Micro-local analysis of contact Anosov flows and band

structure of the Ruelle spectrum

Frédéric Faure
Institut Fourier, UMR 5582, Laboratoire de Mathématiques
Université Grenoble Alpes, CS 40700, 38058 Grenoble cedex 9, France
[Link]@[Link]

Masato Tsujii
Department of Mathematics, Kyushu University,
Moto-oka 744, Nishi-ku, Fukuoka, 819-0395, Japan
tsujii@[Link]
5 September 14, 2023

Abstract
We develop a geometrical micro-local analysis of contact Anosov flow, such as geodesic
flow on negatively curved manifold. This micro-local analysis is based on wave-packet
transform discussed in [18]. The main result is that the transfer operator is well approxi-
10 mated (in the high frequency limit) by the quantization of the Hamiltonian flow naturally
defined from the contact Anosov flow and extended to some vector bundle over the sym-
plectization set. This gives a few important consequences: the discrete eigenvalues of the
generator of transfer operators, called Ruelle spectrum, are structured into vertical bands.
If the right-most band is isolated from the others, most of the Ruelle spectrum in it con-
15 centrate along a line parallel to the imaginary axis and, further, the density satisfies a
Weyl law as the imaginary part tend to infinity. Some of these results were announced in
[15].

Remark 0.1. On this pdf file, you can click on the colored words, they contain an hyper-link
20 to wikipedia or other multimedia contents. Appendix A contains some convention of notations
used in this paper.

1
2010 Mathematics Subject Classification:
37D20 Uniformly hyperbolic systems (expanding, Anosov, Axiom A, etc.)
37D35 Thermodynamic formalism, variational principles, equilibrium states
37C30 Zeta functions, (Ruelle-Frobenius) transfer operators, and other functional analytic techniques in
dynamical systems
81Q20 Semiclassical techniques, including WKB and Maslov methods
81Q50 Quantum chaos
Keywords: Transfer operator; Ruelle resonances; decay of correlations; Semi-classical analysis.

1
 Contents
1 Introduction 4
1.1 Main result: emergence of an effective quantum dynamics . . . . . . . . . . . . . 5
25 1.2 Sketch of proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Consequences and other results . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.4 Related papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5 Organization of the paper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2 Micro-local analysis of a non vanishing vector field X 19

30 2.1 Vector field X and pull back operator etX . . . . . . . . . . . . . . . . . . . . . 19
2.2 More general pull back operator etXF with generator XF . . . . . . . . . . . . . 19
2.3 Lifted flow ϕ̃t : T ∗ M → T ∗ M and its generator X̃ . . . . . . . . . . . . . . . . . 20
2.4 Metric g on T ∗ M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5 Wave-packet transform T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
35 2.6 Description of the operator etX with the bundle T T ∗ M . . . . . . . . . . . . . . 24

3 Micro-local analysis of a contact vector field X on (M, A) near Σ = RA\ {0} 30

3.1 The symplectization Σ = RA\ {0} . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Linearized dynamics on the set Σ . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 The symplectic normal bundle N → Σ . . . . . . . . . . . . . . . . . . . . . . . 32
40 3.4 Equivalent family of operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.5 Description of the operator etX near the set Σ with the bundle N . . . . . . . . 37
3.6 Description of the operator etX near the set Σ with the bundle TΣ T ∗ M . . . . . 40
3.7 Emergence of quantum mechanics near the set Σ for a general contact vector field 40

4 Micro-local analysis of a contact Anosov vector field X near Σ = RA\ {0} 41

45 4.1 Definitions and notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 Anisotropic Sobolev space, decay of norm outside the trapped set and discrete
Ruelle spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Description of the operator etX near the set Σ with the bundle Ns . . . . . . . . 47
4.4 The bundle map etXF and sub-bundle Fk (Ns ) . . . . . . . . . . . . . . . . . . . 49
50 4.5 Symbols and F.I.O. on Σ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5 Proof of Theorem 1.6 (band spectrum) 59

5.1 Contribution of Ω0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2 Contribution of Ω1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.3 Contribution of Ω2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
55 5.4 Sum of contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6 Proof of Theorem 1.4 (emergence of quantum dynamics) 63

7 Proof of Theorem 1.10 (Weyl law) 64

7.1 Approximate projector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.2 Truncated resolvent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
60 7.3 Counting resonances by argument principle . . . . . . . . . . . . . . . . . . . . . 69
7.4 Relation with the symplectic volume . . . . . . . . . . . . . . . . . . . . . . . . 70
7.5 Weyl law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

8 Proof of Theorem 1.12 (accumulation of eigenvalues on narrower bands) 73

9 Proof of Theorem 1.17 (horocycle operators) 74

65 A General notations used in this paper 75

2
 B More information about flows 76
B.1 Transfer operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
B.2 More general pull back operators etXF . . . . . . . . . . . . . . . . . . . . . . . . 77

C Bargmann transform and Metaplectic operators 78

70 C.1 Weyl Heisenberg group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
C.2 Bargmann Transform on a Euclidean vector space (E, g) . . . . . . . . . . . . . 79
C.3 Linear map ϕ : (E2 , g2 ) → (E1 , g1 ) . . . . . . . . . . . . . . . . . . . . . . . . . . 82
C.4 Compatible triple g, Ω, J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
C.5 Bergman projector on a symplectic vector space (F, Ω, g) . . . . . . . . . . . . . 85
75 C.6 The metaplectic decomposition of F ⊕ F ∗ . . . . . . . . . . . . . . . . . . . . . 85
C.7 Metaplectic decomposition of a linear symplectic map . . . . . . . . . . . . . . . 87
C.8 Some useful decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
C.9 Taylor operators Tk on S (E) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
C.10 Analysis on T (E ⊕ E ∗ ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

80 D Linear expanding maps 94

D.1 Anisotropic Sobolev space HW (E) . . . . . . . . . . . . . . . . . . . . . . . . . 94
D.2 Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
D.3 Discrete Ruelle spectrum in a simple toy model . . . . . . . . . . . . . . . . . . 97

3
 1 Introduction
85 Contact Anosov flows: In this paper we consider a smooth contact Anosov vector field
X on a closed contact manifold (M, A) with contact one form A. A particular and important
example is with (N , gN ) a closed Riemannian manifold with negative curvature, the geodesic
vector field X is a contact Anosov vector field on the unit cotangent bundle M = T1∗ N , where
the contact one form A = pdq is the Liouville one-form.
90 The vector field X is considered as a derivation X : C ∞ (M ) → C ∞ (M ) and generates the
pullback operators by the flow ϕt : M → M , i.e. etX u = u ◦ ϕt with u ∈ C ∞ (M ) and t ∈ R.
More generally, if F is a vector bundle over M , we will consider a derivation XF acting on
smooth sections C ∞ (M ; F ) of F over2 X, generator of the pull back operator

etXF : C ∞ (M ; F ) → C ∞ (M ; F ) . (1.1)

As an example, we may consider F = Λ• (M ), bundle of differential forms, and XF being the

95 Lie derivative.

Question of long time behavior: Anosov flows are a typical model of chaotic dynamics
with sensitivity to initial conditions. A typical question in dynamical systems theory that is
addressed in this paper is to describe the long time action of the flow generated by X on
smooth sections, i.e. for any given smooth sections u, v ∈ C ∞ (M ; F ), describe the correlation
100 ⟨v|etXF u⟩L2 for t → ∞? Here L2 = L2dm (M ; F ) with the invariant contact volume dm (3.2)
on M and an arbitrary hermitian metric on F . In other words we are looking for a “good
description” of the pull back operator etXF for t → ∞.
In the case of a trivial bundle F = C, it is known from Sinaï [35] that contact Anosov
flows are mixing, i.e. for any smooth functions u, v ∈ C ∞ (M ; C), we have ⟨v|etX u⟩L2 →
t→+∞
105
1
⟨v|1⟩L2 ⟨ Vol(M )
|u⟩L2 . This mixing property dominates the long time behavior and says that the
operator e converges (in the weak sense) to the rank one operator Π0 = 1⟨ Vol(M
tX 1
)
|.⟩L2 .

Problem 1.1. (Aim) In this paper we will be interested in finding a description of the operator
etXF in the limit of large t, up to an error that decays as e−Λt with an arbitrary large rate Λ > 0.

Discrete Ruelle spectrum and anisotropic Sobolev space: Although the space L2 (M ; F )
110 may be natural to consider in order to describe the operator etXF , this is not what we will do. We
will instead consider a family of generalized Sobolev spaces HW (M ; F ), defined from a weight
function W on T ∗ M . By choosing a good weight W , we can get that the essential spectrum of
the generator XF is in {z ∈ C, Re (z) < −Λ} with arbitrary large Λ > 0 [18, thm 2.11], reveal-
ing some intrinsic discrete spectrum called Ruelle resonances in {z ∈ C, Re (z) > −Λ}.
115 See Figure 1.2. This fact has been essentially first obtained by Butterley and Liverani in [5].
The function space HW (M ; F ) is called anisotropic Sobolev space [23][2][13, 14] because it is a
Hilbert space of distributional sections that contains smooth sections C ∞ (M ; F ) and its order
depends on the stable/unstable directions of the dynamics.
This discrete spectrum governs the long time behavior of etXF but only little3 is known
120 about it for a general Anosov flow (or Axiom A flow). In the special case of a contact Anosov
vector field X, much more can be said about the Ruelle spectrum of XF and about the effective
action of the operator etXF . This is the subject of this paper.
2
See Definition (2.4).
3
Ruelle spectrum can be calculated explicitly in only exceptional cases. Appendix D.3 presents the Ruelle
spectrum of a very simple toy model on R. See [15, Prop. 4.1] for the Ruelle spectrum of the geodesic flow on
a compact hyperbolic surface.

4
 1.1 Main result: emergence of an effective quantum dynamics
1.1.1 Known results beyond the exponential mixing description
125 In order to motivate the analysis and few results presented in this paper, let us discuss again
about the question of long time behavior of the dynamics etX in the case of a contact Anosov vec-
tor field X (i.e. trivial bundle F = C for this discussion) and known results so far. Beyond the
result of mixing given above that is etX → Π0 = 1⟨ Vol(M 1
)
|.⟩L2 weakly, there is exponential
t→+∞
mixing obtained by C. Liverani [31], that is ∃ϵ > 0, ∃C > 0, ∀t ≥ 0, etX − Π0 H (M ) ≤ Ce−tϵ
M
130 and even better, an explicit description of the exponential small corrections [41] in terms
of a finite number of Ruelle eigenvalues of X in HW (M ), zj = aj + iωj ∈ C with real
part γ0+ + ϵ < . . . aj+1 ≤ aj ≤ . . . < a1 < a0 = 0 and associated finite rank spectral projectors
Πj , that is
+
∃ϵ > 0, Cϵ > 0, ∀t ≥ 0, etX − Rt H (M ) ≤ Cϵ et(γ0 +ϵ) , (1.2)
M

with a finite rank operator 4

X
Rt = etX Πj (1.3)
zj ,Re(zj )>γ0+ +ϵ

135 and with the explicit threshold

1/t
γ0+ := lim log etXF0 L∞
<0 (1.4)
t→+∞

where XF0 is the Lie derivative of the vector field acting on section of the line bundle F0 (Es ) :=
|detEs |−1/2 → M , where Es ⊂ T M is the vector bundle over M of the stable directions of the
Anosov flow, detEs = Es∧  is the determinant bundle from Es , |detEs |1/2 is the half densities
n

 ∗
bundle and |detEs |−1/2 = |detEs |1/2 its dual.
140 Remark 1.2.

• We have γ0+ < 0 because Es is the stable bundle and therefore under the pullback oper-
ator etX , the induced action of the line bundle |detEs | is expanding hence on F0 (Es ) =
|detEs |−1/2 it is contracting. We have the estimate γ0+ < − d2 λmin with the minimal Lya-
pounov exponent λmin > 0 and d = dimEs = 12 (dimM − 1).

145 • Except for the obvious eigenvalue z0 = 1 of X, the set of other eigenvalues {zj }j that
enter in the sum (1.3) is quite unknown in general and may be empty.

• Considering an arbitrary vector bundle F → M changes the operator Rt and the estimate
γ0+ , see below.

1.1.2 Looking for a better description of the dynamics up to arbitrary small rate
150 e−Λt
In this paper our aim 1.1 is to give a description of the dynamics etX similar to (1.2) but with
an error term Ce−Λt with arbitrary large rate Λ > 0. So we have to overcome the threshold
γ0+ . For this we do some micro-local analysis of the pullback operator etX which means that
we describe precisely its action on functions (or sections) with high frequencies (i.e. roughly
155 speaking considering in any local chart only Fourier components with large Fourier variable
|ξ| ≫ 1, ξ ∈ RdimM ).
With the micro-local approach, it is a general fact that we can describe the operator
etX in terms of the action of the flow induced on the cotangent bundle T ∗ M (in other terms
etX is considered as a Fourier Integral Operator F.I.O.). In the special case of contact Anosov
4
In case of a simple eigenvalue zj this gives etX Πj = ezj t Πj . Notice also that z0 = 0 is simple, hence
etX Π0 = Π0 corresponding to the “mixing main term” so called “equilibrium”.

5
160 flow we show that only a neighborhood of the symplectic subset Σ = R∗ A ⊂ T ∗ M has to be
considered, where A is the contact one-form, because (in some sense explained in this paper)
“outside of Σ, the norm of etX ” decays faster than e−Λt with arbitrary large Λ > 0. As a result,
we obtain an effective description of etX in terms of an Hamiltonian flow on the symplectic
subset Σ but considering also its action on the symplectic normal bundle N → Σ, needed
165 to get the neighborhood of Σ. This micro-local description of the operator etX is in
term of
quantization of the Hamiltonian “classical” action etXF giving an operator OpΣ etXF .
The main drawback of the micro-local approach is its inability to describe the “low fre-
quency subspace of functions” (that corresponds effectively to a finite dimensional subspace
since M is compact) and for this reason, our approximation below in theorem 1.4 will contain
170 an unknown finite rank operator Rσ,t , similar to the operator Rt in (1.2). Optimistically we
can consider that a
finite rank operator is negligible compared to the infinite rank quantum
operator OpΣ etXF that will enter in the description.
Let us now explain more precisely this micro-local approach, the effective classical dynamics
that appears in a first result presented in theorem 1.4 below.

175 1.1.3 Effective classical dynamics

Bundle Fk (Es ) → M and its dynamics: We now go to our first aim 1.1 and consider again
a general vector bundle F → M . For this, for every k ∈ N, we introduce the Hölder continuous
finite rank vector bundle over M :

Fk (Es ) := |detEs |−1/2 ⊗ Polk (Es ) ⊗ F → M (1.5)

where Polk (Es ) is the bundle of homogeneous polynomials of degree k on Es . The bundle
180 Es → M is Hölder continuous but smooth along the flow and Es directions. Let C β (M ; Fk (Es ))
be the space of β−Hölder continuous sections. Then we consider XFk : C β (M ; Fk (Es )) →
C β (M ; Fk (Es )) being the derivation of sections of Fk (Es ) over X (C 1 along the flow direction).
XFk is naturally induced from the initial derivation XF and is the generator of the group
of operators etXFk : C β (M ; Fk (Es )) → C β (M ; Fk (Es )) , t ∈ R. We introduce the related
185 quantities that generalize (1.4)
1/t
γk± := lim log etXFk L∞ (M ;Fk (Es ))
. (1.6)
t→±∞

Remark 1.3. We can give obvious (but rough) estimates for γk± as follows. If 0 < λmin ≤ λmax
denote the minimal and maximal Lyapounov exponents of the Anosov flow ϕt , we have
   
d d
− − +
+ k λmax + cF ≤ γk ≤ γk ≤ − + k λmin + CF , (1.7)
2 2
where dimM = 2d + 1 and cF ≤ CF depend on XF . For example, if XF = X is the vector field
itself with F = C, we have cF = CF = 0 and from (1.7), the pinching condition λλmax
min
< 1 + k1
190 implies γk+1
+
< γk− that we call the “spectral gap assumption” later.

Taylor projectors Tk : The homogeneous polynomials Polk (Es ) are obtained as the image
of the “Taylor projector” (see section C.9 for a definition) 5 Tk : S (Es ) → Polk (Es ) ⊂ S ′ (Es )
that we extend to a bundle map Tk ∈ L (F (Es ) , F ′ (Es )),

Tk : F (Es ) := |detEs |−1/2 ⊗ S (Es ) ⊗ F → F ′ (Es ) := |detEs |−1/2 ⊗ S ′ (Es ) ⊗ F (1.8)

For any given K ∈ N, we denote T[0,K] := K k=0 Tk .

L

5
For every point m ∈ M , S (Es (m)) is the space of Schwartz functions on the vector space Es (m).
For simplicity we denote S (Es ) := C β (M ; S (Es )) the space of continuous sections. Similarly S ′ (Es ) :=
C β (M ; S ′ (Es )) is the space of sections valued in Schwartz
distributions, or more precisely distributional exten-
sions of S (Es ), the dual of Schwartz forms S Es ; Λd (Es ) , so that polynomials are included in S ′ (Es ).

6
195 Bundle Fk (Ns ) → Σ: We consider the symplectic sub-manifold

Σ := R∗ A = {ωA (m) , ω ∈ R∗ , m ∈ M }

being the symplectization of the contact manifold (M, A). This manifold Σ, as a sub-manifold
∗
of T ∗ M , is also the trapped set (or non wandering set) of the induced dynamics ϕ̃t = (dϕt ) on
T ∗ M and therefore invariant. We consider the pull back of Fk (Es ) → M under the projection
π : Σ → M . More precisely we consider the symplectic normal bundle N = (T Σ)⊥ , its stable
200 component Ns ⊂ N and construct the bundle Fk (Ns ) := |detNs |−1/2 ⊗ Polk (Ns ) ⊗ F → Σ.
The operator etXFk defined above extends to

etXFk : C β (Σ; Fk (Ns )) → C β (Σ; Fk (Ns )) (1.9)

∗
over the Hamiltonian flow ϕ̃t = (dϕt ) : Σ → Σ. The family of operators etXFk t∈R in

(1.9), will be called the classical dynamics. For any t ∈ R, we also denote the flow etXF :
(Σ; F (Ns )) with similar definition F (Ns ) := |detNs |−1/2 ⊗ S (Ns ) ⊗ F
C β (Σ; F (Ns )) → C β L
205 and that etXF T̃[0,K] = K k=0 e
tXFk
with Taylor projectors T̃k : S (Ns ) → Polk (Ns ) and T̃[0,K] :=
k=0 T̃k .
LK

1.1.4 General remarks about quantization

Before continuing the presentation, we first make a pause to say few important words about
quantization. These remarks will be important to understand the content of the next theorem.

210 • In the most common situation, on a given manifold M , quantization denoted Op (.) maps
a function a ∈ S (T ∗ M ; C) called “symbol” to a bounded operator Op (a) : L2 (M ) →
L2 (M ) called Pseudo Differential Operator P.D.O. There are many different definitions
for Op (.), and it is important to known that they are all equivalent at leading order
in the limit of high frequencies. At leading order the map Op (.) satisfies some
215 very interesting and useful universal properties as the “boundness theorem”, “composition
theorem” that is homomorphism of algebra, “trace formula”, etc (under some additional
hypothesis called “symbol classes” [38, (1.4) p.3]) and our analysis will rely on these
properties. For these reasons the precise definition of Op (.) is not very crucial for the
moment.

220 • The simplest example of Op (.) called ordinary quantization, is given in local charts x ∈ Rn
on M with dual coordinates ξ ∈ Rn on Tx∗ M , by [38, (1.3) p.2]
Z
(Op (a) u) (x) = a (x, ξ) (Fu) (ξ) ei⟨ξ|x⟩ dξ,

where (Fu) (ξ)is Fourier transform of u ∈ S (Rn ). Weyl quantization is different and has
the advantage that a real valued symbol a gives a self-adjoint operator Op (a), [38, (14.1)
p.67].

225 • In geometric quantization people introduce an operator T : C ∞ (M ) → S (T ∗ M ), called

Bargman transform or FBI transform, wavelet transform, wave-packet transform (de-
pending on the context, and additional structures for its construction) and they define
Toeplitz quantization by
Op (a) := T † aT
where a on the right hand side stands for the multiplication operator by a and T † :
230 S (T ∗ M ) → C ∞ (M ) is its L2 adjoint satisfying Op (1) = T † T = Id. We will mainly use
this quantization in this paper.

7
 • It is also possible to define
  the quantization of a symplectomorphism ϕ̃ : T M → T M
∗ ∗

giving an operator Op ϕ̃ : L2 (M ) → L2 (M ) called Fourier Integral Operator F.I.O.

(e.g. [43]). In geometric quantization the definition is
 
1/2
Op ϕ̃ := T † Υϕ̃ ϕ̃−◦ T (1.10)
1/2
235 where ϕ̃−◦ : S (T ∗ M ) → S (T ∗ M ) is the push-forward operator and Υϕ̃ ∈ C ∞ (T ∗ M ; R+ )
is the “metaplectic correction” (C.17) to get an almost unitary operator.
• It is possible to consider a vector bundle F → M and define the quantization of a symbol
a ∈ S (T ∗ M ; End (F )) giving Op (a) : L2 (M ; F ) → L2 (M ; F ).

1.1.5 Quantization in our setting

240 In this paper in order to study the operator etXF : C ∞ (M ; F ) → C ∞ (M ; F ) in (1.1), we
will need all these previous and different variations around quantization and moreover another
variation: instead of considering the whole cotangent bundle T ∗ M as the symplectic manifold,
we will restrict to a vicinity of the symplectic sub-manifold Σ := R∗ A ⊂ T ∗ M . To have a
good description of a vicinity of Σ, we will use its normal bundle N = (T Σ)⊥ → Σ and the
245 exponential map expN : N → T ∗ M that is a local diffeomorphism from a vicinity of the zero
section of N to a vicinity of Σ in T ∗ M . See Figure 1.1.
Then we will use a wave-packet transform given by some operators TNs : C ∞ (M ; F ) →
C β (Σ; F (Ns )) and a converse TN∆ : C β (Σ; F (Ns )) → C ∞ (M ; F ) defined by
†
TNs = BN ^)◦ T
χµ (exp
s ,Nu Σ N
(4.34)

and
−1 ◦ µ
TN∆s = T † exp
^

N χΣ BNs ,Nu ,
(4.35)

250
^)◦ : S (T ∗ M ; F ) →
where T : C ∞ (M ; F ) → S (T ∗ M ; F ) is a wave-packet transform, (exp N
S (Σ; S (N )) is the pull back of expN with some additional phase in (3.35), χµΣ is some cut-
off function to a vicinity of the zero section of N at distance ω µ/2 in (3.34) and BNs ,Nu :
C β (Σ; S (Ns )) → C β (Σ; S (N )) in (4.30) is a bundle-wise Bargman transform. With this, we
define the quantization of the classical dynamics (1.9) with the additional symbol T̃[0,K] by
 
1/2
OpΣ etXF T̃[0,K] = TN∆s Υt etXF T̃[0,K] TNs : C ∞ (M ; F ) → C ∞ (M ; F ) (1.11)
(4.55)

255 called quantum evolution operators. In the analysis, the semi-classical parameter is the
frequency ω along the flow direction (equivalently the parameter along Σ) and the semi-
classical limit is |ω| → ∞ (usually denoted 2πℏ = 1/ |ω| → 0).

1.1.6 Emergence of an effective quantum dynamics

To express the result, we also need an operator Op (χΣ,σ ) : C ∞ (M ; F ) → C ∞ (M ; F ) defined
260 in (1.20) that restricts functions microlocally to their vicinity of Σ at distance σ > 0. The next
theorem shows how the pull-back operator etXF in (1.1) is well approximated by the quantum
evolution operator OpΣ etXF T[0,K] in (1.11).

Theorem 1.4 (Emergence of quantum dynamics). For any K ∈ N and ϵ > 0, we

can choose an anisotropic Sobolev space HW (M ; F ), such that ∃C > 0, ∀t > 0, ∃σt >
0, ∀σ > σt , there exists a finite rank operator Rσ,t and
  +
etXF − Op (χΣ,σ ) OpΣ etXF T̃[0,K] Op (χΣ,σ ) − Rσ,t ≤ Cet(γK+1 +ϵ) . (1.12)
HW (M ;F )

8
 The proof of Theorem 1.4 is given in Section 6.
+
265 Remark 1.5. From (1.7) we have γk+ → −∞, hence the term et(γK+1 +ϵ) in (1.12) decays very
k→∞
fast for t → +∞ when K is large, so this theorem is a kind of answer to our aim 1.1. One
interpretation of Theorem 1.4 is that for large time t ≫ 1, an effective quantum dynamics
emerges from the contact Anosov dynamics. This quantum dynamics is the quantization
of the Hamiltonian dynamics (1.9) that takes place on Σ.
270 To show (1.12) we will use microlocal analysis directly on T ∗ M following the approach
proposed in [18], using a metric g on T ∗ M compatible with the symplectic form, that measures
the size of wave-packets (sometimes called coherents states) in accordance with the uncertainty
principle.

1.1.7 Few comments about the result (1.12)

275 • As explained above, in the case of a trivial bundle F = C, contact Anosov flows are
[
mixing etX t→ +∞]weak→Π0 = 1⟨ Vol(M 1
)
|.⟩L2 . Although this mixing property dominates
the long time behavior, the limit rank one operator 1⟨ Vol(M
1
)
|.⟩L2 , belongs to the finite rank
operator Rσ,t in (1.12) that we do not describe in this paper. We instead consider terms
that are exponentially small compared to it, but that belong to an infinite dimensional
280 effective space. In our analysis the rank of the operator Rσ,t may increase with t because
we use micro-local analysis (i.e. high frequency analysis) where the cutoff ω we need in
frequency increases with t. The operator Rσ,t represents the remainder in low frequencies
that we are not able to describe with this approach.
• In (1.5), the vector bundle F → M is arbitrary but notice that the special choice of the
285 bundle F = |detEs |1/2 is particularly interesting because it gives
F0 = |detEs |−1/2 ⊗ Pol0 (Es ) ⊗ |detEs |1/2 = C (1.13)
(4.27)

i.e. the trivial bundle over Σ. The paper [17] is devoted to that case6 . Eq.(1.6) gives
γ0± = 0 and γ1+ < 0. For that case and taking K = 0, Eq.(1.12) shows that for t ≫
1, the operator e is well described by the dominant term OpΣ e T0 that is the
tXF tXF0

“quantization” of the dynamics X itself. From Theorem 1.6 below, for ω = Im (z) → ∞,
290 the eigenvalues of XF accumulate on the imaginary axis Re (z) = γ0± = 0 with density
given by the Weyl law, separated by a uniform spectral gap γ1+ < 0. See Figure 1.2(b).
• In the special case of the geodesic flow on a surface N of constant and negative curva-
ture, i.e. N = Γ\SL2 (R) /SO2 (R), giving M = Γ\SL2 (R), with Γ being a co-compact
subgroup of SL2 (R) (that contains −Id),
 there is no remainder operator Rt in (1.12) and
295 the dominant quantum operator OpΣ etXF T̃0 has the same spectrum as the wave oper-
 q 
ators exp ±it ∆ − 4 on S (N ), [15, 9], see Figure 4.1(a) in [15, Fig 4.1(a) p.389]. This
1

operator is indeed considered in physics as the Schrödinger evolution operator, giving the
quantum description of a free particle. For the case of geodesic flow on a non-constant neg-
atively curved Riemannian manifold (N , g), let us observe that the symplectic phase space
300 Σ is (symplectically) isomorphic to a double cover of the cotangent bundle T ∗ N \ {0} and

that the principal symbol of the generator of the leading quantum operator OpΣ etXF T̃0
√
in (1.12) is the frequency ω, equal to the principal symbol of ∆ on T ∗ N under this iso-
morphism, where ∆ = d† d is the Laplace Beltrami operator on S (N ). However we do
not expect that the spectra of both operators coincide in general. We will investigate this
305 question in a future work.
6 1/2
F = |detEs | is not a smooth bundle but only Hölder continuous and technically we consider an smooth
extension to a Grassmanian bundle in [17].

9
 • In semi-classical analysis, the Egorov Theorem [38, p.26] or the WKB approximation [4,
p.11] show that classical Hamiltonian dynamics emerges in the high frequency
limit (and finite time) of quantum dynamics. This is how in physics, geometrical
optics is derived from electromagnetic waves, Newtonian (and Hamilton) mechanics is
310 derived from quantum waves mechanics of Schrödinger, etc. Inspired from these phys-
ical phenomena this is how quantization has been defined in mathematics, for example
standard quantization [38, p.2] that defines a pseudo-differential operator on S (Rn ) from
a Hamiltonian function (symbol) on R2n , or geometric quantization in a more geometric
setting [42]. In this paper we have exhibited the converse and maybe unexpected phe-
315 nomena: how quantum mechanics emerges from the classical mechanics when
this later is chaotic. From the mathematical point of view, an interesting consequence is
that it furnishes a natural quantization of a given classical dynamics among all pos-
sible quantizations. This natural quantization has indeed the preferable properties (that
characterize it) that the semi-classical Van-Vleck formula or semi-classical Trace formula
320 are asymptotically exact, i.e. they have error terms that decay exponentially fast with
t → ∞ at large but fixed ω. This provides a kind of generalization of Selberg theory to
non constant curvature. For more discussions, see [16, Section 1.5,1.6,1.7] and [17]. From
a physical point of view one may wonder if quantum phenomena observed in experimental
data may emergence from an underlying deterministic but chaotic dynamics, see e.g. the
325 last paragraph in [33]. In section 3.7 we give theorem 3.21 that concerns a general contact
flow (i.e. without Anosov assumption) so this may concern a general geodesic flow and
show the emergence of quantum dynamics near the invariant set Σ.

1.2 Sketch of proof

The proof of the emerging quantum dynamics in (1.12) can be summarized by the following
330 mechanisms that will be explained in details in this paper and illustrated in Figure 1.1.

1. In the limit of high frequencies, evolution of functions (or sections) by the pull-back
operator etXF is well described on the cotangent bundle T ∗ M with the induced flow
∗
ϕ̃t := (dϕt ) , t ∈ R. This is because etXF is a Fourier integral operator (i.e. has
micro-local property (2.39)). We introduce a specific metric g on T ∗ M , compatible
335 with the symplectic form and then we define an L2 -isometric “wave-packet transform”
T : C ∞ (M ; F ) → S (T ∗ M : F ), that allows to use micro-local analysis on T ∗ M for
the pull back operator etXF . The unit boxes for the metric g correspond to the effective
size of wave-packets and reflect the uncertainty principle.

2. For an Anosov contact flow, the dynamics induced on the cotangent bundle T ∗ M is a
340 “scattering dynamics” on the trapped set Σ = R∗ A ⊂ T ∗ M and Σ is symplectic and
normally hyperbolic. See Figure 1.1. In terms of dynamics, the subset Σ is the non
wandering set for the flow ϕ̃t on T ∗ M and the orbits on the outside of Σ go to infinity
either as t → +∞ or t → −∞. As a consequence, for large time |t| ≫ 1, the outer part
of the trapped set Σ has a negligible contribution, because information escapes to infinity
345 (i.e. the Sobolev norm measured by the specific weight W decays). This will be given in
Theorem 4.6. So only the dynamics on Σ plays a role for our purpose. But due to the
uncertainty principle in T ∗ M , we still have to consider a neighborhood of Σ in T ∗ M . We
will hence consider a neighborhood of Σ of a given size ⟨ω⟩µ/2 in (3.34) (measured by the
metric g), with some 0 < µ < 1. It is important to remark that for large frequencies ω this
350 neighborhood of size ⟨ω⟩µ/2 projected down on M is scaled by ⟨ω⟩−1/2 (from the definition
of the metric g on T ∗ M in (2.25), see remark 2.7) and get size ≍ ω −(1−µ)/2 → 0 that
ω→+∞
goes to zero as ω goes to infinity. This will allow us to use the linearization of the
dynamics ϕ̃t as a local approximation.

10
 Ns (ρ)

Es
< ω >µ/2 ω
Σ = (Eu ⊕ Es )⊥
X symplectic Trapped set
ρ = ωA(m)
m flow ϕt 0

Eu Nu (ρ)
t
Lifted flow ϕ̃

M T ∗M N (ρ)
Size ≍ ω −(1−µ)/2
≪1

Figure 1.1: The dynamics induced on T ∗ M scatters on the trapped set Σ ⊂ T ∗ M defined in
(3.7). Σ is a line bundle over M , a symplectic sub-manifold of T ∗ M and at every point ρ =
ωA (m) ∈ Σ, where ω called frequency is the coordinate along the line, the symplectic-normal
bundle N (ρ) = (Tρ Σ)⊥ (a symplectic linear subspace of Tρ T ∗ M ) splits into unstable/stable
subspaces, N (ρ) = Nu (ρ) ⊕ Ns (ρ), see (3.11). The main geometrical object considered
in this paper is this fibration Ns → Σ → M . Beware that for a geodesic flow on (N , g), this
fibration sequence continues with M = (T ∗ N )1 → N .

3. In the neighborhood of size ⟨ω⟩µ/2 of the trapped set Σ (i.e. the non wandering set)
355 that matters, there is a micro-local decoupling between the directions tangent to Σ and
those (symplectically) normal to Σ, represented by a normal vector bundle denoted N =
(T Σ)⊥ → Σ. The dynamics on the normal direction N is hyperbolic and responsible
for the emergence of polynomial functions along the stable direction Ns (that projects
to Es ). By considering
  the Taylor operators T̃k as symbols, we introduce approximate
360 projectors OpΣ T̃k on C ∞ (M ; F ) that micro-locally (i.e. seen in S (T ∗ M ; F ) after the
wave-packet transform T ) restricts functions to the symplectic trapped set Σ and that
are polynomial valued along Ns with degree k. This projector plays a similar role as
the Bergman projector (or Szegö projector) in geometric quantization. What remains for
large time, is an effective Hilbert space of “quantum waves” that live on the trapped set
365 Σ, valued in the vector bundle Fk defined in (1.5).

The very simple toy model that is useful to have in mind is given in Section D.3. It explains in
particular the emergence of polynomials Polk (Es ) as in (4.27).

1.3 Consequences and other results

370 There are many consequences of the effective description (1.12) of the dynamics. In this
paper we describe a few of them that are illustrated on Figure 1.2. We will follow general
ideas from classical-quantum correspondence principles for quantization in the case where the
classical symbol is an operator valued function on a symplectic manifold Σ (this situation is
also present in physics for elastic waves, electromagnetic waves, Dirac equation etc), and we
375 will derive the following results for the operator XF .

1.3.1 Discrete spectrum in vertical bands

The generator of the classical dynamics (1.9) is XFk and its spectrum
 − + (some essential spectrum
in fact) in L (Σ; Fk ) is contained in the vertical band Bk := γk , γk × iR with γk defined in
2 ±

11
 Spectrum that controls
Essential Intrinsic emerging behavior
ω = Im(z)
spectrum discrete spectrum ω = Im(z)

bounded resolvent
bounded resolvent
r≫1

bounded resolvent
bounded resolvent

ωε ωε

Re(z) Re(z)

H
γ2+ γ1− γ1+ γ0− γ0+
−2λmax r + C ′ B2 B1 B0 γ1+ γ0+ = γ0− = 0
−λmin r + C γ̌0
(a) (b)

Figure 1.2: The dots represent the intrinsic Ruelle discrete spectrum of the derivation XF in a
Sobolev space HW (M ; F ). From [18, thm 2.11], the essential spectrum is in a (brown) vertical
band that can be moved arbitrarily far on the left by changing the weight W , and reveals
this intrinsic discrete spectrum. The right most eigenvalues in the first band B0 dominate
the emerging behavior of etXF for t ≫ 1. On figure (b) for the special case of the bundle
F = |detEs |1/2 , the first band coincides with the imaginary axis.

(1.6). We always have γk+1 −

< γk− and γk+1
+
< γk+ but this does not guarantee that the bands
380 are separated by gaps i.e. that γk+1 < γk , Bk+1 ∩ Bk = ∅.
+ −

In the semi-classical limit, i.e. for ω = Im (z) → ±∞ on the spectral plane z ∈ C, theorem
1.6 below shows S that the spectrum of XF is discrete and asymptotically contained in these
vertical bands k∈N Bk and the resolvent is uniformly bounded in the gaps between the bands
(if they exist). This is illustrated on Figure 1.2.

Theorem 1.6 (Band structure of the Ruelle spectrum). For any ϵ > 0, C > 0,
there exists Cϵ > 0, ωϵ > 0 such that the Ruelle eigenvalues spect (XF ) are contained in
the following spectral domain that consists of a union of a “low frequency horizontal
band” and “vertical bands”:
!
[  − +

(spect (XF ) ∩ {Re (z) > −C}) ⊂ {|Im (z)| ≤ ωϵ } ∪ Re (z) ∈ γk − ϵ, γk + ϵ ,
k∈N
(1.14)
and the resolvent operator is uniformly bounded in the “gaps”:

(z − XF )−1 + ϵ, γk− − ϵ .
 + 
HW
≤ Cϵ , ∀z ∈ C s.t. |Im (z)| > ωϵ , Re (z) ∈ γk+1
(1.15)
385

The proof is given in Section 5.

Notice that if the gap do not exist, i.e. if the band overlap (γk+1+
≥ γk− for every k ∈ N)
then the content of theorem 1.6 is empty except for the right most limit γ0+ .
From a general theorem in semi-group theory [10, p.276, Thm1.18 p.307, prop 1.15p.305],
390 Theorem 1.6 that concerns the generator XF is equivalent to the following result that concerns
the semi-group e and illustrated on Figure 1.3.
tXF


t≥0

12
 Im(z)

−
etγ0
+
etγ0 Re(z)

+
etγ1
−
etγ1

Figure 1.3: Band spectrum of etXF in HW (M ; F ) for t > 0, from Theorem 1.7. For k ∈ N, and
+ −
γk± defined in (1.6), the operator etXF has discrete spectrum in the annulus etγk+1 < |z| < etγk
+
(if γk+1
+
< γk− ) called “spectral gap”. It has also discrete spectrum on |z| > etγ0 and possibly
essential spectrum elsewhere (the colored annuli). This picture is somehow the exponential of
Figure 1.2.

