Shrinkage corrections of sample linear estimators in the small sample size regime
We are living in a data deluge era where the dimensionality of the data gathered by inexpensive sensors is growing at a fast pace, whereas the availability of independent samples of the observed data is limited. Thus, classical statistical inference methods relying on the assumption that the sample...
| Author: | |
|---|---|
| Format: | doctoral thesis |
| Publication Date: | 2016 |
| Country: | España |
| Institution: | Universitat Politècnica de Catalunya (UPC) |
| Repository: | UPCommons. Portal del coneixement obert de la UPC |
| Language: | English |
| OAI Identifier: | oai:upcommons.upc.edu:2117/106290 |
| Online Access: | https://hdl.handle.net/2117/106290 https://dx.doi.org/10.5821/dissertation-2117-106290 |
| Access Level: | Open access |
| Keyword: | Fotònica Tractament del senyal Àrees temàtiques de la UPC::Enginyeria de la telecomunicació |
| Summary: | We are living in a data deluge era where the dimensionality of the data gathered by inexpensive sensors is growing at a fast pace, whereas the availability of independent samples of the observed data is limited. Thus, classical statistical inference methods relying on the assumption that the sample size is large, compared to the observation dimension, are suffering a severe performance degradation. Within this context, this thesis focus on a popular problem in signal processing, the estimation of a parameter, observed through a linear model. This inference is commonly based on a linear filtering of the data. For instance, beamforming in array signal processing, where a spatial filter steers the beampattern of the antenna array towards a direction to obtain the signal of interest (SOI). In signal processing the design of the optimal filters relies on the optimization of performance measures such as the Mean Square Error (MSE) and the Signal to Interference plus Noise Ratio (SINR). When the first two moments of the SOI are known, the optimization of the MSE leads to the Linear Minimum Mean Square Error (LMMSE). When such statistical information is not available one may force a no distortion constraint towards the SOI in the optimization of the MSE, which is equivalent to maximize the SINR. This leads to the Minimum Variance Distortionless Response (MVDR) method. The LMMSE and MVDR are optimal, though unrealizable in general, since they depend on the inverse of the data correlation, which is not known. The common approach to circumvent this problem is to substitute it for the inverse of the sample correlation matrix (SCM), leading to the sample LMMSE and sample MVDR. This approach is optimal when the number of available statistical samples tends to infinity for a fixed observation dimension. This large sample size scenario hardly holds in practice and the sample methods undergo large performance degradations in the small sample size regime, which may be due to short stationarity constraints or to a system with a high observation dimension. The aim of this thesis is to propose corrections of sample estimators, such as the sample LMMSE and MVDR, to circumvent their performance degradation in the small sample size regime. To this end, two powerful tools are used, shrinkage estimation and random matrix theory (RMT). Shrinkage estimation introduces a structure on the filters that forces some corrections in small sample size situations. They improve sample based estimators by optimizing a bias variance tradeoff. As direct optimization of these shrinkage methods leads to unrealizable estimators, then a consistent estimate of these optimal shrinkage estimators is obtained, within the general asymptotics where both the observation dimension and the sample size tend to infinity, but at a fixed rate. That is, RMT is used to obtain consistent estimates within an asymptotic regime that deals naturally with the small sample size. This RMT approach does not require any assumptions about the distribution of the observations. The proposed filters deal directly with the estimation of the SOI, which leads to performance gains compared to related work methods based on optimizing a metric related to the data covariance estimate or proposing rather ad-hoc regularizations of the SCM. Compared to related work methods which also treat directly the estimation of the SOI and which are based on a shrinkage of the SCM, the proposed filter structure is more general. It contemplates corrections of the inverse of the SCM and considers the related work methods as particular cases. This leads to performance gains which are notable when there is a mismatch in the signature vector of the SOI. This mismatch and the small sample size are the main sources of degradation of the sample LMMSE and MVDR. Thus, in the last part of this thesis, unlike the previous proposed filters and the related work, we propose a filter which treats directly both sources of degradation. |
|---|