Since HeI $\lambda$ 4471 is the most extensively covered absorption line in our data set and since it is also usually the least blended one, we decided to use the RVs of this line to determine a new orbital solution. Other lines are either weak (e.g. HeII $\lambda$ 4542), blended with other weaker absorptions (HeII $\lambda$ 4200 blended with NIII $\lambda$ 4196) or affected by a rather strong emission component (HeI $\lambda$ 5876). These features are thus not well suited to measure the RVs of the binary components.

We notice that the intensities of the absorption lines (including HeI $\lambda$ 4471, see Fig.2 and Nazé et al. 2000) change with phase. This phenomenon is not unexpected since HD149404 is known to display variations of the line strength in the UV that could be related to the so-called "Struve-Sahade effect" (Stickland 1997).

The radial velocities were determined by fitting two Gaussians whenever it was possible to do so. However, the lines remain heavily blended over the main part of the orbit and even around quadrature the lines are not completely separated. Therefore, two different techniques were tested to determine the radial velocities when Gaussian fitting turned out to be impossible. In both cases we use a template line profile built from the spectra at maximum separation.

Given the difficulties due to the severe blending of the lines, we also used another method developed at the Geneva observatory to determine the radial velocities. The velocities are obtained by cross-correlation of the FEROS and Coralie spectra with a mask built from a synthetic spectrum. This method is well known for less massive stars, but this is the first attempt to extend it to the optical spectra of O-type stars. A synthetic O-star spectrum in the 3875-6820 Å spectral domain was generated using the SYNSPEC (Hubeny et al. 1994) code with the model atmospheres interpolated from the Kurucz ATLAS9 (1994) grid. The Vienna Atomic Line Database was used to create a line list for the spectrum synthesis (Kupka et al. 1999). The spectrum synthesis programme uses an LTE-model which is certainly not appropriate to determine the properties of O-stars, but should be sufficient for the radial velocity determinations. Many tests were conducted and eventually, a synthetic spectrum with $T_{\rm eff} = 42\,500$ K was adopted because it provided the most narrow cross-correlation peak. Blended lines as well as emission lines or lines that occur in the synthetic spectrum but are absent from the actual spectrum were removed from the mask. Therefore, the final mask no longer corresponds to a single $T_{\rm eff}$ but should be valid over a rather broad temperature range from about 35000 to 42500 K. The radial velocities were finally obtained from the cross-correlation function by fitting two Gaussians. The results are presented in the last two columns of Table2.

4.2 Period determination

We used the generalized Fourier spectrogramme technique of Heck et al. (1985, hereafter HMM, see their Eq. (1)) and the methods of Renson (1978) and Lafler & Kinman (1965, hereafter LK) to determine an improved value of the orbital period. To this aim, we applied the period search techniques to the time series of the RV₁ - RV₂ radial velocity differences as determined from the HeI $\lambda$ 4471 line in our data. To improve the accuracy of the result, we included also the data of Stickland & Koch (1996). The HMM method yields an orbital period of 9.81475days (in agreement with the results of the Renson method) whilst the LK method yields 9.81431days. The estimation of the error on a period is not a straightforward problem. A secure upper limit on the error can be estimated from the width of the peak (which regardless of the method is of the order $1/\Delta T$ where $\Delta T$ is the total time spanned by the data set) i.e. 0.014day in the present case. Empirically, the error is usually a substantial fraction of it. Under the hypothesis that the data are well distributed over $\Delta T$ and that the sampling does not suffer from any particular problem, we can use Eq. (20) of Lucy & Sweeney (1971) to compute a formal error. In the present case this yields $\sigma_{\rm P} = 0.00084$ day. The three methods hereabove yield values of the period that overlap within the errors and in the following we will adopt the period obtained with the HMM technique.

Table 3: Orbital solutions adopting a period of P = 9.81475days and assuming e = 0.0. T₀ corresponds to the conjunction with the secondary in front. The quoted errors correspond to $1\,\sigma$ uncertainties
HeI $\lambda$ 4471 cross-correlation

Prim. Second. Prim. Second.

