Hello
if your data frequencies are not integer factor of your sampling frequency , you cannot use a rectangular (or alike) windows; it must be a window that has zero amplitude at both ends and max amplitude in the center, like a gaussian peak; a Hanning or Hamming window is the typical choice then
for more info - about what is the "best" window it all depends of what your aim is :
There is absolute no way to chose a "best" window function without a rigorous statement of your criterion-of-goodness. This problem was recognized in the 1970s as an issue in the processing of scientific data where the observer could introduce bias into the research results merely by selecting the appropriate windowing function. The unbiased choice is the rectangular window but it produces high sidelobes in the spectrum.
The answer to the problem then, as now, is the Maximum Entropy Method (MEM) by John Parker Burg. If you must use the FFT and a window function, you have many to choose from: (http://en.wikipedia.org/wiki/Window_function) Since they all introduce observer bias, you just pick the one that gives you an output you like. The truth of the matter is, you can get any spectrual result you desire by creating the appropriate window function.
In conclusion, there is no rigorous way except in the sense that you can chose to accept the criterion-of-goodness associated with a particular window.
