## A Touch of Topological Quantum Computation 3: Categorical Interlude

Welcome back, friend.

In the last two posts, I described the basics of how to build and manipulate the Fibonacci anyon vector space in Haskell.

As a personal anecdote, trying to understand the category theory behind the theory of anyons is one of the reasons I started learning Haskell. These spaces are typically described using the terminology of category theory. I found it very frustrating that anyons were described in an abstract and confusing terminology. I really wondered if people were just making things harder than they have to be. I think Haskell is a perfect playground to clarify these constructions. While the category theory stuff isn’t strictly necessary, it is interesting and useful once you get past the frustration.

but I hope everyone can get something out of it. Give it a shot if you’re interested, and don’t sweat the details.

#### The Aroma of Categories

I think Steve Awodey gives an excellent nutshell of category theory in the introductory section to his book:

“What is category theory? As a first approximation, one could say that category theory is the mathematical study of (abstract) algebras of functions. Just as group theory is the abstraction of the idea of a system of permutations of a set or symmetries of a geometric object, so category theory arises from the idea of a system of functions among some objects.”

For my intuition, a category is any “things” that plug together. The “in” of a thing has to match the “out” of another thing in order to hook them together. In other words, the requirement for something to be a category is having a notion of composition. The things you plug together are called the morphisms of the category and the matching ports are the objects of the category. The additional requirement of always having an identity morphism (a do-nothing connection wire) is usually there once you have composition, although it is good to take especial note of it.

Category theory is an elegant framework for how to think about these composing things in a mathematical way. In my experience, thinking in these terms leads to good abstractions, and useful analogies between disparate things.

It is helpful for any abstract concept to list some examples to expose the threads that connect them. Category theory in particular has a ton of examples connecting to many other fields because it is a science of analogy. These are the examples of categories I usually reach for. Which one feels the most comfortable to you will depend on your background.

• Hask. Objects are types. Morphisms are functions between those types
• Vect. Objects are vector spaces, morphisms are linear maps (roughly matrices).
• Preorders. Objects are values. Morphisms are the inequalities between those values.
• Sets. Objects are Sets. Morphisms are functions between sets.
• Cat. Objects are categories, Morphisms are functors. This is a pretty cool one, although complete categorical narcissism.
• Systems and Processes.
• The Free Category of a directed graphs. Objects are vertices. Morphisms are paths between vertices

#### Generic Programming and Typeclasses

The goal of generic programming is to run programs that you write once in many way.

There are many ways to approach this generic programming goal, but one way this is achieved in Haskell is by using Typeclasses. Typeclasses allow you to overload names, so that they mean different things based upon the types involved. Adding a vector is different than adding a float or int, but there are programs that can be written that reasonably apply in both situations.

Writing your program in a way that it applies to disparate objects requires abstract ways of talking about things. Mathematics is an excellent place to mine for good abstractions. In particular, the category theory abstraction has demonstrated itself to be a very useful unified vocabulary for mathematical topics. I, and others, find it also to be a beautiful aesthetic by which to structure programs.

In the Haskell base library there is a Category typeclass defined in base. In order to use this, you need to import the Prelude in an unusual way.

The Category typeclass is defined on the type that corresponds to the morphisms of the category. This type has a slot for the input type and a slot for the output type. In order for something to be a category, it has to have an identity morphisms and a notion of composition.

The most obvious example of this Category typeclass is the instance for the ordinary Haskell function (->). The identity corresponds to the standard Haskell identity function, and composition to ordinary Haskell function composition.

Another example of a category that we’ve already encountered is that of linear operators which we’ll call LinOp. LinOp is an example of a Kliesli arrow, a category built using monadic composition rather than regular function composition. In this case, the monad Q from my first post takes care of the linear pipework that happens between every application of a LinOp. The fish <=< operator is monadic composition from Control.Monad.

A related category is the FibOp category. This is the category of operations on Fibonacci anyons, which are also linear operations. It is LinOp specialized to the Fibonacci anyon space. All the operations we've previously discussed (F-moves, braiding) are in this category.

The "feel" of category theory takes focus away from the objects and tries to place focus on the morphisms. There is a style of functional programming called "point-free" where you avoid ever giving variables explicit names and instead use pipe-work combinators like (.), fst, snd, or (***). This also has a feel of de-emphasizing objects. Many of the combinators that get used in this style have categorical analogs. In order to generically use categorical typeclasses, you have to write your program in this point free style.

It is possible for a program written in the categorical style to be a reinterpreted as a program, a linear algebra operation, a circuit, or a diagram, all without changing the actual text of the program. For more on this, I highly recommend Conal Elliot's  compiling to categories, which also puts forth a methodology to avoid the somewhat unpleasant point-free style using a compiler plug-in. This might be an interesting place to mine for a good quantum programming language. YMMV.

### Monoidal Categories.

Putting two processes in parallel can be considered a kind of product. A category is monoidal if it has this product of this flavor, and has isomorphisms for reassociating objects and producing or consuming a unit object. This will make more sense when you see the examples.

We can sketch out this monoidal category concept as a typeclass, where we use () as the unit object.

#### Instances

In Haskell, the standard monoidal product for regular Haskell functions is (***) from Control.Arrow. It takes two functions and turns it into a function that does the same stuff, but on a tuple of the original inputs. The associators and unitors are fairly straightforward. We can freely dump unit () and get it back because there is only one possible value for it.

The monoidal product we'll choose for LinOp is the tensor/outer/Kronecker product.

Otherwise, LinOp is basically a monadically lifted version of (->). The one dimensional vector space Q () is completely isomorphic to just a number. Taking the Kronecker product with it is basically the same thing as scalar multiplying (up to some shuffling).

Now for a confession. I made a misstep in my first post. In order to make our Fibonacci anyons jive nicely with our current definitions, I should have defined our identity particle using type Id = () rather than data Id. We'll do that now. In addition, we need some new primitive operations for absorbing and emitting identity particles that did not feel relevant at that time.

With these in place, we can define a monoidal instance for FibOp. The extremely important and intriguing F-move operations are the assoc operators for the category. While other categories have assoc that feel nearly trivial, these F-moves don't feel so trivial.

#### This is actually useful

The parC operation is extremely useful to explicitly note in a program. It is an opportunity for optimization. It is possible to inefficiently implement parC in terms of other primitives, but it is very worthwhile to implement it in new primitives (although I haven't here). In the case of (->), parC is an explicit location where actual computational parallelism is available. Once you perform parC, it is not longer obviously apparent that the left and right side of the tuple share no data during the computation. In the case of LinOp and FibOp, parC is a location where you can perform factored linear computations. The matrix vector product $(A \otimes B)(v \otimes w)$ can be performed individually $(Av)\otimes (Bw)$. In the first case, where we densify $A \otimes B$ and then perform the multiplication, it costs $O((N_A N_B)^2)$ time, whereas performing them individually on the factors costs $O( N_A^2 + N_B^2)$ time, a significant savings. Applied category theory indeed.

#### Laws

Like many typeclasses, these monoidal morphisms are assumed to follow certain laws. Here is a sketch (for a more thorough discussion check out the wikipedia page):

• Functions with a tick at the end like assoc' should be the inverses of the functions without the tick like assoc, e.g. assoc . assoc' = id
• The parC operation is (bi)functorial, meaning it obeys the commutation law parC (f . f') (g . g') = (parC f g) . (parC f' g') i.e. it doesn't matter if we perform composition before or after the parC.
• The pentagon law for assoc: Applying leftbottom is the same as applying topright
• The triangle law for the unitors:

#### String Diagrams

String diagrams are a diagrammatic notation for monoidal categories. Morphisms are represented by boxes with lines.

Composition g . f is made by connecting lines.

The identity id is a raw arrow.

The monoidal product of morphisms $f \otimes g$ is represented by placing lines next to each other.

The diagrammatic notion is so powerful because the laws of monoidal categories are built so deeply into it they can go unnoticed. Identities can be put in or taken away. Association doesn't even appear in the diagram. The boxes in the notation can naturally be pushed around and commuted past each other.

This corresponds to the property

$(id \otimes g) \circ (f \otimes id) = (f \otimes id) \circ (id \otimes g)$

What expression does the following diagram represent?

Is it $(f \circ f') \otimes (g \circ g')$ (in Haskell notation parC (f . f') (g . g') )?

Or is it $(f \otimes g) \circ (f' \otimes g')$ (in Haskell notation (parC f g) . (parC f' g')?

Answer: It doesn't matter because the functorial requirement of parC means the two expressions are identical.

There are a number of notations you might meet in the world that can be interpreted as String diagrams. Three that seem particular pertinent are:

• Quantum circuits
• Anyon Diagrams!

#### Braided and Symmetric Monoidal Categories: Categories That Braid and Swap

Some monoidal categories have a notion of being able to braid morphisms. If so, it is called a braided monoidal category (go figure).

The over and under morphisms are inverse of each other over . under = id. The over morphism pulls the left morphism over the right, whereas the under pulls the left under the right. The diagram definitely helps to understand this definition.

These over and under morphisms need to play nice with the associator of the monoidal category. These are laws that valid instance of the typeclass should follow. We actually already met them in the very first post.

If the over and under of the braiding are the same the category is a symmetric monoidal category. This typeclass needs no extra functions, but it is now intended that the law over . over = id is obeyed.

When we draw a braid in a symmetric monoidal category, we don't have to be careful with which one is over and under, because they are the same thing.

The examples that come soonest to mind have this symmetric property, for example (->) is a symmetric monoidal category..

Similarly LinOp has an notion of swapping that is just a lifting of swap

However, FibOp is not symmetric! This is perhaps at the core of what makes FibOp so interesting.

#### Automating Association

Last time, we spent a lot of time doing weird typelevel programming to automate the pain of manual association moves. We can do something quite similar to make the categorical reassociation less painful, and more like the carefree ideal of the string diagram if we replace composition (.) with a slightly different operator

Before defining reassoc, let's define a helper LeftCollect typeclass. Given a typelevel integer n, it will reassociate the tree using a binary search procedure to make sure the left branch l at the root has Count l = n.

Once we have LeftCollect, the typeclass ReAssoc is relatively simple to define. Given a pattern tree, we can count the elements in it's left branch and LeftCollect the source tree to match that number. Then we recursively apply reassoc in the left and right branch of the tree. This means that every node has the same number of children in the tree, hence the trees will end up in an identical shape (modulo me mucking something up).

It seems likely that one could write equivalent instances that would work for an arbitrary monoidal category with a bit more work. We are aided somewhat by the fact that FibOp has a finite universe of possible leaf types to work with.

### Closing Thoughts

While our categorical typeclasses are helpful and nice, I should point out that they are not going to cover all the things that can be described as categories, even in Haskell. Just like the Functor typeclass does not describe all the conceptual functors you might meet. One beautiful monoidal category is that of Haskell Functors under the monoidal product of Functor Composition. More on this to come, I think. https://parametricity.com/posts/2015-07-18-braids.html

We never even touched the dot product in this post. This corresponds to another doodle in a string diagram, and another power to add to your category. It is somewhat trickier to work with cleanly in familiar Haskell terms, I think because (->) is at least not super obviously a dagger category?

You can find a hopefully compiling version of all my snippets and more in my chaotic mutating Github repo https://github.com/philzook58/fib-anyon

See you next time.

#### References

The Rosetta Stone paper by Baez and Stay is probably the conceptual daddy of this entire post (and more).

Bartosz Milewski's Category Theory for Programmer's blog (online book really) and youtube series are where I learned most of what I know about category theory. I highly recommend them (huge Bartosz fanboy).

https://www.math3ma.com/blog/what-is-category-theory-anyway

There are fancier embeddings of category theory and monoidal categories than I've shown here. Often you want constrained categories and the ability to choose unit objects. I took a rather simplistic approach here.

https://parametricity.com/posts/2015-07-18-braids.html

## Compiling to Categories 3: A Bit Cuter

Ordinary Haskell functions form a cartesian closed category. Category means you can compose functions. Cartesian basically means you can construct and deconstruct tuples and Closed means that you have first class functions you can pass around.

