gootee said:
Hi Tom,
I've also experimented a bit with parasitics in simulation, without any clear cut conclusions at this point. However, doing so suggests a specific implementation obviously, so I'm wondering do you model a particular amp or some sort of best practice say where the entire amp including the outputs are on a single PC board where you can use an assumed length for traces? Different decoupling caps will have different parasitics, it gets quite complicated.
Just to mention, we do this in ASIC design, where by the way there are often more than 100,000 transistors, and every length of metal or interconnect is accounted for in the timing model. Placeholders are used pre-layout, and then the actual physical layout numbers are back annotated into the simulation model.
Pete B.
Tom, if you will just note the measurements of the JC-1 Amp done by 'Stereophile' you will find very little 'compromise' My slew rate is over 100V/us, but I probably could make it 500V/us if I used a coil. I have done it in the past with a similar design.
My output impedance with frequency is LOWER than the Halcro (with coil) by a factor of 2. My amp is unconditionally stable with virtually any load. At least, it is supposed to be. However, I did get someone (John Atkinson) to get the JC-1 to oscillate, on one occasion. However, I have never heard back as to why, or what conditions this occurred. The techs at Parasound have never duplicated or heard of it happening, nor my boss at Parasound who handled the JC-1 that is now in operation for the last month at John Atkinson's residence. It is still a mystery to me.
My output impedance with frequency is LOWER than the Halcro (with coil) by a factor of 2. My amp is unconditionally stable with virtually any load. At least, it is supposed to be. However, I did get someone (John Atkinson) to get the JC-1 to oscillate, on one occasion. However, I have never heard back as to why, or what conditions this occurred. The techs at Parasound have never duplicated or heard of it happening, nor my boss at Parasound who handled the JC-1 that is now in operation for the last month at John Atkinson's residence. It is still a mystery to me.
john curl said:
Hi John,
The JC-1 is indeed a very good amp. I think 100 V/us is plenty.
One thing I'm not connecting on in your comments here is why you would be able to have a higher slew rate approaching 500 V/us if you used an output coil? Is it because the use of the coil would give you more freedom in how you compensated it?
Thanks,
Bob
Discussing about negative feedback, theres an interesting thing about behavior of global FB amps.
This is proposed by PMA here : http://web.telecom.cz/macura/mixtest.html
Look at the residual graphs when there is 31khz square signal present and when it is unpresent (testing of Sony TA-FE320R)
This could be one reason why a single tone THD fails to predict how one particular amp will sound.
This is proposed by PMA here : http://web.telecom.cz/macura/mixtest.html
Look at the residual graphs when there is 31khz square signal present and when it is unpresent (testing of Sony TA-FE320R)
This could be one reason why a single tone THD fails to predict how one particular amp will sound.
Absolutely interesting!
Thanks Lumanauw for pointing this out, and thanks to PMA for very clear graphs.
I recall a published test a decade ago where inserting a sharp cut-off filter at 23 KHz between a CD player and an amplifier was preferred to the non-filtered version in a blind test. This makes it clear perhaps why.
Thanks Lumanauw for pointing this out, and thanks to PMA for very clear graphs.
I recall a published test a decade ago where inserting a sharp cut-off filter at 23 KHz between a CD player and an amplifier was preferred to the non-filtered version in a blind test. This makes it clear perhaps why.
Device and simple circuit simulations
I noticed in the discussion about NFB and distortion that AndyC simulation was beginning with essentially no starting distortion. What has been done to simulate real circuits where the inherent distortion properties of the devices themselves are added into the mix? For example there was a study done by Boyk and Sussman showing distortion for simple single devices and then complementary (or pushpull for a triode) and a differential circuit.
They go far beyond Baxandall in simulating out to I believe the 20th harmonic.
http://www.its.caltech.edu/~musiclab/feedback-paper-acrobat.pdf
I think it is interesting that for both the BJT and for the Triode they show a full spectrum of distortion products. What is interesting in their simulation is that for a given input voltage the amplitudes of the distortion products for the triode are much lower than for the BJT. It is also interesting to note that they show with NFB that the BJT all harmonics reduce in level.
