Variogram¶

A geostatistical lag function measuring expected squared differences between field values, used to model spatial dependence, anisotropy, nugget, range, and kriging weights.

Version: v1 · 2026-08-30 · History
Domain-specific #: 3062
Origin domain: statistics

Core Idea¶

A variogram describes how dissimilarity between values of a spatial or temporal random field changes with separation. Under intrinsic stationarity, the semivariogram is conventionally

\[ \gamma(h)=\frac12 E\left[(Z(s+h)-Z(s))^2\right], \]

where the expectation depends on lag vector \(h\) but not absolute location \(s\). Some literature calls \(2\gamma(h)\) the variogram and \(\gamma(h)\) the semivariogram; others use “variogram” for the half form. A reference-grade use must declare the convention.^[1]

The abstraction turns sampled pairwise differences into a lag-dependent dependence model used in geostatistics, especially kriging. Its role structure includes support, lag, stationarity assumptions, directional dependence, estimator bins, an admissible fitted model, and features such as nugget, sill, and range.^[2] It is not merely a plot of variance against distance.

Structural Signature¶

Recognition roles:

the regionalized/random field \(Z(s)\) — values indexed by spatial or temporal location;
the sampling support — point, block, interval, or other measurement footprint;
the lag vector \(h\) — separation magnitude and potentially direction;
the increment \(Z(s+h)-Z(s)\) — change over the lag;
the squared-increment expectation — the theoretical dependence function;
the intrinsic-stationarity assumption — increment mean/variance depend on lag, not absolute position;
the empirical estimator — average squared differences for observed pairs in a lag bin;
the admissible model — a conditionally negative-definite function suitable for kriging;
the structural features — nugget, partial sill, range, anisotropy, and nested scales.^[3]

Recognition test. State the convention, field, support, lag/direction, pair selection, and stationarity model. Verify the fitted function is variogram-admissible. A raw scatter of squared differences without lag aggregation is a variogram cloud, not yet a fitted variogram.

What It Is Not¶

A variogram is not ordinary sample variance: it concerns increments between locations and is a function of lag. It is not a covariance function, though under second-order stationarity \(\gamma(h)=C(0)-C(h)\). It is not a correlogram, which typically normalizes covariance. It is not the same as spatial autocorrelation statistics such as Moran's I.

It is not automatically isotropic. Replacing vector \(h\) with distance \(\|h\|\) assumes rotational invariance. It is not automatically bounded by a sill; power-law variograms may be unbounded. A nugget discontinuity at the origin can reflect measurement error or microscale variation and should not be interpreted as a physical nugget without evidence.

Scope of Application¶

Variograms are central in mining geostatistics, soil and environmental mapping, hydrogeology, remote sensing, ecology, epidemiology, image analysis, and spatiotemporal prediction.^[3] They guide kriging weights and prediction variance, compare continuity by direction, and support simulation of spatial random fields.

They are useful when local differences are more stable than a global mean, which is why intrinsic stationarity can be weaker than full second-order stationarity. Directional variograms diagnose anisotropy; cross-variograms extend to multiple variables. Block support and change-of-support problems require integrating or regularizing the point-support model.

The method is not appropriate without adequate spatial coverage and pair counts, or when strong nonstationary trend has been mistaken for dependence. Detrending or a nonstationary model may be required before interpretation.

Clarity¶

The variogram clarifies that “near things are similar” is a testable scale-dependent claim. Small \(\gamma(h)\) means increments at lag \(h\) are usually small; growth with lag indicates declining similarity. A plateau suggests that beyond a range, additional separation no longer increases expected squared difference under the model.

It also clarifies that direction matters. Two locations 100 meters apart east-west and north-south share distance but may cross different geological structures. A scalar distance plot can average away this anisotropy. Finally, a reported range depends on model form and convention: spherical models reach a finite sill, exponential models approach it asymptotically and use a practical range.

Manages Complexity¶

A dataset with \(n\) locations contains \(n(n-1)/2\) pairs. Variogram estimation compresses this pair cloud into lag/direction bins, then into a valid parametric or nonparametric model. Nugget, sill, range, and anisotropy summarize patterns used by kriging.

The compression discards local pair identity, sampling imbalance, and uncertainty unless these are reported separately. Wide bins stabilize estimates but blur scale; narrow bins retain scale but may have few pairs. Fitted-model simplicity trades against nested or directional structure. Responsible use retains bin definitions, pair counts, support, and fitting criterion.

Abstract Reasoning¶

For observed values, a classical method-of-moments estimator at lag bin \(h\) is

\[ \widehat\gamma(h)=\frac{1}{2N(h)} \sum_{(i,j)\in P(h)}[Z(s_i)-Z(s_j)]^2, \]

where \(P(h)\) is the selected pair set and \(N(h)=|P(h)|\).^[2] The factor one-half matches the semivariogram convention. Outliers have large influence because differences are squared, motivating robust estimators in some analyses.

Under constant mean and covariance \(C(h)\), expanding the square gives \(\gamma(h)=C(0)-C(h)\). Thus variogram growth corresponds to covariance decline, but the variogram can exist under intrinsic assumptions where covariance is not conveniently defined. A valid variogram satisfies conditional negative definiteness so kriging systems produce coherent variances.^[3]

Knowledge Transfer¶

Literal transfer occurs across spatial fields—ore grade, rainfall, soil concentration, elevation residuals—and temporal increments when the same lag/increment structure applies. Practitioners transfer diagnostics about support, anisotropy, pair counts, and admissible models.

