[Rob Watts]

I define a transient as everything that is not steady state - and the fact that most transients in music does not have huge HF content is not relevant. The pertinent issue here is the reconstruction of the timing of non-steady state signals (transients!) and whether that timing depends upon program material (i.e. something happening in the future or the past that is signal related) - and it's clear that the only way too guarantee transient timing recovery that is totally independent is by an ideal sinc function interpolation filter (assuming the signal is ideally bandwidth limited).

My view on the timing of transients recovery issue is that any timing error (that is a timing of transients that is either too early or too late and constantly changes) irrespective of how small that error is, is audible and significant - and this is based upon two observations. Firstly, that going from half a million taps to one million with a WTA filter is easily audible - in spite of the fact that the timing error in terms of transient uncertainty is already very small, and doubling the tap length will only half the error, which is already tiny. Secondly, replacing the IIR filter from 16FS to 256FS with a WTA filter is audible - and here the timing uncertainty effect would be logically vanishingly small.

I suppose what I am trying to get across is that errors that are vanishingly small have important subjective consequences. This sentiment will profoundly annoy the sound science lobbyists; but carefully controlled listening tests do reveal that some errors that are converging to zero are indeed still audible. I am thinking here on the noise shaper listening tests, where 350 dB performance noise shapers had better depth perception than 330 dB noise shapers - and yet the amplitude non-linearity errors introduced by the 330 dB noise shaper is almost (but not quite) zero. Although I do not have numbers to show timing error, my current assumption is that any transient timing error that is constantly varying is likely to be audible, no matter how small that error is.

 

[bhobba]

A transient any faster than what is bandwidth limited to audible frequencies would be of a higher frequency than is audible. That's obvious, isn't it?

Read the link I posited previously - it has to do with what the band-limiting does. It is claimed by some that the higher frequencies when chopped off cause a sort of ringing in the audio band. There are a number of ways of handling it - one is to sample at higher rate above what you can hear - so high at 16 bits resolution is all just noise above it, so while it still rings its of no importance - 96k is all that's needed for most recordings. If the music doesn't peter out until above that what the MQA guys do is use a gentle filter of a very unusual type called a spline that makes sure its does. A simple example of such a filter is a triangle - see figure 6:
https://www.soundonsound.com/techniques/mqa-time-domain-accuracy-digital-audio-quality#para4

But it has other effects such as aliasing - however by analyzing the music it is ensured it is is below 16 bits so if you chop off the bottom bits there is no problem. A triangle is a simple spline - MQA analyses the music to determine what spline to use. In this way 99.9% of music at least are limited to 96k;. MQA is cagey on how it handles the few recordings it does not work on.

Rob can explain what he does not like about this method. I think he would prefer to just chop in off at 96k using his highly accurate sharp cutoff filters - spline filters do other things like muck around with timing, which in tum, according to some, mucks around with localisation and depth perception,

BTW using modern compression methods 96k can be lossless encoded at about the same rate as a normal CD or you can use bit stacking tricks MQA uses - Rob can comment on why he may not like that one.

Thanks
Bill

 

[Currawong]

I don't agree. The Fourier reconstruction of that transient contains all frequencies - and you're only throwing away those that are inaudible. I have some test tracks of two impulse signals varying deltas apart (down to 5 microseconds). I'll try to dig these out and pm them to you. You'll be surprised how you can hear above 20 kHz (based on your arguments above). I don't think we understand enough of how the ear/brain process sound to conclude exactly what's audible and what's not.

Maybe it's of low priority, but I've never liked the argument that we don't need to worry about it because "sharp transients don't occur in music". There's no law preventing anybody from recording an album of dirac delta functions - and of course, the classic follow-up, the square-wave album. You might not think it sounds very good, but then I don't think rap sounds very good An ADC -> DAC should be able to work perfectly (or as well as we can possibly make it) in all cases, even extreme limiting cases.

I've found that some DACs will alias very high frequencies into the audio band, so I'm not quite so sure about hearing above 20 kHz. I'll check out what you sent though.

Well that's weird. @Currawong - I just tried to send you a pm, but headfi says I'm not allowed to. I hope that's just a headfi bug and that none of my prior rantings have caused you to block me?

Anyway, here are those impulse test tracks: https://www.dropbox.com/sh/qmkr6429gqpxfhe/AABxRdfmfgkOxoyTdYBsqZmba?dl=0

There's a readme file in the folder.

Sorry, I had PMs switched off as I was getting flooded with requests for help with gear I have no experience with.

It is claimed by some that the higher frequencies when chopped off cause a sort of ringing in the audio band.

I think we are confusing two separate things here. I wasn't talking about what happens at the filter cut-off. That is something I don't fully understand yet and am not talking about. Nor was I talking about the importance of high-frequency info.

What I am talking about is, say, how a transient with a frequency response inside the sample rate of the audio band at any rate cannot contain frequency information outside that band. So you cannot have, say, a 16/44 square wave because, by its very nature, the transient of a square wave contains frequency information above 22.050 kHz. Likewise an impulse response, step function or triangular function. So to say that "Oh look, linear phase filters cause ringing before and after an [out of band] impulse response and this hurts the music" seems to be bull-crap to me (again, ignoring filter stop-band issues) because the square wave or impulse response is demonstrating a behaviour that does not exist in actual music.

 

[csglinux]

I define a transient as everything that is not steady state - and the fact that most transients in music does not have huge HF content is not relevant. The pertinent issue here is the reconstruction of the timing of non-steady state signals (transients!) and whether that timing depends upon program material (i.e. something happening in the future or the past that is signal related) - and it's clear that the only way too guarantee transient timing recovery that is totally independent is by an ideal sinc function interpolation filter (assuming the signal is ideally bandwidth limited).

My view on the timing of transients recovery issue is that any timing error (that is a timing of transients that is either too early or too late and constantly changes) irrespective of how small that error is, is audible and significant - and this is based upon two observations. Firstly, that going from half a million taps to one million with a WTA filter is easily audible - in spite of the fact that the timing error in terms of transient uncertainty is already very small, and doubling the tap length will only half the error, which is already tiny. Secondly, replacing the IIR filter from 16FS to 256FS with a WTA filter is audible - and here the timing uncertainty effect would be logically vanishingly small.

I suppose what I am trying to get across is that errors that are vanishingly small have important subjective consequences. This sentiment will profoundly annoy the sound science lobbyists; but carefully controlled listening tests do reveal that some errors that are converging to zero are indeed still audible. I am thinking here on the noise shaper listening tests, where 350 dB performance noise shapers had better depth perception than 330 dB noise shapers - and yet the amplitude non-linearity errors introduced by the 330 dB noise shaper is almost (but not quite) zero. Although I do not have numbers to show timing error, my current assumption is that any transient timing error that is constantly varying is likely to be audible, no matter how small that error is.

Hi Rob, I agree with all the above and appreciate your attention to these tiny details. I know you'd previously mentioned the importance of the ADC step. Do you have any current insights or predictions on what X would look like in the case of the following:

Heaviside step function -> Davina -> Dave 2 -> X ?

Any interpolation across x=0 is going to create ringing, right? So, will X ring or will X be somehow filtered/dissipated/damped?

P.S. I don't want to re-hash the argument of whether this might or might not exist in an actual musical recording. Those of you that don't believe in electronic/digital music, please just regard this as a hypothetical

 

[castleofargh]

I define a transient as everything that is not steady state - and the fact that most transients in music does not have huge HF content is not relevant. The pertinent issue here is the reconstruction of the timing of non-steady state signals (transients!) and whether that timing depends upon program material (i.e. something happening in the future or the past that is signal related) - and it's clear that the only way too guarantee transient timing recovery that is totally independent is by an ideal sinc function interpolation filter (assuming the signal is ideally bandwidth limited).

My view on the timing of transients recovery issue is that any timing error (that is a timing of transients that is either too early or too late and constantly changes) irrespective of how small that error is, is audible and significant - and this is based upon two observations. Firstly, that going from half a million taps to one million with a WTA filter is easily audible - in spite of the fact that the timing error in terms of transient uncertainty is already very small, and doubling the tap length will only half the error, which is already tiny. Secondly, replacing the IIR filter from 16FS to 256FS with a WTA filter is audible - and here the timing uncertainty effect would be logically vanishingly small.

I suppose what I am trying to get across is that errors that are vanishingly small have important subjective consequences. This sentiment will profoundly annoy the sound science lobbyists; but carefully controlled listening tests do reveal that some errors that are converging to zero are indeed still audible. I am thinking here on the noise shaper listening tests, where 350 dB performance noise shapers had better depth perception than 330 dB noise shapers - and yet the amplitude non-linearity errors introduced by the 330 dB noise shaper is almost (but not quite) zero. Although I do not have numbers to show timing error, my current assumption is that any transient timing error that is constantly varying is likely to be audible, no matter how small that error is.

not planning to be Sound Science ambassador as I'm merely the incompetent mall cop of the section(also if we're a lobby, where is the money?!!!!!!!!), but I at least have personal issues with some subjective stuff you say and their more or less intuitive implications on human hearing. about the objective part of your posts, I'm usually in agreement with you and honestly learned a good deal from you over the years. so more than ever, I find important to properly separate objective and subjective conversation.

I can believe in the possibility that changes in the settings of very extreme stuff, noise shapers or anything else, could end up having various consequences at the output signal. some maybe less expected than others depending on gear and processing.
and if some of those consequences manifest at audible levels, then of course I believe that you could hear them. presented that way there is nothing too strange about your anecdotes. except you're never telling them that way. instead you use those anecdotes to repeatedly suggest that human hearing goes way beyond what the actual knowledge on the subject would consider realistic. let's say something like this with simple stimuli for starters http://www2.bcs.rochester.edu/courses/crsinf/221/13.pdf and maybe this about masking which is so very relevant if you're going to make alusion at audible stuff down by 300dB http://www2.bcs.rochester.edu/courses/crsinf/221/14.pdf

if you actually have an experience demonstrating the need to revise the boundaries of what we consider human hearing, then obviously you should describe it exhaustively so that other people would point out your mistakes, or maybe replicate your findings and help the facts be clearly established. instead, you suggesting extraordinary stuff in a perfectly casual way, makes me think that you're talking from sighted impressions(and then anything goes), or that you're only doing this for marketing effect. in which case, can you really blame me for thinking that the all super hearing idea is making as much sense as the scenario of Sharknado3?


about sampling and reconstruction, now I'm with you and math. although I have little care for extreme tap length because it's a lot of efforts for very small improvement. at the same time, so long as it's not detrimental to other stuff, I can't really complain about you trying to increase accuracy ^_^. so as I said, on the objective side of things, we're unsurprisingly good. in general, clean cut for higher fidelity within the band limited range and as little out of band crap as possible, that's pretty rational to me.
on the other hand, most brands advocating that good sound is a good looking Dirac pulse are IMO willingly misleading consumers with that visual trickery. it's the same abuse of intuitive but false demonstration that has been used so often by showing a staircase signal in a graph. when I see those stuff my trust in the brand instantly drops to zero and I move into my "skeptical whiner" mode AKA Monday morning.

 

[zettelsm]

not planning to be Sound Science ambassador as I'm merely the incompetent mall cop of the section(also if we're a lobby, where is the money?!!!!!!!!), but I at least have personal issues with some subjective stuff you say and their more or less intuitive implications on human hearing. about the objective part of your posts, I'm usually in agreement with you and honestly learned a good deal from you over the years. so more than ever, I find important to properly separate objective and subjective conversation.

I can believe in the possibility that changes in the settings of very extreme stuff, noise shapers or anything else, could end up having various consequences at the output signal. some maybe less expected than others depending on gear and processing.
and if some of those consequences manifest at audible levels, then of course I believe that you could hear them. presented that way there is nothing too strange about your anecdotes. except you're never telling them that way. instead you use those anecdotes to repeatedly suggest that human hearing goes way beyond what the actual knowledge on the subject would consider realistic. let's say something like this with simple stimuli for starters http://www2.bcs.rochester.edu/courses/crsinf/221/13.pdf and maybe this about masking which is so very relevant if you're going to make alusion at audible stuff down by 300dB http://www2.bcs.rochester.edu/courses/crsinf/221/14.pdf

if you actually have an experience demonstrating the need to revise the boundaries of what we consider human hearing, then obviously you should describe it exhaustively so that other people would point out your mistakes, or maybe replicate your findings and help the facts be clearly established. instead, you suggesting extraordinary stuff in a perfectly casual way, makes me think that you're talking from sighted impressions(and then anything goes), or that you're only doing this for marketing effect. in which case, can you really blame me for thinking that the all super hearing idea is making as much sense as the scenario of Sharknado3?.

So, Mr. Mall Cop, I invite your attention to:

https://m.phys.org/news/2013-02-human-fourier-uncertainty-principle.html

Which is only to point out that the boundaries of what is accepted as the limits of human hearing are indeed changing, and this article is already 5 years old. Neuroscience is far from a mature science, particularly where the perceptual mechanisms of the sensory input/brain interface are concerned. If the guy who designed the best, most realsitic sounding DAC I have ever heard says his listening tests have shown him that performance at the 330-350 dB down level is audible and have informed his work, then I have to ask who am I to question the veracity of his statements, let alone suggest as you do that he is either mistaken because he isn’t doing double-blind testing to prove it to you, or worse that he is shilling for Chord’s marketing department.

Really, after witnessing your seeming inability to understand the concept of a pixel or sampling theory, I have to wonder if you understand this stuff as much as you think you do.

If this sounds rough, then I suggest you go back and re-read your third paragraph above.

Steve Z

 

[Rob Watts]

I don't want to put my moderators hat on - but please let's keep this thread civilised. This should be a safe place to talk, even when there are strongly held differences of opinion, so please concentrate on the facts not the personalities involved. There have been a few posts recently that has given me cause for concern...

 

[Arpiben]

So, Mr. Mall Cop, I invite your attention to:

https://m.phys.org/news/2013-02-human-fourier-uncertainty-principle.html

Which is only to point out that the boundaries of what is accepted as the limits of human hearing are indeed changing, and this article is already 5 years old. Neuroscience is far from a mature science, particularly where the perceptual mechanisms of the sensory input/brain interface are concerned...

Steve Z

A complementary information about Fourier uncertainty.

The information available from Fourier analysis is bound by an uncertainty relation called the Gabor limit. This says that you cannot know the timing of a sound and its frequency – or pitch – beyond a certain degree of accuracy. The more accurate the measurement of the timing of a sound, the less accurate the measurement of its pitch and vice versa.
Unlike the Heisenberg uncertainty principle, the Gabor limit is not an intrinsic property of the signal but is a result of the method used to analyse it. If you can find a way to analyse a complex waveform without decomposing it into sine waves, you can in theory track the frequency at a particular time to much greater accuracy. However, whatever analytical technique you choose must be nonlinear because any technique that represents the waveform as a sum of simpler waveforms will be bound by the Gabor limit.

Your linked paper says that human hearing can beat the Gabor limit.
It doesn't say we can not measure it.

Hope it helps among all those uncertainties.

 

[Triode User]

withdrawn

 

[Rob Watts]

Hi Rob, I agree with all the above and appreciate your attention to these tiny details. I know you'd previously mentioned the importance of the ADC step. Do you have any current insights or predictions on what X would look like in the case of the following:

Heaviside step function -> Davina -> Dave 2 -> X ?

Any interpolation across x=0 is going to create ringing, right? So, will X ring or will X be somehow filtered/dissipated/damped?

P.S. I don't want to re-hash the argument of whether this might or might not exist in an actual musical recording. Those of you that don't believe in electronic/digital music, please just regard this as a hypothetical

Of course a signal that has a discontinuity will never appear in real life using acoustic instruments, and would not meet the requirements of being bandwidth limited for sampling theory. This will be taken care of with the Davina project decimation filter - and I have two filters designed. As too the DAC side, with a M scaler and Dave, there will be no extra ringing whatsoever, as it will reconstruct the un-sampled signal to a better than 16 bit accuracy under all conditions. But we still have the ringing in the decimation filters to worry about.

Now sampling theory is very clear - you must use a sinc function filter to perfectly reconstruct a perfectly bandwidth limited sampled signal; but sampling theory says nothing about how that bandwidth limiting occurs, it just states that at FS/2 and greater there must be zero signal energy. I have designed two filters - an IIR type, and an FIR type, both with >250 dB rejection at FS/2 and above. The IIR type will have no pre-ringing, but is phase non-linear. The FIR type is phase linear, but has lots of pre-ringing. Which one will sound more like no filtering at all? I will know when the listening tests starts, and we will be able to answer this question completely. My guess is that bandwidth limiting with FIR will be close to completely transparent, as I see the pre-ringing issue at 22.05 kHz as a complete red-herring. The filters have been designed to bandwidth limit but preserve the 705 kHz sample rate, so one will hear only the effect of bandwidth limiting.

The test files will be published too, so people can listen for themselves on a 768 kHz capable DAC.

 

[dac64]

withdrawn

It's time for a good breakfast and read!

 

[Rob Watts]

Here are the plots from the decimation filter using actual filter data from a Verilog simulation:

This is for 768 kHz, and aliasing is guaranteed to be below -250 dB.
The transition zoomed in is:

So again at FS/2 (24 kHz) we have 250 dB rejection. Conventional decimation filters are half band, which means at 24 kHz it would be -6dB. To convert to 22.05 it is just a linear change of scale.

 

[Triode User]

It's time for a good breakfast and read!

I meant comment withdrawn ie deleted.

 

[castleofargh]

warning, this is a massive #mylife post. and most of it ends up being about them pixels. so don't feel bad for not reading, you're not missing much.

So, Mr. Mall Cop, I invite your attention to:

https://m.phys.org/news/2013-02-human-fourier-uncertainty-principle.html

Which is only to point out that the boundaries of what is accepted as the limits of human hearing are indeed changing, and this article is already 5 years old. Neuroscience is far from a mature science, particularly where the perceptual mechanisms of the sensory input/brain interface are concerned. If the guy who designed the best, most realsitic sounding DAC I have ever heard says his listening tests have shown him that performance at the 330-350 dB down level is audible and have informed his work, then I have to ask who am I to question the veracity of his statements, let alone suggest as you do that he is either mistaken because he isn’t doing double-blind testing to prove it to you, or worse that he is shilling for Chord’s marketing department.

Really, after witnessing your seeming inability to understand the concept of a pixel or sampling theory, I have to wonder if you understand this stuff as much as you think you do.

If this sounds rough, then I suggest you go back and re-read your third paragraph above.

Steve Z

to answer everything in one word: context.

Spoiler: and to answer in too many words

doing slightly better than previously tested because of a change in the testing conditions themselves, it's not out of question. 0dB SPL was defined as the threshold of audibility for a single tone in the midrange. turns out many kids can perceive below that. but about 10dB below, the margin of error was pretty much always allowing for people to ear things at that level(so long as they're in an anechoic chamber and no other louder sound is played!!!!).
things are constantly made more accurate of course, including tests. but it won't be discovered that we can jump up 500meters at sea level gravity with only the help of our legs. no test is going to discover that we can in fact see the mouse moving in the grass on the mountain a few kilometers away with the naked eye. and no test is going to confirm that we can notice changes 300dB below music(that would be more than superman level of hearing). there is a difference between keeping an open mind, and accepting impossible stuff as being facts. magnitudes here are very relevant and a log unit perhaps fail to convey the true enormity of a number like 300dB. just the idea of having music playing at the same time renders the possibility of noticing 0dB SPL pretty difficult and borderline wishful because of masking, because of the increased damping of the ear in presence of louder sounds(making it lose some sensitivity). and stepping out of the anechoic chamber and into the real world, is also going to make our hearing much less impressive that what it may look like under nominal conditions. so consider for a second 300dB below music. even if we push the music to 120dB where it may start to be painful and damaging for the ear, we're still talking -180dB SPL ^_^. we don't have anything with the ability to resolve that accurately in a playback chain. thermal noise from components is going to be often be louder than that. we can have some fun in the digital domain at 64bit, but it's back to fantasy magnitudes once we're back to the analog domain with real gear. I mean no DAC I know of, resolves full 24bit, so I don't feel like what I'm saying is anything new here. this is all pretty much the consensus on the matter AFAIK.
as for believing something because of who said it, we all do that at various levels, but of course a fact doesn't care about who's saying it. the quality of the demonstration and the repeatability of the results should be what makes it relevant. no the guy saying it.


