ALL GROUPS ARE SURJUNCTIVE

JAN CANNIZZO

Abstract. We prove that for any finitely generated group G and any k > 1, the space
of k-colorings of G does not admit a strict self-embedding. This settles the Gottschalk
arXiv:1912.00541v1 [math.DS] 2 Dec 2019

surjunctivity conjecture and, consequently, Kaplansky’s direct finiteness conjecture.

1. Introduction

Let G be a finitely generated group and k := {0, . . . , k−1} a collection of k > 1 symbols,
or colors. The shift action of G on kG , the space of k-colorings of G, is a widely studied
system of fundamental importance in symbolic dynamics, ergodic theory, and other
fields, notably the theory of cellular automata. In 1973, Gottschalk [3] posed a natural
question about this system, namely whether it may contain a proper copy of itself,
and conjectured a negative answer. To be more precise, Gottschalk conjectured that
for any group G and any k > 1, there cannot exist a continuous, G-equivariant map
f : kG → kG that is injective but not surjective.
Groups that satisfy Gottschalk’s conjecture came to be known as surjunctive groups,
and there is no shortage of examples. It is obvious that finite groups are surjunctive.
Residually finite groups are also surjunctive, since for such groups periodic k-colorings
(the ones that are constant on the cosets of a finite-index subgroup) are dense in the
space of all k-colorings. Moreover, amenable groups can be shown to be surjunctive
via an argument that exploits topological entropy (see, for instance, [7]). In 1999,
Gromov [4] introduced a class of groups that is a common generalization of the classes
of residually finite and amenable groups and proved that this class of groups, which
was given the name sofic by Weiss [7], is surjunctive as well. As of this writing, it
is unknown whether every group is sofic, a fact that demonstrates that the class of
surjunctive groups is extremely large.
The Curtis-Hedlund-Lyndon theorem (see [5]) asserts that cellular automata on a finitely
generated group G (the most famous example of which is perhaps Conway’s game of life
on Z2 ) correspond precisely to continuous, G-equivariant self-maps of kG , so the question
of whether every group is surjunctive has a direct bearing on the theory of cellular
automata. Moreover, it is known that surjunctive groups satisfy Kaplansky’s direct
finiteness conjecture (see, for instance, [1]), which asks whether, given two elements x
and y that satisfy xy = 1 in the group ring F[G], where F is an arbitrary field, it is true
that yx = 1 as well.
Surjunctive groups are also connected with ideas in ergodic theory. In a recent series of
influential papers, Seward introduced the Rokhlin entropy of the action of a countable
group by measure-preserving transformations on a Lebesgue space (also known as a
standard probability space), and showed in [6] that if the action of G on (kG , µk ),
1
2 JAN CANNIZZO

where µk is the Bernoulli measure determined by the uniform distribution on k, has

Rokhlin entropy log(k) for all k > 1, then G is surjunctive. (Recall that the classical
Kolmogorov-Sinai entropy of the action of Z on (kZ , µk ) is equal to log(k).)
Our aim in this paper is to prove that the class of surjunctive groups is in fact as large
as imaginable. Namely, we prove the following theorem.
Theorem. All groups are surjunctive. That is, for any finitely generated group G and
any k > 1, if a continuous, G-equivariant map f : kG → kG is injective, then f is
surjective.
Our main tool is an object that we call a generalized de Bruijn graph which, as its name
suggests, generalizes classical de Bruijn graphs defined with respect to one-dimensional
strings over k. As we show, these graphs serve as finite models of kG and its subspaces,
and homomorphisms between them serve as finite models of endomorphisms of kG .
This allows us to translate questions about endomorphisms of kG into questions about
homomorphisms between finite structures.
This paper is organized into five sections, including the introduction. In Section 2, we
discuss background knowledge and basic definitions. Section 3 formalizes the notion
of a generalized de Bruijn graph and gives basic examples, definitions, and results. In
Section 4, we show how generalized de Bruijn graphs naturally serve as finite models
of kG and its subspaces, and how homomorphisms between such graphs are naturally
finite models of endomorphisms of kG .
Finally, in Section 5, we establish two lemmas and proceed to the proof of our main
theorem. The key idea behind the proof is to associate an endomorphism φ of a finite
generalized de Bruijn graph to a given injective endomorphism f of kG . The map φ has
the property that if f is not surjective, then φ is not surjective either, which leads to a
contradiction upon considering the iterates of φ.

