REVIEW 1 major objections 4 minor 8 references
When the Gromov-Hausdorff distance between finite-dimensional space and its subset is finite?
T0 review · 1 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read The paper proves that a subset A of R^n is at finite Gromov–Hausdorff distance from R^n exactly when it is an ε-net, and that this equivalence fails in infinite-dimensional Euclidean space.
desk verdict The epsilon-net characterization of finite GH distance to R^n is new and mostly proved; one local gap in the final contradiction needs fixing but looks fixable. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The proof is carried by an external inequality from reference [6], used as Corollary 1: for bounded subsets $X, Y \subset \mathbb{R}^n$, the Euclidean Gromov–Hausdorff distance — the minimal Hausdorff distance between $X$ and an isometric image of $Y$ under the isometry group of $\mathbb{R}^n$ — is at most $c'_n (\max\{\mathrm{diam}\, X, \mathrm{diam}\, Y\})^{1/2} d_{\mathrm{GH}}(X,Y)^{1/2}$. This inequality turns a finite-distortion correspondence into an approximate Euclidean isometry. Given a correspondence with distortion $c < \infty$, the proof fixes a point $p$, takes a large sphere $S_N(p)$ together with its center, and uses the inequality to make the image of $A$ nearly coincide with that sphere in Hausdorff distance. The distortion bound then forces every cone of fixed angle with vertex $p$ to meet $A$ within an annulus of radii $N-c$ and $N+c$. If the complement of $A$ contained a large ball, a suitably placed cone would lie inside that ball, forcing a point of $A$ inside the empty ball — a contradiction. The cone and sphere geometry is where the Euclidean isometry group enters, which is why the theorem does not follow for arbitrary finite-dimensional normed spaces.
What would settle it
A concrete test: let $A_k$ be $\mathbb{R}^n$ with $k$ disjoint closed balls of radius $k$ removed, placed far apart, and compute $d_{\mathrm{GH}}(\mathbb{R}^n, A_k)$; the theorem forces these distances to grow without bound as $k \to \infty$, so bounded values would refute it.
Extended reading notes
Core claim
The central discovery is Theorem 3.1: for a subset $A \subset \mathbb{R}^n$, let $t = \sup\{r : \exists B_r(x) \subset \mathbb{R}^n \setminus A\}$. Then $d_{\mathrm{GH}}(\mathbb{R}^n, A) < \infty$ if and only if $t < \infty$. Since $t < \infty$ is the same as saying that $A$ is an $\varepsilon$-net for some $\varepsilon > 0$, the finite-distance subsets of $\mathbb{R}^n$ are exactly its $\varepsilon$-nets. Finite dimensionality is essential: Example 3.1 shows that the hyperplane $x_1 = 0$ in $\ell_2$ is isometric to $\ell_2$, hence has Gromov–Hausdorff distance $0$ from $\ell_2$, but is not an $\varepsilon$-net for any $\varepsilon > 0$.
Load-bearing premise
The proof relies on the inequality from reference [6] that for bounded subsets of $\mathbb{R}^n$ the Euclidean Gromov–Hausdorff distance is at most a dimension-dependent constant times the square root of the maximum diameter times the Gromov–Hausdorff distance; the paper imports this theorem without reproving it, and if that inequality failed the argument would lose the approximate isometries it needs.
Editorial extensions
If this is right
- If $A \subset \mathbb{R}^n$ is at finite Gromov–Hausdorff distance from $\mathbb{R}^n$, then $A$ is automatically an $\varepsilon$-net, so the Gromov–Hausdorff cloud of $\mathbb{R}^n$ is exactly the class of $\varepsilon$-nets.
- Finite distance also bounds the size of holes: the complement of $A$ cannot contain closed balls of arbitrarily large radius.
- For every $\varepsilon$-net $A$, the distance $d_{\mathrm{GH}}(\mathbb{R}^n, A)$ is at most $\varepsilon$, so on the class of subsets the net radius controls the Gromov–Hausdorff distance.
- The characterization is dimension-sensitive: in infinite-dimensional Euclidean space, an isometric hyperplane has distance zero but is not an $\varepsilon$-net.
- The paper leaves open whether the analogous statement holds for arbitrary finite-dimensional normed spaces, since the proof uses the isometry group of Euclidean space.
Reading between the lines
- A quantitative version the paper does not state likely holds: $d_{\mathrm{GH}}(\mathbb{R}^n, A)$ should be bounded by a function of the empty-ball radius $t$ and the dimension $n$, not merely finite.
