Bi-Form Formalism: Paired-Structure Frameworks

• Bi-Form Formalism is a set of frameworks that integrates paired components such as axiomatic reasoning and computation, left/right differentials, or additive and multiplicative structures within a single formal object.
• It underpins diverse applications including mechanized mathematics with biform theories, information geometry using bi-forms, BI and separation logic for resource reasoning, and gauge-theoretic methods.
• Its design resolves common misconceptions by disambiguating context-dependent usages and by providing practical architectures for automated reasoning and knowledge management.

Bi-form formalism denotes a family of technical frameworks in which two distinct but coupled layers are treated within a single formal object. In current arXiv usage, closely related expressions refer to several non-equivalent constructions: biform theories in mechanized mathematics, where axiomatic and algorithmic components are combined to integrate reasoning and computation; bi-forms on $M\times M$ in information geometry, where left and right differential structure are organized cohomologically; BI-form in the setting of Bunched Implications and Separation Logic, where additive and multiplicative resource structure coexist; and a nearby gauge-theoretic usage involving two-form gauge fields and gaugeon methods (Carette et al., 2018, Ciaglia et al., 16 Jun 2026, Jipsen et al., 2017, Aochi et al., 2017). A precise account therefore requires disambiguation rather than a single uniform definition.

1. Terminological scope and disambiguation

The literature represented here uses related terminology for distinct formal programs.

Usage	Core object	Characteristic role
Biform theory	triple $T=(L,\Pi,\Gamma)$	integrates axiomatic meaning and symbolic algorithms
Bi-form in information geometry	$(p,q)$-bi-form on $M\times M$	accommodates torsion-full statistical structures
BI-form	formulas and GBI-algebras for BI/SL	combines additive and multiplicative resource connectives
Bi-form gauge field	Abelian two-form field with gaugeons	shifts the gauge-fixing parameter

In the mechanized-mathematics usage, the central problem is the integration of formal deduction and symbolic computation. In the information-geometric usage, the central problem is the generation of metrics and conjugate connections from two-point data, now extended beyond the torsion-free case. In the BI/SL usage, the central problem is local reasoning about resources via additive and multiplicative structure. In the gauge-theoretic usage, the term is tied to two-form fields rather than to the preceding logical or geometric formalisms.

A common misconception is that these usages are interchangeable because they share the expression “bi-form” or “biform.” They are not. The commonality is structural rather than doctrinal: each formalism organizes paired components—syntax and semantics, left and right differentials, additive and multiplicative conjunction, or gauge field and gaugeon freedom—within one framework. This suggests a family resemblance, not a single theory.

2. Biform theories in mechanized mathematics

A biform theory is defined as a triple

$T=(L,\Pi,\Gamma),$

where $L$ is a language of some underlying logic, $\Pi$ is a set of transformers implementing functions on syntactic expressions of $L$, and $\Gamma$ is a collection of axioms that both fix the mathematical meaning of the non-logical symbols of $L$ and contain meaning formulas relating each transformer’s computational behavior to its intended semantics (Carette et al., 2018). If $T=(L,\Pi,\Gamma)$0, the theory is axiomatic; if $T=(L,\Pi,\Gamma)$1, it is algorithmic.

The intended setting is a reflective logic. The project description works in a version of simple type theory with an inductive type $T=(L,\Pi,\Gamma)$2 of syntactic values, a quotation operator $T=(L,\Pi,\Gamma)$3, and an evaluation operator $T=(L,\Pi,\Gamma)$4 that is a partial inverse to quotation. Each transformer $T=(L,\Pi,\Gamma)$5 is named in $T=(L,\Pi,\Gamma)$6 by a constant of type

$T=(L,\Pi,\Gamma)$7

This lets the theory speak simultaneously about object-language expressions and their denotations.

The distinctive feature is the meaning formula. In general form, it quantifies over syntactic values, checks that they represent expressions of appropriate sorts, and equates the semantics of the transformed expression with an abstract semantic function $T=(L,\Pi,\Gamma)$8 applied to the semantics of the inputs. The formalism is therefore not merely a way of attaching code to symbols. Its characteristic claim is that the computational action of a transformer and the mathematical content of that action are both expressed internally.

A standard example is polynomial normalization in the language of rings with identity. Here $T=(L,\Pi,\Gamma)$9, with

$(p,q)$0

and $(p,q)$1 contains the ring axioms together with a meaning formula stating that if $(p,q)$2 is a polynomial expression, then the denotation of $(p,q)$3 is the standard mathematical normal form of the denotation of $(p,q)$4 (Carette et al., 2018). This makes explicit why biform theories are particularly suited to algorithms that manipulate mathematical expressions.

3. Theory graphs, implementations, and the MathScheme/MMT program

The broader methodology organizes mathematical knowledge as a network of biform theories connected by theory morphisms. In the MathScheme project, this is presented as a five-step layered architecture: Logic Design (Log), Implementation (Impl), Transformers, Theory Graphs, and Generic Transformers (Carette et al., 2018). Log starts from Church’s simple type theory enriched with syntax, quotation, and evaluation; Impl realizes the logic in a proof-assistant kernel; the transformer layer links external symbolic algorithms and formalizes their meaning formulas; the theory-graph layer uses morphisms and combinators to extend, rename, combine, and instantiate theories; the generic-transformer layer specializes parametric algorithms through morphisms and partial evaluation.

