- The paper establishes that polarised syntax for the resource modality ! and effect modality ◇ corresponds to comonads and monads relative to CBPV polarity shifts, resolving a completeness gap in ordinary models.
- It develops correspondence theorems linking relative coadjunctions and relative strengths with positive and linear-non-linear adjunctions over duploids, providing an axiomatic account of the syntactic model.
- The results show that ! must be positive and ◇ negative with only weak focalisation, while the framework supports sound interpretation and completeness despite open questions about strength, composition, and non-associative bases.
Overview
The paper develops the proof-theoretic and categorical underpinnings of polarised sequent calculi for intuitionistic linear logic extended with a resource (exponential) modality ! and a strong effect modality ◊. Its central claim is that the correct semantic notion for these modalities in linear call-by-push-value (CBPV) models is not an ordinary comonad on the value category and monad on the computation category, but rather a comonad relative to the positive polarity shift ↑ and a monad relative to the negative polarity shift ↓ (2606.14652). The authors show that a polarised term syntax for these modalities, previously derived from focalisation considerations in work by Danos–Joinet–Schellinx and Curien–Herbelin-style sequent calculi, is precisely a syntax for such relative (co)monads — a link that had not been noticed when the syntax was first introduced.
Background: polarisation, focalisation, and linear CBPV
The paper works with a polarised version of intuitionistic multiplicative-additive linear sequent calculus ($\system{IMALL}$), in which formulæ are classified as positive (⊗, ⊕, $1$, $0$) or negative (⊸, ◊0, ◊1), mirroring the value/computation distinction in CBPV models. Semantically, a linear CBPV model is a strong adjunction between a symmetric monoidal distributive category of values ◊2 and a category of computations ◊3 that is cartesian closed relative to ◊4. The deductive system (duploid) associated to such an adjunction satisfies the rules of ◊5, and the calculus ◊6 presented here is a proof-relevant syntax whose ◊7-reduction is confluent and noetherian, whose normal-long forms coincide with focused normal forms (yielding completeness of focusing proof search), and which itself forms a syntactic linear CBPV model via its subcategories of thunkable and linear maps.
The design rationale is traced to Danos, Joinet and Schellinx's proof-relevant ◊8: cut-elimination critical pairs are resolved statically by polarity annotations, and ◊9-expansions must be compatible with reductions. This forces polarisation (expansions must preserve the reduction order determined by polarity) and focalisation (hidden cuts in compound values must be revealed), the latter achieved by restricting primitive constructors to values and co-values.
The modalities and the completeness gap
Extending ↑0 with ↑1 and ↑2 (the latter required to be strong with respect to ↑3, via the schema ↑4) yields ↑5. The corresponding syntax ↑6 introduces a promotion constructor in the shape of pattern matching ↑7 and dereliction as a pattern ↑8, reflecting a weak invertibility of ↑9 on the right and of ↓0 on the left. The authors argue that polarisation forces ↓1 to be positive (call-by-value reduction, so duplication/erasure only applies to boxes) and ↓2 negative (call-by-name reduction), and that only weak focalisation properties hold for these modalities — the unrestricted focalisations are shown to be unsound in general.
The key technical observation is that interpreting ↓3 and ↓4 requires an extra polarity shift:
↓5
This corresponds to Andreoli's observation that introduction of ↓6 terminates a synchronous phase and of ↓7 an asynchronous phase. The consequence is a completeness failure: with ordinary comonads on ↓8 and monads on ↓9, one cannot recover a resource modality on the thunkable subcategory nor an effect modality on the linear subcategory of the syntactic model.
The relative (co)monad resolution
The resolution is to replace the ordinary comonad $\system{IMALL}$0 on $\system{IMALL}$1 and monad $\system{IMALL}$2 on $\system{IMALL}$3 by:
- a $\system{IMALL}$4-relative comonad $\system{IMALL}$5 arising from a $\system{IMALL}$6-relative coadjunction $\system{IMALL}$7 with $\system{IMALL}$8 strong monoidal, and
- a $\system{IMALL}$9-relative monad ⊗0 from a ⊗1-relative adjunction, equipped with a strength ⊗2.
