Published October 26, 2026 | Version 2.0.0

OTAZ-II: Mathematical and AI Extensions of Exponentially Weighted Fusion — Duplicate-Record Consistency, Hilbert Sensitivity Bounds, Truncation Certificates and Constrained Calibration

Description

The earlier preprint in this series defined a pair-valued fusion operator: a value V that is a
softmax-weighted mean, and a score W that is a scaled log-sum-exp. That work established
closure, commutativity and associativity at a fixed coefficient, geometric and score
bounds, derivatives, entropy identities, and first-order perturbation bounds. This paper
resumes from that endpoint and contributes four guarded extensions, each with a written
proof and independent numerical evidence:
Duplicate-record consistency. A theorem showing that the full pair (V,W) can be
reproduced from deduplicated storage, in either of two opposite directions (w ± β−1ln
k), together with a bound on the error induced when multiplicities are estimated
incorrectly. The mathematics is simple; the contribution is the pair of executable rules
plus an a-priori damage bound, which turns a known hazard into a controlled
procedure.
Sharp sensitivity. The Hilbert distance between weight vectors never exceeds |β|·(max
δ − min δ). This yields a globally valid total-variation bound, TV ≤ tanh(|β|·range(δ)/4),
valid without a small-perturbation assumption and proved sharp by an explicit twopoint
family. A vector-valued counterpart and a cross-entropy ceiling follow.
Certificates, and an honest negative result. Truncating the score tail can be made
auditable: dropping mass τ bounds the value error by D·τ and makes the score gap
exactly computable. The experiments confirm the certificate in every case — and show
that certificates do not imply savings. With uniform scores, a 2 % tail budget keeps
98 % of records.
Constrained calibration. Minimum-variance weights are optimal only on the simplex,
and the unconstrained solution can be negative: an explicit analytic counterexample
gives weights (1.25, −0.25) with objective 0.875 against 1.000 on the simplex. This
paper gives an exact face-enumeration oracle for small problems, an equivalence
condition for the widely used inverse-MSE rule, and a calibration study of 2,400
replications that reports where weighting wins and where uniform averaging wins.
Every theoretical statement is proved or explicitly attributed to prior work. Every
reported number is regenerated by seeded code, reproduces byte-for-byte across reruns,
and is checked by an independent verifier of 27 checks that does not import the package
under test. Thirty-six unit tests accompany the reference implementation. The paper
states plainly what is prior art (softmax, log-sum-exp, the Hilbert–total-variation
inequality and its constant, forecast-combination weighting) and makes no priority, speed
or efficacy claim.

Files

OTAZ_II_Math_AI_Extensions_EN.pdf

Files (4.2 MB)

Name Size
md5:ea4f370317c93cd5a956c2e12da9a841
753.0 kB Preview Download
md5:944515520e09353d74a447276936d550
1.6 MB Preview Download
md5:9b7ef15789f7dd580ba72391af8d2ddd
1.8 MB Preview Download

Additional details

Related works

Is supplement to
Preprint: 10.5281/zenodo.23648070 (DOI)