2. Preliminaries

Let G be a finitely generated group and A a finite set of generators of G. We will

frequently identify G with the Cayley graph determined by A, namely the graph whose
vertex set is G and in which two vertices g, h ∈ G are connected with an edge that is
directed from g to h and labeled with a generator a ∈ A if and only if h = ag. The
Cayley graph structure on G equips the group with a natural metric (which is identical
to the word metric determined by A). Given an integer r > 0, we denote by Gr the
ball of radius r, or r-neighborhood, centered at the group identity e ∈ G with respect
to this metric. To be more precise, we define Gr to be the subgraph of G induced by
those elements whose distance from the identity is less than or equal to r.
Denoting by
k := {0, . . . , k − 1}
a collection of k > 1 symbols, or colors, which we identify with the integers from 0 to
k − 1, consider the space kG of k-colorings of G. This space is endowed with a natural
topology (the product topology), as well as a natural action of G by homeomorphisms
 ALL GROUPS ARE SURJUNCTIVE 3

(the shift). When G is marked with a finite generating set A, the space kG may be
realized as the projective limit
kG = lim kGr ,
←−
where for r > q the connecting maps πr,q : kGr → kGq are projections that send a k-
coloring x : Gr → k to the k-coloring obtained by restricting x to Gq . It is not difficult
to see that kG is a compact, totally disconnected Hausdorff space.
We equip kG with the ultrametric
d(x, y) = exp(−r),
where r is the largest radius such that the k-colorings x and y coincide on the r-
neighborhood Gr (and where we take G∞ = G and exp(−∞) = 0). With respect to the
metric d (which generates the product topology on kG ), the ε-neighborhood Uε (x) of a
k-coloring x ∈ kG is a cylinder set
Uε (x) = {y ∈ kG | π∞,r (x) = π∞,r (y)}
for some r > 0. That is to say, the distance between two k-colorings is small if they
are identical when restricted to Gr for some large r. It follows that cylinder sets, which
serve as basic open sets in kG , may be identified with elements of kGr .
We will be interested in maps from the space kG to itself that respect both the topology
and G-action on kG , i.e. that are continuous and G-equivariant, and we will often refer
to such maps simply as endomorphisms.
By the Curtis-Hedlund-Lyndon theorem, endomorphisms correspond precisely to cellular
automata on the group G. We remind the reader that a cellular automaton on G is a
function, or rule, f : kA → k defined with respect to some finite generating set A ⊆ G.
Given a k-coloring of G, this rule may be used to change the color of each element of
G based on the colors of its neighbors in the Cayley graph determined by A. A cellular
automaton therefore defines a map f : kG → kG which, abusing notation, we also
denote by f . The Curtis-Hedlund-Lyndon theorem asserts that f is in fact continuous
and G-equivariant and, conversely, that every continuous, G-equivariant endomorphism
of kG is a cellular automaton on G for some finite generating set A ⊆ G.
Nearly all of the structures with which we will be working are graphs whose edges are
given a special orientation and labeling. We call these objects A-labeled graphs, and we
define them as follows.

Definition 2.1. Let A be a collection of symbols. An A-labeled graph is a graph ∆,

which may have loops (cycles of length one) and multi-edges (multiple edges that join
the same pair of vertices), such that each edge of ∆ is directed and labeled with an
element of A. We denote the edge set of ∆ by E(∆) and write (x, a, y) ∈ E(∆) if ∆
has an edge directed from the vertex x to the vertex y and labeled with a ∈ A.

We likewise define what it means for a map between A-labeled graphs to be structure-
preserving.

Definition 2.2. A homomorphism of A-labeled graphs ∆ and ∆′ is a map φ : ∆ → ∆′

with the property that if (x, a, y) ∈ E(∆), then (φ(x), a, φ(y)) ∈ E(∆′ ). If φ is bijective
4 JAN CANNIZZO

and the inverse φ−1 : ∆′ → ∆ is also a homomorphism of A-labeled graphs, then φ is

an isomorphism of A-labeled graphs.

