GRAPH ISOMORPHISM AND REPRESENTATION THEORY

DANIEL LITT

1. Introduction
How many isomorphism classes of (unlabeled) graphs are there with n vertices
and k edges? (Here a graph is a set V of vertices, and a set E of (unordered) pairs
of distinct elements of V , which we think of as edges between two vertices. In
particular, we do not allow loops or multiple edges between two vertices. We say
two graphs (V1 , E1 ), (V2 , E2 ) are isomorphic if there is a bijection V1 ' V2 sending
E1 bijectively to E2 .)
For example, the following two graphs with 8 vertices and 12 edges are isomorphic.

Below we record
the number of graphs with n vertices, as the number of edges
range from 0 to n2 :

n # of graphs with n vertices and 0, 1, 2, · · · edges

1 1
2 1, 1
3 1, 1, 1, 1
4 1, 1, 2, 3, 2, 1, 1
5 1, 1, 2, 4, 6, 6, 6, 4, 2, 1, 1
6 1, 1, 2, 5, 9, 15, 21, 24, 24, 21, 15, 9, 5, 2, 1, 1
7 1, 1, 2, 5, 10, 21, 41, 65, 97, 131, 148, 148, 131, 97, 65, 41, 21, 10, 5, 2, 1, 1
Let gn,k be the number of isomorphism classes of (unlabeled) graphs with n
vertices and k edges.
Observe: each row of the table above is
(1) Symmetric (i.e. gn,k = gn,(n)−k for all k, n), and
2
(2) Unimodal (i.e. gn,k increases, and then decreases).

Proof of (1). There is a bijection between (unlabeled) graphs with k edges and
graphs with n2 − k edges, given by sending a graph to its complement.




The rest of this talk will be devoted to a proof of (2), that is, that the sequence
{gn,k } for fixed n is unimodal. The idea of the proof is to turn this into a problem
of linear algebra — in particular, we will use the representation theory of the Lie
algebra sl2 .
1
2 DANIEL LITT

2. The Lie algebra sl2 (C)

As a vector space, we define
sl2 (C) = {A ∈ Mat2×2 (C) | Tr(A) = 0}.
That is, sl2 (C) consists of the set of 2 × 2 traceless matrices. The product of two
traceless matrices need not be traceless, so this space does not have a natural
product. That said, the commutator of two traceless matrices
[A, B] = AB − BA
is traceless. We view sl2 (C) as the above vector space, with this commutator as a
binary operation. The commutator (which we call the Lie bracket from now on) is
bilinear and satisfies
(1) Anti-commutativity:
[A, B] = −[B, A],
(2) The Jacobi identity:
[A, [B, C]] + [C, [A, B]] + [B, [C, A]] = 0.
A vector space V with a bilinear map
[−, −] : V × V → V
satisfying the above properties is called a Lie algebra.
Let us describe sl2 (C) in terms of a basis. We set
     
0 1 0 0 1 0
e= , f= , h= .
0 0 1 0 0 −1
The set {e, f, h} is evidently a basis for sl2 (C). A short computation gives
[e, f ] = h, [h, f ] = −2f, [h, e] = 2e,
which determines the Lie bracket in general by bilinearity and anti-commutativity.
Definition 2.1. A representation of a Lie algebra g is a linear map
ρ : g → gln (C) := Matn×n (C)
such that
ρ([A, B]) = [ρ(A), ρ(B)]
for all A, B.
In particular, an n-dimensional representation of sl2 (C) is a set of three n × n
matrices E, F, H such that
[E, F ] = H, [H, F ] = −2F, [H, E] = 2E.

3. Representation theory of sl2 (C)

We now describe explicitly the finite-dimensional representations of sl2 (C). Let V
be a finite-dimensional C-vector space and E, F, H : V → V linear maps satisfying
the commutator relations above.
Theorem 3.1. E, F are nilpotent. If v is an eigenvector of H with eigenvalue λ,
then H(Ev) = (λ + 2)Ev; if F v is non-zero, H(F v) = (λ − 2)F v.
 GRAPH ISOMORPHISM AND REPRESENTATION THEORY 3

Proof. Let v be an eigenvector of H with eigenvalue λ, i.e.

