Regularity of functions:
Genericity and multifractal analysis
Dissertation presented by
Céline ESSER
for the degree of Doctor in Sciences
University of Liège – Institute of Mathematics
Liège – October 22, 2014
Advisor: Françoise BASTIN(University of Liège)
Introduction Weierstraß function. W (x) := +∞ X n=0 ancos(bnπx), a∈ (0, 1), ab > 1. 1 −2 −1 0 1 2 1
Figure:Weierstraß function fora = 0.5andb = 3
Two questions.
• Are theremany such functions? Or is this example atypical?
−→Notions of genericity
• Is it possible tocharacterize the local behaviorof such functions?
−→Hölder exponent and multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lν
Introduction Weierstraß function. W (x) := +∞ X n=0 ancos(bnπx), a∈ (0, 1), ab > 1. Two questions.
• Are theremany such functions? Or is this example atypical?
−→Notions of genericity
• Is it possible tocharacterize the local behaviorof such functions?
−→Hölder exponent and multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lν
spaces
Introduction Weierstraß function. W (x) := +∞ X n=0 ancos(bnπx), a∈ (0, 1), ab > 1. Two questions.
• Are theremany such functions? Or is this example atypical?
−→Notions of genericity
• Is it possible tocharacterize the local behaviorof such functions?
−→Hölder exponent and multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lν
Introduction Weierstraß function. W (x) := +∞ X n=0 ancos(bnπx), a∈ (0, 1), ab > 1. Two questions.
• Are theremany such functions? Or is this example atypical?
−→Notions of genericity
• Is it possible tocharacterize the local behaviorof such functions?
−→Hölder exponent and multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lν
spaces
Introduction Weierstraß function. W (x) := +∞ X n=0 ancos(bnπx), a∈ (0, 1), ab > 1. Two questions.
• Are theremany such functions? Or is this example atypical?
−→Notions of genericity
• Is it possible tocharacterize the local behaviorof such functions?
−→Hölder exponent and multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lν
Introduction Weierstraß function. W (x) := +∞ X n=0 ancos(bnπx), a∈ (0, 1), ab > 1. Two questions.
• Are theremany such functions? Or is this example atypical?
−→Notions of genericity
• Is it possible tocharacterize the local behaviorof such functions?
−→Hölder exponent and multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lν
spaces
Notions of genericity Residuality, prevalence and lineability
Notions of genericity
• Residuality. LetX be a Baire space. A subsetM ofXisresidualinX ifM
contains a countable intersection of dense open sets inX.
• Prevalence (Christensen, 1974 / Hunt, Sauer, Yorke, 1992). LetXbe a
complete metrizable vector space. A Borel subsetM ofXisshyif there exists a
Borel measureµonX with compact support such that
µ(M + x) = 0, x∈ X.
More generally, a subsetV is called shy if it is contained in a shy Borel set. The
complement of a shy set is called aprevalentset.
• Lineability (Aron, Gurariy, Seoane-Sepúlveda, 2005). LetXbe a topological
vector space andµa cardinal number. A subsetM ofXis(dense-)lineableif
M∪ {0}contains an infinite dimensional vector subspace (dense) inX. If the
Notions of genericity Residuality, prevalence and lineability
Notions of genericity
• Residuality. LetX be a Baire space. A subsetM ofXisresidualinX ifM
contains a countable intersection of dense open sets inX.
• Prevalence (Christensen, 1974 / Hunt, Sauer, Yorke, 1992). LetXbe a
complete metrizable vector space. A Borel subsetM ofXisshyif there exists a
Borel measureµonX with compact support such that
µ(M + x) = 0, x∈ X.
More generally, a subsetV is called shy if it is contained in a shy Borel set. The
complement of a shy set is called aprevalentset.
• Lineability (Aron, Gurariy, Seoane-Sepúlveda, 2005). LetXbe a topological
vector space andµa cardinal number. A subsetM ofXis(dense-)lineableif
M∪ {0}contains an infinite dimensional vector subspace (dense) inX. If the
dimension of this subspace isµ,M is said to beµ-(dense-)lineable.
Notions of genericity Residuality, prevalence and lineability
Notions of genericity
• Residuality. LetX be a Baire space. A subsetM ofXisresidualinX ifM
contains a countable intersection of dense open sets inX.
• Prevalence (Christensen, 1974 / Hunt, Sauer, Yorke, 1992). LetXbe a
complete metrizable vector space. A Borel subsetM ofXisshyif there exists a
Borel measureµonX with compact support such that
µ(M + x) = 0, x∈ X.
More generally, a subsetV is called shy if it is contained in a shy Borel set. The
complement of a shy set is called aprevalentset.
• Lineability (Aron, Gurariy, Seoane-Sepúlveda, 2005). LetXbe a topological
vector space andµa cardinal number. A subsetM ofX is(dense-)lineableif
M∪ {0}contains an infinite dimensional vector subspace (dense) inX. If the
Notions of genericity Denjoy-Carleman classes
Existence of nowhere analytic functions. An example was given by Cellérier (1890)
by the function f (x) := +∞ X n=1 sin(anx) n! , x∈ R
whereais a positive integer larger than1.
Results.
• Genericity of the set of nowhere analytic functions inC∞([0, 1]).
• Extension of these results using Gevrey classes.
Question. Similar results in the context of classes of ultradifferentiable functions?
Notions of genericity Denjoy-Carleman classes
Existence of nowhere analytic functions. An example was given by Cellérier (1890)
by the function f (x) := +∞ X n=1 sin(anx) n! , x∈ R
whereais a positive integer larger than1.
Results.
• Genericity of the set of nowhere analytic functions inC∞([0, 1]).
• Extension of these results using Gevrey classes.
Notions of genericity Denjoy-Carleman classes
Existence of nowhere analytic functions. An example was given by Cellérier (1890)
by the function f (x) := +∞ X n=1 sin(anx) n! , x∈ R
whereais a positive integer larger than1.
Results.
• Genericity of the set of nowhere analytic functions inC∞([0, 1]).
• Extension of these results using Gevrey classes.
Question. Similar results in the context of classes of ultradifferentiable functions?
Notions of genericity Denjoy-Carleman classes
Denjoy-Carleman classes
An arbitrary sequence of positive real numbersM = (Mk)k∈N0 is called aweight
sequence.
Definition
LetΩbe an open subset ofRandM be a weight sequence. The spaceE{M }(Ω)is
defined by
E{M }(Ω) :=f ∈ C∞(Ω) :∀K ⊆ Ωcompact∃h > 0such thatkfkMK,h< +∞ ,
where kfkM K,h:= sup k∈N0 sup x∈K |Dkf (x) | hkM k .
