Low Density Positions in Diverse Fields

• Low density positions are context-driven constructs defined as regions or points with underdensity relative to a domain-specific baseline.
• They are employed in cosmology, galaxy topology, and muography to detect weak-lensing centers, empty cells, and geological anomalies, respectively.
• In labor-market NLP, low density positions reveal sparse yet semantically coherent job clusters that predict emerging occupational trends.

“Low density positions” does not denote a single invariant object across the technical literature. In the supplied sources, it functions as an operational label for locations, cells, embedding points, or density sectors that are underdense relative to a chosen reference construction. In observational cosmology, it refers to projected sky positions far from bright foreground galaxies and used as weak-lensing centers (Huang et al., 14 Jul 2025). In labor-market NLP, it denotes HDBSCAN noise or outlier postings in embedding space, interpreted as sparse but potentially semantically coherent precursors of occupational emergence (Rawat, 22 Jun 2026). In 2D galaxy topology, low-density regions are empty hexagonal cells in a balanced counts-in-cells analysis (Wu et al., 2013). In muographic inversion, low-density zones are cavities or fractured regions with reduced density-length relative to intact host rock (Balázs et al., 2023). By contrast, some papers warn that the phrase would be misleading if read literally: finite-temperature amorphous-solid vibrational anomalies are tied to thermally averaged particle positions rather than to any identified low-density structural motif (Das et al., 2020).

1. Semantic scope and operational definitions

In the supplied literature, the term is best understood as a family of context-specific constructions. The decisive question is not whether the underlying medium is “objectively sparse,” but how a method identifies underdensity relative to a domain-specific baseline.

Context	Operational object	Low-density criterion
Weak lensing	Sky positions on a HEALPix grid	Sufficiently far from bright foreground galaxies
Labor-market NLP	Embedding-space noise points or groups	HDBSCAN label $-1$ and sparse local neighborhood
Galaxy topology	Hexagonal cells	Empty cell, i.e. no galaxies in the cell
Muography	Voxels or fractured zones	Reduced density-length relative to homogeneous rock
Amorphous solids	Not a valid structural label in the paper’s sense	Low-frequency modes are not attributed to low-density regions

This heterogeneity has methodological consequences. In some settings, low density is a signal to be exploited rather than discarded; this is explicit both in weak lensing around underdense positions and in density-based outlier analysis of job-posting embeddings. In other settings, the phrase must be rejected as an interpretive shortcut. The amorphous-solids study is especially explicit: it links universal quasi-localized modes to the vibrational spectrum around stable thermally averaged configurations, not to low-density positions or void-like motifs (Das et al., 2020).

2. Projected low-density positions in weak-lensing cosmology

In the cosmological usage, low density positions are projected sky locations constructed from the DESI Legacy Imaging Survey DR9. The foreground sample is defined by an $r$-band absolute-magnitude cut $M_r < -21.5$ and photometric-redshift range $0.18 < z < 0.28$. For any sky position, $R_s$ is the shortest projected angular distance to any foreground bright galaxy, and the analysis uses three LDP bins,

$$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$

The positions are placed on a HEALPix grid with $N_{\rm side}=4096$, and the weak-lensing observable is the excess surface density

$$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$

The construction is explicitly intended as a practical alternative to traditional 3D void finders. It uses only projected galaxy positions and photo-$z$, yields many more usable regions than strict void catalogs, and can be reproduced directly in SHAM-based mock catalogs. The measurement pipeline combines LDP lensing with galaxy-galaxy lensing and the galaxy-galaxy two-point correlation function, while jointly fitting $\sigma_{\lg M}$, $r$0, and $r$1 under a Gaussian likelihood with simulation-based model predictions.

The reported quantitative result is that lensing around low density positions is detected at $r$2, which is the largest signal-to-noise ratio among the three probes used in the fit. Including LDP lensing improves the $r$3 constraint by $r$4 relative to galaxy-galaxy lensing and galaxy-galaxy clustering alone. For flat $r$5CDM with other cosmological parameters fixed to Planck values, the combined analysis yields

$r$6

with $r$7 and $r$8 (Huang et al., 14 Jul 2025).

This usage makes “low density positions” a deliberately observational construct. They are not centers of formally identified voids, but projected underdense locations defined by exclusion from bright-galaxy neighborhoods. A plausible implication is that their value lies precisely in this operational simplicity: they probe underdense structure without requiring a full 3D void-finding pipeline.

3. Empty-cell topology and diffuse structure in galaxy surveys

A second astronomical usage appears in 2D topological analysis of 2MASS. The sky is projected with an equal-area Lambert azimuthal projection and tiled with hexagonal cells. A filled cell contains one or more galaxies and is treated as a high-density region; an empty cell contains no galaxies and is treated as a low-density region. The cell size is adjusted so that the numbers of filled and empty cells are equal. In the northern 2MASS sky, after trimming boundary cells, the analysis uses 172,320 filled cells and 172,320 empty cells.

