Scale-Normalised Density Overlap

• Scale-normalised density overlap is a framework that normalizes raw density or overlap measures against a reference scale, separating intrinsic structure from effects like resolution and noise.
• Its applications span diverse fields such as lattice gauge theory, density functional theory, network science, and MRI intensity normalization, each employing tailored normalization techniques.
• Normalization methods include smoothing time calibration, hypergeometric null models, and doubly stochastic scaling, ensuring meaningful comparisons across varying scales and conditions.

Scale-normalised density overlap denotes, across several technical literatures, a family of constructions in which a density, overlap, or local structural signal is not interpreted in raw form but only after normalization against a reference scale. That reference may be a smoothing time, a uniform-gas kinetic-energy density, a size-conditioned null model, a unit-mean local density profile, a doubly stochastic scaling, or an affine standardisation of intensity coordinates. Taken together, these usages suggest a common aim: to separate intrinsic structural content from nuisance variation induced by resolution, sampling density, embedding scale, or global location-scale mismatch (Cheng et al., 2020, Sun et al., 2012, Magner et al., 2016, Kazanskii, 3 May 2026, Landa et al., 2022, Castro et al., 2018).

1. General structure of the concept

Taken together, the relevant works suggest that “scale-normalised density overlap” is best treated as an umbrella description rather than a single canonical formalism. In each case, the object of interest is a density-like or overlap-like quantity, but its interpretation depends on a normalization that fixes the comparison scale. In lattice gauge theory, the normalization is operational and is supplied by matching the overlap topological charge density to a flowed gluonic density at a “proper flow time” $\tau_{\rm pr}$. In semilocal density functional theory, the normalization is intrinsic and dimensionless, since the overlap descriptor $\alpha$ divides the excess kinetic-energy density by the local uniform-gas kinetic-energy density. In graph overlap statistics, the normalization is probabilistic, since overlap significance is judged relative to the hypergeometric scale set by $|A|$, $|B|$, and $n$. In density-preserving embedding and manifold methods, normalization removes arbitrary global rescaling or endpoint-specific noise inflation. In MRI intensity normalisation, global translation and rescaling are removed before nonlinear density matching is performed (Cheng et al., 2020, Sun et al., 2012, Magner et al., 2016, Kazanskii, 3 May 2026, Landa et al., 2022, Castro et al., 2018).

This suggests a recurring tripartite structure. First, there is a local object: $q_{\rm ov}(x)$, $\alpha$, $\delta(Z)$, $\rho_i$, $W_{ij}$, or a histogram density $\alpha$0. Second, there is a reference scale: Wilson-flow time, $\alpha$1, a community-size-conditioned null, the mean of local densities, Sinkhorn scaling factors, or affine moment matching. Third, there is a comparison rule: maximizing $\alpha$2, minimizing an $\alpha$3 mismatch, evaluating a joint tail probability, or enforcing row-stochastic or doubly stochastic constraints. The significance of the phrase therefore lies less in a single formula than in a methodological principle: overlap or density is made comparable only after explicit control of scale.

2. Overlap topological charge density in lattice gauge theory

In lattice gauge theory, the most direct realization of scale-normalised density overlap is the calibration of the overlap-definition topological charge density against an explicit smoothing scale. The overlap formalism uses a Wilson kernel with hopping parameter $\alpha$4 related to the Wilson mass parameter $\alpha$5 by

$\alpha$6

The local fermionic topological charge density is defined from the overlap operator by

$\alpha$7

with total charge $\alpha$8. Because point-source evaluation is expensive on large lattices, the density is estimated with the symmetric multi-probing (SMP) method; the approximation neglects off-diagonal spacetime terms and is justified by locality. The paper verifies this estimator against a point-source calculation and reports an almost perfect agreement, with matching parameter $\alpha$9, so the observed scale dependence is attributed to the overlap kernel parameter rather than to SMP artifacts (Cheng et al., 2020).

The central empirical result is that increasing the Wilson mass removes finer topological structures. In the paper’s $|A|$0-based convention, smaller $|A|$1 means larger $|A|$2, and “the non-trivial topological objects are removed as the Wilson mass is increased.” The comparison quantity for matching overlap and Wilson-flowed gluonic densities is

$|A|$3

with $|A|$4 and $|A|$5. The best match is the flow time where $|A|$6 is closest to $|A|$7; this defines the proper flow time $|A|$8. On the $|A|$9, $|B|$0 ensemble, the map is monotone: $|B|$1 The physical smoothing scale is then expressed as the proper flow radius $|B|$2. Using averaged $|B|$3, the paper quotes $|B|$4 for $|B|$5 at $|B|$6, $|B|$7 at $|B|$8, and $|B|$9 at $n$0, corresponding to $n$1. As the lattice spacing $n$2 decreases, the proper flow radius also decreases, “as expected” (Cheng et al., 2020).

