Particle approximation to probability density functions: Dirac delta function representation

In the previous post, I have briefly shown the idea of using dirac delta function for discrete data representation. In the second example there, a histogram locations for a given set of points are presented as spike trains, where as heights are somehow given in a second sum. This is hard to follow and visualise, of course if you are not that good in reading formulation with different indexes. Due to pedagocial reasons, an easier representation of arbitrary probability density function (PDF), $p(x)$, one would simply need to couple each discrete points with a corresponding weight.

Hence, a set  $\{x_{i}, \omega^{i}\}_{i=1}^{N}$ would be an estimation of PDF, $\hat{p}(x)$ . At this point we can invoke dirac delta function,

$ \hat{p}(x) = \sum_{i=1}^{N} \omega^{i} \delta(x-x_{i})$

Let's revisit the R code given there, this time let's draw uniform numbers between $[-2, 2]$ to get 100  $x_{i}$ values. Simply these numbers will indicate the locations on the x-axis, a spike train.  For simplicity, let's use Gausian distribution for target PDF, $ \mathcal{N}(0, 1)$.  Than, for weights we need to draw numbers using the spike locations. This approach is easier to understand compare to my previous double index notation.

R Example code

Above explained procedure is trivial to implement in R.

1 2 3 4 5 6 7 8 9 10

# Generate locations 100 x locations # out out 1000 points in [-2.0, 2.0] set.seed(42) # Domain where Dirac comb operates Xj = seq(-2,2,0.002) Xi = sample(Xj, 100) # Now generate weights from N(0,1) at those given locations Wi = dnorm(Xi) # Now visualise plot(Xi, Wi, type="h",xlim=c(-2.0,2.0),ylim=c(0,0.6),lwd=2,col="blue",ylab="p")

Conclusion

Above notation introduces second  abuse of notation while actually there must be a secondary regular grid that pics $x_{i}$ values using dirac delta in practice. Because, the argument of $\hat{p}(x)$, is in the discrete domain. So a little better notation that reflects the above code would be

$ \hat{p}(x_j) = \sum_{i=1}^{N} \omega^{i} \delta(x_{j}-x_{i})$

The set $x_j$ is simply defined in a certain domain, for example regularly. Hence I only recommend not to introduce dirac delta for explaining a particle approximation to PDFs for novice students in the class. It will only confuse them even more.

Figure: Spike trains with weights $\hat{p}(x) = \sum_{i=1}^{N} \omega^{i} \delta(x-x_{i})$

Demystify Dirac delta function for data representation on discrete space

Dirac delta function is an important tool in Fourier Analysis. It is used specially in electrodynamics and signal processing routinely.  A function over set of data points
is often shown with a delta function representation. A novice reader relying on integral properties of the delta function may found this notation quite confusing.  Probably, the notation itself is an example of  abuse of notation.

One dimensional function/distribution: Sum of delta functions

Let's define a one dimensional function, $f(x)$ as follows, $x \in \mathbb{R}$ and $a$ being constant:

$ f(x) = a \sum_{i=-n}^{n} \delta(x - x_{i})$

This representation is inspired from Dirac comb and used in spike trains.  Note that set of data points in one dimension $\{x_{i} \}$ will determine the graph of this function. Using the shifting property of delta function, the value of the function will be zero every where except on data points. The constant $a$ will simply be the height of the graph at the data point.

Figure: A spike train.

Numeric Example

Let's plot $f(x)$ for some specific values of the set $\{x_{i} \} = {-0.5, -0.2, -0.1, 0.2, 0.4}$ and $a=0.5$. Here is the R code for plotting this spike train. 

x_i = c(-0.5, -0.2, -0.1, 0.2, 0.4)
a   = c(0.5, 0.5, 0.5, 0.5, 0.5)
plot(x_i,a,type="h",xlim=c(-0.6,0.6),ylim=c(0,0.6),lwd=2,col="blue",ylab="p") 

Representing Histograms: One dimensional example

Particularly convenient representation of histograms can be developed similarly. Consider set of points $\{x_{i}\}_{i=1}^{n}$ where we would like to establish a histogram out of this set, let's say $h(x)$. If we set our histogram intervals as $\{x_{j}\}_{j=1}^{m}$. The histogram $h(x)$ then can be written as

$h(x_{j}) =  \sum_{i=1}^{n} \sum_{j=1}^{m} \delta(x_{j}- x_{i}^{min})$
where set $x_{i}^{min}$  represents the value from set $\{x_{j}\}_{j=1}^{m}$ that is closest to given $x_{i}$. Where as, second sum determines the height at a given point, i.e., frequency. This is just a confusing mathematical representation and practical implementation only counts the frequency of $x_{i}^{min}$ directly.


Conclusion

However it is quite trivial, the above usage of sum of delta functions appear in mathematical physics as well, not limited to statistics.