Multiple Tangent Spaces: Theory & Applications

Updated 28 February 2026

Multiple Tangent Spaces are defined as the use of several distinct tangent spaces to locally approximate non-linear, singular, or non-Euclidean structures.
They are applied in manifold learning and geometric data analysis to robustly capture local geometry for function approximation and classification.
This framework bridges theoretical advancements in diffeological spaces with practical methods in high-dimensional and noisy data environments.

A multiple tangent space framework describes the association or utilization of several distinct tangent spaces rather than a single one, typically in non-Euclidean, non-linear, or singular geometric contexts. This concept is central both in the development of manifold learning and geometric data analysis methods that estimate local approximation spaces at many points, and in the abstract theory of diffeological and singular spaces where the very notion of tangent space may bifurcate into different, non-equivalent constructions. Multiple tangent spaces are exploited for function approximation on manifolds, robust local geometric estimation, and, more abstractly, for distinguishing non-manifold behavior in generalized spaces.

1. Tangent Spaces: Classical and Generalized Constructions

In classical differential geometry, the tangent space $T_pM$ at $p$ on a smooth $n$ -manifold $M$ is uniquely defined and provides a linearization of $M$ near $p$ . This classical picture extends to smooth submanifolds embedded in $\mathbb{R}^D$ , where tangent planes provide the best linear approximation of the manifold at each point.

In generalized settings (e.g., metric, diffeological, or singular spaces), the meaning of tangent space can diverge drastically. Several functorial constructions have been developed:

Internal tangent space: defined via colimits of velocity vectors along smooth curves or plots, yielding $T_x^{\mathrm{int}}(X)$ [(Christensen et al., 2014); (Taho, 2024); (Taho, 22 Nov 2025)].
External tangent space: as the space of smooth derivations on the algebra of germs of smooth functions at $x$ , denoted $T_x^{\mathrm{ext}}(X)$ [(Christensen et al., 2014); (Taho, 2024); (Taho, 22 Nov 2025)].
Right-Kan extension tangent space: constructed via the right Kan extension of the tangent functor, often nearly isomorphic to (or coinciding with) the external tangent space under regularity conditions (Taho, 2024).

On smooth manifolds, all these notions are canonically isomorphic. In the broader category of diffeological spaces, however, they can differ, and one may even construct infinitely many distinct tangent functors, indexed by test objects such as irrational tori or orbit spaces, that agree only on smooth cases (Taho, 22 Nov 2025).

2. Multiple Tangent Spaces in Manifold-Valued Learning and Data Analysis

In practical geometric data analysis and machine learning, especially for manifold-valued data, one estimates or employs a field of tangent spaces, i.e., a collection $p$ 0 at data locations $p$ 1. Diffusion geometry methods, for instance, estimate a local Gram matrix at each $p$ 2 to robustly recover multiple tangent spaces from data, each capturing the manifold's local linear structure (Jones, 2024). Such approaches are essential for high-dimensional, noisy, or non-uniformly sampled datasets, and allow practitioners to exploit the geometry in downstream analysis, including dimension estimation, curvature analysis, and function approximation.

The simultaneous estimation and utilization of tangent spaces across many points enables parameter-free, noise-robust algorithms, a significant advancement over methods using a single global or local linearization.

3. Multiple Tangent Spaces in Manifold-Valued Function Approximation

The Multiple Tangent Space Model (MTSM) is foundational for manifold-valued function approximation where the output lies nonlinearly on a Riemannian manifold $p$ 3 (Wang et al., 17 Apr 2025). Rather than restricting to a single anchor tangent space or solving a global Riemannian Fréchet mean (with computational cost scaling linearly with sample size), MTSM selects $p$ 4 anchor points $p$ 5. Each anchor defines its own tangent space $p$ 6, where local regression or interpolation is carried out. The local predictions are then aggregated via a weighted Fréchet mean over $p$ 7 elements, allowing significant flexibility and computational speed.

This multiple-tangent approach, in contrast to single-tangent or full weighted Fréchet regression, achieves:

Stronger approximation accuracy, especially on manifolds with significant curvature,
Online evaluation cost scaling as $p$ 8 (much less than $p$ 9 for global methods),
Robustness to data spread and distribution.

