Annales Fennici Mathematici

Volumen 47, 2022, 645–694

Fast approximation of the affinity dimension for

dominated affine iterated function systems
Ian D. Morris

Abstract. In 1988 Falconer introduced a formula which predicts the value of the Hausdorff
dimension of the attractor of an affine iterated function system. The value given by this formula—
sometimes referred to as the affinity dimension—is known to agree with the Hausdorff dimension
both generically and in an increasing range of explicit cases. It is however a nontrivial problem
to estimate the numerical value of the affinity dimension for specific iterated function systems. In
this article we substantially extend an earlier result of Pollicott and Vytnova on the computation of
the affinity dimension. Pollicott and Vytnova’s work applies to planar invertible affine contractions
with positive linear parts under several additional conditions which among other things constrain
the affinity dimension to be between 0 and 1. We extend this result by passing from planar self-affine
sets to self-affine sets in arbitrary dimensions, relaxing the positivity hypothesis to a domination
condition, and removing all other constraints including that on the range of values of the affinity
dimension. We provide explicit examples of two- and three-dimensional affine iterated function
systems for which the affinity dimension can be calculated to more than 30 decimal places.

Dominoitujen affiinien iteroitujen funktiosysteemien

affiinisuusulottuvuuden nopea arvioiminen

Tiivistelmä. Vuonna 1988 Falconer esitteli kaavan, joka ennustaa affiinin iteroidun funktio-
systeemin kiintojoukon Hausdorffin ulottuvuuden. Tämän kaavan antaman arvo – jota toisinaan
kutsutaan affiinisuusulottuvuudeksi – tiedetään yhtä suureksi Hausdorffin ulottuvuuden kanssa sekä
geneerisesti että kasvavassa joukossa suoria esimerkkejä. On kuitenkin epätriviaali ongelma arvioida
affiinin ulottuvuuden numeerista arvoa määrätyille iteroiduille funktiosysteemeille. Tässä työssä
laajennamme oleellisesti Pollicottin ja Vytnovan aiempaa affiinisuusulottuvuuden laskentaa koske-
vaa tulosta. Pollicottin ja Vytnovan työ soveltuu tason kääntyviin affiineihin kutistuksiin, joiden
lineaariset osat ovat positiivisia, kun lisäksi oletetaan useita muita ehtoja, jotka mm. rajoittavat
affiinin ulottuvuuden nollan ja yhden välille. Laajennamme tätä tulosta siirtymällä tasosta yleisen
ulottuvuuden itseaffiineihin joukkoihin, lieventämällä positiivisuusoletusta dominointiehdoksi sekä
luopumalla kaikista muista rajoitteista – erityisesti affiinin ulottuvuuden arvojoukkoa rajoittavasta
ehdosta. Esitämme suoria esimerkkejä kaksi- ja kolmiulotteisista affiineista iteroiduista funktiosys-
teemeistä, joiden affiini ulottuvuus voidaan laskea yli 30 desimaalin tarkkuudella.

1. Introduction
1.1. Background and context. If T1 , . . . , TN : Rd → Rd are contractions it
is well-known
SN that there exists a unique nonempty compact set X ⊂ Rd such that
X = i=1 Ti X. In this case (T1 , . . . , TN ) is called an iterated function system and the
set X its attractor. When each transformation Ti is a similitude with contraction ratio
ri ∈ (0, 1) and the distinct images Ti X ∩Tj X do not overlap too strongly it is classical

https://doi.org/10.54330/afm.116153
2020 Mathematics Subject Classification: Primary 28A80; Secondary 37D30, 37D35.
Key words: Iterated function system, self-affine set, affinity dimension.
c 2022 The Finnish Mathematical Society
646 Ian D. Morris

that the box dimension and HausdorffP dimension of the attractor are both equal to the
unique real number s > 0 such that N s
i=1 ri = 1 (see for example [19, Theorem 9.3]
or the original article [31]). In the case where each Ti is instead an affine map Ti x =
Ai x+vi the Hausdorff dimension and box dimension of the attractor X—which in this
context we call a self-affine set—are more challenging to calculate. The problem of
determining the Hausdorff dimension of such sets, even implicitly, has been an active
topic of research since the 1980s and has received particularly intense research interest
within the last decade (see for example the classic articles [10, 16, 17, 18, 30, 45] and
more recent contributions such as [4, 5, 13, 14, 21, 23, 24, 39, 50]). In the landmark
article [17] Falconer defined an implicit formula which is known to give the correct
value for the Hausdorff dimension of a wide variety of self-affine sets. The subject of
this article is the numerical estimation of the value predicted by Falconer’s formula.
In order to define Falconer’s formula we require a few preliminary definitions.
Let Md (R) denote the set of all real d × d matrices. If A ∈ Md (R) we recall that
the singular values of A are defined to be the square roots of the eigenvalues of the
positive semidefinite matrix A⊤ A. We denote the singular values of A ∈ Md (R) by
σ1 (A), . . . , σd (A) in decreasing order of absolute value. For each A ∈ Md (R) and
s ≥ 0 let us define
(
σ1 (A) · · · σ⌊s⌋ (A)σ⌈s⌉ (A)s−⌊s⌋ if 0 ≤ s ≤ d,
ϕs (A) := s
| det A| d if s ≥ d.
It was shown in [17] that for each s ≥ 0 we have ϕs (AB) ≤ ϕs (A)ϕs (B) for all
A, B ∈ Md (R). The affinity dimension of the iterated function system Ti x := Ai x =
vi , where 1 ≤ i ≤ N, is then defined to be the quantity
( ∞ N
)
X X
dimaff (T1 , . . . , TN ) := inf s > 0 : ϕs (Ai1 · · · Ain ) < ∞ .
n=1 i1 ,...,in =1

Since dimaff (T1 , . . . , TN ) depends only on A1 , . . . , AN and not on the additive part of
the transformations Ti we will also denote it by dimaff (A1 , . . . , AN ). If the matrices
A1 , . . . , AN are assumed to be invertible and contracting with respect to some norm
on Rd then the affinity dimension is the unique s > 0 such that the quantity
N
1 X
P (A1 , . . . , AN ; s) := lim log ϕs (Ai1 · · · Ain )
n→∞ n
i ,...,i =1
1 n

is equal to zero.
Let k · k denote the Euclidean norm on Rd . It was shown in [17] that when
max1≤i≤N kAi k < 1 the affinity dimension dimaff (A1 , . . . , AN ) is well-defined and
is an upper bound for the box dimension of the attractor. (This argument may
easily be adapted to the case where max1≤i≤N |||Ai ||| < 1 in the operator norm in-
duced by some norm |||·||| on Rd .) It was additionally shown that when matrices
A1 , . . . , AN satisfying max1≤i≤N kAi k < 31 are fixed, then for Lebesgue-a.e. choice
of (v1 , . . . , vN ) ∈ (Rd )N the attractor of the affine transformations T1 , . . . , TN given
by Ti x := Ai x + vi has Hausdorff dimension equal to min{d, dimaff (A1 , . . . , AN )}.
Subsequent research focused on providing explicit examples for which the Hausdorff
dimension of the attractor equals the affinity dimension of the defining iterated func-
tion system, with explicit special cases being given in articles such as [21, 24, 30, 50].
Recently, Bárány, Hochman and Rapaport have shown that the Hausdorff dimension
of a planar self-affine set is always equal to the affinity dimension of the defining
 Fast approximation of the affinity dimension for dominated affine iterated function systems 647

iterated function system as long as the matrices Ai are invertible, the affine transfor-
mations satisfy the strong open set condition, and the matrices | det Ai |−1/2 Ai neither
belong to a compact subgroup of GL2 (R) nor preserve a finite subset of RP1 . At the
present time, however, results on higher-dimensional self-affine sets additional to that
of Falconer are essentially unavailable.
Despite its prominent rôle in the dimension theory of self-affine sets, the prop-
erties of the affinity dimension itself have been investigated only very recently. In
the 2014 article [23] Feng and Shmerkin showed for the first time that the affinity
dimension dimaff (A1 , . . . , AN ) depends continuously on the entries of the matrices
A1 , . . . , AN , and in [46] it was shown that the affinity dimension is computable in
principle in the sense that for any given ε > 0 we may algorithmically compute an
explicit approximation to dimaff (A1 , . . . , AN ) which is guaranteed to be accurate to
within the prescribed error ε and which requires only finitely many arithmetical oper-
ations to calculate. However, the method of [46] does not result in an algorithm which
is fast enough to be useful in practical computations. Further general properties of
the affinity dimension were investigated in [13, 38].
At the present time there are very few practical techniques available for the
computation of the affinity dimension. In the article [48] the author gave a simple
closed-form expression for the affinity dimension in the very special case where the
matrices Ai are generalised permutation matrices, that is, matrices having exactly one
nonzero entry in every row and column. Closed-form expressions are also available in
the case of diagonal and upper-triangular matrices [22, 38]. To the best of the author’s
knowledge there so far exists only one result in the literature which is powerful enough
to be able to estimate the affinity dimension for a nonempty open set of examples in a
practicable time frame. The following result was proved by Pollicott and Vytnova in
[56]. Here and throughout this article ρ(A) denotes the spectral radius of the matrix
or linear operator A.
Theorem 1. Let A1 , . . . , AN be 2 × 2 matrices which satisfy the following con-
ditions:
(i) We have σ1 (Ai )2 < σ2 (Ai ) < 1 for all i = 1, . . . , N.
(ii) If Q2 is defined to be the open second quadrant {(x, y) ∈ R2 : x < 0 < y},
then the sets A−1 −1
1 Q2 , . . . , AN Q2 are subsets of Q2 and have pairwise disjoint
closures in Q2 .
(iii) All entries of the matrices Ai are strictly positive1.
For each n ≥ 1 and s ∈ C define
N
X ρ(Ai1 · · · Ain )2+s
tn (s) = ,
i1 ,...,in =1
ρ(Ai1 · · · Ain )2 − det Ai1 · · · Ain
n k
X (−1)k X Y tni (s)
an (s) :=
k=1
k! k i=1
ni
(n1 ,...,nk )∈N
Pk
i=1 ni =n

1, and for each n ≥ 1 let sn ∈ R denote the smallest positive real number
and a0 (s) :=P
s such that ni=0 ai (s) = 0. Then dimaff (A1 , . . . , AN ) ∈ (0, 1), sn is well-defined for

1This hypothesis is invoked in Pollicott and Vytnova’s section 3 but is not explicitly stated in their
introduction. It does not follow automatically from the other hypotheses unless the determinants
are assumed positive.
648 Ian D. Morris

all sufficiently large n, and there exists γ > 0 such that

|dimaff (A1 , . . . , AN ) − sn | = O exp(−γn2 ) .

Remark. The quantity an (s) may be alternatively characterised as

 
t1 (s) n−1 0 ··· 0 0
 t2 (s) t1 (s) n−2 ··· 0 0 
..
 
(−1)n  t (s)

t2 (s) t1 (s) . 0 0 

det  3 . .. .. .. .. ..  ,
n!  .. . . . . . 
 
tn−1 (s) tn−2 (s) tn−3 (s) · · · t1 (s) 1 
tn (s) tn−1 (s) tn−2 (s) · · · t2 (s) t1 (s)
and we will prefer this format in our exposition.
The methods underlying the proof of Theorem 1 will be described in more
detail in the following section. We remark that condition (i) above implies that
the matrices are invertible, and the combination of the three conditions implies
0 < dimaff (A1 , . . . , AN ) < 1 (see [30] for details).
In fact the only condition which is really essential to Pollicott and Vytnova’s
argument is that the matrix entries are positive, although in cases where we have
dimaff (A1 , . . . , AN ) ∈ (1, 2) the formula for tn (s) must be replaced with
N
X ρ(Ai1 · · · Ain )4−s | det Ai1 · · · Ain |s−1
tn (s) := 2 − det A · · · A
.
i1 ,...,in =1
ρ(A i 1 · · · Ai n ) i 1 i n

In this article we aim to prove as comprehensive as possible an extension of The-

orem 1. In particular, as well as removing hypotheses (i)–(ii) from Theorem 1 we
will establish a version of that theorem which is valid for affine iterated function
systems in dimensions higher than two, in which dimaff (A1 , . . . , AN ) may take any
value in the range (0, d), and in which the hypothesis of positivity is weakened to
one of domination. In order to state our results in full we will require a number of
definitions, which relate to multilinear algebra, to positivity and to domination.
1.2. Multilinear algebra. In extending Theorem 1 one of our concerns will be
to allow matrices of arbitrary dimension. Whereas in two dimensions the function
ϕs (A) admits the simple characterisation
(
s kAks if 0 ≤ s ≤ 1,
ϕ (A) = s−1 2−s
| det A| kAk if 1 ≤ s ≤ 2,
when s > 1 and d > 2 the analogous formula involves exterior powers of the matrix
A. In order to study the singular value function ϕs in dimensions higher than two
we therefore need to recall some concepts and notation from multilinear algebra.
Recall that when 1 ≤ k ≤ d the real vector space ∧k Rd is the vector space
spanned by the formal expressions {v1 ∧ v2 ∧ · · · ∧ vk : v1 , . . . vk ∈ Rd } subject to the
identifications
λ(v1 ∧ v2 ∧ · · · ∧ vk ) = (λv1 ) ∧ v2 ∧ · · · ∧ vk ,
(u1 ∧ v2 ∧ · · · ∧ vk ) + (v1 ∧ v2 ∧ · · · ∧ vk ) = (u1 + v1 ) ∧ v2 ∧ · · · ∧ vk ,
v1 ∧ v2 ∧ · · · ∧ vk = (−1)sign(π) vπ(1) ∧ vπ(2) ∧ · · · vπ(k)
for all v1 , . . . , vk , u1 ∈ Rd ,
λ ∈ R and permutations π : {1, . . . , k} → {1, . . . , k}.
The vector space ∧k Rd is kd -dimensional and if v1 , . . . , vd is any basis for Rd , then
 Fast approximation of the affinity dimension for dominated affine iterated function systems 649

{vi1 ∧ · · · ∧ vik : 1 ≤ i1 < i2 < · · · < ik ≤ d} is a basis for ∧k Rd . The kd -dimensional

vector space ∧k Cd may be constructed analogously.

The space ∧k Rd inherits an inner product h·, ·i∧k Rd from the standard inner prod-
uct h·, ·i on Rd which satisfies
 
hu1 ∧ · · · ∧ uk , v1 ∧ · · · ∧ vk i∧k Rd = det [hui , vj i]ki,j=1 .

If A ∈ Md (R) then we may define a linear map A∧k : ∧k Rd → ∧k Rd by A∧k (v1 ∧ · · · ∧

vk ) = Av1 ∧ · · · ∧ Avk . If v1 , . . . , vd is a basis for Cd consisting of eigenvectors and
generalised eigenvectors for A then the vectors vi1 ∧· · ·∧vik form a basis for ∧k Cd and
it is not hard to see that if
λ1 , . . . , λd are the eigenvalues of A then the eigenvalues
of A∧k are precisely the kd different products λi1 · · · λik with 1 ≤ i1 < · · · < ik ≤ d.
It is clear from the definition of the inner product on ∧k Rd that (A∧k )⊤ = (A⊤ )∧k .
Combining these observations we may easily see that
∧k

⊤

∧k  21
A ∧k Rd = ρ A A = σ1 (A) · · · σk (A)

for all A ∈ Md (R). By convention we also define ∧0 Rd = R and A∧0 = 1. It follows

easily that we may write
1+s−⌊s⌋ ⌈s⌉−s
ϕs (A) = A∧⌊s⌋ A∧⌈s⌉
for all A ∈ Md (R) and s ∈ [0, d].
1.3. Positivity and domination. As well as increasing the dimension of the
matrices to be considered in our extension of Theorem 1 we would like to weaken as
much as possible the hypothesis that the matrices have positive entries. To this end
we introduce the following definition:
Definition 1.1. Let A ⊂ Md (R) be nonempty. We say that (K1 , . . . , Km ) is a
multicone for A if the following properties hold:
(i) Each Kj is a closed, convex subset of Rd with nonempty interior such that
λKj ⊆ Kj for every non-negative real number λ.
d
Smexists a unit vector w ∈ R such that hu, wi > 0 for all nonzero vectors
(ii) There
u ∈ j=1 Kj . In particular Kj ∩ −Kj = {0} for all j = 1, . . . , m.
(iii) For every A ∈ A and j ∈ {1, . . . , m} there exists ℓ = ℓ(j, A) ∈ {1, . . . , m}
such that A(Kj \ {0}) ⊂ (Int Kℓ ) ∪ (− Int Kℓ ).
(iv) For all distinct j1 , j2 ∈ {1, . . . , m} we have Kj1 ∩ Kj2 = {0}.
When (ii) holds we say that w is a transverse-defining
Sm vector for (K1 , . . . , Km ) since
the hyperplane normal to w is transverse to j=1 Kj . If a multicone for A exists then
we say that A is multipositive.
We shall say that a set A ⊂ Md (R) is k-multipositive if the set {A∧k : A ∈ A}
is multipositive. By abuse of notation we shall say that a tuple of matrices is k-
multipositive if and only if the corresponding set is. We observe that a tuple of d × d
matrices with all entries positive is multipositive since we may take m = 1 and K1
to be the closed positive orthant in Rd . It follows that every tuple of d × d matrices
is 0-multipositive. We also observe that every tuple of d × d invertible matrices is
d-multipositive.
In generalising Theorem 1 we will adopt the hypothesis that (A1 , . . . , AN ) is k-
multipositive for certain integers k depending on dimaff (A1 , . . . , AN ). In the invertible
case this hypothesis may be related to the concept of domination as follows. If
650 Ian D. Morris

1 ≤ k < d then a tuple of invertible matrices (A1 , . . . , AN ) ∈ GLd (R)N is called

k-dominated if there exist C, γ > 0 such that
σk+1 (Ai1 · · · Ain ) ≤ Ce−γn σk (Ain · · · Ai1 )
for all i1 , . . . , in ∈ {1, . . . , n} and n ≥ 1. By convention we will say that every
(A1 , . . . , AN ) ∈ GLd (R)N is both 0- and d-dominated. It is not difficult to show using
the observations made in the previous subsection that (A1 , . . . , AN ) is k-dominated if
and only if (A∧k ∧k
1 , . . . , AN ) is 1-dominated. Various characterisations of domination—
in terms of invariant splittings, singular values, contraction on projective spaces and
contraction on Grassmannians—were explored by Bochi, Gourmelon, Barnsley and
Vince in [12, 9]. In particular it was shown in [12] that a compact set of invertible
matrices is 1-dominated if and only if it satisfies a slightly weakened form of multi-
positivity in which the criteria of Definition 1.1 all hold except that the sets Kj are
not assumed to be convex. By repeating iteratively the operations of replacing each
set Kj with its convex hull (which may introduce overlaps) and uniting overlapping
sets Kj (which may introduce non-convexity but reduces the number of sets Kj to be
considered) one may prove the following result by inductive descent on the number
of sets Kj :
Proposition 1.2. [9] Let A ⊂ Md (R) be compact, and suppose that every A ∈ A
is invertible. Then A is 1-dominated if and only if it is multipositive.
An obvious consequence of this proposition is that for every k = 0, . . . , d every
compact set of invertible matrices A ⊂ Md (R) is k-dominated if and only if it is
k-multipositive.
1.4. The main theorem. In order to state our main theorem we require just
a few more items of notation. For each N ≥ 1 let us define
∞
[
Σ∗N := {1, . . . , N}n .
n=1

If i = (ik )nk=1 ∈ Σ∗N we write |i| = n and refer to |i| as the length of i. If i, j ∈ Σ∗N
we let ij ∈ Σ∗N denote the sequence of length |i| + |j| obtained by running first
through the symbols of i and then through those of j in the obvious fashion. Clearly
Σ∗N is a semigroup with respect to the operation (i, j) 7→ ij. If A1 , . . . , AN ∈ Md (R)
and i = (ik )nk=1 ∈ Σ∗N then we write Ai := Ain · · · Ai1 . We observe that Ai Aj = Aji
for all i, j ∈ Σ∗N .
If B is a linear transformation of a finite-dimensional real vector space we let
λ1 (B), . . . , λd (B) denote the eigenvalues of B listed with repetition according to
multiplicity and listed in decreasing order of absolute value. While this notation a
priori introduces ambiguities when distinct eigenvalues of the same modulus exist,
we will see that this consideration does not affect the statements of our results.
We may now present the following generalisation of Pollicott and Vytnova’s result:
Theorem 2. Let d, N ≥ 2, let (A1 , . . . , AN ) ∈ Md (R)N and let 0 ≤ k < d.
Suppose that (A1 , . . . , AN ) is both k-multipositive and (k + 1)-multipositive. For
each integer n ≥ 1 and s ∈ R define
 d −1

d
∧k (k)−1 ∧(k+1) (k+1)
k+1−s  ∧(k+1) s−k

X λ1 Ai λ1 Ai ρ A∧k i ρ Ai
tn (s) :=   
′ ∧k


′ ∧(k+1)
|i|=n p A∧k
λ 1 A i p ∧(k+1) λ 1 Ai
i Ai
 Fast approximation of the affinity dimension for dominated affine iterated function systems 651

where p′B (x0 ) denotes the first derivative of the characteristic polynomial pB (x) :=
det(xI − B) evaluated at the point x0 . Define also
 
t1 (s) n−1 0 ··· 0 0
 t2 (s) t1 (s) n−2 ··· 0 0 
..
 
