Date post: | 22-Dec-2015 |
Category: |
Documents |
View: | 214 times |
Download: | 0 times |
Subdivision Analysis via JSR
We already know the z-transform formulation of schemes:
To check if the scheme generates a continuous limit curve
( the scheme is ) , we have to check if the difference
scheme is contractive.
We hereby present the ‘joint spectral radius’ analysis (JSR) of schemes, and examine the relation between the two methods. Let us consider a univariate scheme with support [0,s].
)()()( 21 zFzazF mm
0C
mi
iij
mj faf
21 i
s
ii zaza
0
)(
)1/()()( zzazb
2005 SubdivisionSummer School
2005 Subdivision Summer School
Since the support of the basic limit function is also [0,s],
and the limit function is , the values
of on [0,1] are fully determined by
f ts ffu ),...,( 0
00
10
)()( 0 ixfxf
ii
101-2-..1-
. 00
01
02
02
01
s
fffff ss
Since the support of the basic limit function is also [0,s],
and the limit function is , the values
of on [0,1] are fully determined by
The values on [0,1/2] are determined by
The values on [1/2,1] are determined by
f ts ffu ),...,( 0
00
10
)()( 0 ixfxf
ii
101-2-..1-
. 00
01
02
02
01
s
fffff ss
ts ffu ),...,( 1
01
10,0
21
21
21
10
11
12
12
11
0-1-..
.
s
ss fffff
ts ffu ),...,( 1
11
21,0
10-..
.
21
21
22
11
10
11
13
12
s
ss fffff
Let us recall the matrix representation of subdivision schemes, and use this representation to compute limit values of the subdivision process, and to analyze the convergence and the smoothness of the limit function.
We present everything for a mask with support [0,4], i.e.
the scheme coefficients are
and the symbol of the scheme is .
The scheme is:
The key model is the scheme generating the cubic B-spline:
43210 ,,,, aaaaai
i i zaza
4
0)(
.
,
131112
241201
2
mj
mj
mj
mj
mj
mj
mj
fafaf
fafafaf
.8
)1()(,
8
1,
2
1,
4
3,
2
1,
8
1 4
43210
zzaaaaaa
The subdivision matrix S: Infinite – 2 slanted
00000
000000
00000
000000
00000
000000
00000
000000
420
31
420
31
420
31
420
31
aaa
aa
aaa
aa
aaa
aa
aaa
aa
S
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m Sf
f
f
f
f
f
f
f
f
aa
aaa
aa
aaa
aa
aaa
aa
aaa
aa
f
f
f
f
f
f
f
f
f
f
.000000
00000
000000
00000
000000
00000
000000
00000
000000
2
1
0
1
2
3
4
5
13
024
13
024
13
024
13
024
13
13
12
11
10
11
12
13
14
15
1
The values of on [0,1] are determined by
The values on [0,1/2] are determined by
The values on [1/2,1] are determined by
f
101-2-..1-
. 00
01
02
02
01
s
fffff ss
ts ffu ),...,( 1
01
10,0
21
21
21
10
11
12
12
11
0-1-..
.
s
ss fffff
ts ffu ),...,( 1
11
21,0
10-..
.
21
21
22
11
10
11
13
12
s
ss fffff
ts ffu ),...,( 0
00
10
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m Sf
f
f
f
f
f
f
f
f
aa
aaa
aa
aaa
aa
aaa
aa
aaa
aa
f
f
f
f
f
f
f
f
f
f
.000000
00000
000000
00000
000000
00000
000000
00000
000000
2
1
0
1
2
3
4
5
31
420
31
420
31
420
31
420
31
13
12
11
10
11
12
13
14
15
1
The sub-matrix : 0At
st
s ffff ),...,(),...,( 10
11
00
01
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m Sf
f
f
f
f
f
f
f
f
aa
aaa
aa
aaa
aa
aaa
aa
aaa
aa
f
f
f
f
f
f
f
f
f
f
.000000
00000
000000
00000
000000
00000
000000
00000
000000
2
1
0
1
2
3
4
5
31
420
31
420
31
420
31
420
31
13
12
11
10
11
12
13
14
15
1
The sub-matrix : 1A ts
ts ffff ),...,(),...,( 1
11
20
00
1
The vector of initial values determines
the limit values at [0,1].
determines the values at [0,1/2] while determines the
values at [1/2,1] .
