2
$\begingroup$

I am trying to model a joint probability constraint of a union of correlated / uncorrelated exponential functions.

Formally, assume we have random variables $A_i \sim exp(\theta_i), \forall i \in {1…n}$. The rate of each exponential distribution is a random variable itself (suppose we have a general distribution, G) thus $\theta_i \sim G_i(.)$. The objective is to get a closed form expression for the following chance constraint:

$Prob( \bigcup_{i=1}^{i=m}(A_i > s_i )) \leq \epsilon$, $\forall s_i$ constants

  1. If we assume that random variables $A_i$ are independent then the joint constraint becomes :$\sum_{i=1}^{i=m} Prob(A_i \geq s_i) \leq \epsilon $. In that case each $ Prob(A_i \geq s_i)$ can be modeled as a mixture of exponential distribution. If $G_i(.)$ is a discrete probability distribution, in that case we could easily obtain a closed form expression for the constraint . (Is this correct?)

  2. If variables $\theta_i$’s are correlated, I am stuck on what could be possible. If I don’t model dependence and assume independence in that case I guess I could get an upper bound on the joint constraint. Also, I was wondering if assuming linear dependence between $\theta_i$’s would make sense or not. Hence I could pose $\theta_i = a_i*\lambda + b_i$ with $a_i$ and $b_i$ different linear coefficients, and $\lambda$ a common random variable. I am again confused whether this would make sense or not. I have never used copula’s but I guess they could be useful in that case (Any comments appreciated)

I would appreciate your comments.

$\endgroup$
  • $\begingroup$ First of all you made an incorrect statement which is the basis of your argument. It is NOT TRUE that independence of events implies P(UA$_i$)=∑P(A$_i$). That holds if the events A$_i$ are dsijoint. So I think you need to reformulate your question. $\endgroup$ – Michael Chernick Aug 26 '12 at 1:14
2
$\begingroup$

The union bound is true whether the variables are independent or not. The inequality

$$P(\bigcup_i (A_i \gt s_i)) \le \sum_i P(A_i \gt s_i)$$

does not require the $A_i$ to be independent, so if that's good enough then you have what you need already. If you assume independence, you get an equation,

$$P(\bigcup_i (A_i \gt s_i)) = 1 - \prod_i (1-P(A_i \gt s_i)) $$

but in many situations where you are trying to bound the probability by something small, you might use the union bound anyway even if you know the variables are independent. $\sum_i P(A_i \gt s_i)$ is simpler than $1 - \prod_i (1-P(A_i \gt s_i))$ and when the probabilities are small they don't differ by much .

$\endgroup$

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.