Quantile-Conditional Variance Ratio Estimators

• The paper demonstrates that QCVR estimators robustly quantify tail indices and scale parameters via trimmed variance ratios, outperforming traditional methods.
• They use order statistics on quantile-trimmed subsamples to efficiently infer parameters for α-stable and Lévy-type distributions while maintaining computational simplicity.
• Empirical studies confirm that QCVR techniques enable high-powered goodness-of-fit tests and reliable estimation even in small-sample settings.

Quantile-Conditional Variance Ratio Estimators (QCVR) are a class of statistical techniques designed to quantify tail characteristics or scale parameters in heavy-tailed distributions, primarily through ratios of conditional variances calculated over quantile-trimmed subsamples. Distinctive for their robustness, computational simplicity, and versatility across symmetric α-stable as well as Lévy-type distributions, QCVR estimators leverage the order statistics of observed samples to infer distributional indices such as stability index α or scale c, and form the basis for high-powered goodness-of-fit testing. Their construction, statistical properties, implementation strategies, and empirical performance are summarized below, with particular reference to developments in symmetric α-stable settings (Pączek et al., 2022) and the one-sided Lévy law (Pączek et al., 2023).

1. Formal Definition and Statistical Construction

The QCVR framework centers on quantile-conditional variances. Given a sample $X_1,\ldots,X_n$ from a distribution $F_X$ with quantile function $Q_X(u)=F_X^{-1}(u)$, select quantile intervals $0 < a < b < 1$. The trimmed conditional variance

$$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$

uses only the data between empirical quantiles $Q_X(a)$ and $Q_X(b)$, where $\hat\mu_X(a, b)$ is the corresponding empirical mean. For symmetric $\alpha$-stable settings, the core QCVR statistic is the ratio

$$\hat N = \frac{\hat\sigma_X^2(a,b) + \hat\sigma_X^2(1-b,1-a)}{\hat\sigma_X^2(d, 1-d)}$$

with user-specified central quantile $F_X$0 and tail intervals $F_X$1. In the one-sided Lévy law, the corresponding scale-ratio is

$F_X$2

where $F_X$3 standardizes the observed conditional variance using its population counterpart at unit scale (Pączek et al., 2023).

2. Theoretical Rationale and Properties

The rationale for QCVR estimators lies in the monotonic functional relationship between the tail behavior (e.g., $F_X$4-stable index) and quantile-conditional variances in the tails, relative to those in the center. For symmetric $F_X$5-stable distributions, the mapping $F_X$6 is strictly decreasing and numerically one-to-one on $F_X$7, enabling inversion to yield tail index estimates. The conditional variance in the tails, $F_X$8, is highly sensitive to $F_X$9 for properly chosen quantile intervals (e.g., $Q_X(u)=F_X^{-1}(u)$0 in symmetric coordinates), while the central variance is only weakly $Q_X(u)=F_X^{-1}(u)$1-dependent (Pączek et al., 2022). For the Lévy law, the ratios of conditional variances or scales over distinct quantile intervals form pivotal quantities independent of the unknown scale $Q_X(u)=F_X^{-1}(u)$2 under the null, supporting their use in inference and testing (Pączek et al., 2023).

3. Estimation Algorithms and Quantile Selection

Application of QCVR estimators involves three procedural components:

1. Choice of Splits: For $Q_X(u)=F_X^{-1}(u)$3-stable distributions, recommended splits are $Q_X(u)=F_X^{-1}(u)$4 for moderately heavy tails and $Q_X(u)=F_X^{-1}(u)$5 for near-Gaussian cases. For the Lévy law, $Q_X(u)=F_X^{-1}(u)$6 and $Q_X(u)=F_X^{-1}(u)$7 are effective. Ensuring sufficient sample size within each slice ($Q_X(u)=F_X^{-1}(u)$8) is necessary for stable moment estimation.
2. Computation of QCV and Ratios: Empirically compute trimmed variances and the relevant ratio statistic.
3. Parameter Recovery: Invert the numerically pre-computed map (e.g., $Q_X(u)=F_X^{-1}(u)$9) to obtain point estimates. For scale-invariant ratios, direct comparison to the expected null value yields scale-free test statistics.

For both estimation and testing in small samples, quantile intervals can be centered to maintain subsample sizes, with adjustments based on the application focus—tail or body deviations (Pączek et al., 2022, Pączek et al., 2023).

4. Asymptotic Theory and Statistical Inference

Under mild regularity, each quantile-trimmed variance estimator is $0 < a < b < 1$0-consistent for its population target and satisfies a central limit theorem. The QCVR ratio estimator inherits asymptotic normality via the multivariate delta method: $0 < a < b < 1$1 for the $0 < a < b < 1$2-stable case, and similarly,

$0 < a < b < 1$3

for scale-invariant ratios under the Lévy law, where the limiting variance $0 < a < b < 1$4 is explicit and does not depend on nuisance parameters such as $0 < a < b < 1$5 (Pączek et al., 2023). The final index or scale estimator exhibits consistency, asymptotic normality, and small finite-sample bias that vanishes as $0 < a < b < 1$6 (Pączek et al., 2022).

5. Monte Carlo Performance and Empirical Behavior

Extensive simulation studies confirm the robust finite-sample performance of QCVR estimators. For symmetric $0 < a < b < 1$7-stable estimation, the $0 < a < b < 1$8 construction uniformly outperforms McCulloch's quantile-ratio estimator and often improves upon characteristic function regression when $0 < a < b < 1$9. The near-Gaussian-tuned $$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$0 estimator is superior for $$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$1. Maximum-likelihood approaches are generally slower and less accurate in moderate-to-high tail regimes (Pączek et al., 2022). For goodness-of-fit testing in the one-sided Lévy context, the QCVR-based test ($$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$2) demonstrates power close to 1 across a range of alternatives, especially in moderate samples ($$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$3), and maintains advantages in small-sample or light-tailed alternatives compared to MLE-based and transform-domain tests (Pączek et al., 2023).

6. Robustness, Ensembles, and Extensions

QCVR estimators exhibit empirical robustness to moderate departures from perfect symmetry or the presence of skewness, with empirical shifts in $$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$4 bounded by 0.05 for skewness parameters up to 1 and near-horizontal RMSE surfaces as $$\hat\sigma_X^2(a, b) = \frac{1}{\,\lfloor n b\rfloor-\lfloor n a\rfloor\,} \sum_{i=\lfloor n a\rfloor+1}^{\lfloor n b\rfloor} \big(X_{(i)} - \hat\mu_X(a, b)\big)^2$$5 (Pączek et al., 2022). Saliently, QCVR statistics extract sample features orthogonal to those targeted by characteristic function regression; thus, ensemble estimators—formed via simple averaging of, e.g., regression-CF and QCVR estimates—yield additional root-mean-square-error reductions. Extensions under consideration include:

• Joint estimation of location and scale via multiple QCV ratios,
• Fully Bayesian formulations using QCV statistics as summary data,
• Multivariate generalizations by componentwise or radial QCV,
• Adaptive or optimized split selection depending on preliminary estimates (Pączek et al., 2022).

7. Applications and Summary of Practical Advantages

QCVR estimators have been validated on real datasets, e.g., plasma-turbulence time series from fusion experiments, where they successfully detect and distinguish small departures from Gaussianity in temporal intervals, agree with bootstrap confidence intervals, and are congruent with standard goodness-of-fit diagnostics. Competing classical and semi-parametric methods display higher bias or fail to resolve near-Gaussian regimes (Pączek et al., 2022). The practical procedure is computable via order statistics and trimmed variances, with no need for special-function evaluation or density inversion, and is directly applicable to distributions with infinite raw moments.

In conclusion, QCVR estimators offer a transparent, statistically robust, and operationally efficient toolset for parameter estimation and hypothesis testing in heavy-tailed and Lévy-type distributions, with notable strengths in small-sample regimes, skewness tolerance, and ensemble construction. Their continued development opens further prospects in multivariate modeling and adaptive inference (Pączek et al., 2022, Pączek et al., 2023).