of the exterior unclassified ieieeit siope-e]hee · aa96 6b5 wisconsin univ madison mathematics...
TRANSCRIPT
AA96 6b5 WISCONSIN UNIV MADISON MATHEMATICS RESEARCH
CENTER FI 12/ION THE CHOICE OF THE EXTERIOR KNOTS IN THE B-SPLINE BASIS FOR A--ETC(U)DEC 80 J KOZAK OAAB29-BO-C-0041
UNCLASSIFIED MRC-TSR-2148 NL
sIopE-E]hEEIEIEEIT
MRC Technical Summary Report # 2148
ON THE CHOICE OF THE EXTERIOR KNOTS
IN THE B-SPLINE BASIS
FOR A SPLINE SPACE
~J. Kozak
CO \
Mathematics Research Center
University of Wisconsin-Madison
610 Walnut StreetMadison, Wisconsin 53706 AAR23 i9 8 i
December 1980
Received May 29, 1980
"' Approved for public releaseLLJ-.-3 Distribution unlimited
Sponsored by
U. S. Army Research Office Council for International
P. 0. Box 12211 Exchange of Scholars
Research Triangle Park Suite 300, Eleven DuPont Circle
North Carolina 27709 Washington, D. C. 20036
81 3 19 0824
S
ON THE CHOICE OF THE EXTERIOR KNOTS IN THE B-SPLINE BASIS TcT'
FOR A SPLINE SPACE
J. KOZAK __
Av:.I, -. ,. Codes
Technical Summary Report #2148 bist
Decemb.. 1980
ABSTRACT
The B-spline representation of a spline on some interval [a,b] requires
the introduction of additional knots outside ]a,b[ that have nothing in
common with the spline space itself. In this note, it is shown that locating
all these additional knots at the endpoints a and b of the interval
minimizes the matrix norm II.I1 of the matrix (X N )- where X are given00 1i j,k
linear interpolation conditions, and so is preferable when the B-spline
representation of a spline interpolant is to be constructed. Such a choice
usually also simplifies the algorithms. In particular, one is able to compute
stably the B-spline coefficients of a complete spline interpolant by Gauss
elimination without pivoting though the corresponding matrix fails to be
totally positive or diagonally dominant.
AMS(MOS) Subject Classification - 65DI0, 41A15
Key Words: B-spline basis, exterior knots, spline interpolation
Work Unit No. 3 - Numerical Analysis and Computer Science
Sponsored by the United States Army under Contract No. DAAG29-80-C-0041 andFulbright Grant No. 79-045-A.
SIGNIFICANCE AND EXPLANATION
The B-spline representation of a spline on some interval [a,b requires
the introduction of additional knots outside ]a,b[ that have nothing in
common with the spline space itself. Nevertheless, their choice influences
the numerical accuracy when the B-spline representation
of a spline approximant is to be constructed. In this note we prove that
locating all the additional knots at the endpoints a and b of the interval
minimizes the norm
i ,k, )
and is therefore preferable. Here X (Xi ) is an arbitrary sequence of
linear functionals with support in I except for the assumption that the
interpolation problem is correct, i.e. (A N nonsingular. The abovei j,k,t
choice usually also simplifies the algorithms. In particular, one is able to
compute in a stable way the B-spline coefficients of a complete spline
interpolant by Gauss elimination without pivoting though the corresponding
matrix fails to be totally positive or diagonally dominant.
The proof relies on the observation that
(N i,k,) = (N i,k,t)Qt on [a,b]
with T,t being two knot sequences that coincide in [a,b]. The matrix
Qt T for the particular choice of additional knots t described above turns
to be totally positive and its columns sum up to 1 , i.e. Qt T is composed
of (discrete) B-splines.
The responsibility for the wording and views expressed in this descriptive
summary lies with MRC, and not with the author of this report.
ON THE CHOICE OF THE EXTERIOR KNOTS IN THE B-SPLINE BASISFOR A SPLINE SPACE
J. KOZAK
1. The Result.
In a practical computation one is rarely able to make statements about the inverse of
a given matrix, particularly if the linear system to be solved depends on several free
parameters. This note is intended to demonstrate that in an important linear problem, that
of spline interpolation, the properties of a chosen basis allow us to draw an interesting
conclusion.
