Analysis and geometry of shape spaces.pdf
-
Upload
nhellcerna -
Category
Documents
-
view
233 -
download
0
Transcript of Analysis and geometry of shape spaces.pdf
-
8/11/2019 Analysis and geometry of shape spaces.pdf
1/220
Analysis and geometry of shape spacesI - V
Peter W. Michor
Winterschool Geometry and Physics, Srni, Jan. 1724, 2009
-
8/11/2019 Analysis and geometry of shape spaces.pdf
2/220
Content:Introduction: What are shape spaces and what
are they good for. A little bit of history.Shape spaces of plane curves:The topology of shape space.Hamiltonian background and conserved momenta.A bunch of metrics, their geodesics and curvatures:The L2 -metric and its vanishing of geodesic dis-tance.Almost local metrics.Immersion Sobolev metrics.The scale invariant Sobolov H 1 -metric and its rela-tion to the Grassmannian of 2-planes in an innite
dimensional space, and Neretin geodesics.A covariant formula for curvature and its rela-tion to ONeills curvature formulas.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
3/220
Shape spaces as quotients of diffeomorphism
groups:
Right invariant metrics on diffeomorphism groups,their geodesics and curvatures.Landmark space, geodesics and curvatures.Spaces of submanifolds (plane curves and higherdimensional ones)High dimensional shape space Imm( M, N ) / Diff( M )Vanishing geodesic distance on diffeomorphism
groups. Burgers equation corresponds to this phe-nomenon. The Camassa-Holm equation has topositive geodesic distance. For the Korteweg-deVries equation we do not know.The universal Teichmueller space with the Weil-Peterssen metric as shape space. (not done)
-
8/11/2019 Analysis and geometry of shape spaces.pdf
4/220
Based on:P.M. and D. Mumford. Riemannian geometries on spaces of plane curves.J. Eur. Math. Soc. (JEMS) 8 (2006), 1-48, arXiv:math.DG/0312384.
H. Kodama, P.M. The homotopy type of the space of degree 0 im-mersed curves. Revista Matematica Complutense 19 (2006), 227-234.arXiv:math/0509694.P.M. and D. Mumford. An overview of the Riemannian metrics on spacesof curves using the Hamiltonian approach. Appl. Numerical Harmonic Analysis 23 (2007), 74-113.arXiv:math.DG/0605009P.M., David Mumford, Jayant Shah, Laurent Younes: A Metric on ShapeSpace with Explicit Geodesics. Rend. Lincei Mat. Appl. 9 (2008) 25-57.arXiv:0706.4299Mario Micheli, P.M., David Mumford: Landmarks. In preparation.David Mumford: Lectures at the Chennai Mathematical Institute.
P.M., David Mumford: Vanishing geodesic distance on spaces of subman-ifolds and diffeomorphisms, Documenta Math. 10 (2005), 217245.arXiv:math.DG/0409303V.Cervera, F.Mascaro, P.M.: The action of the diffeomorphism group onthe space of immersions. Diff. Geom. Appl. 1 (1991), 391401For background material: Peter W. Michor: Some Geometric EvolutionEquations Arising as Geodesic Equations on Groups of Diffeomorphism,Including the Hamiltonian Approach. IN: Phase space analysis of Par-tial Differential Equations. Birkhauser Verlag 2006. Pages 133-215.arXiv:math/0609077
-
8/11/2019 Analysis and geometry of shape spaces.pdf
5/220
Introduction:
What are shapes, why are they interesting, andhow are they arranged in shape spaces.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
6/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
7/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
8/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
9/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
10/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
11/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
12/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
13/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
14/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
15/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
16/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
17/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
18/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
19/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
20/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
21/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
22/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
23/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
24/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
25/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
26/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
27/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
28/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
29/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
30/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
31/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
32/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
33/220
Shape spaces of plane curves:
Some spaces:Diff( S 1 ) a regular Lie group, = Diff + ( S 1 )Diff ( S
1 ).
Emb = Emb( S 1 , R 2 ), the manifold of all smooth
embeddings S 1 R 2 .T Emb( S 1 , R 2 ) = Emb( S 1 , R 2 ) C ( S 1 , R 2 ).Imm = Imm( S 1 , R 2 ), the manifold of all smooth
immersions S 1
R 2
.T Imm( S 1 , R 2 ) = Imm( S 1 , R 2 ) C ( S 1 , R 2 ).Imm free = Imm free ( S 1 , R 2 ), the manifold of all free
smooth immersions S 1
R 2 , i.e., those with triv-
ial isotropy group for the right action of Diff( S 1 )on Imm( S 1 , R 2 ).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
34/220
B e = Be( S 1 , R 2 ) = Emb( S 1 , R 2 ) / Diff( S 1 ), the man-ifold of 1-dimensional connectedsubmanifolds of R 2 ,
B i = B i( S 1 , R 2 ) = Imm( S 1 , R 2 ) / Diff( S 1 ), an in-nite dimensional orbifold
B i, free = Imm free ( S 1 , R 2 ) / Diff( S 1 ), a manifold, the
base of a principal ber bundle,
-
8/11/2019 Analysis and geometry of shape spaces.pdf
35/220
Notation. We work mostly with arclength ds , ar-clength derivative Ds and the unit tangent vector
v to the curve:ds = |c|d
D s = / |c|v = c
/
|c
|Attention: Given a family of curves c( , t ), then and t commute but Ds and t dont. Rota-tion through 90 degrees (complex multiplication by
1) will be denoted by:J = 0 11 0 .
The unit normal vector to the image curve is thusn = Jv.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
36/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
37/220
The degree of immersions. The degree or rota-tion degree of an immersion c : S 1 R 2 is the wind-ing number around 0 of the tangent c : S 1 R 2 .Imm( S
1,
R 2) decomposes into the disjointunion of the open submanifolds Imm k( S 1 , R 2 ) for
k Z according to the degree k. These are con-nected according to a theorem of Whitney and
Graustein (1931-32)
-
8/11/2019 Analysis and geometry of shape spaces.pdf
38/220
Theorem. The manifold Imm k( S 1 , R 2 ) of immersed
curves of degree k contains S 1 as a strong smoothdeformation retract.For k = 0 the manifold B ki ( S
1 , R 2 ) := Imm k( S 1 , R 2 ) / Diff + ( S 1 )is contractible.For k = 0 we have (surprise, Kodama-M.)
1 ( B 0 ( S 1 , R 2 )) = Z ,2 ( B 0 ( S 1 , R 2 )) = Z ,k( B
0 ( S 1 , R 2 )) = 0 for k > 2 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
39/220
The tangent bundle isT Imm( S 1 , R 2 ) = Imm( S 1 , R 2 ) C ( S 1 , R 2 ), thecotangent bundle isT Imm( S 1 , R 2 ) = Imm( S 1 , R 2 ) D( S 1 ) 2where the second factor consists of periodic distri-butions.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
40/220
We consider smooth Riemannian metricson Imm( S 1 , R 2 ), i.e., smooth mappings
G : Imm( S 1 , R 2 ) C ( S 1 , R 2 ) C ( S 1 , R 2 ) R( c,h,k ) Gc( h, k ) , bilinear in h, kGc( h, h ) > 0 for h
= 0 .
Each such metric is weak in the sense that Gc,viewed as bounded linear mapping
Gc : T c Imm( S 1
,R 2
) = C ( S 1
,R 2
) T c Imm( S 1 , R 2 ) = D( S 1 ) 2G : T Imm( S 1 , R 2 ) T Imm( S 1 , R 2 )G ( c, h ) = ( c, G c( h, ))
is injective, but can never be surjective.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
41/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
42/220
The fundamental symplectic form onT Imm( S 1 , R 2 ) pulled back from the canonical sym-plectic form on the contangent bundle via the map-ping G : T Imm( S 1 , R 2 ) T Imm( S 1 , R 2 ) is then:
( c,h ) (( k1 , 1 ) , ( k2 , 2 )) == dG c( k1 )( h, k 2 ) Gc( 1 , k2 )
+ dG c( k2 )( h, k 1 ) + Gc( 2 , k1 )
= Gc( k2 , H c( h, k 1 ) K c( k1 , h ))+ Gc( 2 , k1 ) Gc( 1 , k2 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
43/220
The geodesic equation. The Hamiltonian vectoreld of the Riemann energy function
E ( c, h ) = 12
Gc( h, h ) , E : T Imm( S 1 , R 2 ) Ris the geodesic vector eld:
grad 1 ( E )( c, h ) = hgrad 2 ( E )( c, h ) =
12 H c( h, h ) K c( h, h )
and the geodesic equation becomes:
ct = hh t = 12 H c( h, h ) K c( h, h )
ctt = 12 H c( ct , ct ) K c( ct , ct )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
44/220
The momentum mapping for a G-isometric group
action. Consider a (possibly innite dimensionalregular) Lie group with Lie algebra g with a rightaction g r g by isometries on Imm( S 1 , R 2 ). Fun-damental vector eld mapping : g
X(Imm( S 1 , R 2 ))
a bounded Lie algebra homomorphism, given by
X ( c) = t |0 r exp( tX ) ( c) .momentum map j : g
C
G (T Imm( S 1 , R 2 ) , R ):
jX ( c, h ) = Gc( X ( c) , h ) .
J : T Imm( S 1 , R 2 )
g,
J ( c, h ) , X = jX ( c, h ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
45/220
It ts into the following commmutative diagram
and is a homomorphism of Lie algebras:
0 H 0 i C G grad
X H 1 0
g j
T Imm
J is equivariant for the group action. Along anygeodesic t c( t, ) this momentum mapping isconstant, thus for any X
g
J ( c, c t ) , X = jX ( c, c t ) = Gc( X ( c) , ct )is constant in t.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
46/220
We can apply this construction to the followinggroup actions on Imm( S 1 , R 2 ).
