Minimum-variance multitaper spectral estimation on the sphere

• Mapping large spherical regions to a plane can make spectral estimates increasingly unreliable as the region approaches a significant fraction of the sphere.Source: §1 · p. 2
• Localized spherical spectra serve either global-spectrum estimation from stationary data or local-spectrum estimation under local stationarity.Geophysical and geodetic analyses may focus on regions because of local geologic history, while cosmological measurements mask galactic emissions.Source: §1 · p. 2
• The paper extends zonal spherical tapers to non-zonal windows, increasing the number of component estimates and significantly reducing multitaper variance.It describes choosing weights to minimize that variance.Source: §1 · p. 2
• The paper develops global-spectrum estimation theory, window design, minimum-variance multitaper estimation, and heuristic bias and variance analyses for white and red spectra.It also presents a single-realization example and identifies future research directions.Source: §1 · pp. 2–3

2.3 Minimum-variance multitaper spectral estimation

• 60°?Source: §2.3 · p. 10

3. White and Red Stochastic Processes

• Localized spectra depend on the global spectrum within l±L, so bias can be appreciable near or below the window bandwidth, especially for red spectra.At degrees near zero the localized spectrum is biased downward, while above about L/2 the bias is positive; red spectra show stronger bias because high low-degree power leaks upward.Source: §3 · pp. 11–12
• For white spectra, individual-window estimates are dominated by diagonal covariance terms, and their approximate statistical independence improves with increasing degree.The covariance comparisons cover l = 30 and 65.Source: §3 · p. 12, Figure 4 · p. 13
• Adding 30 non-zonal windows to four zonal windows reduces white-spectrum uncertainty by an additional 50%, although the estimates are not fully independent.For white spectra, optimal and equal weights produce nearly indistinguishable uncertainty.Source: §3 · pp. 12, 14
• For white processes, uncertainty decreases as more windows are used and with increasing degree, though it can exceed 30% when only a few windows are used.The displayed white-process uncertainty covers degrees l = 30 to 100.Source: §3 · p. 14, Figure 5 · p. 14
• For red spectra, l = 65 has covariance and uncertainty similar to the white case, whereas l = 30 differs substantially as higher-number tapers have larger uncertainty.At low degree, off-diagonal covariance terms also grow in importance as taper indices increase.Source: §3 · pp. 15–16, Figure 6 · p. 15
• For red spectra at l = 30, equal weighting increases uncertainty after the first five windows, while optimal weights effectively use only those five.Adding further windows yields no appreciable variance decrease with optimal weighting.Source: Figure 6 · p. 15, §3 · p. 16

4. Single Realizations of Stochastic Processes

• Because individual windowed estimates are somewhat uncorrelated, they can be treated approximately as separate realizations, so adding windows is expected to reduce estimate variance.This addresses settings with only one available realization, such as Earth’s gravitational field or the cosmic microwave background.Source: §4 · p. 16
• Iteratively estimating the global spectrum, covariance, and optimal weights could approach a minimum-variance estimate, but repeated covariance calculations can be computationally infeasible.The proposed practical approach assumes individual spectral estimates at each degree are independent and Gaussian with identical variance.Source: §4 · p. 17
• The practical uncertainty estimate assumes independent, identically distributed Gaussian tapered estimates and is computed from the data alone, unlike the global-spectrum-dependent analytic variance.With equal weights, the estimated variance is σ²/(K−1); omitting variability among windowed expectations makes it an overestimate, while dependence can make it an underestimate.Source: §4 · p. 17
• With more tapers, the difference between each estimate and its known expectation almost everywhere decreases, while estimated uncertainty is generally comparable but slightly underestimates the true value.The authors attribute the underestimation most likely to correlations between individual spectral estimates.Source: §4 · p. 18, Figure 7 · p. 19
• Recovering the global spectrum from localized multitaper estimates is underdetermined because the global-spectrum vector is larger than the estimate vector, and the mapping matrix is never full-rank.Several linear inversions were unsatisfactory; a solution with positive values and reasonable variances required spectrum bins wider than about 15 degrees.Source: Inverting M(mt) to obtain Sff from S(mt) · pp. 18–19

5. Concluding Remarks

• A weighted multitaper estimate built from orthogonal windows with good spatiospectral localization reduces variance versus single-window estimates and permits weights chosen to minimize estimation variance.The windows balance spatial concentration in the target region against small effective spectral bandwidth.Source: §5 · pp. 19–20
• The study restricts localization windows to spherical-cap concentration, although the techniques generalize to other concentration domains and the authors expect the general character of results to remain similar.Spherical-cap windows are described as sufficient for applications such as planetary gravity and topography.Source: §5 · p. 20
• Using different windows for the two fields could provide up to K(K + 1)/2 cross-power estimates instead of K, but their covariance and variance benefit remain uninvestigated.The study used a single localization window for localized cross-power spectra.Source: §5 · p. 20
• Covariance-matrix calculations are computationally demanding and become increasingly infeasible as spherical harmonic degree increases, motivating alternative methods for covariance, optimal weights, and uncertainties.Reported calculations ranged from about 12 hours and 2 days for Figure 4 panels to about one month for Figure 5.Source: §5 · p. 20
• Linear inversion of the localized-spectrum relation yielded mixed results for recovering an unbiased global spectrum; nonlinear or Monte Carlo methods with bounds and positivity constraints remain potential alternatives.The paper’s discussion primarily quantifies multitaper bias and uncertainty.Source: §5 · p. 20

A. Real and Complex Spherical Harmonics

• Using complex spherical harmonics and their orthogonality simplifies the derivations and relates the complex coefficients to those of the real function.The complex harmonics omit the Condon–Shortley phase factor; including or excluding it does not affect the paper’s results.Source: A. Real and Complex Spherical Harmonics · pp. 20–21
• The appendix specifies the complex-harmonic identity and relates complex coefficients to real coefficients using the paper’s convention that excludes the Condon–Shortley phase.The stated phase choice does not affect the results presented.Source: A. Real and Complex Spherical Harmonics · p. 21
• A generalized Parseval relation connects a real function’s total power to its complex spectral coefficients, with the power spectrum defined in the same framework.The appendix also gives corresponding relations for cross-power and the cross-power spectrum of two real functions.Source: A. Real and Complex Spherical Harmonics · pp. 21–22
• For zero-mean real functions, each degree’s auto-spectrum and cross-spectrum represent that degree’s contributions to variance and covariance, respectively..Source: A. Real and Complex Spherical Harmonics · p. 22

B. Bias of a Localized Spectral Estimate

• The derivation assumes zero-mean real spherical-harmonic coefficients and isotropic power spectra for the fields.For two fields, their cross-power at each degree is also assumed isotropic.Source: B. Bias of a Localized Spectral Estimate · p. 22
• For two fields localized by different tapers of equal spectral bandwidth, the appendix derives the expected localized cross-power spectrum.When both fields use the same window, the expectation is expressed after a change of variables.Source: B. Bias of a Localized Spectral Estimate · pp. 22, 24
• With K tapers, up to K(K + 1)/2 cross-spectral estimates can be obtained, even when the tapers are orthogonal.The sum over one harmonic order may be zero, but generally is not.Source: B. Bias of a Localized Spectral Estimate · p. 24

C. Variance of a Localized Spectral Estimate

• A localized multitaper cross-power estimate is a weighted average of direct estimates from K orthogonal tapers, with weights constrained to sum to one.The variance is derived for this weighted estimate.Source: C. Variance of a Localized Spectral… · pp. 24–25
• The variance derivation proceeds for identical tapers; the benefit of using two different tapers in a spectral estimate remains for further research.The appendix states that the relevant quantity is real when the weight matrix is symmetric.Source: C. Variance of a Localized Spectral… · p. 25
• The variance derivation uses Gaussian localized coefficients, justified either by a central-limit assumption or by Gaussian unwindowed coefficients.It applies Isserlis’ theorem to zero-mean Gaussian complex variables.Source: C. Variance of a Localized Spectral… · pp. 25–26
• For spherical-cap concentration windows, each window has nonzero real spherical-harmonic coefficients at only one angular order, simplifying computation of the covariance matrix.The nonzero complex coefficients occur at the positive and negative orders corresponding to each window.Source: C. Variance of a Localized Spectral… · p. 27

D. Correlation of Multitaper Spectral Estimates

• Estimates separated by less than 2L degrees are expected to be partially correlated, where L is the localization-window bandwidth.The appendix quantifies this correlation through the covariance of the two multitaper estimates.Source: D. Correlation of Multitaper Spectral… · p. 27
• The covariance between multitaper estimates is derived using the preceding covariance methodology.The expected estimate at degree l depends on the global power spectrum over l±L.Source: D. Correlation of Multitaper Spectral… · p. 27
• For cross-power spectra of a single function, the matrix F can be simplified.Source: D. Correlation of Multitaper Spectral… · p. 27