Uncertainty quantification in the Bradley-Terry-Luce model

Abstract: The Bradley-Terry-Luce (BTL) model is a benchmark model for pairwise comparisons between individuals. Despite recent progress on the first-order asymptotics of several popular procedures, the understanding of uncertainty quantification in the BTL model remains largely incomplete, especially when the underlying comparison graph is sparse. In this paper, we fill this gap by focusing on two estimators that have received much recent attention: the maximum likelihood estimator (MLE) and the spectral estimator. Using a unified proof strategy, we derive sharp and uniform non-asymptotic expansions for both estimators in the sparsest possible regime (up to some poly-logarithmic factors) of the underlying comparison graph. These expansions allow us to obtain: (i) finite-dimensional central limit theorems for both estimators; (ii) construction of confidence intervals for individual ranks; (iii) optimal constant of $\ell_2$ estimation, which is achieved by the MLE but not by the spectral estimator. Our proof is based on a self-consistent equation of the second-order remainder vector and a novel leave-two-out analysis.

• The paper develops uniform non-asymptotic expansions for the MLE and spectral estimators in sparse Bradley-Terry-Luce models.
• It introduces a novel leave-two-out analysis to achieve sharp ℓ2 and ℓ∞ error bounds alongside finite-sample central limit theorems.
• The results enable constructing tailored confidence intervals for individual merits, validating the MLE’s optimal efficiency in high-dimensional sparse regimes.

Uncertainty Quantification in the Bradley-Terry-Luce Model

Introduction

The Bradley-Terry-Luce (BTL) model provides a fundamental statistical framework for modeling pairwise comparisons between individuals, widely applied in areas such as web search, sports analytics, and social network analysis. Although the estimation accuracy of key procedures for the BTL model—including the Maximum Likelihood Estimator (MLE) and the spectral estimator (rank centrality)—has been thoroughly analyzed, uncertainty quantification, especially in sparse regimes, remains incomplete. The discussed paper (2110.03874) advances this front by developing non-asymptotic expansions for the MLE and spectral estimators under the sparsest possible Erdős-Rényi comparison graphs and analyzing the resulting statistical properties.

BTL Model Formulation and Existing Estimation Methods

The BTL model assumes latent merit parameters $\theta^*_i$ for $n$ individuals, with comparisons dictated by an Erdős-Rényi graph with edge probability $p$. On each edge, $L$ repeated pairwise comparisons are made, and outcomes are modeled with Bernoulli draws using the logistic function $\psi(\theta^*_i-\theta^*_j)$.

The main estimators are:

• Maximum Likelihood Estimator (MLE): The negative log-likelihood function derived from observed comparisons leads to a convex optimization problem, with efficient numerical solvers available.
• Spectral Estimator (Rank Centrality): Constructs a Markov chain whose stationary distribution captures ranking information, building on the spectral properties of the graph-based transition matrix.

Previously, global $\ell_2$ and $\ell_\infty$ risk guarantees for both estimators in sparse graphs were established, but the limiting distribution theory and finite-sample uncertainty quantification remained relatively unexplored, particularly at minimal connectivity ($p \sim n^{-1}$).

Main Theoretical Contributions

Non-Asymptotic Estimator Expansions

The central achievements are uniform, non-asymptotic expansions for the MLE and spectral estimator in the regime $p \sim n^{-1}$ (up to polylog factors):

• MLE Expansion:

${\theta}_i - \theta_i^* \approx \frac{b_i}{d_i}$

where $b_i$ aggregates the score deviations and $d_i$ normalizes them through the Hessian-like structure. This approximation is uniform over $i$, and the error terms are controlled to be negligible at the appropriate normalization.

• Spectral Estimator Expansion:

$${\theta}_i - \theta_i^* \approx \frac{b^\mathrm{spec}_i}{d^\mathrm{spec}_i}$$

Analogously, $b^\mathrm{spec}_i$ and $d^\mathrm{spec}_i$ incorporate the stationary distribution $\pi^*$ of the constructed Markov chain, resulting in a closely related but statistically distinct expansion.

These results address the critical regime where $np$ barely exceeds $\log n$, leading to sparse but connected comparison graphs. Previous work only proved asymptotic normality for much denser graphs ($p \gtrsim n^{-1/10}$).

Proof Strategy: Remainder Self-Consistency and Leave-Two-Out Analysis

The methodology diverges from earlier inverse Hessian approximations. Instead, sharp $\ell_2$ and $\ell_\infty$ bounds are obtained by constructing a self-consistent equation for the remainder vector, then employing a leave-two-out analysis. This approach decorrelates estimation errors from comparison graph realizations, which is essential for precise entrywise bounds in sparse graphs.

Statistical Applications

Finite-Dimensional Central Limit Theorem

The expansions directly yield finite-sample CLTs for any fixed-dimensional collection of estimated parameters. Both the MLE and spectral estimator satisfy

$$(\rho_1(\theta^*)({\theta}_1 - \theta^*_1), ..., \rho_k(\theta^*)({\theta}_k - \theta^*_k)) \leadsto \mathcal{N}_k(0, I_k)$$

where $\rho_i(\theta^*)$ enables data-driven uncertainty quantification. These CLTs hold in the sparse regime, extending previous results limited to dense graphs.

Construction of Confidence Intervals for Individual Ranks

The explicit expansions allow individualized confidence intervals for merit parameters, where interval lengths adapt to observed comparison counts. This enables credible intervals for ranks, useful in practical tasks such as sports analytics. The procedure leverages both finite-sample CLT results for high-priority individuals and conservative concentration bounds for others.

Optimal Constant in $\ell_2$ Estimation Risk

The paper gives precise formulas for $\ell_2$ estimation risk constants:

• For MLE:

$$\frac{1}{pL}\sum_{i=1}^n\left(\sum_{k \neq i} \psi'(\theta_i^* - \theta_k^*)\right)^{-1}$$

• For Spectral Estimator:

$$\frac{1}{pL}\sum_{i=1}^n\frac{\sum_{j \neq i}(e^{\theta_i^*}+e^{\theta_j^*})^2\psi'(\theta_i^*-\theta_j^*)}{\left[\sum_{j \neq i}(e^{\theta_i^*} + e^{\theta_j^*})\psi'(\theta_i^*-\theta_j^*)\right]^2}$$

Only the MLE attains the local minimax optimal constant; the spectral estimator is strictly suboptimal unless all parameters are equal.

Theoretical and Practical Implications

• Estimator Efficiency: The MLE maintains optimal efficiency with respect to local minimax risk, confirming its asymptotic properties from classical theory even in high-dimensional, sparse scenarios. The spectral estimator, widely used for scalability, entails increased uncertainty.
• Sparse Networks: The results bridge theoretical understanding in minimal-comparison regimes relevant for real-world networks, where connectivity is barely above the identifiability threshold.
• Statistical Inference: The derived finite-sample normal approximations justify sophisticated inference procedures, including simultaneous interval estimation for ranking and merit scores, which are robust against the sparsity inherent in practical applications.

Future Directions

Potential developments include:

• Improved analysis for regimes even closer to the connectivity threshold ($np \gtrsim \log n$).
• Extension to heterogeneous or non-Erdős-Rényi graphs, e.g., stochastic block models tailored to community detection.
• Adaptation to other pairwise comparison models and network inference contexts.

Conclusion

This work rigorously establishes sharp expansions and uncertainty quantification for the main estimators in the BTL model under sparsest possible comparison graphs. The results deliver both theoretical insights—refined CLTs, confidence intervals, and minimax optimality—and practical tools, notably for applications in ranking and networked data analysis. The introduced proof strategies suggest broader applicability for high-dimensional inference under structural sparsity.