Classical and quantum aspects of tomography

Paolo Facchi∗,† and Marilena Ligabò∗

∗
Dipartimento di Matematica, Università di Bari, I-70125 Bari, Italy
†
INFN, Sezione di Bari, I-70126 Bari, Italy
arXiv:1001.5169v1 [math-ph] 28 Jan 2010

Abstract. We present here a set of lecture notes on tomography. The Radon transform and some of
its generalizations are considered and their inversion formulae are proved. We will also look from a
group-theoretc point of view at the more general problem of expressing a function on a manifold in
terms of its integrals over certain submanifolds. Finally, the extension of the tomographic maps to
the quantum case is considered, as a Weyl-Wigner quantization of the classical case.
Keywords: Radon transform; integral geometry
PACS: 03.65.Wj, 42.30.Wb, 02.30.Uu

1. INTRODUCTION
We present here the notes of three lectures given by one of us at the XVIII International
Fall Workshop on Geometry and Physics in Benasque. The course considers some
aspects of tomographic mappings, a vast and very active area of different applications of
integral geometry.
As a motivation, we start with the physical problem of the tomographic imaging of
a two dimensional body by the measurements of the intensity attenuation of a beam
of particles passing through it. Then, we will introduce the Radon transform, which is
the key mathematical tool for reconstructing the tomographic map of both the Wigner
quasi-distribution of a quantum state and the probability distribution on the phase space
of a classical particle. The original transform was introduced by J. Radon [22], who
proved that a differentiable function on the three-dimensional Euclidean space can be
determined explicitly by means of its integral over the planes. We will consider the
straightforward generalization of the Radon transform to codimension-one hyperplanes
in Rn and will study its main properties. Moreover, we will prove the original Radon
inversion formula, by making use of potential theory.
We will then look at a first generalization of the Radon transform: the M 2 -transform
which, roughly speaking, is defined on a larger space. We will show how it can be
inverted very easily by using the Fourier transform. Then, we will unveil the relation
between Radon and M 2 , thus exhibiting an alternative, much simpler inversion formula.
Then, we will consider a broader framework and look at the more general problem of
expressing a function on a manifold in terms of its integrals over certain submanifolds.
This has become an important topic in integral geometry with many applications ranging
from partial differential equations, group representations, and X-ray technology. The
focus will be on invariant (or equivariant) transformations under some symmetry groups
from the space of functions on one geometrical space to the space of functions on
another geometrical space. We will also exhibit two examples that can be framed in
 I
detector
!
λ

I0
λ object
source

FIGURE 1. Tomographic experimental setup.

the above setting and work out their respective inversion formulae: the Radon transform
on codimension-d planes and the John transform on unit spheres in R3 .
We will then show a simple mechanism that allows to generalize the Radon trans-
form, and its inverse, to more general families of codimensione-one submanifolds. As
examples, we will look at two families of curves in the plane: the family of all circles
passing through a fixed point and at the family of all equilateral hyperbolas with one
fixed asymptote.
Finally, we will show how the above techniques can be used in the quantum case.
A straightforward generalization derives from the phase-space description of quantum
mechanics through Wigner quasi-distribution functions. We will see how the homodyne
detection scheme can be considered as the Weyl quantization of the Radon transform.
There is a huge literature on tomography and it is impossible to be complete in
the references. We apologize in advance for any shortcomings in this respect. There
are several classical references: we suggest for a group-theoretic approach the book
by Helgason [9] and for the connection with partial differential equations the book by
John [12]. Additional general references are Gel’fand and Shilov [8] and Strichartz [23].
Finally, for an up-to-date overview of the many applications of the Radon transform we
refer to the book by Deans [4], which includes also an English translation of the 1917
paper by Radon, that announced the transform’s discovery.

2. TOMOGRAPHIC IMAGING IN THE PLANE

Consider a body in the plane R2 , and consider a beam of particles (neutrons, electrons,
X-rays, etc.) emitted by a source. Assume that the initial intensity of the beam is I0 .
When the particles pass through the body they are absorbed or scattered and the intensity
of the beam traversing a length ∆s decreases by an amount proportional to the density of
the body µ, namely
∆I = −∆s µ(s), (1)
so that  Zs 
I(s) = I0 exp − µ(r) dr . (2)
0

A detector placed at the exit of the body measures the final intensity I(s), and then from
Z s
I(s)
− ln = µ(r) dr (3)
I0 0

one can record the value of the density integrated on a line. If another ray with a different
direction is considered, with the same procedure one obtains the value of the integral of
the density on that line. See Fig 1.
The mathematical model of the above setup is the following: Given a normalized
positive and smooth function f = f (q, p), namely f (q, p) ≥ 0 and
Z
f (q, p) dq dp = 1, (4)
R2

define for all possible lines λ in the plane R2 the quantity

Z
]
f (λ ) = f (x) dm(x), (5)
λ

where dm is the Euclidean measure on the line λ . In this way, we have defined an
operator ] that maps a smooth function f on the plane R2 into a function f ] on P2 , the
manifold of the lines in R2 . Along with the transform f 7→ f ] , it is convenient to consider
also the dual transform g 7→ g[ which to a function g on P2 associates the function g[ on
R2 given by Z
g[ (x) = g(λ ) dµ(λ ), (6)
x∈λ

where dµ is the measure on the compact set {λ ∈ P2 |x ∈ λ } which is invariant under

the group of rotations around x and for which the measure of the whole set is 1.
We ask the following question: If we know the family of integrals ( f ] (λ ))λ ∈P2 , can we
reconstruct the total density function f ? The answer is affirmative and in the following
we will see how to obtain this result.
In order to write the integral (5) in a more convenient form, let us parameterize P2 in
the following way
R × S → P2 , (X, ξ ) 7→ λ (7)
where X ∈ R is the (signed) distance of the line λ from the origin, and
ξ = (cos θ , sin θ ) ∈ S is a unit vector perpendicular to λ . This means that the Cartesian
equation of λ is
λ = {(q, p) ∈ R2 | X − q cos θ − p sin θ = 0}. (8)
By considering a pair (X,Y ) of Cartesian axes associated to the line λ , where X is
parallel to ξ , and Y is perpendicular to ξ , we obtain the following parametric equation
for λ
q = X cos θ −Y sin θ , p = X sin θ +Y cos θ . (9)
 p

λ
Y X

ϑ
q

FIGURE 2. Parametrization of λ using its signed distance X from the origin and a unit vector ξ =
(cos θ , sin θ ) perpendicular to the line.

See Fig. 2. With a little abuse of notation, we have

Z
] ]
f (λ ) = f (X, ξ ) = dY f (X cos θ −Y sin θ , X sin θ +Y cos θ ), (10)
R

that is called the Radon transform of f . Observe that the (7) is a double covering of P2 ,
because the pairs (X, ξ ) and (−X, −ξ ) correspond to the same line λ . However, one has
f ] (X, ξ ) = f ] (−X, −ξ ) (11)
and thus f ] depends only on the line λ and not on its parameterization.

3. RADON TRANSFORM
Let us generalize the above definitions to the case of a n-dimensional space. Let f be a
function defined on Rn , integrable on each hyperplane in Rn and let Pn be the manifold
of all hyperplanes in Rn . The Radon transform of f is defined as follows
Z
f ] (λ ) = f (x) dm(x), (12)
λ

where dm is the Euclidean measure on the hyperplane λ . Thus we have an operator ],

the Radon transform, that maps a function f on Rn into a function f ] on Pn , namely
f → f ] . Its dual transform, also called back projection operator, g 7→ g[ associates to a
function g on Pn the function g[ on Rn given by
Z
[
g (x) = g(λ ) dµ(λ ), (13)
x∈λ

where dµ is the unique probability measure on the compact set {λ ∈ Pn |x ∈ λ } which

is invariant under the group of rotations around x.
 Rn

λ
ξ
X

FIGURE 3. Parametrization of the hyperplane λ using its signed distance X from the origin and a unit
vector ξ perpendicular to λ .

Let us consider the following covering of Pn

R × Sn−1 → Pn , (X, ξ ) 7→ λ , (14)

where Sn−1 is the unit sphere in Rn . Thus, the equation of the hyperplane λ is

λ = {x ∈ Rn | X − ξ · x = 0}, (15)

with a · b denoting the Euclidean inner product of a, b ∈ Rn . See Fig. 3. Observe that
the pairs (X, ξ ), (−X, −ξ ) ∈ R × Sn−1 are mapped into the same hyperplane λ ∈ Pn .
Therefore (14) is a double covering of Pn . Thus Pn has a canonical manifold structure
with respect to which this covering mapping is differentiable. We thus identify contin-
uous (differentiable) functions g on Pn with continuous (differentiable) functions g on
R × Sn−1 satisfying g(X, ξ ) = g(−X, −ξ ).
We will work in the Schwartz space S (Rn ) of complex-valued rapidly decreasing
functions on Rn . We recall that f ∈ S (Rn ) if and only if f ∈ C∞ (Rn ) and for all
multiindices α = (α1 , . . . , αn ) ∈ Nn and m ≥ 0 results

sup ||x|m Dα f (x)| < +∞, (16)

x∈Rn

where |x| denotes the Euclidean x, and

Dα f (x) := (∂xα11 . . . ∂xαnn f )(x). (17)

In analogy with S (Rn ) we define S (R × Sn−1 ) as the space of C∞ functions g on

R × Sn−1 which for any integers m, α ≥ 0 and any differential operator D on Sn−1 satisfy

∂ α (Dg)
sup |X|m (X, ξ ) < +∞. (18)
X∈R,ξ ∈Sn−1 ∂ Xα
The space S (Pn ) is then defined as the set of g ∈ S (R × Sn−1 ) satisfying g(−X, −ξ ) =
g(X, ξ ).
If f ∈ S (Rn ) its Radon transform (12) can be written in a simple way, by using the
Dirac δ distribution, as
Z
]
f (X, ξ ) = f (x) δ (X − ξ · x) dx = hδ (X − ξ · x)i f , (19)
Rn

In the same way if g ∈ S (Pn ) one has

Z
[
g (x) = g(X, ξ ) δ (X − ξ · x) dXdξ . (20)
R×Sn−1

Using (19) and (20) it is easy to check the duality relation between the Radon transform
] and its dual [, namely for every f ∈ S (Rn ) and g ∈ S (Pn ) one gets
Z Z
]
dXdξ f (X, ξ ) g(X, ξ ) = dx f (x) g[ (x). (21)
R×Sn−1 Rn

Indeed,
Z Z Z
]
dXdξ f (X, ξ ) g(X, ξ ) = dXdξ dx f (x) δ (X − ξ · x) g(X, ξ )
R×Sn−1 R×Sn−1 Rn
Z
= dx f (x) g[ (x). (22)
Rn

Now, consider a unit vector ξ ∈ Sn−1 . We define the tomogram of f along ξ the
function
fξ] (X) := f ] (X, ξ ). (23)
Physically, this means that we fix a direction ξ and we look at the variation of the
intensity of all the rays that interact with the body in that direction. We also define the
family ( fξ] )ξ ∈Sn−1 of tomograms associated to f .
The following result shows how the Radon transform maps a joint probability den-
sity function f of n random variables into a family of probability density functions
( fξ] )ξ ∈Sn−1 of a random variable.
Lemma 1. Let f ∈ S (Rn ) be a nonnegative and normalized function, namely
Z
f ≥0 and f (x) dx = 1. (24)
Rn

Then, for all ξ ∈ Sn−1 it results that

Z
fξ] ≥ 0 and fξ] (X) dX = 1. (25)
R

Proof. Let us fix ξ ∈ Sn−1 . Observe that for all X ∈ R

Z
fξ] (X) = ]
f (X, ξ ) = f (x) δ (X − ξ · x) dx ≥ 0. (26)
Rn
Moreover,
Z Z Z Z
fξ] (X) dX = dX f (x) δ (X − ξ · x) dx = f (x) dx = 1. (27)
R R Rn Rn

The Radon transform has a peculiar behavior under translation. Given a displacement
a ∈ Rn , consider the translation operator defined by

(τa f )(x) = f (x − a). (28)

One has the following property

Lemma 2. Let f ∈ S (Rn ) and a ∈ Rn . Then for any ξ ∈ Sn−1

(τa f )]ξ = τξ ·a fξ] . (29)

Proof. By setting y = x − a we get

Z Z
(τa f )]ξ (X) = f (x − a) δ (X − ξ · x) dx = f (y) δ (X − ξ · y − ξ · a) dy
Rn Rn
= fξ] (X − ξ · a).

The following lemma, which is the infinitesimal version of Lemma 2, shows how
the Radon transform and its dual intertwine the Laplacian on Rn , ∆ = ∑i ∂x2i with the
Laplacian on R, ∂X2 .
Lemma 3. Let f ∈ S (Rn ) and g ∈ S (Pn ). Then
[
∂2 f] ∂ 2g

]
(∆ f ) = and = ∆g[ . (30)
∂ X2 ∂ X2
Proof. By Lemma 2 we get

(τa f )] (X, ξ ) = f ] (X − ξ · a, ξ ). (31)

Therefore, by considering an orthonormal basis (e1 , . . . , en ) of Rn , for all X ∈ Rn and

ξ ∈ Sn−1 one has that
] ]
f − τhei f ∂ f]
 
∂f
(X, ξ ) = lim (X, ξ ) = ξi (X, ξ ) (32)
∂ xi h→0 h ∂X
and thus
] ]
∂2 f ∂2 f]
 
∂ ∂f
(X, ξ ) = ξi (X, ξ ) = ξi2 (X, ξ ). (33)
∂ xi2 ∂X ∂ xi ∂ X2
Therefore, since ξ ∈ Sn−1 , by summing over i we have the first equality in (30).
As for the second relation, observe that
Z Z
[
g (x) = g(X, ξ ) δ (X − ξ · x) dXdξ = g(ξ · x, ξ ) dξ , (34)
R×Sn−1 Sn−1

whence

∂ g[ ∂g ∂ 2 g[ ∂ 2g
Z Z
(x) = (ξ · x, ξ ) ξi dξ , (x) = (ξ · x, ξ ) ξi2 dξ . (35)
∂ xi Sn−1 ∂ X ∂ xi2 Sn−1 ∂ X 2

Therefore,
[
∂ 2 g[
n
∂ 2g ∂ 2g
Z 
[
∆g (x) = ∑ 2
(x) = 2
(ξ · x, ξ ) dξ = (x). (36)
i=1 ∂ xi Sn−1 ∂ X ∂ X2

The Radon transform is closely connected with the Fourier transform. Let us recall
some basic facts about the Fourier transform. For any function g ∈ S (Rm ) with m ∈ N∗ ,
we define its Fourier transform as
Z
ĝ(k) = g(x) e−ik·x dx, (37)
Rm

with k ∈ Rm . It is not difficult to prove that the Fourier transform is a bijection of S (Rm )
onto itself, and its inverse reads
dk
Z
ǧ(x) = g(k) eik·x . (38)
Rm (2π)m

In short, for any g ∈ S (Rm ), one has that ĝˇ = ǧˆ = g. Using standard duality arguments
the Fourier transform can be extended to the space of temperated distributions S 0 (Rm )
(and thus also to its subspace of square summable functions L2 (Rn ), where it can also
be proved to be unitary). In particular, we recall that δ̂ = 1 and thus
dk
Z
δ (x) = eik·x . (39)
Rn (2π)n

We will show that the n-dimensional Fourier transform of a function f on Rn is the 1-

dimensional Fourier transform of its tomogram fξ] for all ξ ∈ Sn−1 . This result is known
as the Fourier slice theorem.
Lemma 4. Consider f ∈ S (Rn ), τ ∈ R and ξ ∈ Sn−1 . Then

fˆ(τξ ) = fξ] (τ) (40)

b
Proof. The thesis follows immediately from the definitions. Indeed
Z Z Z
fξ] (τ) = dX e −iτX
fξ] (X) = dX e −iτX
dx f (x) δ (X − ξ · x)
b
ZR Z R Rn Z

= dx f (x) dX e−iτX δ (X − ξ · x) = dx f (x) e−iτξ ·x = fˆ(τξ ). (41)

Rn R Rn

Now we will address the problem of the range of the Radon transform with domain
S (Rn ). The following lemma is crucial.
Lemma 5. If f ∈ S (Rn ), then for any integer k ∈ N and any ξ = (ξ1 , . . . , ξn ) ∈ Sn−1 ,
the integral of its tomogram,
Z
I f (ξ ) = fξ] (X) X k dX, (42)
R

is a homogeneous polynomial of degree k in the Cartesian components ξ1 , . . . , ξn .

Proof. This is immediate from the relation
Z Z Z Z
] k k
f (X, ξ ) X dX = dX X dx f (x) δ (X − ξ · x) = f (x) (ξ · x)k dx. (43)
R R Rn Rn

In accordance with this lemma we define the space SH (Pn ) as follows: g ∈ SH (Pn )
if and only if g ∈ S (Pn ) and for any integer k ∈ N the integral
Z
Ig (ξ ) = g(X, ξ ) X k dX (44)
R

is a homogeneous polynomial of degree k in ξ1 , . . . , ξn .

Theorem 1. The Radon transform is a linear one-to-one mapping of S (Rn ) onto
SH (Pn ).
The proof uses the well-known fact that the Fourier transform maps S (Rn ) onto itself
and Lemma 5. For more details see [9]. In the next section we will give a constructive
proof, by explicitly exhibiting the inverse mapping.

