A 2025 preprint by Vimalajeewa, Müller, and Vidakovic (arXiv:2510.20828) proposed a double-threshold stochastic-resonance estimator that recovers a weak constant or time-varying signal from threshold-exceedance data by adding controlled noise, and showed that applying the estimator to wavelet approximation coefficients reduces the noise scale required for recovery. All of its experiments use Gaussian added noise and its analysis does not characterize how the noise distribution affects recovery, the two single-threshold estimators are combined by inverse-variance weights that ignore their covariance, and the wavelet decomposition level is chosen empirically. The problem is to extend the analysis to general noise distributions FF as the aim, including at least one two-parameter non-Gaussian family, and characterize their effect on the bias–variance trade-off, the optimal noise scale, and the relative efficacy of data-domain versus wavelet-domain stochastic resonance; and to practically implement the covariance-aware optimal-weight estimator with usable finite-sample covariance estimation — together with a principled, data-driven rule for selecting the wavelet decomposition level J0J_0 and the wavelet basis, which the paper's discussion leaves to empirical testing. None of these components has been resolved in the literature, so the problem remains open.

Background

Stochastic resonance (SR) is the counterintuitive phenomenon in which adding noise to a nonlinear system improves, rather than degrades, the detection or transmission of a weak signal. It was introduced by Benzi, Sutera, and Vulpiani (1981) to explain the periodic recurrence of ice ages and is surveyed in the review of Gammaitoni et al. (1998). One statistical branch of SR, developed by Greenwood, Ward, Wefelmeyer, Müller and coauthors since 1999, studies threshold detectors: a receiver cannot record the raw signal, only whether it crosses fixed thresholds. For two thresholds a<ba<b the recorded data are the indicators Yi=1(Xi>b)−1(Xi<a)Y_i=\mathbf{1}(X_i>b)-\mathbf{1}(X_i<a), and a subthreshold signal θ∈(a,b)\theta\in(a,b) is invisible without noise. If the signal can be convolved with independent identically distributed (i.i.d.) added noise ϵi\epsilon_i with a known invertible distribution function FF, so that Xi=θ+ϵiX_i=\theta+\epsilon_i, then the exceedance probabilities pa=F(a−θ)p_a=F(a-\theta) and pb=1−F(b−θ)p_b=1-F(b-\theta) identify θ\theta by quantile inversion: θ=a−F−1(pa)\theta=a-F^{-1}(p_a) and θ=b−F−1(1−pb)\theta=b-F^{-1}(1-p_b). Estimating pap_a and pbp_b by relative frequencies (or, for time-varying signals θ(ti)\theta(t_i), by Nadaraya–Watson kernel smoothers, which form local weighted averages of the indicator data) yields two estimators of the signal; their approximate variances follow from the central limit theorem and the delta method, a standard technique for approximating the distribution of a nonlinear function of an asymptotically normal statistic. Recovery quality is poor both at vanishing noise (no exceedances occur) and at overwhelming noise (the signal is drowned out), and best at an intermediate noise scale σ\sigma, the SR point. Greenwood, Ward, and Wefelmeyer (1999) quantified information retention of single- versus dual-threshold systems, Greenwood, Müller, Ward, and Wefelmeyer (2003) derived asymptotic mean squared errors and optimal bandwidths for kernel estimators from thresholded data, and Müller (2000) treated nonparametric regression from threshold data — in each case under a known noise distribution, with normal noise as the worked example.

The preprint of Vimalajeewa, Müller, and Vidakovic (2025) extends this line in two directions. First, it combines the two single-threshold estimators θ^a\hat{\theta}_a and θ^b\hat{\theta}_b into a double-threshold estimator using inverse-variance weights that treat the two estimators as uncorrelated. The preprint's Appendix A gives the covariance-corrected optimal weights of its Eq. (20), with the covariance obtained from analytic formulas based on the multinomial distribution of the indicator data or from bootstrap resampling. Second, it performs the whole procedure in a multiscale domain: the noisy signal is transformed by the discrete wavelet transform (DWT), a decomposition of a signal into time-localized frequency components computed by repeated filtering and downsampling, and thresholding is applied to the smooth approximation coefficients at a chosen decomposition level J0J_0. Because deeper decomposition concentrates low-frequency signal content into fewer, larger coefficients, higher J0J_0 raises the probability of threshold exceedance and lowers both the noise scale needed for detection and the average mean squared error (AMSE), the mean squared recovery error averaged over time points; but excessive decomposition destroys signal features and degrades recovery. Experiments on simulated 1D and 2D signals and on benchmark signals confirmed both improvements, using exclusively Gaussian added noise. Three elements of the framework are left unresolved by that preprint, and they jointly constitute the problem addressed here.

Problem Statement

The common objective is to turn the double-threshold multiscale stochastic-resonance recovery procedure of Vimalajeewa, Müller, and Vidakovic (arXiv:2510.20828) from a variance-weighted instantiation that has been experimented with only Gaussian noise and uses an empirically chosen decomposition level into a principled and robust method, by resolving three linked open issues in this workflow — the two the authors raise as future work, and the decomposition-level selection their discussion leaves to empirical testing. (1) Non-Gaussian noise: for i.i.d. added noise with a general invertible distribution FF beyond the normal distribution — for which the preprint's data-domain asymptotic variance formulas already hold — characterize the behavior of the double-threshold estimator and of its wavelet-domain variant, in particular the bias–variance trade-off as a function of the noise scale σ\sigma, the existence and value of an optimal noise scale, the relative efficacy of performing SR in the data domain versus the wavelet domain, and the consequences of a misspecified FF (the authors note that signal structure may remain inferable after a scale transformation), together with practical guidance for choosing or adapting FF when the true noise distribution is unknown. (2) Covariance-aware weighting: the current estimator combines θ^a\hat{\theta}_a and θ^b\hat{\theta}_b by inverse-variance weights that ignore their covariance C=cov(θ^a,θ^b)C=\mathrm{cov}(\hat{\theta}_a,\hat{\theta}_b); implement in practice the estimator using the optimal weights ω∗=(vb−C)/(va+vb−2C)\omega^{*}=(v_b-C)/(v_a+v_b-2C) of the preprint's Eq. (20), where va,vbv_a,v_b are the variances of θ^a,θ^b\hat{\theta}_a,\hat{\theta}_b and ω∗\omega^{*} is the weight on θ^a\hat{\theta}_a; develop computationally practical covariance estimators, and characterize their finite-sample behavior. (3) Decomposition-level selection: develop a principled, data-driven rule for choosing the wavelet decomposition level J0J_0 (and the wavelet basis) for general time-varying signals — the preprint's Section IV-C chooses levels by empirical testing — quantifying the trade-off whereby higher J0J_0 reduces the required noise scale and AMSE while excessive decomposition degrades recovery. Each characterization identifies the mechanism driving the effect and is supported either analytically under stated regularity conditions on FF or by a simulation study that reports its design and uncertainty. The non-Gaussian analysis is carried out for at least one two-parameter non-Gaussian family of FF (such as the Laplace or Student-tt law), and the guidance on choosing or adapting FF and the rule for the level J0J_0 and wavelet basis each state an explicit decision procedure together with the conditions under which it fails, rather than heuristic advice.

References

  1. Dixon Vimalajeewa, Ursula U. Müller, Brani Vidakovic, 'A Multiscale Approach for Enhancing Weak Signal Detection', arXiv:2510.20828 (2025). DOI: 10.48550/arXiv.2510.20828, https://arxiv.org/abs/2510.20828
  2. R. Benzi, A. Sutera, A. Vulpiani, 'The mechanism of stochastic resonance', Journal of Physics A: Mathematical and General 14, L453 (1981). DOI: 10.1088/0305-4470/14/11/006, https://doi.org/10.1088/0305-4470/14/11/006
  3. Luca Gammaitoni, Peter Hänggi, Peter Jung, Fabio Marchesoni, 'Stochastic resonance', Reviews of Modern Physics 70, 223 (1998). DOI: 10.1103/RevModPhys.70.223, https://doi.org/10.1103/RevModPhys.70.223
  4. Priscilla E. Greenwood, Lawrence M. Ward, Wolfgang Wefelmeyer, 'Statistical analysis of stochastic resonance in a simple setting', Physical Review E 60, 4687 (1999). DOI: 10.1103/PhysRevE.60.4687, https://doi.org/10.1103/PhysRevE.60.4687
  5. Priscilla E. Greenwood, Ursula U. Müller, Lawrence M. Ward, Wolfgang Wefelmeyer, 'Statistical Analysis of Stochastic Resonance in a Thresholded Detector', Austrian Journal of Statistics 32(1&2), 49–70 (2003). https://www.ajs.or.at/index.php/ajs/article/view/vol32%2C%20no1%262%20-%203 (no DOI registered for this article)
  6. Ursula U. Müller, 'Nonparametric regression for threshold data', Canadian Journal of Statistics 28(2), 301–310 (2000). DOI: 10.2307/3315980, https://doi.org/10.2307/3315980
  7. N. G. Stocks, 'Suprathreshold Stochastic Resonance in Multilevel Threshold Systems', Physical Review Letters 84, 2310 (2000). DOI: 10.1103/PhysRevLett.84.2310, https://doi.org/10.1103/PhysRevLett.84.2310
  8. Yong Xu, Juanjuan Li, Jing Feng, Huiqing Zhang, Wei Xu, Jinqiao Duan, 'Lévy Noise-Induced Stochastic Resonance in a Bistable System', arXiv:1207.3939 (2012). DOI: 10.48550/arXiv.1207.3939, https://arxiv.org/abs/1207.3939
  9. Qingbo He, Jun Wang, 'Effects of multiscale noise tuning on stochastic resonance for weak signal detection', Digital Signal Processing (2012). DOI: 10.1016/j.dsp.2012.02.008, https://doi.org/10.1016/j.dsp.2012.02.008