Remark 2.3. When referring to an A-labeled graph, especially an r-neighborhood Gr ,

we are at times only interested in its isomorphism class.

The group G itself, when marked with a finite generating set A, inherits the structure
of an A-labeled graph when viewed as a Cayley graph, and to avoid having to introduce
and switch between different notation for what we view as essentially the same object,
it will be our standing assumption that G represents both a group and an A-labeled
graph. Recall, furthermore, that whenever G acts on a set X, it is possible to endow X
with a Schreier graph structure for which (x, a, y) ∈ E(X) if and only if y = ax. In this
way, the space kG naturally has the structure of an A-labeled graph as well. In fact,
to say that a map f : kG → kG is G-equivariant is to say that f is a homomorphism of
A-labeled graphs.

3. Generalized de Bruijn graphs

In this section, we introduce the central objects with which we will be working. These
objects, which we call generalized de Bruijn graphs, are inspired by the classical and
well-studied graphs that are named after de Bruijn and were independently discovered
by various authors (see [2]). Intuitively speaking, generalized de Bruijn graphs are A-
labeled graphs that locally model the space kG , not unlike how a Riemannian manifold
locally models Rn .
Recall that a classical de Bruijn graph is a graph whose vertex set is a collection of strings
x1 . . . xr ∈ kr of length r > 1 and in which two strings x = x1 . . . xr and y = y1 . . . yr
are joined with an edge directed from x to y if there exists a string z1 . . . zr+1 ∈ kr+1
such that x = z1 . . . zr and y = z2 . . . zr+1 (similar constructions, e.g. where x and y are
allowed to have smaller overlap, are possible as well). From our point of view, such a
graph is a de Bruijn graph for the infinite cyclic group Z equipped with the standard
generating set A = {1}. We would like to define the notion of a de Bruijn graph for an
arbitrary finitely generated group G equipped with an arbitrary finite generating set A.
We begin by defining the edge relation of such a graph.

Definition 3.1. Let x, y ∈ kGr be k-colorings of Gr and a ∈ A a generator. We say

that it is possible to move from x to y via a if there exists a k-coloring of G whose
restriction to the r-neighborhood of the identity e ∈ G is isomorphic to x and whose
restriction to the r-neighborhood of a ∈ G is isomorphic to y.

We now come to the main definition of this section. Here, as elsewhere, G is a finitely
generated group with finite generating set A, and we identify G with the Cayley graph
determined by A.

Definition 3.2. A generalized de Bruijn graph for (kG , A, r) is a pair (∆, α), where ∆
is an A-labeled graph and α is an atlas for ∆, namely a map α : ∆ → kGr with the
property that for every edge (x, a, y) ∈ E(∆), it is possible to move from α(x) to α(y)
via the generator a ∈ A.
 ALL GROUPS ARE SURJUNCTIVE 5

Remark 3.3. For the sake of brevity, we may omit mention of the atlas of a generalized
de Bruijn graph in our notation if doing so is unlikely to cause confusion.

