---
title: Monoidal Differential Modality in Linear Logic
url: https://www.emergentmind.com/topics/monoidal-differential-modality
type: topic
---

# Monoidal Differential Modality in Linear Logic

A monoidal differential modality is a differential modality whose underlying coalgebra modality is monoidal, equivalently a monoidal coalgebra modality equipped with a deriving transformation \(\mathsf d_A:\oc A\otimes A\to \oc A\). In the standard additive symmetric monoidal setting, it is the categorical structure that models the exponential modality of intuitionistic differential linear logic and supports the passage from a “linear” category with \(!\)-structure to a coKleisli category with differential structure [2002.02554]. Closely related terminology is not uniform across the literature: some papers speak primarily of a **monoidal coalgebra modality** and reserve “differential” for the additional deriving transformation, while others use “monoidal differential modality” for the combined package [1808.08513].

## 1. Definition and terminological scope

In an additive symmetric monoidal category \((\mathbb X,\otimes,K,+,0)\), a coalgebra modality consists of a comonad \((\oc,\delta,\varepsilon)\) together with a cocommutative comonoid structure on each \(\oc A\),
\[
\Delta_A:\oc A\to \oc A\otimes \oc A,\qquad e_A:\oc A\to K,
\]
natural in \(A\), such that \(\delta\) preserves comultiplication; equivalently, each \(\oc(f)\) is a comonoid morphism [1806.04804]. A monoidal coalgebra modality adds a symmetric monoidal comonad structure
\[
m_\otimes:\oc A\otimes \oc B\to \oc(A\otimes B),\qquad m_K:K\to \oc K,
\]
with compatibility between the monoidal, comonadic, and comonoid structures [1806.04804].

A differential modality is obtained by adjoining a deriving transformation
\[
\mathsf d_A:\oc A\otimes A\to \oc A
\]
satisfying the constant rule, product rule, linear rule, chain rule, and interchange rule [1806.04804]. A monoidal differential modality is therefore the monoidal version of this structure: the comonad \(\oc\) already behaves as a monoidal exponential modality, and \(\mathsf d\) supplies the differential operator [2508.14320].

The terminology requires some care. One source explicitly notes that the standard term is **monoidal coalgebra modality**, not “monoidal differential modality,” and that a differential linear category is a differential category whose coalgebra modality is monoidal [1808.08513]. A plausible implication is that “monoidal differential modality” is best read as a convenient composite name for the monoidal exponential structure together with its differential operator, rather than as a separate primitive distinct from monoidal coalgebra modalities.

## 2. Structural data and the logic-semantics interface

The monoidal part of the structure is what makes the exponential modality behave multiplicatively:
\[
\oc(A\otimes B)\simeq \oc A\otimes \oc B.
\]
In the presence of finite products, this is equivalent to the Seely-isomorphism presentation,
\[
\chi_{A,B}:\oc(A\times B)\xrightarrow{\cong}\oc A\otimes \oc B,\qquad \oc(\mathsf T)\xrightarrow{\cong} K,
\]
so monoidal coalgebra modalities are the standard categorical semantics of the exponential modality of linear logic [1808.08513]. The same equivalence is stated in the additive setting as
\[
\text{monoidal coalgebra modality} \;\Longleftrightarrow\; \text{additive bialgebra modality},
\]
which identifies the monoidal presentation with a bialgebraic one [1806.04804].

The differential part is encoded by \(\mathsf d_A\), and coKleisli morphisms \(f:\oc A\to B\) are interpreted as the nonlinear or smooth maps. Their derivative is
\[
\mathsf D[f]:=\mathsf d_A f:\oc A\otimes A\to B
\]
in the differential-category sense [1808.08513]. In this way the modality controls both duplicability and differentiation.

This structure is precisely what is needed so that the coKleisli category becomes a cartesian differential category [2002.02554]. The induced differential combinator is given explicitly by
\[
{!}(A\times A) \xrightarrow{\chi} {!}A\otimes {!}A \xrightarrow{1\otimes\varepsilon} {!}A\otimes A \xrightarrow{\mathsf d} {!}A \xrightarrow{f} B,
\]
which is the bridge from linear differential structure to Cartesian differential structure [2002.02554]. If the ambient category is moreover monoidal closed, the coKleisli category is cartesian closed differential, hence a model of the differential \(\lambda\)-calculus [2002.02554].

