2
$\begingroup$

I was reading through this one page paper on using Gibbs sampling for detecting a change point in a time series like data. While I understand the part where the $\lambda$ and $\phi$ are chosen from a gamma distribution I do not understand how the author generated the posterior distribution for $k$. My thinking is, if the first $k$ points came from a Poisson distribution then the probability of $P(k|\lambda,\phi,Y_i)$ $$ P(k|\lambda,\phi,Y_i) = \Pi_{i=1}^{k}\frac{e^{-\lambda}\lambda^{Y_i}}{Y_i!} \times \Pi_{i=k+1}^{N}\frac{e^{-\phi}\phi^{Y_i}}{Y_i!} $$ The author shows an expression of the form $\frac{A}{B}$ where $B$ is the normalizing term. This part I understand. However it is unclear to me how he found the numerator term which appears to be a log of $$ P(k|\lambda,\phi,Y_i) = \frac{\Pi_{i=1}^{k}\frac{e^{-\lambda}\lambda^{Y_i}}{Y_i!}}{\Pi_{i=k+1}^{N}\frac{e^{-\phi}\phi^{Y_i}}{Y_i!}} $$ How does he arrive at that? Any help appreciated. Thanks in advance

$\endgroup$
  • $\begingroup$ It doesn't look to me like the author is saying what you say he seems to show. $\endgroup$ – Glen_b -Reinstate Monica Aug 9 '14 at 1:51
  • $\begingroup$ He has the term $\frac{A}{B}$ where $ A = e^{k(\phi - \lambda) + \ldots} $. This appears to imply he is taking the ratio $\endgroup$ – broccoli Aug 9 '14 at 4:58
  • $\begingroup$ He has a ratio of two terms (and he should, since the denominator normalizes the numerator so that it sums to one). He doesn't have the ratio you state. $\endgroup$ – Glen_b -Reinstate Monica Aug 9 '14 at 5:44
  • $\begingroup$ But the normalizer is $B$ , which is already in the form of $\sum A$, which I understand. How did he get the term in the numerator, i.e. $A$? That is the part that is confusing me $\endgroup$ – broccoli Aug 9 '14 at 6:07
  • 1
    $\begingroup$ Nope. A good lesson in why you should give full references for anything. From my initial comment it looks like I may have briefly glanced at the paper (the general issue of normalizing such terms is quite familiar though) but I have no recollection of any of the specifics of this instance. Does the wayback machine archive github? $\endgroup$ – Glen_b -Reinstate Monica Oct 16 '18 at 23:06
1
+50
$\begingroup$

I believe the glitch is in your second equation, as the author's formula for $p(k|\textbf{y}, \lambda, \phi)$ doesn't depend on $N$. Manipulating his formula gives a quotient of products, but indexed slightly differently:

$$ e^{k(\phi - \lambda) + log(\frac{\lambda}{\phi})\sum_{i=1}^ky_i} = e^{k(\phi - \lambda)}e ^{ log(\frac{\lambda}{\phi})\sum_{i=1}^ky_i} $$

$$ = \frac{e^{-\lambda k}}{e^{-\phi k}}\left(\frac{\lambda}{\phi}\right)^{\sum_{i=1}^ky_i} $$

$$ = \frac{e^{-\lambda k}}{e^{-\phi k}}\left(\frac{\lambda^{\sum_{i=1}^ky_i}}{\phi^{\sum_{i=1}^ky_i}}\right) $$

$$ = \frac{\Pi_{i=1}^{k}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!}}{\Pi_{i=1}^{k}\frac{e^{-\phi}\phi^{y_i}}{y_i!}} $$

To find our way to this equation, first factor the joint probability $p(\lambda, \phi, k, \textbf{y})$ two ways:

$$ p(\lambda, \phi, k, \textbf{y}) = p(\textbf{y}|\lambda, \phi, k)p(\lambda, \phi, k) = p(k|\lambda, \phi, \textbf{y})p(\lambda, \phi, \textbf{y}) $$

To solve for $p(k|\lambda, \phi, \textbf{y})$, divide both factored probabilities by $p(\lambda, \phi, \textbf{y})$:

$$ p(k|\lambda, \phi, \textbf{y}) = \frac{p(\textbf{y}|\lambda, \phi, k)p(\lambda, \phi, k)}{p(\lambda, \phi, \textbf{y})} $$

As succinctly put on p. 4 of this similar treatment of the same data and approach, "Trick: in conditional posterior distribution, the parameters that are conditioned on are treated as constants." Using this and the fact that $\lambda, \phi$, and $k$ have independent priors, and the prior for $k$ is a uniform constant, we now have a slight variation on your first equation:

$$ p(k|\lambda, \phi, \textbf{y}) \propto p(\textbf{y}|\lambda, \phi, k) $$ $$ p(k|\lambda, \phi, \textbf{y}) \propto \Pi_{i=1}^{k}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!} \times \Pi_{i=k+1}^{N}\frac{e^{-\phi}\phi^{y_i}}{y_i!} $$ Because $\phi$ and $\textbf{y}$ are known, we can divide both sides by $\prod_{i=1}^{N}\frac{e^{-\phi}\phi^{y_i}}{y_i!}$ and maintain proportionality.

$$ p(k|\lambda, \phi, \textbf{y}) \propto \frac{\prod_{i=1}^{k}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!} \times \prod_{i=k+1}^{N}\frac{e^{-\phi}\phi^{y_i}}{y_i!}}{\prod_{i=1}^{N}\frac{e^{-\phi}\phi^{y_i}}{y_i!}} = \frac{\prod_{i=1}^{k}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!} \times \prod_{i=k+1}^{N}\frac{e^{-\phi}\phi^{y_i}}{y_i!}}{\prod_{i=1}^{k}\frac{e^{-\phi}\phi^{y_i}}{y_i!} \times \prod_{i=k+1}^{N}\frac{e^{-\phi}\phi^{y_i}}{y_i!}} $$

The rightmost products in the numerator and the denominator cancel, and we've solved for for $p(k|\lambda,\phi, \textbf{y})$ up to a proportionality constant:

$$ p(k|\lambda, \phi, \textbf{y}) \propto \frac{\Pi_{i=1}^{k}\frac{e^{-\lambda}\lambda^{y_i}}{y_i!}}{\Pi_{i=1}^{k}\frac{e^{-\phi}\phi^{y_i}}{y_i!}} = e^{k(\phi - \lambda) + log(\frac{\lambda}{\phi})\sum_{i=1}^ky_i} $$

To yield a proper conditional posterior, we normalize over $k$, giving the author's formula.

$\endgroup$
  • $\begingroup$ Thanks, this is very useful. Could another way of looking at this be the following (?)... Engineer a likelihood function $p(k|\lambda,\phi,y)$ such that it is maximum just when $i = k$, the point where the change point happens? $\endgroup$ – broccoli Aug 17 '14 at 14:50
  • $\begingroup$ Quite welcome! I'm not quite sure I follow your alternate view. The actual change point is never observed, so we can't really compare the posterior for $k$ to an actual value. We can view the posterior marginal for $k$ as the best compromise between the prior and the data, and in some cases a maximum likelihood estimate will be equivalent to the mode of the posterior using an uninformative prior. Is that helpful, or is your question something different? $\endgroup$ – Sean Easter Aug 17 '14 at 15:55

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.