📊 statistics

One Truncated Likelihood Expansion: Estimation, Testing, and Classification as a Single Captured-Fraction Functional

This paper unifies parameter estimation, hypothesis testing, and signal classification by demonstrating that all three tasks can be evaluated through a single functional—the captured fraction of a log-likelihood expansion in a fixed basis—which reveals their shared efficiency, termination conditions, and optimality across diverse distributions, including those where traditional orthogonal bases fail.

Original authors: Serhii Zabolotnii

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Serhii Zabolotnii

Original paper licensed under CC BY 4.0 (https://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 trying to understand a complex mystery, like a weather pattern or a stock market trend. Traditionally, statisticians use three different "flashlights" to solve different parts of the puzzle:

Estimation: One flashlight tries to find the exact center of the data (like finding the average temperature).
Testing: Another flashlight tries to decide if two things are different (like asking, "Is today's weather a fluke, or is the climate changing?").
Classification/Decomposition: A third flashlight tries to break the data down into its simplest building blocks to see how well it can be rebuilt.

Usually, these three flashlights are built by different teams, use different batteries, and give you three different numbers to judge how well they are working.

The Paper's Big Idea: One Master Flashlight

This paper argues that you don't need three different tools. You only need one single "Master Flashlight" that can be used in three different ways to solve all three problems at once.

The author, Serhii Zabolotnii, proposes a method where you take a complex mathematical description (called a "log-likelihood") and break it down into a series of simpler pieces, like building a tower out of blocks.

The "Captured Fraction": The core of the paper is a single number called the "Captured Fraction." Think of this as a "Percentage of Truth Captured."
- If your tower of blocks perfectly reconstructs the mystery, you have captured 100% of the truth.
- If your tower is missing some blocks, you have captured, say, 70% of the truth.

The paper shows that whether you are trying to find an average, test a hypothesis, or classify data, you are actually just measuring this same "Percentage of Truth" on three different objects.

When estimating, you measure how much of the "score" (the clue) is captured.
When testing, you measure how much of the "difference" between two scenarios is captured.
When classifying, you measure how much of the "shape" of the data is captured.

The "Magic" Twist: Heavy Tails and Broken Ladders

Here is where the paper gets really clever.

Imagine you are trying to build a ladder to reach a high shelf.

The Old Way (Polynomials/Hermite): Most statisticians use a standard ladder made of straight, rigid rungs (polynomials). This works great for normal, predictable data (like a Gaussian bell curve).
The Problem: But what if the data is "heavy-tailed"? This means the data has wild, extreme outliers (like a stock market crash or a massive storm). In these cases, the standard ladder breaks because the rungs don't fit the shape of the shelf. The math literally explodes or becomes undefined.

The Paper's Solution:
The author introduces a "Fractional Ladder." Instead of rigid, straight rungs, this ladder uses flexible, curved pieces (called "fractional elements" or $sgn(x)|x|^a$ ) that can bend to fit the wild, heavy tails of the data.

The Claim: On these wild, heavy-tailed datasets, the old ladder falls apart, but the new fractional ladder holds firm. It can still capture a high percentage of the truth (up to 96% in some tests) where the old method captures nothing because it can't even be built.

The "Perfect Match" (The Anchor)

The paper also identifies a special case called the "Gaussian Anchor."
Imagine a specific type of puzzle (a Gaussian distribution with a shift in variance) where the "Captured Fraction" hits exactly 100% very quickly (at just the second level of the tower).

Because this match is perfect, the author was able to use a computer proof system (Lean 4) to mathematically verify that their logic is 100% correct for this specific case. It's like having a blueprint that a super-computer has double-checked to ensure no errors exist.

What the Paper Does NOT Claim

It does not claim that this method works for every possible problem in the world.
It does not say that the three numbers (Estimation, Testing, Classification) are always the same number. They are usually different, but they rise and fall together. They only become identical when the data fits the "perfect" model or when the difference between two scenarios is tiny.
It does not apply this to real-world medical or clinical data in this paper; the tests were done on "synthetic" (computer-generated) data to prove the theory works.

In Summary

The paper unifies three separate statistical tasks into one single concept: "How much of the truth have we captured with our current set of building blocks?"

It proves that by using a flexible, "fractional" set of blocks instead of rigid ones, we can solve problems involving extreme, wild data that previously broke traditional statistical tools. It's like discovering that while a standard ladder is great for a flat wall, you need a flexible, custom-shaped ladder to climb a jagged, stormy cliff.

Technical Summary: One Truncated Likelihood Expansion

Problem Statement
In statistical inference, parameter estimation, hypothesis testing, and signal classification (decomposition) are traditionally treated as distinct tasks, each governed by separate theoretical machinery and quality metrics: estimator efficiency, error probability/detection delay, and reconstruction error, respectively. The paper posits that these tasks are often studied in isolation, obscuring their shared structural limits. The central problem addressed is whether these three tasks can be unified under a single object and a single measure of "completeness," particularly in regimes where classical orthogonal expansions (e.g., Hermite polynomials) fail, such as under heavy-tailed distributions.

