This post is to illustrate what the problems are with measuring amplifier using a reactive load.
Here is a simple model of an amplifier connected to a speaker driver in LTspice. This model is copied from the MATLAB "
Loudspeaker Modeling with Simscape" tutorial (the "simplified model"). The driver parameters used are from SB Acoustics SB10PGC21-4 3" full range driver. The voltage source V1 is the "theoretically perfect" amplifier, but with a 0.1 Ω source resistance. It generates a 3 cycle 100 Hz sine wave tone burst.
The green trace shows the amplifier voltage output. The red trace "V_out" is the voltage across the amplifier output (i.e. V1 in series with the source resistance). You can see that the red trace does not go to zero immediately when the tone burst terminated. The second picture had the vertical axis magnified to show the overshoot and decay of the red trace. The cause of the back EMF of the driver as it is a reactive load which can store and subsequently release energy. The voltage reading "V_out" is a combination of both "V1" and the back EMF.
The 0.1 Ω source resistance gives a damping factor of 40 into a 4 Ω load, while not super high, is fully adequate according to Dr Toole's "
Damping, Damping Factor, and Damn Nonsense" article (20 was enough), if he is to be believed. Before anyone complains that amplifiers should have zero output impedance, let's consider taking the measurements at the speaker terminals instead. Now the speaker wire resistance is part of the source resistance, and the 0.1 Ω value is not unrepresentative of real-life setups.
Here are the FFT of both V1 and V_out, which shows the differences in their spectral compositions. The deviations of V_out from V1 will be interpreted by the analyzer as distortions (as linear and/or nonlinear distortions).
So, here we have a theoretically perfect voltage source (amplifier), but the measurements won't show it as perfect. It is obvious that attributing the "measured distortions" to the perfect voltage source/amplifier is a faulty interpretation of the results. Also, this analysis used a highly simplified linear speaker model. In real-life, speaker drivers are nonlinear — the spider/surround is a nonlinear spring, mechanical damping is nonlinear, the back EMF is nonlinear as the electromagnetics are nonlinear, etc. Therefore, it is likely that using a real speaker driver as the load will worsen these effects.
There are good reasons why the industry doesn't standardize on measuring amplifier distortions with reactive loads. If Occam's razor is to be believed, the simple reason is that it's usefulness is rather limited.
Regarding the assertion that amplifier performance is different when the load is complex, I'll point out the various TPA3255 amplifiers as example. The load dependent behavior of the frequency response can be solely explained by the response of the
output LC filter. Because of that, we can deduced that the input to the filter is reasonably frequency independent. It is easy to understand how. The TPA3255 chip has its feedback taken at the chip output pins, and the feedback corrections work. However, without PFFB, the feedback circuit is blind to the effect of the load on the LC filter. The chip has no problem with "the correct output" at its output pins, even as the output LC filter together with the speaker/dummy load is always a complex load for the chip. It is the LC filter that causes the load/frequency dependency. So the amplifier chip works as intended and is always outputting to a complex load.
With basic PFFB, the load dependency can be, to a large extent, corrected in the audible range. However, for example, as in the cases of the Aiyimas, they are not fully corrected. The PFFB implementation per the
TI app report is limited by the amount of feedback available. The TPA3255 chip has a gain of 21.5 dB. The app report implementation used 5.6 dB for feedback, and left the amplifier with a gain of 15.9 dB. For more complete correction, higher amounts of feedback (and possibly different implementation methods) are necessary.
