Abstract
In this note, we study the operator norm of the generalized spherical Radon transform, defined by a smooth measure on the underlying incidence variety. In particular, we prove that for small perturbations of the measure, the spherical Radon transform remains an isomorphism between the corresponding Sobolev spaces.
Mathematical Subject Classifications (2010): 44A12, 53C65
Access provided by Autonomous University of Puebla. Download chapter PDF
Similar content being viewed by others
Key words
1 Introduction and Background
Throughout the note, we fix a Euclidean space V = ℝ d + 1, and consider the Euclidean spheres X = S d ⊂ V, and Y = S d ⊂ V ∗ . For p ∈ Y, C p ⊂ X will denote the copy of S d − 1 ⊂ X given by C p = { q ∈ X : ⟨q, p⟩ = 0}. Let σ d − 1(q) denote the SO(d)-invariant probability measure on C p . The set C q ⊂ Yand the measure σ d − 1(p) on it are defined similarly. Then the spherical Radon transform is defined as follows:
Let σ be the unique SO(d + 1)-invariant probability measure on the incidence variety Z = { (q, p) ∈ X ×Y: ⟨q, p⟩ = 0}. Assume one is given a smooth, not necessarily positive measure dμ on Z, given by μ(q, p)σ where μ ∈ C ∞(Z), and which satisfies μ( ± q, ± p) = μ(q, p) (call such μ symmetric). Introduce \(\mathcal{R}_{\mu } : {C}^{\infty }(X) \rightarrow {C}^{\infty }(Y )\) by
Introduce also the dual Radon transform \(\mathcal{R}_{\mu }^{T} : {C}^{\infty }(Y ) \rightarrow {C}^{\infty }(X)\) which is formally adjoint to \(\mathcal{R}_{\mu }\) and given by
Let \(L_{s}^{2}(\mathbb{P}X)\) and \(L_{s}^{2}(\mathbb{P}Y )\) denote the Sobolev space of even functions on X and Y, respectively. It is well known (see [2]) that the spherical Radon transform extends to an isomorphism of Sobolev spaces:
for every \(s \in \mathbb{R}\). For general μ as above, \(\mathcal{R}_{\mu }\) is a Fourier integral operator of order \(\frac{d-1} {2}\) (see [3, 5]), and so extends to a bounded map \(\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+\frac{d-1} {2} }^{2}(\mathbb{P}Y )\). We look for conditions on μ so that this is again an isomorphism.
It follows from Guillemin’s theorem on general Radon transforms associated to double fibrations [1], that \(\mathcal{R}_{\mu }^{T}\mathcal{R}_{\mu } : {C}^{\infty }(\mathbb{P}X) \rightarrow {C}^{\infty }(\mathbb{P}X)\) is an elliptic pseudo-differential operator of order d − 1 for all smooth, positive, symmetric measures μ on Z (for completeness, this is verified in the Appendix). The dependence of the principal symbol of \(\mathcal{R}_{\mu }^{T}\mathcal{R}_{\mu }\) on μ was investigated in [5]. In this note, we analyze the dependence on μ of the operator norm of \(\mathcal{R}_{\mu }^{T}\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+(d-1)}^{2}(\mathbb{P}X)\). We then give a sufficient condition on a perturbation μ of μ0 so that \(\mathcal{R}_{\mu } : {C}^{\infty }(\mathbb{P}X) \rightarrow {C}^{\infty }(\mathbb{P}Y )\) remains an isomorphism. Namely, we prove the following
Theorem. The set of C ∞ measures μ on Z for which \(\mathcal{R}_{\mu } : {C}^{\infty }(\mathbb{P}X) \rightarrow {C}^{\infty }(\mathbb{P}Y )\) is an isomorphism, is open in the C 2d+1 (Z) topology.
