gibbs.if.usp.brgibbs.if.usp.br/~marchett/fismat3/asymptotics_appendices.pdfgibbs.if.usp.br
TRANSCRIPT
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Publishers’ page
i
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Publishers’ page
ii
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Publishers’ page
iii
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Publishers’ page
iv
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Preface: a description of contents
Decay of various quantities (return or survival probability, correlation functions)
in time are the basis of a multitude of important and interesting phenomena in
quantum physics, ranging from spectral properties, resonances, return and approach
to equilibrium, to dynamical stability properties and irreversibility and the “arrow of
time”. This monograph is devoted to a clear and precise, yet (hopefully) pedagogical
account of the associated concepts and methods. It is aimed at graduate students
and researchers in the fields of mathematical physics and mathematics.
v
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
vi Asymptotic Decay in Quantum Physics
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Contents
Preface: a description of contents v
Appendix A A survey of classical ergodic theory 1
Appendix B Transfer matrix, Prufer variables and spectral analysis
of sparse models 23
Appendix C Symmetric Cantor sets and related subjects 41
Bibliography 49
vii
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Appendix A
A survey of classical ergodic theory
In this appendix we discuss some of the elements of classical ergodic theory, together
with some rudiments of number theory, which are complementary to the discussion
in chapters ?? and ??. For complete and readable expositions, see [36], [54] and [7],
and for a discussion in the spirit of this appendix, which emphasizes the relationship
with number theory and probability, see [25].
The ergodic problems of classical mechanics have been covered in various clas-
sical monographs, notably [2] and [50]. A lucid treatment of statistical mechanics
aspects may be found in [50] and the still pedagogically valuable [34].
To Francois Viete (1540-1603), who is by some considered as the “father of
algebra” (see the fascinating history in the article by J. J. O’Connor and E. F.
Robertson in Wikipedia) is attributed the formula (Vieta’s formula):
sinx
x=
∞∏k=1
cos(2−kx) (A.1)
This formula is used twice in this book: in the beginning of chapter ?? as the
basis of the most elementary example that the convolution of two s.c. measures
may be a.c., and in the approach to equilibrium of the spin model in chapter ??.
It is therefore quite adequate to use it, following [25], in order to provide a first
illustration of the connection between number theory, analysis and probability.
Given a real number x ∈ [0, 1), and an integer q ≥ 2, there exists one and only
one representation of x in the basis q given by the expansion
x =
∞∑k=1
εkqk
=ε1q
+ε2q2
+ · · ·+ εkqk
+ · · · (A.2)
where εk are integers such that
0 ≤ εk < q for k = 1, 2, . . . (A.3)
and
εk < q − 1 for any infinity of k’s (A.4)
If q = 10, (A.2) coincides with the decimal representation
x = 0.ε1ε2 · · · εk · · · (A.5)
1
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
2 Asymptotic Decay in Quantum Physics
We shall use the same notation to represent x in basis q, with q 6= 10. Except when
explicitly mentioned, we shall use q = 2. In this case, (A.2), (A.3,b) establishes a
1-1 correspondence between a real number x ∈ [0, 1] and a binary sequence
ω = (εj)j≥1 with εj ∈ 0, 1 (A.6)
with an infinity of digits εj equal to zero: note that the latter ((A.4) in general)
chooses one among the two possible representations, e.g.,
3/4 = 1/2 + 1/22 + 0/23 + 0/24 + · · · = 1/2 + 0/22 + 1/23 + 1/24 + · · ·
Note also that in (A.6), εjj≥1 varies with x, and we denote this by
x⇔ εj(x)j≥1 (A.7)
For example,
ε1(x) =
0 if x ∈ [0, 1/2) ,
1 if x ∈ [1/2, 1) ,(A.8)
ε2(x) =
0 if x ∈ [0, 1/4) ∪ [1/2, 3/4) ,
1 if x ∈ [1/4, 1/2) ∪ [3/4, 1) ,(A.9)
etc. We have
Definition A.1. For each integer k, the k–th Rademacher function is given by
rk(x) = 1− 2εk(x) (A.10)
For example, from (A.8,b)
r1(x) =
1 if x ∈ [0, 1/2) ,
−1 if x ∈ [1/2, 1) ,(A.11)
r2(x) =
1 if x ∈ [0, 1/4) ∪ [1/2, 3/4) ,
−1 if x ∈ [1/4, 1/2) ∪ [3/4, 1) ,(A.12)
etc. Whatever n, rn(x) is constant and equal to ±1 in the intervals
In,j ≡ [j/2n, (j + 1)/2n) (A.13)
with j = 0, 1, . . . , 2n − 1, alternating sign when x varies from one interval to the
next, with rn(0) = 1.
(A.10) may thus be expressed as
rj(x) = sign(sin 2jπx) (A.14)
where
sign(y) ≡
1 if y > 0 ,
0 if y = 0 ,
−1 if y < 0 ,
. (A.15)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 3
From the definition or (A.14), rj(x) changes sign an odd number of times in each
interval Ii,k = [k/2i, (k + 1)/2i) of continuity of ri(x) if i < j (look at (A.11,b)). It
follows that ∫ 1
0
ri(x)rj(x)dx = δij (A.16)
The Rademacher functions do not form a complete orthonormal basis of L2(0, 1),
because they may be expressed as a particular linear combination of the so-called
Haar functions, which do form a complete orthonormal basis, see [26].
We now use (A.2) with q = 2 and (A.10) to write
1− 2x = 1− 2
∞∑k=1
εk/2k =
∞∑k=1
1− 2εk2k
=
∞∑k=1
rk(x)
2k
On the other hand,∫ 1
0
exp (iξ(1− 2x)) dx =exp(iξ)(1− exp(−2iξ))
2iξ=
sin ξ
ξ
and therefore
sin ξ
ξ=
∫ 1
0
dx
∞∏k=1
exp(iξ2−jrj(x)
)(A.17)
Problem A.1. Use the identity sinα = 2 sin(α/2) cos(α/2) successively to obtain
sin ξ = 2n sin(ξ/2n) cos(ξ/2n) cos(ξ/2n−1) · · · cos(ξ/2)
and then use the above to obtain Vieta’s formula (A.1).
Problem A.2. Prove that∫ 1
0
exp[iξ2−krk(x)]dx = cos(x/2k) . (A.18)
Equations (A.1), (A.17) and (A.18) suggest the validity of
Proposition A.1.∫ 1
0
∞∏k=1
exp[iξckrk(x)]dx =
∞∏k=1
∫ 1
0
exp[iξckrk(x)]dx . (A.19)
Proposition A.1 with ck = 2−k is Vieta’s formula (A.1) written in terms of the
orthogonal (see (A.16)) Rademacher functions.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
4 Asymptotic Decay in Quantum Physics
Probabilistic interpretation We shall provide two proofs of proposition A.1.
Initially, let us remark that the linear combination Ψ(x) =∑nk=1 ckrk(x) is a con-
stant function in each interval In,j = [j/2n, (j+1)/2n), for all j = 0, 1, . . . , 2n−1, of
length |In,j | = 2−n. Observe that the intervals in which rk : k < n assume constant
values ±1 are compatible with the intervals In,j and contain them. Using (A.2) and
(A.10), we may establish a 1-1 relation between each value of j which labels the
interval In,j and the vector ~σ = (σ1, . . . , σn) of components σm ∈ −1, 1:
j/2n =1− σ1
2
1
2+
1− σ2
2
1
22+ · · ·+ 1− σn
2
1
2n. (A.20)
Problem A.3. Derive the dependence of j in ~σ = ~σ(j) given by (A.20) explicitly.
If x ∈ In,j , Ψ(x) is a constant equal to
Ψ(x) = Ψ(j) =
n∑k=1
ckσk = ~c · ~σ
and therefore, for F : R→ R,∫ 1
0
F (Ψ(x))dx =
2n−1∑j=0
|In,j |F (Ψ(j)) =
=1
2n
∑~σ∈(−1,1)n
F (~c · ~σ)
where in the second line we used the relation j ⇔ σ shown in problem A.3 and
(A.20). Substituting F (Ψ) = exp(iξΨ), we obtain∫ 1
0
exp
(iξ
n∑k=1
ckrk(x)
)dx =
1
2n
∑~σ∈(−1,1)n
exp(iξ~c · ~σ)
=
n∏k=1
1
2
∑σ∈−1,1
exp(iξckσ)
=
n∏k=1
cos(ξck) =
n∏k=1
∫ 1
0
exp (iξckrk(x)) dx(A.21)
by (A.18). We conclude the proof of proposition A.1 upon setting ck = 2−k and
then taking the limit n→∞.
q.e.d.
We now introduce the probability space (Ω,B, µ) with Ω = [0, 1), B the Borel
algebra generated by the subintervals of [0, 1), and µ Lebesgue measure on [0, 1).
A random variable (r.v.) f : Ω→ R is a measurable function, i.e., s.t. f−1(A) ∈ Bfor any Borel set A. We see that r1(x), . . . , rn(x) are r.v. defined on the space
([0, 1),B, µ). The probability that a r.v. f assumes some value in the interval
I = [a, b) ⊂ R is the Lebesgue measure of the set
f−1(I) = x ∈ [0, 1) : a ≤ f(x) < b= P(a ≤ f < b)
= µ(x ∈ [0, 1) : a ≤ f(x) < b) = |f−1(I)|
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 5
Since the Rademacher functions rk(x) assume only the values ±1,
P(rk = ±1) = µ(x ∈ [0, 1) : rk(x) = ±1) =
∫ 1
0
1∓ rk(x)
2dx = 1/2
Vieta’s formula (A.1) is satisfied due to the following property of the Rademacher
functions. Consider the event Ek that rk(x) assumes the value σk ∈ −1, 1 for
each k = 1, · · · , n individually:
Ek = x ∈ [0, 1) : rk(x) = σk
and the event E(n) that rk(x) assumes the value σk for k = 1, · · · , n jointly:
E(n) = x ∈ [0, 1) : rk(x) = σk for all k = 1, . . . , n =
n⋂k=1
Ek
It follows from the definition of the rk that the probability of these events fulfills
the relation
P(E(n)) = µ(x ∈ [0, 1) : rk(x) = σk, k = 1, · · · , n
=
n∏k=1
µ(x ∈ [0, 1) : rk(x) = σk) =
n∏k=1
P(Ek) =1
2n(A.22)
As an example, in the case n = 3, we may verify by inspection that the Lebesgue
measure of the set of x s.t. r1(x) = 1, r2(x) = −1, r3(x) = −1 is the length of the
interval I3,3 = [3/8, 1/2) (see (??) et ff.):
1/8 = |I3,3| = |I1,0| × |I2,1 ∪ I2,3| × |I3,1 ∪ I3,3 ∪ I3,5 ∪ I3,7| = 1/2× 1/2× 1/2
In chapter ?? subsection ?? we have actually shown (A.22) in general.
(A.22) shows that r1, · · · , rn are (by definition, see, e.g., [6]) independent
r.v., and therefore the integrand on the r.h.s. of (A.21) may be written as the
expectation E (in the language of probability, see again [6]):
E exp
(iξ
n∑k=1
ckrk
)=
∫ 1
0
exp
(iξ
n∑k=1
ckrk(x)
)dx
=∑
~σ∈(−1,1)nexp
(iξ
n∑k=1
ckrk
)µ(x : rk(x) = σk, k = 1, . . . , n)
=∑
~σ∈(−1,1)nexp(iξ~c · ~σ)
n∏k=1
µ(x : rk(x) = σk)
=
n∏k=1
E exp(iξckrk)
which equals the r.h.s. of (A.21), concluding the second proof of proposition A.1.
q.e.d.
By (??), definition A.1 and (A.22) provide an alternative wording (in terms of
Rademacher functions) of (??).
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
6 Asymptotic Decay in Quantum Physics
Proposition A.1 expresses a connection between probability theory and elemen-
tary number theory. We now come to the interplay between ergodic theory and
dynamical systems; some connection between ergodic theory and number theory
appears in theorem A.5, and the final paragraphs of the appendix are reserved to
some further elementary notions of number theory which were used in chapter ??.
Given a probability space (Ω,B, µ), and a group Φt; t ∈ R of transformations
with Φt+s = ΦtΦs,Φ0 = I - a flow (resp. semiflow), which preserves the measure
µ:
µ(Φt(A)) = µ(A) for all t ∈ R and A ∈ B
(resp. for all t ∈ R+), we say that A ⊂ B is invariant w.r.t Φt iff Φt(A) = A
for all t ∈ R (resp. t ∈ R+). In case of discrete times, we define T = Φ1, Tn =
Φn1 = Φ1 · · · Φ1 ≡ Φn where the indicates composition, and the · · · refer
to composition n times. In this case, A is invariant if T (A) = A = T−1(A) if T
is an automorphism, i.e., a bijection T : B → B such that µ(A) = µ(T (A)) =
µ(T−1(A)). If T is an endomorphism, i.e., a surjection s.t. for all A ∈ B,
T−1(A) ∈ B and µ(A) = µ(T−1(A)), A is said to be invariant if T−1(A) = A.
Definition A.2. A dynamical system is a group or semigroup of transformations
which preserves the measure µ of a probability space (Ω,B, µ). It is said to be
ergodic if for all A invariant, either µ(A) = 0 or µ(A) = 1.
Sometimes ergodic dynamical systems are also referred to as quasiergodic.
This terminology originates from the following basic theorem:
Theorem A.1 (Birkhoff’s ergodic theorem). If f ∈ L1(Ω, dµ), the “time av-
erage”
limT→∞
1
T
∫ T
0
dtf(Φt(x)) ≡ f(x) (A.23)
exists for almost every x with respect to µ. Moreover, if the system is ergodic, f is
a.e. a constant and thus equal to its “space average”
f(x) =
∫Ω
fdµ (A.24)
a.e. with respect to µ.
