Professional Documents
Culture Documents
Artigo - Matemática Pura
Artigo - Matemática Pura
Abstract. The moment problem in probability theory asks for criteria for when there ex-
ists a unique measure with a given tuple of moments. We study a variant of this problem for
arXiv:2210.06279v1 [math.PR] 12 Oct 2022
random objects in a category, where a moment is given by the average number of epimor-
phisms to a fixed object. When the moments do not grow too fast, we give a necessary and
sufficient condition for existence of a distribution with those moments, show that a unique
such measure exists, give formulas for the measure in terms of the moments, and prove that
measures with those limiting moments approach that particular measure. Our result applies
to categories satisfying some finiteness conditions and a condition that gives an analog of
the second isomorphism theorem, including the categories of finite groups, finite modules,
finite rings, as well as many variations of these categories. This work is motivated by the
non-abelian Cohen-Lenstra-Martinet program in number theory, which aims to calculate
the distribution of random profinite groups arising as Galois groups of maximal unramified
extensions of random number fields.
1. Introduction
The kth moment of a measure ν on the real numbers is defined to be X X k dν. The
R
classical moment problem asks whether, given a sequence Mk , whether there exists a measure
whose kth moment is Mk for all k, and whether that measure is unique. For the existence
part, a natural necessary and sufficient condition is known, whereas for uniqueness, multiple
convenient sufficient conditions are known [Sch17]. An important auxiliary result allows us
to apply the uniqueness result to determine the limit of a sequence of distributions from the
limit of their moments.
There have been recent works, in number theory, probability, combinatorics, and topol-
ogy that have used analogs of the moment problem to help identify distributions, but not
distributions of numbers–rather distributions of algebraic objects such as groups [EVW16,
Woo17, Woo19, Més20, NW21] or modules [LWZ19]. Even more recent works require dis-
tributions on more complicated categories, such as groups with an action of a fixed group
[BW17, Saw20], groups with certain kinds of pairings [LST20], groups with other additional
data [Woo18, Liu22], or tuples of groups [Lee22, NVP22]. The distributions that have been
identified this way include distributions of class groups of function fields, profinite comple-
tions of fundamental groups of 3-manifolds, sandpile groups of random graphs, and cokernels
of random matrices. These works have often required new proofs of moment problem analogs
in each different category considered.
The goal of this paper is to prove a theorem on the moment problem for distributions
of objects of a rather general category. We expect that future works of many authors will
consider distributions on new categories and we have tried to make our result as general as
possible so that it can be applied in these new settings. In a subsequent paper, we ourselves
will consider profinite groups G together with an action of a fixed group Γ and a class in
the group homology H 3 (G, Z/n), and will apply the main results of the current paper to
1
understand the distributions of class groups and fundamental groups of curves over finite
fields, and their influence under roots of unity in the base field.
In the categories in which we work, we give a criterion for there to exist a measure on
isomorphism classes of that category with a given sequence of moments, and a criterion for
the measure to be unique. We also give explicit formulas for the measure, when it exists, in
terms of the moments. This is a significant advance, as previous work has always relied on a
priori knowing a distribution with certain moments. Being able to construct a distribution
from its moments opens new avenues such as in our subsequent paper mentioned above.
Working at a high level of generality requires us to use some notation we expect to be
unfamiliar. Thus, for clarity, we begin with the statement of our main theorem in the
category of groups. In many cases of interest, the probability distributions of interest are on
the set of profinite groups (which includes finite groups), so we work there.
Let C be the category of finite groups. A level C is a finitely generated formation of
finite groups (a formation is a set of isomorphism classes of groups closed under taking
quotients and finite subdirect products). For example, the group Z/pZ generates the level
of elementary abelian p-groups. We consider the set P of isomorphism classes of profinite
groups that have finitely many (continuous) homomorphisms to any finite group. A group
X ∈ P has a pro-C completion X C which is the inverse limit of the (continuous) quotients
of X in C. We consider a topology generated on P generated by {X | X C ≃ F } as C and F
vary (and the associated Borel σ-algebra). For a finite group G, the G-moment of a measure
ν on P is defined to be Z
| Sur(X, G)|dν.
X∈P
Theorem 1.2. For each finite group G, let MG be a real number such that the sequence MG
is well-behaved.
(Existence): There is a measure on P with G-moment MG for each finite group G if and only if
the vC,F defined above are non-negative for each C and F .
(Uniqueness): When such a measure ν exists, it is unique and is given by the formulas
ν({X | X C ≃ F }) = vC,F .
(Robustness): If ν t are measures on P such that for each finite group G,
Z
lim | Sur(X, G)|dν t = MG ,
t→∞ X∈P
is absolutely convergent. (In Section 3 we will give formulas for µ(F, G) and Z(π) that are
easy to compute in practice.) We say (MG )G is well-behaved if it is well-behaved at level
C for all C. We will provide some explanation in Remark 1.9, after the statements of the
theorems in which the notion of well-behaved is used, to motivate why this is the rough
shape of what one needs for moments “not growing too fast”.
We define the level C quotient GC of G to be the maximal quotient of G that is in C, i.e.
the join of all level C quotients of G. By Lemma 2.8, we have that C is countable, and we
use the discrete σ-algebra for any measure on C.
Then our main theorem, below, says that for well-behaved moments, a distributions exists
with those moments if and only if certain formulas in the moments are positive, and when
a distribution exists with those moments, it is given by those certain formulas. Moreover,
this uniqueness is robust even through taking limits. We write Epi(F, G) for the set of
epimorphisms from F to G. If ν is a distribution on C/≃ , then G-moment of ν is defined to
be
Z
(1.4) | Epi(F, G)|dν.
F ∈C/≃
We also define
X µ̂(F, G)
(1.5) vC,F := MG .
G∈C
| Aut(G)|
Theorem 1.6 (Distributions on C from moments). Let C be a diamond category and let C
be a level. For each G ∈ C, let MG be a real number so that (MG )G is well-behaved at C.
(Existence): A measure νC on C with G-moment MG for all G ∈ C exists if and only if vC,F ≥ 0
for all F ∈ C.
(Uniqueness): If such a measure exists, then it is given by νC ({F }) = vC,F .
(Robustness): If νCt are measures on C such that for each G ∈ C,
Z
lim | Epi(F, G)|dνCt = MG ,
t→∞ F ∈C
4
then limt→∞ νCt ({F }) = vC,F .
1.2. The main result in the profinite case. If we use MG for each G ∈ C/≃ to define
measures νC as above for each level C, then by the uniqueness in Theorem 1.6, for levels
C ⊂ C ′ and the functor C ′ → C that sends X to X C , the measure νC ′ pushes forward to νC .
This compatible system of measures for each level gives rise to a measure ν. However, in
most of our main examples of interest, this measure ν is not on the category C, but rather
on pro-objects of C. We define a pro-isomorphism class X of C, to be, for each level C, an
isomorphism class X C ∈ C such that whenever C ⊂ C ′ we have that (X C )C ∼
′
= X C . (Here
′ ′ ′
X C ∈ C, so (X C )C takes the level C quotient of X C .)
The reader might wonder how this definition compares to the usual notion of a pro-
object. We will show in Lemma 5.7 that pro-isomorphism classes are naturally in bijection
with isomorphism classes of pro-objects satisfying an additional condition we call small
(Definition 5.4), which in the case where C is the category of finite groups, so that pro-
objects are profinite groups, restricts to the usual notion of small profinite groups (i.e groups
with finitely many open subgroups of index n for each n). When R is a Noetherian local
commutative ring with finite residue field, and R̂ its completion at its maximal ideal, then
pro-isomorphism classes of finite R-modules are naturally in bijection with isomorphism
classes of finitely generated R̂-modules (Lemma 5.10).
Let P be the set of pro-isomorphism classes of C. We define a topology on P with a basis
given by the open sets UC,F = {X ∈ P | X C ∼ = F } for each level C and F ∈ C.
For X a pro-isomorphism class, G ∈ C, and C a level containing G, we have that
Epi(X C , G) is independent of the choice of C. (This is because if D is the level gener-
ated by G, any epimorphism X C → G factors through (X C )D ≃ X D , and so Epi(X C , G) ∼
=
Epi(X D , G).) We define |Epi(X, G)| = Epi(X C , G) for any such C.
We take a σ-algebra (according to which we consider any measures) on P to be generated
by the UC,F for each level C and F ∈ C.
Now, we give the consequence of Theorem 1.6 for measures on P.
Theorem 1.7. Let C be a diamond category. For each G ∈ C/≃ , let MG be a real number
so that (MG )G is well-behaved.
(Existence): There exists a measure ν on P with G-moment MG for all G ∈ C if and only if for
every level C and every F ∈ C, we have vC,F ≥ 0.
(Uniqueness): If such a measure ν exists, then it is unique, and for every level C and F ∈ C, we
have ν(UC,F ) = vC,F .
If C has countably many isomorphism classes, then our σ-algebra on P is equivalent to
the Borel σ-algebra of the topology on P generated by the UC,F (because then there are only
countably many pairs C, F , so arbitrary unions of finite intersections of them are the same
as countable unions).
Theorem 1.8. Let C be a diamond category with at most countably many isomorphism
classes. For each G ∈ C/≃ , let MG be a real number so that (MG )G is well-behaved. If ν t is
a sequence of measures on P such that for all G ∈ C
Z
(Robustness): lim |Epi(X, G)|ν t = MG ,
t→∞ X∈P
then a measure ν as in Theorem 1.7 exists and as t → ∞ the ν t weakly converge to ν.
5
Remark 1.9. We explain some of the motivation for the definition of well-behaved. In both
Theorem 1.6 and Theorem 1.7, the sum
X µ̂(F, G)
MG
G∈C
| Aut(G)|
appears as the formula for a probability. For this formula to be well-defined, the sum must be
convergent, and since there is no obvious ordering of the terms, the sum must be absolutely
convergent. The condition that the MG be well-behaved is a strictly stronger condition than
this, because it requires the absolute convergence of the sum with an additional Z(π)3 factor.
This is technical restriction required by our method to prove the existence theorems, but we
do not know if it is necessary for the statement of the theorems to hold. However, for the
applications we are considering, the Z(π)3 factor is not large enough to turn the sum from
convergent to divergent, and thus causes no serious difficulty.
1.3. Motivation from number theory and prior work. When C is the category of
Fq -vector spaces for some prime power q, then the moments in our sense are just a finite
upper triangular transformation of the classical moments of the size of the group, since
| Hom(G, Fkq )| = |G|k . However, the bounds on the classical moment problem have not been
sufficient for many authors studying distributions on C, and they have had to develop their
own moment theorems. This includes work on the distribution of Selmer groups of congruent
number elliptic curves [Hea94], ranks of (non-uniform) random matrices over finite fields
[Ale99, Coo00], and 4-ranks of class groups of quadratic fields [FK06a, FK06b].
Cohen and Lenstra [CL84] made conjectures on the distribution of class groups of quadratic
number fields that have been a major motivating and explanatory force for work on class
groups, including that on 4-ranks mentioned above. Ellenberg, Venkatesh, and Westerland
[EVW16] showed that the analog of Cohen and Lenstra’s conjectures for Sylow p-subgroups
of class groups of imaginary quadratic fields, when Q is replaced by Fq (t), holds as q → ∞.
To do this, they found the limiting moments of the distributions of interest and showed
uniqueness and robustness in the moment problem for the particular distribution on abelian
p-groups conjectured by Cohen and Lenstra. This work introduced the idea that considering
averages of # Sur(−, G) was a useful way to access distributions on finite abelian groups. (For
finite abelian groups, these averages are a finite upper triangular transformation of classical
mixed moments of a tuple of integral invariants that determine the group, see [CKL+ 15,
Section 3.3].)
This approach has been followed by many papers, seeking to show function field (q → ∞)
analogs of the conjectures of Cohen and Lenstra and their many generalizations. In [Woo18],
the second author shows such an analog for real quadratic function fields, which follows from
a more refined result that gives the distribution of pairs (Pic0 , d∞ ), where Pic0 is the group of
degree 0 divisors on the associated curve and d∞ is an element of Pic0 given by the difference
between the two points above ∞. The quotient Pic0 /d∞ gives the desired class group. This
work included a moment problem uniqueness theorem for the category of pairs (A, s) where
A is an abelian p-group and a ∈ A. In [BW17], Boston and the second author, proved
function field analogs of some conjectures of Boston, Bush, and Hajir [BBH17] on p-class
tower groups of imaginary quadratic fields, which generalize the conjectures of Cohen and
Lenstra to a non-abelian analog. This work included a moment problem uniqueness theorem
for the category of p-groups with a generator inverting automorphism.
6
Liu, Zureick-Brown, and the second author [LWZ19] found the moments of the distribu-
tions of the fundamental groups of Γ-covers of P1Fq split at infinity, or equivalently of the
Galois groups of the maximal unramified extensions of totally real Γ-extensions of Fq (t), as
q → ∞. For the part of these groups of order relatively prime to q − 1 (the order of the
roots of unity in Fq (t)) and q|Γ|, they were able to construct an explicit distribution on
profinite groups with an action of Γ that had these moments. In particular their distribution
abelianizes to the distribution of finite abelian Γ-groups (i.e. groups with an action of Γ)
conjectured by Cohen and Martinet [CM90] for the class groups of totally real Γ-number
fields, and thus one can apply the uniqueness and robustness results on the moment problem
for finite abelian Γ-groups of Wang the the second author [WW21] to obtain a q → ∞ analog
of the Cohen-Martinet conjectures. The first author [Saw20] then showed a uniqueness and
robustness result on the moment problem for (not necessarily abelian) Γ-groups, which then
gave a q → ∞ theorem that the prime-to-(q − 1)q|Γ|-part of the fundamental groups of
Γ-covers of P1Fq approached the distribution constructed in [LWZ19].
For the part of the groups that has order not relatively prime to the roots of unity in
the base field, even though moments can be found with the methods of [LWZ19], in general
there is no known distribution with those moments. For the case of quadratic extensions,
Lipnowski, Tsimerman, and the second author [LST20] showed a q → ∞ function field result
for the class groups of quadratic function fields, even at primes dividing the order of roots
of unity in the base field. Their work considered additional pairings on the class groups
(a structure they call a bilinearly enhanced group), and proved uniqueness and robustness
theorem on the moment problem for bilinearly enhanced abelian groups.
In combinatorics, work to find the distribution of Sylow-p subgroups of cokernels of random
integral matrices [Woo19], the distribution of the entire cokernel of such matrices [NW21],
the distribution of Sylow-p subgroups of sandpile groups of Erdős-Rényi random graphs
[Woo17], and the distribution of Sylow-p subgroups of sandpile groups of d-regular random
graphs [Més20] has all involved applying a uniqueness and robustness theorem for the moment
problem on finite abelian groups.
