My group is working on related structures collected back-to-back with the same microscope configuration. One lab member processed with Relion, and the other with cryoSPARC (downsampling from 512 to 360). When refining the pixel size of the two reconstructions vs. an independent crystallographically-obtained atomic model, there is a difference in pixel size of 0.3% which happens to correspond to the error if the box dimension was off by one resampled pixel. I am reminded of a 1990s problem in crystallography with different programs assuming that maps were “open” or “closed” (last pixel = 1st).
The downsampling was performed by Fourier cropping that implicitly makes the volume periodic. Say that we have a pixel length of 1A and a box dimension of 100A, then we either have 100 unique grid points (pixels) for an open cell running 0 to 99 at 0A to 99A (not 0A to 100A), or we have 101 grid points of 1A separation for a closed cell, running from 0A to 100A, where the 0th point is repeated as the 100th. I believe that most FFT routines expect input of the unique / open cell. If so, the inverse FFT should be mapped back to a downsampled open cell. So, if we were downsampling this example by two, one would map back to real grid points from 0A to 49A. If the new grid was assumed to be running from 0A to 50A (half the original box), then there would be a pixel size error of the magnitude that we see.
Of course, this could be a big coincidence, but I am wondering whether this might have been the origin of interpolation errors discussed a couple of years ago with reference to model-map correlations. Can anyone comment on what is going on under the hood, and whether we should be expecting to make small corrections to the pixel size when downsampling?
Thanks, Michael.