Thermal noise entropy is probably good enough for most practical uses, but it's still fundamentally producing a seed value that can be captured, misused, or bruteforced, right? Also curious if there are monte-carlo models looking at this to see "how uniformly random" they look.
Edit: I think figure 3 in this study is what I'm looking for. They define the inconsistency I described as "spectral pivoting".
> This discrepancy is because the Mermin-Wagner-Hohenberg theorem holds in the thermodynamic limit, while these simulations are for finite lattices
I think thermodynamic limit here means, it needs to be way too hot?
In practice it would be very difficult to predict RDRAND outputs. Even so I believe the truly paranoid can use RDSEED to skip the PRNG step. Not qualified at all to talk about how they de-bias the measurements.
Edit: I think figure 3 in this study is what I'm looking for. They define the inconsistency I described as "spectral pivoting".
> This discrepancy is because the Mermin-Wagner-Hohenberg theorem holds in the thermodynamic limit, while these simulations are for finite lattices
I think thermodynamic limit here means, it needs to be way too hot?
https://arxiv.org/html/2403.09078v1