Formalising and Computing the Fourth Homotopy Group of the $3$-Sphere in Cubical Agda

Homotopy theory originated in algebraic topology, but is by now a central tool in many branches of modern mathematics, such as algebraic geometry and category theory. One of the central notions of study in homotopy theory is that of the homotopy groups of a space $X$, denoted $\pi_{n}(X)$. These groups constitute a topological invariant, making them a powerful tool for establishing whether two given spaces can or cannot be homotopy equivalent. The first two such groups of a space are easy to understand: $\pi_{0}(X)$ characterises the connected components of $X$ and $\pi_{1}(X)$ is the fundamental group, i.e. the group of equivalence classes consisting of the loops contained in $X$ up to homotopy. This idea generalises to higher values of $n$, for which $\pi_{n}(X)$ consists of $n$-dimensional loops up to homotopy. For many spaces, these groups tend to become increasingly esoteric and difficult to compute for large $n$. This is true also for seemingly tame spaces like spheres, for which $\pi_{n}(\mathbb{S}^{m})$ in general is highly irregular when $n>m\geq 2$.¹¹1See [Bru16a, Figure 2.1] for a table of $\pi_{n}(\mathbb{S}^{m})$ for small $n$ and $m$. This paper concerns the first computer formalisation of the classical result that $\pi_{4}(\mathbb{S}^{3})\cong\mathbb{Z}/2\mathbb{Z}$, a result which is particularly interesting because it gives the whole first stable stem of homotopy groups of spheres, i.e. $\pi_{n+1}(\mathbb{S}^{n})$ for $n\geq 3$. The fact that $\pi_{4}(\mathbb{S}^{3})\cong\mathbb{Z}/2\mathbb{Z}$ was proved already in the 1930’s by Pontryagin using cobordism theory, but we instead follow the synthetic approach to homotopy theory developed in Homotopy Type Theory (HoTT) and popularised by the HoTT Book [Uni13]. In this new approach to homotopy theory, spaces are represented directly as (higher inductive) types and homotopy groups are computed using Voevodsky’s univalence axiom [Voe10a]. This gives a logical approach to homotopy theory, suitable for computer formalisation in proof assistants based on type theory, while also making it possible to interpret results in any suitably structured $(\infty,1)$-topos [Shu19].

The basis for our formalisation is the 2016 PhD thesis of Brunerie [Bru16a] which contains the first synthetic proof in HoTT that $\pi_{4}(\mathbb{S}^{3})\cong\mathbb{Z}/2\mathbb{Z}$. The proof is one of the most impressive pieces of synthetic homotopy theory to date and uses advanced machinery from classical algebraic topology developed synthetically, including the symmetric monoidal structure of smash products, (integral) cohomology rings, the Mayer-Vietoris and Gysin sequences, the Hopf invariant, Whitehead products, etc. The formalisation of Brunerie’s proof has since remained open, primarily due to the highly technical nature of some of the details. In this paper, we will present such a formalisation in $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Cubical}$ $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Agda}$ [VMA21], a cubical extension of the $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Agda}$ proof assistant [Agd24] with native support for computational univalence and higher inductive types (HITs).

In addition to being a very impressive proof in synthetic homotopy theory, Brunerie’s proof is particularly interesting as it is fully constructive. The proof consists of two parts, with the first one culminating in Chapter 3 with the definition of a number $\beta:\mathbb{Z}$ such that $\pi_{4}(\mathbb{S}^{3})\cong\mathbb{Z}/\beta\mathbb{Z}$. Since then, this $\beta$ has been commonly referred to as the Brunerie number. Brunerie writes the following about it:

This result is quite remarkable in that even though it is a constructive proof, it is not at all obvious how to actually compute this [$\beta$]. At the time of writing, we still haven’t managed to extract its value from its definition. [Bru16a, Page 85]

In fact, [Bru16a, Appendix B] contains a complete and concise definition of $\beta$ as the image of $1$ under a sequence of $12$ maps:

By implementing this number in a proof assistant with computational support for univalence and HITs, one should be able to normalise it using a computer to establish that $\beta=\pm 2$ and hence that $\pi_{4}(\mathbb{S}^{3})\cong\mathbb{Z}/2\mathbb{Z}$. In 2016, by the time Brunerie was finishing his thesis, there were some experimental proof assistants based on the cubical type theory of [CCHM18], but these were too slow to perform such a complex computation. So, instead of relying on normalisation, Brunerie spends the second part of the thesis (Chapters 4–6) to prove, using a lot of the advanced machinery mentioned above, that $|\beta|$ is propositionally equal to $2$. However, if one were instead able to compute the number automatically in a proof assistant, this equality would hold definitionally—effectively reducing the complexity and length of the proof by an order of magnitude.

