Distinguishing two distributions In my application I'm receiving a continuous stream of 50 changing numbers, mostly these samples are contained with a nice Gaussian distribution, but now and then there are two+ distributions within the samples. The samples are always positive and within the range of 0 to 1. If there is more than one distribution present, the distribution I'm always interested has the lowest mean. 
So ideally I want to identify the lower bound and roughly the mean of the main distribution. The lower bound is straight-forward, its always the minimum value of the samples, but I'm stumped as to how find roughly the mean of the main distribution when there is more than one distribution present.
The really tricky thing is that its a real time system and I have very little time to process the samples before the next set arrives, so other than finding the min, max, mean and at a push median I don't have time to carry out much more processing.  
Anyone have any suggestions for how I could achieve my goals?! Many thanks and sorry for the stats-newbie nature of my question 
 A: When you have a sample where different points potentially come from different distributions, the data generating distribution is often envisaged as a finite mixture. That is, the data comes from a density
$$ f(x) = \sum_{j=1}^{M} p_{j} f_{j}(x) $$ 
such that $\sum p_{j} = 1$ where each $f_{j}$ is a density. Typically in finite mixture modeling the purpose is to estimate the number of mixture components, $M$, under some assumptions about the constituent densities, $f_{j}$. Also, posterior probabilities that particular observations came from component $j$ are usually of interest, giving a means for classifying the observations. 
In the case you described you know a priori there are only two groups, so one could imagine observed data comes from a density $f$ such that 
$$ f(x) = p f_{1}(x) + (1-p) f_{0}(x) $$ 
That is, $p \in (0,1)$ proportion of the samples come from a population with density $f_{1}$ while the rest come from a density $f_{0}$. It may suffice to model your observed data as a mixture of Gaussians with different means (despite them being bounded between 0 and 1). That and more complex procedures related to mixture models can be found in the FlexMix package in R. Here is some documentation: 
http://cran.r-project.org/web/packages/flexmix/vignettes/flexmix-intro.pdf
The package mixtools also appears to be pretty popular as well: 
http://www.stat.psu.edu/~dhunter/papers/mixtools.pdf 
although I've never used it