The smooth right action of the group Diff( S 1 ) onImm( S 1 , R 2 ), given by composition from the right:c c for Diff( S 1 ). For X X( S 1 ) thefundamental vector eld is then given by
Diff X (c) = X ( c) = t |0 ( c Fl X t ) = c.X.The reparametrization momentum , for any vector
eld X on S 1
is thus: jX ( c, h ) = Gc( c.X,h ) .
Assuming the metric is reparametrization invariant,
it follows that on any geodesic c( , t ), the expres-sion Gc( c.X,c t ) is constant for all X .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
47/220
The left action of the Euclidean motion groupM (2) = R 2 SO (2) on Imm( S 1 , R 2 ) given by c eaJ c+ B for ( B, e aJ ) R 2SO (2). The fundamentalvector eld mapping is
( B,a ) ( c) = aJc + BThe linear momentum is thus Gc( B, h ) , B R 2 andif the metric is translation invariant, Gc( B, c t ) will
be constant along geodesics. The angular momen-tum is similarly Gc( Jc,h ) and if the metric is rota-tion invariant, then Gc( Jc,c t ) will be constant alonggeodesics.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
48/220
The action of the scaling group of R given byc er c, with fundamental vector eld a ( c) = a.c .If the metric is scale invariant, then the scaling momentum Gc( c, c t ) will also be invariant alonggeodesics.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
49/220
If the Riemannian metric G on Imm is invariantunder the action of Diff( S 1 ) it induces a metric on
the quotient B i as follows. For any C 0 , C 1 B i ,consider all liftings c0 , c1 Imm such that ( c0 ) =C 0 , ( c1 ) = C 1 and all smooth curves t ( c( t, )) in Imm( S 1 , R 2 ) with c(0 , ) = c0 and c(1 , ) =c1 . Since the metric G is invariant under the actionof Diff( S 1 ) the arc-length of the curve t ( c( t, ))in B i( S 1 , R 2 ) is given by
LhorG (c) := LG ( ( c( t,
)))
= 10 G ( c) ( T c.c t , T c.c t ) dt=
1
0
Gc( ct , ct ) dt
dist B i ( S 1 ,R 2 )
G (C 1 , C 2 ) = inf c LhorG (c) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
50/220
The simplest ( L2 -) metric.
G0c ( h, k ) = S 1 h, k ds = S 1 h, k |c|dWe compute the G0 -gradients of c G0c ( h, k ):dG
0
( c)( m )( h, k ) = G0c ( K
0c ( m, h ) , k ) = G
0c m, H
0c ( h, k ) ,
K 0c ( m, h ) = D s ( m ) , v h, D s =
|c|, v =
c
|c|.
H 0c ( h, k ) =
D s h, k v
Geodesic equation
ctt =
1
2 |c| |ct |
2 c
|c|
1
|c|2 ct , c ct .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
51/220
Horizontal Geodesics for G0
ct , c = 0 and ct = an = aJ c
|c
| for a
C ( S 1 , R ).
We use functions a, s = |c|, and , only holonomicderivatives:
s t =
as, a t = 12 a
2 ,
t = a 2 + 1s
a s
= a 2 + as2
ass3
.
We may assume s|t=0 1. Let v( ) = a(0 , ), theinitial value for a. Thens ts = a = 2 a ta , so log( sa 2 ) t = 0, thuss ( t, ) a ( t, ) 2 = s(0 , ) a (0 , ) 2 = v( ) 2 ,a conserved quantity along the geodesic. We sub-
stitute s = v2
a 2 and = 2 a ta 2 to get
-
8/11/2019 Analysis and geometry of shape spaces.pdf
52/220
a tt 4 a2t
a a6
a 2 v4
+ a6
a vv5 a
5a
2
v4 = 0 ,
a (0 , ) = v( ) ,
a nonlinear hyperbolic second order equation. Notethat wherever v = 0 then also a = 0 for all t. Sosubstitute a = vb. The outcome is
( b3
) tt = v2
2 (b3
) 2 vv( b3
) 3vv
2 b3
,b(0 , ) = 1 .
This is the codimension 1 version where
Burgers equation is the codimension 0 version.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
53/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
54/220
The general almost local metric G .
Gc (h, k ) := S 1 ( c, c( )) h ( ) , k ( ) ds.The metric G is invariant under the reparame-tization group Diff( S 1 ) and under the Euclideanmotion group.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
55/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
56/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
57/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
58/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
59/220
The horizontal geodesic equation.
Let c( , t ) be a horizontal geodesic for the metricG . Then ct ( , t ) = a( , t ) .n ( , t ). Denote the in-tegral of a function over the curve with respect toarclength by a bar. Then the geodesic equation forhorizontal geodesics is:
a t = 12 +
2 2 a 2
D 2s 2 a 2 + 2 2 aD 2s ( a )2 1 ( a ) a + ( 1 a 2 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
60/220
Curvature on B i for G .
Let W ( 1 , 2 ) = h( 1 ) m ( 2 )
h ( 2 ) m ( 1 )
so that its second derivative 2 W ( 1 , 1 ) = W 2 ( 1 , 1 ) = h( 1 ) m ( 1 )h ( 1 ) m ( 1 )is the Wronskian of h and m .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
61/220
R 0 ( m,h,m,h ) = G0 ( R 0 ( m, h ) m, h ) =
= . 2 2 + 2 . 2 2( 2 )
2
( 2 )2
2( 1 ) W 2 ( 1 , 1 ) 2 d1
+ 22 ( 1 )2 W 22 ( 1 , 1 ) 2 d1+
1 2
1 2 1 2 ( 1 ) W 2 ( 1 , 1 ) W (
1 , 2 ) ( 2 ) d2 d1
+ 1 2 12 ( 1 ) W 22 ( 1 , 1 ) W ( 1 , 2 ) ( 2 ) d2 d1+
1 ( 1 )
21
2 .
(2 ) W 1 ( 1 , 2 ) 2 d2 d1
+ 2 . 3 2 .4 24 + 2 .2 + 28 . 1 ( 1 ) 1 ( 2 ) W ( 1 , 2 ) 2 d2 d1
+
112
21
4 ( 1 ) 1 ( 1 )
12 ( 2 )
( 2 ) ( 3 ) W ( 1 , 2 ) W ( 1 , 3 ) d2 d1 d3
-
8/11/2019 Analysis and geometry of shape spaces.pdf
62/220
Special case: the metric GA .If we choose ( c, c) = 1 + A 2c then we obtainthe metric we have investigated before:
GAc (h, k ) = S 1 (1 + A c( ) 2 ) h ( ) , k ( ) ds.The horizontal geodesic equation for the GA-metricreduces to
a t = 11 + A 2c 12 ca
2
+ A a 2 (D 2s ( c) + 12 3c )
4 D s ( c) aD s ( a ) 2 cD s ( a ) 2
-
8/11/2019 Analysis and geometry of shape spaces.pdf
63/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
64/220
Area swept out bound.If c is any path from C 0 to C 1 , then
area of the regionswept out by the
variation c max
t ( c( t, )) Lhor GA ( c) .Maximum distance bound.Consider < min { A/ 4 , 3 / 4 / 8} and let = 4( 3 / 4 A
1 / 4 + 1 / 4 ) . Then for any path c
starting at C 0 whose length Lhor GA is , the nal curve lies in the tubular neighborhood of C 0 of width .More precisely, if we choose the path c( t, ) to be
horizontal, thenmax |c(0 , ) c(1 , ) | < .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
65/220
Corollary.For any A > 0, the map from B i( S 1 , R 2 ) in the GA
metric to the space B conti (S 1 , R 2 ) in the Frechet
metric is continuous, and, in fact, uniformly contin-uous on every subset where the length is bounded.In particular, GA is a separating metric on B i( S 1 , R 2 ) .Moreover, the completion B i( S 1 , R 2 ) of B i( S 1 , R 2 )
in this metric can be identied with a subset of B lipi (S 1 , R 2 ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
66/220
Explicit equicontinuity bounds, under appropriateparameterization.
Corollary.If a path c( , t ) , 0 t 1 satises:
|c( , t ) | ( t ) / 2 for all , t , ct , c (0 , t ) 0 in a base point 0 for all t C t (1 + A 2C t ) | ct , ic |2 d/ |c| L2 for all t,then|c( 1 , t 1 ) c( 2 , t 2 ) |
max2 |1 2 |+
+ 7( 3 / 4max /A 1 / 4 + 1 / 4max ) L ( t1 t2 ) (1)
whenever |t1 t2 | min(2 Amin , 3 / 2min ) / (8 L ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
67/220
A numerical simulation of the geodesic connect-ing two circles. Minimize E horG1 (c) for variations cwith initial and end curves unit circles at distance3 produced the following image for the geodesic:
The geodesic joining 2 random shapes of sizeabout 1 at distance 5 apart with A = .25 (using20 time samples and a 48-gon approximation forall curves).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
68/220
The forward integration of the geodesic equationwhen A = 0, starting from a straight line in thedirection given by a smooth bump-like vector eld.Note that two corner like singularities with curva-ture going to are about to form.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
69/220
Top Row: Geodesics in 3 metrics joining the same two ellipses. Ellipseshave eccentricity 3, same center and are rotated at 60 degree.
A = 1; A = 0 .1; A = 0 .01.
Bottom Row: Geodesic triangles in Be formed by joining three ellipses atangles 0, 60 and 120 degrees, for the same three values of A. Here the
intermediate shapes are just rotated versions of the geodesic in the toprow but are laid out on a plane triangle for visualization purposes.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
70/220
The sectional curvature on B i
R 0 ( a,b,a,b ) = GA0 ( R 0 ( a, b ) a, b ) =
= S 1 12 ( A 2 1)( aba b) 2 + A( aba b) 2 d+ S 1
A 2
A2 4 + 2 A2
4 A2
2
1 + A 2 (aba b) 2 d= S 1 ( A 2 1) 2 + 4 A2 8 A2 22(1 + A 2 ) W ( a, b ) 2 d+ S 1 A W ( a, b ) 2 dwhere W ( a, b ) = aba b is the Wronskian of a and
b.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
71/220
Special case: the conformal metrics( ( c) , ( c)) = ( ( c)), metric proposed by Menucciand Yezzi and, for linear, independently by Shah:
Gc (h, k ) = ( c)
S 1
h, k ds = ( c) G0c ( h, k ) .
