DCSmooth

Bastian Schaefer

18.10.2021

1 Introduction

This vignette describes the use of the DCSmooth-package and its functions. The DCSmooth package provides some tools for non- and semiparametric estimation of the mean surface \(m\) of an observed sample of some function \[y(x, t) = m(x, t) + \varepsilon(x, t).\]

This package is accompanying Schäfer and Feng (2021). Boundary modification procedures for local polynomial smoothing are considered by Feng and Schäfer (2021). For a more comprehensive and detailed description of the algorithms than in section 4, please refer to these papers.

The DCSmooth contains the following functions, methods and data sets:

Functions
set.options() Define options for the dcs()-function.
dcs() Nonparametric estimation of the expectation function of a matrix Y. Includes automatic iterative plug-in bandwidth selection.
surface.dcs() 3d-plot for the surfaces of an "dcs"-object.
sarma.sim() Simulate a SARMA-model.
sarma.est() Estimate the parameters of a SARMA-model.
sfarima.sim() Simulate a SFARIMA-model.
sfarima.est() Estimate the parameters of a SFARIMA-model.
kernel.assign() Assign a pointer to a kernel function.
kernel.list() Print list of kernels available in the package.
Methods/Generics
summary.dcs() Summary statistics for an object of class "dcs".
print.dcs() Print an object of class "dcs".
plot.dcs() Plot method for an "dcs"-object, returns contour plot.
residuals.dcs() Returns the residuals of the regression from an "dcs"-object.
print.summary_dcs() Print an object of class "summary_dcs", which inherits from summary.dcs().
print.set_options() Prints an object of class "dcs_options", which inherits from set.options()
summary.sarma() Summary statistics for an object of class "sarma"
print.summary_sarma() Prints an object of class "summary_sarma", which inherits from summary.sarma()
summary.sfarima() Summary statistics for an object of class "sfarima"
print.summary_sfarima() Prints an object of class "summary_sfarima", which inherits from summary.sfarima()
Data
y.norm1 A surface with a single gaussian peak.
y.norm2 A surface with two gaussian peaks.
y.norm3 A surface with two gaussian ridges.
temp.nunn Temperatures in Nunn, CO observed in 2020 in 5 minute intervals. (Source: NOAA)
temp.yuma Temperatures in Yuma, AZ observed in 2020 in 5 minute intervals. (Source: NOAA)
wind.nunn Wind speed in Nunn, CO observed in 2020 in 5 minute intervals. (Source: NOAA)
wind.yuma Wind speed in Yuma, AZ observed in 2020 in 5 minute intervals. (Source: NOAA)
returns.alv 5 minute returns of Allianz SE from 2007 to 2010
volumes.alv 5 minute volumes of Allianz SE from 2007 to 2010

2 Details of Functions, Methods and Data

2.1 Functions

set.options()

This auxiliary function is used to set the options for the dcs function. An object of class dcs_options is created and should be used as dcs_options- argument in the dcs function.

Arguments of set.options() are

  • type Specifies the regression type. Supported methods are kernel regression ("KR") and local polynomial regression ("LP"), which is the default value.
  • kerns A character vector of length 2 stating the identifiers for the kernels in each dimension to use. The first element corresponds to the smoothing conditional on rows, the second conditional on columns. The identifiers are of the form \(X\_k\mu\nu\), where \(X\) indicates the smoothing method to use, either one of M, MW or T. The value \(k\) is the kernel order, \(\mu\) is the smoothness degree and \(\nu\) the derivative estimated by the kernel, which must match the order of derivative drv. For more information on the kernels see section 4.3, a list of available kernels is given in A.1. The default kernels are "MW_220" for both dimensions.
  • drv Derivative \((\nu_x, \nu_t)\) of \(m(x,t)\) to be estimated. Note that \(k \geq \nu + 2\), hence, only estimation of derivatives corresponding to kernels available is possible.
  • var_model Specifies the model assumption and estimation method for the errors/innovations \(\varepsilon(x, t)\) in the regression model. The form is given by "model_method" Currently available are
    • "iid" (independently identically distributed errors with variance estimation from the residuals, set as default).
    • "sarma_sep" (separable spatial ARMA (SARMA) process, two univariate processes in both directions are estimated via stats::arima).
    • "sarma_HR" (fast estimation of an SARMA process by the Hannan-Rissanen algorithm).
    • "sarma_RSS (estimation of an separable SARMA process by numerical minimization of the RSS).
    • "sfarima_RSS" (estimation of a separable spatial FARIMA (SFARIMA) model by numerical minimization of the RSS).

The models and estimation methods are described in more detail in Section 4.4.

The option var_est with values c("iid", "qarma", "sarma", "lm") has been replaced by var_model. For compatibility reasons, the option can still be used in set.options but will be converted to var_model

  • Additional arguments passed to set.options. The default values of these options typically depend on other options and thus are put in an ellipsis. Accepted arguments are
    • IPI_options Advanced options for tuning the parameters of the iterative plug-in algorithm of the bandwidth selection. These options include 2-element vectors for the inflation parameters (infl_par), the inflation exponents (infl_exp) and trimming parameters for stabilized estimation of the necessary derivatives (trim). Another option to further stabilize estimation of derivatives at the boundaries is the use of a constant estimation window at the boundaries setting the logical flag const_window to TRUE. The default values for the IPI-options depend partly on the regression type and error model selected and are given below. For convenience, the contents of IPI_options can be passed directly into the ellipsis ... of set.options().
    • model_order controls the order of the parametric error term model if an SARMA or SFARIMA model is used. This can be either a list of the form list(ar = c(1, 1), ma = c(1, 1)) (the default for SARMA and SFARIMA) specifying the model order, or any of c("aic", "bic", "gpac") specifying an order selection criterion. Note that gpac does not work under SFARIMA errors.
    • order_max Controls the maximum order if an order selection process is chosen in model_order. Is a list of the form list(ar = c(1, 1), ma = c(1, 1)) (the default).

set.options() returns an object of class "dcs_options including the following values

  • type Inherited from input.
  • kerns Inherited from input.
  • drv Inherited from input.
  • p_order A numeric vector of length 2, computed from drv by \(p_k = \nu_k + 1, k = x, t\).
  • var_model Inherited from input.
  • IPI_options Options for the iterative-plug in algorithm for bandwidth selection. If unchanged, values are set conditional on type (see default values for KR and LP below).
  • add_options A list containing the additional options model_order and order_max if available.

Every argument of the set.options function has a default value. Hence, just using set.options() will produce a complete set of options for double conditional smoothing regression in dcs (which is also implemented as default options in dcs, if the argument dcs_options is omitted).

Default options for kernel regression (type = "KR") are

#> dcs_options 
#> --------------------------------------- 
#> options for DCS   rows    cols 
#> --------------------------------------- 
#> type: kernel regression 
#> kernels used:        MW_220  MW_220
#> derivative:      0   0
#> variance model:  
#> ---------------------------------------
#> IPI options:
#> inflation parameters     2   1
#> inflation exponents  0.5 0.5
#> trim             0.05    0.05
#> constant window width    FALSE
#> ---------------------------------------

Default options for local polynomial regression (type = "LP") are

#> dcs_options 
#> --------------------------------------- 
#> options for DCS   rows    cols 
#> --------------------------------------- 
#> type: local polynomial regression 
#> kernel order:        MW_220  MW_220
#> derivative:      0   0
#> polynomial order:    1   1
#> variance model:  iid
#> --------------------------------------- 
#> IPI options:
#> inflation parameters     2   1
#> inflation exponents  auto     
#> trim             0.05    0.05
#> constant window width    FALSE
#> ---------------------------------------

dcs()

The dcs()-function is the main function of the package and includes IPI-bandwidth selection and non-parametric smoothing of the observations Y using the selected bandwidths. This function creates an object of class dcs, which includes the results of the DCS procedure.

Arguments of dcs() are

  • Y The matrix of observations to be smoothed via the DCS procedure. This matrix should only contain numeric values and no missing observations. For computational reasons, Y has to have at least five rows and columns, however, for reliable results the size should be larger.
  • dcs_options The options used for the smoothing and bandwidth selection. This should be an object of class "dcs_options" created by set.options(). This argument is optional, if omitted, all options will be set to their default values from the set.options() function.
  • h Either a two-value vector of positive numeric bandwidths or "auto" if bandwidth selection should be employed (the default).
  • parallel A logical flag if parallelization should be used for computation of the smoothed surfaces and its derivatives. If the order of the variance model is automatically selected, parallelization affects also this.
  • Further arguments to be passed to the function. This includes the equidistant covariates X and T which should be ordered numerical vectors whose length matches the number of rows of Y for X and the number of columns of Y for T.

dcs returns an object of class "dcs including the following values

  • X, T Vectors of covariates inherited from input or calculated to be equidistant on \([0,1]\) if these are omitted in the input.
  • Y Matrix of observations inherited from input.
  • M Matrix of smoothed values. If the argument h = "auto is used in dcs, the bandwidths are optimized via the IPI-algorithm, if h is set to fixed values, these bandwidth are used.
  • R Matrix of residuals computed from \(R = Y - M\).
  • h Bandwidths used for smoothing of Y. Either obtained by IPI bandwidth selection or given as argument in dcs.
  • c_f The estimated variance factor used in the last iteration of the bandwidth selection algorithm. Is set to NA, if no bandwidth selection is used.
  • var_est The estimated model obtained for the error terms (residuals) \(\varepsilon(x, t)\), i.e. the matrix $R. Output depends on the model specification used in dcs_options$var_model. For var_model = "iid", it contains the estimated standard deviation of the residuals and an indicator for stationarity, which is true by assumption. For dcs_options$var_model = c("sarma_sep", "sarma_RSS", "sarma_HR"), it contains the estimated model in an object of class "sarma" including the coefficient matrices $ar and $ma, the standard deviation $sigma as well as an stationarity indicator $stnry. For dcs_options$var_model = "sfarima_RSS", the output is of class "sfarima", with similar contents as "sarma" and the addition of the estimated long memory parameter vector $d.
  • dcs_options An object of class "dcs_options" containing the options used in the function.
  • iterations An integer reporting the number of iterations of the IPI algorithm. Is set to NA, if no bandwidth selection is used.
  • time_used A number reporting the time (in seconds) used for the IPI algorithm (total including all iterations). Is set to NA, if no bandwidth selection is used.

surface.dcs()

This function is a convenient wrapper for the plotly::plot_ly() function of the plotly package, for easy displaying of the considered surfaces. Direct plotting is available for any object of class "dcs" or any numeric matrix Y.

Arguments of surface.dcs() are

  • Y Either an object of class "dcs", inheriting from a call to dcs() or a numeric matrix, which is then directly passed to plotly::plot_ly().
  • trim A two-value vector which gives the percentage (between 0 and 0.5) of boundaries to leave out in the plot. Useful, if estimation at boundaries is unstable and has too high values compared to the inner, e.g. useful when estimation of derivatives is considered.
  • plot_choice Only used, if Y is an object of class "dcs". Specifies the surface to be plotted, 1 for the original observations, 2 for the smoothed surface and 3 for the residual surface. If plot_choice is omitted and Y is an "dcs"-object, a choice dialogue will be prompted to the console, which asks to state one of the available options.
  • Further arguments to be passed to the plotly::plot_ly() function.

surface.dcs() returns an object of class "plotly".

sarma.sim()

Simulation of a spatial ARMA process (SARMA). This function returns an object of class "sarma" with attribute "subclass" = "sim". The simulated innovations are created from a normal distribution with specified standard deviation \(\sigma\). This function uses a burn-in period for more consistent results.

Arguments of sarma.sim() are

  • n_x, n_t The dimension of the resulting matrix of observations, where n_x specifies the number of rows and n_t the number of columns. Initially, a matrix Y' of size \(2 n_x \times 2 n_t\) is simulated, for which simulation points with \(i \leq n_x\) or \(j \leq n_t\) are discarded (burn-in period).
  • model A list containing the model parameters to be used in the simulation. The argument should be a list of the form list(ar, ma, sigma). The values ar and ma are matrices of size \((p_x + 1) \times (p_t + 1)\) respective \((q_x + 1) \times (q_t + 1)\) and contain the coefficients in ascending lag order, so that the upper left entry is equal to 1 (for lag 0 in both dimensions). The standard deviation of the iid. innovations with zero mean is sigma, which should be a single positive number. See the examples in the application part 3.3 and the notation of SARMA models in Section 4.4.

sarma.sim() returns an object of class "sarma" with attribute "subclass" = "sim" including the following values:

  • Y The matrix of simulated values with size \(n_x \times n_t\) (determined by function arguments n_x, n_t). The matrix \(Y\) is the lower left \(n_x \times n_t\) submatrix of the actually simulated matrix \(Y'\) of size \(2 n_x \times 2 n_t\) to avoid effects from setting the initial values (burn-in period).
  • innov The \(n_x \times n_t\) matrix of iid. normally distributed innovations/errors of the SARMA model with zero mean and variance \(\sigma^2\) (determined by the function argument model\$sigma). As with Y, the original matrix has size \(2 n_x \times 2 n_t\).
  • model The model used for simulation, inherited from input.
  • stnry A flag indicating whether the simulated process is stationary.

sarma.est()

Estimate the parameters of an SARMA of given order. It returns an object of class "sarma" with attribute "subclass" = "est". For estimation, three methods are available.

Arguments of sarma.est() are

  • Y A (demeaned) matrix of observations, which contains only numeric values and no missing observations.
  • method A character string specifying the method for estimation. Currently supported methods are the Hannan-Rissanen algorithm ("HR"), a separable model using two univariate estimations via stats::arima ("sep") and a separable model which minimizes the residual sum of squares (RSS) of the model ("RSS").
  • model_order A list specifying the order of the SARMA to be estimated. This list should be of the form list(ar = c(1, 1), ma = c(1, 1)), where all orders should be non-negative integers. A SARMA\(((1,1), (1,1))\) model is estimated by default, if model_order is omitted.

sarma.est() returns an object of class "sarma" with attribute "subclass" = "est" including the following values:

  • Y The matrix of observations inherited from input.
  • innov The matrix of estimated innovations (residuals).
  • model A list of estimated model coefficients containing the matrices ar of autoregressive coefficients, the matrix ma of moving average coefficients as well as the standard deviation of residuals sigma.
  • stnry A flag indicating whether the estimated process is stationary.

sfarima.sim()

Simulation of a (separable) spatial fractional ARIMA (SFARIMA) process. This function returns an object of class "sfarima" with attribute "subclass" = "sim". The simulated innovations are created from a normal distribution with specified standard deviation \(\sigma\). This function uses a burn-in period for more consistent results.

Arguments of sfarima.sim() are

  • n_x, n_t The dimension of the resulting matrix of observations, where n_x specifies the number of rows and n_t the number of columns. Initially, a matrix Y' of size \(2 n_x \times 2 n_t\) is simulated, for which simulation points with \(i \leq n_x\) or \(j \leq n_t\) are discarded (burn-in period).
  • model A list containing the model parameters to be used in the simulation of the form list(ar, ma, d, sigma). The values ar and ma are matrices of size \((p_x + 1) \times (p_t + 1)\) respective \((q_x + 1) \times (q_t + 1)\) and containing the coefficients in ascending lag order, so that the upper left entry is equal to 1 (for lag 0 in both dimensions). The long-memory parameters \(d_x, d_t\) are stored in d, a numerical vector of length 2, with \(0 < d_x, d_t < 0.5\). The standard deviation of the iid. innovations with zero mean is sigma, which should be a single positive number. See the examples in the application part 3.4 and the notation of the short memory SARMA part in Section 4.4.

sfarima.sim() returns an object of class "sfarima" with attribute "subclass" = "sim" including the following values:

  • Y The matrix of simulated values with size \(n_x \times n_t\) (determined by function arguments n_x, n_t). The matrix \(Y\) is the lower left \(n_x \times n_t\) submatrix of the actually simulated matrix \(Y'\) of size \(2 n_x \times 2 n_t\) to avoid effects from setting the initial values (burn-in period).
  • innov The \(n_x \times n_t\) matrix of iid. normally distributed innovations/errors of the SFARIMA model with zero mean and variance \(\sigma^2\) (determined by the function argument model\$sigma). As with Y, the original matrix has size \(2 n_x \times 2 n_t\).
  • model The model used for simulation, inherited from input.
  • stnry A flag indicating whether the simulated process is stationary.

sfarima.est()

Estimation of an SFARIMA process. This function minimizes the residual sum of squares (RSS) to estimate the SFARIMA-parameters of a given order. It returns an object of class "sfarima" with attribute "subclass" = "est".

Arguments of sfarima.est() are

  • Y A (demeaned) matrix of observations, which contains only numeric values and no missing observations.
  • model_order A list specifying the order of the SFARIMA to be estimated. This list should be of the form list(ar = c(1, 1), ma = c(1, 1)). All orders should be non-negative integers. A SFARIMA\(((1,1), (1,1))\) model is estimated by default, if model_order is omitted.

sfarima.est() returns an object of class "sfarima" with attribute "subclass" = "est" including the following values:

  • Y The matrix of observations inherited from input.
  • innov The matrix of estimated innovations (residuals).
  • model A list of estimated model coefficients containing the matrices ar of autoregressive coefficients, the matrix ma of moving average coefficients as well as the vector d holding the long-memory parameters and the standard deviation of residuals sigma.
  • stnry A flag indicating whether the estimated process is stationary.

kernel.assign()

This function sets an external pointer to a specified boundary kernel available in the DCSmooth package. These kernels are functions \(K(u, q)\), where \(u\) is a vector on \([q, -1]\) and \(q \in [0, 1]\). The boundary kernels are as proposed by Müller and Wang (1994), Müller (1991) and constructed via the method described in Feng and Schäfer (2021). Available types are Müller-Wang (MW), Müller (M) and truncated kernels (T).

Arguments of kernel.assign() are

  • kernel_id The identifier for the kernel to be assigned. It is a character string of the form "X_abc", where X is specifies the type (MW, M, T), a i the kernel order \(k\), b is the degree of smoothness \(\mu\) and c is the order of derivative \(\nu\). A list of currently usable kernel identifiers can be accessed with the function kernel.list().

kernel.assign() returns an object of class "function", which points to a precompiled kernel function.

kernel.list()

kernel.list() prints the available identifiers for use in kernel.assign().

The argument of kernel.list() is

  • print A logical value indicating if the list of available kernels should be printed to the console.

kernel.list() returns a list including the available kernels as character strings, if the argument is print = FALSE.

Available kernels are

\(k\) \(\mu\) \(\nu\) Truncated Kernels Müller Kernels Müller-Wang Kernels
2 0 0 M_200 MW_200
2 1 0 M_210 MW_210
2 2 0 T_220 M_220 MW_220
3 2 1 T_321 M_321 MW_320
4 2 0 T_420 M_420 MW_420
4 2 1 M_421 MW_421
4 2 2 T_422 M_422 MW_422

2.2 Methods

The DCSmooth package contains the following methods

Function Methods/Generics available
dcs_options print, summary
dcs plot, print, print.summary, residuals, summary
sarma print.summary, summary
sfarima print.summary, summary

2.3 Data

This package contains three simulated example data sets and six data sets of environmental spatial time series.

Each of the three simulated example data sets is a matrix of size \(101 \times 101\) computed on \([0,1]^2\) for the following functions, where \(N(\mu, \Sigma)\) is the bivariate normal distribution with mean vector \(\mu\) and covariance matrix \(\Sigma\):

# surface.dcs(y.norm1)
# surface.dcs(y.norm2)
# surface.dcs(y.norm3)

The environmental application data sets feature the temperature and wind speed surfaces of Nunn, CO (temp.nunn, wind.nunn) and Yuma, AZ (temp.yuma, wind.yuma). The observations are taken in 2020 in 5-minute intervals. The temperatures are given in Celsius and wind speed is given in m/s. All data sets consist therefore of 288 columns (the intraday observations) and 366 rows (the days). The data is taken from the U.S. Climate Reference Network database at www.ncdc.noaa.gov. (see Diamond et al., 2013)

# surface.dcs(temp.nunn)
# surface.dcs(temp.yuma)
# surface.dcs(wind.nunn)
# surface.dcs(wind.yuma)

For examples of financial applications, the return and volume data of German insurance company Allianz SE is available in the package. The data is aggregated to the 5-minute level over the years 2007-2010, hence, the financial crisis 2008 is covered. The matrices consist of 1016 rows representing the days and 101 (returns) respective 102 (volumes) columns for the intraday 5-minute intervals.

# surface.dcs(returns.alv)
# surface.dcs(volumes.alv)

3 Application

The application of the package is demonstrated at the example of the simulated function y.norm1, which represents a gaussian peak on \([0,1]^2\) with \(n_x = n_t = 101\) evaluation points. Different models are simulated and estimation using dcs is demonstrated. Whenever default options are used, they are not explicitly used as function arguments, instead only when deviating from the defaults, the options are changed.

In order to keep the file size manageable, the surface.dcs commands in this vignette are commented out. Run the complete code used in section 3 of this vignette by demo("DCS_demo", package = "DCSmooth"), but note that it might take some time.

3.1 Defining the Options

In order to set specific options use the set.options() function to create an object of class "dcs_options".

opt1 = set.options(type = "KR", kerns = c("M_220", "M_422"), drv = c(0, 2),
                   var_model = "sarma_RSS",
                   IPI_options = list(trim = c(0.1, 0.1), infl_par = c(1, 1),
                                      infl_exp = c(0.7, 0.7), 
                                      const_window = TRUE),
                   model_order = list(ar = c(1, 1), ma = c(0, 0)))
summary(opt1)
#> dcs_options 
#> --------------------------------------- 
#> options for DCS   rows    cols 
#> --------------------------------------- 
#> type: kernel regression 
#> kernels used:        M_220   M_422
#> derivative:      0   2
#> variance model:  
#> ---------------------------------------
#> IPI options:
#> inflation parameters     1   1
#> inflation exponents  0.7 0.7
#> trim             0.1 0.1
#> constant window width    TRUE
#> ---------------------------------------
class(opt1)
#> [1] "dcs_options"

The contents of the advanced option IPI_options can be set directly as argument in set.options(). Changing these options might lead to non convergent bandwidths.

opt2 = set.options(type = "KR", kerns = c("M_220", "M_422"), drv = c(0, 2),
                   var_model = "sarma_RSS", trim = c(0.1, 0.1),
                   infl_par = c(1, 1), infl_exp = c(0.7, 0.7), 
                   const_window = TRUE,
                   model_order = list(ar = c(1, 1), ma = c(0, 0)))
summary(opt2)
#> dcs_options 
#> --------------------------------------- 
#> options for DCS   rows    cols 
#> --------------------------------------- 
#> type: kernel regression 
#> kernels used:        M_220   M_422
#> derivative:      0   2
#> variance model:  
#> ---------------------------------------
#> IPI options:
#> inflation parameters     1   1
#> inflation exponents  0.7 0.7
#> trim             0.1 0.1
#> constant window width    TRUE
#> ---------------------------------------
class(opt2)
#> [1] "dcs_options"

When using a model selection procedure, the additional option order_max is available:

opt3 = set.options(var_model = "sarma_sep", model_order = "bic",
                   order_max = list(ar = c(0, 1), ma = c(2, 2)))
summary(opt3)
#> dcs_options 
#> --------------------------------------- 
#> options for DCS   rows    cols 
#> --------------------------------------- 
#> type: local polynomial regression 
#> kernel order:        MW_220  MW_220
#> derivative:      0   0
#> polynomial order:    1   1
#> variance model:  sarma_sep
#> --------------------------------------- 
#> IPI options:
#> inflation parameters     2   1
#> inflation exponents  auto     
#> trim             0.05    0.05
#> constant window width    FALSE
#> ---------------------------------------

3.2 Application of the DCS with iid. errors

The example data set is simulated by using iid. errors:

# surface.dcs(y.norm1)

y_iid = y.norm1 + matrix(rnorm(101^2), nrow = 101, ncol = 101)
# surface.dcs(y_iid)

Kernel Regression with iid. errors

While local linear regression has some clear advantages over kernel regression, kernel regression is the faster method.

opt_iid_KR = set.options(type = "KR")
dcs_iid_KR = dcs(y_iid, opt_iid_KR)

# print results
dcs_iid_KR
#> dcs 
#> ------------------------------------------ 
#> DCS with automatic bandwidth selection:
#> ------------------------------------------ 
#>  Selected Bandwidths:
#>      h_x: 0.18875 
#>      h_t: 0.19282 
#>  Variance Factor:
#>      c_f: 1 
#> ------------------------------------------

# print options used for DCS procedure
dcs_iid_KR$dcs_options
#> dcs_options 
#> --------------------------------------- 
#> options for DCS  rows    cols 
#> --------------------------------------- 
#> type: kernel regression 
#> kernels used:        MW_220  MW_220
#> derivative:      0   0
#> variance model:  iid
#> ---------------------------------------

# plot regression surface
# surface.dcs(dcs_iid_KR, plot_choice = 2)

The summary of the "dcs"-object provides some more detailed information:

summary(dcs_iid_KR)
#> summary_dcs 
#> ------------------------------------------ 
#> DCS with automatic bandwidth selection:
#> ------------------------------------------ 
#> Results of kernel regression:
#> Estimated Bandwidths:     h_x:    0.1888 
#>           h_t:    0.1928 
#> Variance Factor:      c_f:    1 
#> Iterations:           4 
#> Time used (seconds):      0.03123 
#> ------------------------------------------ 
#> Variance Model:   iid 
#> ------------------------------------------ 
#> sigma:    0.9968943 
#> stationary:  TRUE
#> ------------------------------------------ 
#> See used parameter with "$dcs_options".

Local Polynomial regression with iid. errors

This is the default method, specification of options is not necessary. Note that local polynomial regression requires the bandwidth to cover at least the number of observations of the polynomial order plus one. For small bandwidths or too few observation points in one dimension, local polynomial regression might fail (“Bandwidth h must be larger for local polynomial regression.”). It is suggested to use kernel regression in this case.

dcs_LP_iid = dcs(y_iid)
dcs_LP_iid
#> dcs 
#> ------------------------------------------ 
#> DCS with automatic bandwidth selection:
#> ------------------------------------------ 
#>  Selected Bandwidths:
#>      h_x: 0.1941 
#>      h_t: 0.20778 
#>  Variance Factor:
#>      c_f: 1.0003 
#> ------------------------------------------

summary(dcs_LP_iid)
#> summary_dcs 
#> ------------------------------------------ 
#> DCS with automatic bandwidth selection:
#> ------------------------------------------ 
#> Results of local polynomial regression:
#> Estimated Bandwidths:     h_x:    0.1941 
#>           h_t:    0.2078 
#> Variance Factor:      c_f:    1 
#> Iterations:           4 
#> Time used (seconds):      0.04686 
#> ------------------------------------------ 
#> Variance Model:   iid 
#> ------------------------------------------ 
#> sigma:    0.9970611 
#> stationary:  TRUE
#> ------------------------------------------ 
#> See used parameter with "$dcs_options".

# plot regression surface
# surface.dcs(dcs_LP_iid, plot_choice = 2)

3.3 Application of the DCS with SARMA errors

A matrix containing innovations following a SARMA\(((p_x, p_t), (q_x, q_t))\) process can be obtained by using the sarma.sim() function. We use the following SARMA\(((1, 1), (1, 1))\)-process as example: \[ AR = \begin{pmatrix} 1 & 0.4 \\ -0.3 & -0.12 \end{pmatrix}, \quad MA = \begin{pmatrix} 1 & -0.5 \\ -0.2 & 0.1 \end{pmatrix} \quad \text{ and } \quad \sigma^2 = 0.25 \]

ar_mat = matrix(c(1, -0.3, 0.4, 0.12), nrow = 2, ncol = 2)
ma_mat = matrix(c(1, -0.2, -0.5, 0.1), nrow = 2, ncol = 2)
sigma  = sqrt(0.25)
model_list = list(ar = ar_mat, ma = ma_mat, sigma = sigma)
sim_sarma = sarma.sim(n_x = 101, n_t = 101, model = model_list)

# SARMA observations
y_sarma = y.norm1 + sim_sarma$Y
# surface.dcs(y_sarma)

Estimation of an SARMA process for a given order is implemented via the sarma.est() function (note that the simulated matrix can be accessed via $Y):

est_sarma = sarma.est(sim_sarma$Y, method = "HR",
                      model_order = list(ar = c(1, 1), ma = c(1, 1)))

summary(est_sarma)
#> ------------------------------------------ 
#> Estimation of SARMA((1,1),(1,1))
#> ------------------------------------------ 
#> sigma:    0.5008 
#> stationary:   TRUE 
#> ar:
#>         lag 0  lag 1
#> lag 0  1.0000 0.4042
#> lag 1 -0.3051 0.1291
#> 
#> ma:
#>         lag 0    lag 1
#> lag 0  1.0000 -0.49690
#> lag 1 -0.1804  0.07785

Local Polynomial regression with specified SARMA order

The dcs()-command is used with the default SARMA\(((1, 1), (1, 1))\) model (correctly specified) and with an SARMA\(((1, 1), (0, 0))\) (i.e. SAR\((1, 1)\)) model. The chosen estimation procedure is "sep":

# SARMA((1, 1), (1, 1))
opt_sarma_1 = set.options(var_model = "sarma_sep")
dcs_sarma_1 = dcs(y_sarma, opt_sarma_1)

summary(dcs_sarma_1$var_est)
#> ------------------------------------------ 
#> Estimation of SARMA((1,1),(1,1))
#> ------------------------------------------ 
#> sigma:    0.5441 
#> stationary:   TRUE 
#> ar:
#>         lag 0   lag 1
#> lag 0  1.0000  0.5003
#> lag 1 -0.5246 -0.2624
#> 
#> ma:
#>         lag 0    lag 1
#> lag 0  1.0000 -0.47000
#> lag 1 -0.1524  0.07165

# SARMA((1, 1), (0, 0))
# Use "sarma_HR" as it includes fast Yule-Walker estimation
opt_sarma_2 = set.options(var_model = "sarma_HR",
                          model_order = list(ar = c(1, 1), ma = c(0, 0)))
dcs_sarma_2 = dcs(y_sarma, opt_sarma_2)

summary(dcs_sarma_2$var_est)
#> ------------------------------------------ 
#> Estimation of SARMA((1,1),(0,0))
#> ------------------------------------------ 
#> sigma:    0.5331 
#> stationary:   TRUE 
#> ar:
#>        lag 0   lag 1
#> lag 0  1.000 0.63580
#> lag 1 -0.176 0.09775
#> 
#> ma:
#>       lag 0
#> lag 0     1

Local Polynomial regression with automated order selection

Automated order selection is used with model_order = c("aic", "bic", "gpac") in set.options(). The first one minimizes the AIC, the second one the BIC and the third uses a generalized partial autocorrelation function for order selection (not available for SFARIMA estimation). Order selection for large data sets is slowly in general, however, the "gpac" might be slightly faster than the other two. If automatic order selection is used, the argument order_max sets the maximum orders to be tested in the same way as model_order is usually defined as a list.

# BIC
opt_sarma_3 = set.options(var_model = "sarma_HR", model_order = "bic",
                          order_max = list(ar = c(2, 2), ma = c(2, 2)))
dcs_sarma_3 = dcs(y_sarma, opt_sarma_3)

summary(dcs_sarma_3$var_est)
#> ------------------------------------------ 
#> Estimation of SARMA((1,1),(1,1))
#> ------------------------------------------ 
#> sigma:    0.4997 
#> stationary:   TRUE 
#> ar:
#>         lag 0  lag 1
#> lag 0  1.0000 0.4031
#> lag 1 -0.3083 0.1271
#> 
#> ma:
#>         lag 0    lag 1
#> lag 0  1.0000 -0.50190
#> lag 1 -0.1879  0.07439

# gpac
opt_sarma_4 = set.options(var_model = "sarma_HR", model_order = "gpac",
                          order_max = list(ar = c(2, 2), ma = c(2, 2)))
dcs_sarma_4 = dcs(y_sarma, opt_sarma_4)

summary(dcs_sarma_4$var_est)
#> ------------------------------------------ 
#> Estimation of SARMA((2,2),(2,0))
#> ------------------------------------------ 
#> sigma:    0.5061 
#> stationary:   TRUE 
#> ar:
#>          lag 0    lag 1     lag 2
#> lag 0  1.00000  0.85200  0.291300
#> lag 1 -0.46200 -0.14640 -0.051140
#> lag 2  0.01744 -0.02795 -0.009702
#> 
#> ma:
#>          lag 0
#> lag 0  1.00000
#> lag 1 -0.33880
#> lag 2 -0.01287

3.4 Modeling errors with long memory

This package includes a bandwidth selection algorithm when the errors \(\varepsilon(x, t)\) follow an SFARIMA\(((p_x, p_t), (q_x, q_t))\) process with long memory. Order selection for SFARIMA models works exactly as in the SARMA case. We use the same SARMA model as in 3.3 with long-memory parameters \(d = (0.3, 0.1)\):

ar_mat = matrix(c(1, -0.3, 0.4, 0.12), nrow = 2, ncol = 2)
ma_mat = matrix(c(1, -0.2, -0.5, 0.1), nrow = 2, ncol = 2)
d = c(0.3, 0.1)
sigma  = sqrt(0.25)

model_list = list(ar = ar_mat, ma = ma_mat, d = d, sigma = sigma)
sim_sfarima = sfarima.sim(n_x = 101, n_t = 101, model = model_list)

# SFARIMA surface observations
y_sfarima = y.norm1 + sim_sfarima$Y

# surface.dcs(y_sfarima)

opt_sfarima = set.options(var_model = "sfarima_RSS")
dcs_sfarima = dcs(y_sfarima, opt_sfarima)

summary(dcs_sfarima$var_est)
#> ------------------------------------------ 
#> Estimation of SFARIMA((1,1),(1,1))
#> ------------------------------------------ 
#> d:        0.3146 0.09794 
#> SD (sigma):   0.4978 
#> stationary:   TRUE 
#> ar:
#>         lag 0    lag 1
#> lag 0  1.0000  0.38940
#> lag 1 -0.1575 -0.06132
#> 
#> ma:
#>          lag 0    lag 1
#> lag 0  1.00000 -0.50950
#> lag 1 -0.07279  0.03708

3.5 Estimation of derivatives

Local polynomial estimation is suitable for estimation of derivatives of a function or a surface. While estimation of derivatives works as well under dependent errors, the example uses the iid. model from 3.2. Derivatives can be computed for any derivative vector drv, if the values are non-negative and an appropriate kernel function is chosen (such that the derivative order of the kernel matches the derivative order in drv). Note that the order of the polynomials for the \(\nu\)th derivative is chosen to be \(p_i = \nu_i + 1, i = x, t\). As bandwidths increase with the order of the derivatives, the bandwidth might be large for higher derivative orders.

The estimator for \(m^{(1, 0)}(x, t)\) is

opt_drv_1 = set.options(drv = c(1, 0), kerns = c("MW_321", "MW_220"))
opt_drv_1$IPI_options$trim = c(0.1, 0.1)
dcs_drv_1 = dcs(y_iid, opt_drv_1)

dcs_drv_1
#> dcs 
#> ------------------------------------------ 
#> DCS with automatic bandwidth selection:
#> ------------------------------------------ 
#>  Selected Bandwidths:
#>      h_x: 0.24868 
#>      h_t: 0.41024 
#>  Variance Factor:
#>      c_f: 1.002 
#> ------------------------------------------
# surface.dcs(dcs_drv_1, trim = c(0.1, 0.1), plot_choice = 2)

The estimator for \(m^{(0, 2)}(x, t)\) is given by

opt_drv_2 = set.options(drv = c(0, 2), kerns = c("MW_220", "MW_422"))
opt_drv_2$IPI_options$trim = c(0.1, 0.1)

dcs_drv_2 = dcs(y_iid, opt_drv_2)

dcs_drv_2
#> dcs 
#> ------------------------------------------ 
#> DCS with automatic bandwidth selection:
#> ------------------------------------------ 
#>  Selected Bandwidths:
#>      h_x: 0.28678 
#>      h_t: 0.37157 
#>  Variance Factor:
#>      c_f: 1.002 
#> ------------------------------------------

# surface.dcs(dcs_drv_2, trim = c(0.1, 0.1), plot_choice = 2)

4 Mathematical Background

4.1 Double Conditional Smoothing

The double conditional smoothing (DCS, see Feng (2013)) is a spatial smoothing technique which effectively reduces the two-dimensional estimation to two one-dimensional estimation procedures. The DCS is defined for kernel regression as well as for local polynomial regression.

Classical bivariate (and multivariate) regression has been considered e.g. by Herrmann et al. (1995) (kernel regression) and Ruppert and Wand (1994) (local polynomial regression). The DCS provides now a faster and, especially for equidistant data, more efficient smoothing scheme, which leads to reduced computation time. For the DCS procedure implemented in this package, consider a \((n_x \times n_t)\)-matrix \(\mathbf{Y}\) of non-empty observations \(u_{i,j}\) and equidistant covariates \(X\), \(T\) on \([0,1]\), where \(X\) has length \(n_x\) and \(T\) has length \(n_t\). The model is then \[ y_{i,j} = m(x_i, t_j) + \varepsilon_{i,j} \] where \(m(x, t)\) is the mean or trend function, \(x_i \in X\), \(t_j \in T\) and \(\varepsilon\) is a random error function with zero mean. The model in matrix form is \(\mathbf{Y} = \mathbf{M}_1 + \mathbf{E}\) at the observation points.

The main assumption of the DCS is that of product kernels, i.e. the weights in the respective methods are constructed by \(K(u,v) = K_1(u) K_2(v)\). Now, a two stage smoother can be constructed by either the kernel weights directly (kernel regression) or by using locally weighted regression with kernels \(K_1\), \(K_2\), in any case, the weights are called \(\mathbf{W}_x\) and \(\mathbf{W}_t\). The DCS procedure implemented in DCSmooth smoothes over rows (conditioning on \(X\)) first and then over columns (conditioning on \(T\)), although switching the smoothing order is exactly equivalent. Hence, the DCS is given by the following equations: \[ \begin{align*} \mathbf{\widehat{M}}_0[ \ , j] &= \mathbf{Y} \cdot \mathbf{W}_t[ \ , j] \\ \mathbf{\widehat{M}}_1[i, \ ] &= \mathbf{W}_x [i, \ ] \cdot \mathbf{\widehat{M}}_0 \end{align*} \]

4.2 Bandwidth Selection

The bandwidth vector \(h = (h_x, h_t)\) is selected via an iterative plug-in (IPI) algorithm (Gasser et al., 1991). The IPI selects the optimal bandwidths by minimizing the mean integrated squared error (MISE) of the estimator. As the MISE includes derivatives of the regression surface \(m(x, t)\), auxiliary bandwidths for estimation of these derivatives are calculated via an inflation method. These inflation method connects the bandwidths of \(m(x, t)\) with that of a derivative \(m^{(\nu_x, \nu_t)}(x, t)\) by \[ \widetilde{h}_k = c_k \cdot h_k^{\alpha}, \quad k = x, t \] and is called exponential inflation method (EIM). The values of \(c_k\) are chosen on simulations, that of \(alpha\) are subject to the derivative of interest. The IPI now starts with an initial bandwidth \(h_0\) (chosen to be \(h_0 = (0.1, 0.1)\)) and calculates in each step \(s\) the auxiliary bandwidths \(\widetilde{h}_{k,s}\) from \(h_{s-1}\) and \(h_s\) from the smoothed derivative surfaces using \(\widetilde{h}_{k,s}\). The iteration process finishes until a certain threshold is reached.

4.3 Boundary Modification

In kernel regression, the boundary problem exists, which leads to biased estimated at the boundaries of the regression surface. This problem can (partially) be solved by means of suitable boundary kernels as introduced by Müller (1991) and Müller and Wang (1994). These boundary kernels differ in their degrees of smoothness and hence lead to different estimation results at the boundaries. However, all kernels are similar to the classical kernels in the interior region of the regression.

Following Feng and Schäfer (2021), a boundary modification is also defined for local polynomial regression. In the DCSmooth package, the local polynomial regression is always with boundary modification weights. Kernel types available (either for kernel regression or local polynomial regression) are Müller-type, Müller-Wang-type and truncated kernels, denoted by M, MW and T. In most applications, the Müller-Wang type are the preferred weighting functions.

For observations \(X = x_i\), \(x_i, i = 1, \dots, n_x\) and a given bandwidth \(h\), define \(u_r = \frac{x_r - X}{h} \in [-1, q]\). A (left) boundary kernel function \(K^l_q(u)\) of order \((k, \mu, \nu)\) is defined on \([-1, q]\), for \(q \in [0,1]\) and has the following properties \[ \int_{-1}^q u^j K^l_q(u) \text{d} u = \begin{cases} 0 & \text{ for } 0 \leq j < k, j \neq \nu\\ (-1)^\nu \nu! & \text{ for } j = \nu\\ \beta \neq 0 & \text{ for } j = k \end{cases} \] The corresponding right boundary kernels can be calculated by \(K^r_q(u) = (-1)\nu K_q(-u)\). The boundary kernels assigned by kernel.assign() are left boundary kernels.

4.4 Spatial ARMA processes

The SARMA process \(\varepsilon_{i,j}\) is given by the following equations: \[ \phi(B_1, B_2)\varepsilon_{i,j} = \psi(B_1, B_2)\eta_{i,j},\\ \] where the lag operators are \(B_1 \varepsilon_{i,j} = \varepsilon_{i-1, j}\) and \(B_2 \varepsilon_{i,j} = \varepsilon_{i,j-1}\), \(\xi \underset{iid.}{\sim} \mathcal{N}(0,\sigma^2)\) and \[ \phi(z_1, z_2) = \sum_{m = 0}^{p_1} \sum_{n = 0}^{p_2} \phi_{m, n} z_1^m z_2^n, \ \psi(z_1, z_2) = \sum_{m = 0}^{q_1} \sum_{n = 0}^{q_2} \psi_{m, n} z_1^m z_2^n. \] The coefficients \(\psi_{m,n}\) and \(\phi_{m,n}\) are written in matrix form \[ \begin{align*} \boldsymbol{\phi} = \begin{pmatrix} \phi_{0,0} & \dots & \phi_{0, p_2} \\ \vdots & \ddots & \\ \phi_{p_1, 0} & & \phi_{p_1, p_2} \end{pmatrix} \ \text{ and } \ \boldsymbol{\psi} = \begin{pmatrix} \psi_{0, 0} & \dots & \psi_{0, q_2} \\ \vdots & \ddots & \\ \psi_{q_1, 0} & & \psi_{q_1, q_2} \end{pmatrix}, \end{align*} \] where \(\Phi\) is the AR-part ($var_model$ar) and \(\Psi\) is the MA-part ($var_model$ma). The example from 3.3, \[ \begin{align} \boldsymbol{\phi} = \begin{pmatrix} 1 & 0.4 \\ -0.3 & -0.12 \end{pmatrix} \ \text{ and } \ \boldsymbol{\psi} = \begin{pmatrix} 1 & -0.2 \\ -0.5 & 0.1 \end{pmatrix} \end{align}, \] would then reduce to the process \[ \varepsilon_{i,j} = 0.4\varepsilon_{i,j-1} - 0.3 \varepsilon_{i-,j} + 0.2\varepsilon_{i-1,j-1} + 0.2 \xi_{i,j-1} + 0.2 \xi_{i-1,j} -0.5 \xi_{i-1,j-1} + \xi_{i,j}. \] Note that this process can be written as product of two univariate processes in the sense that \[ \phi_1(B_1)\phi_2(B_2)^T \varepsilon_{i,j} = \psi_1(B_1) \psi_2(B_2)^T \eta_{i,j},\\ \] with \[ \phi_1 = \begin{pmatrix} 1 \\ -0.3 \end{pmatrix}, \quad \phi_2 = \begin{pmatrix} 1 \\ 0.4 \end{pmatrix}, \quad \psi_1 = \begin{pmatrix} 1 \\ -0.5 \end{pmatrix}, \quad \psi_2 = \begin{pmatrix} 1 \\ -0.2 \end{pmatrix}. \] Hence, these process forms a separable SARMA. Estimation of separable SARMA models can be reduced to estimation of univariate ARMA models.

4.5 Estimation of SARMA processes

For estimation of SARMA models, three methods are implemented in DCSmooth:

Estimation of a separable SARMA by ML-estimation

This method is only available under the assumption of a separable model. Define two univariate time series \[ \varepsilon_{1,r} = \{\varepsilon_1\}_r = \{\varepsilon_{i,j}\}_{i + n_t (j - 1)}, \quad \varepsilon_{2,s} = \{\varepsilon_2\}_s = \{\varepsilon_{i,j}\}_{j + n_x * (i - 1)} \] for \(r,s = 1, \dots, n_x \cdot n_t\), \(i = 1, \dots, n_x\), \(j = 1, \dots, n_t\). The parameters \(\phi_1, \psi_1\) of \(\varepsilon_{1,r}\) and \(\phi_2, \psi_2\) of \(\varepsilon_{2,s}\) can then be estimated by well-known maximum likelihood estimators.

