There is definitely a change in loop gain, but no reduction in loop gain. That is, suppose you compare the loop gains of two amplifiers, one with and one without a properly-sized capacitor across the feedback resistor but otherwise equal. The one with the capacitor will then have a higher loop gain at high frequencies than its colleague, while it also has a higher phase margin/better damped poles. That's unusual, because most frequency compensation techniques improve stability at the expense of high-frequency loop gain.
In a distant past, I learned quite a lot from Ernst Nordholt. Among many other things, he categorized frequency compensation techniques. I think it was something like this, in order of decreasing preference:
1. Phantom zero techniques, such as this capacitor across the feedback resistor: improved stability without any loss of overall loop gain, in fact with an increase of overall loop gain. In some cases, an inductor in series with a resistive load or a capacitor across a resistive source can have a similar effect.
2. Techniques that exchange overall loop gain for loop gain in a local loop and do so frequency-selectively. Example: Miller compensation.
3 or 4. Techniques that exchange overall loop gain for loop gain in a local loop and do so over the whole band. Example: emitter or cathode degeneration, common base or grid stages, common collecter or anode stages, although it also depends on the impedance that drives those stages.
4 or 3. Techniques that reduce overall loop gain and do so frequency-selectively, but without any local feedback. Example: dominant pole capacitor or pole-zero compensation network between base and emitter or between grid and cathode.
5. Techniques that reduce overall loop gain, but not frequency-selectively and without any local feedback. Example: resistor between base and emitter or between grid and cathode to throw away loop gain.
The idea was to use techniques from the top of the list as much as feasible (without being dogmatic about it and while using common sense); unfortunately, it is almost never possible to compensate an amplifier with only phantom zeros. He didn't cover multipath techniques. That is a pity, because I think they belong high on the list.
In a distant past, I learned quite a lot from Ernst Nordholt. Among many other things, he categorized frequency compensation techniques. I think it was something like this, in order of decreasing preference:
1. Phantom zero techniques, such as this capacitor across the feedback resistor: improved stability without any loss of overall loop gain, in fact with an increase of overall loop gain. In some cases, an inductor in series with a resistive load or a capacitor across a resistive source can have a similar effect.
2. Techniques that exchange overall loop gain for loop gain in a local loop and do so frequency-selectively. Example: Miller compensation.
3 or 4. Techniques that exchange overall loop gain for loop gain in a local loop and do so over the whole band. Example: emitter or cathode degeneration, common base or grid stages, common collecter or anode stages, although it also depends on the impedance that drives those stages.
4 or 3. Techniques that reduce overall loop gain and do so frequency-selectively, but without any local feedback. Example: dominant pole capacitor or pole-zero compensation network between base and emitter or between grid and cathode.
5. Techniques that reduce overall loop gain, but not frequency-selectively and without any local feedback. Example: resistor between base and emitter or between grid and cathode to throw away loop gain.
The idea was to use techniques from the top of the list as much as feasible (without being dogmatic about it and while using common sense); unfortunately, it is almost never possible to compensate an amplifier with only phantom zeros. He didn't cover multipath techniques. That is a pity, because I think they belong high on the list.
Friends, resurrecting this stale thread.
AX published my little essay about bandwidth in this weeks' AudioVoice. Attached.
Comments?
Jan
AX published my little essay about bandwidth in this weeks' AudioVoice. Attached.
Comments?
Jan
Reality check , for tube power amps with output power transformer high open loop gain/low bandwidth variant is impossible because such power amp will be prone to oscillation and instability , only exception can be OTL tube amp. , J.Futterman designed such power amp already in early 50` .
Not if you know what you are doing. It's all a matter of correct compensation.
Bruno Putzeys once took the challenge and designed such an amp with 60dB of feedback.
Apparently it even sounded good ;-)
But yes, if you want to play it safe, take the fb from the xformer primary.
But maybe Bruno is not a fair example - if you know how to wrap a stable 6th order feedback loop around a class D amp, you're an exception.
Jan
Bruno Putzeys once took the challenge and designed such an amp with 60dB of feedback.
Apparently it even sounded good ;-)
But yes, if you want to play it safe, take the fb from the xformer primary.
But maybe Bruno is not a fair example - if you know how to wrap a stable 6th order feedback loop around a class D amp, you're an exception.
Jan
Can you take a low bandwidth system (like class D) and make it control as well or even better than a high bandwidth system, simply by inserting a brick-wall filter and increasing the gain?
No.
No.
