ARMA(p,q), given p,q, then it generally uses m.l.e. to estimate the model parameters(phi_hat,theta_hat, and white noise variance),  probably(you should check maybe), I think they use the R base built-in optim() to get the nonlinear optimization solution. In optim(), if flag returns 1, which means that the algorithm does Not converge, in another words, it's not the m.l.e. estimate you want.

Hope this helps.