📊 statistics

Orthogonal parametrisations of Extreme-Value distributions

This paper introduces and evaluates several orthogonal reparametrisations for the Generalised Extreme-Value, Generalised Pareto, and Gumbel distributions to address inference instability and improve interpretability, particularly when modeling rare events with small samples.

Original authors: Nathan Huet, Ilaria Prosdocimi

Published 2026-02-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nathan Huet, Ilaria Prosdocimi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster trying to predict the most extreme storms, floods, or heatwaves of the century. To do this, you use special mathematical tools called Extreme-Value Distributions. Think of these tools as "risk calculators" that help societies prepare for rare, dangerous events.

However, these calculators have a major flaw: they are notoriously difficult to tune.

The Problem: The "Wobbly Tripod"

Imagine trying to balance a camera on a tripod where the three legs are made of rubber and are all tangled together. If you try to adjust one leg (a parameter) to get a better picture, the other two legs wiggle uncontrollably.

In statistics, these "legs" are the parameters (numbers like location, scale, and shape) that define the distribution.

The Issue: When you try to estimate these numbers from a small amount of data (which is common for rare events), the estimates are highly correlated. Changing one number messes up the others.
The Result: The calculations become unstable, slow to converge, and hard to interpret. It's like trying to solve a puzzle where moving one piece shifts three others.

The Solution: Untangling the Knots

The authors of this paper, Nathan Huet and Ilaria Prosdocimi, propose a clever trick: Orthogonal Reparametrisation.

In everyday language, "orthogonal" means "at right angles" or "independent." Imagine untangling those rubber legs of the tripod so that each one stands perfectly straight and independent of the others. Now, if you adjust one leg, the others stay perfectly still.

They achieved this by applying a mathematical framework developed by Cox and Reid in 1987. They took the standard, messy formulas for these extreme-value distributions and rewrote them into new versions where the parameters don't "fight" each other.

The Three Main Tools They Fixed

The paper focuses on three specific types of "risk calculators":

The Gumbel Distribution: Used for things like annual maximum temperatures.
- The Fix: They found a way to rewrite the formula so that the "average" (location) and the "variability" (scale) are completely independent.
The Two-Parameter GEV (Generalized Extreme Value): Used for block maxima (e.g., the highest flood level in a decade).
- The Fix: They created a version where the "shape" of the curve and the "size" of the scale are independent.
The Generalized Pareto (GP): Used for things that exceed a high threshold (e.g., every flood above 5 meters).
- The Fix: They confirmed a method that separates the "threshold" from the "shape" of the extreme tail.

Why Does This Matter? (The Simulation)

The authors didn't just do the math on paper; they ran a simulation (a computer experiment) to prove it works.

The Experiment: They generated 1,000 fake datasets and tried to estimate the parameters using the old, tangled method and the new, untangled method.
The Result:
- Old Method: The estimates were all over the place, heavily linked to each other (like a tangled knot).
- New Method: The estimates were clean, stable, and independent (like straight, separate lines).

The Big Picture

Why should a regular person care?

Better Predictions: When these models are used in climate science or insurance, having stable parameters means more accurate risk assessments.
Faster Computing: In modern data science, we often use complex computer algorithms (like MCMC) to find answers. These algorithms run much faster and more reliably when the parameters are "orthogonal" (independent).
Simplicity: It turns a confusing, jumbled equation into a clean, understandable one.

Summary

Think of this paper as an instruction manual for reorganizing a messy toolbox. The tools (mathematical models) were always there, but they were so tangled that using them was a nightmare. The authors have untangled the wires, labeled the tools clearly, and shown that when you use them this new way, they work much faster, more accurately, and with much less frustration.

Note: The authors admit that for the most complex, three-parameter version of the main tool (the full GEV), the math is so tangled that it might be impossible to fully untangle without making approximations. But for the most common, simplified versions, they have found the perfect solution.

1. Problem Statement

Extreme Value (EV) distributions, specifically the Generalised Extreme-Value (GEV) and Generalised Pareto (GP) distributions, are fundamental tools for modeling rare events in fields like environmental science. However, statistical inference for these models faces significant challenges:

Instability: Parameter estimation is often unstable, particularly when fitting to small samples, which is common in extreme event analysis due to data scarcity.
Parameter Interdependence: The standard parameters (location $\mu$ , scale $\sigma$ , and shape $\xi$ ) exhibit strong correlations. This high sensitivity to outliers and parameter interaction leads to slow convergence in estimation algorithms (e.g., MCMC in Bayesian settings) and makes interpretation difficult.
Lack of Unified Reference: While orthogonal reparametrisations (which reduce parameter correlation) have been applied to specific distributions like the GP or Weibull, the necessary Fisher information matrix elements for EV distributions are scattered in the literature, and a comprehensive derivation for the main EV families is missing.

