Homework No. 2 Solutions TA: Theo Drivas

Transcription

1 Homework No. Solutions TA: Theo Drivas Problem 1. (a) Consider the initial-value problems for the following two differential equations: (i) dy/dt = y, y(0) = y 0, (ii) dy/dt = y, y(0) = y 0. Use separation of variables (or any other technique) to solve these two problems exactly as explicit functions y(t; y 0 ) of time t and initial condition y 0. Are these problems well-posed? Explain your answer. (b) For each of the two problems in (a), calculate both the absolute condition number and the relative condition number Abs.cond(y(t; y 0 )) := y(t; y 0) y 0 Rel.cond(y(t; y 0 )) := y 0 y(t; y 0 ) y(t; y 0 ) y 0 For the specific time t = 100, is each problem absolutely well-conditioned? Relatively well-conditioned? Explain your answers. (c) Later in the course we shall study the Euler method to approximately solve the problem (ii) in part (a): ŷ(t n+1 ) ŷ(t n ) h = ŷ(t n ), t n = nh, n = 0, 1,,...; ŷ(0) = y 0, which gives the approximation ŷ(t n ) y(t n ) by solving a simple first-order linear difference equation. Find the explicit expression for the approximate solution ŷ(t n ; y 0, h) as a function of t n, y 0, and h. Also calculate the absolute and relative condition numbers of the approximation, using similar definitions as in part (b). For what values of h is the absolute condition number always less than 1? (Sol a) The general solution to ẏ = ±y is y(t; y 0 ) = y 0 e ±t These problems are well posed because there (i) is a solution, (ii) it is unique and (iii) it depends continuously on the data (here, y 0 ). (Sol b) The absolute conditions numbers are Abs.cond(y(t; y 0 )) := y(t; y 0) y 0 = { e t e t for ẏ = y for ẏ = y The relative conditions numbers are identical for both problems because y(t;y0) y 0 Rel.cond(y(t; y 0 )) := y 0 y(t; y 0 ) = 1 y(t; y 0 ) y 0 = y(t;y0) y 0 : 1

2 Clearly both problems are relatively well-conditioned since both relative condition numbers are unity for all time. Because { e 100 for ẏ = y Abs.cond(y(100; y 0 )) = e for ẏ = y! the first problem is not absolutely well-conditioned at time t = 100, but the second is absolutely well-conditioned at t = 100. (Sol c) For t n = nh, n = 0, 1,,... ŷ(t n+1 ) ŷ(t n ) = ŷ(t n ), h = ŷ(t n ) = (1 h) tn/h y 0 The condition numbers are: Abs.cond(ŷ(t n ; y 0 )) := ŷ(t n; y 0 ) y 0 Rel.cond(ŷ(t n ; y 0 ))) := For h satisfying 1 < (1 h) < 1, i.e. necessarily less than 1. = (1 h) tn/h y 0 ŷ(t n ; y 0 ) = 1 ŷ(t n ; y 0 ) y 0 Problem. (a) Use bisection method to find the roots of 0 < h <, the absolute condition number is f(x) = sin(x) ln(x 4 + 1/) x with tolerance T OL = Make certain that you find all of the roots of the given function. Record the final endpoints of the bracketing interval, the achieved tolerance, the number of iterations, and the computational time for each root. (You may use the prepared matlab script bisect.m, if you wish.) (b) Repeat part (a) using instead the Newton-Fourier method. This method combines the Newton iteration for x n with another iteration equation y n+1 = y n f(y n) f (x n ). Assuming that f is increasing and convex on the interval [x 0, y 0 ], i.e. that f (x) > 0, f (x) > 0, x [x 0, y 0 ] ( ) and that x 0, y 0 bracket a root, or f(x 0 ) < 0 and f(y 0 ) > 0, then the iterates [x n, y n ] converge quadratically and always bracket the root. Write your own matlab script newtfour.m to implement this algorithm, e.g. by modifying newton.m. (c) If the condition (*) is not satisfied, but f and f do not change sign on an interval around the root, then one of the following modified functions f (x) = f(x), f 3 (x) = f( x), f 4 (x) = f( x) will satisfy (*). Use the Newton-Fourier method to find all of the roots of the function in part (a) to tolerance T OL = For each root use the same initial interval [x 0, y 0 ] as you did for the bisection method in part (a). Note that the function may not satisfy the condition (*) for each of the roots and, in that case, you will need to use one of the transformed functions discussed above

