grchk function

Run tests, where possible, on user objective function and (optionally) gradient and hessian

grchk checks a user-provided R function, ffn.

grchk(xpar, ffn, ggr, trace=0, testtol=(.Machine$double.eps)^(1/3), ...)

Arguments

• xpar: parameters to the user objective and gradient functions ffn and ggr
• ffn: User-supplied objective function
• ggr: User-supplied gradient function
• trace: set >0 to provide output from grchk to the console, 0 otherwise
• testtol: tolerance for equality tests
• ...: optional arguments passed to the objective function.

Details

Package:	grchk
Depends:	R (>= 2.6.1)
License:	GPL Version 2.

numDeriv is used to numerically approximate the gradient of function ffn

and compare this to the result of function ggr.

Returns

grchk returns a single object gradOK which is TRUE if the differences between analytic and approximated gradient are small as measured by the tolerance testtol.

This has attributes "ga" and "gn" for the analytic and numerically approximated gradients, and "maxdiff" for the maximum absolute difference between these vectors.

At the time of preparation, there are no checks for validity of the gradient code in ggr as in the function fnchk.

Author(s)

John C. Nash

Examples

# Would like examples of success and failure. What about "near misses"?
cat("Show how grchk works\n")
require(numDeriv)
# require(optimx)

jones<-function(xx){
  x<-xx[1]
  y<-xx[2]
  ff<-sin(x*x/2 - y*y/4)*cos(2*x-exp(y))
  ff<- -ff
}

jonesg <- function(xx) {
  x<-xx[1]
  y<-xx[2]
  gx <-  cos(x * x/2 - y * y/4) * ((x + x)/2) * cos(2 * x - exp(y)) - 
    sin(x * x/2 - y * y/4) * (sin(2 * x - exp(y)) * 2)
  gy <- sin(x * x/2 - y * y/4) * (sin(2 * x - exp(y)) * exp(y)) - cos(x * 
              x/2 - y * y/4) * ((y + y)/4) * cos(2 * x - exp(y))
  gg <- - c(gx, gy)
}

jonesg2 <- function(xx) {
  gx <- 1
  gy <- 2
  gg <- - c(gx, gy)
}

xx <- c(1, 2)

gcans <- grchk(xx, jones, jonesg, trace=1, testtol=(.Machine$double.eps)^(1/3))
gcans

gcans2 <- grchk(xx, jones, jonesg2, trace=1, testtol=(.Machine$double.eps)^(1/3))
gcans2