Package ‘mvtnorm’

October 8, 2021
Title Multivariate Normal and t Distributions
Version 1.1-3
Date 2021-10-05
Description Computes multivariate normal and t probabilities, quantiles,
random deviates and densities.
Imports stats, methods
Depends R(>= 3.5.0)
License GPL-2
URL http://mvtnorm.R-forge.R-project.org
NeedsCompilation yes
Author Alan Genz [aut],
Frank Bretz [aut],
Tetsuhisa Miwa [aut],
Xuefei Mi [aut],
Friedrich Leisch [ctb],
Fabian Scheipl [ctb],
Bjoern Bornkamp [ctb] (<https://orcid.org/0000-0002-6294-8185>),
Martin Maechler [ctb] (<https://orcid.org/0000-0002-8685-9910>),
Torsten Hothorn [aut, cre] (<https://orcid.org/0000-0001-8301-0471>)
Maintainer Torsten Hothorn <[email protected]>
Repository CRAN
Date/Publication 2021-10-08 09:50:02 UTC

R topics documented:
algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
Mvnorm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Mvt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
pmvnorm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
pmvt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
qmvnorm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
qmvt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1
2 algorithms

Index 17

algorithms Choice of Algorithm and Hyper Parameters

Description
Choose between three algorithms for evaluating normal (and t-) distributions and define hyper pa-
rameters.

Usage
GenzBretz(maxpts = 25000, abseps = 0.001, releps = 0)
Miwa(steps = 128, checkCorr = TRUE, maxval = 1e3)
TVPACK(abseps = 1e-6)

Arguments
maxpts maximum number of function values as integer. The internal FORTRAN code
always uses a minimum number depending on the dimension. (for example 752
for three-dimensional problems).
abseps absolute error tolerance; for TVPACK only used for dimension 3.
releps relative error tolerance as double.
steps number of grid points to be evaluated; cannot be larger than 4097.
checkCorr logical indicating if a check for singularity of the correlation matrix should be
performed (once per function call to pmvt() or pmvnorm()).
maxval replacement for Inf when non-orthant probabilities involving Inf shall be com-
puted.

Details
There are three algorithms available for evaluating normal (and two algorithms for t-) probabilities:
The default is the randomized Quasi-Monte-Carlo procedure by Genz (1992, 1993) and Genz and
Bretz (2002) applicable to arbitrary covariance structures and dimensions up to 1000.
For normal probabilities, smaller dimensions (up to 20) and non-singular covariance matrices, the
algorithm by Miwa et al. (2003) can be used as well. This algorithm can compute orthant prob-
abilities (lower being -Inf or upper equal to Inf). Non-orthant probabilities are computed from
the corresponding orthant probabilities, however, infinite limits are replaced by maxval along with
a warning.
For two- and three-dimensional problems and semi-infinite integration region, TVPACK implements
an interface to the methods described by Genz (2004).

Value
An object of class "GenzBretz", "Miwa", or "TVPACK" defining hyper parameters.
Mvnorm 3

References
Genz, A. (1992). Numerical computation of multivariate normal probabilities. Journal of Compu-
tational and Graphical Statistics, 1, 141–150.
Genz, A. (1993). Comparison of methods for the computation of multivariate normal probabilities.
Computing Science and Statistics, 25, 400–405.
Genz, A. and Bretz, F. (2002), Methods for the computation of multivariate t-probabilities. Journal
of Computational and Graphical Statistics, 11, 950–971.
Genz, A. (2004), Numerical computation of rectangular bivariate and trivariate normal and t-probabilities,
Statistics and Computing, 14, 251–260.
Genz, A. and Bretz, F. (2009), Computation of Multivariate Normal and t Probabilities. Lecture
Notes in Statistics, Vol. 195. Springer-Verlag, Heidelberg.
Miwa, A., Hayter J. and Kuriki, S. (2003). The evaluation of general non-centred orthant probabil-
ities. Journal of the Royal Statistical Society, Ser. B, 65, 223–234.
Mi, X., Miwa, T. and Hothorn, T. (2009). mvtnorm: New numerical algorithm for multivariate
normal probabilities. The R Journal 1(1): 37–39. https://journal.r-project.org/archive/
2009-1/RJournal_2009-1_Mi+et+al.pdf

Mvnorm Multivariate Normal Density and Random Deviates

Description
These functions provide the density function and a random number generator for the multivariate
normal distribution with mean equal to mean and covariance matrix sigma.

