HAL Id: hal-03107403
https://hal.archives-ouvertes.fr/hal-03107403v2
Submitted on 6 Jul 2021
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
2D random magnetic Laplacian with white noise magnetic field
Léo Morin, Antoine Mouzard
To cite this version:
Léo Morin, Antoine Mouzard. 2D random magnetic Laplacian with white noise mag- netic field. Stochastic Processes and their Applications, Elsevier, 2022, 142, pp.160-184.
�10.1016/j.spa.2021.10.004�. �hal-03107403v2�
2D random magnetic Laplacian with white noise magnetic field
Léo MORIN and Antoine MOUZARD
Abstract
We define the random magnetic Laplacian with spatial white noise as magnetic field on the two-dimensional torus using paracontrolled calculus. It yields a random self-adjoint operator with pure point spectrum and domain a random subspace of nonsmooth functions in L
2. We give sharp bounds on the eigenvalues which imply an almost sure Weyl-type law.
MSC 2020–35J10; 60H25; 47A52
Keywords –Magnetic Laplacian; White noise; Paracontrolled calculus; Spectral theory.
Contents
1 Definition of the operator 5
1.1 Construction of the domain . . . . 5
1.2 First properties of H . . . . 7
1.3 Form comparison between H and ∆ . . . 10
2 Self-adjointness and spectrum 11 2.1 The operator is symmetric . . . 11
2.2 The operator is self-adjoint . . . 12
2.3 H is the resolvent-limit of H
ε. . . 13
2.4 Comparison between the spectrum of H and ∆ . . . 14 3 Stochastic renormalisation of the singular potential 15
A Paracontrolled calculus 17
Introduction
The magnetic Laplacian associated to a magnetic potential A in two dimensions H = (i∂
1+ A
1)
2+ (i∂
2+ A
2)
2is of interest in the description of a number of physical models. For example, it describes the behavior of a particule in a magnetic field B related to A via
B = ∇ × A = ∂
2A
1− ∂
1A
2.
While the case of constant magnetic field has been largely studied, the analysis of the magnetic Laplacian with nonconstant magnetic field gives rise to a number of interesting questions. Apart from the analogy with the electric Laplacian −∆ + V , the spectral study of the magnetic Laplacian is also motivated by supraconductivity, where H plays a specific role in the third critical field of Ginzburg-Landau theory, see [7]. This work is dedicated to the study of the magnetic Laplacian with random singular magnetic field given by the space white noise B = ξ on the two-dimensional torus T
2. It can be constructed as a distribution with independant random Fourier coefficients with centered normal law of unit variance. In two dimensions, the space white noise belongs almost surely to the Sobolev spaces H
−1−κor the Besov-Hölder spaces C
−1−κfor any κ > 0 . Since the associated potential A satisfies the equation
ξ = ∂
2A
1− ∂
1A
2,
each component A
1, A
2is expected to belong to C
−κfor any κ > 0 . Thus the potential is not a measurable function and the associated magnetic Laplacian falls out of the range of the classical theory: It is a singular random operator. The study of PDEs or operators where rough stochastic terms give rise to ill-defined operations such as product of distributions has intensified since the 2010 ’s with the parallel introduction of regularity structures in [10] by Hairer and the paracontrolled calculus in [8] by Gubinelli, Imkeller and Perkowski. A famous example of such singular operator is the Anderson Hamiltonian
−∆ + ξ
which naturally appears in the parabolic Anderson model describing random walks in random environment, see [13] for a survey on the topic. The random operator introduced and studied in this work is the magnetic analog of the Anderson Hamiltonian with a first order perturbation of the Laplacian.
The class of singular stochastic PDEs consists of equations with rough stochastic terms which give rise to ill-defined terms, classicaly products of distributions. Examples are given by the KPZ equation
(∂
t− ∂
x2)u = ζ + (∂
xu)
2with ζ a spacetime white noise in dimension 1 + 1 or the Φ
4dequation (∂
t− ∆)φ = −φ
3+ ξ
with ξ a space white noise in dimension d ≥ 2 . These equations do not make sense a priori since solutions are expected to be too irregular for the terms (∂
xu)
2or φ
3to make sense. The recently developped approach consists in the construction of a random subspace of classical function spaces through a stochastic renormalisation procedure from the noise in which one can make sense almost surely of the equation. The different approaches of regularity structures and paracontrolled calculus differ by the tools used to study the singular products. In the first one, distributions are locally described by generalised Taylor expansions while the second one relies on tools from harmonic analysis with a description of distributions by frequency “schemes”. In both approaches, solutions are described by a richer stochastic object than the noise called the enhanced noise, the renormalisation procedure corresponds to its construction. This idea takes root in the theory of Lyons’ rough paths and Gubinelli’s controlled paths developped as a pathwise approach to stochastic integration. This has been used to study the Anderson Hamiltonian
−∆ + ξ
which is singular in dimension d ≥ 2 , the first time by Allez and Chouk using paracontrolled calculus on the two-dimensional torus, see [1]. See also [9, 16, 18] for more general settings including dimension d ∈ {2, 3} , different boundary conditions or a more general geometrical framework.
Following the approach for singular SPDEs, the idea is to construct a random subspace D
Ξof L
2as domain for the operator given an enhancement Ξ of the noise ξ such that (H, D
Ξ) is a well-defined unbounded operator in L
2. The singularity is dealed with through a renormalisation procedure in the construction of the enhanced noise Ξ . In this work, we construct such a domain for the random magnetic Laplacian with white noise magnetic field.
