2 Observations and data reduction

The ESO spectra were collected at the ESO 0.52-m telescope with the Heidelberg fiber-linked echelle spectrograph FLASH/HEROS (cf. Kaufer et al. 1996 and references therein). The spectra have a spectral resolution of about 20000 with a fiber diameter of 100 $\mu$m, corresponding to 2$^{\prime\prime}$  projected on the sky. The reduced spectra have been binned to 0.1Å/pixel. The low state spectrum was obtained on 20 June 1992, exactly on the center of the 1992 spectroscopic event, and covers the spectral range 4040 - 6790Å. The low state spectrum is a combination of seven spectra in the blue (3450 - 5540Å) and eight spectra in the red (5800 - 8600Å). We averaged the 1995 spectra to enhance the signal-to-noise ratio (S/N), as no variability was detected at the time spanned by those observations.

LNA spectra were collected at the Coudé focus of the 1.6-m telescope with a $1152\times750$ GEC/CCD; the dispersion was 0.4Å/pixel and the measured fwhm of the comparison lamp was 0.7Å. The slit width was close to 1$^{\prime\prime}$. Several spectra were collected in the 3600 - 11000Å  wavelength range, but only the subset of data complementary to the ESO (yellow and near-infrared) spectra are presented here. The low state spectra were collected on 3 June 1992 and the high state spectra on May 1995 at LNA observatory. Spectra collected on 20, 24, and 25 June 1992 show that some lines hadn't yet reached the very minimum in 3 June 1992. However, the differences are small and the spectral coverage at LNA was much more complete in 3 June 1992 than on mid-June, allowing spectra collected on that date to represent the low excitation phase to a reasonable approximation.

We degraded ESO spectra to 0.4Å/pixel in order to compare it with that of LNA, collected in the same epochs. No measurable differences were found in the line profiles. Low resolution spectra collected at LNA on 24-31 May 1992 and 15-16 June 1992 (not reported here) confirm that that [Ne III]$\lambda$3868 line was absent on the low state spectrum. This is important information, as no high resolution spectrum shorter than 4000Å was collected during the 1992.5 spectroscopic event.

ESO spectra were reduced with the MIDAS package and that of LNA with IRAF. All line measurements were performed with the IRAF package installed on CASA computers (University of Colorado). After normalization of the spectrum to the continuum level, measurements of isolated lines and individual components inside blends were performed in the standard way. Extracting line parameters from $\eta$ Car's spectrum is not a trivial task, because of asymmetries in the broad line components and lines with extended wings. Gaussian fit was adopted in all cases except for too complicated blends. For these, only the total flux is meaningful and integration under the normalized line profile was used. When measuring the low state spectrum, we always overplotted the high state spectrum, and vice-versa, in order to adopt the same criteria for both. Many training trials were done until a high degree of repeatability in the results was attained. These steps were crucial, as the criteria had to be kept unchanged for months, during which many thousands of measurements were performed. The data presented here have a high degree of internal consistency. When comparing our measurements with that of other authors, however, care must be taken, as assumptions regarding the beginning and ending of the spectral features and the level of the continuum are likely to be different from one author to another.