Assured of a reliable method of solving the normal equations, we now consider
the equations of condition.
The observed magnitude *m* of a star at *M* air masses
with true instrumental magnitude *m*_{0} is modelled as ([10], [28])

where

Here

On general numerical-analysis grounds, we initially assume the best pair of
passbands to use for *C*_{0} and *R* to be the pair that flank the passband in
which *m* was observed.
Thus, for the B band of UBV, we use the (U-V) color.
This is not conventional practice; however, it provides a general rule that
works reasonably well for all photometric systems.
Such a rule is required in a general-purpose reduction program.
For a well-sampled system, this produces a strikingly accurate representation
of the extinction correction [30].
For undersampled systems,
we can use a somewhat more accurate gradient estimator, using a linear
combination of the two adjacent color indices, as described above
(see subsection 13.5.4, ``Choosing a gradient estimator'').
For bands at the extreme wavelengths of our system, we adopt just the
neighboring color index (e.g., (U-B) for the U magnitude and (B-V) for the V
magnitude of UBV).

Regarding the use of (U-V), we may note Bessell's remark [2] (in connection with the problems created by the wide and variable B band) that ``in retrospect, it would have been much better had (U-V) rather than (U-B) been used by Johnson in his establishment of the UBV system.'' In any case, there are special problems in UBV, both due to the inadequate definition of the original system, and the neglect of the transformation from inside to outside the atmosphere entirely in the (U-B) index, which made the ``B'' of (B-V) and the ``B'' of (U-B) have different effective wavelengths; in principle, it is incorrect to add the two indices to a V magnitude and come up with a ``U magnitude'' as a result. Consequently, one must be very careful in doing UBV photometry, no matter how it is reduced; the only safe course is to observe at least the 20 standard stars recommended by Johnson (more would be preferable), and to look very carefully for systematic trends in the transformation residuals.

Although, for astrophysical reasons, there may be partial correlations of the true stellar gradient at a given band with color indices that are remote in wavelength from the band in question, such correlations will depend on metallicity, reddening, and other peculiarities of individual stars. If correlations obtained from one group of stars are applied to another group, the results may easily be worse than if the additional correlation had been ignored in the first place. Thus, it is exceedingly dangerous to employ distant color indices unless the calibration and program stars are very similar in all these respects. We cannot rely on such good matching in general, so these partial correlations are not used in the present package. However, one can expect that very good results will be obtained if the program and extinction/standard stars cover the same region of parameter space. That is, they should have the same ranges of spectral types, reddening, chemical composition, etc.

Because filter photometers observe different bands at different times, we have
to do the reduction in terms of magnitudes (which are measured),
rather than colors (which are not -- see [10],
pp. 152 - 154).
This also allows isolated observations of extinction stars in a single filter,
or a partial subset of filters, to be used.
Furthermore, the best estimate of the extra-atmospheric color *C*_{0} is used to
reduce every observation, so that errors in individual measurements have the
least effect on the reduction of each particular observation.

Of course, it may be necessary to allow the extinction *A*_{0} to be a function
of time; we always solve for individual values on different nights, as many
careful investigations have demonstrated the inadequacy of ``mean extinction''.
The instrumental zero-point terms *Z* may likewise be functions of time,
temperature, or relative humidity.
Finally, the zero-points may be kept the same in each passband for all nights
in a run if the instrument is sufficiently stable.