Usage
dmvnorm(x, mean = rep(0, p), sigma = diag(p), log = FALSE, checkSymmetry = TRUE)
rmvnorm(n, mean = rep(0, nrow(sigma)), sigma = diag(length(mean)),
method=c("eigen", "svd", "chol"), pre0.9_9994 = FALSE, checkSymmetry = TRUE)

Arguments
x vector or matrix of quantiles. If x is a matrix, each row is taken to be a quantile.
n number of observations.
mean mean vector, default is rep(0,length = ncol(x)).
sigma covariance matrix, default is diag(ncol(x)).
log logical; if TRUE, densities d are given as log(d).
method string specifying the matrix decomposition used to determine the matrix root
of sigma. Possible methods are eigenvalue decomposition ("eigen", default),
singular value decomposition ("svd"), and Cholesky decomposition ("chol").
The Cholesky is typically fastest, not by much though.
4 Mvt

pre0.9_9994 logical; if FALSE, the output produced in mvtnorm versions up to 0.9-9993 is

reproduced. In 0.9-9994, the output is organized such that rmvnorm(10,...)
has the same first ten rows as rmvnorm(100,...) when called with the same
seed.
checkSymmetry logical; if FALSE, skip checking whether the covariance matrix is symmetric or
not. This will speed up the computation but may cause unexpected outputs when
ill-behaved sigma is provided. The default value is TRUE.

Author(s)
Friedrich Leisch and Fabian Scheipl

See Also
pmvnorm, rnorm, qmvnorm

Examples
dmvnorm(x=c(0,0))
dmvnorm(x=c(0,0), mean=c(1,1))

sigma <- matrix(c(4,2,2,3), ncol=2)

x <- rmvnorm(n=500, mean=c(1,2), sigma=sigma)
colMeans(x)
var(x)

x <- rmvnorm(n=500, mean=c(1,2), sigma=sigma, method="chol")

colMeans(x)
var(x)

plot(x)

Mvt The Multivariate t Distribution

Description
These functions provide information about the multivariate t distribution with non-centrality pa-
rameter (or mode) delta, scale matrix sigma and degrees of freedom df. dmvt gives the density
and rmvt generates random deviates.

Usage
rmvt(n, sigma = diag(2), df = 1, delta = rep(0, nrow(sigma)),
type = c("shifted", "Kshirsagar"), ...)
dmvt(x, delta = rep(0, p), sigma = diag(p), df = 1, log = TRUE,
type = "shifted", checkSymmetry = TRUE)
Mvt 5

Arguments
x vector or matrix of quantiles. If x is a matrix, each row is taken to be a quantile.
n number of observations.
delta the vector of noncentrality parameters of length n, for type = "shifted" delta
specifies the mode.
sigma scale matrix, defaults to diag(ncol(x)).
df degrees of freedom. df = 0 or df = Inf corresponds to the multivariate normal
distribution.
log logical indicating whether densities d are given as log(d).
type type of the noncentral multivariate t distribution. type = "Kshirsagar" corre-
sponds to formula (1.4) in Genz and Bretz (2009) (see also Chapter 5.1 in Kotz
and Nadarajah (2004)). This is the noncentral t-distribution needed for calculat-
ing the power of multiple contrast tests under a normality assumption. type =
"shifted" corresponds to the formula right before formula (1.4) in Genz and
Bretz (2009) (see also formula (1.1) in Kotz and Nadarajah (2004)). It is a lo-
cation shifted version of the central t-distribution. This noncentral multivariate
t distribution appears for example as the Bayesian posterior distribution for the
regression coefficients in a linear regression. In the central case both types co-
incide. Note that the defaults differ from the default in pmvt() (for reasons of
backward compatibility).
checkSymmetry logical; if FALSE, skip checking whether the covariance matrix is symmetric or
not. This will speed up the computation but may cause unexpected outputs when
ill-behaved sigma is provided. The default value is TRUE.
... additional arguments to rmvnorm(), for example method.

