Day Convolution¶
Lift an indexing category's tensor to a product of functors by left Kan extending their value-wise tensor along it.
Core Idea¶
Day convolution makes a tensor product of functors from the tensor products already present in their indexing and value categories. If \(F,G:\mathcal C\to\mathcal V\) are covariant functors, first form the external tensor \((a,b)\mapsto F(a)\otimes_{\mathcal V}G(b)\). Then left Kan extend it along \(\otimes_{\mathcal C}:\mathcal C\times\mathcal C\to\mathcal C\). The resulting functor \(F\star G\) gathers contributions from ways pairs of indices can combine into an output index. Under suitable size, colimit and tensor-preservation hypotheses, this construction gives the functor category a monoidal product; symmetry additionally requires symmetric premises.[1]
For an ordinary small \(\mathcal C\), the covariant objectwise picture, when the needed copowers and coends exist, is $$ (F\star G)©\cong\int^{a,b\in\mathcal C}\mathcal C(a\otimes b,c)\cdot\bigl(F(a)\otimes_{\mathcal V}G(b)\bigr). $$ The hom-set points from \(a\otimes b\) to \(c\). For presheaves \(\mathcal C^{\mathrm{op}}\to\mathcal V\) the variance and corresponding hom direction change; the two formulas must not be spliced together. The coend expresses left-Kan aggregation, not a demand that every instance be evaluated by symbolic coend calculus.[1]
Structural Signature¶
Sig role-phrases: monoidal indexing category → cocomplete monoidal target → input functors and external tensor → left-Kan aggregation → coherent functor-category product.
- Monoidal indexing category. \(\mathcal C\) has a tensor and unit. Its tensor determines which pairs of indices can contribute to a result: disjoint union for finite-set species is a concrete instance. Without this structure there is no specified transport to perform.[1][2]
- Cocomplete monoidal target. \(\mathcal V\) supplies the tensor on values and the colimits needed for left Kan extension. Coherent monoidal lift requires appropriate preservation of these colimits by the target tensor; it does not follow for every arbitrary target.[1]
- Input functors and external tensor. \(F\) and \(G\) carry indices into \(\mathcal V\); \(F\boxtimes G\) pairs \(F(a)\) with \(G(b)\) before the indexing tensor is applied. This differs from evaluating both at the same \(c\) and taking a pointwise tensor.[1]
- Left-Kan aggregation. Transport \(F\boxtimes G\) along \(\otimes_{\mathcal C}\), collecting compatible pairs. In the covariant presentation, morphisms \(a\otimes b\to c\) supply the weights; reversing them without changing variance changes the construction.[1]
- Coherent functor-category product. \(F\star G\) is a functor, with associative and unital constraints inherited under the hypotheses; symmetric input structures yield symmetric coherence. An objectwise recipe without naturality or coherence is not yet the claimed monoidal tensor.[1][3]
What It Is Not¶
It is not pointwise tensor: \((F\otimes_{\mathrm{pt}}G)(c)=F(c)\otimes G(c)\) pairs only values at the same index. Species Cauchy product instead uses every ordered split of a label set. It is not functor composition: \(F\) and \(G\) need not have composable source and target types. The left Kan extension is along the indexing tensor, not along an input functor.[1][2]
It is also not automatically a numerical sliding-kernel convolution. The live prime Convolution has a fixed local kernel and translation-invariant sliding combination; these are not required here. The smash product of symmetric spectra is not simply the bare convolution of symmetric sequences: Hovey, Shipley and Smith first build the symmetric-sequence tensor and then take a relative tensor over the sphere monoid \(S\).[3]
Scope of Application¶
Glasman's treatment constructs a Day convolution symmetric monoidal \(\infty\)-category of functors under stated conditions, including preservation of colimits separately by the target tensor. Its introduction gives the ordinary covariant left-Kan-extension pattern. One cannot remove the existence assumptions or conclude that an arbitrary functor category carries this product. The coend notation is an ordinary-category presentation, while \(\infty\)-categorical colimits require their own coherence.[1]
Two literal realizations have directly inspected original sources. For set-valued combinatorial species, the indexing category is finite sets and bijections under disjoint union, and the value tensor is Cartesian product; the resulting Day product is the Cauchy product. For pointed-simplicial symmetric sequences, the same indexing combinatorics meets pointed smash in the target, yielding degree-split summands with permutation actions. Their value categories and consequences differ even though the lifting operation is the same.[2][3]
Clarity¶
The phrase “convolution of functors” can conceal two choices: the variance of the functors and the tensor whose decompositions are aggregated. For covariant \(F,G:\mathcal C\to\mathcal V\), a map \(a\otimes b\to c\) contributes to \((F\star G)(c)\); presheaf conventions reverse the orientation. Declaring these choices prevents a plausible-looking but wrong coend formula.[1]