The intriguing possibility of a computer assisted formal proof made many people interested and countless attempts to normalise Brunerie’s $\beta$ have been made using increasingly powerful computers. However, to date, no one has succeeded and it is still unclear whether it is normalisable in a reasonable amount of time. In light of this, it is natural to wonder whether it is possible to simplify Brunerie’s number in order to be able to compute it. For example, Brunerie’s original definition only involves 1-HITs, as the status of higher HITs was still quite understudied at the time. With a better understanding of higher HITs developed in [LS20, CHM18, CH19], one quickly sees that the first $3$ maps can be combined into one sending $1$ to the 3-cell of $\mathbb{S}^{3}$ defined as a 3-HIT and not as an iterated suspension as in Brunerie’s thesis. Unfortunately, simple optimisations like this do not seem to reduce the complexity of the computation enough and all attempts to run it have thus far failed.

After several unsuccessful attempts at optimising the computation, we instead decided to formalise the second half of Brunerie’s thesis. However, this is by no means straightforward. The first issue appears already in the beginning of Chapter 4, a chapter concerning smash products of spheres. The main result of the section is Proposition 4.1.2, which says that the smash product is a 1-coherent symmetric monoidal product on pointed types. However, the proof of this result is just a sketch and Brunerie writes the following about it:

The following result is the main result of this section even though we essentially admit it. [Bru16a, Page 90]

Unfortunately, this result is then used to construct integral cohomology rings, $H^{*}(X)$, whose cup product, $\smile$, appears in the definition of the so-called Hopf invariant which is crucially used to prove that $|\beta|$ is $2$. While one might be convinced that Brunerie’s informal proof sketch is correct, it is not obvious how one convinces a proof assistant of this. A complete formalisation would either have to fill in the holes in the sketch or find an alternative construction which avoids Proposition 4.1.2. In fact, Brunerie tried very hard to fill these holes using $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Agda}$ metaprogramming [Bru18]. However, he never managed to typecheck his computer generated proof of the pentagon identity. Hence, this approach also seems infeasible with current proof assistant technology.

Luckily, Brunerie, Ljungström and Mörtberg [BLM22] recently gave an alternative synthetic definition of the cup product on $H^{*}(X)$ which completely avoids smash products. This has allowed us to completely skip the problematic Chapter 4 and, in particular, Proposition 4.1.2, while still following the proofs in Chapters 5 and 6. Having a strategy for a formal proof, we were then able to embark on able to embark on the ambitious project of formalising Brunerie’s proof. Even though we do not need any theory about smash products, there was still a lot left to formalise and our final formalisation closely follows Brunerie’s proof, except for various smaller simplifications and adjustments which we discuss in the paper.

In addition to this, we have also formalised a new proof by Ljungström [Lju22] which completely circumvents Chapters 4–6. This major simplification builds on manually calculating the image of the element $\eta:\pi_{3}(\mathbb{S}^{2})$, corresponding to $\beta$ under the isomorphism $\pi_{3}(\mathbb{S}^{2})\cong\mathbb{Z}$, by dividing this isomorphism into several maps, tracing $\eta$ in each step. In particular, the new proof is completely elementary and does not rely on advanced tools such as cohomology. The elements that one obtains while tracing $\eta$ are all new “Brunerie numbers” that should normalise to $\pm 2$. In fact, one of these normalises, in just under $4$ seconds on a regular laptop, to $-2$ in $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Cubical}$ $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Agda}$ at the time of writing. So, despite still not being able to compute the original $\beta$, this work can be seen as an alternative solution to Brunerie’s conjecture about obtaining a computational proof that $\pi_{4}(\mathbb{S}^{3})\cong\mathbb{Z}/2\mathbb{Z}$ which relies on simplifying the Brunerie number until it becomes effectively computable.

In this section, we concisely summarise the key HoTT concepts needed for the proofs and their formalisation in $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Cubical}$ $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Agda}$. This roughly corresponds to [Bru16a, Chapter 1]. For a more in-depth introduction, see the HoTT Book [Uni13] which also serves as a reference for the formal language “Book HoTT”. In this paper, we will present many things with cubical notations, but almost all of the results also hold with minor changes in Book HoTT where paths are represented using Martin-Löf’s inductive Id-types [ML75] instead of cubical path types. In section 7 we discuss in more detail which proofs crucially rely on cubical features.

All of the results presented in this section were already part of the $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}agda/cubical}$ library before we began our formalisation and, while useful as a resource for our notations, experts on HoTT and $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Cubical}$ $\mathtt{\color[rgb]{.25,.25,.25}\definecolor[named]{pgfstrokecolor}{rgb}{.25,.25,.25}\pgfsys@color@gray@stroke{.25}\pgfsys@color@gray@fill{.25}Agda}$ can safely skim this section.

In this section, we cover [Bru16a, Chapter 2], which introduces some elementary results on the homotopy groups of spheres. All of these results can also be found in the HoTT Book [Uni13]. Before even stating them, we need homotopy groups: {defi}[Homotopy groups] For $n:\mathbb{N}$, we define the $n$th homotopy group of a pointed type $A$ by:

The name homotopy group should be taken with a grain of salt: it, in general, only has a group structure when $n\geq 1$ (abelian when $n\geq 2$). The structure may be defined, much like in [BHF18, Section 5], by considering the equivalence ${(\sphere{n}\to_{\star}A)}\,\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{≃}}\,\,{(\sphere{n-1}\to_{\star}{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\Omega$}}\,{A})}$, where the latter type has a multiplication given by pointwise path composition. An alternative definition of ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\pi$}}_{n}(A)$ is via loop spaces. There is an equivalence $\omega_{n}:{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\Omega$}}^{n}\,{A}}\,\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{≃}}\,\,{(\sphere{n}\to_{\star}A)}$ and, hence, we could equivalently have defined ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\pi$}}_{n}(A)$ by setting ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\pi$}}_{n}(A)={\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\Omega$}}^{n}\,{A}\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{0}$. This makes the group structure on ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\pi$}}_{n}(A)$ more transparent: it is simply path composition. This is the definition used in the HoTT Book [Uni13]. Brunerie uses both definitions in his thesis and often passes between the two without comment.

An elementary but crucial result for the computation of homotopy groups is the existence of the long exact sequence of homotopy groups. Its proof is usually phrased using the loop space definition of homotopy groups as in e.g. [Uni13, Theorem 8.4.6]. For ease of notation, let us write $\textnormal{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}{fib}}}\,{f}$ for the fibre of a pointed function $f:A\to_{\star}B$ over the basepoint of $B$.

Above, we have implicitly taken the kernel and image of a group homomorphism $\phi:G\to H$ to be defined by

When analysing loop spaces and homotopy groups of suspensions, the following function is of great importance. It will be used in many constructions to come. {defi}[The suspension map] Given a pointed type $A$, there is a canonical map ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\sigma$}}:A\to{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\Omega$}}\,{(}\textnormal{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}{Susp}}{$\,A$}})$ given by

This induces a homomorphism on homotopy groups by post-composition:

We will often, with some abuse of notation, simply write ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\sigma$}}_{*}$ for this composition. We also define ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\sigma$}}_{n}:{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,A\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{n}\to{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\Omega$}}\,{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,\textnormal{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}{Susp}}{$\,A$}}\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{n+1}}$ by

We will soon see the suspension map in action, but first we need the following elementary result.

In fact, as we will see in section 6, there is more to say about this equivalence. We make a forwards reference to 32 and the preceding discussion for a detailed account of its construction.

In particular, 2 gives us an equivalence ${{\sphere{1}}{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{\text{\textasteriskcentered}}}\,{\sphere{1}}}\,\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{≃}}\,\,{\sphere{3}}$. Using this fact, we define the following map, which will play a crucial role in the analysis of ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{$\pi$}}_{4}(\sphere{3})$. {defi}[Hopf map] We define $\textnormal{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}{hopf}}}:\sphere{3}\to\sphere{2}$ by the composition $\sphere{3}\xrightarrow{\sim}{\sphere{1}}{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{\text{\textasteriskcentered}}}\,{\sphere{1}}\xrightarrow{\textnormal{{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}{h}}}}\sphere{2}$ where h is given by \ExecuteMetaData[agda/latex/BruneriesProof.tex]hopf where $-$ is defined using the H-space and inversion structure on $\sphere{1}$. It turns out that the following is true [Uni13, Theorem 8.5.1].

3 gives us a fibration sequence $\sphere{1}\to\sphere{3}\to\sphere{2}$ which, in particular, will allow us to connect homotopy groups of $\sphere{2}$ with those of $\sphere{3}$ and $\sphere{1}$. For this, we need to introduce the notion of connectedness. We say that a type $A$ is $n$-connected if ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,A\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{n}$ is contractible. Similarly, we say that a function $f:A\to B$ is $n$-connected if all of its fibres are $n$-connected. This means, in particular, that the induced function ${\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,f\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{n}:{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,A\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{n}\to{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}\,B\,{\color[rgb]{0,0,0.8046875}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0.8046875}\textsf{∥}}_{n}$ is an equivalence. The following is an immediate consequence of the definition of $n$-truncations.

Using 4, we can easily prove the following: