Neighbourhood-Exclusion Framework

• The Neighbourhood-Exclusion Framework is a formal pattern where local mechanisms—such as affordability gradients, homophily, and local constraints—systematically restrict access, leading to macro-level reorganization.
• In urban models, the framework quantifies exclusion using thresholded dissatisfaction indices and asymmetric entry/exit rules, accurately predicting phenomena like ghetto formation from local compositional disparities.
• Extensions of the framework apply to binary constraint satisfaction and continuous Bayesian inference, demonstrating its versatility in measuring exclusion through local loss evaluations and connectivity analyses.

Searching arXiv for papers directly relevant to the "Neighbourhood-Exclusion Framework" and its main instantiations. The Neighbourhood-Exclusion Framework is a family of formalizations in which exclusion is generated by local rules defined on a neighborhood, adjacency relation, or excluded vicinity, and then propagated to larger-scale structure. In the urban segregation literature, it denotes mechanisms by which disadvantaged groups are prevented from entering, remaining in, or co-residing within particular areas because of tolerance thresholds, affordability gradients, homophily, low exposure, or weak connectivity (Ortega et al., 2022, Sadurní et al., 27 Jan 2025, Neira et al., 20 May 2025, Zhang et al., 20 Jun 2026). In other domains, the same label has been used for local elimination in binary constraint satisfaction problems and for inferential loss induced by removing a neighborhood in continuous parameter spaces (Cooper, 2020, Villa, 21 Apr 2026). Across these settings, the common structure is local exclusion plus global reorganization: tract-level admissions generate ghettos, co-residence deficits generate clustered communities, component-level divergence plus weak connectivity generates multiscalar exclusion, value elimination shrinks CSP domains, and excluded parameter regions induce objective priors.

1. Conceptual core

In the urban and spatial literature, neighborhood exclusion is typically defined as systematic underrepresentation, low exposure, or denied access at the local level. One formulation treats exclusion as the condition in which population groups are under-represented relative to a wider reference and/or poorly connected to higher-level opportunities, yielding coupled compositional and relational dimensions (Neira et al., 20 May 2025). Another formulation defines it through low inter-group co-residence probabilities relative to intra-group concentrations and homophily, with exclusion visible in low cross-group exposure, high isolation, and dense within-cluster ties (Sadurní et al., 27 Jan 2025). A third formulation shifts from residence to daily behavior and defines exposure segregation as the lack of shared activity spaces, with built-form cues such as fences, gated enclosures, or monofunctional zoning acting as physical regulators of exclusion (Zhang et al., 20 Jun 2026).

In the Schelling-derived formulations, exclusion is operationalized directly as an entry and exit asymmetry. Agents compare local composition and place-specific affordability against a tolerance threshold, and admission occurs only when dissatisfaction is non-positive. This produces systematic affordability-based exclusion of disadvantaged agents from expensive neighborhoods and concentration in affordable zones (Ortega et al., 2022). Related lattice formulations similarly combine local diversity fractions with a house-price field and a monetary gap, so that the disadvantaged group is more likely to be excluded from high-price cells and to form connected blue clusters interpreted as ghettos (Ortega et al., 2022).

Outside urban studies, the same structural idea appears in formally different objects. In binary CSPs, neighbourhood substitution removes a value when another value can substitute for it across all neighboring constraints, and strengthened forms such as snake substitution and conditioned neighbourhood substitution extend this local elimination principle without increasing the $O(e d^3)$ complexity bound for convergence of the strengthened rules (Cooper, 2020). In objective Bayesian analysis, point exclusion is vacuous in continuous parameter spaces, so inferential loss is defined by excluding a small neighborhood $R_\delta(\theta)$ around each parameter value, with the induced prior determined by local Kullback–Leibler geometry (Villa, 21 Apr 2026).

A concise cross-domain summary is therefore possible.

