Hi, This field is new for me and I am not longer programming in matlab, but I have compared the code in adassp.m to Algorithm 2 in the paper and it seems to be some differences.
-
The identity matrix (eye(d)) is added to XTX, a this seems to be different from Algorithm 2.
-
When calculation lamb_min, logsod/(epsilon/3) is not multiplied with BX^2 as in Algorithm 2. ( Although BX = 1 in the function. Data scaled beforehand… )
-
Matrix Z is calculated by Z=randn(d,d); Z=0.5*(Z+Z');. Thus diagonal elements in the resulting symmetric matrix is sampled form N(0,1), but the variance of the off-diagonal elements become 0.5 (= 0.5^2 + 0.5^2) instead of 1. In Algorithm 2 every element have variance 1.
Hi, This field is new for me and I am not longer programming in matlab, but I have compared the code in adassp.m to Algorithm 2 in the paper and it seems to be some differences.
The identity matrix (eye(d)) is added to XTX, a this seems to be different from Algorithm 2.
When calculation lamb_min, logsod/(epsilon/3) is not multiplied with BX^2 as in Algorithm 2. ( Although BX = 1 in the function. Data scaled beforehand… )
Matrix Z is calculated by Z=randn(d,d); Z=0.5*(Z+Z');. Thus diagonal elements in the resulting symmetric matrix is sampled form N(0,1), but the variance of the off-diagonal elements become 0.5 (= 0.5^2 + 0.5^2) instead of 1. In Algorithm 2 every element have variance 1.