1
$\begingroup$

I am trying to calculate a posterior predictive distribution for the magnitude (Euclidean norm) of a 3D displacement vector. Displacement in each dimension is independent and normally distributed (but each may have different means and variances).

$$ X_i \sim N(\mu_i,\sigma_i^2); (i = 1:3) $$

$$ D = \sqrt{\sum\limits_{i=1}^3 X_i^2} $$

There is plenty of literature on how to deal with normally distributed data with both frequentist and bayesian approaches.

However the magnitude isn't normally distributed, and I believe it should follow a noncentral chi distribution. This is defined by number of degrees of freedom k = 3, and the lambda parameter, which combines the ratio of the mean and standard deviation for each dimension (en.wikipedia.org/wiki/Noncentral_chi_distribution).

$$ \lambda = \sqrt{\sum\limits_{i=1}^k \left(\frac{\mu_i}{\sigma_i}\right)^2} $$

I don't know any population parameters - but I have informative priors on the parameters governing displacement in each dimension, and would like to use them:

$$ \mu_i \sim N(m_i,a_i^2) $$ $$ \sigma_i^2 \sim N(s_i^2,b_i^2) $$ Where $m_i$ and $s_i^2$ are my best guesses for the mean and variances of data in each displacement dimension, and $a_i^2$ and $b_i^2$ are hyperparameters quantifying my uncertainty in these beliefs.

How can I convert prior information on the gaussian displacement parameters into a prior on the lambda parameter of the non-central chi distribution?

$$ \lambda \sim f(m_i,a_i^2,s_i^2,b_i^2)? $$

Next I will combine this prior with observed data (either in each dimension, or the magnitude directly) to obtain a posterior distribution for lambda. I will then integrate over all lambda to get my posterior predictive distribution for the magnitude. I think this is a sensible approach, but I need a decent prior first.

$\endgroup$
  • 1
    $\begingroup$ @Xi'an Edited to add equations - I've done my best (I'm not very familiar with LaTeX), hopefully that makes the question clearer. $\endgroup$ – Arun Jan 6 '17 at 10:27
2
$\begingroup$

Note that

  1. the distribution of $D^2$ does not depend directly on $\lambda$ but rather on $||\mathbf{\mu}||$ and $(\sigma_1,\ldots,\sigma_k)$. For instance,$$\mathbb{E}[D^2]=||\mathbf{\mu}||^2+\sum_{i=1}^k \sigma^2_i$$
  2. this distribution is definitely not a non-central $\chi^2$ distribution (*) as explained in this SE answer

Since $\lambda$ is a transform of $(\mu_1,\sigma_1,\ldots,\sigma_3)$, $$\lambda=T(\mu_1,\sigma_1,\ldots,\sigma_3)$$ the prior density of $\lambda$ is the (mathematical) projection of the joint density of $(\mu_1,\sigma_1,\ldots,\sigma_3)$, which is given by the Jacobian formula. If this proves impossible to derive in a closed-form formula, a simulation-based representation can be achieved by simulating from the joint prior and constructing the resulting $\lambda$'s.

If one observes the $X_i$'s (rather than $D$), the posterior of $(\mu_1,\sigma_1,\ldots,\sigma_3)$ given the $X_i$'s is available in closed form. And can again be simulated to produce a projected posterior on $\lambda$.

If one only observes the norm $D$, the posterior of $(\mu_1,\sigma_1,\ldots,\sigma_3)$ given $D$ is harder to manage if only because the distribution of $D$ is not standard and I would suggest treating the $X_i$'s as missing data in a simulation that would then condition on those $X_i$'s [in a Gibbs sampler] to simulate $(\mu_1,\sigma_1,\ldots,\sigma_3)$.


(*) The reason why $D^2$ is not distributed as a non-central $\chi^2$ variate is that it can be expressed as $$D^2=\sum_{i=1}^3 X_i^2=\sum_{i=1}^3 (\mu_i+\sigma_i\epsilon_i)^2\qquad\epsilon_i\sim\text{N}(0,1)$$thus that the different scales $\sigma_i$ prevent a non-central $\chi^2$ representation.

To achieve a non-central $\chi^2$ distribution, as defined on the Wikipedia page, one would need to divide the $X_i^2$'s by the \sigma_i^2$'s, which is impossible since those are unknown.

$\endgroup$
  • $\begingroup$ Thank you for your response. My primary goal is to create a posterior predictive distribution for $D$. In the case of perfect parameter knowledge, (all $\mu_i$ and $\sigma_i^2$ known), is $D$ chi-distributed as defined by $\chi$ and $k$? $\endgroup$ – Arun Jan 9 '17 at 14:02
  • $\begingroup$ I'm afraid I am confused - in the case of known sigma's, why is the equation for lambda not valid? $\endgroup$ – Arun Jan 9 '17 at 17:12
  • $\begingroup$ @Xi'an would you say that a normal prior for $\sigma^2$ could be considered dangerous/inconvenient/ temerarious? $\endgroup$ – mugen Jan 9 '17 at 23:18
  • 1
    $\begingroup$ @mugen: a normal prior on $\sigma^2$ needs to be truncated at zero. Other than that, it is a proper prior, hence leads to a valid Bayesian analysis of the problem. I do not see how qualifications like dangerous and others apply to a prior: a prior is a prior (is a prior). $\endgroup$ – Xi'an Jan 10 '17 at 8:51
  • 1
    $\begingroup$ @Xi'an That makes much more sense, thank you. I think my original question is probably now tangential to what I want to do - calculate a posterior predictive interval for future $D$, given either acquired measurements of $D$, or measurements of $X_i$ with priors on the parameters underlying the generation of $X_i$. Time to try some simulations I think. $\endgroup$ – Arun Jan 13 '17 at 16:53

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.