For a relatively simple proof, see [54], p. 33. If the system is ergodic, each
integral of motion (i.e., a function g s.t. g (Φt(x)) = g(x)) is constant a.e.. Indeed,
let Ca = x : g(x) < a, then Ca is invariant for all a ∈ R. Hence g(x) < a a.e. or
g(x) ≥ a a.e. Since this is true for all a, g(x) = const. a.e.. This explains the last
assertion of Birkhoff’ theorem. In the discrete case, the time averages are
limN→∞
1
N
N−1∑i=0
f(T ix) (A.25)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 7
Putting f = χA, the characteristic function of A ⊂ B, the time average, i.e., the
mean time spent by the system in region A, equals µ(A).
It may also happen that (A.25) exists for x in some manifold M , and f : M → Ra continuous function, and is independent of the point x in some set B ⊂ M : call
it Ex(f). Then f → Ex(f) defines for all x ∈ B a nonnegative linear operator
on the space C0(M,R) of real continuous functions from M to R, which, by the
Riesz theorem ?? may be written in the form Ex(f) =∫fdµ for a Borel measure
µ on M , for all f ∈ C0(M,R), and for all B ⊂ M , which we assume is of positive
Lebesgue measure (on M). B is the called the ergodic basin of µ, and µ a SRB
measure (Sinai–Ruelle–Bowen measures, see [49], [5]). Note that such a measure
may be “physically observed” by computing time–averages of continuous functions
for randomly chosen points x ∈ M (positive probability of getting x ∈ B). SRB
measures are expected to exist in great generality and are of foremost importance
in nonequilibrium statistical mechanics (see [12] Chapter 13 for a concise but lucid
exposition and further references). In the following exposition, the measure µ may
be thought to stand for a SRB measure.
Definition A.3. An automorphism (or flow) is mixing iff, for all f, g ∈ L2(Ω, dµ),
limn→±∞
∫Ω
f(Tnx)g(x)dµ =
∫Ω
fdµ
∫Ω
gdµ (A.26)
in the case of an automorphism, or
limt→±∞
∫Ω
f(Φt(x))g(x)dµ =
∫Ω
fdµ
∫Ω
gdµ (A.27)
in the case of a flow. Setting f = χA and g = χB , A, B both in B, we obtain
limn→±∞
µ(Tn(A) ∩B) = µ(A)µ(B) (A.28)
which is sometimes used as definition of mixing. In the case of an endomorphism,
the definition is
limn→±∞
µ(T−n(A) ∩B) = µ(A)µ(B) (A.29)
Conversely, one may pass from (A.28) to (A.26) using the fact that finite linear
combinations of characteristic functions are dense in L2(Ω, dµ).
The best “physical” interpretation of the mixing property, e.g., (A.28), has been
given in the famous example of Arnold and Avez [2], a glass containing initially
20 per cent of rum and 80 per cent of coca-cola. Initially, the rum is in region A,
which is disjoint of a region B, fully occupied by coca-cola. After n “mixings” the
percentage of rum in B will be µ(Tn(A) ∩B)/µ(B). In this situation, one expects
that for n→∞ all parts of the glass will contain approximately 20 per cent of the
rum, which is the content of (A.28).
It is very useful to express the mixing property in other, occasionaly more man-
ageable forms. For this purpose, we define:
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
8 Asymptotic Decay in Quantum Physics
Definition A.4. The Koopman operator Ut is defined by
(Utf)(x) = f(Φt(x)) (A.30)
for flows, and similarly in the other cases.
Clearly, for each t, Ut is isometric, i.e., preserves the norm in L2(Ω, dµ):
||Utf ||2 =
∫Ω
|f(Φt(x))|2dµ(x) =
∫Ω
|f(Φt(x))|2dµ(Φt(x)) =
∫Ω
|f(x)|2dµ(x) = ||f ||2
and it is a bijection: each g ∈ L2 may be written as some Uf : f(x) = g(Φ−1t (x)).
Therefore, the Koopman operator is unitary. (Φ, µ) is ergodic iff 1 is a simple
eigenvector of U , because f is invariant under Φt if Utf = f : Φ is ergodic iff the
invariant functions are constant a.e.. Since the latter are scalar multiples of one
another, Φ is ergodic iff the subspace of solutions of Uf = f has dimension one.
Related to this, and basic to ergodic theory, is the following mean ergodic theorem
of von Neumann:
Theorem A.2. Let U be a unitary operator on a Hilbert space H, and P be the
orthogonal projection onto Ψ : Ψ ∈ H and UΨ = Ψ. Then, for any f ∈ H,
limn→∞
1
n
n−1∑i=0
U if = Pf . (A.31)
Problem A.4. Prove theorem A.2, using the decomposition
H = Ran(I− U)⊕Ker(I− U†)Ker(I− U†) = Ker(I− U)
Alternatively, see [42], Theorem II.11.
Together with our remarks preceding theorem A.2, we have shown that T is
ergodic iff, ∀f, g ∈ L2(Ω, dµ),
limn→∞
1
n
n−1∑i=0
(U if, g) = (f, 1)(1, g) (A.32)
The assertion in the form corresponding to (A.26) is
limn→∞
1
n
n−1∑i=0
∫f(T ix)g(x)dµ =
∫fdµ
∫gdµ (A.33)
or, in the continuum case,
limT→∞
1
T
∫ T
0
∫f(φt(x))g(x)dµ(x) =
∫fdµ
∫gdµ (A.34)
We now come to the important problem of relating the spectrum of U on 1⊥to ergodic properties. We need a final definition:
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 9
Definition A.5. We say that an endomorphism T is weak-mixing iff
limn→∞
1
n
n−1∑i=0
|µ(T−i(A) ∩B)− µ(A)µ(B)| = 0
with analogous definitions in the other cases (automorphism, flow). As usual, equiv-
alent definitions similarly to (A.26) and (A.27) follow, and therefore T is weakly
mixing if, ∀f, g ∈ L2(Ω, dµ),
limn→∞
1
n
n−1∑i=0
|(U if, g)− (f, 1)(1, g)| = 0
We now have:
Theorem A.3. If T is an automorphism and U the corresponding Koopman oper-
ator, T is weak mixing iff 1 is the only eigenvalue of U , and on 1⊥ the spectrum
of U is continuous.
Problem A.5. Prove theorem A.3 using the spectral theorem for U and Fubini’s
theorem, in a very similar fashion to the proof of Wiener’s theorem ??.
For a mixing automorphism, (A.26) may be written in terms of the Koopman
operator U as
limt→±∞
(Utf, g) = (f, 1)(1, g) for all f, g ∈ L2(Ω, dµ) (A.35)
As a consequence of (A.35), the Riemann-Lebesgue lemma ?? and the spectral
theorem we have (see [42], Theorem VII.15, p. 241):
Theorem A.4. Let T be an automorphism and U the associated Koopman operator.
If U has purely a.c. spectrum on 1⊥, then T is mixing.
We have mentioned previously the “physical” significance of the mixing property.
We now inquire into its relation with the ergodic property and the related signif-
icance in a real physical framework, that of statistical mechanics. It is clear from
(A.28) or (A.29) that mixing implies ergodicity, for, take A invariant, T−1(A) = A,
and let B = A in (A.29), for instance, then the latter yields µ(A) = µ(A)2 and
therefore µ(A) = 0 or µ(A) = 1, i.e., the system is ergodic. The converse does not
hold, however, as we now show. Let M be the circle z ∈ C : |z| = 1, µ Lebesgue
measure over M , Φ the translation Φ(z) = θz with θ = exp(2πiω) where ω ∈ R.
Consider the orthonormal basis of L2(M,dµ) given by zp : p ∈ Z; we have
(Unzp, zp) = θpn (A.36)
Φ is ergodic iff 1 is a simple eigenvalue of U , i.e., iff pω is not in Z if p 6= 0, i.e.,
if ω is irrational. However, Φ is not mixing under the latter condition, because,
taking f = g = zp in (A.35), we obtain, for p 6= 0, limn→±∞(Unzp, zp) = 0 as the
condition for mixing, but, by (A.36), it is not satisfied because limn→±∞ θpn does
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
10 Asymptotic Decay in Quantum Physics
not exist. These results extend immediately to higher dimensional tori and show
that ergodicity does not imply “approach to equilibrium”: the translations of the
torus do not deform a region A, they are such that the intersections of Tn(A) with
B are alternately empty or of positive measure. We now show that mixing does
imply approach to equilibrium in the framework of classical statistical mechanics.
Consider the dynamical system generated by the motion of N matter points (gas
molecules) in a fixed volume V , and let Γ = x ≡ (qi, pi), i = 1, . . . , 3N be the
corresponding phase space, with qi denoting the generalized coordinates, and pi the
momenta. If the system is isolated (fixed total energy E), Γ will be compact and
there exists [51] an invariant measure µ which may be interpreted as the distribution
in thermodynamic equilibrium. Starting from initial conditions in a certain volume
V1, the set of admissible coordinates and momenta of the molecules forms a subset
A ⊂ Γ. The formula
µ0(B) =µ(A ∩B)
µ(A)(A.37)
defines a initial distribution (which is not the equilibrium distribution), interpreted
as conditional distribution relatively to the system’s known initial condition. We
may now relate µ0 to the knowledge of the state of the system at time t by the
measure µt defined by
µt(B) = µ0(Tt(B)) (A.38)
where T is the flow which leaves µ invariant. In the given example,
µ(F ) = µE(F ) =
∫F
δ(H(x)− E)dx (A.39)
the microcanonical Gibbs measure where H(x) is the Hamiltonian describing
the system and (A.38) is Liouville’s theorem. Suppose, now, that µ0 is a.c. with
respect to µ, i.e., there exists ρ0 ∈ L1(M = Γ, dµ) such that
dµ0 = ρ0(x)dµ . (A.40)
Then,
µt(B) =
∫M
χB(Ttx)dµ0 =
∫M
χB(Ttx)ρ0(x)dµ (A.41)
If the system is mixing, we obtain from (A.35), as t→∞,
limt→∞
µt(B) =
∫M
χBdµ
∫M
ρ0dµ = µ(B)1 = µ(B) (A.42)
which means that, whatever the initial distribution µ0, normalized and a.c. w.r.t.
µ, the time–translates µt of µ0 under Tt converge, for t→∞, to the equilibrium dis-
tribution. (A.42) may be taken as the definition of the approach to equilibrium .
It means that mixing systems are “memoryless”, i.e., they posess a stochastic char-
acter which justifies equilibrium statistical mechanics. The microscopic mechanism
of this process of loss of memory is the sensitive (exponential) dependence on initial
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 11
conditions (later on precisely defined) produced by “defocalizing shocks” between
the gas molecules, first pointed out by Krylov, see [50] and references given there.
It may also be felt in the exponential rate at which the mixing condition takes
place, a topic to which we also return later.
We may picture the gas molecules as a system of hard spheres enclosed in a cube
with perfectly reflecting walls or periodic b.c.. This is supposed to be a K–system
(see [54], definition 4.7, p. 101). Rising still one step in this so–called ergodic
hierarchy (see [34]), we come to Bernoulli systems , such as the one we presently
introduce along the lines of [2].
Define Tr by Tr : X → Y , where X = [0, 1] and Y = [0, 1], by
Trx = fr(rx) ≡ rx mod1 (A.43)
Note that this mapping is not 1-1, it is an endomorphism. In fact, if r = 2, we
see that the inverse image of a point x is x/2 or (x + 1)/2. Tr is called the r–adic
transformation and leaves Lebesgue measure µ invariant. Indeed, if
A = [m
rn,m+ 1
rn] for m ∈ [0, rn − 1] we have
T−1r (A) =
r−1⋃s=0
[m+ s
rn+1,m+ s+ 1
rn+1
]and
µ(A) = r−n = µ(T−1r (A)) = rr−(n+1)
Consider the case r = 2. T2 has a simple alternative description: writing x =
.ε1ε2 · · · , i.e., x =
∞∑n=1
εn2n
with εn ∈ 0, 1 as in (A.5), then T2x = .ε2ε3 · · · , Tn2 x =
.εn+1εn+2 · · · (check!) and therefore T2 is also known as the one-sided shift. Let
M = [0, 1]Z be the probability space. The sigma–algebra of subsets of M generated
by the cylinders C = (xn)n∈Z : xi = εis , s = 1, . . . , k, where is is an increasing
finite sequence of integers, and εis ∈ 0, 1 coincides with the Borel sigma–algebra
of M , equipped with the product topology. If µ0 is a probability measure s.t.
µ0(0) = p1 and µ0(1) = p2, with p1 + p2 = 1, p1 > 0, p2 > 0, we take as µ the
product measure; in particular, µ(C) =∏ks=1 pis . The mixing property (A.29) need
only be verified for A, B cylinders. But, for n sufficiently large, T−n2 (A) and B are
disjoint and, since µ(T−n2 (A)) = µ(A), we have that µ(T−n2 (A) ∩ B) = µ(A)µ(B)
for n sufficiently large. Thus we have proven that T2 is mixing.
Take, now, as our dynamical system the set of autonomous n first order difference
equations
~xn+1 = T (~xn) (A.44)
where ~x ≡ (x1, . . . , xn), T (~x) ≡ (T1(~x), . . . , Tn(~x)). The mapping T : M → Rn,
where M is a manifold (which we often identify with an open set in Rn) is a
vector field: we assume it is an automorphism. Let J(~x) denote the Jacobian
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
12 Asymptotic Decay in Quantum Physics
matrix Ji,j(~x) =∂Ti∂xj
(~x) associated with (A.44). Defining Tn(~x) ≡ T (T · · ·T ) (~x)
as before and using the chain rule we have
J(Tn)(~x) = J(Tn−1)(~x) · · · J(T )(~x)J(~x) (A.45)
Let µ denote the measure invariant under the automorphism T , which completes
the definition of the dynamical system. We shall assume it has compact support
and that it is ergodic under the automorphism T . Then, Oseledec’s multiplicative
ergodic theorem ([39], see also [15]) implies the existence, for almost all ~x with
respect to µ, of the limit
limn→∞
[J(Tn)†(~x)J(Tn)(~x)] ≡ Λ~x (A.46)
The positive matrix Λ~x has eigenvalues λ1 > λ2 > · · · .
Definition A.6. λi, with multiplicity mi, is called a Lyapunov exponent.
By ergodicity of µ it follows that these exponents are constants for almost all
values of ~x (for SRB measures this will hold for almost all ~x in the ergodic basin
of µ). If Ei~x is the subspace of Rn corresponding to the eigenvalues smaller than
exp(λi), Rn = E1~x ⊃ E2
~x ⊃ · · · , one can show that for almost all ~x with respect to
µ,
limn→∞
log ||J(Tn(~x))u||n
= λi for i = 1, 2, . . . (A.47)
if u ∈ Ei~x\Ei+1~x . In (A.47) the norm is the Euclidean norm in Rn. In particular, for
all u ∈ Rn\E2~x, the limit equals the largest Lyapunov exponent λ1. We shall see
explicit examples shortly. But note that existence of a positive largest Lyapunov
exponent is a precise definition of what is meant by sensitive or exponential depen-
dence on initial conditions, which is an important element of chaotic behavior: see
[15] for a detailed discussion. It is however easy to illustrate the phenomenon on
the basis of the example T2: consider two x1, x2 both in [0, 1], close in the sense that
the first n digits are identical. By the interpretation of T2 as one-side shift, it fol-
lows immediately that Tn2 x1 and Tn2 x2 differ already in the first digit: an initially
exponentially small difference 2−n is magnified by the evolution to one of order
O(1). This is the best way to approach exponential sensitivity to initial conditions
for quantum systems: see chapter ??.
It turns out, however, that T2 has more than sensitive dependence on initial
conditions: the sequence of iterations Tn2 x0 has, for almost every x0, the same
random character as the sequence of successive tossings of a coin; this is what
characterizes Bernoulli systems and places them at the top of the ergodic hierarchy:
Theorem A.5 (Borel’s theorem on normal numbers). For a.e. x0 ∈ [0, 1],
the n–th digit in its binary expansion has relative frequency 1/2, i.e., almost all x0
is normal to base 2.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 13
Proof. Let f : [0, 1]→ [0, 1] be defined by
f(x) =
0 if x < 1/2 ,
1 if x ≥ 1/2 ,
Since T2 is ergodic (being mixing),
limn→∞
1
n
n−1∑i=0
f(T i2x) =
∫ 1
0
f(x)dx = 1/2
by Birkhoff’s theorem A.1, for a.e x ∈ [0, 1], but
f(T i2x) =
0 if εi = 0 ,
1 if εi = 1 ,
which proves the theorem.
q.e.d.
Problem A.6. Which numbers comprise the set of zero Lebesgue measure in theo-
rem A.5?
The two-sided shift (i.e. in both directions) is a.e. an automorphism of [0, 1]→[0, 1]: the baker’s map [32]. It is a caricature of the Smale horseshoe, which describes
the local behavior of Hamiltonian systems near homoclinic points (which we shall
define shortly): see [2, 9]. We now introduce it because it plays an important
illustrative role in chapter ??.
Let X be the unit square X = [0, 1] × [0, 1]. The Borel sigma algebra B is
now generated by all possible rectangles of the form [0, a] × [0, b], and the Borel
measure µ is the unique measure on B such that µ([0, a]× [0, b]) = ab. We define a
transformation S : X → X - the baker’s transformation by
S(x, y) =
(2x, y/2) if 0 ≤ x < 1/2 and 0 ≤ y ≤ 1 ,
(2x− 1, y/2 + 1/2) if 1/2 ≤ x ≤ 1 and 0 ≤ y ≤ 1 ,(A.48)
The reader might wish to look at fig. 5 of [34] or verify for himself the effect of S
on some figure symmetrically disposed w.r.t. the line x = 1/2: as in the kneading
of a piece of dough, one first squashes the unit square to half its original height and
twice its original width, and then cuts the resulting in half and moves the right half
of the rectangle above the left. Repeating this operation several times, the result is
to very quickly scramble, or mix, various parts of the original figure: this provides
a very colourful illustration of the mixing property!
The Koopman operator, defined by (A.30), is unitary from L2(Ω, dµ) to itself,
and leads (for automorphisms) to a characterization of the mixing property by
(A.35). It is interesting, from the point of view of approach to equilibrium, to con-
sider an alternative description. Consider, to fix ideas, the case of endomorphisms,
and define the Koopman operator (we shall keep the same name for simplicity) now
as an operator from L∞(Ω, dµ) to L∞(Ω, dµ) by (see [32], p. 42):
(Uf)(x) = f(Tx) (A.49)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
14 Asymptotic Decay in Quantum Physics
It is well–defined because f1(x) = f2(x) a.e. implies f1(Tx) = f2(Tx) a.e. by the
definition of endomorphism. It has the property
||Uf ||∞ ≤ ||f ||∞ (A.50)
If f ∈ L1, the functional
g ∈ L∞ → (f, Ug) (A.51)
defines a continuous linear functional on L∞, with
|(f, Ug)| ≤ ||f ||1||g||∞ (A.52)
We may thus define a bounded linear operator - the Ruelle-Perron-Frobenius
operator (see e.g. [45, 32]) P : L1 → L1 by
(Pf, g) = (f, Ug) (A.53)
From (A.53), for g = χA, where A is any Borel subset of Ω, we obtain∫A
(Pf)(x)dµ(x) =
∫T−1(A)
f(x)dµ(x) (A.54)
Conversely, (A.54) defines P uniquely by the Radon-Nikodym theorem, see [32],
p.37. In [32], P is mentioned as being the adjoint of the Koopman operator U
(defined by (A.49)), but the Banach space adjoint (see, e.g., [42], p. 185) is an
operator defined from the dual space (the space of continuous linear functionals on
the given space) to the dual space: in the present case, the dual of L∞ is a huge
space, which contains L1 properly. In fact, by exercise 8 (b) of [42], p. 86, there
exists a bounded linear functional λ on L∞(R) such that λ(f) = f(0) ∀f ∈ C(R),
i.e., the dual space of L∞(R) contains the Dirac measure at the origin. But, as
we shall see, it is absolutely crucial for the developments here and in chapter ??
that P be considered as an operator mapping densities (in contrast to individual
orbits, formally characterized by the evolution of delta measures at some point) to
densities!
The baker transformation is a.e. invertible (it is not invertible on the line y =
1/2): it is only a.e. a diffeomorphism. Taking A = [0, x] × [0, y] we have for
0 ≤ x ≤ 1 and 0 ≤ y < 1/2, T−1(A) = [0, x/2]× [0, 2y] and thus, by (A.54),
(Pf)(x, y) =∂2
∂x∂y
∫ x/2
0
ds
∫ 2y
0
dtf(s, t) = f(x/2, 2y) , if 0 ≤ y < 1/2 (A.55)
and for 1/2 ≤ y ≤ 1 and 0 ≤ x < 1
T−1(A) = [0, x/2]× [0, 1] ∪ [1/2, 1/2 + x/2]× [0, 2y − 1]
and hence, again by (A.54),
(Pf)(x, y) =∂2
∂x∂y
[∫ x/2
0
ds
∫ 1
0
dtf(s, t) +
∫ 1/2+x/2
1/2
ds
∫ 2y−1
0
dtf(s, t)
]= f(1/2 + x/2, 2y − 1) , if 1/2 ≤ y ≤ 1 (A.56)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 15
Note that, by (A.54), the Ruelle–Perron–Frobenius operator maps densities to
densities: a density (function) f any f : f ∈ L1 and f ≥ 0 a.e.. We denote the
class of density functions by D. The positivity preserving character of P , i.e.,
Pf ≥ 0 a.e. if f ≥ 0 a.e. (A.57)
is immediate from the Radon-Nikodym theorem (see, again, [32], p. 37). It leads
to the inequality
||Pf ||1 ≤ ||f ||1 (A.58)
Let f+(x) ≡ max0, f(x) and f−(x) ≡ min0, f(x). We have that (Pf)+ =
(Pf+−Pf−)+ ≤ (Pf)+ and (Pf)− ≤ Pf−. From these inequalities it follows that
|Pf | ≤ P |f |, and it is now simple to show:
Problem A.7. Prove (A.58).
Analogously to problem A.7, it may also be proved that equality in (A.58), i.e.,
||Pf ||1 = ||f ||1, occurs iff Pf+ and Pf− have disjoint supports; in particular,
||Pf ||1 = ||f ||1 if f ≥ 0 a.e. (A.59)
Properties (A.57) and (A.59) define a Markov operator. If P is a Markov operator
and for some f ∈ L1, Pf = f , then f is called a fixed point of P . If Pf = f , then
Pf+ = f+ and Pf− = f− (see [32], proposition 3.1.3). If f ∈ D, and Pf = f , f is
called a stationary density of P . We have the following important theorem ([32],
theorem 4.4.1):
Theorem A.6. T is mixing iff Pnf is weakly convergent to 1 for all f ∈ D, i.e.,
limn→∞
(Pnf, g) = (f, 1)(1, g) ∀f ∈ D and ∀g ∈ L∞ . (A.60)
If P is a Markov operator with stationary density 1, i.e.,
P1 = 1 (A.61)
then Pnf converges weakly to the stationary density: the weak limit is a fixed
point of P and, being unique, it is the function f = 1. This is an approach to
equilibrium in the formerly defined sense.
How does all this apply to the baker’s transformation? Property (A.57) and
hence (A.59) follow from (A.55,b). We now consider the mixing property. The latter
is simplest to see by noting that, if (x0, y0) ∈ [0, 1] × [0, 1], s.t. x0 = .a1a2a3a4 · · ·and y0 = .b1b2b3b4 · · · , by justaposing the representations in the following way:
· · · b4b3b2b1a1a2a3a4 · · · , the baker transformation may be seen to correspond to
the shift of the decimal point to the right - the so–called two–sided shift, see [54],
p.18 (check!). Then mixing follows by the same proof as for the one–sided shift.
Finally, (A.61) also follows from (A.55,b). Thus, by theorem A.6, there is approach
to equilibrium for the baker’s map. Note that f ∈ L1 excludes a delta measure at
some point in the square: this is expected because the baker map is (a.e.) invertible,
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
16 Asymptotic Decay in Quantum Physics
and individual orbits are therefore reversible. We see hereby that approach to
equilibrium in the theory of dynamical systems in the sense defined above depends
on two key features: Q1) One looks only at the evolution of densities, and not
individual orbits: this is a reduced description of the system, similar to Boltzmann’s
approach (see chapter ??); Q2) the initial state (density) is not the (invariant,
stationary) density, which is the uniform distribution on the unit square. We return
to these points in chapter ?? section ??.
We finish this appendix with two subjects: Anosov systems and rate of mixing
(with some remarks on non-uniformly hyperbolic systems) and a few basic results
of number theory used in the main text.
The baker transformation is a prototype of an important class of transforma-
tions, the Anosov systems [1]. Another special example of Anosov system is the
Arnold cat map [2]:
TA(x, y) = (x+ y, x+ 2y) mod 1 (A.62)
The fixed points F of TA are given by (x + y, x + 2y) = (x, y) hence F = (0, 0).
This is an example of a hyperbolic point. In the more general case (A.44), but
now with T a mapping from a complex manifold M to Cn, let ~x0 be an invariant
or equilibrium point of the vector field T :
Definition A.7. A linear vector field on Cn given by ~x→ L~x, where L is a n× nmatrix with complex entries, is said to be hyperbolic iff the spectrum σ(L) of L has
empty intersection with the imaginary axis. An equilibrium point ~x0 of the vector
field T is said to be hyperbolic iff the Jacobian matrix J( ~x0) is a hyperbolic vector
field. By the Jordan decomposition (see [20]), it may be shown that σ(exp(L)) =
exp(σ(L)): hence, ~x0 is hyperbolic iff J( ~x0) does not have any eigenvalue of unit
modulus.
Due to the Hartman–Grobman theorem (see [19]) the phase portrait of a vector
field near a hyperbolic point has the same topological structure as its linearization
at this point: therefore a hyperbolic fixed point of a vector field is unstable. Such
is not the case for elliptic fixed points, for which J( ~x0) has modulus one: their
stability depends on the character of the non-linear terms [19]. The Hartman–
Grobman theorem also guarantees the existence of the so–called local (stable and
unstable) manifolds at any hyperbolic point (see also [50], lecture 17). We now
return to the cat map. The Jacobian matrix(1 1
1 2
)has eigenvectors v± given, respectively, by: 1
1±√
5
2
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 17
corresponding to eigenvalues r± =(3±√
5)/2. The corresponding invariant (local)
manifolds are W±(~0)
= E±~0 with E± generated by v±, with angular coefficients
s and −s−1, with s = (1 +√
5)/2. Since the latter are irrational numbers, these
manifolds cover the torus densely. Because r+ > 1, (0, 0) is a hyperbolic point.
Problem A.8. Identify the sets E1~0
and E2~0
in the definition of Lyapunov exponents
and compute them for the cat map.
The sum of the Lyapunov exponents for an area preserving map, such as the
cat map above, is zero: this follows in a straightforward way from their definition.
In Hamiltonian and other measure preserving maps chaos has its origins in homo-
clinic points [2] defined as points of intersection of the local (stable and unstable)
manifolds associated to a given hyperbolic point, other than the point itself (see also
[9]). The usual proofs of the existence of the “homoclinic tangle” rely on the (often
verified, such as in the cat map) hypothesis that the manifolds intersect transver-
sally, i.e., are not tangential at the homoclinic point (Melnikov’s method, see [21]
for a pedagogic discussion). See also the monograph of Palis and Takens for the
discussion of homoclinic tangencies [40]. In the case of the cat map, considering the
image of a region B after a large number n of iterations of the map TA, it will consist
of the product of rn+ expansions in the unstable direction v+, and rn− contractions in
the stable direction v−, i.e., a highly ”stretched” band in direction v+ (mod 1): for
n large TnAB covers the torus and the mixing property holds (see [32] Example 4.4.3
p. 71 for the proof). The two invariant manifolds intersect in homoclinic points;
the image of a homoclinic point is, by definition, also a homoclinic point, and, thus,
the cat map posesses an infinity of homoclinic points; it is also immediate (check!)
that the manifolds intersect transversally at these points.
Problem A.9. Which points of the unit square yield periodic orbits of the cat map?
What can one say about the asymptotic distribution of their periods?
Hint for the second part: see the discussion in [9] on the Ozorio–Hannay uni-
formity principle.
Anosov systems such as the cat map (A.62) are “uniformly hyperbolic”. For
such systems there is exponential decay of correlations∣∣∣∣∫Ω
f(Tnx)g(x)dµ−∫
Ω
fdµ
∫Ω
gdµ
∣∣∣∣ ≤ Crn for some r < 1 (A.63)
(see [53] for a nice review). Much of the recent work in dynamical systems has been,
however, devoted to systems displaying only a weak form of hyperbolicity. One of
them is that T is expanding in its critical orbit, see theorem 5.1, p. 125, of [53]
or the original early article of Baladi and Viana [4] for a readable account for an
important class of models. Together with the other assumptions, it is proved that
T admits a unique absolutely continuous invariant measure which is ergodic (and
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
18 Asymptotic Decay in Quantum Physics
so is an SRB measure for T ) and exhibits exponential decay of correlations in the
sense of (A.63).
Physically more “realistic” systems remain, however, very difficult to analyse.
A prototype of the latter is the famous standard or Chirikov-Taylor map ([50],
p. 138), which is the classical analogue of the kicked rotor (??), (??): it is a map
of the cylinder R× T to itself, defined by:
zn+1 = zn + k sin(2πφn)
φn+1 = φn + zn+1 mod1 (A.64)
The importance of this map is reviewed by Sinai in [50], p. 138. There exist very
few rigorous results for this map: the beautiful Aubry–Mather theory shows that
for k > 1 the standard map has no invariant curves which may be represented by
a continuous function z = f(φ) (for a simple proof of this, see [50], p. 142): the
structures occurring for k > 1 are fractal objects denominated Cantori. Another
result, much more difficult, due to Duarte [13, 14] proves, very roughly speaking,
the existence of an abundance of of elliptic islands even deep inside the region k > 1
of “hard chaos” in the model (actually this region is conjctured to be k > kcr with kcr
given by Greene’s conjecture, see the discussion in [50]). Duarte’s remarkable result
throws considerable doubt on the validity of exponential decay of correlations (A.63)
in the chaotic region: it may be algebraic instead, see chapter ?? for further remarks.
Estimates on the rate of decay of correlations are among the most challenging open
problems, both in the theory of dynamical systems and in the theory of “quantum
chaos”.
We conclude this appendix with a brief review of some of the most fundamental
aspects of number theory used in the main text, in particular chapter ?? sections ??
and ??. The reader will find a complete and very readable account of these features
in Khinchin’s book [30].
Irrational numbers may be approximated by rationals with arbitrary precision.
For example, π is approximated by the sequence
s/r = 3/1, 31/10, 314/100, 3142/1000, 31416/10000, · · ·which yields the better approximants the greater the value of r. For such decimal
approximations, we have for an irrational number µ:
∣∣∣∣µ− snrn
∣∣∣∣ < 1/rn, with rn =
10n and a similar inequality is valid for any base. These approximants are not,
however, the best possible, because they are contaminated by the intrinsic properties
of the base. We are thus led to search a sequence of approximants µnn≥1 which
is independent of the base: such a sequence is proportionated by the continued
fractions:
µn = a0 +1
a1 +1
a2 + · · ·
(A.65)
The approximants in (A.65):
µn = sn/rn (A.66)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 19
obey the recursion relations
sn = ansn−1 + sn−2 for n ≥ 2
rn = anrn−1 + rn−2 for n ≥ 2 (A.67)
In order to prove (A.67) by induction, suppose it valid up to order (n−1). Defining
s′n−1
r′n−1
= a1 +1
a2 +1
a3 +1
· · ·+ 1
an
we have:
sn = a0s′n−1 + r′n−1
rn = s′n−1 (A.68)
and by the induction hypothesis
s′n−1 = ans′n−2 + s′n−3
r′n−1 = anr′n−2 + r′n−3 (A.69)
Introducing (A.69) into (A.68)
sn = a0(ans′n−2 + s′n−3) + anr
′n−2 + r′n−3
rn = ans′n−2 + s′n−3
or
sn = an(a0s′n−2 + r′n−2) + a0s
′n−3 + r′n−3 = ans
′n−1 + s′n−2
rn = anr′n−1 + r′n−2
Clearly (A.67) is valid for n = 2 and thus the result is demonstrated.
The difference between two approximants µn and µn−1 is
sn−1
rn−1− snrn
=sn−1rn − snrn−1
rnrn−1=
(−1)n
rnrn−1(A.70)
In fact, multiplying the first formula in (A.67) by rn−1, the second one by sn−1 and
subsequently subtracting the resulting first one from the resulting second one, we
obtain
sn−1rn − rn−1sn = −(sn−2rn−1 − rn−2sn−1)
and, since (s−1 = 1, r−1 = 0), r0s−1 − s0r−1 = 1, and therefore it follows that
sn−1rn − rn−1sn = (−1)n
from which (A.70) follows.
Problem A.10. Show that for all n ≥ 2,
sn−2
rn−2− snrn
=(−1)n−1anrnrn−2
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
20 Asymptotic Decay in Quantum Physics
The two results above show that the convergents of even order form an increas-
ing sequence (we assume an > 0 for all n > 1), and those of odd order form a
decreasing sequence. It follows therefore that both sequences have the same limit.
Indeed, since, by (A.67), each convergent of odd order is greater than the immedi-
ately subsequent convergent of even order, it follows that each convergent of odd
order is greater than each convergent of even order, and that the two sequences of
convergents (of odd and even order) tend to the same limit.
The above result culminates (see [30]) in
Theorem A.7 (Hurwitz’s theorem). For any µ, there exist rational approxi-
mants sn/rn such that ∣∣∣∣µ− snrn
∣∣∣∣ ≤ 1√5r2n
. (A.71)
Clearly the approximants converge the quicker the an diverge (for a rational
number, some an = ∞ after a n = n0 finite). As an illustration, we suggest that
the reader show the explicit results for π (precise up to the kn–th digit):
s0/r0 = 3
s1/r1 = 22/7 ' 3.1429 (k1 = 2)
s2/r2 = 333/106 ' 3.1415 (k2 = 3)
s3/r3 = 355/113 ' 3.1515929 (k3 = 6)
(the last result, according to M.Berry, was known to Lao–Tse (604–531 b.C.)). The
slowest convergence, corresponding to the irrational number most poorly approxi-
mated by rationals, corresponds to the golden mean: a0 = a1 = · · · = 1,
1 +11
1 + · · ·
=
√5− 1
2.
This number is important for several reasons: it appears in problem ?? in con-
nection with the P.V. numbers, and is also conjectured to be the rotation number
corresponding to the critical coupling kcr in the standard map, which signals the
disappearance of the last “KAM torus”, i.e., in the standard map the last continuous
invariant curve. It saturates the inequality (A.71) provided by Hurwitz’s theorem.
This last remark brings us to our last topic: diophantine numbers.
Definition A.8. An irrational number µ is said to be diophantine of type σ iff
there exists a γ > 0 such that∣∣∣µ− s
r
∣∣∣ ≥ γ
rσfor all s/r ∈ Q (A.72)
In passing, we note that µ is said to be a Liouville number if it is neither rational
nor diophantine. Alternatively, µ is Liouville iff there exists a sequence sn/rn ∈ Q
such that
∣∣∣∣µ− snrn
∣∣∣∣ < r−nn for all n ≥ 1. By a theorem due to Liouville, every
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
A survey of classical ergodic theory 21
algebraic number (zero of a polynomial with rational coefficients) is diophantine.
On the other hand, the number µ =
∞∑n=1
2−n! is by definition Liouville. Thus there
exist non-algebraic numbers: they are the transcendental numbers. For much more
about this, see [38].
The set of diophantine numbers of type σ has full Lebesgue measure if σ > 2:
for almost all µ ∈ [0, 1] and for all τ > 0, there exists a number K = K(µ, σ = τ+2)
such that ∣∣∣µ− s
r
∣∣∣ ≥ K
r2+τ(A.73)
for all s/r ∈ Q. In order to show (A.73), fix r and s. Consider the set of numbers
µ ∈ [0, 1] such that (A.73) is violated. They form a set of length ≤ 2K
r2+τ, and,
since s ≤ r, the union of all such intervals with the same r has measure ≤ 2K
r1+τ.
Summing over all the r we obtain a set with Lebesgue measure ≤ CK, where
C = 2∑∞r=1 r
−1−τ <∞ since τ > 0. Choosing K arbitrarily small, we may render
the measure of the set of numbers violating (A.73) as small as wished. Thus, the
set of numbers violating (A.73) for all K has zero Lebesgue measure.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
22 Asymptotic Decay in Quantum Physics
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Appendix B
Transfer matrix, Prufer variables andspectral analysis of sparse models
Weyl–Titchmarsh m–function of a Jacobi matrix We may think of a Ja-
cobi matrix
J =
v1 p1 0 0 · · ·p1 v2 p2 0 · · ·0 p2 v3 p3 · · ·0 0 p3 v4 · · ·...
......
.... . .
, (B.1a)
as an operator acting on the space of complex–valued square–summable sequences
u = (un)n≥0, denoted by l2(Z+), and the Schrodinger equation associated with J
reads
((J − zI)u)n = pnun+1 + pn−1un−1 + (vn − z)un = 0 (B.1b)
for n ≥ 1 with p0 ≡ 1 and z ∈ C. A Jacobi matrix Jφ is said to satisfy a φ–boundary
at 0 if
u0 cosφ− u1 sinφ = 0 , (B.1c)
for some φ ∈ [0, π); (B.1a), as an operator, satisfies Dirichlet 0–boundary phase
condition u0 = 0 and we write J = J0. The phase boundary plays an important
role on the characterization of singular part of the spectrum, as we shall see in the
following. We call J an admissible Jacobi matrix if it is of the form (B.1a) and
satisfies 0 < pn ≤ 1 with∑n≥0 p
−1n = ∞ and |vn| ≤ M < ∞. Sooner, we shall
restrict the admissible class to two sparse models dealt in this monography (see ??)
et seq.): (i) pn = 1 ∀n and vn = 0 except for a lacunary subsequence A = (aj)j≥1;
(ii) vn = 0 ∀n and pn = 1 except for a lacunary subsequence A = (aj)j≥1.
The set of all solutions uφ(z) of (B.1b) and (B.1c) form a two–dimensional vector
space whose base: y(z) = uα(z) and w(z) = uα+π/2(z) may be chosen by fixing the
initial conditions
y0 = sinα , y1 = cosα
w0 = cosα , w1 = − sinα (B.2)
23
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
24 Asymptotic Decay in Quantum Physics
for some α ∈ [0, π). For =z 6= 0, there are two alternatives: either we have the
limit–circle (in which case both y(z), w(z) ∈ l2(Z+)) or the limit–point case, for
which there exist exactly one linear independent l2(Z+)–solution. We have
Proposition B.2. If J is an admissible Jacobi matrix, then J is limit–point.
Proof. Let y(z) and w(z) be a base for (B.1b). By Green’s formula (see e.g. (1.20)
of [52])
0 = (w, Jy)n − (Jw, y)n = −Wn[y, w] +W0[y, w] (B.3a)
where Wn[f, g] is the (weighted) Wronskian function:
Wn [f, g] = pn(fngn+1 − fn+1gn) (B.3b)
of two sequences f = (fn)n≥0 and g = (fn)n≥0 at n and (f, g)N =
N∑n=1
fngn is the
inner product restricted to 1, . . . , N. (B.3a) implies that the Wronskian of two
linearly independent (L.I.) solutions of the Schrodinger equation is constant and
does not vanish:
Wn[y, w] = W0[y, w] = p0 (y0w1 − y1w0) = −1 (B.3c)
by (B.2). Together with (B.3b) and the Schwartz inequality, we have
∞∑n=0
1
pn=
∞∑n=0
−1
pnWn[y, w] =
∣∣∣∣∣∞∑n=1
(ynwn+1 − yn+1wn)
∣∣∣∣∣ ≤ 2 ‖y‖ ‖w‖ (B.3d)
where ‖·‖ is the l2(Z+)–norm. If J is admissible, then the l.h.s. of (B.3d) diverges
and necessarily one of the two L. I. solutions: y or w does not belong to l2(Z+),
concluding the proof.
q.e.d
We set α = 0 in the following propositions and refer to chapter 2 and appendix
B of [52] for references and further directions.
Proposition B.3. The l2(Z+)–solution of (B.1b) with =z > 0 can be written as
χ(z) = w(z)−m(z)y(z) (B.4)
where m(z) = m0(z), the Weyl-Titchmarsh m–function related to Jα with α =
0, is defined by
m(z) :=
(e1,
1
J0 − zIe1
)(B.5)
and ej , j ≥ 0 denotes the canonical base of l2(Z+): (ej)i = δij.
Proof. Let χ(z) be the l2(Z+)–solution of (B.1b). Since y(z) is not an eigenvector
of J when =z > 0, it cannot be an l2(Z+)–solution of (B.1b) so, χ(z) is a linear
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 25
combination of y(z) and w(z) of the form (B.4). The Green’s function associated
with J
gij(z) =
(ei,
1
J − zIej
)can be written in terms of the two L.I. solutions of (B.1b) y(z) and χ(z):
gij(z) =
yi(z)χj(z)
−Wj [y, χ]if i < j
yj(z)χi(z)
−Wj [y, χ]if j < i
. (B.6)
Setting i = j = 1 in (B.6), together with Wj [y, χ] = W0 [y, w] = −1 and
y1(z)χ1(z) = m0, the proof is concluded.
q.e.d.
Remark B.1. Since Jφ is a rank–one perturbation of J0:(Jφu
)n
=(J0u
)n
+
δ0,nu0 tanφ, the Weyl–Titchmarsh m–function mα(z), defined by (B.5) with J0
replaced by Jα, satisfies
mα(z) =m(z)
1−m(z) tanα. (B.7)
Definition (B.5) implies that m(z) is holomorphic in C\σ(J) and, since it maps
the upper–half plane H into itself, it is, in addition, a Herglotz function (called
also Pick or Nevanlinna–Pick function).
Problem B.11. Prove the above statements.
Hint. Use the first resolvent equation (J − zI)−1 − (J − ζI)−1= (z −
ζ) (J − zI)−1(J − ζI)−1
for the former.
As a consequence, m(z) admits a unique canonical integral representation (the-
orem I, chapter 2 of [24]):
m(z) = az + b+
∫ ∞−∞
(1
λ− z− λ
λ2 + 1
)dρ(λ) , =z > 0
where a ≥ 0, b ∈ R are constants and dρ(λ) is a Borel–Stieltjes measure such that∫ ∞−∞
dρ(λ)
λ2 + 1<∞ .
It follows from (B.5) that
m(z) = −z−1 +O(z−2)
which implies a = 0. Since the spectrum of a bounded operator is compact, the
spectral measure µ of J is supported in a compact set Σ,
m(z) =
(e1,
1
J0 − zIe1
)= b−
∫Σ
λ
λ2 + 1dρ(λ) +
∫Σ
1
λ− zdρ(λ)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
26 Asymptotic Decay in Quantum Physics
where ∫Σ
1
λ− zdρ(λ) = −z−1
∫Σ
dρ(λ) +O(z−2)
as z →∞, from which we conclude that b−∫
Σ
λ
λ2 + 1dρ(λ) = 0 and
m(z) =
∫ ∞−∞
1
λ− zdρ(λ) . (B.8)
is a Borel transform of the spectral measure µ = dρ.
The Weyl m–function can be approached, as originally has been done by Weyl,
as a sequence limit: m(z) = limN→∞mN (z) of holomorphic functions in C\σ(J0,β).
Here, J0,β is a finite Jacobi matrix satisfying 0–Dirichlet boundary condition at 0
and β–boundary condition at N :
uN cosβ − uN+1 sinβ = 0 . (B.9)
A solution χ(z) = χ(z;N) of Schrodinger equation (B.1b),
χ(z) := w(z)−mN (z)y(z)
satisfies the phase β–boundary condition (B.9) at N iff
mN (z) =wN (z)− ζwN+1(z)
yN (z)− ζyN+1(z), ζ = tanβ . (B.10)
Since the r.h.s. of (B.10) is a linear fractional map L : C −→ C, L =
(aζ + b) / (cζ + d) with ad− bc 6= 0, as ζ varies over R, mN (z), for fixed N ∈ N and
z ∈ C with =z 6= 0, varies over a circle KN in C, called Weyl circle. The fact that
χ(z) satisfies (B.9) may also be expressed as ζ = χN (z)/χN+1(z).
Proposition B.4. The inequality
‖χ(z)‖2N ≤=mN (z)
=z(B.11)
is precisely the condition for mN to lie inside a Weyl circle KN :
|cN −m|2 = r2N (B.12)
of center cN = WN [w, y]/WN [y, y] and radius rN = |WN [y, y]|−1.
Proof. Let us show that (B.12) is the Weyl circle. We write the circle equation
=ζ = =(χN/χN+1) = 0 as
0 = WN [χ, χ] = WN [v, v]−mNWN [y, v]−mNWN [v, y]+ |mN |2WN [y, y] . (B.13a)
Together with (B.3d), WN [g, h] = −WN [h, g] = −WN [h, g] and
|WN [g, h]|2 = WN [g, h]WN [g, h] +WN [g, g]WN [h, h] (B.13b)
with g = y and h = w, (B.13a) multiplied by 1/WN [y, y] (WN [y, y] 6= 0, by (B.13e)
below) reads
− 1
|WN [y, y]|2+WN [y, w]
WN [y, y]
WN [y, w]
WN [y, y]−mN
WN [y, w]
WN [y, y]− mN
WN [y, w]
WN [y, y]+ |mN |2 = 0
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 27
which is exactly −r2N + |cN −mN |2 = 0. Hence, by (B.13a), mN lies in KN iff
WN [χ, χ]
WN [y, y]= 0 . (B.13c)
Using again the fact that mN is a linear fractional map, either mN satisfies
|cN −mN | ≤ rN or it is outside the Weyl circle KN . Replacing mN in the circle
equation (B.13c) by its center cN , the l.h.s. of (B.13c) is the negative number −r2N .
So, mN is inside or at KN iffWN [χ, χ]
WN [y, y]≤ 0 . (B.13d)
Applying Green’s identity (B.3a) with w replaced by y, together with W0[y, y] = 0,
yield
WN [y, y] = −=z ‖y‖2N . (B.13e)
Similarly, with W0[χ, χ] = =mN , we have
WN [χ, χ] = =mN −=z ‖χ(z)‖2N . (B.13f)
Replacing (B.13e,f) into (B.13d), we conclude that mN is inside or at KN iff
=z ‖χ(z)‖2N −=mN
=z ‖y‖2N≤ 0
which is equivalent to (B.11).
q.e.d.
Problem B.12. Show (B.13b).
The radius rN = rN (z) =(=z ‖y(z)‖2N
)−1
is, by (B.13e) and definition (B.12),
a monotone decreasing function of N and the circles KN (z), N ≥ 1, contract to a
limit point by proposition B.2. It thus follows that mN (z) converges, as N → ∞,
to the Titchmarsh–Weyl m–function m(z), uniformly in compact sets of H, and
w(z)−m(z)y(z), =z > 0, is the (only one) l2(Z+)–solution of (B.1b).
Spectral decomposition in terms of the values of m(λ+ i0) The spectral
measure µ = dρ may be decomposed into absolutely continuous, singular, singular
continuous and pure point components: µ = µac + µs, µs = µsc + µpp, according to
the boundary value =m+(λ) = limε↓0=m(λ + iε) for which the Stieltjes inversion
formula (lemma 1, chapter 2 of [24])
ρ (λ+)− ρ (λ−) = limε↓0
1
π
∫ λ+
λ−
=m(λ+ iε)dλ
formula plays a fundamental role.
A function f(z) is said to possess a normal limit at point λ ∈ R if f(z)
converges to a (finite or infinite) limit as z ↓ λ along perpendicular to real axis
direction. Since m(z) is Herglotz, we have (see [17])
Lemma B.1. If the Radon–Nikodym derivative (dµ/dλ) (λ) of the spectral measure
µ at λ w.r.t. Lebesgue measure dλ exists, finite or infinitely, then =m+(λ) exists
and (dµ/dλ) (λ) = (1/π)=m+(λ).
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
28 Asymptotic Decay in Quantum Physics
Definition B.9. A set Σ ⊂ R is called minimal (or essential) support of a
measure ν in R if
i. ν(R/Σ) = 0 (i.e. Σ is the support of ν)
ii. Any subset Σ0 ⊂ Σ which does not support Σ has ν and Lebesgue measure 0:
ν(Σ0) = L(Σ0) = 0 .
It follows from de la Vallee–Poisson (theorem 9.6, chapter IV of [46]), Lebesgue–
Radon–Nikodym theorem (theorem 6.9 of [44]) and lemma B.1 that (see proposition
1 of [17])
Proposition B.5. The minimal supports Σ, Σac, Σs, Σsc and Σpp of µ, µac, µs,
µsc and µpp, the spectral measure µ of a Jacobi matrix J , and the absolutely con-
tinuous (ac), singular (s), singular continuous (sc) and pure point (pp) parts, are
respectively given by
i. Σ = λ ∈ E : 0 < =m+(λ) ≤ ∞ii. Σac = λ ∈ E : 0 < =m+(λ) <∞
iii. Σs = λ ∈ E : 0 < =m+(λ) =∞iv. Σsc = λ ∈ E : 0 < =m+(λ) =∞, L(λ) = 0v. Σpp = λ ∈ E : 0 < =m+(λ) =∞, L(λ) > 0
where E = λ ∈ R : =m+(λ) exists.
Now, we introduce a key notion in the Gilbert–Pearson theory relating solutions
of the Schrodinger equation (B.1b) and the decomposition of the spectral measure
µ of J .
Definition B.10. A solution u = (un)n≥0 of(Jφ − λI
)u = 0 (regardless the phase
boundary φ) is said to be subordinate iff
limL→∞
‖u‖L‖w‖L
= 0 (B.14)
holds for any linearly independent solution w = (wn)n≥0 of the equation, where
‖u‖2L =
L∑n=1
|un|2 denotes the norm over an interval of length L.
Remark B.2. Gilbert–Pearson’s theory deals with boundary values of the Weyl–
Titchmarsh m–function. Theorems 1 and 2 of [17] (see [29] for the discrete case)
prove that m+(λ) = limε↓0m(λ + iε) exists if, and only if, a subordinate solution
of (J − λI)u exists which (when it exists) is given by χ(λ) = w(λ)−m+(λ)y(λ) if
|m+(λ)| <∞ and by y(λ) if |m+(λ)| =∞. By (B.7), if m+(λ) exists, then m+α (λ)
exists for every α ∈ (0, π); if it exists and |m+(λ)| = ∞, then m+α (λ) is finite for
every α ∈ (0, π); if m+(λ) exists and is finite, then there is only one α ∈ (0, π) such
that |m+α (λ)| =∞.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 29
It turns out that the absolutely continuous part dρac of spectral measure dρ is
supported on a set Σac of λ’s for which no subordinate solution of(Jφ − λI
)u exists
and the singular part dρs is supported on a set Σs of λ’s for which u obey the phase
boundary φ and the Radon–Nikodym derivative dρ/dλ = (1/π)=m+φ (λ) diverges.
More precisely, we have
Proposition B.6. Let Jφ be an admissible operator in l2(Z+) satisfying φ–phase
b.c. (B.1c) and whose Schodinger equation is (B.1b). Let µ = dρ be its spectral
measure (B.8). Then, the minimal supports Σ, Σac, Σs, Σsc and Σpp of µ, µac, µs,
µsc and µpp are respectively given by
i. Σ = R/Σ0 where Σ0 =
λ ∈ R : there exists a subordinate solution but it does not satisfy (B.9)ii. Σac = λ ∈ R : no subordinate solution exists
iii. Σs = λ ∈ R : there exists a subordinate solution that satisfies φ-phase b.c.iv. Σsc =
λ ∈ R : there exists a subordinate solution satisfying φ-phase b.c. but does not belong to l2(Z+)v. Σpp =
λ ∈ R : there exists a l2(Z+) subordinate solution that satisfies φ-phase b.c..
Transfer matrix We restrict ourselves to the two sparse models we have dealt
with. For our convenience, we consider off–diagonal (??)-(??) analogous version of
the model Jω = J0 + V ω defined by (??-c) but the formulas derived for the latter
can be translated to the former without difficulties.
Definition B.11. Given λ ∈ R and n ∈ N, a (n–step) transfer matrix T (n;λ) is
a product
T (n;λ) := T (n, n− 1;λ)T (n− 1, n− 2;λ) · · ·T (1, 0;λ) (B.15)
of the (one–step) 2× 2 transfer matrix
T (k, k − 1;λ) =1
pk
(λ −pk−1
pk 0
)(B.16)
(replaced by
(λ− vk −1
1 0
)for the diagonal model (??-c)).
Any solution u = (un)n≥0 of the Schrodinger equation((Jφ − λI
)u)n
= pnun+1 + pn−1un−1 − λun = 0 (B.17)
for n ≥ 1, p0 ≡ 1, with φ–phase boundary condition at 0 (B.1c) satisfies(uk+1
uk
)= T (k, k − 1;λ)
(ukuk−1
). (B.18)
We denote by uφ =(uφn)n≥0
, the solution of (B.17) “normalized” by u20 + u2
1 = 1.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
30 Asymptotic Decay in Quantum Physics
Proposition B.7. Let (pn)n≥0 be a sparse sequence:
pn =
p if n = aj ∈ A1 otherwise
with 0 < p < 1 and A defined by (??). Then, there exists a real 2× 2 matrix U so
that the n–step transfer matrix conjugated by U reads
UT (n;λ)U−1 = R ((n− aj)ϕ)P+−R (βjϕ) · · ·P+−R (β1ϕ) , (B.19)
for aj < n < aj+1 and λ = 2 cosϕ, ϕ ∈ (0, π), where
R(θ) =
(cos θ sin θ
− sin θ cos θ
)(B.20)
and
P+− =
(p 0
(1/p− p) cotϕ 1/p
). (B.21)
Proof. Let λ = 2 cosϕ and let
U =
(0 sinϕ
1 − cosϕ
)(B.22)
be a 2 × 2 nonsingular matrix that conjugates the free transfer matrix T0(λ) =(λ −1
1 0
)into a clockwise rotation matrix:
Problem B.13. Show that UT0(λ)U−1 = R(ϕ).
(B.19) follows by (B.15), (B.20) and P+− = R (−ϕ)UT (aj + 1, aj ;λ)T (aj , aj −1;λ)U−1R (−ϕ) for each aj ∈ A.
q.e.d.
Prufer variables Definition B.12. Let uφn and U be as in definition B.11 and
(B.22). The real valued functions Rj = Rj(ϕ, φ) and θj = θj(ϕ, φ), j = 0, 1, 2, . . .,
given by
Rj
(cos θjsin θj
):= U
(uφaj+1
uφaj
), (B.23)
are called Prufer variables.
Proposition B.8 ( [37], pp. 770, 771 and 776). Let (pn)n≥1 be as in Proposi-
tion B.7. Then,
θj = g(ϕ, θj−1)− (βj + ωj − ωj−1)ϕ (B.24a)
1
R2j
= F (ϕ, θj)1
R2j−1
(B.24b)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 31
hold for any j ∈ N with (R0, θ0) ∈ R+ × [0, π], βj given by (??), ωjj≥1 i.i.d.
random variables,
g(ϕ, θ) = tan−1
(1
p2(tan θ + cotϕ)− cotϕ
), (B.24c)
F (ϕ, θ) =p2
a+ b cos 2θ + c sin 2θ, (B.24d)
and a, b and c are functions of p and ϕ:
2a =(1− p2
)2cot2 ϕ+ 1 + p4
2b =(1− p2
)2cot2 ϕ− 1 + p4
c =(1− p2
)cotϕ .