Conal Elliott’s Compiling to Categories is a paradigm for reinterpreting ordinary functions as the equivalent in other categories. At an abstract level, I think you could describe it as a mechanism to build certain natural law abiding Functors from Hask to other categories. It’s another way to write things once and have them run many ways, like polymorphism or generic programming. The ordinary function syntax is human friendly compared to writing raw categorical definitions, which look roughly like point-free programming (no named variables). In addition, by embedding it as a compiler pass within GHC, he gets to leverage existing GHC optimizations as optimizations for other categories. Interesting alternative categories include the category of graphs, circuits, and automatically differentiable functions. You can find his implementation here

I’ve felt hesitance at using a GHC plugin plus I kind of want to do it in a way I understand, so I’ve done different versions of this using relatively normal Haskell (no template haskell, no core passes, but a butt ton of hackery).

The first used explicit tags. Give them to the function and see where they come out. That is one way to probe a simple tuple rearrangement function.

The second version worked almost entirely at the typelevel. It worked on the observation that a completely polymorphic tuple type signature completely specifies the implementation. You don’t have to look at the term level at all. It unified the polymorphic values in the input with a typelevel number indexed Tag type. Then it searched through the input type tree to find the piece it needed. I did end up passing stuff in to the term level because I could use this mechanism to embed categorical literals. The typeclass hackery I used to achieve this all was heinous.

I realized today another way related to both that is much simpler and fairly direct. It has some pleasing aesthetic properties and some bad ones. The typeclass hackery is much reduced and the whole thing fits on a screen, so I’m pleased.

Here are the basic categorical definitions. FreeCat is useful for inspection in GHCi of what comes out of toCCC.

And here is the basic toCCC implementation

Here is some example usage

What we do is generate a tuple to give to your function. The function is assumed to be polymorphic again but now not necessarily totally polymorphic (this is important because Num typeclass usage will unify variables). Once we hit a leaf of the input tuple, we place the categorical morphism that would extract that piece from the input. For example for the input type (a,(b,c)) we pass it the value (fstC ,(fstC . sndC, sndC . sndC )). Detecting when we are at a leaf again requires somehow detecting a polymorphic location, which is a weird thing to do. We use the Incoherent IsTup instance from last time to do this. It is close to being safe, because we immediately unify the polymorphic variable with a categorial type, so if something goes awry, a type error should result. We could make it more ironclad by unifying immediately to something that contains the extractor and a user inaccessible type.

We apply the function to this input. Now the output is a tuple tree of morphisms. We recursively traverse down this tree with a fanC for every tuple. This all barely requires any typelevel hackery. The typelevel stuff that is there is just so that I can traverse down tuple trees basically.

One benefit is that we can now use some ordinary typeclasses. We can make a simple implementation of Num for (k z a) like how we would make it for (z -> a). This let’s us use the regular (+) and (*)  operators for example.

What is not good is the performance. As it is, the implementation takes many global duplication of the input to create all the fanCs. In many categories, this is very wasteful.This may be a fixable problem, either via passing in more sophisticated objects that just the bare extraction morphisms to to input (CPS-ified? Path Lists?) or via the GHC rewrite rules mechanism. I have started to attempt that, but have not been successful in getting any of my rewrite rules to fire yet, because I have no idea what I’m doing. If anyone could give me some advice, I’d be much obliged. You can check that out here. For more on rewrite rules, check out the GHC user manual and this excellent tutorial by Mark Karpov here.

Another possibility is to convert to FreeCat, write regular Haskell function optimization passes over the FreeCat AST and then interpret it. This adds interpretation overhead, which may or may not be acceptable depending on your use case. It would probably not be appropriate for automatically differentiable functions, but may be for outputting circuits or graphs.

