I'm having trouble understanding the following. Say I have a quantity of objects arranged spatially, with distribution $dn/dx$. For one reason or another*, I want to sum auto-pairs of spatial points, but only pairs at the same place. Thinking about this discretely, we have

$$ \sum_{i,j} (\bar{n}_i \Delta x) (\bar{n}_j \Delta x) \delta_{ij} = \sum_i (\bar{n}_i \Delta x)^2 $$

Clearly this is dimensionless on both sides. However, moving into the continuous limit, I expect to have

$$ \int\int \frac{dn}{dx_i}\frac{dn}{dx_j}\delta(x_i-x_j)dx_idx_j = \int \left(\frac{dn}{dx}\right)^2 dx. $$

Here both sides have the same dimensions (inverse length), but this is different from the sum, which is dimensionless. What am I doing wrong??

Thanks for any help!

* For those who are interested, this problem arises (or at least I think it does) when finding the variance of a sum of correlated variables. In particular, let the number of objects be a function of $x$ and $y$, and let them be spatially correlated in $y$, but not $x$. Then

$$ {\rm Var}(\sum N_i) = \sum_{x_i}\sum_{x_j}\sum_{y_i}\sum_{y_j} (\bar{n}_{x_i} \Delta x \Delta y) (\bar{n}_{x_j} \Delta x \Delta y)C_{y_i y_j} \delta_{x_i x_j} = \sum_{x_i}\sum_{y_i}\sum_{y_j} (\bar{n}_i \Delta x \Delta y)^2 C_{y_i y_j} .$$

The rest follows as above (note that I have that the spatial covariance in $y$ at a given $x$ is $(\bar{n}_x \Delta x \Delta y)^2 C_{y_i y_j}$).

  • $\begingroup$ You have to consider the differences of the discrete Dirac-Delta operator and the continuous Dirac-Delta distribution. If your discretization widths $\Delta x_i,\Delta x_j$ shrink your sum converges to zero since $\delta_{ij} = 1$ when $i=j$ and $\Delta x_i,\Delta x_j\rightarrow 0$. You have to replace the discrete Dirac-Delta by a density like $\frac{\delta_{i,j}}{\Delta x_i}$ to get something that converges to the Dirac-Delta distribution. $\endgroup$
    – Tobias
    Nov 18 '16 at 7:11
  • $\begingroup$ @tobias -- I think this is the correct answer. Would you like to post it as an answer so I can accept? In my case, the discrete version is "correct" while the integrated version is wrong, so insead I'd have to replace the continuous Dirac with something like $\delta(x-k) dx$. $\endgroup$ Nov 21 '16 at 2:53

I think that the discrepancy is that your discrete case is not the strict analogue of your continuous case.

If I were summing up just values of $n$, and using the Kronecker delta to pull out one particular value, say $k$, then we'd have $$ \sum_i a_i\delta_{ik} = a_k, $$ and its continuous analogue would be $$ \int a(x)\delta(x-k)\,dx = a(k). $$ In all cases, the units are the same (whatever units $a$ has). Notice that the discrete version lacks the extra $\Delta x$ that you included.

Working backwards from your continuous version, we have \begin{align} \int \frac{dn}{dx_i}\frac{dn}{dx_j}\delta(x_i-x_j)\,dx_i\,dx_j &= \int \left(\frac{dn}{dx}\right)^2\,dx \\ \text{(Going to discrete case)}&\Rightarrow \sum_i \bar{n}^2_i \Delta x \\ &= \left[\sum_{i,j}\bar{n}_i\bar{n}_j \delta_{ij}\right]\Delta x, \end{align} so both versions have units of inverse length.

  • $\begingroup$ You could be right, but I'm not absolutely convinced. The sum of $a_i$ does not converge to the integral over $a$, rather the sum of $a_i \Delta x$ does. Unless you are saying what @tobias did -- i.e. that $\delta_{ik}$ in the continuous limit literally translates to $\delta(x-k) dx$? $\endgroup$ Nov 21 '16 at 2:56
  • $\begingroup$ Yep, I think @Tobias is correct too. I think we're ultimately saying the same thing, but I was only working by analogy, and Tobias provided the underlying reason why. $\endgroup$
    – robotopia
    Nov 22 '16 at 0:04

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.