n
(−1)  t3 (s)

t2 (s) t1 (s) . 0 0 

an (s) := det  . .. .. .. .. .. 
n!  .. . . . . . 
 
tn−1 (s) tn−2 (s) tn−3 (s) · · · t1 (s) 1 
tn (s) tn−1 (s) tn−2 (s) · · · t2 (s) t1 (s)
for all n ≥ 1, and a0 (s) := 1. For each s ∈ [k, Pnk + 1] leti rn (s) denote the smallest
positive real root of the polynomial pn,s (x) := i=0 an (s)x . Then there exists n0 ∈ N
such that rn (s) is well-defined for all s ∈ [k, k + 1] and n ≥ n0 , and we have
1
eP (A1 ,...,AN ;s) − ≤ K exp (−γnα )
rn (s)
for some constants K, γ > 0 not depending on s ∈ [k, k + 1], where
d+1


−1
α := k+1
d+1

> 1.
k+1
− 2
Suppose additionally that there is a norm |||·||| on Rd such that max1≤i≤N |||Ai ||| < 1,
and that dimaff (A1 , . . . , AN ) ∈ (k, k + 1). Then for all sufficiently large n the function
s 7→ 1/rn (s) is strictly decreasing and convex on [k, k + 1] and there exists a unique
sn ∈ [k, k + 1] such that rn (sn ) = 1. There exist constants K ′ , γ ′ > 0 depending on
A1 , . . . , AN such that for all such n we have
|dimaff (A1 , . . . , AN ) − sn | ≤ K ′ exp (−γ ′ nα ) .
Since every matrix tuple is 0-multipositive, in the case k = 0 the hypothesis
of Theorem 2 reduces to the requirement that dimaff (A1 , . . . , AN ) is 1-multipositive
and dimaff (A1 , . . . , AN ) ∈ (0, 1). Since B ∧0 is the identity map on R the expressions
involving A∧k
i reduce to 1 in the case k = 0, resulting in the formula
X λ1 (Ai )d−1 ρ (Ai )s
tn (s) := .
p′Ai (λ1 (Ai ))
|i|=n

In particular when d = 2, k = 0 and the matrices Ai have positive entries we may

recover the conclusion of Theorem 1. Similarly, since every tuple in GLd (R)N is
∧(k+1)
d-multipositive and B ∧d = det B, the expressions involving Ai simplify when
k = d − 1 yielding
 d−1  d−s
∧(d−1) ∧(d−1)
X λ1 Ai ρ Ai | det Ai |s+1−d
tn (s) :=    .
′ ∧(d−1)
|i|=n p ∧(d−1) λ1 Ai
Ai

and the hypotheses are reduced to the requirement that (A1 , . . . , AN ) is (d − 1)-
multipositive and dimaff (A1 , . . . , AN ) ∈ (d − 1, d). We remark that hypotheses of
domination and positivity analogous to those in Theorem 2 have been a feature of
numerous recent works on affine iterated function systems such as [6, 7, 8, 20, 21] as
well as the older article [30].
If it is known that the tuple (A∧k ∧k k d
1 , . . . , AN ) preserves a single cone in ∧ R and
∧(k+1) ∧(k+1)
similarly (A1 , . . . , AN ) preserves a single cone in ∧k+1 Rd then the condition
652 Ian D. Morris

dimaff (A1 , . . . , AN ) ∈ (k, k + 1) may be easily checked. A theorem of Protasov [58]

implies that if B1 , . . . , BN preserve a cone then
N
! n1 N
!
X X
lim kBi1 · · · Bin k =ρ Bi ,
n→∞
i1 ,...,in =1 i=1

and so in this case

N
! n1 N
!
X X
lim ϕk (Ai1 · · · Ain ) =ρ A∧k
i ,
n→∞
i1 ,...,in =1 i=1

N
! n1 N
!
∧(k+1)
X X
lim ϕk+1 (Ai1 · · · Ain ) =ρ Ai
n→∞
i1 ,...,in =1 i=1

using the identity ϕℓ (B) = kB ∧ℓ k for ℓ = 0, . . . , d. It follows that in this situation

Theorem 2 is applicable if
N
! N
!
∧(k+1)
X X
ρ Ai <1<ρ A∧k
i .
i=1 i=1

An example of this situation is presented in §7 below.

In the situation where (A1 , . . . , AN ) fails to be both k- and (k + 1)-multipositive
we believe it to be unlikely that any analogue of Theorem 2 can be proved. The
precise role of the multipositive hypothesis is discussed in more detail in the following
section, and in the final section §8.

2. Overview of the method and statement of the main technical theorem

The method underlying Theorem 2 is, like Theorem 1, based on Fredholm deter-
minants of transfer operators, and in broad terms resembles many other arguments
of this type such as [33, 34, 36, 51, 54, 56, 57]. Both in order to give a sense of the
organisation of this article and to indicate those complications present in the proof of
Theorem 2 which do not occur in the context of Theorem 1 let us briefly describe this
strategy. For simplicity we will specialise our description to the situation in which the
transfer operators act on a Hilbert space, although this is not a strict requirement.
We recall that an operator L on an infinite-dimensional Hilbert space is called
trace-class if the sequence of approximation numbers
sn (L ) := inf {kL − F k : rank F < n}
is summable; we observe in particular that such an operator is compact (being a
limit in the norm topology of a sequence of finite-rank operators) and cannot be
invertible. We also observe that clearly sn (L ℓ ) ≤ kL ℓ−1 ksn (L ) for every n, ℓ ≥ 1
and consequently every power of a trace-class operator is also trace-class. The notion
of trace-class operator is reviewed in detail for the reader’s convenience in §4. Suppose
then that H is a separable complex Hilbert space and L : H → H a trace-class
linear operator, and let (λℓ )∞ℓ=1 be the sequence of nonzero eigenvalues of L listed
with repetition according to their algebraic multiplicity. (If only M < ∞ nonzero
eigenvalues exist then define λℓ = 0 for ℓ > M.) It is a classical fact that the function
 Fast approximation of the affinity dimension for dominated affine iterated function systems 653

z 7→ det(I − zL ) which may be defined2 by

M
Y
det(I − zL ) := (1 − zλℓ )
ℓ=1

is an entire function from

P∞ C toℓ C, and moreover one may show that in the power
series det(I − zL ) = n=0 aℓ z the coefficients are given by a0 = 1 and

ℓ−1 ···

tr L 0 0 0
 tr L 2 tr L ℓ−2 ··· 0 0 
.
 
(−1)ℓ  tr L 3

tr L 2 tr L .. 0 0 

aℓ = det  . .. .. .. .. .. 
ℓ!  .. . . . . . 
 
ℓ−1 ℓ−2 ℓ−3
tr L tr L tr L · · · tr L 1 
tr L ℓ tr L ℓ−1 tr L ℓ−2 · · · tr L 2 tr L
for ℓ ≥ 1. If we write
∞
X M
Y
ℓ
aℓ z = det(I − zL ) = (1 − zλℓ )
ℓ=0 ℓ=1

then by equating coefficients of z n we find (at least informally) that also

X
(1) an = (−1)n λ i1 · · · λ in .
i1 <i2 <···<in

for each n ≥ 1. Suppose now that we wished to calculate the spectral radius ρ(L ),
knowing the values of the traces L ℓ for ℓ = 1, . . . , n, say, and knowing also that the
spectral radius is an eigenvalue of L . The roots of det(I − zL ) are precisely the
reciprocals
P∞ of the eigenvalues of L and therefore ρ(L )−1 P
is the smallest positive root
ℓ n ℓ
of ℓ=0 aℓ z . In particular, the smallest P∞positive root of ℓ=0 aℓ z should be a good
−1
approximation to ρ(L ) as long as ℓ=n+1 |aℓ | is small. But if we are able to show
that the eigenvalues (λn ) decay exponentially (or even just stretched-exponentially)
in n, then the expression (1) implies a super-exponential decay estimate for the
coefficients an . Such an estimate will hold in particular if the approximation numbers
of L decay stretched-exponentially. In such a situation we may therefore reasonably
hope that the approximation procedure just outlined provides an estimate which
becomes super-exponentially more accurate as n increases.
In order to implement this line of reasoning we need therefore to construct,
for each s ∈ [k, k + 1], a trace-class operator Ls on a Hilbert space H such that
eP (A1 ,...,AN ;s) is an eigenvalue of Ls and is equal to the spectral radius of Ls , such
that Ls is trace-class, such that the sequence of approximation numbers of Ls de-
cays rapidly to zero, and such that the sequence of traces tr Lsn is easy to compute.
Once such a family of operators has been constructed the result follows by relatively
straightforward manipulations which, while they do not correspond precisely to any
prior work, share a degree of familial resemblance with calculations occurring in
numerous earlier articles such as [3, 32, 33, 34, 35, 36, 37, 40, 52, 53, 54, 55, 56, 57].
If V is a finite-dimensional real vector space let P V denote the real projective
space of lines through the origin in V . Intuitively, in order to construct an operator
2The Fredholm determinant is more usually defined first by its power series and shown later to
equal the infinite product given here, see e.g. [62]; we adopt this characterisation for simplicity of
presentation and because of its more direct connection with the problems being studied.
654 Ian D. Morris

Ls with spectral radius eP (A1 ,...,AN ;s) , we might consider an operator acting on some
space of continuous functions P (∧k Rd ) × P (∧k+1Rd ) → C defined by
!k+1−s  ∧(k+1) s−k
N ∧k
Ai u Ai v  
∧(k+1)
X
(Ls f ) (u, v) =   f A∧k i u, Ai v
i=1
kuk kvk

where for v ∈ V the notation v represents the one-dimensional subspace spanned by

the vector v. Since we would then have
!k+1−s  ∧(k+1) s−k
∧k
Ai u Ai v  
∧(k+1)
X
(Lsn f ) (u, v) =   f A∧k i u, Ai v
kuk kvk
|i|=n

for each n ≥ 1 we might then reasonably expect that

  n1   n1
1 k+1−s s−k
∧(k+1)
X X
lim kLsn k n = lim  A∧k
i Ai  = lim  ϕs (Ai )
n→∞ n→∞ n→∞
|i|=n |i|=n

so that eP (A1 ,...,AN ;s) is equal to the spectral radius of Ls . Indeed, such operators were
successfully constructed by Guivarc’h and Le Page on spaces of Hölder continuous
functions P (∧k Rd ) × P (∧k+1Rd ) → C in the article [29].
However, notwithstanding the (rather minor) additional complications posed by
the fact that the spaces defined above are not Hilbert, there is no reason to believe
that Ls acting on such a space should have a summable sequence of approxima-
tion numbers sn (Ls ). Indeed, Ls as constructed is equal to a sum of weighted
composition operators f 7→ g · f ◦ T where T is an invertible transformation of
P (∧k Rd ) × P (∧k+1Rd ) and g is nowhere zero. Such an operator might reasonably be
expected to be invertible, and there is certainly no reason to believe that Ls should
be trace-class.
The problem is thus to define Ls approximately as above in such a way that it is
a sum of trace-class, non-invertible operators. It is here that the hypothesis of k- and
(k + 1)-multipositivity becomes relevant: this hypothesis implies that for ℓ = k, k + 1
the matrices A∧ℓ ∧ℓ ℓ d
1 , . . . , AN map a finite union of patches of P (∧ R ) strictly inside
itself. By taking H to be a suitable Hilbert space of functions defined only on the
patches, composition with the projective action of the matrices should then induce
an operator which is non-invertible and hopefully trace-class. It transpires that
composition operators on spaces of holomorphic functions are reliably trace-class
subject to moderate geometrical conditions, and as such our strategy will involve
passing to a space of holomorphic functions defined on complex extensions of the
patches in real projective space. Once we have verified that such an extension can
be constructed in such a way that the operator Ls is well-defined on the patches we
may proceed to prove Theorem 2 along the lines outlined above.
In the two-dimensional context of Theorem 1 the construction of these complex
patches is very straightforward. Since Theorem 1 is restricted to affine transforma-
tions whose linear parts contract the positive cone in R2 , it is sufficient to consider
the projective action of those linear maps on the interval {(x, 1 − x) : x ∈ [0, 1]},
which is an action by linear fractional transformations. A finite collection of linear
fractional transformations each of which maps an interval strictly inside itself can
easily be shown to also map a corresponding complex disc inside itself, and this
complex disc can be used as the domain of the holomorphic functions on which the
 Fast approximation of the affinity dimension for dominated affine iterated function systems 655

operator Ls acts. In higher dimensions and using multicones instead of cones, the
corresponding problem is to understand (in place of one-dimensional intervals) a
family of (d − 1)-dimensional sections of cones in Rd – in effect, a finite collection
of arbitrary (d − 1)-dimensional convex bodies – and a collection of linear fractional
transformations between them, and to contrive a system of extensions of those con-
vex bodies into Cd−1 which is also preserved by the same family of linear fractional
transformations. This much more involved procedure is undertaken in §3 and lays
the foundation for following technical theorem which is obtained subsequently:
Theorem 3. Let d, N ≥ 2 and let (A1 , . . . , AN ) ∈ Md (R)N be both k-multipositive
and (k + 1)-multipositive, where 0 ≤ k < d. Then there exist a separable complex
Hilbert space H and a family of bounded linear operators Ls : H → H defined for
all s ∈ C with the following properties:
(i) There exist C, κ, γ
> 0 such that for all s ∈ C and n ≥ 1 we have sn (L ) ≤
C exp κ|s| − γnβ , where
1
β := d+1

∈ (0, 1].
k+1
−2
In particular each Ls is trace-class.
(ii) For every s ∈ C and n ≥ 1 we have
 d −1

(kd)−1 ∧(k+1) (k+1)
k+1−s  ∧(k+1) s−k

X λ1 A∧k
i λ1 Ai ρ A∧k
i ρ Ai
tr Lsn =   
′ ∧k


′ ∧(k+1)
|i|=n pA∧k λ1 Ai p ∧(k+1) λ1 Ai
i Ai

where pB (x) := det(xI − B) denotes the characteristic polynomial of B and

p′B (x0 ) its derivative evaluated at x0 .
(iii) For every s ∈ R the spectral radius of Ls is equal to
1 X k+1−s ∧(k+1)
s−k
lim log A∧k
i Ai .
n→∞ n
|i|=n

In particular the above limit exists for all s ∈ R, and for every s ∈ [k, k + 1]
the spectral radius of Ls is equal to eP (A1 ,...,AN ;s). For all s ∈ R the spectral
radius of Ls is a simple eigenvalue of Ls and there are no other eigenvalues
of the same modulus.
Theorem 3 is a special case of a slightly more general result, Theorem 11, which
will be proved later. Theorem 11 is also applied in the sequel article [49] to the
estimation of a related invariant of tuples of matrices.
The remainder of this article is structured as follows. In §3 we undertake the
construction of the complex extensions of the patches in real projective space. We
then review in §4 the properties of trace-class operators which will be needed in this
article and extend a standard result from this context in view of the fact that we will
be working with spaces of holomorphic functions defined on a non-connected region.
We then proceed in §5 to establish the properties of the operator Ls and deduce
Theorem 3. In §6 we derive Theorem 2 from Theorem 3 above. Some examples of
the application of Theorem 2 are presented in §7. In §8 we consider the problem of
calculating the affinity dimension in situations where the hypotheses of Theorem 2
do not apply.
656 Ian D. Morris

We remark that sections 6–8 depend only on the statement of Theorem 3 and
the material presented in sections 1 and 2 and as such may be read independently of
sections 3–5 in which the proof of Theorem 3 is prepared for and presented.

3. Complex domains for linear semigroups acting on a multicone

Our first task in proving Theorem 2 is to translate the matter from the context of
linear maps between real cones to the context of holomorphic maps between complex
domains. We will prove the following:
Theorem 4. Let d ≥ 1 and let (K1 , . . . , Km ), (K1′ , . . . , Km′
) be multicones in Rd ,
d ′
both with transverse-defining vector w ∈ R , such that Kj \ {0} ⊂ Int Kj for each
j = 1, . . . , m. Define
( m
! m
)
[ [
Kj′ ∪ −Kj′


A := A ∈ Md (R) : A Kj ⊆
j=1 j=1
∗
and let A denote the set of all nonzero elements of A. We observe that A is a
semigroup. Then there exists a subset Ω of the complex hyperplane {z ∈ Cd : hz, wi =
1} such that the following properties are satisfied by A∗ and Ω:
(i) There is a constant τ > 0 such that kA1 A2 k ≥ τ kA1 k·kA2 k for every A1 , A2 ∈
A. In particular A∗ is a subsemigroup of A.
(ii) The set Ω is open and bounded and is symmetric with respect to complex
conjugation. Every connected component of Ω intersects Rdi . The closures of
the connected components of Ω are disjoint.
(iii) There exists C > 0 such that
C −1 kAk ≤ |ℜ(hAz, wi)| ≤ |hAz, wi| ≤ CkAk
for all A ∈ A and z ∈ Ω.
(iv) Every A ∈ A∗ induces a well-defined holomorphic transformation φA : Ω → Ω
defined by φA (z) := hAz, wi−1 Az. The set
[
φA (Ω)
A∈A∗

is a compact subset of Ω.
(v) There exist a metric d on Ω which is bi-Lipschitz equivalent to the standard
metric and a constant θ ∈ (0, 1) such that d(φA (z1 ), φA (z2 )) ≤ θd(z1 , z2 ) for
every A ∈ A∗ .
(vi) Let A ∈ A∗ . Then the largest eigenvalue λ1 (A) of A is algebraically simple,
is real, is strictly larger in modulus than all of the other eigenvalues of A,
and has a corresponding eigenvector zA ∈ Ω ∩ Rd which is the unique fixed
point of φA : Ω → Ω. The eigenvalues of the derivative DzA φA are precisely
the numbers λj (A)/λ1 (A) for j = 2, . . . , d, and in particular
p′A (λ1 (A))
det(I − DzA φA ) = 6= 0
λ1 (A)d−1
where pA (x) := det(xI − A) denotes the characteristic polynomial of A and
p′A its first derivative.
Theorem 4 is trivial in the case d = 1 and for the remainder of this section we shall
ignore this case, assuming at all times that d ≥ 2. (When d = 1 the determinant in
(4) above will be interpreted as being equal to 1.) Here and throughout the remainder
 Fast approximation of the affinity dimension for dominated affine iterated function systems 657

of this article we use the notation z ∗ to denote the complex conjugate of z ∈ C and
reserve the notation z for the one-dimensional subspace spanned by z.
Using the machinery of complex cones and gauges (see [15, 60]) it is possible to
obtain Theorem 4 by extending each real cone Kj to a complex cone
KjC := {λ((u + v) + i(u − v)) : λ ∈ C and u, v ∈ Kj }
and considering the projective action on a slice through the complex extension of the
union of the cones K1 , . . . , Km ,
( m
)
[
Ω := z ∈ Cd : z ∈ Int KjC and hz, wi = 1 .
j=1

This procedure has the advantage of explicitness and may be a useful direction of
research in the event that effective versions of Theorem 2 are sought. It is on the other
hand somewhat laborious to implement, and since our interest is only in establishing
the correctness of the formulas in Theorem 2 and giving a super-exponential bound
for the error term, we pursue a simpler but less explicit construction along the lines
of [2, §2].
3.1. The action on the real multicone. We begin by establishing some
preliminary results concerning the action of A on the real cones K1 , . . . , Km and
proceed to prove Theorem 4 in the following subsection.
Lemma 3.1. Let d ≥ 1 and let (K1 , . . . , Km ), (K1′ , . . . , Km′
) be multicones in Rd ,
d ′
both with transverse-defining vector w ∈ R , such that Kj \ {0} ⊆ Int Kj for each
j = 1, . . . , m. Define
( m
! m
)
[ [
Kj′ ∪ −Kj′


A := A ∈ Md (R) : A Kj ⊆
j=1 j=1

and observe that A is a semigroup. Then there exists τ ∈ (0, 1] such that:
(i) For every u ∈ m
S
j=1 Kj we S
have τ kuk ≤ hu, wi ≤ kuk.
(ii) For every A ∈ A and u ∈ m ′
j=1 Kj we have kAuk ≥ τ kAk · kuk.
(iii) For every A1 , A2 ∈ A we have kA1 A2 k ≥ τ kA1 k · kA2 k. In particular the set
of all nonzero elements of A is a subsemigroup of A.
Proof. We will allow the constant τ > 0 to be different in each of (i),(ii) and (iii),
which obviously suffices. To prove (i) it is sufficient, by homogeneity, to consider
Sm kuk = 1. The function u 7→ hu, wi is obviously continuous
only those cases in which
on the set of all u ∈ j=1 Kj such that kuk = 1 and is positive everywhere on this
set by the definition of a multicone. Since this set is compact this function attains
its minimum, so this minimum is positive; call it τ . We have 0 < τ ≤ hu, wi ≤ 1 for
all u ∈ m
S
j=1 j with kuk = 1 and the result follows.
K
By homogeneity in A and u it is sufficient to prove (ii) in the case kAk = kuk = 1.
By a similar S
compactness argument it suffices to show that Au may not be zero when
A ∈ A, u ∈ m ′
j=1 Kj and kAk = kuk = 1. For a contradiction suppose that we may
find such A and u satisfying Au = 0. Since A 6= 0 there exists a unit vector v such
that Av 6= 0. Since u is a nonzero element of some Kj′ it is an interior point of the
corresponding cone Kj and therefore there exists ǫ > 0 such that u + ǫv and u − ǫv
both belong to Kj . But this implies that A(u + ǫv) = ǫAv and A(u − ǫv) = −ǫAv
are both nonzero elements of AKj . Since AKj ⊆ Ki ∪ −Ki for some i we deduce that
658 Ian D. Morris

Av ∈ (Ki ∩ −Ki ) \ {0} contradicting the definition of a multicone.