Proof. It follows directly from (B.19) and (B.23).
q.e.d.
Problem B.14. Prove the recursive relations (B.24b,b).
An explicit computation yields
a+ b cos 2θ + c sin 2θ ≥ minθ
(a+ b cos 2θ + c sin 2θ) = a−√a2 − p4 > 0 , (B.25)
and implies that F (ϕ, θ) is uniformly bounded in [0, π]× [0, π] and has unit mean:
F :=1
π
∫ π
0
F (ϕ, θ)dθ =p2
√a2 − b2 − c2
= 1 (B.26)
by equations (26) and (29) of [41] together with b2 + c2 = a2 − p4.
Write
v0 ≡ R0
(cos θ0
sin θ0
):= U
(cosφ
sinφ
), (B.27)
i.e., R0 =√
1− sin 2φ cosϕ and tan θ0 = cotφ/ sinϕ − cosϕ with 0 < θ0 < π. For
n and j such that aN < n < aN+1, the Euclidean norm∥∥UT (n;λ)U−1v0
∥∥2= R2
N , (B.28)
can be written as
R2N =
R2N
R2N−1
· · · R21
R20
R20 = exp
(N∑k=1
f (ϕ, θk)
)R2
0 , (B.29)
where
f(ϕ, θ) = − logF (ϕ, θ) (B.30)
by (B.24b), (B.24d) and (B.25), is uniformly bounded in [a, b]×[0, π], 0 < a < b < π,
and has mean
f =1
π
∫ π
0
f(ϕ, θ)dθ = log
(a+√a2 − b2 − c2
2p2
). (B.31)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
32 Asymptotic Decay in Quantum Physics
The quantity between parenthesis r = r(p, ϕ) =a+√a2 − b2 − c2
2p2, is explicitly
given by
r = 1 +v2
4csc2 ϕ
with v = (1− p2)/p or, using λ = 2 cosϕ, by
r = 1 +v2
4− λ2, (B.32)
and coincides with the analogous expression for the diagonal model (??-c). We shall
employ it in the following for both diagonal and off–diagonal sparse models.
Spectral nature criteria We gather a collection of results on spectral analysis
useful for the sparse models in consideration. We begin with the following (lemma
3.1 of [33])
Proposition B.9. Let uφ and T (n;λ) be defined as in definition B.11. Then∥∥uφ+π/2∥∥2
L+1
‖uφ‖2L+1
≤ 2
(1
L
L∑n=1
‖T (n;λ)‖2)2
. (B.33)
Proof. By definition B.11, uφ and uφ+π/2 satisfy the initial conditions (B.2) with
α = φ and it thus follows that
Tφ(n;λ) := T (n;λ)R (−φ) =
(uφn+1 u
φ+π/2n+1
uφn uφ+π/2n
)(B.34a)
and
1 = Wn
[uφ, uφ+π/2
]= −pn detTφ(n;λ), (B.34b)
by (B.18), (B.20), (B.3b) and (B.3c). Let the first and second columns of (B.34a)
be denoted by uφn and vφn. In terms of these vectors, the Wronskian reads
Wn
[uφ, uφ+π/2
]= pnuφn ·J vφn, with J =
(0 1
−1 0
)and by (B.34b), Cauchy-Schwarz
inequality, pn ≤ 1 and ‖J ‖ = 1∥∥uφn∥∥ ∥∥vφn∥∥ ≥ pn ∣∣uφn · J vφn∣∣ = 1 . (B.34c)
By definition (B.34a)
1
L
L∑n=1
∥∥vφn∥∥2 ≤ 1
L
L∑n=1
‖T (n;λ)‖2
holds (term–by–term) and, by ,
1 ≤
(1
L
L∑n=1
∥∥uφn∥∥∥∥vφn∥∥)2
≤ 1
L
L∑n=1
∥∥uφn∥∥2 1
L
L∑n=1
∥∥vφn∥∥2
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 33
by (B.34c) and Cauchy-Schwarz inequality. From the last two equations, one con-
cludes ∑Ln=1
∥∥vφn∥∥2∑Ln=1
∥∥∥uφn∥∥∥2 ≤
(1
L
L∑n=1
‖T (n;λ)‖2)2
. (B.34d)
and this, together with
L+1∑n=1
∣∣∣uφ+π/2n
∣∣∣2 ≤ L∑n=1
∥∥vφn∥∥2
L+1∑n=1
∣∣uφn∣∣2 ≥ 1
2
L∑n=1
∥∥uφn∥∥2, (B.34e)
yields (B.33).
q.e.d.
Let µ = dρ be the spectral measure of Jφ:(e1, f
(Jφ)e1
)=
∫f(λ)dρ(λ) (B.35)
for every bounded measurable function with compact support in (−2, 2). The vector
e1 is cyclic for Jφ in the sense that(Jφ)ke1 : k ∈ N
is a dense set in the Hilbert
space H and any other spectral measure µΨ is absolutely continuous with respect
to µ. According to proposition B.6,
Σac = λ ∈ [−2, 2] : there is no subordinate solution (B.36)
is an essential support of µac and has zero measure with respect to the singular part
µs. Proposition B.9 gives one–half of the following
Theorem B.1 (Theorem 1.1 of [33]). The essential support Σac of the a.c. part
µac is given by
Σac =
λ : lim inf
L→∞
1
L
L−1∑n=0
‖T (n;λ)‖2 <∞
. (B.37)
Proof. Let us denote by S the r.h.s. of (B.37). If Jφ has a subordinate solution
at λ then the r.h.s. of (B.33) diverge. So, S is contained into Σac and we quote
proposition 3.3 of [33] for a proof that it is an essential support of dρac, i.e., for
almost every λ ∈ [−2, 2] with respect to µac, we have λ ∈ S.
q.e.d.
Now, we come back to equations (B.28)–(B.32). Suppose that f = f(ϕ, θ) in
(B.29) is replaced by its average f = f(ϕ) given by (B.31) and
EN (ϕ) =1
N
N∑k=1
(f (ϕ, θk)− f(ϕ)
)−→ 0 , as N →∞ for a.e. ϕ ∈ [0, π) .
(B.38)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
34 Asymptotic Decay in Quantum Physics
Since by definition (B.32), r > 1 if v > 0 and by (B.28)
‖T (n;λ)‖2 ∼ rj
for aj < n < aj+1 a.e. λ ∈ (−2, 2) (see equations (B.42a) and (B.42c) below and
(??), for a derivation), we conclude by theorem B.1 the following
Theorem B.2. Let µ = dρ be the spectral measure of Jφ and suppose that the
hypothesis (B.38) holds. Then the essential support of its a.c. part µac, Σac = ∅.
The conclusions is not affected by the exclusion of a set A ⊂ [−2, 2] of Lebesgue
measure zero, by definition B.9 of essential support Σac. The presence of randomness
ω on the sparse models in consideration allows the hypothesis to be rigorously proved
(theorem ??).
Metric properties of spectral measure We state without proof an extension
to Gilbert-Pearson theory, due Jitomirskaya–Last [23], which relates the Hausdorff
decomposition of the spectral measure to a generalized subordinacy solution:
Theorem B.3 (Theorem 1.2 of [23]). Let µ = dρ be given by (B.35), λ ∈ [−2, 2]
and α ∈ (0, 1). Then, the upper Hausdorff derivative
Dαµ(λ) := lim sup
ε→0
ρ ((λ− ε, λ+ ε))
(2ε)α =∞
iff
lim infL→∞
∥∥uφ∥∥L∥∥uφ+π/2
∥∥α/(2−α)
L
= 0 .
One consequence of Theorem B.3 (see Corollary 4.4 of [23]) may be restated in
terms of the transfer matrix T (n;λ).
Corollary B.1 (Corollary 3.7 of [10]). Suppose that for some α ∈ [0, 1) and ev-
ery λ in a Borel set A,
lim supl→∞
1
l2−α
l∑n=0
‖T (n;λ)‖2 <∞ . (B.39)
Then, the restriction ρ(A ∩ ·) of ρ to the set A is α–continuous.
The other consequence is (see Corollary 4.5 of [23])
Corollary B.2 ((Corollary 3.8 of [10])). Suppose that
lim infl→∞
∥∥uφ∥∥2
l
lα= 0
holds for every λ in some Borel set A. Then, the restriction ρ(A∩ ·) is α–singular.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 35
The proof of both corollaries relates the behavior of the eigenvectors uφn and
uφ+π/2n with the norm of T (n;λ) and the Prufer radius Rk. Note that
R2j (φ) =
∥∥∥∥UT (n;λ)
(cosφ
sinφ
)∥∥∥∥2
=(uφn+1
)2
+(uφn)2 − 2uφn+1u
φn cosϕ (B.40a)
holds with aj < n < aj+1, φ ∈ (0, π) and λ = 2 cosϕ ∈ (−2, 2), by (B.23), (B.34a),
(B.22) and (B.20). From (B.40a) we deduce
(1− |cosϕ|)((
uφn+1
)2
+(uφn)2) ≤ R2
j (φ) ≤ (1 + |cosϕ|)((
uφn+1
)2
+(uφn)2)
(B.40b)
Problem B.15. Let ‖A‖U :=∥∥UAU−1
∥∥ be a matrix norm defined by (B.22) and
‖B‖ = supv∈C2 ‖Bv‖ / ‖v‖. Show that
C−1 ‖A‖ ≤ ‖A‖U ≤ C ‖A‖ (B.41)
holds with C =√
(1 + |cosϕ|) / (1− |cosϕ|).
Problem B.15 together with (B.28), imply
‖T (n;λ)‖ ≥ C−1 ‖v0‖−1 ∥∥UT (n;λ)U−1v0
∥∥ = C ′Rj(φ) (B.42a)
and
‖T (n;λ)‖ ≤ C supv:‖v‖=1
∥∥UT (n;λ)U−1v∥∥ = C ′′Rj(φ) . (B.42b)
where v =(cos θ, sin θ
)is the (unique) unit vector for which the supremum is
attained and φ solves θ0(φ) = θ for φ. To get rid of the sup, an artifact introduced
in theorem 2.3 of [31] can be used (see (B.51d)) to replace equation (B.42b) by
‖T (n;λ)‖ ≤ C max(Rj(φ
1), Rj(φ2))
(B.42c)
with C = C/∣∣sin (θ1
0 − θ20
)/2∣∣ and θi0 = θ0(φi), i ∈ 1, 2. See quoted reference for
completion of the proof.
Proposition B.10. If (βj)j≥1 satisfies a super–exponential sparseness condition
limj→∞ βj−1/βj = 0, then the spectrum σ(Jφ) of Jφ has Hausdorff dimension 1.
Proof. It follows from (B.42c) and (??) that, if (βj)j≥1 satisfies a super–exponential
sparseness, then for any Borel set A ⊂ [−2, 2] and ε > 0 there exist n0 = n0(ε,A)
such that
‖T (n;λ)‖2 ≤ Cjrj ≤ nε
holds for n ≥ n0, uniformly in A. So,
1
l2−α
l∑n=1
‖T (n;λ)‖2 ≤ Clα−1+ε
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
36 Asymptotic Decay in Quantum Physics
holds for l sufficiently large and the lim-sup is finite for any α > 1. This, to-
gether with Corollary B.1 and definition of Hausdorff dimension (see definition ??),
concludes the proof.
q.e.d.
We turn now to the point spectrum, which matters only for sequence (βj)j≥1
exponentially sparse
βj = βj (B.43)
for some integer β > 1, asymptotically as j →∞. By Theorems 1.6 and 1.7 of [33](Jφ − λI
)u = 0 has no l2 (N)–solutions if
∞∑n=1
‖T (n;λ)‖−2=∞ (B.44)
(see also Theorem 2.1 in [48]) and an l2 (N)–solution if
∞∑n=1
‖T (n;λ)‖2( ∞∑k=n
‖T (k;λ)‖−2
)2
<∞ . (B.45)
Condition (B.45) can be optimized for sparse models by improving Theorems
8.1 and 8.2 of [33] on the existence of subordinate and l2 (N)–solutions. By equation
(B.19), for n and N s.t. aN < n < aN+1, we have
‖T (n, λ)‖U ≤N∏k=1
‖P+−R (βkϕ)‖ ≤ ‖P+−‖N .
This together with (B.41) imply that ‖T (n;λ)‖ cannot growth exponen-
tially fast in n. As a consequence, the largest Lyapunov exponent, γ =
limn→∞(1/n) log ‖T (n;λ)‖, vanishes for T (n;λ) as a product of one-step trans-
fer matrices but it is strictly positive as a product of the P+−R (βkϕ):
limN→∞(1/N) log ‖P+−R (βNϕ) · · ·P+−R (β1ϕ)‖ > 0 a.e. ϕ.
We follow closely Lemma 4.1 of [11] and Proposition 3.9 of [10], where decay at
infinity of a subordinate solution were established (see also lemma 2.1 of [55]).
Proposition B.11. Let Jφ be given by (B.17) with sparse sequence (βn)n≥1 and
let tn denote
‖T (an + 1;λ)‖. Suppose that∞∑n=1
t−2n <∞ (B.46)
holds for some λ ∈ [−2, 2]. Then, there exist a unit vector v∗ = (cosφ∗, sinφ∗) with
φ∗ ∈ [0, π) so that
‖T (n;λ) v∗‖2 ≤ B
( ∞∑M=N
t−2M
)2
t2N + t−2N (B.47)
is satisfied for some constant B <∞ and every aN < n < aN+1 with N ∈ N.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 37
Corollary B.3. Under the assumptions of Proposition B.11, the vector uφ∗
n :=
T (n;λ) v∗ =
(uφ∗
n+1
uφ∗
n
)defines a strong subordinate solution uφ
∗=(uφ∗
n
)n≥0
in
the sense that
limn→∞
∥∥uφ∗n ∥∥∥∥∥uφn∥∥∥ = 0
holds for any φ ∈ [0, π) with φ 6= φ∗.
