Extensions of outlier-robust time-of-flight estimation beyond the narrow-band Gaussian regime: multi-sidelobe and non-Gaussian analytic models, robustness to training-deployment mismatch, and phase-shift bank design

Apartsin, Cooper, and Intrator (IEEE TGRS, 2014) introduced a biosonar-inspired method for time-of-flight (ToF) estimation from a single echo in the threshold-SNR regime, where matched-filter maximum-likelihood estimates suffer outliers caused by autocorrelation sidelobes: a bank of phase-shifted unmatched filters yields partially decorrelated biased estimators; a weak classifier labels which correlation peak produced each estimate; and a confusion-matrix-based bias correction removes most of the outlier-induced error. The paper's acknowledged open issues split into two groups. The deferred multi-ping items - fusion across independent echoes with per-measurement outlier-uncertainty weighting - were resolved by the authors' own sequels (Part II, which introduced optimal weighted averaging over per-class conditional mean-square errors, and a companion paper in IEEE JSTARS (Journal of Selected Topics in Applied Earth Observations and Remote Sensing) with adaptive pulse control). Three items survive untouched: (i) the analytic theory (outlier probability, phase-shifted-estimator statistics, inter-estimator correlation) rests on narrow-band two-point Taylor expansions and Gaussian filtered-noise assumptions, with no extension to wider fractional bandwidths, multi-sidelobe outlier models, or non-Gaussian/non-white noise; (ii) the robustness of the classifier-driven fusion to mismatch between training and deployment statistics (confusion-matrix stability, SNR nonstationarity, the observed filter-bank-size dependence of generalization) is uncharacterized; and (iii) no principled procedure exists for optimizing the phase-shift bank (filter count, phase range, spacing) for a given operational SNR distribution. The problem is these three linked extensions: the rigorous general-regime theory, robustness characterization, and design rules for the method.

Background

Time-of-flight (ToF) estimation - recovering the arrival time $t_0$ of a known pulse $s(t)$ from a noisy echo $r(t)=k\,s(t-t_0)+n(t)$ - is the basic ranging primitive of sonar, radar, ultrasonic sensing, and subsurface exploration. The standard estimator is the matched filter: cross-correlate $r(t)$ with $s(t)$ and take the location of the maximum, which coincides with the maximum-likelihood estimator (MLE) of $t_0$ under additive Gaussian noise. Below a critical signal-to-noise ratio (SNR), the mean-square error (MSE) of this estimator rises abruptly - the classical threshold effect. For narrow-band pulses, quantified by the fractional bandwidth $V = B_c/F_c$ (the ratio of the centralized mean-square bandwidth $B_c$ to the central frequency $F_c$), the threshold effect is dominated by outliers: the pulse autocorrelation has secondary peaks (sidelobes) at time shifts $\pm 1/F_c, \pm 2/F_c, \dots$, and when filtered noise raises a sidelobe above the main peak the estimator locks onto the wrong peak, contributing a large bias-squared error. Operating near or below threshold is often unavoidable, because power-constrained platforms and bandwidth-limited channels cap pulse energy and usable spectrum. The biological motivation is bat biosonar: echolocating bats achieve reliable ranging across wide dynamic ranges with low-power emissions, and biologically inspired filter-bank echo processing has been studied since the early 2000s (reviewed by Simmons, Gaudette, and Warnecke, 2017).

Within this setting, Apartsin, Cooper, and Intrator developed a machine-learning estimation program across three papers (2012-2014), following an earlier regression-based precursor (JASA, 2012). Part I (IEEE TGRS, 2014) treats a single echo. Its analytic core is a two-point approximation - keep only the main lobe and the nearest sidelobe at $\pm 1/F_c$ of the ensemble-average correlation and Taylor-expand in the small quantities $V$ and the applied phase shift - under which the outlier probability of the matched-filter MLE for a pulse with no dc (zero-frequency) component of energy $E$ in white Gaussian noise of standard deviation $\sigma$ is approximately $\Pr(\text{outlier})\approx Q(\pi V E/\sigma)$, where $Q$ is the standard normal upper-tail function. On the constructive side, Part I replaces the single matched filter by a bank of $n$ phase-shifted unmatched filters - filters whose response applies a uniform spectral phase shift $\varphi$ to the pulse, called unmatched because the induced correlation peak is shifted and attenuated (peak location $t_\varphi^{M}\approx \frac{\varphi}{2\pi}\frac{1}{F_c}\frac{1}{1+V^2}$, with second-order attenuation in $\varphi$), so each filter yields a deliberately biased ToF estimator. Different shifts give partially decorrelated biased estimators - semicorrelated multiple estimators (SMEs) - with inter-estimator correlation approximately $\rho_{\varphi,\theta}\approx 1-\frac{(\theta-\varphi)^2}{2}\frac{V^2}{1+V^2}$. The $n$ biased peak locations form a feature vector for a weak multiclass classifier (a decision tree) trained to label which correlation peak - main lobe or one of the largest sidelobes - produced the matched-filter estimate. Using the classifier's confusion matrix $G=\{g_{i,j}\}$ (the probability that a measurement of true class $i$ is labeled class $j$), class priors, and class-conditional mean errors estimated from labeled training data, the method subtracts the label-conditional expected bias from each measurement, provably reducing the bias-squared part of the MSE via the identity $e_m^2 = e_o^2 + (\sum_j q_j b_j)^2 - \sum_j q_j b_j^2$, where $q_j$ are posterior label probabilities, $b_j$ the class-conditional expected biases, and $e_o^2$ the original MSE. In simulation (10,000 mixed-SNR samples, post-filter SNR 3-8 dB, $n=11$ filters with phase range $a=\pi/2$, $m=5$ classes), this reduced the root-MSE by up to about 30% relative to the classical MLE near threshold; the same simulations revealed an unexplained design tradeoff, with the larger bank ($n=11$) better at higher SNR and the smaller bank ($n=7$) generalizing better at lower SNR.

