Géométries de l’information et des matrices de...
Transcript of Géométries de l’information et des matrices de...
Air Systems Division
Géométries de l’information et des matrices de covariances : application en RadarFrédéric BarbarescoPoitiers, 5 Mai 2010
Air Systems Division2
Acknowledgement
Marc Arnaudon, Équipe de probabilités, Laboratoire de Mathématiques et Applications (LMA),CNRS : UMR 6086, université de Poitiers Yang Le, PhD student (in collaboration with THALES),
université de Poitiers Léon Brillouin Seminar on « Geometric Science of
Information » launched in collaboration with : Frank Nielsen, LIX, Ecole Polytechnique Arshia Cont, IRCAM
Air Systems Division3
Séminaire Léon Brillouin « Science géométrique de l’information »
« La Science et la Théorie de l’Information », L. Brillouin, 1956
« Théorie scientifique de l’information d’une part,
mais aussi application de la théorie de l’information
à des problèmes de science pure. »
Théorie & géométrie de l’information
Probabilité & statistique
Géométrie Riemannienne (symétrique, Kählerien, métrique)
Géométrie & Groupes de Lie
Géométrie symplectique
Géométrie complexeGéométrie
discrète
Géométrie algébrique & arithmétique
Géométrie non commutative
Physique Statistique Thermo-
dynamique & Géométrie (Souriau,
Ruppeiner)
Science géométrique
de l’information
Physique quantique
Air Systems Division4
Léon Brillouin Seminar Corpus : « Geometric Science of Information » Place : IRCAM (Arshia Cont), Paris Animation : F. Barbaresco (Thales), Arshia Cont (IRCAM), Frank Nielsen
(X/LIX) Laboratories : IRCAM (Arshia Cont, Gerard Assayag, Arnaud Dessein) Polytechnique (Frank Nielsen) Mines ParisTech (Pierre Rouchon, Silvere Bonnabel, Jesus Angulo) Telecom ParisTech (Hugues Randriam) SUPELEC (Mérouane Debbah, Romain Couillet) UTT Troyes (Hichem Snoussi) Univ. Poitiers (Marc Arnaudon, Le Yang) Univ. Montpellier (Michel Boyom, Paul Bryand) Observation de Nice (Cédric Richard) INRIA (Jean-Paul Zolesio, Rama Cont, Xavier Pennec) Thales (Frédéric Barbaresco, Jean-Francois Marcotorchino, François Gosselin) …http://www.informationgeometry.org/Seminar/seminarBrillouin.html
Air Systems Division5
La SEE en quelques mots
Lieu de rencontres science-industrie-société Reconnue d’utilité publique (avantage fiscal) Environ 2000 adhérents et 5000 personnes impliquées Présence industrielle forte (~50%) 19 clubs techniques et 12 groupes régionaux Organise des conférences et journées d’études Suscite et attire des congrès internationaux sur le sol
français Membre institutionnel français de l’IFAC et de l’IFIP Remet des distinctions et médailles Produit 3 revues (REE, 3E.I, eSTA) Produit 3 monographies par an Présente sur le Web: http://www.see.asso.fr et groupe
LinkedIn SEE
Air Systems Division6
Méthodologie Signal, Image,
Information, Décision Automatique et
automatisation industrielle Génie logiciel Systèmes informatiques
de confiance Systèmes complexes
Applications Systèmes électriques Transports terrestres Systèmes aérospatiaux Sécurité Globale Ingénierie biomédicale
Les Clubs Techniques
Physique Ondes et propagation Foudre
Technologie Métrologie Composants Télécommunications Systèmes radars, sonars
et radioélectriques Systèmes optroniques
d’observation et de surveillance Matériels électriques Systèmes industriels
+ discussions avec IFIP pour meilleure prise en compte de leurs activités
Air Systems Division7
Publications SEE
Revues REE, Revue de l'Electricité et de l'Electronique 10 numéros par an. Couvre tous les domaines d’activité de la SEE: Electricité, Signal, Automation,
Electronique, Computing, Composants, Communications, etc. Première partie « nouvelles de nos métiers » et un dossier thématique Vente par abonnement ou au numéro. Archives consultables sur le site
3EI, Revue de l'Enseignement de l'Electrotechnique et de l'Electronique Industrielle Support à l’Education – cible les professeurs et les industriels des professions
concernées. La 3E.I développe des dossiers scientifiques, techniques et historiques. Vente par abonnement ou au numéro.
e-STA, Revue des Sciences et Technologies de l'Automatique Monographies SEE/CNRS Editions Bulletin SEE Informations sur la vie de l ’association et annonces et programmes des
événements.
Air Systems Division8
MONOGRAPHIE SEE/CNRS EDITIONS
Monographies Gestion de la Complexité et de l’information dans les
grands systèmes critiques Sécurité globale Micro et nanoélectronique À venir: « Le traitement de l'Information en interaction
avec les mathématiques et la physique », et « Le développement du réseau de transport d’électricité en France »
Comment commander? www.see.asso.fr
Prix : 39 EUR + 5 EUR (frais de port)
Air Systems Division
Preamble : Radar Data Structure
Air Systems Division10
Technology breakthroughs open new future for RADAR : Low Frequency Radar HFSWR Radar (HF Surface Wave Radar, HF Skywave Radar) VHF/UHF Radar (Alerteur,…)
Bi/Multi-static Radar : Passive Radar (A/D Radio, TNT, …) MIMO Radar (RIAS, …)
Bi-Band & Ultra-Wide-Band Radar Shared Aperture Antenna
Large deployable/deformable/conformal Antenna Modular Integrated Mast Multifunction/Multimission/Multi-Use Intelligent Radar « Cognitive & Software » Radar Full-Digital Radar Antenna « Waveform Diversity & Design » : coloured waveforms, sparse
sampling, Phase conjugation & Time Reversal
RENEWEL OF RADAR
Air Systems Division11
Radar Antenna measures the EM Wave to extract: Reflectivity Information
Doppler Information
Polarimetric Information
Radar Signal
Doppler fouillis atmosphérique Doppler cible (vitesse)
Ellipse polarisation sur sphère de Poincaré
Doppler relatif cible (ISAR)
Air Systems Division12
itffi
tf
r
cVr
rDoppler
eecRtuAtz
etuts
VfcV
cVf
Doppler
r
22
2
1/0
0
0
2.)( :Reception
).()( :Emission
.2/1
/2
ChristianDoppler
011
1
01
11011
cccc
ccccc
z
z
z
zER
n
n
nn
Time Covariance Matrix :Toeplitz Hermitian Positive Definite
HippolyteFizeau
Radar Antenna measures the EM Wave to extract: Reflectivity Information
Doppler Information
Polarimetric Information
Radar Signal
2
arctan1
2
ss
2
arctan22
21
3
sss
2sin2sin2cos2cos2cos
Im2Re2
0
0
0
0
*12
*12
22
21
22
21
3
2
1
0
ssss
zzzz
zzzz
ssss
S
n
kktéreflectivi
n
zn
Rz
zZ
0
21 1
Air Systems Division13
STAP (Space-Time Adaptive Processing) Principle
STAP Principle
Nml
ml
ml
ml
Ml
l
l
lt
z
zz
z
z
zz
Z2
1
2
1
, with
gate range lpulse m
element n
th
th
th
nmlz
jammerclutternoise
tt
t
RRRRVZVZER
nVZ
,
Air Systems Division14
Doppler-Fizeau : Poincaré or Einstein Model ?
"L'histoire des Sciences montre que les progrès de la Science ont constamment été entravés par l'influence tyrannique de certaines
conceptions que l'on avait fini par considérer comme des dogmes.Pour cette raison, il convient de soumettre périodiquement à un examen très approfondi les principes que l'on a fini par admettre
sans plus les discuter."
Louis de Broglie - Nouvelles Perspectives en Microphysique.
We show that the image by the Lorentz Transformation of a spherical light wavefront, emitted by a moving source, is not a spherical light wavefront but an ellipsoidal light wavefront. Poincaré’s elongated ellipse is the direct geometrical representation of Poincaré’s relativity of simultaneity. Einstein’s circles are the direct geometrical representation of Einstein’s convention of synchronisation. Poincaré’s ellipse supposes another convention for the definition of space-time units involving that the Lorentz Transformation (LT) of an unit of length is directly proportional to the LT of an unit of time: This is Poincaré’s definition of isotropic elongated distance (with dilated travel time). The historical (polemical) problem of priorities is therefore scientifically solved because Einstein’s explicit kinematics and Poincaré’s implicit kinematics are not the same.
Special Relativity: Einstein’s Spherical Waves versus Poincaré’s Ellipsoidal Waves, Yves Pierseaux, Annales de la Fondation Louis de Broglie, Volume 30 no 3-4, 2005
Air Systems Division15
Doppler Sagnac Experiment : 1913
GYROLASER
On ne saurait, à aucun point de vue, comparer cette expérience àcelle de M. Michelson. Celle-ci est du second ordre en fonction de
la vitesse de translation et son importance tient à ce qu’elle est venue mettre en évidence de manière aigüe la nécessité
d’introduire un cinématique nouvelle...
Paul Langevin
Sagnac non null result (1913) :
The main problem of rotating platform with Einstein’s kinematics is precisely Einstein’s invariance of one way speed of light, t+ = t− or t = 0 in the proper system.
This is the reason why Langevin solves the problem in the framework of General Relativity. In Poincaré’s relativistic kinematics we can have in the system of the source
t+ t− . With Poincaré’s elongated ellipse and Poincaré’s group with rotations [Reignier J.], we predict immediately (at the second order k) the experimentally measured difference of time.
2
4cωAγtt
2
2
02
2cvRtdx
cvdtt
R
Air Systems Division
New Geometric Foundation of Radar Processing
Air Systems Division17
Non-Euclidean Geometry & Lie Group Theory
“On pourrait dire d’une manière paradoxale que le point de départ du non-euclidisme réside dans l’épuration d’une notion pure, dans la simplification d’une notion simple…Le groupe apporte la preuve d’une mathématique fermée sur elle-même. Sa découverte clos l’ère des conventions, plus ou moins indépendantes, plus ou moins cohérentes” Gaston BACHELARD, La Philosophie géométrique, in « Le
nouvelle Esprit scientifique », 1934
Air Systems Division18
Objective
We propose new robust Radar algorithms based on : Geometry of Covariance Matrices Intrinsic Geometry of HPD(n) & SO(n) Information Geometry deduced from data statistical distributions
Mean/Barycenter/Center-of-Mass of sample covariance matrices Robust Median Estimation of sample covariance matrices in
inhomogeneous data Anisotropic Diffusion on a graph of covariance matrices
We apply this approach for : Doppler/Array Processing STAP Processing Phase Processing Polarimetry Processing
Air Systems Division19
Robust Detector should be defined based on Geometry of Covariance matrices
Signal Processing is classically based on flat metric & normed space. The flat metric can still be used because the set of SPD matrices is convex but
vector space assumption can yield to degraded algorithms. the set of SPD matrices with flat metric is not a geodesically complete space
since the geodesic A+t(B-A) is not a positive matrix for all t
We propose to use geometry of symmetric cone This geometry in the framework of Information Geometry , takes into account
matrix variance
Flat Metric versus Information Geometry Metric
)()( ,
2
))((,2
ABtAtBABA
BABATrBABABA TF
2/1 2/12/12/12/12/12/12/12/1
2/12/1
)( ,
, with log
ABAAAtABAAABA
XXTrXXXBAAt
TFF
Air Systems Division20
From Linear Algebra to Lie Group Theory & Metric Space
Since the mid-20th century, Radar Signal Processing is mainly based on : Euclidean Geometry : Euclidean geometry corresponds to
the choice of the group E(3) of distance-preserving transformations of the Euclidean space R3. Normed Space : It is a theorem of P. Jordan and J. von
Neumann that each normed space satisfying parallelogram law is Euclidean. Therefore, the Euclidean spaces are the normed spaces satisfying the semi-parallelogram law. Linear Algebra : Linear Algebra is concerned with the
study of vector spaces and transformation functions represented by matrices. A fundamental role in linear algebra is played by the notions of linear combination, span, and linear independence.
Air Systems Division21
From Linear Algebra to Lie Group Theory & Metric Space
Geometry of Metric Space : Maurice René Fréchet has introduced the entire concept of metric spaces and functional theory on this space, where a notion of distance (called a metric) between elements of the set is defined. Any normed vector space is a metric space by defining d(x, y) = ||y − x|| but not the contrary. Every metric space is a topological space. We are more interested by Jacques HadamardSpace, called complete CAT(0) space. CAT(k) spaces were named byMikhail Gromov and they are metric space with curvature that is bounded from above by k.
