![Page 1: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/1.jpg)
The Karcher Mean of Points on SOn
Knut Huper
joint work with
Jonathan Manton (Univ. Melbourne)
National ICT Australia Ltd.
CESAME LLN, 15/7/04 – p.1/25
![Page 2: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/2.jpg)
Contents
Introduction
Centroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 3: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/3.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet mean
The Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 4: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/4.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 5: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/5.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 6: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/6.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 7: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/7.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 8: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/8.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 9: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/9.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 10: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/10.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlook
CESAME LLN, 15/7/04 – p.2/25
![Page 11: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/11.jpg)
Contents
IntroductionCentroids, Karcher mean, Fréchet meanThe Euclidean case
Motivation, why SOn
Radii of convexity and injectivity
Karcher mean on SOn
Cost function, gradient and Hessian
Newton-type algorithm
Convergence results
Discussion, outlookCESAME LLN, 15/7/04 – p.2/25
![Page 12: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/12.jpg)
Several ways todefine a centroid xC
Given x1, . . . , xk ∈ Rn.
1) As the sum
xc :=1
n
k∑
i=1
xi.
2) Equivalently, to ask the vector sum
−→xx1 + · · · + −→xxk
to vanish.
CESAME LLN, 15/7/04 – p.3/25
![Page 13: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/13.jpg)
Several ways todefine a centroid xC
Given x1, . . . , xk ∈ Rn.
1) As the sum
xc :=1
n
k∑
i=1
xi.
2) Equivalently, to ask the vector sum
−→xx1 + · · · + −→xxk
to vanish.
CESAME LLN, 15/7/04 – p.3/25
![Page 14: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/14.jpg)
Several ways todefine a centroid xC
Given x1, . . . , xk ∈ Rn.
1) As the sum
xc :=1
n
k∑
i=1
xi.
2) Equivalently, to ask the vector sum
−→xx1 + · · · + −→xxk
to vanish.
CESAME LLN, 15/7/04 – p.3/25
![Page 15: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/15.jpg)
Several ways todefine a centroid xC
3) (Appolonius of Perga) As unique minimum of
xc := argminx∈R
k∑
i=1
‖x − xi‖2.
4) More generally, assign to each xi a mass mi,∑
mi = 1. By induction
xc1,2=
m1x1 + m2x2
m1 + m2
xc1,2,3=
(m1 + m2)xc1,2+ m3x3
m1 + m2 + m3, . . .
Also works on spheres.
CESAME LLN, 15/7/04 – p.4/25
![Page 16: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/16.jpg)
Several ways todefine a centroid xC
3) (Appolonius of Perga) As unique minimum of
xc := argminx∈R
k∑
i=1
‖x − xi‖2.
4) More generally, assign to each xi a mass mi,∑
mi = 1. By induction
xc1,2=
m1x1 + m2x2
m1 + m2
xc1,2,3=
(m1 + m2)xc1,2+ m3x3
m1 + m2 + m3, . . .
Also works on spheres. CESAME LLN, 15/7/04 – p.4/25
![Page 17: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/17.jpg)
Several ways todefine a centroid xC
5) Axiomatically:Let Φ : R
n × · · · × Rn ⊃ Ξ → R
n be a rule mappingpoints to its centroid.
Axioms:(A1) Φ is symmetric in its arguments.(A2) Φ is smooth.(A3) Φ commutes with the induced action of SEn
on Rn × · · · × R
n.(A4) If Ω ⊂ R
n is an open convex ball then Φ mapsΩ × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.5/25
![Page 18: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/18.jpg)
Several ways todefine a centroid xC
5) Axiomatically:Let Φ : R
n × · · · × Rn ⊃ Ξ → R
n be a rule mappingpoints to its centroid.
Axioms:(A1) Φ is symmetric in its arguments.(A2) Φ is smooth.(A3) Φ commutes with the induced action of SEn
on Rn × · · · × R
n.(A4) If Ω ⊂ R
n is an open convex ball then Φ mapsΩ × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.5/25
![Page 19: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/19.jpg)
Axioms
(A1) Centroid is independent of the ordering ofthe points.
(A2) Small changes in the location of the pointscauses only small changes in xc.
(A3) Invariance w.r.t. translation and rotation.(A4) Centroid lies in the "same region" as the
points themselves. Especially, Φ(x, ., x) = x.