And so on, while
determines the limit function on the interval ,
determine the limit function on the left half and the right half
of respectively. It thus follows that at any point in [0,1],
the limit value of a convergent subdivision, starting with initial values,
is given by
tffffu ),,,( 00
01
02
03
0
00uA 0
1uA
tmj
mj
mj
mj
m ffffu ),,,( 123 ]2)2/1(,2[ mm jjI
mm uAuA 10 ,I
},1,0{,.....,]1,0[ 4321 kiiiiixx0f
....))(),(),(),(( 0
1234uAAAAxfxfxfxf iiii
t
First, a scheme must reproduce constants, therefore ,
Having that, we should make sure that the components in the subspace complementary to the constant eigenvector tend to 0.
Let us show how to check it, and let us also try to understand the relation to the difference scheme analysis.
The idea is to represent the operators and in another basis, the vectors comprising this basis are (e.g.) the columns of the following matrix:
0C
1A
1111
0111
0011
0001
V
.1eigenvaluewith1,1,1,1reigenvectohaveand 10tAA
0A
In the new basis, the two operators are:
Since is an eigenvector of and it follows that
Note that determine the behavior of the last 3 components in the new basis.
VAVBVAVB 11
101
0 ,
t)1,1,1,1( 10 AA
0
0
0
0
0***1
QB
1
1
0
0
0***1
QB
10 andonly QQ
Recalling the products , which in the
new basis appear as
The last three coordinates of this product will converge to
zero if the joint spectral radius is less than 1.
Each matrix has its own spectral radius. The
joint spectral radius is defined as :
If then the last three components of any vector of the form (#) will tend to zero.
0
1234... uAAAA iiii
#,... 0100
1234uVvvBBBB iiii
10 , QQ),( 10 QQ
1),( 10 QQ
mjiii
m
iQQQQQm
/110 }})1,0{:||...(max{||suplim),(
21
We already know that a scheme is convergent to a limit if the differences of the generated values tend to zero. In the new basis, namely, the columns of , it is clear that differences within the first column are zero. The differences within the other columns tend to zero simply because the corresponding coefficients tend to zero.
The number tells us how fast the differences are decaying to zero, and from it we can derive the Holder exponent of the limit function:
0C
1111
0111
0011
0001
V
V
),( 10 QQ
),(log 102 QQ
The choice of and the relation to the difference scheme.
As an exercise, show that with the
specific choice of as :
the two matrices
are just the two matrices generating
the limit values on [0,1] for the difference scheme related with
the generating polynomial
In fact, any other completion of the constant vector into a basis may be used here.
1111
0111
0011
0001
V
10 and QQ
V
V
.)1/()()( zzazb
The analysis of higher smoothness is very similar. First we have to show that if a scheme is , it must reproduce polynomials up to degree k. That is,
To show it we consider eigenvectors of the infinite matrix S
kC
}:)({ Zjjxspank
00000
000000
00000
000000
00000
000000
00000
000000
420
31
420
31
420
31
420
31
aaa
aa
aaa
aa
aaa
aa
aaa
aa
S
The non-zero eigenvalues of the infinite matrix S, due to its two-slanted structure, are just the non-zero eigenvalues of the finite matrix marked below.
00000
000000
00000
000000
00000
000000
00000
000000
420
31
420
31
420
31
420
31
aaa
aa
aaa
aa
aaa
aa
aaa
aa
S
tuuuuuu )...,,,,,,..(. 21012 Let be an eigen-vector of S,
and let denote the limit function of the subdivision process, starting with . It thus follows that
Exercise: Show that if then
Also, the entries of the eigen-vector are of the form
It can further be shown that if the scheme is non-degenerate, then all the monomials up to degree k are eigen-functions of the subdivision scheme.
,uSu uSfu
u
.)2()( xfxf uu ,0,0)0()( krf r
u
.)(,2 ru
r xxfand
01
1 ... ciciu rr
ri
In particular ,
To verify that a given scheme is , we first check that it satisfies the necessary conditions. I.e., its symbol has the factor , or, in the matrix language, it has the polynomial eigenvectors up to degree k.
Then, we have to show the decay to zero of all the
differences of the k-th order divided differences
One way is by showing that the scheme is contractive .
The other way
.0,2seigenvaluewith
rseigenvectopolynomialdegreehaveand 10
kr
rAAr
1)1( kz
1)1/()(2)( kk zzazb
, . . . why do we need the other way ?
kC
Here again we represent the operators and in another basis, where the first k+1 vectors in the basis are the polynomial
vectors (for a scheme with support [0,s]):
The differences of the k-th order divided differences within these vectors are all zero.