Sm The result follows.
′
To deduce (iii) we observe that for any unit vector u ∈ j=1 Kj we have
kA1 A2 k ≥ kA1 A2 uk ≥ τ kA1 k · kA2 uk ≥ τ 2 kA1 k · kA2 k · kuk = τ 2 kA1 k · kA2 k
by repeated application of (ii). 
The following Perron–Frobenius result does not follow in a completely direct
manner from standard statements of the Perron–Frobenius theorem for cones since
it is possible for A(Ki \ {0}) to include the zero vector, preventing the direct use of
off-the-shelf results.
Lemma 3.2. Let d, w, (K1 , . . . , Km ), (K1′ , . . . , Km
′
) and A∗ be as in the statement
of Theorem 4. Suppose that A ∈ A satisfies AKi ⊆ Ki′ for some i ∈ {1, . . . , m}.
∗

Then ρ(A) is an algebraically simple eigenvalue of A with corresponding eigenvector

in Ki′ and all other eigenvalues of A are of strictly smaller absolute value.
Proof. Choose a cone Ki′′ such that Ki′′ \ {0} ⊂ Int Ki and Ki′ \ {0} ⊆ Int Ki′′ . We
observe that the single matrix A, the one-element multicone (Ki ) and the one-element
multicone (Ki′′ ) together satisfy the hypotheses of Lemma 3.1, and by part (ii) of that
lemma it follows that Av is not the zero vector for any nonzero v ∈ Ki′′ . In particular
A(Ki′′ \ {0}) ⊆ Ki′ \ {0} ⊆ Int Ki′′ and standard versions of the Perron–Frobenius
Theorem such as [11, Theorem 1.3.26] may be applied to the action of A on Ki′′ . The
result follows. 
Proposition 3.3. Let d, w, (K1 , . . . , Km ), (K1′ , . . . , Km
′
) and A∗ be as in the
statement of Theorem 4. Then there exist C > 0 and θ ∈ (0, 1) such that for all
j = 1, . . . , m and n ≥ 1, for all nonzero v1 , v2 ∈ Kj′ ,
A1 · · · An v1 A1 · · · An v1
sup − ≤ Cθn kv1 − v2 k.
A1 ,...,An ∈A∗ hA1 · · · An v1 , wi hA1 · · · An v1 , wi
Proof. For every nonzero v ∈ Rd let v denote the one-dimensional subspace of Rd
spanned by v, and let Kj / ∼ denote the set of one-dimensional subspaces spanned
by an element of Int Kj . For each j = 1, . . . , m define
α(v1 , v2 ) := sup {λ ≥ 0 : v2 − λv1 ∈ Kj }
and
β(v1 , v2 ) := inf {λ ≥ 0 : λv1 − v2 ∈ Kj }
for all v1 , v2 ∈ Int Kj ; then the formula
β(v1 , v2 )
dKj (v1 , v2 ) := log
α(v1 , v2 )
defines a metric on Kj / ∼ called the S Hilbert projective metric. It follows from
Lemma 3.1(i) that the set of all v ∈ m j=1 Kj such that hv, wi = 1 is bounded. By
compactness it follows that there exists ε ∈ (0, 1] such that for every j = 1, . . . , m,
if v ∈ Kj′ with hv, wi = 1 then the open Euclidean ε-ball centred at v is a subset
of Kj . We deduce that if v1 , v2 ∈ Kj′ with hv1 , wi = hv2 , wi = 1 then since kv1 k,
kv2 k ≤ τ −1 by Lemma 3.1(i) we have α(v1 , v2 ) ≥ ετ and β(v1 , v2 ) ≤ ε−1 τ −1 , and
hence the quantity
∆ := max sup dKj (v1 , v2 )
1≤j≤m v ,v ∈K′
1 2 j
v1 ,v2 6=0
 Fast approximation of the affinity dimension for dominated affine iterated function systems 659

is finite. In particular if v1 , v2 ∈ Kj \ {0} for some j ∈ {1, . . . , m}, and A ∈ A∗ ,

then dKi (Av1 , Av2 ) ≤ ∆ where i is the unique integer such that AKj ⊆ Ki ∪ −Ki . It
follows by e.g. [42, Theorem 1.1] that if A ∈ A∗ , v1 , v2 ∈ Kj / ∼ and AKj ⊆ Ki ∪ −Ki
then we have dKi (Av1 , Av2 ) ≤ θdKj (v 1 , v2 ) where θ := tanh(∆/4) ∈ (0, 1).
We claim that there exists C1 > 0 such that if v1 , v2 ∈ Kj′ with hv1 , wi = hv2 , wi =
1 then  
C1−1 kv1 − v2 k ≤ edKj (v1 ,v2 ) − 1 ≤ C1 kv1 − v2 k.
Indeed, given such vectors v1 , v2 ∈ Kj′ with v1 6= v2 let α := α(v1 , v2 ) and β :=
β(v1 , v2 ). Since Kj is closed the supremum in the definition of α is attained, and
therefore we have v2 − αv1 ∈ Kj . Similarly we have βv1 − v2 ∈ Kj . From the
maximality of α and the minimality of β it follows that v2 − αv1 and βv1 − v2 are
boundary points of Kj . Since hv1 −v2 , wi = 0 and v1 −v2 6= 0 neither v1 −v2 nor v2 −v1
can belong to Kj , so neither α nor β may equal 1 and we deduce that α < 1 < β.
To obtain the first of the two claimed inequalities we observe that βv1 − v2 and
(1 −α)v1 belong to Kj , where we have used 0 < α < 1. Hence (β −α + 1)v1 −v2 ∈ Kj .
If τ > 0 is as given by Lemma 3.1(i) then we have
τ kv1 − v2 k ≤ τ k(β − α + 1)v1 − v2 k + τ k(β − α)v1 k
≤ |h(β − α + 1)v1 − v2 , wi| + (β − α)|hv1 , wi|
 
β  
≤ 2(β − α) < 2 − 1 = 2 edKj (v1 ,v2 ) − 1
α
where we have again used 0 < α < 1 in the final line. This yields the first inequality.
1 1
To obtain the second inequality define u1 := β−1 (βv1 − v2 ) and u2 := 1−α (v2 − αv1 ).
We observe that both u1 and u2 belong to the boundary of Kj , which implies ku1 −
v1 k, ku2 − v2 k ≥ ετ by the definition of ε and the bound kv1 k, kv2 k ≤ τ −1 . We now
observe that
β ku1 − v2 k · ku2 − v1 k
edKj (v1 ,v2 ) = =
α ku1 − v1 k · ku2 − v2 k
  
ku1 − v1 k + kv1 − v2 k ku2 − v2 k + kv2 − v1 k
≤
ku1 − v1 k ku2 − v2 k
2
≤ 1 + ε−1 τ −1 kv1 − v2 k

and therefore
   
dKj (v1 ,v2 ) 2 1 4
e −1≤ + 2 2 kv1 − v2 k kv1 − v2 k ≤ kv1 − v2 k
ετ ετ ε2 τ 3
where we have again used kv1 k,kv2 k ≤ τ −1 . The claim follows.
We may now prove the proposition. Given n ≥ 1, nonzero v1 , v2 ∈ Kj′ and
A1 , . . . , An ∈ A∗ , let i ∈ {1, . . . , m} be the integer such that An · · · A1 Kj ⊆ Ki ∪ −Ki .
We have
An · · · A1 v1 An · · · A1 v2  
− ≤ C1 edKi (An ···A1 v1 ,An ···A1 v2 ) − 1
hAn · · · A1 v1 , wi hAn · · · A1 v2 , wi
 n 
≤ C1 eθ dKj (v1 ,v2 ) − 1
≤ C12 θn kv1 − v2 k
and the proposition is proved. 
660 Ian D. Morris

While Proposition 3.3 will provide us with a vital contraction estimate for maps
between specific cones Kj , in order to apply it we will need the following combinatorial
lemma which allows us to reduce the action of a specific matrix product on the
multicone to that on a single cone:
Lemma 3.4. Let d, w, (K1 , . . . , Km ), (K1′ , . . . , Km
′
) and A∗ be as in the statement
of Theorem 4. Let k ≥ 2m − m − 1. Then for every A1 , . . . , Ak ∈ A∗ there exists
i ∈ {1, . . . , m} such that
m
!
[
Ak · · · A1 Kj ⊆ Ki ∪ −Ki .
j=1

Proof. It is clearly sufficient to consider the case k = 2m − m − 1 only. Let

I0 := {1, . . . , m} and for each n = 1, . . . , 2m − m − 1 let In denote the intersection
of all sets I ⊆ {1, . . . , n} such that
m
!
[ [
(2) An · · · A1 Kj ⊆ (Ki ∪ −Ki )
j=1 i∈I

where the union over an empty set of indices i is understood to be {0}. We observe
that I = In itself satisfies (2). By Lemma 3.1(iii) the product A2m −m−1 · · · A1 is not
the zero matrix and therefore In is nonempty. We observe that the cardinality of In
is non-increasing as a function of n.
We wish to prove that I2m −m−1 has cardinality 1, so for a contradiction let us
suppose that its cardinality is at least 2. This implies that every preceding In also
has cardinality at least 2, and also that m ≥ 2. Since the number of subsets of
{1, . . . , m} with cardinality at least 2 is 2m − m − 1, by the pigeonhole principle there
exist integers n1 , n2 with 0 ≤ n1 < n2 ≤ 2m − m − 1 such that In1 = In2 . The matrix
B := An2 · · · An1 +1 therefore takes each cone Ki such that i ∈ In1 to a nontrivial
subset of some cone Kj such that j ∈ In1 , inducing a permutation on the elements
of In1 = In2 . It follows that the matrix B #In1 induces the identity permutation on
In1 : for every i ∈ In1 we have B #In1 Ki ⊆ (Ki ∪ −Ki ). Hence B 2#In1 Ki ⊆ Ki′ for
every i ∈ In1 . By Lemma 3.2, for every i ∈ In1 the matrix B 2#In1 has a simple
positive leading eigenvalue with a one-dimensional eigenspace which intersects Ki
nontrivially: but since #In1 ≥ 2 and distinct cones Ki do not intersect this implies
that the leading eigenvalue is not simple, which is a contradiction. 
3.2. Proof of Theorem 4. Throughout the proof we fix d, A, (K1 , . . . , Km ),
(K1′ , . . . , Km
′
and w as in the statement of the theorem. Part (i) of Theorem 4 follows
)
directly from Lemma 3.1 so we concentrate on parts (ii) to (vi).
Define H := {z ∈ Cd : hz, wi = 1} and let Kj := Kj′ ∩ H for each j = 1, . . . , m.
Each Kj is closed by definition and is bounded as a consequence of Lemma 3.1(i).
For each n ≥ 1 define
A∗n := {A1 · · · An : A1 , . . . , An ∈ A∗ } .
Define a function M : H → [0, +∞) by
M(z) := inf {|hAz, wi| : A ∈ A and kAk = 1} .
Clearly M(z) is well-defined and
M(z) = inf kAk−1 |hAz, wi| : A ∈ A∗ .

 Fast approximation of the affinity dimension for dominated affine iterated function systems 661

We observe that M is 1-Lipschitz continuous: given z1 , z2 ∈ U and A ∈ A with

kAk = 1 we have

|hAz2 , wi| ≥ |hAz1 , wi| − |hA(z1 − z2 ), wi| ≥ M(z1 ) − kz1 − z2 k

and taking the infimum over A and rearranging easily yields M(z1 ) − M(z2 ) ≤
kz1 − z2 k. The result followsSby symmetry. The set U := {z ∈ H : M(z) 6= 0} is
consequently open. We have m j=1 Kj ⊆ U by Lemma 3.1(i) and (ii) and in particu-
lar U is nonempty.
We now claim that if A ∈ A∗ and z ∈ U then necessarily hAz, wi−1Az ∈ U. If
this is not the case for some A and z then by compactness there exists B ∈ A∗ with
kBk = 1 such that hB(hAz, wi−1 Az), wi = 0, but then necessarily hBAz, wi = 0
which contradicts z ∈ U since obviously BA ∈ A∗ by the semigroup property of
A∗ . The claim is proved. We deduce that for every nonzero A ∈ A the formula
φA (z) := hAz, wi−1 Az gives rise to a well-defined holomorphic function φA : U → U.
We observe that φA ◦ φB = φAB for all A, B ∈ A∗ and that φtA = φA for all real t > 0
and all A ∈ A∗ .
Let τ > 0 be as given by Lemma 3.1 and observe that
( m
)
[
sup kzk : z ∈ Kj ≤ τ −1
j=1

by Lemma 3.1(i) and

( m
)
[
inf kAk−1 |ℜ(hAz, wi)| : z ∈ Kj and A ∈ A∗ ≥ τ2
j=1

by Lemma 3.1(i) and (ii). For each j = 1, . . . , m define

 
Uj := z ∈ U : inf kz − ωk < ǫ
ω∈Kj

where ǫ > 0 is chosen small enough that the following properties hold: the sets
τ2 −1
Sm pairwise disjoint closures; |ℜ(hAz, wi)| ≥ 2 kAk and kzk ≤ 2τ
Uj have for all
z ∈ j=1 Uj and all A ∈ A; and

1
(3) (256τ −10 + 4τ −4 )ǫ < .
4
The second condition is possible since the function z 7→ inf{kAk−1 |ℜ(hAz, wi)| : A ∈
A∗ } is 1-Lipschitz continuous for the same reasons as M. Each Kj is convex as a
consequence of the definition of a multicone, so each Uj is convex also.
Now let C1 , θ1 be the constants given by Proposition 3.3 and let n1 ≥ 1 be large
enough that C1 θ1n1 < 41 . We claim that for every A ∈ A∗n1 the map φA satisfies
kDz φA k ≤ 21 for all z ∈ m ∗ n1 1
S
Sm j=1 Uj . Fix A ∈ An1 and observe that kDω φA k ≤ C1 θ1 < 4
for all ω ∈ j=1 Kj by Proposition 3.3.
By simple direct calculation, for all v ∈ Cd such that hv, wi = 0 and all z ∈ U
we have
(Dz φA )(v) = hAz, wi−2 (hAz, wiAv − hAv, wiAz) .
662 Ian D. Morris

It follows that if z1 , z2 ∈ Uj , then

(Dz1 φA − Dz2 φA )(v)
= hAz1 , wi−2hAz2 , wi−2 hAz2 , wi2 − hAz1 , wi2 (hAz1 , wiAv − hAv, wiAz1)

+ hAz2 , wi−2 (hA(z1 − z2 ), wiAv − hAv, wiA(z1 − z2 ))

for all v in the tangent space {v ∈ Cd : hz, wi = 0}. Since |hAz, wi|−2 ≤ 4τ −4 kAk−2
for all z ∈ Uj by the definition of Uj , this yields the estimate
kDz1 φA − Dz2 φA k ≤ (16τ −8 (kz1 k + kz2 k)(2kz1 k) + 4τ −4 )kz1 − z2 k
≤ (256τ −10 + 4τ −4 )kz1 − z2 k
where we have used the bound kzk ≤ 2τ −1 which applies to all z ∈ Uj . In particular
if z ∈ Uj is arbitrary and ω ∈ Kj is chosen such that kz − ωk < ǫ, taking z1 := z and
z2 := ω and applying (3) together with kDω φA k < 41 yields kDz φA k < 12 . We conclude
that max1≤j≤m supz∈Uj kDz φA k ≤ 21 . Since each Uj is convex it follows by the mean
value inequality that for every A ∈ A∗n1 and j = 1, . . . , m the map φA : Uj → U is
1
2
-Lipschitz continuous with respect to the Euclidean metric. (We observe that this
does not imply 21 -Lipschitz continuity on m
S
j=1 Uj .) It follows that for every k ≥ 1
∗ 1
and A ∈ Akn1 the map φA is 2k -Lipschitz on each Uj , which will be used later.
We next observe that for every A ∈ A∗ and z1 , z2 ∈ U there holds the Lipschitz
continuity estimate
(4) kφA (z1 ) − φA (z2 )k ≤ M(z1 )−1 M(z2 )−1 min{kz1 k, kz2 k}kz1 − z2 k.
Clearly it is sufficient to prove this in the case kAk = 1. In this case we observe that
kφA (z1 ) − φA (z2 )k = |hAz1 , wihAz2, wi|−1 khAz2 , wiAz1 − hAz1 , wiAz2 k
≤ M(z1 )−1 M(z2 )−1 khAz2 , wiAz1 − hAz1 , wiAz2k
= M(z1 )−1 M(z2 )−1 khAz2 , wiA(z1 − z2 ) + hA(z2 − z1 ), wiAz2 k
≤ M(z1 )−1 M(z2 )−1 kz2 k · kz1 − z2 k,
and performing the same calculation with z1 and z2 interchanged obviously yields
(4). As a consequence we have
(5) kφA (z1 ) − φA (z2 )k ≤ C2 kz1 − z2 k
Sm
for every z1 , z2 ∈ j=1 Uj and A ∈ A∗ where C2 := 8τ −5 , using the inequalities
M(z) ≥ τ 2 /2 and kzk ≤ 2τ −1 which follow from the definition of the sets Uj .
Let n2 ≥ 2m − m − 1 be an integer such that for every A ∈ A∗ the map φA |Uj
is 1-Lipschitz continuous for every j = 1, . . . , m. (Note that every sufficiently large
multiple of n1 has this property.) Fix k large enough that C2 2−k ≤ 21 and define a
metric d on U by
kn1X
+n2 −1
n
d(z1 , z2 ) := 2 kn1 +n2 sup kφA (z1 ) − φA (z2 )k
n=0 A∈A∗n

where the summand corresponding to n = 0 is understood as kz1 − z2 k. It follows

from (4) that d(z1 , z2 ) is well-defined for all z1 , z2 ∈ U and its property of being a
metric is obvious. For z1 , z2 ∈ m
S
j=1 j we additionally have
U
!
C2
kz1 − z2 k ≤ d(z1 , z2 ) ≤ 1 kz1 − z2 k
2 kn1 +n2 − 1
 Fast approximation of the affinity dimension for dominated affine iterated function systems 663

by applying (5) and summing the geometric series, soSd is bi-Lipschitz equivalent to
the Euclidean distance when considered as a metric on m j=1 Uj . We observe that since
every A ∈ A∗ is real, the metric d is symmetric with respect to complex conjugation:
d(z1 , z2 ) = d(z1∗ , z2∗ ) for all z1 , z2 ∈ U.S
We claim that for every z1 , z2 ∈ m ∗
j=1 Uj and B ∈ A we have
1
−
(6) d(φB (z1 ), φB (z2 )) ≤ 2 kn1 +n2 d(z1 , z2 ).
Sm
To see this let z1 , z2 ∈ j=1 Uj and B ∈ A∗ . We have
kn1X
+n2 −1
n
d(φB (z1 ), φB (z2 )) = 2 kn1 +n2 sup kφA (φB (z1 )) − φA (φB (z2 ))k
n=0 A∈A∗n

knX
1 +n2
n−1
≤ 2 kn1 +n2 sup kφA (z) − φA (ω)k
n=1 A∈A∗n

kn1X
+n2 −1
1 n
− kn
=2 1 +n2 2 kn1 +n2 sup kφA (z1 ) − φA (z2 )k
n=1 A∈A∗n
kn1 +n2 −1
+2 kn1 +n2
sup kφA (z1 ) − φA (z2 )k.
A∈A∗kn
1 +n2

