Re: Statistical approximations, calculator style (2)

sales_at_kt-algorithms.com
Date: 03/21/05 

Date: 21 Mar 2005 15:47:39 -0800

Correction:

A thorough analysis of the critical t approximation above has indicated
that the error bounds have to be widened slightly:

Max. |rel. error|<0.0071 for df>=4 and 0.10>=A>=0.001

with the best set of rounded-to-4-decimals coefficients being:

t(A,df) = [(0.8497 (df-3) (-Ln(A))^-0.3179 + 2.6457
A^0.2094)/df]^-1.6829 - 0.4188

The problem was that the significance dimension was sampled too
sparsely during optimisation. The updated error bounds, on the other
hand, are stable when the number of sampled logarithmically equidistant
significance values are increased from 1,000 to 10,000 and 100,000
logarithmically equidistant positions!

I hope that this hasn't caused any inconvenience.

Regards,

Knud Thomsen