In each iteration the coefficients are given as output, and the cost, based on some constraints (add to 1, cannot be 0, coefficient less than domain , numeric it is less than max value....)
Every iteration is better than the previous and it leads to convergence.