Details
If X denotes a random vector following a t distribution with location vector 0 and scale matrix Σ
(written X ∼ tν (0, Σ)), the scale matrix (the argument sigma) is not equal to the covariance matrix
Cov(X) of X. If the degrees of freedom ν (the argument df) is larger than 2, then Cov(X) =
Σν/(ν − 2). Furthermore, in this case the correlation matrix Cor(X) equals the correlation matrix
corresponding to the scale matrix Σ (which can be computed with cov2cor()). Note that the scale
matrix is sometimes referred to as “dispersion matrix”; see McNeil, Frey, Embrechts (2005, p. 74).
For type = "shifted" the density
c(1 + (x − δ)0 S −1 (x − δ)/ν)−(ν+m)/2
is implemented, where
c = Γ((ν + m)/2)/((πν)m/2 Γ(ν/2)|S|1/2 ),
S is a positive definite symmetric matrix (the matrix sigma above), δ is the non-centrality vector
and ν are the degrees of freedom.
df=0 historically leads to the multivariate normal distribution. From a mathematical point of view,
rather df=Inf corresponds to the multivariate normal distribution. This is (now) also allowed for
rmvt() and dmvt().
Note that dmvt() has default log = TRUE, whereas dmvnorm() has default log = FALSE.
6 pmvnorm

References
McNeil, A. J., Frey, R., and Embrechts, P. (2005). Quantitative Risk Management: Concepts,
Techniques, Tools. Princeton University Press.

See Also
pmvt() and qmvt()

Examples
## basic evaluation
dmvt(x = c(0,0), sigma = diag(2))

## check behavior for df=0 and df=Inf

x <- c(1.23, 4.56)
mu <- 1:2
Sigma <- diag(2)
x0 <- dmvt(x, delta = mu, sigma = Sigma, df = 0) # default log = TRUE!
x8 <- dmvt(x, delta = mu, sigma = Sigma, df = Inf) # default log = TRUE!
xn <- dmvnorm(x, mean = mu, sigma = Sigma, log = TRUE)
stopifnot(identical(x0, x8), identical(x0, xn))

## X ~ t_3(0, diag(2))
x <- rmvt(100, sigma = diag(2), df = 3) # t_3(0, diag(2)) sample
plot(x)

## X ~ t_3(mu, Sigma)
n <- 1000
mu <- 1:2
Sigma <- matrix(c(4, 2, 2, 3), ncol=2)
set.seed(271)
x <- rep(mu, each=n) + rmvt(n, sigma=Sigma, df=3)
plot(x)

## Note that the call rmvt(n, mean=mu, sigma=Sigma, df=3) does *not*
## give a valid sample from t_3(mu, Sigma)! [and thus throws an error]
try(rmvt(n, mean=mu, sigma=Sigma, df=3))

## df=Inf correctly samples from a multivariate normal distribution

set.seed(271)
x <- rep(mu, each=n) + rmvt(n, sigma=Sigma, df=Inf)
set.seed(271)
x. <- rmvnorm(n, mean=mu, sigma=Sigma)
stopifnot(identical(x, x.))

pmvnorm Multivariate Normal Distribution

pmvnorm 7

Description
Computes the distribution function of the multivariate normal distribution for arbitrary limits and
correlation matrices.

