Gilles Sadowski schrieb: >>> L1 norm: equals(a, b, tolerance) = sum(abs(a-b)) < tolerance >>> L2 norm: equals(a, b, tolerance) = sqrt(sum((a-b)^2)) < tolerance >>> L-infinity norm: equals(a, b, tolerance) = max(abs(a-b)) < tolerance >>> >>> All are useful. >> I'll guess for 3D vector the L2-norm is the more useful as it is >> invariant when physical orthonormal frames are changed. >> I'll check in the variant you prefer. > > [Cf. my other reply.] > Should I still create an issue?
I'm currently not invovled in commons-math, but I've written substatntial amount of numeric software. My thought on this point is, that it is generally more feasible to implement a bunch on distance methods like a.distance1(b) ... L1-Norm a.distance(b) ... L1-Norm (this is usually what the user wants as default) a.distanceInf(b) ... L∞-Norm and have the user decide on what is equal in his particular situation. There is a very good article on ieee754 equality under http://www.cygnus-software.com/papers/comparingfloats/comparingfloats.htm which contains an excessive discussion about various "equality" approaches. IMHO it should be moreover considered to include the "AlomostEquals2sComplement" method of this paper into commons-math if not done so far. Best regards, Wolfgang --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