And, since these vectors span the polynomial eigenvectors of and , the operators now take the form:
.0,)1,2,...,)1(,()( kjssv tjjjj
0A
0A
1A
1A.2,1,
0000
0000
0000
0000****
****
****
****
000
*00
**0
***1
21
21
i
Q
B
i
ik
.2,1i,
0000
0000
0000
0000****
****
****
****
000
*00
**0
***1
i
21
21
i
Q
B k
To show that the scheme is it remains to verify thatkC
kQQ 2),( 10
kQQ ),(log 102 Holder exponent of the k-th derivative of the limit function:
An example : The 4-point scheme for w=1/16
0A 1A
V
m
mjiii
m
iQQQQQm
/110 }})1,0{:||...(max{||suplim),(
21
mjiii iQQQ
m
/1}})1,0{:||...(max{||:ofGraph21
It can be shown here that .25.0),( 10 QQ
Limit values at grid points
To represent the curve or surface at the m-th refinement level, we prefer to use the limit values and derivatives . Can we ?
We would like to use only values at the m-th level .
It is enough to show how to do it for the 0-th level : Using
)()( 0 ixfxfi
i
.0,,)()( )(0)( kZjijfjfi
i
Therefore, it is enough to know the values of the basic limit function and its derivatives at the integers!
Zjijfjfi
i ,)()( 0
2005 SubdivisionSummer School
2005 Subdivision Summer School
Here we use the refinement equation:
.],0[)(supp,)2()( sjxaxj
j
.0,
)4(
)3(
)2(
)1(
)0(
2
)4(
)3(
)2(
)1(
)0(
0000
00
00
0000
)(
)(
)(
)(
)(
)(
)(
)(
)(
)(
4
234
01234
012
0
k
a
aaa
aaaaa
aaa
a
The values are left eigenvectors of the subdivision matrix S
.0,0,)2(2)( )()( ksijiaij
j
Example 1 : The limit values of the cubic B-spline scheme.4
81 )1()( zza
;
0
0
0
0
0000
00
00
0000
61
32
61
61
32
61
81
43
21
81
81
21
43
21
81
81
21
43
81
0
0
0
0
0
0
0000
00
00
0000
21
21
21
21
21
81
43
21
81
81
21
43
21
81
81
21
43
81
Normalization : Using the polynomial reproduction property
.0,!)()( kjjj
Example 2 : The limit derivatives of the 4-point scheme.
0
0
0
0
0
0
000000
00000
1000
010
0001
00000
000000
21
21
21
21
21
21
21
21
21
21
21
21
21
w
w
w
w
w
ww
www
wwww
www
ww
w
)()(2
1
41
2)2(' 2211
mi
mi
mi
mi
mm ffwff
wif
Approximation order and Quasi-interpolation
A local method has approximation order n+1 if it is exact on
2005 SubdivisionSummer School
2005 Subdivision Summer School
:n
kn
k
k
n jhxk
jhfxThjjhx )(
!
)()( and,])1(,[Let :Proof
0
)(
bounded be let and,)R(,)(Let 1n0 Cfihffi
.,)()()( and,],[)supp(with i
n pxpixipsr
.)(|)(-)(|Then i
10 ni hOihxfxf
ii
0 |)()]()([)(-)(| |)(-)(| ihxihTihfxTxfihxfxf nni
.)(|)()(| termsfor these and
],1[for only 0)(])1(,[1
n
n hOihTihf
srijihxhjjhx
If the symbol of a convergent scheme has the factor then
1)1( nz)}({ jxspann
.,)( ,i.e.,Let n0
n0 pipff i
n000 )})({(,)())(( jfSZjijfjfS
ii
To see this let us start with an initial polynomial data
There exists a scheme transforming into
kn f1 11 kn f
n0
n1 .,0
fS
kfkf kkn
If the symbol of a convergent scheme has the factor then
1)1( nz)}({ jxspann
.)())})(({())((but,)})({( 0n
0 jpjipSjfSjfS
.)())})(({())((but, Also, 0n
0 xpxipSxfSfS
Interpolatory subdivision schemes reproduce polynomials. Non-interpolatory scheme usually do not.
