algorithmic treatment of nonlinear datamishra/systmod_presentations/...p. grohs jun-15-2012, liege...

139
Algorithmic Treatment of Nonlinear Data P. Grohs ETH Zurich, Seminar for Applied Mathematics Jun-15-2012, Liege University

Upload: others

Post on 02-Aug-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Algorithmic Treatment of Nonlinear DataP. GrohsETH Zurich, Seminar for Applied Mathematics

Jun-15-2012, Liege University

Page 2: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Outline

MotivationWeighted Averages in ManifoldsWaveletsGeodesic Finite Elements

P. Grohs Jun-15-2012, Liege University p. 2

Page 3: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Nonlinear Data

Source: X. Pennec

P. Grohs Jun-15-2012, Liege University p. 3

Page 4: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Nonlinear Data

Source: I. UrRahman

P. Grohs Jun-15-2012, Liege University p. 4

Page 5: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Nonlinear Data

P. Grohs Jun-15-2012, Liege University p. 5

Page 6: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Nonlinear Data

Source: A. Ivancevic

P. Grohs Jun-15-2012, Liege University p. 6

Page 7: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Nonlinear Data

Source: G. Sundaramoorthi

P. Grohs Jun-15-2012, Liege University p. 7

Page 8: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

And many more!

Spherical data (SAR inferogram)Grassmann manifold (array signal processing)Stiefel manifold (low rank approximations)Quaternions (motion design). . .

P. Grohs Jun-15-2012, Liege University p. 8

Page 9: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

And many more!Spherical data (SAR inferogram)

Grassmann manifold (array signal processing)Stiefel manifold (low rank approximations)Quaternions (motion design). . .

P. Grohs Jun-15-2012, Liege University p. 8

Page 10: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

And many more!Spherical data (SAR inferogram)Grassmann manifold (array signal processing)

Stiefel manifold (low rank approximations)Quaternions (motion design). . .

P. Grohs Jun-15-2012, Liege University p. 8

Page 11: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

And many more!Spherical data (SAR inferogram)Grassmann manifold (array signal processing)Stiefel manifold (low rank approximations)

Quaternions (motion design). . .

P. Grohs Jun-15-2012, Liege University p. 8

Page 12: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

And many more!Spherical data (SAR inferogram)Grassmann manifold (array signal processing)Stiefel manifold (low rank approximations)Quaternions (motion design)

. . .

P. Grohs Jun-15-2012, Liege University p. 8

Page 13: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

And many more!Spherical data (SAR inferogram)Grassmann manifold (array signal processing)Stiefel manifold (low rank approximations)Quaternions (motion design). . .

P. Grohs Jun-15-2012, Liege University p. 8

Page 14: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goals/Challenges

Interpolation and Approximation (also for inverseproblems)CompressionNoise RemovalFeature Extraction

P. Grohs Jun-15-2012, Liege University p. 9

Page 15: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goals/ChallengesInterpolation and Approximation (also for inverseproblems)

CompressionNoise RemovalFeature Extraction

P. Grohs Jun-15-2012, Liege University p. 9

Page 16: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goals/ChallengesInterpolation and Approximation (also for inverseproblems)Compression

Noise RemovalFeature Extraction

P. Grohs Jun-15-2012, Liege University p. 9

Page 17: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goals/ChallengesInterpolation and Approximation (also for inverseproblems)CompressionNoise Removal

Feature Extraction

P. Grohs Jun-15-2012, Liege University p. 9

Page 18: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goals/ChallengesInterpolation and Approximation (also for inverseproblems)CompressionNoise RemovalFeature Extraction

P. Grohs Jun-15-2012, Liege University p. 9

Page 19: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Desiderata

Objectivity (respecting natural invariances)Universality (valid for general manifolds)Theoretical Guarantees (same theoretical results asfor linear case)Tractability (same overall complexity as linear case)

In short, generalize linear tools to manifold-valued case.

P. Grohs Jun-15-2012, Liege University p. 10

Page 20: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

DesiderataObjectivity (respecting natural invariances)

Universality (valid for general manifolds)Theoretical Guarantees (same theoretical results asfor linear case)Tractability (same overall complexity as linear case)

In short, generalize linear tools to manifold-valued case.

P. Grohs Jun-15-2012, Liege University p. 10

Page 21: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

DesiderataObjectivity (respecting natural invariances)Universality (valid for general manifolds)

Theoretical Guarantees (same theoretical results asfor linear case)Tractability (same overall complexity as linear case)

In short, generalize linear tools to manifold-valued case.

P. Grohs Jun-15-2012, Liege University p. 10

Page 22: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

DesiderataObjectivity (respecting natural invariances)Universality (valid for general manifolds)Theoretical Guarantees (same theoretical results asfor linear case)

Tractability (same overall complexity as linear case)In short, generalize linear tools to manifold-valued case.

P. Grohs Jun-15-2012, Liege University p. 10

Page 23: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

DesiderataObjectivity (respecting natural invariances)Universality (valid for general manifolds)Theoretical Guarantees (same theoretical results asfor linear case)Tractability (same overall complexity as linear case)

In short, generalize linear tools to manifold-valued case.

P. Grohs Jun-15-2012, Liege University p. 10

Page 24: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

DesiderataObjectivity (respecting natural invariances)Universality (valid for general manifolds)Theoretical Guarantees (same theoretical results asfor linear case)Tractability (same overall complexity as linear case)

In short, generalize linear tools to manifold-valued case.

P. Grohs Jun-15-2012, Liege University p. 10

Page 25: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Central Problem

f1 : Ω→ M f2 : Ω→ M

P. Grohs Jun-15-2012, Liege University p. 11

Page 26: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Central Problem

f1 : Ω→ M f2 : Ω→ M

f1 + f2 =??

P. Grohs Jun-15-2012, Liege University p. 12

Page 27: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Central Problem

f1 : Ω→ M f2 : Ω→ M

f1 + f2 =??linear functionalanalysis!

classicalapproximationtheory!anything based onvector spaces!

