A new method for source localization is described that is based on a modifi
cation of the well-known MUSIC algorithm. In classical MUSIC, the array man
ifold vector is projected onto an estimate of the signal subspace. Errors i
n the estimate of the signal subspace can make localization of multiple sou
rces difficult, Recursively applied and projected (RAP) MUSIC uses each suc
cessively located source to form an intermediate array gain matrix and proj
ects both the array manifold and the signal subspace estimate into its orth
ogonal complement. The MUSIC projection to find the next source is then per
formed in this reduced subspace, Special assumptions about the array manifo
ld structure, such as Vandermonde or shift invariance, are not required. Us
ing the metric of principal angles, we describe a general form of the RAP-M
USIC algorithm for the case of diversely polarized sources. Through a unifo
rm linear array simulation with two highly correlated sources, we demonstra
te the improved Monte Carlo error performance of RAP-MUSIC relative to MUSI
C and two other sequential subspace methods: S and IES-MUSIC, We then demon
strate the more general utility of this algorithm for multidimensional arra
y manifolds in a magnetoencephalography (MEG) source localization simulatio
n.