The Walsh Hadamard transform of uniform random numbers.

And also random projections from sign flipping.

After filling an array with random numbers from the uniform distibution go pair-wise

through the array to get an x and y coordinate to plot. Do the same for the

WHT transformed random data. Notice that the WHT transforms the data into the

Gaussian distribution due to the central limit theorem, which applies not only

to sums of random numbers but also sums and differences as found in the WHT.

Natural images when transformed by the WHT have a small number of high magnitude

coefficents and a large number of low magnitude coefficents. Randomly sign flipping

the data in the natural image make all the WHT outputs have a Gaussian distibution.

This makes a very fast random projection algorithm with many uses (eg. locality

sensitive hashing, neural networks and associative memory.)

You can improve the quality of the random projections by repeating the sign flip, WHT

process a number of times or by including random permutations for even higher

random projection entropy.