For the case of the MOSFET they show that in Class A with no feedback one gets only low order and even harmonics for a single ended case but for pushpull if the circuit is not pure Class A there are problems and feedback doesn't improve the situation. For pure Class A with perfectly matched FETS they show no distortion at all and adding feedback doesn't result in the generation of any. Again for a given input voltage the distortion amplitude is higher for a FET than for a tube.
If these results could be combined into one of the more complicated circuit design simulations do you think we would then get a more realistic picture of what is going on audibly?? The problem is, and I agree with John Curl here, that what the meter reads and what we hear don't agree very well with normal distortion measurements.
Someone also mentioned that Self believes that once the distortion is in the noise floor it is no longer audible but is this really true and how low is really low enough? After all distortion is signal correlated and true noise is not. We hear below the noise floor all the time when listening to analog tape (the hiss being true random noise) to correlated music signal. Has this been answered conclusively? If not then perhaps knowing the limit first would be useful for optimizing the designs.
Finally, I remember reading an article on distortion by Crowhearst where he showed mathematically that feedback would multiply the order of the distortion as its fed back with the conclusion that what might appear to be noise floor is in fact a myriad of harmonic and intermodulation distortion peaks making essentially a signal correlated "noise" floor, which in reality was distortion and not noise. I am sure most have read this article but is it discredited or has it still some relevance to the discussion? If it has been discredited then please explain to me where the flaw in his logic lies.
I noticed in the discussion about NFB and distortion that AndyC simulation was beginning with essentially no starting distortion. What has been done to simulate real circuits where the inherent distortion properties of the devices themselves are added into the mix? For example there was a study done by Boyk and Sussman showing distortion for simple single devices and then complementary (or pushpull for a triode) and a differential circuit.
They go far beyond Baxandall in simulating out to I believe the 20th harmonic.
http://www.its.caltech.edu/~musiclab/feedback-paper-acrobat.pdf
I think it is interesting that for both the BJT and for the Triode they show a full spectrum of distortion products. What is interesting in their simulation is that for a given input voltage the amplitudes of the distortion products for the triode are much lower than for the BJT. It is also interesting to note that they show with NFB that the BJT all harmonics reduce in level.
For the case of the MOSFET they show that in Class A with no feedback one gets only low order and even harmonics for a single ended case but for pushpull if the circuit is not pure Class A there are problems and feedback doesn't improve the situation. For pure Class A with perfectly matched FETS they show no distortion at all and adding feedback doesn't result in the generation of any. Again for a given input voltage the distortion amplitude is higher for a FET than for a tube.
If these results could be combined into one of the more complicated circuit design simulations do you think we would then get a more realistic picture of what is going on audibly?? The problem is, and I agree with John Curl here, that what the meter reads and what we hear don't agree very well with normal distortion measurements.
Someone also mentioned that Self believes that once the distortion is in the noise floor it is no longer audible but is this really true and how low is really low enough? After all distortion is signal correlated and true noise is not. We hear below the noise floor all the time when listening to analog tape (the hiss being true random noise) to correlated music signal. Has this been answered conclusively? If not then perhaps knowing the limit first would be useful for optimizing the designs.
Finally, I remember reading an article on distortion by Crowhearst where he showed mathematically that feedback would multiply the order of the distortion as its fed back with the conclusion that what might appear to be noise floor is in fact a myriad of harmonic and intermodulation distortion peaks making essentially a signal correlated "noise" floor, which in reality was distortion and not noise. I am sure most have read this article but is it discredited or has it still some relevance to the discussion? If it has been discredited then please explain to me where the flaw in his logic lies.
Re: Device and simple circuit simulations
Actually, the distortion of the output stage I simulated earlier was just under 0.1 percent without feedback, at the power level where the crossover distortion is maximized. To get a feel for this, have a look at figure 4 here, showing the distortion vs. power for Charles Hansen's amp without global feedback. See the hump at about 10-40 W? That's the max of the crossover distortion, where the transition from class A to class AB occurs. I found where that hump was in simulation for my circuit, and chose that as the signal level to test. Note that to get the distortion over 0.1 percent in Charles' circuit, you have to load it with 2 Ohms. With 4 Ohms, it's quite a bit lower - lower in fact than the starting distortion I used in my simulation.
