A zero in the precision matrix Ω = Σ−1 is a missing edge of the Gaussian graph. The graphical lasso minimises −log det Ω + tr(SΩ) + λ‖Ω‖1, so a larger penalty λ removes more edges. Choose a data set, slide λ, and compare the graph with the truth when it is known.
The algorithm is block coordinate descent on the covariance: each variable is regressed on the others by a lasso whose Gram matrix is the current estimate of Σ, as in the lecture notes. The penalty includes the diagonal, as in the book; the covariance of the flow cytometry data is the one the book's penalties refer to. Edge width is the absolute partial correlation.
code/ch30_graph_learning.py (glasso, Figure 30.1).