A monoidal category is a category equipped with some notion of ‘tensor product’. A good example is the category Vect, where we can take the tensor product, not only of vector spaces, but also of linear maps: given linear maps and , we get a linear map
The same category can often be made into a monoidal category in more than one way. For example the category Set can be made into a monoidal category with cartesian product or disjoint union (i.e. coproduct) as the ‘tensor product’. We can also make Vect into a monoidal category with direct sum as the ‘tensor product’ — this may seem perverse, but it’s actually very useful.
For any monoidal category , the operation of tensor product is actually a functor
This functor, which we can think of as a kind of ‘multiplication’, makes into a vertically categorified version of a monoid. This explains the term ‘monoidal category’.
The original list of coherence axioms for monoidal categories given by Mac Lane in 1963 was longer; Max Kelly showed they could be whittled down to just the pentagon and triangle identities. We reproduce his arguments here.
In the proofs below, monoidal product symbols will be suppressed, to save space.
(Kelly 64) In a monoidal category, the equation holds, i.e., the diagram
commutes. Similarly, the following equation holds: .
We prove only the first equation; the proof of the second is entirely analogous. Since the functor is an equivalence (being isomorphic to the identity functor), it suffices to show that the triangle on the right in the diagram below commutes:
where the square in the middle commutes by naturality of , and the triangle on the left commutes by a unit coherence triangle (tensored by on the right). Since all the arrows are isomorphisms, it suffices to show that the diagram formed by the perimeter commutes. But this follows from the commutativity of the diagram
which uses the pentagon coherence condition, naturality of , and a unit coherence condition.
(Kelly 64) The equation holds in a monoidal category.
Since is an equivalence, it suffices to show . But we have the equations
where the first equation follows from Lemma 1 and the second from a unit coherence triangle. One concludes by composing each side of the equations above by .
Strict monoidal categories
A monoidal category is said to be strict if the associator, left unitor and right unitors are all identity morphisms. In this case the pentagon and triangle commute automatically.
The definition of monoidal category looks rather complicated at first sight, so it is natural to wonder if there is some magic wand we can wave that makes it appear automatically. For example, one might wonder if we can define monoidal categories using internalization.
In fact a strict monoidal category is just a monoid internal to the category Cat. Unfortunately this definition is circular, since to define a monoid internal to Cat, we need to use the fact that Cat is a monoidal category! Furthermore, hardly any of the monoidal categories in nature are strict.
Ronnie Brown I entirely understand that most monoidal categories in nature are not strict, and CWM gives an example to show that you cannot even get strictness for the cartesian product. On the other hand, for the cartesian product we get coherence properties directly from the universal property.
Now the tensor product in many monoidal categories in nature comes from the cartesian product, but with more elaborate morphisms. Thus the tensor product of vector spaces comes from bilinear maps. The associativity of this tensor product comes from looking at trilinear maps, and so derives from the associativity of the cartesian product. In a sense, this tensor product is as coherently associative as the cartesian product, which could means that in a rough and ready way we do not need to worry.
My query is whether there is a study of this kind of argument in categorical generality?
Peter LeFanu Lumsdaine: The setting for a statement like this would presumably be the connections between monoidal categories and multicategories, which are discussed very nicely in Chapters 2 and 3 of Tom Leinster’s book?. As far as I remember he doesn’t give anything that would quite make this argument, and I don’t know the literature of these well enough to say whether it’s been done elsewhere, but I’d guess it has, or at least that it would be fairly straightforward to give in that terminology. The statement would look something like:
“If is a multicategory generated by its nullary, unary and binary arrows, its underlying category, and , are functors on representing the nullary and binary arrows of , then and form the tensor and unit of a monoidal structure on .”
The ugly part of this is the generation condition, which will be needed since we only start with and (indeed, some stronger presentation condition might be needed, actually). The unbiased version, where we have not just and but an -ary tensor product for every , is essentially given in Leinster’s book, iirc, and doesn’t require such a condition.
So, we need to weaken the definition of monoidal category, and this is where the subtleties come in: we need the associator, left unitor, and right unitor to satisfy some ‘coherence laws’ — e.g. the pentagon identity.
But where do the coherence laws come from?
In fact, these are precisely what we need to make any diagram built solely by tensoring, associators, and unitors commute. This fact is another version of Mac Lane’s Coherence Theorem. Mac Lane proved it in the same paper where he originally defined the concept of monoidal category.
There are indeed ‘magic wands’ that automatically produce the definition of monoidal category, but most of these magic wands are so heavy that only more advanced wizards can lift them.
For example, you can define a monoidal category to be a pseudomonoid internal to the 2-category Cat — but nobody knew how to define these concepts until they knew what a monoidal category is!
To make the first one work, the magic words to say are “there is a 2-monad whose algebras are strict monoidal categories, and a (non-strict) monoidal category is a pseudo-algebra for that 2-monad.” This doesn’t give you the definition of monoidal categories that we’re used to, though; it gives you the unbiased version.
To make the second magic wand work, the magic words to say are “there is a monad/operad/etc. in Cat whose algebras are strict monoidal categories, and the monad/operad/etc. whose algebras are (non-strict) monoidal categories is a cofibrant replacement for that one.” Since cofibrant replacements are usually defined only up to equivalence, this one also doesn’t determine the usual definition uniquely. There is a “canonical” choice, but again it gives you the unbiased version. The equation “cofibrant = flexible” says that these two magic wands are doing essentially the same thing.
Of course, both are also sort of a cheat, since in order to prove that the biased and unbiased definitions are equivalent, you need to have the coherence theorem for the biased definition. However, it’s only because of the coherence theorem that we can say definitely that the usual set of complicated-looking diagrams is “correct.” The approach using lax -functors really only postpones this question, since you also need a coherence theorem to show that the definition of lax -functor is “correct.” So perhaps there is no magic wand after all, at least not one that produces the specific diagrams in the usual biased definition of monoidal category.
However, if we temporarily ignore the unitors and focus on the associator, we may ask where does the pentagon identity come from? And one answer to this is provided by the Stasheff polytopes, which can be nicely obtained using Ross Street’s theory of orientals. For instance the pentagon diagram above is nothing but the 4th oriental! The tensor product itself is the second oriental, and the associator the third. The following section explains this in a bit more detail.
Relation to lax functors, orientals and descent
One can understand the structure of a monoidal category as a special simple case of the general notion of “lax -functor”, also known – up to the issue of invertible versus non-invertible structure morphisms – as the notion of -categorical descent and as the notion of infinity-anafunctor.
This may be familiar from the special simple case of a monoid in any bicategory , which can be identified with a lax functor
from the point to . This lax functor sends the point to some object of , sends the identity morphism on the point to some endomorphism of that object. The unitor of the lax functor gives the product on that endomorphism and the coherence of the unitor is the associativity condition on this product.
This is part of a more general principle. A lax monoid in any tricategory would again be a lax functor from the point to that tricategory.
And a monoidal category can be regarded as a pseudomonoid in the tricategory , which has a single object, categories as 1-morphisms with the composition of 1-morphisms being the standard cartesian tensor product on categories.
Evidently, in the fully general context of weak -categories it becomes increasingly hard to state what a lax functor into a given -category should be: it will involve a plethora of structure morphisms and their coherences. One task of higher category theory is to organize this mess into something pretty and then to deal with this problem.
But before being intimidated by the problem in its most general form, it may pay to understand it in slightly simplified situations. One such slightly simplified setup is that of strict -categories, usually known as -categories or strict omega-categories.
For that case, Ross Street has given a general combinatorial formula for the -coherence law of the general monoidal structure: this is encoded in the orientals, which are nothing but the standard simplicial simplices, but equipped with extra information about source and targets of all faces.
See the picture of the first five orientals. We can read off the above definition of a monoidal category from them as follows:
identify the monoidal category itself with the first oriental, just an arrow;
identify the ambient product with the juxtaposition of two such arrows;
identify the tensor product with the second oriental: a triangular cell going from the concatenation of two arrows to a single arrow;
identify the associator with the third oriental, the tetrahedron: a map from one way to compose three arrows (=copies of ) to the other way of doing this;
identify the pentagon identity with the fourth oriental. In general, the fourth oriental is itself a nontrivial 4-cell, but assume now that the big arrow in the middle of that is the identity. This makes what in general would be the pentagonator, the pentagon identity in this case.
This shows that it is a bit of an illusion to think of a pentagon identity: the full geometric shape is really a 4-dimensional tetrahedron (the 4-simplex) whose five tetrahedral faces are the five vertices of the pentagon identity.
We can formulate this identification of structure morphisms and coherence laws with orientals more formally using the general notion of descent, which was indeed the original motivation for conceiving the orientals. The descent -category (constructed in terms of orientals) can be regarded as a way to formalize “lax -functor from to ”.
Indeed, using observations pretty much as just sketched, one finds that for a 2-category that
and for the 3-category we have
where the 2-category on the right is defined as above, but with the associator not required to be an isomorphism.
Remark: pseudo versus lax, orientals versus unorientals
In closing, it should be remarked that the fact that everything here is lax instead of pseudo is related to a curious property of the orientals: the th oriental for fails to be weakly equivalent to the point. As a result, the objects of are not quite -anafunctors from the point to , since they do not map out of a proper hypercover of . In the strict notion of descent as used in most of the literature, the orientals would hence provide something more general than ordinary descent, which in its generality is lacking some properties usually required of descent.
We can remedy this by replacing in the definition of the descent -category the orientals by another cosimplicial -category, one which is equivalent to the point in each degree. Doing so and then going through the above discussion will make all the structure maps appeaing have inverses. But this will also apply to the monoidal product itself, then, which is usually not desired.
A more detailed tour of monoidal categories, also using string diagrams, and including autonomous, balanced, braided, compact closed, pivotal, ribbon, rigid, sovereign, spherical, tortile, and traced monoidal categories: