arxiv: 2605.12568 · v1 · submitted 2026-05-12 · 🧮 math.ST · math.PR· stat.ML· stat.TH

Recognition: 2 theorem links

· Lean Theorem

Non-asymptotic quantisation of spherically symmetric distributions

Luc Pronzato , Anatoly Zhigljavsky

Authors on Pith no claims yet

Pith reviewed 2026-05-14 20:47 UTC · model grok-4.3

classification 🧮 math.ST math.PRstat.MLstat.TH

keywords spherically symmetric distributionsquantizationrandom quantizersexpected distortionextreme value theoryhigh-dimensional quantizationnon-asymptotic analysis

0 comments

The pith

For spherically symmetric distributions, random quantizers placed uniformly on a sphere of optimal radius achieve low expected distortion even with moderate numbers of points in high dimensions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Zador's theorem describes how optimal quantizers behave for very large n, but in high dimensions this regime requires impractically enormous samples. The paper shows that for spherically symmetric distributions, drawing points uniformly at random on a sphere of suitable radius r produces surprisingly good performance at moderate n. The expected L_s distortion reduces to a triple integral that can be evaluated numerically to arbitrary precision, and the best r can be found by direct optimization. When n grows with dimension d, extreme-value theory supplies simple approximations for the optimal r, which may tend to zero or to a positive constant depending on the growth rate.

Core claim

For spherically symmetric distributions, random quantisers uniformly distributed on a sphere of suitable radius r achieve exceptional performance. The expected distortion is a triple integral computable with arbitrary precision, and the optimal r can be determined numerically. Leveraging results from extreme-value theory, approximations for r are derived when n scales with d; depending on the growth rate, r may converge to zero or approach a limiting value independent of s.

What carries the argument

A random n-point quantizer whose points are drawn uniformly from the sphere of radius r centered at the origin, with expected distortion expressed as a triple integral over the radial distribution, the angular measure, and the point locations.

If this is right

The distortion value can be obtained to any desired accuracy by numerical quadrature without Monte Carlo simulation.
The optimal radius r is found by a low-dimensional numerical search for any fixed distribution, n, d, and s.
When n grows linearly or logarithmically with d, explicit limiting formulas for r become available from extreme-value theory.
The same construction yields explicit non-asymptotic bounds that remain useful long before the super-exponential n required by classical asymptotics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach may extend directly to other rotationally invariant error measures or to mixtures of spherically symmetric components.
In practice, one could initialize a Lloyd iteration with these sphere points to reach even lower distortion while retaining the computational simplicity of the initial placement.
The triple-integral representation could serve as a benchmark for testing new quantization algorithms on high-dimensional symmetric data.

Load-bearing premise

The target distribution must be spherically symmetric, and the claimed advantage is asserted only for sample sizes n that have not yet reached the asymptotic regime where Zador's theorem applies.

What would settle it

For a standard multivariate Gaussian in dimension d=20 with n=500, compute the triple-integral distortion for the numerically optimized r and compare it against the distortion of an n-point k-means quantizer trained on a large sample; if the sphere random quantizer does not show lower or equal error, the performance claim does not hold.

Figures

Figures reproduced from arXiv: 2605.12568 by Anatoly Zhigljavsky, Luc Pronzato.

**Figure 1.** Figure 1: Sphere. Value a ∗ minimising Dµ,s(P [n] a ) (left column), D 1/s µ,s (P [n] a∗ ) (central column) and ratio D 1/s µ,s (P [n] a∗ )/D1/s µ,s (P [n] 1 ) (right column) as functions of d for s = 1 (top row) and s = 10 (bottom row) and different values of n: n = 10 (•), n = 102 (♦), n = 103 (▼), n = 104 (⋆), and n = 105 (■). The next figure ( [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

**Figure 2.** Figure 2: Left: value α ∗ minimising Dµ,s(Q [n] α,1 ) as function of n. Center: D 1/s µ,s (Q [n] α∗,a) (- - -) and D 1/s µ,s (Q [n] 1,a) (—) as functions of a for n = 20 and α ∗ = 0.846. Right: D 1/s µ,s (Q [n] α∗,1 )/D1/s µ,s (P [n] 1 ) (•) and D 1/s µ,s (Q [n] α∗,1 )/D1/s µ,s (P [n] a∗ ) (♦) as functions of n; s = 10, d = 10. 3.2 Comparison between random quantisers uniform on a sphere and optimised full factorial… view at source ↗

**Figure 3.** Figure 3: compares the optimal full factorial designs Xn(b ∗ s ) for s = 2 and 4 with optimised random quantisers P [n] a∗ on Sd−1(1) with the same sample size n = 2d . The left panel shows the values of the radii of the various spheres on which the points lie, as functions of d. The right panel presents D 1/s µ,s (P [n] a∗ ) for the random quantisers and Eµ,s[Xn(b ∗ s )] for the full factorial designs, for s = 2 an… view at source ↗

**Figure 4.** Figure 4: presents the same information as [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: n ∗ (d, s) as a function of d for s = 2 (♦), s = 4 (⋆) and s = 10 (•). Left: µ is uniform on Sd−1(1) and n ∗ (d, s) is given by (15). Right: µ is spherically normal and n ∗ (d, s) is given by (17). 3.4 µ is a spherically symmetric normal distribution Any spherically symmetric normal distribution in Rd can be renormalised as N (0d, Id/d), with Id the d-dimensional identity matrix. In this section, we assume… view at source ↗

**Figure 6.** Figure 6: illustrates that σ ∗ = σ ∗ (d, n, s) increases with s, similarly to σ ∗ ∞, and that a ∗ = a ∗ (d, n, s) also increases with s, contrary to Sections 3.1 and 3.3. The reason is that large values of r = ∥U∥ get increasing importance as s increases, while r = 1 in Section 3.1 and r ≤ 1 in Section 3.3. Note that σ ∗ (d, n, s) is significantly smaller than σ ∗ ∞ = 1 + s/d ( [PITH_FULL_IMAGE:figures/full_fig_p01… view at source ↗

**Figure 7.** Figure 7: D 1/s µ,s (P [n] a=σ) (- - -) and D 1/s µ,s (µ [n] σ ) (—) as functions of σ for different d with n = 1 000 and s = 2 [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

**Figure 8.** Figure 8: Left: Optimal values a ∗ (d, n, s) (- - -) and σ ∗ (d, n, s) (—) as functions of d. Right: ratios D 1/s µ,s (P [n] a∗ )/D1/s µ,s (µ [n] σ∗∞ ) (- - -) and D 1/s µ,s (µ [n] σ∗ )/D1/s µ,s (µ [n] σ∗∞ ) (—) as functions of d for s = 2 and n = 10 000. In view of the right panel of [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗

**Figure 8.** Figure 8: The expressions for ab∗ = ab∗(d, n, 4) and Eb4 µ,4 (n; ab∗) can also be written in analytic form (by finding the roots of a third-degree polynomial), but the expressions are cumbersome. 4.2 Three asymptotic regimes for n growing with d In view of Lemma 4.1, the asymptotic behaviour of κn,d is the key element in understanding that of d 2 (u, Rn). For d = 3, β1,1 is uniform on [0, 1] and we simply have κn,3… view at source ↗

**Figure 9.** Figure 9: Values of κn,d (left) and of ab∗(d, n, 2) − a ∗ (d, n, 2) (right) as functions of d for different values of n: n = 102 (♦), n = 103 (▼), n = 104 (⋆) and n = 105 (■). The left and right panels of [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 10.** Figure 10: Relative error 1 − Db1/s µ,s (P [n] a )/D1/s µ,s (P [n] a ) of the approximation of D 1/s µ,s (P [n] a ) (left for s = 2 and right for s = 4) based on extreme-value theory, as a function of d for different values of n: n = 102 (♦), n = 103 (▼), n = 104 (⋆) and n = 105 (■); a is the optimal value a ∗ (d, n, s) [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

**Figure 11.** Figure 11: n 1/d Db1/s µ,s (P [n] a ) (♦) and n 1/d D 1/s µ,s (P [n] a ) (•) as functions of n for s = 4 and d = 20. Left: a = 0.75 for top curves and a = a ∗ (d, n, s) for bottom curves. Right: a = a ∗ (d, n, s), the values of n 1/d Eµ,s(Rn) computed numerically for 100 random quantisers Rn d∼ P [n] a∗ are shown as magenta dots. the limiting behaviour indicated in Proposition 4.1. Therefore, limd→∞ ab∗(d, n, s) = l… view at source ↗

**Figure 12.** Figure 12: displays the following quantities as functions of d for two values of n: a ∗ (d, n, 2), its approximation ab∗(d, n, 2), and b ∗ (d, n, 2), which minimises Dµ,2(P [n] 0,b) (see Section 3.3). The comparative behaviour of a ∗ (d, n, 2) and b ∗ (d, n, 2) has already been discussed in Section 3.3; the plots of a ∗ (d, n, 2) and ab∗(d, n, 2) are practically indistinguishable [PITH_FULL_IMAGE:figures/full_fig… view at source ↗

**Figure 13.** Figure 13: displays the efficiencies D 1/s µ,s (P [n] a )/D1/s µ,s (P [n] 0,b∗ ) as functions of d for a = a ∗ (d, n, 2) and a = ab∗(d, n, 2). These efficiencies are quasi indistinguishable, and P [n] a∗ and P [n] ac∗ both outperform P [n] 0,b∗ for the values of n and d considered. Additionally, P [n] 0,b∗ significantly outperforms P [n] 0,1 ; see [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗

**Figure 14.** Figure 14: (left) displays σ ∗ (d, n, s), a ∗ (d, n, s), ae∗(d, n, s) and Mψ,1ae∗(d, n, s) as functions of d for s = 4, together with ac∗ λ (indicated by an horizontal line) and ab∗(d, n, s). The right panel in the same figure shows the efficiency of the optimal normal distribution µσ∗ compared to Pa uniform on Sd−1(a), that is, D 1/s µ,s (P [n] a )/D1/s µ,s (µ [n] σ∗ ), for the five choices of a considered: Pa perf… view at source ↗

read the original abstract

Zador's celebrated theorem is a cornerstone of optimal quantisation, establishing both the weak limit of the empirical distribution of an $n$-point optimal quantiser in $R^d$ and the decay rate of the associated $L_s$-mean quantisation error. However, for large dimensions $d$, observing this asymptotic behaviour demands an astronomically large sample size $n$, which grows super-exponentially with $d$. Through a detailed analysis of the quantisation problem for spherically symmetric distributions, we demonstrate that for moderate $n$ random quantisers uniformly distributed on a sphere of suitable radius $r$ achieve exceptional performance. The expected distortion, expressed as a triple integral, can be computed with arbitrary precision, and the optimal radius $r$ can be efficiently determined numerically. Leveraging results from extreme-value theory, we derive approximations for $r$, particularly in scenarios where $n$ scales with $d$. Depending on the growth rate of $n$, $r$ may either converge to zero or approach a limiting value that is independent of $s$.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

For spherically symmetric distributions this gives a clean reduction of finite-n distortion to a triple integral plus EVT approximations for the radius, but the performance claim sits unanchored without comparisons.

read the letter

The main takeaway is that for spherically symmetric distributions you can get usable non-asymptotic quantization by drawing points uniformly on a sphere whose radius r is chosen to minimize a triple integral for the expected L_s distortion. The integral is over the radial density, the sphere, and the per-point error, and it can be evaluated to high precision while r itself is found numerically. Extreme-value approximations then handle the case where n grows with dimension d, showing when r shrinks to zero or settles to a positive limit independent of s. This is a direct, practical step past Zador's asymptotic result, which needs impractically large n in high d. The reduction itself is clean and the numerical route is straightforward, so the work supplies something concrete that people can actually compute for moderate n. The soft spot is exactly the one the stress-test flags: no head-to-head numbers against sampling the n points straight from the target law, against lattices, or against any finite-n lower bound derived from the same radial measure. Without those anchors the word 'exceptional' remains an assertion rather than a measured claim, and it is possible the main advantage is analytic tractability rather than actual superiority. The paper is aimed at people who need explicit finite-n quantization rules in high-dimensional statistics or ML. It shows clear, grounded thinking and makes honest use of existing extreme-value tools, so it is worth sending to a serious referee even if the revision letter will ask for some benchmark comparisons.

Referee Report

2 major / 2 minor

Summary. The manuscript analyzes non-asymptotic quantization for spherically symmetric distributions in high dimensions. It shows that random quantizers drawn uniformly from a sphere of suitably chosen radius r achieve low expected L_s distortion for moderate n (where Zador asymptotics have not yet set in). The expected distortion reduces to a triple integral over the radial density, the sphere, and the per-point quantization error; this integral is claimed to be evaluable to arbitrary precision, the optimal r is found numerically, and extreme-value approximations for r are derived when n grows with d (with r either tending to 0 or to an s-independent limit depending on the growth rate).

Significance. If the derivations hold, the work supplies an explicit, numerically tractable expression for the finite-n distortion of a concrete construction that is easy to sample, together with extreme-value approximations that become useful precisely when d is large and n is only moderate. This is a genuine contribution to the non-asymptotic regime, where standard asymptotic results are inapplicable. The reduction to a triple integral and the parameter-free character of the extreme-value limits (once the radial law is fixed) are clear strengths.

major comments (2)

[Abstract and §1] Abstract and §1: the repeated claim of 'exceptional performance' is not anchored by any quantitative comparison. No tables or figures compare the spherical construction to (i) n i.i.d. draws from the target distribution, (ii) any lattice or deterministic quantizer, or (iii) a finite-n lower bound obtained from the same radial measure. Without such anchors the adjective 'exceptional' remains unsupported even if the triple-integral formula is correct.
[§3] §3 (or wherever the triple-integral expression appears): the reduction of the expected distortion to the triple integral is presented without an explicit statement of the measurability or integrability conditions on the radial density that guarantee the integral is finite for all finite n. A short lemma establishing this would remove any doubt about the domain of applicability.

minor comments (2)

Notation for the radial density and the sphere radius r is introduced without a consolidated table of symbols; a short notation table would improve readability.
[final section] The extreme-value approximations in the final section are stated for several growth regimes of n(d); a single corollary collecting the limiting expressions for r would make the results easier to cite.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive overall assessment and for the detailed, constructive comments. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation.

read point-by-point responses

Referee: [Abstract and §1] Abstract and §1: the repeated claim of 'exceptional performance' is not anchored by any quantitative comparison. No tables or figures compare the spherical construction to (i) n i.i.d. draws from the target distribution, (ii) any lattice or deterministic quantizer, or (iii) a finite-n lower bound obtained from the same radial measure. Without such anchors the adjective 'exceptional' remains unsupported even if the triple-integral formula is correct.

Authors: We agree that the adjective 'exceptional' would be better supported by explicit quantitative comparisons. The manuscript's primary contribution is the closed-form triple-integral expression for the expected distortion together with the extreme-value approximations for the optimal radius; these already allow direct numerical evaluation for any fixed radial law. Nevertheless, to address the referee's point we will add, in the revised version, a new subsection (or appendix) containing a table and accompanying figure that compare the spherical quantizer (with numerically optimized r) against (i) n i.i.d. samples from the same distribution and (ii) the best available finite-n lower bound derived from the radial measure, for representative cases (standard Gaussian and uniform on the unit ball) and moderate n,d pairs. This will anchor the performance claim without altering the theoretical focus of the paper. revision: yes
Referee: [§3] §3 (or wherever the triple-integral expression appears): the reduction of the expected distortion to the triple integral is presented without an explicit statement of the measurability or integrability conditions on the radial density that guarantee the integral is finite for all finite n. A short lemma establishing this would remove any doubt about the domain of applicability.

Authors: We thank the referee for this observation. While the finiteness follows immediately from the fact that the radial density integrates to one and the per-point quantization error is at most polynomial in the radius (combined with the finite-s-moment assumption implicit in the L_s distortion), we agree that an explicit statement removes any ambiguity. In the revision we will insert a short lemma (new Lemma 3.1) immediately preceding the triple-integral formula. The lemma will state that if the radial density ρ is measurable, non-negative, integrates to one against the surface measure on R_+, and the random vector satisfies E[||X||^s] < ∞, then the triple integral is finite for every finite n. This will precisely delineate the domain of applicability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper reduces the expected L_s distortion of its proposed random spherical quantizer construction to an explicit triple integral over the radial density, sphere surface measure, and per-point quantization error. This integral is evaluated directly and optimized numerically over r; large-n/d approximations invoke standard external extreme-value theory. No load-bearing step equates a claimed performance quantity to a fitted parameter, self-citation, or input by construction. The derivation remains self-contained once spherical symmetry is granted.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The analysis rests on the domain assumption of spherical symmetry and standard results from extreme value theory; no free parameters are fitted to data and no new entities are postulated.

axioms (2)

domain assumption The target distribution is spherically symmetric
Invoked to reduce the quantization problem to a radial integral over a sphere of radius r.
standard math Standard results from extreme value theory apply to the radius selection
Used to derive approximations for optimal r when n grows with d.

pith-pipeline@v0.9.0 · 5487 in / 1312 out tokens · 43393 ms · 2026-05-14T20:47:50.672722+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

for spherically symmetric distributions... random quantisers uniformly distributed on a sphere

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

13 extracted references

[1]

Alon and J.H

N. Alon and J.H. Spencer.The Probabilistic Method. Wiley, 2000. Second edition

2000
[2]

K.-T. Fang, S. Kotz, and K.W. Ng.Symmetric Multivariate and Related Dis- tributions. Chapman and Hall/CRC, 1990

1990
[3]

Feller.An Introduction to Probability Theory and Its Applications, vol

W. Feller.An Introduction to Probability Theory and Its Applications, vol. 2. Wiley, New York, 1971. Second edition

1971
[4]

Graf and H

S. Graf and H. Luschgy.Foundations of Quantization for Probability Distribu- tions. Springer, Berlin, 2000

2000
[5]

Kotz and S

S. Kotz and S. Nadarajah.Extreme Value Distributions: Theory and Applica- tions. Imperial College Press, 2000

2000
[6]

Luschgy and G

H. Luschgy and G. Pag` es. Greedy vector quantization.Journal of Approxima- tion Theory, 198:111–131, 2015

2015
[7]

Nogales G´ omez, L

A. Nogales G´ omez, L. Pronzato, and M.-J. Rendas. Incremental space-filling design based on coverings and spacings: improving upon low discrepancy se- quences.Journal of Statistical Theory and Practice, 15(4):77, 2021

2021
[8]

Noonan and A

J. Noonan and A. Zhigljavsky.High-Dimensional Optimization: Set Explo- ration in the Non-Asymptotic Regime. Springer, 2024

2024
[9]

G. Pag` es. A space quantization method for numerical integration.Journal of Computational and Applied Mathematics, 89(1):1–38, 1997

1997
[10]

Pronzato and W

L. Pronzato and W. M¨ uller. Design of computer experiments: space filling and beyond.Statistics and Computing, 22:681–701, 2012

2012
[11]

Pronzato and A.A

L. Pronzato and A.A. Zhigljavsky. Quasi-uniform designs with asymptotically optimal and near-optimal uniformity constant.Journal of Approximation The- ory, 294(105931), 2023. 23

2023
[12]

Vershynin.High-Dimensional Probability: An Introduction with Applica- tions in Data Science

R. Vershynin.High-Dimensional Probability: An Introduction with Applica- tions in Data Science. Cambridge University Press, 2018

2018
[13]

P.L. Zador. Asymptotic quantization error of continuous signals and the quan- tization dimension.IEEE Trans. Inform. Theory, 28:139–149, 1982. 24

1982