Additive transversality of fractal sets in the reals

and the integers

Daniel Glasscock Joel Moreira Florian K. Richter

July 13, 2020
arXiv:2007.05480v1 [math.NT] 10 Jul 2020

Abstract
By juxtaposing ideas from fractal geometry and dynamical systems, Hillel Furstenberg
proposed a series of conjectures in the late 1960’s that explore the relationship between
digit expansions with respect to multiplicatively independent bases. In this work, we
introduce analogues of some of the notions and results surrounding Furstenberg’s work in
the discrete setting of the integers. In particular, we define a new class of fractal sets of
integers that parallels the notion of ×r-invariant sets on the 1-torus, and we investigate the
additive independence between these fractal sets when they are structured with respect
to different bases. We obtain
• an integer analogue of a result of Furstenberg regarding the classification of all sets
that are simultaneously ×2 and ×3 invariant (see Theorem B);
• an integer analogue of a result of Lindenstrauss-Meiri-Peres on iterated sumsets of
×r-invariant sets (see Theorem C);
• an integer analogue of Hochman and Shmerkin’s solution to Furstenberg’s sumset
conjecture regarding the dimension of the sumset X + Y of a ×r-invariant set X
and a ×s-invariant set Y (see Theorem D).
To obtain the latter, we provide a quantitative strengthening of a theorem of Hochman
and Shmerkin which provides a lower bound on the dimension of λX + ηY uniformly
in the scaling-parameters λ and η at every finite scale (see Theorem A). Our methods
yield a new combinatorial proof of the theorem of Hochman and Shmerkin that avoids
the machinery of local entropy averages and CP-processes.

Contents
1 Introduction 2
1.1 History and context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.1 Sums of multiplicatively invariant subsets of the reals . . . . . . . . . . . . . . . . 6
1.2.2 Multiplicatively invariant subsets of the non-negative integers . . . . . . . . . . . . 7
1.3 Overview of the paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Preliminary definitions and results 10
2.1 Continuous and discrete fractal geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Multiplicatively invariant subsets of the reals and their finite approximations . . . . . . . . 13
2.3 Dimension of subsets of integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3 Sums of multiplicatively invariant subsets of the reals 18
3.1 A discrete Marstrand projection theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Trees and a regularity result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 A quantitative equidistribution lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Outline of the proof of Theorem A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Proof of Theorem A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

1
 3.6 Proof of Claim 3.14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4 Multiplicatively invariant subsets of the non-negative integers 33
4.1 Connections to symbolic dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Connections to fractal subsets of the reals . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Proof of Theorem B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.4 Proof of Theorem C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.5 Proofs of Theorems D and E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.6 An example that shows R-invariance does not suffice . . . . . . . . . . . . . . . . . . . . . 43
5 Open directions 45
5.1 Positive density for sumsets of full dimension . . . . . . . . . . . . . . . . .. . . . . . . . 45
5.2 Difference sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . 46
5.3 Analogous results for the counting dimension . . . . . . . . . . . . . . . . .. . . . . . . . 46
5.4 Polynomial functions of multiplicatively invariant sets . . . . . . . . . . . .. . . . . . . . 47
5.5 Geometric transversality in the integers . . . . . . . . . . . . . . . . . . . .. . . . . . . . 47
5.6 Multiplicatively invariant sets in relation to other arithmetic sets in the integers . . . . . . 47
5.7 Transversality of multiplicatively invariant sets in the rs-adics . . . . . . . . . . . . . . . . 48

1. Introduction
The purpose of this paper is to investigate the additive independence between fractal sets that
are structured with respect to multiplicatively independent bases. We explore this topic in two
different regimes: the unit interval [0, 1] and the non-negative integers N0 = {0, 1, 2, 3, . . .}.
We will begin by explaining the precedent for this inquiry at the intersection of combinatorial
number theory, fractal geometry, and ergodic theory.

1.1. History and context

Number theorists in the first half of the 20th century were among the first to consider the
degree to which base 2 and base 3 representations of real numbers are independent: an open
conjecture attributed to Mahler [MF] postulates, for example, P∞ that −n if (an )∞P ⊆ {0, −n
n=1 1} is
not eventually periodic, then at least one of the numbers n=1 an 2 and ∞ a
n=1 n 3 is
transcendental. Cassels [Cas], answering a question of Steinhaus about Cantor’s middle thirds
set C, proved that almost every number in C/2 (with respect to the log 2/ log 3-dimensional
Hausdorff measure) is normal to every base which is not a power of 3.
In the language of fractal geometry and dynamical systems, Furstenberg [Fur1, Fur2]
established a number of conjectures and results that explore the relationship between mul-
tiplicative structures with respect to different bases. The notion of structure particularly
relevant to this work is that of multiplicative invariance: a set X ⊆ [0, 1] is ×r-invariant if it
is closed and Tr X ⊆ X, where Tr : [0, 1] → [0, 1] denotes the map
Tr : x 7→ rx mod 1.
One of Furstenberg’s first and most well-known results concerning multiplicatively invari-
ant sets is the following theorem, the measure-theoretic analogue of which is the infamous
×2, ×3 conjecture, a central open problem in ergodic theory.

Theorem 1.1 ([Fur1, Theorem 4.2]). If X ⊆ [0, 1] is simultaneously ×2- and ×3-invariant
then either X is finite or X = [0, 1].

The numbers 2 and 3 in Theorem 1.1 can be replaced by any pair of multiplicatively
independent positive integers, i.e., integers r, s ∈ N for which log r/ log s ∈
/ Q. Furstenberg
conjectured that if r and s are multiplicatively independent and X, Y ⊆ [0, 1] are ×r- and
×s-invariant, respectively, then X and Y are transverse in a sense made precise below. While
some of Furstenberg’s “transversality conjectures” remain open, two of them were resolved
recently by Hochman and Shmerkin [HS], Shmerkin [Shm], and Wu [Wu]. The transversality
conjecture resolved by Hochman and Shmerkin is of particular relevance to this work, so we
will expound on it further now.
In Euclidean geometry, linear subspaces U, V ⊆ Rd are said to be in general position (or
transverse) if


dim(U + V ) = min dim U + dim V, d . (1.1)
In analogy, Furstenberg conjectured that if r and s are multiplicatively independent and X
and Y are ×r- and ×s-invariant subsets of [0, 1], then


dimH (X + Y ) = min dimH X + dimH Y, 1 , (1.2)
where dimH denotes the Hausdorff dimension. Besides the obvious geometric analogy between
(1.1) and (1.2), this expression gives meaning to the statement that the sets X and Y are
additively combinatorially independent. Indeed, the size of the sumset X +Y is a rudimentary
measure of the additive structure shared between X and Y . If X and Y are finite sets of real
numbers, then it is easy to check that
|X| + |Y | − 1 6 |X + Y | 6 |X||Y |. (1.3)
Equality holds on the left if and only if X and Y are arithmetic progressions of the same
step size. When |X + Y | is near this lower bound, inverse theorems in combinatorial number
theory provide additive structural information on the sets X and Y . At the other end of the
spectrum, equality holds on the right in (1.3) if and only if none of the sums x + y, with
x ∈ X and y ∈ Y , coincide. In this case, the sets X and Y lie in general position from an
additive combinatorial point of view.
When X and Y are (infinite) subsets of [0, 1], their sizes and the size of the sumset X + Y
can be measured by the Hausdorff dimension. If X and Y are such that dimH (X × Y ) =
dimH X + dimH Y , the analogues of the inequalities in (1.3) are




max dimH X, dimH Y 6 dimH X + Y 6 min dimH X + dimH Y, 1 .
Equality in the lower bound happens when there are significantly many coincidences amongst
the sums x + y, an indication that X and Y share mutual additive structures. Equality in the
upper bound happens when the sums x+ y are mostly as unique as they can be, an indication
that X and Y are additive combinatorially transverse. It is precisely this equality that was
conjectured by Furstenberg to hold in (1.2) in the case that X and Y are multiplicatively
structured with respect to multiplicatively independent bases.
With no structural assumptions on the sets X, Y ⊆ [0, 1], it is not difficult to find examples
for which the equality in (1.2) does not hold. Nevertheless, it is a consequence of Marstrand’s
projection theorem that for all Borel sets X and Y , the typical dilated sets λX and ηY are
additive combinatorially transverse in the sense of (1.2).

3
Theorem 1.2 ([Mar, Theorem II]1 ). Let X and Y be Borel subsets of [0, 1]. For Lebesgue-
a.e. (λ, η) ∈ R2 ,



dimH λX + ηY = min dimH (X × Y ), 1 . (1.4)

In this context, Furstenberg’s conjecture in (1.2) says that the multiplicative structure
of the sets X and Y can be leveraged to change the result in Marstrand’s theorem from
concerning the typical sumset λX + ηY to concerning the specific one X + Y . In fact,
Furstenberg conjectured that for ×r- and ×s-invariant sets X and Y , the equality in (1.4)
holds for all non-zero λ and η. Hochman and Shmerkin resolved this conjecture by proving
a stronger result for multiplicatively invariant measures; the following theorem is a corollary
of their main result, [HS, Theorem 1.3].

Theorem 1.3 ([HS]). Let r and s be multiplicatively independent positive integers, and let
X, Y ⊆ [0, 1] be ×r- and ×s-invariant sets, respectively. For all λ, η ∈ R\{0},



dimH λX + ηY = min dimH X + dimH Y, 1 . (1.5)

A number of partial results preceded Theorem 1.3 both for multiplicatively invariant
sets and for attractors of iterated function systems (IFSs). Carlos Moreira [Mor] considered
sumsets of attractors of IFSs with certain irrationality and non-linearity conditions. Peres
and Shmerkin [PS] proved (1.5) for attractors of IFSs with rationally independent contraction
ratios; this resolved Theorem 1.3 in the special case that X and Y are restricted digit Cantor
sets with respect to multiplicatively independent bases. This work of Peres and Shmerkin
is particularly relevant to the arguments in this paper, as we will explain further in the
next section and in Section 3. Hochman and Shmerkin [HS] developed Furstenberg’s CP
processes [Fur2] and introduced local entropy averages to prove (1.5) both for invariant sets
and measures and for attractors of IFSs satisfying some general minimality conditions.
In an effort to better understand the role that the multiplicative independence between the
bases plays in Theorem 1.3, it is natural to ask about the sum of sets that are all structured
with respect to the same base r. Taking X ⊆ [0, 1] to be those numbers that can be written
in decimal with only the digits 0, 1, and 2, we see that the equality in (1.2) need not hold:
log 5 2 log 3
= dimH (X + X) < 2 dimH X = .
log 10 log 10
Nevertheless, it is a consequence of the following theorem of Lindenstrauss, Meiri, and Peres
that the dimension of the iterated sumset X+· · ·+X approaches 1 as the number of summands
increases.

Theorem 1.4 ([LMP, Corollary 1.2]). Let (Xi )∞

P∞ i=1 be a sequence of ×r-invariant subsets of
[0, 1]. If i=1 dimH Xi /| log dimH Xi | diverges, then


lim dimH X1 + · · · + Xn = 1.
n→∞

This theorem demonstrates that the multiplicative structure captured by multiplicative

invariance sits transversely to the additive structure captured by additive closure: because
1
Marstrand’s projection theorem originally concerns orthogonal projections of subsets of the plane. Images
of the Cartesian product X × Y under orthogonal projections are, up to affine transformations which preserve
dimension, sumsets of the form λX + ηY . Also note that for sufficiently regular sets X and Y , dimH (X × Y ) =
dimH X + dimH Y ; see, for example, [Mat, Corollary 8.11].

4
the sumset X1 + · · · + Xn fills out the entire space (with respect to the Hausdorff dimension),
the sets Xi are not contained in an additively closed set of dimension less than 1. A stronger
conclusion is reached under the assumption of multiplicative independence of the bases with
respect to which X1 and X2 are structured: Theorem 1.3 gives that relatively few of the sums
x1 + x2 , with xi ∈ Xi , coincide.
While there is a strong historical precedent for the study of ×r-invariant subsets of the
unit interval, less seems to be known in the integer and r-adic settings, despite the fact that
many of the same objects and questions can be naturally formulated there. Furstenberg
[Fur2], assuming a positive answer to one of his yet-unresolved transversality conjectures in
the reals, drew a connection between the real regime and the integers by showing that given
any finite collection of finite strings from the symbols {0, . . . , 9}, the number 2n , written in
decimal, contains all of those strings provided that n is sufficiently large. In the same work,
he also proved an analogue of Theorem 1.1 in the rs-adics showing that no non-trivial set is
structured with respect to multiplicatively independent bases.
Integer restricted digit Cantor sets, to which we now turn our focus, are sets of integers
which exhibit a special case of the structure we consider in this paper and which have received
some attention in the literature. An integer base-r restricted digit Cantor set is a set of
non-negative integers whose base-r expansion includes only digits from a fixed set D ⊆
{0, 1, . . . , r − 1}, i.e.,
( n  )
X 

ai r i  n ∈ N0 , a0 , . . . , an ∈ D . (1.6)

i=0

Given the direct analogy between these sets and their real counterparts, integer restricted digit
Cantor sets are natural first candidates for analogues of the transversality results mentioned
above. While a number of arithmetic properties of restricted digit Cantor sets in the positive
integers are well studied – divisibility [BS], distribution in arithmetic progressions [EMS,
Kon], number of prime factors [KMS], character sums [BCS] – much less appears to be
known about the relationship between integer restricted digit Cantor sets with respect to
different bases.
An unresolved conjecture of Erdős [Erd] posits that for all but finitely many positive
integers n, the number 2n requires the digit 1 in order to be expressed in base 3; see [DW, Lag]
for some recent progress. In the same vein, it is a folklore conjecture in number theory [Inc]
that {0, 1, 82000} is equal to the set A of non-negative integers that can be written in bases
2,
 3, 4, and 5 usingε only the digits 0 and 1; Burrell and Yu [BY] proved that for all ε > 0,
A ∩ [0, N ] 6 Cε N (that is, the set has zero upper mass dimension, as defined in the next
section). These statements are all profitably understood in terms of intersections of restricted
digit Cantor sets; as such, they strongly resemble the intersection transversality conjectures
of Furstenberg in [Fur2]. The recent resolution of Furstenberg’s intersection conjecture in
the reals by Shmerkin [Shm] and Wu [Wu] perhaps lends some evidence in favor of analogous
intersection transversality statements in other regimes.
Much less appears to be known regarding additive transversality in the integers. Yu [Yu]
achieves some results on the number of solutions to the equation x + y = z in which the
variables come from different integer restricted digit Cantor sets. Our main results, to which
we turn next, concern solutions to the equation x1 + y1 = x2 + y2 , where x1 , x2 ∈ X and
y1 , y2 ∈ Y , for much more generally structured sets X and Y in the real and the integer
settings.

5
 1.2. Main results
The main results in this article are best categorized by two settings: the unit interval and
the non-negative integers. For subsets of the unit interval, we enhance Theorem 1.3 by
demonstrating uniformity in the quantifiers on λ and η. We provide a new, combinatorial
proof of Theorem 1.3 that avoids the machinery of local entropy averages and CP-processes.
For subsets of the non-negative integers, we introduce a natural class of ×r-invariant sets. We
prove a result on iterated sumsets of ×r-invariant sets, and we demonstrate how invariance
with respect to multiplicatively independent bases leads to analogues of transversality results
of Furstenberg and Hochman-Shmerkin.

1.2.1. Sums of multiplicatively invariant subsets of the reals

In order to capture the way in which the conclusion of Theorem 1.3 is uniform in the pa-
rameters λ and η, and in order to define a discrete Hausdorff dimension for subsets of the
non-negative integers, we make use of the discrete Hausdorff content, defined for X ⊆ Rd to
be
(  )
X  [
γ 
H>ρ X) = inf δiγ  X ⊆ Bi , Bi open ball of diameter δi > ρ .

i∈I i∈I

The discrete Hausdorff content is discussed at more length in Section 2.1 (see Definition 2.3).
For the current discussion, it is helpful to know that for compact sets X,

  


dimH X = sup γ > 0  lim Hγ X > 0 ; (1.7)
 ρ→0+ >ρ

see Lemma 2.4. Thus, bounding the discrete Hausdorff content from below at all scales uni-
formly across a family of sets allows us to quantify a “uniform lower bound on the Hausdorff
dimension” across the family.

Theorem A. Let r and s be multiplicatively independent positive integers, and let X,
Y ⊆
[0, 1] be ×r- and ×s-invariant sets, respectively. Define γ = min dimH X + dimH Y, 1 . For
all compact I ⊆ R\{0} and all γ < γ,
γ

lim inf H>ρ λX + ηY > 0. (1.8)
ρ→0+ λ,η∈I

We use Theorem A in a critical way in our main sumset result for the integers, Theorem D.
Beyond that, utilizing the fact in (1.7), Theorem 1.3 follows as an immediate corollary to
Theorem A. Thus, we provide a new proof of [HS, Theorem 1.3] for sets that avoids use of
the main tools in [HS], namely CP-processes and local entropy averages.
The core of our argument in the proof of Theorem A can be traced back to Peres and
Shmerkin [PS], who showed, among other things, that Theorem 1.3 holds for restricted digit
Cantor sets. We are able to modify and strengthen their argument to achieve uniformity in the
parameters λ and η and extend the result to sets which are only assumed to be multiplicatively
invariant. The latter is achieved by combining a flexible combinatorial discrete Marstrand
theorem in Section 3.1 with a combinatorial result on trees in Section 3.2.

6
 1.2.2. Multiplicatively invariant subsets of the non-negative integers
A primary goal of this article is to introduce the study of multiplicatively invariant subsets
of the non-negative integers and demonstrate some transversality results analogous to those
on the real line. To that end, we begin by introducing an analogue of a ×r-invariant set for
the integers. Let r ∈ N, r > 2. Define Rr : N0 → N0 and Lr : N0 → N0 by
Rr : n 7→ ⌊n/r⌋ and Lr : n 7→ n − r k ⌊n/r k ⌋,
where k = ⌊log n/ log r⌋ when n > 1 and ⌊ · ⌋ denotes the floor function. The maps Rr
and Lr are best understood using the base-r representations of non-negative integers: if
n = ak r k + · · · + a1 r + a0 is the base-r representation of n, then
Rr (n) = ak r k−1 + · · · + a2 r + a1 and Lr (n) = ak−1 r k−1 + · · · + a1 r + a0 .
In other words, the map Rr “forgets” the least significant digit (the right-most digit,
hence the letter R) while the map Lr “forgets” the most significant digit (the left-most digit,
hence the letter L) in base r. For example, in base r = 10 we have that R10 (71393) = 7139
and L10 (71393) = 1393.
Definition 1.5. A set A ⊆ N0 is ×r-invariant if Rr (A) ⊆ A and Lr (A) ⊆ A.
It may be helpful to note that a ×r-invariant set A need not satisfy rA ⊆ A and that there
are examples showing that the condition rA ⊆ A does not yield a natural integer analogue
of the notion of ×r-invariance on the unit interval; see Section 4.6.
There are many natural examples of ×r-invariant subsets of N0 . Integer base-r restricted
digit Cantor sets, defined in (1.6), are clearly ×r-invariant. More general examples arise
from symbolic subshifts of {0, 1, . . . , r − 1}N0 . For any closed and left-shift-invariant set
Σ ⊆ {0, 1, . . . , r − 1}N0 , the corresponding language set is defined by
 
L(Σ) = (w0 , w1 , . . . , wk )  (w0 , w1 , . . .) ∈ Σ, k ∈ N0 .
Any language set naturally embeds into the non-negative integers as
 
wk r k + · · · + w1 r + w0  (w0 , w1 , . . . , wk ) ∈ L(Σ) ,
yielding a set that is ×r-invariant. For more details, see Definition 4.1 and Proposition 4.3,
and for more such examples, see Examples 4.2. As yet another source of ×r-invariant subsets
of the non-negative integers, we note that if X is a ×r-invariant subset of [0, 1], then the set
[  
⌊r k x⌋  x ∈ X
k∈N0

can be shown to be ×r-invariant; see Section 4.2 for more details.

