Sampling in multivariate distribution with constraints on the sum of the result

219 Views Asked by At

I am trying to use a multidimensional Gaussian to sample material compositions in their probability distributions.

Let's assume I have a material composed of aluminum, carbon and hydrogen. The average composition with associated std deviation is:

  • Alu: average 44 %, std dev 19.6 %
  • Carb: average 4.5 %, std dev 13 %
  • Hydr: average 51.5 %, std dev 24.8 %

I construct a vector mu =[44, 4.5, 51.5] and a covariance matrix (I have some measurement to reconstruct the covariances)

sigma = [ 384.2 -199.2    -216]
        [-199.2    169     -68]
        [ -216     -68  615.04]

I am performing a sampling with n=2000 with computer script. When I compute the mean and std dev of the 2000 samples, I found back again the correct mean and std dev. However, one constraint is that the sum of alu, carbon and hydrogen is 100 %. So, I renormalize for the sum to be 100 % but this has the effect to change my std deviation.

I illustrate my pb with an example, but my question is more the following:

How can I sample in a multidimensional gaussian, knowing mu and sigma, but adding a constraint on the sum of my output vector?

Many thanks for your help