Sie sind auf Seite 1von 11

International Journal of Emerging Trends in Signal Processing

ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013



http://ijetsp.info/article/IJETSPV1I301.pdf


Abstract: This paper deals with geometry of covariance
matrices to define new advanced Radar Doppler Processing
based on Metric Space tools. Information Geometry has been
introduced by C.R.Rao, and axiomatized by N. Chentsov, to
define a distance between statistical distributions that is
invariant to non-singular parameterization transformations. For
Doppler/Array/STAP Radar Processing, Information Geometry
Approach will give key role to Homogenous Symmetric bounded
domains geometry. For Radar, we will observe that Information
Geometry metric could be related to Khler metric, given by
Hessian of Khler potential (Entropy of Radar Signal given by
log[det(R)]). To take into account Toeplitz structure of
Time/Space Covariance Matrix or Toeplitz-Block-Toeplitz
structure of Space-Time Covariance matrix, Parameterization
known as Partial Iwasawa Decomposition could be applied
through Complex Autoregressive Model or Multi-channel
Autoregressive Model. Then, Hyperbolic Geometry of Poincar
Unit Disk or Symplectic Geometry of Siegel Unit Disk will be
used as natural space to compute p-mean (p=2 for mean,
p=1 for median) of covariance matrices via Karcher Flow
derived from Weiszfeld algorithm extension on Cartan-
Hadamard Manifold. This new mathematical framework will
allow development of OS (Ordered Sta-tistic) concept for
Hermitian Positive Definite Covariance Space/Time Toeplitz
matrices or for Space-Time Toeplitz-Block-Toeplitz matrices.
We will define OS-HDR-CFAR (Ordered Statistic High Doppler
Resolution CFAR) and OS-STAP (Ordered Statistic Space-Time
Adaptive Processing).


Index Terms: Median, p-mean, Center of mass, Cartan-
Hadamard Manifold, Cartan Symmetric spaces, Homogenous
bounded domains, Siegel Upper-half plane, Poincar disk, Siegel
disk, Frchet metric space, Berger fibration, Mostow
decomposition, Radar, CFAR, STAP

Manuscript received September 15, 2012. This work was supported in part
by US Air Force Office of Scientific Research (AFOSR) through European
Office of Aerospace Research and Development (EOARD) Grant n FA8655-
11-1-3031, and by THALES Key Technology Domain Processing Control &
Cognition (KTD PCC) funding.
Frdric BARBARESCO is Senior Scientist and Advanced Studies Manager
in THALES AIR SYSTEMS, Surface Radar Domain, Technical Directorate,
Advanced Developments Department, Voie Pierre-Gilles de Gennes, F-91470
Limours FRANCE (corresponding author to provide phone: +33 (0)630071419
; fax: +33 (0)164916766; e-mail: frederic.barbaresco@ thalesgroup.com).
I. INTRODUCTION
NFORMATION Geometry has been introduced by
C.R.Rao, and axiomatized by N. Chentsov, to define a
distance between statistical distributions that is invariant to
non-singular parameterization transformations. For
Doppler/Array/STAP Radar Processing, Information
Geometry Approach will give key role to Symmetric spaces
and Homogenous bounded domains geometry. For Radar, we
will observe that Information Geometry metric could be
identified to Khler metric, given by Hessian of Khler
potential (Entropy of Radar Signal given by log[det(R)]). To
take into account Toeplitz structure of Time/Space
Covariance Matrix or Toeplitz-Block-Toeplitz structure of
Space-Time Covariance matrix, Parameterization known as
Partial Iwasawa Decomposition could be applied through
Complex Autoregressive Model or Multi-channel
Autoregressive Model. Then, Hyperbolic Geometry of
Poincar Unit Disk or Symplectic Geometry of Siegel Unit
Disk will be used as natural space to compute p-mean (p=2
for mean, p=1 for median) of covariance matrices via
Karcher Flow for Weiszfeld algorithm extension on Manifold.
This new mathematical framework will allow development
of (Ordered Statistic) concept [3] for Hermitian Positive
Definite Covariance Space/Time Toeplitz matrices or for
Space-Time Toeplitz-Block-Toeplitz matrices. We will define
OS-HDR-CFAR (Ordered Statistic High Doppler Resolution
CFAR) and OS-STAP (Ordered Statistic Space-Time
Adaptive Processing) algorithms for radar detection.
II. INFORMATION GEOMETRY MANIFOLD OF COVARIANCE
HERMITIAN POSITIVE DEFINITE MATRICES
In 1945, Rao has introduced Information Geometry for
parameterized density of probability
( ) / . p
with the metric
given by the formula
| | ( ) d I d d p p K ds
+
= + = ) / (. ), / (.
2

where
| | ) ( ) (
ij
g I =
is the Fisher Information matrix. If we
model Signal by complex circular multivariate Gaussian
Information Geometry Manifold of Toeplitz
Hermitian Positive Definite Covariance
Matrices: Mostow/Berger Fibration and Berezin
Quantization of Cartan-Siegel Domains
Barbaresco F. , Emeritus Member SEE, SEE Ampere Medal 2007, member IEEE
I
1
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013

http://ijetsp.info/article/IJETSPV1I301.pdf

distribution of zero mean :
( ) ( )
n n n n n
m X R m X
n
n
n n
e R R X p

+
=
.
1
1
. . ) ( ) / (
(1)
then for
0 =
n
m
:
( ) ( ) | |
2
2 / 1 2 / 1
2
1 2
F
n n n n n
R dR R dR R Tr d I d ds
+
= = =

(2)
This metric has been integrated by Siegel and the distance is :
( ) ( ) ( )

=

= =
n
k
k
F
R R R R R dist
1
2
2
2 / 1
1 2
2 / 1
1 2 1
2
log . . log ,

(3) with
( ) 0 det
1 2
= R R
k


This is a complete simply connected metric space of
negative curvature with geodesic between
1
R
and
2
R
:
( )
( )
2 / 1
1
2 / 1
1 2
2 / 1
1
2 / 1
1
2 / 1
1
log . 2 / 1
1
2 / 1
1 2
2 / 1
1
) ( R R R R R R e R t
t
R R R t
= =