Our first result in the integer setting is a natural analogue of Theorem 1.1 that demon-
strates that there are no non-trivial examples of sets which exhibit structure simultaneously
with respect to multiplicatively independent bases.

Theorem B. Let r and s be multiplicatively independent positive integers. If A ⊆ N0 is

simultaneously ×r- and ×s-invariant then either A is finite or A = N0 .

To measure the size of multiplicatively invariant subsets of N0 and their sumsets, we

make use of two notions of dimension in the integers that parallel the classical Minkowski

7
and Hausdorff dimensions from geometric measure theory. The discrete analogue of the lower
and upper Minkowski dimension are the lower and upper mass dimension, defined for A ⊆ N0
as
(    )
log |A ∩ [0, N )|  A ∩ [0, N )

dimM A = lim inf = sup γ > 0  lim inf >0 ,
N →∞ log N  N →∞ Nγ
(    )
log |A ∩ [0, N )|  A ∩ [0, N )

dimM A = lim sup = sup γ > 0  lim sup >0 .
N →∞ log N  N →∞ Nγ

Whenever dimM A = dimM A, we say that the mass dimension of A exists and denote it by
dimM A. In analogy to (1.7), the lower and upper discrete Hausdorff dimension of A are
defined to be
( 
)
 H γ
A ∩ [0, N )

dimH A = sup γ > 0  lim inf >1 >0 ,
 N →∞ Nγ
( 
)
 γ
H>1 A ∩ [0, N )

dimH A = sup γ > 0  lim sup >0 ,
 N →∞ Nγ

and if these two quantities agree then we say that the discrete Hausdorff dimension of A,
dimH A, exists and is equal to this quantity.
The mass dimension and the upper discrete Hausdorff dimension are systematically stud-
ied along with a host of other discrete dimensions in [BT2]. We discuss these notions of
dimension and the interplay between them at greater length in Section 2.3. For the current
discussion, it is helpful to know that
dimH 6 dimM 6 dimM and dimH 6 dimH 6 dimM , (1.9)
and that for any ×r-invariant set A ⊆ N0 , both the mass dimension dimM A and the discrete
Hausdorff dimension dimH A exist and coincide; see Lemma 2.18 and Remark 4.7.
Our second main result in the integer setting is an analogue of Theorem 1.4 concerning
the dimension of iterated sumsets of ×r-invariant sets.
P∞
Theorem C. Let (Ai )∞ i=1 be a sequence of ×r-invariant subsets of N0 . If i=1 dimH Ai /
| log dimH Ai | diverges, then


lim dimH A1 + · · · + An = 1.
n→∞

In the same way as in the continuous regime, this theorem demonstrates that the struc-
ture captured by ×r-invariance in N0 sits transversely to the additive structure captured by
additive closure. It also demonstrates the connection between ×r-invariant subsets of the in-
tegers and ×r-invariant subsets of [0, 1], and it will serve to emphasize the role multiplicative
independence plays in the other results in this section.
Our final results in the integer setting concern the dimension of sumsets of ×r- and
×s-invariant sets. To better frame these results, consider the following discrete analogue
of Marstrand’s theorem from [Gla, Theorem 1.4]: for all A, B ⊆ Z satisfying a necessary

8
dimension condition 2 and for Lebesgue-a.e. (λ, η) ∈ R2 ,



dimM ⌊λA + ηB⌋ = min dimM (A × B), 1 , (1.10)
 
where ⌊λA + ηB⌋ := ⌊λa + ηb⌋  a ∈ A, b ∈ B . When A and B are sufficiently regular
– for example, when A and B are ×r- and ×s-invariant, respectively – dimM (A × B) =
dimM A+dimM B, and thus the typical sumset has dimension as large as possible. In the same
way that Theorem 1.3 improves Marstrand’s theorem under the assumption of multiplicative
invariance, the following theorem improves this discrete version of Marstrand’s theorem under
the analogous structural assumptions.

Theorem D. Let r and s be multiplicatively independent positive integers, and let A, B ⊆

N0 be ×r- and ×s-invariant sets, respectively. For all λ, η > 0, the mass and discrete
Hausdorff dimensions of ⌊λA + ηB⌋ exist and




dimH ⌊λA + ηB⌋ = dimM ⌊λA + ηB⌋ = min dimH A + dimH B, 1 .

In particular, Theorem D tells us that for multiplicatively independent bases r and s, any
×r-invariant set A and ×s-invariant set B are additively combinatorially transverse in the
sense that



dimH A + B = min dimH A + dimH B, 1 , (1.11)
thus realizing a natural analogue to Furstenberg’s sumset conjecture in the integers.
Bounding dimH (A + B) from above is accomplished by a straight-forward combination of
the trivial upper sumset estimate in (1.3), the dimension bounds in (1.9), and the fact that
the mass and discrete Hausdorff dimensions coincide for ×r- and ×s-invariant sets. Much
more work goes into bounding dimH (A + B) from below. The following theorem gives the
lower bound required for (1.11) by demonstrating more: the same sort of uniformity in the
real regime described in Theorem A is present in the integer regime.

B ⊆ N0
Theorem E. Let r and s be multiplicatively independent positive integers, and let A,
be ×r- and ×s-invariant sets, respectively. Define γ = min dimH A + dimH B, 1 . For all
compact I ⊆ (0, ∞) and all γ < γ,
γ

H>1 ⌊λA + ηB⌋ ∩ [0, N )
lim inf inf > 0.
N →∞ λ,η∈I Nγ
Following the outline in the previous paragraph, Theorem D follows from Theorem E; a
complete proof is given in Section 4.5. Since Theorem E is, in turn, derived from Theorem A,
one can view Theorem D as a consequence of Theorem A. It is natural to ask whether
Theorem D can be derived from Theorem 1.3 directly, but it seems to the authors that this
is not possible and that one needs the full strength of Theorem A to obtain Theorem D.
2
The condition is that the upper mass dimension of A×B is equal to the upper counting dimension of
 A×B.
The upper mass dimension of A × B is dimM (A × B) := lim supN→∞ log (A × B) ∩ {−N, . . . , N }2  log N ,


while the upper  counting dimension of A × B is equal to lim supN→∞ maxz∈Z2 log (A × B) ∩ (z +

{−N, . . . , N }2 ) log N .

9
 1.3. Overview of the paper
In Section 2, we present several preliminary results that are used later, including some basic
facts from discrete and continuous fractal geometry, some properties of ×r-invariant subsets of
[0, 1], and notions of dimension for subsets of N0 . Section 3 contains a proof of Theorem A,
adapting the argument from [PS] and combining it with two new crucial ingredients: a
discrete version of Marstrand’s projection theorem which includes topological information on
the exceptional set and handles projections of “large subsets”; and a combinatorial theorem
about existence of regular subtrees. In Section 4, we study ×r-invariant sets of integers and
their dimensions. In particular, we prove Theorems B, C, D and E. Theorem B is proved
with a self contained elementary argument, and Theorem C is derived from Theorem 1.4.
Theorem E, from which Theorem D follows as a corollary, is proved using Theorem A and
a correspondence between ×r-invariant sets of [0, 1] and ×r-invariant sets of N0 . Finally, in
Section 5, we collect a number of open questions concerning ×r-invariant subsets of N0 .

1.4. Acknowledgements
The authors extend a debt of gratitude to Pablo Shmerkin and Mark Pollicott, whose in-
sightful questions led to improvements in the formulations of Theorems D and E. The third
author is supported by the National Science Foundation under grant number DMS 1901453.

2. Preliminary definitions and results

The positive and non-negative integers are denoted by N and N0 , respectively. For x ∈ R,
denote the fractional part by {x} and the integer part (or floor) by ⌊x⌋. The Lebesgue
measure on the real line is denoted by Leb. Throughout the paper, Rd is equipped with the
Euclidean norm which we denote by | · |. Given two positive-valued functions f and g, we
write f ≪a1 ,...,ak g or g ≫a1 ,...,ak f if there exists a constant K > 0 depending only on the
quantities a1 , . . . , ak for which f (x) 6 Kg(x) for all x in the domain common to both f and
g. We write f ≍a1 ,...,ak g if both f ≪a1 ,...,ak g and f ≫a1 ,...,ak g.

2.1. Continuous and discrete fractal geometry

In this section, we lay out the notation, tools, and results we need from continuous and
discrete fractal geometry. A good general reference for the standard material in this section
is [Mat, Ch. 4]. In the definitions that follow, ρ, γ, c > 0, d ∈ N, and X ⊆ Rd is non-empty.
Definition 2.1.
• The set X is ρ-separated if for all distinct x1 , x2 ∈ X, |x1 − x2 | > ρ.
• The metric entropy of X at scale ρ is
 
N (X, ρ) = sup |X0 |  X0 ⊆ X is ρ-separated .
• The lower Minkowski dimension of X is
log N (X, δ)
dimM X = lim inf . (2.1)
δ→0+ log δ−1
The upper Minkowski dimension, dimM X, is defined analogously with a limit supremum
in place of the limit infimum. If dimM X = dimM X, then this value is the Minkowski

10
 dimension of X, dimM X.
It is a well-known fact that we will
 use without further mention that if ρ < 1, then

dimM X = lim inf N →∞ log N (X, ρ−N ) log ρN and dimM X = lim supN →∞ log N (X, ρ−N )
log ρN .
Definition 2.2.
• The unlimited Hausdorff content at dimension γ of X is
(  )
X γ  [
γ
H∞ (X) = inf δi  X ⊆ Bi , Bi open ball of diameter δi .

i∈I i∈I

Note that when X is compact, the index set I may be taken to be finite.
• The Hausdorff dimension of X is
γ
dimH X = sup{γ ∈ R | H∞ (X) > 0}
γ
= inf{γ ∈ R | H∞ (X) = 0}.
In the following definition, we introduce two notions meant to capture the dimensionality
of discrete sets.
Definition 2.3.
• (cf. [KT, Definition 1.2]) The set X is a (ρ, γ)c -set if it is ρ-separated and for all δ > ρ
and all open balls B of diameter δ,
 γ
 
X ∩ B  6 c δ . (2.2)
ρ
• The discrete Hausdorff content of X at scale ρ and dimension γ is
(  )
γ
X γ  [
H>ρ (X) = inf δi  X ⊆ Bi , Bi open ball of diameter δi > ρ .

i∈I i∈I

Note that when X is compact, the index set I may be taken to be finite.
In the definition of a (ρ, γ)c -set, we think of ρ as being positive and close to 0, γ ∈ [0, d] as
the “dimension” of the set, and c > 0 as an uninteresting parameter that exists only to make
our arguments explicit. The inequality in (2.2) guarantees that the points of a (ρ, γ)c -set
cannot be too concentrated in any ball. It follows from that inequality that the maximum
cardinality of a (ρ, γ)c set in [0, 1]d is on the order of ρ−γ . A (ρ, γ)c -set with cardinality
≫ ρ−γ can be thought of as a discrete approximation to a set with Hausdorff dimension γ;
this is made more precise in Remark 2.5 below and is realized in Lemma 2.13.
The discrete Hausdorff content at scale ρ is a “ρ-resolution” analogue of the unlimited
Hausdorff content. The discrete Hausdorff contents of two sets that look the same at scale ρ
are approximately equal; the more formal statement can be found in Lemma 2.7.
The following lemma provides a connection between the discrete and the continuous
regimes and proves the equality in (1.7) from the introduction.

Lemma 2.4. Let X ⊆ Rd be compact. For all γ > 0,

γ γ
lim H>ρ (X) = H∞ (X). (2.3)
ρ→0+

11
 γ
Consequently, if limρ→0 H>ρ (X) > 0, then dimH X > γ.
γ
Proof. Let γ > 0. The limit in (2.3) exists because the function ρ 7→ H>ρ (X) is non-increasing
+ γ
as ρ tends to 0 and is bounded from below by H∞ (X). Equality in the limit follows from
γ
the fact that X is compact, allowing for the index set in the definition of H∞ (X) to be taken
γ γ
to be finite. If limρ→0 H>ρ (X) > 0, then H∞ (X) > 0, and it follows from the definition of
the Hausdorff dimension that dimH X > γ.

Remark 2.5. It would be natural to define the metric entropy at scale ρ and dimension γ
of the set X as
 
N (X, (ρ, γ)c ) = sup |X0 |  X0 ⊆ X is a (ρ, γ)c -set .
Using a max flow, min cut argument similar to the one in [BP, Ch. 3], it can be shown that
for X compact,


N X, (ρ, γ)c γ
−γ
≍c,d H>ρ (X). (2.4)
ρ
Thus, (ρ, γ)c -sets of cardinality ≫ ρ−γ can be thought of as discrete fractal sets of dimension
γ. We will not need (2.4), so we omit the details.
The following is a discrete version of the well-known mass distribution principle, cf. [BP,
Lemma 1.2.8].

Lemma 2.6. Let µ be a Borel probability measure on Rd , and let ρ, κ > 0. If for all balls B
γ
of diameter δ > ρ, µ(B) 6 κδγ , then the support supp µ of µ satisfies H>ρ (supp µ) > κ−1 .

Proof.
P Let ε > 0, and let {Bi }i∈I be a cover of supp µ with ball Bi of diameter δi > ρ and with
γ γ
i∈I i 6 H>ρ (supp µ) + ε. Then the conclusion follows because ε > 0 was arbitrary.
δ

Denote by [X]δ the closed δ-neighborhood of X:

 
[X]δ := z ∈ [0, 1]  ∃x ∈ X with |z − x| 6 ε .
The Hausdorff distance between two non-empty, compact sets X, Y ⊆ R is
 
dH (X, Y ) := inf ε > 0  X ⊆ [Y ]ε and Y ⊆ [X]ε .
By the Blaschke selection theorem, the set of all non-empty, compact subsets of R equipped
with the Hausdorff distance is a complete metric space.

Lemma 2.7. Let a > 1 and ρ > 0. If X, Y ⊆ R are compact and X ⊆ [Y ]aρ , then



N X, ρ ≪a 1 + N Y, ρ
and
γ γ
H>ρ (X) ≪a H>ρ (Y ).

Proof. If n = N (X, ρ) then, by the definition of metric entropy, there exist u1 , . . . , un ∈ X

such that ui+1 − ui > ρ for all i ∈ {1, . . . , n − 1}. Since X ⊆ [Y ]aρ , we can find for every
i ∈ {1, . . . , n} a point vi ∈ Y such that |ui − vi | 6 aρ. Then, by the triangle inequality,

12
  
we have vi+2a+1 − vi > ρ. Therefore the set Y ′ := v(2a+1)i  1 6 i 6 ⌊n/(2a + 1)⌋ is a
ρ-separated subset of Y . This implies
 


n N X, δ
N Y, ρ > |Y ′ | > > − 1.
2a + 1 2a + 1
γ γ
For the proof of H>ρ (X) ≪a H>ρ let {Bi }i∈I be a collection of open balls that
(Y ),P S covers
Y and where Bi has diameter ri > ρ and i∈I riγ < 2H>ρ γ
(Y ). It follows that
P X ⊆ i∈I [Bi ]aρ
γ
and [Bi ]aρ is a ball of diameter ri +2aρ 6 (2a+1)ri . Therefore H>ρ (X) 6 i∈I ((2a+1)ri )γ 6
γ
2(2a + 1)H>ρ (Y ).

2.2. Multiplicatively invariant subsets of the reals and their finite

approximations
Multiplicatively invariant subsets of [0, 1] and N0 are the main objects of study in this paper.
In this section, we record some basic facts about multiplicatively invariant subsets of [0, 1]
and their discrete approximations.
Definition 2.8. Let r ∈ N and X ⊆ [0, 1].
• The map Tr : [0, 1] → [0, 1] is defined by Tr x = {rx}, the fractional part of the real
number rx.
• The set X is ×r-invariant if it is closed and Tr X ⊆ X.
The Hausdorff and Minkowski dimensions of a multiplicatively invariant set coincide. As
a consequence of this regularity, the Hausdorff dimension of products of such sets is also
well-behaved. We record these facts here for later use.

Theorem 2.9 ([Fur1, Proposition III.1]). If X ⊆ [0, 1] is ×r-invariant, then dimH X =

dimM X.

Lemma 2.10. If X, Y ⊆ [0, 1] are ×r, ×s-invariant, respectively, then dimH (X × Y ) =

dimH X + dimH Y .

Proof. This follows immediately from [Mat, Corollary 8.11] and the fact that dimH X =
dimM X.

Since we will work almost exclusively with finite approximations to multiplicatively in-
variant sets, we establish some useful notation.
Definition 2.11. Let X ⊆ [0, 1] be ×r-invariant. For n ∈ N0 , the set Xn denotes the set X
rounded down to the lattice r −n Z. That is, the point i/r n is an element of Xn if and only if
X ∩ [i/r n , (i + 1)/r n ) is non-empty.
The next results show that finite approximations to a multiplicatively invariant set are
multiplicatively invariant and are discrete models of fractal sets as captured by Definition 2.3.

Lemma 2.12. Let X ⊆ [0, 1] be ×r-invariant. For all n ∈ N0 , Tr Xn ⊆ Xn−1 .

Proof. Let n ∈ N0 , and let i/r n ∈ Xn with i ∈ {0, . . . , r n − 1}. Write i = i0 + dn−1 r n−1
with i0 ∈ {0, . . . , r n−1 − 1} and dn−1 ∈ {0, . . . , r − 1}. Note that Tr (i/r n ) = i0 /r n−1 and
Tr ((i + 1)/r n ) = (i0 + 1)/r n−1 . We must show that i0 /r n−1 ∈ Xn−1 .

13
 Since i/r n ∈ Xn , there exists x ∈ X ∩ [i/r n , (i + 1)/r n ). Since Tr x ∈ X, Tr x ∈ X ∩
[i0 /r n−1 , (i0 + 1)/r n−1 ). It follows by the definition of Xn−1 that i0 /r n−1 ∈ Xn−1 , as was to
be shown.

Lemma 2.13. Let r > 2, and let X ⊆ [0, 1] be ×r-invariant with dimH X > 0. For all
0 < γ4 < dimH X < γ5 , there exists c > 0 such that for all sufficiently large N ∈ N, the set
XN is a (r −N , γ5 )c -set satisfying r N γ4 6 |XN | 6 r N γ5 .

