A doubling Hilbert subset with no finite-dimensional embedding

Source and verification

This article explains family 098 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ Doubling spaces and Euclidean embeddings

A metric space is doubling with constant $\lambda$ if every ball of radius $r$ can be covered by at most $\lambda$ balls of radius $r/2$, with centers in the space. This is a quantitative restriction on the number of distinguishable points at each scale.

A bi-Lipschitz embedding into ${\mathbb{R}}^k$ has constants $s>0$ and $D<\infty$ such that

$$\begin{array}{rl}sd(x,y)& \le |f(x)-f(y)|\\ & \le Dsd(x,y)\end{array}$$

for every pair. The target dimension and distortion are fixed for the entire space.

Finite-dimensional Euclidean subsets are doubling. The converse becomes subtle even if the original space already sits isometrically inside an infinite-dimensional Hilbert space.

🏷️ A fixed-set obstruction

Statement in the manuscript

There is a single subset $S$ of real ${\ell }_2$, with doubling constant at most $76800$, that admits no bi-Lipschitz embedding into any finite-dimensional Euclidean space at any finite distortion. (OpenAI, 2026)

The quantifiers matter: $S$ is selected first, independently of $k$ and $D$. The claim is stronger than constructing a different difficult finite set for each proposed target dimension.

The paper also constructs a compact obstruction inside every infinite-dimensional real Banach space, with a universal doubling constant. That compact construction has its own constant; the numerical value $76800$ belongs to the stated Hilbert construction.

Doubling spaces have positive embedding results after replacing the distance $d$ by a snowflake $d^{\alpha }$, $0<\alpha <1$. Such a replacement changes the metric, so it does not contradict an obstruction for exponent one.

🏷️ Sparse orthogonal displacements

The source begins with a plane and adds small orthogonal displacements at scales

$$r_j=1000^{-j}.$$

At a given level, horizontal strips and vertical strips receive colors. A displacement in an additional Hilbert coordinate is permitted when the base point lies in the corresponding row or column. Finite-support tuples of allowed displacements define the set.

The number of colors and the width ratio of the strips vary with the level. Every required parameter pair occurs infinitely often. Thus, after a hypothetical embedding fixes $k$ and $D$, the proof can select a suitable recurring level without changing the space itself. (OpenAI, 2026)

Why can this still be doubling? At a given observation radius, much coarser displacement coordinates are fixed within a sufficiently small ball, while the sum of much finer tails is negligible. Only one intermediate scale requires a detailed covering argument. The large separation between successive $r_j$ keeps these three contributions uniformly bounded.

This is a scale-local geometric property. It does not prevent the full infinite collection of levels from imposing mutually incompatible requirements on a fixed-dimensional target.

🏷️ Differentiation and near-extremal sheets

Assume an embedding into ${\mathbb{R}}^k$ with distortion $D$ exists and rescale its lower constant to one. Each planar sheet inherits a Lipschitz map, so Rademacher’s theorem supplies derivatives almost everywhere.

The manuscript collects pairs of squared derivative norms in a compact set and chooses a lexicographic maximum: first maximize the horizontal value, then the vertical value among horizontal maximizers. Near-maximal horizontal energy then constrains vertical energy without requiring an explicit rate for the second maximization.

The underlying Hilbert identity is

$$\begin{array}{rl}\mathbb{E}|V-v{|}^2& =\mathbb{E}|V{|}^2-|\mathbb{E}V{|}^2\\ & +|\mathbb{E}V-v{|}^2.\end{array}$$

If the mean is close to a maximal-norm vector and the mean squared norm is bounded by the same maximum, the variance is small. This converts near-extremal derivative information into nearly constant relative offsets between sheets.

The proof uses long thin rectangles. Horizontal errors must be small enough to survive transport over their growing width; vertical errors are controlled on a fixed normalized height. A qualitative error in the wrong direction would be insufficient, which is why the two derivative selections are treated separately.

🏷️ Crossings and a packing contradiction

At crossings of selected differently colored rows and columns, two allowed source displacements differ by $\sqrt{2}{r}_j$. The lower Lipschitz bound keeps their images separated. Near-constant offset estimates transport this separation to a common family of vectors in the target.

The result is a collection of $N$ mutually separated vectors inside a ball whose radius depends only on $D$. Euclidean volume comparison bounds the possible number by a quantity depending only on $k$ and $D$. Choosing the recurring color parameter $N$ larger than this capacity yields a contradiction. (OpenAI, 2026)

The mechanism needs both ingredients: local differentiation makes offsets comparable across the rectangle, and the crossing pattern supplies many independent separation requirements.

🏷️ Compactness and Banach-space transfer

If all finite subsets of $S$ embedded with the same $k,D$, a product-compactness argument would assemble a global embedding. Hence the obstruction has finite witnesses for each pair of bounds.

The classical finite-dimensional Dvoretzky theorem places each finite witness’s Hilbert span almost isometrically inside any infinite-dimensional Banach space. The manuscript rescales these witnesses into geometrically shrinking clusters and adjoins their limit point. The union is compact.

