Hi Jim,

It's different. The FWHM (or rms or rmm) blur size is used to
define the width of a gaussian kernel that, when centered over
(around - in 3D) an output voxel, is used to compute the value
placed there. The relative height of that kernel over all of
the surrounding voxels represents the fraction of that voxel's
value that is applied to the output voxel. Note that these are
not just immediate neighbors, but neighbors close enough where
the gaussian curve is above some small threshold.

So aside from going toward the limits (of 0 and infinity), a
small change in the kernel width should have a likewise small
change in the result.

---

AFNI measures smoothing using 3dFWHMx (or perhaps 3dBlurToFWHM,
which will blur not _by_ 6 mm, say, but will blur until the
estimated blur _is_ 6 mm). These programs estimate the blur
(along each axis) using the ratio of the variance of the first
differences to the overall data variance. It is important to
apply a mask for this computation.

When computed over many sub-bricks (such as in EPI data), these
values are averaged.

Note that the afni_proc.py script has new options to include this
computation after running 3dDeconvolve (-regress_est_blur_errts
and -regress_est_blur_epits). These values can then be averaged
across subjects and applied to all AlphaSim computations (assuming
no more blurring is done :).

- rick