Iff ∈ E{M }(Ω), we say thatf isM-ultradifferentiable of Roumieu typeonΩ.
Particular case. The weight sequences(k!)k∈N0and((k!)α)
Notions of genericity Denjoy-Carleman classes
Denjoy-Carleman classes
An arbitrary sequence of positive real numbersM = (Mk)k∈N0 is called aweight
sequence.
Definition
LetΩbe an open subset ofRandM be a weight sequence. The spaceE{M }(Ω)is
defined by
E{M }(Ω) :=f ∈ C∞(Ω) :∀K ⊆ Ωcompact∃h > 0such thatkfkMK,h< +∞ ,
where kfkM K,h:= sup k∈N0 sup x∈K |Dkf (x) | hkM k .
Iff ∈ E{M }(Ω), we say thatf isM-ultradifferentiable of Roumieu typeonΩ.
Particular case. The weight sequences(k!)k∈N0and((k!)α)
k∈N0 withα > 1.
Notions of genericity Denjoy-Carleman classes
Denjoy-Carleman classes
An arbitrary sequence of positive real numbersM = (Mk)k∈N0 is called aweight
sequence.
Definition
LetΩbe an open subset ofRandM be a weight sequence. The spaceE{M }(Ω)is
defined by
E{M }(Ω) :=f ∈ C∞(Ω) :∀K ⊆ Ωcompact∃h > 0such thatkfkMK,h< +∞ ,
where kfkM K,h:= sup k∈N0 sup x∈K |Dkf (x) | hkM k .
Iff ∈ E{M }(Ω), we say thatf isM-ultradifferentiable of Roumieu typeonΩ.
Particular case. The weight sequences(k!)k∈N0and((k!)α)
Notions of genericity Denjoy-Carleman classes
Definition
LetΩbe an open subset ofRandM be a weight sequence. The spaceE(M )(Ω)is
defined by
E(M )(Ω) :=f ∈ C∞(Ω) :∀K ⊆ Ωcompact,∀h > 0, kfkMK,h< +∞ .
Iff ∈ E(M )(Ω), we say thatf isM-ultradifferentiable of Beurling typeonΩand we
use the representation
E(M )(Ω) = proj ←−−− K⊆Ω proj ←−− h>0 EM,h(K)
to endowE(M )(Ω)with a structure of Fréchet space.
Questions.
• When do we haveE{M }(Ω)⊆ E(N )(Ω)?
• In that case, “how small” isE{M }(Ω)inE(N )(Ω)?
Notions of genericity Denjoy-Carleman classes
Definition
LetΩbe an open subset ofRandM be a weight sequence. The spaceE(M )(Ω)is
defined by
E(M )(Ω) :=f ∈ C∞(Ω) :∀K ⊆ Ωcompact,∀h > 0, kfkMK,h< +∞ .
Iff ∈ E(M )(Ω), we say thatf isM-ultradifferentiable of Beurling typeonΩand we
use the representation
E(M )(Ω) = proj ←−−− K⊆Ω proj ←−− h>0 EM,h(K)
to endowE(M )(Ω)with a structure of Fréchet space.
Questions.
• When do we haveE{M }(Ω)⊆ E(N )(Ω)?
Notions of genericity Denjoy-Carleman classes
General assumptions.
• We assume that any weight sequenceM is logarithmically convex, i.e.
Mk2≤ Mk−1Mk+1, ∀k ∈ N .
It implies that the spaceE{M }(Ω)is an algebra.
• We assume that any weight sequenceM is such thatM0= 1.
• We usually assume that any weight sequenceM is non-quasianalytic, i.e.
+∞
X
k=1
(Mk)−1/k< +∞.
By Denjoy-Carleman theorem, it implies that there exists non-zero functions with
compact support inE{M }(R).
Notions of genericity Denjoy-Carleman classes
General assumptions.
• We assume that any weight sequenceM is logarithmically convex, i.e.
Mk2≤ Mk−1Mk+1, ∀k ∈ N .
It implies that the spaceE{M }(Ω)is an algebra.
• We assume that any weight sequenceM is such thatM0= 1.
• We usually assume that any weight sequenceM is non-quasianalytic, i.e.
+∞
X
k=1
(Mk)−1/k< +∞.
By Denjoy-Carleman theorem, it implies that there exists non-zero functions with
Notions of genericity Denjoy-Carleman classes
General assumptions.
• We assume that any weight sequenceM is logarithmically convex, i.e.
Mk2≤ Mk−1Mk+1, ∀k ∈ N .
It implies that the spaceE{M }(Ω)is an algebra.
• We assume that any weight sequenceM is such thatM0= 1.
• We usually assume that any weight sequenceM is non-quasianalytic, i.e.
+∞
X
k=1
(Mk)−1/k< +∞.
By Denjoy-Carleman theorem, it implies that there exists non-zero functions with
compact support inE{M }(R).
Notions of genericity Denjoy-Carleman classes
Inclusions between Denjoy-Carleman classes
Notation. Given two weight sequencesM andN, we write
M N ⇐⇒ lim k→+∞ Mk Nk 1k = 0.
Proposition
LetM, N be two weight sequences and letΩbe an open subset ofR. Then
M N ⇐⇒ E{M }(Ω)⊆ E(N )(Ω)
and in this case, the inclusion is strict.
Keys.
• IfM N, then there exists a weight sequenceLsuch thatM L N.
• There existsθ∈ E{M }(R)such that|Dkθ(0)| ≥ Mkfor allk∈ N0. In particular,
Notions of genericity Denjoy-Carleman classes
Inclusions between Denjoy-Carleman classes
Notation. Given two weight sequencesM andN, we write
M N ⇐⇒ lim k→+∞ Mk Nk 1k = 0.
Proposition
LetM, N be two weight sequences and letΩbe an open subset ofR. Then
M N ⇐⇒ E{M }(Ω)⊆ E(N )(Ω)
and in this case, the inclusion is strict.
Keys.
• IfM N, then there exists a weight sequenceLsuch thatM L N.
• There existsθ∈ E{M }(R)such that|Dkθ(0)| ≥ Mkfor allk∈ N0. In particular,
this function does not belong toE(M )(R).
Notions of genericity Denjoy-Carleman classes
Inclusions between Denjoy-Carleman classes
Notation. Given two weight sequencesM andN, we write
M N ⇐⇒ lim k→+∞ Mk Nk 1k = 0.
Proposition
LetM, N be two weight sequences and letΩbe an open subset ofR. Then
M N ⇐⇒ E{M }(Ω)⊆ E(N )(Ω)
and in this case, the inclusion is strict.
Keys.
• IfM N, then there exists a weight sequenceLsuch thatM L N.