Domain	Excluded object	Governing local quantity
Urban segregation on tract or lattice networks	Entry, residence, or co-residence in places	Dissatisfaction, exposure, homophily, affordability
Multiscalar regional analysis	Inclusion in connected reference systems	Normalised GJSD and connectivity loss
Spatial perception and boundary drawing	Inclusion of blocks in subjective neighborhoods	Distance-decay, homophily, contiguity
Binary CSP	Domain values	Neighbourhood substitution relations
Continuous Bayesian inference	Local parameter regions	KL loss over excluded neighborhoods

This suggests that “neighbourhood” functions less as a single ontological scale than as a local exclusion operator whose exact meaning depends on the graph, lattice, tract system, or parameter manifold under study.

2. Thresholded exclusion in Schelling-type urban models

A central urban instantiation appears in the Washington, D.C. tract model, where census tracts form the nodes of an undirected contiguity graph $G=(V,E)$, with 179 nodes and 535 unweighted edges (Ortega et al., 2022). The local neighborhood of node $i$ is

$N(i)=\{j\in V:(i,j)\in E\},$

and the fraction of dissimilar occupied neighbors is

$f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$

The dissatisfaction index is

$I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$

with $-H$ for red agents and $+H$ for blue agents. Lower $I_{\mathrm{dis}}$ means higher satisfaction, and an agent is happy iff $R_\delta(\theta)$0. In the D.C. case, $R_\delta(\theta)$1, $R_\delta(\theta)$2, and $R_\delta(\theta)$3 is a latitude-based affordability proxy with northern tracts most expensive (Ortega et al., 2022).

The exclusion mechanism is asymmetric entry. A blue agent may enter a vacant node only if

$R_\delta(\theta)$4

whereas for red agents the condition is

$R_\delta(\theta)$5

In high-$R_\delta(\theta)$6 tracts, even socially acceptable surroundings can still leave blue agents dissatisfied because of the $R_\delta(\theta)$7 affordability penalty. Exit is similarly asymmetric: if a blue agent’s dissatisfaction becomes positive and no weakly improving internal relocation exists, the agent leaves the city. The resulting equilibrium labels persistent blue-occupied nodes in low-$R_\delta(\theta)$8 tracts as ghettos (Ortega et al., 2022).

The update process is open-city and asynchronous. Starting from equal red and blue shares with $R_\delta(\theta)$9 vacancies, one event occurs per step: with probability $G=(V,E)$0, an internal relocation from an occupied node to a vacant node is attempted and accepted if dissatisfaction does not increase; with probability $G=(V,E)$1, a vacant node may admit a random incoming color if $G=(V,E)$2, or an occupied unhappy agent may exit. Convergence is reached when all proposed internal moves and external entries or exits are rejected (Ortega et al., 2022).

The empirical D.C. outcome is strongly directional. Blue agents concentrate in the south and southeast, especially Wards 7 and 8 across the Anacostia, while red agents concentrate in the north and northwest. Comparison against EM-derived binary labels across 1000 stochastic runs yields mean accuracy $G=(V,E)$3 with standard deviation $G=(V,E)$4 (Ortega et al., 2022). The same study reports that a J48 tree classifies $G=(V,E)$5 tracts correctly using a single retained split, $G=(V,E)$6, identifying ghettos in that dataset (Ortega et al., 2022).

A related lattice formulation in stylized cities uses the dissatisfaction

$G=(V,E)$7

with $G=(V,E)$8 for blue and $G=(V,E)$9 for red, on an $i$0 square lattice with $i$1, Moore neighborhoods, open boundary conditions, and $i$2 (Ortega et al., 2022). Five house-price maps are examined: flat, vertical, suburban, core, and grid. Across all five, blue-cluster sizes follow a power law $i$3, with fitted exponents ranging from $i$4 to $i$5, and segregated population ratios depend largely on the monetary gap rather than city type (Ortega et al., 2022). The reported approximation

$i$6

over $i$7 states explicitly that the disadvantaged equilibrium share is near-linearly sensitive to the gap $i$8 (Ortega et al., 2022).

These models formalize neighborhood exclusion as a thresholded admission problem with local composition and local affordability jointly determining whether a disadvantaged agent can remain in, enter, or is ejected from a place.