Usage
pmvnorm(lower=-Inf, upper=Inf, mean=rep(0, length(lower)),
corr=NULL, sigma=NULL, algorithm = GenzBretz(), keepAttr=TRUE, ...)

Arguments
lower the vector of lower limits of length n.
upper the vector of upper limits of length n.
mean the mean vector of length n.
corr the correlation matrix of dimension n.
sigma the covariance matrix of dimension n less than 1000. Either corr or sigma can
be specified. If sigma is given, the problem is standardized. If neither corr nor
sigma is given, the identity matrix is used for sigma.
algorithm an object of class GenzBretz, Miwa or TVPACK specifying both the algorithm to
be used as well as the associated hyper parameters.
keepAttr logical indicating if attributes such as error and msg should be attached to
the return value. The default, TRUE is back compatible.
... additional parameters (currently given to GenzBretz for backward compatibility
issues).

Details
This program involves the computation of multivariate normal probabilities with arbitrary corre-
lation matrices. It involves both the computation of singular and nonsingular probabilities. The
implemented methodology is described in Genz (1992, 1993) (for algorithm GenzBretz), in Miwa
et al. (2003) for algorithm Miwa (useful up to dimension 20) and Genz (2004) for the TVPACK
algorithm (which covers 2- and 3-dimensional problems for semi-infinite integration regions).
Note the default algorithm GenzBretz is randomized and hence slightly depends on .Random.seed
and that both -Inf and +Inf may be specified in lower and upper. For more details see pmvt.
The multivariate normal case is treated as a special case of pmvt with df=0 and univariate problems
are passed to pnorm.
The multivariate normal density and random deviates are available using dmvnorm and rmvnorm.

Value
The evaluated distribution function is returned, if keepAttr is true, with attributes

error estimated absolute error

msg status message(s).
algorithm a character string with class(algorithm).
8 pmvnorm

Source
http://www.sci.wsu.edu/math/faculty/genz/homepage

References
Genz, A. (1992). Numerical computation of multivariate normal probabilities. Journal of Compu-
tational and Graphical Statistics, 1, 141–150.
Genz, A. (1993). Comparison of methods for the computation of multivariate normal probabilities.
Computing Science and Statistics, 25, 400–405.
Genz, A. (2004), Numerical computation of rectangular bivariate and trivariate normal and t-probabilities,
Statistics and Computing, 14, 251–260.
Genz, A. and Bretz, F. (2009), Computation of Multivariate Normal and t Probabilities. Lecture
Notes in Statistics, Vol. 195. Springer-Verlag, Heidelberg.
Miwa, A., Hayter J. and Kuriki, S. (2003). The evaluation of general non-centred orthant probabil-
ities. Journal of the Royal Statistical Society, Ser. B, 65, 223–234.

See Also
qmvnorm

Examples

n <- 5
mean <- rep(0, 5)
lower <- rep(-1, 5)
upper <- rep(3, 5)
corr <- diag(5)
corr[lower.tri(corr)] <- 0.5
corr[upper.tri(corr)] <- 0.5
prob <- pmvnorm(lower, upper, mean, corr)
print(prob)

stopifnot(pmvnorm(lower=-Inf, upper=3, mean=0, sigma=1) == pnorm(3))

a <- pmvnorm(lower=-Inf,upper=c(.3,.5),mean=c(2,4),diag(2))

stopifnot(round(a,16) == round(prod(pnorm(c(.3,.5),c(2,4))),16))

a <- pmvnorm(lower=-Inf,upper=c(.3,.5,1),mean=c(2,4,1),diag(3))

stopifnot(round(a,16) == round(prod(pnorm(c(.3,.5,1),c(2,4,1))),16))