Several recent works, particularly in probability, have been concerned with the distribution
of tuples of finite abelian groups. Boreico [Bor16, Corollary 3.8.5], while studying cokernels
of random integral matrices with uniform, bounded coefficients, proves uniqueness and ro-
bustness results for r-tuples (V1 , . . . , Vr ) where Vi is an Fqid -vector space for some (fixed)
di ∈ N. Nguyen and Van Peski [NVP22, Theorem 9.1] prove a uniqueness and robustness
result for r-tuples of abelian groups in order to prove new universality results on the joint
distribution of cokernels of products of random matrices. Lee [Lee22, Theorem 1.3] has
also proven a uniqueness and robustness result for r-tuples of abelian groups, in order to
apply it to several universality results for joint distributions of cokernels of random matrices
perturbed in various ways. These uniqueness and robustness results for tuples would also
follow from the argument in [WW21, Proof of Theorem 6.11] (since (R1 × · · · × Rr )-modules
correspond to tuples (M1 , . . . , Mr ), where Mi is an Ri -module). Lee [Lee22, Theorem 1.10]
also proves a uniqueness and robustness result for r-tuples of finite groups with an action of
a fixed group Γ, with an application to the joint distribution of the quotient of free profinite
groups by Haar random relations and perturbations of those relations.
Given the ever growing number of different categories that such moment problem results
are required in, we have made every attempt here to give a result that will be applicable as
7
broadly as possible. All the moment problem results mentioned above are special cases of
the main theorems in this paper (except that, unlike the current paper, the moment problem
results of [Saw20] do not always require absolute convergence of the sums involved).
In all of the above discussed work, a known explicit distribution was given (or was found by
trying various distributions) with certain moments, and so uniqueness and robustness were
the only aspects of the moment problem that were necessary. Yet there remained settings
where moments are known, but no distribution with those moments was known. In one
setting like this, the authors [SW22b] found the distribution of profinite completions of 3-
manifold groups in the Dunfield-Thurston model of random Heegaard splittings, constructing
the distribution itself from the knowledge of the moments. This required working on the
moment problem in the category of pairs (G, h), where G is a finite group and h ∈ H 3 (G, Z)
and proving theorems on existence, uniqueness, and robustness.
The current paper makes completely general the approach used in [SW22b] on reconstruc-
tion of a distribution from its moments, via explicit formulas, without any a priori knowledge
of a distribution with those moments. One key motivation is our forthcoming work, which
will give a q → ∞ theorem on the distribution of the part of fundamental groups of Γ-covers
of P1Fq that is relatively prime to q|Γ| (but not necessarily q − 1). In particular, this will give
a correction to the conjectures of Cohen and Martinet at primes dividing the order of roots
of unity in the base field, which will be provided in a short forthcoming paper [SW22a]. We
will obtain the desired distributions using the results of the current paper.
Before this paper, many papers constructed distributions as limits of distributions given
by random matrix models or models of random generators and relations. In this setting, it
is usually very easy to compute the limiting moments of these distributions. It is somewhat
more challenging, but doable, to compute the limiting distributions. As the categories got
more complicated, showing that the limiting distribution had those limiting moments be-
came increasingly more difficult. Now Theorem 1.6 can be used to not only compute the
limiting distributions of such models in a straightforward way from their moments, but also
immediately obtain that the limiting distribution has moments given by the limiting mo-
ments. Moreover, if one’s only goal is to have explicit formulas for a distribution with certain
moments, one no longer has to guess possible models that may have the desired moments,
since Theorem 1.6 gives those formulas directly.
1.4. Strategy of proof. To understand the proof of Theorem 1.6, it is helpful to first
consider the special case where (MG )G is finitely supported, i.e. MG = 0 for G not iso-
morphic to any object in some finite list G1 , . . . , Gn . It follows that, if a measure ν
with moments MG exists, then ν({G}) = 0 unless G is isomorphic to one of G1 , . . . , Gn ,
since certainly MG ≥ ν({G}) as there is always at least one epimorphism from G to it-
self. Thus, in this case, the problem of existence and uniqueness of a measure can be
expressed as thePn existence and uniqueness of a solution to the finitely many linear equa-
tions MGi = j=1 |Epi(Gj , Gi )|ν({Gj }) in finitely many variables ν({Gj }), together with
the inequalities ν({Gj }) ≥ 0.
The linear equations can be expressed concisely in terms of a matrix A with Aij =
|Epi(Gj , Gi )|. The necessary observation to handle this case is that A is invertible. In
fact, it is possible to order G1 , . . . , Gn so that there are no epimorphisms from Gj to Gi if
j < i. This forces A to be upper-triangular, with nonzero diagonal, so an inverse always
8
exists. Thus the linear equations always have a unique solution, and a measure exists if and
only if that solution has ν({Gj }) ≥ 0 for all j.
A key observation that powers our proof is that one can explicitly calculate this inverse
µ̂(Gi ,Gj )
in great generality. In fact, the entry A−1 ij is given by the formula |Aut(Gj )| (Lemma 4.1).
Using this formula, we immediately obtain a guess vC,F for the unique measure with a
given moments (in Equation (1.4)). Proving that guess is correct, both that vC,F produces a
measure with those moments (Existence) and produces the only measure with those moments
(Uniqueness) requires only that we exchange the order of summation in a certain sum.
Using our assumption that the quotients of any object form a finite modular lattice, we note
that µ̂(Gi , Gj ) vanishes unless there exists a semisimple epimorphism from Gj to Gi . This
allows us to restrict attention to semisimple morphisms in sums involving µ̂. Furthermore, we
can calculate the exact value of µ̂ using known results on the structure of finite complemented
modular lattices. From this we can prove the sum relevant for (Uniqueness) is absolutely
convergent using the well-behavedness hypothesis, justifying the exchange of summation.
The sum relevant for (Existence), on the other hand, is trickier. It is not absolutely con-
vergent, under well-behavedness or even much stronger assumptions, in reasonable categories
like the category of finite groups. Instead we need a more intricate argument where we break
each sum into a series of multiple sums and exchange them one at a time, proving absolute
convergence only of certain restricted sums, which does follow from well-behavedness.
More precisely, we can place any level C into a series of levels C1 ⊂ C2 ⊂ · · · ⊂ Cn = C
where C1 consists (in the category of groups) only of the trivial group or (in general) only
of objects with no nontrivial quotients. We can express a sum over objects of level C as a
sum over objects H1 of level C1 , then over objects H2 of level C2 whose maximal level-C1
quotient is isomorphic to H1 , i.e. such that H2C1 ∼ = H1 , then over objects H3 of level C3 with
C2 ∼
H3 = H2 , and so on. Crucially, we can choose Ci to be “close together" in the sense that
for every H ∈ Ci , the natural epimorphism H → H Ci−1 is semisimple. This, again, lets us
restrict attention to semisimple morphisms, which are controlled by the well-behavedness
condition, when exchanging a given pair of sums.
The proof of (Robustness) requires, instead of an exchange of sums, the exchange of a
sum with a limit. We apply a similar strategy, breaking the sum into several smaller sums,
each over semisimple morphisms, and exchanging each sum with the limit in turn, using a
dominated convergence argument.
The approach of attacking problems in level C by an inductive argument using a sequence
of levels C1 , . . . , Cn , where each element of Ci has a semisimple morphism to an element of
Ci−1 , is something that we expect to be relevant to future problems about random groups and
other random objects. As a first example, for the problem of efficiently sampling from the
probability distributions we have constructed, we believe the best method will typically be
to choose a random object of level C1 , then a random lift to C2 , and so on, which reduces the
problem to a series of steps of sampling from a conditional probability distribution, which
will often be given by a simple probability distribution on an explicit set. For the exis-
tence problem, it is the combination of this inductive strategy with our Möbius calculations
which together allow us to restrict entirely to semisimple morphisms for the most difficult
calculations in the proof.
1.5. Organization of the paper. In Section 2, we show basic facts about modular lattices
and levels that will be used throughout the paper. In Section 3, we show how the Möbius
9
function and Z(π) can be given by simple formulas, and give examples in the categories of
groups, rings, and modules. In Section 4, we prove our main result, Theorem 1.6, which
operates at a single level. In fact, we prove slightly stronger results, in which we give the
minimal hypotheses separately for each piece of the theorem. In Section 5, we prove the
results regarding distributions on pro-isomorphism classes (Theorems 1.7 and 1.8), relate
pro-isomorphism classes to classical pro-objects, and give examples in groups and modules
to show what these are concretely. In Section 6, we give examples of the calculation of
measures from moments, including showing that moments of size O(|G|v ) are well-behaved
for any real v in the categories of groups or R-modules. Our examples give a blueprint for
how one evaluates, in practice, the sums defining vC,F and well-behavedness. We also show
many examples of diamond categories, including simple ones like (the opposites of) finite
sets and finite graphs, and how various modifications to diamond categories lead to new
diamond categories.
by the defining property of the Möbius function and another well-known property ([Sta12,
Proposition 3.7.1] for f the characteristic function of H).
Semisimple morphisms play a crucial role in the study of the Möbius function.
Lemma 2.6 (Philip Hall). Let G be an object of C and F a quotient of G. If G → F is not
semisimple, then µ(F, G) = 0.
Proof. This follows from [Rot64, Corollary p. 349] but we give the simple proof below.
Let Λ be the sublattice of quotients H of G such that G → H is semisimple. By Lemma 2.4,
we have that Λ is meet-closed, and thus for every quotient H of G, there exists H ss , the
minimal semisimple quotient of G that is ≥ H. Then we have
X X
0= µ(H, G) µ(F, K) (each interior sum vanishes by defn. of µ)
H∈Λ K∈[F,H]
X X
= µ(F, K) µ(H, G) (rearrange sum)
K∈[F,G] H∈Λ∩[K,G]
i.e., H∈[K ss ,G]
= µ(F, G),
12
where the last equality holds because the inner sum vanishes unless K ss = G by (2.5), and
K ss = G if and only if K = G (as if K < G then K is less than some simple quotient of
G).
Let π : G → F be an epimorphism. Let H be the join of all the minimal non-F elements
in the interval [F, G]. We call the epimorphism G → H the radical of π. Since H is the join
of the minimal non-F elements, H → F is semisimple, and it is the maximal semisimple
morphism through which π factors. We call H → F the semisimplification of π.
The finiteness in Definition 1.3 (1) ensures that we can write any epimorphism π as a
finite composition of simple morphisms. The modularity assumption means that, when we
do this, the number of simple morphisms depends only on π [Bir48, V, Theorem 3 on p. 68].
Definition 2.7. We define the dimension of π to be the number of simple morphisms it is a
composition of (or 0 if π is an isomorphism) and the dimension of G to be the dimension of
its morphism to its unique minimal quotient.
Note the dimension of an epimorphism G → H is the difference between the dimension of
G and the dimension of H. Also, if A < B, them dim A < dim B.
We will use many inductive arguments, which are all a form of induction on the dimension.
The first one is:
Lemma 2.8. For a level C, and F ∈ C, and d a natural number, there are finitely many
G ∈ C with an epimorphisms G → F of dimension d.
Proof. This follows by induction on d, with Definition 1.3 (3), and the induction step using
Definition 1.3 (3) and the fact that any epimorphism G → F of dimension d factors (in at
least one way) as G → G′ → F with G → G′ of dimension 1 (i.e. simple) and G′ → F of
dimension d − 1.
Proof. Every module K ∈ [N, M] defines a submodule of ker π, the kernel of the natural
morphism M → K. Conversely, every submodule S ⊆ ker π gives a module M/S ∈ [N, M].
This bijection sends inclusion of submodules to the ≥ relation on [N, M], so π is semisimple
if and only if the lattice of submodules of ker π is complemented. By Lemma 3.8, this occurs
only if ker π is a product of finite simple R-modules. The lattice [N, M] is dual to the lattice
of submodules of ker π, which by Lemma 3.7 has the stated description. The Möbius formula
then follows from Lemma 3.6
Lemma 3.9. Let C be the category of finite rings. Then a morphism π : R → S of finite
rings is semisimple if and only if ker π is a product of finite simple R-bi-modules.
In this case, there must exist finite simple R-bi-modulesQM1 , . . . , Mω , pairwise non-isomorphic,
and natural numbers n1 , . . . , nω ≥ 1, such that ker π ∼ ω
= i=1 Mini . Let qi be the order of the
field of endomorphisms of Mi . Then the lattice [S, R] is the product of i from 1 to ω of the
(ni )
lattice of subspaces of Fnqii . Thus µ(S, R) = ωi=1 (−1)ni qi 2 .
Q
In particular, if π is a local morphism of finite local commutative rings with residue field
κ, then π is semisimple if and only if ker π ∼ = κn for some n, and, then [S, R] is isomorphic
n
to the lattice of subspaces of κn , so µ(S, R) = (−1)n |κ|( 2 ) .
Proof. Any ring T ∈ [S, R] defines a sub-R-bi-module of ker π, its kernel, and conversely, any
sub-R-bi-module of ker π gives a two-sided ideal of R, whose quotient is a ring in [S, R]. This
bijection sends inclusion of submodules to the ≥ relation on [S, R], so π is semisimple if and
only if the lattice of sub-bi-modules of ker π is complemented. Since a bi-module is simply
an R ⊗ Rop -module, we can apply Lemma 3.7 to show that this happens if and only if ker π
is a product of simple bi-modules. The lattice [S, R] is dual to the lattice of submodules of
ker π, which by Lemma 3.7 has the stated description, and the Möbius formula then follows
from Lemma 3.6
For commutative rings, the left and right actions of R on R are identical, so the same is
true for ker π, so we can express ker π and every submodule of it as modules rather than
bimodules. For local commutative rings with residue field κ, the unique simple module is κ,
and so the expression ωi=1 Mini simplifies to κn , and similarly for the Möbius expression.
Q
17
Lemma 3.10. Let C be the category of finite groups. Then a morphism π : G → H of finite
groups is semisimple if and only if ker π is a product of finite simple G-groups.
In this case, there must exist pairwise non-isomorphic finite simple abelian G-groups,
V1 , . . . , Vn , and finite simple
∼ Qm G-groups N1 , . . . , Nm , as well as exponents e1 , . . . , en ≥
Qn non-abelian
ei
1, such that ker π = i=1 Vi × i=1 Ni . Let qi be the field of endomorphisms of Vi for i
from 1 to n. Then the lattice [H, G] is the product for i from 1 to n of the lattice of subspaces
of Feqii together with m copies of the two-element lattice.
(ei )
Thus µ(H, G) = (−1)m ni=1 (−1)ei qi 2 .
Q
(2) If C has the property that for any semisimple morphism G → H of level−C objects,
the number of simple factors in the lattice [H, G] is bounded by a bound depending
only on H and C, then for π : G → H with G of level C, we have
X
|µ(K, G)| ≪ |µ(H, G)|
K∈[H,G]
18
with the implicit constant depending only on H and C.
(3) If C has the property that for any semisimple morphism G → H of level−C objects,
the number of simple factors in the lattice [H, G] which are not isomorphic to the two-
element lattice is bounded by a bound depending only on H and C, then for π : G → H
with G of level C, we have
X
|µ(K, G)| ≪ |µ(H, G)|Z(π)
K∈[H,G]
is absolutely convergent, where either a = 3; C satisfies the hypotheses of Lemma 3.12 (2)
and a = 0; or C satisfies the hypotheses of Lemma 3.12 (3) and a = 1.
Proof. We handle case (1) first. Both sides depend only on the lattice [H, G] and are multi-
plicative under direct products of lattices. Since [H, G] is complemented modular, we may
reduce to checking each of the three forms of Theorem 3.5.
In the case of the lattice of subspaces of a n-dimensional vector space over Fq , the left side
is the sum over d from 0 to n of the number of (n − d)-dimensional subspaces of Fq times
d d
q (2) , and the right side is 23 q (2) by Lemmas 3.6 and 3.11. Since the number of subspaces of
dimension d is nd q , using the q-binomial theorem, the left side is
n n−1 ∞ ∞
n d n n n n
q (2 ) = q ( 2 ) (1 + q −j ) ≤ q ( 2 ) (1 + q −j ) ≤ q ( 2 ) (1 + 2−j ) < q ( 2 ) 8,
X Y Y Y
d=0
d q j=0 j=0 j=0
where b is number of simple factors of [H, G] with more than two elements. By assumption,
4b can be absorbed into our implicit constant.
The assumption of part (3) of Lemma 3.12 is satisfied for most examples arising in ordinary
mathematics, such as groups, rings, and modules. However, there are cases where it is not
satisfied. We give an example:
Example 3.13. Let C be the category whose objects are pairs consisting of a finite-dimensional
vector
L space V over F2 and a set S of nontrivial subsets of V such that the natural map
W ∈S W → V is an isomorphism, where morphisms (V1 , S1 ) → (V2 , S2 ) are given by maps
f : V1 → V2 such that f (W ) ∈ S2 ∪ {0} for all W ∈ S and if f (W ) = f (W ′) then W = W ′
or f (W ) = 0.
We can think of these as graded vector spaces, but instead of graded by integers, they are
graded by an arbitrary finite set.
It is not so hard to check a morphism is an epimorphism if and only if f is surjective, and
the poset of quotients of (V, S) is given by the product over W ∈ S of the poset of quotients
of W , and thus is a finite modular lattice. One can similarly check the other assumptions:
automorphism groups are subgroups of GLn (F2 ) and thus finite, and a simple morphism either
raises the dimension of one element of S by one or adds a new one-dimensional summand
in S, so there are finitely many simple morphisms.
By our description of the lattice of quotients, we say every object is a meet of its one-
dimensional quotients, and every one-dimensional object of C is isomorphic to (F2 , {F2}),
so every object is level-C for C = {(F2 , {F2 })}. Similarly, this meet property implies every
morphism is semisimple.
Thus the assumption of part (3) of Lemma 3.12 is not satisfied even for H = (0, ∅),
since [G, H] can be a product of an arbitrary finite number of lattices of quotients of finite-
dimensional vector spaces over F2 .
20
Lemma 4.1. Let G, H ∈ C, then
(
X X 1 if G = H
TF,H SG,F = TG,F SF,H =
F ∈I F ∈I
0 6 H.
if G =
Proof. The quotients of G that are (in C) isomorphic to F correspond exactly to the Aut(F )
orbits on Epi(G, F ) (which all have size |Aut(F )|). Thus, given G, H ∈ C, we have an
epimorphism
{F ∈ I, πFG ∈ Epi(G, F ), ρ ∈ Epi(F, H)} → {πH
G
∈ Epi(G, H), F ∈ [H, G]}
(F, πFG , ρ) 7→ (ρπFG , (F, πFG )),
where the fiber of (ρπFG , (F, πFG )) has size | Aut(F )|. Thus,
X X X 1 1 X
SG,F TF,H = µ(ρ)
F ∈I F ∈I G
| Aut(F )| |Aut(H)|
πF ∈Epi(G,F ) ρ∈Epi(F,H)
X 1 X
= µ(H, F )
G ∈Epi(G,H)
| Aut(H)|
πH F ∈[H,G]
(
1 if G = H
=
0 if G 6= H,
where the last equality is by (2.5). The second statement follows similarly, using the other
part of (2.5).
We are now ready to prove (Uniqueness) with a slightly weaker hypothesis, which we
restate here.
Theorem 4.2. Let C be a diamond category, C be a level of C, and (MG )G ∈ RI be so that
for all F ∈ C, we have
X X
(4.3) |µ(F, G)mG | < ∞.
G∈C π∈Epi(G,F )
Suppose for all F ∈ C we have values pC,F ≥ 0, such that for all G ∈ C we have
X
SF,G pC,F = mG
F ∈C
Then, for all F ∈ C, X
pC,F = TG,F mG
G∈C
where the first equality is by assumption, the last is from Lemma 4.1 (using that anything
with an epimorphism from H is in C). To verify that we can exchange
P the order
P of summa-
tion, as in the middle equality, it suffices to check that the sum G∈C TG,FP H∈C SH,G pH,C
is absolutely convergent. Since SH,G pH,C ≥ 0, it suffices to show that G∈C TG,F mG is
absolutely convergent, which is precisely our assumption.
21
Note that (Uniqueness) implies the “only if” direction of (Existence). Similarly, we could
prove the “if” direction of (Existence), conditional on absolute convergence of the below
sums,
!
X X X X
(4.4) SF,H TG,F mG = mG TG,F SF,H = mH .
F ∈C G∈C G∈C F ∈C
Unfortunately, in most cases of interest the absolute convergence of the above sums will
not hold (see Example 4.5). This requires us to make more delicate arguments to exchange
the order of summation one piece at a time, relying on further identities satisfied by the TG,F
as well as bounds on their size.
Example 4.5. One of the nicest examples of moments are that of the distribution µ on
profinite groups from [LW20], which has G-moment MG = 1 for every finite group G. (We
will see in Lemma 6.12 that these are well-behaved.) The category C here is the category of
finite groups. However, even in this case, the sum
XX XX 1
TG,F mG = TG,F
F ∈C G∈C F ∈C G∈C
| Aut(G)|
does not converge absolutely when C is generated by {F3 , S3 } (we write F2 , F3 for the cyclic
groups of order 2,3, respectively). (This sum is the left side of (4.4) in the special case
H = 1.) Indeed, to check that
XX
|TG,F || Aut(G)|−1 = ∞,
F ∈C G∈C
We will see that, summed over the structures on H as a quotient of G (i.e. over Epi(G, H)/ Aut(H)),
this final sum is precisely the sum in the lemma.
There is an | Aut(F )| − 1 correspondence between pairs (F, π) of F ∈ I and π ∈ Epi(G, F )
and F ≤ G. Given π and φ, when F ≥ E then there is one choice of ρ ∈ Epi(F, E) such that
ρπ = φ, and otherwise there is no possible ρ such that ρπ = φ. Finally, F C = H in the sum
above means that F C , as a quotient of G, is isomorphic to H, and so obtain the sum over all
F ∈ I with F C isomorphic to H in C, as in the lemma, we need to sum over all G-quotient
structures on H up to isomorphism.
Now we will see that well-behaved moments will provide absolute convergence on more
limited sums.
Proposition 4.8. Let C be a level generated by a finite set B, and let B̄ ⊂ B be a subset
containing every proper quotient of an element of B, and C¯ the level generated by B̄. Let
E ∈ C . and H ∈ C.¯ If (MG )G is behaved at the level generated by C and E, then
X X X X
Tπ mG < ∞.
G∈I F ∈I π∈Epi(G,F ) ρ∈Epi(F,E)
F C̄ ≃H ρπ rel. level C
In the second inequality, we can drop the sum over ρ because ρ is determined as ψκ (and
κ is unique since π is an epimorphism). The last inequality is by Lemma 3.12, where a is
as in the definition of behaved. The implicit constant in the ≪ depends on the Ki and C,˜
which in turns depends on only C, H, and E. The final sum converges by the definition of
behaved and the face that the first two sums are finite.
We combine Lemma 4.7 and Proposition 4.8 to give a general result from which (Existence)
will follow.
Proposition 4.9. Let C be a regime, and C¯ ⊂ C a sub-regime. Let E be an object of C, and
¯ If (MG )G is behaved at the level generated by C and E and vC,F ≥ 0 (defined in
let H̄ ∈ C.
(1.5)) for all F ∈ C, then
X X X X X X X X X
Tπ mG = Tπ mG
G∈I F ∈I ρ∈Epi(F,E) π∈Epi(G,F ) H∈IC G∈I F ∈I ρ∈Epi(F,E) π∈Epi(G,F )
F C̄ =H̄ ρπ rel. level C¯ H C̄ =H̄ F C =H ρπ rel. level C
By Proposition 4.8, we have that the sum defining SH,i,j is absolutely convergent. Using
Lemma 4.7, we can subtract from SH,i,i+1 the terms such that ρπ is not level Ci and conclude
that
(4.10) SH,i,i+1 = SH,i,i .
Since the sum defining SH,i,i+1 is absolutely convergent, we can rearrange the order of
summation to obtain
X X X X X X X X X
Tπ mG = Tπ mG ,
G∈I F ∈I ρ∈Epi(F,E) π∈Epi(G,F ) H ′ ∈Ci+1 G∈I F ∈I ρ∈Epi(F,E) π∈Epi(G,F )
F Ci ≃H ρπ rel. level Ci+1 H ′ Ci =H F Ci+1 ≃H ′ ρπ rel. level Ci+1
i.e.
X
(4.11) SH,i,i+1 = SH ′ ,i+1,i+1 .
H ′ ∈Ci+1
(H ′ )Ci ≃H
We have
X X X X
SH,t,t = Tπ mG (Lemma 4.6)
G∈Ct F ∈I ρ∈Epi(F,E) π∈Epi(G,F )
F Ct ≃H
X X X
= TG,F mG (absolute convergence)
F ∈I ρ∈Epi(F,E) G∈Ct
F Ct ≃H
≥0 (vC,F ≥ 0).
Using (4.11) and (4.10), we inductively see that for all i, we have SH,i,i , SH,i,i+1 ≥ 0. This
non-negativity allows us to rearrange the following sum and obtain, for any H̄ ∈ C0 .
X X X
(4.12) SH ′ ,i+1,i+1 = SH ′ ,i+1,i+1.
H ′ ∈Ci+1 H∈Ci H ′ ∈Ci+1
(H ′ )C0 ≃H̄ H C0 ≃H̄ (H ′ )Ci ≃H
If we expand SH ′ ,i+1,i+1 with its definition, the resulting sums in (4.12) are not necessarily
absolutely convergent, as we can see in Example 4.5. This requires us to take a great deal
of care in this argument. However, the sums in (4.12) are sums of the same non-negative
terms SH ′ ,i+1,i+1 , and thus the sums both either equal some non-negative number or ∞.
Combining (4.12), (4.11), (4.10), we obtain
X X
(4.13) SH ′ ,i+1,i+1 = SH,i,i .
H ′ ∈Ci+1 H∈Ci
(H ′ )C0 ≃H̄ H C0 ≃H̄
26
From (4.13), we see inductively that for any n, we have
X X
(4.14) SH,n,n = SH ′ ,0,0 = SH̄,0,0 ,
H∈Cn H ′ ∈C0
H C0 ≃H̄ H C0 ≃H̄
and in particular the sum on the left hand side in (4.14) converges. If B0 generates C¯ and
Bn generates C then this establishes the proposition.
Now we will choose B0 ⊂ B1 ⊂ · · · ⊂ Bt so as the prove the proposition. If B̂ is the
closure under taking quotients of a set B of objects, then we note that B and B̂ generate
the same level. Thus to prove the proposition, we can assume without loss of generality that
C and C¯ are generated by finite sets B, B̄ respectively, both closed under taking quotients.
Let Bt = B\ ∪ {E}. We construct Bt−j for j ≥ 1 by removing from Bt−j+1 one element that
is not a proper quotient of any element of Bt−j+1 , and is not in B. Since B is closed under
taking quotients, this is possible unless Bt−j+1 = S, and eventually we get to that point and
Bn = B. Now we similarly construct a chain from Bn = B to B0 = B̄ such that Bi−1 ⊂ Bi
and every proper quotient of an element of Bi is in Bi−1 . Thus we have B0 ⊂ B1 ⊂ · · · ⊂ Bt
satisfying the assumptions at the start of the proof, and also such that B0 generates C¯ and
Bn generates C, which proves the proposition.
We now can prove the “if” part of (Existence), which we restate here.
Theorem 4.15. Let C be a diamond category, C be a level of C, and let (MG )G ∈ RC be
behaved at C with vC,F ≥ 0 for all F ∈ C. Then νC ({F }) = vC,F defines a measure νC on C
with moments (MG )G .
Proof. Let E ∈ C. We apply Proposition 4.9 with C¯ containing only the minimal objects.
Note that an epimorphism F → E is relative level C¯ if and only if it is an isomorphism. The
left-hand side of Proposition 4.9 is
XX X X
Tπ mG = ME .
G∈I F ∈I ρ∈Epi(F,E) π∈Epi(G,F )
ρπ an isom.
XX X X pnC,F
= sup
n∈N
G∈C F ∈C φ∈Epi(G,H) π∈Epi(F,G)
|Aut(G)|
semisimple
dim φ≤2
28
X X X X X 1
= sup pnC,F .
n∈N
F ∈C ρ∈Epi(F,H) G∈C φ∈Epi(G,H) π∈Epi(F,G)
|Aut(G)|
semisimple φπ=ρ
dim φ≤2
Note that given ρ as in the sum, if for π ∈ Epi(F, G) there exists a φ such that φπ = ρ, then
that φ is unique. Thus we can evaluate the inner sum above by
X X X 1
= |{G ∈ [H, F ] | G → H semisimple of dim. ≤ 2}.|
G∈C
|Aut(G)|
φ∈Epi(G,H) π∈Epi(F,G)
semisimple φπ=ρ
dim φ≤2
Let N be the sum above. Then by restricting the sum over d in the definition of N to a
particular value, we obtain for any d ∈ N that
X X N
sup pnC,F ≤ d+1
.
n∈N
F ∈C ρ∈Epi(F,H) 2
+1
dim sρ =d
This gives
X X X X N
sup pnC,F ≤ d+1
≪ N < ∞,
d∈N
n∈N
F ∈C ρ∈Epi(F,H) d∈N 2
+1
dim sρ =d
as desired.
