Infinity-Operads Demystified


The purpose of this post is to record the definition of ∞-operads and explain why it works like that. For this I’m using Lurie’s definition of ∞-operads; there is also a definition by Cisinski–Moerdijk–Weiss using dendroidal sets, about which I might talk later.

Indeed, the definition on an ∞-operad is a bit mysterious taken “as-is” – see [HA, §]. My goal is to explain how to reach this definition, mostly for my own sake. Most of what follows is taken either from the book Higher Algebra, the nnLab, or the semester-long workshop about hgiher category theory in Lille in 2015.

Cartesian morphisms

Grothendieck fibrations

The first thing to explain would be Cartesian (and coCartesian) morphisms. They are generalizations of Grothendieck fibrations from ordinary category theory. The basic idea of a Grothendieck fibration p:EBp : E \to B is that the fibers Eb=p1(b)E_b = p^{-1}(b) depend contravariantly on bb, i.e. given a morphism f:bbf : b \to b', there exists a lift (a functor) fˉ:EbEb\bar{f} : E_{b'} \to E_b. The definition of a Grothendieck fibration is exactly what’s needed for all this to work correctly.

Following the nnLab, a morphism ϕ:ee\phi : e' \to e is said to be **pp-Cartesian lift ϕ:ee\phi : e' \to e. Then p^(f)(b):=eEb\hat{p}(f)(b) := e' \in E_{b'}. The unique factorizations in the definition of a fibration makes this into a functor p^(f)\hat{p}(f), and this becomes a (pseudo)functor BCatopB \to \mathsf{Cat}^{\rm op}.


The study of ∞-operads obviously involves higher category theory, and in Lurie’s setting higher categories are quasi-categories (I think some people call them quategories). There are plenty of explanations of quasi-categories online, so I will just record the bare definitions here.

A quasi-category is a simplicial set CΔopSetC_\bullet \Delta^{\rm op} \to \mathsf{Set} such that all inner horns can be filled, i.e. every morphism ΛknC\Lambda_k^n \to C can be extended to Δn\Delta^n for 1<k<n1 < k < n (compare with Kan complexes, where outer horns can be filled too).

The set C0C_0 is to be understood as the set of objects of the quasi-categories; the set C1C_1 is to be understood as the set of morphisms. Faces d0,d1:C1C0d_0, d_1 : C_1 \to C_0 give the target and the source of a morphism, and the degeneracy s0:C0C1s_0 : C_0 \to C_1 gives the identity of an object. The set C2C_2 gives information about composition: a composable pair of arrows is the same thing as a morphism σ:Λ12C\sigma : \Lambda_1^2 \to C, which extends to σ~C2\tilde\sigma \in C_2; then d1σ~d_1\tilde\sigma is “a composite” of the pair of composable morphisms. Higher CnC_n give coherence conditions for composition, and everything works out (the space of possible compositions for a pair of composable morphisms is contractible, idx- \circ \operatorname{id}_x is homotopic to the identity and so on).

The nerve of a plain category is an example of a quasi-category.

(co)Cartesian morphisms

The definition of a Cartesian morphism essentially mimics the definition of a Grothendieck fibration, except that the categories are replaced with quasi-categories. I won’t dwell too much on this (see the nnLab article), since as far as I can tell there’s no hidden surprise here, the most difficult thing being to determine how to translate the axioms in terms of quasi-categories. Let’s just note that a morphism p:EBp : E \to B is said to be a coCartesian morphism (AKA opfibration) if pop:EopBopp^{\rm op} : E^{\rm op} \to B^{\rm op} is Cartesian. This means that the fibers EbE_b depends covariantly instead of contravariantly on bb. I am not 100% sure why a different name is used for plain categories and for quasi-categories; maybe it’s just a historical artefact?

Symmetric monoidal ∞-categories

Rather than plain operads, ∞-operads are a generalization of colored operads, AKA multicategories. As such, they have multiple objects, and behave more like a ∞-category where morphisms can have multiple inputs instead of just one. Let’s first define symmetric monoidal ∞-categories, where a morphism c1cndc_1 \otimes \dots \otimes c_n \to d can be thought of as a morphism with multiple inputs. (This section is taken straight from the introduction of Chapter 2 of Higher Algebra and adapted to my notations).