• There existsθ∈ E{M }(R)such that|Dkθ(0)| ≥ Mkfor allk∈ N0. In particular,
Notions of genericity Denjoy-Carleman classes
Construction
Definition
We say that a function isnowhere inE{M }if its restriction to any open and non-empty
subsetΩofRnever belongs toE{M }(Ω).
Proposition
Assume thatM andNare two weight sequences such thatM N. IfM is
non-quasianalytic, there exists a function ofE(N )(R)which is nowhere inE{M }.
Idea. Construct a sequence(L(p))
p∈Nof weight sequences such that
M L(1) L(2) · · · L(p) · · · N.
For everyp∈ N, consider a functionfp∈ E{L(p)}(R)such that|Dkfp(0)| ≥ L
(p) k ,
∀k ∈ N0.If{xp: p∈ N}is a dense subset ofR, consider f (x) =
+∞
X
p=1
fp(x− xp)Φp(x), x∈ R
whereΦpis a compactly supported function well chosen.
Notions of genericity Denjoy-Carleman classes
Construction
Definition
We say that a function isnowhere inE{M }if its restriction to any open and non-empty
subsetΩofRnever belongs toE{M }(Ω).
Proposition
Assume thatM andN are two weight sequences such thatM N. IfM is
non-quasianalytic, there exists a function ofE(N )(R)which is nowhere inE{M }.
Idea. Construct a sequence(L(p))
p∈Nof weight sequences such that
M L(1) L(2) · · · L(p) · · · N.
For everyp∈ N, consider a functionfp∈ E{L(p)}(R)such that|Dkfp(0)| ≥ L
(p) k ,
∀k ∈ N0.If{xp: p∈ N}is a dense subset ofR, consider f (x) =
+∞
X
p=1
fp(x− xp)Φp(x), x∈ R
Notions of genericity Denjoy-Carleman classes
Construction
Definition
We say that a function isnowhere inE{M }if its restriction to any open and non-empty
subsetΩofRnever belongs toE{M }(Ω).
Proposition
Assume thatM andN are two weight sequences such thatM N. IfM is
non-quasianalytic, there exists a function ofE(N )(R)which is nowhere inE{M }.
Idea. Construct a sequence(L(p))
p∈Nof weight sequences such that
M L(1) L(2) · · · L(p) · · · N.
For everyp∈ N, consider a functionfp∈ E{L(p)}(R)such that|Dkfp(0)| ≥ L
(p) k ,
∀k ∈ N0.If{xp: p∈ N}is a dense subset ofR, consider f (x) =
+∞
X
p=1
fp(x− xp)Φp(x), x∈ R
whereΦpis a compactly supported function well chosen.
Notions of genericity Denjoy-Carleman classes
Construction
Definition
We say that a function isnowhere inE{M }if its restriction to any open and non-empty
subsetΩofRnever belongs toE{M }(Ω).
Proposition
Assume thatM andN are two weight sequences such thatM N. IfM is
non-quasianalytic, there exists a function ofE(N )(R)which is nowhere inE{M }.
Idea. Construct a sequence(L(p))
p∈Nof weight sequences such that
M L(1) L(2) · · · L(p) · · · N.
For everyp∈ N, consider a functionfp∈ E{L(p)}(R)such that|Dkfp(0)| ≥ L
(p) k ,
∀k ∈ N0.
If{xp: p∈ N}is a dense subset ofR, consider
f (x) =
+∞
X
p=1
fp(x− xp)Φp(x), x∈ R
Notions of genericity Denjoy-Carleman classes
Construction
Definition
We say that a function isnowhere inE{M }if its restriction to any open and non-empty
subsetΩofRnever belongs toE{M }(Ω).
Proposition
Assume thatM andN are two weight sequences such thatM N. IfM is
non-quasianalytic, there exists a function ofE(N )(R)which is nowhere inE{M }.
Idea. Construct a sequence(L(p))
p∈Nof weight sequences such that
M L(1) L(2) · · · L(p) · · · N.
For everyp∈ N, consider a functionfp∈ E{L(p)}(R)such that|Dkfp(0)| ≥ L
(p) k ,
∀k ∈ N0.If{xp: p∈ N}is a dense subset ofR, consider f (x) =
+∞
X
p=1
fp(x− xp)Φp(x), x∈ R
whereΦpis a compactly supported function well chosen.
Notions of genericity Denjoy-Carleman classes
Generic results
Proposition
Assume thatN andM are two weight sequences such thatM N. IfM is non
quasianalytic, the set of functions ofE(N )(R)which are nowhere inE{M }is
• prevalent,
• residual,
• c-dense-lineable.
More with countable unions
LetN be a weight sequence and let(M(n))
n∈Nbe a sequence of weight sequences
such thatM(n) N for everyn
∈ N. If there isn0∈ Nsuch that the weight
sequenceM(n0)is non quasianalytic, the set of functions of
E(N )(R)which are
nowhere inSn∈NE{M(n)}is prevalent, residual andc-dense-lineable inE(N )(R).
Idea. Construct a weight sequenceP such that
[
n∈N
Notions of genericity Denjoy-Carleman classes
Generic results
Proposition
Assume thatN andM are two weight sequences such thatM N. IfM is non
quasianalytic, the set of functions ofE(N )(R)which are nowhere inE{M }is
• prevalent,
• residual,
• c-dense-lineable.
Idea. The set of functions ofE(N )(R)which are nowhere inE{M }is the complement of
[ I⊆R [ m∈N [ s∈N f ∈ E(N )(R) : sup x∈I|D kf (x) | ≤ smkM k, ∀k ∈ N0 | {z }
closed set with empty interior
| {z }
proper linear subspace ofE(N )(R)
.
More with countable unions
LetN be a weight sequence and let(M(n))
n∈Nbe a sequence of weight sequences
such thatM(n) N for everyn
∈ N. If there isn0∈ Nsuch that the weight
sequenceM(n0)is non quasianalytic, the set of functions of
E(N )(R)which are
nowhere inS
n∈NE{M(n)}is prevalent, residual andc-dense-lineable inE(N )(R).
Idea. Construct a weight sequenceP such that
[
n∈N
E{M(n)}(Ω)⊆ E{P }(Ω) (E(N )(Ω).
Notions of genericity Denjoy-Carleman classes
Generic results
Proposition
Assume thatN andM are two weight sequences such thatM N. IfM is non
quasianalytic, the set of functions ofE(N )(R)which are nowhere inE{M }is
• prevalent,
• residual,
• c-dense-lineable.
More with countable unions
LetN be a weight sequence and let(M(n))
n∈Nbe a sequence of weight sequences
such thatM(n) N for everyn
∈ N. If there isn0∈ Nsuch that the weight
sequenceM(n0)is non quasianalytic, the set of functions of
E(N )(R)which are
nowhere inSn∈NE{M(n)}is prevalent, residual andc-dense-lineable inE(N )(R).
