Source-linked AI summary
Precision-Aware Hierarchical Spectrum Search for Finite-Precision MUSIC DOA Estimation
Linyun Yue, Chuang Yang, Jierui Chen, Mugen Peng
TL;DR
Dense-grid MUSIC spectrum search is computationally costly, and low-precision evaluation can distort candidate selection. PAH-MUSIC coordinates stage-dependent precision, angular grids, and candidate windows; in the tested two-source scenario, selected low-precision coarse searches with FP64 refinement had RMSE close to the FP64 two-stage baseline while reducing weighted spectrum-evaluation cost.
Problem
MUSIC’s spectrum search over a dense angular grid is computationally costly, while low-precision evaluation can introduce errors that distort candidate selection.
Source: §I · p. 1Method
PAH-MUSIC coordinates arithmetic precision, angular-grid spacing, and candidate-window selection across stages, using low precision for coarse discovery and higher precision for finer-grid refinement.
Source: §I · p. 2Results
28.36%–44.66% weighted spectrum-evaluation cost reductions versus the FP64 two-stage baseline, with RMSE close to that baseline in the tested separated two-source scenario.
Source: §V · p. 5Takeaways & Limitations
The tested FP8, FP16, and FP32 coarse-search configurations with FP64 refinement provide RMSE close to FP64 two-stage search at lower weighted spectrum-evaluation cost.
Source: §V · p. 5
Abstract
from arXiv · showhide
Dense-grid spectrum search is computationally costly for multiple signal classification (MUSIC) direction-of-arrival (DOA) estimation, and low-precision evaluation can introduce numerical errors that distort candidate selection. This paper proposes precision-aware hierarchical MUSIC (PAH-MUSIC), a multistage framework that coordinates arithmetic precision, angular grids, and candidate windows in spectrum evaluation, while covariance estimation and eigendecomposition remain in FP64. Stage-dependent precision is assigned from coarse candidate discovery to final DOA estimation. Under bounded perturbations and local spectral growth, a conditional displacement bound gives a sufficient window-width condition for retained, correctly associated candidates. For a two-stage continuous cost model with fixed precision weights and window coefficients, a closed-form optimal coarse step and a cost-gain condition are derived. Simulations with two separated sources and 500 frames at each tested signal-to-noise ratio (SNR) from -10 to 15 dB show that FP8, FP16, and FP32 coarse searches followed by FP64 refinement achieve DOA root-mean-square errors close to a baseline using FP64 at both search stages. Relative to that baseline, the selected configurations reduce weighted spectrum-evaluation cost by 28.36\%--44.66\% under the adopted precision weights.
I. INTRODUCTION
MUSIC offers high-resolution DOA estimation but requires covariance eigendecomposition and spectrum search across the angular domain. PAH-MUSIC coordinates precision, grid spacing, and candidate windows across search stages to reduce search cost while retaining peaks for refinement.
I. INTRODUCTION
- MUSIC requires covariance-matrix eigendecomposition and spectrum-peak search over the full angular domain, creating two computational burdens.Subspace methods such as MUSIC use subspace orthogonality for high-resolution DOA estimation.Source: §I · p. 1
- PAH-MUSIC coordinates arithmetic precision, angular-grid spacing, and candidate-window selection across search stages.The proposed framework uses low precision for coarse candidate discovery and higher precision on finer grids for accurate DOA estimation.Source: §I · pp. 1–2
- A conditional displacement bound guides window selection for retained, correctly associated candidates under numerical perturbations.The paper also derives an optimal coarse-grid spacing and a cost-reduction condition for a two-stage cost model.Source: §I · p. 2
A. Signal Model and MUSIC Spectrum Search
- The framework’s stated evaluation keeps covariance estimation and eigendecomposition in FP64 while applying stage-dependent precision to spectrum evaluation.The reference MUSIC denominator uses the noise-subspace eigenvectors from the FP64 sample covariance.Source: A. Signal Model and MUSIC Spectrum… · p. 2
B. Finite-Precision Spectrum Denominator Model
The paper models finite-precision perturbations in spectrum evaluation separately from sampling error shared through the FP64 sample subspace. Its search design uses progressively finer grids and nondecreasing precision, aiming to reduce weighted cost while keeping RMSE close to an FP64 two-stage baseline.
B. Finite-Precision Spectrum Denominator Model
- Finite-precision spectrum evaluation adds numerical perturbation to the sampling error shared across strategies by the common FP64 sample subspace.The model distinguishes operand quantization with exact subsequent operations from additional arithmetic error.Source: B. Finite-Precision Spectrum… · p. 2
- The finite-precision perturbation model assumes finite local perturbation bounds; nonfinite values lie outside its scope.Candidate displacement and candidate omission during selection are treated as distinct failure mechanisms.Source: B. Finite-Precision Spectrum… · p. 2
C. Stage-Wise Precision Configuration and Problem Formulation
- Stage formats use matching operand and arithmetic formats, with precision nondecreasing from stage to stage.Each real operation, including sequential accumulation, is rounded in that format, without fused multiply-add or block scaling.Source: C. Stage-Wise Precision Configuration… · p. 2, B. Finite-Precision Spectrum… · p. 2
- Each stage evaluates its grid within the retained region, and each candidate defines a next-stage window of width κ_m s_m.Overlapping windows are merged and coincident grid points are evaluated once per stage.Source: C. Stage-Wise Precision Configuration… · p. 2
- The design objective is to reduce precision-weighted spectrum-evaluation cost while maintaining RMSE within an allowed degradation of the FP64 two-stage baseline.The stated objective is a design goal; conditional local-minimum coverage is analyzed separately.Source: C. Stage-Wise Precision Configuration… · p. 2
III. PROPOSED PRECISION-AWARE HIERARCHICAL MUSIC
PAH-MUSIC refines the grid and does not decrease precision across stages, evaluating only within regions retained from the preceding stage. Candidate selection accounts for local peaks and separation, while final estimates are produced within the retained windows.
A. Precision-Aware Hierarchical Spectrum Search
- Each stage evaluates a refined grid within the retained region, using operands freshly quantized from the original FP64 subspace and steering vectors.Both operands and arithmetic use that stage’s format.Source: A. Precision-Aware Hierarchical… · pp. 2–3
- Candidate selection detects local peaks within each connected interval and greedily retains N representatives subject to minimum separation Δ_m.Plateaus are represented by the grid point nearest their midpoint; representatives are ranked by peak height.Source: A. Precision-Aware Hierarchical… · p. 3
- Quantization can create multiple local minima within one spectral basin, so ranking minima alone can assign several candidates to one source and omit another.A finer search cannot recover a reference peak excluded from its search region.Source: A. Precision-Aware Hierarchical… · p. 3
- Setting Δ_m to zero disables nearby-candidate suppression, while positive values may exclude closely spaced sources.Fewer than N accepted candidates constitute a search failure.Source: A. Precision-Aware Hierarchical… · p. 3
- Each candidate defines a next-stage window, and overlapping windows are merged for evaluation while original windows are retained for output association.At the final stage, overlapping windows may produce repeated estimates, so coverage alone does not ensure correct association.Source: A. Precision-Aware Hierarchical… · p. 3
B. Conditional Candidate Displacement and Window Design
The conditional displacement analysis bounds how far a retained candidate can lie from a reference grid minimum when both lie in the specified local region and the candidate’s perturbed denominator is no larger. This bound yields a sufficient window condition, with larger errors or flatter spectra requiring wider windows.
B. Conditional Candidate Displacement and Window Design
- If a retained candidate and reference grid minimum lie in the local region and the candidate’s perturbed denominator is no larger, Proposition 1 bounds their displacement.The proof compares the perturbed candidate value with the reference value and uses the error bound to constrain the distance.Source: B. Conditional Candidate Displacement… · p. 3
- The sufficient next-window condition includes a safety margin; larger errors or flatter spectra require wider windows.Using the grid-approximation bound g_k,m(s_m) ≤ s_m/2 requires an additional assumption, and a multistage application requires the conditions at every stage.Source: B. Conditional Candidate Displacement… · p. 3
C. Two-Stage Step-Size and Cost Analysis
The two-stage continuous cost model balances coarse-grid work against refinement work expanded by the candidate window. For fixed weights and window coefficient, it gives an unconstrained optimal coarse step and a cost-gain comparison with an FP64 two-stage baseline.
C. Two-Stage Step-Size and Cost Analysis
- Noise-subspace search complexity scales as O(M(M − N) sum_m G_m); the continuous approximation ignores overlap, angular-boundary clipping, and discrete endpoint counts.Actual costs instead use unique-point counts.Source: C. Two-Stage Step-Size and Cost Analysis · p. 3
- In the two-stage case, s is the coarse step, h the final step, κ the window coefficient, and ρc and ρf the coarse and final cost weights.Source: C. Two-Stage Step-Size and Cost Analysis · p. 3
- Increasing the coarse step reduces the first cost term but increases the second through window expansion.Source: C. Two-Stage Step-Size and Cost Analysis · p. 3
- For fixed positive weights and κ, Proposition 2 gives an unconstrained continuous minimum; at its optimal step, coarse and fine stages contribute equally to weighted search cost.This balances the number of coarse evaluations against the extent of local refinement.Source: C. Two-Stage Step-Size and Cost Analysis · p. 4
- With FP64 final evaluation and κref = 2, the cost-gain condition compares configurations through both precision weight and window coefficient.The prescribed-configuration ratio is ρcκ/2.Source: C. Two-Stage Step-Size and Cost Analysis · p. 4
A. Experimental Setup
The experiments test two-stage PAH-MUSIC in a separated two-source ULA scenario across six SNRs. Coarse-search formats and window settings are compared with an FP64 two-stage baseline, using FP64 final evaluation.