Theorem 1.7 (Ring spectrum). Let k ∈ N. For any t > 0, the op-
tXF
erator
n e : H W (M ; F )
o → H W (M ; F ) has discrete spectrum on the ring
+ − +
z ∈ C, etγk+1 < |z| < etγk (if non empty, i.e. if γk+1
+
< γk− ) and on |z| > etγ0 .

Remark 1.8. An immediate consequence of Theorem 1.7, if γk+1 +

< γk− for some k ∈ N, is the
possibility to perform
n a contour integral of othe resolvent of e tXF
at fixed value of t > 0, on
+
t(γk+1 +ϵ)
395 the circle Ct,ϵ := z ∈ C s.t. |z| = e with ϵ > 0 such that γk+1 +
+ ϵ < γk− and such
that there is no eigenvalue on this circle. Using Cauchy u integral
and holomorphic functional
−1
calculus, we get a spectral projector Π[0,k] = Id − 2πi Ct,ϵ z − etXF
1
dz on the first k external
bands and one can approximate e tXF
for t > 0 by the spectral restriction etXF Π[0,k] with the
following estimate:
+
etXF − etXF Π[0,k] H (M ;F ) ≤ Cϵ et(γk+1 +ϵ) ,
W

400 where etXF Π[0,k] is compared to a quantum evolution operator in this paper. This operator can
be used to describe with great accuracy the decay of correlations ⟨v|etXF u⟩L2 (M ;F ) as t → ∞,
that is the question raised in the beginning of this paper.

Remark 1.9. The discrete spectrum of XF (and eigenspace) are said “intrinsic” because they
405 do not depend on the space HW (M ; F ), see [14, Thm 1.5]. However more refined spectral
properties as the norm of the resolvent (z − XF )−1 H (M ;F ) depend on the choice of the space
W
HW (M ; F ) and play an important role in this paper. For example the uniform boundness of
the norm of the resolvent in the gaps given in theorem 1.6 (implying theorem 1.7) is obtained
with the weight function W and Sobolev space HW (M ; F ) from [18] and defined below in (4.15)
410 because we have decay outside a parabolic neighborhood of the trapped set Σ, saturating the
uncertainty principle. This property of uniform boundness is not true with the weight function
and Sobolev space defined in [14] where the decay takes place only outside a (much bigger)
conical neighborhood of the trapped set Σ.

13
 1.3.2 Weyl law
415 We
S have dimM  = 2d + 1, with d = dimEs,u = dimNu,s . For some given k ∈ N, suppose Bk ∩
k′ ̸=k Bk′ = ∅, i.e. that the band Bk is isolated. Theorem 1.10 below shows that the density of
d
discrete eigenvalues of XF in band Bk in the limit ω → ∞ converges to rank (Fk ) Vol (M ) (2π)
ω
d+1

with  
k+d−1
rank (Fk ) = rank (F ) . (1.16)
(1.5) d−1
It is analogous to the usual Weyl law.

Theorem 1.10 (Weyl law for isolated bands). For any k ∈ N and ϵ > 0, such that
+
γk+1 < γk− − ϵ and γk+ + ϵ < γk−1
−
(this last condition is only for k ≥ 1)

1 rank (Fk ) Vol (M )

lim lim sup d
♯ {spect (XF ) ∩ Boxω,δ } − =0 (1.17)
δ→+∞ ω→±∞ |ω| δ (2π)d+1

with the spectral box

Boxω,δ := z ∈ γk− − ϵ, γk+ + ϵ × i [ω, ω + δ] .

  

420

The proof is obtained in Section 7.

Remark 1.11. As the proof shows, the result holds true more generally for a group of isolated
bands indexed by k ∈ [k1 , k2 ] with k1 ≤ k2 , i.e. assuming γk+2 +1 < γk−2 and γk+1 < γk−1 −1 . For the
case k1 = 0 we only have to assume γk+2 +1 < γk−2 . Then in (1.17), the interval γk− − ϵ, γk+ + ϵ
has to be replaced by γk−2 − ϵ, γk+1 + ϵ and rank (Fk ) has to be replaced by rank k=k1 Fk .
  Lk2
425

1.3.3 Ergodic concentration of the spectrum

We define in (8.1) the maximal and minimal exponents  − γ̌+
−
 k − of +the
k  ≤ γ̌
+
 bundle maps
etXFk
with respect to the contact volume on Σ, satisfying γ̌k , γ̌k ⊂ γk , γk . The following
theorem in addition to Theorem
 − 1.10  shows that most of the eigenvalues in band Bk belong to
430 the narrower band Re (z) ∈ γ̌k , γ̌k .
+

Theorem 1.12 (Ergodic concentration of the spectrum). Let k ∈ N. Assume that

+
γk+1 < γk− and γk+ < γk−1
−
(this last condition is only for k ≥ 1). For any ϵ > 0 small
enough, we have
1
(1.18)

lim lim sup ♯ spect (XF ) ∩ Stripsω,δ = 0
δ→+∞ ω→±∞ |ω|d δ

with
Stripsω,δ := z ∈ γk− − ϵ, γk+ + ϵ \ γ̌k− − ϵ, γ̌k+ + ϵ × i [ω, ω + δ] .
    

The proof is given in Section 8. For example, in the special case of band k = 0 and
rank (F ) = 1 we have rank (F0 ) = 1, and from ergodicity of X, most of eigenvalues accumu-
(1.16)
late on the vertical line Z
1
γ̌0− = γ̌0+ = Ddm
Vol (M ) M

435 being the space average of the “damping function” D ∈ C (M ; R) defined by D (m) := V (m) +
1
2
divX/Es (m) with the potential function V given in (2.5), see [16].

14
 1.3.4 Emerging quantum dynamics on vector bundles
The band structure of the Ruelle spectrum and Weyl law described above reflect in fact a
deeper geometric phenomenon that can be explained in terms of geometric quantization. Let
440 us explain this.

Wave-packet transform T : In paper [18, def. 4.23], see also (2.33), we introduce a wave-
packet transform (here S (T ∗ M ; F ) is the space of Schwartz sections of the bundle F → T ∗ M )

T : C ∞ (M ; F ) → S (T ∗ M ; F ) , (1.19)

satisfying T † T = IdL2 (M ;F ) and that is the basic tool used for micro-local analysis. For example,
using the characteristic function χΣ,σ = 1n∥ρ +ρ ∥ ≤σo : T ∗ M → [0, 1] defined in (3.30), for the
u s gρ

445 neighborhood of the set Σ = R A ⊂ T M at a distance σ > 0 measured by a specific metric g

∗ ∗

(given in Lemma 2.8), we define the operator

Op (χΣ,σ ) = T † χΣ,σ T : C ∞ (M ; F ) → C ∞ (M ; F ) (1.20)

(3.31)

that restricts functions to their micro-local components near Σ.

Theorem 4.6 will show that the norm decays exponentially fast outside the trapped set Σ
with an exponential rate, i.e. we have etXF (Id − Op (χΣ,σ )) H (M ) ≤ Ce−tΛ for t ≥ 0, with
W
450 a rate Λ > 0 arbitrarily large if σ is large enough (depending on t). This implies that for our
study, we can consider only the component of the dynamics etX Op (χΣ,σ ) near the trapped set.

Quantization: Recall the vector bundle F (Es ) → M defined in (1.8) and the lifted bundle
F (Ns ) → Σ. For a map a ∈ C β (M ; End (F (Es ))) (or even a ∈ C β (M ; L (F (Es ) , F ′ (Es ))))
and time t ∈ R, we will define a quantum operator in Definition 4.27

OpΣ etXF ã : C ∞ (M ; F ) → C ∞ (M ; F )

455 where ã ∈ C β (Σ; L (F (Ns ))) is the lifted map.

For two operators A, B, we define in (3.32) an equivalence relation A ≈ B if their differ-
ence in any σ−neighborhood of the trapped set Σ and for large enough frequencies becomes
“negligible”. A consequence is that for any σ > 0,

∥Op (χΣ,σ ) (A − B) Op (χΣ,σ ) (Id − Op (χω ))∥HW (M ) → 0, (1.21)

ω→∞

where Op (χΣ,σ ) is given in (1.20) and for ω > 0, Op (χω ) : S ′ (M ; F ) → S (M ; F ) is a smoothing

460 operator hence compact that removes frequencies larger than ω, defined as follows. Let gM be
an arbitrary smooth metric on M . Let χω : T ∗ M → R+ given by

χω (.) = 1{∥.∥ } (1.22)

gM ≤ω

for ρ ∈ T ∗ M and let

Op (χω ) := T † χω T . (1.23)
In Theorem 4.29 we will obtain
Theorem 1.13 (Approximation of the dynamics by quantum operator). For any
t ∈ R,

etXF ≈ OpΣ etXF . (1.24)

465 In Theorem 4.37 we will obtain7

7
Recall from section 1.1 that a symbol a ∈ C β M ; End F[0,K] (Es ) can be extended to a ◦ T[0,K] ∈
C β (M ; L (F, F ′ )) using Taylor projectors T[0,K] : F (Es ) → F[0,K] (Es ) ⊂ F ′ (Es ).

15
 Theorem 1.14

(Composition formula). For any symbols a, b ∈
C β M ; L F[0,K] (Es ) , any t, t′ ∈ R, we have
 ′   ′

OpΣ etXF ã OpΣ et XF b̃ ≈ OpΣ etXF ãet XF b̃ . (1.25)

In particular, since for any k, k ′ ∈ N we have for the Taylor projectors [Tk , Tk′ ] = δk=k′ Tk
and Tk , e = 0, we will get in corollary 4.41 that
tXF


   ′   ′

OpΣ etXF T̃k OpΣ et XF T̃k′ ≈ δk=k′ OpΣ e(t+t )XF T̃k . (1.26)

We will get a general continuity theorem 4.31,

Theorem 1.15 (continuity theorem

for F.I.O.). There exists C > 0, such that for any
β
symbol a ∈ C M ; L F[0,K] (Es ) and t ∈ R,

OpΣ etXF ã H (M ) ≤ C etXF ã H (F ) . (1.27)

W W
470

As a corollary we will obtain in corollary 4.35 some estimates that reflect Eq.(1.6) at the
quantum level, such as

Corollary 1.16. For any k ∈ N, ∀ϵ > 0, ∃Ck,ϵ > 0,∀t ≥ 0,

  +
OpΣ etXF T̃k ≤ Ck,ϵ et(γk +ϵ) , (1.28)
HW (M )
  −
OpΣ e−tXF T̃k ≤ Ck,ϵ e−t(γk −ϵ) . (1.29)
HW (M )

The band spectrum of the Ruelle spectrum in theorem 1.6 and theorem 1.7 is a direct
475 manifestation of these boundness estimates together with the algebraic structure of (1.26).
In theorem 4.43 we will get some trace formula that will be used to prove Weyl law in
section 7.

1.3.5 Horocycle operators

In the special case of the geodesic flow on M = Γ\SL2 (R), the Lie algebra of the sl2 (R) provides
480 left invariant vector fields u ∈ C ∞ (M ; Eu ), s ∈ C ∞ (M ; Es ) that satisfies
[s, u] = X (1.30)
and called respectively unstable/stable horocycle vector fields (and more generally for the al-
gebra of so (n, 1)), see [19, 9, 25] where it is shown that these operators s, u map respectively
band Bk to band Bk−1 , Bk+1 .
In the general case of a contact Anosov vector field X, we propose below some “approxi-
485 mate horocycle operators” OpΣ (ιs ) , OpΣ (u) that satisfy some commutator relation simi-
lar to (1.30) but in the semiclassical limit ω ≫ 1. They also respectively map band Bk to Bk∓1
(at first order).
If u ∈ C β (M ; Eu ), s ∈ C β (M ; Es ) are respectively Hölder continuous sections of the un-
stable/stable bundles of the Anosov dynamics, we naturally associate to them the following
490 “operator-valued symbols” (see Section 9 for a more detailed definition)
ιs : C β (Σ; Fk (Ns )) → C β (Σ; Fk−1 (Ns )) (1.31)
by point-wise tensor-contraction with the polynomial in (1.5) and
u : C β (Σ; Fk (Ns )) → C β (Σ; Fk+1 (Ns )) (1.32)

16
 by point-wise symmetric tensor product. We will see in Lemma 9.2 that they satisfy the point-
wise Weyl algebra (also called symplectic Clifford algebra)

[ιs , u] T̃k = ω (.) ((dA) (s, u)) T̃k . (1.33)

where ω (.) : ωA (m) ∈ Σ → ω ∈ R is the frequency map, dA is the symplectic form on Eu ⊕ Es

495 and T̃k the projector on Fk .

Theorem 1.17 (Quantized Weyl algebra). Let K ∈ N and T̃[0,K] the projector on
F[0,K] . For k < K we have
h    i    
OpΣ ιs T̃[0,K] , OpΣ (u) T̃[0,K] OpΣ T̃k ≈ OpΣ ω (.) (dA) (s, u) T̃k (1.34)

and for any t ∈ R,    

etXF OpΣ ιs T̃k ≈ OpΣ ιdϕt s T̃k etXF (1.35)
  


etXF OpΣ (u) T̃k ≈ OpΣ dϕt u  T̃k etXF (1.36)

The proof is obtained in Section 9.

Remark 1.18. To compare (1.34) with (1.30), notice that ω (.) is the principal symbol of X, see
[18, prop.4.42].

500 1.4 Related papers

We present here some related previous works specifically on the topics of band structure of
the Ruelle spectrum for contact Anosov dynamics and the emergence of an effective quantum
dynamics as expressed in Eq.(1.12) using micro-local analysis. This paper is a continuation
along the following series of papers.
 
2 1
505 • In [11], the case of a contact U (1)-extension of the Arnold’s cat map M = on
1 1
T2 is treated (or any hyperbolic M ∈ SL2 (Z)). Band structure of the Ruelle spectrum
and emergence of the quantum cat map dynamics has been shown. This model is also
called “prequantum cat map” since its construction follows the prequantization proce-
dure of Kostant, Souriau, Kirillov that is equivalent to a contact U (1)-extension of the
510 Hamiltonian dynamics.
Remark 1.19. The symplectic form Ω on T2 is not exact so we need to consider a non
trivial line bundle L → T2 with curvature Ω, called prequantum line bundle. On the
opposite, in the model of contact Anosov flows considered in this paper the symplectic
form Ω = dθ on Σ ⊂ T ∗ M is exact, θ being the Liouville one-form (2.12), hence the
515 prequantum line bundle L → Σ is trivial and we have decided to ignore it, although its
presence manifests all along the paper, for example in the definition of the twist operator
(2.42). For a more correct geometrical description, we have to twist by L in the tensor
product F ⊗ L → Σ in (4.27).
• In [16], the case of a U (1)-extension of an arbitrary symplectic Anosov map ϕ : M → M
520 is considered (also called prequantum Anosov map as explained before). Band structure
of the Ruelle spectrum and emergence of the quantum dynamics has been shown. The
models in [11, 16] can be considered as toy models (or simplified models) for contact
Anosov flows considered in this paper because they are analogous (but not equivalent) to
a time one contact Anosov flow map.
525 • The paper [15] announced the band spectrum for contact Anosov flows, (less precise)
Weyl law and the method presented in this paper.

17
 • In the paper [17], one has considered the interesting choice of the bundle F = |detEs |1/2 ,
giving the trivial bundle (1.13). This model is treated in relation with the “semi-classical
zeta function” that generalizes the Selberg zeta function to non constant curvature. Tech-
530 nically, the case of the Hölder continuous bundle F = |detEs |1/2 is more tricky than the
case a smooth bundle F and needs to consider an extension to a Grassmanian bundle.
• C. Guillarmou and M. Cekic in [24] prove the first band for contact Anosov flows in
dimension 3 using horocycle operators. For the special case of hyperbolic manifolds,
endowed with an algebraic structure, band spectrum of Anosov dynamics has been studied
535 in [9], [25], [28], [3].
• In different situations, band structure and Weyl law for the spectrum of resonances has
been studied in [37],[36] for convex obstacles and by S. Dyatlov in [8] for regular normally
hyperbolic trapped sets.

1.5 Organization of the paper

540 First part, setup of a symbolic calculus for contact Anosov vector field. The first
part of paper is organized by starting with a general model and then adding more and more
assumptions to arrive at the model we are interested in and get more specific results.
In section 2 we present the micro-local analysis of a general non vanishing vector field X
on a manifold M . The strategy is to lift the analysis of the pullback operator etX on the
545 cotangent bundle T ∗ M using the wave-packet transform T : C ∞ (M ) → S (M ). In the high
frequency limit (semi-classical limit), the linearization of the flow ϕ̃t induced on T ∗ M , gives a
good description of the pullback operator etX in term of the differential map dϕ̃t on T T ∗ M .
This is obtained in theorem 2.16.
In section 3 we assume furthermore that X is a Reeb vector field, i.e. defined from smooth
550 contact one-form A on M . The purpose of this section is to obtain a precise description of
the dynamics etX not on all T ∗ M but only in a vicinity of the symplectization subset Σ :=
(R\ {0}) A ⊂ T ∗ M . This subset Σ is an invariant set for the dynamics ϕ̃t induced on T ∗ M . Our
description of the operator etX in theorem 3.19 uses the differential dϕ̃t on the vector bundle
T T ∗ M restricted to the base Σ ⊂ T ∗ M and denoted TΣ T ∗ M . We get an alternative description
555 in theorem 3.15 in terms of the normal bundle N := (T Σ)⊥ → Σ. Both descriptions will be
used later.
In section 4 we assume furthermore that X is a contact Anosov vector field on M, the
main subject of this paper. First, from this assumption, we have that the outside of the set Σ
is negligible for the description of etX for large time. This is theorem 4.6 and obtained using
560 escape functions, anisotropic Sobolev norm, results from previous papers. Secondly, the normal
bundle over Σ splits N = Ns ⊕ Nu in stable and unstable invariant sub-bundles. It is there that
we introduce the vector bundles Fk (Ns ) = |detNs |−1/2 ⊗ Polk (Ns ) ⊗ F → Σ for every k ∈ N. In
theorem 4.16 we obtain a description of the operator etX in terms of the dynamics induced on
these bundles Fk . We show that this description is a “geometric quantization framework” with
565 some adapted symbolic calculus. We introduce symbols in definition 4.24 and quantization in
definition 4.27 giving PDO and FIO. The rest of the section is to obtain classical and useful
results in symbolic calculus, as continuity theorem 4.31, composition theorem 4.37, Egorov
theorem 4.41 Trace formula 4.43.

Second part, some consequences. In the second part of the paper we use the setup of
570 the first part to deduce some specific spectral properties of a contact Anosov vector field X.
In section 5, we show how to use these properties to deduce the band structure of the Ruelle
spectrum of the vector field X. In section 6, we deduce the theorem 1.4 about emergence
of quantum dynamics. Section 7 gives the proof of the Weyl law and section 8 explains the
concentration in narrower bands. Section 9 gives definitions and proofs of theorem 1.17 for
575 approximate horocycle operators.

18
 Appendices: Appendix A gives some convention of notations used in this paper. Appendix B
contains additional comments about flows. Appendices C and D are quite important and sustain
the main argument used in this paper: for the analysis of the differential dϕ̃t on TΣ T ∗ M , we use
Bargmann transform,
  i.e. wave-packet transform on an Euclidean vector space with metaplectic
580 operators Op dϕ̃t obtained by quantization of the bundle map of linear symplectic maps
dϕ̃t : T T ∗ M → T T ∗ M . All the details of this are given in appendix C. We also use results for
linear expanding maps given in appendix D.
Acknowledgement. F. Faure acknowledges Claude Gignoux, Colin Guillarmou, Victor Maucout,
Malik Mezzadri, Stéphane Nonnenmacher for their support, for interesting and motivating dis-
585 cussions. F. Faure acknowledges M.S.R.I. and organizers of the micro-local semester 2020 where
a part of this work has been developed. F. Faure acknowledges partially supported by French
ANR Adyct, Grant Number ANR-20-CE40-0017 during this work. M. Tsujii acknowledges
partially supported by JSPS KAKENHI Grant Number 15H03627 and 22340035 during this
work.

590 2 Micro-local analysis of a non vanishing vector field X

In this section we consider a general smooth and non vanishing vector field X (not necessarily
Anosov) on a smooth closed manifold M (not necessarily contact). We review some results given
in [18]. We define the wave packet transform that is a linear operator T : S (M ) → S (T ∗ M )
and an isometry from L2 (M ) to L2 (T ∗ M ), i.e. satisfies T † T = Id where T † is the L2 -adjoint
595 operator. Then we recall the theorem of propagation of singularities 2.13, saying that the
Schwartz kernel of the lifted pull back operator T etX T † on S (T ∗ M ), with t ∈ R, decays very
∗
fast outside the graph of the flow ϕ̃t := (dϕt ) : T ∗ M → T ∗ M acting on the cotangent space
T ∗ M . Moreover by linearization, we give in theorem 2.16 an approximate description of this
Schwartz kernel in the neighborhood of the graph of ϕ̃t .

600 2.1 Vector field X and pull back operator etX

Let M be a C ∞ closed connected manifold. Let X ∈ C ∞ (M ; T M ) be a C ∞ vector field on
M , considered as a first order differential operator acting on smooth functions C ∞ (M ), i.e. in
local coordinates y = (y1 , y2 , . . . , ydimM ), X = j=1 Xj (y) ∂y∂ j . For t ∈ R, let
PdimM

ϕt : M → M (2.1)
d(u◦ϕt )
be the C ∞ flow defined by dt
= Xu, ∀u ∈ C ∞ (M ), i.e.

◦
etX u := u ◦ ϕt = ϕt u, (2.2)
◦
605 where (ϕt ) = etX denotes the pull back operator acting on functions (following notations in
Appendix A). See Appendix B.1 for additional remarks about the transfer operator.

2.2 More general pull back operator etXF with generator XF

It will be interesting to consider the following more general situation. Let π : F → M be a
smooth complex vector bundle of finite rank over M . Let XF a derivation over X, i.e. XF
610 is a first order differential operator acting on sections of F

XF : C ∞ (M ; F ) → C ∞ (M ; F ) (2.3)

satisfying the Leibniz rule

XF (f u) = X (f ) u + f XF (u) , ∀f ∈ C ∞ (M ; C) , ∀u ∈ C ∞ (M ; F ) . (2.4)

19
 Example 2.1. If F is the bundle of differential forms F = Λ• (T M ) (or a more general tensor
field). The Lie derivative XF (obtained from the differential dϕt ) satisfies (2.4), so XF is a
derivation over X.

615

Example 2.2. A simple but useful example is the trivial rank one bundle F = M × C. Let
V ∈ C ∞ (M ; C) called the potential function. The operator

XF = X + V (2.5)

satisfies (2.4), where V is seen as the multiplication operator by the function V . By integration
we get, with u ∈ C ∞ (M ), Rt s
etXF u = e| 0 V{z◦ϕ ds} u ◦ ϕt . (2.6)
| {z }
amplitude transport

620 More remarks about general pull back operators etXF are given in Section B.2.

2.3 Lifted flow ϕ̃t : T ∗ M → T ∗ M and its generator X̃

∗
In this section we define and describe the flow ϕ̃t = (dϕt ) : T ∗ M → T ∗ M induced from the
flow ϕt in (2.1). It will appear in Theorem 2.13 below, that for the analysis of etX , this flow ϕ̃t
on T ∗ M plays a major role.

625 2.3.1 Frequency function ω

We define the frequency function ω ∈ C ∞ (T ∗ M ; R) by

∀ρ ∈ T ∗ M, ω (ρ) := ρ (X) . (2.7)

that measures the oscillations of ρ along the flow direction X.

Remark 2.3. In terms of micro-local analysis, the function iω is the principal symbol of the
vector field X, see [14, footnote page 332.].

630 2.3.2 The lifted flow ϕ̃t : T ∗ M → T ∗ M

For m ∈ M , we denote dϕt : Tm M → Tϕt (m) M , the differential of the flow map ϕt : M → M .
By duality we get a map
∗
ϕ̃t := dϕt : Tm∗ M → Tϕ∗−t (m) M. (2.8)
defined by

∗
⟨v| dϕt (ρ)⟩Tm M = ⟨dϕt v|ρ⟩Tϕ∗t (m) M , ∀ρ ∈ Tϕ∗t (m) M, v ∈ Tm M, m ∈ M. (2.9)

Notice that ϕ̃t is a lift of the inverse map ϕ−t map, i.e.
 
π ϕ̃t (ρ) = ϕ−t (π (ρ)) , (2.10)

635 with the bundle projection map π : T ∗ M → M .

20
 2.3.3 The canonical Liouville one form θ and the symplectic two form Ω
Let θ be the canonical Liouville one form on the vector bundle π : T ∗ M → M [32, p.90][6,
p.7][7, p.56]: at any point ρ ∈ T ∗ M , θρ is defined by

∀V ∈ Tρ T ∗ M, θρ (V ) := ρ (dπ (V )) . (2.11)

Using the local coordinates (yj )j=1,...,dimM on M and dual coordinates (ηj )j on T ∗ M , this gives

dimM
X
θ= ηj dyj . (2.12)
j=1

640 Let Ω be the canonical symplectic form

dimM
X
Ω := dθ = dηj ∧ dyj , (2.13)
j=1

that gives a bundle isomorphism Ω̌ : T T ∗ M → T ∗ T ∗ M defined by

∀V ∈ T T ∗ M, Ω̌ (V ) := Ω (V, .) ∈ T ∗ T ∗ M.

Using notation in section A for the pullback, the definition (2.11) of θ on T ∗ M is equivalent to
the fact that for any smooth one form α ∈ C ∞ (M ; T ∗ M ), considered as a map α : M → T ∗ M ,
considering its differential dα : T M → T T ∗ M , we have

(dα)◦ θ = α, (2.14)

645 and Kerθ = Ker (dπ) away from the zero section. Taking the differential we get

(dα)◦ Ω = dα. (2.15)

We define the Euler (or Liouville) vector field on T ∗ M

E := Ω̌−1 (θ) . (2.16)

In coordinates, E = ηj ∂η∂ j . Using Cartan formula we get for the Lie derivatives
PdimM
j=1

LE θ = θ, LE Ω = Ω.

2.3.4 The vector field X̃ on T ∗ M

Let us write X̃ for the vector field on T ∗ M that generates the flow ϕ̃t (2.8). Notice from (2.10)
650 that  
(dπ) X̃ = −X. (2.17)

Lemma 2.4. [21, thm 2.124p.112, in coordinates],[34, prop. 1.21 p.15]. The Lie deriva-
tive of θ by the vector field X̃ vanishes:

LX̃ θ = 0, (2.18)

and X̃ is determined by
X̃ = Ω̌−1 (dω) , (2.19)
i.e. X̃ is the Hamiltonian vector field on (T ∗ M, Ω), with Hamiltonian function ω,
Eq.(2.7).

21
 Proof. [21, Thm 2.124 p.112, for a proof using coordinates],[34, Prop. 1.21 p.15] Let ρ ∈ T ∗ M
and V ∈ Tρ (T ∗ M ). For any t ∈ R we have
 ∗       
dϕ̃t θϕ̃t (ρ) (V ) = θϕ̃t (ρ) dϕ̃t V = ϕ̃t (ρ) dπ dϕ̃t (V )
(2.11)
∗
= ⟨ dϕt (ρ) |dϕ−t (dπ (V ))⟩ = ρ (dπ (V )) = θρ (V )


(2.8) (2.9)
 ∗
⇔ dϕ̃t θϕ̃t (ρ) = θρ ⇔ etLX̃ θ = θ ⇔ LX̃ θ = 0.
(B.3)

This gives (2.18). Notice that

    
ω (ρ) = ρ (X) = −ρ dπ X̃ = −θρ X̃ = −ιX̃ θρ . (2.20)
(2.7) (2.17) (2.11)

Hence, using Cartan formula LX̃ θ = d (ιX̃ θ) + ιX̃ dθ we get

 
dω = −d (ιX̃ θ) = −LX̃ θ + ιX̃ dθ = Ω X̃, . ,
(2.20)

655 that gives (2.19).

 
Remark 2.5. As a consequence of (2.19) we have X̃ (ω) = Ω X̃, X̃ = 0, i.e. for every
(2.19)
ω ∈ R, the frequency level
Σω := ω −1 (ω) = {ρ ∈ T ∗ M, X (ρ) = ω} ⊂ T ∗ M (2.21)
that is a affine sub-bundle of T ∗ M is preserved by ϕ̃t .

660 Remark 2.6. For ω ̸= 0, the one form − ω1 θ is a contact one form on the frequency level
 Σω and

 
 
X̃ is its Reeb vector field, because − ω θ X̃
1
= 1 and d − ω θ X̃, . = − ω Ω X̃, . =
1 1
(2.20)
− ω1 dω = 0.
(2.21)

2.4 Metric g on T ∗ M
Let us suppose that we have an atlas M = Jj=1 Uj where Uj are open subset on M such that
S
665 for j = 1, . . . J,
κj : m ∈ Uj → y = (x, z) ∈ Vj ⊂ RdimM
x
−1
× Rz (2.22)
are flow box coordinate, i.e. (dκj ) (X) = ∂z ∂
. On each chart Uj , let us denote η = (ξ, ω) ∈
R dimM −1
×R the dual
coordinates associated to (x, z). Hence (y, η) = (x, z, ξ, ω) are coordinates
on T ∗ RdimM
x
−1
× Rz .
Let us recall the definition and properties of the metric g on T ∗ M introduced in [18, section
670 4.1.2] with parameters that we take here α⊥ = 21 and α∥ = 0. For s ∈ R, we set

1/2
⟨s⟩ := 1 + s2 ∼ |s| . (2.23)
|s|≫1

For η ∈ RdimM , let

δ ⊥ (η) := ⟨|η|⟩−1/2 . (2.24)
In coordinates, we define the metric g on T ∗ RdimM = R2dimM
ϱ at a point ϱ = (y, η) ∈ RdimM ×
RdimM by  2  2
dx dξ
g (ϱ) := + −1 + dz 2 + dω 2 . (2.25)
δ ⊥ (η) ⊥
δ (η)
This gives a metric gj := (dκj )∗ g on T ∗ Uj . It is shown in [18, Lemma 4.6] that on each
675 intersection Uj ∩ Uj ′ , the metric gj and gj ′ are relatively bounded by each other uniformly in
η. Each metric gj is compatible with the canonical symplectic form Ω on T ∗ M given in (2.13).

22
 Remark 2.7. an important property of the metric g that will be essential in our analysis is that
the size of a unit ball projects on the coordinates x as a set of size of order δ ⊥ (η) = ⟨|η|⟩−1/2
that goes to zero as |η| → +∞. See [18, Figure 2.3].
680 We recall the definition that Ω, g are compatible in Definition C.13.

Lemma 2.8. There exists a global smooth metric g on T ∗ M that is also compatible with
Ω and relatively bounded with respect to every metric gj on any chart T ∗ Uj .

Proof. Let G be a global metric on T ∗ M defined by

J
X
|detdκj | χ2j ◦ κj ◦ π · gj (2.26)


G :=
j=1

where π : T ∗ M → M denotes P the bundle projection and χj ∈ C0∞ (κj (Uj ) ; R+ ) form a quadratic
partition of unity on M , i.e. j χ2j ◦ κj |detdκj | = 1, see [18, Lemma 4.2]. Recall that on each
685 intersection Uj ∩ Uj ′ , the metric gj and gj ′ are uniformly equivalent (i.e. relatively bounded)
that we write gj ≍ gj ′ . Consequently, from finitness of the sum (2.26), on any chart T ∗ Uj , we
have
G ≍ gj . (2.27)
We define
K := Ω̌−1 Ǧ : T T ∗ M → T T ∗ M, (2.28)
where Ǧ, Ω̌ : T T ∗ M → T ∗ T ∗ M are defined from G, Ω√as in (A.1). We consider the polar
690 decomposition of K written K = J |K| where |K| = K †g K is positive definite Hermitian
(bundle maps over T ∗ M ) and J is an almost complex structure because, using that Ǧ∗ = Ǧ
and Ω̌∗ = −Ω̌ we get K †g = Ǧ−1 K ∗ Ǧ = Ǧ−1 Ǧ∗ Ω̌∗−1 Ǧ = −Ω̌−1 Ǧ = −K, hence J 2 =
(A.3) (2.28)
K2
= −Id. Since gj is compatible with Ω then ǧj = Ω̌Jj with Jj2 = −Id. Thus Ω̌K =
( †g K )
K (2.28)
Ǧ ≍ gj = Ω̌Jj gives K ≍ Jj then
(2.27)
|K| ≍ Id. (2.29)
695 We define the metric g by
ǧ := Ω̌J : T T ∗ M → T ∗ T ∗ M. (2.30)
By construction, the global metric g is smooth, compatible with Ω and relatively bounded with
respect to every gj on any chart T ∗ Uj because ǧ = Ω̌J ≍ Ω̌J |K| = Ω̌K = Ǧ ≍ ǧj .
(2.29) (2.28)

We will denote distg (ρ′ , ρ) the distance between two points ρ, ρ′ ∈ T ∗ M according to the
metric g. It is shown in [18, Lemma 4.5] that the metric g is geodesically complete.
700 We will also use a property of slow variation of the metric g given in [18, Lemma 4.12] and
recalled later in (2.59). Finally we have from [18, Lemma 4.6] the “Lipschitz property of g”,
namely, for any t ∈ R, there exists a constant Ct > 0 such that for any ρ, ρ′ ∈ T ∗ M ,
 
distg ϕ̃ (ρ), ϕ̃ (ρ ) ≤ Ct distg (ρ, ρ′ ) .
t t ′
(2.31)

where ϕ̃t : T ∗ M → T ∗ M is defined in (2.8).

23
 2.5 Wave-packet transform T
705 As in [18, def. 4.23], for every coordinate chart Uj ⊂ M with j ∈ {1, . . . , J} and ρ ∈ T ∗ Uj we
define a wave packet8 Φj,ρ ∈ C ∞ (M ; C) and the “wave packet transform”
(
C ∞ (M ; C) → S T ∗ M ; CJ


T :
, (2.33)
u → ⟨Φj,ρ |u⟩L2 (M ) ρ∈T ∗ M,j∈{1...J}

that satisfies the ”resolution of identity on C ∞ (M )” [18, prop. 4.24]

Id/C ∞ (M ) = T † T (2.34)
where !
dρ
T † : L2 T ∗ M ⊗ CJ ; → L2 (M ; dm)
(2π)dimM
is the L2 -adjoint of T where dρ ≡ dydη is the canonical measure on T ∗ M . We will introduce
the wave packet projector onto the image of T , see [18, prop. 4.27]:
P := T T † . (2.35)
710 Remark 2.9. Below for simplicity, we will ignore ⊗CJ in the notation, i.e. ignore chart compo-
nents, though keeping it in mind.

2.6 Description of the operator etX with the bundle T T ∗ M

Remark 2.10. From now on we will often use the knowledge about Bargmann transform that
is summarized in Section C of the appendix and we ask the readers to check the notation and
715 their knowledge. We will often refer to this Section C.
We first introduce a definition that will be used to express that some operator R is “under
control” or “negligible” in our analysis. It will mean that the Schwartz kernel of T RT † decays
very fast outside the graph of ϕ̃t defined in (2.8) and moreover that on the graph, the Schwartz
kernel is bounded by the polynomial growth ⟨|ρ|⟩m .

Definition 2.11. Let t ∈ R and m ∈ R. We define Ψm ϕ̃t

as the set of operators R :
S (M ) → S (M ) such that for any N > 0, there exists a constant CN,t > 0 such that for
′

any ρ, ρ′ ∈ T ∗ M
D  E−N
⟨δρ′ |T RT † δρ ⟩L2 (T ∗ M ) ≤ CN,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩m . (2.36)
720

A useful consequence of (2.36) is the following Lemma that uses the truncation operator
Op (χω ) defined in (1.23).

Lemma 2.12. If R ∈ Ψm
ϕ̃t
with some m < 0, then ∃Ct > 0,

∥R (Id − Op (χω ))∥L2 (M ) ≤ Ct ω m → 0. (2.37)

ω→+∞

8
We refer to [18, def. 4.23] for the precise expression of a wave packet Φj,ρ . Here it is enough to say that
on a chart Uj , for a given ϱ = (y, η) ∈ T ∗ Rn+1 and in local coordinates y ′ ∈ Rn+1 , then for large |η| ≫ 1, the
function Φj,ρ is equivalent to a Gaussian wave packet in “vertical Gauge”:
 
2
Φj,ϱ (y ′ ) ∼ aϱ χ (y ′ − y) exp iη. (y ′ − y) − ∥y ′ − y∥gϱ (2.32)

with y ′ = (x′ , z ′ ) ∈ Rn+1 and where χ ∈ C0∞ Rn+1 is some cut-off function with χ ≡ 1 near the origin and


2
2 2 (x′ −x)
aϱ > 0 is such that ∥Φj,ϱ ∥L2 (Rn+1 ) = 1 and ∥y ′ − y∥gϱ = |z ′ − z| + δ ⊥ (η)
is obtained from the metric (2.25).

24
 Proof. We follow the same notations and techniques as in [18, section 4.2]. We have that

∥R (Id − Op (χω ))∥L2 (M ) = T R (Id − Op (χω )) T † L2 (T ∗ M )

(2.34)

= T RT † (P − Pχω P) L2 (T ∗ M )
.
(2.35)

From [18, Prop. 4.27], we have ∀N ≥ 0, ∃CN > 0, ∀ρ, ρ′ ∈ T ∗ M ,

−N
⟨δρ′ |Pδρ ⟩L2 (T ∗ M ) ≤ CN ⟨distg (ρ′ , ρ)⟩ , (2.38)

725 i.e. the Schwartz kernel of P decays fast outside the diagonal and is uniformly bounded on the
diagonal. Using definition of χω in (1.22) we have (the exponent 1/2 comes from the metric g
in (2.25)), ∀N ≥ 0, ∃CN > 0, ∀ρ, ρ′ ∈ T ∗ M ,
D   E−N
−N
⟨δρ′ |Pχω Pδρ ⟩L2 (T ∗ M ) ≤ CN ⟨distg (ρ′ , ρ)⟩ max 0, ω −1/2 ω − ∥ρ∥gM .

