The lamps used for flat fields have an uneven spectral emissivity.
Combined with the uneven spectral sensitivity of the spectrograph and detector,
this results in flat fields which usually have very different intensities
in different spectral regions.
Therefore a direct flat-fielding, *i.e.*, division by
the original dark subtracted
flat-field, produces spectra which are artificially enhanced in some parts and
depressed in others. In principle, this should not be a problem because the
instrumental response curve should include these variations and permit them
to be corrected. However, the response curve is established by integration
over spectral intervals of small but finite width. At this stage, strong
gradients obviously introduce severe problems. Also the wish, at a later stage
to evaluate the statistical significance of features argues strongly against
distorting the signal scale in a basically uncontrolled way.

Therefore, it is better first to remove the low spatial frequencies
along the dispersion axis (but not perpendicular to it!) from the
flat-field, using the command `NORMALIZE/FLAT`. This command
first averages the original image along the slit (assumed to be along
the Y-axis) and fits the resulting one-dimensional image by a
polynomial of specified degree. Then the fitted polynomial grows to a
two-dimensional image and divides the original flat-field by this
image to obtain a ``normalized'' flat-field.

In some cases, *e.g.*, EFOSC in its B300 mode which gives a very
low intensity flat-field at the blue end, the polynomial fit is not
satisfactory and it is advisable to do the sequence manually. Instead
of fitting a polynomial one could fit a spline using the command `INTERPOLATE/II` or `NORMALIZE/SPECTRUM`. The latter belongs to the
low-level context `SPEC` and is interactively operated on a
graphical display of the spectrum.