.,)( ,i.e.,Let n0
n0 pipff i
n000 )})({(,)())(( jfSZjijfjfS
ii
.,,})())})(({({haveweYet 1 npxpxipS
Example: For the cubic B-spline scheme
We would like to find a local operator Q such that
;}{;}{;}{;1}1{ 333122 xxxSxxSxxSS
I.e., Q is the inverse of nS on
nppQpS ,
The ‘method’ will be local and exact on nQS
1:where,,On RRITn
)}{;}{exampletheRecall( 333122 xxxSxxS
.since
,onthen ,... Define1n
nn2
RITQ
ITQRRRIQ
.)( n ppQpS
))((;)())(( 000 jiToeplitzTfTjififSj
j
Constructing Q
Example : For the cubic B-spline scheme ( n=3 )
)(
))(]([))((016
10310
1610
0016
10320
1610
iiii
iii
ffff
ifRIfffiTf
.)( 3 ppQpS
016
10340
161
016
10310
161000 )())(]([))((
iii
iiii
fff
ffffifRIiQf
.)1(;)0(;))(()())(( 61
32000 ifTjififS
jj
016
10340
16100 ))(( iiii fffiQfg
Approximation order: Given values
This suggests ‘lifting’ the given data points, or control points, by the quasi-interpolation operator Q before applying subdivision.
That works if the control points are taken on a ‘uniform grid’ on a smooth curve (surface)
)R(,)( 40 Cfihffi
.|)()(| 1n0 chxfxQfS
.|)()(|,Then 40 chxfxgS
.)(|)()(|while 20 hOxfxfS
)R(,)(:casegeneraltheIn 1n0 Cfihffi
Subdivision for Surfaces
2005 SubdivisionSummer School
2005 Subdivision Summer School
Refinement on regular rectangular meshes
Refinement on regular triangular meshes
Refinement on regular triangular meshes
Can we use a univariate scheme on each of the three directions?
Tensor Product Refinement
A univariate refinement along one direction
A univariate refinement along the second direction
MLK
cba
zzzzz
zzzzz
zzz
zzz
yxyx
zczbz
MLK )1()1()1(),a(z
)1()1()1(),a(z
:schemes splineBox
)1()1(),a(z
)1)(1(),a(z
:schemesproduct Tensor Some
)()(),(
)()(),a(z
:productTensor
212124
21
221
22
2116
121
22
214
121
2121
2121
The tensor product bi-cubic spline scheme with the symbol:
64
)1()1(),(
42
41
21
zzzza
41
41
41
41
161
161
161
161
161
161
161
161
169
83
83
83
83
323
323
323
323
641
641
641
641
The butterfly scheme on a grid:
12
1
21
21 w-
w-
w-
w-
w-
w-
w-
w-
w-
w- w-
w-
w2
w2
w2
w2w2
w2
21
21
21
The averaged butterfly scheme (w=1/16):
1
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
21
21
21
21
165
165
165
165
Analysis of bivariate schemes on regular grids
Let us consider a biivariate scheme
with the symbol
The generating functions
satisfy
),(),(),( 22
2121211 zzFzzazzF mm
21
221
21 21),(
,21 ),( ii
Zii
miim zzfzzF
21
21
21 2121ii
)σ(,ii,ii zza),za(z
a
mii
iiijij
mjj faf
212
21
221121 ,),(
2,21
,
We have four subdivision rules, according to the parity of
and the symbol should satisfy
Using the Laurent polynomial formulation
starting with we have
mii
iiijij
mjj faf
212
21
221121 ,),(
2,21
,
21
21
21 2121ii
)σ(,ii,ii zza),za(z
a
),( 21 jj
,1),( 210 zzF
,),(),(),( 22
2121211 zzFzzazzF mm
,),(),(),(),(),(11 2
221
42
41
22
212121
mm
zzazzazzazzazzFm
21
221
21 21),(
,21 ),( ii
Zii
miim zzfzzF
Plot coefficients’ array?
.0)1,1()1,1()1,1(;4)1,1( aaaa
An example . :
This scheme satisfies the necessary conditions for convergence. Does it converge to a continuous limit? How to check it?
Let us try to find a difference schemes :
If the symbol has the factor , we have a difference scheme transforming differences in one direction from one level
to the next.
)1( 1z
To verify convergence to a continuous limit we have to show that differences in two independent directions tend to zero.
In the example above, we do have a nice difference scheme for differences in the diagonal direction. Let us check it!
)1()1(),( 212
221
121
21
21 zzzzzza
),()1(1
),(),()1( 2
221
21
1
212111 zzFz
z
zzazzFz mm
.0)1,1()1,1()1,1(;4)1,1( aaaa
),()1(1
),(),()1( 2
221
22
21
21
2121121 zzFzz
zz
zzazzFzz mm
Hence, is the symbol of the
scheme taking differences into differences of the
same kind in the next level.
}{ 1,1,m
jimji ff
}{ 11,1
1,
mji
mji ff
222
112
121
21 )1(),( zzzzb
2141
221
1212
2812
181
21
21 ),( zzzzzzzzb
41
,2,22
1
,2,22
1
,2,124
3
,2,2 ||;||;||;||
jiji
jiji
jiji
jiji bbbb
Therefore, differences in the diagonal direction tend to zero .