The doubling estimate separates clusters comparable to the observation radius from the small tail. Only a bounded number of comparable clusters can meet a ball, and the tail fits in one additional ball. Any fixed-dimensional bi-Lipschitz embedding of the union would restrict to all the finite witnesses, contradicting their choice.

The paper also realizes the original Hilbert metric isometrically in real $L_p[0,1]$, $1\le p<\infty$, through normalized Gaussian series. This statement concerns function-space sources, not an identification of coordinate ${\ell }_p$ with Hilbert space.

🏷️ Key steps and proof outlines

Scale separation gives a uniform local covering number

Fix a source ball of radius $t$. If an orthogonal displacement coordinate has step size $r_j>t$, every point in the ball has that coordinate fixed to the center’s value. Because $r_j/r_{j+1}=1000$, at most one level lies in the intermediate range $t/64<r_j\le t$. All finer levels have squared tail length bounded by a geometric series.

The base-plane projection lies in a square of side $2t$. At the intermediate level, an interval of that length meets at most $\lfloor 2t/r_j\rfloor +2\le 129$ relevant strips. The construction’s row/column rules and a fixed planar grid cover this finite intermediate complexity by a universal number of half-radius balls. The fine tail is absorbed into the remaining radius margin. This is the source’s explicit doubling calculation. The number of colors can increase indefinitely because only the locally permitted displacements at the one intermediate level enter the count. (OpenAI, 2026)

A derivative extremum provides a common comparison vector

Assume a map to ${\mathbb{R}}^k$ has lower Lipschitz constant one and upper constant $D$. Restrict it to each planar sheet. Almost-everywhere derivatives have norms between the appropriate lower and upper bounds. The source considers the closure of the possible pairs of squared horizontal and vertical derivative norms and chooses a lexicographic maximum.

First choose a sheet and a differentiation point with horizontal energy close to its maximum. Blow up around that point to obtain an approximately affine map on a long thin rectangle. Among almost horizontally maximizing choices, select vertical energy close to the second maximum. This order matters: the rectangle’s horizontal aspect ratio will be large, so its horizontal error needs a stronger quantitative tolerance than its vertical error.

The variance identity above then shows that neighboring sheets have almost constant relative offsets. To see the mechanism, write a derivative field as its mean plus fluctuation. Its integrated squared norm is the squared mean plus the nonnegative variance. If the mean already nearly attains the maximal allowed energy, little room remains for fluctuation. Integrating along horizontal and vertical segments transports this control across the rectangle.

Crossings create many separated target offsets

Choose a recurring scale whose color parameter is large compared with the proposed target packing capacity. A displacement of size $r_j$ in one color coordinate and one in another differ by $\sqrt{2}{r}_j$ in Hilbert norm. The lower Lipschitz bound preserves that separation up to the common normalization.

At a selected row-column crossing, both relevant sheets exist. Their local separation becomes a comparison between their target offsets. Near constancy transports those offsets to one common reference location; the total transport error is chosen smaller than a fixed fraction of $r_j$. Dividing offsets by $r_j$ therefore produces many uniformly separated vectors in a target ball of radius depending only on $D$.

If their separation is at least $c>0$ and their norms are at most $R$, disjoint Euclidean balls of radius $c/2$ centered at them fit inside radius $R+c/2$. Consequently their number is at most

$${\left(1+\frac{2R}{c}\right)}^k.$$

The preselected source contains recurring levels with more colors than this number, contradicting the embedding. Differentiation supplies comparable offsets; the crossing geometry supplies their pairwise separation; finite-dimensional volume supplies the contradiction.

Finite witnesses can be assembled into a compact obstruction

If every finite subset embedded with common $k,D$, fix one point at the origin in the target and place every other image in its bounded admissible ball. Product compactness and the closed pairwise distance inequalities produce a global embedding. Therefore each prescribed $(k,D)$ has a finite obstructing subset.

Take a sequence of such witnesses with increasing bounds. Dvoretzky’s theorem places each finite-dimensional Hilbert span almost isometrically in any chosen infinite-dimensional Banach space. Rescale the witnesses into clusters of rapidly decreasing diameter, position them with controlled separation, and add their limit point. Compactness follows because all but finitely many clusters lie in any neighborhood of that point.

The doubling proof again separates clusters larger than the observation scale, comparable clusters, and a negligible tail. A fixed target embedding of the union would restrict to every witness, eventually violating its chosen obstruction. This compact construction is a second assembly theorem; it is not obtained merely by taking the closure of the original unbounded planar-sheet set.

See Also

Finite-set dimension reduction allows target dimension to grow with the number of points. Graph embeddings in L1 uses structural graph hypotheses and a different target geometry.

🏷️ Manuscripts and verification scope

• A doubling Hilbert subset with no finite-dimensional bi-Lipschitz embedding — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope describes the fixed Hilbert-space obstruction and its stated transfers; the source manuscript should be consulted for the distinct constants and compactness assertions. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026. A doubling Hilbert subset with no finite-dimensional bi-Lipschitz embedding.