To start, let I -- [a,b] c R be a given real interval, partitioned by the sequence
a -: t k < tk+ 4 . t < tn : b, .)
k k+1 n n+1with ti i , all i and some integer 1 4 k 4 n. For the purpose of using B-splines,
the sequence is extended by
t 1 t2 (tk( tk tn+14 (tn+2 C .tn+k (1.2)_ n+k
To simplify the distinction between both parts of t :- (t ii. put
n+1 : k-1;n+kint(=) :- (ti) i-k' extt) :- (ti) i-1;i-n+2 * The collection of polynomial splines of
order k on I with knot sequence t is defined by
S k,tI) :- Ifftif t i+1 is a polynomial of degree < kJ mp t r) -0,r ( k - card(jlt, - ti}, sit i}
Any f e S k,(I) admits a unique B-spline representationn
t h p i (1.3)
that has proved very successful in practical computations. Here, (N i,k,) is the B-splinei~ktt
partition of the unity, i.e.
Ni,k,t(x) t (ti+k - t, t - x)+k-1
Sponsored by the United States Army under Contract No. DAAG29-80-C-0041 andFulbright Grant No. 79-045-A.
Consider now the following interpolation problem: for given interpolation conditions
X .:= (A) and prescribed numbers r := (r ) F Rn
find f E Sk,(I) such that
Xif = r i all i . (1.4)
There A is an arbitrary sequence of linear functionals with support in I except for the
assumption that the interpolation problem is correct, i.e.
f C S k,(I) and A = 0 , all i , implies f = 0
The particular representation (1.3) leads to the solution
n -;1:(a i), 1 A (1.5)
St
N. nwith At := (Xi Nj,k,)i,j=1 . We note that the solution f does not depend on ext(t)
but A and consequently a do. This fact can influence the numerical accuracy of thet
computed spline.
Cox [8] considered various ways of choosing the additional knots. As he concluded
from numerical evidence, in the case of general-purpose algorithms the coincident choice at
end points is preferable. In fact, he observed in [7] that the spectral condition number
of the matrix obtained in particular spline least-squares problems is considerably smaller
for
ext(t) = (t k t k... t k tn+1, tn+1 t n+) (1.6)
k-l k-l
compared with an equidistant and an average choice. We note also that in the book by
de Boor [5] the choice (1.6) is the rule.
Theorem 1.1. Let the knot sequences T,t satisfy
int(T) = int(t),
T.i i , i =
i ; t i = n+2,n+3,.oo,n+k
and let n ) 2(k-1). If the interpolation problem (1.4) is correct, then
UA I IA- 1 (1.7)
-2-
The inequality (1.7) displays the numerical advantage of using the ext(t) as defined in
(1.6). Also, following E2], consider the interpolation map
P: Rn - S ,t(I) :r Pr :- (X, Pr -r i , all i )
Then
n -
1PM :- sup (suplPr(x)l) = sup (eupl a i Ni,k,t(x)1) 4 IA 1 1
00E 41 xcI lirE 41 xcI i-I
(1.8)
and (1.6) gives the best bound of the form (1.8). It is also clear from (2] that A must
become singular, as T1+ - or T n - . We shall explicitly observe
1 k-1 tk+l-T r k- Tn+r-t n
A- Im Mx(. - , (1.9)- AtH. r-2 tk+i- T r-2 t n+r-t n
Our final observation concerns complete even order spline interpolation. Let k - 2m and
_I := (6l .6(1), ,6 (m-1) 6 ..*,* 6(m-1 1
,..,6n+ (1.10)
t 't t t t t tk k k k+1 n n+-1 n+1 I.
with 6 r)f - fr (t). It Is proved in (4) that the Gauss elimination without pivoting cant
be applied safely if the matrix is totally positive. Unfortunately for X given in (1.10)
the matrix is not totally positive. But (1.6) gives us At of the form
-3-
.... . .. .... ... fl - + +~ l + - -'i~' '
I~iII
m x x
xx xx 0
x x xx x
xx xx x
x x
xi
and we need to factor only the submatrix
rn-t, m+2, ",n-rn
tm4*, mn+2, ",n-rn
and this submatrix is totally positive.
-4-
Wf--:-
2. Its Proof.
Consider two knot sequences t , T that satisfy (1.1), (1.2). Assume that int ()
is strictly increasing. The general case will follow from continuity properties of divided
differences. Of course there exists a matrix Q such that
AT At Qt T* (2.1)
but it is not so obvious that we can find Qt T explicitly when int(T) - int(t). Recall
Marsden's identity [9]
kl n(y-x) k- = ,k,ty) N i,k,t(x) xy E 1. (2.2)
i=l _
Here
k-lt(2.3)
i,k,t t e+i
As observed in [1], (2.2) implies
kl nk-1 I
(y -1
x+
= (y- t +i + Ni,k,t (x) y , int(t) x I.
( 2.4 )
But (2.4) holds for an y 4 tk = a too, since then both sides vanish identically. Hence
k- n k-1 Kly Oyl o..,y k I(. - X) +
-I = [ [y0,y 1 , o ,,,y k 1 (°- xt)+i+Ni,k,t W on
(2.5)
under the assumption
ky (Y i C int(t) U ]-,tk I
Similarly, with y c int(t) u It n 1,_ n+1
X )+ = (yoy...,yk](X- k-1--ly .. ,k +I k +
n k-1
= O £kI (tx i + N i (x) on 1. (2.6)
One finds that (2.5), (2.6) are the identity [3, (5.10)], adjusted to the finite interval
I.
Lemma 2.1. Let n > 2(k-1), int(T) = int(t). Then
(N )n= (N ) Q (2.7)i,k,x i1 i,k,t i1I t T
with
-5-
____ - - - ,. =:*--~-- - - - --- --- ~~- -- r
a nnQt T {iJ'.tI i,J-1
anCT T r ; ( t )/, (t T , 1V.1 j V. k-1,j+k J Lj) i,k,t lt+k j-l,k+2,Tl +k
i
q _ ] -T) .j C i,k,t (t)/S CJ-,k+2,T (T) n-k+2 4 j V i 4 n,
i otherwise.
Proof. Assume for a moment that int(t) is strictly increasing. Choose 1 4 j V k-1.
Since n+l ) 2(k-I) + I - 2k-I we can use (2.5) for any such j to obtain
n k-1
Nj,k,T ( j+k - j )T .j+ J+k t L+i + Ni,k,t on I
and further
k-i
( - ) , (. - t )+)J+k J J+ ' 'Tj+k =+i
k-i
j+k r -t+i)-C j+k - ) J+k
r=j T (Tr- T 9
Er
k-i
J+k (T r " 4i
= 0 + (Tj.I - T - all ij+k j -~ J+k i~
HI (Tr T)"
r T C r E
The proof for n-k + 2 V j V n follows inthe same way from (2.6), and (2.7) obviously holds
for the remaining range of j . The proof is completed for a strictly increasing
int(t), but int (T) = int (t) simplifies gj, say for the range
I < 1. j 4 k-i , tok-i k-1
qiJ't 1 (T J+k - TJ)[T i+k' Ti+k+ .,T j+kIC T . - t)/ T C. - T)) , (2.8)-- ik =ki1 k=.
and the general case follows. U
-6-
2. Its Proof.
Consider two knot sequences t T that satisfy (1.1), (1.2). Assume that int ()
is strictly increasing. The general case will follow from continuity properties of divided
differences. Of course there exists a matrix Q such that
AT =At Qt T (2.1)
but it is not so obvious that we can find Qt T explicitly when int(T) = int(t). Recall
Marsden's identity [91
k-i n(y-x) = . ikt Ni,k,t(x) xy £ Ie (2.2)
i-i ' - -
Herek-l
"oi,k,t n (- - t£+i ) (2.3)
As observed in El], (2.2) implies
n k-1
k-i(y -x)+
I'- f (y -t£+) Nkt(x) , y E int(t) ,x e I.•
(y X) +i +i,k,t_
(2.4)
But (2.4) holds for an y 4 tk = a too, since then both sides vanish identically. Hence
ki n k-i
[yoy 1 ,.. .,y k I(. - x)+ = [ [yo, 1 ,. ..,y kT J c. - + )Ni,k,t (x) on Ii=I =1--
(2.5)
under the assumption
k
Similarly, with y c int(t) u )tn1, [,[y0,Yl, ,yk] + xI k-
k-l k-ily 0 'y I .... y k= y 0 ,y , ... ,YkI(x
n k-i
i=Iy ly' '" 1 k=I (t +i + Nik't (x) on (2.6)i= g1 _
One finds that (2.5), (2.6) are the identity (3, (5.10)], adjusted to the finite interval
I.
Lemma 2.1. Let n ) 2(k-1), int(T) = int(t). Then
(N n. = (N n= Q 12.7)i,k,T)=1 (i,k,t i=1 t T
with
-5-
... ,.
In order to investigate Q further we compute, for 1 4 1 4 j 4 k-1
t T
q ijt!J+k - J' "i~k~tt C k"PJ-1,k+2, (TI~
i j+k it x j i,k,t(tL+k/ "j-1,k+2,T (+k
j-1 !li (rt+k/
i i,k,t t1+k V j-1,k+1,T T+k) + --ik4t+k)/ ; k+1,r (T k
and consequently
n k-1 k-1I q I q P -i,k,t1(t+k)/P'k ,k+1,T (T+k
) wJ=1 qiJ , J i qij ,tT_ £"i
Since additionally int(T) = int(t) , then
k-111 (T£ tr )
2k-I r+iW r=1
X-i+k 2k 2
r-kr
r*£
k-1
k-1 i+k-1 T tr+i[ k'rk, l'** ', 2k 1 ] T C. - t2k-) - X 2k-I
r= 1 £fI k k (T - Tr
r=kr*1
=1-0 1
The following lemma summarizes this observation.
Lemma 2.2. Let n ) 2(k-1), and int(T) - int(t) . Thenn n
I= = 1 , all i .
From Lemma 2.1 we conclude
and to complete now the proof of the theorem it is enough to show that, for the
-7--
particular t and T as specified there,
Q ) 0 := (q ,t ) 0 , all i,j)
since then by Lemma 2.2 and invertible A -
RA-1t t
Also (1.9) is confirmed just by changing At and Qt T in the previous argument. It was
pointed out to me by de Boor (6] that Qt T is known to be totally positive if t is a
refinement of T on I, i.e. i ,t(t) - int (t), T 4 t , t ) tn. The following lemma_-- k 1, n42 > n+k*
indicates that this fact holds in the more general situation of the Theorem 1.1, and the
columns of Qt T deserve to be called discrete B-splines as in [3].
Lemma 2.3. Under assumptions of the Theorem 1.1
q j-iq ij, T t > 0 , all ij . (2.9)qij ,t 7 ij , t
Proof. Choose again 1 4 i 4 j 4 k-1. The assumption on T,t reads
Ti 4 ti r t k < Ti+k , 1 4 i < k-1 . (2.10)
From (2.8) we observe that (2.9) holds for j - i . Let j > i+1. The elements
qij,t Tare linear in any t. , i+11 4 Z k-1. Thus from (2.10)
sgn(q ) orij't T
sgn(qi., ) = tk
or after repeating the argument k-i-1 times
-8-
sgn(q ) " sgn(qij....it+1 tk-1
ti+1 or ,otk-1 or
tk tk
m m+1j sgn((Tj+k I j ETi+k Ti+k+1 ,o.*Tj+k - k ) / 1 (o - T V
for some m , J-i-1 4 m 4 k-i-i, and some j 4 Vj < VJ+1 ( -1 Vi+M < k-1 . For
m - j-i-1 the argument of sgn(*) in (2.11) vanishes, and it is enough to consider
m J-i . Then
[T i+ku T i+k~l '...T J+k]U t k )C l - )
m[T i+k, Ti+k+1,*'' T J+k
( tk / (*T I - 1w (k))
[TT ,T T..T+
i+k Ti+k+l' J+k z m+-+i-j(Z -t )
r-m-i+j-1 m-r
(T i+k' Ti+k+ 1 ,*' T J+kI z W) (W t k)r-O
( itj (J-i (m- 1 - r-I-m m-r + 0
~ ( I )(ii - v^)(J-i) 1 r (m+i-j-r)I k
for some w,z with Tj 4 7 V w V m+i ( Tk- I I TI+ k 4 , j+k . Thus the first
inequality in (2.91 is confirmed. The other follows similarly. U
Acknowledgement.
