Seite - 353 - in Differential Geometrical Theory of Statistics
Bild der Seite - 353 -
Text der Seite - 353 -
Entropy2016,18, 396
theRiemannianvolumemeasure. Let fˆk be the estimatordefined inEquation (7). There exists a constantCf
such that ∫
x∈M Ex1,...,xk[(f(x)− fˆk(x))2]dμ≤Cf( 1
krn +r4). (8)
If r∼ k −1n+4 , ∫
x∈M Ex1,...,xk[(f(x)− fˆk(x))2]dμ=O(k− 4
n+4). (9)
Proof. SeeAppendixA.
It canbecheckedthatontheSiegel space rinj=+∞andthat, foran isometryα,wehave:
fˆk(x,x1,...,xk)= fˆk(α(x),α(x1), ...,α(xk)).
Each locationanddirectionareprocessedassimilarlyaspossible. Thisdensityestimatorcanbeused
fordataclassificationonRiemannianmanifolds, see [26].
Inorder toobtain theexplicit expressionof theestimator,onemusthavetheexplicit expression
of the Riemannian exponential, of its inverse, and of the function θp (see Equations (6) and (7)).
Theseexpressionsaredifficultandsometimes impossible toobtain forgeneralRiemannianmanifolds.
In thecaseof theSiegel space, the symmetric structuremakes thecomputationpossible. Since the
space ishomogeneous, thecomputationcanbemadeat theorigin iI∈Hnor0∈Dn andtransported
to thewholespace. In thepresentwork, therandomvariable lays inDn.However, in the literature,
theCartanandIwasawadecompositionsareusuallygivenfor the isometrygroupofHn. Thus,our
computationstarts inHnbeforemovingtoDn.
TheKilling formontheLiealgebrasp(n,R)of the isometrygroupofHn inducesascalarproduct
onp. This scalarproduct canbe transportedon exp(p)by leftmultiplication. Thisoperationgives
exp(p)aRiemannianstructure. It canbeshownthatonthisRiemannianmanifold, theRiemannian
exponentialat the identitycoincideswith thegroupexponential. Furthermore,
φ : exp(p) → Hn
g → g.iI (10)
isabijective isometry,uptoascalingfactor. Since thevolumechangeθp is invariantunderrescaling
of themetric, thisscalingfactorhasnoimpact. Thus,Hn canbe identifiedwith exp(p)and expiI∈Hn
with exp|p. Theexpressionof theRiemannianexponential isdifficult toobtain ingeneral;however,
itboilsdownto thegroupexponential in thecaseof symmetric spaces. This is themainelementof the
computationofθp. TheRiemannianvolumemeasureon exp(p) isnotedvol′. Let
ψ : K×a → p
(k,H) → Adk(H).
Leta+be thediagonalmatriceswithstrictlydecreasingpositiveeigenvalues. LetΛ+be thesetof
positiverootsofsp(n,R)asdescribed inp. 282 in [20] ,
Λ+={ei+ej, i≤ j}∪{ei−ej, i< j},
where ei(H) is the i-thdiagonal termof thediagonalmatrixH. LetCc(E)be the set of continuous
compactly supported functionson the spaceE. In [27], atpage73, it is given that for all t∈Cc(p),
thereexists c1>0suchthat∫
p t(Y)dY= c1 ∫
K ∫
a+ t(ψ(k,H))∏
λ∈Λ+ λ(H)dkdH, (11)
353
Differential Geometrical Theory of Statistics
- Titel
- Differential Geometrical Theory of Statistics
- Autoren
- Frédéric Barbaresco
- Frank Nielsen
- Herausgeber
- MDPI
- Ort
- Basel
- Datum
- 2017
- Sprache
- englisch
- Lizenz
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03842-425-3
- Abmessungen
- 17.0 x 24.4 cm
- Seiten
- 476
- Schlagwörter
- Entropy, Coding Theory, Maximum entropy, Information geometry, Computational Information Geometry, Hessian Geometry, Divergence Geometry, Information topology, Cohomology, Shape Space, Statistical physics, Thermodynamics
- Kategorien
- Naturwissenschaften Physik