Hv = λv.
Then
HEv = 2Ev + EHv = (λ + 2)Ev
and
HF v = F Hv − 2F v = (λ − 2)F v
as desired. So λ+2, λ−2 are both eigenvalues of H, and F v, Ev are both eigenvectors
(if they are non-zero). Hence if E is not nilpotent, we have that λ+2n is an eigenvalue
for all positive n; likewise if F is not nilpotent, λ − 2n is an eigenvalue for all positive
n. But H has only finitely many eigenvalues (as V is finite-dimensional), so E, F
act nilpotently on any eigenvector of H, and hence on the subspace V ss spanned by
the eigenvectors of H. By the computations above, E, F preserve this subspace, so
we may replace V by V /V ss , whence we are done by induction on the dimension of
V. 

Theorem 3.2. Let λ be an eigenvalue of H with maximal real part, and let v
be a λ-eigenvector. Let N be minimal such that F N v = 0. Then Ev = 0, and
Span(v, F v, F 2 v, · · · , F N −1 v) is a subrepresentation of V . Moreover λ = N − 1.
Proof. Ev = 0 as if not, λ + 2 would be an eigenvalue of H, contradicting the
maximality of the real part of λ. Then
λv = Hv = [E, F ]v = EF v − F Ev = EF v,

(λ − 2)F v = HF v = [E, F ]F v = EF 2 v − F EF v
and hence
EF 2 v = (λ − 2)F v + F λv = (2λ − 2)F v
and in general
(λ − 2(n − 1))F n−1 v = HF n−1 v = [E, F ]F n−1 v = EF n v − F EF n−1 v
and so
EF n v = (λ − 2(n − 1))F n−1 v + F EF n−1 v = (−n2 + (λ + 1)n)F n−1 v
by induction on n (exercise). Thus E sends Span(F n v) to Span(F n−1 (v)), F sends
Span(F n−1 v) to Span(F n v), and H preserves Span(F n v), so the vector space in
question is a subrepresentation as desired.
Now take n = N , so that F N v = 0 and F N −1 6= 0. Then we have
0 = EF N v = (−N 2 + (λ + 1)N )F N −1 v
and hence
−N 2 + (λ + 1)N = 0.
Thus
λ=N −1
as desired. 
4 DANIEL LITT

In fact the proof gives an explicit description of all the irreducible finite-dimensional
representations of sl2 (C). Given any irreducible representation V , we may take λ
with maximal real part as above — then the subrepresentation given by the theorem
must be all of V by irreducibility.
Indeed, the proof shows that there is a unique irreducible sl2 (C)-representation
VN of dimension N , which we may describe explicitly as follows. Namely VN has a
basis {v0 , v1 , · · · , vN −1 } so that
F vn = vn+1 for n < N − 1 and F vN −1 = 0
Hvn = (N − 1 − 2n)vn
Evn = (−n2 + (λ + 1)n)vn−1 .
(In fact, all finite-dimensional representations of sl2 (C) are semisimple, i.e. they
split up into direct sums of copies of the VN above, so this gives a description of all
finite-dimensional representations of sl2 (C). But we will not use this fact.)
Remark 3.3. One may give a more “coordinate-free” description of the irreducible
representations. Namely VN may be viewed as the space of homogeneous polynomials
∂ ∂
in two variables X, Y of degree N − 1. Then E acts via X ∂Y , F via Y ∂X , and H
∂ ∂
via X ∂X − Y ∂Y .
We record the key fact that these results give us about sl2 (C)-representations,
for future use:
Theorem 3.4. Let V be a finite-dimensional sl2 (C)-representation. Let dk be the
dimension of the generalized k-eigenspace of H acting on V . Then the sequences
{dk }k odd , {dk }k even

are both unimodal and symmetric about 0 (i.e. dk = d−k for all k).
Proof. The statement is true for irreducible representations by our classification,
and is preserved under extensions (which have the effect of summing the sequences
in question). Thus it is true for arbitrary representations. 

