Khái niệm cốt lõi
Perturbative analysis explains the superconvergence behavior of the ESPRIT algorithm in estimating spike locations and weights under large noise and large frequency regimes.
Tóm tắt
The content presents a perturbative analysis to explain the superconvergence behavior observed in the ESPRIT algorithm for noisy spectral estimation.
The key highlights are:
The analysis assumes a minimum gap between the spike locations and that the weights are all Θ(1).
Using a perturbative approach, it is shown that the absolute errors of the spike locations {xk} scale like O(n^(-3/2)), while the relative errors of the weights {wk} scale like O(n^(-1/2)), where n is the number of Fourier measurements.
The analysis is extended to the case where the noise magnitude grows with the frequency, modeled as zj ~ j^p * σ * NC(0, 1). As long as p < 1/2, the errors of {wk} decay as n grows, and as long as p < 3/2, the errors of {xk} go to zero with n.
Numerical examples are provided for different noise scaling scenarios, and the observed error asymptotics match well with the predictions from the perturbative analysis.
The content provides an intuitive understanding of the superconvergence behavior of the ESPRIT algorithm in the large noise, large frequency regime, which is highly relevant for quantum phase estimation applications.
Thống kê
Each diagonal entry of the A^T A matrix is 2/3 * n^3 * (1 + O(1/n)).
Each off-diagonal entry of the A^T A matrix is O(n^2/Δ), where Δ is the minimum gap between the spike locations.
Each diagonal entry of the B^T B matrix is 2n * (1 + O(1/n)).
Each off-diagonal entry of the B^T B matrix is O(1/Δ).
Trích dẫn
"The superconvergence result O(n^(-3/2)) for {xk} is quite surprising and the proof in [3] is a tour de force."
"As long as the p < 1/2, the errors of {wk} decay as n grows. As long as p < 3/2, the errors of {xk} goes to zero with n."