@dppalomar I think I'm missing something on the glasso package?
The estimated covariance matrix is similiar to the sample covariance matrix, however, their generalized inverse are quite different. See the following example:

library(spectralGraphTopology)
library(glasso)

set.seed(123)
w <- c(1, 2, 3)
Lw <- L(w)
Y <- MASS::mvrnorm(100, rep(0, nrow(Lw)), MASS::ginv(Lw))

# True Laplacian Matrix
Lw
#      [,1] [,2] [,3]
# [1,]    3   -1   -2
# [2,]   -1    4   -3
# [3,]   -2   -3    5

# Sample Covariance Matrix
cov(Y)
#             [,1]        [,2]        [,3]
# [1,]  0.12937250 -0.07160465 -0.05776785
# [2,] -0.07160465  0.10031397 -0.02870933
# [3,] -0.05776785 -0.02870933  0.08647718

# Naive Estimator
MASS::ginv(cov(Y))
#          [,1]      [,2]      [,3]
# [1,]  3.456331 -1.434415 -2.021916
# [2,] -1.434415  4.690137 -3.255723
# [3,] -2.021916 -3.255723  5.277639

# glasso
gl <- glasso(cov(Y), rho = 1e-2)
gl

# $w (Graphical Lasso Covariance Matrix)
#             [,1]        [,2]        [,3]
# [1,]  0.13937250 -0.06160457 -0.04776738
# [2,] -0.06160457  0.11031397 -0.01870933
# [3,] -0.04776738 -0.01870933  0.09647718
#
# $wi (Graphical Lasso Inverse Covariance Matrix)
#           [,1]      [,2]      [,3]
# [1,] 14.567376  9.676671  9.089026
# [2,]  9.676661 15.801224  7.855262
# [3,]  9.089156  7.855371 16.388597

• refine initial guess of w by solving the following QP: min w: ||vec(L_naive) - vec(Lw)||^2
• add plots of the error criterion as a function of the number of observations
• compare the performance of spectralGraphTopology against glasso(http://statweb.stanford.edu/~tibs/glasso/)
• prepare detailed examples for Jiaxi to investigate why Situation 3 doesn't seem to occur
• look at qgraph or similar packages to draw graphs
• add documentation for the C++ functions
• investigate whether there is a hack to render LaTeX's algorithm2e in the HTML version of the vignette
• general notation improvements
• add input validation to the exported functions

@dppalomar please do add more tasks as you see fit.

• ISCM ---> Naive
• LLQP ---> QP
• SGL ---> SGL (proposed)
• CGL(A) ---> CGL (clairvoyant)
• N ---> n
• T ---> p
• Same orientation for learned graph plots
• F-score for K > 7

Problem description

Package is declared in Suggests, but it is used on package load. As a result, it doesn't work on systems without CVXR.

Example

See the build log:

** R
** inst
** byte-compile and prepare package for lazy loading
Error in library(CVXR) : there is no package called 'CVXR'
Error: unable to load R code in package 'spectralGraphTopology'
Execution halted
ERROR: lazy loading failed for package 'spectralGraphTopology'
* removing '/builddir/build/BUILDROOT/R-CRAN-spectralGraphTopology-0.2.3-1.fc37.copr4853884.x86_64/usr/local/lib/R/library/spectralGraphTopology'
error: Bad exit status from /var/tmp/rpm-tmp.gMNrVw (%install)

Expected behavior

The package should work without CVXR, or CVXR should be promoted to Imports or Depends.

Environment:

• platform (e.g. Linux, OSX, Windows): Linux
• spectralGraphTopology version (e.g. 0.2): 0.2.3
• installation method (e.g. CRAN, source): source

Great work!
Will there be a python version?

Thanks!

Problem description

I tested the data Lymph (number of nodes = 587) using learn_laplacian_gle_mm and got the error:
Error: cannot allocate vector of size ...

Example

library(spectralGraphTopology)
learn_laplacian_gle_mm(S, A_mask = NULL, alpha = 0, maxiter = 20000, reltol = 1e-06, record_objective = TRUE, verbose = TRUE)

Environment:

• platform Windows workstation

This is a note to myself to look into proper ways on how to manage memory allocation in the C++ code. This might impose problems when we work on simulations with large (>400) number of nodes.

The vector w is very sparse specially when the number of components of the graph is large. I should look into exploring this fact and use more efficient data structures probably from the Matrix package.