Understanding REACH in Geometry, Control & More

• REACH is a multifaceted term that defines geometric regularity via unique nearest-point projections and intrinsic manifold properties.
• In control systems and graph algorithms, reachability identifies attainable states and safe trajectories, underpinning robust verification and synthesis.
• Acronymic REACH in cosmology, vehicular communications, and computer vision represents domain-specific projects with tailored methodologies and performance benchmarks.

In current research literature, “reach” and “REACH” denote several distinct concepts. In geometry and manifold learning, reach is the largest radius for which nearby points admit a unique nearest-point projection onto a set (Cholaquidis et al., 2021). In control, verification, and graph algorithms, closely related terminology refers to reachability, reach sets, or reach-avoid sets (Hanauer et al., 2020). In other domains, REACH is an acronym, notably the Radio Experiment for the Analysis of Cosmic Hydrogen in global 21-cm cosmology (Acedo et al., 2022), Relevance-based Explanation and Architectural Compression for cHannel estimators in vehicular communications (Ngorima et al., 10 Jun 2026), and the Room-Environment dataset Annotated with Chest cameras for Hand pose estimation in computer vision (Nakamura et al., 21 May 2026).

1. Terminological scope

The term spans several mature research traditions. In some of them it names a geometric regularity parameter; in others it denotes algorithmic reachability or serves as a project acronym.

Research area	Meaning of “reach” / “REACH”	Representative source
Geometric measure theory, manifold learning	Unique-projection radius and regularity parameter of a set or manifold	(Cholaquidis et al., 2021)
Autoencoders	Local certificate for whether a latent code corresponds to a unique nearest-point projection onto the decoder manifold	(Hauschultz et al., 2022)
Graph algorithms	Static reachability indexing in directed graphs, as in O’Reach	(Hanauer et al., 2020)
Control and formal methods	Reach sets, reachset conformance, and reach-avoid sets	(Haddad et al., 2021)
Radio cosmology	Radio Experiment for the Analysis of Cosmic Hydrogen	(Acedo et al., 2022)
Communications and computer vision	Acronyms for an interpretability/compression framework and a room-scale hand-pose dataset/model	(Ngorima et al., 10 Jun 2026)

A useful distinction follows from this taxonomy. Geometric reach concerns uniqueness of projection and regularity. Reachability concerns existence of paths, trajectories, or controllers. Acronymic uses of REACH are domain-specific and formally unrelated, even when the names overlap.

2. Geometric reach in analysis and manifold inference

For a compact set or manifold $M \subset \mathbb{R}^D$, reach is the largest radius up to which the nearest-point projection remains unique. One formal definition is

$$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$

where $\operatorname{Unp}(M)$ is the set of points with a unique closest point on $M$ (Cholaquidis et al., 2021). Equivalent viewpoints used in recent work include distance to the medial axis, chord-versus-geodesic distortion, and tangent-space or curvature characterizations (Lieutier et al., 2024).

This geometric parameter controls both local curvature and global self-avoidance. The literature repeatedly emphasizes that the reach may be limited either by a local curvature obstruction or by a global bottleneck. In Federer-style formulas, it can be written through tangent-space deviation; in metric form, it is the largest $r$ such that

$d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$

for all sufficiently short chords $|a-b|<2r$ (Lieutier et al., 2024). This connects reach directly to intrinsic-versus-extrinsic metric distortion, which is why it appears in manifold reconstruction, topological inference, and metric learning (Aamari et al., 2022).

Positive reach has strong regularity consequences. A closed embedded topological manifold has positive reach if and only if it can locally be written as the graph of a $C^{1,1}$ function from the tangent space to the normal space (Lieutier et al., 2024). The same work gives an optimal tangent-variation bound: the Lipschitz constant of the tangent-space map is exactly the reciprocal of the local reach. It also states that the reach decomposes as

$rch(M)=\min(rch_{glob.}(M),\,rch_{loc.}(M)),$

so the attained value is either a bottleneck-type global obstruction or a local one (Lieutier et al., 2024).