# Example from R News paper (original by Genz, 1992):

m <- 3
sigma <- diag(3)
sigma[2,1] <- 3/5
sigma[3,1] <- 1/3
sigma[3,2] <- 11/15
pmvt 9

pmvnorm(lower=rep(-Inf, m), upper=c(1,4,2), mean=rep(0, m), corr=sigma)

# Correlation and Covariance

a <- pmvnorm(lower=-Inf, upper=c(2,2), sigma = diag(2)*2)

b <- pmvnorm(lower=-Inf, upper=c(2,2)/sqrt(2), corr=diag(2))
stopifnot(all.equal(round(a,5) , round(b, 5)))

pmvt Multivariate t Distribution

Description
Computes the the distribution function of the multivariate t distribution for arbitrary limits, degrees
of freedom and correlation matrices based on algorithms by Genz and Bretz.

Usage
pmvt(lower=-Inf, upper=Inf, delta=rep(0, length(lower)),
df=1, corr=NULL, sigma=NULL, algorithm = GenzBretz(),
type = c("Kshirsagar", "shifted"), keepAttr=TRUE, ...)

Arguments
lower the vector of lower limits of length n.
upper the vector of upper limits of length n.
delta the vector of noncentrality parameters of length n, for type = "shifted" delta
specifies the mode.
df degree of freedom as integer. Normal probabilities are computed for df=0.
corr the correlation matrix of dimension n.
sigma the scale matrix of dimension n. Either corr or sigma can be specified. If
sigma is given, the problem is standardized. If neither corr nor sigma is given,
the identity matrix is used for sigma.
algorithm an object of class GenzBretz or TVPACK defining the hyper parameters of this
algorithm.
type type of the noncentral multivariate t distribution to be computed. type = "Kshirsagar"
corresponds to formula (1.4) in Genz and Bretz (2009) (see also Chapter 5.1 in
Kotz and Nadarajah (2004)). This is the noncentral t-distribution needed for cal-
culating the power of multiple contrast tests under a normality assumption. type
= "shifted" corresponds to the formula right before formula (1.4) in Genz and
Bretz (2009) (see also formula (1.1) in Kotz and Nadarajah (2004)). It is a lo-
cation shifted version of the central t-distribution. This noncentral multivariate
t distribution appears for example as the Bayesian posterior distribution for the
regression coefficients in a linear regression. In the central case both types co-
incide.
10 pmvt

keepAttr logical indicating if attributes such as error and msg should be attached to
the return value. The default, TRUE is back compatible.
... additional parameters (currently given to GenzBretz for backward compatibility
issues).

Details
This function involves the computation of central and noncentral multivariate t-probabilities with
arbitrary correlation matrices. It involves both the computation of singular and nonsingular prob-
abilities. The methodology (for default algorithm = GenzBretz()) is based on randomized quasi
Monte Carlo methods and described in Genz and Bretz (1999, 2002).
Because of the randomization, the result for this algorithm (slightly) depends on .Random.seed.
For 2- and 3-dimensional problems one can also use the TVPACK routines described by Genz (2004),
which only handles semi-infinite integration regions (and for type = "Kshirsagar" only central
problems).
For type = "Kshirsagar" and a given correlation matrix corr, for short A, say, (which has to be
positive semi-definite) and degrees of freedom ν the following values are numerically evaluated
Z ∞
√ √
I = 21−ν/2 /Γ(ν/2) sν−1 exp(−s2 /2)Φ(s · lower/ ν − δ, s · upper/ ν − δ) ds
0

where
Z b
m −1/2
Φ(a, b) = (det(A)(2π) ) exp(−x0 Ax/2) dx
a

is the multivariate normal distribution and m is the number of rows of A.

For type = "shifted", a positive definite symmetric matrix S (which might be the correlation or
the scale matrix), mode (vector) δ and degrees of freedom ν the following integral is evaluated:
Z upper1 Z upperm
c ... (1 + (x − δ)0 S −1 (x − δ)/ν)−(ν+m)/2 dx1 ...dxm ,
lower1 lowerm

where

c = Γ((ν + m)/2)/((πν)m/2 Γ(ν/2)|S|1/2 ),

and m is the number of rows of S.

