Table 1.

Summary of parameter values used in our application of Algorithm 2.

Parameter	Value	Definition
r	Varies	Maximum rank of J
π1, … , πr	Varies	Constraint signs
ϵmax	0.5	Maximum value of the regularization parameters
ngrid	501	Number of different values the regularization parameter can assume forming a uniform grid from 0 to ϵmax
Ttrain	70% of samples	Indices of data samples that form the training set
TCV	20% of samples	Indices of data samples that form the cross-validation set
Mmax	20	Maximum number of iterations of the block coordinate descent algorithm
δp	0 (machine precision)	Convergence precision
σmax	3	Number of allowed failures to improve cross-validation performance

Since the convergence precision, δ_p, is set to zero, the algorithm converges when the negative log-likelihood evaluated on the cross-validation set does not decrease above machine precision. The constraint signs are assigned to equal numbers of positive and negative components.