A case study formalizing natural-number arithmetic illustrates this program concretely. Eight biform theories, BT1 through BT8, were constructed, with first-order logic underlying BT1–BT7 and simple type theory underlying BT8. Theories such as BT2 and BT3 add transformers for addition and multiplication on syntax, while BT5 and BT6 add decision procedures on closed formulas. The morphisms include inclusions between arithmetic theories and an interlogical morphism from a first-order Peano theory to a typed one (Carette et al., 2017). The study contrasts a global reflection approach in $(p,q)$5, where a single $(p,q)$6–quote–eval infrastructure is shared across the graph, with a local deep embedding approach in Agda, where each language fragment requires its own inductive syntax, recognizer, evaluator, and associated lemmas.

Within the MMT/OMDoc framework, the same idea is recast in terms of theories and views. A biform theory consists of a specification logic theory $(p,q)$7, a programming-language theory $(p,q)$8, a view $(p,q)$9 called a bifoundation, and a realization $M\times M$0 for a specification theory $M\times M$1 over $M\times M$2 (Kohlhase et al., 2013). Because both logics and programming languages are represented uniformly as MMT theories, realizations become first-class morphisms in one theory graph.

The Universal Machine operationalizes these realizations by turning implemented constants into rewrite rules on MMT terms. In the running example, an OpenMath content dictionary arith1 is realized in Scala; the constant minus is implemented by a Scala function, and the Universal Machine rewrites a term such as $M\times M$3 to OMI(4) by invoking that realization (Kohlhase et al., 2013). The associated build workflow has three named processes—extract, integrate, and load—that synchronize specifications, generated stubs, and compiled implementations. This suggests that biform theory graphs are intended not only as a foundational scheme but also as a knowledge-management architecture.

The principal design pressures are explicitly identified. First-order logic does not easily talk about the syntax of its own expressions, hence the reflective infrastructure. Meaning formulas can become lengthy and difficult to render, hence pretty-printing support. Repetition across closely related theories motivates the theory-graph architecture. External code is incorporated through a foreign-function interface model, but trust is maintained by proving meaning formulas internally and by placing generic generation inside the reflective logic (Carette et al., 2018).

4. Bi-forms in information geometry

In information geometry, the term bi-form has a different meaning. For $M\times M$4, a $M\times M$5-bi-form on a smooth manifold $M\times M$6 is a section of

$M\times M$7

where $M\times M$8 are the two projections (Ciaglia et al., 16 Jun 2026). Equivalently, it is blockwise multilinear and alternating in left and right slots separately. Decomposable bi-forms are written $M\times M$9.

The motivating setting is a Lauritzen manifold, that is, a smooth manifold equipped with a pseudo-Riemannian metric $T=(L,\Pi,\Gamma),$0, an affine connection $T=(L,\Pi,\Gamma),$1, and its $T=(L,\Pi,\Gamma),$2-conjugate $T=(L,\Pi,\Gamma),$3, with no requirement that either torsion vanish. Traditional information geometry often derives $T=(L,\Pi,\Gamma),$4 from a scalar two-point potential $T=(L,\Pi,\Gamma),$5 by differentiating in left and right directions. The bi-form formalism replaces the scalar $T=(L,\Pi,\Gamma),$6 by a genuine $T=(L,\Pi,\Gamma),$7-form $T=(L,\Pi,\Gamma),$8, explicitly allowing torsion-full statistical structures and providing a cohomological framework (Ciaglia et al., 16 Jun 2026).

Two anticommuting differentials are defined on $T=(L,\Pi,\Gamma),$9:

$L$0

with

$L$1

Locally near the diagonal $L$2, one also has left and right homotopy operators and associated projectors that split spaces of bi-forms into exact and antiexact components.

A $L$3-bi-form $L$4 is a contrast bi-form if its restriction to the diagonal induces a nondegenerate symmetric $L$5-tensor

$L$6

It then determines a connection $L$7 by

$L$8

In the dually curvature-free case, where both $L$9 and $\Pi$0 are flat but may have torsion, there is a canonical contrast bi-form

$\Pi$1

equivalently

$\Pi$2

with $\Pi$3 the $\Pi$4-parallel transport from $\Pi$5 to $\Pi$6 (Ciaglia et al., 16 Jun 2026).

The torsion content is expressed cohomologically:

$\Pi$7

Moreover, contrast bi-forms with vanishing left restriction are exactly the left-exact forms $\Pi$8, called “pre-contrast,” while those with both left and right restrictions vanishing to first order are bi-exact forms $\Pi$9, corresponding to “contrast functions.” This is the main conceptual shift: scalar contrast functions become one part of a larger bicomplex that also captures partial flatness and torsion.

