For me it seems that NaN values come into the situation in 'calcAvec' function around the line 267 in 'lars_regression_noise.m'. In my case Mm can be a singular matrix, so avec = Mm\b will be a vector with NaN values. Once 'avec' has NaN values, W, Ws,w_dir and Aw will also have, so roots function throws error.
@Kata5 the other thing I can possibly think of is that your dataset has a lot of negative values due to offset adjustment. If that's the case, can you subtract the min scalar value to make it nonnegative and try again?