I guess my explanation wasn't very clear... second try:
Without all the banking and rails, the track is a single curve. To display it in AVS, it's rendered as a series of line segments between dots. That's every-day stuff for any Superscoper.
Now, because the camera always follows the rail, I 'catch' the next camera position when calculating the current points. When the camera position 'u' is close to the point-position 'i', I copy the point's coordinates into the new camera's coordinates. This is done not only for speed reasons, but because the curve-code is too long for 'onframe' sections.
This has the side effect that the camera is always centered on one of those superscope-points and that it moves at exactly one-superscope-point-per-second (1/n).
Now, the camera's angles are calculated off the current position, and the last two positions. However, if the camera is moving at less than one-superscope-point-per-second, then you cannot guarantee that the current point and the last 2 positions will be 3 different points. And if they aren't, the banking and rotation formula's will divide by zero or calculate angles between 2 identical points. That doesn't work out and this will make the camera do very whacky things.
But, you say, why not just interpolate between the two nearest points! Then the camera could move at any speed and it would still look good. That works nicely for the camera movement, but not for the camera angles and rotation. If I use linear interpolation, the smoothness of the curve is lost and it actually becomes a series of line-segments, even internally. This would mean that the camera would rotate in sharp jumps. Just imagine a camera moving along a 12-dot circle and looking along that circle, and compare it to e.g. moving along a 12-sided polygon with a higher dot-resolution. At every corner, the camera's rotation would jump in the direction of the center.
I really think you underestimate the complexity of the rollercoaster problem: given a curve f(x,y,z) through space, define a track that moves along that curve mainly in a horizontal plane and rolls/banks in curves in a realistic and smooth matter. It's harder than you might think: it requires first and second order derivatives and some tricky trigonometry and vector math. Work it out on paper if you don't believe me
And the fact that it's AVS doesn't help much either...