Least squares estimation using the RSS

The SARMA model can be rewritten as \[ \eta_{i,j} = \psi(B_1, B_2)^{-1}\phi(B_1, B_2)\varepsilon_{i,j},\\ \] which allows for an \(AR(\infty)\)-representation of the SARMA model \[ \eta_{i,j} = \sum_{r,s = 0}^{\infty} \theta^{AR}_{r,s} \varepsilon_{i-r, s-j} \]

From this, we can define the residual sum of squares (RSS) and get an estimate for the vector of coefficients \(\theta = c(\phi_{1,0}, \phi_{0,1}, \dots, \psi_{1,0}, \psi_{0,1}, \dots\) by \[ \widehat{\theta} = \arg \min RSS \approx \arg \min \sum_{i,j = 0}^{\infty} \eta_{i,j}^2 \]

Calculation of the $AR() representation of an SARMA model is difficult for a general SARMA but for a separable SARMA, the known univariate formulas hold.

These procedure can be directly used for SFARIMA models, if the long memory parameter \(d\) is included in \(\theta\) (see Beran et al. (2009)).

The Hannan-Rissanen Algorithm

The previously defined estimation methods require numeric optimization of some quantities and hence take more time for calculation on a computer. The Hannan-Rissanen algorithm (Hannan and Rissanen (1982)) provides a much faster estimation procedure. An extension to SARMA models is straightforward:

The main idea Hannan-Rissanen algorithm is to use a high-order SAR auxiliary model for initial estimation of the non observable innovations sequence. Then, a linear regression model is applied, which yields the SARMA-parameters from the observations and estimated innovations.

Let \(\{\varepsilon\}_{i,j}\) be the the ordered observations of the \(SARMA((p_x, p_t), (q_x, q_t))\) process \(\varepsilon(x, t)\) with \(i = 1, \dots, n_x, j = 1, \dots, n_t\). The SARMA parameters \(\mathbf{\phi, \mathbf{\psi}\) are then estimated by the modified Hannan-Rissanen algorithm: 1. Obtain the auxiliary residuals \(\widetilde{\eta}_{i,j}\) by fitting a high-order autoregressive model with \((\widetilde{p}_x, \widetilde{p}_t) \geq (p_x, p_t)\) to the observations: \[ \begin{align} \widetilde{\eta}_{i,j} = \begin{cases} \sum_{m = 0}^{\widetilde{p}_1} \sum_{n = 0}^{\widetilde{p}_2} \widetilde{\phi}_{m,n} \varepsilon_{i - m, j - n}, & \widetilde{p}_1 < i \leq n_1, \widetilde{p}_2 < j \leq n_2\\ 0, &1 \leq i \leq \widetilde{p}_1, 1 \leq j \leq \widetilde{p}_2 \end{cases} \end{align} \] where \(\widetilde{\phi}_{m,n}\) is estimated by the Yule-Walker equations and \(\phi_{0,0} = 1\). 2. Obtain \(\widehat{\phi}_{m,n}\) and \(\widehat{\psi}_{m,n}\) and the estimated innovations \(\widehat{\eta}_{i,j}\) by linear regression from \[ \begin{align} \varepsilon_{i,j} = -\sum_{\substack{m=0\\m \neq n = 0}}^{p_1} \sum_{\substack{n=0\\n \neq m = 0}}^{p_2} \widehat{\phi}_{m,n} \varepsilon_{i-m,j-n} + \sum_{\substack{m=0\\m \neq n = 0}}^{q_1} \sum_{\substack{n=0\\n \neq m = 0}}^{q_2} \widehat{\psi}_{m,n} \widetilde{\xi}_{i-m,j-n} + \widehat{\xi}_{i,j} \end{align} \] The resulting coefficients \(\widehat{\mathbf{\phi}}\), \(\widehat{\mathbf{\psi}}\) are then the estimates for the parameters.

The autocovariance function of the SARMA-process is \(\gamma(s,t) = \mathbb{E}(\varepsilon_{i,j} \varepsilon_{i+s, j+t})\). For \(\widetilde{p}_x, \widetilde{p}_t\), the spatial Yule-Walker equation for the \(SAR(\widetilde{p}_x, \widetilde{p}_t)\) approximation of the SARMA is then \[ \begin{align} \mathbf{\Gamma} \ \left( \text{vec} (\widetilde{\boldsymbol{\phi}}) \setminus \{ \widetilde{\phi}_{0,0}\} \right) = \gamma \end{align} \] where \(\mathbf{\Gamma}\) is the autocovariance matrix and \(\gamma\) the vector of autocovariances when using the following notation: \[ \begin{align} \mathbf{\Gamma} &= \begin{pmatrix} \gamma(0,0) & \gamma(1,0) & \dots & \gamma(\widetilde{p}_1,0) & \gamma(2,0) & \dots & \gamma(\widetilde{p}_1 - 1, \widetilde{p}_2)\\ \gamma(1,0) & \gamma(0,0) & & & & & \gamma(\widetilde{p}_1 - 2, \widetilde{p}_2)\\ \vdots & & \ddots & & & & \vdots \\ \gamma(\widetilde{p}_1 - 1, \widetilde{p}_2) & & \dots & & & & \gamma(0,0) \end{pmatrix} \\ \gamma &= \left(\gamma(0,1), \gamma(0,2), \dots, \gamma(\widetilde{p}_1, 0), \dots, \gamma(\widetilde{p}_1, \widetilde{p}_2) \right)^T. \end{align} \]

References

Beran, J., Ghosh, S. and Schell, D. (2009) On least squares estimation for long-memory lattice processes. Journal of Multivariate Analysis, 100, 2178–2194.
Feng, Y. (2013) Double-conditional smoothing of high-frequency volatility surface in a spatial multiplicative component GARCH with random effects. Center for International Economics Working Paper, 415, 643–652.
Feng, Y. and Schäfer, B. (2021) Boundary modification in local polynomial regression and associated boundary kernel functions. Preprint, Paderborn University.
Gasser, T., Kneip, A. and Köhler, W. (1991) A flexible and fast method for automatic smoothing. Journal of the American Statistical Association, 86, 643–652.
Hannan, E. J. and Rissanen, J. (1982) Recursive estimation of mixed autoregressive-moving average order. Biometrika, 69, 81–94.
Herrmann, E., Engel, J., Gasser, T., et al. (1995) A bandwidth selector for bivariate kernel regression. Journal of the Royal Statistical Society, 57, 171–180.
Müller, H.-G. (1991) Smooth optimum kernel estimators near endpoints. Biometrika, 78, 521–530.
Müller, H.-G. and Wang, J.-L. (1994) Hazard rate estimation under random censoring with varying kernels and bandwidths. Biometrics, 50, 61–76.
Ruppert, D. and Wand, M. P. (1994) Multivariate locally weighted least squares regression. The Annals of Statistics, 22, 1346–1370.
Schäfer, B. and Feng, Y. (2021) Fast computation and bandwidth selection algorithms for smoothing functional time series. Preprint, Paderborn University.