Regression and Inference in Metric Spaces

52 concepts24 papers1 proposal

When we think of regression, we usually picture fitting a line or a curve through a cloud of points. In this classical setting, both our predictors ($X$) and our responses ($Y$) live in Euclidean space—a world where we can freely add vectors, multiply them by scalars, and calculate averages.

But what happens when the data points themselves are complex, structured objects? Consider predicting a patient's brain connectivity network from their age, or forecasting the daily distribution of a city's bike-share usage based on weather conditions. Networks, phylogenetic trees, and probability distributions do not live in a vector space. You cannot add two networks together, nor can you multiply a probability distribution by two without violating the fundamental laws of probability.

Because these complex objects lack a natural linear structure, standard statistical tools break down. To analyze them, we must step away from Euclidean arithmetic and operate within the broader realm of general metric spaces, relying entirely on the concept of distance.

Redefining the Mean via Distance

To perform regression without vector addition, we must first reinvent the most fundamental building block of statistics: the average.

In a standard Euclidean space, the arithmetic mean of a dataset can be defined as the point that minimizes the sum of squared distances to all other points. Mathematically, for a random variable $Y$, its expected value is:

$$\mathbb{E}[Y] = \arg\min_{y} \mathbb{E}[\|Y - y\|^2]$$

This optimization-based definition is incredibly powerful because it requires no addition or scalar multiplication—only a distance metric. By replacing the Euclidean norm with a general metric $d(\cdot, \cdot)$, Maurice Fréchet generalized the concept of a mean to abstract spaces. The Fréchet mean of a random variable $Y$ taking values in a metric space $\mathcal{M}$ is defined as:

$$\arg\min_{y \in \mathcal{M}} \mathbb{E}[d^2(Y, y)]$$

As noted by Song (2025), this formulation allows us to quantify central tendency and dispersion in spaces that completely lack linear algebraic properties. Remarkably, this structural shift is so robust that even the foundational principles of calculus and ordinary differential equations can be recovered in complete metric spaces by substituting linear paths with metric-based path equivalence classes, as demonstrated by Tabor (2002).

The Core Mechanism of Fréchet Regression

Fréchet regression extends this concept from a static mean to a conditional mean. In classical linear regression, we look for a conditional expectation function $m(x) = \mathbb{E}[Y|X=x]$, which minimizes the conditional expected squared error. In a metric space, we define the regression function $m(x)$ as the point in the metric space $\mathcal{M}$ that minimizes the weighted or conditional expected metric distance:

$$m(x) = \arg\min_{y \in \mathcal{M}} \mathbb{E}[d^2(Y, y) \mid X = x]$$

As highlighted by Müller et al. (2022), this simple elegant twist generalizes the concept of a conditional mean to responses residing in completely non-Euclidean spaces.

To estimate this from a finite sample of data, researchers use two primary strategies:

  1. Global Fréchet Regression: This approach optimizes a global objective function across the entire metric space, mapping the predictors globally to the response space, as explored by Song et al. (2023).
  2. Local Fréchet Regression: For a target predictor value $x$, we compute a locally weighted Fréchet mean of the observed responses $Y_i$. The weights $s_i(x)$ are determined by a smoothing kernel, giving more influence to observations whose predictors $X_i$ are close to $x$:

$$\hat{m}(x) = \arg\min_{y \in \mathcal{M}} \sum_{i=1}^n s_i(x) d^2(Y_i, y)$$

This local averaging mechanism allows us to model highly non-linear relationships. For instance, Chen et al. (2020) leveraged this framework to build "Wasserstein Regression," where the response variables are probability distributions evaluated within the intrinsic geometry of Wasserstein space.

Euclidean Regression:
[Predictor X] ---> [Linear / Kernel Weights] ---> Arithmetic Mean (Σ w_i * Y_i) ---> [Euclidean Output Y]

Fréchet Regression:
[Predictor X] ---> [Linear / Kernel Weights] ---> ArgMin_y Σ w_i * d²(Y_i, y)  ---> [Metric Space Output Y]
                                                                                      (Distribution, Network, Tree)

Beating the Curse of Dimensionality

While local Fréchet regression is highly flexible, it suffers from a severe bottleneck: the curse of dimensionality. As the number of predictors in $X$ grows, the volume of the predictor space grows exponentially. Training points become isolated and sparse, causing local smoothing weights to degrade and rendering the model inaccurate (Iao et al., 2024; Cui et al., 2026).