Proof. Put γ = dimH X. It follows from Theorem 2.9 that there exists c0 > 0 such that for
all N ∈ N,
|XN | 6 c0 r N γ5 , (2.5)
and for all sufficiently large N ∈ N,
r N γ4 6 |XN | 6 r N γ5 .
Using the fact that X is ×r-invariant and the bound in (2.5), for all 0 6 n 6 N and for
all i ∈ {0, . . . , r n − 1},
  
 i i + 1 
XN ∩ , n  6 |XN −n | 6 c0 r (N −n)γ5 .
 r n r
Put c = 2r γ5 c0 . To show that XN is a (r −N , γ5 )c -set, let B ⊆ R be a ball of diameter
δ > r −N . Put n = ⌊− logr δ⌋ so that r −(n+1) < δ 6 r −n , and note that a union of two
intervals of length r n of the form above suffice to cover B. Therefore,
 γ5
(N −n)γ5 δ
|XN ∩ B| 6 2c0 r 6 c −N ,
r
as was to be shown.

The following notation, borrowed from [PS], allows us to easily compare powers of r
and powers of s. This is useful when considering the finite approximations to the Cartesian
product of a ×r- and a ×s-invariant set.
Definition 2.14. For n ∈ N0 , we set n′ = ⌊n log r/ log s⌋ to be the greatest integer so that
′
sn 6 r n . (The bases r and s do not appear in this notation but should always be clear from
context.)
Recall from Definition 2.11 that XN is the set X rounded to the lattice r −N Z. Extending
this notation to Y , the set YN is the set Y rounded to the lattice s−N Z. Henceforth, this
discrete approximation will always appear as YN ′ , which is the set Y rounded to the lattice
′
s−N Z.

Corollary 2.15. Let 2 6 r < s, and let X, Y ⊆ [0, 1] be ×r- and ×s-invariant sets with
dimH X, dimH Y > 0. For all 0 < γ4 < dimH (X × Y ) < γ5 , there exist c1 , c2 > 0 such that
for all sufficiently large N ∈ N, the sets XN × YN ′ and XN × YN ′ +1 are (c1 r −N , γ5 )c2 -sets
satisfying r N γ4 6 |XN × YN ′ | 6 r N γ5 and r N γ4 6 |XN × YN ′ +1 | 6 r N γ5 .

Proof. Let 0 < g4 < dimH X < g5 and 0 < h4 < dimH Y < h5 be such that
γ4 < g4 + h4 < dimH (X × Y ) < g5 + h5 < γ5 .

14
Applying Lemma 2.13, there exist c, d > 0 such that for sufficiently large N ∈ N, the set
′
XN is a (r −N , g5 )c -set satisfying r N g4 6 |XN | 6 r N g5 and YN ′ is a (s−N , h5 )d -set satisfying
′ ′
sN h4 6 |YN ′ | 6 sN h5 . It follows that for sufficiently large N ∈ N, r N γ4 6 |XN × YN ′ | 6 r N γ5
and r N γ4 6 |XN × YN ′ +1 | 6 r N γ5 .
′ ′
Set c1 = s−1 and c2 = sg5 cd. Since sN < r N < sN +1 , the sets XN × YN ′ and XN × YN ′ +1
are c1 r −N -separated. Since XN is a (r −N , g5 )c -set, it is a (c1 r −N , g5 )sg5 c -set.3 Let B ⊆ R2
be a ball of diameter δ > c1 r −N . Note that
 g 5  h 5
  δ δ
(XN × YN ′ ) ∩ B  6 sg5 c d −N ′
c1 r −N s
 g 5  h 5  γ5
g5 δ h5 δ δ
6s c dc1 6 c2 ,
c1 r −N c1 r −N c1 r −N
which shows that the set XN × YN ′ is a (c1 r −N , γ5 )c2 -set. By a similar calculation,
 g 5  h 5
  δ δ
(XN × YN ′ +1 ) ∩ B  6 sg5 c d −(N ′ +1)
c1 r −N s
 g 5  h 5  γ5
g5 δ δ δ
6s c d 6 c2 ,
c1 r −N c1 r −N c1 r −N
which shows that the set XN × YN ′ +1 is a (c1 r −N , γ5 )c2 -set.

2.3. Dimension of subsets of integers

To measure the size of subsets of N0 , we will make use of the (upper and lower) mass dimension
and the (upper and lower) discrete Hausdorff dimension, defined below. The upper and
lower mass dimensions and the upper Hausdorff dimension are treated systematically in
[BT2]; we will state the properties we require from these quantities with the aim of making
this presentation self-contained. These dimensions join a bevy of other natural notions of
dimension for subsets of the integers, integer lattices, and more general discrete sets; see
[Nau, BT1, IRUT, LM2].
Definition 2.16. Let A ⊆ N0 be non-empty.
• The lower mass dimension of A is
 
log A ∩ [0, N )
dimM A = lim inf .
N →∞ log N
The upper mass dimension, dimM A, is defined analogously with a limit supremum in
place of the limit infimum. If dimM A = dimM A, then this value is the mass dimension
of A, dimM A.
• The lower Hausdorff dimension of A is
( 
)
 γ
H>1 A ∩ [0, N )

dimH A = sup γ > 0  lim inf >0 .
 N →∞ Nγ

The upper Hausdorff dimension, dimH A, is defined analogously with a limit supremum
3
More generally, if 0 < c1 < 1, then every (δ, γ)c -set is a (c1 δ, γ)c−γ c -set. This is a quick exercise left to
1
the reader.

15
 in place of the limit infimum. If dimH A = dimH A, then this value is the discrete
Hausdorff dimension of A, dimH A.
As the notation suggests, the mass and discrete Hausdorff dimensions are defined in
analogy to the Minkowski and Hausdorff dimensions, respectively. The analogies becomes
clearer on noting that
 
 
A ∩ [0, N ) = N A ∩ [0, N ) , N −1 , (2.6)
N
γ
 
H>1 A ∩ [0, N ) γ A ∩ [0, N )
= H>N −1 , (2.7)
Nγ N
so that the mass and discrete Hausdorff dimensions are capturing, in some sense, the Minkowski
and Hausdorff dimensions of the sequence of sets N 7→ A/N in the unit interval. It should al-
ways be clear from context which dimension is understood: we will never consider Minkowski
or Hausdorff dimension of subsets of N0 .
As a word of caution, note that our terminology does not match exactly with the termi-
nology used in [BT2]. What we call the upper discrete Hausdorff dimension is called dimL in
[BT2] (see Lemma 2.3 in that paper), while the discrete Hausdorff dimension defined in that
work does not appear in our work.
In the following lemmas, we collect the required properties of the mass and discrete
Hausdorff dimensions.

Lemma 2.17. Let A, B ⊆ N0 , λ > 0, and η ∈ R.

(I) For all dim ∈ {dimM , dimM , dimH , dimH }, dim A ∈ [0, 1].

(II) For all dim ∈ {dimM , dimM , dimH , dimH }, dim A = dim ⌊λA + η⌋ .


(III) For all dim ∈ {dimM , dimH }, dim(A ∪ B) = max dim A, dim B .
(IV) For all r ∈ N, r > 2,
( 
)
 γ
H>1 A ∩ [0, r N )

dimH A = sup γ > 0  lim inf >0 ,
 N →∞ rN γ

and the analogous statement with dimH in place of dimH and limit supremum in place
of limit infimum holds.
(V) dimM (A + B) 6 dimM A + dimM B.

Proof. The statements in (I), (II), and (III) follow from straightforward calculations which
are left to the reader. The sets in Examples 2.19 (II) below show that (III) does not hold for
the lower mass and discrete Hausdorff dimensions.
Both of the statements in (IV) follow from the fact that for all γ > 0 and all r K 6 N 6
r K+1 ,
γ
γ
γ

H>1 A ∩ [0, r K ) γ H>1 A ∩ [0, N ) 2γ H>1 A ∩ [0, r
K+1 )
6r 6r .
r Kγ Nγ r (K+1)γ
Indeed, this shows
γ
that the limit infimum (resp. limit supremum) of the sequence N 7→
H>1 A ∩ [0, r N ) /r N γ is non-zero if

and only if the limit infimum (resp. limit supremum) of
γ γ
the sequence N 7→ H>1 A ∩ [0, N ) /N is non-zero.
To show the statement in (V), note that for finite sets F, G ⊆ N0 , |F + G| 6 |F ||G|.

16
Applying this to a finite segment of the set A + B, we see
  

   
(A + B) ∩ [0, N ) 6  A ∩ [0, N ) + B ∩ [0, N )  6 A ∩ [0, N )B ∩ [0, N ).

The statement in (V) follows by taking logarithms and a limit supremum as N tends to
infinity. The sets in Examples 2.19 (II) and (IV) below show that the statement in (V) does
not hold for the lower mass dimension or the upper or lower discrete Hausdorff dimensions.

Lemma 2.18. For all A ⊆ N0 ,

dimH A 6 dimM A 6 dimM A,
dimH A 6 dimH A 6 dimM A,
and no other comparisons are possible in general.

Proof. It is immediate from the definitions that dimM A 6 dimM A and dimH A 6 dimH A,
and the set in Examples 2.19 (I) below shows that neither of these inequalities are, in general,
equalities.
To see that dimH A 6 dimM A and that dimH A 6 dimM A, note that it follows by covering
A ∩ [0, N ) by |A ∩ [0, N )| many balls of diameter 1 that
γ

H>1 A ∩ [0, N ) |A ∩ [0, N )|
γ
6 .
N Nγ
If γ > dimM A (resp. γ > dimM A), then the limit infimum (resp. limit supremum) of the
right hand side is zero, implying that γ > dimH A (resp. γ > dimH A). It follows that
dimH A 6 dimM A and dimH A 6 dimM A. The set in Examples 2.19 (III) below shows that
neither of these inequalities are, in general, equalities.
To see that no other comparisons are possible, it suffices to show that there can in general
be no comparison between dimH and dimM . This is demonstrated by the sets in Exam-
ples 2.19 (I) and (III) below.

We conclude this section with some examples meant to illustrate the extent to which the
mass and discrete Hausdorff dimensions relate. These examples do not feature the type of
structures that we are concerned with in this work, so we leave some of the details to the
reader.
Examples 2.19.
(I) Let (xn )∞
n=0 ⊆ N0 be any sequence which satisfies limn→∞ log(xn+1 − xn )/ log xn+1 = 1,
and define
∞
[
A := {0} ∪ {x2n , x2n + 1, . . . , x2n+1 }.
n=0

It is easy to check that dimM A = dimH A = 0 and
that dimM A = dimH A = 1.

(II) Let A be the set from (I). Put B = {0} ∪ N0 \A . Then dimM B = dimH B = 0 while
dimM B = dimH B = 1, and A + B = A ∪ B = N0 .
(III) Define
∞
[ 
A = {0, . . . , 16} ∪ 2n , . . . , 2n + ⌊2n−n/ log n ⌋ .
n=2

17
 It is quick to check that the mass dimension of A exists and dimM A = 1. On the
other hand, by covering A with the intervals in its definition, it can be shown that the
discrete Hausdorff dimension of A exists and dimH A = 0.
(IV) Let A be the set from (III). Define B ⊆ N0 to contain {0, . . . , 16} and to be such that
for all n > 2, B on the interval [2n , 2n+1 ) is comprised of 2⌊2n/ log n ⌋ many integers, as
equally spaced as possible. It is quick to check that dimM B = 0. By the definitions
of A and B on the interval [2n , 2n+1 ), it is easy to see that [2n+1 , 2n+2 ) is contained in
A + B. Therefore, A + B = N0 .

3. Sums of multiplicatively invariant subsets of the reals

In this section, we prove Theorem A, the main theorem in the first half of this work. Several
auxiliary results go into the proof: a discrete version of Marstrand’s projection theorem in
Section 3.1, a regularity result for finite trees in Section 3.2, and a quantitative equidistribu-
tion result in Section 3.3. We outline the proof of Theorem A in Section 3.4 before presenting
the full details in Sections 3.5 and 3.6.
Theorem A has a geometric formulation in terms of orthogonal projections; while we
will not make any particular use of the theorem in this form, it is worth formulating for its
historical connection to the topic. Let πθ : R2 → R2 be the orthogonal projection onto the
line that contains the origin and forms an angle θ with the positive x-axis.

Y ⊆
Theorem 3.1. Let r and s be multiplicatively independent positive integers, and let X,
[0, 1] be ×r- and ×s-invariant sets, respectively. Define γ = min dimH X + dimH Y, 1 . For
all compact I ⊆ (0, π)\{π/2} and all γ < γ,
γ

lim inf H>ρ πθ (X × Y ) > 0.
ρ→0+ θ∈I

The proof of the equivalence between Theorem A and Theorem 3.1 is standard and not
needed in this work, so it is omitted.

3.1. A discrete Marstrand projection theorem

In this section, we prove a discrete analogue of Marstrand’s projection theorem from geometric
measure theory. The theorem – stated for sumsets in the introduction as Theorem 1.2 – says
that for every Borel set A ⊆ [0, 1]2 , for Lebesgue-a.e. θ ∈ [0, π), dimH πθ A = min(1, dimH A),
where πθ : R2 → R2 is the orthogonal projection onto ℓθ , the line that contains the origin
and forms an angle θ with the positive x-axis. Marstrand’s theorem and its relatives have
enjoyed much recent attention: we refer the interested reader to the survey [FFJ] and to the
end of this section where we put Theorem 3.3 into more context.
The key idea behind Marstrand’s theorem is that of “geometric transversality” and is
captured in the following lemma. An immediate consequence of the lemma is that there are
not many projections which map two distant points close together. The proof follows from a
simple geometric argument and is left to the reader.

Lemma 3.2. For all nonzero x ∈ R2 and all ρ > 0, the set of angles θ ∈ [0, π) for which
|πθ x| 6 ρ is contained in at most two balls of diameter ≪ ρ|x|−1 .

Our discrete analogue of Marstrand’s theorem, Theorem 3.3, reaches a conclusion similar

18
to that of Marstrand’s by quantifying the size of the set E of exceptional directions, those
directions in which the image of the set A is small. On a first reading, it is safe to think of
γ < 1, n ≈ ρ−γ , δ = 1, and m ≈ ρ−(γ−ε) . In this case, the set A is a discrete analogue of a
set of Hausdorff dimension γ and the set E is the set of exceptional directions in which the
set A loses at least a proportion ρε of its points.

Theorem 3.3. Let γ, ρ, c > 0. Put γ = min(γ, 1). If A ⊆ [0, 1]2 is a (ρ, γ)c -set with
n := |A| > − log c, then for all δ > 0 and all 0 6 m 6 δ2 n 4, the set
 
E = θ ∈ [0, π)  ∃A′ ⊆ A, |A′ | > δn, N (πθ A′ , ρ) 6 m (3.1)
satisfies
(
−1 m n1−γ/γ if γ 6= 1
N (E, ρ) ≪γ,c ρ .
δ2 n log n if γ = 1
2
 Let A ⊆ [0, 1] be a (ρ, γ)c -set of cardinality n > − log c. Let δ > 0, and let 0 6 m 6
Proof.
2
δ n 4.  
Define S(θ) = (a1 , a2 ) ∈ A2  |πθ (a1 − a2 )| < ρ . PLet E′ be a maximal ρ-separated
subset of E; thus, |E ′ | = N (E, ρ). The goal is to bound θ∈E ′ S(θ) from above and below
to get the desired bound on |E ′ |.
Let θ ∈ E ′ and A′ be the subset of A corresponding to θ. Since the set πθ A′ lies on a line
and N (πθ A′ , ρ) 6 m, there exists a collection {B}B∈B of no more than 2m closed balls B of
diameter ρ whose union covers πθ A′ . By Cauchy-Schwarz,
!2
X 
2 ′ 2
(δn) 6 |A | 6 {a0 ∈ A | πθ a0 ∈ B}
′

B∈B
 X 
6 B  {a0 ∈ A′ | πθ a0 ∈ B}2
B∈B
X 
6 2m {a ∈ A | πθ a ∈ B}2
B∈B
X 
= 2m {(a1 , a2 ) ∈ A2 | πθ a1 , πθ a2 ∈ B}
B∈B
 
6 2mS(θ).
It follows that
δ 2 n2 ′ X 
|E | 6 S(θ). (3.2)
2m ′ θ∈E

Now we use Lemma 3.2 to bound the right hand side of (3.2) from above: for a1 , a2 ∈
[0, 1]2 , the set
  
Θ(a1 , a2 ) = θ ∈ [0, π)  πθ (a1 − a2 ) < ρ
is contained in at most two balls of diameter ≪ ρ/|a1 − a2 |. Therefore, N (Θ(a1 , a2 ), ρ) ≪
1/|a1 − a2 |, and using the fact that E ′ is ρ-separated, we see that

19
 X X 1
1S(θ) (a1 , a2 ) = 1Θ(a1 ,a2 ) (θ) 6 K
|a1 − a2 |
θ∈E ′ θ∈E ′

for some constant K depending on the result in Lemma 3.2. It follows that
X  X X
S(θ) = 1S(θ) (a1 , a2 )
θ∈E ′ θ∈E ′ a1 ,a2 ∈A
X X
= n|E ′ | + 1Θ(a1 ,a2 ) (θ)
a1 ,a2 ∈A θ∈E ′
a1 6=a2
X
6 n|E ′ | + K |a1 − a2 |−1 ,
a1 ,a2 ∈A
a1 6=a2

and so we are left to bound the second term from above. P

For ℓ ∈ N0 , let Hℓ = {x ∈ R2 | |x| ∈ [ρeℓ , ρeℓ+1 )}. Breaking up the sum |a1 − a2 |−1 by
fixing a1 and partitioning the a2 ’s by shells, and using the fact that A is ρ-separated, we see
X ∞
XX X
−1
|a1 − a2 | = |a1 − a2 |−1
a1 ,a2 ∈A a1 ∈A ℓ=0 a2 ∈A∩(a1 +Hℓ )
a1 6=a2
∞
XX  
6 ρ−1 e−ℓ A ∩ (a1 + Hℓ ).
a1 ∈A ℓ=0
  

Since A is a (ρ, γ)c -set, for all ℓ > 0, A ∩ (a1 + Hℓ ) 6 c 2ρeℓ+1 ρ γ . On the other hand,
P∞  

ℓ=0PA ∩ (a1 + Hℓ ) = |A| − 1. It follows then from the fact that ℓ 7→ e−ℓ is decreasing
P
that ∞ ℓ=0 e
−ℓ A ∩ (a + H ) 6
1 ℓ
ℓ0 γ ℓ(γ−1)+γ , where ℓ = ⌈log((n/c)1/γ )⌉ is the smallest
ℓ=0 2 ce 0
value such that the set A could be contained in a ball of diameter ρeℓ0 about a1 . Therefore,

∞
XX ℓ0
XX
 
ℓ
ρ−1 e−ℓ A ∩ (a1 + Hℓ ) ≪γ,c ρ−1 eγ−1
a1 ∈A ℓ=0 a1 ∈A ℓ=0
(
−1 n1−γ/γ if γ 6= 1
≪γ,c ρ . n
log n if γ = 1
P  
Combining the upper and lower bounds on θ∈E ′ S(θ), we see that there exists a con-
stant K depending on the result in Lemma 3.2, γ, and c such that
(
δ 2 n2 ′ ′ −1 n1−γ/γ if γ 6= 1
|E | 6 n|E | + Kρ n .
2m log n if γ = 1

Dividing both sides by n and using the fact that m 6 δ2 n/4, we see that
 2  (
δ2 n ′ δ n n1−γ/γ if γ = 6 1
|E | 6 − 1 |E ′ | 6 Kρ−1 ,
4m 2m log n if γ = 1
which rearranges to the desired conclusion.

