Before setting up the axiomatics of \(\infty \)-categories used in this book, we take a moment to explain the heuristics behind the theory, which will help motivate our choice of setup.

The fundamental principle of homotopy theory

The first concept we need to discuss is the notion of equality. In traditional mathematics, equality is typically viewed as a mere property: two mathematical objects are either equal or they are not. In homotopy theory and related fields, however, we encounter a notion of identification where it is not just important that two objects are identified, but precisely how they are identified. In particular, the identification between objects becomes a mathematical structure in its own right: think of a homotopy between two continuous maps, or of an isomorphism between two abelian groups. This perspective is so central to understanding \(\infty \)-categories that it deserves to be elevated to a guiding principle:

Fundamental Principle of Homotopy Theory: When expressing that two entities are equal, we must always specify how they are equal by providing an identification between them.

To get a feeling for this principle, let us look at three familiar settings in which it already appears:

Example 1.1.1 (Algebraic topology). Algebraic topology provides the most natural example of the Fundamental Principle:

  • Two continuous maps \(f,g\colon X \to Y\) are considered ‘equal up to homotopy’ if we provide a homotopy \(H\colon X \times [0,1] \to Y\) between \(f\) and \(g\);
  • Two spaces \(X\) and \(Y\) are then considered ‘homotopically the same’ when equipped with a homotopy equivalence: continuous maps \(f\colon X \to Y\) and \(g\colon Y \to X\) together with chosen homotopies \(gf \sim \id _X\) and \(fg \sim \id _Y\).
  • When studying diagrams of spaces, we similarly replace strict commutativity by homotopy commutativity. For instance, given a square
    Commutative square with maps X to Y labeled f, Z to W labeled k, X to Z labeled g, Y to W labeled h.
    we do not ask for a strict equality \(h \circ f = k \circ g\), but instead require a specified homotopy \(H\colon h \circ f \sim k \circ g\) witnessing how the two compositions are ‘the same’. This perspective naturally leads to homotopy coherent diagrams, where the specified homotopies are accompanied by higher homotopies expressing their compatibility.

Example 1.1.2 (Homological algebra). The phenomenon of ‘equality as structure’ also naturally appears in homological algebra, where it leads us to think about chain complexes not as rigid algebraic objects but as representatives of more flexible homotopical entities:

  • Instead of requiring two chain maps \(f,g\colon C_{\bullet } \to D_{\bullet }\) to be equal, we specify a chain homotopy between them: a family of maps \(\phi _n\colon C_n \to D_{n+1}\) satisfying \(f_n - g_n = \partial ^D \circ \phi _n + \phi _{n-1} \circ \partial ^C\). A chain map admitting an inverse up to chain homotopy is called a chain homotopy equivalence.
  • In practice, one often wants to go further and even identify two chain complexes whenever they are connected by a quasi-isomorphism, that is, a chain map inducing isomorphisms on all homology groups. Localizing at the quasi-isomorphisms gives the derived category of chain complexes.

Example 1.1.3 (Category theory). A third example comes from category theory:

  • For two objects \(X\) and \(Y\) in a 1-category \(C\), the natural notion of equality between them is often not the strict set-theoretic equality, but rather a specified isomorphism \(f\colon X \iso Y\). Universal constructions illustrate this: any two products of a fixed pair of objects are related by a unique isomorphism compatible with their projections.
  • This pattern extends to functors: we wish to think of two functors \(F,G\colon C \to D\) as ‘the same’ whenever we are given a natural isomorphism \(\eta \colon F \iso G\). If we think of the individual isomorphisms \(\eta _X \colon F(X) \iso G(X)\) as identifications, then naturality expresses their compatibility with morphisms.
  • In monoidal categories, the associativity and unitality conditions are expressed not as strict equalities but as natural isomorphisms: for objects \(X\), \(Y\), and \(Z\), we have an associator \(\alpha _{X,Y,Z}\colon (X \otimes Y) \otimes Z \iso X \otimes (Y \otimes Z)\) and unitors \(\lambda _X\colon 1 \otimes X \iso X\) and \(\rho _X\colon X \otimes 1 \iso X\). The coherence conditions (like the Mac Lane pentagon) then govern how these isomorphisms relate to each other.