3. Co-residence, exposure, and built-form regulation

A second family of formulations measures exclusion without explicit agent relocation. In Vienna, district-level nationality counts $i$9 are used to reconstruct a co-residence network from a bipartite nationality–district system (Sadurní et al., 27 Jan 2025). District-level co-residence weights are

$N(i)=\{j\in V:(i,j)\in E\},$0

and a multinomial null model preserving district populations and citywide nationality shares defines

$N(i)=\{j\in V:(i,j)\in E\},$1

and

$N(i)=\{j\in V:(i,j)\in E\},$2

Standardized ties are then

$N(i)=\{j\in V:(i,j)\in E\},$3

with adjacency matrix $N(i)=\{j\in V:(i,j)\in E\},$4 (Sadurní et al., 27 Jan 2025).

The Vienna study covers 21 groups, including Austrians, the 19 most populous migrant nationalities, and an “Others” category. Among 210 significant links, 80 are positive and 130 negative, and Infomap on the positive co-residence network identifies two major communities (Sadurní et al., 27 Jan 2025). The “majority cluster” includes examples such as Austria, Ukraine, Germany, Russia, and Hungary; the “minority cluster” includes Serbia, Turkey, Syria, Romania, and Poland. Wealth disparities, district diversity, and nationality-based homophily align with these cluster boundaries (Sadurní et al., 27 Jan 2025).

The same framework uses standard segregation quantities. Dissimilarity is

$N(i)=\{j\in V:(i,j)\in E\},$5

exposure is

$N(i)=\{j\in V:(i,j)\in E\},$6

isolation is

$N(i)=\{j\in V:(i,j)\in E\},$7

and the scaled homophily index is

$N(i)=\{j\in V:(i,j)\in E\},$8

District diversity is measured using Simpson indices, with citywide $N(i)=\{j\in V:(i,j)\in E\},$9 and population-weighted district average $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$0, which the study interprets as relatively low overall district-level segregation in Vienna (Sadurní et al., 27 Jan 2025).

A behaviorally different but conceptually related formulation appears in VISAGE, where exclusion is encoded in the built environment rather than directly in residence counts (Zhang et al., 20 Jun 2026). Exposure segregation for census tract $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$1 is

$f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$2

with $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$3 the share of visits to tract $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$4 from income quartile $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$5, computed from aggregated, de-identified SafeGraph Weekly Patterns visits over 12 weeks in 2019. VISAGE uses 40 Google Street View frames and 7 satellite tiles for each of 10,030 communities in 31 U.S. cities, organizes 44 image-observable cues into a 10-family codebook, and predicts mobility-derived segregation with community-level Pearson correlation $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$6, reported as $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$7 (Zhang et al., 20 Jun 2026).

The strongest positive cue-level associations with segregation include “High Fences or Walls,” “Barbed Wire or Gated Enclosures,” “Monofunctional Residential Zones/Blocks,” and “Vacant Lots or Abandoned Buildings.” The strongest negative associations include “Commercial Activity Zones,” “Visible Shops or Colorful Signs,” “Large Open Public Spaces,” “Parks, Squares, Playgrounds,” “Street Trees and Green Belts,” and “Public Facilities.” Example response-curve correlations reported are $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$8 for “Vacant Lots or Abandoned Buildings” and $f_d(i)=\frac{N_d(i)}{N_s(i)+N_d(i)}.$9 for “Wide Multi-Lane Roads” with predicted $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$0 (Zhang et al., 20 Jun 2026). Inclusionary housing tracts are reported to have ground-truth $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$1 versus $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$2 in non-IH tracts, with adjusted OLS estimate $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$3 (Zhang et al., 20 Jun 2026).

Together, these studies shift the framework from thresholded micro-mobility to measurable co-residence deficits, exposure deficits, and physical regulators of interaction. A plausible implication is that “exclusion” can be observed either as denied residential access or as suppressed intergroup contact even when co-location exists.