To prove the claimed inequality it therefore suffices to show that

1
kφA (z1 ) − φA (z2 )k ≤ kz1 − z2 k
2
1
−
for all A ∈ A∗kn1 +n2 , since then the final term above is bounded by 2 kn1 +n2 kz1 − z2 k
and simple rearrangement yields (6). Now, if A ∈ A∗kn1 +n2 let us write S A = A1 A2
where A1 ∈ Akn1 and A2 ∈ An2 . By Lemma 3.4 there exists i such that A2 ( m
∗ ∗
j=1 Kj ) ⊆
Sm ′ ′ ′
Sm
Ki ∪ −Ki , and this clearly implies A2 ( j=1 Kj ) ⊆ Ki ∪ −Ki . Hence φA2 ( j=1 Kj ) ⊆
Ki . Choose j1 , j2 such that z1 ∈ Uj1 and z2 ∈ Uj2 and choose ω1 ∈ Kj1 and ω2 ∈ Kj2
such that kz1 − ω1 k < ǫ and kz2 − ω2 k < ǫ. We have φA2 (ω1 ), φA2 (ω2 ) ∈ Ki ,
kφA2 (z1 ) − φA2 (ω1 )k < ǫ by the 1-Lipschitz continuity of φA2 restricted to Uj1 , and
likewise kφA2 (z2 ) − φA2 (ω2 )k < ǫ. Thus φA2 (z1 ) and φA2 (z2 ) are both elements of Ui
and they satisfy kφA2 (z1 ) − φA2 (z2 )k ≤ C2 kz1 − z2 k by (5). But φA1 is 21k -Lipschitz
when restricted to Ui , so
kφA1 (φA2 (z1 )) − φA1 (φA2 (z2 ))k ≤ 2−k kφA2 (z1 ) − φA2 (z2 )k
1
≤ C2 2−k kz1 − z2 k ≤ kz1 − z2 k
2
and the claim follows.
For each j = 1, . . . , m define Ωj to be the unique connected component of the set
 
(7) z ∈ U : inf d(z, ω) < ǫ
ω∈Kj

which intersects Kj . Obviously we have Kj ⊂ Ωj . Since kz1 − z2 k ≤ d(z1S , z2 ) for

all z1 , z2 ∈ U we also have Ωj ⊆ Uj for each j = 1, . . . , m. Define Ω := m j=1 Ωj
Sm
and observe that z ∈ Ω if and only if z ∈ Ω by the fact that j=1 Kj ⊂ Rd and
∗

the fact that d is conjugation-symmetric. Since every Uj is bounded, so is every Ωj

and therefore so is Ω. The connected components of Ω are precisely the sets Ωj and
these have disjoint closures since this is true of the sets Uj which contain them. Each
664 Ian D. Morris

Ωj contains the corresponding set Kj ⊂ Rd and in particular intersects Rd . This

completes the proof of (4). We have

τ2
kAk ≤ |ℜ(hAz, wi)| ≤ |hAz, wi| ≤ 2τ −1 kAk
2
for all z ∈ Ω and A ∈ A as a consequence of the definition of U1 , . . . , UmS, and this
completes the proof of (4). Since for every A ∈ A∗ the function φA maps m j=1 Kj to
a subset of itself, and φA contracts distances between points in m
S
U
j=1 j with respect
−1/(kn1 +n2 )
to d by a factor of θ2 := 2 , it follows that
[  
φA (Ω) ⊆ z ∈ U : Sinf m
d(z, ω) ≤ θ2 ǫ
ω∈ j=1 Kj
A∈A∗

which is a compact subset of the set defined in (7). Each φA (Ωj ) is a connected
subset of the set defined above and intersects one of the sets Ki , hence it is a subset
of the set defined in (7) and intersects Ki , hence is a subset of the corresponding set
S
Ωi , hence is a subset of Ω. We conclude that A∈A∗ φA (Ω) is a compact subset of Ω.
This completes the proof of (4) and (4).
remains only to prove (4). Fix A ∈ A∗ . Since An2 ∈ A∗n2 the matrix An2
It S
maps m j=1 Kj into (Ki ∪ −Ki ) for some i ∈ {1, . . . , m} and in particular A
2n2
maps
′ 2n2
Ki into Ki . It follows by Lemma 3.2 that A has an algebraically simple leading
eigenvalue which is real and positive, has corresponding eigenvector vA in Ki′ and is
the unique eigenvalue with maximal modulus. Hence A has an algebraically simple
leading eigenvalue λ1 (A) which is real (but may be negative), is the unique eigenvalue
of maximal modulus, and satisfies AvA = λ1 (A)vA . Defining zA := hvA , wi−1vA we
have zA ∈ Ki ⊂ Ω ∩ Rd . Obviously φA zA = zA and hAzA , wi = λ1 (A). By (4) there
can be no other fixed points for φA in Ω.
Let us now calculate the eigenvalues of the derivative DzA φA . Let u1 , . . . , ud ∈ Cd
be a Jordan basis for A with basis elements listed in descending order of the absolute
value of the corresponding eigenvalue, and with u1 = zA . Since |λ1 (A)| > |λ2 (A)| we
have Au1 = λ1 (A)u1 and Au2 = λ2 (A)u2 . For each j ∈ {3, . . . , d}, let δj ∈ {0, 1}
such that Auj = λj (A)uj + δj uj−1.
For every v in the tangent space {v ∈ Cd : hv, wi = 0} to Ω at zA we have
 
1 A(u1 + εv) Au1
(DzA φA ) v := lim −
ε→0 ε hA(u1 + εv), wi hAu1 , wi
hAu1 , wi · Av − hAv, wi · Au1 1
= = (Av − hAv, wiu1) .
hAu1 , wihAu1, wi λ1 (A)

Clearly the vectors vj := uj − huj , wiu1, where j runs from 2 to d, form a basis of
the tangent space {z ∈ Cd : hz, wi = 0}. We have
1
(DzA φA ) v2 = (Av2 − hAv2 , wiu1)
λ1 (A)
1
= (λ2 (A)u2 − λ1 (A)hu2 , wiu1 − λ2 (A)hu2, wiu1 + λ1 (A)hu2, wiu1)
λ1 (A)
1 λ2 (A)
= (λ2 (A)u2 − λ2 (A)hu2 , wiu1) = v2 ,
λ1 (A) λ1 (A)
 Fast approximation of the affinity dimension for dominated affine iterated function systems 665

and for j = 3, . . . , d we similarly have

1
(DzA φA ) vj = (Avj − hAvj , wiu1)
λ1 (A)
1
= λj (A)uj + δj uj−1 − λ1 (A)huj , wiu1
λ1 (A)


− λj (A)huj , wiu1 − δj huj−1, wiu1 + λ1 (A)huj , wiu1
1
= (λj (A)uj − λj (A)huj , wiu1 + δj uj−1 − δj huj−1, wiu1)
λ1 (A)
λj (A) δj
= vj + vj−1 .
λ1 (A) λ1 (A)
It follows that with respect to the basis v2 , . . . , vd the matrix of DzA φA is upper
triangular with the values λj (A)/λ1 (A) along the diagonal. In particular its eigen-
values are precisely the numbers λj (A)/λ1 (A) for j = 2, . . . , d as claimed. Since
pA (x) = det(xI − A) = dj=1 (x − λj (A)) we have
Q

d
X Y
p′A (x) = (x − λj (A))
ℓ=1 1≤j≤d
j6=ℓ

and therefore
Qd d 
p′A (λ1 (A)) j=2 (λ1 (A) − λj (A))

Y λj (A)
= = 1− = det(I − DzA φA ).
λ1 (A)d−1 λ1 (A)d−1 j=2
λ1 (A)

Since 1 − λj (A)/λ1 (A) is nonzero for all j = 2, . . . , d this quantity is nonzero. This
completes the proof of (4) and hence of the theorem.

4. Operator-theoretic preliminaries
In this section we collect some preliminary results which will underpin the con-
struction of the operators Ls defined in Theorem 3.
4.1. Bergman spaces. If Ω ⊂ Ck is open and nonempty the Bergman space
2
A (Ω) is´defined to be the set of all holomorphic functions f : Ω → C such that the
integral Ω |f (z)|2 dV (z) is finite, where V denotes 2k-dimensional Lebesgue measure
on Ck ≃ R2k . The space A2 (Ω) is a Hilbert space when equipped with the inner
product hf, giA2 (Ω) := Ω f (z)g(z)∗ dV (z). In particular it is a closed subspace of the
´

Hilbert space L2 (Ω) and is therefore separable. We note the following elementary
estimate:
Lemma 4.1. Let Ω ⊆ Ck be a nonempty open set and let K ⊆ Ω be compact.
Then there exists CK > 0 depending on K such that supz∈K |f (z)| ≤ CK kf kA2 (Ω) for
every f ∈ A2 (Ω).
Proof. Choose ε > 0 small enough that for every z ∈ K the open ball Bε (z0 ) is
a subset of Ω. By harmonicity we have
1
ˆ
2
|f (z0 )| = f (z)2 dV (z)
V (Bε (z0 )) Bε (z0 )
1 1 k!
ˆ
≤ |f (z)|2 dV (z) = kf k2A(Ω) = k k kf k2A(Ω)
V (Bε (z0 )) Ω V (Bε (z0 )) π ε
666 Ian D. Morris

for all f ∈ A2 (Ω) and z0 ∈ K. 

We observe in particular that for every z ∈ Ω the evaluation map f 7→ f (z) is a
continuous linear functional A2 (Ω) → C.
In practice we will be interested in the case where Ω is a bounded open subset of
an affine subspace of Cd rather than of Cd itself. Clearly the results of this section
will apply equally well in that context with k being equal to the dimension of the
affine subspace of Cd of which Ω is an open subset.
4.2. Trace-class operators. We define the singular values or approximation
numbers sn (L) of a bounded linear operator L : H → H acting on a separable complex
Hilbert space H to be the quantities
sn (L) := inf {kL − F k : F : H → H is bounded with rank at most n − 1} ,
where n ranges over the positive integers. If L is compact then the values sn (L)2
∗
P∞ of the self-adjoint operator L L (see e.g.
coincide with the sequence of eigenvalues
[27, Theorem IV.2.5]). If L satisfies n=1 sn (L) < ∞ then L is called trace-class.
Any trace-class operator is obviously the limit in the operator norm of a sequence of
finite-rank operators and in particular is compact. It follows easily from the definition
of sn that if L1 and L2 are bounded operators then sn (L1 L2 ) and sn (L2 L1 ) are both
bounded by kL1 ksn (L2 ) for every n ≥ 1, and in particular the composition of a trace-
class operator with a bounded operator is trace-class. In particular every power of a
trace-class operator is trace-class.
The fundamental properties of the trace are summarised in the following result
which combines several statements from [62, §3]:
Theorem 5. Let L be a trace-class operator acting on a complex separable
Hilbert space H and let (λn )M
n=1 be a complete enumeration of the nonzero eigenvalues
of L, listed with repetition according to algebraic multiplicity, where M ∈ N ∪
{0, +∞}. Then for every orthonormal basis (en )∞ n=1 of H we have
∞
X M
X
(8) hLen , en i = λn
n=1 n=1

with both series being absolutely convergent. The common value of (8) is defined to
be the trace of L and is denoted tr L.
It is clear from the definition that s2n−1 (L1 + L2 ) ≤ sn (L1 ) + sn (L2 ) for every
pair of bounded linear operators L1 , L2 : H → H and every n ≥ 1. It follows easily
that if L1 , . . . , Lk are trace-class operators on H then any finite linear combination
P k
i=1 ai Li is also trace-class and satisfies
k
X k
X
tr ai Li = ai tr Li
i=1 i=1

as a consequence of (8).
The following result also combines several statements from [62, §3], with the
exception of the determinant formula for an which may be found instead in, for
example, [61, Theorem 6.8] or [27, Theorem IV.5.2].
Theorem 6. Let L be a trace-class operator on a separable complex Hilbert
space H and let (λn )∞
n=1 be an enumeration of the nonzero eigenvalues of L, repeated
according to algebraic multiplicity. (If only M < ∞ nonzero eigenvalues exist, then
 Fast approximation of the affinity dimension for dominated affine iterated function systems 667

we define λn := 0 for all n > M.) For every n ≥ 1 define

X
an := (−1)n λi1 (L) · · · λin (L)
i1 <···<in

and define also a0 := 1. Then the function

∞
X
det(I − zL) := an z n
n=0

is well-defined
Q∞ and entire, and is equal to the absolutely convergent infinite prod-
uct n=1 (1 − zλn ). The zeros of z 7→ det(I − zL) are precisely the reciprocals of
the nonzero eigenvalues of L and the order of each zero is equal to the algebraic
multiplicity of the corresponding eigenvalue. The coefficients an satisfy
 
tr L n−1 0 ··· 0 0
 tr L2 tr L n −2 ··· 0 0 
..
 
(−1)n 
tr L 3
tr L 2
tr L . 0 0

an = det  . ,
 
n!  .. .. .. .. .. .. 
 . . . . . 

tr Ln−1 tr Ln−2 tr Ln−3 · · · tr L 1 
tr Ln tr Ln−1 tr Ln−2 · · · tr L2 tr L
and X
|an | ≤ si1 (L) · · · sin (L)
i1 <···<in

for all n ≥ 1.
4.3. Weighted composition operators on Bergman spaces. It has long
been known that composition operators on Bergman spaces, and on other Banach
spaces of holomorphic functions, are trace-class under mild conditions (see e.g. [28]).
Historically most results in this context have assumed the set Ω ⊂ Ck to be bounded
and connected but in this article we will need to work with sets having multiple
connected components. We will use the notation Ω0 ⋐ Ω to mean that the closed set
Ω0 is a compact subset of the open set Ω.
The following result is a special case of [1, Theorem 5.9].
Theorem 7. Let Ω ⊆ Ck be a nonempty open set and let Ω0 ⋐ Ω be nonempty.
Suppose that φ1 , . . . , φm : Ω → Ω0 are holomorphic and ψ1 , . . . , ψm : Ω → C are
holomorphic and bounded. Then the operator L : A2 (Ω) → A2 (Ω) given by
m
X
(Lf ) (z) := ψj (z)f (φj (z))
j=1

is a well-defined bounded linear operator on A2 (Ω), and there exist C, γ > 0 depend-
ing only on Ω and Ω0 such that
m
!
X  1

sn (L) ≤ C sup |ψj (z)| exp −γn k
z∈Ω
j=1

for every n ≥ 1. In particular L is trace class.

In this article we will need to calculate explicitly the traces of a family of op-
erators. The following result is a minor variation on a type of result appearing in
668 Ian D. Morris

work of Ruelle ([59, Lemma 1]), Mayer ([43, §III] and remark following [44, Corol-
lary 7.11]), Fried ([25, Lemma 5]) and other authors. The result may be proved easily
by following the second, third and fourth paragraphs of the proof of [2, Theorem 4.2].
Theorem 8. Let Ω ⊂ Ck be a bounded, connected, nonempty open set and
suppose that φ : Ω → Ω is a holomorphic function such that φ(Ω) ⋐ Ω. Let ψ : Ω →
C be holomorphic and bounded. Then φ has a unique fixed point z0 ∈ Ω, the
eigenvalues of the derivative Dz0 φ are all strictly less than 1 in modulus, and the
operator L : A2 (Ω) → A2 (Ω) defined by (Lf )(z) := ψ(z)f (φ(z)) is trace-class and
has trace equal to ψ(z0 )/ det(I − Dz0 φ).
Since we will in general need to study operators on Bergman spaces A2 (Ω) for
which Ω is not connected, we prove the following simple extension of Theorem 8
which does not seem to have been previously stated elsewhere:
Theorem 9. Let Ω ⊆ Ck be a bounded nonempty open set and suppose that
φ : Ω → Ω is a holomorphic function such that φ(Ω) ⋐ Ω. Let ψ : Ω → C be
holomorphic and bounded. Then the set of fixed points Fix φ := {z ∈ Ω : φ(z) = z}
is either finite or empty, and each connected component of Ω contains at most one
fixed point of φ. At each fixed point z ∈ Fix φ the eigenvalues of the derivative Dz φ
are all strictly less than 1 in modulus. The operator L : A2 (Ω) → A2 (Ω) defined by
(Lf )(z) := ψ(z)f (φ(z)) is trace-class and satisfies
X ψ(z)
(9) tr L = .
det(I − Dz φ)
z∈Fix φ

Additionally, if Ω is connected, then Fix φ is a singleton.

Proof. The number of connected components of Ω is at most countably infinite
since otherwise the separability of Ck would be contradicted. Let (Ωm )M m=1 be an
enumeration of the connected components of Ω where M ∈ N ∪ {∞}. For each m,
by connectedness and continuity we have either φ(Ωm ) ⋐ Ωm or φ(Ωm ) ∩ Ωm = ∅. In
the former case there is a unique fixed point of φ in Ωm and the derivative of φ at
the fixed point has all eigenvalues strictly less than 1 in modulus by Theorem 8. In
the latter case there is obviously no fixed point in Ωm . It follows in particular that if
m = 1 then Fix φ is a singleton as required. Moreover we observe that Fix φ consists
entirely of isolated points, is closed, and is compact since it is contained in φ(Ω); it
is therefore finite or empty, as required.
The operator L meets the hypotheses of Theorem 7 and hence is trace-class, so it
remains to calculate its trace. For each integer m such that 1 ≤ m ≤ M let (fm,n )∞ n=1
be an orthonormal basis for A2 (Ωm ). Extend each fm,n to a function f˜m,n : Ω → C by
defining f˜m,n (z) := fm,n (z) when z ∈ Ωm and f˜m,n (z) := 0 otherwise. Clearly (f˜m,n )
is an orthonormal basis for A2 (Ω), so by Theorem 5 we have
∞
M X ∞ ˆ
M X
hLf˜m,n , f˜m,n iA2 (Ω) = ψ(z)f˜m,n (φ(z))f˜m,n (z)∗ dV (z)
X X
(10) tr L =
m=1 n=1 m=1 n=1 Ω
M X
∞ ˆ
ψ(z)f˜m,n (φ(z))f˜m,n (z)∗ dV (z)
X
=
m=1 n=1 Ωm

using the fact that each f˜n,m is supported on Ωm , and these series are absolutely
convergent.
 Fast approximation of the affinity dimension for dominated affine iterated function systems 669

Let us evaluate the final term of (10) by considering the contribution of each
m. For m such that φ(Ωm ) ∩ Ωm = ∅ the integrand is clearly identically zero for
every n ≥ 1 and the contribution of that m to the total is zero. On the other
hand for each m such that φ(Ωm ) ⋐ Ωm let us define Lm : A2 (Ωm ) → A2 (Ωm ) by
(Lm f )(z) := ψ(z)f (φ(z)). By Theorem 8 there is a unique fixed point zm of φ in Ωm ,
the operator Lm is trace-class and
∞
ψ(zm ) X
= tr Lm = hLm fm,n , fm,n iA2 (Ωm )
det(I − Dzm φ) n=1
X∞ ˆ
= ψ(z)fm,n (φ(z))fm,n (z)∗ dV (z)
n=1 Ωm
∞ ˆ
ψ(z)f˜m,n (φ(z))f˜m,n (z)∗ dV (z).
X
=
n=1 Ωm

We have shown that for all m

∞ ˆ
ψ(z)
ψ(z)f˜m,n (φ(z))f˜m,n (z)∗ dV (z) =
X X

n=1 Ωm det(I − Dz φ)
z∈Ωm ∩Fix φ

and combining this with (10) yields (9). 

