Relational Exponentiation in Categories and Databases

• Relational exponentiation is the construction of function-like spaces over relationally specified domains, bridging categorical and relational algebra approaches.
• In categorical logic, it produces exponential objects and partial products through convexity and safeness conditions, underpinning cartesian and locally cartesian closed structures.
• In database theory, it defines the higher-order operator Decision^Base to generate all candidate functions, supporting NP-complete query formulations and optimization workflows.

Searching arXiv for papers on relational exponentiation and related exponentiability in relational structures. Relational exponentiation denotes two distinct but structurally related notions in current arXiv literature. In categorical logic and semantics, it concerns exponentiable objects and morphisms in categories of models of relational Horn theories, yielding exponential objects, partial products, and sufficient conditions for cartesian closure, local cartesian closure, and quasitopos structure in categories such as preorders, posets, and certain $V$-categories (Parker, 2022). In database theory, it denotes the higher-order operation $Decision^{Base}$ in solution-set relational algebra, where the result is the set of all functions $f:Base\to Decision$ and therefore the full search space for subset selection and optimisation queries, together with translation semantics into ordinary relational algebra and complexity-theoretic consequences (Pratten et al., 8 Sep 2025). The shared theme is the construction of function-like spaces over relationally specified domains, but the semantic roles are different: one is a categorical exponentiation problem for relational structures, the other a query-algebraic operator for candidate relations.

1. Two technical uses of the term

In the cited literature, the term appears in two non-identical settings. One is the study of exponentiability in categories $\mathbb{T}\text{-}\mathsf{Mod}$ of models of a relational Horn theory $\mathbb{T}$. The other is the introduction of relational exponentiation as a primitive operation in a higher-order relational algebra for subset selection and optimisation.

Setting	Core object	Role
Relational Horn theories	Exponential object $Y^X$ or partial product $P(Y,f)$	Right adjoint to $X\times -$ or to pullback along $f$
Solution-set relational algebra	$Decision^{Base}$	Search space of all candidate relations $Decision^{Base}$0

The categorical usage is formulated over $Decision^{Base}$1-structures, $Decision^{Base}$2-morphisms, and Horn axioms. The database usage is formulated over active-domain relations, complete-domain relations, and solution sets equipped with characteristic functions. This suggests a common abstract motif—function space formation over relationally specified domains—without collapsing the two formalisms into a single theory.

2. Relational Horn theories and categorical exponentials

A relational signature $Decision^{Base}$3 is a set of relation symbols, each $Decision^{Base}$4 of finite arity $Decision^{Base}$5. A relational Horn formula over $Decision^{Base}$6 is an implication

$Decision^{Base}$7

where $Decision^{Base}$8 is a finite set of atomic $Decision^{Base}$9-edges in distinct variables, each $f:Base\to Decision$0 of the form $f:Base\to Decision$1, and $f:Base\to Decision$2 is either another $f:Base\to Decision$3-edge or an equality $f:Base\to Decision$4. A relational Horn theory $f:Base\to Decision$5 is a set of such formulas; it is said to be without equality if no axiom concludes an equality.

A $f:Base\to Decision$6-structure $f:Base\to Decision$7 consists of a set $f:Base\to Decision$8 together with, for each $f:Base\to Decision$9 of arity $\mathbb{T}\text{-}\mathsf{Mod}$0, a relation $\mathbb{T}\text{-}\mathsf{Mod}$1. A $\mathbb{T}\text{-}\mathsf{Mod}$2-morphism $\mathbb{T}\text{-}\mathsf{Mod}$3 is a function $\mathbb{T}\text{-}\mathsf{Mod}$4 preserving all relations: whenever $\mathbb{T}\text{-}\mathsf{Mod}$5, then $\mathbb{T}\text{-}\mathsf{Mod}$6. A $\mathbb{T}\text{-}\mathsf{Mod}$7-model is a $\mathbb{T}\text{-}\mathsf{Mod}$8-structure satisfying every axiom of $\mathbb{T}\text{-}\mathsf{Mod}$9. The category $\mathbb{T}$0 has these models and morphisms; it is complete and cocomplete, and the forgetful functor $\mathbb{T}$1 creates limits (Parker, 2022).

