Problems of MinibatchCKA

Hello, I find the value of unbiased HSIC is always negative, which will result in NAN when I calculate the MinibatchCKA. Will you meet this situation? Here is my code, I have reshaped the features X, Y to B * (C * H * W):

def unbiased_HSIC(X, Y):
"""Unbiased estimator of Hilbert-Schmidt Independence Criterion
Song, Le, et al. "Feature selection via dependence maximization." 2012.
"""
kernel_XX = torch.mm(X, X.T)
kernel_YY = torch.mm(Y, Y.T)

tK = kernel_XX - torch.diag(kernel_XX)
tL = kernel_YY - torch.diag(kernel_YY)

N = kernel_XX.shape[0]

hsic = (
torch.trace(tK @ tL)
+ (torch.sum(tK) * torch.sum(tL) / (N - 1) / (N - 2))
- (2 * torch.sum(tK, 0).dot(torch.sum(tL, 0)) / (N - 2))
)
return hsic / (N * (N - 3))