4.4. An operator Perron–Frobenius theorem. The last general functional-
analytic result which we will require is the following:
Theorem 10. (Krasnoselskiı̆) Let X be a real Banach space and C ⊆ X a subset
such that:
(i) C is closed and convex and satisfies λC = C for all real λ > 0,
(ii) C ∩ −C = {0},
(iii) The interior of C is nonempty.
Suppose that L : X → X is a compact linear operator which is strongly positive:
for every nonzero x ∈ C there exists n ≥ 1 such that Ln x ∈ Int C. Then ρ(L) is
nonzero and is a simple eigenvalue of L whose corresponding eigenspace intersects
Int C. Moreover there exist no other eigenvalues of L with modulus ρ(L).
Proof. The strong positivity of the operator permits the application of [41, The-
orem 2.5] which implies that there exists an eigenvector in the cone C with positive
real eigenvalue λ; by strong positivity this eigenvector must belong to Int C. In the
terminology of [41, §2.1.1] the strong positivity of the operator L implies that L is
u0 -positive for every u0 ∈ Int C, so [41, Theorem 2.10] may be applied to show that
the eigenvalue λ is simple and [41, Theorem 2.13] shows that it is maximal in absolute
value (hence equal to ρ(L)) and that no other eigenvalues of the same absolute value
exist. 

5. Proof of Theorem 3
We will follow [13] in analysing the singular value function
1+⌊s⌋−s s−⌊s⌋
∧⌊s⌋ ∧⌈s⌉
ϕs (Ai ) = Ai Ai
Qℓ (j) tj (1) (1)
by treating it as a product of the form j=1 kAi k where (A1 , . . . , AN ), . . . ,
(ℓ) (ℓ)
(A1 , . . . , AN ) are a priori unrelated tuples of matrices with respective dimensions
670 Ian D. Morris

∧⌊s⌋ ∧⌊s⌋
d1 , . . . , dℓ ≥ 1, essentially ignoring the fact that the two tuples (A1 , . . . , AN ) and
∧⌈s⌉ ∧⌈s⌉
(A1 , . . . , AN ) are related by the property of being exterior powers of the same
tuple. Besides the established utility of this approach in [13, 47], we suspect that
other results of a similar character such as [23, 29] could in principle be rewritten in
these terms.
Theorem 3 is a special case of the following more general statement which is also
applied in [49]:
Theorem 11. Let ℓ ≥ 1 and t = (t1 , . . . , tℓ ) ∈ Cℓ , for j = 1, . . . , ℓ let mj , dj ≥ 1,
(j) (j) (j) (j)
let (A1 , . . . , AN ) ∈ Mdj (R), and let (K1 , . . . , Kmj ) be a multicone with transverse-
defining vector wj ∈ Rdj . Suppose that not every dj is equal to 1. Then there exists
a bounded open subset Ω of the ℓj=1 (dj − 1)-dimensional affine space
P
( ℓ ℓ
)
M M
(11) z= zj ∈ Cdj : hzj , wj i = 1 for all j = 1, . . . , ℓ
j=1 j=1

such that the operator

ℓ
! N Y
ℓ
!tj ℓ
!
(j)
M X hAi zj , wj i M (j) (j)
(Lt f ) zj := (j)
f hAi zj , wj i−1 Ai zj
j=1 i=1 j=1 sign ℜ(hAi zj , wj i) j=1

is a well-defined bounded linear operator on A2 (Ω) and:

(i) There exist constants C, κ, γ > 0 such that the approximation numbers sn (Lt )
satisfy sn (Lt ) ≤ C exp(κktk − γn1/(d1 +···+dℓ −ℓ) ) for every n ≥ 1 and t =
(t1 , . . . , tℓ ) ∈ Cℓ .
(ii) For each n ≥ 1 the trace of the operator Ltn is equal to
 d −1  t
(j) j (j) j
X Y λ1 Ai
ℓ ρ Ai
(12)   
′ (j)
|i|=n j=1 p (j) λ 1 Ai
Ai

where pB (x) := det(xI − B) denotes the characteristic polynomial of B ∈

Md (R) and p′B (x0 ) its derivative evaluated at x0 .
(iii) If t ∈ Rℓ , then
  n1
m
X Y (j) tj
(13) ρ(Lt ) = lim  Ai 
n→∞
|i|=n j=1

and in particular this limit exists. Furthermore in this case ρ(Lt ) is a simple
eigenvalue of Lt , and Lt has no other eigenvalues with modulus equal to
ρ(Lt ).
(i) (i)
Proof of Theorem 11. Fix i ∈ {1, . . . , ℓ}. Since (A1 , . . . , AN ) strictly preserves
(i) (i)
the multicone (K1 , . . . , Kmi ) with transverse-defining vector wi we may choose a
(i) (i) (i)
multicone (K̂1 , . . . , K̂1 ) with the same transverse-defining vector such that K̂j \
(i) (i) (i) (i)
{0} ⊆ Int Kj for each j = 1, . . . , mi and such that Ak ( m mi
S S
j=1 Kj ) ⊆ j=1 (K̂j ∪
i

(i)
−K̂j ) for every k = 1, . . . , N. Let
( m
! m 
)

(i) (i) (i)
[ [
A(i) := A ∈ Mdi (R) : A Kj ⊆ K̂j ∪ −K̂j
j=1 j=1
 Fast approximation of the affinity dimension for dominated affine iterated function systems 671

and let A∗(i) denote the set of all nonzero elements of A(i) . By Theorem 4 A∗(i) is a
(i) (i) (i)
semigroup. Since obviously A1 , . . . , AN ∈ A∗(i) we have Ai ∈ A∗(i) for every i ∈ Σ∗N .
Theorem 4 implies that there exists a bounded open set Ωi ⊂ {zi ∈ Cdi : hzi , wi i =
(i) (i) (i)
1} such that for every i ∈ Σ∗N the map φi : Ωi → Ωi defined by φi (z) := hAi zi , wi i−1
(i) (i)
·Ai zi is well-defined. For each i ∈ Σ∗N we have ℜ(hAi z, wi i) 6= 0 for all zi ∈ Ωi by
(i)
Theorem 4(4), so for each i ∈ Σ∗N the function zi 7→ sign ℜ(hAi zi , wi i) ∈ {±1} is
well-defined and is constant on every connected component of Ωi . In particular it is
a holomorphic function on Ωi . For each i ∈ Σ∗N define
!ti !!
(i) (i)
(i) hAi zi , wi i hAi zi , wi i
ψi,t (zi ) := (i)
:= exp ti log (i)
.
sign ℜ(hAi zi , wi i) sign ℜ(hAi zi , wi i)
(i) (i)
Since hAi zi , wi i/ sign ℜ(hAi zi , wi i) has positive real part for all zi ∈ Ωi its logarithm
is a well-defined holomorphic function of zi ∈ Ωi and has imaginary part confined to
the range (− π2 , π2 ) throughout Ωi .
For all zi ∈ Ωi and i ∈ Σ∗N we have
!!
(i)
hAi zi , wi i (i)
(14) ℜ log (i)
− log Ai ≤ log C1
sign ℜ(hAi zi , wi i)
for some constant C1 > 1 using Theorem 4(4), where C1 may be chosen independent
of i ∈ {1, . . . , m} by taking the maximum of its possible values as i varies. Hence
!! !!
(i) (i)
hAi zi , wi i hAi zi , wi i
ℜ ti log (i)
= ℜ(ti )ℜ log (i)
sign ℜ(hAi zi , wi i) sign ℜ(hAi zi , wi i)
!!
(i)
hAi zi , wi i
− ℑ(ti )ℑ log (i)
sign ℜ(hAi zi , wi i)
(i) π
≤ ℜ(ti ) log Ai + |ℜ(ti )| log C1 + |ℑ(ti )|
2
∗
for all zi ∈ Ωi and i ∈ ΣN and therefore
(i)
|ti | (i) ℜ(ti )
sup ψi,t (zi ) ≤ C1 eπ/2 Ai
zi ∈Ωi

for all i ∈ Σ∗N .

Now define Ω := Ω1  × · · · × Ωℓ . We observe that each Ωi is a bounded, open
subset of the hyperplane zi ∈ Cdi : hzi , wi i = 1 which is symmetric with respect to
complex conjugation and therefore Ω also has these properties as a subset of the affine
space defined in (11). For each i ∈ Σ∗N define a holomorphic function φi : Ω → Ω by
 
(1) (ℓ)
φi (z1 , . . . , zℓ ) := φi (z1 ), . . . , φi (zℓ )
in accordance with the statement of the theorem. As a consequence of Theorem 4(4)
the set [ (1) (ℓ)
[
φi (Ω) = φi (Ω1 ) × · · · × φi (Ωℓ )
i∈Σ∗N i∈Σ∗N

is a compact subset of Ω. For each i ∈ Σ∗N define also ψi,t : Ω → C by

ℓ
(i)
Y
ψi,t (z1 , . . . , zℓ ) := ψi,t (zi )
i=1
672 Ian D. Morris

and observe that

ℓ
ℜ(ti )
ktk (i)
Y
(15) sup |ψi,t (z)| ≤ C2 Ai
z∈Ω
i=1
√
for all i ∈ Σ∗N and t ∈ Cℓ , where C2 := (C1 e π/2 ℓ
) . In particular
N N Y
ℓ
ℜ(ti )
ktk (i) ktk
X X
(16) sup |ψj,t (z)| ≤ C2 Aj ≤ NC3 ,
z∈Ω
j=1 j=1 i=1
ℓ
say, for every t ∈ C .
We may now define the operator Lt by
N
X
(Lt f ) (z) := ψj,t (z)f (φj (z))
j=1

for all f ∈ A2(Ω) and z ∈ Ω in accordance with the statement of the theorem. By
Theorem 7 it follows that each Lt is a well-defined bounded linear operator acting
on A2 (Ω) and that there exist C4 , γ > 0 such that for all t ∈ Cℓ we have
N
!
X  Pℓ 
sn (Lt ) ≤ C4 sup |ψj,t (z)| exp −γn1/( i=1 (di −1))
z∈Ω
j=1
 Pℓ 
≤ C4 N exp κktk − γn1/ i=1 (di −1))

as a consequence of (16), where κ := log C3 . We have proved (i).

It follows from (i) that Lt is a trace-class operator. For each i ∈ Σ∗N and t ∈ Cℓ
let us define an auxiliary operator Li,t by
(Li,t f ) (z) := ψi,t (z)f (φi (z)).
Theorem 7 shows in the same manner as before that each Li,t is a well-defined
trace-class operator on A2 (Ω). The reader may easily verify the equations
ψji,t (z) = ψi,t (φj (z))ψj,t (z), φji (z) = φi (φj (z))
all i, j ∈ Σ∗N and t ∈ Cℓ . It follows by a simple
and therefore Lji,t = Li,t Lj,t forP
n ℓ
inductive argument that Lt = |i|=n Li,t for every n ≥ 1 and t ∈ C , so in
particular
X X
(17) tr Lt = tr Li,t = tr Li,t
|i|=n |i|=n

for every n ≥ 1 and t ∈ Cℓ by the linearity of the trace.

Let us now compute tr Li,t for fixed i ∈ Σ∗N and t ∈ Cℓ . By Theorem 4(4)
(i) (i)
each φi has a unique fixed point zi ∈ Ωi and it follows directly that zi :=
(1) (ℓ) (i) (i) (i)
(zi , . . . , zi ) ∈ Ω is the unique fixed point of φi in Ω. Since hAi zi , wi i = λ1 (Ai )
(i) (i) (i)
for each i = 1, . . . , ℓ it follows easily that ψi,t (zi ) = ℓi=1 ψi,t (zi ) = ℓi=1 ρ(Ai )ti .
Q Q
(i) (i) (i)
By Theorem 4(4) the derivative Dz (i) φi of φi at zi satisfies
i
  
(i)
  p′ (i) λ1 Ai
(i) A
det I − Dz (i) φi = i  d −1
i (i) i
λ1 Ai
 Fast approximation of the affinity dimension for dominated affine iterated function systems 673

where both sides of this expression are interpreted as 1 if di = 1, and since clearly
(1) (ℓ)
Dzi φi = Dz (1) φi ⊕ · · · ⊕ Dz (ℓ) φi we easily obtain
i i
  
(i)
ℓ ℓ p′ (i) λ1 A
  Y
Ai i
(i)
Y
det (I − Dzi φi ) = det I − Dz (i) φi =  d −1
.
i (i) i
i=1 i=1 λ1 Ai
It follows by Theorem 9 that
 di −1  t
(i) (i) i
ℓ
Y λ1 Ai ρ Ai
(18) tr Li,t =   
(i)
i=1 p′ (i) λ1 Ai
Ai

for every t = (t1 , . . . , tℓ ) ∈ C and every i ∈ Σ∗N , and combining (17) with (18) yields
ℓ

(12) which completes the proof of (ii).

The proof of (iii) requires some preparatory steps. For the remainder of the
proof we fix t = (t1 , . . . , tℓ ) ∈ Rℓ . We begin with the existence of the limit in (13).
(i) (i) (i) (i)
By Theorem 4(4) there exists τ ∈ (0, 1] such that τ kAi k · kAj k ≤ kAi Aj k ≤
(i) (i)
kAi k · kAj k for all i = 1, . . . , ℓ and i, j ∈ Σ∗N , which clearly implies
ti ti ti
(i) (i) (i)
Aji ≥ τ |ti | Ai Aj
for all i, i and j. The inequality
  
ℓ Pℓ ℓ ℓ
ti ti ti
(i) (i) (i)
X Y XY X Y
|ti |
Ai ≥τ i=1  Ai  Ai 
|i|=n+m i=1 |i|=n i=1 |i|=m i=1

for all n, m ≥ 1 follows, so by superadditivity the limit

  n1   n1
ℓ ℓ
X Y (i) ti Pℓ ti
(i)
XY
eP(t) := lim  Ai  = sup τ i=1 |ti | Ai 
n→∞ n≥1
|i|=n i=1 |i|=n i=1

is well-defined. We obtain in particular the inequality

ℓ
ti Pℓ
(i) ktk
XY
(19) Ai ≤ τ− i=1 |ti | nP(t)
e ≤ C5 enP(t) ,
|i|=n i=1

say, for all n ≥ 1.

We next introduce a subset of Ω which will be useful in describing the behaviour
of the eigenfunctions of Lt . By Theorem 4(4) there exist for each i = 1, . . . , ℓ
a metric di on Ω which is equivalent to the standard metric and a real number
(i)
θi ∈ (0, 1) such that every φi is a θi -contraction with respect to di . Clearly if d
is the product metric derived from d1 , . . . , dℓ and θ := maxi θi then every φi is a
θ-contraction on Ω with respect to d and d is equivalent to the standard metric on
Ω. It follows
S that φ1 , . . . , φN defines an iterated function system on the compact
′
set Ω := i∈Σ∗ φi (Ω) in the sense of Hutchinson [31] and therefore there exists a
N

unique nonempty compact set Λ ⊆ Ω′ with the property Λ = N

S
S j=1 φj (Λ). Clearly
Λ = |i|=n φi (Λ) for every n ≥ 1 by a straightforward induction and it follows
T S
easily by contractivity that Λ = n=1 |i|=n φi (Ω). On the other hand φ1 , . . . , φN
clearly also defines an iterated function system on Ω′ ∩ ℓi=1 Rdi since each map
L
674 Ian D. Morris

(i) (i)
zi 7→ hAi z, wi i−1 Ai zi obviously preserves Ωi ∩ Rdi . There therefore exists a unique
nonempty compact set Λ′ ⊆ Ω′ ∩ ℓi=1 Rdi with the same property Λ′ = N ′
L S
j=1 φj (Λ ).
By uniqueness we have Λ = Λ′ and we deduce that Λ ⊆ Ω ∩ ℓi=1 Rdi .
L
We claim that Λ has the following transitivity property: for every open set U ⊂ Ω
having nonempty intersection with Λ there exists j ∈ Σ∗N such that φj (Ω) ⊆ U. To
demonstrate this choose ω ∈ U ∩ Λ arbitrarily, let ε > 0 be small enough that the
ball of centre ω and radius ε in the metric d is a subset of U, and letSn be large
enough that θn diam Ω < ε in the sense of the metric d. Since ω ∈ Λ = |i|=n φi (Λ)
there exists j ∈ Σ∗N with length n such that ω ∈ φj (Λ). Clearly ω ∈ φj (Ω) and every
other point of φj (Ω) is within distance θn diam Ω < ε of ω, so φj (Ω) is contained in
the ε-ball around ω and is therefore a subset of U as required. The claim is proved.
We make one final preliminary claim: there exists C6 > 1 such that for every
f ∈ A2 (Ω),
ktk
(20) lim sup sup e−nP(t) |(Ltn f )(z)| ≤ C6 sup |f (z)|.
n→∞ z∈Ω z∈Λ

To prove the claim let z0 ∈ Ω and n ≥ 1 be arbitrary: we have

X
e−nP(t) |(Ltn f ) (z0 )| = e−nP(t) ψi,t (z0 )f (φi (z0 ))
|i|=n
ℓ
!
ti
ktk (i)
X Y
≤ C2 e−nP(t) Ai |f (φi (z0 ))|
|i|=n i=1
ktk ktk
≤ C2 C5 sup |f (z)|
S
z∈ |i|=n φi (Ω)
T S
using (15) and (19) and the result follows easily since Λ = n=1 |i|=n φi (Ω). We
observe immediately that ρ(Lt ) ≤ eP(t) since obviously (20) prevents Lt from having
an eigenfunction which corresponds to an eigenvalue of modulus strictly greater than
eP(t) . We also observe that as a consequence of (20) an eigenfunction of Lt with
eigenvalue of modulus eP(t) cannot vanish identically on Λ.
In order to apply Theorem 10 we wish to study the action of Lt on a real Hilbert
space. Let us define
( ℓ
)
M
H := f ∈ A2 (Ω) : f (z) ∈ R for all z ∈ Ω ∩ Rdi
i=1
2
and note that H is a closed subset of A (Ω) as a consequence of Lemma 4.1. It is
clear that H is also a real Hilbert space when equipped with the norm k · kA2 (Ω) .
We observe that the complexification HC is precisely A2 (Ω). Indeed, since z ∈ Ω if
and only if z ∗ ∈ Ω by Theorem 4, for every f ∈ A2 (Ω) the holomorphic function f ∗
defined by f ∗ (z) := f (z ∗ )∗ is also an element of A2 (Ω); thus every f ∈ A2 (Ω) can be
written as f = 12 (f +f ∗ )+ 21 (f −f ∗ ) = g+ih, say, where f, g ∈ H. This decomposition
is moreover unique sinceL if g + ih is the zero function with g, h ∈ H then g and h are
identically zero on Ω ∩ ℓi=1 Rdi , hence all of their derivatives vanish there, hence
they are zero on every connected component of Ω which intersects ℓi=1 Rdi , hence
L
they are zero throughout Ω by Theorem 4(4).
We wish to apply Theorem 10 in order to study the spectrum of Lt on H and
hence on its complexification A2 (Ω). The natural mechanism for doing this is to
 Fast approximation of the affinity dimension for dominated affine iterated function systems 675

consider the cone SNof elementsL of H which are non-negative on a convenient compact
subset such as j=1 φj (Ω) ∩ ℓi=1 Rdi and show that every nonzero element of the
cone is eventually mapped to an interior point (which is precisely a function which
is positive throughout the compact subset) by some power of Lt . However, in the
full generality of Theorem 2 it is possible that φi (Ω) may be extremely small, indeed
even a singleton set. In such cases it is not necessarily the case that every nonzero
holomorphic function on Ω is eventually mapped to a function which is positive on a
prescribed set and Theorem 10 may not be directly applicable. To resolve this issue
we will pass to a suitable quotient Hilbert space.
It is clear from the definition of ψj,t and φj that (Lt f )(z) is real when f ∈ H and
z ∈ Ω∩ ℓi=1 Rdi , so Lt acts on H. Define Z := {f ∈ H : f (z) = 0 for all z ∈ Λ} and
L
note that Z is a vector subspace of H and is closed as a consequence of Lemma 4.1. We
observe that by similar reasoning Lt preserves the subspace Z. The quotient space
H/Z is a Hilbert space when equipped with norm k[f ]kH/Z := inf{kf − gkA2 (Ω) : g ∈
Z}, being isometrically isomorphic to the orthogonal complement of Z in H. It is not
difficult to see that the operator Lt induces a compact operator on the real Hilbert
space H/Z which we also denote by Lt .
We observe that for each z ∈ Λ the functional [f ] 7→ f (z) is a well-defined
continuous linear functional H/Z → R. Indeed, if [f ] ∈ H/Z and g ∈ Z then we
have f (z) = (f + g)(z) and
|f (z)| = |(f + g)(z)| ≤ CΛ kf + gkA2 (Ω)
where CΛ > 0 is the constant given by Lemma 4.1 in respect of the nonempty compact
set Λ. In particular f (z) is independent of the choice of representative f ∈ [f ] and
(21) |f (z)| ≤ CΛ inf{kf + gkA2 (Ω) : g ∈ Z} = CΛ k[f ]kH/Z
so that the functional [f ] 7→ f (z) is continuous as claimed. Now define
\
C := {[f ] ∈ H/Z : f (z) ≥ 0 for all z ∈ Λ} = {[f ] ∈ H/Z : f (z) ≥ 0} .
z∈Λ

