Table 2.

The Frobenius loss and the entropy loss estimated by the probit graphical model, the oracle method and the naive method. The oracle method and the naive method simply apply the graphical lasso algorithm to the latent continuous data Z and the observed discrete data X, respectively. The results are averaged over 50 repetitions and the corresponding standard deviations are recorded in the parentheses.

Example	n	Frobenius Loss			Entropy Loss
		Gaussian	Oracle	Probit	Gaussian	Oracle	Probit
Scale-free	50	2.3 (0.12)	0.7 (0.05)	2.2 (0.13)	12.0 (0.73)	3.1 (0.29)	23.1 (1.83)
	100	2.2 (0.13)	0.4 (0.08)	1.7 (0.09)	9.4 (0.68)	1.9 (0.29)	10.1 (0.45)
	200	1.7 (0.12)	0.3 (0.02)	1.2 (0.04)	6.4 (0.33)	1.1 (0.10)	5.4 (0.26)
	500	0.9 (0.05)	0.1 (0.01)	0.7 (0.04)	3.3 (0.19)	0.5 (0.05)	2.7 (0.19)
Hub	50	1.2 (0.06)	0.3 (0.02)	1.1 (0.04)	21.2 (1.32)	5.8 (0.70)	29.4 (1.76)
	100	1.1 (0.10)	0.1 (0.01)	0.8 (0.03)	15.9 (1.03)	3.2 (0.27)	15.1 (0.64)
	200	0.8 (0.05)	0.1 (0.01)	0.6 (0.01)	11.9 (0.39)	1.8 (0.23)	10.4 (0.33)
	500	0.6 (0.02)	0.0 (0.00)	0.5 (0.01)	9.1 (0.16)	0.7 (0.06)	7.5 (0.16)
Nearest-neighbor	50	1.4 (0.04)	0.6 (0.02)	1.3 (0.06)	16.5 (0.80)	5.6 (0.30)	25.6 (2.04)
	100	1.3 (0.08)	0.4 (0.02)	1.0 (0.02)	12.1 (0.52)	3.5 (0.36)	12.4 (0.76)
	200	1.0 (0.04)	0.2 (0.01)	0.7 (0.03)	8.6 (0.32)	2.0 (0.11)	7.5 (0.17)
	500	0.6 (0.03)	0.1 (0.01)	0.5 (0.02)	5.5 (0.12)	0.8 (0.02)	4.5 (0.19)
Random-block	50	1.8 (0.05)	0.7 (0.05)	1.7 (0.04)	14.8 (1.04)	4.7 (0.46)	23.5 (1.76)
	100	1.6 (0.16)	0.4 (0.02)	1.3 (0.03)	10.7 (1.10)	2.9 (0.27)	11.3 (0.46)
	200	1.3 (0.05)	0.2 (0.03)	0.9 (0.05)	7.2 (0.19)	1.6 (0.11)	6.3 (0.32)
	500	0.7 (0.03)	0.1 (0.01)	0.6 (0.03)	4.1 (0.15)	0.7 (0.06)	3.5 (0.13)