1
$\begingroup$

I had recently posted a question on applying powers on a sum of Gaussians (here) to enhance signal resolution artificially. The discussion has given rise to another query. Consider this problem. We have two Gaussian time functions $G_1$ and $G_2$. Suppose we wish sum these Gaussians and raise it to a power of 2.

Situation no. 1: $(G_1+G_2)^2$, we should get another function $G$= $G_1^2$ + $2G_1G_2$ + $G_2^2$. The product of two $G_1G_2$ is another Gaussian. We should see three peaks in this case.

Situation no. 2: We sum the two Gaussians element wise i.e., $S_i$ = $g(t)_{1i}$ + $g(t)_{2i}$ where $g(t)_i$ 's are the corresponding elements of $G_1$ + $G_2$. Now each summed element $S_i$ is raised to power 2. We get only two peaks as a result.

I discussed this issue with some of my colleagues including the person who had proposed this method; they seek agree that both operations are equivalent. It seems they are not, because in situation 2, we will never see $G_1G_2$. What would the discrete version of situation 1?

I feel there is some fallacy here and some mathematical insight should be able to settle this problem.

Thanks.

For Situation no. 2: Here is the MATLAB code.

t=[0:1/200:60]';

u1= 19; % mu the mean time peak i

u2= 22; % mu the mean time peak j

A= 250; %Area

s_ij=[0.84 0.87];; %s_ij represent the sigma of i,j pair; Choose elements by indexing

G_ij=A*normpdf(t, u1, s_ij(1,1))+ A*normpdf(t, u2, s_ij(1,2)); %Sum normpdf is a built in function for a unit area Gaussian from statistical package.

Power= (G_ij).^2;

subplot(2,1,1); plot(t,G_ij);title('Unresolved Gaussians');

subplot(2,1,2); plot (t, Power);title ('Gaussian Sum Raised to Power 2');
$\endgroup$
  • $\begingroup$ i might make a stab at answering this question, but i still don't see this question described adequately. there is more to a random signal in the time domain than it's p.d.f. (which is gaussian). there are also the spectral characteristics which are often described with a power spectral density or with the autocorrelation of a stochastic signal. anyway, to be consistent with electrical engineering convention, you should denote the time functions as "$g_1(t)$" and "$g_2(t)$". and define how they are related to each other (if they are at all), but we need the p.d.f. and autocorrelation at least $\endgroup$ – robert bristow-johnson Sep 17 at 3:10
  • $\begingroup$ @robertbristow-johnson, I am afraid this is not as complicated. I am a chemist, and the sum of two simple Gaussians are just a simulation of two chromatogram peaks. There is no correlation concept involved here, as far I can understand and the position of two peaks is not all related -just two arbitrary peaks. $\endgroup$ – M. Farooq Sep 17 at 3:27
  • $\begingroup$ I don't see a definition for G_ij in the code. $\endgroup$ – MBaz Sep 17 at 13:24
  • $\begingroup$ Corrected. G_ij $\endgroup$ – M. Farooq Sep 17 at 13:32
  • $\begingroup$ In the first plot you have the sum of two Gaussians. In the second, you have the sum squared. I don't understand why you think the two should be the same? $\endgroup$ – MBaz Sep 17 at 14:59
2
$\begingroup$
([a b c] .+ [d e f]).^2 = [a+d b+e c+f].^2
                        = [(a+d)^2 (b+e)^2 (c+f)^2]
                        = [a^2+2ad+d^2 b^2+2be+e^2 c^2+2cf+f^2]

I suspect that:

  • The problem may be in your assumptions: "We should see three peaks in this case" sounds like a hypothesis that may be false.
  • You may have a bug in your code, which might be why "We get only two peaks as a result".

Here is some code to illustrate that (A.+B).^2 = A.^2 .+ 2A.*B .+ B.^2:

t=0:1/200:60;
gauss1 = 250*normpdf(t,19,0.84);
gauss2 = 250*normpdf(t,22,0.87);
m1 = gauss1.^2+2*gauss1.*gauss2+gauss2.^2;
m2 = (gauss1+gauss2).^2;
fprintf('Difference: %1.2f\n', sum(abs(m1-m2)));
subplot(2,1,1);plot(m1);title('Method 1');
subplot(2,1,2);plot(m2);title('Method 2');

I get:

Difference: 0.00

and

enter image description here

So, the math is correct. You may still be wondering, how come the third Gaussian $G_1G_2$ is not apparent in the plot? This is the point where it's useful to challenge your assumption that the third Gaussian should be visible. If we plot $G_1G_2$ by itself we obtain:

plot(2*gauss1.*gauss2);

enter image description here

Look at the amplitude and the position in the x-axis: this third Gaussian is not apparent in the plot because it is completely swamped by the other two, much larger, surrounding peaks.

$\endgroup$
  • $\begingroup$ I am afraid there is no fancy code, I am just doing it in Excel or Matlab. I think there is some other fundamental fallacy. $\endgroup$ – M. Farooq Sep 17 at 2:15
  • 1
    $\begingroup$ I doubt it. The problem is not complex enough to hide a fallacy. If you share your code I'll take a look at it. $\endgroup$ – MBaz Sep 17 at 2:40
  • $\begingroup$ added the code for the sake of completeness. Thanks. $\endgroup$ – M. Farooq Sep 17 at 2:55
  • 1
    $\begingroup$ I was just about to post an answer when I saw that @MBaz nailed it. I got the same thing, as expected. I already upvoted this answer. So I will e-mail what I have separately to M. Farooq. $\endgroup$ – Ed V Sep 17 at 15:16
  • 1
    $\begingroup$ So we do generate a new peak when we square element-wise. Thanks for double checking and confirming this problem. $\endgroup$ – M. Farooq Sep 17 at 17:18

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.