This set is clearly well-defined, positively homogenous, convex, and closed. If [f ] ∈

C ∩ −C then f (z) = 0 for all z ∈ Λ so that f ∈ Z and therefore the only element of
C ∩ −C is [0]. Since the function [f ] 7→ inf z∈Λ f (z) is continuous as a consequence of
(21) it is not difficult to see that [f ] ∈ H/Z is an interior point of C if and only if
inf z∈Λ f (z) > 0. In particular the set C satisfies conditions (i)–(iii) of Theorem 10. We
observe also that Lt C ⊆ C since by construction each ψj,t is positive on Ω ∩ ℓi=1 Rdi
L
and in particular on Λ.
In order to be able to apply Theorem 10 we must show that for every [f ] ∈ C with
[f ] 6= [0] there exists n ≥ 1 such that inf z∈Λ (Lnt f )(z) > 0. Given [f ] ∈ C with [f ] 6= [0]
there necessarily exists z0 ∈ Λ such that f (z0 ) > 0 and hence there exists an open
set U ⊂ Ω intersecting Λ such that f (z) > 0 for all z ∈ U ∩ Λ. By the transitivity
property of Λ remarked earlier there exists a word j ∈ Σ∗N with some length n such
that φj (Ω) ⊂ U, so in particular φj (Λ) ⊂ U ∩ Λ and therefore f (φj (z)) > 0 for all
z ∈ Λ. Hence
X
(Ltn f )(z) = ψi,t (z)f (φi (z)) ≥ ψj,t (z)f (φj (z)) > 0
|i|=n

for all z ∈ Λ since each ψi,t is real and positive throughout Λ, each f ◦ φi is real and
non-negative throughout Λ, and f ◦ φj is real and positive throughout Λ. We have
obtained inf z∈Λ (Lnt f )(z) > 0 and therefore [Lnt f ] ∈ C as required.
676 Ian D. Morris

We may now apply Theorem 10 to the action of Lt on H/Z. By that theorem

the spectral radius R of Lt on H/Z is positive and there exists [ξt ] ∈ Int C such that
[Lt ξt ] = R[ξt ]. Now let z0 ∈ Λ be arbitrary. It follows from (14) that
ℓ ℓ
ti ti
−ktk (i) ktk (i)
Y Y
C1 Ai ≤ ψi,t (z0 ) ≤ C1 Ai
i=1 i=1

for every i ∈ Σ∗N , so

ℓ  
ti
−ktk (i)
X XY
(Ltn ξt ) (z0 ) = ψi,t (z0 )ξt (φi (z0 )) ≥ C1 Ai inf ξt (z)
z∈Λ
|i|=n |i|=n i=1

and in a similar fashion

ℓ  
ti
ktk (i)
XY
(Ltn ξt ) (z0 ) ≤ C1 ≤ Ai sup ξt (z) .
z∈Λ
|i|=n i=1

Since [Ltn ξ] = Rn [ξt ] and the function [f ] 7→ f (z0 ) is continuous, the left-hand
side of each of the last two displayed equations is simply Rn ξt (z0 ) > 0. Taking
the power 1/n and letting n → ∞ it follows that R = eP(t) , and we previously
observed that eP(t) ≥ ρ(Lt ). On the other hand it is clear that necessarily ρ(Lt ) ≥
limn→∞ |(Ltn ξt )(z0 )|1/n = R and we conclude that R = eP(t) = ρ(Lt ).
If an eigenvalue of Lt acting on A2 (Ω) has absolute value ρ(Lt ) then by (20)
its corresponding eigenfunction ηt cannot be identically zero on Λ. Consequently
[ηt ] 6= [0] and therefore [ηt ] is an eigenfunction of Lt on H/Z (or its complexification)
with the same eigenvalue. But by Theorem 10 this is only possible if the eigenvalue
is ρ(Lt ) itself, and we conclude that ρ(Lt ) is the only eigenvalue of Lt on A2 (Ω)
which has maximum modulus. Moreover this eigenvalue is simple: if two linearly
independent eigenfunctions ξt1 , ξt2 ∈ A2 (Ω) exist then by (20) neither function can
be identically zero on Λ; by multiplying each by a complex unit if necessary, we
may assume that each takes a nonzero real value somewhere on Λ; and replacing ξt1
and ξt2 with the functions ξt1 + (ξt1 )∗ and ξt + (ξt2 )∗ if necessary we may assume that
ξt1, ξt2 ∈ H and [ξt1 ], [ξt2 ] 6= 0. Since Lt acting on H/Z has a simple eigenvalue at eP(t)
by Theorem 10, the equivalence classes [ξt1 ] and [ξt2] must be exact, nonzero scalar
multiples of one another. This is precisely to say that some linear combination of
ξt1 and ξt2 vanishes identically on Λ but is not the zero element of H; but since that
linear combination is an eigenfunction with eigenvalue eP(t) this contradicts (20).
To complete the proof it remains only to show that ρ(Lt ) is an algebraically sim-
ple eigenvalue. Let ξt ∈ A2 (Ω) be an eigenfunction corresponding to this eigenvalue
and observe that by (20) ξt is not identically zero on Λ. If eP(t) is not algebraically
simple, there exists nonzero ηt ∈ A2 (Ω) such that Lt ηt = eP(t) (ηt + ξt ) and therefore
Ltn ηt = enP(t) (ηt + nξt ) for every n ≥ 1, but this is only compatible with (20) if ξt is
identically zero on Λ, a contradiction. The proof is complete. 

6. Proof of Theorem 2
Before starting the proof of Theorem 2 we require two preliminary lemmas, one
concerning the behaviour of the leading eigenvalue of the operator Ls of Theorems 3
and 11 and one an abstract result concerning sequences of implicit functions in two
complex variables.
 Fast approximation of the affinity dimension for dominated affine iterated function systems 677

Lemma 6.1. Let (A1 , . . . , AN ) ∈ Md (R)N be k- and (k + 1)-multipositive with

N, d ≥ 2 and 0 ≤ k < d, and for each s ∈ C let Ls : H → H be as given by
Theorem 3. Define
 
1 X k+1−s ∧(k+1)
s−k
p(s) := log ρ(Ls ) = lim log  A∧k
i Ai 
n→∞ n
|i|=n

for all s ∈ R. Then p : R → R is convex. If additionally there exists a norm |||·||| on

Rd with respect to which max1≤i≤N |||Ai ||| < 1, then there exists c > 0 such that
p(s2 ) − p(s1 )
≤ −c < 0
s2 − s1
for all pairs of distinct points s1 , s2 ∈ R.
Proof. If s1 , s2 ∈ R, λ ∈ (0, 1) and n ≥ 1, then
k+1−λs1 −(1−λ)s2 λs1 +(1−λ)s2 −k
∧(k+1)
X
A∧k
i Ai
|i|=n

X λ  1−λ
k+1−s1 s1 −k k+1−s2 s2 −k
∧(k+1) ∧(k+1)
= A∧k
i Ai A∧k
i Ai
|i|=n
 λ  1−λ
k+1−s1 s1 −k k+1−s2 s2 −k
∧(k+1) ∧(k+1)
X X
≤ A∧k
i Ai   A∧k
i Ai 
|i|=n |i|=n

using Hölder’s inequality with p = 1/λ and q = 1/(1 − λ). Taking nth roots and
letting n → ∞ it follows directly that ρ(Lλs1 +(1−λ)s2 ) ≤ ρ(Ls1 )λ ρ(Ls2 )1−λ and the
convexity of p follows by taking logarithms.
To complete the proof suppose that there exists a norm |||·||| on Rd with respect
to which max1≤i≤N |||Ai ||| < 1, and choose C > 0 such that kBk ≤ C|||B||| for all
B ∈ Md (R). Observe that in particular σk+1 (Ai ) ≤ σ1 (Ai ) = kAi k ≤ C|||Ai ||| for all
i ∈ Σ∗N . If s1 < s2 ∈ R and n ≥ 1, then
k+1−s2 s2 −k
∧(k+1)
X
A∧k
i Ai
|i|=n
X
= σ1 (Ai ) · · · σk (Ai )σk+1 (Ai )s2 −k
|i|=n
X
= σ1 (Ai ) · · · σk (Ai )σk+1 (Ai )s1 −k σk+1 (Ai )s2 −s1
|i|=n
 s2 −s1 X
≤ max σk+1 (Ai ) σ1 (Ai ) · · · σk (Ai )σk+1 (Ai )s1 −k
|i|=n
|i|=n
 s2 −s1 X
≤ max C|||Ai ||| σ1 (Ai ) · · · σk (Ai )σk+1 (Ai )s1 −k
|i|=n
|i|=n
 n(s2 −s1 ) X
k+1−s1 s1 −k
∧(k+1)
≤ C s2 −s1 max |||Ai ||| A∧k
i Ai
1≤i≤N
|i|=n
678 Ian D. Morris

so that by taking the nth root and letting n → ∞ we obtain

 s2 −s1
ρ(Ls2 ) ≤ max |||Ai ||| ρ(Ls1 )
1≤i≤N

for all such s1 and s2 . Taking logarithms and rearranging yields the claim with
c := − log max1≤i≤N |||Ai ||| > 0. 
Similarly to §4 we shall say that X1 is compactly contained in X2 if the closure
of X1 is a compact subset of the interior of X2 , and express this relation with the
notation X1 ⋐ X2 .
Lemma 6.2. Let D1 , D2 ⊂ C be open discs, let fn : D1 × D2 → C be a bounded
holomorphic function for each n ≥ 1, and let f : D1 × D2 → C be bounded and
holomorphic. Suppose that there exists a holomorphic function g : D1 → D2 such
that for all s ∈ D1 , g(s) is a simple zero of the function z 7→ f (s, z) and is the unique
zero of that function in D2 . Suppose also that
lim sup sup |fn (s, z) − f (s, z)| = 0.
n→∞ s∈D1 z∈D2

Let D1′ be any open disc which is compactly contained in D1 . Then there exist a disc
D2′ ⊆ D2 , which may be chosen concentric with D2 and with radius arbitrarily close
to that of D2 , an integer n0 ≥ 1 and holomorphic functions gn : D1′ → D2′ defined for
all n ≥ n0 such that:
(i) For all n ≥ n0 and s ∈ D1′ , gn (s) is a simple zero of z 7→ fn (s, z) and is the
unique zero of that function in D2′ .
(ii) For every integer ℓ ≥ 0 there exists Cℓ > 0 such that
sup gn(ℓ) (s) − g (ℓ) (s) ≤ Cℓ sup sup |fn (s, z) − f (s, z)|
s∈D1′ s∈D1 z∈D2

for all n ≥ n0 , where h(ℓ) denotes the ℓth derivative of the function h.
Proof. Throughout the proof let D3 be an open disc such that D1′ ⋐ D3 ⋐ D1 .
By compactness and continuity we have g(D3 ) ⋐ D2 . Let D2′ ⋐ D2 be any disc which
is concentric with D2 and has radius large enough that g(D3 ) ⋐ D2′ . By compactness
and continuity we obtain
inf inf ′ |f (s, z)| > 0
s∈D3 z∈∂D2
and hence by uniform convergence there exists n1 ≥ 1 such that for all n ≥ n1
sup sup |f (s, z) − fn (s, z)| < inf inf |f (s, z)|.
s∈D3 z∈∂D2′ s∈D3 z∈∂D2′

It follows by Rouché’s theorem that for every s ∈ D3 and n ≥ n1 there exists a

unique zero gn (s) of the function z 7→ fn (s, z) in D2′ and this zero is simple. Since
each fn is holomorphic it follows by the holomorphic implicit function theorem (see
e.g. [26, p. 34]) that each gn : D3 → D2′ is holomorphic on D3 .
We claim now that
lim sup |gn (s) − g(s)| = 0.
n→∞
s∈D3

Indeed, let ε > 0 be any number which is small enough that for every s ∈ D3 the
closed ε-ball centred at g(s) is a subset of D2′ . By compactness and the absence of
zeros of z 7→ f (s, z) in D2 \ {g(s)} we have
inf inf |f (s, z)| > 0
s∈D3 |z−g(s)|=ε
 Fast approximation of the affinity dimension for dominated affine iterated function systems 679

so that in the same manner if n is large enough

sup sup |f (s, z) − fn (s, z)| < inf inf |f (s, z)|.
s∈D3 |z−g(z)|=ε s∈D3 |z−g(s)|=ε

Applying Rouché’s theorem again it follows that if n is sufficiently large then for
all s ∈ D3 there is a unique zero of the function z 7→ fn (s, z) in the region 0 ≤
|z − g(s)| < ε. This zero belongs to D2′ and hence is necessarily equal to gn (s), and
we therefore have sups∈D3 |gn (s) − g(s)| ≤ ε. Since ε was arbitrary we conclude that

(22) lim sup |gn (s) − g(s)| = 0

n→∞
s∈D3

as claimed.
For each s ∈ D1 the value g(s) is a simple zero of the function z 7→ f (s, z), so we
have ∂f
∂z
(s, g(s)) 6= 0 for all s ∈ D1 . Define

∂f
c := inf (s, g(s)) > 0.
s∈D3 ∂z

Since g(D3 ) ⋐ D2′ ⋐ D2 we may choose τ > 0 small enough that for every z ∈ ∂D2′
the closed ball of radius 2τ centred at z is a subset of D2 which does not intersect
g(D3). Using (22) take n2 ≥ n1 large enough that
sup |gn (s) − g(s)| < τ
s∈D3

for all n ≥ n2 . Observe that if s ∈ D3 and n ≥ n2 then |gn (s) − g(s)| < τ and
|g(s) − ω| > 2τ and therefore |gn (s) − ω| > τ for all ω ∈ ∂D2′ . Using Cauchy’s
integral formula, for any two distinct points z1 , z2 ∈ D2′ we have
f (s, z1 ) − f (s, z2 ) ∂f
− (s, z2 )
z1 − z2 ∂z
1 f (s, ω) f (s, ω) f (s, ω)
ˆ
= − − dω
2πi ∂D2′ (z1 − z2 )(ω − z1 ) (z1 − z2 )(ω − z2 ) (ω − z2 )2
1 f (s, ω)((ω − z2 )2 − (ω − z1 )(ω − z2 ) − (z1 − z2 )(ω − z1 ))
ˆ
= dω
2πi ∂D2′ (z1 − z2 )(ω − z1 )(ω − z2 )2
1 f (s, ω)(z12 − 2z1 z2 + z22 )
ˆ
= dω
2πi ∂D2′ (z1 − z2 )(ω − z1 )(ω − z2 )2
1 f (s, ω)(z1 − z2 )
ˆ
= dω.
2πi ∂D2′ (ω − z1 )(ω − z2 )2

Hence if s ∈ D3 , n ≥ n2 and gn (s) 6= g(s), then since g(s), gn (s) ∈ D2′

f (s, gn (s)) − f (s, g(s)) ∂f R|gn (s) − g(s)|
− (s, g(s)) ≤ · sup sup |f (t, z)|
gn (s) − g(s) ∂z τ3 t∈D1 z∈D2

where R denotes the radius of D2′ . Now take n3 ≥ n2 large enough that additionally
! 
R c
sup |gn (s) − g(s)| 3
sup sup |f (s, z)| < .
s∈D3 τ s∈D1 z∈D2 2
680 Ian D. Morris

If n ≥ n3 , s ∈ D3 and gn (s) 6= g(s), then since fn (s, gn (s)) = 0 = f (s, g(s)) we have
f (s, gn (s)) − fn (s, gn (s)) f (s, gn (s)) − f (s, g(s))
=
gn (s) − g(s) gn (s) − g(s)
∂f f (s, gn(s)) − f (s, g(s)) ∂f c
≥ (s, g(s)) − − (s, g(s)) > .
∂z gn (s) − g(s) ∂z 2
It follows that when n ≥ n3
2
sup |gn (s) − g(s)| ≤ sup sup |fn (s, z) − f (s, z)|.
s∈D3 c s∈D3 z∈D2
To complete the proof of the lemma let δ > 0 be small enough that for every s ∈ D1′
the closed δ-ball centred at s is a subset of D3 . By the Cauchy integral formula we
have for each integer ℓ ≥ 0 and every n ≥ n3
ℓ! gn (t) − g(t)
ˆ
(ℓ) (ℓ)
sup gn (s) − g (s) ≤ sup ℓ+1
dt
s∈D1′ s∈D1′ 2πi |s−t|=δ (t − s)

≤ δ −ℓ ℓ! sup |gn (s) − g(s)|

s∈D3
2ℓ!
≤ sup sup |fn (s, z) − f (s, z)|
cδ ℓ s∈D1 z∈D2
as required. The proof is complete. 
Proof of Theorem 2. Let (A1 , . . . , AN ) ∈ Md (R)N be k- and (k +1)-multipositive
where N, d ≥ 2 and 0 ≤ k < d. For all s ∈ C let Ls : H → H be as given by
Theorem 3. Let tn (s) and an (s) be as defined in the statement of Theorem 2. We
claim that there exist K̃, γ̃, κ > 0 such that
(23) |an (s)| ≤ K̃ n enκ|s| exp (−γ̃nα )
for all n ≥ 1 and s ∈ C, where K̃, γ̃ and κ do not depend on s or n and where
d+1


k+1
−1 1
α := d+1
= 1 + d
d

.
k+1
−2 k
+ k+1 −2
By Theorem 3 there exist constants C, γ, κ > 0 such that
sn (Ls ) ≤ C exp κ|s| − γnβ


(24)
d+1


for all n ≥ 1 and s ∈ C where β := ( k+1 − 2)−1 = α − 1, and Ls is trace-class with
tr Lsn = tn (s) for all s ∈ C. By Theorem 6 we have
X
(25) |an (s)| ≤ si1 (Ls ) · · · sin (Ls )
i1 <···<in

for all n ≥ 1. In order to proceed further we require two elementary inequalities. We

first note that for every integer m ≥ 2
∞ ˆ ∞
1 ∞
ˆ
1
−γℓβ −γtβ
X
(26) e ≤ e dt = u β −1 e−γu du
ℓ=m m−1 β (m−1)β
ˆ ∞
K γ 2K − γ (m−1)β 2K − 1+β
γ β
≤ e− 2 u du ≤ e 2 ≤ e 2 m
β (m−1)β βγ βγ
 Fast approximation of the affinity dimension for dominated affine iterated function systems 681
1
where K := sup{x β −1 e−γx/2 : x ≥ 21 } > 0 depends only on β and γ, and by increasing
γ β
K if necessary we have ∞ −γℓβ
≤ 2K e− 21+β m also for m = 1. Secondly we notice
P
ℓ=m e βγ
that
m ˆ m
X
β m1+β
(27) ℓ ≥ tβ dt =
ℓ=1 0 1+β

for all integers m ≥ 1 since the series is an upper Riemann sum for the integral.
Combining (24), (25), (26) and (27) we may now obtain
n
X Y
|an (s)| ≤ C exp(κ|s| − γiβℓ )
i1 <···<in ℓ=1

n X   
= Ceκ|s| exp −γ iβ1 + · · · + iβn
i1 <···<in
∞ X ∞ ∞   
κ|s| n
X X
exp −γ iβ1 + · · · + iβn


