Univariate spline quasi-interpolants and applications to numerical analysis
Paul Sablonnière INSA and IRMAR, Rennes
Abstract
We describe some new univariate spline quasi-interpolants on uniform partitions of bounded intervals. Then we give some applications to numerical analysis: integration, differentiation and approximation of zeros. AMS classification: 41A15, 65D07, 65D25, 65D32.
1 Introduction
Univariate spline quasi-interpolants (abbr. QIs) can be defined as operators of the form
where
is the B-spline basis of some space of splines, say of degree
, on a bounded interval
endowed with some partition
of
in
subintervals. We denote by
the space of polynomials of total degree at most
. In general we impose that
is exact on the space
, i.e.
for all
. Some authors impose further that
is a projector on the space of splines itself (see e.g.[6] , [11] ,[12] ). As a consequence of this property, the approximation order is
on smooth functions,
being the maximum steplength of the partition
. The coefficients
are local linear functionals which are in general of one of the following types:
(i) differential type :
is a linear combination of values of derivatives of
, of order at most
, at some point in
(see e.g. [1] , [2] ). The associated quasi-interpolant is called a differential quasi-interpolant (abbr. DQI). (ii) integral type :
is a linear combination of weighted mean values of
, i.e. of quantities
where
can be, for example, a linear combination of B-splines (see e.g. [1] , [2] , [18] ). The associated quasi-interpolant is called an integral quasi-interpolant (abbr. iQI).
(iii) discrete type :
is a linear combination of discrete values of
at some points in the neighbourhood of
(see e.g. [3] , [11] , [12] , [14] ). The associated quasi-interpolant is called a discrete quasi-interpolant (abbr. dQI).
The main advantage of QIs is that they have a direct construction without solving any system of linear equations. Moreover,they are local, in the sense that the value of
depends only on values of
in a neighbourhood of
. Finally, they have a rather small infinity norm, so they are nearly optimal approximants.
In this paper, we only consider dQs, neither DQIs nor iQIs. We also restrict our study to splines defined on uniform partitions of
. Our aim is to give explicit formulas for dQIs of degrees
and some applications to three classical problems in numerical analysis: approximate integration and derivation and location of zeros of functions. The paper is organised as follows. In sections 2, we recall some facts about splines and quasi-interpolants. In section 3, we describe dQIs of degrees
and we give their infinity norms and their approximation orders. In sections 4 and 5, we give the associated quadrature formulas and some numerical examples. In section 6, we give the differentiation matrices for quadratic and cubic splines with numerical examples. Finally, in section 7, we show, on a simple polynomial example, how quadratic dQIs can be applied to the location of zeros of functions.
2 Spline spaces on uniform partitions and dQIs
For
, we denote by
the space of splines of degree
and class
on the uniform partition
with meshlength
.
A basis of this space is
, with
. With these notations,
, and
is the set of the
interior knots in the support of
. As usual, we add multiple knots at the endpoints:
and
. We recall the represention of monomials in terms of symmetric functions of knots in
[6] ,[19] .
In particular, the Greville points
are the coefficients of
and the vertices of the control polygon of
are the control points
. The Schoenberg-Marsden operator is the simplest discrete quasi-interpolant which is exact on the space
:
—————A discrete quasi-interpolant (abbr. dQI) of degree
is a spline operator of the form:
whose coefficients
are linear combinations of values of
on either the set
( for
even) or on the set
(for
odd), where
Therefore, for
even, we set
, and for
odd, we set
. Moreover,
is exact on
:
For the construction of dQIs, i.e. for the determination of functionals
, the exactness of
on
amounts to solve a system of linear equations for interior B-splines and a finite number of specific linear systems for boundary B-splines. The determinants of these systems being Vandermonde determinants, there is existence and unicity of dQIs with the above assumptions (see also [3] for more general cases). For the sake of completeness, we give below complete formulas for degrees
. Moreover, we give exact values or upper bounds of
and their approximation order on smooth functions. Actually, it is well known (see e.g. [6] , chapter 5) that for any subinterval
, and for any function
where the distance of
to polynomials is defined by
Here, as usual,
. Therefore, for
smooth enough, e.g.
, this implies that
.
3 Discrete Quasi-Interpolants of degrees
3.1
Quadratic dQI
For the
-quadratic dQI
, the coefficient functionals are easy to compute (details are given in [15] ,[16] ):
,
,
,
, and for
The exact value
has been computed in [16] . Therefore, for
for example, we have the following error estimates
3.2
Cubic dQI
For the
cubic dQI
, the coefficient functionals are respectively:
,
,
, and for
As
and
for
, we obtain the upper bound
. It is possible to improve that result by writing the operator in the ”quasi-Lagrange” form:
where the fundamental functions are linear combinations of B-splines, e.g. for
,
. It is well known that
is equal to the Chebyshev norm of the Lebesgue function:
In each interval of the uniform partition,
is bounded above by the cubic polynomial whose Bernstein-Bézier (abbr. BB-) coefficients are sums of absolute values of BB-coefficients of fundamental functions. This allows to see that the maximum of
is attained in the interval
and we obtain:
From that we deduce for
for example, we have the following error estimates
3.3 Quartic dQI
For the
quartic dQI
, the coefficient functionals are respectively:
,
,
,
,
,
,
, and for
Let us give some details on the computation of functionals
. As
for
, we determine the five coefficients of the discrete functional
as solutions of the three corresponding linear systems (
) of
linear equations
They have the same Vandermonde determinant
since the
are distinct. Therefore they have unique solutions. The same technique is applied to the computation of other coefficient functionals.
As
,
,
,
for
, we can conclude that
and that for
for example, we have the following error estimates
3.4 Quintic dQI
For the
quintic dQI
, the coefficient functionals are respectively:
,
( symmetric formulas for
), and for
:
As
,
,
, and
for
, we deduce that
Using a similar technique as for cubics, we find that
Therefore, for
for example, we have the following error estimates
4 Application to numerical integration
Newton-Cotes formulas are obtained by integrating interpolation polynomials (see e.g.
[4] ,[7] ,[9] ). In the same way, integrating spline quasi-interpolants give interesting quadrature formulas (abbr. QF) which are easily deduced from the above computations. We use the notations
As
and
are known explicitly, we can compute the following quadrature formulas. Moreover, as QIs give the best approximation order, we can conclude that
for
, where
is the meshlength. Moreover, as for Newton-Cotes formulas, we get a higher approximation order for even degrees.
(i) QF for quadratics
Error : for
,
This result is proved in [16] .
Error for Simpson:
By comparing the two above errors, we see that the linear combination (extrapolation):
is such that
.
(ii) QF for cubics
Error:
for
. Numerical experiments show that this formula is not as good as the preceding one.
(iii) QF for quartics
Error:
for
. This is a remarkable formula, which can be compared to the Newton-Cotes formula of the same order. Numerical experiments show that the error for the former QF has also the opposite sign of the error for the latter, as in the quadratic case. The proof will be given elsewhere.
(iv) QF for quintics
Error:
for
. Numerical experiments show that this formula is not as good as the preceding one.
5 Numerical examples
We compare numerical results on QF applied to the computation of
(i) QF/dQI degrees 2 and 3
,
for
Simpson's QF
for
Example1:
| |
| |
| |
| |
Example 2:
| |
| |
| |
| |
(ii) QF/dQI degree 4
for
.
Newton-Cotes QF of degree 4:
for
.
Example 1:
| |
| |
| |
| |
Example 2:
| |
| |
| |
| |
(iii) QF/dQI degree 4:
for
.
QF/dQI degree 5:
for
, Example 1:
| |
| |
| |
| |
Example 2:
| |
| |
| |
| |
6 Application to numerical differentiation
Differentiating interpolation polynomials leads to classical finite differences for the approximate computation of derivatives. Therefore, it seems natural to approximate derivatives of
by derivatives of
as long as it is possible, i.e. up to the order
. The general theory will be developed elsewhere. Here we only give results for the first derivative and
. We evaluate
at points
for
even and at points
for
odd.
(i) Differentiation matrix for quadratics The derivation matrix
is defined as follows: setting
for the vector with components
and
for the vector with components
, we simply write:
(ii) Differentiation formula for cubics The derivation matrix
is defined as follows: setting
for the vector with components
and
for the vector with components
, we obtain:
(iii) Some numerical results Again we use the two functions
and
on the interval
. For
, we set
where
(resp.
) for
(resp.
) and
where
is the classical centered approximation of
of order
(with standard modifications at the endpoints).
For quadratics, we obtain the following results
| |
| |
| |
| |
| |
We see that the orders are all
. However, the errors for the derivatives of the quadratic QI (
and
) are between
and
times less than the errors for the centered finite differences (
and
).
For cubics, we obtain the following results
| |
| |
| |
| |
| |
Of course,
and
are both
and
and
are both at least
. However, for the function
, a superconvergence phenomenon occurs because we have
instead of
. We shall study this kind of results in a further paper.
————–
7 Approximating zeros of a function by those of a quadratic dQI
Let
be a continuous function defined on
. In order to locate the zeros of
in this interval, we approximate
by its
quadratic dQI
and we compute the exact zeros of
: this is quite possible because
is piecewise quadratic. The complete study will be done elsewhere. Here we take a simple example: we want to approximate the zeros of the Legendre polynomial
in the interval
. The five zeros of
are respectively
, with
The following array gives the errors
where
is the zero of
nearest to
.
| |
| |
| |
Acknowledgements: the author thanks Professor Catterina Dagnino and the Department of Mathematics of the University of Turin for their kind invitation to deliver this seminar during his stay from January 12 to 20, 2005.
References
-
C.de Boor, A practical guide to splines, Revised edition. Springer-Verlag, New-York (2001).
-
C.de Boor, Splines as linear combinations of B-splines, Approximation Theory II, G.G. Lorentz et al. (eds), Academic Press, New-York (1976), 1-47.
-
G.Chen, C.K. Chui, M.J. Lai, Construction of real-time spline quasi-interpolation schemes, Approx. Theory Appl. 4 (1988), 61-75.
-
P.J. Davis, P. Rabonowitz, Numerical integration. Blaisdell, Waltham (1967).
-
S.A. De Swardt, J.M. De Villiers, Gregory type quadrature based on quadratic nodal spline interpolation. Numer. Math. 85 (2000), 129-153.
-
R.A. DeVore, G.G. Lorentz, Constructive approximation, Springer-Verlag, Berlin (1993)
-
H. Engels: Numerical quadrature and cubature. Academic Press (1980).
-
W. Gautschi: Orthogonal polynomials: applications and computation. Dans Acta Numerica 1996, A. Iserles (ed.), CUP 1996.
-
A. Krommer, Ch. W. Ueberhuber, Computational integration. SIAM, Philadelphia (1998).
-
V. Lampret, An invitation to Hermite's integration and summation: a comparison between Hermite's and Simpson's rules. SIAM Review 46, No 2 (2004), 329-345.
-
B.G. Lee, T. Lyche, L.L. Schumaker, Some examples of quasi-interpolants constructed from local spline projectors. In Math methods for CAGD Oslo II, 243-252.
-
T. Lyche, L.L. Schumaker, Local spline approximation, J. Approx. Theory 15 (1975), 294-325.
-
M.J.D. Powell, Approximation theory and methods , Cambridge University Press (1981).
-
P. Sablonnière: Quasi-interpolantes splines sobre particiones uniformes. First Meeting in Approximation Theory of the University of Jaén (Ubeda, June 29-July 2, 2000). Prépublication IRMAR 00-38, Rennes (June 2000).
-
P. Sablonnière: On some multivariate quadratic spline quasi-interpolants on bounded domains. In Modern Developments in Multivariate Approximation, W. Haussmann et al. (eds), ISNM Vol. 145, Birkhäuser Verlag (2003), 263-278.
-
P. Sablonnière: Quadratic spline quasi-interpolants on bounded domains of
. In Spline and radial functions, Rend. Sem. Univ. Pol. Torino, Vol. 61 (2003), 61-78.
-
P. Sablonnière: A quadrature formula associated with a univariate quadratic spline quasi-interpolant. Prépublication IRMAR, Rennes, April 2005 (submitted).
-
P. Sablonnière, D. Sbibih: Integral spline operators exact on polynomials. Approx. Theory Appl. 10, No 3 (1994), 56-73.
-
L.L. Schumaker, Spline functions: basic theory, John Wiley and Sons, New-York (1981).
Author's address:
Paul Sablonnière, Centre de mathématiques, INSA de Rennes, 20 avenue des Buttes de Coësmes, CS 14315, F-35043-RENNES Cédex, France e-mail: psablonn@insa-rennes.fr