(4)
with 1 0 t
III. FROM CARTAN CENTER OF MASS TO FRCHET MEDIAN
Robust Covariance Matrix Mean Estimation can be
addressed by Riemannian center of mass. Elie Cartan has
proved back in the 1920s existence and uniqueness of center
of mass for simply connected complete manifolds of negative
curvature for any compact subset. In Euclidean space, the
center of mass is defined for finite set of points
{ }
M i i
x
,..., 1 =
by
arithmetic mean:

=
=
M
i
i center
x
M
x
1
1
or by
0
1
=

=
M
i
i center
x x
. This
point also minimizes:

=
=
M
i
i
x
center
x x d Min x
1
2
) , ( arg
. This has been
extended to general Riemannian manifolds by Elie Cartan
that has proved that the function


A
da a m d m f ) , (
2
1
:
2

is
strictly convex (its restriction to any geodesic is strictly
convex as a function of one variable) and achieves a unique
minimum at a point called the center of mass of A for the
distribution da. Moreover, this point is characterized by being
the unique zero of the gradient vector field:
) ( exp
1


=
A
m
da a f
, where
exp(.)
is the exponential map and
) ( exp
1
a
m

is the tangent vector at


m
of the geodesic from
m to a :

m m
T a) ( exp
1
. Hermann Karcher has then
introduced a gradient flow intrinsic on the Manifold that
converges to the center of mass, called Karcher Barycenter:
( ) ) ( ) 0 ( with ) ( . exp ) (
1 n n n n m n n n
m f m f t t m
n
= = =
+

(5)
In the discrete case, the center of mass for set of points is:
) ( exp . exp
1
1
1
|
.
|

\
|
=

=

+
M
i
i m n m n
x t m
n n

(6)

Figure 1. Frchet-Karcher Flow on Cartan-Hadamard Manifold
But, center of mass is not useful for robust statistic.
Replacing L
2
square geodesic distance by L
1
geodesic
distance, we can extend this approach to estimate a Median
| | m x E Min m
m
median
=
in
metric space (called Fermat-Webers point in Physic). Frchet
studied Median statistic using compared
to
| |
2
m x E Min m
m
mean
=
. Classically, in Euclidean space,
Median point minimizes

=
=
M
i
i
x
median
x x d Min x
1
) , ( arg
,
equivalently
0 /
1
=

=
M
i
i median i median
x x x x
, with its Riemannian
extension

=
A m
m
Min
A
da
a
a
h da a m d m h
) ( exp
) ( exp
) , (
2
1
:
1
1


(7)
We cannot directly extend Karcher Flow to median
computation in the discrete case:
( )

) ( exp
exp
. exp
1
1
1
1
|
|
.
|

\
|
=

=

+
M
k
k m
k m
n m n
x
x
t m
n
n
n

(8)
because
) ( exp
1
k m
x
n

could vanish if
x m
k n
=
. The
uniqueness of the geometric median of a probability measure
on a complete Riemannian manifold [4] has been
investigated. By regarding the Weiszfeld algorithm as a sub-
gradient procedure, a sub-gradient algorithm has been
introduced [1] to estimate the median and to prove that this
algorithm always converges :
( )
|
|
.
|

\
|
=

+
n m
n
n
n
G k
k m
k m
n m n
x
x
t m
) ( exp
exp
. exp
1
1
1
with
{ }
n k m
m x k G
n
= /
(9)
Then, the median A of the N matrices B
k
can be computed
by sub-gradient Karcher flow :
( )
( )
2 / 1
log
log
2 / 1
1
2 / 1 2 / 1
2 / 1 2 / 1
n
A B A
A B A
t
n n
A e A A
n A G k
F
n k n
n k n
n
|
|
|
.
|

\
|
+

=




(10)
with
{ }
n k A
A B k G
n
= /

2
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013


http://ijetsp.info/article/IJETSPV1I301.pdf


Figure 2. Frchet-Karcher Flow on space HPD Matrices
IV. FOURIER HEAT FLOW EQUATION ON 1D GRAPH OF
HERMITIAN POSITIVE DEFINITE MATRICES
We can replace Median computation by anisotropic
diffusion. In normed vector space in 1D, if we note
( ) 2 /
1 1 +
+ =
n n n
u u u
, Fourier diffusion Equation is given by :

2
2
x
u
t
u

| |
t n t n t n t n t n t n t n n,t
u u u u u u
x
t
u u
, , , , , , 2 , 1
. ). 1 (
2

= + =

+ =
+
(11)
By analogy, we can define diffusion equation on a 1D
graph of HPD(n) by:
( )
( )
( )
2 / 1
, 1
/2 1
2 / 1
, 1 , 1
2 / 1
, 1
2 / 1
, 1 , 2
, ,
2 / 1
,
2 / 1
, ,
2 / 1
,
2 / 1
,
2 / 1
,

log
2
2 / 1
, 1 ,

and
x
t 2
with

2 / 1
, ,
2 / 1
,
2
t n t n t n t n t n t n
t n t n t n t n t n t n t n t n
A A A
x
t
t n t n
A A A A A A
A A A A A A A A e A A
t n t n t n
+

+ +

+
=

=
= = =


(12)
Obiously, we can introduce anisotropy by making adaptive
the parameter

.
From each Time covariance matrix, we can compute
Doppler Spectrum. In the following exemple, we give image
with range on X axis and Doppler frequency on Y axis.
Fourier heat Diffusion is applied on covariance matrices and
then, we draw associated Doppler spectrum of results :

Figure 3. Fourier Heat Equation on a 1D graph of covariance matrices:
isotropic diffusion

Figure 4. Fourier Heat Equation on a 1D graph of covariance matrices:
anisotropic diffusion
V. TOEPLITZ-CONSTRAINED FLOW FOR HERMITIAN POSITIVE
DEFINITE COVARIANCE MATRICES OF STATIONARY SIGNAL
Previous flow has a major drawback, it will not constrained
Toeplitz structure of covariance matrices M for stationary
signal
| |
*
, k n n k k n n
z z E r r