The two principal statistics are the like-neighbor count $r$9, where

$M_r < -21.5$0

and the connectivity class $M_r < -21.5$1, where $M_r < -21.5$2 is the number of connected groups of inner edges and $M_r < -21.5$3 the number of connected groups of outer edges. The random baseline is binomial with $M_r < -21.5$4, giving

$M_r < -21.5$5

for the $M_r < -21.5$6 distribution, and

$M_r < -21.5$7

for $M_r < -21.5$8, $M_r < -21.5$9, $0.18 < z < 0.28$0, $0.18 < z < 0.28$1, and $0.18 < z < 0.28$2. Departures are summarized by

$0.18 < z < 0.28$3

The empirical conclusion is that low-density and high-density regions are topologically distinct, with both showing significant $0.18 < z < 0.28$4 shifts from the binomial or random baseline. Filled cells are more strongly clustered than empty cells, and empty cells are interpreted as having less fiber structure than filled ones. The method therefore treats low-density positions not as absence of information but as a topological object with its own measurable connectivity statistics (Wu et al., 2013).

A related large-scale-structure study begins from 7,596 nearby galaxies and identifies 3,168 field galaxies, i.e. $0.18 < z < 0.28$5 of the sample, after removing galaxies already assigned to pairs, groups, or clusters. Applying friends-of-friends percolation with $0.18 < z < 0.28$6 Mpc yields 226 diffuse agglomerates with $0.18 < z < 0.28$7. The richest systems have a median radial-velocity dispersion of about $0.18 < z < 0.28$8, linear size around $0.18 < z < 0.28$9 Mpc, integral $R_s$0-band luminosity of $R_s$1, formal virial-mass-to-luminosity ratio of about $R_s$2, mean density contrast only $R_s$3, and crossing times of $R_s$4–$R_s$5 Gyr. These are explicitly characterized as non-virialized diffuse agglomerates rather than relaxed clusters (Karachentsev et al., 2012).

Taken together, these works show two distinct astronomical meanings. In one, low-density positions are discrete empty cells in a topology test. In the other, low-density environments are diffuse agglomerates embedded in the field population. The common element is relational underdensity, but the mathematical objects differ: occupancy cells in one case, percolated galaxy associations in the other.

4. Low-density zones in muographic inversion

In muography, low-density positions are geological density-decrease anomalies embedded in otherwise denser rock. The target structures include cavities or voids, low-density fractured zones, and karstic crack zones. The relevant observable is the density-length

$R_s$6

which converts muon attenuation into a line integral of density. After discretizing the subsurface into voxels with densities $R_s$7, the forward model becomes

$R_s$8

where $R_s$9 is the path length of trajectory $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$0 through voxel $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$1.

Because the inverse problem is a cone-beam limited-angle transmission tomography problem with more voxels than measurements, the reconstruction is underdetermined and ill-conditioned. The paper therefore uses linearized density-length inversion and a Bayesian MAP objective,

$$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$2

with a homogeneous-rock Gaussian prior. The first-order estimate is

$$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$3

and the prior-induced bias toward the homogeneous model is made explicit. Because the detector positions lie along a straight tunnel, the full 3D inversion is reduced to a set of independent 2D planes, described as a 2+1D reduction.

The synthetic benchmark uses a 3 m diameter zero-density cavity at 20 m depth, with detectors 50 m below a flat surface. In the noise-free case, the cavity is detected but partially filled in by the prior, with radial artifacts along projection lines. In a noisy one-month case, the anomaly remains visible, and detectability is reported at about the $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$4 confidence level. The field application uses a high-quality dataset from 7 positions, a $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$5 detector, $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$6 angular bins, and about one month per position. The reconstructed anomaly is not a simple empty cavity but a complex karstic crack or fissure zone. Drill holes of 5.4 m, 5.8 m, and 9.2 m validate the presence of large fissures filled with altered dolomite powder of density about $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$7 or lower, compared with intact cherty dolomite at $$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$8–$$\text{I: } 2' < R_s \le 5', \qquad \text{II: } 5' < R_s \le 7', \qquad \text{III: } R_s > 7' .$$9, and place the contact between high- and low-density zones within about 20 cm of the muographic prediction (Balázs et al., 2023).

Here, low-density positions are neither binary empty cells nor projected lensing centers. They are volumetric anomalies inferred from line-integral data, and their significance depends on inversion geometry, prior structure, and external geological validation.

5. Low-density positions in embedding space and occupational emergence

A markedly different usage appears in labor-market text mining. Standard clustering pipelines based on HDBSCAN often discard the 10–15% of job postings assigned to the noise class. The paper argues that this is a mistake in rapidly evolving domains, because low posting density may indicate novelty rather than incoherence. This is formalized as the Emergence-Density Inversion hypothesis: for genuinely new occupations, early postings fail to form clusters not because they are semantically incoherent, but because their volume is below the minimum cluster size.

The pipeline embeds job descriptions with INSTRUCTOR-xl, reduces them with UMAP, and clusters with HDBSCAN. Points labeled $N_{\rm side}=4096$0 are treated as noise. A secondary HDBSCAN is then applied to the noise points within each quarter, with min_cluster_size = 5, and only groups with $N_{\rm side}=4096$1 are retained. A noise group is considered emerged if a later stable HDBSCAN cluster satisfies cluster-group SDC $N_{\rm side}=4096$2 and group overlap $N_{\rm side}=4096$3.