Idea. Construct a weight sequenceP such that
[
n∈N
Notions of genericity Denjoy-Carleman classes
Generic results
Proposition
Assume thatN andM are two weight sequences such thatM N. IfM is non
quasianalytic, the set of functions ofE(N )(R)which are nowhere inE{M }is
• prevalent,
• residual,
• c-dense-lineable.
Idea. Construct for everyt∈ (0, 1)a weight sequenceL(t)such that
M L(t) N and L(t) L(s)ift < s.
Then, we have for everyt∈ (0, 1)
M L(t2) L(2t3) L(3t4) · · · L(t) N
and we construct as before a function ofE(N )(R)which is nowhere inE{M }.
More with countable unions
LetN be a weight sequence and let(M(n))n∈Nbe a sequence of weight sequences
such thatM(n) N for everyn∈ N. If there isn0∈ Nsuch that the weight
sequenceM(n0)is non quasianalytic, the set of functions ofE
(N )(R)which are
nowhere inS
n∈NE{M(n)}is prevalent, residual andc-dense-lineable inE(N )(R).
Idea. Construct a weight sequenceP such that
[
n∈N
E{M(n)}(Ω)⊆ E{P }(Ω) (E(N )(Ω).
Notions of genericity Denjoy-Carleman classes
Generic results
Proposition
Assume thatN andM are two weight sequences such thatM N. IfM is non
quasianalytic, the set of functions ofE(N )(R)which are nowhere inE{M }is
• prevalent,
• residual,
• c-dense-lineable.
More with countable unions
LetN be a weight sequence and let(M(n))
n∈Nbe a sequence of weight sequences
such thatM(n) Nfor everyn
∈ N. If there isn0∈ Nsuch that the weight
sequenceM(n0)is non quasianalytic, the set of functions of
E(N )(R)which are
nowhere inSn∈NE{M(n)}is prevalent, residual andc-dense-lineable inE(N )(R).
Idea. Construct a weight sequenceP such that
[
n∈N
Notions of genericity Denjoy-Carleman classes
An important example of ultradifferentiable functions of Roumieu type is given by the
classes ofGevrey differentiable functionsof orderα > 1. They correspond to the
weight sequences
Mk := (k!)α, k∈ N0.
Particular case of Gevrey classes
Letα > 1. The set of functions ofE((k!)α)(R)which are nowhere inE{(k!)β}for every
β∈ (1, α), is prevalent, residual andc-dense-lineable inE((k!)α)(R).
It suffices to take the weight sequencesM(n)(n
∈ N) given by
Mk(n):= (k!)βn, k∈ N0,
where(βn)n∈Nis an increasing sequence of(1, α)that converges toα.
Proposition (Schmets, Valdivia, 1991)
Letα > 1. The set of functions ofE((k!)α)(R)which are nowhere inE{(k!)β}for every
β∈ (1, α)is residual inE((k!)α)(R).
Notions of genericity Denjoy-Carleman classes
An important example of ultradifferentiable functions of Roumieu type is given by the
classes ofGevrey differentiable functionsof orderα > 1. They correspond to the
weight sequences
Mk := (k!)α, k∈ N0.
Particular case of Gevrey classes
Letα > 1. The set of functions ofE((k!)α)(R)which are nowhere inE{(k!)β}for every
β∈ (1, α), is prevalent, residual andc-dense-lineable inE((k!)α)(R).
It suffices to take the weight sequencesM(n)(n
∈ N) given by
Mk(n):= (k!)βn, k∈ N0,
where(βn)n∈Nis an increasing sequence of(1, α)that converges toα.
Proposition (Schmets, Valdivia, 1991)
Letα > 1. The set of functions ofE((k!)α)(R)which are nowhere inE{(k!)β}for every
Notions of genericity Denjoy-Carleman classes
An important example of ultradifferentiable functions of Roumieu type is given by the
classes ofGevrey differentiable functionsof orderα > 1. They correspond to the
weight sequences
Mk := (k!)α, k∈ N0.
Particular case of Gevrey classes
Letα > 1. The set of functions ofE((k!)α)(R)which are nowhere inE{(k!)β}for every
β∈ (1, α), is prevalent, residual andc-dense-lineable inE((k!)α)(R).
It suffices to take the weight sequencesM(n)(n
∈ N) given by
Mk(n):= (k!)βn, k∈ N0,
where(βn)n∈Nis an increasing sequence of(1, α)that converges toα.
Proposition (Schmets, Valdivia, 1991)
Letα > 1. The set of functions ofE((k!)α)(R)which are nowhere inE{(k!)β}for every
β∈ (1, α)is residual inE((k!)α)(R).
Notions of genericity Denjoy-Carleman classes
Other results.
• Similar results have been obtained with classes of ultradifferentiable functions
defined using weight functions and weight matrices.
Perspectives.
• What about the algebrability?
• Other notions of genericity (such as porosity)?
Multifractal analysis
Content of the presentation.
1. Notions of genericity
a) Residuality, prevalence and lineability b) Denjoy-Carleman classes
2. Multifractal analysis
a) Hölder regularity and multifractal spectrum b) Multifractal formalism
c) Leaders profile method d) Lνspaces
Multifractal analysis Hölder regularity and multifractal spectrum
Hölder regularity and multifractal spectrum
Recall. Is it possible to characterize the local regularity of an irregular function?
1 -100 -80 -60 -40 -20 0 20 40 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.91 1 -25 -20 -15 -10 -5 0 5 10 15 20 25 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1 -500 0 500 1000 1500 2000 2500 3000 3500 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Multifractal analysis Hölder regularity and multifractal spectrum
Hölder regularity and multifractal spectrum
Recall. Is it possible to characterize the local regularity of an irregular function?
Definition
Letf : R→ Rbe a locally bounded function,α≥ 0andx∈ R. The functionf
belongs to theHölder spaceCα(x)if there exist a constantC > 0and a polynomial
Pof degree strictly smaller thanαsuch that
|f(y) − P (y)| ≤ C|y − x|α
for allyin a neighborhood ofx. Then, theHölder exponenthf(x)offatxis defined
by
hf(x) := sup{α ≥ 0 : f ∈ Cα(x)}.
Weierstraß function.hf(x) =−log alog b, ∀x ∈ R.
Multifractal analysis Hölder regularity and multifractal spectrum
Hölder regularity and multifractal spectrum
Recall. Is it possible to characterize the local regularity of an irregular function?
Definition
Letf : R→ Rbe a locally bounded function,α≥ 0andx∈ R. The functionf
belongs to theHölder spaceCα(x)if there exist a constantC > 0and a polynomial
Pof degree strictly smaller thanαsuch that
|f(y) − P (y)| ≤ C|y − x|α
for allyin a neighborhood ofx. Then, theHölder exponenthf(x)offatxis defined
by
hf(x) := sup{α ≥ 0 : f ∈ Cα(x)}.
Multifractal analysis Hölder regularity and multifractal spectrum
• Sincehf(x)can change widely from a point to another, we will characterize the
size of the sets of points which have the same local regularity. • Theiso-Hölder setsoff areEh:={x ∈ R : hf(x) = h}.
Definition
Themultifractal spectrumdf off is defined by
df(h) := dimHEh, ∀h ∈ [0, +∞],
with the convention thatdimH∅ = −∞.
−→ dfgives a geometrical idea about the distribution of the singularities off
Multifractal analysis Hölder regularity and multifractal spectrum
• Sincehf(x)can change widely from a point to another, we will characterize the
size of the sets of points which have the same local regularity. • Theiso-Hölder setsoff areEh:={x ∈ R : hf(x) = h}.
Definition
Themultifractal spectrumdf offis defined by
df(h) := dimHEh, ∀h ∈ [0, +∞],
with the convention thatdimH∅ = −∞.
Multifractal analysis Hölder regularity and multifractal spectrum
Examples
Riemann function 1 0 10− log2(1− p1)− log2(1− p2)1 − log2(p2) 2 − log2(p1)
Sum of two cascades
1 0 1 0 − log2(1− p)1 − log2(p) 2 Cascade 1 0 1
0 − log2(1− p) 1 γ− log2(p) htmax
Threshold of a cascade
Multifractal analysis Hölder regularity and multifractal spectrum
Examples
Riemann function 1 0 10− log2(1− p1)− log2(1− p2)1 − log2(p2) 2 − log2(p1)
Sum of two cascades
1 0 1 0 − log2(1− p)1 − log2(p) 2 Cascade 1 0 1
0 − log2(1− p) 1 γ− log2(p) htmax
Multifractal analysis Hölder regularity and multifractal spectrum
Examples
Riemann function 1
0 1
0− log2(1− p1)− log2(1− p2)1 − log2(p2) 2 − log2(p1)
Sum of two cascades
1 0 1 0 − log2(1− p)1 − log2(p) 2 Cascade 1 0 1
0 − log2(1− p) 1 γ− log2(p) htmax
Threshold of a cascade
Multifractal analysis Hölder regularity and multifractal spectrum
Examples
Riemann function 1
0 1
0− log2(1− p1)− log2(1− p2)1 − log2(p2) 2 − log2(p1)
Sum of two cascades
1 0 1 0 − log2(1− p)1 − log2(p) 2 Cascade 1 0 1
0 − log2(1− p) 1 γ− log2(p) htmax
Multifractal analysis Multifractal formalism
Multifractal formalism
Amultifractal formalismis a method which is expected to give the multifractal spectrum of a function, from “global” quantities which are numerically computable.
Several multifractal formalisms based on a decomposition off ∈ L2([0, 1])in a
wavelet basis f = X j∈N0 2j−1 X k=0 cj,kψj,k + C
have been proposed to estimatedf, where the mother waveletψbelongs toS(R).
Characterization of the Hölder exponent using wavelet coefficients
Iff is uniformly Hölder, the Hölder exponent off atxis
hf(x) = lim inf
j→+∞k∈{0,...,2infj−1}
log(|cj,k|)
log(2−j+|k2−j− x|).
Advantage. Easy to compute and relatively stable from a numerical point of view.
Multifractal analysis Multifractal formalism
Multifractal formalism
Amultifractal formalismis a method which is expected to give the multifractal spectrum of a function, from “global” quantities which are numerically computable.
Several multifractal formalisms based on a decomposition off ∈ L2([0, 1])in a
wavelet basis f = X j∈N0 2j−1 X k=0 cj,kψj,k + C
have been proposed to estimatedf, where the mother waveletψbelongs toS(R).
Characterization of the Hölder exponent using wavelet coefficients
Iff is uniformly Hölder, the Hölder exponent off atxis
hf(x) = lim inf
j→+∞k∈{0,...,2infj−1}
log(|cj,k|)
log(2−j+|k2−j− x|).
Multifractal analysis Multifractal formalism
Multifractal formalism
Amultifractal formalismis a method which is expected to give the multifractal spectrum of a function, from “global” quantities which are numerically computable.
Several multifractal formalisms based on a decomposition off ∈ L2([0, 1])in a
wavelet basis f = X j∈N0 2j−1 X k=0 cj,kψj,k + C
have been proposed to estimatedf, where the mother waveletψbelongs toS(R).
Characterization of the Hölder exponent using wavelet coefficients
Iff is uniformly Hölder, the Hölder exponent off atxis
hf(x) = lim inf
j→+∞k∈{0,...,2infj−1}
log(|cj,k|)
log(2−j+|k2−j− x|).
Advantage. Easy to compute and relatively stable from a numerical point of view.
Multifractal analysis Multifractal formalism
• The Frisch-Parisi formalism (1985) and the classical use of Besov spaces lead to
a loss of information (only concave hull and increasing part of spectra can be recovered).
• Wavelet leaders method (S. Jaffard, 2004): Modification of the Frisch-Parisi
formalism using the wavelet leaders of the function instead of wavelet coefficients.
−→Detection of increasing and decreasing parts of concave spectra.
• Introduction of spaces of typeSν (J.M. Aubry, S. Jaffard, 2005), based on
histograms of wavelet coefficients.
−→Detection of concave and non-concave parts of increasing spectra.
• Combination of the two previous methods to obtain theleaders profile method
and the spaces of typeLν.
−→Detection of increasing and decreasing parts of concave and non-concave
Multifractal analysis Multifractal formalism
• The Frisch-Parisi formalism (1985) and the classical use of Besov spaces lead to
a loss of information (only concave hull and increasing part of spectra can be recovered).
• Wavelet leaders method (S. Jaffard, 2004): Modification of the Frisch-Parisi
formalism using the wavelet leaders of the function instead of wavelet coefficients.
−→Detection of increasing and decreasing parts of concave spectra.
• Introduction of spaces of typeSν (J.M. Aubry, S. Jaffard, 2005), based on
histograms of wavelet coefficients.
−→Detection of concave and non-concave parts of increasing spectra.
• Combination of the two previous methods to obtain theleaders profile method
and the spaces of typeLν.
−→Detection of increasing and decreasing parts of concave and non-concave
spectra.
Multifractal analysis Multifractal formalism
• The Frisch-Parisi formalism (1985) and the classical use of Besov spaces lead to
a loss of information (only concave hull and increasing part of spectra can be recovered).
• Wavelet leaders method (S. Jaffard, 2004): Modification of the Frisch-Parisi
formalism using the wavelet leaders of the function instead of wavelet coefficients.
−→Detection of increasing and decreasing parts of concave spectra.
• Introduction of spaces of typeSν (J.M. Aubry, S. Jaffard, 2005), based on
histograms of wavelet coefficients.
−→Detection of concave and non-concave parts of increasing spectra.
• Combination of the two previous methods to obtain theleaders profile method
and the spaces of typeLν.
−→Detection of increasing and decreasing parts of concave and non-concave
Multifractal analysis Multifractal formalism
• The Frisch-Parisi formalism (1985) and the classical use of Besov spaces lead to
a loss of information (only concave hull and increasing part of spectra can be recovered).
• Wavelet leaders method (S. Jaffard, 2004): Modification of the Frisch-Parisi
formalism using thewavelet leadersof the function instead of wavelet coefficients.
−→Detection of increasing and decreasing parts of concave spectra.
• Introduction of spaces of typeSν (J.M. Aubry, S. Jaffard, 2005), based on
histograms of wavelet coefficients.
−→Detection of concave and non-concave parts of increasing spectra.
• Combination of the two previous methods to obtain theleaders profile method
and the spaces of typeLν.
−→Detection of increasing and decreasing parts of concave and non-concave
spectra.
Multifractal analysis Multifractal formalism
• The Frisch-Parisi formalism (1985) and the classical use of Besov spaces lead to
a loss of information (only concave hull and increasing part of spectra can be recovered).
• Wavelet leaders method (S. Jaffard, 2004): Modification of the Frisch-Parisi
formalism using thewavelet leadersof the function instead of wavelet coefficients.
−→Detection of increasing and decreasing parts of concave spectra.
• Introduction of spaces of typeSν (J.M. Aubry, S. Jaffard, 2005), based on
histogramsof wavelet coefficients.
−→Detection of concave and non-concave parts of increasing spectra.
• Combination of the two previous methods to obtain theleaders profile method
and the spaces of typeLν.
−→Detection of increasing and decreasing parts of concave and non-concave
Multifractal analysis Multifractal formalism
• The Frisch-Parisi formalism (1985) and the classical use of Besov spaces lead to
a loss of information (only concave hull and increasing part of spectra can be recovered).
• Wavelet leaders method (S. Jaffard, 2004): Modification of the Frisch-Parisi
formalism using thewavelet leadersof the function instead of wavelet coefficients.
−→Detection of increasing and decreasing parts of concave spectra.
• Introduction of spaces of typeSν (J.M. Aubry, S. Jaffard, 2005), based on
histogramsof wavelet coefficients.
−→Detection of concave and non-concave parts of increasing spectra.
• Combination of the two previous methods to obtain theleaders profile method
and the spaces of typeLν.
−→Detection of increasing and decreasing parts of concave and non-concave
spectra.
Multifractal analysis Multifractal formalism
Wavelet leaders
Standard notation. Forj∈ N0, k∈0, . . . , 2j− 1 ,
λ(j, k) :=x ∈ R : 2jx− k ∈ [0, 1[ = k 2j, k + 1 2j ,
and for allj∈ N0,Λjdenotes the set of all dyadic intervals (of[0, 1)) of length2−j.
Ifλ = λ(j, k), we use both notationscj,korcλto denote the wavelet coefficients.
Definition
Thewavelet leadersof a functionf ∈ L2([0, 1])are defined by
dλ:= sup λ0⊆3λ|cλ
0|, λ ∈ Λj, j∈ N0.
Multifractal analysis Multifractal formalism
Wavelet leaders
Standard notation. Forj∈ N0, k∈0, . . . , 2j− 1 ,
λ(j, k) :=x ∈ R : 2jx− k ∈ [0, 1[ = k 2j, k + 1 2j ,
and for allj∈ N0,Λjdenotes the set of all dyadic intervals (of[0, 1)) of length2−j.
Ifλ = λ(j, k), we use both notationscj,korcλto denote the wavelet coefficients.
Definition
Thewavelet leadersof a functionf ∈ L2([0, 1])are defined by
dλ:= sup λ0⊆3λ|cλ
0|, λ ∈ Λj, j∈ N0.
−→their decay properties are directly related with the Hölder exponent.
Multifractal analysis Multifractal formalism
Ifx∈ [0, 1), letλj(x)denote the dyadic interval of length2−j which containsx.
Hölder regularity and wavelet leaders
Iff is uniformly Hölder, the Hölder exponent off atxis given by
hf(x) = lim inf j→+∞ log dλj(x) log 2−j . Interpretation. dλj(x)∼ 2 −hf(x)j
Multifractal analysis Multifractal formalism
Ifx∈ [0, 1), letλj(x)denote the dyadic interval of length2−j which containsx.
Hölder regularity and wavelet leaders
Iff is uniformly Hölder, the Hölder exponent off atxis given by
hf(x) = lim inf j→+∞ log dλj(x) log 2−j . Interpretation. dλj(x)∼ 2 −hf(x)j
Multifractal analysis Multifractal formalism
Ifx∈ [0, 1), letλj(x)denote the dyadic interval of length2−j which containsx.
Hölder regularity and wavelet leaders
Iff is uniformly Hölder, the Hölder exponent off atxis given by
hf(x) = lim inf j→+∞ log dλj(x) log 2−j . Interpretation. dλj(x)∼ 2 −hf(x)j
Multifractal analysis Multifractal formalism
Ifx∈ [0, 1), letλj(x)denote the dyadic interval of length2−j which containsx.
Hölder regularity and wavelet leaders
Iff is uniformly Hölder, the Hölder exponent off atxis given by
hf(x) = lim inf j→+∞ log dλj(x) log 2−j . Interpretation. dλj(x)∼ 2 −hf(x)j
Multifractal analysis Multifractal formalism
Method based on
S
νspaces
Thewavelet profileνf of a locally bounded functionfis defined for everyh≥ 0by
νf(h) := lim
ε→0+lim supj→+∞
log #λ ∈ Λj : |cλ| ≥ 2−(h+ε)j
log 2j .
Interpretation.
• There are approximatively2νf(h)j coefficients greater in modulus than2−hj.
Properties.
• νf is a right-continuous increasing function.
• νf is independent of the chosen wavelet basis.
• Iff is uniformly Hölder, df(h)≤ dνf(h) := min ( h sup h0∈(0,h] νf(h0) h0 , 1 ) , ∀h ≥ 0.
Multifractal analysis Multifractal formalism
Method based on
S
νspaces
Thewavelet profileνf of a locally bounded functionfis defined for everyh≥ 0by
νf(h) := lim
ε→0+lim supj→+∞
log #λ ∈ Λj : |cλ| ≥ 2−(h+ε)j
log 2j .
Interpretation.
• There are approximatively2νf(h)j coefficients greater in modulus than2−hj.
Properties.
• νf is a right-continuous increasing function.
• νf is independent of the chosen wavelet basis.
• Iff is uniformly Hölder, df(h)≤ dνf(h) := min ( h sup h0∈(0,h] νf(h0) h0 , 1 ) , ∀h ≥ 0.
Multifractal analysis Multifractal formalism
Definition
Take0≤ a < b ≤ +∞. A functiong : [a, b]7→ [0, +∞)is withincreasing-visibilityifg
is continuous ataandsupy∈(a,x]
g(y)
y ≤
g(x)
x for allx∈ (a, b].
In other words, a functiongis with increasing-visibility if for allx∈ (a, b], the segment
[(0, 0), (x, g(x))]lies above the graph ofgon(a, x]. 1
0 1
0
Example ofνf (---) anddνf (—)
−→The passage fromνftodνf transforms the functionνfinto a function with
Multifractal analysis Multifractal formalism
Definition
Take0≤ a < b ≤ +∞. A functiong : [a, b]7→ [0, +∞)is withincreasing-visibilityifg
is continuous ataandsupy∈(a,x]
g(y)
y ≤
g(x)
x for allx∈ (a, b].
In other words, a functiongis with increasing-visibility if for allx∈ (a, b], the segment
[(0, 0), (x, g(x))]lies above the graph ofgon(a, x]. 1
0 1
0
Example ofνf (---) anddνf (—)
−→The passage fromνftodνf transforms the functionνfinto a function with
increasing-visibility.
Multifractal analysis Multifractal formalism
Definition
Take0≤ a < b ≤ +∞. A functiong : [a, b]7→ [0, +∞)is withincreasing-visibilityifg
is continuous ataandsupy∈(a,x]
g(y)
y ≤
g(x)
x for allx∈ (a, b].
In other words, a functiongis with increasing-visibility if for allx∈ (a, b], the segment
[(0, 0), (x, g(x))]lies above the graph ofgon(a, x]. 1
0 1
0
Example ofνf (---) anddνf (—)
−→The passage fromνftodνf transforms the functionνfinto a function with
Multifractal analysis Multifractal formalism
Definition
Take0≤ a < b ≤ +∞. A functiong : [a, b]7→ [0, +∞)is withincreasing-visibilityifg
is continuous ataandsupy∈(a,x]
g(y)
y ≤
g(x)
x for allx∈ (a, b].
In other words, a functiongis with increasing-visibility if for allx∈ (a, b], the segment
[(0, 0), (x, g(x))]lies above the graph ofgon(a, x]. 1
0 1
0
Example ofνf (---) anddνf (—)
−→The passage fromνftodνf transforms the functionνfinto a function with
increasing-visibility.
Multifractal analysis Multifractal formalism
Definition
Take0≤ a < b ≤ +∞. A functiong : [a, b]7→ [0, +∞)is withincreasing-visibilityifg
is continuous ataandsupy∈(a,x]
g(y)
y ≤
g(x)
x for allx∈ (a, b].
In other words, a functiongis with increasing-visibility if for allx∈ (a, b], the segment
[(0, 0), (x, g(x))]lies above the graph ofgon(a, x]. 1
0 1
0 hmax
Example ofνf (---) anddνf (—)
−→The passage fromνftodνf transforms the functionνfinto a function with
Multifractal analysis Multifractal formalism
Particular case
Assumption. Assume that the wavelet coefficients off are given bycλ= µ(λ)where
µis a finite Borel measure on[0, 1].
Notation. Letfβdenote the function with wavelet coefficients given byc β
λ = 2−βjcλ.
In this case, one has
• dfβ(h) = df(h− β)for allh≥ β. • νfβ(h) = νf(h− β)for allh≥ β. Moreover, if inf νf(x)− νf(y) x− y : x, y∈ [hmin, h 0 max], x < y > 0,
wherehmin= inf{α : νf(α)≥ 0},h0max= inf{α : νf(α) = 1}, then there exists
β > 0such that the functionνfβ is with increasing-visibility on[hmin, h0max]. In this
case,dνfβ = ν
fβ approximatesdfβ. Therefore the increasing part ofdfcan be
approximated byνf.
Multifractal analysis Multifractal formalism
Particular case
Assumption. Assume that the wavelet coefficients off are given bycλ= µ(λ)where
µis a finite Borel measure on[0, 1].
Notation. Letfβdenote the function with wavelet coefficients given byc β
λ = 2−βjcλ.
In this case, one has
• dfβ(h) = df(h− β)for allh≥ β. • νfβ(h) = νf(h− β)for allh≥ β. Moreover, if inf νf(x)− νf(y) x− y : x, y∈ [hmin, h 0 max], x < y > 0,
wherehmin= inf{α : νf(α)≥ 0},h0max= inf{α : νf(α) = 1}, then there exists
β > 0such that the functionνfβ is with increasing-visibility on[hmin, h0max]. In this
case,dνfβ = ν
fβ approximatesdfβ. Therefore the increasing part ofdfcan be
Multifractal analysis Multifractal formalism
Particular case
Assumption. Assume that the wavelet coefficients off are given bycλ= µ(λ)where
µis a finite Borel measure on[0, 1].
Notation. Letfβdenote the function with wavelet coefficients given byc β
λ = 2−βjcλ.
In this case, one has
• dfβ(h) = df(h− β)for allh≥ β. • νfβ(h) = νf(h− β)for allh≥ β. Moreover, if inf νf(x)− νf(y) x− y : x, y∈ [hmin, h 0 max], x < y > 0,
wherehmin= inf{α : νf(α)≥ 0},h0max= inf{α : νf(α) = 1}, then there exists
β > 0such that the functionνfβ is with increasing-visibility on[hmin, h0max]. In this
case,dνfβ = ν
fβ approximatesdfβ. Therefore the increasing part ofdfcan be
approximated byνf.
Multifractal analysis Multifractal formalism
Multifractal analysis Leaders profile method
Wavelet leaders density
Thewavelet leaders densityoffis defined for everyh≥ 0by
e
ρf(h) := lim
ε→0+lim supj→+∞
log #λ ∈ Λj: 2−(h+ε)j ≤ dλ< 2−(h−ε)j
log 2j .
Interpretation. There are approximatively2ρef(h)jcoefficients of size2−hj.
Heuristic argument. We consider the pointsxsuch thathf(x) = h.
• dλj(x)∼ 2−hjand there are about2eρf(h)j such dyadic intervals.
• If we cover each singularityxby dyadic intervals of size2−j, from the definition of
the Hausdorff dimension, there are about2df(h)j such intervals.
=⇒ρef(h) = df(h)
Problems.
• The wavelet leaders density may depend on the chosen wavelet basis.
• The definition of the wavelet leaders density is numerically extremely unstable.
Multifractal analysis Leaders profile method
Wavelet leaders density
Thewavelet leaders densityoffis defined for everyh≥ 0by
e
ρf(h) := lim
ε→0+lim supj→+∞
log #λ ∈ Λj: 2−(h+ε)j ≤ dλ< 2−(h−ε)j
log 2j .
Interpretation. There are approximatively2ρef(h)jcoefficients of size2−hj.
Heuristic argument. We consider the pointsxsuch thathf(x) = h.
• dλj(x)∼ 2−hjand there are about2eρf(h)j such dyadic intervals.
• If we cover each singularityxby dyadic intervals of size2−j, from the definition of
the Hausdorff dimension, there are about2df(h)j such intervals.
=⇒ρef(h) = df(h)
Problems.
• The wavelet leaders density may depend on the chosen wavelet basis.
Multifractal analysis Leaders profile method
Wavelet leaders density
Thewavelet leaders densityoffis defined for everyh≥ 0by
e
ρf(h) := lim
ε→0+lim supj→+∞
log #λ ∈ Λj: 2−(h+ε)j ≤ dλ< 2−(h−ε)j
log 2j .
Interpretation. There are approximatively2ρef(h)jcoefficients of size2−hj.
Heuristic argument. We consider the pointsxsuch thathf(x) = h.
• dλj(x)∼ 2−hjand there are about2eρf(h)j such dyadic intervals.
• If we cover each singularityxby dyadic intervals of size2−j, from the definition of
the Hausdorff dimension, there are about2df(h)j such intervals.
=⇒ρef(h)≥df(h)
Problems.
• The wavelet leaders density may depend on the chosen wavelet basis.
• The definition of the wavelet leaders density is numerically extremely unstable.
Multifractal analysis Leaders profile method
Wavelet leaders density
Thewavelet leaders densityoffis defined for everyh≥ 0by
e
ρf(h) := lim
ε→0+lim supj→+∞
log #λ ∈ Λj: 2−(h+ε)j ≤ dλ< 2−(h−ε)j
log 2j .
Interpretation. There are approximatively2ρef(h)jcoefficients of size2−hj.
Heuristic argument. We consider the pointsxsuch thathf(x) = h.
• dλj(x)∼ 2−hjand there are about2eρf(h)j such dyadic intervals.
• If we cover each singularityxby dyadic intervals of size2−j, from the definition of
the Hausdorff dimension, there are about2df(h)j such intervals.
=⇒ρef(h)≥df(h)
Problems.
• The wavelet leaders density may depend on the chosen wavelet basis.
Multifractal analysis Leaders profile method
Wavelet leaders profile
Lethsbe the smallest positive real number such thatρef(hs) = 1. Thewavelet leaders
profileoff is defined by e νf(h) := lim ε→0+lim supj→+∞ log #λ ∈ Λj : dλ≥ 2−(h+ε)j log 2j ifh≤ hs, lim ε→0+lim supj→+∞ log #λ ∈ Λj : dλ≤ 2−(h−ε)j log 2j ifh≥ hs. Properties. • e
νf is independent of the chosen wavelet basis.
•
e
νf takes values in{−∞} ∪ [0, 1], it is increasing and right-continuous on[0, hs],
decreasing and left-continuous on[hs, +∞),eνf(hs) = 1and the function
h∈ [hs, +∞) 7→ e
νf(h)− 1
h
is decreasing.
• Moreover, any functionνwhich satisfies these properties is the wavelet leaders
profile of a function.
Multifractal analysis Leaders profile method
Wavelet leaders profile
Lethsbe the smallest positive real number such thatρef(hs) = 1. Thewavelet leaders
profileoff is defined by e νf(h) := lim ε→0+lim supj→+∞ log #λ ∈ Λj : dλ≥ 2−(h+ε)j log 2j ifh≤ hs, lim ε→0+lim supj→+∞ log #λ ∈ Λj : dλ≤ 2−(h−ε)j log 2j ifh≥ hs. Properties. • e
νf is independent of the chosen wavelet basis.
•
e
νf takes values in{−∞} ∪ [0, 1], it is increasing and right-continuous on[0, hs],
decreasing and left-continuous on[hs, +∞),eνf(hs) = 1and the function
h∈ [hs, +∞) 7→ e
νf(h)− 1
h
is decreasing.
• Moreover, any functionνwhich satisfies these properties is the wavelet leaders
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Leaders profile method
Leaders profile method. It is based on the estimation of the multifractal spectrumdf
off by the functioneνf.
Results.
• Our method allows to detect some multifractal spectra thatall other methods
proposed were not able to detect;
• It gives the correct multifractal spectrum for somespecific functions;
• It always gives anupper boundfor the multifractal spectrum;
• From a theoretical point of view, it gives as good results as the wavelet leaders
method in the concave case, andbetter results in the non-concave case;
• From a theoretical point of view, it gives better results than the method based on
theSν spaces and in particular, it allows to detect spectra which arenot with
increasing visibility.
• Animplementationof this method has been proposed and tested on several examples.
Multifractal analysis Lνspaces
L
νspaces
Letνbe a function which has the same properties as any wavelet leaders profile.
Definition
ThespaceLνis the set of functionsf ∈ L2([0, 1])such that
e νf ≤ ν.
This space has been endowed with a complete metrizable topology.
Results. If there isαmin> 0such thatν(α) =−∞ifα < αmin, then
• Lνis also separable;
• The set of functionsf such thateνf = νis residual and dense-lineable inLν.
Perspectives.
• Generic validity of the leaders profile method;
Multifractal analysis Lνspaces
L
νspaces
Letνbe a function which has the same properties as any wavelet leaders profile.
Definition
ThespaceLνis the set of functionsf ∈ L2([0, 1])such that
e νf ≤ ν.
This space has been endowed with a complete metrizable topology.
Results. If there isαmin> 0such thatν(α) =−∞ifα < αmin, then
• Lνis also separable;
• The set of functionsf such thateνf = νis residual and dense-lineable inLν.
Perspectives.
• Generic validity of the leaders profile method;
• More with oscillating singularities.
Multifractal analysis Lνspaces
L
νspaces
Letνbe a function which has the same properties as any wavelet leaders profile.
Definition
ThespaceLνis the set of functionsf ∈ L2([0, 1])such that
e νf ≤ ν.
This space has been endowed with a complete metrizable topology.
Results. If there isαmin> 0such thatν(α) =−∞ifα < αmin, then
• Lνis also separable;
• The set of functionsf such thateνf = νis residual and dense-lineable inLν.
Perspectives.
• Generic validity of the leaders profile method;