5
$\begingroup$

Let $n \sim \mathcal{N}(\mu, \sigma^2)$ and let $u \sim \mathcal{U}(a,b)$, with $b>a>0$, and suppose that $n$ and $u$ are independent random variables. Let $g = n + u$. The probability density function (pdf) of $g$, denoted with $p(\cdot)$, is the convolution between the gaussian distribution and the uniform distribution and, if my derivation is correct, it should be given by the following expression $$ (*) \quad \quad p(t) = \frac{1}{(b-a)\sqrt{2 \pi \sigma^2}} \int_{a+\mu}^{b+\mu} \exp \left\{-\frac{(\tau -t)^2}{2\sigma^2} \right\} d\tau $$


Besides the verification if $(*)$ is correct, I'm looking for a generalization to $N$-dimensional spaces. Let $\mathbf{n} \sim \mathcal{N}_N(\mathbf{m}, \mathbf{C})$ and let $\mathbf{u}$ be a vector of indipendent uniform random variables, each defined in a (possibly) different interval $(a,b)$, i.e. $u_1 \sim \mathcal{U}(a_1, b_1)$, $u_2 \sim \mathcal{U}(a_2, b_2)$, and so on, with the only condition that $b_i > a_i > 0$. The vectors $\mathbf{a,b}$ cointain these scalars.

As before, $\mathbf{n} \in \mathbb{R}^{N \times 1}$ and $\mathbf{u} \in \mathbb{R}^{N \times 1}$ are indipendent; let $\mathbf{g} = \mathbf{n} + \mathbf{u}$. Is the pdf of $\mathbf{g}$ of the type $$ (**) \quad \quad p(\mathbf{t}) \propto \frac{1}{\sqrt{\det(\mathbf{C})}} \int_{\mathbf{a}+\mathbf{m}}^{\mathbf{b}+\mathbf{m}} \exp \left\{ -\frac{1}{2}(\mathbf{v-t})^T \mathbf{C}^{-1}(\mathbf{v-t}) \right\} d\mathbf{v} $$ where $^T$ denotes transpose? If $\mathbf{C}$ is a diagonal matrix (making the gaussian variables indipendent from each other), can the integral be splitted in smaller pieces, in the sense that the $N$-dimensional integral becomes the product of $N$ monodimensional integrals?

$\endgroup$
5
+50
$\begingroup$

The first formula is correct. This is a direct consequence of the definition of the convolution and the definition of the normal and uniform laws.

For the second formula, there are 2 missing factors: $$ \frac 1{\prod_{k=1}^N (b_k - a_k)} $$ which is the inverse of the total weight of the uniform distribution, and $$ \frac 1{(2\pi)^{N/2}} $$ to make sure the weight of the normal distribution is 1.

For the third question, if $C = \text{diag}(\sigma_1^2, \dots \sigma_N^2)$ then the integrand splits and becomes $$ \prod_{k=1}^N \int_{m_k + a_k}^{m_k + b_k} \frac 1{(b_k-a_k)\sqrt{2\pi\sigma_k^2}} \exp\left(-\frac {(v_k - t_k)^2}{2\sigma_k^2}\right) dt_k $$ so the coordinates remain independent.

$\endgroup$
  • $\begingroup$ Very good answer. Thanks! $\endgroup$ – PseudoRandom Oct 15 '14 at 13:04

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.