T₀(HJD-2450000) $1680.188 \pm 0.531$ $1680.279 \pm 0.174$

$\gamma$ (kms^-1) $-54.0 \pm 5.2$ $-46.4 \pm 11.5$ $-46.6 \pm 1.7$ $-41.6 \pm 3.5$

K(kms^-1) $62.8 \pm 6.0$ $99.1 \pm 9.0$ $59.7 \pm 2.0$ $98.7 \pm 3.2$

$a\,\sin{i}$ ( $R_{\odot}$ ) $12.2 \pm 1.2$ $19.2 \pm 1.9$ $11.6 \pm 0.4$ $19.1 \pm 0.6$

$m\,\sin^3{i}$ ( $M_{\odot}$ ) $2.64 \pm 0.66$ $1.68 \pm 0.42$ $2.52 \pm 0.21$ $1.52 \pm 0.13$

q = m₂/m₁ $0.634 \pm 0.085$ $0.605 \pm 0.027$

$R_{\rm RL}/(a_1 + a_2)$ $0.42 \pm 0.01$ $0.34 \pm 0.01$ $0.42 \pm 0.01$ $0.34 \pm 0.01$

$R_{\rm RL}\,\sin{i}$ ( $R_{\odot}$ ) 13.2 10.7 13.0 10.3

rms (O-C) (kms^-1) 18.3 6.1

**Table 3:** Orbital solutions adopting a period of P = 9.81475days and assuming e = 0.0. T₀ corresponds to the conjunction with the secondary in front. The quoted errors correspond to $1\,\sigma$ uncertainties
	HeI $\lambda$ 4471	cross-correlation
	Prim.	Second.	Prim.	Second.
T₀(HJD-2450000)	$1680.188 \pm 0.531$	$1680.279 \pm 0.174$
$\gamma$ (kms^-1)	$-54.0 \pm 5.2$	$-46.4 \pm 11.5$	$-46.6 \pm 1.7$	$-41.6 \pm 3.5$
K(kms^-1)	$62.8 \pm 6.0$	$99.1 \pm 9.0$	$59.7 \pm 2.0$	$98.7 \pm 3.2$
$a\,\sin{i}$ ( $R_{\odot}$ )	$12.2 \pm 1.2$	$19.2 \pm 1.9$	$11.6 \pm 0.4$	$19.1 \pm 0.6$
$m\,\sin^3{i}$ ( $M_{\odot}$ )	$2.64 \pm 0.66$	$1.68 \pm 0.42$	$2.52 \pm 0.21$	$1.52 \pm 0.13$
q = m₂/m₁	$0.634 \pm 0.085$	$0.605 \pm 0.027$
$R_{\rm RL}/(a_1 + a_2)$	$0.42 \pm 0.01$	$0.34 \pm 0.01$	$0.42 \pm 0.01$	$0.34 \pm 0.01$
$R_{\rm RL}\,\sin{i}$ ( $R_{\odot}$ )	13.2	10.7	13.0	10.3
rms (O-C) (kms^-1)	18.3	6.1

4.3 Orbital elements

We used our radial velocity measurements to determine a new orbital solution. As a first step, we considered the possibility of an eccentric solution. The RVs determined by cross-correlation of our echelle spectra with a mask yield a very small eccentricity of $e = 0.02 \pm 0.02$ . We notice that the RVs of the HeI $\lambda$ 4471 line alone yield a larger value of $e = 0.20 \pm 0.05$ , but this result is most probably due to the systematic deviation of the HeI $\lambda$ 4471 RVs at certain orbital phases with respect to the RVs derived by cross-correlation. In the following, we will focus on the non-eccentric solutions. The orbital elements determined using the various datasets are listed in Table 3 and the curves shown in Figs.4 and 5.

A simple inspection of the solutions listed in Table 3 shows that most of the orbital elements overlap within their errors. The errors on the cross-correlation solution are however much smaller and the data points show little scatter around the best fit solution. This illustrates the power of the mask cross-correlation technique to handle the orbital solutions of heavily blended early-type binaries and, in the following, we will use the orbital elements determined from this solution. The orbital phases used throughout this paper refer to the ephemerides from this latter solution.

The RVs determined for the HeI $\lambda$ 4471 line are usually in rough agreement with those determined by mask cross-correlation if we account for the slight difference in the $\gamma$ velocities. A major disagreement is apparent around $\phi \sim 0.75$ for the RVs of the primary star that cluster systematically some 40kms^-1 above the best fit curve (Fig.4). It is interesting to note that the most discrepant HeI $\lambda$ 4471 RVs occur precisely in the range of orbital phases when the primary's lines display a reduced EW. The discrepant RVs of the primary near $\phi \sim 0.75$ lead to a somewhat larger mass ratio inferred from the RVs of the HeI $\lambda$ 4471 line than from the cross-correlation RVs. In this context, we note that an orbital solution derived from the RVs of the HeII $\lambda$ 5412 line yields a mass ratio of 0.619 (Nazé et al. 2000), in better agreement with the value obtained from the cross-correlation method.

Our orbital solutions yield apparent systemic velocities for the two components that are somewhat different with the primary showing the more negative $\gamma$ . This could indicate that the primary's wind is slightly stronger than that of the secondary. However, the effect is much smaller for HD149404 than for some other evolved early-type binaries (e.g. Rauw et al. 1999, 2000b).

