# densityAdaptiveKernel

##### Adaptive Kernel Estimate of Intensity of Point Pattern

Computes an adaptive estimate of the intensity function of a point pattern using a variable-bandwidth smoothing kernel.

##### Usage

`densityAdaptiveKernel(X, …)`# S3 method for ppp
densityAdaptiveKernel(X, bw, …,
weights=NULL,
at=c("pixels", "points"),
edge=TRUE, ngroups)

##### Arguments

- X
Point pattern (object of class

`"ppp"`

).- bw
Numeric vector of smoothing bandwidths for each point in

`X`

, or a pixel image giving the smoothing bandwidth at each spatial location, or a spatial function of class`"funxy"`

giving the smoothing bandwidth at each location. The default is to compute bandwidths using`bw.abram`

.- …
Arguments passed to

`bw.abram`

to compute the smoothing bandwidths if`bw`

is missing, or passed to`as.mask`

to control the spatial resolution of the result.- weights
Optional vector of numeric weights for the points of

`X`

.- at
String specifying whether to compute the intensity values at a grid of pixel locations (

`at="pixels"`

) or only at the points of`x`

(`at="points"`

).- edge
Logical value indicating whether to perform edge correction.

- ngroups
Number of groups into which the bandwidth values should be partitioned and discretised.

##### Details

This function computes a spatially-adaptive kernel estimate of the
spatially-varying intensity from the point pattern `X`

using the partitioning technique of Davies and Baddeley (2018).

The argument `bw`

specifies the smoothing bandwidths to be
applied to each of the points in `X`

. It may be a numeric vector
of bandwidth values, or a pixel image or function yielding the
bandwidth values.

If the points of `X`

are \(x_1,\ldots,x_n\)
and the corresponding bandwidths are
\(\sigma_1,\ldots,\sigma_n\)
then the adaptive kernel estimate of intensity at a location \(u\) is
$$
\hat\lambda(u) = \sum_{i=1}^n k(u, x_i, \sigma_i)
$$
where \(k(u, v, \sigma)\) is the value at \(u\)
of the (possibly edge-corrected) smoothing kernel with bandwidth \(\sigma\)
induced by a data point at \(v\).

Exact computation of the estimate above can be time-consuming: it takes \(n\) times longer than fixed-bandwidth smoothing.

The partitioning method of Davies and Baddeley (2018)
accelerates this computation by partitioning the range of
bandwidths into `ngroups`

intervals,
correspondingly subdividing the points of the pattern `X`

into
`ngroups`

sub-patterns according to bandwidth,
and applying fixed-bandwidth smoothing to each sub-pattern.

The default value of `ngroups`

is the integer part of the square root of
the number of points in `X`

, so that the computation time is
only about \(\sqrt{n}\) times slower than fixed-bandwidth
smoothing. Any positive value of `ngroups`

can be specified by the user. Specifying `ngroups=Inf`

enforces exact
computation of the estimate without partitioning. Specifying
`ngroups=1`

is the same as fixed-bandwidth smoothing with
bandwidth `sigma=median(bw)`

.

##### Value

If `at="pixels"`

(the default), the result is a pixel image.
If `at="points"`

, the result is a numeric vector with one entry
for each data point in `X`

.

##### Bandwidths and Bandwidth Selection

The function `densityAdaptiveKernel`

computes one adaptive estimate of the intensity,
determined by the smoothing bandwidth values `bw`

.

Typically the bandwidth values are computed by first computing
a pilot estimate of the intensity, then using `bw.abram`

to compute the vector of bandwidths according to Abramson's rule.
This involves specifying a global bandwidth `h0`

.

The default bandwidths may work well in many contexts, but for optimal
bandwidth selection, this calculation should be performed repeatedly with
different values of `h0`

to optimise the value of `h0`

.
This can be computationally demanding; we recommend
the function `multiscale.density`

in the sparr package
which supports much faster bandwidth selection, using the FFT
method of Davies and Baddeley (2018).

##### References

Davies, T.M. and Baddeley, A. (2018)
Fast computation of spatially adaptive kernel estimates.
*Statistics and Computing*, **28**(4), 937-956.

Hall, P. and Marron, J.S. (1988)
Variable window width kernel density estimates of probability
densities.
*Probability Theory and Related Fields*, **80**, 37-49.

Silverman, B.W. (1986)
*Density Estimation for Statistics and Data Analysis*.
Chapman and Hall, New York.

##### See Also

`density.ppp`

,
`adaptive.density`

,
`densityVoronoi`

,
`im.object`

.

See the function `bivariate.density`

in the sparr package
for a more flexible implementation, and
`multiscale.density`

for an implementation that is more
efficient for bandwidth selection.

##### Examples

```
# NOT RUN {
Z <- densityAdaptiveKernel(redwood, h0=0.1)
plot(Z, main="Adaptive kernel estimate")
points(redwood, col="white")
# }
```

*Documentation reproduced from package spatstat, version 1.63-0, License: GPL (>= 2)*