Dvoretzky’s Theorem Overview

• Dvoretzky’s theorem is a fundamental result in convex geometry and Banach space theory, establishing almost Euclidean sections in high-dimensional centrally-symmetric convex bodies.
• Quantitative forms reveal that the maximal subspace dimension scales as ε²n, with precise asymptotic estimates impacting random projections and polynomial approximations.
• Extensions of the theorem include randomized, metric, and measure-theoretic generalizations, linking it to high-dimensional probability, quantum information, and modern asymptotic analysis.

Dvoretzky’s theorem is a foundational result in convex geometry and the local theory of Banach spaces, characterizing the existence of almost Euclidean sections of high-dimensional convex bodies and normed spaces. The classical statement and its numerous generalizations—and analogs for polynomials, measures, random projections, nonlinear metric structures, and random matrix models—form a central pillar of asymptotic geometric analysis and have deep connections with concentration of measure, the geometry of Banach spaces, and high-dimensional probability.

1. Classical Dvoretzky Theorem and Its Sharp Quantitative Form

Let $K \subset \mathbb{R}^n$ be a centrally-symmetric convex body. The classical Dvoretzky theorem asserts the following: for every $\varepsilon > 0$ and integer $k$, there exists $n_0 = n_0(k, \varepsilon)$ such that if $n \geq n_0$, one can find a $k$-dimensional subspace $V \subset \mathbb{R}^n$ with

$$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$

where $B_V$ is the Euclidean unit ball in $V$ (Dol'nikov et al., 2010). Thus, high-dimensional convex bodies admit almost spherical lower-dimensional sections. This existential form underpins modern asymptotic geometric analysis.

Quantitative versions reveal that the maximal such $\varepsilon > 0$0—the Dvoretzky dimension—is proportional to $\varepsilon > 0$1 up to absolute constants; specifically, for any $\varepsilon > 0$2 centrally symmetric and $\varepsilon > 0$3, there exist universal constants $\varepsilon > 0$4 such that

$\varepsilon > 0$5

where $\varepsilon > 0$6 is the maximal $\varepsilon > 0$7 such that a $\varepsilon > 0$8-dimensional subspace $\varepsilon > 0$9 satisfies the $k$0-sphericity above (Mendelson, 2020).

2. Extensions: Polynomials, Bundles, and the Polynomial Dvoretzky Number

Dvoretzky-type phenomena persist beyond linear sections. For homogeneous real polynomials $k$1 of even degree $k$2, there exists $k$3 such that for $k$4, there is a $k$5-dimensional subspace $k$6 where $k$7 is a round quadratic form to the $k$8 power: $k$9 for some constant $n_0 = n_0(k, \varepsilon)$0. This number $n_0 = n_0(k, \varepsilon)$1, the polynomial–Dvoretzky number, is also fundamental for the odd-degree “Birch” case and complex analogs: $n_0 = n_0(k, \varepsilon)$2 for both odd $n_0 = n_0(k, \varepsilon)$3 (real case) and all $n_0 = n_0(k, \varepsilon)$4 in the complex case (Dol'nikov et al., 2010).

A stronger but generally false conjecture stated that every section of the symmetric power bundle $n_0 = n_0(k, \varepsilon)$5 must be round on some $n_0 = n_0(k, \varepsilon)$6-plane for large $n_0 = n_0(k, \varepsilon)$7 for even $n_0 = n_0(k, \varepsilon)$8. Counterexamples arise for certain $n_0 = n_0(k, \varepsilon)$9, though the conjecture holds for $n \geq n_0$0 when $n \geq n_0$1, all odd $n \geq n_0$2, and for all $n \geq n_0$3 in the complex Grassmannian (Dol'nikov et al., 2010).

3. Randomized, Isomorphic, and Metric Generalizations

Randomized Dvoretzky’s theorem strengthens the existential form by establishing that random subspaces of appropriate dimension are almost always $n \geq n_0$4-Euclidean for a wide class of norms. In $n \geq n_0$5, the critical dimension $n \geq n_0$6 for $n \geq n_0$7-spherical random sections exhibits precise asymptotics depending on $n \geq n_0$8:

• For $n \geq n_0$9: $k$0.
• For $k$1: $k$2.
• For $k$3: $k$4 (Paouris et al., 2015).

Paouris and Valettas established a two-level Gaussian–$k$5 tail phenomenon for subspaces of $k$6, yielding optimal bounds for the dimension $k$7 of almost-spherical sections: $k$8 with $k$9 absolute in $V \subset \mathbb{R}^n$0 (Paouris et al., 2015). Enhanced concentration arises in the $V \subset \mathbb{R}^n$1-position with $V \subset \mathbb{R}^n$2-unconditional bases: the critical dimension improves to

$V \subset \mathbb{R}^n$3

with high probability (Tikhomirov, 2017).

In the isomorphic regime, Dvoretzky–Milman phenomena hold for a broad class of random matrices, including non-Gaussian, heavy-tailed, and product ensembles, under small-ball and subgaussian assumptions (Mendelson, 2020).

Beyond linearity, Mendelson showed that subgaussian coordinate projections of convex function classes admit large almost-Euclidean sections, with critical dimension determined by the Gaussian mean-width $V \subset \mathbb{R}^n$4 and $V \subset \mathbb{R}^n$5 diameter $V \subset \mathbb{R}^n$6: $V \subset \mathbb{R}^n$7 (Mendelson, 2014).