The common pattern in all these examples is that it is not enough to know that two things are the same: one must also remember the identifications, and often also the compatibilities between different such identifications.

Animae

Once equality is no longer a mere property, the usual notion of a collection also has to change. In traditional set-based mathematics, the primitive notion of a collection is that of a set. The behavior of sets is governed by axioms which strongly reflect the traditional perspective on equality: given a set \(X\), two of its elements \(x\) and \(y\) are either equal or not equal. Their equality carries no additional structure.

If we take the Fundamental Principle fully seriously, we are thus led to a different notion of collection, one in which equality between elements is a structure rather than a property. In recent years, the homotopy theory community has started to adopt the term anima for this concept. In terms of traditional mathematics, one may think of animae as weak homotopy types of topological spaces; we will discuss this viewpoint at length in Chapter 2. But the philosophy we adopt in this book is that animae are primitive objects in their own right, whose role is similar to that of sets in traditional mathematics. In particular, the familiar constructions on sets still have analogues for animae:

  • Given animae \(X\) and \(Y\), there is a product \(X \times Y\), whose elements are pairs \((x,y)\) consisting of an element \(x\) of \(X\) and an element \(y\) of \(Y\);
  • Given animae \(X\) and \(Y\), there is a disjoint union \(X \sqcup Y\), an element of which is either an element of \(X\) or an element of \(Y\);
  • There is an empty anima \(\emptyset \), which does not have any elements;
  • There is a one-point anima \(*\), which has a unique element in the homotopical sense explained below.

The main difference between animae and sets lies in what it means for two elements \(x\) and \(y\) of an anima \(X\) to be equal. Rather than equality being a yes/no question, we should regard identifications between \(x\) and \(y\) as mathematical structure that can be manipulated. The collection of all such identifications should then itself form an anima, which we denote by \((x=_X y)\), or simply by \(x = y\) if the ambient anima is clear. Given two such identifications \(\alpha \) and \(\beta \), we may in turn form the anima \((\alpha =_{(x=_X y)}\beta )\) of identifications between them. This process may be repeated indefinitely, forming identifications between identifications at every level.

The following are some examples of animae:

Example 1.1.4. Every set \(X\) gives rise to an anima, which we will again denote by \(X\). Given two elements \(x,y \in X\), the identification anima \((x=_X y)\) is the one-point anima if \(x\) and \(y\) are equal as elements of the set \(X\), and the empty anima otherwise.

Example 1.1.5. Every 1-category \(C\) gives rise to an anima \(C^{\simeq }\), called its groupoid core. The elements of \(C^{\simeq }\) are the objects of \(C\). The identification anima \((X=_{C^{\simeq }}Y)\) between two objects \(X\) and \(Y\) is the set of isomorphisms between \(X\) and \(Y\) (regarded as an anima via the previous example).

Example 1.1.6. Every topological space \(X\) gives rise to an anima \(\Pi _{\infty }(X)\), called its underlying anima (or fundamental \(\infty \)-groupoid). Every point in \(X\) defines an element of \(\Pi _{\infty }(X)\). Given points \(x,y \in X\), every path \(p\colon [0,1] \to X\) from \(x\) to \(y\) defines an identification in \(\Pi _{\infty }(X)\) between \(x\) and \(y\). Given two such paths \(p\) and \(q\), every homotopy between \(p\) and \(q\) relative to their endpoints defines an identification between them in \((x=_{\Pi _{\infty }(X)}y)\). Homotopies between these homotopies, relative to their boundaries, define identifications between identifications. This pattern repeats ad infinitum. We will study the assignment \(X \mapsto \Pi _{\infty }(X)\) in more detail in Chapter 2.

Just as sets come with maps between them, animae come with maps between them. A map of animae \(f\colon X \to Y\) assigns to every element \(x\) of \(X\) an element \(f(x)\) of \(Y\), to every identification between \(x\) and \(x'\) an identification between \(f(x)\) and \(f(x')\), to every identification between two such identifications an identification between their images, and so on, compatibly with identities and composition at every level. The role played by bijections of sets is now played by equivalences of animae: those maps which admit an inverse, together with the identifications witnessing that both composites are the identity.

