Gamma function's mean and standard deviation through shape and rate

17.5k Views Asked by At

I'm currently learning how to program neural networks. In the most basic ones, you initialize your matrix with random numbers that are distributed using Gamma function.

I'm using a math library, that lets me initialize a Gamma function by giving values for shape and rate.

The most basic way to initialize it for the neural networks is to do it with a Gamma function with mean 0 and standard deviation 1. I'm struggling on how to connect these properties though.

What I have already learned, is that I can obtain my desired mean value with shape and rate like this:

$μ = k/β$

where k is the shape and β is the rate.

So far, i couldn't find / understand how to get a desired standard deviation though.

Extra: If you could explain using 1yr college math (if possible), that would be amazing :)

1

There are 1 best solutions below

1
On BEST ANSWER

A gamma distribution has a strictly positive mean. If $X$ is gamma distributed with shape $a$ and rate $b$, then the mean of $X$ is $$\mu = \operatorname{E}[X] = a/b,$$ and the standard deviation is $$\sigma = \sqrt{\operatorname{Var}[X]} = \sqrt{a}/b.$$ Note that $a$ and $b$ must be positive.

It follows from the above that, given a desired mean $\mu$ and standard deviation $\sigma$, the shape and rate that produce a gamma distribution with that desired $\mu$ and $\sigma$ are: $$a = (\mu/\sigma)^2, \quad b = \mu/\sigma^2.$$

For example: If the desired mean is $\mu = 5$ and the desired standard deviation is $\sigma = 2$, then the shape would be $a = (5/2)^2 = 25/4$, and the rate would be $b = 5/2^2 = 5/4$.