That's exactly what the last umpteen pages of this thread have been all about. The circuit simulator SPICE has the equations for transistor nonlinearity as part of its code, and the distortion is computed using FFT techniques. These things are already taken into account in the simulations and plots. That's what they are. If there are some things you don't understand about what you're reading, just ask.
They do more harmonics than Baxandall does, but only do about three different feedback values. And they don't attempt to show in a clear way the effect of feedback on each harmonic individually.
It only appears that way because they did not take data for enough different levels of feedback to show what Baxandall's analysis shows.
I'd be interested in seeing a PDF of the Crowhurst article if you have one.
morricab said:
Actually, the distortion of the output stage I simulated earlier was just under 0.1 percent without feedback, at the power level where the crossover distortion is maximized. To get a feel for this, have a look at figure 4 here, showing the distortion vs. power for Charles Hansen's amp without global feedback. See the hump at about 10-40 W? That's the max of the crossover distortion, where the transition from class A to class AB occurs. I found where that hump was in simulation for my circuit, and chose that as the signal level to test. Note that to get the distortion over 0.1 percent in Charles' circuit, you have to load it with 2 Ohms. With 4 Ohms, it's quite a bit lower - lower in fact than the starting distortion I used in my simulation.
That's exactly what the last umpteen pages of this thread have been all about. The circuit simulator SPICE has the equations for transistor nonlinearity as part of its code, and the distortion is computed using FFT techniques. These things are already taken into account in the simulations and plots. That's what they are. If there are some things you don't understand about what you're reading, just ask.
They do more harmonics than Baxandall does, but only do about three different feedback values. And they don't attempt to show in a clear way the effect of feedback on each harmonic individually.
It only appears that way because they did not take data for enough different levels of feedback to show what Baxandall's analysis shows.
I'd be interested in seeing a PDF of the Crowhurst article if you have one.
(Sorry I took so long to respond and this is now so out-of-sequence.)
John,
First, I probably should say that whatever (technical) information or opinions I post here usually shouldn't count for much, since I am, relative to most others here, grossly-inexperienced and am still mostly just trying to learn (or re-learn).
But the 0.5uH coil value that you mentioned reminded me that back when I was doing all of those spice amplifier simulations that I mentioned, it seemed like whenever I tried really-hard to max-out the performance (i.e. trying for "perfect" squarewave response with large and small capacitive loads, and THD-20 less than, say, .0001%), the optimal output-coil value (at least for the ones where I was determined to use a coil) often eventually converged to something around 0.5uH. But, even with very low parallel resistances, some things were usually still too-sensitive to the coil's value. (However, I was often probably pushing the design past where it should have gone, "just for fun".)
I don't know (and doubt) if any of that "means" anything, for any broader context. I just mentioned it because I noticed the now-familiar "0.5uH" value.
Wait. "Oxygen Free"?! How do they get oxygen out of copper?
Thanks, very much. I really appreciate that you even took the time to take a look. It's still just a work in progress. I have sort-of a love/hate relationship with it, and alternate between thinking "maybe someday it'll be pretty good" and thinking it's almost embarrassing.
Ah! So maybe I was under a bit of a wrong impression, about the seriousness of any tradeoffs. (I didn't do any homework. Sorry.)
Regarding the reported oscillation episode, it's probably frustrating, to say the least, not getting any information about the conditions for which it occurred. But, since they apparently became reticent after reporting it, maybe it was just something bone-headed, like out-->in.
Thanks for responding, John.
- Tom Gootee
http://www.fullnet.com/~tomg/index.html
John,
john curl said:
First, I probably should say that whatever (technical) information or opinions I post here usually shouldn't count for much, since I am, relative to most others here, grossly-inexperienced and am still mostly just trying to learn (or re-learn).
But the 0.5uH coil value that you mentioned reminded me that back when I was doing all of those spice amplifier simulations that I mentioned, it seemed like whenever I tried really-hard to max-out the performance (i.e. trying for "perfect" squarewave response with large and small capacitive loads, and THD-20 less than, say, .0001%), the optimal output-coil value (at least for the ones where I was determined to use a coil) often eventually converged to something around 0.5uH. But, even with very low parallel resistances, some things were usually still too-sensitive to the coil's value. (However, I was often probably pushing the design past where it should have gone, "just for fun".)
I don't know (and doubt) if any of that "means" anything, for any broader context. I just mentioned it because I noticed the now-familiar "0.5uH" value.
Wait. "Oxygen Free"?! How do they get oxygen out of copper?
john curl said:
Thanks, very much. I really appreciate that you even took the time to take a look. It's still just a work in progress. I have sort-of a love/hate relationship with it, and alternate between thinking "maybe someday it'll be pretty good" and thinking it's almost embarrassing.
john curl said:
Ah! So maybe I was under a bit of a wrong impression, about the seriousness of any tradeoffs. (I didn't do any homework. Sorry.)
Regarding the reported oscillation episode, it's probably frustrating, to say the least, not getting any information about the conditions for which it occurred. But, since they apparently became reticent after reporting it, maybe it was just something bone-headed, like out-->in.
Thanks for responding, John.
- Tom Gootee
http://www.fullnet.com/~tomg/index.html
Thanks for responding, Tom. 0.5uH is pretty much the default minimum value mentioned on this website and is used by Bob Cordell, as well. I didn't say much about coils, until someone mentioned that he was using 5.0uH or so. I still think that is way too much, since 2uH was the standard 30 years ago.
Moreover, you will find that several high end designers, who contribute to this website, do not use any coil at all in their latest designs. I thought they were 'over the top' 20 years ago, but I have since, changed my mind.
I find your website very useful.
Moreover, you will find that several high end designers, who contribute to this website, do not use any coil at all in their latest designs. I thought they were 'over the top' 20 years ago, but I have since, changed my mind.
I find your website very useful.
PB2 said:
Hi Pete,
(Your ASIC design stuff sounds coool!)
It depends. If it's just a quickie design, just "playing around", with no pcb layout yet, then I have to somehow just estimate what typical PCB trace widths and lengths might be, and use a corresponding impedance for them (often just the LR components). For i/o and power supply connections, I usually assume wires or cables for interconnections and use a small set of spice .PARAM statements for each one, so I can just set the lengths in inches and the impedances are automatically calculated and their values inserted.
Even though the trace and wire characteristics might be "just estimates", note that they can then be used to help determine, for example, which ones' lengths or thicknesses might be critical.
For inductors, I always include at LEAST the series resistance. For through-hole resistors, I usually just put something like 0.5pF in parallel with each one. For capacitors, it's often very important to include at least the ESR (equiv series R).
For film and ceramic caps, I try to get the datasheets, and look at the manufacturers' websites. If they only give the %dissipation factor, which is tan(delta) x 100 (where tan(delta) is the tangent of the loss angle), then I use tan(delta) = 2 x Pi x f x C x ESR and solve for ESR. If a max value for DF is all they give, I have to use that. Also, Kemet's website has software that will give good capacitor models, but not for electrolytics.
For electrolytic capacitors, they often only give ESR at one f, such as 100 kHz. But it varies significantly with f. Usually I can get lucky and find a figure for DF at some other f, on the manuf website, from which I can calculate the ESR for a second f. If so, for transient simulations at least, I can assume a linear ESR vs f (the best I can do without more data) and then use a spice .PARAM statement for each type and value of capacitor, to calculate the ESR for the freq param being used in a transient run. For leakage current, they usually say something like "I=.01CV or 3uA, whichever is greater". So I usually just use a spice .param statement to calculate Rleak=1/(.01C), for each capacitor type and value. I posted an example of a similar procedure for modeling an electrolytic's ESR and leakage, here: http://www.diyaudio.com/forums/showthread.php?s=&postid=1228677&highlight=#post1228677 .
For "AC Analysis" spice runs, the param statements to calculate ESR vs f cannot be used, unfortunately. But you can put a G-Laplace source is series with a capacitor, to get the ESR to vary with f, using something like the linear model for ESR vs f mentioned above. You can look in the message archive of the LT-SPICE group at http://www.yahoogroups.com , for a more-detailed answer about that, which was posted on June 22, 2007. Also, look in the group's Files area, mainly in the Tut section, to get many examples for modeling R, L, and C, with parasitics and with temperature dependence, monte carlo runs, sensitivity analysis, etc., etc., etc.
Also, I have posted a two-winding transformer model that calulates its L and R parameters based on simple RMS voltage measurements that the user can enter, which is at http://www.fullnet.com/~tomg/gooteesp.htm . But, it doesn't model the interwinding capacitance or the two intra-winding capacitances. However, if you can short both the primary and secondary and measure, somehow (another whole can of worms), the C between them (interwinding), and then short each one in turn and measure the other's C, which is actually measuring each intra-winding C in parallel with the interwinding C, then you can calculate each intra-winding C.
- Tom Gootee
http://www.fullnet.com/~tomg/index.html
comments
Thanks for the reply, I didn't go back and read all 57 pages before responding so sorry if my comments had already been addressed earlier.
"Thus we observe that adding negative feedback to a FET amplifier, while decreasing the overall amount of distortion, significantly changes the distribution of the distortion products. In two-tone tests, feedback introduces new tiers of products, most very weak, but not necessarily insignificant perceptually, as they produce a noise floor correlated with the program material. And not only does the amplitude of this noise floor rise and fall with the amplitude of the program material, but its character changes as it rises and falls, higher order products being more volatile than lower-order ones."
I found this comment from boyk and sussman interesting though as it is similar to what Crowhearst pointed out (for tube amps I might add)
Actually they do look at individual harmonics and note in particular that higher order harmonics are affected more than lower order ones.
They say, "If we increase the drive of the FET amplifier by 12 dB, from 0.05 peak volts to 0.2 peak volts, the distortion at frequency 2 rises by 24 dB...But the components at frequencies 17 and 19 increase by 60 dB...As the level rises, the distortion is thus weighted toward
the higher-order terms, and mostly toward higher frequencies"
It is not clear to me why three levels of feedback are not adequate, you have the zero case, the extreme high case and a point in the middle. In any case for FETS they show that the distortion is both feedback AND level dependent.
For BJTs: "The nonlinearity of a BJT is extremely sharp, so we drive the input with a peak voltage of only 0.0004 Volts in each frequency. If we used 0.05 Volts, as with the FETs, the distortion of the no-feedback BJT amplifier would be very bad and we would not be able to understand the effect of the feedback"
"The spectrum of d, the BJT amplifier without feedback (figure 7), is complicated, and remarkably similar to the spectrum of the FET amplifier with feedback, showing several tiers of distortion products."
" The spectrum of e, the BJT amplifier with feedback (figure 8) is like
the spectrum without feedback except that all of the distortion products are substantially attenuated. When we increase the feedback, as in circuit f, we see (figure 9) that the performance of the amplifier again improves dramatically, so that the principal distortion lines are suppressed below −100 dB and the next tier is suppressed below −180 dB. Thus, in the case of the BJT,
feedback improves the nonlinear distortion of the amplifier."
So it appears at least in single ended circuits that BJTs behave differently to negative feedback than FETS as they are much more non-linear to begin with and all harmonics are already present.
However; Boyk and Sussman point out that the same shift in the distortion pattern towards higher harmonics occurs just like the case of FETS with feedback.
" Comparison of figure 9 and figure 10 illustrates another interesting point: For the BJT stage with or without feedback, as for the FET stage with feedback, higher-order distortion products are emphasized as the amplitude of the input signals is increased. This is generally true of any nonlinear system that shows tiers of intermodulation products."
The triode gives predictable results and gives similar distortion to BJT with a lot of feedback at higher input levels with fewer high order products.
What about for the output stage (of more importance to this discussion I believe)?
