This idea is quite obvious so I'm sure someone has done it before, but I've never seen it mentioned anywhere so I'll post it anyway. I'm also not sure what it should be called; "active feedback" would be a fitting name, but searching on both this forum and Google reveals that name already being used to describe a variety of disparate techniques which are not this.
On to the heart of the matter: Negative feedback subtracts the output from the input, reducing nonlinearity and increasing bandwidth at the expense of gain. If instead of feeding back the output, the difference between the output and input is fed back, then linearity and bandwidth can be improved without sacrificing gain.
The attached schematic shows a practical implementation of the concept, with X1 being the main amplifier and X2 being the error amplifier. The error amplifier should be faster than the main amplifier to avoid ringing, or even instability. It should also be as linear as possible, which goes without saying, but it doesn't have to supply any current (ideally).
Results to follow...
On to the heart of the matter: Negative feedback subtracts the output from the input, reducing nonlinearity and increasing bandwidth at the expense of gain. If instead of feeding back the output, the difference between the output and input is fed back, then linearity and bandwidth can be improved without sacrificing gain.
The attached schematic shows a practical implementation of the concept, with X1 being the main amplifier and X2 being the error amplifier. The error amplifier should be faster than the main amplifier to avoid ringing, or even instability. It should also be as linear as possible, which goes without saying, but it doesn't have to supply any current (ideally).
Results to follow...
Attachments
Distortion could theoretically be made arbitrarily low by increasing the gain of the error amplifier, but the main amplifier will need to be increasingly slower or the corrections of the error amplifier will cause ringing at frequencies where both the phase shift of the error amplifier is significant and the gain of the main amplifier still high enough. If my thinking is correct, then the reduction in distortion is ideally equal to the gain of the error amplifier (20dB as shown). Simulations show about this much improvement.
In reality I can't measure low distortion, so to see how it works in practice I added an antiparallel pair of diode in series with R1 to simulate massive odd harmonic distortion easily visible on an oscilloscope. The attached picture shows what this looks like before error correction.
More results in the next post...
In reality I can't measure low distortion, so to see how it works in practice I added an antiparallel pair of diode in series with R1 to simulate massive odd harmonic distortion easily visible on an oscilloscope. The attached picture shows what this looks like before error correction.
More results in the next post...
Attachments
After error correction it looks like the picture attached to this post.
Doing a quick test on a breadboard with a dual op-amp shows that in reality it does exactly what the simulations say, surprise surprise. I haven't bothered showing anything about the increased bandwidth because I don't have a signal generator capable of testing an op-amp to the limit, but simulations show an improvement of similar magnitude to the distortion.
So has anyone made anything like this before, or read about it? Any drawbacks I haven't thought of (increased complexity and care needed for stability is all I can think of)?
Doing a quick test on a breadboard with a dual op-amp shows that in reality it does exactly what the simulations say, surprise surprise. I haven't bothered showing anything about the increased bandwidth because I don't have a signal generator capable of testing an op-amp to the limit, but simulations show an improvement of similar magnitude to the distortion.
So has anyone made anything like this before, or read about it? Any drawbacks I haven't thought of (increased complexity and care needed for stability is all I can think of)?
Attachments
re: Unnamed feedback method explored
Mr. Evil,
Here is a circuit which looks similar, although it does not incorporate feed back from the main opamp output.
The link is:
http://www.lsionline.co.uk/lsi/features/techfocus/article.asp?ID=-I9LOCH
If you go to electronicsweekly.com and put "sandman" in the search tool, it displays three related articles.
Maybe it helps.
Tom
Mr. Evil,
Here is a circuit which looks similar, although it does not incorporate feed back from the main opamp output.
The link is:
http://www.lsionline.co.uk/lsi/features/techfocus/article.asp?ID=-I9LOCH
If you go to electronicsweekly.com and put "sandman" in the search tool, it displays three related articles.
Maybe it helps.
Tom
Interesting. Sandman's topology does look similar in principle. However, it doesn't improve bandwidth and I think it doesn't decrease distortion by as much as the one I posted, although it would take further analysis to determine that.