Note that both -Inf and +Inf may be specified in the lower and upper integral limits in order to
compute one-sided probabilities.
Univariate problems are passed to pt. If df = 0, normal probabilities are returned.

Value
The evaluated distribution function is returned, if keepAttr is true, with attributes
error estimated absolute error and
msg status message (a character string).
algorithm a character string with class(algorithm).
pmvt 11

Source
http://www.sci.wsu.edu/math/faculty/genz/homepage

References
Genz, A. and Bretz, F. (1999), Numerical computation of multivariate t-probabilities with applica-
tion to power calculation of multiple contrasts. Journal of Statistical Computation and Simulation,
63, 361–378.
Genz, A. and Bretz, F. (2002), Methods for the computation of multivariate t-probabilities. Journal
of Computational and Graphical Statistics, 11, 950–971.
Genz, A. (2004), Numerical computation of rectangular bivariate and trivariate normal and t-probabilities,
Statistics and Computing, 14, 251–260.
Genz, A. and Bretz, F. (2009), Computation of Multivariate Normal and t Probabilities. Lecture
Notes in Statistics, Vol. 195. Springer-Verlag, Heidelberg.
S. Kotz and S. Nadarajah (2004), Multivariate t Distributions and Their Applications. Cambridge
University Press. Cambridge.
Edwards D. and Berry, Jack J. (1987), The efficiency of simulation-based multiple comparisons.
Biometrics, 43, 913–928.

See Also
qmvt

Examples

n <- 5
lower <- -1
upper <- 3
df <- 4
corr <- diag(5)
corr[lower.tri(corr)] <- 0.5
delta <- rep(0, 5)
prob <- pmvt(lower=lower, upper=upper, delta=delta, df=df, corr=corr)
print(prob)

pmvt(lower=-Inf, upper=3, df = 3, sigma = 1) == pt(3, 3)

# Example from R News paper (original by Edwards and Berry, 1987)

n <- c(26, 24, 20, 33, 32)

V <- diag(1/n)
df <- 130
C <- c(1,1,1,0,0,-1,0,0,1,0,0,-1,0,0,1,0,0,0,-1,-1,0,0,-1,0,0)
C <- matrix(C, ncol=5)
### scale matrix
cv <- C %*% V %*% t(C)
### correlation matrix
dv <- t(1/sqrt(diag(cv)))
12 qmvnorm

cr <- cv * (t(dv) %*% dv)

delta <- rep(0,5)

myfct <- function(q, alpha) {

lower <- rep(-q, ncol(cv))
upper <- rep(q, ncol(cv))
pmvt(lower=lower, upper=upper, delta=delta, df=df,
corr=cr, abseps=0.0001) - alpha
}

### uniroot for this simple problem

round(uniroot(myfct, lower=1, upper=5, alpha=0.95)$root, 3)

# compare pmvt and pmvnorm for large df:

a <- pmvnorm(lower=-Inf, upper=1, mean=rep(0, 5), corr=diag(5))

b <- pmvt(lower=-Inf, upper=1, delta=rep(0, 5), df=300,
corr=diag(5))
a
b

stopifnot(round(a, 2) == round(b, 2))

# correlation and scale matrix

a <- pmvt(lower=-Inf, upper=2, delta=rep(0,5), df=3,

sigma = diag(5)*2)
b <- pmvt(lower=-Inf, upper=2/sqrt(2), delta=rep(0,5),
df=3, corr=diag(5))
attributes(a) <- NULL
attributes(b) <- NULL
a
b
stopifnot(all.equal(round(a,3) , round(b, 3)))

a <- pmvt(0, 1,df=10)

attributes(a) <- NULL
b <- pt(1, df=10) - pt(0, df=10)
stopifnot(all.equal(round(a,10) , round(b, 10)))

qmvnorm Quantiles of the Multivariate Normal Distribution

Description
Computes the equicoordinate quantile function of the multivariate normal distribution for arbitrary
correlation matrices based on inversion of pmvnorm, using a stochastic root finding algorithm de-
scribed in Bornkamp (2018).
qmvnorm 13