CESAME LLN, 15/7/04 – p.6/25
![Page 20: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/20.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physics
statistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)
smoothing dataplate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 21: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/21.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifolds
pose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)
smoothing dataplate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 22: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/22.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and robotics
shape analysis and shape trackingfuzzy control on manifolds (defuzzification)
smoothing dataplate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 23: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/23.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape tracking
fuzzy control on manifolds (defuzzification)
smoothing dataplate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 24: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/24.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)
smoothing dataplate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 25: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/25.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)smoothing data
plate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 26: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/26.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)smoothing dataplate tectonics
sequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 27: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/27.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)smoothing dataplate tectonicssequence dep. continuum modeling of DNA
comparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 28: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/28.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)smoothing dataplate tectonicssequence dep. continuum modeling of DNAcomparison theorems (diff. geometry)
stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 29: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/29.jpg)
Why centroids onmanifolds?
Engineering, Mathematics, Physicsstatistical inferences on manifoldspose estimation in vision and roboticsshape analysis and shape trackingfuzzy control on manifolds (defuzzification)smoothing dataplate tectonicssequence dep. continuum modeling of DNAcomparison theorems (diff. geometry)stochastic flows of mass distributions onmanifolds (jets in gravitational field)
CESAME LLN, 15/7/04 – p.7/25
![Page 30: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/30.jpg)
The specialorthogonal groupSOn
SOn := X ∈ Rn×n|X>X = I, det X = 1.
Facts:a) SOn is a Lie group,b) is in general not diffeomorphic to a sphere,c) can be equipped with a Riemannian metric,
therefore notion of distance is available,d) is compact and connected, but in general
not simply connected.
CESAME LLN, 15/7/04 – p.8/25
![Page 31: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/31.jpg)
The specialorthogonal groupSOn
SOn := X ∈ Rn×n|X>X = I, det X = 1.
Facts:
a) SOn is a Lie group,b) is in general not diffeomorphic to a sphere,c) can be equipped with a Riemannian metric,
therefore notion of distance is available,d) is compact and connected, but in general
not simply connected.
CESAME LLN, 15/7/04 – p.8/25
![Page 32: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/32.jpg)
The specialorthogonal groupSOn
SOn := X ∈ Rn×n|X>X = I, det X = 1.
Facts:a) SOn is a Lie group,
b) is in general not diffeomorphic to a sphere,c) can be equipped with a Riemannian metric,
therefore notion of distance is available,d) is compact and connected, but in general
not simply connected.
CESAME LLN, 15/7/04 – p.8/25
![Page 33: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/33.jpg)
The specialorthogonal groupSOn
SOn := X ∈ Rn×n|X>X = I, det X = 1.
Facts:a) SOn is a Lie group,b) is in general not diffeomorphic to a sphere,
c) can be equipped with a Riemannian metric,therefore notion of distance is available,
d) is compact and connected, but in generalnot simply connected.
CESAME LLN, 15/7/04 – p.8/25
![Page 34: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/34.jpg)
The specialorthogonal groupSOn
SOn := X ∈ Rn×n|X>X = I, det X = 1.
Facts:a) SOn is a Lie group,b) is in general not diffeomorphic to a sphere,c) can be equipped with a Riemannian metric,
therefore notion of distance is available,
d) is compact and connected, but in generalnot simply connected.
CESAME LLN, 15/7/04 – p.8/25
![Page 35: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/35.jpg)
The specialorthogonal groupSOn
SOn := X ∈ Rn×n|X>X = I, det X = 1.
Facts:a) SOn is a Lie group,b) is in general not diffeomorphic to a sphere,c) can be equipped with a Riemannian metric,
therefore notion of distance is available,d) is compact and connected, but in general
not simply connected.
CESAME LLN, 15/7/04 – p.8/25
![Page 36: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/36.jpg)
Geometry of SOn
a) We think of SOn as a submanifold of Rn×n.
b) Tangent spaceTXSOn
∼= XA|A ∈ Rn×n, A> = −A.
c) (Scaled) Frobenius inner product on Rn×n
〈U, V 〉 =1
2tr(V >U)
restricts to
〈XU, XV 〉 =1
2tr(V >U), U, V ∈ TXSOn.
Gives Riemannian metric on SOn.
CESAME LLN, 15/7/04 – p.9/25
![Page 37: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/37.jpg)
Geometry of SOn
a) We think of SOn as a submanifold of Rn×n.
b) Tangent spaceTXSOn
∼= XA|A ∈ Rn×n, A> = −A.
c) (Scaled) Frobenius inner product on Rn×n
〈U, V 〉 =1
2tr(V >U)
restricts to
〈XU, XV 〉 =1
2tr(V >U), U, V ∈ TXSOn.
Gives Riemannian metric on SOn.
CESAME LLN, 15/7/04 – p.9/25
![Page 38: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/38.jpg)
Geometry of SOn
a) We think of SOn as a submanifold of Rn×n.
b) Tangent spaceTXSOn
∼= XA|A ∈ Rn×n, A> = −A.
c) (Scaled) Frobenius inner product on Rn×n
〈U, V 〉 =1
2tr(V >U)
restricts to
〈XU, XV 〉 =1
2tr(V >U), U, V ∈ TXSOn.
Gives Riemannian metric on SOn.CESAME LLN, 15/7/04 – p.9/25
![Page 39: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/39.jpg)
Geometry of SOn
cont’d
d) Let X ∈ SOn, Ω> = −Ω ∈ Rn×n.
γ : R → SOn,
t 7→ X · et·Ω
is a geodesic through X = γ(0).
∫ T
0〈γ(t), γ(t)〉
12 d t
is minimal (for T not too large..)
CESAME LLN, 15/7/04 – p.10/25
![Page 40: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/40.jpg)
Geometry of SOn
cont’d
d) Let X ∈ SOn, Ω> = −Ω ∈ Rn×n.
γ : R → SOn,
t 7→ X · et·Ω
is a geodesic through X = γ(0).
∫ T
0〈γ(t), γ(t)〉
12 d t
is minimal (for T not too large..)
CESAME LLN, 15/7/04 – p.10/25
![Page 41: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/41.jpg)
Geometry of SOn
cont’d
d) Let X ∈ SOn, Ω> = −Ω ∈ Rn×n.
γ : R → SOn,
t 7→ X · et·Ω
is a geodesic through X = γ(0).
∫ T
0〈γ(t), γ(t)〉
12 d t
is minimal (for T not too large..)
CESAME LLN, 15/7/04 – p.10/25
![Page 42: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/42.jpg)
Geometry of SOn
cont’d
e) Squared distance between any two pointsX, Y ∈ SOn
d2(X, Y ) =1
2min
A>=−Aexp(A)=X>Y
tr(AA>)
= −1
2tr(log(X>Y ))2
CESAME LLN, 15/7/04 – p.11/25
![Page 43: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/43.jpg)
Geometry of SOn
cont’d
e) Squared distance between any two pointsX, Y ∈ SOn
d2(X, Y ) =1
2min
A>=−Aexp(A)=X>Y
tr(AA>)
= −1
2tr(log(X>Y ))2
CESAME LLN, 15/7/04 – p.11/25
![Page 44: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/44.jpg)
Centroid of SOn byaxioms
Let
Ξ ⊂ SOn × · · · × SOn
be open and consider Φ : Ξ → SOn.
(A1) Φ is symmetric in its arguments.(A2) Φ is smooth.(A3) Φ commutes with left and right translation.(A4) If Ω ⊂ SOn is an open convex ball then Φmaps Ω × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.12/25
![Page 45: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/45.jpg)
Centroid of SOn byaxioms
Let
Ξ ⊂ SOn × · · · × SOn
be open and consider Φ : Ξ → SOn.(A1) Φ is symmetric in its arguments.
(A2) Φ is smooth.(A3) Φ commutes with left and right translation.(A4) If Ω ⊂ SOn is an open convex ball then Φmaps Ω × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.12/25
![Page 46: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/46.jpg)
Centroid of SOn byaxioms
Let
Ξ ⊂ SOn × · · · × SOn
be open and consider Φ : Ξ → SOn.(A1) Φ is symmetric in its arguments.(A2) Φ is smooth.
(A3) Φ commutes with left and right translation.(A4) If Ω ⊂ SOn is an open convex ball then Φmaps Ω × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.12/25
![Page 47: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/47.jpg)
Centroid of SOn byaxioms
Let
Ξ ⊂ SOn × · · · × SOn
be open and consider Φ : Ξ → SOn.(A1) Φ is symmetric in its arguments.(A2) Φ is smooth.(A3) Φ commutes with left and right translation.
(A4) If Ω ⊂ SOn is an open convex ball then Φmaps Ω × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.12/25
![Page 48: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/48.jpg)
Centroid of SOn byaxioms
Let
Ξ ⊂ SOn × · · · × SOn
be open and consider Φ : Ξ → SOn.(A1) Φ is symmetric in its arguments.(A2) Φ is smooth.(A3) Φ commutes with left and right translation.(A4) If Ω ⊂ SOn is an open convex ball then Φmaps Ω × · · · × Ω into Ω.
CESAME LLN, 15/7/04 – p.12/25
![Page 49: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/49.jpg)
Notion of convexity
Ω ⊂ SOn is defined to be convex if for anyX, Y ∈ SOn there is a unique geodesic whollycontained in Ω connecting X to Y and such that itis also the unique minimising geodesic in SOn
connecting X to Y .
A function f : Ω → R is convex if for any geodesicγ : [0, 1] → Ω, the function f γ : [0, 1] → R isconvex in the usual sense, that is,
f(γ(t)) ≤ (1 − t)f(γ(0)) + tf(γ(1)), t ∈ [0, 1].
CESAME LLN, 15/7/04 – p.13/25
![Page 50: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/50.jpg)
Notion of convexity
Ω ⊂ SOn is defined to be convex if for anyX, Y ∈ SOn there is a unique geodesic whollycontained in Ω connecting X to Y and such that itis also the unique minimising geodesic in SOn
connecting X to Y .A function f : Ω → R is convex if for any geodesicγ : [0, 1] → Ω, the function f γ : [0, 1] → R isconvex in the usual sense, that is,
f(γ(t)) ≤ (1 − t)f(γ(0)) + tf(γ(1)), t ∈ [0, 1].
CESAME LLN, 15/7/04 – p.13/25
![Page 51: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/51.jpg)
Notion of convexitycont’d
Maximal convex ball (centered at the identity In)
B(I, r) = X ∈ SOn|d(I, X) < r.
rconv is the largest r s.t. B(I, r) is convex andd(I, X) is convex on B(I, r).
Theorem: For SOn it holds rconv = π2 .
CESAME LLN, 15/7/04 – p.14/25
![Page 52: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/52.jpg)
Notion of convexitycont’d
Maximal convex ball (centered at the identity In)
B(I, r) = X ∈ SOn|d(I, X) < r.
rconv is the largest r s.t. B(I, r) is convex andd(I, X) is convex on B(I, r).
Theorem: For SOn it holds rconv = π2 .
CESAME LLN, 15/7/04 – p.14/25
![Page 53: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/53.jpg)
Notion of convexitycont’d
Maximal convex ball (centered at the identity In)
B(I, r) = X ∈ SOn|d(I, X) < r.
rconv is the largest r s.t. B(I, r) is convex andd(I, X) is convex on B(I, r).
Theorem: For SOn it holds rconv = π2 .
CESAME LLN, 15/7/04 – p.14/25
![Page 54: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/54.jpg)
Injectivity radius
For son := A ∈ Rn×n|A> = −A let
exp : son → SOn,
Ψ 7→ exp(Ψ),
and
B(0, ρ) = A ∈ son|12 tr A>A < ρ2.
The injectivity radius rinj of son is the largest ρ s.t.exp |B(0,ρ) is a diffeomorphism onto its image.Theorem: For son it holds rinj = π.
CESAME LLN, 15/7/04 – p.15/25
![Page 55: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/55.jpg)
Injectivity radius
For son := A ∈ Rn×n|A> = −A let
exp : son → SOn,
Ψ 7→ exp(Ψ),
and
B(0, ρ) = A ∈ son|12 tr A>A < ρ2.
The injectivity radius rinj of son is the largest ρ s.t.exp |B(0,ρ) is a diffeomorphism onto its image.
Theorem: For son it holds rinj = π.
CESAME LLN, 15/7/04 – p.15/25
![Page 56: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/56.jpg)
Injectivity radius
For son := A ∈ Rn×n|A> = −A let
exp : son → SOn,
Ψ 7→ exp(Ψ),
and
B(0, ρ) = A ∈ son|12 tr A>A < ρ2.
The injectivity radius rinj of son is the largest ρ s.t.exp |B(0,ρ) is a diffeomorphism onto its image.Theorem: For son it holds rinj = π.