B. Putzeys is certainly not a representative example of DIYers, (Arf!) but the principles remain. Thanks for the comprehensive paper. It seems to both summarize and conclude the Curl/Cordell dilemma from the 1980s? that we seem to have so much trouble getting past.
In the Cordell/Putzeys model (more is more) the unstated goal is to force errors to below some (very likely) minimum (ie: not important) level, which is always best approached by input signal bandwidth limiting and maximum loop feedback. This is also the best approach from a noise standpoint. Extra loop bandwidth at low frequencies is a bonus, so what do we trade off for this? I have this Second Law thing hanging over my head - don't judge me. Just how I was raised.
All good fortune,
Chris
In the Cordell/Putzeys model (more is more) the unstated goal is to force errors to below some (very likely) minimum (ie: not important) level, which is always best approached by input signal bandwidth limiting and maximum loop feedback. This is also the best approach from a noise standpoint. Extra loop bandwidth at low frequencies is a bonus, so what do we trade off for this? I have this Second Law thing hanging over my head - don't judge me. Just how I was raised.
All good fortune,
Chris
It's quite clear. It is essentially what Peter Garde wrote in 1977/1978, but looked at in the frequency rather than the time domain.
Hi Chris, can't say I disagree.
Maybe there is a tradeoff I/we miss, if so I am not aware.
Any hunch? 😎
Jan
There are references to it in http://linearaudio.net/sites/linearaudio.net/files/volume1ltemvdg.pdf They are originally IREE Australia articles, but republished by the AES.
Feedback models always assume monotonicity of errors with signal level, but that's only true for valves and inductor permeability if biased away from zero-crossing. Not including B. Putzeys' whacko modern Time Lords kind of amplifier, which has its own monotonicity things to deal with (better him than us, and Jesus Take the Wheel!), regular farm-raised linear amplifiers can't differ too much from the classic model of transconductance stage/integrator/follower (somewhat basterdized and smeared in classic valve amplifiers, but still all there) without some performance compromises. The simple model seems to be the best, if a lot of (obtainable - hey, it's just audio) constraints are observed.
The tricky thing seems to me to be a thermal one; monotonicity costs heat (in conventional linear amplifiers). This is the great strength of vacuum valve amplifiers, their ability to operate at high temperatures without complaint, so monotonicity is easy. Open loop linearity is another advantage, but poor impedance matching to expected loads (requiring output transformers) gives us timing problems for feedback. If this issue of monotonicity is made insignificant, then how do we define remaining issues?
Is it possible to (monotonically!) reduce errors to below perceptability? If so, done, where's my check? If not, is there a remaining error in our model? My personal (only) thought is that our human perception machinery is so malleable as to make definitive live-or-die conclusions foolish at best. We do not experience the world; we experience a model of the world that we generate from limited data and update as possible. Our human model was designed iteratively over many generations to keep us alive, and really shouldn't be expected to carry any heavier burdens.
Way off topic, sorry, love y'all,
Chris
The tricky thing seems to me to be a thermal one; monotonicity costs heat (in conventional linear amplifiers). This is the great strength of vacuum valve amplifiers, their ability to operate at high temperatures without complaint, so monotonicity is easy. Open loop linearity is another advantage, but poor impedance matching to expected loads (requiring output transformers) gives us timing problems for feedback. If this issue of monotonicity is made insignificant, then how do we define remaining issues?
Is it possible to (monotonically!) reduce errors to below perceptability? If so, done, where's my check? If not, is there a remaining error in our model? My personal (only) thought is that our human perception machinery is so malleable as to make definitive live-or-die conclusions foolish at best. We do not experience the world; we experience a model of the world that we generate from limited data and update as possible. Our human model was designed iteratively over many generations to keep us alive, and really shouldn't be expected to carry any heavier burdens.
Way off topic, sorry, love y'all,
Chris
This is an important question. Jan's article says you might as well use the excess loop gain of an integrator at low frequencies rather than make the open loop gain flat across the audio band. And there are real benefits that Jan describes. I'd watch out for motor-boating type effects in some designs, but that aside.
So Putzey does something akin to an extreme version of this. Rather than have a 20dB/decade, regular integrator roll-off, he implements more like a brick-wall filter roll-off, over a narrow bandwidth. It's a way to boost the excess gain at 20kHz when the unity gain has to be relatively low frequency due to the switching speed of class D. But is this sleight of hand from a control point of view?
If the steep filter technique is valid, it begs the question why linear amplifiers don't also use it? Think about it.
