To bring the theory into practice, you are saying that at some point in stacking; read noise becomes the dominant source of noise. Are you suggesting the SNR stops improving at this point? Not just that it follows the curve asymptotically(diminishing returns)?
Technically it will be asymptotic. When one term dominates another term doesn't mean that the dominated term suddenly vanishes at a certain point, but it will continue to get smaller and smaller (basically that is the definition of an asymptote).
So at what point does read noise become an issue?
It depends, I'm not trying to calculate it here, I'm just trying to show that there is such a point, as many ppl have implied that there is not such a point. Surely this point will vary and will depend on quite a number of things: sensor performance, temperature, light pollution, etc. I imagine it is quite large (many hundreds if not thousands of subs) for a decent setup and dark skies. It's also very likely that most ppl will stop well before ever reaching this point, esp. with bright objects, b/c they would have deemed their result sufficiently noise free (Even the extreme guys who take days worth of total integration time).
Most counter arguments have been something like "well this guy's image is 40 hours". But all of those amazing images, while really long, typically have rather reasonable number of subs usually < 100 (per filter of course, when collecting more subs with a different filter, you are collecting a different signal, you don't stack images with different filters anyway, you merge them in different ways, like by assigning them to different color channels). They just have incredibly long sub-exposures (on the order of 30 minutes).
I'm not saying there is a point of diminishing returns on total integration time, I'm saying there's a point of diminishing returns on increasing the number of subs to get more total integration time. At some point the only way to get more useful
total integration time is by increasing subexposure length, not by taking more subs.
Also, if we use dark frames, we are removing the read noise. Does this further push back the point where the read noise becomes significant?
Yes it does, but it doesn't eliminate it completely, because it is impossible to eliminate read noise completely (we can do a very good job, esp. with modern sensor with really low read noise, but perfection is impossible). So there will still be a point somewhere where it does happen, it may be many hundreds of subs later, but it will surely happen.
Also, if we have the read noise within the SNR formula, then why wouldn't a sub be the exact same thing as an interval of time?
If a sub were exactly the same as an interval of time, we could divide a given total integration time into as many subs as we wanted and never loose any quality -> I could take many thousands of 1/10 sec subs and achieve the same result as an equivalent single exposure. Clearly I can't do that. That is the proof that more subs is not exactly
the same as more integration time. It's close, but it's not exactly, and the difference is important, the difference is what puts an upper bound on the number of subs.