Geometry of Symmetric space: A symmetric space is a smooth manifold whose group of symmetries contains an inversion symmetry about every point. For Riemannian Symmetric space, these inversion are isometric geodesic symmetries. More generally for Lie theory a symmetric space is a homogeneous space G/H for a Lie group G such that the stabilizer H of a point is an open subgroup of the fixed point set of an involution of G. They were classified by Élie Cartan and Marcel Berger. More specifically, we are interested by Cartan–Hadamard-van-Mangoldt manifold, a Riemannian manifold that is complete and simply-connected, and has everywhere non-positive sectional curvature.
Air Systems Division22
From Linear Algebra to Lie Group Theory & Metric Space
Geometry of Kähler Manifold : A Hermitian symmetric space is a Kähler manifold which is a Riemannian symmetric space. Kähler Geometry was introduced by Erik Kähler and this is a manifold with unitary structure satisfying an integrability condition and where Kähler metric is given by Hessian of real function called Kähler potential. Especially, the non-compact Hermitian symmetric spaces can be realized as bounded domains in complex vector spaces.
Geometry of Lie Groups : a Lie group is a group and a differentiable oriented manifold where there is a bundle isomorphism between its tangent bundle and the product of itself with the tangent space at the identity. We are interested by Geometry on Lie Groups.
Geometry of Symplectic Manifold and Siegel Space : In his book “Symplectic Geometry”, C.L. Siegel has defined distance between complex matrices whose the imaginary part is Positive Definite. Siegel Upper-Half Space is an extension for complex matrix of the Poincaré’s upper half space (this is also a symmetric space). The associate metric and distance is invariant under generalized Möbius transform. A particular case is space of Symmetric or Hermitian Positive Definite Matrices.
Air Systems Division23
From Linear Algebra to Lie Group Theory & Metric Space
Geometry on Symmetric Cones : Symmetric Hermitian Space is equivalent to a tube domain, where we can apply Semi-Simple Lie Group Theory and Jordan Algebra. It could envisaged as a particular case of Bruhat-Tits space.
Geometry of Robust Statistic on abstract spaces & Fréchet-Karcher Mean or Median : Statistics on manifolds is a critical aspect of different fields of applied mathematic. Maurice René Frechet imprimatur has been appropriated within this domain. The Fréchet mean, alsoknown as the Karcher mean, is the point that minimizes the sum of the squared geodesic distance to each point in a set of points on a manifold. More generally, on a metric space, the Fréchet mean still be defined. Replacing square geodesic distance by geodesic distance, we can extend this approach to estimate a median in metric space (called Fermat-Weber point in Physic). We investigate the geometric median of a probability measure on a complete Riemannian manifold and prove the uniqueness. By regarding the Weiszfeld algorithm as a subgradient procedure, we introduce a sub-gradient algorithm to estimate the median and prove that this algorithm always converges without condition of the sign of curvatures.
Air Systems Division24
From Linear Algebra to Lie Group Theory & Metric Space
Geometry of Information : Information Geometry has been introduced by C.R.Rao, and axiomatized by N. Chentsov, with same roots that the well-known Cramer-Rao bound (Cramer-Rao bound has been introduced by Maurice René Fréchet in 1939, published in 1943 and extended to multivariate case by Georges Darmois in 1945), allows to build a distance between statistical distributions that is invariant to non-singular parameterization transformations.
Air Systems Division25
Geometry on Symmetric Cones Information Geometry & Geometry on Symmetric Cones
Close links with Carl Ludwig Siegel (1896-1981) works in Symplectic Geometry (Siegel nth upper-half space SHn) Geometry on Symmetric Cone and on Symmetric Space (Bruhat-Tits
Space, Cartan-Hadamard Manifold) have been studied in many fields
Air Systems Division26
German Mathematical Works
H. Karcher, “Riemannian center of mass and mollifier smoothing”, Comm. Pure Applied Math., °30,pp.509-541,1977
E. Kähler, “Über eine bemerkenswerte Hermitesche Metrik”, Abh., Math. Sem. Hamburg
Univ., n°9, pp.173-186, 1933
K. T. Sturm, “Probability measures on metric spaces of nonpositive curvature”. In vol.: Heat kernels and analysis on manifolds, graphs, and metric spaces, Contemp. Math. 338 ,
357-390, 2003
C.L. Siegel, "Symplectic Geometry", Academic Press, New York, 1964
H.C.F. von Mangoldt, “Über diejenigen Punkte auf positiv gekrümmten flächen, welche die
eigenschaft haben, dass die von ihnen ausgehenden geodätischen Linien nie aufhören, kürzeste Linien zu sein“, J. Reine Angew.Math.,
n°91, pp.23-52, 1881
Air Systems Division27
French Mathematical Works
M. Gromov,"Hyperbolic groups". Essays in group theory. Math. Sci. Res. Inst. Publ. 8. New York: Springer. pp. 75–263, 1987
M. Fréchet, “L’intégrale abstraite d’une fonction abstraite d’une variable abstraite et son application à la
moyenne d’un élément aléatoire de nature quelconque”, Revue Scientifique, pp. 483-512, 1944
J. Hadamard, « Les surfaces à courbures opposées et leurs lignes géodésiques », J. Math, sér. 5, 4, pp.27-73, 1898
E. Cartan, « Leçons sur la géométrie des espaces de Riemann », 2nd édition, revue et
augmentée. Paris, Gauthier-Villars, 1946
F. Bruhat and J. Tits, « Groupes réductifs sur un corps local », IHES, n°41, pp.5-251, 1972
Air Systems Division28
« At one point Siegel thought that too many unnecessary things were being published, so he decided not to publish anything at all »George PolyaThe Polya Picture Album, Encounters of a Mathematician, Birkäuser
Carl Ludwig SiegelWith George Polya
iYXZZdYdZYTrdsSiegel with 112
Carl Ludwig siegel
Air Systems Division
Center of Mass : From Appolonius of Perga toElie Cartan
Air Systems Division30
From Linear Algebra to Lie Group Theory & Metric Space
Elie Cartan & Center of Mass : Historically, von Mangoldt, E. Cartan & J. Hadamard proved that a simply connected surface of non-negative curvature is homeomorphic to the plane. For this kind of Manifold, any two points are joined by a unique geodesic, and we can define notion of center of mass and its uniqueness. German Hermann Karcher has introduced Riemannian center of mass, and proved that for negative curvature, the minimum is unique (it is named Karcher’s barycentre). The existence of a center of mass in the large for manifolds with non-positive curvature was proven and used by Elie Cartan back in the 1920’s. He used this to prove that maximal compact subgroups of Lie groups are always conjugate. The general case was employed by Eugenio Calabi in an unpublished note. This holds because a symmetric space of non-positive curvature is nothing but the quotient of a non-compact Lie group by one of its maximal compact subgroups. All these irreductible symmetric spaces have been classified by E. Cartan& M. Berger.
Air Systems Division31
Center of Mass : from Appolonius of Perga to Elie Cartan In Euclidean space, the center of mass is defined for finite set of points
by arithmetic mean:
Appolonius of Perga was the first to discover that this point minimizes the function of distances:
This extends to general Riemannian manifolds. Elie Cartan has proved that the function :
is strictly convex (its restriction to any geodesic is strictly convex as a function of one variable), achieves a unique minimum at a point called the center of mass of A for the distribution da.
Miix ,...,1
M
iicenter x
Mx
1
1
M
ii
xcenter xxMinx
1arg
M
ii
xcenter xxdMinx
1
2 ),(arg
A
daamdmf ),(21: 2
E. Cartan, “Groupes simples clos et ouverts et géométrie Riemannienne”, J. Math Pures & Appl., n°8, pp.1-33, 1929
Air Systems Division32
Center of Mass : from Elie Cartan to Herman Karcher Center of Mass is characterized by being the unique zero of the
gradient vector field:
where is the “exponential map” and is the tangent vector at of the geodesic from to :
Herman Karcher has introduced a gradient flow intrinsic on the Manifold that converges to the center of mass, called Karcher Barycenter :
In the discrete case, the center of mass for finite set of points is given by:
)(exp 1 A
m daaf
exp(.) )(exp 1 am
m m a mm Ta)(exp 1
)()0( with )(.exp)(1 nnnmnn mfmfttmn
Pushed by Normal Jacobi Field (Sum of Tangent vectors of
Geodesics)
Karcher Barycenter :
Normal Jacobi Field Equal to zero (Sum of Tangent vectors of
Geodesics = 0)
)(exp.exp1
11
M
iimmn xtm
nn
H. Karcher, “Riemannian Center of Mass and Mollifier Smoothing ”, Com. Pure & Applied Math. Vol30, 1977
Air Systems Division33
Center of Mass : from Herman Karcher to Maurice Fréchet Maurice René Fréchet, inventor of Cramer-Rao bound in 1939, has
also introduced the entire concept of Metric Spaces Geometry and functional theory on this space (any normed vector space is a metric space by defining but not the contrary). On this base, Fréchet has then extended probability in abstract spaces.
In this framework, expectation of an abstract probabilistic variable where lies on a manifold is introduced by Emery as an exponential barycenter :
In Classical Euclidean space, we recover classical definition ofExpectation E[.] :
xyyxd ),(
M. R. Fréchet, “Les éléments aléatoires de nature quelconque dans un espace distancié”, Annales de l’Institut Henri Poincaré, n°10, pp.215-310, 1948
xgEb )(xg x
0)(exp 1 dxPxgM
b
nn RX
Rp
n dxxpxgdxPxgxgEpqqRqp )()()()()()(exp, 1
M. Emery & G. Mokobodzki, “Sur le barycentre d’une probabilité sur une variété”, Séminaire de Proba. XXV, Lectures note in Math. 1485, pp.220-233, Springer, 1991
Air Systems Division34
Laplace Median Definition : « valeur probable »
mxEMinmediandxxPmedian
mX 5.0).(0
Laplace has proved in 1774 that : let F be the cumulative distribution function of an (absolute continuous) random variable; the median is defined as the value μ such that F(μ) = 0.5. Laplace proved that μ is also the value minimizing the average of the absolute deviations, where the deviation between two values is their L1 (called also Manhattan) distance i.e., the absolute value of their differences. Laplace called this value "le milieu de probabilité" or "la valeur probable". The term median has been introduced by Cournot in l’Exposition de la théorie des chances in 1883.
Air Systems Division35
Median/Fermat’s Point : From Laplace to Maurice Fréchet
Statistics on manifolds is a critical aspect of different fields of applied mathematic.
Center of Mass is not useful for robust statistic. Replacing L2 square geodesic distance by L1 geodesic distance, we can extend this approach to estimate a Median in metric space (called Fermat-Weber ‘s point in Physic). Fréchet studied Median statistic using
compared to :
Classically, in Euclidean space, Median point minimizes
or equivalently :
mxEMinmmmedian
2mxEMinmmmean
M
ii
xmedian xxdMinx
1),(arg
M
iii
xmedian xxxxMinx
1
/arg
Air Systems Division36
Median/Fermat’s Point : From Fréchet to Wesfeld
More generally, on a metric space, the Fréchet mean still be defined. Replacing square geodesic distance by geodesic distance, we can extend this approach to estimate a median in metric space
For Riemannian extension :
We cannot directly extend the Karcher Flow to median computation in the discrete case :
because could vanish if
A m
m
MinA
daaahdaamdmh
)(exp)(exp),(
21:
1
1
)(exp
exp.exp
11
1
1
M
k km
kmmn x
xtm
n
n
n
)(exp 1km x
n
xm kn
Air Systems Division37
Wesfeld Algorithm for Median in metric space
Yang Le & Marc Arnaudon have investigated the geometric median of a probability measure on a complete Riemannian manifold and prove the uniqueness. By regarding the Weiszfeld algorithm as a subgradient procedure, they have introduced a sub-gradient algorithm to estimate the median and prove that this algorithm always converges without condition of the sign of curvatures.
with
LE Yang, "Riemannian Median and its Estimation”, LMS Journal of Computation and Math., 17 Nov. 2009
nm n
n
nGk km
kmmn x
xtm
)(exp
exp.exp
1
1
1 nkm mxkG
n /
Air Systems Division
Information Geometry
Air Systems Division39
Combinatorial/Variational Foundation of Kullback Divergence
Combinatorial Fundation of Kullback Divergence Kullback Divergence can be naturally introduced by combinatorial elements
and stirling formula :
)(log),(
eEESupqpK qp
in
M
i i
ni
MMM nqNqqnnnP
i
1121 !
!,...,/,...,,
iq
M
ii Nn
1 Nnp i
i
n when ..2..! nenn nn
),(log.log11
qpKqppP
NLim
M
i i
iiMN
Let multinomial Law of N elements spread on M levels
with priors , and
Sirling formula gives :
Variational Foundation of Kullback Divergence Donsker and Varadhan have proposed a variational definition of Kullback
divergence :
Air Systems Division40
Kullback Divergence & VARADHAN’s Variational Approach
)(log),(
eEESupqpK qp
Donsker and Varadhan have proposed a variational definition of Kullback divergence :
),()1ln(),()()()(ln
)()(ln)()(ln)(
)()(ln)( :Consider
qpKqpKqpq
qppeEE
qp
qp
0)()(ln)()(ln)(),(
)()()(with
)()(ln)(
)(ln)(ln)(
)(
)(
qppeEEqpK
eqeqq
qqp
eEeEeEE
qp
qpqp
This proves that the supremum over all is no smaller than the divergence
Using the divergence inequality,
Link with « Large Deviation Theory » & Fenchel-Legendre Transform which gives that logarithm of generating function are dual to Kullback Divergence :
)(
(.)
)(
(.)
)(
log)(),(
)(log)()(),(
),()()()(log
xVqp
V
xV
V
p
xV
eEVESupqpK
dxxqedxxpxVSupqpK
qpKdxxpxVSupdxxqe
Air Systems Division41
Chentsov/Rao GeometryInformation Geometry and Kullback Divergence (Rao, 1945) Riemannian metric of Information is deduced from 2nd order Taylor
expansion of Kullback divergence :
3
,.).(
!21)/(),/( dOddgdxpxpK
jijiij
jijiij θ.θ
xpE.dxθ.θxpxpg
/log/log/
22
Link with Fisher Information Matrix Riemannian metric of Information is given by Fisher matrix I() :
jiijjiij θ
xp.θ
xpEnd g gI
/log/log)(a)(,
22 /log/log.. xpdExpdVardIdds T Cramer-Rao bound given by : 1ˆˆ
IE
T
)()()( 22 dswdsWw
Air Systems Division42
Chentsov Information Geometry Information Geometry & Kullback divergence Chentsov (student of Smirnov) was the first to identify Kullback geometric
properties. He has proposed an analogue of Pythagoras theorem :
qpE.dx
xqxp.xpnd K(p,q)
.K(m,p)h D(m,p)(q,p) witD(m,q)D(m,p)then, D
.dξpqm).(qR or .dξ
pqm.: such that
ns p,q,m istributio and the d, a K-bowlrR(p,R),Let B
p
KKK
K
ln//log/a
2
0lnln
210
222
2
mq
p
Chentsov has named it « Asymmetric PythagorianGeometry ». Locally, this geometry could be identified as a symmetric Riemannian Geometry : Information Geometry
Air Systems Division43
Example : Gaussian scalar distributionGaussian Law Fisher Information matrix is given by :
mIEI
T and )(ˆˆ with
2001
.)( 12
2
22
2
2
2
22
2.2.2..
ddmddmdIdds T
Fisher matrix induced the following differential metric :
.2
imz 1
iziz
22
22
1.8
dds
Poincaré model of hyperbolic Space :
Geometry ofGaussian LawIs Geometry of
Hyperbolic Poincare Space
Air Systems Division44
Example : Gaussian scalar distribution Gaussian Law Metric :
If we set , we can integrate along one radial :
Homeomorphisme is used then to compute distance between two arbitrary points in the unit disk :
Distance between two Gaussian Law is then given by :
rrd
rdrds
11ln.2
1.8
2
22
)(0 and .1
)( .
je
and .2
1),( with
),(1),(1ln.2,,,
2
22112
izizimz
mmD
r
Air Systems Division45
Information Geometry of Multivariate Gaussian PDF
Multivariate Gaussian Distribution of zero mean Riemannian Information metric is given by Rao Metric
T
ZRZn
ABTrBAAAA
RdRRdRRTrdseR
RZp
, and ,with
.. det1/
2
22/12/12121
Multivariate Gaussian of non zero mean Isometries are given by following homeomorphisms :
2112 dRRTrdmdmds T
),(x, with '
isometry ,',',222 RnGLRAadsdsds
RAAamARmRmn
TT
22
111
1then As
RR dsdsdRRRdRIRR Invariance by inversion
12
1121 ,, RRDRRD
Air Systems Division46
Distance between Multivariate Gaussian Law of zero Mean From Rao Metric to Rao distance
For , Rao distance is given by extented eigenvalues :
0det...detwith
log..log,
122/1
122/1
1
1
222/112
2/1121
2
RRIRRR
RRRRRDn
kk
TABTrBA , with ,2
James Theorem If , then is driven by a Khi-2 law with n.(n+1)/2
degrees of freedom.21 RR
n
kk
n1
2log2
Swain Generalization (contraste function) Distance from contraste function C3 on [0,[ :
n
kkvRRD
121
2 )(, 1 , 0 if 0)( and2/1)('' , 0)1(')1(
v
vvv
)()( if ,, 121
221
2 vwRRDRRD wv
22/12/12 .. RdRRds
Rao metric
1
1.)1()log(n
nn
nIAA
21 mm
Air Systems Division47
Information Metric & sphericity testSphericity Test Spherecity test based on Information Geometry is different of classical one :
0det
log21,
:0 with :
1
22
1
0
nin
n
i
inn
nn
nn
I
IDIHIH
n
i
n
iiinn
nn
iiopt n
ID1
2
1
22/1
1
log1log21,
Sphericity test based on Kullback divergence :
n
ii
n
iinnnopt
nnnn
nnIKtr
n
nntrIK
11log1log
21,1
logdetlog121,
Air Systems Division
Geometry of Covariance Matrix
Air Systems Division49
Problem : « Barycentre of N HPD matrices »
Mean of N HPD matrices
Median of N HPD matrices (Fermat-Weber Point)
Matrix Mean/Median Barycenter : Optimisation criteria
N
kk
XXBXdXfX
1,arg)(arg minmin
2
N
kk
XXBXdXfX
1
22 ,arg)(arg minmin
VTBXfgrad X
N
kkX
1
12 exp)(
1
N
kk
AXBXdXfX
11 ,arg)(arg minmin
VT
BXdBXfgrad X
N
k k
kX
1
1
1 ,exp)(
nXn XfgradXn 21 .exp
NkkB 1
nXn XfgradXn 11 .exp
Air Systems Division50
Fréchet-Karcher Center of Mass Gradient Flow Gradient flow on Manifold
Pushed by Normal Jacobi Field (Sum of Tangent vectors of
Geodesics)
Fréchet-Karcher Barycenter :
Normal Jacobi Field Equal to zero (Sum of Tangent vectors of
Geodesics = 0)
Air Systems Division51
Information Geometry for Gaussian Multivariate Law of zero mean and Intrinsic Geometry of Hermitian Positive Definite matrices (particular case of Siegel Space) lead to same metric and distance
Information Geometry :
Geometry of Siegel Upper-Half Space :
Metric and Geometry of Covariance Matrices
0Im/),( Y(Z)CnSymiYXZSH n
iYXZZdYdZYTrdsSiegel with 112
nRYX 0
212nn dRRTrds
212nn dRRTrds
nn
nnnnn
RRTrn
nnn
RRE
mZmZR
eRRZp nn
ˆ and
.ˆwith
..)()/(1.ˆ1
*
2
.)/(ln)(
ji
nnij
ZpEg
0nm
with
Air Systems Division52
Information Geometry is given by : Advantages based on Information Geometry : invariante metric by non singular parametrization change
Metric taking into account parameters statistics
« Good Metric » for Information Geometry Meaning
212nn dRRTrds
)()()( 22 dswdsWw
dRddIdds
IER
..).(.et
Rao-Cramer de borne : ˆ
12
1
Air Systems Division53
Intrinsic metric is given by : Advantage based on Siegel Geometry : Siegel Space Isometries are given by quotient group
with symplectic group :
Only one metric is invariant by , Siegel metric :
« Good Metric » for Intrinsic Geometry Meaning
212nn dRRTrds
nSH nIRnSpRnPSp 2/),(),( ),( FnSp
1)(
DCZBAZZM
DCBA
M
nTT
TT
IBCDADBCA
FnSpDCBA
M
symmetric et ),(
),2(0
0 , /),2(),( RnSL
II
JJJMMFnGLMFnSpn
nT
)(ZM ZdYdZYTrdsSiegel
112 iYXZ avec
nRYX 0
212nn dRRTrds
Air Systems Division54
Metric is given by : Associate distance is : Obtained by integration :
In general case, for Siegel Space :
From metric to Distance 212 dRRTrds
n
kkRRRRRd
1
222/112
2/1121
2 log..log,
0det 12 RR avec
0 with XSHiYXZ n
n
n
k k
kSiegel SHZZZZd
21
1
221
2 , with 11
log,
with 0.),(det 21 IZZR
12121
1212121,
ZZZZZZZZZZR
Air Systems Division55
In case of Siegel Space We can observe that 2nd derivative of in is given by :
From Metric to Distance
ZZRZ ,1 ZZ 1
11112 .2/1)(2 YZddZYZZZdZZdZRD
111
1111,
ZZZZZZZZZZR
RDTrZddZYYTrds 2112 .2
Air Systems Division56
Poincaré & Siegel Upper Half Plane & Disk
0et avec
*112
2
2
2
222
yiyxzdzdzyyds
ydz
ydydxds
- *1*1*2
22
22
11
1
dwwwdwwwds
w
dwds
),(et ),( avec
*112
CnHDPYCnHermXiYXZ
dZdZYYTrds
*1*1*2 11 dWWWdWWWTrds
izizw
1 iIZiIZW
Poincaré Upper Half Plane Poincaré Unit disk
Siegel Upper Half Plane Siegel Unit disk
Air Systems Division57
Poincaré Upper-half plane
Air Systems Division58
Hyperbolic Geometry : Poincaré, Klein, Minkowski
Air Systems Division59
Poincaré Upper Half Plane & Disk in Art (Escher )
Air Systems Division60
Hyperbolic Geometry in Art
Irène Rousseau
Air Systems Division61
Siegel Upper Half Space
0Im/),( Y(Z)CnSymiYXZSH n
Siegel Upper Half Space
X
0Y
ZdYdZYTrds 112
n
iiiZZd
1
221
2 1/1log,
kkk YiXZ .
1k
2k
212 dRRTraceds
kkkk RNWRiZ ,0 if .
1k
2k
n
kkRRd
1
221
2 log,
12121
1212121,
ZZZZZZZZZZR
0...det 2/112
2/11 IRRR
0.,det 21 IZZR
Air Systems Division
Covariance Matrix Mean
Air Systems Division63
Covariance Matrix Mean
2
N
kk
XXBXdXfX
1
22 ,arg)(arg minmin
VTBXfgrad X
N
kkX
1
12 exp)( nXn XfgradX
n 21 .exp
2/12/1
2/12/12/12/11
2/12/1
)(exp
logexp
XeXV
XUXXXUVXX
X
X
2/1
log2/1
11
2/12/1
n
XBX
nn XeXX
N
knkn
with
2/12/12/12/12 log,log, XBXXBXBXd kkkwith
Air Systems Division64
Geodesic between two matrices X and Y given by :
Mean is given by Karcher Barycenter :
X
Y
Geodesic and Symmetrized Geometric Mean
2/1 2/12/12/12/1log2/1 2/12/1
)( XYXXXXeXt tYXXt
2/1log2/1
2/1..2/1
2/12/12/12/11
2/12/1
2/12/1
),(exp
),(exp
logexp)(
XeXtv
XeXtv
XYXXXVVgradv
YXXtYXX
XvXtYXX
XYX
YX
YX
2/1/21 2/12/12/12/1 XYXXXYX
YX )1( and )0( 1,0with t
Air Systems Division65
Gradient Flow on Manifold Gradient flow on Manifold
Tangent at X of geodesic from X to Bk
Sum of tangents at X of geodesics from X to
2/12/12/12/10
2/1log2/12/1 2/12/12/1
log)()(
2/12/1
XXBXXdt
tdXeXXXBXXt
ktk
k
XBXttkk
k
kkk
2/1
1
2/12/12/1
10 log )( XXBXXG
dttdG
N
kkX
N
kt
k
kX
NkkB 1
Air Systems Division66
Karcher Barycenter & Jacobi Field
2/1log2/12/1 2/12/12/1 2/12/1
)( XeXXXBXXt XBXttkk
k
2/12/12/12/10 log)( XXBXX
dttd
ktk
N
k
N
kkt
k XXBXXdt
td1
2/1
1
2/12/12/10 0log)(
1B 2B
3B
4B
5B
6B
Air Systems Division67
Let M be a connected Riemannian manifold and p a point of M. A map f defined on a neighborhood of p is said to be a geodesic symmetry, if it fixes the point p and reverses geodesics throughthat point. It follows that the derivative of the map at p is minus the identity map on the tangent space of p. On a general Riemannian manifold, f need not be isometric, nor can it be extended, in general, from a neighbourhood of p to all of M. M is said to be locally Riemannian symmetric if its geodesic
symmetries are in fact isometric, and (globally) Riemannian symmetric if in addition its geodesic symmetries are defined on all of M. A Riemannian Manifold (M,g) is symmetric when, for all point x
of M, there exist one isometry σx:MM such that : σx(x) = x ; dσx(x) = − Id.
This isometry σx is called involution at x.
Riemannan Symmetric Space
Air Systems Division68
For considered space, geodesic between two matrices A et B is given by :
For Symmetric space of Symmetric or Hermitian Definite Positive Matrices, for each couple (A,B), there exist one bijective isometry such that :
and This isometry has a unique fixed point Z which is given by
barycenter of (A,B), the geometric mean previously defined :
with
And properties : and This different of classical approach in signal processing with
assumption of an euclidean vectorial normed space :
with and
Riemannian Symmetric Space
1,0 avec )( 2/1log.2/12/1 2/12/12/1 2/12/1
tAeAABAAAt BAAtt
),( BAG
),( BAG BA ),( ABG BA
ZXdXXGd BA ,2,),( )(X -1
),( BABAXG BA 2/12/12/12/12/1 ABAAABA
X- ),( BABAXG BA
2BABA
FBA
),( BABAG BA ),( IBAdG BA
Air Systems Division69
Geodesic and Symmetrized Geometric Mean
),( AGB BA
BGA BA ),( BAXBAXG(A,B)
X
2
BABA
Normed Space : Isometry(classical signal processing approach)
),( AGB BA
BGA BA ),(
)(X -1),( BABAXG BA
X
2/1/21 2/12/12/1 ABAAABA
Metric space : Isometry
1,0)( 2/1 2/12/12/1
tABAAAt t
Normed Space : Geodesic
A)0(
B)1(
A)0(
B)1(
1,0)(
tABtAt
Metric space : Geodesic
Null curvature Negative curvature
Air Systems Division70
2/12/12/12/12/12/12/12/12/12/1
2/12/12/12/12/12/1
2/12/12/12/12/12/11
))(()(
ABAAAXBAAXAA
BAAXAAXAABAAAXAXAABXXA
Symmetrized Geometric Mean Involved Mean is a Symmetric or Hermitian Positive Definite Matrix
given by :
This is a Karcher Barycenter, that minimize :
In case where matrices can commute, we recover classical geometric mean :
which is not symmetric positive definite An other way of computaion of « symmetrized »Geometric Mean is
given by solution of following Ricatti Equation :
2/12/12/12/12/1 ABAAABA
22/12/122/12/1 loglog XBBXAAMin
X
2/1ABBA
Air Systems Division
Covariance Matrix Median
Air Systems Division72
N
kk
XXBXdXfX
11 ,arg)(arg minmin
Covariance Matrix Median
2/12/1
2/12/12/12/11
2/12/1
)(exp
logexp
XeXV
XUXXXUVXX
X
X
0det and
loglogwith 1
22/12/1
2/1loglog
2/11
12/12/1
2/12/1
kn
M
iiFnkn
nXBXXBX
nn
BX
XBX
XeXX
N
k Fnkn
nkn
1
VT
BXdBXfgrad X
N
k k
kX
1
1
1 ,exp)( nXn XfgradX
n 11 .exp
Air Systems Division73
Gradient Flow Convergence Problem Gradient Flow for Median Computation
Cannot be differentiate at point X : Convergence Problem In theory : regularization in neighborhood of dat sets (algorithm of
Huiling Le, but slow convergence) In practice : Very often, data sets are around median point : convergence Take an adpated iterative step
0det and
loglogwith 1
22/12/1
2/1loglog
2/11
12/12/1
2/12/1
kn
M
iiFnkn
nXBXXBX
nn
BX
XBX
XeXX
N
k Fnkn
nkn
Air Systems Division74
Median sub-gradient algorithm Proof : See THALES PhD Thesis (Yang Le supervised by Marc
Arnaudon)
otherwise 1
allfor if : criteria Stopping
0det and
loglogwith
,1,loglog
1
22/12/1
2/12/1
2/12/1
2/12/11
NS
Bk, XS
BX
XBX
XBNkXBXXBXS
XeXX
Fn
knFn
kn
M
iiFnkn
nkk
Fnkn
nknn
nSS
nnn
Fn
n
Air Systems Division75
Illustration of Time Covariance Matrices (Doppler= Raw Doppler
Spectrum
Mean Doppler Spectrum
Median Doppler Spectrum
Air Systems Division
Anisotropic Diffusion on a graph of covariance matrices
Air Systems Division77
Diffusion Heat Fourier Equation on matrices graph Diffusion Fourier heat equation on 1D graph of scalar values In a normed vectorial space in 1D case, diffusion equation is
approximated by discret Laplacian :
with arithmetic mean : Discret Fourier heat equation can be written :
Based on analogy with Karcher Flow, diffusion equation on 1D graph of Symmetric or Hermitian matrices can be written with geometric mean between :
nnnnnn uu
xxuu
xuu
xtu
xu
tu
ˆ21
211
2
2
2/ˆ 11 nnn uuu
2,,,,2,12 with ˆ.).1(ˆ2
xtuuuu
xtuu tntntntntnn,t
tnA ,ˆ tntn AA ,1,1 ,
2/1
,1/21 2/1
,1,12/1,1
2/1,1,2
2/1,
2/1,,
2/1,
2/1,
2/1,
ˆlog22/1
,1,
ˆ and x
t2with
ˆ2/1
,,2/1
,2
tntntntntntn
tntntntntntn
XXXx
t
tntn
XXXXXX
XXXXXXeXXtntntn
J. Fourier
Air Systems Division78
Isotropic Diffusion based on Fourier Heat Equation
2
2
2
2
yu
xuu
tu
Air Systems Division79
Diffusion Heat Fourier Equation on matrices graph Diffusion Equation on a 1D graph of matrices
Geometric Details : colinearity of tangent vectors for geodesics and between , and :
As in Euclidean case weighted mean is replaced by geodesic weighted barycenter :
2/1
,1/21 2/1
,1,12/1,1
2/1,1,2
2/1,
2/1,,
2/1,
2/1,
2/1,
ˆlog22/1
,1,
ˆ and x
t2with
ˆ2/1
,,2/1
,2
tntntntntntn
tntntntntntn
XXXx
t
tntn
XXXXXX
XXXXXXeXXtntntn
)(1,
, tn
tn
XX )(,
,
ˆ tn
tn
XX tnX , 1, tnX tnX ,
ˆ
2/1,,
2/1,2
2/1,1,
2/1,
ˆlog2log
tntntntntntn XXXx
tXXX
tn
tntnXtnX
tn
tn
tn
tn XX
XX
XX TT
dd
xt
dd
,
,1,,
,
,
1,
, ˆ
0
ˆ
2
0
//)(
.2)(
yxyx )1(.
2/12/12/12/1 ABAAABA
J. Fourier
Air Systems Division80
Diffusion Heat Fourier Equation on scalar 1D graph Diffusion Equation on a 1D graph of scalar values
J. Fourier
tnu ,
tntntn uuu ,12/1,1,ˆ
tnu ,1tnu ,1
1,0.).1(
vuvu
t
t+1
range
tnu ,tnu ,1 tnu ,1
tnu ,ˆ
1n,tu
time
tntntn uuu ,,1, ˆ
1
0
2ˆ and 2with
ˆ.).1(
,1,1,2
,,1
tntntn
tntnn,t
uuu
xt
uuu
Air Systems Division81
Diffusion Heat Fourier Equation on matrices 1D graph Diffusion Equation on a 1D graph of matrices
2/1
,1/21 2/1
,1,12/1,1
2/1,1,2
2/1,
2/1,,
2/1,
2/1,
2/1,
ˆlog22/1
,1,
ˆ and x
t2with
ˆ2/1
,,2/1
,2
tntntntntntn
tntntntntntn
XXXx
t
tntn
XXXXXX
XXXXXXeXXtntntn
J. Fourier
tnX ,
tntntn XXX ,12/1,1,ˆ
tnX ,1tnX ,1
1,0
2/1 2/12/12/1
ABAAABA
t
t+1
range
tnX ,tnX ,1 tnX ,1
tnX ,ˆ
1n,tX
time
tntntn XXX ,,1,ˆ
1
0
Air Systems Division82
Diffusion Equation : Discret & continuous Diffusion Equation on 1D graph of Symmetric or Hermitian
Positive Definite Matrices :
To recover diffusion equation in continuous case, Campbell-Hausdorff equation should be used :
2/1
,1/21 2/1
,1,12/1,1
2/1,1,2
2/1,
2/1,,
2/1,
2/1,
2/1,
ˆlog22/1
,1,
ˆ and x
t2with
ˆ2/1
,,2/1
,2
tntntntntntn
tntntntntntn
XXXx
t
tntn
XXXXXX
XXXXXXeXXtntntn
YXXYYXterms
XYYYXXYXYXee YX
,with 4
,,121,,
121,
21log
Air Systems Division83
Diffusion Equation : Discret & continuous Discret Diffusion Equation by mean of Campbell-Hausdorff
Equation :
2/1
,,2/1
,2/1
,1,2/1
,2
2/1,,
2/1,2
2/1,1,
2/1,
2/1,,
2/1,
2/1,1,
2/1,
2/1,,
2/1,
2/1,1,
2/1,
ˆloglogˆloglog
ˆlog,log1
ˆlog2log
0ˆlog,log21
ˆloglog
0log2/1
,,2/1
,2/1
,1,2/1
,2/1
,,2/1
,2/1
,1,2/1
,
tntntntntntn
tntntntntntn
tntntntntntn
tntntntntntn
XXXXXXXXXXXX
XXXXXXx
XXXxt
XXX
XXXXXX
XXXXXX
eeIee tntntntntntntntntntntntn
2
2
2
2 log,log2
loglogx
Xt
Xdtx
Xt
X
Air Systems Division84
Anisotropic Diffusion Equation Classical Anisotropic Diffusion : Diffusion Equation on a Manifold is given by Laplace-Beltrami
equation : Riemannian Metric :
Diffusion Equation :
1D isotropic case :
1D anisotropic case :such that :
2
222
xu
tudxds
Njiji
N
kjiji ggdxdxgds
1,,1
,2 with
Njiji
ji j
ij
i
ggxugg
xgtu
1,,1
, with )det(
)det(1
22
222 1 dxxududxds
2
1xug
xu
xu
xxu
tu
2/122/12
11
Air Systems Division85
Anisotropic Diffusion Equation : Scalar 1D Case 1D anistropic case :
Discret anisotropic diffusion equation :
with
xu
xu
xxu
tu
2/122/12
11
xuu
xuu
xt
uu tntntn
tntntn
tntnn,t
,1,,
,,1,
,,1
.
2
,1,1, 1
xuu tntn
tn
2
,,1, 1
xuu tntn
tn
2
,1,, 1
xuu tntn
tn
Air Systems Division86
Anisotropic Diffusion : scalar case
IsotropicDiffusion
AnisotropicDiffusion
Air Systems Division87
Anisotropic Diffusion : scalar case
In Blue, signal during anisotropic diffusion
In Greenn, substraction of blue signal to original
signal(targets detection)
Air Systems Division88
Anisotropic Diffusion : 1D Matrices Graph Case In 1D anisotropic Case, by analogy with scalar case, extension
for graph of matrices :
For distance , we use :
0detwith
loglog,
21
1
22/112
2/1121
XX
XXXXXdM
iiF
2/1
,1 2/1
,1,12/1,1
2/1,1
2/1,1
log2/1,1,
2/1,
2/1,,
2/1,
2/1,
2/1,
ˆlog2/1,1,
2/1,1,1
2/1,1
2/1,,
2/1,
ˆwith
ˆ
tntntntntntnXXX
tntn
tntntntntntnXXX
tntn
XXXXXXeXX
XXXXXXeXXtntntn
tntntn
2/12,1,
,
2/12,,1
,
2/12,1,1
,2,,,
,,
,
,1 and
,1
,1,
. ,
xXXd
xXXd
xXXd
xt
tntntn
tntntn
tntntn
tntntn
tntn
tn
Air Systems Division89
Diffusion Isotrope Simulation
Air Systems Division90
Diffusion anisotrope Simulation
Air Systems Division
new processing chain in Doppler Processing domain
Air Systems Division92
Rationale for Use of Diffusive CFAR : Low Altitude Threats New Operational Requests Detection of new low altitude threats (small and/or stealth, agile, asymmetric,…) Increase Reaction Time against “Ultra Critical threats”
Classical CFAR are not edge-preserving & poorly take into account Clutter statistic : probability of detection is not optimal close to clutter transitions
Clutter transitions are most threatening areas : crest-line & unmasking areas (threats : furtive helicopter pop-up with missile shooting, low altitude cruise missile & UAV, asymmetric threats, rockets/batteries,…)
Coast-Lines
Crest-Lines
Example of Naval radarradial in Littoral Area
UAV Heli ULA Cruise
Define Edge-preserving CFAR based on clutter statistic
Air Systems Division93
Rationale for Use of Diffusive CFAR : Low Altitude Threats
BEHIND THISCREST LINE
…
Increase Survivability against Lethal Threatwith Limited Time Exposure behind Crest-Lines :
Improve Critical Reaction Time
…LETHAL THREAT
WITH LIMITEDTIME EXPOSURE
Classical Helicopter
Helicopter withLimited Exposure
Potential ThreatBehind each Crest-lines
Air Systems Division94
Principle of Doppler CFAR based on “Matrix Mean”
If we can define distance between two HPD matrices and Mean of N HPD matrices : Detection is given by threshold of
distance between matrix of cell under test and « Matrix Mean » of the neighborhood
For short time series waveforms, Classical FFT or Doppler Filter Banks are not efficient, and suffer of the following drawbacks : Poor Doppler Resolution If Target Doppler is in between two
Doppler filters, detection is sub-optimal High intensity of Ground Clutter is not
limited to zero-Doppler filter but pollution is spread over all filters due to poor Filter-Banks Resolution & Doppler Filter side lobes.
Air Systems Division95
DOPPLER MATRIX CFAR
Main Goal Improve Doppler Detection of small/stealth targets in inhomogeneous clutter
using Doppler Matrix CFAR based on Differential Geometry of Covariance Matrices space (CFAR Mean or CFAR median)
Main advantages Improve performances compared with classical FFT (or Doppler Filter Bank)
in case of Waveform Bursts with few number of pulses (<16) Regularized High Resolution Doppler Analysis Based on statistics of parameters (Cramer-Rao Bound via Fisher matrix) Robust Environment assessment based on Median
Improve detection in inhomogeneous clutter (e.g. : Low altitude target on Ground Clutter, small target on Sea clutter)
Improve detection for closely spaced targets
Possible Extensions Use for sparse regular sampling waveforms (for pulses interleaving) Use for STAP (Spatio-Temporal Adaptative Processing) on spatio-temporal
covariance matrices Use for DORT method (Time Reversal & retrodirective techniques)
Air Systems Division96
Doppler Matrix CFAR by Matrix Mean
FFTI&Q
log|.|
Sliding Mean
>S
log|.|
Sliding Mean
>S
OR
I&QCovariance
Matrix
R
-
-
Robust
Matrices
Distance
Sliding Matrix Mean
>S
1B 2B kB 1kB NB
Air Systems Division97
Doppler Matrix CFAR by Matrix Median or Anisotropic Diffusion
I&QCovariance
Matrix
R
Robust
Matrices
Distance
Anisotropic Diffusion
on 1D graph of matrices
>S
I&QCovariance
Matrix
R
Robust
Matrices
Distance
Sliding Matrix Median
>S
Anisotropic Diffusion
Air Systems Division98
Doppler Matrix CFAR by Matrix or Complex Autoregressive Medians
I&QComplex
Autoregressive
Models (CAM)
Robust
CAM
Distance
Sliding CAM median
>S
Median of Complex Autoregressive Models
I&QCovariance
Matrix
R
Robust
Matrices
Distance
Sliding Matrix Median
>S
Air Systems Division99
Doppler CFAR by Matrices Mean
I&QCovariance
Matrix
R
Robust
Matrices
Distance
Sliding Matrix Mean
>S
2
N
kk
XXBXdXfX
1
22 ,arg)(arg minmin
VTBXfgrad X
N
kkX
1
12 exp)( nXn XfgradX
n 21 .exp
2/12/1
2/12/12/12/11
2/12/1
)(exp
logexp
XeXV
XUXXXUVXX
X
X
2/1
log2/1
11
2/12/1
n
XBX
nn XeXX
N
knkn
avec
2/12/12/12/12 log,log, XBXXBXBXd kkkwith
Air Systems Division100
N
kk
XXBXdXfX
11 ,arg)(arg minmin
Matrix Median CFAR
I&QCovariance
Matrix
R
Robust
Matrices
Distance
Sliding Matrix Median
>S
2/12/1
2/12/12/12/11
2/12/1
)(exp
logexp
XeXV
XUXXXUVXX
X
X
0det and
loglogwith 1
22/12/1
2/1loglog
2/11
12/12/1
2/12/1
kn
M
iiFnkn
nXBXXBX
nn
BX
XBX
XeXX
N
k Fnkn
nkn
1
VT
BXdBXfgrad X
N
k k
kX
1
1
1 ,exp)( nXn XfgradX
n 11 .exp
Air Systems Division101
Anisotropic Diffusion
I&QCovariance
Matrix
R
Robust
Matrices
Distance
Anisotropic Diffusion
On 1D graph of matrices
>S
2/1
,1 2/1
,1,12/1,1
2/1,1
2/1,1
log2/1,1,
2/1,
2/1,,
2/1,
2/1,
2/1,
ˆlog2/1,1,
2/1,1,1
2/1,1
2/1,,
2/1,
ˆwith
ˆ
tntntntntntnXXX
tntn
tntntntntntnXXX
tntn
XXXXXXeXX
XXXXXXeXXtntntn
tntntn
Air Systems Division
Matrix CFAR byComplex Autoregressive Model Median
Air Systems Division103
Covariance Matric Toeplitz structure is not conserved by previous Gradient Flow. Need for Using Complex Autoregressive Model Parametrization Gradient Flow driving reflection coefficients to converge toward
CAR Median
Median of Complex Autoregressive Model
I&QComplex
Autoregressive
Model (CAR)
Robust
CAR
Distance
Sliding CAR Median
>S
Tnn P 110
)(
1
122
22
0
02
1)(.
n
ii
in
din
PdPnds
,arg1
min
N
kkmed d
with
And metric
Air Systems Division104
Kähler Geometry Erich Kähler Geometry is an extension of classical
Riemannian Geometry for complex Manifold : Positive Definite Riemannian Form that defines Kähler metric is given
by :
Kähler condition : There exist locally a Kähler Potential function, (and Pluri-harmonic equivalents) such that :
Erich Kähler has prooved that Ricci tensor is then given by :
And scalar curvature :
n
ji
jiji dzdzgds
1,
2 .2
jiji zzg
2
ji
lkji zz
gR
detlog2
n
lklk
lk RgR1,
.
Air Systems Division105
Complex Autoregressive ModelRadar Signal Model : Multivariate Gaussian Complex Circular Model :
Radar Model :
Complex Autoregressive Model :
Link with Issai Schur’s algorithm (1875-1941) [Alpay] D. Alpay, « Algorithme de Schur, espaces à noyau reproduisant et théorie des
systèmes », Panoramas et synthèse, n°6, Société Mathématique de France, 1998
nnnnnnn
RRTrn
nnn
RREmZmZR
eRRZp nn
ˆ and .ˆwith
..)()/(1.ˆ1
ki
kkkkT
nn
nnn
n
eiyxzzzZ
ZZER
m
with
Positive DefiniteHermitian Toeplitz
processmean zero 0
1
TNN
NNkknn
N
knkn
Nkn aaAbbEbzaz )()(
12
0,*
1
)( ... and with
Air Systems Division106
Regularized Burg Algorithm for Robust CAR estimation Regularized Burg Algorithm (THALES Patent : F. Barbaresco, « Procédé et
dispositif de détermination du spectre de fréquence d’un signal », brevet n° 95 06983, Juin 1995)
)(.)1()( )1(.)()(
11 , .
1
).()2.( with ..2)1()(1
...2)1().(2 to1For : (n) tep .
1
)(.1ech.) nb. : (N 1 , )()()(f
:tion Initialisa .
1*
1
11
)(
)*1()1()(
)(0
221
)(
1
1
0
2)1()(21
21
1
1
1
)1()1()(*11
)0(0
1
20
00
kfkbkbkbkfkf
a
,...,n-k=aaa
a
nkakbkf
nN
aakbkfnN
MnSa
kzN
P
,...,N k=kzkbk
nnnn
nnnn
nn
n
nknn
nk
nk
n
nkN
nk
n
k
nk
nknn
N
nk
n
k
nkn
nk
nknn
n
N
k
Air Systems Division
Block Structure of Covariance Matrix and Partial Iwasawa Decomposition
Air Systems Division108
Partial Iwasawa Decomposition
Partial Iwasawa decomposition : the components of a positive definite or semi-definite matrix in the Iwasawa coordinates : Iwasawa, K., « On some types of topological groups », Ann. Math. Vol.
50, n°3, pp. 507–558, 1949 Every n×n positive definite matrix G can be uniquely expressed
using its Iwasawa components as follows.
where W & V are HPD matrices of size k×k and m×m respectively By computing the matrix multiplications in previous equation, we
derive the following parametrization of positive semidefinite matrices :
ABBBAIXI
VW
Gm
k
where
000
VWXXWXWXW
G
Air Systems Division109
Partial Iwasawa Decomposition
by using the Cholesky decomposition of , where A is a lower triangular matrix with non-negative diagonal elements, we can establish equivalence by defining B to be the matrix that satisfies the equation .
We derive the following parametrization for the Gram matrix :
If W is positive definite, then the Cholesky factor A is unique,is also unique, V=0 and therefore, the parametrization
in is unique
BA
BA
BBBAABAA
VWXXWXWXW
G
AAW
WXBA
AXB
Air Systems Division110
Autoregressive Model & Structured Covariance Matrix
Covariance Matrix and its inverse can be expressed by mean of complex autoregressive model : Block Structure of Covariance Matrix :
1111111
1111
....
nnnnnn
nnnn AARA
AR
111
1111111
....
nnn
nnnnnnn RAR
RAARAR
*)()(111
121 .
00100100
where1
.0
and .1with VVAA
A nn
nnnnn
Air Systems Division111
Non-Symmetric Square Root of Siegel Group If we consider Cholesky decomposition of covariance matrix :
Cholesky decomposition (Goldberg inversion algorithm) :
All distribution of n-dimensionnal variable is associated with Affine Group. It is the element such that its action on vector
Is transformed to random vector :
This representation of Affine Group elements could be considered as non symmetric square root of Siegel Group element :
2/11
2/1112/1
11
2n
1111
121
. and 01
1with
.
1.1..
nnnnn
n
nnnn
nnnnnnn
AW
AAAA
WWR
),0(~ nn INZ),(~ nnn ANX
XAZZA nnnn
1.
11.
01
12/11
2/111
1111
1
.1
nnnn
n
AAAA
Air Systems Division112
Block matrix structure & extended eigen-values Matrix or that is
used for Rao Metric computation has same block structure
Invariance of Block structure :
This block structure to compute extended eigen-values :
2/1)1(1)2(2/1)1( .. nnn RRR 2/1)1()2(2/1)1( .. nnn RRR
)1(1
)2(1
1-n)1(1
)2(1
2/1)1(1
)1(11
111111
1112/1)1(1)2(2/1)1(
and ..with
....
..
n
nnnnnn-
nnnnnn
nnnnnnn
AARW
WWWW
RRR
111
1)(
11)()(
1,
)(
1
1)()1(
2)1(1)(
11)()()(
.....1
0.
..
nnnn
knnn
kn
k
nk
n
in
kn
i
ninn
knnn
kn
kn
WUIUXX
XWF
Air Systems Division113
We can deduce an algorithm that could be parallelized according to CAR model order for each extended eigen-values :
111
1)(
11)()(
1,
)(
1
1)()1(
2)1(1)(
11)()()(
.....1
0.
..
nnnn
knnn
kn
k
nk
n
in
kn
i
ninn
knnn
kn
kn
WUIUXX
XWF
..
.1 )(1,
)(2/1
1
12)()1(
2)1(12)(
)(1,
)(
nk
nk
n
in
kn
i
ninn
knk
nk
XXXW
XX
1
12)1(
2)1(1
)1(
12)(1,
)(
1)(
1..
.1.
n
kn
k
nkn
nk
nnk
nk
nn XW
X
F
nnnnn
nn
nn
n Tr
)(1
)1(1
)(2
)(1
)1(1
)( ...0
1111
11.
nnnnn WW
TrTr
)()4( F
stricly increasing curve on each interval :
)3(1
)3(2
)3(3
4Tr
)4(1
Renormalisation at each iteration :
avec
)4(3
)4(2
)4(4
.A
1
1
n
1=k)(
1,
)(1)()(
1-n
1
1)()(
nk
nk
nk
n
n
k
nk
n
XXF
F
Interpretation in term of projection of CAR vector :
0
Block matrix structure & extended eigen-values
Air Systems Division
Information Geometry Metric for Complex Autoregressive Model
Air Systems Division115
Kähler metric for Complex Autoregressive Model In the framework of Affine Information Geometry, Kähler
metric is given by Hessian of Entropy, considered as Kähler Potential Function :
Entropy of Multivariate Gaussian Law of Zero Mean :
Entropy can be parametrized by mean of reflection coefficients
enRRΦ logdetlog~
-RHΗΗΦg
jiij
et ~2
0
1
1
2 ..ln.1ln).(~ Penkn)(RΦn
kkn
n
kk
nnn
zn
P1
20
10
11
21
1 avec
.1
1
1
20
1
0
1 1detn
k
kn
kn
n
kknR
nn
nnnnn
RRTrn
nnn
RRE
mZmZR
eRRZp nn
ˆet
.ˆ avec
..)()/(1.ˆ1
Air Systems Division116
Erich Kähler…
Entropy U=-logdet[R] of complex autoregressive model can be considered as a Kähler Potential function. If we use reflection coefficient for Entropy parametrization, we recover metric initially proposed by Erich Kähler.
« Kähler Erich, Mathematical Works », Edited by R. Berndt and O. Riemenschneider, Berlin, Walter de Gruyter, ix, 2003
Seminal Erich Kähler Paper, 1932, Hambourg
Air Systems Division117
Kähler Metric : hyper-Abelian Case We recover Kähler metric proposed in his seminal paper
of 1932 : Kähler called this case, « hyper-Abelian » :
Other metric proposed by E. Kähler, called « hyper-fuchian » :
),(ln1ln.1
1
2 zzKzΦ D
n
kkk
,...1 1 / and
1),(kernelBregman
with1
1
2
nkzz
zzz : K
k
n
kkD
k
1 / with 1ln.1k
2
1
2n
k
n
kk zzzΦ
Air Systems Division118
We define a « Doppler » metric in case of complex autoregressive case as Hessian of Kähler Potential function, where this potential is given by entropy in framework of affine Information geometry :
Kähler potential parametrized by :
Metric can be computed according to reflection coefficients :
10
1
1
2 ..ln.1ln).(~
enkn)(RΦn
kkn
Tnn
nTn
n P )()(1110
)(
20
2011
nPng 221
).(
i
ijij
ing
1
122
22
0
02
1)(.
n
ii
in
din
PdPnds
Kähler metric for Complex Autoregressive Model
TnP 110
Air Systems Division119
Scalar Curvature of Complex Autoregressive Model
We use Ricci Tensor given by Erich Kähler in case of complex Manifold :
In Kähler Geometry, Ricci Tensor is given by :
In case of complex autoregressive model, we have :
We can compute negative scalar curvature :
ji
lkji zz
gR
detlog2
1,...,2for 1
2
12
22
20
11
nkR
PR
k
kllk
lk
lklk RgR
,.
n
n
j jnR
1
0 )(1.2
Air Systems Division120
Autoregressive Model # Kähler-Einstein Metric The metric is not of Kähler-Einstein kind, but a close
« matricial » structure : A metric is called Kähler-Einstein metric if its Ricci tensor is
proportional to the metric :
In case of Kähler-Einstein metric, Kähler potential is solution of Monge-Ampère equation :
For complex autoregressive model, we have :
,....,2 where
)(1.2 with
1)(
1
0
)()(
indiagB
jnBTrRgBR
n
n
j
nij
nij
ji
ji
lklkji zz
kzz
gkgkR
2
0
2
00 .detlog
constant : with .
function cholomorphiPotentialKähler
with )det( 02
ψ : Φ :
eg klk
Air Systems Division121
Median Algorithm on Reflection coefficients 1/2We have , where is a n uplet of reflection coefficients for all order and with signal power . We compute first , in R (classical method on scalar median).We compute then median of complex reflection coefficients : First, reflection coefficients are transformed from unit disk to upper half space with :
where Geodesic are circle orthogonal to real axis
1)()(0
)(1
)(1
)(0 ),(),...,,(
nkkkn
kkk DRPP N ,...1
HDC :1
)(
)()()(
11
kj
kjk
jk
j iz
)( )(1)( kk Cz
)))log(),...,log(),(log(exp( )(0
)2(0
)1(00
Nmedian PPPmedianeP
)(1
)(1
)( ,, kn
kk
)(0
kP
c
Air Systems Division122
Median Algorithm on Reflection coefficients 2/2We initiate algorithm with arbitrary point Z in H :
At step p, displacement is driven by gradient given by:
where Z is estimated and ck is circle origin of geodesic between Zand zk.
Integer p, initiated at 1, decreases if difference between to steps is lower than 1/p.After convergence, we compute value in unit disk :
N
kk
lp
l
kl
plk
lp
lplZ pcZ
cZzZsigneit1
)()(
)()()()()(
, /1)(
)Re(2 )()(
2)(2)()(
pl
kl
pl
klk
l ZzZz
c
TnZZZ )0(1
)0(1
)0(
with
)Im()Re()Im(
)Re( )()(
,
)(,)(
)(p
lplZ
plZp
lZZ
tt
Zc pl
)arg( )()()(
plZ
pl
pl cZ
pcZ
cZcZsigne
MkicZcZ
pl
pl
pl
pl
pl
Zp
l
Zp
l
Zp
lp
lZp
lZp
l/1
))(Re(2
1exp)(
)(
)()()()(
)()()()()1(
iZiZ)C(Zμ
PPPmedianePZCP
convl
convlconv
lmedianl
Nmedian
)(
)()(
)(0
)2(0
)1(00 )))log(),...,log(),(log(exp(
with ))(,(
Air Systems Division123
Gradient Flow in Unit Disk
)( plZC
)1( plZC
)1()1(ll zC
)2()2(ll zC
)3()3(ll zC )4()4(
ll zC
Air Systems Division124
Results on Reflection Coefficients
Mean in blue
Median in Black
Air Systems Division125
« Doppler Matrix CFAR» by median autoregressive model
FFTI&Q
log|.|
Scalar CFAR
>S
log|.|
Scalar CFAR
>S
OU
I&QRegularized
AR
Model
-
-
Robust
Siegel
Distance
Median CFAR on
>S
CfenêtreTFAiimiiP
,,1,0 ,...,,
2
1*
,,
,,
2
,0
,0,,1,0,,1,0 ].[1
arglog,...,,,,...,,
m
n mediannjn
mediannjn
j
medianmedianmmedianmedianimii μ
thn)(mP
PmPPdist
Classical
Chain
New Chain
Air Systems Division126
Iterated median in Poincaré Disk for reflection coefficient
Disk isometries (conforme transformation : angles preservation) :
Algorithm :*
,
,*, )(
)(1
)(wzewez
zzwwzez
wi
iwi
w
*nmedian,n
nmedian,nmedian,n
*nk,n
nk,nk,n
l,n
m
lkk k,n
k,nnn
n
mm,,median,
wμwμ
μ
.wμwμ
hen μ,...,m tFor k
εμl/ with μμ
γw
as w n as longIterate on,...,μμ,...,μμ and μ
Init :
1
11
00
1
1
1
10010
n ofonOptimisati
Air Systems Division127
Iterated median in Poincaré Disk for reflection coefficient
12
3
0,median
30,3
10,1 20,2
00, median Problem is simplified : we start from All geodesics are radials and space is quasi-euclidean close to 0
Initialisation
00, median
Classical Approach of Median Flow New Paradigm of Dual Median Flow
Fixed
Points
Point Driven
By median flow
Points Driven
By Dual Median Flow
Fixed point
Air Systems Division128
Iterated median in Poincaré Disk for reflection coefficient
n,3
n,1n,2
nmedian,
Classical Median Flow
1,3 n
1,1 n
Dual Median Flow
εμl/ avec μμ
γw l,n
m
lkk k,n
k,nnn
1
nnmedian w1,
*nk,n
nk,nk,n .wμ
wμμ
11
1,2 n
*nmedian,n
nmedian,nmedian,n wμ
wμμ
11
Air Systems Division129
Regularized Burg Algorithm
)(.)1()( )1(.)()(
11 , .
1
).()2.( with ..2)1()(1
...2)1().(2 to1For : (n) .
1
)(.1ech.) nb. : (N 1 , )()()(f
:Init .
1*
1
11
)(
)*1()1()(
)(0
221
)(
1
1
0
2)1()(21
21
1
1
1
)1()1()(*11
)0(0
1
20
00
kfkbkbkbkfkf
a
,...,n-k=aaa
a
nkakbkf
nN
aakbkfnN
MnIteratea
kzN
P
,...,N k=kzkbk
nnnn
nnnn
nn
n
nknn
nk
nk
n
nkN
nk
n
k
nk
nknn
N
nk
n
k
nkn
nk
nknn
n
N
k
Air Systems Division130
Iterated Mean in Poincaré Disk Algorithm :
*nmedian,n
nmedian,nmedian,n
*nk,n
nk,nk,n
l,n
m
lkk k,n
k,nk,nnn
n
mm,,median,
wμwμ
μ
.wμwμ
hen μ,...,m tfor k
εμl/ with μμ
μarcthγw
n until wIterate on,...,μμ,...,μμ et μ
tion :Initialisa
1
11
.
00
1
1
1
10010
n Optimise
Air Systems Division131
Iterated Computation of median in Poincaré disk
Air Systems Division132
Doppler smoothing by diffusion
Air Systems Division133
Figures of Merit
COR curve
Red : Matric Doppler CFAR based on CAR median
Black : Classical FFT
Gain of
New processing
chain
Air Systems Division134
Figure of Merit for closed targets
COR Curves
Red : Matrix Median CFAR
Black : Classical FFT
Gain of
New processing
Chain
Air Systems Division135
Results
Normal conditions
Air Systems Division135
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
taux
de
cibl
es d
étec
tés
taux de faus s es alarmes
algorithme médianFft
Air Systems Division136
Results
Closely separated targets
Air Systems Division136
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
taux
de
cibl
es d
étec
tés
taux de faus s es alarmes
algorithme médianFft
Air Systems Division137
Median versus Mean
Air Systems Division137
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.7
0.75
0.8
0.85
0.9
0.95
taux de faus s es alarmes
taux
de
déte
ctio
n
médianmoyenne
Air Systems Division138
Result
Traget close to clutter edges
Air Systems Division138
0 0.01 0.02 0.03 0.04 0.05
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
taux
de
cibl
es d
étec
tés
taux de faus s es alarmes
algorithme médianFft
Air Systems Division139
Effet diffusion-1
Air Systems Division139
0 0.01 0.02 0.03 0.04 0.05 0.06
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
taux de faus s es alarmes
taux
de
déte
ctio
n
s ans diffus ionavec diffus ion
Pour l’algorithme médian:
Air Systems Division140
Médian du coefficient de réflexion
COR curbes Cyan : Median in Poiincaré Disk Red : Median in Poincaré Half-Plane Blue : Mean in Poincaré Disk Yellow : Mean in Poincaré half-Plane Black : FFT
0 0.5 1 1.5 2 2.5 3
x 10-3
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
taux
de
cibl
es d
étec
tés
taux de faus s es alarmes
Air Systems Division141
Médian du coefficient de réflexion
COR Curves (closely separated targets) Cyan : Median in Poiincaré Disk Red : Median in Poincaré Half-Plane Blue : Mean in Poincaré Disk Yellow : Mean in Poincaré half-Plane Black : FFT
0 1 2 3 4 5 6 7 8 9 10
x 10-3
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
taux
de
cibl
es d
étec
tés
taux de faus s es alarmes
Air Systems Division
Radar STAP Processing
Air Systems Division143
STAP (Space-Time Adaptive Processing) Principle
STAP Principle
ttoutput ZSVSZwZ 2/12/1ˆ
Nml
ml
ml
ml
Ml
l
l
lt
z
zz
z
z
zz
Z2
1
2
1
, with
gate range lpulse m
element n
th
th
th
nmlz
jammerclutternoise
tt
t
RRRRVZVZER
nVZ
,
Air Systems Division144
Secondary Data Covariance Matrix Estimation
Sample Data & CFAR Detector
ˆ 1VSw
1
ˆ011
2
tt
t
ZSZVSVZw
ttoutput ZSVSZwZ 2/12/1ˆ
M
kkZkZRMS
1)()(ˆ.
