Fourier Interpolation

In [2]: 
import numpy as np
import numpy.linalg as la
import matplotlib.pyplot as pt

Let's fix the number of points and the number of functions as n.

Make sure n is odd.

In [49]: 
n = 5

assert n % 2 == 1
In [50]: 
x = np.linspace(0, 2*np.pi, n, endpoint=False)

Next, fix the values of $k$ in $\cos(kx)$ as cos_k and in $sin(kx)$ as sin_k so that there are exactly $n$ altogether:

In [51]: 
cos_k = np.arange(0, n//2 + 1, dtype=np.float64)
sin_k = np.arange(1, n//2 + 1, dtype=np.float64)
In [52]: 
print(cos_k)
print(sin_k)

[ 0.  1.  2.]
[ 1.  2.]

Next, build the generalized Vandermonde matrix.

Make sure to order the matrix by increasing $k$:

In [53]: 
V = np.zeros((n,n))
V[:, ::2] = np.cos(cos_k*x[:, np.newaxis])
V[:, 1::2] = np.sin(sin_k*x[:, np.newaxis])
V
Out[53]:
array([[ 1.        ,  0.        ,  1.        ,  0.        ,  1.        ],
       [ 1.        ,  0.95105652,  0.30901699,  0.58778525, -0.80901699],
       [ 1.        ,  0.58778525, -0.80901699, -0.95105652,  0.30901699],
       [ 1.        , -0.58778525, -0.80901699,  0.95105652,  0.30901699],
       [ 1.        , -0.95105652,  0.30901699, -0.58778525, -0.80901699]])

Now let's try and do interpolation with this. Here are a few functions:

In [54]: 
if 1:
    def f(x):
        return x
elif 0:
    def f(x):
        return np.abs(x-np.pi)
elif 1:
    def f(x):
        return (x<=np.pi).astype(np.int32).astype(np.float64)

Find the coefficients as coeffs:

In [55]: 
coeffs = la.solve(V, f(x))
In [56]: 
plot_x = np.linspace(0, 2*np.pi, 1000)
In [57]: 
interpolant = 0 * plot_x
for i, k in enumerate(cos_k):
    interpolant += coeffs[2*i] * np.cos(k * plot_x)
for i, n in enumerate(sin_k):
    interpolant += coeffs[2*i+1] * np.sin(n * plot_x)
In [58]: 
pt.plot(plot_x, interpolant)
pt.plot(plot_x, f(plot_x), "--", color="gray")
pt.plot(x, f(x), "or")
Out[58]:
[<matplotlib.lines.Line2D at 0x7f9157556470>]
  • For $f(x)=x$, why does it do what it does at the interval end?
  • What happens if we increase the number of points?
  • What if the function is actually periodic (e.g. the function with abs above?)?
  • What if the function has jumps?
In [59]: 
# Answers
#
# * Because we're interpolating with periodic functions--so the interpolant is forced to be periodic.
# * We observe a distinct "overshoot". This overshoot is called "Gibbs phenomenon".
# * Periodic: no Gibbs at interval end.
# * Gibbs can also happen in the middle of the interval.

Computing the coefficients

In [60]: 
B = V.T.dot(V)
In [61]: 
B[np.abs(B)<1e-12] = 0
In [62]: 
B
Out[62]:
array([[ 5. ,  0. ,  0. ,  0. ,  0. ],
       [ 0. ,  2.5,  0. ,  0. ,  0. ],
       [ 0. ,  0. ,  2.5,  0. ,  0. ],
       [ 0. ,  0. ,  0. ,  2.5,  0. ],
       [ 0. ,  0. ,  0. ,  0. ,  2.5]])
  • What do you observe?
  • How could this be useful for computing the inverse?
  • What is the normalization?
  • This is a pretty special matrix. What is the cost of computing $Ax$ with it?
  • The so-called Fast Fourier transform (FFT) computes a version of $Ax$ (with complex exponentials) in $O(n\log n)$!
In [19]: 
# Answers:
# 
# * V's columns are orthogonal. (though not normalized)
# * The transpose of $V$ (with appropriate normalization) its inverse. This makes Fourier coefficients cheap to compute.
# * The normalization is n for the first entry, and n/2 for all the ones after that.
# * Computing Ax costs n**2 operations.