20
 The proof of Theorem A will feature oblique projections instead of orthogonal ones. The
following corollary concerns oblique projections and is stated in a way that will make it
immediately applicable in the proof of Theorem A.
Denote by Πt : R2 → R the oblique projection Πt (x, y) = x + ty. Let ϕ : (0, π/2) → R
be the diffeomorphism ϕ(θ) = log tan θ. Note that Πeϕ(θ) is the oblique projection that is the
“continuation” of the orthogonal projection πθ , meaning that the points (x, y), (Πeϕ(θ) (x, y), 0),
and πθ (x, y) are collinear.

Corollary 3.4. Let 0 < γ2 < γ3 < γ4 < γ5 be such that γ2 < 1 and
2(γ5 − γ3 ) < γ4 − γ2 . (3.3)
For all compact I ⊆ R, all ε, c1 , c2 , c3 > 0, all sufficiently small ρ > 0, and all (c1 ρ, γ5 )c2 -sets
A ⊆ [0, 1]2 with |A| > ρ−γ4 , there exists T ⊆ I with the following properties:
(I) the set I\T is covered by a disjoint union of finitely many half-open intervals of total
Lebesgue measure less than ε;
(II) for all t ∈ T and all A′ ⊆ A with |A′ | > ρ−γ3 , there exists a subset A′t ⊆ A′ with
|A′t | > ρ−γ2 such that the points of Πet A′t are distinct and c3 ρ-separated.


Proof. Let I ⊆ R be compact and ε, c1 , c2 , c3 > 0. Let σ ∈ γ5 − γ3 , (γ4 − γ2 )/2 . Let ρ > 0 be
sufficiently small (to be specified later). Let A ⊆ [0, 1]2 be a (c1 ρ, γ5 )c2 -set with |A| > ρ−γ4 .
Put γ5 = min(1, γ5 ), n = |A|, δ = σ −γ2 . Note that since A is a (c ρ, γ ) -set
√ ρ , and m = 2c−γ 3ρ 1 5 c2
contained in a ball of diameter 2, n 6 2c2 (c1 ρ) 5 .
We want to apply Theorem 3.3 with γ5 as γ, c1 ρ as ρ, c2 as c, and with A, n, δ, and m
as they are. We see that the inequality n > − log c2 holds for ρ sufficiently small, as does
m 6 δ2 n/4 since σ < (γ4 −γ2 )/2. Since the conditions of Theorem 3.3 hold, the set E ⊆ [0, π)
defined in (3.1) satisfies
m
N (E, ρ) ≪γ5 ,c2 ρ−1 2 n1−γ5 /γ5 log n
δ n
ρ−γ2
(3.4)
≪γ5 ,c1 ,c2 ,c3 ρ−1 2σ −γ γ /γ log ρ−γ5 .
ρ ρ 4 5 5
Let J = ϕ−1 (I), and put T = I\ϕ|J (E). Since the map ϕ|J is bi-Lipschitz,
N (ϕ|J (E), ρ) ≍I N (E, ρ).
Combining this with (3.4) and the fact that σ < (γ4 − γ2 )/2, we have that for sufficiently
small ρ, N (I\T, ρ) 6 ερ−1 /6. It follows that the set I\T can be covered by a disjoint union
of not more than ερ−1 /2-many half-open intervals of length ρ, a cover of total measure less
than ε. This establishes (I).
To prove (II), let t ∈ T , and let A′ ⊆ A with |A′ | > ρ−γ3 . Since n 6 2c2 (c1 ρ)−γ5
and σ > γ5 − γ3 , for sufficiently small ρ, ρ−γ3 > δn. It follows that |A′ | > δn. Because
θ := ϕ−1 (t) 6∈ E, N (πθ A′ , ρ) > m. It follows that N (πθ A′ , c3 ρ) > ρ−γ2 . By choosing points
in A′ in each fiber of a maximally ρ-separated set of the projection, we see that there exists a
subset A′t ⊆ A′ of cardinality at least ρ−γ2 such that the orthogonal projection of the points in
A′θ onto ℓθ are disjoint and c3 ρ-separated. Since the oblique projection Πet increases distances
between points that lie on ℓθ , the images of points of A′t under Πet are c3 ρ-separated.

The results in this section add to a number of other discrete Marstrand-type theorems in

21
the recent literature: [LM2, Lemma 5.2], [LM1, Prop. 3.2], [Gla, Lemma 3.8], [PS, Prop. 7],
to name a few. Let us highlight some distinguishing features of Lemma 3.2 and Theorem 3.3
that play an important role in this work. Analogues of Lemma 3.2 more commonly found
in the literature, such as the one in [Mat, Lemma 3.11], bound the measure of the set
of projections which map x close to 0. The result in Lemma 3.2 uses coverings to capture
topological information on the set of projections. This information is carried into Theorem 3.3
and is important in the application to Theorem A. Another useful feature of Theorem 3.3
is the allowance of a subset A′ in (3.1); this will allow us to treat sets in Theorem A that
exhibit multiplicative invariance without necessarily being self-similar.

3.2. Trees and a regularity result

Trees are combinatorial objects that are convenient for describing fractal sets. We will be
concerned solely with finite trees throughout this work. After giving the main definitions, we
motivate their importance by explaining how they will be used in the proof of Theorem A.
We move then to prove the main result in this section.
The following definitions describe the familiar notion of a rooted tree, a graph with no
cycles whose vertices can be arranged on levels and whose edges only connect vertices on
adjacent levels.
Definition 3.5.
• A tree of height N ∈ N0 is a finite set of nodes Γ together with a partition Γ =
Γ0 ∪ · · · ∪ ΓN with |Γ0 | = 1 and a parent function P : Γ\Γ0 → Γ\ΓN such that for every
n ∈ {1, . . . , N }, P (Γn ) = Γn−1 .
• The nodes in Γn have height n. The single node with height 0 is the root and the nodes
with height N are called leaves.
• The node Q is the parent of each of its children, nodes in the set CΓ (Q) := P −1 (Q).
N −n n+i
• If Q is a node of height n, the induced tree based at Q is the tree ΓQ := ∪i=0 CΓ (Q)
of height N − n with root Q and the same parent function as Γ, restricted to the set
ΓQ .
• A subtree of Γ is a tree Γ′ ⊆ Γ of the same height as Γ with parent function P |Γ′ \Γ′ .
0
(A subtree is uniquely determined by its non-empty set of leaves Γ′N ⊆ ΓN .)
Continuing with terminology inspired by genealogy trees, the ancestors of a node Q are
those nodes that lie between Q and the root. For the reasons described below in Remark 3.8,
it will be important to count the number of ancestors of Q that have many children. To this
end, we introduce the following terminology and notation.
Definition 3.6. Let Γ be a tree, c > 0, and ω ∈ [0, 1].
• The ancestry of Q ∈ Γn is the set
AΓ (Q) := {P k (Q) | 1 6 k 6 n}.
Note that |AΓ (Q)| is equal to the height of Q.
• The node Q is c-fertile if |CΓ (Q)| > c. The set of c-fertile ancestors of Q is denoted
FΓ,c (Q) := {A ∈ AΓ (Q) | A is c-fertile}.
A node Q has (c, ω)-fertile ancestry if |FΓ,c (Q)| > ω|AΓ (Q)|.
The following definitions allow us to capture the dimension of a finite tree by giving costs

22
to the nodes and measuring the cost of the least expensive cut.
Definition 3.7. Let Γ be a tree, r ∈ N, r > 2, and γ > 0.

• A cut of Γ is a subset C ⊆ Γ such that for every leaf L of Γ, {L} ∪ AΓ (L) ∩ C 6= ∅.
• The Hausdorff content of Γ with base r at dimension γ is
  
X  

Hrγ (Γ) := min r −height(Q)γ  C is a cut of Γ .
  
Q∈C

The main result in this section, Theorem 3.11, says, roughly speaking, that any tall
enough tree with Hausdorff content bounded from below and with a uniform upper bound on
the number of children of any node has a subtree in which most nodes have fertile ancestry.
Before making this statement precise and beginning with the details of the proof, let us make
two observations about the concept of fertile ancestry that will help explain why it will be
useful later on in the proof of Theorem A.
Remark 3.8.
(I) The property of having fertile ancestry is preserved under a type of tree thinning process
that we will employ in the proof of Theorem A. More specifically, suppose that Γ is a tree
in which every node has either one child or at least c many children and in which every
node has (c, ω)-fertile ancestry. Suppose further that for every node Q, there
exists
a subset C̃(Q) ⊆ CΓ (Q) of the children of Q with |C̃(Q)| > min c̃, |CΓ (Q)| . These
subsets naturally give rise to a subtree Γ̃ obtained by thinning the tree Γ: the subtree
Γ̃ is uniquely defined by the property that if Q is a node of Γ̃, then CΓ̃ (Q) = C̃(Q). It
is not hard to see that every node in Γ̃ has (c̃, ω)-fertile ancestry, regardless of how the
subsets of children C̃(Q) were chosen.
(II) A tree in which every node has fertile ancestry necessarily has large Hausdorff content.
This is a simple consequence of the mass distribution principle (or the max flow-min cut
theorem) for trees, the real analogue of which is stated in Lemma 2.6. More specifically,
let Γ be a tree, and consider a “flow” through Γ of magnitude 1 starting at the root
that splits equally amongst children. The value of the flow at any node Q with fertile
ancestry can be bounded from above using the fact that many times, much of the flow
is split amongst a large set of children before reaching Q. If all nodes of Γ have fertile
ancestry, then the flow is not concentrated too highly at any node. According to the
mass distribution principle, the Hausdorff content of a tree that supports such a flow
is high.
We now proceed with the main results in this subsection. In the next two results, fix
r > 2 and 0 < γ3 < γ4 < γ5 such that setting
A := γ5 − γ4 + logr 2,
B := γ4 − γ3 − logr 2,
ensures the quantity B is positive. The following lemma describes the fundamental dichotomy
behind Theorem 3.11.

Lemma 3.9. If Γ is a tree with the property that

every node in the tree has at most r γ5 many children, (3.5)

23
then at least one of the following holds:
(I) there are at least r γ3 many children Q of the root, each of which satisfies
Hrγ4 (ΓQ ) > Hrγ4 (Γ)r −A ;
(II) there is at least one child Q of the root satisfying
Hrγ4 (ΓQ ) > Hrγ4 (Γ)r B .

Proof. Let Γ be a tree satisfying (3.5). Let Q1 , Q2 , . . . , QI be the children of the root of
Γ, ordered so that Hrγ4 (ΓQi ) > Hrγ4 (ΓQi+1 ). If neither (I) nor (II) holds, then Hrγ4 (ΓQ1 ) <
Hrγ4 (Γ)r B and Hrγ4 (ΓQ⌈rγ3 ⌉ ) < Hrγ4 (Γ)r −A . It follows by the ordering of the Qi ’s and the
definition of the Hausdorff content and induced trees that
I
X
Hrγ4 (Γ) 6r −γ4
Hrγ4 (ΓQi )
i=1
⌊r γ3 ⌋ I
X X
−γ4
= r Hrγ4 (ΓQi ) + r −γ4 Hrγ4 (ΓQi )
i=1 i=⌈r γ3 ⌉
γ3 −γ4
<r r Hrγ4 (Γ)r B +r r γ5 −γ4
Hrγ4 (Γ)r −A = Hrγ4 (Γ),
a contradiction.

Lemma 3.10. Every finite tree Γ that satisfies (3.5) has a subtree Γ′ with the property that
for all nodes Q in Γ′ ,
|AΓ′ (Q)|B + logr Hrγ4 (Γ)
|FΓ′ ,rγ3 (Q)| > . (3.6)
A+B
Proof. We will prove the lemma by induction on the height N of the tree Γ. To verify the
base case, let Γ be the tree of height N = 0: a single node with no children. Taking Γ′ = Γ,
the inequality (3.6) for this single node follows from the fact that logr Hrγ4 (Γ) = 0.
Suppose that N ∈ N is such that the theorem holds for all trees of height N − 1. Let Γ
be a tree of height N that satisfies (3.5). By Lemma 3.9, at least one of Case (I) or Case (II)
holds.
Suppose Case (I) of Lemma 3.9 holds. Let Q be any one of the r γ3 -many children guar-
anteed by Case (I). By the induction hypothesis, there exists a subtree Γ′Q of ΓQ in which
every node satisfies (3.6) with ΓQ in place of Γ and Γ′Q in place of Γ′ . Define the subtree Γ′
of Γ to be the root node of Γ with the collection of at least r γ3 many children Q, each of
those children followed by its subtree Γ′Q .
We will now verify that (3.6) holds for all nodes of Γ′ . Let Q be any node of Γ′ . If Q
is the root node of Γ′ , then (3.6) holds because logr Hrγ4 (Γ) 6 0. If Q is a non-root node of
Γ′ , then it belongs to one of the subtrees Γ′S for some child S of the root of Γ′ . By property
(3.6) for the subsubtree Γ′S , we see
|FΓ′ ,rγ3 (Q)| − 1 = |FΓ′S ,rγ3 (Q)|
|AΓ′Q (Q)|B + logr Hrγ4 (ΓS )
>
A+B
(|AΓ′ (Q)| − 1)B + logr Hrγ4 (Γ) − A
> .
A+B

24
This simplifies to the inequality in (3.6), verifying the inductive step if Case (I) of Lemma 3.9
holds.
Suppose Case (II) of Lemma 3.9 holds. Let Q be the child guaranteed by Case (II). By
the induction hypothesis, there exists a subtree Γ′Q of ΓQ in which every node satisfies (3.6)
with ΓQ in place of Γ and Γ′Q in place of Γ′ . Define the subtree Γ′ of Γ to be the root of Γ
with only the child Q followed by its subtree Γ′Q .
We will now verify that (3.6) holds for all nodes of Γ′ . Let Q be any node of Γ′ . If Q is
the root node of Γ′ , then (3.6) holds because logr Hrγ4 (Γ) 6 0. If Q is a non-root node of Γ′ ,
then by property (3.6) for the subtree containing Q, we see
(|AΓ′ (Q)| − 1)B + Hrγ4 (Γ) + B
|FΓ′ ,rγ3 (Q)| > .
A+B
This simplifies to the inequality in (3.6), verifying the inductive step if Case (II) of Lemma 3.9
holds. The proof of the inductive step is complete, and the theorem follows.

Theorem 3.11. For all 0 < ε < 1, for all 0 < γ3 < γ4 < γ5 < γ4 + ε(γ4 − γ3 ), for all
sufficiently large r ∈ N, and for all V > 0, there exists N0 ∈ N for which the following holds.
For all N > N0 and for all trees Γ of height N with Hrγ4 (Γ) > V that satisfy (3.5), there exists
a subtree Γ′ of Γ such that all nodes Q ∈ Γ′ with height at least N0 have (r γ3 , 1 − ε)-fertile
ancestry in Γ′ .

Proof. Let 0 < ε < 1 and 0 < γ3 < γ4 < γ5 < γ4 + ε(γ4 − γ3 ). Let r ∈ N be sufficiently large
so that γ4 − γ3 − logr 2 > (1 − ε)(γ5 − γ3 ). Define A = γ5 − γ4 + logr 2 and B = γ4 − γ3 − logr 2,
and note by the inequality in the previous sentence, B/(A + B) > (1 − ε). Let V > 0. Choose
N0 ∈ N such that
N0 B + logr V
> 1 − ε, (3.7)
N0 (A + B)
and note that for all N > N0 , the inequality in (3.7) holds with N0 replaced by N .
Let N > N0 , and let Γ be a tree of height N with Hrγ4 (Γ) > V that satisfies (3.5). By
Lemma 3.10, there exists a subtree Γ′ of Γ such that for all nodes Q of Γ′ , the inequality in
(3.6) holds.
Let Q be a node of Γ′ with height at least N0 . By (3.6) and (3.7), we see that
|FΓ′ ,rγ3 (Q)| |AΓ′ (Q)|B + logr V
> > 1 − ε.
|AΓ (Q)|
′ |AΓ′ (Q)|(A + B)
It follows that Q has (r γ3 , 1 − ε)-fertile ancestry in Γ′ , as was to be shown.

3.3. A quantitative equidistribution lemma

The main result in this short section, Lemma 3.13, gives a lower bound on the number of
visits of an equidistributed sequence to a set as a function only of the measure and topological
complexity of the set’s complement. This result is certainly not new; we state it explicitly
here for convenience in a way that highlights the uniformity in the quantifiers.
For U ∈ N, denote by IU the collection of those subsets of [0, 1) that are a union of no
more than U disjoint intervals of the form [a, b).

25
Lemma 3.12. For any uniformly distributed sequence (xn )n∈N0 ⊆ [0, 1), U ∈ N, and ε > 0,
there exists N0 ∈ N such that for all N > N0 and all B ∈ IU ,
1  
{0 6 n 6 N − 1 | xn ∈ B} 6 Leb(B) + ε.
N
Proof. Let (xn )n∈N0 ⊆ [0, 1) be uniformly distributed, U ∈ N, and ε > 0. The discrepancy of
N −1
(xn )n=0 (cf. [KN, Def. 1.3]) is
 
 {0 6 n 6 N − 1 | xn ∈ I} 
DN = sup   − Leb(I) ,
I N
where the supremum is taken over all half-open intervals I in [0, 1). Because (xn )n is uniformly
distributed, DN → 0 as N → ∞. By the definition of discrepancy, for any half-open interval
I ⊆ [0, 1),
1  
{0 6 n 6 N − 1 | xn ∈ I} 6 Leb(I) + DN .
N
It follows that for every B ∈ IU ,
1  
{0 6 n 6 N − 1 | xn ∈ B} 6 Leb(B) + U DN .
N
Let N0 ∈ N be large enough so that for all N > N0 , U DN 6 ε. The conclusion follows.

Lemma 3.13. Let β > 0. For any uniformly distributed sequence (xn )n∈N0 ⊆ [0, β) with
respect to the Lebesgue measure, U ∈ N, and ε > 0, there exists N0 ∈ N such that for all
N > N0 and all J ⊆ [0, β) whose complement is covered by a union of no more than U many
disjoint, half-open intervals of total Lebesgue measure less than εβ/2,
1  
{0 6 n 6 N − 1 | xn ∈ J} > 1 − ε.
N
Proof. Let (xn )n∈N0 ⊆ [0, β) be uniformly distributed, U ∈ N, and ε > 0. Let N0 be from
Lemma 3.12 with (xn /β)n∈N0 , U , and ε/2.
Let N > N0 and J ⊆ [0, β). Put B = [0, β)\J, and note that by assumption, B/β ∈ IU
and Leb(B/β) < ε/2. It follows from Lemma 3.12 that
1  
{0 6 n 6 N − 1 | xn /β ∈ B/β} < ε.
N
Therefore,
1  
{0 6 n 6 N − 1 | xn ∈ J} > 1 − ε,
N
as was to be shown.

