1
$\begingroup$

It is said that we may use the binomial coefficients ( a layer from Pascal's triangle) to approximate the 1-D Gaussian kernel with certain $\sigma$, where $\frac{n}{4} = \sigma$ and $n$ is the index of the layer.

This works really nice when we want to generate a gaussian quickly. But is there a way to prove this? Or just coincidence?

And how do we decide the variance of certain layer of Pascal's triangle? Say the third layer is [1 3 3 1] (consider [1 1] as the first layer), and it is used to approximate gaussian kernel of $\sigma = \sqrt{\frac{3}{4}} = \frac{\sqrt{3}}{2}$. But how do we prove this?

$\endgroup$
  • $\begingroup$ maybe should ask the math SE this question. we be just dumb electrical engineers. we don't know which end of the integral symbol is up. $\endgroup$ – robert bristow-johnson Mar 6 '18 at 8:56
  • 1
    $\begingroup$ I think this is the reverse of the De Moivre-Laplace Theorem . $\endgroup$ – Shukant Pal Feb 8 at 15:10
2
$\begingroup$

Let $X \sim B(n, p)$. At the third layer $\begin{bmatrix}1 & 3 & 3 & 1\end{bmatrix}$ of the Pascal's triangle, we have $X \sim B(3, \frac{1}{2})$. Which means $P(X = k) = {3 \choose k} (\frac{1}{2})^k (\frac{1}{2})^{3-k} = \frac{{3 \choose k}}{2^3}$. The variance is given by $\sigma^2 = n p q = 3 \cdot \frac{1}{2} \frac{1}{2} = \frac{3}{4}$ and the std by $\sigma = \sqrt{\frac{3}{4}}$. You can derive the variance by looking at $n$ independent Bernoulli trials.

In general, a 1d binomial filter/kernel of size $n$ with $p = \frac{1}{2}$ is given by $\begin{bmatrix}P(X = 0) & P(X = 1) & \cdots & P(X = n-1)\end{bmatrix}$. As $n \to \infty$, this kernel will approximate the normal distribution with mean $\frac{n}{2}$ and variance $\sigma^2 = \frac{n}{4}$. If you want a different variance/std, you should change $p \neq 0.5$.

As somebody already wrote in the comments, you should look at the De Moivre-Laplace theorem to prove this. However, this is only a special case of the more general central limit theorem. By showing that the binomial distribution is asymptotically equivalent to the normal distribution as $n\to\infty$, I would argue that we have shown already both directions.

| improve this answer | |
$\endgroup$
0
$\begingroup$

Hi: The CLT for the sum says that

$\sum_{i-1}^{n} X_{i} \rightarrow Norm(n \mu, n \sigma^2)$ where $\mu$ is the mean of $X_{i}$ and $\sigma^2$ is the variance of $X_{i}$.

And, when $X_{i}$ is binomial, it can be shown that, $\mu = p$ and $\sigma^2 = p(1-p)$.

So, you put those into the CLT for the sum and that's why.

| improve this answer | |
$\endgroup$

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.