Sharedweek 13 / examples.ipynbOpen in CoCalc
Examples UQ
import numpy as np
import scipy.linalg as la
import matplotlib.pyplot as plt

Generate data

# grid
n = 500
x = np.linspace(0,1,n)
xx,yy = np.meshgrid(x,x)

# generate u with covariance C
Cm = np.exp(-np.abs(xx - yy)**1/(0.5))

u = np.random.multivariate_normal(np.zeros(n),Cm)
# Define G
K = np.linalg.matrix_power(0.5*np.diag(np.ones(n)) + 0.25*np.diag(np.ones(n-1),k=-1) + 0.25*np.diag(np.ones(n-1),k=1),250)

# data
sigma = 1e-3
Cd = sigma*np.eye(n)
f = np.random.multivariate_normal(K@u,Cd)
plt.plot(x,f)
[<matplotlib.lines.Line2D at 0x7f37d03aa828>]

Inversion

# define operators
Cm = np.exp(-np.abs(xx - yy)**1/(0.5))
Cd = np.eye(n)
K = np.linalg.matrix_power(0.5*np.diag(np.ones(n)) + 0.25*np.diag(np.ones(n-1),k=-1) + 0.25*np.diag(np.ones(n-1),k=1),50)
alpha = 0.01

# posterior mean
up = Cm@K.T@np.linalg.inv(K@Cm@K.T + alpha*Cd)@f

# posterior covariance
Cp = Cm - Cm@K.T@np.linalg.inv(K@Cm@K.T + alpha*Cd)@K@Cm

# draw samples
us = np.random.multivariate_normal(up,Cp,10)

# plot
plt.plot(x,us.T)
plt.plot(x,up,'k')
plt.plot(x,up + np.sqrt(np.diag(Cp)),'k--')
plt.plot(x,up - np.sqrt(np.diag(Cp)),'k--')
[<matplotlib.lines.Line2D at 0x7f37cea2ea58>]
# L-curve
ns = 10
a = np.logspace(-6,0,ns)
nk = np.zeros(ns)
rk = np.zeros(ns)
for k in range(ns):
    uk = Cm@K.T@np.linalg.inv(K@Cm@K.T + a[k]*Cd)@f
    nk[k] = np.dot(np.linalg.inv(Cm)@uk,uk)
    rk[k] = np.linalg.norm(K@uk - f)
plt.loglog(nk,rk,'k-o')
plt.loglog(nk[6],rk[6],'ro')
[<matplotlib.lines.Line2D at 0x7f37ceb5ab38>]
a[6]
0.01
# residual
r = K@up-f
plt.hist(r,10)
(array([ 4., 18., 39., 78., 92., 113., 83., 43., 24., 6.]), array([-0.08762024, -0.07050998, -0.05339971, -0.03628945, -0.01917919, -0.00206892, 0.01504134, 0.03215161, 0.04926187, 0.06637214, 0.0834824 ]), <a list of 10 Patch objects>)
# Resolution operator
R = K.T@np.linalg.inv(K@Cm@K.T + alpha*Cd)@K

plt.plot(x,R[:,int(n/2)])
[<matplotlib.lines.Line2D at 0x7f4cf5dca240>]

Final comparison

# ground truth
plt.plot(x,u,'r')
plt.plot(x,up,'k')
plt.plot(x,up + np.sqrt(np.diag(Cp)),'k--')
plt.plot(x,up - np.sqrt(np.diag(Cp)),'k--')
[<matplotlib.lines.Line2D at 0x7f37ce9750f0>]