My friend TIm has been thinking a lot this year about the design of audio amplifiers. His first idea (later abandoned) was that MOSFETs are the most linear devices in their most linear region, which is typically the high current region surrounding 1/2 the maximum current. A Class A MOSFET amplifier biased to that point would have negligible distortion even without feedback. This optimal linear operation does not just reduce distortion to the 2nd order that some tube circuits do following the square-law operation of their devices, it would virtually eliminate distortion.
With regards to feedback, he follows the results of Peter Baxandall, who showed that either very little feedback and a tremendous amount is optimal--following an inverse goldilocks distribution. For a typical transistor amp, this would be less than 1dB or greater than 35 dB. Anywhere in the middle, from 1-35dB, and the feedback isn't actually reducing distortion, it's increasing distortion by trading away some distortion harmonics for others. Often meter readers (such as myself) may not see this because the added distortion is in very high order harmonics which are cut off by the bandwidth limitations of the amplifier. But the mere attenuation of this distortion doesn't eliminate its effect, which can be an added graininess. (Subjective speculation here is my own.) To really understand what's going on, you have to do the math, as Peter Baxandall did.
The problem with using the larger amount of useful feedback, 35dB or higher, is that you must first build an amplifier with 35dB extra gain. For a large power amplifier, this is not an easy task. Far easier to make the amplifier low distortion in other ways, such as using very linear devices, class a operation, and complementary operation (where complementariness is useful--this requires inversely matching PNP and NPN transistors, which even the best only approximate, and the worst are worse than useless).
So you can see this is pretty heavy stuff.
For awhile, Tim was fascinated by the Adcom amplifiers designed by Nelson Pass and others, particularly the 5400, 5500, 5800, and 5802 (the last not being a Pass design). They use MOSFET outputs, and are biased into the high Class AB+, the 5802 consuming more than 400W at idle, for example, more than many so-called Class A amplifiers.
But by the time he actually got around to figuring the correct bias points, Tim realized these amplifiers don't even come close to the optimal linear area of MOSFETS. (Actually, some of my questions led him to this disappointing analysis.) For the devices in these Adcom amplifiers, something like 3-8 amp quiescent bias would be required. And in contrast, despite the high dissipation in these amplifiers, it's spread over a large number of devices (for power handling reasons), resulting in less than 1amp per device. The high quiescent power results not from super high bias but from lots of devices with large rail voltages, and relatively high, but not optimally high bias.
Further thinking revealed that it's basically not possible to use MOSFETS effectively at their most linear region in practical power amplifiers. Either they would have to be especially low output super Class A amplifiers, or you would have to use liquid cooling of some sort with custom devices. MOSFETS seem to be made for voltage amplifying circuits, not power providing circuits, at least if you are mainly thinking about taking advantage of the most linear operating region.
When NOT in their most linear region, MOSFETS actually have little to offer that's better than bipolar transistors. The linear region in MOSFETS ends abruptly, and you don't really want an amplifier in those cutoff areas at all, as it may be worse than a bipolar cutoff region. But that is exactly where all practical MOSFET amplifiers operate much of the time, notably the aforementioned Adcom amplifiers, but even Nelson Pass's most cherished designs like the F5. His more commercial Pass Labs amplifiers are worse. For example, the XA30.5 is really a high bias Class AB amp with 150W maximum power output. It's not even close to the Super Class A that a truly linear MOSFET amplifier would use.
So then he got about to comparing the linearity of MOSFETS with transistors used in some transistor amplifiers, such as the Parasound amplifiers I own (I use a Parasound HCA-1500A in my master bedroom…where I used to use a Parasound HCA-1000A). He quickly came to like the devices and the designs of these amplifiers very much. They use very linear output devices in optimal circuits. He particularly liked the HCA-1500A. (This made me feel good.) But all the amplifiers in the series (850, 1000, 1200, 1500) were about equally good. He was not so impressed with my cherished Aragon 8008BB, noting the bipolars used in the front end and driver are not as good as the JFETs and MOSFETS used in corresponding positions in the Parasound amplifiers. All this came as a big surprise, since previously Tim was anything but a worshipper of John Curl (though he did think Curl was right about a lot of things back when he designed the famed JC2 preamp for Mark Levinson).