Currently, the pseudodata are generated again each time even one of them is negative (and not ASY) before going on. From a theoretical point of view this is strange, because it is equivalent to cut the gaussians distribution before zero. Moreover, from a performance point of view, this easily becomes a bottleneck (for example in the case of fit with thcovmat in which the shifts are bigger).
We want to try doing something slightly different, namely taking the absolute value of each negative pseudodata. From a theoretical point of view this makes no sense as the actual method, but, at least, it requires only a single iteration in every case.
