arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

Additive Fréchet Regression Gives a Curse‑of‑Dimensionality Free Tool for Complex Data | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Statistical MethodologyEnglishPublished

Additive Fréchet Regression Gives a Curse‑of‑Dimensionality Free Tool for Complex Data

November 22, 2026arXiv: 2611.26614v1

Researchers have introduced a new framework called additive Fréchet regression that applies to random objects living in any metric space, such as network graphs, probability distributions, or compositional data. The method builds on Fréchet regression, which estimates a conditional mean in a metric space, but it adds an additive structure to the predictor variables. This change removes the curse of dimensionality that plagues local Fréchet regression when the predictor vector has many components.

The estimator is built by projecting the conditional Fréchet variance onto an additive subspace. For each target point the algorithm solves a linear system of weight equations that guarantees a unique solution under mild density conditions. These smooth backfitting weights are then combined with an empirical weighted sum of squared distances to find the regression function. The resulting estimator converges at the non‑parametric rate O(n^{-2/5}), the same rate as a univariate kernel regression, and this rate does not depend on the number of predictors.

Simulation studies show that when the true regression surface truly is additive, the additive Fréchet estimator dramatically outperforms alternative methods such as network regression, single‑index Fréchet regression, and manifold‑valued regression. In scenarios where the underlying function is not additive, single‑index Fréchet regression shows the best median performance, while additive Fréchet regression remains second best.

The approach unifies many existing intrinsic regression methods. By choosing appropriate metric spaces – graph Laplacians for networks, Wasserstein distances for distributions, or spherical metrics for compositional data – the additive Fréchet framework reproduces and extends these special cases.

Limitations remain. The theoretical guarantees rely on the conditional Fréchet variance having an additive structure, assumptions on bandwidth and kernel smoothness, and regularity conditions on the metric space’s geometry. In spaces that are highly curved or not totally bounded, or when the smooth backfitting weights become negative, the optimisation may become non‑convex and minimisers may not be unique.

Source paper

Additive Fréchet Regression for Random Objects

Statistical MethodologyMathematics
View on arXiv