Since the Anderson Hamiltonian appears, at least formally, as the continuum limit of discrete Anderson Hamiltonians with independant and identically distributed random potentials with finite variance, the random magnetic Laplacian naturally describes the formal continuous limit of Hamil- tonians with magnetic random fields. It is also of interest to the study of different PDEs as for the deterministic magnetic Laplacian. For example, it can be used to solve dispersive PDEs like the Schrödinger equation
∂
tu = (i∂
1+ A
1)
2u + (i∂
2+ A
2)
2u − |u|
2u
or the wave equation. In particular, see [19] where Strichartz inequalities are obtained for the Anderson Hamiltonian as well as the random magnetic Laplacian constructed in the present work.
Main results
The 2D magnetic Laplacian with potential A : T
2→ R
2is formally given by H = −∆ + 2iA · ∇ + A · A + i div A
with associated magnetic field
B = ∇ × A.
From the physical point of view, one considers a given magnetic field B : T
2→ R and chooses a potential A satisfying this equation. Different choices of A yield different operators H
A, this is the choice of gauge. However, the physical quantities are expected to be independent of this choice and there is a partial invariance by gauge change. Indeed, if A : T
2→ R
2is a given potential associated to B , then the new potential A + ∇f is still a potential associated to B for functions f : T
2→ R.
The two operators are related formally by the relation H
A+∇f= e
ifH
Ae
−ifhence the spectral properties of H
Aand H
A+∇fare the same. While on R
2, every pair (A, A
0) of smooth potential that satifies
∇ × A = ∇ × A
0are related by
A = A
0+ ∇f,
this is no longer the case on T
2due to topological holes. One can show that the choice of a potential A can influence the spectrum of H
A, see for example the Aharonov-Bohm effect in [12]. In this work, we consider the magnetic field B = ξ the space white noise and we choose the Lorentz gauge defined as follows. Take
A = ∇
⊥ϕ where ϕ := ∆
−1ξ.
Then A satifies indeed
ξ = ∇ × A
with the addition of div (A) = 0 . Our method could also deal with a gauge A + ∇f with A the Lorentz gauge and f a smooth deterministic function, see Section 3 for a discussion about this. It would be interesting to study the influence of a more general gauge but the implict definition
B = ∇ × A
makes more difficult the study of singularities. This leaves us with H = −∆ + 2iA · ∇ + A · A
where A ∈ C
α−1almost surely for any α < 1 . The term A · A is singular and one has to give a meaning to it using probabilistic arguments. This is done in Section 3 and yields the associated enhanced potential
A = (A, A
2) ∈ C
α−1( T
2, R
2) × C
2α−2( T
2, R ).
Remark that since A is a distribution of negative Hölder regularity, the singular product A · A is expected to worsen the regularity. Given such an enhanced potential A , we construct in Section 1 a dense subspace D
A⊂ L
2such that
u ∈ D
A⊂ L
2= ⇒ Hu ∈ L
2.
In Section 2, we show that (H, D
A) is almost surely a self-adjoint operator with pure point spectrum.
We also prove that it is the resolvent-limit of
H
ε= −∆ + 2iA
ε· ∇ + A
2εfor any regularisation A
ε= (A
ε, A
2ε) ∈ C
∞( T
2, R
2) × C
∞( T
2, R ) such that
ε→0
lim kA − A
εk
Cα−2+ kA
2− A
2εk
C2α−2= 0.
Finally, we construct in Section 3 the enhanced potential A associated to the random magnetic field B = ξ . In particular, it is described by the limit in probability of
A
ε, A
2ε− c
εas ε goes to 0 where A
εis a regularisation of A and
c
ε= E
A
ε(0) · A
ε(0) .
In particular, the almost sure singularity of the product A · A implies the need to substract a
diverging constant c
εas ε goes to 0 and a singular random operator has to be interpreted as the
description of the limiting behavior of a diverging system. One is interested in the fluctuations of
this system in this diverging frame, as the central limit Theorem for a simple random walk. For the case of the Anderson Hamiltonian, see the work [17] of Martin and Perkowski for a nice example.
Our construction is done in the same spirit as in [18] for the Anderson Hamiltonian with the additional difficulty of the first order perturbation here. In particular, this work illustrates that the heat paracontrolled calculus deals naturally with such terms and its flexibility to solve singular stochastic PDEs. In particular, this shows that the method could be used to deal with a general class of operators of the form
−∆ + a
1· ∇ + a
2with rough stochastic fields a
1, a
2and therefore associated time-dependant PDEs. A very important example of such operators in probability corresponds to the case a
2= 0 where one recovers the infinitesimal generator of an SDE with drift a
1. For example, it makes possible the study of processes as the Brox diffusion in a more general framework, see for example [15] by Kremp and Perkowski. This is current investigation and will be done in a subsequent paper. Finally since we only define and prove self-adjointness of the operator, no major differences appear in the study of the Anderson Hamiltonian or the random magnetic Laplacian with white noise magnetic field. As one does for smooth potentials, it is natural to pursue the study of both operators to understand their differences. For example, an interesting first question it to investigate the asymptotics of the eigenvalues as the volume of the torus goes to infinity.
In both the Anderson and the magnetic case, it would be interesting to study the influence of a scaling λ > 0 in the noise. For instance, the behavior of
−∆ + λξ
as λ to 0 should not be a problem however the limit λ to ∞ is much more delicate and of great interest. It is equivalent to the large-volume limit obtained by taking a torus of large diameter √
λ . It was studied for example in [5] by Chouk and Van Zuijlen using the large deviation principle, see also [14] for motivations from Probability. This is also equivalent to the semiclassical limit h goes to 0 in
−h∆ + ξ.
