
Covariance models and their parameter settings
magCompCovModels.RdLists the covariance families, parameter names, and domains supported by GP
and NNGP terms. The same covariance calculations are used for fitting and
prediction. uniform() specifies a prior; fixed() marks a starting
value as fixed throughout fitting.
Details
Specify a family with cov_model inside gp(), nngp(),
factor_gp(), or factor_nngp(). The default is "exp".
For space-time families, the final coordinate is numeric time, for example
nngp(easting, northing, year, cov_model = "sep_exp", m = 15).
Time gaps are used as supplied. Let \(h\) be spatial distance, \(u\) absolute
time separation, and \(d\) the number of spatial coordinates. Correlations are:
exp\(\exp(-\phi h)\), with
phi > 0.matern\((\phi h)^\nu K_\nu(\phi h)/(2^{\nu-1}\Gamma(\nu))\), with
phi > 0andnu > 0. Atnu = 0.5this is exponential.sep_exp\(\exp(-\phi h-\lambda u)\), with positive
phiandlambda.multi_res_sep_exp\(\alpha\exp(-\phi_1h-\lambda_1u) +(1-\alpha)\exp(-\phi_2h-\lambda_2u)\). The mixing weight
alphais between zero and one; all four decays are positive.gneitingWrite \(T=1+a u^{2\alpha}\). Correlation is \(T^{-\delta-d/2}\exp(-c h^{2\gamma}/T^{\beta\gamma})\). Parameters
aandcare positive;alphaandgammaare in \((0,1]\);betais in \([0,1]\);deltais nonnegative.
Magnitude process variance is separate from these correlations and has an inverse-gamma Gibbs update. Composition factors have unit marginal variance.
All estimated covariance parameters have independent bounded uniform priors.
Put overrides in a term's covariance list under priors,
starting, or tuning. The same parameter names are used in
samples() and summary(). For composition, a common scalar setting
applies to all factors; a named vector of starts/tuning values or a named list
of priors can override individual factors. A named list of starting values can
mix fixed(value) and ordinary numeric values across factors.
Omitted entries retain defaults. magCompSettings() shows the resolved
supports and starts. Fixed parameters do not have priors or proposals.
By default, Matérn smoothness has prior uniform(0.05, 2). Gneiting
alpha = fixed(1), gamma = fixed(0.5), and delta = fixed(0);
a, c, and beta are estimated. An ordinary starting value or
a prior override releases a default-fixed parameter.
Other default supports are calibrated to spatial distances and the time span.
For the mixture, the default supports of phi_1 and phi_2 meet at
\(30/D\), where \(D\) is maximum spatial separation; component 1 describes
longer spatial dependence. Temporal supports are calibrated to dependence up
to twice the observed time span. User overrides are direct parameter bounds,
not range parameters. Overlapping mixture supports can permit label switching.
When a coordinate domain has no positive span, supply the corresponding
support or fix its parameter. The technical guide gives the default formulas.
Covariance parameters are proposed jointly within each magnitude process and
each composition factor, after bounded-logit transformation, with Jacobian
correction. The tuning entries are initial transformed-scale standard
deviations (default 0.1). magCompControl(metropolis = ) controls adaptation.
Value
magCompCovModels() returns a named list of registry entries.
uniform() returns a prior specification and fixed() a marked
numeric value for a covariance starting-value list.
Examples
magCompCovModels()$sep_exp
~ nngp(easting, northing, year, cov_model = "sep_exp", name = "space_time")
magCompPriors(magnitude = list(terms = list(space_time = list(
covariance = list(phi = uniform(0.1, 4), lambda = uniform(0.02, 2))))))
# A matching starting-value list for magComp(starting = ...).
list(magnitude = list(terms = list(space_time = list(
covariance = list(phi = 1, lambda = fixed(0.3))))))