4. INVERSION FORMULA
Now we want to obtain the inversion formula, namely we want to prove that one can
recover a function f on Rn if all its tomograms ( fξ] )ξ ∈Sn−1 are known. In order to get
this result we need three preliminary lemmata.
Lemma 6. Let f ∈ S (Rn ) and V (x) = 1/|x|, x ∈ Rn , x 6= 0 . Then

( f ] )[ (x) = an f ∗V, (45)

 Rn
potential (f � )� (x)
x
an
y |x − y|

charge distribution f (y)

FIGURE 4. ( f ] )[ (x) is the potential at x generated by the charge distribution f .

where an depends only on the dimension n, and ∗ denotes the convolution product,
Z
( f ∗ g)(x) = f (y) g(x − y) dy. (46)
Rn

Proof. Observe that

Z Z Z
( f ] )[ (x) = dξ f ] (ξ · x, ξ ) = dξ dy f (y) δ (ξ · x − ξ · y)
Sn−1 Sn−1 Rn
ds isξ ·(x−y)
Z Z Z
= dξ dy f (y) e . (47)
Sn−1 Rn R 2π

Now, with the substitution x − y = |x − y|ω, ω ∈ Sn−1 , and t = s|x − y|, the integral (47)
becomes
dt 1
Z Z Z Z
dξ dy f (y) eitξ ·ω = an dy f (y)V (x − y), (48)
Sn−1 Rn R 2π |x − y| Rn

where
dt
Z Z
an = dξ eitξ ·ω . (49)
R 2π Sn−1
By symmetry, it is apparent that an is independent of ω.
The physical meaning of Lemma 6 is the following: if f is a charge distribution,
then the potential at the point x generated by that charge is exactly ( f ] )[ (x), see Fig. 4.
Notice, however, that the potential of a point charge scales always as the inverse distance
independently of the dimension n, and thus it is Coulomb only for n = 3. The only
dependence on n is in the strength of the elementary charge an . This fact will be crucial in
the following: we will see that the associated Poisson equation involves an n-dependent
(fractional) power of the Laplacian.
Lemma 7. Let V (x) = 1/|x|, x ∈ Rn , x 6= 0 . Then its Fourier transform is
bn
V̂ (k) = . (50)
|k|n−1
where bn depends only on n
Proof. One gets
1 −ik·x 1 −i|k|ω·x dy 1 −iω·y bn
Z Z Z
V̂ (k) = dx e = dx e = n−1
e = n−1 , (51)
Rn |x| Rn |x| Rn |k| |y| |k|

where we have set k = |k|ω with ω ∈ Sn−1 , y = |k|x, and

e−iω·x
Z
bn = dx . (52)
Rn |x|
Moreover, bn is independent of ω by symmetry.
Lemma 8. Let an and bn as in Lemmata 6 and 7. Then
an bn = 2(2π)n−1 . (53)
Proof. Since the Fourier transform is a bijection, we use the trick of going back and
forth. From Lemma 7 we have
1 dk bn ik·x dk
Z Z
V (x) = = V̂ˇ (x) = V̂ (k) eik·x = e . (54)
|x| Rn (2π)n Rn |k|
n−1 (2π)n

By setting x = |x|ω and p = |x|k = tξ , with ω, ξ ∈ Sn−1 and t ∈ R+ , Eq. (54) becomes
1 bn dk ik·ω|x| bn dp ip·ω
Z Z
= n n−1
e = e
|x| (2π) Rn |k| (2π) |x| Rn |p|n−1
n

bn dξ itξ ·ω
Z Z
n−1
= n
dt t n−1
e
(2π) |x| R+ Sn−1 t
bn dt an bn 1
Z Z
= n−1
dξ eitξ ·ω = n−1
, (55)
2(2π) |x| R 2π Sn−1 2(2π) |x|
where in the last equality we used Eq. (49) of Lemma 6. Therefore,
an bn
=1 (56)
2(2π)n−1
and the thesis follows.
Now we are ready to prove the inversion formula for the Radon transform.
Theorem 2. Let f ∈ S (Rn ). Then
1 n−1
f (x) = (−∆) 2 ( f ] )[ (x) (57)
2(2π)n−1
where (−∆)α , with α > 0, is a pseudodifferential operator whose action is
dk
Z
α
((−∆) f )(x) = |k|2α fˆ(k) eik·x . (58)
Rn (2π)n
Proof. First of all observe that if n = 1 statement (62) is trivial, so assume that n > 1.
Let p > 0, we get by Lemma 6
dk dk
Z Z
p
(−∆) V (x) = n
|k|2p V̂ (k) eik·x = bn n
|k|2p−n+1 eik·x , (59)
Rn (2π) Rn (2π)

so that, if 2p = n − 1, we have
n−1
(−∆) 2 V (x) = bn δ (x). (60)

Therefore, again by Lemma 6, we have

n−1
 n−1

(−∆) 2 ( f ] )[ (x) = an f ∗ (−∆) 2 V (x) = an bn ( f ∗ δ )(x) = an bn f (x). (61)

From (53) we get the inversion formula (62).

n−1
Equation (57) says that, modulo the final action of (−∆) 2 , the function f can be
recovered from its Radon transform f ] by the application of the dual mapping [: first
one integrates over the set of points in a hyperplane and then one integrates over the set
of hyperplanes passing through a given point. Explicitly we get
1
Z
n−1
f (x) = n−1
(−∆) 2 f ] (ξ · x, ξ ) dξ , (62)
2(2π) Sn−1

which has the following remarkable interpretation. Note that if one fixes a direction
ξ ∈ Sn−1 , then the function f ] (ξ · x, ξ ) is a plane wave, i.e. a function constant on each
plane perpendicular to ξ . Therefore, Eq. (62) gives a representation of f in terms of
a continuous superposition of plane waves. A well-known analogous decomposition is
given by Fourier transform. When n = 3, one recovers the inversion formula originally
found by Radon [22]
1
Z
f (x) = − 2 ∆ f ] (ξ · x, ξ ) dξ . (63)
8π S2

4.1. Localization
Now we consider the following question on the localization properties of the Radon
transform: given a function f can we reconstruct it in a compact region Ω, using only
the tomograms of f on hyperplanes λ that pass through Ω? By looking at the inversion
formula (62) one would naively think that the answer be positive. By following a nice
argument given by John [12], we will show that in fact this cannot be true in two (and,
in general, even) dimensions. The reason is the well-known fact that waves cannot be
localized in two dimensions: a pebble dropped in a pond produces ripples that propagate
along expanding disks. On the other hand, in one dimension waves remain localized: a
local perturbation of the pressure in an organ pipe propagates by translating its support.
This is a fortunate accident that allows us to enjoy a Bach’s fugue without hearing an
endless annoying echo of each tune.
 ξ

−t
0
t X

FIGURE 5. A function f supported in a ball and its tomogram in the direction ξ .

Consider the following Cauchy problem for the wave equation in R2 :

∂t2 f − ∆ f = 0, f (0, x) = g(x), ∂t f (0, x) = 0, (64)
and assume that the initial condition g has its support in the ball B(0, ε) = {|x| < ε} with
center 0 and radius ε > 0. Let us fix a time t > ε. The solution to this equation is given
by the celebrated Poisson’s formula [5]
1 g(y) + ∇g(y) · (y − x)
Z
f (t, x) = p dy, (65)
2πt B(x,t) t 2 − |y − x|2
where ∇ denotes the gradient in R2 . Therefore the support of f (t, ·) is a subset of
B(0,t + ε). On the other hand, if we consider the Radon transform of (64) we get, for
every ξ ∈ Sn−1 , a Cauchy problem for the wave equation in R

∂t2 fξ] − ∂X2 fξ] = 0, fξ] (0, X) = g]ξ (X), ∂t fξ] (0, X) = 0. (66)

Observe that g]ξ has its support in (−ε, ε). The solution of Eq. (66) is

1 ] 
fξ] (t, X) = ]
g (X − t) + gξ (X + t) , (67)
2 ξ
thus the function fξ] (t, ·) has its support in (−t − ε, −t + ε) ∪ (t − ε,t + ε). So we see that
if one wants to reconstruct the function f (t, ·) in the set B(0,t − ε) one cannot consider
only the tomograms on the lines that intersect B(0,t − ε), because they are all equal to
zero. See Fig. 5. Indeed, from f ] (t, X, ξ ) = 0 for |X| < t − ε we have that
Z
] [
( f ) (t, x) = f ] (t, ξ · x, ξ ) dξ = 0, for |x| < t − ε. (68)
S
On the other hand, the inversion formula (62) for n = 2 yields
1
f (t, x) = (−∆)1/2 ( f ] )[ (t, x) 6= 0, for |x| < t − ε. (69)
4π
The seeming contradiction derives from the fact that the pseudodifferential operator
(−∆)1/2 is nonlocal. In fact, by (58) it is an integral operator that depends also on the
values of ( f ] )[ (t, x) for |x| > t − ε. Of course, this is a peculiar behavior of the even
dimension. For odd n in the inversion formula (62) we have an integer power of the
Laplacian that is a bona fide differential operator and thus is local.