## 3. Deriving transformations, coderelictions, and additive structure

A central structural result is that, in the presence of a monoidal coalgebra modality, the two traditional presentations of differentiation are equivalent. Besides deriving transformations, one may specify a codereliction \(\eta:A\to \oc A\). For a bialgebra modality, the correspondences are
\[
\mathsf d = (1\otimes \eta)\circ \nabla,\qquad \eta = (u\otimes 1)\circ \mathsf d,
\]
and in an additive linear category deriving transformations, coderelictions, and Fiore’s creation operators all define the same structure [1806.04804]. The slogan stated in the literature is that, for linear logic settings, there is only one notion of differentiation [1806.04804].

The additive enrichment traditionally assumed in differential linear categories is also more rigid than it first appears. One recent result shows that if a symmetric monoidal category carries a monoidal bialgebra modality and a pre-codereliction, then the category becomes an additive symmetric monoidal category via bialgebra convolution [2502.14134]. The induced sum and zero are
\[
f+g = A \xrightarrow{\eta_A} \oc(A) \xrightarrow{\Delta_A} \oc(A)\otimes \oc(A) \xrightarrow{\oc(f)\otimes\oc(g)} \oc(B)\otimes\oc(B) \xrightarrow{\nabla_B} \oc(B) \xrightarrow{\varepsilon_B} B,
\]
and
\[
0:A\to B \quad\text{by}\quad
A \xrightarrow{\eta_A} \oc(A) \xrightarrow{\mathsf e_A} I \xrightarrow{\mathsf u_B} \oc(B) \xrightarrow{\varepsilon_B} B
\]
[2502.14134]. This shows that the additive structure needed for the Leibniz rule is not merely background decoration; it can be reconstructed from the monoidal-bialgebraic differential data itself.

The same paper proves that pre-coderelictions, and hence coderelictions, are unique when they exist [2502.14134]. This suggests that, once the modality is fixed and the relevant monoidal-bialgebraic hypotheses are present, the differential operator is essentially canonical.

## 4. Integration, antiderivatives, and calculus in the monoidal setting

The integration theory associated with differential categories is organized around the coderiving transformation
\[
d_A^\circ := \Delta_A(1 \otimes \varepsilon_A): !A \to !A \otimes A
\]
and the natural transformations
\[
\mathsf K := d^\circ d + \oc 0,\qquad \mathsf J := d^\circ d + 1
\]
[1707.08211]. A differential category has antiderivatives when \(\mathsf K\) is invertible, and in that case the induced integral transformation is built from \(\mathsf K^{-1}\) [1707.08211].

The monoidal case is especially rigid. In a differential category with a monoidal coalgebra modality, if \(\mathsf K\) is invertible, then the integral transformation is given by antiderivatives and is uniquely determined by the differential structure [1707.08211]. This uniqueness is one of the main conceptual differences between the monoidal and merely coalgebraic settings.

A further refinement identifies the decisive role of the tensor unit. In a differential linear category, antiderivatives can be characterized by what happens at the monoidal unit \(R\): the category has antiderivatives iff there exists a map
\[
\mathsf s_R:\oc R\to \oc R
\]
such that
\[
\mathsf s_R\mathsf d_R + \oc(0)=1_{\oc R}
\]
[1808.08513]. The paper emphasizes that the monoidal coalgebra modality is exactly the structure that allows one to reduce a global integration problem to a one-object calculation at \(R\) [1808.08513]. This suggests that, in monoidal differential settings, global calculus is often controlled by the behavior of the modality at the unit object.