An earlier calibration study reached the same conclusion through stout-link smearing rather than Wilson flow. There, larger $n$3 reveals more visible topological structure, while smaller $n$4 yields a smoother, more filtered density. The overlap density was matched to a gluonic density after over-improved stout-link smearing with $n$5 and $n$6, using both multiplicative amplitude matching and the normalization-independent correlator $n$7. The resulting calibration was approximately

$n$8

The same work also showed that applying the overlap operator to a pre-smeared field performs further filtering: for $n$9, a 25-sweep pre-smeared configuration yielded an overlap density resembling roughly 45 smearing sweeps (Moran et al., 2010).

The principal caveat is that the calibration is empirical and operator dependent. The total overlap charge can change with $q_{\rm ov}(x)$0, especially on coarser lattices, and the best-match $q_{\rm ov}(x)$1 is defined by local density correlation rather than by equality of integrated charge. The $q_{\rm ov}(x)$2 relation also depends on the kernel convention, the choice of improved field-strength tensor, and the flow discretisation. A plausible implication is that “scale-normalised overlap density” in this setting is not a universal renormalization scheme, but an operational map from the overlap kernel mass to an effective smoothing scale (Cheng et al., 2020, Moran et al., 2010).

3. Dimensionless orbital-overlap descriptors in semilocal density functional theory

In semilocal density functional theory, the central scale-normalised overlap variable is the meta-GGA inhomogeneity parameter

$q_{\rm ov}(x)$3

where

$q_{\rm ov}(x)$4

$q_{\rm ov}(x)$5

The paper explicitly describes $q_{\rm ov}(x)$6 as a “dimensionless inhomogeneity parameter” and interprets it as characterizing the “extent of orbital overlap.” Because $q_{\rm ov}(x)$7, $q_{\rm ov}(x)$8, and $q_{\rm ov}(x)$9 have the same dimensions, $\alpha$0 is dimensionless; because the denominator is the kinetic-energy density of the uniform electron gas, $\alpha$1 is normalized by the local homogeneous reference scale (Sun et al., 2012).

The physical regime structure is explicit. The single-orbital regime satisfies $\alpha$2, hence $\alpha$3. The slowly varying density regime is characterized by $\alpha$4. Large $\alpha$5 corresponds to strong overlap and pronounced inhomogeneity, particularly intershell regions. In the paper’s analysis of 10-electron and 12-electron hydrogenic densities, shell regions have $\alpha$6 and intershell regions have $\alpha$7. The paper further notes that, in intershell regions with large $\alpha$8, even when $\alpha$9 is small, the exchange hole is likely not tightly centered near the electron but spread over neighboring shells, making the exchange energy density less negative than in a slowly varying or uniform region (Sun et al., 2012).

The scale-normalised role of $\delta(Z)$0 becomes operational in the exchange enhancement factor

$\delta(Z)$1

with

$\delta(Z)$2

$\delta(Z)$3

This separates the gradient-driven variable $\delta(Z)$4 or $\delta(Z)$5 from the orbital-overlap variable $\delta(Z)$6. The paper argues that GGAs depending only on $\delta(Z)$7 cannot distinguish one-orbital regions from overlapping-orbital regions at fixed density and gradient, whereas $\delta(Z)$8 can distinguish one-orbital, slowly varying, and intershell or strong-overlap environments (Sun et al., 2012).

For the phrase “scale-normalised density overlap,” the paper supports a specific but qualified interpretation. It is reasonable, based on this work, to describe $\delta(Z)$9 as a scale-normalised orbital-overlap measure or a scale-normalised inhomogeneity descriptor tied to orbital overlap. The caveat, stated explicitly, is that $\rho_i$0 is not a direct overlap integral and not a direct measure of density overlap in the usual integral sense; it is a kinetic-energy-density-based semilocal indicator. Its significance lies in the fact that its monotonic dependence in the exchange enhancement factor has major consequences for compact versus less compact systems, affecting lattice constants, atomization energies, and surface energetics (Sun et al., 2012).

