Thanks. I guess it’s pretty obvious for the totally anti-correlated case. The error goes down as 1/N rather than 1/sqrt(N).
Good point about the bias in NUTS—it’s a subtle end tweak that makes a big difference in efficiency.
Thanks. I guess it’s pretty obvious for the totally anti-correlated case. The error goes down as 1/N rather than 1/sqrt(N).
Good point about the bias in NUTS—it’s a subtle end tweak that makes a big difference in efficiency.