tuto9_sol_3230

pdf

School

The Chinese University of Hong Kong *

*We aren’t endorsed by this school

Course

3230

Subject

Mathematics

Date

Nov 24, 2024

Type

pdf

Pages

5

Uploaded by MinisterGorilla8297

Report
MATH3230 Numerical Analysis Tutorial 9 with solution 1 Recall: 1. Discrete and continuous least-squares approximations: (a) Discrete least-squares approximations: Let A be a m × n matrix, with m > n and now we consider a general linear system Ax = b. (1) The least-squares solution seeks for some vector x that minimizes the error ( Ax b ) in the least-squares sense, that is min x R n ஀? Ax b ஀? 2 . We assume that the columns of A are linearly independent. We define f ( x ) = ஀? Ax b ஀? 2 . The minimizer of f ( x ) satisfies the normal equation A T Ax = A T b. (b) Continuous least-squares approximations: This method is an approximation of a general function f on an interval [ a, b ] . For any two functions f, g C [ a, b ] , we define their inner product as f, g = ˆ b a f ( x ) g ( x ) dx, and the norm as ஀? f ஀? 2 = f, g 1 / 2 = ஀? ˆ b a f 2 ( x ) dx ஀? 1 / 2 . Let P n be the set of all polynomials of degree n on [ a, b ] . Then we can define the continuous least-squares approximation of a general function f on an interval [ a, b ] as follows: Find p n P n such that ஀? p n f ஀? 2 = min p P n ஀? p f ஀? 2 . p n P n is the least-squares approximation of f ( x ) on [ a, b ] if and only if the error function p n f is orthogonal to the space P n . 2. Sensitivity of linear systems: Consider solving the linear system (1). Because of the rounding errors, or the observation data errors, the actual problem we are solving by computer is the perturbed system: ˜ A ˜ x = b. Let A be a nonsingular matrix, and ˜ A = A + E . If Ax = b, ˜ A ˜ x = b, b = 0 , 1
then we have ஀? ˜ x x ஀? ஀? ˜ x ஀? ≤ ஀? A 1 E ஀? = ஀? A 1 ˜ A I ஀? . In addition, if ஀? A 1 E ஀? < 1 , we have ஀? ˜ x x ஀? ஀? x ஀? ஀? A 1 E ஀? 1 − ஀? A 1 E ஀? . The real number κ ( A ) given by κ ( A ) = ஀? A ஀?஀? A 1 ஀? is called the condition number of the matrix A . Note that κ ( A ) 1 . Then we have ஀? ˜ x x ஀? ஀? x ஀? c ( E ) κ ( A ) ஀? E ஀? ஀? A ஀? = c ( E ) κ ( A ) ஀? ˜ A A ஀? ஀? A ஀? , where c ( E ) = 1 1 κ ( A ) ஀? E ஀? ஀? A ஀? > 1 . The condition number gives a bound on how inaccurate the solution x will be after approximation. If the condition number is large, even a small error in b will cause a large error in x . If the condition number is small, then the solution x will be stable. 2 Exercises: Please do the star problems (*) in the tutorial class and finish the rest after the class. 1. Let A be a m × n matrix and consider a general linear system Ax = b. (2) (a) Write down the di ff erences between the two systems: the square system with m = n and the non-square system with m > n . When do these systems have a solution? (b) Assume that A with m > n is full rank, we can find the least square solution x which minimizes ஀? Ax b ஀? 2 . Prove that solving for x is equivalent to solving the equation A T Ax = A T b . And explain briefly the relation between the error vector Ax b and the space spanned by the column vectors of A . (c) * Solve the following linear system using the Cholesky factorization in the least-squares sense. Ax = b, where A = ஀? ஀? ஀? ஀? 1 1 1 1 1 1 0 1 1 0 1 1 ஀? ஀? ஀? ஀? and b = ஀? ஀? ஀? ஀? 2 1 1 2 ஀? ஀? ஀? ஀? . Solution. (a)-(b) See lecture notes. (c) Note that A T A = ஀? ஀? 2 2 0 2 4 2 0 2 4 ஀? ஀? Then A T A = ஀? ஀? 2 2 0 0 2 2 0 0 2 ஀? ஀? T ஀? ஀? 2 2 0 0 2 2 0 0 2 ஀? ஀? = R T R First we solve Ry = A T b , we have y = ஀? ஀? ஀? 1 2 1 2 3 2 ஀? ஀? ஀? 2
then we solve Rx = y , we have x = ஀? ஀? 3 2 1 3 2 ஀? ஀? 2. * Apply the continuous least-squares approximation to the function f ( x ) = e x for x [0 , 1] . The basis of P n is { φ k ( x ) = x k } n k =0 and consider the case n = 1 . Find p 1 P 1 such that ஀? p 1 f ஀? 2 = min p P 1 ஀? p f ஀? 2 . Solution. Let p 1 = c 0 + c 1 x . Then the error function is given by E ( c 0 , c 1 ) := ஀? f ( x ) ( c 0 + c 1 x ) ஀? 2 L 2 = ˆ b a ( f ( x ) c 0 c 1 x ) 2 d x = ˆ b a ஀? f ( x ) 2 2 f ( x ) ( c 0 + c 1 x ) + ஀? c 2 0 + 2 c 0 c 1 x + c 2 1 x 2 ஀?஀? d x = ˆ b a f ( x ) 2 d x 2 c 0 ˆ b a f ( x )d x 2 c 1 ˆ b a xf ( x )d x + c 2 0 ( b a ) + c 0 c 1 ஀? b 2 a 2 ஀? + 1 3 c 2 1 ஀? b 3 a 3 ஀? . Then we optimize E over c 0 and c 1 , i.e., we find the values of c 0 and c 1 for which E c 0 = E c 1 = 0 . We have c 0 ( b a ) + 1 2 c 1 ஀? b 2 a 2 ஀? = ˆ b a f ( x )d x, 1 2 c 0 ஀? b 2 a 2 ஀? + 1 3 c 1 ஀? b 3 a 3 ஀? = ˆ b a xf ( x )d x. That is ஀? ( b a ) 1 2 ஀? b 2 a 2 ஀? 1 2 ஀? b 2 a 2 ஀? 1 3 ஀? b 3 a 3 ஀? ஀? ஀? c 0 c 1 ஀? = ஀? ´ b a f ( x )d x ´ b a xf ( x )d x ஀? . Hence ஀? 1 1 2 1 2 1 3 ஀? ஀? c 0 c 1 ஀? = ஀? e 1 1 ஀? . The desired solution is c 0 = 4e 10 , c 1 = 18 6e . 3. Given an invertible n × n matrix A . Let b, b δ , x, x δ R n \{ 0 } be four non-zero vectors such that Ax = b and Ax δ = b δ . (a) Show that there exists κ ( A ) > 0 such that 1 κ ( A ) ஀? b b δ ஀? ஀? b δ ஀? ஀? x x δ ஀? ஀? x δ ஀? κ ( A ) ஀? b b δ ஀? ஀? b δ ஀? , where ஀? · ஀? is a given norm. (b) Let A = ஀? 20 25 15 20 ஀? . Find ஀? κ ( A ) ஀? in the following norm 3
Your preview ends here
