Miller-capacitor compensation elegantly solves several problems at once, and the decision to adopt it is simple. However the question of whether to include the output stage in the Miller feedback loop is less easy. Such inclusion (see Figure 7.1c) presents the alluring possibility that local feedback could linearise both the VAS and the output stage, with just the input stage left out in the cold as frequency rises and global NFB falls. This idea is most attractive as it would greatly increase the total feedback available to linearise a distortive Class-B output stage.

There is certainly some truth in this, as I have shown[4], where applying Cdom around the output as well as the VAS reduced the peak (not rms) 1 kHz THD from 0.05% to 0.02%. However I must say that the output stage was deliberately under-biased to induce crossover spikes, because with optimal bias the improvement, although real, was too small to be either convincing or worthwhile. A vital point is that this demonstration used a model amplifier with TO-92 output transistors, because in my experience the technique just does not work well with real power bipolars, tending to intractable HF oscillation. There is evidence that inclusive compensation, when it can be made stable, is much less effective at dealing with ordinary crossover distortion than with the spikes produced by deliberate under-biasing.

The use of local NFB to linearise the VAS demands a tight loop with minimal extra phase-shift beyond...