Reach (Mathematics)¶

Measure the largest open Euclidean tube around a closed set in which every point has a unique nearest point on the set, exposing the first scale at which metric projection becomes ambiguous.

Version: v3 · 2026-09-06 · History
Domain-specific #: 2628
Origin domain: mathematics
Subdomain: geometric measure theory
Aliases: Reach of a set, Federer reach

Core Idea¶

The reach of a nonempty closed set \(A\subseteq\mathbb R^n\) is the supremal radius of an open offset neighborhood on which nearest-point projection onto \(A\) is single-valued. Write

\[ d_A(x)=\inf_{a\in A}\|x-a\|, \qquad \Pi_A(x)=\{a\in A:\|x-a\|=d_A(x)\}. \]

Then

\[ \operatorname{reach}(A) =\sup\{r\ge 0:|\Pi_A(x)|=1\text{ whenever }d_A(x)<r\}. \]

Closedness matters: in finite-dimensional Euclidean space it ensures that a nearest point exists, so the test isolates uniqueness rather than mixing uniqueness with nonattainment of the distance minimum. Federer introduced positive-reach sets as a class broad enough to include convex sets and compact closed sufficiently regular embedded manifolds, or manifolds with a positive global normal-tube radius, while still supporting curvature measures and tube formulas.^[1]

Equivalently, let the exterior medial axis be

\[ \operatorname{Med}(A)=\{x\in\mathbb R^n\setminus A:|\Pi_A(x)|>1\}. \]

With the convention \(\inf\varnothing=\infty\),

\[ \operatorname{reach}(A)=\inf_{x\in\operatorname{Med}(A)}d_A(x). \]

This second form reveals the abstraction's meaning: reach is the clearance from the set to the first possible ambiguity in metric projection. The number can be zero, a finite positive value, or infinity. Zero reach means ambiguous-projection points occur arbitrarily close to the set. Infinite reach means every ambient point has one nearest point. A finite reach is a supremum/infimum threshold; its value need not be attained by a particular medial-axis point, and uniqueness at distance exactly equal to the reach is not promised. The open inequality \(d_A(x)<r\) is therefore load-bearing.

The locked identity is:

a nonempty closed Euclidean set + its distance function and nearest-point correspondence + the open tubes on which that correspondence is single-valued -> the supremal admissible tube radius, with zero, finite, infinite, and nonattained thresholds kept distinct.

Structural Signature¶

Sig role-phrases:

the closed carrier set — the geometric subset whose projection regularity is being measured
the ambient Euclidean metric — the norm distance that determines offsets and nearest points
the distance-to-set function — \(d_A(x)\), converting ambient location into clearance from the set
the nearest-point correspondence — \(\Pi_A(x)\), potentially set-valued even though a minimum exists
the open offset tube — all points satisfying \(d_A(x)<r\)
the uniqueness predicate — \(|\Pi_A(x)|=1\) throughout the entire tube
the competing-footpoint locus — the exterior medial axis where uniqueness fails
the supremal radius — the largest threshold in the supremal sense, not necessarily a witnessed maximum
the endpoint convention — no assertion is made on the boundary \(d_A(x)=\operatorname{reach}(A)\)
the extended-value regimes — zero, finite positive, and infinite reach encode qualitatively different projection behavior
the geometric failure mechanism — local curvature/focal behavior or a global bottleneck can make projection ambiguous^[2]

Recognition test. A quantity instantiates Reach when it is defined from a set's ambient metric by asking for the supremal radius of an open neighborhood in which every point has exactly one nearest point on the set. It fails the test if it merely bounds the set's diameter, measures how far the set extends, records a generic curvature radius, or assumes a single projection without quantifying the tube on which uniqueness holds. A finite numerical estimate must also identify whether it is an upper bound, a lower bound, or the reach itself.

What It Is Not¶

Not the diameter, circumradius, or size of the set. An unbounded affine subspace has infinite reach, while a compact cusp can have zero reach.
Not measure in the measure-theoretic sense. Reach is a single extended nonnegative length scale, not an additive size assignment to subsets.
Not generic algebraic projection. The live Projection prime concerns idempotent reduction onto a chosen lower-dimensional target; reach concerns uniqueness of a metric nearest-point correspondence onto the set being studied.
Not convexity. Every nonempty closed convex set has infinite reach in finite-dimensional Euclidean space, but many nonconvex smooth sets have finite positive reach.
Not boundedness. Reach can be infinite for a bounded convex body or an unbounded affine subspace, and zero for a bounded set.
Not merely a curvature bound. Local curvature can limit reach, but globally separated parts of a manifold can form a narrower bottleneck.
Not the injectivity radius of a Riemannian manifold. Injectivity radius concerns the intrinsic exponential map; reach is extrinsic and depends on the chosen embedding and ambient metric.
Not a promise at the threshold. The definition controls \(d_A(x)<\operatorname{reach}(A)\), not every point at equality.
Not necessarily attained. The infimum distance to the medial axis may be approached without one medial-axis point realizing it.

Scope of Application¶

Reach belongs to geometric measure theory and extrinsic differential geometry. It applies to nonempty closed subsets of Euclidean space, including closed convex sets, sufficiently regular embedded manifolds, and more general positive-reach sets that may have boundary or nonsmooth but controlled features. Federer used the condition to extend curvature measures beyond smooth submanifolds.^[1]

In computational geometry and manifold inference, a positive lower bound on reach supplies a geometric conditioning scale. Sampling and noise must resolve distances smaller than that scale if a point cloud is to support reliable reconstruction of a manifold's topology. Niyogi, Smale, and Weinberger formulate sampling guarantees using the inverse reach as a condition number, while later statistical work treats the reach itself as a difficult quantity to estimate from data.^[3]^[2]

The cleanest definition here is Euclidean. Analogues exist in Hilbert, Riemannian, and other ambient spaces, but existence and behavior of nearest points depend on the ambient geometry. Those extensions should not be silently imported into this node. Nor should an empirical estimate from finitely many samples be confused with the underlying set's exact reach unless an estimation theorem and its regularity assumptions justify the inference.

Clarity¶

Reach clarifies four layers that informal geometric language often collapses.

Existence: does a nearest point on the set exist? Closedness in finite-dimensional Euclidean space handles this.
Uniqueness: is that nearest point the only one? This is what fails on the medial axis.
Uniform clearance: how far from the entire set does uniqueness hold simultaneously? This is the tube quantified by \(r\).
Threshold status: is the reported radius a supremum, and is it finite, zero, infinite, or attained?

The circle shows why the open-tube wording is necessary. For a circle of radius \(R\), points at distance less than \(R\) from the circle have unique nearest points, but the center is exactly distance \(R\) away and every point of the circle is nearest. Thus the reach is \(R\), although uniqueness fails at the threshold itself.

Reach also separates local and global explanations. High curvature can make neighboring normal directions intersect quickly, while distant parts of a folded manifold can approach one another and create competing nearest points even when each part bends gently. Aamari and colleagues organize estimation around this curvature-versus-bottleneck alternative.^[2]

Manages Complexity¶

The geometry of an arbitrary embedded set can involve infinitely many points, directions, curvature values, and self-approach relations. Reach compresses that field into one operational scale with a precise guarantee: below it, nearest-point assignment is globally unambiguous. Once a positive lower bound \(\tau\) is established, arguments can work inside a certified tube rather than re-proving projection uniqueness point by point.

That compression is useful but lossy. The same reach may arise from a sharp local bend, a thin global bottleneck, or a limiting sequence of near-ambiguities. Algorithms that need diagnosis rather than mere conditioning must retain more information, such as local feature size, curvature profiles, or the medial-axis geometry. The scalar manages complexity by answering “how far is safe everywhere?”; it does not answer “where and why does safety first fail?”

Abstract Reasoning¶

Reach supports reasoning through the duality between a universal tube property and an obstruction set. To prove \(\operatorname{reach}(A)\ge r\), establish that every point with \(d_A(x)<r\) has a unique footpoint. To prove \(\operatorname{reach}(A)\le r\), exhibit ambiguous-projection points at distance at most \(r\), or a sequence whose distances decrease to \(r\). The sequential route is essential when the infimum is not attained.

The invariant obeys clean Euclidean symmetries. Orthogonal transformations and translations preserve reach, and for \(\lambda>0\),

\[ \operatorname{reach}(\lambda A)=\lambda\operatorname{reach}(A). \]

These laws make it a genuine geometric length scale. They also prevent a common category error: reach is not “scale invariant” as a numerical value. It is scale-equivariant; dilation multiplies it.

Counterfactual reasoning becomes concrete. If a sampled manifold is rescaled by two, every absolute noise and sampling threshold based on reach must rescale by two. If two distant sheets move closer while their local curvature remains fixed, a global bottleneck may lower the reach. If a corner is rounded, reach can move from zero to positive even if the overall diameter changes negligibly.

Knowledge Transfer¶

The transferable core is a radius of unambiguous decoding: objects near a structured set can be assigned to one source point until competing explanations become possible. This pattern helps interpret robust reconstruction, tubular coordinates, denoising to a manifold, and condition-number arguments. It also suggests a disciplined question in adjacent settings: what is the largest perturbation neighborhood in which the inverse assignment remains unique?

The transfer has a firm boundary. Euclidean nearest-point geometry, closed-set existence, and the medial-axis obstruction are not decorative terminology. A classification margin, error-correcting radius, or optimization basin may resemble reach, but it is not this abstraction unless its carrier, distance, nearest-point correspondence, and uniform open-tube test literally match.

Examples¶

Mapped example 1 — a circle. Let \(A\) be the circle of radius \(R>0\) centered at the origin in \(\mathbb R^2\).

carrier: the circumference, not the filled disk
metric: ordinary Euclidean distance
nearest-point correspondence: radial projection is unique away from the origin
first ambiguity: the origin has every circle point as a nearest point
clearance: \(d_A(0)=R\)
result: \(\operatorname{reach}(A)=R\)
endpoint diagnostic: the open tube of every radius at most \(R\) is unambiguous, while equality at the center fails.

Mapped example 2 — a closed convex set. Let \(A\) be a nonempty closed convex subset of \(\mathbb R^n\), such as a filled square, a line, or a single point.

carrier: a closed convex set
metric: Euclidean distance
projection: the Hilbert-space projection theorem gives a unique nearest point for every ambient point
medial axis: empty
convention: \(\inf\varnothing=\infty\)
result: \(\operatorname{reach}(A)=\infty\)
diagnostic: corners do not force zero reach when they belong to a filled convex set; an inward re-entrant corner or a V-shaped one-dimensional graph behaves differently.

Worked intervention — auditing a manifold-reconstruction claim. A paper claims topology recovery from a noisy point cloud under “feature size \(0.05\).” Ask whether \(0.05\) is a proved lower bound on reach, a point estimate, or a tuning parameter. Check the ambient units, whether the underlying set is closed, what regularity class is assumed, and whether noise and sampling radii satisfy the cited theorem's strict inequalities relative to the reach bound. Then test both failure modes: local curvature and global near-self-intersection. A dense sample cannot rescue a guarantee if the geometric lower bound was inferred from local curvature while a narrower unseen bottleneck controls the true reach.^[3]^[4]

Structural Tensions¶

Local curvature ↔ global bottleneck. Either mechanism can set the same scalar threshold. Diagnostic: search both for a high-curvature region and for distinct sheets sharing nearly equidistant ambient points.
Open guarantee ↔ boundary failure. Reach certifies all distances below a threshold without certifying equality. Diagnostic: on a circle, test the center at distance exactly \(R\).
Supremum ↔ attained maximum. A finite reach need not have a single realizing witness. Diagnostic: determine whether a sequence of medial-axis distances converges to the infimum without attaining it.
Zero reach ↔ useful geometry. A set can be familiar and structured yet have no uniform positive projection tube. Diagnostic: the graph of \(|x|\) admits ambiguous projections arbitrarily close to its corner.
Infinite reach ↔ finite extent. The extended value describes projection uniqueness, not spatial unboundedness. Diagnostic: both a single point and an unbounded line have infinite reach.
Global scalar ↔ local feature variation. One narrow defect controls the reach of an otherwise well-separated set. Diagnostic: compare the global reach with pointwise distance to the medial axis.
Geometric truth ↔ finite-sample estimate. Samples hide unsampled curvature and bottlenecks. Diagnostic: state the model class and confidence guarantee before treating an estimator as a lower bound.^[2]^[4]
Autonomous abstraction ↔ reduction to Metric plus Projection. Reach depends on distance and nearest points but adds a uniform maximal uniqueness radius and extended-value conventions not supplied by either neighbor. Diagnostic: ask whether the proposed reduction can distinguish a circle's reach \(R\), a V-graph's reach zero, and a convex body's reach infinity.

Structural–Framed Character¶

Reach qualifies as a structural-framed domain abstraction under five criteria:

Stable roles. Carrier set, ambient metric, nearest-point correspondence, open tube, uniqueness predicate, obstruction locus, and supremal radius recur unchanged.
Relational compression. One extended nonnegative number summarizes when a many-to-one nearest-point relation first appears.
Counterfactual leverage. Rescaling, bending, rounding, or narrowing a bottleneck predicts how the safe projection tube changes.
Operational diagnostics. Lower bounds come from uniform uniqueness; upper bounds come from ambiguous points or convergent obstruction sequences.
Boundary discipline. The open inequality, closedness assumption, extended-value conventions, and possible nonattainment sharply separate valid from invalid uses.

Its character: a geometric conditioning scale defined by the maximal uniform domain of unique nearest-point projection.

Structural Core vs. Domain Accent¶

Structural core. A structured set admits an unambiguous inverse assignment throughout a neighborhood until competing assignments appear; the smallest obstruction distance determines a robustness radius.

Domain accent. In Reach, the set is a nonempty closed subset of Euclidean space, assignment means metric nearest-point projection, the obstruction locus is the exterior medial axis, and the robustness radius is an extended-valued geometric length.

Substitution test. Replacing Euclidean points with codewords and nearest projection with decoder output preserves an analogy but changes theorems, existence conditions, and obstruction geometry. Replacing a smooth manifold with another closed Euclidean set preserves the identity. Thus the structural skeleton travels, but the node remains domain-specific.

Metric — prospective strict parent (composition / presupposes). Reach literally requires ambient Euclidean distance, distance-to-set, and metric balls. The Metric prime supplies those notions but not the maximal unique-projection threshold.
Convexity — strong related sufficient condition. Nonempty closed convex sets in finite-dimensional Euclidean space have infinite reach, but positive reach neither requires convexity nor reduces to it.
Boundedness — related contrast only. A finite threshold sounds like a bound, yet reach may be infinite and does not bound the spatial extent of the carrier.
Projection — lexical and mathematical neighbor, not proposed parent. Nearest-point assignment is often called metric projection, but the live prime fixes Projection to idempotent reduction onto a chosen lower-dimensional target. It does not literally cover arbitrary-set nearest-point correspondence.
Measure — rejected false friend. Federer developed reach within geometric measure theory, but reach is not an additive measure.
Scale Invariance — rejected. Reach scales linearly under dilation rather than remaining numerically invariant.

No structured DAG edge is encoded in this isolated draft. The prospective edge is documented for independent review in CATALOG_MATCH_AND_DAG_PLACEMENT.md.

Relationships to Other Abstractions¶

Current abstraction Reach (Mathematics) Domain-specific

Parents (1) — more general patterns this builds on

Reach (Mathematics) presupposes Metric Prime

Metric — prospective strict parent (composition / presupposes). Reach literally requires ambient Euclidean distance, distance-to-set, and metric balls.

Hierarchy path (1) — routes to 1 parentless root

Reach (Mathematics) → Metric → Function (Mapping)

Neighborhood in Abstraction Space¶

Reach (Mathematics) sits in a sparse region of the domain-specific corpus (67^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Diameter: maximum pairwise separation within a set. Tell: it measures extent, not projection uniqueness outside the set.
Circumradius: radius of a containing sphere. Tell: it depends on enclosure rather than the medial axis.
Injectivity radius: intrinsic radius for which a manifold's exponential map is well behaved. Tell: reach changes with extrinsic embedding even when intrinsic geometry is preserved.
Radius of curvature: reciprocal local curvature scale. Tell: a global bottleneck can make reach smaller than every local curvature radius.
Local feature size: often the distance from a point on a shape to its medial axis. Tell: reach is the global infimum of the relevant feature scale.
Medial axis: locus of competing nearest points. Tell: it is the obstruction set; reach is its infimal clearance from the carrier.
Hausdorff distance: distance between two sets. Tell: reach is a property of one embedded set's projection tube.
Convexity: closure under line segments. Tell: it implies infinite reach for closed sets but is not equivalent to positive reach.
Metric projection: the nearest-point map or correspondence itself. Tell: reach quantifies the largest neighborhood where that correspondence is single-valued.
Measure: an additive set function. Tell: reach has units of length and no additivity axiom.
Reachability: which states a system can enter. Tell: mathematical reach here concerns nearest-point uniqueness, not state transitions.
The image or range of a function: values a map attains. Tell: that lexical sense of “reach” contains no offset radius.

References¶

[1] Federer, Herbert. “Curvature Measures.” Transactions of the American Mathematical Society 93, no. 3 (1959): 418–491. https://doi.org/10.1090/S0002-9947-1959-0110078-1. registry ↩a ↩b

[2] Aamari, Eddie, Jisu Kim, Frédéric Chazal, Bertrand Michel, Alessandro Rinaldo, and Larry Wasserman. “Estimating the Reach of a Manifold.” Electronic Journal of Statistics 13, no. 1 (2019): 1359–1399. https://doi.org/10.1214/19-EJS1551. registry ↩a ↩b ↩c ↩d

[3] Niyogi, Partha, Stephen Smale, and Shmuel Weinberger. “Finding the Homology of Submanifolds with High Confidence from Random Samples.” Discrete & Computational Geometry 39 (2008): 419–441. https://doi.org/10.1007/s00454-006-1250-7. registry ↩a ↩b

[4] Berenfeld, Clément, John Harvey, Marc Hoffmann, and Krishnan Shankar. “Estimating the Reach of a Manifold via Its Convexity Defect Function.” Discrete & Computational Geometry 67 (2022): 403–438. https://doi.org/10.1007/s00454-021-00290-8. registry ↩a ↩b