as for having fun about pixels and sampling, again, context. I was answering about the relevance of a specific analogy and considering real life application. while my new best friend decided that the only relevant story was about repeating "pIXelS AreN't SquAReS!" in loop to pretend that I didn't even have the ability to understand that, even after being told a few times. also he cherry picking an isolated situation with totally specific conditions nobody uses in photography, just to support yet another captain obvious point, that sampled data follow sampling theory. well what can I say, he's right both times obviously. bravo captain obvious. but at no point that makes the analogy relevant. you don't make an analogy relevant by disregarding the 10 stuff that are different just for the sake of saying that we can apply the same math on 2 sets of sampled data. which again I wasn't trying to contest.
if it makes everybody happy to think that I'm an idiot, good for you all, but there are enough situations where I'm saying stupid stuff on my own, not to need to force fake ideas into my mouth.
I wrote something stupid and admitted it, I'm not sure why I wrote that in the first place but hey! I still did it. then he got in my head and to make things worst I formulated something poorly again in the next post. and again I admitted it. so maybe if I insist on something else, there is a reason beyond my ego? I took the care of saying that I was trying to make a specific point and that it wasn't necessary to mention the 20 ways sampled data could be captured and processed following sampling theory(to no avail, obviously as it's all he talked about, that and pixels aren't squares. and it absolutely got in my head. I very much lost it and my posts don't come close to portray how pissed I was reading that crap repeated like a broken record. I'm not proud but it's true. which led me to dig my own grave and even end up writing "pixel size" instead of photosite, resolution or whatever concept of paired lines per millimeter(the autocorrect didn't want photosite and at the time I was way too mad to stop and go google to make sure it's also used in English). so once more I proved my ability to fail, by posting a reply while angry, which is by far the dumbest thing I've done on this topic.
so really if the deep desire of someone is to point a finger at me and say "ah ah!", the opportunities are everywhere. and if you follow my posts, I humiliate myself pretty much on a weekly basis(it's fun because it's true). does it mean that the analogy between resizing a picture and oversampling audio is a valid one? no it's not! which was really the only thing I was trying to defend all along.

I shouldn't have gotten mad, I should have looked for the proper translation of technical terms instead of reformulating my sentences on the fly with what felt "close enough, I don't give a F anyway". I shouldn't have posted anything while mad. that's really all this is about.
it just happened to hit really too close to home to leave me unaffected. because treating me like a 5year old who never thought about pixels, sampling theory, digital cameras or processing, when you've had my life, it's really maddening. as a teen I was a nolife gamer for a few years obsessed with specs and tech and how to get that one extra frame, pixel, ms. to the point where I ended modo of the tech section in the main French forum of Enemy Territory( FPS game where I wasted too many hours). I was the nolife geek people would query on IRC about monitors and video configurations(and mice and DLLs to change the USB refresh rate and the quake engine and...). not something where it would have ever occurred to me to think about pixels, aliasing or anything of the sort, right? and I was just a teen at the time. at school I was on a math and physic curriculum until I was 20, why would I have ever hear of sampling theory, Fourier transform and all the fun of considering sampled data as sums of sines? I mean it not like I learned that at school... oh wait! but you want the real funny part? at 20 I abandoned all that for my passion, photography...
I got in a photo school full time for 3 years, and you've guessed it, why would we ever learn or discussed anything about digital processing... oh wait! after that I spent about 10 years trying to become the geekiest geek at photoshop, to the point where I once ended up telling the guy doing the demo on the latest version in Paris, which shortcuts to use so he would stop wasting everybody's time with the old stuff and stayed with him for the rest of the demo.
and I don't say all that to brag, I'm not sure I ever mentioned those aspects of my life on Headfi before, and I'm not convinced it's anything to be proud of(at best I was a photo and post processing nerd). I'm saying all this hoping that you can put yourself in my shoes when a random dude on the web started talking to me like my all life didn't happen, and like he knew what I was thinking better that I did. oh the fury!
well now I'm passed it and see how silly and futile it's been. I'm not even mad at Jawed, as for anything he did, I went and reacted to it like a dumbass.