In the metric setting, infinite versions assert that every countably infinite metric space contains an infinite $V \subset \mathbb{R}^n$8-embeddable subset into Hilbert space, and analytic sets of infinite Hausdorff dimension contain infinite-dimension $V \subset \mathbb{R}^n$9-ultrametric embeddable closed subsets (Funano, 2011).

4. Measure-Theoretic, Lipschitz, and Quantum Applications

Dvoretzky-type results extend to high-dimensional probability measures: for a random vector $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$0 in $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$1 with isotropic and mild radial/directional control, most $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$2-dimensional marginals are nearly Gaussian in bounded-Lipschitz distance provided

$$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$3

with this threshold being sharp (Meckes, 2011). Under log-concavity assumptions, the threshold can be higher, paralleling the volume/cotype-improved regimes in Banach space theory.

A Lipschitz variant, crucial for applications in quantum information, states that for a circled $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$4-Lipschitz function $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$5 on the unit sphere in $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$6, and $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$7, a random $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$8-dimensional subspace $$(1-\varepsilon) B_V \subseteq K \cap V \subseteq (1+\varepsilon) B_V,$$9 (for $B_V$0) satisfies

$B_V$1

with overwhelming probability, for any central value $B_V$2 (Aubrun et al., 2010). This underlies Hastings' construction of counterexamples to the additivity of the minimal output von Neumann entropy: the sharp Lipschitz Dvoretzky theorem guarantees that, with explicit dimension and distortion, certain quantum channels achieve nearly minimal output entropy and violate additivity constraints (Aubrun et al., 2010).

5. Proof Methods and Structural Lemmas

Core proof techniques involve concentration of measure (Lévy's lemma and Gaussian concentration), chaining/net arguments (Dudley's inequality), variance bounds (e.g., Talagrand's $B_V$3–$B_V$4 inequality), symmetrization (e.g., Carathéodory averaging for forms, Sylow group averaging for symmetry), and entropy methods. For polynomials, key steps include:

• Forcing symmetry via group and net arguments to ensure invariance under large subgroups of $B_V$5.
• Averaging (in the Carathéodory sense) to guarantee that pullbacks become proportional to powers of the standard quadratic form.
• Application of Borsuk–Ulam-type results for $B_V$6-toral groups to assert the existence of subspaces with desired invariance or vanishing properties (Dol'nikov et al., 2010).

In subgaussian settings, chaining and small-ball estimates compensate for scarcity of rotation invariance (Mendelson, 2020). In measure-theoretic settings, entropy bounds, truncation, and high-dimensional anti-concentration provide tight control over projections (Meckes, 2011).

6. Sharpness, Limitations, and Open Problems

Sharpness of Dvoretzky-type bounds is evidenced by extremal examples: For the cross-polytope, the phase transition in measure-theoretic Dvoretzky occurs exactly at $B_V$7 (Meckes, 2011); in random $B_V$8 sections, the known upper and lower estimates match and interpolate precisely between the two extreme cases $B_V$9 (cross-polytope: $V$0) and $V$1 (cube: $V$2) (Paouris et al., 2015).

Some conjectures, notably the general polynomial field conjecture, are provably false in specific regimes, particularly for bundle-valued sections when $V$3 and $V$4 odd, or even $V$5 with small even $V$6 (Dol'nikov et al., 2010). Further, while the infinite nonlinear results for metric spaces achieve distortion $V$7 in countable regimes, they only achieve $V$8 distortion for analytic sets of infinite Hausdorff dimension—a barrier inherited from finite-dimensional nonlinear Dvoretzky theory (Funano, 2011).

Outstanding research directions include optimization of constants and dependence on parameters in the isomorphic theorem for general (especially heavy-tailed) ensembles, extension to non-symmetric bodies or other target spaces, explicit quantitative bounds for more structured convex sets (e.g., Schatten classes, polytopes), and the delineation of precise thresholds in measure and subgaussian Dvoretzky-type theorems (Mendelson, 2020, Mendelson, 2014).

7. Structural and Geometric Impact

Dvoretzky's theorem and its descendants have transformed the understanding of high-dimensional convexity, the geometry of Banach spaces, and random matrix theory. They provide universal euclidean and gaussian substructures in high dimension, underpin probabilistic geometric functional analysis, and are crucial for constructions in quantum information theory (including entropy counterexamples), metric embedding theory, numerical linear algebra, compressed sensing, and the analysis of random projections. The interplay between dimension, symmetry, and concentration phenomena—as formalized by the Dvoretzky program—remains central across geometric analysis and high-dimensional probability.

---

References:

• Dvoretzky type theorems for multivariate polynomials and sections of convex bodies (Dol'nikov et al., 2010)
• On Dvoretzky's theorem for subspaces of $V$9 (Paouris et al., 2015)
• Random version of Dvoretzky's theorem in $\varepsilon > 0$00 (Paouris et al., 2015)
• Superconcentration, and randomized Dvoretzky's theorem for spaces with 1-unconditional bases (Tikhomirov, 2017)
• An isomorphic Dvoretzky-Milman Theorem using general random ensembles (Mendelson, 2020)
• Projections of probability distributions: A measure-theoretic Dvoretzky theorem (Meckes, 2011)
• Dvoretzky type theorems for subgaussian coordinate projections (Mendelson, 2014)
• Hastings' additivity counterexample via Dvoretzky's theorem (Aubrun et al., 2010)
• Two infinite versions of nonlinear Dvoretzky's theorem (Funano, 2011)