arXiv:1510.00209v1 [math.

FA] 1 Oct 2015

GENERIC PROPERTIES OF THE LOWER SPECTRAL RADIUS

FOR SOME LOW-RANK PAIRS OF MATRICES
IAN D. MORRIS
Abstract. The lower spectral radius of a set of d d matrices is defined to
be the minimum possible exponential growth rate of long products of matrices
drawn from that set. When considered as a function of a finite set of matrices
of fixed cardinality it is known that the lower spectral radius can vary discontinuously as a function of the matrix entries. In a previous article the author
and J. Bochi conjectured that when considered as a function on the set of all
pairs of 2 2 real matrices, the lower spectral radius is discontinuous on a set
of positive (eight-dimensional) Lebesgue measure, and related this result to an
earlier conjecture of Bochi and Fayad. In this article we investigate the continuity of the lower spectral radius in a simplified context in which one of the
two matrices is assumed to be of rank one. We show in particular that the set
of discontinuities of the lower spectral radius on the set of pairs of 2 2 real
matrices has positive seven-dimensional Lebesgue measure, and that among
the pairs of matrices studied, the finiteness property for the lower spectral
radius is true on a set of full Lebesgue measure but false on a residual set.

Key words and phrases: joint spectral radius; lower spectral radius; generalised spectral radius. MSC2010: 15A60 (primary), 15A18, 47D03 (secondary)

1. Introduction
The joint spectral radius (A) of a set A of d d matrices, introduced by G.C. Rota and G. Strang in 1960 [25, 26], is defined to be the maximum possible
exponential growth rate of a product of matrices drawn from the set A:
n
o
1
1
(A) := lim sup kAn A1 k n : Ai A =
sup lim inf kAn A1 k n .
n

N n
(Ai )
i=1 A

In the last two decades the joint spectral radius has been the subject of a proliferation of research articles which we will not attempt to list here exhaustively; some
prominent recent contributions include [4, 5, 9, 14, 15, 16, 18, 20, 22]. By contrast
the lower spectral radius, defined by the expression
o
n
1
1
inf
lim sup kAn A1 k n
(A) := lim inf kAn A1 k n : Ai A =

(Ai )i=1 AN n

has been studied relatively little, despite its relevance to a range of pure and applied
mathematical problems (see e.g. [3, 11, 12, 16, 23, 24, 27]).
By comparison with the joint spectral radius, the behaviour of the lower spectral
radius is delicate. The joint spectral radius of a finite set of matrices depends
continuously on the matrix entries [19] and there exists an algorithm to determine
whether or not (A) = 0 using a number of steps which grows polynomially with
the dimension of the matrices and with the cardinality of the set of matrices being
1

IAN D. MORRIS

considered ([17], see also [20, 2.3.1]). Conversely the lower spectral radius can
exhibit discontinuities (see [7, 20]) and the problem of determining whether or
not a pair of 15 15 integer matrices has lower spectral radius equal to zero is
computationally undecidable [10].
A set of matrices A is said to satisfy the finiteness property for the joint spectral
radius a situation which in this article will be called the upper finiteness property
if there exists a finite sequence A1 , . . . , An of elements of A, possibly including
repetitions, such that (A) = (A1 An )1/n , where denotes the ordinary spectral
radius of a square matrix. The upper finiteness property has been investigated
extensively since it was first defined explicitly in [22] (and implicitly in [16]) and
is believed to hold for most finite sets of matrices (though see e.g. [5, 9, 18] for
examples where the property fails). By analogy we will say that A has the lower
finiteness property if (A) = (A1 An )1/n for some finite sequence A1 , . . . , An
of elements of A. The most powerful methods presently available for computing
the upper and lower spectral radii of a finite set of matrices proceed by proving
that the set of matrices in question has the upper or lower finiteness property, and
by exhibiting precisely a product whose normalised spectral radius (A1 An )1/n
realises the upper or lower spectral radius [14]. In the interest of the computation of
these joint spectral characteristics it is therefore important to understand exactly
how frequently these finiteness properties can be expected to hold.
In the article [8], J. Bochi and the present author investigated the continuity
properties of the lower spectral radius both in the context of compact sets of matrices (with respect to the Hausdorff metric) and as a function of finite sets of matrices
with fixed cardinality. As a corollary, results were also obtained on the genericity
of the lower finiteness property. In order to state these results fully we require a
definition:
Definition 1.1. Let A1 , A2 be 2 2 real matrices. We will say that (A1 , A2 ) is a
resistant pair if there exist c, , > 0 with the following property: for all n 1 and
for all choices of integers i1 , . . . , in {1, 2}, if at most n of the integers ik are
equal to 2, then kAi1 Ain k cen .
The following conjecture was formulated in [6]:
Conjecture 1.2 (Bochi-Fayad Conjecture). Let H denote the set of all 2 2 real
matrices with unit determinant and unequal real eigenvalues, and let R denote the
set of all 2 2 real matrices unit determinant and non-real eigenvalues. Then the
set of all resistant pairs (H, R) H R has full Lebesgue measure.
Some partial results in the direction of this conjecture may be found in [1, 2, 13];
we note that resistant pairs (H, R) H R can easily be constructed explicitly in
the special case where Rn is equal to the identity for some integer n 1, see [8].
Let M2 (R) denote the set of all 2 2 real matrices. J. Bochi and the author have
proved the following result:
Theorem 1.3 ([8]). Define
U := {(H, R) : H H, R R and 0 < < } M2 (R)2 .
Then there is a dense G subset of U in which the lower finiteness property fails.
If additionally the Bochi-Fayad conjecture is true, then the set of discontinuities of
the lower spectral radius on U has full Lebesgue measure.

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

Since U is open, this result motivates in particular the following conjecture:

Conjecture 1.4. The set of discontinuities of : M2 (R)2 R has Hausdorff
dimension 8.
At the present time the resolution of Conjecture 1.2 seems remote. The purpose
of this article is to investigate what partial results in the direction of Conjecture
1.4 may be obtained without the resolution of Conjecture 1.2. In the process of
this investigation we have also obtained results on the lower finiteness property.
Let us write P for the set of all 2 2 real matrices with rank one which are
not nilpotent, and E for the set of all 2 2 real matrices with non-real eigenvalues.
In this article we investigate the size of the set of discontinuities of and the
prevalence of the lower finiteness property on the 7-dimensional manifold P E
M2 (R)2 , where the lower spectral radius is much more easily understood than on
the set investigated in Theorem 1.3. The situation which emerges is surprisingly
subtle: we find that the set of discontininuities of on P E, and the set of pairs
(A1 , A2 ) P E which satisfy the lower finiteness property, are both dense subsets
of P E with full Lebesgue measure; but on the other hand their complements are
dense G sets. The set of discontinuities in this set may therefore be seen as large
in the sense of measure theory, but small in the sense of point-set topology, and
similarly for the set in which the lower finiteness property holds. Yet furthermore,
there exist dense subsets of P E where is continuous but the lower finiteness
property holds, and where is discontinuous but the lower finiteness property fails.
We prove the following theorem:
Theorem 1.5. Let P M2 (R) denote the set of all non-nilpotent rank-one matrices and E M2 (R) the set of all matrices with non-real eigenvalues, and define
four subsets U1 , U2 , U3 , U4 of P E as follows:
U1 is the set of all A P E such that (A) = 0, is continuous at A, and
A does not have the lower finiteness property.
U2 is the set of all A P E such that (A) = 0, is continuous at A, and
A does have the lower finiteness property.
U3 is the set of all A P E such that (A) > 0, is discontinuous at A,
and A does not have the lower finiteness property.
U4 is the set of all A P E such that (A) > 0, is discontinuous at A,
and A does have the lower finiteness property.
Then P E is equal to the union of the four sets Ui , and each of the four sets Ui
is dense in P E. The set U1 is equal to a countable intersection of dense open
subsets of P E: conversely, the set U4 is of full Lebesgue measure.
Since P E is a seven-dimensional submanifold of the eight-dimensional linear space M2 (R)2 , we immediately obtain the following result in the direction of
Conjecture 1.4:
Corollary 1.6. The set of discontinuities of the lower spectral radius : M2 (R)2
R has Hausdorff dimension at least 7.
The proof of Theorem 1.5 is essentially divided into four parts: some common
preliminaries, an analysis of the sets U1 and U2 , an analysis of the set U4 , and an
analysis of the set U3 . The last of these parts is by far the most intricate, and
indeed is similar in length to the other three parts combined. The remainder of

IAN D. MORRIS

this paper follows the structure of this proof: in 2 we prove certain preliminary
statements; in 3 we examine the case (A) = 0; in 4 and 5 we examine U4 and U3
respectively; and in 6, all of these results are synthesised into the proof of Theorem
1.5.
2. Preliminaries
In this section we show that a general pair (H, R) P E may be reduced by
rescaling and a change of basis to the situation in which R is a rotation and H has
only zero entries in its bottom row. Using this reduction we obtain a formula for
the lower spectral radius which will be used in all subsequent sections. We begin
with the following fundamental observation:
Lemma 2.1. Let (H, R) P E. Then there
simultaneously similar to a pair of matrices of

 

cos
,
0 0
sin

exists > 0 such that 1 (H, R) is

the form

sin
cos

where , , R and 6= 0.
Proof. Let H, R P E. By dividing both matrices by (det R)1/2 if necessary
we may assume that det R = 1; subject to this assumption we will show that the
pair (H, R) is similar to a pair of the desired form.
Since R is a real matrix with unit determinant and complex eigenvalues, there
exists an invertible matrix A such that ARA1 is a rotation matrix. Let us
write H := AHA1 . Since H has trace 6= 0 and determinant 0, it has a
one-dimensional eigenspace corresponding to the eigenvalue . It is clear that by
conjugating H by a suitable rotation B we may obtain a matrix H := BH B 1
for which the horizontal axis is an eigenspace corresponding to the eigenvalue ;
since rotations commute with one another, this additional conjugation leaves the
rotation matrix ARA1 unchanged. Since H preserves the horizontal axis we may
write



H =
0
for real numbers and , but since det H = det H = 0 and 6= 0 we must have
= 0.

Lemma 2.2. Let H P and R E, and suppose that we have





cos sin
(1)
H=
,
R=
0 0
sin cos
for some , , R, where 6= 0. Then for all finite sequences n1 , . . . , nk and
m1 , . . . , mk of positive integers
tr (H

nk

Rmk

n1

Rm1 )

Pk

i=1

ni

k
Y

i=1

cos mi + 1 sin mi

and hence in particular

(H nk Rmk H n1 Rm1 ) = ||

Pk

i=1

ni

k
Y


cos mi + 1 sin mi  .
i=1

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

Proof. By the Cayley-Hamilton theorem or direct computation we have H 2 = H

and hence by a trivial induction H n = n1 H for every n 1. It follows that for
each n, m 1
 n



n1
cos m sin m
H n Rm =
0
0
sin m cos m
 n

1
(cos m + sin m) n (1 cos m sin m)
=
.
0
0
It follows directly that the lower-right and upper-left entries of the matrix
H nk Rmk H n1 Rm1
are respectively zero and
Pk

i=1

ni

k
Y

cos mi + 1 sin mi

i=1

and the result follows.

Lemma 2.3. Let (H, R) P E, and let H nk Rmk H n1 Rm1 be a product of

these two matrices in which every exponent ni , mi is a non-negative integer, and
not all of these integers are zero. Then there exists n 0 such that
o
n
1
1
Pk
(2)
(H nk Rmk H n1 Rm1 ) i=1 (ni +mi ) min (R), (HRn ) n+1 .
Proof. It is clear that the quantities appearing in the statement of the lemma are
homogenous in H and R and are unaffected by a change-of-basis transformation,
so by Lemma 2.1 it is sufficient to consider the case where (H, R) has the form (1).
Let us assume that the product H nk Rmk H n1 Rm1 has been presented in such
a manner that the integer k 1 is minimal. If all of the mi are zero then we have
 Pk 1
 Pk
1
Pk
n
(H nk Rmk H n1 Rm1 ) i=1 (ni +mi ) = H i=1 ni i=1 i = (H)
so that the conclusion of the lemma holds with n := 0. If all of the indices ni are
zero then the left-hand side of (2) similarly reduces to (R) and the conclusion holds
trivially. For the remainder of the proof we may assume without loss of generality
Pk
Pk
that i=1 mi and i=1 ni are nonzero.
If ni is equal to zero for some integer i < k then we may write


(H nk Rmk H n1 Rm1 ) = H nk Rmk H ni+1 Rmi+1 +mi H ni1 H n1 Rm1
which requires only k 1 indices, contradicting the minimalitiy of k. Similarly if
nk = 0 then


(H nk Rmk H n1 Rm1 ) = H nk1 Rmk1 H n1 Rm1 +mk

by the standard identity (AB) = (BA), and the minimality of k is again contradicted. Analogous considerations show that each mi must also be nonzero. We
may therefore apply Lemma 2.2 to obtain
Pk

(H nk Rmk H n1 Rm1 ) = ||

i=1 (ni 1)

k
Y

| cos mi + sin mi | .

i=1

Let us define

r := min | cos mi + sin mi | 1+mi .

1ik

IAN D. MORRIS

Obviously, there exists an integer m 1 such that

1

r = | cos m + sin m| 1+m = (HRm ) 1+m .

