Q: What do you get when you put two pedants into one thread?
A:
I had hoped my "control theory" mention would've cued onto the concept of a controller and a plant (the plant generally being some squirrelly thing with multiple poles and/or delay), the canonical problem of control theory. But oh well.
Actually, the audio amp example is still proof that the amp must be slow: there's always a Miller cap or dominant pole or some kind of compensation in the loop, so that the amp itself can be stable at the desired gain. There is no such thing as an op-amp or audio amp without this feature, explicitly (e.g., a cap C-B) or parasitically (e.g., reduced bias allowing junction Ccb to play the role).
So, shut up.
To be perfectly precise (oh boy here we go...), suppose we have a plant with transfer function P = 1 / (1 + s/a) (i.e., unity gain at DC, and a cutoff frequency of 2*pi*a). This is Laplace domain, or s --> j*w if you prefer Fourier/analytic domain.
Now suppose we have a controller A = k / s, whose input is the difference between source (setpoint) and output (feedback). The loop is then:
(Q = output, I = input, D = difference, L = O/I = overall loop transfer function)
Q = D * (k / s)*[1 / (1 + s/a)]
D = I - Q
0 = (I - Q) * (k / s)*[1 / (1 + s/a)] - Q
0 = I * (k / s)*[1 / (1 + s/a)] - O * (1 + (k / s)*[1 / (1 + s/a)])
O * (1 + (k / s)*[1 / (1 + s/a)]) = I * (k / s)*[1 / (1 + s/a)]
O / I = (k / s)*[1 / (1 + s/a)] / (1 + (k / s)*[1 / (1 + s/a)])
L = 1 / (1 + 1 / {(k / s)*[1 / (1 + s/a)]})
= 1 / [1 + (s*(1+s/a)/k)]
= 1 / [1 + s/k + s^2/(a*k)]
Which has the form of 1 / (s^2 + a*s + b^2), which has poles:
s = -1/2k +/- sqrt(1/(4*k^2) - 1/(a*k))
In other words, when 1/(4*k) > 1/a, the roots are real, meaning the system is stable and overdamped; otherwise, the roots are complex, and the system rings (underdamped), but still remains stable.
This is the most marginal case for an integrating controller, a single pole in the plant. If there are additional poles, the loop gain must roll off much sooner than those poles, otherwise the loop will go into the right half-plane where it is unstable (oscillates uncontrollably). A pole-zero compensator can be tuned for better performance, or to tolerate additional poles, but the fact remains: loop gain cannot be unlimited into a load with more than zero poles, and must roll off below 4x the plant's lowest pole. QED.
TLDR: jerk uses math on the internet. On the plus side, you get to see an actual application of the quadratic equation, so, if you were ever wondering...
Tim