The grad function in both the {pracma} and the {numDeriv} libraries of R gives erroneous results

Question

I am interested in the 1st order numerical derivative of a self-defined function pTgh_y(q,g,h) with respect to q. For a special case, pTgh_y(q,0,0) = pnorm(q). In other words pTgh_y(q,g,h) is reduced to the CDF of the standard normal when g=h=0 (see picture below).

This means that d pTgh_y(0,0,0)/dq should equal to the following

dnorm(0)

0.3989423

grad(pnorm,0)

0.3989423

Here are some of my attempts with the grad function in the {pracma} library.

library(pracma)
# load pTgh and all relevant functions
grad(function(x){pTgh_y(x,0,0)},0)

0

grad(function(x){pTgh_y(x,0,0)},0,heps=1e-10)

0

Here are some of my attempts with the grad function in the {numDeriv} library.

library(numDeriv)
# load pTgh and all relevant functions
grad(function(x){pTgh_y(x,0,0)},0,method='simple')

0.3274016

grad(function(x){pTgh_y(x,0,0)},0,method='Richardson')

-0.02505431

grad(function(x){pTgh_y(x,0,0)},0,method='complex')

Error in pmin(x, .Machine$double.xmax) : invalid input type Error in grad.default(function(x) { : function does not accept complex argument as required by method 'complex'.

None of these functions is giving correct result.

My pTgh_y(q,g,h) function is defined as follows

qTgh_y = function(p,g,h){                             
  zp = qnorm(p)
  if(g==0) q = zp                                    
  else     q = (exp(g*zp)-1)*exp(0.5*h*zp^2)/g       
  q[p==0] = -Inf
  q[p==1] = Inf
  return(q)
}

pTgh_y = function(q,g,h){                      
  if (q==-Inf) return(0)
  else if (q==Inf) return(1)
  else { 
    p = uniroot(function(t){qTgh_y(t,g,h)-q},interval=c(0,1))
    return(p$root)
  } 
}

Hans W. · Accepted Answer · 2016-02-26T11:12:28.870

Your function is flat around 0, thus computing a derivative of 0 is correct:

f = function(x){pTgh_y(x,0,0)}
h = 0.00001; f(0+h); f(0-h)
# [1] 0.5
# [1] 0.5

library(pracma)
grad(f, 0, heps=1e-02); grad(f, 0, heps=1e-03);
grad(f, 0, heps=1e-04); grad(f, 0, heps=1e-05)
# [1] 0.3989059
# [1] 0.399012
# [1] 0.4688766
# [1] 0

You need to increase the accuracy of your function pTgh_y:

pTgh_y = function(q,g,h){                      
    if (q==-Inf) return(0)
    else if (q==Inf) return(1)
    else { 
        p = uniroot(function(t){qTgh_y(t,g,h)-q},interval=c(0,1),
                    tol = .Machine$double.eps)
        return(p$root)
    } 
}

And now you get the result you wanted to have:

f = function(x){pTgh_y(x,0,0)}
grad(f, 0)
[1] 0.3989423

The grad function in both the {pracma} and the {numDeriv} libraries of R gives erroneous results

1 Answers1