Statistical work on the topic has developed several estimation regimes. “Estimating the Reach of a Manifold” introduced a plug-in estimator based on Federer’s formula in an oracle setting with known tangent spaces and showed that reach estimation differs in difficulty depending on whether the true obstruction is a bottleneck or local curvature (Aamari et al., 2017). “Universally consistent estimation of the reach” proposed a graph-based estimator that is consistent under the minimal assumption of positive reach and also proved that one cannot determine from a finite sample whether the reach of a support is zero or not (Cholaquidis et al., 2021). “Optimal Reach Estimation and Metric Learning” then derived optimal nonasymptotic bounds, with adaptive rates depending on whether the reach is curvature-dominated or bottleneck-dominated, and coupled this with optimal geodesic metric estimation (Aamari et al., 2022).

In highly structured algebraic settings, the reach can even be computed exactly. For the Segre–Veronese variety intersected with the unit sphere, the reach is

$\tau= \begin{cases} \frac{\pi}{4}, & d\le 5,\[0.3em] \sqrt{\frac{d}{2(d-1)}}, & d>5, \end{cases}$

where $$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$0 is the total degree; notably, the result depends on the total degree and not on the individual dimensions $$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$1 (Breiding et al., 2023).

3. Reach in autoencoder geometry

In representation learning, reach has been used to formalize when an autoencoder’s latent code corresponds to a trustworthy geometric projection. For an encoder $$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$2 and decoder $$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$3, the decoder spans the manifold

$$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$4

For a fixed decoder, the mathematically optimal encoder is the nearest-point projection onto $$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$5,

$$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$6

but this projection need not be unique (Hauschultz et al., 2022).

The key issue is that an encoder network always outputs a single latent vector, whereas the true nearest-point projection can have infinitely many values. The relevant geometric object is therefore

$$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$7

the region in which a unique latent representation can be trusted (Hauschultz et al., 2022).

The same paper uses Federer’s reach and then introduces a local refinement, the pointwise normal reach. With

$$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$8

it proves that

$$\operatorname{reach}(M,x)=\sup\{r>0:\mathring{\mathcal B}(x,r)\subset \operatorname{Unp}(M)\}, \qquad \operatorname{reach}(M)=\inf_{x\in M}\operatorname{reach}(M,x),$$9

This makes $\operatorname{Unp}(M)$0 a local uniqueness certificate rather than a global worst-case certificate (Hauschultz et al., 2022).

A practical decision rule follows. If a reconstruction $\operatorname{Unp}(M)$1 of an observation $\operatorname{Unp}(M)$2 satisfies

$\operatorname{Unp}(M)$3

then the optimal encoder exists in the sense of a unique latent representation. Using the bound $\operatorname{Unp}(M)$4, the paper proposes the conservative criterion

$\operatorname{Unp}(M)$5

Thus reconstruction quality alone does not certify uniqueness of the code (Hauschultz et al., 2022).

To make the quantity computable, the paper replaces the infimum over the manifold by a minimum over a finite sample $\operatorname{Unp}(M)$6: $\operatorname{Unp}(M)$7 where $\operatorname{Unp}(M)$8 is the decoder Jacobian at $\operatorname{Unp}(M)$9. This estimator is an upper bound, $M$0, and is differentiable, which enables its use as a regularizer (Hauschultz et al., 2022).

The proposed penalty is

$M$1

leading to the full objective

$M$2

Empirically, the paper reports that many trained autoencoders produce latent representations that are not guaranteed to be unique; on CelebA, almost all validation samples lie outside the pointwise normal reach, while on MNIST the regularizer increases the fraction of points inside reach with only a slight reconstruction penalty (Hauschultz et al., 2022).

4. Reachability, reach sets, and reach-avoid

In graph algorithms, reachability is the problem of deciding whether a directed path exists from $M$3 to $M$4. O’Reach addresses the static version by combining enhanced topological-order-based pruning with a static version of supportive vertices. It computes $M$5 extended topological orderings, stores high/max or low/min indices, and supplements them with full in- and out-reachability summaries for $M$6 supportive vertices. The resulting index is linear in space, answers a large fraction of queries in constant time, and can also serve as a preprocessor for earlier methods such as PReaCH, PPL, IP, or BFL (Hanauer et al., 2020).

In continuous and hybrid dynamical systems, reach sets denote attainable-state sets under dynamics and uncertainty. For integrator chains in Brunovský normal form with box-valued inputs, the forward reach set factorizes across blocks and admits an exact support function, exact boundary parameterizations, and implicit algebraic boundary equations given by vanishing Hankel determinants. These sets are shown to be semialgebraic translated zonoids, which makes them benchmark objects for reachability algorithms (Haddad et al., 2021). For large affine systems, decomposition methods have been proposed that perform set operations in low-dimensional blocks while retaining full-dimensional matrix operations; the dense-time case was demonstrated with more than $M$7 variables (Bogomolov et al., 2018).

