Shaping probability distribution function

[SunnyBoyNY]

Homework Statement

Incoming signal has normal distribution, xmin is equal to -sigma, xman is equal to +sigma. What is the governing equation of the nonlinearity through which the signal has to be passed in order to make its pdf uniform?

Homework Equations

http://en.wikipedia.org/wiki/Normal_distribution

The Attempt at a Solution

I have already found out that the signal needs to be passed through erf(x/sqrt(2)), which is very relevant to the CDF of normal distribution. The problem is that I cannot find a mathematical proof.

 

[Ray Vickson]

Homework Statement

Incoming signal has normal distribution, xmin is equal to -sigma, xman is equal to +sigma. What is the governing equation of the nonlinearity through which the signal has to be passed in order to make its pdf uniform?

Homework Equations

http://en.wikipedia.org/wiki/Normal_distribution

The Attempt at a Solution

I have already found out that the signal needs to be passed through erf(x/sqrt(2)), which is very relevant to the CDF of normal distribution. The problem is that I cannot find a mathematical proof.

It's very easy. Say you have a continuous random variable X with a strictly increasing cumulative distribution F(x) on an x-interval [a,b] (possibly a = -∞ and b = +∞). The probability density of X is f(x) = (d/dx) F(x). Now look at Y = F(X); that is, for each observation x of X we let the observation of Y be y = F(x). What is the distribution of Y? For x < X < x + dx the probability is f(x)*dx, so if x <--> x and y + dy <--> x + dx, we have
P{y < Y < y + dy} = f(x)*dx. If g(y) is the probability density of Y, we therefore have g(y)*dy = f(x)*dx. But dy/dx = (d/dx) F(x) = f(x), so dy = f(x) dx, hence we must have g(y) = 1; that is, Y is uniform on (0,1).

All this is very standard in Monte-Carlo simulation, where it is used to generate samples from non-uniform distributions: we generate Y uniform on (0,1), then obtain our sample of X from x = F^-1(y) (at least, in those cases where the latter function is known and not too hard to compute).

RGV

 

[SunnyBoyNY]

RGV, thanks for your feedback. I believe I have already proved it using a non-linearly distributed substitute variable "a" as opposed to the linear variable "x". Those two variables could be unambiguously mapped from one to the other to warp the "x" axis to ultimately render the pdf uniform.

I dare to say I understand your reasoning here, I will just need a little bit of time to absorb it.

The integral of (pdf):

<br /> f(x) = \frac{1}{\sqrt{2\pi}\sigma}e^{-\frac{1}{2}(\frac{x}{\sigma})^2}<br />

is (cdf):

<br /> \frac{1}{2}erf(\frac{x}{\sqrt{2}\sigma})<br />

Which is a normalized function with zero mean. I assume the mapping function needs to be multiplied by \sigma to achieve the desired uniform mapping. This is just a detail.

Quite interesting info about Monte-Carlo simulations. It would make sense to transform a linear distribution to resemble other kinds of distributions. I wonder whether MatLab does it the same way.

 

[jbunniii]

Quite interesting info about Monte-Carlo simulations. It would make sense to transform a linear distribution to resemble other kinds of distributions. I wonder whether MatLab does it the same way.

I'm not sure how Matlab does it, but a common way to generate Gaussian (normal) random numbers from uniform ones is the following trick: if u_1 and u_2 are independent random variables, uniformly distributed over (0,1], then
n_1 = \sqrt{-2 \log(u_1)} \cos(2\pi u_2)
and
n_2 = \sqrt{-2 \log(u_1)} \sin(2\pi u_2)
are independent Gaussian random variables with zero mean and unit variance. This is the so-called Box-Muller transformation:

http://en.wikipedia.org/wiki/Box–Muller_transform

 

[SunnyBoyNY]

jbuniniii,

Thanks for the info about the transform. That's why I love this forum. One question brings together many ideas and perspectives.