Given a symmetric monoidal category C\mathsf{C}, one can construct its category of operators C\mathsf{C}^\otimes. This is a category over Γ\Gamma, the category of finite pointed sets with objects n+={,1,,n}n_+ = \{ *, 1, \dots, n \}.

The objects of C\mathsf{C}^\otimes are finite sequences [c1,,cn][c_1, \dots, c_n] of objects of C\mathsf{C}. The morphisms [c1,,cn][d1,,dm][c_1, \dots, c_n] \to [d_1, \dots, d_m] consist of:

Then p[c1,,cn]p[c_1, \dots, c_n] is n+n_+, and the image of a morphism by pp is the “α\alpha” part. Consistenly with the previous notation for fibers, let Cn:=p1(n+)\mathsf{C}^\otimes_n := p^{-1}(n_+). Note that Γ\Gamma is the category of operators of the terminal category (equipped with its unique symmetric monoidal structure).

Then CpΓ\mathsf{C}^\otimes \xrightarrow{p} \Gamma satisfies the two fundamental properties:

If we unroll the definition, this means that for every [c1,,cn]C[c_1, \dots, c_n] \in \mathsf{C}^\otimes and for every f:n+m+f : n_+ \to m_+, there is some morphism fˉ:[c1,,cn][d1,,dm]\bar{f} : [c_1, \dots, c_n] \to [d_1, \dots, d_m] that covers ff and such that for all morphism gˉ:[c1,,cn][d1,,dk]\bar{g} : [c_1, \dots, c_n] \to [d_1, \dots, d_k] and every factorization of ff through g=p(gˉ)g = p(\bar{g}), there is a factorization of fˉ\bar{f} through gˉ\bar{g} inducing it. Indeed, we can choose djd_j to be f(i)=jci\bigotimes_{f(i) = j} c_i and see that everything works. Recall that this means the fibers Cn\mathsf{C}^\otimes_n depend covariantly on nn.

Indeed, a morphism [c1,,cn][d1,,dn][c_1, \dots, c_n] \to [d_1, \dots, d_n] in C\mathsf{C}^\otimes which covers the identity of n+n_+ is uniquely determined by morphism cidic_i \to d_i; one just has to see that this morphism is given by p^(ρi)=ρi\hat{p}(\rho^i) = \rho^i_*.

Conversely, if DpΓ\mathsf{D} \xrightarrow{p} \Gamma is a functor that satisfies (M1) and (M2), let C=D1=p1(1+)\mathsf{C} = \mathsf{D}_1 = p^{-1}(1_+). Then C\mathsf{C} becomes a symmetric monoidal category.

So axioms (M1) and (M2) capture exactly what it means to be a symmetric monoidal category. There are many possible equivalent definitions of monoidal categories, but Lurie’s insight was to find one that could be adapted to quasi-categories. Indeed, he defines:

Definition. [HA,] A symmetric monoidal ∞-category is a coCartesian fibration p:CNΓp : \mathsf{C}^\otimes \to N\Gamma such that the maps ρi:n+1+\rho^i : n_+ \to 1_+ induce an equivalence Cn(C1)×n\mathsf{C}^\otimes_n \to (\mathsf{C}^\otimes_1)^{\times n}.

In this definition, C\mathsf{C}^\otimes should be thought of as the ∞-category of operators, and the underlying ∞-category is really C1=p1(1+)\mathsf{C}^\otimes_1 = p^{-1}(1_+). The monoidal product and the unit are induced as above.


Following the same pattern, Lurie defines the category of operators of a colored operad. This satisfies a bunch of axioms, which allow one to recover the colored operad from the category of operators. The axioms are also laid out in such a way that generalizing them to ∞-categories is possible, thus giving the definition of ∞-operads.