The ranking statistic is the Emerging Occupation Score. Its base form is

$N_{\rm side}=4096$4

The extended version adds Temporal Velocity and Cross-Platform Convergence, and a learned-weight logistic regression assigns the largest coefficients to Temporal Velocity $N_{\rm side}=4096$5 and Cross-Platform Convergence $N_{\rm side}=4096$6, with Cohesion $N_{\rm side}=4096$7, Novelty $N_{\rm side}=4096$8, Distinctiveness $N_{\rm side}=4096$9, TaxGap $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$0, and Intercept $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$1.

The dataset contains 84,988 English-language job postings over 8 quarters from Q4 2022 to Q3 2024. Among 412 noise groups with $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$2, 87 eventually emerged, i.e. $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$3. Emerged groups have SDC $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$4, versus $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$5 for non-emerged groups, with Mann–Whitney $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$6, $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$7, and AUC $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$8 for SDC alone. The headline longitudinal result is that high-EOS outlier groups transition to stable clusters in $$\Delta\Sigma(R)=\gamma_t \Sigma_c = \bar\Sigma(\le R)-\Sigma(R).$$9 quarters, versus $z$0 for low-EOS groups, with $z$1. At the 2-quarter horizon, LR-EOS reaches $z$2 and AUC $z$3, outperforming Isolation Forest, LOF, GLOSH, BERTrend, and frequency count. A held-out annotator panel with $z$4 finds that EOS $z$5 corresponds to coherent emerging occupations with $z$6 precision and $z$7 recall. Prompt Engineer, AI Safety Researcher, Foundation Model Engineer, and Agent Systems Engineer are reported as the top-4 candidates in Q3 2024, absent from O*NET and forming stable clusters by Q1 2025 (Rawat, 22 Jun 2026).

In this setting, low-density positions are sparse neighborhoods in embedding space. Their epistemic status is inverted relative to ordinary anomaly detection: density extremity alone is uninformative, but semantic cohesion within the sparse region can be predictive of future cluster formation. This suggests a broader methodological principle: underdensity can be structurally informative when the clustering threshold is stricter than the underlying semantic coherence.

6. Related but distinct low-density concepts

Several supplied papers use “low density” in ways that are adjacent to, but not equivalent to, “low density positions.” The distinction is important because conflating them obscures the object under study.

In finite-temperature amorphous solids, the relevant structural entity is the thermally averaged configuration,

$z$8

with covariance

$z$9

and effective Hessian

$\sigma_{\lg M}$0

The universal law

$\sigma_{\lg M}$1

is shown to persist up to $\sigma_{\lg M}$2, provided the average configuration remains stable long enough for the calculations to converge. The paper explicitly states that interpreting the result in terms of low-density positions would be misleading; the controlling ingredient is a stable thermally averaged cage structure, not a density-based motif (Das et al., 2020).

In neutron-star EOS inference, “low density” refers to the matching sector below

$\sigma_{\lg M}$3

not to spatial positions. Branch A matches a low-density input to a $\sigma_{\lg M}$4EFT branch through parameters $\sigma_{\lg M}$5 and $\sigma_{\lg M}$6, while Branch B continues the low-density input directly to $\sigma_{\lg M}$7. Current NICER-based inference mainly constrains the shared continuation above $\sigma_{\lg M}$8, but also indirectly restricts the low-density matching sector, with posterior-weighted values $\sigma_{\lg M}$9 and $r$00 in Branch A (Nola, 5 Jun 2026).

Other low-density usages are similarly non-positional. A zero-temperature FNDMC study of $r$01He monolayers on graphite finds a gas-liquid phase transition around $r$02, with coexistence windows $r$03 for clean graphite and $r$04 for graphite preplated with $r$05He (Ruggeri et al., 2015). In the sharp-interface unscreened Ohta–Kawasaki or liquid-drop model, the low-volume-fraction regime is specified by $r$06, with $r$07-convergence to a limit functional whose minimizer is the uniform measure $r$08; the result concerns asymptotic multi-droplet phases rather than localized low-density positions (Knuepfer et al., 2015). In dilute Hubbard systems near a Lifshitz transition, low-density ferromagnetism is shown to depend not on band-minimum degeneracy alone but on the singularity of the density of states at the band bottom, with the one-diagonal model losing its apparent low-density ferromagnetism in the thermodynamic limit while the two-diagonal model retains it (Wang et al., 15 Jun 2026). In hard-sphere geometry, Dionysian packings show that strictly jammed packings can satisfy $r$09 while maintaining finite bulk and shear moduli, so the minimum density of mechanically stable purely repulsive sphere packings is zero in 2D and 3D (Dennis et al., 2020).

The common lesson is terminological rather than substantive: “low density” may describe positions, cells, phases, matching branches, asymptotic regimes, or band-bottom limits. Only some of these usages refer to localized underdense positions in the geometric or embedding-space sense. Where the literature is explicit, that distinction should be preserved.