Air Systems Division145
Classical STAP Equation : Likelihood Ratio Test Joint Probability Density Function of primary & 2nd data set
set dataprimary fromVector Input Single :set Data 2nd fromVector Input Single :)(
tZkZ
M
ktt
M
ktt
i
MRRTr
Nti
M
ktt
kZkZVZVZM
R
i
kZkZZZM
R
i
R
kZkZEk , Re(R)π
,Z(M)),,Z(Zp
kZpZpMZ),Z(Zp
i
11
10
11
)()(1
1
: hypothesis noise-plus-signal , 1
)()(1
1
: hypothesis alone noise , 0
)()(with det11
)()(,,1
1
Air Systems Division146
Classical STAP Equation : Likelihood Ratio Test Likelihood ratio test : The PDF of the inputs is maximized over all unknown
parameters separately for each of the two hypothesis The ratio of these maxima is the detection statistic The maximizing parameter values are the maxium likelihood
(ML) estimators of these parameters : the maximized PDFs are obtained by replacing the unknown parameters by their ML estimators The ML estimator of a covariance matrix is equal to the sample
covariance matrix
NRRTrRR
Re,Z(M)),,Z(ZpMax
iiML
iNtiR
1 ˆwith
det11
Air Systems Division147
Maximum Likelihood Hermitian covariance matrix N.R. Goodman’s Theorem : Consider M independent
identically distributed N-variate complex Gaussian random variable , k=1,…,M as a sample of size N from a population with PDF . Let DH be the set of NxN Hermitian positive definite matrices. Over the domain DH the maximum likelihood estimator of the Hermitian covariance matrix is :
Proof :
)()(1ˆ
1kZkZ
MR
M
kZ
ZRZp /)(kZ
ZR̂ZR
M
kZZ
kZRkZ
MZ
NM
M
kZZ
kZRkZRMNML
eR
RkZpRMZZpL
M
kZ
1
1
)()(
1
)()(detlog)log(ln
det1/)(/)(),...,1( 1
1
Air Systems Division148
Classical STAP Equation : Likelihood Ratio Test
Gerelarizd Likelihood Ratio Test :
We consider : (M times the sample covariance matrix of the secondary
data) that satisfy Wishart distribution
0
1
0
1
0
0
1
detdet
detdet
)(,),1(,
)(,),1(,
RMinR
RRMax
MZZZpMax
MZZZpMaxMax
tR
tR
M
kkZkZS
1)()(
VZSVZSRM
ZSZSRMZZSRM
ttN
ttN
tt
11
100
1detdet1
1detdet1)1(
Air Systems Division149
Classical STAP Equation : Likelihood Ratio Test
Minimisation over :
Likelihood ratio is given by :
We put :
VSVZSV
VSVZSVVSVZSZ
ZSZVSVZSZVZSVZ
tttt
tttttt
1
212
1
111
11211
.
.Re2
1
1detdet
1
211
1
1
0
VSVZSV
ZSZ
ZSZRRMax
ttt
tt
10 with 1
11
1
00
000
11
21
η
ιZSZVSV
ZSV
tt
t
Air Systems Division150
Classical STAP Equation : Likelihood Ratio Test
GLRT : Generalized Likelihood Ratio Test :
.ˆ11ˆ
ˆ
1ˆ
011
21
MZRZ
MVRV
ZRV
SM
R
tt
t
.ˆ11ˆ
ˆ
ˆˆ
011
2
1
MZRZ
MVRV
Zw
VRw
tt
t
Air Systems Division151
Classical STAP Equation : AMF
AMF (Adaptive Matched Filter) Test :
1ˆ with ˆ
ˆ1
21
SM
RVRV
ZRV t
VRVZRV
VRVZRV
VRVZRV
Λ
VRVZRVΛMax
VRVZRVΛeΛ
,Z(M)),,Z(Zp
,Z(M)),,Z(ZpMax
,Z(M)),,Z(Zp,Z(M)),,Z(ZpMaxΛ
ttt
t
α
tZRZαVZRαVZ
t
tα
t
t
tttt
1
21
1
21
1
21
1
1
121*
0
1
0
1
.2log
logargˆ
Re2log
1
1
11
11
Air Systems Division152
Classical STAP Equation : Likelihood Ratio Test
GLRT : Generalized Likelihood Ratio Test :
Adaptive Matched Filter (AMF) :
1
ˆ
)()(ˆ. with ˆ
011
21
1
tt
t
M
k
ZSZVSVZw
kZkZRMSVSw
ˆˆ
ˆ
)()(ˆ. with ˆ
01
21
1
wSwZw
kZkZRMSVSw
t
M
k
• CFAR Property
• better Pd at low SINR
• CFAR Property
• More robust to sidelobes targets
ttoutput ZSVSZwZ 2/12/1ˆ
Air Systems Division153
Optimum STAP interpretation
Optimum STAP :
Space-Time Steering Vector :
filter Matched : SZ
NoiseceInterferen whitten :
ˆ
ˆ
1/2
2/1
2/12/1
2/12/11
yV
ZSy
ZSVSZwZ
VSSVSw
output
t
ttoutput
)1(2
2)1(2
)1(2
22
)1(2
2
1
1
1
,
s
t
s
s
Nj
jNj
Nj
jj
Nj
j
e
ee
e
ee
e
e
V
mkkm
NNNN
N
N
zzEQ
QQQ
QQQQQQ
ZZER
tttt
t
t
with
21
22221
11211
Air Systems Division
Parametric Adaptive Matched Filter
Air Systems Division155
Parametric Adaptive Matched Filter (PAMF) Parametric Adaptive Matched Filter (PAMF) methodology for STAP
and detection : This methodology is based on approximating the interference spectrum
with a multichannel autoregressive (AR) model of low order. a low-order AR model provides an accurate representation of
simulated and measured interference for a wide variety of system and scenario conditions leads to reduced computational requirements
The modeling fidelity is attained using a small fraction of the Reed-Brennan rule training data set, thus presenting reduced secondary data requirements
the method offers dramatic improvement in detection performance over the conventional adaptive matched filter (AMF)
the method offers dramatic improvement in detection performance over the conventional adaptive matched filter (AMF)
the PAMF provides significantly improved detection performance over the AMF using only a small fraction of the secondary data required bythe AMF
J. R. Román, M. Rangaswamy, D. W. Davis, Q. Zhang, B. Himed,and J. H. Michels., « Parameteric adaptive matched filter forairborne radar applications. IEEE Transactions Aerospace Electronic Systems, 36(2) :677–692, April 2000
Air Systems Division156
Parametric Adaptive Matched Filter (PAMF)
Multichannel autoregressive (AR) model : Nuttall, A., « Multivariate linear predictive spectral analysis
employing weighted forward and backward averaging: A generalization of Burg’s algorithm. », Technical report TR-5501, Naval Underwater Systems Center, New London, CT, 1976 Strand, O. N., « Multichannel complex maximum entropy (auto-
regressive) spectral analysis », IEEE Transactions on Automatic Control, AC-22, n°4, pp 634-640, 1977
Rami KANHOUCHE, « Méthodes mathématiques en traitement du signal pour l’estimation spectrale », thèse de doctorat de l’ENS Cachan, 21 Décembre 2006 (tel-00136093, version 1 - 12 Mars 2007)
Air Systems Division157
Parametric Adaptive Matched Filter (PAMF)
MF and AMF :
LDU decomposition :
M
k
tAMF kZkZRMS η
VSVZSV
Λ1
01
21
)()(ˆ. with
ηVRVZRV
Λ tMF
1
21
2
2/12/1
22/12/12/12/1 and
VRRVZRRV
ZRVR tt
diagramblock main thealong 10 matricesHermitian matrix diagonal-block a :
diagonalblock main thealong matrixIdentity JxJh matrix wit triangular-blockLower :
with
x
x
x
),...,N-(iCDCD
CA
ADAR
JJi
JNJN
JNJN
Air Systems Division158
Parametric Adaptive Matched Filter (PAMF) the matrix coefficients of the nth-order multichannel
forward linear predictor for the process:
JNNNN
JNNN
J
J
J
IANANANAINANANA
IAAIA
I
A
)1()3()2()1(0)4()3()2(
00)1()2(000)1(0000
1111
222
22
1
1
1
12
11
1
0000000000
ND
DD
D
)1(),...,1(),( nnn AnAnA
Air Systems Division159
Parametric Adaptive Matched Filter (PAMF) MF based on LDU decomposition :
the multichannel element vectors are given by :
12/12/111 ADDAR VADVAD
ZADVADΛ
t
MF 12/112/1
12/112/1
tMF ZA
VAu
uDuD
DuDΛ
1
1
2/12/1
2/12/1
with
nIAknZkADnDn
NnknZkAn
Jnn
ktnnn
n
ktn
)0( with
)()()()(
1,...,1 , )()()(
0
2/12/1
0
TT
tTt
Ttt
Jtt
NZZZZ
NnCnZZ
)1()1()0(
data theof series time1,...,1,0/)(:
Air Systems Division160
Parametric Adaptive Matched Filter (PAMF) the multichannel element vectors are given by :
Parametric Matched Filter by approximation to the MF with a simplified structure : retain only the vector sequence for the filter of order P (1P N-1) :
filter ingock whitenspatial)bl(or ation transform whiteningspatial a is step This instant. each timeat
dimension spatial thealong elements eduncorrelat generates :
matrix Identity theismatrix covariance its : )(matrix covarianceor with error vect predictionnth theis : )(
2/1
n
J
n
D
InDn
covarianceerror predictor theand tscoefficien
predictionlinear order -Pth : ,1,...,1,0),(
1,...,1,0 and )0(with
)()()()(0
2/12/1
P
J
P
ktPP
DPNkkA
PNnIA
PknZkADnDn
Air Systems Division161
Parametric Adaptive Matched Filter (PAMF) PMF & PAMF schemes:
1,...,1,0 and )0(with
)()()()(
)()()()(
0
2/12/1
0
2/12/1
PNnIA
PknVkADnuDn
PknZkADnDn
J
P
kPP
P
ktPP
1
0
21
0
)()(
)()(
PN
n
PN
nPMF
nn
nnΛ
The inverse of an AR model is an MA model : the MA whitening
filter is implemented as a multichannel tapped delay line
Air Systems Division162
Parametric Adaptive Matched Filter (PAMF) Multivariate Autoregressive Model and Identification
Algorithms : A complex-valued Gaussian-distributed zero-mean stationary,
vector random process is a multichannel AR process of order P, denoted as AR(P), if it satisfies a relation of the form :
For stability, all the system poles must lie inside the unit circle in the complex plane. The multivariable poles of an AR(P) system are the
multivariable zeros of its inverse MA(P) system.