The notion of reachset conformance plays a different role in formal robotics. A model is reachset conformant when all measured outputs of the implementation are enclosed by the reachable output sets of the abstract model; this is presented as necessary and sufficient for transferring safety properties. For robotic contact tasks, recent work extends this idea from continuous systems to hybrid automata with linear dynamics, nondeterministic flows, measurement error, and uncertain resets, with zonotopes used for $M$8, $M$9, and transition uncertainty $r$0 (Tang et al., 2024). In a related but older control tradition, reach control decomposes a safe polytope into simplices and synthesizes piecewise affine feedback laws that prevent trajectories from crossing restricted facets while forcing exit through designated facets; a quadrocopter experiment showed safe and robust maneuvers without a predefined open-loop reference trajectory (Vukosavljev et al., 2016).

A closely related concept is the reach-avoid set, the set of states from which a controller can reach a target while remaining in a safe set beforehand. For deterministic systems in unknown environments, the controlled reach-avoid set

$r$1

acts as a barrier separating feasible from infeasible initial states, motivating inner-approximation and online controller synthesis methods (Ding et al., 2023). In model predictive control, reach-avoid analysis has been used to replace discrete terminal samples by a continuous reach-avoid set $r$2 derived from a guidance-barrier function, thereby removing integer variables from the online problem (Ren et al., 2023). For Bayesian neural network dynamics, probabilistic reach-avoid verification and synthesis have been built around backward recursion, interval propagation, and conservative lower bounds on the probability of satisfying a finite-horizon reach-avoid specification (Wicker et al., 2023).

Symbolic verification uses the same language at a different abstraction level. A decision-diagram operation named REACH computes the reflexive-transitive closure $r$3 directly, without requiring the transition relation to be partitioned as in saturation. Sequential and parallel versions were given for both BDDs and MDDs, and the method often outperforms saturation on weakly local transition relations (Brand et al., 2022). In flight safety, the backward reachable tube plays the central role: the unsafe helicopter operating region after engine failure is defined as the complement of the BRT, and a grid-free trajectory-optimization reformulation was used to compute both safe operating envelopes and autonomous autorotation landings for a nonlinear helicopter model (Kirchner et al., 2020).

5. REACH in global 21-cm cosmology

In radio astronomy, REACH is the Radio Experiment for the Analysis of Cosmic Hydrogen, a ground-based global 21-cm radiometer experiment designed to detect the sky-averaged neutral-hydrogen signal from the Cosmic Dawn and the Epoch of Reionization (Acedo et al., 2022). It targets the band 50–170 MHz, corresponding approximately to redshift $r$4 to $r$5, and is explicitly motivated by the need to separate a very faint cosmological signal from much brighter foregrounds and instrumental spectral structure (Acedo et al., 2022).

The experiment’s design emphasizes joint modeling rather than simple foreground subtraction. REACH Phase I uses two independent radiometers observing simultaneously in the same sky region but with different antennas: a hexagonal dipole covering 50–130 MHz and a conical log-spiral antenna covering 50–170 MHz. The instrument also includes an ultra-wideband receiver, a digital backend based on the SKA1-LOW TPM FPGA board, high-resolution 6 kHz channels for RFI excision, and an in-field calibrator using a compact VNA and the noise-waves formalism (Acedo et al., 2022).

Methodologically, the experiment is built around Bayesian joint inference of the cosmological signal, the beam-convolved foreground sky, and instrumental systematics. The measured spectrum is modeled through the sky decomposition

$r$6

and the corresponding observation

$r$7

Nested sampling with PolyChord is used because the parameter space is degenerate and multimodal (Acedo et al., 2022).

Subsequent work has used simulated REACH observations to assess foreground recovery and component-separated mapping across the same 50–170 MHz band. In that framework the sky is divided into $r$8 regions, the anchor map is the 2008 Global Sky Model at 230 MHz, and the data model is

$r$9

The injected global 21-cm signal is a Gaussian absorption feature with $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$0 MHz, $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$1 MHz, and $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$2 K (Robins et al., 1 Jul 2026).

