Lab 1
A Pseudospectral method
for periodic functions
Lab Objective: We look at a pseudospectral method with a Fourier basis, and
numerically solve the advection equation using a pseudospectral discretization in
space and a Runge-Kutta integration scheme in time.
Let f be a periodic function on [0, 2π]. Let x1 , . . . , xN be N evenly spaced grid
points on [0, 2π]. Since f is periodic on [0, 2π], we can ignore the grid point xN = 2π.
We will further assume that N is even; similar formulas can be derived for N odd.
Let h = 2π/N ; then {x0 , . . . , xN −1 } = {0, h, 2h, . . . , 2π − h}.
The discrete Fourier transform (DFT) of f , denoted by fˆ or F(f ), is given by
N
X −1
fˆ(k) = h e−ikxj f (xj ) where k = −N/2 + 1, . . . , 0, 1, . . . , N/2.
j=0
The inverse DFT is then given by
N/2
1 X eikxj ˆ
f (xj ) = f (k), j = 0, . . . , N − 1, (1.1)
2π ck
k=−N/2
where
(
2 if k = −N/2 or k = N/2,
ck = (1.2)
1 otherwise.
The inverse DFT can then be used to define a natural interpolant (sometimes called
a band-limited interpolant) by evaluating (1.1) at any x rather than xj :
N/2
1 X
p(x) = eikx fˆ(k). (1.3)
2π
k=−N/2
The interpolant for f 0 is then given by
N/2−1
1 X
p0 (x) = ikeikx fˆ(k). (1.4)
2π
k=−N/2+1
1
2 Lab 1. A Pseudospectral method for periodic functions
Consider the function u(x) = sin2 (x) cos(x) + e2 sin(x+1) . Using (1.4), the deriva-
tive u0 may be approximated with the following code. 1 We note that although we
only approximate u0 at the Fourier grid points, (1.4) provides an analytic approxi-
mation of u0 in the form of a trigonometric polynomial.
import numpy as np
from scipy.fftpack import fft, ifft
import matplotlib.pyplot as plt
N=24
x1 = (2.*np.pi/N)*np.arange(1,N+1)
f = np.sin(x1)**2.*np.cos(x1) + np.exp(2.*np.sin(x1+1))
# This array is reordered in Python to
# accomodate the ordering inside the fft function in scipy.
k = np.concatenate(( np.arange(0,N/2) ,
np.array([0]) , # Because hat{f}'(k) at k = N/2 is zero.
np.arange(-N/2+1,0,1) ))
# Approximates the derivative using the pseudospectral method
f_hat = fft(f)
fp_hat = ((1j*k)*f_hat)
fp = np.real(ifft(fp_hat))
# Calculates the derivative analytically
x2 = np.linspace(0,2*np.pi,200)
derivative = (2.*np.sin(x2)*np.cos(x2)**2. -
np.sin(x2)**3. +
2*np.cos(x2+1)*np.exp(2*np.sin(x2+1))
)
plt.plot(x2,derivative,'-k',linewidth=2.)
plt.plot(x1,fp,'*b')
plt.savefig('spectral2_derivative.pdf')
plt.show()
Problem 1. Consider again the function u(x) = sin2 (x) cos(x) + e2 sin(x+1) .
Create a function that approximates 12 u00 − u0 on the Fourier grid points for
N = 24.
The advection equation
Recall that the advection equation is given by
ut + cux = 0 (1.5)
where c is the speed of the wave (the wave travels to the right for c > 0). We
will consider the solution of the advection equation on the circle; this essentially
amounts to solving the advection equation on [0, 2π] and assuming periodic bound-
ary conditions.
1 See Spectral Methods in MATLAB by Lloyd N. Trefethen. Another good reference is Cheby-
shev and Fourier Spectral Methods by John P. Boyd.
3
8
0 1 2 3 4 5 6 7
Figure 1.1: The derivative of u(x) = sin2 (x) cos(x) + e2 sin(x+1) .
A common method for solving time-dependent PDEs is called the method of
lines. To apply the method of lines to our problem, we use our Fourier grid points in
[0, π]: given an even N , let h = 2π/N , so that {x0 , . . . , xN −1 } = {0, h, 2h, . . . , 2π −
h}. By using these grid points we obtain the collection of equations
ut (xj , t) + cux (xj , t) = 0, t > 0, j = 0, . . . N − 1. (1.6)
−1
Let U (t) be the vector valued function given by U (t) = (u(xj , t))N j=0 . Let
F(U )(t) denote the discrete Fourier transform of u(x, t) (in space), so that
N/2
F(U )(t) = (û(k, t))k=−N/2+1 .
Define F −1 similarly. Using the pseudospectral approximation in space leads to the
system of ODEs
Ut + ~cF −1 i~kF(U ) = 0 (1.7)
where ~k is a vector, and ~kF(U ) denotes element-wise multiplication. Similarly ~c
could also be a vector, if the wave speed c is allowed to vary.
4 Lab 1. A Pseudospectral method for periodic functions
Problem 2. Using a fourth order Runge-Kutta method (RK4), solve the
initial value problem
ut + c(x)ux = 0, (1.8)
2
where c(x) = .2 + sin2 (x − 1), and u(x, t = 0) = e−100(x−1) . Plot your
numerical solution from t = 0 to t = 8 over 150 time steps and 100 x steps.
Note that the initial data is nearly zero near x = 0 and 2π, and so we can
use the pseudospectral method. a Use the following code to help graph.
t_steps = 150 # Time steps
x_steps = 100 # x steps
'''
Your code here to set things up
'''
sol = # RK4 method. Should return a t_steps by x_steps array
X,Y = np.meshgrid(x_domain, t_domain)
fig = plt.figure()
ax = fig.add_subplot(111, projection="3d")
ax.plot_wireframe(X,Y,sol)
plt.show()
a This problem is solved in Spectral Methods in MATLAB using a leapfrog discretization
in time.
5
Figure 1.2: The solution of the variable speed advection equation; see Problem 2.