3.4. Outline of the proof of Theorem A

Before beginning with the details of the proof of Theorem A, we explain the main ideas behind
it. To understand the argument, it helps to begin by assuming that the set X×Y is self-similar
in the sense that for every n ∈ N0 , it is a union of approximately r n(dimH X+dimH Y ) many
′ ′
translates of the set r −n X × s−n Y . (Recall that n′ = ⌊n log r/ log s⌋ so that s−n ≈ r −n .)
This is the case, for example, if X and Y are both restricted digit Cantor sets. In this

26
case, Peres and Shmerkin [PS] proved that for all λ, η ∈ R\{0}, dimH (λX + µY ) = γ. Our
argument follows along the same lines as theirs.
Recall that Πt : R2 → R is the oblique projection Πt (x, y) = x + ty. A quick calculation
shows that
′
Πet (r −n X × s−n Y ) = r −n Πet rn /sn′ (X × Y ),
′
which implies that the images of the translates of r −n X × s−n Y under the map Πet are
affinely equivalent to the image of the full set X × Y under the map Πet rn /sn′ . It follows that
the set Πet (X × Y ) contains affine images of the sets Πet rn /sn′ (X × Y ) and hence that

dimH Πet (X × Y ) > sup dimH Πet rn /sn′ (X × Y ).

n∈N0

Thus, to bound dimH Πet (X × Y ) from below, it suffices to show that there is some n ∈ N0 for
′
which et r n /sn is a “good angle” for X × Y , in the sense that dimH Πet rn /sn′ (X × Y ) > γ − ε.
It follows from Marstrand’s theorem that the set of such “good angles” for X × Y (indeed,
′
for any set) has full measure in R, and it will be shown that the sequence n 7→ log(et r n /sn )
has image in [t, t + log s) and is the orbit of t under the irrational x 7→ x + log r (mod log s)
translated by t. When combined, these facts fall just short of allowing us to conclude the
′
existence of n ∈ N0 for which et r n /sn is a good angle: it is possible that the image of an
equidistributed sequence misses a set of full measure.
To make use of the above outline, one needs to gain some topological information on
the set of good angles from Marstrand’s theorem. This can be accomplished by moving the
argument to a discrete setting. Discretizing introduces a number of technical nuisances, but
the core of the argument remains the same. Recall that Xn and Yn′ are the sets X and Y
′
rounded to the lattices r −n Z and s−n Z, respectively. The discrete analogue of Marstrand’s
theorem in Theorem 3.3 tells us that the complement of the set of “good angles” for a finite set
such as Xn ×Yn′ can be covered by a disjoint union of few half-open intervals. This topological
information combines with the equidistribution of the irrational rotation described above to
′
allow us to find many n ∈ N0 for which et r n /sn is a good angle for Xn × Yn′ .
The argument described thus far is essentially due to Peres and Shmerkin in [PS] and
allows them to conclude that for all t ∈ R\{0}, dimH Πet (X × Y ) = γ. We will now de-
scribe the two primary modifications we make to this argument in the course of the proof of
Theorem A.
The first modification allows us to show that the discrete Hausdorff content of Πet (X ×Y )
at all small scales is uniform in t. Ultimately, this uniformity stems from the fact that the
irrational rotation described above is uniquely ergodic: changing t in the argument above
changes only the point whose orbit we consider. Exposing the uniformity in the argument
after this is then mainly a matter of taking care with the quantifiers in the auxiliary results.
The second modification allows us to handle sets X and Y which are only assumed to
be ×r- and ×s-invariant. Such sets need not be self-similar, but they do exhibit some “near
self similarity” in the following sense. Consider  the discrete set
Xm for some large m ∈ N.
Because X is ×r-invariant, the set X(n+1)m ∩ i/r nm , (i + 1)/r nm , when dilated by r mn and
considered modulo 1, is a subset of Xm . While this set is generally not equal to Xm , it is,
by an averaging argument, very often of cardinality greater than r −ε |Xm |. This is profitably
re-interpreted in the language of trees: in the tree with levels Xnm × Y(nm)′ , n ∈ N0 , many
nodes have nearly the maximum allowed number of children. The tree thinning result in

27
Theorem 3.11 exploits this abundance by finding a sufficiently “regular” subtree on which we
focus our attention. Then, we invoke our discrete analogue of Marstrand’s theorem – which
provides information on the set of angles that are good not only for the original set Xm × Ym′ ,
but also for large subsets of it – to further thin the subtree. Following the reasoning given in
Remark 3.8, the resulting subtree has fertile ancestry and hence has large Hausdorff content.
By the construction of the subtree, its image under Πet is large, and this yields the lower
bound on the Hausdorff dimension in the conclusion of the theorem.

3.5. Proof of Theorem A

In this section and the next, let r, s, X, Y , and γ be as given as in the statement of
Theorem A. The proof of Theorem A begins with a number of reductions, the last of which
in Claim 3.14 is a statement about the existence of measures on the images of the discrete
product sets under oblique projections. We prove Claim 3.14 in the next subsection.
By Lemma 2.10, dimH (X × Y ) = dimH X + dimH Y . Note that if dimH X = 0, then the
conclusion is clear by considering, for any x ∈ X, images of the set {x} × Y . The same is true
if dimH Y = 0. Thus, we will proceed under the assumption that dimH X, dimH Y > 0. Note
that the set 1 − X is ×r-invariant and that −λX + ηY is a translate of the set λ(1 − X) + ηY .
The analogous statement holds for Y . Combining these facts, it is easy to see that it suffices
to prove Theorem A in the case that I ⊆ (0, ∞).
The next step is to formulate a statement sufficient to prove Theorem A in terms of
oblique projections of discrete sets. Recall that n′ = ⌊n log r/ log s⌋ and that Xn , Yn′ are the
′
sets X and Y rounded to the lattices r −n Z and s−n Z, respectively. For n ∈ N0 , define
Qn = Xn × Yn′ and en = Xn × Yn′ +1 .
Q (3.8)

Claim 3.14. For all compact I ⊆ R and all 0 < γ < γ, there exists m, N0 ∈ N such that
for all N > N0 and all t ∈ I, there exists a probability measure µ supported on the finite set
Πet QN m with the property that for all balls B ⊆ R of diameter δ > r −N m , µ(B) 6 r N0 m δγ .

To deduce Theorem A from  Claim 3.14, let I ⊆ (0, ∞) be compact and 0 < γ < γ.
Apply Claim 3.14 with I ˜ := log(η/λ)  η, λ ∈ I as I and γ as it is. Let m, N0 ∈ N be as
guaranteed by Claim 3.14.
γ
Note that the limit in (1.8) is guaranteed to exist because the function ρ 7→ inf λ,η∈I H>ρ


λX + ηY is non-increasing (as ρ decreases) and is bounded from below by zero. Therefore,
to show that (1.8) holds, it suffices to prove that
γ

lim inf H>r −Nm λX + ηY > 0. (3.9)
N →∞ λ,η∈I

It follows from the fact that

dH (λXN m + ηY(N m)′ , λX + ηY ) ≪I,r,s r −N m
and Lemma 2.7 that for all λ, η ∈ I,
γ
γ

H>r −Nm λX + ηY ≍I,r,s H>r −Nm λXN m + ηY(N m)′

γ log(η/λ)

(3.10)
≍I,r,s H>r −Nm XN m + e Y(N m)′ .

28
Therefore, to show (3.9), it suffices to prove that
γ

lim inf H>r −Nm Πet QN m > 0. (3.11)
N →∞ t∈I˜

Combining the conclusion
of Claim 3.14 with Lemma 2.6, we see that for all N > N0 and
˜ Hγ −Nm Πet QN m > r −N0 m . This shows that the limit in (3.11) is positive and
t ∈ I, >r
completes the deduction of Theorem A from Claim 3.14.

3.6. Proof of Claim 3.14

Choosing the parameter m and scale ρ. Recall that r, s, X, Y , and γ are as given as
in the statement of Theorem A. Without loss of generality, we can assume that r < s. Put
β = log s, let 0 < γ < γ, and define ε := γ − γ and γ1 := γ.
We claim that there exist γ2 , γ3 , γ4 , and γ5 such that
(I) 0 < γ1 /(1 − ε/2) < γ2 < γ3 < γ4 < dimH X + dimH Y < γ5 ;
(II) γ5 < γ4 + ε(γ4 − γ3 )/6;
(III) γ2 < 1;
(IV) 2(γ5 − γ3 ) < γ4 − γ2 (this is the inequality in (3.3)).
To see why, note that if we put γ2 = γ1 /(1 − ε/2), γ4 = γ5 = dimH X + dimH Y , and
γ3 = γ2 /3 + 2γ4 /3, then the inequalities in (I) holds with “<” replaced by “6”, while the
inequalities in (II), (III), and (IV) hold as written. It follows that γ2 and γ5 can be increased
and γ4 can be decreased (with the corresponding change γ3 = γ2 /3 + 2γ4 /3) so that all of
the inequalities hold.
Let c1 and c2 be the constants guaranteed in the statement of Corollary 2.15. Let I ⊆
(0, ∞) be compact, and define Iβ = I + [0, β]. Let P > 0 be a Lipschitz constant for all of
the maps Πet , t ∈ Iβ , and let c3 = 4P s−1 + 1. Choose m ∈ N large enough so that we can
apply
• Theorem 3.11 with ε/6 as ε and r m as r;
• Corollary 3.4 with Iβ as I, εβ/12 as ε and r −m as ρ;
• Corollary 2.15 with m as N .
Put ρ = r −m .
′

A uniformly distributed sequence. Let α = log r m /sm and let R : [0, β) → [0, β) be
the transformation R : x 7→ x + α (mod β). As β = log s and m′ = ⌊m log r/ log s⌋, we have

α/β = m log r/ log s − m′ = m log r/ log s . (3.12)
Since log r/ log s is irrational, we conclude that α/β is irrational, whereby the sequence
(Rn (0))n∈N0 is uniformly distributed on [0, β).

Claim 3.15. For all n ∈ N0 ,

(V) Rn (0) + (nm)′ log s = nm log r; (
(VI)
′ (nm)′ + m′ if Rn (0) + α < β
(n + 1)m = .
(nm)′ + m′ + 1 if Rn (0) + α > β

Proof. Sincefor all n ∈ N, n n

 R (0)
 = nα (mod
β), using (3.12), we′ can write R (0)/β =
nα/β = n{m log r/β} = nm log r/β . Recalling that (nm) = ⌊nm log r/β⌋, this
establishes (V).

29
 Next, note that for any real numbers x, y,
(
⌊x⌋ + ⌊y⌋ if {x} + {y} < 1
⌊x + y⌋ = .
⌊x⌋ + ⌊y⌋ + 1 if {x} + {y} > 1

The equality in (VI)

 follows from
this by substituting x = nm log r/β and y = m log r/β and
n
using R (0)/β = nm log r/β and (3.12).

Choosing the parameter N0 . em (defined in

From Corollary 2.15, the sets Qm and Q
(3.8)) are (c1 ρ, γ5 )c2 -sets and satisfy
em | 6 ρ−γ5 .
ρ−γ4 6 |Qm |, |Q (3.13)
Let T1 (resp. T2 ) be the subset of Iβ obtained from applying Corollary 3.4 with Iβ as I,
εβ/12 as ε and Qm (resp. Q em ) as A. Put T = T1 ∩ T2 . It follows from Corollary 3.4 that
Iβ \T is covered by a disjoint union of finitely many half-open intervals of Lebesgue measure
less than εβ/6. Let U ∈ N be the number of half-open intervals which suffice to cover the set
Iβ \T .
Let N0 ∈ N be the larger of
γ4
• the N0 from Theorem 3.11 with ε/6 as ε, r m as r, and 2−γ4 H∞ (X × Y ) as V ;
• the N0 from Lemma 3.13 with (Rn (0))n∈N0 as (xn )n∈N0 and ε/3 as ε.

Fixing the parameters N and t. To prove Claim 3.14, we will show that for all N > N0
and all t ∈ I there exists a probability measure µ supported on the set Πet QN m with the
property that for all balls B ⊆ R of diameter δ > ρN , µ(B) 6 ρ−N0 δγ . Let N > N0 and
t ∈ I. From this point on, all new quantities and objects can depend on N and t.

Constructing the tree Γ. Let Γ be the tree (see Definition 3.5) of height N with node
′
set at height n ∈ {0, 1, . . . , N } equal to Qnm . Associating the point (i/r mn , j/s(mn) ) ∈ Qnm
with the rectangle
   
i i+1 j j+1
, × , ,
r mn r mn s(mn)′ s(mn)′
parentage in the tree Γ is determined by containment amongst associated rectangles.
Denote by CΓ (Q) the children of the node Q in Γ. Denote by ⊙ : R2 × R2 → R2 the
binary operation of pointwise multiplication.

Claim 3.16. Let n < N and Q ∈ Qnm .

′
(VII) If Rn (0) + α < β, then CΓ (Q) ⊆ Q + (r −nm , s−(nm) ) ⊙ Qm .
′
em .
(VIII) If Rn (0) + α > β, then CΓ (Q) ⊆ Q + (r −nm , s−(nm) ) ⊙ Q
γ4 γ
(IX) Hrm (Γ) > 2−γ4 H∞4 (X × Y ).

Proof. We first prove parts (VII) and (VIII). By Lemma 2.12, rXn
⊆ Xn−1 (mod 1) and
′
sYn′ ⊆ Yn′ −1 (mod 1). By (VI), if Rn (0) + α < β, then (n + 1)m = (nm)′ + m′ , and
′ ′
hence (r nm , s(nm) ) ⊙ Q(n+1)m ⊆ Qm (mod 1), and in particular (r nm , s(nm) ) ⊙ CΓ (Q) ⊆ Qm

′ ′
(mod 1). If Rn (0) + α > β, then (n + 1)m = (nm)′ + m′ + 1, and hence (r nm , s(nm) ) ⊙
Q(n+1)m ⊆ Q em (mod 1), and in particular (r nm , s(nm)′ ) ⊙ CΓ (Q) ⊆ Q
em (mod 1).

30
 ′
Write Q = (i/r nm , j/s(nm) ) and let Q′ ∈ CΓ (Q). Because Q′ is a child of Q, we can write
′ ′
Q′ = Q + (i0 /r (n+1)m , j0 /s((n+1)m) ) where 0 6 i0 < r m and 0 6 j0 < sm . It follows that
′ ′
(r nm , s(nm) )⊙(CΓ (Q)−Q) ⊆ Qm (in the first case Rn (0)+α < β) or (r nm , s(nm) )⊙(CΓ (Q)−
Q) ⊆ Q em (in the second case Rn (0) + α > β), where the containment now is understood
without reducing modulo 1.
To prove (IX), take a cut {Q1 , . . . , Qℓ } ⊆ Γ of Γ with node Qi at height ni . Then, by
construction of Γ, there exists a cover X × Y ⊆ ∪ℓi=1 Bi where ball Bi has diameter at most
2ρni . Since the cut was arbitrary, it follows that Hrγm4 γ4
(Γ) > 2−γ4 H∞ (X × Y ).

Constructing the tree Γ′ . Combining (3.13) with (VII) and (VIII), it follows that
|CΓ (Q)| 6 r mγ5 for every non-leaf node Q of Γ. The tree Γ has now been shown to satisfy
γ4
all the hypothesis of Theorem 3.11 (with ε/6 as ε, r m as r, and 2−γ4 H∞ (X × Y ) as V ), thus
′
there exists a subtree Γ of Γ with the property that every node with height at least N0 has
(r mγ3 , 1 − ε/6)-fertile ancestry in Γ′ .

Constructing the tree Γ′′ . Now we will use Corollary 3.4, the corollary to the discrete
version of Marstrand’s theorem, to further thin out the tree Γ′ ; an outline for this step was
described in Remark 3.8 (I). For each non-leaf node Q ∈ Γ′ , we will define a subset CΓm′ (Q)
of CΓ′ (Q). Define J = (T − t) ∩ [0, β). Since Iβ \T is covered by at most U many half-open
intervals of measure less than εβ/6, the same is true for the set [0, β)\J. Define J = {0 6 n 6
N −1 | Rn (0) ∈ J}. Note that for all n > N0 , by Lemma 3.13, |J ∩{0, . . . , n−1}| > (1−ε/3)n.
Let Q be a non-leaf node of Γ′ , and let n ∈ {0, . . . , N − 1} be the height of Q. Consider
the following cases:
(X) n 6∈ J or |CΓ′ (Q)| < ρ−γ3 . Select a single child Q′ of Q and put CΓm′ (Q) = {Q′ }.
(XI) n ∈ J , |CΓ′ (Q)| > ρ−γ3 , and Rn (0) + α < β. By Theorem 3.11 and (VII), the set
′
A′ := (r nm , s(nm) ) ⊙ (CΓ′ (Q) − Q) is a subset of Qm of cardinality at least ρ−γ3 . Since
n ∈ J , we have that t + Rn (0) ∈ T . Applying Corollary 3.4 (II) with t + Rn (0) in the
role of t, there exists a subset A′t ⊆ A′ with |A′t | > ρ−γ2 and such that the points of
′
Πet+Rn (0) A′t are distinct and c3 ρ-separated. Define CΓm′ (Q) = Q + (r −nm , s−(nm) ) ⊙ A′t
′
so that (r nm , s(nm) ) ⊙ (CΓm′ (Q) − Q) = A′t .
(XII) n ∈ J , |CΓ′ (Q)| > ρ−γ3 , and Rn (0)+α > β. We do exactly as in (XI) with Qm replaced
by Qem and using (VIII) to get the set C m′ (Q).
Γ
Let Γ be the subtree of Γ′ with the property that if Q is a non-leaf node of Γ′′ , then
′′

CΓ′′ (Q) = CΓm′ (Q). We claim that

every node of Γ′′ with height at least N0 has (r mγ2 , 1 − ε/2)-fertile ancestry. (3.14)
Indeed, let Q be a node of Γ′′ with height n > N0 . The ancestry of Q in Γ′ is (r mγ3 , 1 − ε/6)-
fertile. Each r mγ3 -fertile
 ancestor of Q in Γ′ with height in the set J is an r mγ2 -fertile ancestor
of Q in Γ . Since J ∩ {0, . . . , n − 1} > (1 − ε/3)n, there are at least (1 − ε/2)n many r mγ3 -
′′

fertile ancestor of Q in Γ′ with height in the set J . It follows that Q has (r mγ2 , 1− ε/2)-fertile
ancestry in Γ′′ .

Claim 3.17. If L1 and L2 are two distinct leaves of Γ′′ and n is maximal such that L1 and
L2 have a common ancestor at height n, then |Πet L1 − Πet L2 | > ρn+1 .

Proof. Let Q be the common ancestor of L1 and L2 in Γ′′ of height n. Note that by the
definition of Γ′′ and maximality of n, it must be that Q has more than one child and hence

31
that n ∈ J . Let Q1 and Q2 be the children of Q in Γ′′ that are ancestors of L1 and L2 ,
respectively. Note that Q1 6= Q2 but that Qi may be equal to Li .
We will show first that Πet Q1 and Πet Q2 are c3 ρn+1 -separated. Write Q = (p, q) and
Qi = (pi , qi ). Suppose that Rn (0) + α < β. It follows from (V) that
Πet Qi = r −nm (r nmΠet (Qi − Q)) + Πet Q
 n ′

= ρn r nm (pi − p) + et+R (0) s(nm) (qi − q) + Πet Q (3.15)
 ′

= ρn Πet+Rn(0) (r nm , s(nm) ) ⊙ (Qi − Q) + Πet Q.
′

By (XI), the points of Πet+Rn (0) (r nm , s(nm) ) ⊙ (Qi − Q) , i = 1, 2, are c3 ρ-separated. It
follows then from (3.15) that the points of Πet Qi , i = 1, 2, are c3 ρn+1 -separated. A similar
argument works to reach the same conclusion if Rn (0) + α > β using (XII).
By the definition of the Qnm sets, |Qi − Li | 6 2s−1 ρn+1 . By the triangle inequality and
the fact that c3 = 4P s−1 + 1,
       
Πet L1 − Πet L2  > Πet Q1 − Πet Q2  − Πet (Q1 − L1 ) − Πet (Q1 − L1 )
> (4P s−1 + 1)ρn+1 − 4P s−1 ρn+1 > ρn+1 .
 