Uniqueness

The homotopical perspective on equality also forces us to reconsider uniqueness. Classically, saying that a set \(X\) contains a unique element with some given property means that the subset of all valid choices has exactly one element, or equivalently that the map from this subset to the one-point set \(*\) is a bijection. Replacing sets by animae and bijections by equivalences, the homotopical analogue reads: the anima of all valid choices is contractible, meaning that the map from it to the one-point anima \(*\) is an equivalence. In this sense, contractibility is the homotopical incarnation of ‘there is exactly one’: it says that a choice exists, that any two choices are equal, that any two identifications between the same two choices are themselves identified, and so on.

For example, given elements \(x\) and \(y\) of the terminal anima \(*\), we can express that there exists a unique identification between \(x\) and \(y\) by saying that the anima \((x =_* y)\) is contractible.

The homotopical interpretation of uniqueness appears under various names in the classical literature. In category theory, the phrase ‘up to unique isomorphism’ is often used for objects satisfying a universal property. It means that the anima (in this case: groupoid) of such objects, together with the isomorphisms compatible with their universal structures, is contractible. In homotopy theory, a common phrase is ‘up to contractible choice’. For example, to concatenate \(n\geq 1\) loops in a pointed space, one may first choose a subdivision \(0=t_0<t_1<\cdots <t_n=1\) of the interval. The space of such subdivisions is convex, so its underlying anima is contractible. In this book, all uniqueness statements should be interpreted in the homotopical sense, and we simply say unique.

\(\infty \)-categories

Having replaced sets by animae, we can now make the same replacement in the notion of a category. Instead of a set of objects and sets of morphisms between them, we now allow the objects to form an anima, and for each pair of objects the morphisms between them to form an anima as well. Composition is then accompanied by identifications expressing associativity and unitality, together with higher identifications expressing their coherence. This is the heuristic origin of the notion of an \((\infty ,1)\)-category, or just an \(\infty \)-category for short.1

One should not take this heuristic too literally as a definition: replacing sets by animae suggests the right shape of the theory, but turning this into a workable notion of \(\infty \)-category requires encoding highly non-trivial coherence data for composition. This is one of the reasons why, later in this chapter, we will take \(\infty \)-categories as primitive and describe their behavior axiomatically rather than trying to define them directly from this heuristic.

Let us now go through some basic families of examples of \(\infty \)-categories:

Example 1.1.7. Every classical category \(C\) defines an \(\infty \)-category, whose anima of objects is the groupoid core \(C^{\simeq }\) from Example 1.1.5, and whose hom animae are the hom sets of \(C\).

Example 1.1.8. Every topologically enriched category also defines an \(\infty \)-category, whose hom animae are obtained by applying \(\Pi _{\infty }\) to its hom spaces. We construct this \(\infty \)-category in Section 25.3.

Example 1.1.9. Every anima \(X\) defines an \(\infty \)-category, again denoted \(X\), whose objects are the elements of \(X\) and whose animae of morphisms from \(x\) to \(y\) are the animae \((x=_X y)\) of identifications. Identity morphisms and categorical composition arise from identity identifications and composition of identifications. Inversion of identifications makes every morphism invertible.

In our axiomatic setup, animae form a primitive class of \(\infty \)-categories. Once morphisms and their inverses have been introduced, we will impose the axiom that this class consists exactly of those \(\infty \)-categories in which every morphism is invertible. The animae \((x=_X y)\) from before are then identified with the hom animae \(\Hom _X(x,y)\) of this \(\infty \)-category.

Many familiar constructions from classical category theory also make sense for \(\infty \)-categories, such as products, coproducts, pullbacks, pushouts, functor categories, groupoid cores, subcategories, and localizations.

Notes

1The terminology ‘\(\infty \)-category’ is standard but in some ways unfortunate: all morphisms in degrees greater than 1 are invertible, so the notion does not include arbitrary 2-categories, 3-categories, etcetera.