P. Grohs Jun-15-2012, Liege University p. 13

Page 28: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Central Problem

f1 : Ω→ M f2 : Ω→ M

f1 + f2 =??linear functionalanalysis!classicalapproximationtheory!

anything based onvector spaces!

P. Grohs Jun-15-2012, Liege University p. 13

Page 29: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Central Problem

f1 : Ω→ M f2 : Ω→ M

f1 + f2 =??linear functionalanalysis!classicalapproximationtheory!anything based onvector spaces!

P. Grohs Jun-15-2012, Liege University p. 13

Page 30: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Weighted Two-Point Averages

f1 : Ω→ M f2 : Ω→ M

P. Grohs Jun-15-2012, Liege University p. 14

Page 31: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Weighted Two-Point Averages

f1 : Ω→ M f2 : Ω→ M

tf1(x) + (1− t)f2(x) = γf2(x)f1(x) (t)

(γqp (·) is the (locally) unique geodesic connecting the points p ∈ M and q ∈ M)

P. Grohs Jun-15-2012, Liege University p. 15

Page 32: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Weighted Two-Point Averages

f1 : Ω→ M f2 : Ω→ M

tf1(x) + (1− t)f2(x) = γf2(x)f1(x) (t)

(γqp (·) is the (locally) unique geodesic connecting the points p ∈ M and q ∈ M)

P. Grohs Jun-15-2012, Liege University p. 15

Page 33: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Observation: Actually, many constructions can be formulatedentirely using weighted averages∑

i

wipi ,∑

i

wi = 1,wi ∈ R!

P. Grohs Jun-15-2012, Liege University p. 16

Page 34: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Quadratic B-Splines

p−1

p0

p1

p2

P. Grohs Jun-15-2012, Liege University p. 17

Page 35: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Quadratic B-Splines

q−2

q−1

q0

q1

q2

q2i = 34 pi + 1

4 pi+1

q2i+1 = 14 pi + 3

4 pi+1

P. Grohs Jun-15-2012, Liege University p. 18

Page 36: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Quadratic B-Splines

q−2

q−1

q0

q1

q2

q2i = 34 pi + 1

4 pi+1

q2i+1 = 14 pi + 3

4 pi+1

P. Grohs Jun-15-2012, Liege University p. 18

Page 37: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Quadratic B-Splines

Sp(i) =∑

j∈Z a(i − 2j)p(j) a = [. . . ,0, 14 ,

34 ,

34 ,

14 ,0, . . . ]

P. Grohs Jun-15-2012, Liege University p. 19

Page 38: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Quadratic B-Splines∑

j∈Z a(i − 2j) = 1

P. Grohs Jun-15-2012, Liege University p. 20

Page 39: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Quadratic B-Splines

limit function with control polygon

P. Grohs Jun-15-2012, Liege University p. 21

Page 40: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Example:motion design/interpolation

limit functionwith control polygon??

P. Grohs Jun-15-2012, Liege University p. 22

Page 41: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Example: finite elements

0 1

1

f (x) =∑

i ϕi (x)fiwith shape functions ϕi , satisfying∑

i ϕi (x) ≡ 1.

x-dependent weighted average of points fi !

P. Grohs Jun-15-2012, Liege University p. 23

Page 42: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Example: finite elements

0 1

1

f (x) =∑

i ϕi (x)fiwith shape functions ϕi , satisfying∑

i ϕi (x) ≡ 1.

x-dependent weighted average of points fi !

P. Grohs Jun-15-2012, Liege University p. 23

Page 43: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Weighted Averages in Manifolds

P. Grohs Jun-15-2012, Liege University p. 24

Page 44: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goal of this Section

Come up with geometric notion of weighted mean

p∗ =N⊕

i=1

wipi ∈ M

of points pi ∈ M with weights wi ∈ R, such that∑N

i=1 wi = 1.

Should also be efficiently computable!

P. Grohs Jun-15-2012, Liege University p. 25

Page 45: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Goal of this Section

Come up with geometric notion of weighted mean

p∗ =N⊕

i=1

wipi ∈ M

of points pi ∈ M with weights wi ∈ R, such that∑N

i=1 wi = 1.

Should also be efficiently computable!

P. Grohs Jun-15-2012, Liege University p. 25

Page 46: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Linear Averages Revisited

In Rd we have

p∗ =N∑

i=1

wipi ⇔N∑

i=1

wi−−→p∗pi = 0

Necessary: Computationof difference vectorbetween two points.

p∗

p1

p2

p3

p4

p∗ = 14 (p1 + p2 + p3 + p4)

P. Grohs Jun-15-2012, Liege University p. 26

Page 47: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Linear Averages Revisited

In Rd we have

p∗ =N∑

i=1

wipi ⇔N∑

i=1

wi−−→p∗pi = 0

Necessary: Computationof difference vectorbetween two points.

p∗

p1

p2

p3

p4

p∗ = 14 (p1 + p2 + p3 + p4)

P. Grohs Jun-15-2012, Liege University p. 26

Page 48: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Linear Averages Revisited

In Rd we have

p∗ =N∑

i=1

wipi ⇔N∑

i=1

wi−−→p∗pi = 0

Necessary: Computationof difference vectorbetween two points.

p∗

p1

p2

p3

p4

p∗ = 14 (p1 + p2 + p3 + p4)

P. Grohs Jun-15-2012, Liege University p. 26

Page 49: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Computation of Difference Vectors

(a) Retraction pair basedon exponential map

(b) Retraction pair basedon closest point projection

(c) Retraction pair basedon vertical projection

Figure: Different retractions for the circle.

P. Grohs Jun-15-2012, Liege University p. 27

Page 50: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Retraction Pairs

Definition (G (2012))

Two functions (f ,g),

f : TM → M, g : M ×M → TM,

such that

f (p,0) = p, f (p,g(p,q)) = q for all p, q ∈ M,

and∂

∂ξf (p,0) = Id

are called a retraction pair for the manifold M.

Standard example: f = expM and g = logM .

P. Grohs Jun-15-2012, Liege University p. 28

Page 51: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Retraction Pairs

Definition (G (2012))

Two functions (f ,g),

f : TM → M, g : M ×M → TM,

such that

f (p,0) = p, f (p,g(p,q)) = q for all p, q ∈ M,

and∂

∂ξf (p,0) = Id

are called a retraction pair for the manifold M.

Standard example: f = expM and g = logM .

P. Grohs Jun-15-2012, Liege University p. 28

Page 52: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs

Point-vector addition p ⊕ ξ for ξ ∈ TpMdefined by f (p, ξ).

Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q). Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 53: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs Point-vector addition p ⊕ ξ for ξ ∈ TpM

defined by f (p, ξ). Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q). Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 54: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs Point-vector addition p ⊕ ξ for ξ ∈ TpM

defined by f (p, ξ).

Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q). Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 55: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs Point-vector addition p ⊕ ξ for ξ ∈ TpM

defined by f (p, ξ). Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q). Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 56: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs Point-vector addition p ⊕ ξ for ξ ∈ TpM

defined by f (p, ξ). Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q).

Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 57: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs Point-vector addition p ⊕ ξ for ξ ∈ TpM

defined by f (p, ξ). Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q). Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 58: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Constructing operations from retraction pairs Point-vector addition p ⊕ ξ for ξ ∈ TpM

defined by f (p, ξ). Point-point difference q p ∈ TpM such that

p ⊕ (q p) = q

defined by g(p,q). Weighted average

p∗ =⊕i∈I

wipi ,∑i∈I

wi = 1, and pi ∈ M.

defined implicitly by ∑i∈I

wig(p∗,pi ) = 0.

P. Grohs Jun-15-2012, Liege University p. 29

Page 59: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Well-Definedness

Theorem (G 2012)

The notion of average is locally well-defined for any retraction pair(f ,g).

M Riemannian manifold, f = expM , g = logM ; Karcher mean(Karcher ∼ 1970, Cartan ∼ 1930).

Computation is efficient as long as the compuation of f and g isefficient (simple algorithm: fixed-point iterationpn+1∗ ← f (pn

∗,∑

i wig(pn∗,pi )))!

P. Grohs Jun-15-2012, Liege University p. 30

Page 60: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Well-Definedness

Theorem (G 2012)

The notion of average is locally well-defined for any retraction pair(f ,g).

M Riemannian manifold, f = expM , g = logM ; Karcher mean(Karcher ∼ 1970, Cartan ∼ 1930).

Computation is efficient as long as the compuation of f and g isefficient (simple algorithm: fixed-point iterationpn+1∗ ← f (pn

∗,∑

i wig(pn∗,pi )))!

P. Grohs Jun-15-2012, Liege University p. 30

Page 61: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Well-Definedness

Theorem (G 2012)

The notion of average is locally well-defined for any retraction pair(f ,g).

M Riemannian manifold, f = expM , g = logM ; Karcher mean(Karcher ∼ 1970, Cartan ∼ 1930).

Computation is efficient as long as the compuation of f and g isefficient (simple algorithm: fixed-point iterationpn+1∗ ← f (pn

∗,∑

i wig(pn∗,pi )))!

P. Grohs Jun-15-2012, Liege University p. 30

Page 62: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Examples of Retraction Pairs

P. Grohs Jun-15-2012, Liege University p. 31

Page 63: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Sphere

M = x ∈ Rd : ‖x‖2 = 1TpM = ξ ∈ Rd : ξ>p = 0

expM(p, ξ) = cos(‖ξ‖2)p + sin(‖ξ‖2)ξ

‖ξ‖2

logM(p,q) = arccos(p>q)q − (p>q)p‖q − (p>q)p‖2

P. Grohs Jun-15-2012, Liege University p. 32

Page 64: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Lie Group of Euclidean Motions

SE(d) = A =

(R x0 1

): x ∈ Rd ,R ∈ SO(3)

TASE(d) = ξ =

(ξ1 ξ20 0

): ξ1 ∈ Rd , ξ2 ∈ sl3

expSE(d)(A, ξ) = A exp(ξ)

logSE(d)(A,B) = log(A−1B)

P. Grohs Jun-15-2012, Liege University p. 33

Page 65: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Symmetric Space of SPD-Matrices

SPD(d) = A ∈ Rd×d : A 0

TASPD(d) = ξ ∈ Rd×d : ξ> = ξ

expM(A, ξ) = A1/2 exp(A−1/2ξA−1/2)A1/2

logM(A,B) = A1/2 log(A−1/2BA−1/2)A1/2

P. Grohs Jun-15-2012, Liege University p. 34

Page 66: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Stiefel Manifold

St(n,p) = A ∈ Rn×p : A>A = Ip

TASt(n,p) = ξ ∈ Rn×p : ξ>A+A>ξ = 0p

expM(A, ξ) = some complicated expression

logM(A,B) = no explicit expression known!

Important: All operations computable inorder pn operations (not order n2).

Sources: TOP: [Lubich, Nonnenmacher

2010], Bottom: [Sundaramoorthi etal

2011]

P. Grohs Jun-15-2012, Liege University p. 35

Page 67: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Stiefel Manifold

St(n,p) = A ∈ Rn×p : A>A = Ip

TASt(n,p) = ξ ∈ Rn×p : ξ>A+A>ξ = 0p

expM(A, ξ) = some complicated expression

logM(A,B) = no explicit expression known!

Important: All operations computable inorder pn operations (not order n2).

Sources: TOP: [Lubich, Nonnenmacher

2010], Bottom: [Sundaramoorthi etal

2011]

P. Grohs Jun-15-2012, Liege University p. 35

Page 68: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Stiefel Manifold

St(n,p) = A ∈ Rn×p : A>A = Ip

TASt(n,p) = ξ ∈ Rn×p : ξ>A+A>ξ = 0p

expM(A, ξ) = some complicated expression

logM(A,B) = no explicit expression known!