It follows that Πet L1 − Πet L2  > ρn+1 , as was to be shown.

Constructing the measure µ. The proof of Claim 3.14 will be concluded by demon-
strating that 1) the fertile ancestry property of Γ′′ in (3.14) guarantees that Γ′′ supports
a “measure” which is not too concentrated on any node (an outline for this step was de-
scribed in Remark 3.8 (II)); and 2) by Claim 3.17, the projection of this measure is not too
concentrated on any ball.
Let ν : Γ′′ → [0, 1] be the unique function that takes 1 on the root of Γ′′ and has
the ′′ CΓ′′ (Q) and ν(Q) =
P properties that for all non-leaf nodes Q of Γ , ν is constant on ′′ and spreads down the
C∈CΓ′′ (Q) ν(C). (Colloquially, a mass of 1 begins at the root of Γ
tree by splitting equally amongst the children of each node.) Let νN be the function ν
restricted to Γ′′N , the set of leaves of Γ′′ . By the defining properties of ν, the function νN is
a probability measure on Γ′′N .
Since Γ′′N ⊆ QN m , the measure µ = Πet νN , the push-forward of νN through the map
Πet , is a probability measure supported on the set Πet QN m . We will conclude the proof of
Claim 3.14 by verifying that for all balls B ⊆ R of diameter δ > ρN , µ(B) 6 ρ−N0 δγ1 . (Recall
that γ1 = γ.)
Let B ⊆ R be an interval of length δ > ρN . Put n = ⌊logρ δ⌋ + 1 and note that
ρ < δ 6 ρn−1 . It follows from Claim 3.17 that there exists a node Q of Γ′′ with height at
n

least n with the property that if L is a leaf of Γ′′ with Πet L ∈ B, then Q is an ancestor of L.
This implies that µ(B) 6 ν(Q), and so it suffices to show that
ν(Q) 6 ρ−N0 δγ1 . (3.16)
If n 6 N0 , then ρ−N0 δγ1 > 1 and (3.16) holds trivially. If n > N0 , then by the definition
of ν and the fact that Q has (r mγ2 , 1 − ε/2)-fertile ancestry (cf. (3.14)),
1
ν(Q) 6 = ργ2 (1−ε/2)n 6 ρ−N0 δγ1 ,
r mγ2 (1−ε/2)n
since (1 − ε/2)γ2 > γ1 . This verifies (3.16), completing the proof of Claim 3.14 and hence of

32
Theorem A.

4. Multiplicatively invariant subsets of the non-negative

integers
This section is devoted to the study of ×r-invariant sets in the non-negative integers (see
Definition 1.5) and contains the proofs of Theorems B, C, D, and E.
As was already mentioned in Section 1.2.2, ×r-invariant subsets of N0 are closely related
to symbolic subshifts in r symbols. This connection is explored in more detail in Section 4.1,
where we establish numerous fundamental properties about ×r-invariant sets, including the
existence of their mass dimension (see Proposition 4.3). Thereafter, in Section 4.2, we show
that the study of ×r-invariant sets in the integers also connects to fractal geometry in the
unit interval [0, 1]. Among other things, we describe a natural way of identifying ×r-invariant
subsets of N0 with ×r-invariant subsets of [0, 1] which respects the Hausdorff content at every
finite scale (see Proposition 4.6). This plays a crucial role in the derivations of Theorems
C and E from their continuous counterparts in the later subsections. In Section 4.3 we
give an elementary and self-contained proof of Theorem B. Finally, Theorem C is proven in
Section 4.4, and Theorems D and E are proven in Section 4.5.

4.1. Connections to symbolic dynamics

Throughout this subsection, we use σ to denote the left-shift on {0, 1, . . . , r − 1}N0 , which is
defined by
σ : (wn )n∈N0 7→ (wn+1 )n∈N0 .
In the language of symbolic dynamics, any closed subset of {0, 1, . . . , r − 1}N0 satisfying
σ(Σ) ⊆ Σ is called a subshift. The language set associated to a subshift Σ is the set of all the
finite words appearing in the elements of Σ, i.e.,
 
L(Σ) = (w0 , w1 , . . . , wk )  (w0 , w1 , w2 , . . .) ∈ Σ, k ∈ N0 .
The language set of any subshift can be naturally embedded into the integers by identifying
finite words in the alphabet {0, 1, . . . , r − 1} with the base-r expansion in the integers. This
gives rise to the following definition.
Definition 4.1. The r-language set associated to a subshift Σ ⊆ {0, 1, . . . , r − 1}N0 is the
set AΣ ⊆ N0 defined by
 
AΣ = {0} ∪ w0 + w1 r + . . . + wk−1 r k−1  (w0 , w1 , . . . , wk−1 ) ∈ L(Σ) .
Note that the set AΣ is a ×r-invariant subset of N0 (recall Definition 1.5). Indeed,
the invariance of AΣ under the map Rr follows directly from the shift-invariance of Σ,
whereas the invariance of AΣ under Lr follows from the simple observation that for any word
(w0 , w1 , . . . , wk ) ∈ L(Σ), all of the prefixes (w0 , w1 , . . . , wi ), i 6 k, belong to L(Σ). Therefore,
the notion of r-language sets provides us with a natural way of constructing from any subshift
of the full symbolic shiftspace in r letters a ×r-invariant subsets of the non-negative integers.
Examples 4.2.
• Let F be an arbitrary collection of finite words from the alphabet {0, 1, . . . , r − 1}.

33
 The shift of finite type with forbidden words F is the subshift of {0, 1, . . . , r − 1}N0
consisting of all infinite words that do not contain a word from F as a sub-word. The
r-language sets corresponding to shifts of finite type form a natural class of ×r-invariant
subsets of N0 which contains the class of restricted digit Cantor sets defined in (1.6)
as a rather special subclass. Integer sets corresponding to subshifts of finite type were
also considered by Lima and Moreira in [LM2].
• The classical golden mean shift is the subshift of {0, 1}N0 consisting of all binary se-
quences with no two consecutive 1’s. This leads to a natural example of a ×2-invariant
set Agolden ⊆ N0 consisting of all integers whose binary digit expansion does not contain
two consecutive 1’s.√ Since the topological entropy of the golden mean shift is known
the equal log((1 + 5)/2) (cf. [LM3, Example 4.1.4]), it √ follows from Proposition 4.3
below that the mass dimension of Agolden equals log((1 + 5)/2)/ log 2.
• The even shift is the subshift of {0, 1}N0 consisting of all binary sequences so that
between any two 1’s there are an even number of 0’s. The corresponding ×2-invariant
set Aeven ⊆ N0 consists of all integers whose binary digit expansion has an even number
of 0’s between any two 1’s. Since the topological entropy of the golden mean shift
coincides with the topological entropy of the even shift (cf. [LM3, Example 4.1.6]), we
conclude that Aeven and Agolden have the same mass dimension.
• The prime gap shift is the subshift of {0, 1}N0 consisting of all binary sequences such
that there is a prime number of 0’s between any two 1’s. This corresponds to the
×2-invariant set Aprime ⊆ N0 of all those numbers written in binary in which there
is a prime number of 0’s between any two 1’s. For example, the first 17 elements of
Aprime are: 0, 1, 2, 4, 8, 9, 16, 17, 18, 32, 34, 36, 64, 65, 68, 72, 73. The entropy of the prime
gap shift is approximately 0.30293, (cf. [LM3, Exercise 4.3.7]) which implies that the
dimension of Aprime is approximately 0.437.
As we have observed, every r-language set is a ×r-invariant set. The converse is also
true: Every ×r-invariant subset A ⊆ N0 coincides with the r-language set associated to
some subshift Σ ⊆ {0, 1, . . . , r − 1}N0 . Additionally, the dimension of A coincides with the
normalized topological entropy of (Σ, σ).

Proposition 4.3. For any ×r-invariant set A ⊆ N0 , there exists a closed and shift-invariant
set Σ ⊆ {0, 1, . . . , r − 1}N0 such that A coincides with AΣ , the r-language set associated to
Σ. Moreover, the mass dimension of A exists and equals the normalized topological entropy
of the symbolic subshift (Σ, σ), i.e.,
htop (Σ, σ)
dimM A = .
log r
We remark that Proposition 4.3 is a generalization of some of the results in [LM2, Section
3], where subsets of integers arising from shifts of finite type are defined and studied. Also
note that Proposition 4.3 only gives the existence of the mass dimension, not the discrete
Hausdorff dimension, for ×r-invariant sets. The discrete Hausdorff dimension of such sets
also always exists, but this is only proved in the next subsection using different methods (see
Proposition 4.6 below). Finally, we remark that the identification of ×r-invariant subsets
of N0 and subshifts of {0, 1, . . . , r − 1}N0 given by Proposition 4.3 is not bijective. In fact,
for every ×r-invariant set A ⊆ N0 , there exists an infinite family of subshifts Σ such that
A = AΣ .

34
Proof of Proposition 4.3. Let Σ(k) denote the set of all infinite words w = (w0 , w1 , . . .) ∈
{0, 1, . . . , r − 1}N0 for which w0 + w1 r + . . . + wk r k is an element of A, and define
\
Σ := Σ(k) .
k∈N0

Being an intersection of closed sets, Σ is closed. From Rr (A) ⊆ A, it follows that σ(Σ) ⊆ Σ,
which proves that (Σ, σ) is a subshift. From the construction it is clear that the language
set AΣ of Σ is contained in A. On the other hand, if a = w0 + · · · + wk r k ∈ A then,
using Lr (A) ⊆ A, it follows that the word (w0 , . . . , wk , 0, 0, . . . ) belongs to Σ. It follows that
a ∈ AΣ , showing that A = AΣ .
It remains to verify that dimM A = htop (Σ, T )/ log r. Let LN (Σ) denote the set of words
of length N appearing in the language set L(Σ), i.e.,
 
LN (Σ) = (w0 , w1 , . . . , wN −1 )  (w0 , w1 , w2 , . . .) ∈ Σ .
It is well known (see, for instance, [Wal, Theorem 7.13 (i)]) that the topological entropy of
(Σ, σ) is given by
1
htop (Σ, σ) = lim log |LN (Σ)|,
N →∞ N
where the limit as N → ∞ on the right hand side is known to exist. Since |LN (Σ)| =
|A ∩ [0, r N )|, we have
log |LN (Σ)| log |A ∩ [0, r N )|
= . (4.1)
N N
It follows that
log |A ∩ [0, r N )|
htop (Σ, σ) = lim = (log r)(dimM A),
N →∞ N
which implies dimM A = htop (Σ, σ)/ log r.

As a corollary of Proposition 4.3 we obtain the following result, which plays an important
role in our proof of Theorem D.

Corollary 4.4. For any ×r-invariant A ⊆ N0 , the set

\ \
A′ := Rkr Lℓr (A)
k∈N0 ℓ∈N0

satisfies Rr (A′ ) = Lr (A′ ) = A′ (in particular, A′ is ×r-invariant) and dimM A′ = dimM A.

Proof. Note that A′ is the largest subset of A satisfying Rr (A′ ) = Lr (A′ ) = A′ ; in particular,
it is ×r-invariant. Therefore, to prove dimM A′ = dimM A, it suffices to find any subset
A′′ ⊆ A satisfying Rr (A′′ ) = Lr (A′′ ) = A′′ and dimM A′′ = dimM A. If dimM A = 0, then
there is nothing to show, so let us proceed under the assumption that dimM A > 0.
According to Proposition 4.3, we can find a closed and shift-invariant set Σ ⊆ {0, 1, . . . , r−
1}N0 such that A coincides with the r-language set AΣ associated to Σ. Let µ be an ergodic σ-
invariant Borel probability measure on Σ of maximal entropy (the existence of such a measure
follows from, eg. [Wal, Theorem 8.2 + Theorem 8.7 (v)]). Let Σ′′ denote the support of µ
and observe that (Σ′′ , σ) is a subshift of (Σ, σ) with htop (Σ, σ) = htop (Σ′′ , σ). Moreover, since

35
µ is ergodic, almost every point in Σ′′ has a dense orbit (by Birkhoff’s ergodic theorem) and
almost every point is recurrent (by Poincaré’s recurrence theorem). Therefore there exists a
point x ∈ Σ′′ which visits every non-empty open set in Σ′′ infinitely often.
Let A′′ ⊆ N0 be the r-language set associated to Σ′′ . Since Σ′′ ⊆ Σ, we have A′′ ⊆ A. Also,
dimM A = htop (Σ, σ) log r, dimM A′′ = htop (Σ′′ , σ) log r, and htop (Σ, σ) = htop (Σ′′ , σ), which
implies dimM A = dimM A′′ . All that remains to be shown is that Rr (A′′ ) = Lr (A′′ ) = A′′ .
Since A′′ is a r-language set, we already have the inclusions
Rr (A′′ ) ⊆ A′′ and Lr (A′′ ) ⊆ A′′ .
To prove the inclusion A′′ ⊆ Rr (A′′ ), let n ∈ A′′ be arbitrary. We can write this number as n =
w0 +w1 r+. . .+wk r k where (w0 , w1 , . . . , wk ) is a word in the language L(Σ′′ ). Since the point x
visits every open set of Σ′′ infinitely often, the word (w0 , . . . , wk ) appears in x infinitely often.
Therefore, there exists a letter u ∈ {0, 1, . . . , r − 1} such that the word (u, w0 , w1 , . . . , wk )
appears in x and hence belongs to L(Σ′′ ). Take n1 := u + w0 r + w1 r 2 + . . . + wk r k+1 and note
that n1 ∈ A′′ and that Rr (n1 ) = n, which proves n ∈ Rr (A′′ ).
Finally, to prove the inclusion A′′ ⊆ Lr (A′′ ), let n = w0 + w1 r + . . . + wk r k ∈ A′′
be arbitrary. Since htop (Σ′′ , σ) > 0, some non-zero letter must appear in L(Σ′′ ). In-
voking again the fact that x visits every open set infinitely often, a word of the form
(w0 , w1 , . . . , wk , v1 , v2 , . . . , vℓ ) must appear in x (and hence belong to L(Σ′′ )), where v1 , v2 ,
. . . , vℓ ∈ {0, 1, . . . , r − 1} for some ℓ ∈ N, with vℓ 6= 0 but vi = 0 for all i < ℓ. Defining
n2 := w0 + w1 r + . . . + wk r k + v1 r k+1 + . . . + vℓ r k+ℓ ∈ A′′ , we see that Lr (n2 ) = n and hence
n ∈ Lr (A′′ ).

Here is another corollary of Proposition 4.3, which may be of independent interest.

Corollary 4.5. If A ⊆ N0 is ×r-invariant and dimM A = 1, then A = N0 .

Proof. Suppose A is ×r-invariant with dimM A = 1. There exists a closed and shift-invariant
set Σ ⊆ {0, 1, . . . , r − 1}N0 such that A coincides with the r-language set coming from Σ and
htop (Σ, σ) = log r. However, the only subshift of {0, 1, . . . , r − 1}N0 with full entropy is the
full shift. Hence Σ = {0, 1, . . . , r − 1}N0 , which implies A = N0 .

4.2. Connections to fractal subsets of the reals

The purpose of this subsection is to establish a connection between ×r invariant subsets of
the non-negative integers and ×r-invariant subsets of [0, 1]. Recall that X ⊆ [0, 1] is called
×r-invariant if it is closed and Tr X ⊆ X, where Tr : x 7→ rx mod 1.
First, we remark that every ×r-invariant subset of [0, 1] can be “lifted” to a ×r-invariant
subset of N0 . Indeed, if X ⊆ [0, 1] is ×r-invariant, then one can show that the set
 k
⌊r x⌋ | x ∈ X, k ∈ N0
is ×r-invariant. We will not make use of this fact, so we leave the details to the interested
reader. Of more importance to us is the converse direction, stated in the following proposition.
Recall from Section 2.1 the definition of Hausdorff distance.

Proposition 4.6. For any ×r-invariant set A ⊆ N0 , the sequence Xk := (A ∩ [0, r k ))/r k
converges with respect to the Hausdorff metric dH as k → ∞ to a ×r-invariant set X ⊆ [0, 1]

36
satisfying dimH X = dimM A = dimH A. In particular, the discrete Hausdorff dimension of A
exists and equals the discrete mass dimension of A.

Remark 4.7. It follows from Proposition 4.6 that for a ×r-invariant set A ⊆ N0 , dimM A =
dimH A. Therefore Proposition 4.3, Corollary 4.4, and Corollary 4.5 remain true when the
mass dimension dimM is replaced by the discrete Hausdorff dimension dimH .
For the proof of Proposition 4.6 we will need two technical lemmas.

Lemma 4.8. Let A ⊆ N0 , and define Xk := (A ∩ [0, r k ))/r k .

(I) If Rr (A) ⊆ A, then for any k, l ∈ N with l > k, we have Xl ⊆ [Xk ]r−k .
(II) If Rr (A) ⊇ A, then for any k, l ∈ N with l > k, we have Xk ⊆ [Xl ]r−k .
In particular, if Rr (A) = A then for all l > k, we have dH (Xl , Xk ) 6 r −k .

Proof. It is helpful to note first that for all n, l, k ∈ N with l > k,

 l−k

n 
 − Rr (n)  6 1 . (4.2)
 rl r k  rk

This inequality follows easily from the fact that Rl−k r (n) = ⌊n/r
l−k ⌋. For the proof of

part (I), let y ∈ Xl and write y = m/r l for some m ∈ A. Note that m̃ := Rl−k r (m) belongs to
A ∩ [0, r k ) because Rr (A) ⊆ A. Then, setting ỹ := m̃/r k , we see that ỹ ∈ Xk and, by (4.2),
d(y, ỹ) 6 r −k . This proves Xl ⊆ [Xk ]r−k .
Next, we prove part (II). For any x ∈ Xk we can find n ∈ A ∩ [0, r k ) such that x = n/r k .
Since A ⊆ Rl−k l
r (A), there exists ñ ∈ A ∩ [0, r ) such that

Rl−k
r (ñ) = n.

Now x̃ := ñ/r l belongs to Xl and it follows from (4.2) that d(x, x̃) 6 r −k . This proves
Xk ⊆ [Xl ]r−k .
T
Lemma 4.9. Suppose A ⊆ N0 satisfies Rr (A) ⊆ A, and define A′ := k∈N Rkr (A). Also, set
Xk := (A ∩ [0, r k ))/r k and Xk′ := (A′ ∩ [0, r k ))/r k . Then limk→∞ dH (Xk , Xk′ ) = 0.