5. M 2 -TRANSFORM
In this section we introduce another covering of the homogeneous space Pn that uses an
additional parameter, which is absent in the classical tomographic framework of Radon.
This choice of parametrization of Pn gives an easier inversion formula, although this
approach is completely equivalent to that considered in the previous section.
We define the following covering
R × Rn∗ → Pn , (X, µ) 7→ λ , (70)
where the Cartesian equation for λ is
λ = {x ∈ Rn | X − µ · x = 0}. (71)
Of course the additional parameter with respect to Radon is |µ|. Now, if f ∈ S (Rn ) we
can define the M 2 -transform of f as follows
Z
M2
f (X, µ) = f (x) δ (X − µ · x) dx = hδ (X − µ · x)i f . (72)
Rn
This tomographic mapping was introduced by Man’ko and Marmo, who gave seminal
contribution towards its significance [16, 17, 18]. Also for this transform we can define
2 2
the tomograms, precisely for every µ ∈ Rn∗ we define f µM (X) := f M (X, µ). It easy
to check that also this transform maps a probability density on Rn into a family of
probability density functions R, namely if f ∈ S (Rn ) is a positive and normalized
2
function, then for all µ ∈ Rn∗ , f µM is positive and normalized.
Let us now consider the relations between Radon and M 2 transform. Obviously, from
2
f M (X, µ) one can immediately recover f ] (X, ξ ) by setting µ = ξ ∈ Sn−1 :
2
f ] (X, ξ ) = f M (X, ξ ). (73)
2
However, notice that, although f ] is the restriction of f M to the unit sphere Sn−1 , there is
actually a bijection between the two transforms. Therefore, they carry exactly the same
information. Indeed, since the Dirac distribution is positive homogeneous of degree −1,
i.e. δ (αx) = |α|−1 δ (x), for every α 6= 0, one gets from Eq. (72)
 
M2 1 M2 X µ
f (X, µ) = f , , (74)
|µ| |µ| |µ|
 2
for µ 6= 0. In words, the tomogram f M (X, µ) at a generic point µ ∈ Rn∗ is completely
determined by the tomogram at µ/|µ| ∈ Sn−1 . But the latter is nothing but the Radon
transform, by Eq. (73). Therefore, we get the bijection
 
] M2 M2 1 ] X µ
f (X, ξ ) = f (X, ξ ), f (X, µ) = f , . (75)
|µ| |µ| |µ|

Finally, let us look at an inversion formula as in the case of the Radon transform.
Theorem 3. Let f ∈ S (Rn ). Then

dXdµ
Z
2
f (x) = f M (X, µ) ei(X−µ·x) . (76)
Rn+1 (2π)n

Proof. Observe that

dτ iτ(X−µ·x)
Z Z Z
2
f M (X, µ) = dx f (x) δ (X − µ · x) = dx f (x) e
Rn Rn R 2π
dy dτ i(τX−µ·y)  y  (2π)n dy dτ i(τX−µ·y)
Z Z
= n+1
e f n
= n+1
e g(τ, y), (77)
Rn+1 (2π) τ |τ| Rn+1 (2π)

2
where g(τ, y) = (2π)n f (y/τ)/|τ|n . Therefore, f M (X, µ) = ǧ(X, −µ), and so

2
f M (τ, −x).
g(τ, x) = d (78)

Finally,

2
g(−1, −x) f M (−1, x) dX dµ M2
d Z
f (x) = n
= n
= n
f (X, µ) ei(X−µ·x) . (79)
(2π) (2π) R n+1 (2π)

Summarizing, the inversion formulae for the transforms (62) and (76) read
1
Z
n−1
f (x) = n−1
(−∆) 2 f ] (ξ · x, ξ ) dξ , (80)
2(2π) Sn−1
dXdµ
Z
2
f (x) = f M (X, µ) ei(X−µ·x) , (81)
Rn+1 (2π)n

respectively. While formula (81), which is nothing but a Fourier transform, is quite easy
to handle, the inversion formula (80) is in general very hard to tackle, especially for even
n, due to the presence of a fractional Laplacian. For an explicit example, see [6].
 6. GROUP-THEORETIC FRAMEWORK
The general problem posed by the Radon transform is that of determining a function on a
manifold by means of its integrals over a certain family of submanifolds. Here, following
Helgason [9] we will construct a quite general framework for the above problem.
Let us start from the simple observation that the Euclidean group E(n) acts transitively
in a natural way both on the affine space Rn and on the hyperplane space Pn . The
Euclidean group is the group of isometries associated with the Euclidean metric and
is the semidirect product of the orthogonal group O(n) of rotations, extended by the
abelian group of translations Rn , namely E(n) = O(n) n Rn .
Fix an origin 0 in Rn (a reference point) and an origin λ0 in Pn (a reference hyper-
plane). It will be convenient to choose an hyperplane λ0 passing through the origin 0.
Transitivity means that any element of Rn (Pn ) can be obtained by an Euclidean motion
of the origin 0 (λ0 ), namely

∀x ∈ Rn : x = g · 0, for some g ∈ E(n),

∀λ ∈ Pn : λ = γ · λ0 , for some γ ∈ E(n), (82)

where the dot · denotes the action of E(n). The action is not free, i.e. one can have
g · x = h · x for some x ∈ Rn and g, h ∈ E(n) with g 6= h. Therefore, let us consider the
isotropy subgroup of the origin 0 (also called the stabilizer subgroup) as the set of all
elements in E(n) that fix 0:

E(n)0 = {g ∈ E(n) | g · 0 = 0} = O(n). (83)

It is given by the rotation group (around 0). On the other hand, the isotropy subgroup of
the reference hyperplane λ0 passing through the origin is

E(n)λ0 = {g ∈ E(n) | g · λ0 = λ0 } = Z2 × E(n − 1), (84)

where E(n − 1) is the Euclidean group of the (n − 1)-dimensional affine subspace λ0

(translations and rotations in λ0 ), while Z2 ' O(1) is the group whose elements are
the identity and the reflection in λ0 . See Fig. 6. Now, given an element x ∈ Rn , by
transitivity we get x = g · 0, for some g ∈ E(n). Therefore we can identify x with
the left coset gE(n)0 = gO(n). Analogously, for all λ ∈ Pn we get the identification
λ ' γE(n)λ0 = γ(Z2 × E(n − 1)), for some γ ∈ E(n). Therefore, we get that

Rn ' E(n)/E(n)0 = E(n)/O(n),

Pn ' E(n)/E(n)λ0 = E(n)/(Z2 × E(n − 1)), (85)

are homogeneous spaces of E(n).

We want to describe the incidence relation in this framework. Namely, given an
hyperplane λ ∈ Pn and a point x ∈ Rn , when x lies in λ , that is x ∈ λ ? It is not difficult
to see that this is true if and only if their corresponding cosets, considered as subsets of
E(n), have a point in common. In equations, given x ' gO(n) and λ ' γ(Z2 × E(n − 1)),
one has
x∈λ ⇔ gO(n) ∩ γ(Z2 × E(n − 1)) 6= 0. / (86)
 Rn
λ0

E(n)λ0

E(n)0

FIGURE 6. Origins 0 and λ0 and their isotropy groups.

Indeed, one gets

g · 0 ∈ γ · λ0 ⇔ g · 0 = γh · 0, for some h ∈ Z2 × E(n − 1)

⇔ gk = γh, for some k ∈ O(n)
⇔ gO(n) ∩ γ(Z2 × E(n − 1)) = {gk}. (87)

In the Cartesian product Rn × Pn we can consider the following set, called the inci-
dence relation,
I = {(x, λ ) ∈ Rn × Pn | x ∈ λ } . (88)
Consider now the restrictions to I of the natural projections p : I ⊂ Rn × Pn → Rn and
π : I ⊂ Rn × Pn → Pn . Obviously, since every point of Rn lies in some hyperplane of Pn
and every hyperplane passes through some point, we get p(I) = Rn and π(I) = Pn , so
that p and π are onto. Now, for any x ∈ Rn and λ ∈ Pn consider the sets

x[ = {λ ∈ Pn | (x, λ ) ∈ I} , λ ] = {x ∈ Rn | (x, λ ) ∈ I} . (89)

They represent the set of all hyperplanes passing through x and the set of all points lying
in λ , respectively. The crucial fact is that the mappings

x 7→ x[ = π ◦ p−1 (x), λ 7→ λ ] = p ◦ π −1 (λ ), (90)

are in fact injective. Indeed, we obviously have

\ [
x= λ, λ= x. (91)
λ ∈x[ x∈λ ]