3 (d) Compare the results of the bisection and Newton-Fourier methods (final endpoints of the bracketing interval, achieved tolerance, number of iterations, and computational time). Which method is preferable and why? (Sol a) 1 >> f=inline('sin(x)-log(x.ˆ4+1/)-x') >> tol=10ˆ(-15) 3 4 >> figure; fplot(@(x) [f(x), 0*x], [-10,10]) 5 >> pause 6 7 >> format long 8 9 >> % BISECTION >> a=0;b=1; 1 >> bisect k = ans = Elapsed time is seconds. 17 >> pause >> a=-1;b=0; 0 >> bisect 1 k = 50 3 ans = Elapsed time is seconds. 5 6 >> pause 7 8 >> a=-10;b=-9; 9 >> bisect k = 47 3 ans = Elapsed time is seconds. (Sol b) Here is a version of newtfour.m: 1 tic 3 itmax=100; 4 5 if sign(f(x))==sign(f(y)) 6 'failure to bracket root' 7 return 8 end 9 10 k=0 11 [x y y-x] 1 13 while abs(x-y)>tol*max(abs(y),1.0) 14 if k+1>itmax 15 break 16 end 17 xold=x; 18 yold=y; 19 k=k+1 3

4 Figure 1: The function f(x) = sin(x) ln(x 4 + 1/) x. 0 DF=Df(y); 1 x=x-f(x)/df; y=y-f(y)/df; 3 [x y y-x] 4 end 5 6 toc (Sol c) We now apply this algorithm: 1 % FOURIER-NEWTON 3 >> fold=f; 4 >> figure; fplot(@(x) [fold(x), 0*x], [0,1]) 5 >> pause 6 >> % f(x)=-f(x) 7 >> f=inline('-(sin(x)-log(x.ˆ4+1/)-x)') 8 >> figure; fplot(@(x) [f(x), 0*x], [0,1]) 9 >> Df=inline('-cos(x)+4*x.ˆ3./(x.ˆ4+1/)+1') >> x=0;y=1; 1 >> newtfour k = 7 15 ans = Elapsed time is seconds >> pause 19 0 >> figure; fplot(@(x) [fold(x), 0*x], [-1,0]) 1 >> pause >> % f(x)=-f(-x) 3 >> f=inline('sin(x)+log(x.ˆ4+1/)-x') 4 >> figure; fplot(@(x) [f(x), 0*x], [0,1]) 5 >> Df=inline('cos(x)+4*x.ˆ3./(x.ˆ4+1/)-1') 6 7 >> x=0;y=1; 8 >> newtfour 9 30 k = 7 31 ans =

5 3 Elapsed time is seconds >> pause >> figure; [fold(x), 0*x], [-10,-9]) 38 >> pause 39 >> % f(x)=f(-x) 40 >> f=inline('-sin(x)-log(x.ˆ4+1/)+x') 41 >> figure; fplot(@(x) [f(x), 0*x], [9,10]) 4 >> Df=inline('-cos(x)-4*x.ˆ3/(x.ˆ4+1/)+1') >> x=9;y=10; 45 >> newtfour k = 4 48 ans = Elapsed time is seconds. Figure : The function f f(x) = sin(x) + ln(x 4 + 1/) + x (left) and the derivative f f (x) = cos(x) (right) on the interval x [0, 1]. 4x3 x 4 +1/ Similar plots to these are generated by the code for the corresponding other modified functions. (Sol d) Newton-Fourier method is preferable since it requires less computational time, less number of iterations, while achieves approximately the same bracketing interval and tolerance, than that of bisection method. 5

6 Problem 3. a) Use Newton s method to find the smallest positive root of f(x) = sin(x 3 ) x/, again with tolerance T OL = Record the final approximate root, the achieved tolerance, the number of iterations, and the computational time. (b) Repeat part (a) using the secant method. Use the same initial x 0 as you did for Newton s method in (a) and make one Newton iteration to generate x 1. (c) Repeat part (a) using the method of inverse quadratic interpolation (IQI). Use the same initial x 0, x 1 as you did for the secant method in (b), and make one secant iteration to generate x. (d) Compare the results of the three methods (final endpoints of the bracketing interval, achieved tolerance, number of iterations, and computational time). Which method is preferable and why? (Sol a) Figure 3: The function f(x) = sin(x 3 ) x/. 1 Ntry=1000; >> f=inline('sin(x.ˆ3)-x/') 3 >> Df=inline('3*x.ˆ.*cos(x.ˆ3)-1/') 4 >> figure; fplot(@(x) [f(x), 0*x], [0,1]) 5 >> pause 6 7 >> timen=0; 8 >> for nt=1:ntry 9 >> x=1; 10 >> tic 11 >> newton 1 >> timen=timen+toc; 13 >> end 14 >> timen=timen/ntry; 15 >> k=k 16 >> [x abs(x-xold)] 17 >> disp(['elapsed time is ', numstr(timen), ' seconds']) 18 6

7 19 k = 6 0 ans = Elapsed time is seconds (Sol b) 1 >> times=0; >> for nt=1:ntry 3 >> a=1; 4 >> clear b 5 >> tic 6 >> secant 7 >> times=times+toc; 8 >> end 9 >> times=times/ntry; 10 >> k=k 11 >> [b abs(b-a)] 1 >> disp(['elapsed time is ', numstr(times), ' seconds']) k = 7 15 ans = Elapsed time is seconds (Sol c) 1 >> timei=0; >> for nt=1:ntry 3 >> a=1; 4 >> clear b 5 >> clear c 6 >> tic 7 >> iquadi 8 >> timei=timei+toc; 9 >> end 10 >> timei=timei/ntry; 11 >> k=k 1 >> [c abs(c-b)] 13 >> disp(['elapsed time is ', numstr(timen), ' seconds']) k = 7 16 ans = Elapsed time is seconds (Sol d) The Newtons method is preferable since it requires less computational time than those of IQL and secant methods. Note: These 3 methods give roughly the same final approximate root and achieved tolerance. The number of iterations in Newtons method is the least, but every iteration in Newtons method requires the computer to evaluate f (x). This can cause this method to be slower that secant or IQl at each iteration for some problems since they do not need to do this function evaluation. 7

8 Problem 4. Derive the error formula for the secant method x x n+1 = (x x n )(x x n 1 ) f[x n 1, x n, x ] f[x n 1, x n ] in terms of the divided-differences. By direct computation: (x n x n 1 ) x x n+1 = x x n + f(x n ) f(x n ) f(x n 1 ) ( = (f(x n ) f(x n 1 )) (x ) x n ) (x n x n 1 ) f(x (xn x n 1 ) n) f(x n ) f(x n 1 ) ( = (x x n ) (f(x n) f(x n 1 )) f(x ) n) (xn x n 1 ) (x n x n 1 ) (x x n ) f(x n ) f(x n 1 ) ( = (x x n ) (f(x n) f(x n 1 )) + f(x ) ) f(x n ) (xn x n 1 ) (x n x n 1 ) (x x n ) f(x n ) f(x n 1 ) where we have made use of the fact that f(x ) = 0. Continuing on, we have x x n+1 = (x x n ) ( f[x n 1, x n ] + f[x n, x ]) /f[x n 1, x n ] ( ) f[xn, x ] f[x n 1, x n ] = (x x n )(x x n 1 ) /f[x n 1, x n ] (x x n 1 ) = (x x n )(x x n 1 ) f[x n 1, x n, x ]. f[x n 1, x n ] Problem 5. Numerically calculate the Newton iterates for solving x 1 = 0, and use x 0 = 5. Identify and explain the resulting speed of convergence. (Sol) 1 >> f=inline('xˆ-1'); >> Df=inline('*x'); 3 >> tol=eps; 4 5 >> MM=5; 6 >> x=ˆmm; 7 8 >> k=0; 9 >> xold=0; 10 >> itmax=50; 11 >> xx(k+1)=x; 1 >> kk(k+1)=k; >> while abs(x-xold)>tol*max(abs(x),1.0) 15 >> if k+1>itmax 16 >> break 17 >> end 18 >> xold=x; 19 >> k=k+1 0 >> x=x-f(x)/df(x); 1 >> xx(k+1)=x; 8

