An S4 class to enable forward differentiation of multivariate computations.
Think of ‘madness’ as ‘multivariate automatic differentiation -ness.’
There is also a constructor method for madness
objects, and a
wrapper method.
# S4 method for madness
initialize(
.Object,
val,
dvdx,
xtag = NA_character_,
vtag = NA_character_,
varx = matrix(nrow = 0, ncol = 0)
)madness(val, dvdx = NULL, vtag = NULL, xtag = NULL, varx = NULL)
An object of class madness
.
a madness
object, or proto-object.
an array
of some numeric value, of arbitrary
dimension.
a matrix
of the derivative of
(the vector of) val
with respect to some independent
variable, \(X\).
an optional name for the \(X\) variable.
an optional name for the \(val\) variable.
an optional variance-covariance matrix of the independent variable, \(X\).
val
an array
of some numeric value. (Note that
array
includes matrix
as a subclass.) The numeric
value can have arbitrary dimension.
dvdx
a matrix
of the derivative of
(the vector of) val
with respect to some independent
variable, \(X\).
A Derivative is indeed a 2-dimensional matrix.
Derivatives have all been 'flattened'. See the details.
If not given, defaults
to the identity matrix, in which case \(val=X\),
which is useful to initialization. Note that the derivative
is with respect to an 'unrestricted' \(X\).
xtag
an optional name for the \(X\) variable.
Operations between two objects of the class with distinct
xtag
data will result in an error, since they are
considered to have different independent variables.
vtag
an optional name for the \(val\) variable. This will be propagated forward.
varx
an optional variance-covariance matrix of the independent variable, \(X\).
Steven E. Pav shabbychef@gmail.com
A madness
object contains a (multidimensional)
value, and the derivative of that with respect to some independent
variable. The purpose is to simplify computation of multivariate
derivatives, especially for use in the Delta method. Towards this
usage, one may store the covariance of the independent variable
in the object as well, from which the approximate variance-covariance
matrix can easily be computed. See vcov
.
Note that derivatives are all implicitly 'flattened'. That is, when we talk of the derivative of \(i \times j\) matrix \(Y\) with respect to \(m \times n\) matrix \(X\), we mean the derivative of the \(ij\) vector \(\mathrm{vec}\left(Y\right)\) with respect to the \(mn\) vector \(\mathrm{vec}\left(X\right)\). Moreover, derivatives follow the 'numerator layout' convention: this derivative is a \(ij \times mn\) matrix whose first column is the derivative of \(\mathrm{vec}\left(Y\right)\) with respect to \(X_{1,1}\). Numerator layout feels unnatural because it makes a gradient vector of a scalar-valued function into a row vector. Despite this deficiency, it makes the product rule feel more natural. (2FIX: is this so?)
Petersen, Kaare Brandt and Pedersen, Michael Syskind. "The Matrix Cookbook." Technical University of Denmark (2012). http://www2.imm.dtu.dk/pubdb/pubs/3274-full.html
Magnus, Jan R. and Neudecker, H. "Matrix Differential Calculus with Applications in Statistics and Econometrics." 3rd Edition. Wiley Series in Probability and Statistics: Texts and References Section (2007).
obj <- new("madness", val=matrix(rnorm(10*10),nrow=10), dvdx=diag(100), xtag="foo", vtag="foo")
obj2 <- madness(val=matrix(rnorm(10*10),nrow=10), xtag="foo", vtag="foo^2")
Run the code above in your browser using DataLab