So Putzey does something akin to an extreme version of this. Rather than have a 20dB/decade, regular integrator roll-off, he implements more like a brick-wall filter roll-off, over a narrow bandwidth. It's a way to boost the excess gain at 20kHz when the unity gain has to be relatively low frequency due to the switching speed of class D. But is this sleight of hand from a control point of view?
If the steep filter technique is valid, it begs the question why linear amplifiers don't also use it? Think about it.
It is known as conditional stability and it was impopular because conditionally stable valve amplifiers tended to oscillate when the cathodes were not yet hot enough. I wonder what Putzeys did to prevent that in the valve amplifier Jan wrote about.
Some also need special tricks (state variable limiting) to properly recover from clipping, but then again, some don't, and some amplifiers with a more conventional frequency compensation also need special tricks to recover from clipping.
One of the first threads I ended up in at diyAudio was about conditional stability:
https://www.diyaudio.com/community/threads/opamp-with-open-loop-gain-of-1-000-000-000-000-000-p.16936/
It's also been used recently in the solid-state forum:
https://www.diyaudio.com/community/threads/three-pole-compensated-blameless-clone.397740/
Some also need special tricks (state variable limiting) to properly recover from clipping, but then again, some don't, and some amplifiers with a more conventional frequency compensation also need special tricks to recover from clipping.
One of the first threads I ended up in at diyAudio was about conditional stability:
https://www.diyaudio.com/community/threads/opamp-with-open-loop-gain-of-1-000-000-000-000-000-p.16936/
It's also been used recently in the solid-state forum:
https://www.diyaudio.com/community/threads/three-pole-compensated-blameless-clone.397740/
I'm actually repeating myself... https://www.diyaudio.com/community/threads/negative-feedback.393351/post-7211926
We'll need a new acronym for this old chestnut. I'm leaning towards including Otala and Curl for the late 1970s proposal,** Cordell for the 1980(?) rebuttal, and Jan D, Marcel vdG and B. Putzeys for latter day saints. Any catchy acronyms come to mind?
edit: must include Garde late 1970s rebuttal
Keep it clean; this is a family show!
Chris
edit: must include Garde late 1970s rebuttal
Keep it clean; this is a family show!
Chris
Last edited:
May I ask you more experienced people what do you think about feeding back only the amplified distortion?
The feedback path has a gain stage that subtracts the input signal from the scaled down and rephased output signal, then amplifies it and feeds it back to the driver or phase splitter.
Common names I’ve found for them is ODNF or Corr.Diff.
Thanks in advance.
Roberto
The feedback path has a gain stage that subtracts the input signal from the scaled down and rephased output signal, then amplifies it and feeds it back to the driver or phase splitter.
Common names I’ve found for them is ODNF or Corr.Diff.
Thanks in advance.
Roberto
Perhaps the math whiz folks could elaborate, but a fundamental concept is that any analog real time sharp filter produces a massive phase shift, which would make negative feedback into positive feedback. You ~cannot make an analog "brick-wall filter" without massive phase shift. Digital filters can do magic, but they are not "real time", ie they require a time delay, so that "future" data is available for the calculations. The classic "single dominant pole" works because it has a maximum phase shift of 90 degrees while eventually pulling the OLG down below 0dB where it is no longer a threat of instability.
Basically it is three possible variant from where or how to take fb signal ,
1) from OPT secondary coil which feed (load)speaker ,
2) from OPT primary coil
3) from OPT separate coil
,this third variant I have seen in many domestic made Iskra hi-fi tube amps but also in many other comerciall tube amps ,
btw ,IMHO when on some tube amp all amplifying stages including output power stage work in non switching A-class than OL wide bandwidth and relative low level or even no GNFB is good design solution , but when output power stage work closer to class B than higher level of GNFB correction is mandatory which again automatically pulls for higher OLG , tube amps with OPS which work closer to energy economic class B (usually PP amps) one of the main design idea is to significantly extend output power tubes life also .
Attachments
See https://www.diyaudio.com/community/threads/opamp-with-open-loop-gain-of-1-000-000-000-000-000-p.16936/ post #9 for a demonstration circuit showing that it is not quite that simple.
This is feedback, only more convoluted.
Take a fraction of Vout, subtract from Vin, insert somewhere at the input stage.
As long as you keep the phase correct, either in the summing or with a separate phase flipper, it's classic feedback.
The gain can be in the forward path or in the feedback path, no conceptual diffference.
Set up the equation for Vout/Vin and it will reduce to the well-known feedback equation.
jan