Now, it is clear that we may write
(H nk Rmk H n1 Rm1 ) = ||
||

Pk

i=1 (ni 1)

Pk

i=1 (ni 1)

k
Y

| cos mi + sin mi |

i=1
Pk

i=1 (1+mi )

If r > || = (H) then we have

(H nk Rmk H n1 Rm1 ) > ||

Pk

i=1 (ni +mi )

Pk

(H)

i=1 (ni +mi )

and the conclusion of the lemma is is satisfied with n := 0. Otherwise we have

(H nk Rmk H n1 Rm1 ) r

Pk

i=1 (ni +mi )

and therefore
(H nk Rmk H n1 Rm1 )

1
Pk
(ni +mi )
i=1

r = (HRm ) 1+m

as required. The proof is complete.

We immediately deduce the following:

Lemma 2.4. Let (H, R) P E. Then
1

(H, R) = inf (HRn ) n+1 .

n0

Proof. We have
(HRn ) kHRn k kHk kRn k
for every n 0, and therefore
1

inf (HRn ) 1+n lim sup (HRn ) 1+n lim sup (kHk kRn k) 1+n = (R)

n0

using Gelfands formula. It follows by this inequality together with Lemma 2.3 that
1

(H, R) inf (HRn ) n+1 ,

n0

and the general identity

(A1 , . . . , Ak ) = inf

min

n1 1i1 ,...,in k

(Ai1 Ain ) n

(see e.g. [20, 1.2.1]) yields the other direction of inequality.

Finally, the following lemma will be used on two occasions in which it is helpful
to estimate small quantities of the form | cos + sin |.
Lemma 2.5. Let , R where is nonzero.
If = 0 then define 0 := /2;
otherwise, define 0 := tan1 (/) 12 , 12 , so that in either event we have
cos 0 + sin 0 = 0. Then there exists a constant C0 > 1 such that for every
R
1
dist(, 0 + Z) | cos + sin | C0 dist(, 0 + Z),
C0
where dist(, a + bZ) denotes the distance from the real number to the nearest real
number of the form a + b with Z.

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

Proof. Define f, g : R R by f () := | cos + sin | and g() := dist(, 0 + Z).

We must show that the ratio f ()/g() is uniformly bounded away from zero and
infinity with respect to R \ (0 + Z). Since both f and g are periodic with
period , it is sufficient to show that f ()/g()
 is bounded away from zero and
infinity for belonging to theinterval 2 , 2 \ {0 }.
The function f /g : 2 , 2 \ {0 } R is clearly well-defined, continuous and
nowhere zero, so to show that this function is bounded away from zero and infinity
it is sufficient to show that lim0 f ()/g() exists and is nonzero. We have
lim

f ()
f (0 + h)
= lim
g() h0 g(0 + h)
| cos(0 + h) + sin(0 + h)|
= lim
h0
|h|


 cos(0 + h) + sin(0 + h) cos(0 ) sin(0 ) 


= lim 

h0
h
= | cos 0 sin 0 |

so the result follows if we can show that cos 0 sin 0 is nonzero. If cos 0
sin 0 = 0 and cos 0 6= 0 then tan 0 = / and therefore / = /, implying
the equation 2 + 2 = 0 which is impossible since both numbers are real and
6= 0. If instead cos 0 sin 0 = 0 and also cos 0 = 0 then we have = 0, a
contradiction.

3. On the existence of zero products
In this section we investigate those pairs (H, R) P E such that there exists
a finite product of the two matrices H and R which is equal to the zero matrix.
We begin with the following result, which shows that this situation is, in a certain
topological sense, uncommon.
Lemma 3.1. For every two finite sequences of non-negative integers n1 , . . . , nk and
m1 , . . . , mk+1 , with all integers except possibly m1 and mk+1 being nonzero, the set
Z := {(A, B) P E : B mk+1 Ank B mk Ank1 An1 B m1 = 0}
is closed and has empty interior.
Proof. Since every B E is invertible we note that the equation
B mk+1 Ank B mk Ank1 An1 B m1 = 0
for some (A, B) P E is equivalent to the equation
Ank B mk Ank1 An1 B m1 + = 0
for the same A and B and every non-negative integer , so we may assume without
loss of generality that mk+1 = 0 and m1 > 0. By continuity the set Z is obviously
closed. For each A P let us define a function A : M2 (R) M2 (R) by
A (B) := Ank B mk An1 B m1 .
Let us suppose for a contradiction that the set Z contains a nonempty open set, and
let (A0 , B0 ) be an interior point of Z. Each of the four entries of the matrix A0 (B)
is a polynomial function of the entries of B, and by hypothesis is identically zero in
a neighbourhood of B0 . It follows that each of these entries is the zero polynomial,

IAN D. MORRIS

and therefore A0 is identically zero throughout M2 (R); however, we obviously

have
Pk
n +m
(A0 ) = A0 i=1 i i 6= 0
since A0 is not nilpotent, and we obtain a contradiction. We conclude that Z has
empty interior as claimed.

In the converse direction, the existence of a product of the matrices (H, R) which
is zero is in fact a dense property in P E:
Lemma 3.2. The set of all (H, R) P E such that (H, R) = 0 and (H, R)
satisfies the lower finiteness property is a dense set.
R
arbiProof. Given (H, R) P E we must show that there exist matrices H,

trarily close to (H, R) such that (H, R) = 0 and such that H, R satisfies the lower
finiteness property. Clearly this property is unaffected by applying a change of
basis or multiplying both matrices H and R by the same scalar, so using Lemma
2.1 we may assume without loss of generality that the pair (H, R) has the form





cos sin
H=
,
R=
0 0
sin cos
for some real numbers , , R, where 6= 0. If (H, R) is as above for some
, , R then using Lemma 2.2 we have
trRm H = cos m + sin m
and if this trace is zero then by the Cayley-Hamilton theorem we have (Rm H)2 = 0,
which implies HRm H = 0 since R is invertible. To prove the lemma it is thus
sufficient to prove the following: for every fixed , R the set

{ R : cos m + sin m = 0}

m=1

is dense in R. Given and , it is clearly sufficient to show that for every m 1,

every interval in R with length greater than /m contains a number such that
cos m + sin m = 0. But this result is straightforward to prove: if we define
m : R R by m () := | cos m + sin m| then this function is periodic with
period /m, and it has at least one zero, since either m (/2m) = 0 if = 0, or
m (m1 tan1 (/)) = 0 if 6= 0. In particular every interval of length at least
/m contains a zero of m as required.