If we have an epimorphism F → R0 , we can take its radical F → R1 , and then take the
radical of F → R1 , etc., which leads to the following definition. Given a sequence of maps
φ ρk ρ1
F → Rk → Rk−1 · · · → R0 , we say φ, ρk , . . . , ρ1 is a radical sequence if for all 1 ≤ i ≤ k,
we have that ρi+1 ◦ · · · ◦ ρk ◦ φ : F → Ri is a radical of ρi ◦ ρi+1 ◦ · · · ◦ ρk ◦ φ : F → Ri−1 .
Lemma 4.18 will allow us to inductively prove the following result using the Fatou-Lebesgue
theorem, moving the limit further and further past the sum as k increases.
29
Lemma 4.20. For some level C, for all F ∈ C and n ∈ N let pnC,F ≥ 0 be a real number.
Suppose that, for all G ∈ C, we have that (4.19) holds. Then, for any R0 ∈ C, for all natural
numbers k we have
X X X X pnC,F
lim sup SF,R0 pnC,F ≤ lim sup Qk .
n→∞
F ∈C d1 ,...,dk ∈N
n→∞
F,R1 ,...,Rk ∈C φ,ρk ,...,ρ1 rad. seq. i=0 |Aut(Ri )|
φ∈Epi(F,Rk )
ρi ∈Epi(Ri ,Ri−1 )
dim(ρi )=di
Proof. The proof is by induction on k. The case k = 0 is trivial. Now we assume the lemma
is true for k. An epimorphism F → Rk has a radical F → Rk+1 , where the object Rk+1 is
unique up to isomorphism, the map F → Rk+1 is unique up to composition with elements
of Aut(Rk+1 ), and the resulting map Rk+1 → Rk is semisimple. Thus we have
X X pnC,F X X X pnC,F
Qk = Qk+1 .
F,R1 ,...,Rk ∈C φ,ρk ,...,ρ1 rad. seq. i=0 |Aut(Ri )| dk+1 ∈N F,R1 ,...,Rk+1 ∈C b,ρk+1 ,...,ρ1 rad. seq. i=0 |Aut(Ri )|
φ∈Epi(F,Rk ) φ′ ∈Epi(F,Rk+1 )
ρi ∈Epi(Ri ,Ri−1 ) ρi ∈Epi(Ri ,Ri−1 )
dim(ρi )=di dim(ρi )=di
(where the sums are over the same sets as in the previous equation). Given d1 , . . . , dk ,
Lemma 2.8 says there are only finitely many possible isomorphism classes of objects that can
be R1 , . . . , Rk in the above sums, as dim(ρi ) = di . Thus the sum over R1 , . . . , Rk , ρ1 , . . . , ρk
can be freely exchanged with the lim sup. Fixing R1 , . . . , Rk , ρ1 , . . . , ρk , we apply the Fatou-
Lebesgue Theorem to the remaining sum, and the hypothesis is checked by applying Lemma 4.18
with ρ = ρk+1 φ′ and H = Rk . The condition that φ′ is the radical of ρk+1 φ′ determines Rk+1 ,
φ′ (up to simply transitive Aut(Rk+1 ) orbit), and then ρk+1 from ρ. The Fatou-Lebesgue
theorem thus gives the inequality above, and we conclude the lemma by induction.
Lemma 4.20 will be useful because it turns out that at a particular level, radical sequences
cannot go on arbitrarily long without stabilizing.
Lemma 4.21. Let C be a level. Then there exists some k, depending on C, such that for
all tuples R0 , . . . , Rk ∈ C with semisimple ρi : Ri → Ri−1 and F ∈ C with φ ∈ Epi(F, Rk ), if
φ, ρk . . . ρ1 is a radical sequence, then φ is an isomorphism.
Proof. Let B be a finite set generating C, and let k be the maximal dimension of an element
of B. For 0 ≤ j ≤ k, let Bj be the set of all quotients of elements of C with dimension at
most j and let Cj be the level generated by Bj .
Let us check that Ri ≥ F Ci (as quotients of F ) for all i, by induction on i. It will follow
that Rk ≥ F Ck = F so Rk ≃ F , as desired.
The case i = 0 is automatic because F C0 is the minimal quotient of F and thus is ≤ R0 .
For the induction step, assume Ri ≥ F Ci . By Lemma 2.1, [Ri , F Ci+1 ∨ Ri ] is isomorphic
to [F Ci+1 ∧ Ri , F Ci+1 ], and the latter is a subinterval of [F Ci , F Ci+1 ] since F Ci ≤ F Ci+1 and
F Ci ≤ Ri . By Lemma 2.9, the natural morphism F Ci+1 → F Ci is semisimple for all i, so
30
[F Ci , F Ci+1 ] is a complemented modular lattice, and thus [F Ci+1 ∧ Ri , F Ci+1 ] is complemented
by Lemma 2.3, and hence [Ri , F Ci+1 ∨ Ri ] is as well. Because F Ci+1 ∨ Ri → Ri is semisimple,
we have Ri+1 ≥ F Ci+1 ∨ Ri , since Ri+1 is the maximal quotient that F → Ri factors through
that is semisimple over Ri . Thus, Ri+1 ≥ F Ci+1 , verifying the induction step.
Combining Lemmas 4.20 and 4.21, we show that under the assumptions (4.19) (“lim sup
of moments is finite”) and that a limit distribution exists, we can take a limit before or after
taking moments (“limit moments exist and are moments of the limit”).
Lemma 4.22. Let C be a level and for all F ∈ C and n ∈ N let pnC,F ≥ 0 be a real number.
Suppose that, for each F ∈ C, we have that the limit limn→∞ pnC,F exists and for all G ∈ C,
we have
X
(4.23) sup SF,G pnC,F < ∞
n
F ∈C
Then for all G ∈ C we have
X X
(4.24) lim SF,G pnC,F = SF,G lim pnC,F .
n→∞ n→∞
F ∈C F ∈C
Proof. Choose k as in Lemma 4.21, the maximal dimension of an element of C. Lemma 4.20
with R0 = G gives
X
n
X X X pnC,F
lim sup SF,G pC,F ≤ lim sup Qk .
n→∞
F ∈C d1 ,...,dk ∈N
n→∞
F,R1 ,...,Rk ∈C φ,ρk ,...,ρ1 rad. seq. i=0 |Aut(Ri )|
φ∈Epi(F,Rk )
ρi ∈Epi(Ri ,Ri−1 )
dim(ρi )=di
Lemma 4.21 says that is in the above sum, F is determined by Rk and φ must be an
isomorphism. Then, Lemma 2.8 implies that, given d1 , . . . , dk , the remaining sums over
F, R1 , . . . , Rk , φ, ρk , . . . , ρ1 only have finitely many terms and can be exchanged with the
lim sup. Thus the right-hand side above is equal to
X X X 1 X X limn→∞ pnC,F
Qk lim pnC,F = .
i=0 |Aut(Ri )| n→∞ |Aut(G)|
d1 ,...,dk ∈N F,R1 ,...,Rk ∈C φ,ρk ,...,ρ1 rad. seq. F ∈C ρ∈Epi(F,G)
φ∈Epi(F,Rk )
ρi ∈Epi(Ri ,Ri−1 )
dim(ρi )=di
The equality above follows because if we take ρ = ρ1 ◦ · · · ◦ ρk ◦ φ, then, inductively each Ri−1
and ρi ◦ · · · ◦ ρk ◦ φ for 2 ≤ i ≤ k + 1 is determined, up to simply transitive Aut(Ri−1 ) orbit,
from the radical sequence condition, which then determines ρi−1 (from ρ and ρi ◦ · · · ◦ ρk ◦ φ).
Combining, we obtain
X X
lim sup SF,G pnC,F ≤ SF,G lim pnC,F .
n→∞ n→∞
F ∈C F ∈C
Fatou’s lemma gives
X X
lim inf SF,G pnC,F ≥ SF,G lim pnC,F ,
n→∞ n→∞
F ∈C F ∈C
and the lemma follows.
Finally, this allows us to prove (Robustness) conditional on (Uniqueness).
31
Theorem 4.25. Let C be a diamond category and C be a level. Let (MG )G ∈ RC such that
the following implication holds:
• For any (qC,F )F ∈ [0, 1]CP
P
, if for all G ∈ C we have F ∈C SF,G qC,F = mG , then for all
F ∈ C, we have qC,F = G∈C TG,F mG .
For all F ∈ C and n ∈ N let pnC,F ∈ [0, 1]. Suppose that for each G ∈ C, we have
X
(4.26) lim SF,G pnC,F = mG .
n→∞
F ∈C
Proof. By a diagonal argument the pnC,F have a weak limit p∞C,F in n, i.e. there is a subsequence
nj
nj of N such that for all F , we have limj→∞ pC,F = p∞
C,F . By Lemma 4.22 and (4.26), for all
G ∈ C, X
SF,G p∞
C,F = mG .
F ∈C
By the assumed implication, this implies that for all F ∈ C,
X
p∞
C,F = TG,F mG .
G∈C
pnC,F
has p∞
P
The fact that in every weak limit of C,F equal to the same value G∈C TG,F mG
implies that X
lim pnC,F = TG,F mG .
n→∞
G∈C
Finally, we record how Theorem 1.6 follows from the three theorems of this section.
Proof of Theorem 1.6. Well-behavedness implies the assumption (4.3) of Theorem 4.2, which
gives (Uniqueness) and the “only if” part of (Existence). Theorem 4.15 gives the “if” part of
(Existence). Theorem 4.25, using (Uniqueness) to satisfy the hypothesis, gives (Robustness).
Since the elements of Lℓ are a countable partition of X, we also have that X \B is a countable
disjoint union of the Vℓ,i ∈ Lℓ not among the Uℓ,j . Hence for every M, by finite additivity
and the non-negativity of ν, we have
M
X M
[ M
X
ν(Uℓ,j ) ≤ ν(B) ≤ ν(X \ Vℓ,i ) = b − ν(Vℓ,i ).
j=1 i=1 i=1
Since any open set in our topology is a disjoint union of basic opens, using Fatou’s lemma
and then the Portmanteau theorem proves Theorem 1.8.
5.2. Pro-isomorphism classes and pro-objects. We now explain the relationship be-
tween pro-isomorphism classes and the usual notion of pro-object.
One defines a pro-object of C to be a functor F from a small category I to C, where the
category I is cofiltered in the sense that for any two objects of I there is an object with a
morphism to both, and for any two morphisms f, g between objects A, B ∈ I there is an
object C ∈ I with a morphism h : C → A such that f ◦ h = g ◦ h [KS06, Definition 6.1.1
and Definition 3.1.1]. We think of a pro-object as the formal inverse limit of the diagram
35
(I, F ). The morphisms between two pro-objects (I, F ) → (J, G) is given by the formula
[KS06, (2.6.4)]
Hom((I, F ), (J, G)) = lim lim Hom(F (i), G(j)).
←− −→
j∈J i∈I
It is not hard to define the identity morphism and composition for morphisms, making the
pro-objects into a category (the pro-completion of C.) We can view an object of C as a
pro-object using a category I with one object and only the identity morphism.
Definition 5.4. We say a pro-object is small if it arises (up to isomorphism) from a diagram
(I, F ) where every F (f ) is an epimorphism for every morphism f of I, and if for each G ∈ C,
the set of epimorphisms from (I, F ) to G is finite.
This definition is in analogy to the definition of small profinite groups (see [FJ08, Section
16.10], and also Lemma 5.8 below).
We now construct the bijection between isomorphism classes of small pro-objects and pro-
isomorphism classes. We first prove a quick lemma describing the quotients of a pro-object
Lemma 5.5. Let C be a diamond category and X be a small pro-object of C. The poset of
quotients of X in C is a modular lattice.
Proof. Let X arise from a diagram (I, F ) with all morphisms in the image F epimorphisms.
We first note that, for G ∈ C, we have Hom(X, G) = limi∈I Hom(F (i), G) so every morphism
−→
X → G arises from some morphism F (i) → G. Next observe that a morphism X → G is an
epimorphism if and only if it arises from an epimorphism F (i) → G, i.e. that Epi(X, G) =
limi∈I Epi(F (i), G). The case “only if" is immediate and “if" uses that all morphisms in the
−→
image of F are epimorphisms.
It follows that the poset of quotients of X is the forward limit over i ∈ I of the poset of
quotients of F (i). For F (f ) : F (i) → F (j) an epimorphism, the induced map from the poset
of quotients of F (j) to the poset of quotients of F (i) is an inclusion of modular lattices (in
particular, respects meets and joins). It follows that the limit is a modular lattice.
Lemma 5.6. Let C be a diamond category and X be a small pro-object of C. Then, for any
level C, the object X has a maximal level-C quotient X C , i.e. an object of C of level C that
is a quotient of X such that every other such quotient factors through it.
Proof. Let B be a generating set of C and m the maximum dimension of an object of B,
and let Bn be the set of quotients of objects of B of dimension at most n. Let Cn be the
level generated by Bn . We will prove for all n that X has a maximal level-Cn quotient by
induction on n, and then the claim follows from the n = m case. For the base case, we note
that the cofiltered and modular lattice properties imply that all F (i) have the same minimal
quotient, and we see that this is X C0 , the maximal level-C0 quotient of X.
For the induction step, let X Cn−1 be a maximal level-Cn−1 quotient of X. Let K be a
level-Cn quotient of X. Then, by Lemma 2.9, K → K Cn−1 is semisimple. Since K Cn−1
is a level-Cn−1 quotient of X, we have K Cn−1 ≤ X Cn−1 , and thus K ∨ X Cn−1 → X Cn−1 is
semisimple by Lemma 2.1 for modular lattices. It suffices to show that there is a maximal
level-Cn -quotient of X with a semisimple morphism to X Cn−1 , as then K ∨ X Cn−1 , which has
those properties, will factor through it, and then K will as well.
To do this, consider a level-Cn quotient H of X such that H → X Cn−1 is semisimple of
fixed dimension d. There are finitely many possible isomorphism classes of such objects
36
H by Lemma 2.8. Each one admits finitely many morphisms from X by the smallness
assumption, so there are finitely many such quotients of X. In particular, taking d = 1,
there are finitely many level-Cn quotients of X with a simple epimorphism to X Cn−1 . If
H → X Cn−1 is semisimple of dimension d, then it factors through at least d different simple
morphisms Gi → X Cn−1 of quotients of K (by Lemma 4.16) and hence quotients of X, and
so the dimension of any semisimple morphism from a level-Cn quotient of X to X Cn−1 is
bounded in terms of X, C, and n. Consider the level-Cn quotients of X with a semisimple
morphism to X Cn−1 . There are finitely many possible dimensions, and then finitely many
of each dimension. Thus, such quotients form a finite set and the join of two elements of
this set is an element of this set (by Lemmas 5.5 and 2.3), it must have a maximal member,
completing the induction step.
Lemma 5.7. Let C be a diamond category. For each pro-isomorphism class X of C, there
exists a small pro-object Y of C, unique up to isomorphism, such that Y C ∼ = X C for each
level C.
In other words, the map from the set of isomorphism classes of small pro-objects to P that
sends Y to (Y C ) is a bijection.
Proof. First, for each level C we chose an object in the isomorphism class X C , and by slight
abuse of notation we will now write X C for this object, and we make analogous choices
throughout this proof. We will next show the existence of an assignment to each pair C1 , C2
of levels with C1 ⊂ C2 an epimorphism πCC12 : X C2 → X C1 such that whenever C1 ≤ C2 ≤ C3
we have πCC12 ◦ πCC23 = πCC13 . Since there are finitely many epimorphisms X C2 → X C1 for each
C2 , C1 , to check that these infinitely many conditions can be satisfied, by the compactness of
an infinite product of finite sets it suffices to check that every finite subset of them can be
satisfied. Any finite set of equations of the form πCC12 ◦ πCC23 = πCC13 involves only finitely many
regimes Ci , and these finitely many regimes Ci are contained in a single regime C.
For each Ci , fix an epimorphism πi : X C → (X C )Ci realizing the definition of (X C )Ci . Then
for every i, j with Ci ⊂ Cj , there is a unique epimorphism πij : (X C )Cj → (X C )Ci such that
C
πij πj = πi . We define πCij for each i, j with Ci ⊂ Cj by composing πij with fixed isomorphisms
C
(X C )Ck ≃ X Ck chosen for each k. One can check these πCij satisfy all of the required equations
involving only Ci ⊆ C, completing the argument for existence of the πCC12 .
Any epimorphism X C1 → X C1 is an isomorphism by the finiteness of quotient posets. Thus
any epimorphism πCC12 : X C2 → X C1 realizes X C1 as the maximal level-C1 quotient of X C2 , since
it factors through the maximal level-C1 quotient of X C2 (which we know, by the definition
of a pro-isomorphism class, is abstractly isomorphic to X C1 ).
Define a category I to have one object for each level C and a single morphism C2 → C1
if and only if C1 ⊂ C2 . Let Y be the pro-object defined using the index category I and the
functor sending C to X C , and the unique morphism C2 → C1 to πCC12 . Let’s check that Y is
a small pro-object and that Y C ≃ X C . It suffices to check that Y C ≃ X C for each level C,
since taking C to be generated by H, the number of epimorphisms Y → H is then bounded
by the number of epimorphisms X C → H, which is finite.
′
To do this, let F be any level-C quotient of Y . Then F must be a quotient of X C for
some C ′ by definition of Y . Let C˜ be the level generated by C ′ and C. Every epimorphism
˜
X C̃ → F factors through πCC̃ : X C → X C by our observation above that this is the maximal
37
level C quotient of X C̃ . Thus every level-C quotient of Y factors through Y → X C , which
completes the proof of existence.
It remains to show uniqueness up to isomorphism. In other words, we fix a small pro-
object Z with Z C ∼ = XC ∼ = Y C for all C and must construct an isomorphism Y → Z. As
above, we can choose compatible maps πCC12 : X C2 → X C1 for all C1 ⊂ C2 . For each level C,
we choose a map τC : Z → Z C (which realizes Z C as the maximal level C quotient of Z),
and then define τCC12 : Z C2 → Z C1 for all C1 ⊂ C2 so that τC1 = τCC12 τC2 . Next we will choose
a system of isomorphisms X C → Z C (for each C) that are compatible in the sense that, for
C2 ⊂ C1 , they form a commutative diagram with πCC12 and τCC12 . This again is choosing an
element of an infinite product of finite sets satisfying infinitely many conditions that each
depend on only finitely many terms of the product, so it suffices to check that any finite
subset of the conditions can be satisfied. To do this, we chose a level C containing all levels
appearing in a finite list of conditions, fix an isomorphism φ : X C → Z C , and note that
since πCCi : X C → X Ci and τCCi φ : X C → Z Ci both give the maximal level-Ci quotient of X C ,
then they must be isomorphic, and there is a unique isomorphism φCi : X Ci → Z Ci such that
φCi πCCi = τCCi φ. Then these φCi are compatible in the above sense.
Now we construct the isomorphism Y → Z. Writing Z = (I, F ), we must for each i ∈ I
choose a level C and a map X C → F (i), in a way compatible with all the morphisms of I.
We can choose C to be the level generated by F (i). Then there is a unique epimorphism
Z C → F (i) commuting with τC and the natural map Z → F (i). Taking the composition
with φC constructed above, we have X C → F (i), and one can check these are compatible
with the morphisms of I.
To map Z → Y , we must for each regime C choose i ∈ I and a map F (i) → X C , compatible
with all inclusions between regimes. We have φC τC : Z → X C , so there exists i ∈ I with a
morphism F (i) → X C , and choose that i and morphism.
It is straightforward to check that these two maps are inverses, and so Y ≃ Z.
∞ n
!
1 X Y (−1)ei X
vCS ,N = MNα .
|Aut(N)| e ,...,e =0 i=1 |Hom(N, Ki )|ei ej=1 (qij − 1)
Qi
1
Qn ei
1 n α∈Ext (N, i=1 Ki )
We have that (MN )N is well-behaved at level CS if and only if the sum above is absolutely
convergent for every finite S-module N.
Proof. By definition we have
X X µ(π)ML
vCS ,N = .
L∈CS π∈Epi(L,N )
|Aut(N)||Aut(L)|
∞ n X
1 X Y ei
ei ( 2 )
X ML
vCS ,N = (−1) qi .
|Aut(N)| e ,...,e =0 i=1 L∈C
|Aut(L)|
1 n s π∈Epi(L,N )
Q ei
ker π ∼
= ni=1 Ki
Now ker π ∼
Qn Qn
= i=1 Kiei if and only if there exists an injection f : i=1 Kiei → N such that
im f = ker π, and the number of such f ’s is
ei ei
!
n Y n n n Y n
e
q ( 2 ) (q j − 1) .
Y Y Y Y i
Y
e
Aut( Ki i ) = |Aut(Kiei )| = |GLei (Fqi )| = (q ei −q ei −j ) =
i=1 i=1 i=1 i=1 j=1 i=1 j=1
Thus
∞ n
!
1 X Y (−1)ei X X ML
vCS ,N = Qei j .
|Aut(N)| e ,...,e =0 i=1 j=1 (qi − 1) L∈Cs
|Aut(L)|
1 n π∈Epi(L,N )
Qn ei
f: i=1 Ki →L
Im f =ker π
f injective
Associated to eachQtriple L, f, π there is a class α ∈ Ext1 (N, ni=1 Kiei ), the class of the
Q
exact sequence 0 → ni=1 Kiei →f K →π N → 0. For this α we have ∼
Qn Nα = L so MNeαi = ML .
Next we check that each class α arises from exactly Aut(Nα )/ i=1 |Hom(N, Ki )| triples
L, f, π. Two triples give the same α, if and only if have the same L and are related by an
automorphism of L. Thus the number of different triples giving α is the number of orbits
of Aut(L) on maps f, π. The stabilizer is the group of automorphisms fixing the submodule
42
Qn
Kiei and its quotient module. The group of such automorphisms is Hom(N, ni=1 Kiei )
Q
i=1
with cardinality
Yn Y n
Hom(N, ei
Ki ) = |Hom(N, Ki )|ei
i=1 i=1
which cancels the Aut(L) term and gives the desired statement.
The MN are well-behaved if and only if the sum defining vCS ,N with an additional Z(π)3
factor is absolutely convergent for all N. We have 2 ≤ Z(π) ≤ 2n , so that absolute conver-
gence happens if and only if the sum defining vCS ,N (in which n is constant) is absolutely
convergent for all N.
Now we give a commonly occurring example of moments (e.g. the Cohen-Lenstra-Martinet
distributions have such moments [WW21, Theorem 6.2]), for which we find the measure
explicitly.
Lemma 6.3. Let R be a ring, C be the category of finite R-modules and u be a real number.
For each N ∈ C, let MN = |N|−u . Then (MN )N is well-behaved.
Let S be a finite quotient ring of R, and let K1 , . . . , Kn be representatives of the isomor-
phism classes of finite simple S-modules. Let q1 , . . . , qn be the cardinalities of the endomor-
phism fields of Ki . For N a finite S-module, we have
n Y ∞
!
Ext1 (N, Ki )
1 Y S
vCS ,N = 1− |qi |−j
|Aut(N)||N|u i=1 j=1 |Hom(N, Ki )||Ki |u
and
n
Y
−u −u
MNα = |Nα | = |N| |Ki |−ei u
i=1
43
so
∞ n
! n
1 X Y (−1)ei
1
Y
e
vCS ,N = Ext (N, Ki )
i
|Aut(N)||N|u |Hom(N, Ki )|ei |Ki |ei u ej=1 (qij − 1)
Qi
e1 ,...,en =0 i=1 i=1
∞ n
! n
1 X Y (−1)ei Y
Ext1 (N, Ki )ei
= u
|Aut(N)||N| e ,...,e =0 i=1 |Hom(N, Ki )|ei |Ki |ei u ej=1 (qij − 1) i=1
Q i
1 n
∞ n
e !
(−1)ei Ext1 (M, Ki ) i
1 X Y
=
|Aut(N)||N|u e ,...,e =0 i=1 |Hom(N, Ki )|ei |Ki |ei u ej=1 (qij − 1)
Qi
1 n
n X ∞
e !
(−1)e Ext1 (N, Ki )
1 Y
=
|Aut(N)||N|u i=1 e=0 |Hom(N, Ki )|e |Ki |eu ej=1 (qij − 1)
Q
n Y ∞
!
Ext1 (N, Ki )
1 Y S −j
= 1− |qi |
|Aut(N)||N|u i=1 j=1 |Hom(N, Ki )||Ki |u
e
by the q-series identity ∞
Q∞
Q v −j
P
e=0 ej=1 (q j −1) = j=1 (1 + vq ) (from the q-binomial theorem).
This q-series sum is absolutely convergent because the denominator grows superexponen-
tially in e while the numerator grows only exponentially. All our manipulations preserve
absolute convergence in the reverse direction because we never combine two terms in the
sum with opposite signs – in fact, we never combine two terms except in the very first step,
where we group together identical terms.
|Ext1S (N,Ki )| 1
If u is an integer, then |Hom(N,K i )||Ki |
u is an integer power of qi since ExtS (N, Ki ), Hom(N, Ki ),
and Ki are all vector spaces over Fqi and hence have cardinality an integer power of qi . If
|Ext1S (N,Ki )|
|Hom(N,Ki )||Ki |u
is a positive integer power of qi then one of the terms in the product over j
vanishes, so the product is 0 and thus is nonnegative, and if it is a nonnegative integer power
of qi then all the terms in the product over j are positive, so the product is nonnegative.
Thus vCS ,N ≥ 0.
Corollary 6.5. Let R be a ring and let C be the category of finite R-modules. Let (MN )N ∈
RC/≃ . If MN = O(|N|n ) for some real n, then (MN )N is well-behaved.
Proof. This follows immediately form Lemma 6.3, taking u = −n, and Lemma 2.11.
Many special cases of both the uniqueness and robustness that follows from Corollary 6.5,
and the formulas for particular measures in Lemma 6.3 have appeared in previous work. For
example, Fouvry and Klüners prove uniqueness for the case R = Z/pZ [FK06a, Proposition
3] and give formulas for the probabilities when the moments are all 1 [FK06a, Proposition
2]. Ellenberg, Venkatesh, and Westerland prove uniqueness and give formulas for R = Zp
when all moments are 1, as well as prove robustness [EVW16, Section 8.1]. When R = Z,
the second author proved uniqueness and robustness given by Corollary 6.5, as well as giving
the formulas for moments |N|−u (see [Woo17, Theorem 8.3] and [Woo19, Section 3]). When
R is a maximal order in a semi-simple Q-algebra, Wang and the second author did the same
[WW21, Section 6]. For a R with finitely many (isomorphism classes of) simple R-modules
and such that Ext1R between any two finite R-modules is finite, the first author [Saw20,
Theorem 1.3] proved the uniqueness and robustness given by Corollary 6.5.
44
Under mild hypotheses, the measures from Lemma 6.3 have a nice description in terms of
a limit of random modules from generators and relations.
Lemma 6.6. Let R be a ring such that there are only countable many isomorphism classes
of finite R-modules (such as a Noetherian local commutative ring with finite residue field).
Let R̂ be the profinite completion of R.
As n → ∞, the distribution of the quotient of R̂n by n + u random relations taken from
Haar measure on R̂n weakly converges to µu (defined above as the measure on pro-finite
R-modules with moments |N|−u ).
Proof. A finitely generated R̂-module N naturally gives a small pro-finite R-module. Each
continuous finite quotient ring R̂/I is a finite quotient of R, and the functor from the category
of continuous finite quotients of R̂ taking R̂/I to the finite R-module R̂/I ⊗ N gives the
pro-object. Thus we can apply Theorem 1.8 to show that the distribution of the quotient
of R̂n by n + u random relations taken from Haar measure converges to µu as long as its
N-moment converges to |N|−u for all N.
This is true since, for each surjection π from R̂n to N, the probability that π descends to
the quotient of R̂n by n + u Haar random relations is |N|−n−u , since this is equivalent to
the probability that the images of the n + u random relations in N are all trivial. So the
expected number of surjections from the quotient of R̂n by n + u Haar random relations is
Sur(R̂n , N)|N|−n−u . As n goes to ∞, the ratio Sur(R̂n , N)/|N|n converges to 1 since it is
equal to the probability that n random elements of N generate N, which converges to 1.
Hence the expected number of surjections from the quotient of N converges to |N|−u , as
desired.
For the rest of Section 6.1, we restrict attention to commutative rings.
When R is a local ring, the formulas for µu simplify even further, and we can make more
computations about the measure.
Lemma 6.7. Let R be a Noetherian local commutative ring with finite residue field k. We
have
∞
!
Ext1 (L, k)
Y S
µu (US,L ) = |Aut(L)|−1 |L|−u 1− |k|−j−u
j=1
|Hom(L, k)|
for any nonzero finite quotient S of R.
Proof. Since R is a local ring, the only simple R-module is the residue field k, and for S 6= 0,
the residue field is an S-module, so we can take n = 1 and N1 = k in Lemma 6.3. In this
case, the formula of Lemma 6.3 simplifies to the stated one.
One interesting set whose measure we can calculate in the local rings case is the set of
finite R-modules. This is a measurable set since it is a countable union of closed points.
Lemma 6.8. Let R be an infinite Noetherian local commutative ring with finite residue field
k. We have
0
if u < 0
Q∞ −j−u
µu ({N | N finite}) = µu ({0}) = j=1 (1 − |k| ) if 0 ≤ u < dim R − 1
1 if u ≥ dim R − 1
45
where dim R is the Krull dimension. In the middle case, saying that the measure is µu ({0})
is equivalent to saying each nonzero finite R-module has measure 0.
Proof. Let N be a finite R-module. Then N has a minimal presentation in terms of d
generators and m relations for d = dim Hom(N, k) and m = dim Ext1R (N, k). We can choose
a finite quotient S of R such that N factors through S, as do all finite extensions of N by k,
so that Ext1S (N, k) = Ext1R (N, k). It follows that for an R module L, we have L ⊗R S ∼
=N
∼
if and only if L = N, with “if" obvious and “only if" because otherwise some quotient of L
would be an extension of N by k and thus an S-module, a contradiction.
Then
∞
Y
µu ({N}) = µu (US,N ) = |Aut(N)|−1 |N|−u (1 − |k|m−d−j−u )
j=1
and this upper bound satisfies the assumption of the dominated convergence theorem since
X ∞
Y Y∞ Y∞
−j −1 −j −1
µu ({N}) (1 − |k| ) ≤ µu (P) (1 − |k| ) = (1 − |k|−j )−1 < ∞.
N finite R-module j=1 j=1 j=1
For a discrete valuation ring, we can give a simple condition for a collection of moments to
be well-behaved that is almost sharp. For an R-module N, we write ∧2 N for the quotient of
the R-module N ⊗ N by the R-module generated by elements of the norm n ⊗ n for n ∈ N.
Lemma 6.9. Let R be a discrete valuation ring with finite residue field. Let MN be an
assignment of a real number to each isomorphism class of R-modules. Then MN is well-
behaved if there exists ǫ > 0 and c such that MN ≤ c|N|1−ǫ |∧2 (N)| for all finite R-modules
N.
We then obtain corollaries for uniqueness of measures from Theorems 1.6 and 1.7. A
slightly weaker uniqueness result appears in [Woo17, Theorems 8.2 and 8.3], without the
|N|1−ǫ in the upper bound (though stated there over Zp , the argument works over other
DVR’s with finite residue field, e.g. see [WW21, Theorem 6.11]).
47
Lemma 6.9 is almost sharp because, if we set MN = |N||∧2 (N)| = Sym2 (N) then the
moments are attained for two distinct measures µ: The probability distribution the quotient
of R̂n by the columns of a Haar random n×n skew-symmetric matrix converges as n → ∞ for
n even to one measure, as n → ∞ for n odd to a different measure, which both have the
and
same moments Sym2 (N). (See [CKL+ 15, Theorems 10 and 11] for an argument showing
moments in a different, but analogous case, and see [BKL+ 15, Theorems 3.9 and 3.11] for
the limiting distributions in the case R̂ = Zp .) So if we allow ǫ = 0 then MN ≤ c|N||∧2 (N)|
are not necessarily well-behaved since the conclusion of Theorem 1.7 is not necessarily true.
Proof. Let k be the residue field of R and q its cardinality. Let π be the uniformizer of R.
From Proposition 6.2, since k is the only simple R-module, it suffices to prove for each
finite R-module N that
∞
!
1 X 1 X
c|Nα |1−ǫ ∧2 (Nα ) < ∞.
e Qe j
|Aut(N)| e=0 |Hom(N, k)| j=1 (q − 1)
α∈Ext1 (N,k e )
F3 /F2 ∼
= ∧2 N,
a ∧ b ∈ πNα ∧ b = Nα ∧ πb = Nα ∧ 0 = 0,
2 (e−rank(α))(e−rank(α)−1)
∧ (Nα ) = |F3 /F2 ||F2 /F1 ||F1 | ≤ ∧2 (N)q en ∧2 (k e / Im(α)) = ∧2 (N)q en q
2 .
is absolutely convergent.
Proof. By definition we have
X X µ(π)MG
vC,F = .
G∈C π∈Epi(G,F )
|Aut(G)||Aut(F )|
The µ(π) term vanishes unless π is semisimple, in which case ker π is a product of finite
simple [F ]-groups, of which the abelian ones are irreducible representations of F over prime
fields, So, we can write ker π as ni=1 Viei × m fi
Q Q
i=1 Ni for unique tuple e1 , . . . , en , f1 , . . . fm
(ei ) Q
of nonnegative integers. In this case, µ(π) = ni=1 (−1)ei qi 2 · m fi
Q
i=1 (−1) by Lemma 3.10.
Thus
n Ym
1 X Y ei
ei ( 2 )
X X MG
vC,F = (−1) qi (−1)fi ,
|Aut(F )| e∈Nn ,f ∈Nm i=1 i=1 G∈C
|Aut(G)|
π∈Epi(G,F )
Q ei Qm fi
ker π ∼
= ni=1 Vi × i=1 Ni
For each triple G, π, ι we can quotient G by all the factors of ker π except ι(Viei ), obtaining
an extension of F by Viei , with an extension class αi ∈ H 2 (F, Vi )ei . We can quotient further
to obtain ei individual extensions of F by Vi , each of which must be level-C, hence αi ∈ W ei .
There is a natural isomorphism from G to the fiber product, over F , of each of the minimal
extensions of F obtained by quotienting G by all of the Vi and Ni factors except one. One
can check that the minimal extension by Ni is naturally isomorphic to Aut(Ni ) ×Out(Ni ) F ,
where G maps to Aut(Ni ) by its conjugation action and to F by the given map (using that
the center of Ni is trivial). Thus we have an isomorphism G → Ge,f ,α , respecting both the
map from ni=1 Viei × m fi
Q Q
i=1 Ni and the maps to F .
We next check that, for given e, f, each class α arises from exactly
n ei
Y |H 0 (F, Vi )|
Aut(Ge,f ,α)
i=1
|Vi ||H 1 (F, Vi )|
triples G, π, ι. First, we note that each class α arises from G = Ge,f ,α and taking π, ι to be
the natural maps, as Ge,f ,α ∈ C. We take the group Ge,f ,α to be the representative of its
isomorphism class we use for the sum. From the last paragraph, we see that if Ge,f ,α , π ′ , ι′
gives α, then there is some automorphism β of Ge,f ,α such that ι′ = βι and π ′ = πβ. So
Aut(Ge,f ,α ) acts transitively on the non-empty set of triples G, π, ι that give α, and the
stabilizer is the subgroup of automorphisms preserving ι and π.
So it suffices to prove ei the number of automorphisms of Ge,f ,α that preserve ι and π
Qn |Vi ||H (F,Vi )|
1
is i=1 |H 0 (F,Vi )|
. To do this, note that any such automorphism sends g ∈ G
to the product of g with an element of ni=1 Viei × m fi
Q Q
i=1 Ni , which since the automor-
phism fixes Nifi and thus fi
Qnfixesei the action of g on Ni 1 must Qncommute with Nifi for all
i and thus must lie in i=1 Vi , giving a cocycle in C (F, i=1 Viei ), and every cocycle
gives such
Qn an automorphism. The number of cocycles is the size of the cohomology Qn group
Qn ei
1 ei 1
|H (F, i=1 Vi )| = i=1 |H (F, Vi )| times the number Q of coboundaries, which is i=1 |Vi |ei
e
divided by the number of cocycles in degree 0, which is ni=1 |H 0 (F, Vi )| i , giving the stated
formula.
Thus, we can replace the sum over G, π, ι with a sum over α after multiplying by
n ei
Y |H 0 (F, Vi )|
Aut(Ge,f ,α) ,
i=1
|Vi ||H 1 (F, Vi )|
51
which cancels the Aut(G) term and gives the desired statement.
By definition, the assignment MG is well-behaved if and only if the sum, with an additional
Z(π)3 factor,Pm is absolutely convergent. We have Z(π) = 2
ω(π)
where ω(π) = #{1 ≤ iP≤ n |
m
ei > 0} + i=1 fi , by Lemma 3.10, and thus Z(π) is bounded by a constant 2n times 2 i=1 fi .
Thus,
Pm the MG are well-behaved if and only if the sum, with the signs removed and an
fi
8 i=1 factor inserted, is absolutely convergent. We may also ignore the Aut(F ) factor as
it is bounded.
Lemma 6.12. Let C be the category of finite groups. Let u be a real number, and for each
G ∈ C, let MG = |G|−u . Then (MG )G is well-behaved.
Keeping the notation of Proposition 6.11, we have
n ∞ ! Ym −
|F |−u Y Y
1
|H 0 (F, Vi )||Wi | −j |Ni |u |Aut[F ] (Ni )|
vC,F = 1− qi e .
|Aut(F )| i=1 j=1 |Vi |u+1 |H 1 (F, Vi )| j=1
for every finite group G. So then Theorem 1.8 implies that the Xg weakly converge in
distribution to a probability measure on the space of profinite groups, indeed the one given
by the formulas in Lemma 6.12. This recovers [LW20, Theorem 1.1]. One can reconcile the
formulas in Lemma 6.12 with [LW20, Equation (3.2)] using observations from the proof of
Proposition 6.11.
Proof. We observe that
n
Y m
Y n
Y m
Y
−u ei fi −u −u −uei
MGe,f ,α = |Ge,f ,α | = (|F | · |Vi | · |Ni | ) = |F | · |Vi | · |Ni |−ufi
i=1 i=1 i=1 i=1
is. e
|H 0 (F,Vi )| i |Vi |−uei |Wi |ei
Now note that every term in the ratio |V |ei |H 1 (F,v )|ei Qei (qj −1) is exponential in ei , ex-
i i j=1 i
cept for ej=1 (qij − 1), which is superexponential, so the ratio decays superexponentially.
Qi
52
8fi |Ni |−ufi
Similarly, every term in f is exponential in fi except for (fi )!, which is super-
|Aut[F ] (Ni )| i (fi )!
exponential, so the ratio decays superexponentially. Thus, the term being summed decays
superexponentially in all the ei , fi variables, and hence the sum is absolutely convergent.
To calculate vC,F , we similarly obtain
n
! m !
e
|F |−u X Y (−1)ei |H 0(F, Vi )| i |Vi |−uei Y (−1)fi |Ni |−ufi X
vC,F = 1
|Aut(F )| e∈Nn ,f ∈Nm i=1 |Vi |ei |H 1 (F, vi )|ei ej=1 (qij − 1) i=1 Aut[F ] (Ni )fi (fi )!
Q i
Qn ei
α∈ i=1 Wi
n
! m
!
e
|F |−u X Y (−1)ei |H 0 (F, Vi )| i |Vi |−uei |Wi |ei Y (−1)fi |Ni |−ufi
=
|Aut(F )| e∈Nn ,f ∈Nm i=1 |Vi |ei |H 1 (F, vi )|ei ej=1 (qij − 1) i=1 Aut[F ] (Ni )fi (fi )!
Qi
n ∞
! m ∞
!
e
|F |−u Y X (−1)e |H 0 (F, Vi )| |Vi |−ue |Wi |e Y X (−1)f |Ni |−uf
=
|Aut(F )| i=1 e=0 |Vi |e |H 1 (F, Vi )|e ej=1 (qij − 1) i=1 f =0 Aut[F ] (Ni )f f !
Q
n ∞ ! Ym −
|F |−u Y Y
1
|H 0 (F, Vi )||Wi | −j |Ni |u |Aut[F ] (Ni )|
= 1− qi e
|Aut(F )| i=1 j=1 |Vi |u+1 |H 1 (F, Vi )| j=1
P∞ e Q∞
by the q-series identity Q v −j
e=0 ej=1 (q j −1) = j=1 (1 + vq ) and the Taylor series for the
exponential function.
For u an integer, this is nonnegative because, on the one hand, the exponential terms are
|H 0 (F,Vi )||Wi |
nonnegative, and, on the other hand, |V |u+1 |H 1 (F,V )| is an integer power of q because every
i i
term is the cardinality of a vector space over Fqi raised to an integer power. (One can check
that C closed under fiber products and quotients implies that Wi is a vector space over the
endomorphism field of Vi .) If this power is positive then one of the terms in the product over
j is zero, making this factor nonnegative, and otherwise, each term in the product over j is
nonnegative, making this factor nonnegative.
Corollary 6.13. Let C be the category of finite groups. Let (MG )G ∈ RC/≃ . If MG =
O(|G|n ) for some real n, then (MG )G is well-behaved.
Proof. This follows immediately form Lemma 6.12, taking u = −n, and Lemma 2.11.
Remark 6.14. It may be possible to prove, following the calculations in [SW22b, §7], that
(MG )G is well-behaved as long as
!
|G|1−ǫ |H2 (G, Z)|
MG = O
|H1 (G, Z)|
for some ǫ > 0, but we do not pursue that here.
6.3. General tools for constructing examples. In this section, we give a general theorem
which gives a large class of examples of diamond categories, as well as tools for constructing
new diamond categories from old ones.
A Mal’cev variety [Smi76, 142 Theorem] is a variety in universal algebra (a set of abstract
n-ary operations for different natural numbers n together with a set of equations they satisfy)
such that some composition of the n-ary operations is an operation T (x, y, z) which satisfies
T (x, x, z) = z and T (x, z, z) = x. The most familiar example of such an operation is given
53
by T (x, y, z) = xy −1 z in the variety of groups, or in any variety which contains a group
operation. Given a Mal’cev variety, one has a category of finite algebras for that Mal’cev
variety. An object in this category is a finite set S with functions S n → S for each abstract n-
ary operation of the Mal’cev variety (for each n) such that the functions satisfy the equations
of the Mal’cev variety. A morphism in this category is a function S → S ′ that respects all the
operations. The congruences in any algebra for a Mal’cev variety (i.e. equivalence relations
respecting the algebra structure, analogous to normal subgroups of a group or ideals of a
ring) form a modular lattice [Bir67, Ch. VII: Theorem 4 and Theorem of Malcev].
Lemma 6.15. For any Mal’cev variety with operations of finitely many different arities, the
category of finite algebras for that Mal’cev variety, with morphisms given by surjections, is
a diamond category.
This includes the cases of finite rings, finite commutative rings, finite R-modules for a ring
R, finite groups, finite groups with the action of a fixed group Γ, finite quasigroups, finite
Lie algebras over a finite field, finite Jordan algebras over a finite field, etc.
Proof. In this proof, we use “algebra” to refer to an algebra for the given Mal’cev variety.