= =
but only the following
structure: M JMJ = with J an anti-diagonal matrix.
{ } M JMJ C HPD M E
n
= = / ) (
is a closed sub-
manifold of
) (C HPD
n
.To take into account this constraint,
Partial Iwasawa decomposition should be considered. This is
equivalent for time or space signal to Complex
AutoRegressive (CAR) Model decomposition (see Trench
Theorem [5]) :
( ) ( )
(

+
= = =
+

+
+
1 1 1 1
1
2 1
.
1
. 1 . .
n n n n
n
n n n n n n
A A A
A
W W R

(13)
(

=

2 / 1
1 1
2
n
0 1
1
n n
n
A
W
with
+

=
2 / 1
1
2 / 1
1 1
.
n n n

(14)
where
| |
1
1
2
1
. 1

=
n n n

,
(

+
(

=


1 0
1 1
) (
n
n
n
n
A

A
A
and
* ) (
.V J V =


In the framework of Information Geometry, Information
metric could be introduced as Khlerian metric where Khler
potential is given by the process Entropy
,
~
0
) P (R
n

:
( ) ( ) | | | | . . log . 1 log ). ( . log det log ,
~
0
1
1
2
1
0
P e n k n e R ) P (R
n
k
k n n
= =

Information metric is then given by hessian of Entropy


)* ( ) (
2
~
n
j
n
i
ij
g

where
| |
T
n
n
P
1 1 0
) (

=

(15)
with
{ }
1
1

=
n
k k

Regularized [2] Burg reflection coefficient


[6,7] and
1
0 0

= P
mean signal Power. Khlerian metric is
finally :
| |
( )

=
+

+
|
|
.
|

\
|
= =
1
1
2
2
2 2
0
0 ) ( ) ( 2
1
) ( .
n
i
i
i n
ij
n
n
d
i n
P
dP
n d g d ds



(16)
For Median autoregressive model computation, Karcher
flow could be trivial. For
0
P
, we use classical median on real
value. For
{ }
1
1

=
n
k k

, we use homeomorphism of Poincars unit


3
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013


http://ijetsp.info/article/IJETSPV1I301.pdf

disk
*
n k,n
n k,n
k,n
.w
w


=
+
1
1
, to fix the point under action of karcher
flow at the origin from where all geodesics are radials and
where the space is locally Euclidean. Dual Karcher Flow is:
{ } 0 with
0
,
,
0
= =

k,n
G k n k
n k
n n
k/ G w


(17)
Median is deduced from each
n
w
at each iteration:
*
n median,n
n median,n
median,n
w
w

+
+
=
+
1
1

(18)

Figure 5. Modified Karcher Flow in Poincar disk by homeomorphism

Figure 6. Classical Karcher flow in Poincars unit disk

Figure 7. Dual Karcher Flow centered at the origin
VI. BLOCK-TOEPLITZ CONSTRAINED KARCHER FLOW
We will extend previous approach to space-time covariance
matrix for STAP. Based on generalization of Trench
Algorithm [5], if we consider Toeplitz-block-Toeplitz
Hermitian Positive Definite matrix:
(

=
(
(
(
(

=
+
+ +
+
+
0
,
0 1
1
0 1
1 0
1 , ~
~
R R
R R
R R R
R
R R
R R R
R
n
n n p
n
n
n p


(19)
with
*
1
~
(
(
(

=
n
n
R
R
V R
and
(
(
(
(
(

=
0 0
0
0
0 0

p
p
p
J
J
J
V

(20)
From Burg-like parameterization [11], we can deduced this
inversion of Toeplitz-Block-Toeplitz matrix :
(
(

+
=
+
+

+
n n n n p n n
n n n
n p
A A R A
A
R

. . .
.
1
,
1
1 ,



(21)
and
(
(

+
=
+ +
+
n p n n p
n p n n n p n n
n p
R A R
R A A R A
R
, ,
, ,
1
1 ,
.
. . .


(22)
| |
(
(
(
(
(

+
(
(

=
(
(
(

=
= =

+
p
p
n
p
p
n
n p
n
n
p
n
n
n
n
-
n
n
n
n
n n
I
J A J
J A J
A
A
A
A
A
R A A
* 1
1
* 1
1
1
1
1
0
1
0
1
1
1
.
0
and
, . 1 with



(23)
Where we have the following Burg-like generalized
forward and backward linear prediction :
1
1 1 1
1
1
1
0
0 0
* 1
1
0
* 1
1
1
1
1
0
1
1
) ( ) ( ) ( ) ( ) 1 ( ) ( 2
) ( ) ( ) (
with
) ( ) 1 ( ) ( ) ( ) (
) 1 ( ) ( ) ( ) ( ) (

+
=
+
+
=
+
+
=
+ +
+
+
+
+
=
+
+
+
+
+
=
+
+
(

+
(

=
= =

+ = + =
+ = =

n N
k
b
n
b
n
n N
k
f
n
f
n
n N
k
b
n
f
n
n
n
p
n
b f
f
n
n
n
b
n
n
l
n
l
b
n
b
n
n
n
f
n
n
l
n
l
f
n
k k k k k k A
I A
k Z k k
k J JA k l n k JZ k JA k
k A k l k Z k A k





(24)
Using Schwarzs inequality, it is easily to prove that
1
1
+
+
n
n
A

Burg-Like reflection coefficient matrix lies in Siegel Disk
p
n
n
SD A
+
+
1
1
.
VII. CARTAN-SIEGEL HOMOGENEOUS DOMAINS : SIEGEL DISK
To solve median computation of Toeplitz-Block-Toeplitz
matrices, Karcher-Frchet Flow has to be extended in Siegel
Disk. Siegel Disk has been introduced by Carl Ludwig Siegel
[8] through Symplectic Group
R Sp
n 2
that is one possible
generalization of the group
R Sp R SL
2 2
=
(group of invertible
matrices with determinant 1) to higher dimensions. This
generalization goes further; since they act on a symmetric
homogeneous space, the Siegel upper half plane, and this
action has quite a few similarities with the action of
R SL
2
on
the Poincars hyperbolic plane. Let F be either the real or
the complex field, the Symplectic Group is the group of all
matrices
F GL M
n 2

satisfying :
4
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013

http://ijetsp.info/article/IJETSPV1I301.pdf

{ }
) , 2 (
0
0
/ ) , 2 ( ) , (
R n SL
I
I
J
J JM M F n GL M F n Sp
n
n
T

|
|
.
|

\
|

=
=
(25)
n
T T
T T
I B C D A
D B C A F n Sp
D C
B A
M
=

|
|
.
|

\
|
=
and
symmetric and ) , ( or
(26)
The Siegel upper half plane is the set of all complex
symmetric nxn matrices with positive definite imaginary part:
{ } 0 Im / ) , ( > = + = = Y (Z) C n Sym iY X Z SH
n

(27)
The action of the Symplectic Group on the Siegel upper
half plane is transitive. The group
{ }
n
I R n Sp R n PSp
2
/ ) , ( ) , (

is group of
n
SH
biholomorphisms via generalized Mbius
transformations:
( )( )
1
) (

+ + =
|
|
.
|

\
|
= D CZ B AZ Z M
D C
B A
M

(28)

) , ( R n PSp
acts as a sub-group of isometries. Siegel has
proved that Symplectic transformations are isometries for the
Siegel metric in
n
SH
. It can be defined on
n
SH
using the
distance element at the point iY X Z + = , as defined by:
( ) ( ) ( ) iY X Z dZ Y dZ Y Tr ds
Siegel
+ = =
+
with
1 1 2
(29)
with associated volume form :
( )
+
= dZ Y dZ Y Tr
1 1


C.L. Siegel has proved that distance in Siegel Upper-Half
Plane is given by :
( )
n
n
k
k
k
Siegel
SH Z Z
r
r
Z Z d
|
|
.
|

\
|
|
|
.
|

\
|

+
=

=
2 1
1
2
2 1
2
, with
1
1
log ,
(30)
and r
k
eigenvalues of the cross-ratio :
( ) ( )( ) ( )( )
1
2 1 2 1
1
2 1 2 1 2 1

+ + +

+
= Z Z Z Z Z Z Z Z ,Z Z R
. (31)
This is deduced from the 2
nd
derivative of
( ) Z Z R Z ,
1

in
Z Z =
1
given by :
( ) ( )
1 1 1
1
2
. 2 / 1 ) ( 2
+ + +

+
= = Y dZ dZY Z Z dZ Z Z dZ R D
(32)
and
( ) ( ) R D Tr dZ dZY Y Tr ds
2 1 1 2
. 2 = =
+
(33)
In parallel, in China in 1945, Hua Lookeng has given the
equations of geodesic in Siegel upper-half plane :
0
1
2
2
= +

ds
dZ
Y
ds
dZ
i
ds
Z d
(34)
Using generalized Cayley transform
( )( )
1
+ =
n n
iI Z iI Z W
,
Siegel Upper-half Plane
n
SH
is transformed in unit Siegel
disk
{ }
n n
I WW W SD < =
+
/
where the metric in Siegel Disk is
given by :
( ) ( ) | |
+

+
= dW W W I dW WW I Tr ds
n n
1 1
2
(35)

Figure 8. Geometry of Siegel Upper Half-Plane
Contour of Siegel Disk is called its Shilov boundary
{ }
n n n
I WW W SD 0 / = =
+
. We can also defined horosphere.
Let
n
SD U
and
+

*
R k
, the following set is called
horosphere in siegel disk :
( ) ( ) ( )( ) { }
1 1
1
/ 0 / ,
+
<
)
`

+
= < =
+ + +
k
k
U
k
Z Z Z U I U Z I Z Z I k Z U k H
(36)
Hua Lookeng and Siegel have proved [8] that the previous
positive definite quadratic differential is invariant under the
group of automorphisms of the Siegel Disk. Considering
(

=
D C
B A
M
such that
(

=
(

n
n
n
n
I
I
M
I
I
M
0
0
0
0
* :
( )( )
( ) ( )
( )( ) ( ) ( )
2 2
1 1 1
1 1
1
) (
W V
n n
n n
ds ds
A BW dW W W I dW WW I A BW
dV V V I dV VV I
D CZ B AZ W M V
=
+ +
=
+ + = =

+ +

+ +
+

(36)
Complementary, Hua Lookeng has also proved that, let
V,W be complex-valued matrices, if
0 >
+
VV I
and
0 >
+
WW I , then the following identity holds :
( ) ( ) ( )
2
det det det
+ + +
VW I WW I VV I
(37)
( ) ( ) ( )
2
det det det B A I B B I A A I
+ + +

is based on Huas
matrix identity :
( ) ( ) ( ) ( ) ( )( ) ( ) B A I A A I A B I B A AA I B A B B I
+

+ + +
= +
1
* *
1
(38)
using the intermediate equalities:
( )( ) ( ) ( ) ( ) ( )( ) A B AA I A B B B I B A I A A I A B I =
+ + + +

+ +
1
(39)
( ) ( ) A AA I A I A A I
1 1
+ +

+
+ =

and
( ) ( )
1 1
+ + +

+
= AA I A A A A I
(40)
Same kind of inequality is true for the trace :
( ) ( ) ( )
2
B A I Tr B B I Tr A A I Tr
+ + +

(41)
5
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013


http://ijetsp.info/article/IJETSPV1I301.pdf

To go further to study Siegel Disk, we need now to define
what are the automorphisms of Siegel Disk
n
SD
. They are all
defined by:
( )
t
Z n
U Z U Z C n U U SD Aut ) ( ) ( / , ), (
0
=
(42)
with
( ) ( )( ) ( )
2 / 1
0 0
1
0 0
2 / 1
0 0
) (
0
Z Z I Z Z I Z Z Z Z I Z
Z
+

+
= =
(43)
and its inverse :
( ) ( ) ( )( )
( )
( ) ( )

=
+ + = =

= =

+ +
+

+ +
2 / 1
0 0
2 / 1
0 0
0
1
0
1
1
0 0
2 / 1
0 0
2 / 1
0 0
with
) ( ) (
0
Z Z I Z Z I G
Z G I GZ Z
Z Z I Z Z Z Z I Z Z I G
Z