9 >> kk(k+1)=k; 3 >> [x abs(x-xold)] 4 >> end 5 6 k = 31 7 ans = >> plot(kk,log(xx),'-b',kk,mm-kk,'-r',kk,0*kk,'-k') 30 >> axis([0 k MM-k MM]) 31 >> xlabel('$k$','fontsize',15,'interpreter','latex') 3 >> ylabel('$\log (x k)$','fontsize',15,'interpreter','latex') 33 >> title('newton iterates $x k$ for $f(x)=xˆ-1$ and... $x 0=ˆ{5}$','FontSize',18,'Interpreter','latex') 34 >> legend('x k','ˆ{5-k}','1') We can understand the rate of convergence as follows. We have: x = x f(x) f (x) = x x 1 = 1 ( x + 1 ) x for x 1. x x thus the rate of convergence is 1/. Problem 6. Define an iteration formula by z n+1 = x n f(x n) f (x n ) x n+1 = z n+1 f(z n+1) f (x n ) This is like two Newton steps, with the second re-using the derivative from the first. (a) Show that the order of convergence of x n to x is at least 3. Hint: Exploit the theorem proved in class on order of convergence of fixed-point methods (generalizing Theorem.9 in Burden & Faires), and let h(x) = x f(x) f (x) g(x) = h(x) f(h(x)) f (x) 9

10 (b) Compare this method to Newton s method in terms of the number of function and derivative evaluations. If τ f is the time required to evaluate f and τ f is the additional time to evaluate f, then how large must τ f /τ f be in order for the new iteration to converge faster than Newton? (Sol a) We first calculate the derivatives of h: h (x) = f(x)f (x) (f (x)) h (x) = f (x) f (x) + f(x) d dx [ f ] (x) (f (x)) so h (x ) = 0 and h (x ) = f (x )/f (x ). Now we calculate the derivatives of g, ( g (x) = h (x) 1 f ) (h(x)) f + f(h(x))f (x) (x) f (x) ( g (x) = h (x) 1 f ) (h(x)) f + h (x) d ( 1 f ) (h(x)) (x) dx f (x) + f (h(x))h (x) f (x) f (x) + f(h(x)) d ( f ) (x) dx f (x) So that g (x ) = 0 and g (x ) = h (x) 0+0 d ( 1 f ) (h(x)) dx f +0 f (x) f (x) (x) f (x) +0 d dx x=x ( f ) (x) f (x) = 0! Since g(x ) = x 0 but g (x ) = g (x ) = 0, we can apply the theorem proved in class (pg 6 of Ch. of the notes), to say that we have order of convergence at least 3. (Sol b) The convergence times for Newton and the 3rd order method we have here are: τ newt = (τ f + τ f ) log K log τ 3rd = (τ f + τ f ) log K log 3 Thus τ 3rd τ newt = ( τf + τ f τ f + τ f ) log log 3 < 1 iff ( ) τf + τ f τ f + τ f < log 3 log = log 3 iff τ f τ f > log 3 log

11 Problem 7. Given below is a table of iterates from a linearly convergent iteration x n+1 = g(x n ). Estimate (a) the rate of linear convergence, (b) the fixed point x and (c) the error x 7 x. n x n (d) Use x 7 and the error estimate in (c) to make an improved estimate of the fixed point value x to which the sequence converges. 1 % g=inline('x+cos(x/)-sin(x/)') >> xi(1)=1; 3 >> xi()= ; 4 >> xi(3)= ; 5 >> xi(4)= ; 6 >> xi(5)= ; 7 >> xi(6)= ; 8 >> xi(7)= ; 9 >> xi(8)= ; >> x=xi(3:8); 1 >> xp=xi(:7); 13 >> xpp=xi(1:6); (Sol a) We estimate the rate of convergence as follows: λ n x n+1 x n x n x n 1, λ >> lamest=(x-xp)./(xp-xpp); 3 lamest = >> lamest=lamest(6) 6 lamest = (Sol b) The naive estimate is x 7 = (Sol c) We estimated the error as: x x 7 =

12 1 >> errest=lamest.*(x-xp)./(lamest-1); 3 errest = >> errest=errest(6) 6 errest = (Sol d) Exploiting our estimate of the error, we use Aitken extrapolation to obtain an improved estimate: x n = x n + = x n + λ n 1 λ n (x n x n 1 ) x n + (x n x n 1 ) (x n 1 x n ) (x n x n 1 ) x x n x n 1 x n 1 x n 1 xn xn 1 x n 1 x n (x n x n 1 ) 1 >> xx=x-errest; 3 xx = >> fixest=xx(6) 6 fixest = You may recognize this number x as π/ = ! Clearly the Aitken extrapolate x 7 gives a much more accurate estimate (relative error ) than does the naive guess x 7 (relative error ). 1