We view the atlas of a generalized de Bruijn graph (∆, α) as loosely analogous to the
atlas of coordinate charts of a Riemannian manifold. Just as the atlas of a Riemannian
manifold M shows how the manifold locally models Euclidean space by associating
neighborhoods in Rn to points in M, the atlas of a generalized de Bruijn graph shows
how the graph locally models kG by associating neighborhoods in kG (that is, cylinder
sets in kGr ) to points in ∆. Moreover, if two points x, y ∈ M are sufficiently close to
one another, then it is possible to transition between neighborhoods in Rn associated to
x and y, just as it is possible to move between the k-colored r-neighborhoods associated
to two points x, y ∈ ∆ if x and y are sufficiently close to one another. We hasten to
add, however, that the r-neighborhood of a point x ∈ ∆ within the graph ∆ is by no
means guaranteed to be isomorphic to α(x) ∈ kGr . In this regard, a generalized de
Bruijn graph is more analogous to an immersed submanifold of Rn .
The following are examples of generalized de Bruijn graphs.
i. An arbitrary A-labeled graph ∆ may be realized as a generalized de Bruijn graph
for (kG , A, r) for any r > 0 by equipping it with a constant atlas that associates
to every vertex of ∆ a fixed monochromatic coloring of Gr . A priori, such graphs
are of course not very interesting.
ii. A k-coloring of G itself is an example of a generalized de Bruijn graph for
(kG , A, r) for any r > 0. Here the atlas assigns to each g ∈ G the k-coloring of
the r-neighborhood of g.
iii. Generalizing the previous example, if S ⊆ G is a subgraph of G, then any k-
coloring x : G → k realizes S as a generalized de Bruijn graph for (kG , A, r)
for any r > 0. Here the atlas assigns to each g ∈ S the restriction of x to the
r-neighborhood of g within G.
iv. Suppose H is a quotient of G with the property that the quotient map φ : G → H
is injective when restricted to Gr . When G and H are treated as A-labeled
graphs, φ is a covering map, and any k-coloring of H naturally realizes H as a
generalized de Bruijn graph for (kG , A, r), as the r-neighborhood of each h ∈ H
is isomorphic to Gr .
v. If ∆ is a sofic approximation of G, namely an A-labeled graph with the property
that for some (large) r > 0 and some (small) ε > 0,
|{x ∈ ∆ | Br (x) ∼
= Gr }|
> 1 − ε,
|∆|
where Br (x) is the r-neighborhood of x ∈ ∆, then any k-coloring of the subgraph
∆′ ⊆ ∆ induced by the set {x ∈ ∆ | Br (x) ∼= Gr } realizes ∆′ as a generalized de
G
Bruijn graph for (k , A, r).
vi. When endowed with the Schreier graph structure induced by the action of G,
whose connected components are the orbits of the action, the space kG , and in
fact any subspace of kG , is realized as a generalized de Bruijn graph for (kG , A, r)
for any r > 0.
6 JAN CANNIZZO

vii. Given a nonempty subset S ⊆ kGr , denote by ∆(S) the graph whose vertex set
is S and in which two vertices x, y ∈ S are joined with an edge (x, a, y) if and
only if it is possible to move from x to y via a. Then ∆(S) is a generalized de
Bruijn graph for (kG , A, r) when equipped with the atlas defined by α(x) = x
for all x ∈ S.
As Examples iii. and vi. already suggest, any subgraph of a generalized de Bruijn graph
(∆, α) is also a generalized de Bruijn graph, as one need only restrict the atlas α to the
subgraph. Note that if q 6 r, then a generalized de Bruijn graph (∆, α) for (kG , A, r)
may be regarded as a generalized de Bruijn graph for (G, A, q) as well, as the k-colorings
in the image of the atlas α : ∆ → kGr may be restricted to Gq . Abusing notation, we
may also denote such a restricted atlas by α. In case q = 0, the restricted atlas is simply
a k-coloring of ∆, so α endows ∆ with a k-coloring in a canonical way.
In the language of category theory, atlases are contravariant, in the sense that if (∆, α)
is a generalized de Bruijn graph for (kG , A, r) and φ : ∆′ → ∆ is a homomorphism of
A-labeled graphs, then the atlas α may be pulled back in a natural way to determine
an atlas for ∆′ that realizes ∆′ as a generalized de Bruijn graph for (kG , A, r). We are
interested in the conditions under which atlases are covariant, which leads us to the
following definition.

Definition 3.4. Let (∆, α) be a generalized de Bruijn graph for (kG , A, r), and let
φ : ∆ → ∆′ be a homomorphism of A-labeled graphs. We say that φ is q-compatible
with the atlas α if for some q 6 r it is possible to construct a commutative diagram as
follows:
∆ α kGr
φ πr,q

φ∗ α
∆′ kGq

In other words, denoting by A := Im(φ) the image of φ, we say that φ is q-compatible

with α if it is possible to define an atlas φ∗ α : A → kGq by putting
(φ∗ α)(x) = πr,q (α(y)),
where y ∈ φ−1 (x) is any element in the fiber over x and where πr,q : kGr → kGq is the
natural projection.
We will be concerned primarily with the generalized de Bruijn graphs ∆(S) defined in
Example vii. above. The case when S = kGr is the most important, and to simplify
notation we set
∆r := ∆(kGr ). (3.1)
Once again, ∆r is the A-labeled graph whose vertex set is identified with kGr and for
which (x, a, y) ∈ E(∆r ) if and only if it is possible to move from x to y via the generator
a ∈ A. The following proposition asserts that, in the category of generalized de Bruijn
graphs for (kG , A, r), the graph ∆r acts as a terminal object, as every other object
admits a canonical homomorphism into it.

