MATH 333: Partial Differential Equations

Transcription

1 MATH 333: Partial Differential Equations Problem Set 9, Final version Due Date: Tues., Nov. 29, 2011 Relevant sources: Farlow s book: Lessons 9, MacCluer s book: Chapter 3 44 Show that the Poisson problem with Dirichlet BCs on a bounded domain Ω is stable. Specifically, suppose that Ω R n is open, bounded and connected, and consider the two (nearly identical they both have the same forcing function f ) problems u(x) = f (x), x Ω, subject to u(x) = g 1 (x), for x Ω, (1) and v(x) = f (x), x Ω, subject to v(x) = g 2 (x), for x Ω. (2) If boundary data for the two problems are similar that is, if g 1 g 2 := max { g 1 (x) g 2 (x) : x Ω } is small show that the difference in solutions is small throughout Ω. 45 Set up the BVP corresponding to the thermal equilibrium of a rectangular plate that is insulated on two of its sides, has 0 at its top edge and 100 at the bottom edge. Where do you expect the maximum temperature to be located? What is its value? Can you find a formula for temperature inside the plate? 46 [This problem gives a partial answer to the question of what forms harmonic functions of 2 variables can take.] Find an harmonic function u(x, y) defined on the annulus 1 < x < 1 (where x = (x, y)) subject to the constant Dirichlet BCs u = a on x = and u = b on x = 1. You may assume, as seems physically reasonable given the BCs, that the solution is radially symmetric, so that u = u(r) is a function of r = x 2 + y 2 alone (independent of θ), and so Laplace s equation reduces to u = 2 u r + 1 u 2 r r = 0. Hint: Let v = u r and solve the 1 st order ODE for v. Then find u.

2 MATH 333 Problem Set Find a series solution of the heat problem u t = u xx, 0 < x < 1, t > 0, subject to BCs: u(t, 0) = 0, u(t, 1) = sin t, IC: u(0, x) = Consider the Poisson problem with Dirichlet BCs in a square region of the plane u(x, 0) = sin(πx), u = f (x, y), 0 < x < 1, 0 < y < 1, subject to BCs u(x, 1) = e 2 sin(πx), u(0, y) = 0 = u(1, y). where f (x, y) = (4 π 2 )e 2y sin(πx). (a) Employing a finite difference scheme, solve this problem on a grid that uses N = 4 interior values on the x-axis and M = 4 interior values on the y-axis to obtain the steady-state temperature distribution. Graph the resulting approximate solution as a surface in 3D. I encourage you to rely heavily on my class examples which used scripts like psset1.m, psset2.m, psset3.m, psset4.m, poissonsolver.m and discretelaplacian.m, available to download from scofield/courses/m333/materials/octave/ Pay special attention to how a forcing function is implemented in psset4.m (i.e., the form of its inputs and outputs). Hand in a printout of your code and the graphs you produce, and send me electronic versions as well. (b) Demonstrate that the true solution of the problem is u(x, y) = e 2y sin(πx), and explain why the problem is, therefore, well-posed. (c) Compare the values of the true solution with those from your approximate solution in part (a). Among all interior grid points find both the maximum error and the maximum relative error of the numerical solution. (d) Redo parts (a) and (c), this time doubling N and M so that they are both 8.

3 MATH 333 Problem Set 9 3 Work with a partner and choose one problem pair from 49A-B or 50A-B to do. 49A In class we considered Poisson s equation with Dirichlet BCs u = f, inside R = { (x, y) 0 < x < a, 0 < y < a }, subject to u = g on R. We replaced the PDE with a linear system of difference equations u j 1,m u j+1,m u j,m 1 u j,m+1 + 4u j,m = h 2 f j,m, 1 j N, 1 m M, (3) vectorized the unknowns U = [ u 1,1 u 2,1... u N,1 u 1,2 u 2,2... u N,2... u 1,M u 2,M... u N,M ] T, and then determined the matrix A such that AU corresponds to the left-hand side of Equation (3). (a) Suppose, instead of Dirichlet BCs, we have periodic BCs. That is, as you approach the right edge of the mesh of points (x j, y m ), you wrap around to the left side; the top wraps around to the bottom (like in the classic video game PacMan). For simplicity, assume there are only points (x j, y m ) for j = 1, 2,..., N, m = 1, 2,..., M, so that (x 1, y m ) is to the immediate right of (x N, y m ) and (x j, y 1 ) is immediately above (x j, y M ). Determine the form of the matrix A (give a picture) so that AU is the discrete Laplacian operator on u, taking into account these periodic BCs. [Note: The size of the correct matrix should be NM rows by NM columns.] (b) Write an Octave function that, given inputs N and M, returns the matrix A of part (a). The.m-file I wrote to generate the matrix A for class is discretelaplacian.m, and it actually contains two different ways to construct (my) A. 49B Obtain a small black-and-white image file (say, something on the order of M = 60 pixels by N = 60 pixels, so NM isn t horribly large), and place it in your working directory. I ve prepared the file testim.png which you may use, but you may substitute your own. (Most standard image formats are supported by Octave.) Read this file into Octave, convert it to double precision, and then display it using the commands image = double ( imread ( testim. png ) ) / ; imshow ( image ) Now add some Gaussian noise to the image and display the result via a command like:

4 MATH 333 Problem Set 9 4 nimage = image randn ( s ize ( image ) ) ; figure ( 2 ) imshow ( nimage ) (If you have the clean image displayed, the figure() command creates a 2 nd figure window so the two can be viewed side-by-side.) We will use diffusion to smooth out the noise in the image. discretization of the diffusion equation On a uniform spatial grid with x = y = h, one u t (t, x, y) = u(t, x, y) is or u(t + t, x, y) u(t, x, y) t = u(t, x h, y) + u(t, x + h, y) + u(t, x, y h) + u(t, x, y + h) 4u(t, x, y) h 2, u(t + t, x, y) = u(t, x, y) + t h 2 [ u(t, x h, y) + u(t, x + h, y) + u(t, x, y h) + u(t, x, y + h) 4u(t, x, y) ]. (4) Here, consider t to be the current time (so values at time t refer to the current state of the image, while those at time (t + t) refer to the next state) and, relative to a pixel at position (x, y), positions (x h, y), (x + h, y), (x, y h) and (x, y + h) are neighboring pixels to the left, right, below and above, respectively. You can vectorize the noisy image with a command like U = nimage ( : ) ; and return it to its original M-by-N shape with reshape (U, M, N) Now, if we call the vectorized form of your current image U (k), then Equation (4), which describes the evolution of your image at the local level, can be turned into the global (entire image at a time) scheme (where we have taken h = 1) U (k+1) = U (k) + tau (k). Here, A, the matrix you found for the previous exercise, has periodic BCs built into it. Assuming U (0) is the starting noisy image, you could allow the image intensity values to evolve up until, say, time t = 1 in just one step (so U (1) is the resulting image) by taking t = 1. You could also get to the same time via smaller steps, t = 0.1 for instance, so that the final image is U (10).

5 MATH 333 Problem Set 9 5 (a) Implement this global evolution scheme in Octave or Matlab. Here is pseudocode for what you are to do: 1. Set U to be the vectorized form of the noisy image. 2. Build your matrix A of appropriate size. 3. Set the value of dt, and the number of time steps to carry out. 4. Begin loop Let U = U + dt*a*u Plot the current state of the image (reshaped U). Pause briefly for the viewer to see the result. End loop You should hand in a hardcopy of your code and send me an electronic version as well. (b) Try out your code, independently altering the number of iterations to carry out and the time step dt. Comparing your final image to the original clean image, find a combination of these two settings that best pleases you in terms of removing noise but not blurring the image too badly. The code you send for part (a) should have these settings. (c) Along with a tendency to become blurred in time, what do you notice about the evolving image produced by your routine? Explain these observations. 50A Consider the ordinary differential equation/bvp (a 1-dimensional version of the Poisson problem) y (x) = f (x), subject to BCs y(0) = 0, y (1) = 1. (5) Let us subdivide the interval [0, 1] into n equally-spaced subintervals, 0 = x 0 < x 1 <... < x n = 1, with each x m x m 1 = h := 1 n. (a) If we use a centered difference approximation for y (x) then, writing y m as an approximation to the value of y(x m ), derive the finite difference equation y m 1 + 2y m y m+1 = h 2 f (x m ), (6) holding for m = 1, 2,..., n 1. (b) While equation (6) provides most of the particulars for a finite difference scheme to approximate the solution of Problem (5), we will not have a complete scheme until we specify how to handle the Neumann boundary condition at x = 1. Consider two different schemes, based on the following two ways of dealing with

6 MATH 333 Problem Set 9 6 this Neumann condition. In the simpler of the two schemes S 1, we approximate y (1) using a backward difference: y n y n 1 h = 1. In scheme S 2, we imagine there to be a ghost node x n+1 = 1 + h, outside the right end point (where the Neumann condition occurs) of our interval [0, 1], require equation (6) to hold at the right end point x n (i.e., to hold with m = n, which gives rise to an equation involving the ghost node x n+1 ), and introduce the constraint (approximate equation for y (1) = 1) y n+1 y n 1 2h = 1. Schemes S 1 and S 2 both may be turned into matrix problems A 1 v 1 = b 1 and A 2 v 2 = b 2, where v 1 = (y 1, y 2,..., y n ) and v 2 = (y 1, y 2,..., y n, y n+1 ) represent the vector of unknowns. Determine the entries of the matrices A 1, A 2 and the right-hand sides b 1, b 2. {c} This problem part is optional. Are the matrices A 1, A 2 symmetric? positive definite? (The definition of the matrix A is positive definite is that v T Av 0 for all vectors v of appropriate size, with equality only when v = 0.) If the answers are not obvious, then you may resort to use of numerical evidence. Be clear about what you know to be true (from numerical evidence), and what you only guess at being true. Note: Positive definite matrices are always nonsingular, making a problem like Av = b uniquely solvable. (d) Let f (x) = e x 1, so that the resulting true solution of (5) is y(x) = e 1 (e x 1). Write algorithms to compute finite difference solutions (with this f ) both for scheme S 1 and S 2. Compute the maximum error between true and approximate solutions (i.e, maximum difference y(x m ) y m taken over all grid points x = x m ) using both schemes at several different choices of h (say, h = 0.1 and h = 0.05, at the least). What can you say about the order of convergence for each scheme? 50B (a) Derive an implicit scheme for the following Neumann problem: BCs : u x (t, 0) = α(t), t > 0 u t = u xx, 0 < x < l, t > 0, subj. to u x (t, l) = β(t) IC : u(0, x) = f (x), 0 < x < l (7)

7 MATH 333 Problem Set 9 7 (b) Use your scheme to solve Problem (7) in the case where l = 4, α(t) = 0, β(t) = 100(1 e t ), and f (x) = 100, if 1 x 2, 0, otherwise.