Proof. Let ε > 0, and let m ∈ N such that 2r −m < ε. Since Rr (A) ⊆ A, we have
A ∩ [0, r m ) ⊇ Rr (A) ∩ [0, r m ) ⊇ R2r (A) ∩ [0, r m ) ⊇ R3r (A) ∩ [0, r m ) ⊇ . . . .
The sequence k 7→ Rkr (A) ∩ [0, r m ) eventually stabilizes. This happens exactly when Rkr (A) ∩
[0, r m ) = A′ ∩ [0, r m ), or equivalently, when Rkr (A) ∩ [0, r m ) = Rkr (A′ ) ∩ [0, r m ), because
Rkr (A′ ) = A′ . It follows from (4.2) that the Hausdorff distance between Xk and (Rrk−m (A) ∩
[0, r m ))/r m is bounded from above by r −m . The same holds for Xk′ and (Rrk−m (A′ ) ∩
[0, r m ))/r m . Therefore, for all k > m for which Rrk−m (A) ∩ [0, r m ) = Rrk−m (A′ ) ∩ [0, r m ), we
have dH (Xk , Xk′ ) 6 2r −m by the triangle inequality. Since there exists a cofinite set of k for
which Rrk−m (A)∩[0, r m ) = Rrk−m (A′ )∩[0, r m ), we conclude that lim supk→∞ dH (Xk , Xk′ ) < ε.
The conclusion follows since ε > 0 was arbitrary.
T
Proof of Proposition 4.6. Define A′ := k∈N0 Rkr (A) and Xk′ := (A′ ∩ [0, r k ))/r k . In view of
Lemma 4.9, the sequence k 7→ Xk converges with respect to the Hausdorff metric if and only
if the sequence k 7→ Xk′ converges. Since A′ = Rr (A′ ), it follows from Lemma 4.8 that
dH (Xk′ , Xl′ ) 6 r −k , for all k, l ∈ N with l > k.

37
This implies that k 7→ Xk′ is a Cauchy sequence, and hence it is convergent (recall that by
the Blaschke selection theorem, the set of all non-empty, compact subsets of [0, 1] equipped
with the Hausdorff distance is a complete metric space).
Next, let us show that X is ×r-invariant. Since Lr (A) ⊆ A, a simple computation shows
Tr (Xk ) ⊆ Xk−1 . Therefore, using X = limk→∞ Xk , we get that Tr (X) ⊆ X.
Finally, we have to show dimH X = dimH A = dimM A. It follows form Theorem 2.9 that
the Minkowski and Hausdorff dimensions of ×r-invariant subsets of [0, 1] coincide. In other
words, we have
dimM X = dimH X. (4.3)
It therefore suffices to show that
dimM X = dimM A, (4.4)
dimH X = dimH A. (4.5)
We begin with (4.4). As guaranteed by Corollary 4.4, dimM A = dimM A′ . By combining
part (I) of Lemma 4.8 with Lemma 2.7, we see that


!
log N Xk , r −k log N X, r −k
0 6 lim inf −
k→∞ k log r k log r

(4.6)
log N X, r −k
= dimM A − lim sup ,
k→∞ k log r
1 −k ) (cf.
where the equality follows from the fact that dimM A = limk→∞ k log r log N (Xk , r
equation (2.6)). On the other hand, using part (II) of Lemma 4.8, Lemma 2.7, and the fact
1
that dimM A′ = limk→∞ k log ′
r log N (Xk , r
−k ), we see

!
log N X ′ , r −k log N Xk′ , r −k
0 6 lim inf −
k→∞ k log r k log r

(4.7)
log N X ′ , r −k
= lim inf − dimM A′ .
k→∞ k log r
Combining (4.6) and (4.7) with the fact that X ′ ⊆ X, we see
′ , r −k

−k


log N X log N X, r
dimM A′ 6 lim inf 6 lim sup 6 dimM A.
k→∞ k log r k→∞ k log r
Since dimM A = dimM A′ and X ′ ⊆ X, we conclude that dimM X exists and is equal to
dimM A.
Next, let us turn to the proof of (4.5). In view of (4.3), (4.4), and Lemma 2.18, instead
of (4.5) it suffices to show
dimH X 6 dimH A. (4.8)
Note, however, that dimH A′ 6 dimH A because A′ ⊆ A, and that dimH X = dimH X ′ because
dimM A = dimM A′ . Also, we have dimH X = dimM X = dimM A and dimH X ′ = dimM X ′ =
dimM A′ , where dimH X ′ = dimM X ′ follows from the fact that X ′ is also ×r-invariant and
dimM X ′ = dimM A′ follows from the work above with A′ in place of A and X ′ in place of X.

38
This means that (4.8) will follow from
dimH X ′ 6 dimH A′ . (4.9)
Next, recall from equation (1.7) that
  


′
dimH X = sup γ > 0  lim Hγ X ′ > 0 . (4.10)
 ρ→0+ >ρ

Also, by definition,
  


′
dimH A = sup γ > 0  lim inf Hγ A′ ∩ [0, N ) /N γ > 0 ,
 N →∞ >1

which in view of part (IV) of Lemma 2.17 and equation (2.7) can be rewritten as
  
 γ

′  ′
dimH A = sup γ > 0  lim inf H>r−k Xk > 0 . (4.11)
k→∞

By combining part (I) of Lemma 4.8 with Lemma 2.7 we see that
γ ′

γ ′


H>r −k X ≪ H>r −k Xk ,

which togehter with (4.10) and (4.11) implies (4.9).

4.3. Proof of Theorem B

In this subsection we give a proof of Theorem B. For w = (w0 , . . . , wℓ−1 ) ∈ {0, . . . , r − 1}ℓ ,
define
(w)r := w0 r ℓ−1 + w1 r ℓ−2 + · · · + wℓ−2 r 1 + wℓ−1 .
We will say that a non-negative integer n begins with w in base r if there exists d ∈ N0 and
n0 ∈ [0, r d ) such that
n = (w)r r d + n0 . (4.12)
If w0 6= 0, this means that the n most significant digits in the base-r expansion of n are w0 ,
w1 , . . . , wℓ−1 , in order.

Lemma 4.10. For all w ∈ {0, . . . , r − 1}ℓ , there is an arc Iw ⊆ [0, 1) modulo 1 (meaning
that I is an interval when 0 and 1 are identified) with the property that for all x > (w)r , if
{log x/ log r} ∈ Iw , then ⌊x⌋ begins with w in base r.

Proof. Let w ∈ {0, . . . , r − 1}ℓ+1 . It follows from (4.12) that a positive integer n begins with
w in base r if and only if there exists d ∈ N0 such that
(w)r r d 6 n < (w)r + 1)r d .
Therefore, a positive real number x has the property that ⌊x⌋ begins with w in base r if and
only if
(w)r r d 6 x < (w)r + 1)r d .

39
The previous inequality is equivalent to


log(w)r log x log (w)r + 1
+d6 < + d. (4.13)
log r log r log r
Let Iw be the

modulo 1 arc from the fractional part of log(w)r / log r to the
 fractional part of
log (w)r +1 / log r in the positive direction. We see that if x > (w)r and log x/ log r ∈ Iw ,
then (4.13) holds, and ⌊x⌋ begins with w in base r.

Lemma 4.11. Let r and s be multiplicatively independent positive integers, and let A ⊆ N0
be ×s invariant and infinite. For all w ∈ {0, . . . , r − 1}ℓ , there exists an element of A that
begins with w in base r.

Proof. Let w ∈ {0, . . . , r−1}ℓ , and let δ be half the length of the interval Iw from Lemma 4.10.
Define K0 = ⌈log(w)  r / log s⌉ and α = log s/ log r. Since α is irrational, there exists K ∈ N
such that the set {iα} | i ∈ {K0 , . . . , K} is δ-dense in [0, 1).
Since A is infinite, there exists n ∈ A such that k := ⌊log n/ log s⌋ > K. Since A is
Rs -invariant, n, ⌊n/s⌋, . . . , ⌊n/sk ⌋ are all elements of A. Define x = n/sk , and note that
⌊x⌋, ⌊sx⌋, . . . , ⌊sk x⌋ is the same list of integers which are, therefore, all elements of A.
We will show that there exists 0 6 i 6 k for which si x > (w)r and {log(si x)/ log r} ∈ Iw .
It will follow by Lemma 4.10 that ⌊si x⌋ is an element of A that begins with w in base r.
Note that log(si x)/ log r = iα + log x/ log r. Since {iα} | i ∈ {K0 , . . . , K} is δ-dense
in [0, 1) and Iw is an interval of length equal to 2δ, there exists i ∈ {K0 , . . . , K} such that
{log(si x)/ log r} ∈ Iw . Since i > K0 , we have that si x > (w)r . We have found 0 6 i 6 k for
which si x > (w)r and {log(si x)/ log r} ∈ Iw , as was to be shown.

Proof of Theorem B. Let A ⊆ N0 be simultaneously ×r- and ×s-invariant, and suppose A

is infinite. To show that A = N0 , we will show that for all w ∈ {0, . . . , r − 1}ℓ , (w)r ∈ A.
Let w ∈ {0, . . . , r − 1}ℓ . By Lemma 4.11, there exists n ∈ A that begins with w in base r.
Writing n as in (4.12), we see that Rdr (n) = (w)r . Since A is Rr -invariant, (w)r ∈ A, as was
to be shown.

4.4. Proof of Theorem C

In this section, we will prove Theorem C. The strategy is to use tools from Section 4.2 to
derive Theorem C from the theorem of Lindenstrauss-Meiri-Peres, Theorem 1.4. Throughout
this section, r > 2 is fixed and all of the asymptotic notation may implicitly depend on it.
Remark 4.12. There are some useful remarks to make before the proof. Let X1 , X2 , . . . ,
Xn ⊆ [0, 1] be ×r-invariant sets. The sumset X1 + · · · + Xn may be interpreted in R/Z or
in R. Denote temporarily by Wn the set X1 + · · · + Xn interpreted modulo 1 as a subset of
[0, 1] and by Yn the set X1 + · · · + Xn interpreted in R as a subset of [0, n]. Two facts of
particular relevance to us are: 1) set Wn is ×r-invariant, and 2) dimH Wn = dimH Yn . The
first fact follows easily from the fact that multiplication by r is a group endomorphism of
(R/Z, +). (In contrast, note that the sumset of ×r-invariant subsets of N0 is not necessarily
×r-invariant: if A is the base-10 restricted digit Cantor set with allowed digits 0 and 5, then
A + A contains 10 but does not contain R10 (10) = 1, for
example). The second fact follows
n−1
immediately by writing Wn = ∪i=0 (Yn ∩ [i, i + 1]) − i and using the translation-invariance
and finite (countable) stability under unions of the Hausdorff dimension.

40
Proof of Theorem C. Recall that (Ai )∞ i=1 is a sequence of ×r-invariant subsets of N0 . For
′
each i ∈ N, let Ai be the set described in Corollary 4.4, and define Xi ⊆ [0, 1] to be the
′ N N ∞
P∞ (Ai ∩ [0, r )/r )N =1 as in Proposition 4.6. Since
Hausdorff limit of the sequence P dimH Xi =
dimH Ai = dimH Ai and i=1 dimH Ai /| log dimH Ai | diverges, we have that ∞
′
i=1 dimH Xi /
| log dimH Xi | diverges. It follows by Theorem 1.4 that


lim dimH X1 + · · · + Xn = 1. (4.14)
n→∞

According to Remark 4.12, we can and will interpret the sum X1 + · · · + Xn to be in R.

We claim now that for all n ∈ N, the discrete Hausdorff dimension of the set A′1 + · · · + A′n
exists and



dimH A′1 + · · · + A′n = dimH X1 + · · · + Xn . (4.15)
′
Combined with (4.14), this
′ ′

suffices to conclude the
proof of Theorem C since Ai ⊆ Ai implies
that dimH A1 + · · · + An 6 dimH A1 + · · · + An .
To show (4.15), let n ∈ N, and define k = ⌊log n/ log r⌋ + 1. Define Bn = A′1 + · · · + A′n
and Yn = X1 + · · · + Xn , where the sum defining Yn is understood to be in R. Note that for
all N > k,
n
X n
A′i ∩ [0, r N −k ) Bn ∩ [0, r N ) X A′i ∩ [0, r N )
⊆ ⊆ , (4.16)
rN rN rN
i=1 i=1

where the sums indicate sumsets. The goal now is to compare the discrete Hausdorff contents
of each of these sets at scale r −N .
By the definition of the set Xi , it follows from Lemma 4.8 that
 ′ 
Ai ∩ [0, r N )
dH , Xi ≪ r −N , (4.17)
rN
which implies by Lemma 2.7 that for all γ ∈ [0, 1],
n
!
γ
X A′i ∩ [0, r N ) γ

H>r−N ≍ n H −N Y n . (4.18)
rN >r
i=1

It also follows from (4.17) that

 
A′i ∩ [0, r N −k ) Xi
dH , k ≪n r −N ,
rN r
which implies by Lemma 2.7 that
n
!  
γ
X A′i ∩ [0, r N −k ) γ Yn
H>r −N ≍n H>r −N . (4.19)
rN rk
i=1

Combining (4.16) with (4.18) and (4.19), we see that

    γ

γ Yn γ Bn ∩ [0, r N ) H>1 Bn ∩ [0, r N ) γ

H>r−N k
≪n H>r−N N
= N γ
≪n H>r −N Yn .
r r r
Letting N tend to ∞ and noting that n, and hence k, are fixed, these inequalities combine
with Lemma 2.4, Lemma 2.17 (IV), (4.14), and the fact that dimH (Yn /r k ) = dimH Yn to

41
prove the equality in (4.15).

4.5. Proofs of Theorems D and E

Proof of Theorem D assuming Theorem E. Suppose A, B ⊆ N0 are ×r- and ×s-invariant,
where r and s are multiplicatively independent, and let λ, η > 0 be fixed. Define γ =
min(dimH A + dimH B, 1), which according to Proposition 4.6 is the same as min(dimM A +
dimM B, 1). By Lemma 2.17, parts (II) and (V), we have the upper bound


dimM ⌊λA + ηB⌋ 6 γ.
On the other hand, from Theorem E we obtain for all γ < γ the lower bound


dimH ⌊λA + ηB⌋ > γ.
In view of Lemma 2.18, this proves that the discrete mass dimension and the discrete Haus-
dorff dimension of ⌊λA + ηB⌋ both exist and equal γ.

Proof that Theorem A implies Theorem E. Suppose A, B ⊆ N0 are ×r- and ×s-invariant,
where r and s are multiplicatively independent, and I ⊆ (0, ∞) is compact. Assuming
Theorem A, we want to show that
γ

H>1 ⌊λA + ηB⌋ ∩ [0, N )
lim inf inf > 0, (4.20)
N →∞ λ,η∈I Nγ
for all γ < γ := min(dimH A + dimH B, 1).
First, let us make the observation that
γ
1 γ

H>1 ⌊λA + ηB⌋ ∩ [0, N ) > H>1 (λA + ηB) ∩ [0, N ) .
2
Next, note that if M ∈ N is chosen sufficiently large depending on I, then for every λ, η ∈ I,
the set λ(A ∩ [0, N/M )) + η(B ∩ [0, N/M )) is a subset of (λA + ηB) ∩ [0, N ). This implies
γ
γ

H>1 (λA + ηB) ∩ [0, N ) H>1 λ(A ∩ [0, N/M )) + η(B ∩ [0, N/M ))
>
Nγ Nγ
γ
!
−γ H>1 λ(A ∩ [0, N/M )) + η(B ∩ [0, N/M ))
> M .
(N/M )γ

Therefore, (4.20) is implied by

γ

H>1 λ(A ∩ [0, N )) + η(B ∩ [0, N ))
lim inf inf > 0,
N →∞ λ,η∈I Nγ
which according to equation (2.7) is the same as
!
 A ∩ [0, N )   B ∩ [0, N ) 
γ
lim inf inf H>N −1 λ +η > 0. (4.21)
N →∞ λ,η∈I N N

Define for every k, ℓ ∈ N the sets Xk := (A ∩ [0, r k ))/r k and Yℓ := (B ∩ [0, sℓ ))/sℓ .
Define kN := ⌊log N/ log r⌋ and ℓN := ⌊log N/ log s⌋, and note that N = r kN r {log N/ log r} =

42
sℓN s{log N/ log s} . Since A ∩ [0, N ) ⊇ A ∩ [0, r kN ) and B ∩ [0, N ) ⊇ B ∩ [0, sℓN ), we have
!
 A ∩ [0, N )   B ∩ [0, N ) 
γ
H>N −1 λ +η
N N
!
 A ∩ [0, r kN )   B ∩ [0, sℓN ) 
γ
> H>N −1 λ k {log N/ log r} + η ℓ {log N/ log s}
r Nr s Ns
 
γ −{log N/ log r}
> H>N −1 λr XkN + ηs−{log N/ log s} YℓN .

Since I ⊆ (0, ∞) is compact, we can choose t > 0 such that I ⊆ [t−1 , t]. So if λ and η belong
to I then λr −{log N/ log r} and ηs−{log N/ log s} belong to the interval J := [min(r −1 , s−1 ) · t−1 , t].
We conclude that
!
 A ∩ [0, N )   B ∩ [0, N )   
γ γ
inf H>N −1 λ +η > inf H>N −1 λXkN + ηYℓN . (4.22)
λ,η∈I N N λ,η∈J

Next, let X = limk→∞ Xk and Y = limℓ→∞ Yℓ in the Hausdorff metric. The existence of
these limits is guaranteed by Proposition 4.6, which also gives that dimH X = dimH A and
dimH Y = dimH B. Since r kN 6 N < r kN +1 and sℓN 6 N < sℓN +1 , it follows from part (I) of
Lemma 4.8 that
   
X ⊆ XkN rN −1 and Y ⊆ YℓN sN −1 .
Therefore there exists a > 1 such that
h i
λX + ηY ⊆ λXkN + ηYℓN
aN −1

uniformly over all λ, η ∈ J. In view of Lemma 2.7 this implies

   
γ γ
inf H>N −1 λXkN + ηYℓN ≫a inf H>N −1 λX + ηY . (4.23)
λ,η∈J λ,η∈J

Since γ < γ, Theorem A gives

 
γ
lim inf H>N −1 λX + ηY > 0. (4.24)
N →∞ λ,η∈J

The claim in (4.21) now follows from (4.22), (4.23), and (4.24).

4.6. An example that shows R-invariance does not suffice

In this subsection, we exhibit sets A, B ⊆ N0 that demonstrate that Rr - and Rs -invariance
alone does not suffice to reach the conclusions in Theorem D. This is in contrast to Theorem B:
the conclusion holds under the weaker assumption that A is simultaneously Rr - and Rs -
invariant. We do not know whether L-invariance alone suffices in either Theorem B or
Theorem D, but invariance under multiplication by r and s alone does not suffice to reach
the conclusions in either theorem: the set of squares is invariant under multiplication by
both 4 and 9 simultaneously, but has dimension equal to 1/2, while the sets in the example
below demonstrate that Theorem D does not hold under the assumption of invariance under
multiplication.
Fix 2 6 r < s. We will construct two sets A, B ⊆ N0 which satisfy the following properties:

43
 (I) the mass dimensions of A and B exist and dimM A = dimM B = 1/2;
(II) rA ⊆ A and sB ⊆ B;
(III) Rr (A) = A and Rs (B) = B; and
(IV) dimM (A + B) 6 4/5.
This shows that neither R-invariance alone nor multiplication-invariance alone suffice to
obtain the result in Theorem D.
In what follows, the interval notation [a, b] is understood to mean [a, b]∩ N0 . For i, j ∈ N0 ,
let
Ii = [r i , r i + r (i+1)/2 ], Jj = [sj , sj + s(j+1)/2 ],
and then define
[ [
A = {0} ∪ r ℓ Ii , B = {0} ∪ sm Jj .
i,ℓ>0 j,m>0

First we will verify (I) by showing that the mass dimension of A exists and is equal to
1/2; the argument for B is the same. It is easy to see that for all N > 1,
[
IN −1 ⊆ A ∩ [1, r N ) ⊆ r ℓ Ii ,
i,ℓ>0
i+ℓ6N

from which it follows that

 
r N/2 6 A ∩ [0, r N ) 6 (N + 1)2 (r (N +1)/2 + 1).

