Are you referring to the guard bands, ADDED to the front and end of a block of samples being processed, that ramp up and down from zero, and designed to maintain bandwidth limiting and pretend that the block is doubly infinite?
If so, technically there is no distortion at the front and end of the actual band, but processing it as such would treat it like a square wave (where not zero), which would generate frequency components well beyond the bandwidth to which the signal is required (according to the Nyquist et al theorem) to be constrained. Careful design of the guard band slope patterns would keep bandwidth, but would also add to the coefficient values for each of the conversion frequencies, which presumably would have to be factored out before using the coefficients in any calculations. The people that come up with this stuff are SMART!
I am not familiar with all the maths, so maybe you can stop being a smart-a** and let us all know THE ANSWER.