pith. sign in

arxiv: 2604.10482 · v2 · submitted 2026-04-12 · 📊 stat.ME

The Fr\'echet correlation coefficient for heterogeneous random objects

Shuaida He , Yangzhou Chen , Xin Chen This is my paper

Pith reviewed 2026-05-10 16:24 UTC · model grok-4.3

classification 📊 stat.ME
keywords Fréchet correlationmetric space regressionnon-Euclidean dataFréchet varianceconditional dependencebootstrap testheterogeneous objects
0
0 comments X p. Extension

The pith

The Fréchet correlation coefficient measures the relative reduction in Fréchet variance of a response after conditioning on a predictor.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper introduces the Fréchet correlation coefficient for regression problems in which responses and predictors take values in the same or different metric spaces. It defines the coefficient as the proportion of Fréchet variance removed from the response by conditioning on a given predictor, providing a directional and scale-free analogue to R-squared that requires no parametric model. The authors supply a partition-based estimator that sidesteps direct estimation of the conditional Fréchet mean function, together with a wild bootstrap test for Fréchet conditional mean dependence and supporting asymptotic results. The approach therefore lets analysts rank heterogeneous predictors by explanatory strength in settings where classical correlation tools do not apply.

Core claim

We introduce the Fréchet correlation coefficient (FCC), defined as the relative reduction in the Fréchet variance of the response after conditioning on a specific predictor. FCC is directional, model-free, and interpretable on a unit-scale, attaining one under almost sure functional dependence and zero when the Fréchet mean is invariant to conditioning. We propose a novel partition-based estimator that avoids explicit nonparametric estimation of the conditional Fréchet mean function, thereby improving both computational efficiency and flexibility. A tailored wild bootstrap algorithm is further developed for testing the Fréchet conditional mean dependence.

What carries the argument

The Fréchet correlation coefficient, defined as one minus the ratio of the conditional Fréchet variance of the response given the predictor to the unconditional Fréchet variance.

If this is right

  • FCC reaches exactly one when the response is a deterministic function of the predictor almost surely.
  • FCC equals zero whenever conditioning on the predictor leaves the Fréchet mean of the response unchanged.
  • The partition estimator converges to the population FCC without requiring consistent estimation of the full conditional Fréchet mean function.
  • The wild bootstrap test controls type-I error and detects dependence with increasing power as sample size grows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same variance-reduction construction could be applied to other Fréchet functionals such as quantiles or higher moments.
  • Feature-ranking pipelines that mix scalar, functional, and graph-valued variables could adopt FCC for automatic predictor ordering.
  • The testing procedure might be extended to settings with many competing predictors by suitable multiple-testing adjustment.

Load-bearing premise

The underlying metric spaces admit well-defined and unique Fréchet means so that variances and conditional variances are meaningful.

What would settle it

Generate independent response and predictor objects whose unconditional and conditional Fréchet means coincide; if the estimated FCC then fails to concentrate at zero, the coefficient or its estimator is invalid.

Figures

Figures reproduced from arXiv: 2604.10482 by Shuaida He, Xin Chen, Yangzhou Chen.

Figure 1
Figure 1. Figure 1: Washington bike-sharing example. Panel (a) plots mean normalized daily rental [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Empirical rejection rate at the nominal level [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical rejection rates at the nominal level [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
read the original abstract

Modern regression analysis often involves responses and predictors taking values in the same or distinct metric spaces. To rank non-Euclidean heterogeneous predictors in regression by explanatory strength, analogous to the classical $R^2$, we introduce the Fr\'echet correlation coefficient (FCC), defined as the relative reduction in the Fr\'echet variance of the response after conditioning on a specific predictor. FCC is directional, model-free, and interpretable on a unit-scale, attaining one under almost sure functional dependence and zero when the Fr\'echet mean is invariant to conditioning. We propose a novel partition-based estimator that avoids explicit nonparametric estimation of the conditional Fr\'echet mean function, thereby improving both computational efficiency and flexibility. A tailored wild bootstrap algorithm is further developed for testing the Fr\'echet conditional mean dependence. We establish asymptotic theory and evaluate power through extensive simulations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the Fréchet correlation coefficient (FCC) as a directional, model-free measure of dependence between a response and predictor in (possibly distinct) metric spaces, defined as the relative reduction in Fréchet variance of the response after conditioning on the predictor. It attains 1 under almost-sure functional dependence and 0 when the Fréchet mean is invariant to conditioning. The authors propose a partition-based estimator that avoids explicit nonparametric estimation of the conditional Fréchet mean, develop a tailored wild bootstrap for testing Fréchet conditional mean dependence, establish asymptotic theory, and evaluate performance via simulations.

