I'm dealing with an echo signal from a hybrid. An echo path includes DAC, the hybrid, ADC and the line infrastructure. I can see, that after using a digital linear echo-canceller the residual echo signal is still significant and I suspect that echo-path has additional non-linear components. Now I want to estimate, how much of the incoming signal is an actual echo and not a noise, and I want to do it before actually implementing or inferring any specific non-linear function (e.g. without assumptions regarding the order of Volterra series).

I've found the ITU-T O.42 recommendation, which suggests using the 4-tone intermodulation method to measure nonlinear distortion impairments on analogue circuits. But it doesn't suit me, since my echo path includes the DAC.

So is there a way to measure or estimate the level of the non-linearity in my echo path? I would highly appreciate any references. Also, if this question is more related to electronics, please feel free to migrate it to the corresponding site.


Definition of non-linearity: based on the source, a non-linearity for an echo signal is anything that cannot be modeled with a linear filter: a dc component, a long tail and truly non-linear dependencies. So the question becomes: is there a way to measure or estimate the level of those components compared to noise? The peculiarity here is that both a dc component and a long tail are actually linear in their nature.

Signal measurements: Following are the data that I got by changing the power of transmitted signal (from 1 to 6).

Power spectral density plots (regular and normalized), where frequencies are limited by a sampling rate:

PSD PSD normalized

Left: echo mean z mean and variance z var and residual echo mean e mean and variance e var for a regular (random) transmitted signal (using some arbitrary length of a linear filter). Right: echo mean and variance and residual echo mean and variance (using filter weights from training on the regular signal) for a constant transmitted symbol.

Regular tx Constant tx


  1. The PSD plots look identical, except for unclear shift of a plot for the lowest power level (1).

  2. Mean and variance plots do not resemble any obvious non-linearity, which would have shown as a power law dependency everywhere. Based on increased e mean for a constant transmitted symbol, there is a tail in the echo.

  3. What puzzles me the most is a linear growth of a residual echo variance, which I cannot explain.

So the original question is still relevant. Additionally, can I actually use obtained plots to estimate the echo non-linearity and if so, how should I do it?


excellent, you have PAM-4!

Send a long sequence of your least-amplitude symbol, and integrate the received power to calculate the energy of your received signal.

Do the same with the three other symbols.

Now, you can solve the equation

$$E_{RX} = \alpha E_{TX} + E_N + F(P_{TX}),$$

with $\alpha$ being the channel gain, which has to be the same for all 4 of your measurements, $E_{RX},E_{TX}$ being the transmitted and received energies (hint: normalize them to the least-power symbol), $E_N$ is noise energy, i.e. noise power times number of samples (which becomes asymptotically constant if you just average enough symbols), and $F$ is the nonlinear error term.

With $E$ known, and $\alpha$ and $E_N$ overdetermined determined from four measurements, $F$ is the reason for the error.

  • $\begingroup$ Great suggestion! The main concern: doesn't this approach require that I make an assumption about the non-linearity in order to solve the equation system? E.g. I set $F(P_{TX}) = a * P_{TX}^2 + b * P_{TX}^3$, and then solve the system for $a$, $b$, $\alpha$ and $E_N$. This is somewhat limiting. What if I want to estimate all non-linear components? $\endgroup$ – megasplash Jan 25 at 10:13
  • $\begingroup$ all the nonlinear components are in $F$; to determine it, you might need more than 4 measurements with different powers , though. $\endgroup$ – Marcus Müller Jan 25 at 10:38
  • $\begingroup$ So did I understand you correctly, that I have to use an explicit form of $F$? $\endgroup$ – megasplash Jan 25 at 10:56
  • $\begingroup$ If that is true, it will limit the results, as function $F$ will include only the explicitly defined non-linear components, and not all of them. $\endgroup$ – megasplash Jan 27 at 11:27
  • $\begingroup$ no, you only get an intergral of "all the power that doesn't fit in the linear model" as F, not a model of any kind. $\endgroup$ – Marcus Müller Jan 27 at 12:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.