Usage
qmvnorm(p, interval = NULL, tail = c("lower.tail",
"upper.tail", "both.tails"), mean = 0, corr = NULL,
sigma = NULL, algorithm = GenzBretz(),
ptol = 0.001, maxiter = 500, trace = FALSE, ...)

Arguments
p probability.
interval optional, a vector containing the end-points of the interval to be searched. Does
not need to contain the true quantile, just used as starting values by the root-
finder. If equal to NULL a guess is used.
tail specifies which quantiles should be computed. lower.tail gives the quantile
x for which P [X ≤ x] = p, upper.tail gives x with P [X > x] = p and
both.tails leads to x with P [−x ≤ X ≤ x] = p.
mean the mean vector of length n.
corr the correlation matrix of dimension n.
sigma the covariance matrix of dimension n. Either corr or sigma can be specified. If
sigma is given, the problem is standardized. If neither corr nor sigma is given,
the identity matrix is used for sigma.
algorithm an object of class GenzBretz, Miwa or TVPACK specifying both the algorithm to
be used as well as the associated hyper parameters.
ptol, maxiter, trace
Parameters passed to the stochastic root-finding algorithm. Iteration stops when
the 95% confidence interval for the predicted quantile is inside [p-ptol, p+ptol].
maxiter is the maximum number of iterations for the root finding algorithm.
trace prints the iterations of the root finder.
... additional parameters to be passed to GenzBretz.

Details
Only equicoordinate quantiles are computed, i.e., the quantiles in each dimension coincide. The
result is seed dependend.

Value
A list with two components: quantile and f.quantile give the location of the quantile and the
difference between the distribution function evaluated at the quantile and p.

References
Bornkamp, B. (2018). Calculating quantiles of noisy distribution functions using local linear re-
gressions. Computational Statistics, 33, 487–501.

See Also
pmvnorm, qmvt
14 qmvt

Examples
qmvnorm(0.95, sigma = diag(2), tail = "both")

qmvt Quantiles of the Multivariate t Distribution

Description
Computes the equicoordinate quantile function of the multivariate t distribution for arbitrary corre-
lation matrices based on inversion of pmvt, using a stochastic root finding algorithm described in
Bornkamp (2018).

Usage
qmvt(p, interval = NULL, tail = c("lower.tail",
"upper.tail", "both.tails"), df = 1, delta = 0, corr = NULL,
sigma = NULL, algorithm = GenzBretz(),
type = c("Kshirsagar", "shifted"),
ptol = 0.001, maxiter = 500, trace = FALSE, ...)

Arguments
p probability.
interval optional, a vector containing the end-points of the interval to be searched. Does
not need to contain the true quantile, just used as starting values by the root-
finder. If equal to NULL a guess is used.
tail specifies which quantiles should be computed. lower.tail gives the quantile
x for which P [X ≤ x] = p, upper.tail gives x with P [X > x] = p and
both.tails leads to x with P [−x ≤ X ≤ x] = p.
delta the vector of noncentrality parameters of length n, for type = "shifted" delta
specifies the mode.
df degree of freedom as integer. Normal quantiles are computed for df = 0 or df =
Inf.
corr the correlation matrix of dimension n.
sigma the covariance matrix of dimension n. Either corr or sigma can be specified. If
sigma is given, the problem is standardized. If neither corr nor sigma is given,
the identity matrix in the univariate case (so corr = 1) is used for corr.
algorithm an object of class GenzBretz or TVPACK defining the hyper parameters of this
algorithm.
type type of the noncentral multivariate t distribution to be computed. type = "Kshirsagar"
corresponds to formula (1.4) in Genz and Bretz (2009) (see also Chapter 5.1 in
Kotz and Nadarajah (2004)) and type = "shifted" corresponds to the formula
before formula (1.4) in Genz and Bretz (2009) (see also formula (1.1) in Kotz
and Nadarajah (2004)).
qmvt 15