All these metrics are conformally equivalent to thebasic L2 -metric G0 .As they show, the inmum of path lengths in thismetric is positive so long as saties an inequality( ) C. for some C > 0.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
72/220
More precisely (Shah), if Area( c) is area swept overby the path c,
distG
( C 0 , C 1 ) = inf c
Area( c)
Ae. inf c Area( c) dist GeA( C 0 , C 1 ) Ae.e Amax inf c Area( c)
-
8/11/2019 Analysis and geometry of shape spaces.pdf
73/220
The horizontal geodesic equation reduces to:
a t = 2
a 2 + 1
12 a2 .ds .a.ds a
If we change variables and writeb( s, t ) = ( ( t )) .a ( s, t ), then this equation simpli-es to:
bt = 2 b
2
1
b2
ds
-
8/11/2019 Analysis and geometry of shape spaces.pdf
74/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
75/220
Curvature on B i for the conformal metrics.Sectional curvature has been computed by J. Shah.
Let g, h be orthonomal, then
Curv. in plane g.h=
2 ( g.D s ( h )
h.D s ( g)) 2 +
1
4 g2 . 2 + h2 . 2
+ 3 1 2 2 . 21
4 2 ( g. )2 + ( h. )
2
1
2 D s ( g) 2 + D s ( h ) 2 +
1
2 2 2
Note that the rst two lines are positive while thelast line is negative. The rst term is the curvatureterm for the H 0 -metric. The key point about thisformula is how many positive terms it has.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
76/220
Special case: the smooth scale invariant met-ric GSI
( , ) = 3 + A 2 gives the metric:
GSI c (h, k ) = S 1 13c + A 2c
ch, k ds.
The beauty of this metric is that (a) it is scaleinvariant and (b) log( ) is Lipschitz, hence the in-mum of path lengths is always positive.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
77/220
Horizontal geodesics in this metric as special case
of the equation for G :
a t = 1
1 + A( ) 2 1 + A( )2 a 2
2
2 A2
D s ( a )2
4 A2
D s ( ) aD s ( a )+ 3 + A( ) 2 ( a ) a
32
( a 2 )
A2
2 (a ) 2
where the overline stands now for the average of a function over the curve, i.e.
ds/ .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
78/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
79/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
80/220
Immersion-Sobolev metrics on Imm( S 1 , R 2 ) and
on B iNote that Ds = |c| is anti self-adjoint for the metric
G0 , i.e., for all h, k C ( S 1 .R 2 ) we have
S 1 D s ( h ) , k ds = S 1 h, D s ( k) dsThe metric:G imm ,nc (h, k ) =
S 1
( h, k + A. D ns h, Dns k ) .ds
= S 1 Ln ( h ) , k ds whereLn ( h ) or Ln,c ( h ) = I + ( 1) n A.D 2 ns (h )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
81/220
Geodesics in the H imm ,n -metric
( Ln ( ct )) t = Ln ( ct ) , D s ( ct ) v
|ct
|2 ( c)2 n D s ( ct ) , v Ln ct
+ A2
.2 n1 j =1
(1) n + j D 2 n js ct , D js ct ( c) n
-
8/11/2019 Analysis and geometry of shape spaces.pdf
82/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
83/220
ct = u =: X 1 ( c, u )
u t = L1n,c Ln,c ( u ) , D s ( u ) D s ( c)
|ct
|2 ( c)2 JD s ( c) D s ( u ) , D sc u
+ A2
.2 n1 j =1
(1) n + j D 2 n js u, D js u ( c) JD s ( c)
+ ( 1) n A.2 n1 j =1
D js D s ( u ) , D s ( c) D2 n js (u )
=: X 2 ( c, u )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
84/220
The conserved momenta of Gimm ,n along anygeodesic t c( t, ):
c, L n,c ( ct )
|c( )
| X( S 1 ) repar. moment.
S 1 Ln,c ( ct ) ds R 2 linear moment. S 1 Jc,L n,c ( ct ) ds R angular moment.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
85/220
Horizontality for Gimm ,n h T c Imm( S 1 , R 2 ) isG imm ,nc -orthogonal to the Diff( S 1 )-orbit through cif and only if
0 = Gimm ,nc (h, X ( c)) = S 1 X. Ln,c ( h ) , c dsfor all X X( S 1 ). So the Gimm ,n -normal bundle isgiven by
N nc = {h C ( S, R 2 ) : Ln,c ( h ) , v = 0 }.The Gn -orthonormal projection T c Imm N nc , de-noted by h h = h,G
nand the complemen-
tary projection h h T c( c Diff( S 1 )) are 1-dimensional pseudo-differential operators.
They are determined as follows:
-
8/11/2019 Analysis and geometry of shape spaces.pdf
86/220
They are determined as follows:
h = X ( h ) .v where Ln,c ( h ) , v = Ln,c ( X ( h ) .v) , vThus we are led to consider the linear differentialoperators associated to Ln.c
Lc , Lc : C ( S 1 ) C ( S 1 ) ,Lc ( f ) = Ln,c ( f.v ) , v = Ln,c ( f.n ) , n ,Lc ( f ) = Ln,c ( f.v ) , n = Ln,c ( f.n ) , v .
The operator Lc is of order 2 n and also unbounded,self-adjoint and positive on L2 ( S 1 ,
|c
|d). In par-
ticular, Lc is injective. Lc , on the other hand is of order 2 n 1 and is skew-adjoint. For example, if n = 1, then one nds that:Lc = A.D
2s + (1 + A.
2
) .I Lc = 2 A..D s A.D s ( ) .I
-
8/11/2019 Analysis and geometry of shape spaces.pdf
87/220
The operator Lc : C ( S 1 ) C ( S 1 ) is invertible.This is by deformation invariance of the index.
We want to go back and forth between the naturalhorizontal space of vector elds a.n and the Gimm ,n -horizontal vector elds {h | Lh,v = 0 }: We useC c : C ( S 1 , R 2 ) C ( S 1 ) given by
C c( h ) := ( Lc )1
Lc ,
a pseudo-differential operator of order -1 so that
a.n + C ( a ) .v is H imm ,n -horizontal
-
8/11/2019 Analysis and geometry of shape spaces.pdf
88/220
The restriction of the metric Gimm ,n to horizontalvector elds hi = ai .n + bi .v can be computed likethis:
G imm ,nc
(h1 , h 2 ) =
S 1
Lh 1 , h 2 .ds
= S 1 L + LC a 1 .a 2 .ds.Thus the metric restricted to horizontal vector elds
is given by the pseudo differential operator Lred =L + L ( L)1 L.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
89/220
The metric on the cotangent space to B i , is simple.On the smooth cotangent spaceC ( S 1 , R 2 ) = G0c ( T c Imm( S 1 , R 2 )) D( S 1 ) 2the dual metric is given by convolution with theelementary kernel K n .
Gnc ( a 1 , a 2 ) = S 1S 1 K n ( s1 s2 ) .. n c( s1 ) , n c( s2 ) .a 1 ( s1 ) .a 2 ( s2 ) .ds 1 ds 2 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
90/220
Horizontal geodesicsFor any smooth path c in Imm( S 1 , R 2 ) there exists a smooth path in Diff( S 1 ) with ( t, ) = Id S 1depending smoothly on c such that the path e given
by e( t, ) = c( t, ( t, )) is horizontal: Ln,c ( et ) , e =0 .
We may specialize the general geodesic equation to
horizontal paths and then take the v and n partsof the geodesic equation. For a horizontal path wemay write Ln,c ( ct ) = an for a ( t, ) = Ln,c ( ct ) , n .The v part of the equation turns out to vanish
-
8/11/2019 Analysis and geometry of shape spaces.pdf
91/220
identically and then n part gives us
a t = |ct |2 ( c)
2 D sct , v a++
( c)
2
2 n1 j =1
(
1) n + j D 2 n js ct , D js ct
A Lipschitz bound for arclength in Gimm ,n
| ( C 1 ) ( C 0 ) | C ( A, n )2 dist B iGn ( C 1 , C 0 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
92/220
The scale invariant Sobolov H 1 -metric and itsrelation to the Grassmannian of 2-planes in an in-nite dimensional space, and Neretin geodesics.
Gc( h, k ) = limA
1
AG imm ,scal ,1
c (h, k )
= 1( c) S 1 D sh, D sk ds
= 1
( c) S 1
h,
D 2
sk ds
on Imm / translations or {c Imm : c(1) = 0 }.
Geodesics in this metric
-
8/11/2019 Analysis and geometry of shape spaces.pdf
93/220
ctt = 12 D 2s cn c ct 2Gc 12 D 1s |D sct |2 vc 1c c ct , n c ds ct D 1s D sct , vc D sct
The conserved momenta of Gimm ,n along anygeodesic t c( t, ):1( c)
c, D2s ( ct ) |c( ) | X( S 1 ) repar. moment.
1( c) S 1 D 2s ( ct ) ds = 0 R 2 linear moment.1( c) S 1 ic,D 2s ( ct ) ds R angular moment.1( c) S 1 c, D 2s ( ct ) ds = t log( ( t )) scaling moment.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
94/220
Thm. For each k 3/ 2 this geodesic equationhas unique local solutions in the Sobolev space of H k-immersions. The solutions depend C on t and on the initial conditions c(0 , . ) and ct (0 , . ) . The domain of existence (in t)is uniform in k and thus this also holds in Imm
:=
{c
Imm( S 1 , R 2 ) : c(1) = 0
}.
