betaFeedback

WorksMathematics

Calibrating the Brody exponent as a quantitative measure of short-range exclusion in 2D spatial point processes

Abstract

The Brody distribution, originally a phenomenological interpolation between Poisson and Wigner level-spacing statistics in quantum chaos, is calibrated here as a quantitative measure of short-range exclusion in 2D spatial point processes. Two results form the core. First, the 2D complete-spatial-randomness baseline is recalibrated to β=0.96±0.15β=0.96\pm0.15, correcting the inappropriate 1D Poisson reference. Second, an empirical ββ–rexclr_{\text{excl}} calibration is validated against the effective hard-core radius with Spearman ρ=0.988ρ=0.988. The framework is demonstrated on 58 manufactured surfaces (10 materials, 10 processes), phase-extracted interferometric profilometry of a certified roundness standard, and 2D binary embeddings of prime numbers. A sparse-integer control proves the prime β=2.15β=2.15 signal is genuinely arithmetic (Δβ=+0.68Δβ=+0.68 over random-integer control), while a Cantor-embedding null result (β=1.40β=1.40, TOST p<0.01p<0.01) demonstrates that 2D exclusion is embedding-created rather than intrinsic. Density-thinning experiments establish that ββ captures exclusion strength rather than point density, while absolute values are density-dependent. A distinct CSR baseline for binary fields at low fill fraction is identified, with a decision table provided. The ββ–rexclr_{\text{excl}} calibration, the CSR baseline correction, and the control protocols together constitute a calibrated measurement framework for reproducible characterisation of short-range exclusion in 2D spatial point processes.

This is a preprint submitted to a peer-reviewed journal. The most up-to-date version is available at arXiv:2606.16393.

Introduction

A practical question in the statistical characterisation of spatial point patterns is whether a single, interpretable parameter can quantify short-range exclusion strength on a continuous scale, analogous to how the hard-sphere packing fraction parameterises excluded volume in liquid-state theory. In quantum chaos, the Brody distribution Pβ(s)=(β+1)asβexp⁡(−asβ+1)P_{\beta}(s) = (\beta+ 1)as^{\beta}\exp(-as^{\beta+1}), a=[Γ((β+2)/(β+1))]β+1a = [\Gamma((\beta+ 2)/(\beta+ 1))]^{\beta+1}, was introduced as a phenomenological one-parameter interpolation between the Poisson (β=0\beta= 0, uncorrelated levels) and Wigner (β=1\beta= 1, GOE; β=2\beta= 2, GUE) limits of random matrix theory (RMT) [1]. Its single parameter β\beta quantifies the degree of level repulsion—the suppression of probability density at small spacings—on a continuous scale. In the present work, β\beta is employed as a phenomenological exclusion parameter rather than a symmetry-class indicator; values exceeding the GUE limit (β>2\beta> 2) are interpreted as progressively stronger short-range exclusion (up to β≈10\beta\approx10 for near-crystalline hard-core processes). Values of β>2\beta> 2 are not interpreted as evidence of higher-order random-matrix symmetry classes; they are used solely as a phenomenological measure of exclusion strength on a continuous scale. This distinction is essential: the Brody distribution is used here as an empirical descriptor of spatial order, not as a claim about underlying matrix symmetries. Beyond quantum spectra, the Brody distribution has been applied to quantum chaos [2], quantum dots [3], acoustic cavities [4], and financial correlations [5], establishing it as a general-purpose tool for characterising spacing statistics in 1D sequences.

In a celebrated development, Montgomery [6] conjectured and Odlyzko [7] numerically confirmed that the pair correlation function of the non-trivial zeros of the Riemann zeta function coincides with the GUE prediction—connecting prime number theory to RMT universality. Berry [8] first proposed the Riemann zeta function as a model for quantum chaos; Bogomolny and Keating [9] developed the semiclassical framework beyond the diagonal approximation, and Keating and Snaith [10] established the connection to random-matrix moments. This raises a practical question: can the Brody parameter, when properly calibrated for two dimensions, serve as a transferable measure of short-range exclusion strength across different classes of 2D point processes?

The gap. Despite the breadth of Brody applications to 1D spectra, its extension to 2D spatial point processes raises two unresolved issues. First, the appropriate CSR baseline for 2D Brody statistics has not been established: the 2D nearest-neighbour spacing distribution under CSR is Rayleigh, P(s)=(πs/2)exp⁡(−πs2/4)P(s) = (\pi s/2)\exp(-\pi s^{2}/4), which is not a member of the Brody family, so the 1D Poisson reference β=0\beta= 0 does not apply. Second, no calibration exists to translate the dimensionless β\beta into a physically interpretable exclusion measure. Sakhr and Nieminen [11] demonstrated a Poisson-to-Wigner crossover in the nearest-neighbour statistics of random points on fractals, establishing that RMT-style spacing distributions can characterise spatial point patterns outside quantum mechanics. The extensive literature on spatial point-process statistics provides well-established tools—Ripley’s KK, pair correlation functions, and Voronoi tessellation—documented in the monographs of Illian et al. [12], Baddeley et al. [13], and Chiu et al. [14], with specialised treatments for Cox-process inference by Møller and Waagepetersen [15]; but these methods do not directly yield a single continuous parameter quantifying short-range exclusion strength. Recent work on spatial form factors for Coulomb gas and vortex statistics [16] and Fourier analysis of spatial point processes [17] has begun bridging spectral and spatial approaches, but a framework connecting these to the Brody one-parameter family is lacking.

A manufactured surface contains a spatial point process formed by its peaks—features arising from cutting tools, abrasive grains, plastic deformation, and fracture mechanics [18]. Similarly, binary embeddings of arithmetic sequences produce 2D point patterns governed by multiplicative constraints rather than physical interactions. Phase-extracted interferometric profiles [19] introduce yet a third generative domain: optical wavefront reconstruction. Whether the Brody exponent can be calibrated as a transferable exclusion measure across these domains—and what the necessary corrections (CSR baseline, density matching, embedding specification) are—has not been systematically investigated.

This work. This study develops and validates a calibrated measurement framework for Brody statistics in 2D spatial point processes. The framework comprises three components: (i) the correct 2D CSR baseline (β=0.96±0.15\beta= 0.96 \pm0.15, density-stable), (ii) an empirical βexcl/⟨NN⟩\beta_{\rm excl}/\langle\mathrm{NN}\rangle calibration curve derived from synthetic hard-core and soft-core point processes, and (iii) a set of control protocols (density-thinning, embedding-robustness, spatial-permutation, and peak-detection sensitivity tests) that isolate genuine exclusion from confounding effects. The framework is demonstrated on three data classes: 58 manufactured surface topographies (10 materials, 10 processes, focus-variation microscopy); phase-extracted interferometric profilometry of a certified roundness standard (JENOPTIK FN 111, 24 000 frames, 5 continuous 400° revolutions); and 2D binary embeddings of prime numbers as a mathematically controlled benchmark. The aim is not to claim that β\beta reveals a universal exclusion mechanism—any process that prevents point coincidence (material continuity, optical resolution limits, arithmetic gap constraints) will produce β>βCSR\beta> \beta_{\mathrm{CSR}}—but to establish the measurement tools and correction protocols required for quantitative, reproducible Brody analysis in two dimensions.

Observation versus interpretation. Throughout this work, a clear distinction is maintained between what is measured and what is inferred. Measured: β\beta values, nearest-neighbour spacing distributions, g2(r)g_{2}(r) profiles, D2D_{2} excesses, and their associated confidence intervals. Inferred: exclusion strength, effective hard-core radii, and spatial ordering. The β\beta–rexclr_{\mathrm{excl}} relation is calibration-based rather than derived from first principles; it is an empirical mapping established through synthetic point-process experiments, not a theoretical necessity. Density dependence and embedding sensitivity demonstrate that absolute β\beta values reflect both the exclusion mechanism and the measurement conditions (point density, embedding geometry).

Throughout, the Brody distribution is employed as a phenomenological fitting function; no claim of random-matrix universality is made for surface peaks. The three domains—manufacturing processes, multiplicative number theory, and optical interferometry—all yield β\beta values exceeding their density-matched CSR baselines. This is expected for any process with short-range repulsion, and the contribution of the present work is not the observation of exclusion but the provision of the calibrated tools to quantify it: the 2D CSR baseline, the β\beta–rexclr_{\mathrm{excl}} mapping, and the control protocols that separate genuine structure from density, embedding, and detection artefacts.

Methods

The unified framework rests on three methodological pillars: (i) Brody β\beta estimation from 2D nearest-neighbour spacings, applicable to any spatial point process (§ 2.1–§ 2.3); (ii) synthetic point-process generation for exclusion-physics calibration (§ 2.4); and (iii) complementary D2D_{2} correlation dimension and Banach-space descriptors for binary patterns where peak detection is inapplicable. Full details of the D2D_{2} box-counting procedure, multifractal analysis, singularity spectra, and Banach norms are provided in Supplementary Information (SI, Sections S1.1–S1.7); the key methodological choices are summarised below. All three pillars share a common null-model philosophy: structured data are compared against matched-density Bernoulli or CSR ensembles to isolate genuine spatial order from trivial density effects.

Arithmetic surface construction

Let p1<p2<⋯<pπ(N)p_{1} < p_{2} < \cdots< p_{\pi(N)} be the primes not exceeding NN. The binary prime surface of side W=⌊N⌋W = \lfloor\sqrt{N}\rfloor is the Boolean matrix

Zij={1if i⋅W+j+1 is prime,0otherwise,(1)Z_{ij} = \begin{cases} 1 & \text{if } i \cdot W + j + 1 \text{ is prime},\\ 0 & \text{otherwise}, \end{cases} \tag*{(1)}

for 0≤i,j<W0 \le i,j < W. At N=105N = 10^{5}, W=316W = 316 and the surface contains 9592 non-zero entries (density ρ=0.0959\rho= 0.0959). For the scale-up analysis, surfaces are constructed at N∈{103,2×103,5×103,104,2×104,5×104,105,2×105,5×105,106,5×106,107}N \in\{10^{3}, 2 \times10^{3}, 5 \times10^{3}, 10^{4}, 2 \times10^{4}, 5 \times10^{4}, 10^{5}, 2 \times10^{5}, 5 \times10^{5}, 10^{6}, 5 \times10^{6}, 10^{7}\}.

Five additional arithmetic sequences are constructed analogously for discrimination testing:

  1. Square-free numbers: nn is square-free iff μ(n)≠0\mu(n) \ne0, where μ\mu is the Möbius function. Density ρ→6/π2≈0.6079\rho\to6/\pi^{2} \approx0.6079 [20].

  1. Sums of two squares: n=a2+b2n = a^{2} + b^{2} for integers a,ba,b. Density ∼K/log⁡n\sim K/\sqrt{\log n} (Landau–Ramanujan constant K≈0.7642K \approx0.7642).

  1. Twin primes (upper): pp such that p−2p - 2 is also prime.

  1. Primes ≡1(mod4)\equiv1 \pmod{4} and primes ≡3(mod4)\equiv3 \pmod{4}: testing for Chebyshev bias.

Two-dimensional embeddings

Three deterministic embeddings are compared. The Ulam spiral is not introduced as a visualisation device but as a deterministic mapping known to amplify arithmetic correlations through quadratic residue structure [21]: integers falling on the same diagonal satisfy n=4k2+bk+cn = 4k^{2} + bk + c, concentrating primes on lines where bb and cc produce many quadratic residues modulo small primes. Row-major embedding preserves the natural integer ordering and is the simplest reshaping. Cantor pairing provides a diagonal sweep that tests whether the exclusion signal depends on the specific geometry of the embedding or on the arithmetic sequence alone.

