Work

One truncated likelihood expansion: estimation, testing, and classification as a single captured-fraction functional

Year: 2026
Type: article
Abstract: Parameter estimation, hypothesis testing, and signal classification are usually studied with separate machinery and separate notions of quality (estimator efficiency, error probability, reconstruction error). We show that, when each task is built from the truncated expansion of a log-likelihood in a fixed basis, all three are read off by one functional: the fraction of the relevant object captured by an order-m projection. Writing κm(z; ν) = ∥ projVm ˚z∥2 ν/∥˚z∥2 ν for the captured fraction of an object z in L2(ν), estimation reads κm on the score, detection on the log-likelihood ratio (its sequential change-point form reads the same object as a normalized information functional), and decomposition on the log-density. These are the same functional on different objects, not a single number; they reach the value one simultaneously exactly when the truncated series captures the centered log-density itself—the sufficient statistic and the carrier (lo˚g p ∈ Vm)—and otherwise lie strictly below one and increase as the basis is enriched. The paper gives a selfcontained projection-efficiency identity for the estimation branch; a per-branch termination theorem on exponential families; a completeness theorem showing that, under a heavy tail, the non-orthogonal fractional basis is dense in L2 exactly where the polynomial/Hermite system is not even square-integrable, so the basis restores the closure the orthogonal system loses; a bridge identifying the captured fraction of the score with the asymptotic relative efficiency and local power of both the estimation and detection branches; a local cross-branch equality (c(δ) → g as a two-hypothesis gap δ → 0 for location families); and a Gaussian covariance-shift anchor whose order-two projection is exact, so the sequential statistic equals the exact likelihood-ratio and classical change-point optimality applies directly. The detection identities of that anchor are machine-checked in Lean 4/Mathlib. A controlled study (fifteen pre-specified checks, all passing) confirms simultaneous termination on the Gaussian, joint rise of all readings on the Laplace, the local equality numerically, and—on Student-t and Cauchy laws— that the non-orthogonal fractional element captures Fisher information where a polynomial/Hermite basis is not even square-integrable. The same construction yields the reduction principle as a corollary. Mathematics Subject Classification (2020): 62F12 (primary); 62F03; 62L10; 62H30; 42C05.
Language: English
Open Access status: green
OpenAlex ID: W7167220516
FWCI: 5.867
Cites: 32
Cited by: 3
Related to: 10