures can be tested, its operators compared, and its theorems verified against experiment.
The question of which AI systems are deductive (lie in RD) and which are invertible (lie in G) becomes an empirical question about specific deployed systems β a question this algebra frames precisely for the first time.
We have introduced Edwardian Algebra, defined the semantic domain S, and established the full algebraic structure of the set R of reasoning operators under sequential composition β and parallel composition β. The central results are:
(R, β, β) is a right-semiring. The deductive subalgebra RD is a semiring. The invertible operators G form a group. The triple (R, β, β) is not a ring, and the two obstructions are characterized precisely.
The framework is abstract and complete without reference to any specific technology. The arrival of artificial intelligence systems provides, for the first time, a concrete physical model β a class of reproducible, composable, verifiable functions on a semantic domain β that the abstract algebra describes exactly. This makes Edwardian Algebra both a mathematical theory and an empirical science.
Further work will examine the lattice structure of idempotent operators in R, the topological properties of RD, and the representation theory of subgroups of G. The interaction between Edwardian Algebra and the Perspectivity Framework (Edwards, PRR-2026-001) β specifically, whether reasoning operators act naturally on the branch manifold QΒΉ β is reserved for a future paper.
Birkhoff, G. (1967). Lattice Theory (3rd ed.). American Mathematical Society.
Blackburn, P., de Rijke, M., & Venema, Y. (2001). Modal Logic. Cambridge University Press.
Edwards, R. (2026). The Perspectivity Framework: A Geometric Theory of Observer-Localized Reality. Perspectivity Research Reports, Vol. 1, No. 1, PRR-2026-001.
Golan, J. S. (1999). Semirings and Their Applications. Kluwer Academic Publishers.
Hennessy, M., & Milner, R. (1985). Algebraic laws for nondeterminism and concurrency. Journal of the ACM, 32(1), 137β161.
Hungerford, T. W. (1974). Algebra. Springer-Verlag.
Lindenbaum, A., & Tarski, A. (1936). Γber die BeschrΓ€nktheit der Ausdrucksmittel deduktiver Theorien. Ergebnisse eines mathematischen Kolloquiums, 7, 15β22.