Likelihood Quantile (LQ) Methods

• Likelihood Quantile (LQ) is a family of robust statistical methods that replace standard log-likelihood with quantile, deformed, or weighted transformations to mitigate outlier influence.
• LQ approaches introduce a tuning parameter that balances the trade-off between sensitivity to extreme values and efficiency, enhancing performance across parametric, Bayesian, and generative models.
• These methods have practical applications in outlier detection, decentralized algorithms, and robust model selection by improving estimation stability and computational tractability.

The Likelihood Quantile (LQ) encompasses a family of statistical methodologies that generalize classical likelihood-based inference and learning through the use of quantile, deformed, or weighted transformations of the likelihood function. This approach spans robust parametric inference, Bayesian computation for intractable models, neural generative modeling for outlier detection, and decentralized algorithms for multi-agent systems. LQ methods typically introduce a tuning parameter (e.g., a quantile level or distortion constant) that mediates the trade-off between sensitivity to outliers and efficiency, or, equivalently, tail versus bulk fit to the observed data. By embedding quantile logic into the likelihood principle, LQ supplant arithmetic averaging with rank-based, reweighted, or pivotal-statistic-driven information, yielding estimators and predictive procedures with provable robustness, tractable computation, or improved boundary behavior.

1. Fundamental Definitions and Core Variants

The LQ paradigm admits several instantiations unified by an underlying motif: the replacement or augmentation of the standard log-likelihood $L(\theta) = \sum_{i=1}^n \log f(x_i;\theta)$ with an alternative aggregating function emphasizing quantiles, power-distortion, or pivotal statistics.

1.1. Lq-likelihood (Deformed Log-likelihood):

For data $x_1,\ldots,x_n$ drawn from parametric density $f(x;\theta)$ and a robustness/efficiency parameter $q>0$, the order-$q$ deformed log-likelihood is

$\ell_q(\theta) = \sum_{i=1}^n L_q(f(x_i; \theta)),$

where

$$L_q(u) = \begin{cases} \frac{u^{1-q} - 1}{1-q}, & q \neq 1\ \log u, & q = 1 \end{cases}$$

so that as $q \to 1$, one recovers the classical log-likelihood (Guney et al., 2018, Ribeiro et al., 2020, Osorio et al., 2024).

1.2. Likelihood Quantile for Generative Objectives:

In LQ-based generative modeling (notably for outlier detection), the aggregated log-likelihood is replaced by a lower quantile: $$\mathcal{L}_{q\mathrm{NLL}}(\theta) = -\mathcal{Q}_{q}\left(\{\log p_\theta(x_i)\}_{i=1}^N\right),$$ where $\mathcal{Q}_{q}$ denotes the $x_1,\ldots,x_n$0-quantile of log-likelihoods (Taghikhah et al., 2023).

1.3. Quantile Implied Likelihood (QIL):

For Bayesian inference under intractable or computationally expensive likelihoods, the LQ is the density of a pivotal statistic $x_1,\ldots,x_n$1: $x_1,\ldots,x_n$2 with $x_1,\ldots,x_n$3 and $x_1,\ldots,x_n$4 denoting model quantiles at grid points (Karabatsos et al., 2018).

A commonality across these formulations is explicit parameterization by $x_1,\ldots,x_n$5 or quantile level, mediating the statistical influence of tails and the bulk of the observed data.

2. Theoretical Motivation and Robustness Properties

LQ methodologies are motivated by the limitations of traditional maximum likelihood (ML), including susceptibility to outliers, overemphasis of high-leverage points, and computational infeasibility for complex models.

• Robustness vs. Efficiency: For $x_1,\ldots,x_n$6 in Lq-likelihood, low-density observations (potential outliers) are exponentially downweighted via $x_1,\ldots,x_n$7. As $x_1,\ldots,x_n$8, classical ML is recovered, maximizing efficiency under the exact model; as $x_1,\ldots,x_n$9 decreases, robustness increases (Guney et al., 2018, Ribeiro et al., 2020, Osorio et al., 2024).
• Bounded Influence Function: In bounded-density families such as beta regression, the influence function under LQ remains bounded for $f(x;\theta)$0, preventing domination by extreme observations (Ribeiro et al., 2020).
• Trade-off Tuning: Practically, the tuning parameter $f(x;\theta)$1 governs this balance. Data-driven selection methods (e.g., grid searches for stability, thresholding of quadratic variation in parameter estimates, or minimization of efficiency loss relative to ML) are empirically validated (Ribeiro et al., 2020, Osorio et al., 2024).

3. Estimation and Inferential Procedures

The LQ approach modifies classical estimation and testing, resulting in reweighted score functions, alternative optimization criteria, and model selection tools.

