Introduction

The availability of immediate, on-line reduction for radial velocities at the ELODIE echelle-spectrograph of the Observatoire de Haute-Provence, led to the obvious idea that it would be very useful to extend this type of fast analysis to other stellar parameters. Being ourselves involved in the determination of the metallicity of stars of known proper motions in two galactic directions (Soubiran 1992, Ohja et al. 1994, Perrin et al. 1995) we were tempted to try to determine on-line the metallicity, effective temperature and gravity of a star observed with ELODIE. Our former experience with this type of determination from spectra at a lower resolution from a grid of synthetic spectra ( Cayrel et al. 1991a & b, Perrin et al. 1995) allowed us to estimate that for old stars (with slow rotation) of solar type or cooler, a mean accuracy of 200 K in $T_\mathrm{eff}$, 0.4 dex in gravity and 0.3 dex in metallicity was possible from a spectrum obtained with a signal to noise ratio (S/N) of 50 on a limited spectral interval. From theoretical considerations (Cayrel 1991) we thought that with the higher resolution and larger spectral interval of ELODIE's spectra, it would be possible to obtain comparable or better precision on low S/N ($\sim$ 10) spectra.

We could have followed the same approach as before, and tried to compare the observed spectrum with a grid of synthetic spectra, generated for example with the ATLAS9, SYNTHE (Kurucz 1993) codes made generously available to us by R.L. Kurucz. However we thought that there was a simpler method, i.e. comparing the spectrum of the target star with a library of spectra of reference stars, with accurately determined parameters, taken with the same spectrograph. The disadvantage is that an error on the parameters of reference stars affects the result. Also the optimal extraction of the parameters possible with a grid of synthetic spectra, where the sensitivity of any spectral feature to the parameters is known explicitely, is lost. The advantage is that a very large spectral range can be used, without the very tedious work of fine-tuning the oscillator strengths and damping constants of an extremely large number of atomic or molecular lines. This approach would also avoid the empirical corrections of synthetic to reference star spectra that have proven to be necessary (Cuisinier et al. 1994).

The principle sounds very simple. But, before the target spectrum can be meaningfully compared to the spectra of the library, many steps are required. First of all it is necessary to remove all features which are not specific to the object, but instrumental in nature, or associated to particular conditions of observation (Sect. 3). The main instrumental feature is the modulation of the spectra by the blaze profile of each order. If the two objects to be compared had the same radial velocity with respect to the instrument, at the time of the exposures, this modulation would cancel out. But most of the time there is a significant difference, and the blaze efficiency is shifted in wavelength between the two exposures. This modulation must be corrected (Sect. 3.2). Cosmic rays are a big nuisance in all instruments using CCDs as detectors. A first treatment is made in the radial-velocity software (Baranne et al. 1996), following Horne's algorithm (1986). Unfortunately many cosmics escape the trap, because too few pixels are illuminated along the direction perpendicular to the dispersion. Therefore the remaining cosmics must be chased (Sect. 3.3). Equally disturbing are the telluric lines which change in intensity and position with time in the rest frame of the object. The pixels affected by these wandering disturbers must be eliminated (Sect. 3.6). If the night is spoiled by the Moon, it may be necessary to subtract a sky exposure, for which all the steps already described (except telluric line removal) must be carried out too (Sect. 3.4).

After these different steps, three other actions remain to be performed. Two spectra to be compared must be brought :

(i) $\ \$ to the same spectral resolution
(ii) $\ \$ to a common wavelength scale
(iii)$\ \$ to a common level of flux

If the target star and a reference star of the library have a different line-broadening, because they have a different projected rotational velocity $v\sin i$, we must not be fooled into considering them as objects of different effective temperature, gravity and metallicity. Also it is not guaranteed that the instrumental resolution is exactly the same for all observing runs (e.g. because of focus variations). Therefore all the spectra of the library and the target's star are forced to a common resolution (action (i) listed above), in order to eliminate spectral differences originating from projected rotation, macroturbulence or instrumental resolution (Sect. 3.5).

Action (ii) is easy to perform because the radial velocity of the target star and of each comparison star is accuratly known from the radial velocity software (Baranne et al. 1996), by cross-correlation with a mask (Sect. 4.1). Action (iii) is done by least squares (Sect. 4.2) .

A large fraction of our observing time was devoted to the aquisition of the library of reference stars, which includes 211 spectra at the present time. Sect. 2 describes the observational material. A detailed description of the library is available in a companion paper (Soubiran et al. 1998, hereafter paper II). The spectra of this library are available at CDS of Strasbourg, together with their revised atmospheric parameters (Sect. 5).

Different tests have been performed to evaluate the consistency and the accuracy of the method (Sect. 6).

The software is named TGMET, for Temperature, Gravity, METallicity.