Significance. If the central claims hold, the FCC supplies a practical, unit-scale analogue of R² for heterogeneous random objects that permits ranking predictors by explanatory strength without parametric assumptions. The partition estimator's computational advantages and the wild bootstrap for inference are concrete strengths, as is the provision of asymptotic guarantees. These elements address a genuine gap in methodology for non-Euclidean regression settings.

major comments (2)
  1. [§2 (Definition of FCC)] §2 (Definition of FCC): The interpretive claim that FCC attains zero precisely when 'the Fréchet mean is invariant to conditioning' presupposes uniqueness of the Fréchet mean. In general metric spaces the argmin of m ↦ E[d(Y,m)²] need not be unique (e.g., any point on a geodesic segment or antipodal points on a circle under symmetric mass). While the conditional Fréchet variance itself remains well-defined, the zero-attainment statement becomes ambiguous without an explicit uniqueness assumption or a separate argument for the non-unique case. The partition estimator and wild bootstrap derivations appear to maintain this uniqueness tacitly; the paper should either add the assumption or clarify the behavior when multiple minimizers exist.
  2. [§4 (Asymptotic theory)] §4 (Asymptotic theory): The consistency and asymptotic normality results for the partition estimator are stated under the maintained assumption that Fréchet means exist and are unique. It is unclear whether the rates or limiting distributions continue to hold (or require modification) when the uniqueness condition is relaxed, which directly affects the validity of the bootstrap test for conditional mean dependence in spaces where non-uniqueness can occur.
minor comments (2)
  1. [§2] The notation distinguishing unconditional and conditional Fréchet variances could be made more explicit (e.g., by adding a subscript for the conditioning variable) to improve readability when comparing the ratio definitions.
  2. [Simulation section] Simulation section: The power curves in the figures would benefit from explicit reporting of the number of Monte Carlo replications and the precise metric spaces used in each scenario to facilitate reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and thoughtful review of our paper. The comments highlight important points about the assumptions underlying the Fréchet correlation coefficient and its asymptotic properties. We have carefully considered each point and plan to make revisions to strengthen the manuscript accordingly.

read point-by-point responses

  1. Referee: The interpretive claim that FCC attains zero precisely when 'the Fréchet mean is invariant to conditioning' presupposes uniqueness of the Fréchet mean. In general metric spaces the argmin of m ↦ E[d(Y,m)²] need not be unique (e.g., any point on a geodesic segment or antipodal points on a circle under symmetric mass). While the conditional Fréchet variance itself remains well-defined, the zero-attainment statement becomes ambiguous without an explicit uniqueness assumption or a separate argument for the non-unique case. The partition estimator and wild bootstrap derivations appear to maintain this uniqueness tacitly; the paper should either add the assumption or clarify the behavior when multiple minimizers exist.

    Authors: We appreciate the referee's observation regarding the potential non-uniqueness of Fréchet means in general metric spaces. The Fréchet variance is defined as the infimum of the expected squared distance, which is uniquely determined even if the minimizing point is not. Consequently, the FCC, being the relative reduction in this variance, attains zero whenever the conditional Fréchet variance equals the unconditional one, without requiring uniqueness. However, the phrasing 'the Fréchet mean is invariant to conditioning' does implicitly assume uniqueness for its interpretation. To address this, we will revise Section 2 to explicitly state the assumption of uniqueness of the Fréchet mean (both unconditional and conditional) and clarify that the zero-attainment holds via equality of variances. We will also add a brief discussion on the non-unique case, noting that the variance-based definition remains valid. This assumption is standard in Fréchet regression literature and is maintained in our derivations for the estimator and bootstrap. revision: yes

  2. Referee: The consistency and asymptotic normality results for the partition estimator are stated under the maintained assumption that Fréchet means exist and are unique. It is unclear whether the rates or limiting distributions continue to hold (or require modification) when the uniqueness condition is relaxed, which directly affects the validity of the bootstrap test for conditional mean dependence in spaces where non-uniqueness can occur.

    Authors: The asymptotic theory in Section 4 is indeed established under the assumption of unique Fréchet means, which ensures the well-posedness of the conditional Fréchet mean function and the differentiability conditions used in the proofs. When uniqueness is relaxed, the rates and limiting distributions may not hold in the same form, as the estimator could converge to a set rather than a point, potentially affecting the bootstrap approximation. We will revise the manuscript to make this assumption explicit in the statements of the theorems and add a remark acknowledging that extending the theory to non-unique cases is beyond the current scope and left for future work. Under the maintained assumptions, the wild bootstrap remains valid for testing the null of no conditional mean dependence. revision: yes

Circularity Check

0 steps flagged

No significant circularity; FCC is a direct definitional extension of relative variance reduction

full rationale

The paper introduces the Fréchet correlation coefficient explicitly as the relative reduction in Fréchet variance of the response after conditioning on the predictor, attaining 1 under almost sure dependence and 0 when the Fréchet mean is invariant. This construction is self-contained and does not reduce by the paper's own equations to a fitted parameter, a self-citation chain, or an ansatz smuggled from prior work. The maintained assumption that metric spaces admit unique Fréchet means is stated upfront as a prerequisite for the variance quantities to be well-defined, rather than derived from the coefficient itself. No load-bearing step equates a prediction to its own input or renames a known result as a new derivation. The partition estimator and bootstrap are presented as computational tools under this assumption, without circular justification.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claims rest on the existence of Fréchet means and variances in the metric spaces and on the validity of the stated asymptotic results for the estimator.

axioms (1)
  • domain assumption Random objects take values in metric spaces that admit well-defined Fréchet means and variances.
    Required to define the Fréchet variance and the correlation coefficient itself.

pith-pipeline@v0.9.0 · 5442 in / 1248 out tokens · 78310 ms · 2026-05-10T16:24:22.012910+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [1]

    A new coefficient of correlation

    22 Sourav Chatterjee. A new coefficient of correlation. Journal of the American Statistical Asso- ciation, 116(536):2009–2022,

    work page 2009

  2. [2]

    Association and independence test for random objects

    Hang Zhou and Hans-Georg Müller. Association and independence test for random objects. arXiv preprint arXiv:2505.01983 ,

    work page arXiv