NC-ALPHEE: Noise-Controlled Hurst Estimation

• The paper introduces NC-ALPHEE, a method that extends ALPHEE with explicit additive noise correction to robustly estimate the Hurst exponent.
• It employs a neural network to adaptively aggregate multiple level-pairwise estimates, overcoming limitations of traditional averaging in noisy data.
• Empirical results indicate that NC-ALPHEE maintains nearly unbiased Hurst estimations across both noise-free and noisy regimes with low mean squared error.

Noise-Controlled ALPHEE (NC-ALPHEE) is a robust, noise-aware method for Hurst exponent estimation that extends the Average Level-Pairwise Hurst Exponent Estimator (ALPHEE) to settings in which additive noise contaminates self-similar signals. It incorporates noise mitigation, generates multiple level-pairwise estimates from signal energy pairs, and replaces traditional averaging with a neural network (NN) that combines these estimates adaptively. In noise-free data, NC-ALPHEE matches ALPHEE’s accuracy using both averaging and NN-based methods; under noise, it improves performance relative to traditional averaging and to existing techniques without requiring impractical level restrictions (Premarathna et al., 6 Oct 2025).

1. Definition and methodological position

NC-ALPHEE is situated within wavelet-based estimation of self-similarity. In this setting, self-similarity is quantified by the Hurst exponent $H$, which reveals long-term dependencies, and wavelet-based methods are effective because of their multi-scale analysis capability. ALPHEE, the immediate precursor, averages level-pairwise $H$ estimates derived from the energies of wavelet coefficients between pairs of scales, assuming negligible noise, that is, clean signals (Premarathna et al., 6 Oct 2025).

The defining modification introduced by NC-ALPHEE is explicit treatment of additive noise. Rather than assuming that fine-scale wavelet energies reflect only the underlying self-similar process, the method models the observed signal in a noise-aware way and corrects pairwise estimates at the level of both bias and variance. This shifts the estimator from a clean-signal regime to a robust estimator intended for real-world measurements in which additive noise would otherwise degrade accuracy.

A central implication is methodological rather than merely incremental: NC-ALPHEE preserves the level-pairwise logic of ALPHEE, but alters the statistical object being estimated. This suggests that the method is best understood not as a denoising wrapper around ALPHEE, but as a reformulation of the estimator under an explicit noise model.

2. Statistical formulation under additive Gaussian noise

NC-ALPHEE considers the observed signal

$Y(t) = X(t) + \varepsilon(t),$

with

$\varepsilon(t) \sim N(0, \sigma_{\varepsilon}^2).$

This explicit noise model is the basis for the method’s statistical correction scheme (Premarathna et al., 6 Oct 2025).

Under this model, the sum of squared wavelet coefficients at level $j$ follows a scaled chi-squared distribution: $$\sum_{k=1}^{n_j} d_{j, k}^2 \sim \chi^2_{n_j}(\sigma_X^2 2^{-j(2H+1)} + \sigma_{\varepsilon}^2).$$ The formulation gives the estimator access to the distributional consequences of noise at each scale, rather than treating noise as an unspecified perturbation.

This exact distributional behavior is the key extension beyond ALPHEE. The method derives closed-form expressions for both the $H$ estimate and its variance while accommodating the effects of noise. A practical consequence is that broader level ranges can be used in noisy signals, reducing bias relative to procedures that rely on ad hoc exclusion of noisy scales.

3. Pairwise NC-ALPHEE estimator and variance correction

For any two levels $j_1, j_2$, NC-ALPHEE derives the pairwise estimator

$$\widehat{H}^{NC}_{j_1,j_2} = \frac{1}{2(j_1-j_2)} \left[ \frac{ \psi(n_{j_1}/2)-\psi(n_{j_2}/2) }{ \log 2 } - \log_2 \left( \frac{ n_{j_1} \overline{d^2}_{j_1} - 2 \sigma_\varepsilon^2 \exp(\psi(n_{j_1}/2)) } { n_{j_2} \overline{d^2}_{j_2} - 2 \sigma_\varepsilon^2 \exp(\psi(n_{j_2}/2)) } \right) \right] - \frac{1}{2},$$

where $\psi$ is the digamma function, $H$0 is the number of coefficients at level $H$1, $H$2 is average wavelet energy, and $H$3 is noise variance estimated from finest-level wavelet coefficients (Premarathna et al., 6 Oct 2025).

The estimator is explicitly noise corrected at the level-pair level. ALPHEE’s formula is recovered by setting $H$4, so NC-ALPHEE contains the clean-signal method as a special case. This is important because it explains why the method maintains ALPHEE’s behavior in noise-free cases while changing performance materially in noisy regimes.

The paper also derives a variance formula for each pairwise estimate: $H$5 where $H$6 and $H$7 are explicit formulas involving digamma and trigamma functions, variance estimates, and noise corrections. This variance expression is not an ancillary result; it enables variance-weighted averaging and provides the basis for more sophisticated aggregation strategies.

In methodological terms, NC-ALPHEE treats each scale pair as a corrected local estimator of $H$8, then delegates the global estimate to an aggregation layer. The estimator’s architecture is therefore naturally modular: pairwise correction first, aggregation second.

4. Aggregation: from variance-weighted averaging to neural networks

The traditional aggregation strategy uses a weighted mean or median of pairwise $H$9 estimates, with weights inversely proportional to variance: $Y(t) = X(t) + \varepsilon(t),$0 In low-noise settings this is adequate, but in noisy settings some level-pairs become extremely unreliable, forcing practitioners to restrict analyses to a subset of level-pairs. The paper characterizes this as ad hoc level selection and notes that it is often insufficient and suboptimal, especially in high-noise regimes (Premarathna et al., 6 Oct 2025).

NC-ALPHEE’s principal aggregation innovation is a neural network-based strategy. All pairwise $Y(t) = X(t) + \varepsilon(t),$1 estimates $Y(t) = X(t) + \varepsilon(t),$2, across all valid level pairs, are input to a fully-connected feedforward neural network with 2–5 hidden layers, with nodes and other hyperparameters optimized via Bayesian optimization. The network is trained via cross-validation and early stopping to map the set of candidate $Y(t) = X(t) + \varepsilon(t),$3 estimates to the true $Y(t) = X(t) + \varepsilon(t),$4 supplied in simulations, minimizing mean-squared error.

The function of the network is nonlinear, context-dependent aggregation. Unlike fixed averaging rules, it can adapt the effective weighting of different level-pairs and learn to discount or amplify their contributions based on joint reliability, context, and interaction. The data describe this as particularly useful in noise, where individual pairwise variances and covariances are complex and non-Gaussian.

A common misconception is that the neural network merely replaces a mean with a more flexible combiner. The paper’s formulation is narrower and more specific: the NN is introduced because traditional averaging deteriorates under noise and requires impractical level restrictions, whereas the NN can leverage the entire ensemble of pairwise estimates without manual restriction of scale ranges.

5. Empirical behavior in noise-free and noisy regimes

In the noise-free regime, with $Y(t) = X(t) + \varepsilon(t),$5, Standard Wavelet Spectrum, ALPHEE, and NC-ALPHEE are reported to be comparable, accurate, and unbiased across a range of $Y(t) = X(t) + \varepsilon(t),$6 from 0.1 to 0.8. The analytical correction in NC-ALPHEE marginally reduces variance compared to ALPHEE, and NN-based and traditional averaging aggregations are equivalent in this regime (Premarathna et al., 6 Oct 2025).

The noisy regime is where the methods separate. Standard methods and ALPHEE suffer severe bias and underestimate $Y(t) = X(t) + \varepsilon(t),$7, especially at moderate-to-high noise or if level ranges are not aggressively restricted; in extreme noise, estimates can be negative or meaningless. NC-ALPHEE with traditional averaging outperforms Standard and ALPHEE by explicit noise correction and is less biased, but it remains sensitive to noise if aggregation includes unreliable fine levels. The paper therefore states that its best performance still requires careful, data-dependent range selection.

The strongest reported result concerns NC-ALPHEE with neural network aggregation. It maintains accurate, almost unbiased Hurst exponent estimation across all noise levels

$Y(t) = X(t) + \varepsilon(t),$8

without restricting to handpicked noiseless ranges. For moderate noise, mean squared error remains low, in the range $Y(t) = X(t) + \varepsilon(t),$9 to $\varepsilon(t) \sim N(0, \sigma_{\varepsilon}^2).$0. Figure 1 is described as showing predicted versus actual $\varepsilon(t) \sim N(0, \sigma_{\varepsilon}^2).$1 with the NN trace nearly perfectly aligned with the ideal $\varepsilon(t) \sim N(0, \sigma_{\varepsilon}^2).$2, and Figure 2 is described as showing much higher resilience across $\varepsilon(t) \sim N(0, \sigma_{\varepsilon}^2).$3 bins, with most samples classified in the correct bin even under strong noise.

These results support a specific interpretation of robustness. In this work, robustness does not simply mean lower bias at a fixed subset of scales; it means retaining access to multiscale information without arbitrary, fixed exclusions of certain scale pairs.

6. Scope, practical significance, and nomenclature

NC-ALPHEE is presented as a theoretically justified, practically robust estimator of the Hurst exponent for self-similar processes in domains where signals are unavoidably noisy, including natural phenomena analysis, financial modeling, biomedicine, finance, and climate (Premarathna et al., 6 Oct 2025). Its practical value lies in reducing the ad hoc intervention associated with level-range selection while preserving the multiscale structure of wavelet-based inference. The paper also notes that code is available in the Data Availability Statement.

The terminology can generate confusion because a distinct 2015 paper on controlled remote state preparation discusses a “noise-control philosophy” in a quantum communication context. That paper does not explicitly discuss “Noise-Controlled ALPHEE (NC-ALPHEE)” by name; rather, it studies how noisy environments can be engineered and exploited so that noise can be used as an important resource for quantum information tasks (Li et al., 2015). This suggests a conceptual analogy in the use of noise as a controllable element, but not an identity of framework, estimator, or application domain.

Within the Hurst estimation literature, the more precise characterization is therefore the following: NC-ALPHEE is an enhancement of ALPHEE that combines explicit additive-noise correction with adaptive aggregation over multiple level-pairwise estimates. Its distinctive claim is not merely that it handles noise, but that it maintains ALPHEE’s behavior in noise-free cases while substantially improving estimation reliability in noisy conditions without restrictive scale exclusions.