The parent prime:comparison carries the portable operation of placing paired values in a shared frame and reading off a relation. The variogram fixes that frame to lag and reads squared dissimilarity rather than generic likeness or normalized co-movement. prime:correlation remains a close statistical neighbor, but the variogram can be defined under intrinsic stationarity where covariance and correlation need not exist. Using “variogram” for any curve of variation is metaphor unless lagged increments and admissibility remain.

Examples¶

Linear field boundary. For deterministic \(Z(s)=as\) on a line, increments equal \(ah\), so the squared difference is \(a^2h^2\). This apparent variogram reflects trend rather than stationary residual dependence, warning that detrending matters.

Nugget plus spherical model. An empirical curve jumps near zero, grows, and reaches a plateau. A fitted nugget accounts for unresolved/measurement variation, the partial sill for spatial structure, and the finite range for the modeled correlation distance. Each interpretation requires domain evidence.

Directional anisotropy. Samples along a sedimentary layer change slowly, while samples across layers change rapidly. Directional variograms show a longer range along strike than across it. Averaging directions would obscure the geometry.

Covariance relation. If a stationary field has variance \(C(0)=10\) and covariance \(C(h)=6\), the semivariogram is \(4\) in squared units. The example demonstrates dissimilarity/dependence duality.

Sparse-bin failure. A distant lag bin contains only three pairs clustered in one corner. Its empirical value may be numerically computable but weakly representative. Pair count and spatial layout are part of evidence quality.

Structural Tensions¶

Stationarity vs. real heterogeneity. Pooling by lag gains power but can erase trend or local regimes. Diagnostic: inspect residual maps and directional/location subsets before assuming one variogram.
Resolution vs. stability. Narrow bins preserve scale; broad bins increase pair counts. Diagnostic: report sensitivity to bin width and minimum pair count.
Nugget interpretation vs. ambiguity. A discontinuity may be measurement error or microscale structure. Diagnostic: use replicate/precision evidence before assigning cause.
Fit quality vs. admissibility. An arbitrary smooth curve may match points yet be invalid for kriging. Diagnostic: use a known admissible model or prove conditional negative definiteness.
Autonomy vs. reduction. Variogram constitutively uses Comparison and empirical Measurement, but its lagged squared increments and kriging role are stable. Diagnostic: remove lag or increments; if the object still qualifies, it has collapsed into generic variability.

Structural–Framed Character¶

The function is mathematically structural, but sampling support, stationarity, directional bins, and fitted-model conventions frame every empirical use. The units are squared units of \(Z\), which makes a bare ordinate meaningless without the measured attribute.

Geostatistical vocabulary reflects a history in mining and earth science, yet the mechanism is not limited to ore. “Nugget,” “sill,” and “range” are model features, not universal physical entities.

Structural Core vs. Domain Accent¶

The portable core is dependence as a function of separation, expressed by expected squared differences. The domain accent is regionalized variables, intrinsic stationarity, spatial support, anisotropy, kriging admissibility, and empirical lag bins.

The candidate remains domain-specific. Comparison travels across domains; the variogram's exact dissimilarity function and geostatistical consequences do not become a new prime merely because temporal uses also exist.

Variogram presupposes prime:comparison: field values separated by a selected lag are co-framed, their squared increment is aligned under a declared support and convention, and a lag-conditioned dissimilarity relation is read off. It also relates to prime:correlation, because it describes statistical dependence without causal claim, to prime:measurement, because empirical estimates depend on support and procedure, and to prime:scale, because structure changes across lags. Comparison is the minimal proposed constitutive parent; the variogram is an output function built through that operation, not a subtype of generic Correlation.

Relationships to Other Abstractions¶

Current abstraction Variogram Domain-specific

Parents (1) — more general patterns this builds on

Variogram presupposes Comparison Prime

Variogram presupposes prime:comparison: field values separated by a selected lag are co-framed, their squared increment is aligned under a declared support and convention, and a lag-conditioned dissimilarity relation is read off.

Hierarchy path (1) — routes to 1 parentless root

Variogram → Comparison → Self Checking

Neighborhood in Abstraction Space¶

Variogram sits in a sparse region of the domain-specific corpus (74^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Sample variance: one-number dispersion about a mean.
Covariance function: expected product of centered values; related by \(C(0)-C(h)\) under stronger stationarity.
Correlogram: usually normalized autocorrelation versus lag.
Variogram cloud: individual pair squared differences before binning/modeling.
Kriging: prediction algorithm that consumes a variogram/covariance model.
Nugget, sill, range: features of particular fitted models, not the entire variogram.

References¶

[1] Georges Matheron, Matheron's Theory of Regionalised Variables, edited by Vera Pawlowsky-Glahn and Jean Serra, Oxford University Press, 2019, based on 1970 lectures, doi:10.1093/oso/9780198835660.001.0001. registry ↩

[2] Noel Cressie, Statistics for Spatial Data, rev. ed., Wiley, 1993, ISBN 9780471002550. registry ↩a ↩b

[3] Jean-Paul Chilès and Pierre Delfiner, Geostatistics: Modeling Spatial Uncertainty, 2^nd ed., Wiley, 2012, ISBN 9780470183151. registry ↩a ↩b ↩c