0
$\begingroup$

Given two independent normal random variables $X \sim \mathcal{N_{\mu_1,\sigma_1^2}}$ and $Y \sim \mathcal{N_{\mu_2,\sigma_2^2}}$, I want to compute this quantity:

$\hat{J} = \sum_{v \in \mathbb{Z}} \left( \int_{v-0.5}^{v+0.5} f(x | \mu_1,\sigma_1^2) \; dx \cdot \int_{v-0.5}^{v+0.5} f(y | \mu_2,\sigma_2^2) \; dy\right)$

$f(x | \mu_1,\sigma_1^2)$ denotes the normal PDF. Note that, here, functions are integrated and the results are multiplied.

The following quantity is much easier to compute:

$\tilde{J} = \sum_{v \in \mathbb{Z}} \left( \int_{v-0.5}^{v+0.5} f(x | \mu_1,\sigma_1^2) \; \cdot f(x | \mu_2,\sigma_2^2) \; dx\right)$

$ = \int_{- \infty}^{+\infty} f(x | \mu_1,\sigma_1^2) \; \cdot f(x | \mu_2,\sigma_2^2) \; dx$ $ = f(\mu_1 | \mu_2,\sigma_1^2+\sigma_2^2)$

Note that here the functions are multiplied first and then integrated. The last transformation is not trivial: It can be shown, that the product of two normal densities is a normal density (integrates to one in the previous equation) and a scale factor that is normal as well, but depends only on the distribution parameters (given here).

Obviously, in general, $\tilde{J} \neq \hat{J}$. However, when $\sigma_1$ and $\sigma_2$ are restricted to values such that function values are below 1, $\tilde{J}$ appears to approximate $\hat{J}$ quite well.

Comparison of integrate-then-multiply to multiply-then-integrate

Intuitively, this makes sense to me, as changing the PDFs to densitites that are constant on $(v-0.5,v+0.5], v \in \mathbb{Z}$ (continuous representation of a discrete distribution ), makes the approximation hold with equality .

Is there a way to derive an error bound for $\tilde{J} \approx \hat{J}$? For example, an upper bound on the absolute error depending on the distribution parameters?

  • 0
    What is the difficulty in evaluating the cdf of the Normal: $F(v+\frac12) - F(v-\frac12)$? It's a standard function in every package.2017-01-15
  • 0
    Computing the CDF for a single $v$ is no problem, but doing this for all $v \in \mathbb{Z}$ is. Of course, I could use the standard deviations to restrict on those $v$ that have a significant effect on $\hat{J}$, but computing the approximation is still cheaper. My application needs to perform a whole lot of these computations and its time-critical.2017-01-15

0 Answers 0