Another problem is dealing with boolean operations. The booleans operators and comparison operators are not sufficiently polymorphic in the Prelude. We could define new operators that work as drop in replacements in the original context, but I don’t really have the ability to overload the originals. It is tough because if we do things like this, it feels like we’re really kind of building a DSL more than we are compiling to categories. We need to write our functions with the DSL in mind and can’t just import and use some function that had no idea about the compiling to categories stuff.

I should probably just be using Conal’s concat project. This is all a little silly.

## Resources on String Diagrams, and Adjunctions, and Kan Extensions

I’ve been trying to figure out Kan Extensions

Ralf Hinze on Kan Extensions

https://www.cs.ox.ac.uk/ralf.hinze/Kan.pdf

But while doing that I went down a rabbit hole on String Diagrams

This post is the first one on String Diagrams that made sense to me.

https://parametricity.com/posts/2015-07-18-braids.html

I had seen this stuff before, but I hadn’t appreciated it until I saw what Haskell expressions it showed equivalence between. They are not obvious equivalences

This seems like a very useful video on this topic.

https://skillsmatter.com/skillscasts/8807-categories-and-string-diagrams

In Summary, it is an excellent notation for talking about transformations of a long sequence of composed Functors  F G H … into some other long sequence of Functors. The conversion of functors runs from up to down. The composition of functors procedes left to right.  F eta is the fmap of eta, and eta F is eta with the forall’ed type unified to be of the form F a.

Adjunctions L -| R are asymmetric between cups and caps. L is on the left in cups and on the right in caps. That’s what makes squiggles pull straightable

I think I have an interesting idea for a linear algebra library based on this stuff

John Baez and Mike Stay’s Rosetta Stone (A touch stone I keep returning to)

math.ucr.edu/home/baez/rosetta.pdf

Dan Piponi gave a talk which is another touch stone of mine that I come back to again and again. There is a set of corresponding blog posts.

Other resources:

NCatLab article

https://ncatlab.org/nlab/show/string+diagram

John Baez hosted seminars

http://www.math.ucr.edu/home/baez/QG.html

Catsters

Dan Marsden’s Article

https://arxiv.org/abs/1401.7220

Marsden and Hinze have been collaborating

events.inf.ed.ac.uk/wf2016/slides/hinze.pdf

A Section From an old Oregon Programming Language Summer School (a rich set of resources)

Marsden and Hinze have been collaborating

events.inf.ed.ac.uk/wf2016/slides/hinze.pdf

Mike Stay doing a very interesting series of Category Theory in Javascript. He uses contracts in place of types. Defeats one of the big points of types (static analysis), but still pretty cool

Oh yeah, there is the whole Coecke and Abramsky categorical quantum mechanics stuff too.

## Contracts and Category Theory

Contracts are both less cryptic and less impressive than a type system.

A type system somehow proves that things will never happen at compile time. You’ll never get a string when you expected an int or even more powerful facts. Contracts just guarantee that at least if certain problems occur at least the program will freak out and tell you at runtime.

I had been vaguely aware of contracts which I considered to be kind of a band aid Racket thing that is their solution to type safety (besides typed racket), but I did not go into depth. And I kind of viewed the thing more as a methodology for showing loop invariants and algorithmic correctness rather than type correctness. I do not know if this is an accurate portrayal of what is in Racket and I’m pretty sure contracts do not actually originate there (Eiffel?).

Mike Stay (who you may know as the co-author of the Rosetta Stone paper https://arxiv.org/abs/0903.0340)made a bunch of videos which I don’t know how I didn’t come across before (they’re kind of old by skateboarding millennial mountain dew front end developer standards. Did Node even exist? Barely.). Javascript (well maybe after python) was my language of greatest comfort a couple years ago and I would have loved this. I absolutely loved Bartosz Milewski’s Category Theory for Programmer’s series. There is a crap-ton of line noise  that kind of muddies the waters though in javascript. I wonder if it makes sense to me because I mostly already understand what he’s going for from a Haskell context. He has some cool patterns here like using objects as typeclasses/interfaces.

https://jscategory.wordpress.com/

The really neat thing he discusses is higher order contracts which I’d bet is a central piece of contract based programming but I had never gotten that far.

I’m still skimming over the vids, but nice job.