Sparse Inverse Covariance Estimation

Introduction

Assume we are given i.i.d. observations \(x_i \sim N(0,\Sigma)\) for \(i = 1,\ldots,m\), and the covariance matrix \(\Sigma \in {\mathbf S}_+^n\), the set of symmetric positive semidefinite matrices, has a sparse inverse \(S = \Sigma^{-1}\). Let \(Q = \frac{1}{m-1}\sum_{i=1}^m (x_i - \bar x)(x_i - \bar x)^T\) be our sample covariance. One way to estimate \(\Sigma\) is to maximize the log-likelihood with the prior knowledge that \(S\) is sparse (Friedman, Hastie, and Tibshirani 2008), which amounts to the optimization problem:

\[ \begin{array}{ll} \underset{S}{\mbox{maximize}} & \log\det(S) - \mbox{tr}(SQ) \\ \mbox{subject to} & S \in {\mathbf S}_+^n, \quad \sum_{i=1}^n \sum_{j=1}^n |S_{ij}| \leq \alpha. \end{array} \]

The parameter \(\alpha \geq 0\) controls the degree of sparsity. The problem is convex, so we can solve it using CVXR.

Example

We’ll create a sparse positive semi-definite matrix \(S\) using synthetic data

set.seed(1)
n <- 10      ## Dimension of matrix
m <- 1000    ## Number of samples

## Create sparse, symmetric PSD matrix S
A <- rsparsematrix(n, n, 0.15, rand.x = stats::rnorm)
Strue <- A %*% t(A) + 0.05 * diag(rep(1, n))    ## Force matrix to be strictly positive definite

We can now create the covariance matrix \(R\) as the inverse of \(S\).

R <- base::solve(Strue)

As test data, we sample from a multivariate normal with the fact that if \(Y \sim N(0, I)\), then \(R^{1/2}Y \sim N(0, R)\) since \(R\) is symmetric.

x_sample <- matrix(stats::rnorm(n * m), nrow = m, ncol = n) %*% t(expm::sqrtm(R))
Q <- cov(x_sample)    ## Sample covariance matrix

Finally, we solve our convex program for a set of \(\alpha\) values.

Version 1.0 Note: Positive semi-definite variables are now designated using PSD = TRUE rather than the Semidef function!

alphas <- c(10, 8, 6, 4, 1)
if (packageVersion("CVXR") > "0.99-7") {
    S  <- Variable(n, n, PSD = TRUE)
} else {
    S <- Semidef(n)    ## Variable constrained to positive semidefinite cone
}

obj <- Maximize(log_det(S) - matrix_trace(S %*% Q))

S.est <- lapply(alphas,
                function(alpha) {
                    constraints <- list(sum(abs(S)) <= alpha)
                    ## Form and solve optimization problem
                    prob <- Problem(obj, constraints)
                    result <- solve(prob)
                    
                    ## Create covariance matrix
                    R_hat <- base::solve(result$getValue(S))
                    Sres <- result$getValue(S)
                    Sres[abs(Sres) <= 1e-4] <- 0
                    Sres
                })

In the code above, the Semidef constructor restricts S to the positive semidefinite cone. In our objective, we use CVXR functions for the log-determinant and trace. The expression matrix_trace(S %*% Q) is equivalent to `sum(diag(S %*% Q))}, but the former is preferred because it is more efficient than making nested function calls.

However, a standalone atom does not exist for the determinant, so we cannot replace log_det(S) with log(det(S)) since det is undefined for a Semidef object.

Results

The figures below depict the solutions for the above dataset with \(m = 1000, n = 10\), and \(S\) containing 26% non-zero entries, represented by the dark squares in the images below. The sparsity of our inverse covariance estimate decreases for higher \(\alpha\), so that when \(\alpha = 1\), most of the off-diagonal entries are zero, while if \(\alpha = 10\), over half the matrix is dense. At \(\alpha = 4\), we achieve the true percentage of non-zeros.

do.call(multiplot, args = c(list(plotSpMat(Strue)),
                            mapply(plotSpMat, S.est, alphas, SIMPLIFY = FALSE),
                            list(layout = matrix(1:6, nrow = 2, byrow = TRUE))))

Session Info

sessionInfo()

## R version 4.2.1 (2022-06-23)
## Platform: x86_64-apple-darwin21.6.0 (64-bit)
## Running under: macOS Ventura 13.0
## 
## Matrix products: default
## BLAS:   /usr/local/Cellar/openblas/0.3.21/lib/libopenblasp-r0.3.21.dylib
## LAPACK: /usr/local/Cellar/r/4.2.1_4/lib/R/lib/libRlapack.dylib
## 
## locale:
## [1] en_US.UTF-8/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## attached base packages:
## [1] grid      stats     graphics  grDevices datasets  utils     methods  
## [8] base     
## 
## other attached packages:
## [1] expm_0.999-6  Matrix_1.5-1  ggplot2_3.3.6 CVXR_1.0-11  
## 
## loaded via a namespace (and not attached):
##  [1] tidyselect_1.2.0   xfun_0.34          bslib_0.4.0        slam_0.1-50       
##  [5] lattice_0.20-45    Rmosek_10.0.25     colorspace_2.0-3   vctrs_0.5.0       
##  [9] generics_0.1.3     htmltools_0.5.3    yaml_2.3.6         gmp_0.6-6         
## [13] utf8_1.2.2         rlang_1.0.6        jquerylib_0.1.4    pillar_1.8.1      
## [17] glue_1.6.2         Rmpfr_0.8-9        withr_2.5.0        DBI_1.1.3         
## [21] Rcplex_0.3-5       RColorBrewer_1.1-3 bit64_4.0.5        scs_3.0-1         
## [25] lifecycle_1.0.3    stringr_1.4.1      munsell_0.5.0      blogdown_1.13     
## [29] gtable_0.3.1       gurobi_9.5-2       codetools_0.2-18   evaluate_0.17     
## [33] labeling_0.4.2     knitr_1.40         fastmap_1.1.0      fansi_1.0.3       
## [37] cccp_0.2-9         highr_0.9          Rcpp_1.0.9         scales_1.2.1      
## [41] cachem_1.0.6       jsonlite_1.8.3     farver_2.1.1       bit_4.0.4         
## [45] digest_0.6.30      stringi_1.7.8      bookdown_0.29      dplyr_1.0.10      
## [49] Rglpk_0.6-4        cli_3.4.1          tools_4.2.1        magrittr_2.0.3    
## [53] sass_0.4.2         tibble_3.1.8       pkgconfig_2.0.3    rcbc_0.1.0.9001   
## [57] assertthat_0.2.1   rmarkdown_2.17     R6_2.5.1           compiler_4.2.1

Source

R Markdown