nLab
operad

Contents

The idea

An operad is a gadget used to describe algebraic structures in symmetric monoidal categories. An operad is like a Lawvere theory in that it can be used to describe structures having finitary operations obeying equational laws. However, unlike Lawvere theories, operads can be applied to general symmetric monoidal categories where the tensor product might not be the cartesian product.

Actually the notion of operad (and allied notions such as PROP, club?, multicategory and so on) come in many flavors. Originally used in algebraic topology to provide a systematic formalism for describing the internal operations which exist on iterated loop spaces, the basic idea is quite flexible and adaptable to many categorical situations, and the importance of operads continues to grow.

The rough definition

The original definition is due to J.P. May and was given in his book The Geometry of Iterated Loop Spaces. Since the detailed definition is available from many sources, we will just sketch May’s definition; in the section following this, we give a more detailed higher-level description which generalizes in a number of directions.

Let V be a symmetric monoidal category. A ( permutative or symmetric ) operad in V consists of objects F(n) of V indexed over the natural numbers n=0,1,2, which we intuitively think of as “objects that parametrize the n-ary operations of an algebraic theory” equipped with the following extra structure:

  • Right actions of symmetric groups ρ n:S nhom(F(n),F(n));

  • An unit e:IF(1) which we think of as picking out the identity map as unary operation;

  • Composition operations

    F(k)F(n 1)F(n 2)F(n k)F(n 1++n k)F(k) \otimes F(n_1) \otimes F(n_2) \otimes \cdots \otimes F(n_k) \to F(n_1 + \ldots + n_k)

    which we think of as the result of plugging the outputs of operations θ 1,,θ k into a k-ary operation θ, to produce a new operation θ(θ 1θ k).

These data are subject to obvious identities such as associativity of composition, unit laws, and compatibility of composition with symmetric group actions. For example, the unit laws say that the evident composite

F(n)IF(n)e1F(1)F(n)compF(n)F(n) \cong I \otimes F(n) \stackrel{e \otimes 1}{\to} F(1) \otimes F(n) \stackrel{comp}{\to} F(n)

is the identity map, as is

F(n)F(n)I n1e nF(n)F(1) ncompF(n)F(n) \cong F(n) \otimes I^{\otimes n} \stackrel{1 \otimes e^{\otimes n}}{\to} F(n) \otimes F(1)^{\otimes n} \stackrel{comp}{\to} F(n)

Compatibility with symmetric group actions means that for each element σS k, the composition operation

F(k) i=1 nF(n i)F(n 1++n k)F(k) \otimes \bigotimes_{i = 1}^n F(n_i) \to F(n_1 + \ldots + n_k)

coequalizes a pair of automorphisms

ρ(σ)1,1λ(σ):F(k) i=1 nF(n i)F(k) i=1 nF(n i)\rho(\sigma) \otimes 1, 1 \otimes \lambda(\sigma): F(k) \otimes \bigotimes_{i=1}^n F(n_i) \stackrel{\to}{\to} F(k) \otimes \bigotimes_{i=1}^n F(n_i)

where σ acts on the big tensor product on the left by permuting tensor factors in the obvious way. If V has suitable colimits, this condition could be expressed in terms of tensor products over S n.

The associativity condition will be left for others to fill in.

Algebras

An algebra over an operad F in V is just a semantics for interpreting the F(n) as objects of actual n-ary operations on an object v. That is, an F-algebra structure on an object v in V consists of a collection of maps

F(n)v nvF(n) \otimes v^{\otimes n} \to v

which intuitively is a mapping like this:

θx 1x nθ(x 1,,x n)\theta \otimes x_1 \otimes \ldots \otimes x_n \mapsto \theta(x_1, \ldots, x_n)

so that “elements” of F(n) are interpreted as as n-ary operations on v. These data are subject to some natural conditions which implement this idea.

