OctoInformBeta

OpenAI preprint, September 2026 · §7. Compact obstructions in infinite-dimensional Banach spaces

Placement and cluster doubling: Auerbach vectors, linear maps of distortion 3636 on finite subsets of SS, and 76800876800^8-doubling placements near rvrv if SS is 7680076800-doubling

Finite subsets of SS placed in infinite-dimensional normed spaces.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Let SS be the subset of the Hilbert space ℓ2\ell^2 built in the next section, and let BB be an infinite-dimensional real normed space. The question is how faithfully a finite subset of SS can be copied into BB, and whether the copy can be put near a prescribed point of BB without losing control of its doubling constant.

The Constructed Set

The Hilbert space ℓ2\ell^2 is the space of real sequences a=(an)n≥0a = (a_n)_{n \ge 0} with ∑nan2<∞\sum_n a_n^2 < \infty, normed by ∥a∥=(∑nan2)1/2\|a\| = (\sum_n a_n^2)^{1/2}.✦ The base plane ℝ2\mathbb{R}^2 is the set of pairs p=(x,y)p = (x, y) of real numbers with the sup distance max⁡(|x−x′|,|y−y′|)\max(|x - x'|, |y - y'|).✦ The base vector of p=(x,y)p = (x, y) is the vector of ℓ2\ell^2 with entry xx at index 00, entry yy at index 11 and all other entries zero, and pp also stands for it.✦

Levels are indexed by j≥0j \ge 0. The scale of level jj is rj=1000−(j+1)r_j = 1000^{-(j+1)}.✦ Fix a bijection ⟨⋅,⋅⟩:ℕ×ℕ→ℕ\langle \cdot, \cdot \rangle \colon \mathbb{N} \times \mathbb{N} \to \mathbb{N}, the pairing bijection. The number of colors of level jj is Nj=a+1N_j = a + 1 and the width factor of level jj is Wj=b+1W_j = b + 1, where the inverse of the pairing bijection sends jj to a pair (m,k)(m, k) and mm to the pair (a,b)(a, b). The horizontal strips of level jj are qrj<y<(q+1)rjq r_j < y < (q+1) r_j and the vertical strips are qWjrj<x<(q+1)Wjrjq W_j r_j < x < (q+1) W_j r_j, for q∈ℤq \in \mathbb{Z}; the strip of index qq has color qmod⁡Njq \bmod N_j✦, and Uj,iU_{j,i} is the union of the open strips of color ii of both kinds.✦ For a level jj and a color ii, the unit vector ej,ie_{j,i} is the coordinate vector of ℓ2\ell^2 at index 2+⟨j,i⟩2 + \langle j, i \rangle; these vectors are orthogonal to each other and to the base plane, and rjej,ir_j e_{j,i} is the level vector of level jj and color ii.

An admissible displacement at a point pp of the base plane is a finitely supported sequence w=(wj)j≥0w = (w_j)_{j \ge 0} in which each wjw_j is either 00 or a level vector rjej,ir_j e_{j,i} of a color i∈{0,…,Nj−1}i \in \{0, \dots, N_j - 1\} with p∈Uj,ip \in U_{j,i}.✦ The constructed set SS is the subset of ℓ2\ell^2 of all vectors (p,w)=p+∑j≥0wj,(p, w) = p + \sum_{j \ge 0} w_j,(1)✦ with pp in the base plane and ww an admissible displacement at pp; it carries the distance of ℓ2\ell^2.

Doubling Subsets

For a natural number NN, a subset KK of a metric space MM with distance dd is NN-doubling if, for every point xx of KK and every radius r>0r > 0, there is a set CC of at most NN points of KK such that K∩{y:d(y,x)<r}⊂⋃c∈C{y:d(y,c)<r/2}.K \cap \{ y : d(y, x) < r \} \subset \bigcup_{c \in C} \{ y : d(y, c) < r/2 \}.(2)✦

The Claim

Theorem 1. Let BB be an infinite-dimensional real normed space.

  1. For every natural number nn there are Auerbach vectors v1,…,vnv_1, \dots, v_n in BB: for all real numbers c1,…,cnc_1, \dots, c_n and every index ii, ∥vi∥≤1,|ci|≤∥∑j=1ncjvj∥.\|v_i\| \le 1, \qquad |c_i| \le \Big\| \sum_{j=1}^{n} c_j v_j \Big\|.(3)✦

  2. For every finite subset XX of SS there is a linear map T:ℓ2→BT \colon \ell^2 \to B such that, for all x,y∈Xx, y \in X, 16∥x−y∥≤∥Tx−Ty∥≤6∥x−y∥.\tfrac{1}{6}\, \|x - y\| \le \|Tx - Ty\| \le 6\, \|x - y\|.(4)✦

  3. Suppose that SS is 7680076800-doubling. For every finite subset XX of SS, every vector vv of BB and every r>0r > 0, there are a map φ:X→B\varphi \colon X \to B whose image φ(X)\varphi(X) is 76800876800^8-doubling and a number a>0a > 0 such that, for all x,y∈Xx, y \in X, ∥φ(x)−rv∥≤r100,a∥x−y∥≤∥φ(x)−φ(y)∥≤36a∥x−y∥.\|\varphi(x) - rv\| \le \frac{r}{100}, \qquad a\, \|x - y\| \le \|\varphi(x) - \varphi(y)\| \le 36\, a\, \|x - y\|.(5)✦

  4. Let LL be a natural number, EE an LL-doubling subset of a metric space MM, M′M' a metric space, ZZ a set, and π:Z→E\pi \colon Z \to E and g:Z→M′g \colon Z \to M' two maps such that, for some a>0a > 0 and all z,z′∈Zz, z' \in Z, ad(π(z),π(z′))≤d(g(z),g(z′))≤36ad(π(z),π(z′)).a\, d(\pi(z), \pi(z')) \le d(g(z), g(z')) \le 36\, a\, d(\pi(z), \pi(z')).(6)✦ Then the image g(Z)g(Z) is L8L^8-doubling.

In the paper, the result reads: each finite witness XjX_j, which plays the part of XX, is placed in BB by Dvoretzky’s theorem with distortion less than 22, so that each image YjY_j is doubling with constant at most 76800476800^4.

What Is Checked

The four parts of the claim are stated over an infinite-dimensional real normed space BB and, in part 4, over arbitrary metric spaces; over the Hilbert space ℓ2\ell^2 and its subset SS of points (1), built from the base plane, the scales rjr_j, the colors NjN_j, the width factors WjW_j, the strips Uj,iU_{j,i} and the unit vectors ej,ie_{j,i}; and over the doubling condition (2).