You may derive this from standard curves. Here is one example:

https://openi.nlm.nih.gov/detailedresult.php?img=1524913_gkl400f1&req=4https://openi.nlm.nih.gov/detailedresult.php?img=1524913_gkl400f1&req=4

You can see that the intersept of the regression is at Ct=35 (that's assay-specific, but this range is reasonable). This means: if one would amplify a single molecule (copy), the expected Ct would be 35. Values above 35 would already indicate that you measured less than a single molecule, what is obviousely nonsense. At a Ct around 30 the expected number of molecules is around 2^(35-30) = 2^5 = 32. By Poisson's law the variance of this estimate equals the estimate, so the 95% prediction interval for the true copy numbers in samples with 35 expected copies is (28, 48), so this is the best available precision you can get for such samples (ignoring variability introduced by the measurement; this is a purely statistical feature!). If you go down to 10 copies, the range is (4, 17), and for a single copy it is (0, 3). So it is difficult to get a sufficient precision for estimates of low copy numbers already by Poisson's law. This becomes an even more severe problem when the amplification is not 100% specific. If side-products are co-amplified you get a competition on primers and enzyme what can have a considerable effect on the ct value.

You may derive this from standard curves. Here is one example:

https://openi.nlm.nih.gov/detailedresult.php?img=1524913_gkl400f1&req=4https://openi.nlm.nih.gov/detailedresult.php?img=1524913_gkl400f1&req=4

You can see that the intersept of the regression is at Ct=35 (that's assay-specific, but this range is reasonable). This means: if one would amplify a single molecule (copy), the expected Ct would be 35. Values above 35 would already indicate that you measured less than a single molecule, what is obviousely nonsense. At a Ct around 30 the expected number of molecules is around 2^(35-30) = 2^5 = 32. By Poisson's law the variance of this estimate equals the estimate, so the 95% prediction interval for the true copy numbers in samples with 35 expected copies is (28, 48), so this is the best available precision you can get for such samples (ignoring variability introduced by the measurement; this is a purely statistical feature!). If you go down to 10 copies, the range is (4, 17), and for a single copy it is (0, 3). So it is difficult to get a sufficient precision for estimates of low copy numbers already by Poisson's law. This becomes an even more severe problem when the amplification is not 100% specific. If side-products are co-amplified you get a competition on primers and enzyme what can have a considerable effect on the ct value.

there are some polymerases more reliable for weak samples than others, some newer and improved polymerases can also provide better results than older ones. In my opinion you cannot categorize in bulk all the results over 30 circles as unreliable.

As previous authors have noted, quantitation past >30 cycles has been unreliable also in my experiences.

In my experience, I have achieved values of ~38 Cq that end up being emblematic of single-copy reactions.  Also, recall the paper that had proven by sequencing, that (in a ~70% efficient PCR reaction) a specific single copy amplification is detectable out to 48 cycles (see attached).  Also, in my work on a certain pig virus (with a private company), we have precision down to ~3.5 copies/rxn with Cq values of ~36-37... so, depending on how many technical replicates of the suspected low target-copy reactions you decide to run also determines the extent of your ability to interpret the nether region of the assay. All standard curves for this latter work included 8 technical replicates (minimum) at every dilution along the way.