Page - 415 - in Differential Geometrical Theory of Statistics
Image of the Page - 415 -
Text of the Page - 415 -
Entropy2016,18, 425
In (16), thecurvature termsappear in thecovariantacceleration fornormalMPPs,whilecubic
splines leads to thecurvature termappearing in the thirdorderderivative.
5.CometricFormulationandLow-RankGenerator
Wenowinvestigateacometric gFkM+λgR,where gR isRiemannian, gFkM is a rank kpositive
semi-definite innerproductarisingfromk linearly independent tangentvectors,andλ>0aweight.
WeassumethatgFkM is chosenso thatgFkM+λgR is invertible, eventhoughgFkM is rank-deficient.
The situation corresponds to extracting the first k eigenvectors in Euclidean space PCA. If the
eigenvectorsareestimatedstatistically fromobserveddata, thisallowstheestimationtoberestricted
toonly thefirstkeigenvectors. Inaddition,an importantpractical implicationof theconstruction is
thatanumerical implementationneednot transporta fulld×dmatrix for the frame,butapotentially
much lowerdimensionald×kmatrix. Thispoint is essentialwhendealingwithhigh-dimensional
data, examplesofwhichare landmarkmanifoldsasdiscussed inSection6.
Whenusingthe framebundle tomodelcovariances, thesumformulation isnatural toexpressas
acometric comparedtoametricbecause,with thecometric formulation,gFkM+λgR representsasum
of covariancematrices insteadofa sumofprecisionmatrices. Thus, gFkM+λgR canbe intuitively
thoughtofasadding isotropicnoiseofvarianceλ to thecovariancerepresentedbygFkM.
To pursue this, let FkM denote the bundle of rank k linear mapsRk → TxM. We define
acometricby 〈
ξ, ξ˜ 〉
= δαβ(ξ|hu(uα))(ξ˜|hu(uβ))+λ 〈
ξ, ξ˜ 〉
gR
forξ, ξ˜∈T∗uFkM. Thesumoverα,β is forα,β=1,. . . ,k. Thefirst termisequivalent to the lift (9)of
thecometric 〈
ξ, ξ˜ 〉
= ( ξ|gu(ξˆ) )
givenu :Rk→TxM.Note that in thedefinition(6)ofgu, themapu is
not inverted; thus, thedefinitionof themetric immediatelycarriesover to therank-deficientcase.
Let (xi,uiα),α=1,. . . ,kbeacoordinatesystemonFkM. Theverticaldistribution is in thiscase
spannedby the dkvectorfieldsDjβ = ∂ujβ . Except for index sumsbeingover k insteadof d terms,
thesituation is thussimilar to the full-rankcase.Note that (ξ|π−1∗ w)=(ξ|wjDj)=wiξi. Thecometric
incoordinates is 〈
ξ, ξ˜ 〉
= δαβuiαξiu j
βξ˜j+λg ij
Rξiξ˜j= ξi (
δαβuiαu j
β+λg ij
R )
ξ˜j= ξiWijξ˜j
withWij = δαβuiαu j
β+λg ij
R. We can thenwrite the corresponding sub-Riemannianmetric gFkM in
termsof theadaptedframeD
gFkM(ξhD h+ξhγD hγ)=WihξhDi (17)
because (ξ|gFkM(ξ˜))= 〈
ξ, ξ˜ 〉
= ξiWijξ˜j. That is, thesituation isanalogousto(11), except the termλg ij
R
isaddedtoWij.
ThegeodesicsystemisagaingivenbytheHamilton–Jacobiequations.As in the full-rankcase,
thesystemisspecifiedbythederivativesofgFkM:
∂ylg ij
FkM=W ij
,l , ∂ylg ijβ
FkM=−Wih,lΓ jβ
h −WihΓ jβ
h,l , ∂ylg iαj
FkM=−Γ iα
h,lW hj−ΓiαhW hj
,l ,
∂ylg iαjβ
FkM=Γ iα
k,lW khΓ jβ
h +Γ iα
kW kh
,lΓ jβ
h +Γ iα
kW khΓ jβ
h,l ,
∂ ylζ gijFkM=W ij
,lζ , ∂ ylζ g ijβ
FkM=−Wih,lζΓ jβ
h −WihΓ jβ
h,lζ , ∂ ylζ giαjFkM=−Γ iα
hW hj
,lζ −Γiαh,lζW hj ,
∂ ylζ g iαjβ
FkM=Γ iα
k,lζ WkhΓ jβ
h +Γ iα
kW kh
,lζΓ jβ
h +Γ iα
kW khΓ jβ
h,lζ ,
Γiαh,lζ = ∂ylζ (
Γihku k
α )
= δζαΓihl , W ij
,l =λg ij
R ,l , W ij
,lζ = δilujζ+δ jluiζ .
415
Differential Geometrical Theory of Statistics
- Title
- Differential Geometrical Theory of Statistics
- Authors
- Frédéric Barbaresco
- Frank Nielsen
- Editor
- MDPI
- Location
- Basel
- Date
- 2017
- Language
- English
- License
- CC BY-NC-ND 4.0
- ISBN
- 978-3-03842-425-3
- Size
- 17.0 x 24.4 cm
- Pages
- 476
- Keywords
- Entropy, Coding Theory, Maximum entropy, Information geometry, Computational Information Geometry, Hessian Geometry, Divergence Geometry, Information topology, Cohomology, Shape Space, Statistical physics, Thermodynamics
- Categories
- Naturwissenschaften Physik