Proposition 3.5. If (∆, α) is a generalized de Bruijn graph for (kG , A, r), then there
exists a canonical r-compatible homomorphism φ : ∆ → ∆r .
 ALL GROUPS ARE SURJUNCTIVE 7

Proof. Consider the map φ : ∆ → ∆r defined by setting φ(x) = y if α(x) = α′ (y),

where α′ is the atlas of ∆r . Since α′ : ∆r → kGr is bijective, φ is well defined. Let
(x, a, y) ∈ E(∆) be an edge. Then by definition it is possible to move from α(x) to α(y)
via a, so by construction it is possible to move α′ (φ(x)) to α′ (φ(y)) via a. Therefore, φ
is a homomorphism of A-labeled graphs. It is clear that φ is r-compatible. 

Note that, in light of Proposition 3.5 and the fact that atlases are contravariant, we
could equivalently define a generalized de Bruijn graph for (kG , A, r) to be an A-labeled
graph that admits a homomorphism into ∆r .

4. Finite models of endomorphisms of kG

Given a generalized de Bruijn graph (∆, α), consider the set Hom(G, (∆, α)) of all
homomorphisms of A-labeled graphs ι : G → ∆. Intuitively, we think of such maps as
immersions of G into ∆. For each immersion ι, the atlas α determines a k-coloring of
G, so there is a natural map from Hom(G, (∆, α)) to kG . In this way, generalized de
Bruijn graphs correspond to subspaces of kG . In fact it is not difficult to see that if ∆
is finite, then any such subspace (which may be empty) is closed and G-invariant.
The following proposition shows that the generalized de Bruijn graphs ∆r defined by
(3.1) correspond to the space kG itself.

Proposition 4.1. There exists a natural bijection β : kG → Hom(G, ∆r ) for any r > 0.

Proof. Let x : G → k be a k-coloring of G. This k-coloring determines a natural atlas

α′ : G → kGr defined by taking α′ (g) to be the restriction of x to the r-neighborhood
of g ∈ G. It follows that (G, α′ ) is a generalized de Bruijn graph for (kG , A, r), so by
Proposition 3.5 we obtain a canonical homomorphism ι : G → ∆r .
On the other hand, if ι : G → ∆r is a homomorphism, then the atlas α of ∆r can be
pulled back to define an atlas α′ for G by setting α′ (g) = α(ι(g)). The atlas α′ , in turn,
endows G with a canonical k-coloring. It is not difficult to see that the associations just
described are inverses of one another and therefore define the desired bijection β. 

Remark 4.2. To simplify notation, and because it is unlikely to cause confusion, we

will denote the bijection of Proposition 4.1 by β regardless of the value of r > 0.

Now suppose φ : ∆ → ∆′ is a homomorphism of A-labeled graphs and the set Hom(G, ∆)

is nonempty. Then φ naturally induces a map
φ∗ : Hom(G, ∆) → Hom(G, ∆′ ) (4.1)
given by ι 7→ φ ◦ ι. In case φ is a homomorphism φ : ∆r → ∆q , then via the bijection
β of Proposition 4.1, φ∗ corresponds to a continuous, G-equivariant map f : kG → kG .
Here we think of ∆r (and ∆q ) as a finite model of kG , and of φ as a finite model of f .
More generally, if A ⊆ kG is a closed, G-invariant subspace, then we would like to define
the notion of a finite model of A and a finite model of a continuous, G-equivariant map
f : A → kG .
8 JAN CANNIZZO

To this end, we will work with the natural projections π∞,r : kG → ∆r and treat each of
them as a homomorphism of A-labeled graphs, where kG is endowed with its Schreier
graph structure. Given a subspace A ⊆ kG , we put
Ar := π∞,r (A)
and emphasize that Ar is not merely collection of points or the subgraph of ∆r induced
by the image of the set A. Rather, we have (x, a, y) ∈ E(Ar ) if and only if there exist
elements x′ , y ′ ∈ kG such that π∞,r (x′ ) = x, π∞,r (y ′ ) = y, and (x′ , a, y ′) ∈ E(kG ).

