Sparse covariance estimation for Gaussian variables¶
A derivative work by Judson Wilson, 5/22/2014.
Adapted (with significant improvements and fixes) from the CVX example of the same name, by Joelle Skaf, 4/24/2008.
Topic References:
- Section 7.1.1, Boyd & Vandenberghe "Convex Optimization"
Introduction¶
Suppose $y \in \mathbf{\mbox{R}}^n$ is a Gaussian random variable with zero mean and covariance matrix $R = \mathbf{\mbox{E}}[yy^T]$, with sparse inverse $S = R^{-1}$ ($S_{ij} = 0$ means that $y_i$ and $y_j$ are conditionally independent). We want to estimate the covariance matrix $R$ based on $N$ independent samples $y_1,\dots,y_N$ drawn from the distribution, and using prior knowledge that $S$ is sparse
A good heuristic for estimating $R$ is to solve the problem $$\begin{array}{ll} \mbox{minimize} & \log \det(S) - \mbox{tr}(SY) \\ \mbox{subject to} & \sum_{i=1}^n \sum_{j=1}^n |S_{ij}| \le \alpha \\ & S \succeq 0, \end{array}$$ where $Y$ is the sample covariance of $y_1,\dots,y_N$, and $\alpha$ is a sparsity parameter to be chosen or tuned.
Generate problem data¶
import cvxpy as cvx
import numpy as np
import scipy as scipy
# Fix random number generator so we can repeat the experiment.
np.random.seed(0)
# Dimension of matrix.
n = 10
# Number of samples, y_i
N = 1000
# Create sparse, symmetric PSD matrix S
A = np.mat(np.random.randn(n, n)) # Unit normal gaussian distribution.
A[scipy.sparse.rand(n, n, 0.85).todense().nonzero()] = 0 # Sparsen the matrix.
Strue = A*A.T + 0.05 * np.matrix(np.eye(n)) # Force strict pos. def.
# Create the covariance matrix associated with S.
R = np.linalg.inv(Strue)
# Create samples y_i from the distribution with covariance R.
y_sample = scipy.linalg.sqrtm(R) * np.matrix(np.random.randn(n, N))
# Calculate the sample covariance matrix.
Y = np.cov(y_sample)
Solve for several $\alpha$ values¶
# The alpha values for each attempt at generating a sparse inverse cov. matrix.
alphas = [10, 2, 1]
# Empty list of result matrixes S
Ss = []
# Solve the optimization problem for each value of alpha.
for alpha in alphas:
# Create a variable that is constrained to the positive semidefinite cone.
S = cvx.semidefinite(n)
# Form the logdet(S) - tr(SY) objective. Note the use of a set
# comprehension to form a set of the diagonal elements of S*Y, and the
# native sum function, which is compatible with cvxpy, to compute the trace.
# TODO: If a cvxpy trace operator becomes available, use it!
obj = cvx.Maximize(cvx.log_det(S) - sum([(S*Y)[i, i] for i in range(n)]))
# Set constraint.
constraints = [cvx.sum_entries(cvx.abs(S)) <= alpha]
# Form and solve optimization problem
prob = cvx.Problem(obj, constraints)
prob.solve()
if prob.status != cvx.OPTIMAL:
raise Exception('CVXPY Error')
# If the covariance matrix R is desired, here is how it to create it.
R_hat = np.linalg.inv(S.value)
# Threshold S element values to enforce exact zeros:
S = S.value
S[abs(S) <= 1e-4] = 0
# Store this S in the list of results for later plotting.
Ss += [S]
print 'Completed optimization parameterized by alpha =', alpha
Result plots¶
import matplotlib.pyplot as plt
# Show plot inline in ipython.
%matplotlib inline
# Plot properties.
plt.rc('text', usetex=True)
plt.rc('font', family='serif')
# Create figure.
plt.figure()
plt.figure(figsize=(12, 12))
# Plot sparsity pattern for the true covariance matrix.
plt.subplot(2, 2, 1)
plt.spy(Strue)
plt.title('Inverse of true covariance matrix', fontsize=16)
# Plot sparsity pattern for each result, corresponding to a specific alpha.
for i in range(len(alphas)):
plt.subplot(2, 2, 2+i)
plt.spy(Ss[i])
plt.title('Estimated inv. cov matrix, $\\alpha$={}'.format(alphas[i]), fontsize=16)