CESAME LLN, 15/7/04 – p.15/25
![Page 57: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/57.jpg)
Karcher mean onSOn
Let Ω ⊂ SOn be open.
A Karcher mean of Q1, . . . , Qk ∈ SOn is defined tobe a minimiser of
f : Ω → R,
f(X) =k∑
i=1
d2(Qi, X).
Existence, uniqueness?
CESAME LLN, 15/7/04 – p.16/25
![Page 58: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/58.jpg)
Karcher mean onSOn
Let Ω ⊂ SOn be open.A Karcher mean of Q1, . . . , Qk ∈ SOn is defined tobe a minimiser of
f : Ω → R,
f(X) =k∑
i=1
d2(Qi, X).
Existence, uniqueness?
CESAME LLN, 15/7/04 – p.16/25
![Page 59: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/59.jpg)
Karcher mean onSOn
Let Ω ⊂ SOn be open.A Karcher mean of Q1, . . . , Qk ∈ SOn is defined tobe a minimiser of
f : Ω → R,
f(X) =k∑
i=1
d2(Qi, X).
Existence, uniqueness?
CESAME LLN, 15/7/04 – p.16/25
![Page 60: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/60.jpg)
Results
Theorem (MH’04):The critical points of
f : Ω → R,
f(X) =k∑
i=1
d2(Qi, X)
are precisely the solutions of
k∑
i=1
log(Q>i X) = 0.
CESAME LLN, 15/7/04 – p.17/25
![Page 61: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/61.jpg)
Results
Theorem (MH’04):The Karcher mean is well defined and satisfiesaxioms (A1)-(A4) of a centroid on the open set
Ξ =⋃
Y ∈SOn
B(Y, π/2) × · · · × B(Y, π/2).
CESAME LLN, 15/7/04 – p.18/25
![Page 62: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/62.jpg)
Results
Theorem (MH’04):The Hessian of f represented along geodesics
d2
d t2(f γ)(t)
∣∣∣∣∣∣t=0
is always positive definite.
CESAME LLN, 15/7/04 – p.19/25
![Page 63: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/63.jpg)
f , grad f andHessian explicitly
f : Ω → R,
f(X) =k∑
i=1
d2(Qi, X) = −k∑
i=1
1
2tr(log(X>Qi))
2.
D f(X)XA = −k∑
i=1
tr(log(Q>i X)A)
=
⟨
2Xk∑
i=1
log(Q>i X)
︸ ︷︷ ︸
=grad f(X)
, XA
⟩
.
CESAME LLN, 15/7/04 – p.20/25
![Page 64: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/64.jpg)
f , grad f andHessian explicitly
f : Ω → R,
f(X) =k∑
i=1
d2(Qi, X) = −k∑
i=1
1
2tr(log(X>Qi))
2.
D f(X)XA = −k∑
i=1
tr(log(Q>i X)A)
=
⟨
2Xk∑
i=1
log(Q>i X)
︸ ︷︷ ︸
=grad f(X)
, XA
⟩
.
CESAME LLN, 15/7/04 – p.20/25
![Page 65: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/65.jpg)
f , grad f andHessian explicitly
d2
d ε2f(
X eεA)
ε=0= vec> A · H(X) · vec A
with (n2 × n2)−matrix
H(X) :=k∑
i=1
Zi(X) coth(Zi(X))
and
Zi(X) :=In ⊗ log(Q>
i X) + log(Q>i X) ⊗ In
2
CESAME LLN, 15/7/04 – p.21/25
![Page 66: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/66.jpg)
f , grad f andHessian explicitly
d2
d ε2f(
X eεA)
ε=0= vec> A · H(X) · vec A
with (n2 × n2)−matrix
H(X) :=k∑
i=1
Zi(X) coth(Zi(X))
and
Zi(X) :=In ⊗ log(Q>
i X) + log(Q>i X) ⊗ In
2
CESAME LLN, 15/7/04 – p.21/25
![Page 67: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/67.jpg)
f , grad f andHessian explicitly
d2
d ε2f(
X eεA)
ε=0= vec> A · H(X) · vec A
with (n2 × n2)−matrix
H(X) :=k∑
i=1
Zi(X) coth(Zi(X))
and
Zi(X) :=In ⊗ log(Q>
i X) + log(Q>i X) ⊗ In
2CESAME LLN, 15/7/04 – p.21/25
![Page 68: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/68.jpg)
Algorithm
Given Q1, ., Qk ∈ SOn, compute a local minimumof f .Step 1: Set X ∈ SOn to an initial estimate.