Therefore, any point x ∈ Rn can be uniquely identified with a set of hyperplanes x[ ⊂ Pn

(those incident to x), and conversely any hyperplane λ ∈ Pn can be uniquely identified
with a set of points λ ] ⊂ Rn (those incident to λ ). We say that Rn and Pn are homoge-
neous spaces in duality. In this setting we can rewrite the Radon transform and its dual
as follows Z Z
] [
f (λ ) = f (x) dm(x), ϕ (x) = ϕ(λ ) dµ(λ ). (92)
λ] x[
Notice that from Eq. (87) we get for any pair (x, λ ) ∈ I
(x, λ ) ' (gO(n), gk(Z2 × E(n − 1))) = (glO(n), gkl(Z2 × E(n − 1))) , (93)
for some k ∈ O(n) and for all l ∈ L, where
L = E(n)0 ∩ E(n)λ0 = O(n) ∩ (Z2 × E(n − 1)) = Z2 × O(n − 1) (94)
is the subgroup that fixes both 0 and λ0 . Therefore, we get that also the incidence relation
I is a homogeneous space, namely
I ' E(n)/L = E(n)/(Z2 × O(n − 1)). (95)
The Radon transform on the homogeneous spaces in duality Rn and Pn can be conve-
niently described by the following double fibration:

I ' E(n)/(Z2 × O(n − 1))

 HH
 H
p  HH π
 H
  HH
H

 HH
j
R ' E(n)/O(n)
n
P ' E(n)/(Z2 × E(n − 1))
n

(96)
Now let us generalize the above construction. Let G be a locally compact group, let K
and H be two closed subgroups of G, and let L = K ∩ H. Assume that KH ⊂ G is closed.
Let us consider the following two left coset spaces
X = G/K and Λ = G/H, (97)
and define the incidence relation
I = {(x, λ ) ∈ X × Λ | x ∩ λ 6= 0}
/ . (98)
Define also for all x ∈ X and λ ∈ Λ
x[ = π ◦ p−1 (x) = {λ ∈ Λ | (x, λ ) ∈ I} , λ ] = p ◦ π −1 (λ ) = {x ∈ X | (x, λ ) ∈ I} , (99)
where p and π are the restrictions to I of the natural projections of X × Λ. It is not
difficult to prove that x[ and λ ] are closed subsets, and that the sets
KH = k ∈ K | kH ∪ k−1 H ⊂ HK , HK = h ∈ H | hK ∪ h−1 K ⊂ KH
 
(100)
are subgroups of K and H, respectively. Moreover, Helgason [9] proved the following
Theorem 4. The following propositions are equivalent:
1.
L = KH = HK (101)
 2. The mappings
x 7→ x[ , λ 7→ λ ] (102)
are injective.
Therefore, under the transversality assumption (101), any element x of X can be
uniquely identified with the subset x[ of Λ, and any element λ ∈ Λ can be uniquely
identified with the subset λ ] ⊂ X. Thus, X and Λ are homogeneous spaces in duality.
Exactly as in the case of hyperplanes in Rn , we have the identification I = G/L, with
L = K ∩ H, that is easily proved via the bijection G/L → I : gL 7→ (gK, gH). As a
consequence, also in the general case we have the following double fibration, which
is the generalization of (96):

I = G/(K ∩ H)
p  HH π
 H (103)

 H
j
X = G/K Λ = G/H

Let x0 and λ0 be the origins of X and Λ, respectively. Let us assume that the sets x0[ =
K/L and λ0] = H/L have positive measures invariant under K and H, respectively. Under
the transversality assumption (101), by using the action of G one can uniquely induce
a K-invariant (H-invariant) positive measure µ (m) on each x[ (λ ] ). Therefore, one can
define a Radon transform and its dual as in Eq. (92) for the double fibration (103).
A final remark is in order. The above construction in terms of homogeneous spaces in
duality is quite suggestive and represents a good framework for encompassing various
examples of tomographic mappings. However, as noted by Helgason himself [9], it is
doubtful that it could capture the large variety of results in this field. In particular, there
remain several open problems that should be addressed case by case. In particular, one
should find suitable function spaces on X and Λ where the Radon transform f 7→ f ] is
invertible, and – an even more difficult task – find an inversion formula.

7. EXAMPLES
In this section we consider two examples of tomographics mappings a functions f on Rn ,
in terms of its integral over certain submanifolds of Rn and prove the related inversion
formulae. In the first example we consider the planes of codimension d, 0 < d < n, in
Rn , while in the second one we consider the unit spheres in R3 .

7.1. (n-d)-planes in Rn
Let us consider the space Pnd of all the affine subspaces λ of Rn with dimension n − d,
0 < d < n. In this example the homogeneous spaces in duality are Rn and Pnd and the
double fibration in Eq. (103) reads

I = E(n)/(O(d) × O(n − d))

 HH
 H
p  HH π
 H
  HH
H



HH
j
H
R = E(n)/O(n)
n
Pnd
= E(n)/(O(d) × E(n − d))
(104)
Indeed, G = E(n) and K = O(n), the isotropy group of the origin 0, hence X = Rn =
E(n)/O(n). Moreover, the isotropy group of a reference subspace λ0 ∈ Pnd passing trough
the origin is H = O(d) × E(n − d), and thus Λ = Pnd = E(n)/(O(d) × E(n − d)). Finally,
L = K ∩ H = O(d) × O(n − d), so that I = E(n)/(O(d) × O(n − d)).
The Radon transform has the same form (12) of the codimension-one case, where
dm(x) is the Euclidean measure on the codimension-d hyperplane λ . First of all notice
that, since a codimension-one hyperplane can be viewed as a disjoint union of parallel
codimension-d planes, it is obvious that the transform f 7→ f ] is injective. In other
words, the information recorded is redundant and we can recover the function f by
integrating away d − 1 degrees of freedom and then inverting the usual Radon transform
on hyperplanes. We will follow here a more direct, different approach.
Observe that λ can be obtained by the intersection of d orthogonal hyperplanes in Rn ,
namely λ can be described as the space of the solutions of the following linear system
λ = {x ∈ Rn | X j = ξ j · x, j = 1, . . . , d}. (105)

where X j ∈ R and ξ j ∈ Sn−1 , ξ j · ξk = δ j,k , with j, k = 1, . . . , d. Of course, this repre-

sentation of λ is not unique, in fact we can obtain λ as the solution of another linear
system
λ = {x ∈ Rn |Y j = η j · x, j = 1, . . . , d}. (106)
where for all i = 1, . . . , d
d d
ηi = ∑ ai j ξ j , Yi = ∑ ai j X j , (107)
j=1 j=1

and A = (ai j ) ∈ GL(d, R). Let us compute

d
ηi · η j = ∑ aik a jl ξk · ξl = (AAT )i j . (108)
k,l=1

Note that, since ηi ∈ Sn−1 , one must have that (AAT )ii = 1 for all i = 1, . . . , d.
Let f ∈ S (Rn ) and define the tomographic transform of f on λ ∈ Pnd as follows (with
the usual abuse of notation)
Z d
] ]
p
f (λ ) = f (Y, η) = f (x) ∏ δ (Y j − η j · x) det(ηk · ηl ) dx, (109)
Rn j=1
where Y = (Y1 , . . . ,Yd ) ∈ Rd and η = (η1 , . . . , ηd ) ∈ (Sn−1 )d . The definition of the Radon
transform in (109) does not depend on the parametrization of λ . Indeed, by writing
the parameterization (Y, η) in terms of a parameterization with orthogonal hyperplanes
(X, ξ ) through (107), we get
d
dt
Z
ei ∑ j t j (Y j −η j ·x) |det A|
p
∏ δ (Y j − η j · x) det(ηk · ηl ) = Rd (2π)d
j=1
dt dt
Z Z
= e i ∑ j t j ∑l a jl (Xl −ξl ·x)
|det A| = ei ∑l (∑ j a jl t j )(Xl −ξl ·x) |det A|
Rd (2π)d Rd (2π)d
d
ds
Z
= ei ∑l sl (Xl −ξl ·x) = ∏ δ (Xl − ξl · x). (110)
Rd (2π)d l=1

where we set s = AT t. Therefore,

Z d Z
] ]
f (Y, η) = f (X, ξ ) = f (x) ∏ δ (X j − ξ j · x) dx = f (x) dm(x) = f ] (λ ), (111)
Rn j=1 λ

as given in (12).
Let us consider also the dual of the Radon transform. If g ∈ S (Pnd ), we define
d
dY dη
Z
[
g (x) = g(Y, η) ∏ δ (Y j − η j · x) p . (112)
(R×Sn−1 )d j=1 det(ηk · ηl )

Theorem 5. Let f ∈ S (Rn ). Then

n−1 d
Γ d2



n−d Γ
f (x) = cn,d (−∆) 2 ( f ] )[ (x), where cn,d = 2
(113)
2n π (dn−d+n)/2 Γ n−d

,
2

with Γ denoting the Euler gamma function.