  1. Row-major: n↦(⌊(n−1)/W⌋,(n−1) mod W)n \mapsto(\lfloor(n - 1)/W\rfloor,(n - 1) \bmod W), the standard array reshaping.

  1. Ulam spiral: integers arranged in an expanding square spiral centred at the middle of the array. Position (x,y)(x,y) for integer nn follows the spiral algorithm of [21].

  1. Cantor pairing: π(i,j)=(i+j)(i+j+1)/2+j\pi(i,j) = (i + j)(i + j + 1)/2 + j, a diagonal sweep of the matrix from top-left to bottom-right.

For the surrogate embedding test, 30 random permutations of the 1D→2D mapping are generated by shuffling the linear index before reshaping. This destroys any structure specific to the embedding while preserving the 1D sequence statistics.

Null-model ensembles

Five types of null-model surfaces are generated, each with Nens=20N_{\mathrm{ens}} = 20 independent realisations at the same shape as the arithmetic surface:

  1. Gaussian: Zij∼N(0,1)Z_{ij} \sim\mathcal{N}(0,1) i.i.d., maximal entropy, zero spatial correlation.

  1. Poisson: Zij∼Poisson⁡(λ=0.01)Z_{ij} \sim\operatorname{Poisson}(\lambda= 0.01), sparse spike process.

  1. 1/f1/f noise: PSD ∝∣k∣−1\propto|k|^{-1}, constructed via frequency-domain filtering of white noise, yielding long-range power-law correlations.

  1. Brownian: cumulative sum of white noise along both axes, PSD ∝∣k∣−2\propto|k|^{-2}, non-stationary.

  1. Bernoulli: Zij∈{0,1}Z_{ij} \in\{0,1\} with P(Zij=1)=ρP(Z_{ij} = 1) = \rho, the exact Cramér-type null for binary arithmetic surfaces at matched density.

The Bernoulli null at matched density is the primary comparison, as it isolates the effect of arithmetic structure from the trivial effect of sparsity. The other four null models provide additional reference points for the metrological parameter space.

Statistical power and multiple-comparison correction

With m=14m = 14 sequence types tested against the Bernoulli null, the family-wise error rate requires correction. The Benjamini–Hochberg procedure is applied to control the false discovery rate at α=0.05\alpha= 0.05. A power analysis establishes the minimum detectable effect size: for the average null standard deviation σnull=0.0034\sigma_{\rm null} = 0.0034 (estimated from 50 independent Bernoulli realisations at N=105N = 10^5), the Bonferroni-corrected critical value is zα/2m=2.91z_{\alpha/2m} = 2.91, yielding a minimum detectable ∣ΔD2∣=0.013|\Delta D_2| = 0.013 at 80% power (z0.80=0.84z_{0.80} = 0.84). The previously used heuristic threshold ∣ΔD2∣>0.003|\Delta D_2| > 0.003 corresponds to only ∼40%\sim40\% power and is replaced throughout by the corrected threshold of 0.013 for claims of statistical significance. Results exceeding ∣ΔD2∣=0.003|\Delta D_2| = 0.003 but below 0.013 are reported as “nominally above the heuristic threshold” and are not interpreted as statistically significant without confirmation at larger NN.

Choice of D2D_2 as spatial descriptor

The correlation dimension D2D_2 is selected in preference to standard spatial-statistical tools for the following reasons. Moran’s II and Ripley’s KK function detect departures from CSR but do not provide a single scale-invariant parameter comparable across different densities and sequence lengths. The pair correlation function g2(r)g_2(r) is the natural RMT analogue but is sensitive to binning and requires ensemble averaging for stable estimation on discrete binary fields. Nearest-neighbour statistics require peak detection, which is not applicable to binary patterns. D2D_2 offers three advantages: (i) it is a single number on a fixed scale (D2∈[0,2]D_2 \in[0, 2] for 2D data); (ii) it is estimated via box-counting regression, which is robust to discretisation; and (iii) it has a direct interpretation as the scaling exponent of the probability that two randomly chosen occupied sites belong to the same box, making it sensitive to spatial clustering and anti-clustering at all length scales simultaneously.

For a surface ZZ of size H×WH \times W with total mass M=∑i,jZijM = \sum_{i,j} Z_{ij}, a grid of boxes of linear size ε\varepsilon is overlaid. The partition function (generalised statistical sum) of order qq is

Zq(ε)=∑k=1Nbox(ε)(Mk(ε)M)q,(2)Z_q(\varepsilon) = \sum_{k=1}^{N_{\mathrm{box}}(\varepsilon)} \left(\frac{M_k(\varepsilon)}{M}\right)^q, \tag*{(2)}

where Mk(ε)M_k(\varepsilon) is the mass in box kk and Nbox(ε)≈HW/ε2N_{\mathrm{box}}(\varepsilon) \approx HW/\varepsilon^2. The generalised Rényi dimension of order qq is [22, 23]

Dq=1q−1lim⁡ε→0log⁡Zq(ε)log⁡ε,(q≠1),(3)D_q = \frac{1}{q-1}\lim_{\varepsilon\to0}\frac{\log Z_q(\varepsilon)}{\log\varepsilon}, \qquad(q \ne1), \tag*{(3)}

with the information dimension D1=lim⁡q→1DqD_1 = \lim_{q \to1} D_q obtained by L’Hôpital’s rule: D1=lim⁡ε→0∑kμklog⁡μk/log⁡εD_1 = \lim_{\varepsilon\to0} \sum_k \mu_k \log\mu_k / \log\varepsilon, where μk=Mk/M\mu_k = M_k/M.

The correlation dimension D2D_2 is the special case q=2q = 2 and has the intuitive interpretation [24]: C(ε)∝εD2C(\varepsilon) \propto\varepsilon^{D_2}, where C(ε)C(\varepsilon) is the probability that two randomly chosen units of surface “mass” lie within the same box of size ε\varepsilon. For a uniform distribution in a dd-dimensional embedding space, Dq=dD_q = d for all qq (monofractal). Deviations of D2D_2 from the embedding dimension d=2d = 2 indicate fractal heterogeneity in the distribution of mass.

In practice, DqD_q is estimated by linear regression of log⁡Zq(ε)\log Z_q(\varepsilon) versus log⁡ε\log\varepsilon over box sizes ε∈{2,4,8,…,2m}\varepsilon\in\{2, 4, 8, \ldots, 2^m\} where m=⌊log⁡2min⁡(H,W)⌋−1m = \lfloor\log_2 \min(H, W) \rfloor- 1. The quality of the linear fit is monitored via the adjusted R2R^2 statistic; all reported fits satisfy R2>0.95R^2 > 0.95.

Singularity spectrum

The Legendre transform of DqD_q yields the Hölder exponent spectrum α(q)=dτ/dq\alpha(q) = d\tau/dq and the singularity spectrum f(α)=qα−τ(q)f(\alpha) = q\alpha- \tau(q), where τ(q)=(q−1)Dq\tau(q) = (q - 1)D_q [23]. The width Δα=αmax⁡−αmin⁡\Delta\alpha= \alpha_{\max} - \alpha_{\min} quantifies the degree of multifractality: Δα=0\Delta\alpha= 0 for a monofractal (uniform scaling), and Δα>0\Delta\alpha> 0 for a multifractal (spatially heterogeneous scaling). A narrower f(α)f(\alpha) spectrum for the prime surface relative to the Bernoulli null would indicate that arithmetic constraints reduce the multifractal heterogeneity.

Validation procedures

Four validation tests are employed to ensure that the observed D2D_2 signal is genuine rather than a methodological artefact:

  1. Surrogate embedding test: The 1D-to-2D mapping is randomly permuted Nsurr=30N_{\mathrm{surr}} = 30 times. Under the null hypothesis that the D2D_2 signal is an artefact of the specific embedding geometry, the permuted D2D_2 distribution should overlap with the original value. A significant deviation (∣z∣>3|z| > 3) indicates that the signal is intrinsic to the 1D sequence, not the embedding.

  1. Bootstrap stability: The surface is randomly subsampled Nboot=50N_{\mathrm{boot}} = 50 times at half linear size (W/2×W/2W/2 \times W/2). The coefficient of variation CV=σ(D2)/μ(D2)\mathrm{CV} = \sigma(D_2)/\mu(D_2) quantifies the estimator stability.

  1. Controlled-correlation sequences: Three surrogate binary sequences are generated at matched density to establish the D2D_2 ordering: (a) block-clustered (consecutive blocks of b=5b = 5 ones), (b) repulsive (minimum distance dmin⁡=4d_{\min} = 4 between ones, plus random jitter), and (c) Poisson (independent, sparse). These span the range from strong clustering to strong anti-clustering, providing a calibrated scale for interpreting D2(primes)D_2(\mathrm{primes}).

  1. Negative control: Square-free numbers, which share the multiplicative structure of integers but lack the primality constraint, are tested. A non-significant D2D_2 for square-free numbers relative to Bernoulli would confirm that the method does not indiscriminately flag all arithmetic sequences as structured.

Finite-size bias analysis

The box-counting estimator of D2D_2 is known to exhibit finite-size bias [24, 25]: even for a purely random point set (theoretical D2=dD_2 = d), the estimated D^2(N)<d\widehat{D}_2(N) < d for finite NN. To quantify this effect, the box-counting procedure is applied to Bernoulli surfaces at the same sizes {N}\{N\} as the arithmetic surfaces. The bias function B(N)=d−D^2(Bern)(N)B(N) = d - \widehat{D}_2^{(\mathrm{Bern})}(N) is fitted, and the bias-cancelled excess correlation is defined as

ΔD2(N)=D2(prime)(N)−D2(Bern)(N),(4)\Delta D_2(N) = D_2^{(\mathrm{prime})}(N) - D_2^{(\mathrm{Bern})}(N), \tag*{(4)}

where both surfaces are embedded identically and at matched density. If the scaling of D2(prime)(N)D_2^{(\mathrm{prime})}(N) is dominated by estimator bias, then ΔD2(N)\Delta D_2(N) should be approximately constant.

Remark on density dependence. Density-thinning experiments were performed to test whether β\beta values are stable under changes in point density. The 2D prime embedding (N=105N = 10^5, ρ=0.096\rho= 0.096, β=2.15\beta= 2.15) was randomly thinned to match point counts corresponding to densities ρ∈{0.01,0.02,0.032,0.05,0.096}\rho\in\{0.01, 0.02, 0.032, 0.05, 0.096\} (20 trials each). At the PSI density (ρ=0.032\rho= 0.032), prime β\beta decreased to 1.46±0.021.46 \pm0.02 [1.42, 1.49], and at the CSR calibration density (ρ=0.02\rho= 0.02) to 1.30±0.051.30 \pm0.05.

Critically, β\beta at all densities remained significantly above the density-matched CSR baseline (βCSR=0.98±0.05\beta_{\mathrm{CSR}} = 0.98 \pm0.05 at ρ=0.032\rho= 0.032; 0.97±0.040.97 \pm0.04 at ρ=0.02\rho= 0.02), confirming genuine exclusion structure beyond density effects. The CSR baseline itself is density-stable across all tested densities (0.95–0.99). These results establish two points: (i) β\beta captures exclusion strength, not density; (ii) absolute β\beta values are density-dependent, and cross-domain comparisons require density matching.

