Generate the basis matrix for a particular
N, W Slepian sequence
family member, with the additional property that the smoother passes constants
without distortion. Can be quite slow execution due to the latter property.
ns for implementation with
Parallel implementation for
mgcv included in package as
the predictor variable. Missing values are allowed. Assumed to be contiguous;
if not, then converted to a contiguous series to determine appropriate
the time bandwidth. Computed as the frequency domain analogue of the maximum period of
interest for a time series-regression problem using “smooth functions of time”. For example,
a period choice of 2 months converts to 60 days and
the number of basis vectors requested. If not provided, then
the time step for the input
a flag for returning the naive (default) Slepian basis vectors
a flag for choosing between a SLP2 or SLP3 basis. Type-2 bases capture (absorb) means of target series, while Type-3 bases ignore (pass) means.
a flag for using the built-in
a flag for forced computation of the basis vectors. Several combinations of commonly
a flag for returning the projection matrix
slp is based around the routine
.dpss, which generates a family of Discrete
Prolate Spheroidal (Slepian) Sequences. These vectors are orthonormal, have alternating
even/odd parity, and form the optimally concentrated basis set for the subspace of
R^N corresponding to the bandwidth
W. Full details are given
in Slepian (1978). These basis functions have natural boundary conditions, and lack any form of
knot structure. This version is returned for
naive = TRUE.
dpss basis vectors can be adapted to provide the additional
useful property of capturing or passing constants perfectly. That is, the smoother matrix
S formed from the returned rectangular matrix will either reproduce constants
at near round-off precision, i.e.,
S %*% rep(1, N) = rep(1, N),
naive = FALSE with
intercept = TRUE, or will pass constants,
S %*% rep(1, N) = rep(0, N), for
naive = FALSE with
intercept = FALSE.
The primary use is in modeling formula to directly specify a Slepian time-based smoothing term in a model: see the examples.
N this routine can be very slow. If you are computing models with
N, we highly recommend pre-computing the basis object, then using it
in your models without recomputation. The third example below demonstrates this approach.
A matrix of dimension
length(x) * K or
length(x) * (K-1) where
K was supplied, or
W was supplied and
K converted. Note that the
basis vectors are computed on a contiguous grid based on
x, and then
back-converted to the time structure of
Attributes are returned that correspond to the arguments to
and explicitly give
Thomson, D.J (1982) Spectrum estimation and harmonic analysis. Proceedings of the IEEE. Volume 70, number 9, pp. 1055-1096.
Slepian, David (1978) Prolate Spheroidal Wave Functions, Fourier Analysis, and Uncertainty V: the Discrete Case. Bell System Technical Journal. Volume 57, pp. 1371-1429.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
# Examples using pkg:gam library("gam") library("slp") N <- 730 W <- 14 / N K <- 28 # will actually use 27 df when intercept = FALSE x <- rnorm(n = N, sd = 1) y <- x + rnorm(n = N, sd = 2) + 5.0 t <- seq(1, N) # note: all three examples share identical results # example with in-call computation, using K (df) fit1 <- gam(y ~ x + slp(t, K = K, forceC = TRUE), family = gaussian) # example with in-call computation, using W fit2 <- gam(y ~ x + slp(t, W = W, forceC = TRUE), family = gaussian) # example with out-of-call computation, using K timeBasis <- slp(t, K = K, forceC = TRUE) fit3 <- gam(y ~ x + timeBasis, family = gaussian) # the same computations can be done using pre-computed basis vectors # for significant speed-ups, especially for large N - see `checkSaved' # for more details fit4 <- gam(y ~ x + slp(t, W = W, forceC = FALSE))