0
$\begingroup$

For a research project I work with continuous distributions, like the normal distribution. In my use case however the random variable Z generally follows a normal distribution, though it can only take discrete values (..., -3, -2, -1, 0, 1, 2, 3, ...)$\in \mathbb{Z}$.

For the "discretization" I use a simple heuristic. An easy example would be a standard normal distribution:

$$ Z \sim \mathcal{N}(0,1) \\ f(Z) = \sum_{i\in \mathbb{Z}} \mathbb{1}_{(i-0.5, i+0.5]}(Z) \\ X = f(Z) $$

My question now is: Is there a special name for the resulting distribution X? Is there any research on the resulting distribution?

Best, Simon

$\endgroup$
1
  • $\begingroup$ This is called the discretized normal distribution on p. 2 of arxiv.org/pdf/1111.3162.pdf . See the paper for various theoretical results. $\endgroup$ Jul 14 '16 at 0:17
1
$\begingroup$

In lattice cryptography, discrete gaussian distribution is used a lot. However, there, it is defined in a different way: $$ \mathbb P(X = k) = \dfrac{1}{S} e^{-k^2/2}, \quad S = \sum_{k \in \mathbb Z} e^{-k^2 / 2}. $$ Probably, they use this distribution because it has the maximal entropy among discrete distributions with fixed variance. I refer to ''Compact and Side Channel Resistant Discrete Gaussian Sampling'' for applications in cryptography (they consider problems of sampling with Knuth-Yao algorithm, and some more computational issues). I also refer to ''Probability distributions and maximum entropy'' for principle of maximum entropy - I didn't find the proof of maximal entropy for discrete gaussian mentioned above, but probably using the tools described in the article, you can do that.

What you use here, is rather $\lfloor \mathcal N(0,1) + \tfrac12 \rfloor$ with probability 1 (I mean, deciding on the value at the half-integers).

$\endgroup$

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.