Proof. First of all observe that if n = 1 then necessarily d = 1, and the statement (113)
is trivial. So assume that n ≥ 1. Let 0 < α < n, then if V (x) = 1/|x|, with x ∈ Rn∗ , one
can easily prove that [14]
α


Γ
V[n−α = b V ,
n,α
α
where bn,α = 2 π α n/2 2

. (114)
Γ n−α
2

Let p > 0 and let us compute the following quantity

dk
Z
p d
(−∆) V (x) = n
|k|2p Vcd (k) eik·x (115)
Rn (2π)
so if 2p = n − d, we obtain
n−d
(−∆) 2 V d (x) = bn,n−d δ (x). (116)
Now let us compose the Radon transform with its dual. As in Lemma 6 we obtain
n−1
ds 2bn,1 2π 2
Z Z
] [ isξ ·ω
(f ) (x) = adn ( f d
∗V )(x), an = dξ e = =
, (117)
R 2π Sn−1 2π Γ n−1
2

and again an is independent of ω ∈ Sn−1 . Therefore, we have

n−d n−d
 n−d

(−∆) 2 ( f ] )[ = adn (−∆) 2 ( f ∗V d ) = adn f ∗ (−∆) 2 V d
= adn bn,n−d f ∗ δ = adn bn,n−d f (118)

and so we have Eq. (113) with cn,d = 1/(adn bn,n−d ), and this concludes the proof.

7.2. Spherical means in R3

Let us consider the case, studied by John [12], of a tomographic transform that maps a
function f on Rn into a function on the spheres in Rn with a fixed radius r. For simplicity
we consider the case n = 3 and r = 1.
The homogeneous spaces in duality, R3 and R̃3 = set of all unit spheres in R3
(parametrized by their centers), are described by the following double fibration:

I = E(3)/O(2)
p
 H
HHπ
 H (119)

 HH
j
R3 = E(3)/O(3) R̃3 = E(3)/Õ(3)

Indeed, G = E(3) and K = O(3), the isotropy group of the origin 0, hence X = R3 =
E(3)/O(3). Moreover, the isotropy group of a reference unit sphere λ0 ∈ R̃3 passing
trough the origin is H = Õ(3) (rotation around the center of λ0 ), and thus Λ = R̃3 =
E(3)/Õ(3). Finally, L = K ∩ H = O(3) ∩ Õ(3) = O(2) (rotation around the line joining
0 and the center of λ0 ), so that I = E(3)/O(2).
We define the tomogram of f as its average over the unit sphere with center x
1
Z
f ] (x) = f (x + ξ ) dξ , (120)
|S2 | S2

where S2 = 4π is the surface of S2 . We look for an inversion formula, namely we want

to reconstruct the function f using its tomograms on all possible unit spheres in R3 . This
problem is equivalent to the following integral equation for the function f :
1
Z
f (x + ξ ) dξ = g(x). (121)
|S2 | S2

In Ref. [12] John proved that the solution of the integral equation (121) is unique, under
mild regularity assumptions on f and g. Therefore, also this tomographic mapping is one
FIGURE 7. Inversion formula (128). Dotted lines are spheres with an odd radius, while dashed lines
are spheres with an even radius.

to one on S (R3 ). We will now explicitly construct its inverse. To this purpose, consider
the following Cauchy problem for the wave equation
∂t2 v − ∆v = 0, v(x, 0) = 0, ∂t v(x, 0) = g(x). (122)
The solution of this problem is
1
Z
(g) (g)
v(x,t) = tI (x,t), I (x,t) = 2 g(x + tξ ) dξ . (123)
|S | S2

We construct the function u(x,t) as follows

+∞
u(x,t) = − ∑ [∂t v(x, 2k + 1 + t) − ∂t v(x, 2k + 1 − t)] (124)
k=0

and define f (x) := ∂t u(x, 0). Observe that u(x, 0) = 0, thus u solves the Cauchy problem
∂t2 u − ∆u = 0, u(x, 0) = 0, ∂t u(x, 0) = f (x), (125)
whence u(x,t) = tI ( f ) (x,t). In particular, we have
+∞
u(x, 1) = I ( f ) (x, 1) = − ∑ [∂t v(x, 2k + 2) − ∂t v(x, 2k)] = ∂t v(x, 0) = g(x) (126)
k=0

and so we have proved that the function f solves Eq. (121). Moreover,
+∞
f (x) = ∂t u(x, 0) = −2 ∑ ∂t2 v(x, 2k + 1)
k=0
+∞ +∞
= −2 ∑ ∆v(x, 2k + 1) = −2∆ ∑ (2k + 1)I (g) (x, 2k + 1)
k=0 k=0
+∞
2k + 1
Z Z
= −2∆ ∑ dξ dη f (x + (2k + 1)ξ + η). (127)
k=0 |S2 |2 S2 S2
 y p

x q

FIGURE 8. Diffeomorphism of the plane: (q, p) ∈ R2 → (x, y) = ϕ(q, p) ∈ R2 .

Therefore, we finally obtain

+∞
2k + 1
Z
f (x) = −2∆ ∑ 2
dξ f ] (x + (2k + 1)ξ ), (128)
k=0 |S | S2

which is a sum of the spherical means over all the spheres centered at x and with odd
radius. See Fig. 7.

8. TOMOGRAMS ON HYPERSURFACES
A simple mechanism that allows nonlinear generalizations of the Radon transform is
the combination of the standard transform with a diffeomorphism of the underlying
Rn space [2]. Let us consider a function f (q) on the n-dimensional space q ∈ Rn . The
problem is to reconstruct f from its integrals over an n-parameter family of submanifolds
of codimension one.
We can construct such a family by diffeomorphic deformations of the hyperplanes (in
the x ∈ Rn space)
X − µ · x = 0, (129)
with X ∈ R and µ ∈ Rn . Let us consider a diffeomorphism of Rn
q ∈ Rn 7→ x = ϕ(q) ∈ Rn . (130)
The hyperplanes (129) are deformed by ϕ into a family of submanifolds (in the q space)
X − µ · ϕ(q) = 0. (131)
The case n = 2 is displayed in Fig. 8.
Given a probability density f˜(x) on the x space, the M 2 -transform can be rewritten as
Z Z
˜M2
f (X, µ) = f˜(x) δ (X − µ · x) dx = f˜(ϕ(q)) δ (X − µ · ϕ(q)) J(q) dq, (132)
Rn Rn
where
∂ xi ∂ ϕi (q)
J(q) = = (133)
∂qj ∂qj
is the Jacobian of the transformation.
Observe now that f˜(x) dx = f˜(ϕ(q)) J(q) dq, whence
f (q) = f˜(ϕ(q)) J(q) (134)
is a probability density. Therefore the tomograms along the codimension-1 submani-
folds (131) are given by
Z
f ϕ (X, µ) = hδ (X − µ · ϕ(q))i f = f (q) δ (X − µ · ϕ(q)) dq, (135)
Rn

with X ∈ R and µ ∈ Rn∗ .

The inverse transform follows by (76):
dX dµ
Z
f (q) = f˜(ϕ(q)) J(q) = f ϕ (X, µ) J(q) ei(X−µ·ϕ(q)) , (136)
Rn+1 (2π)n
with a modified kernel
∂ ϕi (q) i(X−µ·ϕ(q))
Kϕ (q; X, µ) = J(q) ei(X−µ·ϕ(q)) = e . (137)
∂qj
Therefore, a probability density distribution on Rn
Z
f (q) ≥ 0, f (q) dq = 1, (138)
Rn

produces tomograms f ϕ (X, µ) that are probability densities

Z
ϕ
f (X, µ) ≥ 0, f ϕ (X, µ) dX = 1, ∀µ ∈ Rn∗ . (139)
R

We will now consider two applications of these generalizations of the Radon transform.
For additional examples see [1, 2].

8.1. Circles in the plane

In the punctured (x, y) plane without the origin (0, 0), the conformal inversion
 
q p
(x, y) = ϕ(q, p) = , , (140)
q2 + p2 q2 + p2
maps the family of lines X − µx − νy = 0 into a family of circles

X(q2 + p2 ) − µq − ν p = 0, (141)
centered at C = (µ/2X, ν/2X) and passing through the origin (see Fig. 9). When X = 0
they degenerate into lines through the origin. The Jacobian reads
J(q, p) = (q2 + p2 )−2 , (142)
 p

q
q

FIGURE 9. Circular tomography (left panel) and hyperbolic tomography (right panel). All circles pass
through the origin and all hyperbolas have the vertical axis as asymptote.

whence the transformation is a diffeomorphism of the punctured plane R2∗ onto itself.
The singularity of the transformation at the origin (0, 0) is irrelevant for tomographic
integral transforms because it only affects a zero measure set.
Equations (135)-(136) become
  
ϕ µq νp
f (X, µ, ν) = δ X − 2 − , (143)
q + p2 q2 + p2 f
 
i X− µq
− νp
e q2 +p2 q2 +p2
Z
f (q, p) = f ϕ (X, µ) dX dµ dν. (144)
R3 (2π)2 (q2 + p2 )2

8.2. Hyperbolas in the plane

In the (x, y) plane the family of lines X − µx − νy = 0 is mapped into a family of
hyperbolas
µ
X − − ν p = 0, (145)
q
with asymptotes q = 0 and p = X/ν , by the transformation

(x, y) = ϕ(q, p) = q−1 , p .