Methodology
The paper proposes a unified framework based on the truncated expansion of the log-likelihood in a fixed, possibly non-orthogonal basis $\{\phi_i\}$ . The core methodology involves:

The Master Object: A truncated Kunchenko likelihood series defined as:
$\log p(x; \theta) = \psi_0(\theta) + \sum_{i \le m} k_i(\theta) (\phi_i(x) - \psi_i) + R_m(x; \theta)$
where coefficients are solved via a normal system (generalized least squares) and $R_m$ is the orthogonal truncation residual.
Three Tasks from One Series: The three statistical tasks are derived as elementary operations on this single series:
- Estimation: Differentiation with respect to the parameter ( $\partial_\theta$ ) yields the score function, leading to a polynomial-maximization estimator.
- Detection: Differencing between two hypotheses ( $\Delta$ ) yields the log-likelihood ratio (LLR), leading to a Neyman–Pearson test. A sequential version is formed by composing this difference with time.
- Decomposition: The norm of the truncation residual provides the reconstruction error.
The Unifying Functional: The quality of all three tasks is measured by a single functional, the captured-fraction functional $\kappa_m$ :
$\kappa_m(z; \nu) = \frac{\| \text{proj}_{V_m} \mathring{z} \|_\nu^2}{\| \mathring{z} \|_\nu^2}$
where $V_m$ is the subspace spanned by the centered basis functions, and $\mathring{z}$ is the centered object of interest (score, LLR, or log-density).
- For estimation, $z$ is the score $S$ ; the reading is the asymptotic relative efficiency $g$ .
- For detection, $z$ is the LLR $\Lambda$ ; the reading is the captured-ratio $c$ .
- For decomposition, $z$ is the log-density; the reading is $R^2$ .
Basis Selection: The framework utilizes non-orthogonal fractional bases (e.g., $\rho_a(x) = \text{sgn}(x)|x|^a$ ) specifically to address heavy-tailed distributions where polynomial/Hermite systems are not square-integrable.

Key Contributions

Structural Unification (C1): The paper demonstrates that estimation, detection, and decomposition are not three separate methods but one object read in three ways. Their quality is governed by the same functional $\kappa_m$ applied to different objects. Crucially, the paper clarifies that these readings are generally different numbers (e.g., on the Laplace law, $g$ and $c$ rise at different rates), but they reach the value of 1 simultaneously if and only if the basis captures the sufficient statistic (Theorem 1).
Completeness via Basis Switch (C2): A major theoretical contribution is a completeness theorem (Theorem 3) showing that under heavy tails (tail index $\alpha$ ), a non-orthogonal fractional system with exponents accumulating in $(0, \alpha/2)$ is dense in $L^2(\nu)$ . This restores the $L^2$ closure that orthogonal polynomial systems lose in these regimes, allowing the captured fraction to approach 1 where Hermite expansions diverge.
Operating Characteristics (C3): The captured fraction is identified as the operating characteristic for both estimation and detection.
- Estimation: $g$ equals the asymptotic relative efficiency (ARE) of the estimator against the maximum likelihood estimator.
- Detection: $g$ determines the local asymptotic power and the Pitman relative efficacy of the test.
- This provides a unified power statement for heavy-tailed inference where orthogonal bases cannot define efficacy.
Local Cross-Branch Equality (C4): The paper proves that in the local limit of a vanishing hypothesis gap ( $\delta \to 0$ ), the detection reading $c(\delta)$ converges to the estimation reading $g$ (Proposition 2), formalizing the captured-fraction form of Local Asymptotic Normality (LAN).
Gaussian Anchor and Machine Verification (C5): A Gaussian covariance-shift model serves as an "anchor" where the order-two projection is exact. The paper provides machine-checked proofs (in Lean 4/Mathlib) verifying that on this anchor, the sequential statistic equals the exact likelihood ratio, ensuring classical CUSUM optimality applies directly.
Reduction Principle (C7): The framework yields an invariance corollary: any transformation leaving the basis evaluation invariant leaves all branch readings (score, LLR, reconstruction error) unchanged. This recovers the reduction principle as a generative consequence of the single-series factoring.

Results
The paper validates its theory through a controlled synthetic study with 15 pre-specified checks:

Termination: On the Gaussian family, all readings ( $g, c, R^2$ ) reach 1 exactly when the basis includes the sufficient statistic (e.g., $x^2$ for variance shift).
Joint Rise: On the Laplace distribution, all readings rise toward 1 as the basis is enriched, though at different rates depending on the object.
Local Equality: Numerical experiments confirm that as the hypothesis gap $\delta$ shrinks, the detection reading converges to the estimation reading.
Heavy-Tailed Advantage: On Student- $t_3$ and Cauchy distributions, the fractional basis captures significant Fisher information (e.g., $g \approx 0.96$ for $t_3$ ) where the polynomial Gram matrix diverges. The orthogonal basis fails to provide defined efficacy in these regimes.
Completeness: As fractional exponents densify, the captured fraction approaches 1 for both light-tailed (Laplace) and heavy-tailed (Student- $t_3$ ) targets.

Significance and Claims
The paper modestly claims to provide a structural unification rather than a numerical identity. It argues that the "single scalar" view of inference is only valid on specific terminating anchors (exponential families) or in local limits. The primary significance lies in:

Unifying the Theory: Showing that three distinct statistical branches share a single functional measure of completeness.
Extending to Heavy Tails: Demonstrating that non-orthogonal fractional bases are not merely convenient but necessary to restore $L^2$ completeness and define inference efficacy on heavy-tailed laws where classical polynomial systems fail.
Operational Clarity: Linking the captured fraction directly to asymptotic relative efficiency and local power, providing a single metric to evaluate the "goodness" of a basis across estimation, detection, and decomposition tasks.

The work does not claim that the three readings are identical numbers in general, but rather that they are structurally linked, rise together with basis enrichment, and coincide under specific conditions (sufficiency or local limits). The machine-checked core is limited to the Gaussian anchor, while the heavy-tail and general-basis results rely on analytical proofs and numerical verification.

Technical Summary: One Truncated Likelihood Expansion

More like this