Finally, we note the following implication of the existence of a zero product:
Lemma 3.3. If H, R M2 (R) and (H, R) = 0 then : M2 (R)2 R is continuous
at (H, R).
Proof. Since for any A1 , A2 M2 (R) we have
(A1 , A2 ) = inf

inf

n1 i1 ,...,in {1,2}

kAi1 . . . Ain k n

(see e.g. [20, 1.2.2]) the lower spectral radius of (A1 , A2 ) is equal to the infimum of
a sequence of continuous functions of (A1 , A2 ) and hence is upper semi-continuous.
Since it is also non-negative, this implies that every zero of the lower spectral radius
is a point of continuity.


LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

4. Behaviour on a set of full Lebesgue measure

In this section we shall prove that the set of all (H, R) P E such that the lower
finiteness property is satisfied and such that the lower spectral radius of (H, R) is
nonzero is a set of full Lebesgue measure. In this section we shall find it convenient
to consider angles as belonging to the quotient space R/2Z on which Lebesgue
measure is a finite measure as opposed to R, which has infinite Lebesgue measure.
Our first objective shall be to prove the following proposition:
Proposition 4.1. Fix , R with 6= 0 and define





cos sin
H :=
,
R :=
0 0
sin
cos
for every R/2Z. Then for Lebesgue almost every R/2Z there exists n 0
such that
1
(H, R ) = (HRn ) n+1 > 0.
Let us make some observations on the proof of Proposition 4.1. By Lemma 2.4
we find that
1
(H, R ) = inf (HRn ) n+1
n0

for every R/2Z, and by Lemma 2.2 we have

1

(HRm ) m+1 = | cos m + sin m| m+1

for every R/2Z and m 0. To prove the proposition it is therefore sufficient
to show that for Lebegue almost all R/2Z the infimum
1

inf | cos m + sin m| m+1

m0

is attained for some integer m 0, and is nonzero. The proposition thus follows
from the combination of Lemmas 4.2, 4.3 and 4.4 below:
Lemma 4.2. Let , R. If R is not a rational multiple of , then
(3)

inf | cos m + sin m| m+1 < 1.

m0

In particular (3) holds for Lebesgue almost every R/2Z.

Proof. For each integer m the value of | cos m + sin m| is unaffected by adding
an integer multiple of to , so we may freely identify R with its equivalence
class modulo 2Z. To prove (3) it is clearly sufficient to show that
(4)

inf {| cos m + sin m| : m 0} < 1.

When is not a rational multiple of the set {m R/2Z : m 0} is dense in

R/2Z, so by continuity we have
(5)

inf {| cos m + sin m| : m 0} = min {| cos + sin | : R/2Z} .

If = 0 then | cos(/2) + sin(/2)| = 0, and if 6= 0 then we have | cos 0 +

sin 0 | = 0 where 0 R/2Z satisfies tan 0 = /. In either case the
minimum in (5) is zero and hence the inequality (4) is satisfied. Since for Lebesgue
almost every R/2Z, is not a rational multiple of , the result follows.


10

IAN D. MORRIS

Lemma 4.3. Let , R. Then for Lebesgue almost every R/2Z,

1

lim | cos m + sin m| m+1 = 1.

Proof. Since trivially | cos m + sin m| || + || for every m 0 we have

1

lim sup | cos m + sin m| m+1 1

m

for every R/2Z. To prove the lemma we must show that the limit inferior is
greater than or equal to one for Lebesgue almost every R/2Z.
By Lemma 2.5 there exist constants C0 > 1 and 0 ( 2 , 2 ) such that for every
R we have
min | 0 k| C0 | cos + sin |.
kZ

In particular if 0 < 2 and | cos + sin | < t < /2C0 , then must belong
to the set
(0 C0 t, 0 + C0 t) (0 + C0 t, 0 + + C0 t) (0 + 2 C0 t, 0 + 2 + C0 t).

We deduce that for every t (0, 2C

) the set
0

{ [0, 2) : | cos + sin | < t}

has Lebesgue measure not greater than 6C0 t, and clearly this implies
({ R/2Z : | cos + sin | < t}) 6C0 t
for the same range of real numbers t, where denotes Lebesgue measure on R/2Z.
Since for every m 1 the transformation R/2Z R/2Z defined by 7 m

)
preserves Lebesgue measure, it follows that for all m 1 and t (0, 2C
0
({ R/2Z : | cos m + sin m| < t}) 6C0 t.
In particular
n
o

R/2Z : | cos m + sin m| < e m

6C0 e m

for every sufficiently large integer m. By the Borel-Cantelli Lemma we deduce

that

for Lebesgue almost every R/2Z we have | cos m + sin m| e m for

every sufficiently large m, and therefore
1
  m+1
1
=1
lim inf | cos m + sin m| m+1 lim inf e m
m

as required to complete the proof.

Lemma 4.4. Let R. Then for Lebesgue almost every R/2Z

1

inf | cos m + sin m| m+1 > 0.

m0

Proof. It follows from the combination of Lemma 4.2 and Lemma 4.3 that for
Lebesgue almost every R/2Z the above infimum is attained at some integer
n 0 depending on . To prove the lemma it is therefore sufficient to show that
the set
{ R/2Z : m 0 such that cos m + sin m = 0}
has measure zero; but this set is countable, since for each m 0 the set
{ R/2Z : cos m + sin m = 0}
is finite, and therefore it has Lebesgue measure zero as required.

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

11

As was remarked previously, the combination of Lemmas 4.2, 4.3 and 4.4 proves
Proposition 4.1. We may now deduce the following result, which is the main result
of this section:
Proposition 4.5. The set
n
o
1
X := (H, R) P E : (H, R) = (HRn ) n+1 > 0 for some n 0

has full Lebesgue measure as a subset of P E.

2
Proof. Let us define SL
2 (R) (R \ {0}) R R/2Z to be the set of all
(A, , , , ) such that the pair

 


cos sin
,
0 0
sin
cos

does not belong to X. By Proposition 4.1 this set has zero Lebesgue measure
2
as a subset of SL
2 (R) (R \ {0}) R R/2Z. Let us now define a function

2
: SL2 (R) (R \ {0}) R R/2Z P E by

 




cos sin
(A, , , , ) = A
A1 , A
A1 .
0 0
sin cos
Using Lemma 2.1 it is easy to see that (H, R) P E fails to belong to X if and
only if (H, R) (). The function is a smooth map between 7-dimensional
manifolds, so it is locally Lipschitz and maps sets whose 7-dimensional Lebesgue
measure is equal to zero onto other sets whose 7-dimensional Lebesgue measure is
equal to zero. In particular () has Lebesgue measure zero as required.