This shows that dimM A = dimM A = dimM A = 1/2.

It is clear from the definition of the sets A and B that (II) holds.
Next we will verify (III) by showing that Rr (A) = A; the same argument works to show
that Rs (B) = B. Since rA ⊆ A, we have that
[
A = Rr (rA) ⊆ Rr (A) = {0} ∪ Rr (r ℓ Ii ).
i,ℓ>0

Since 0 ∈ A, we need only to verify that for all i, ℓ > 0, Rr (r ℓ Ii ) ⊆ A. If ℓ > 1, then
Rr (r ℓ Ii ) = r ℓ−1 Ii ⊆ A. If ℓ = 0 and i = 0, then we see Rr (I0 ) = {0} ⊆ A. If ℓ = 0 and i > 1,
then we see Rr (Ii ) = [r i−1 , r i−1 + r (i−1)/2 ] ⊆ Ii−1 ⊆ A. Thus, Rr (A) = A.
Finally we will verify (IV) by showing that for all N sufficiently large,
 
(A + B) ∩ [0, r N ) 6 4N 4 r 4N/5 . (4.25)
Let σ = log s/ log r. Because
[
B ∩ [1, r N ) ⊆ sm Jj ,
i,ℓ>0
σ(j+m)6N

we have that
  X
(A + B) ∩ [0, r N ) 6 1 + |r ℓ Ii + sm Jj |, (4.26)
i,j,ℓ,m

where the sum is over all i, j, ℓ, m > 0 for which i + ℓ 6 N and σ(j + m) 6 N . We will

44
estimate this sum from above by splitting the sum indices into two sets depending on the
“type” of the pair (i, j), which we now define.
A pair (i, j) is of Type I if
i+1 j+1 4N
+σ 6 . (4.27)
2 2 5
Using the trivial bound |C + D| 6 |C||D| for finite sets C, D ⊆ N0 , we see that if i, j, ℓ, and
m are such that (i, j) is of Type I, then
|r ℓ Ii + sm Jj | 6 |Ii ||Jj | = r (i+1)/2 s(j+1)/2 6 r 4N/5 . (4.28)
A pair (i, j) is of Type II if it is not of Type I, that is, if
i+1 j+1 4N
+σ > . (4.29)
2 2 5
Using the fact that σj 6 N and that N is sufficiently large, we see from (4.29) that (i−1)/2 >
N/4. It follows then from the fact that i + ℓ 6 N that
i+1 4N
ℓ+ < . (4.30)
2 5
Similarly, using that i 6 N and the fact that N is sufficiently large, we see from (4.29) that
σ(j − 1)/2 > N/4. It follows from the fact that σ(j + m) 6 N that
 
j+1 4N
σ m+ < . (4.31)
2 5
Now we are in a position to use the following fact: if C, D ⊆ N0 are contained in intervals of
length L, M , respectively, then C + D is contained in an interval of length L + M and hence
|C + D| 6 L + M + 1. If i, j, ℓ, and m are such that (i, j) is of Type II, then
|r ℓ Ii + sm Jj | 6 r ℓ+(i+1)/2 + sm+(j+1)/2 + 1.
Using (4.30) and (4.31), we have that
|r ℓ Ii + sm Jj | 6 3r 4N/5 . (4.32)
Finally, by splitting up the sum in (4.26) into tuples for which the pairs (i, j) are of Type
I or Type II, we see by combining (4.28) and (4.32) that the desired inequality in (4.25) holds.

5. Open directions
We collect in this section a number of interesting and potentially fruitful open questions con-
cerning multiplicatively invariant subsets of the non-negative integers. Though these ques-
tions and conjectures are stated for arbitrary ×r-invariant subsets of N0 , many are already
open and interesting for the special case of base-r restricted digit Cantor sets.

5.1. Positive density for sumsets of full dimension

In [Hoc, Problem 4.10], Hochman asks whether the sumset X + Y of a ×r-invariant set
X ⊆ [0, 1) and a ×s-invariant set Y ⊆ [0, 1) satisfying dimH X + dimH Y > 1 has positive
Lebesgue measure. We remark that a projection theorem of Marstrand [Mar, Theorem I]

45
implies that λX +ηY has positive Lebesgue measure for a.e. (λ, η) ∈ R2 , suggesting a possible
affirmative answer. In [Gla, Theorem 1.4], a version of Marstrand’s projection theorem for
subsets of the integers was obtained, with Lebesgue measure replaced by the notion of upper
natural density.4 It therefore makes sense to consider the following integer analogue of the
the question.

Question 5.1. Let r, s ∈ N be multiplicatively independent, and let A, B ⊆ N0 be ×r- and

×s-invariant, respectively. If dimM A+dimM B > 1, then does the sumset A+B have positive
upper natural density?

5.2. Difference sets

For closed subsets X, Y ⊆ [0, 1], working with the difference set X − Y is no harder than
working with the sumset X + Y . In particular, proving that



dimM X − Y = min dimM X + dimM Y, 1
in Theorem 1.3 requires no additional work. The story changes in the setting of the non-
negative integers, where difference sets are much more cumbersome to handle, ultimately
because the fibers of the map (a, b) 7→ a − b are not compact. This is why our main results
in the integer setting only deal with sumsets λA + ηB with λ and η both positive. This
naturally leads us to the following question.

Question 5.2. Let r and s be multiplicatively independent positive integers, and let A, B ⊆
N0 be ×r- and ×s-invariant, respectively. Is it true that


dimM (A − B) = min dimM A + dimM B, 1 ?

The methods used in Section 4 allow us to establish the lower bound dimM (A − B) >
min(dimM A + dimM B, 1). However, the upper bound dimM (A − B) 6 min(dimM A +
dimM B, 1), which is straightforward for sums, remains open for differences.

5.3. Analogous results for the counting dimension

The upper Banach dimension (or upper counting dimension, cf. [LM2] and [Gla]) of a set
A ⊆ N0 is
 
∗ log A ∩ [M, N ]
dim A := lim sup .
N −M →∞ log(N − M )

In general, we only have the inequality dim∗ A > dimM A, but if A ⊆ N0 is ×r-invariant, then
it can be shown that dimM A = dimH A = dim∗ A.

Question 5.3. Let r and s be multiplicatively independent positive integers, and let A, B ⊆
N0 be ×r- and ×s-invariant, respectively. Is it true that


dim∗ (A + B) = min dim∗ A + dim∗ B, 1 ?


Note that the lower bound dim∗ (A + B) > min dim∗ A + dim∗ B, 1 follows from The-
orem D using the fact that dim∗ > dimM .
4 ¯
Given a set E ⊆ Z, its upper natural density is defined by d(E) := lim supN→∞ |E ∩{−N, . . . , N }|/(2N +1).

46
 5.4. Polynomial functions of multiplicatively invariant sets
Our main results for integer sets, Theorems D and E, concern the dimension of the sumset
A + B. It is natural to ask about different functions of A and B. The following conjecture is
a (special case of a) natural polynomial extension of Theorem D.

Conjecture 5.4. Fix n, m ∈ N and let ∆(A, B) := {an + bm | a ∈ A, b ∈ B} for A, B ⊆ N0 .

Let r, s ∈ N be multiplicatively independent, and let A, B ⊆ N0 be a ×r- and a ×s-invariant
subsets, respectively. Then
 
1 1
dimM ∆(A, B) = min dimM A + dimM B , 1 . (5.1)
n m
It is easy to see that for any A ⊆ N0 , the set An := {an | a ∈ A} has dimension
dimM An = n1 dimM A, however it is not true in general that An is ×r-invariant when A is.
Setting ∆(A, B) = ∆λ,η (A, B) := {λan + ηbm | a ∈ A, b ∈ B}, it follows from (1.10) that
the equality in (5.1) holds for Lebesgue almost every (λ, η) ∈ R2 when A and B satisfy a
natural dimension condition (see footnote 2).

5.5. Geometric transversality in the integers

In [Fur2], Furstenberg conjectured that given ×r- and a ×s-invariant subsets X and Y of
[0, 1], where r, s ∈ N are multiplicatively invariant, the intersection satisfies



dimH X ∩ Y 6 max dimH X + dimH Y − 1, 0 .
This conjecture was recently proved, independently, by Shmerkin [Shm] and Wu [Wu]. It is
natural to formulate an analogous conjecture for subsets of N0 .

Conjecture 5.5. Let r, s ∈ N be multiplicatively independent, and let A, B ⊆ N0 be ×r-

and ×s-invariant, sets respectively. Then



dimM A ∩ B 6 max dimM A + dimM B − 1, 0 .

This conjecture and some related problems will be addressed in a forthcoming paper.

5.6. Multiplicatively invariant sets in relation to other arithmetic sets in

the integers
In this paper, we are concerned with transversality between ×r- and ×s-invariant sets when-
ever r and s are multiplicatively independent. In principle, it makes sense to inquire about
transversality (or independence) between any two sets which are structured in different ways.
To keep the discussion short, we restrict to infinite arithmetic progressions (or congruence
classes), the set of perfect squares, and the set of primes.

Question 5.6. Let A ⊆ N0 be a ×r-invariant set, and let P be an infinite arithmetic

progression. Is is true that dimM (A ∩ P ) is either 0 or dimM (A)?

The answer is yes for restricted digit Cantor sets. In fact, it is proved in [EMS] that such
sets satisfy “good equidistribution properties” in residue classes.
More generally, one could ask about the sum or the intersection of a ×r-invariant set and

47
the image of an arbitrary polynomial with integer coefficients, for instance the set of perfect
squares, S = {n2 | n ∈ N0 }. Note that dimM S = 1/2.

Question 5.7. Let A ⊆ N0 be a ×r-invariant set. Is is true that

dimM A + S = min dimM A + 1/2, 1
and/or



dimM A ∩ S 6 max dimM A − 1/2, 0 ?

In a similar vein, one can ask about intersections with the set of prime numbers, P. Note
that dimM P = 1.

Question 5.8. Let A ⊆ N0 be a ×r-invariant set. Is is true that dimM (A ∩ P) is either 0 or

dimM (A)?

Maynard showed in [May] that the answer to Question 5.8 is positive when A is a restricted
digit Cantor set where the number of restricted digits is small enough with respect to the
base. In fact, he obtains a Prime Number Theorem in such sets, which is stronger than
simply dimM (A ∩ P) = dimM A. Question 5.8 is open for general restricted digit Cantor sets,
and may be very difficult in general. The methods in this paper do not appear to shed new
light on this line of inquiry.

5.7. Transversality of multiplicatively invariant sets in the rs-adics

The rs-adics is a non-Archimedean regime in which it is easy to ask questions analogous to
those asked in this work. Following Furstenberg [Fur2], note that the maps Rr and Rs , with
domains extended to Z, are uniformly continuous with respect to the rs-adic metric on Z,
and therefore extend to continuous transformations of the set of rs-adic integers, Zrs . As a
compact metric space, there is a natural Hausdorff dimension to measure the size of subsets
of Zrs . Let us call a set X ⊆ Zrs ×r-invariant if it is closed and Rr X ⊆ X.

Question 5.9. Let r and s be multiplicatively independent positive integers,

and let X, Y ⊆
Zrs be ×r- and ×s-invariant

sets, respectively. Is it true that dim H X +Y = min dimH X +
dimH Y, dimH Zrs ?

Conjecture 5.10 ([Fur2, Conjecture 3]). Let r and s be multiplicatively independent posi-
tive integers, and let X, Y ⊆ Zrs be ×r- and ×s-invariant sets, respectively. One has



dimH X ∩ Y 6 max dimH X + dimH Y − dimH Zrs , 0 .

Furstenberg [Fur2, Theorem 3] proved an analogue of Theorem 1.1 in the rs-adics; positive
answers to the previous question and conjecture would combine with that result to bring
transversality results in the rs-adics in line with those in the real and integer settings.

References
[BCS] William D. Banks, Alessandro Conflitti, and Igor E. Shparlinski. Character sums
over integers with restricted g-ary digits. Illinois J. Math., 46(3):819–836, 2002.

48
[BP] Christopher J. Bishop and Yuval Peres. Fractals in probability and analysis, volume
162 of Cambridge Studies in Advanced Mathematics. Cambridge University Press,
Cambridge, 2017.

[BS] William D. Banks and Igor E. Shparlinski. Arithmetic properties of numbers with
restricted digits. Acta Arith., 112(4):313–332, 2004.

[BT1] Martin T. Barlow and Samuel J. Taylor. Fractional dimension of sets in discrete
spaces. J. Phys. A, 22(13):2621–2628, 1989. With a reply by Jan Naudts.

[BT2] Martin T. Barlow and Samuel J. Taylor. Defining fractal subsets of Zd . Proc. London
Math. Soc. (3), 64(1):125–152, 1992.

[BY] Stuart A. Burrell and Han Yu. Digit expansions of numbers in different bases. arXiv
e-prints, page arXiv:1905.00832, May 2019.

[Cas] John William Scott Cassels. On a problem of Steinhaus about normal numbers.
Colloq. Math., 7:95–101, 1959.

[DW] Taylor Dupuy and David E. Weirich. Bits of 3n in binary, Wieferich primes and a
conjecture of Erdős. J. Number Theory, 158:268–280, 2016.

[EMS] Paul Erdős, Christian Mauduit, and András Sárközy. On arithmetic properties of
integers with missing digits. I. Distribution in residue classes. J. Number Theory,
70(2):99–120, 1998.

[Erd] Paul Erdős. Some unconventional problems in number theory. Math. Mag., 52(2):67–
70, 1979.

[FFJ] Kenneth Falconer, Jonathan Fraser, and Xiong Jin. Sixty years of fractal projections.
In Fractal geometry and stochastics V, volume 70 of Progr. Probab., pages 3–25.
Birkhäuser/Springer, Cham, 2015.

[Fur1] Harry Furstenberg. Disjointness in ergodic theory, minimal sets, and a problem in
Diophantine approximation. Math. Systems Theory, 1:1–49, 1967.

[Fur2] Harry Furstenberg. Intersections of Cantor sets and transversality of semigroups.

pages 41–59, 1970.

[Gla] Daniel Glasscock. Marstrand-type theorems for the counting and mass dimensions
in Zd . Combin. Probab. Comput., 25(5):700–743, 2016.

[Hoc] Michael Hochman. Dimension theory of self-similar sets and measures. In Proceedings
of the International Congress of Mathematicians—Rio de Janeiro 2018. Vol. III.
Invited lectures, pages 1949–1972. World Sci. Publ., Hackensack, NJ, 2018.

[HS] Michael Hochman and Pablo Shmerkin. Local entropy averages and projections of
fractal measures. Ann. of Math. (2), 175(3):1001–1059, 2012.

[Inc] OEIS Foundation Inc. The On-Line Encyclopedia of Integer Sequences.

http://oeis.org/A146025, 2020.

49
[IRUT] Alexander Iosevich, Michael Rudnev, and Ignacio Uriarte-Tuero. Theory of dimen-
sion for large discrete sets and applications. Math. Model. Nat. Phenom., 9(5):148–
169, 2014.

[KMS] Sergei Konyagin, Christian Mauduit, and András Sárközy. On the number of prime
factors of integers characterized by digit properties. Period. Math. Hungar., 40(1):37–
52, 2000.

[KN] Lauwerens Kuipers and Harald Niederreiter. Uniform distribution of sequences.

Wiley-Interscience [John Wiley & Sons], New York-London-Sydney, 1974. Pure and
Applied Mathematics.

[Kon] Sergei Konyagin. Arithmetic properties of integers with missing digits: distribution
in residue classes. Period. Math. Hungar., 42(1-2):145–162, 2001.

[KT] Nets Hawk Katz and Terence Tao. Some connections between Falconer’s distance
set conjecture and sets of Furstenburg type. New York J. Math., 7:149–187, 2001.

[Lag] Jeffrey C. Lagarias. Ternary expansions of powers of 2. J. Lond. Math. Soc. (2),
79(3):562–588, 2009.

[LM1] Yuri Lima and Carlos Gustavo Moreira. A combinatorial proof of Marstrand’s the-
orem for products of regular Cantor sets. Expo. Math., 29(2):231–239, 2011.

[LM2] Yuri Lima and Carlos Gustavo Moreira. A Marstrand theorem for subsets of integers.
Combinatorics, Probability and Computing, 23:116–134, 1 2014.

[LM3] Douglas Lind and Brian Marcus. An introduction to symbolic dynamics and coding.
Cambridge University Press, Cambridge, 1995.

[LMP] Elon Lindenstrauss, David Meiri, and Yuval Peres. Entropy of convolutions on the
circle. Ann. of Math. (2), 149(3):871–904, 1999.

[Mar] John M. Marstrand. Some fundamental geometrical properties of plane sets of frac-
tional dimensions. Proc. London Math. Soc. (3), 4:257–302, 1954.

[Mat] Pertti Mattila. Geometry of sets and measures in Euclidean spaces, volume 44 of
Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cam-
bridge, 1995. Fractals and rectifiability.

[May] James Maynard. Primes with restricted digits. Invent. Math., 217(1):127–218, 2019.

[MF] Michel Mendès France. Sur les décimales des nombres algébriques réels. In Seminar
on Number Theory, 1979–1980 (French), pages Exp. No. 28, 7. Univ. Bordeaux I,
Talence, 1980.

[Mor] Carlos Gustavo T. Moreira. Sums of regular Cantor sets, dynamics and applica-
tions to number theory. In International Conference on Dimension and Dynamics
(Miskolc, 1998), volume 37, pages 55–63. 1998.

[Nau] Jan Naudts. Dimension of discrete fractal spaces. J. Phys. A, 21(2):447–452, 1988.

50
[PS] Yuval Peres and Pablo Shmerkin. Resonance between Cantor sets. Ergodic Theory
Dynam. Systems, 29(1):201–221, 2009.

[Shm] Pablo Shmerkin. On Furstenberg’s intersection conjecture, self-similar measures, and

the Lq norms of convolutions. Ann. of Math. (2), 189(2):319–391, 2019.

[Wal] Peter Walters. An introduction to ergodic theory, volume 79 of Graduate Texts in

Mathematics. Springer-Verlag, New York-Berlin, 1982.

[Wu] Meng Wu. A proof of Furstenberg’s conjecture on the intersections of ×p- and ×q-
invariant sets. Ann. of Math. (2), 189(3):707–751, 2019.

[Yu] Han Yu. A higher dimensional version of Furstenberg’s intersection problem

and additive properties of numbers with restricted digits. arXiv e-prints, page
arXiv:2004.05926, April 2020.

Daniel Glasscock
University of Massachusetts Lowell
daniel [email protected]

Joel Moreira
University of Warwick
[email protected]

Florian K. Richter
Northwestern University
[email protected]

51