The responses to his circuit mentioned such techniques being used for a long time, but no details given, so I still don't know why I never see anything like this actually used.
The responses to his circuit mentioned such techniques being used for a long time, but no details given, so I still don't know why I never see anything like this actually used.
Evil,
It's been thought of before (I can't remember where,
but it probably wasn't original then, either).
It works, but you would have to have a pretty lousy
gain stage for X1 to want to bother, and as you point
out yourself, it will be rife with stability issues.
On the other hand, it might sound great and you can
drive Halcro out of business. Now that would be really
evil
It's been thought of before (I can't remember where,
but it probably wasn't original then, either).
It works, but you would have to have a pretty lousy
gain stage for X1 to want to bother, and as you point
out yourself, it will be rife with stability issues.
On the other hand, it might sound great and you can
drive Halcro out of business. Now that would be really
evil
I believe this is just another one of the many possible composite amplifiers that in the end can be shown to have lower distortion due to added loop gain
Walt Jung has shown composite op amps with gain in the loop, Jerald Graeme discusses them in his books&articles, Soliman in WW, Randall Geiger in IEEE…
Walt Jung has shown composite op amps with gain in the loop, Jerald Graeme discusses them in his books&articles, Soliman in WW, Randall Geiger in IEEE…
Stability needs extra attention, but isn't really a big problem. X1 should have a GBP lower than X2 by a factor of the gain of X2. So if X2 has a GBP of 10MHz and a gain of 10, then if X1 has a GBP of 1MHz everything is fine. That's how it is in the one I tested and it shows no overshoot on transients. The good point however is that even though X1 has reduced bandwidth, the overall GBP is still 10MHz. Higher bandwidth for X1 is possible if some overshoot is permissible.Nelson Pass said:Evil,
It's been thought of before (I can't remember where,
but it probably wasn't original then, either).
It works, but you would have to have a pretty lousy
gain stage for X1 to want to bother, and as you point
out yourself, it will be rife with stability issues.
On the other hand, it might sound great and you can
drive Halcro out of business. Now that would be really
evil
Of course if the main amplifier is already good then it might not be worth the effort, but one attractive aspect is that it seems to reduce 3rd harmonic distortion a lot. 2nd and 3rd are reduced at lower frequencies, but as the phase shift of the main amp increases then the effect reduces to only 3rd harmonic reduction. Who wouldn't like to be able to put on the spec sheet that 3rd harmonic distortion is down 140dB at 20kHz, even if it doesn't make an audible difference!
I'm not in the commercial electronics business, so I won't be putting anyone out of business, even if it did prove to have practical advantages.
Thanks for the info. Do you have any particular links, or specific books I could look at to learn more about it?jcx said:
The schematic I posted can be used for a real audio amplifier, if it's inverting. If it's non inverting then it requires a slight rearrangement.
Yes, you can use two different amplifiers to do it. In fact it needs the two amplifiers to be different. The main amplifier can be anything; the error amplifier should be something wide bandwidth, linear and it can be low-power. An op-amp would be a good, simple way of doing it. A highly linear class-A amp would be ideal though.
Yes, you can use two different amplifiers to do it. In fact it needs the two amplifiers to be different. The main amplifier can be anything; the error amplifier should be something wide bandwidth, linear and it can be low-power. An op-amp would be a good, simple way of doing it. A highly linear class-A amp would be ideal though.
Evil,
It is similar, but not identical to the Hawksford error correction, explored elsewhere in this forum.
Hawksford also takes the difference between input and output of a (gain = 1) stage, but adds this EXACT difference to the source.
This exact difference is significant. In Mr. Evils' graph after correction, you still see some remnant of the xover distortion. In theory, this can be made arbitrarily small but not zero, by arbitrarily increasing the gain of the correction amp, as has been shown, but you quickly get into stability issues.
The beauty of the Hawksford system is that by copying the error to the source, the error in theory becomes zero.
Of course, there's no free lunch here either, because the limiting factor becomes how accurately you can set the exact times-one adding of the error to the source.
