forum.alglib.net

It is almost surely not sqrt(eps), but some other root (may be, cubic root or higher). Numerical Recipes describes procedure for determining correct power, but I can't reproduce these calculations now - may be tomorrow.... In any case, I think that 0.0001*angle1 will work.

Any news on when the other language versions will be released in Ver 3 format (VBA specifically)?

_________________
Doug Jenkins
http://newtonexcelbach.wordpress.com/

Python is planned to be released in the November, FreePascal/Delphi may follow in the December. As for VBA, it will wait for 2011, January or February I think. VBA is a "hard" language, because it does not support function pointers, nor it supports console output. So it will need special treatment.

I have another question regarding the estimation of the gradient. When the values of angle1 + h, angle +2h etc exceed the bound limit should I replace the value with the min - max bound?

For example the bound limit might be [-60, 60] and the angle + 2h is 60.1 or higher. Should I replace the value with 60?

That depends on problem you want to solve.

If your function is defined for all values of its arguments (i.e. there is no such angle which leads to overflow, division by zero, etc.), I recommend you to leave angle+2h unchanged. You will get centered formula with quadratic accuracy.