Suppose $y_i$ is a random variable generated by sampling $y_i' \sim \mathcal{N}(0,1)$ and setting $y_i = y_i'$ if $c_1 \leq y_i' \leq c_2$, $y_i = c_1$ if $y_i' < c_1$, and $y_i = c_2$ if $y_i' > c_2$. In other words, take a normally distributed sample $y_i'$ and threshold it at some lower and upper limit to give $y_i$. The "pdf" of $y_i$ then looks like a normal distribution for $c_1 < y_i < c_2$, and is $0$ everywhere else except at $c_1,c_2$ where it is some other value.

Let $Y_j \in \mathbb{R}^T$ be a vector of iid samples $(y_1, \ldots ,y_T)$ where each $y_i$ is distributed in this way, and suppose I have a set of vectors $\{Y_j\}, j = 1, \ldots ,N$.

I want to calculate $p(Y_j)$ = $p(y_{j,1},\ldots ,y_{j,T})$ for each $Y_j$ in a way that I can compare this value for different $Y_j$ in a meaningful way. If there was no discontinuity in the pdf of $y_i$ this would be easy - just multiply the pdf of the iid components: $$ p(Y_j) = \prod_{i=1}^T p(y_{j,i}) $$

I'm confused as to what to do when the pdf of each $y_i$ is not continuous. If I work it out as a product like this, what value do I use for $p(y_i)$ if $y_i = c_1$ or $c_2$? If I use for instance: $$ p(c_2) = \int_{c_2}^{\infty} \mathcal{N}(y_i; 0,1) \mathbb{d} y_i $$

This integral could be much lower than $p(c_2 - \epsilon)$ even though the "probability of seeing $c_2$" is much higher than a value slightly less than $c_2$. When I take the product I think the value for $p(y_i = c_2)$ should be larger than for $p(y_i = c_2 - \epsilon)$, otherwise comparing different $p(Y_j$) is meaningless.

My actual situation is that I have a set of measurements $Y_j$ in a file, and need to calculate $p(Y_j)$ for each.


Not really an answer but worth mentioning:

The rv $y_i$ as described in top of your question has no PDF.

If it would have - let's call if $f$ - then $P(y_i\in A)=\int_Af(y)dy$ should be true for each measurable set $A$.

However, if $\emptyset\neq A\subseteq\{c_1,c_2\}$ then the LHS is positive and the RHS is zero.

  • $\begingroup$ Seems like an answer to me, probably even the answer. $\endgroup$ – Did Oct 3 '14 at 8:41
  • $\begingroup$ What does this mean for $P(Y_j)$? $\endgroup$ – akxlr Oct 3 '14 at 8:47
  • $\begingroup$ To avoid misunderstandings: what exactly is meant by $P(Y_j)$ (or $p(Y_j)$ in the question)? You are saying that you want to calculate it without mentioning what it is. $\endgroup$ – drhab Oct 3 '14 at 8:53
  • $\begingroup$ I mean it to be the probability of observing the set of $T$ measurements $y_{j,1}, \ldots ,y_{j,T}$ - i.e. each $Y_j$ is a vector of $T$ iid measurements and I want to work out the probability of observing all of them assuming they are independent. $\endgroup$ – akxlr Oct 3 '14 at 9:13
  • $\begingroup$ Probability of "observing..." sounds vague. Its probability can only be calculated if "observing..." is some sort of event, i.e. a measurable subset of a probability space. Can you describe this event in clear terms of probability theory? $\endgroup$ – drhab Oct 3 '14 at 9:23

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.