To generalize the method described in Paper I, let us suppose
the observed spectrum I(x,t) (the variable
is proportional to logarithm of
wavelength expressed in an arbitrary unit 
; cf.
Simkin 1974) to be composed of spectra 
of
n stars which have no intrinsic variability apart from an
overall change of their intensities proportional to functions
which may be caused e.g., by ellipticity or eclipses
of binary components or by air-mass or humidity for the telluric
spectrum.
If the instantaneous radial velocity (of the star j at
the time t) 
,
the composite spectrum is given by the convolution in x
The Fourier transform (
) of this equation reads
If we have k spectra (k>n) observed at times 
corresponding to various values of 
,
we can - in principle - fit them by searching for appropriate
values of 
, 
and
. The velocities
can be treated either as independent values, or
to be given functions of time and certain parameters p, e.g.
the orbital elements of the spectroscopic binary.
Using the standard method of least squares we arrive at the
condition
where
Here 
is the weight of each Fourier mode y in the
exposure l. In practice, we suppose
The weight 
of each exposure can be chosen e.g. in
dependence on the number of photon counts. The function w(y)
can be introduced as a frequency filter, e.g. to cut off the low
frequency modes influenced by the rectification of the spectra.
Figure 1: Continuum and line intensities of uneclipsed
components (left) and in the primary eclipse (right)
Because S is bilinear in 
, the conditions
obtained for them from Eq. (3 (click here)) varying with respect to
(i.e. the variation of complex
conjugate of 
) are linear equations
(m=1,...,n), which are moreover independent for each y.
Just this independence
of Fourier components, which is the consequence of the form of
Eq. (2 (click here)) local in the variable y (unlike the form of
Eq. (1 (click here)) integral in the variable x), makes the
disentangling of the observed spectra
easier in Fourier transform than in the wavelength space.
Solving this system of equations for each y and substituting
into Eq. (4 (click here)), S can be optimized
only with respect to other parameters, which are much less
numerous.
It is obvious that Eq. (6 (click here)) is singular for y=0.
This corresponds to the fact that the contributions of individual
stars to the constant term of I(x) cannot be directly
distinguished. The continua of stars are almost constant and in
practice unaffected by the Doppler shift. Consequently, a large
error of 
can be expected for small values of
y. It can thus be convenient to cut-off this range of y
choosing here w(y)=0. An indirect method of distinguishing the
contributions to the continuum is described in Sect. 3 (click here).
S is bilinear also in the coefficients 
. Hence, varying
with respect to 
, we get the linear set of equations
for these coefficients. Because these coefficients are generally
still quite numerous (but less than the Fourier modes of
the component spectra), it is advantageous to solve for them
directly from these equations before optimizing S with respect
to either 
or p, in which it is non-linear.
It is important to keep in mind that the solutions of orbital
elements (or individual independent radial velocities),
the decomposition of the spectrum and the solution of component
intensities are inter-related and their self-consistent solution
should be found. This can be achieved iteratively starting from
some initial estimate of orbital parameters and line intensities.
An inner loop solving successively Eq. (6 (click here)) for
and then Eq. (7 (click here)) for 
is used
in KOREL, while the orbital parameters are solved in an outer
loop by simplex method. Practical experience shows, that slower,
but more stable convergence of some of coefficients 
can
be advantageous to get a better initial estimate. A successive
increasing of the number k of exposures can also help the
convergence.
