Can a compact kernel be smooth?

Yes. Suitable constructions provide different levels of smoothness while remaining exactly zero beyond a boundary.

Must it integrate to one?

Not as a covariance kernel. Unit integral is a requirement for certain density-estimation kernels, a different use of the word kernel.

Is the support radius the same as observation noise?

No. The radius governs pairwise covariance range; observation noise represents variability in measurements or observations.

AI Glossary / Probabilistic machine learning / Intermediate

Compactly supported kernel

Definition

A compactly supported kernel is a kernel whose nonzero values are confined to a bounded neighborhood. For the covariance functions discussed here, it assigns exactly zero covariance once the distance between two inputs reaches or exceeds a specified boundary. The term is synonymous with compact-support kernel.

What it means

The support of a function is the closure of the region where it is nonzero. In ordinary Euclidean space, compact support means that this region is closed and bounded. A distance-based covariance kernel uses this idea to limit the range of its covariance.
The support radius answers how far the kernel extends. Smoothness answers a different question: how gently the function varies, including near its boundary. A kernel can have finite support and still be smooth. Neither property should be confused with the number of observations available for training.

How it works

Select a valid kernel family, accounting for the dimension of the input space. Wendland functions are an example of carefully constructed, compactly supported radial covariance functions. Their parameters control aspects of smoothness and dimensional validity.
For a radius R, form the scaled distance r = distance / R. Evaluate the selected function for r below 1 and use zero at or beyond the boundary when the function is defined that way. A variance multiplier controls the covariance magnitude separately.
Build the covariance matrix, include any specified observation-noise model, and fit the statistical model. Evaluate the support radius as a modeling parameter. A larger radius typically connects more pairs, but does not guarantee better prediction or more appropriate uncertainty.

Three practical examples

Spatial interpolation: predict a measured quantity from nearby locations using a valid compact covariance, checking errors at held-out locations rather than only at training points.
Covariance tapering: multiply an existing covariance by a valid compact taper to create an approximation with finite-range pairwise covariance. This changes the modeled covariance and requires assessment.
Local kernel learning: use a compact similarity structure in an appropriate kernel method, taking care that feature scaling does not make physically unrelated inputs appear close.

Think of it this way

Think of a local delivery service with a clearly defined service radius. Moving the boundary changes which addresses are connected, while the quality of service within the boundary is a separate issue. The analogy does not prove a statistical kernel is valid: covariance functions must meet mathematical consistency requirements across every collection of inputs, not merely follow a plausible distance rule.

Worked example

Consider an original example using the normalized Wendland function phi(r) = (1 − r)^4 × (4r + 1) for 0 ≤ r < 1, and zero for r ≥ 1. This particular function is valid as a radial covariance in Euclidean dimensions up to three.
Take unit marginal variance and support radius R = 2. Two points distance 1 apart have r = 0.5, so their covariance is (0.5)^4 × 3 = 0.1875. Points distance 2 apart have r = 1 and covariance zero. At distance zero, the covariance is 1.
Now increase R to 4 while keeping the physical distance 1. The scaled distance becomes 0.25, giving covariance (0.75)^4 × 2 = 0.6328125. Thus a radius change affects covariance values inside the support as well as which pairs have zero entries.
These values describe the chosen prior covariance. They are not probabilities that the observations are correct, and they do not by themselves specify a posterior prediction.

Illustrative case study

Fictional case: Larch Mapping estimates local variation across a two-dimensional site. Its analysts compare two support radii using the worked-example kernel and consistent coordinate units. The shorter radius leaves many observation pairs unconnected in the covariance matrix.
During review, they discover that one coordinate was recorded in meters and the other in centimeters. They correct the units before comparing predictive performance. They then report the selected radius alongside assumptions about the mean trend and measurement noise. The case illustrates why a compact kernel's practical meaning depends on the input geometry, not just its formula.

Common misconceptions and mistakes

A compactly supported function is not automatically a positive semidefinite kernel. Multiplying a covariance by an arbitrary cutoff can make an invalid covariance matrix; use a construction with the needed validity guarantee.
Do not confuse marginal zero covariance with conditional independence. Even in a Gaussian model, those are different statements. Furthermore, local covariance does not imply every aspect of the fitted model is local: a shared mean function or learned parameters may connect distant predictions.
Do not present this term and compact-support kernel as distinct algorithms. Their separate glossary records reflect alternative wording.

Business use

The business decision is whether finite-range covariance is a useful approximation or an appropriate assumption for the problem. It can support scalable modeling when sparse numerical methods exploit many exact zeros.
Require evidence about prediction, uncertainty, and computational cost across plausible radii. A kernel matrix may become nearly dense when most points lie within the support, eliminating expected savings. Conversely, overly local modeling may leave regions weakly informed. Record coordinate units, dimensional restrictions, and the interpretation of the radius so another team can reproduce the analysis.

Check your understanding

In the worked example with R = 2, what happens to covariance between points distance 3 apart if the marginal variance is multiplied by 5?

Answer: It remains zero, because distance 3 is outside the support. Multiplying the kernel's amplitude does not extend its support radius.

Related terms

Compact-support kernel
Radial basis function
Gaussian Process Regression
Wendland covariance
support
covariance tapering
length scale
sparse covariance matrix

Learn next

Compare support radius, variance amplitude, and smoothness by changing one at a time in a small covariance matrix. Then examine the distinction between a sparse covariance matrix and a sparse precision matrix. Study covariance tapering after understanding why products of valid covariance kernels remain valid.

Covariance matrix

Sources

Rasmussen and Williams, Gaussian Processes for Machine Learning, chapter 4, equation 4.21 and compact-support discussion: https://gaussianprocess.org/gpml/chapters/RW4.pdf fields authors, Wendland family documentation: https://rpkgs.github.io/fields/reference/Wendland.html Independent source review: GPML equation 4.21 and fields documentation were read for support, smoothness and dimensional restrictions.

Previous: Compact-support kernel Next: Companding