Why DSD quantization errors are always correlated with the signal

Since the term “uncorrelated DSD quantization/modulation noise/error” keeps coming up in discussions, I thought it would be useful to put this issue to rest and demonstrate there is no such thing. The question we want to answer is: can we design a 1-bit quantizer (or modulator) so that the quantization error of the two-level output signal is not correlated with the continuous input signal? According to the signal theory, the answer is “no”. This is the mathematical proof.

First, a couple of definitions.

  1. The average power of a digital signal S over N consecutive samples is defined as:

image

where:

  • N is the number of consecutive samples
  • s[n] is the signal’s n-th sample

That is, the average power over N samples is the average of the squared magnitude of those samples. To find the total power of a signal, all we need to do is take more and more samples (i.e. increase N) so that we cover more and more of the time domain and compute the limit of the above formula when N goes to infinity. If our digital signals are well-behaved and have a finite Fourier transform, then, according to Parseval, the average power will converge, so the total power is well-defined.

For a 1-bit signal, there are only two possible sample values, +1 and -1, so no matter now many samples we take, the average power - and thus the total power - is always 1.

  1. The quantization error of a signal that undergoes quantization is define as:

E = Q - C,

where:

  • C is the continuous signal that enters the quantizer
  • Q is the discrete output signal of the quantizer

We can rewrite that as:

Q = C + E

That is, the output of the quantizer is the original continuous signal plus the quantization error.

Before getting into the proof, we need one theorem:

Given two uncorrelated signals A and B, the power of the sum of those signals is the sum of the individual signal powers. That is:

P(A + B) = P(A) + P(B)

To demonstrate that no 1-bit (two-level) quantization/modulation can produce uncorrelated errors, I will use reductio ad absurdum, or proof by contradiction. So, let’s assume that:

We can design a modulator that takes the continuous signal C and produces the output Q so that the error E is uncorrelated with C.

Then, according to the two formulas above:

P(Q) = P(C + E) = P(C) + P(E)

Since Q is a two-level signal, its power is always 1, so we can write:

1 = P(C) + P(E)

P(E) = 1 - P(C)

The last formula shows a clear correlation between the power of E and the power of C. That contradicts our assumption that E and C are uncorrelated, so the assumption must be incorrect. This concludes the demonstration.

To summarize in less formal terms, if we assume that a modulator completely decouples the quantization noise from the signal, we find that the input signal’s power modulates the quantization error’s power, which is a contradiction. This is inherent to the nature of two-level signals, regardless of what modulator is used.

Don’t get me wrong, I’m not saying that DSD modulators are not useful (e.g. during rendering). Due to noise shaping, the error in the 0 - 20kHz band can be brought well under audibility, even if it’s in theory still correlated with the signal. My point is simply that the phrase “uncorrelated DSD quantization error” is wrong and should not be used.

As usual, comments are most welcome.

You are thinking about way oversimplified modulator implementation and that’s what makes your assumptions wrong.

You are sort of equivalent trying to prove that it is impossible to create encryption algorithm by using over simplified logic.

Output of (proper) encryption algorithm is uncorrelated with the input data. ROT-13 is not proper encryption algorithm.

I think I made it clear that I’m not assuming anything about the modulator. It’s all about the two-level signal.

Number of “levels” doesn’t matter. 1-bit stream can also be multi-level, etc etc. What matters is how the data came to be.

I’m not talking about 1-bit streams, I’m talking about a two-level digital signal. That’s exactly what DSD is. The fact that there are only two levels is the very issue.

Both wrong…

ESS DAC chips are also two level, Mola-Mola is two level. The multi-level implementations invariably have worse performance.

By the way, remember that for example class-D amplifiers also operate on two levels. Analog or digital.

These are all “DSD” devices (I still dislike the marketing term).

P.S. Your formulas and writeup already contains way too many errors to begin with.

How various DAC chips or class D amps (which are not even digital devices) work are all off-topic. Unless you or anyone else comes up with a math rebuttal, I will politely ignore.

They are very much on-topic. Class-D amps are simplest to understand and constantly switch between two levels, and the control can be implemented in analog or digital way. The switching transistor doesn’t know if the control comes from analog or digital source (same would apply for a DSD DAC as well).

Class-D amps are like super simplified DSD DACs. No multi-element and bare bones post-filter. Which of course manifests in much worse performance than what you’d get from a line level DSD DAC.

Ok, so they all introduce correlated quantization errors. That’s the topic.

No they don’t, proper modulator doesn’t. But so far you have missed a modulator altogether from your write up.

But PCM has series of correlated images over multiples of the sampling rate to the inifinity.

That’s the whole point: the conclusion is modulator-independent.

It is not, because your thinking is all fundamentally mathematically wrong because of that.

I still don’t see a math rebuttal. I only see a rebuttal of math.

You wanna pay for math classes?

Look, I’ve graduated from physics and I’m well versed in math. Besides, the math I presented is very simple, no lessons needed. If you don’t mind, I’ll stop here.

Still you got that wrong too.

I’d be also curious to see some of your analysis of these things from DAC analog outputs.

Off-topic. You won’t see anything analog here, since I’m talking about digital modulators.

Which are used to drive actual D/A conversion stages, otherwise they are useless. Reason for using delta-sigma modulators in first place are technical limitations of the physical world. So you get into your physics there.

Ok boys that’s quite enough thank you.
Rather than back and forth on the forum please take this to PM, I think it would be better served that way.
Thank you.