2. Methodology

The authors apply the theoretical framework developed by Cox and Reid (1987) to derive orthogonal reparametrisations.

Orthogonality Definition: A reparametrisation $(\psi, \lambda)$ is orthogonal if the Fisher cross-information between the parameter of interest ( $\psi$ ) and the nuisance parameters ( $\lambda$ ) is zero. Mathematically, this requires $E[\frac{\partial^2 l}{\partial \psi \partial \lambda_j}] = 0$ .
Derivation Process:
1. The authors calculate the elements of the Fisher information matrix for the GEV, Gumbel, and GP distributions.
2. They solve the resulting partial differential equations to find transformations that decouple the parameters.
3. Scope of Application:
  - Gumbel Distribution: Derived for both cases where location or scale is the parameter of interest.
  - Two-Parameter GEV: Derived for the case where the shape parameter $\xi \neq 0$ and the distribution's finite bound (lower or upper) is known (fixing the location parameter $\mu$ relative to the bound).
  - GP Distribution: Re-derives and extends existing results (Chavez-Demoulin and Davison, 2005).
- Exclusion: The authors explicitly note that deriving an orthogonal reparametrisation for the full three-parameter GEV distribution is intractable due to the complex structure of its Fisher information matrix and the resulting coupled system of partial differential equations.

3. Key Contributions

Compilation of Fisher Information: The paper consolidates scattered results regarding the Fisher information matrix for extreme value distributions into a single reference, facilitating future inference work.
New Orthogonal Parametrisations:
- Gumbel: Provides explicit transformations for both $\mu$ and $\sigma$ . For example, when $\sigma$ is of interest, the new location parameter is $\tilde{\mu} = (1-\gamma)\sigma + C_1(\nu)$ , where $\gamma$ is the Euler-Mascheroni constant.
- Two-Parameter GEV: Derives an orthogonal form where the scale parameter is transformed: $\tilde{\sigma} = C(\rho)\xi \exp((1-\gamma)\xi)$ .
- GP Distribution: Confirms and extends the orthogonal form for the shape parameter, offering an alternative to the standard scale-based transformation.
Theoretical Clarification: The paper clarifies why orthogonal parametrisations are possible for the GP and Gumbel distributions but generally impossible for the full three-parameter GEV without approximation, citing Huzurbazar (1956).

4. Results (Simulation Study)

The authors conducted a simulation study with $d=1000$ replications of sample size $n=100$ to validate the theoretical properties.

Decorrelation: The study measured the cross-correlation between estimated parameters using Maximum Likelihood Estimation (MLE) under both standard and orthogonal parametrisations.
Findings:
- The orthogonal parametrisations drastically reduced the correlation between parameter estimates compared to the standard forms.
- The empirical distributions of the correlations for the reparametrised estimates were concentrated around zero.
- Exception: In the GP case, reparametrising the shape parameter (keeping scale fixed) resulted in higher residual correlation compared to reparametrising the scale parameter. This supports the choice of previous literature (Chavez-Demoulin and Davison, 2005) to prioritize scale reparametrisation for the GP distribution.
Implication: These results suggest that even with moderate sample sizes ( $n=100$ ), orthogonal parametrisations significantly improve the stability and convergence properties of estimation algorithms.

5. Significance

Improved Inference Stability: By reducing the correlation between parameter estimates, these reparametrisations mitigate the instability issues inherent in fitting extreme value models to small datasets.
Computational Efficiency: Orthogonal parameters are known to accelerate the convergence of MCMC algorithms in Bayesian inference, making complex modeling of rare events more computationally feasible.
Interpretability: While the primary goal is statistical efficiency, orthogonal parameters often allow for a clearer separation of the effects of location, scale, and shape, aiding in the interpretation of risk models.
Practical Utility: The paper provides a "ready-to-use" reference for researchers and practitioners working with extreme value theory, particularly in environmental risk management, by offering explicit formulas for the most tractable and common sub-cases of EV distributions.