I thank Professor Carl de Boor for his careful reading and questioning of the
manuscript. His suggestions helped me very much in developing the final draft of the
paper.
-9-
References
1. C. de Boor, On calculating with B-splines, J. Approximation Theory 6 (1972), 50-62.
2. C. de Boor, On bounding spline interpolation, J. Approximation Th.ory 14 (1975),
191-203.
3. C. de Boor, Splines as linear combinatins of .-splines. A survey, MRC Technical
Summary Report No. 1667, 1976.
4. C. de Boor, A. Pinkus, Backward error analysis for totally positive linear systems,
Numer. Math. 27 (1977), 485-490.
5. C. de Boor, A Practical Guide to Splines, Springer-Verlag, 1978.
6. C. de Boor, Private communication.
7. M. G. Cox, Numerical methods for the interpolation and approximation of data by
spline functions, Ph. D. Thesis, 1975.
8. M. G. Cox, The incorporation of boundary conditions in spline approximation problems,
NPL NAC Report No. 80, 1977.
9. M. Marsden, An identity for spline functions and its application to variation
diminishing spline approximations, J. Approximation Theory 3 (1970), 7-49.
JK/db
-10-
...............................-!
SECURITY CLASSIFICATION OF THIS PAGE (hsn De. En CsREAD IWSTRSCTIO14SREPORT DOCUMENTATION PAGE BEFORE COMPLETING FORM
1. REPORT N MBER 2. GOVT ACCESSION NO 3. RECIPIENT'S CATALOG NUMBER
2148 (N malej4.:TITL - S. TYPE OF REPORT & PERIOD COVERED
,9N THE CHOICE OF THE EXTERIOR NOTS IN THE Summary Report - no specificB-SPLINE JASIS FOR A SPLINE SPACE - reporting period
6. PERFORMING ORG. REPORT NUMBER
7( I 6. CONTRACT OR GRANT NUMBER(&)
/ I ia DAAG29-8 -C-#41'J-' 79-045-A
9, PERFORMING ORGANIZATIO E AND ADDRESS 10. PROGRAM ELEMENT. PROJECT, TASK
Mathematics Research Center, University of - Numerical Analysis and
610 Walnut Street Wisconsin
Madison. Wisconsin 53706 Computer Science
II. CONTROLLING OFFICE NAME AND ADDRESS
See Item 18 below I. NUMBER OF PAGES10
14. MONITORING AGENCY NAME & ADDRESS(IIdiflerent from ControillnS Office) 15. SECURITY CLASS. (of thls report)
UNCLASSIFIED"S. DECL ASSI FICATION/DOWNGRADING
Al t M a I iSCHEDULE
16. OISTRGUTION STATEMENT (of Ihis Report) -1
Approved for public release; distribution unlimited. sI
17. DISTRIBUTION STATEMENT (of the abst,act entered in Block 20, it different trmiu Report)
IS. SUPPLEMENTARY NOTES
U. S. Army Research Office Council for International
P. 0. Box 12211 Exchange of Scholars
Research Triangle Park Suite 300, Eleven DuPont CircleNorth Carolina 27709 Washington, D. C. 20036
19. KEY WORDS (Continue on reverse side it necessary ind identifv by block nhumber)
B-spline basis, exterior knots, spline interpolation
20 AUDST itACT (Continue on reveres side if necessary and identify by block number)
The B-spline representation of a spline on some interval [a,b] requires the in-troduction of additional knots outside ]a,b[ that have nothing in common with thespline space itself. In this note, it is shown that locating all these additionalknots at the endpoints 1 a and b of the interval minimizes the matrix norm II .II
). the matrix (XiN ) where i. are given linear interpolation conditions, andso is preferable wen the B-spline representation of a spline interpolant is to beconstructed. Such a choice usually also simplifies the algorithms. In particulalone is able to compute stably the B-spline coefficients of a complete spline
FORM i (continued)DO I 1473 EDITION OF I NOV Of15 OBSOLETE UNCLASSIFIED
SECURITY CLASSIFICATION OF THIS PAGE (When Data SEn rta
'- -___... ........ "- .., .,-. -- -o . .
20. Abstract (continued)
interpolant by Gauss elimination without pivoting though the corresponding matrixfails to be totally positive or diagonally dominant.
Ii
,DA7FE
,ILMED,