4. Measurement, multiscalarity, and neighborhood delineation

A third line of work turns neighborhood exclusion into a measurement problem. One approach defines exclusion zones statistically through representation and exposure under a null model of random allocation of households across areal units (Louf et al., 2015). For category $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$4 in unit $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$5,

$I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$6

with $I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$7 and

$I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$8

A unit is significantly overrepresented at 99% confidence if

$I_{\mathrm{dis}}(i)=f_d(i)-T+D(i)\pm H,$9

and underrepresented if

$-H$0

Exposure between categories $-H$1 and $-H$2 is

$-H$3

with $-H$4 indicating attraction and $-H$5 indicating repulsion (Louf et al., 2015).

Applied to ACS 2014 block-group data, this framework derives three emergent income classes from the original 16 categories: lower-income categories 0–8, middle-income categories 9–10, and higher-income categories 11–15 (Louf et al., 2015). Neighborhoods for class $-H$6 are then defined by contiguity clustering of significantly overrepresented units. The clustering coefficient

$-H$7

summarizes whether those units form checkerboards or single contiguous regions (Louf et al., 2015). The same study argues that density, rather than distance to a center, is the more meaningful organizing variable in polycentric cities.

Another measurement-oriented formulation defines neighborhoods as subjective, contiguity-constrained inclusions of census blocks by respondents rather than as fixed administrative units (McCartan et al., 2021). The city is a graph $-H$8 on census blocks with rook contiguity, and block inclusion is modeled sequentially. For respondent $-H$9 and block $+H$0,

$+H$1

where $+H$2 enforces contiguity and

$+H$3

Equivalently,

$+H$4

The empirical result is that White respondents are more likely to include blocks with more White residents, and Democrats and Republicans are more likely to include co-partisan areas, with stronger homophily in the New York City “community of interest” task than in the neighborhood task (McCartan et al., 2021). Out-of-sample F1 improvements over matched-radius circles are about $+H$5, and in the NYC COI task median $+H$6 versus tracts is approximately $+H$7 for the baseline and $+H$8 for the full model out-of-sample (McCartan et al., 2021).

A multiscalar formulation goes further by defining neighborhoods through percolation on transport networks rather than through fixed buffers or administrative units (Neira et al., 20 May 2025). Thresholding weighted edges at distance or travel-time $+H$9 yields components $I_{\mathrm{dis}}$0; the giant component ratio is

$I_{\mathrm{dis}}$1

Population compositions are assigned to nodes in the resulting dendrogram, and segregation at each parent node is measured by the normalised Generalised Jensen–Shannon Divergence

$I_{\mathrm{dis}}$2

For each leaf unit $I_{\mathrm{dis}}$3, segregation is $I_{\mathrm{dis}}$4, connectivity is

$I_{\mathrm{dis}}$5

and composite exclusion is

$I_{\mathrm{dis}}$6

In Ecuador, the road network contains 387,797 nodes and 1,027,462 edges, with salient percolation scales around 200 m, 1,000 m, 4,000 m, 6,000 m, and 8,000 m. Normalised GJSD is maximal near $I_{\mathrm{dis}}$7 km, indicating strongest segregation at a regional scale, and high exclusion hotspots include Amazonian border areas and the Guaranda–Guanujo region (Neira et al., 20 May 2025).

These measurement-oriented variants replace the question “who is excluded by local dynamics?” with “how should neighborhoods be defined so that exclusion is measurable at the relevant scale?” The shared answer is that neighborhood boundaries should be derived from statistical overrepresentation, subjective block inclusion, or network connectivity rather than assumed a priori.

5. Dynamical-systems and connected-neighborhood formulations

A continuous-time version of neighborhood exclusion appears in dynamical-systems reformulations of Schelling’s bounded neighborhood model. In a single neighborhood, with densities $I_{\mathrm{dis}}$8 and $I_{\mathrm{dis}}$9, one linear-tolerance formulation is

$R_\delta(\theta)$00

After rescaling,

$R_\delta(\theta)$01

with $R_\delta(\theta)$02 and $R_\delta(\theta)$03 (Haw et al., 2017). Stable mixed equilibria exist only in specific parameter regions. For unlimited movement, if $R_\delta(\theta)$04, stable integration occurs when

$R_\delta(\theta)$05

and if $R_\delta(\theta)$06, when

$R_\delta(\theta)$07

where

$R_\delta(\theta)$08

Outside these regions, only segregated equilibria are stable (Haw et al., 2017). The same analysis treats tipping as basin crossing across the stable manifolds of saddle equilibria.

