[Rd] Exchange of information between functions aside from argument/return-value interface

[Peter Ruckdeschel]

Hi R-devels,

I would like to ask for your advice on the preferred way to exchange
information between functions aside from the regular argument/return-value
interface.

I am well aware that, S being a functional language, this is to be avoided
where possible. There are exceptions, though:

Let myF() and myH() be functions provided by me, and notmyG() be a function
provided by someone else.

The calling hierarchie is myF() calls notmyG() which calls --- possibly
with other functions being called in between ---  myH().

notmyG() takes as argument a function with return value of type numeric
--- typically a function for optimizing, integrating or zero-finding ---
and is a function which I do not want to modify (and subsequently maintain
the modification).

As myH() is expensive as to evaluation time and, as a side effect, also
provides important side information, I would like to pass this information
back to myF(). (Also see example below).

Using information from myF() in myH() by lexical scoping is straight
forward, whereas to pass information from myH() to myF() directly, the
solution I found so far is to use assign() with a corresponding
envir=./pos=.-argument.

(Q1) Is there a better way to do so?

While it is not too difficult to find out the correct value for the
envir/pos argument when using "pure" R-functions for notmyG(), the
situation is a little more complicated as soon as notmyG() calls myH()
through intermediate calls to .Internal() as in case of optim(),
uniroot().

Is there any way to

(Q2a) find out a corresponding correct value for argument envir/pos in assign
      ---on run-time---, i.e. how many layers of calls are between
      myF() and myH()?

 --- my first guess was to use sys.frame() but I am not convinced...

(Q2b) to make myH() when called from within a call to .Internal()
      perform assignments in the environment of myF()

?

Or should we proceed in a completely different way?

Thank you already for your advice,

Peter

--------------------------------------------------------------------------
--- Example:

To given clipping height b, myH() calculates an "optimal"  influence curve
which involves determining certain Lagrange multipliers in a fixed-point type
algorithm.
To determine b according to a certain criterium, moreover, myH() evaluates
a function H0 in b, the zero of which will be the "optimal" b.

So to use notmyG()=uniroot() as zero-finder the return value of myH()
should be H0(b), while myF() should also get to know about the Lagrange
multipliers determined in myH().