One key difference between feed-forward (FF) and feed-back (FB) AGC: almost any darn detector and gain cell can be "made to work" in FB, but FF requires that the detector and gain cell track very well.
It is suggested that Gyraf's vari-Mu limiter is FF. I would like to hear from someone who knows what is inside (Jakob?), but I doubt it. The Voltage/GR curve for tubes is not any simple law, and no ordinary detector could give the right signal to work in FF mode. Especially with tube variations and replacements. The advertising text does say "no feedback", but I bet it means "no audio feedback". While it is possible to devise a FF vari-Mu (I've played with it), it is not easy, and really wants a matched tube in the GR path, which this unit does not have.
Is FB "unstable"? At modest Ratio, no, or not so you'd notice, unless you have a very unsuitable choice of time constants. It comes down to basic feedback theory: error and overshoot are easily understood and often can be tamed. It does get harder at high ratio, and impossible to get "infinite compression". But how often do we hear a difference between 10:1 and infinity:1 (neglecting artifacts)? The ear is not such a precise judge of level, and musical signals are highly dynamic and imperfectly predictable.
I don't think there is anything basically wrong with FB for ordinary music limiting. And it does seem that its theoretical weaknesses can be musically useful.
If you are going to mess with levels, aiming for less dynamics, you probably do not want to do "perfect limiting". The ear is not a great judge of level, but when every peak is -exactly- the same, on sounds that should be varied, the ear senses that this is "not normal". You get that very bland flat sound that some of the older DBXes gave. The minor "errors" of a slightly bobbly FB limiter let the peaks vary a bit, leaving some life in the sound while allowing much higher average for the same peaks. The perfect FF limiter could maybe get a couple dB higher average at the same peak, but the ruler-flat peaks lose "2dB of liveliness" and may not "sound better" or significantly louder.
For a PWM gain cell, with a linear ramp/triangle switcher drive, the detector must be linear. If we actually want 40dB range (a lot), and the top is 10V, then the rectifier must be linear down to 0.1V with error less than about 10mV for 1dB error. That is far from impossible but much more precision than you need in a FF limiter.
Because DBX works in the exponential realm, the problem is less. Their error budget is still on the order of a few mV, but the max GR signal is only about 1V, and error distributes smoothly over the whole 40dB range. All you do is trim offset or unity-gain. But exponential action gets odd when you use R-C networks for time constants; DBX uses a wacky "non linear capacitor" to give a useful (and oddly specified) time-domain action.
If everybody here had milliVoltmeters and knew what they said, we might have some fun with FF. But for solder-pokers without a wall of test gear, a FB limiter is fairly self-correcting without precision action.
It is suggested that Gyraf's vari-Mu limiter is FF. I would like to hear from someone who knows what is inside (Jakob?), but I doubt it. The Voltage/GR curve for tubes is not any simple law, and no ordinary detector could give the right signal to work in FF mode. Especially with tube variations and replacements. The advertising text does say "no feedback", but I bet it means "no audio feedback". While it is possible to devise a FF vari-Mu (I've played with it), it is not easy, and really wants a matched tube in the GR path, which this unit does not have.
Is FB "unstable"? At modest Ratio, no, or not so you'd notice, unless you have a very unsuitable choice of time constants. It comes down to basic feedback theory: error and overshoot are easily understood and often can be tamed. It does get harder at high ratio, and impossible to get "infinite compression". But how often do we hear a difference between 10:1 and infinity:1 (neglecting artifacts)? The ear is not such a precise judge of level, and musical signals are highly dynamic and imperfectly predictable.
I don't think there is anything basically wrong with FB for ordinary music limiting. And it does seem that its theoretical weaknesses can be musically useful.
If you are going to mess with levels, aiming for less dynamics, you probably do not want to do "perfect limiting". The ear is not a great judge of level, but when every peak is -exactly- the same, on sounds that should be varied, the ear senses that this is "not normal". You get that very bland flat sound that some of the older DBXes gave. The minor "errors" of a slightly bobbly FB limiter let the peaks vary a bit, leaving some life in the sound while allowing much higher average for the same peaks. The perfect FF limiter could maybe get a couple dB higher average at the same peak, but the ruler-flat peaks lose "2dB of liveliness" and may not "sound better" or significantly louder.
For a PWM gain cell, with a linear ramp/triangle switcher drive, the detector must be linear. If we actually want 40dB range (a lot), and the top is 10V, then the rectifier must be linear down to 0.1V with error less than about 10mV for 1dB error. That is far from impossible but much more precision than you need in a FF limiter.
Because DBX works in the exponential realm, the problem is less. Their error budget is still on the order of a few mV, but the max GR signal is only about 1V, and error distributes smoothly over the whole 40dB range. All you do is trim offset or unity-gain. But exponential action gets odd when you use R-C networks for time constants; DBX uses a wacky "non linear capacitor" to give a useful (and oddly specified) time-domain action.
If everybody here had milliVoltmeters and knew what they said, we might have some fun with FF. But for solder-pokers without a wall of test gear, a FB limiter is fairly self-correcting without precision action.