ON THE STABILITY/SENSITIVITY OF RECOVERING VELOCITY FIELDS
FROM BOUNDARY MEASUREMENTS
By
Hai Zhang

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Mathematics - Doctor of Philosophy
2013

ABSTRACT
ON THE STABILITY/SENSITIVITY OF RECOVERING VELOCITY
FIELDS FROM BOUNDARY MEASUREMENTS
By
Hai Zhang
The thesis investigates the stability/sensitivity of the inverse problem of recovering
velocity fields in a bounded domain from the boundary measurements. The problem
has important applications in geophysics where people are interested in finding the inner
structure (the velocity field in the elastic wave models) of earth from measurements on
the surface. Two types of measurements are considered. One is the boundary dynamic
Dirichlet-to-Neumann map (DDtN) for the wave equation. The other is the restricted
Hamiltonian flow induced by the corresponding velocity field at a sufficiently large time
and with domain the cosphere bundle of the boundary, or its equivalent form the scattering
relation. Relations between these two type of data are explored. Three main results
on the stability/sensitivity of the associated inverse problems are obtained: (1). The
sensitivity of recovering scattering relations from their associated DDtN maps. (2). The
sensitivity of recovering velocity fields from their induced boundary DDtN maps. (3). The
stability of recovering velocity fields from their induced Hamiltonian flows. In addition,
a stability estimate for the X-ray transform in the presence of caustics is established.
The X-ray transform is introduced by linearizing the operator which maps a velocity field
to its corresponding Hamiltonian flow. Micro-local analysis are used to study the X-ray
transform and conditions on the background velocity field are found to ensure the stability
of the inverse transform. The main results suggest that the DDtN map is very insensitive

to small perturbations of the velocity field, namely, small perturbations of velocity field can
result changes to the DDtN map at the same level of large perturbations. This differs from
existing H¨lder type stability results for the inverse problem in the case when the velocity
o
fields are simple. It gives hint that the methodology of velocity field inversion by DDtN
map is inefficient in some sense. On the other hands, the main results recommend the
methodology of inversion by Hamiltonian flow (or its equivalence the scattering relation),
where the associated inverse problem has Lipschitz type stability.

ACKNOWLEDGMENTS

I am deeply indebted to my advisor Prof. Gang Bao, for giving me the opportunity to join
his research group at MSU, for his constant support and encouragement for my research,
and for his deep insight on the various research topics which guided my research and
shaped my understanding of both pure and applied mathematics.
I would like to thank Prof. Tien-Yien Li, Prof. Jianliang Qian, Prof. Moxun Tang
and Prof. Zhengfang Zhou for serving in my thesis committee.
I would like to thank the Professors who taught me during my study at MSU, especially, Prof. Jianliang Qian for valuable discussions on Gaussian beam method and the
collaboration which resulted a beautiful paper, and Prof. Zhengfang Zhou for sharing his
understanding of nonlinear partial differential equations.
Thanks also go to former and current members of Prof. Bao’s research group at MSU.
Justin Droba, Guanghui Hu, Jun Lai, Peijun Li, Junshan Li, Songting Luo, Yuanchang
Sun, Yuliang Wang, Eric Wolf, Xiang Xu, and KiHyun Yun, for the many interesting
discussions on various topics and many happy time together.
I would like to thank Prof. Gengsheng Wang at Wuhan University for his constant
encouragement on my research, and Prof. Jun Zou at the Chinese University of Hong
Kong for introducing me to the exciting field of inverse problems.
Finally, I would like to delicate this thesis to my parents, who are always the strongest
support behind me.

iv

TABLE OF CONTENTS

Chapter 1 Introduction . . . . . . . . .
1.1 The main inverse problem . . . . .
1.2 Overview of the approach of solving
1.3 Outline of the thesis . . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

. 8
. 8
. 10
. 13

Chapter 3 Weighted X-ray transform for scale functions
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 X-ray transform in the Euclidean space . . . . . . . . . .
3.3 X-ray transform with weight in the Euclidean space . . .
3.4 X-ray transform with weight on Riemannian Manifold . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

Chapter 2 The inverse kinematic
2.1 Introduction . . . . . . . . . .
2.2 Mukhometov’s solution in 2D
2.3 The lens rigidity problem . . .

. . . . . . . . . . .
. . . . . . . . . . .
the main problem
. . . . . . . . . . .

problem of seismic
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .

.
.
.
.

1
1
4
7

15
15
16
18
21

Chapter 4 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Chapter 5 Sensitivity of recovering scattering relations from
duced DDtN maps . . . . . . . . . . . . . . . . . . . .
5.1 Gaussian beam solutions to the wave equation . . . . . . . . .
5.2 The sensitivity result . . . . . . . . . . . . . . . . . . . . . . .

their
. . . .
. . . .
. . . .

in. . . 27
. . . 27
. . . 39

Chapter 6 Stability of X-ray transform in the presence of caustics . . .
6.1 The X-ray transform resulted from linearizing Hamiltonian flow with respect to velocity field . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Statement of the main results for the stability of the geodesic X-ray transform Ic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Local properties of the normal operator N . . . . . . . . . . . . . . . . .
6.4 Singularities of the map ϕ(x, ·) . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Discussions on the concept of Fold-regular . . . . . . . . . . . . . . . . .
6.6 Proof of Theorem 6.2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.7 Proof of Theorem 6.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 44
. 44
.
.
.
.
.
.

47
52
57
59
62
69

Chapter 7 Stability of of recovering velocity fields from their induced
Hamiltonian flows . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

v

Chapter 8 Sensitivity of recovering velocity fields from their induced
DDtN maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
BIBLIOGRAPHY

. . . . . . . . . . . . .

vi

. . . . . . . . . . . . . . . . . 78

Chapter 1
Introduction

1.1

The main inverse problem

The work in the thesis mainly comes from [9], where the sensitivity of the inverse problem
of recovering velocity fields from boundary DDtN maps is investigated. The result on the
stability of recovering velocity fields from Hamiltonian flows are obtained as a byproduct.
For this reason, we refer the inverse problem of recovering velocity fields from boundary
DDtN maps as our main problem, which we formulate now. Let Ω be a bounded strictly
convex smooth domain in Rd , d ≥ 2, with boundary Γ. Let c(x) be a velocity field in
Ω which characterizes the wave speed in the medium and let T be a sufficiently large
positive number. Consider the following wave equation system:

1

(x, t) ∈ Rd × (0, T )

u(0, x) = ut (0, x) = 0,

(1.1)

x ∈ Ω,

utt − ∆u = 0,
c2 (x)

(1.2)

u(x, t) = f (x, t),

(x, t) ∈ Γ × (0, T ).

(1.3)

1
For each f ∈ H0 ([0, T ] × Γ), it is known that (see for instance [31]) there exists an unique

solution u ∈ C 1 (0, T ; L2 (Ω))

∩

C(0, T ; H 1 (Ω)), and furthermore ∂u ∈ L2 ([0, T ] × Γ),
∂ν

1

where ν is the unit outward normal to the boundary. The DDtN map Λc is defined by

Λc (f ) :=

∂u
|
.
∂ν [0,T ]×Γ

The inverse problem is to recover the velocity function c from the DDtN map Λc .
The uniqueness of the inverse problem is solved by the boundary control method first
introduced by Belishev in [10]. The method can also be used to solve the uniqueness for
more general problems, for instance, the anisotropic medium case. We refer to [11], [13],
[12], [30] and the references therein for more discussions.
We are interested in the sensitivity of the above inverse problem. Namely, we want to
investigate how sensitive or stable is it to recover the velocity field from the DDtN map
and characterize how small changes in the DDtN map affect the recovered velocity field.
The inverse problem of recovering velocity field is closely related to the inverse kinematic problem in geophysics, which we shall briefly review in chapter and we refer to
[43] for more discussions. It also can be viewed as a special case of the inverse problem of
recovering a Riemannian metric on a Riemannian manifold. Indeed, it corresponds to the
case when the metrics are restricted to the class of those which are conformal to the Euclidean one. The inverse problem of recovering a Riemannian metric has been extensively
studied in the literature. The uniqueness is proved by Belishev and Kurylev in [13] by
using the boundary control method. However, it is still unclear that if their approach can
give a stability estimate since it uses in an essential way an unique continuation property
of the wave equation.
The first stability result on the determination of the metric from the DDtN map was

2

given by Stefanov and Uhlmann in [46], where they proved conditional stability of H¨lder
o
type for metrics close enough to the Euclidean one in C k for k ≫ 1 in three dimensions.
Later, they extended the stability result to generic simple metrics, [49]. Here we recall
the definition of simple manifold.
Definition 1.1.1. A compact Riemannian manifold (M, g) with boundary ∂M is called
simple if ∂M is strictly convex with respect to g, and for any x ∈ M , the exponential map
expx : exp−1 (M ) → M is a diffeomorphism.
x
An important feature of their approach is to first derive a stability estimate of recovering the boundary distance function from the DDtN map and then apply existing
results from the boundary rigidity problem in geometry. Their approach was extended by
Montalto in [33] to study the more general problem of determine a metric, a co-vector and
a potential simultaneously from the DDtN map, and a similar H¨lder type conditional
o
stability result was obtained. The stability of the inverse problem of determining the
conformal factor to a fixed simple metric was studied by Bellassoued and Ferreira in [14].
They proved the H¨lder type conditional stability result for the case when the conformal
o
factors are close to one. We comment that the result in [14] holds for all simple metrics.
For other stability results on the related problems, we refer to the references in [33].
We emphasize that all of the above stability results deal with the case when the metrics
are simple. To our best knowledge, no stability result is available for the general case when
the metrics are not simple.
The thesis is devoted to the study of the general case when the metrics induced by the
velocity fields are not simple. To avoid technical complications due to the boundary, we

3

restrict our study to situations when the velocity fields are equal to one near the boundary.
From this point of view, our results can be regarded as interior estimates. We refer to
[49], [51] and the references therein for useful boundary estimates. In contrast to existing
results mentioned above, where H¨lder type stability estimate are suggested for the case
o
when the velocity fields are simple, our result shows that the inverse problem of recovering
velocity fields from their induced DDtN maps is insensitivity to small perturbations of
the data. In fact, we showed that for a quite general velocity field, which we call “foldregular” (see definition 6.2.3, 6.2.4, 8.0.1), if another velocity field is sufficiently close to
it and satisfies a certain orthogonality condition, then the two must be equal if the two
corresponding DDtN maps are sufficiently close. On the other hand, we showed that the
inverse problem of recovering velocity fields from their induced Hamiltonian flows at a
sufficiently large time is well-posed, in the sense that a local Lipschitz stability estimate
for the inverse problem can be established.

1.2

Overview of the approach of solving the main
problem

We now briefly review the approach we used to solve the main problem introduced in the
previous section. We first derive a sensitivity result of recovering the scattering relation
from the DDtN map. Our result shows that two scattering relations must be identical if the
two corresponding DDtN maps are sufficiently close in some suitable norm. Equivalently,
any arbitrarily small change in the scattering relation can imply a certain change in the
DDtN map. To our best knowledge, this is the first sensitivity result for the problem in the
4

non-simple metric case. Moreover, our result is fundamentally different from those in the
literature where Lipschitz, H¨lder or logarithmic estimates are derived , see for examples
o
[8], [2], [46], [47], [48], [49], [33] and [28]. This is the reason the term “sensitivity” is used
instead of “stability” whenever it is more proper. We remark that when the geometry
induced by the velocity field is simple, the scattering relation is equivalent to the boundary
distance function. In that case, a H¨lder type interior stability estimate for recovering the
o
boundary distance function from DDtN map has been established in [49]. Compared to
the H¨lder type result, our result is much stronger. Our approach is based on Gaussian
o
beam solutions to the wave equation, which are capable of dealing with caustics, major
obstacles to the construction of classic geometric-optics solutions. We refer to [37] and
[30] for more discussions on Gaussian beams and its applications.
t
We observe that for any velocity filed c, the induced Hamiltonian flow Hc when re-

stricted to the unit cosphere bundle S ∗ Rd determines the scattering relation Sc . We
t
linearize the operator which maps c to Hc |S ∗ Rd and obtain a geodesic X-ray transform

operator Ic with matrix-valued weight. Note that the scattering relation (or Hamiltonian
flow) is the natural object to study when the metric is not simple. It is related to the
lens rigidity problem in geometry. We refer to [53] for more discussions on the topic.
The boundary distance function (global or local) has received extensive attention in the
literature for the problem where the metrics are “simple” or “regular”, see for instance
[48], [53]. However, it is unlikely to work for the case of general non-simple metrics. In
the thesis, we attempt to overcome the difficulty by analyzing the scattering relation (or
Hamiltonian flow).
We study the inverse problem of recovering a vector-valued function f from its weight5

ed geodesic transform Ic f . For a fixed interior point x, we use a carefully selected set
∗
of geodesics whose conormal bundle can cover the cotangent space Tx Rd to recover the

singularity of f at x. We allow fold caustics along these geodesics, but require that these
caustics contribute to a smoother term in the transform than x itself. It is still an open
problem to show that such a set of geodesics exists generically for a general velocity field
with caustics. But we draw evidence from the classification result on caustics and regularity theory of Fourier Integral Operators (FIOs) to show that it is the case under some
natural assumptions in the dimensions equal to or greater than three. We call the interior
point which has the above set of geodesics “fold-regular”. A local stability estimate is
derived near fold-regular points.
We define a velocity field to be fold-regular if every interior point is fold-regular with
respect to the Hamiltonian flow induced by it. As a consequence of the above local
stability result, we obtain a Lipschitz stability result, up to a finite dimensional space, for
the X-ray transform in a fold-regular background velocity field, or the linearized inverse
problem of recovering velocity fields from their induced Hamiltonian flows at a foldregular background velocity field. By standard arguments of linearization, it yields a
similar stability estimate for the nonlinear inverse problem. We remark that it is still
an open problem to show whether the finite dimensional space is empty or not. This is
closely related to the injectivity of the X-ray transform Ic f .
Finally, We combine the stability result on the X-ray transform and the sensitivity
result on recovering scattering relations from DDtN maps to deduce a sensitivity result
for the inverse problem of recovering velocity fields from their induced DDtN maps.

6

1.3

Outline of the thesis

The thesis is organized as follows. In the first two chapters, chapter 2 and 3, we present
some background material for the problem we contributed in the thesis. The results
shown therein are standard and well-known in the field. In chapter 2, we give a brief
introduction to the inverse kinematic problems of seismic (or travel time tomography in
seismic). It serves as a major motivation for our investigation of the stability of recovering
velocity field from Hamiltonian flow, which can be viewed as a generalization of the travel
time tomography to the more general case when the velocity fields are not simple (see
chapter 7). In chapter 3, we introduce the X-ray transform for scalar functions in the case
when the background metric is simple. It prepares necessary preliminaries for the more
general results in chapter 6, which works for the case when the background metric is not
simple. Starting from chapter 4, we investigate the main problem in the thesis. This is
where new ideas and results are developed. We first give some preliminaries in chapter
4, fixing notations and conventions. We then derive a sensitivity result for recovering
scattering relations from their corresponding DDtN maps in chapter 5. In chapter 6, we
show the equivalence of the scattering relation and the Hamiltonian flow. We linearize
the Hamiltonian flow with respect to the velocity field. This leads to a X-ray transform.
We study properties of the X-ray transform and establish some stability estimate for the
transform. In chapter 7, we apply the stability estimate to the nonlinear inverse problem of
recovering velocity fields from their induced Hamiltonian flows, a Lipschitz type stability
result is obtained. Finally, in chapter 8, we combining the results from chapter 5 and 6
to study the sensitivity of recovering velocity fields from their induced DDtN maps.

7

Chapter 2
The inverse kinematic problem of
seismic

2.1

Introduction

In geophysics, a basic problem is to study the earth’s inner structure by making observations of various wave fields on the surface. In the elastic model, one assumes that the
earth is an elastic body and the inner structure of earth is characterized by the velocity
field of the elastic waves. Geophysicists are interested in finding the velocity field by measuring the travel time of seismic waves between points on the surface of the earth. The
problem is referred to the inverse kinematic problem of seismic or travel time tomograph
in seismic. The first result on the problem was obtained by G.Herglotz, E.Wiechert and
K.Zeoppritz in 1905. They modeled the earth by a ball with spherically symmetric metric
and studied the following mathematical problem: Let M = {x ∈ R3 : |x| ≤ R}. Assume
that the metric of M is given by

dτ 2 = n2 (r)dx2

8

(2.1)

where n(r) is a positive function on [0, R]. The mathematical problem then becomes to
determine the function n(r) from the induced boundary distance function.
They solved the problem and constructed solution to the inverse problem under the
following “non-trapping” assumption

(rn′ (r))′ > 0.

The general problem is to recover a metric of the following form

dτ 2 = n2 (x1 , x2 , x3 )dx2

(2.2)

from the boundary distance function induced by it. Linearization of the problem leads
to a linear integral geometry problem. Both the linear and the nonlinear problem in two
dimensions were solved by Mukhometov [20, 21] under assumptions which can be interpreted as the “simple metric” assumption. In addition Romanov studied the linearized
problem near a spherically symmetric metric in 1967 [41].
Generalizations of the above two-dimensional case to multinational case were studied
by Mukhometov [22, 23], Anikonov and Romanov [4], Bernstein and Gerver [15], Romanov
[42], Pestov and Sharafutdinov [34].
We remark that all the results above require that the velocity fields are “simple”, in
the sense that the metrics induced by them are simple. In the general case when the
velocity fields are not simple, the inverse problem of travel time tomography is no longer
well-posed. In fact, in that case, the geodesics may have conjugate points (or caustics)

9

and hence are not necessarily length minimizing. Therefore, linearization of the travel
time tomography problem may not be well-defined. For the general case without “simple
metric” assumption, we need more data than the travel time in order to “regularize”
the problem. This leads to the lens rigidity problem where additional information about
directions of geodesics (the scattering relation) are used to recover the associated metric
(or the velocity field in our case). We shall briefly introduce the lens rigidity problem in
Section 2.3 and its connection with our result in chapter 6.

2.2

Mukhometov’s solution in 2D

In this section, we present Mukhometov’s solution to the inverse kinematic problem in the
two dimensional case. We choose his solutions for two main reasons: one is the simplicity
of the results; the other is its importance in the development of the field.
We first introduce the concept “regular family of curves” which is a key assumption in
his results. Let M be a bounded simply connected domain on the plane with C 1 smooth
boundary ∂M . Parameterize ∂M by

x1 = τ1 (t), x2 = τ2 (t),

0 ≤ t ≤ T,

where t is the arc length parameter and T is the length of ∂M . We call that a family of
curves Γ in M are “regular” if the following four conditions are satisfied:
(1). For any two different points on ∂M , there exists a unique curve γ ∈ Γ joining
them.
(2). The end points and inner points of any γ ∈ Γ belong to ∂M and M respectively.
10

Moreover, the lengths of all curves are uniformly bounded.
(3). For each point z ∈ M , direction θ, there exists a unique curve γ ∈ Γ passing
through z with direction θ. We parameterize the curve by γ = γ(z, θ, s) where s is the
standard length parameter.
(4). The vector-valued function γ in (3) is C 3 in M and satisfies

∂γ
≥ sC
∂(θ, s)

for some C > 0.
We remark that in the case when the family of curves are the geodesics (with respect
to some metric equipped to M ) passing through M , the “regular” condition can be shown
to be equivalent to the “simple metric” assumption on the underlining metric.
Now, we formulate and present Mokhometov’s solution to the 2D inverse kinematic
problem of seismic. Let n(x, y) > 0 be a function defined in M . Consider the Riemannian
metric
dτ 2 = n2 (x, y)(dx2 + dy 2 )
and the corresponding functional
∫
J(γ) =

∫

√
n(x, y) dx2 + dy 2 .

dτ =
γ

γ

Assume that the family of extremals of the functional J(γ) is regular in the sense just

11

defined. We can thus define the boundary distance function
∫
Γ(t1 , t2 ) =

nds,
γ(t1 ,t2 )

where γ(t1 , t2 ) is the unique extremal of the functional J(γ) joining the boundary points
(τ1 (t1 ), τ2 (t1 )) and (τ1 (t2 ), τ2 (t2 )).
Mukhometov established the following result.
¯
Theorem 2.2.1. Let n ∈ C 4 (M ) be such that the corresponding family of extremals for J
is regular. Then n(x, y) can be uniquely determined from Γ(t1 , t2 ), moreover, the following
stability estimate holds:
1 ∂(Γ1 − Γ2 )
∥n1 − n2 ∥L2 (M ) ≤ √ ∥
∥L2 ([0,T ]×[0,T ]) .
∂t1
2π
Proof. See [21].
In addition to the nonlinear inverse problem considered above, Mukhometov investigated the linearized problem, which is a X-ray transform, and derived a similar Lipschitz
stability estimate.
¯
Theorem 2.2.2. Let f ∈ C 2 (M ), define
∫
I(f )(t1 , t2 ) =

γ(t1 ,t2 )

f (x1 , x2 ) ds,

0 ≤ t1 , t2 ≤ T,

where γ(t1 , t2 ) is the unique curve in Γ joining the boundary points (τ1 (t1 ), τ2 (t1 )) and
(τ1 (t2 ), τ2 (t2 )). If the family of curves {γ(t1 , t2 )}0≤t1 ,t2 ,≤T are regular, then the following
12

estimate holds
1 ∂g(t1 , t2 )
∥f ∥L2 (D) ≤ √ ∥
∥L2 ([0,T ]×[0,T ]) .
∂t1
2π
Proof. See [20].

2.3

The lens rigidity problem

Let (M, g) be a compact Riemannian manifold with boundary ∂M . Let Ht be the geodesic
flow on the tangent bundle T M and let SM be the unit tangent bundle. Denote

S+ ∂M = {(x, ξ) : x ∈ ∂M, |ξ|g = 1, ⟨ξ, ν(x)⟩ > 0};
S− ∂M = {(x, ξ) : x ∈ ∂M, |ξ|g = 1, ⟨ξ, ν(x)⟩ < 0}.

where ν is the unit out-normal and ⟨·, ·⟩ stands for the inner product. We now define the
scattering relation. There are several definitions. Here, we only introduce the simplest
one. We assume that the metric g is known for all points on the boundary. For each
(x, ξ) ∈ S− ∂M , define L(g)(x, ξ) > 0 to be the first positive moment at which the unit
speed geodesic passing through (x, ξ) hits the boundary ∂M . If L(g)(x, ξ) does not exist,
we define formally Lg (x, ξ) = ∞ and call the corresponding geodesic trapped. We define
Σ(g) : S− ∂M → S+ ∂M by
Σ(g)(x, ξ) = ΦL(g) (x, ξ).
We call the manifold (M, g) non-trapping if there exists T > 0 such that L(g)(x, ξ) ≤ T
for all (x, ξ) ∈ S− ∂M . We call the pair (Σ(g), L(g)) the full scattering relation induced
by the metric g.
13

The lens rigidity problem can be formulated as recovering the metric g from its induced
full scattering relation (Σ(g), g). It is known that there is no uniqueness to this problem.
The first obstruction comes from the diffeomorphisms which leave the boundary ∂M fixed.
In addition to this, trapping of geodesics also prevents the uniqueness of the problem. See
the counterexamples constructed in [17]. Therefore, a more natural formulation for the
lens rigidity problem is as follows:
Conjecture 1. Given a compact non-trapping Riemannian manifold (M, g) with boundary ∂M , can we recover the metric g by its induced full scattering relation (Σ(g), L(g)),
up to an action of diffeomorphism which leaves the boundary fixed.
It was observed by Michel that the lens rigidity and boundary rigidity problem are
equivalent for simple metrics. We refer to [48] and the reference therein for the uniqueness
and stability results for the boundary rigidity problem. With regarding to the lens rigidity
problem, there are only two uniqueness results available:
1. Croke (2005) [16]: The finite quotient space of a lens rigid manifold is lens rigid.
2. Stefanov and Uhlmann (2009) [53]: Uniqueness up to diffeomorphism fixing the
boundary for metrics a priori close to a generic regular metrics.
To our best knowledge, there is no stability result on the lens rigidity problem. In
Chapter 7, we obtained a stability estimate for an equivalent problem in the case when
the metrics are conformal to the Euclidean metric, i.e. the inverse problem of recovering
velocity fields from their induced Hamiltonian flows. By exploring the relation between
the scattering relation and the information we used in the Hamiltonian flow, we can
establish a Lipschitz stability estimate for the corresponding lens rigidity problem.
14

Chapter 3
Weighted X-ray transform for scale
functions

3.1

Introduction

The X-ray transform is an integral transform first introduced by Fritz John in 1938. In
the simplest form, it can be stated as follows. Let f be a compactly supported function
in Rd . for each straight line passing through point z with direction ω, define
∫
I(f )(z, ω) =

f (z + tω)dt,

z ∈ Rd , ω ∈ Sd−1 .

I(f ) is called the X-ray transform of the function f . There are two nature questions for
the transform. One is the uniqueness, namely, given I(f )(z, ω) for all or some (z, ω), can
we recover f ? The other is the stability, namely, how perturbations in the data I(f ) affect
the reconstructed f .
The X-ray transform has important applications in the field of medical imaging. Xray computed tomography, also called computed tomography (CT), utilizes computerprocessed X-rays to produce tomographic images or slices of specific areas of human body.
In that case, f represents the density of an inhomogeneous medium and I(f ) represents
15

the scattered data of a tomographic scan. Besides medical imaging, the X-ray transform
also finds applications in the inverse kinematic problems of seismic and nondestructive
materials testing.
The X-ray transform coincides with the Radon transform in the two dimensional case.
The Radon transform was first introduced by Johann Radon in 1917, who also derived
a formula for the inverse transform. We remark that both the X-ray transform and the
Radon transform belong to the broad field of integral geometry. We refer to [44] for more
detail.
We shall present some basics about the X-ray transform in this chapter. Section
3.2 and 3.3 studies the case in the Euclidean space with constant and variable weight
respectively, while the case in simple Riemannian manifold is studied in Section 3.4.
These results gives necessary preliminaries for us to understand the stability results we
derived for an X-ray transform in the case when the Riemannian manifold is not simple
in Chapter 6.

3.2

X-ray transform in the Euclidean space

We study the X-ray transform in the Euclidean space with constant weight in this section.
We first derive a formula for the inverse of the transform. We start by parameterizing the
set of straight lines by

Sd−1 × Rd−1 = {(θ, x) : θ ∈ Sd−1 , ⟨x, θ⟩ = 0}.

16

We then define the X-ray transform in the following form
∫
I(f )(θ, x) =

f (x + tθ)dt.

We have the following important theorem which is also called the Projection-Slice
Theorem.
Theorem 3.2.1. The following identity holds

I(f )∧ (θ, ξ) = (2π)1/2 f ∧ (ξ), ξ ⊥ θ,

where the fourier transform on the left hand side is the (d-1)-dimensional Fourier transform in x ∈ θ⊥ , namely,

I(f )∧ (θ, ξ) =

∫
x∈θ⊥

e−ix·ξ I(f )(θ, x)dx,

for ξ ⊥ θ.

while the Fourier transform on the right hand side is the usual d-dimensional Fourier
transform in x ∈ Rd .
We now introduce the adjoint of the X-ray transform. We define

I † g(x) =

∫
Sd−1

g(θ, Eθ x)dθ,

where Eθ is the orthogonal projection onto θ⊥ , i.e. Eθ x = x − ⟨x, θ⟩θ.
With the help of the adjoint operator I † , we obtain the following recovering formula.

17

Theorem 3.2.2.
f=

1
I † R(I(f )),
2π|Sd−1 |

where the operator R is defined by (Rh)∧ (θ, ξ) = |ξ|h∧ (θ, ξ).
As a consequence of the recovering formula above, we have the following uniqueness
result on the inverse of the X-ray transform.
Theorem 3.2.3. Let d = 3, and assume that each great circle on S2 meets S0 . Then, f
is uniquely determined by I(f )(θ, ·) for θ ∈ S0 .
The condition on S0 above is called Orlov’s completeness condition. An obvious
example of S0 satisfying Orlov’s completeness condition is the great circle.

3.3

X-ray transform with weight in the Euclidean space

We now study the X-ray transform with weight in Rd . Let M be a convex domain in Rd
and let w = w(x, θ) be a smooth function defined in M × Sd−1 . Assume that suppf ⊂ M .
Define the weighted X-ray transform by
∫
Iw (f )(x, θ) =

w(x + tθ, θ)f (x + tθ)dt,

, x ∈ Rd , θ ∈ Sd−1 .

We are interested in the uniqueness and stability of recovering f from Iw f . Unfortunately, for some weight, the uniqueness fails, as is shown by the following counterexample
first constructed by Boman in 1993.
18

Theorem 3.3.1. Let B be the unit disk in the plane. There exist functions w ∈ C ∞ (B)
and f ∈ C ∞ (B) with the property that w > c0 > 0, f ̸= 0, but

Iw (f ) ≡ 0.

Although the uniqueness does not hold in the case of a general weight. We still can
study the stability of the inverse transform up to a finite dimensional space. This is
because of the fact that the kernel of the X-ray transform can be shown to be of finite
dimensional for general smooth weight.
We now focus on the stability of the X-ray transform. Define

S− ∂M = {(x, ω) : x ∈ ∂M, |ω| = 1, ⟨ω, ν(x)⟩ < 0},
SM = {(x, θ) : x ∈ M, |θ| = 1}.

The set S− ∂M parameterizes all straight lines passing through M . We define a measure µ on S− ∂M by
dµ(x, ω) = |⟨ω, ν(x)⟩|dS(x)dσ(ω),
where dS and dσ are the Lebesgue measures in ∂M and Sd−1 . We have the following
elementary results about the X-ray transform and its adjoint.
Lemma 3.3.1. Iw is a bounded linear operator from L2 (M ) to L2 (S− ∂M, dµ).
†

Lemma 3.3.2. The adjoint operator Iw has the following representation

I † g(x) =

∫
Sd−1

g ♯ (x, θ)w(x, θ)dσ(θ),
19

where g ♯ (x, θ) is defined as the function that is constant along the ray and is equal to
g(x, θ) on S− ∂M .
†

We define the normal operator Nw to be Iw Iw . We can show that Nw has the following
expression

∫
Nw f (x) = cn

W (x, y)f (y)
dy,
|x − y|n−1

where

W (x, y) = w(x, −
¯

x−y
x−y
x−y
x−y
)w(y, −
) + w(x,
¯
)w(y,
).
|x − y|
|x − y|
|x − y|
|x − y|

We now present some properties of the normal operator Nw .
Theorem 3.3.2. Nw is a elliptic ΨDO of order −1, provided the following elliptic condition on w is satisfied:

∀ (x, ξ) ∈ M × Sd−1 ,

∃ θ ∈ Sd−1 such that θ ⊥ ξ and w(x, θ) ̸= 0.

Moreover, in that case there exists a constant C > 0 such that

∥f ∥L2 (M ) ≤ C(∥Nw f ∥H 1 (M ) + ∥f ∥H −s (M ) ),

∀f ∈ L2 (M )

for any s > 0.
Proof. See [52].
As a consequence of the above theorem, we have the following corollary.

20

Theorem 3.3.3. Let w ∈ C ∞ (M × Sd−1 ) satisfy the elliptic condition. Assume that Iw
is injective on C ∞ (M ). Then Nw : L2 (M ) → H 1 (M ) is onto, and there exists a constant
C > 0 such that
∥f ∥L2 (M ) ≤ C∥Nw f ∥H 1 (M ) ,

∀f ∈ L2 (M ).

Moreover, the above estimation remains true under small C 1 perturbations of w.

3.4

X-ray transform with weight on Riemannian Manifold

We consider the X-ray transform with weight defined on a Riemannian manifold in this
section. Let (M, g) be a Riemannian Manifold with boundary ∂M . Assume that ∂M is
strictly convex. Define

S− ∂M = {(x, ω) ∈ T M : x ∈ ∂M, |ω| = 1, ⟨w, ν(x)⟩ < 0},
SM = {(x, ω) ∈ T M : x ∈ M, |ω| = 1}.

Define measure µ on S− ∂M by

dµ(x, ω) = |⟨ω, ν(x)⟩|dSx (x)dx σ(ω),

where dx S(x) and dx σ(ω) are in surface measure in ∂M and Sx M in the metric, respectively. dσx (ω) = (det g)1/2 dσ0 (ω), where dσ0 (ω) is the Lebesgue measure in Sd−1 .
For each (x, ω) ∈ S− ∂M , denote γx,ω (t) the unit speed geodesic through (x, ω). We
21

call M non-trapping if γx,ω (t) hit ∂M at finite time for all (x, ω) ∈ S− ∂M .
We define the X-ray transform on M with weight w by
∫
Iw f (z, ω) =

w(γx,ω (t), γx,ω (t))f (γx,ω (t))dt.
˙

We now present some basic results on the uniqueness and stability of the inverse of
the above X-ray transform.
†

Theorem 3.4.1. Let (M, g) be a simple Riemannian manifold, then Iw Iw is an elliptic
ΨDO of order −1, provided the following elliptic condition on w is satisfied:

∀ (x, ζ) ∈ T ∗ M,

∃ θ ∈ Tx M such that θ ⊥ ζ and w(x, θ) ̸= 0.

Theorem 3.4.2. Let (M, g) be a simple Riemannian manifold, and let w ≡ 1, then
1. I is injective;
2. ∥f ∥L2 (M ) ≤ C∥Nw f ∥H 1 (M ) .
Proof. See [22, 23, 15] and [47].
Theorem 3.4.3. Let (M, g) be a simple Riemannian manifold. Assume the ellipticity
condition on the weight w, then
1. Iw has a finitely dimensional smooth kernel;
2. ∥f ∥L2 (M ) ≤ C∥Nw f ∥H 1 (M ) ,

∀f ∈ (KerI)⊥ ;

3. If I is injective, then ∥f ∥L2 (M ) ≤ C∥Nw f ∥H 1 (M )
Proof. See [47].
22

∀f ∈ L2 (M ).

Chapter 4
Preliminaries
Starting from this chapter, we study the main inverse problem introduced in section
1.1 in the following chapters: 5, 6, 7 and 8. For the convenience of reading, we first fix
notations and definitions in this chapter. They will be used throughout the rest of the
thesis. Let Ω be a strictly convex smooth domain in Rd with boundary Γ. Let c be a
smooth velocity field defined in Ω which is equal to one near the boundary. Then c has
natural extension to Rd . Throughout the paper, we always use the natural coordinate
system of the cotangent bundle T ∗ Rd in which we write (x, ξ) for the co-vector ξj dxj in
∗
Tx Rd . For ease of notation, we also use ξ for the co-vector ξj dxj . The meaning of ξ

should be clear from the context. The velocity field c introduces a Hamiltonian function
∗
Hc (x, ξ) = 1 c2 (x)|ξ|2 to T ∗ Rd . It also defines a norm to each cotangent space Tx Rd by
2

|ξ|c = c(x)|ξ|,

∗
for ξ ∈ Tx Rd .

Here, | · | stands for the usual Euclidean norm in Rd , while | · |c stands for the norm
in T ∗ Rd induced by the function c. When there is no other velocity field in the context,
we drop the subscript c and write ∥ · ∥ instead.
t
Denote the corresponding Hamiltonian flow by Hc , i.e. for each (x0 , ξ0 ) ∈ T ∗ Rd ,

23

t
Hc (x0 , ξ0 ) = (x(t, x0 , ξ0 ), ξ(t, x0 , ξ0 )) solves the following equations:

∂Hc
= c2 · ξ, x(0) = x0 ,
∂ξ
∂Hc
1
˙
= − ∇c2 · |ξ|2 , ξ(0) = ξ0 .
ξ = −
∂x
2

x =
˙

(4.1)
(4.2)

We call (x(·, x0 , ξ0 ), ξ(·, x0 , ξ0 )) the bicharateristic curve emanating from (x0 , ξ0 ) and
t
x(·, x0 , ξ0 ) the geodesic. By the assumptions on c, the flow Hc is defined for all t ∈ R.
t
Note that the flow Hc is also well-defined on the cosphere bundle S ∗ Rd = {(x, ξ) : x ∈

R, |ξ|c = 1}.
We say that a velocity field c is non-trapping in Ω for time T > 0 if the following
condition is satisfied:
T
Hc (S ∗ Ω)

∩

S ∗ Ω = ∅.

(4.3)

Denote

∗
S+ Γ = {(x, ξ) : x ∈ Γ, |ξ|c = 1, ⟨ξ, ν(x)⟩ > 0};
∗
S− Γ = {(x, ξ) : x ∈ Γ, |ξ|c = 1, ⟨ξ, ν(x)⟩ < 0}.

Assume that the velocity field c is non-trapping in Ω for time T ; we now define the
∗
∗
∗
scattering relation Sc : S− Γ → S+ Γ. For each (x0 , ξ0 ) ∈ S− Γ, let l(x0 , ξ0 ) be the first

moment that the geodesic x(·, x0 , ξ0 ) hits the boundary Γ. Define

l(x ,ξ )
Sc (x0 , ξ0 ) = Hc 0 0 (x0 , ξ0 ).

24

For future reference, we define l− : S ∗ Ω → (−∞, 0] by letting l− (x, ξ) be the first
∗
negative moment that the bicharacteristic curve Ht (x, ξ) hits the boundary S− Γ and
∗
τ : S ∗ Ω → S− Γ by

τ (x, ξ) = Hl− (x,ξ) (x, ξ).
We remark that l− (·) and τ (·) are well-defined by the assumption (4.3).
We now introduce the class of admissible velocity fields that are considered in the
paper.
Definition 4.0.1. Let M0 , ϵ0 and T be positive numbers. A velocity field c is said to
belong to the admissible class A(M0 , ϵ0 , Ω, T ) if and only if the following three conditions
are satisfied:
1
1. c ∈ C 3 (Rd ), 0 < M ≤ c ≤ M0 , and ∥c∥C 3 (Rd ) ≤ M0 ;
0

2. the support of c − 1 is contained in the set Ω0 =: {x ∈ Ω : dist(x, Γ) > ϵ0 };
3. the Hamiltonian Hc is non-trapping in Ω for time T .
By Condition 2 above, we can find two small positive constants ϵ∗ and ϵ1 , both de∗
pending on ϵ0 , such that for any (x0 , ξ0 ) ∈ S− Γ, if

{Ht (x0 , ξ0 ) : t ∈ (0, l(x0 , ξ0 ))}

25

∩

S ∗ Ω0 ̸= ∅,

then

⟨ξ0 , ν(x0 )⟩ ≤ −ϵ∗ ,

(4.4)

⟨ξ1 , ν(x1 )⟩ ≥ ϵ∗ ,

(4.5)

l(x0 , ξ0 ) ≥ ϵ1 ,

(4.6)

where (x1 , ξ1 ) = Sc (x0 , ξ0 ).
Finally, we remark that we set up the discussion in the paper in the cotangent space
T ∗ Rd . But one can also set up the discussion in the tangent space T Rd , see for instance
[45], [53]. The equivalence of the two setups can be seen from the procedure of “raising and
lowing indices” in Riemannian geometry. We choose the cotangent setup mainly because
the following three reasons. First, it is more natural to the construction of Gaussian
beams. Second, the classification result of singular Lagrangian maps is more complete
than that of singular exponential maps in the literature, though these two problems are
equivalent in Riemannian manifold. Finally, it is more natural to study caustics in the
cotangent space.

26

Chapter 5
Sensitivity of recovering scattering
relations from their induced DDtN
maps
We study the sensitivity of recovering scattering relations from their induced Boundary
DDtN maps in this chapter.

5.1

Gaussian beam solutions to the wave equation

We first construct Gaussian beam solutions to the wave equation system (1.1)-(1.3) in this
section. Let c be a velocity field in the class A(ϵ0 , Ω, M0 , T ). We now construct a Gaussian
∗
beam in Rd . Following [36], we define G(x, ξ) = c(x)|ξ|. For a given (x0 , ξ0 ) ∈ S− Γ, let

27

(x(t), ξ(t), M (t), a(t)) be the solution to the following ODE system:

x = Gp ,
˙
˙
ξ = −Gx ,

x(t0 ) = x0 ,
ξ(t0 ) = ξ0 ,

√
†
˙
M = −Gxξ M − M Gξx − M Gξξ M − Gxx , M (t0 ) = −1 · Id,
d
a
†
†
a = − (c2 trace(M ) − Gx Gξ − Gξ M Gξ ), a(t0 ) = λ 4 .
˙
2G

(5.1)
(5.2)

The corresponding Gaussian beam with frequency λ (λ ≫ 1) is given as follows

g(t, x, λ) = a(t)eiλτ (t,x)

where τ (t, x) = ξ(t) · (x − x(t)) + 1 (x − x(t))† M (t)(x − x(t)).
2
Let the beam g impinge on the surface Γ, we want to construct the reflected beam
g − . Without loss of generality, we may assume that the ray x(t) hits Γ at the point
x(t1 ) = x1 . Write ξ(t1 ) = ξ1 . We parameterize Γ in a neighborhood of x1 , say V (x1 ),
by a smooth diffeomorphism F : U (x1 ) → V (x1 ), where U (x1 ) is a neighborhood of the
origin in Rd−1 . We require that F (0) = x1 . With the coordinate x = F (y), we can
rewrite functions restricted to the boundary Γ. For example, we rewrite

g(t, x) = g(t, F (y)) = g (t, y),
ˆ

τ (t, x) = τ (t, F (y)) = τ (t, y),
ˆ

for x ∈ V (x1 ).

We next derive formulas for τ (t, y) and g (t, y). For this, we need to calculate τ (t1 , 0),
ˆ
ˆ
ˆ

28

∂ τ (t , 0), ∂ τ (t , 0) and M (t ) =: ∂ 2 τ (t , 0). In fact, by direct calculation, we have
ˆ
ˆ
ˆ
ˆ 1
∂t 1
∂y 1
∂(t,y)2 1

∂τ
ˆ
(t , 0) = −1,
∂t 1

∂τ
ˆ
∂F
(t1 , 0) = (
(0))† ξ1 .
∂y
∂y

Moreover, the imaginary part and real part of the matrix


ˆ
ℑM (t1 ) = 

∂ 2 τ (t , 0) are given below
ˆ
∂(t,y)2 1



†
c4 (x1 )ξ1 ℑM (t1 )ξ1

†
−c2 (x1 )ξ1 ℑM (t1 ) · ∂F (0) 
∂y

−c2 (x1 )( ∂F (0))† ℑM (t1 )ξ1
∂y

( ∂F (0))† ℑM (t1 ) ∂F (0)
∂y
∂y



= R(t1 , x1 )† ℑM (t1 )R(t1 , x1 ),


ˆ
ℜM (t1 ) = R(t1 , x1 )† ℜM (t1 )R(t1 , x1 ) + 



−(∇ ln c(x1 ))† ∂F (0) 
∂y
,

2F
∂
−( ∂F (0))† ∇ ln c(x1 )
(0)ξ1
2
∂y
1
2
2 ∇c (x)ξ1

∂y

where R(t1 , x1 ) = (c2 (x1 )ξ1 , ∂F (0)).
∂y
We claim that
ˆ
ℑM (t1 ) > 0.
Indeed, note that the column vectors in the matrix ∂F (0) are linearly independent and
∂y
hence span the tangent space of the surface Γ at the point x1 . By (4.5), ξ1 forms a nonzero
angle with the tangent space and thus is linearly independent with all the column vectors
in the matrix ∂F (0). Therefore the matrix R(t1 , x1 ) is invertible, and our claim follows.
∂y
Using (4.4) and (4.5), we further have

ˆ
ℑM (t1 ) > C

29

(5.3)

for some C > 0 depending on ϵ0 and M0 .
2ˆ
ˆ
ˆ
Now, we have calculated τ (t1 , 0), ∂ τ (t1 , 0), ∂ τ (t1 , 0) and ∂ τ 2 (t1 , 0). It follows that
ˆ
∂t
∂y
∂(t,y)

τ (t1 , y) = τ (t1 , 0) + (
ˆ
ˆ

∂τ
ˆ
∂ 2τ
ˆ
(t1 , 0))† (t − t1 , y) + (t − t1 , y)
(t , 0)(t − t1 , y)†
2 1
∂(t, y)
∂(t, y)

+O(|(t − t1 , y)|3 )
= ⟨(−1,

∂F †
ˆ
(0) ξ1 ), (t − t1 , y)⟩ + (t − t1 , y)M (t1 )(t − t1 , y)† + O(|(t − t1 , y)|3 ).
∂y

We proceed to construct the reflected beam g − . Write
−
g − (t, x, λ) = a− (t)eiλτ (t,x)

with
1
τ − (t, x) = ξ − (t) · (x − x− (t)) + (x − x− (t))† M − (t)(x − x− (t)).
2
We need to find (x− (t1 ), ξ − (t1 ), a− (t1 ), M − (t1 )) such that the g − +g ≈ 0 on the boundary.
Following [3], we impose the following condition

α ˆ
α ˆ
∂t,y τ (t1 , 0) = ∂t,y τ − (t1 , 0),

for all |α| ≤ 2.

(5.4)

The above condition with |α| = 0 gives that τ − (t1 , 0) = τ (t1 , 0) = 0; with |α| = 1 gives
ˆ
ˆ
that
(

∂F
∂F
−
(0))† ξ1 = (
(0))† ξ1 ,
∂y
∂y

(5.5)

−
where ξ1 = ξ − (t1 ). Since the column vectors in ( ∂F (0))† spans the tangent space Tx1 Γ,
∂y

30

−
we see that the tangential component of ξ1 and ξ1 are equal. Besides, note that |ξ1 | =
−
|ξ1 | = 1. Thus,
−
ξ − (t1 ) = ξ1 = ξ1 − 2⟨ξ1 , ν(x1 )⟩ν(x1 ).

ˆ
ˆ
Condition (5.4) with |α| = 2 gives that M − (t1 ) = M (t1 ). Recall the relation between
ℑM − (t1 ) and ℑM (t1 ), ℜM − (t1 ) and ℜM (t1 ), we have the following two identities:

R(t1 , x1 )† ℑM (t1 )R(t1 , x1 ) = R(t1 , x1 )† ℑM − (t1 )R(t1 , x1 ),
R(t1 , x1 )† ℜM (t1 )R(t1 , x1 ) = R(t1 , x1 )† ℑM − (t1 )R(t1 , x1 )




−
1
0
 − 2 ∇c2 (x)(ξ1 − ξ1 )


.
+

∂ 2 F (0)(ξ − − ξ )
0
1
1
∂y 2

Solving the above equations, we obtain ℑM − (t1 ) and ℜM − (t1 , ) and hence M − (t1 ). Finally, set a− (t1 ) = −a(t1 ). Then all of the four components of (x− (t1 ), ξ − (t1 ), a− (t1 ), M − (t1 ))
are constructed. We then solve an ODE system to get (x− (t), ξ − (t), M − (t), a− (t)) as we
did for the beam g. This completes the construction for the reflected beam g − .
We now present some properties about the constructed beam. The following lemma
is crucial in the subsequent estimates. We refer to [37] for the proof.
Lemma 5.1.1. Both the matrices M (t) and M − (t) are uniformly bounded for t ∈ [0, T +
ϵ1 ]. Moreover, there exists C > 0, depending on M0 and ϵ0 , such that ℑM (t) > C and
ℑM − (t) > C for all t ∈ [0, T + ϵ1 ].
We next introduce two auxiliary beams below

τ
g∗ (t, y, λ) = a(t1 )eiλˆ∗ ,
ˆ

−

τ
g∗ (t, y, λ) = a− (t1 )eiλˆ∗ ,
ˆ−

31

where
⟨
⟩
∂F †
ˆ
(−1,
(0) ξ1 ), (t − t1 , y) + (t − t1 , y)M (t1 )(t − t1 , y)† ,
∂y
⟩
⟨
∂F † −
ˆ
(0) ξ1 ), (t − t1 , y) + (t − t1 , y)M − (t1 )(t − t1 , y)† .
τ∗ = (−1,
ˆ−
∂y
τ∗ =
ˆ

It is clear that τ∗ = τ∗ and g∗ = −ˆ∗ .
ˆ
ˆ−
ˆ
g−
Lemma 5.1.2.
√
g (t, y, λ) = g∗ (t, y, λ) + O( λ) in H 1 ((3ϵ1 /4, t1 + ϵ1 /2) × U (x1 )),
ˆ
ˆ
√
g − (t, y, λ) = g∗ (t, y, λ) + O( λ)
ˆ
ˆ−

in H 1 ((3ϵ1 /4, t1 + ϵ1 /2) × U (x1 )).

(5.6)
(5.7)

Proof:. We only show (5.6), since (5.7) follows in a similar way. For simplicity, denote
D = (3ϵ1 /4, t1 + ϵ1 /2) × U (x1 ). We first show that
1
g (t, y, λ) = g∗ (t, y, λ) + O( √ ) in L2 (D).
ˆ
ˆ
λ

(5.8)

Indeed, by direct calculation,

τ
τ
τ
g (t, y, λ) − g∗ (t, y, λ) = (a(t) − a(t1 ))eiλˆ∗ + a(t)(eiλˆ − eiλˆ∗ ).
ˆ
ˆ

It suffices to show that

τ
R1 := ∥(a(t) − a(t1 ))eiλˆ∗ ∥L2 (D)
τ
τ
R2 := ∥a(t)(eiλˆ − eiλˆ∗ )∥L2 (D)

32

1
√ ,
λ
1
√ .
λ

(5.9)

d

We first estimate R1 . By Lemma 3.1 in [7], we have |a(t)| ≈ λ 4 . By equation (5.2), we
d

further derive that |a(t)| ≈ λ 4 , thus
˙

a(t) − a(t1 ) =

∫ 1
0

d

a(t1 + s(t − t1 )) ds(t − t1 ) = O(λ 4 )|t − t1 |.
˙

Therefore,

τ
∥(a(t) − a(t1 ))eiλˆ∗ ∥2 2
L (D)

This proves R1

∫
D

d

†

λ 2 (t − t1 )2 e−λ(t−t1 ,y)ℑM (t1 )(t−t1 ,y) dtdy
ˆ

1
.
λ

1
√ .
λ

We next estimate R2 . Write τ = τ∗ + δˆ, then δˆ = O(|(t − t1 , y)|3 ) and hence
ˆ
ˆ
τ
τ
τ
|1 − eiλδˆ |

λ · O(|(t − t1 , y)|3 ). It follows that
∫
R2 ≤

∫D
D

τ
τ
|a(t)eiλˆ∗ |2 · |1 − eiλδˆ | dtdy
d

†

λ 2 · λ · |(t − t1 , y)|3 e−2λ(t−t1 ,y)ℑM (t1 )(t−t1 ,y) dtdy
ˆ

1
.
λ

This completes the proof of (5.8).
We now proceed to show (5.6). By direct calculation,

g
∂τ
ˆ
∂ τ∗
ˆ
∂ˆ ∂ˆ∗
g
−
= iλ
· g − iλ
ˆ
· g∗
ˆ
∂y
∂y
∂y
∂y
∂τ
ˆ ∂ τ∗
ˆ
∂ τ∗
ˆ
= iλ(
−
) · g + iλ
ˆ
· (ˆ − g∗ )
g ˆ
∂y
∂y
∂y
ˆ
ˆ
One can check that ∂ τ − ∂ τ∗ = O|(t − t1 , y)|2 , then a similar argument as used in the
∂y
∂y

33

estimate of R1 above shows that
∂τ
ˆ ∂ τ∗
ˆ
−
) · g ∥2 2
ˆ
L (D)
∂y
∂y

1.

∂ τ∗
ˆ
· (ˆ − g∗ )∥2 2
g ˆ
L (D)
∂y

λ.

∥λ(

Besides, (5.8) implies that

∥λ

Combining these two estimates together, we conclude that

∥

∂ˆ ∂ˆ∗ 2
g
g
−
∥
∂y
∂y L2 (D)

λ.

∥

∂ˆ ∂ˆ∗ 2
g
g
−
∥
∂t
∂t L2 (D)

λ.

Similarly, we can show that

This completes the proof of (5.6) and hence the lemma.
Note that ∥ˆ(t, y, λ)∥L2 ((t −ϵ /2,t +ϵ /2)×U (x )) ≈ 1. As a direct consequence of
g
1 1
1 1
1
Lemma 5.1.2, we obtain the following norm estimate for the beam g restricted to the
boundary Γ.
Lemma 5.1.3.
∥g(·, ·, λ)∥L2 ((t −ϵ /2,t +ϵ /2)×V (x )) ≈ 1.
1 1
1 1
1

(5.10)

We now present an H 1 -norm estimate for g − + g and an approximation for the Neu∂g
mann data ∂ν

−

∂g
+ ∂ν on the boundary.

34

Lemma 5.1.4.
√
g − (t, x, λ) + g(t, x, λ) = O( λ)

in H 1 ((3ϵ1 /4, t1 + ϵ1 /2) × V (x1 ));

√
∂g − ∂g
= 2iλg · ⟨ξ1 , ν(x1 )⟩ + O( λ)
+
∂ν
∂ν

(5.11)
(5.12)

in L2 ((3ϵ1 /4, t1 + ϵ1 /2) × V (x1 )).

Proof:. Denote D = (3ϵ1 /4, t1 + ϵ1 /2) × U (x1 ) again. We first show (5.11). Since x
is restricted to V (x1 ) ⊂ Γ, it suffices to show that
√
g − (t, y, λ) + g (t, y, λ) = O( λ) in H 1 (D).
ˆ
ˆ

But this is a direct consequence of Lemma 5.1.2 and the fact that g∗ = −ˆ∗ .
ˆ−
g
We now prove (5.12). By direct calculate

∂g
∂
∂τ
(t, x) =
(a(t)eiλτ (t,x) ) = iλg ·
∂ν
∂ν
∂ν
= iλg · ⟨ξ(t) + M (t)(x − x(t), ν(x)⟩
= iλg · ⟨ξ(t1 ), ν(x1 )⟩ + iλg · (⟨ξ(t), ν(x)⟩ − ⟨ξ(t1 ), ν(x1 )⟩)
+iλg · ⟨M (t)(x − x(t), ν(x)⟩.

Note that in the coordinate x = F (y),

|⟨M (t)(x − x(t)), ν(x)⟩| = O(|(t − t1 , y)|),
|⟨ξ(t), ν(x)⟩ − ⟨ξ(t1 ), ν(x1 )⟩| = O(|(t − t1 , y)|).

35

It follows that

∥g · (⟨ξ(t), ν(x)⟩ − ⟨ξ(t1 ), ν(x1 )⟩)∥2 2

L (D)

∥g · ⟨M (t)(x − x(t), ν(x)⟩∥2 2

L (D)

1
,
λ
1
.
λ

Thus
√
∂g
(t, x) = iλg · ⟨ξ(t1 ), ν(x1 )⟩ + O( λ).
∂ν
Similarly,
√
∂g −
(t, x) = iλg − · ⟨ξ − (t1 ), ν(x1 )⟩ + O( λ).
∂ν
Finally, using (5.11) and the fact that ⟨ξ − (t1 ), ν(x1 )⟩ = −⟨ξ(t1 ), ν(x1 )⟩, we conclude that
(5.12) holds. This completes the proof of the lemma.
Now, we are ready to construct Gaussian beam solutions to the initial boundary
∞
value problem of the wave system (1.1)-(1.3). We first choose χϵ1 (t) ∈ C0 (R) such
∪
that χϵ1 (t) = 1 for t ∈ (ϵ1 /4, ϵ1 /2) and χϵ1 (t) = 0 for t ∈ (−∞, 0) (3ϵ1 /4, ∞). Let
ϵ

1
ϵ ·ξ
∗
∗
(x0 , ξ0 ) ∈ S− Γ and (x∗ , ξ0 ) = H− 4 (x0 , ξ0 ) = (x0 − 14 1 , ξ0 ). Let g be the Gaussian beam
0
d

∗
constructed with the initial data x(0) = x∗ , ξ(0) = ξ0 , M (0) = i · Id and a(0) = λ 4 . The
0
l(x ,ξ )
ϵ1
beam g is reflected by Γ at (x1 , ξ1 ) = Sc (x0 , ξ0 ) = Hc 0 0 (x0 , ξ0 ) at t1 = l(x0 , ξ0 ) + 4 .

We construct the reflected beam g − by the preceding procedure. Let u be the exact
solution to the wave system (1.1)-(1.3) with

f (t, x, λ) = g(t, x, λ) · χϵ1 (t).

36

Then u = g + g − + R, where the remaining term R satisfies the following equation
system

PR = −P(g + g − ),

(t, x) ∈ Ω × (0, t1 + ϵ1 /2),

R(0, x, λ) = −(g + g − )(0, x, λ),
−
Rt (0, x, λ) = −(gt + gt )(0, x, λ),

x ∈ Ω,
x ∈ Ω,

R(t, x, λ) = −g(t, x, λ)(1 − χϵ1 (t)) − g − (t, x, λ),

(t, x) ∈ (0, t1 + ϵ1 /2) × Γ.

Here P stands for the wave operator 21 ∂tt − ∆.
c (x)
Lemma 5.1.5.
∥

√
∂R
∥L2 ([0,t +ϵ /2]×Γ) ≤ C λ
1 1
∂ν

for some constant C > 0 depending on ϵ0 and M0 .
Proof. We apply Theorem 4.1 in [31] to derive the estimate. Note that the compatibility condition is satisfied on the boundary at time t = 0. It remains to show that the
following four estimates hold:
√
λ,

(5.13)

∥(g + g − )(0, ·, λ)∥H 1 (Ω)

√
λ,

(5.14)

−
∥(gt + gt )(0, ·, λ)∥L2 (Ω)

√
λ,

(5.15)

√
λ.

(5.16)

∥P(g + g − )∥C([0,t +ϵ /2];L2 (Ω))
1 1

∥g(t, x, λ)(1 − χϵ1 (t)) − g − (t, x, λ)∥H 1 ([0,t +ϵ /2]×Γ)
1 1

First, (5.13) follows from the standard estimate for Gaussian beams, see for example [7].
37

We next show (5.14). By Lemma 5.1.1, there exists a constant C > 0 depending on M0
and ϵ0 such that the following two inequalities hold

|g(t, x, λ)|

d
− 2
λ 4 · e−Cλ·|x−x (t)| ,

|g − (t, x, λ)|

d
− 2
λ 4 · e−Cλ·|x−x (t)| .

Thus the beam g and g − are exponentially decaying away from the ray x(t) and x− (t)
respectively. Using this property, it is straightforward to show that ∥g(0, ·, λ)∥H 1 (Ω)
and ∥g − (0, ·, λ)∥H 1 (Ω)

1, whence (5.14) and (5.15) follows.

Now, we show (5.16). We divide the domain (0, t1 + ϵ1 /2) × Γ into three parts:

Σ1 = (0, ϵ1 /2) × Γ,

Σ2 = (ϵ1 /2, t1 − ϵ1 /2) × Γ,

Σ3 = (t1 − ϵ1 /2, t1 + ϵ1 /2) × Γ.

We show that inequality (5.16) holds on each part.
For (t, x) ∈ Σ1 , we have 1 − χϵ1 (t) = 0. Consequently,
g(t, x)(1 − χϵ1 (t)) − g − (t, x, λ) = g − (t, x, λ).
By the exponential decaying property of g − , we obtain that

∥g(t, x)(1 − χϵ1 (t)) − g − (t, x, λ)∥H 1 (Σ )
1

38

√
1.

1

For (t, x) ∈ Σ2 , by the exponential decaying property for both g and g − again, we obtain

∥g(t, x, λ)(1 − χϵ1 (t)) − g − (t, x, λ)∥H 1 (Σ )
2

1.

ϵ1
ϵ1
ϵ1
3ϵ
Finally, for (t, x) ∈ Σ3 , note that t1 − 2 = l(x0 , ξ0 ) + 4 − 2 ≥ 41 . We can apply

Lemma 5.1.4 to the part x ∈ V (x1 ) and the exponential decaying property for both g and
g − to the remaining part to conclude that

∥g(t, x, λ)(1 − χϵ1 (t)) − g − (t, x, λ)∥H 1 (Σ )
3

√
λ

This completes the proof of (5.16) and hence the lemma.

5.2

The sensitivity result

It is known that the DDtN map Λc determines the scattering relation Sc uniquely [35].
We show that the following sensitivity result of recovering the scattering relation from
the DDtN map holds.
Theorem 5.2.1. Let c and c be two velocity fields in the class A(ϵ0 , Ω, M0 , T ). Then
˜
there exists a constant δ > 0 such that

Sc = Sc
˜

if ∥Λc − Λc ∥H 1 [0,3ϵ /4]×Γ→L2 ([0,T +ϵ ]×Γ) ≤ δ.
˜
1
1
0

39

l(x ,ξ )
∗
Proof. For any (x0 , ξ0 ) ∈ S− Γ, let (x1 , ξ1 ) = Sc (x0 , ξ0 ) = Hc 0 0 (x0 , ξ0 ) and
˜

l(x ,ξ )
(˜1 , ξ1 ) = Sc (x0 , ξ0 ) = Hc 0 0 (x0 , ξ0 ). We need to show that
x ˜
˜
˜

(l(x0 , ξ0 ), x1 , ξ1 ) = (˜ 0 , ξ0 ), x1 , ξ1 )
l(x
˜ ˜

if ∥Λc − Λc ∥ is sufficiently small. We do this in the following steps.
˜
ϵ1
ϵ1
˜
Step 1. Let t1 = l(x0 , ξ0 ) + 4 and t1 = ˜ 0 , ξ0 ) + 4 . Without loss of generality, we
l(x

˜
may assume that t1 ≤ t1 . Let V (x1 ) be a neighborhood of x1 in Γ which is parameterized
by a smooth function F : U (x1 ) → V (x1 ) as before. We may assume that x1 ∈ V (x1 ).
˜
Let x1 = F (δy). We construct the initial beam g, the reflected beam g − , the boundary
˜
Dirichlet data f , the solution u to the wave equation with velocity field c and remanning
˜
term R as in the previous section. We similarly construct g , g − , u and R to the system
˜ ˜ ˜
˜
with velocity field c and with boundary Dirichlet data f = f .
˜
˜
Step 2. Denote by I(t1 , ϵ1 /2) the interval (t1 − ϵ1 /2, t1 + ϵ1 /2). Since t1 ≤ t1 and
˜
l(x0 , ξ0 ) ≥ ϵ1 , we have I(t1 , ϵ1 /2) ⊂ (3ϵ1 /4, t1 + ϵ1 /2) and I(t1 , ϵ1 /2) ⊂ (3ϵ1 /4, t1 + ϵ1 /2).
Then we can apply (5.12) and Lemma 5.1.5 to obtain

∂u ∂ u
˜
−
∂ν ∂ν
˜
∂(g + g − ) ∂(˜ + g − ) ∂R ∂ R
g ˜
=
−
+
−
∂ν
∂ν
∂ν
∂ν
√
{
}
˜
= 2iλ · ⟨ξ1 , ν(x1 )⟩ · g − ⟨ξ1 , ν(˜1 )⟩ · g + O( λ)
x
˜

(Λc − Λc )f =
˜

in L2 (I(t1 , ϵ1 /2) × V (x1 )).

40

It follows that
⟨

[

⟩

(Λc − Λc )f, g L2 (I(t ,ϵ /2)×V (x )) = 2iλ ·
˜
1 1
1

⟨

⟩ ⟨ ⟩
ξ1 , ν(x1 ) · g, g L2 (I(t ,ϵ /2)×V (x ))
1 1
1

⟨
⟩ ⟨ ⟩
˜
− ξ1 , ν(˜1 ) · g , g L2 (I(t ,ϵ /2)×V (x ))
x
˜
1 1
1

]

√
+ O( λ).

Note that

|⟨(Λc − Λc )f, g⟩L2 (I(t ,ϵ /2)×V (x )) |
˜
1 1
1
≤ ∥(Λc − Λc )f ∥L2 (I(t ,ϵ /2)×V (x )) · ∥g∥L2 (I(t ,ϵ /2)×V (x ))
˜
1 1
1
1 1
1
≤ ∥(Λc − Λc )f ∥L2 ((0,T +ϵ )×Γ) · ∥g∥L2 (I(t ,ϵ /2)×V (x ))
˜
1
1 1
1
≤ ∥Λc − Λc ∥H 1 ([0,3ϵ /4]×Γ)→L2 ([0,T +ϵ ]×Γ) · ∥f ∥H 1 ([0,3ϵ /4]×Γ)
˜
1
1
1
0
0
·∥g∥L2 (I(t ,ϵ /2)×V (x ))
1 1
1
λ · ∥Λc − Λc ∥H 1 ([0,3ϵ /4]×Γ)→L2 ([0,T +ϵ ]×Γ) .
˜
1
1
0
Thus the following inequality holds

˜
|⟨ξ1 , ν(x1 )⟩ · ⟨g, g⟩L2 (I(t ,ϵ /2)×V (x )) | − |⟨ξ1 , ν(˜1 )⟩ · ⟨˜, g⟩L2 (I(t ,ϵ /2)×V (x )) |
x
g
1 1
1
1 1
1
C
(5.17)
≤ √ + C · ∥Λc − Λc ∥H 1 ([0,3ϵ /4]×Γ)→L2 ([0,T +ϵ ]×Γ)
˜
1
1
λ
0
for some constant C > 0.
Step 3. We now estimate the two terms on the left hand side of the inequality (5.17).

41

First, by (4.5) and Lemma 5.1.3, we have
⟨ ⟩
|⟨ξ1 , ν(x1 )⟩ · | g, g L2 (I(t ,ϵ /2)×V (x )) | ≈ 1.
1 1
1

(5.18)

We next estimate ⟨˜, g⟩L2 (I(t ,ϵ /2)×V (x )) . In the coordinate x = F (y), by Lemma 5.1.2,
g
1 1
1
we have

1
ˆ ˆ
ˆ ˆ
⟨g , g ⟩L2 (I(t ,ϵ /2)×U (x )) = ⟨g∗ , g∗ ⟩L2 (I(t ,ϵ /2)×U (x )) + O( √ ).
˜
˜
1 1
1
1 1
1
λ

(5.19)

ˆ
τ
˜
ˆ
Here we recall that g∗ = a(t1 )eiλˆ∗ and g∗ = a(t1 )eiλτ∗ with
ˆ
˜
˜

∂F †
ˆ
(0) ξ1 ), (t − t1 , y)⟩ + (t − t1 , y)M (t1 )(t − t1 , y)† ;
∂y
∂F
ˆ
˜
˜
ˆ
˜
˜
˜
τ∗ = ⟨(−1,
˜
(δy)† ξ1 ), (t − t1 , y − δy)⟩ + (t − t1 , y − δy)M (t1 )(t − t1 , y − δy)† .
∂y
τ∗ = ⟨(−1,
ˆ

By Lemma 3.7 in [7], we have

ˆ ˆ
|⟨g∗ , g∗ ⟩L2 (I(t ,ϵ /2)×U (x )) |
˜
1 1
1

e−c0 λ|δz|

where c0 is a positive constant depending only on ∥c∥C 3 + ∥˜∥C 3 and
c
∂F † 2
∂F
˜
˜
(δy)† ξ1 −
(0) ξ1 | .
|δz| = |t1 − t1 |2 + |δy|2 + |
∂y
∂y

42

(5.20)

It follows from (5.19) and (5.20) that

|⟨˜, g⟩L2 (I(t ,ϵ /2)×V (x )) |
g
1 1
1

1
e−c0 λ|δz| + O( √ ).
λ

(5.21)

Step 4. Combining (5.17), (5.18) and (5.21), we see that

e−c0 λ|δz|

1
C1 − C2 ∥Λc − Λc ∥H 1 ([0,3ϵ /4]×Γ)→L2 ([0,T +ϵ ]×Γ) − C3 √
˜
1
1
λ
0

for some positive constants C1 , C2 and C3 which are independent of (x0 , ξ0 ). By letting
λ → ∞, we conclude that δz = 0 if
C
∥Λc − Λc ∥H 1 ([0,3ϵ /4]×Γ)→L2 ([0,T +ϵ ]×Γ) < 1 .
˜
C2
1
1
0
C
˜
˜
Set δ = C1 . From δz = 0 it follows that t1 = t1 , δy = 0, and ∂F (0)† ξ1 − ∂F (0)† ξ1 = 0.
∂y
∂y
2

˜
˜
It remains to show that ξ1 = ξ1 . Indeed, ∂F (0)† ξ1 − ∂F (0)† ξ1 = 0 implies that the
∂y
∂y
˜
˜
tangential component of ξ1 and ξ1 are equal. Besides, ∥ξ1 ∥ = ∥ξ1 ∥. These together with
˜
(4.5) yield that ξ1 = ξ1 . This completes the proof of the theorem.

43

Chapter 6
Stability of X-ray transform in the
presence of caustics

6.1

The X-ray transform resulted from linearizing
Hamiltonian flow with respect to velocity field

We begin with the following observation.
Lemma 6.1.1. Let c and c be two velocity fields in the class A(ϵ0 , Ω, M0 , T ), then Sc =
˜
T
T
Sc if and only if Hc |S ∗ Γ = Hc |S ∗ Γ .
˜
˜ −
−

The above lemma shows the equivalence of the Hamiltonian flow and the scattering
t
relation. The next lemma shows that Hc satisfies an equivalent ordinary differential

equation (ODE) system in S ∗ Rd .
Lemma 6.1.2. Let (x0 , ξ0 ) ∈ S ∗ Rd = {(x, ξ) ∈ R2d : c(x)|ξ| = 1}, and let (x(t), ξ(t)) =
t
Hc (x0 , ξ0 ), then (x(t), ξ(t)) satisfies the following ODE system

ξ
,
|ξ|2

(6.1)

˙
ξ = b(x).

(6.2)

x =
˙

44

where b(x) = − 1 ∇ ln c2 . Conversely, if (x(t), ξ(t)) ∈ S ∗ Rd satisfies the ODE system
2
t
(6.1)-(6.2), then (x(t), ξ(t)) = Hc (x0 , ξ0 ).

We next linearize the operator which maps each velocity field to its induced Hamiltonian flow restricted to the cosphere bundle. Let c be a fixed smooth background velocity
field. Denote the perturbed velocity field and Hamiltonian flow at time T as c2 = c2 + δc2
˜
1
T
T
T
˜
and Hc = Hc + δHc respectively. Denote also that δb = − 2 ∇(ln c2 − ln c2 ) and
˜





 0
A(x, ξ) = 


ξ
∂
∂ξ ( |ξ|2 ) 


∂b
∂x

0

.

∗
For each (x0 , ξ0 ) ∈ S− Γ, let Φ(t, x0 , ξ0 ) be the solution of the following ODE system

t
˙
Φ(t) = −Φ(t)A(Hc ),

Φ(0) = Id.

By the standard linearization argument, we have

T
δHc =

T
δHc
(δb) + r(δb),
δb

where


∫ T



T
0
δHc


(δb)(x0 , ξ0 ) =
Φ−1 (T, x0 , ξ0 ) · Φ(s, x0 , ξ0 ) 
 ds
δb
0
δb(x(s, x0 , ξ0 ))

and ∥r(δb)∥L∞ ≤ C∥δb∥2 1 for some constant C > 0 depending only on ∥c∥C 3 (Rd ) .
C

45

(6.3)

Formula (6.3) motivates us to define the following geodesic X-ray transform operator
∫ T
Ic (f )(x0 , ξ0 ) =

Φ(s, x0 , ξ0 )f (x(s, x0 , ξ0 )) ds,

0

f ∈ E ′ (Ω, R2d ).

(6.4)

δHT

Then δbc (δb)(x0 , ξ0 ) = Φ−1 (T, x0 , ξ0 ) · Ic (f )(x0 , ξ0 ) with


f =


0
1 ∇(ln c2 − ln c2 )
˜
2


.

(6.5)

We introduce a matrix Φ(x, ξ) for each (x, ξ) ∈ S ∗ Ω. Let (x0 , ξ0 ) = τ (x, ξ) =
Hc−

l (x,ξ)

(x, ξ). We then define

Φ(x, ξ) = Φ(−l− (x, ξ), τ (x, ξ)).

It is clear that the following identity holds

s
Φ(Hc (x0 , ξ0 )) = Φ(s, x0 , ξ0 )

s
for all s ∈ R+ such that Hc (x0 , ξ0 ) ∈ S ∗ Ω. We can rewrite the X-ray transform operator

Ic in the following standard form
∫ T
Ic f (x0 , ξ0 ) =
=

0

s
s
Φ(Hc (x0 , ξ0 ))f (π(Hc (x0 , ξ0 ))) ds

∫ l(x ,ξ )
0 0
0

s
s
Φ(Hc (x0 , ξ0 ))f (π(Hc (x0 , ξ0 ))) ds.

(6.6)

Remark 6.1.1. Formula (6.6) is derived in the coordinate of T ∗ Rd . Hence it may not
46

be geometrically invariant.
Lemma 6.1.3. Assume that Sc = Sc , let f be defined as in (6.5), then
˜

∥f ∥2 1

∥Ic f ∥L∞

6.2

C (Ω)

.

Statement of the main results for the stability of
the geodesic X-ray transform Ic

We consider the stability estimate of the operator Ic . For simplicity, we drop the subscript
c. Define β : T ∗ Rd \{(x, 0) : x ∈ Rd } → S ∗ Rd by
)
(
ξ
.
β(x, ξ) = x,
∥ξ∥
Let π : T ∗ Rd → Rd be the natural projection onto the base space. We define ϕ :
T ∗ Rd → Rd by
ϕ(x, ξ) = π ◦ Ht=1 (x, ξ),

(x, ξ) ∈ T ∗ Rd .

We remark that ϕ defined above is equivalent to the exponential map in Riemannian
manifold.
The following result about the normal operator N = I† I is well-known.
Lemma 6.2.1. The normal operator N : L2 (Ω, R2d ) → L2 (Ω, R2d ) is bounded and has

47

the following representation
∫
Nf (x) =

∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ)) dσx (ξ),

f ∈ L2 (Ω, R2d )

(6.7)

∗
where dσx denotes the measure in the space Tx Rd induced by the velocity field c, i.e.

dσx (ξ) = c(x)d dξ, and W is defined as

W (x, ξ) =

1
{Φ† ◦ β(x, ξ) · Φ ◦ β ◦ H(x, ξ) + Φ† ◦ β(x, −ξ) · Φ ◦ β ◦ H−1 (x, −ξ)}. (6.8)
∥ξ∥d−1

Proof. See [54] or [47].
We see from (6.7) that the local property of the normal operator N restricted to a small
∗
neighborhood of x ∈ Ω is determined by the lagrangian map ϕ(x, ·) : Tx Rd → Rd . When

the map is a diffeomorphism, it is known that the operator N near x is a pseudo-differential
operator (ΨDO). However, in general case, the map may not be a diffeomorphism and
may have singular points which are called caustic vectors. The value of the map at caustic
vectors are called caustics. When caustics occur, the Schwartz kernel of the operator N
has two singularities, one is from the diagonal which contributes to a ΨDO N1 , and the
other is from the caustics which contributes to a singular integral operator N2 . The
property of N2 depends on the type of caustics. The case for fold caustics is investigated
in [54], where it is shown that fold caustics contribute a Fourier Integral Operator (FIO)
to N2 . Little is known for caustics of other type. Here we recall the following definition
of fold caustics.
Definition 6.2.1. Let f : Rd → Rn be a germ of C ∞ map at x0 , then x0 is said to be a
48

fold vector and f (x0 ) a fold caustic if the following two conditions are satisfied:
1. the rank of df at x0 equals to n − 1 and det df vanishes of order 1 at x0 ;
2. the kernel of the matrix df (x0 ) is transversal to the manifold {x : det df (x) = 0} at
x0 .
We now introduce the following concept of “operator germ” to characterize the contribution of an infinitesimal neighborhood of a caustic or a regular point to the normal
operator N.
∗
Definition 6.2.2. For each ξ ∈ Tx Rd \0, the operator germ Nξ is defined to be the

equivalent class of operators in the following form
∫
Nξ f (y) =

∗
Ty Ω

W (y, η)f (ϕ(y, η))χ(y, η) dσy (η).

(6.9)

where χ is a smooth function supported in a small neighborhood of (x, ξ) in R2d . Two
operators with χ1 and χ2 are said to be equivalent if there exists a neighborhood B(x, ξ)
∞
of (x, ξ) such that χ1 = χ2 · χ3 for some χ3 ∈ C0 (B(x, ξ)) with χ3 (x, ξ) ̸= 0.

The operator germ Nξ is said to has certain property if there exists a neighborhood
B(x, ξ) of (x, ξ) in T ∗ Rd such that the property holds for all operators of the form (6.9)
∞
with χ ∈ C0 (B(x, ξ)).

Properties of the above defined operator germ will be given in Section 6.3.
We note from the preceding discussion that it is complicated to analyze the full operator N which contains information from all geodesics. However, for a given interior point
x, to recover f or the singularity of f at x from its geodesic transform, we need only
49

∗
to select a set of geodesics whose conormal bundle can cover the cotangent space Tx Rd .

Caustics may be allowed along these geodesics as long as they are of the simplest type,
i.e fold type so that we can analyze their contributions. This idea can be carried out by
introducing a cut-off function for the set of geodesics as we do now. We remark that this
∞ ∗
idea is motivated by the work [51]. For any α ∈ C0 (S− Γ), we define

Iα f (x0 , ξ0 ) = α(x0 , ξ0 )

∫ l(x ,ξ )
0 0
0

s
s
Φ(Hc (x0 , ξ0 ))f (π(Hc (x0 , ξ0 ))) ds

(6.10)

∗
where (x0 , ξ0 ) ∈ S− Γ. Let α♯ be the unique lift of α to S ∗ Ω which is constant along

bicharacteristic curves, i.e. α♯ (x, ξ) = α ◦ τ (x, ξ) for (x, ξ) ∈ S ∗ Ω. Then α♯ is smooth in
S ∗ Ω and we have

Iα f (x0 , ξ0 ) =

∫ l(x ,ξ )
0 0
0

s
s
(α♯ · Φ)(Hc (x0 , ξ0 ))f (π(Hc (x0 , ξ0 ))) ds

(6.11)

With the original weight Φ being replaced by the new one α♯ ·Φ, we similarly can define
Nα . In fact, it is easy to check that Nα is defined as in (6.8) with W being replaced by

Wα (x, ξ) =

1
|α ◦ τ ◦ β(x, ξ)|2 Φ† ◦ β(x, ξ) · Φ ◦ β ◦ H(x, ξ)
∥ξ∥d−1
1
+
|α ◦ τ ◦ β(x, −ξ)|2 Φ† ◦ β(x, −ξ) · Φ ◦ β ◦ H−1 (x, −ξ).
∥ξ∥d−1

It can be shown that with properly chosen α, the analysis of the operator Nα becomes
possible and we can recover the singularity of f from Nα f .
We now give two definitions whose discussions are postponed to Section 6.
∗
Definition 6.2.3. A fold vector ξ ∈ Tx Rd is called fold-regular if there exists a neighbor-

50

hood U (x) of x such that the operator germ Nξ is compact from L2 (Ω0 , R2d ) to H 1 (U (x), R2d )
(or from H s (Ω0 , R2d ) to H s+1 (U (x), R2d ) for all s ∈ R).
Definition 6.2.4. A point x is called fold-regular if there exists a compact subset Z2 (x) ⊂
∗
Sx Rd such that the following two conditions are satisfied:

1. For each ξ ∈ Z2 (x), there exist only singular vectors of fold-regular type along the
ray {tξ : t ∈ R} for the map ϕ(x, ·);
∗
2. ∀ ξ ∈ Sx Rd , ∃ θ ∈ Z2 (x), such that θ ⊥ ξ.

We remark that Z2 (x) parameterizes a subset of geodesics that pass through x and
along which there exist only fold-regular caustics.
We now present the main results on the stability estimate for the geodesic X-ray
transform operator. The proofs are given in Section ?.
Theorem 6.2.1. Let x∗ be a fold-regular point, then there exist a cut-off function α ∈
∞ ∗
C0 (S− Γ), a neighborhood U (x∗ ) of x∗ , a compact operator N2,α from L2 (Ω0 , R2d ) to

H 1 (U (x∗ ), R2d ) and a smoothing operator R from E ′ (Ω, R2d ) into C ∞ (U (x∗ ), R2d ), such
that for any U0 (x∗ )

∥f ∥H s (U (x∗ ))
0

U (x∗ ) the following holds

∥Nα f ∥H s+1 (U (x )) + ∥N2,α f ∥H s+1 (U (x )) + ∥Rf ∥H s (U (x∗ ))
∗
∗

(6.12)

for all f ∈ D′ (Ω0 , R2d ) and s ∈ R.
Theorem 6.2.2. Assume that the background velocity field c is strong fold-regular. Then
∞
there exist U (xj ) ⊂ Ω, αj ∈ C0 (S− ∂Ω), j = 1, 2...N , and a finite dimensional space

51

L0 ∈ L2 (Ω, R2d ) such that the following estimate holds for any f ∈ L2 (Ω, R2d ) with
support in Ω0 :

∥f ∥L2 (M )

N
∑
j=1

∥Nαj f ∥H 1 (U (x )) + ∥N2,αj f ∥H 1 (U (x )) + ∥Rj f ∥L2 (U (x ))
˜
j
j
j

(6.13)

where each N2,αj is a compact operator from L2 (Ω0 , R2d ) to H 1 (U (xj ), R2d ), and Rj
smoothing from E ′ (Ω, R2d ) into C ∞ (U (xj ), R2d ). If we denote by P ⊥ the orthogonal
L

0
2 (Ω, R2d ) onto the subspace which is perpendicular to L , we have the
projection in L
0

following Lipschitz estimate

∥P ⊥ f ∥L2 (M )
L
0

6.3

N
∑
j=1

∥Nαj f ∥H 1 (U (x )) .
j

(6.14)

Local properties of the normal operator N

In this subsection, we present some results about the local properties of the normal operator N (see (6.7)).
From now on, we fix x∗ ∈ Ω. We first decompose N locally into two parts based
on the separation of singularities of its Schwartz kernel. Note that the map ϕ(x∗ , ·) :
Rd → Rd is a diffeomorphism in a neighborhood of the origin. In fact, we can check
that

∂ϕ(x∗ ,·)
∂ξ (0) = c(x∗ ) · Id.

Similar to the proof of existence of uniformly normal

neighborhood in Riemannian manifold [32], we can find ϵ2 > 0 and a neighborhood of x∗ ,
˜
˜
say U (x∗ ) ⊂ Rd , such that ϕ(x, ·) are diffeomorphisms for each x ∈ U (x∗ ) in the domain
{ξ : ∥ξ∥ < 2ϵ2 }.

52

∞
Let χ∗ ∈ C0 (R) be such that χ(t) = 1 for |t| < ϵ2 and χ(t) = 0 for |t| > 2ϵ2 . We

then define
∫
N1 f (x) =
N2 f (x) =

T ∗Ω
∫ x
∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ))χ∗ (∥ξ∥) dσx (ξ),

(6.15)

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) dσx (ξ).

(6.16)

Note that for any f supported in Ω, f (ϕ(x, ξ)) = 0 for all ∥ξ∥ > T . Thus we have
∫
N2 f (x) =

∗
ξ∈Tx Ω, ϵ2 <∥ξ∥<T

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) dσx (ξ).

It is clear that Nf = N1 f + N2 f . This gives the promised decomposition of N. We
next study N1 and N2 separately.
∞ ˜
˜
Lemma 6.3.1. N1 : C0 (U (x∗ ), R2d ) → D′ (U (x∗ ), R2d ) is an elliptic ΨDO of order −1

and its principle symbol is
∫
σp (N1 )(x, ξ) = 2π ·

∗
Sx Ω

δ(⟨ξ, θ⟩)Φ† (x, θ) · Φ(x, θ) dσx (θ).

Proof. See [51] or [54].
We now proceed to study the operator N2 whose property is determined by the La∗
grangian map ϕ(x∗ , ·). We shall study the operator germ N2,ξ∗ for each ξ ∈ Tx∗ Rd . We

first consider the case when ξ∗ is not a caustic vector, i.e. ξ∗ is a regular vector.
∗
Lemma 6.3.2. Let ξ∗ ∈ Sx∗ Rd be a regular vector, then there exists a neighborhood U (x∗ )
∞
of x∗ and a neighborhood B(x∗ , ξ∗ ) of (x∗ , ξ∗ ) such that for any χ ∈ C0 (B(x∗ , ξ∗ )) the

53

following operator
∫
N2,ξ∗ f (x) =

∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · χ(x, ξ) dσx (ξ)

is a smoothing operator from E ′ (Ω, R2d ) into C ∞ (U (x∗ ), R2d ).
∗
Proof. Since ξ∗ ∈ Sx∗ Rd is regular, there exist a neighborhood V (x∗ ) of x∗ in Rd and

a neighborhood B(x∗ , ξ∗ ) of (x∗ , ξ∗ ) in R2d of the form B(x∗ , ξ∗ ) = V (x∗ ) × B0 (ξ∗ ) for
some open set B0 (ξ∗ ) in Rd such that the map ϕ(x, ·) is a diffeomorphism between B0 (ξ∗ )
and its image for all x ∈ V (x∗ ). We denote the inverse of the map ϕ(x, ·) by ϕ−1 (x, ·). By
a change of coordinate ξ = ϕ−1 (x, y) and use some cut-off function, we can write N2,ξ∗
in the following form
∫
N2,ξ∗ f (x) =

Ω

K(x, y)f (y) dy, f ∈ E ′ (Ω, R2d )

for some smooth function K in Ω × Ω. The Lemma follows immediately.
We next consider the case when ξ∗ is a fold vector. We have the following slightly
modified result from [54].
Lemma 6.3.3. Let ξ∗ be a fold vector of the map ϕ(x∗ , ·). Then there exists a small
neighborhood U (x∗ ) of x∗ and a small neighborhood B(x∗ , ξ∗ ) of (x∗ , ξ∗ ) in R2d such that
∞
for any χ ∈ C0 (B(x∗ , ξ∗ )), the operator N2,ξ∗ : E ′ (Ω, R2d ) → D′ (U (x∗ ), R2d ) defined by

∫
N2,ξ∗ f (x) =

∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · χ(x, ξ) dσx (ξ),

f ∈ E ′ (Ω, R2d )
(6.17)

54

is an FIO of order − d whose associated canonical relation is compactly supported in the
2
following set
{
(x, ξ, y, η);

x ∈ U (x∗ ), y = ϕ(x, ω), (x, ω) ∈ B(x∗ , ξ∗ ), det dω ϕ(x, ω) = 0,
}
∂ϕi (x, ω)
, η ∈ Coker (dω ϕ(x, ω)).
ξ = −ηi
(6.18)
∂x

Proof. We sketch a proof here and refer to [54] for detail. We first note that by the
fold condition, there exists a small neighborhood B1 (x∗ , ξ∗ ) of (x∗ , ξ∗ ) in R2d such that
ξ∗ is the only singular vector of the map ϕ(x∗ , ·) along the ray {tξ∗ : t ∈ R} in B1 (x∗ , ξ∗ ).
Define

S = {(x, w) : det dω ϕ(x, ω) = 0} ⊂ R2d ,
Σ = {(x, y) : y = ϕ(x, ω), (x, ω) ∈ S} ⊂ R2d .

By shrinking B1 (x∗ , ξ∗ ) if necessary, we can show that S

∩

B1 (x∗ , ξ∗ ) is a smooth

(2d − 1)-dimensional manifold in R2d , and ϕ is a diffeomorphism between S

∩

B1 (x∗ , ξ∗ )

and its image. Denote

S1 = S

∩

B1 (x∗ , ξ∗ ),

Σ1 = ϕ(S1 ).

Note that Σ1 is a smooth (2d − 1)-dimensional manifold in a neighborhood of (x∗ , y∗ )
in Ω × Ω. Let π2 be the projection from Ω × Ω to its second component. By the foldcondition for ξ∗ and the fact that the matrix dx,ξ ϕ(x∗ , ξ∗ ) is surjective, we can show that
dπ2 : T(x∗ ,ξ∗ ) Σ1 → Ty∗ Rd is surjective. Thus there exists a neighborhood V1 (y∗ ) of y∗
55

in Rd such that V1 (y∗ ) ⊂ π2 (Σ1 ). We remark that the surjectivity of the linear map dπ2
implies that the conormal bundle of Σ1 belongs to (T ∗ Rd \ 0) × (T ∗ Rd \ 0), where 0 stands
for the zero section of the normal bundle T ∗ Rd .
Let U1 (x∗ ) be a small neighborhood of x∗ such that U1 (x∗ ) ⊂ π(S1 ) and χ ∈
˜
∞
C0 (B1 (x∗ , ξ∗ )). Consider the Schwartz kernel of operator

˜
N2,ξ∗ : E ′ (Ω, R2d ) → D′ (U1 (x∗ ), R2d )

defined by
∫
˜
N2,ξ∗ f (x) =

∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · χ(x, ξ) dσx (ξ),
˜

f ∈ E ′ (Ω, R2d ).

We can show that it has conormal singularity supported in the set Σ1 . Moreover, the
conormal bundle N ∗ Σ1 is given by
{
N ∗ Σ1 = (x, ξ, y, η);

x ∈ U (x∗ ), y = ϕ(x, ω), (x, ω) ∈ B1 (x∗ , ξ∗ ),
ξ = −ηi

}
∂ϕi (x, ω)
, η ∈ Coker (dω ϕ(x, ω)), det dω ϕ(x, ω) = 0.
∂x

By analyzing the singularity of the Jacobian determinant of dω ϕ(x, ω)), we can show that
d
˜
the Schwartz kernel of N2,ξ∗ belongs to the conormal class I − 2 (Ω × Ω, Σ1 , M2d×2d ),

where M2d×2d denotes the vector bundle of matrices from R2d to R2d over Ω. Especially,
˜
when the domain of N2,ξ∗ is restricted to distribution sections supported in V1 (y∗ ), the
˜
operator N2,ξ∗ is a FIO of order − d from E ′ (V1 (y∗ ), R2d ) to D′ (U1 (x∗ ), R2d ).
2
∞
˜
The above N2,ξ∗ with χ ∈ C0 (B1 (x∗ , ξ∗ )) requires the domain to be E ′ (V1 (y∗ ), R2d ),
˜

56

we now show that this condition can be relaxed by further decreasing B1 (x∗ , ξ∗ ). Indeed,
let U (x∗ ) be a neighborhood of x∗ such that U (x∗ )
hood of (x∗ , ξ∗ ) such that B(x∗ , ξ∗ )

U1 (x∗ ) and B(x∗ , ξ∗ ) be a neighbor-

B1 (x∗ , ξ∗ ). By choosing U (x∗ ) and B(x∗ , ξ∗ ) to be

sufficiently small, we can assume that the set V (y∗ ) = { ϕ(x, ξ) : x ∈ U (x∗ ), ξ ∈ B(x∗ , ξ∗ )}
∞
is compactly supported in V1 (y∗ ). We then choose χV ∈ C0 (V1 (x∗ )) such that χV (x) = 1
∞
for x ∈ V (x∗ ). One can check that for any χ ∈ C0 (B(x∗ , ξ∗ )), the operator N2,ξ∗ defined

by
∫
N2,ξ∗ f (x) =

∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · χ(x, ξ) dσx (ξ),

f ∈ E ′ (Ω, R2d )

satisfies N2,ξ∗ f (x) = N2,ξ∗ (χV · f )(x) for all x ∈ U (x∗ ). Thus N2,ξ∗ is well-defined from
E ′ (Ω, R2d ) to D′ (U (x∗ ), R2d ), and is a FIO of order − d with canonical relation compactly
2
supported in the set (6.18). This completes the proof of the lemma.

6.4

Singularities of the map ϕ(x, ·)

In this subsection, we present some properties about the map ϕ(x, ·) which is equivalent
to the exponential map in Riemannian manifold.
By the classification result for Lagrangian maps (see [5] and [6] for detail), there are
only a finite number of stable and simple singular Lagrangian map germs in dimensions
between three and five and they are generic. In three dimensions, there are four types:
fold, cusp, swallow-tail and D4. The others are unstable and can be removed by using

57

arbitrarily small perturbations. We define

∗
K(x) = {ξ ∈ Tx Rd : the map germ ϕ(x, ·) at ξ is singular};
∗
K1 (x) = {ξ ∈ Tx Rd : the map germ ϕ(x, ·) at ξ has singularity of fold type};
∗
K2 (x) = {ξ ∈ Tx Rd : the map germ ϕ(x, ·) at ξ has singularity of cusp type};
∗
K3 (x) = {ξ ∈ Tx Rd : the map germ ϕ(x, ·) at ξ has simple and stable singularities

of types other than fold and cusp};
∗
K4 (x) = {ξ ∈ Tx Rd : the map germ ϕ(x, ·) at ξ has singularity

which are either not simple or stable}.

It is clear that K(x) =

∪4

j=1 Kj (x). Denote

∗
S(x) = {ξ ∈ Sx Rd : rξ ∈ K for some r ∈ R},
∗
Sj (x) = {ξ ∈ Sx Rd : rξ ∈ Kj for some r ∈ R}, j = 1, 2, 3, 4.

We say that the map ϕ(x, ·) is in a general position(or generic) if the map germ ϕ(x, ·) is
simple and stable at all caustic vectors in K(x), i.e. K4 (x) = ∅. It is possible that the map
ϕ(x, ·) can be brought to a general position by adding an arbitrarily small perturbation
to the velocity field c. By the classification result of Lagrangian maps, see for instance
[5], the following result holds for the set K(x).
Propsition 1. Assume that the map ϕ(x, ·) is in a general position, then the sets K1 (x)
and K2 (x) are smooth manifolds of dimensions d − 1 and d − 2, respectively. The set
K3 (x) is a union of smooth manifolds of dimensions not greater than d − 3. Especially,
58

for d = 3, the sets K1 (x), K2 (x) and K3 (x) consists of smooth surfaces, smooth curves
and isolated points, respectively.
In the case when the map ϕ(x, ·) is not in a general position, it is known that
K1 (x)

∪

K2 (x)

∪

K3 (x) is open and dense in K(x).

∗
Note that Sj (x) are the images of Kj (x) under the map β which sends ξ ∈ Tx Rd to
ξ
∗
∈ Sx Rd for ξ ̸= 0. We conclude that the following result holds.
∥ξ∥

Lemma 6.4.1. Assume that the map ϕ(x, ·) is in general position, then the sets S2 (x)
and S3 (x) are of finite d − 2 and d − 3 dimensional Hausdorff measures, respectively.
Especially, for d = 3, the set S2 (x) is a curve (not necessarily smooth) of finite length in
∗
Sx R3 and S3 (x) consists of a finite number of points.

6.5

Discussions on the concept of Fold-regular

In this subsection, we discuss the concept “fold-regular”. We show that for a general
velocity field in Ω whose induced metric is not simple, a given point in Ω is fold-regular
under some natural assumptions.
We begin with the concept “fold-regular vector”. It is still an open problem to find
a complete characterization for it, i.e. what are the necessary and sufficient conditions
for the map germ ϕ(x∗ , ·) at ξ∗ for ξ∗ to be fold-regular. We have the following partial
answer in the form of remarks.
Remark 6.5.1. In dimension d = 2, the set of fold-regular vectors is generally empty. Indeed, for a fold vector ξ∗ , the operator germ N2,ξ∗ is a FIO of order −1, and hence the best

59

estimate is that it is bounded from L2 (Ω0 , R2d ) to H 1 (U (x∗ ), R2d ) for some neighborhood
U (x∗ ) of x∗ .
Remark 6.5.2. In dimension d ≥ 3, a sufficient condition for a fold vector ξ∗ to be
fold-regular is that the following condition is satisfied

d2 ϕ(x∗ , ξ∗ )(Nx∗ (ξ∗ ) \ 0 × ·)|T S(x∗ )
ξ
ξ∗

is of full rank.

(6.19)

∗
where Nx∗ (ξ∗ ) denotes the kernel of dξ ϕ(x∗ , ξ∗ ) and S(x∗ ) the set of all vectors ξ ∈ Tx∗ Rd

such that det dξ ϕ(x∗ , ξ) = 0. Indeed, in that case, it is shown in [54] that the canonical
relation associated with the operator germ N2,ξ∗ is locally a canonical graph and hence
d

N2,ξ∗ is bounded from L2 (Ω0 , R2d ) to H 2 (U (x∗ ), R2d ) for some neighborhood U (x∗ ) of
d

x∗ . Note that for d ≥ 3, H 2 (U (x∗ ), R2d ) is compactly embedded in H 1 (U (x∗ ), R2d ), so
N2,ξ∗ is compact from L2 (Ω0 , R2d ) to H 1 (U (x∗ ), R2d ) and we can conclude that ξ∗ is
fold-regular.
The set of fold-regular vectors Z1 (x∗ ) contains more elements than those which satisfy the graph condition (6.19). In fact, let C ⊂ T ∗ Ω × T ∗ Ω be the canonical relation
associated with the operator germ N2,ξ∗ defined in Lemma 6.3.3. We have shown that
C is homogeneous and C ⊂ (T ∗ Ω\0) × (T ∗ Ω\0). By the main result in [24], N2,ξ∗ is
d

1

bounded from L2 (Ω0 , R2d ) to H 2 − 3 (U (x∗ ), R2d ) for some neighborhood U (x∗ ) of x∗ , if
the only singularity of the projection of C to its first or second component at the point
d

1

associated with (x∗ , ξ∗ ) is fold or cusp. Since H 2 − 3 (U (x∗ ), R2d ) is compactly embedded
in H 1 (U (x∗ ), R2d ), we see that N2,ξ∗ is compact from L2 (Ω0 , R2d ) to H 1 (U (x∗ ), R2d ) and
hence ξ∗ is fold-regular.
60

We now consider the concept “fold-regular point”. We denote

∗
Z1 (x∗ ) = {ξ ∈ Sx∗ Rd : ∀ r ∈ R, rξ is either regular or fold-regular}.

It is clear that

∗
Z1 (x∗ ) ⊂ Z(x∗ ) =: Sx∗ Rd \ (S2 (x∗ )

∪

S3 (x∗ ))

∪

S4 (x∗ ).

We remark that Z(x∗ ) characterizes the set of geodesics that pass through x∗ and along
which the map ϕ(x∗ , ·) only has singularities of fold type. By Morse’s index theorem (a
fold vector for the map ϕ(x, ·) corresponds to a fold conjugate vector for the exponential
map expx (·)), for each ξ ∈ S(x∗ ), there are at most finitely many fold vectors along the
geodesic π ◦ Ht (x∗ , ξ). Using the definition of “fold-regular”, we can conclude that Z1 (x∗ )
∗
is open in Sx∗ Rd .

Recall that x∗ is fold-regular if there exists a compact subset Z2 (x∗ ) ⊂ Z1 (x∗ ) such
that the following completeness condition is satisfied

∗
∀ ξ ∈ Sx∗ Rd , ∃ θ ∈ Z2 (x∗ ),

such that θ ⊥ ξ.

Remark 6.5.3. A sufficient condition for the completeness of Z1 (x∗ ) is that Z1 (x∗ )
contains a set
∗
θ⊥ = {ξ ∈ Sx∗ Rd , ξ ⊥ θ}
∗
for θ ∈ Sx∗ Rd .

61

∗
Remark 6.5.4. If the completeness condition fails for Z1 (x∗ ), then there exists θ ∈ Sx∗ Rd

such that
θ⊥ ⊂ S2 (x∗ )

∪

S3 (x∗ )

∪

S4 (x∗ ).

(6.20)

Assume that the map ϕ(x∗ , ·) is in a general position. By Lemma 6.4.1, the set on the
right hand side of (6.20) is of finite d − 2-dimensional Hausdorff measure, so is the set
∗
θ⊥ for each θ ∈ Sx∗ Rd . Thus, we conclude that there exists at most a finite number of θ

such that (6.20) holds.

6.6

Proof of Theorem 6.2.1

We prove Theorem 6.2.1 in this subsection. The proof can be divided into two major
stages: in the first stage, we present some preliminaries and construct a cut-off function
∞ ∗
α ∈ C0 (S− Γ) which selects a complete set of geodesics with only fold-regular caustics,
†

see Lemma 6.6.1; in the second stage, we study the normal operator Nα = Iα Iα , see
Lemma 6.6.2 and 6.6.3. Theorem 6.2.1 is then a direct consequence of Lemma 6.6.2 and
6.6.3.
We now present some preliminaries that are necessary for the construction of α. Let
∗
x∗ be a fold-regular point with the compact subset Z2 (x∗ ) ⊂ Sx∗ Rd in Definition 6.2.4.

Denote CZ2 = {rξ; ξ ∈ Z2 (x∗ ), r ∈ R and ϵ2 ≤ |r| ≤ T }. For each ξ∗ ∈ CZ2 , by Lemma
6.3.2 and Lemma 6.3.3, there exist a neighborhood U (x∗ , ξ∗ ) of x∗ and a neighborhood
∞
B(x∗ , ξ∗ ) of (x∗ , ξ∗ ) such that for any χ ∈ C0 (B(x∗ , ξ∗ )) the following operator

∫
N2,ξ∗ f (x) =

∗
Tx Ω

W (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · χ(x, ξ) dσx (ξ)
62

is compact from H s (Ω0 , R2d ) to H s+1 (U (x∗ , ξ∗ ), R2d ). Let B0 (x∗ , ξ∗ ) be another neighborhood of (x∗ , ξ∗ ) in R2d such that B0 (x∗ , ξ∗ )

B(x∗ , ξ∗ ). Since CZ2 is compact, there

exists a finite number of ξ∗ ’s in CZ2 , say ξ1 , ξ2 ,..., ξM , such that

CZ2 ⊂

M
∪

B0 (x∗ , ξj ).

j=1

We can then find smooth functions χ1 , χ2 ,...,χM with suppχj ⊂ B(x∗ , ξj ) for each j such
that
M
∑

χj (x, ξ) = 1 for all (x, ξ) ∈

j=1

M
∪

B0 (x∗ , ξj ).

j=1

Denote by A0 be the greatest connected open symmetric subset in

∪M

j=1 B0 (x∗ , ξj )

which contains CZ2 . Here and after, we say that a set B in R2d is symmetric if (x, ξ) ∈ B
implies that (x, −ξ) ∈ B. Define

Aϵ = {(x, ξ) ∈ R2d : |x − x∗ | ≤ ϵ, ϵ2 ≤ ∥ξ∥ ≤ T }

for each ϵ > 0. It is clear that Aϵ is compact in R2d , so is the set Aϵ \A0 .
∞ ∗
Lemma 6.6.1. There exist ϵ3 > 0 and α ∈ C0 (S− Γ) such that the following two condi-

tions are satisfied:

α(x0 , ξ0 ) = 1

for all (x0 , ξ0 ) ∈ τ ◦ β(CZ2 (x∗ )),

(6.21)

α(x0 , ξ0 ) = 0

for all (x0 , ξ0 ) ∈ τ ◦ β(Aϵ3 \A0 ).

(6.22)

Proof. Note that both β and τ are continuous. Since CZ2 (x∗ ) and Aϵ \A0 are
63

compact, so are the sets τ ◦ β(CZ2 (x∗ )) and τ ◦ β(Aϵ \A0 ). We claim that there exists
ϵ3 > 0 such that
τ ◦ β(CZ2 (x∗ ))

∩

τ ◦ β(Aϵ \A0 ) = ∅

for all ϵ ≤ ϵ3 . Indeed, assume the contrary, then

τ ◦ β(CZ2 (x∗ ))

∩

τ ◦ β(Aϵ \A0 ) ̸= ∅

for all ϵ > 0. Note that the collection of compact subsets τ ◦ β(CZ2 (x∗ ))

∩

τ ◦ β(Aϵ \A0 )

is decreasing with respect to ϵ, so it satisfies the finite intersection property and we can
thus conclude that
τ ◦ β(CZ2 (x∗ ))

∩

τ ◦ β(Aϵ \A0 ) ̸= ∅.

ϵ>0

But on the other hand, we can check that
∩

τ ◦ β(Aϵ \A0 ) = τ ((Aϵ \A0 )

ϵ>0

τ ◦ β(CZ2 (x∗ )) = τ (CZ2 (x∗ )

∩
∩

∗
Sx∗ Rd )
∗
Sx∗ Rd ).

∗
Using the fact that τ is injective on Sx∗ Rd and CZ2 ⊂ A0 , we obtain

τ ((Aϵ \A0 )

∩

∗
Sx∗ Rd )

∩

τ (CZ2 (x∗ )

∩

∗
Sx∗ Rd ) = ∅.

Thus,
τ ◦ β(CZ2 (x∗ ))

∩

τ ◦ β(Aϵ \A0 ) = ∅.

ϵ>0

64

This contradiction completes the proof of our claim.
Now, we have
τ ◦ β(CZ2 (x∗ ))

∩

τ ◦ β(Aϵ3 \A0 ) = ∅.

By decreasing ϵ3 if necessary, we may assume that

{x : |x − x∗ | ≤ ϵ3 } ⊂ π(A0 ).

∗
Since both the sets τ ◦ β(CZ2 (x∗ )) and τ ◦ β(Aϵ3 \A0 ) are compact in S− Γ, we can find
∞ ∗
α ∈ C0 (S− Γ) as desired. This concludes the proof of the lemma.

The construction of α above completes the first stage of the proof of Theorem 6.2.1,
we are now at the second stage. We define the truncated geodesic X-ray transform Iα f
as in (6.10) or (6.11). By replacing the weight Φ with the new one α♯ · Φ, we obtain Nα ,
N1,α and N2,α from the corresponding formulas of N, N1 and N2 . It is clear that Lemma
6.3.2, 6.3.3 still hold with the new weight.
Lemma 6.6.2. There exist a neighborhood U (x∗ ) of x∗ and a smoothing operator R from
E ′ (Ω, R2d ) into C ∞ (U (x∗ ), R2d ), such that for for any s ∈ R and any neighborhood U0 (x∗ )
of x∗ with U0 (x∗ )

U (x∗ ), the following estimate holds

∥f ∥H s (U (x ),R2d )
0 ∗

∥N1,α f ∥H s+1 (U (x ),R2d ) + ∥Rf ∥H s (Ω,R2d ) .
∗

(6.23)

Proof. We first show that N1,α is an elliptic ΨDO. Indeed, as in Lemma 6.3.1, N1,α

65

∞ ˜
˜
is a ΨDO of order −1 from C0 (U (x∗ ), R2d ) to D′ (U (x∗ ), R2d ) with principle symbol

∫
σp (N1 )(x, ξ) = 2π ·

∗
Sx Ω

δ(⟨ξ, θ⟩)|α♯ (x∗ , θ)|2 Φ† (x, θ) · Φ(x, θ) dσx (θ).

∗
∗
By the construction of α, for any ξ ∈ Sx∗ Rd , we have α♯ (x∗ , θ) = 1 for some θ ∈ Sx∗ Rd

with θ ⊥ ξ. Thus
∫
σp (N1,α )(x∗ , ξ) = 2π ·

∗
θ∈Sx∗ Rd , θ⊥ξ

|α♯ (x∗ , θ)|2 Φ† (x∗ , θ) · Φ(x∗ , θ) dσx∗ (θ) > 0

in the sense of symmetric positive definite matrix. By continuity, we can find a neighbor∗
˜
hood U (x∗ ) ⊂ U (x∗ ) of x∗ such that σp (N1,α )(x, ξ) > 0 for all x ∈ U (x∗ ) and ξ ∈ Sx Rd .
∞
Thus we can conclude that N1,α is an elliptic ΨDO of order −1 from C0 (U (x∗ ), R2d ) to

D′ (U (x∗ ), R2d ).
Now, let B be the pseudo-inverse of N1,α restricted to U (x∗ ). Then B is ΨDO of
order 1 and there is a smoothing operator R1 : E ′ (U (x∗ ), R2d ) → C ∞ (U (x∗ ), R2d ) such
that
g = B ◦ N1,α g + R1 g

(6.24)

for all g supported in U (x∗ ).
Next, let U0 (x∗ ) be any given neighborhood of x∗ with U0 (x∗ )

U (x∗ ). For later

convenience, we write U3 (x∗ ) for U (x∗ ). Then there exist two neighborhoods of x∗ , say
U1 (x∗ ) and U2 (x∗ ) such that U0 (x1 )

U1 (x∗ )

U2 (x∗ )

U3 (x∗ ). We choose three

smooth cut-off functions χ0 , χ1 and χ2 such that suppχj ⊂ Uj+1 and χj |Uj = 1 for
j = 0, 1, 2. Then both χ0 B(1 − χ1 ) and χ1 N1,α (1 − χ2 ) are smoothing operators.

66

Note that χ2 f is compact supported in U3 (x∗ ), so we have by (6.24) that

χ2 f = B ◦ N1,α (χ2 f ) + R1 (χ2 f ).

Thus,

χ0 f = χ0 · χ2 f = χ0 · B ◦ N1,α χ2 f + χ0 R1 (χ2 f )
(
)
= χ0 Bχ1 N1,α χ2 · f + χ0 B(1 − χ1 )N1,α χ2 + χ0 R1 χ2 f
(
)
= χ0 Bχ1 N1,α f + χ0 Bχ1 N1,α (1 − χ2 ) + χ0 B(1 − χ1 )N1,α χ2 + χ0 R1 χ2 f
= χ0 Bχ1 N1,α f + Rf

where R = χ0 Bχ1 N1,α (1 − χ2 ) + χ0 B(1 − χ1 )N1,α χ2 + χ0 R1 χ2 . We can check that R
is a smoothing operator from E ′ (Ω, R2d ) to C ∞ (U (x∗ ), R2d ).
Finally, we conclude that

∥f ∥H s (U (x ),R2d )
0 ∗

∥χ0 Bχ1 N1,α f ∥H s (U (x ),R2d ) + ∥Rf ∥H s (U (x ),R2d )
0 ∗
0 ∗
∥Bχ1 N1,α f ∥H s (U (x ),R2d ) + ∥Rf ∥H s (U (x ),R2d )
3 ∗
0 ∗
∥χ1 · N1,α f ∥H s+1 (U (x ),R2d ) + ∥Rf ∥H s (U (x ),R2d )
3 ∗
0 ∗
∥N1,α f ∥H s+1 (U (x ),R2d ) + ∥Rf ∥H s (U (x ),R2d ) .
3 ∗
0 ∗

This completes the proof of the Lemma.
We now study the operator N2,α .
Lemma 6.6.3. There exists a small neighborhood of x∗ , say U (x∗ ), such that the following
67

decomposition holds for the operator N2,α : E ′ (Ω, R2d ) → D′ (U (x∗ ), R2d )

N2,α =

M
∑

(6.25)

N2,j

j=1

where each N2,j is compact from H s (Ω0 , R2d ) to H s+1 (U (x∗ ), R2d ).
Proof. Recall that N2,α has the following representation
∫
N2,α f (x) =

∗
Tx Ω

Wα (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · dσx (ξ),

where

Wα (x, ξ) =

1
|α ◦ τ ◦ β(x, ξ)|2 Φ† ◦ β(x, ξ) · Φ ◦ β ◦ H(x, ξ)
∥ξ∥d−1
1
+
|α ◦ τ ◦ β(x, −ξ)|2 Φ† ◦ β(x, −ξ) · Φ ◦ β ◦ H−1 (x, −ξ).
∥ξ∥d−1

By (6.22) and the fact that A0 is symmetric, we see that supp Wα ⊂ A0 for all x with
|x − x∗ | ≤ ϵ3 .
Now, let χj ’s be as in the first stage. Define N2,j : E ′ (Ω, R2d ) → D′ (U (x∗ , ξj ), R2d )
by

∫
N2,j f (x) =
Let U (x∗ ) =

∩M

∗
Tx Ω

Wα (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) · χj (x, ξ) dσx (ξ).

j=1 (U (x∗ , ξj ))

∩
{x : |x − x∗ | < ϵ3 }. Then U (x∗ ) is a neighborhood of

x∗ and each N2,j is compact from H s (Ω0 , R2d ) into H s+1 (U (x∗ ), R2d ).
∑M
We claim that N2,α =
j=1 N2,j when both sides are viewed as operators from
∞
E ′ (Ω, R2d ) to D′ (U (x∗ ), R2d ). Indeed, for any f ∈ C0 (Ω, R2d ), since

68

∑M

j=1 χj = 1 on

A0 and supp Wα ⊂ A0 , we have

Wα (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) = Wα (x, ξ)f (ϕ(x, ξ))(1 − χ∗ (∥ξ∥)) ·

M
(∑

)
χj (x, ξ)

j=1

for all x ∈ U (x∗ ). Thus N2,α f =

∑M

j=1 N2,j f and the claim follows. This completes the

proof of the lemma.
Finally, note that Nα = N1,α + N2,α . Theorem 6.2.1 follows from Lemma 6.6.2 and
6.6.3.

6.7

Proof of Theorem 6.2.2

Proof of Theorem 6.2.2: Step 1. We show that (6.13) holds. For each x ∈ Ω0 , by
Theorem 6.2.1, there exist neighborhoods U (x)

˜
U (x∗ ) of x and a smooth function

∞ ∗
α ∈ C0 (S− Γ) such that the estimate (5.13) holds. Since Ω0 is compact, we can find
∪
finite number of points, say x1 , x2 , ... xN , such that Ω0 ⊂ M U (xj ) and the following
j=1

estimate holds for each j:

∥f ∥L2 (U (x ))
j

N
∑
j=1

∥Nαj f ∥H 1 (U (x )) + ∥N2,αj f ∥H 1 (U (x ),R2d ) + ∥Rj f ∥L2 (U (x ),R2d )
˜
j
j
j

where f ∈ L2 (Ω) has support in Ω0 . The estimate (6.13) follows by observing that
∑
∥f ∥L2 (Ω) ≤ N ∥f ∥L2 (U (x )) .
j=1
j

Step 2. From now on, we show that the second part of the theorem holds. Denote by

69

H the Hilbert space

∏M

1
2d
j=1 H (U (xj ), R ). We consider the following three operators

T f = (Nα1 f, Nα2 f, ..., NαM f ),
T1 f = (N2,α1 f, N2,α2 f, ..., N2,αM f ),
T2 f = (Rα1 f, Rα2 f, ..., RαM f ).

It is clear that all three operators are bounded from L2 (Ω0 ) to H. Moreover, T1 and T2
are also compact and the following estimate holds

∥f ∥L2 (Ω,R2d )

∥T f ∥H + ∥T1 f ∥H + ∥T2 f ∥H .

(6.26)

Step 3. Let L0 ⊂ L2 (Ω0 , R2d ) ⊂ L2 (Ω, R2d ) be the kernel of T . We claim that
L0 ⊂ L2 (Ω0 , R2d ) is of finite dimension. We prove by contradiction. Assume the contrary,
then there exists an infinity number of orthogonal vectors in L0 ⊂ L2 (Ω0 , R2d ), say, e1 ,
e2 , ..., such that ∥ej ∥L2 (Ω ,R2d ) = 1 and T ej = 0 for all j ∈ N. Since the sequence
0
{ej }∞ is bounded in L2 (Ω0 , R2d ) and the operators T1 and T2 are compact, we can
j=1
find a subsequence, still denoted by {ej }∞ , such that both the sequences {T1 ej }∞
j=1
j=1
and {T2 ej }∞ are Cauchy in H. By applying Inequality (6.26) to the vectors ei − ej
j=1
and recall that T (ei − ej ) = 0, we conclude that the sequence {ej }∞ is also Cauchy in
j=1
L2 (Ω0 , R2d ). This contradicts to the fact that ∥ei − ej ∥L2 (Ω ,R2d ) > 1 for all i ̸= j. This
0
contradiction proves the claim.

70

Step 4. We claim that

∥f ∥L2 (Ω ,R2d )
0

∥T f ∥H

for all f ∈ L⊥ .
0

(6.27)

Indeed, assume the contrary, there exists a sequence {fn }∞ ⊂ L⊥ such that
n=1
1
∥fn ∥L2 (Ω ,R2d ) = 1, and ∥T fn ∥H ≤
for all n.
n
0
By the same argument as in Step 3, we can find a subsequence, still denoted by {ej }∞ ,
j=1
such that both the sequences {T1 ej }∞ and {T2 ej }∞ are Cauchy in H. By Inequality
j=1
j=1
1
(6.26) and the fact that ∥T fn ∥H ≤ n , we can conclude that {fn }∞ is also Cauchy in
n=1

L2 (Ω0 , R2d ). Let f0 = limn→∞ fn , then ∥T f0 ∥H = limn→∞ ∥T fn ∥H = 0. This implies
that f0 ∈ L0 . However, note that L⊥ is closed, as the limit of a sequence of functions in
0
L⊥ , f0 must belong to L⊥ . Therefore, we see that f0 = 0. But this contradicts to the
0
0
fact that ∥f0 ∥L2 (Ω ,R2d ) = limn→∞ ∥fn ∥L2 (Ω ,R2d ) = 1. The claim is proved, whence
0
0
the estimate (6.14) follows.

71

Chapter 7
Stability of of recovering velocity
fields from their induced
Hamiltonian flows
We investigate the stability of the inverse problem of recovering velocity fields from
their induced Hamiltonian flows.
Let c be the smooth background velocity field and let c be a perturbation. We require
˜
that c is in the admissible class of velocity fields. Denote by X the X-ray transform
˜
operator obtained from linearizing the map which associate velocity fields to their induced
δHT

1
Hamiltonian flows, More precisely, X = δbc , where b(x) = − 2 ∇ ln c2 . It is clear that

(see section 6.1)
Xf (x0 , ξ0 ) = Φ−1 (T, x0 , ξ0 ) · I(f )(x0 , ξ0 ).

(7.1)

We recall the following result on the linearization.
Lemma 7.0.1. The following estimate holds

∥HT (˜) − HT (c) − Xf ∥C 1 (S ∂M )
c
−

72

∥f ∥2 2

C (Ω)

.

(7.2)

where f is defined as in (6.5).
Theorem 7.0.1. Let d = 3. Let c be a fold-regular admissible velocity field. Then there
exist ϵ > 0, a finite dimensional subspace L ⊂ L2 (Ω, R3 ), and a finite number of smooth
functions αj ∈ C ∞ (S− ∂Ω), j = 1, 2, ...N , such that the following estimate holds

∥˜2 − c2 ∥H 1 (Ω)
c

N
∑
j=1

∥ˆ j (HT (˜) − HT (c))∥H 1 (S ∂Ω) ,
α
c
−

(7.3)

for all c satisfying ∥˜ − c∥H 9 (Ω) ≤ ϵ and ∇(ln c2 − ln c2 ) ⊥ L.
˜
c
˜
Proof. Let f be as in (6.5). Denote by L the projection of L0 from L2 (Ω, R6 ) to the
space L2 (Ω, R3 ) by taking the last three components. Note that the first three components
of f are zero. Thus the condition ∇(ln c2 − ln c2 ) ⊥ L implies that f ∈ L⊥ . Therefore
˜
0

∥f ∥L2

N
∑
j=1
N
∑
j=1
N
∑
j=1
N
∑
j=1
N
∑
j=1

∥Nαj f ∥H 1 (U (x ))
j

(by Theorem 6.2.2)

∥Iαj f ∥H 1 (S Γ)
−

(by the result that I† is bounded from H 1 to H 1 )

∥Xαj f ∥H 1 (S Γ)
−

(by (7.1))

∥HT (˜) − HT (c)∥αj f ∥H 1 (S Γ) + ∥f ∥2 2
c
C (Ω)
−

(by lemma 7.0.1)

∥αj (HT (˜) − HT (c))∥H 1 (S Γ) + ∥f ∥L2 (Ω) · ∥f ∥H 8 (Ω) .
c
−

(by interpolation inequality)

73

By choosing ϵ to be sufficiently small, we can show that

∥f ∥L2

N
∑
j=1

∥αj (HT (˜) − HT (c))∥H 1 (S Γ) .
c
−

Using the formula for f , we obtain

∥ ln c2 − ln c2 ∥H 1
˜

N
∑
j=1

∥αj (HT (˜) − HT (c))∥H 1 (S Γ) .
c
−

Finally, note that
2
˜2
c2 − c2 = c2 · (eln c −ln c − 1).
˜

The estimate (7.3) follows. This completes the proof of the theorem.
Remark 7.0.1. Similar result also holds for d ≥ 3.

74

Chapter 8
Sensitivity of recovering velocity
fields from their induced DDtN maps
In this chapter, we investigate the sensitivity of the inverse problem of recovering
velocity fields from their induced DDtN maps. We first present a lemma which is a direct
consequence of Theorem 5.2.1 and Lemma 6.1.3.
Lemma 8.0.2. Let c and c be two velocity field in A(ϵ0 , Ω, M0 , T ), and let f be as in
˜
(6.5). Then there exists δ > 0 such that if ∥Λc − Λc ∥H 1 [0,3ϵ /4]×Γ→L2 ([0,T +ϵ ]×Γ) ≤ δ,
˜
1
1
0
then
∥If ∥L∞ (S ∗ Γ,R2d ) ≤ C∥f ∥2 1

C (Ω,R2d )

−

(8.1)

for constant C > 0 depending M0 .
Definition 8.0.1. An admissible velocity field c is called fold-regular if all points in Ω
t
are fold-regular with respect to the Hamiltonian flow Hc .

We have established the following sensitivity result. For simplicity we only consider
the case d = 3, similar results also hold for d > 3.
Theorem 8.0.2. Let c and c be two velocity fields in the class A(ϵ0 , Ω, M0 , T ). Assume
˜
that the velocity field c is smooth and is fold-regular. Then there exist a finite dimensional
75

subspace L ⊂ L2 (Ω0 , R3 ), and a constant δ > 0 such that for all c sufficiently close to c in
˜
17

H 2 (Ω) and satisfying ∇(ln c2 − ln c2 ) ⊥ L, ∥Λc − Λc ∥H 1 [0,3ϵ /4]×Γ→L2 ([0,T +ϵ ]×Γ) ≤ δ
˜
˜
1
1
0
implies that c = c.
˜
Proof. The proof is divided into the following three steps.
Step 1. Let L0 be defined as in Let Theorem 6.2.2. Let f be as in (6.5). Assume that
∞
f ⊥ L0 . By Theorem 6.2.2, there exist U (xj ) ⊂ Ω, αj ∈ C0 (S− ∂Ω), j = 1, 2...N such

that the following estimate holds

∥f ∥L2 (Ω)

N
∑
j=1

∥Nαj f ∥H 1 (U (x )) .
j

(8.2)

Step 2. We show that
N
∑
j=1

∥Nαj f ∥H 1 (U (x ))
j

2
∥f ∥ 5

· ∥f ∥L2 (Ω,R2d ) .
15
H 2 (Ω,R2d )

(8.3)

†

Indeed, for each Iαj , by Lemma 8.0.2, we have ∥Iαj f ∥L∞ (S ∗ Γ) ≤ C1 ∥f ∥2 1 . Apply Iαj
C
−
†
to both sides and use the fact that Iαj is bounded from L2 to L2 (see [45]), we obtain

∥Nαj f ∥L2 (Ω,R2d )

76

∥f ∥2 1

.
C (Ω,R2d )

(8.4)

Then,
2

1

· ∥Nα f ∥ 3 2
H (U (xj ),R2d )
L (Ω,R2d )

∥Nαj f ∥H 1 (U (x ),R2d )
j

∥Nαj f ∥ 3 3

(by interpolation inequality)
1

4

∥Nαj f ∥ 3 3

· ∥f ∥ 3 1
H (U (xj ),R2d )
C (Ω,R2d )

1

4

∥f ∥ 3 3

· ∥f ∥ 3 1
H (Ω,R2d )
C (Ω,R2d )

(by Lemma 6.6.2, 6.6.3)

4

1

∥f ∥ 3 3

· ∥f ∥ 3 3
H (Ω,R2d )
H (Ω,R2d )

=

(by (8.4))

5
∥f ∥ 3 3
H (Ω,R2d )
2
∥f ∥ 5 15
· ∥f ∥L2 (Ω,R2d ) .
H 2 (Ω,R2d )

(by interpolation inequality)

(by interpolation inequality)

It follows that (8.3) holds.
Step 3. Denote by L the projection of L0 from L2 (Ω, R2d ) to the space L2 (Ω, Rd ) by
taking the last three components. Note that the first three components of f are zero,
see (6.5). Thus the condition ∇(ln c2 − ln c2 ) ⊥ L implies that f ∈ L⊥ . Consequently,
˜
0
Inequality (8.2) holds. Combining this with (8.3), we see that

∥f ∥L2 (Ω,R2d )

2

∥f ∥ 5 15

H 2 (Ω,R2d )

· ∥f ∥L2 (Ω,R2d ) .

2
Therefore, we must have f = 0 for ∥f ∥ 5 15
sufficiently small. Finally, note that
H 2 (Ω,R2d )

∥f ∥ 15

H 2 (Ω,R2d )

∥c − c∥ 17
˜

and that both c and c vanishes near the boundary, we
˜

H 2 (Ω)

conclude that f = 0 implies c = c. This completes the proof of the theorem.
˜

77

BIBLIOGRAPHY

78

BIBLIOGRAPHY
[1] G. Alessandrini, J. Sylvester and Z. Sun. Stability for a multidimensional inverse spectral theorem, Comm.PDE, 15(1990), 711-736.
[2] G. Alessandrini and J. Sylvester. Stability for a multidimensional inverse spectral theorem, Comm.PDE, 15(1990), 711-736.
[3] R. Alexandre, J. Akian and S. Bougacha. Gaussian beams summation for the wave
equation in a convex domain, Commun. Math. Sci, 7(4)(2009), 973-1008.
[4] Y. E. Anikonov and V. G. Romanov. On uniqueness of definition of first-order form
by its integrals over geodesics, Ill-posed Math. and Geophys. Problems, Novosibirsk,
in Russian, (1976), 22-27.
[5] V. I. Arnold. Singularity Theory, London Mathematical Society Lecture Note Series(53), Cambrige Univeristy Press, 1981.
[6] V. I. Arnold, S. M. Gusein-Zade and A. N. Varchenko. Singularites of Differentiable
Maps I, volume 82 of monographs in mathematics. Birkh¨user, Boston, 1985.
a
[7] G. Bao, J. Qian, L. Ying and H. Zhang. A convergent multiscale Gaussian-beam
parametrix for wave equations, Comm.PDE, 38(2013), 92-134.
[8] G. Bao and K. Yun. On the stability of an inverse problem for the wave equation,
Inverse Problems, 25(4)(2009), R1-R7.
[9] G. Bao and H. Zhang. Sensitivity analysis of an inverse problem for the wave equation
with caustics, submitted.
[10] M. Belishev. An approach to multidimensional inverse problems for the wave equation, Dokl. Akad. Nauk. SSR, 297(1987), 524-527.
[11] M. Belishev. Boundary control in reconstruction of manifolds and metrics (the BC
method), Inverse Problems, 13(1997), R1-R45.

79

[12] M. Belishev. Recent progress in the boundary control method, Inverse Problems,
23(2007), R1-R67.
[13] M. Belishev and Y. V. Kurylev. To the reconstruction of a Riemannian manifold via
its spectral data (BC-method), Comm.P.D.E, 17(1992), 765-804.
[14] M. Bellassoued and D. D. S. Ferreira. Stability estimates for the anisotropic wave
equation from the Dirichlet-to-Neumann map, Inverse Problems and Imaging,
5(4)(2011), 745-773.
[15] I. N. Bernstein and M. L. Gerver. Conditions of distinguishability of metrics by
godographs, Methods and Algoritms of Interpretation of Seismological Information,
Computerized Seismology, 13(1980), 50-73.
[16] C. Croke. Boundary and lens rigidity of finite quotients, Proc. AMS, 133(12) (2005),
3663-3668.
[17] C. Croke and B. Kleiner. Conjugacy and rigidity for manifolds with a parallel vecotor
feild, J. Diff. Geom, 39(1994), 659-680.
[18] G. Eskin. A new approach to hyperbolic inverse problems, Inverse Problems, 22
(2006), 815-831.
[19] G. Eskin. A new approach to hyperbolic inverse problems II: global step, Inverse
Problems, 23 (2007), 2343-2356.
[20] R. G. Mukhometov. On the problem of integral gometry, Math. Problems of Geophysics, Akad. Nauk. SSSR, Sibirsk. Otdel, Vychisl. Tsentr, Novosibirsk, 6(2) (1975),
212-242.
[21] R. G. Mukhometov. Inverse kinetic problem of seismic on the plane, Math. Problems
of Geophysics, Akad. Nauk. SSSR, Sibirsk. Otdel, Vychisl. Tsentr, Novosibirsk, 6(2)
(1975), 212-242.
[22] R. G. Mukhometov, The reconstruction problem of a two-dimensional Riemannian
metric, and integral geometry, Soviet Math. Dokl, 18(1), 27-37.
[23] R. G. Mukhometov. On a problem of reconstructing Riemannian metrics, Siberian
Math. J, 22(3), 420-433.
80

[24] A. Greenleaf and A. Seeger. Fourier integral operators with cusp singularites, American Journal of Mathematics, 120(5)(1998), 1077-1119.
[25] L. H¨mander. The Analysis of Linear Partial Differential Operators, III and IV,
o
volume 274 and 275 of Grundelehren der Mathematischen Wissenschaften, SpringVerlag, Berlin, 1985.
[26] V. Isakov. An inverse hyperbolic problem with many boundary measurements, Comm. PDE, 16 (1991), 1183-1195.
[27] V. Isakov and Z. Sun. Stability estimates for hyperbolic inverse problems with local
boundary data, Inverse Problems, 8 (1992), 193-206.
[28] V. Isakov. Inverse Problems for Partial Differential Equations, volume 127 of Applied
Mathematical Sciences, second edition, Springer Science and Business Media, Inc,
New York, 2006.
[29] Y. V. Kurylev and M. Lassas, Hyperbolic inverse problem with data on a part of
the boundary, in “Differential Equations and Mathematical Physics”, AMS/IP Stud.
Adv. Math. 16, Amer. Math. Soc., Providence, (2000), 259-272.
[30] A. Katchalov, Y. Kurylev and M. Lassas. Inverse Boundary Spectral Problems, Chapman Hall/CRC, Boca Raton, 2001.
[31] I. Lasiecka, J. L. Lions and R. Triggiani. Nonhomogeneous boundary value problems
for second order hyperbolic operators, J. Math. Pures et Appl, 65(1986), 149-192.
[32] J. M. Lee. Riemannian manifolds, an introduction to curvature, volume 176 of Graduate Texts in Mathematics, Spring-Verlag, New York, 1997.
[33] C. Montalto. Stable determination of a simple metric, a covector field and a potential
from the hyperbolic Dirichlet-to-Nuemann map, Arxiv, 2012.
[34] L. N. Pestov and V. A. Sharafutdinov. Integral geometry of tensor fields on a manifold
of negative curvative, Siberian Math. J, 29(3), (1988), 427-441.
[35] L. Pestov and G. Uhlmann. The scattering relation and the Dirichlet-to-Neumann
map, Contemporary Math, 412(2006), 249-262.

81

[36] J. Qian and L. Ying. Fast multiscale Gaussian wavepacket transforms and multiscale
Gaussian beams for the wave equation, SIAM. MMS, 8 (2010).
[37] J. Ralston. Gassian beams and the propagation of singularities, Studies in Partial Differential Equations, MAA Studies in Mathematics, Edited by W. Littman, 23(1983),
206-248.
[38] Rakesh. Reconstruction for an inverse problem for the wave equation with constant
velocity, Inverse Problems, 6 (1990), 91-98.
[39] Rakesh and W. Symes. Uniqueness for an inverse problems for the wave equation,
Comm. PDE, 13 (1988), 87-96.
[40] A. Ramm and J. Sjostrand. An inverse problem of the wave equation, Math. Z, 206
(1991), 119-130.
[41] V. G. Romanov. Reconstructing a function by means of integral along a family of
curves, Sibirsk. Mat. Z, 8(5) (1967), 1206-1208.
[42] V. G. Romanov. Integral geometry on geodesics of an isotropic Riemannian metric,
Soviet Math. Dokl, 19(4) (1979), 847-851.
[43] V. G. Romanov. Invere Problems of Mathematical Physics, VNU Science Press, Utreht, the Netherlands, 1987.
[44] V. A. Sharafutdinov. Integral Geometry of Tensor Fields, VSP, Utrech, the Netherland, 1994.
[45] V. A. Sharafutdinov. Ray Transform on Riemannian Manifolds, Lecture notes, University of Oulu, 1999.
[46] P. Stefanov and G. Uhlmann. Stability estimates for the hyperbolic Dirichlet to Neumann map in anisotropic media, J. Funct. Anal, 154(2)(1998), 330-358.
[47] P. Stefanov and G. Uhlmann. Stability estimates for the X-ray transform of tensor
fields and boundary rigidity, Duke Math. J, 123(2004), 445-467.
[48] P. Stefanov and G. Uhlmann. Boundary rigidity and stability for generic simple metrics, Journal Amer. Math. Soc, 18(2005), 975-1003.
82

[49] P. Stefanov and G. Uhlmann. Stable determination of generic simple metrics from the
hyperbolic Dirichlet-to-Neumann map, IMRN, 17(2005), 1047-1061.
[50] P. Stefanov and G. Uhlmann. The X-Ray transform for a generic family of curves and
weights, J. Geom. Anal, 18(1)(2008), 81-97.
[51] P. Stefanov and G. Uhlmann. Integral geometry of tensor fields on a class of nonsimple Riemannian manifolds, American J. of Math, 130 (2008), 239-268.
[52] P. Stefanov and G. Uhlmann. Linearizing non-linear inverse problems and its applications to inverse backscattering, Journal Functional Analysis, 256 (2009), 2842-2866.
[53] P. Stefanov and G. Uhlmann. Local lens rigidity with incomplete data for a class of
non-simple Riemannian manifolds, Journal Diff. Geometry, 82 (2009), 383-409.
[54] P. Stefanov and G. Uhlmann. The geodesic X-ray transform with fold caustics, Anal.
and PDE, 5(2012), 219-260.
[55] Z. Sun. On continuous dependence for an inverse initial boundary value problem for
the wave equation, J. Math. Anal. Appl, 150(1990), 188-204.
[56] G. Uhlmann. The cauchy data and the scattering relation, IMA Publications,
Springer-Verlag, 137(2003), 263-288.

83