I computed the occurrence of higher-order modes up to order m + n = 20 as a function of g factor for a ring cavity.

In the first set of plots of plots, I've fixed the cavity half-length L and chosen several values of modulation frequency f_PDH. In the second set of plots, I've fixed f_PDH and chosen several values of L. Green is the carrier, red is the lower sideband, and blue is the upper sideband. The takeaway messages from these plots are that

1. there are two or three "best" regions to place g: near 0.06, near 0.46, or near 0.54 (although 0.06 is sort of close to instability);
2. the locations of these regions are independent of the modulation frequency, at least for the frequency range we are interested in; and
3. a lower modulation frequency widens these best regions.

So I think we should go for as low a crystal frequency as possible that is consistent with having shot-noise limited intensity and a high loop speed. I know the number 20 MHz has been thrown around as the lowest reasonable PDH frequency, but I don't understand quantitatively why this is.