Latin hypercube sampling function: how does iterative minimization of the correlative criterion work?

I am using the Latin hypercube function for an experimental design lhsdesign() with the correlation criterion that iteratively minimizesthe sum of between-column squared correlations, as described in the documentation in https://de.mathworks.com/help/stats/lhsdesign.html.

It is not clear to me how the function iteratively creates the best design that minimizes the sum of between-column squared correlations. It can't be that it tests all possible designs, since the possibilities are too many, and the correlation varies each time I use the function. But then how does the function minimize them? I can't find any documentation or information on this.

Thanks in advance