## 5. Canonical constructions, free completions, and filtered variants

A major recent development shows that, in a suitably well-behaved \(k\)-linear symmetric monoidal category with finite biproducts and algebraically-free commutative monoids, every coalgebra modality can be freely completed to a differential modality, and every monoidal coalgebra modality can be freely completed to a monoidal differential modality [2508.14320]. The central construction is
\[
\oc^\partial X = \oc X\otimes SX,
\]
where \(SX\) is the algebraically-free commutative monoid on \(X\) [2508.14320]. The universal map from the original coalgebra modality is
\[
\zeta_X = \oc X\otimes \mathsf u_X : \oc X\to \oc X\otimes SX,
\]
and the deriving transformation is
\[
\mathsf d^\partial_X = \oc X\otimes \mathsf d^S_X:\oc X\otimes SX\otimes X\to \oc X\otimes SX
\]
[2508.14320].

This yields, in particular, an initial monoidal differential modality. The construction proceeds by first forming the initial monoidal coalgebra modality \(P\), then freely completing it to
\[
P^\partial = P\otimes S,
\qquad
P^\partial X=\bigoplus_{x:I\to X} SX
\]
[2508.14320]. In simple examples such as the category of sets and relations, the resulting initial monoidal differential modality is distinct from the familiar multiset exponential [2508.14320].

Another recent direction extracts degree information from an existing differential modality. Under mild cokernel hypotheses, every differential modality yields an \(\mathbb N\)-filtered differential modality with endofunctors \(\oc_{\le n}\), where \(\oc_{\le n}A\) is obtained as the cokernel of the \((n+1)\)-st iterated derivative \(\partial^{n+1}\) [2604.16016]. A morphism
\[
f:\oc_{\le n}A\to B
\]
corresponds to a map whose \((n+1)\)-st derivative is zero, so the filtered modality captures polynomial behavior of bounded degree [2604.16016]. This suggests that monoidal differential modalities admit a systematic internal theory of degree truncation analogous to Taylor-degree bounds.

## 6. Examples, variants, and boundaries of the notion

Several standard examples realize monoidal differential structure. The comonad \(Q\) on commutative monoids, or more generally on modules over a commutative rig \(k\), is presented as an example of a differential modality; indeed, it is the initial monoidal differential modality on \(Mod\) [2002.02554]. Its deriving transformation is
\[
\mathsf d_A({x_0,\dots,x_n}\otimes y)={x_0,\dots,x_n,y}
\]
[2002.02554]. This modality controls the skew-monoidal enrichment used to characterize cartesian differential categories as \(Mod^Q\)-enriched categories [2002.02554].

The category \(\mathsf{CON}\) of convenient vector spaces and bounded linear maps is a differential linear category with antiderivatives [1808.08513]. Here the monoidal unit is \(\mathbb R\), coKleisli maps are exactly smooth maps, and the deriving transformation is the classical derivative of smooth maps [1808.08513]. Weighted relational models \(R^\Pi\), given as biproduct completions of complete semirings, are also differential linear categories with antiderivatives under the assumption that all positive integers are invertible in \(R\) [1808.08513]. The polynomial model \(\mathsf{MOD}_R^{op}\), with \(\oc\) the symmetric algebra functor, likewise has antiderivatives when all positive integers are invertible in \(R\) [1808.08513].

The boundaries of the notion are equally important. Monoidal coalgebra modality does not imply differential category: differential algebras provide a monoidal coalgebra modality but do not induce a differential category [1806.04804]. Conversely, Rota–Baxter algebras give a differential category whose coalgebra modality is non-monoidal [1806.04804]. This makes clear that monoidal differential modalities are only one part of the broader landscape of categorical differentiation.

Related structures may sit near this boundary without crossing it. One paper on Lambek calculus with a relevant modality uses a monoidal biclosed category equipped with a lax monoidal endofunctor, comultiplication, counit, and copying maps, and explicitly describes this as reminiscent of coalgebra modalities of differential categories; however, it does not define a deriving transformation and therefore does not present a differential category or a monoidal differential modality in the full sense [2005.03074]. In a different direction, linear exponential comonads have been extended to non-symmetric monoidal categories by replacing ambient symmetry with a derived swap map \(\sigma\), providing the exponential substrate on which differential structure could plausibly be added later [1701.04919]. A plausible implication is that the phrase “monoidal differential modality” should be reserved for settings in which both the monoidal exponential structure and the differential operator are present, rather than for coalgebra-modality-like systems that stop short of differentiation.

Source: https://www.emergentmind.com/topics/monoidal-differential-modality