Given that the amplitude of the polarization is so small the question of foregrounds is even more important than for the temperature anisotropy. Unfortunately, the level and structure of the various foreground polarization in the CMB frequency bands is currently not well known. We review some of the observations in the adjacent radio and IR bands (a more complete discussion can be found in [Keating et al.] 1997). Atmospheric emission is believed to be negligibly polarized ([Keating et al.] 1997), leaving the main astrophysical foregrounds: free-free, synchrotron, dust, and point source emissions. Of these the most important foreground is synchrotron emission.
Free-free emission (bremsstrahlung) is intrinsically unpolarized ([Rybicki & Lightman] 1979) but can be partially polarized by Thomson scattering within the HII region. This small effect is not expected to polarize the emission by more than 10% ([Keating et al.] 1997). The emission is larger at low frequencies but is not expected to dominate the polarization at any frequency.
The polarization of dust is not well known.
In principle, emission from dust particles could be highly polarized,
however [Hildebrand & Dragovan] (1995) find that in their observations
the majority of dust is polarized
at the 
level at 
m with a small fraction of regions
approaching 
polarization.
Moreover [Keating et al.] (1997) show that even at 1000olarization,
extrapolation of the IRAS 100 
m map with the COBE FIRAS index
shows that dust emission is negligible below 80GHz.
At higher frequencies it will become the dominant foreground.
Radio point sources are polarized due to synchrotron emission at
level. For large angle experiments, the random contribution
from point sources will contribute negligibly, but may be of more
concern for the upcoming satellite missions.
Galactic synchrotron emission is the major concern. It is potentially
highly polarized with the fraction dependent on the spectral index and
depolarization from Faraday rotation and non-uniform magnetic fields.
The level of polarization is expected to lie between 10% of a total
intensity which itself is approximately 
K at 30GHz.
This estimate follows from extrapolating the [Brouw & Spoelstra] (1976)
measurements at 1411 MHz with an index of 
.
Due to their different spectral indices, the minimum in the foreground polarization, like the temperature, lies near 100GHz. For full sky measurements, since synchrotron emission is more highly polarized than dust, the optimum frequency at which to measure intrinsic (CMB) polarization is slightly higher than for the anisotropy. Over small regions of the sky where one or the other of the foregrounds is known a priori to be absent the optimum frequency would clearly be different. However as with anisotropy measurements, with multifrequency coverage, polarized foregrounds can be removed.
It is also interesting to consider whether the spatial as well as frequency signature of the polarization can be used to separate foregrounds. Using angular power spectra for the spatial properties of the foregrounds is a simple generalization of methods already used in anisotropy work. For instance, in the case of synchotron emission, if the spatial correlation in the polarization follows that of the temperature itself, the relative contamination will decrease on smaller angular scales due to its diffuse nature. Furthermore the peak of the cosmic signal in polarization occurs at even smaller angular scales than for the anisotropy.
One could attempt to exploit the additional properties of polarization, such as its E- and B-mode nature. We mentioned above that in a wide class of models where scalars dominate on small angular scales, the polarization is predicted to be dominantly E-mode ([Kamionkowski et al.] 1997 [Zaldarriaga & Seljak] 1997). [Seljak] (1997) suggests that one could use this to help eliminate foreground contamination by ``vetoing'' on areas of B-mode signal. However in general one does not expect that the foregrounds will have equal E- and B-mode contribution, so while this extra information is valuable, its use as a foreground monitor can be compromised in certain circumstances. Specifically, if a correlation exists between the direction of polarization and the rate of change (curvature) of its amplitude, the foreground will populate the two modes unequally. Two simple examples: either radial or tangential polarization around a source with the amplitude of the polarization dropping off with radius, or polarization parallel or perpendicular to a ``jet'' whose amplitude drops along the jet axis. Both examples would give predominantly E-mode polarization.