5. Positive lower spectral radius without the lower finiteness
property
In this section we investigate those pairs (H, R) P E for which the lower
spectral radius is positive but the lower finiteness property is not satisfied, showing
that the set of all such pairs in dense in P E. This outcome is obtained as a
corollary of the following result:
Proposition 5.1. Fix R \ {0}. For each R and each R/2Z define





cos sin
H :=
,
R :=
.
0 0
sin
cos
Then the set of all (, ) R2 such that the pair (H , R ) has positive lower spectral
radius and does not have the lower finiteness property is a dense subset of R2 .
If , R are fixed, with 6= 0, then Lemma 2.4 together with Lemma 2.2 tells
us that we have
1

(H , R ) = inf (H Rm ) 1+m = inf | cos m + sin m| 1+m .

m0

m0

On the other hand the combination of Lemma 2.3 and Lemma 2.2 tells us that
the lower finiteness property for (H , R ) holds if and only if either (H , R ) =
(R ) = 1, or there exists m 0 such that
1

(H , R ) = | cos m + sin m| 1+m .

Now, Lemma 2.5 tells us that the quantity | cos m+ sin m| is well-approximated
by the quantity dist(m, 0 + Z), where 0 is the unique zero of the function

12

IAN D. MORRIS

7 | cos + sin | in ( 2 , 2 ]. The behaviour of | cos m + sin m|1/(1+m)

for large m is thus linked with the following problem of inhomogenous Diophantine
approximation: for large m, how close to an integer is 1 (m 0 )?
To investigate this problem we apply the theory of continued fractions. In the
sequel we will refer to the classic treatment of the subject by A. Ya. Khinchin for
specific results. We briefly recall some relevant definitions. Recall that if a1 , a2 , . . .
is a finite or infinite sequence of positive integers then we define the sequence of
convergents pk /qk by
p3
p0
0
1
p2
1
1
p1
,
:= ,
:= ,
:=
:=
,
q0
1
q1
a1
q2
q3
a1 + a12
a1 + a2 +1 1
a3

pn
:=
qn
a1 +

1
1

a2 +

1
a3 + 1
an

for each n for which the integers a1 , . . . , an are defined. The above representations
pk /qk in least terms can equivalently be defined by the recurrence relations
(6)

pk+1 := ak+1 pk + pk1 ,

qk+1 := ak+1 qk + qk1

for every k 0, where we additionally define p1 := 1, q1 := 0. If (0, 1)

is irrational then there exists a unique sequence of positive integers (an )
n=1 for
which the associated sequence of convergents pk /qk converges to , and for every
sequence of positive integers (an )
n=1 the corresponding sequence of convergents is
a convergent sequence with an irrational limit. If (an ) is such a sequence then we
denote the limit of the corresponding sequence of convergents by [a1 , a2 , . . .]; more
generally, if (an )
n=0 is a sequence such that a0 Z and an N for all n 1, then
we write [a0 ; a1 , a2 , . . .] := a0 + [a1 , a2 , . . .]. The signifance of continued fractions for
this work is their property of being good rational approximations: if = [a1 , a2 , . . .]
then we have
1
1
< |qn pn | <
qn+1 + qn
qn+1
for every n 1, and if qn < < qn+1 then
1
| k|
2
for every k Z which shares no common factors with , see for example Theorems
9, 13 and 19 of [21]. The proof of Proposition 5.1 will be executed by constructing,
for each fixed and in the case where R belongs to a suitable dense set, a
dense set of R whose convergents satisfy a certain precise growth condition.
The proof of Proposition 5.1 begins with the following preliminary result:
Lemma 5.2. Let 0 ( 2b , 2b ) \ Q and ab Q in least terms, where b is prime and
a is nonzero, and let > 0. Then there exist integers N 1 and a0 , a1 , . . . , aN with
the following properties: the associated sequence of convergents (pk /qk )N
k=0 satisfies
(7)

pN + a0 qN pN 1 + a0 qN 1 a

mod b,

and if (an )
n=0 is any infinite extension of the finite sequence a0 , a1 , . . . , aN then
|[a0 ; a1 , a2 , . . .] 0 | < .
Proof. Let (bn )
n=0 be the unique sequence of integers such that 0 = [b0 ; b1 , b2 , . . .],
where bn 1 for every n 1. Let (Pn /Qn )
n=0 be the associated sequence of
convergents, and let n0 1 be large enough that 1/Q2n0 < /2. Define ak := bk for

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

13

n0
0 k n0 . We will show that there exists a finite extension (an )N
n=0 of (an )n=0
N
such that the associated sequence of convergents (pk /qk )k=0 satisfies (7). We will
subsequently deduce that the second desired property also holds.
Let us first show that it is impossible for the integers pn0 + a0 qn0 and pn0 1 +
a0 qn0 1 to both be divisible by b. We prove this claim by contradiction. If pk +a0 qk
and pk1 + a0 qk1 are both divisible by b for some integer k 1, then it follows
from the relation (6) that

pk + a0 qk = ak (pk1 + a0 qk1 ) + pk2 + a0 qk2

and therefore pk2 +a0 qk2 is also divisible by b. If pn0 +a0 qn0 and pn0 1 +a0 qn0 1
are both divisible by b, then by inductive descent it follows that p1 + a0 q1 is
divisible by b, but p1 + a0 q1 = 1, which is a contradiction since b is prime. This
contradiction proves the claim.
Let us consider the possibilities which remain. If pn0 + a0 qn0 is not divisible by
b, then since b is prime we may choose an integer k 1 such that
k(pn0 + a0 qn0 ) + pn0 1 + a0 qn0 1 a

mod b.

Taking an0 +1 := k we find that pn0 +1 + a0 qn0 +1 a mod b. If we now take

an0 +2 := b then
pn0 +2 + a0 qn0 +2 = b(pn0 +1 + a0 qn0 +1 ) + pn0 + a0 qn0 a

mod b

so we can now take N := n0 + 2 and we have obtained the desired extension of

0
(an )nn=0
in this case. If on the other hand pn0 + a0 qn0 is divisible by b, then by the
previous claim pn0 1 + a0 qn0 1 is not. Taking an0 +1 := 1 we have
pn0 +1 + a0 qn0 +1 = pn0 + a0 qn0 + pn0 1 + a0 qn0 1 pn0 1 + a0 qn0 1

mod b.

This reduces the problem to the preceding case but with n0 + 1 in place of n0 . In
n0
either event we are able to construct a finite extension (an )N
n=0 of (an )n=0 such that
N
the associated sequence of convergents (pk /qk )k=0 satisfies (7) as desired.
Let us now suppose that (an )
n=0 is an infinite extension of the finite sequence
with
a

1
for
every
n
>
N , and let (pk /qk )
(an )N
n
n=0
k=0 be the associated sequence
of convergents. Let := [a0 ; a1 , a2 , . . .] R. The sequence of convergents satisfies
pk /qk = Pk /Qk for k n0 since the sequences (an ) and (bn ) coincide up to this
point. By [21, Theorem 16] it follows that
 




1
1
Pn0   Pn0

+
0  2 + 2 <
| 0 | 
Qn   Qn
Q
Q
0

n0

since by the definition of n0 we have

1/Q2n0

n0

< /2. The proof is complete.

The core of the proof of Proposition 5.1 rests in the following long lemma:
Lemma 5.3. Let ab Q in least terms where b is prime and a is nonzero, let
K > 1, let 0 ( 2b , 2b ) \ Q and let > 0. Then there exists an infinite sequence
of integers (an )
n=0 , where an 1 for every n 1, such that the irrational number
:= [a0 ; a1 , a2 , . . .] belongs to ( 2b , 2b ), satisfies | 0 | < , and has the following
additional properties: for every n 0 there exists q > n such that
(8)

dist(n, a + bZ) n+1 > K n+1 dist(q, a + bZ) q+1 ,

and furthermore
(9)

inf dist(n, a + bZ) n+1 > 0.

n0

14

IAN D. MORRIS

Proof. By replacing > 0 with a smaller value if necessary, we may ensure that
| 0 | < implies that ( 2b , 2b ). By Lemma 5.2 there exist integers N 1
and a0 Z, a1 , . . . , aN N such that every infinite extension (an )
n=0 of the finite
sequence (an )N
has
the
property
that
|[a
;
a
,
.
.
.]

|
<
,
and
such that the
0
1
0
n=0
associated sequence of convergents (pk /qk )N
satisfies
k=0
pN + a0 qN pN 1 + a0 qN 1 a

mod b.

We will show that this sequence (an )N

n=0 may be extended to an infinite sequence
in
such
a
way
that
the
associated
number := [a0 ; a1 , . . .] satisfies proper(an )
n=0
ties (8) and (9).
Let aN +1 be an integer which is divisible by b and is so large that the integer
qN +1 := aN +1 qN + qN 1 satisfies
(10)

1
1+q

qN +1N > 2KqN > (2KqN ) 1+qN 1 .

Since aN +1 is a multiple of b, and by hypothesis pN + a0 qN pN 1 + a0 qN 1 a

mod b, the equation
pN +1 + a0 qN +1 = aN +1 (pN + a0 qN ) + pN 1 + a0 qN 1
implies
pN +1 + a0 qN +1 pN + a0 qN a

mod b.

Now let C > 1 be so large that

qN +1 C 1+qN .
We claim that there exists an infinite sequence (an )
n=0 which extends the finite
sequence a0 , a1 , . . . , aN +1 such that for each n N + 1,
(11)

pn + a0 qn pn1 + a0 qn1 a

mod b

and
(12)

1
1+qn1

(2Kqn1 ) 1+qn2 < qn

Pn2

C(4Kb)

1
k=N 1+qk

The sequence is constructed inductively. Let us assume that the sequence (an )m
n=0
has been constructed in such a way that the above properties are satisfied for every
n in the range N + 1 n m. (Clearly this already holds in the particular case
m = N + 1.) We will show that an integer am+1 may be chosen such that the
extended sequence (an )m+1
n=0 has the above properties for all integers n in the range
N + 1 n m + 1.
Given the sequence (an )m
n=0 , we define the integer am+1 by


1+qm
1
1+qm1
.
am+1 := b qm (2Kqm )
Since
pm+1 + a0 qm+1 = am+1 (pm + a0 qm ) + pm1 + a0 qm1
and b divides am+1 , we have
pm+1 + a0 qm+1 pm1 + a0 qm1 a

mod b,

and we know already that pm + a0 qm a mod b, so the property (11) is also

satisfied for n := m + 1. For the integer qm+1 := am+1 qm + qm1 we have the

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

15

inequality
1
1+qm

qm+1

 1+q1

1+qm
m
1+qm1
qm1 + bqm + b(2Kqm )

 1+q1
1+qm
m
1+qm1
< ((2Kb + b + 1) qm )
Pm1

1
k=N 1+qk

< (4Kbqm ) 1+qm1 C(4Kb)

where we have used the induction hypothesis to obtain the upper bound
1
1+qm1

qm

Pm2

C(4Kb)

1
n=N 1+qk

The second inequality of (12) is thus satisfied for n = m + 1. Finally we have



 1+q1

1
1+qm
m
1+qm
1
1+qm1
qm+1 = bqm qm (2Kqm )
+ qm1
1

> b 1+qm (2Kqm ) 1+qm1

1

> (2Kqm ) 1+qm1

so that the left-hand side of (12) is also satisfied for n = m + 1. We conclude
that the choice of am+1 given above provides the requisite extension of the finite
sequence (an )m
n=0 , and by induction we deduce the existence of the desired infinite

sequence (an )
n=0 . For the remainder of the proof we fix the sequence (an )n=0
defined in the preceding manner.
Let us show that the number := [a0 ; a1 , a2 , . . .] thus constructed has the properties required in the statement of the lemma. By [21, Theorem 9] we have for
every n 0
(13)

|qn ( a0 ) pn | <

1
qn+1

If n N then necessarily qn+1 q2 = a2 q1 + 1 2, so (11) and (13) together

imply that the nearest integer to qn is congruent to a modulo b. For n N we
thus have
(14)
1

 1+q
 1+q1

n
n1
1
1
1
1
1+qn
1+qn
<
dist (qn , a + bZ)
= |qn ( a0 ) pn |
<
qn+1
2Kqn
using (12).
We now make the following claim: if k 0 is any integer, and n 1 is the
unique integer such that qn k < qn+1 , then
dist(k, a + bZ) >

1
.
2qn+1

We consider several cases. Firstly, if k = 0 then n = 1 and we have

dist(k, a + bZ) = min{|a|, |b a|} 1 >

1
1
=
2
2qn+1

as required. Suppose next that k 1. Let Z such that dist(k, a+bZ) = |k|,
and let d 1 denote the greatest common divisor of k and ||. Let k = k d, = d

16

IAN D. MORRIS

where k and are coprime, and let m 0 such that qm k < qm+1 . If k > qm
then by [21, Theorem 19] we have |k | 1/2k and therefore
dist(k, a + bZ) = |k | = d|k |

d
1
1

>

2k
2k
2qn+1

as required. If on the other hand k = qm then since k 1 we have qm+1 > 1

and therefore qm+1 2. It follows using (13) that the nearest integer to qm is
pm + a0 qm . By [21, Theorem 13] we may therefore deduce
dist(k, a+ bZ) = |k | = d|qm | d|qm ( a0 ) pm | >

d
1
>
qm+1 + qm
2qn+1

and this completes the proof of the claim.

We may now prove that has the properties required in the statement of the
lemma. Let k 0 be an integer, and let n 1 such that qn k < qn+1 . We will
show that there exists q max{qN , k + 1} such that
1

dist(k, a + bZ) 1+k > K 1+k dist(q, a + bZ) 1+q .

We consider two cases. If n < N , then by the previous claim together with (10) we
have
1
 1+k

1
1
1

dist(k, a + bZ) 1+k >

2qn+1
2qN


 1+q1
 1+q1
N
N
1
1
1
K 1+k
>K
qN +1
qN +1
and therefore
dist(k, a + bZ)

1
1+k

>K

1
1+k

 1+q1

qN +1

> K 1+k dist(qN , a + bZ) 1+qN

where we have used (14). If on the other hand n N , then by the previous claim
together with (14)
1

 1+k
1
1
dist(k, a + bZ) 1+k >
2qn+1
1

 1+q
n
1
1
1+k
K
2Kqn+1
1

> K 1+k dist(qn+1 , a + bZ) 1+qn+1 .

This establishes the first of the two properties required in the statement of the
lemma. It remains to prove (9). We note that the preceding argument implies the
inequality
1

 1+q
n
1
1
1+k
inf
,
inf dist(k, a + bZ)
nN
k0
2qn+1
so to establish (9) it is sufficient to prove the inequality
1
 1+q

n
1
inf
> 0.
nN
qn+1

LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

By [21, Theorem 12] we have qn 2

k=N

1
<
1 + qk

If n N then by (12)

 q1
P
n
1
n1
k=N
C 1 (4Kb)
qn+1

n1
2

k=1

1
1+qk

17

for every n 0 so in particular

1

1+2

k1
2

<

2 2 < 4.

k=0

> C 1 (4Kb)

1
k=N 1+qk

> C 1 (4Kb)4 > 0

and since the last two terms do not depend on n, the infimum in question is nonzero.
This completes the proof of (9) and hence completes the proof of the lemma. 
We may now give the proof of Proposition 5.1, and deduce from it the result
which will be used in the following section.
Proof of Proposition 5.1. Let ,
R, and let > 0; we will define , such that

|
| and | | are smaller than and such that (H , R ) has positive lower
spectral radius but lacks the lower finiteness property. Since R = R+2n for
every R and n Z we may freely assume that ( 2 , 2 ], and by making
an arbitrarily small adjustment to if necessary we may also assume that 1 is
irrational.
Since the set of rationals with prime denominator is dense in R, and the function
| <
tan : ( 2 , 2 ) R is a homeomorphism, we may choose R such that |
)
for
some
nonzero
integer
a
and
prime
natural
number
b5
and 1 = tan( a
b
such that ab ( 12 , 12 ). By Lemma 2.5 we may choose a constant C0 > 1 such that

 a
+ Z C01 | cos + sin |
(15)
C0 | cos + sin | dist ,
b
for every R. By Lemma 5.3 there exist an infinite sequence of integers (an )
n=0
with an 1 for all n 1, and a real number > 0, such that the irrational
< b 1 , and such that
number := [a0 ; a1 , . . . , ] ( 2b , 2b ) satisfies | b 1 |
the inequality
 2 


1
1
C0 b 1+n
1+n
1+m
(16)
dist(n, a + bZ)
>0
>
inf dist(m, a + bZ)
m>n

holds for every integer n 0, where we have used C0 b/ 5C0 / > C0 > 1.

Define :=
b ( 2 , 2 ) so that | | < . Since is irrational, is not a
rational multiple of , so we have
1

inf (H Rn ) 1+n = inf | cos n + sin n| 1+n < 1 = (R )

n0

n0

using Lemmas 2.2 and 4.2. If (H , R ) has the lower finiteness property then
necessarily (H , R ) = (H Rn )1/(n+1) for some integer n 0, since otherwise
Lemma 2.3 together with the previous inequality yields a contradiction. To show
that (H , R ) does not have the lower finiteness property it is therefore sufficient
to show that the infimum
1
inf (H Rn ) n+1
n0

is not attained. By Lemma 2.4, to show that (H , R ) > 0 it is sufficient to show

that the same infimum is nonzero.

18

IAN D. MORRIS

Let us first prove that this infimum is not attained. Given n 0, using Lemma
2.2, (15) and (16) we may find m > n such that
1

(H Rn ) 1+n = | cos n + 0 sin n| 1+n

1

 1+n
a
1
+ Z
C0 1+n dist n,
b
1
 1+n

1

dist(n, a + bZ) 1+n

=
C0 b
 1

1
C0 b 1+n
dist(m, a + bZ) 1+m
>

 1

1
C0 b 1+m
dist(m, a + bZ) 1+m
>

1

 1+m
1
a
= C01+m dist m,
+ Z
b
1
1
| cos m + 0 sin m| 1+m = (H Rm ) 1+m
proving our assertion. To see that (H , R ) is nonzero we note that by a similar
chain of inequalities
1

(H , R ) = inf (H , Rn ) 1+n inf

n0

n0

C0 b

The proof of the proposition is complete.

1
 1+n

dist(n, a + bZ) 1+n

> 0.
C0 b


We now easily deduce the following result which will be used in the proof of
Theorem 1.5.
Corollary 5.4. The set of all (H, R) P E such that (H, R) > 0 and such that
(H, R) does not have the lower finiteness property is a dense subset of P E.
Proof. We argue similarly to the proof of Proposition 4.5. Let us define
2
2
SL
2 (R) (R \ {0}) R to be the set of all (A, , , , ) such that the pair

 


cos sin
,
0 0
sin
cos
has nonzero lower spectral radius but fails to have the lower finiteness property. By
2
2
Proposition 5.1 this set is a dense subset of SL
2 (R) (R \ {0}) R . If we define

2
2
: SL2 (R) (R \ {0}) R P E by

 




cos sin
1
1
(A, , , , ) := A
A ,A
A
0 0
sin cos
then by Lemma 2.1 the function is surjective. Clearly is also continuous. It is
straightforward to see that if (H, R) () then (H, R) has nonzero lower spectral
radius and does not have the lower finiteness property, and since is continuous
and surjective, () is dense in P E.