Aspect	Classical MLE	LQ/Lq-likelihood
Score Equation	$f(x;\theta)$2	$f(x;\theta)$3
Posterior (Bayes)	$f(x;\theta)$4	$f(x;\theta)$5 (QIL)
Outlier Detection	Average-based log-likelihood	Loss based on $f(x;\theta)$6-quantile of log-densities
Model Selection	AIC: $f(x;\theta)$7	$f(x;\theta)$8

• Generalized Linear Models (GLMs): The Lq approach seamlessly extends to exponential family GLMs, yielding estimating equations and IWLS implementation with observation-level weights determined by the value of $f(x;\theta)$9 (Osorio et al., 2024).
• Robust Beta Regression: Explicit reparameterization ensures Fisher-consistency for Lq-ML estimators in the beta family, with estimation executed by solving weighted score equations for the regression and precision parameters (Ribeiro et al., 2020).
• Quantile Implied Likelihood (QIL): Estimation is performed by penalized maximum a posteriori (MAP) or via adaptive Metropolis-Hastings sampling on an explicitly tractable confidence-distribution-driven pseudo-likelihood (Karabatsos et al., 2018).
• Neural Flow-based Outlier Detection: In QuantOD, a normalizing flow is trained to maximize a low quantile of likelihoods over inlier batches, leading to improved separation at the inlier/outlier boundary. The inference threshold is also set via the empirical $q>0$0-quantile (Taghikhah et al., 2023).

4. Applications and Empirical Evidence

LQ methods have been empirically validated across parametric regression, generative modeling, Bayesian computation, and decentralized RL.

• Beta Regression: Under contamination, standard maximum likelihood estimation exhibits catastrophic MSE inflation, while Lq-based estimators maintain nominal error with negligible loss of efficiency for clean data. Data-driven selection picks $q>0$1 near 1 when no outliers are present and $q>0$2 under 5% contamination (Ribeiro et al., 2020).
• GLMs: Simulation and real-data analyses confirm that for moderate deviation from $q>0$3, robust AIC and deviance carry over, and the full suite of score, Wald, and likelihood-ratio tests remain asymptotically valid (Osorio et al., 2024).
• QuantOD Outlier Detection: On complex image datasets, quantile-based negative-log-likelihood ($q>0$4) improves FPR95 and AUROC over mean-based MLE and previous state-of-the-art unsupervised baselines, without requiring observed outliers for training (Taghikhah et al., 2023).
• Quantile Implied Likelihood: In Bayesian inference for models with intractable likelihoods (e.g., Student's $q>0$5 with large $q>0$6, g-and-h distributions, exponential random graphs), QIL achieves posterior inferences nearly matching exact or best-available approximations at a fraction of the computational cost (Karabatsos et al., 2018).
• Multi-Agent RL: In decentralized RL, the Likelihood Quantile (Time-Difference Likelihood) provides a per-sample weight that downweights “unusual” transitions due to other agents’ explorations, yielding marked improvements in convergence and stability for multi-agent planning (Lyu et al., 2018).

5. Extensions, Model Selection, and Diagnostic Tools

LQ methodology supports robustification of residual analysis, deviance and AIC-type criteria, as well as inference under linear hypotheses for both parametric and semi-parametric models.

• Residual Diagnostics: LQ-weighted residuals provide diagnostic clarity unattainable via classical MLE, clearly identifying outliers masked in standard approaches (Ribeiro et al., 2020, Osorio et al., 2024).
• Robust Model Selection: LQ-patched AIC ($q>0$7) and deviance metrics adapt classical model selection to the robustified setting, adjusting information criteria by functions of $q>0$8 and model dimension (Osorio et al., 2024).
• Hypothesis Testing: Wald, score, and likelihood-ratio analogues for LQ-based estimators retain $q>0$9 limiting distributions, provided appropriate scaling of the variance-covariance matrices is maintained (Osorio et al., 2024).

6. Theoretical Guarantees and Computational Considerations

Across formulations, LQ-based estimators and posteriors possess well-characterized large-sample properties.

• Consistency and Asymptotic Normality: Under regularity, LQ estimators are consistent, and their sampling distributions approach normality with covariance given by explicit information matrices (functions of $q$0), reducing to Fisher information as $q$1 (Guney et al., 2018, Ribeiro et al., 2020, Osorio et al., 2024, Karabatsos et al., 2018).
• Computational Tractability: In high-dimensional, intractable, or massive-data regimes, QIL and related LQ approaches collapse per-iteration cost by evaluating the likelihood only on quantile grids or batch-minima, allowing analysis even when exact likelihoods are unavailable (Karabatsos et al., 2018, Taghikhah et al., 2023).
• Tuning Parameter Selection: Data-driven criteria—such as stabilization of parameter paths or efficiency-loss thresholds—have been shown to select $q$2 near 1 when models are uncontaminated, and adapt automatically under contamination (Ribeiro et al., 2020, Osorio et al., 2024).

7. Domain-Specific Implementations and Variants

LQ methodology appears under a variety of more specialized guises or application frameworks.

• Block Maxima and Extreme Value Analysis: Lq-likelihood and its tilting variants improve parameter estimation for block maxima under the scarcity of extremes, as shown in risk assessment for natural hazards and public health settings (Jeffree et al., 2018).
• RL Coordination via TDL: In decentralized environments, LQ as time-difference likelihood discriminates between self- and other-agent driven non-stationarity, modulating learning rates adaptively and maintaining performance as system complexity scales (Lyu et al., 2018).
• Multivariate and Hierarchical Models: Via QIL, LQ logic supports scalable inference for multivariate skew-normal models and hierarchical choice models, leveraging pivotal-normality of quantile or depth-transformed statistics (Karabatsos et al., 2018).

LQ approaches thus enjoy breadth across robust parametric estimation, computational Bayesian modeling for intractable problems, unsupervised learning for high-dimensional data, and distributed control in dynamic systems. The central technical advance remains the explicit incorporation of quantile-based or tail-sensitive logic into the likelihood framework, yielding theoretical and empirical improvements across diverse statistical and machine learning domains.