Feedback is a fascinating subject! Think what all those non-global-feedback guys are missing
Jan Didden
It is similar, but not identical to the Hawksford error correction, explored elsewhere in this forum.
Hawksford also takes the difference between input and output of a (gain = 1) stage, but adds this EXACT difference to the source.
This exact difference is significant. In Mr. Evils' graph after correction, you still see some remnant of the xover distortion. In theory, this can be made arbitrarily small but not zero, by arbitrarily increasing the gain of the correction amp, as has been shown, but you quickly get into stability issues.
The beauty of the Hawksford system is that by copying the error to the source, the error in theory becomes zero.
Of course, there's no free lunch here either, because the limiting factor becomes how accurately you can set the exact times-one adding of the error to the source.
Feedback is a fascinating subject! Think what all those non-global-feedback guys are missing
Jan Didden
Mr Evil said:
I have seen your pictures. It works but I have to comment a little bit. The image (curve) that we can see after error correction is still highly distorted. The correction circuit must be able to remove cross-over distortion to that extent that it is invisible for the oscilloscope. Then you start with spectral analysis and will see your results.
Regards,
Pavel
If only I was still at university, where they had a large backcatalogue of those. Maybe my local library will have some.name said:
Indeed, I've just been reading some of his publications. Very interesting, and again, something that seems as though it should be used a lot more than it is. The difference is that it is feedforward, where this is feedback.janneman said:
Yes it's still distorted. That's because it started of with an unholy amount of distortion in the first place. I've attached a fourier transform of the simulation results matching the two posted earlier. You can see that distortion is reduced by about 20dB for all harmonics, which is the gain of the error amp, as mentioned before.PMA said:
I didn't think it would be new, since it is an obvious step to take with feedback. I'm just surprised I haven't seen it before. I was thinking there might be some compelling reason why not to use it.Upupa Epops said:Nothing new on this planet
Attachments
Eric, (?)
A possible reason that it is not used more could be that the traditional feedback gives the same reduction of 20dB if you add 20dB gain to the main amp.
And don't underestimate the power of tradition. We (the engineering world) is so much used to using normal feedback that you have to come up with something truly spectecular like saving 50% in manufacturing to even get them to take notice.
Jan Didden
A possible reason that it is not used more could be that the traditional feedback gives the same reduction of 20dB if you add 20dB gain to the main amp.
And don't underestimate the power of tradition. We (the engineering world) is so much used to using normal feedback that you have to come up with something truly spectecular like saving 50% in manufacturing to even get them to take notice.
Jan Didden
PMA got me thinking about the limits of the error correction, so I tried increasing the gain of the error amp to 40dB. Here's a fourier transform of those simulation results, showing a 40dB reduction in distortion. Now it's brought the distortion down to almost reasonable levels. Even harmonic distortion is off the scale, below -140dB.
The second graph is the same test performed at 20kHz. Here you can see the effect of the reduced bandwidth of the main amp (1/10th the bandwidth it was when the error amp only had a gain of 10). Odd harmonic distortion is still reduced by the same amount, but the effect on even harmonics is not as big anymore, with 2nd harmonic only just below -120dB, which is about the same as without error correction at all.
The second graph is the same test performed at 20kHz. Here you can see the effect of the reduced bandwidth of the main amp (1/10th the bandwidth it was when the error amp only had a gain of 10). Odd harmonic distortion is still reduced by the same amount, but the effect on even harmonics is not as big anymore, with 2nd harmonic only just below -120dB, which is about the same as without error correction at all.
I'm not called Eric!janneman said:
Attachments
Mr Evil said:
Yes, this is what I would expect from the error correction. According to my experience it is possible to get 40 - 50 dB distortion reduction, and this is worthwhile. The combination of output stage error correction + NFB leads to excellent measured and sonic results. There are several advantages of an error correction: it reduces high harmonics caused by cross-over distortion, reduces and linearizes output impedance, extends bandwidth (reduces turn-on and turn-off spikes and phase shifts).
- Status
- This old topic is closed. If you want to reopen this topic, contact a moderator using the "Report Post" button.