≤ Ce ···
i1 =1 i2 =2 in =n
n X ∞

n Y
= Ceκ|s| exp −γℓβ


m=1 ℓ=m
 κ|s|
n Y n
2KCe  γ 
≤ exp − 1+β mβ
βγ m=1
2
n
2KCeκ|s|
   
γ 1+β
≤ exp − n
βγ (1 + β)21+β

which establishes the claimed inequality (23) with γ̃ := γ/(21+β (1 + β)) and K̃ :=
2KC/βγ. Pn
2 m
Now define a function dP n : C → C for each n ≥ 1 by dn (s, z) := m=0 am (s)z ,
∞ m
and define also d∞ (s, z) := m=0 am (s)z , the convergence of the series being guar-
anteed by (23). As a consequence of (23) it is clear that
∞   
X
m γ̃ α
(28) |dn (s, z) − d∞ (s, z)| = am (s)z = O exp − n
m=n+1
2

uniformly on compact subsets of C2 . It is clear by inspection that each dn is holo-

morphic, and using the convergence of dn to d∞ uniformly on compact sets together
with Cauchy’s theorem and Morera’s theorem it follows easily that d∞ : C2 → C is
holomorphic. By Theorem 6 we have d∞ (s, z) = det(I − zLs ) for every (s, z) ∈ C2 .
In particular for every s ∈ C the zeros of z 7→ d∞ (s, z) are precisely the reciprocals
of the nonzero eigenvalues of Ls , with the degree of each zero being equal to the
algebraic multiplicity of the corresponding eigenvalue.
For each s ∈ R define r∞ (s) := ρ(Ls )−1 ∈ (0, +∞). We observe that p(s) =
− log r∞ (s) is a continuous function of s by Lemma 6.1 since it is a convex function
of s ∈ R, so r∞ (s) : R → (0, +∞) is continuous. By the combination of Theorem 3
and Theorem 6, for each s ∈ R the function z 7→ d∞ (s, z) has a simple zero at r∞ (s)
and has no zeroes with equal or smaller absolute value. We claim that there exist
n0 ≥ 1, an open set U ⊂ C containing [k, k + 1], a holomorphic extension of r∞ |[k,k+1]
to U and a sequence of holomorphic functions rn : U → C defined for all n ≥ n0 such
682 Ian D. Morris

that
  
γ̃
(29) sup rn(ℓ) (s) − (ℓ)
r∞ (s) = O exp − nα
s∈U 2
for all integers ℓ ≥ 0 and such that for all n ≥ n0 and s ∈ [k, k + 1], rn (s) is the
smallest positive real number x such that dn (s, x) = 0.
To prove the claim it is clearly sufficient, by the compactness of [k, k + 1], to
show that every s0 ∈ [k, k + 1] admits an open neighbourhood U(s0 ) such that r∞
extends holomorphically from U(s0 ) ∩ [k, k + 1] to all of U(s0 ), such that there exists
a sequence of functions rn : U(s0 ) → C defined for all large enough n such that for
all s ∈ [k, k + 1] ∩ U(s0 ), rn (s) is the smallest positive real number x such that
dn (s, x) = 0, and such that
  
(ℓ) (ℓ) γ̃ α
sup rn (s) − r∞ (s) = O exp − n
s∈U (s0 ) 2
for all integers ℓ ≥ 0. The open set U can then be taken equal to the union of a
finite cover of [k, k + 1] by different sets U(s), and the characterisation of rn (s) as
the smallest positive root of dn (s, x) = 0 ensures that for each n the local functions
rn : U(s) → C extend consistently to a single well-defined function rn : U → C.
Let us therefore prove this local version of the preceding claim. Fix s0 ∈ [k, k +1].
Since z 7→ d∞ (s0 , z) has a unique zero in the closed disc with centre 0 and radius
r∞ (s0 ), and all of its zeros are isolated, we may choose an open disc D2 (s0 ) with
centre z0 ∈ R and radius R > 0 such that [0, r∞ (s0 )] ⊂ D2 (s0 ) and such that D2 (s0 )
contains no other zeros of z 7→ d∞ (s0 , z). A simple argument using compactness
shows that we may choose a small open disc D1 (s0 ) centred at s0 such that
sup sup |d∞ (s, z) − d∞ (s0 , z)| < inf |d∞ (s0 , z)|
s∈D1 (s0 ) |z−z0 |=R |z−z0 |=R

and by shrinking the neighbourhood D1 (s0 ) further if necessary we may assume using
continuity that additionally r∞ (s) ∈ D2 (s0 ) for all s ∈ D1 (s0 ) ∩ [k, k + 1].
By Rouché’s theorem, for all s ∈ D1 (s0 ) the function z 7→ d∞ (s, z) has a unique
zero in D2 (s0 ) and this zero is simple. When s ∈ D1 (s0 ) ∩ [k, k + 1] this zero must
be equal to r∞ (s) ∈ D2 (s0 ) by uniqueness. Extend r∞ : D1 (s0 ) ∩ [k, k + 1] → R to
a function D1 (s0 ) → C by defining r∞ (s) to be the unique zero of z 7→ d∞ (s, z) in
D2 (s0 ) for each s ∈ D1 (s0 ). By the holomorphic implicit function theorem and the
simplicity of the zero r∞ : D1 (s0 ) → D2 (s0 ) is holomorphic. Applying Lemma 6.2 we
find, shrinking D1 (s0 ) and D2 (s0 ) if necessary, that there exist constants Cℓ > 0, an
integer n1 ≥ 1 and holomorphic functions rn : D1 (s0 ) → D2 (s0 ) defined for all n ≥ n1
such that
sup rn(ℓ) (s) − r∞
(ℓ)
(s) ≤ Cℓ sup sup |dn (s, z) − d∞ (s, z)|
s∈D1 (s0 ) s∈D1 (s0 ) z∈D2 (s0 )
  
γ̃
= O exp − nα
2
for every integer ℓ ≥ 0, such that rn (s) is the unique zero of z 7→ dn (s, z) in D2 (s0 )
for all s ∈ D1 (s0 ) and n ≥ n1 and is a simple zero for all such s and n, such that
[0, r∞ (s0 )] ⊆ D2 (s0 ), and such that D2 (s0 ) is an open disc centred on the real axis.
For all s ∈ D1 (s0 ) ∩ [k, k + 1] and n ≥ n0 the numbers rnP (s) and rn (s)∗ both lie
in D2 (s0 ) and are both zeros of the polynomial dn (s, z) = nm=0 an (s)z m since the
coefficients of that polynomial are real and since D2 (s0 ), being a disc centred on
 Fast approximation of the affinity dimension for dominated affine iterated function systems 683

the real axis, is symmetric with respect to complex conjugation. By the uniqueness
of the zero rn (s) in D2 (s0 ) this is possible only if rn (s) = rn (s)∗ , which is to say
if rn (s) is real. Since D2 (s0 ) contains the interval from 0 to P rn (s), it follows that
if rn (s) is positive then it is the smallest positive real root of nm=0 an (s)xm for all
s ∈ D1 (s0 ) ∩ [k, k + 1]. To complete the proof of the claim it therefore suffices to show
that if n is sufficiently large then rn (s) > 0 for all s ∈ D1 (s0 ). To see this choose
δ ∈ (0, r∞ (s0 )) small enough that the open δ-ball centred at r∞ (s0 ) is contained in
D2 (s0 ), and observe that by shrinking D1 (s0 ) further if necessary we may obtain
inf inf |d∞ (s, z)| > 0
s∈D1 (s0 ) |z−r∞ (s0 )|=δ

and hence for all large enough n

sup sup |dn (s, z) − d∞ (s, z)| < inf inf |d∞ (s, z)|.
s∈D1 (s0 ) |z−r∞ (s0 )|=δ s∈D1 (s0 ) |z−r∞ (s0 )|=δ

By Rouché’s theorem this implies that there exists n0 ≥ n1 such that for all n ≥ n0
and all s ∈ D1 (s0 ) there is a unique zero of z 7→ dn (s, z) inside the circle of radius
δ and centre r∞ (s0 ), and since this region is a subset of D2 (s0 ) this root must equal
rn (s) by the uniqueness of that root in D2 (s0 ). In particular for all n ≥ n0 and
s ∈ D1 (s0 ) ∩ [k, k + 1] we have rn (s) > r∞ (s0 ) − δ > 0 and no other P root lies in
(0, rn (s)) ⊂ D2 (s0 ). Hence rn (s) is the smallest positive real root of nm=0 an (s)xm
for all s ∈ D1 (s0 ) ∩ [k, k + 1] as required to prove the local version of the claim with
U(s0 ) := D1 (s0 ). The full statement of the claim follows.
We may now complete the proof of the theorem. Define Pn (s) := rn (s)−1 > 0 for
all s ∈ [k, k + 1] and n ≥ n0 , and P (s) := r∞ (s)−1 > 0 for all s ∈ R. Observe that
by Theorem 3 we have eP (A1 ,...,AN ;s) = P (s) for all s ∈ [k, k + 1]. Since r∞ : U → C is
holomorphic, P is real-analytic at least on a neighbourhood of [k, k + 1]. Since r∞ (s)
is positive for all real s and [k, k + 1] is compact it follows that
(30) inf r∞ (s) > 0
s∈[k,k+1]

and by the case ℓ = 0 of (29) we deduce that

(31) lim inf rn (s) > 0.
n→∞ s∈[k,k+1]

Using (29), (30), (31) and the expressions

1 1
|Pn (s) − P (s)| = − ,
rn (s) r∞ (s)
r ′ (s) r ′ (s)
|Pn′ (s) − P ′(s)| = n 2 − ∞ 2
rn (s) r∞ (s)
and
rn′′ (s)rn (s) − rn′ (s)2 r∞
′′ ′
(s)r∞ (s) − r∞ (s)2
|Pn′′ (s) − P ′′ (s)| = −
rn (s)4 r∞ (s)4
it follows by elementary manipulations that
  
γ̃ α
(32) sup |Pn (s) − P (s)| = O exp − n ,
s∈[k,k+1] 2
  
′ ′ γ̃ α
(33) sup |Pn (s) − P (s)| = O exp − n
s∈[k,k+1] 2
684 Ian D. Morris

and
  
γ̃ α
(34) sup |Pn′′ (s) ′′
− P (s)| = O exp − n .
s∈[k,k+1] 2
In the case where we do not assume that max1≤i≤N |||Ai ||| < 1 for some norm on Rd
the estimate (32) already completes the proof of Theorem 2. Otherwise, we claim
that inf s∈[k,k+1] P ′′(s) > 0 and sups∈[k,k+1] P ′ (s) < 0. Let p(s) := log P (s) for s ∈ R
so that P ′ (s) = p′ (s)P (s) and P ′′ (s) = p′′ (s)P (s) + p′ (s)2 P (s). Obviously p is real-
analytic on [k, k + 1] since P is positive and real-analytic there, and p is convex by
Lemma 6.1, so necessarily p′′ (s) ≥ 0 for all s ∈ [k, k + 1]. By Lemma 6.1 we have
p′ (s) < 0 for all s ∈ [k, k + 1] and therefore
(35) sup P ′(s) = sup p′ (s)P (s) < 0.
s∈[k,k+1] s∈[k,k+1]

Similarly we observe that inf s∈[k,k+1] |p′ (s)| > 0, and since P ′′ (s) = p′′ (s)P (s) +
p′ (s)2 P (s) ≥ p′ (s)2 P (s) we likewise deduce that inf s∈[k,k+1] P ′′ (s) > 0 as claimed.
Combining the previous claim with (34) we find in particular that inf s∈[k,k+1] Pn′′ (s)
> 0 for all large enough n, which proves that each such function Pn : [k, k + 1] → R is
convex. By the hypothesis dimaff (A1 , . . . , AN ) ∈ (k, k + 1) of Theorem 2 there exists
a solution s ∈ (k, k + 1) to P (s) = 1, and since P has negative derivative on [k, k + 1]
this implies that P (k) > 1 > P (k + 1). Combining this observation with (32) we find
that Pn (k) > 1 > Pn (k + 1) for all large enough n, and by the combination of (35)
and (33) we find that sups∈[k,k+1] Pn′ (s) ≤ −c < 0 for all large enough n where c > 0
is some positive constant. It follows that for all large enough n there exists a unique
sn ∈ [k, k + 1] such that Pn (sn ) = 1. Let s∞ := dimaff (A1 , . . . , AN ) ∈ [k, k + 1] be the
unique solution to P (s∞ ) = 1. If sn 6= s∞ then by the Mean Value Theorem there
exists t strictly between sn and s∞ such that
P (sn ) − P (s∞ )
P ′ (t) =
sn − s∞
and therefore since Pn (sn ) = 1 = P (s∞ ) we obtain
|P (sn ) − P (s∞ )| |P (sn ) − Pn (sn )|
|sn − s∞ | = ′
= ≤ c−1 |P (sn ) − Pn (sn )|.
|P (t)| |P ′(t)|
The inequality |sn − s∞ | ≤ c−1 |P (sn ) − Pn (sn )| obviously also holds when sn = s∞ ,
so   
γ̃ α
|sn − s∞ | = O exp − n
2
as n → ∞ using (32). The proof of the theorem is complete. 

7. Examples
7.1. Methodology. There are two intuitively natural mechanisms by which to
make the approximations given in Theorem 2 yield an approximation to the affinity
dimension. On the one hand since eP (A1 ,...,An ;s) is decreasing in s and since the affinity
dimension is the unique s ∈ [k, k + 1] such that 1 is the leading eigenvalue of Ls , the
affinity dimension corresponds to the smallest s ∈ [k, k +1] P∞such that det(I −Ls ) = 0,
which is to say the smallest s ∈ [k, k + 1] such that m=0 am (s) = 0. One might
therefore attempt to approximate
Pn the affinity dimension by looking for the smallest
solution s to the equation m=0 am (s) = 0 for each fixed n. In practice this is
impractical since Ls may in general have infinitely many positive real eigenvalues
 Fast approximation of the affinity dimension for dominated affine iterated function systems 685

and the number of solutions to nm=0 am (s) = 0 may therefore be extremely large
P
and the function itself highly oscillatory.

n Approximation to affinity dimension CPU time

2 1.14341 79598 76019 95000 60486 91827 85789 60135 0.043s
3 1.11827 23247 08006 28499 89060 66409 13091 47143 0.044s
4 1.11538 89736 67461 99644 51849 00512 18003 54788 0.053s
5 1.11560 42107 66261 56209 11669 09958 04069 77087 0.075s
6 1.11560 31850 39305 08475 98379 83168 80085 68510 0.11s
7 1.11560 32522 24751 03699 38823 87724 66623 37012 0.16s
8 1.11560 32579 27402 64806 11546 27227 11083 45893 0.30s
9 1.11560 32577 86505 71154 77556 50836 85812 53178 0.39s
10 1.11560 32577 87028 88533 65835 00045 83936 61000 0.67s
11 1.11560 32577 87030 91898 36777 33249 49956 17495 1.2s
12 1.11560 32577 87030 89197 97928 71446 51257 73313 2.0s
13 1.11560 32577 87030 89218 88050 96492 48585 23429 4.3s
14 1.11560 32577 87030 89218 84942 17623 75680 33697 8.8s
15 1.11560 32577 87030 89218 84937 14660 75123 27001 20s
16 1.11560 32577 87030 89218 84937 14840 85419 85122 44s
17 1.11560 32577 87030 89218 84937 14840 24544 08248 100s
18 1.11560 32577 87030 89218 84937 14840 24574 24137 210s
19 1.11560 32577 87030 89218 84937 14840 24574 25551 440s
20 1.11560 32577 87030 89218 84937 14840 24574 25551 990s
Table 1. Approximations to the affinity dimension of Example 1 calculated using Theorem 2
and the secant method as described in §7.1, implemented in Wolfram Mathematica. The CPU
time used in each approximation is as reported by Mathematica’s Timing function. For n = 1 the
approximation to the pressure function has no root in (1, 2) and this line is therefore omitted from
the table. Digits which are empirically observed to have converged to a stable value are underlined.
In practice we therefore adopt the following
Pn alternative approach. For large n the
smallest positive real root x = rn (s) of m=0 am (s)xm approximates the reciprocal
of the leading eigenvalue of Ls . Moreover, for large n the function s 7→ rn (s)−1 is
convex and strictly decreasing with a unique root in [k, k + 1] by virtue of Theo-
rem 2. Computing the unique root of a convex decreasing function is a far more
tractable enterprise than finding the smallest root of an oscillating function, and for
this reason our application of Theorem 2 follows the approach of solving rn (s) = 1.
For this problem we use the secant method. Since rn−1 is convex and decreasing the
convergence of the sequence of approximations√ generated by the secant method is
(1+ 5)/2
guaranteed with super-exponential rate O(θm ) for some θ ∈ (0, 1). In practical
instances we found that the sequence (sm ) consistently converged empirically to 40
decimal places by around m ≃ 12 independently of n. The results of this proce-
dure applied to some examples of two- and three-dimensional affine iterated function
systems are presented in this section.
For large n one may show that the trace tn (s) appearing in Theorem 2 approx-
imates the value enP (A1 ,...,AN ;s) whereas the coefficients an (s) are shown in Theorem
2 to decrease to zero with super-exponential speed. The small size of an (s) is thus
attributable to additive cancellation between potentially very large summands. It is
therefore likely to be necessary in implementation to record the traces tn (s) to sig-
nificantly more decimal places than are desired for the ultimate approximation. In
686 Ian D. Morris

the computations which follow the traces tn (s) were calculated in arbitrary precision,
reducing to finite precision only for the outcome of the calculation of the coefficients
an (s).
7.2. Example 1: a pair of dominated matrices. Define
 4 5  1 
−7 7 0
A1 := , A2 := 7 5 .
0 17 − 7 − 47
We claim that the pair (A1 , A2 ) is 1-dominated. Indeed, define
  
x 2
C1 := ∈ R : |x| ≥ 2|y| ,
y
  
x 2
C2 := ∈ R : |y| ≥ 2|x| .
y
If (x, y)⊤ ∈ C1 , then
5 4 4 5 3 2
y − x ≥ |x| − |y| ≥ |y| ≥ y
7 7 7 7 7 7
and equality of the first and last terms is only possible if y = 0 and consequently
x = 0. In particular if (x, y)⊤ ∈ C1 is nonzero we obtain A1 (x, y)⊤ ∈ Int C1 . Moreover
for (x, y)⊤ ∈ C1 we also have
4 5 5 4 3 2
y + x ≥ |x| − |y| ≥ |x| ≥ x
7 7 7 7 7 7
which yields A2 (x, y)⊤ ∈ Int C2 when (x, y)⊤ is nonzero. In a similar manner, if
(x, y)⊤ ∈ C2 then
5 4 5 4 3 2
y − x ≥ |y| − |x| ≥ |y| ≥ y
7 7 7 7 7 7
and
4 5 4 5 3 2
y + x ≥ |y| − |x| ≥ |x| ≥ x
7 7 7 7 7 7
which respectively give A1 (x, y)⊤ ∈ Int C1 and A2 (x, y)⊤ ∈ Int C2 when (x, y)⊤ is
nonzero.
If we now let w = (1, 1)⊤ then hu, wi is never zero for any nonzero u ∈ C1 ∪ C2 ,
so defining
Ki := {u ∈ Ci : hu, wi > 0}
for i = 1, 2, it is not difficult to see that (K1 , K2 ) is a multicone for (A1 , A2 ). In
particular Theorem 2 may be applied to estimate the affinity dimension of the pair
(A1 , A2 ). Let (B1 , B2 ) := (A1 , −A2 ). Since
  n1
X
eP (A1 ,A2 ;1) = eP (B1 ,B2 ;1) = lim  kBi k
n→∞
|i|=n
1
n √
X
n
1 50
≥ lim Bi = lim k(B1 + B2 ) k = ρ(B1 + B2 ) =
n >1
n→∞ n→∞ 7
|i|=n

and
8
eP (A1 ,A2 ;2) = | det A1 | + | det A2 | = <1
49
 Fast approximation of the affinity dimension for dominated affine iterated function systems 687

we infer that dimaff (A1 , A2 ) ∈ (1, 2). The first 20 approximations to the affinity
dimension of (A1 , A2 ) are tabulated in Table 1.
7.3. Example 2: a three-dimensional iterated function system.

n Approximation to affinity dimension CPU time

3 1.74010 38961 34544 64381 66016 57752 82592 79145 0.067s
4 1.53612 13489 34570 18769 13237 56458 61628 45041 0.10s
5 1.58779 31446 44939 17928 98900 28708 16065 92496 0.15s
6 1.58459 23810 06597 43285 21249 54866 32813 68839 0.22s
7 1.58477 97771 44149 34557 48903 92413 22985 52229 0.33s
8 1.58477 17757 07488 53767 71488 42424 52891 52003 0.63s
9 1.58477 20386 65944 76377 72361 85895 44529 09738 0.80s
10 1.58477 20318 53062 52952 58955 36166 25319 46959 1.4s
11 1.58477 20319 95110 47059 43620 26740 31575 13317 2.4s
12 1.58477 20319 92686 60697 00747 19778 01115 41015 5.4s
13 1.58477 20319 92720 93370 05697 62846 36869 58071 12s
14 1.58477 20319 92720 52545 02878 00445 78535 74528 27s
15 1.58477 20319 92720 52956 88351 89418 63989 50927 59s
16 1.58477 20319 92720 52953 32862 81715 84179 24019 130s
17 1.58477 20319 92720 52953 35507 79078 84111 41677 270s
18 1.58477 20319 92720 52953 35490 71502 87276 30757 560s
19 1.58477 20319 92720 52953 35490 81124 12318 84553 1200s
20 1.58477 20319 92720 52953 35490 81076 56294 07542 2800s
21 1.58477 20319 92720 52953 35490 81076 77018 06325 5900s