The two-neighborhood extension shows that stable integration is harder, not easier, when a connected neighborhood is added (Haw et al., 2019). In the symmetric case, the compact system is

$R_\delta(\theta)$09

$R_\delta(\theta)$10

The key thresholds are

$R_\delta(\theta)$11

and

$R_\delta(\theta)$12

The study concludes that stable integration in two connected neighborhoods is possible only when the minority is small and combined tolerance is large, that limiting one population does not necessarily create integration and may destroy it, and that a growing minority can remain integrated only if the majority increases its own tolerance (Haw et al., 2019).

A networked-individuals formulation adds friendship externalities to Schelling relocation (Cerqueti et al., 2020). Utility at candidate location $R_\delta(\theta)$13 is

$R_\delta(\theta)$14

or equivalently

$R_\delta(\theta)$15

The paper reports that moving costs have a monotonic pro-status-quo effect, while friendships attenuate segregation when friend location matters and degrees are not too large; with large degree, the friend term becomes nearly location-invariant and the model approaches Schelling-with-costs (Cerqueti et al., 2020).

These continuous and networked formulations preserve the neighborhood-exclusion logic but translate it into bifurcation structure, basins of attraction, and nonlocal social externalities. A plausible implication is that exclusion should be understood not only as a static map of denied places but also as a stability property of the state space.

6. Extensions beyond urban segregation

The term acquires a formally different meaning in binary CSPs, where neighborhood refers to the set of variables constrained with a given variable. For $R_\delta(\theta)$16, classical neighbourhood substitution removes $R_\delta(\theta)$17 if

$R_\delta(\theta)$18

equivalently $R_\delta(\theta)$19, where $R_\delta(\theta)$20 is the support set of value $R_\delta(\theta)$21 in constraint $R_\delta(\theta)$22 (Cooper, 2020). The paper introduces two strict strengthenings: snake substitution and conditioned neighbourhood substitution. Snake substitution uses

$R_\delta(\theta)$23

to allow coordinated repair of neighbor values, and conditioned neighbourhood substitution allows the replacement value $R_\delta(\theta)$24 to depend on a supporting value $R_\delta(\theta)$25 in a conditioning neighbor. Both subsume classical neighbourhood substitution, both can be applied until convergence in $R_\delta(\theta)$26 time, and the combined rule SCSS also preserves satisfiability (Cooper, 2020). By contrast, finding an optimal elimination sequence for CNS, SS, or SCSS is NP-hard (Cooper, 2020).

In objective Bayesian inference, the neighborhood-exclusion framework addresses the fact that excluding a single point in a continuous parameter space induces no inferential loss (Villa, 21 Apr 2026). With exclusion sets

$R_\delta(\theta)$27

the $R_\delta(\theta)$28-worth of $R_\delta(\theta)$29 is

$R_\delta(\theta)$30

and the loss-based prior is

$R_\delta(\theta)$31

Under regularity and ellipsoidal exclusion

$R_\delta(\theta)$32

the local KL expansion gives

$R_\delta(\theta)$33

and therefore

$R_\delta(\theta)$34

In one dimension the resulting prior coincides with Jeffreys’ prior, $R_\delta(\theta)$35; in higher dimensions, the family is indexed by the geometry $R_\delta(\theta)$36, with Fisher-isotropic exclusion recovering $R_\delta(\theta)$37 (Villa, 21 Apr 2026).

These non-urban uses preserve the same abstract logic: one removes a local neighborhood, quantifies the resulting loss, and uses that loss to simplify or weight the global system. This suggests that the phrase “Neighbourhood-Exclusion Framework” is best read as a transferable formal pattern rather than as the name of a single discipline-specific model.

