| """Finite reference distributions for verification, never model selection.""" |
| import torch |
|
|
|
|
| def inverse_empirical_cdf(values,probabilities): |
| """Quantiles of a finite uniform law; no interpolation between support points.""" |
| assert values.ndim==1 and probabilities.ndim==1 and len(values)>0 |
| assert values.device.type=='cuda' and probabilities.device==values.device |
| assert torch.all((probabilities>=0)&(probabilities<=1)) |
| ordered=values.sort().values |
| indices=torch.ceil(probabilities*len(values)).long().clamp(1,len(values))-1 |
| return ordered[indices] |
|
|
|
|
| def sign_flip_pvalues(daily_differences): |
| """Enumerate the exact sign-symmetric null for independent paired days.""" |
| assert daily_differences.ndim==1 and daily_differences.device.type=='cuda' |
| n=len(daily_differences) |
| code=torch.arange(2**n,device=daily_differences.device) |
| bit=torch.arange(n,device=daily_differences.device) |
| signs=(((code[:,None]>>bit)&1)*2-1).to(daily_differences.dtype) |
| simulated=(signs*daily_differences).mean(1);observed=daily_differences.mean() |
| one=(simulated>=observed-1e-12).double().mean() |
| two=(simulated.abs()>=observed.abs()-1e-12).double().mean() |
| return one,two |
|
|