That study tests four foreground families of increasing complexity: a pure synchrotron power law, a variable-amplitude power law, a curved power law, and a synchrotron-plus-free-free model with a fixed free-free spectral index of 2.1. Its central conclusion is that the model complexity must match the true sky complexity: simple models are best for simple skies, whereas richer skies require curvature, amplitude freedom, or separate synchrotron and free-free components. The trade-off is stronger parameter degeneracy, especially $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$3–$d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$4 anti-correlation in the curved model and $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$5–$d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$6 trade-offs in the component model (Robins et al., 1 Jul 2026).

The same simulations indicate that REACH can function as both a cosmology experiment and a low-frequency Galactic mapping experiment. Synchrotron is recovered well across the sky, while free-free recovery is limited because it is subdominant, has lower signal-to-noise, and is not well aligned with the original percentile-based region split. Modified regioning strategies, especially a mixed split using synchrotron-informed regions for synchrotron parameters and free-free-informed regions for free-free parameters, yield lower foreground RMSE, improved 21-cm recovery, and better component separation (Robins et al., 1 Jul 2026).

6. REACH as an acronym in communications and computer vision

In vehicular communications, REACH stands for Relevance-based Explanation and Architectural Compression for cHannel estimators. It is a gradient-based interpretability framework for IEEE 802.11p multi-channel vehicular channel estimation, designed to explain why mixed-SNR multi-channel training improves out-of-distribution generalization and to use that explanation for architecture compression (Ngorima et al., 10 Jun 2026). Applied to the DPA-RDCNN estimator, REACH performs attribution at two levels. At the input level, it identifies 1,020 of 5,200 features, or 19.6% of the input, as consistently relevant across all six channel models. At the filter level, it classifies the $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$7 filters into 23.8% universal, 58.0% environment-specific, and 18.1% redundant, thereby motivating width reduction rather than block pruning (Ngorima et al., 10 Jun 2026).

The compression results are quantitative. With width $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$8, the compact model achieves 60.2% parameter reduction, 60.4% FLOP reduction, and $d_M(a,b)\le 2r\arcsin\!\frac{|a-b|}{2r}$9 dB NMSE degradation; the paper also reports that OOD generalization degrades more slowly than within-distribution performance under increasing compression (Ngorima et al., 10 Jun 2026). A plausible implication is that the framework’s main scientific value is not only compression but a representational account of why a single shared internal filter structure generalizes across channel conditions.

In computer vision, REACH means Room-Environment dataset Annotated with Chest cameras for Hand pose estimation and also names the associated model REACH-Net (Nakamura et al., 21 May 2026). The dataset targets room-scale, far-view hand-pose estimation from fixed cameras at room corners, where hands are often about 20–25 pixels wide and heavily occluded. It contains 50 participants, 1M frames from fixed cameras, and 540K frames of close hand imagery from concealed chest cameras. Annotation is produced with two GoPro 12 fisheye cameras mounted on a chest strap under a jacket, together with ceiling ArUco markers, MediaPipe, HaMeR, and triangulated wrist keypoints (Nakamura et al., 21 May 2026).

REACH-Net is a Transformer-based multiview autoregressive architecture built to exploit hand-body coordination, multiview geometry, and temporal continuity. The published implementation uses a 6-layer Pose Encoder with 8 attention heads and hidden dimension 256, and a 6-layer encoder-decoder REACH-Net Transformer with 4 attention heads and hidden dimension 1024. Training uses AdamW, learning rate $|a-b|<2r$0, weight decay $|a-b|<2r$1, and a two-stage schedule of 60k plus 60k iterations (Nakamura et al., 21 May 2026).

On the benchmark defined by the paper, REACH-Net outperforms the compared baselines. It achieves PA-MPJPE = 14.47 mm and joint angle error = 14.48°, compared with 14.88 mm and 16.35° for ArcticNet-LSTM and 16.01 mm and 24.31° for HaMeR. It is also best across the Near, Medium, and Distant evaluation bins, with 12.96, 13.44, and 14.86 mm PA-MPJPE, respectively (Nakamura et al., 21 May 2026).

Across these literatures, the shared label conceals very different formal objects. In geometry, reach is a quantitative certificate of unique projection and regularity. In algorithms and control, reachability concerns paths, reachable states, or safe target attainment. In astronomy, communications, and computer vision, REACH names specific instruments, frameworks, or datasets whose technical meanings are fixed by their domains rather than by a common mathematical definition.