These are utility functions for working with matrices.
as.indicator.matrix(x)
as.inverse(x)
as.parm.matrix(x, k, parm, Data, a=-Inf, b=Inf, restrict=FALSE, chol=FALSE)
as.positive.definite(x)
as.positive.semidefinite(x)
as.symmetric.matrix(x, k=NULL)
is.positive.definite(x)
is.positive.semidefinite(x)
is.square.matrix(x)
is.symmetric.matrix(x)
Cov2Cor(Sigma)
CovEstim(Model, parm, Data, Method="Hessian")
GaussHermiteCubeRule(N, dims, rule)
Hessian(Model, parm, Data, Interval=1e-6, Method="Richardson")
Jacobian(Model, parm, Data, Interval=1e-6, Method="simple")
logdet(x)
lower.triangle(x, diag=FALSE)
read.matrix(file, header=FALSE, sep=",", nrow=0, samples=0, size=0, na.rm=FALSE)
SparseGrid(J, K)
TransitionMatrix(theta.y=NULL, y.theta=NULL, p.theta=NULL)
tr(x)
upper.triangle(x, diag=FALSE)The as.indicator.matrix function creates an indicator matrix
from a vector. This function is useful for converting a discrete
vector into a matrix in which each column represents one of the
discrete values, and each occurence of that value in the related
column is indicated by a one, and is otherwise filled with
zeroes. This function is similar to the class.ind function in
the nnet package.
The as.inverse function returns the matrix inverse of
x. The solve function in base R also returns the matrix
inverse, but solve can return a matrix that is not symmetric,
and can fail due to singularities. The as.inverse function
tries to use the solve function to return a matrix inverse, and
when it fails due to a singularity, as.inverse uses eigenvalue
decomposition (in which eigenvalues below a tolerance are replaced
with the tolerance), and coerces the result to a symmetric
matrix. This is similar to the solvcov function in the fpc
package.
The as.parm.matrix function prepares a correlation, covariance,
or precision matrix in two important ways. First,
as.parm.matrix obtains the parameters for the matrix specified
in the x argument by matching the name of the matrix in the
x argument with any parameters in parm, given the
parameter names in the Data listed in parm.names. These
obtained parameters are organized into a matrix as the elements of the
upper-triangular, including the diagonal. A copy is made, without the
diagonal, and the lower-triangular is filled in, completing the
matrix. Second, as.parm.matrix checks for
positive-definiteness. If matrix x is positive-definite, then
the matrix is stored as a variable called LaplacesDemonMatrix
in a new environment called LDEnv. If matrix x is not
positive-definite, then LaplacesDemonMatrix in LDEnv is
sought as a replacement. If this variable exists, then it is used to
replace the matrix. If not, then the matrix is replaced with an
identity matrix. Back in the model specification, after using
as.parm.matrix, it is recommended that the user also pass the
resulting matrix back into the parm vector, so the sampler or
algorithm knows that the elements of the matrix have changed.
The as.positive.definite function returns the nearest
positive-definite matrix for a matrix that is square and symmetric
(Higham, 2002). This version is intended only for covariance and
precision matrices, and has been optimized for speed. A more
extensible function is nearPD in the matrixcalc package, which
is also able to work with correlation matrices, and matrices that are
asymmetric.
The as.positive.semidefinite function iteratively seeks to
return a square, symmetric matrix that is at least
positive-semidefinite, by replacing each negative eigenvalue and
calculating its projection. This is intended only for covariance and
precision matrices. A similar function is makePsd in the RTAQ
package, though it is not iterative, and returns matrices that fail a
logical check with is.positive.semidefinite.
The as.symmetric.matrix function accepts either a vector or
matrix, and returns a symmetric matrix. In the case of a vector, it
can be either all elements of the matrix, or the lower triangular. In
the case of a x being entered as a matrix, this function
tolerates non-finite values in one triangle (say, the lower), as long
as the corresponding element is finite in the other (say, the upper)
triangle.
The Cov2Cor function converts a covariance matrix into a
correlation matrix, and accepts the covariance matrix either in matrix
or vector form. This function may be useful inside a model
specification and also with converting posterior draws of the elements
of a covariance matrix to a correlation matrix. Cov2Cor is an
expanded form of the cov2cor function in the stats
package, where Cov2Cor is also able to accept and return a
vectorized matrix.
The CovEstim function estimates a covariance matrix with one of
several methods. This is mainly used by
LaplaceApproximation, where the parm argument
receives the posterior modes. See the CovEst argument for
more details.
The GaussHermiteCubeRule function returns a matrix of nodes and
a vector of weights for a dims-dimensional integral given
\(N\) univariate nodes. The number of multivariate nodes will differ
from the number of univariate nodes. This function is for use with
multivariate quadrature, often called cubature. This has been adapted
from the multiquad function in the NominalLogisticBiplot
package. The GaussHermiteQuadRule function is a
univariate version. A customized univariate rule may be
supplied when constraints necessitate that one or more nodes and
weights had to be altered.
The Hessian returns a symmetric, Hessian matrix, which is a
matrix of second partial derivatives. The estimation of the Hessian
matrix is approximated numerically using Richardson extrapolation by
default. This is a slow function. This function is not intended to be
called by the user, but is made available here. This is essentially
the hessian function from the numDeriv package, adapted to
Laplace's Demon.
The is.positive.definite function is a logical test of whether
or not a matrix is positive-definite. A \(k \times k\)
symmetric matrix \(\textbf{X}\) is positive-definite if all of
its eigenvalues are positive (\(\lambda_i > 0, i \in k\)). All main-diagonal elements must be positive. The
determinant of a positive-definite matrix is always positive, so a
positive-definite matrix is always nonsingular. Non-symmetric,
positive-definite matrices exist, but are not considered here.
The is.positive.semidefinite function is a logical test of
whether or not a matrix is positive-semidefinite. A \(k x k\)
symmetric matrix \(\textbf{X}\) is positive-semidefinite if all
of its eigenvalues are non-negative (\(\lambda_i \ge 0, i \in
k\)).
The is.square.matrix function is a logical test of whether or
not a matrix is square. A square matrix is a matrix with the same
number of rows and columns, and is usually represented as a \(k
\times k\) matrix \(\textbf{X}\).
The is.symmetric.matrix function is a logical test of whether
or not a matrix is symmetric. A symmetric matrix is a square matrix
that is equal to its transpose, \(\textbf{X} = \textbf{X}^T\). For example, where \(i\) indexes rows and \(j\) indexes
columns, \(\textbf{X}_{i,j} = \textbf{X}_{j,i}\). This differs from the isSymmetric function in base R
that is inexact, using all.equal.
The Jacobian function estimates the Jacobian matrix, which is
a matrix of all first-order partial derivatives of the Model.
The Jacobian matrix is estimated by default with forward
finite-differencing, or optionally with Richardson extrapolation. This
function is not intended to be called by the user, but is made
available here. This is essentially the jacobian function from
the numDeriv package, adapted to LaplacesDemon.
The logdet function returns the logarithm of the determinant of
a positive-definite matrix via the Cholesky decomposition. The
determinant is a value associated with a square matrix, and was used
historically to determine if a system of linear equations has a
unique solution. The term determinant was introduced by Gauss,
where Laplace referred to it as the resultant. When the determinant is
zero, the matrix is singular and non-invertible; there are either no
solutions or many solutions. A unique solution exists when the
determinant is non-zero. The det function in base R works well
for small matrices, but can return erroneously return zero in larger
matrices. It is better to work with the log-determinant.
The lower.triangle function returns a vector of the lower
triangular elements of a matrix, and the diagonal is included when
diag=TRUE.
The read.matrix function is provided here as one of many
convenient ways to read a numeric matrix into R. The most common
method of storing data in R is the data frame, because it is
versatile. For example, a data frame may contain character, factor,
and numeric variables together. For iterative estimation, common in
Bayesian inference, the data frame is much slower than the numeric
matrix. For this reason, the LaplacesDemon package does not use data
frames, and has not traditionally accepted character or factor
data. The read.matrix function returns either an entire numeric
matrix, or row-wise samples from a numeric matrix. Samples may be
taken from a matrix that is too large for available computer memory
(RAM), such as with big data.
The SparseGrid function returns a sparse grid for a
\(J\)-dimensional integral with accuracy \(K\), given
Gauss-Hermite quadrature rules. A grid of order \(K\) provides an
exact result for a polynomial of total order of \(2K - 1\) or less.
SparseGrid returns a matrix of nodes and a vector of weights.
A sparse grid is more efficient than the full grid in the
GaussHermiteCubeRule function. This has been adapted from the
SparseGrid package.
The TransitionMatrix function has several uses. A user may
supply a vector of marginal posterior samples of a discrete Markov
chain as theta.y, and an observed posterior transition matrix
is returned. Otherwise, a user may supply data (y.theta) and/or
a prior (p.theta), in which case a posterior transition matrix
is returned. A common row-wise prior is the dirichlet distribution.
Transition probabilities are from row element to column element.
The tr function returns the trace of a matrix. The trace of a
matrix is the sum of the elements in the main diagonal of a square
matrix. For example, the trace of a \(k \times k\) matrix
\(\textbf{X}\), is \(\sum_{k=1} \textbf{X}_{k,k}\).
The upper.triangle function returns a vector of the lower
triangular elements of a matrix, and the diagonal is included when
diag=TRUE.
Higham, N.J. (2002). "Computing the Nearest Correlation Matrix - a Problem from Finance". IMA Journal of Numerical Analysis, 22, p. 329--343.
BayesianBootstrap,
Cov2Prec,
cov2cor,
ddirichlet,
GaussHermiteQuadRule,
isSymmetric,
LaplaceApproximation,
LaplacesDemon,
lower.tri,
MISS,
Prec2Cov,
solve, and
upper.tri.