Thanks (Dennis?).
So in absolute terms, these measurements are meaningless. But is there anything to them in relative terms? I.e. do they suggest that #9 is closer than #8 to #1, or is that stretching things a bit too far?
Mani.
Yes it is Dennis.
No not meaningless. 8 and 9 share about as close to identical FR as possible or the null couldn't have reached below 90 db. They apparently have as close to perfect similarity as possible. Once in the analog world there is thermal noise. These files likely are so similar any differences are below thermal noise. Meaning if you listen to it one time and listen a minute later the thermal noise is different, but the file playback differences are below that if there are any. Nothing different to hear.
The issue with comparing to the original file is timing and FR. The null test is so sensitive extremely minor FR differences reduce the null. We already know your DAC/ADC loop droops a bit below 10 hz. Even that can alter the results. It is quite possible they also droop some tiny fraction at the upper end near 20 khz.
Next is timing. Ray's post above is showing that. The clocks of both recordings run very slightly fast. However, your sampling is delayed by about 3 nanoseconds for every meter of cable bewteen DAC and ADC. So comparing the original to a recording it probably starts out behind, eventually catches up and is in phase with the original. The difference will be a minimum at that point. Then slowly the recording gets ahead of the original file and the difference grows larger. With a long enough file you can see that repeat over a regular sequence. Another tell tale to that condition is comparing FFT of the difference file. If it has a 6 db per octave upward tilt vs the original it probably is a mis-timing issue. Higher frequencies are more effected than lower frequencies.
Now Diffmaker can, and sometimes is successful at fixing the timing between files not just at one point, but throughout the file. It also will match gain very precisely, but here is where an FR difference can corrupt that process. So when it works it fixes shifts in time, drifts in clocking and level differences to get the deepest possible null. However this is a complex process and sometimes it gets tripped up.
So the original vs 8 or 9 being in the mid 30 db range probably isn't fully correct. Ray's post above shows at some point it gets to a fairly deep null. It probably is much more than 30 db, but whether it is 90 db nulls or not we couldn't say. Also even a .1 db FR issue raises the null, but assessing that vs audibility isn't clearcut. It probably isn't audible even though it reduces the real null depth. If a device is causing distortion and that was enough for a 30 db null the result is rather audible. So really deep null numbers tell you not much of anything is different. You get to moderate or low null numbers and what can be deduced depends on what is causing the differences.