Numerical Methods I: Interpolation (cont ed)

Transcription

1 1/20 Numerical Methods I: Interpolation (cont ed) Georg Stadler Courant Institute, NYU November 30, 2017

2 Interpolation Things you should know 2/20 I Lagrange vs. Hermite interpolation I Conditioning of interpolation I Uniform vs. non-uniform points, Lebesgue constant I Polynomial bases: Lagrange, Newton, Monomial

3 Classical polynomial interpolation Newton polynomial basis 3/20 The Newton basis! 0,...,! n is given by Yi 1! i (t) := (t t j ) 2 P i. j=0 The leading coe cient a n of the interpolation polynomial of f P (f t 0,...,t n )=a n x n +... is called the n-th divided di erence, [t 0,...,t n ]f := a n.

4 Classical polynomial interpolation Newton polynomial basis 4/20 Theorem: For f 2 C n, the interpolation polynomial P (f t 0,...,t n ) is given by P (t) = nx [t 0,...,t i ]f! i (t). i=0 If f 2 C n+1,then f(t) =P (t)+[t 0,...,t n,t]f! n+1 (t). This property allows to estimate the interpolation error.

5 Classical polynomial interpolation Divided di erences 5/20 The divided di erences [t 0,...,t n ]f satisfy the following properties: I [t 0,...,t n ]P =0for all P 2 P n 1. I If t 0 =...= t n : nodes. [t 0,...,t n ]f = f (n) (t 0 ) n!

6 Classical polynomial interpolation Divided di erences 6/20 I The following recurrence relation holds for t i 6= t j (nodes with a hat are removed): [t 0,...,t n ]f = [t 0,...,ˆt i,...,t n ]f [t 0,..., ˆt j,...,t n ]f t j t i I If f 2 C n [t 0,...,t n ]f = 1 n! f (n) ( ) with an a apple apple b, and the divided di erences depend continuously on the nodes.

7 Classical polynomial interpolation Divided di erences 7/20 Let us use divided di erences to compute the coe cients for the Newton basis for the cubic interpolation polynomial p that satisfies p(0) = 1, p(0.5) = 2, p(1) = 0, p(2) = 3. t i 0 [t 0 ]f =1 0.5 [t 1 ]f =2 [t1]f [t 0 t 1 ]f = t 1 [t0]f t 0 =2 1 [t 2 ]f =0 [t2]f [t 1 t 2 ]f = t 2 [t1]f t 1 = 4 [t 0 t 1 t 2 ]f = 6 2 [t 3 ]f =3 [t3]f [t 2 t 3 ]f = t 3 [t2]f t 2 =3 [t 1 t 2 t 3 ]f = 14 3 Thus, the interpolating polynomial is 16 3 p(t) =1 + 2t + ( 6)t(t 0.5) + 16 t(t 0.5)(t 1). 3

8 Classical polynomial interpolation Divided di erences 8/20 Let us now use divided di erences to compute the coe cients for the Newton basis for the cubic interpolation polynomial p that satisfies p(0) = 1, p 0 (0) = 2, p 00 (0) = 1, p(1) = 3. t i 0 [t 0 ]f =1 0 [t 0 ]f =1 [t 0 t 1 ]f = p 0 (0) = 2 0 [t 0 ]f =1 [t 1 t 2 ]f = p 0 (0) = 2 [t 0 t 1 t 2 ]f = p00 (0) 2! = 1 2 [t3]f [t0]f 1 1 [t 3 ]f =3 [t 2 t 3 ]f = t 3 t 0 =2 0 2 Thus, the interpolating polynomial is p(t) =1 + 2t t2 + ( 1 2 )t3

9 Classical polynomial interpolation Approximation error If f 2 C (n+1),then 9/20 f(t) P (f t 0,...,t n )(t) = f (n+1) ( ) (n + 1)!! n+1(t) for an appropriate = (t), a< <b. In particular, the error depends on the choice of the nodes. For Taylor interpolation, i.e., t 0 =...= t n,thisresultsin: f(t) P (f t 0,...,t n )(t) = f (n+1) ( ) (n + 1)! (t t 0) n+1

10 Classical polynomial interpolation Approximation error 10 / 20 Consider functions {f 2 C n+1 ([a, b]) : sup f n+1 ( ) apple M(n + 1)!} 2[a,b] for some M>0, then the approximation error depends on! n (t), and thus on t 0,...,t n. Thus, one can try to minimize max! n+1(t), aappletappleb which is achieved by choosing the nodes as the roots of the Chebyshev polynomial of order (n + 1).

11 Classical polynomial interpolation Approximation error 11 / 20 Summary on pointwise convergence: I If an interpolating polynomial is close/converges to the original function depends on the regularity of the function and the choice of interpolation nodes I For a good choice of interpolation nodes, fast convergence can be obtained for almost all functions

12 Classical polynomial interpolation Interpolation/Least square approximation/splines I Polynomial interpolation I Least squares with polynomials I Splines (i.e., piecewise polynomial interpolation): 12 / 20

13 Splines 13 / 20 Assume (l + 2) pairwise disjoint nodes: a = t 0 <t 1 <...<t l+1 = b. A spline of degree k 1 (order k) is a function in C k 2 which on each interval [t i,t i+1 ] coincides with a polynomial in P k 1. Most important examples: I linear splines, k =2 I cubic splines, k =4

14 Splines Cubic splines look smooth: So s a = to 14 / 20

15 Splines B-splines 15 / 20 B-splines are a basis in the spline space that: I has local support I satisfies a 3-term recursion I non-negative

16 Splines B-splines 16 / 20 I Coe cients for interpolation with the B-spline basis can be computed e ciently using the De Boor algorithm. I Splines are essential in Computer Aided Design (CAD). I Also important in CAD: Bezier curves (these do not interpolate points and have useful geometrical properties).

17 Trigonometric Interpolation For periodic functions 17 / 20 Instead of polynomials, use sin(jt), cos(jt) for di erent j 2 N. For N 1, we define the set of complex trigonometric polynomials of degree apple N 1 as 8 9 < NX 1 = T N 1 := c j e ijt,c j 2 C : ;, j=0 where i = p 1. Complex interpolation problem: Givenpairwisedistinctnodes t 0,...,t N 1 2 [0, 2 ) and corresponding nodal values f 0,...,f N 1 2 C, find a trigonometric polynomial p 2 T N 1 such that p(t i )=f i,fori =0,...,N 1.

18 Trigonometric Interpolation I There exists exactly one p 2 T N interpolation problem. 1, which solves this I Choose the equidistant nodes t k := 2 k N for k =0,...,N 1. Then, the trigonometric polynomial that satisfies p(t i )=f i for i =0,...,N 1 has the coe cients c j = 1 N NX 1 k=0 e 2 ijk N fk. I For equidistant nodes, the linear map from C N! C N defined by (f 0,...,f N 1 ) 7! (c 0,...,c N 1 ) is called the discrete Fourier transformation (DFT). 18 / 20

19 19 / 20 iscrete Fourier transform The interpolation problem (f 0,...,f N 1 ) 7! (c 0,...,c N 1 ) and its inverse require the multiplication or solution with a dense n n system, i.e., at least O(n 2 ) flops. However, the special structure of the system matrix allows performing those operations using a much faster algorith, the Fast Fournier Transform (FFT).

20 Trigonometric Interpolation 20 / 20 The Fast Fourier Transform (FFT) is a (very famous!) algorithm that computes the DFT and its inverse in O(n) flops. I Note that uniform nodes are used (and even required for the FFT). I Tensor products on square domains can be used for two dimensional approximations, i.e., p(x)p(y). I Can be used to approximate and solve di erential equations (see Numerical Methods II).