(146)

For µ > 0 the hyperbolas are in the second and fourth quadrants, while for µ < 0 they
are in the first and third quadrants (see Fig. 9). When µ = 0 or ν = 0 they degenerate
into horizontal or vertical lines, respectively. The Jacobian reads

J(q, p) = q−2 (147)

whence the transformation is a diffeomorphism in the cut plane without the axis (0, y).
 Equations (135)-(136) become
  
ϕ µ
f (X, µ, ν) = δ X − − ν p , (148)
q f
 
1 i dX dµ dν
Z
ϕ X− µq −ν p
f (q, p) = f (X, µ) 2 e . (149)
R3 q (2π)2

9. QUANTUM TOMOGRAPHY
We now consider a possible generalization of the above classical techniques to the
quantum case. The most direct path is that of using the phase-space description of
quantum mechanics through Wigner quasi-distribution functions [25, 20, 10]. To this
purpose, let us consider the phase space T ∗ Rn ' R2n of a classical system with n
degrees of freedom. Let us assume that the (statistical) state of the system is given by a
probability density function f (q, p) in S (R2n ). The expectation value of an observable,
which is represented by a function σ (q, p) on the phase space, in the state f is given by
Z
hσ i f = σ (q, p) f (q, p) dq dp. (150)
R2n

The M 2 -transform (72) of the density f , and its inverse (76) read
Z
M2
f (X, µ, ν) = hδ (X − µ · q − ν · p)i f = f (q, p) δ (X − µ · q − ν · p) dq dp,
R2n
dXdµdν
Z
2
f (q, p) = f M (X, µ, ν) ei(X−µ·q−ν·p) . (151)
R2n+1 (2π)2n
2
with X ∈ R and (µ, ν) ∈ R2n M
∗ . Note that the tomogram f (µ,ν) is nothing but the prob-
ability density function of the observable X(q, p) = µ · q + ν · p. A very special set
of observables are the canonical coordinates: position q = (q1 , . . . , qn ) and momentum
p = (p1 , . . . , pn ). One obtains their density directly from the tomograms (151) along the
Cartesian axes. At µi = δi, j and ν = 0 one has X(q, p) = q j and thus
Z
M2
f (X, (δi, j ), 0) = f (q, p) δ (X − q j ) dq dp
Z R2n

= f ((q1 , . . . , q j−1 , X, q j+1 , . . . , qn ), p) dq1 . . . dq j−1 dq j+1 . . . dqn dp, (152)

R2n−1

that is the marginal density of the position component q j . Analogously, at µ = 0 and

νi = δi, j one has X(q, p) = p j and
Z
M2
f (X, 0, (δi, j )) = f (q, p) δ (X − p j ) dq dp (153)
R2n

is the marginal density of p j .

 In quantum mechanics (more precisely, in its Schrödinger representation) the Carte-
sian canonical coordinates q and p are promoted to operators, q̂ and p̂, that act in the
Hilbert space of square summable functions L2 (Rn ) endowed with the scalar product
Z
hψ, φ i = ψ(x) φ (x) dx, (154)
Rn

with ψ, φ ∈ L2 (Rn ) and the overline denoting complex conjugation. Precisely, q̂ is the
multiplication operator (q̂ψ)(x) = xψ(x), and the p̂ = −ih̄∇ is a differential operator,
both with maximal domains. Here, h̄ is the Planck constant. Pure quantum states of
the system are associated to normalized elements ψ ∈ L2 (Rn ), called wave functions,
kψk22 = hψ, ψi = 1. The expectation value of position (momentum) in the wave function
ψ is given by the sandwich hψ, q̂ψi (hψ, p̂ψi). It easily follows that x 7→ |ψ(x)|2 is
the marginal probability density of position (correctly normalized, since kψk2 = 1).
Moreover, by Fourier transform it is also not difficult to show that the marginal of
momentum is given by p 7→ |ψ̂(p/h̄)|2 /(2π h̄)n (prove it!).
Now we want to define the quantization of any classical observable σ (also called
symbol), that for simplicity we assume in S (R2n ). Observe that

dη dy
Z
σ (q, p) = σ̂ (η, y) ei(η·q+y·p) , (155)
R2n (2π)2n

so, by following Weyl, we can construct an operator out of σ by substituting q with the
position operator q̂ and p with the momentum operator p̂. Thus, we are led to consider
the unitary operators
 
(y, η) 7→ ei(η·q̂+y· p̂) , ei(η·q̂+y· p̂) ψ (x) = eiη·(x+h̄y/2) ψ(x + h̄y) (156)

(the Schrödinger representation on L2 (Rn ) of the Heisenberg group [7]), with (y, η) ∈
R2n and ψ ∈ L2 (Rn ). In conclusion, for any σ ∈ S (R2n ) we define the Weyl quantization
of the symbol σ , the operator
dηdy
Z
Op(σ ) = σ̂ (η, y) ei(η·q̂+y· p̂) , (157)
R2n (2π)2n

whose action on any ψ ∈ L2 (Rn ) is

Z
(Op(σ )ψ)(x) = Kσ (x, y) ψ(y) dy (158)
Rn

with a kernel  
x+y dη
Z
Kσ (x, y) = σ ,η eiη·(x−y)/h̄ . (159)
Rn 2 (2π h̄)n
Observe that the kernel Kσ defined in (159) is obtained from σ by a partial Fourier
transform followed by an invertible and measure-preserving change of variables. These
operations make sense when σ is an arbitrary temperated distribution and define Kσ as
a temperated distribution. Therefore, it is possible to define the quantization of a generic
temperated distribution. Later we will use the quantization of the Dirac δ distribution.
Now we want the inverse mapping of the Weyl quantization, namely we want to
obtain the symbol starting from the operator. From (159) it is easy to check that the
dequantization of the operator S whose kernel is K(x, y) is given by the symbol

σ (q, p) = Op−1 (S)(q, p) = (2π h̄)n W (K)(q, p), (160)

where
y y  −iy·p/h̄ dy
Z

W (F)(q, p) = F q+ ,q− e (161)
Rn 2 2 (2π h̄)n
is known as the Wigner transform of F ∈ L2 (R2n ). This is the fundamental connection
between Weyl quantization and Wigner transform. By a straightforward direct computa-
tion, that we leave as an exercise, one has the following remarkable relation:
dq dp
Z Z
Tr(Op (σ )Op(τ)) := Kσ (x, y) Kτ (y, x) dx dy = σ (q, p) τ(q, p) , (162)
R2n R2n (2π h̄)n

for any σ , τ ∈ S (R2n ). This property will be crucial for our tomographic framework.
The statistical states of a quantum system are described by positive (self-adjoint)
operators with unit trace on the Hilbert space L2 (Rn ), namely ρ ≥ 0, ρ ∗ = ρ and Trρ = 1,
that are called density operators. It is easy to check that a density operator ρ is an integral
operator on L2 (Rn ), and if
ρ = ∑ λk ψk hψk , ·i (163)
k

is the spectral resolution of ρ, with {ψk } an orthonormal eigenbasis and 1 ≥ λk ≥ λk+1 ≥

0 the corresponding eigenvalues, the trace condition is equivalent to ∑k λk = 1, and the
kernel of ρ is (with a little abuse of notation)

ρ(x, y) = ∑ λk ψk (x)ψk (y). (164)

k

The expectation value in the state ρ of an observable associated to a bounded operator S

is given by
hSiρ = Tr(S ρ) = ∑ λk hψk , Sψk i. (165)
k
Therefore, the density matrix ρ is given by a convex combination of wave functions ψk
with weights λk . Physically, it represents a classical statistical mixture of pure quantum
states extracted with probabilities λk . By making use of Eqs. (160) and (162), the
quantum expectation value (165) of S = Op(σ ) reads
Z
hOp(σ )iρ = Tr (Op(σ ) ρ) = σ (q, p)W (ρ)(q, p) dq dp = hσ iW (ρ) . (166)
R2n

