Closed Oblynx closed 3 years ago
Let r_err = 1 - k() + s*a
(s: sparsity, a: area), then r_err < 1
wins per area is the probability to maximize. That means tuning t
to make the mean of the probability distribution P(X=x) -> r_err.
Since for each tied minicolumn (how many?) whether it's selected is an independent experiment, it's a binomial distribution {n,p}.
In general n!=a
since most minicolumns are either far above or far below the tie and don't contend. t=1-p
A very rough estimate of n
could be k/2
.
n*p = r_err => p= r_err/n => t= 1-r_err/n
Estimate the tied minicolumns: tieEst(o,Z)= count(o .== Z) * area()/prod(szₛₚ)
27 assumed that 1 of the tied minicolumns should win and maximized that probability. Actually, < 1 should win on average:
1 - k() + s*area
to be precise.With this in mind redo the calculations and account for the rounding error.