A compact metric set that still needs infinitely many dimensions

← All digests

Family 098 · Convex and metric geometry
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-14

1. The question

A set may sit in an infinite-dimensional space while having very modest local complexity. Suppose every ball in the set can be covered by a fixed number of balls of half the radius. Does that mean the set can be represented faithfully in some finite-dimensional Euclidean space?

For subsets of Hilbert space, this is the Lang–Plaut problem. The hypothesis already provides Euclidean-style distances, just with infinitely many possible coordinates. The issue is whether a uniform covering bound makes infinitely many coordinates unnecessary.

The manuscript in family 098 claims a negative answer. It first constructs one doubling subset of Hilbert space that has no finite-dimensional bi-Lipschitz embedding at any finite distortion. It then turns finite pieces of that obstruction into a compact example inside every infinite-dimensional real Banach space.

Compactness does not remove the difficulty: the claimed obstruction can be bounded and contain all its limit points, with increasingly intricate configurations appearing at smaller and smaller scales.

2. Background and essential definitions

A metric space $(X,d)$ is doubling with constant $\lambda$ if every ball of radius $r$ can be covered by at most $\lambda$ balls of radius $r/2$, centered in $X$. The same number works at every location and scale.

Repeatedly applying this condition gives a polynomial bound on covering numbers as the radius shrinks. It is a notion of finite metric complexity, but does not prescribe a coordinate system or a differentiable structure.

A map $f:X\to\mathbb R^k$ is bi-Lipschitz with distortion at most $D$ if some $a>0$ satisfies

$$a\,d(x,y)\leq|f(x)-f(y)|\leq Da\,d(x,y)$$

for all pairs. The lower bound prevents collapsing distinct points; the upper bound prevents excessive stretching. Both $k$ and $D$ must be finite, but they need not be prescribed in advance.

A Hilbert space has a norm coming from an inner product. The sequence space $\ell_2$, with $\sum_jx_j^2<\infty$, is the standard infinite-dimensional example. A Banach space is a complete normed vector space and need not have an inner product.

A set is compact if every sequence in it has a convergent subsequence with limit in the set. For an infinite-dimensional ambient space, boundedness and closedness alone do not guarantee compactness; the final construction explicitly arranges shrinking finite clusters to obtain it.

3. Development history and previous partial results

Lang and Plaut asked the Hilbert-subset question in Bilipschitz embeddings of metric spaces into space forms (2001). Gupta, Krauthgamer and Lee independently raised it in an algorithmic dimension-reduction context.

Assouad’s embedding theorem gives a positive result after changing the metric. If $0<\alpha<1$, the snowflake distance $d(x,y)^\alpha$ on a doubling space admits a finite-dimensional Euclidean bi-Lipschitz embedding. This power expands small distances relative to large ones. It does not preserve the original metric up to a uniform factor across arbitrarily small scales.

Naor and Neiman’s dimension-independent snowflaking theorem bounds the target dimension in terms of the doubling constant alone when $\alpha$ is near one. Its distortion still depends on $\alpha$. One cannot simply let $\alpha\to1$ and conclude a bounded-distortion embedding of the original distance.

For ambient $L^p$ spaces with $p>2$, Lafforgue and Naor constructed doubling sets that resist finite-dimensional Euclidean embeddings. Bartal, Gottlieb and Neiman obtained related obstructions through thin Laakso configurations, and later work gave an elementary fixed-set version. Those results leave the Hilbert exponent $p=2$ untreated.

The source also records a 2017 announcement by Schioppa, subsequently withdrawn after a possible problem in its duality argument. It is historical context, not an established solution used by M1.

The new manuscript uses ordinary differentiation on planar sheets and a packing contradiction. Its compact Banach-space extension then invokes the classical Dvoretzky theorem, which supplies almost Euclidean finite-dimensional subspaces inside every infinite-dimensional Banach space.

4. What the new manuscripts claim

M1 first states that there is a fixed subset $S$ of real $\ell_2$, with its inherited Hilbert distance, whose doubling constant is at most $76\,800$, such that no embedding

$$S\longrightarrow\mathbb R^k$$

is bi-Lipschitz for any positive integer $k$ and any finite distortion. The set is chosen once; it is not reconstructed after the target dimension or distortion is announced.

A later corollary strengthens the setting in another direction. There is a universal finite constant $\Lambda$ such that every infinite-dimensional real Banach space $B$ contains a compact subset $K_B$ that is $\Lambda$-doubling and has no finite-distortion bi-Lipschitz embedding into any finite-dimensional real normed space.

The compact set may depend on $B$, but its covering constant has a bound independent of $B$. The explicit $76\,800$ bound belongs to the initial Hilbert construction; the compact assembly uses a larger universal bound. No optimality of these constants is claimed.

M1 also observes that the original Hilbert metric can be realized isometrically in real $L^p([0,1])$ for each $1\leq p<\infty$, using normalized Gaussian series. A linear combination of independent standard Gaussians has its $p$-th moment proportional to the corresponding Euclidean norm to the power $p$, which supplies this isometry.

The no-embedding conclusion allows arbitrary nonlinear maps. It is not merely an obstruction to linear projection or to one proposed coordinate-selection algorithm.

