reg.dim.est: Estimate the optimal dimension in linear regression problem

In freqdom: Frequency Domain Analysis for Multivariate Time Series

Description

Consider a linear regression problem for a multivariate stationary time series X_t:

Y_t = P X_t + \varepsilon_t.

Estimator based on formula

EY_0X_0 (EX_0^2)^{-1} = P

is fragile on the eigendirections of EX_0^2 with small eigenvalues. It is therefore desired to truncate the inversion at a level where eigenvalues are estimated consistently. Procedure dim.est suggest such level by taking only the eigenvalues which are greater and equal than 1/√{n}K_{const}. It is designed for √{n} consistent matrix estimator and can serve as one of heuristics for matrix inverion problems.

Usage

reg.dim.est(eigenvalues, n, Kconst = 1)

Arguments

eigenvalues	vector of eigenvalues
n	used for estimation
Kconst	parameter for fitting the convergence rate to 1/(Kconst*n^1/2)

Value

number of 'safe' eigendirections

References

Siegfried Hormann and Lukasz Kidzinski A note on estimation in Hilbertian linear models Research report, 2012

See Also

reg.est, pseudoinverse

Examples

n = 100
X = rar(n)
Y = rar(n)
CV = lagged.cov(X,Y)
E = eigen(CV)
K = reg.dim.est(E$values, n)

freqdom documentation built on May 2, 2019, 5:55 p.m.