I have three random variables:

  • $Y$: my data
  • $Y_n$: my data corrupted by additive white Gaussian noise (AWGN)
  • $Y_{nc}$: my noisy data corrupted by a non-linear transformation $\mathcal{C}$.

I have computed the histograms of the noise terms:

  • $N = Y_n - Y$: the AWGN
  • $N_c = Y_{nc} - Y_n$: the noise added by the transformation $\mathcal{C}$.
  • $N_{tot} = Y_{nc} - Y$: the total noise (AWGN + transformation noise)

I have found that the histograms of $N$ and $N_c$ are Gaussian with variances $\sigma_{N}^2$ and $\sigma_{N_c}^2$, respectively, and that $N$ and $N_c$ are uncorrelated based on their covariance analysis. Furthermore, I have found that $N_{tot}$ also has a Gaussian histogram with variance $\sigma_{N_{tot}}^2$.

However, I am surprised to find that $\sigma_{N_{tot}}^2 < \sigma_{N}^2 + \sigma_{N_c}^2$, and even more surprisingly, I have $\sigma_{N_{tot}}^2 < \sigma_{N}^2$. For example, I have computed these values from an example signal:

  • $\sigma_{N}^2 = 3.97277792$
  • $\sigma_{N_c}^2 = 0.26705326$
  • $\sigma_{N_{tot}}^2 = 3.71623896$

This phenomenom always happens even if I change the input data.

Covariance matrix of $N$ and $N_c$ is

$\begin{bmatrix} 3.97277792 & 0.25886258 \\ 0.25886258 & 0.26705326 \end{bmatrix}$

Can someone help me understand why this is happening? I was pretty sure that the result should have been $\sigma_{N_{tot}}^2 = \sigma_{N}^2 + \sigma_{N_{c}}^2$. I am actually closer to $\sigma_{N_{tot}}^2 = \sigma_{N}^2 - \sigma_{N_{c}}^2$

  • 1
    $\begingroup$ Since you are applying a nonlinear transformation to a noisy signal, there is not a simple relation between the noise on the input side of the nonlinearity and the noise on the output side of the nonlinearity. Furthermore, there are cross-terms (involving both signal and noise) that need to be accounted for. Thus, there are too many unknowns here. Can you reveal the nonlinearity in more detail, or is that proprietary information? $\endgroup$ Mar 2, 2023 at 18:54

1 Answer 1


I was pretty sure that the result should have been $\sigma_{N_{tot}}^2 = \sigma_{N}^2 + \sigma_{N_{c}}^2$.

Why would you think this? Noise variances do only add if the processes (and the resulting noise) are uncorrelated. A non-linear transform will in almost cases produce noise that's correlated with the original signal and hence all bets are off.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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