Jack-knifing a Multitaper SDF estimator¶
Assume there is a parameter \(\theta\) that parameterizes a distribution, and that the set of random variables \(\lbrace Y_1, Y_2, ..., Y_n \rbrace\) are i.i.d. according to that distribution.
The basic jackknifed estimator \(\tilde{\theta}\) of some parameter \(\theta\) is found through forming pseudovalues \(\hat{\theta}_i\) based on the original set of samples. With n samples, there are n pseudovalues based on n “leave-one-out” sample sets.
General JN definitions¶
Now the jackknifed esimator is computed as
\(\tilde{\theta} = \dfrac{1}{n}\sum_i \hat{\theta}_i = n\hat{\theta} - \dfrac{n-1}{n}\sum_i \hat{\theta}_{-i}\)
This estimator is known (?) to be distributed about the true parameter theta approximately as a Student’s t distribution, with standard error defined as
\(s^{2} = \dfrac{n-1}{n}\sum_i \left(\hat{\theta}_i - \tilde{\theta}\right)^{2}\)
General Multitaper definition¶
The general multitaper spectral density function (sdf) estimator, using n orthonormal tapers, combines the n \(\lbrace \hat{S}_i^{mt} \rbrace\) sdf estimators, and takes the form
\(\hat{S}^{mt}(f) = \dfrac{\sum_{k} w_k(f)^2S^{mt}_k(f)}{\sum_{k} |w_k(f)|^2} = \dfrac{\sum_{k} w_k(f)^2S^{mt}_k(f)}{\lVert \vec{w}(f) \rVert^2}\)
For instance, using discrete prolate spheroidal sequences (DPSS) windows, the \(\rbrace w_i \lbrace\) set, in their simplest form, are the eigenvalues of the spectral concentration system.
A natural choice for a leave-one-out measurement is (leaving out the dependence on argument f)
\(\ln\hat{S}_{-i}^{mt} = \ln\dfrac{\sum_{k \neq i} w_k^2S^{mt}_k}{\lVert \vec{w}_{-i} \rVert^2} = \ln\sum_{k \neq i} w_k^2S^{mt}_k - \ln\lVert \vec{w}_{-i} \rVert^2\)
where \(\vec{w}_{-i}\) is the vector of weights with the ith element set to zero. The natural log has been taken so that the estimate is distributed below and above \(S(f)\) more evenly.
Multitaper Pseudovalues¶
I’m not quite clear on the form of the pseudovalues for multitaper combinations.
One Option¶
The simple option is to weight the different leave-one-out measurements equally, which leads to
\(\ln\hat{S}_{i}^{mt} = n\ln\hat{S}^{mt} - (n-1)\ln\hat{S}_{-i}^{mt}\)
And of course the estimate of \(S(f)\) is given by
\(\ln\tilde{S}^{mt} (f) = \dfrac{1}{n}\sum_i \ln\hat{S}_i^{mt}(f)\)
Another Option¶
Another approach seems obvious which weights the leave-one-out measurements according to the length of \(\vec{w}_{-i}\). It would look something like this
Then the pseudovalues are
\(\ln\hat{S}_i^{mt} = \left(\ln\hat{S}^{mt} + \ln g\right) - \left(\ln\hat{S}_{-i}^{mt} + \ln g_i\right)\)
and the jackknifed estimator is
\(\ln\tilde{S}^{mt} = \sum_i \ln\hat{S}_i^{mt} - \ln g\)
and I would wager, the standard error is estimated as
\(s^2 = \dfrac{1}{n}\sum_i \left(\ln\hat{S}_i^{mt} - \ln\tilde{S}^{mt}\right)^2\)