Let P\mathtt{P} be a colored operad. Its category of operators CP\mathsf{C}_\mathtt{P} is given by:

Like before, there’s a functor p:CPΓp : \mathsf{C}_\mathtt{P} \to \Gamma given by p[c1,,cn]=n+p[c_1, \dots, c_n] = n_+. This is a Grothendieck opfibration, which allows us the recover P\mathtt{P} from CP\mathsf{C}_\mathtt{P}.

Let Pn:=p1(n+)\mathtt{P}_n := p^{-1}(n_+). In particular, P1\mathtt{P}_1 is the category of unary operations in P\mathtt{P} (also called “underlying category” of P\mathtt{P}). The colors of P\mathtt{P} are given by the objects of P1\mathtt{P}_1. As before, iρi:Pn(P1)×n\prod_i \rho^i_* : \mathtt{P}_n \to (\mathtt{P}_1)^{\times n} is an isomorphism. The operations of type P(c1,,cn;d)\mathtt{P}(c_1, \dots, c_n; d) are recovered as the morphisms CP([c1,,cn],d)\mathsf{C}_\mathtt{P}([c_1, \dots, c_n], d) which cover n+1in_+ \to 1_i, i1i \mapsto 1.

Two types of morphisms appear in the previous discussion, which lead to the following definition: a morphism f:n+m+f : n_+ \to m_+ in Γ\Gamma is

Definition. [HA,] An ∞-operad (really, the category of operators of an ∞-operad) is a functor of quasi-categories p:CNΓp : \mathsf{C}^\otimes \to N\Gamma such that:

  1. For every inert morphism f:m+n+f : m_+ \to n_+ and every CCmC \in \mathsf{C}^\otimes_m, there is a pp-coCartesian morphism fˉ:CC\bar{f} : C \to C' lifting ff, which induces a functor f!:CmCnf_! : \mathsf{C}^\otimes_m \to \mathsf{C}^\otimes_n. Recall that the inert morphisms are those who “forget” points. The functor f!f_! is the functor which, given an mm-uple, forgets some of the factors. The object CC is a sequence of colors, and CC' is the same sequence with some colors forgotten.
  2. For f:n+m+f : n_+ \to m_+, let Cf(,)C(,)\mathsf{C}^\otimes_f(-,-) \subset \mathsf{C}^\otimes(-,-) be the connected components lying over ff. Then

Cf(C,C)1kmCρif(C,Ci)\mathsf{C}^\otimes_f(C, C') \to \prod_{1 \le k \le m} \mathsf{C}^\otimes_{\rho^i \circ f}(C, C'_i)

is a homotopy equivalence. This means that an “operation” [c1,,cn][d1,,dn][c_1, \dots, c_n] \to [d_1, \dots, d_n] is “the same” as a collection of operations P((ci)f(i)=j,dj)\mathtt{P}((c_i)_{f(i) = j}, d_j). 3. For every collection of objects c1,,cnC1c_1, \dots, c_n \in \mathsf{C}^\otimes_1, there exists an object CCnC \in \mathsf{C}^\otimes_n and pp-Cartesian morphisms CciC \to c_i covering ρi\rho^i. This means that ρ!i:Cn(C1)×n\prod \rho^i_! : \mathsf{C}^\otimes_n \to (\mathsf{C}^\otimes_1)^{\times n} is an equivalence.

From this data, the quategory of unary operations of the ∞-operad is given by C1\mathsf{C}^\otimes_1. The colors of the ∞-operad are the object of this quategory. Finally, the operations of type (c1,,cn;d)(c_1, \dots, c_n; d) are the morphisms c1cndc_1 \oplus \dots c_n \to d lying over the unique active morphism n+1+n_+ \to 1_+, where c1cnc_1 \oplus \dots \oplus c_n is “the” object of Cn\mathsf{C}^\otimes_n corresponding to (c1,,cn)(C1)×n(c_1, \dots, c_n) \in (\mathsf{C}^\otimes_1)^{\times n} under the equivalence of (3).

And voilà! An ∞-operad. I’m not as scared of the definition as I was when I first saw it, and I hope you aren’t anymore either.