Together with (2.36) we deduce that ∀N ≥ 0, ∃CN,t > 0, ∀ρ, ρ′ ∈ T ∗ M ,

D  E−N
†
⟨δρ′ |T R (Id − Op (χω )) T δρ ⟩L2 (T ∗ M ) ≤ CN,t distg ρ , ϕ̃ (ρ) ′
⟨|ρ|⟩m
t

D   E−N
−1/2
max 0, ω ω − ∥ρ∥gM .

Using Shur Lemma [18, Lemma 4.38], that estimates the L2 norm operator from the Schwartz
kernel, we deduce that if m < 0, then ∃Ct > 0,

∥R (Id − Op (χω ))∥L2 (M ) ≤ Ct ω m → 0.

ω→+∞

730

2.6.1 Propagation of singularities. First estimate for the wave front set.
The next theorem is similar to the description of evolution of the wave-front set in micro-local
analysis as in [38, Prop. 9.5, page 29.] and often called “propagation of singularities”. It gives
some first description of the pull back operator etX that we will need to improve later. Recall
735 the notation Ψ0ϕ̃t from definition 2.11.

Theorem 2.13 (Propagation of singularities). [18, Thm 4.51]. For each t ∈ R,

etX ∈ Ψ0ϕ̃t . (2.39)

Remark 2.14. If ρ′ and ϕ̃t (ρ) are in bounded distance from each other, theorem 2.13 says
nothing very informative for the value of the Schwartz kernel ⟨δρ′ |T etX T † δρ ⟩L2 (T ∗ M ) . The next
Theorem 2.16 below will complete this lack of information by giving an approximate expression
740 for the Schwartz kernel of T etX T † in the neighborhood of the graph of ϕ̃t , and this will be
useful later to get Theorem 3.15. We first need to introduce some operators.

2.6.2 Definition of some operators

Below we view the (usual) space of Schwartz functions S (T T ∗ M ) as the set of functions u :
ρ ∈ T ∗ M → uρ ∈ S (Tρ T ∗ M ), i.e. we have a natural identification, with induced topology,

S (T T ∗ M ) ≡ S (ρ ∈ T ∗ M ; S (Tρ T ∗ M )) . (2.40)

25
745 The operator exp g◦ . The Riemannian manifold (T ∗ M, g) is geodesically complete [18, Lemma
4.5]. Let us write
exp : T T ∗ M → T ∗ M (2.41)
for the exponential map associated to the metric g on T ∗ M of Lemma 2.8. From this map we
have the pull back operator exp◦ : S (T ∗ M ) → S (T T ∗ M ) but we consider instead the twisted
pull back operator defined as follows
g◦ := eiφ exp◦
exp : C ∞ (T ∗ M ) → C ∞ (T T ∗ M ) , (2.42)
750 with a phase function φ : T T ∗ M → R that corresponds to a change of trivialization to pass
from the vertical gauge given by the wave-packet transform (2.32) to the local radial gauge as
explained in section C.2. Explicitly, this phase function φ is given by, for ρ ∈ T ∗ M , v ∈ Tρ T ∗ M ,
1
φ (v) := −θ (v) − D2 f (v) (2.43)
2
with the Liouville one form θ (v) = ρ (dπ (v)) given in (2.11) and in the second part the function
f := θ ◦ (d exp)ρ − θρ : Tρ T ∗ M → R made with the differential (d exp)ρ : Tρ T ∗ M → T T ∗ M
755 (with the canonical identification of vector spaces T Tρ T ∗ M ≡ Tρ T ∗ M ). One has f (0) = 0 and
Df (0) = 0, so the quadratic form D2 f : Tρ T ∗ M → R called Hessian, is well defined.
Remark 2.15. Geometrically the phase φ reflects the existence of a (trivial) complex line bundle
L over T ∗ M with connection given by the canonical Liouville one form with respect to a global
trivialization. This line bundle is usually called the prequantum line bundle. As mentioned
760 before in remark 1.19, we can ignore this line bundle since it is trivial so we can use a global
section, giving phases like in (2.42).To understand better the phase φ in (2.42) it may be useful
to express it in local coordinates. Let y ∈ RdimM be local coordinates on M as in (2.22)
and η ∈ RdimM be dual coordinates on T ∗ M . For a (local) flat metric g on T ∗ M one has
ρ′ = expρ (v) = (yρ + yv , ηρ + ηv ) hence v ′ = (d exp)ρ (v) = (yv , ηv ). The Liouville one-form
765 is θρ (v ′ ) = ηρ yv and θρ′ (v ′ ) = ηρ′ yv = (ηρ + ηv ) yv giving the function f (v) = ηv yv which is
quadratic, i.e. D2 f = f (due to absence of non linearities in this simple setting of flat metric).
The first term in (2.43) is e−iθρ (v) = e−iηρ yv and is there to remove high oscillations coming from
i 2 i
wave-packet transform at point ρ ∈ T ∗ M . The second term in (2.43) is e− 2 D f (v) = e− 2 ηv yv
and is used to pass from vertical gauge to radial gauge in accordance with formula (C.6). In
770 summary, for Euclidean metric the phase function is
1
φρ (yv , ηv ) = −ηρ yv − ηv yv . (2.44)
2

Truncation operator χλ . Let 0 < λ < 1/2 that is fixed9 in this paper. The operator χλ
truncates functions in a ball of radius ⟨|ρ|⟩λ/2 centered on the zero section of T T ∗ M :

χλ := 1n∥.∥ λ/2
o : S (T T ∗ M ) → S ′ (T T ∗ M ) , (2.45)
gρ ≤⟨|ρ|⟩

i.e. for u ∈ S (T T ∗ M ), ρ ∈ T ∗ M , v ∈ Tρ T ∗ M ,
(
u (v) if ∥v∥gρ ≤ ⟨|ρ|⟩λ/2
χλ u (v) = (2.46)


.
0 otherwise.

The restriction operator r0 . Let

r0 : S (T T ∗ M ) → S (T ∗ M )
775 be the map that restricts a function u ∈ S (T T ∗ M ) (written u (ρ, v), ρ ∈ T ∗ M , v ∈ Tρ T ∗ M ),
to its value at the zero section (r0 u) (ρ) := u (ρ, 0).
9
From the result (2.49) and estimate (2.37), we see that λ close to zero gives better results, but in this paper
it is enough to take λ = 1/4.

26
  
The metaplectic operator Op dϕ̃t . Recall the map ϕ̃t : T ∗ M → T ∗ M defined in (2.8).
˜
 −◦
Its differential dϕ̃ : T T M → T T M gives a push-forward operator that we denote dϕ̃t
t ∗ ∗
:
S (T T M ) → S (T T M ). As in (C.17) we define the “metaplectic correction”
∗ ∗

!!!1/2
  1
 −1 †  −1
Υ dϕ̃t := det Id + dϕ̃t dϕ̃t ,
2
that is a positive function T ∗ M . We also denote P : S (T T ∗ M ) → S (T T ∗ M ) the Bergman
projector in radial gauge as defined in (C.26), that is here a fiber-wise operator. As in definition
C.33, we define the fiber-wise metaplectic operator over the map ϕ̃t :
    1/2  −◦
˜ dϕ̃t
Op = Υ dϕ̃t P dϕ̃t P. (2.47)
(C.33)

780 2.6.3 Linear approximation of the kernel

In the next theorem we use and compose the operators defined in the previous section 2.6.2
with the operator T in (2.33), and obtain a good approximation of the pull back operator etX
that improves the information given in Theorem 2.13. The rough idea of this approximation is
based on two properties: first, for ρ ∈ T ∗ M , with |ρ| ≫ 1, the metric is slowly varying hence
785 the metric g can be approximated by a Euclidean metric in a neighborhood of ρ. Second, the
unit balls of the metric g, transversely to the flow, have very small size projected on the base
M (see remark 2.7), hence the dynamics can be approximated by its linearization i.e. by the
action of the differential dϕ̃t .

Theorem 2.16 (“More precise expression for the propagation of singularities”).

For any 0 < λ < 21 , for any t ∈ R, we have
 
˜ dϕ̃t χλ exp
etX = T † r0 Op (2.48)



g◦ T + Rt

with some operator Rt that satisfies

− 1 +λ
Rt ∈ Ψϕ̃t2 . (2.49)

790 Remark 2.17. Since − 21 + λ < 0, we can apply Lemma 2.12 that guaranties that the remainder
operator Rt restricted to high frequencies is arbitrary small in operator norm.

Remark 2.18. The proof below shows that we also have the simpler expression
  1/2  −◦
etX = T † Υ dϕ̃t ϕ̃t T + Rt (2.50)
 −◦
with the push-forward operator ϕ̃t : S (T ∗ M ) → S (T ∗ M ) and the multiplication oper-
  1/2
795 ator by the metaplectic correction Υ dϕ̃t : S (T ∗ M ) → S (T ∗ M ). We already have
commented this expression in (1.10). One reason for us to later use (2.48) instead of (2.50) is
  1/2  
that (2.48) does not contains Υ dϕ̃ t
, but contains instead the operator Op dϕ̃t that
˜
    1/2
is fiberwise unitary in L2 , see appendix C.7 (in fact Op
˜ dϕ̃t contains Υ dϕ̃t ). On the
other hand, one advantage of (2.50) compare to (2.48) is that it deals directly with function in
800 S (T ∗ M ) instead of the more elaborate space S (T T ∗ M ). Later we will meet a similar situation,
see remark 3.20.

27
 Proof. Let 0 < λ < 1
2
and t ∈ R. We consider the operator Rt defined in (2.48), lifted to the
cotangent space:

R̃t := T Rt T †
 
˜ dϕ̃t χλ exp
= T etX T † − T T † r0 Op g◦ T T † . (2.51)




(2.48)

According to definition 2.11, to get (2.49), we have to show the following estimate for the
Schwartz kernel of R̃t , that for any N > 0, there exists a constant CN,t > 0 such that for any
ρ, ρ′ ∈ T ∗ M
D  E−N 1
′ t
⟨δρ′ |R̃t δρ ⟩ ≤ CN,t distg ρ , ϕ̃ (ρ) ⟨|ρ|⟩− 2 +λ . (2.52)

We will split the computation in two parts: (A) far from the graph of ϕ̃t and (B) near the
805 graph.

Step (A), far from the graph of ϕ̃t . Let us show that ∀λ′ > λ, ∃Ct,λ′ , ∀ρ, ρ′ ∈ T ∗ M ,
  ′
 
distg ρ′ , ϕ̃t (ρ) > Ct,λ′ ⟨|ρ|⟩λ /2 ⇒ ⟨δρ′ |r0 Op
˜ dϕ̃t χλ exp (2.53)


g◦ δρ ⟩ = 0.

From definition of the exponential map we have ∀ρ′′ ∈ T ∗ M, v ′′ ∈ Tρ′′ T ∗ M ,

distg (ρ′′ , exp ((ρ′′ , v ′′ ))) = ∥v ′′ ∥gρ′′ . (2.54)

From definitions of χλ and exp

g◦ , we have that ⟨δρ′′ ,v′′ | χλ exp ̸ 0 iff exp ((ρ′′ , v ′′ )) = ρ and


g◦ δρ ⟩ =
λ/2
distg (ρ′′ , ρ) = ∥v ′′ ∥gρ′′ ≤ ⟨|ρ′′ |⟩ . (2.55)
(2.54) (2.45)
   
The operator Op ˜ dϕ̃t δρ′′ ,v′′ ⟩ vanishes
˜ dϕ̃t is a bundle map over the map ϕ̃t hence ⟨δρ′ ,v′ |Op
810 if ρ′ ̸= ϕ̃t (ρ′′ ). Hence, setting ρ′′ = ϕ̃−t (ρ′ ), we have that
     
λ/2
˜ t λ g◦
̸ 0 ⇒ distg ρ , ϕ̃ (ρ) ≍ distg ϕ̃ (ρ ) , ρ ≤ ⟨|ρ′′ |⟩ .
′ t −t ′


⟨δρ′ |r0 Op dϕ̃ χ exp δρ ⟩ =
(2.31) (2.55)
(2.56)
From [18, Lemma 4.14] we have that ∀0 < λ < 1, ∀ϵ > 0, ∃Cϵ > 0, ∀ρ, ρ , ′

λ/2
distg (ρ′′ , ρ) ≤ ⟨|ρ′′ |⟩
⇒ ⟨|ρ′′ |⟩ ≤ Cϵ ⟨|ρ|⟩1+ϵ ,
  ′
hence the right hand side of (2.56) becomes distg ρ , ϕ̃ (ρ) ≤ Ct,λ′ ⟨|ρ|⟩λ /2 for any λ′ > λ and
′ t

we get (2.53).   ′
We assume that distg ρ′ , ϕ̃t (ρ) > Ct,λ′ ⟨|ρ|⟩λ /2 i.e. points (ρ, ρ′ ) ∈ T ∗ M × T ∗ M are “far”
 
from the graph of ϕ̃t . We have that etX ∈ Ψ0ϕ̃t and T † r0 Op ˜ dϕ̃t χλ exp g◦ T ∈ Ψ0 t . This



ϕ̃
implies that Rt ∈ Ψ0ϕ̃t . For any N > 0, take M > 1
λ′
and N ′ = N + M . We have
D  E−N ′ D  E−N D  E−M
⟨δρ′ |R̃t δρ ⟩ ≤ CN′ ′ ,t ′ t
distg ρ , ϕ̃ (ρ) ′′ ′ t
≤ CN ′ ,t distg ρ , ϕ̃ (ρ) ′ t
distg ρ , ϕ̃ (ρ)
(2.39)
D  E−N  ′
−M
≤ CN′′ ′ ,t distg ρ′ , ϕ̃t (ρ) Ct,λ′ ⟨|ρ|⟩λ /2
(hyp.)
D  E−N ′
≤ CN′′′′ ,t,λ′ distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩−M λ /2
D  E−N
≤ CN,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩−1/2

that gives (2.52).

28
815 Step (B), near the graph of ϕ̃t . We assume now that
  ′
distg ρ′ , ϕ̃t (ρ) ≤ Ct,λ′ ⟨|ρ|⟩λ /2 (2.57)

with λ < λ′ < 1, i.e. points (ρ, ρ′ ) ∈ T ∗ M × T ∗ M are “near” the graph of ϕ̃t . Let us explain
the strategy that we will pursue. Using the slow variation of the metric g given in [18, Lemma
4.12], we will (1) approximate the metric
 gρ′ by the Euclidean metric at point ϕ̃t (ρ). From
(2.25), the projected points π (ρ′ ) , π ϕ̃t (ρ) on M are at distance from each other that is
 
λ′ /2−1/2
820 O ⟨|ρ|⟩ transversely to the flow direction and this distance goes to zero if |ρ| → ∞.
So we will (2) approximate the map ϕ̃t by its differential, i.e. neglect non-linear terms. Finally
(3), even for a linear map on an Euclidean space, we need to show that the effect of the cutoff
χλ is negligible.
Having these three approximations in mind we write

⟨δρ′ |R̃t δρ ⟩ ≤ R1 + R2 + R3 ,
(2.51)

with  
R1 = ⟨δρ′ | T etX T † − Bϕ̃t (ρ) etX Bρ† δρ ⟩
825    
R2 = ⟨δρ′ | Bϕ̃t (ρ) etX Bρ† − e−iφϕ̃t (ρ) Op
˜ ρ dϕ̃tρ eiφρ δρ ⟩ (2.58)
    

R3 = ⟨δρ′ | e−iφϕ̃t (ρ) Op
˜ ρ dϕ̃t eiφρ − T T † r0 Op ˜ dϕ̃t χλ exp †



◦ T T δρ ⟩ ,
ρ
g

where Bρ = B(V ),ρ : S RdimM → S R2dimM is the Bargman transform in vertical gauge (C.8)

using local charts on M and defined from the Euclidean metric gρ . We have the metaplectic
    1/2  −◦
operator Op
˜ ρ dϕ̃t = Υ dϕ̃t
ρ ρ P t dϕ̃t
ϕ̃ (ρ) Pρ with the Bergman projector in radial
ρ
(C.33)
†
830 gauge Pρ = B(R)ρ B(R),ρ . We have the phase function φρ defined in (2.43) and expressed in
(C.12)
(2.44) for the Euclidean metric, that is to pass from vertical gauge to radial gauge.
1. In the term R1 , the operator T is constructed from the metric g whereas the operator Bg
is constructed from the local Euclidean metric gρ in Bρ . We use [18, Lemma 4.12] that
the metric g varies slowly on T ∗ M : for any 0 ≤ γ < 1, there exist N > 0 and C > 0 such
that for any ϱ, ϱ′ ∈ R2(n+1) and v ∈ R2(n+1) ,
( )
∥v∥gϱ′ ∥v∥gϱ D EN
max , ≤ 1 + C ⟨|ρ|⟩−(1−γ)/2 ⟨|ϱ|⟩−γ/2 ∥ϱ′ − ϱ∥gϱ . (2.59)
∥v∥gϱ ∥v∥gϱ′
1
If we take γ = λ + ϵ with ϵ > 0, we get that for ∥ϱ′ − ϱ∥gϱ < ⟨|ϱ|⟩ 2 (λ+ϵ) then

∥v∥gϱ′ 1
− 1 ≤ C ⟨|ρ|⟩− 2 (1−λ−ϵ) ,
∥v∥gϱ
D  E−N 1
hence |R1 | ≤ CN,t distg ρ , ϕ̃ (ρ)
′ t
⟨|ρ|⟩− 2 +λ .

2. For the term R2 , we start from

◦
Bϕ̃t (ρ) etX Bρ† = Bϕ̃t (ρ) ϕt Bρ†
(2.2)
 
835 and in local chart in a small neighborhood of π (ρ) , π ϕ̃t (ρ) ∈ M of size
λ′
   ′

Ot ⟨|ρ|⟩ 2 δ ⊥ (ρ) = Ot ⟨|ρ|⟩−(1−λ )/2 , (2.60)
(2.24)

29
 ◦ ◦
we approximate ϕt by its differential dϕtρ at ρ, replace (ϕt ) by (dϕt )ρ . In the next lines,
(V ) , (R) denote respectively vertical and radial gauge defined in section C.2. In the first
equality, we substract the main oscillatory term e−iθρ (.) . In the last equality we use that
det (dϕt ) = 1. We write
(3.2)

◦ iθϕ̃t (ρ) (.) (V ) t ◦ †(V ) −iθρ (.) (R)
◦

Bϕ̃t (ρ) dϕt †
e−iφϕ̃t (ρ) Bϕ̃t (ρ) dϕt Bρ†(R) eiφρ


ρ
Bρ = e B t
ϕ̃ (ρ)
dϕ B
ρ ρ
e = ρ
(C.6,C.8,2.44)
 
−iφϕ̃t (ρ) ˜
= e Opρ dϕ̃tρ eiφρ
(C.43)
 
In the approximation replacing Bϕ̃t (ρ) etX Bρ† by e−iφϕ̃t (ρ) Op
˜ ρ dϕ̃tρ eiφρ , we have neglected
the non linearity of the map ϕt . From Taylor expansion this gives that
D  E−N  ′
2
|R2 | ≤ CN,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩−(1−λ )/2 ,
(2.60)

smaller than previous terms.

3. For the term R3 , one first operation is to pass to the tangent bundle with exp g◦ and with
840 a truncation χ and past restriction r0 . For this, in the linear setting, Lemma C.36
λ

shows that in case of a linear map A : E → E and Euclidean metric g on E ⊕ E ∗ , a

cutoff χσ gives an error term bounded  by CN σ −N , ∀N . Here σ = ⟨|ρ|⟩λ/2 , so this gives
D  E−N
ON,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩−N , smaller than previous terms. Another operation is
to pass again to the manifold with the metric g as in step 1. We have seen that it gives
D  E−N 1
845 |R3 | ≤ CN,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩− 2 +λ .

We have checked (2.52) that gives (2.49).

3 Micro-local analysis of a contact vector field X on (M, A)

near Σ = RA\ {0}
In this section we pursue the analysis of a smooth non vanishing vector field X on a closed
850 manifold M done in section 2) and furthermore we assume that X is a Reeb vector field,
i.e. there is a smooth contact one form A on M such that for every m ∈ M , the linear space
Ker (A (m)) endowed with the two form (dA) (m) is a linear symplectic space and
A (X) = 1, dA (X, .) = 0. (3.1)
This implies that A is invariant under the flow ϕt generated by X. Indeed the Lie derivative
vanishes: LX A = ιX dA + dιX A = dA (X, .) + d1 = 0. Also dimM = 2d + 1, with 2d =
855 dim (KerA). We will write
1
dm := A ∧ (dA)∧d (3.2)
d!
for the corresponding smooth and non-degenerate volume form on M invariant by the flow ϕt .
Remark 3.1. For the moment we do not assume that X is Anosov, but later in section 4, this
one form A will be determined by KerA = Eu ⊕Es in (4.4). A typical example of contact vector
field X on (M, A) is a general geodesic vector field on the unit cotangent bundle M = (T ∗ N )1
860 of a Riemannian manifold N with A being the Liouville one-form.
In T ∗ M we define
E0∗ := (KerA)⊥ = {ρ ∈ T ∗ M, Kerρ ⊃ KerA} (3.3)
= RA = {ωA (m) , m ∈ M, ω ∈ R} . (3.4)
(3.1)

30
 E0∗ is a rank 1 sub-bundle of T ∗ M over M . We have

T ∗ M = KerX ⊕ E0∗ , (3.5)

where  
KerX = ∗
ρ ∈ T M, ω (ρ) = ρ (X) = X (ρ) = 0 (3.6)
(2.7)

is a rank 2d sub-bundle of T ∗ M over M .

3.1 The symplectization Σ = RA\ {0}

865 We first recall the important following lemma. It shows that Σ := RA\ {0} = E0∗ \ {0} is a
symplectic sub-manifold of T ∗ M , called the symplectization of M [6, Section 11.2][1, Appendix
4.]. We will denote π : T ∗ M → M the projection and π ◦ (dm) the form dm (3.2) pulled back
on T ∗ M .
Lemma 3.2. The set

Σ := RA\ {0} := {ωA (m) ∈ T ∗ M | m ∈ M, ω ∈ R\ {0}} (3.7)

is a smooth symplectic sub-manifold of T ∗ M with dimΣ = 2 (d + 1). The induced

volume form on Σ is given by
1
dϱ := (dθ)∧(d+1) = ω d (dω) ∧ π ◦ (dm) (3.8)
(d + 1)!

where θ is the Liouville form (2.11).

870 Proof. For m ∈ M , ω ∈ R\ {0}, at point ρ = ωA (m) ∈ Σ ⊂ T ∗ M we have

θρ = ω (ρ) · π ◦ A (3.9)
(2.11)

where π ◦ is the pull back map on forms. Hence dθ = d (ω (π ◦ A)) = dω ∧ (π ◦ A) + ω (π ◦ dA)

giving the following volume form on E0∗
1 ∧(d+1) 1 d

◦ ◦ ∧d

dϱ : = (dθ) = (d + 1) ω · dω ∧ (π A) ∧ (π dA)
(d + 1)! (d + 1)!
= ω d (dω) ∧ π ◦ (dm)
(3.2)

which does not vanish on Σ = E0∗ \ {0} since for the line bundle E0∗ → M , dm is a measure
on the base M and dω is a measure on the fibers. Consequently dθ restricted to Σ is non
degenerate, Σ is a smooth sub-manifold of T ∗ M .

3.2 Linearized dynamics on the set Σ

875 Since the one form A is invariant under the flow ϕt , the set Σ is invariant under the lifted flow
ϕ̃t in T ∗ M defined in (2.8). The purpose of this section is to provide a useful approximation
of the pullback operator etX restricted to a micro-local neighborhood of Σ (i.e. vicinity of Σ
in T ∗ M ) in terms of the differential map dϕ̃tN that is dϕ̃t restricted to the symplectic normal
bundle N of Σ.
880 Remark 3.3. Later in section 4 we will consider a contact Anosov flow and see in Lemma 4.2
that Σ is the trapped set or non wandering set for the lifted flow ϕ̃t . Consequently we will
obtain in Theorem 4.6 that the dynamics etX outside Σ is negligible in some sense, so the only
important part is indeed a micro-local neighborhood of Σ that we consider in this section.

31
 3.3 The symplectic normal bundle N → Σ
⊥Ω
885 3.3.1 Metaplectic decomposition K ⊕ N
We have seen in Lemma 3.2 that Σ := E0∗ \ {0} is symplectic. We will denote

TΣ T ∗ M := {Tρ (T ∗ M ) | ρ ∈ Σ}

that is the tangent bundle T (T ∗ M ) restricted to the base space Σ. The following Lemma
provides a decomposition of TΣ T ∗ M into symplectic sub-bundles, invariant under the flow
map dϕ̃t : TΣ T ∗ M → TΣ T ∗ M with ϕ̃t defined in (2.8). For the one form A seen as a map
890 A : M → T ∗ M , we will use its differential map dA : T M → T T ∗ M . Recall that X̃ defined in
section 2.3.4.

Lemma 3.4. Let Ẽ0 := RX̃ and Ẽ0∗ := RA be rank 1 sub-bundles of TΣ T ∗ M . Then

K0 := Ẽ0 ⊕ Ẽ0∗ (3.10)

is a symplectic sub-bundle of TΣ T ∗ M and we have a decomposition of TΣ T ∗ M as an

orthogonal sum of symplectic spaces K, K0 , N :
⊥Ω ⊥Ω
TΣ T ∗ M = K ⊕ K0 ⊕ N (3.11)
| {z }
TΣ

dimK = 2d, dimK0 = 2, dimN = 2d,

invariant under the flow map dϕ̃t : TΣ T ∗ M → TΣ T ∗ M .

⊥Ω
Remark 3.5. Notice that N = (T Σ) ⊥Ω is the Ω-symplectic orthogonal to T Σ = K ⊕ K0 in
TΣ T ∗ M . To explain the notation of Ẽ0∗ , notice that K0 = Ẽ0 ⊕ Ẽ0∗ is Ω-symplectic, hence Ẽ0∗ is
895 isomorphic to the dual of Ẽ0 .
Proof. For fixed ω ∈ R\ {0}, let T (ωA) ⊂ T T ∗ M be the tangent space to the graph of ωA:

T (ωA) := (d (ωA)) (T M ) .

For any ρ = ωA (m) ∈ Σ, m = π (ρ) ∈ M , we have that dπ : T (ωA) (ρ) → T M (m) is an

isomorphism. We decompose
T (ωA) = K ⊕ Ẽ0 (3.12)
with Ẽ0 = RX̃ = T (ωA) ∩ dπ −1 (RX) , dimẼ0 = 1, and K := T (ωA) ∩ dπ −1 (KerA), dimK =
900 2d. We have Ẽ0∗ = RA = T Σ ∩ dπ −1 ({0}), dimẼ0∗ = 1 and

K0 = Ẽ0 ⊕ Ẽ0∗ = T Σ ∩ dπ −1 (RX) . (3.13)

We have T Σ = T (ωA) ⊕ Ẽ0∗ = K ⊕ K0 . We have to prove that K, K0 are symplectic and that
K ⊥Ω K0 . We first show the following Lemma.
Lemma 3.6. At point ωA ∈ Σ, the linear map
1
p := √ (d (ωA)) ◦ (dπ) : TΣ T ∗ M → TΣ T ∗ M (3.14)
ω
is a projector with
Kerp = Ker (dπ) , Imp = T (ωA) . (3.15)
905 We have  ⊥ Ω
(dπ)−1 (KerA) = (RA)⊥Ω = Ẽ0∗ . (3.16)
(def)

32
 Proof. We have π ◦ A = IdM hence (dπ) ◦ (dA) = IdT M . Then p is a projector because
1
p2 = (d (ωA)) (dπ) (d (ωA)) (dπ) = p.
ω | {z }
ωId

We have
Ω/Ker(dπ) = 0. (3.17)
(2.13)

For any U ∈ TωA T ∗ M , we use (2.14) for α = ωA giving

1 1
A ((dπ) (U )) = θ ((d (ωA)) ◦ (dπ) (U )) = Ω (E, (d (ωA)) ◦ (dπ) (U ))
(2.14) ω (2.16) ω
1 1 1
= √ Ω (E, p (U )) = √ Ω (E, (p (U ) − U ) + U ) = √ Ω (E, U ) .
(3.14) ω ω (3.15,3.17) ω

On Σ = RA we have RE = RA, we deduce (3.16).

We have an isomorphism

d (ωA) : T M = KerA ⊕ RX → T (ωA) = K ⊕ Ẽ0

(3.12)

910 from (2.15) applied to α = ωA, we get

(d (ωA))◦ Ω = ω (dA) . (3.18)

In particular
d (ωA) : (KerA, ωdA) → (K, Ω)
is a symplectomorphism for the respective symplectic structures.
 Since
 KerA is dA−symplectic,
this implies that K is Ω−symplectic. This also gives that Ω K, Ẽ0 = ωdA (KerA, RX) = 0.
By definition
 we have Ẽ0 = RA and K = T (ωA) ∩ dπ (KerA). From (3.16) we have that
∗ −1

915 Ω K, Ẽ0∗ = 0. Since K0 = Ẽ0 ⊕ Ẽ0∗ , we have obtained that K ⊥Ω K0 and that K0 is
Ω−symplectic.
⊥Ω
The next Lemma concerns specifically the subspace K ⊕ N in (3.11). We show that this
space is also the sum of canonical Lagrangian spaces H ⊕ V (horizontal and vertical). This is
illustrated on figure 3.1. Recall the frequency function ω : ρ ∈ T ∗ M → ω (ρ) = ρ (X) ∈ R
920 defined in (2.7) and that (KerA, dA) is a linear symplectic space from section 4.1.2.

Definition 3.7. For u ∈ KerA, we define the maps

K (u) := dπ −1 (u) ∩ K, N (u) := dπ −1 (u) ∩ N

1 1
(3.19)



H (u) := K (u) + N (u) , V (u) := K (u) + N (−u) ,
2 2
that define the horizontal and the vertical spaces:

H := H (u) , u ∈ KerA ,

V := V (u) , u ∈ KerA

⊥Ω
that are subspaces of K ⊕ N .

33
 V
K
N̄ (−u) V̄ (u)
K̄(u)
H̄(u) H
N̄ (u) KerA
^
N symplectic for Ω

dπ
KerA
−u 0 u symplectic for dA
π
Figure 3.1: Illustration of Lemma 3.8 about the vector bundle KerA
^ ⊂ TΣ T ∗ M → M . We have
dπ
the linear map KerA → KerA ⊂ T M and KerA is a sum of two symplectic subspaces K ⊕ N
^ ^
and two Lagrangian subspaces H ⊕ V . This structure is preserved by the dynamics.

Lemma 3.8 (Metaplectic decomposition). At any point ωA ∈ Σ with ω ̸= 0, let

^ := dπ −1 (KerA) ∩ Ker (dω) . (3.20)

KerA

^ is a Ω−symplectic subspace of TΣ T ∗ M and we have the decompositions

KerA
⊥
^ = K ⊕Ω N = H ⊕ V,
KerA (3.21)

with Ω−symplectic and orthogonal subspaces K, N and Ω−Lagrangian subspaces H, V

with dimV = dimH = 2d.
We have (dπ) (H) = KerA and (dπ) (V ) = {0}. The linear maps
√
(3.22)


dπ : (K, Ω) → KerA, ωdA
√
(3.23)


dπ : (N, Ω) → KerA, − ωdA
are symplectomorphism for the respective symplectic structures.

Proof. We have
 ⊥ Ω
dπ −1 (KerA) ∩ Ker (dω) = Ẽ0∗


KerA
^ = ∩ Ker (dω)
(3.20) (3.16)
 
∗
= K ⊕ Ẽ0 ⊕ N ∩ Ker (dω) (3.24)
(3.11)

   
We have dω (N ) = Ω X̃, N = Ω Ẽ0 , N = 0. This gives
(2.19) (def) (3.11)

N ∩ Ker (dω) = 0.

and similarly
K ∩ Ker (dω) = 0.
925 We have Ẽ0∗ = RE and
   
dω (E) = Ω X̃, E = θ X̃ = −ω ̸= 0,
(2.19) (2.16) (2.20)

⊥
^ = K ⊕Ω N in (3.21).
hence Ẽ0∗ ∩ Ker (dω) = 0. We deduce the decomposition KerA

34
 On the space K we have
1
Ω/K = p◦ Ω/K = √ (dπ)◦ ◦ (d (ωA))◦ Ω/K
(3.14,3.12) (3.14) ω
√ ◦
= ω (dπ) (dA) .
(3.18)

This gives (3.22). For u ∈ KerA, we have (dπ) (V (u)) = 1

(u − u) = 0, hence V is vertical
(3.19) 2
hence Ω−Lagrangian.
For the map V : KerA → V , we have V ◦ dπV →KerA = Id/V and V is Ω−Lagrangian hence

◦ 1 
◦
◦ 
0 = Ω/V = (dπV →KerA )◦ ◦ V Ω = (dπV →KerA )◦ ◦ K (Ω) + (−Id)◦ ◦ N (Ω)
(3.19) 2

930 hence since K = (dπ)−1 ◦

KerA→K and (−Id) = Id on bilinear forms,

◦
◦ √
N (Ω) = − K (Ω) = − (dπ)−1 KerA→K (Ω) = − ωdA,
(3.22)

giving (3.23) because N = (dπ)−1 KerA→N .

For the map H : KerA → H, we have H ◦ dπH→KerA = Id/H hence

◦ 1 
◦
◦ 
Ω/H = (dπH→KerA )◦ ◦ H Ω = (dπH→KerA )◦ ◦ K (Ω) + N (Ω)
(3.19) 2
1 √ √
(dπH→KerA )◦ ωdA − ωdA = 0.


=
(3.22,3.23) 2

Hence H is Ω−Lagrangian.

3.3.2 Modified metric g̃ on the set Σ

⊥Ω ⊥Ω
In (3.11), the decomposition K ⊕ K0 ⊕ N is orthogonal for the symplectic form Ω. But a priori
935 it is not orthogonal for the metric g given in (2.25). Since this last property would be useful
later, we will modify the metric g according to the next lemma to get g−orthogonality as well.
Lemma 3.9. On the set Σ, we construct a modified metric g̃ from the existing metric g
⊥g̃ ⊥g̃
as follows. Let g̃/N = g/N , g̃/K = g/K , g̃/K0 = g/K0 and assume K ⊕ K0 ⊕ N . Then

1. g̃ is compatible with Ω.

2. The decomposition (3.11) is orthogonal with respect to the metric g̃ (and w.r.t. Ω
as before):
⊥Ω,g̃ ⊥Ω,g̃
TΣ T ∗ M = K ⊕ K0 ⊕ N (3.25)

3. The modified metric g̃ is equivalent to the metric g, uniformly with respect to ρ ∈ Σ

for |ω| ≥ 1.

Proof. By construction we have claim 1,2. In local coordinates (x, z, ξ, ω) used in section 2.4,
1/2 −1/2
and for some ω0 ≫ 1, consider the change of coordinates x′ = ω0 x, z ′ = z, ξ ′ = ω0 ξ, ω ′ = ω
940 so that Ω = dx′ ∧ dξ ′ + dz ′ ∧ dω ′ . Let ρ = ωA (m) ∈ Σ with |ω| ≍ ω0 . From (2.24) we have
−1/2
δ ⊥ (ρ) ≍ |ρ|−1/2 ≍ |ω|−1/2 ≍ ω0 . Then g ≍ dx′2 + dξ ′2 + dz ′2 + dω ′2 , where the equivalence
(2.25)
are uniform w.r.t. |ω| ≥ 1. Using this new system of coordinates we get that the subspaces
K, K0 , N are uniformly apart form each over with respect to ρ ∈ Σ, i.e. for the pair of spaces
(K, N ), we have ∃0 < C < 1, ∀ρ ∈ Σ,∀u ∈ Kρ ,∀v ∈ Nρ ,
|gρ (u, v)| < C ∥u∥gρ ∥v∥gρ , (3.26)
945 and similarly for other pair of space (K, K0 ) , (K0 , N ). This gives Claim 3.

35
 For simplicity of notations, from now on, we assume that the metric g on T ∗ M has
the properties 1,2,3 of Lemma 3.9 on Σ.
  
3.3.3 ˜
Factorization formula for Op t
dϕ̃
/Σ
 
We consider now the operator Op˜ dϕ̃t that appears on the right hand side of (2.48). Since Σ
 
950 in invariant by ϕ̃t and Op
˜ dϕ̃t is fiber-wise, we can consider its restriction to S (TΣ T ∗ M ). It
is denoted and given by
     1/2  −◦
˜
Op dϕ̃ t
= Υ dϕ̃t P dϕ̃t P : S (TΣ T ∗ M ) → S (TΣ T ∗ M )
/Σ (2.47)

The decomposition (3.11) is orthogonal for both the symplectic form and the modified metric
g̃, and preserved by the differential of the flow, i.e.
 
dϕ̃t = dϕ̃t K + dϕ̃t K0 +dϕ̃t N . (3.27)
/Σ | {z }
dϕ̃tT Σ

Remark 3.10. Eq. (3.11) implies that on the set Σ we have a natural identification

S (TΣ T ∗ M ) = S (K) ⊗Σ S (K0 ) ⊗Σ S (N ) . (3.28)

955 where the right hand side is a notation for the space of sections S (Σ; S (K) ⊗ S (K0 ) ⊗ S (N )),
i.e. the tensor product is fiber-wise over Σ and not global.

Theorem 3.11 (Factorization formula). On the set Σ, with respect to the identifica-
tion (3.28) we have
        
˜
Op dϕ̃t ˜ dϕ̃tK ⊗Σ Op
= Op ˜ dϕ̃tK ⊗Σ Op
0
˜ dϕ̃tN . (3.29)
/Σ | {z }
˜ (dϕ̃t )
Op TΣ

Proof. We have to check that we can apply Theorem C.25 to get (3.29). At every point ρ ∈ Σ,
⊥Ω,g̃ ⊥Ω,g̃
we have an orthogonal decomposition TΣ T ∗ M = K ⊕ K0 ⊕ N for the metric g̃ and the
(3.25)
960 symplectic form Ω. This decomposition is preserved by the dynamics dϕ̃t . From invariant
Lagrangian decomposition (3.21), (3.10) and Lemma C.22 we have

P = PK ⊗ PK0 ⊗ PN .

and get (3.29).

3.4 Equivalent family of operators

As we have said, we want to describe only the restriction of the operator etX to a small neigh-
965 borhood of the set Σ ⊂ T ∗ M , i.e. to consider Schwartz kernel of operators in this neighborhood
and not on the whole manifold T ∗ M . For this purpose, we introduce here an equivalence re-
lation between operators denoted ≈ that we will often use later. The equivalence A ≈ B will
mean roughly that the two operators A, B are approximately the same at high frequency and
near the set Σ.

36
 Let σ > 0 that will be chosen large enough later. Define the following characteristic function
χΣ,σ : T ∗ M → R+ as follows. We write ρ = ρ∗ + ρ0 ∈ T ∗ M with transverse component
ρ∗ ∈ KerX and frequency component ρ0 ∈ E0∗ , see (3.5). We set

χΣ,σ (ρ) := 1n∥ρ o, (3.30)

∗ ∥gρ ≤σ

970 and
Op (χΣ,σ ) := T † χΣ,σ T : C ∞ (M ) → C ∞ (M ) (3.31)
be the corresponding P.D.O. operator, as defined in [18, def. 4.28] that extracts components at
distance less than σ from the set Σ.
Recall Ψmϕ̃t
that has been defined in definition 2.11.

Definition 3.12. For two operators A, B : S (M ) → S ′ (M ) we write

A≈B (3.32)

if there exists m < 0, t ∈ R, such that for any σ > 0,

Op (χΣ,σ ) (A − B) Op (χΣ,σ ) ∈ Ψm
ϕ̃t
.

975 Remark 3.13. From m < 0, we recall that consequently we have the estimate (2.37) for R =
Op (χΣ,σ ) (A − B) Op (χΣ,σ ), where the constant Ct depends on σ also.

3.5 Description of the operator etX near the set Σ with the bundle N
We first define some operators. The construction that we pursue below is represented on Figure
1.1. Using the metric g on T ∗ M , and considering the sub-bundle N ⊂ TΣ T ∗ M , we have the
980 exponential map
expN : N → T ∗ M. (3.33)
Due to the slow variation of the metric g at high frequencies, given in (2.59), we have that for
any 0 ≤ µ < 1, and any frequency |ω| ≥ ω0 with ω0 large enough, expN is a diffeomorphism on
the neighborhood ∥v∥g ≤ ⟨ω⟩µ/2 . To express this, let 0 < µ < 1 and ω0 ≥ 1 for this property.
We introduce the following cut-off function in frequency χµΣ : N → [0, 1] defined as follows. For
985 ρ ∈ Σ, v ∈ Nρ and ω = ω (ρ),

χµΣ (v) = 1{∥v∥ µ/2

,|ω|≥ω0 } . (3.34)
g ≤⟨ω⟩

The multiplication operator by χµΣ is also denoted χµΣ : S (N ) → S ′ (N ). Later, in Theorem

4.37, we will need to take µ small enough depending in particular on the Hölder exponents of
the Anosov dynamics.
Similarly to (2.42), we define the twisted pull back and push forward operators

^)◦ : S (T ∗ M ) → S ′ (N ) ,
χµΣ (expN
990
−1 ◦ µ
^ χΣ : S (N ) → S (T ∗ M ) .


expN

Namely, for u ∈ S (T ∗ M ), w ∈ S (N ), let ρ ∈ T ∗ M with ρ = expN (v2 ), v2 ∈ Nρ2 , ρ2 ∈ Σ,

|ω (ρ2 )| ≥ ω0 , and ∥v2 ∥g ≤ ⟨ω (ρ2 )⟩µ/2 ,
 
^)◦ u (v2 ) := eiφ(v2 ) u (ρ) ,
(exp (3.35)
N

 
−1 ◦
expN w (ρ) := e−iφ(v2 ) w (v2 ) ,
^ (3.36)

37
 with phase function φ defined in (2.43), to pass from vertical gauge to radial gauge. In the
995 next definition, we combine these previous operators together with T : C ∞ (M ) → S (T ∗ M ) in
(2.33).

Definition 3.14. We define the operators

^)◦ T
TN := χµΣ (exp : C ∞ (M ) → S ′ (N ) (3.37)
N

−1 ◦ µ
TN∆ := T † exp
^ χΣ : S (N ) → C ∞ (M ) . (3.38)


N

Using the operators (3.37),(3.38), the next theorem gives a good approximation of the
operator etX in a neighborhood of the set Σ in terms of the differential map dϕ̃tN on the normal
1000 bundle N . We will use the notation ≈ defined in (3.32) that compare operators near Σ. We
also use the shortened notation  
Υt := Υ dϕ̃tT Σ , (3.39)

that is the metaplectic correction (C.17) of the differential restricted to T Σ, dϕ̃tT Σ = dϕ̃t K +dϕ̃t K0
1/2
that appears in (3.27). In the next theorem we view Υt > 0 as a multiplication operator over
Σ.
Theorem 3.15. For any t ∈ R, we have
 
1/2 ˜
etX ≈ TN∆ Υt Op dϕ̃tN TN (3.40)
1005

Remark 3.16. Let us compare formula (3.40) with (2.48). Both formulaexpress  the pull back
operator etX from some quantum (metaplectic) operator Op ˜ dϕ̃t or Op ˜ dϕ̃t acting on some
N
vector bundle over T M . The difference is that (2.48) is valid microlocally on the whole T ∗ M
∗

(but at high frequencies) and general to any non vanishing vector fields X, whereas (3.40)
1010 is specific to contact (or Reeb) vector field and valid microlocally near Σ ⊂ T ∗ M (and high
frequencies).

Remark 3.17. As a particular case, taking t = 0 in (3.40) gives that

TN∆ TN ≈ Id. (3.41)

Proof. As we already did in the proof of Theorem 2.16, we use T : C ∞ (M ) → S (T ∗ M ) to lift

1015 the analysis to T ∗ M and consider the Schwartz kernel
   
∆ 1/2 ˜
R̃ := ⟨δρ′ |T e − TN Υt Op dϕ̃N TN T † δρ ⟩ ,
tX t

that we will estimate for points ρ′ , ρ ∈ T ∗ M close to the set Σ as needed by definition 3.12.
For this, let σ > 0. We write ρ = ρ∗ + ρ0 with transverse component ρ∗ ∈ KerX and frequency
component ρ0 ∈ E0∗ , see (3.5), similarly for ρ′ and we assume that

∥ρ∗ ∥gρ ≤ σ, ∥ρ′∗ ∥gρ ≤ σ. (3.42)

We also assume that points (ρ, ρ′ ) ∈ T ∗ M × T ∗ M are “near” the graph of ϕ̃t , i.e. that (2.57)
1020 holds true. Otherwise the proof of Theorem 2.16, (step A), shows that far from the graph, the
Schwartz kernel is totally negligible.
As before, the strategy is to approximate the dynamical operator etX to its linear part,
approximate the metric to an Euclidean metric, and then check that formula holds true exactly
in this linear and Euclidean setting. For this, we use local coordinates in R2(2d+1) on T ∗ M

38
  
1025 and use the operator Opρ dϕ̃tρ and phase φρ that have already been defined in (2.58). We
˜
introduce intermediary operators
 
O1 := T etX T † O2 := e−iφϕ̃t (ρ) Op
˜ ρ dϕ̃tρ eiφρ ,
 
−1 ◦ 1/2 ˜ ^◦
O3 := e−iφϕ̃t (ρ) Pϕ̃t (ρ) exp
^ t iφρ


N Υt Op ρ d ϕ̃N (expN ) Pρ e ,

 
−1 ◦ µ 1/2 ˜ ^)◦ T T †
O4 := T T expN χΣ Υt Op dϕ̃N χµΣ (exp
^ † t


N
 
1/2 ˜
= T TN∆ Υt Op dϕ̃tN TN T † ,
(3.37,3.38)

and their differences

R1 := |⟨δρ′ | (O1 − O2 ) δρ ⟩| , R2 := |⟨δρ′ | (O2 − O3 ) δρ ⟩| , (3.43)

R3 := |⟨δρ′ | (O3 − O4 ) δρ ⟩| .
1030 So we have
R̃ ≤ R1 + R2 + R3 .
For the term R1 , in the proof of Theorem 2.16, (step B), we have already shown that for any
0 < λ < 12 ,
D  E−N 1
R1 ≤ CN,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩− 2 +λ .
For the term R2 , we decompose in O2 the term
        
˜ ρ dϕ̃t
Op = Op ˜ ρ dϕ̃t
˜ ρ dϕ̃t ⊗ Op ⊗ ˜ ρ dϕ̃t .
Op
ρ K N K0
(3.29)

For the transverse part Φ̃ = dϕ̃tK + dϕ̃tN , Lemma C.31 in the appendix gives that (and this is
1035 the main relation that underlies theorem 3.15)
     
  t 1/2
−1 ◦
 
^◦
˜ ρ dϕ̃tK ⊗Op
Op ˜ ρ dϕ̃tN = Op
˜ ρ Φ̃ = P t exp ^ Υ dϕ̃ K
˜
Op ρ d ϕ̃t
N (expN ) Pρ .
ϕ̃ (ρ) N
(C.46)

 
If we add the neutral component Op
˜ ρ dϕ̃t
K0 that has no effect this gives
   
˜ t ^ −1 ◦ 1/2 ˜ t ^)◦ Pρ .


Opρ dϕ̃ρ = Pϕ̃t (ρ) expN Υt Opρ dϕ̃N (expN

This shows that R2 = [Link] for the term R3 , we add the cut-off χµΣ that has no ef-
(3.43)
fect from our assumptions on points ρ, ρ′ , we pass from the Euclidean metric to the met-
ric g repeating arguments with the slow variation of the metric (2.59) and get that R3 ≤
D  E−N 1
1040 CN,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩− 2 +λ for any 0 < λ < 21 , as well.
From these estimates and assumption (3.42) on ρ, ρ′ , we can add the cut-off Op (χΣ,σ ) and
deduce that for any 0 < λ < 21 ,

− 1 +λ
   
tX 1/2 ˜
Op (χΣ,σ ) e − TN∆ Υt Op dϕ̃tN TN Op (χΣ,σ ) ∈ Ψϕ̃t2 .

From definition 3.12, this gives (3.40).

39
 3.6 Description of the operator etX near the set Σ with the bundle
1045 TΣ T ∗ M
In this section we provide another expression for the operator etX similar to (3.40) but sometimes
more convenient, see remark 3.20 below. The little difference with (3.40) is that we will use
the bundle TΣ T ∗ M = T Σ ⊕ N → Σ instead of the sub-bundle N → Σ.
(3.11)
Similarly to (3.34), we introduce the following cut-off function in frequency χµΣ : TΣ T ∗ M →
1050 [0, 1] defined as follows. For v ∈ TΣ T ∗ M , and ω = ω (π (v)),

χµΣ (v) = 1{∥v∥ µ/2

,|ω|≥ω0 } . (3.44)
g ≤⟨ω⟩

The multiplication operator by χµΣ is also denoted χµΣ : S (TΣ T ∗ M ) → S ′ (TΣ T ∗ M ).

We denote r/Σ : S (T T ∗ M ) → S (TΣ T ∗ M ) the restriction operator to the base space Σ ⊂
T ∗ M . We denote
rN : S (TΣ T ∗ M ) → S (N ) (3.45)
the restriction operator to the sub-bundle N ⊂ TΣ T ∗ M .

Definition 3.18. We define the operators

TΣ := χµΣ r/Σ exp : C ∞ (M ) → S ′ (TΣ T ∗ M ) (3.46)

g◦ T

−1 ◦
TΣ∆ := T † exp
^ rN χµΣ : S (TΣ T ∗ M ) → C ∞ (M ) . (3.47)


N
1055

Theorem 3.19. For any t ∈ R, we have

  
e tX
≈ ˜
TΣ∆ Op dϕ̃ t
TΣ (3.48)
/Σ

Remark 3.20. This remark is very similar to the previous remark 2.18. One advantage of (3.48)
1/2
compared to (3.40) is that it does not contain the time dependent metaplectic
 correction
 Υt ,
but contains instead, through the decomposition (3.29), the operator Op ˜ dϕ̃t
TΣ that itself
1/2
1060 contains Υt . Later in the proof of theorem 4.31, we will use some unitary properties of this
operator Op˜ dϕ̃t . On the other hand, one advantage of (3.40) compare to (3.48) is that it
TΣ
deals directly with the normal sub-bundle N on which we aim to work, so it is easier at first
sight.
Proof. The proof follows the same lines as the proof of Theorem 3.15. The only difference is
1065 that instead of using Lemma C.31, we use Lemma C.33 in the appendix for the Euclidean and
linear case.

3.7 Emergence of quantum mechanics near the set Σ for a general

contact vector field
A direct consequence of theorem 3.15 is the following theorem that shows that at any time t the
1070 dynamics of the pull back operator etXF is well approximated microlocally near theset Σ (at
1/2 ˜
distance less than σ) and at high frequencies ω, by a quantum operator TN∆ Υt Op dϕ̃tN TN
(this is a F.I.O., compare with (1.10)).
This result is less precise than theorem 1.4 that gives an approximation of the full operator
e , whereas here we have an approximation for a restriction to a micro-local neighborhood
tXF

1075 of Σ. However this may be interesting because Σ is an invariant set for the dynamics. We

40
 may wonder if with some additional generic assumptions, this result can be improved towards
theorem 1.4.
Theorem 3.21. ∀t ∈ R, ∃m < 0, ∀σ > 0,∃Ct,σ > 0 ∀ω > 0,
   
tXF ∆ 1/2 ˜ t
Op (χΣ,σ ) e − TN Υt Op dϕ̃N TN Op (χΣ,σ ) (Id − Op (χω )) ≤ Ct ω m → 0.
L2 (M ;F ) ω→+∞

Proof. From theorem 3.15, definition 3.12 and Lemma 2.12.

1080 4 Micro-local analysis of a contact Anosov vector field X

near Σ = RA\ {0}
In this section, as a last step, we consider the model of interest for this paper: a smooth contact
Anosov vector field X on a closed manifold M . In other words, we add the hypothesis of Anosov
dynamics to the previous section 3. We first recall the precise definitions and notations.

1085 4.1 Definitions and notations

4.1.1 Anosov vector field
We make the hypothesis that X is a smooth Anosov vector field on a closed manifold M .
This means that for every point m ∈ M the tangent vector X (m) is non zero and we have a
continuous splitting of the tangent space

Tm M = Eu (m) ⊕ Es (m) ⊕ E0 (m) (4.1)

| {z }
RX(m)

1090 that is invariant under the action of the differential of the flow map dϕt : T M → T M and there
exist λmin > 0, C > 0 and a smooth metric gM on M such that

∀t ≥ 0, ∀m ∈ M, dϕ−t
/Eu (m) ≤ Ce−λmin t , dϕt/Es (m) gM
≤ Ce−λmin t . (4.2)
gM

See Figure 4.1. The linear subspace Eu (m) , Es (m) ⊂ Tm M are called the unstable/stable
spaces and the one dimensional space E0 (m) := RX (m) is called the neutral direction or
flow direction. For a general Anosov vector field, the maps m → Eu (m), m → Es (m) and
1095 m → Eu (m) ⊕ Es (m) are only Hölder continuous with (a priori different) Hölder exponents
respectively [29]:
βu , βs , β0 ∈]0, 1], β := min (βu , βs ) . (4.3)
Let A ∈ C 0 (M ; T ∗ M ) be the continuous one form on M called Anosov one form defined
for every m ∈ M by the conditions

A (m) (X (m)) = 1 and Ker (A (m)) = Eu (m) ⊕ Es (m) . (4.4)

From this definition, A is preserved10 by the flow ϕt and the map m → A (m) is Hölder
1100 continuous with exponent β0 .
10
Notice from (4.2), that A is the unique continuous one form preserved by the flow with the normalization
condition A (X) = 1.

41
 Es (m) Es
stable
Vector field E0 (m) Flow
X
ϕt (m)
m X E0

Eu (m) Eu M
unstable

Figure 4.1: Anosov flow ϕt (in solid line) generated by a vector field X on a compact manifold
M.

4.1.2 Contact Anosov vector field

In addition to section 4.1.1 we assume in this paper that the Anosov vector field X is contact,
i.e. that the Anosov one form A defined in (4.4), is a smooth contact one form on M . This
means that the distribution Eu (m) ⊕ Es (m) is smooth (i.e. C ∞ ) w.r.t. m ∈ M (hence with
1105 Hölder exponent β0 = 1) and that the linear space Ker (A (m)) = Eu (m) ⊕ Es (m) endowed
with the two form dA (m) is a linear symplectic space for every m ∈ M .
Since A

is invariant by the flow, for any u1 , u2 ∈ Es (m) we have ∀t ∈ R, (dA) (u1 , u2 ) =
∗
d (ϕt ) A (u1 , u2 ) = (dA) (dϕt (u1 ) , dϕt (u2 )) → 0 hence dA/Es = 0 meaning that Es (m)
t→+∞
is dA−isotropic. Similarly Eu (m) is dA−isotropic hence Eu (m) , Es (m) are dA−Lagrangian
1110 and
d := dimEu (m) = dimEs (m) , (4.5)
hence dimM = 2d + 1 is odd.
Remark 4.1. The vector field X is the Reeb vector field of A, i.e. X is determined from A
by A (X) = 1 and11
(dA) (X, .) = 0, (4.6)
as the setting of section 3.

1115 4.1.3 Dual decomposition

The dual bundle decomposition of (4.1) is

T ∗ M = Eu∗ ⊕ Es∗ ⊕ E0∗ (4.7)

with

Eu∗ := (Es ⊕ E0 )⊥ = {ρ ∈ T ∗ M, Kerρ ⊃ Es ⊕ E0 } , (4.8)

⊥
Es∗ := (Eu ⊕ E0 ) = {ρ ∈ T M, ∗
Kerρ ⊃ Eu ⊕ E0 } , (4.9)

E0∗ := (Eu ⊕ Es )⊥ = {ρ ∈ T ∗ M, Kerρ ⊃ Eu ⊕ Es } (4.10)

= RA = {ωA (m) , m ∈ M, ω ∈ R} . (4.11)
(4.4)

Hence
dimEu∗ (m) = dimEs∗ (m) = d, dimE0∗ (m) = 1.
(4.5)

11
From Cartan formula and invariance of A: 0 = LX A = ιX dA + dιX A and ιX A = 1 give (dA) (X, .) = 0.

42
 From (4.3), the map m ∈ M → Eu∗ (m) is βs -Hölder continuous and m ∈ M → Es∗ (m) is
βu -Hölder continuous. However the maps m ∈ M → E0∗ (m) = RA (m) and m ∈ M →
1120 (Eu∗ ⊕ Es∗ ) (m) = E0⊥ (m) = Ker (X) (m) are smooth and have already been defined in (3.3)
and (3.6).
A cotangent vector ρ ∈ T ∗ M is decomposed accordingly to the dual decomposition (4.7) as

ρ = ρu + ρs + ρ0 (4.12)

with components
ρu ∈ Eu∗ , ρs ∈ Es∗ , ρ0 = ωA (m) ∈ E0∗
with m = π (ρ) and the frequency ω = ω (ρ) ∈ R.
(2.7)
By duality, the hyperbolicity assumption (4.2) gives that the components (4.12) of ρ (t) =
ϕ̃ (ρ) satisfy
t

∃C > 0, ∀ρ (0) ∈ T ∗ M, ∀t ≥ 0,
1
∥ρu (t)∥gM ≥ eλmin t ∥ρu (0)∥gM , ∥ρs (t)∥gM ≤ Ce−λmin t ∥ρs (0)∥gM , ω (t) = ω (0) .
C
(4.13)

1125 See figure 4.2.

4.1.4 Trapped set

From (4.13) we deduce the following Lemma.

Lemma 4.2. The set E0∗ = RA is the trapped set (or non wandering set) of the flow
(4.10)
t
ϕ̃ in the sense that
n o
E0∗ = RA = Σ ∪ {0} = ρ ∈ T ∗ M | ∃K ⊂ T ∗ M compact, s.t. ϕ̃t (ρ) ∈ K, ∀t ∈ R ,
(4.10) (3.7)
(4.14)

and E0∗ is transversely hyperbolic.

Proof. From (4.12), the trapped set E0∗ is characterized by ρu = 0, ρs = 0 and from (4.13), E0∗
1130 is transversely hyperbolic.

4.2 Anisotropic Sobolev space, decay of norm outside the trapped set
and discrete Ruelle spectrum
In Lemma 4.2 we have observed that the trapped set E0∗ is transversely hyperbolic. This will
permit us to define a weight function or escape function W : T ∗ M → R+ that decays outside
1135 the trapped set (i.e. that is a Lyapounov function for the flow ϕ̃t ), following the constructions
in [14] or [18, Def 5.6]. Then we define the anisotropic Sobolev space HW (M ) and obtain that
the norm of the pull back operator etX restricted to the outside of E0∗ decays exponentially fast
with an arbitrarily large rate. This is the important property that will enable us to restrict
the analysis of the operator etX for t > 0, to a neighborhood of the trapped set E0∗ in the
1140 forthcoming sections. In this section we recall all these results.

43
 ρ
Trapped set
E0∗ (m) ϕ̃t
ωA(m)
A(m)

0 ρ(t) 0
X̃
Eu∗ (m)

Σω (m)
Es∗ (m)

∗
Tm M Tϕ∗t (m) M

ϕt
X ϕt (m)
u
m
Eu
exp(tX)u
M Es
 
Figure 4.2: The Anosov flow ϕt = exp (tX) on M induces a Hamiltonian flow ϕ̃t = exp tX̃ in
cotangent space T ∗ M . The magenta lines on M represent “internal oscillations” of a function
u, that is a wave packet and is supposed here to have a small support near ϕt (m). These
oscillations correspond to a cotangent vector ρ ∈ Tϕ∗t (m) M . Transported by the flow the os-
cillations of etX u = u ◦ ϕt increase and the wave front of these oscillations become parallel to
Es ⊕ E0 equivalently ρ (t) = ϕ̃t (ρ) converges to the direction of Eu∗ ⊂ T ∗ M and remains in the
frequency level Σω := ω −1 (ω). The trapped set of the lifted flow ϕ̃t is the rank one vector
bundle E0∗ = RA (green line) where A is the Anosov one form. From Lemma 3.2, Σ := E0∗ \ {0}
is a symplectic sub-manifold of T ∗ M , with dimΣ = 2 (d + 1).

44
 4.2.1 Definition and properties of the escape function W
Following [18, Def 5.6], we consider the positive valued continuous function W ∈ C (T ∗ M ; R+ \ {0})
called weight function defined as follows. Let R ∈ R, γ ∈ [1 − β, 1[ with β given in (4.3). For
ρ ∈ T ∗ M , with stable/unstable components ρs ∈ Es∗ , ρu ∈ Eu∗ given in (4.12), we define
D ER
hγ (ρ) ∥ρs ∥gρ
W (ρ) := D ER (4.15)
hγ (ρ) ∥ρu ∥gρ
1145 with  1/2 −γ
hγ (ρ) := ∥ρu ∥2gρ + ∥ρs ∥2gρ (4.16)

Remark 4.3. In the previous expressions, the vertical vectors ρu , ρs ∈ Tm∗ M are naturally iden-
tified with vectors in the tangent space Tρ (T ∗ M ) of T ∗ M at point ρ ∈ T ∗ M , so we can get
their norm, ∥ρu ∥gρ , ∥ρs ∥gρ .
The function W : ρ ∈ T ∗ M → W (ρ) ∈ R in (4.15) is positive and Hölder continuous, since
1150 the decomposition (4.12) is Hölder continuous. W has a few important properties given in [18,
Thm 5.9] as decay property and slow variation properties.

4.2.2 Anisotropic spaces HW (M ; F )

We will define the anisotropic Sobolev spaces HW (M ; F ) used in this paper, in order to get the-
orem 4.8 below, saying that the operator XF has intrinsic discrete spectrum in HW (M ; F ).
1155 Remark 4.4. Let us remark why to get discrete spectrum, we cannot consider the Hilbert
space L2 (M ; F ). If we choose an arbitrary smooth Hermitian metric on the vector bundle
F → M and using the measure dm on M , (3.2), we can define the Hilbert space L2 (M ; F )
of sections of F . One can show that the operator XF , Eq.(2.3), acting in L2 (M ; F ) has only
essential spectrum contained in some vertical band (the detailed argument will be explained in
1160 a forthcoming paper):

spec XF : L2 (M ; F ) → L2 (M ; F ) = z ∈ C, Re (z) ∈ γL−2 , γL+2 ,

  

with
1/t
γL±2 = lim log etXF L∞
.
t→±∞

For example, considering the vector field X itself and the trivial bundle F = M × C, since
divdm X = 0, we have that X = −X † is skew symmetric in L2 (M ) and X has continuous
spectrum on the imaginary axis iR.
1165 For simplicity of notations, we will forget the vector bundle F , that plays no role for this
construction, i.e. we first consider only a trivial bundle F = M × C. For any u ∈ C ∞ (M ) we
define its HW (M )-norm
∥u∥HW (M ) := ∥W T u∥L2 (T ∗ M ) (4.17)
where W is used as a multiplication operator on S (T ∗ M ). Let HW (M ) be the Hilbert space
obtained by completion of C ∞ (M ) with respect to this norm and called anisotropic Sobolev
1170 space:
n o
HW (M ) := u ∈ C ∞ (M ) , ∥u∥HW (M ) < ∞ . (4.18)

In other words W T : HW (M ) → L2 (T ∗ M ) is an isometry. To define HW (M ; F ) we pro-

ceed similarly by completion of the space of smooth sections C ∞ (M ; F ). For more details or
information, we refer to [18, def 4.37].

45
 Remark 4.5. The function W in (4.15) satisfies
D E−|r| D E|r|
∃C > 0, ∀ρ ∈ T ∗ M, C −1 |ρ|gM ≤ W (ρ) ≤ C |ρ|gM . (4.19)

1175 with
R
r= (1 − γ) > 0. (4.20)
2
Consequently
H r (M ) ⊂ HW (M ) ⊂ H −r (M ) . (4.21)
D E−r 
where H (M ) := Op
r
|ξ|gM (L2 (M )) is the standard Sobolev space of order r.

4.2.3 Decay of norm outside the trapped set

Recall the truncation operator Op (χΣ,σ ) at distance σ from the set Σ defined in (3.31).

Theorem 4.6 (Decay outside the trapped set). [18, Thm 5.13]For any Λ > 0, we
can choose R ≫ 1 large enough as a parameter of the function W in (4.15), such that
∃C > 0, ∀t ≥ 0, ∃σt , ∀σ > σt we have

etX (Id − Op (χΣ,σ )) HW (M ;F )

≤ Ce−Λt , (4.22)

(Id − Op (χΣ,σ )) etX HW (M ;F )

≤ Ce−Λt . (4.23)
1180

Remark 4.7. Consider the decomposition etX = etX (Id − Op (χΣ,σ )) + etX Op (χΣ,σ ). Theorem

4.6 means that the first component etX (Id − Op (χΣ,σ )) decays exponentially fast as O e−Λt
with an arbitrarily large rate Λ ≫ 1. We deduce that in order to describe the dominant effect of
the operator etX and Ruelle eigenvalues on the spectral domain Re (z) > −Λ, one has to study
1185 the second component etX Op (χΣ,σ ) namely the dynamics etX restricted to a neighborhood of
 set Σ = E0 \ {0}. Then, from Theorem 3.15, we only need to study the operator
the trapped ∗

˜ dϕ̃t acting on the normal bundle over the trapped set Σ. This is the subject of the
Op N
section 4.3 below.

4.2.4 Discrete Ruelle spectrum in anisotropic Sobolev spaces HW (M ; F )

1190 In paper [18, thm 5.14], Theorem 4.6 is used to prove the following theorem. Recall that the
function W in (4.15) depends on the exponent R and that the order r = R2 (1 − γ) is defined
from R in (4.20).

Theorem 4.8. “Group property and discrete spectrum”[18, thm 2.11]. For a gen-
eral Anosov vector field X on M , the family of operators

etXF : HW (M ; F ) → HW (M ; F ) , t ∈ R,

form a strongly continuous group and the generator XF has discrete spectrum
denoted spect (XF ), on C − rλmin ≤ Re (z) ≤ C ′ with C, C ′ independent of the parameter
R. This spectrum is intrinsic (i.e. does not depend on W ) and called future Ruelle
spectrum in [18, Thm 2.11].

See Figure 1.2.

46
1195 4.3 Description of the operator etX near the set Σ with the bundle Ns
4.3.1 Decomposition N = Nu ⊕ Ns
From the Anosov property, we have the decomposition KerA = Eu ⊕Es into Hölder continuous
(4.4)
sub-bundles, that gives a refined decomposition of (3.11) into isotropic subspaces:

Lemma 4.9. Let

Nu := N ∩ dπ −1 (Eu ) , Ns := N ∩ dπ −1 (Es ) ,

N = Nu ⊕ Ns , (4.24)
Ku := K ∩ dπ −1 (Eu ) , Ks := K ∩ dπ −1 (Es ) , K = Ku ⊕ Ks , (4.25)
dimKu = dimKs = dimNu = dimNs = d, dimK0 = 2,
Each component Ku , Ks , Ẽ0 , Ẽ0∗ , Nu , Ns is a Ω−isotropic vector space, invariant under the
dynamics dϕ̃t . The indices u, s, 0 denotes respectively instability, stability or neutrality.
See Figure 1.1.

1200 Remark 4.10.

 
• Notice that due to (2.10) that reverses the direction of dynamics, Nu is stable for dϕ̃ t
t≥0
and Ns is unstable.

• The symplectic normal bundle N → Σ is smooth, but the isotropic sub-bundle Ns → Σ

is only βs -Hölder continuous.

1205 Proof. From the fact explained in section 4.1.2 that Es , Eu are dA−isotropic and that dπ is a
symplectomorphism in (3.22) and (3.23), we deduce that Nu , Ns , Ku , Ks are Ω−isotropic.

4.3.2 Transfer operators on Ns

We start from theorem 3.15. We have the splitting N = Ns ⊕ Nu in (4.24) of the  symplectic

bundle N over Σ into Lagrangian sub-bundles. We consider the operator Op dϕ̃N on the
˜ t

1210 right hand side of (3.40) and our purpose is to express it as an operator acting only on the
Lagrangian sub-bundle Ns . The bundles Nu , Ns are invariant under the dynamics dϕ̃t , hence
the flow map splits accordingly (recall remark 4.10)

dϕ̃tN = dϕ̃tNu + dϕ̃tNs : N → N.

Notation 4.11. The bundle Ns → Σ defined in (4.25) is Hölder continuous with exponent β.
For simplicity we denote

S (Ns ) := C β (Σ; S (Ns (.))) (4.26)

as the set of functions f ∈ C β (Σ; S (Ns (.))), β-Hölder continuous on Σ with fast decay in
frequency ω, valued in the Schwartz space of functions S (Ns (ρ)) for each ρ ∈ Σ. This notation
1215 will also concerns later other space of sections over Σ. Notice that S (Ns (.)) → Σ is an infinite
rank bundle vector over Σ.

4.3.3 The bundle F (Ns ) → Σ

Let us denote |detNs |−1/2 the dual of the half density bundle in Ns [26, p.509].

47
 Definition 4.12. Let
F (Ns ) := |detNs |−1/2 ⊗ S (Ns ) (4.27)
and
XF : C β (Σ; F (Ns )) → C β (Σ; F (Ns )) (4.28)
be the derivation of sections of F over the vector field X̃, that is generator of the group
of operators, ∀t ∈ R,
  −1/2  −◦
etXF := det dϕ̃tNs dϕ̃tNs : C β (Σ; F (Ns )) → C β (Σ; F (Ns )) . (4.29)

1220 Remark 4.13. The derivation XF over X̃ in (4.28) is well defined since the bundle F (Ns ) is
constructed from the bundle Ns and the given vector field X on M induces a natural derivation
on Ns hence XF on  s ). Recall from remark 4.10 that the linear map dϕ̃Ns is expanding
 F (N
t

on Ns , hence det dϕ̃tNs > 1. The operator XF has a domain that includes sections smooth
along the flow direction. Recall that for simplicity of notation we have ignored the bundle F
1225 in (4.27).  
Now we will relate the operator Op
˜ dϕ̃t on the right hand side of (3.40) to the operator
N
etXF defined in (4.29). In (C.8) we define the Bargman transform

BNs : S (Ns ) → S (Ns ⊕ Ns∗ ) .

As explained in Lemma C.16, we have an (Ω, g) −isomorphism

Ψ : N = Ns ⊕ Ns⊥g → Ns ⊕ Ns∗ .
⊥
But Nu ̸= Ns g in general, so we introduce SNs ,Nu : N → N the linear (shear) symplectic map
⊥
1230 defined as identity on Ns and any vector of Ns g is mapped to its projection on Nu parallel to
Ns :
SNs ,Nu (Ns ) = Ns , SNs ,Nu Ns⊥g = Nu .

Notice that the bundle of maps SNs ,Nu , Ψ is uniformly bounded over Σ. The operator Op
˜ (SNs ,Nu Ψ−1 ) :
S (Ns ⊕ Ns∗ ) → S (N ) is defined as in (C.33) and we set

˜ SNs ,Nu Ψ−1 BNs (4.30)

BNs ,Nu := Op : S (Ns ) → S (N ) .

Lemma 4.14. We have

 
˜ dϕ̃tN = BNs ,Nu etXF B †
Op (4.31)
Ns ,Nu .

   ∗−1
Proof. Let ΦNs := dϕ̃tNs ⊕ dϕ̃tNs : Ns ⊕ Ns∗ → Ns ⊕ Ns∗ . From Lemma C.30 with the
 −1
setting ϕ = dϕ̃Ns
t
, we have
  −1/2  −◦
˜ (ΦNs ) = †
Op det dϕ̃tNs BNs dϕ̃tNs BN s
(C.43)
†
= BNs etXF BN s
: S (Ns ⊕ Ns∗ ) → S (Ns ⊕ Ns∗ ) (4.32)
(4.29)

1235 By construction, we have the conjugation

−1
dϕ̃tN = SNs ,Nu Ψ−1 ◦ ΦNs ◦ SNs ,Nu Ψ−1


:N →N

48
 that gives
 
˜ dϕ̃t ˜ SNs ,Nu Ψ−1 †
˜ SNs ,Nu Ψ−1 BNs etXF B † Op



Op N = Op Ns
(C.40,4.32)
†
= BNs ,Nu etXF BN s ,Nu
. (4.33)
(4.30)

We finish this section with the description of the operator etX near the set Σ using the
bundle Ns .
Definition 4.15. We define the operators

†
TNs := BN †
T = BN ^)◦ T
χµ (exp : C ∞ (M ) → C β (Σ; F (Ns )) (4.34)
s ,Nu N s ,Nu Σ N
(3.37)

−1 ◦ µ
TN∆s := TN∆ BNs ,Nu = T † exp
^ : C β (Σ; F (Ns )) → C ∞ (M ) . (4.35)


N χΣ BNs ,Nu
(3.38)

Theorem 4.16. For any fixed t ∈ R, we have

1/2
etX ≈ TN∆s Υt etXF TNs (4.36)
1240
1/2
Recall that the metaplectic correction Υt > 0 defined in (3.39) is a function on Σ consid-
ered as a multiplication operator.
Proof. We write
 
1/2 ˜
etX ≈ TN∆ Υt Op dϕ̃tN TN (4.37)
(3.40)
1/2
= TN∆s Υt etXF TNs .
(4.31,4.34,4.35)

4.4 The bundle map etXF and sub-bundle Fk (Ns )

1245 The expression (4.36) will be convenient to get some properties of the dynamical operator etX
from the study of properties of the operator etXF . For example, one important property
 that −◦
we will first exploit is that dϕ̃tNs : Ns → Ns is a linear map hence the operators dϕ̃tNs and
etXF in (4.29) preserve the finite rank sub-bundle of homogeneous polynomials of degree k ∈ N.
This property is responsible for the band structure of the spectrum.

1250 4.4.1 Taylor projectors Tk

 −◦
Since dϕ̃tNs : Ns → Ns is a vector bundle homomorphism, the push-forward operator dϕ̃tNs
in (4.29) keeps invariant the space of homogeneous polynomial functions of order k ∈ N denoted
Polk (Ns ) ⊂ S ′ (Ns ). To express this property, we first consider the vector bundle Es → M and
the bundle map of finite rank “Taylor projectors” over Id, from definition C.34, denoted

Tk : S (Es ) → Polk (Es ) ⊂ S ′ (Es ) (4.38)

1255 and
K
X
T[0,K] := Tk , T≥(K+1) = Id − T[0,K] , (4.39)
k=0

49
 satisfying
Tk Tk′ = δk=k′ Tk . (4.40)
Using the projection map dπ : Ns → Es in (3.23), we lift these operators to the vector bundle
S (Ns ) → Σ:
T̃k : S (Ns ) → Polk (Ns ) ⊂ S ′ (Ns ) . (4.41)
We define T̃[0,K] and T̃≥(K+1) similarly to (4.39). We have (in the next equation and after, a
1260 strict notation should be Id|detNs |−1/2 ⊗ T̃k instead of T̃k )
h i
tXF
e , T̃k = 0. (4.42)

Definition 4.17. We denote

Fk (Ns ) := |detNs |−1/2 ⊗ Polk (Ns ) (4.43)

that is a finite rank bundle over Σ. The space of sections C β (Σ; Fk (Ns )) is the image of
the bundle map T̃k : C β (Σ; F (Ns )) → C β (Σ; Fk (Ns )).
We denote XFk the generator XF in (4.28) restricted to the invariant space of sections
C β (Σ; Fk (Ns )).

More generally, for 0 ≤ k1 ≤ k2 , we denote the direct sum bundle

k2
M
F[k1 ,k2 ] (Ns ) := Fk (Ns ) , (4.44)
k=k1

and denote XF[k1 ,k2 ] the generator in C β Σ; F[k1 ,k2 ] (Ns ) .

4.4.2 Spectrum of the derivation XFk in L2 (Σ; Fk (Ns ))

The operators etXFk t∈R in (4.29) form a group of point-wise linear bundle maps over ϕ̃t :


1265

Σ → Σ that preserves the volume. Consequently all Lp norms etXFk Lp (Σ;F ) for p ∈ [1, ∞]
k
are equal, as it is for general multiplication operators.

Definition 4.18. For any k ∈ N, let

1/t
γk± := lim log etXFk L2 (Σ;Fk (Ns ))
(4.45)
t→±∞

Remark 4.19. The values γk± defined in (4.45) with L2 (Σ; Fk (Ns )) coincide with the definition
1270 (1.6) using L∞ (M ; Fk (Es )).

Lemma 4.20. We have the estimates ∀ϵ > 0, ∃Cϵ > 0, ∀t ≥ 0,

+ −
etXFk L2 (Σ;Fk (Ns ))
≤ Cϵ et(γk +ϵ) , e−tXFk L2 (Σ;Fk (Ns ))
≤ Cϵ e−t(γk −ϵ) . (4.46)

Equivalently ∀ϵ > 0, ∃Cϵ > 0,∀z ∈ C, Re (z) > γk+ + ϵ or Re (z) < γk− − ϵ,

(z − XFk )−1 L2 (Σ;Fk (Ns ))

≤ Cϵ . (4.47)

Consequently the spectrum of the operator XFk in L2 (Σ; Fk (Ns )) is contained in the
vertical band
Bk := Re (z) ∈ γk− , γk+ . (4.48)
  

50
 Remark 4.21. We will explain in a forthcoming paper that the spectrum of XFk is made by
essential spectrum in this band Bk .
Proof. Eq.(4.45) implies (4.46), (4.47) and (4.48) by writing the convergent expression for
1275 Re (z) > γk+ + ϵ: Z ∞
−1
(z − XFk ) = et(XFk −z) dt
0

that gives, if ϵ < ϵ,

′

Z ∞ Z ∞
+
e−Re(z)t Cϵ′ et(γk +ϵ ) dt ≤ Cϵ .
−1 ′
−Re(z)t tXFk
(z − XFk ) ≤ e e dt ≤
0 0

et(XFk −z) dt, and get also (4.47).

R0
Similarly for Re (z) < γk− −ϵ, we have (z − XFk )−1 = − −∞

4.4.3 Sobolev space HW (Ns ) and remainder term

The operator etXF acts in the infinite rank bundle F (Ns ) in (4.27). We explain here how to
“extract” the finite codimension part etXF T̃≥(K+1) . We define a weight function W : N → R+
1280 similar to W : T ∗ M → R+ in (4.15), with the same parameters 0 < γ < 1, R > 0, as follows.
For ρ ∈ Σ, v = vu + vs ∈ N (ρ) = Nu (ρ) ⊕ Ns (ρ) with vu ∈ Nu (ρ), vs ∈ Ns (ρ), let
D ER
hγ (v) ∥vs ∥gρ
W (v) := D ER (4.49)
hγ (v) ∥vu ∥gρ

with D E−γ
hγ (v) = ∥v∥gρ .

The metric g on T ∗ M is moderate [18, Lemma 4.12]. This implies that under the exponential
map expN : N → T ∗ M defined in (3.33), the functions W and W are equivalent in the
neighborhood of the trapped set defined in (3.34), i.e. ∃C ≥ 1, ∃ω0 > 0, ∀v ∈ N with ∥v∥g ≤
⟨ω⟩µ/2 , ω = ω (π (v)), |ω| ≥ ω0 , we have
1
W (v) ≤ W (expN (v)) ≤ CW (v) . (4.50)
C
We define the Sobolev norm of u ∈ S (Ns ) = C β (Σ; S (Ns )) by
(4.26)

∥u∥HW (Ns ) := ∥WBNs ,Nu u∥L2 (N ) ,

and the anisotropic Sobolev space bundle by completion

n o
HW (Ns ) := u ∈ S (Ns ) , ∥u∥HW (Ns ) < ∞ . (4.51)

Lemma 4.22. For any ϵ > 0, K ∈ N, we can choose R in (4.49) large enough so that
∃CK,ϵ > 0, ∀t ≥ 0,
+
etXF T̃≥(K+1) χ|ω|≥ω0 ≤ CK,ϵ e(γK+1 +ϵ)t . (4.52)
HW (Ns )

51
1285 Proof. Let K ∈ N, ϵ > 0 and t ≥ 0. Recall that etXF is a bundle map over ϕ̃t : Σ → Σ. Let
ρ ∈ Σ with ω (ρ) ≥ ω0 and m = π (ρ) ∈ M . We first observe that the operator
  
etXF T̃≥(K+1) : S (Ns (ρ)) → S Ns ϕ̃t (ρ)

is isomorphic to
etXF T̃≥(K+1) : S (Es (m)) → S Es ϕt (m) .

with m ∈ M on a compact space providing uniformity over Σ. We apply Proposition D.3 in

the appendix giving that ∀ϵ > 0,∃CK,ϵ > 0, ∀t ≥ 0, ∀u ∈ S (Ns ),
2
Z   2
tXF
e T̃≥(K+1) χ|ω|≥ω0 u = etXF T̃≥(K+1) χ|ω|≥ω0 u (ρ) dρ
HW (Ns ) Σ HW (Ns )
 +
2 Z
≤ CK,ϵ e(γK+1 +ϵ)t

2
χ|ω|≥ω0 u (ρ) H (Ns ) dρ
W
(D.9) Σ
 +
2
≤ CK,ϵ e(γK+1 +ϵ)t ∥u∥2HW (Ns ) ,

giving (4.52).

Proposition 4.23. For any K ∈ N, we can choose R in (4.49) large enough so that for
every k ∈ [0, K], there exists Ck such that

T̃k ≤ Ck . (4.53)
HW (Ns )

1290 Proof. As in the proof of Lemma 4.22, we use the fact that we have a bundle map and at each
point of the base, we use Lemma D.6 that is uniform over Σ.

4.5 Symbols and F.I.O. on Σ

Recall Fk (Es ) = |detEs |−1/2 ⊗ Polk (Es ) that is a C β continuous bundle over M and more
(1.5)
generally F[k1 ,k2 ] (Es ) in (4.44).

Definition 4.24. A symbol is a function on M operator valued in L F[0,K] (Es ) (or

L (F (Es ))):
a ∈ C β M ; L F[0,K] (Es ) .

Using the projection map dπ : Ns → Es in (3.23) that is invertible, we obtain a lifted

function denoted
ã ∈ C β Σ; L F[0,K] (Ns ) ,

considered below as a multiplicative operator.

1295

Remark 4.25. Here is precisely how we get ã from a: for ρ ∈ Σ, m = π (ρ) ∈ M , one has
dπρ◦ : S (Es (m)) → S (Ns (ρ)) and set

◦
ã : ρ ∈ Σ → ã (ρ) := dπρ◦ a dπρ−1 .

Example 4.26. Later we will use definition (4.55) only for symbols of the form

a = Tk (giving ã = T̃k ), or a = T[0,K] or a ∈ End F[0,K] or a = T≥(K+1) or a = IdF . (4.54)

52
 Recall the operators TNs : C ∞ (M ) → C β (Σ; F (Ns )) and TN∆s : C β (Σ; F (Ns )) → C ∞ (M )
1300 defined in (4.34) and (4.35).

Definition 4.27 (F.I.O., quantum operator). For a symbol a ∈ C β (M ; L (F (Es )))

and t ∈ R, the quantization of the symplectic bundle map etXF ã : C β (Σ; F (Ns )) →
C β (Σ; F (Ns )) is the F.I.O. operator
1/2
OpΣ etXF ã := TN∆s Υt etXF ãTNs : C ∞ (M ) → C ∞ (M ) (4.55)

Remark 4.28. The term F.I.O. stands for Fourier integral operator, that is the usual name for
this kind of operator.
Using the new notation (4.55), previous Theorem 4.16 can be rephrased as follows.

Theorem 4.29 (Approximation of the dynamics by a quantum operator). For

any t ∈ R,

etX ≈ OpΣ etXF . (4.56)

1305

4.5.1 Continuity theorem, boundness estimates with respect to time

Remark 4.30. In the rest of this section we take any K ∈ N and fix it. Then we choose a
weight function W with


exponent R such that (4.52) holds true and consider symbols a ∈
C M ; L F[0,K] (Es ) .
β

Theorem 4.31 (continuity theorem

for F.I.O.). There exists C > 0, such that for any
symbol a ∈ C β M ; L F[0,K] (Es ) and any t ∈ R,

OpΣ etXF ã H (M ) ≤ C etXF ã H (F ) . (4.57)

W W
1310

Proof. We first give an equivalent expression for the operator as we already saw in Theorem
3.19. This expression is more complicated in appearance but more useful for the proof to work
uniformly w.r.t. time t.
Lemma 4.32 (Equivalent expression). We have
    
1/2
†
˜ dϕ̃t
OpΣ etXF ã = TN∆s Υt etXF ãTNs ≈ TΣ∆ Op tXF
TΣ (4.58)


TΣ ⊗ BNs ,N u e ã BNs ,Nu
(4.55)

Proof. We have already obtained

  
1/2 ˜
TN∆s Υt etXF TNs ≈ TΣ∆ Op TΣ t
dϕ̃
(4.36,3.48) /Σ
    
= TΣ∆ Op˜ dϕ̃t ⊗Σ
˜
Op dϕ̃t
TΣ
TΣ N
(3.29)
    
∆ ˜ t tXF †
= TΣ Op dϕ̃T Σ ⊗ BNs ,Nu e BNs ,Nu TΣ .
(4.33)

1315 By inserting the bundle map ã we obtain (4.58).

We have defined HW (Ns ) in (4.51). Similarly, we can define HW (N ) := L2 (N ; W 2 ) with
the weight W on N in (4.49) and define HW (TΣ T ∗ M ) by the weight W extended trivially to
TΣ T ∗ M = T Σ ⊕ N by W (u, v) = W (v) for any u ∈ Tρ Σ, v ∈ Nρ .
(3.11)

53
 Lemma 4.33. ∃C > 0,

∥TΣ ∥HW (M )→HW (TΣ T ∗ M ) ≤ C, TΣ∆ HW (TΣ T ∗ M )→HW (M )

≤ C. (4.59)

Proof. Recall the expression TΣ = χµΣ r/Σ exp

g◦ T where, from (4.50), the operator χµΣ r/Σ exp



1320 g◦
(3.46)
satisfies ∃C > 0,
χµΣ r/Σ exp


g◦ ≤ C.
L2 (T ∗ M ;W )→HW (TΣ T ∗ M )

Since T : HW (M ) → L2 (T ∗ M ; W ) is an isometry, we deduce the first bound in (4.59). Similarly

−1 ◦
from the expression T ∆ = T † exp^ r χµ , we deduce the second claim.


Σ N N Σ
(3.47)

Lemma 4.34. We have

  
† 
˜ t tXF
Op dϕ̃T Σ ⊗ BNs ,Nu e ã BNs ,Nu = etXF ã HW (F )
. (4.60)
HW (TΣ T ∗M )

Proof. From the definition of the norms above, we have

       
˜ dϕ̃tT Σ ⊗ BNs ,Nu etXF ã B †

†
˜ dϕ̃tT Σ BNs ,Nu etXF ã BN


Op Ns ,Nu = Op s ,Nu
HW (TΣ T ∗ M ) L2 HW (N )
tXF
= e ã HW (F )

1325

We come back to the proof of Theorem 4.31. Then

OpΣ etXF ã H (M ) C etXF ã

≤ HW (F )
W
(4.58,4.59,4.60)

giving (4.57).
Here are some example of Theorem 4.31 that we will use later.

Corollary 4.35. We have ∀ϵ > 0, ∃CK,ϵ > 0,∀t ≥ 0, (positive time)

  +
OpΣ etXF T̃≥(K+1) ≤ CK,ϵ et(γK+1 +ϵ) (4.61)
HW (M )

and for any 0 ≤ k1 ≤ k2 ≤ K, ∀ϵ > 0, ∃CK,ϵ > 0,∀t ≥ 0,

  +
OpΣ etXF T̃[k1 ,k2 ] ≤ CK,ϵ et(γk1 +ϵ) . (4.62)
HW (M )

Also ∀t ≤ 0, (negative time)

−
 
OpΣ etXF T̃[k1 ,k2 ] ≤ CK,ϵ et(γk2 −ϵ) . (4.63)
HW (M )

1330 Proof. To get (4.61) we have that ∃C, ∀ϵ > 0, ∃CK,ϵ > 0,∀t ≥ 0,

 
tXF
OpΣ e T̃≥(K+1) ≤ C etXF T̃≥(K+1)
HW (M ) (4.57) HW (F )
+
≤ CK,ϵ e( γK+1 +ϵ )t .
(4.52)

To get (4.62), in the equations below we will use (*) that γk+1
+
≤ γk+ , ∀k ∈ N. We have that
∃C, ∀ϵ > 0, ∃CK , CK,ϵ > 0,∀t ≥ 0,

54
  
tXF[k
OpΣ etXF T̃[k1 ,k2 ] ≤ C etXF T̃[k1 ,k2 ] ≤C e 1 ,k2 ]
HW (M ) (4.57) HW (F ) L2 (Σ;F[k1 ,k2 ] )
+
≤ CK,ϵ max et(γk +ϵ) = CK,ϵ et(γk1 +ϵ)
+

(4.46) k∈[k1 ,k2 ] (∗)

Finally to get (4.63), we repeat the same first lines as above and write ∀ϵ > 0, ∃CK,ϵ > 0,∀t ≥ 0,
 
−tXF −tX
OpΣ e T̃[k1 ,k2 ] ≤C e F[k1 ,k2 ]
HW (M ) L2 (Σ;F[k1 ,k2 ] )
− −
≤ C max e−t(γk −ϵ) = C e−t(γk2 −ϵ) .
K,ϵ K,ϵ
(4.46) k∈[k1 ,k2 ] (∗)

Remark 4.36. In particular, taking t = 0 we get for any k ∈ [0, K],

 
OpΣ T̃k ≤ CK . (4.64)
HW (M ) (4.62)

1335 4.5.2 Composition formula

Recall F[0,K] defined in (4.44). Recall that the operators TNs , TN∆s in definition 3.14 depends on
the exponent 0 < µ < 1 introduced in the cut-off (3.34). Consequently the quantization OpΣ (.)
defined in (4.55) also depends on µ. Recall that β is the Hölder exponent of Eu , Es in (4.3).

Theorem 4.37 (Composition formula). Take µ < β. For any symbols a, b ∈

′
β



C M ; L F[0,K] (Es ) , any t, t ∈ R, we have
 ′   
t′ XF
tXF t XF tXF
(4.65)


OpΣ e ã OpΣ e b̃ ≈ OpΣ e ãe b̃ .

1/2
(where the right hand side term has the factor Υt+t′ in its definition (4.55)).

Proof. For convenience, we first naturally extend the definition 3.12 of the equivalence ≈ in
(3.32) to operators on S (T ∗ M ) and also C β (Σ; F) (instead of S (M ) only). We define

ŤNs := BN† ^)◦ : S (T ∗ M ) → C β (Σ; F) ,

χµ (exp (4.66)
s ,Nu Σ N

−1 ◦ µ
^
ŤN∆s := exp χΣ BNs ,Nu : C β (Σ; F) → C (T ∗ M ) .


N

1340 These operators enter in the following expressions

TNs = ŤNs T , TN∆s = T † ŤN∆s . (4.67)
(4.34) (4.35)

For an operators A : C (Σ; F) → C (Σ; F), we denote

Ø := ŤN∆s AŤNs : S (T ∗ M ) → S ′ (T ∗ M ) (4.68)

.
Definition 4.38. For two operators A, B : S (T ∗ M ) → S ′ (T ∗ M ) we write
A≈B (4.69)
if there exists m < 0, t ∈ R, for any N > 0, any σ > 0, there exists a constant Cσ,N,t > 0 such
1345 that for any ρ, ρ′ ∈ T ∗ M ,
D  E−N
|⟨δρ′ |χΣ,σ (A − B) χΣ,σ δρ ⟩| ≤ Cσ,N,t distg ρ′ , ϕ̃t (ρ) ⟨|ρ|⟩m .

For two operators A, B : C (Σ; F) → C (Σ; F) we write A ≈ B if Ø ≈ B̃.

˜

55
 Using operators TNs , TN∆s defined in (4.34,4.35), let

PNs := TNs TN∆s : C β (Σ; F) → C β (Σ; F) . (4.70)

Lemma 4.39. We have

PN2 s ≈ PNs , PNs TNs ≈ TNs . (4.71)
Proof. Taking t = 0 in (4.36) gives
TN∆s TNs ≈ Id. (4.72)
1350 Then
PN2 s = TNs TN∆s TNs TN∆s ≈ TNs TN∆s = PNs .


(4.70) (4.72) (4.70)

PNs TNs = TNs TN∆s TNs

≈ TNs .
(4.70) (4.72)

We show the following Lemma (similar to [18, Lemma 4.47])

Lemma 4.40 (Basic Lemma). For any symbol a ∈ C β M ; L F[0,K] (Es ) , we have for µ < β

1355 that
[ã, PNs ] ≈ 0. (4.73)
Proof. Observe that
ŤNs ŤN∆s ≈ Id, ŤN∆s ŤNs ≈ PN . (4.74)
(4.66) (4.66)

We have ∀N ≥ 0, ∃CN > 0, ∀ρ, ρ′ ∈ T ∗ M ,

⟨δρ′ |P̃˜Ns δρ ⟩ = ⟨δρ′ |ŤN∆s PNs ŤNs δρ ⟩ = ⟨δρ′ |ŤN∆s ŤNs T T † ŤN∆s ŤNs δρ ⟩ (4.75)
(4.68) (4.70,4.67)
−N
≈ |⟨δρ′ |Pδρ ⟩| ≤ CN ⟨distg (ρ′ , ρ)⟩ .
(4.74,2.35) (2.38)

Then
˜ P̃˜Ns δρ ⟩
h i
⟨δρ′ |ŤN∆s [ã, PNs ] ŤNs δρ ⟩ ≈ ⟨δρ′ | ã, (4.76)
(4.74,4.68)
Z
= ˜ ρ′′ ⟩⟨δρ′′ |P̃˜Ns δρ ⟩ − ⟨δρ′ |P̃˜Ns δρ′′ ⟩⟨δρ′′ |ãδ
⟨δρ′ |ãδ ˜ ρ⟩ .
ρ′′ ∈T ∗ M

The bundle Es is Hölder continuous with some exponent β > 0 and consequently ã satisfies the
following property (called slowly varying property in [18, def 4.40]), using any local trivializa-
tion, for any ϱ, ϱ′ ∈ Σ, setting x = π (ϱ) , x′ = π (ϱ′ ) ∈ M , we assume ω = ω (ϱ) ≍ ω (ϱ′ ), we
1360 use |x′ − x| ≲ ⟨ω⟩−1/2 distg (ϱ′ , ϱ), and we have
(2.25)

∥ã (ϱ′ ) − ã (ϱ)∥ ≍ ∥a (x′ ) − a (x)∥ ≲ |x′ − x| ≍ ⟨ω⟩−β/2 (distg (ϱ′ , ϱ)) .
β β
(4.77)

Hence for ρ, ρ′ ∈ T ∗ M , due to the cutoff χµΣ in (3.44) at distance ⟨ω⟩µ/2 from Σ, writing
ρ ≡ (ϱ, v) ∈ T ∗ M with ϱ ∈ Σ, v ∈ Nϱ , expN (v) = ρ, we have for distg (ρ′ , ρ) ≲ 1 and
distg (ρ′′ , ρ) ≲ 1 that
˜ ρ ⟩ ≤ ⟨ω⟩µ/2 |ã (ϱ′ ) − ã (ϱ)| ≤ ⟨ω⟩µ/2 ⟨ω⟩−β/2 .
˜ ρ′′ ⟩ − ⟨δρ′′ |ãδ
⟨δρ′ |ãδ (4.78)
(4.77)

We deduce
−N
⟨δρ′ |ŤN∆s [ã, PNs ] ŤNs δρ ⟩ ≤ C ⟨ω⟩−(β−µ)/2 CN ⟨distg (ρ′ , ρ)⟩ .
(4.73,4.75)

1365 From definition 4.38, we get that [ã, PNs ] ≈ 0 if β > µ.

56
 Observe that
   
t′ XF tXF t′ XF (t+t′ )XF (t+t′ )XF
tXF
(4.79)


OpΣ e OpΣ e ≈ e e =e ≈ OpΣ e .
(4.56) (4.56)

Also    
OpΣ (ã) OpΣ b̃ = TN∆s ãPNs b̃TNs ≈ TN∆s ãb̃TNs = OpΣ ãb̃ . (4.80)
(4.55,4.70) (4.73,4.71)

and similarly
OpΣ etXF OpΣ (ã) ≈ OpΣ etXF ã (4.81)

OpΣ (ã) OpΣ etXF ≈ OpΣ ãetXF . (4.82)

−t′ XF t′ XF
From these relations, we deduce (4.65) as follows. We set ã′ = e ãe .
 ′
  ′
  
OpΣ etXF ã OpΣ et XF b̃ OpΣ etXF OpΣ et XF OpΣ ã′ b̃



≈
(4.81,4.82,4.80)
     
tXF t′ XF ′ tXF t′ XF
≈ OpΣ e e OpΣ ã b̃ ≈ OpΣ e ãe b̃ .
(4.79) (4.81,4.82,4.80)

1370

The next corollary follows as special cases of (4.65).

Corollary 4.41. Take µ < β. We have

• Egorov formula: for a symbol a ∈ C β M ; End F[0,K] (Es ) , t ∈ R,

OpΣ etXF OpΣ (ã) ≈ OpΣ etadXF ã OpΣ etXF . (4.83)

with the natural action etadXF ã = et[XF ,.] ã := etXF ãe−tXF .

• Approximate projectors: for any k, k ′ ≤ K,

     
OpΣ T̃k OpΣ T̃k′ ≈ δk=k′ OpΣ T̃k . (4.84)

• Decomposition of the dynamics in components: for any k, k ′ ≤ K,

   ′   ′

OpΣ etXF T̃k OpΣ et XF T̃k′ ≈ δk=k′ OpΣ e(t+t )XF T̃k . (4.85)

Remark 4.42. The algebraic structure of (4.85) will manifest itself later in the band structure
of the Ruelle spectrum.
Proof. To get (4.83), we write
OpΣ etXF OpΣ (ã) ≈ OpΣ etXF ã = OpΣ etadXF ã etXF





(4.65)

≈ OpΣ etadXF ã OpΣ etXF

(4.65)

1375 To get (4.84), we write

       
OpΣ T̃k OpΣ T̃k′ ≈ OpΣ T̃k T̃k′ = δk=k′ OpΣ T̃k .
(4.65) (4.40)

To get (4.85), we write

   ′   ′

OpΣ etXF T̃k OpΣ et XF T̃k′ ≈ OpΣ etXF T̃k et XF T̃k′
(4.65)
   
tXF t′ XF (t+t′ )XF
= OpΣ e e T̃k T̃k′ = δk=k′ OpΣ e T̃k .
(4.42) (4.40)

57
 4.5.3 Trace formula
We present some trace formula for PDO (Pseudo Differential Operators), i.e. FIO (Fourier
Integral Operators) at t = 0.

We will use this formula in section 7.1. It is also possible to
1380 express Tr OpΣ 1[ω,ω′ ] e ã for t ̸= 0, but we don’t need it in this paper.
tXF

Theorem 4.43 (Trace formula). Take µ

< β. There exists C > 0, for any ω ≥ δ ≥ 1,
for any symbol a ∈ C β M ; L F[0,K] (Es ) ,

≤ Cω d δ. (4.86)


OpΣ 1ω(ρ)∈[ω−δ,ω+δ] ã TrHW (M )
!Z
ωd
(4.87)



Tr OpΣ 1ω(ρ)∈[ω−δ,ω+δ] ã = (2δ) Tr (a (m)) dm + r (ω, δ) ,
(2π)d+1 M

with remainder  
d −(β−µ)/2 δ
|r (ω, δ)| ≤ Cω δ |ω| + + Cω d .
ω

Remark 4.44. Later in (7.39) we will use (4.87) as follows. For any small c′ > 0 we will take
δ ≫ 1 and ω ≫ δ large enough so that the remainder r (ω, δ) is smaller than c′ ω d δ, i.e. negligible
w.r.t. the first term.
Proof. Since the symbol 1ω(ρ)∈[ω−δ,ω+δ] ã is fiber-wise, we have

OpΣ 1ω(ρ)∈[ω−δ,ω+δ] ã = TN∆s 1ω(ρ)∈[ω−δ,ω+δ] ãTNs

(4.55)
Z
dρ
= A (ρ) 1ω(ρ)∈[ω−δ,ω+δ] ã (ρ) B (ρ) (4.88)
ρ∈Σ (2π)d+1

with for ρ ∈ Σ, the finite dimensional vector space F[0,K] (Ns (ρ)) and the finite rank linear
operators

A (ρ) := TN∆s T̃[0,K] (ρ) : F[0,K] (Ns (ρ)) → C ∞ (M ) (4.89)

B (ρ) := T̃[0,K] (ρ) TNs : C ∞ (M ) → F[0,K] (Ns (ρ)) (4.90)

1385 Where T̃[0,K] (ρ) is the bundle map projector onto F[0,K] (Ns (ρ)), defined in (4.41). We have
∃C > 0, ∀ρ ∈ Σ,
∥A (ρ)∥ ≤ C, ∥B (ρ)∥ ≤ C.
and
1ω(ρ)∈[ω−δ,ω+δ] ã (ρ) TrHW (M )
≤C (4.91)

For ω ≥ δ > 0, we have

Z Z ω+δ   
δ
1ω(ρ)∈[ω−δ,ω+δ] dρ = Vol (M ) ′d ′
ω dω = Vol (M ) ω (2δ) 1 + O d
. (4.92)
Σ (3.8) ω−δ ω

Hence
Z

dρ
OpΣ 1ω(ρ)∈[ω−δ,ω+δ] ã ≤ C 1ω(ρ)∈[ω−δ,ω+δ] ã (ρ)
TrHW (M )
(4.88) Σ
TrHW (M )
(2π)d+1
≤ Cω d δ,
(4.91,4.92)

58
 giving (4.86). Then
Z



dρ
Tr OpΣ 1ω(ρ)∈[ω−δ,ω+δ] ã = Tr 1ω(ρ)∈[ω−δ,ω+δ] ã (ρ) B (ρ) A (ρ) ,
(4.88) Σ (2π)d+1
where
† ^)◦ T T † exp−1 ◦ µ
χµ (exp ^

B (ρ) A (ρ) = T̃[0,K] (ρ) BN s ,Nu Σ N N χΣ BNs ,Nu T̃[0,K] (ρ)
(4.89,4.90,4.35,4.34)
 
−(β−µ)/2
= T̃[0,K] (ρ) + O ⟨ω⟩ ,
1
1390 where ω (ρ)− 2 (β−µ) comes from Hölder irregularity as in (4.78). Hence
Z



dρ
Tr OpΣ 1ω(ρ)∈[ω−δ,ω+δ] ã = Tr 1ω(ρ)∈[ω−δ,ω+δ] ã (ρ) +R
Σ (2π)d+1
with remainder
Z
|R| ≤ C 1ω(ρ)∈[ω−δ,ω+δ] ω (ρ)−(β−µ)/2 dρ + Cω d
Σ
≤ Cω d δω −(β−µ)/2 + Cω d .
(4.92)

Cω d is due to sharp cutoff in frequency, and main term

ω d (2δ)
Z    Z

dρ δ
Tr 1ω(ρ)∈[ω−δ,ω+δ] ã (ρ) =
d+1 (3.8,4.92) d+1
1+O Tr (a (m)) dm.
Σ (2π) (2π) ω M

We have obtained (4.87).

5 Proof of Theorem 1.6 (band spectrum)

1395 We will follow the strategy presented in the proof of [18, Lemma 5.16]. Let K ∈ N, ϵ > 0 be
fixed and consider z ∈ C in the spectral plane such that
+
γK+1 −
+ ϵ < Re (z) < γK − ϵ. (5.1)

We put
ω = Im (z) . (5.2)
Let δ > 0. We consider the frequency intervals

Jω := [ω − 1, ω + 1] , ′
Jω,δ := [ω − 1 − δ, ω + 1 + δ] , (5.3)

so that Jω ⊂ Jω,δ
′
. Let σ > 0. We consider the following partition of the cotangent bundle

T ∗ M = Ω0 ∪ Ω1 ∪ Ω2 (5.4)

1400 with, as in (3.30),

n o
Ω0 := ρ ∈ T ∗ M | ∥ρu + ρs ∥gρ ≤ σ and ω (ρ) ∈ Jω,δ
′
, (5.5)
n o
Ω1 := ρ ∈ T ∗ M | ∥ρu + ρs ∥gρ > σ and ω (ρ) ∈ Jω,δ
′
, (5.6)

Ω2 := ρ ∈ T ∗ M | ω (ρ) ∈ ′

/ Jω,δ .

59
 For j = 0, 1, 2, we define 1Ωj : T ∗ M → [0, 1] as being the characteristic function of the set Ωj
above and as in (3.31), we set
Op 1Ωj := T † 1Ωj T


: S (M ) → S (M ) ,
giving a resolution of identity
IdHW (M ) = Op (1Ω0 ) + Op (1Ω1 ) + Op (1Ω2 )
(5.4)

(5.7)



=OpΣ T[0,K] Op (1Ω0 ) + OpΣ T≥(K+1) Op (1Ω0 ) + Op (1Ω1 ) + Op (1Ω2 )
+ (Id − OpΣ (IdF )) Op (1Ω0 ) . (5.8)
1405 We will now construct the resolvent of X at z from the contribution of each term in (5.7),
except for the last term. For this last term, we observe that
Id − OpΣ (IdF ) ≈ 0.
(4.56)

Since σ is fixed in (5.5) we have that

lim ∥(Id − OpΣ (IdF )) Op (1Ω0 )∥HW (M ) = 0. (5.9)
|ω|→∞ (2.37)

5.1 Contribution of Ω0
Recall that Op (1Ω0 ) defined from Ω0 in (5.5) depends on ω defined in (5.2).

Lemma 5.1. We have (for negative time) ∀ϵ > 0, ∃Cϵ > 0,∀t ≤ 0, ∃ωt > 0,∀ω > ωt ,
  −
tX
e OpΣ T̃[0,K] Op (1Ω0 ) ≤ Cϵ et(γK −ϵ) , (5.10)
HW (M )

and (for positive time) ∀ϵ > 0, ∃Cϵ > 0,∀t ≥ 0, ∃ωt > 0,∀ω > ωt ,
  +
etXF OpΣ T̃[0,K] Op (1Ω0 ) ≤ Cϵ et(γ0 +ϵ) , (5.11)
HW (M )
  +
etX OpΣ T̃≥(K+1) Op (1Ω0 ) ≤ Cϵ et(γK+1 +ϵ) . (5.12)
HW (M )
1410

Proof. Let t ≤ 0. We have

   
etX OpΣ T̃[0,K] ≈ OpΣ etXF OpΣ T̃[0,K]


(4.56)
 
≈ OpΣ etXF T̃[0,K] (5.13)
(4.65)

Now we use the cut-off in frequency Op (1Ω0 ). We have

∥Op (1Ω0 )∥HW (M ) ≤ C, (5.14)
and
   
tX tXF
e OpΣ T̃[0,K] Op (1Ω0 ) ≤ OpΣ e T̃[0,K] Op (1Ω0 ) + r (ω)
HW (M ) (5.13,2.37) HW (M )
 
≤ OpΣ etXF T̃[0,K] ∥Op (1Ω0 )∥HW (M ) + r (ω)
HW (M )
−
≤ CK,ϵ′ et( γK −ϵ′ ) + r (ω) ,
(4.63,5.14)

with some function r (ω) that satisfies lim|ω|→∞ r (ω) = 0. This gives (5.10). Similarly, using
(4.62) we get (5.11) and using (4.61) we get (5.12).

60
 (0) (0)
Approximate resolvent Rin (z) and Rout (z): For T ≥ 0 we define
Z 0   
(0)
Rin (z) := − −tz tX
e e dt OpΣ T̃[0,K] Op (1Ω0 ) . (5.15)
−T

Then
Z T Z T
−
(0) −t(γK −ϵ′ ) ′
Rin (z) ≤ e tRe(z)
Cϵ′ e dt ≤ Cϵ′ e−(ϵ−ϵ )t dt ≤ Cϵ . (5.16)
HW (M ) (5.15,5.10) 0 (5.1) 0

where we have taken ϵ′ = 21 ϵ. We also have

 
(0)
Id − eT z e−T X OpΣ T̃[0,K] Op (1Ω0 )


(z − X) Rin (z) =
(5.15)

and
  −
≤ eT Re(z) Cϵ′ e−T (γK −ϵ )
(0) ′
(z − X) Rin (z) − OpΣ T̃[0,K] Op (1Ω0 ) (5.17)
HW (M ) (5.10)
′ 1
≤ Cϵ′ e−(ϵ−ϵ )T ≤ Cϵ e− 2 ϵT (5.18)
(5.1)

1415 with ϵ = 21 ϵ′ . Similarly we define

Z T   
(0)
Rout (z) := e −tz tX
e dt OpΣ T̃≥(K+1) Op (1Ω0 ) (5.19)
0

and get
(0)
Rout (z) ≤ Cϵ (5.20)
HW (M ) (5.12,5.1)

and  
(0) 1
(z − X) Rout (z) − OpΣ T̃≥(K+1) Op (1Ω0 ) ≤ Cϵ e− 2 ϵT . (5.21)
HW (M ) (5.12,5.1)

5.2 Contribution of Ω1
Recall that Op (1Ω1 ) defined from Ω1 in (5.6) depends on σ > 0. Let σ ′ > 0.

etX Op (1Ω1 ) = etX (Id − Op (χΣ,σ′ )) Op (1Ω1 ) + etX Op (χΣ,σ′ ) Op (1Ω1 )

1420 For the first term, from Theorem 4.6, we have ∀Λ > 0, ∃C > 0, ∀t ≥ 0, ∃σt , ∀σ ′ > σt ,
1 −Λt
etX (Id − Op (χΣ,σ′ )) Op (1Ω1 ) HW (M )
≤ Ce .
(4.22) 2

For the second term, since the support of the symbols are disjoint, we have ∀Λ > 0, ∃C >
0, ∀t ≥ 0, ∀σ ′ , ∃σ > σ ′ ,∃ωt > 0, ∀ |ω| > ωt ,
1
∥Op (χΣ,σ′ ) Op (1Ω1 )∥HW (M ) ≤ Ce−Λt
2
Taking Λ > 0 large enough, we deduce that ∃C > 0, ∀t ≥ 0, ∃σt , ∀σ > σt , ∃ωt > 0, ∀ |ω| > ωt ,
+
etX Op (1Ω1 ) HW (M )
≤ CetγK+1 . (5.22)

61
 Approximate resolvent: For T ≥ 0 we define
Z T
(1)
R (z) := e−tz etX Op (1Ω1 ) dt (5.23)
0

1425 and similarly to (5.16) and (5.17) we get

R(1) (z) HW (M )
≤ Cϵ , (5.24)
(5.22,5.1)

and
(z − X) R(1) (z) − Op (1Ω1 ) HW (M )
≤ Ce−ϵT . (5.25)
(5.22,5.1)

5.3 Contribution of Ω2
We define  
(2) 1
R (z) := Op 1Ω
z − iω (.) 2
We have the following properties. Recall that δ > 0 enters in the definition (5.3) of Jω,δ
′
, hence
1430 for Ω2 .
Lemma 5.2. [18, Lemma 5.17]There exists C > 0 such that for any ω ∈ R, δ > 0,
C
R(2) (z) HW (M )
≤ , (5.26)
δ
and
C
(z − X) R(2) (z) − Op χΩ2,σ (5.27)


HW (M )
≤ .
δ

5.4 Sum of contributions

The final approximate resolvent is defined by
(0) (0)
R (z) := Rin (z) + Rout (z) + R(1) (z) + R(2) (z) .
1435 From previous estimates we have
C
∥R (z)∥HW (M ) ≤ 3Cϵ + ,
(5.16,5.20,5.24,5.26) δ
and
1 C
∥(z − X) R (z) − Id∥HW (M ) ≤ r (ω) + 2Cϵ e− 2 ϵT + Ce−ϵT + ,
(5.7,5.9,5.17,5.21,5.25,5.27) δ
with lim|ω|→∞ r (ω) = 0. Hence we can take in this order T ≫ 1, σ ≫ 1, δ ≫ 1,|ω| ≫ 1, large
enough so that
(z − X) R (z) = Id − r2 (z)
with
1
∥r2 (z)∥HW (M ) < .
2
1440 Hence (Id − r2 (z))−1 is bounded. We set R̃ (z) := R (z) (Id − r2 (z))−1 and get that
(z − X) R̃ (z) = Id.
With a similar construction on the left R̃l (z) (z − X) = Id, we deduce that
(z − X)−1 ≤ Cϵ
is bounded by Cϵ > 0 that does not depend on ω. We have shown (1.14) and (1.15). This
finishes the proof of Theorem 1.6.

62
 6 Proof of Theorem 1.4 (emergence of quantum dynamics)
1445 Recall that the space HW (M ) in (4.18) depends on the weight W in (4.15) that itself depends
on a parameter R ∈ R. Let K ∈ N and γK+1 +
defined in (4.45). Using Theorem 4.8, we can
take R large enough such that the generator X in HW (M ) has discrete Ruelle spectrum on
+
Re (z) > γK+1 .  
We will use the approximate projector OpΣ T̃[0,K] defined from the symbol (4.39) and
1450 (4.55), the operator Op (χω ) in (1.23) that selects the low frequencies. We have the decompo-
sition

 
e tX tX
= Op (χΣ,σ ) e OpΣ T̃[0,K] Op (χΣ,σ ) (Id − Op (χω )) (6.1)
  
+ Op (χΣ,σ ) etX Id − OpΣ T̃[0,K] Op (χΣ,σ ) (Id − Op (χω )) (6.2)
+ (Id − Op (χΣ,σ )) etX (Id − Op (χω )) + Op (χΣ,σ ) etX (Id − Op (χΣ,σ )) (Id − Op (χω ))
(6.3)
+ etX Op (χω ) (6.4)

For the first term in the right hand side of (6.1) we have
   
etX OpΣ T̃[0,K] ≈ OpΣ etXF OpΣ T̃[0,K]


(4.56)
 
≈ OpΣ etXF T̃[0,K] (6.5)
(4.65)

hence
 
Op (χΣ,σ ) etX OpΣ T̃[0,K] Op (χΣ,σ ) (Id − Op (χω )) (6.6)
 
= Op (χΣ,σ ) OpΣ etXF T̃[0,K] Op (χΣ,σ ) (Id − Op (χω )) + rt,ω (6.7)

with
lim ∥rt,ω ∥HW (M ) = 0. (6.8)
|ω|→∞ (2.37)

For the second term in the right hand side of (6.1), we proceed exactly as for the proof of
1455 Eq.(5.12) in Lemma 5.1 above, with the difference that we replace Op (1Ω0 ) by Op (χΣ,σ ) (Id − Op (χω )).
We get that ∀ϵ > 0, ∃Cϵ > 0,∀t ≥ 0, ∃ωt > 0,∀ω > ωt ,
   +
Op (χΣ,σ ) etX
Id − OpΣ T̃[0,K] Op (χΣ,σ ) (Id − Op (χω )) ≤ Cϵ et(γK+1 +ϵ)
HW (M )

Let Λ ≥ −γK+1
+
. For the terms in line (6.3), we have ∃C > 0, ∀t ≥ 0, ∃σt , ∀σ > σt ,
+
(Id − Op (χΣ,σ )) etX (Id − Op (χω )) HW (M )
≤ Ce−Λt ≤ CetγK+1 ,
(4.23)

+
Op (χΣ,σ ) etX (Id − Op (χΣ,σ )) (Id − Op (χω )) HW (M )
≤ Ce−Λt ≤ CetγK+1 .
(4.23)

The operator Op (χω ) is compact hence we can decompose the last term in the right hand side
1460 of (6.1) and the last term in (6.5) as
   
etX − Op (χΣ,σ ) OpΣ etXF T̃[0,K] Op (χΣ,σ ) Op (χω ) = Rt + r′′ (6.9)

63
 +
where Rt is finite rank and ∥r′′ ∥HW (M ) ≤ Cϵ et(γK+1 +ϵ) . We get
 
e tX
= Op (χΣ,σ ) OpΣ e tXF
T̃[0,K] Op (χΣ,σ ) + Rt + r′′ + rt,ω ,
(6.1,6.6,6.9)

+
where |ω| is taken large enough in (6.8) so that ∥rt,ω ∥HW (M ) ≤ Cet(γK+1 +ϵ) also. We have
obtained ∃C > 0, ∀t > 0, ∃σt > 0,∀σ > σt , ∃Rσ,t finite rank operator,
  +
etX − Op (χΣ,σ ) OpΣ etXF T̃[0,K] Op (χΣ,σ ) + Rσ,t ≤ Cet(γK+1 +ϵ) .
HW (M )

giving theorem 1.4.

 
1465 Remark 6.1. by the effects of weight W we can replace Op (χΣ,σ ) OpΣ etXF T̃[0,K] Op (χΣ,σ )
 
by OpΣ etXF T̃[0,K] in (1.12).

7 Proof of Theorem 1.10 (Weyl law)

The aim of this section is to prove Theorem 1.10. We do this in three steps.

7.1 Approximate projector

1470 For ω ∈ R, δ > 0 and k ∈ N, we define an approximate projector P for the union of band
k ′ ∈ [0, k] and frequency interval [ω − δ, ω + δ]
as follows. We will

relate its trace to a symplectic
volume. We denote 1[ω−δ,ω+δ] : C Σ; F[0,k] → C Σ; F[0,k] the multiplication operator by
β β

the characteristic function for ρ ∈ Σ,

(
1 if ω (ρ) ∈ [ω − δ, ω + δ]
1[ω−δ,ω+δ] (ρ) := .
0 if not.

We define the operator

 
P := OpΣ 1[ω−δ,ω+δ] T̃[0,k] : C ∞ (M ) → C ∞ (M ) . (7.1)

1475 The next lemma shows that the trace of P is related to the symplectic volume of the support
of function 1[ω−δ,ω+δ] on the trapped set Σ.

Lemma 7.1. The operator P is trace class in HW (M ) and ∃C > 0,∀ω ≥ δ ≥ 1,

∥P ∥Tr ≤ Cω d δ, ∥P ∥HW ≤ C, (7.2)

Vol (M ) ω d
 
−(β−µ)/2 δ
Tr (P ) − rank F[0,k] (2δ) ≤ Cω δ |ω| d
+ + Cω d (7.3)
(2π)d+1 ω

Remark 7.2. Later in (7.39) we will use (7.3) as follows. For any small c′ > 0 we will take
δ ≫ 1 and ω ≫ δ large enough so that the right hand side of (7.3) is smaller than c′ ω d δ.
1480 Proof. We have  
∥P ∥Tr = OpΣ 1[ω−δ,ω+δ] T̃[0,k] ≤ Cω d δ,
(7.1) Tr (4.86)
 
∥P ∥HW (M ) = OpΣ 1[ω−δ,ω+δ] T̃[0,k] ≤ C.
(7.1) HW (M ) (4.57)

64
 Eq. (7.3) comes from (4.87) because T[0,k] (m) is a projector onto Pol[0,k] (Es (m)) hence




Tr T[0,k] (m) = dim Pol[0,k] (Es (m)) = rank F[0,k]
(4.38) (4.27)

and Z



Tr T[0,k] (m) dm = Vol (M ) rank F[0,k] .
M

1485 7.2 Truncated resolvent

Definition 7.3. For z ∈ C, Λ > 0, let

R (z) := (z − (X − ΛP ))−1 (7.4)

D (z) := (z − X)−1 − R (z) (7.5)

Remark 7.4. A similar construction of truncated resolvent R (z) has been used in [18, Proof of
Prop. 5.14].
Let δ ′ > 0 and define the rectangle in the spectral plane

R̃ (ω, δ, δ ′ ) := γk− − 2ϵ, γ0+ + 2ϵ × i [ω − δ + δ ′ , ω + δ − δ ′ ] (7.6)

 

1490 The next lemma is similar to [18, Lemma 5.16] so we omit the proof.

Lemma 7.5. There exists C > 0 such that, if δ ′ > 0 in (7.6) and Λ > 0 in (7.4) are
sufficiently large, then for any ω ∈ R, δ > δ ′ > 0, z ∈ R̃ (ω, δ, δ ′ ),

∥R (z)∥HW (M ) ≤ C. (7.7)

Remark 7.6. A consequence of Lemma 7.5 and expression (7.5) is that the poles of D (z) in
R̃ (ω, δ, δ ′ ) coincide with the poles of the resolvent (z − X)−1 , i.e. Ruelle resonances.
Let us consider the following union of two horizontal bands (see Figure 7.1)

B± (ω, δ, δ ′ ) := γk− − ϵ, γ0+ + ϵ × i [ω ± (δ − 2δ ′ ) , ω ± (δ − δ ′ )] , (7.8)

 

B (ω, δ, δ ′ ) := B+ (ω, δ, δ ′ ) ∪ B− (ω, δ, δ ′ ) . (7.9)

1495 We will need later (in the proof of Lemma 7.11), the following proposition.

Proposition 7.7. The operator D (z) : HW (M ) → HW (M ) is meromorphic w.r.t. z ∈

R̃ (ω, δ, δ ′ ) and trace class outside of its poles of first order. In particular Tr (D (z)) ∈ L1loc
is locally integrable and with setting z := x + iy, we have ∃C > 0, ∀δ > 1, ∀ω > δ,
Z
|Tr (D (z))| dxdy ≤ Cδ ′ ω d ln δ. (7.10)
B(ω,δ,δ ′ )

Proof. We first show few lemmas.

65
 Γ− Γ+ Im(z)
ω+δ
B+ (ω, δ, δ ′ ) δ′
R̃(ω, δ, δ ′ )
R(ω, δ, δ ′ ) ω

B+ (ω, δ, δ ′ )
ω−δ
ϵ Re(z)

γk− γ0+

Figure 7.1: Picture in the spectral plane for the rectangular domain R̃ (ω, δ, δ ′ ) defined in (7.6),
the sub-domains B± (ω, δ, δ ′ ) defined in (7.8) and R (ω, δ, δ ′ ) defined later in (7.30).

Proof.

Lemma 7.8. D (z) is trace class outside of its poles and

Tr (D (z)) = ∂z ln F (z) (7.11)

with

A (z) := Id − R (z) ΛP, (7.12)

F (z) := det (A (z)) . (7.13)

D (z) is meromorphic because both (z − X)−1 and R (z) are meromorphic, see [18, Lemma
5.16]. R (z) is even holomorphic on the rectangle R̃ (ω, δ, δ ′ ), from Lemma 7.5. Then

D (z) = (z − X)−1 − R (z)

(7.5)

= (z − (X − ΛP ))−1 ((z − (X − ΛP )) − (z − X)) (z − X)−1

=R (z) ΛP (z − X)−1 . (7.14)

In this last expression P is trace class. (z − X)−1 and R (z) are bounded outside of their poles
hence D (z) is trace class outside the poles of (z − X)−1 and R (z). We have

D (z) = (z − X)−1 − R (z) = (z − (X − ΛP ) + ΛP )−1 − R (z)

(7.5)

= R (z) (Id − R (z) ΛP )−1 − R (z) = R (z) A (z)−1 − Id . (7.15)

(7.4) (7.12)

1500 We compute

∂z A (z) = R (z)2 ΛP = R (z) (Id − A (z)) = D (z) A (z) , (7.16)

(7.12) (7.12) (7.15)

then we have

∂z ln A (z) = (∂z A (z)) A (z)−1 = D (z) ,

(7.16)

ln F (z) = ln det (A (z)) = Tr (ln A (z)) ,

(7.13)

∂z ln F (z) = Tr (∂z ln A (z)) = Tr (D (z)) .

We have obtained (7.11).

66
 For later use, we also notice that

R (z) = A (z) (z − X)−1 . (7.17)

(7.14)

Lemma 7.9. There exists C > 0, such that for every z ∈ B (ω, δ, δ ′ ),
d
|F (z)| ≤ eC(ln δ)ω . (7.18)

and for every z ∈ B (ω, δ, δ ′ ) with Re (z) = γk− − ϵ,

d
|F (z)| ≥ e−C(ln δ)ω . (7.19)
1505

Proof. For z ∈ B− (ω, δ, δ ′ ), writing λj (z), j = 1, 2, · · · , J, for the eigenvalues of the trace class
operator R (z) ΛP , we have by Lidskii’s theorem
X
log |F (z)| = log |det (Id − R (z) ΛP )| = log |1 − λj (z)|
j
X X
≤ log (1 + |λj (z)|) ≤ |λj (z)| ≤ ∥R (z) ΛP ∥Tr . (7.20)
j j

For the last inequality, see e.g. [22, p.64]. Notice that ∥R (z) ΛP ∥Tr ≤ Cδω d but this is not
(7.2,7.7)
enough to get (7.18). We will improve the last bound as follows. We decompose the frequency
interval [ω − δ, ω + δ] as the union of ⌊δ⌋ (the integer part of δ) intervals of bounded length
2δ
l = [δ] , so that
[
[ω − δ, ω + δ] = Iw (7.21)
w∈{1,2,...,⌊δ⌋}

1510 with
Iw := [ω − δ + (w − 1) l, ω − δ + wl] . (7.22)
Parallel to (7.1) we define
ˇ Σ T[0,k] χIw Ťσ
Pw := Ťσ∆ Op : C ∞ (M ) → C ∞ (M )

i.e. an approximate projector on frequency interval of size l and we write

X
P = Pw . (7.23)
w

Similarly to (7.2) we have

∥Pw ∥Tr ≤ Cω d
and
1
∥R (z) ΛPw ∥Tr ≤ Cω d (7.24)
w
1515 uniformly for 1 ≤ w ≤ ⌊δ⌋. Hence
[δ]
X
∥R (z) ΛP ∥Tr ≤ ∥R (z) ΛPw ∥Tr ≤ Cω d ln δ.
w=1

We get
d
|F (z)| ≤ e∥R(z)ΛP ∥Tr ≤ eCω ln δ
.
(7.20)

67
 We have obtained (7.18). To prove the second claim (7.19), we consider an arbitrary z ∈
B (ω, δ, δ ′ ) with Re (z) = γk− − ϵ. From Corollary 1.6 we have (z − X)−1 ≤ C. Then
(Id − R (z) ΛP )−1 = A (z)−1 = (z − X)−1 R (z)−1 = Id + (z − X)−1 ΛP


(7.17) (7.4)

is uniformly bounded from (7.2). Hence |1 − λj (z)| > c with some c > 0 independent on ω and
1520 therefore we have |log |1 − λj (z)|| < C ′ |λj (z)| for some C ′ > 0. Thus
X X
− log |F (z)| = − log |1 − λj (z)| ≤ C ′ |λj (z)| ≤ C ′ ∥R (z) ΛP ∥Tr ≤ C ′ ω d ln δ,
j j

giving (7.19).
Let spect (X) denotes the discrete Ruelle spectrum of X.

Lemma 7.10. We have ∃C > 0, ∀ω,∀z ∈ B (ω, δ, δ ′ ) ,

X 1
∂z ln F (z) − ≤ C (ln δ) ω d (7.25)
z − zj
zj ∈spect(X)∩B(ω,δ,δ ′ )

where the sum over zj ∈ spect (X) ∩ B (ω, δ, δ ′ ) is counted with multiplicity of the Ruelle
eigenvalues.

Proof. This is lemma α in [39, p.56]. We reproduce the proof here. From (7.7), F (z) is a
1525 well defined holomorphic function in the domains B (ω, δ, δ ′ ) and from (7.17), the zeroes of
F (z) coincide up to multiplicities with the poles of (z
− X)−1 , i.e. Ruelle eigenvalues zj ∈
spect (X) ∩ B (ω, δ, δ ′ ). Using Weyl upper bound O ω d on the density in [18, Thm.3.6], note
that the number of such Ruelle eigenvalues is bounded by Cδ ′ ω d ≤ Cδω d . From (7.19), we can
′
fix some z0 ∈ B (ω, δ, δ ′ ) with Re (z0 ) = γk− − ϵ so that |F (z0 )| ≥ e−C δω . The function
d

(7.19)
Q
F (z) z (z0 − zj )
G (z) := Qj . (7.26)
F (z0 ) zj (z − zj )
1530 G (z) is holomorphic on B (ω, δ, δ ′ ) and from estimate (7.18) on the boundary of B (ω, δ, δ ′ ) and
maximum modulus principle we get
d
∀z ∈ B (ω, δ, δ ′ ) , |G (z)| ≤ eCω ln δ
. (7.27)
(7.18)

Moreover G (z) has no zero on on the simply connected domain B (ω, δ, δ ′ ) hence
g (z) := ln G (z) (7.28)
is well defined as a holomorphic function on B (ω, δ, δ ′ ) with g (z0 ) = 0 and satisfies Reg (z) ≤
(7.27)
Cω ln δ. By Borel-Carathéodory theorem
d 12
this implies
|g (z)| < C ′ ω d ln δ
1535 on a region slightly larger than B (ω, δ, δ ′ ) and therefore by Cauchy integral formula
X 1
|∂z g (z)| = ∂z ln F (z) − < C ′′ ω d ln δ
(7.28,7.26)
zj
z − zj

giving the conclusion.

12
Here we do not need to give the constants and the proof is simple: g maps B (ω, δ, δ ′ ) on the half-plane
Re (z) < Cδω d . We post-compose by a Moebius transformation f which maps this half-plane to the unit disk
and maps g (z0 ) = 0 to 0. Then the image of B (ω, δ, δ ′ ) by f ◦ g is contained in a disk |z| < r < 1. This implies
that the image g (B (ω, δ, δ ′ )) is contained in a (large) disk |z| < C ′ δω d .

68
 Now we finish with the proof of Lemma 7.7. Using Weyl upper bound on the density in [18,
Thm. 2.3], we have, with z = x + iy, that
Z
X 1
dxdy < C δ ′ ω d ln δ ′ . (7.29)


′ B(ω,δ,δ ′ ) |z − zj |
zj ∈spect(X)∩B(ω,δ,δ )

Then
Z Z
|Tr (D (z))| dxdy = |∂z ln F (z)| dxdy
B(ω,δ,δ ′ ) (7.11) B(ω,δ,δ ′ )

≤ Cδ ω ln δ + Cδ ′ ω d ln δ ′ ≤ 2Cδ ′ ω d ln δ.
′ d
(7.25,7.29)

We have obtained (7.10).

1540 7.3 Counting resonances by argument principle

From Lemma 7.5, the truncated resolvent R (z) is bounded on the rectangle R̃ (ω, δ, δ ′ ) and
hence by the definition (7.5) the poles of z → Tr D (z) coincides with the poles of (z − X)−1 ,
i.e., the Ruelle resonance. In the following, we count the number of Ruelle resonance in the
rectangle R̃ (ω, δ, δ ′ ) by using a slightly generalized version of the argument principle. We
1545 learned this method from the paper of S. Dyatlov [8, Section 11].
In the following, we suppose that δ > 0, δ ′ > 0 is fixed so that 2δ ′ < δ and we consider the
limit |ω| → ∞. We take a C ∞ function fδ,δ′ : R → [0, 1] that is fδ,δ′ (ω ′ ) = 1 for |ω ′ | < δ − 2δ ′
and fδ,δ′ (ω ′ ) = 0 for |ω ′ | > δ − δ ′ and |∂ω′ f | < Cδ ′−1 . Then for ω ∈ R, z ∈ C, let

fω,δ,δ′ (z) := fδ,δ′ (Im (z − iω))

Let us define the rectangle in the spectral plane (see Figure 7.1)

R (ω, δ, δ ′ ) := γk− − ϵ, γ0+ + ϵ × i [ω − δ + 2δ ′ , ω + δ − 2δ ′ ] . (7.30)

 

1550 In the following we suppose that the k-th band is separated from (k + 1)-st band i.e.
+
γk+1 < γk− .

Note that R (ω, δ, δ ′ ) is a proper subset of R̃ (ω, δ, δ ′ ) defined in (7.6). We have that ∀z ∈
R (ω, δ, δ ′ ), fω,δ,δ′ (z) = 1 and ∀z ∈ R̃ (ω, δ, δ ′ ),

|∂z fω,δ,δ′ (z)| < Cδ ′−1 . (7.31)

We consider the following paths Γ := Γ− ∪ Γ+ on C, parameterized as

Γ− : ω ∈ R → z = γk− − ϵ + iω ∈ C,

Γ+ : ω ∈ R → z = γ0+ + ϵ + iω ∈ C.
1555 From (7.7) and (1.15), there is no pole of D(z) on Γ∩supp (fω,δ,δ′ )Rsupposing that |ω| is suffi-
ciently large. We know that D (z) is in trace class. Hence the integral Γ+ −Γ− fω,δ,δ′ (z) Tr (D (z)) dz
is well defined. The next lemma shows that this integral is close to the number of Ruelle reso-
nances in the rectangle R (ω, δ, δ ′ ).

Lemma 7.11. We have ∃C > 0,∀δ > 1, ∀ω > δ, ∀0 < δ ′ < δ,

 Z 
1  
fω,δ,δ′ (z) Tr (D (z)) dz − ♯ spect (X) ∩ R̃ (ω, δ, δ ′ ) < Cω d max (ln δ, δ ′ )
2πi Γ+ −Γ−
(7.32)

69
 Proof. From the definition of fω,δ,δ′ above, we have fω,δ,δ′ (z) = 0 if |Im (z) − ω| > δ − δ ′ , so we
can close the contour integral on horizontal segments Im (z) = ω ± (δ − δ ′ ). The residues of the
poles of D (z) inside the contour are the spectral projector onto eigenspaces. Hence Cauchy
integral formula for smooth functions gives
Z
1 X
fω,δ,δ′ (z) Tr (D (z)) dz = fω,δ,δ′ (zj )
2πi Γ+ −Γ−
zj ∈spect(X)∩R̃(ω,δ,δ ′ )
Z
1
+ (∂z fω,δ,δ′ ) (z) Tr (D (z)) dxdy
π R̃(ω,δ,δ′ )

with z = x + iy. We have ∃C > 0,∀δ > 0, ∀δ ′ < 12 δ, ∀ω > δ,

Z Z
′−1
(∂z fω,δ,δ′ ) (z) Tr (D (z)) dxdy ≤δ |Tr (D (z))| dxdy
R̃(ω,δ,δ ′ ) B(ω,δ,δ ′ )

≤ Cδ ′−1 ′
(ln δ) δ ω d = Cω d ln δ. (7.33)
(7.31,7.10)

1560 From properties of fω,δ,δ′ and using Weyl upper bound on the density in [18, Thm.2.3] we have

X n o
′
fω,δ,δ′ (zj ) − ♯ zj ∈ spect (X) ∩ R̃ (ω, δ, δ ) ≤ Cω d δ ′ .
zj ∈spect(X)∩R̃(ω,δ,δ ′ )

We have obtained (7.32).

7.4 Relation with the symplectic volume

The last step is to relate the integral 2πi
1
f ′ (z) Tr (D (z)) dz in (7.32) to the trace of
R
Γ+ −Γ− ω,δ,δ
P in (7.3). Recall Definition (7.1) of P . We define the operator


P0 := OpΣ 1[ω−δ+δ′ /2,ω+δ−δ′ /2] T[0,k] .
1565 For z ∈ R (ω, δ, δ ′ ), let
d(z) = min {Im(z) − (ω + δ + δ ′ /2), Im(z) + (ω − δ − δ ′ /2)} ≥ δ ′ /2.

Lemma 7.12 (Approximate expressions of D (z)). There exists C > 0 such that for
any T > 0, ω > δ > δ ′ > 1 and any z ∈ C with Im (z) ∈ [ω − δ + δ ′ , ω + δ − δ ′ ] and
Re (z) = γ0+ + ϵ, we have
Z T Z T
δ′
   
−tz t(X−Λ) −tz tX d −ϵT
D (z) − e e dt (ΛP0 ) e e dt ≤ Cω e +
0 0 Tr (d (z))2
(7.34)
If instead Re (z) = γk− − ϵ, we have
T 0
δ′
Z  Z   
−tz t(X−Λ) −tz tX
D (z) − e e dt (ΛP0 ) e e dt ≤ Cω e−ϵT +
d
0 −T Tr (d (z))2
(7.35)

Proof. Let z ∈ C with Im (z) ∈ [ω − δ + δ ′ , ω + δ − δ ′ ]. In the following wePlet T > 0 be

sufficiently large. Suppose Re (z) = γ0+ + ϵ. As in (7.23), we decompose P = w∈W Pw with
W := {1, 2, . . . , ⌊δ⌋}. As in (7.21) with interval Iw defined in (7.22) we decompose
[
[ω − δ + δ ′ /2, ω + δ − δ ′ /2] = Iw
w∈W0

70
1570 with a subset W0 ⊂ W . So we can write
X
P0 = Pw .
w∈W0

We have X
D (z) = Dw (z)
(7.14)
w∈W

with
Dw (z) := ΛR (z) Pw (z − X)−1 .
For w ∈ W0 , let
Z T  Z T 
−tz t(X−Λ) −tz tX
D̃w (z) := Λ e e dt Pw e e dt .
0 0

Then
T T
(z − X + Λ) D̃w (z) (z − X) = Λ e−tz et(X−Λ) 0 Pw e−tz etX 0
 

= ΛPw + Λrw ,

with
rw = eT (−z+X−Λ) Pw + Pw eT (−z+X) + eT (−z+X−Λ) Pw eT (−z+X) .
1575 We deduce that

D̃w (z) = Dw (z) + ΛR (z) rw (z − X)−1

with

R (z) rw (z − X)−1 = R (z) eT (−z+X−Λ) Pw (z − X)−1 + R (z) Pw (z − X)−1 eT (−z+X)

+ R (z) eT (−z+X−Λ) Pw (z − X)−1 eT (−z+X) .

For the right hand side of the last equality, we have the following estimates. We use the notation
OTr (∗) for an operator whose trace norm is bounded by a constant multiple of ∗. If w ∈ W0
then as in (7.24), we have
 
−1 d −ϵT 1
R (z) rw (z − X) = OTr ω e ,
dist (Iw , Im (z))2
where dist (Iw , Im (z)) is the distance between Iw and Im (z). If w ∈ W \W0 then
 
d 1
Dw (z) = OTr ω .
(d (z))2
1580 Taking the sum with respect to w ∈ W0 , we see
X −1
X  1

O ω d e−ϵT = O ω d e−ϵT ,


R (z) rw (z − X) =
w∈W0
Tr
w∈W0
dist (Iw , Im (z))2

δ′
   
X 1
OTr ω d = OTr ω d .
w∈W \W0
(d (z))2 (d (z))2
Hence
T T
δ′
Z  Z    
−tz t(X−Λ) −tz tX
D (z) − e e dt (ΛP0 ) e e dt = O ω d e−ϵT + .
0 0 Tr (d (z))2

71
 Now we give the key estimate.

Lemma 7.13. For any arbitrarily small c > 0, if we take δ large enough, δ ′ > 1 large
enough and δ/δ ′ large enough, we have
 Z 
1
fω,δ,δ′ (z) Tr (D (z)) dz − Tr (P ) < cω d δ (7.36)
2πi Γ+ −Γ−

1585 Proof. We first consider the integration on Γ+ . From the last lemma, we have

Z Z
fδ,δ′ (ω ′ − ω) Tr D γ0+ + ϵ + iω ′ d (iω ′ )



fω,δ,δ′ (z) Tr (D (z)) dz =
Γ+ R
Z Z T  Z T 
′ −t′ (γ0+ +ϵ+iω ′ ) t′ (X−Λ) ′ −t(γ0+ +ϵ+iω ′ ) tX
= iTr fδ,δ′ (ω − ω) e e dt (ΛP0 ) e e dt dω ′
(7.34) ω ′ ∈R t′ =0 t=0
(7.37)
d −ϵT ′



+ OT r ω δe +δ .

Let
Z
+ +
e−t(γ0 +ϵ+iω ) fδ,δ′ (ω ′ − ω) dω ′ = e−t(γ0 +ϵ+iω) fˆδ,δ′ (t) ,
′
f˜ω,δ,δ′ (t) :=
R

′
where fˆδ,δ′ (t) := e−itω fδ,δ′ (ω ′ ) dω ′ satisfies a fast decay for |t| ≫ δ ′−1 as follows: ∀N >
R
R
0, ∃CN > 0, ∀t,
Z
−N
fˆδ,δ′ (t) ≤ CN ⟨δ ′ t⟩ and fˆδ,δ′ (t) dt = 2πfδ,δ′ (0) = 2π. (7.38)
R

Then, with the change of variables (t, t′ ) → (t, s = t + t′ ), we see

Z
fω,δ,δ′ (z) Tr (D (z)) dz
Γ+
Z T Z T  ′ 
f˜ω,δ,δ′ (t + t′ ) Tr et (X−Λ) P0 ΛetX dt′ dt + OT r ω d δe−ϵT + δ ′



= i
(7.37) 0 0
!
Z 2T Z min{T,s}
+
e−s(γ0 +ϵ+iω) fˆδ,δ′ (s) Tr e(s−t)(X−Λ) P0 ΛetX dt ds


=i
0 t=max{0,s−T }

+ OT r ω d δe−ϵT + δ ′



.

From (7.38), by letting δ ′ be large (and fˆδ,δ′ (t) be more concentrate to 0), we get
Z
fω,δ,δ′ (z) Tr (D (z)) dz = OT r δ ′−1 ω d δ + OT r ω d δe−ϵT + δ ′ .




Γ+

72
 Next we consider the integration on Γ− . Similarly to the argument above, we see
Z Z
fω,δ,δ′ (z) Tr (D (z)) dz = i fδ,δ′ (ω ′ − ω) Tr D γ0− − ϵ − iω ′ dω ′



−Γ−
Z Z T  Z 0 
′ −t′ (γ0− −ϵ−iω ′ ) t′ (X−Λ) ′ −t(γ0− −ϵ−iω ′ ) tX
= iTr fδ,δ′ (ω − ω) e e dt (ΛP0 ) e e dt dω ′
(7.34) 0 −T
+ OTr ω d δe−ϵT + δ ′



Z T Z min{0,s} !
−
e−s(γ0 −ϵ−iω ) fˆδ,δ′ (s) Tr es(X−Λ) e−t(X−Λ) ΛP0 etX dt ds
′

=i
s=−T t=max{−T,s−T }

+ OTr ω δe−ϵT + δ ′
d



Z 0
Tr e−t(X−Λ) ΛP0 etX dt + OT r δ ′−1 ω d δ + OTr ω d δe−ϵT + δ ′





=2πi
t=−T
 0 
=2πiTr e P0 t=−T + OT r δ ′−1 ω d δ + OTr ω d δe−ϵT + δ ′
tΛ

=2πiTr (P0 ) + OT r (e−ϵT ω d ) + OT r δ ′−1 ω d δ + OTr ω d δe−ϵT + δ ′

=2πiTr (P ) + OTr ω d δ ′ + OTr (e−ϵT ω d ) + OTr δ ′−1 ω d δ + OTr ω d δe−ϵT + δ ′ .

Therefore, by letting T large δ ≫ δ ′ ≫ 1 and then letting ω large, we obtain the required
1590 estimate.

7.5 Weyl law

Using the previous estimates we get that

 
Vol (M ) ω d
♯ spect (X) ∩ R̃ (ω, δ, δ ′ ) − rank F[0,k] (2δ)
(2π)d+1
   
d ′ d d −β/2 δ d −N −β/2 δ
< Cω max {δ , ln δ} + cω δ + Cω δ |ω| + + Cω CN σ + |ω| σ+
(7.32,7.36,7.3) ω ω
′ d
<c ω δ. (7.39)

with arbitrarily small c′ > 0 and ω ≫ δ ≫ δ ′ ≫ 1. Comparing this estimate for adjacent k’s,
we obtain Theorem 1.10.

1595 8 Proof of Theorem 1.12 (accumulation of eigenvalues on

narrower bands)
The aim of this section is to prove Theorem 1.12. We consider the operator XF defined in Section
2.2 where X is a contact Anosov flow. For k ∈ N, consider the vector bundle π : Fk → Σ, with
Fk = |detNs |−1/2 ⊗ Polk (Ns ) ⊗ F defined in (4.43), twisted by F , and the flow induced from
1600 XF ,
ϕtk : Fk → Fk , t ∈ R,
that is a linear bundle map over ϕ̃t : Σ → Σ. For a point ρ ∈ Σ, t ∈ R, let
1/t !
∥ϕtk (v)∥

gk,t (ρ) := ln max ,
v∈π −1 (ρ) ∥v∥

that defines a continuous function gk,t on Σ. Let dϱ be the Liouville measure (3.8) on Σ and
◦ †


µk,t := gk,t (dϱ)

73
 be the push-forward measure on R by the map gk,t . We have already defined γk± in (4.45) that
can be written as
γk+ := lim max gk,t (ρ) = lim max (supp (µk,t )) ,
t→+∞ ρ∈Σ t→+∞
1605
γk− := lim min gk,t (ρ) = lim min (supp (µk,t )) .
t→−∞ ρ∈Σ t→−∞

Now we define the maximal and minimal Lyapounov exponents with respect to the
Liouville measure by   
+
γ̌k = max supp lim µk,t , (8.1)
t→+∞
  
−
γ̌k = min supp lim µk,t .
t→−∞

One has
γk− ≤ γ̌k− ≤ γ̌k+ ≤ γk+ .
From the definition of γ̌k− , γ̌k+ , for any small ϵ > 0, if Iϵ := γ̌k− − ϵ, γ̌k+ + ϵ and
 
1610

−1


Σω,δ,Iϵ ,t := gk,t (Iϵ ) ∩ [ω, ω + δ] ⊂ Σ,

then we have that for any c > 0, there exists Tc > 0 such that

(dϱ) (Σ\Σω,δ,Iϵ ,Tc ) ≤ cδω d . (8.2)

For the proof of Theorem 1.12, we follow [16, Thm 1.3.11 and Chapter 8,]. Let c > 0. Take
Tc > 0 large enough such that (8.2) holds true. Then similarly to (4.62) one gets ∀ϵ > 0, ∃Cϵ >
0,∀t ≥ 0, ∃σt > 0, ∀σ > σt , ∃ωσ > 0, ∀ω > ωσ ,
+
etX OpΣ 1Σω,δ,Iϵ ,Tc Tk ≤ Cϵ et(γ̌k +ϵ) ,


HW (M )
1615
−
e−tX OpΣ 1Σω,δ,Iϵ ,Tc Tk ≤ Cϵ e−t(γ̌k −ϵ) ,


HW (M )

= cδω d + Cω d .


OpΣ 1Σ\Σω,δ,Iϵ ,Tc Tk Tr

Using arguments with perturbation of the resolvent as in previous proofs, we deduce the The-
orem.

9 Proof of Theorem 1.17 (horocycle operators)

1620 Suppose s ∈ C β (M ; Es ) is a Hölder continuous section of the stable bundle over M and
u ∈ C β (M ; Eu ) a section of the unstable bundle. Let π : T ∗ M → M be the canonical
projection. We have seen in (4.25) that for ρ ∈ Σ, the differential dπ gives a linear isomorphism
dπ : N (ρ) = Ns (ρ) ⊕ Nu (ρ) → Eu (m) ⊕ Es (m) with m = π (ρ). By pull back we get sections
s̃ := (dπ)−1 (s) ∈ C (Σ; Ns ) and ũ := (dπ)−1 (u) ∈ C (Σ; Nu ). Recall from Section 4.1.2 that
1625 dA is a symplectic form on Eu ⊕ Es and from Definition 3.4 that N (ρ) = Ns (ρ) ⊕ Nu (ρ) is a
Ω-symplectic vector space and that Ωρ (ũ, s̃) = ω (ρ) . (dA) (u, s), hence ũ (so then u) defines
(3.9)
a dual vector
ũ∗ := Ω̌ũ = Ω (ũ, .) = ω (.) (dA) (u, dπ (.)) ∈ Ns∗ (ρ) . (9.1)
(A.1) (3.9)

For k ∈ N, let p ∈ C (Σ; Fk ) a continuous section of Fk (Ns ) = |detNs |−1/2 ⊗Polk (Ns )⊗F
(4.27)
over Σ where Polk (Ns ) ≡ Symk (Ns∗ ⊗ . . . ⊗ Ns∗ ) is equivalent to the symmetric tensor algebra
1630 of degree k.

74
 Definition 9.1. For s ∈ C β (M ; Es ) we define
(
C β (Σ; Fk ) → C β (Σ; Fk−1 )
ιs : (9.2)
p → ιs̃ p = kp (s̃, ., . . . , .)

as the interior product, i.e. the pointwise (i.e. independently in every fiber) contraction
in the first entry of p by s̃ := (dπ)−1 (s) ∈ C β (Σ; Ns ), but with a factor k. For u ∈
C β (M ; Eu ) we define
(
C β (Σ; Fk ) → C β (Σ; Fk+1 )
u :
p → ũ∗  p = Sym (ũ∗ ⊗ p)

as the symmetric point-wise tensor product of p by ũ∗ := ω (dA) (u, dπ (.)) ∈ C β (Σ; Ns∗ ).

Lemma 9.2 (Weyl algebra). For s ∈ C β (M ; Es ), u ∈ C β (M ; Eu ) and k ∈ N we have

[ιs , u] = ω (.) (dA) (s, u) IdFk : C β (Σ; Fk ) → C β (Σ; Fk ) (9.3)

where ω is the frequency function (2.7) on Σ and IdFk is the identity operator in fibers.

Proof. This is point wise relation so we consider a point ρ ∈ Σ with frequency ω = ω (ρ) and
(2.7)
1635 the vector space Ns (ρ) defined in (4.25). We use the isomorphism Polk (Ns ) ≡P Symk (Ns∗ ⊗ . . . ⊗ Ns∗ )
to derive (9.3) that goes as follows. If (ei )i is a basis of Ns (ρ), we write s̃ = di=1 σi ei ∈ Ns (ρ)
with components σ = (σ
i )i ∈ Rd . For p ∈ Symk (Ns∗ ⊗ . . . ⊗ Ns∗ ), we associate a degree k
polynomial P ∈ Polk Rd by P (σ) := p (s̃, s̃, . . . s̃). We define e∗i  p := Sym (e∗i ⊗ p) and
 (∂σi P ) (σ). Then the “Weyl algebra relation
observe that (e∗i  p) (s̃) = σi P (σ) and (ιei p) (s̃) =
1640 ” on polynomials [∂σi , σj ] = δi=j Id gives ιei , e∗j  = δi=j Id hence for s̃ ∈ Ns , ũ∗ ∈ Ns∗ we have
[ιs̃ , ũ∗ ] = ũ∗ (s̃) Id. With (9.1), we get (9.3).

Proof of theorem 1.17: Let K ∈ N. We will use T̃[0,K] to restrict to the finite rank bundle
of homogeneous polynomials of degree less than K. For k < K, let T̃k the projector on
homogeneous polynomials of degree k. We have
h    i      
OpΣ ιs T̃[0,K] , OpΣ (u) T̃[0,K] OpΣ T̃k ≈ OpΣ [ιs , u] T̃k = OpΣ ω (.) (dA) (s, u) T̃k
(4.65) (9.3)

that gives (1.34). We have

     
etXF OpΣ ιs T̃k ≈ OpΣ etXF OpΣ ιs T̃k ≈ OpΣ etXF ιs T̃k


(4.56) (4.65)
   
tXF
OpΣ ιdϕ s T̃k OpΣ etXF


= OpΣ ιdϕ s e T̃k
t ≈ t
(4.42,4.65)
 
≈ OpΣ ιdϕt s T̃k etXF ,
(4.56)

giving (1.35) and similarly to get (1.36).

A General notations used in this paper

• Dual ∗. If E is a finite dimensional vector space and E ∗ its dual space, x ∈ E, ξ ∈ E ∗ we
1645 denote the duality by ⟨ξ|x⟩ ∈ R. If E1 , E2 are vector spaces and A : E1 → E2 is a linear

75
 map, we will denote A∗ : E2∗ → E1∗ the induced dual map on the dual spaces defined by
⟨A∗ ξ|x⟩ = ⟨ξ|Ax⟩, ∀x ∈ E1 , ξ ∈ E2∗ .
• Bilinear map. Let Ω : E × E → R be a bilinear map on a vector space E. It defines a
linear map Ω̌ : E → E ∗ by
Ω̌ (x) (.) = Ω (x, .) . (A.1)

1650 – If Ω̌ is invertible (Ω is said to be non degenerated), it induces a bilinear form
on

E ∗ denoted Ω−1 : E ∗ × E ∗ → R and defined by Ω−1 (ξ1 , ξ2 ) := Ω Ω̌−1 ξ1 , Ω̌−1 ξ2 =
ξ1 Ω̌−1 ξ2 .

– The bilinear map Ω : E × E → R is anti-symmetric iff

Ω̌∗ = −Ω̌. (A.2)
The bilinear map Ω : E × E → R is symmetric iff Ω̌∗ = Ω̌.

1655 • Adjoint †. Suppose A : E1 → E2 is a linear map and Ω1 (respect. Ω2 ) is a non

degenerated bilinear form on E1 (respect. E2 ). The Ω-adjoint of A is A†Ω : E2 → E1
defined by
A†Ω = Ω̌−1 ∗
1 A Ω̌2 (A.3)
equivalently given by
Ω1 A†Ω x1 , x2 = Ω2 (x1 , Ax2 ) , ∀x1 ∈ E1 , ∀x2 ∈ E2 .

In this paper bilinear forms will be either a symplectic form Ω (., .) or a metric g (., .).
1660 • Pull-back ◦. If f : M → N is a smooth map between two manifolds,
– We denote f ◦ the pull back operator
(
C ∞ (N ) → C ∞ (M )
◦
f : (A.4)
u →u◦f
◦
and if f is a diffeomorphism, we denote f −◦ := (f −1 ) the push forward operator,
so that f −◦ f ◦ = Id.
– We denote df : T M → T N the differential of f that is a linear bundle map and
1665 denote (df )∗ : T ∗ N → T ∗ M its dual.

B More information about flows

B.1 Transfer operator
◦
Suppose that X is a vector field on M and (ϕt ) = etX is the pull back operator defined
in (2.2). ForR any smooth measure∞dµ on M , the L (M,
2
dµ) scalar product is defined by
◦† ◦
1670 ⟨v|u⟩L2 := M uvdµ with u, v ∈ Cc (M ). Then the L -adjoint (ϕt ) of (ϕt ) , defined by
2
◦† ◦
⟨v| (ϕt ) u⟩L2 = ⟨(ϕt ) v|u⟩L2 , ∀u, v ∈ Cc∞ (M ) is given by

◦†
◦ †
ϕt = det dϕ−t . ϕ−t = etX , with X † = −X − divµ X.

◦†
(ϕt ) is called the Ruelle-Perron-Frobenius operator or transfer operator. Observe that
◦†
(ϕt ) pushes forward probability distributions because
Z Z
t ◦† t ◦† t ◦




ϕ udµ = ⟨1| ϕ u⟩L2 = ⟨ ϕ 1 |u⟩L2 = udµ.
M | {z } M
1
◦†
In particular the evolution of Dirac measures gives (ϕt ) δm = δϕt (m) and is equivalent to
1675 evolution of points under the flow map ϕt .

76
 F (m) etXF
etXF u X˜F
u
ϕ̃tF
X
m ϕt (m) M
ϕt

Figure B.1: pull back operator etXF

B.2 More general pull back operators etXF

Suppose that X is a smooth vector field on M and F → M is a smooth vector bundle over M .
Suppose that XF : C ∞ (M ; F ) → C ∞ (M ; F ) is a derivation over X fulfilling (2.4).

B.2.1 Expression in local frames

1680 For a general vector bundle F of rank r, with
Pr respect to a local frame (e1 , . . . er ), a section
u ∈ C ∞ (M ; F ) is expressed as u (m) = j=1 uj (m) ej (m) with m ∈ M and components
uj (m) ∈ C. We introduce a matrix of potential functions Vk,j (m) ∈ C defined by
X
(XF ek ) (m) = Vk,j (m) ej (m) . (B.1)
j

Then the operator XF in (2.3) is expressed as

r r
!
X X
(XF u) (m) = (Xuj ) (m) + Vk,j (m) uk (m) ej (m) . (B.2)
(2.4)
j=1 k=1

We see that the expression (B.2) generalizes (2.5).

1685 B.2.2 Expression of etXF from a flow ϕ̃tF on F

An equivalent way to present the operator XF in (2.4), is to consider a vector field X̃F on F ,
a lift of X, i.e. (dπ) X̃F = X, with the projection π : F → M , and such that the flow map
generated by X̃F
ϕ̃tF : F → F
 ◦
i.e. ϕ̃tF = etX̃F : S (F ) → S (F ), is a smooth linear bundle map over ϕt which means that for
1690 every m ∈ M, t ∈ R, ϕ̃tF (m) : F (m) → F (ϕt (m)) is a linear map. See Figure B.1. This defines
a group of pull back operators with generator XF acting on a section u ∈ C ∞ (M ; F ) of
the vector bundle F :
etXF : C ∞ (M ; F ) → C ∞ (M ; F ) , t ∈ R.
by
etXF u := ϕ̃−t u ◦ ϕt . (B.3)


F

We get etXF (f u) = ϕ̃−t etX f (u ◦ ϕt ) = etX f etXF u and deduce (2.4) by derivation w.r.t.





F
1695 t. Conversely the flow map ϕ̃tF and its generator X̃F is defined from XF by (B.3) and use of
any section u.

77
 C Bargmann transform and Metaplectic operators
In this section we collect results that concerns the Bargmann transform on a vector space E,
1700 the quantization of affine and linear symplectic map, called Heisenberg group and metaplectic
operators. All these definitions and results are well known and important in many fields of
mathematics and physics. They are at the core of micro-local analysis. We present these
definitions and results in a form that is adapted to our paper that relies strongly on them.
References are [20],[27],[30, appendix 4.4],[16, chap.3, chap.4].

1705 C.1 Weyl Heisenberg group

Let E be a real vector space, n = dimE. For x ∈ E, ξ ∈ E ∗ we denote ⟨ξ|x⟩ ∈ R the duality.
Let dx ∈ |Λn (E)| be a density on E let dξ be the induced density on E ∗ . Let us denote the
Fourier transform (
S (E) → S (E ∗ )
F:
→ v (ξ) = (2π)1n/2 e−i⟨ξ|x⟩ u (x) dx
R
u

Its L2 -adjoint is (
S (E ∗ ) → S (E)
F† : 1
R
v → u (x) = (2π)n/2
ei⟨ξ|x⟩ v (ξ) dξ

1710 and we have that F † F = IdS(E) hence F −1 = F † .

For x ∈ E, ξ ∈ E ∗ we define the translation maps
( (
E →E E∗ → E∗
Tx : , Tξ :
y →y+x ξ′ → ξ′ + ξ

And denote Tx−◦ , Tξ−◦ the push forward operators. We will denote

T̂x,ξ := Tx−◦ F −1 Tξ−◦ F (C.1)

Lemma C.1. On S (E), one has

F −1 Tξ−◦ F Tx−◦ = ei⟨ξ|x⟩ Tx−◦ F −1 Tξ−◦ F . (C.2)

′
T̂x,ξ T̂x′ ,ξ′ = ei⟨ξ|x ⟩ T̂x+x′ ,ξ+ξ′ .
In particular T̂x,ξ is a unitary operator in L2 (E, dx) and
†
−1
T̂x,ξ = T̂x,ξ = T̂−(x,ξ) eiξx . (C.3)

Remark C.2. The family of operators T̂x,ξ with (x, ξ) ∈ T ∗ E = E⊕E ∗ are called the Schrödinger
1715 unitary representation of the Weyl Heisenberg group.
Proof. Let x ∈ E. Then
1 ′ 1 ′
(Fδx ) (ξ ′ ) = e−iξ x , Tξ−◦ Fδx (ξ ′ ) = e−i(ξ −ξ)x ,


n/2 n/2
(2π) (2π)
Z
1 ′ ′
−1
Tξ−◦ Fδx ′
eiξ x e−i(ξ −ξ)x dξ ′ = eiξx δx (x′ )


F (x ) =
(2π)n
Hence F −1 Tξ−◦ Fδx = eiξx δx . Hence
′
Tx−◦ F −1 Tξ−◦ F δx′ = eiξx δx+x′ ,

78
 ′
F −1 Tξ−◦ F Tx−◦ δx′ = eiξ(x+x ) δx+x′ = ei⟨ξ|x⟩ Tx−◦ F −1 Tξ−◦ F δx′ .

Then

T̂x,ξ T̂x′ ,ξ′ = Tx−◦ F −1 Tξ−◦ F Tx−◦ F −1 Tξ−◦

′ ′ F
′
= ei⟨ξ|x ⟩ Tx−◦ Tx−◦ F −1 Tξ−◦ F F −1 Tξ−◦



′ ′ F
′
= ei⟨ξ|x ⟩ T̂x+x′ ,ξ+ξ′ .

1720

C.2 Bargmann Transform on a Euclidean vector space (E, g)

Let E be a vector space with an Euclidean metric g. We have the canonical identification

T ∗E = E ⊕ E ∗,

with the canonical symplectic form Ω on E ⊕ E ∗ given by: for (x1 , ξ1 ) , (x2 , ξ2 ) ∈ E ⊕ E ∗ ,

Ω ((x1 , ξ1 ) , (x2 , ξ2 )) = ⟨ξ1 |x2 ⟩ − ⟨ξ2 |x1 ⟩. (C.4)

We denote dx the density on E induced by g.

Definition C.3. A Gaussian wave packet in vertical gauge (V) with parameters
(V)
(x, ξ) ∈ E ⊕ E ∗ , is the function φx,ξ ∈ S (E) given by: for y ∈ E,

(V) n 1 2
φx,ξ (y) := π − 4 ei⟨ξ|y−x⟩ e− 2 ∥y−x∥g (C.5)

Similarly a Gaussian wave packet in radial gauge (R) is

(R) i (V) n 1 1 2
φx,ξ (y) := e 2 ⟨ξ|x⟩ φx,ξ (y) == π − 4 ei⟨ξ|y− 2 x⟩ e− 2 ∥y−x∥g (C.6)
1725

Remark C.4. We have

• The terms ’vertical gauge’ or ’radial gauge’ come from a geometrical construction of these
waves packets using prequantization that we do not describe here. But we can explain
as follows. The wave packet is seen as a map φ(V ) : (x, ξ) ∈ E ⊕ E ∗ → H = L2 (E) that
induces a map to the projective space φ(V ) : E ⊕ E ∗ → P (H) (i.e. ignoring phase) and

1730

a pull back line bundle L → E ⊕ E ∗ with Hermitian connection (Levi Civita). The map
φ(V ) can be seen as a trivialization of this line bundle L. Its covariant derivative is given
by Dφ(V ) = P dφ(V ) = η (V ) φ(V ) , with projector onto the fiber P = φ(V ) ⟨φ(V ) |.⟩ hence
connection one-form
η (V ) = ⟨φ(V ) |dφ(V ) ⟩ = −iξdx.
(C.5)

1735 We observe that the kernel of this one-form is along x = cste on phase space E ⊕ E ∗ ,
i.e. ’vertical’. Similarly for the wave packet (C.6), considered as another trivialization,
we compute the connection one-form
i
η (R) = ⟨φ(R) |dφ(R) ⟩ = (−ξdx + xdξ) ,
(C.6) 2

for which the kernel at point (x, ξ) contains the ’radial’ vector (x, ξ).
(V) (R)
• φx,ξ = φx,ξ = 1 with dx the density on E associated to the metric g.
L2 (E,dx) L2 (E,dx)

79
1740 • We have
(V) (V)
φx,ξ = T̂x,ξ φ0,0 (C.7)
(C.1)

Definition C.5. The Bargmann transform in vertical gauge is the continuous operator
(
S (E ⊕ E ∗ )
S (E) → 
BE,(V) : (V)
 (C.8)
u → (x, ξ) → ⟨φx,ξ |u⟩L (E;dx) .
2

1 ∗
Lemma C.6. With respect to the density (2π) n dxdξ on E ⊕ E , the Bargmann transform
 
BE,(V) : L2 (E; dx) → L2 E ⊕ E ∗ ; (2π)1
n dxdξ is an isometry. Its L2 -adjoint operator
 
†
B(V) : L2 E ⊕ E ∗ ; (2π)
1
n dxdξ → L2 (E; dx) is given by
Z

†

(V) dxdξ
B(V) v (y) = φx,ξ (y) v (x, ξ) (C.9)
E⊕E ∗ (2π)n

and satisfies the resolution of identity

Z
† dxdξ
IdL2 (E) = B(V) B(V) = π (x, ξ) (C.10)
E⊕E ∗ (2π)n
(V) (V)
with π (x, ξ) := φx,ξ ⟨φx,ξ |.⟩ that is the rank one orthogonal projector in L2 (E) onto
(V)
Cφx,ξ .

Proof. We first check (C.9). For u ∈ S (E) and v ∈ S (E ⊕ E ∗ ) we have

Z
† (V) dxdξ
⟨u|B(V) v⟩L2 = ⟨B(V) u|v⟩L2 = ⟨φx,ξ |u⟩L2 v (x, ξ)
E⊕E ∗ (2π)n
Z Z
(V) dxdξ
= u (y) φx,ξ (y) v (x, ξ) dy.
E E⊕E ∗ (2π)n
2
Now we check that B(V) u = ∥u∥2 that is equivalent to (C.10). We have
Z 2 dxdξ
2 (V)
B(V) u = ⟨φx,ξ |u⟩L2
E⊕E ∗ (2π)n
Z
n 1 2 ′ 1 ′ 2 dxdξ
= π− 2 e−i⟨ξ|y−x⟩ e− 2 ∥y−x∥g u (y) dyei⟨ξ|y −x⟩ e− 2 ∥y −x∥g u (y ′ )dy ′
(2π)n

But Z
′
ei⟨ξ|y −y⟩ dξ = (2π)n δ (y ′ − y)
Z
2 n
e−∥y∥ dy = π 2
E
1745 hence Z Z
2 −n −∥y−x∥2g 2
B(V) u =π 2 e |u (y)| dydx = |u (y)|2 dy = ∥u∥2 .

80
 (R) (V) †
By using φx,ξ in radial gauge instead of φx,ξ , we can define similarly operators B(R) , B(R) .
Let
†
P(V) := B(V) B(V) : S (E ⊕ E ∗ ) → S (E ⊕ E ∗ ) (C.11)
†
P(R) := B(R) B(R) : S (E ⊕ E ∗ ) → S (E ⊕ E ∗ ) (C.12)
be the orthogonal projector in L2 (E ⊕ E ∗ ) onto their closed images Im B(V) , Im B(R) re-



1750

spectively, called the Bergman projector.

The metric g on E induces a canonical metric on E ⊕ E ∗ denoted g := g ⊕ g −1 .

Lemma C.7. The Schwartz kernel of P(V) , P(R) are given by

 
i ′ 1 2
⟨δx′ ,ξ′ |P(V) δx,ξ ⟩ = exp ′ ′ ′
(ξ + ξ) (x − x) − ∥(x , ξ ) − (x, ξ)∥g⊕g−1 (C.13)
2 4
 
i 1 2
′ ′ ′ ′
⟨δx′ ,ξ′ |P(R) δx,ξ ⟩ = exp − Ω ((x , ξ ) , (x, ξ)) − ∥(x , ξ ) − (x, ξ)∥g⊕g−1 (C.14)
2 4

Proof. We have
Z
(V) (V) −n ′ ′ 1 ′ 2 1 2
⟨δx′ ,ξ′ |P(V) δx,ξ ⟩ = ⟨φx′ ,ξ′ |φx,ξ ⟩ e−i⟨ξ |y−x ⟩ e− 2 ∥y−x ∥g ei⟨ξ|y−x⟩ e− 2 ∥y−x∥g dy
=π 2

Z
−n i(⟨ξ ′ |x′ ⟩−⟨ξ|x⟩)− 12 ∥x′ ∥2g − 12 ∥x∥2g 2
=π e
2 e−∥y∥g +⟨b|y⟩ dy

with b = ǧ (x′ + x) + i (ξ − ξ ′ ) ∈ E ∗ . We use the Gaussian integral

Z
2
 π  n2 1 2
e−α∥y∥g +⟨b|y⟩ dy = e 4α ∥b∥g−1 , (C.15)
E α
with α = 1, giving
′ 2 1 2 1 2 ′ 2
⟨δx′ ,ξ′ |P(V) δx,ξ ⟩ = ei(⟨ξ |x ⟩−⟨ξ|x⟩)− 2 ∥x ∥ − 2 ∥x∥ e 4 (∥x +x∥ −∥ξ−ξ ∥ +2i⟨ξ−ξ |x +x⟩)
′ ′ 1 ′ ′ ′

1 1 2 ′ 2
= ei(⟨ξ |x ⟩−⟨ξ|x⟩+ 2 ⟨ξ−ξ |x +x⟩) e− 4 (∥x −x∥ +∥ξ−ξ ∥ )
′ ′ ′ ′ ′

1 ′ ′ 1 ′ ′ 2
= ei 2 ⟨ξ +ξ|x −x⟩ e− 4 ∥(x ,ξ )−(x,ξ)∥g+g−1
We have
1 ′ ′
⟨δx′ ,ξ′ |P(R) δx,ξ ⟩ = ei 2 (⟨ξ|x⟩−⟨ξ |x ⟩) ⟨δx′ ,ξ′ |P(V) δx,ξ ⟩
1 ′ ′ 1 ′ ′ 2
= ei 2 (−⟨ξ |x⟩+⟨ξ|x ⟩) e− 4 ∥(x ,ξ )−(x,ξ)∥g+g−1
1 ′ ′ 1 ′ ′ 2
= e−i 2 Ω((x ,ξ ),(x,ξ)) e− 4 ∥(x ,ξ )−(x,ξ)∥g+g−1
1755

Lemma C.8. Let π : E ⊕ E ∗ → E denotes the projector on the first component, and π ◦ :
S (E) → S (E ⊕ E ∗ ) the pull back operator. We have
n
B(V) = π 4 P(V) π ◦

Proof. We compute the Schwartz kernel of both sides for y ∈ E, (x′ , ξ ′ ) ∈ E ⊕ E ∗ . We have
(V)
⟨δx′ ,ξ′ |B(V) δy ⟩ = φx′ ,ξ′ (y) and
Z
i 1 2
+∥ξ−ξ ′ ∥2 ) dxdξ
e 2 (ξ +ξ)(x −x)− 4 (∥x −x∥
′ ′ ′
◦
⟨δx′ ,ξ′ |P(V) π δy ⟩ = δ (x − y)
(2π)n
Z
−n ξ (x −y)− 14 ∥x′ −y∥2 − 14 ∥ξ ′ ∥2
i ′ ′ i ′ 1 ′ 1 2
= (2π) e 2 e 2 ξ(x −y)+ 2 ξ.ξ − 4 ∥ξ∥ dξ

81
 We use the Gaussian integral (C.15) with α = 14 , b = 2i (x′ − y) + 21 ξ ′ ,

1 1 ′ 2 2
 i
∥b∥2 = ∥ξ ∥ − ∥x′ − y∥ + (x′ − y) ξ ′
4α 4 2

 π  n2
◦ −n ξ (x −y)− 14 ∥x′ −y∥2 − 14 ∥ξ ′ ∥2
i ′ ′ 1 2
⟨δx′ ,ξ′ |P(V) π δy ⟩ = (2π) e 2 e 4α ∥b∥
α
n
−n n
−n iξ ′ (x′ −y)− 12 ∥x′ −y∥2
= (4π) (2π) 2 π π
4 4 e
n (V)
= π − 4 φx′ ,ξ′ (y)

C.3 Linear map ϕ : (E2 , g2 ) → (E1 , g1 )

1760 Suppose ϕ : (E2 , g2 ) → (E1 , g1 ) is a linear invertible map between two metric spaces. Let

Φ := ϕ−1 ⊕ ϕ∗ : E1 ⊕ E1∗ → E2 ⊕ E2∗

be the induced map (i.e. pull back of differential forms Φ : T ∗ E1 → T ∗ E2 ). We have defined in
(C.8), BE1 : S (E1 ) → S (E1 ⊕ E1∗ ) and BE2 : S (E2 ) → S (E2 ⊕ E2∗ ) where BEj is either B(V) or
B(R) .

Lemma C.9. Suppose ϕ : (E2 , g2 ) → (E1 , g1 ) is a linear invertible map between two
Euclidean vector spaces. We have that

ϕ◦ = Υ (ϕ) BE† 2 Φ−◦ BE1 (C.16)

with   
1 1/2
−1 † −1
(C.17)


Υ (ϕ) := det Id + ϕ ϕ
2
† †
where (ϕ−1 ) is the metric-adjoint of ϕ−1 defined by ⟨x1 | (ϕ−1 ) x2 ⟩g1 = ⟨ϕ−1 x1 |x2 ⟩g2 for
any x1 ∈ E1 , x2 ∈ E2 .

1765 Remark C.10. If ϕ is an isometry then ϕ† ϕ = Id and Υ (ϕ) = 1.

Proof. We compute the Schwartz kernels. For y ′ ∈ E1 , z ′ ∈ E2 ,

⟨δz′ |BE2 Φ−◦ BE1 δy′ ⟩ = ⟨BE2 δz′ |Φ−◦ BE1 δy′ ⟩

For y ∈ E1 , η ∈ E1∗ ,
n ′ 1 ′ 2
− 4 −i⟨η|y −y⟩ − 2 ∥y −y∥g1
(BE1 δy′ ) (y, η) = ⟨φ(1)
y,η |δy ′ ⟩ = π e e

For z ∈ E2 , ξ ∈ E2∗ ,
n ∗−1 ′ 1 ′ 2
Φ−◦ BE1 δy′ (z, ξ) = π − 4 e−i⟨ϕ ξ|y −ϕz⟩ e− 2 ∥y −ϕz∥g1

n ′ 1 ′ 2
(BE2 δz′ ) (z, ξ) = π − 4 e−i⟨ξ|z −z⟩ e− 2 ∥z −z∥g2
1770 Hence
Z
−n ′ 1 ′ 2 ∗−1 ξ|y ′ −ϕz⟩ 1 ′ 2 dzdξ
−◦
⟨BE2 δz′ |Φ BE1 δy′ ⟩ = π 2 ei⟨ξ|z −z⟩ e− 2 ∥z −z∥g2 e−i⟨ϕ e− 2 ∥y −ϕz∥g1
(2π)n

82
  
∗−1
We do the symplectic change of variables (z, ξ) ∈ E2 ⊕E2∗ →(y, η) =Φ −1
(z, ξ) = ϕ (z) , ϕ ξ ∈
(E1 ⊕ E1∗ ), i.e. ξ = ϕ η, z = ϕ y, and get
∗ −1

2
Z
− 1 ∥z ′ −ϕ−1 y ∥ 1 ′ 2 dydη
† −◦ −n ∗ ′ −1 ′
g2 −i⟨η|y −y⟩ − 2 ∥y −y∥g1
⟨δz′ |BE2 Φ BE1 δy′ ⟩ = π 2 ei⟨ϕ η|z −ϕ y⟩ e 2 e e
(2π)n
2
Z
n ′ ′ − 1 ∥ϕ−1 (ϕz ′ −y)∥ 1 ′ 2 dydη
g2 − 2 ∥y −y∥g1
= π− 2 ei⟨η|ϕz −y ⟩ e 2 e
(2π)n

We have Z
′ ′ dη
ei⟨η|ϕz −y ⟩ = δ (ϕz ′ − y ′ ) = ⟨δz′ |ϕ◦ δy′ ⟩
(2π)n
hence
2
Z
−n − 12 ∥ϕ−1 (y ′ −y)∥ 1 ′ 2
⟨δz′ |BE† 2 Φ−◦ BE1 δy′ ⟩ ◦
= ⟨δz′ |ϕ δy′ ⟩π 2 e g2
e− 2 ∥y −y∥g1 dy
2
Z
−n − 12 ∥ϕ−1 Y ∥ 1 2
◦
= ⟨δz′ |ϕ δy′ ⟩π 2 e g2
e− 2 ∥Y ∥g1 dY
ZE  †

◦ −n − 12 ⟨Y | Id+(ϕ−1 ) ϕ−1 Y ⟩
= ⟨δ |ϕ δ ⟩π
z′ y′ 2 e dY
 1/2
n (2π)n
= ⟨δz′ |ϕ◦ δy′ ⟩π − 2   
det Id + (ϕ−1 )† ϕ−1
= ⟨δz′ |ϕ◦ δy′ ⟩Υ (ϕ)−1

Giving (C.16).
Later we will use the following Lemma, where the determinant is measured with respect to
local densities.
Lemma C.11. With Υ (ϕ) defined in (C.17), we have

Υ ϕ−1 = |detϕ| Υ (ϕ) . (C.18)

⊥
If ϕ = ϕ1 ⊕ ϕ2 on E = E1 ⊕ E2 , then

Υ (ϕ) = Υ (ϕ1 ) Υ (ϕ2 ) . (C.19)

1775

Proof. Notice that det ϕ† ϕ = |detϕ|2 . Then

   
1

−1 1 †
2
2 −1† −1 †
ϕ ϕ + Id = |detϕ|−2 Υ ϕ−1



Υ (ϕ) = det Id + ϕ ϕ = det ϕ ϕ det
2 2

83
 Corollary C.12. If E is endowed with two metrics g1 , g2 then
†
IdS(E) = Υg2 ,g1 Bg2g2 Bg1

with   1/2 n  1/2

1 Y 1 + Ij
Id + Id† (C.20)


Υg2 ,g1 = det =
2 j=1
2

(Ij )j ∈ Rn are inertial moments of g2 with respect to g1 i.e. in suitable coordinates,

g1 = j dyj2 , g2 = j Ij dyj2 . Consequently
P P

†
Pg1 ,g2 := Υg2 ,g1 Bg1 Bg2g2 : S (E ⊕ E ∗ ) → S (E ⊕ E ∗ )

is a projector, with Im (Pg1 ,g2 ) = Im (Bg1 ), Ker (Pg1 ,g2 ) = (Im (Bg2 ))⊥ .

Proof. We apply Lemma C.9 with ϕ = Id : (E, g2 ) → (E, g1 ) and Φ = Id. We have

†
⟨u| ϕ−1 v⟩g1 = ⟨ϕ−1 u|v⟩g2 = ⟨u|v⟩g2

In coordinates such that g1 = j dyj2 , g2 = j Ij dyj2 ,

P P

†
(Id)†j,k = ϕ−1 j,k
= Ij δj=k

1780 giving (C.20).

C.4 Compatible triple g, Ω, J

Let us recall the definition of a compatible triple g, Ω, J. Let F be a vector space endowed with
a symplectic bilinear form Ω and an Euclidean metric g. Let

J := ǧ −1 Ω̌ :F →F (C.21)

with Ω̌ : F → F ∗ , ǧ : F → F ∗ defined in Section A.

Definition C.13. We say that Ω, g are compatible structures on F if J := ǧ −1 Ω̌ : F → F

is an almost complex structure on F , i.e.

J 2 = −Id (C.22)

.
1785

Proposition C.14. From (C.21), we have that for any u, v ∈ F

Ω (u, v) = g (Ju, v) . (C.23)

If moreover (C.22) holds true, we have that for any u, v ∈ F ,

g (Ju, Jv) = g (u, v) , Ω (Ju, Jv) = Ω (u, v) . (C.24)

Proof. We have
g (Ju, v) = ⟨ǧǧ −1 Ω̌u|v⟩ = Ω (u, v) .
(C.21)

84
 Suppose J 2 = ǧ −1 Ω̌ǧ −1 Ω̌ = −Id hence ǧ = −Ω̌ǧ −1 Ω̌. Then

g (Ju, Jv) = ⟨ǧǧ −1 Ω̌u|ǧ −1 Ω̌v⟩ = ⟨−Ω̌ǧ −1 Ω̌u|v⟩ = g (u, v)

Ω (Ju, Jv) = ⟨Ω̌ǧ −1 Ω̌u|ǧ −1 Ω̌v⟩ = ⟨−Ω̌ǧ −1 Ω̌ǧ −1 Ω̌u|v⟩ = Ω (u, v)

1790

Example C.15. Let (E, g) be a vector space with Euclidean metric g. On E⊕E ∗ , the canonical
symplectic form Ω in (C.4), is compatible with the induced metric g := g ⊕ g −1 .

Lemma C.16. Let (F, Ω, g) be a vector space with a symplectic structure Ω and com-
patible metric g. If E ⊂ F is a linear Lagrangian subspace, let E ⊥g be the orthogonal
subspace with respect to g. Then E ⊥g = J (E) is also Lagrangian and E = J E ⊥g .

Hence we have an isomorphism Ψ that is symplectic (for Ω) and isometric (for g)

(
F = E ⊕ E ⊥g → E ⊕ E ∗
Ψ:
(C.25)
(v, u) 7→ v, Ω̌u

where E ⊕ E ∗ is endowed with the canonical symplectic form (C.4) and with the metric
g = g ⊕ g −1 .

Proof. Let v1 , v2 ∈ E Lagrangian. Then

g (Jv1 , v2 ) = Ω (Jv1 , Jv2 ) = Ω (v1 , v2 ) = 0.

(C.23) (C.24) (Lag.)

Considering dimensions, dimE = dimE ⊥g = 21 dimF , we deduce E ⊥g = J (E) and E = J E ⊥g ,

1795

since J 2 = −Id. From compatibility J is a symplectic map hence E ⊥g is Lagrangian as well.

C.5 Bergman projector on a symplectic vector space (F, Ω, g)

Let (F, Ω, g) be a vector space with a symplectic structure Ω and compatible metric g.

Definition C.17. Let PF : S (F ) → S (F ) be the operator defined by its Schwartz

kernel: for any ρ, ρ′ ∈ F ,
 
i 1 ′ 2
′
⟨δρ′ |PF δρ ⟩ = exp − Ω (ρ , ρ) − ∥ρ − ρ∥g . (C.26)
2 4

 
dρ
Lemma C.18. PF is an orthogonal projector in L2 F, (2π) n , called the Bergman pro-

jector.
1800

Proof. We choose any E ⊂ F Lagrangian linear subspace. Then PF = PF,(R) is the Bergman
projector in radial gauge defined in (C.14).

C.6 The metaplectic decomposition of F ⊕ F ∗

Let (F, Ω) be a linear symplectic space. Let us denote Ω the canonical symplectic form on
1805 F ⊕ F ∗ as in (C.4). Recall Ω̌ : F → F ∗ defined in Section A and that Ω̌∗ = −Ω̌.

85
 F∗ (x, ξ) K
ξ
N Ω̌(ν)

πN Ω̌∗ (ζ) πK
ζ x ν F

⊥
Figure C.1: Picture for the orthogonal decomposition F ⊕ F ∗ = K ⊕ N in (C.28).

Lemma C.19 (Metaplectic decomposition). Let (F, Ω) be a linear symplectic space and

(C.27)



K := graph Ω̌ , N := graph −Ω̌ .

We have
⊥Ω
F ⊕ F ∗ = K ⊕ N. (C.28)
where the right hand side is an Ω-orthogonal decomposition into linear symplectic sub-
spaces. See Figure C.1. Explicitly for any (x, ξ) ∈ F ⊕ F ∗ , (x, ξ) = ν, Ω̌ (ν) +

ζ, −Ω̌ (ζ) ∈ K ⊕ N is given by

1
x + Ω̌−1 (ξ) (C.29)


ν=
2
1
x − Ω̌−1 (ξ) (C.30)


ζ=
2

Remark C.20. The metaplectic decomposition (C.28) is central in this paper. It has also been
used in [11, Prop.6] and in [16, Prop.2.2.9].
Proof. If (x, ξ) = x, Ω̌ (x) ∈ graph Ω̌ and (x′ , ξ ′ ) = x′ , −Ω̌ (x′ ) ∈ graph −Ω̌ then

Ω ((x, ξ) , (x′ , ξ ′ )) = ⟨ξ|x′ ⟩ − ⟨x|ξ ′ ⟩ = ⟨Ω̌ (x) |x′ ⟩ + ⟨x|Ω̌ (x′ )⟩

= Ω (x, x′ ) + Ω (x′ , x) = 0

Moreover if (x, ξ) = x, Ω̌ (x) = x, −Ω̌ (x) ∈ graph Ω̌ ∩ graph −Ω̌ then Ω̌ (x) = −Ω̌ (x)

hence Ω̌ (x) = 0 hence x = 0. This orthogonality and transversality implies that graph Ω̌ , graph −Ω̌



1810

are symplectic. Let us prove (C.29). We write

( (
ν = 21 x + Ω̌−1 (ξ)


x =ν+ζ
⇔
ζ = 21 x − Ω̌−1 (ξ)


ξ = Ω̌ (ν) − Ω̌ (ζ) = Ω̌ (ν − ζ)

Lemma C.21. Let (F, Ω, g) be a vector space with a symplectic structure Ω and a com-
patible metric g. Then (C.28) is orthogonal for the metric g = g ⊕ g −1 as well. The
maps
(C.31)


πK : x, Ω̌ (x) ∈ (K, Ω, g) → x ∈ (F, 2Ω, 2g)
πN : x, Ω̌∗ (x) ∈ (N, Ω, g) → x ∈ (F, −2Ω, 2g)

are isomorphism for the respective bilinear forms.

86
 ⊥
Proof. K ⊕ N is also orthogonal for the metric g because

g ((x, ξ) , (x′ , ξ ′ )) = g (x, x′ ) + g −1 (ξ, ξ ′ ) = g (x, x′ ) + g −1 Ω̌ (x) , −Ω̌ (x′ )

= g (x, x′ ) − g −1 Ω̌ (x) , Ω̌ (x′ ) = 0

because g −1 Ω̌ (x) , Ω̌ (x′ ) = g (x, x′ ) from compatibility between Ω and g.

If (x, ξ) = x, Ω̌ (x) ∈ K and (x′ , ξ ′ ) = x′ , Ω̌ (x′ ) ∈ K then

Ω ((x, ξ) , (x′ , ξ ′ )) = ⟨ξ|x′ ⟩ − ⟨x|ξ ′ ⟩ = ⟨Ω̌ (x) |x′ ⟩ − ⟨x|Ω̌ (x′ )⟩ = ⟨Ω̌ (x) |x′ ⟩ − ⟨Ω̌∗ (x) |x′ ⟩
= ⟨Ω̌ (x) |x′ ⟩ + ⟨Ω̌ (x) |x′ ⟩ = 2Ω (x, x′ )

g ((x, ξ) , (x′ , ξ ′ )) = g (x, x′ ) + g −1 (ξ, ξ ′ ) = g (x, x′ ) + g −1 Ω̌ (x) , Ω̌ (x′ ) = 2g (x, x′ )

1815 and a similar computation on N , if Ω̌ is replaced by −Ω̌.

Let BF : S (F ) → S (F ⊕ F ∗ ) the Bargman transform in radial Gauge. Let PF ⊕F ∗ := BF BF† .
Let PK , PN defined as in (C.26) for the respective spaces (K, ΩK , gK ), (N, ΩN , gN ) defined in
(C.27) with metrics induced from Ω, g. From (C.28) we have the natural identification

S (F ⊕ F ∗ ) = S (K ⊕ N ) = S (K) ⊗ S (N ) .

Lemma C.22. We have

PF ⊕F ∗ = PK ⊗ PN (C.32)

1820 Proof. Since ΩF ⊕F ∗ = ΩK ⊕ ΩN and g = gK ⊕ gN , (C.32) follows from the expression of the
kernels (C.14) and (C.26) that coincide for both sides.

C.7 Metaplectic decomposition of a linear symplectic map

In this section, let Φ : (F1 , Ω1 ) → (F2 , Ω2 ) a linear invertible symplectic map and gj a compatible
metric on Fj for j = 1, 2.
1825 Let us first give a definition of metaplectic operator that will be useful. The metaplectic
correction Υ (Φ) > 0 has been defined in (C.17). The reader may consult [40, section 13.2] for
a general reference about metaplectic operators.

Definition C.23. • We define the metaplectic operator (or quantization) of

Φ by
˜ (Φ) := (Υ (Φ))1/2 PF2 Φ−◦ PF1
Op : Im (PF1 ) → Im (PF2 ) (C.33)
with the metaplectic correction Υ (Φ) > 0 defined in (C.17).

• If E1 ⊂ F1 , E2 ⊂ F2 are Lagrangian linear subspace, and using the identification

(C.25), we define

Op (Φ) : = (Υ (Φ))1/2 BE† 2 Φ−◦ BE1 : S (E1 ) → S (E2 ) (C.34)

that is conjugated to Op
˜ (Φ)
/ImBE →ImBE2 since
1

˜ (Φ) = BE2 Op (Φ) B † .

Op (C.35)
E1

Remark C.24. We will show in Proposition C.28 that Op

˜ (Φ) : L2 (F1 ) → L2 (F2 ) is L2 −unitary
1830 (hence Op (Φ) : L (E1 ) → L (E2 ) is also unitary).
2 2

87
 Let
Φ̃ := Φ−1 ⊕ Φ∗ : F2 ⊕ F2∗ → F1 ⊕ F1∗ (C.36)
the induced map on cotangent spaces. Since Φ is symplectic we have that Φ̃ preserves the
decompositions (C.28), Fj ⊕ Fj∗ = Kj ⊕ Nj , with j = 1, 2 and we denote its components by

Φ̃ = ΦK ⊕ ΦN . (C.37)

Theorem C.25. [11][16]We have

 
BF1 Φ◦ B † = Op
˜ Φ̃ = Op
˜ (ΦK ) ⊗ Op
˜ (ΦN ) (C.38)



F2 : Im PF2 ⊕F2∗ → Im PF1 ⊕F1∗ .

1835 Proof. We have  

Υ Φ−1 Υ (Φ∗ ) = (Υ (Φ))2 . (C.39)


Υ Φ̃ =
(C.36,C.19) (C.18)

Hence

BF1 Φ◦ BF† 2 = Υ (Φ) PF1 ⊕F1∗ Φ̃−◦ PF2 ⊕F2∗

(C.16,,C.12)
  −1/2    
= Υ (Φ) Υ Φ̃ ˜ Φ̃ = Op
Op ˜ Φ̃ ,
(C.33) (C.39)

and
    1/2
˜
Op Φ̃ = Υ Φ̃ PF2 ⊕F2∗ Φ̃−◦ PF1 ⊕F1∗
(C.33)
   
† †
= (Υ (ΦK ) Υ (ΦN ))1/2 PK 1
Φ−◦
K PK2 ⊗ P Φ
N1 N
−◦
PN2
(C.37,C.32,C.19)
˜ (ΦK ) ⊗ Op
= Op ˜ (ΦN ) .
(C.33)

Remark C.26. Lemma C.25 is central to the analysis in this paper. It is a factorization formula
for the map Φ and somehow gives a square root of Φ◦ .

Proposition C.27. For any linear symplectic maps Φ2,1 : (F1 , Ω1 ) → (F2 , Ω2 ), Φ3,2 :
(F2 , Ω2 ) → (F3 , Ω3 ) with compatible metrics gj on Fj , j = 1, 2, 3, we have
˜ (Φ3,2 Φ2,1 ) = Op
Op ˜ (Φ3,2 ) Op
˜ (Φ2,1 ) . (C.40)

Proof. We have
˜ (Φ3,2 Φ2,1 ) ⊗ Op
˜ (Φ3,2 Φ2,1 ) = BF3 (Φ3,2 Φ2,1 )−◦ BF† 1



Op K N
(C.38)
  
−◦ † −◦ † −◦ †
= BF3 Φ−◦ Φ B
3,2 2,1 F1 = B Φ B
F3 3,2 F2 B Φ B
F2 2,1 F1
(C.10)
  

˜ (Φ3,2 ) Op ˜ (Φ2,1 ) ˜ (Φ3,2 ) Op ˜ (Φ2,1 )




= Op K K ⊗ Op N N
(C.38)

1840 Considering each factor gives (C.40).

˜ (Φ) : Im (PF1 ) → Im (PF2 ) defined in (C.33) is

Proposition C.28. The operator Op
L2 −unitary.

88
 Remark C.29. Since BEj : L2 (Ej ) → ImBEj = ImPFj is unitary this implies that Op (Φ) :
L2 (E1 ) → L2 (E2 ) is also unitary.
Proof. We have Υ (Id) = 1 hence Op
˜ (Φ) (Id) = P = Id/ImP . Also
(C.33)

Υ Φ−1 (C.41)


= |detΦ| Υ (Φ) = Υ (Φ) .
(C.18)

1845 We have
˜ Φ−1 Op
˜ (Φ) = Op
˜ (Id) = Id/ImP ,


Op
(C.40)

˜ (Φ)−1 = Op
so Op ˜ (Φ−1 ) and


1/2
−◦
˜ (Φ)† = (Υ (Φ))1/2 PΦ◦ P = Υ Φ−1
Op P Φ−1 P
(C.33) (C.41)
˜ (Φ)−1 .
˜ Φ−1 = Op


= Op
(C.33)

The next lemma considers a special case for the symplectic map Φ.

Lemma C.30. [16, proof of Prop 4.3.1 p.79]Let (E1 , g1 ),(E2 , g2 ) be Euclidean vector
spaces, and ϕ : E2 → E1 an invertible linear map. We denote ϕ◦ : S (E1 ) → S (E2 ) the
pull-back operator and Φ := ϕ−1 ⊕ ϕ∗ : E1 ⊕ E1∗ → E2 ⊕ E2∗ the induced symplectic map
˜ (Φ) in (C.33). We have
on cotangent spaces. Op (Φ) has been defined in (C.34) and Op

Op (Φ) = |detϕ|1/2 ϕ◦ . (C.42)

˜ (Φ) = |detϕ|1/2 BE2 ϕ◦ B †

Op (C.43)
E1

Proof. We have

Υ (Φ) = Υ (ϕ) Υ ϕ−1 = |detϕ| (Υ (ϕ))2 (C.44)

(C.19) (C.18)

Then

Op (Φ) = (Υ (Φ))1/2 BE† 2 Φ−◦ BE1 = (Υ (Φ))1/2 (Υ (ϕ))−1 ϕ◦ = |detϕ|1/2 ϕ◦

(C.34) (C.16) (C.44)

1850

C.8 Some useful decompositions

Here we give a relation that is simple here in the linear setting and is used in the main text
as an approximation, in the proof of Theorem 3.15. Let us denote F̃ = F ⊕ F ∗ = K ⊕ N
(C.28)
and exp : T F̃ → F̃ the exponential map. Now we restrict the base to K ⊂ F̃ and consider
1855 N ⊂ TK F̃ as a sub-bundle N → K of TK F̃ → K. We denote expN : N → F̃ the exponential
map that is an isomorphism.    
As before, we have the operators BF : S (F ) → S F̃ , BF† : S F̃ → S (F ), exp] ◦
N :
   
−1 ◦
^ : S (N ) → S F̃ .


S F̃ → S (N ), exp N
Let Φ : F → F be a linear symplectic map. We have Φ◦ : S (F ) → S (F ). The  map

1860 Φ̃ : F̃ → F̃ that can be seen as the bundle map Φ̃N : N → N over ΦK . Then Op ˜ Φ̃N :

89
 S (N ) → S (N ) is a bundle map over ΦK . The metaplectic correction (Υ (ΦK ))1/2 > 0 has been
defined in (C.17) and in fact we have Υ (ΦK ) = Υ (ΦN ) = Υ (Φ).

Lemma C.31. We have

 
−1 ◦
Φ◦ = BF† exp
^ 1/2 ˜
(C.45)

◦
(Υ (Φ )) Op Φ̃N expN BF .
N K
]

Equivalently
   
−1 ◦
˜ Φ̃ = PF exp
^ 1/2 ˜
(C.46)

◦
Op (Υ (Φ )) Op Φ̃N expN PF .
N K
]
(C.38)

Remark C.32. As the proof belows

 shows, the metaplectic correction (Υ (ΦK )) > 0 in (C.45)
1/2

1865 is due to the fact that Op

˜ Φ̃N over ΦK does not preserves the space ImPK .

Proof. We have
 
Φ◦ = BF† BF Φ◦ BF† BF = BF† Op ˜ (ΦK ) ⊗ Op
˜ (ΦN ) BF
(C.10) (C.38)
  
† 1/2 −◦ ˜
= BF (Υ (ΦK )) PK ΦK PK ⊗ Op (ΦN ) BF
(C.33)

We use that BF = PF BF = (PK ⊗ PN ) BF and similarly BF† = BF† (PK ⊗ PN ). Hence

(C.12,C.10) (C.32)
we can remove the operators PK and get
 
◦ † 1/2 −◦ ˜
Φ = BF (Υ (ΦK )) ΦK ⊗ Op (ΦN ) BF
−1
 
= (Υ (ΦK ))1/2 BF† exp
] ◦
N
˜ Φ̃N exp
Op ] ◦
N BF .

 
where the last line expresses that Op Φ̃N : S (N ) → S (N ) is a bundle map over ΦK . Using
˜
(C.38) and (C.12) we deduce (C.46).
   
In the next Lemma, we use the (twisted) operator exp g◦ : S F̃ → S T F̃ . We also use
         
the restriction operators r0 : S T F̃ → S F̃ , r/K : S T F̃ → S TK F̃ , rN : S TK F̃ →
1870 S (N ). We have
^−1 ◦


g◦ = exp
IdS (F̃ ) = r0 exp N
g◦ ,
rN r/K exp
where the second equality can be seen as a generalization of the first equality for the respective
decompositions F̃ = {0} ⊕ F̃ = K ⊕ N . More generally we have the following Lemma.

Lemma C.33. For the symplectic map Φ̃ = ΦK ⊕ ΦN : F̃ → F̃ , using Op ˜ (Φ) :

(C.37)
       
S T F̃ → S T F̃ and Op ˜ /K (Φ) : S TK F̃ → S TK F̃ as bundle map operators, we
have the decomposition

−1 ◦
˜ (Φ) exp ^ ˜ /K (Φ) r/K exp (C.47)


r0 Op g◦ = expN rN Op g◦ .

C.9 Taylor operators Tk on S (E)

1875 Let
Pn E an Euclidean vector space, n = dimE. If (e1 , . . . en ) is a basis of E, we write x =
i=1 xi ei ∈ E. A polynomial on E is P (x) := p (x1 , . . . xn ) with p ∈ C [x1 , . . . xn ] a polynomial

90
 on Rn . For α = (α1 , . . . , αn ) ∈ Nn , we write xα := xα1 1 . . . xαnn a monomial of degree |α| :=
α1 + . . . αn . For k ∈ N, we denote Polk (E) the space of homogeneous polynomials of degree k,
that is independent on the basis. P ∈ Polk (E) can be written P (x) = α∈Nk ,|α|=k Pα xα with
P
(α ) (α )
1880 components Pα ∈ C. We denote α! := α1 ! . . . αn ! and δ (α) := δ0 1 (x1 ) . . . δ0 n (xn ) the Dirac
(α′ )
distribution on E (with α-derivatives). We have that for
any α, α′ ∈ Nn , ⟨ α1′ ! δ0 |xα ⟩ = δα′ =α .
Hence the set (xα )α forms a basis of Pol (E) and ⟨δ (α) |.⟩ α is the dual basis. See Section 9 for
 
∗ ⊗k
comments about the equivalence Polk (E) ≡ Sym (E ) .

Definition C.34. For k ∈ N, let

Tk : S (E) → Polk (E) ⊂ S ′ (E)

be the projector onto Polk (E) with kernel ⊕k′ ̸=k Polk (E). We have

∀k, k ′ , Tk′ Tk = Tk δk′ =k ,

and explicitly, with respect to a basis of E,

X 1 (α)
Tk = xα ⟨ δ |.⟩. (C.48)
α! 0
α∈Nk ,|α|=k

1885 The next lemma shows how this relation is changed by a small norm operator if we insert
a truncation at large distance σ from the origin. We put χσ (ρ) = 1 for |ρ|g ≤ σ, χσ (ρ) = 0 for
|ρ|g > σ and Op (χσ ) := B † Mχσ B.

Lemma C.35. For k, k ′ ∈ N, we have ∀N, ∃CN > 0, ∀σ > 0,

Rσ := Op (χσ ) (Tk′ Op (χσ ) Tk − Tk δk′ =k ) Op (χσ )

satisfies
−N −N
⟨δρ′ |B † Rσ Bδρ ⟩ ≤ distg/σ2 (ρ′ , ρ) CN σ −N distg/σ2 (ρ, 0) ,

Consequently ∥Rσ ∥ ≤ CN σ −N for any N > 0, with CN > 0.

Proof. We have

⟨δρ′ |BRB † δρ ⟩ = χσ (ρ′ ) χσ (ρ) (C.49)

(C.48)
 
X
α′ 1 (α′ ) α 1 (α)
⟨φρ′ |x ⟩ ⟨ ′ δ0 |Op (χσ ) x ⟩ − δα′ =α ⟨ δ0 |φρ ⟩.
α! α!
α′ ∈Nk ,|α′ |=k′ ,α∈Nk ,|α|=k,
(C.50)

We have ∀N, ∃CN , ∀σ > 0,

1 (α′ ) †
⟨ δ |B Mχσ Bxα ⟩ − δα′ =α ≤ CN σ −N ,
α′ ! 0

1890 and due to truncations χσ (ρ′ ) χσ (ρ) in (C.49) we deduce that

−N −N
⟨δρ′ |B † RBδρ ⟩ ≤ distg/σ2 (ρ′ , ρ) CN σ −N distg/σ2 (ρ, 0) ,

and ∥Rσ ∥ ≤ CN σ −N from example (2.37).

91
 C.10 Analysis on T (E ⊕ E ∗ )
In this section we “lift” the analysis from E ⊕E ∗ to T (E ⊕ E ∗ ). This has no much meaning and
interest for vector spaces, but it will be useful in this paper as a linearized model for manifolds.

1895 C.10.1 Bargman transform

Let (E, g) be a linear Euclidean space. Let F := E ⊕ E ∗ with induced metric g = g ⊕ g −1 . One
has T F = T (E ⊕ E ∗ ) = (E ⊕ E ∗ ) ⊕ (E ⊕ E ∗ ). The exponential map is
(
TF →F
exp :
(ρ, ρ ) → ρ + ρ′
′

Let us define the “twisted pull back operator” exp g◦ : S (F ) → S (T F ) by its Schwartz kernel as
follows. For (ρ1 , ρ′1 ) ∈ T F , ρ2 ∈ F , ρ1 = (x1 , ξ1 ),ρ′1 = (x′1 , ξ1′ ),
g◦ δρ2 ⟩ = δρ − ρ +ρ′ e−iξ1 x′1
⟨δρ1 ,ρ′1 |exp 2 ( 1 ) 1

1900 We define the “restriction” operator r0 : S (T F ) → S (F ) by its Schwartz kernel

⟨δρ2 |r0 δρ1 ,ρ′1 ⟩ = δρ2 −ρ1 δρ′1

Let the “horizontal space” be

H := F ⊕ E. (C.51)
If B : S (E) → S (F ) is the Bargman transform defined in (C.8), we define the partial Bargman
transform
B : S (H) → S (T F )
by B (u (ρ) ⊗ v (x′ )) = u (ρ) ⊗ (Bv) (x′ ), i.e. action on the second part only.
1905 Let σ > 0 and let us introduce the cutoff function that truncates at distance σ in the fiber
F: 
S (T F ) → (
 S ′ (T F )
χσ : ′ u (ρ, ρ′ ) if ∥ρ′ ∥g ≤ σ
u (ρ, ρ ) →
otherwise

0


Let
†
B∆ g◦
χ := B χσ exp : S (F ) → S (H)
B := r0 B : S (H) → S (F )

C.10.2 Linear map

1910 Let A : E → E a linear invertible map and à := A−1 ⊕ A∗ the induced map on F := E ⊕ E ∗ .
Let A◦ : S (E) → S (E) the pull back operator and AH := Ã−1 ⊕ A : H → H the induced
map on H in (C.51).

Lemma C.36. Let us consider the difference operator

R := BA◦H B ∆ ◦ †
χ − BA B : S (F ) → S (F ) .

Then ∀N > 0, ∃CN > 0 such that for any σ > 0, ρ, ρ′ ,

D E−N
|⟨δρ′ |Rδρ ⟩| ≤ CN ρ − Ã−1 ρ′ σ −N .

Remark C.37. In particular, for σ = ∞, i.e. no cut-off, then R = 0. Another interesting

1915 particular case is A = Id.

92
 Proof. We compute the Schwartz kernel.

⟨δρ′ |BA◦H B ∆ ◦ † g◦
χ δρ ⟩ = ⟨δρ′ |r0 BAH B χσ exp δρ ⟩
Z
= ⟨δρ |r0 δρ1 ,ρ′1 ⟩⟨δρ1 ,ρ′1 |BA◦H B † δρ2 ,ρ′2 ⟩χσ (ρ′2 ) ⟨δρ2 ,ρ′2 |exp
g◦ δρ ⟩
Z  
′
δρ′ −ρ1 δρ′1 δÃρ2 −ρ1 ⟨φρ′1 |A◦ φρ′2 ⟩ χσ (ρ′2 ) δρ−(ρ2 +ρ′ ) e−iξ2 x2



=
2

−1 ′
 
= ⟨φ0 |A◦ φρ−Ã−1 ρ′ ⟩e−i((A ) ξ )(x−Ax ) χσ ρ − Ã−1 ρ′
∗ ′

because Dirac measures gave ρ′2 = ρ − ρ2 = ρ − Ã−1 ρ′ , ξ2 = (A∗ )−1 ξ ′ and x′2 = x − Ax′ . Then
 
iξ ′ (x′ −A−1 x′ )
⟨δρ′ |BA◦H B ∆ δ
χ ρ ⟩ = ⟨φ 0 |A◦
φ ρ−Ã−1 ρ′ ⟩e χ σ ρ − Ã −1 ′
ρ
   
◦ −◦ −◦ iξ ′ (x′ −A−1 x′ ) −1 ′
= ⟨φ0 |A Tx−Ax′ FTξ−A∗−1 ξ′ F φ0 ⟩e χσ ρ − Ã ρ
(C.7)
 
−◦ −◦
= ⟨φ0 |A◦ FT−A −◦
FTξ−◦ F φ0 ⟩


∗−1 ξ ′ F T−Ax ′ Tx
(C.2)
 
e−i(−A ξ )x eiξ (−A x ) χσ ρ − Ã−1 ρ′
∗−1 ′ ′ −1 ′

 
−◦

−◦ ◦ −◦ −◦

−1 ′
= = ⟨φ0 | FT−ξ ′ F T −x′ A Tx FT ξ F φ 0 ⟩χ σ ρ − Ã ρ
(C.2)
   
◦ −1 ′ ◦ † −1 ′
= ⟨φρ′ |A φρ ⟩χσ ρ − Ã ρ = ⟨δρ′ |BA B δρ ⟩χσ ρ − Ã ρ

If ρ − Ã−1 ρ′ ≥ σ then ⟨δρ′ |Rδρ ⟩ = 0 otherwise

∀N > 0, ∃CN > 0, ⟨φ0 |A◦ φρ−Ã−1 ρ′ ⟩ ≤ CA CN σ −N .

C.10.3 Taylor projectors

For ρ ∈ F , T̂ρ has been defined in (C.1). Let us define
(k)
TH := T̂−ρ Tk T̂ρ : S (H) → S (H)

1920 where ρ ∈ F denotes the first variable and the operators acts on the second variable x′ ∈ E
only.

Lemma C.38. Let us consider the difference operator

(k) ∗
R := BTH B ∆
χ − BTk B : S (F ) → S (F ) .

Then for ∥ρ − ρ′ ∥ ≤ σ we have ⟨δρ′ |Rδρ ⟩ = 0.

Remark C.39. In particular for σ = ∞, i.e. no cut-off, then R = 0.

Proof. We repeat the lines of proof of Lemma C.36.
(k) ∗ (k)
⟨δρ′ |BTH B ∆ g◦
χ δρ ⟩ = ⟨δρ′ |r0 BTH B χσ exp δρ ⟩
′ ′
= ⟨φ0 |T̂−ρ′ Tk T̂ρ′ φρ−ρ′ ⟩eiξ (x −x) χσ (ρ − ρ′ )
′ ′ ′ ′
= ⟨φ0 |T̂ρ†′ e−iξ x Tk T̂ρ′ φρ−ρ′ ⟩eiξ (x −x) χσ (ρ − ρ′ )
(C.3)

= ⟨φρ′ |Tk φρ ⟩χσ (ρ − ρ′ ) = ⟨δρ′ |BTk B † δρ ⟩χσ (ρ − ρ′ )

93
1925 D Linear expanding maps
Let (E, g) a finite dimensional vector space with Euclidean metric g. In this section we consider
a linear invertible and contracting map ϕ : (E, g) → (E, g) i.e.
1/t
λ± := lim log ϕt (D.1)
t→±∞

satisfy λ− ≤ λ+ < 0. The pull back operator ϕ◦ : S (E) → S (E) has been defined in (A.4).
The purpose of this section is to study the spectrum of the following operator

Op (Φ) = |detϕ|1/2 ϕ◦ : S (E) → S (E) (D.2)

(C.42)

1930 on an adequate Hilbert space that contains S (E) (notice that Op (Φ) is unitary in L2 (E),
but has essential spectrum on the unit circle). Recall that Op (Φ) : = (Υ (Φ))1/2 B † Φ−◦ B
(C.34)
with Φ := ϕ−1 ⊕ ϕ∗ : E ⊕ E ∗ → E ⊕ E ∗ being the induced map on cotangent space and
B : S (E) → S (E ⊕ E ∗ ) defined in (C.8).
For k ∈ N, consider Tk the finite rank projector defined in (C.48). The vector space
1935 Im (Tk ) ⊂ S ′ (E) is finite dimensional. Since ϕ is a linear map, we have for any k ∈ N,

[ϕ◦ , Tk ] = 0, (D.3)

hence Op (Φ) : Im (Tk ) → Im (Tk ) is invariant and finite rank (we can compute explicitly its
spectrum from the spectrum of ϕ). Let
1/t
γk± := lim log Op Φt (D.4)


/Im(Tk )
.
t→±∞

We can13 compute γk± from the eigenvalues of ϕ, see remark [16, rem. 3.4.7]. As in (1.7), we
have    
d d
−
+ k λ− ≤ γk ≤ γk ≤ +
+ k λ+ , (D.5)
2 2
1940 With d = dimE. For every k ∈ N, we have γk− ≤ γk+ , γk+1 ±
≤ γk± . The spectrum of Op
n o (Φt ) :
− +
Im (Tk ) → Im (Tk ) is discrete and contained in the annulus z ∈ C, etγk ≤ |z| ≤ etγk . How-
ever we want to understand the action of Op (Φ) on every function in S (E). For K ∈ N,
let !
XK
T≥(K+1) := IdS(E) − Tk .
k=0
+
We will show below in Proposition D.3 that Op (Φt ) T≥(K+1) HW (E)
≤ Cϵ e(γK+1 +ϵ)t for some
1945 adequate norm ∥.∥HW (E) that we first define.

D.1 Anisotropic Sobolev space HW (E)

Let 0 < γ < 1, R ∈ R. We define a weight function W : E ⊕ E ∗ → R+ similar to (4.49) as
follows. For ρ = (x, ξ) ∈ E ⊕ E ∗ , let
D E−γ
hγ (ρ) = ∥ρ∥g⊕g−1

and D ER
hγ (ρ) ∥ξ∥g−1
W (ρ) := D ER (D.6)
hγ (ρ) ∥x∥g

In the simple 1 − dim case ϕ (x) = eλ x on R, with λ± = λ < 0, we get γk+ = γk− = 1
13


2 + k λ.

94
 Definition D.1. For u ∈ S (E), we define the norm

∥u∥HW (E) := ∥WBu∥L2 (E⊕E ∗ ) , (D.7)

(where W denotes multiplication operator by W) and the Sobolev space

HW (E) := {u ∈ S (E)} (D.8)

where the completion is with the norm ∥u∥HW (E) .

1950

Remark D.2.

• For R = 0 we have W ≡ 1 hence HW (E) = L2 (E).

• We have that W (0, ξ) ≤ ⟨|ξ|⟩r and W (x, 0) ≤ ⟨|x|⟩−r with the order r = R (1 − γ) as in
(4.20) (except for the factor 21 that came from the metric).

1955 For a non linear map (Axiom A diffeomorphisms), it has been explained in [18, section 4] that
⊥
it is necessary to use a metric on phase space as (2.25) with δ ⊥ (η) = ⟨|η|⟩−α and exponent
α⊥ ≥ 1/2 to perform the micro-local analysis and get Ruelle spectrum. In this section, we
consider the special case of a linear map ϕ : E → E is linear and for that reason we can use
the constant Euclidean metric g ⊕ g −1 on T ∗ E = E ⊕ E ∗ (i.e. exponent α⊥ = 0) to get the
1960 Ruelle spectrum of the operator Op (Φ) in (D.2).

D.2 Result
The result that we will show is
Proposition D.3. For any ϵ > 0, K ∈ N, R in (D.6) large enough so that
+
λ+ R (1 − γ) < γK+1 ,

∃Cϵ > 0, ∀t ≥ 0,
+
Op Φt T≥(K+1) ≤ Cϵ e(γK+1 +ϵ)t , (D.9)


HW (E)
+
with γK+1 defined in (D.4).

D.2.1 Proof of Proposition D.3

1965 We have Op
˜ (Φ) = (Υ (Φ))1/2 PΦ−◦ P and let
(C.33)

˜ W (Φ) := W Op
Op ˜ (Φ) W −1 .

From (C.35) and (D.7) R 2we have the commutative diagram (with the notation of weighted norm
∥u∥L2 (E⊕E ∗ ;W 2 ) := W (x, ξ) u (x, ξ) (2π)dimE )
dxdξ

B W
HW (E) −→ L2 (E ⊕ E ∗ ; W 2 ) −→ L2 (E ⊕ E ∗ )
↓ Op (Φ) ˜ (Φ)
↓ Op ↓ Op˜ W (Φ) (D.10)
B W
HW (E) −→ L2 (E ⊕ E ∗ ; W 2 ) −→ L2 (E ⊕ E ∗ )
where horizontal arrows are isometries by definition. Hence, the study of Op (Φt ) : HW (E) →
HW (E) is equivalent to study Op
˜ W (Φt ) : L2 (E ⊕ E ∗ ) → L2 (E ⊕ E ∗ ).

95
 Lemma D.4. “Decay property of W with respect to Φ”. There exists C > 0, such that
for any t ≥ 0, there exists Ct such that for any ρ ∈ E ⊕ E ∗

W (Φt (ρ))
≤C
W (ρ)
≤ Ce−Λt if ∥ρ∥g ≥ Ct (D.11)

with
Λ = −λ+ R (1 − γ) > 0. (D.12)
1970

Proof. Similar to the proof in [18, Thm 5.9]. Write ρ = (x, ξ) ∈ T ∗ E = E ⊕ E ∗ . We consider
different zones of T ∗ E:

1. If |x| , |ξ| ≤ 1 then W (ρ) ≍ 1.

2. If 1 ≤ |ξ| ≤ |x|γ then W (ρ) ≍ |x|−R(1−γ) .

1975 3. If 1 ≤ |x| ≤ |ξ|γ then W (ρ) ≍ |ξ|R(1−γ) .

 R
|ξ|
4. Otherwise W (ρ) ≍ |x| .

We observe that for t ≥ 0, |ϕ−t (x)| ≥ e−λ+ t |x| and |ϕ∗t (ξ)| ≤ eλ+ t |ξ| from which we deduce
Lemma D.4.

Truncation in phase space near the trapped set Let χ ∈ Cc∞ (R+ ; [0, 1]) such that

χ (x) = 1 if x ≤ 1,
χ (x) = 0 if x ≥ 2,

Let σ > 0. For ρ ∈ E ⊕ E ∗ let

∥ρ∥g
 
χσ (ρ) := χ .
σ

Lemma D.5. We have ∀R > 0,∀ϵ > 0,∃C > 0, ∀t ≥ 0, ∃σt > 0, ∀σ > σt ,

˜ W Φt (1 − χσ ) ≤ Ce(−Λ+ϵ)t . (D.13)


Op
L2

where Λ is given in (D.12). The operator Op ˜ W (Φt ) χσ is Trace class in L2 . Consequently

for any t > 0, the essential spectral radius is
 
˜ W Φt 2 2 ≤ e−Λt . (D.14)


ress. Op L →L
1980

Proof. For (D.13), see [18, thm 5.13].

Lemma D.6. Let k ∈ N. If k + 1 < r in (4.20), then Tk : HW (E) → HW (E) is a

bounded operator.

(α)
Proof. We check that in (C.48), ∥xα ∥HW (E) < ∞ and δ0 < ∞.
HW (E)∗

96
 +
Lemma D.7. Let K ∈ N and R > 0 large enough such that −Λ < γK+1 in (D.12). For
t > 0, the spectral radius is
  +
rspec Op Φt T≥(K+1) H (E) = etγK+1



W

1985 Remark D.8. Lemma D.7 is similar to [16, Claim (2) in Prop. 3.4.6].
Proof. First, from (D.14) and because Op (Φt ) Tk is finite rank, we have that
 
ress Op Φt T≥(K+1) H (E) ≤ e−Λt ,



W

i.e. Op (Φt ) T≥(K+1) : HW (E) → HW (E) has discrete spectrum outside the disk of radius e−Λt .
From Taylor-Lagrange remainder formula, for any u, v ∈ S (E), we have
+
⟨v|Op Φt T≥(K+1) u⟩L2 ≤ CetγK+1 ∥u∥C K+1 xK+1 v (D.15)


L1
.
+
i.e. correlation functions decay faster than etγK+1 . We deduce that Op (Φt ) T≥(K+1) has no
+
1990 spectrum on |z| > etγK+1 .
Finally Lemma D.7 is equivalent to Proposition D.3.

D.3 Discrete Ruelle spectrum in a simple toy model

We give here a very simple example that illustrates Section D and some mechanisms that play
a role in this paper. (See the lecture notes [12] for further details). On Rx let us consider the
1995 vector field
X = −x∂x (D.16)
that gives a flow ϕt (x) = e−t x that is contracting for t > 0,
hence we think Es = Rx as a
stable direction. In order to study the pull back operator etX u (x) = u (ϕt (x)) = u (e−t x) on
functions u ∈ S (R), we consider the induced (pull-back) flow on the cotangent space T ∗ Es =
T ∗ R given by
ϕ̃t (x, ξ) = et x, e−t ξ .

2000 We observe that for any k ∈ N, the monomial xk is an eigenfunction of X (hence of etX ) with
eigenvalue (−k):
Xxk = (−k) xk
and the spectral projector is
1 (k)
Tk = xk ⟨ δ |.⟩,
k!
′
where δ (k)
is the k-th derivative of the Dirac distribution (notice indeed from ⟨ k!1 δ (k) |xk ⟩ = δk=k′
′
that δ (k) k forms a dual basis to xk k′ ).

2005 However xk , δ (k) do not belong to L2 (R). In the Hilbert space L2 (R) one has X † = −X+1 ⇔

†
X − 12 = − X − 12 that implies that the spectrum of X is on the vertical axis 12 + iR with

some essential spectrum. Some better Hilbert space HW (R) is constructed as in (D.8). For
0 < h ≪ 1, R ≥ 0 we define the “escape function” or “Lyapounov function” for ϕ̃t that is
D√ ER
hξ
W (x, ξ) = D√ ER .
hx

97
 Im(z)
R → +∞

−k, ... −1 0 Re(z)

X = −x∂x
−R + 1/2 + o(1)
0 x ∈ R ≡ Es −2R + 1/2 + o(1)

Figure D.1: In blue, discrete spectrum of the operator (D.16), X = −x∂x in HW (R). In brown,
the essential spectrum moves far away if R → +∞.

t W◦ϕ̃t
Indeed, it satisfies W◦
W
ϕ̃
≤ C everywhere and W
≤ e−Rt far from the “trapped set” or “non
2010 wandering set” (0, 0). In the Hilbert space

HW (R) := Op W −1 L2 (R) ,

for −R + 12 < −k, we have that xk ∈ HW (R) and ∥Tk ∥HW ≤ C is a bounded operator. See
figure [Link] large time t ≫ 1, the emerging behavior of etX u is given by

etX u = Π0 u + O e−t

= 1 u (0) + O e−t

i.e. projection onto the constant function, the remainder is in operator norm.

References
[1] V.I. Arnold. Les méthodes mathématiques de la mécanique classique. Ed. Mir. Moscou,
2015 1976.

[2] V. Baladi. Anisotropic Sobolev spaces and dynamical transfer operators: C ∞ foliations. In
Algebraic and topological dynamics, volume 385 of Contemp. Math., pages 123–135. Amer.
Math. Soc., Providence, RI, 2005.

[3] Sonja Barkhofen, Philipp Schütte, and Tobias Weich. Semiclassical formulae for wigner
2020 distributions. Journal of Physics A: Mathematical and Theoretical, 55(24):244007, 2022.

[4] S. Bates and A. Weinstein. Lectures on the Geometry of Quantization, volume 8. American
Mathematical Soc., 1997.

[5] O. Butterley and C. Liverani. Smooth Anosov flows: correlation spectra and stability. J.
Mod. Dyn., 1(2):301–322, 2007.

2025 [6] A. Cannas Da Silva. Lectures on Symplectic Geometry. Springer, 2001.

[7] J.J. Duistermaat. Fourier integral operators, volume 130. Birkhauser, 1996.

[8] S. Dyatlov. Resonance projectors and asymptotics for r-normally hyperbolic trapped sets.
Journal of the American Mathematical Society, 28(2):311–381, 2015.

[9] S. Dyatlov, F. Faure, and C. Guillarmou. Power spectrum of the geodesic flow on hyperbolic
2030 manifolds. Analysis and PDE link, 8:923–1000, 2015.

[10] K.J. Engel and R. Nagel. One-parameter semigroups for linear evolution equations, volume
194. Springer, 1999.

98
 [11] F. Faure. Prequantum chaos: Resonances of the prequantum cat map. arXiv:nlin/0606063.
Journal of Modern Dynamics, 1(2):255–285, 2007.

2035 [12] F. Faure. From classical chaos to quantum chaosspectrum. In link, Lectures notes for the
school 22-26 April 2019 at CIRM. 2018.

[13] F. Faure, N. Roy, and J. Sjöstrand. A semiclassical approach for Anosov diffeomorphisms
and Ruelle resonances. Open Math. Journal. link, 1:35–81, 2008.

[14] F. Faure and J. Sjöstrand. Upper bound on the density of Ruelle resonances for Anosov
2040 flows. a semiclassical approach. Comm. in Math. Physics, Issue 2. link, 308:325–364, 2011.

[15] F. Faure and M. Tsujii. Band structure of the Ruelle spectrum of contact Anosov flows.
Comptes rendus - Mathématique 351 , 385-391, (2013) link, 2013.

[16] F. Faure and M. Tsujii. Prequantum transfer operator for symplectic Anosov diffeomor-
phism. Asterisque 375 (2015), link, pages ix+222 pages, 2015.

2045 [17] F. Faure and M. Tsujii. The semiclassical zeta function for geodesic flows on negatively
curved manifolds. Inventiones mathematicae. link, 2016.

[18] F. Faure and M. Tsujii. Fractal Weyl law for the Ruelle spectrum of Anosov flows. Annales
Henri Lebesgue, arXiv:1706.09307 link, 6:331–426, 2023.

[19] L. Flaminio and G. Forni. Invariant distributions and time averages for horocycle flows.
2050 Duke Mathematical Journal, 119(3):465–526, 2003.

[20] G. Folland. Harmonic Analysis in phase space. Princeton University Press, 1988.

[21] S. Gallot, D. Hulin, and J. Lafontaine. Riemannian geometry, volume 3. Springer, 1990.

[22] I. Gohberg, S. Goldberg, and N. Krupnik. Traces and Determinants of Linear Operators.
Birkhauser, 2000.

2055 [23] S. Gouëzel and C. Liverani. Banach spaces adapted to Anosov systems. Ergodic Theory
and dynamical systems, 26:189–217, 2005.

[24] C. Guillarmou and C. Cekic. First band of ruelle resonances for contact anosov flows in
dimension 3, 2020.

[25] C. Guillarmou, J. Hilgert, and T. Weich. Classical and quantum resonances for hyperbolic
2060 surfaces. arXiv preprint arXiv:1605.08801, 2016.

[26] V. Guillemin. Lectures on spectral theory of elliptic operators. Duke Math. J., 44:485–517,
1977.

[27] B. Hall. Holomorphic Methods in Mathematical Physics. Contemp. Math., 260:1–59, 1999.

[28] J. Hilgert, T. Weich, and L. L Wolf. Higher rank quantum-classical correspondence. arXiv
2065 preprint arXiv:2103.05667, 2021.

[29] S. Hurder and A. Katok. Differentiability, rigidity and Godbillon-Vey classes for Anosov
flows. Publ. Math., Inst. Hautes étud. Sci., 72:5–61, 1990.

[30] N. Lerner. Metrics on the phase space and non-selfadjoint pseudo-differential operators,
volume 3. Springer, 2011.

2070 [31] C. Liverani. On contact Anosov flows. Ann. of Math. (2), 159(3):1275–1312, 2004.

99
 [32] D McDuff and D Salamon. Introduction to symplectic topology, 2nd edition. clarendon
press, Oxford, 1998.

[33] E. Nelson. Review of stochastic mechanics. In Journal of Physics: Conference Series,

volume 361, page 012011. IOP Publishing, 2012.

2075 [34] G. Paternain. Geodesic flows, volume 180. Springer Science & Business Media, 2012.

[35] Y. G. Sinai. Geodesic flows on compact surfaces of negative curvature. In Doklady Akademii
Nauk, volume 136, pages 549–552. Russian Academy of Sciences, 1961.

[36] J. Sjöstrand and M. Zworski. Asymptotic distribution of resonances for convex obstacles.
Acta Mathematica, 183(2):191–253, 1999.

2080 [37] P. Stefanov, G. Vodev, et al. Distribution of resonances for the neumann problem in linear
elasticity outside a strictly convex body. Duke Mathematical Journal, 78(3):677–714, 1995.

[38] M. Taylor. Partial differential equations, Vol II. Springer, 1996.

[39] E. C. Titchmarsh et al. The theory of the Riemann zeta-function. Oxford University Press,
1986.

2085 [40] F. Treves et al. Analytic partial differential equations. Springer, 2022.

[41] M. Tsujii. Quasi-compactness of transfer operators for contact Anosov flows. Nonlinearity,
arXiv:0806.0732v2 [[Link]], 23(7):1495–1545, 2010.

[42] N.M.J. Woodhouse. Geometric quantization. Clarendon Press, Oxford, 1992.

[43] S. Zelditch. Index and dynamics of quantized contact transformations. Ann. Inst. Fourier
2090 (Grenoble), 47(1):305–363, 1997.

100