7. Common themes, policy relevance, and limitations

Across the urban literature, several recurrent themes are explicit. First, exclusion is usually generated locally but diagnosed globally. Local dissatisfaction rules generate tract-level ghetto maps in Washington, D.C. (Ortega et al., 2022); co-residence deficits generate nationality clusters in Vienna (Sadurní et al., 27 Jan 2025); physical barriers and land-use separation predict tract-level exposure segregation across 31 U.S. cities (Zhang et al., 20 Jun 2026); and transport-derived hierarchy plus information divergence identifies exclusion hotspots at city, regional, and national scales in Ecuador (Neira et al., 20 May 2025).

Second, affordability, connectivity, and homophily are treated as complementary rather than interchangeable mechanisms. In D.C., a coarse affordability proxy and a group-level economic gap suffice to generate tract-level ghetto predictions with $R_\delta(\theta)$38 agreement against EM-derived labels (Ortega et al., 2022). In Vienna, district wealth and diversity align with nationality homophily but do not exhaust it (Sadurní et al., 27 Jan 2025). In VISAGE, daily social mixing is conditioned by the visual grammar of the built environment rather than only by residential composition (Zhang et al., 20 Jun 2026). In Ecuador, exclusion can arise from low connectivity even where local composition is not extreme, because compositional divergence and network fragmentation are jointly necessary for high $R_\delta(\theta)$39 (Neira et al., 20 May 2025).

Third, scale matters. The multiscalar framework explicitly warns that single-scale metrics produce scale bias and can hide places that look locally integrated but regionally excluded (Neira et al., 20 May 2025). The block-level subjective-neighborhood model similarly shows that tract or ZIP proxies can miss the boundary logic actually used by residents (McCartan et al., 2021). The representation-and-exposure approach argues that density, not radial distance to a center, is the relevant organizing variable in polycentric cities (Louf et al., 2015).

The policy implications reported in the papers are correspondingly scale-sensitive. Adjusting $R_\delta(\theta)$40 or reducing $R_\delta(\theta)$41 in the D.C. Schelling model changes the blue entry condition and can open excluded neighborhoods (Ortega et al., 2022). The Vienna study identifies mixed-income housing, inclusionary zoning, spatial allocation of public housing, anti-discrimination enforcement, migrant support services, and boundary redesign as levers for raising exposure and reducing homophily (Sadurní et al., 27 Jan 2025). VISAGE frames planners’ interventions in terms of softening defensive boundaries, reconnecting fragmented street networks, legalizing mixed-use frontages, converting vacant lots, and co-locating transit, schools, and amenities (Zhang et al., 20 Jun 2026). The multiscalar Ecuador analysis distinguishes city-scale levers such as inclusionary zoning and feeder links from regional-scale levers such as inter-municipal corridors and strategic transport investments (Neira et al., 20 May 2025).

The main limitations are equally recurrent. Several studies are cross-sectional or quasi-static rather than fully causal (Sadurní et al., 27 Jan 2025, Zhang et al., 20 Jun 2026). Validation often relies on proxy or internally derived labels rather than independent ground truth, as in the D.C. comparison to EM-derived ghetto labels (Ortega et al., 2022). Fine-scale heterogeneity is abstracted away in one-agent-per-tract models (Ortega et al., 2022), stylized price fields (Ortega et al., 2022), and administrative district co-residence networks (Sadurní et al., 27 Jan 2025). Measurement remains sensitive to unit definitions, privacy thresholds, and data availability (Louf et al., 2015, Sadurní et al., 27 Jan 2025, Neira et al., 20 May 2025).

Taken together, the literature does not present a single canonical Neighbourhood-Exclusion Framework. Instead, it presents a coherent family of local-exclusion formalisms: thresholded dissatisfaction on graphs and lattices, co-residence deficits under null models, exposure deficits induced by built form, multiscalar divergence plus weak connectivity, contiguity-constrained subjective inclusion, neighborhood-based domain reduction in CSPs, and KL-neighborhood exclusion in continuous parameter spaces. The unifying principle is exact: remove access to a local neighborhood, quantify the resulting loss or asymmetry, and study the macrostructure that emerges.