Full disclosure: I didn't spend the time to reverse engineer the circuit diagram for my Cosmos ADC. So, these comments are based on more general ideas. (Is LKA's schematic shown in post 1556 accurate??)
From PMA's measurements and observations, I wonder if the issue isn't so much the cap distortion, per se, but the common mode distortion characteristics of the Cosmos input. I ask this because Pavel has shown that a sort of balanced input signal from a power amplifier made by using an external voltage divider that is balanced gives better results. He's also measured the CMRR for the Cosmos ADC input to be ~60 dB, which is what I measured on mine. That sounds like a lot, but might not be. (Much higher than most audio devices, though!)
Lots of authors have discussed the issue of common mode distortion in opamp like circuits. Here's a couple examples:
Douglas Self
Walt Jung
ADI Crew
The general idea is that in order to minimize distortions in an opamp circuit, you need symmetry and balance. That way the symmetry and balance of the opamp itself allows for cancellation of common mode signals of any kind. This means that you need for the opamp inputs to "see" identical impedances at both inputs. Otherwise you get mode conversion in the circuit as described by Whitlock and Floru:
Whitlock and Floru
This is why you need really tight precision resistors in an instrumentation amplifier system.
Nash
OK, enough of the links, most of which you guys already know about.
This really becomes an issue when you have blocking capacitors at the input of a differential amplifier where you hope to maintain really high CMRR. At high frequencies where the caps are pretty close to being shorts, you're usually limited by the resistors in the opamp "bridge". But, as you go lower in frequency, the caps add a significant impedance in series with the bridge. If the caps are identical in capacitance, that's not a problem. But, how often do you get large caps like electrolytics where they match to like .01%?
I think this might be what is being demonstrated here.
One solution is to DC couple. Duh. No caps means no imbalance as a function of frequency.
Another is to match the caps. Hard.
A third is to make the caps really large in value so that their impedance in the band of interest is as close to zero as possible. Matthias Carstens modified his AutoRanger with larger caps to get better low frequency distortion. Note that these are film caps in the AR, which should have close to zero distortion to begin with.
Carstens' AutoRanger
Yet another solution is to increase the impedance of the input circuit so that the difference in impedance caused by cap mismatch is relatively smaller compared to the "bridge." Of course, this adds noise that you observe when there's no source device at the input shorting out the divider resistor impedance with regard to noise. Plus, you can bring on bias issues for the opamp and so on.
Or, you can use a different circuit at the input that is less sensitive to impedance differences between the non-inverting and inverting inputs to the ADC. That's what PMA's divider does to a degree.
So, I'm NOT suggesting that electrolytic capacitor distortion isn't a real thing. There, larger values means that the voltage drop across the caps is lower, which reduces the distortion as Cyril Bateman and others here (
Capacitor Distortion) have noted. Basic capacitor distortion could well be the problem here. But, PMA's use of a balanced divider to get better measurements kind of belies that.
Now here's my editorial comment: This is another example of why evaluations of various circuits need to be done in the proper context. Using an Audio Precision test system can mask some deficiencies. Not because the AP is bad in any way - it's because it's great in almost every way. (Except price - what a strange correlation...) The AP outputs are really balanced and are designed to have really high common mode rejection. Same for the inputs. The AC mains connection is highly isolated for signal currents. That means the AP has minimum impact on the device under test. Minimal Heisenberg effects, except as you add them in yourself. But, what other audio gear is so perfect in that regard? So, it's quite likely that various pieces of gear will behave very differently in actual use than they will when tested in an ideal test system. This is especially so when you add in various digitally based products like computers and DACs, which have common mode noise currents galore that propagate between boxes with the AC main systems as the very imperfect current return. Basic stuff that is usually dismissed. OK, end of my editorializing.