I suspect that you could encode your problem into a polynomial multiplication. It might be worth examining the cost of doing this computation using exact rationals rather than bfloats. It is also possible that you could do this by a bfloat version of an FFT. RJF