For regular potentials V , the semiclassical limit is deeply related to the classical dynamics described by the Hamiltonian
(x, v) ∈ T
∗( T
2) 7→ |v|
2+ V (x)
with T
∗( T
2) the cotangent bundle of T
2. Similarily, would it be possible to find an effective
“classical” equation describing the semiclassical limit in the case of a singular potential V = ξ or B = ξ ? See the book [12] from Helffer for details on the semiclassical limit and the work [6]
by Dumaz and Labbé for a very precise description of the asymptotics in one dimension for the Anderson Hamiltonian.
Organisation of the paper
In the first section, we construct the domain and prove density in L
2. We compare the graph norm and the natural norms of the domain which gives closedness of the operator. We also give an explicit form comparison between the random magnetic Laplacian H and the Laplacian ∆ . In the second section, we show that the operator is symmetric as a weak limit of the regularised operator.
The form comparison of H and ∆ with the Babuška-Lax-Milgram Theorem gives self-adjointness.
Finally, we show that H is the resolvent-limit of the regularised operator H
εin Proposition 2.4 and compare the spectrum of H and ∆ . In particular, this implies an almost sure Weyl-type law for the random magnetic Laplacian. The third Section deals with the construction of the enhanced potential A built from the noise ξ through a renormalisation procedure. We gather in Appendix A all the results we need for the paracontrolled calculus and refer to [18] for the details. In particular, P and Π respectively denotes the paraproduct and the resonent term such that
f g = P
fg + P
gf + Π(f, g) and P e is a paraproduct intertwinned with P via the relation
∆ ◦ P e = P ◦ ∆
and satisfies the same continuity properties as P . We also denote
∆
−1:= − Z
10
e
t∆dt
an inverse for the Laplacian up to the regularising operator e
∆.
1 – Definition of the operator
In this section, we first construct the domain and show that its natural norms are equivalent to the graph norm of H . In particular, this guarantees the closedness of the operator. Finally, we compare the respective forms associated to H and L .
1.1 – Construction of the domain
Fix α ∈ (
23, 1) and let A be an enhanced magnetic potential
A = (A, A
2) ∈ X
α:= C
α−1( T
2, R
2) × C
2α−2( T
2, R ) with its natural norm
kAk
Xα:= kAk
Cα−1+ kA
2k
C2α−2.
For A ∈ L
∞, the term A
2can be interpreted as A·A while it is not defined if A is only a distribution.
It is enhanced in the sense that one does not have a natural interpretation of A · A for singular potential, this is specified by the additional data A
2. Section 3 is devoted to the particular case of magnetic white noise where A
2is constructed through a probabilistic renormalisation procedure.
Thus we refer as noise-dependent a quantity that depends on this enhanced potential A . For any regular function u ∈ C
∞( T
2) , we have
2iA · ∇u + A
2u = P
∇u2iA + P
uA
2+ (])
where (]) ∈ C
∞( T
2) with P
u2iA ∈ H
α−1and P
∇uA
2∈ H
2α−2. Assuming Hu ∈ L
2yields
−∆u = Hu − 2iA · ∇u + A
2u ∈ H
2α−2since 2α−2 < α−1 hence u is expected to belong to H
2αby elliptic regularity theory. For u ∈ H
2α, we have
2iA · ∇u + A
2u = 2iP
∇uA + 2iP
A∇u + 2iΠ(∇u, A) + P
uA
2+ P
A2u + Π(u, A
2)
= (α − 1) + (3α − 2) + (2α − 2) + (4α − 2)
= P
uA
2+ P
∇u2iA + (3α − 2)
where (β ) denotes a term of formal regularity H
βfor any β ∈ R. Following the paracontrolled calculus approach, we want to consider a paracontrolled function of the form
u = P e
uX
1+ e P
∇uX
2+ u
]with u
]a smoother remainder such that Hu ∈ L
2. Thus we take
X
1:= ∆
−1(A
2) and X
2:= ∆
−1(2iA) and define the domain of H as follows.
Definition. We define the set D
Aof functions paracontrolled by A as D
A:= {u ∈ L
2; u − P e
uX
1− P e
∇uX
2∈ H
2}.
The domain is defined as
D
A= Φ
−1(H
2) with
Φ(u) := u − e P
uX
1− P e
∇uX
2.
However the domain could be anything from trivial to dense in L
2a priori. For s ∈ (0, 1) , we introduce the map Φ
sas
Φ
s:
D
A→ H
2u 7→ u − P e
suX
1− P e
s∇uX
2with e P
sthe paraproduct truncated at scale s ; see Appendix A for the definition. In particular, the map
Φ
s: H
β→ H
βis a perturbation of the identity for any β ∈ [0, 2α) invertible for s small enough, we denote its inverse Γ . Since P e
v− P e
svX is a smooth function, the domain is also given by D
A= (Φ
s)
−1(H
2) = Γ(H
2).
The reader should keep in mind that Γ implicitely depends on s , we do not keep it in the notation to lighten this work. This parametrisation of the domain will be crucial to prove that the domain is dense in L
2and to study H . In particular, sharp bounds on the eigenvalues of H are needed to get a Weyl-type law for H . To do so, we need to keep a careful track of the different constants.