Eager to read complete document? Join bartleby learn and gain access to the full version
  • Access to all documents
  • Unlimited textbook solutions
  • 24/7 expert homework help
i. 1 -norm. ii. sup-norm. (c) Based on the statement in ( a ) , explain the importance of the condition number κ ( A ) for solving the system Ax = b . Solution. (a) Let us first recall that ஀? Ax ஀? ≤ ஀? A ஀?஀? x ஀? x R n . Using this inequality and the fact that b = Ax b δ = Ax δ , we have i. ஀? b b δ ஀? ≤ ஀? A ஀?஀?஀? x x δ ஀? ii. ஀? b δ ஀? ≤ ஀? A ஀?஀? x δ ஀? iii. ஀? x x δ ஀? ≤ ஀? A 1 ஀?஀? b b δ ஀? iv. ஀? x δ ஀? ≤ ஀? A 1 ஀?஀? b δ ஀? Using (i) and (iv), we have ஀? b b δ ஀? · 1 ஀? b δ ஀? 1 ஀? A ஀?஀? A 1 ஀? ≤ ஀? x x δ ஀? · 1 ஀? x δ ஀? . which is the first inequality required. Then using (ii) and (iii), we have ஀? x x δ ஀? · 1 ஀? x δ ஀? ≤ ஀? A 1 ஀?஀? b b δ ஀?஀? A ஀? · 1 ஀? b δ ஀? , which is the second inequality required. (b) i. κ ( A ) = ஀? A ஀? 1 ஀? A 1 ஀? 1 = 45 × 1 . 8 = 81 . ii. κ ( A ) = ஀? A ஀? ஀? A 1 ஀? = 45 × 1 . 8 = 81 . (c) See lecture notes. 4. Consider Ax = b . THe perturbation occurs on both A and b , which means Ax = b ( A + A ) ( x + x ) = b + b. (a) If ஀? A ஀? < 1 , ஀? I ஀? = 1 , prove that ஀? ஀? ஀? ( I A ) 1 ஀? ஀? ஀? 1 1 − ஀? A ஀? . (b) If ஀? A ஀? < 1 , prove that ஀? I + A ஀? is invertible. (c) * Suppose ஀? ஀? A 1 ஀? ஀? ஀? A ஀? < 1 . Using the result of (a) and (b), prove that ஀? x ஀? ஀? x ஀? κ ( A ) 1 κ ( A ) ஀? A ஀? ஀? A ஀? ஀? ஀? A ஀? ஀? A ஀? + ஀? b ஀? ஀? b ஀? ஀? . Solution. (a) Since ஀? A ஀? < 1 and ஀? A n ஀? ≤ ஀? A ஀? n , we have ஀? A n ஀? → 0 when n → ∞ . Let s N = N ஀? k =0 A k 4
Then s N ( I A ) = I A N +1 s N = ( I A ) 1 ( I A N +1 ) when N → ∞ , A N +1 0 (Since ஀? A N ஀? = 0 i ff A N 0 ) lim N →∞ s N = ( I A ) 1 . So ஀? ஀? k =0 A k ஀? = ஀? ( I A ) 1 ஀? . Moreover, we have ஀? ஀? k =0 A k ஀? ஀? k =0 ஀? A k ஀? = ஀? I ஀? + ஀? A ஀? + · · · = 1 1 − ஀? A ஀? (b) x = 0 , ஀? ( I + A ) x ஀? = ஀? x + Ax ஀? ≥ ஀? x ஀? − ஀? Ax ஀? ≥ ஀? x ஀? − ஀? A ஀?஀? x ஀? = ஀? x ஀? (1 − ஀? A ஀? ) > 0 . So I + A is invertible. (c) Since ( A + A )( x + x ) = b + b , we have ( A + A ) x = b Ax (3) We have ஀? ( I + A 1 A ) 1 ஀? ≤ 1 1 − ஀? A 1 ஀?஀? A ஀? (4) By (3), x = ( A + A ) 1 ( b Ax ) = ( I + A 1 A ) 1 A 1 ( b Ax ) ⇒ ஀? x ஀? ஀? ( I + A 1 A ) 1 ஀?஀? A 1 ஀? ( ஀? b ஀? + ஀? A ஀?஀? x ஀? ) By (4), ஀? x ஀? ஀? A 1 ஀? 1 − ஀? A 1 ஀?஀? A ஀? ( ஀? b ஀? + ஀? A ஀?஀? x ஀? ) ஀? x ஀? ஀? x ஀? ஀? A 1 ஀? 1 − ஀? A 1 ஀?஀? A ஀? ( ஀? A ஀? + ஀? b ஀? ஀? x ஀? ) ஀? A 1 ஀? 1 − ஀? A 1 ஀?஀? A ஀? ( ஀? A ஀? + ஀? b ஀?஀? A ஀? ஀? b ஀? ) κ ( A ) 1 κ ( A ) ஀? A ஀? ஀? A ஀? ( ஀? A ஀? ஀? A ஀? + ஀? b ஀? ஀? b ஀? ) 5