Under/over sampling is dependent on your seeing conditions. What I have read puts this under/over point somewhere in the 3-4x range. There are other factors that would further limit your effective resolution limit, as well. I think an easy starting point is seeing/3. For me, my typical seeing is 1.5", so going as fine as 0.5"/px will be beneficial for resolution. I have proven that 0.3”/px (moderately oversampled) shows more detail than 0.9”/px (moderately undersampled) on M13 for my system and local conditions. For those with 3” typical seeing, the cut-off would be more like 1”/px.
But there is a SNR trade-off (all things being equal) to finer image scales. It is generally inadvisable to push it to oversampling, as you will get a lower SNR for no resolution advantage. Also, sampling does not come into consideration when imaging large targets, as FoV considerations dominate getting the entire target on the sensor. Sampling is a concern when imaging a small target that covers a small portion of sensor, as extra resolution is beneficial for cropping in tight.
Following is more info than you requesting, but I am including some more background info for other who may be reading this:
For imaging, the main concerns are Field of View (FoV) for composition, image scale for detail of small targets, and Signal to Noise Ratio (SNR). For large targets, FoV normally the dominates image scale as the critical factor, as these targets will generally require large FoV to fit on the sensor and this necessitates undersampling (e.g. using shorter FL leading to coarser image scales). For small targets, image scale becomes more important in order to resolve the detail on a highly magnified subject that composes a small part of the sensor.
FoV is determined by the combination of the scope's focal length (FL) and the sensor's sensor size. You can get larger FoV by a shorter FL and/or larger sensor size.
Image Scale (arcsec/px) is determined by the combination of the scope's FL and the sensor's pixel size. You can get finer image scale (smaller number, more resolution) by a longer FL and/or smaller pixel size.
There is no advantage to a finer image scale than determined by Nyquist Critical Sampling, as it results in no extra achievable resolution. Nyquist Critical Sampling is rate (frequency no higher than, or arcsec/px no smaller than) at which ALL of the embedded information (filtered signal) can be reconstructed.
Nyquist Critical Sampling is a fraction of the size of the finest detail (or multiple of the highest frequency component) of the filtered signal. I refer to this divisor of size or multiplier of frequency as the Nyquist Factor. Estimates of this Nyquist Factor are between 3 to 4 for 2-D images, and 2 for 1-D waveforms.
An estimate of a filtered signal's finest detail is the smallest measured FWHM of any star. The true signal is filtered (or blurred) by seeing conditions, resolving power of the telescope, and the tracking ability of the mount.
Critical Sampling ~ SQRT(Seeing^2 + Dawes^2+ ...) / Nyquist Factor. (I am not sure whether to add the mounts error in quadrature into the prior equation, or to treat the equation as the limit with a theoretically perfect mount and degrade the estimate, since seeing conditions would also show up in the measurement of the guiding accuracy (e.g. double counting). This could also be the reason for the Nyquist Factor range.)
Seeing is a measure of blurring caused turbulence in the Earth's atmosphere, measured in arcsec. This is what causes stars to “twinkle”, like the image of a coin on the bottom of a swimming pool jiggles with the ripples on the water surface.
Dawes Limit is a measure blurring due to wave diffraction caused by size of the telescope’s aperture edge, measured in arcsec. The larger the diameter of the scope the sharper the image, and the lower the Dawes Limit. Dawes Limit = 116 / Aperture Diameter, in mm.