years have passed since school(too many for my taste), and I certainly forgot a great deal about a great deal of things. the other day I had to think hard for about 20seconds on how to do a long division by hand^_^. but not enough to ever think that pixels are squares or that sampled data isn't sampled data. there isn't even any need for any actual knowledge to get that captain obvious level of argument.
so yes we can take a digital picture and oversample the data if we like, then go back to the previous dimensions(if I say "pixel dimensions", you think he'll come back to say that pixels aren't square?), and really not lose much of anything. the same way we can do it with an audio signal. but of course pretty much nobody nowadays actually uses that to interpolate a picture. the judicious occasions to increase the image size and then reduce it again are really limited outside of some automation in filters. and the opposite like jawed assumed as his default situation for the sake making a strawman point about perfect reconstruction, is something no average photograph would do to his workflow because it would mean losing information for no reason. once again, context. it should have mattered.
then there is the little detail where a picture will have to be considered as having more dimensions than sound(not really an issue with Fourier, we just go with spacial frequency or whatever term is used in English). or we could even make yet another fake situation and consider all the pixels lined up one rank after the other. the problem here is that the oversampling, if anybody cared to do it, would occur only horizontally, or only vertically. so even while trying to find circumstances to make it work, the logical conclusion is that the analogy is wrong.
also how the sum of sines showing the function of sampled audio, happens to be the analog audio signal itself. that's not a small detail. something you can't transpose to the censor of a camera, where the analog voltage of each cell is already pretty much the sampled data, and as such we don't really have anything to "reconstruct" anything, just get a RGB value for the areas where the pixels(pixels aren't squares!!!) are.
then there is the issue of having specific photosite layout per color channel, so any forced analogy should at least limit the context to one color channel at a time.
then of course we have to disregard all the cosmetic choices taking precedence over objective fidelity in photography. including for interpolation! it's been many years since the most used option stopped being basic oversampling. the interpolations typically used are very much destructive like I said at the beginning of that photography mess. and it's logical because nobody cares about having a reversible operation at this point, people who change the dimension of their pictures, do it with subjective visual result as number one priority. the best options to increase size, often aren't the best to decrease image size(pixels aren't squares!).
I'm most likely forgetting a few other reasons why making an analogy between audio oversampling and enlarging a picture is a very bad idea, but as I got dragged into this once again (when I though I was out, they pull me back in!), I hope I've given enough, not to keep saying that it's a valid analogy. TBH analogies in science and engineering are almost always a bad idea. I can never understand why people feel the need to hang on to them once various flaws and limitations have clearly been exposed.

 

[HeeBroG]

@robwatts

Hi Rob,

DAVE has selectable "Crossfeed" to emulate what we hear in 2ch audio with speakers which I and many find more natural.
This is especially so on recordings where instruments are unnaturally panned hard left or right which clearly does not happen in real life.

Yet there are several companies eg. Dirac/Acourate/HomeAudioFidelity etc. who offer DSP convolution filters to suppress inherent stereo crosstalk in 2ch audio to help improve imaging and soundstage/depth.
It seems some in 2ch audio are working at cross purposes with the head-fi community.
Can both sides be correct?

DSP however can't eliminate stereo crosstalk completely and the signal to the DAC is no longer bit perfect.
How does this affect the function of DAVE/M scaler in a 2ch system if implemented?