It also separates the bare lift from structures built upon it. Hovey, Shipley and Smith's \(X\otimes Y\) is the symmetric-sequence product; their symmetric-spectrum smash is \(X\otimes_S Y\). Calling the latter simply “the Day convolution” hides the extra balancing imposed by the sphere-spectrum action. The species Cauchy product is a direct Day instance, not a module-relative quotient.[3][2]
Manages Complexity¶
Instead of inventing associativity, unit and symmetry maps for each functor product independently, the construction transports monoidal structure through a universal left Kan extension. The compact question becomes: what is the indexing tensor, what is the target tensor, and do the needed colimits and preservation properties hold? In species, this replaces label-splitting recipes with one functorial product rule. In symmetric sequences, it packages degree splits and \(\Sigma_n\)-actions into a tensor with a coherent twist.[1][2][3]
That compression does not eliminate size or existence issues. A coend can range over many pairs and maps; a non-cocomplete target may not support it, and a target tensor that fails to preserve required colimits may not give the asserted monoidal coherence. The construction organizes this work, not a guarantee of cheap evaluation.[1]
Abstract Reasoning¶
To recognize a candidate instance, start with \(\otimes_{\mathcal C}\) and ask what it means for \(a\otimes b\) to contribute to \(c\). Then form the external value tensor and left-extend it. In a groupoid of finite sets under disjoint union, maps \(A\sqcup B\to S\) correspond, up to relabeling, to ordered partitions of \(S\); hence the abstract construction predicts the species formula \(\coprod_{T\subseteq S}F(T)\times G(S-T)\). This is a derivation from the source's categorical identification and equation, not a mere shared name.[2]
For symmetric sequences, replace Cartesian product with pointed smash and retain bijection equivariance. At degree \(n\), \(p+q=n\) produces summands induced from \(\Sigma_p\times\Sigma_q\) to \(\Sigma_n\). The induced action matters: omitting its shuffles would not produce the twist Hovey, Shipley and Smith define. If the objects are then made \(S\)-modules, take a further relative tensor step before claiming the spectrum smash.[3]
Knowledge Transfer¶
The transfer within category theory is literal: changing the target from sets to pointed simplicial sets changes the value tensor and colimit, but the external-tensor-then-left-Kan-extension operation survives. Re-check the target's colimit properties and the input symmetry each time rather than importing the species formula verbatim.[1][2][3]
Outside categorical mathematics, numerical convolution may have a decomposition sum, but resemblance alone does not identify this operation. A genuine new Day instance requires an indexing monoidal category, functors, target tensor and compatible Kan extension. The portable universal transport of operations is a possible future-prime question; the named Day construction remains domain-specific.
Examples¶
Canonical: Cauchy product of combinatorial species¶
Let \(F\) and \(G\) be set-valued species on finite sets and bijections. Baez writes their Cauchy product at a finite label set \(S\) as $$ (F\cdot_C G)(S)=\coprod_{T\subseteq S}F(T)\times G(S-T). $$ For \(S=\{1,2,3\}\), the summand indexed by \(T=\{1\}\) pairs an \(F\)-structure on \(\{1\}\) with a \(G\)-structure on \(\{2,3\}\). Every subset supplies an ordered split. This is Day convolution from disjoint union, not the pointwise Hadamard product \(F(S)\times G(S)\) that Baez separately lists.[2]
Mapped back: Monoidal indexing category → finite sets/bijections under disjoint union; cocomplete monoidal target → sets with Cartesian product and coproducts; input functors and external tensor → species \(F,G\) and \(F(A)\times G(B)\); left-Kan aggregation → coproduct over all ordered splits \(S=T\sqcup(S-T)\); coherent functor-category product → the species Cauchy product \(F\cdot_C G\).
Applied: tensor of symmetric sequences¶
Let \(X\) and \(Y\) be symmetric sequences of pointed simplicial sets. Hovey, Shipley and Smith define $$ (X\otimes Y)n=\bigvee(\Sigma_n)+\wedge(X_p\wedge Y_q). $$ Equivalently, on a finite labelled set \(C\) one wedges \(X(A)\wedge Y(B)\) over ordered disjoint partitions \(C=A\sqcup B\). The permutation action transports these summands, so a degree split is not merely an unlabelled pair of integers. This is the bare symmetric-sequence tensor; the smash of symmetric spectra is the further relative product \(X\otimes_S Y\).[3]
Mapped back: Monoidal indexing category → finite sets/bijections under disjoint union; cocomplete monoidal target → pointed simplicial sets with smash and wedge coproduct; input functors and external tensor → symmetric sequences \(X,Y\) and \(X_p\wedge Y_q\); left-Kan aggregation → induced \(\Sigma_n\)-equivariant wedge over \(p+q=n\); coherent functor-category product → \(X\otimes Y\) on symmetric sequences, before sphere-module balancing.
Structural Tensions¶
T1: Pointwise simplicity versus decomposition-sensitive multiplication. A pointwise tensor is straightforward to evaluate but cannot combine structures on complementary parts of an index. Day convolution captures those cross-index combinations, at the cost of Kan-colimit existence and potentially many contributing decompositions. Diagnostic: Is the intended product only \(F(c)\otimes G(c)\), or must values at \(a\) and \(b\) contribute when \(a\otimes b\) maps to \(c\)?[1][2]
T2: Bare sequence tensor versus module-relative spectrum tensor. Retaining the unbalanced \(X\otimes Y\) makes degree splits and permutation data explicit. Imposing \(S\)-module compatibility gives the symmetric-spectrum smash \(X\otimes_S Y\), but identifies terms through the sphere action and changes the typed category of objects. Diagnostic: Are the operands merely symmetric sequences, or spectra whose module actions must be respected?[3]
Structural–Framed Character¶
Day convolution is near the structural end within categorical mathematics, yet its category-theoretic frame is constitutive. Evaluative weight: elegance or computational ease is an appraisal, not membership; a cumbersome valid Kan extension is still an instance. Human-practice dependence: a mathematician chooses indexing and target categories, but the induced product is constrained by their tensors and universal extension. Institutional origin: the named construction arose in category theory; a particular source's notation is not necessary. Vocabulary travel: “convolution” also names numerical and signal operators, but the word's travel does not transfer their sliding-kernel signature here. Import versus recognition: recognize an instance from its external tensor, left Kan extension and coherence, rather than importing the name from a similar-looking sum. Its character: a domain-specific categorical tensor construction with multiple literal mathematical realizations and an essential monoidal/Kan-extension frame.[1][2][3]
Structural Core vs. Domain Accent¶
Skeletal core. An operation on inputs is transported along a combination map by a universal aggregation. Whether that operation-lifting skeleton deserves a cross-domain prime is an explicit future-prime question; it is not a claim that live Convolution already supplies the parent.
Domain-bound mechanism. Here the inputs are functors; the combination map is the indexing-category tensor; aggregation is left Kan extension; and output must satisfy monoidal coherence under existence assumptions. Species use Cartesian product and label partitions, while symmetric sequences use pointed smash and equivariant degree partitions.[1][2][3]
Why not a prime. Remove categories, functors and Kan extensions and the identity becomes the looser idea of combining components. The named operation cannot be imported into scheduling or signal processing merely because both sum over decompositions.
Instantiates / Related Primes¶
No strict DAG edge is proposed yet. Live prime Convolution requires a fixed kernel sliding across an input with translation-invariant alignment; those roles fail here. Live domain-specific Kan Extension describes a universal extension operation used to construct Day convolution, but an individual extension and a monoidal product on a functor category are not automatically the same kind of node.
Under the higher-categorical conditions Glasman proves, commutative algebra objects for Day convolution correspond to lax symmetric monoidal functors. This is a structural consequence, not an unconditional axiom for arbitrary functor categories.[1]
Neighborhood in Abstraction Space¶
Day Convolution sits in a moderately populated region (56th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Categories, Sheaves & Homotopy (21 abstractions)
Nearest neighbors
- Direct Limit — 0.87
- Mac Lane's coherence theorem — 0.87
- Monoidal Natural Transformation — 0.86
- Monoidal Monad — 0.85
- Categorical Trace — 0.85
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
- Pointwise tensor or Hadamard product. Combines \(F(c)\) and \(G(c)\) at one index. Tell: Are all indexing-tensor decompositions of \(c\) aggregated?[2]
- Live numerical Convolution. Slides a fixed local kernel across a signal. Tell: Is there translation-invariant input/kernel alignment, or a Kan extension of functors?
- Left Kan extension in general. Extends a suitably typed functor along a map; Day convolution is a specialized product built using one. Tell: Is the extension specifically along an indexing tensor?
- Functor composition. Requires composable codomain and domain. Tell: Are two same-source functors multiplied rather than composed?
- Symmetric-spectrum smash. It is \(X\otimes_S Y\) for \(S\)-modules, not simply bare symmetric-sequence tensor. Tell: Has sphere-module balancing occurred?[3]
- Presheaf-oriented formula. Uses opposite variance and reversed hom direction. Tell: Are functor variance and hom weight consistent?[1]
References¶
[1] Saul Glasman, “Day convolution for ∞-categories,” revised original author version v4, introduction, §2 Definition 2.8 and Proposition 2.12. This revision corrects substantial errors in an earlier version and states colimit-preservation hypotheses. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r
[2] John C. Baez, “Dirichlet Species and Arithmetic Zeta Functions”, original author PDF, §3.1 pp.3–4, especially Eq. (2) and comparison with pointwise Hadamard product. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l
[3] Mark Hovey, Brooke Shipley and Jeff Smith, “Symmetric Spectra”, original author PDF, §2.1 Definition 2.1.3 and Remark 2.1.5; §2.2 Definition 2.2.3 and Corollary 2.2.4. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l