2 Bounding the Norm of \(\mathcal{R}_{\nu }^{T}\mathcal{R}_{\mu }\)
We start by recalling an equivalent description of the Radon transform. Consider the double fibration
Let σ X = π ∗ σ and σ Y = ρ ∗ σ be the rotation-invariant probability measures on X and Y, respectively. Then for f ∈ C ∞(X), \((\mathcal{R}f)\sigma _{Y } = \rho _{{\ast}}(\sigma ({\pi }^{{\ast}}f))\). For smooth symmetric measures dμ, dν on Z, given by μ(q, p)σ and ν(q, p)σ we can define \(\mathcal{R}_{\mu } : {C}^{\infty }(X) \rightarrow {C}^{\infty }(Y )\), \(\mathcal{R}_{\nu }^{T} : {C}^{\infty }(Y ) \rightarrow {C}^{\infty }(X)\) respectively by
and
It follows from [1] that both \(\mathcal{R}_{\mu }\) and \(\mathcal{R}_{\nu }^{T}\) are Fourier integral operators of order \(\frac{d-1} {2}\). Thus we restrict to even functions and consider \(\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+\frac{d-1} {2} }^{2}(\mathbb{P}Y )\) and \(\mathcal{R}_{\nu }^{T} : L_{s+\frac{d-1} {2} }^{2}(\mathbb{P}Y ) \rightarrow L_{s+(d-1)}^{2}(\mathbb{P}X)\).
As before, q will denote a point in X and p a point in Y. We will often write q ∈ p instead of ⟨q, p⟩ = 0 ⇔ q ∈ C p ⇔ p ∈ C q . In the following, the functions f, g are even. We also assume d ≥ 2.
Proposition 1.
The Schwartz kernel of \(\mathcal{R}_{\nu }^{T}\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+(d-1)}^{2}(\mathbb{P}X)\) is
that is,
Here c d is a constant, and α(q,q′) is the average over all p ∈ Y s.t. q,q′∈ p of μ(q,p)ν(q′,p). More precisely,
where SO(d − 1) ={ g ∈ SO(d + 1) : gq = q,gq′ = q′}, \(C_{p_{0}}\) is any fixed copy of S d through q,q′, and dM is the Haar probability measure on SO(d − 1).
Proof.
Fix some q′ ∈ X, and p 0 ∈ Ys.t. q′ ∈ p 0. Let SO(d) ⊂ SO(d + 1) be the stabilizer of q′ ∈ X. For g ∈ C ∞(Y) we may write
where dM is the Haar probability measure on SO(d). Then taking
we get
Denote \(\theta = \text{dist}(\tilde{q},q^{\prime})\), and S θ d − 1 = { q : dist(q′, q) = θ}. Let dσ d − 1 θ(q) denote the rotationally invariant probability measure on S θ d − 1. The inner integral may be written as
Here α(q, q′) = ∫ SO(d − 1)μ(q, Mp 0)ν(q′, Mp 0)dM with SO(d − 1) = Stab(q) ∩ Stab(q 0) is just the average of μ(q, p)ν(q′, p) over all (d − 1)-dimensional spheres C p containing both q and q′. Then
and since the inner integral only depends on \(\theta = \text{dist}(\tilde{q},q^{\prime})\), this may be rewritten as
Finally, dσ d = c d sind − 1θdθdσ d − 1 θ, and so
We conclude that the Schwartz kernel is
We proceed to estimate the norm of \(\mathcal{R}_{\nu }^{T}\mathcal{R}_{\mu }\). Our main tool will be the following proposition proved in Sect. C
Proposition.
Consider a pseudodifferential operator P of order m
between Sobolev spaces with x−compactly supported symbol p(x,ξ) in \(K \subset {\mathbb{R}}^{n}\) s.t.
There exists a constant C(n,s) such that
Proposition 2.
The norm of \(\mathcal{R}_{\nu }^{T}\mathcal{R}_{\mu } : L_{-(d-1)}^{2}(\mathbb{P}X) \rightarrow L_{0}^{2}(\mathbb{P}X)\) is bounded from above by
for some constant C dependent on the double fibration.
Proof.
First introduce coordinate charts. Choose a partition of unity χ i (q′) corresponding to a covering of X by charts U i , and a function \(\rho : [0,\infty ) \rightarrow \mathbb{R}_{+}\) with support in [0, 1] s.t. ρ(r) = 1 for \(r \leq \frac{1} {2}\). Write
and
Let \(\mathcal{R}_{\nu }^{T}\mathcal{R}_{\mu } =\sum _{i}T_{K_{i}} + T_{L_{i}}\) be the corresponding decomposition for the operators.
First we will bound the norm of the diagonal terms, i.e., the operators defined by K i . Fix i, and choose some point q′ ∈ U i . Introduce polar coordinates (r, ψ) around q′ so that ψ ∈ S 1 d − 1(q′) and r = sinθ for \(r \leq \frac{1} {2}\), θ = dist(q, q′). Note that α(q′, (r, ψ)) = α(q′, (r, − ψ)). By Proposition 5, the corresponding symbol is
For a given q′, introduce spherical coordinates ψ = (ϕ, ϕ1, …, ϕ d − 2) 0 ≤ ϕ ≤ π, on S 1 d − 1(q′) in such a way that cosϕ = ψ1; Take ξ0 = (1, 0, …, 0). Then
where m(q′, r, ϕ) = χ i (q′)∫ 0 π …∫ 0 π ∫ 0 2π dϕ1 …dϕ d − 2α(q′, r, ϕ, ϕ1, …, ϕ d − 2). We then have m(q′, r, π ∕ 2 + ϕ) = m(q′, r, π ∕ 2 − ϕ). Take t = cosϕ and M(q′, r, t) = m(q′, r, arccost). Then M(q′, r, t) = M(q′, r, − t) and
Since M is even, we may write
and so for all multi-indices α
Define as in Appendix A
then we can write
and conclude by Proposition 3 that
and similarly for all multi-indices α
It is also immediate that
and similarly
So we can write
for some universal constant C = C(d). Then choosing s = 0 and | α | = d + 1 in Proposition 4 we get that
Now we bound the norm of the off-diagonal term, namely the sum of operators corresponding to L i . They constitute a smoothing operator T L (μ, ν); its Schwartz kernel k(q′, q) = (1 − ρ(dist(q′, q)))K(q′, q) is a smooth function in both arguments. Denoting by ∇ j : C ∞(X) → (T ∗ X) ⊗ j the j-th derivative obtained from the Levi-Civita connection,
and
So
It is easy to see that the adjoint operator \(T_{L}{(\mu,\nu )}^{{\ast}} : L_{(d-1)}^{2}{(\mathbb{P}X)}^{{\ast}}\rightarrow L_{0}^{2}{(\mathbb{P}X)}^{{\ast}}\) equals T L (ν, μ) after the isomorphic identification \(L_{s}^{2}{(\mathbb{P}X)}^{{\ast}}\simeq L_{-s}^{2}(\mathbb{P}X)\) for s = 0, d − 1. Since the bound above is symmetric in μ, ν we conclude
Finally
Theorem 1.
Assume d ≥ 2, and let μ 0 ∈ C ∞ (Z) be such that \(\mathcal{R}_{\mu _{0}} : {C}^{\infty }(\mathbb{P}X) \rightarrow {C}^{\infty }(\mathbb{P}Y )\) is an isomorphism. Then there exists ε 0 > 0 (depending on the double fibration), such that if \(\|\mu - \mu _{0}\|_{{C}^{2d+1}(Z)} < \epsilon _{0}\) then \(\mathcal{R}_{\mu } : {C}^{\infty }(\mathbb{P}X) \rightarrow {C}^{\infty }(\mathbb{P}Y )\) is an isomorphism (for all s).
Proof.
Since \(\mathcal{R}_{\mu _{0}}^{T}\mathcal{R}_{\mu _{0}} : L_{-(d-1)}^{2}(\mathbb{P}X) \rightarrow L_{0}^{2}(\mathbb{P}X)\) (and likewise for Y) is elliptic, it is an isomorphism. Let us verify that both of the maps \(\mathcal{R}_{\mu }^{T}\mathcal{R}_{\mu } : L_{-(d-1)}^{2}(\mathbb{P}X) \rightarrow L_{0}^{2}(\mathbb{P}X)\) and \(\mathcal{R}_{\mu }\mathcal{R}_{\mu }^{T} : L_{-(d-1)}^{2}(\mathbb{P}Y ) \rightarrow L_{0}^{2}(\mathbb{P}Y )\) remain an isomorphism for small perturbations μ of μ0 in the C 2d + 1(Z) norm:
so by Corollary 2, there is an ε0 > 0 s.t. all norms are indeed small when \(\|\mu - \mu _{0}\|_{{C}^{2d+1}(Z)} < \epsilon _{0}\). The operator \(\mathcal{R}_{\mu }\mathcal{R}_{\mu }^{T}\) is treated identically, and we take the minimal of the two ε0.
Since both \(\mathcal{R}_{\mu }^{T}\mathcal{R}_{\mu }\) and \(\mathcal{R}_{\mu }\mathcal{R}_{\mu }^{T}\) are elliptic operators, the dimension of the kernel and cokernel are independent of s. It follows that \(\mathcal{R}_{\mu }^{T}\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+d-1}^{2}(\mathbb{P}X)\) and \(\mathcal{R}_{\mu }\mathcal{R}_{\mu }^{T} : L_{s}^{2}(\mathbb{P}Y ) \rightarrow L_{s+d-1}^{2}(\mathbb{P}Y )\) are isomorphisms for all s. In particular, \(\text{Ker}(\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+\frac{d-1} {2} }^{2}(\mathbb{P}Y )) = 0\) and \(\text{Coker}(\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+\frac{d-1} {2} }^{2}(\mathbb{P}Y )) = 0\), implying by the open mapping theorem that \(\mathcal{R}_{\mu } : L_{s}^{2}(\mathbb{P}X) \rightarrow L_{s+\frac{d-1} {2} }^{2}(\mathbb{P}Y )\) is an isomorphism for all s. The result follows.
Remark 1.
It is unlikely that the result is sharp. For instance, in the case of d = 1 one only needs \(\|\mu - 1\|_{{C}^{0}}\) to be small to conclude that \(\mathcal{R}_{\mu }\) is an isomorphism, while the statement (although non-applicable for d = 1) would suggest bounding the C 3-norm.
3 Appendix
3.1 A Some Integral Estimates
Fix some real T > 0. For an integer d ≥ 2, a smooth function \(\rho : [0,\infty ) \rightarrow \mathbb{R}\) compactly supported in [0, 1), and smooth functions \(m(r,\phi ),n(r,\phi ) : [0,\infty ) \times [0,\pi ]\mathbb{R}\) we define the integrals
and
where M(r, t) = m(r, ϕ) and N(r, t) = n(r, ϕ) for t = cosϕ. We assume m is even w.r.t. \(\frac{\pi } {2}\), namely \(m(r, \frac{\pi } {2} + \phi ) = m(r, \frac{\pi } {2} - \phi )\;\Longleftrightarrow\;M(r,t) = M(r,-t)\); while n is odd, i.e. \(n(r, \frac{\pi } {2} + \phi ) = -n(r, \frac{\pi } {2} - \phi )\;\Longleftrightarrow\;N(r,t) = -N(r,-t).\)
Proposition 3.
There exists a constant C = C(d,ρ) such that for d ≥ 2 and all even functions m
and for odd functions n
Proof.
Induction on d. Start by verifying the bounds for d = 2. To bound
we first integrate the inner integral by parts:
Let us bound separately
and
Now
and since \(\frac{\partial M} {\partial t} = -\frac{\partial m} {\partial \phi } \frac{1} {\sqrt{1-{t}^{2}}}\),
Now since Si is bounded, it follows that
Thus
Similarly,
Tracing back,
Next we bound | J(2, ρ, n) |
Integrate the inner integral by parts:
Let us bound separately
and
Now
and since \(\frac{\partial N} {\partial t} = -\frac{\partial n} {\partial \phi } \frac{1} {\sqrt{1-{t}^{2}}}\) and N(r, 0) = 0 we get that \(\vert \frac{\partial } {\partial t}N(r,t)\vert \leq C\sup \vert \frac{\partial n} {\partial \phi }\vert \) for \(0 \leq t \leq \frac{1} {2}\) so \(\vert N(r,t)\vert \leq C\sup \vert \frac{\partial n} {\partial \phi }\vert t\) and
Thus
Similarly,
and putting all together,
as required.
Next consider the case d = 3. We bound I, J simultaneously. Apply integration by parts to the inner integrals:
and similarly
These first summands are
and
the second summand for I
is bounded by
Similarly since N(r, 0) = 0, also the second summand for J
thus we showed that
and
and plugging the already proved estimates for d = 2 concludes the case d = 3.
Finally, for d > 3 we will apply induction. Again consider both integrals simultaneously. Start by integrating by parts the inner integral: the boundary term is zero (for J since n is odd), so
Thus
and
The first terms are \(\frac{1} {T}J(d - 1,\rho, \frac{\partial m} {\partial \phi } )\) and \(-\frac{1} {T}I(d - 1,\rho, \frac{\partial n} {\partial \phi })\), respectively.
In the second term, first change the order of integration:
Now apply integration by parts to the inner integral. Since d − 3 > 0 and ρ(1) = 0, again there is no boundary term:
Thus
and the corresponding term for J:
and we conclude by induction.
3.2 B Guillemin’s Condition
For q ∈ X, we denote by \(\bar{q} \in X\) the unique point proportional to q and distinct from it. We will consider the projective space \(\mathbb{P}X = \mathbb{R}{\mathbb{P}}^{d}\), \(\mathbb{P}Y = \mathbb{R}{\mathbb{P}}^{d}\) and the projectivized incidence variety \(\mathbb{P}Z =\{ (q,p) \in \mathbb{P}X \times \mathbb{P}Y :\langle q,p\rangle = 0\}\).Consider the projectivized double fibration
Then any two fibers F p (ℙX) intersect transversally (since before projectivization, the only non-transversal intersection was between fibers over antipodal points). Denote N W ⊂ T ∗ (ℙX ×ℙY ), N E ⊂ T ∗ (X ×Y) the conormal bundles of W, E respectively. Since dimE = 2d − 1 and dim(X ×Y) = 2d, the fibers of N E , N W are one-dimensional. Recall that T (q, p) E = { (ξ, η) ∈ T q X ×T p Y: ⟨q, η⟩ + ⟨ξ, p⟩ = 0}. Therefore, N E over (q, p) ∈ E has its fiber spanned by (p, q) ∈ T q ∗ X ×T p ∗ Y. One thus has N E ∖ 0 ⊂ (T ∗ X ∖ 0) ×(T ∗ Y ∖ 0), and ρ : N E ∖ 0 → T ∗ Y ∖ 0 given by ((q, p), t(p, q))↦(p, tq) is an immersion, which is two-to-one since \(\rho ((q,p),t(p,q)) = \rho ((\bar{q},p),(-t)(p,\bar{q}))\). The corresponding map ρ : N W ∖ 0 → T ∗ ℙY ∖ 0 is already an injective immersion. Thus Guillemin’s condition is satisfied, and we conclude
Corollary 1.
For any smooth positive measure \(\mu \in {\mathcal{M}}^{\infty }(\mathbb{P}Z)\) , ℛ μ T ℛ μ : C ∞ (ℙX) → C ∞ (ℙX) is an elliptic pseudodifferential operator.
3.3 C Pseudo-Differential Operators
For a survey of the subject, see for instance [4].We will study the norm of a pseudodifferential linear operator \(P : {C}^{\infty }({\mathbb{R}}^{n}) \rightarrow {C}^{\infty }({\mathbb{R}}^{n})\) which is given by its symbol p(x, ξ)
where p ∈ Sym m(K), i.e.,
-
1.
