Why is it counterintuitive? When you look at the plots of the inplane rotations and correction factors you notice that they follow a trend, they are not random. So, if the "estimate" parameter is true, then the inplane rotation and correction factors of the neighboring image are good starting values, especially so for the initial cycle where you set the inplane rotation to zero, for example. What you probably observed is that the optimization algorithm converges faster when the initial values are closer to the ideal values. I would expect that behavior for the first few cycles. For later cycles, when the geometric parameters do not change substantially any more, it might be better to use the stored values from the previous cycle as a starting point for area matching.
The "estimate" parameter only affects the initial geometric values of a particular cycle. If it is true, the inplane rotation and corrections are simply copied from the neighbor image that has just been aligned (in the same cycle). Thinking about it, a better way than just copying the values would be to extrapolate from the already measured corrections from all images at lower tilt that have already been matched.
The choice of the estimated initial values could have an influence on the outcome of the optimization if different local minima would exist and the algorithm gets caught in the wrong one. I haven't found any indications so far that this is happening, though. A slightly different problem that we observed, however, is double or multiple peaks in the correlation function, and the true peak is not necessarily the highest one. This is very obvious when displaying the peaks and is usually related to incorrect values for the backprojection body size.
