This is an old revision of the document!

A PCRE internal error occured. This might be caused by a faulty plugin

I use quaternions if I need to span the whole space (i.e. my initial estimate could be far off the answer). If finding a solution using optimization, e.g. R^*=argmin_(R∈SO(3) )⁡〖f(R)〗 Then given the function quat2mat(q)=(■(q_00^2-…&&@&&@&&)) I optimize the function g(q)=f(quat2mat(q/(\|q\|))) The division by norm doesn't actually make the derivatives much worse, and is much better than optimization subject to \|q\|=1. I tend not to worry too much about the gauge freedom g(q) = g(λq), but one should renormalize q after each descent step/linesearch. If I believe I have a decent initial estimate R_init (sounds like your problem), then I tend to use the exponential map R=R_init exp⁡([a]_× ) where exp is the matrix exponential, [v]_×is the 3x3 matrix which effects cross-product by v, and a is, oddly, the rotation axis multiplied by the angle of rotation. http://en.wikipedia.org/wiki/Rodrigues%27_rotation_formula It has pleasant derivatives, its singularities live far from the origin, and it’s 3 parameters, so you don’t need to get inside the optimizer to renormalize.