Th calibration lamp spectra are commonly used at resolutions at which line blending displaces the majority of the measured features from the laboratory wavelength of their principal component by an amount which is large compared to the random noise limited centering accuracy. Earlier, we derived improved wavelengths in a restricted wavelength and resolution domain and obtained through their use more accurate wavelength calibrations. Here we derive improved resolution-dependent wavelengths for the full subset of at most weakly blended Th and Ar features in the whole wavelength and resolution domain of interest. This work should encourage the search for more robust wavelength calibration techniques.
The use of higher order polynomial approximations in wavelength calibrations should be discouraged. First, proper care should be given to the elimination of systematic effects that dominate over random noise. Otherwise, the higher order terms will be determined by the trend in the systematic effects towards the ends of the wavelength range; as a consequence, and in spite of the lower internal scatter with respect to the calibration line positions, the model is then likely to deviate further from reality than a low order polynomial. In fact, the use of appropriate wavelengths strongly enhances the power to discern significant terms from spurious ones.
The elimination of the calibration bias due to blended lines leads to a
better accuracy, even while using less numerous calibration points. When the
calibration relation itself depends on a low number of fit parameters, then
the full advantage of the higher accuracy of the input data can be conserved.
Unfortunately, in echelle spectroscopy
the practice of fitting a polynomial approximation to each order independently
is still common. This leads to many parameters to be determined, i.e.
m.n for a (m-1)th degree approximation on n orders, which is
usually of the order of 102! In general, such approach will prevent to make
the proper line selection (our experience at 
/pixel-scale of
3 to 
gave
densities of selected lines of 7 per spectral order on the average, but
some orders had very few lines). This is however not a fundamental problem,
as we could reduce the number of free parameters by more than a factor 10 once
we took care of the blending effects by deriving an adequate 7-parameter 2-D
relation (Hensberge & Verschueren 1989).
The argument that the use of much more, partly fainter and stronger blended lines would lead to a comparable final calibration accuracy, because the blending effects will statistically largely cancel, appears wrong: first of all, the total number of quasi-unblended weak lines does not increase sufficiently fast with decreasing intensity in order to compensate for the larger random errors on individual lines. Secondly, even when numerous blended lines are included, the line density is insufficient to smooth away significant chance fluctuations in the distribution of the line displacements: large calibration errors at particular locations are to be expected, while the global calibration accuracy might be appropriate for many purposes.
Once the accuracy of the calibration is limited mainly by random errors, further gains can be expected from coupling the calibration from different frames (see Verschueren et al. 1997). This aspect is still under investigation and falls beyond the scope of the current paper.