1
$\begingroup$

I have the following problem. I have I finite family $\mathcal{F}=\{ P_1,...,P_2 \} $ of discrete probability distributions on the same set $S$. Now I observe experiments from a random variable X which is distributed as one of the probabilities in $\mathcal{F}$. I want to compute the probability that $X \sim P_i$ for each $i$, or an estimate of these probabilities with some thoeretical guarantee.

The real problem I have is that I have to keep a belief on these probabilities and take decisions based on this belief, update it as I observe new samples, and change my decisions accordingly.

The two approaches I thought are with a Dirichlet-categorical Bayesian model, but it is continuous so the probabilities of single probabilities are 0, or with concentration inequalities, but cannot manage to find a solution.

$\endgroup$
0

1 Answer 1

1
$\begingroup$

This seems to be a straightforward application of Bayes' Theorem. Your initial belief on the distributions' probabilities yields your priors $P(P_i)$. Given a distribution $P_i$, you can calculate the conditional probability $$P(X|P_i) = \prod_{j=1}^n P_i(x_j),\text{ where } X=\{x_1, \dots, x_n\},$$ and then you simply update $$ P(P_i|X) \propto P(X|P_i)P(P_i). $$


As an illustration, assume your discrete distributions are binary, and that we have three candidates

$$ \mathcal{F} = \{(0.2,0.8), (0.4,0.6), (0.9,0.1)\} $$

with equal prior probabilities. Assume further that the true distribution is $(0.5,0.5)$, so not one of the candidates, although the second one comes closest (but this doesn't matter). Here is how the posterior probabilities develop as we draw 100 samples $x$ and update the probabilities after each observation:

plot

distributions <- rbind(c(0.2,0.8),c(0.4,0.6),c(0.9,0.1))
nn <- 100
probabilities <- matrix(NA,nrow=nn,ncol=nrow(distributions))
probabilities[1,] <- 1/3

for ( ii in 2:nn ) {
    set.seed(ii)    # for reproducibility
    xx <- sample(c(1,2),size=1,prob=c(0.5,0.5))
    probabilities[ii,] <- probabilities[ii-1,]*distributions[,xx]       # update
    probabilities[ii,] <- probabilities[ii,]/sum(probabilities[ii,])    # normalize
}

plot(c(1,nn),c(0,1),type="n",xlab="Step",ylab="Probabilities")
for ( ii in 1:ncol(probabilities) ) lines(1:nn,probabilities[,ii],col=ii)
legend("right",lwd=1,col=1:ncol(probabilities),
    legend=paste0("(",apply(distributions,1,paste,collapse=","),")"))
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.