layout | mathjax | author | affiliation | e_mail | date | title | chapter | section | topic | theorem | sources | proof_id | shortcut | username | |||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
proof |
true |
Joram Soch |
BCCN Berlin |
joram.soch@bccn-berlin.de |
2023-09-15 09:33:38 -0700 |
Approximation of log family evidences based on log model evidences |
Model Selection |
Bayesian model selection |
Family evidence |
Approximation of log family evidences |
|
P415 |
lfe-approx |
JoramSoch |
Theorem: Let
- Then, the log family evidences can be approximated as
where
- Under the condition that prior model probabilities are equal within model families, the approximation simplifies to
where
Proof: The log family evidence is given in terms of log model evidences as
Often, especially for complex models or many observations, log model evidences are highly negative, such that calculation of the term
- As a solution, we select the maximum LME within each family
and define differences between LMEs and maximum LME as
In this way, only the differences
Using the relation \eqref{eq:LME-diff}, equation \eqref{eq:LFE-LME} can be reworked into
$$ \label{eq:LFE-approx-v1-qed} \begin{split} \mathrm{LFE}(f_j) &= \log \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i) + \mathrm{L}^{}(f_j)] \cdot p(m_i|f_j) \ &= \log \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i)] \cdot \exp[\mathrm{L}^{}(f_j)] \cdot p(m_i|f_j) \ &= \log \left[ \exp[\mathrm{L}^{}(f_j)] \cdot \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i)] \cdot p(m_i|f_j) \right] \ &= \mathrm{L}^{}(f_j) + \log \left[ \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i)] \cdot p(m_i|f_j) \right] ; . \end{split} $$
- Under uniform within-family prior model probabilities, we have
such that the approximated log family evidences becomes
$$ \label{eq:LFE-approx-v2-qed} \begin{split} \mathrm{LFE}(f_j) &= \mathrm{L}^{}(f_j) + \log \left[ \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i)] \cdot \frac{1}{M_j} \right] \ &= \mathrm{L}^{}(f_j) + \log \left[ \frac{1}{M_j} \cdot \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i)] \right] \ &= \mathrm{L}^{*}(f_j) + \log \sum_{m_i \in f_j} \exp[\mathrm{L}'(m_i)] - \log M_j ; . \end{split} $$