\(p \in {C}^{\infty }({\mathbb{R}}^{n} \times {\mathbb{R}}^{n})\)
-
2.
p has compact x − support \(K \subset {\mathbb{R}}^{n}\)
-
3.
| D ξ β D x α p(x, ξ) | ≤ C αβ(1 + | ξ | )m − | β |
It is well known that for all \(s \in \mathbb{R}\), P extends to a bounded operator between Sobolev spaces
We will trace the proof of this fact to understand the dependence on p of the operator norm \(\|P\|\).
Proposition 4.
There exists a constant C(n,s) such that
Proof.
All the integrals in the following are over \({\mathbb{R}}^{n}\). Start by integrating by parts:
So
where C α = 2 | α | (C 00 + C α0). We want to bound
Take \(v \in L_{-s}^{2}({\mathbb{R}}^{n})\), then
so denoting
we have
Now
Therefore,
where
similarly
implying
and this holds for all α s.t. A(n, | s | − | α | ) = ∫dξ(1 + | ξ | ) | s | − | α | < ∞, i.e. | s | − | α | < − n ⇔ | α | > n + | s | . We thus choose α s.t. | α | = ⌊ | s | ⌋ + n + 1, and recall that C α = 2 | α | (C 00 + C α0) to obtain the stated estimate.
We will also need the relation between the Schwartz kernel and the symbol.
Proposition 5.
Suppose the Schwartz kernel of P is given by K(x,y), namely
Then the symbol p(x,ξ) of P is given by
Proof.
Write for smooth compactly supported f, g
That is, K(x, y) = ∫dξe i⟨x − y, ξ⟩ p(x, ξ), and ⟨Pf, g⟩ = ∫f(y)g(x)K(x, y)dydx. Denoting by \(\check{h}(x) =\int d\xi h(\xi ){e}^{i\langle x,\xi \rangle }\) the inverse Fourier transform, we can also write \(K(x,y) =\check{ p}(x,\bullet )(x - y)\;\Longleftrightarrow\;K(x,x - y) =\check{ p}(x,\bullet )(y)\), so
as claimed.
References
V. Guillemin, On some results of Gelfand in integral geometry, in Pseudodifferential Operators and Applications. Proceedings of Symposia in Pure Mathematics, vol. 43 (American Mathematical Society, Provindence, 1985), pp. 149–155
S. Helgason, The Radon Transform, 2nd edn. (Birkhuser, Boston, 1999)
L. Hörmander, Fourier integral operators. I. Acta Math. 127(1–2), 79–183 (1971)
H.B. Lawson Jr., M.L. Michelsohn, Spin Geometry. Princeton Mathematical Series, vol. 38 (Princeton University Press, Princeton, 1989)
E.T. Quinto, The dependence of the generalized Radon transform on defining measures. Trans. Am. Math. Soc. 257, 331–346 (1980)
Acknowledgements
I would like to thank Semyon Alesker for numerous illuminating discussions, Victor Palamodov for his help with Fourier integral operators, and Alon Nishry for a useful suggestion. Finally, I want to thank Vitali Milman for his encouragement and interest in this work.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2013 Springer Science+Business Media New York
About this chapter
Cite this chapter
Faifman, D. (2013). On a Stability Property of the Generalized Spherical Radon Transform. In: Ludwig, M., Milman, V., Pestov, V., Tomczak-Jaegermann, N. (eds) Asymptotic Geometric Analysis. Fields Institute Communications, vol 68. Springer, New York, NY. https://doi.org/10.1007/978-1-4614-6406-8_5
Download citation
DOI: https://doi.org/10.1007/978-1-4614-6406-8_5
Published:
Publisher Name: Springer, New York, NY
Print ISBN: 978-1-4614-6405-1
Online ISBN: 978-1-4614-6406-8
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)