Bimodal MaxEnt distributions? What kind of constrains give rise to bimodal distributions in the Maximum Entropy formalism? Are there any known results in this topic?
 A: This might be related to the support of the constraints. With the following notation: let X be a discrete random variable on states, let P be a probability distribution with P(X=x_i)=p_i for i=1,...,n where n=length(states), and let H(X)=-sum(p_i*log(p_i) for i=1,...,n) be the entropy of X.
Define constraints by:
A_j: R^n -> R 

as a column vector of length n, j=1,...,m and demand that the conditions
A_j %*% X == bvec[j]  for j=1,...,m 

are satisfied.
Now if the constraints are supported on different subsets of states, a bimodal distribution may arise. For example:
n <- 40
p <- 0.5
mu <- n*p
A_1 <- c(1:n, rep(0,n))
A_2 <- c((1:n-mu)^2, rep(0,n))
A_3 <- c(rep(0,n), 1:n)
A_4 <- c(rep(0,n), (1:n-mu)^2)
bvec <- c(mu, n*p*(1-p), mu, n*p*(1-p))/2

The first two conditions result in a binomial distribution on 1:40 and the second two conditions result in a binomial distribution on 41:80. The maximum entropy distribution on states=1:80 has then two modi:

I believe this generalizes to continuous random variables.