Part II (IEEE TGRS, 2014) extends the method to multiple pings: per-ping bias-corrected estimates are fused by optimal weighted averaging (OWA) with weights determined by per-class conditional MSEs (inlier versus outlier labels), estimated from training data. The companion JSTARS paper (2014) makes the scheme adaptive (adaptive OWA, AOWA): a $k$-$N$ stopping rule keeps pinging only until $k$ inlier-labeled returns have been observed, saving transmitted energy at fixed accuracy. What the sequels did not do - the deferred issues that survive - is the subject of this problem.

Problem Statement

The problem is the set of deferred issues from Part I that its own sequels (Part II and the JSTARS companion) left unresolved. It has three linked extensions - the rigorous general-regime theory, the robustness characterization, and the design rules - whose joint resolution places the outlier-robust, classifier-fused ToF estimation framework on rigorous, general ground beyond its narrow-band Gaussian proof-of-concept:

(1) Analytic model extensions. Extend the analytic theory - the closed-form outlier-probability model, the deterministic statistics of phase-shifted unmatched-filter estimators (peak location, peak attenuation, sidelobe asymmetries), the filtered-noise covariances, and the inter-estimator correlation law - beyond the two-point (main-lobe plus nearest sidelobe) Taylor expansions and Gaussian filtered-noise assumptions to which Part I and Part II are restricted. The target regimes, which the extension must cover and may treat separately (each regime on its own, rather than requiring one unified result valid in their intersection), are: wider fractional bandwidths, where the two-point narrow-band expansion fails and multiple sidelobes compete; multi-sidelobe outlier models beyond the nearest sidelobe (outlier-class analytics beyond the roughly five classes used in Part I's simulations); and non-Gaussian or non-white additive noise. The extension should include its consequences for the bias-correction and weighted-averaging fusion, not only the outlier probability of the matched-filter MLE.

(2) Robustness under training-deployment mismatch. The fusion requires class-conditional means, variances, and priors estimated from labeled training data and assumes the classifier's confusion matrix remains stable in deployment. Characterize - with guarantees, bounds, or a validated methodology - the degradation of the bias-corrected and optimal-weighted-averaging (OWA)-fused estimator under mismatch between training and operational signal/noise statistics: confusion-matrix instability, SNR mismatch or nonstationary SNR across pings, and the empirically observed dependence of generalization on filter-bank size (the $n=11$ versus $n=7$ tradeoff).

(3) Phase-shift bank design. Provide a principled procedure - an analytic design rule or a convergent optimization with an explicitly stated objective - for selecting the number of filters $n$, the phase range $a$, and the phase spacing of the phase-shift bank, balancing the degradation of the individual biased estimators against their partial decorrelation (via the correlation law) for a specified operational SNR distribution, in place of the empirical trial-and-error of Part I's simulations.

An answer that merely re-tunes the existing method inside the original narrow-band Gaussian regime does not answer this problem.

