Unbiased L<sub>1</sub>and L<sub>∞</sub>estimation (Q3707159)
From MaRDI portal
!
This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:
scientific article; zbMATH DE number 3936266
| Language | Label | Description | Also known as |
|---|---|---|---|
| default for all languages | No label defined |
||
| English | Unbiased L<sub>1</sub>and L<sub>∞</sub>estimation |
scientific article; zbMATH DE number 3936266 |
Statements
Unbiased L<sub>1</sub>and L<sub>∞</sub>estimation (English)
0 references
1985
0 references
regression quantiles
0 references
L sub p-estimators
0 references
unbiasedness
0 references
symmetric distribution
0 references
non-uniqueness problem
0 references
0 references
0.8237954378128052
0 references
0.8108227252960205
0 references
0.8054191470146179
0 references
0.7999102473258972
0 references
In this paper the linear regression model \(y_ i=\sum^{m}_{j=1}\beta_ jx_{ji}+\epsilon_ i\), \(i=1,...,n\), is considered. For a fixed p, \(1\leq p\leq \infty\) and for a set of fixed positive weights \(w_ i\), \(i=1,...,n\), the \(L_ p\) estimator b of \(\beta\) minimizes \[ \sum^{n}_{i=1}w^ p_ i| y_ i- \sum^{m}_{j=1}\beta_ jx_{ji}|^ p\quad if\quad 1\leq p<\infty,\quad and\quad \max_{i}w_ i| y_ i- \sum^{m}_{j=1}\beta_ jx_{ji}| \quad if\quad p=\infty. \] The paper restates that \(1<p<\infty\) defines a strictly convex optimization problem and hence has a unique solution. If the error distribution is symmetric about zero then b is symmetrically distributed about \(\beta\) and if E(b) exists, b is an unbiased estimator of \(\beta\). However, in the cases \(p=1\) and \(p=\infty\) the solution is not necessarily unique. Four distinct linear programming problems for the computation of \(L_ 1\) and \(L_{\infty}\) estimates are given. It is shown that in a case of non-uniqueness the solution from one algorithm is not symmetrically distributed about \(\beta\), but if the algorithm is selected by random the estimator is symmetrically distributed about \(\beta\) (always assuming symmetric error distribution). A drawback of the paper is that the author is more concerned with stating conditions for when the estimators are symmetrically distributed about the true value than with unbiasedness (as one might expect after reading the title of the paper). For unbiasedness a necessary condition is that the expectation of the estimators exists. Therefore it would have been interesting to see under which conditions this expectation exists. Furthermore, the condition of symmetric distribution of the estimator, which the author is concerned about, is not necessary for unbiasedness! An estimator that is not symmetrically distributed about the true value can still be unbiased. I also feel uncomfortable with solving the non-uniqueness problem by randomly selecting algorithms. A far more appealing alternative is to use a linear combination of the solutions of several algorithms. If the author had tried to describe the set of solutions he would find it to be a closed convex set. The solutions of the algorithms defined in the paper are always on the edge of the solution set. By randomly selecting algorithms we get ''in the mean'' an interior point of the solution set, thereby the symmetry property. Nevertheless, the solution from the selected algorithm is still at the edge. A linear combination on the other hand yields always an inner point of the solution set. Finally, a description of the solution set will give an idea of how important (or disastrous) a non-uniqueness problem is in a particular application. Thus, if all solutions are ''approximately equal'' the non- uniqueness problem is unimportant and need not to be considered at all. This will probably happen in large samples, but is still to be more investigated.
0 references