ptol, maxiter, trace

Parameters passed to the stochastic root-finding algorithm. Iteration stops when
the 95% confidence interval for the predicted quantile is inside [p-ptol, p+ptol].
maxiter is the maximum number of iterations for the root finding algorithm.
trace prints the iterations of the root finder.
... additional parameters to be passed to GenzBretz.

Details
Only equicoordinate quantiles are computed, i.e., the quantiles in each dimension coincide. The
result is seed dependend.

Value
A list with two components: quantile and f.quantile give the location of the quantile and the
difference between the distribution function evaluated at the quantile and p.

References
Bornkamp, B. (2018). Calculating quantiles of noisy distribution functions using local linear re-
gressions. Computational Statistics, 33, 487–501.

See Also
pmvnorm, qmvnorm

Examples
## basic evaluation
qmvt(0.95, df = 16, tail = "both")

## check behavior for df=0 and df=Inf

Sigma <- diag(2)
set.seed(29)
q0 <- qmvt(0.95, sigma = Sigma, df = 0, tail = "both")$quantile
set.seed(29)
q8 <- qmvt(0.95, sigma = Sigma, df = Inf, tail = "both")$quantile
set.seed(29)
qn <- qmvnorm(0.95, sigma = Sigma, tail = "both")$quantile
stopifnot(identical(q0, q8),
isTRUE(all.equal(q0, qn, tol = (.Machine$double.eps)^(1/3))))

## if neither sigma nor corr are provided, corr = 1 is used internally

df <- 0
set.seed(29)
qt95 <- qmvt(0.95, df = df, tail = "both")$quantile
set.seed(29)
qt95.c <- qmvt(0.95, df = df, corr = 1, tail = "both")$quantile
set.seed(29)
qt95.s <- qmvt(0.95, df = df, sigma = 1, tail = "both")$quantile
stopifnot(identical(qt95, qt95.c),
16 qmvt

identical(qt95, qt95.s))

df <- 4
set.seed(29)
qt95 <- qmvt(0.95, df = df, tail = "both")$quantile
set.seed(29)
qt95.c <- qmvt(0.95, df = df, corr = 1, tail = "both")$quantile
set.seed(29)
qt95.s <- qmvt(0.95, df = df, sigma = 1, tail = "both")$quantile
stopifnot(identical(qt95, qt95.c),
identical(qt95, qt95.s))
Index

∗ distribution rmvnorm (Mvnorm), 3

algorithms, 2 rmvt (Mvt), 4
Mvnorm, 3 rnorm, 4
Mvt, 4
pmvnorm, 6 TVPACK, 7, 9, 10, 13, 14
pmvt, 9 TVPACK (algorithms), 2
qmvnorm, 12
qmvt, 14
∗ multivariate
Mvnorm, 3
Mvt, 4
.Random.seed, 7, 10

algorithms, 2
attributes, 7, 10

character, 7, 10
cov2cor, 5

dmvnorm, 5, 7
dmvnorm (Mvnorm), 3
dmvt (Mvt), 4

GenzBretz, 7, 9, 13–15
GenzBretz (algorithms), 2

logical, 5, 7, 10

Miwa, 7, 13
Miwa (algorithms), 2
Mvnorm, 3
Mvt, 4

pmvnorm, 4, 6, 12, 13, 15

pmvt, 5–7, 9, 14
pnorm, 7
pt, 10

qmvnorm, 4, 8, 12, 15
qmvt, 6, 11, 13, 14

rmvnorm, 5, 7

17