Although the speed is limited there is no real limit on how fast it can change direction. Suppose you make something 2.4mm wide with 0.5mm filament. E.g. a Mendel spring: -
Each wall will be 0.6mm wide leaving a 1.2mm gap in the middle. That gets filled with a zigzag infill where the head moves to within 0.3mm of each wall, so the head moves about 0.6mm on each stroke. At 36mm/s that makes 30 complete oscillations every second. 30Hz is a pretty high frequency for a mechanical system!
What actually happens is my y-axis starts to resonate. Over a few cycles the amplitude of the oscillation builds up and the infill overshoots the outline leaving a serrated edge.
The torque of a stepper motor is zero at rest and increases as it is displaced, so in that respect it behaves like a spring. That springiness together with the inertia of the rotor gives a resonance at hundreds of Hertz, known as mid band resonance. When the load is rigidly coupled, as in this case, the mass of the load brings the resonant frequency down.
As I don't get any missed steps I think the springiness might actually be in the belt rather than the motor. Timing belts have metal cables in them so that they don't stretch, but that makes them stiff, so they don't like to bend round a tight radius. That means the belt has some springiness being pulled round the pulley. A bigger pulley would be better but that would reduce the effective stiffness of the motor, so might actually make things worse. A lighter bed would be good but I haven't found a way to ensure it is flat without going to 6mm tooling plate.
I fixed the problem in software by slowing down the infill that has a high frequency content. I examine each infill path, one axis at a time, and convert it into a list of lengths between changes in direction. I then find the shortest wavelength over three cycles (less than three cycles is not long enough for the resonance to build up). I do this for X and Y directions and save the shortest of the two wavelengths. When I extrude the path I work out the frequency from the pre-calculated wavelength and the desired speed. I then compare that with a limit for each machine and reduce the speed if the frequency limit would be exceeded. I could have a separate frequency limit for each axis but I don't like the idea that the orientation of an object affects how it builds, so I pick the worst axis when deciding the limit.
I set the frequency limit to 20 Hz on my Mendel and 16 Hz on HydraRaptor. HydraRaptor does not show the overshoot problem, but it makes horrible growling noises and shakes the house. The machines make more interesting noises now because each infill run that hits the limit is extruded at an arbitrary lower speed. The overshoot is completely cured.
The builds are a bit slower and in some cases a long infill path will be slowed down by a short section that is high frequency, often a section between a hole and the outline. A more complicated solution would be to isolate the high frequency section and extrude the rest of the path at full speed.