Seems like a mismatch between implementation code & mathematical formulation in the paper

[Buhua-Liu]

My understanding:

• RFF: random_fourier_features_gpu() adds a new dimension of size num_f rather than performs dimension transformation.
• lossb_expect(cfeaturec, num_f) computes the square of the Frobenius norm of the original feature vector (of size batch_size * feature_dimemsion ) at each num_f dimension, then accumulating them minus the trace of covariance matrix gives the final loss.

Question:

• Does cfeaturec refer to A_i and B_i, and n the batch size in Eq. (3)?
• num_f refer to n_A and n_B in Eq. (4)?
  

[xxgege]

1. Actually, A and B refer to two variables and and refer to features of the representation, as described in Section 3.1 of the paper. Thus cfeaturec does not refer to A_i and B_i, but (A_1, ..., A_n) and (B_1, ... ,B_n) refer to two variables in cfeature. n refers to the batch size.
2. Actually, n_A and n_B are twice of num_f, because when we randomly sample one set of w and b for RFF, both sine and cosine of the feature are calculated.