Results

Roadmap. The Results demonstrate the calibrated measurement framework on three data classes and establish the correction protocols required for valid cross-domain comparison. § 3.1 presents β\beta for 58 manufactured surfaces. § 3.2 establishes the synthetic β\beta–rexclr_{\mathrm{excl}} calibration—the interpretive key for all data—together with robustness checks (§ 3.3). § 3.4 presents g2(r)g_2(r) as an internal consistency check. § 3.5 presents density-thinning controls, establishing that β\beta captures exclusion rather than density. § 3.6–§ 3.7 present 2D prime embeddings, with emphasis on embedding dependence and the sparse-integer control that isolates arithmetic structure from sparsity. § 3.8 summarises a proof-of-concept application to interferometric profilometry (full analysis in Supplementary Information). Complementary D2D_2 and Banach descriptors are developed in Supplementary Information (SI, Sections S1.1–S1.7) and summarised in one paragraph (§ 3.9).

Peak spacing statistics in manufactured surfaces

Fifty-eight areal surface texture maps obtained via focus-variation microscopy [18, 26] were analysed, spanning 8 engineering materials (1.4301 stainless steel, C45 carbon steel, Ti and Ti6Al4V titanium alloys, Al and Al7075 aluminium alloys, brass (MO58A and generic brass), ELLOR tool steel, and graphite) and 6 manufacturing processes (turning, grinding, bead blasting, honing, milling, and WEDM electrical discharge machining). Local maxima (peaks) were detected using a 2D prominence-based algorithm (a pixel is a peak if it is the strict maximum within a (2k+1)×(2k+1)(2k + 1) \times(2k + 1) neighbourhood, k=max⁡(3,⌊min⁡(nx,ny)/40⌋)k = \max(3, \lfloor\min(n_x,n_y)/40\rfloor), and its elevation above the local baseline—the mean of the same neighbourhood—exceeds 3% of the global height range).

Watershed peak detection for dense surfaces

For surfaces where the prominence-based detector produced grid-limited results (> 85% column occupancy), an alternative watershed-based detector was applied. The surface is inverted (Z→−ZZ \rightarrow-Z) and a maximum filter with kernel size k=max⁡(3,⌊min⁡(nx,ny)/30⌋)k = \max(3, \lfloor\min(n_x,n_y)/30\rfloor) is used to identify regional maxima. An hh-minima suppression of 3% of the height range eliminates shallow noise peaks, and a minimum inter-peak distance of 5 pixels prevents fragmentation of broad summits. Peaks are validated by a prominence threshold of 2% of the global height range. This detector suppresses the dense, low-amplitude roughness features that saturate the prominence-based detector in bead-blasted, ground, and WEDM surfaces, while preserving well-separated topographic peaks.

Surface filtering and exclusion criteria

Of 58 measured surfaces, 37 (64%) were classified as grid-limited (> 85% column occupancy) under the original 3% prominence, kernel-based peak detector. To recover these surfaces, a watershed-based detector with hh-minima suppression (h=0.03h = 0.03 of height range) and minimum inter-peak distance of 5 pixels was applied to all 37 grid-limited surfaces. All 37 produced valid β\beta estimates (range β=0.71\beta= 0.71–4.614.61, mean 2.89±1.102.89 \pm1.10), with column occupancy reduced to 0.020.02–0.250.25 (all below the 85% threshold). The watershed detector suppresses dense, low-amplitude noise peaks while preserving genuine topographic features, making it suitable for surfaces where the prominence-based detector saturates.

Combined with the original 21 valid surfaces, the full analysable dataset comprises all 58 surfaces spanning 10 materials and 10 processes. The qualitative conclusion—that burnishing and honing produce higher β\beta than turning and grinding—was unchanged under all threshold and detector combinations tested. (30–770 peaks per surface, 18–428 valid nearest-neighbour pairs after outlier rejection; outlier rejection used a median-absolute-deviation threshold of 5 MAD, retaining 85–98% of detected NN pairs across surfaces. A sensitivity analysis varying the threshold from 3 to 7 MAD changed β\beta estimates by < 0.2 for surfaces with n>50n > 50 pairs).

For each surface, the 2D Euclidean nearest-neighbour distances between peak positions were computed via KD-tree, normalised to unit mean (unfolded), and fitted to the Brody distribution Pβ(s)=(β+1)asβexp⁡(−asβ+1)P_\beta(s) = (\beta+ 1)as^\beta\exp(-as^{\beta+1}), where a=[Γ((β+2)/(β+1))]β+1a = [\Gamma((\beta+ 2)/(\beta+ 1))]^{\beta+1}. The Brody parameter β\beta was estimated by maximum likelihood with the constraint β≥0\beta\ge0 (the Brody distribution is defined for β>−1\beta> -1 and is physically motivated as an interpolation between the Poisson limit at β=0\beta= 0 and Wigner limits at β=1,2\beta= 1, 2; negative unconstrained estimates are reported as β=0\beta= 0, interpreted as “indistinguishable from Poisson”). The Brody distribution interpolates between the Poisson limit (β=0\beta= 0, P(s)=exp⁡(−s)P(s) = \exp(-s), uncorrelated levels), the GOE Wigner surmise (β=1\beta= 1, P(s)=(πs/2)exp⁡(−πs2/4)P(s) = (\pi s/2)\exp(-\pi s^2/4), linear level repulsion), and the GUE Wigner surmise (β=2\beta= 2, P(s)=(32s2/π2)exp⁡(−4s2/π)P(s) = (32s^2/\pi^2)\exp(-4s^2/\pi), quadratic repulsion). The Brody distribution is employed here in preference to standard spatial-statistical descriptors (Ripley’s K, pair correlation function) because its single parameter β\beta directly quantifies the degree of short-range order in the spacing distribution on a continuum from uncorrelated (β=0\beta= 0, Poisson limit) to ordered (β>1\beta> 1, approaching the Wigner limits at β=1,2\beta= 1, 2). The Poisson, GOE, and GUE limits serve as convenient reference points for interpreting β\beta values, not as claims that surface peaks obey random-matrix statistics. Goodness-of-fit against these limits is assessed via the Kolmogorov–Smirnov distance DKSD_{\mathrm{KS}}; a surface is classified as “closest to” the limit with the smallest DKSD_{\mathrm{KS}}, without implying statistical consistency with that limit.

Figure 1 and Table 1 present the fitted Brody parameters for all 58 analysable surfaces. The Brody exponent spans nearly the full theoretical range: from β=−0.20\beta= -0.20 (ground Ti; negative values are fitted but should be interpreted as indistinguishable from Poisson, β=0\beta= 0) to β=+4.61\beta= +4.61 (rough turned C45 steel, watershed). The full dataset (58 surfaces across 10 materials and 10 processes, observational design without full factorial crossing) reveals the following trends (descriptive only, owing to process–material confounding):

Figure 1. Brody exponent β\beta across all 58 analysable surfaces. Left: histogram with kernel density estimate. Right: β\beta by manufacturing process (box plots with individual points). The CSR baseline (β=0.96\beta= 0.96, grey band) is indicated. The earlier bar chart for 21 prominence-based surfaces (Figure S2) provides an alternative visualisation.

Figure S2. Correlation dimension D2D_2 scaling for six arithmetic sequences at N=105N = 10^5, compared with matched-density Bernoulli null models.

Table 1: Brody exponent β\beta for 58 manufactured surfaces ordered by decreasing β\beta. Watershed-recovered surfaces are marked with w. Fitting is constrained to β≥0\beta\ge0; unconstrained negative estimates are reported as β=0\beta= 0 (indistinguishable from Poisson). Bootstrap 95% confidence intervals (500 resamples) are reported for each surface. CIs narrower than ±0.3\pm0.3 indicate reliable estimates; CIs wider than ±1.5\pm1.5 (surfaces with nNNS<30n_{\mathrm{NNS}} < 30) are indicative only.

MaterialProc.npeaksn_{\mathrm{peaks}}nNNSn_{\mathrm{NNS}}β\beta95% CI
MO58A brassBn41414.30[3.40, 5.76]
1.4301 steelHo49494.00[3.22, 5.31]
Al7075Mi44433.70[2.88, 5.08]
1.4301 steelMi29183.10[2.12, 5.18]†^{\dagger}
Al7075Bn30293.10[2.29, 4.48]†^{\dagger}
Ti6Al4VBn28282.60[1.89, 3.80]†^{\dagger}
AlBB25232.40[1.66, 3.97]†^{\dagger}
1.4301 steelGr32212.20[1.47, 3.68]†^{\dagger}
1.4301 steelBn68651.80[1.37, 2.45]
TiHo32271.40[0.91, 2.50]†^{\dagger}
Al7075Tu34331.30[0.85, 2.15]
C45 steelBn89881.30[0.97, 1.70]
AlGr1691671.20[0.98, 1.49]
GraphiteHo54520.90[0.58, 1.38]
MO58A brassWE86610.80[0.51, 1.27]
MO58A brassTu3352790.70[0.56, 0.86]
1.4301 steelTu4584280.40[0.29, 0.51]
Ti6Al4VMi109840.40[0.21, 0.69]
MO58A brassMi1831800.00[0.00, 0.13]
TiBB97940.00[0.00, 0.06]
TiGr98960.00[0.00, 0.00]
Watershed-recovered (representative; full 58-surface table in SI):
C45 steelRTw^{\mathrm{w}}1991924.61[3.98, 5.48]
GraphiteBBw^{\mathrm{w}}2011984.49[4.03, 5.21]
Ti6Al4VWErw^{\mathrm{rw}}2282254.16[3.72, 4.71]
BrassHow^{\mathrm{w}}72712.00[1.64, 2.50]
BrassBBw^{\mathrm{w}}82821.71[1.36, 2.32]
1.4301 steelBBw^{\mathrm{w}}21170.71[0.36, 1.22]†^{\dagger}

Table 1. †\dagger nNNS<30n_{\mathrm{NNS}} < 30: CI width >1.5> 1.5, estimate indicative only. w^{\mathrm{w}} Watershed detector. Complete 58-surface table in Supplementary Information (Table S1).

  • Burnishing: β=1.30\beta= 1.30–4.304.30 (median 2.6, n=5n = 5). Burnished surfaces in the present dataset exhibited the highest median exclusion among finishing processes.

  • WEDM: β=0.80\beta= 0.80–4.264.26 (median 3.4, n=12n = 12)—consistently moderate-to-high exclusion.

  • Rough turning: β=0.70\beta= 0.70–4.614.61 (median 4.0, n=5n = 5). The widest range from any single process category in the present dataset.

  • Honing: β=0.90\beta= 0.90–4.004.00 (median 2.2, n=6n = 6)—intermediate to high values.

  • Grinding and bead blasting: β=−0.20\beta= -0.20–+4.49+4.49, spanning nearly the full range.

The wide within-process variability and overlap between categories confirm that factors beyond the manufacturing process alone (material, surface preparation history, measurement scale) influence β\beta.

Benchmarking against standard spatial statistics

To evaluate whether β\beta captures information beyond existing spatial descriptors, four standard spatial-statistical metrics were computed for all 21 valid surfaces: the nearest-neighbour index (NNI), Ripley’s KK normalised by the CSR expectation, Voronoi cell area variance, and the pair correlation function g(r)g(r). The Brody exponent β\beta correlates negatively with NNI (Spearman ρ=−0.64\rho= -0.64, p=0.002p = 0.002) and positively with normalised Ripley’s KK (ρ=+0.60\rho= +0.60, p=0.004p = 0.004), indicating that β\beta captures similar spatial-ordering information. However, β\beta provides a single interpretable parameter on a fixed continuum (CSR at ∼0.96\sim0.96, strong repulsion at >2> 2), whereas NNI and Ripley’s KK are scale-dependent and require reference to CSR expectations. The pair correlation function g2(r)g_2(r) is the richest descriptor but does not reduce to a single number; Moran’s II is sensitive primarily to global spatial autocorrelation rather than local exclusion; the correlation dimension D2D_2 probes multi-scale heterogeneity orthogonal to exclusion (Pearson r=−0.31r = -0.31 with β\beta). β\beta thus occupies a specific niche: a single-parameter, calibration-based measure of short-range exclusion strength, complementary to existing spatial-statistical tools rather than replacing them.