A. Experimental Setup
- The setup uses a 20-element half-wavelength-spaced ULA, two equal-power sources at −10.45° and 10.88°, and 40 snapshots.The source signals are independent circularly symmetric complex Gaussian signals with unit expected power per source.Source: A. Experimental Setup · p. 4
- Each SNR from −10 to 15 dB uses 500 Monte Carlo frames, with identical input data and shared FP64 covariance and eigendecomposition across strategies.Source: A. Experimental Setup · p. 4
- FP8 (E4M3FN), FP16, and FP32 coarse searches are compared with an FP64 two-stage baseline, and all use FP64 final evaluation.Candidate selection uses Δ1 = 5°; each original window returns one denominator minimum on the common final grid.Source: A. Experimental Setup · p. 4
- Table I lists the coarse-stage formats, window coefficients, coarse-grid steps, and refinement-window widths used in the comparison.Source: A. Experimental Setup · p. 4, Table I · p. 4
- The adopted per-evaluation weights are proportional to format bit widths and normalized to FP64; they proxy search cost rather than hardware time or energy.Source: A. Experimental Setup · p. 4
- For FP32, FP16, and FP8, selected window coefficients are 2, 3, and 5, respectively, below break-even values of 4, 8, and 16.The configurations are fixed across SNRs and use the corresponding optimal coarse steps.Source: A. Experimental Setup · p. 4
B. DOA Estimation and Candidate Coverage
Across the tested SNRs, the low-precision coarse searches produce RMSE curves close to the two-stage FP64 baseline. Paired outputs and true-source window coverage provide additional diagnostics of candidate retention and agreement.
B. DOA Estimation and Candidate Coverage
- All-frame RMSE is computed from source errors after one-to-one matching estimated and true DOAs to minimize the summed squared angular errors.For this one-dimensional criterion, pairing sorted angle lists attains the minimum.Source: B. DOA Estimation and Candidate Coverage · pp. 4–5
- At −10 dB, FP8 yields 0.342707° RMSE versus 0.343753° for the two-stage FP64 baseline.Fig. 2 plots all-frame DOA RMSE versus per-source SNR using 500 frames per SNR; all strategies use FP64 final evaluation.Source: B. DOA Estimation and Candidate Coverage · p. 5, Figure 2 · p. 5
- Paired output-difference RMSE is zero for FP16 and FP32 at all six SNRs; for FP8 it is zero from −5 to 15 dB and 0.015492° at −10 dB.Estimates are compared frame by frame with the two-stage FP64 outputs.Source: B. DOA Estimation and Candidate Coverage · p. 5
- At −10 dB, true-source joint window coverage is 94.8% for FP8, 99.2% for FP16, and 98.4% for FP32.Coverage is the fraction of frames in which all true DOAs lie in R2; it can differ from paired output agreement when an estimated peak remains in-window while a true DOA lies just outside.Source: B. DOA Estimation and Candidate Coverage · p. 5
C. Spectrum-Search Cost and Model Comparison
The selected configurations reduce observed weighted search counts relative to two-stage FP64 search, while theoretical cost comparisons help characterize the savings. The continuous-model decomposition predicts equal coarse- and fine-stage contributions at its analytic optimum.
C. Spectrum-Search Cost and Model Comparison
- For the adopted weights, predicted cost ratios are 0.5590 for FP8, 0.6124 for FP16, and 0.7071 for FP32.These ratios compare selected coefficients below their respective break-even thresholds with the FP64 baseline.Source: C. Spectrum-Search Cost and Model… · p. 5
- The continuous model predicts costs of 94.8683 for FP8, 103.9230 for FP16, and 120.0000 for FP32 versus 169.7056 for the FP64 baseline.Fig. 3 decomposes predicted cost into coarse and fine contributions, which are equal at the analytic optimum.Source: C. Spectrum-Search Cost and Model… · p. 5, Figure 3 · p. 5
- Observed weighted-count reductions are 44.54%–44.66% for FP8, 39.04%–39.16% for FP16, and 28.36%–28.51% for FP32 versus same-SNR mean two-stage FP64 counts.Fine points in overlapping windows are counted once; the selected configurations retain similar observed RMSE in the tested separated-source setting.Source: C. Spectrum-Search Cost and Model… · p. 5
V. CONCLUSION
PAH-MUSIC coordinates precision, angular grids, and candidate windows in hierarchical MUSIC search; selected low-precision coarse searches with FP64 refinement reduce weighted cost while keeping RMSE close to FP64 search.
V. CONCLUSION
- 28.36%–44.66% lower weighted spectrum-evaluation cost than FP64 two-stage search, with RMSE close to that baseline, in the tested separated two-source scenario.The selected configurations used FP8, FP16, or FP32 for coarse search and FP64 for refinement, across −10 to 15 dB.Source: §V · p. 5
- PAH-MUSIC coordinates stage-wise precision, angular grids, and candidate windows for MUSIC spectrum search.A conditional displacement bound relates local numerical perturbations to candidate-window coverage.Source: §V · p. 5
- With fixed precision weights and window coefficients, a two-stage continuous cost model yields an optimal coarse step and a cost-gain condition.Source: §V · p. 5