The strength is defined relative to ⊗3 rather than via a monoidal structure on ⊗4; this is a deliberate choice, since such monoidal structure is unavailable in certain categories of algebras of interest, and it contrasts with the monoidal-target assumption made in RMM-LNL models, which the paper shows to be a special case (non-linear CBPV with an effect modality). The interpretation of values of type ⊗5 as ⊗6-coalgebra morphisms and covalues of type ⊗7 as ⊗8-algebra morphisms, via the comparison functor of relative adjunctions, yields a sound interpretation of ⊗9 into these models — and crucially, the syntactic model itself carries such a relative comonad and monad, restoring completeness.
The paper connects this to two independent recent lines of work that arrived at the same relative structures: Jiang et al.'s interface for user-defined effects in CBPV via Church encodings and coinductive types (yielding negative-typed monads, i.e. monads relative to ⊕0 in the paper's terminology), and Melliès' extension of tensorial logic and dialogue categories to exponentials via relative symmetric monoidal comonads, motivated by obstacles in functorial game semantics.
Adjunctions over duploids and the correspondence theorems
The second half develops the axiomatic, non-associative account. Working with duploids (non-associative categories in which every object is positive or negative, with a positive left adjoint ⊕1 and negative right adjoint ⊕2 to the identity functor), and with adjunctions between bare functors (reflexive graph morphisms not necessarily preserving composition), the paper establishes a correspondence:
- Every ⊕3-relative coadjunction over an adjunction ⊕4 induces a positive adjunction on the duploid ⊕5; conversely, every positive adjunction from an associative category into a duploid induces a ⊕6-relative coadjunction.
- This is lifted to linear-non-linear adjunctions: ⊕7-relative LNL coadjunctions (with cartesian ⊕8 and strong monoidal ⊕9) correspond to direct LNL adjunctions (positive adjunctions from a cartesian category into a symmetric monoidal duploid).
- For the combined resource-and-effect setting, a relative monad strong with respect to $1$0 in the indirect situation corresponds to a bare endofunctor $1$1 strong relative to $1$2 in the direct situation, with the four coherence diagrams matching.
A corollary is that any $1$3 model over an adjunction induces one over its thunkable/linear completion — the construction underpinning the syntactic model. The appendix verifies in detail that the syntactic duploid has finite sums, with the required naturality equations holding via the $1$4- and $1$5-equational theory, and proves the relative (co)monad laws for $1$6 and $1$7 syntactically.
Limitations and open questions
The paper is candid that several points remain open. The notion of strength for the relative monad requires $1$8 to have a right adjoint; the alternative of assuming a monoidal structure on $1$9 (as in Liell-Cock et al.) is discussed but the general comparison between the two approaches beyond the LNL special case is not fully developed. The treatment of the lax monoidal structure on $0$0 and of the strength in the interpretation of the rules is stated as "omitted for simplicity" in places. A composition theorem for the duploid constructions is flagged as an outstanding todo in the source. The weak focalisation properties for the modalities are inherently partial — the unrestricted focalisations are provably unsound — so the calculus does not enjoy full focusing for $0$1 and $0$2 in the way it does for the additive and multiplicative connectives. Finally, the correspondence results are established for adjunctions over associative base categories; the behaviour of relative (co)adjunctions internal to non-associative categories is not addressed.
Conclusion
The paper establishes that the polarised syntax for $0$3 and $0$4 arising from focalisation is exactly a syntax for (co)monads relative to the CBPV polarity shifts, resolving a completeness gap in the semantics of linear CBPV with resource modalities. Its correspondence theorems between relative LNL coadjunctions and direct LNL adjunctions over duploids, and between relative and direct notions of strength, give the non-associative categorical framework of polarised logic a precise axiomatic account of resource and effect modalities, unifying proof-theoretic motivations with the recent relative-monad approaches to effects and to game semantics.