📊 statistics

Closed-form fractional radial links for elliptical Mahalanobis discriminant analysis

This paper introduces a closed-form, $\sqrt{n}$ -consistent fractional radial link for elliptical Mahalanobis discriminant analysis that achieves asymptotic Bayes optimality and empirically outperforms both QDA and tuned global GAMs on heavy-tailed financial and UCI datasets by leveraging a finite fractional-power stochastic-polynomial projection instead of generic spline smoothing.

Original authors: Serhii Zabolotnii

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Serhii Zabolotnii

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 security guard trying to sort people into two groups: "Friends" and "Strangers." You have a special radar that measures how far each person is from the center of their respective group. In the world of statistics, this distance is called the Mahalanobis radius.

For decades, the standard rule (called QDA) has been: "If the distance is small, they are a Friend; if it's large, they are a Stranger." This rule works perfectly if the groups are shaped like perfect, smooth balls (Gaussian distributions). But in the real world, data often looks more like a lumpy potato or a spiky star (heavy-tailed distributions). When the shape is weird, the old "ball" rule starts making mistakes.

This paper introduces a new, smarter way to sort people that adapts to the actual shape of the data. Here is the breakdown:

1. The Old Way vs. The New Way

The Old Way (QDA): Imagine you are using a rigid, pre-made template to judge people. It assumes everyone is a perfect circle. If the data is a circle, it's great. If the data is a spiky star, the template doesn't fit, and you misclassify people.
The Competitor (The "Fitted" GAM): Other researchers suggested using a flexible, stretchy rubber sheet (a spline) to mold around the data. You stretch and shrink this sheet until it fits perfectly. It works well, but it's like hiring a master tailor who takes a long time to measure, cut, and sew every single time. It requires a lot of tuning and guesswork.
The New Way (The Derived Link): This paper says, "Wait a minute! If we know the shape of the data (the generator), we don't need to guess or stretch a rubber sheet. We can write down the exact mathematical formula for the rule."
- The Analogy: Instead of hiring a tailor to measure a spiky star, the author realized that if you know the star's geometry, you can just print the exact template needed. No measuring, no stretching, no guesswork. It's a "closed-form" solution—a recipe you can write down on a napkin.

2. The Core Discovery: The "Radial Link"

The paper proves that for these weird, lumpy shapes, the rule for sorting isn't just about the distance; it's about a specific twist applied to that distance.

The Twist (The Link): In the perfect circle world, the twist is a straight line (simple). In the lumpy world, the twist is a curve.
The Breakthrough: The author shows that this curve isn't a mystery. It is mathematically locked to the shape of the data. You don't need to "learn" the curve from scratch; you can derive it directly from the data's shape.
The "Fractional Power" Trick: To get this curve right, the author uses a special set of building blocks called "fractional powers" (like square roots or 1.5 powers). Think of these as a set of Lego bricks that can snap together to build the exact curve needed, without needing the heavy machinery of the tailor's rubber sheet.

3. Why It Matters (The Results)

The author tested this new "printed template" against the "master tailor" (the fitted rubber sheet) and the "rigid ball" (the old QDA).

Vs. The Master Tailor: The new method was just as accurate as the tailor, but it was much faster and cheaper because it didn't need any tuning. It didn't lose any accuracy, but it saved a lot of time.
Vs. The Rigid Ball: When the data was "heavy-tailed" (very spiky, like financial market crashes or oil prices), the rigid ball failed. The new method crushed it, making significantly fewer mistakes.
The "Adaptive" Superpower: The best part is that the new method is smart.
- If the data is a perfect circle (Gaussian), it acts exactly like the old rigid ball (no harm done).
- If the data is a spiky star (heavy-tailed), it instantly switches to the complex, curved rule to save the day.
- It's like a chameleon that looks like a rock when it's on a rock, but turns into a bird when it's in the sky.

4. The "Heavy Tail" Reality Check

The paper tested this on real-world financial data (like oil prices and stock markets), which are notorious for being "spiky" and unpredictable.

The Finding: On these spiky datasets, the new method was significantly better than the old rigid rule.
The Caveat: The author admits that real-world data isn't perfectly symmetrical (it has a slight tilt). Even though the theory assumes perfect symmetry, the new method still worked better than the old one, proving it's robust enough to handle real-world messiness.

5. The "Math Police" (Lean 4)

To be absolutely sure their math wasn't full of holes, the author used a computer program called Lean 4 to check every single step of their logic. It's like having a robot lawyer verify that the blueprint for a bridge is mathematically sound before you build it. The robot gave a "green light," confirming the logic is flawless.

Summary

This paper gives us a new tool for sorting data.

Old Tool: A rigid ruler (fails on weird shapes).
Competitor Tool: A flexible tape measure (works well, but slow and fussy).
New Tool: A 3D printer that instantly prints the perfect shape based on the data's geometry.

It is just as accurate as the fussy tool, much faster to use, and significantly better than the rigid ruler when dealing with the messy, "heavy-tailed" data we see in finance and real life. It works automatically, adapting to the data without needing a human to tweak the settings.

Technical Summary: Closed-form fractional radial links for elliptical Mahalanobis discriminant analysis

Problem Statement
The paper addresses binary classification under the assumption of shared-generator elliptical class-conditional distributions. While Quadratic Discriminant Analysis (QDA) is the Bayes-optimal rule for Gaussian classes, it is suboptimal when classes are elliptical but non-Gaussian. Existing approaches, such as the Mahalanobis-distance Generalized Additive Model (GAM) proposed by Ghosh et al. (2025), treat the relationship between the Mahalanobis radii and the class label as an unknown smooth function to be fitted nonparametrically (e.g., via penalized splines). This paper asks whether, for elliptical distributions, the optimal link function can be derived analytically rather than fitted, thereby avoiding the computational cost and tuning complexity of nonparametric smoothing while retaining statistical efficiency.

Methodology
The core theoretical insight is that for two elliptical classes sharing a radial generator $g$ , the log-likelihood ratio (LLR) is an additive function of the squared Mahalanobis radii, $D^2_c(x) = (x - \mu_c)^T \Sigma_c^{-1} (x - \mu_c)$ . Specifically, the LLR takes the form:
$\Lambda(x) = \phi(D^2_1(x)) - \phi(D^2_0(x)) + \text{const}$
where $\phi = \log g$ is the radial link.

Structural Derivation: The paper establishes that the Bayes rule is a GAM in the two radii. Crucially, it proves a dichotomy: the link $\phi$ is affine (reducing the model to QDA) if and only if the generator is Gaussian. For any non-Gaussian elliptical generator (e.g., Student- $t$ ), the link is non-affine and non-identity.
Identifiability: The paper demonstrates that the radial link $\phi$ is identifiable from the within-class distribution of the squared Mahalanobis radius. This allows the link to be estimated as a functional of the observable radial law.
Estimation Strategy: Instead of fitting a generic spline, the authors propose a parametric "sieve" estimator. They approximate the derived link $\phi$ using a finite linear combination of fractional powers of the radius (a "fractional-power sieve"). The coefficients are estimated via logistic regression on these radial features. This approach leverages the stochastic-polynomial apparatus (Kunchenko, 2002; Zabolotnii, 2026d) to project the link onto a basis of powers $t^p$ (e.g., $p \in \{1, 0.5, 1.5\}$ ).
Formal Verification: The structural algebraic claims (the bridge between the LLR and the radial link, and the Gaussian/non-Gaussian dichotomy) are formally verified in Lean 4 without unproven placeholders.

Key Contributions

Structural Dichotomy (C1): The paper proves that the identity link (QDA) is Bayes-optimal exactly for Gaussian carriers. For all other normalizable shared-generator elliptical families, a non-identity link is strictly required. This bridges the gap between QDA and flexible GAMs.
Identifiability (C2): It establishes that the radial link is identifiable from the radial distribution, making the estimation problem well-posed and dictating a basis of powers of the radius.
Estimation Theory (C3): The paper proves that the plug-in estimator using a fractional-power sieve is $\sqrt{n}$ -consistent and asymptotically normal under finite-moment regularity conditions. Furthermore, the induced classifier is shown to be asymptotically Bayes-optimal in an iterated sieve limit (where sieve dimension $m \to \infty$ after sample size $n \to \infty$ ).
Empirical Performance (C4): The derived closed-form link is compared against a re-implemented, tuned global Mahalanobis-GAM (Ghosh et al., 2025) and standard QDA.

Results

Synthetic Data: On synthetic elliptical data, the identity link (QDA) fails to capture the true radial log-density for non-Gaussian generators (e.g., $R^2_{lin}$ drops to 0.24 for $t_4$ ), while the fractional-power basis captures the link with high fidelity ( $\kappa_2 = 0.95$ ).
UCI Benchmarks: On the Wine and Breast Cancer datasets (which are relatively light-tailed), the derived link performs statistically indistinguishably from the fitted spline GAM (no significant difference in accuracy) but significantly outperforms the identity link. It ties or slightly loses to QDA on these light-tailed sets, consistent with the theory that the advantage vanishes near the Gaussian limit.
Financial Time Series: On six heavy-tailed financial series (FX, Oil, S&P 500), the derived link is never significantly worse than the fitted GAM and significantly outperforms QDA on the heaviest-tailed series (Oil, S&P 500, JPY/USD). The performance advantage tracks tail-heaviness and covariance heterogeneity.
Robustness: Although the elliptical assumption is technically rejected on real financial data due to asymmetry (skewness), the derived link maintains its parity with the fitted GAM and its edge over QDA, suggesting robustness to mild deviations from elliptical symmetry.
Rate Simulation: Simulations confirm the $\sqrt{n}$ convergence rate and the theoretical "excess-risk dichotomy": QDA exhibits a non-vanishing approximation error floor in non-Gaussian regimes, whereas the derived link's excess risk vanishes as the sieve dimension increases.

Significance and Claims
The paper claims that for elliptical classification, the optimal decision boundary is not a matter of nonparametric modeling choice but a known functional form derived from the generator. The primary contribution is a derived, closed-form radial link that:

Avoids Tuning: It eliminates the need for spline smoothing parameter selection or bandwidth tuning required by GAMs.
Maintains Parity: It achieves statistical parity with a tuned, nonparametric GAM on real-world benchmarks.
Improves on QDA: It provides a provable and empirical advantage over QDA specifically in regimes characterized by heavy tails and covariance heterogeneity, where the curvature of the generator is significant.
Theoretical Guarantees: It offers $\sqrt{n}$ -consistency and asymptotic Bayes-optimality, grounded in a structural framework verified by machine-checked formal proofs.

The authors position this work not as a replacement for all nonparametric methods, but as a targeted, theoretically grounded solution for elliptical data that recovers the benefits of flexibility without the cost of nonparametric estimation complexity. The method is adaptive: it behaves like QDA when the data is Gaussian and diverges to capture heavy-tailed structure when necessary.