LOWER SPECTRAL RADIUS FOR LOW-RANK MATRICES

19

6. Proof of Theorem 1.5

We first claim that : P E R is continuous at (H, R) if and only if (H, R) =
0. On the one hand, by Proposition 4.5 the set of all (H, R) P E such that
(H, R) > 0 has full Lebesgue measure, and in particular this set is dense. On the
other hand by Lemma 3.2 the set of all (H, R) P E such that (H, R) = 0 is
also dense. In particular if (H, R) P E and (H, R) > 0 then (H, R) is a point
of discontinuity of ; but by Lemma 3.3, if (H, R) P E and (H, R) = 0 then
is continuous at (H, R), which proves the claim. We deduce from the claim that
P E is equal to the disjoint union of the four sets Ui .
By Lemma 3.2, Corollary 5.4 and the above considerations, the sets U2 and U3
are dense in P E. For every n 1 the set


1
Xn := Int (H, R) P E : (H, R) <
n
is by definition open, and by Lemma 3.3 this set contains every zero of in P E.
By Lemma 3.2 each Xn is dense. On the other hand, for each n1 , . . . , nk 0 and
m1 , . . . , mk 0, by Lemma 3.1 the set
Y(n1 ,...,nk ;m1 ,...,mk ) := {(H, R) P E : H nk Rmk H n1 Rm1 6= 0}
is open and dense. If there exists a product H nk Rmk H n1 Rm1 with zero spectral
radius then by the Cayley-Hamilton theeorem the product (H nk Rmk H n1 Rm1 )2
must be zero, so the set
Y := {(H, R) P E : no product of H and R is zero}
is equal to the set

Y(n1 ,...,nk ,m1 ,...,mk )

k=1 n1 ,...,nk 0
m1 ,...,mk 0

which is the intersection of countably many open dense sets. Since furthermore

\
U1 = Y
Xn
n=1

we conclude that U1 is equal to the intersection of countably many open dense

subsets of P E. In particular it is dense by Baires theorem. We finally note that
the set U4 has full Lebesgue measure by Proposition 4.5, and this completes the
proof of Theorem 1.5.
7. Acknowledgements
This research was supported by EPSRC grant EP/L026953/1.
References
[1] Jeffrey Allen, Benjamin Seeger, and Deborah Unger, On the size of the resonant set
for the products of 2 2 matrices, Involve, 4 (2011), pp. 157166.
[2] Artur Avila and Thomas Roblin, Uniform exponential growth for some SL(2, R) matrix
products, J. Mod. Dyn., 3 (2009), pp. 549554.
[3] Vincent D. Blondel, Julien Cassaigne, and Rapha
el M. Jungers, On the number of power-free binary words for 2 < 7/3, Theoret. Comput. Sci., 410 (2009), pp. 28232833.
[4] Vincent D. Blondel and Yurii Nesterov, Computationally efficient approximations of
the joint spectral radius, SIAM J. Matrix Anal. Appl., 27 (2005), pp. 256272 (electronic).

20

IAN D. MORRIS

[5] Vincent D. Blondel, Jacques Theys, and Alexander A. Vladimirov, An elementary

counterexample to the finiteness conjecture, SIAM J. Matrix Anal. Appl., 24 (2003), pp. 963
970 (electronic).
[6] Jairo Bochi and Bassam Fayad, Dichotomies between uniform hyperbolicity and zero Lyapunov exponents for SL(2, R) cocycles, Bull. Braz. Math. Soc. (N.S.), 37 (2006), pp. 307349.
[7] Jairo Bochi and Ian D. Morris, Continuity properties of the lower spectral radius. arXiv
preprint, 2013.
, Continuity properties of the lower spectral radius, Proc. Lond. Math. Soc. (3), 110
[8]
(2015), pp. 477509.
[9] Thierry Bousch and Jean Mairesse, Asymptotic height optimization for topical IFS, Tetris
heaps, and the finiteness conjecture, J. Amer. Math. Soc., 15 (2002), pp. 77111 (electronic).
[10] Julien Cassaigne, Vesa Halava, Tesa Harju, and Franc
ois Nicolas, Tighter undecidability bounds for matrix mortality, zero-in-the-corner problems, and more. arXiv preprint
1404.0644, 2014.
[11] F. Michel Dekking and Bram Kuijvenhoven, Differences of random Cantor sets and lower
spectral radii, J. Eur. Math. Soc., 13 (2011), pp. 733760.
[12] Jean Marie Dumont, Nikita Sidorov, and Alain Thomas, Number of representations
related to a linear recurrent basis, Acta Arith., 88 (1999), pp. 371396.
el Krikorian, Exponential growth of product of matrices in
[13] Bassam Fayad and Rapha
SL(2, R), Nonlinearity, 21 (2008), pp. 319323.
[14] Nicola Guglielmi and Vladimir Protasov, Exact computation of joint spectral characteristics of linear operators, Found. Comput. Math., 13 (2013), pp. 3797.
[15] N. Guglielmi, F. Wirth, and M. Zennaro, Complex polytope extremality results for families
of matrices, SIAM J. Matrix Anal. Appl., 27 (2005), pp. 721743.
[16] Leonid Gurvits, Stability of discrete linear inclusion, Linear Algebra Appl., 231 (1995),
pp. 4785.
, Stability of linear inclusions Part 2. NECI technical report TR96-173, 1996.
[17]
[18] Kevin G. Hare, Ian D. Morris, Nikita Sidorov, and Jacques Theys, An explicit counterexample to the Lagarias-Wang finiteness conjecture, Adv. Math., 226 (2011), pp. 4667
4701.
[19] Christopher Heil and Gilbert Strang, Continuity of the joint spectral radius: application
to wavelets, in Linear algebra for signal processing (Minneapolis, MN, 1992), vol. 69 of IMA
Vol. Math. Appl., Springer, New York, 1995, pp. 5161.
[20] Rapha
el M. Jungers, The joint spectral radius: theory and applications, vol. 385 of Lecture
Notes in Control and Information Sciences, Springer-Verlag, Berlin, 2009.
[21] Aleksandr Ya. Khinchin, Continued fractions, Dover Publications, Inc., Mineola, NY, russian ed., 1997. With a preface by B. V. Gnedenko, Reprint of the 1964 translation.
[22] Jeffrey C. Lagarias and Yang Wang, The finiteness conjecture for the generalized spectral
radius of a set of matrices, Linear Algebra Appl., 214 (1995), pp. 1742.
[23] Ian D. Morris, On Falconers formula for the generalised q-dimension of a self-affine measure. preprint, 2015.
[24] Vladimir Yu. Protasov, Asymptotics of the partition function, Mat. Sb., 191 (2000), pp. 65
98.
[25] Gian-Carlo Rota, Gian-Carlo Rota on analysis and probability, Contemporary Mathematicians, Birkh
auser Boston, Inc., Boston, MA, 2003. Selected papers and commentaries, Edited
by Jean Dhombres, Joseph P. S. Kung and Norton Starr.
[26] Gian-Carlo Rota and Gilbert Strang, A note on the joint spectral radius, Nederl. Akad.
Wetensch. Proc. Ser. A 63 = Indag. Math., 22 (1960), pp. 379381.
[27] John N. Tsitsiklis and Vincent D. Blondel, The Lyapunov exponent and joint spectral
radius of pairs of matrices are hardwhen not impossibleto compute and to approximate,
Math. Control Signals Systems, 10 (1997), pp. 3140.