Consider a random variable *x* taking discrete values $$x \in \{x_1,x_2,\ldots,x_n\},$$ with probability mass function $$p(x_i).$$ If all the possible values of *x* are positive, it turns out that the following is also a valid probability mass function (all positive, sums to 1): $$q(x_i) = p(x_i) \frac{x_i}{E[X]},$$ where $$E[X] = \sum_i x_i p(x_i)$$ is the expected value of *x*.

In fact *q*(*x*) is a posterior distribution for an “intensity” channel. Suppose we have a binary channel output $$y \in \{0,1\},$$ with *x* as the channel input. Further suppose, given *x*, we have $$\mathrm{Pr}(y = 1 \:|\: x_i) = k x_i$$ for some constant *k*; that is, the probability of turning *y* “on” is proportional to the input *x*. Then $$\mathrm{Pr}(y = 1) = \sum_i p(x_i) k x_i = k E[X]$$ and $$\Pr(x_i \:|\: y = 1) = \frac{p(x_i) k x_i}{k E[X]} = q(x_i).$$

I call this an “intensity” channel because the intensity of the input *x* is the key parameter in the input-output relation. A physical example would be a light sensor, where *x* represents the intensity of the light source (i.e., density of photons) and *y* = 1 represents the arrival of a photon at the sensor.

I feel like distributions of the form of *q*(*x*) must have a name and must be well studied. Drop me a line if you know more.

### Like this:

Like Loading...

*Related*

Joel Gouse

Precisely what I was looking for, thankyou for putting up.