Corollary B.4. The strong subordinate solution uφ∗
=(uφ∗
n
)n≥0
is an l2 (Z+)–
solution of(Jφ − λI
)u = 0 with φ = φ∗, provided
∞∑n=1
βnt−2n <∞ and
∞∑n=1
( ∞∑m=n
t−2m
)2
βnt2n <∞ (B.48)
are verified in addition to the assumptions of proposition B.11.
Corollary B.5. Suppose, in addition to the assumptions of proposition B.11, that
C−1n rn ≤ t2n ≤ Cnrn (B.49)
holds with r > 1 and C1/nn 1 as n tends to ∞. Then∥∥∥uφ∗n ∥∥∥2
≤ Cnr−n (B.50)
holds with C1/nn 1 as n tends to ∞.
Remark B.3. As the Prufer angles (θωn)n≥1 are uniformly distributed mod π (the-
orem ??), equation (B.49) actually holds with r given by (B.32) a.e. λ. Theorem
?? excludes a set of λ’s which is countable and independent of the initial Prufer
angle θ0. These properties are crucial for proving pure point spectrum. Note that
(B.42a) and (B.42a) hold for some θi0, i = 1, 2, where θ0, by (B.27), depends on λ
and on the phase boundary φ.
Proof of proposition B.11. Let λ = 2 cosϕ and write
T (an + 1;λ) = An(λ) · · ·A1(λ)
where
UAkU−1 = P+−R (βkϕ) .
Let sn = ‖An(λ)‖ denote the spectral norm of An(λ). It follows from (B.15)
sn ≤ C ‖An(λ)‖U ≤ C ‖P+−‖ ≤ D (B.51a)
with D = C
√1 + (p− 1/p)
2csc2 ϕ, by (B.21), uniformly in n. As a consequence,
∞∑n=1
s2n+1
t2n≤ D2
∞∑n=1
1
t2n<∞ , (B.51b)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
38 Asymptotic Decay in Quantum Physics
by hypothesis (B.46), verifies the assumption of Theorem 8.1 of [33].
The transfer matrices Ak’s are 2×2 unimodular real matrices. Since unimodular
matrices form an algebra, the product of T (an + 1;λ) with its adjoint T ∗(an +
1;λ) is a 2 × 2 unimodular symmetric real matrix with eigenvalues t2n and t−2n
and corresponding (orthonormal) eigenvectors v+n and v−n : v+
n · v−n = 0. Write
vφ =
(cosφ
sinφ
)and define φn by
vφn = v−n . (B.51c)
Clearly, v+n = vφn+π/2 and
‖T (an + 1;λ)vφ‖2 = vφ · T ∗(an + 1;λ)T (an + 1;λ)vφ
= t2n∣∣vφ · v+
n
∣∣2 + t−2n
∣∣vφ · v−n ∣∣2= t2n sin2 (φ− φn) + t−2
n cos2 (φ− φn) . (B.51d)
by the spectral theorem.
The completion now follows exactly the steps of the proof of Theorem 8.1 of
[33]. The conclusion of proposition B.11, equation (B.47), is the combination of
equations (8.5) and (8.7) of [33]. We review the main steps. Firstly, the sequence
(φn)n≥1 converges to φ∗ ∈ [0, π] under the condition (B.51b). The key estimate
|φn − φn+1| ≤π
2
s2n+1
t2n, (B.51e)
established using properties of a matrix norm, (B.51d) for n + 1 and (B.51c), to-
gether with (B.51b) implies that (φn)n≥1 is a Cauchy sequence. Next, the telescope
estimate
|φn − φ∗| ≤∞∑m=n
|φm − φm+1| ≤π
2
∞∑m=n
s2m+1
t2m
replaced into (B.51d) yields
‖T (an + 1;λ)vφ∗‖2 ≤ t2n (φ∗ − φn)2
+ t−2n
≤ Bt2n
( ∞∑m=n
t−2m
)2
+ t−2n (B.51f)
with B =π
2D4, which concludes the proof.
q.e.d.
Proof of Corollary B.3. Let v∗ = vφ∗ and v∗ = vφ∗+π/2. By (B.51d),
‖T (an + 1;λ)v∗‖2 ≥ t2n/2 is satisfied for sufficiently large n and for any φ 6= φ∗
we have vφ = av∗ + bv∗, with a2 + b2 = 1, a 6= 0, and ‖T (an + 1;λ)vφ‖2 ≥ a2t2n/2
holds for some n sufficiently large. By (B.51d),∥∥uφ∗n ∥∥2∥∥∥uφn∥∥∥2 ≤2
a2
B( ∞∑m=n
t−2m
)2
+ t−4n
−→ 0
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Transfer matrix, Prufer variables and spectral analysis of sparse models 39
by (B.46), concluding the proof.
q.e.d.
Proof of Corollary B.4. We prove that the strongly subordinate solution is a l2(Z+)–
solution under assumptions (B.46) and (B.48). For any k ∈ N let n be such that
an + 1 ≤ k < an+1 holds. By (B.41),
‖T (k;λ)vφ∗‖2U ≤ C ‖vφ∗‖2
which, by (B.43), (B.40a) and (B.40b), together with (B.47), yields
∞∑k=1
∥∥∥uφ∗k ∥∥∥2
≤ 1
2(1− |cosϕ|)
∞∑k=1
‖UT (k;λ)vφ∗‖2
≤ B′∞∑n=1
βnt2n
( ∞∑m=n
t−2m
)2
+B′′∞∑n=1
βnt−2n
for some constants B′ and B′′, concluding the proof.
q.e.d.
Proof of Corollary B.5. Equation (B.50) is a direct consequence of (B.47) and
(B.49).
q.e.d.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
40 Asymptotic Decay in Quantum Physics
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Appendix C
Symmetric Cantor sets and relatedsubjects
Let AB be a closed segment of length l and let
0 < ξ < 1/2 (C.1)
be a number (the “ratio”). Following [28], consider a trisection of the segment AB
in parts respectively equal to lξ, l(1 − 2ξ), and lξ, and remove the open central
interval (“black” interval) of length l(1 − 2ξ); there remain two closed intervals
(“white” intervals) of common length lξ. Such a dissection of the given interval
AB will be said to be a dissection of type (2, ξ), the number 2 recalling that after
dissection there remain two white intervals.
We now start from a fundamental interval [a, b] (often the interval [0, 2π]). Let
us operate a dissection of type (2, ξ1) on this interval, then one of type (2, ξ2) over
each of the two remaining white intervals; further, a dissection of type (2, ξ3) over
each of the 22 white intervals obtained, and so on, the infinite sequence ξk∞1 being
such that
0 < ξk < 1/2 , for all k = 1, 2, . . . (C.2)
At the k–th step, we shall have a set Ek consisting of 2k white intervals of common
length
(b− a)ξ1 · · · ξk . (C.3)
The intersection
E ≡∞⋂k=1
Ek (C.4)
is a perfect set: E is closed and every point of E is a limit point of Ek. Further,
it has no interior points, i.e., its closure is nowhere dense; it is a Cantor set. Its
contiguous points are the edges of the black intervals obtained in the course of all
the dissections. We have (L will denote Lebesgue measure in this appendix), by
(C.3):
L(E) = 0⇔ limk→∞
2kξ1 · · · ξk = 0 (C.5)
41
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
42 Asymptotic Decay in Quantum Physics
Special cases of the symmetric Cantor sets described above are prototypes of
self-similar sets. In the case ξk = ξ < 1/2 above one has a homogeneous perfect
set E of type (2, ξ), but this notion may be generalized, see [28], p. 16. This set
E may be written as E = ∪2i=1fi(E), where f1(x) = x/3 and f2(x) = x/3 + 2/3.
The generalization of this structure to T = ∪Ni=1fi(T ), where T is a nonempty
compact set and (f1, . . . , fN ) are similitudes, i.e., functions of type fi(x) = λix+bi,
with 0 < λi < 1 is called an iterated function system [16](IFS) for T , and
T the attractor or the invariant set for the IFS. If there exists an open set V
such that fi(V ) ⊂ V and fi(V ) ∩ fj(V ) = ∅ for i 6= j, we say that the IFS
satisfies the open set condition. This is the case for the Cantor set of type (2, ξ),
with V = (0, ξ). Even further generalizations to Rn exist: see [16]. Measures
supported by these sets are called self-similar measures, and their construction
has been described in detail by Guarneri and Schulz-Baldes in a beautiful review
[18]: the latter authors also describe the construction of Schrodinger operators (one-
dimensional Jacobi matrices) having the previously mentioned self-similar measures
as spectral measures µΨ with Ψ = δ0 as cyclic vector (see theorem ??). This
construction was used in models of quasicrystals in three dimensions, see chapter
?? section ??.
Claim C.1. The 2k origins of the white intervals of which Ek consists are given by
the basic formula
a+ l[ε1(1− ξ1) + ε2ξ1(1− ξ2) + · · ·+ εkξ1 · · · ξk−1(1− ξk)] (C.6)
with εi ∈ 0, 1 ∀i = 1, . . . , k and l = b− a.
Proof. By induction: for k = 1, a+ l[ε1(1− ξ1)], for ε1 ∈ 0, 1 is s.t.
a+ 0 = a and a+ l(1− ξ1) = a+ lξ1 + l(1− 2ξ1)
Suppose, now, k ≥ k0 ≥ 1, and that the 2k0 origins are given by
a+ l[ε1(1− ξ1) + · · ·+ εk0ξ1 · · · ξk0−1(1− ξk0)] (C.7a)
For k = k0 + 1, one has, to each (C.7a), to add lξ1 · · · ξk0(1 − ξk0+1), but this
corresponds to build
a+ l[ε1(1− ξ1) + · · ·+ εk0ξ1 · · · ξk0−1(1− ξk0) + εk0+1ξ1 · · · ξk0(1− ξk0+1)] (C.7b)
with εk0+1 ∈ 0, 1. This ends the induction.
q.e.d.
Since limk→∞ ξ1 · · · ξk = 0 (because of (C.2)), the points of E given by (C.4) are
given by the infinite series
x = a+ l[ε1(1− ξ1 + ε2ξ1(1− ξ2) + · · ·+ εkξ1 · · · ξk−1(1− ξk) + · · · ] (C.8)
with εk ∈ 0, 1 for all integers k ≥ 1. Define
rk = ξ1 · · · ξk−1(1− ξk) (C.9)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Symmetric Cantor sets and related subjects 43
Then, by (C.2),
rk > rk+1 + rk+2 + · · · (C.10)
We associate to each perfect symmetric set E the Lebesgue function which is
constructed in the following way. Let Lk(x) be the continuous function which
equals zero for x ≤ a, one for x ≥ b, increasing linearly by 12k
on each white interval
building Ek, and constant in each black interval contiguous to Ek. If k → ∞,
Lk(x) tends uniformly (see below) to a function L(x), continuous, constant in each
interval contiguous to E, monotone increasing, with L(a) = 0, L(b) = 1: the
Lebesgue function constructed on E. The same simple induction leading from
(C.7a) to (C.7b) shows that for all x ∈ E, where E is given by (C.4),
L(x) =ε12
+ε222
+ · · ·+ εk2k
+ · · · (C.11)
In order to see that the convergence is uniform, let Lp(x) =
p∑i=1
εi/2i; then
|Lp(x)− Lp+1(x)| ≤ 2−p; thus the limp→∞ Lp(x) is uniform.
The variation of L(x) along an interval will be called the L–measure of this
interval, denoted by dL; the Stieltjes integral of a continuous function f(x) w.r.t.
dL will be denoted by
∫ b
a
f(x)dL(x); in particular, the Fourier-Stieltjes transform
of dL,
Γ(u) ≡∫ ∞−∞
exp(iux)dL(x) =
∫ b
a
exp(iux)dL(x) (C.12)
Observing that L(x) grows by 1/2k on every white interval building up Ek, we have
Γ(u) = limk→∞
Γk(u) (C.13a)
where
Γk(u) ≡ 1
2k
∑εj∈0,1∀1≤j≤k
exp(iu[a+ l(ε1r1 + · · ·+ εkrk)]) (C.13b)
with rk defined by (C.9). Thus
Γk(u) = exp(iua)
k∏j=1
1 + exp(iulrj)
2
= exp
iua+ l/2
k∑j=1
rj
k∏j=1
cos(ulrj/2) (C.14a)
from which, using
∞∑j=1
rj = 1 and b− a = l,
Γ(u) = exp[iu(a+ b)/2]
∞∏j=1
cos(ulrj/2) . (C.14b)
Problem C.16. Prove the convergence of the infinite product on the r.h.s. of
(C.14b).
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
44 Asymptotic Decay in Quantum Physics
In particular, if [a, b] is centered at the origin, with a = −d, b = d,
Γ(u) =
∞∏j=1
cos(udrj) . (C.15)
Definition C.13. If all ξj have a common value ξ, E = Eξ is said to be a symmetric
Cantor set of constant ratio ξ.