Table 4 provides an overview of the orbital solutions of HD149404 available in the literature. We notice that the mass ratio determined from our solution is intermediary between the values proposed by Massey & Conti (1979) and by Stickland & Koch (1996). Several authors argue that the actual mass ratio might be close to unity. In fact, Massey & Conti (1979) suggested a value of $q \sim 1$ based on the RVs of the NIII $\lambda \lambda$ 4634-41 emission lines which they attributed to the primary star. Though it has been shown that the NIII emissions might actually form in the photosphere of Of-type stars (Mihalas & Hummer 1973), our observations of HD149404 reveal a rather complex behaviour of these lines that is not consistent with a purely photospheric origin (see below). A mass ratio near unity was also suggested by Hutchings & van Heteren (1981) based on a preliminary analysis of four IUE spectra and by Penny et al. (1996) from a comparison of the location of the components of HD149404 in the H-R diagram with theoretical evolutionary tracks of single stars. It is worth mentioning that all the prominent absorption lines in the spectrum of HD149404 (HeII $\lambda$ 4542, HeI $\lambda$ 4713, HeI $\lambda$ 4921, NIII $\lambda \lambda$ 4510-24, ...) yield a mass ratio in agreement with our orbital solution.

Finally, we note that including the Stickland & Koch (1996) data in our orbital solution yields orbital elements very similar to those found from our data alone.

4.4 Inclination

Massey & Conti (1979) report a photometric study of HD149404 by Dr. N. Morrison that revealed some variability at a rather low level (0.03mag). The lightcurve could probably be explained by ellipsoidal variations but there are not enough data points to draw any firm conclusion (Morrison 2000).

Luna (1988) obtained polarimetric observations of HD149404 that yield only an upper limit of $50^{\circ}$ on the orbital inclination. Vanbeveren & de Loore (1980) propose an inclination of $28^{\circ}$ based on the orbital solution of Massey & Conti (1979) and on a comparison of the primary parameters with typical masses of O7 stars. Penny et al. (1996) derived an upper limit of $31^{\circ}$ from the condition that neither of the two stars should fill its Roche lobe, and a lower limit of $15^{\circ}$ by setting an upper limit of 150 $M_{\odot}$ on the primary's mass.

Table 4: Orbital solutions from the literature. T₀ corresponds to the conjunction with the less massive star in front
Massey & Conti Stickland & Koch

Prim. Second. Prim. Second.

P (days) 9.813 9.81452

T₀ (HJD-2440000) 2503.607 4446.849

e 0 (adopted) 0 (adopted)

$\gamma$ (kms^-1) -28.0 -37.0 -60.0 -52.5

K (kms^-1) 60.0 101.0 64.3 98.1

$a\,\sin{i}$ ( $R_{\odot}$ ) 11.60 19.63 12.47 19.04

$m\,\sin^3{i}$ ( $M_{\odot}$ ) 2.7 1.6 2.64 1.73

q = m₂/m₁ 0.595 0.655

Spectral types O7III(f) O8.5I

**Table 4:** Orbital solutions from the literature. T₀ corresponds to the conjunction with the less massive star in front
	Massey & Conti	Stickland & Koch
	Prim.	Second.	Prim.	Second.
P (days)	9.813	9.81452
T₀ (HJD-2440000)	2503.607	4446.849
e	0 (adopted)	0 (adopted)
$\gamma$ (kms^-1)	-28.0	-37.0	-60.0	-52.5
K (kms^-1)	60.0	101.0	64.3	98.1
$a\,\sin{i}$ ( $R_{\odot}$ )	11.60	19.63	12.47	19.04
$m\,\sin^3{i}$ ( $M_{\odot}$ )	2.7	1.6	2.64	1.73
q = m₂/m₁	0.595	0.655
Spectral types	O7III(f)	O8.5I

Comparing our minimum masses from Table 3 with "typical" masses of O9.7I and O7.5I supergiants (e.g. Howarth & Prinja 1989), we find an inclination of $20.3^{\circ}$ for the primary and $21.0^{\circ}$ for the secondary. Adopting $21^{\circ}$ as a "realistic" estimate, the Roche lobe radii are equal to 36 $R_{\odot}$ and 29 $R_{\odot}$ for the primary and secondary respectively. According to Howarth & Prinja (1989), the typical radii of O9.7I and O7.5I supergiants are 21 and 26 $R_{\odot}$ respectively. With that inclination, it seems a priori unlikely that any of the two components fills its Roche lobe. However, we will come back to this point in Sect.7.1.

4 Orbital solution

4.1 Radial velocity determination

4.2 Period determination

4.3 Orbital elements

4.4 Inclination