Re: Understanding Mathematica's evaluation procedure
- To: mathgroup at smc.vnet.net
- Subject: [mg117421] Re: Understanding Mathematica's evaluation procedure
- From: Leonid Shifrin <lshifr at gmail.com>
- Date: Fri, 18 Mar 2011 05:59:13 -0500 (EST)
Alexey,
> As I understand, by default no replacement rules are associated with
> these functions, and they work "themselves".
>
> I was always confounded by this aspect of the system. The edge between
> replacement and "the true functions which do the job" always remained
> stumbling-block for my understanding.
>
> Can anybody clarify this moment?
>
My understanding is that there isn't much mystery here. Many built-in
functions in
Mathematica can work with both symbolic and numeric arguments. The symbolic
part
is governed by rules, just as for the user-defined functions. Some of these
rules are written
in the top-level Mathematica and can be inspected. Most are written on the
system level
and can not be inspected, but they are still replacement rules. So, many
functions in
Mathematica are inter-connected with a large number of rules governing their
symbolic
relationships, simplifications etc.
On numerical (or, generally, atomic) arguments, the rules may just invoke
the implementations
that would return a number, list of numbers, etc. Here is an example to
illustrate: we define our
own sine function, and add just one simplification rule, as follows:
Clear[sin];
sin[x_ + y_] := sin[x]*cos[y] + sin[y]*cos[x];
sin[x_?NumericQ] := sinC5[x];
where sinC is a Compiled to C function retaining only the first 5 terms of
the sine expansion
(to keep everything within a machine precision - my point here is just to
illustrate):
sinC5 = Compile[{{num, _Real}},
Sum[(-1)^k num^(2 k + 1)/(2 k + 1)!, {k, 0, 5}],
CompilationTarget -> "C"]
Now, when we call:
In[162]:= sin[a + 1]
Out[162]= 0.841471 cos[a] + cos[1] sin[a]
So, what happened is that when the rule matched for a numeric argument, it
*invoked* a lower-level
C function which did the computation. The system functions work in a very
similar way - the rules are
applied, and when they match on "terminals" - lower-level function calls,
they invoke them and substitute
back the result of their evaluation. Apart from the somewhat special way
that the rules work (the main
evaluation loop), what happens here is no different from what happens in any
language interpreter. You
need a way to map your higher-level language statements and expressions onto
an "instruction set"
available to you in the implementation language. What makes Mathematica
perhaps really different is
that some computations may be purely symbolic and involve only
simplifications based on rules (term
rewriting), which is indeed probably not how most programming languages
work.
Regards,
Leonid