Step 2: Computek∑
i=1log(Q>
i X).
Step 3: Stop if ‖k∑
i=1log(Q>
i X)‖ is suff. small.
Step 4: Compute the update direction
vec Aopt = −(H(X))−1 k∑
i=1vec(log(Q>
i X))
Step 5: Set X := X eAopt.Step 6: Go to Step 2. CESAME LLN, 15/7/04 – p.22/25
![Page 69: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/69.jpg)
Results
Theorem (MH’04):The algorithm is an intrinsic Newton method.
Theorem:If the algorithm converges, then it convergeslocally quadratically fast.
CESAME LLN, 15/7/04 – p.23/25
![Page 70: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/70.jpg)
Results
Theorem (MH’04):The algorithm is an intrinsic Newton method.Theorem:If the algorithm converges, then it convergeslocally quadratically fast.
CESAME LLN, 15/7/04 – p.23/25
![Page 71: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/71.jpg)
Discussion, outlook
Need simple test to ensure that update step inalgorithm remains in open convex ball ⇒global convergence.
Different RM, e.g. Cayley-like, gives differentfunction, geodesics, etc.., but typically‖KMcay − KMexp‖ << 1.
(H(X))−1 via EVD.
Quasi-Newton (rank-one updates).
CESAME LLN, 15/7/04 – p.24/25
![Page 72: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/72.jpg)
Discussion, outlook
Need simple test to ensure that update step inalgorithm remains in open convex ball ⇒global convergence.
Different RM, e.g. Cayley-like, gives differentfunction, geodesics, etc.., but typically‖KMcay − KMexp‖ << 1.
(H(X))−1 via EVD.
Quasi-Newton (rank-one updates).
CESAME LLN, 15/7/04 – p.24/25
![Page 73: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/73.jpg)
Discussion, outlook
Need simple test to ensure that update step inalgorithm remains in open convex ball ⇒global convergence.
Different RM, e.g. Cayley-like, gives differentfunction, geodesics, etc.., but typically‖KMcay − KMexp‖ << 1.
(H(X))−1 via EVD.
Quasi-Newton (rank-one updates).
CESAME LLN, 15/7/04 – p.24/25
![Page 74: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/74.jpg)
Discussion, outlook
Need simple test to ensure that update step inalgorithm remains in open convex ball ⇒global convergence.
Different RM, e.g. Cayley-like, gives differentfunction, geodesics, etc.., but typically‖KMcay − KMexp‖ << 1.
(H(X))−1 via EVD.
Quasi-Newton (rank-one updates).
CESAME LLN, 15/7/04 – p.24/25
![Page 75: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/75.jpg)
Discussion, outlook
Linear convergent algorithm(joint work with Robert Orsi, ANU)
Xi+1 = Xi e1k
∑k
j=1 log(X>
i Qj)
Centroids on homogeneous (symmetric)spaces.
Project with NICTA vision/robotic program(Richard Hartley) to treat SE3 case.
CESAME LLN, 15/7/04 – p.25/25
![Page 76: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/76.jpg)
Discussion, outlook
Linear convergent algorithm(joint work with Robert Orsi, ANU)
Xi+1 = Xi e1k
∑k
j=1 log(X>
i Qj)
Centroids on homogeneous (symmetric)spaces.
Project with NICTA vision/robotic program(Richard Hartley) to treat SE3 case.
CESAME LLN, 15/7/04 – p.25/25
![Page 77: The Karcher Mean of Points on SOcis610/Huper-Karcher-mean-SO.pdf · Centroids, Karcher mean, Fréchet mean The Euclidean case Motivation, why SOn Radii of convexity and injectivity](https://reader033.vdocument.in/reader033/viewer/2022051607/60300cc42ab2ae74fa6b472a/html5/thumbnails/77.jpg)
Discussion, outlook
Linear convergent algorithm(joint work with Robert Orsi, ANU)
Xi+1 = Xi e1k
∑k
j=1 log(X>
i Qj)
Centroids on homogeneous (symmetric)spaces.
Project with NICTA vision/robotic program(Richard Hartley) to treat SE3 case.
CESAME LLN, 15/7/04 – p.25/25