Since we choose morphisms to be surjections, every morphism is an epimorphism, and
the epimorphisms are dual to the lattice of congruences [Bir67, Ch. VI, Sec. 4] which is a
modular lattice as mentioned above. The lattice is finite since the number of congruences is
at most the number of partitions of the underlying set.
The automorphism group of a finite algebra is a subgroup of the permutation group of the
underlying set, and thus is finite.
The last finiteness step is more difficult. If A and B are both quotients of D, then the
elements of A ∨ B are equivalence classes of elements of D, where the equivalence is that
the equivalence defining A and the equivalence defining B both hold. Since we are in a
Mal’cev variety, the elements of A ∧ B are equivalence classes of elements of D, where the x
is equivalent to y if there is some z such that x and z are equivalent in A and z and y are
equivalent in B ([Bir67, Ch. VII: Theorem 4]). Thus, in the setwise fiber product A ×A∧B B
(which has a natural algebra structure), every element is in the image of the map from D.
Thus D → A ×A∧B B is a surjection of algebras, and the equivalence relation giving the
fibers is precisely the same as that defining the fibers of the surjection D → A ∨ B, and
hence A ∨ B = A ×A∧B B.
For any simple surjection between algebras, let the width be the maximum size of a fiber
of the underlying map of sets. For any algebra A, let the width of A be the maximum width
of any simple surjection between two quotients of A. Let n be the maximum width of an
element of a finite generating set of C.
Since A ∨ B = A ×A∧B B, we have that the sizes of the fibers of A ∨ B → A are all sizes
of fibers of B → A ∧ B and vice versa. We next will see that under certain conditions, the
width of E → F is the same as the width of E ∨ G → F ∨ G or E ∧ G → F ∧ G. Suppose
E → F is simple such that F ≥ G ∧ E (all as quotients of some object). Then ∨G gives
an isomorphism [G ∧ E, E] → [G, E ∨ G] by Lemma 2.1, and so E ∨ G → F ∨ G is simple.
Moreover, (F ∨G)∨E = E ∨G and (F ∨G)∧E = F ∨(G∧E) = F (by the modular property
and then the assumption that F ≥ G ∧ E). Thus, E ∨ G → F ∨ G is (F ∨ G) ∨ E → F ∨ G
and has the same width as E → (F ∨ G) ∧ E, which is E → F . Similarly, if E → F is simple
such that E ≤ G ∨ F , then E ∧ G → F ∧ G is simple has has the same width as E → F .
54
Our first step is to check that all level-C algebras have width at most n. By the definition
of level-C, it suffices to check that the set of algebras of width at most n is downward-stable
and join-stable. That it is downward-stable is clear by definition. For join-stable, consider a
simple surjection B → A of quotients of D ∨ E (in some lattice of quotients), where D and
E have width at most n. We have B ≥ (D ∧ B) ∨ A ≥ A so, by simplicity (D ∧ B) ∨ A, is
either equal to B or A.
If (D ∧ B) ∨ A = B, then B = (D ∧ B) ∨ A ≤ D ∨ A, so by the above D ∧ B → D ∧ A is
simple and has the same width as B → A. Since D ∧ B → D ∧ A is a simple surjection of
quotients of D, we conclude B → A has width at most n.
If (D ∧ B) ∨ A = A, then A ≥ (D ∧ B). By the above, B → A has the same width
as B ∨ D → A ∨ D. Moreover, since B ∨ D ≤ E ∨ D = E ∨ (A ∨ D), we have that
(B ∨ D) ∧ E → (A ∨ D) ∧ E is simple and also has the same width as B → A. However,
(B ∨ D) ∧ E → (A ∨ D) ∧ E is a simple map of quotients of E, and hence has width at most
n. Thus B → A has width at most n.
This concludes the proof that all level-C algebras have width at most n. It follows that, for
G of level-C, any H of level C with a simple surjection H → G has cardinality |H| ≤ |G|n.
If there are finitely many operations in the variety, we are done, as there are finitely many
algebra structure on any given finite set.
Otherwise, we need a slightly more complicated argument. Let K be the product of all
the elements of a finite generating set of C. Any equation (involving the algebra operations)
that holds in K holds in every level-C algebra, since equations are preserved by quotients
and fiber products. Since K is a finite set, there can only be finitely many distinct r-ary
maps K → K for each r, so sufficient equations hold in K so that there is some finite set of
operations, such that the equations holding on K, and hence for level-C algebras, determine
the value of every operation in terms of that finite set of operations. It then holds that there
are finitely many level-C algebras on sets of cardinality ≤ |G|n, completing the proof of the
final finiteness step.
One we have a diamond category, there are many ways to modify it and get further
diamond categories.
Lemma 6.16. Let C1 and C2 be diamond categories. Then the product category C1 × C2 ,
whose objects are ordered pairs of an object of C1 and an object of C2 and whose morphisms
are ordered pairs of morphisms, is a diamond category.
Proof. We verify the three assumptions in order:
(1) The poset of quotients of an ordered pair (G1 , G2 ) is the product of the posets of
quotients of G1 and G2 , and the product of two finite modular lattices is a finite
modular lattice.
(2) The automorphism group of an ordered pair (G1 , G2 ) is the product of the automor-
phism groups, and the product of two finite groups is a finite group.
(3) Since the poset of quotients is the product of quotients, the product of two downward-
closed sets is downward-closed, and the product of two join-closed sets is join-closed.
Thus, the level-C objects are contained in the set of ordered pairs (G1 , G2 ) where G1
is level-C1 and G2 is level-C2 , for C1 the set of first terms of objects of C and C2 the
set of second terms. A morphism to (G1 , G2 ) is simple if and only if it is simple in
one factor and an isomorphism to the other, so the finiteness of simple morphisms
55
from level-C objects follows from the finiteness of simple morphisms from level-C1 and
level-C2 objects.
The following result will cover many kinds of additional structure we may add to a diamond
category to get another diamond category.
Proposition 6.17. Let F : C1 → C2 be a faithful functor that sends epimorphisms to epi-
morphisms. Assume:
• C2 is a diamond category.
• The natural map from the poset of quotients of G to the poset of quotients of F (G) is
the inclusion of a sublattice (i.e. we have H1 ≤ H2 if F (H1 ) ≤ F (H2 ) and the image
is closed under joins and meets)
• That there is some n such that F applied to any simple epimorphism gives a morphism
of dimension at most n.
• Given a level C of C1 , there are finitely many isomorphism classes of objects G ∈ C
with F (G) isomorphic to a given object in C2 .
Then C1 is a diamond category.
Remark 6.18. When applying Proposition 6.17, it is very helpful when checking the condi-
tions to note that F faithful implies that if F (π) is an epimorphism then π is an epimorphism.
Proof. We verify the three assumptions in order:
(1) The poset of quotients of G is a sublattice of the poset of quotients of F (G). A
sublattice of a finite modular lattice is modular since a modular lattice is defined by
an algebraic equation involving joins and meets.
(2) The automorphism group of G is a subgroup of the automorphism group of F (G),
hence a finite group.
(3) Let C be a level of C1 and let C ′ be the level of C2 generated by the application of
F to a generating set of C. Then the objects G ∈ C1 such that F (G) is level C ′ are
downward-closed and join-closed by the assumption on the natural map of posets,
so F (C) ⊂ C ′ . So it suffices to prove for an object H that there are finitely many
simple morphisms G → H with F (G) ∈ C ′ . By Lemma 2.8 and assumption, there
are finitely many possibilities for F (G) up to isomorphism. By assumption, there
are finitely many possibilities for G ∈ C up to isomorphism, and then each one has
finitely many surjections to H by the previous two properties.
One such structure is adding the action of a finite group, where F is the forgetful functor.
Lemma 6.19. Let C be a diamond category. Let Γ be a finite group. Define the category
Γ-C in which objects are pairs of an object G ∈ C and a homomorphism Γ → Aut(G) and
morphisms are epimorphisms in C compatible with the action of Γ. Then Γ-C is a diamond
category.
Proof. We verify the conditions of Proposition 6.17 for the forgetful functor F : Γ-C → C
forgetting the Γ action, and then the lemma will follow. We have that F is faithful and takes
all morphisms to epimorphisms (by definition of the morphisms in Γ-C). We have assumed
56
that C is a diamond category. We can check that the poset of quotients of an object G
with an action of Γ is the subposet of Γ-invariant elements of the poset of quotients of G.
Since the join and meet of Γ-invariant elements is Γ-invariant, the Γ-invariant elements of
the poset of quotients of an object of C is a sublattice as required.
A simple morphism of objects with an action of Γ has an underlying morphism of objects
of dimension at most |Γ|. Indeed, let G → H be a morphism of objects, and suppose we
have compatible actions of Γ on G and H. The induced morphism of objects with an action
of Γ is simple if there is no K ∈ [H, G] that is Γ-invariant other than H and G. Take any
element of [H, G] that is simple over H and take the join of its orbit under Γ. Since this
join depends only on the orbit, it is invariant under Γ, and it cannot be H, so it must be G.
Thus G is the join of at most |Γ| objects simple over H and thus must have dimension at
most |Γ| over H.
Since Γ and the automorphism group of any object of C are finite, there are finitely many
isomorphism classes of objects in Γ-C mapping to an object of C.
Several previous works have considered Γ-groups. If we let C be the category of finite
Γ-groups whose order is relatively prime to |Γ|, Liu, Zureick-Brown and the second author
proved the existence of, and gave explicit formulas for, a measure on the pro-category with
moments [G : GΓ ]−u [LWZ19, Theorem 6.2]. The first author [Saw20, Theorem 1.2] proved
uniqueness and robustness in this setting for moments that are O(|G|n) for some n. The
results of the current paper, along with following the “blueprints” above allow one to give a
different proof of these explicit formulas and uniqueness and robustness results.
Another structure is the addition of a map to a fixed object.
Lemma 6.20. Let C be a diamond category. Let K be an object of C. The category K-Ext
of K-extensions whose objects are objects of C together with an epimorphism to K and where
morphisms are epimorphisms of C respecting the epimorphisms to K is a diamond category.
This construction is a variant of the notion of slice category, restricted to epimorphisms.
Applied to the category of finite groups, this produces the category of groups with a
surjection onto a fixed finite group K. This is closely related to the category of groups with
an action of K, as if G is a group with an action of K then G⋊K is a group with a surjection
onto K.
Proof. We verify the conditions of Proposition 6.17 for the forgetful functor F taking an
object of C together with an epimorphism to K to the underlying object, and then the lemma
will follow. This is faithful and takes all morphisms to epimorphisms (by construction of the
category). We have assumed that C is a diamond category. For G an object and f : G → K
a surjection, the poset of quotients of (G, f ) is simply the interval [K, G], which is indeed
a sublattice. A morphism (G, f ) → (H, g) is simple if and only if G → H is simple, so F
applied to a simple morphism is always simple. There are finitely many isomorphism classes
of pairs (G, f ) for any G because there are finitely many epimorphisms f : G → K. in C
since C being a diamond category implies that quotient lattices and automorphism groups
are finite.
We can also add any kind of finite data that can pushforward along morphisms.
Lemma 6.21. Let C be a diamond category. Let G be a functor from C to the category of
finite sets. Then the category (C, G) of pairs of an object G ∈ C together with an element
57
s ∈ G(G), with morphisms (G, sG ) → (H, sH ) given by morphisms f : G → H in C such that
G(f )(sG ) = sH , is a diamond category.
One example of such a functor on the category of finite groups is provided by the kth
group homology for any fixed k. The k = 3 case was important in [SW22b]. The “bilinearly
enhanced groups" studied by Lipnowski, Tsimerman, and the first author in [LST20] can also
be expressed this way, as the set of bilinearly enhanced group structures on a finite abelian
group is a functor from the category of finite abelian groups to finite sets. Our results then
can be used to give a new proof of [LST20, Theorem 8.17], which proves uniqueness and
robustness for finite abelian ℓ-groups with bilinearly enhanced group structures. It should
also be possible to give a new proof of the existence result and formulas of [LST20, Theorem
8.14] using our main results and following the “blueprints” above. In forthcoming work, the
authors will use the results from this paper in the category of Γ-groups with additional finite
data as in Lemma 6.21 (a class in H 3 (G, Z/n)) to give conjectures about the distribution of
Galois groups of maximal unramified extensions of Γ-number fields, as well as q → ∞ results
in the function field case.
Note that, if C has fiber products, then the category of pairs constructed here need not,
unless F (G ×K H) = F (G) ×F (K) F (H). For example, the category of finite groups together
with a kth homology class does not have fiber products. It is for this and similar reasons
that we avoid using fiber products and use joins instead, even though the familiar examples
like groups, rings, and modules have fiber products.
Proof. We verify the conditions of Proposition 6.17 for the (faithful) forgetful functor F
taking a pair (G, sG ) to G, and then the lemma will follow.
If a morphism f : G → H in C such that G(f )(sG ) = sH is an epimorphism (G, sG ) →
(H, sH ), F is another object of C, and h1 , h2 are two morphisms H → F such that h1 ◦ f =
h2 ◦ f . Then
G(h1 )(sH ) = G(h1 )(G(f )(sG )) = G(h1 ◦f )(sG ) = G(h2 ◦f )(sG ) = G(h2 )(G(f )(sG )) = G(h2 )(sH ).
Let sF = G(h1 )(sH ) = G(h2 )(sH ). Then h1 , h2 are two morphisms (H, sH ) → (F, sF ) whose
compositions with f are equal, so they are equal, thus h1 = h2 . Hence f is an epimorphism
G → H in C.
We have assumed that C is a diamond category. The poset of quotients of (G, sG ) is iden-
H
tical to the poset of quotients of G since each quotient H has a unique element G(πG )(sG ) ∈
H
G(H) compatible with the morphism πG from G. A morphism (G, sG ) → (H, sH ) is simple
if and only if G → H is simple, and the dimension of such morphisms is bounded by one.
Since G is valued in finite sets, there are finitely many isomorphism classes of pairs (G, sG )
for a fixed G.
Moreover,in the situation of Lemma 6.21, we can easily related well-behavedness between
(C, G) and C.
Lemma 6.22. Let C be a diamond category. Let G be a functor from C to the category of
finite sets.
Let (MH )H ∈ R(C,G)/≃ . Then (MH )H is well-behaved (for (C, G)) if and only if ( s∈G(G) M(G,s) )G ∈
P
PThis
∞ Qis a version of the classical moment problem. Note that the factorial
n−1 P∞moments
n
m=0 i=0 (m − i) νm carry the same information as the power moments m=0 m νm
since we can write either in terms of the other using the Stirling numbers.
Proof. Let C be the opposite of the category of finite set. First note that every object of this
category is level-C where C is generated by the one-element set, because joins correspond to
unions of sets and every set is the union of one-element sets.
Since the isomorphism classes of finite sets are parameterized by natural numbers, a mea-
sure on C is the same thing as a measure on N, i.e. a sequence νm of nonnegative real
numbers. We write [n] for the object corresponding to a set with n elements.
The epimorphisms from [m] to [n] are the injections from an n-element
Qn−1 Qn−1 set to an m-
element set. The number of these is i=0 (m − i), so Epi([m], [n]) = i=0 (m − i). Thus the
[n]-moment of the measure νm is given by ∞
P Qn−1
m=0 i=0 (m − i) νm .
60
For an epimorphism π : [n] → [m], the interval [[n], [m]] is the lattice of subsets of an
n-element set containing a fixed m-element set, i.e. the lattice of subsets of an n − m-
element set, which is the product of n − m copies of a two-element lattice. This implies
n
µ([n], [m]) = (−1)n−m , and µ̂([n], [m]) = (−1)n−m m , as well as Z(π) = 2n−m .
Thus ∞ n
(−1)n−m m
X X (−1)n−m Mn
vC,m = Mn = .
n=0
n! n≥m
m!(n − m)!
Since [[n], [m]] is a product of copies of the two-element lattice, we can use a = 1 in the
definition of “behaved.” We have
∞
X X |µ([n], [m])| X X |(−1)m−n | n−m
Z(π)Mn = 2 Mn
| Aut([n])| n=0
n!
[n]∈C π∈Epi([n],[m]) π∈Epi([n],[m])
∞ ∞
X 2n−m X 2n−m Mn
= Mn = m! < ∞,
n=m
(n − m)! n=m
m!(n − m)!
so our moments are behaved. Then, the “if” of the first statement follows from Theorem 4.15.
The final statement of the proposition follows from Theorem 4.2. The “only if” of the first
statement follows from the final statement.
If G is a contravariant functor from finite sets to finite sets, we can apply Lemma 6.21 to
obtain a new category to work in, starting from the opposite category of finite sets.
For example, if G(S) is given by the set of relations on S which are reflexive and symmetric
but not necessarily transitive, then we obtain the opposite of the category of finite simple
graphs. We represent each graph by its set of vertices together with the relation of two
vertices either connected by an edge or identical, and morphisms of graphs G1 → G2 given
by maps from the vertices of G2 to the vertices of G1 that sends every pair of vertices
connected by an edge to either a pair of vertices connected by an edge or the same vertex.
(One could also use the opposite convention where the relation never includes a vertex and
itself.)
Epimorphisms in this category correspond to inclusions of induced subgraphs. We can
thus express MG /|Aut(G)| as the expected number of induced subgraphs isomorphic to G
in a random graph.
A similar notion of moments has been considered in graph theory, with work surveyed in
[LS10]. These works studied the probability that a subgraph induced on a set of vertices
sampled uniformly at random from an infinite graph (a graphon) is isomorphic to a fixed
one. Because these moments are associated to a single graphon rather than a probability
measure on the set of finite graphs, the existence and uniqueness statements obtained for
the moment problem in that context differ from those that can be obtained from Theorem
1.6
There are many models of random graphs, which one could hope to check by our method
are uniquely determined by their moments, and one could possibly even create new models
of random graphs from suitably chosen moments. Note, however, that any model of random
graphs with a fixed number of vertices is automatically determined by its moments, so the
main interest is in models of random graphs with varying number of vertices (perhaps by
first generating the number of vertices using an existing distribution on the natural numbers
such as a Poisson or exponential distribution).
61
For now we give just one example:
Lemma 6.25. For λ a positive real number and α ∈ [0, 1], there exists a unique prob-
ability measure on the set of isomorphism classes of finite graphs such that the expected
number of induced subgraphs isomorphic to a given finite graph G is λ#V (G) α#E(G) (1 −
#V (G)
α)( 2 )−#E(G) /|Aut(G)|, which is an Erdős-Rényi random graph, with edge probability α,
on a number of vertices given by a Poisson distribution with mean λ.
Proof. Let C be the opposite of the category of finite simple graphs as described above.
First, we check that the given random graph has Gth moment (in this category) MG =
#V (G)
λ#V (G) α#E(G)(1−α)( 2 )−#E(G) . If we condition on the number n of vertices of the random
graph, then there are n(n − 1) · · · (n − #V (G) + 1) ways to give an injection from the vertices
#V (G)
of G to the vertices of our random graph, and each of these has α#E(G) (1 − α)( 2 )−#E(G)
probability of realizing G as an induced subgraph. The expectation of n(n − 1) · · · (n −
#V (G) + 1) for n Poisson with mean λ is λ#V (G) . Thus the expected number of inclusion
maps from G to our random graph, as an induced subgraph, (which is the Gth moment), is
#V (G)
λ#V (G) α#E(G)(1 − α)( 2 )−#E(G) . Two different realizations of G as an induced subgraph
of our random graph have the same image, i.e. correspond to the same induced subgraph,
if and only if they differ by an automorphism of G, and so the average mentioned in the
lemma statement is always | Aut(G)|−1 times the Gth moment.
For uniqueness, we apply Theorem 4.2. We have seen above that C is a diamond category,
and we can check that the whole category is generated, as a level C, by a graph on one vertex
and the path on two vertices, as each graph is the join of such graphs. As noted in the proof
of Lemma 6.21, the poset [F, G] is the same as the poset [V (F ), V (G)] in the opposite of
the category of finite sets, so µ(F, G) = (−1)#V (G)−#V (F ) . Let F be a fixed graph with m
vertices and e edges. We now check the hypothesis of Theorem 4.2, and have
X X |µ(F, G)MG |
G∈C i:F →G
|Aut(G)|
inclusion as an
induced subgraph
X X X |µ(F, G)MG |
=
n≥m G on n i:F →G
n!
labelled vertices inclusion as an
induced subgraph
n
X X X λn α#E(G) (1 − α)( 2 )−#E(G)
= .
n≥m G on n i:F →G
n!
labelled vertices inclusion as an
induced subgraph
There are n!/(n − m)! ways to map to vertices of F to n labelled vertices. For a graph on
those n vertices to include F as an induced subgraph according to one of these maps, the
edges between the m vertices corresponding to F are determined, but all other edges can
n
− m
appear or not appear. Thus given one of these n!/(n − m)! maps, there are ( 2 ) e′( 2 ) graphs
G on n vertices with e′ + e edges such that the given map realized F as an induced subgraph.
62
Thus, the above expressions are equal to
X n − m λn αe′+e (1 − α)(n2 )−e′ −e
X n! 2 2
n≥m
(n − m)! ′ e′ n!
e
n X n − m ′
m λ n m
=αe (1 − α)( 2 )−e αe (1 − α)( 2 )−( 2 )−e
′
X
2 2
(n − m)! e′ e′
n≥m
m X λn
=(1 − α)( 2 ) < ∞.
n≥m
(n − m)!
This verifies the hypothesis of Theorem 4.2, and thus we have a unique distribution of finite
graphs with the given moments.
References
[Ale99] A. N. Alekseı̆chuk. Conditions for the uniqueness of the moment problem in the class of $q$-
distributions. Diskretnaya Matematika, 11(4):48–57, 1999.
[AS68] Abraham Adrian Albert and Reuben Sandler. An Introduction to Finite Projective Planes. Holt,
Rinehart and Winston, 1968.
[Awo06] Steve Awodey. Category Theory. Oxford Logic Guides. Oxford University Press, Oxford, 2006.
[BBH17] Nigel Boston, Michael R. Bush, and Farshid Hajir. Heuristics for $p$-class towers of imaginary
quadratic fields. Mathematische Annalen, 368(1-2):633–669, 2017.
[Bir48] Garrett Birkhoff. Lattice Theory, Revised Edition (2nd), volume 25 of AMS Colloquium Publica-
tions. American Mathematical Society, 1948.
[Bir67] Garrett Birkhoff. Lattice Theory. Colloquium Publications (American Mathematical Society) ;
Volume 25. American Mathematical Society, Providence, Rhode Island, third edition. edition,
1967.
[BKL+ 15] Manjul Bhargava, Daniel M. Kane, Hendrik W. Lenstra, Jr., Bjorn Poonen, and Eric Rains.
Modeling the distribution of ranks, Selmer groups, and Shafarevich-Tate groups of elliptic curves.
Cambridge Journal of Mathematics, 3(3):275–321, 2015.
[Bor16] Iurie Boreico. Statistics of Random Integral Matrices. PhD thesis, Stanford University, United
States – California, 2016.
[BV88] Winfried Bruns and Udo Vetter. Determinantal Rings, volume 1327 of Lecture Notes in Mathe-
matics. Springer-Verlag, Berlin, 1988.
[BW17] Nigel Boston and Melanie Matchett Wood. Non-abelian Cohen–Lenstra heuristics over function
fields. Compositio Mathematica, 153(7):1372–1390, July 2017.
[CKL+ 15] Julien Clancy, Nathan Kaplan, Timothy Leake, Sam Payne, and Melanie Matchett Wood. On
a Cohen–Lenstra heuristic for Jacobians of random graphs. Journal of Algebraic Combinatorics,
pages 1–23, May 2015.
[CL84] Henri Cohen and Hendrik W. Lenstra, Jr. Heuristics on class groups of number fields. In Number
Theory, Noordwijkerhout 1983 (Noordwijkerhout, 1983), volume 1068 of Lecture Notes in Math.,
pages 33–62. Springer, Berlin, 1984.
[CM90] Henri Cohen and Jacques Martinet. étude heuristique des groupes de classes des corps de nombres.
Journal für die Reine und Angewandte Mathematik, 404:39–76, 1990.
[Coo00] C. Cooper. On the distribution of rank of a random matrix over a finite field. Random Structures
& Algorithms, 17(3-4):197–212, October 2000.
[EVW16] Jordan S. Ellenberg, Akshay Venkatesh, and Craig Westerland. Homological stability for Hurwitz
spaces and the Cohen-Lenstra conjecture over function fields. Annals of Mathematics. Second
Series, 183(3):729–786, 2016.
[FJ08] Michael D. Fried and Moshe Jarden. Field Arithmetic, volume 11 of Ergebnisse Der Mathe-
matik Und Ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results
63
in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics].
Springer-Verlag, Berlin, third edition, 2008.
[FK06a] Étienne Fouvry and Jürgen Klüners. Cohen–Lenstra Heuristics of Quadratic Number Fields. In
Florian Hess, Sebastian Pauli, and Michael Pohst, editors, Algorithmic Number Theory, number
4076 in Lecture Notes in Computer Science, pages 40–55. Springer Berlin Heidelberg, January
2006.
[FK06b] Étienne Fouvry and Jürgen Klüners. On the 4-rank of class groups of quadratic number fields.
Inventiones mathematicae, 167(3):455–513, November 2006.
[Hea94] D. R. Heath-Brown. The size of Selmer groups for the congruent number problem. II. Inventiones
Mathematicae, 118(2):331–370, 1994.
[KS06] Masaki Kashiwara and Pierre Schapira. Categories and Sheaves. Springer Berlin Heidelberg, 2006.
[Lee22] Jungin Lee. Mixed moments and the joint distribution of random groups, October 2022.
[Liu22] Yuan Liu. Non-abelian Cohen–Lenstra Heuristics in the presence of roots of unity.
(arXiv:2202.09471), February 2022.
[LS10] László Lovász and Balázs Szegedy. The graph theoretic moment problem.
https://arxiv.org/abs/1010.5159, 2010.
[LST20] Michael Lipnowski, Will Sawin, and Jacob Tsimerman. Cohen-Lenstra heuristics and bilinear
pairings in the presence of roots of unity. arXiv:2007.12533 [math], July 2020.
[LW20] Yuan Liu and Melanie Matchett Wood. The free group on n generators modulo n + u random
relations as n goes to infinity. Journal für die reine und angewandte Mathematik, 2020(762):123–
166, May 2020.
[LWZ19] Yuan Liu, Melanie Matchett Wood, and David Zureick-Brown. A predicted distribution for Galois
groups of maximal unramified extensions. arXiv:1907.05002 [math], July 2019.
[Més20] András Mészáros. The distribution of sandpile groups of random regular graphs. Transactions of
the American Mathematical Society, 373(9):6529–6594, September 2020.
[Neu67] Hanna Neumann. Varieties of Groups. Springer Berlin Heidelberg, 1967.
[NVP22] Hoi H. Nguyen and Roger Van Peski. Universality for cokernels of random matrix products,
September 2022.
[NW21] Hoi H. Nguyen and Melanie Matchett Wood. Random integral matrices: Universality of surjec-
tivity and the cokernel. Inventiones mathematicae, October 2021.
[Rot64] Gian Carlo Rota. On the foundations of combinatorial theory I. Theory of Möbius Functions.
Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete, 2(4):340–368, January 1964.
[RZ00] Luis Ribes and Pavel Zalesskii. Profinite Groups. Springer Berlin Heidelberg, 2000.
[Saw20] Will Sawin. Identifying measures on non-abelian groups and modules by their moments via re-
duction to a local problem. arXiv:2006.04934 [math], June 2020.
[Sch17] Konrad Schmüdgen. The Moment Problem, volume 277 of Graduate Texts in Mathematics.
Springer, Cham, 2017.
[Smi76] Jonathan D. H. Smith. Mal’cev Varieties. Lecture Notes in Mathematics, Vol. 554. Springer-
Verlag, Berlin-New York, 1976.
[Sta12] Richard P. Stanley. Enumerative Combinatorics. Volume 1, volume 49 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, Cambridge, second edition, 2012.
[SW22a] Will Sawin and Melanie Matchett Wood. Conjectures for distributions of class groups of G-
extensions of fields with roots of unity. forthcoming, 2022.
[SW22b] Will Sawin and Melanie Matchett Wood. Finite quotients of 3-manifold groups.
https://arxiv.org/abs/2203.01140, 2022.
[Woo17] Melanie Matchett Wood. The distribution of sandpile groups of random graphs. Journal of the
American Mathematical Society, 30(4):915–958, 2017.
[Woo18] Melanie Matchett Wood. Cohen-Lenstra heuristics and local conditions. Research in Number
Theory, 4(4):41, September 2018.
[Woo19] Melanie Matchett Wood. Random integral matrices and the Cohen-Lenstra heuristics. American
Journal of Mathematics, 141(2):383–398, 2019.
[WW21] Weitong Wang and Melanie Matchett Wood. Moments and interpretations of the Co-
hen–Lenstra–Martinet heuristics. Commentarii Mathematici Helvetici, 96(2):339–387, June 2021.
64
Department of Mathematics, Columbia University, 2990 Broadway, New York, NY 10027
USA
Email address: sawin@math.columbia.edu
65