Table 2. Approximations to the affinity dimension of Example 3 calculated using Theorem 2

and the secant method as described in §7.1, implemented in Wolfram Mathematica. The CPU time
used in each approximation is as reported by Mathematica’s Timing function. Digits which are
empirically observed to have converged to a stable value are underlined. Convergence is noticeably
slower than for two-dimensional examples: in this context our bound for the error in the nth
approximation is O(exp(−γn5/4 )) as opposed to O(exp(−γn2 )) in the other examples. For n = 1, 2
the approximation to the pressure function has no root in (1, 2) and these lines are therefore omitted.
Consider (A1 , A2 ) where
   
5 4 1 5 5 0
1  1 
A1 := 5 5 4 , A2 := 4 5 1 = A⊤
1
12 0 1 5 12 1 4 5

and note that A1 and A2 are contractions in the Euclidean norm. It is easily checked
that (A1 A1 , A1 A2 , A2 A1 , A2 A2 ) is a tuple of positive invertible matrices and is there-
fore 1-dominated. By the characterisation of domination in terms of singular values
this clearly implies that 1-domination holds also for (A1 , A2 ).
We identify each Ai with the corresponding linear map R3 → R3 defined by
Ai with respect to the standard basis e1 , e2 , e3 of R3 . With respect to the basis
e1 ∧ e2 , e1 ∧ e3 , e2 ∧ e3 for ∧2 R3 we have
   
5 15 11 5 5 5
1  1 
A∧21 = 5 25 19 , A∧2 2 = 15 25 25 .
144 5 25 21 144 11 19 21
688 Ian D. Morris

Since (A∧2 ∧2
1 , A2 ) is thus representable by a pair of positive matrices we see that
(A1 , A2 ) is both 1-and 2-dominated. Using non-negativity it follows by a theorem of
Protasov [58] that
  n1
X
lim  kAi k = ρ(A1 + A2 ) > 1
n→∞
|i|=n

and   n1
X
A∧2  = ρ A∧2 ∧2


lim  i 1 + A2 < 1.
n→∞
|i|=n

Thus P (A1 , A2 ; 1) > 0 > P (A1 , A2 ; 2) and consequently dimaff (A1 , A2 ) ∈ (1, 2), and
we conclude that Theorem 2 is applicable to the computation of dimaff (A1 , A2 ). The
first 21 approximations to dimaff (A1 , A2 ) are presented in Table 2. An illustration of
the attractor of the iterated function system
      
x 5 4 1 x 1
1 
T1 y :=
  5 5 4   y + 0
 
z 12 0 1 5 z 0
      
x 5 5 0 x 0
1 
T2 y :=
  4 5 1   y + 0
 
z 12 1 4 5 z 1
is given in Figure 1.

Figure 1. A projection of the attractor of the iterated function system defined by Example 3.
Approximations to the affinity dimension computed using Theorem 2 are listed in Table 2. It is
known from work of Falconer [17, §5] that the upper box dimension dimB X is bounded above by
dimaff (A1 , A2 ), but unlike the case of planar affine iterated function systems current techniques are
not powerful enough to determine whether or not dimH X = dimaff (A1 , A2 ).
 Fast approximation of the affinity dimension for dominated affine iterated function systems 689

Figure 2. This self-affine set was shown in [50, §6.6] to have Hausdorff dimension equal to the
affinity dimension of the defining iterated function system. However, the linear parts of the defining
affine transformations have non-real eigenvalues and Theorem 2 is not applicable. Non-rigorous
estimates using the discretisation method described in §8 as tabulated in Table 4 suggest that the
affinity dimension is equal to approximately 1.522688.

8. Non-dominated matrices
If (A1 , . . . , AN ) ∈ M2 (R)N is a tuple of invertible matrices which is not 1-
dominated then by a line of reasoning due to Avila [63] there exist tuples (A′1 , . . . , A′N )
arbitrarily close to (A1 , . . . , AN ) with the property that some product A′i1 · · · A′in has
complex eigenvalues. For such matrices the formula for tn (s) in Theorem 2 has no
clear meaning, and also for such matrices no open subset of RP1 may be found which
is mapped strictly inside itself by the action of the matrices A′i , preventing the con-
struction of a trace-class transfer operator in direct mimicry of Theorem 2. For such
matrices it is therefore difficult to see how any reasonable adaptation of Theorem
2 could be made. In this sense we believe that 1-domination, or multipositivity, is
the weakest open condition on the matrices A1 , . . . , AN which permits a version of
Theorem 2 to be proved.
However, for non-dominated matrices it is still possible to obtain non-rigorous
estimates of the affinity dimension by other techniques. Given A1 , . . . , AN ∈ GL2 (R)
and s ∈ [0, 1] we may define an operator Ls : C α (RP1 ) → C α (RP1 ) by
N  s
X kAi uk

(Ls f ) (u) := f Ai u ,
i=1
kuk
and for s ∈ [1, 2] by
N  2−s
X kAi uk
| det Ai |s−1 f Ai u ,


(Ls f ) (u) :=
i=1
kuk
690 Ian D. Morris

in such a manner that

N
! n1
X
ρ(Ls ) = lim ϕs (Ai1 · · · Ain )
n→∞
i1 ,...,in =1

and such that ρ(Ls ) is a simple eigenvalue of Ls , as long as α ∈ (0, 1) is chosen

suitably small (in a manner which in general will depend on s) and mild algebraic
non-degeneracy conditions on (A1 , . . . , AN ) are met. (These spectral properties are
guaranteed by, for example, [29, Théorème 8.8].) We could then hope to estimate
the spectral radius of Ls for different values of s by discretising the phase space RP1 ,
constructing a large matrix representing a discretised action of Ls , and working on
the supposition that the spectral radius of the matrix is a good approximation to
ρ(Ls ) and hence to eP (A1 ,...,AN ;s) . In practical experiments we were able to obtain
around five decimal places of accuracy for the affinity dimension by discretising RP1
into approximately 104 evenly-spaced mesh points: see Tables 3 and 4. We observe in
particular that the results obtained in Table 3 show good agreement with Theorem
2 when tested on the multipositive matrix set described in Example 2. However, we
have not been able to make this method of estimation rigorous. This approach could
also be applied to higher-dimensional affine iterated function systems but we have
not investigated the matter of finding suitable discretisations of the more complicated
phase spaces required in this context.

Approximation to
Mesh size CPU time
affinity dimension
2 1.02591849 0.010s
22 1.07532743 0.0065s
23 1.11171266 0.018s
24 1.11715797 0.036s
25 1.11608327 0.053s
26 1.11557816 0.80s
27 1.11537306 0.46s
28 1.11561123 0.35s
29 1.11559940 0.65s
210 1.11561053 1.8s
211 1.11558601 2.7s
212 1.11560216 4.8s
213 1.11560441 24s
214 1.11560185 21s
215 1.11560275 67s
216 1.11560321 270s
217 1.11560315 4100s

Table 3. Estimates of the affinity dimension of Example 1 calculated using the non-rigorous
discretisation method described in §8. Even at small mesh sizes the first few decimal places show
good agreement with Table 1 but convergence in subsequent decimal places is markedly slower.
Digits which are empirically observed to have converged to a stable value are underlined.
 Fast approximation of the affinity dimension for dominated affine iterated function systems 691

Approximation to
Mesh size CPU time
affinity dimension
2 1.50000000 0.0028s
22 1.51578683 0.0025s
23 1.51254065 0.0047s
24 1.52070716 0.033s
25 1.52415711 0.059s
26 1.52305542 0.079s
27 1.52290806 0.13s
28 1.52262668 0.26s
29 1.52269395 0.61s
210 1.52270408 1.1s
211 1.52269152 2.2s
212 1.52268717 4.5s
213 1.52268810 7.7s
214 1.52268795 18s
215 1.52268780 55s
216 1.52268780 220s
217 1.52268782 1400s

Table 4. Estimates of the affinity dimension of the iterated function system defined in [50,
§6.6] and illustrated in Figure 2, calculated using the non-rigorous discretisation method described
in §8. Digits which are empirically observed to have converged to a stable value are underlined. No
rigorous estimate of the affinity dimension of this IFS is currently available.
Acknowledgements. This research was supported by the Leverhulme Trust (Re-
search Project Grant number RPG-2016-194). The author thanks O. Bandtlow for
helpful comments and suggestions. The author additionally thanks an anonymous
reviewer for suggesting several economies of argument.

References
[1] Bandtlow, O. F., and O. Jenkinson: Explicit eigenvalue estimates for transfer operators
acting on spaces of holomorphic functions. - Adv. Math. 218:3, 2008, 902–925.
[2] Bandtlow, O. F., and O. Jenkinson: On the Ruelle eigenvalue sequence. - Ergodic Theory
Dynam. Systems 28:6, 2008, 1701–1711.
[3] Bandtlow, O. F., O. Jenkinson, and M. Pollicott: Periodic points, escape rates and
escape measures. - In: Ergodic theory, open dynamics, and coherent structures, Springer Proc.
Math. Stat. 70, Springer, New York, 2014, 41–58.
[4] Barański, K.: Hausdorff dimension of self-affine limit sets with an invariant direction. -
Discrete Contin. Dyn. Syst. 21:4, 2008, 1015–1023.
[5] Bárány, B., M. Hochman, and A. Rapaport: Hausdorff dimension of planar self-affine sets
and measures. - Invent. Math. 216:3, 2019, 601–659.
[6] Bárány, B., and A. Käenmäki: Ledrappier–Young formula and exact dimensionality of
self-affine measures. - Adv. Math. 318, 2017, 88–129.
[7] Bárány, B., and A. Käenmäki, and H. Koivusalo: Dimension of self-affine sets for fixed
translation vectors. - J. Lond. Math. Soc. (2) 98:1, 2018, 223–252.
[8] Bárány, B., and M. Rams: Dimension maximizing measures for self-affine systems. - Trans.
Amer. Math. Soc. 370:1, 2018, 553–576.
692 Ian D. Morris

[9] Barnsley, M. F., and A. Vince: Real projective iterated function systems. - J. Geom. Anal.
22:4, 2012, 1137–1172.
[10] Bedford, T.: Crinkly curves, Markov partitions and box dimensions in self-similar sets. -
Ph.D. Thesis, The University of Warwick, 1984.
[11] Berman, A., and R. J. Plemmons: Nonnegative matrices in the mathematical sciences. -
Classics in Applied Mathematics 9, revised reprint of the 1979 original, Society for Industrial
and Applied Mathematics (SIAM), Philadelphia, PA, 1994.
[12] Bochi, J., and N. Gourmelon: Some characterizations of domination. - Math. Z. 263:1,
2009, 221–231.
[13] Bochi, J., and I. D. Morris: Equilibrium states of generalised singular value potentials and
applications to affine iterated function systems. - Geom. Funct. Anal. 28:4, 2018, 995–1028.
[14] Das, T., and D. Simmons: The Hausdorff and dynamical dimensions of self-affine sponges: a
dimension gap result. - Invent. Math. 210:1, 2017, 85–134.
[15] Dubois, L.: Projective metrics and contraction principles for complex cones. - J. Lond. Math.
Soc. (2) 79:3, 2009, 719–737.
[16] Edgar, G. A.: Fractal dimension of self-affine sets: some examples. - Rend. Circ. Mat. Palermo
(2) Suppl. 28, 1992, 341–358.
[17] Falconer, K. J.: The Hausdorff dimension of self-affine fractals. - Math. Proc. Cambridge
Philos. Soc. 103:2, 1988, 339–350.
[18] Falconer, K. J.: The dimension of self-affine fractals. II. - Math. Proc. Cambridge Philos.
Soc. 111:1, 1992, 169–179.
[19] Falconer, K.: Fractal geometry. Third edition. - Mathematical foundations and applications,
John Wiley & Sons, Ltd., Chichester, 2014.
[20] Falconer, K., and T. Kempton: The dimension of projections of self-affine sets and mea-
sures. - Ann. Acad. Sci. Fenn. Math. 42:1, 2017, 473–486.
[21] Falconer, K., and T. Kempton: Planar self-affine sets with equal Hausdorff, box and affinity
dimensions. - Ergodic Theory Dynam. Systems 38:4, 2018, 1369–1388.
[22] Falconer, K., and J. Miao: Dimensions of self-affine fractals and multifractals generated by
upper-triangular matrices. - Fractals 15:3, 2007, 289–299.
[23] Feng, D.-J., and P. Shmerkin: Non-conformal repellers and the continuity of pressure for
matrix cocycles. - Geom. Funct. Anal. 24:4, 2014, 1101–1128.
[24] Fraser, J. M.: On the packing dimension of box-like self-affine sets in the plane. - Nonlinearity
25:7, 2012, 2075–2092.
[25] Fried, D.: The zeta functions of Ruelle and Selberg. I. - Ann. Sci. École Norm. Sup. (4) 19:4,
1986, 491–517.
[26] Fritzsche, K., and H. Grauert: From holomorphic functions to complex manifolds. - Grad.
Texts in Math. 213, Springer-Verlag, New York, 2002.
[27] Gohberg, I., S. Goldberg, and N. Krupnik: Traces and determinants of linear operators.
- Operator Theory: Advances and Applications 116, Birkhäuser Verlag, Basel, 2000.
[28] Grothendieck, A.: Produits tensoriels topologiques et espaces nucléaires. - Mem. Amer.
Math. Soc. 16, 1955.
[29] Guivarc’h, Y., and É. Le Page: Simplicité de spectres de Lyapounov et propriété d’isolation
spectrale pour une famille d’opérateurs de transfert sur l’espace projectif. - In: Random walks
and geometry, Walter de Gruyter, Berlin, 2004, 181–259.
[30] Hueter, I., and S. P. Lalley: Falconer’s formula for the Hausdorff dimension of a self-affine
set in R2 . - Ergodic Theory Dynam. Systems 15:1, 1995, 77–97.
[31] Hutchinson, J. E.: Fractals and self-similarity. - Indiana Univ. Math. J. 30:5, 1981, 713–747.
 Fast approximation of the affinity dimension for dominated affine iterated function systems 693

[32] Jenkinson, O., and M. Pollicott: Computing the dimension of dynamically defined sets:
E2 and bounded continued fractions. - Ergodic Theory Dynam. Systems 21:5, 2001, 1429–1445.
[33] Jenkinson, O., and M. Pollicott: Calculating Hausdorff dimensions of Julia sets and
Kleinian limit sets. - Amer. J. Math. 124:3, 2002, 495–545.
[34] Jenkinson, O., and M. Pollicott: Orthonormal expansions of invariant densities for ex-
panding maps. - Adv. Math. 192:1, 2005, 1–34.
[35] Jenkinson, O., and M. Pollicott: A dynamical approach to accelerating numerical inte-
gration with equidistributed points. - Tr. Mat. Inst. Steklova 256, 2007, 290–304.
[36] Jenkinson, O., and M. Pollicott: Rigorous effective bounds on the Hausdorff dimension
of continued fraction Cantor sets: a hundred decimal digits for the dimension of E2 . - Adv.
Math. 325, 2018, 87–115.
[37] Jenkinson, O., M. Pollicott, and P. Vytnova: Rigorous computation of diffusion coeffi-
cients for expanding maps. - J. Stat. Phys. 170:2, 2018, 221–253.
[38] Käenmäki, A., and I. D. Morris: Structure of equilibrium states on self-affine sets and strict
monotonicity of affinity dimension. - Proc. Lond. Math. Soc. (3) 116:4, 2018, 926–956.
[39] Käenmäki, A., and P. Shmerkin: Overlapping self-affine sets of Kakeya type. - Ergodic
Theory Dynam. Systems 29:3, 2009, 941–965.
[40] Kagiso, D., and M. Pollicott: Computing multifractal spectra. - Dyn. Syst. 30:4, 2015,
404–425.
[41] Krasnosel’skiı̆, M. A.: Positive solutions of operator equations. - P. Noordhoff Ltd. Gronin-
gen, 1964.
[42] Liverani, C.: Decay of correlations. - Ann. of Math. (2) 142:2, 1995, 239–301.
[43] Mayer, D. H.: On composition operators on Banach spaces of holomorphic functions. - J.
Funct. Anal. 35:2, 1980, 191–206.
[44] Mayer, D. H.: Continued fractions and related transformations. - In: Ergodic theory, symbolic
dynamics, and hyperbolic spaces (Trieste, 1989), Oxford Sci. Publ., Oxford Univ. Press, New
York, 1991, 175–222.
[45] McMullen, C.: The Hausdorff dimension of general Sierpiński carpets. - Nagoya Math. J.
96, 1984, 1–9.
[46] Morris, I. D.: An inequality for the matrix pressure function and applications. - Adv. Math.
302, 2016, 280–308.
[47] Morris, I. D.: Some observations on Käenmäki measures. - Ann. Acad. Sci. Fenn. Math. 43:2,
2018, 945–960.
[48] Morris, I. D.: An explicit formula for the pressure of box-like affine iterated function systems.
- J. Fractal Geom. 6:2, 2019, 127–141.
[49] Morris, I. D.: Fast approximation of the p-radius, matrix pressure or generalised Lyapunov
exponent for positive and dominated matrices. - SIAM J. Matrix Anal. Appl. 43:1, 2022, 178–
198.
[50] Morris, I. D., and P. Shmerkin: On equality of Hausdorff and affinity dimensions, via
self-affine measures on positive subsystems. - Trans. Amer. Math. Soc. 371:3, 2019, 1547–1582.
[51] Pollicott, M.: Maximal Lyapunov exponents for random matrix products. - Invent. Math.
181:1, 2010, 209–226.
[52] Pollicott, M.: Computing entropy rates for hidden Markov processes. - In: Entropy of
hidden Markov processes and connections to dynamical systems, London Math. Soc. Lecture
Note Ser. 385, Cambridge Univ. Press, Cambridge, 2011, 223–245.
[53] Pollicott, M., and P. Felton: Estimating Mahler measures using periodic points for the
doubling map. - Indag. Math. (N.S.) 25:4, 2014, 619–631.
694 Ian D. Morris

[54] Pollicott, M., and O. Jenkinson: Computing invariant densities and metric entropy. -
Comm. Math. Phys. 211:3, 2000, 687–703.
[55] Pollicott, M., and P. Vytnova: Estimating singularity dimension. - Math. Proc. Cam-
bridge Philos. Soc. 158:2, 2015, 223–238.
[56] Pollicott, M., and P. Vytnova: Linear response and periodic points. - Nonlinearity 29:10,
2016, 3047–3066.
[57] Pollicott, M., and H. Weiss: How smooth is your wavelet? Wavelet regularity via ther-
modynamic formalism. - Comm. Math. Phys. 281:1, 2008, 1–21.
[58] Protasov, V. Yu.: When do several linear operators share an invariant cone? - Linear Algebra
Appl. 433:4, 2010, 781–789.
[59] Ruelle, D.: Zeta-functions for expanding maps and Anosov flows. - Invent. Math. 34:3, 1976,
231–242.
[60] Rugh, H. H.: Cones and gauges in complex spaces: spectral gaps and complex Perron–
Frobenius theory. - Ann. of Math. (2) 171:3, 2010, 1707–1752.
[61] Simon, B.: Notes on infinite determinants of Hilbert space operators. - Adv. Math. 24:3, 1977,
244–273.
[62] Simon, B.: Trace ideals and their applications. - London Math. Soc. Lecture Note Ser. 35,
Cambridge Univ. Press, Cambridge-New York, 1979.
[63] Yoccoz, J.-C.: Some questions and remarks about SL(2, R) cocycles. - In: Modern dynamical
systems and applications, Cambridge Univ. Press, Cambridge, 2004, 447–458.

Received 19 August 2020 • Accepted 12 September 2021 • Published online 13 April 2022

Ian D. Morris
Queen Mary University of London
School of Mathematical Sciences
Mile End Road, London E1 4NS, United Kingdom
[email protected]