Solution Manual for: Numerical Computing with MATLAB by Cleve B. Moler

Transcription

1 Solution Manual for: Numerical Computing with MATLAB by Cleve B. Moler John L. Weatherwax July 5, 007 Chapter 7 (Ordinary Differential Equations) Problem 7.1 Defining the vector y as y Then we have for its time derivative the following v dt u v ü v u v u y y 1 (1) y 3 y 4 + cos(r) 1+t sin(r) 1+t () Where r given in terms of components of y is r y4 + y 3 (3) Given the initial conditions in terms of u, v, u, and v we have 0 y (4) 0 wax@alum.mit.edu 1

2 Problem 7. See the matlab code in the file prob 7.m. Problem 7.3 The algorithm BS3 can be represented by the following sequence of steps s 1 f(t n, y n ) (5) s f(t n + h, y n + h s 1) (6) s 3 f(t n h, y n s ) (7) y n+1 y n + h 9 (s 1 + 3s + 4s 3 ) (8) Part (a): The specific ODE s (and their exact solutions) to consider are given by dt dt dt dt 1 y t + C (9) t y t + C (10) t y t3 3 + C (11) t 3 y t4 4 + C (1) For the first system, /dt 1, so f(t, y) 1 then the BS3 procedure gives the following s 1 1 (13) s 1 (14) s 3 1 (15) y n+1 y n + h 9 ( ) y n + h (16) Does this equal the true solution at t n + h? We evaluate the above to obtain y(t n + h) (t n + h) + C t n + C + h y n + h (17) and the answer is yes. For the second function f(t, y) t we see that the BS3 procedure is s 1 t n (18) s t n + h (19) s 3 t n h (0) y n+1 y n + h (t n + 3(t n + h ) + 4(t n + 3 h)) (1) 4

3 Simplifying this last equation we have y n + h 9 (9t n + ( 3 + 3)h) () y n + ht n + h 9 (9 )h (3) y n + ht n + h Again we ask, does this equal the true solution at t n+1? Lets check by evaluating (4) y(t n+1 ) y(t n + h) 1 (t n + h) + C t n + C + ht n + h y(t n) + ht n + h and we see that the answer is yes. For the second function f(t, y) t, then the BS3 gives (5) s 1 t n (6) s (t n + h ) (7) s 3 (t n h) (8) y n+1 y n + h 9 (t n + 3(t n + h ) + 4(t n h) ) (9) Now simplifying the last equation we have y n+1 y n + h 9 (t n + (3 + 6)ht n + ( )h ) (30) y n + h 9 (9t n + 9 ht n h ) (31) y n + ht n + h t n + h Does this equal the exact solution at t n + h? To decide we compute (3) y(t n + h) (t n + h) 3 + C (t3 n + 3t n h + 3t nh + h 3 ) + C (33) t3 n 3 + C + (t n h + t nh + h3 3 ) (34) and we see that the answer is yes. Finally, for the function f(t, y) t 3, the BS3 algorithm gives s 1 t 3 n (35) s (t n + h )3 t 3 n + 3 t n h t nh h3 (36) s 3 (t n h)3 t 3 n t n h t nh h3 (37) y n+1 y n + h 9 (t3 n + 3(t3 n + 3 t n h t nh h3 ) (38) +4(t 3 n t n h t nh h3 )) (39)

4 Simplifying this last term gives y n+1 y n + ht 3 n + 3 t n h + t n h 3 + h4 6 The true solution at t n + h however is equal to (40) y(t n + h) (t n + h) 4 + C (t4 n + 4t3 n h + 6t n h + 4t n h 3 + h 4 ) + C (41) 1 4 t4 n + C + t 3 nh + 3 t nh + t n h 3 + h4 4 (4) y n + ht 3 n + 3 t nh + t n h 3 + h4 4. (43) Since this differs from the the approximate solution by ( ) h 4 we see that this method is not exact for f(t) t 3. Part (b): The BS3 error estimate is given by e n+1 h 7 ( 5s 1 + 6s + 8s 3 9s 4 ) (44) where s i for i 1,, 3, 4 is computed as above. When one computes s 1, s, s 3, and s 4 for f(t) 1 and f(t) t one obtains a zero estimation error. Considering the function f(t) t we have an estimated error given by e n+1 h 7 ( 5t n + 6(t n + h ) + 8(t n h) 9(t n + h) ) (45) when one expands the above expression one obtains e n+1 h3 4 which is not zero. Thus the estimated error is exact for f(t) 1 and f(t) t only. (46) Problem 7.4 See the matlab code in the file prob 7 4.m. Problem 7.5 Part (a): See the matlab code in the file myrk4.m. Part (b): The local truncation error of RK4 is given by O(h 4+1 ), so the global truncation error should be O(h 4 ). When you decrease the step size h by this should reduce the global error by 1/ 4 1/16. An experiment illustrating this behaviour can be found in

5 the routine/code prob 7 5.m. There we sequentially decrease the stepsize parameter h and compute the global error between the final timestep and the known true solution value. We then plot the log of stepsize parameter v.s. the log of the global error. On this plot one can see that in this space we have a linear decrease in error representing a powerlaw. A linear function is fitted to to this data with the Matlab function polyfit. The slope of this line is given by 4 empirically displaying the known relation between the error and the stepsize of error h 4 Part (c): For this problem formulation our differential equation is given by ÿ y. The system formulation for this equation is given by defining [ ] [ ] y1 (t) y(t) Y (47) y (t) ẏ(t) then d dt Y (t) d [ y(t) dt ẏ(t) ] [ ẏ(t) ÿ(t) ] so in total in terms of y 1 and y the differential equation is [ ] [ ] d y1 y dt y y 1 [ y (t) y 1 (t) ] (48) (49) This is implemented in the Matlab code prob 7 5 fn.m. In addition, in the driver routine prob 7 5.m we have implemented the numerical experiments mentioned in the book and have verified the statements there. Problem 7.6 Part (a): The differential equation ẏ 1000y sin(t) + cos(t) with y(0) 1 (50) has a homogeneous solution given by y(t) Ce 1000t and a particular solution given by a function of the form A sin(t)+b cos(t). By putting this particular solution into the equation above we can solve for the coefficients A and B. When we do this we get the following A cos(t) B sin(t) 1000A sin(t) 1000B cos(t) sin(t) + cos(t) (51) Grouping the coefficients of cos(t) and sin(t) we obtain (A B 1) cos(t) + ( B A 1000) sin(t) 0. (5) When we require that the coefficients of cos(t) and sin(t) vanish we obtain the following system of equations A B 1 0 (53) B A (54)

6 The solution to the first equation in terms of B is given by A B. When substituted into the second equation we obtain 1000(1 1000B) B 1000 (55) Which has as its solution B 0. Thus A 1 giving in total the solution y of y(t) sin(t) + Ce 1000t (56) using the initial condition y(0) 1 we obtain C 1 and the total analytical solution is given by y(t) e 1000t + sin(t) (57) I should mention that heuristically this problem is midly stiff because of the various time scales present in the solution. The exponential term has a timescale O(1/1000) while the sin term has a timescale like O(1) or effectivly three orders of magnitude different. Part (b)-(f): See the routines prob 7 6.m and prob 7 6 fn.m for an implementation of this problem. Problem 7.7 Part (a): Each equation has as its solution y(t) sin(t). Part (b): The corresponding first order system for the third given differential equations is (defining y 1 (t) y(t)) [ ] [ ] [ ] [ ] d y1 y 0 1 y1 (58) dt y y y The corresponding system for the fourth differential equation is given by [ ] [ ] [ ] [ ] [ ] d y1 y 0 1 y1 0 + dt y sin(t) 0 0 y sin(t) Note that the first representation above is autonomous while the second is not. Part (c): The Jacobian for each of the functions f involved in the definitions of each differential equation are given by f 1 y f y f 3 y f 4 y 0 y 1 y [ ] [ ] From the above we can see that as t π that WWX finish (59)

7 Problem 7.9 Consider the matrix A given by A 0 η 0 σ σ η ρ 1 (60) Expanding the determinant of the above matrix we have or or A σ σ ρ 1 η 0 η σ σ (61) A (σ σρ) η(ση) (6) A η ρ + (63) Setting this equal to zero to find the values where the matrix is singular we see that η ± (ρ 1). (64) Which was the expression we were looking to find. To find the vectors in the nullity of A when η is one of these special values, insert this into the matrix above and reduce it to row-echelon form. Inserting this expression into A means we are looking for a vector v such that Av 0 ± (ρ 1) 0 σ σ (ρ 1) ρ 1 v 1 v v 3 0 (65) dividing the first row in the above expression by and the second row by σ we obtain (ρ 1) 1 0 v v 0 (66) (ρ 1) ρ 1 v 3 Now multiplying the first row by ± (ρ 1) and adding it to the third row we obtain (ρ 1) ρ 1 (ρ 1) (ρ 1) v 1 v v 3 0 (67) which simplifies to (ρ 1) ρ ρ v 1 v v 3 0. (68)

8 which demonstrates that the third row is indeed a multiple of the second row. Eliminating this row we obtain the following requirements among the coefficients of v (ρ 1) v 1 ± v 3 (69) v v 3 (70) Letting v 3 η (ρ 1) and assuming that ρ 1 > 0 we see that the above null vector has components given by as desired. v 1 ρ 1 (71) v η (7) v 3 η (73) Problem 7.10 (the Jacobian of the Lorenz equations) The Lorenz equations are given by ẏ Ay, with y 1 (t) y(t) y (t) and A y 3 (t) 0 y (t) 0 σ σ y (t) ρ 1 Performing the matrix multiplication we have for the full non-linear system representing the Lorenz equations, the following y 1 y 1 (t) + y (t)y 3 (t) y σy (t) + σy 3 (t) y 3 y 1 (t)y (t) + ρy (t) y 3 (t). This expression makes it easier to compute the required Jacobian of our mapping f( ). We find that y 3 (t) y (t) J 0 σ σ. y (t) y 1 (t) + ρ 1 The fixed points of the Lorentz equations are given by solving for y in f(y) 0, which gives the following system of equations y 1 + y y 3 0 σy + σy 3 0 y 1 y + ρy y 3 0. The second equation above has the simple solution of y y 3, which when used in the first and third equaions above gives y 1 + y 0 y 1 y + ρy y 1 0.,

9 The last equation above can be factored as y ( y 1 + ρ 1) 0, which has as solutions y 0, and y 1 ρ 1. Following the consequences of each of these choices, if y 0, then y 1 0, while if y 1 ρ 1 then the first equation above gives (ρ 1) + y 0, which has a solution given by y ± (ρ 1). Thus in summary then the three fixed points of the Jacobian of the Lorentz equations are (y 1, y, y 3 ) (0, 0, 0) (y 1, y, y 3 ) (ρ 1, (ρ 1), (ρ 1)) (y 1, y, y 3 ) (0, 0, 0).