2.1 Infinite fit function diagnostics

From Mike Neale:

For 2.1 I think error messages need a lot of work. Perhaps the most dreaded is the infinite fit function. Under most circumstances, we could do much of the diagnostics - or at least tell the user which R commands to run to diagnose the problem. Internally we should be able to figure whether 1) it is at the starting values that the problem occurred; 2) if a non-positive definite covariance matrix was involved. The second is much more difficult since we may have non-pd only for certain rows of raw data. Keeping the likelihoods around (so that those that are NA or less than SAFELOG where this is the smallest number safe to take the log of without underflow are easily tracked.) At least suggesting R commands along the way — look at eval(parse(text=min(eigen(yourModelObject$expCov$result)$values))) - if this is less than zero then your expected covariance matrix is not positive definite. for example, would go a long way to helping the user past such problems. I note that because $values applies to matrices, and $result to algebras it isn’t so easy to generalize this type of support, but we should still try.