(44)
By analogy with Poincars unit Disk, C.L. Siegel has
deduced geodesic distance in
n
SD
:
|
|
.
|

\
|
+
+
=
) ( 1
) ( 1
log
2
1
) , ( , ,
W
W
W Z d SD W Z
Z
Z
n

(45)
VIII. MOSTOW/BERGERS FIBRATION OF SIEGEL DISK
Information metric will be introduced as a Khler potential
defined by Hessian of multi-channel entropy
( )
1 ,
~
+ n p
R
:
( ) ( ) ( ) ( ) | |
n p j i n p n p n p
R Hess g cste R Tr cte R R
, ,
1
, ,
log det log
~
= + = + =

Using partitioned matrix structure of Toeplitz-Block-Toeplitz
matrix
1 , + n p
R
, recursively parametrized by Burg-Like
reflection coefficients matrix
{ }
n
k
k
n
k
k
k
SD A A

=
with
1
1
, we can
give Multi-variate entropy [3], matrix extension of previous
Entropy :
| | | |
0
1
1
,
det . . log . 1 det log ). (
~
R e n A A k n ) (R
n
k
k
k
k
k n p
=

=
+ (46)
Paul Malliavin [3] has proved that this form is a Khler
Potential of an invariant Khler metric (Information
Geometry metric in our case) that is given by matrix
extenstion of (14):
( ) | | ( ) ( ) | |

=
+

+
+ =
1
1
1 1 2
0
1
0
2
) (
n
k
k
k
k
k
k
k n
k
k
k
k
k
k n
dA A A I dA A A I Tr k n dR R nTr ds

(47)
As we have defined a metric space, we can extend
Karcher/Frechet flow in Unit Siegel Disk to compute the
Median of N Toeplitz-Block-Toeplitz Hermitian Positive
Definite matrices. These matrices are parameterized by Burg-
Like generalized Reflection matrices
{ }
n
k
k
n
k
k
k
SD A A

=
with
1
1

and Karcher/Frechet Flow in Siegel Disk will be solved by
analogy of our scheme used in Poincar unit Disk, by mean of
Mostow Decomposition Theorem : every matrix M of
) , ( C n GL
can be decomposed in
S iA
e Ue M = where S is
symmetric
( ) ( ) M M P P P P P P S
+
= = with log
2
1
2 / 1
2 / 1
2 / 1 * 2 / 1 2 / 1 , A
is antisymmetric
( )
S S
Pe e
i
A

= log
2
1
and finally,
iA S
e Me U

= is
unitary. Median in Siegel disk could be then obtained by
analogy with scheme developed for median in Poincars
disk:
{ } { }
( ) ( )
{ }
( )
( ) ( )( ) ( )
( ) ( )
( ) ( )
2 / 1
,
2 / 1
1
1
2 / 1 1 2 / 1
1
1
,
1
,
, , ,
2 / 1
,
2 / 1
2 / 1
,
*
,
2 / 1
,
2 / 1
, ,
2
,
2
, , , ,
1 0 0 1 0
with

then 1 For
with
with log . 2 / 1 with
until on Iterate
0
, ,
, , , ,

+ +

+
+
+

+
+
+

=
+

=
+ + =
=
= =
< =
= =
= = = =
<
= =

n n n median n n
n n median,n
n n k,n n n k,n n n k,n
k,n G k,n
F
n l
m
l k
k
n k n n
n k n k n k n k n k n k n k n k n k
S
n k
S
S
n k
iA
n k k,n
S iA
n k n k
F
n
m m, , median,
G G I W G G I G
G G I GG W
G G I W G I G W G G I W
W W ,...,m k
H l/ H G
W W P P P P P P S
e W e e W e U H e e U W
G n
,...,W W ,...,W W and W tion : Initialisa
n
n k n k
n k n k n k n k


IX. BEREZIN QUANTIZATION OF CARTAN-SIEGEL DOMAINS
Symmetric Bounded Domains of C
n
are key spaces for all
these approaches and are particular symmetric spaces of non-
compact type. Elie Cartan [4] has proved that there are only 6
types:
2 exceptionnal types (E6 et E7)
4 Classical Symmetric bounded Domains (extension of
Poincar Unit disk):
0 2 ZZ 1 , 1 ZZ
: line 1 and rows n with matrices complex : IV Type
p order of matrices symmetric skew complex : III Type
p order of matrices symmetric complex : II Type
rows q and lines p with matrices complex : I Type
) : (
Matrix r Rectangula Complex :
2
t t
> + <
<
+
+ +
ZZ

conjugate transposed I ZZ
Z
IV
n
III
p
II
p
I
p,q
(48)
kernel function for all these domains were established by
Lookeng Hua:
( )
( )
( )
( )
( )
( )
( ) domain. the of volume euclidean is where
, : IV Type for 2 ZZ 1
1
,
1 , : III Type
1 , : II Type
, : I Type
for det
1
,
-
* * t *
*

n ZW W W W Z K
p
p
q p
ZW I W Z K
IV
n
III
p
II
p
I
p,q
= + =

=
+ =
+ =
=
+

+

(49)
For the case (p=q=n=1), all these domains are reduced to
the classical Poincar unit disk :
{ } ( )
( )
2
*
* *
1 1 1 1 1
1
1
, , 1 /
zw
w z K zz C z
IV III II I
,

= < = = = =
(50)
Groups of analytic automorphisms of these domains are
locally isomorphic to the group of matrices which preserve
following forms:
6
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013

http://ijetsp.info/article/IJETSPV1I301.pdf

|
|
.
|

\
|
= = =
|
|
.
|

\
|
=
|
|
.
|

\
|

= = =
|
|
.
|

\
|

=
|
|
.
|

\
|

= = =
=
|
|
.
|

\
|

= =
n
t IV
n
p
p
p
p t III
p
p
p
p
p t II
p
p
p I
p,q
I
I
H H AHA H AHA
I
I
L
I
I
H L ALA H AHA
I
I
K
I
I
H K AKA H AHA
A
I
I
H H AHA
0
0
, , , : IV Type
0
0
,
0
0
, , , : III Type
0
0
,
0
0
, , , : II Type
1 det ,
0
0
, , : I Type
2 *
*
*
*
(51)
All classical domains are circular and considered in the
general framework of Elie Cartan Theory, where the origin is
a distinguished point for the potential:
( )
( )
( )
( )


+
=
(

= ZZ I
K
Z Z K
Z Z det log
0 , 0
,
log ,
*
* (52)
F.A. Berezin [3] has introduced on these Cartan-Siegel
domains the concept of quantization based on construction of
Hilbert spaces of analytical functions:
Z
gZ
Z g j Z Z K Z g j z g j gZ gZ K
Z Z d
K
Z Z K
h c
Z Z d
K
Z Z K
Z g Z f h c g f
h
h

= =
(

=
(

) , ( with ) , ( ) , ( ) , ( ) , (
) , (
) 0 , 0 (
) , (
) (
) , (
) 0 , 0 (
) , (
) ( ) ( ) ( ,
* * *
*
/ 1
*
1
*
/ 1
*

(53)
One example is given in dimension 1 for Poincar unit disk
{ }
1
/ ) 1 , 1 ( 1 / S SU z C z D = < =
with volume element
* 2
2
) 1 .( 2 / 1 dz dz z i

:
( )
( )
*
2
*
2
2
* *
2
2 2
* *
) ( ) (
) ( log Re 2 ) (
1 log ) ( : potential Khler with
1 where with ) 1 , 1 (
z z
z F
z z
gz F
ds
z F a z b gz F
z z F
b a
a b
b a
g SU g

=
+ + =
=
=
(

=
(54)
It results from the last equation that the Khlerian metric is
invariant under the action
G g
(automorphisms of unit
disk). The transform of the base point 0 = z of the disk by
G g
is given by
( )
1
*
) 0 (

= a b g
. It defines a lifting that allows
to associate to all paths in disk a lift in G. In the same way,
we can define a geometric lift of potential K in G:
( )
( ) ( )
) ( ) (
1 log 1 log )) 0 ( (
) 0 (
1
*
*
1
2
1
2
1
*
1
*
2 2
g F g F
a b
b a
g
b a b g F
a b g
b a
=
|
|
.
|

\
|


=
+ = |
.
|

\
|
=
=

=


(55)
Obviously, all these lifts could be extended to Cartan-
Siegel Domains
{ } I ZZ Z SD
n
< =
+
/
:
( )( )
( ) ( ) | |
( ) ( ) ( )
( ) ) ( ) (
log Re 2 ) ( )) ( (
log det log ) (
: Potential Khler of
) (
0
fy that veri B and A
0
0
with and Let
* *
* *
1
* *
*
*
* *
Z F Z g F
Z B A Tr Z F Z g F
ZZ I Tr ZZ I z F
A Z B B AZ Z g
B A A B
I B B A A
I
I
J J Jg g
B B
B A
g
t
t
t
=
+ + =
= =
+ + =

=
=
(

= =
(

=
+ +

+
+

(56)
Geometric Lift in Cartan-Siegel domain is then given by
the following:
( )
( ) ( ) | |
+ +

+ = + =
=
BB I Tr BB I g F
A B g
log det log )) 0 ( (
) 0 (
1
*
(57)
F.A. Berezin [3] has proved that for every symmetric
Riemannian space, there exist a dual space being compact.
The isometry groups of all the compact symmetric spaces are
described by block matrices (the action of the group in terms
of special coordinates is described by the same formula as the
action of the group of motions of the dual domain).
( )( )
|
|
.
|

\
|
= =
+ + =
|
|
.
|

\
|
=

I iI
iI I
C C C
A W A A W A W
A A
A A
2
1
with : Isometry
) (
1
1
22 21 12 11
22 21
12 11


(58)
Berezin coordinates for Siegel domain are given by
|
|
.
|

\
|
= = =
|
|
.
|

\
|
=
|
|
.
|

\
|
=
+
+
+

0
0
with , : ly equivalent or
,
1
* *
I
I
L L L I
A B
B A
A B
B A
t
t
t



(59)
( ) ( ) ( )
+ +

+ = + = = BB I trace BB I g F A B g ln det ln )) 0 ( ( ) 0 (
1
*
(60)
For this dual space, the volume and the metric are
invariant:
( ) ( )
( )
( )
( )




+
+ =

= =
=

WW I W W H
W W
W W H
g dW dW g ds
W W d
W W H W W d
n
L
det ) , ( where
, log
with
,
, ,
*
*
* 2
,
*
,
2
*
* *
(61)
For arbitrary Khlerian homogeneous space, the logarithm
of the density for the invariant measure is the potential of the
metric.
X. MEDIAN AND CONFORMAL BUSEMANN BARYCENTERS
The Frchet-median barycentre, in Poincar/Siegel unit
disk is equivalent to conformal Douady/Busemann
barycenters [12]. The principle is to assign to every
probability measure

on
{ } 1
1
= = z S
a point
( ) { } 1 < = z D B
so that the map
( ) B
is conform and
7
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013


http://ijetsp.info/article/IJETSPV1I301.pdf

satisfies:
0 ) ( . 0 ) (
1
= =

S
d B

There is a unique conformally way to assign to each
probability measure

on
1
S
a vector field

on D such
that:
0 ) ( . ) 0 (
1
= =

S
d

. For general
w
in D, if we set
z w
w z
z g
w
* 1
) (

=
, the assignment is given by


:
( )
( )
( ) 0 ) (
* 1
1 ) 0 (
) ( '
1
) (
1
2
) ( *
=
|
|
.
|

\
|


= =

S
g
w
d
w
w
w
w g
w
w



(62)

Douady-Earle definition of conformal barycenter
) ( B
of

is the unique zero of

in D.
) (z

can be written
according to
) (z

that is the unit tangent vector of geodesic


at D z pointing toward:
1
S
:

=
1
) ( ) ( ) (
S
d z z


(63)
In Poincar geometry of unit disk, the vector field

is the
gradient of a function

h
whose level lines are the horocycles
tangent to
1
S
at
1
S
:

=
1
) ( ) ( : with
S
d z h z h h



| |

=
|
|
.
|

\
|

==

1 1
) ( ) , ( ) , 0 ( ) (
1
log
2
1
) (
1
2
2
S
r
S
d r z d r d Lim d
z
z
z h

(64)
with
( ) w z d ,
the Poincar distance from z to w in D. We can
then observe that the Frchet median of N points in D,
{ }
N
w w w w ,..., , ,
3 2 1
, is the conformal barycenter of associated
push forward N points
{ }
N
,..., , ,
3 2 1
on
1
S
:
{ } { }
) ( ) (
,..., , , _ ,..., , ,
) (
3 2 1 3 2 1
1
t Lim t Lim
Barycenter Conformal w w w w Median
i
w
z
t
z
t
i
N N Frchet

+ +
= =
=
where
) ( ) (
) (
1
t Lim t Lim
i
w
z
t
z
t
i

+ +
= =
the push forward
association is given by the limit point when t tends to infinity
along the geodesic from the barycenter toward
i
w
.

Figure 9. Frechet Median Barycenter = Busemann Barycenter
All these results could be extended for Busemann
barycenter on Shilov boundary of Cartan Hadamard
manifolds. Following functional associated to median
barycenter
( )

n
H
z d z x d x ) ( ,
is simply connected and reached
its minimum. When z goes to infinity and converges to a
point of
n
H , distance function to z normalized converges
to Busemann function:
) , ( ) , ( lim ) ( z O d z x d x B
z
=

.Fixing the
origin
n
H O , let

finite measure on
n
H
. The following
strictly convex function reaches its minimum at a unique
point, independent of origin and called Busemann
barycenter:

=
n
H
d B x ) ( ) (

(65)
XI. RADAR APPLICATIONS FOR ROBUST ORDERED-STATISTIC
PROCESSING: OS-HDR-CFAR AND OS-STAP
In the following, we will apply previous tools to built
Robust Ordered-Statistic (OS) processing. Ordered-Statistic is
a very useful tool used in Radar for a long time to be robust
againt outliers on scalar data from secondary data. We will
define an OS-HDR-CFAR (Ordered-Statistic High Doppler
Resolution Constant False Alarm Rate) algorithm jointly
taking into account robustness of matrices median and high
Doppler resolution of regularized Complex Auto-Regressive
model. We will define also an OS-STAP (Ordered Statistic
Space-Time Adaptive Processing), based on median
computation of secondary data space-time covariance matrix
with Mostow/Berger fibration applied on Multichannel
Autoregressive Model.
This paper will not address Polarimetric Data processing,
but obviously these tools could be extended to compact
manifold to define Ordered statistic for Polarimetric
covariance matrices.

Figure 10. Doppler Processing, Antenna Processing, Space-Time Processing &
Polar processing
XII. ROBUST DOPPLER PROCESSING : OS-HDR-CFAR
The regularized Burg algorithm [2] is an alternative
Bayesian composite model approach to spectral estimation.
The reflection coefficients, defined in classical Burg
algorithm are estimated through a regularized method, based
on a Bayesian adaptative spectrum estimation technique,
8
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013

http://ijetsp.info/article/IJETSPV1I301.pdf

proposed by Kitagawa & Gersch, who use normal prior
distributions expressing a smoothness priors on the solution.
With these priors, autoregressive spectrum analysis is reduced
to a constrained least squares problem, minimized for fixed
tradeoff parameters, using Levinson recursion between
autoregressive parameters. Then, a reflection coefficient is
calculated, for each autoregressive model order, by
minimizing the sum of the mean-squared values of the
forward and backward prediction errors, with spectral
smoothness constraints . Tradeoff parameters balance
estimate of the autoregressive coefficients between infidility
to the data and infidility to the frequency domain smoothness
constraint. This algorithm conserves lattice structure
advantages, and could be brought in widespread use with a
multisegment regularized reflection coefficient version. The
regularized Burg algorithm lattice structure offers
implementation advantages over tapped delay line filters
because they suffer from less round-off noise and less
sensivity to coefficient value perturbations.

+ =
+ =

=
+ =
=
=
+ +

=
=
= =
= =

+ =


+ =

) ( . ) 1 ( ) (
) 1 ( . ) ( ) (
and
1 1 , .
1
) .( ) 2 .( with

. . 2 ) 1 ( ) (
1
. . . 2 ) 1 ( ). (
2
to 1 for : (n) .
1 and ) ( .
1
burst) per pulses nb. : (N 1 , ) ( ) ( ) ( f
1
*
1
1 1
) (
* ) 1 ( ) 1 ( ) (
) (
0
2 2
1
) (
1
1
0
2
) 1 ( ) (
2
1
2
1
1
1
1
) 1 ( ) 1 ( ) ( *
1 1
) 0 (
0
1
2
0
0 0
k f k b k b
k b k f k f
a
,...,n- k= a a a
a
n k
a k b k f
n N
a a k b k f
n N
M n Iteration
a k z
N
P
,...,N k= k z k b k
sation : . Initiali
n n n n
n n n n
n
n
n
n
k n n
n
k
n
k
n
n
k
N
n k
n
k
n
k
n
k n n
N
n k
n
k
n
k n
n
k
n
k n n
n
N
k

(66)
In the following figures, regularization property is
illustrated with deletion of spurious peaks. We select the AR
model of maximum order (number of pulses minus one).


Figure 11. (up) Non-regularized & (bottom) Regularized Doppler AR Spectrum
We conserve the sliding window structure of classical
CFAR : we compare the AR model under test by computing
its Information Geometry distance with median AR model of
secondary data in the neighborhood. Median Autoregressive
model is computed by :
For
0 , median
P
, we use classical median on real values
k
P
, 0

For
{ }
1
1

=
n
k k

:
{ } 0 with
0
,
,
0
= =

k,n
G k n k
n k
n n
k/ G w

(67)
*
n k,n
n k,n
k,n
.w
w


=
+
1
1
(68)
*
n median,n
n median,n
median,n
w
w

+
+
=
+
1
1

(69)
The detection test is finally based on computation of the
robust Information Geometry distance:
( ) ( ) | |
( )
*
, ,
*
, ,
1
1
2
, 0
0 , 2
, 1 , 1 0 , , 1 , 1 , 0
2
1
with
1
1
log
2
1
log
,
median i k i
median i k i
i
N
i i
i
k
median
median N median median k N k k
k N
P
P
n
P P d


=
|
|
.
|

\
|
|
|
.
|

\
|

+
+
|
|
.
|

\
|
=

=


(70)
In the following figure, we compare the classical
processing chain with new OS-HDR-CFAR

Figure 12. (top figure) Classical OS-CFAR after filter banks, (bottomfigure)
OS-HDR-CFAR
9
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013

http://ijetsp.info/article/IJETSPV1I301.pdf


Figure 13. OS-HDR-CFAR Algorithmwith illustration of two first reflection
coefficients
We have tested OS-HDR-CFAR on real recorded ground
Radar clutter with ingestion of synthetic slow targets.




Figure 14. Reflectivity of ground clutter in recorded sector (top),Comparison of
FFT Doppler spectrum(middle) and High Resolution Regularized Doppler
spectrum(down)
In the following figures, we give ROC curves with
Probability of detection versus probability of false alarm. We
observe that OS-HDR-CFAR is better (Pd =0.8) than OS-
CFAR/Doppler-Filters (Pd =0.65) for arbitrary fixed Pfa. We
could also observe that Information Geometry approach
provides better results than Optimal Transport Theory
approach (based on Wasserstein distance/barycenter : black
curve).

Figure 15. ROC curves for 3 approaches : OS-HDR-CFAR, OS-
CFAR/Doppler-Filters, & method based on Wasserstein barycenter/distance
(optimal transport theory)
10
International Journal of Emerging Trends in Signal Processing
ISSN(2319-9784) IJETSP , Volume 1 , Issue 3 March 2013


http://ijetsp.info/article/IJETSPV1I301.pdf


Figure 16. Probability of detection versus SNR for fixed Pfa=10
-5
, with
relative position in Doppler of the target normalized by Doppler Clutter
SpectrumWidth (=1 means that the target is positioned on the edge in Doppler
of the ground clutter)


Figure 17. Comparison of Mean in red and Median in blue for clutter level
estimation (top) and through Pd/Pfa curves (down).
XIII. CONCLUSION
Frchet Median with Information Geometry and Geometry
of of HPD(n) matrices is a new tool for Radar Signal
Processing that could improve drastically performance and
robustness of classical methods, in Doppler processing and in
STAP. Obviously, these approaches could be extended to
Array Processing and Polar Data Processing in the same way
on respectively spatial covariance matrix and Polar
covariance matrix. Future works will be dedicated to deepen
close relations of Information Geometry with Lagrange
Symplectic Geometry and Geometric Quantization.
ACKNOWLEDGMENT
We would like to thank Leon Brillouin Seminar members
for fruitful discussions on Geometric Sciences of
Information, hosted by Arshia Cont at IRCAM in Paris since
2009. We would more specifically thank Yang Le, PhD
student of Poitiers University, supervised by Prof. Marc
Arnaudon and myself, that has proved convergence of all
flows introduced in this paper. Yang Le has win the Thales
PhD Award 2012 for this work, with support of Cdric
Villani (Fields Medal 2010, Institut Henri Poincar Director)
Referee Letter.
REFERENCES
[1] L. Yang, Mdianes de mesures de probabilit dans les varits
riemanniennes et applications la dtection de cibles radar , Thse de
lUniversit de Poitiers, tel-00664188, 2011, Thales PhD Award 2012
[2] F. Barbaresco, Algorithme de Burg Rgularis FSDS. Comparaison avec
lalgorithme de Burg MFE , GRETSI conference , pp.2932 ,1995
[3] F. Barbaresco,Information Geometry of Covariance Matrix, Matrix
Information Geometry Book, F. Nielsen & R. Bhatia Ed., Springer, 2012
[4] M. mery, G. Mokobodzki, Sur le barycentre d'une probabilit dans une
varit, Sminaire de probabilit Strasbourg, n25,p.220-233, 1991
[5] W.F. Trench,, An algorithmfor the inversion of finite Toeplitz matrices,
J . Soc. Indust. Appl. Math. 12, pp.515522, 1964
[6] N.H. Bingham,Szegs Theorem and its Probabilistic Descendants,
2012, http://arxiv.org/abs/1108.0368v2
[7] H.J . Landau, MaximumEntropy and The Moment Problem, Bulletin of
the American Mathematical Society, pp.47-77, vol.16,n1, J an. 1987
[8] C. L. Siegel,Symplectic geometry,Amer. J . Math., (65), pp.1-86, 1943
[9] J .L. Koszul, Exposs sur les espaces homognes symtriques ,
Publicacao da Sociedade de Matematica de Sao Paulo, 1959
[10] P. Delsarte, Y.V. Genin,Orthogonal Polynomial Matrices on the Unit
Circle, IEEE Trans on CS., vol.25, n3, pp.149-160, March 1978
[11] R. Kanhouche, A Modified Burg AlgorithmEquivalent In Results to
Levinson Algorithm, http://hal.archives-ouvertes.fr/ccsd-00000624
[12] C.J . Douady, C.J . Earle,Conformally natural extension of
homeomorphisms of circle, Acta Math. 157, pp.23-48, 1986

Frdric Barbaresco (graduated from French High Engineering School
SUPELEC in 1991) became a Member (M) of SEE in 1991, a Senior Member
(SM) in 2000, and an Emeritus Member (EM) in 2011. The author become a
Member (M) of IEEE in 2012. F. Barbaresco was awarded by SEE Ampere
medal in 2007 and NATO SET Lecture Award in 2012 for his NATO Lecture
from2008 to 2011 in Europe and North America on WaveformDiversity and
Design for Advanced Radar Systems.
F. Barbaresco has worked for THALES group (www.thalesgroup.com) since
1991. He is Senior Scientist and Advanced Studies Manager in Advanced
Developments Department of Technical Directorate in Surface Radar Domain of
THALES AIR SYSTEMS.
11

Das könnte Ihnen auch gefallen