A Kruskal–Wallis test for β\beta across manufacturing processes yields H=6.02H = 6.02, p=0.197p = 0.197, with effect size η2=0.34\eta^2 = 0.34 (medium). The non-significant pp-value reflects both the sample size (58 surfaces across 10 processes) and the substantial within-process variability noted above. Methodological limitations—including peak-detection sensitivity, classification performance, and D2D_2 bias correction—are discussed in § 4.6.

Exclusion physics: synthetic point processes

The β\beta–rexclr_{\mathrm{excl}} mapping is empirical. Throughout this work, the relationship between the Brody exponent and the exclusion radius is established through calibration against synthetic point processes.

cesses (hard-core and Strauss soft-core). It is not derived from first principles. The mapping is monotonic and reproducible, which makes it useful, but it should be understood as an empirical calibration curve rather than a theoretical law.

To test whether β\beta quantifies effective exclusion strength, synthetic 2D point processes with controlled hard-core exclusion radii were generated by sequential rejection sampling (N=200N = 200 points, box size 100 units, 30 realisations per radius). The exclusion radius rexclr_{\mathrm{excl}} is expressed in units of the mean nearest-neighbour distance ⟨NN⟩\langle\mathrm{NN} \rangle of the resulting point set, providing a dimensionless, scale-invariant measure of exclusion strength. The Brody exponent was fitted to the NN spacings using the same MLE procedure applied to the FV surfaces. A 2D Poisson (complete spatial randomness, CSR) process and the 2D Ginibre ensemble (determinantal repulsion, spatial analogue of GUE) provide reference points. The Ginibre ensemble yields β=2.47±0.18\beta= 2.47 \pm0.18 (50 realisations, N=500N = 500 points, 95% CI [2.15, 2.75]), establishing a precise 2D determinantal-repulsion benchmark.

The 2D Poisson CSR process yields β=0.96±0.15\beta= 0.96 \pm0.15 (30 realisations), not β=0\beta= 0. All β\beta values should be referenced to this 2D CSR baseline, not to A direct comparison against the analytic 2D Rayleigh distribution (P(s)=(πs/2)exp⁡(−πs2/4)P(s) = (\pi s/2)\exp(-\pi s^{2}/4), the exact CSR limit) confirms that the Brody distribution provides a better description of CSR data than the Rayleigh distribution itself: across 50 CSR realisations, the Kolmogorov–Smirnov distance is DKS(Brody)=0.053±0.013D_{\mathrm{KS}}(\mathrm{Brody}) = 0.053 \pm0.013 versus DKS(Rayleigh)=0.061±0.015D_{\mathrm{KS}}(\mathrm{Rayleigh}) = 0.061 \pm0.015, and ΔAIC\Delta\mathrm{AIC} (Brody − Rayleigh) =−0.5= -0.5, indicating that the one-parameter Brody fit is preferred over the zero-parameter Rayleigh despite the AIC penalty.

At β≈10\beta\approx10, the Brody distribution has a coefficient of variation ≈0.09\approx0.09, corresponding to a nearly crystalline NN spacing. A profile-likelihood analysis of the hard-core process at rexcl=5r_{\mathrm{excl}} = 5 (yielding βMLE=6.3\beta_{\mathrm{MLE}} = 6.3) confirms that the MLE is well-identified: the observed Fisher information gives SE(β)=0.21\mathrm{SE}(\beta) = 0.21 and a 95% CI of [5.85, 6.66], ruling out optimisation artefacts at large β\beta.

Synthetic hard-core processes produce a monotonic β\beta–rexcl/⟨NN⟩r_{\mathrm{excl}}/\langle\mathrm{NN} \rangle mapping (Figure 2, left), reproducing the full range of observed β\beta values. An exclusion radius of rexcl/⟨NN⟩≈0.55r_{\mathrm{excl}}/\langle\mathrm{NN} \rangle\approx0.55 yields β≈2.0\beta\approx2.0, matching the 2D Ginibre ensemble (β=2.2±0.4\beta= 2.2 \pm0.4) and the values observed for 2D prime embeddings (β=2.15\beta= 2.15) and phase-extracted PSI profilometry (β=2.00\beta= 2.00). Burnished surfaces (β≈3.1\beta\approx3.1) correspond to rexcl/⟨NN⟩≈0.70r_{\mathrm{excl}}/\langle\mathrm{NN} \rangle\approx0.70, while turned and ground surfaces (β<1\beta< 1) are indistinguishable from the CSR baseline.

Figure 2. Exclusion physics: the Brody exponent β\beta as a measure of short-range exclusion. Left: synthetic hard-core point processes with exclusion radius rexclr_{\mathrm{excl}} normalised by the mean nearest-neighbour distance (blue), 2D Poisson CSR baseline (β=0.96±0.15\beta= 0.96 \pm0.15, grey band), 2D Ginibre ensemble (green), and observed β\beta values from manufactured surfaces (grey dots, n=21n = 21), 2D primes (purple), and PSI ceramic (red). The 2D CSR baseline is β≈0.96\beta\approx0.96, not β=0\beta= 0. Right: β\beta across synthetic reference processes and real-data domains.

These results support the hypothesis that the Brody exponent β\beta—when referenced to the correct 2D CSR baseline (β≈0.96\beta\approx0.96, not β=0\beta= 0)—quantifies effective exclusion strength across diverse generative mechanisms. The three domains occupy overlapping regions of the β\beta–rexclr_{\mathrm{excl}} parameter space, but density-thinning experiments (§ 3.6) show that absolute β\beta values are density-dependent, establishing that cross-domain comparison requires density matching.

Two additional analyses support the robustness of β\beta as an exclusion measure. First, the Brody exponent for 2D prime embeddings was computed across N∈[103,107]N \in[10^{3}, 10^{7}]. At N=103N = 10^{3}, β=2.20\beta= 2.20; at N=5×103N = 5 \times10^{3}, β=1.69\beta= 1.69; at N=105N = 10^{5}, β=2.15\beta= 2.15; at N=106N = 10^{6}, β=1.75\beta= 1.75; at N=107N = 10^{7}, β=1.26\beta= 1.26. The β(N)\beta(N) trend is non-monotonic and does not support a simple convergence to a fixed asymptotic value at accessible NN. The full β(N)\beta(N) series (12 values) is provided in Supplementary Information. The key observation is that β\beta remains substantially above the CSR baseline at all N≤5×106N \leq5 \times10^{6}, ruling out a finite-size artefact, while the decrease at N=107N = 10^{7} may reflect the approach to the sparse limit (ρ→0\rho\to0 as N→∞N \to\infty). Second, a Strauss soft-core process (exclusion penalty γ∈[0.001,1.0]\gamma\in[0.001, 1.0]) produces β=0.93±0.12\beta= 0.93 \pm0.12 (γ=1.0\gamma= 1.0) to β=2.42±0.27\beta= 2.42 \pm0.27 (γ=0.001\gamma= 0.001), demonstrating that the β\beta–exclusion mapping extends beyond the hard-core mechanism. Third, an AIC comparison across all 21 FV surfaces confirms the Brody distribution is preferred over Gamma (17/21 surfaces, median ΔAIC=+3.5\Delta\mathrm{AIC} = +3.5) and Weibull (17/21, median ΔAIC=+2.0\Delta\mathrm{AIC} = +2.0), justifying its selection over alternative two-parameter families.

Robustness of the synthetic calibration

The β\beta–rexclr_{\mathrm{excl}} calibration was subjected to four robustness tests. (i) Density independence of the CSR baseline: as reported in § 3.6, βCSR\beta_{\mathrm{CSR}} varies by less than 0.04 across ρ∈[0.01,0.10]\rho\in[0.01, 0.10], establishing that the baseline is a stable reference. (ii) Finite sample-size convergence: for the hard-core process at rexcl/⟨NN⟩=0.55r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle= 0.55, β\beta converges to within 5% of its asymptotic value at N≳100N \gtrsim100 points (see SI, Fig. S10). (iii) Observation window effects: repeating the calibration on a 50×\times50 window (half the linear size) changes β\beta by <0.15< 0.15 for all rexclr_{\mathrm{excl}}, indicating weak sensitivity to window size when NN is held constant. (iv) Edge corrections: nearest-neighbour distances were computed without edge correction; for N=200N = 200 points in a 100 ×\times 100 box, the fraction of points within one mean NN distance of the boundary is <8%< 8\%, and applying a toroidal-edge correction changes β\beta by <0.05< 0.05. The calibration is therefore robust to the edge effects present in the synthetic data.

Physical origin of > 1

The Brody exponent exceeds the CSR baseline (β>0.96\beta> 0.96) whenever short-range exclusion is present. Three physical mechanisms contribute. First, hard-core repulsion: an exclusion zone of radius rexclr_{\mathrm{excl}} around each point suppresses the probability density of the NN spacing distribution at s<rexcl/⟨NN⟩s < r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle, shifting probability mass to larger separations and increasing β\beta. The monotonic β\beta–rexclr_{\mathrm{excl}} mapping (Figure 2, left) directly quantifies this effect. Second, soft-core and many-body exclusion: the Strauss soft-core process demonstrates that the β\beta–exclusion relationship extends beyond the hard-core limit; even when points may approach arbitrarily closely (with an energetic penalty), the suppression of close pairs elevates β\beta above the CSR baseline. Third, effective exclusion from continuity constraints: in manufactured surfaces, material continuity prevents two peaks from being arbitrarily close—the surface cannot have two summits at the same location. In prime embeddings, the divisibility constraint (nn and n+1n+1 cannot both be prime for n>2n > 2) produces an analogous exclusion. These physically distinct mechanisms produce statistically similar NN spacing distributions because the Brody exponent is sensitive only to the effective exclusion strength, not to its microscopic origin.

Pair correlation function g2(r)

The pair correlation function g2(r)g_2(r) provides an internal consistency check on the β\beta-exclusion interpretation. For a point process with density ρ\rho, g2(r)g_2(r) is the normalised probability of finding two points separated by distance rr. The 1D GUE pair correlation is g2(1D)(r)=1−[sin⁡(πr)/(πr)]2g_2^{\mathrm{(1D)}}(r) = 1 - [\sin(\pi r)/(\pi r)]^2, with g2(r)→0g_2(r) \to0 as r→0r \to0 (quadratic repulsion at small separations). The formal 2D analogue is the Ginibre ensemble, a determinantal point process whose g2(r)g_2(r) also vanishes as r→0r \to0 and which provides the correct dimensional reference for 2D spatial data [27]. Figure 3 shows g2(r)g_2(r) for four representative manufactured surfaces spanning the Brody β\beta range, alongside the 2D Ginibre prediction and the 2D prime embedding. Surfaces with β≳2\beta\gtrsim2 (burnished, Ginibre-like) exhibit the characteristic dip at small rr, while surfaces with β≈0\beta\approx0 (ground Ti, Poisson-like) show g2(r)≈1g_2(r) \approx1 at all rr, consistent with complete spatial randomness. The 2D prime embedding (β=2.15\beta= 2.15) displays a g2(r)g_2(r) profile consistent with the Ginibre ensemble, indicating that the chosen 2D embeddings of primes produces nearest-neighbour statistics comparable to Ginibre-type point processes—a detectable spatial analogue of the Montgomery pair-correlation connection established for ζ\zeta zeros in 1D. The ensemble-mean g2(r)g_2(r) across 14 surfaces with sufficient peak counts for reliable g2(r)g_2(r) estimation (npeaks≥50n_{\mathrm{peaks}} \ge50; 7 of 21 valid surfaces were excluded from the ensemble mean to avoid noise-dominated contributions) (Figure 3, bottom-left) lies between the Poisson and Ginibre limits, with a reduced but detectable dip at small rr.

