kaashif's blog

Programming, with some mathematics on the side

Decomposing representations: a slice of computational group theory

2020-04-19

For my Master's degree, I (helped greatly by my supervisor) implemented some algorithms and even invented some new algorithms to decompose representations of finite groups. I wrote an extremely long (well, relative to other things I've written) and technical thesis about this, but I find myself increasingly unable to understand what any of it means or why I even have a degree.

I thought being forced into a short-form blog post would help me remember whatever it is I spent a few years studying to do. There are some foundational questions:

• What is a group?
• What is a representation?
• What is a decomposition of a representation?

And some more interesting questions, involving some computational tricks relevant to a wider audience:

• Why is this useful?
• How do you get a computer to do it?
• How do you get a computer to do it, quickly?

These are the questions I'll attempt to answer in this blog post. It'll be fun!

What is a group?

Any undergraduate mathematician should know this, but if you don't do algebra and need a refresher, read the Wikipedia page). For non-mathematicians, a group is:

• A set $G$ (for example, the real numbers or the rational numbers)

• With a binary operation $\ast$ (like addition or multiplication) which takes two elements of $G$ and gives you another element of $G$

• Such that $\ast$ is associative, meaning bracketing doesn't matter - $(a \ast b) \ast c = a \ast (b \ast c)$. True for multiplication or addition, but NOT for subtraction

• Such that $G$ has an identity element with respect to $\ast$ - an element $e$ such that applying it with $\ast$ does nothing, like multiplying by 1 or adding 0

• Such that elements of $G$ all have inverses with respect to $\ast$, meaning any element $g \in G$ has another element $g^{-1} \in G$ such that combining them with $\ast$ gives the identity. For addition, the inverse of $x$ is $-x$. For multiplication, it's $\frac1{x}$.

Examples of groups include the real numbers $\mathbb{R}$ with addition and the non-zero rational numbers $\mathbb{Q}^\ast$ with multiplication.

There are also finite examples, like the set ${+1, -1}$ with multiplication, and the set of permutations of $n$ elements with composition - $S_n$, the symmetric group on $n$ elements.

What is a representation?

A representation $\rho : G \to \text{GL}(V)$ of a group $G$ is a homomorphism (a function respecting the group structure) from $G$ to the group of linear automorphisms of a vector space $V$. You can imagine "vector space" to mean $\mathbb{C}^n$ (the space of vectors with $n$ entries in the complex numbers), since that's usually what we're using.

"Linear automorphism" is a coordinate-independent way of saying "invertible matrix". So you can rephrase the definition of a representation as a map from a group $G$ to $n \times n$ matrices (the $n$ is fixed for all elements), such that the group structure is respected.

Sue me if you don't like it, but you'll never get your vector space into a computer without picking a basis, meaning you are forced to think of linear maps as matrices.

What is a decomposition of a representation?

An irreducible representation is a representation that doesn't have any subrepresentations. What does that mean? We could go through the definitions, but for complex representations of finite groups, irreducible means indecomposable: you can't write the representation as a direct sum of other representations. That is the same thing as saying you can't simultaneously block diagonalise all $\rho(g)$ for $g \in G$ such that the blocks are smaller than the whole matrix.

Now that the basics are out of the way, we can ask an interesting question:

Why is this useful?

Irreducible representations come up in all sorts of places:

• Solving problems in quantum chemistry involving symmetries (oh, hello $S_n$)

• Solving completely computationally intractable optimization problems by spotting interesting symmetries ($S_n$ again?), then reducing them

• Turning groups into linear algebra, then studying them through how their representations behave (this is just a vague statement encompassing all of representation theory)

Anything with symmetries probably has something to do with groups and thus, something to do with representations.

I keep mentioning $S_n$, but the people who paid attention in algebra might note that it is kind of misleading to talk about it as if it is special. In fact, all finite groups appear as a subgroup of some symmetric group due to Cayley's theorem. The proof is actually fairly simple (for such a deep-seeming result) and I encourage you to look at it. So really, saying $S_n$ pops up everywhere in group theory is the same thing as saying groups pop up everywhere in group theory - not surprising.

Sadly, physicists are interested in infinite groups and in particular, Lie groups. You might have heard of the Lorentz group $\text{O}(1,3)$, a certain special unitary group $\text{SU}(2)$, or some other Lie groups - these pop up all the time in physics.

I say "sadly" because my algorithms are not directly applicable to those cases, since they're too continuous to fit inside my poor computer. It's very happy that there is such strong motivation to study Lie groups coming from physics, but my work isn't really relevant there.

How do you get a computer to compute a decomposition?

The first question to ask is: how do you do it by hand?

Let's say we have a finite group $G$ and a representation $\rho : G \to \text{GL}(V)$ where $V$ is a finite dimensional complex vector space (I'm skipping over something very important here, see if you can guess what it is, or skip to the end for the answer).

If you already have the complete list of irreducible representations of $G$ (up to isomorphism), $(\rho_i)$, you can just combine all $\rho_i$ as blocks in all possible ways that give matrices of the correct size. That is, you try to find some $i_j$ such that:

$$\rho(g) = \begin{pmatrix} \rho_{i_1}(g) & & \\ & \ddots & \\ & & \rho_{i_m}(g) \end{pmatrix}$$

Where anything not labelled above is zero. This is easy for some small groups. For example, if you have the representation of ${\pm 1}$ given by:

$$\pm 1 \mapsto \begin{pmatrix} 1 & 0 \\ 0 & \pm 1 \end{pmatrix}$$

It's easy to spot this is given by two $1 \times 1$ blocks. But that's because we already know all of the irreducible representations (let's shorten that to irrep, as is standard), there are only two. Even better, the matrices are already given in the nicest basis possible. It's not really obvious how to "observe" the structure of a representation like this in general.

