It's on page 94.
They're referring to the peak change in gain. The error averaged over a full waveform is smaller, by about that much.
For example, suppose the output waveform were a triangle wave (rather than the "barn roof" it actually is). We know the triangle wave has harmonics with amplitudes that go as 1/N^2 (for N odd). This sums to 1.47% THD. But the peak error is [assumption].
Hmm.
Excuse me while I step over to the blackboard and doodle something for a moment...
Ah, the closest sine wave to a triangle wave is 0.81246152 amplitude (best fit, 100 samples evenly spaced). That is, for a triangle of peak amplitude 1 and a sine of peak 0.81..., the RMS error is minimal.
In this case, the peak error is 18.7% while the RMS is 7%, which is, very approximately, the factor of 3 mentioned.
Crude demo:
https://docs.google.com/spreadsheets/d/1gBjter_vGnuW1iC6h7usQKf3WjN0YbGgO2lV13uEJVE/edit?usp=sharingYou may also find this helpful for seeing how such a calculation can be done?
If one were to input an exponential waveform instead of a linear ramp (of some given amplitude, because again, the distortion of an exponential is proportional to its amplitude), they'd get closer to the result H&H got. The waveform is peakier (a poorer fit to a sine), so the peak-to-RMS error can be expected to be higher, probably above 3 for the amplitude pictured.
(Ed: didn't really have to go and write a spreadsheet for that. It's sufficient to note that the harmonics are orthogonal; the RMS test is equivalent to an inner product and the inner product of fundamental with any harmonic is zero. So the "best fit" of course has to be the ratio of the triangle's fundamental amplitude, or 8/pi^2 ~= 0.81. Not that these terms are probably meaningful to the beginner!)
Tim