Approximate Bayesian computations

with symbolic integration






  Lubomír Soukup
  Department of Image Processing
  Institute of Information Theory and Automation
  Academy of Sciences of the Czech Republic
  Pod vodárenskou vezí 4, 182 08 Praha 8
  Czech Republic
phone: + 42 2 6605 2551
fax: + 42 2 688 4903
e-mail: soukup@utia.cas.cz








In Bayesian statistics it is to evaluate Bayes formula to obtain probability distribution of estimated parameters. Bayes formula for independent observations is as follows.


 \begin{displaymath}
g(\mathbf{x}\,\vert\,\widetilde{\mathbf{y}}) = \frac{\displa...
...}) - \widetilde{y}_j)\,p(\mathbf{t})\,d\mathbf{t}} \; \mbox{.}
\end{displaymath} (1)

Here
x  ...  vector of estimated parameters, $\mathbf{x}\stackrel{\triangle}{=}
({x}_1,\,{x}_2,\,\dots\,,\,{x}_m)$, $\mathbf{x}\in {\cal X}$,
       ${\cal X} \subseteq I\!\! R^m$, $m \in I\!\! N$,
$\widetilde{\mathbf{y}}$  ...  vector of measured values (observations), $\widetilde{\mathbf{y}} \stackrel{\triangle}{=}
(\widetilde{y}_1,\,\widetilde{y}_2,\,\dots\,,\,\widetilde{y}_n)$,
aj  ...  functions that are measured for certain fixed argument $\mathbf{x}\in {\cal X}$,
       $j \in \{1,\,\dots\,,n\}$, $n \in I\!\! N$,
$g(\mathbf{x}\,\vert\,\widetilde{\mathbf{y}})$  ...  a posterior density function,
$f_j(\widetilde{y}_j-a_j(\mathbf{x}))$  ...  probability density function of the error of measurement $\widetilde{y}_j$,
p(x)  ...  a prior density function.




The problem is how to evaluate the integral in the denominator in ([*]). In the general case when the functions aj are nonlinear and density functions fj are not normal, the integral have to be evaluated approximately. Unfortunately, no universal eficient method of integration in Bayes formula exist. The most general approach uses simulations based on Monte-Carlo method (see e.g. [#!j:MfAIiSwSEoBIP!#]), but it is too much time consuming for large m. Standard techniques of numerical integration (e.g. Simpson's method and its modifications) are also too slow even for the common number of estimated parameters (m > 4). The most efficient approach utilize symbolic integration, but the applicability of this approach depends on the analytical form of functions aj, fj, p.

When the functions aj are only slightly nonlinear and probability distributions fj, p are close to normal, assumptions of linearity and normality are adopted to evaluate the integral symbolically. It actually represents approximation of the posterior distribution by normal distribution and the estimation problem can be solved by linear least-squares method. In such an approach the computation is very fast but it is hard to estimate the influence of the simplifications to the outcoming posterior probability density function.

The assumptions of linearity and normality can be justified with the aid of modification of density functions fj (see [#!c:PDfENE!#]). This modification corrects the nonlinearity while the mean values and standard deviations of the original distributions remain unchanged. To meet these requests another approximate integration have to be performed, that is easier then integration in ([*]) and can be done symbolically. This integration appears in equations of the form

 \begin{displaymath}
\int_{I\!\! R} b_k(x) \, \gamma_k \,
\exp\left(-\frac{1}{2...
...uad k \in
\{1,\,2,\,\dots\,,m\} \,,\; l \in \{0,\,1,\,2\} \;,
\end{displaymath} (2)

where bk are given analytical functions, ckl known constants, and $\gamma_k$, $\mu_k$, $\sigma_k$ are unknown.

System of equations ([*]) can be solved by means of infinite series. In fact, only few terms of the resulting series have to be evaluated. Nevertheless, the symbolic treatment with series is very tedious work and some computer algebra system has to be used. System of equations ([*]) was solved with the aid of Mathematica software by Wolfram Research. The crucial operations that has been frequently used were differentiation (D), series inversion (SeriesInversion), and composition of series (ComposeSeries).

The designed method can be applied to such scientific and technical branches where slightly nonlinear models are usually used. This case is typical for geosciences such as geodesy, cartography, remote sensing, geology etc.



 

IMACS ACA'98 Electronic Proceedings