Here's the dirty trick. Define $\chi_\rho(g) := \text{trace}(\rho(g))$. $\chi_\rho$ is called the character of $\rho$. It has the useful property that for a finite $G$, a finite dimensional representation (meaning the matrices are finite size), and coefficients in $\mathbb{C}$: $\rho$ is isomorphic to $\tau$ (as representations) if and only if their characters are the same.

This gives rise to this simple algorithm:

1. List all irreps of $G$ using e.g. Dixon's algorithm (IrreducibleRepresentationsDixon in GAP)

2. Put them together in all possible ways adding up to the correct matrix dimension.

3. Compute all of the characters - there will be exactly one matching the character of $\rho$, that's your decomposition!

This involves a lot of listing and trial and error - it is highly inefficient and intractable for even small groups and small representations.

Even ignoring efficiency, this tells you the decomposition up to isomorphism. Telling you the basis change from $\rho$ to actually get the smallest possible blocks is another problem. There's one algorithm due to Serre to do that, with a full proof of correctness in his textbook on Linear Representations of Finite Groups.

How do you get a computer to compute it, quickly?

I (and my supervisor) came up with a bag of tricks to speed up the computation by taking advantage of some extra information that's usually available.

$G$ is always isomorphic to a permutation group, but in the real world, $G$ is usually actually already a permutation group. This lets us specialise our algorithms using the vast array of tools available for computing with permutation groups. The main trick is that a good base and strong generating set) can be computed with the Schreier-Sims algorithm, letting us sum over $G$ quickly (computing $\sum_g f(g)$ for certain $f$). This lets us compute some of Serre's formulas much, much faster.

Let's say we have two representations $\rho$ and $\tau$ as earlier, which are isomorphic but we don't know the basis change (the isomorphism). $G$ is a permutation group, so summing over it is fast. But how can we reduce finding an isomorphism to a sum over $G$? It's quite a neat trick.

We're looking for a matrix $A$ such that $A^{-1} \tau(g) A = \rho(g)$ for all $g \in G$. the "dumb" way would be to forget the representation structure and just try to simultaneously diagonalise the $\tau(g)$ and $\rho(g)$. I spent a whole six months hammering away before noticing the smart way, which is to notice that there's another representation $\alpha$ defined by:

$$g \mapsto (X \mapsto \tau(g)X\rho(g^{-1}))$$

If we say that $\rho$ and $\tau$ map to $\text{GL}(V)$ then this is a representation $G \to \text{GL}(\text{GL}(V))$. Crazy! The funny thing is that this representation is actually $\tau \otimes \rho^\ast$, the Kronecker product of $\tau$ and the conjugate transpose of $\rho$ (tensor product and dual representation, if you're still coordinate-free). The key thing here is that $A$ is a suitable basis change matrix if and only if $\alpha(g)A = A$ for all $g \in G$.

Even better, the projection from $\text{GL}(V)$ to the subspace of the suitable $A$ is given by the map (a specialisation of a more general theorem from Serre):

$$p = \sum_g \alpha(g)$$

This is where the key speed gain is: we can use the summation trick from before. The sacrifice is that the matrices $\alpha(g)$ are absolutely huge. If $n$ is the dimension of $V$, this uses $\text{O}(n4)$ space. Terrible unless you come up with another trick to cut down the space usage.

Be lazy where possible to get the space back

We don't actually care what the matrices $\alpha(g)$ are, we just need to know how to apply them to compute the image of the $p$ above, then pick an invertible element from this image.

Even better, you can pick a random $B$ and $pB$ will almost always be invertible in a rigorous sense: the determinant map $B \mapsto \text{det}(pB)$ is polynomial, nonzero (we know $pB$ is invertible for some $B$), so its zero set is measure zero. In practice, you can just generate a random matrix and it will always work as a basis change matrix.

So in the end we don't need to ever compute the matrices $\tau \otimes \rho^\ast$ in full all at the same time, we can just compute:

$$A = pB = \sum_g (\tau(g) \otimes \rho^\ast(g))(B)$$

The tensor product has the neat property that $(A \otimes B)(C \otimes D) = AC \otimes BD$. We can write the matrix $B$ as $\sum_{i,j} B_{ij} e_i \otimes e_j$, that is we "vectorise" the matrix $B$ into an $n^2$ length vector. Then the equation above simplifies to:

$$A = \sum_g \sum_{i,j} B_{ij} \tau(g)e_i \otimes \rho^\ast(g)e_j$$

So we have ended up with a formula for $A$ (the basis change matrix) which doesn't require the computation of any matrices with $n^4$ entries, and is amenable to the fast group summing method vaguely mentioned earlier.

Since we still have to do matrix multiplications many times, this whole algorithm is still at least $\text{O}(n^3)$ but using clever tensor product tricks, we got it down from $\text{O}(n^4)$. Not bad!

One thing to mention is that the independence of the run time from $|G|$ frees us to compute with huge groups. My algorithm can operate on small degree representations of groups like $S_{10}$, with $10!$ elements (3.6 million elements).

Hopefully this has been some kind of insight into what computational representation theory might involve. Take a look at the source code for the gritty details.

What did you gloss over?

There's one huge thing I haven't explained. Have you guessed what it is?

You can't represent $\mathbb{C}$ in a computer! Floating point is useless for algebra! So how were we multiplying matrices and adding up coefficients in $\mathbb{C}$? There's another trick: restrict your attention to the cyclotomic numbers.

But why is that good enough? Is multiplication still constant time? Was it ever constant time?

I'll answer these questions at some point.