The relevant setting for exponentiability is a reflexive relational Horn theory, meaning that each $\mathbb{T}$2 is assumed reflexive via the axiom $\mathbb{T}$3. Under this assumption constant maps are automatically morphisms. The paper then studies when an object $\mathbb{T}$4 or a morphism $\mathbb{T}$5 admits the appropriate exponential or slice-exponential structure.

3. Convexity, partial products, and exponentiable morphisms

For a morphism $\mathbb{T}$6 in $\mathbb{T}$7 and a fixed $\mathbb{T}$8, the paper constructs a candidate partial product $\mathbb{T}$9 in $Y^X$0. Its underlying set is

$Y^X$1

where $Y^X$2 with the induced $Y^X$3-structure. For $Y^X$4 of arity $Y^X$5, one sets

$Y^X$6

iff first $Y^X$7, and second, for every choice $Y^X$8 with $Y^X$9, one has

$P(Y,f)$0

There are canonical maps $P(Y,f)$1, $P(Y,f)$2, and $P(Y,f)$3, $P(Y,f)$4.

The sufficient condition ensuring that this candidate lies in $P(Y,f)$5 is convexity. A morphism $P(Y,f)$6 is convex when, for each Horn axiom $P(Y,f)$7 beyond reflexivity and any valuation $P(Y,f)$8 satisfying the premises in $P(Y,f)$9, every compatible choice of points $X\times -$0 with $X\times -$1 and $X\times -$2 extends to a valuation $X\times -$3 such that $X\times -$4 for all $X\times -$5 and $X\times -$6. Under this hypothesis, $X\times -$7 satisfies all axioms of $X\times -$8 and $X\times -$9 exhibits $f$0 as a partial product of $f$1 over $f$2. By the Dyckhoff–Tholen criterion, $f$3 is exponentiable (Parker, 2022).

When $f$4 is the terminal projection $f$5, convexity reduces to a condition on $f$6 itself. In that case one sets

$f$7

obtains $f$8, and the relational structure is given by

$f$9

The evaluation map $Decision^{Base}$0, $Decision^{Base}$1, exhibits $Decision^{Base}$2 as the right adjoint to $Decision^{Base}$3.

4. Safe axioms, closure properties, and canonical examples

The categorical results are sharpened by syntactic conditions on axioms. An axiom $Decision^{Base}$4 is safe if, roughly, whenever $Decision^{Base}$5 one automatically has $Decision^{Base}$6 for a suitable trivial valuation; it is very safe if moreover $Decision^{Base}$7 only mentions $Decision^{Base}$8. If the non-reflexive axioms of $Decision^{Base}$9 are all safe, then every object of $Decision^{Base}$00 is convex and hence exponentiable, so $Decision^{Base}$01 is cartesian closed. If they are all very safe, then every morphism is convex and hence exponentiable, so $Decision^{Base}$02 is locally cartesian closed. If in addition $Decision^{Base}$03 is topological over $Decision^{Base}$04 and well-fibred, which is automatic when reflexive and without equalities, then local cartesianness makes $Decision^{Base}$05 a quasitopos, even a topological universe (Parker, 2022).

The preorder and poset examples clarify both the strength and the limits of these criteria. For $Decision^{Base}$06, the only non-reflexivity axiom is transitivity $Decision^{Base}$07. A monotone map $Decision^{Base}$08 is convex iff whenever $Decision^{Base}$09 in $Decision^{Base}$10 and $Decision^{Base}$11 in $Decision^{Base}$12, there exists $Decision^{Base}$13 with $Decision^{Base}$14 and $Decision^{Base}$15. This recovers Niefield’s characterization of exponentiable poset-maps. For preorders and posets, the transitivity axiom is safe, but in the poset case not very safe because it has the extra variable $Decision^{Base}$16; accordingly, both $Decision^{Base}$17 and $Decision^{Base}$18 are cartesian closed, but not locally so.

The $Decision^{Base}$19-categorical examples connect the theory to generalized metric semantics. When $Decision^{Base}$20 is a commutative unital quantale whose underlying lattice is a complete Heyting algebra, the standard relational Horn theory for $Decision^{Base}$21 includes reflexivity and $Decision^{Base}$22-composition

$Decision^{Base}$23

These composition axioms are safe exactly when $Decision^{Base}$24 is a Girard quantale, for example totally ordered. In general the results recover Clementino–Hofmann’s sufficient conditions for cartesian closure of $Decision^{Base}$25. With symmetry axioms for $Decision^{Base}$26 or the metric axiom $Decision^{Base}$27 for $Decision^{Base}$28, one obtains further quasitoposes under the very safe hypotheses. In particular, $Decision^{Base}$29 is cartesian closed whenever $Decision^{Base}$30’s multiplication distributes over arbitrary joins in each variable and $Decision^{Base}$31 is a complete Heyting algebra; the categories of extended metric spaces and symmetric metrics arise similarly and are quasitoposes under the classical Lawvere quantale conditions.

5. Relational exponentiation in solution-set relational algebra

In the database-theoretic sense, relational exponentiation is introduced to complete the fundamental algebraic operations alongside union and cross product. The setting has three layers: active-domain relational algebra over ordinary finite relations, complete-domain relational algebra over characteristic-function relations, and Solution-Set RA, written $Decision^{Base}$32, over sets of candidate relations. Let $Decision^{Base}$33 be an active-domain relation with attribute set $Decision^{Base}$34 and finite extension $Decision^{Base}$35, and let $Decision^{Base}$36 be a complete-domain relation with attribute set $Decision^{Base}$37 and characteristic function $Decision^{Base}$38. Then

$Decision^{Base}$39

is defined as the solution set of all functions $Decision^{Base}$40 (Pratten et al., 8 Sep 2025).

Equivalently, $Decision^{Base}$41 is the set of all active-domain relations $Decision^{Base}$42 satisfying two constraints: first, $Decision^{Base}$43; second, $Decision^{Base}$44 enforces the functional dependency $Decision^{Base}$45. In set-theoretic form,

$Decision^{Base}$46

Writing $Decision^{Base}$47, each $Decision^{Base}$48 corresponds to a tuple $Decision^{Base}$49 with $Decision^{Base}$50, and one reconstructs

$Decision^{Base}$51

If $Decision^{Base}$52 is the characteristic function of $Decision^{Base}$53, then the solution-set characteristic function is

$Decision^{Base}$54

Since there are $Decision^{Base}$55 choices for each of the $Decision^{Base}$56 base tuples, the search space has cardinality

$Decision^{Base}$57

This operation is explicitly presented as the exponential member of an algebraic trio:

• union $Decision^{Base}$58 as “addition,”
• cross-product $Decision^{Base}$59 as “multiplication,”
• exponentiation $Decision^{Base}$60.

The arithmetic analogy is exact at the level of search-space size: a decision space of size $Decision^{Base}$61 over a base of size $Decision^{Base}$62 yields $Decision^{Base}$63 candidate assignments.

6. Algebraic laws, translation semantics, and evaluation

The database formulation establishes several algebraic laws, all up to isomorphism of solution sets. There is an identity-on-exponent law

$Decision^{Base}$64

a product-in-base law

$Decision^{Base}$65

and a product-in-decision law

$Decision^{Base}$66

By contrast, open-domain union does not distribute:

$Decision^{Base}$67

because functions may mix values from both decision relations. The operators $Decision^{Base}$68 and $Decision^{Base}$69 are the solution-set analogues of join or product and union, defined by combining base and decision relations and conjoining or disjoining characteristic functions (Pratten et al., 8 Sep 2025).

A solution set $Decision^{Base}$70 is treated as a first-class algebraic object $Decision^{Base}$71 closed under the operators of $Decision^{Base}$72. Its elements are full assignments $Decision^{Base}$73. Exponentiation encodes the search space, and subsequent operators act on that space:

$Decision^{Base}$74

These respectively filter with a Boolean aggregate over each candidate $Decision^{Base}$75, order candidates by an objective $Decision^{Base}$76, limit to the top $Decision^{Base}$77 candidates, and materialise $Decision^{Base}$78 as an active-domain relation.

A central technical claim is that every $Decision^{Base}$79 expression can be homomorphically translated via $Decision^{Base}$80 into standard relational algebra over two helper relations. The first, $Decision^{Base}$81, is a complete-domain relation whose schema is $Decision^{Base}$82 replicated $Decision^{Base}$83 times, capturing $Decision^{Base}$84 as a giant cross-product. The second, $Decision^{Base}$85, is an active-domain relation of exactly $Decision^{Base}$86 tuples, each tuple $Decision^{Base}$87 extended with symbolic references $Decision^{Base}$88 into the corresponding copy of $Decision^{Base}$89. Under this translation, selections, joins, and aggregations in $Decision^{Base}$90 are lifted symbolically into constraints on $Decision^{Base}$91 and $Decision^{Base}$92, preserving the original $Decision^{Base}$93 structure.

The proposed evaluation architecture is mixed. It first translates $Decision^{Base}$94 over $Decision^{Base}$95 and $Decision^{Base}$96, then generates a constraint model, for example MiniZinc, from the flattened relational algebra, and finally dispatches either to a set-based backend such as SQL when data dominates combinatorics or to a constraint solver, IP, or SAT backend when combinatorics dominates data. The paper gives worst-case time $Decision^{Base}$97, while also stating that solver-based pruning typically yields much better performance in practice.

7. Expressiveness, limits, and conceptual comparison

The database paper assigns strong expressiveness to relational exponentiation. Because $Decision^{Base}$98 captures all functions $Decision^{Base}$99, $f:Base\to Decision$00 can encode any $f:Base\to Decision$01-second-order property and matches the power of SolveDB and NP-Alg, including NP-complete decision problems, NP-hard optimisation, multi-objective queries, and infinite domains such as INT and FLOAT. At the same time, query safety is preserved: as long as the inputs are active-domain relations, $f:Base\to Decision$02 queries return finite active-domain relations upon projection. The complexity boundary is correspondingly sharp: deciding non-emptiness of

$f:Base\to Decision$03

is NP-complete in general combined complexity, and counting or enumerating solutions is $f:Base\to Decision$04-hard (Pratten et al., 8 Sep 2025).

The categorical paper addresses a different limit phenomenon. Exponentials do not arise automatically in arbitrary categories of relational structures; instead, one needs sufficient syntactic conditions such as convexity, safeness, or very safeness. This is why $f:Base\to Decision$05 and $f:Base\to Decision$06 are cartesian closed but not locally cartesian closed, and why the behavior of $f:Base\to Decision$07 depends on properties of the quantale $f:Base\to Decision$08 such as being a complete Heyting algebra or a Girard quantale (Parker, 2022).

A common misconception is to treat the two uses of “relational exponentiation” as direct reformulations of one another. The database construction is an algebraic operator that generates candidate relations and supports optimisation workflows through $f:Base\to Decision$09, $f:Base\to Decision$10, $f:Base\to Decision$11, and $f:Base\to Decision$12. The categorical construction studies when function-space objects or slice-wise partial products exist inside a category of relational models. A plausible implication is that both lines of work elevate function spaces to first-class relational objects, but they do so for different ends: one for categorical closure and semantics, the other for unified query languages and optimiser reasoning.