- The same cone-and-sphere strategy would characterize finite-distance subsets of any homogeneous metric space whose isometry group is rich enough to admit a square-root estimate like the one imported from reference [6].
- For computational geometry, the theorem reduces a question about an abstract distance to a coverage check: a point cloud is at finite Gromov–Hausdorff distance from $\mathbb{R}^n$ exactly when it is an $\varepsilon$-net, which can be tested by locating the largest empty ball.
- The $\ell_2$ example suggests that in infinite-dimensional settings, zero Gromov–Hausdorff distance can coexist with arbitrarily large holes, so a different kind of invariant would be needed there.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proves Theorem 3.1: for a non-empty subset A of R^n, the Gromov–Hausdorff distance d_GH(R^n, A) is finite if and only if A is an epsilon-net for some epsilon > 0, equivalently if and only if t = sup{r : there is a closed ball B_r(x) contained in R^n \ A} is finite. The forward direction is elementary. The reverse direction assumes a finite-distortion correspondence R, chooses a large sphere X = S_N(p) around a point p paired with a fixed a in A, uses Memoli's Euclidean Gromov–Hausdorff inequality to produce an approximate isometry f, and derives a cone property: for every a in A and every cone with vertex a of the stated shape, some point of A lies in the shell B_{N+c}(a) \ B_{N-c}(a) inside that cone. The contrapositive then tries to contradict this property using an empty ball B_r(x). Example 3.1 shows that the statement fails in l_2.
Significance. If the gap described below is repaired, the theorem gives a clean characterization of which subsets of finite-dimensional Euclidean space lie at finite Gromov–Hausdorff distance from R^n, and it sharpens the cloud picture from Gromov's book. The main estimates are coherent: the distortion bound c, the choice of N with 3T sqrt(N) < N/2, the diameter control, and the cone absorption argument all check out algebraically. The proof depends essentially on Memoli's Theorem 2.1, and the paper is explicit about this dependence and about the resulting failure in infinite dimension. The result is modest but appropriate for a short paper in math.MG, and the presentation is mostly readable.
major comments (1)
- [Section 3, final paragraph of Theorem 3.1] The step 'Without loss of generality, suppose that there exists a point a in A such that a in S_r(x)' is impossible under the paper's own convention that B_r(x) is the closed ball {u : |ux| <= r}. Indeed, B_r(x) subset R^n \ A implies S_r(x) subset B_r(x) subset R^n \ A, so A and S_r(x) are disjoint. If B_r were read as an open ball, the claim would still fail for non-closed A, as the example A = {u : u_n > 0} shows. This step is load-bearing because the cone-shell contradiction requires a point a in A whose ray to the center x forms the cone axis. I recommend replacing this with a limiting argument: from t = infinity, choose a center x with d(x,A) arbitrarily large, set r = d(x,A) - epsilon, and choose a in A with |ax| < d(x,A) + delta; with delta and epsilon small relative to N - c, the shell pieces of the cone with vertex a and axis through x lie inside B_r(x), yielding the same contradiction. As written, the contrapositive does not rule out arbitrarily large empty balls, so Theorem 3.1 is not established.
minor comments (4)
- [Section 3, after the proof that Q^{-1}(p) = {p}] The sentence 'Since R'(p) = {a}' is not justified: the restriction of the correspondence R to X' and Y' may pair p with several elements of Y. However, the subsequent conclusion U(a) = {p} follows from the weaker and correct fact that a is in R'(p) together with Q(p) = {p}, so this is a local repair rather than a fatal flaw.
- [Section 3, final paragraph] The text first says 'Let us choose some ball B_r(x)' and then says 'Let us choose r so large that the following inclusion holds.' Since r is fixed by the choice of the ball, the second instruction is logically inverted; the proof should choose an empty ball of sufficiently large radius at the outset, which is possible because t = infinity.
- [Section 2, Definition 3 and Theorem 3.1] The theorem should state explicitly that A is non-empty, since the Gromov–Hausdorff distance and the correspondence machinery are defined only for non-empty spaces.
- [Section 3, Corollary 1 proof] The equalities d_GH(X,Y) = d_GH(cl X, cl Y) and d_EH(X,Y) = d_EH(cl X, cl Y) are standard but are used without proof or citation; a one-line justification would help the reader.
