Density Random Points Models

• Density random points models use parameters (e.g., z, λ, f) to control point generation, spatial structure, and connectivity.
• They influence phenomena such as percolation, topological transitions, and optimized sampling in statistical mechanics and geometric inference.
• Applications include variable-density Poisson-disc sampling, random geometric graph construction, and the study of density-induced metric deformations.

Density random points comprise a family of models in which randomness is assigned to point configurations while a density-like quantity controls either the generation of points, the geometry induced by the sample, or the observables extracted from the configuration. Across continuum statistical mechanics, random geometric graphs, discrepancy theory, stochastic topology, and geometric inference, the relevant density may appear as an activity parameter $z$, an intensity $\lambda$, a sampling density $f$ or $\rho$, a product density $\varrho_d$, or a density functional $\psi_k(t)$ defined from coverings of a point set. The common theme is that point density is not merely descriptive: it determines connectivity, effective metric structure, regularity, inferential identifiability, and asymptotic phase behavior (Aristoff, 2012, Hwang et al., 2012, Novak et al., 9 Dec 2025).

1. Density as a control parameter for random point configurations

In continuum point-process models, density is often encoded indirectly through an intensity-like or activity parameter rather than by an explicit count-per-volume identity. For the hard-disk Gibbs model, admissible configurations are

$$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$

so the point process is constrained by hard-core exclusion. In finite volume $\Lambda_n=[-n,n]^2$, the Gibbs distribution is

$$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$

and the factor $z^N$ favors configurations with more particles. The parameter $\lambda$0 therefore plays the same qualitative role as the intensity $\lambda$1 of a Poisson point process, while remaining moderated by exclusion (Aristoff, 2012).

This distinction between independent and interacting random points recurs in several settings. In Poisson models, disjoint sets are independent and intensity directly determines mean density. In hard-core or Poisson-disc constructions, density is coupled to a minimum-separation rule, so higher density must be achieved subject to repulsion or non-overlap (Aristoff, 2012, Dwork et al., 2020). In the three-dimensional random-packing framework based on quenched Poisson points $\lambda$2, randomness enters through a Poisson random point field, but the admissible configuration $\lambda$3 is chosen by minimizing squared displacement subject to the hard-sphere constraint

$\lambda$4

In the strong-coupling limit $\lambda$5, this becomes a constrained optimization problem defining the packing closest to the quenched Poisson configuration (Song, 2023).

A related but algorithmic use of density appears in variable-density Poisson-disc sampling. There, the point set is random, but the local exclusion radius $\lambda$6 varies spatially, so density is highest where $\lambda$7 is smallest. The specific radius function

$\lambda$8

produces samples dense near the origin and sparser toward the corners, with

$\lambda$9

This is a direct way to prescribe nonuniform random point density under a minimum-distance constraint (Dwork et al., 2020).

These models show that “density random points” does not denote a single formalism. It encompasses at least three distinct but related notions: density as an intensity parameter in a stochastic process, density as a target spatial law for sampling, and density as an emergent descriptor of how random points occupy space.

2. Connectivity, percolation, and jamming at high density

A central question is when a density-controlled random point configuration produces a macroscopic connected structure. In the hard-disk model, the relevant occupied set is

$f$0

equivalently a graph connecting points within distance $f$1. The main theorem states that if $f$2, then for sufficiently large $f$3, every Gibbs measure $f$4 satisfies

$f$5

where $f$6 is the event that the union of radius-$f$7 disks has an infinite connected component. The proof proceeds through a Peierls-type contour argument: finite isolated components would create contours, but at high activity large empty gaps are probabilistically suppressed, so contour probabilities decay exponentially in contour size (Aristoff, 2012).

In random Čech complexes built on a Poisson point cloud with intensity $f$8, density enters through the scaling parameter $f$9. For the distance-to-set function

$\rho$0

the number $\rho$1 of index-$\rho$2 critical points with critical value at most $\rho$3 exhibits three asymptotic regimes: subcritical when $\rho$4, critical when $\rho$5, and supercritical when $\rho$6. In the dense regime the Čech complex becomes asymptotically contractible, while the critical-point counts still scale linearly in $\rho$7 and satisfy Gaussian fluctuation theory (Bobrowski et al., 2011).

Three-dimensional random packing supplies an out-of-equilibrium analogue of a density-driven transition. With packing fraction $\rho$8, the framework identifies a sharp crossover at

$\rho$9

Above this threshold, the system responds through local rearrangements and satisfies

$\varrho_d$0

where $\varrho_d$1 is the mean-square displacement from the quenched points and $\varrho_d$2 is the reduced pressure. Below $\varrho_d$3, this relation fails and global rearrangements occur. The transition is supported by the behavior of the kissing number $\varrho_d$4, which shows a plateau near $\varrho_d$5 and satisfies $\varrho_d$6 at $\varrho_d$7, matching the Maxwell isostatic criterion $\varrho_d$8 in $\varrho_d$9 (Song, 2023).

Taken together, these results indicate that increasing density can drive qualitatively distinct phenomena: infinite-cluster formation in interacting continuum percolation, topological saturation in random geometric complexes, and jamming in constrained packings. A plausible implication is that density serves as a unifying order parameter across both equilibrium and nonequilibrium random-point models, even when the mechanisms—overlap, coverage, or contact—isostaticity—are different.

3. Density-induced geometry and random path optimization

When random points are sampled from a nonuniform density, the induced discrete geometry need not approximate the ambient Euclidean or Riemannian geometry; it may converge to a density-deformed metric. For i.i.d. points $\psi_k(t)$0 on a complete $\psi_k(t)$1-dimensional Riemannian manifold $\psi_k(t)$2 with smooth density $\psi_k(t)$3, the power-weighted path length

$\psi_k(t)$4

defines a shortest-path functional $\psi_k(t)$5. The limiting metric tensor is

$\psi_k(t)$6

with induced distance

$\psi_k(t)$7

Under compactness and positivity assumptions, there exists $\psi_k(t)$8 such that

$\psi_k(t)$9

Because $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$0, high-density regions have lower effective cost, so asymptotic shortest paths are attracted to dense regions (Hwang et al., 2012).

A broader approximation theory for geodesics via random points considers i.i.d. samples in a bounded convex domain $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$1 with density $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$2 satisfying

$$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$3

A random geometric graph connects $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$4 and $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$5 when $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$6, with $$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$7 and

$$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$8

For continuum path costs

$$\Omega=\{\omega\subset\mathbb{R}^2:\ |x-y|\ge 2r\ \forall x\neq y\in\omega\},$$9

the paper introduces three discrete approximations—linear interpolation $\Lambda_n=[-n,n]^2$0, quasinormal interpolation $\Lambda_n=[-n,n]^2$1, and a Riemann-sum cost $\Lambda_n=[-n,n]^2$2—and proves almost sure convergence of minimum costs and, under the stated hypotheses, convergence of minimizing paths to continuum geodesics in uniform and Hausdorff senses (Davis et al., 2017).

These two lines of work address related but distinct geometric questions. The first shows that density changes the limiting metric itself. The second shows that a dense enough random sample from a bounded positive density supports probabilistic $\Lambda_n=[-n,n]^2$3-convergence of discrete path problems to continuum variational problems. This suggests that sampling density is not merely a nuisance parameter for graph construction; it enters directly into the asymptotic geometry of optimization over random points.

4. Regularity, discrepancy, and density-optimized sampling

In discrepancy theory, random points are often assessed by how well their empirical distribution matches a target measure. For i.i.d. uniform points $\Lambda_n=[-n,n]^2$4, the one-dimensional star-discrepancy is

$\Lambda_n=[-n,n]^2$5

The standard high-probability bound is

$\Lambda_n=[-n,n]^2$6

However, if one may delete at most $\Lambda_n=[-n,n]^2$7 points, there exists a subset $\Lambda_n=[-n,n]^2$8 with $\Lambda_n=[-n,n]^2$9 such that, with high probability,

$$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$0

The construction is online and proceeds by binning, thinning each bin to equal occupancy, and applying the Dvoretzky–Kiefer–Wolfowitz inequality to the residual within-bin process. By the probability integral transform, the same conclusion extends to any absolutely continuous distribution on $$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$1 (Bilyk et al., 23 Jan 2025).