This equation expresses the expectation value of a quantum observable in the quantum
state ρ as a classical expectation value of its symbol in the “classical” state W (ρ). In
this respect, W (ρ) is called Wigner quasi-distribution function of ρ and is a substitute
for the nonexistent joint probability density of momentum and position in the quantum
state ρ. Indeed, since the uncertainty principle imposes a limit on the precision with
which momentum and position can be determined at the same time in the state ρ, it does
not make sense to speak of a joint probability distribution for these observables. In fact,
W (ρ) is not a genuine probability density, because it may assume negative values (e.g.
if ρ = ψhψ, ·i, with ψ ∈ L2 (Rn ) odd, we have that W (ρ)(0, 0) = −kψk22 /(π h̄)n < 0).
It is, however, always real and in some sense it tries very hard to be a joint density for
momentum and position. The above discussion provides supporting evidence for this
heuristic assertion. Moreover, it is easy to check that W (ρ) has the right marginals:

1 1
Z
W (ρ)(q, p) dq = ρ̂(p/h̄, p/h̄) = λk |ψ̂k (p/h̄)|2 ,
Rn (2π h̄)n (2π h̄)n ∑
k
Z
W (ρ)(q, p) dp = ρ(q, q) = ∑ λk |ψk (q)|2 . (167)
Rn k

Now we can apply all the results obtained in Secs. 3 and 5 to reconstruct the state of
a quantum system using a tomographic approach. In the quantum framework the object
that we want to reconstruct is the state ρ of the system. Therefore, let us compute the
expectation of
δ (X − µ · q̂ − ν · p̂) := Op (δ (X − µ · q − ν · p)) (168)
in the state ρ. By (166) and (151) one has
2
hδ (X − µ · q̂ − ν · p̂)iρ = hδ (X − µ · q − ν · p)iW (ρ) = W (ρ)M (X, µ, ν). (169)

Therefore, by measuring the expectation in the state ρ of the family of operator-valued

distributions (168), with X ∈ R, (µ, ν) ∈ R2n ∗ , we in fact are gathering all the tomograms
of the Wigner function W (ρ). Therefore, by applying the inversion formula (151) we can
recover W (ρ), that is, up to a normalization factor, the symbol of the density operator ρ
[see Eq. (160)]. Thus, by quantizing it we can reconstruct the quantum state ρ. Explicitly,
by using Eqs. (160), (159) and (151), we get
 
x+y
Z
ρ(x, y) = W (ρ) , η eiη·(x−y)/h̄ dη
Rn 2
dX dµ dν
Z Z
iη·(x−y)/h̄ 2
= dη e 2n
W (ρ)M (X, µ, ν) ei(X−µ·(x+y)/2−ν·η)
Rn R2n+1 (2π)
dX dµ dν
Z
2
= W (ρ)M (X, µ, ν) ei(X−µ·(x+y)/2) δ (ν − (x − y)/h̄)
R 2n+1 (2π)n
dX dµ
Z
2
= W (ρ)M (X, µ, (x − y)/h̄) ei(X−µ·(x+y)/2) . (170)
R n+1 (2π)n
2
We recall that the tomogram W (ρ)M (µ,ν) is the marginal probability density of the quan-
tum observable X(µ,ν) (q̂, p̂) = µ · q̂ + ν · p̂. When ξ = (µ, ν) ∈ S2n−1 the observable
Xξ is nothing but the component of a generalized position, called quadrature, in a ro-
tated frame of the phase space. From the experimental point of view, quadratures of
photons can be measured in a typical quantum optics setup by a homodyne detection
scheme. A different scheme was proposed for the longitudinal motion of neutrons [3].
Recently, in a spectacular experiment [13] quadratures were measured for reconstructing
the transversal motional states of helium atoms, by using a free evolution of the wave
packet followed by a position sensitive measurement.
The determination of the quantum state represents a highly nontrivial problem, whose
history can be traced back to the early days of quantum mechanics, namely to the
Pauli problem of reconstructing a wave function from its marginals in position and
momentum. The experimental validation had to wait until quantum optics opened a new
era. In this section we have followed the idea of Vogel and Risken [24]. Since then
many improvements and new techniques have been proposed [15, 11, 19]. An up-to-
date overview can be found in Ref. [21].

Notation

f ] , g[ Radon transform and its dual, see Eq. (12) and Eq. (13)
fξ] Tomogram of f along the direction ξ , see Eq. (23)
M2
f M 2 transform, see Eq. (72)
S (M) Schwartz space of rapidly decreasing functions on M, see Eq. (16)
δ (x) Dirac delta distribution at 0
fˆ, fˇFourier transform and its inverse, see Eqs. (37)-(38)
W (ρ) Wigner transform of the state ρ, see Eq. (161)
Op(σ ) Weyl quantization of the symbol σ , see Eq. (157)
Pd
n
Space of all the (n − d)-dimensional affine subspaces of Rn . (Pn = Pn1 )
Rn∗ = Rn \ {0}, R+ = [0, +∞), N∗ = N \ {0}
Z +∞
Γ(z) = λ z+1 e−λ dλ , Euler gamma function
0

ACKNOWLEDGMENTS
We would like to thank Beppe Marmo, Manolo Asorey, Volodya Man’ko, and Saverio
Pascazio for stimulating discussions. P. F. would like to thank the organizers, J. F.
Cariñena, E. Martínez, J. Clemente-Gallardo, J. N. da Costa, and D. Martín de Diego, for
their kindness in inviting him and for the effort they exerted on the organization of the
workshop. This work is partly supported by the European Union through the Integrated
Project EuroSQIP.
 REFERENCES
1. M. Asorey, P. Facchi, V. I. Man’ko, G. Marmo, S. Pascazio and E. C. G. Sudarshan, Phys. Rev. A 76,
012117 (2007).
2. M. Asorey, P. Facchi, V. I. Man’ko, G. Marmo, S. Pascazio and E. C. G. Sudarshan, Phys. Rev. A 77,
042115 (2008).
3. G. Badurek, P. Facchi, Y. Hasegawa, Z. Hradil, S. Pascazio, H. Rauch, J. Řeháček, and T. Yoneda,
Phys. Rev. A 73, 032110 (2006).
4. S. R. Deans, The Radon transform and some of its applications, Dover Publications, New York, 2007.
5. L. C. Evans, Partial Differential Equations, Graduate Studies in Mathematics, Vol. 19, American
Mathematical Society, Providence, Rhode Island, 2001
6. P. Facchi, M. Ligabò and S. Pascazio, “On the inversion of the Radon transform: standard versus M 2
approach", Journal of Modern Optics, First published on: 01 December 2009 (iFirst).
7. G. B. Folland, Harmonic Analysis in Phase Space (AM-122), Princeton University Press, Princeton,
1989.
8. I. M. Gel’fand and G. E. Shilov, Generalized Functions: Properties and Operations, Vol. 5, Academic
Press, 1966.
9. S. Helgason, The Radon Transform, Birkhauser, Boston, 1980.
10. M. Hillary, R. F. O’Connell, M. O. Scully, and E. Wigner, Phys. Rep. 106, 121 (1984).
11. Z. Hradil, Phys. Rev. A 55, R1561 (1997).
12. F. John, Plane waves and spherical means: Applied to Partial Differential Equations, Wiley Inter-
science, New York, 1955.
13. C. Kurtsiefer, T. Pfau, and J. Mlynek, Nature 386, 150 (1997).
14. H. Lieb and M. Loss, Analysis, Graduate Studies in Mathematics, Vol. 14, American Mathematical
Society, Providence, Rhode Island, 2001.
15. S. Mancini, V. I. Man’ko and P. Tombesi, Quantum Semiclass. Opt. 7, 615 ( 1995).
16. O. V. Man’ko, V. I. Man’ko, G. Marmo, Phys. Scr. 62, 446 (2000).
17. O. V. Man’ko, V. I. Man’ko, G. Marmo, J. Phys. A: Math. Gen. 35, 699 (2002)
18. V. I. Man’ko, G. Marmo, A. Simoni, A. Stern, E. C. G. Sudarshan, Phys. Lett. A 35, 351 (2005).
19. V. I. Man’ko, G. Marmo, A. Simoni, F. Ventriglia, Open Sys. & Information Dyn. 13, 239 (2006).
20. J Moyal, Proc. Camb. Phil. Soc. 45, 99 (1949).
21. M. G. A. Paris and J. Řeháček, Quantum State Estimation, Lecture Notes in Physics Vol. 649,
Springer, Berlin, 2004.
22. J. Radon, “Über die bestimmung von funktionen durch ihre integralwerte längs dewisse
mannigfaltigkeiten", Breichte Sachsische Akademie der Wissenschaften, Leipzig, Mathematische-
Physikalische Klasse, 69 S. 262 (1917).
23. R. S. Strichartz, American Mathematical Monthly 89, 377 (1982).
24. K. Vogel and H. Risken, Phys. Rev. A 40, 2847 (1989).
25. E. P. Wigner, Phys. Rev. 40, 749 (1932).