Jacobi vs Conjugate Gradient

Copyright (C) 2010-2020 Luke Olson
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.

Let's consider solving the discretized 2D Poisson equation with iterative methods.

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

n=16
T = 2*np.eye(n)-np.diag(np.ones(n-1),1)-np.diag(np.ones(n-1),-1)
A = np.kron(np.eye(n),T)+np.kron(T,np.eye(n))
pt.spy(A)

<matplotlib.image.AxesImage at 0x7f11e955ae48>

Define a right-hand side and solve the resulting system directly.

h = 1/(n-1)
b = h*np.arange(0,n*n)
x = la.solve(A,b)

Split the matrix into its diagonal and strictly lower/upper triangular parts.

d = np.diag(A)
D = np.diag(d)
L = np.tril(A,-1)
U = np.triu(A,1)
la.norm(A-(D+L+U))

0.0

Jacobi iteration proceeds using $$\boldsymbol x^{(i+1)} = \boldsymbol D^{-1}(\boldsymbol b- (\boldsymbol L+\boldsymbol U)\boldsymbol x^{(i)}).$$

def jacobi(niter,x0):
    xi = x0
    for i in range(niter):
        xi = np.diag(1./d)@(b-(L+U)@xi)
    return xi

niters = np.asarray(2**np.arange(4,12),dtype=int)

x0 = np.random.random(n*n)

jacobi_results = []
err = []

for niter in niters:
    jacobi_results.append(jacobi(niter,x0.copy()))
    err.append(la.norm(jacobi_results[-1]-x))

pt.plot(niters,err)
pt.yscale('log')
pt.xscale('log')

def cg(A,b,niter,x0):
    rk = b - A @ x0
    sk = rk
    xk = x0
    for i in range(niter):
        alpha = np.inner(rk,rk)/np.inner(sk, A @ sk)
        xk1 = xk + alpha * sk
        rk1 = rk - alpha * A @ sk
        beta = np.inner(rk1,rk1)/np.inner(rk,rk)
        sk1 = rk1 + beta*sk
        rk = rk1
        xk = xk1
        sk = sk1
    return xk

iters = np.asarray(2**np.arange(2,7),dtype=int)

x0 = np.random.random(n*n)

cg_results = []
err = []

for niter in iters:
    cg_results.append(cg(A,b,niter,x0.copy()))
    err.append(la.norm(cg_results[-1]-x))

pt.plot(iters,err)
pt.yscale('log')
pt.xscale('log')