Figure 3. Pair correlation function g2(r)g_2(r) across domains. Top row: four representative manufactured surfaces spanning the Brody β\beta range, compared with the Ginibre prediction (solid black). β\beta values are from Table 1. Middle left: 2D binary embedding of the primes at N=105N = 10^5 (β=2.15\beta= 2.15) and matched-density Bernoulli null. Middle right: ensemble-mean g2(r)g_2(r) across 14 manufactured surfaces (±1σ\pm1\sigma band). Bottom left: Brody exponent β\beta across the three domains at native density. Density-thinning controls (§ 3.6) show that absolute β\beta values are density-dependent; cross-domain comparison requires density matching.

Complementary descriptors for binary patterns

For binary patterns where peak detection is inapplicable, complementary spatial descriptors—the correlation dimension D2D_2 and Banach-space invariants ∥∇Z∥2\lVert\nabla Z \rVert_2, total variation (TV)—are developed in Supplementary Information (SI, Sections S1.1–S1.7). These descriptors are empirically orthogonal to β\beta (Pearson r=−0.31r = -0.31), confirming that the Brody exponent probes a specific axis—exclusion strength—within a broader spatial-order parameter space. Square-free numbers and the Liouville function serve as negative controls indistinguishable from the Bernoulli null, demonstrating specificity of the D2D_2 analysis; full validation includes embedding-independence tests, finite-size bias correction, and Benjamini–Hochberg multiplicity adjustment across 14 arithmetic sequences.

Density-thinning controls

To test whether the observed β\beta values reflect genuine exclusion structure rather than point density, the 2D prime embedding (N=105N = 10^5, ρ=0.096\rho= 0.096, β=2.15\beta= 2.15) was randomly thinned to match point counts corresponding to ρ∈{0.01,0.02,0.032,0.05,0.096}\rho\in\{0.01, 0.02, 0.032, 0.05, 0.096\} (20 trials each). At each density, β\beta was compared against a density-matched CSR baseline (20 realisations each).

At the PSI density (ρ=0.032\rho= 0.032), prime β\beta decreased to 1.46 ±\pm 0.02 [1.42, 1.49], and at the CSR calibration density (ρ=0.02\rho= 0.02) to 1.30 ±\pm 0.05. Critically, prime β\beta remained significantly above the density-matched CSR baseline at all densities (CSR: β=0.98±0.05\beta= 0.98 \pm0.05 at ρ=0.032\rho= 0.032; 0.97 ±\pm 0.04 at ρ=0.02\rho= 0.02; 0.99 ±\pm 0.02 at ρ=0.096\rho= 0.096), confirming that the prime embedding possesses genuine exclusion structure beyond that expected from density alone. The CSR baseline itself is density-stable across the tested range (0.95–0.99).

Two conclusions follow. First, β\beta captures exclusion strength rather than point density: the CSR baseline is flat, while structured data lie above it at all densities. Second, absolute β\beta values are density-dependent: the same underlying exclusion mechanism (prime gap constraints) produces β=2.15\beta= 2.15 at ρ=0.096\rho= 0.096 but β=1.46\beta= 1.46 at ρ=0.032\rho= 0.032. Cross-domain comparisons of absolute β\beta therefore require density matching. The synthetic hard-core calibration (§ 3.4) is density-referenced through the normalisation rexcl/⟨NN⟩r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle, which provides a density-invariant measure of exclusion strength.

2D prime embeddings: embedding geometry determines exclusion

Prime-number embeddings provide a mathematically controlled benchmark: any exclusion signal must originate from arithmetic divisibility constraints rather than physical interactions.

Whether the Brody exponent detects such arithmetic exclusion—and whether this depends on how the 1D prime sequence is mapped to 2D—is tested here.

β\beta was computed under three deterministic 2D embeddings at N=105N = 10^{5}: row-major (β=2.15\beta= 2.15), Ulam spiral (β=1.87\beta= 1.87), and Cantor pairing (β=1.40\beta= 1.40). A spatially random Bernoulli control at matched density (30 realisations) yields βBern=1.47±0.02\beta_{\mathrm{Bern}} = 1.47 \pm0.02.

The Cantor null result. Under Cantor pairing, primes yield β=1.40\beta= 1.40, which is slightly below the Bernoulli baseline (Δβ=−0.07\Delta\beta= -0.07, two one-sided test for equivalence against Bernoulli: pTOST<0.01p_{\mathrm{TOST}} < 0.01 at equivalence bound ±0.1\pm0.1). This is a decisive finding: the same prime sequence that produces strong exclusion under row-major embedding (β=2.15\beta= 2.15) produces no detectable exclusion under Cantor pairing. The conclusion is unambiguous: 2D exclusion in prime embeddings is not an intrinsic property of the prime sequence; it is created or destroyed by the embedding geometry. The row-major embedding amplifies the prime gap structure because consecutive integers—which cannot both be prime (for n>2n > 2)—map to horizontally adjacent array positions, translating the 1D gap constraint into 2D nearest-neighbour exclusion. Ulam spiral produces intermediate exclusion (β=1.87\beta= 1.87) by concentrating primes along quadratic-residue diagonals while leaving off-diagonal regions depleted. Cantor pairing destroys this structure by sweeping the array diagonally, decorrelating the prime gap constraint from 2D adjacency. The embedding dependence indicates that β\beta for arithmetic sequences reflects both the 1D sequence statistics and the 2D embedding geometry; claims about “prime point processes” must therefore specify the embedding. Figure 4 summarises these results.

Figure 4. Embedding dependence of the Brody exponent for 2D primes at N=105N = 10^{5}. Left: β\beta under three deterministic embeddings and the Bernoulli 2D control (matched density). Row-major and Ulam embeddings produce β\beta significantly above the spatially random control; Cantor pairing does not. Right: distribution of β\beta for 30 Bernoulli 2D realisations, with the row-major prime value indicated. The embedding dependence demonstrates that β\beta for arithmetic sequences reflects both sequence structure and embedding geometry.

Sparse-integer control: sparsity versus arithmetic structure. To test whether the elevated β\beta for row-major primes (β=2.15\beta= 2.15) reflects the arithmetic constraint of primality or merely the sparsity of the prime set (π(105)=9592\pi(10^{5}) = 9592, ρ=0.096\rho= 0.096), a sparse-integer control was performed: π(N)\pi(N) random integers uniformly distributed in [1,N][1,N] were embedded in the identical row-major geometry and β\beta was computed (30 trials). The sparse-integer control yields β=1.47±0.02\beta= 1.47 \pm0.02 [1.43, 1.50], which is indistinguishable from the Bernoulli CSR baseline (β=1.46±0.02\beta= 1.46 \pm0.02) and substantially below the prime value (Δβ=+0.68\Delta\beta= +0.68). This result establishes that the elevated β\beta for row-major primes is a genuine arithmetic signal—not a sparsity artefact—and that the CSR baseline for binary fields at this density (ρ=0.096\rho= 0.096) is βCSR(binary)≈1.46\beta_{\mathrm{CSR}}^{(\mathrm{binary})} \approx1.46, distinct from the continuous 2D CSR baseline of β=0.96±0.15\beta= 0.96 \pm0.15 (Section 3.4). The binary CSR baseline shift reflects the discrete-grid NNS distribution at low fill fraction; all comparisons between binary arithmetic surfaces and their null models use the appropriate binary CSR reference. Figure 5 presents the density-thinning and sparse-integer control results.

Figure 5. Density-thinning and sparse-integer controls. Left: Brody exponent β\beta for row-major primes thinned to five densities (ρ=0.01\rho= 0.01–0.0960.096), compared with density-matched CSR baselines (grey band). β\beta decreases with density (β=2.15\beta= 2.15 at ρ=0.096\rho= 0.096 to β=1.30\beta= 1.30 at ρ=0.01\rho= 0.01) but remains above CSR at all densities. Right: sparse-integer control—30 trials of π(N)\pi(N) random integers embedded identically to primes yield β=1.47±0.02\beta= 1.47 \pm0.02, indistinguishable from the Bernoulli CSR baseline (1.46±0.021.46 \pm0.02) and substantially below the prime value (β=2.15\beta= 2.15). The prime β\beta signal is therefore a genuine arithmetic effect, not a sparsity artefact.

Peak-detection sensitivity

To assess the robustness of β\beta to the peak-detection threshold, a synthetic surface with 50 embedded Gaussian peaks (known positions, amplitudes 0.5–2.0, additive Gaussian noise σ=0.1\sigma= 0.1) was analysed under prominence thresholds ranging from 1% to 15% of the height range. The Brody exponent is stable across the 1%–11% range (β=1.26\beta= 1.26–1.27, variation <0.01< 0.01), then decreases monotonically to β=1.07\beta= 1.07 at 15% as genuine peaks are increasingly excluded (peak count drops from 425 to 63). The plateau region (1%–11%) corresponds to thresholds that retain all genuine peaks while suppressing noise; the 3% threshold used throughout this work lies within this stable plateau. A systematic sensitivity analysis across all 58 empirical surfaces remains to be performed; the present five-surface analysis (Section 3.1) and synthetic validation establish plausible robustness but do not fully characterise the threshold-dependence of β\beta for arbitrary surface textures. Figure 6 provides the complete sensitivity profile.

Figure 6. Peak-detection sensitivity analysis on a synthetic surface with 50 embedded Gaussian peaks. Left: Brody β\beta versus prominence threshold (% of height range). β\beta is stable (< 0.01 variation) across 1%–11% thresholds, then decreases as genuine peaks are excluded. Right: number of detected peaks versus threshold. The 3% threshold used throughout this work (dashed line) lies within the stable plateau.

Proof-of-concept: interferometric profilometry

A proof-of-concept application of the Brody pipeline to phase-extracted interferometric profilometry of a certified roundness standard (JENOPTIK FN 111) is presented in Supplementary Information (SI, Section S4). The concatenated five-revolution profile (629 peaks) yields β\beta = 2.00 [1.90, 2.10]; per-revolution instability (5–9 peaks, β\beta = 1.00–3.68) qualifies the result as exploratory. Scale-invariance analysis of ΔD2(N)\Delta D_2(N) for primes across N∈[104,107]N \in[10^4, 10^7] is also reported in SI (Sections S1.5–S1.6).

Discussion

What measures—and what it does not

The β\beta–rexclr_{\mathrm{excl}} mapping is empirical. The relationship between the Brody exponent and the normalised exclusion radius is established through calibration against synthetic hard-core and Strauss soft-core point processes (Section 3.2). It is not derived from first principles. The mapping is monotonic and reproducible, which makes it practically useful, but it should be understood as an empirical calibration curve rather than a theoretical law. Whether a deeper connection exists between the Brody functional form and 2D exclusion physics—for instance, through determinantal point-process kernels [28]—is an open question.

The calibration provides a physical interpretation: β≈0.96\beta\approx0.96 corresponds to no exclusion (CSR), β≈2\beta\approx2 to Ginibre-level repulsion (rexcl≈0.55⟨NN⟩r_{\mathrm{excl}} \approx0.55 \langle\mathrm{NN} \rangle), and β≳5\beta\gtrsim5 to near-crystalline exclusion (rexcl≳0.8⟨NN⟩r_{\mathrm{excl}} \gtrsim0.8 \langle\mathrm{NN} \rangle). Values of β>2\beta> 2 are not interpreted as higher-order RMT symmetry classes; they serve solely as a phenomenological measure of exclusion strength. Conceptually, the β\beta–rexclr_{\mathrm{excl}} relation plays a role analogous to the hard-sphere equation of state [29]: both compress complex spatial correlations into a single interpretable quantity. Unlike the Carnahan–Starling equation, however, the β\beta–rexclr_{\mathrm{excl}} calibration is purely empirical.

The effective hard-core radius reffr_{\mathrm{eff}}—the 5th percentile of the NN distance distribution—provides independent validation. Across the synthetic calibration data, reff/⟨NN⟩r_{\mathrm{eff}}/\langle\mathrm{NN} \rangle correlates with β\beta at Spearman ρ=0.988\rho= 0.988 (p<10−4p < 10^{-4}). This near-perfect correlation establishes that β\beta and the effective exclusion radius capture essentially identical information, providing the quantitative link between Brody-NNS analysis and standard spatial point-process theory [13].

Critically, β\beta does not identify the generative process, replace g2(r)g_2(r), detect long-range order, or serve as a general-purpose surface descriptor. Adding β\beta to ISO 25178 parameters did not improve manufacturing-process classification accuracy (§ 3.1). This null result is expected: β\beta measures short-range exclusion, while manufacturing processes differ in many aspects (roughness amplitude, skewness, waviness, material).

Validity range and interpretation of large β\beta

The Brody MLE is well-identified across the full reported range (β=0\beta= 0 to +4.6+4.6 for surfaces, to β≈10\beta\approx10 for synthetic hard-core processes). A profile-likelihood analysis at β=6.3\beta= 6.3 (hard-core, rexcl=5r_{\mathrm{excl}} = 5) gives Fisher-information SE(β\beta) = 0.21 and a 95% CI of [5.85, 6.66], ruling out optimisation artefacts. Bootstrap stability is confirmed across 500 resamples (§ 3.4). Goodness-of-fit assessed by Kolmogorov–Smirnov distance remains acceptable (DKS<0.10D_{\mathrm{KS}} < 0.10) for all reported fits.

Values of β>2\beta> 2 are emphatically not interpreted as evidence of higher-order RMT symmetry classes. The Brody distribution was introduced as a phenomenological one-parameter interpolation; its functional form P(s)∝sβexp⁡(−asβ+1)P(s) \propto s^\beta\exp(-a s^{\beta+1}) remains mathematically well-defined and MLE-identifiable for arbitrarily large β\beta. In the large-β\beta limit, the distribution approaches a narrow Gaussian centred at s=1s = 1 (coefficient of variation ≈0.09\approx0.09 at β=10\beta= 10), corresponding to nearly crystalline NN spacings. This is the expected behaviour for strong hard-core exclusion and does not imply any new universality class. The interpretation is purely phenomenological: β\beta quantifies the degree of short-range order on a continuous scale, with β≈0.96\beta\approx0.96 (CSR), β≈2\beta\approx2 (Ginibre-level), and β≳5\beta\gtrsim5 (near-crystalline) serving as convenient reference points.

Scope and limitations of the calibration

The β\beta–rexclr_{\mathrm{excl}} calibration is not universal. It depends on:

  • Point density: absolute β\beta values are density-dependent (§ 3.5). Cross-domain comparison requires density matching or the density-invariant rexcl/⟨NN⟩r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle calibration.

  • Embedding geometry: β\beta for arithmetic sequences varies from 2.15 (row-major) to 1.40 (Cantor) for the same prime set. The embedding must be specified.

  • Peak-detection strategy: the watershed detector shifted absolute β\beta values while preserving relative ordering. Detector choice must be reported.

  • Binary versus continuous domains: the CSR baseline differs (≈1.46\approx1.46 versus 0.96). The appropriate baseline must be used (Table 2).

Data typeCSR baselineExample
Continuous 2DβCSR=0.96±0.15\beta_{\mathrm{CSR}} = 0.96 \pm0.15FV surfaces, PSI
Binary, ρ<0.2\rho< 0.2Density-matched BernoulliPrimes, arithmetic
UncertainGenerate own (20 reps)Any new class

Table 2. CSR baseline selection for 2D Brody analysis.

These dependencies do not invalidate β\beta as an exclusion measure—they define the measurement conditions under which it is valid. The contribution of this work is to identify and quantify these conditions, providing the correction protocols required for reproducible Brody analysis in two dimensions.

The two CSR baselines

A necessary finding is that the 2D CSR baseline is β=0.96±0.15\beta= 0.96 \pm0.15, not β=0\beta= 0. This correction is essential: the 2D NNS distribution under CSR is Rayleigh, P(s)=(πs/2)exp⁡(−πs2/4)P(s) = (\pi s/2)\exp(-\pi s^2/4), which is not a member of the Brody family. The Brody distribution at β≈0.96\beta\approx0.96 provides an excellent approximation (KS distance ≈0.05\approx0.05), outperforming the analytic Rayleigh in 22 of 30 CSR trials. The baseline is density-stable (βCSR=0.95\beta_{\mathrm{CSR}} = 0.95–0.990.99 across ρ∈[0.01,0.10]\rho\in[0.01, 0.10]).

A second, distinct CSR baseline emerges for binary fields at low fill fraction. For the Bernoulli null at ρ=0.096\rho= 0.096 (matching the prime surface density), βCSR(binary)=1.46±0.02\beta_{\mathrm{CSR}}^{(\mathrm{binary})} = 1.46 \pm0.02, substantially above the continuous CSR value. This shift reflects the discrete-grid NNS distribution at low occupancy. The practical implication is summarised in Table 2.

Using the wrong baseline produces spurious exclusion claims. All comparisons in the present work use density-matched CSR references.

Density dependence and cross-domain comparison

Density-thinning experiments (§ 3.6) reveal that β\beta captures exclusion strength rather than point density—the CSR baseline is flat, while structured data consistently exceed it—but that absolute β\beta values are density-dependent. The 2D prime embedding yields β=2.15\beta= 2.15 at native density ρ=0.096\rho= 0.096 but β=1.46±0.02\beta= 1.46 \pm0.02 when thinned to the PSI density (ρ=0.032\rho= 0.032). This density dependence has two implications.

First, it resolves the apparent convergence of primes and PSI at β≈2.0\beta\approx2.0: at matched density, primes (β=1.46\beta= 1.46) and PSI (β=2.00\beta= 2.00) occupy distinct regions of the β\beta scale. The similarity of their native-density β\beta values is partially a density artefact, not evidence of identical exclusion mechanisms. Both domains nevertheless exhibit β\beta values significantly exceeding their respective CSR baselines, confirming genuine exclusion structure.

Second, and more constructively, the synthetic hard-core calibration (§ 3.4) uses the normalised exclusion radius rexcl/⟨NN⟩r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle, which is density-invariant by construction. Cross-domain comparisons should therefore be performed in the (rexcl/⟨NN⟩,βr_{\mathrm{excl}}/\langle\mathrm{NN}\rangle, \beta) parameter space rather than by comparing absolute β\beta values. When referenced to this density-invariant calibration, all three domains occupy a consistent region of the parameter space: exclusion radii in the range 0.30.3–0.7⟨NN⟩0.7\langle\mathrm{NN}\rangle, corresponding to β≈1.3\beta\approx1.3–4.64.6 depending on density.

Physical interpretation by manufacturing process

In the present dataset, burnished surfaces exhibited the highest median β\beta (β=1.3\beta= 1.3–4.34.3, rexcl/⟨NN⟩≈0.5r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle\approx0.5–0.750.75). Surfaces produced by turning and grinding (β=0.0\beta= 0.0–2.22.2, rexcl/⟨NN⟩<0.55r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle< 0.55) remained broadly indistinguishable from CSR-indistinguishable to moderate exclusion. Milled surfaces showed the widest β\beta range (β=0.0\beta= 0.0–3.73.7). Whether these differences reflect the manufacturing process, the material, or their interaction cannot be determined from the present observational design, in which process and material are confounded (e.g., all burnished surfaces are ductile metals). A factorial experiment crossing process ×\times material is required for causal attribution. The qualitative observation that finishing operations produced higher β\beta than abrasive operations in the present dataset was robust to peak-detection algorithm choice and threshold variation.

Internal consistency with g2(r)g_2(r)

The pair correlation function g2(r)g_2(r) provides an internal consistency check: the exclusion radius inferred from β\beta (via the hard-core calibration) and from the g2(r)g_2(r) contact value (smallest rr where g2(r)>0.5g_2(r) > 0.5) are systematically related across 7 FV surfaces (ratio 0.24 ±\pm 0.17). The systematic offset reflects the different definitions of “exclusion”: β\beta captures the shape of the full NNS distribution, while the g2(r)g_2(r) threshold is a single-point measure at contact. Both measures consistently identify the same surfaces as having high or low exclusion, supporting the interpretation of β\beta as an exclusion measure.

Complementary descriptors

Binary patterns where peak detection is inapplicable are characterised via D2D_2 correlation dimension and Banach-space invariants (SI, Sections S1.1–S1.7). These descriptors are empirically orthogonal to β\beta (Pearson r=−0.31r = -0.31), confirming that β\beta probes a specific axis—exclusion strength—within a broader spatial-order parameter space. Square-free numbers and the Liouville function serve as negative controls indistinguishable from the Bernoulli null, while primes exhibit ΔD2>0\Delta D_2 > 0 driven by the fundamental theorem of arithmetic (consecutive integers cannot both be prime).

A noteworthy tension exists between the D2D_2 and β\beta results. D2D_2 for primes is stable across row-major, Ulam, and Cantor embeddings (ΔD2≈+0.05\Delta D_2 \approx+0.05–0.090.09; SI, Section S1.1), indicating that the multi-scale mass distribution is an intrinsic property of the prime sequence. In contrast, β\beta varies from 2.15 (row-major) to 1.40 (Cantor), indicating that short-range exclusion is embedding-dependent. This tension is resolved by recognising that D2D_2 and β\beta probe fundamentally different spatial scales: D2D_2 responds to mass distribution across all box sizes simultaneously, whereas β\beta is dominated by the shortest nearest-neighbour distances. An embedding that preserves the prime gap constraint at nearest-neighbour scales (row-major) yields high β\beta; an embedding that decorrelates the gap constraint from 2D adjacency (Cantor) yields β\beta indistinguishable from CSR, even though the multi-scale arithmetic structure remains detectable by D2D_2. The complementarity of β\beta, D2D_2, and Banach descriptors suggests that spatial point patterns may be classified by their coordinates in a low-dimensional descriptor space whose axes correspond to physically interpretable order parameters.

Limitations

Five limitations qualify the present findings. (1) The β\beta–rexclr_{\mathrm{excl}} calibration is empirical, not derived from first principles; the Brody functional form is justified by AIC comparison (§ 3.2) but its optimality for 2D exclusion is not established analytically. (2) Absolute β\beta values are detector-dependent: the watershed detector recovered 37 grid-limited surfaces but shifted absolute β\beta values (range 0.71–4.61, mean 2.89 ±\pm 1.10); relative ordering is robust. A systematic threshold sensitivity analysis across all 58 surfaces is needed. (3) Surfaces with nNNS<30n_{\mathrm{NNS}} < 30 have wide confidence intervals (±\pm1.0–1.5) and are indicative only. (4) The PSI result is exploratory: a single standard, 5–9 peer revolution, per-revolution pp spanning 1.00–3.68. Replication on additional standards is required. (5) The surface dataset is observational (process ×\times material confounded); causal attribution requires a factorial experiment. Test–retest reliability has not been established. (6) The β\beta scale is calibrated only on the anti-clustered side (CSR through near-crystalline). The behaviour of β\beta for clustered point processes (Thomas, Matérn) is not established here; β\beta values significantly below the CSR baseline may indicate clustering, but this requires dedicated study with appropriate benchmarks.

Outlook

The calibrated framework enables several directions. In spatial statistics, connecting β\beta to formal point-process model parameters (Strauss γ\gamma, area-interaction radius, DPP kernels) would bridge RMT-inspired NNS analysis with established spatial point-process theory [28]. In surface engineering, β\beta may serve as a target parameter where peak proximity is the quantity of interest, subject to test–retest validation. In number theory, the embedding-dependence of β\beta for primes—particularly the Cantor null result—provides an empirical constraint on how 1D arithmetic structure manifests in 2D spatial statistics. Extension to 3D (additive-manufacturing porosity, granular packings) and to additional point-process domains (galaxy distributions, vortex matter, neural spike patterns) would test the generality of the calibration framework.

Conclusion

The Brody exponent β\beta, calibrated against the 2D complete-spatial-randomness baseline (β=0.96±0.15\beta= 0.96 \pm0.15, not β=0\beta= 0), provides an empirically robust measure of short-range exclusion strength in 2D spatial point processes. Synthetic hard-core and soft-core experiments establish a monotonic β\beta–rexcl/⟨NN⟩r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle mapping spanning β≈1\beta\approx1 (CSR) to β≈10\beta\approx10 (near-crystalline exclusion). This calibration—empirical rather than derived from first principles—translates the dimensionless Brody parameter into a physically interpretable exclusion radius and provides the correct 2D reference frame for any application of Brody statistics to spatial data. Values of β>2\beta> 2 are not interpreted as higher-order RMT symmetry classes but solely as a phenomenological measure of exclusion strength.

Three unrelated generative mechanisms were analysed within this calibrated framework. Manufactured surface peaks (β=0\beta= 0 to +4.6+4.6, 58 surfaces, 10 processes) exhibit process-dependent exclusion; in the present dataset, burnished surfaces showed the highest median exclusion and turned surfaces the weakest (β≈0.4\beta\approx0.4, indistinguishable from CSR). 2D binary embeddings of the primes yield β=2.15\beta= 2.15 (row-major), 1.87 (Ulam), and 1.40 (Cantor) at N=105N = 10^{5}, demonstrating that β\beta depends on the embedding geometry as well as on the arithmetic sequence. At matched density (ρ=0.032\rho= 0.032), the row-major prime embedding yields β=1.46±0.02\beta= 1.46 \pm0.02, significantly exceeding the density-matched CSR baseline (0.98±0.050.98 \pm0.05) but below the native-density value. Phase-extracted interferometric profilometry of a certified roundness standard—included as an independently generated spatial point process from optical phase reconstruction—yields β=2.00\beta= 2.00 [1.90, 2.10] at ρ=0.032\rho= 0.032, also well above CSR.

Density-thinning and embedding-robustness experiments establish three findings. (i) β\beta captures exclusion rather than density: the CSR baseline is density-stable (0.95–0.99 across ρ∈[0.01,0.10]\rho\in[0.01, 0.10]) while structured data consistently exceed it. (ii) Absolute β\beta values are density-dependent: cross-domain comparisons require density matching or use of the density-invariant rexcl/⟨NN⟩r_{\mathrm{excl}}/\langle\mathrm{NN}\rangle calibration. (iii) β\beta for 2D arithmetic embeddings is embedding-dependent: the same prime sequence produces different β\beta under different 2D mappings, so claims about the spatial statistics of primes must specify the embedding.

For binary patterns where peak detection is inapplicable, complementary D2D_2 and Banach-space descriptors capture orthogonal aspects of spatial structure (Pearson r=−0.31r = -0.31), with square-free numbers and the Liouville function serving as decisive negative controls (Supplementary Information).

The Cantor embedding null result is particularly instructive: the same prime sequence yields β=2.15\beta= 2.15 (row-major), strong exclusion, and β=1.40\beta= 1.40 (Cantor), no exclusion (Δβ=−0.07\Delta\beta= -0.07 relative to the Bernoulli baseline, pTOST<0.01p_{\mathrm{TOST}} < 0.01). This demonstrates that 2D exclusion in arithmetic embeddings is not an intrinsic property of the arithmetic sequence—it is created or destroyed by the embedding geometry. The framework therefore detects both the presence and the absence of exclusion, which is essential for any measurement tool.

The β\beta–rexclr_{\rm excl} calibration (Spearman ρ=0.988\rho= 0.988 with reffr_{\rm eff}), the 2D CSR baseline correction (β=0.96±0.15\beta= 0.96 \pm0.15 for continuous processes; βCSR(binary)≈1.46\beta_{\rm CSR}^{\rm(binary)} \approx1.46 for binary fields at low density), and the density-thinning and embedding-robustness protocols together constitute a calibrated measurement framework for systematic, reproducible characterisation of short-range exclusion in 2D spatial point processes. The framework is applicable across physical, mathematical, and engineering domains, provided the necessary corrections—density matching, embedding specification, and reference to the appropriate CSR baseline—are applied.

Acknowledgements

The author gratefully acknowledges the support of Poznan University of Technology, Poland. This work was funded by the grant 0614/SBAD/1603. The open-source scientific Python ecosystem (NumPy [30], SciPy [31], Matplotlib [32], mpmath [33]) enabled this computational study. The author declares no competing interests.

Supplementary Information

Extended validation of D2D_2 and Banach descriptors

The main text (§ 3.6–§ 3.7) presents the essential findings from the D2D_2 and Banach-space descriptor analysis. This Supplementary Information provides the complete set of validation results.

Embedding independence

The D2D_2 signal for binary primes persists across three deterministic 2D embeddings (row-major, Ulam spiral, Cantor diagonal) with ΔD2=+0.053\Delta D_2 = +0.053 to +0.089+0.089, ruling out embedding-geometry artefacts. Under random permutation of the 1D-to-2D mapping (30 surrogates), D2D_2 drops from 1.687 to 1.642±0.0031.642 \pm0.003 (z=13.5z = 13.5), confirming the signal is intrinsic to the prime sequence.

Estimator bias

The box-counting estimator of D2D_2 exhibits finite-size bias: 2−D2(Bern)≈12.9/log⁡N−0.862-D_2^{(\mathrm{Bern})} \approx12.9/\log N - 0.86. At N=105N = 10^5, this accounts for ∼250%\sim250\% of the apparent deviation of D2(prime)D_2^{(\mathrm{prime})} from 2. The bias-cancelled excess ΔD2≈0.04\Delta D_2 \approx0.04–0.050.05 is stable across N∈[104,107]N \in[10^4, 10^7].

Controlled-correlation ordering

A calibrated scale of surrogate binary sequences at matched prime density establishes: D2(block-clustered)=1.504<D2(Bernoulli)=1.645<D2(primes)=1.687<D2(repulsive)=1.797D_2(\mathrm{block\text{-}clustered}) = 1.504 < D_2(\mathrm{Bernoulli}) = 1.645 < D_2(\mathrm{primes}) = 1.687 < D_2(\mathrm{repulsive}) = 1.797, confirming that primes occupy an intermediate anti-clustered position.

Arithmetic sequences beyond primes

The D2D_2 analysis was applied to additional arithmetic sequences at N=5×104N = 5 \times10^4–10510^5, selected to span a range of multiplicative constraints: twin primes (upper), primes ≡1(mod4)\equiv1 \pmod{4} and ≡3(mod4)\equiv3 \pmod{4} (Chebyshev bias test), sums of two squares, and square-free numbers (negative control). The Möbius μ±(n)\mu^{\pm}(n) and Liouville λ(n)\lambda(n) functions were also tested. All sequences except square-free numbers and the Liouville function survive Benjamini–Hochberg correction (FDR = 0.05). Square-free numbers and the Liouville function are indistinguishable from the Bernoulli null, serving as negative controls. Beatty sequences [αn] mod 2[\alpha n] \bmod2 for selected irrationals are discussed in S1.5; binary digit sequences of fundamental constants and the Champernowne and Copeland–Erdős constructions did not produce statistically significant ΔD2\Delta D_2 after bias correction and are omitted for brevity.

S1.5 Convergence analysis. ΔD2(N)\Delta D_{2}(N) for primes persists across N∈[5×103,106]N \in[5 \times10^{3}, 10^{6}], while Möbius μ−\mu^{-} fluctuates around zero and Beatty φ\varphi oscillates with Moiré alignment resonances.

S1.6 Analytic bias function. An analytic approximation for the box-counting bias in sparse binary surfaces is derived: D2(N,p)=log⁡(1/p+W2/4)/log⁡(W/4)−log⁡(1/p+4)/log⁡(W/4)D_{2}(N,p) = \log(1/p+W^{2}/4)/\log(W/4)-\log(1/p+4)/\log(W/4), with empirical correction factor 1.24±0.151.24 \pm0.15.

Complete Brody β\beta table (all 58 surfaces)

tbl:S1: Brody exponent β\beta for all 58 analysable surfaces (21 prominence-based + 37 watershed-recovered), ordered by decreasing β\beta.

Supplementary figures

Figure S1. Multifractal DqD_q spectra and f(α)f(\alpha) singularity spectrum for the binary prime surface and matched Bernoulli null.

MaterialProc.npeaksn_{\mathrm{peaks}}nNNSn_{\mathrm{NNS}}β\beta95% CI
C45 steelRTw1991924.61[3.98, 5.48]
ELLORRTw2312294.54[4.15, 5.11]
GraphiteBBw2011984.49[4.03, 5.21]
MO58A brassBn41414.30[3.38, 5.81]
Al7075WEr``w2092084.26[3.80, 4.73]
Ti6Al4VWEr``w2282254.16[3.72, 4.71]
P1-1.4301-t.zgrubw1971784.04[3.64, 4.54]
1.4301 steelHo49494.00[3.22, 5.12]
ELLORTuw1981953.99[3.52, 4.71]
Al7075RTw1351163.97[3.54, 4.39]
Ti6Al4VWEf``w1811793.91[3.44, 4.66]
ELLORRMw1961953.90[3.51, 4.39]
MO58A brassWEr``w1861803.89[3.52, 4.36]
Al7075Mi44433.70[2.94, 4.91]
Ti6Al4VTuw107993.68[3.28, 4.18]
C45 steelRMw1111093.58[3.09, 4.35]
C45 steelTuw1391323.34[2.99, 3.78]
C45 steelWEr``w1801803.30[2.88, 3.86]
1.4301 steelWEr``w1471453.28[2.93, 3.83]
Ti6Al4VMiw1331303.25[2.78, 3.84]
C45 steelWEf``w1781763.20[2.83, 3.74]
1.4301 steelMi29183.10[2.03, 4.86]†
Al7075Bn30293.10[2.35, 4.49]†
GraphiteGrw1321313.03[2.65, 3.71]
Al7075WEf``w97922.98[2.48, 3.62]
ELLORWEr``w88872.78[2.28, 3.70]
MO58A brassRMw71682.70[2.36, 3.24]
Ti6Al4VBn28282.60[1.85, 3.86]†
1.4301 steelWEf``w98952.59[2.20, 3.14]
MO58A brassTuw30252.46[1.80, 3.70]†
AlBB25232.40[1.68, 3.74]†
AlHow1101102.26[1.88, 2.69]
1.4301 steelRMw100992.25[2.01, 2.59]
1.4301 steelGr32212.20[1.44, 3.46]†
Ti6Al4VRTw1201202.20[1.94, 2.59]
C45 steelHow37352.18[1.53, 3.36]
BrassHow72712.00[1.64, 2.50]
1.4301 steelBn68651.80[1.36, 2.38]
C45 steelGrw24211.73[1.15, 2.74]†
BrassBBw82821.71[1.36, 2.32]
C45 steelMiw25221.64[1.24, 2.43]†
Ti6Al4VHo32271.40[0.87, 2.34]†
C45 steelBBw14141.34[1.09, 1.72]†
Al7075Tu34331.30[0.87, 2.02]
C45 steelBn89881.30[0.99, 1.72]
AlGr1691671.20[0.99, 1.46]
ELLORMiw42391.18[0.78, 1.84]
ELLORWEf``w21171.14[0.79, 1.60]†
GraphiteHo54520.90[0.58, 1.32]
BrassGrw39380.85[0.56, 1.36]
MO58A brassWEf86610.80[0.52, 1.21]
1.4301 steelBBw21170.71[0.36, 1.22]†
MO58A brassRT3352790.70[0.56, 0.87]
1.4301 steelTu4584280.40[0.30, 0.52]
Ti6Al4VRM109840.40[0.19, 0.66]
MO58A brassMi1831800.00[-0.10, 0.12]
Ti6Al4VBB9794-0.10[-0.22, 0.07]
Ti6Al4VGr9896-0.20[-0.31, -0.05]

Table 3. Complete β\beta values for all 58 manufactured surfaces. †nNNS<30n_{\mathrm{NNS}} < 30: CI width >1.5> 1.5, estimate indicative only. wWatershed detector.

Generalised dimension spectra and correlation dimensions for representative arithmetic surfaces

Figure S3. Generalised DqD_q spectra for representative manufactured surfaces.

Figure S4. Forest plot of ΔD2\Delta D_2 for all arithmetic sequences with Benjamini–Hochberg-corrected significance.

Figure S5. Power spectral density of the binary prime gap surface across scales N∈[103,107]N \in[10^3, 10^7].

Figure S6. Convergence of ΔD2(N)\Delta D_{2}(N) for the binary prime surface and Möbius μ−\mu^{-} across N∈[104,107]N \in[10^{4},10^{7}].

Figure S7. Comparison of physical (FV) and mathematical (arithmetic) surface descriptors.

Brody exponent plots for prominence-based manufactured surfaces

Figure S8. Brody exponent β\beta for 21 prominence-based manufactured surfaces with nNNSn_{\mathrm{NNS}} indicated.

Figure S9. Banach-space descriptor fingerprints for six arithmetic sequences at N=105N = 10^{5}.

Figure S10. AIC-based model selection for the Brody distribution across N∈[103,107]N \in[10^{3}, 10^{7}] for the binary prime embedding.

β(N)\beta(N) convergence data

The Brody exponent β\beta for the row-major binary prime embedding was computed at twelve values of NN: N=103N = 10^{3} (β=2.20\beta= 2.20), N=2×103N = 2 \times10^{3} (β=2.10\beta= 2.10), N=5×103N = 5 \times10^{3} (β=1.69\beta= 1.69), N=104N = 10^{4} (β=1.97\beta= 1.97), N=2×104N = 2 \times10^{4} (β=2.19\beta= 2.19), N=5×104N = 5 \times10^{4} (β=1.94\beta= 1.94), N=105N = 10^{5} (β=2.15\beta= 2.15), N=2×105N = 2 \times10^{5} (β=2.05\beta= 2.05), N=5×105N = 5 \times10^{5} (β=1.78\beta= 1.78), N=106N = 10^{6} (β=1.75\beta= 1.75), N=5×106N = 5 \times10^{6} (β=1.86\beta= 1.86), N=107N = 10^{7} (β=1.26\beta= 1.26). The trend is non-monotonic; no simple convergence to a fixed asymptotic value is observed at accessible NN. The decrease at N=107N = 10^{7} may reflect the approach to the sparse limit (ρ=π(N)/N∼1/log⁡N→0\rho= \pi(N)/N \sim1/\log N \to0 as N→∞N \to\infty).

Data Availability

All analysis code, processed data (JSON format), and figure-generation scripts are archived in a public GitHub repository at https://github.com/dawidkucharski/brody-2d-exclusion. The repository includes the complete 58-surface Brody-β\beta dataset, all synthetic point-process calibration data, and a reproducible computational environment specified via requirements.txt. Surface metrology raw data (focus-variation .sur files, PSI interferogram frames) are available from the author upon reasonable request. The aurora_lab Python package (v2.0.0) provides reproducible implementations of the surface construction, null-model generation, box-counting, and validation procedures described in this work.

References

  1. [1]T. A. Brody, J. Flores, J. B. French, P. A. Mello, A. Pandey, S. S. M. Wong, Random-matrix physics: spectrum and strength fluctuations, Rev. Mod. Phys. 53 (1981) 385–479. doi:10.1103/RevModPhys.53.385.DOI
  2. [2]H.-J. Stöckmann, Quantum chaos: an introduction, Cambridge University Press (1999). doi:10.1017/CBO9780511524622.DOI
  3. [3]Y. Alhassid, The statistical theory of quantum dots, Rev. Mod. Phys. 72 (2000) 895–968. doi:10.1103/RevModPhys.72.895.DOI
  4. [4]R. L. Weaver, Spectral statistics in elastodynamics, J. Acoust. Soc. Am. 85 (1989) 1005–1013. doi:10.1121/1.397484.DOI
  5. [5]L. Laloux, P. Cizeau, J.-P. Bouchaud, M. Potters, Noise dressing of financial correlation matrices, Phys. Rev. Lett. 83 (1999) 1467–1470. doi:10.1103/PhysRevLett.83.1467.DOI
  6. [6]H. L. Montgomery, The pair correlation of zeros of the zeta function, Proc. Symp. Pure Math. 24 (1973) 181–193.DOI
  7. [7]A. M. Odlyzko, On the distribution of spacings between zeros of the zeta function, Math. Comp. 48 (1987) 273–308.DOI
  8. [8]M. V. Berry, Riemann’s zeta function: a model for quantum chaos?, Lecture Notes in Physics 263 (1986) 1–17. doi:10.1007/3-540-17171-1_1.
  9. [9]E. B. Bogomolny, J. P. Keating, Gutzwiller’s trace formula and spectral statistics: beyond the diagonal approximation, Phys. Rev. Lett. 77 (1996) 1472–1475. doi:10.1103/PhysRevLett.77.1472.DOI
  10. [10]J. P. Keating, N. C. Snaith, Random matrix theory and ζ(1/2 + it), Commun. Math. Phys. 214 (2000) 57–89. doi:10.1007/s002200000261.DOI
  11. [11]J. Sakhr, J. M. Nieminen, Poisson-to-Wigner crossover transition in the nearest-neighbor statistics of random points on fractals, Physical Review E 72 (2005) 045204. doi:10.1103/PhysRevE.72.045204.DOI
  12. [12]J. Illian, A. Penttinen, H. Stoyan, D. Stoyan, Statistical Analysis and Modelling of Spatial Point Patterns, Wiley, 2008. doi:10.1002/9780470725160.DOI
  13. [13]A. Baddeley, E. Rubak, R. Turner, Spatial Point Patterns: Methodology and Applications with R, CRC Press, 2015. doi:10.1201/b19708.DOI
  14. [14]S. N. Chiu, D. Stoyan, W. S. Kendall, J. Mecke, Stochastic Geometry and its Applications, 3rd Edition, Wiley, 2013. doi:10.1002/9781118658222.DOI
  15. [15]J. Møller, R. P. Waagepetersen, Statistical inference for Cox processes, in: Spatial Cluster Modelling, Chapman and Hall/CRC, 2002, pp. 37–60.
  16. [16]M. Massaro, A. D. Campo, Spatial form factor for point patterns: Poisson point process, Coulomb gas, and vortex statistics, Physical Review Research 7 (2025) 023107. doi:10.1103/physrevresearch.7.023107.DOI
  17. [17]J. U. Yang, Y. Guan, Fourier analysis of spatial point processes, Bernoulli 32 (2026) 1–28. doi:10.3150/25-bej1862.DOI
  18. [18]F. Helmli, Focus variation instruments, Optical Measurement of Surface Topography (2011) 131–166doi:10.1007/978-3-642-12012-1_7.DOI
  19. [19]D. Kucharski, M. Wieczorowski, Radial image processing for phase extraction in rough-surface interferometry, Measurement 242 (2025) 117102. doi:10.1016/j.measurement.2025.117102.DOI
  20. [20]G. H. Hardy, E. M. Wright, An Introduction to the Theory of Numbers, 6th Edition, Oxford University Press, 2008.
  21. [21]M. Stein, S. M. Ulam, An observation on the distribution of primes, Amer. Math. Monthly 74 (1967) 43–44. doi:10.2307/2314055.DOI
  22. [22]A. Rényi, On measures of entropy and information, Proc. 4th Berkeley Symp. Math. Stat. Prob. 1 (1961) 547–561.
  23. [23]H. G. E. Hentschel, I. Procaccia, The infinite number of generalized dimensions of fractals and strange attractors, Physica D 8 (1983) 435–444. doi:10.1016/0167-2789(83)90235-x.DOI
  24. [24]P. Grassberger, I. Procaccia, Characterization of strange attractors, Phys. Rev. Lett. 50 (1983) 346–349. doi:10.1103/PhysRevLett.50.346.DOI
  25. [25]J. Theiler, Statistical precision of dimension estimators, Phys. Rev. A 41 (1990) 3038–3051. doi:10.1103/physreva.41.3038.DOI
  26. [26]R. Danzl, F. Helmli, S. Scherer, Focus variation—a robust technology for high resolution optical 3D surface metrology, Strojniški vestnik – Journal of Mechanical Engineering 57 (2011) 245–256. doi:10.5545/sv-jme.2010.175.DOI
  27. [27]J. Ginibre, Statistical ensembles of complex, quaternion, and real matrices, Journal of Mathematical Physics 6 (1965) 440–449. doi:10.1063/1.1704292.DOI
  28. [28]P. J. Forrester, Log-Gases and Random Matrices, Princeton University Press, 2010. doi:10.1515/9781400835416.
  29. [29]S. Torquato, Random Heterogeneous Materials: Microstructure and Macroscopic Properties, Springer, 2002. doi:10.1007/978-1-4757-6355-3.
  30. [30]C. R. Harris, et al., Array programming with NumPy, Nature 585 (2020) 357–362. doi:10.1038/s41586-020-2649-2.DOI
  31. [31]P. Virtanen, et al., SciPy 1.0: fundamental algorithms for scientific computing in Python, Nat. Methods 17 (2020) 261–272. doi:10.1038/s41592-019-0686-2.DOI
  32. [32]J. D. Hunter, Matplotlib: a 2D graphics environment, Comput. Sci. Eng. 9 (2007) 90–95. doi:10.1109/MCSE.2007.55.DOI
  33. [33]F. Johansson, et al., mpmath: a Python library for arbitrary-precision floating-point arithmetic (2023). URL http://mpmath.org

Paper details

vlnc:W_bgn888wf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents