Nothing new here. Higher harmonics are more readily detectable than lower, and that's exactly what was shown in the table
@q3cpma posted. When he says, 'If the designer pays attention solely to the THD value, he may not realize that sometimes the "good-sounding, lower harmonics" get suppressed at the expense of the "bad-sounding, higher harmonics" ' he's just waving his hands without looking at the numbers, and the same holds true for the Olson article he links. At 50Hz the 5th harmonic becomes detectable around -56dB, yet all the 5th harmonics in Olson's examples are at least 20dB below that (and the worst setup is the transformer-coupled design he goes on to praise). It would be nice to see some actual data to back up his claim that feedback
raises the level of higher harmonics, but I suspect this only actually applies to poor feedback implementations that break down at higher frequencies (which can certainly happen using valves).
The 'listening tests' at the end are just the usual unconvincing hokum, I'm afraid.
The underlying problem here, a source of disagreement, is what I have already highlighted in a previous post: both audio devices and our ears are non-linear (and non-permanent), i.e. they behave in "strange ways“. I will try to explain what I mean with "strange ways” comparing them with linear ones, simplifying as much as possible.
Let's say that we have a system that process one signal: one input, one output. If we introduce on it the input signal A, we will have on its output the signal A'; if we introduce the B signal, we will have a B' signal at the output. A,B and A',B' for an amplifier could be respectively two input signals becoming from a DAC and two output signals to the loudspeakers; for our ear, they are the signals becoming from the loudspeakers and the sensations at listening (a bit vague, but we'll see better later).
If the system is
linear (and permanent), with the introduction on it of the sum S=A+B, we will have the sum S'=A'+B' at the output. This regardless of the instant we introduce the signal. Even increasing or decreasing the levels of signals A or B at the input, we will have a variation of the same amount for A' or B' at the output. Moreover, if we connect in sequence more linear systems, the resulting composite system will be still linear. As final consequence, if we play with specific elementary signals, i.e. pure tones, we can understand practically everything about a linear system and predicting its behaviour for any complex signal. The linear distortions of the input signal that can be produced are simple forms of time nature or of level, for different frequencies; new tones are never created.
If the system is
non-linear (and non-permanent), the introduction of the sum S=A+B does not produce the sum A'+B' at all, but a signal S' that can be very different. It could happen for example that small variations in B cause large variations in S', even if B alone in the input causes small changes in B', or vice versa. Here, new pure tones are generated by the interaction of those presents in A and B. Furthermore, S' could depend on the past signals processed and finally, even an increase in the input level is not found in the same proportions at the output. Thus, for non-linear system the law modelling input/output relationship for any complex signal can be of high complexity, and it can't be decomposed by the combination of elementary signals to predict with certainty its behaviour.
Now, music is made up of very complex signals and both amplifiers and our ear are non-linear (I think nature has many good reasons for this). For amplifiers we can obtain with relative effort non-linear models, but usually they are modelled for simplicity as linear ones, quantifying the deviations from this model with specific parameters or graphs, like the classics THD, IMD and many others. In turn, linear distortions are generally less critical to manage, though audible in some situations. Intuitively, small distortions would seem desirable. For our ears this simplification can't be done: it is much more complex, involving acoustical, mechanical, hydrodynamic and neurological subsystems (in one word, the psychoacoustics) all acting in sequence and intrinsically non-linear. Moreover, the countless facets of "sound perception" are bit difficult to model with a high level of precision: after all, we are humans, not systems.
So, when the output of an amplifier (via loudspeakers or headphones, highly non-linear too) is proposed to our ears, "strange" and unexpected things can happen for our sound perception. For example, amplifiers with very small THD/IMD values could sound to our ears dull and artificial, while others with much higher THD/IMD values could be very pleasant and realistic. The reason should be clear based on what has just been described: our non-linear ear behaves in a complex way, hardly predictable. It is therefore equally clear what it is important in the evaluation of an amplifier (or any other audio device): not the absolute values of its THD/IMD, but they form. It is now established that certain types of distortion are much better than others and, as stated by GedLee work, even preferable respect to not have at all. Masking effect of our ear seems the main responsible, hiding the less pleasant distortions, however insignificant they results to classic measures. As far as I know, studies are still ongoing, looking for metrics that are able to approximate always better our sensations of listening.
In summary, the measurements of audio devices are valuable but by themselves, without the psychoacoustics aspects, are not sufficient. Due to the missing of appropriate and agreed metrics at the date, the final judge of the listening pleasure can be only our ear. I agree, the set up of listening tests, however controlled, is complex and full of insides. This requires competence, severity and the possibility to repeat tests for anyone. The results, involving people, will be always of statistical type. But there are not other ways to investigate on these aspects.