The key step in the extraction operation is the determination of the
centroid of each echelle order to subpixel accuracy. In principle, the
global shifting operation of *ORDERG*, the results of which are
incorporated into the high-dispersion SI, accomplishes this same
function, but this routine does not always provide the shifting accuracy
needed (e.g., for images having weak continua and for extended sources).
The input for this refined centroid operation is provided by the
high-dispersion SI and SF, the high-dispersion noise model and a set of
fiducial line positions for each order which are given in
Table 10.1.

The centroid order positions in Table 10.1 were determined empirically
for a collection of well-exposed images for each camera. In each case
the separations of the initial determinations were compared with echelle
grating theory (i.e., they should vary as 1/*m*_{ech}^{2}). The final
positions show smooth undulating departures from this distribution
amounting to a few tenths of a pixel in most cases.

For each of the echelle orders, fluxes are extracted from a cut in the spatial direction of the high-dispersion SI, summing the fluxes of all ``good'' (i.e., nonflagged) pixels in a swath between sample positions 150 and 450. The summed fluxes for each spatial line are normalized to a common number of contributing pixels to account for the exclusion of flagged pixels. Next, an rms scatter is computed after excluding points in the spatial profile with significant flux above an initial rms value calculated by including all profile points. A local background is also fit through the low-flux points and subtracted from the profile array. A gaussian model is then cross-correlated through the net-flux profile if there are at least two points with fluxes above the rms value. A default value (i.e., the appropriate fiducial value from Table 10.1) is assigned for the order's centroid position according to any of the following conditions:

- fewer than two spatial profile points have fluxes above the local rms level,
- the centroiding cross-correlation coefficient is unsatisfactory,
- the centroid determined by cross-correlation is beyond a tolerance value (ranging from ±&nbns;0.5 to ±&nbns;3.0 pixels for short- and long-wavelength orders, respectively) away from the fiducial value.