Proportions instead of differences for delta models

[vadimnazarov]

This could be more efficient for pan length models.

To overcome problems with cumulative errors we can start proportions from the middle positions to both directions.