Abstract
We present the first minimax risk bounds for estimators of the spectral measure in multivariate linear factor models, where observations are linear combinations of regularly varying latent factors. Non-asymptotic convergence rates are derived for the multivariate Peak-over-Threshold (POT) estimator in terms of the $p$-th order Wasserstein distance, and information-theoretic lower bounds for the minimax risks are established. The convergence rate of the estimator is shown to be minimax optimal under a class of Pareto-type models analogous to the standard class used in the setting of one-dimensional observations known as the Hall-Welsh class. When the estimator is minimax inefficient, a novel two-step estimator is introduced and demonstrated to attain the minimax lower bound. Our analysis bridges the gaps in understanding trade-offs between estimation bias and variance in multivariate extreme value theory.
Introduction: Capturing the Limits of Extreme Dependency
In Multivariate Extreme Value Theory (MEVT), understanding how multiple variables behave together during rare, catastrophic events—such as simultaneous financial crashes or extreme weather patterns—is critical. While estimating the "tails" of individual variables is well-established, capturing the dependency structure (how variables fail in tandem) is significantly harder. This structure is typically described by a "spectral measure" on a unit sphere.
A major challenge in this field is the scarcity of data in extreme regions. Existing methods, such as the Peak-over-Threshold (POT) estimator, have lacked a rigorous understanding of their fundamental performance limits, known as minimax risks. This research addresses the gap by analyzing the non-asymptotic trade-offs between estimation bias (from thresholding) and variance (from data sparsity).
Theoretical Framework and Methodology
The study focuses on multivariate linear factor models ($X = AZ$), where observations $X$ are linear combinations of latent, heavy-tailed factors $Z$. To quantify the error between estimated and true dependency structures, the authors utilize the $p$-th order Wasserstein distance. This metric is particularly effective because it allows for comparing discrete estimated measures with continuous theoretical truths, even when they do not share the same support.
The analysis is conducted over a "Pareto-like" model class (the Hall-Welsh class), which allows latent factors to deviate from a perfect Pareto distribution by a parameter $s$. This parameter controls the "smoothness" or the amount of noise in the tail.
Minimax Optimality and the "s" Regime Split
The core finding of the paper is that the optimal estimation strategy depends on the deviation parameter $s$:
- Small $s$ (High Perturbation): When the "noise" from the non-Pareto bulk is high, the standard POT estimator—which thresholds samples and uses the empirical distribution of their angles—is actually minimax optimal.
- Large $s$ (Low Perturbation): When the data is very close to a Pareto distribution, the POT estimator becomes "minimax inefficient." It cannot fully exploit the cleaner tail structure. To address this, the authors introduce a novel two-step estimator. This method first identifies the "directions" of the spectral measure using k-means clustering at a high threshold, then solves one-dimensional equations to find the "magnitudes" (the columns of matrix $A$).
Key Results and Validation
The paper establishes several landmark results:
- Lower Bounds: The best possible convergence rate for any estimator in this setting is proven to be $O(n^{-s})$.
- Estimator Performance: The POT estimator achieves a rate of $O(n^{-1/(2+\max\{1,\alpha\})})$ in the small $s$ regime, matching the lower bound. In the large $s$ regime, the two-step estimator successfully hits the $O(n^{-s})$ target.
- Numerical Confirmation: Synthetic experiments with varying tail indices ($\alpha$) and deviation parameters ($s$) validated these theoretical rates, showing clear linear relationships on log-log plots of error versus sample size.
Conclusion and Future Directions
This work provides the first comprehensive minimax analysis for spectral measure estimation in linear factor models. However, some challenges remain. The current analysis assumes parameters like $\alpha$ and $s$ are known, whereas they must be estimated in practice. Future work will focus on developing "adaptive" procedures that can achieve optimal rates without prior knowledge of these underlying distribution parameters, as well as refining the two-step estimator for broader general-purpose use.
Summary updated