The reader interested only in the construction of the operator and its self-adjointess can skip these computations. Obtaining sharp bounds requires explicit constants with respect to the parameter s and the regularity exponent in the paracontrolled calculus. We recall the needed bounds in Appendix A, see [18] for details and proofs. For β ∈ [0, 2α) , let
s
β(A) :=
β
∗mkAk
Xα 2α−β4where β
∗= 1 − β if β ∈ [0, 1) and β
∗= 2α − β if β ∈ [1, 2α) and m > 0 is a universal constant, see Appendix A. The following Proposition gives regularity estimates for Φ
sand Γ .
Proposition 1.1. Let β ∈ [0, 2α) and s ∈ (0, 1). We have
kΦ
s(u) − uk
Hβ≤ m s
2α−β4β
∗kAk
Xαkuk
Hβ.
In particular, s < s
β(A) implies that the map Φ
sis invertible and its inverse Γ verifies the bound
kΓu
]k
Hβ≤ 1 1 − m
s2α−β 4
β∗
kAk
Xαku
]k
Hβ.
Proof : If β < 1 , the bounds on Φ
sfollow directly from
ke P
suX
1+ e P
s∇uX
2k
Hβ≤ m s
2α−β41 − β kuk
L2kX
1k
C2α+ m s
α+1−β41 − β k∇uk
Hβ−1kX
2k
Hα+1≤ m s
2α−β41 − β kuk
HβkX
1k
C2α+ kX
2k
Cα+1≤ m s
2α−β41 − β kuk
HβkAk
Xα. For β ∈ [1, 2α) , we have
ke P
suX
1+ P e
s∇uX
2k
Hβ≤ m s
2α−β42α − β kuk
L2kX
1k
Cα+ m s
α+1−β2α + 1 − β k∇uk
L2kX
2k
Hα+1≤ m s
2α−β42α − β kuk
H1kX
1k
C2α+ kX
2k
Cα+1≤ m s
2α−β42α − β kuk
H1kAk
Xα. The result for Γ follows since 2α − β > 0 .
We also consider the associated maps Φ
sεand Γ
εfor a regularisation A
εof the enhanced poten- tial. It is defined as
Φ
sε(u) := u − e P
uX
1(ε)− e P
∇uX
2(ε)and
Γ
εu
]= P e
Γεu]X
1(ε)+ P e
∇Γεu]X
2(ε)+ u
]where
X
1(ε):= ∆
−1(A
2ε) and ∆X
2(ε):= 2iA
ε.
These satisfy the same continuity properties as Φ
sand Γ with bounds uniform with respect to ε .
Moreover, we have the following approximation Lemma holds.
Lemma 1.2. Let β ∈ [0, 2α) and s ∈ (0, 1). If s ≤ s
β(A), we have
k Id − ΓΓ
−1εk
L2→Hβ.
A,skA − A
εk
Xα. In particular, this implies the convergence of Γ
εto Γ with the bound kΓ − Γ
εk
Hβ→Hβ.
A,skA − A
εk
Xα.
Proof : Given any u ∈ H
β, we have u = ΓΓ
−1(u) = Γ(u − e P
suX
1− e P
s∇uX
2) . We get ku − ΓΓ
−1ε(u)k
Hβ=
Γ u − P e
suX
1− P e
s∇uX
2− Γ u − e P
suX
1(ε)− e P
s∇uX
2(ε)Hβ
= Γ
e P
suX
1(ε)− X
1+ e P
s∇uX
2(ε)− X
2Hβ
.
A,sP e
suX
1(ε)− X
1+ P e
s∇uX
2(ε)− X
2Hβ
.
A,skA
ε− Ak
X2αkuk
L2since s < s
β(A) implies the continuity of Γ : H
β→ H
βand X
i(ε)− X
idepends linearly on A
ε− A for i ∈ {1, 2} . The result on Γ − Γ
εfollows from the bound on Γ
εuniform with respect to ε .
This allows to prove density of the domain.
Corollary 1.3. The domain D
Ais dense in H
βfor every β ∈ [0, 2α).
Proof : Given f ∈ H
2, Γ(g
ε) ∈ D
Awhere g
ε= Γ
−1εf ∈ H
2thus we can conclude with the previous Lemma that
ε→0
lim kf − Γ(g
ε)k
Hβ= 0.
The density of H
2in H
βcompletes the proof.
1.2 – First properties of H
Since H is formally given by
H = −∆ + 2iA · ∇ + A · A,
we are able to define (H, D
A) as an unbounded operator in L
2associated to the enhanced potential A .
Definition 1.4. We define H : D
A⊂ L
2→ L
2as
Hu := −∆u
]+ R(u) where u
]= Φ(u) and
R(u) := P
2iA∇u + Π(∇u, 2iA) + P
A2u + Π(u, A
2).
The definition of H is independant of the parameter s ∈ (0, 1) . It is a very usefull tool to get differents bounds on the operator with the different representations
Hu = −∆u
]s+ R(u) + Ψ
s(u) where u
]s:= Φ
s(u) and
Ψ
s(u) := −∆ e P
u− P e
suX
1− ∆ P e
∇u− P e
s∇uX
2∈ C
∞( T
2).
For example, we can compare the graph norm of H
kuk
2H:= kuk
2L2+ kHuk
2L2and the natural norms of the domain
kuk
2DA:= kuk
2L2+ kΦ
s(u)k
2H2with the following Proposition provided s is small. Let β :=
12(
43+ 2α) and δ > 0 . For s ∈ (0, 1) such that s < s
β(A) , we introduce the constant
m
2δ(A, s) := ks
α−22kAk
Xα+ kδ
−2−ββ
kAk
Xα1 − m
s2α−β 4
β∗
kAk
Xα
2 2−β
1 + s
α2kAk
Xαwith k > 0 a large enough constant depending. In particular, m
2δ(A, s) diverges as s goes to 0 or s
β(A) or as δ goes to 0 .
Proposition 1.5. Let u ∈ D
Aand s ∈ (0, 1) such that s < s
β(A). Then for any δ > 0, we have (1 − δ)ku
]sk
H2≤ kHuk
L2+ m
2δ(A, s)kuk
L2and
kHuk
L2≤ (1 + δ)ku
]sk
H2+ m
2δ(A, s)kuk
L2with u
]s= Φ
s(u).
Proof : Recall that for any s ∈ (0, 1) , the operator is given by Hu = −∆u
]s+ R(u) + Ψ
s(u) thus we need to bound R and Ψ
s. For u ∈ D
A, we have
kP
2iA∇u + Π(∇u, 2iA)k
L2. k2iAk
Cα−1kuk
HβkP
A2u + Π(u, A
2)k
L2. kA
2k
C2α−2kuk
Hβhence
kR(u)k
L2. kAk
Xαkuk
Hβ. We also have
kΨ
s(u)k
L2. k(e P
u− e P
su)X
1+ (e P
∇u− e P
s∇u)X
2k
H2. s
α−22kAk
Xαkuk
L2. For s < s
β(A) , Proposition 1.1 gives
kuk
Hβ≤ 1 1 − m
s2α−β 4
β∗
kAk
Xαku
]sk
Hβthus we get
kHu + ∆u
]sk
L2. kAk
Xα1 − m
s2α−β 4
β∗
kAk
Xαku
]sk
Hβ+ s
α−22kAk
Xαkuk
L2.
Since 0 < β < 2 , we have for any t > 0
ku
]sk
Hβ.
Z
t 0(−t
0∆)e
t0∆u
]sdt
0t
0Hβ
+ e
t∆u
]s Hβ. t
2−β2ku
]sk
H2+ t
−β2ku
]sk
L2. t
2−β2ku
]sk
H2+ t
−β21 + s
2α4kAk
Xαkuk
L2.
For any δ > 0 , take
t =
δ 1 − m
s2α−β 4
β∗
kAk
XαkkAk
Xα
2 2−β
with k the constant from the previous inequality. This yields
kHu + ∆u
]sk
L2. m
2δ(A, s)kuk
L2+ δku
]sk
H2. and completes the proof.
Remark : In comparison to the work [18] on the Anderson Hamiltonian
u 7→ −∆u + uξ
where the space white noise can be interpreted as an electric potential, one needs s small for these bounds to hold here. In fact, one could perform the same kind of expansion with
∆X
2= P
∇X12iA
at the price of a nonlinear dependance of X
2with respect to A in order to bypass the smallness condition on s. This would change the different bounds one get for Φ
sand Γ but still yield a self- adjoint operator that is the limit of the regularised H
ε. Theorem XIII .26 from [21] guarantees that the different choice of construction coincide provided that H is self-adjoint.
In particular, this implies that (H, D
A) is a closed operator in L
2. Proposition 1.6. The operator H is closed on its domain D
A.
Proof : Let (u
n)
n≥0⊂ D
Abe a sequence such that
u
n→ u in L
2and Hu
n→ v in L
2. Proposition 1.5 gives that Φ
s(u
n)
n≥0
is a Cauchy sequence in H
2hence converges to u
]s∈ H
2for s < s
β(A) . Since Φ
s: L
2→ L
2is continuous, we have Φ
s(u) = u
]shence u ∈ D
A. Finally, we have
kHu − vk
L2≤ kHu − Hu
nk
L2+ kHu
n− vk
L2.
Aku
]s− Φ
s(u
n)k
H2+ ku − u
nk
L2+ kHu
n− vk
L2hence Hu = v and H is closed on D
A.
We conclud this Section by computing the Hölder regularity of the functions in the domain.
Corollary. We have
D
A⊂ C
1−κfor any κ > 0.
Proof : The Besov embedding in two dimensions implies H
2, → B
1∞,2, → L
∞and Φ
s: L
∞→ L
∞is also invertible for s small enough hence D
A= (Φ
s)
−1(H
2) ⊂ L
∞. First for u ∈ D
A, we have
kuk
Cα. kuk
L∞kX
1k
Cα+ k∇uk
C−1kX
2k
Cα+1+ ku
]k
Cα.
Akuk
L∞+ ku
]k
H2.
Finally, this gives
kuk
C1−κ. kuk
L∞kX
1k
C1−κ+ k∇uk
Cα−1kX
2k
C2−α+κ+ ku
]k
C1−κ.
Akuk
L∞+ kuk
Cα+ ku
]k
H2and the proof is complete.
1.3 – Form comparison between H and ∆
We have proven in Theorem 1.5 that Hu can be seen as a small perturbation of −∆u
]in norm.
Here, we prove a similar statement in the quadratic form setting. Let η :=
α4and δ > 0 . For s ∈ (0, 1) such that s < s
1−η(A) , define
m
1δ(A, s) := (1 + s
α−22)kAk
Xα+ δ
−1−ηηkAk
Xα(1 − η
−1s
2α+η−14kAk
Xα)
2!
1η(1 + s
α2kAk
Xα)
with k > 0 a large enough constant. In particular, m
1δ(A, s) diverges as s goes to 0 or s
1−η(A) or as δ goes to 0 .
Proposition 1.7. Let u ∈ D
Aand s ∈ (0, 1) such that s < s
1−α4
(A). For any δ > 0, we have (1 − δ)h∇u
]s, ∇u
]si ≤ hu, Hui + m
1δ(A, s)kuk
2L2and
(1 − δ)h∇u
]s, ∇u
]si ≤ hu, H
εui + m
1δ(A, s)kuk
2L2where u
]s= Φ
s(u).
Proof : For u ∈ D
A, recall that
Hu = −∆u
]s+ R(u) + Ψ
s(u) where u
]s= Φ
s(u) ∈ H
2. We have
u, −∆u
]s=
e P
suX
1, −∆u
]s+
P e
s∇uX
2, −∆u
]s+
u
]s, −∆u
]s= −
P
su∆X
1, u
]s−
P
s∇u∆X
2, u
]s+
∇u
]s, ∇u
]sthus
hu, Hui = −
P
su∆X
1, u
]s−
P
s∇u∆X
2, u
]s+
∇u
]s, ∇u
]s+
u, R(u) +
u, Ψ
s(u) .
For η ≤
α2, we have
P
su∆X
1, u
]s. kP
su∆X
1k
H2α−2ku
]sk
H1−η. kAk
Xαkuk
L2ku
]sk
H1−η,
P
s∇u∆X
2, u
]s. kP
s∇u∆X
2k
Hα−1−ηku
]sk
H1−η. kAk
Xαkuk
H1−ηku
]sk
H1−η,
u, P
2iA∇u
. kuk
H1−ηkP
2iA∇uk
Hα−1−η. kAk
Xαkuk
2H1−η,
u, P
A2u + Π(u, A
2)
. kuk
L2kP
A2u + Π(u, A
2)k
L2. kAk
Xαkuk
L2kuk
H1−η,
u, Ψ
s(u)
. kuk
L2k(P
u− P
su)LX
1+ (P
∇u− P
s∇u)LX
2k
L2. s
α−22kAk
Xαkuk
2L2. The only term that is not a priori controlled is
u, Π(2iA, ∇u) since the resonant term is singular if we only suppose that u ∈ H
1; this is where the almost duality property comes into play. We have
u, Π(2iA, ∇u)
=
P
u2iA, ∇u
+ F(u, 2iA, ∇u)
with the corrector F(u, 2iA, ∇u) controlled if u ∈ H
1−ηwith η <
α2. The paraproduct is not singular however one can not use better regularity than L
2for u thus we use an integration by part to get
P
u2iA, ∇u
= −
div (P
u2iA), u
= −
P
udiv (2iA), u +
B(u, 2iA), u with
B a, (b
1, b
2)
:= div P
a(b
1, b
2)
− P
adiv (b
1, b
2), see Appendix A for continuity estimates on B . We have
u, Π(2iA, ∇u) .
B(u, 2iA), u
+ |F(u, 2iA, ∇u)| . kAk
Xαkuk
2H1−ηsince div (A) = 0 . Since s < s
1−η(A) , we get
hu, Hui − h∇u
]s, ∇u
]si
. (1 + s
α−22)kAk
Xαkuk
2L2+ kAk
Xα1 − η
−1s
2α+η−14kAk
Xα2ku
]sk
2H1−η.
To complete the proof, one only has to interpolate the H
1−ηnorm of u
]sbetween its H
1norm and its L
2norm which is controlled by the L
2norm of u , as in the proof of Proposition 1.5. Since 0 < 1 − η < 1 , we have for any t > 0
ku
]sk
H1−η.
Z
t 0(−t
0∆)e
t0∆u
]sdt
0t
0H1−η
+ e
t∆u
]s H1−η. t
η2ku
]sk
H1+ t
−1−η2ku
]sk
L2. t
η2ku
]sk
H1+ t
−1−η21 + s
2α4kAk
Xαkuk
L2.
For any δ > 0 , take
t = δ 1 − η
−1s
2α+η−14kAk
Xα2kkAk
Xα!
η2with k the constant from the previous inequality. This yields hu, Hui − h∇u
]s, ∇u
]si
. m
2δ(A, s)kuk
L2+ δku
]sk
H1. and completes the proof.
2 – Self-adjointness and spectrum
In this section, we prove that H is self-adjoint with pure point spectrum. It is symmetric since H Γ is the limit in norm of the regularised H
εΓ
εas proved in Section 2.1. Hence it is enough to prove that
H + k : D
A→ L
2is surjective for some k ∈ R, this is the content of Section 2.2. In Section 2.3, we prove that H
εconverges to H in the stronger resolvent sense. Finally, we give in Section 2.4 bounds for the eigenvalues of H using the different representation
H = −∆Φ
s+ R + Ψ
sparametrised by s ∈ (0, 1) from the eigenvalues of −∆ . In particular, it implies a Weyl-type law for H .
2.1 – The operator is symmetric
To prove that H is symmetric, we use the regularised operator H
ε. Recall that (H
ε, H
2) is self- adjoint and that Φ
sε: H
2→ H
2is continuous. In some sense, the operator H should be the limit of
H
ε:= −∆ + 2iA
ε· ∇ + A
2εas ε goes to 0 with A
ε:= (A
ε, A
2ε) a smooth approximation of A in X
α. Since D(H
ε) = H
2, one can not compare directly the operators. However given any u ∈ L
2, we have
u = Γ ◦ Φ
s(u) = lim
ε→0
Γ
ε◦ Φ
s(u).
Thus for u ∈ D
A, the approximation u
ε:= Γ
ε◦ Φ
s(u) belongs to H
2and one can consider the difference
kHu − H
εu
εk
L2= k(H Γ − H
εΓ
ε)u
]k
L2with u
]:= Φ
s(u) . The following Proposition assures the convergence of H
εΓ
εto H Γ provided s < s
β(A) where β =
12(
43+ 2α) .
Proposition 2.1. Let u ∈ D
Aand s ∈ (0, 1) such that s < s
β(A). Then kHu − H
εu
εk
L2.
A,sku
]sk
H2kA − A
εk
Xαwith u
]s= Φ
s(u) and u
ε:= Γ
εu
]s. In particular, this implies that H
εΓ
εconverges to HΓ in norm
as ε goes to 0 as operators from H
2to L
2.
Proof : We have
H
εu
ε= −∆u
]s+ R
ε(u
ε) + Ψ
sε(u
ε)
where R
εand Ψ
sεare defined as R and Ψ
swith A
εinstead of A . Since
43< β < 2α , we have kR(u)−R
ε(u
ε)k
L2≤ kR(u − u
ε)k
L2+ k(R − R
ε)(u
ε)k
L2.
s,Aku − u
εk
Hβ+ kA − A
εk
Xαku
εk
Hβand
kΨ
s(u) − Ψ
sε(u
ε)k
L2.
s,Aku − u
εk
L2+ kA − A
εk
Xαkuk
L2and the proof is complete since s < s
β(A) implies
kuk
Hβ.
s,Aku
]sk
Hβ.
The symmetry of H immediately follows.
Corollary. The operator H is symmetric.
Proof : Let u, v ∈ D
Aand consider u
]:= Φ
s(u) and v
]:= Φ
s(v) for s < s
β(A) . Since H
εis a symmetric operator, we have
hHu, vi = lim
ε→0
hH
εΓ
εu
]s, Γ
εv
s]i = lim
ε→0
hΓ
εu
]s, H
εΓ
εv
]si = hu, Hvi using that H
εΓ
εconverges to H Γ and Γ
εto Γ in norm convergence.
2.2 – The operator is self-adjoint
In this Section, we prove that (H, D
A) is self-adjoint. Being closed and symmetric, it is enough to prove that
(H + k)u = v
admits a solution for some k ∈ R, see Theorem X .1 in [22]. This is done using the Babuška-Lax- Milgram Theorem, see [2], and Theorem 1.7 which implies that H is almost surely bounded below for any δ ∈ (0, 1) and s small enough.
Proposition 2.2. Let δ ∈ (0, 1) and s ∈ (0, 1) such that s < s
1−α4
(A). For k > m
1δ(A, s), the operators H + k and H
ε+ k are invertibles as unbounded operator in L
2. Moreover the operators
H + k
−1: L
2→ D
AH
ε+ k
−1: L
2→ H
2are bounded.
Proof : Since s < s
1−α4
(A) and k > m
1δ(A, s) , Proposition 1.7 gives k − m
1δ(A, s)
kuk
2L2<
(H + k)u, u for u ∈ D
A. Considering the norm
kuk
2DA= kuk
2L2+ ku
]sk
2H2on D
A, this yields a weakly coercive operator using Proposition 1.5 in the sense that kuk
DA.
Ak(H + k)uk
L2= sup
kvkL2=1
(H + k)u, v
for any u ∈ D
A. Moreover, the bilinear map
B : D
A× L
2→ R (u, v) 7→
(H + k)u, v
is continuous since Proposition 1.5 implies
|B(u, v)| ≤ k(H + k)uk
L2kvk
L2.
Akuk
DAkvk
L2for u ∈ D
Aand v ∈ L
2. The last condition we need is that for any v ∈ L
2\{0} , we have sup
kukDA=1
|B(u, v)| > 0.
Let assume that there exists v ∈ L
2such that B(u, v) = 0 for all u ∈ D
A. Then
∀u ∈ D
A, hu, vi
DA,D∗A
= 0.
hence v = 0 as an element of D
∗A. By density of D
Ain L
2, this implies v = 0 in L
2hence the property we want. By the Theorem of Babuška-Lax-Milgram, for any f ∈ L
2there exists a unique u ∈ D
Asuch that
∀v ∈ L
2, B(u, v) = hf, vi.
Moreover, we have kuk
DA.
Akf k
L2hence the result for (H + k)
−1. The same argument works for H
ε+ k since Peroposition 1.7 also holds for H
εwith bounds uniform in ε .
As explained before, this immediatly implies that H is self-adjoint. Moreover, the resolvent is a compact operator from L
2to itself since D
A⊂ H
βfor any β ∈ [0, 2α) hence it has pure point spectrum.
Corollary 2.3. The operator H is self-adjoint with discret spectrum λ
n(A)
n≥1
which is a nonde- creasing diverging sequence without accumulation points. Moreover, we have
L
2= M
n≥1
Ker H − λ
n(A)
with each kernel being of finite dimension. We finally have the min-max principle λ
n(A) = inf
D
sup
u∈D;kukL2=1
hHu, ui
where D is any n-dimensional subspace of D
Athat can also be written as λ
n(A) = sup
v1,...,vn−1∈L2
inf
u∈Vect(v1,...,vn−1)⊥
kukL2 =1
hHu, ui.
2.3 – H is the resolvent-limit of H
εSince the intersection of the domains of H and H
εis trivial, the natural convergence of H
εto H is in the resolvent sense, this is the following Proposition. In particular, this result explains why our operator H is natural since the regularised operator satisfies
i∂
1+ A
(ε)1 2+ i∂
2+ A
(ε)2 2= H + c
ε+ o
ε→0(1) in the norm resolvent sense.
Proposition 2.4. Let δ > 0 and s ∈ (0, 1) such that s < s
β(A). Then for any constant k > m
1δ(A, s), we have
k(H + k)
−1− (H
ε+ k)
−1k
L2→L2.
A,skA − A
εk
Xα.
Proof : Let v ∈ L
2. Since H + k : D
A→ L
2is invertible, there exists u ∈ D
Asuch that v = (H + k)u
thus
k(H + k)
−1v − (H
ε+ k)
−1vk
L2= ku − (H
ε+ k)
−1(H + k)uk
L2.
We introduce u
ε:= Γ
εΦ
s(u) which converges to u in L
2and we have
ku − (H
ε+ k)
−1(H + k)uk
L2≤ ku − u
εk
L2+ ku
ε− (H
ε+ k)
−1(H + k)uk
L2. Since Lemma 1.2 gives
ku − u
εk
L2.
A,skA − A
εk
Xα, we only have to bound the second term. We have
ku
ε− (H
ε+ k)
−1(H + k)uk
L2= k(H
ε+ k)
−1(H
ε+ k)u
ε− (H + k)u k
L2. k(H
ε+ k)u
ε− (H + k)uk
L2. kH
εu
ε− Huk
L2+ kku
ε− uk
L2using Proposition 2.1. In the end, we have
k(H + k)
−1v − (H
ε+ k)
−1vk
L2. ku
]sk
H2kA − A
εk
Xαhence the result since (H + k)
−1: L
2→ D
Ais continuous.
2.4 – Comparison between the spectrum of H and ∆
The following Proposition provides sharp bounds for the eigenvalues of H from the eigenvalues of
∆ , we denote by (λ
n)
n≥1the non-decreasing positive sequence of the eigenvalues of −∆ since it corresponds to the case A = 0 . For δ ∈ (0, 1) and s ∈ (0, 1) , introduce the constant
m
+δ(A, s) := (1 + δ)(1 + ms
α2kAk
Xα).
For s < s
0(A) , we also introduce
m
−δ(A, s) := 1 − δ 1 − ms
α2kAk
Xα.
Recall β =
12(
43+ 2α) . Write a, b ≤ c to mean that we have both a ≤ c and b ≤ c . Proposition 2.5. Let δ ∈ (0, 1) and s ∈ (0, 1) such that s < s
β(A) ∧ s
1−α4
(A). Given any n ∈ Z
+, we have
λ
n(A), λ
n(A
ε) ≤ m
+δ(A, s)λ
n+ 2 + 2ms
α2kAk
Xα+ m
2δ(A, s).
If moreover s < s
0(A), we have
λ
n(A), λ
n(A
ε) ≥ m
−δ(A, s)λ
n− m
1δ(A, s).
Proof : Let u
]1, . . . , u
]n∈ H
2be an orthonormal family of eigenfunctions of −∆ associated to λ
1, . . . , λ
nand consider
u
i:= Γu
]i∈ D
Afor 1 ≤ i ≤ n . Since Γ is invertible, the family (u
1, . . . , u
n) is free thus the min-max representation of λ
n(A) yields
λ
n(A) ≤ sup
u∈Vect(u1,...,un)
kukL2 =1
hHu, ui.
Given any normalised u ∈ Vect (u
1, . . . , u
n) , we have
hHu, ui ≤ kHuk
L2≤ (1 + δ)ku
]sk
H2+ m
2δ(A, s) for u
]s= Φ
s(u) using Proposition 1.5 and s < s
β(A) . Moreover
ku
]sk
H2≤ (1 + λ
n)ku
]sk
L2≤ (1 + λ
n)
1 + ms
α2kAk
Xαhence the upper bound
λ
n(A) ≤ m
+δ(A, s)λ
n+ 2 + 2ms
α2kAk
Xα+ m
2δ(A, s).
For the lower bound, we use the min-max representation of λ
n(A) under the form λ
n(A) = sup
v1,...,vn−1∈L2
inf
u∈Vect(v1,...,vn−1)⊥
kukL2 =1
hHu, ui.
Introducing
F := Vect (u
m; m ≥ n),
we have that F
⊥is a subspace of L
2of finite dimension n − 1 thus there exists a orthogonal family (v
1, . . . , v
n−1) such that F
⊥= Vect (v
1, . . . , v
n−1) . Since F is a closed subspace of L
2as an intersection of hyperplans, we have F = Vect (v
1, . . . , v
n−1)
⊥hence
λ
n(A) ≥ inf
u∈F
kukL2 =1
hHu, ui.
Let u ∈ F with kuk
L2= 1 . Using Proposition 1.7, we have
hHu, ui ≥ (1 − δ)h∇u
]s, ∇u
]si − m
1δ(A, s)
≥ (1 − δ)hu
]s, −∆u
]si − m
1δ(A, s)
≥ (1 − δ)λ
nku
]sk
2L2− m
1δ(A, s).
Finally using Proposition 1.1 for s < s
1−α4
(A) , we get hHu, ui ≥ 1 − δ
1 − ms
α2kAk
Xαλ
n− m
1δ(A, s)
and the proof is complete.
In particular, taking
s = δ
mkAk
Xα α2gives the simpler bounds
λ
n− m
1δ(A) ≤ λ
n(A) ≤ (1 + δ)
2λ
n+ m
2δ(A)
for any δ small enough. This is sharp enough to get an almost sure Weyl-type law from the Weyl law for the Laplacian ∆ , we denote by N (λ) the number of eigenvalues of −∆ lower than λ ∈ R.
Corollary 2.6. We have
λ→∞