To scale metric space regression to high-dimensional inputs, researchers have developed three core paradigms:

1. Sufficient Dimension Reduction (SDR)

Instead of performing regression on the raw high-dimensional predictors, SDR aims to find a low-dimensional subspace of the predictors that contains all the necessary information to predict the metric-space response. Weng et al. (2023) and Weng et al. (2024) developed sparse Fréchet SDR frameworks that utilize nonconvex optimization to perform feature selection and dimension reduction simultaneously, even when the number of features vastly exceeds the number of observations. For distribution-to-distribution mappings, Zhang et al. (2022) bypassed these limitations by pairing Wasserstein metrics with universal kernels to capture complex nonlinear dependencies.

2. Single-Index Models

A classic way to compromise between flexibility and dimensionality is the single-index model, which projects high-dimensional predictors onto a one-dimensional index via a parameter vector $\beta$. Bhattacharjee (2023) introduced Single Index Fréchet Regression (SIFR), allowing researchers to project predictors down to a single dimension before executing local Fréchet regression. This approach drastically improves statistical efficiency and enables interpretable hypothesis testing for non-Euclidean responses.

3. Deep Fréchet Regression

To handle highly complex, high-dimensional inputs without manual feature engineering, recent frameworks merge deep learning with metric geometry. Iao et al. (2024) introduced Deep Fréchet Regression (DFR), which utilizes deep neural networks to map structured objects to a low-dimensional manifold, employing a reverse-mapping technique to output valid metric-space objects. To preserve interpretability, Cui et al. (2026) combined deep learning with the single-index paradigm in their DeSI (Deep Single-Index) framework, using a neural network to learn the optimal 1D projection before performing local Fréchet smoothing.

Inverting the Problem: Metric Covariates

The challenges of metric spaces are not restricted to the response variable $Y$; sometimes, the predictors themselves are the non-Euclidean objects. For example, you might want to predict a binary clinical outcome (healthy vs. diseased) using a patient's brain connectivity matrix as the input covariate.

Standard classification models like logistic regression expect vector inputs. To bridge this gap, Lin et al. (2025) extended binary regression and classification to accommodate covariates in general metric spaces.

Rather than trying to force the metric object into a vector space, a powerful alternative is to use distance profiles. As formalized by Dubey et al. (2022), a distance profile represents the one-dimensional distribution of distances from a single object to all other data points in the sample. By shifting the focus from the abstract objects themselves to their distance profiles, researchers can perform ranking, quantile estimation, and classification using purely coordinate-free, distance-based operations.

Temporal Dynamics and Regularization

Data objects in metric spaces often evolve over time, introducing temporal dependencies and structural shifts. Analyzing these behaviors requires adapting classical time-series and regularization concepts.

For instance, when tracking a brain connectivity matrix or a distribution over time, the underlying trend might experience sudden jumps or sharp discontinuities due to external shocks. Standard smoothing techniques blur these transitions. To solve this, Lin et al. (2021) developed a total variation regularized Fréchet regression framework. By penalizing variation directly within the metric space, their method successfully isolates and preserves sharp discontinuities over time.

Standard Smoothing:      [Continuous Curve]  --> Blurs sudden jumps
Total Variation Fréchet: [Step-like Changes] --> Preserves sharp discontinuities in Metric Space

Furthermore, when dealing with full non-Euclidean time series, traditional inference tools that assume independent data fail. Jiang et al. (2025) extended "self-normalization" techniques to non-Euclidean time series, providing a robust mathematical framework for two-sample hypothesis testing and change-point detection under complex temporal dependencies.

By substituting the rigid requirements of linear algebra with flexible, distance-centric optimization, metric space regression provides a unified, mathematically rigorous toolkit for modeling the rich, non-linear structures inherent in modern data.

Key Foundational Work

Wasserstein Regression (2020) — contributed 1 concept elements to this direction.
Differential equations in metric spaces (2002) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_metric_space_regression_edd27555,
  title = {metric space regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/metric-space-regression/rdr_edd27555e1214206beaa389199049f48},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_metric_space_regression_edd27555_v1,
  title = {metric space regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/metric-space-regression/rdr_edd27555e1214206beaa389199049f48/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake