Orthogonal polynomials

Copyright (C) 2020 Andreas Kloeckner

MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

import numpy as np
import numpy.linalg as la
import matplotlib.pyplot as pt

Mini-Introduction to sympy

import sympy as sym

# Enable "pretty-printing" in IPython
sym.init_printing()

Make a new Symbol and work with it:

x = sym.Symbol("x")

myexpr = (x**2-3)**2
myexpr

$$\left(x^{2} - 3\right)^{2}$$

myexpr = (x**2-3)**2
myexpr
myexpr.expand()

$$x^{4} - 6 x^{2} + 9$$

sym.integrate(myexpr, x)

$$\frac{x^{5}}{5} - 2 x^{3} + 9 x$$

sym.integrate(myexpr, (x, -1, 1))

$$\frac{72}{5}$$

Orthogonal polynomials

Now write a function inner_product(f, g):

def inner_product(f, g):
    return sym.integrate(f*g, (x, -1, 1))

Show that it works:

inner_product(1, 1)

$$2$$

inner_product(1, x)

$$0$$

Next, define a basis consisting of a few monomials:

basis = [1, x, x**2, x**3]
#basis = [1, x, x**2, x**3, x**4, x**5]

And run Gram-Schmidt on it:

orth_basis = []

for q in basis:
    for prev_q in orth_basis:
        q = q - inner_product(prev_q, q)*prev_q / inner_product(prev_q,prev_q)
    orth_basis.append(q)

legendre_basis = [orth_basis[0],]

#to compute Legendre polynomials need to normalize so that q(1)=1 rather than ||q||=1
for q in orth_basis[1:]:
    q = q / q.subs(x,1)
    legendre_basis.append(q)

legendre_basis

$$\left [ 1, \quad x, \quad \frac{3 x^{2}}{2} - \frac{1}{2}, \quad \frac{5 x^{3}}{2} - \frac{3 x}{2}, \quad \frac{35 x^{4}}{8} - \frac{15 x^{2}}{4} + \frac{3}{8}, \quad \frac{63 x^{5}}{8} - \frac{35 x^{3}}{4} + \frac{15 x}{8}\right ]$$

These are called the Legendre polynomials.

---

What do they look like?

mesh = np.linspace(-1, 1, 100)

pt.figure(figsize=(8,8))
for f in legendre_basis:
    f = sym.lambdify(x, f)
    pt.plot(mesh, [f(xi) for xi in mesh])

---

These functions are important enough to be included in scipy.special as eval_legendre:

import scipy.special as sps

for i in range(10):
    pt.plot(mesh, sps.eval_legendre(i, mesh))

What can we find out about the conditioning of the generalized Vandermonde matrix for Legendre polynomials?

#keep
n = 20
xs = np.linspace(-1, 1, n)
V = np.array([
    sps.eval_legendre(i, xs)
    for i in range(n)
]).T

la.cond(V)

$$7274.598185486346$$

The Chebyshev basis can similarly be defined by Gram-Schmidt, but now with respect to a different inner-product weight function, $$w(x) = 1/\sqrt{1-x^2}.$$

w = 1 / sym.sqrt(1-x**2)
def cheb_inner_product(f, g):
    return sym.integrate(w*f*g, (x, -1, 1))

orth_basis = []

for q in basis:
    for prev_q in orth_basis:
        q = q - cheb_inner_product(prev_q, q)*prev_q / cheb_inner_product(prev_q,prev_q)
    orth_basis.append(q)

cheb_basis = [1,]

#to compute Legendre polynomials need to normalize so that q(1)=1 rather than ||q||=1
for q in orth_basis[1:]:
    q = q / q.subs(x,1)
    cheb_basis.append(q)
cheb_basis

$$\left [ 1, \quad x, \quad 2 x^{2} - 1, \quad 4 x^{3} - 3 x, \quad 8 x^{4} - 8 x^{2} + 1, \quad 16 x^{5} - 20 x^{3} + 5 x\right ]$$

for i in range(10):
    pt.plot(mesh, np.cos(i*np.arccos(mesh)))

Chebyshev polynomials achieve similar good, but imperfect conditioning on a uniform grid (but perfect conditioning on a grid of Chebyshev nodes).

#keep
n = 20
xs = np.linspace(-1, 1, n)
V = np.array([
    np.cos(i*np.arccos(xs))
    for i in range(n)
]).T

la.cond(V)

$$4846.7105682362735$$