For a BJT output stage: " We introduce feedback by boosting the prescalar gain to 100, and feeding back 0.089 of the output signal. With this we get the spectrum of the output that appears in figure 16. As we would expect, the feedback improves the result by suppressing all of the distortion products. Whether or not we apply
feedback in this class-A amplifier the relative distortion decreases with the signal amplitude."
"However, the pair without feedback, compared to the single-ended BJT stage without feedback, generates frequencies 17, 25, 27 and 29 at levels much higher than expected. And the pair with feedback, compared to the single-ended BJT with feedback, generates 17, 19, 21, 23 and 25 at much higher levels than expected."
So, again high order odd harmonics are affected more than other harmonics.
For FETS: "In class A, the pair is distortion-free even without feedback, and feedback’s only role is to set gain."
"In class B, even-numbered distortion products are absent due to circuit symmetry; but frequencies 23 and 25 get worse with feedback! Also in class B with or without feedback, relative distortion goes up as input level goes down."
Triodes: "Even-numbered distortion products are absent due to circuit symmetry. Feedback suppresses all products."
For a differential input stage they find the following:
"Unless otherwise stated, (a) Higher input level raises the relative
distortion and (b) emphasizes higher-order distortion products. (c) Adding more feedback lowers all distortion products."
Their conclusions are also interesting but perhaps you find them too speculative?
If I can find the Crowhearst paper I will send it to you.
Thanks for the reply, I didn't go back and read all 57 pages before responding so sorry if my comments had already been addressed earlier.
"Thus we observe that adding negative feedback to a FET amplifier, while decreasing the overall amount of distortion, significantly changes the distribution of the distortion products. In two-tone tests, feedback introduces new tiers of products, most very weak, but not necessarily insignificant perceptually, as they produce a noise floor correlated with the program material. And not only does the amplitude of this noise floor rise and fall with the amplitude of the program material, but its character changes as it rises and falls, higher order products being more volatile than lower-order ones."
I found this comment from boyk and sussman interesting though as it is similar to what Crowhearst pointed out (for tube amps I might add)
Actually they do look at individual harmonics and note in particular that higher order harmonics are affected more than lower order ones.
They say, "If we increase the drive of the FET amplifier by 12 dB, from 0.05 peak volts to 0.2 peak volts, the distortion at frequency 2 rises by 24 dB...But the components at frequencies 17 and 19 increase by 60 dB...As the level rises, the distortion is thus weighted toward
the higher-order terms, and mostly toward higher frequencies"
It is not clear to me why three levels of feedback are not adequate, you have the zero case, the extreme high case and a point in the middle. In any case for FETS they show that the distortion is both feedback AND level dependent.
For BJTs: "The nonlinearity of a BJT is extremely sharp, so we drive the input with a peak voltage of only 0.0004 Volts in each frequency. If we used 0.05 Volts, as with the FETs, the distortion of the no-feedback BJT amplifier would be very bad and we would not be able to understand the effect of the feedback"
"The spectrum of d, the BJT amplifier without feedback (figure 7), is complicated, and remarkably similar to the spectrum of the FET amplifier with feedback, showing several tiers of distortion products."
" The spectrum of e, the BJT amplifier with feedback (figure 8) is like
the spectrum without feedback except that all of the distortion products are substantially attenuated. When we increase the feedback, as in circuit f, we see (figure 9) that the performance of the amplifier again improves dramatically, so that the principal distortion lines are suppressed below −100 dB and the next tier is suppressed below −180 dB. Thus, in the case of the BJT,
feedback improves the nonlinear distortion of the amplifier."
So it appears at least in single ended circuits that BJTs behave differently to negative feedback than FETS as they are much more non-linear to begin with and all harmonics are already present.
However; Boyk and Sussman point out that the same shift in the distortion pattern towards higher harmonics occurs just like the case of FETS with feedback.
" Comparison of figure 9 and figure 10 illustrates another interesting point: For the BJT stage with or without feedback, as for the FET stage with feedback, higher-order distortion products are emphasized as the amplitude of the input signals is increased. This is generally true of any nonlinear system that shows tiers of intermodulation products."
The triode gives predictable results and gives similar distortion to BJT with a lot of feedback at higher input levels with fewer high order products.
What about for the output stage (of more importance to this discussion I believe)?
For a BJT output stage: " We introduce feedback by boosting the prescalar gain to 100, and feeding back 0.089 of the output signal. With this we get the spectrum of the output that appears in figure 16. As we would expect, the feedback improves the result by suppressing all of the distortion products. Whether or not we apply
feedback in this class-A amplifier the relative distortion decreases with the signal amplitude."
"However, the pair without feedback, compared to the single-ended BJT stage without feedback, generates frequencies 17, 25, 27 and 29 at levels much higher than expected. And the pair with feedback, compared to the single-ended BJT with feedback, generates 17, 19, 21, 23 and 25 at much higher levels than expected."
So, again high order odd harmonics are affected more than other harmonics.
For FETS: "In class A, the pair is distortion-free even without feedback, and feedback’s only role is to set gain."
"In class B, even-numbered distortion products are absent due to circuit symmetry; but frequencies 23 and 25 get worse with feedback! Also in class B with or without feedback, relative distortion goes up as input level goes down."
Triodes: "Even-numbered distortion products are absent due to circuit symmetry. Feedback suppresses all products."
For a differential input stage they find the following:
"Unless otherwise stated, (a) Higher input level raises the relative
distortion and (b) emphasizes higher-order distortion products. (c) Adding more feedback lowers all distortion products."
Their conclusions are also interesting but perhaps you find them too speculative?
If I can find the Crowhearst paper I will send it to you.
Crowhurst link
Here is the Crowhurst article (I spelled his name wrong earlier, my mistake).
http://web.mit.edu/cheever/www/crow1.htm
Note it is up on Cheever's website so it is not a pdf, this is just the first page link but the rest is there.
Here is the Crowhurst article (I spelled his name wrong earlier, my mistake).
http://web.mit.edu/cheever/www/crow1.htm
Note it is up on Cheever's website so it is not a pdf, this is just the first page link but the rest is there.
Re: comments
Yes, the rising higher-order harmonics with increasing feedback were also noted by Baxandall, and I don't think anybody is disputing that. In fact, that was one of the things Bob Cordell did to validate his idea of using SPICE to do these simulations. His plots of the JFET amp are shown here. This shows a similar thing. Click the graph to show the results.
I do want to mention something about the variation of distortion with signal level though. There is a region of operation of circuits called "weak nonlinearity", which has certain properties associated with it. These properties hold regardless of whether there is feedback around them or not. They come from the representation of the nonlinear transfer function as a Taylor series (power series) expansion. If you have such a representation for the circuit and assume that the input is a sine wave and the output distorts, but not by an extreme amount (no clipping and negligible compression/expansion for example), you get the following results.
Let's say you increase the input by 1 dB. The following will happen with a weakly nonlinear circuit:
1) The second harmonic will increase by 2 dB
2) The third harmonic will increase by 3 dB
3) The fourth harmonic will increase by 4 dB
etc.) The Nth harmonic will increase by N dB
This is nothing new or revolutionary, and has nothing to do with feedback. It's simply a property of weakly nonlinear circuits. So, you can imagine that when you crank up the signal level, the higher-order harmonics will rise very fast. The higher, the faster.
See my explanation above. A 12 dB increase in input gives a 60 dB increase in the fifth-order product (this holds for IM too). This is common to weakly nonlinear circuits, feedback or not. Note that this represents a 48 dB change in the difference between the fifth-order product and the fundamental.
Have a look at Baxandall's article - part 6 where he takes data on the BJT circuit. Notice that there's some wacky variation of the third-order term with the amount of feedback when the feedback is low. The only way to catch this is by doing lots of points. In fact, earlier in the thread I was initially unable to duplicate this behavior properly in simulation. If I hadn't done many points the problem would not have been noticed.
Exactly. Notice how there's a difference in how the feedback affects harmonics when the amplifier without feedback has higher-order distortion. Right there, that is a counterexample to the "Feedback always increases higher-order harmonics" claim.
Exactly. Or, stated another way, "Each circuit must be considered individually".
Right. This is nothing more than the textbook behavior of weakly nonlinear circuits as I described above. Indeed, this behavior has been known to designers of military radar systems for at least 50 years. It's important there because the behavior determines how far down below clutter the signal level of the "target" can be seen - AKA "sub-clutter visibility". This is not some revolutionary audiophile discovery.
It's important that the results of a given circuit not be generalized to all circuits regarding how feedback affects them. Also, they present results with distortion going down to 200 dBc, yet the models they use are very primitive. The SPICE models take into account many more different types of errors than the models Boyk and Sussman use, yet the SPICE models fall apart at much higher levels than that. So their data can be considered conceptually useful, but one cannot expect to get actual numbers that are anywhere near that in real circuits. Same for the SPICE simulations too. I took my graphs down to 140 dBc but it's extremely unlikely that the data down there are accurate. The idea, like in Boyk and Sussman's analysis, is to show trends at those lower levels.
morricab said:
Yes, the rising higher-order harmonics with increasing feedback were also noted by Baxandall, and I don't think anybody is disputing that. In fact, that was one of the things Bob Cordell did to validate his idea of using SPICE to do these simulations. His plots of the JFET amp are shown here. This shows a similar thing. Click the graph to show the results.
I do want to mention something about the variation of distortion with signal level though. There is a region of operation of circuits called "weak nonlinearity", which has certain properties associated with it. These properties hold regardless of whether there is feedback around them or not. They come from the representation of the nonlinear transfer function as a Taylor series (power series) expansion. If you have such a representation for the circuit and assume that the input is a sine wave and the output distorts, but not by an extreme amount (no clipping and negligible compression/expansion for example), you get the following results.
Let's say you increase the input by 1 dB. The following will happen with a weakly nonlinear circuit:
1) The second harmonic will increase by 2 dB
2) The third harmonic will increase by 3 dB
3) The fourth harmonic will increase by 4 dB
etc.) The Nth harmonic will increase by N dB
This is nothing new or revolutionary, and has nothing to do with feedback. It's simply a property of weakly nonlinear circuits. So, you can imagine that when you crank up the signal level, the higher-order harmonics will rise very fast. The higher, the faster.
See my explanation above. A 12 dB increase in input gives a 60 dB increase in the fifth-order product (this holds for IM too). This is common to weakly nonlinear circuits, feedback or not. Note that this represents a 48 dB change in the difference between the fifth-order product and the fundamental.
Have a look at Baxandall's article - part 6 where he takes data on the BJT circuit. Notice that there's some wacky variation of the third-order term with the amount of feedback when the feedback is low. The only way to catch this is by doing lots of points. In fact, earlier in the thread I was initially unable to duplicate this behavior properly in simulation. If I hadn't done many points the problem would not have been noticed.
Exactly. Notice how there's a difference in how the feedback affects harmonics when the amplifier without feedback has higher-order distortion. Right there, that is a counterexample to the "Feedback always increases higher-order harmonics" claim.
Exactly. Or, stated another way, "Each circuit must be considered individually".
Right. This is nothing more than the textbook behavior of weakly nonlinear circuits as I described above. Indeed, this behavior has been known to designers of military radar systems for at least 50 years. It's important there because the behavior determines how far down below clutter the signal level of the "target" can be seen - AKA "sub-clutter visibility". This is not some revolutionary audiophile discovery
.It's important that the results of a given circuit not be generalized to all circuits regarding how feedback affects them. Also, they present results with distortion going down to 200 dBc, yet the models they use are very primitive. The SPICE models take into account many more different types of errors than the models Boyk and Sussman use, yet the SPICE models fall apart at much higher levels than that. So their data can be considered conceptually useful, but one cannot expect to get actual numbers that are anywhere near that in real circuits. Same for the SPICE simulations too. I took my graphs down to 140 dBc but it's extremely unlikely that the data down there are accurate. The idea, like in Boyk and Sussman's analysis, is to show trends at those lower levels.