COURSE Numerical integration of functions (continuation) 3.3. The Romberg s iterative generation method

Size: px

Start display at page:

Download "COURSE Numerical integration of functions (continuation) 3.3. The Romberg s iterative generation method"

Adrian Stephens
6 years ago
Views:

1 COURSE 7 3. Numerical integration of functions (continuation) 3.3. The Romberg s iterative generation method The presence of derivatives in the remainder difficulties in applicability to practical problems and to computer programs. There are preferred, in this sense, the iterative quadratures. Consider the iterative generation method of a repeated formula by the Romberg s method. In the case of the trapezium formula we have Q T0 (f) = h [f(a) + f(b)], h = b a, Q T0 (f) being the first element of the sequence.

2 We divide the interval [a, b] in two equal parts, of length h and applying to each subinterval [a, a + h ] and [a + h, b] the trapezium formula we get Q T1 (f) = h [ ( f(a) + f a + h ) ] + f(b) 4 or Q T1 (f) = 1 Q T 0 (f) + hf ( a + h ). Dividing now each previous divisions [a, a+b ] and [a+b, b] in two equal parts, we obtain a division of the initial interval in 4 = equal parts, each of length 4 h. Applying the repeated trapezium formula, we get Q T (f) = h 8 f(a) + 3 i=1 = 1 Q T 1 (f) + h [f Continuing in an analogous manner, we get Q Tk (f) = 1 Q T k 1 (f) + h k 1 k f j=1 ( a + ih ) + f(b) (1) 4 (a + 1 ) h + f (a + 3 )] h. f ( a + j 1 ) k h, k = 1,,... ()

3 We obtain the sequence Q T0 (f), Q T1 (f),..., Q Tk (f),... (3) which converges to the value I = a f(x)dx. We approximate the error by Q Tn (f) Q Tn 1 (f). If we want to approximate I with error less than ε, we compute successively the elements of (3) until the first index for which Q Tn (f) being the required value. Q Tn (f) Q Tn 1 (f) ε, Similarly, one may iteratively generate the repeated Simpson s formula. Denoting by Q Sk (f) the Simpson s formula repeated k times, we have Q Sk (f) = 1 3 [ 4QTk+1 (f) Q T k (f) ], k = 0, 1,... where Q S0 (f) = h 6 [ f(a) + 4f ( a + h ) ] + f(b)

4 is the Simpson s quadrature formula Adaptive quadrature methods The repeated integration methods require equidistant nodes. There are problems where the function contains both regions with large variations and with small variations. It is needed a smaller step for the regions with large variations than for the regions with small variations in order that the error to be uniformly distributed. Such methods, which adapt the size of the step in accordance with the need, are called adaptive quadrature methods. We present the method based on the repeated Simpson s quadrature formula. Suppose we want to approximate with precision ε the integral I = a f (x) dx.

5 First step: we apply the Simpson s formula for x 0 = a, x 1 = a+ h, x = b, with h = b a. a f (x) dx = b a 6 ( f(a) + 4f(a + h ) + f(b) ) (b a)5 880 f (4) (ξ) = := S (a, b) h5 90 f (4) (ξ) It may be proved that the error of approximating I by S ( a, a + h ) + S ( a + h, b), h = b a, is 15 times smaller than the expression S (a, b) S ( a, a + h ) S ( a + h, b ). Hence, if (4) S (a, b) S ( a, a + h ) S ( a + h, b ) < 15ε, then (5) a f (x) dx S ( a, a + h ) S ( a + h, b ) < ε. When (5) does not hold, the procedure is applied individually on [a, (a+ b)/] and [(a + b)/, b] in order to determine if the approx. of the integral on each two subintervals is performed with error ε/. If yes,

6 the sum of these two approx. offer an approx. of I with precision ε. If on a subinterval it is not obtained the error ε/, then we divide that subinterval and we analyze if the approx. on the resulted two subintervals has precision ε/4, and so on. This procedure of halfing is continued until the corresponding error is attained on each subinterval.

7 Algorithm: (the idea: divide and conquer ) function I=adquad(a,b,er) I1=Simpson(a,b) I=Simpson(a, a+b )+Simpson(a+b,b) if I1-I <15*er I=I return else end I=adquad(a, a+b,er )+adquad(a+b,b,er ) Remark 1 For example, for evaluating the integral x sin 10 x dx with ε = 10 4, repeated Simpson formula requires 177 function evaluations, nearly twice as many as adaptive quadrature.

8 3.5. General quadrature formulas Using the interpolation formulas, there are obtained a large variety of quadrature formulas. In the case of some concrete applications, the choosing of the quadrature formula is made according to the information about the function f. A general quadrature formula is given by: a f(x)dx = m k=0 j I k A kj f (j) (x k ) + R(f), For example, if we know only the values of f (a) and f(b),

9 using the Birkhoff interpolation formula f(x) = (x b)f (a) + f(b) + (R 1 f)(x) we obtain the quadrature formula with a f(x)dx = A 0f (a) + A 1 f(b) + R 1 (f), A 0 = A 01 = A 1 = A 10 = a a b) (x b)dx = (a dx = b a.

10 3.6. Quadrature formulas of Gauss type All the previous rules can be written in the form a f(x)dx = m k=1 A k f(x k ) + R m (f), (6) where the coefficients A k, k = 1,..., m, do not depend on the function f. We have picked the nodes x k, k = 1,..., m equispaced and have then calculated the coefficients A k, k = 1,..., m. This guarantees that the rule is exact for polynomials of degree m. It is possible to make such a rule exact for polynomials of degree m 1, by choosing also the nodes appropriately. This is the basic idea of the gaussian rules. Let f : [a, b] R be an integrable function and w : [a, b] R + a weight function, integrable on [a, b].

11 Definition A formula of the following form a w(x)f(x)dx = m k=1 A k f(x k ) + R m (f) (7) is called a quadrature formula of Gauss type or with maximum degree of exactness if the coefficients A k and the nodes x k, k = 1,..., m are determined such that the formula has the maximum degree of exactness. Remark 3 The coefficients and the nodes are determined such that to minimize the error, to produce exact result for the largest class of polynomials. A k and x k, k = 1,..., m from (7) are m unknown parameters m equations obtained such that the formula (7) is exact for any polynomial degree at most m 1. It is often possible to rewrite the integral a g(x)dx as a w(x)f(x)dx, where w(x) is a nonnegative integrable function, and f(x) = g(x) w(x) is smooth, or it is possible to consider the simple choice w(x) = 1.

12 For the general case, consider e k (x) = x k ; k = 0,..., m 1 and obtain the system s.t. R m (e k ) = 0 : with m k=1 m k=1.. ṁ k=1 A k e 0 (x k ) = a w(x)e 0 (x) dx A k e 1 (x k ) = a w(x)e 1 (x) dx A k e m 1 (x k ) = a w(x)e m 1 (x) dx A 1 + A A m = µ 0 A 1 x 1 + A x A m x m = µ 1... A 1 x m A x m A m x m 1 m = µ m 1 µ k = a w(x)xk dx. As, the system (8) is difficult to solve, there have been found other ways to find the unknown parameters. (8)

13 If w(x) = 1 (this case was studied by Gauss), then the nodes are the roots of Legendre orthogonal polynomial u(x) = m! (m)! [(x a)m (x b) m ] (m) and for finding the coefficients we use the first m equations from the system (8). For example, for m =, the interval [ 1, 1] and w(x) = 1, we get the system A 1 + A = 1 1 dx = A 1 x 1 + A x = 1 1 xdx = 0 A 1 x 1 + A x = 1 1 x dx = 3 A 1 x A x 3 = 1 1 x 3 dx = 0 with solution A 1 = A = 1 and x 1 = 3 3, x = fomula 1 1 f(x)dx f( 3 3 ) + f( 3 3 ). (9) 3 3, which gives the

14 This formula has degree of precision 3, i.e., it gives exact result for every polynomial of the 3 rd degree or less. Remark 4 The resulting rules look more complicated than the interpolatory rules. Both nodes and weights for gaussian rules are, in general, irrational numbers. But, on a computer, it usually makes no difference whether one evaluates a function at x = 3 or at x = 1/ 3. Once the nodes and weights of such a rule are stored, these rules are as easily used as the trapezium rule or Simpson s rule. At the same time, these gaussian rules are usually much more accurate when compared with the last ones on the basis of number of function values used. Example 5 Consider m = 1 and obtain the following Gauss type quadrature formula a f(x)dx = A 1f(x 1 ) + R 1 (f).

15 The system (8) becomes A 1 = a dx = b a A 1 x 1 = a xdx = b a. The unique solution of this system is A 1 = b a, x 1 = a+b. The same result is obtained considering x 1 the root of the Legendre polynomial of the first degree, u(x) = 1 [(x a)(x b)] = x a + b.

16 The Gauss type quadrature formula with one node is ( ) a + b f(x)dx = (b a)f + R 1 (f), a with (b a)3 R 1 (f) = f (ξ), ξ [a, b] 4 which is called the rectangle quadrature rule (also called the midpoint rule). The repeated rectangle (midpoint) quadrature formula is a f(x)dx = b a n R n (f) = n i=1 f(x i ) + R n (f), (b a)3 4n f (ξ), ξ [a, b] with x 1 = a + b a n, x i = x 1 + (i 1) b a n, i =,..., n.

17 We have R n (f) (b a)3 4n M f, with M f = max x [a,b] f (x). Remark 6 There is another rectangle rule: f(x)dx = (b a)f (a) + R(f), with a R(f) = (b a) f (ξ), ξ [a, b].

18 Romberg s algorithm for the rectangle (midpoint) quadrature formula. Applying successively the rectangle formula on [a, b], we get Q D0 (f) = (b a)f(x 1 ), x 1 = a + b Q D1 (f) = 1 3 Q D 0 (f) + b a 3 [f(x ) + f (x 3 )], x = a + b a 6, x 3 = b b a 6. Continuing in an analogous manner, we obtain the sequence Q D0 (f), Q D1 (f),..., Q Dk (f),... (10) which converges to the value I of the integral a f(x)dx. If we want to approximate the integral I with error less than ε, we compute successively the elements of (10) until the first index for which Q Dm (f) being the required value. Q Dm (f) Q Dm 1 (f) ε,

19 The algorithm for generating the elements of the sequence (10) is: I. Let h := b a,h 1 := h,x 1 : a + h 1 and Q D0 (f) := hf(x 1 ). II. For k := 1,,... do h := h 3,h 1 := h 1 3,h := 4h 1,h 3 := h 1, m := 3 k 1,x 1 := a + h 1 ; for i = 1,..., m 1,do x i := x i 1 + h, x i+1 := x i + h 3 ; x m := x m 1 + h and Q Dk (f) = 1 3 Q D k 1 (f) + h m i=1 f(x i ), (for k = 1(m = 1)the generation of x i,x i+1 is missing). Example 7 Approximate ln, with ε = 10, using the repeated rectangle (midpoint) method.

20 Solution. We have f(x)dx = b a a n so f(x) = 1 x ln = b a n and we get R n (f) = f(a + b a n ) + n i=1 f(x i ) + R n (f), (b a)3 4n f (ξ), ξ [a, b]. ln = n i= 1 dx x, f(a + b a n We have f(x) = 1 x, f (x) = 1 x, f (x) = ξ [1, ] so it follows a + (i 1)b n ) + (b a)3 4n f (ξ) x 3, and f (ξ), for R n (f) 1 4n < 10 1n > 100 n = 3.

21 Therefore, ln = 1 3 ( ) = (real value is ) Example 8 For m =, the Gauss type quadrature formula is a f(x)dx = A 1f(x 1 ) + A f(x ) + R (f). The corresponding Legendre polynomial is u(x) = 4! [ (x a) (x b) ] = x (a + b)x (a + b + 4ab),

22 with the roots x 1 = a + b x = a + b (b a) 3, 6 (b a) For finding A 1 and A we use the first two equations: { A1 + A = b a A 1 x 1 + A x = (b a )/. We get A 1 = A = (b a)/, so the Gauss type quadrature formula with two nodes is f(x)dx = b a [ ( a + b f a b a ) ( a + b 3 + f 6 + b a )] 3 +R (f). 6 Remark 9 a) The coefficients A k, k = 1,..., m of a Gauss type formula are positive.

23 b) The coefficients A k, k = 1,..., m and the roots of the Legendre polynomials can be found in tables, for a = 1, b = 1. For example, for m = and m = 3 : m nodes coefficients c) For different weight functions, tables are available for both the nodes and the coefficients. Example 10 Approximate 1 1 e x cos xdx using a Gauss type quadrature formula for m = 3. Sol. 1 1 ex cos xdx 0.55e 0.77 cos cos e 0.77 cos( 0.77) =

24 (Exact value is ) Absolute error is < The integral a f(x)dx for an arbitrary interval [a, b] could be transfomed in an integral on [ 1, 1] using the change of variable t = x a b x = 1 [(b a)t + a + b]. b a The Gauss type quadrature formulas may be applied on the following way: ( ) 1 (b a)t + (b + a) (b a) f(x)dx = f dt. (11) a 1 Example 11 Consider the integral 3 1 (x 6 x sin(x))dx = a) Compare the result obtained for Newton-Cotes type formula with m = 1 and the gaussian formula for m = ; b) Compare the result obtained for Newton-Cotes type formula with m = and the gaussian formula for m = 3.

25 Sol. a) Each formula needs evaluations of the function f(x) = x 6 x sin(x). We have Trapezium formula (m = 1) : [f(1) + f(3)] = ; and a Gauss type formula for m =, using (11): (x6 x sin(x))dx = ((t + 1 )6 (t + ) sin((t + )))dt f( ) + f( ) = b) Each formula needs 3 evaluations of the function. We have Simpson s F. (m = ) : 1 3 [f(1) + 4f() + f(3)] = 333.3; and a Gauss type formula for m = 3, using (11): (x6 x sin(x))dx = ((t + 1 )6 (t + ) sin((t + )))dt 0.55f( ) f() f( ) =

COURSE Numerical integration of functions

COURSE Numerical integration of functions COURSE 6 3. Numerical integration of functions The need: for evaluating definite integrals of functions that has no explicit antiderivatives or whose antiderivatives are not easy to obtain. Let f : [a,