)1()1(),( 212
221
121
21
21 zzzzzza
The symbol of the butterfly scheme:
12
1
21
21 w-
w-
w-
w-
w-
w-
w-
w-
w-
w- w-
w-
w2
w2
w2
w2w2
w2
21
21
21
2212
212121
1212
11
12
11
12
11
22
11
12
2121
12
112121212
121
2244412
2244422),(
)],(1)[1)(1)(1(),(
zzzzzzzz
zzzzzzzzzzzzzzc
zzzzcwzzzzzzaw
The symbol of the the averaged butterfly scheme:
1
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
321
21
21
21
21
165
165
165
165
)],()1()1(
)],(),([),(
212
22
121
21
12121
21
zzuzz
zzazzazza ww
Theorem 1.
Theorem 2.
.),()1)(1)(1(),( symbol thehave Let 21212121 zzuzzzzzzaSa
.convergent are
,3,2,1,2 symbols with theschemes theif functions
limit generatesThen above. as symbol thehave Let 1
ib
CSS
i
aa
e.contractiv are
schemes thoseof twoifcheck enough to isIt e.contractiv are
,1
),(),(;
1
),(),(;
1
),(),(
masks with theschemes difference theiff convergent is
21
21213
2
21212
1
21211 zz
zzazzb
z
zzazzb
z
zzazzb
Sa
.convergent are schemes thoseof twoifcheck enough to isIt
,),()1)(1)(1(),( symbol with For 21212121 zzuzzzzzzaSa
2212
212121
1212
11
12
11
12
11
22
11
12
2121
12
112121212
121
2244412
2244422),(
)],(1)[1)(1)(1(),(
zzzzzzzz
zzzzzzzzzzzzzzc
zzzzcwzzzzzzaw
For the butterfly scheme on a regular mesh
e.contractiv be tohave
,),()1(2;),()1(2;),()1(2
symbols following with theschemes the is ifcheck to
2121212211
1
zzuzzzzuzzzuz
CSa
)],(1)[1(),( 21121 zzcwzzzq
The scheme related to the following symbol should be contractive
21
21
21 2121Let ii
)σ(,ii,ii zzq),zq(z
q
))) consider Next we 22
212121 ,zq(z,zq(z,z(zq
.1|8||81||||| out that it turns
)||(max|||| norm theuse We,
2,21,0
wwwS
qS
q
Zjijlik
lkq
.08.001|||| out that it turns
)||max|||| norm theuse weNow,
4,43,0
wS
qS
q
Zjijlik
lkq
. cubics reproduces.for is particularIn
.is symbols other two for the holds same The
16116
11
1
aa
a
SwCS
CS
w
w
What if the symbol is not factorizable?
1 .We may use the JSR approach. Here one has to check the joint spectral radius of four local matrices.
2 .If the symbol satisfies the necessary conditions for convergence there still exists a difference scheme, but in general it is a matrix scheme. I.e., a scheme of the form where
For convergence, this matrix scheme has to be contractive .
,21
221
221121 ,),(
2,21
,mii
iiijij
mjj dBd
matrices.22 are}{and ,1,,
,1,,
jim
jimji
mji
mjim
ji Bff
ffd
.0)1,1()1,1()1,1(;4)1,1(satisfy symbol Let the aaaa
We want to show the existence of a difference scheme:
),()]1)(,()1)(,([),()1( 22
21
2221
21212111 zzFzzzczzzbzzFz mm
.),(),()1(),()1( know We 22
212112111 zzFzzazzzFz mm
.)1)(()1,(,)1)(()1,( 21121
21111 zzdzdzzdzd
.0)1,1( have we,),()1(),( Denote 21121 dzzazzzd
.1for 0)]()1()()1[(2
)1(),( 2122112
21
21
zzdzzdzz
zzd
. ),()cz-(1 )]()1()()1[(2
)1(),( 21
22122112
21
21 zzzdzzdzz
zzd
The existence of a matrix difference scheme
)1)(()1,( 21111 zzdzd
An example of a matrix difference scheme
)1()1(),( 212
221
121
21
21 zzzzzza
.)1()(;)3()( 218
112
218
111 zzdzzd
;),()1(),( 21121 zzazzzd
)1)((
)1)((),()1(
221616
3416
3165
2188448
381
211
2212121
22121
31
211
z
zzzazzzzzzz
zzzzzzz
Together with the difference scheme for the y-direction (using the symmetry) we have contraction for two independent directional differences, hence, the scheme is convergent.
Thank you!