RTMB (version 1.6)

ADvector: The AD vector and its methods

Description

An advector is a class used behind the scenes to replace normal R numeric objects during automatic differentiation. An advector has a 'temporary lifetime' and therefore you do not see / need to know it as a normal user.

Usage

advector(x)
# S3 method for advector
Ops(e1, e2)
# S3 method for advector
Math(x, ...)
# S3 method for advector
as.vector(x, mode = "any")
# S3 method for advector
as.complex(x, ...)
# S3 method for advector
aperm(a, perm, ...)
# S3 method for advector
c(...)
# S3 method for advector
[(x, ...)
# S3 method for advector
[(x, ...) <- value
# S3 method for advector
[[(x, ...)
# S3 method for advector
rep(x, ...)
# S3 method for advector
is.nan(x)
# S3 method for advector
is.finite(x)
# S3 method for advector
is.infinite(x)
# S3 method for advector
is.na(x)
# S3 method for advector
sum(x, ..., na.rm = FALSE)
# S3 method for advector
mean(x, ...)
# S3 method for advector
prod(x, ..., na.rm)
# S3 method for advector
is.numeric(x)
# S3 method for advector
as.double(x, ...)
# S3 method for advector
Complex(z)
# S3 method for advector
Summary(..., na.rm = FALSE)
# S3 method for advector
diff(x, lag = 1L, differences = 1L, ...)
# S3 method for advector
print(x, ...)
# S4 method for num,ad,ad
ifelse(test, yes, no)
# S4 method for num,num,num
ifelse(test, yes, no)

Value

Object of class "advector".

Arguments

x

numeric or advector

e1

advector

e2

advector

...

Additional arguments

mode

FIXME might not be handled correctly by as.vector

a

advector with dimension attribute

perm

Permutation as in aperm

value

Replacement value implicitly converted to AD

na.rm

Must be FALSE (default)

z

Complex (not allowed)

lag

As diff

differences

As diff

test

logical vector

yes

advector

no

advector

Functions

• advector(): Construct a new advector

• Ops(advector): Binary operations

• Math(advector): Unary operations

• as.vector(advector): Makes array(x) work.

• as.complex(advector): Convert to ADcomplex. Note that dimensions are dropped for consistency with base R.

• aperm(advector): Equivalent of aperm

• c(advector): Equivalent of c. However note the limitation for mixed types: If x is an AD type, c(x,1) works while c(1,x) does not!

• [: Equivalent of [

• `[`(advector) <- value: Equivalent of [<-

• [[: Equivalent of [[

• rep(advector): Equivalent of rep. Makes outer(x,x,...) work.

• is.nan(advector): Equivalent of is.nan. Check NaN status of a constant advector expression. If not constant throw an error.

• is.finite(advector): Equivalent of is.finite. Check finite status of a constant advector expression. If not constant throw an error.

• is.infinite(advector): Equivalent of is.infinite. Check infinity status of a constant advector expression. If not constant throw an error.

• is.na(advector): Equivalent of is.na. Check NA status of an advector. NAs can only occur directly (as constants) or indirectly as the result of an operation with NA operands. For a tape built with non-NA parameters the NA status of any expression is constant and can therefore safely be used as part of the calculations. (assuming correct propagation of NAs via C-level arithmetic).

• sum(advector): Equivalent of sum. na.rm=TRUE is allowed, but note that this feature assumes correct propagation of NAs via C-level arithmetic.

• mean(advector): Equivalent of mean except no arguments beyond x are supported.

• prod(advector): Equivalent of prod except na.rm not allowed.

• is.numeric(advector): Makes cov2cor() work. FIXME: Any unwanted side-effects with this?

• as.double(advector): Makes as.numeric() work.

• Complex(advector): Complex operations are redirected to adcomplex.

• Summary(advector): Non differentiable Summary operations (e.g. min max) are not allowed and will throw an error.

• diff(advector): Equivalent of diff

• print(advector): Print method

• ifelse(test = num, yes = ad, no = ad): Equivalent of ifelse

• ifelse(test = num, yes = num, no = num): Default method

Details

An AD vector (class='advector') is an atomic R vector of 'codes' that are internally interpretable as 'AD scalars'. A substantial part of R's existing S3 matrix and array functionality can be re-used for AD vectors.

Examples

x <- advector(1:9)
a <- array(x, c(3,3))  ## as an array
outer(x, x, "+") ## Implicit via 'rep'
rev(x)           ## Implicit via '['