5. Why the advance matters

Importance: 5/5. The claim would answer a longstanding basic question about the relationship between metric covering complexity and Euclidean dimension. Doubling is a strong uniform local restriction, yet it would not suffice to compress a Hilbert subset into any finite-dimensional target while preserving the original metric up to a fixed factor.

The compact extension makes the separation particularly clear. The obstruction does not require points escaping to infinity. Instead, finite difficulties accumulate near one limit point, each at a smaller scale. A target that succeeds at every scale simultaneously would have to overcome all of them with one dimension and one distortion.

This fits consistently with finite-set dimension reduction. For any finite collection, a dimension bound may grow with its cardinality. The theorem here asks for one finite target for the entire infinite set. No single bound is forced by applying finite-set theorems to larger and larger pieces.

The Banach-space conclusion also shows that the phenomenon is not peculiar to one carefully selected ambient norm. Dvoretzky’s theorem lets the finite Hilbert obstructions be placed, with controlled distortion, in every infinite-dimensional real Banach space.

6. The main idea

Add many possible directions at widely separated scales

The first construction starts with a base plane. At level $j$, use a small scale $r_j=1000^{-j}$ and a collection of new mutually orthogonal directions. Horizontal and vertical strips are periodically colored. A base point is allowed to acquire a displacement in a color’s direction when it belongs to a strip of that color.

Points carry only finitely many such displacements. The color counts and strip aspect ratios are fixed in advance so that every pair of positive integer choices recurs infinitely often.

Despite unboundedly many colors across the construction, the set remains doubling. At a given observation radius, coarse displacement coordinates are fixed, fine coordinates have very small total diameter, and at most one intermediate level contributes a nontrivial choice of labels. Only a bounded number of its strips meet the local base region. This gives a uniform covering count.

A finite target would force too many separated offsets

Suppose a bi-Lipschitz embedding into $\mathbb R^k$ existed with distortion $D$. Fixing displacement coordinates and varying the base point produces a planar sheet. Its image is a Lipschitz map on an open subset of $\mathbb R^2$, so ordinary almost-everywhere differentiation applies.

The proof examines the two derivative columns across all sheets. It first maximizes the squared size of the horizontal column and then maximizes the vertical one among limiting horizontal maximizers. Near this ordered extremum, a nearly maximal mean derivative has little mean-square variation.

Choose a color count larger than the number of well-separated points that can fit into a radius-$D$ ball in $\mathbb R^k$. Recurring levels with increasingly wide strips let the argument find one horizontal and one vertical line of each color on which relative sheet offsets are nearly constant.

At crossings of differently colored lines, orthogonal source displacements remain separated. Transporting the nearly constant offsets to selected intersections produces too many separated vectors in the bounded target ball. Finite-dimensional volume packing contradicts that possibility.

The source carefully uses different error estimates in the two coordinate directions, because the horizontal lines grow longer as the strip aspect ratio increases. The fixed construction contains all required recurring scales before the hypothetical embedding is chosen.

Assemble finite obstructions into a compact set

If every finite subset of $S$ embedded into one fixed $\mathbb R^k$ with one fixed distortion, a compactness argument would produce an embedding of all of $S$. Therefore finite witnesses already exist against each prescribed dimension and distortion.

For each integer $j$, select a finite witness against dimension $j$ and distortion $j$. Dvoretzky’s theorem places its finite Hilbert span inside $B$ with distortion below two. Scale these copies into tiny clusters near points at distances $2^{-j}$ along a fixed ray, then add the origin.

The clusters shrink to the origin, making the union compact. A ball meets only boundedly many clusters of comparable size; all much smaller clusters fit into one small ball near zero. This preserves a universal doubling bound. Any finite-dimensional embedding of the union would restrict to an embedding of a sufficiently demanding witness, contradicting its choice.

7. Scope, limitations, and open questions

The theorem preserves the original induced metric up to a multiplicative factor. Snowflaking changes that metric, so Assouad-type positive results remain compatible with the claimed obstruction.

The initial explicit Hilbert set and the final compact Banach-space set are different constructions. The latter has a universal but larger doubling bound and may depend on the ambient Banach space. Neither depends on a subsequently proposed target dimension or distortion.

Finite-dimensional normed targets are covered because any two norms on a fixed finite-dimensional space are equivalent with a finite factor. Infinite-dimensional targets, of course, remain possible: the sets already live in them.

The result does not rule out positive dimension reduction under additional geometric assumptions beyond doubling, or for finite data with dimension growing with sample size. Determining which extra conditions restore a finite-dimensional embedding remains a natural direction.

8. Sources and further reading

M1 supplies the sparse-scale Hilbert construction, derivative obstruction, Gaussian realization and compact Banach-space assembly. Its last section separates the finite-witness argument from the use of Dvoretzky’s theorem.

The linked Lang–Plaut paper states the original geometric context. Naor–Neiman explains what snowflaking can recover. The finite-set comparison in family 094 concerns a different quantifier: its target dimension may grow with the number of points.

Source manuscripts

• M1. OpenAI, A doubling Hilbert subset with no finite-dimensional bi-Lipschitz embedding (September 25, 2026).

Sources consulted on 14 October 2026 at repository revision 1581d4c. Download the manuscript citations.