Definition 4.3. Let A be a closed, G-invariant subspace of kG . By a model of A,

we mean any projection Ar := π∞,r (A) of A to ∆r . If f : A → kG is a continuous,
G-equivariant map, then we say that a homomorphism φ : Ar → ∆q is a model of f if
the diagram
f
A kG
π∞,r π∞,q (4.2)
φ
Ar ∆q
commutes.

The next proposition shows that models always exist.

Proposition 4.4. Every continuous, G-equivariant map f : A → kG , where A is a

closed, G-invariant subspace of kG , has a model.

Proof. Since A is compact, f is uniformly continuous. Let ε, δ ∈ (0, 1) be such that

whenever two points x, y ∈ A satisfy d(x, y) < δ, we have d(f (x), f (y)) < ε, and choose
r > q such that δ < exp(−r) and ε < exp(−q). Since f sends exp(−r)-neighborhoods in
A into exp(−q)-neighborhoods in kG , and since the elements of Ar and ∆q correspond,
respectively, to precisely such neighborhoods, we obtain a map φ : Ar → ∆q that makes
the diagram (4.2) commute.
To see that φ is a homomorphism of A-labeled graphs, endow A and kG with the Schreier
graph structure determined by the action of G, and let (x, a, y) ∈ E(Ar ) be an edge.
Since Ar is the projection of A to ∆r , there exists an edge (x′ , a, y ′) ∈ E(A) such that
π∞,r (x′ ) = x and π∞,r (y ′) = y. Since f is G-equivariant, (f (x′ ), a, f (y ′)) ∈ E(kG ), and
(π∞,q (f (x′ )), a, π∞,q (f (y ′))) = (φ(x), a, φ(y)) ∈ E(∆q ). 

Models are by no means unique. Given a model φ : Ar → ∆q of f : A → kG , it is always

possible to find another model φ′ : As → ∆q of f for any s > r. This corresponds to
decreasing the value of δ > 0. To be more concrete, composing φ with the restricted
projection πs,r |As : As → Ar (which is a model of the identity Id : A → A) yields such a
model. On the other hand, it is also always possible to find another model φ′ : Ar → ∆p
of f for any 0 6 p 6 q. This corresponds to increasing the value of ε > 0, and here
composing φ with the projection πq,p : ∆q → ∆p will yield the desired model. It is also
possible to increase q (and hence r), as this corresponds to choosing a smaller value of
ε > 0, for which there exists a smaller value of δ > 0 and thus a larger r.
 ALL GROUPS ARE SURJUNCTIVE 9

Consider next the inverse problem of taking a given homomorphism φ of generalized de

Bruijn graphs and determining the endomorphism of kG that φ models. In principle,
the homomorphism φ may model many different maps f . The function φ∗ defined by
(4.1) corresponds to the maximal such map (that is, the one whose domain is maximal).
Indeed, we may extend the diagram (4.2) as follows:
φ∗
Hom(G, Ar ) Hom(G, ∆q )
β|A β

A
f
kG (4.3)
π∞,r π∞,q
φ
Ar ∆q
Here β is the bijection of Proposition 4.1 and β|A is its restriction to A. The above
considerations lead us to the definition of what we call a full model.

Definition 4.5. Let A be a closed, G-invariant subspace of kG . We say that Ar is a

full model of A if the restriction β|A : A → Hom(G, Ar ) is a bijection. We say that
φ : Ar → ∆q is a full model of f if Ar is a full model of A and the induced map φ∗ is
equivalent to f , in the sense that the diagram (4.3) commutes.

Generally speaking, if (∆, α) is a generalized de Bruijn graph for (kG , A, r), then we
may write
Hom(G, (∆, α)) ∼
=A
to indicate that there is a bijective correspondence between immersions ι : G → (∆, α)
and k-colorings in the subspace A ⊆ kG . If Ar is a full model of A, we might therefore
write Hom(G, Ar ) ∼= A.

5. Main result

Before coming to our main theorem, we will establish two lemmas needed for its proof.
The proofs of both lemmas exploit the same idea, namely composing a model of an
injective endomorphism with a model of its inverse. The first lemma asserts that the
image of an injective endomorphism of kG has a full model.

Lemma 5.1. If a continuous, G-equivariant map f : kG → kG is injective with image

A := Im(f ), then there exists an r > 0 such that Ar is a full model of A.

Proof. Choose r > q > p large enough so that there exists a model ψ : Ar → ∆q of
f −1 : A → kG and a model φ : ∆q → ∆p of f . Note that, by the closed map lemma,
the inverse f −1 is continuous, and since A is closed and G-invariant, Proposition 4.4
ensures that the model ψ exists. We thus consider the diagram
ψ φ
Ar ∆q ∆p .
We claim that Ar is a full model of A. If this were not the case, then there would exist
a homomorphism ι : G → Ar that does not belong to the image of β|A (see the diagram
10 JAN CANNIZZO

(4.3)). Since f −1 is surjective, there exists a homomorphism ι′ : G → Ar that does

belong to the image of β|A and that satisfies ψ ◦ ι = ψ ◦ ι′ , which implies that the map
(φ ◦ ψ)∗ : Hom(G, Ar ) → Hom(G, ∆p )
is not injective. The composition φ ◦ ψ, however, is the restriction of the natural
∗
projection πr,p to Ar , and πr,p is equivalent to the identity Id : kG → kG and hence
injective. This is a contradiction. 

We also record the following corollary, which asserts that as soon as the subspace A of
Lemma 5.1 has a full model Ar , we are free to increase r.

Corollary 5.2. If A is the image of a continuous, G-equivariant map f : kG → kG and

Ar is a full model of A, then As is a full model of A for any s > r.

Proof. Consider the map ψ : Ar → ∆q from Lemma 5.1, and let ψ ′ := ψ ◦ πs,r |As be
its composition with the restricted projection πs,r |As : As → Ar . Then ψ ′ is a model of
f −1 , and repeating the proof of Lemma 5.1 with respect to the diagram
ψ′ φ
As ∆q ∆p
shows that As is a full model of A. 

Our second lemma asserts that any bijective map from a closed, G-invariant subspace
of kG onto kG has a model that satisfies a general compatibility condition (cf. Defini-
tion 3.4).

Lemma 5.3. Let A be a closed, G-invariant subspace of kG . If a continuous, G-

equivariant map f : A → kG is bijective, then for any p > 0, there exists a model
φ : Ar → ∆p of f that is p-compatible.

Proof. Choose r > q > p large enough so that there exists a model φ : Ar → ∆q of f
and a model ψ : ∆q → ∆p of f −1 . We may thus consider the diagram
φ ψ
Ar ∆q ∆p .

Consider the atlas α : Ar → kGp on Ar , and suppose that φ is not p-compatible.

Then there exists an x ∈ ∆q such that for distinct elements y, z ∈ φ−1 (x), we have
α(y) 6= α(z). Since φ(y) = φ(z), it follows that (ψ ◦ φ)(y) = (ψ ◦ φ)(z). However, the
composition ψ ◦ φ is the restriction of the natural projection πr,p to Ar , so it must be
that πr,p (y) 6= πr,p (z). This is a contradiction, so φ is p-compatible. The composition
φ′ := πq,p ◦ φ is in turn a p-compatible model φ′ : Ar → ∆p of f . 

We are now ready to prove our main theorem. The key idea behind the proof is to
show that, associated with any injective endomorphism f : kG → kG , there exists an
endomorphism φ : ∆q → ∆q (for some q) such that φ∗ is injective. The map φ is not a
model of f , but it can be chosen in such a way that if f is not surjective, then φ is not
surjective either, which leads to a contradiction given that φ can be iterated over ∆q
and ∆q is a finite structure.
 ALL GROUPS ARE SURJUNCTIVE 11

Theorem 5.4. Let G be a finitely generated group and f : kG → kG a continuous,

G-equivariant map. If f is injective, then f is surjective.

Proof. Suppose f is not surjective, and put A := Im(f ). Since the subspace A is closed,
the complement kG \A is open and contains a ball of radius exp(−p) for some p. By
Lemma 5.1 and Corollary 5.2, there exists a q > p such that Aq is a full model of A.
By Corollary 5.2 and Lemma 5.3, there exists an r > q such that ψ : Ar → ∆q is a full
model of f −1 : A → kG that is q-compatible.
Consider ∆q equipped with the atlas ψ∗ α : ∆q → kGq . We claim that, when ∆q is
endowed with the k-coloring induced by ψ∗ α,
Hom(G, (∆q , ψ∗ α)) ∼
= A. (5.1)
To see this, let ι : G → Ar be a homomorphism. Since Ar is a full model of A, the
map ι corresponds to an element of A, and since ψ is q-compatible, the homomorphism
ψ ◦ ι : G → (∆q , ψ∗ α) corresponds to precisely the same element of A. The fact that ψ
is a full model of f −1 means that ψ ∗ : Hom(G, Ar ) → Hom(G, (∆q , ψ∗ α)) is a bijection,
which establishes the claim.
By Proposition 3.5, there is a canonical q-compatible map φ : (∆q , ψ∗ α) → ∆q . Since
there exists a ball of radius exp(−p), and hence a ball of radius exp(−q), not contained
in A, the atlas ψ∗ α : ∆q → kGq is not surjective, and therefore neither is φ. Moreover,
φ∗ : Hom(G, (∆q , ψ∗ α)) → Hom(G, ∆q ) is injective. Indeed, the correspondence (5.1)
and the fact that φ is q-compatible imply that Im(φ) = Aq and that φ∗ sends distinct
elements of Hom(G, (∆q , ψ∗ α)) to distinct elements of Hom(G, ∆q ). It follows that there
exists a self-map φ : ∆q → ∆q which is not surjective yet which is a full model of an
injective endomorphism of kG .
Consider the sequence of iterates {φn }n>1 . Since ∆q is finite, this sequence is eventually
periodic. Indeed, there exists a proper subgraph ∆′ ⊂ ∆q such that Im(φn ) = ∆′ for all
sufficiently large n. It follows that there exists an N > 1 such that the restriction of the
N-fold composition φN to ∆′ is the identity Id : ∆′ → ∆′ (that is, φN is a retraction of
∆q onto ∆′ ).
Now choose a homomorphism ι : G → ∆q such that ι(g) ∈ ∆q \∆′ for some g ∈ G. Then
ι′ := φN ◦ ι is a homomorphism of G into ∆′ . We thus find that
(φN )∗ : Hom(G, ∆q ) → Hom(G, ∆q )
is not injective, since φN ◦ ι = φN ◦ ι′ . This is a contradiction, since it is straightforward
to see that (φN )∗ = (φ∗ )N , so the fact that φ∗ is injective implies that (φN )∗ is too. This
completes the proof of the theorem. 

References
[1] V. Capraro and M. Lupini, with an appendix by V. G. Pestov, Introduction to sofic and hyperlinear
groups and Connes embedding conjecture, Lecture Notes in Mathematics, Springer (2015).

[2] N. G. de Bruijn, Acknowledgement of priority to C. Flye Sainte-Marie on the counting of cir-

cular arrangements of 2n zeros and ones that show each n-letter word exactly once, TH report,
Technische Hogeschool Eindhoven Nederland (1975).
12 JAN CANNIZZO

[3] W. Gottschalk, Some general dynamical notions, Recent Advances in Topological Dynamics, Lec-
ture Notes in Mathematics, Springer (1973), p. 120-125.

[4] M. Gromov, Endomorphisms of symbolic algebraic varieties, Journal of the European Mathemat-
ical Society 1 (2) (1999), p. 109-197.

[5] G. A. Hedlund, Endomorphisms and automorphisms of the shift dynamical systems, Mathematical
System Theory, 3 (4) (1969), p. 320-375.

[6] B. Seward, Krieger’s finite generator theorem for actions of countable groups II, Journal of Modern
Dynamics 15 (2019), p. 1-39.

[7] B. Weiss, Sofic groups and dynamical systems, Indian Journal of Statistics, Series A 62 (3) (2000),
p. 350-359.

Department of Mathematical Sciences

Stevens Institute of Technology
1 Castle Point Terrace
Hoboken, NJ 07030
United States

E-mail address: [email protected]