Problem C.17. Prove that for Eξ a symmetric Cantor set of constant ratio ξ
constructed over [−d, d],
Γ(u) =
∞∏j=1
cos[ud(1− ξ)ξj−1] . (C.16)
The Hausdorff dimension dimH of the symmetric Cantor set of ratio ξ is
dimH(E) = | log 1/2| /| log ξ|(see [28]). There are several additional concepts of dimension which are relevant to
describe sets of this type, such as the information dimension or the fractal dimension
(see [15] for details). Let A be a set in Rn, and n(ε) the minimal number of balls
of radius ε which are necessary to cover A. The fractal dimension DF (A) is defined
by
DF (A) = limε→0
n(ε)
| log ε|For the set Eξ, considering the Ek = Ek,ξ in (C.4), take ε = ξ−k; then n(ε) = 2k,
and DF (E) = limk→∞log 2k
log ξk=
log 2
log ξ. For the set Eξ the Hausdorff and fractal
dimensions are equal. For any countable set, it is easy to see that both dimensions
yield zero, and for an interval in R, they yield one. Hence, a nonzero Hausdorff or
fractal dimension is a signal of uncountability (but there are uncountable sets of
zero Hausdorff measure, see chapter ?? for explicit examples). It is easy to see that
Eξ is uncountable; take, e.g., ξ = 1/3. The elements of E1/3 are all the numbers
which in base three representation
∞∑i=1
εi3−i have εi ∈ 0, 2 ∀i. We may, therefore,
define a one–to–one map of E1/3 to the set of numbers
∞∑i=1
εi2−i with εi ∈ 0, 1 ∀i,
i.e., to the whole interval [0, 1], hence the set E1/3 is uncountable.
We now show that the Lebesgue function L, given by (C.11), is UαH (see defi-
nition ??):
Proposition C.12. Let ξk = ξ for k = 1, 2, . . . with ξ satisfying (C.1). Then L,
given by (C.11), is UαH , i.e.,
L(x)− L(x′) ≤ C(x− x′)| log 1/2|| log ξ| (C.17a)
where, for definiteness,
0 < x− x′ < 1 (C.17b)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Symmetric Cantor sets and related subjects 45
Proof. Suppose first that both x and x′ satisfy (C.17b) and are of the form (C.8),
then
x− x′ = l[(ε′p+1 − εp+1)ξp(1− ξ) + (ε′p+2 − εp+2)ξp+1(1− ξ) + · · · ] (C.18)
If the first p εi are equal for x and x′, for definiteness, then ε′p+1 − εp+1 = 1, and∣∣∣ε′p+k − εp+k∣∣∣ ≤ 1. Thus, from (C.18),
x′ − x ≥ lξp(1− ξ)− l[ξp+1(1− ξ) + ξp+2(1− ξ) + · · · ]= lξp(1− ξ)− lξp+1 = lξp(1− 2ξ) = Aξp (C.19a)
where, by (C.1),
0 < A <∞ , with A independent of x and x′ (C.19b)
By (C.11),
L(x′)− L(x) ≤ 2−p + 2−(p+1) + · · · = 2−p
1− 1/2= 22−p (C.19c)
By (C.19a), (x − x′)−1 ≤ A−1ξ−p, whence log(x − x′)−1 ≤ logA−1 + p| log ξ|, and
therefore
p ≥ − log(x− x′)| log ξ|
+logA
| log ξ|(C.19d)
By (C.19c-d),
L(x′)− L(x) ≤ C2− log(x′−x)| log ξ| = C(x− x)α , α ≡ | log 1/2|
| log ξ|(C.19e)
which is (C.17a). If x or x′ or both are not of the form (C.8), let x0 be the right
end point of the black interval containing x, and x′0 the left hand end point of the
black interval containing x′. As x−x′ > x0−x′0, and L(x′)−L(x) = L(x0)−L(x′0)
by the construction of L, (C.17a) subsists.
q.e.d.
We have now
Proposition C.13. Let Jα(L) be defined by (??) (for µ = L). Then, for any
0 < ε < α,
0 < Jα−ε(L) <∞ (C.20)
Proof. By proposition (C.1), L is UαH, and then, by Lemma ?? , (C.20) holds.
q.e.d.
Proposition C.14. Let d = π, and
ξ = 1/p (C.21a)
with p any odd integer such that
p ≥ 3 . (C.21b)
Then (C.20) holds, with α given by (C.19e), but
Γ(u) 9 0 , as u→∞ (C.21c)
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
46 Asymptotic Decay in Quantum Physics
Proof. By (C.16) with d = π,
Γ(u) =
∞∏j=1
cos[πu(1− 1/p)p−(j−1)] =
∞∏j=1
cos
[πu(p− 1)
pj
](C.21d)
Since p is an odd integer ≥ 3, it follows from (C.21d) that
Γ(n) = Γ(pn) ∀n ∈ Z (C.21e)
and thus (C.21c) is proved.
q.e.d.
Proposition C.14 shows that there exist singular measures which are not Rajch-
man measures (see chapter ?? section ??) and propositions C.13 and C.14 together
show that, also in one dimension, finiteness of the integral Jα−ε(µ) does not im-
ply pointwise convergence (see also chapter ?? section ??). The further problem
of whether it could happen that Γ(u) → 0 as u → ∞ except for u in a set of
zero Lebesgue measure, apparently left open by the proof of proposition C.14, is
negatively answered in section ?? (see Proposition ??).
The amusing remark (C.21e) goes back at least to [28]. Lyons [35] also mentions
it and remarks that, to our benefit, (C.21e) was not observed by F. Riesz, who,
instead, constructed his famous Riesz products [43] for precisely the purpose of
providing an example of a continuous measure which is not a Rajchman measure.
The proof of proposition C.12 is patterned after Salem’s paper [47].
We close this appendix with some important remarks. If, instead of (C.2)
1/2 < ξk < 1 (C.22)
the support of dL is the interval [−d, d]. By a theorem of Jessen and Wintner [22],
the corresponding measure may be either singular or a.c.; in the latter case, we have
an a.c. measure with nowhere dense support (see theorem ??). Such measures have
the following property [3]:
Proposition C.15. Suppose µ is an absolutely continuous measure supported by a
Cantor set C. Then µ /∈ L1(R).
Proof ([3], [8], chap. 10, p. 214) dµ(x) = f(x)dx for a function f supported by
C. Since C is a Cantor set, f cannot be continuous; but, if µ ∈ L1, f would be
continuous.
q.e.d.
Example C.1. The famous Vieta’s identity (A.1) may be rewritten as
sin t
t=
( ∞∏k=1
cos(t/22k−1)
)( ∞∏k=1
cos(t/22k)
)(C.23)
By ([27]), each product on the r.h.s. of (C.23) is the Fourier-Stieltjes transform
of a s.c. measure. Indeed, the second term on the r.h.s. of (C.23) is the Fourier-
Stieltjes transform µE of a measure supported by a perfect symmetric set E of
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Symmetric Cantor sets and related subjects 47
constant ratio ξ = 1/4 and is, therefore, s.c. ([27], p. 15). The first term on the
r.h.s. of (C.23) equals µE(2t) = µE′(t), where E′ is the perfect symmetric set of
constant ratio ξ = 1/4 but constructed on the interval [−π/2, π/2], which is, of
course, also s.c. by the same argument. Finally, the l.h.s. of (C.23) belongs to
L2(R,L) and it follows from theorem ?? (for p = 2) that µ is a.c..
Definition C.14 ([3]). Let H be a self–adjoint operator on a separable Hilbert
space H. The vector φ ∈ H is called a transient vector for H if
(φ, exp(−itH)φ) = O(t−N ) ∀N ∈ N (C.24)
The closure of the set of transient vectors is calledHtac (transient a.c. subspace).
One has [3, 8]:
Proposition C.16. Htac is a subspace of H such that Htac ⊂ Hac, and
Htac = φ : µφ ∈ L1(R) (C.25)
where the bar indicates closure. One defines
Hrac = H⊥tac ∩Hac (C.26)
which is called the recurrent a.c. subspace [3]: both Htac and Hrac are invariant
subspaces under H (see [3, 8]).
Chapter ?? section ?? discusses the connection of the above spectra with quan-
tum dynamical stability and provides or mentions explicit notrivial examples of
transient and recorrent a.c. spectra.
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
48 Asymptotic Decay in Quantum Physics
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
Bibliography
[1] D. V. Anosov. Sov. Math. Dokl., 4:1153, 1963.[2] V. I. Arnold and A. Avez. Ergodic problems of classical mechanics. W. A. Benjamin,
Inc., 1968.[3] J. Avron and B.Simon. Transient and recurrent spectrum. J. Func. Anal., 43:1–31,
1981.[4] V. Baladi and M. Viana. Ann. Sci. E. N. S., 29:483–517, 1996.[5] R. Bowen and D. Ruelle. Inv. Math., 29:181, 1975.[6] K. L. Chung. A course in probability theory. Academic Press, 1974.[7] I. P. Cornfeld, S. V. Fomin, and Ya. G. Sinai. Ergodic theory. Springer Verlag, 1981.[8] H. L. Cycon, R. M. Froese, W. Kirsch, and B. Simon. Schrodinger Operators. Springer
Verlag, 1987.[9] A. M. Ozorio de Almeida. Hamiltonian systems, chaos and quantization. Cambridge
University Press, 1988.[10] S. L. de Carvalho, D. H. U. Marchetti, and W. F. Wreszinski. J. Math. Anal. Appl.,
368:218–234, 2010.[11] S. L. de Carvalho, D. H. U. Marchetti, and W. F. Wreszinski. On the uniform dis-
tribution of prufer angles and its implication to a sharp spectral transition of jacobimatrices with randomly sparse perturbations. arXiv: 1006.2849, 2011.
[12] J. R. Dorfman. An introduction to chaos in nonequilibrium statistical mechanics.Cambridge University Press, 1999.
[13] P. Duarte. Ann. Inst. Henri Poinc., 11:359–459, 1994.[14] P. Duarte. Erg. Th. Dyn. Syst., 29:1781–1813, 2008.[15] J. P. Eckmann and D. Ruelle. Rev. Mod. Phys., 57:617, 1985.[16] K. Falconer. The geometry of fractal sets. Cambridge University Press, 1985.[17] D. Gilbert and D. P. Pearson. J. Math. Anal. Appl., 128:30–58, 1987.[18] I. Guarneri and H. Schulz-Baldes. Rev. Math. Phys., 11:1249, 1999.[19] J. Guckenheimer and P. Holmes. Nonlinear oscillations dynamical systems and bifur-
cations of vector fields. Springer Verlag, 1983.[20] M. Hirsch and S. Smale. Differential equations dynamical systems and linear algebra.
Academic Press, 1974.[21] H. R. Jauslin. Stability and chaos in classical and quantum hamiltonian systems. II
Granada seminar on computational physics, World Scientific, 1993.[22] B. Jessen and A. Wintner. Distribution functions and the riemann zeta function.
Trans. Am. Math. Soc., 38:48–88, 1935.[23] S. Jitomirskaya and Y. Last. Comm. Math. Phys., 211:643, 2000.[24] W. F. Donoghue Jr. Monotone matrix functions and analytic continuation. Springer-
49
August 28, 2013 13:52 World Scientific Book - 9.75in x 6.5in asymptotics˙appendices
50 Asymptotic Decay in Quantum Physics
Verlag, original edition, 1974.[25] M. Kac. Statistical independence in probability, analysis and number theory, vol-
ume 12. The Carus Mathematical Monographs, 1959.[26] S. Kaczmarz and H. Steinhaus. Theorie der Orthogonalreihen. Warszawa-Lwow, 1935.[27] J. P. Kahane and R. Salem. Colloquium Mathematicum, 6:193–202, 1958.[28] Jean Pierre Kahane and Raphael Salem. Ensembles Parfaits et series
trigonometriques. Hermann, Paris, 2nd edition, 1994.[29] S. Khan and D. B. Pearson. Helv. Phys. Acta, 65:505–527, 1992.[30] A. I. Khinchin. Continued Fractions. Dover Publ. Inc., 1992.[31] A. Kiselev, Y. Last, and B. Simon. Comm. Math. Phys., 194:1–45, 1998.[32] A. Lasota and M. Mackey. Probabilistic properties of deterministic systems. Cam-
bridge University Press, 1985.[33] Y. Last and B. Simon. Inv. Math., 135:329–367, 1999.[34] J. L. Lebowitz and O. Penrose. Modern ergodic theory. Physics Today, 20, 1973.[35] Russell Lyons. Seventy years of rajchman measures. J. Fourier Anal. Appl., 45:363,
1995.[36] R. Mane. Ergodic theory and differentiable dynamics. Springer Verlag Berlin, 1987.[37] D. H. U. Marchetti, W. F. Wreszinski, L. F. Guidi, and R. M. Angelo. Nonlinearity,
20:765, 2007.[38] Y. Meyer. Algebraic numbers and harmonic analysis. North Holland Publ. Co. Ams-
terdam, 1972.[39] V. I. Oseledec. Trans. Moscow Math. Soc., 19:197, 1968.[40] J. Palis and F. Takens. Hyperbolicity and sentive chaotic dynamics at homoclinic
bifurcations. Cambridge University Press, 1993.[41] D. B. Pearson. Comm. Math. Phys., 60:13–36, 1978.[42] M. Reed and B. Simon. Methods in modern mathematical physics - v.1, Functional
Analysis. Academic Press, 1st edition, 1972.[43] F. Riesz. Math. Zeit., 2:312–315, 1918.[44] W. Rudin. Real and Complex Analysis. McGraw-Hill, 2nd edition, 1974.[45] D. Ruelle. Comm. Math. Phys., 125:239, 1989.[46] J.J. Sakurai. Advanced Quantum Mechanics. Addison Wesley Publ. Co., 1967.[47] R. Salem. Jour. Math. and Phys., 21:69–81, 1942.[48] B. Simon and G. Stolz. Proc. Am. Math. Soc., 124:2073–2080, 1996.[49] Ya. G. Sinai. Russ. Math. Surv., 27:21, 1972.[50] Ya. G. Sinai. Topics in ergodic theory. Princeton University Press, 1994.[51] W. Szlenk. An introduction to the theory of smooth dynamical systems. Wiliey, 1984.[52] G. Teschl. Jacobi operators and completely integrable nonlinear lattices. Number 72
in Mathematical Surveys and Monographs. American Mathematical Society, originaledition, 2000.
[53] M. Viana. Stochastic dynamics of deterministic systems. IMPA lecture notes 1997.[54] P. Walters. Ergodic theory - introductory lectures, volume 458 of Lecture Notes in
Math. Springer, Berlin, Heidelberg, New York, 1965.[55] A. Zlatos. Jour. Func. Anal., 207:216–252, 2004.