Statistical estimation of a subspace in a complex space
Let \(x_j\), \(j=1,\ldots,n\), be a \(p\)-dimensional vector with complex coordinates. These vectors are assumed to satisfy an analogue of the classical linear statistical relationship \(x_j =\mu_j + \varepsilon_j\), \(j=1,\ldots,n\), where \(\mu_j\) represents the mean of \(x_j\) and the unobservable \(\varepsilon_j\) is the random error (\(\varepsilon_j\), \(j=1,\ldots,n\), are \(p\)-dimensional complex normal vectors with mean 0 and covariance matrix \(\sigma^2 I\)). The main assumption is that \(\mu_j\), \(j=1,\ldots,n\), belong to a lower-dimensional linear space \(V\) whose dimension \(m\), \(m<p\), is given. The vectors \(x_j\), \(j=1,\ldots,n\), are orthogonal projections onto \(V\) of some vectors \(\phi_j\) taking values in the original \(p\)-dimensional space \(C^p\), \(\mu_j =P_V \phi_j.\) In this paper the statistical estimation problem of the subspace \(V\) is considered. The author proved that under some loss function the maximum likelihood estimator \[ \hat V =\arg \min \sum_j \text{Tr}\bigl( (I-P_V)x_j x_j^\ast (I-P_V)^\ast \bigr) \] of \(V\) is the (unique) best equivariant estimator. The application of these results to the directions--of--arrival estimators in multivariate sensor array processing is also considered.
- Asymptotic properties of subspace estimators
- scientific article; zbMATH DE number 1047765
- Independent Component Analysis and Blind Signal Separation
- Estimation of the mean vector of a multivariate normal distribution: subspace hypothesis
- scientific article; zbMATH DE number 1058103
- Statistical analysis of novel subspace identification methods
- Invariance of subspace based estimators
- Subgaussian estimates in probability and harmonic analysis
- The asymptotic variance of subspace estimates.
This page was built for publication: Statistical estimation of a subspace in a complex space
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4261527)