Important: All operations computable inorder pn operations (not order n2).

Sources: TOP: [Lubich, Nonnenmacher

2010], Bottom: [Sundaramoorthi etal

2011]

P. Grohs Jun-15-2012, Liege University p. 35

Page 69: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The Stiefel Manifold

St(n,p) = A ∈ Rn×p : A>A = Ip

TASt(n,p) = ξ ∈ Rn×p : ξ>A+A>ξ = 0p

expM(A, ξ) = some complicated expression

logM(A,B) = no explicit expression known!

Important: All operations computable inorder pn operations (not order n2).

Sources: TOP: [Lubich, Nonnenmacher

2010], Bottom: [Sundaramoorthi etal

2011]

P. Grohs Jun-15-2012, Liege University p. 35

Page 70: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Closest Point Projection

Idea: Define f (A, ξ) by closest-point projection onto St(n,p) fromRn×p:

f (A, ξ) := π(A + ξ).

Algorithmic realization using polar decomposition of A + ξ:

π(A + ξ) = U,

whereA + ξ = UP

is the polar decomposition of A + ξ.

Efficient computation possible.

P. Grohs Jun-15-2012, Liege University p. 36

Page 71: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Closest Point Projection

Idea: Define f (A, ξ) by closest-point projection onto St(n,p) fromRn×p:

f (A, ξ) := π(A + ξ).

Algorithmic realization using polar decomposition of A + ξ:

π(A + ξ) = U,

whereA + ξ = UP

is the polar decomposition of A + ξ.

Efficient computation possible.

P. Grohs Jun-15-2012, Liege University p. 36

Page 72: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Computation of the Inverse

Theorem (G 2012)

Assume that −A>B is a Hadamard matrix (only eigenvalues in thenegative half-plane). Then

g(A,B) = 2B∫ ∞

0exp(−tA>B) exp(−tB>A)dt − A.

Fast computation is possible (for instance by alternating projections)!

P. Grohs Jun-15-2012, Liege University p. 37

Page 73: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Computation of the Inverse

Theorem (G 2012)

Assume that −A>B is a Hadamard matrix (only eigenvalues in thenegative half-plane). Then

g(A,B) = 2B∫ ∞

0exp(−tA>B) exp(−tB>A)dt − A.

Fast computation is possible (for instance by alternating projections)!

P. Grohs Jun-15-2012, Liege University p. 37

Page 74: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Summary

We are able to define computationally accessible and flexible notionsof geometric means in virtually all (structured) manifolds of practicalinterest.

P. Grohs Jun-15-2012, Liege University p. 38

Page 75: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

A first Application: Subdivision

P. Grohs Jun-15-2012, Liege University p. 39

Page 76: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Subdivision in SE(3)

P. Grohs Jun-15-2012, Liege University p. 40

Page 77: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Subdivision in SE(3)

P. Grohs Jun-15-2012, Liege University p. 41

Page 78: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Subdivision in SE(3)

P. Grohs Jun-15-2012, Liege University p. 42

Page 79: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Subdivision in SE(3)

P. Grohs Jun-15-2012, Liege University p. 43

Page 80: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Interactive Motion Design

P. Grohs Jun-15-2012, Liege University p. 44

Page 81: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Modeling in the Presence of Obstacles (Pottmann,Wallner 2006)

P. Grohs Jun-15-2012, Liege University p. 45

Page 82: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Smoothness EquivalenceSmoothness equivalence property: nonlinear limit function is assmooth as corresponding linear one.

Theorem (G, SIAM J. Math. Anal. (2010))

The smoothness equivalence property is true for the geometricanalog.

The aim is to establish a (d , f )-proximity condition of the form

‖∆d (T − S)p‖∞ . Ωf (p),

Ωf (p) :=∑j∈Γf

f∏k=1

‖∆k p‖jk∞,

Γf :=

j ∈ Nf

0 :f∑

i=1

iji = f , ‖j‖1 > 1

,

∆p(i) := p(i + 1)− p(i).A perturbation result shows that Cr -smoothness requires(r − 1, r + 1).

P. Grohs Jun-15-2012, Liege University p. 46

Page 83: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Smoothness EquivalenceSmoothness equivalence property: nonlinear limit function is assmooth as corresponding linear one.

Theorem (G, SIAM J. Math. Anal. (2010))

The smoothness equivalence property is true for the geometricanalog.

The aim is to establish a (d , f )-proximity condition of the form

‖∆d (T − S)p‖∞ . Ωf (p),

Ωf (p) :=∑j∈Γf

f∏k=1

‖∆k p‖jk∞,

Γf :=

j ∈ Nf

0 :f∑

i=1

iji = f , ‖j‖1 > 1

,

∆p(i) := p(i + 1)− p(i).A perturbation result shows that Cr -smoothness requires(r − 1, r + 1).

P. Grohs Jun-15-2012, Liege University p. 46

Page 84: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Smoothness EquivalenceSmoothness equivalence property: nonlinear limit function is assmooth as corresponding linear one.

Theorem (G, SIAM J. Math. Anal. (2010))

The smoothness equivalence property is true for the geometricanalog.

The aim is to establish a (d , f )-proximity condition of the form

‖∆d (T − S)p‖∞ . Ωf (p),

Ωf (p) :=∑j∈Γf

f∏k=1

‖∆k p‖jk∞,

Γf :=

j ∈ Nf

0 :f∑

i=1

iji = f , ‖j‖1 > 1

,

∆p(i) := p(i + 1)− p(i).A perturbation result shows that Cr -smoothness requires(r − 1, r + 1).

P. Grohs Jun-15-2012, Liege University p. 46

Page 85: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Smoothness Equivalence

DefinitionS with mask a has polynomial generation degree (PGD) (d , f ) if for allpolynomials p1, . . . ,pk with deg(

∏pj ) < f it holds that∏

j (Spj ) (i)− S(∏

j pj

)(i) = q(i) with some polynomial q of degree

< d .

TheoremAssume S has PGD (d , f ) and

‖∆j p‖∞ . Ω′j (p) j = 1, . . . ,d , (1)

then the geometric analog T satisfies a proximity condition of order(d , f ) with S. The geometric analog of a (stable) linear Cr scheme isCr provided that (1) holds.

P. Grohs Jun-15-2012, Liege University p. 47

Page 86: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Proof Sketch:Step 1 Taylor expansion of ∆d (T − S)p in p.Step 2 Reduction to existence of ’rewriting rule’ of terms∑

i∈Zk

AiΨ (pi1 , . . . ,pik )

as as linear combinations of terms of the form

Ψ(∆j1pi1 , . . . ,∆

jk pik), ‖j‖1 = f ,

Ψ, Ψ k -multilinear forms.Step 3 Translation of rewriting rule to algebraic properties of the

generating function A(x) of (Ai)i∈Zk (vanishing derivatives at(1, . . . ,1)).

Step 4 Show that these algebraic properties are equivalent to PGD oforder (d , f ).

Step 5 Show that stable S ∈ Cr satisfies PGD of order (r − 1, r + 1).

P. Grohs Jun-15-2012, Liege University p. 48

Page 87: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Main Result

TheoremThe smoothness equivalence true is true for the geometric analog.

Proof.Using the previous theorem we need to show that

‖∆j p‖∞ . Ω′j (p) j = 1, . . . , r − 1,

P. Grohs Jun-15-2012, Liege University p. 49

Page 88: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Main Result

TheoremThe smoothness equivalence property is true for the Riemanniananalog.

Proof.Using the previous theorem we need to show that

‖∆jTp‖∞ . Ω′j (p) j = 1, . . . , r − 1,

This is shown using a bootstrapping argument.

P. Grohs Jun-15-2012, Liege University p. 50

Page 89: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Main Result

TheoremThe smoothness equivalence property is true for the Riemanniananalog.

Proof.Using the previous theorem we need to show that

‖∆jTp‖∞ . Ω′j (p) j = 1, . . . , r − 1,

This is shown using a bootstrapping argument.

P. Grohs Jun-15-2012, Liege University p. 50

Page 90: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Approximation Order

DefinitionS has approximation order r if for all f ∈ Cr∥∥∥f − S∞f

∣∣hZ

( ·h

)∥∥∥∞

. hr+1.

Theorem (G, J. Approx. Theory (2010))

Approximation order properties of the linear scheme S carry over tothe nonlinear scheme T .

These results imply that almost anything that can be done withB-splines can be done in general manifolds.

P. Grohs Jun-15-2012, Liege University p. 51

Page 91: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Approximation Order

DefinitionS has approximation order r if for all f ∈ Cr∥∥∥f − S∞f

∣∣hZ

( ·h

)∥∥∥∞

. hr+1.

Theorem (G, J. Approx. Theory (2010))

Approximation order properties of the linear scheme S carry over tothe nonlinear scheme T .

These results imply that almost anything that can be done withB-splines can be done in general manifolds.

P. Grohs Jun-15-2012, Liege University p. 51

Page 92: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Approximation Order

DefinitionS has approximation order r if for all f ∈ Cr∥∥∥f − S∞f

∣∣hZ

( ·h

)∥∥∥∞

. hr+1.

Theorem (G, J. Approx. Theory (2010))

Approximation order properties of the linear scheme S carry over tothe nonlinear scheme T .

These results imply that almost anything that can be done withB-splines can be done in general manifolds.

P. Grohs Jun-15-2012, Liege University p. 51

Page 93: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

LiteratureH. Karcher, “Mollifier Smoothing and Riemannian Center of Mass”,Comm. Pure and Appl. Anal., (1977).

J. Wallner and N. Dyn, “Convergence and C1-analysis of subdivisionschemes on manifolds by proximity.”, CAGD (2005).

P. Grohs, “A general proximity analysis of nonlinear subdivisionschemes”, SIAM J. Math. Anal. (2010).

O. Ebner, “Nonlinear Markov semigroups and refinement schemes onmetric spaces” arxiv:1112.6003 (2012).

P. Grohs, “Geometric multiscale decompositions of dynamic low-rankmatrices.”, SAM Report 2012-04, (2012).

P. Grohs Jun-15-2012, Liege University p. 52

Page 94: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Wavelet Transforms

P. Grohs Jun-15-2012, Liege University p. 53

Page 95: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Interpolatory Wavelets (Decomposition)given dense samples Pn(f ) := (f (i/2n))n∈Zof continuous function f

P. Grohs Jun-15-2012, Liege University p. 54

Page 96: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Interpolatory Wavelets (Decomposition)

downsampling

P. Grohs Jun-15-2012, Liege University p. 55

Page 97: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Interpolatory Wavelets (Decomposition)

prediction by interpolatory subdivision schemep(·) 7→ Sp(·) =

∑j∈Z a(· − 2j)p(j)

P. Grohs Jun-15-2012, Liege University p. 56

Page 98: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Interpolatory Wavelets (Decomposition)

store error

P. Grohs Jun-15-2012, Liege University p. 57

Page 99: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Interpolatory Wavelets (Decomposition)

next scale

P. Grohs Jun-15-2012, Liege University p. 58

Page 100: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Geometric Multiscale Decomposition

Let Pnf (i) := f ( i2n ) and T a manifold-valued analog of the subdivision

scheme S (replace weighted averages by nonlinear averages).

DefinitionLet f : R→ M continuous. The geometric multiscale decompositionis defined as (

P0(f ),d1T (f ), . . .

),

wheredn

T (f ) := Pn(f ) TPn−1(f ).

We have

Pn(f ) = (. . .T (T (Pp0(f )⊕ d1T (f ))⊕ d2

T (f ))⊕ . . . )⊕ dnT (f ).

P. Grohs Jun-15-2012, Liege University p. 59

Page 101: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Norm Equivalences

Theorem (G-Wallner, Appl. Comp. Harmon. Anal. (2009))

Given S interpolatory with polynomial reproduction ≥ d andsmoothness ≥ r . Denote T its geometric analog. Let f : Zk → Mcontinuous with decomposition (P0

T (f ),d1T (f ), . . . ). Then

(i) If f ∈ Bα∞,∞ with α < d, then

‖dnT (f )‖∞ . 2−αn.

(ii) If‖dn

T (f )‖∞ . 2−αn.

with α < r , then f ∈ Bα∞,∞.

P. Grohs Jun-15-2012, Liege University p. 60

Page 102: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Stability

Is the nonlinear reconstruction procedure stable?

Yes, a typical result from (G, Constr. Approx. (2010)) for Riemannianmanifolds:Define for v, w ∈ TM the distance

dist(v, w) := ‖v − Ptπ(v)π(w)

(w)‖2.

Theorem (G, Constr. Approx. (2010))Given f ∈ Bα∞,∞ for some α > 0 with decomposition (P0(f ), d1

T (f ), . . . ). Let (Q0, e1, . . . ) be another sequence of coefficientswith ∥∥∥dist

(dn

T (f )(·), en(·))∥∥∥∞

. 2−αn.

Then there exists g ∈ Bα∞,∞ with

(P0(g), d1T (g), . . . ) = (Q0

, e1, . . . )

and‖f g(·)‖∞ . ‖P0(f ) Q0(·)‖∞ +

∑n≥1

∥∥∥dist(

dnT (f )(·), en(·)

)∥∥∥∞.

P. Grohs Jun-15-2012, Liege University p. 61

Page 103: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Stability

Is the nonlinear reconstruction procedure stable?Yes, a typical result from (G, Constr. Approx. (2010)) for Riemannianmanifolds:Define for v, w ∈ TM the distance

dist(v, w) := ‖v − Ptπ(v)π(w)

(w)‖2.

Theorem (G, Constr. Approx. (2010))Given f ∈ Bα∞,∞ for some α > 0 with decomposition (P0(f ), d1

T (f ), . . . ). Let (Q0, e1, . . . ) be another sequence of coefficientswith ∥∥∥dist

(dn

T (f )(·), en(·))∥∥∥∞

. 2−αn.

Then there exists g ∈ Bα∞,∞ with

(P0(g), d1T (g), . . . ) = (Q0

, e1, . . . )

and‖f g(·)‖∞ . ‖P0(f ) Q0(·)‖∞ +

∑n≥1

∥∥∥dist(

dnT (f )(·), en(·)

)∥∥∥∞.

P. Grohs Jun-15-2012, Liege University p. 61

Page 104: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Best N-term Approximation

Assume M a general manifold with operations ⊕, and weightedaverage induced by retraction pair.

Theorem (G 2012)

Let f : [−1,1]→ M, f∣∣[−1,0)

, f∣∣[0,1]∈ Cs. Then, using N wavelet

coefficients, one can approximate f up to an error of order N−s in theL2 norm.

This is as good as if there were no singularity!

P. Grohs Jun-15-2012, Liege University p. 62

Page 105: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Best N-term Approximation

Assume M a general manifold with operations ⊕, and weightedaverage induced by retraction pair.

Theorem (G 2012)

Let f : [−1,1]→ M, f∣∣[−1,0)

, f∣∣[0,1]∈ Cs. Then, using N wavelet

coefficients, one can approximate f up to an error of order N−s in theL2 norm.

This is as good as if there were no singularity!

P. Grohs Jun-15-2012, Liege University p. 62

Page 106: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Best N-term Approximation

Assume M a general manifold with operations ⊕, and weightedaverage induced by retraction pair.

Theorem (G 2012)

Let f : [−1,1]→ M, f∣∣[−1,0)

, f∣∣[0,1]∈ Cs. Then, using N wavelet

coefficients, one can approximate f up to an error of order N−s in theL2 norm.

This is as good as if there were no singularity!

P. Grohs Jun-15-2012, Liege University p. 62

Page 107: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Proof Sketch

Step 1: Collect coefficients not contributing to the singularity up toscale log2(N) which needs O(N) coefficients. By a result of[G-Wallner ACHA (2009)] these coefficients decay of order 2js, wherej is the scale.

Step 2: For each scale j there are only a fixed number L ofcoefficients which feel the singularity. Collect also these up to scale2s log2(N). Coefficient count: 2Ls log2(N) = O(N).

Step 3: Away from an interval I of size ∼ 2−2s log2(N) we have onlythresholded coefficients which decay of order 2−sj . By a result of [GConstr. Approx. (2010)], the reconstruction precedure is stable, whichimplies that on [0,1] \ I there is a uniform (and thus L2-) error of orderN−s.

Step 4: On I we have a uniformly bounded error. Since I is of size2−2s log2(N) ∼ N−2s, the L2 error is of the same order.

P. Grohs Jun-15-2012, Liege University p. 63

Page 108: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Proof SketchStep 1: Collect coefficients not contributing to the singularity up toscale log2(N) which needs O(N) coefficients. By a result of[G-Wallner ACHA (2009)] these coefficients decay of order 2js, wherej is the scale.

Step 2: For each scale j there are only a fixed number L ofcoefficients which feel the singularity. Collect also these up to scale2s log2(N). Coefficient count: 2Ls log2(N) = O(N).

Step 3: Away from an interval I of size ∼ 2−2s log2(N) we have onlythresholded coefficients which decay of order 2−sj . By a result of [GConstr. Approx. (2010)], the reconstruction precedure is stable, whichimplies that on [0,1] \ I there is a uniform (and thus L2-) error of orderN−s.

Step 4: On I we have a uniformly bounded error. Since I is of size2−2s log2(N) ∼ N−2s, the L2 error is of the same order.

P. Grohs Jun-15-2012, Liege University p. 63

Page 109: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Proof SketchStep 1: Collect coefficients not contributing to the singularity up toscale log2(N) which needs O(N) coefficients. By a result of[G-Wallner ACHA (2009)] these coefficients decay of order 2js, wherej is the scale.

Step 2: For each scale j there are only a fixed number L ofcoefficients which feel the singularity. Collect also these up to scale2s log2(N). Coefficient count: 2Ls log2(N) = O(N).

Step 3: Away from an interval I of size ∼ 2−2s log2(N) we have onlythresholded coefficients which decay of order 2−sj . By a result of [GConstr. Approx. (2010)], the reconstruction precedure is stable, whichimplies that on [0,1] \ I there is a uniform (and thus L2-) error of orderN−s.

Step 4: On I we have a uniformly bounded error. Since I is of size2−2s log2(N) ∼ N−2s, the L2 error is of the same order.

P. Grohs Jun-15-2012, Liege University p. 63

Page 110: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Proof SketchStep 1: Collect coefficients not contributing to the singularity up toscale log2(N) which needs O(N) coefficients. By a result of[G-Wallner ACHA (2009)] these coefficients decay of order 2js, wherej is the scale.

Step 2: For each scale j there are only a fixed number L ofcoefficients which feel the singularity. Collect also these up to scale2s log2(N). Coefficient count: 2Ls log2(N) = O(N).

Step 3: Away from an interval I of size ∼ 2−2s log2(N) we have onlythresholded coefficients which decay of order 2−sj . By a result of [GConstr. Approx. (2010)], the reconstruction precedure is stable, whichimplies that on [0,1] \ I there is a uniform (and thus L2-) error of orderN−s.

Step 4: On I we have a uniformly bounded error. Since I is of size2−2s log2(N) ∼ N−2s, the L2 error is of the same order.

P. Grohs Jun-15-2012, Liege University p. 63

Page 111: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Proof SketchStep 1: Collect coefficients not contributing to the singularity up toscale log2(N) which needs O(N) coefficients. By a result of[G-Wallner ACHA (2009)] these coefficients decay of order 2js, wherej is the scale.

Step 2: For each scale j there are only a fixed number L ofcoefficients which feel the singularity. Collect also these up to scale2s log2(N). Coefficient count: 2Ls log2(N) = O(N).

Step 3: Away from an interval I of size ∼ 2−2s log2(N) we have onlythresholded coefficients which decay of order 2−sj . By a result of [GConstr. Approx. (2010)], the reconstruction precedure is stable, whichimplies that on [0,1] \ I there is a uniform (and thus L2-) error of orderN−s.

Step 4: On I we have a uniformly bounded error. Since I is of size2−2s log2(N) ∼ N−2s, the L2 error is of the same order.

P. Grohs Jun-15-2012, Liege University p. 63

Page 112: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Remarks

This result is optimal and as good as the corresponding result forEuclidean data.

Wavelet trafo is nonadaptive, compression goes simply bythresholding small coefficients; the singularity is foundautomatically.

P. Grohs Jun-15-2012, Liege University p. 64

Page 113: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Remarks

This result is optimal and as good as the corresponding result forEuclidean data.

Wavelet trafo is nonadaptive, compression goes simply bythresholding small coefficients; the singularity is foundautomatically.

P. Grohs Jun-15-2012, Liege University p. 64

Page 114: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Remarks

This result is optimal and as good as the corresponding result forEuclidean data.

Wavelet trafo is nonadaptive, compression goes simply bythresholding small coefficients; the singularity is foundautomatically.

P. Grohs Jun-15-2012, Liege University p. 64

Page 115: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Examples

P. Grohs Jun-15-2012, Liege University p. 65

Page 116: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Flight Data US Air Compression

original data 10 : 1 compression

Data source: www-stat.stanford.edu/$\sim$symmlab

P. Grohs Jun-15-2012, Liege University p. 66

Page 117: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Dynamical Low-Rank MatricesThe matrices of fixed rank form a smooth manifold:

Mn,p :=

A ∈ Rn×n : rank(A) = p.

Piecewise smooth curves γ : t 7→ A(t) ∈Mn,p can be decomposedinto smooth curves on the Stiefel manifold:

Theorem (Smooth Singular Value Decomposition)

Assume that γ as above is a (piecewise) Cs smooth curve. Thenthere exist (piecewise) Cs smooth curves

γU : t 7→ U(t) ∈ St(n,p), γV : t 7→ V(t) ∈ St(n,p), γS : t 7→ S(t) ∈ Rp×p

such thatA(t) = U(t)S(t)V(t)>.

P. Grohs Jun-15-2012, Liege University p. 67

Page 118: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Occurences

Video compression Latent semantic indexing Low-rank (tensor) approximation of high-dimensional evolution

equations

P. Grohs Jun-15-2012, Liege University p. 68

Page 119: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

A Smooth Example

Smooth matrix curve

A(t) = U(t)Σ(t)V(t)> = sin (tE) + noise ∈ R10×10.

Use rank 2 approximation.

P. Grohs Jun-15-2012, Liege University p. 69

Page 120: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 17

6

5

4

3

2

7

6

5

4

3

original data

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 17

6

5

4

3

2

7

6

5

4

3

thresholded coefficients

P. Grohs Jun-15-2012, Liege University p. 70

Page 121: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

0 20 40 60 80 100 120 140

0.16

0.18

0.2

0.22

0.24

0.26

coordinate plot

0 20 40 60 80 100 120 1402

4

6

8

10

12

14

16

18

20

22

two largest singular values

P. Grohs Jun-15-2012, Liege University p. 71

Page 122: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

A singular Example

Singular matrix curve

A(t) = U(t)Σ(t)V(t)> = sin (tE) +(1− t2)1/2

cos (tF) ∈ R10×10.

Use rank 3 approximation.

P. Grohs Jun-15-2012, Liege University p. 72

Page 123: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 17

6

5

4

3

2

7

6

5

4

3

original data

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 17

6

5

4

3

2

7

6

5

4

3

thresholded coefficients

P. Grohs Jun-15-2012, Liege University p. 73

Page 124: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Real Data

A hyperspectral image is an image where intensities corresponding tomany different wavelengths are recorded separately. The result is aparametrized family A(t) of matrices, t being the wavelength. A(t) isa smooth curve!

Applications: Agriculture, Mineraology, Surveillance, ...

P. Grohs Jun-15-2012, Liege University p. 74

Page 125: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Real Data

A hyperspectral image is an image where intensities corresponding tomany different wavelengths are recorded separately. The result is aparametrized family A(t) of matrices, t being the wavelength. A(t) isa smooth curve!Applications: Agriculture, Mineraology, Surveillance, ...

P. Grohs Jun-15-2012, Liege University p. 74

Page 126: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Hyperspectral Image Compression

Given hyperspectal image A ∈ R144×144×128.1. Decompose A into small 12× 12× 128 blocks2. Approximate each block at each wavelength with a rank 2

approximation3. Perform wavelet thresholding on each low-rank approximation

Results in more than 90% saving with a relative error of order 0.01.

P. Grohs Jun-15-2012, Liege University p. 75

Page 127: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Hyperspectral Image Compression

Given hyperspectal image A ∈ R144×144×128.1. Decompose A into small 12× 12× 128 blocks2. Approximate each block at each wavelength with a rank 2

approximation3. Perform wavelet thresholding on each low-rank approximation

Results in more than 90% saving with a relative error of order 0.01.

P. Grohs Jun-15-2012, Liege University p. 75

Page 128: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Further Work

Feature detection, inpainting, ... Wavelet analysis in curve shape space

P. Grohs Jun-15-2012, Liege University p. 76

Page 129: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

LiteratureI. UrRahman etal., “Multiscale Representation of manifold-valueddata”, SIAM Mult. Mod. Sim., (2006).

P. Grohs and J. Wallner, “Interpolatory wavelets for manifold-valueddata”, ACHA (2009).

P. Grohs, “Stability of manifold-valued subdivision schemes andmultiscale transforms”, Constr. Approx. (2010).

O. Koch and Ch. Lubich, “Dynamical low-rank approximation” SIMAX(2007).

P. Grohs, “Geometric multiscale decompositions of dynamic low-rankmatrices.”, SAM Report 2012-04, (2012).

P. Grohs Jun-15-2012, Liege University p. 77

Page 130: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Geodesic Finite Elements1

1ongoing joint work with Prof. Oliver Sander and Hanne Hardening (FU Berlin)

P. Grohs Jun-15-2012, Liege University p. 78

Page 131: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Problem

Seek minimizer Φ : Ω→ M of manifold-valued variational problemargminJ(Φ), J : H1(Ω,M)→ R.

Applications:Liquid crystals: S2, PR2, SO(3)Cosserat shells: S2, SO(3)Image processing: S2

Problem: DiscretizationFinite elements cannot be used, because space M is nonlinearKey problem: enforce nonlinear constraints given by M

P. Grohs Jun-15-2012, Liege University p. 79

Page 132: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Recall...

0 1

1

f (x) =∑

i ϕi (x)fiwith shape functions ϕi , satisfying∑

i ϕi (x) ≡ 1.

x-dependent weighted average of points fi !

P. Grohs Jun-15-2012, Liege University p. 80

Page 133: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Geodesic Finite Elements

Let ϕpi be the p-th order scalar Lagrangian shape functions. ∆

reference simplex with coordinates ξ.

DefinitionGFE function

Υp(v , ξ) =⊕

i

ϕpi (ξ)vi .

GFE space associated to a simplicial decomposition: restriction toeach element must be GFE function. DOF’s: Control points vi ∈ M.

P. Grohs Jun-15-2012, Liege University p. 81

Page 134: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Results (in preparation)

Yields H1-conforming discretization. Nonlinear Bramble-Hilbert lemma (optimal approximation order

of GFE spaces). Nonlinear Céa lemma (coerciveness⇒ minimization of J on

GFE space yields quasioptimal approximation)

P. Grohs Jun-15-2012, Liege University p. 82

Page 135: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Harmonic Mappings

Harmonic energy

J(Φ) =

∫Ω

‖dΦ∣∣Φ(x)‖2

Φ(x)dx

with Dirichlet BC’s.

Used as simple model in liquid-crystal theory, coercivity results e.g.by [Haskins-Speight (2003)].

P. Grohs Jun-15-2012, Liege University p. 83

Page 136: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Ω = [0,1]2, M = S2

1e-06

1e-05

0.0001

0.001

0.01

0.1

1

10

0.01 0.1 1

err

or

mesh size

L2 errorh1 errorO(h^2)O(h^3)

Source: Oliver Sander

P. Grohs Jun-15-2012, Liege University p. 84

Page 137: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

Remarks

Discrete problems solved using Riemannian trust region schemedeveloped by Absil etal.

Method assumes existence of smooth solution which is notalways the case (but in many cases it is!)

Future work will also address time-dependent problems withpossible applications in image processing.

P. Grohs Jun-15-2012, Liege University p. 85

Page 138: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

LiteratureP.-A. Absil, R. Mahoney and R. Sepulchre “Optimization algorithmson matrix manifolds”, Princeton University Press, (2008).

O. Sander., “Geodesic finite elements for Cosserat rods”, Int. J. Num.Meth. Eng., (2011).

O. Sander, “Geodesic finite elements on simplicial grids”, Int. J. Num.Meth. Eng. (2012) to appear.

P. Grohs, “Finite elements of arbitrary order and quasiinterpolation forRiemannian data”, SAM report 2011-56 (2011).

P. Grohs, H. Hardening and O. Sander, “A convergence theory forGFEs” in preparation.

P. Grohs Jun-15-2012, Liege University p. 86

Page 139: Algorithmic Treatment of Nonlinear Datamishra/systmod_presentations/...P. Grohs Jun-15-2012, Liege University p. 10 Desiderata Objectivity (respecting natural invariances) Universality

The End

Questions?

P. Grohs Jun-15-2012, Liege University p. 87