Hi Joel, I it is easy to map it onto d<N>(f,v), but I think it is difficult to be optimized for N=1, 2, 3 ... because to calculate the derivative of d<5>(f,v), actually, d<1>, d<2> ... d<5> will be calculated . but after each derivative is calculated, it will be simplified before next step. now the result of derivative expression can be saved as a functor, so d(d(x*x,x),x), y = d(x*x,x); d(y,x); is one of my purpose. I will implement d<N>(...), that is a good idea and convenience, thank you. Dongfei.

Can't this be mapped onto a d<N>(f,v) function that is optimized for N = 1,2,3,... ? I bet that if you use proto, matching d(d(d( ... )))) and repalcing by d<N>(...) should be easy.

Do you perform your derivation recursively on the derivative degree (ie d(d(x*x,x),x) is y = d(x*x,x); d(y,x);) If yes, you may want to use coarser grain. Let's say you specialized d() d<2>() and d<3>() for ex. Computing d<5> is d<2>(d<3>()) instead of d(d(d(d(d()))). Of course, using d,d2, d4 is maybe a better option so you cna use bit representaiton of derivative degree to find how to combine the derivative.