A higher-dimensional variant modifies not the sample by deletion but the sampling law itself. For the Sobolev space $$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$2, generalized $$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$3-discrepancy is defined by

$$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$4

Instead of uniform sampling, the points are drawn from a product density $$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$5 and weighted by

$$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$6

For the optimal product density $$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$7, the expected generalized discrepancy satisfies

$$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$8

improving the uniform random-point factor

$$G_{n,z,\zeta}(A)=\frac{L_{n,z}(A\cap \Omega_{n,\zeta})}{L_{n,z}(\Omega_{n,\zeta})}, \qquad L_{n,z}(A)=\sum_{N=1}^\infty z^N L_{n,N}(A),$$9

to

$z^N$0

For $z^N$1, the optimal one-dimensional density is explicit: $z^N$2 The interpretation given is importance sampling for $z^N$3, but the curse of dimensionality persists because the base in $z^N$4 remains greater than $z^N$5 (Novak et al., 9 Dec 2025).

Variable-density Poisson-disc generation occupies a different point on the regularity spectrum. There the goal is not empirical-distribution discrepancy but locally regular random spacing. A background grid with cell side $z^N$6 stores indices of points that might affect a candidate; each active point spawns candidates in an annulus between radii $z^N$7 and $z^N$8, and only local conflicts are checked. The method reports runtime improvements of roughly $z^N$9 over the earlier variable-density approach by Tulleken, while producing similar-quality distributions for compressed sensing applications (Dwork et al., 2020).

These results separate three notions that are often conflated. Randomness does not imply irregularity; regularity can be improved by thinning an i.i.d. sample, by choosing a nonuniform importance density and compensating weights, or by imposing local exclusion radii. Conversely, a point set can remain random while being substantially more structured than an unconstrained Poisson or i.i.d. sample.

5. Density estimation and reconstruction from incomplete or transformed random samples

Density random points also arise in inference problems where the observed points are transformed, partially matched, or generated through prescribed density maps. In the batched broken random sample problem, each batch $\lambda$00 contains i.i.d. pairs

$\lambda$01

but only the unordered multisets

$\lambda$02

are observed. The exact permutation-invariant likelihood averages over all $\lambda$03,

$\lambda$04

but is generally infeasible for large $\lambda$05. The proposed pseudo-log-likelihood is

$\lambda$06

with estimator

$\lambda$07

As $\lambda$08, the population objective satisfies

$\lambda$09

so the limiting criterion becomes an $\lambda$10-projection rather than a KL objective. The concentration bounds and estimator convergence rates depend on the number of batches $\lambda$11 and are uniform in $\lambda$12 (Bi et al., 10 Feb 2026).

A different reconstruction paradigm appears in particle generation for general relativistic SPH. There, the normalized rest-mass density on a hypersurface,

$\lambda$13

is treated as a probability density. Under the separability assumptions

$\lambda$14

inverse transform sampling is applied coordinatewise via

$\lambda$15

with $\lambda$16 uniform. The resulting random point cloud is then relaxed by Lloyd’s algorithm in 3D using Voronoi centroids computed with Voro++, according to

$\lambda$17

The reported $\lambda$18 error scaling exponents are $\lambda$19 for random configurations and $\lambda$20 for Voronoi-relaxed configurations, with the relaxed configurations having a smaller prefactor (Pablo et al., 2013).

Both settings concern density estimation or representation under information loss or geometric transformation. One addresses lost pairings in random samples; the other maps uniform random points to a target density and then regularizes them geometrically. A plausible implication is that density-aware modeling can compensate either for combinatorial ambiguity in observation or for excessive noise in direct Monte Carlo placement.