)(nd
sequence noise white: ,0~)()0( and ,parametersmatrix JxJ
luedcomplex va constant, : ,...,2,1/)(
filter whitening: )()()(0
P
J
P
k
DCNnIA
PkkA
kndkAn
Air Systems Division
Multivariate Burg AlgorithmENS Cachan PhD thesis
Air Systems Division164
Multivariate Burg Algorithm MV-Burg Algorithm :
)()(
)()(
)1()(
with
)()()()()1()(2
2
)()()( with )()1()(
)1()()(
and
0010
01100
with )()()(
)()()(
,,...,,0,,...,,,...,,
1
111
11
1*11
*
00*111
111
0
0
*
0
*11
*11
*11
11
12
1121
11
kekeEP
kekeEP
kekeEP
kekekekekekeA
JJPPJJPPAJJPPTrMin
kZkkkJJAkk
kAkk
IAJlnkJZkJAk
lkZkAk
IJJAJJAJJAAAAAAAA
bn
bn
bn
fn
fn
fn
bn
fn
fbn
nN
k
bn
bn
nN
k
fn
fn
nN
k
bn
fn
nn
fn
bn
fbTn
fbn
nn
bn
fn
A
bff
nnn
bn
fn
bn
nn
fn
fn
nn
l
nl
bn
n
l
nl
fn
nn
nn
nn
nn
nn
nnnn
nn
nn
Siegelnn
nn
nn DiskAIAA
11
*11
11
Air Systems Division165
Multivariate Burg Algorithm Open Question : How to estimate Robust Multivariate
Burg Estimation in case in Inhomogeneous area : Regularization of Multivariate Burg (avoid to estimate AR order) Median MV-AR Model estimation by Fréchet Barycenter
See explanation done in Doppler Processing Case
Air Systems Division
Sample Covariance Matrix Estimation by Fréchet-Karcher Barycenter
Air Systems Division167
Inhomogeneous Clutter
Sample Covariance Matrix estimation in inhomogeneous clutter will be the first drawback of these methods due to :Clutter Amplitude variationDoppler Spectral broadening variationClutter EdgesTargets in secondary data
Air Systems Division168
Fréchet-Karcher exponential Barycenter Fréchet définition of Expectation as center of mass or
barycenter :
Karcher’s result :
Karcher’s flow :
)()(,
argˆ
2 dPRTdTG
TGMinRER
Z
TZFréchetFréchetZ
)( , of space tangent : with
)()(exp 1
Z
ZT
RT
TdPRTgradG
)()0(with )(.exp)(
TgradGTgradGtt T
Air Systems Division169
Fréchet-Karcher Barycenter : Discrete Case Fréchet définition of Expectation as center of mass or
barycenter :
Karcher’s result :
Karcher’s flow :
M
k
TZFréchetFréchetZ
kZkZTdTG
TGMinRER
1
2 )()(,
argˆ
)()( , of space tangent : with
)()(exp1
1
kZkZT
TkZkZTgradGM
kT
2/1)()(log
2/11
1
2/12/1
)()0( with )(.exp)(
n
TkZkZTt
nn
T
TeTT
TgradGTgradGttM
knn
M
kkZkZ
MT
1)()(1ˆ
Air Systems Division
Iterated Median in Siegel Disk
Air Systems Division171
Mostow Decomposition Mostow Theorem :
Every matrix of can be decomposed :
whereis unitaryis real antisymmetricis real symmetrix
Can be deduce from
Lemma : Let and two positive definite hermitian matrices, there exist a unique positive definite hermitian matrix such that : Corollary : if is Hermitian Positive Definite, there exist
a unique real symmetric matrix such that :
M CnGL ,SiAeUeM
UAS
A B
X BXAX M
SSS eMeM 1*
Air Systems Division172
Mostow Decomposition Lemma : Let and two positive definite hermitian
matrices, there exist a unique positive definite hermitian matrix such that :
Proof :is unique hermitian positive definite square root of
Observe that is geodesic center of and for symmetric space of hermitian positive definite matrices
A B
X BXAX
2/12/12/12/12/1
2/12/12/12/12/1
2/12/122/12/1
2/12/12/12/12/12/1
ABAAAX
BAAXAA
BAAXAA
BAAXAAAXABXAX
2/1A A
X 1A B
Air Systems Division173
Mostow Decomposition Corollary : if is Hermitian Positive Definite, there exist
a unique real symmetric matrix such that :
Proof :Positive Definite Hermitian Matrix, and with
same property. From previous Lemma, there exist a unique hermitian positive definite matrix such that :
Exponential providing an homeomorphism between symmetric et positive definite symmetric spaces, we have to proof that is positive definite
MSSS eMeM 1*
M *M 1M
XXXMM 1*
X
XXXXMMXMXM
MXXMXMXMM
*1*
*1**
1*1*1**1****
because
Air Systems Division174
MMPPPPPPS
PPPPPeePeP SSS
avec log.2/1
:corrolary Lemma2/12/12/1*2/12/1
2/12/12/1*2/12/1212*
Mostow Decomposition Mostow Theorem :
All matrix of can be decomposed in :
is unitary, is real antisymmetric réelle and is real symmetric
Proof :
M CnGL ,SiAeUeM
U A S
SS
SSiASSSiAS
SiASSiA
ePePeeeeeeeeP
eeeMMPeUeM
212*
2222*
2
MMPPeei
A
Peee
SS
SSiA
avec log21
:y injectivit lleexponentie 2
iASeMeU
Air Systems Division175
Mostow Decomposition Lemma :
Matrix is Hermitian Positive Definite and verify .There exist one particular real antisymmetric
such that : Proof :
SS PeeY IYY * A
iAeY 2
IeePeePeeYYePePePePeeYY
SSSSSSSS
SSSS
212*212*
**
But
0et 2
/**1 *
SiASHHHeYYe
eYHHDPYHH
H
Air Systems Division176
Siegel Isometry To find an isometry of Siegel Disk such that
, We determine an isometry of Siegel half-plane such that , and we use Cayley transform :
Isometric Automorphism of Siegel Half-Plane :
Isometric Automorphism of Siegel Disk :
nSD)(XfW
0)( WfW)ˆ(ˆ
ˆ XfW
nSH iIWfW )ˆ(ˆˆ
)XΦ(X ˆ 1ˆˆˆ
iIXiIX)XΦ(X
iIWfYXYY
M
LL
IXI
Mf
YLSHiYX(W)ΦW
W
W
n
)ˆ(ˆ0
00
0 matrice de ˆation transformLa
et ˆ
ˆ2/1
2/12/1
1
ˆ
2/11
0 avec
matrice de ation transformLa 1
Wf
iIIiII
C
CMCf
W
-W
1)(
DCZBAZZf
DCBA
M
Air Systems Division177
Iterated Computation of Median in Siegel disk
median,nGmedian,n
k,nGk,n
Fnk
m
lkk
nknn
nknknknknknknknknk
S
nk
SS
nkiA
nkk,nSiA
nknk
Fn
mm,,median,
iii
m
WfW
Wf W,...,m k
εHl/ HγG
WWPPPPPPS
ith eWeeWeUHeeUW
Gn
,...,WW,...,WW et Wtion :Initialisa
iIZiIZW,...,mi
,...,ZZ
n
n
nknknknknknk
11
1
,1
,
,,,2/1
,2/12/1
,*,
2/1,
2/1,,
2,
2,,,,
10010
11
then1For
with
with log.2/1
: w
until on Iterate
0
: 1For
Plane-Half Siegelin
,,,,,,
Air Systems Division
Metric for SO(n) Groupmean, median and diffusion
Air Systems Division179
Pulse Compression in Radar : LFM
BT
BTR sin)(
Air Systems Division180
Low Signal to noise ratio
Matched Filter for LFM
SignalWithoutNoise
After MatchedFilter
WithNoise
Remenber :With non Matched filter
Air Systems Division181
Phase Code : Barker code
Barker code of length 13 is longer code, composé with only 0 ou de 1, that maximises, by autocorrélation, the ratio main peak versus secundary peak. This code is made of 13 digit :
[1,1,1,1,1,0,0,1,1,0,1,0,1].
Air Systems Division181
2 4 6 8 10 12-2
-1.5
-1
-0.5
0
0.5
1
1.5
-800 -600 -400 -200 0 200 400 600 800-200
0
200
400
600
800
1000
1200
1400
1600
1800
Air Systems Division182
Distance & geodesic on SO(n) Compact Group SO(n) :
Distance between SO(n) matrices :
Geodesic between SO(n) matrices :
Tangent of geodesic :
Constraints on Karcher barycentre :
XXtraceXRRRRd TFF
T avec log, 2121
21log1211)( RRttT T
eRRRRt
1detet /)()( RIRRnGLRnSO T
2110
log)( RRRdt
td T
t
0log0log)(111 0
N
kk
TN
kk
TN
k t
k RRRRRdt
td
k
Air Systems Division183
Barycentre and median on SO(n) Karcher Barycentre computation :
Median Computation
N
kk
Tn RX
nnn eXX 1
log1
1
FkTnn
RXRX
n
nn RXkSeXX nSk FkTn
kTn
log/ with loglog1
1
Air Systems Division184
Diffusion Heat Fourier Equation on matrices graph Diffusion Fourier heat equation on 1D graph of scalar values In a normed vectorial space in 1D case, diffusion equation is
approximated by discret Laplacian :
with arithmetic mean : Discret Fourier heat equation can be written :
Based on analogy with Karcher Flow, diffusion equation on 1D graph of Symmetric or Hermitian matrices can be written with geometric mean between :
nnnnnn uu
xxuu
xuu
xtu
xu
tu
ˆ21
211
2
2
2/ˆ 11 nnn uuu
2,,,,2,12 with ˆ.).1(ˆ2
xtuuuu
xtuu tntntntntnn,t
J. Fourier
/21
,1,1,1,2
,,,
ˆlog2
,1,
ˆet x
t2 avec
ˆ,,2
tnT
tntntn
tnT
tntn
XXx
t
tntn
XXXX
XXXeXXtn
Ttn
tnX ,ˆ tntn XX ,1,1 ,
Air Systems Division185
Diffusion Heat Fourier Equation on matrices graph Diffusion Equation on a 1D graph of matrices
Geometric Details : colinearity of tangent vectors for geodesics and between , and :
As in Euclidean case weighted mean is replaced by geodesic weighted barycenter :
J. Fourier
)(1,
, tn
tn
XX )(,
,
ˆ tn
tn
XX tnX , 1, tnX tnX ,
ˆ
tnT
tntnT
tn XXx
tXX ,,21,,ˆlog2log
tn
tntnXtnX
tn
tn
tn
tn XX
XX
XX TT
dd
xt
dd
,
,1,,
,
,
1,
, ˆ
0
ˆ
2
0
//)(
.2)(
yxyx )1(.
BAABA T
/21
,1,1,1,2
,,,
ˆlog2
,1,
ˆet x
t2 avec
ˆ,,2
tnT
tntntn
tnT
tntn
XXx
t
tntn
XXXX
XXXeXXtn
Ttn
Air Systems Division186
Diffusion Heat Fourier Equation on scalar 1D graph Diffusion Equation on a 1D graph of scalar values
J. Fourier
tnu ,
tntntn uuu ,12/1,1,ˆ
tnu ,1tnu ,1
1,0.).1(
vuvu
t
t+1
range
tnu ,tnu ,1 tnu ,1
tnu ,ˆ
1n,tu
time
tntntn uuu ,,1, ˆ
1
0
2ˆ and 2with
ˆ.).1(
,1,1,2
,,1
tntntn
tntnn,t
uuu
xt
uuu
Air Systems Division187
Diffusion Heat Fourier Equation on matrices 1D graph Diffusion Equation on a 1D graph of matrices
J. Fourier
tnX ,
tntntn XXX ,12/1,1,ˆ
tnX ,1tnX ,1
t
t+1
range
tnX ,tnX ,1 tnX ,1
tnX ,ˆ
1n,tX
time
tntntn XXX ,,1,ˆ
1
0
1,0
BAABA T
/21
,1,1,1,2
,,,
ˆlog2
,1,
ˆet x
t2 avec
ˆ,,2
tnT
tntntn
tnT
tntn
XXx
t
tntn
XXXX
XXXeXXtn
Ttn
Air Systems Division188
Anisotropic Diffusion on graph of SO(n) matrices In1D case, Anisotropic diffusion is given by :
For distance , we take :
2/12,1,
,
2/12,,1
,
2/12,1,1
,2,,,
,,
,
,1,1,1
log,1,
,,,ˆlog
,1,
,1et
,1
,1,
. ,
ˆwith
ˆ
,1,1
,,
xXXd
xXXd
xXXd
xt
XXXeXX
XXXeXX
tntntn
tntntn
tntntn
tntntn
tntn
tn
tnT
tntnXX
tntn
tnT
tntnXX
tntn
tnT
tn
tnT
tn
XXTraceX
XXXXd
T
F
T
with
log, 2121
Air Systems Division189
Iterated Computation of SO(2) matrices Mean
Iterated Algorithm:
Air Systems Division189
0 0.2 0.4 0.6 0.8
-1
-0.9
-0.8
-0.7
-0.6
-0.5
Air Systems Division190
Iterated Computation of SO(2) matrices Median
Iterated Algorithm :
Air Systems Division190
0.2 0.4 0.6 0.8 1
-0.95
-0.9
-0.85
-0.8
-0.75
-0.7
-0.65
-0.6
-0.55
-0.5
Non unicity
Air Systems Division191
Diffusion
Air Systems Division191
Air Systems Division192
Results
Air Systems Division192
0 100 200 300 400 500 600 700-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 100 200 300 400 500 600 7000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 100 200 300 400 500 600 7000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0 100 200 300 400 500 600 7000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Mean: Median: Anisotropic Diffusion:
Raw signal
Air Systems Division
Extension to POLARIMETRIC DIFFUSIVE CFAR
Air Systems Division194
Unit Sphere Poincaré Model of Polarimetric Data
2;
2
4;
4
2
arctan1
2
ss
2
arctan22
21
3
sss
2sin2sin2cos2cos2cos
Im2Re2
0
0
0
0
*12
*12
22
21
22
21
3
2
1
0
ssss
zzzz
zzzz
ssss
S
23
22
21
20 ssss
Air Systems Division195
Noisy Data
Geometric Diffusion of Polarimetric Data
Beltrami Flow
Noisy Data (Blue) and Data after diffusion (Green) (,)
Air Systems Division196
Questions
« L’eau ronge les montagnes et comble les vallées. Si elle le
pouvait, elle réduirait le monde à une sphère parfaite »
Léonard de Vinci [Cod. Atl. 185 V]