Circularity Check
No circularity: Theorem 3.1 is derived from Memoli's external inequality and standard Gromov–Hausdorff facts; self-citations are motivational only.
full rationale
The derivation chain is not circular. The forward direction dGH(R^n,A) < infinity implying t < infinity rests on standard correspondence-distortion facts (Claims 1, 2, 3, 4), the diameter inequality, and Corollary 1, which is imported from Memoli's Theorem 2.1 [6]; Corollary 1 is a bounded-approximation extension by closure, not a restatement of the theorem. The only place finite dimension enters is Memoli's inequality, an external result whose stated assumptions do not include the target conclusion. No parameter is fitted to data, and no 'prediction' is generated from a fit. The self-citations [1] and [7] appear only in the introduction and are not load-bearing: [1] motivates the cloud question and [7] explains why a possible transfer to normed spaces fails; neither is used in the proof of Theorem 3.1. The example in ell_2 provides an independent falsification of the infinite-dimensional analogue, confirming that the theorem's content is not tautological. Separately, the final step of Theorem 3.1 contains an apparent correctness gap not directly related to circularity: the paper defines B_r(a) as a closed ball, so the assertion 'Without loss of generality, suppose that there exists a point a in A such that a in S_r(x)' after choosing B_r(x) subset R^n\A is impossible under that convention; this is a proof incompleteness, not a self-referential derivation. Overall score 0.
Assumptions & free parameters
assumptions (4)
- standard math Memoli inequality: for bounded X, Y in R^n, d_EH(X, Y) <= c'_n * max(diam X, diam Y)^{1/2} * d_GH(X, Y)^{1/2}.
- standard math Correspondence-distortion duality: 2 d_GH(X, Y) = inf{dis R : R is a correspondence between X and Y}.
- standard math Composition and diameter inequalities: dis(R2 composed with R1) <= dis R1 + dis R2, and |diam X - diam Y| <= 2 d_GH(X, Y) when one diameter is finite.
- standard math Elementary Euclidean geometry of spheres and cones: the sphere S_N(p) is connected, and a cone with aperture controlled by N/2 can be absorbed into a sufficiently large empty ball in the final contradiction step.
Cite this review
Pith. "Pith review of When the Gromov-Hausdorff distance between finite-dimensional space and its subset is finite?." pith.science (2026). https://pith.science/paper/P7QPWHC5
@misc{pith2026241113539,
author = {Pith},
title = {Pith review of: When the Gromov-Hausdorff distance between finite-dimensional space and its subset is finite?},
year = {2026},
howpublished = {\url{https://pith.science/paper/P7QPWHC5}},
note = {Machine review of arXiv:2411.13539}
}
abstract
In this paper we prove that the Gromov--Hausdorff distance between $\mathbb{R}^n$ and its subset $A$ is finite if and only if $A$ is an $\varepsilon$-net in $\mathbb{R}^n$ for some $\varepsilon>0$. For infinite-dimensional Euclidean spaces this is not true. The proof is essentially based on upper estimate of the Euclidean Gromov--Hausdorff distance by means of the Gromov-Hausdorff distance.
Reference graph
Works this paper leans on
-
[1]
S. A. Bogaty, A. A. Tuzhilin, Gromov–Hausdorff class: its completeness and cloud geometr y, ArXiv e-prints, arXiv:2110.06101, 2021
arXiv 2021
-
[2]
Burago, Yu
D. Burago, Yu. Burago, S. Ivanov, A Course in Metric Geometry , Graduate Studies in Mathe- matics 33, AMS, 2001
2001
-
[3]
Gromov, Structures m´ etriques pour les vari´ et´ es riemanniennes, Textes Math
M. Gromov, Structures m´ etriques pour les vari´ et´ es riemanniennes, Textes Math. 1 (1981)
1981
-
[4]
Gromov, Metric structures for Riemannian and non-Riemannian space s, Birkh¨ user (1999)
M. Gromov, Metric structures for Riemannian and non-Riemannian space s, Birkh¨ user (1999)
1999
-
[5]
D. A. Herron, Gromov–Hausdorff Distance for Pointed Metric Spaces , J. Anal., 2016, v. 24, N 1, pp. 1–38
2016
-
[6]
F. Memoli, Gromov-Hausdorff distances in Euclidean spaces , in IEEE Computer Society Confer- ence on Computer Vision and Pattern Recognition Workshops, June (2008), pp. 1–8
work page 2008
-
[7]
I. N. Mikhailov, Gromov–Hausdorff distances between normed spaces , arXiv:2407.01388, 2024 (in print)
arXiv 2024
-
[8]
A. A. Tuzhilin, Lectures on Hausdorff and Gromov–Hausdorff distance geometr y, arXiv:2012.00756, 2019. 6
arXiv 2012
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.