References

1. Alexander Apartsin, Leon N. Cooper, Nathan Intrator, "Time-of-Flight Estimation in the Presence of Outliers Part I—Single Echo Processing", IEEE Transactions on Geoscience and Remote Sensing 52(6):3382-3392, 2014. DOI: 10.1109/TGRS.2013.2272737, https://doi.org/10.1109/TGRS.2013.2272737
2. Alexander Apartsin, Leon N. Cooper, Nathan Intrator, "Time-of-Flight Estimation in the Presence of Outliers. Part II—Multiple Echo Processing", IEEE Transactions on Geoscience and Remote Sensing 52(7):3843-3850, 2014. DOI: 10.1109/TGRS.2013.2276919, https://doi.org/10.1109/TGRS.2013.2276919
3. Alexander Apartsin, Leon N. Cooper, Nathan Intrator, "Energy-Efficient Time-of-Flight Estimation in the Presence of Outliers: A Machine Learning Approach", IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7(4):1306-1313, 2014. DOI: 10.1109/JSTARS.2013.2295324, https://doi.org/10.1109/JSTARS.2013.2295324
4. Alexander Apartsin, Leon N. Cooper, Nathan Intrator, "Semi-coherent time of arrival estimation using regression", The Journal of the Acoustical Society of America 132(2):832-837, 2012. DOI: 10.1121/1.4730885, https://doi.org/10.1121/1.4730885
5. Leon N. Cooper, Nathan Intrator, "Time of Flight Estimation in the Presence of Outliers: A Biosonar-Inspired Machine Learning Approach", U.S. Army Research Office report (DTIC ADA597037), 2013. DOI: 10.21236/ADA597037, https://doi.org/10.21236/ADA597037
6. Kouei Yamaoka, Taishi Nakashima, Yukoh Wakabayashi, Nobutaka Ono, "Minimum-Spanning-Tree-Based Time Delay Estimation Robust to Outliers", IEEE Access, 2023. DOI: 10.1109/ACCESS.2023.3327011, https://doi.org/10.1109/ACCESS.2023.3327011
7. Klaus Brümann, Kouei Yamaoka, Nobutaka Ono, Simon Doclo, "Incremental Averaging Method to Improve Graph-Based Time-Difference-of-Arrival Estimation", 2025 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2025. DOI: 10.1109/WASPAA66052.2025.11230939, https://doi.org/10.1109/WASPAA66052.2025.11230939
8. Brian M. Sadler, Lili Huang, Zhengyuan Xu, "Ziv-Zakai Time Delay Estimation Bound for Ultra-Wideband Signals", 2007 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2007. DOI: 10.1109/ICASSP.2007.366738, https://doi.org/10.1109/ICASSP.2007.366738
9. James A. Simmons, Jason E. Gaudette, Michaela Warnecke, "Biosonar-inspired signal processing and acoustic imaging from echolocating bats", in Biologically-Inspired Radar and Sonar: Lessons from Nature, IET, 2017. DOI: 10.1049/sbra514e_ch2, https://doi.org/10.1049/sbra514e_ch2
10. Anup Das, "Deterministic and Bayesian Sparse Signal Processing Algorithms for Coherent Multipath Directions-of-Arrival (DOAs) Estimation", IEEE Journal of Oceanic Engineering, 2019. DOI: 10.1109/JOE.2018.2851119, https://doi.org/10.1109/JOE.2018.2851119
11. Zhenqian Wu, Youming Li, Xiaolong Zhang, Xiangpei Meng, Xinrong Lv, Yonghong Wu, "Multiple Anchors and RIS-Aided Localization Method in Complex NLOS Environments", IEEE Internet of Things Journal, 2024. DOI: 10.1109/JIOT.2024.3433948, https://doi.org/10.1109/JIOT.2024.3433948
12. Chao Zhou, Yueke Wang, Chunjie Qiao, Shen Zhao, Zhigang Huang, "High-accuracy ultrasonic temperature measurement based on MLS-modulated continuous wave", Measurement, 2016. DOI: 10.1016/j.measurement.2016.03.037, https://doi.org/10.1016/j.measurement.2016.03.037
13. Yuliya Shulgina, Maria A. Kostina, P.V. Sorokin, Marina Polonskaya, O.A. Kozhemyak, Andrey Soldatov, "Quantitative Measurement System of Liquid Products in Sealed Tank", Materials Science Forum 970, 2019. DOI: 10.4028/www.scientific.net/MSF.970.63, https://doi.org/10.4028/www.scientific.net/MSF.970.63
14. Kwang-Yul Kim, Euicheol Jeong, Seongil Kim, Yoan Shin, "Magnetic Anomaly Detection Using Continuous Angle Alignment of Three-Axis Magnetic Signal", IEEE Sensors Journal, 2019. DOI: 10.1109/JSEN.2018.2878890, https://doi.org/10.1109/JSEN.2018.2878890
15. Liron Darhi, Ariel Dvorjetski, Yehudit Aperstein, "Data-Driven Estimation of Helicopter Engine Power Using Regular Flight Data: A Machine Learning Approach", Electronics, 2025. DOI: 10.3390/electronics15010141, https://doi.org/10.3390/electronics15010141
16. Dorel Aiordachioaie, "Some Aspects of Time Delay Estimation Based on Cross-Correlation", Circuits, Systems, and Signal Processing, 2025. DOI: 10.1007/s00034-025-03323-2, https://doi.org/10.1007/s00034-025-03323-2