Learn R Programming

LaplacesDemon (version 16.1.0)

Matrices: Matrix Utility Functions

Description

These are utility functions for working with matrices.

Usage

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)

Arguments

N

This required argument accepts a positive integer that indicates the number of nodes.

x

This is a matrix (though as.symmetric.matrix also accepts vectors).

J

This required argument indicates the dimension of the integral and accepts a positive integer.

k

For as.parm.matrix, this is a required argument, indicating the dimension of the matrix. For as.symmetric.matrix, this is an optional argument that specifies the dimension of the symmetric matrix. This applies only when x is a vector. It defaults to NULL, in which case it calculates k <- (-1 + sqrt(1 + 8 * length(x)))/ 2.

K

This required argument indicates the accuracy and accepts a positive integer. Larger values result in many more integration nodes.

diag

Logical. If TRUE, then the elements in the main diagonal are also returned.

dims

This required argument indicates the dimension of the integral and accepts a positive integer.

Sigma

This is a covariance matrix, \(\Sigma\), and may be entered either as a matrix or vector.

Model

This is a model specification function. For more information, see LaplacesDemon.

parm

This is a vector of parameters passed to the model specification.

Data

This is the list of data passed to the model specification. For more information, see LaplacesDemon.

a,b

These optional arguments allow the elements of x to be bound to the interval \([a,b]\). For example, elements of a correlation matrix are in the interval \([-1,1]\).

restrict

Logical. If TRUE, then x[1,1] is restricted to 1. This is useful in multinomial probit, for example. The variable, LaplacesDemonMatrix, is created in a new environment, LDEnv so as.parm.matrix can keep track of changes from iteration to iteration.

rule

This is an optional argument that accepts a univariate Gauss-Hermite quadrature rule. Usually, this argument is left empty. A rule may be supplied that differs from the traditional rule, such as when constraints have been observed, and one or more nodes and weights were adjusted.

chol

Logical. If TRUE, then x is an upper-triangular matrix.

file

This is the name of the file from which the numeric data matrix will be imported or read.

header

Logical. When TRUE, the first row of the file must contain names of the columns, and will be converted to the column names of the numeric matrix. When FALSE, the first row of the file contains data, not column names.

Interval

This accepts a small scalar number for precision.

Method

This accepts a quoted string. For Hessian, it defaults to Method="Richardson", which uses Richardson extrapolation. For Jacobian, it defaults to Method="simple", which uses finite-differencing. Richardson Richardson extrapolation is more accurate, but slower to calculate. Since error due to finite-differencing propagates through to higher derivatives, finite-differencing should not be used when approximating a Hessian matrix. Another method called automatic differentiation is not currently available here, but should be more accurate, though even slower to calculate. Another popular alternative is to use the BayesianBootstrap on the data. For CovEstim, this accepts Method="Hessian", Method="Identity" (which simply assigns an identity matrix), Method="OPG" (which calculates the sum of outer products of record-level gradients), or Method="Sandwich", which is the sandwich estimator and combines the Hessian and OPG estimates.

nrow

This is the number of rows of the numeric matrix, and defaults to nrow=0. If the number is known, the function will perform noticeably faster when it does not have to check.

p.theta

This accepts a matrix of prior probabilities for a transition matrix, and defaults to NULL. If used, each row must sum to 1.

samples

This is the number of samples to take from the numeric matrix. When samples=0, sampling is not performed and the entire matrix is returned.

sep

This argument indicates a character with which it will separate fields when creating column vectors. For example, a read a comma-separated file (.csv), use sep=",".

size

This is the batch size to be used only when reading a numeric matrix that is larger than the available computer memory (RAM), and only when samples is greater than zero. Sampling of a big data matrix is performed by first determining the records to keep, and then reading batches, one by one, and keeping the matching records.

theta.y

This accepts a vector of posterior samples of a discrete Markov chain, and defaults to NULL. If used, the order of the samples affects the transition probability.

na.rm

Logical. When TRUE, rows with missing values are removed from the matrix after it is read. Rather than removing missing values, the user may consider imputing missing values inside the model, or before the model with the MISS function. Examples of within-model imputation may be found in the accompanying "Examples" vignette.

y.theta

This accepts a vector of data that are samples of a discrete distribution, and defaults to NULL. If used, the order of the samples affects the transition probability.

Details

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 eqnK 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.

References

Higham, N.J. (2002). "Computing the Nearest Correlation Matrix - a Problem from Finance". IMA Journal of Numerical Analysis, 22, p. 329--343.

See Also

BayesianBootstrap, Cov2Prec, cov2cor, ddirichlet, GaussHermiteQuadRule, isSymmetric, LaplaceApproximation, LaplacesDemon, lower.tri, MISS, Prec2Cov, solve, and upper.tri.