Selection of the anchor points is based on manifold clustering (e.g., Riemannian $n$ 0-means), and weights for the Fréchet mean aggregation are constructed as smooth partitions of unity, typically subordinate to geodesic balls around each anchor.

4. Multiple Tangent Spaces in Discriminative Models and Representations

Discriminative models, such as the K-Tangent Spaces (KTS) model, deploy multiple tangent spaces for classification tasks on manifolds, notably symmetric positive-definite matrix spaces arising in image analysis (Sanin et al., 2014). In KTS, $n$ 1 poles (anchor points) are selected on the manifold, each with its own tangent space. Feature descriptors are mapped via the logarithm map to each tangent space, and a separate discriminant is learned in each. The multiple-embedding structure allows for accurate local linearization and aggregation, better capturing manifold geometry and improving downstream classification performance, as empirically demonstrated in pedestrian detection benchmarks.

This paradigm acknowledges the inadequacy of a single linearization in high-curvature or large-support settings and efficiently interpolates local linear models to respect global geometry.

5. Infinitely Many Tangent Functors and the Abstract Theory

A central insight from diffeological geometry is that, outside the category of smooth manifolds, there is no canonical tangent functor (Taho, 22 Nov 2025). By choosing arbitrary test objects (e.g., irrational tori, orbit spaces), one may define associated (internal or external) tangent functors—essentially different ways to "probe" the infinitesimal structure of a space. These multiple tangent functors can give rise to entirely distinct tangent spaces at points, leading to inequivalent geometric and analytic theories. The uniqueness that underpins much of classical differential geometry fails, making the choice of tangent functor a fundamental part of the "structure data" of a space.

This proliferation of tangent functors has practical as well as theoretical implications: in designing analytic or algebraic invariants for singular, stratified, or non-manifold spaces, the tangent functor selected may detect or fail to detect features of the space, depending on the test objects encoded in its construction.

6. Multiple Tangent Spaces in Fractals and Irregular Geometry

In geometric measure theory, especially in the context of Hölder and non-Lipschitz curves, the set of tangent spaces or tangent sets at a typical point can be extremely rich. While Rademacher's theorem guarantees that almost every point on a Lipschitz curve has a unique (line) tangent space, recent research shows that, for Hölder curves, one can construct examples where almost every point admits infinitely many pairwise non-homeomorphic tangent sets (Shaw et al., 2024). These "exotic tangent spaces" do not arise, even up to bi-Lipschitz equivalence, as tangents of any connected self-similar set at a typical point. This starkly contrasts the classical intuition and motivates the use of multiple tangent constructs and new invariants for describing such geometries.

7. Significance and Broader Context

The use of multiple tangent spaces reflects foundational advances at several levels:

In geometric data analysis and machine learning, multiple tangent fields provide scalable, stable, and accurate local geometric models.
In function approximation, MTSM leverages the geometric structure for computational efficiency and superior performance.
In abstract geometry and analysis, the non-uniqueness and richness of tangent functors in the diffeological/singular context require the careful selection of constructions suited to the analytic or topological properties desired.

This multiplicity is both a necessity—driven by curvature, nonlinearity, and singularity—and an opportunity, opening new avenues for theoretical classification and algorithmic development.

Key References Table

Setting	Role of Multiple Tangent Spaces	Main Source
Manifolds (data analysis)	Simultaneous estimation $n$ 2 (PCA/diffusion)	(Jones, 2024)
Manifold approximation/learning	MTSM: anchor-based regression and Fréchet mean aggregation	(Wang et al., 17 Apr 2025)
Riemannian discrimination/classification	K-Tangent model: K poles, Linear models per tangent	(Sanin et al., 2014)
Diffeological/singular spaces (abstract)	Infinitely many distinct functors; non-uniqueness	(Taho, 2024, Taho, 22 Nov 2025)
Fractals/Hölder curves	Infinitely many local tangents (wild geometric behavior)	(Shaw et al., 2024)