4. Joint significance of overlap size and overlap density in networks

In network science, “scale-normalised density overlap” is realized most explicitly by CoDO, “Combining Density and Overlap,” a p-value for the observed overlap of two subgraphs under an Erdős–Rényi null model. For communities $\rho_i$1 and $\rho_i$2, with overlap $\rho_i$3, internal density is

$\rho_i$4

and the CoDO score is

$\rho_i$5

Conditioning on overlap size $\rho_i$6, the paper writes

$\rho_i$7

The first factor is hypergeometric in overlap size; the second is a conditional hypergeometric upper tail for the number of overlap edges required to reach density $\rho_i$8 (Magner et al., 2016).

The normalization is statistical rather than geometric. Overlap size is normalized relative to the community sizes $\rho_i$9, $W_{ij}$0, and ambient graph size $W_{ij}$1, with

$W_{ij}$2

Density is normalized relative to what is achievable for an overlap of size $W_{ij}$3. The score is a joint tail probability, not a product of separate tests. This avoids favoring overlaps that are extreme only in size or only in density. The paper contrasts CoDO with the hypergeometric tail (HGT), which uses only overlap size, and with density-only ERD, which ignores the fact that the dense subgraph arises as the intersection of two given communities (Magner et al., 2016).

Theoretical properties reinforce the scale-normalised interpretation. Fixing overlap density $W_{ij}$4, CoDO decreases monotonically as normalized overlap size $W_{ij}$5 increases, where $W_{ij}$6. Fixing $W_{ij}$7, CoDO decreases monotonically as $W_{ij}$8 increases. The paper defines the threshold

$W_{ij}$9

and proves that the transition to significance occurs not far above the expected normalized overlap size, under the condition $\alpha$00. This is a scale-calibrated significance statement: the meaning of a given overlap depends on where it sits relative to the random scale set by the constituent communities (Magner et al., 2016).

The empirical results are likewise organized around combined scale and density. On synthetic graphs, CoDO detects cases where either overlap size or overlap density is significant, and shows a smoother transition than ERD. On social circles from Facebook, Google+, and Twitter, it outperforms HGT and ERD in ROC/AUC evaluation. On KEGG pathways in a human protein interaction network, the reported Spearman correlations between pathway relatedness and p-values are $\alpha$01 for HGT, $\alpha$02 for ERD, and $\alpha$03 for CoDO. The paper’s main methodological contribution is therefore a formal, null-model-based notion of density overlap significance that is normalized simultaneously by combinatorial scale and internal coherence (Magner et al., 2016).

5. Relative-density preservation in stochastic neighbor embedding

In dimensionality reduction, the most direct analogue of scale-normalised density overlap is DR-SNE, a density-regularized variant of stochastic neighbor embedding that aligns relative density structure through normalized log-density estimates. The baseline structure term is a sparse KL objective over $\alpha$04-nearest-neighbor affinities, while the density term is

$\alpha$05

where local density is estimated by

$\alpha$06

and normalized to unit mean by

$\alpha$07

The full training loss uses a warm-up phase and then optimizes $\alpha$08 (Kazanskii, 3 May 2026).

The scale normalization is explicit. If the embedding is globally rescaled by $\alpha$09, then the embedding-space density surrogate scales by $\alpha$10, but the mean density scales by the same factor, so $\alpha$11 is unchanged. In log space, the global multiplicative constant becomes an additive constant and is removed by normalization. The paper therefore describes the density term as a “simple and scale-invariant mechanism for preserving relative density variations.” Training uses regression-style squared matching of normalized log densities, while evaluation uses density correlation,

$\alpha$12

This is not an overlap integral and not a divergence between full density fields; it is a pointwise alignment of per-sample relative-density statistics (Kazanskii, 3 May 2026).

The theoretical interpretation is given through local volume distortion. For an embedding map $\alpha$13 on a smooth manifold, the appendix derives

$\alpha$14

and, after normalization,

$\alpha$15

Hence minimizing $\alpha$16 encourages local volume changes to be uniform across the manifold, modulo one global scale factor. This suggests that the “overlap” being preserved is the relative density profile over samples rather than absolute metric scale (Kazanskii, 3 May 2026).

The paper positions this against DensMAP and DenSNE, which it characterizes as preserving density through local scale consistency. DR-SNE instead directly aligns normalized density estimates. The reported trade-off is explicit: increasing $\alpha$17 improves density preservation but degrades trustworthiness and continuity, and the method can produce less visually separated clusters and more class overlap. A plausible implication is that, in this context, scale-normalised density overlap is not a substitute for neighborhood preservation but an additional objective controlling relative density distortion (Kazanskii, 3 May 2026).

6. Doubly stochastic scaling as density- and noise-normalised overlap on manifolds

A closely related but theoretically distinct construction appears in robust manifold learning under high-dimensional noise. There, the raw Gaussian kernel

$\alpha$18

is transformed by symmetric Sinkhorn-type scaling: $\alpha$19 The corresponding population scaling function $\alpha$20 is defined so that the continuum kernel is doubly stochastic with respect to the sampling measure. The leading-order asymptotic form is

$\alpha$21

because $\alpha$22 for small $\alpha$23. This is an explicit density-corrected overlap formula: the clean Gaussian neighborhood overlap is divided by the square root of the local sampling masses at its endpoints (Landa et al., 2022).

The scaling factors also absorb pointwise noise magnitude. The theorem gives

$\alpha$24

up to relative error terms. Thus higher density and larger noise magnitude both reduce $\alpha$25, and when the factors enter symmetrically as $\alpha$26, the endpoint-specific inflation due to heterogeneous noise is cancelled at leading order. The paper’s central contribution is therefore a scale-normalised density overlap that is corrected simultaneously for sampling density and heteroskedastic high-dimensional noise (Landa et al., 2022).

Because $\alpha$27 is row-stochastic, density is not estimated from row sums. Instead, the paper defines the Doubly Stochastic Kernel Density Estimator

$\alpha$28

with the entropy-limit form

$\alpha$29

The theorem states

$\alpha$30

From the same normalized overlap, the paper derives estimators of local noise magnitude,

$\alpha$31

clean signal magnitude, and clean squared distance,

$\alpha$32

It also constructs robust Laplacian normalizations converging to manifold differential operators, including the Laplace–Beltrami operator at $\alpha$33 (Landa et al., 2022).

The main significance is that overlap is no longer interpreted as raw kernel mass. Instead, it is the overlap that remains after local sample-volume effects, global kernel mass, and pointwise noise-scale inflation have been normalized away. This provides a particularly literal realization of scale-normalised density overlap: an affinity matrix whose leading-order entries reflect clean manifold proximity rather than nuisance density or noise heterogeneity (Landa et al., 2022).

7. Density matching and mass-conserving transport in MRI intensity normalisation

In MRI intensity normalisation, the relevant analogue is not an affinity or p-value but an overlap-sensitive density registration framework. Source and target histograms are modeled as Dirichlet process Gaussian mixture models,

$\alpha$34

with Gaussian components

$\alpha$35

The comparison criterion is the $\alpha$36 divergence

$\alpha$37

Here $\alpha$38 is the explicit density-overlap term. Minimizing $\alpha$39 increases overlap while reducing discrepancy everywhere, and the objective is available in closed form for Gaussian mixtures (Castro et al., 2018).

The scale-normalisation step is explicit and precedes nonlinear matching. The method begins with a coarse affine alignment based on first and second moments, “accounting for arbitrary translation and rescaling of the values,” and in experiments the images are aligned to zero mean and unit variance before constructing reference densities. After this standardisation, source component means and precisions are optimized while source weights are kept fixed. The gradients of $\alpha$40 with respect to $\alpha$41 and $\alpha$42 are available in closed form, and the transformed parameters define a smooth time-dependent density path (Castro et al., 2018).

Voxel intensities are then transported through a mass-conserving flow. The evolving density $\alpha$43 and velocity field $\alpha$44 satisfy the continuity equation

$\alpha$45

For a single Gaussian component,

$\alpha$46

which decomposes into a translation term and a local scale-like dilation or compression term induced by changing precision. The mixture velocity is the posterior-weighted combination

$\alpha$47

and the final intensity transformation is generated by

$\alpha$48

The paper states that the true solution is diffeomorphic by construction and integrates the flow numerically with fourth-order Runge–Kutta (Castro et al., 2018).

This framework shows how density overlap can be scale-normalised in a practical harmonisation pipeline: first remove global location-scale mismatch, then optimize an explicit overlap-sensitive density objective, then transport data so that the empirical histogram remains consistent with the matched density. The main limitation is also stated explicitly: “A fundamental limitation of any histogram matching scheme is that it is unclear how to proceed when the distributions are genuinely different.” This caveat is broadly relevant. Across all domains surveyed here, scale normalization does not eliminate operator dependence, null-model dependence, or model mismatch; it renders comparisons interpretable only within a specified calibration framework (Castro et al., 2018).