Perhaps the quickest way to define it is to suppose that V is symmetric monoidal closed, and work by way of parallel to how representations or modules work. Just as an R-module (over a ring R) can be defined as a ring homomorphism

Rhom(A,A)R \to \hom(A, A)

where the hom here is an internal hom of abelian groups, called an endomorphism ring, so there is such a thing as an endomorphism operad attached to any object v in a symmetric monoidal closed category, and an F-algebra over an operad F is the same thing as an operad morphism

Fhom(v ,v)F \to \hom(v^{\otimes \bullet}, v)

to an endomorphism operad (also called a tautological operad).

Now that the clue has been given, the rest is not hard to figure out. The components of the endomorphism operad are defined by

End(v)(n)=hom(v n,v)End(v)(n) = \hom(v^{\otimes n}, v)

Certainly S n acts on the right (that is, contravariantly) on the hom-object hom(v n,v). And clearly there is a canonical map e:Ihom(v,v) to play the role of the unit. The operad composition involves an instance of enriched functoriality of iterated tensor products: there is a map

hom(v n 1,v)hom(v n k,v)hom(v n 1++n k,v k)\hom(v^{\otimes n_1}, v) \otimes \ldots \otimes \hom(v^{\otimes n_k}, v) \to \hom(v^{n_1 + \ldots + n_k}, v^{\otimes k})

The endomorphism operad composition is obtained by tensoring this last arrow with hom(v k,v) on the left, and composing the result with ordinary internal hom-composition

hom(v k,v)hom(v n 1++n k,v k)hom(v n 1++n k,v)\hom(v^{\otimes k}, v) \otimes \hom(v^{\otimes n_1 + \ldots + n_k}, v^{\otimes k}) \to \hom(v^{\otimes n_1 + \ldots + n_k}, v)

A closely related way of defining an F-algebra is via the monad attached to an operad, which we will describe below.

A detailed conceptual treatment

We describe here a compact one-sentence definition of operad first worked out by G.M. Kelly, after a few preliminaries which are important in their own right. The treatment is essentially an exercise in enriched category theory and the formalism of Day convolution. We will work this out fully in the case of ordinary category theory first, that is for categories enriched in V=Set; the case for categories enriched in a complete, cocomplete, symmetric monoidal closed V is completely parallel.

Preparation

  1. Let be the groupoid of finite cardinals and bijections thereon. Since is the underlying groupoid of the category Fin of finite cardinals and functions between them, the tensor product given by the coproduct on Fin restricts to a symmetric monoidal product on . Under this symmetric monoidal structure, may be characterized as the free symmetric strict monoidal category on one generator.

  2. The symmetric monoidal product on extends along the Yoneda embedding to a symmetric monoidal product FG on the category Set op of presheaves on ; this is an instance of Day convolution. The presheaf category is cocomplete, and Day convolution is cocontinuous in each of its separate arguments F,G; we say in that case that Set op is symmetric monoidally cocomplete.

is also a skeleton of the category FB of finite sets and bijections; a functor FB opSet (or what is more or less the same thing, a functor FBSet) is called a species. The Day convolution product on the category of species may be described by the rule

(FG)[S]= S=T+UF[T]×G[U](F \otimes G)[S] = \sum_{S = T + U} F[T] \times G[U]

where the sum is over all ways of partitioning the finite set S into two (possibly empty) parts.

  1. According to the yoga of presheaf categories and Day convolution, given a symmetric monoidally cocomplete category D, a symmetric monoidal functor

    X:DX: \mathbb{P} \to D

    extends uniquely up to isomorphism to a symmetric monoidal cocontinuous functor

    X̂:Set opD,\hat{X}: Set^{\mathbb{P}^{op}} \to D,

    taking a presheaf F: opSet to the weighted colimit FX. Putting items 1. and 2. together, we may therefore describe Set op universally up to equivalence as the free symmetric monoidally cocomplete category on a single generator.

In general, weighted colimits may be described by coend formulas; here

F X= k:F(k)X(k)F \cdot_{\mathbb{P}} X = \int^{k: \mathbb{P}} F(k) \cdot X(k)

where Sd denotes the tensoring of a set S with an object d, that is the coproduct of an S-indexed set of copies of d. The coend here indicates a coequalizer

kF(k)S kX(k) kF(k)X(k) kF(k)X(k)\sum_k F(k) \cdot S_k \cdot X(k) \stackrel{\to}{\to} \sum_k F(k) \cdot X(k) \to \int^k F(k) \cdot X(k)

where one of the parallel arrows involves right actions of symmetric groups S k on the F(k), and the other involves left actions of S k on objects X. In other words, the coend in this instance may be described as a sum of tensor products:

F X= kF(k) S kX(k).F \cdot_{\mathbb{P}} X = \sum_k F(k) \otimes_{S_k} X(k).

The aforementioned universal property of Set op with its convolution product may be more explicitly described as follows: given a symmetric monoidally cocomplete category D and an object d therein, there exists up to isomorphism a unique symmetric monoidal cocontinuous functor Set opD which sends hom (,1) to d. This functor takes a presheaf F: opSet to the following object of D:

kF(k) S kd k.\sum_k F(k) \otimes_{S_k} d^{\otimes k}.

For instance, when D is the symmetric monoidally cocomplete category (Set,×) and x is a set, this formula

F̂(x)= nF(k) S kx k\hat{F}(x) = \sum_n F(k) \otimes_{S_k} x^k

is the value at x of what Joyal calls the analytic functor F̂:SetSet associated to a species F, proposed as the categorification of exponential generating functions. The fact that FF̂(x) is symmetric monoidal (cocontinuous) means that there is a canonical isomorphism

(F DayG)̂(x)F̂(x)×Ĝ(x)\hat{(F \otimes_{Day} G)}(x) \cong \hat{F}(x) \times \hat{G}(x)

In other words, FF̂ behaves like a categorified version of Fourier transform, taking convolution products to ordinary (pointwise) products.

  • For symmetric monoidally cocomplete categories C,D, let [C,D] denotes the category of symmetric monoidal cocontinuous functors CD. The universal property of Set op means that we have an equivalence [Set op,D]D. Consequently, we have an equivalence [Set op,Set op]Set op. Since symmetric monoidal cocontinuous functors compose, the category on the left carries a monoidal product given by endofunctor composition. Transferring endofunctor composition across the equivalence produces a monoidal product on Set op called substitution product of species. The substitution product of species F,G is denoted FG.

In detail: a species G: opSet induces a symmetric monoidal cocontinuous functor

Set opSet op:FFG= kF(k) S kG DaykSet^{\mathbb{P}^{op}} \to Set^{\mathbb{P}^{op}}: F \mapsto F \circ G = \sum_k F(k) \otimes_{S_k} G^{\otimes_{Day} k}

The k-fold Day tensor power of G is given (in the language of species) by the formula

G k[S]= S=T 1++T kG[T 1]×G[T 2]××G[T k]G^{\otimes k}[S] = \sum_{S = T_1 + \ldots + T_k} G[T_1] \times G[T_2] \times \ldots \times G[T_k]

where we sum over all ways of breaking up a finite set S into k blocks, some possibly empty. Thus we have an explicit description of the substitution product,

(FG)[n]= kF[k] S k( [n]=T 1++T kG[T 1]××G[T k]),(F \circ G)[n] = \sum_k F[k] \otimes_{S_k} (\sum_{[n] = T_1 + \ldots + T_k} G[|T_1|] \times \ldots \times G[|T_k|]),

and it is clear from our discussion above that substitution is a monoidal product.

Definition as monoid

We are at last ready for the one-sentence definition:

A (Set-based) operad is a monoid in the monoidal category (Set op,).

Remarks

  • We can get different flavors of operad by considering different notions of monoidal category. For instance, for the theory of monoidal categories, the discrete category plays the role of the free (strict) monoidal category on one generator, and Set op the free monoidally cocomplete category on one generator. Similarly, for braided monoidal categories, we have the braid category 𝔹, and Set 𝔹 op is the free braided monoidally cocomplete category on one generator. Again, for cartesian categories, we have Fin op (the opposite of finite sets and functions) as the free cartesian category on one generator, and Set Fin is the free cartesian monoidally cocomplete category on one generator. In each of these cases we get a corresponding notion of operad by following the above treatment mutatis mutandis: nonpermutative operads, braided operads, cartesian operads (better known as Lawvere theories).

  • All of the above carries over to the enriched setting, where we work over a complete, cocomplete symmetric monoidal closed base category V. Here ordinary categories (like ,,𝔹,Fin) are viewed as V-enriched by a simple change of base: change from hom-sets to hom-objects by applying the change of base functor

    SetVSet \to V

    that takes a set S to the S-fold coproduct SI, where I is the monoidal unit of V.

  • There are still other notions of operad; see for example the discussion of T-operads under the entry multicategory for more details. See also the general framework by Hyland, Fiore, and others to be filled in.

  • In still other directions, there are for example notions of cyclic operad and modular operad. Anyone want to take these up?

The monad attached to an operad

Examples

this list of examples should eventually be collected in a table of contents on operad theory

generalizations:

Free operads

For G a discrete group, write V G for the category of objects of V equipped with a G-action. For V symmetric monoidal this is again a symmetric monoidal category and the forgetful functor V GV is symmetric monoidal.

Definition

The category of collections of V is

VColl:= nV Σ n.V Coll := \prod_{n \in \mathbb{N}} V^{\Sigma_n} \,.

Notice that both Σ 0 and Σ 1 are the trivial group.

So a V-operad P is a special V-collection with extra structure relating its components. This gives an evident forgetful functor U:VOperadVColl and its left adjoint, the free operad functor

F:VCollVOperad:U.F : V Coll \stackrel{\leftarrow}{\to} V Operad : U \,.

This is for instance used to define the model structure on operads by transfer along this adjunction from a model struucture on V.

The free operad functor may more explcitly be described as follows (for instance BerMor03 section 5.8)

Let T:=Core(Ω pl) be the core of the category of planar rooted trees. Write

  • t nΩ n for the n-corolla (the tree with a single vertex, n inputs and its unique output root)

  • for T any tree with n-ary root vertex let {T i} i=1 n be the sub-trees such that T=t n(T 1,,T n).

Then every KVColl defines a functor K¯:T opV by the inductive formula

K¯:I\bar K : | \mapsto I
K¯:TK¯(t n(T 1,,T n)):=K(n)K(T 1)K(T n).\bar K : T \mapsto \bar K(t_n(T_1, \cdots, T_n)) := K(n) \otimes K(T_1) \otimes \cdots K(T_n) \,.

Let moreover λ:TSez be the functor that sends a tree to the set of numbering of its incoming edges, and let λ¯:TV be given by postcomposition with S sSI.

Then the free operad on a collection K is the coend

K¯ Tλ¯= TTK¯(T)λ¯(T).\bar K \otimes_{\mathbf{T}} \bar \lambda = \int^{T \in \mathbf{T}} \bar K(T) \otimes \bar \lambda(T) \,.

Examples

Let V be a cartesian monoidal category and K=* the terminal collection, which is the terminal object in each degree, with, necessarily, trivial Σ n-action.

The free operad on this should be the V-A-infinity operad it consists in degree n of precisely Σ n-operations per n-ary planar tree. So every planar n-ary tree is regarded by the operad as one distinct operation to multiply n elements, and freely adjoining to each tree a Σ n-action amounts to not dividing out any commutativity symmetry on these operations.

Model structures on operads

If the symmetric monoidal category V that the operads under consideration are enriched in carries the structure of a monoidal model category, then under suitable conditions there is also the structure of a model category on the category of V-operads. This is important for the notion of homotopy algebra over an operad, such as A - and E -algebras.

See

References

The definition is originally due to

  • Peter May, The geometry of iterated loop spaces