next up previous contents
Next: Maximum posterior approximation Up: Prior models for potentials Previous: Mixtures of Gaussian process   Contents


Average energy

Using a standard Gaussian smoothness prior as in Eq. (34) with zero reference potential $v_0 \equiv 0$ (and, say, zero boundary conditions for $v$) leads to flat potentials for large smoothness factors $\lambda $. Especially in such cases it turned out to be useful to include besides smoothness also a priori information which determines the depth of the potential. One such possibility is to include information about the average energy

\begin{displaymath}
U =
\sum_\alpha p_\alpha E_\alpha
=\,
<E>
.
\end{displaymath} (41)

We may remark, that for fixed boundary values of $v$ a certain average energy cannot be obtained by simply adding a constant to the potential. The average energy can, however, be set to a value $\kappa $ by including a Lagrange multiplier term
\begin{displaymath}
E_U =
\mu (U - \kappa)
,
\end{displaymath} (42)

Similarly, and technically sometimes easier, one can include noisy `energy data' of the form
\begin{displaymath}
p_{{}_{\scriptsize U}} %%p_U
\propto e^{-E_U}
,\quad
E_U =
\frac{\mu}{2} (U - \kappa)^2
.
\end{displaymath} (43)

For $\mu\rightarrow\infty$ this results in $U\rightarrow\kappa$ so both approaches coincide.


next up previous contents
Next: Maximum posterior approximation Up: Prior models for potentials Previous: Mixtures of Gaussian process   Contents
Joerg_Lemm 2000-06-06