Many thanks to Dimitri Danyuk for permission to include these output stage simulations.
Here the open-loop linearity of the output stage is being investigated.
An effect not obvious from the measurements is clearly revealed in this simulation as a small peak in the gain at the point where the lower half cuts off. This appears to be caused by the current through Q4 not being compensated for accurately by the error amplifier because it only passes through half the output resistor. Normally Q4 current is a small proportion of the output current, but Q1 cuts off before Q4 and then current from Q4 alone reaches the load and is not fully cancelled by the error amplifier. Simply increasing R4 from 47ohms to 150ohms reduces the peak as in the lower graph. Unfortunately there are other effects when this resistor is increased, one of which is that the base-emitter voltage of Q4 is reduced, and so the quiescent current increases. Resetting the current to its original value the large signal distortion is then found to have increased. This is because the base-emitter voltage of Q4 now varies over a larger range, and the extra error resulting from this causes the error amplifier to switch off sooner as it tries to compensate. Small signal distortion is reduced a little, but high signal distortion is increased more. If the increase in R4 is considered worthwhile the quiescent current Iq needs increasing to keep large signal distortion close to the original level. The improvement in small signal distortion at 20kHz with R4 increased is visible on the extracted distortion waveform, a sharp edge of the wave becoming more rounded, but there is no obvious change in the total amplitude. A value of 100ohms for R4 seems to give almost the same visible improvement as 150ohms, so this value seems a good compromise, with Iq increased to 120mA.
The step in the gain as the error amplifier cuts off looks bad, but only a small range of gain is shown in the diagram, and actually it is less than a 0.5% change in the slope of the output vs input graph, not a step in the output voltage. With an 8ohm load instead of 4ohms the effect will be smaller and occur only at a higher output voltage. To put the gain step in context a standard class-B amplifier using typical 5% tolerance 0.25 ohm output resistors could have a similar step in gain purely from the resistors having different values within their tolerance, this in addition to the usual crossover nonlinearity, and occuring at small signal levels rather than as here at 3 amps output.
According to the simulation an increase in quiescent current to 600mA will ensure that the error amplifier does not switch off even with the 4ohm load, but of course an efficient heatsink is then needed. An alternative approach is to adjust the values of the two 0.25ohm resistors. Reducing R2, e.g. by about 10%, will increase the negative signal output current from the lower half. If the current is then too high the error amplifier will conduct more and add positive current to compensate. If R2 is reduced R1 should be reduced by the same amount to keep R1 = R2 + R3. Initial experiments suggest that this method does work, and a resistor of 2R7 in parallel with R2 plus another 2R7 in parallel with one of the two 0R25 making up R1 prevented the error amplifier switching off. One unfortunate effect, not present in the unmodified amplifier, is that with moderate clipping a serious latch-up effect was observed. It may be that this is only a problem for this particular design, and with a more conventional input and driver it will be better behaved, so this method may still be possible, but for the present design it is not recommended without further investigation.
Without this modification, and with Iq = 120mA, the distortion at high signal levels appears to be mostly second harmonic, with third harmonic rising only near clipping, as would be expected. Small-signal linearity is comparable to a good class-A design, and a little second harmonic at higher levels (about 0.014% at 20kHz at 3dB below clipping) is nothing to worry about. My conclusion from the test results and simulations is that some further improvement may be possible, but even so the result is already significantly better than would be expected from a typical class-B or class-AB design of similar complexity.