4. Back to graph theory

We now have the tools to return to our original problem. Recall that gn,k was
the number of unlabeled graphs with n vertices and k edges; we wish to show that
for fixed n, the sequence gn,k is unimodal. The idea will to be to construct an
sl2 -representation such that the gn,k appear as dimensions of H-eigenspaces as in
Theorem 3.4.
We fix n, and let Wn be the vector space on the set of labelled graphs on the
vertices {1, · · · , n}. That is, if Gn is the set of labelled graphs on the vertices
{1, · · · , N },
M
Wn = Cg.
g∈Gn
The symmetric group Sn acts on Gn by permuting the labels of a labelled graph,
and hence acts on Wn . We let Wn,k ⊂ Wn be the subspace spanned by g ∈ Gn such
that g has k edges. Observe that Wn,k is a subrepresentation.
Proposition 4.1.
Sn
gn,k = dim Wn,k .
 GRAPH ISOMORPHISM AND REPRESENTATION THEORY 5

Sn
Proof. A basis for Wn,k is given by elements of the form
X
σ(g)
σ∈Sn

where g is a labelled graph with n vertices and k edges. It is clear that this basis is
in bijection with isomorphism classes of graphs with k edges. 

We now define an action of sl2 (C) on Wn . Let ai,j : W → W be the operator

(
g ∪ (i, j) if (i, j) 6∈ g
ai,j : g 7→
0 otherwise
i.e. ai,j adds an edge to g between vertices i and j if there isn’t one there already,
and sends g to 0 otherwise. Let bi,j be the operator
(
g \ (i, j) if (i, j) ∈ g
bi,j : g 7→
0 otherwise
i.e. bi,j removes the edge between i and j if such an edge exists, and sends g to 0
otherwise. Note that if {s, t} =
6 {u, v},
[as,t , bu,v ] = 0,
and (
g if (s, t) is an edge in g
[as,t , bs,t ]g =
−g otherwise
We set X
E= ai,j
i<j

and X
F = bi,j .
i<j

Then for g a labelled graph with k edges,

" #
X X
[E, F ](g) = as,t , bu,v g
s<t u<v
X
= [as,t , bu,v ]g
s<t,u<v
X
= [as,t , bs,t ]g
s<t
X X
= g− g
(s,t)∈g (s,t)6∈g
  
n
= 2k − g.
2
Thus we set Hk : Wn,k → Wn,k equal to the operator
  
n
Hk : g 7→ 2k − g,
2
6 DANIEL LITT

and M
H= Hk .
k
We have already verified that [E, F ] = H. Let us verify the other two commutation
relations for sl2 (C). Indeed, we have that for g a graph with k edges,
     
n n
[H, F ]g = 2k − 2 − F g − F 2k − g
2 2
= −2F g
and
     
n n
[H, E]g = 2k + 2 − Eg − E 2k − g
2 2
= 2Eg
as desired.
Observe that the action of sl2 (C) on Wn commutes with the action of Sn , and so
we get an action on M
Sn
WnSn = Wn,k .
k
We are now ready to prove the theorem.
Theorem 4.2. The sequence {gn,k } (for fixed n and varying k) is unimodal.
Sn
Proof. We already know that gn,k = dim Wn,k . But we have already constructed
an action of sl2 (C) on M
Sn
WnSn = Wn,k
k
such that H acts
diagonalizably, and the eigenspace of H corresponding to the
Sn
eigenvalue 2k − n2 is precisely Wn,k . Thus we are done by Theorem 3.4, as the
numbers 2k − 2 all have the same parity (namely, the parity of n2 ).
n




Remark 4.3. Note that the same argument proves more — namely that for every
irreducible representation χ of Sn , the sequence
χ
dim Wn,k
is unimodal.
Remark 4.4. Note that this gives a fundamentally different proof that gn,k =
gn,(n)−k .
2

Exercise 4.5. Let Pkl (n) be the number of ways of partitioning n into at most k
pieces, each of which has size at most l. Show that for fixed k and l, the sequence
Pkl (n)
is unimodal.
Question 4.6 (For experts only). In geometry, if X is a smooth projective variety,
there are natural sl2 -representations on the vector space
H ∗ (X, C),
associated to ample classes on X. This is the hard Lefschetz theorem. A natural
question is — is there a natural variety whose cohomology is of the form WnSn ?
 GRAPH ISOMORPHISM AND REPRESENTATION THEORY 7

Even better, is there a natural variety with an Sn -action whose cohomology is W ,

viewed as an Sn × sl2 -representation?