nlsr and Other Approaches to Derivatives in R

This long and incomplete vignette is an attempt to catalog and illustrate the various capabilities in the R statistical computing system to perform differentiation. There are many traps and pitfalls for the unwary in doing this. It is hoped that by collecting examples in this long treatment will serve to record these and show how to avoid them, and how to reliably compute the derivatives desired. Derivative capabilities of R are in the base system (essentially the functions D() and deriv()) and in different packages, namely nlsr (or more recently nlsr), Deriv, and Ryacas. General tools for approximations to derivatives are found in the package numDeriv as well as optimx. Other approximations may be embedded in various packages, but not necessarily exported for use in scripts or packages.

As a way of recording where attention is needed either to this document or to the functions and methods described, I have put double question marks in various places. Contributions and collaborations to extend and complete the treatment are welcome.

Note: To distinguish output results (which are prefaced '##' by knitr, I have attempted to put comments in the R code with the preface '#-'.)

Available analytic differentiation tools

R has a number of tools for finding analytic derivatives.

How the tools are used

This is an overview section to give an idea of the capabilities. It is not intended to be exhaustive, but to give pointers to how the tools can be used quickly.

An important issue that may cause a lot of difficulty is the iterating of the tools. That is, we compute a derivative, then want to apply a tool to the derivative to get a second derivative. In doing so, we need to be careful that the type (class??) of the quantity output by the tool is passed back into the tool in a form that will generate a derivative expression. Some examples are presented.

We also note that the Deriv package will give a result in cases when the input is undefined. This is clearly a bug. There is an example below on the section for Deriv.

stats (i.e., the base R installation)

D(), deriv() and deriv3(): As deriv3() is stated to be the same as deriv() but with argument hessian=TRUE, we will for now only consider the first two.

dx2x <- deriv(~ x^2, "x") 
x <- -1:2
eval(dx2x) # This is evaluated at -1, 0, 1, 2, with the result in 
# the "gradient" attribute.
# Note that we cannot (easily) differentiate this again.
firstd <- attr(dx2x,"gradient")
#  ... and the following gives an error
d2x2x <- try(deriv(firstd, "x"))
#- Build a function from the expression
#- # Now try D()
Dx2x <- D(expression(x^2), "x")
x <- -1:2
# We can differentiate aggain
D2x2x <- D(Dx2x,"x")
eval(D2x2x) #- But we don't get a vector -- could be an issue in gradients/Jacobians
#- Note how we handle an expression stored in a string via parse(text=  ))
sx2 <- "x^2"
sDx2x <- D(parse(text=sx2), "x")
#- But watch out! The following "seems" to work, but the answer is not as intended.  
#- The problem is that the first argument is evaluated before being used.  Since 
#- x exists, it fails
Dx2xx <- D(x^2, "x")
#-  Something 'tougher':
trig.exp <- expression(sin(cos(x + y^2)))
( <- D(trig.exp, "x") )
all.equal(D(trig.exp[[1]], "x"),
( dxy <- deriv(trig.exp, c("x", "y")) )
y <- 1
#-  function returned:
deriv((y ~ sin(cos(x) * y)), c("x","y"), func = TRUE)
#- ??#-  Surely there is an error, since documentation says no lhs! i.e.,
#- "expr: a 'expression' or 'call' or (except 'D') a formula with no lhs."
#-  function with defaulted arguments:
(fx <- deriv(y ~ b0 + b1 * 2^(-x/th), c("b0", "b1", "th"),
             function(b0, b1, th, x = 1:7){} ) )
fx(2, 3, 4)
#-  First derivative
D(expression(x^2), "x")
#-  stopifnot(D("x"), "x") == 1) #- A way of testing. 
#- This works by coercing "x" to name/symbol, and derivative should be 1.
#- Would fail only if "x" cannot be so coerced. How could this happen??
#-  Higher derivatives showing deriv3
myd3 <- deriv3(y ~ b0 + b1 * 2^(-x/th), c("b0", "b1", "th"),
     c("b0", "b1", "th", "x") )
myd3(2,3,4, x=1:7)
#- check against deriv()
myd3a <- deriv(y ~ b0 + b1 * 2^(-x/th), c("b0", "b1", "th"),
     c("b0", "b1", "th", "x"), hessian=TRUE )
myd3a(2,3,4, x=1:7)
identical(myd3a, myd3) #- Remember to check things!

#-  Higher derivatives:
DD <- function(expr, name, order = 1) {
   if(order < 1) stop("'order' must be >= 1")
   if(order == 1) D(expr, name)
   else DD(D(expr, name), name, order - 1)
DD(expression(sin(x^2)), "x", 3)
#-  showing the limits of the internal "simplify()" :
#-  -sin(x^2) * (2 * x) * 2 + ((cos(x^2) * (2 * x) * (2 * x) + sin(x^2) *
#-     2) * (2 * x) + sin(x^2) * (2 * x) * 2)


dx2xn <- nlsDeriv(~ x^2, "x")
x <- -1:2
eval(dx2xn) # This is evaluated at -1, 0, 1, 2, BUT result is returned directly,
#-  NOT in "gradient" attribute
firstdn <- dx2xn
d2x2xn <- nlsDeriv(firstdn, "x")
d2x2xnF <- nlsDeriv(firstdn, "x", do_substitute=FALSE)
d2x2xnF # in this case we get the same result
d2x2xnT <- nlsDeriv(firstdn, "x", do_substitute=TRUE)
d2x2xnT # 0 ## WATCH OUT

#- ?? We can iterate the derivatives
nlsDeriv(d2x2xn, "x")

nlsDeriv(x^2, "x")# 0
nlsDeriv(x^2, "x", do_substitute=FALSE)# 0
nlsDeriv(x^2, "x", do_substitute=TRUE) # 2 * x
nlsDeriv(~ x^2, "x") # 2 * x
nlsDeriv(~ x^2, "x", do_substitute=FALSE) # 2 * x
nlsDeriv(~ x^2, "x", do_substitute=TRUE) # 2 * x

#- Some things to check and explain??
#?? firstde <- quote(firstd)
#?? firstde
#?? firstde <- bquote(firstd)
#?? firstde
#?? nlsDeriv(firstde, "x")
d2 <- nlsDeriv(2 * x, "x")
#?? firstc <-
#?? nlsDeriv(firstc, "x")
#- Build a function from the expression
#?? fdx2xn<-function(x){eval(dx2xn)}
#?? fdx2xn(1)
#?? fdx2xn(3.21)
#?? fdx2xn(1:5)

The tool codeDeriv returns an R expression to evaluate the derivative efficiently. fnDeriv wraps it in a function. By default the arguments to the function are constructed from all variables in the expression. In the example below this includes x.

codeDeriv(parse(text="b0 + b1 * 2^(-x/th)"), c("b0", "b1", "th")) 
#- Include parameters as arguments
fj.1 <- fnDeriv(parse(text="b0 + b1 * 2^(-x/th)"), c("b0", "b1", "th")) 
#- Get all parameters from the calling environment
fj.2 <- fnDeriv(parse(text="b0 + b1 * 2^(-x/th)"), c("b0", "b1", "th"),
        args = character())
b0 <- 1
b1 <- 2
x <- 3
th <- 4 

#- Just use an expression
fje <- codeDeriv(parse(text="b0 + b1 * 2^(-x/th)"), c("b0", "b1", "th"))

dx2xnf <- fnDeriv(~ x^2, "x") #- Use tilde
dx2xnf <- fnDeriv(expression(x^2), "x") #- or use expression()
x <- -1:2
#?? eval(dx2xnf) # This is evaluated at -1, 0, 1, 2, BUT result is returned directly,
#-  NOT in "gradient" attribute
# Note that we cannot (easily) differentiate this again.
# firstd <- dx2xnf
# str(firstd)
# d2x2xnf <- try(nlsDeriv(firstd, "x")) #- this APPEARS to work, but WRONG answer 
# str(d2x2xnf)
# d2x2xnf
# eval(d2x2xnf)

# dx2xnfh <- fnDeriv(expression(x^2), "x", hessian=TRUE) #- Try for second derivatives
# dx2xnfh
# mode(dx2xnfh)
# str(dx2xnfh)
# x <- -1
# eval(dx2xnfh) # This is evaluated at -1, 0, 1, 2, BUT result is returned directly,


The following examples are drawn from the example(Deriv) contained in the Deriv package.


f <- function(x) x^2
#- Should see
#- function (x)
#- 2 * x
#- Now save the derivative
f1 <- Deriv(f)
f1 #- print it
f2 <- Deriv(f1) #- and take second derivative
f2 #- print it

f <- function(x, y) sin(x) * cos(y)
f_ <- Deriv(f)
f_ #- print it
#- Should see
#- function (x, y)
#- c(x = cos(x) * cos(y), y = -(sin(x) * sin(y)))
f_(3, 4)
#- Should see
#-              x         y
#- [1,] 0.6471023 0.1068000

f2 <- Deriv(~ f(x, y^2), "y") 
#- This has a tilde to render the 1st argument as a formula object
#- Also we are substituting in y^2 for y
f2 #- print it
#- -(2 * (y * sin(x) * sin(y^2)))
mode(f2) #- check what type of object it is
arg1 <- ~ f(x,y^2)
mode(arg1) #- check the type
f2a <- Deriv(arg1, "y")
f2a #- and print to see if same as before
#- try evaluation of f using current x and y
eval(f2a) #- We need x and y defined to do this.

f3 <- Deriv(quote(f(x, y^2)), c("x", "y"), cache.exp=FALSE) #- check cache.exp operation
#- Note that we need to quote or will get evaluation at current x, y values (if they exist)
f3 #- print it
#- c(x = cos(x) * cos(y^2), y = -(2 * (y * sin(x) * sin(y^2))))
f3c <- Deriv(quote(f(x, y^2)), c("x", "y"), cache.exp=TRUE) #- check cache.exp operation
f3c #- print it
#- Now want to evaluate the results
#- First must provide some data
x <- 3
y <- 4
#- Should see
#- x         y 
#- 0.9480757 0.3250313 
eval(f3) #- check this also
#- or we can create functions
f3cf <- function(x, y){eval(f3c)}
f3cf(x=1, y=2)
#-          x          y 
#- -0.3531652  2.5473094 
f3f <- function(x,y){eval(f3)}
f3f(x=3, y=4)
#-         x         y 
#- 0.9480757 0.3250313 

#- try an expression
Deriv(expression(sin(x^2) * y), "x")
#- should see
#- expression(2 * (x * y * cos(x^2)))

#- quoted string
Deriv("sin(x^2) * y", "x") # differentiate only by x
#- Should see
#- "2 * (x * y * cos(x^2))"

Deriv("sin(x^2) * y", cache.exp=FALSE) #- differentiate by all variables (here by x and y)
#- Note that default is to differentiate by all variables.
#- Should see
#- "c(x = 2 * (x * y * cos(x^2)), y = sin(x^2))"

#- Compound function example (here abs(x) smoothed near 0)
#- Note that this introduces the possibilty of `if` statements in the code
#- BUT (JN) seems to give back quoted string, so we must parse.
fc <- function(x, h=0.1) if (abs(x) < h) 0.5*h*(x/h)**2 else abs(x)-0.5*h
efc1 <- Deriv("fc(x)", "x", cache.exp=FALSE) 
#- "if (abs(x) < h) x/h else sign(x)"
#- A few checks on the results

fc1 <- function(x,h=0.1){ eval(parse(text=efc1)) }
## h=0.1
fc1(0.001, 1)

#- Example of a first argument that cannot be evaluated in the current environment:
  try(suppressWarnings(rm("xx", "yy"))) #- Make sure there are no objects xx or yy
  Deriv(~ xx^2+yy^2)
#- Should show
#- c(xx = 2 * xx, yy = 2 * yy)
#- ?? What is the meaning / purpose of this construct?

#- ?? Is following really AD?  
#- Automatic differentiation (AD), note intermediate variable 'd' assignment
Deriv(~{d <- ((x-m)/s)^2; exp(-0.5*d)}, "x")
# Note that the result we see does NOT match what follows in the example(Deriv) (JN ??)
#   d <- ((x - m)/s)^2
#   .d_x <- 2 * ((x - m)/s^2)
#   -(0.5 * (.d_x * exp(-(0.5 * d))))
#- For some reason the intermediate variable d is NOT included.??

#- Custom derivative rule. Note that this needs explanations??
  myfun <- function(x, y=TRUE) NULL #- do something useful
  dmyfun <- function(x, y=TRUE) NULL #- myfun derivative by x.
  drule[["myfun"]] <- alist(x=dmyfun(x, y), y=NULL) #- y is just a logical
  Deriv(myfun(z^2, FALSE), "z")
  # 2 * (z * dmyfun(z^2, FALSE))

#- Differentiation by list components
  theta <- list(m=0.1, sd=2.) #- Why do we set values??
  x <- names(theta) #- and why these particular names??
  names(x)=rep("theta", length(theta))
  Deriv(~exp(-(x-theta$m)**2/(2*theta$sd)), x, cache.exp=FALSE)
#- Should show the following (but why??)
#- c(theta_m = exp(-((x - theta$m)^2/(2 * theta$sd))) *
#-  (x - theta$m)/theta$sd, theta_sd = 2 * (exp(-((x - theta$m)^2/
#-  (2 * theta$sd))) * (x - theta$m)^2/(2 * theta$sd)^2))
lderiv <-  Deriv(~exp(-(x-theta$m)**2/(2*theta$sd)), x, cache.exp=FALSE)
fld <- function(x){ eval(lderiv)} #- put this in a function
fld(2) #- and evaluate at a value

Deriv has some design choices that can get the user into trouble. The following example shows one such problem.

rm(x) # ensures x is undefined
Deriv(~ x, "x")  # returns [1] 1 -- clearly a bug!
Deriv(~ x^2, "x")   # returns 2 * x
x <- quote(x^2)
Deriv(x, "x") # returns 2 * x

By comparison, nlsr

rm(x) # in case it is defined
try(nlsDeriv(x, "x")  ) # fails, not a formula
try(nlsDeriv(as.expression("x"), "x")  ) # expression(NULL)
try(nlsDeriv(~x, "x")  ) # 1
try(nlsDeriv(x^2, "x"))   # fails
try(nlsDeriv(~x^2, "x")) # 2 * x
x <- quote(x^2)
try(nlsDeriv(x, "x")) # returns 2 * x


There is at least one other symbolic package for R. Here we look at Ryacas. The structures for using yacas tools do not seem at the time of writing (2016-10-21) to be suitable for working with nonlinear least squares or optimization facilities of R. Thus, for the moment, we will not pursue the derivatives available in Ryacas beyond the following example provided by Gabor Grothendieck.

dnlsr <- nlsr::nlsDeriv(~ sin(x+y), "x")
detach("package:nlsr", unload=TRUE)
detach("package:Deriv", unload=TRUE)

## New Ryacas mechanism as of 2019-8-29 from (Mikkel Meyer Andersen)

yac_str("D(x) Sin(x+y)") 

# or if an expression is needed:
ex <- yac_expr("D(x) Sin(x+y)")
expression(cos(x + y))
eval(ex, list(x = pi, y = pi/2))

## Previous syntax for Ryacas was
## x <- Sym("x")
## y <- Sym("y")
## dryacas <- deriv(sin(x+y), x)
## print(dryacas)
## class(dryacas)

detach("package:Ryacas", unload=TRUE)

Derivatives and simplifications -- base R

See specific notes either in comments or at the end of the section.

The help page for D lists the functions for which derivatives are known: "The internal code knows about the arithmetic operators +, -, *, / and ^, and the single-variable functions exp, log, sin, cos, tan, sinh, cosh, sqrt, pnorm, dnorm, asin, acos, atan, gamma, lgamma, digamma and trigamma, as well as psigamma for one or two arguments (but derivative only with respect to the first)."

Derivatives and simplifications -- package nlsr

This package supports the derivatives that D supports, as well as a few others, and users can add their own definitions. The current list is


Derivatives table

Here is a slightly expanded testing of the elements of the nlsr derivatives table.

## Try different ways to supply the log function
aDeriv <- nlsDeriv(~ log(x), "x")
aderiv <- try(deriv( ~ log(x), "x"))
aD <- D(expression(log(x)), "x")
cat("but \n")
try(D( "~ log(x)", "x")) # fails -- gives NA rather than expected answer due to quotes
try(D( ~ log(x), "x"))
interm <- ~ log(x)
interme <- as.expression(interm)
try(D(interme, "x"))
try(deriv(interme, "x"))
try(deriv(interm, "x"))

nlsDeriv(~ log(x, base=3), "x" ) # OK
try(D(expression(log(x, base=3)), "x" )) # fails - only single-argument calls supported
try(deriv(~ log(x, base=3), "x" )) # fails - only single-argument calls supported
try(deriv(expression(log(x, base=3)), "x" )) # fails - only single-argument calls supported
try(deriv3(expression(log(x, base=3)), "x" )) # fails - only single-argument calls supported
fnDeriv(quote(log(x, base=3)), "x" )

nlsDeriv(~ exp(x), "x")
D(expression(exp(x)), "x") # OK
deriv(~exp(x), "x") # OK, but much more complicated
fnDeriv(quote(exp(x)), "x")

nlsDeriv(~ sin(x), "x")
D(expression(sin(x)), "x")
deriv(~sin(x), "x")
fnDeriv(quote(sin(x)), "x")

nlsDeriv(~ cos(x), "x")
D(expression(cos(x)), "x")
deriv(~ cos(x), "x")
fnDeriv(quote(cos(x)), "x")

nlsDeriv(~ tan(x), "x")
D(expression(tan(x)), "x")
deriv(~ tan(x), "x")
fnDeriv(quote(tan(x)), "x")

nlsDeriv(~ sinh(x), "x")
D(expression(sinh(x)), "x")
deriv(~sinh(x), "x")
fnDeriv(quote(sinh(x)), "x")

nlsDeriv(~ cosh(x), "x")
D(expression(cosh(x)), "x")
deriv(~cosh(x), "x")
fnDeriv(quote(cosh(x)), "x")

nlsDeriv(~ sqrt(x), "x")
D(expression(sqrt(x)), "x")
deriv(~sqrt(x), "x")
fnDeriv(quote(sqrt(x)), "x")

nlsDeriv(~ pnorm(q), "q")
D(expression(pnorm(q)), "q")
deriv(~pnorm(q), "q")
fnDeriv(quote(pnorm(q)), "q")

nlsDeriv(~ dnorm(x, mean), "mean")
D(expression(dnorm(x, mean)), "mean")
deriv(~dnorm(x, mean), "mean")
fnDeriv(quote(dnorm(x, mean)), "mean")

nlsDeriv(~ asin(x), "x")
D(expression(asin(x)), "x")
deriv(~asin(x), "x")
fnDeriv(quote(asin(x)), "x")

nlsDeriv(~ acos(x), "x")
D(expression(acos(x)), "x")
deriv(~acos(x), "x")
fnDeriv(quote(acos(x)), "x")

nlsDeriv(~ atan(x), "x")
D(expression(atan(x)), "x")
deriv(~atan(x), "x")
fnDeriv(quote(atan(x)), "x")

nlsDeriv(~ gamma(x), "x")
D(expression(gamma(x)), "x")
deriv(~gamma(x), "x")
fnDeriv(quote(gamma(x)), "x")

nlsDeriv(~ lgamma(x), "x")
D(expression(lgamma(x)), "x")
deriv(~lgamma(x), "x")
fnDeriv(quote(lgamma(x)), "x")

nlsDeriv(~ digamma(x), "x")
D(expression(digamma(x)), "x")
deriv(~digamma(x), "x")
fnDeriv(quote(digamma(x)), "x")

nlsDeriv(~ trigamma(x), "x")
D(expression(trigamma(x)), "x")
deriv(~trigamma(x), "x")
fnDeriv(quote(trigamma(x)), "x")

nlsDeriv(~ psigamma(x, deriv = 5), "x")
D(expression(psigamma(x, deriv = 5)), "x")
deriv(~psigamma(x, deriv = 5), "x")
fnDeriv(quote(psigamma(x, deriv = 5)), "x")

nlsDeriv(~ x*y, "x")
D(expression(x*y), "x")
deriv(~x*y, "x")
fnDeriv(quote(x*y), "x")

nlsDeriv(~ x/y, "x")
D(expression(x/y), "x")
deriv(~x/y, "x")
fnDeriv(quote(x/y), "x")

nlsDeriv(~ x^y, "x")
D(expression(x^y), "x")
deriv(~x^y, "x")
fnDeriv(quote(x^y), "x")

nlsDeriv(~ (x), "x")
D(expression((x)), "x")
deriv(~(x), "x")
fnDeriv(quote((x)), "x")

nlsDeriv(~ +x, "x")
D(expression(+x), "x")
deriv(~ +x, "x")
fnDeriv(quote(+x), "x")

nlsDeriv(~ -x, "x")
D(expression(- x), "x")
deriv(~ -x, "x")
fnDeriv(quote(-x), "x")

nlsDeriv(~ abs(x), "x")
try(D(expression(abs(x)), "x")) # 'abs' not in derivatives table
try(deriv(~ abs(x), "x"))
fnDeriv(quote(abs(x)), "x")

nlsDeriv(~ sign(x), "x")
try(D(expression(sign(x)), "x")) # 'sign' not in derivatives table
try(deriv(~ sign(x), "x"))
fnDeriv(quote(sign(x)), "x")


Simplifying algebraic expressions

nlsr also includes some tools for simplification of algebraic expressions, extensible by the user. Currently these involve the following functions:

#- Remove ##? to see reproducible error
#- ?? For some reason, if we leave packages attached, we get errors.
#- Here we detach all the non-base packages and then reload nlsr
##? require(nlsr)
##? sessionInfo()
##? ##? nlsSimplify(quote(+(a+b)))
##? nlsSimplify(quote(-5))
if ("Deriv" %in% loadedNamespaces()){detach("package:Deriv", unload=TRUE)} 
if ("nlsr" %in% loadedNamespaces() ){detach("package:nlsr", unload=TRUE)}
if ("Ryacas" %in% loadedNamespaces() ){detach("package:Ryacas", unload=TRUE)}
#- require(Deriv)
#- require(stats)
#- Various simplifications
#- ?? Do we need quote() to stop attempt to evaluate before applying simplification






nlsSimplify(quote(a + b + 0))
nlsSimplify(quote(0 + a + b))
nlsSimplify(quote((a+b) + (a+b)))
nlsSimplify(quote(1 + 4))

nlsSimplify(quote(a + b - 0))
nlsSimplify(quote(0 - a - b))
nlsSimplify(quote((a+b) - (a+b)))
nlsSimplify(quote(5 - 3))

nlsSimplify(quote(1L * (a+b)))
nlsSimplify(quote((a+b) * 1))

nlsSimplify(quote((a+b) / 1))
nlsSimplify(quote((a+b) / (-1)))

nlsSimplify(quote((a+b) ^ 1))

nlsSimplify(quote(log(exp(a), 3)))

nlsSimplify(quote(FALSE && b))
nlsSimplify(quote(a && TRUE))
nlsSimplify(quote(TRUE && b))

nlsSimplify(quote(a || TRUE))
nlsSimplify(quote(FALSE || b))
nlsSimplify(quote(a || FALSE))

nlsSimplify(quote(if (TRUE) a+b))
nlsSimplify(quote(if (FALSE) a+b))

nlsSimplify(quote(if (TRUE) a+b else a*b))
nlsSimplify(quote(if (FALSE) a+b else a*b))
nlsSimplify(quote(if (cond) a+b else a+b))


Derivatives and simplifications -- package Deriv

Derivatives table

?? To be added


#- For some reason, if we leave packages attached, we get errors.
#- Here we detach all the non-base packages and then reload nlsr
if ("Deriv" %in% loadedNamespaces()){detach("package:Deriv", unload=TRUE)} 
if ("Deriv" %in% loadedNamespaces() ){detach("package:nlsr", unload=TRUE)}
if ("Deriv" %in% loadedNamespaces() ){detach("package:Ryacas", unload=TRUE)}
#- Various simplifications
#- ?? Do we need quote() to stop attempt to evaluate before applying simplification?






Simplify(quote(a + b + 0))
Simplify(quote(0 + a + b))
Simplify(quote((a+b) + (a+b)))
Simplify(quote(1 + 4))

Simplify(quote(a + b - 0))
Simplify(quote(0 - a - b))
Simplify(quote((a+b) - (a+b)))
Simplify(quote(5 - 3))

Simplify(quote(1L * (a+b)))
Simplify(quote((a+b) * 1))

Simplify(quote((a+b) / 1))
Simplify(quote((a+b) / (-1)))

Simplify(quote((a+b) ^ 1))

Simplify(quote(log(exp(a), 3)))

Simplify(quote(FALSE && b))
Simplify(quote(a && TRUE))
Simplify(quote(TRUE && b))

Simplify(quote(a || TRUE))
Simplify(quote(FALSE || b))
Simplify(quote(a || FALSE))

Simplify(quote(if (TRUE) a+b))
Simplify(quote(if (FALSE) a+b))

Simplify(quote(if (TRUE) a+b else a*b))
Simplify(quote(if (FALSE) a+b else a*b))
Simplify(quote(if (cond) a+b else a+b))

#- This one is wrong... the double minus is an error, yet it works ??.
#- By comparison

Comparison with other approaches

?? to be added

check modelexpr() works with an ssgrfun ??

?? to be added

test model2rjfun vs model2rjfunx ??

?? to be added

Need more extensive discussion of Simplify??

?? to be added

Issues of programming on the language

?? need to explain where Deriv package comes from

One of the key tasks with tools for derivatives is that of taking objects in one or other form (that is, R class) and using it as an input for a symbolic function. The object may, of course, be an output from another such function, and this is one of the reasons we need to do such transformations.

We also note that the different tools for symbolic derivatives use slightly different inputs. For example, for the derivative of log(x), we have

dlogx <- nlsr::nlsDeriv(~ log(x), "x")

Unfortunately, there are complications when we have an expression object, and we need to specify that we do NOT execute the substitute() function. Here we show how to do this implicitly and with an explicit object.

dlogxs <- nlsr::nlsDeriv(expression(log(x)), "x", do_substitute=FALSE)
cat(as.character(dlogxs), "\n")
fne <- expression(log(x))
dlogxe <- nlsr::nlsDeriv(fne, "x", do_substitute=FALSE)

# base R
dblogx <- D(expression(log(x)), "x")

ddlogx <- Deriv::Deriv(expression(log(x)), "x")
cat(as.character(ddlogx), "\n")
ddlogxf <- ~ ddlogx

?? do each example by all methods and by numDeriv and put in dataframe for later presentation in a table.

Do we want examples in columns or rows. Probably 1 fn per row and work out a name for the row that is reasonably meaningful. Probably want an index column as well that is a list of strings. Can we then act on those strings to automate the whole setup?

# require(stats)
# require(Deriv)
# require(Ryacas)

# Various derivatives 

new <- codeDeriv(quote(1 + x + y), c("x", "y"))
old <- deriv(quote(1 + x + y), c("x", "y"))
# Following generates a very long line on output of knitr (for markdown)
newf <- function(x, y){

oldf <- function(x,y){

Unfortunately, the inputs and outputs are not always easily transformed so that the symbolic derivatives can be found. (?? Need to codify this and provide filters so we can get things to work nicely.)

As an example, how could we take object new and embed it in a function we can then use in R? We can certainly copy and paste the output into a function template, as follows,

fnfromnew <- function(x,y){
    .value <- 1 + x + y
    .grad <- array(0, c(length(.value), 2L), list(NULL, c("x", 
    .grad[, "x"] <- 1
    .grad[, "y"] <- 1
    attr(.value, "gradient") <- .grad


However, we would ideally like to be able to automate this to generate functions and gradients for nonlinear least squares and optimization calculations. The same criticism applies to the object old

Another issue:

If we have x and y set such that the function is not admissible, then both our old and new functions give a gradient that is seemingly reasonable. While the gradient of this simple function could be considered to be defined for ANY values of x and y, I (JN) am sure most users would wish for a warning at the very least in such cases.

x <- NA
y <- Inf


We could define a way to avoid the issue of character vs. expression (and possibly other classes) as follows:

safeD <- function(obj, var) {
   # safeguarded D() function for symbolic derivs
   if (! is.character(var) ) stop("The variable var MUST be character type")
   if (is.character(obj) ) {
       eobj <- parse(text=obj)
       result <- D(eobj, var)
   } else {
       result <- D(obj, var)

lxy2 <- expression(log(x+y^2))
clxy2 <- "log(x+y^2)"
try(print(D(clxy2, "y")))
print(try(D(lxy2, "y")))
print(safeD(clxy2, "y"))
print(safeD(lxy2, "y"))

Indexed parameters or variables

Erin Hodgess on R-help in January 2015 raised the issue of taking the derivative of an expression that contains an indexed variable. We show the example and its resolution, then give an explanation. Note that indexed parameters in nonlinear regression models is an open issue for package nlsr, largely because there does not seem to be an agreed mechanism to specify models with indexed parameters.

zzz <- expression(y[3]*r1 + r2)
try(nlsr::nlsDeriv(zzz, c("r1","r2")))
try(fnDeriv(zzz, c("r1","r2")))
newDeriv(`[`(x,y), stop("no derivative when indexing"))
try(nlsr::nlsDeriv(zzz, c("r1","r2")))
try(nlsr::fnDeriv(zzz, c("r1","r2")))

Richard Heiberger pointed out that internally, R stores




that is, as a function. Duncan Murdoch pointed out the availability of nlsr and the use of newDeriv() to redefine the "[" function for the purposes of derivatives.

This is not an ideal resolution, especially as we would like to be able to get the gradients of functions with respect to vectors of parameters, noted also by Sergei Sokol in the manual for package Deriv. The following examples illustrate this.

try(nlsr::nlsDeriv(zzz, "y[3]"))


Try the nlsr package in your browser

Any scripts or data that you put into this service are public.

nlsr documentation built on Sept. 8, 2023, 5:48 p.m.