Sphere Stiefel and Grassmannian
-
8/11/2019 Analysis and geometry of shape spaces.pdf
95/220
Sphere, Stiefel, and GrassmannianV := {f C ( R , R ) : f ( x + 2 ) = f ( x)}below only : odd case. +: even case.f 2 = 2 0 f 2 dx weak inner product on V .Gr (2 , V ) Grassmannian of oriented 2-planes.T W Gr = L( W, W ) with metric
v 2 = tr( vv) = v( e) 2 + v( f ) 2 ,e, f orthonormal basis of W .
For W Gr (2 , V ) letZ ( V ) = {x : f ( x ) = 0 f W }.Gr 0 (2 , V ) =
{W
Gr (2 , V ) : Z ( W ) =
} open in
Gr (2 , V ).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
96/220
The Stiefel manifoldSt
(2 , V ) of orthonormal pairsin V .St 0 (2 , V ) = {( e, f ) St : Z ( e, f ) = } open in St .T ( e,f ) St = {( e,f ) V 2 : 0 = e,de = f ,f =e,f + f ,e }Metric ( e,f ) 2 = e 2 + f 2 .
St (2 , V )
S ( V 2
open) sphere of radius 2.
V open = C ([0 , 2 ], R ).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
97/220
The basic bijection
( e, f ) = c( ) = 12 0 ( e + if ) 2 dx
: S 0 2fold Imm opentransl.,scalings
: St 0 2fold
Imm oddtransl.,scalings
: Gr0
Imm oddtransl.,rot.,scalings
: Gr 0 /U ( V ) B i,oddtransl.,rot.,scalings
-
8/11/2019 Analysis and geometry of shape spaces.pdf
98/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
99/220
Reparameterizations
-
8/11/2019 Analysis and geometry of shape spaces.pdf
100/220
Let U ( V ) be the group of all unitary operators on V of the form
f (f
) for all smooth
: R
R
with ( x) > 0 and ( x + 2 ) = ( x) + 2 , i.e. liftsof Diff + ( S 1 ).The innitesimal action on V of a periodic vector
eld X on R is f 12 X .f + X.f .Prop. ( e, f ) = ( e ) , ( f ) .A tangent vector (
e,f )
T ( e,f )St is
perpendicular to the rotation orbits iff e, df V = f,de V = 0 .
It is perpendicular to the reparameterization orbit
iff W ( e,e ) + W ( f ,f ) = 0where W ( a, b ) = a.b a .b is the Wronskian.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
101/220
Neretin geodesics on Gr (2 , V )Y.A.Neretin: On Jordan angles and the triangle inequality in Grassmannmanifolds, Geom. Dedicata 86 (2001)
If W 0 , W 1 Gr (2 , V ), use the singular value decom-position of the orthonormal projection
p :
W 0 W 1 . This gives ONB ( e0 , f 0 ) of W 0 and ( e1 , f 1 ) of W 1 such that p( e0 ) = cos( ) e1 , p( f 0 ) = cos( ) f 1 ,e0f
1 and f 0e1 for
0 , / 2 the Jordan angles .The metric is then given bydist( W 0 , W 1 ) =
2 + 2
-
8/11/2019 Analysis and geometry of shape spaces.pdf
102/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
103/220
We apply this to compute the distance betweencurves in Imm od / (sim) and B i, od / (sim). We write
c0 = r0 ( ) ei0 ( ) and c1 = r1 ( ) ei
1 ( ) . We put
e0 = 2 r 0 cos 02 f 0 = 2 r 0 sin 02 ,e
1= 2 r 1 cos
12 f
1= 2 r 1 sin
12 ,
lifting the curves to 2-planes in the Grassmannian.The 2 2 matrix M ( c0 , c1 ) of the orthogonal pro- jection from the space
{e0 , f 0
} to
{e1 , f 1
} in these
bases is:
S 1 2 r 0 .r 1 . cos 02 cos 12 d S 1 2 r 0 .r 1 . cos 02 sin 12 d
S 1 2
r 0 .r 1 . sin 02 cos 12 d
S 1 2
r 0 .r 1 . sin 02 sin 12 d
Notations:
-
8/11/2019 Analysis and geometry of shape spaces.pdf
104/220
C :=
S 1
r 0 .r 1 cos
0 12 d= 12 M ( c0 , c1 ) 11 )M ( c
0 , c1 ) 22
S := S 1 r 0 .r 1 sin 0 12 d=
12 M ( c
0, c
1) 21 M ( c
0, c
1) 12
We have to diagonalize this matrix by rotating thecurve c0 by a constant angle 0 , i.e., the basis
{e0 , f 0} by the angle 0 / 2; and similarly c1 by aconstant angle 1 . So replace 0 by 0 0 and 1 by 1 1 such that (for both signs)
0 =
S 1
r 0 .r 1 sin
( 0 0 ) ( 1 1 )2
d
= S . cos 0 12 C . sin
0 12
-
8/11/2019 Analysis and geometry of shape spaces.pdf
105/220
Thus
0 1 = 2 arctan ( S /C ) .In the newly aligned bases, the diagonal elements of the matrix will be the cosines of the Jordan angles.
The following lemma gives you a formula for them:
If M = a bc d , C = 12 ( ad) , S =
12 ( c b) , then
the singular values of M are:
C 2 + S 2 C 2+ + S 2+ .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
106/220
This gives the formula
D od ,rot ( c0 , c1 ) 2 =
= arccos 2
S 2+ + C
2+ +
S 2
+ C 2
+ arccos 2 S 2 + C 2 S 2+ + C 2+ .This is the distance in the space
Imm od ( S 1 , C )/(transl, rot., scalings).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
107/220
Horizontal Neretin distances.If we want the distance in the quotient spaceB i, od / (transl, rot., scalings) by the groupDiff( S 1 ) we have to take the inmum of this dis-
tance over all reparametrizations.To simplify, we assume that the initial curves c0 , c1
are parametrized by arc length so that r 0 r 1 1 / 2 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
108/220
To describe the inf, we can use that geodesics in
-
8/11/2019 Analysis and geometry of shape spaces.pdf
109/220
B i are horizontal geodesics in Imm.Consider the Neretin geodesic t
{e( t ) , f ( t )
} in
Gr (2 , V ) described above
W ( t ) =
e( t ) = sin((1 t ))
sin( ) .e0 +
sin( t )sin( )
.e1
f ( t ) = sin((1 t ) )sin( ) .f 0 + sin( t )
sin( ) .f 1
for
e0 =
cos ( 0) 02 e1 = 1 cos 1 12 ,
f 0 = sin ( 0) 02 f 1 = 1 sin 1 12 ,where the rotations 0 and 1 must be computedfrom c0 and c1 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
110/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
111/220
This is an ordinary differential equation for which
is coupled to the (integral) equations for calculat-ing the s as functions of . If it is non-singular(i.e., the coefficient function of does not van-ish for any ) then there is a solution , at leastlocally. But the non-existence of the inf describedfor open curves above will also affect closed curvesand global solutions may actually not exist. How-
ever, for closed curves that do not double back onthemselves too much geodesics do seem to usuallyexist.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
112/220
The generic way in which a family of open immersions crosses
the hypersurface where Z = . The parametrized straight linein the middle of the family has velocity with a double zero at
the black dot, hence is not an immersion.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
113/220
100fold blowup in middle
This is a geodesic of open curves running from the curve with the kinkat the top left to the straight line on the bottom right. A blow up of thenext to last curve is shown to reveal that the kink never goes away itmerely shrinks. Thus this geodesic is not continuous in the C 1 -topologyon Bopen . The straight line is parametrized so that it stops for a wholeinterval of time when it hits the middle point and thus it is C 1 -continuous
in Imm open .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
114/220
A great circle geodesic on Bod . The geodesic begins at the circle at thetop left, runs from left to right, then to the second row and nally thethird. It leaves Bod twice: at the top right and bottom left, in both of which the singularity of the rst gure occurs in 2 places. The index of the curve changes from +1 to 3 in the middle row.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
115/220
Curvature. Let W Gr (2 , V ) be a xed 2-plane.Let : V V be the isomorphism which equals1 on W and 1 on W satisfying = 1 . ThenGr is the symmetric space O( V ) / ( O ( W ) O ( W ))with involutive automorphism : O( V )
O( V )
given by ( U ) = .U. For the Lie algebra in theV = W W -decomposition we have
1 00 1
x yT y U
1 00 1
= x yT
y U Here x L( W, W ) , y L( W, W ). The xed pointgroup is O( V ) = O( W ) O ( W ).
The reductive decomposition g = k + p is given by
-
8/11/2019 Analysis and geometry of shape spaces.pdf
116/220
x yT y U
= x 0
0 U , x
so(2)
+ 0 yT y 0 , y L( W, W )For the sectional curvature we have (where we as-sume that Y 1 , Y 2 is orthonormal):
kspan( Y 1 ,Y 2 ) = B ( Y 2 , [[Y 1 , Y 2 ], Y 1 ])= tr W ( y
T 2 y2 y
T 1 y1 + y
T 2 y1 y
T 1 y2 2 y
T 2 y1 y
T 2 y1 )= 12 y
T 2 y1 yT 1 y2 2L2 ( W,W )
+ 12 y2 yT 1 y1 yT 2 2L2 ( W ,W ) 0 .
where L2 stands for the space of Hilbert-Schmidtoperators. Note that there are many orthonormal
-
8/11/2019 Analysis and geometry of shape spaces.pdf
117/220
pairs Y 1 , Y 2 on which sectional curvature vanishesand that its maximum value 2 is attained when yi
are isometries and y2 = Jy 1 where J is rotationthrough angle / 2 in the image plane of y1 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
118/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
119/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
120/220
We have explicit formulas for the ONeill term and
thus for the sectional curvature kB i/(sim)span( h1 ,h 2 ) at a
curve C B i/(sim) and tangent vector hi . We alsohave an explicit upper bound for this as a functionof h1 . This shows that geodesics have at least asmall interval before they meet another geodesic.The size of this interval can be controlled by anupper bound that involves the supremum norm of
the rst two derivatives of h1 .
See the paper for this.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
121/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
122/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
123/220
Curve evolution with and without the closednessconstraint. Lower and upper bounds for the geodesic
distance: 0.433 and 0.439
-
8/11/2019 Analysis and geometry of shape spaces.pdf
124/220
Curve evolution with and without the closednessconstraint. Lower and upper bounds for the geodesic
distance: 0.498 and 0.532
-
8/11/2019 Analysis and geometry of shape spaces.pdf
125/220
Curve evolution with and without the closednessconstraint. Lower and upper bounds for the geodesicdistance: 0.513 and 0.528
-
8/11/2019 Analysis and geometry of shape spaces.pdf
126/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
127/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
128/220
The geodesic equation on Diff( R 2 ) is V.Arnoldsequation EPDiff:
t ( t, ) Diff( R 2 )v( t ) = ( t) 1 X( R 2 ) , u ( t ) = L( v( t )) ,u it
+ j
v j .u ix j
+ u j .v j
x i+ div v.u i = 0 .
The quotient Emb( S 1 , R 2 ) .ff( 2 ) b( 1 2 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
129/220
Diff( R 2 ) Emb( S 1 , R 2 ) i , where i : S
1
R
2.
If c = i , the ber through is.{ : i = i} = { : c = c}..The tangent space to the ber is (right translatedby )
{X X ( R 2 ) : X c = 0 }.The horizontal subspace is the translate by of
{Y : R 2 LY,X dx = 0 , if X c = 0 }.If Y is C then Y = 0. So we need
LY = c( p( ) .ds ) for p C ( S 1 , R 2 ), a distribution
carried by c. Thus
Y ( x) = S 1 F ( x c( )) p( ) ds
Y ( x) = S 1 F ( x c( )) p( ) ds
-
8/11/2019 Analysis and geometry of shape spaces.pdf
130/220
Mapped to T c Emb we get( Y c)( ) = S 1 F ( c( ) c( 1 )) .p( 1 ) .|c( 1 ) |d1=: ( F c p)( ) whereF c( 1 , 2 ) := F ( c( 1 )
c( 2 ))
is an elliptic pseudo differential operator kernel of order 2 n + 1 which is real and positive, so theoperator p
F c
p is self-adjoint and positive, so
injective, and by index deformation it is bijectivebetween the Sobolev spaces on S 1 . The inverseoperator ( F c )
1 has kernel Lc( , 1 ) which is a
pseudo differential operator kernel of order 2 n 1.
Write h = Y c T c Emb and express the horizontal
-
8/11/2019 Analysis and geometry of shape spaces.pdf
131/220
lift Y = Y h in terms of h:
h = Y h c = F ( c( p.ds )) = F c p so p = Lch
Y = Y h = F ( c(( Lch ) .ds ))Y h ( x) = S 1 F ( x c( )) .. S 1 Lc( , 1 ) h ( 1 ) |c( 1 ) |d1 |c( ) |d
Finally the metric:Gdiff ,nc (h, k ) = R 2 LY h , Y k dx=
S 1
S 1
Lc( , 1 ) h ( 1 ) , k ( ) ds 1 ds.
We can now compute K and H and the geodesicequation It becomes simpler if written for the 1
-
8/11/2019 Analysis and geometry of shape spaces.pdf
132/220
equation. It becomes simpler if written for the 1-
current Lc
ct = p.|c| =: q:qt ( 0 ) = S 1 F c( 0 , 1 ) q( 0 ) , q( 1 ) d1
where F c( 1 , 2 ) = grad F ( c( 1 ) c( 2 )).Existence of geodesics. Theorem.Let n
1. For each k > 2n
12 the geodesic equa-
tion has unique local solutions in the Sobolev space of H k-embeddings. The solutions are C in t and in the initial conditions c(0 , . ) and ct (0 , . ) . The domain of existence (in t) is uniform in k and thus this also holds in Emb( S 1 , R 2 ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
133/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
134/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
135/220
Geometry of landmark spaceand of spaces of currents
The diffeomorphism groupDiff = Diff S( R
n ): the regular Lie group of all dif-
-
8/11/2019 Analysis and geometry of shape spaces.pdf
136/220
S ( ) g g pfeomorphisms which are rapidly falling towards theidentity.Its Lie algebra is the space XS ( R
n ) of all smoothvector elds which decrease rapidly, with the neg-ative of the usual bracket as Lie bracket.
We consider XS ( R n ) as pre Hilbert space H L withinner productX, Y H L = R n LX,Y dx
where L : XS ( R n ) XS ( R n ) is an invertible lin-ear (elliptic) scalar differential operator or pseudo-differential operator which is self-adjoint with re-spect to the weak inner product
G0 ( X, Y ) = R n X, Y dx
-
8/11/2019 Analysis and geometry of shape spaces.pdf
137/220
on XS ( Rn ) and which is applied to each component
of a vector eld separately.
For example:For the Laplacian = 2i and constant A, let
-
8/11/2019 Analysis and geometry of shape spaces.pdf
138/220
p
i A,
L = (1 A) l =| |l
(A) | |l! ! ( l )! 2
= 1 + + nl
(A) 1 + + n l! 1 ! . . . n ! ( l 1 )! . . . ( l n )!
2 1x1 . . . 2 nxn
The Fourier transform is Lu = (1 + A||2 ) lu ( ).Thus the fundamental solution K of LK = in thespace S ( R n ) of tempered distributions is
K ( x) = 1
(2 ) n R n ei x, 1(1 + ||2 ) l dn which can be expressed in terms of the classicalmodied Bessel functions K l1 ( |x|/ A). It satises
-
8/11/2019 Analysis and geometry of shape spaces.pdf
139/220
( L1 u )( x) = R n K ( xy) u ( y) dn y for each tempereddistribution u.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
140/220
Or:We consider a kernel function K : R n R n R withgood properties (for example smooth and rapidlydecreasing off the diagonal) and its associated op-erator K ( f )( x) = R n K ( x, y ) f ( y) dy which we as-sume to be invertible on C S (R n ) on the space of of smooth functions with compact support, andthen we put L = K
1.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
141/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
142/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
143/220
T 0 ev q0 .( Y 0 ) = Y ( qk) k = i( K ( qk qi) .P i) k
C id P ( P ) T L dN
-
8/11/2019 Analysis and geometry of shape spaces.pdf
144/220
Consider a tangent vector P = ( P k) T q Land N .Its horizontal lift with footpoint 0 is P hor 0where the vector eld P hor on R n is given as follows:Let K 1 ( q) ki be the inverse of the ( N N )-matrixK ( q) ij = K ( qi
q j ). Then
P hor ( x) =i,j
K ( x qi) K 1 ( q) ij P jL ( P hor ( x)) =
i,j( x
qi) K 1 ( q) ij P j
Note that P hor is a vector eld of class H 2 l1 .
The Riemannian metric on Land N induced by thegL metric on Diff ( R n ) is
-
8/11/2019 Analysis and geometry of shape spaces.pdf
145/220
g -metric on Diff S ( R ) isgLq ( P, Q ) = G
L0
( P hor , Q hor ) = R n L ( P hor ) , Q hor dx
= R n i,j ( x qi) K 1 ( q) ij P j , k,l K ( x qk) K 1 ( q) kl Q l dx=
i,j,k,lK 1 ( q) ij K ( qi qk) K 1 ( q) kl P j , Q l
So the metric is given by:
gLq ( P, Q ) = k,l K
1( q) kl P k , Q l .
Recall: K 1 ( q) ki is the inverse of the ( N N )-matrix K ( q)
ij = K ( q
i q j
).
Lemma Let X, Y XS ( R n ) be a vector elds withsupport in a compact box B
R n . Let q1 , q2 , q3 , . . .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
146/220
pp p q1 , q2 , q3 ,be an equidistributed sequence in B : For eachBorel subset U B we require
limN
# {i N : qi U }N
= Vol( U )Vol( B )
.
For each N consider the initial part qN = ( q1 , . . . , q N )as a point in the landmark space Land N of N points in R n . Then we have
limN
Vol( B ) 2
N 2N
i,j =1K 1 ( q) i,j X ( qi) , Y ( q j ) =
=
R n
LX,Y dx.
The geodesic equation on T Land N ( R n ) .Elements of the cotangent bundle
N N N
-
8/11/2019 Analysis and geometry of shape spaces.pdf
147/220
T Land N ( R n ) = Land N ( R n )
(( R n ) N )
are denoted by
( q, ) = ( q1 , . . . , q N ) , 1...
N
=q11 . . . q
1N
. . .qn1 . . . q
nN
, 11 . . .
1n
. . . N 1 . . .
N n
and we shall use this as global coordinates.The metric looks like
( gL )1q (, ) =i,j
K ( q) ij i , j ,
K ( q) ij = K ( qi q j ) .
We consider the the energy function
-
8/11/2019 Analysis and geometry of shape spaces.pdf
148/220
gy
E ( q, ) = 12 ( gL )1q (, ) = 12 i,j K ( q) ij i , j
= 12i,j
K ( q) ij i , j
and its Hamiltonian vector eld (using R n -valuedderivatives to save notation)
H E ( q, ) = 1
2
N
i,j,k =1
K ( q) ij i , j
k
qk
K ( q) ij i , j
qk
k .
=N
i,k =1
K ( qk qi) i
qk+ grad K ( qi qk) i , k
k
.
So the geodesic equation is the ow of this vector
eld:
-
8/11/2019 Analysis and geometry of shape spaces.pdf
149/220
qk = i K ( qi qk) i
k =i
k , i grad K ( qi qk)
A covariant formula for curvature and its rela-tions to ONeills curvature formulas.Mario Micheli in his 2008 thesis derived the thecoordinate version of the following formula for thesectional curvature expression, which is valid forclosed 1-forms , on a Riemannian manifold ( M, g ),
where we view g : T M T M and so g1
is the
-
8/11/2019 Analysis and geometry of shape spaces.pdf
150/220
dual inner product on T M . Here
= g1
( ).gL R ( , ) , = 14 d( g1 ( , )) 2
+ 14 g1 d( 2 ) , d( 2 )+ 34 g [ , ], [ , ]
12 ( 2 ) 12 ( 2 )+ 12 (
+ ) g1 ( , )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
151/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
152/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
153/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
154/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
155/220
Notation for the coordinate formula:A = indices of landmark points in R n
-
8/11/2019 Analysis and geometry of shape spaces.pdf
156/220
A indices of landmark points in
a,b,c, = elements of A, = { a |a A}, { a |a A}, cotangent vectors to L , = the dual tangent vectors, e.g.
a =
b
K ( P a P b) bK ( x) = k( x ) , the kernel dening the metric
note: K ( x) = k( x ) xx
dab =
P a
P b , uab = ( P a
P b) /d ab ,
the unit vector between landmarksK ab = k( dab) ,K ab = DK ( P a P b) = k( dab) uabK ab =
k( dab)k2 ( dab)
1dabk( dab)
Four expressions in the skew form :
-
8/11/2019 Analysis and geometry of shape spaces.pdf
157/220
p
ab,cd ( , ) = a ,K cd b a ,K cd bbcd( , ) =a
( K ac K ad ) ab,cd ( , )= c d,K cd b c d,K cd b
(Note that the terms in angle brackets are discrete strains) ab,cd ( , ) = ( a c) ( a c) , ( b d) ( c d) ,
(Bracket in R nR n , points a, b on left, c, d on right)
bd( , ) = (
b d)( b d) ( b d)( b d) , b d b d
-
8/11/2019 Analysis and geometry of shape spaces.pdf
158/220
With these notations, we get the following formula:
R ( ,,, ) = 12bd
K bd bbd( , ) , dbd( , )
+ 12bcd
( bcb( , ) bcd( , ) ) , cd,bd( , )
34b
bb( , )2K 1 +
12
cd
k bddab cd( , )
18abcd
( K ab K ad K cb + K cd) K ac ,K bd ab,cd ( , ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
159/220
-
8/11/2019 Analysis and geometry of shape spaces.pdf
160/220
The geodesic equation on Diff( R 2 ) is V.Arnoldsequation EPDiff:
t ( t, ) Diff( R 2 )v( t ) = ( t) 1 X( R 2 ) , u ( t ) = L( v( t )) ,u it
+ j
v j .u ix j
+ u j .v j
x i+ div v.u i = 0 .
The quotient Emb( S 1 , R 2 ) .Diff( R 2 ) Emb( S 1 , R 2 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
161/220
i , where i : S
1
R 2.If c = i , the ber through is
.{ : i = i} = { : c = c}..The tangent space to the ber is (right translated
by )
{X X ( R 2 ) : X c = 0 }.The horizontal subspace is the translate by of
{Y :
R
2 LY,X dx = 0 , if X
c = 0
}.
If Y is C then Y = 0. So we needLY = c
( p( ) .ds ) for p C ( S 1 , R 2 ), a distribution
carried by c. Thus
Y ( x) = S 1 K ( x c( )) p( ) ds
-
8/11/2019 Analysis and geometry of shape spaces.pdf
162/220
Write h = Y c T c Emb( S 1 , R 2 ) and express thehorizontal lift Y = Y h in terms of h:h Y K ( ( d )) K L h
-
8/11/2019 Analysis and geometry of shape spaces.pdf
163/220
h = Y h
c = K
( c
( p.ds )) = K c
p so p = Lc
h
Y = Y h = K ( c(( Lch ) .ds ))Y h ( x ) =
= S 1 K ( x c( )) . S
1 Lc( , 1 ) h ( 1 ) |c( 1 ) |d1 |c( ) |dFinally the metric:Gdiff ,nc (h, k ) =
R 2
LY h , Y k dx
= S 1S 1 Lc( , 1 ) h ( 1 ) , k ( ) ds 1 ds.This formula looks innocent, but there is an in-
version of the (nice) operator K c
in it to getLc = ( K c )1
We can now compute K and H and the geodesicequation. It becomes simpler if written for the 1-
-
8/11/2019 Analysis and geometry of shape spaces.pdf
164/220
current Lc
ct = p.
|c
| =: :
t ( 0 ) = S 1 K c( 0 , 1 ) ( 0 ) , ( 1 ) d1where K
c(
1,
2) = grad K ( c(
1)
c(
2)).
Existence of geodesics. Theorem.Let n 1. For each k > 2n 12 the geodesic equa-tion has unique local solutions in the Sobolev space of H k-embeddings. The solutions are C in t and in the initial conditions c(0 , . ) and ct (0 , . ) . The domain of existence (in t) is uniform in k and thus this also holds in Emb( S 1 , R 2 ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
165/220
Horizontal geodesics
-
8/11/2019 Analysis and geometry of shape spaces.pdf
166/220
Horizontal geodesics.A eld h along c is horizontal if Lc h, c = 0.For a horizontal path we have , c = 0, so let = a.n . Then the horizontal geodesic equation is
a t ( ) = t , n ( ) == S 1 K c( , 1 ) , n ( ) a ( ) a ( 1 ) n ( ) , n ( 1 ) d1
Note that also n = Jc/
|c
| appears. It is a strange
equation, but it is well-posed by the theorem above.
Requirements for innite dimensional manifoldsL t ( M ) b k Ri i if ld d
-
8/11/2019 Analysis and geometry of shape spaces.pdf
167/220
Let ( M, g ) be a weak Riemannian manifold mod-elled on convenient locally convex vector spaces.For x M the metric gx : T xM T x M is usuallyonly injective (weak metric). The image g( T M ) T M is called the smooth cotangent bundle asso-ciated to g. Now 1g ( M ) := ( g( T M )) and =g1 X( M ) , X = gX are as above. The exteriorderivative restricts tod : 1g ( M ) 2 ( M ) = ( L2skew ( T M ; R ))since the embedding g( T M ) T M is a smoothber linear mapping.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
168/220
Some constructions above encountered a secondproblem: they lead to vector elds whose values do
-
8/11/2019 Analysis and geometry of shape spaces.pdf
169/220
not lie in T xM , but in the Hilbert space completionT xM with respect to gx . To manipulate theseas in the nite dimensional case, we need to knowthat
x
M T xM forms a smooth vector bundle over
M . In other words, in each coordinate chart on anopen subset U M , T M |U is a trivial bundle U V and all the inner products gx , x U dene innerproducts on one and the same topological vectorspace V . We assume that they are all boundedwith respect to each other, so that the completionV of V with respect to gx does not depend on x
and xU T xM = U V .
This means that xM T xM forms a smooth vectorbundle over M with trivialisations the linear exten-
-
8/11/2019 Analysis and geometry of shape spaces.pdf
170/220
sions of the trivialisations of the bundle T M M .These two properties will be sufficient for all theconstructions we need so we make them into a def-inition:
Denition. A convenient weak Riemannian mani-fold ( M, g ) will be called a robust Riemannian man-
ifold if (1) The metric gx admits gradients in the abovetwo senses,(2) The completions T xM form a vector bundle as
above.
Covariant curvature and ONeills formula ininnite dimensions. Let p : ( E, g E ) (B, g B ) be a
-
8/11/2019 Analysis and geometry of shape spaces.pdf
171/220
p ( , g E ) ( , g B )Riemann submersion between innite dimensionalrobust Riemann manifolds; i.e., for each b Band x E b := p1 ( b) the tangent mapping T x p :( T xE, g E )
(T
bB, g B ) is a surjective metric quo-
tient map so that
b gB := inf X x T xE : T x p.X x = b .The innimum need not be attained in T xE butwill be in the completion T xE . The orthogonalsubspace {Y x : gE ( Y x , T x ( E b) ) = 0 } will thereforebe taken in T x ( E b) in T xE .
If b = gB (
b, ) gB ( T bB ) T b B is an elementi th th d l
-
8/11/2019 Analysis and geometry of shape spaces.pdf
172/220
in the gB
-smooth dual,then p b := ( T x p)( b) = gB ( b, T x p ) : T xE R is in T x M but in general it is not an elementin the smooth dual gE ( T xE ). It is, however, an
element of the Hilbert space completion gE ( T xE )of the gE -smooth dual gE ( T xE ) with respect to thenorm g1E , and the elementg
1E ( p b) = : ( p b)
is in the gE -completionT xE of T xE . We can call g1E ( p b) =: ( p b) thehorizontal lift of b = g1B ( b) T bB .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
173/220
The metric ( gE ) x can be evaluated at elements inthe completion T xE . Moreover, for any smoothsections X, Y ( T E ) the mapping
gE ( X, Y ) : M Ris still smooth, by the smooth uniform boundednesstheorem.
Lemma. If is a smooth 1-form on an open subset U of B with values in the gB -smooth dual gB ( T B ) ,then p is a smooth 1-form on p1 ( U ) E with
-
8/11/2019 Analysis and geometry of shape spaces.pdf
174/220
then p is a smooth 1-form on p ( U )
E with
values in the g1E -completion of the gE -smoothdual gE ( T E ) . Thus also ( p ) is smooth from E into the gE -completion of T E , and it has values in
the gE -orthogonal subbundle to the vertical bundle in the gE -completion. We may continuously ex-tend T x p to the gE -completion, and then we have T p
( p ) =
p. Moreover, the Lie bracket
of two such forms, [( p ) , ( p ) ], is dened. The exterior derivative d( p ) is dened and is applica-ble to vector elds with values in the completion
like ( p )
.That the Lie bracket is dened, is also a non-trivial
-
8/11/2019 Analysis and geometry of shape spaces.pdf
175/220
statement: We have to differentiate in directionswhich are not tangent to the manifold.