6. Density observables derived from point sets

In some problems, density is not the sampling law but the observable computed from the point configuration. For periodic sequences on the line,

$\lambda$21

the $\lambda$22-th density function is defined by the fraction of a unit cell covered by exactly $\lambda$23 intervals of radius $\lambda$24: $\lambda$25 In one dimension each $\lambda$26 is piecewise linear. The function $\lambda$27 depends only on the sorted cyclic gaps, whereas for $\lambda$28,

$\lambda$29

a sum of trapezoid functions determined by local triples $\lambda$30. The full family obeys the symmetries

$\lambda$31

so only finitely many functions are needed to recover the infinite fingerprint (Anosova et al., 2022).

A conceptually different density observable is the expected density of complex critical points of random polynomials. For

$\lambda$32

with real $\lambda$33 or complex $\lambda$34 Gaussian coefficients, the expected critical-point density for the real ensemble satisfies, for $\lambda$35,

$\lambda$36

where

$\lambda$37

In one variable, the complex $\lambda$38 critical-point density is given exactly by

$\lambda$39

and the real ensemble differs by an exponentially small correction away from the real axis (Macdonald, 2010).

Distance distribution problems provide yet another density functional. For two uniformly and independently distributed points in a right-angled triangle, the distance density $\lambda$40 is derived from the chord length distribution using Piefke’s formula,

$\lambda$41

or equivalently

$\lambda$42

Here the density is a distribution over pairwise distances induced by a uniform random point law on a domain, not a spatial point density per se (Bäsel, 2012).

These examples underscore a persistent ambiguity in the term “density.” It may refer to a sampling law, a density of geometric events, or a coverage profile of a point configuration. The literature treats these as formally distinct objects, even when all arise from the same underlying random points.

7. Conceptual scope and recurrent themes

Several recurrent principles unify these otherwise heterogeneous models. First, density frequently deforms geometry. In random shortest-path problems, the conformal factor $\lambda$43 changes the limiting metric itself (Hwang et al., 2012). In geodesic approximation on random graphs, bounded positive density is a structural assumption ensuring graph-level convergence to continuum variational geometry (Davis et al., 2017).

Second, density regulates phase behavior. High activity $\lambda$44 yields percolation of hard disks when $\lambda$45 (Aristoff, 2012); the scaling $\lambda$46 controls the topological regime of random Čech complexes (Bobrowski et al., 2011); and the packing fraction $\lambda$47 governs the onset of local-versus-global rearrangement in random packing, with $\lambda$48 in three dimensions (Song, 2023).

Third, density can be optimized or corrected algorithmically. Nonuniform product densities with inverse-density weights improve generalized discrepancy bounds (Novak et al., 9 Dec 2025); deleting at most $\lambda$49 points can regularize a random sample to discrepancy $\lambda$50 in one dimension (Bilyk et al., 23 Jan 2025); and variable-density Poisson-disc sampling imposes local exclusion to generate random patterns that are both nonuniform and spatially regular (Dwork et al., 2020).

Fourth, inferential tractability depends strongly on how density information is observed. Broken random samples lose pairings within a batch, forcing pseudo-likelihood methods and altering the asymptotic population objective from KL to $\lambda$51 form (Bi et al., 10 Feb 2026). In SPH initial-data generation, the same density may instead be used constructively as a coordinatewise pdf, then refined by centroidal Voronoi relaxation (Pablo et al., 2013).

A common misconception is that higher density always means stronger disorder or clustering. The cited work shows the opposite can occur: larger $\lambda$52 in a repulsive hard-core model still produces macroscopic connectivity (Aristoff, 2012); optimal nonuniform densities can reduce discrepancy rather than exacerbate irregularity (Novak et al., 9 Dec 2025); and Lloyd relaxation can preserve a target density while reducing random noise (Pablo et al., 2013). Another misconception is that density is synonymous with intensity. This is accurate for homogeneous Poisson processes, but only heuristic for interacting Gibbs models, density-deformed metrics, and coverage-based density fingerprints.

In aggregate, density random points form a research domain in which probabilistic sampling laws, geometric constraints, and density-derived observables interact at multiple scales. The resulting theory links continuum percolation, random optimization, discrepancy minimization, topological phase transitions, and statistical recovery, with density functioning alternately as input parameter, asymptotic scaling variable, variational weight, and output invariant.