Why decrease in frequency, increase in distortion?

[David Hess]

What's the thermal time constant of the IC and heat sink? I would have thought it's tens of seconds, possibly even minutes.

The thermal time constant that matters is transmission across the silicon die itself, which can be 100s of microseconds to at least milliseconds.

Thermal feedback might be an issue at very low frequencies and at high power levels, but the graph of TDA7293, shows the THD+N plots for 100mW out and 50W are essentially the same, indicating thermal feedback is insignificant from 20Hz to 20kHz.

That is interesting because there should be a difference.  Since it is a low frequency effect, I doubt it would ever be audible because low frequency transducers tend to have the worst distortion performance anyway, which is why I prefer well tuned base-reflex speakers.

@David Hess, why the open loop decreases by the thermal feedback? BJTs increase the Hfe by temperature.
I am more convinced by the explanation given by @srd1954. The LM3886 has four input transistors, if the thermal wave increases the Hfe of one of them more then the other 3, there is an unbalance. The designer must have privileged the feedback to be negative.
I will harvest the thermal ac component from the mute circuit to inject it in either inputs .

Changes in hfe will affect input bias current creating an error with low frequency imbalance of the input impedances.  In some audio designs where this is a problem, like where input capacitance is high and has common mode variation, the input impedance into the inverting and non-inverting inputs is matched to turn this into a common mode error which will be rejected.

I think the error comes from the change in Vbe with temperature.  It is very apparent with operational amplifiers where precision parts rely on careful symmetrical layouts and low power operation.

Thermal feedback also has a major effect on operational amplifier settling time, and is why settling time to high precision has very little to do with bandwidth.  It does not affect fast but low open loop gain parts because they never settle to high precision.  So if you have settling time data, this can be used to estimate thermal effects, but pole-zero mismatch, if this method of compensation is used, also has a major effect on settling time, which is another way of saying not to rely on your OP-27 type of operational amplifier for fast settling time.

For a long time, discrete or hybrid construction to physically isolate the input stage was used for high open loop gain and fast settling applications.  This method still provides the highest performance in these areas, and there are some applications which cannot use a fully integrated design without significantly compromising performance.  Jim Williams designed a high bandwidth RMS calibrator where this was an issue and the input stage had to use separate discrete parts, although I am not sure if this was due to thermal feedback.

[mzzj]

If not thermal distortion it could be input capacitor distortion if electrolytic caps were used in test circuit

[mawyatt]

What's the thermal time constant of the IC and heat sink? I would have thought it's tens of seconds, possibly even minutes.

Thermal feedback might be an issue at very low frequencies and at high power levels, but the graph of TDA7293, shows the THD+N plots for 100mW out and 50W are essentially the same, indicating thermal feedback is insignificant from 20Hz to 20kHz.


Note the x-axis is in kHz.

Back (few decades) when we were designing silicon ICs in SiGe BiCMOS the thermal time constant was a few milliseconds for adjacent IC components, obviously longer for more distant components. We had extensive device models that included 1st and 2nd order local thermal effects which were also influenced by proximity and layout (which required multiple post layout re-simulations to study such). These can be non-linear effects such as having the waveform envelope affect results due to thermal influence (we were involved with some very fast precision use cases).

Recall an 70s IEEE paper by Solomon ? than actually showed the original 741 op-amp input differential pair reversing polarity (swapping the inverting and non-inverting input) due to thermal feedback from the output, quite interesting!! Of course this wasn't normally "seen" in use as this reversing input effect was only occurring around a few 10s of microvolts of input and totally swamped by normal use high level inputs.

Best

[Kokoriantz]

At low frequencies, the temperature undergoes 4 cycles of hot and cold per signal cycle. This variation becomes less pronounced as the frequency increases. The LTPs in lm3886 have 1.1k emitters degenerated but not in TDA, if Vbe variation is a problem, the LM would be less pronounced distortion.

[2N3055]

In addition to what David said about hfe thermal  modulation, with AB class amps there can also be a modulation of idle bias current, having influence on crossover region. That can happen because there is time lag between power transistors heating up and temp compensation circuit.  Also, the pulsating thermal changes propagate over the crystal as basically acoustic waves (phonons related thermal transfer) influencing input stages with periodic thermal modulation, that lags in phase and can affect actual components areas in a way that can create modulation/distortions. Depending on schematics.

