Slow convergence, or no convergence

For most of the time, this is due to me being lazy (and mathematically unapt) and not coding a more efficient maxent algorithm. In the rest of the cases, it is either your data, or that the decision made by the program about the F₀₀₀ value was completely wrong (see section 8.1). Now, if the problem is the value of F₀₀₀ being too small, you should be seeing an unjustifiably sharp map. If that is the case, stop the program, edit the MAXENT_AUTO.IN file, add a line with the F000 keyword (see 7.1.4), and re-run the program with GraphEnt MAXENT_AUTO.IN.

If you are calculating a 3D isomorphous difference Patterson function you can improve the convergence properties (hopefully without loosing much of the signal) by giving LIMIT 1.0 or LIMIT 2.0 in the MAXENT_AUTO.IN file and re-running the program (see discussion in section 7.3.9).

If you are calculating a native Patterson function, see page .

You can get problems with slow convergence even when you calculate something as simple as a two-dimensional projection. To my experience, slow convergence indicates the presence of some signal (to take the extreme view, if your data are consistent with a uniform map, GraphEnt will stop immediately), but this is not necessarily the signal you expect, or wish to have¹⁹. In other cases, it indicates the presence of outliers in your data (which makes it difficult to find a solution that satisfies the constraints they impose). This last case is easily identified from the contributions to table (and the normal probability plot in the case of difference Patterson functions).

Footnotes

... have¹⁹

For example, you may have measured very good quality native and derivative data, but if they are not isomorphous, GraphEnt will be fitting their (very accurately) measured differences, the only trouble being that the result will appear to be just noise to us.