Recently, while I was working on a screen-space shader effect, I had to do some random sampling over the surface of a sphere. An effective sampling requires a uniform distribution of samples. After a quick googling, I found out a way to generate uniformly distributed samples(), and it showed a decent result for my application. But, still unsure if that was an ideal way, I performed a due research about it later. Following is the result of that short research.
Usually, in graphics application, one can limit it to the three-dimensional space. In that case, there are four possible approaches, all of which guarantee a uniform distribution(BTW, as for what the ‘uniform distribution’ exactly means,  has some explanations). If a n-dimension support is required, one is out, so three remain. Let’s take stock of each.
Rejection sampling ()
One simple way is something called ‘rejection sampling’. For each x, y, z coordinates, choose a random value of a uniform distribution between [-1, 1]. If the length of the resulting vector is greater than one, reject it and try again. Obviously, this method can be generalized to n-dimension. But the bigger the dimension gets, the higher the rejection rate gets, so the less efficient the technique becomes.
Normal deviate ()
This technique chooses x, y and z from a normal distribution of mean 0 and variance 1. Then normalize the resulting vector and that’s it.  shows why this method can generate a uniform distribution over a sphere. In short,
It works because the vector chosen (before normalization) has a density that depends only on the distance from the origin.
as  explains. This also generalizes to n-dimension without a hassle.
Trigonometry method ()
This one works only for a three-dimensional sphere(called 2-sphere in literatures, which means it has two degrees of freedom), but is an easiest one to intuitively grab how it works.  nicely explains why it works from Archimedes’ theorem:
The area of a sphere equals the area of every right circular cylinder circumscribed about the sphere excluding the bases.
The exact steps are as below:
- Choose z uniformly distributed in [-1,1].
- Choose t uniformly distributed on [0, 2*pi).
- Let r = sqrt(1-z^2).
- Let x = r * cos(t).
- Let y = r * sin(t).
This is the one I used for my shader effect. Since I had to use a very small number of samples for the sake of performance, I did a stratified sampling with this method. A straightforward extension to a stratified sampling is another advantage of this technique.
Coordinate approach ()
The last one is applicable to general n-dimensions and  explains its quite math-heavy derivation in detail. This technique first gets the distribution of a single coordinate of a uniformly distributed point on the N-sphere. Then, it recursively gets the distribution of the next coordinate over (N-1)-sphere, and so on. Fortunately, for the usual 3D space(i.e. 2-sphere), the distribution of a coordinate is uniform and one can do a rejection sampling on 2D for the remaining 1-sphere(i.e. a circle). The exact way is explained in  as a variation of the trigonometry method.