Theorem Let p : ( E g E ) (B g B ) be a Riemann
-
8/11/2019 Analysis and geometry of shape spaces.pdf
176/220
Theorem. Let p : ( E, g E )
(B, g B ) be a Riemann
submersion between innite dimensional robust Riemann manifolds. Then for 1-forms , 1g ( B ) ONeills formula holds in the form:
gB R B ( , ) , =
= gE R E (( p ) , ( p ) )( p ) , ( p )
+ 34
[( p
) , ( p
) ]ver 2gE
-
8/11/2019 Analysis and geometry of shape spaces.pdf
177/220
4 Gdiff ,n R ( , ) , =
d ( 1 ) , ( 2 ) ( 1 ) , ( 2 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
178/220
=
( S 1 ) 4
det ( 1 ) , ( 2 )
( 1 ) , ( 2 )
( 3 ) , ( 4 ) ( 3 ) , ( 4 )grad K ( c( 1 ) c( 2 )) , grad K ( c( 3 ) c( 4 )) K c( 1 , 3 ) 2 K c( 1 , 4 ) + K c( 2 , 4 )
+ 3
(S 1 ) 2 L c( 3 , 4 )
S 1 grad K ( c( 3 ) c( 1 )) , ( 3 ) ( 1 ) ( 1 ) grad K ( c( 3 ) c( 1 )) , ( 3 ) ( 1 ) ( 1 ) ,
S 1 grad K ( c( 4 ) c( 2 )) , ( 4 ) ( 2 ) ( 2 )
grad K ( c( 3 ) c( 1 )) , ( 3 ) ( 1 ) ( 1 )+
( S 1 ) 2
2 ( 1 ) , ( 2 ) d2 K ( c( 1 )
c( 2 ))
( 1 ) ( 2 ) , ( 1 ) ( 2 )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
179/220
2 ( 1 ) , ( 2 ) d2 K ( c( 1 ) c( 2 )) ( 1 ) ( 2 ) , ( 1 ) ( 2+ 4 ( 1 ) , ( 2 ) d2 K ( c( 1 ) c( 2 )) ( 1 ) ( 2 ) , ( 1 ) (
High dimensional shape spaceImm( M, N ) / Diff( M ) .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
180/220
M , a compact smooth connected manifold of di-mension m 1.( N, g ) a connected Riemannian manifold of dimen-sion n > m .
Diff( M ), the regular Lie group of all diffeomor-phisms of M .
Diff x0 ( M ), the subgroup of diffeomorphisms xingx0 M .
Emb = Emb( M, N ), the manifold of all smoothembeddings M N .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
181/220
Imm free = Imm free ( M, N ), the manifold of all smoothfree immersions M N (those with trivialisotropy group for the right action of Diff( M )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
182/220
isotropy group for the right action of Diff( M )on Imm( M, N )).
B e = Be( M, N ) = Emb( M, N ) / Diff( M ), the mani-
fold of submanifolds of type M in N , base of a smooth principal bundle.
Bi = B
i( M, N ) = Imm( M, N ) / Diff( M ), an innite
dimensional orbifold.
B i,f = B i, ( M, N ) = Imm f ( M, R 2 ) / Diff( M ), a man-
ifold, the base of a principal ber bundle.
Free immersionsAn immersion f : M N is called free if Diff( M )acts freely on it, i.e., f = f for Diff( M )
-
8/11/2019 Analysis and geometry of shape spaces.pdf
183/220
y , . ., f
f
( M )
implies = Id. We have the following results:
If Diff( M ) has a xed point and if f = f for any immersion f then = Id .
If for f Imm( M, N ) there is a point x c( M )with only one preimage then f is a free immersion.There exist free immersions without such points.
We might view Imm f ( M, N ) as the nonlinear Stiefelmanifold of parametrized submanifolds of type M in N and consequently B i,f ( M, N ) as the nonlinear
Grassmannian of unparametrized submanifolds of type M .
Non free immersions. Since M is compact, theorbit space B i( M, N ) = Imm( M, N ) / Diff( M ) is Haus-dorff. For any immersion f the isotropy group
-
8/11/2019 Analysis and geometry of shape spaces.pdf
184/220
Diff( M ) f is a nite group which acts as groupof covering transformations for a nite coveringqf : M M such that f factors over qf to a freeimmersion f : M
N with f
qf = f .
For each f Imm there exist a slice Q( f ) i n astrong sense: Q( f ) is invariant under the isotropy group Diff( M ) f .
If (Q ( f )
)
Q( f )
=
for
Diff( M ) then is
in the isotropy group Diff( M ) f . Q( f ) Diff( M ) is an invariant open neighbour-hood of the orbit f Diff( M ) in Imm( M, N ) ad-mitting a smooth retraction r onto the orbit. Theber r1 ( f ) equals Q( f ).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
185/220
Volumes of an immersion. For an immersionf Imm( M, N ), we consider the volume density
-
8/11/2019 Analysis and geometry of shape spaces.pdf
186/220
vol g( f ) = vol( f g) Vol( M ) on M given byvol g( f ) |U = det(( f g) ij ) |du 1 du m |for any chart ( U, u : U R m ) of M .Lemma. The derivative of vol g : Imm( M, N ) Vol( M ) is
d vol g( f )( h ) =
Tr f g( g( S f , h
)) vol( f
g)++ div f g( h)( f g)) vol( f g) .
The second summand vanishes when integrated over M .
The metric on Imm . Let h, k C f (M,TN ) betangent vectors with foot point f Imm( M, N ),
-
8/11/2019 Analysis and geometry of shape spaces.pdf
187/220
i.e., vector elds along f . We consider the follow-ing weak Riemannian metric onImm( M, N ), for a constant A 0:
GAf ( h, k ) :=
= M (1 + A Tr f g( S f ) 2gN ( f ) ) g( h, k ) vol( f g)where
Tr f g( S f ) gN ( f ) is the norm of the mean
curvature. The metric GA is invariant for the actionof Diff( M ). This makes the map : Imm( M, N ) B i( M, N ) into a Riemannian submersion (off the
singularities of B i( M, N )).
The tangent vectors to the orbits areT f ( f Diff( M )) = {Tf. : X( M )}. The bundle
-
8/11/2019 Analysis and geometry of shape spaces.pdf
188/220
N Imm( M, N ) of tangent vectors normal to theDiff( M )-orbits is independent of A: N f = {h C ( M,TN ) : g( h,Tf ) = 0 }
= ( f ( T N |M /Tf.TM )) = ( f TN/TM ) ,the space of sections of the normal bundle.A tangent vectorh
T
f Imm( M, N ) = C
f (M,TN ) = ( f
T N ) hasan orthonormal decomposition
h = h + h T f ( f Diff + ( M )) N f into smooth tangential and normal components.
The metric GA on Imm( M, N ) is invariant underDiff( M ) and induces a metric on the quotient B i( M, N For any F 0 , F 1 B i , consider all liftings f 0 , f 1 Imm
-
8/11/2019 Analysis and geometry of shape spaces.pdf
189/220
such that ( f 0 ) = F 0 , ( f 1 ) = F 1 and all smooth curves t f ( t, ) in Imm( M, N ) with f (0 , ) = f 0 and f (1 , ) =f 1 . The length of t
( f ( t,
)) in B i( M, N ) is given
by
LhorGA ( f ) := LGA( ( f ( t, ))) ==
1
0 GA
( f )( T
f .f
t, T
f .f
t) dt =
1
0 GA
f ( f t , f t ) dt
= 10 M (1 + A Tr f g( S f ) 2g ) g( f t , f t ) vol( f g)12dt
In fact the last computation only makes sense onB i,f ( M, N ) but we take it as a motivation.
The metric on B i( M, N ) is dened by taking theinmum of this over all paths f (and all lifts f 0 , f 1 ):
-
8/11/2019 Analysis and geometry of shape spaces.pdf
190/220
dist B iGA( F 1 , F 2 ) = inf f LhorGA ( f ) .
Theorem. For f 0 , f 1 Imm( M, N ) there exists al-ways a path t f ( t, ) in Imm( M, N ) with f (0 , ) =f 0 and ( f (1 , )) = ( f 1 ) such that Lhor G0 ( f ) is arbi-trarily small.
So the lowest order metric is not suitable for vision.Sketch the proof!
-
8/11/2019 Analysis and geometry of shape spaces.pdf
191/220
Area swept out bound. If f is any path from F 0to F then
-
8/11/2019 Analysis and geometry of shape spaces.pdf
192/220
to F 1
, then
( m + 1) volume of the region swept out by the variation f
maxt Volg( f ( t, )) L
hor GA ( f ) .
Together with Lipschitz continuity this shows thatthe geodesic distance LB iGA separates points at leaston Be( M, N ), if A > 0.
Horizontal energy of a path as anisotropic vol-ume We consider a path t f ( t, ) in Imm( M, N ).
-
8/11/2019 Analysis and geometry of shape spaces.pdf
193/220
It projects to a path f in B i whose energy is:E GA( f ) = 12 ba GA ( f ) ( T.f t ,T.f t ) dt =
= 12 ba GAf ( f t , f t ) dt == 12 ba M (1+ A Tr f g( S f ) 2g ) g( f t , f t ) vol( f g) dt.
We now consider the graph f : [a, b ]M ( t, x ) ( t, f ( t, x )) [a, b ]N of the path f and its image f ,an immersed submanifold with boundary of R N .
hor
-
8/11/2019 Analysis and geometry of shape spaces.pdf
194/220
E horGA ( f ) == 12 [a,b ]M (1 + A Tr f g( S f ) 2gN ( f ) )
f t
2
1 + f t 2gvol(
f (dt
2 +g
))
This is intrinsic for the graph f and the brationpr 1 : R N R . To nd a geodesic between theshapes ( f ( a, )) and ( f ( b, )) we look for animmersed surface which is critical for E horGA . This isa Plateau-problem with anisotropic volume.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
195/220
The geodesic equation of G0 in Imm( M, N )
g t f t + div
f g( f t ) f t g( f t , Tr f g( S f )) f t++ 12 T f. grad f g( f t 2g ) + 12 f t 2g Tr f g( S f ) = 0
-
8/11/2019 Analysis and geometry of shape spaces.pdf
196/220
The sectional curvature for G0 in B i( M, N )
kf ( P ( m, h )) =G0f ( R ( m, h ) m, h )
2 2 0 2 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
197/220
m2
h2
G0a ( m, h )
2
We get then for x, y ( N f ):R f ( x,y,x,y ) = G
0f ( R f ( x, y ) x, y ) =
= M vol( f g) 12 Tr( L f L f )( xy) 0
1
4Tr( L f
x) y
Tr( L f
y) x 2
g 0
+ 14 xy2 Tr g( S f ) 2 0
+ g( R g( x, y ) x, y )+ xy
2 Ric( T M, span( x, y ))
12 ( g( x,y) g( y,x) 2 1M 0
1 2
-
8/11/2019 Analysis and geometry of shape spaces.pdf
198/220
+ 2 x
y yx 1M 2 N ( f ) 0 .
Corollary. If M has codimension 1 in N then
all sectional curvatures are non-negative. For any codimension, sectional curvature in the plane spanned by x and y is non-negative if x and y are parallel,i.e., x
y = 0 in
2 T N .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
199/220
The right invariant H 0 -metric on Diff 0 ( N ) is thengiven as follows, where h, k : N T N are vector
-
8/11/2019 Analysis and geometry of shape spaces.pdf
200/220
elds with compact support along and where X =h 1 , Y = k 1 Xc( N ):G0( h, k ) =
N g( h, k ) vol( g)
= N g( X , Y )vol( g)= N g( X, Y ) vol( g)
Theorem. Geodesic distance on Diff 0 ( N ) withrespect to the H 0 -metric vanishes.
x 2retr etr
-
8/11/2019 Analysis and geometry of shape spaces.pdf
201/220
t
ph
3
ph(x)
ph+(x)ph(x)
x
slopea
ph
slopela
ph+
3.5Particle trajectories under , = 0.6
-
8/11/2019 Analysis and geometry of shape spaces.pdf
202/220
0.5 0 0.5 1 1.5 2 2.5 30.5
0
0.5
1
1.5
2
2.5
3
Time t
S p a c e x
-
8/11/2019 Analysis and geometry of shape spaces.pdf
203/220
In our case, for Diff 0 ( N ), we havead( X ) Y = [X, Y ]
-
8/11/2019 Analysis and geometry of shape spaces.pdf
204/220
G0 ( X, Y ) = N g( X, Y ) vol( g)G0 (ad( Y )X, Z ) = G0 ( X, [Y, Z ]) == N g LY X + ( g1LY g) X + div g( Y ) X, Z vol( g)ad( Y ) = LY + g1LY ( g)+d iv g( Y ) = LY + ( Y ) ,
where the tensor eld ( Y ) = g1LY ( g)+d iv g( Y ) : T N T N is self adjoint with respect to g.
Thus the geodesic equation for G0 is
u t = ( g1Lu ( g))( u ) div g( u ) u = ( u ) u,
-
8/11/2019 Analysis and geometry of shape spaces.pdf
205/220
u = t 1 .The main part of the sectional curvature is givenby:
4 G ( R ( X, Y ) X, Y ) =
= N ( X ) Y ( Y ) X + [ X, Y ] 2g4 g([ ( X ) , ( Y )] X, Y ) vol( g)
So sectional curvature consists of a part which isvisibly non-negative, and another part which is dif-
cult to decompose further.
Example. For ( N, g ) = ( R , can) or ( S 1 , can) the
-
8/11/2019 Analysis and geometry of shape spaces.pdf
206/220
p ( , g ) ( , ) ( , )geodesic equation is Burgers equation , a com-pletely integrable innite dimensional system,
ut =
3 ux u, u =
t
1 ,
to which corresponds vanishing geodesic distance.and we get G0 ( R ( X, Y ) X, Y ) = [X, Y ]2 dx sothat all sectional curvatures are non-negative.
Example. For ( N, g ) = ( R n , can) or (( S 1 ) n , can):
(ad( X ) Y ) k =i
(( iX k) Y i X i( iY k))
-
8/11/2019 Analysis and geometry of shape spaces.pdf
207/220
G0 (ad( X ) Y, Z ) = R n dX.Y dY.X,Z dx= R n i,k Y k ( kX i) Z i + ( iX i) Z k + X i( iZ k) dx(ad( X )Z )
k=
=i
( kX i) Z i + ( iX i) Z k + X i( iZ k) ,
so that the geodesic equation is given by
tuk = (ad( u )u ) k == i
( kui) u i + ( iu i) uk + u i( iuk) ,
the n-dimensional analogon of Burgers equation,called the basic Euler-Poincare equation (EPDiff)
-
8/11/2019 Analysis and geometry of shape spaces.pdf
208/220
by Holm. Also here we have vanishing geodesicdistance.
Stronger metrics on Diff 0 ( N ) .A very small strengthening of the weak Riemannian
-
8/11/2019 Analysis and geometry of shape spaces.pdf
209/220
H 0 -metric on Diff 0 ( N ) makes it into a true met-ric. We dene the stronger right invariant weakRiemannian metric by the formula:
GA ( h, k ) = N ( g( X, Y ) + A div g( X ) . div g( Y )) vol( g) .Theorem. For any distinct diffeomorphisms 0 ,1 ,the inmum of the lengths of all paths from 0 to 1 with respect to GA is positive.
Example We consider the groups Diff c( R ) or Diff( S 1 )with Lie algebras Xc( R ) or X( S 1 ) with Lie bracketad( X ) Y = [X, Y ] = X Y XY . The GA-metric
-
8/11/2019 Analysis and geometry of shape spaces.pdf
210/220
equals the H 1 -metric on Xc( R ), and we have:GA( X, Y ) = R ( XY + AX Y ) dx
= R X (1 2x ) Y dx,
ad( X ) = (1 2x )1 (2 X + X x )(1 A 2x )so that the geodesic equation in Eulerian represen-
tation u = ( t) 1 Xc( R ) or X( S 1 ) is tu = ad( u )u
=
(1
2x ) 1 (3 uu
2 Au u
Au u ) ,
u t u txx = Au xxx .u + 2 Au xx .u x 3 ux .u,
-
8/11/2019 Analysis and geometry of shape spaces.pdf
211/220
which for A = 1 is the Camassa-Holm equation ,another completely integrable innite dimensionalHamiltonian system. Here geodesic distance is ametric.
Virasoro-Bott group. Let Diff denote any of thegroups Diff( S 1 ), Diff c( R ) (diffeomorphisms with
-
8/11/2019 Analysis and geometry of shape spaces.pdf
212/220
compact support), or Diff S ( R ). Thenc : Diff Diff R
c(, ) : = 1
2 log((
) ) d log( )
= 12 log( ) d log( )
satises c(,1 ) = 0, c(Id , ) = 0, c(, Id) = 0,and is a smooth Hochschild group cocycle, i.e.,c(2 ,3 )c(12 ,3 )+ c(1 ,23 )c(1 ,2 ) = 0 ,called the Bott cocycle.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
213/220
The corresponding central extension group R cDiff, called the Virasoro-Bott group, is a regularLie group with operations
= + + c(, ) , 1 = 1
1for , Diff and , R .
The Lie algebra of the Virasoro-Bott Lie group isthe central extension R X of X, called the Vira-soro Lie algebra, with bracket:
-
8/11/2019 Analysis and geometry of shape spaces.pdf
214/220
X a
, Y b
= [X, Y ]( X, Y )
= X Y XY ( X, Y )
( X, Y ) = ( X ) Y =
X dY =
X Y dx =
= 12 det X Y X Y dx,is the Gelfand-Fuchs Lie algebra cocycle : X
X
R , which is a bounded skew-symmetric
bilinear mapping satisfying the cocycle condition
([ X, Y ], Z ) + ([ Y, Z ], X ) + ([ Z, X ], Y ) = 0 .
It is a generator of the 1-dimensional bounded Cheval-ley cohomology H 2 ( X, R ) for any of the Lie algebras
-
8/11/2019 Analysis and geometry of shape spaces.pdf
215/220
X = X( S 1 ), Xc( R ), or S ( R ) x .
We shall use the L2 -inner product on R X, whereX = X( S 1 ) , Xc( R ) , S ( R ) x :X , Y
b 0 := XY dx + ab.
-
8/11/2019 Analysis and geometry of shape spaces.pdf
216/220
a b 0 Integrating by parts we getad
X a
Y b
,Z c 0
=X Y XY
( X, Y ),
Z c 0
= (X Y Z XY Z + cX Y ) dx= (2 X Z + XZ + cX ) Y dx= Y
b, ad X
a Z c 0
, where
adX aZ
c =
2 X Z + XZ + cX 0
.
The H 0 geodesic equation on the Virasoro-Bottgroup (Ovsienko-Khesin):
d 3 h
-
8/11/2019 Analysis and geometry of shape spaces.pdf
217/220
u ta t
= ad ua ua = 3 uxu au xxx0 whereu ( t )a ( t )
= s( s) ( s)
.( t ) 1
( t ) s= t
= s ( s ) ( t ) 1
( s ) ( t ) + c(( s) ,( t ) 1 ) s= t=
t 1 t
txxx22x dx
Thus a is a constant in time and the geodesic equa-tion is hence the Korteweg-de Vries equation
u t + 3 uxu + au xxx = 0 .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
218/220
with its natural companions
t = u , t = a +
txxx
22x
dx.
I d k h h h i h i i L2 i
-
8/11/2019 Analysis and geometry of shape spaces.pdf
219/220
I do not know whether the right invariant L2 -metricon the Virasoro-Bott group has vanishing geodesicdistance?On Mondays I think: YESOn Tuesdays I think: NO. . .
-
8/11/2019 Analysis and geometry of shape spaces.pdf
220/220
At the end of last main lecture:Many thanks to the organizers (except one of them)for a great conference, and for the ne weather andgreat snow.