The (abridged) Baxandall conclusions...

Even fets* used without feedback generate high order harmonics - and therefore on programme high order intermodulation products.

A small amount of negative feedback (eg. 6dB) in a single-ended stage, though reducing the second harmonic distortion, and also the total (unweighted) distortion, by about 6dB, will increase the  higher order distortion and the quality of reproduction may well become worse as judged subjectively.

If enough negative feedback is applied all significant harmonics ( and corresponding intermodulation products)  can be reduced to a far lower level than without feedback, though the amount of feedback required to achieve this becomes larger the higher the order of the harmonic considered.

The magnitude of harmonics of extremely high order will be increased by the application of negative feedback, no matter what practical amount of feedback is employed, but this is of no consequence if, when thus increased, they are, say, 120dB below the fundamental.

(paraphrasing) As signal levels drop higher order harmonics drop faster than lower order.

Paul

*I think it's safe to assume that tubes are at least as bad as fets in the raw linearity stakes. The analysis in the article is based on 'square law' transfer characteristics which tend to generate 2nd harmonic.