Douglas Self mentions it...
http://www.douglas-self.com/ampins/thermald/thermald.htm
There is a typo ( in the book also), IC is supposed to be TDA1552Q. TDA1522 is tape head preamp.

But also LM3886 does not have single pure THD plot in a datasheet I looked into. They are all THD+N.
So 1/f contribution is included like it was mentioned before.
Frequency related distortion and noise floor should be independently measured to actually know contribution.

As for bad implementations, since LM3886 protection circuits are fast and work very well, sometimes people mount it with inadequate cooling. It will work, for years if need be. But SpIKe will enforce thermally dictated SOA and there will be distortions in funny places...

[David Hess]

If not thermal distortion it could be input capacitor distortion if electrolytic caps were used in test circuit

Oh, that could definitely do it.  Electrolytic coupling capacitors should have at least 10 times the capacitance needed for the 3dB low frequency cutoff frequency and should never be used in high pass applications where distortion matters.  Otherwise their voltage coefficient of capacitance produces considerable distortion.  High dielectric constant ceramic capacitors have the same problem.

Bob Cordell:

The electrolytic capacitor is problematic because it is in the signal path. Electrolytic
capacitors can be notoriously nonlinear and will cause distortion. Any signal voltage
appearing across the electrolytic will be distorted. At minimum, capacitance is nonlin-
ear with voltage across the capacitor and ESR is nonlinear with current through the
capacitor. Distortion from the capacitor rises at lower frequencies as a larger signal volt-
age appears across the capacitor. ...

We will see that the single biggest reason for employing a DC servo is the elimina-
tion of the electrolytic capacitor. The DC servo is a smaller and less expensive solution
that provides much higher sound quality and performance.

Douglas Self:

Standard aluminium electrolytics create distortion
when they are used for coupling and DC blocking
while driving a significant resistive load. Figure 11.13
is the test circuit; Figure 11.14 shows the resulting
distortion for a 47 mF 25 V capacitor driving þ20
dBm (7.75V rms) into a 680U load, while Figure 11.15 shows how the associated LF roll-off has
barely begun. The distortion is a mixture of second
and third harmonics, and rises rapidly as frequency
falls, at something between 12 and 18 dB/octave.
The great danger of this mechanism is that serious
distortion begins while the response roll-off is barely
detectable; here the THD reaches 0.01% when the
response has only fallen by 0.2 dB. The voltage across
the capacitor is 2.6 V peak, and this voltage is a better
warning of danger than the degree of roll-off.
Further tests showed that the distortion roughly
triples as the applied voltage doubles; this factor
seems to vary somewhat between different capacitor
rated voltages.
The mechanism by which capacitors generate this
distortion is unclear. Dielectric absorption appears to
be ruled out as this is invariably (and therefore presum-
ably successfully) modelled by adding linear compo-
nents, in the shape of resistors and capacitors, to the
basic capacitor model. Reverse-biasing is not the
problem, for capacitors DC biased by up to þ15 V
show slightly increased, not reduced distortion. Non-
polarised electrolytics show the same effect but at
a much greater AC voltage, typically giving the same
distortion at one-tenth the frequency of a conventional
capacitor with the same time-constant; the cost and
size of these components generally rule out their use
to combat this effect. Usually the best solution is
simply to keep increasing the capacitor value until the
LF distortion rise disappears off the left of the THD
graph. Negligible roll-off in the audio band is not a suffi-
cient criterion.

[Zero999]

I don't doubt thermal effects, but the graphs on the data sheets for the LM3886 and TDA7293 do not appear to show it's the main reason for the low frequency distortion. The LM3886 data sheet also shows distortion at output powers of 30W and 60W and the curves aren't different enough to account for that.
https://www.ti.com/lit/ds/symlink/lm3886.pdf

There isn't enough information on the data sheets to indicate whether it's thermal, due to capacitors or is simply 1/f noise.

They don't give any spectral data, so the entire thread is pure speculation.