The examples make the point sharply. On the interior of the probability simplex, the canonical bi-form is

$L$0

with $L$1 and $L$2 for the Kullback–Leibler divergence. In the quantum Umegaki case on faithful density matrices, the canonical bi-form is expressed באמצעות the Morozova–Chentsov operator and again satisfies $L$3. By contrast, for the left and right Cartan connections on a semisimple Lie group, the solution bi-forms have nonvanishing $L$4 or $L$5, exhibiting non-zero torsion directly (Ciaglia et al., 16 Jun 2026). A common misconception is therefore that potentials in information geometry must be scalar or torsion-free; the bi-form formalism is explicitly designed to remove that restriction.

5. BI-form, GBI-algebras, and Separation Logic

A third usage appears in the algebraic study of Bunched Implications (BI) and Separation Logic (SL). Here the relevant structure is not a differential bi-form on $L$6 but a logic with both additive and multiplicative connectives. BI formulas are built from atoms using

$L$7

where $L$8 are intuitionistic additives and $L$9 are multiplicatives (Jipsen et al., 2017). The declarative resource reading interprets $\Gamma$0 as conjunction over disjoint resource pieces and $\Gamma$1 as separating implication.

The algebraic semantics is given by generalized BI-algebras (GBI-algebras):

$\Gamma$2

where the additive reduct is a Heyting algebra, $\Gamma$3 is a monoid, and $\Gamma$4 and $\Gamma$5 are left and right residua of $\Gamma$6, satisfying

$\Gamma$7

Commutative GBI-algebras are BI-algebras; Boolean ones are BBI-algebras. The paper also identifies weakening subvarieties and involutive variants (Jipsen et al., 2017).

The semantic models extend far beyond heaps. They include generalized preordered partial monoids, standard heap models, generalized effect algebras, weakening relations on posets, formal languages with concatenation and quotients, and labeled trees or semistructured data. The corresponding complex algebras of up-sets yield GBI-algebras. This directly counters a frequent simplification in program-logic discussions: heap semantics is important, but it is not exhaustive of the BI formalism.

The framework also has a duality theory. A GBI-space is an Esakia space equipped with an associative, continuous operation $\Gamma$8 and a clopen upset $\Gamma$9 satisfying unit and separation conditions. The paper states a dual equivalence

$L$0

On the proof-theoretic side, sequents are inequalities between bunches, and an algebraic argument via residuated frames yields cut elimination for valid GBI-inequalities. On the decision-theoretic side, the picture is mixed: the equational theory of BI is decidable, many weakening-style subvarieties have decidable universal theories, but the equational theories of BBI and related systems are undecidable, and BI has an undecidable quasi-equational theory (Jipsen et al., 2017).

The section of greatest practical visibility concerns (bi-)abduction. Algebraic bi-abduction asks for $L$1 such that

$L$2

In the symbolic-heap fragment of Separation Logic, the problem becomes: given assertions $L$3 and $L$4, find antiframe $L$5 and frame $L$6 such that

$L$7

The exposition notes that bi-abduction was first implemented in Facebook’s Infer, where it drives compositional static analysis of large code-bases (Jipsen et al., 2017). In this sense, BI-form is both a logical formalism and an operational method for automated local reasoning.

6. Gauge-theoretic neighboring usage

A distinct but terminologically adjacent usage occurs in the gaugeon formalism for the two-form gauge fields. The object of study is an Abelian two-form gauge field $L$8 with field strength

$L$9

together with a BRST-symmetric gaugeon extension (Aochi et al., 2017). The formulation introduces a vector gaugeon field $T=(L,\Pi,\Gamma)$00 as a quantum gauge freedom.

The combined Lagrangian is

$T=(L,\Pi,\Gamma)$01

where $T=(L,\Pi,\Gamma)$02 is Kimura’s covariant gauge-fixed two-form Lagrangian and $T=(L,\Pi,\Gamma)$03 is the gauge-fixed vector-Froissart gaugeon Lagrangian. The total system is BRST invariant, with nilpotent BRST operator $T=(L,\Pi,\Gamma)$04 and physical subspace

$T=(L,\Pi,\Gamma)$05

A central result is that $T=(L,\Pi,\Gamma)$06 satisfies a fourth-order “dipole” wave equation,

$T=(L,\Pi,\Gamma)$07

after elimination of auxiliary fields. This higher-derivative character is what allows the gauge-fixing parameter $T=(L,\Pi,\Gamma)$08 to be shifted by the $T=(L,\Pi,\Gamma)$09-number gauge transformation:

$T=(L,\Pi,\Gamma)$10

Because this transformation commutes with the BRST operator, the physical subspace is unchanged (Aochi et al., 2017).

This usage should not be conflated with biform theories, information-geometric bi-forms, or BI-form logics. The shared word “bi-form” here refers to the underlying two-form gauge field rather than to a formalism that combines two semantic layers. The connection is therefore lexical rather than conceptual.

The broader lesson across the literature is that “bi-form formalism” is best treated as a context-dependent designation for paired-structure frameworks. In mechanized mathematics, the pair is axiomatic and algorithmic content; in information geometry, left and right differential structure; in BI and Separation Logic, additive and multiplicative resource structure; and in the gaugeon setting, the two-form field and its quantum gauge freedom. This suggests that disambiguation is not an editorial convenience but a technical necessity.