One-sided accurate numeric derivatives

[maxchen]

Should it be "One-sided fast numeric derivatives"?

Code: Select all

!N = 1000 wfcreate(page=gm) u 1 !N rndseed(type=mt) 12357 coef b = 1 coef r = 2 genr y = @rgamma(b(1),r(1)) 'random of Gamma dist logl L_g L_g.append @logl L L_g.append L = @log( @dgamma(y,b(1),r(1)) ) L_g.append @derivstep(1) @all 1.49e-8 1e-10 L_g.append @param b(1) 0.05 r(1) 2 freeze(tb0) L_g.ml(showopts) freeze(tb1) L_g.checkderivs

Then check the top of tb1, "One-sided accurate numeric derivatives". I think one-sided for speed, and two-sided for accuracy.

[trubador]

The default is "numeric derivatives" in LogL object. You can control the "Accuracy" and "Speed" for both "Analytical" and "Numerical" derivatives. Therefore, although "One-sided accurate numeric derivatives" are faster than that of "Two-sided accurate numeric derivatives", it is not "One-sided fast numeric derivatives", which requires explicit specification of "Speed" as the method of taking derivatives.

[EViews Glenn]

To elaborate on Trubador's response. The fast numeric derivatives are simple two point (two-sided) evaluations. The accurate numeric derivatives are RIchardson derivatives which evaluate at four points. Note that EViews does not compute one-sided numeric derivatives.

One last comment. The fast/accurate distinction when using analytic derivatives is only relevant for those terms for which we cannot determine the analytic expression for the derivative.

[maxchen]

hmm, but [f(x+h)-f(x)]/h is called one-sided, and [f(x+h)-f(x-h)]/(2h) is called two-sided

[EViews Glenn]

That is correct and does not contradict anything that I wrote. EViews only does two-sided or Richardson numeric derivatives. The two-sided are the fast and the Richardson are the accurate.