refactor symbolic functions

[burcin]

Attached patch refactors the symbolic function code in sage/symbolic/function.pyx.

• evalf() now accepts a parent argument instead of a precision
  This allows us to use the numeric evaluation framework in ginac for
  evaluating things with RIF, CIF as well, not just RealField, or ComplexField with the given precision.

• python arguments passed to custom methods of sfunctions are not
  wrapped in Expression objects any more. No need to call .pyobject()
  to unwrap these.

• custom methods support calling methods on self.
  This would be useful if you need access to other function of the
  defining class, or store tables of data calculated on demand.

• __call__ method supports hold parameter
  This works:

        sage: exp(log(x))
        x
        sage: exp(log(x), hold=True)
        e^log(x)

• Custom methods for symbolic functions (_eval_, _evalf_, _conjugate_,
  _derivative_, etc.) can be
  written in Cython for builtin functions (that are provided by the
  Sage library)

• New class hiearchy:

Function
  GinacFunction
  CustomizableFunction
    BuiltinFunction
    SymbolicFunction

We have 4 different types of functions, those defined by

• ginac (sin, cos, ...),
• the Sage library (cot)
• the user (in a python file, subclassing the new
  SymbolicFunction)
• the command line function_factory (by calling function('f') )

Things we need to do for these functions different for each of these,
perhaps similar for the last two. Normally initializing a function
means checking if it's already defined, if not, initializing a
structure from ginac called function_options, and registering this in
a table. There are also issues with pickling.

For ginac functions, we don't need any of this, since we can't change
it at python level. We only need to look up the serial number (the
indicator in the table) of the function. We don't need to do anything
to pickle or unpickle these either.

Pickling and unpickling library functions only needs an identifier
for the class to initialize it again if necessary.

User defined functions need to lookup if there is an existing
function in the table, since we should try to keep the table small.

There is also a new function_factory() function in sage.symbolic.function_factory
(it needs to be in a python file) that creates NewSymbolicFunction
classes on the fly for the function() calls from the command line.

The pynac package here is required for this patch:

http://sage.math.washington.edu/home/burcin/pynac/pynac-0.1.10.a0.spkg

CC: @mwhansen @jasongrout

Component: symbolics

Keywords: pynac

Author: Burcin Erocal

Reviewer: Mike Hansen

Merged: sage-4.3.alpha1

Issue created by migration from https://trac.sagemath.org/ticket/7490

[jasongrout]

comment:2

After applying:

sage: integrate(e^x,(x,0,2),hold=True)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)

/home/grout/.sage/temp/good/20605/_home_grout__sage_init_sage_0.py in <module>()

/home/grout/sage/local/lib/python2.6/site-packages/sage/misc/functional.pyc in integral(x, *args, **kwds)
    566     """
    567     if hasattr(x, 'integral'):
--> 568         return x.integral(*args, **kwds)
    569     else:
    570         from sage.symbolic.ring import SR

/home/grout/sage/local/lib/python2.6/site-packages/sage/symbolic/expression.so in sage.symbolic.expression.Expression.integral (sage/symbolic/expression.cpp:25362)()

TypeError: integral() got an unexpected keyword argument 'hold'

[burcin]

refactor symbolic functions

[burcin]

comment:3

Attachment: trac_7490-refactor_symbolic_functions.patch.gz

Hi Jason,

This patch does not add a hold method to integrate() since it is not a symbolic function. Golam did some work in this direction at #6465, but more effort is needed to polish those patches.

Cheers,

Burcin

[burcin]

comment:4

I uploaded a new patch with a doctest fix in sage.rings.arith. The patch applies cleanly to 4.2.1 and passes all doctests (even long!). I think it is ready for review.

This also fixes #6523 and #6286. It should provide a basis to #6949, #6961, #6220 and #6465.

Note that the patch touches 42 files. I would appreciate it if we can get it in 4.3.

Mike, while working on this, I had to tackle many of the problems you must have encountered during the symbolics switch. Can you take a look to see if I forgot anything/messed things up?

[burcin]

rebased to 4.3.alpha0

[burcin]

comment:5

Attachment: trac_7490-refactor_symbolic_functions.rebase-4.3.alpha0.patch.gz

I rebased the patch to 4.3.alpha0:

attachment: trac_7490-refactor_symbolic_functions.rebase-4.3.alpha0.patch

[mwhansen]

comment:6

Thanks Burcin. I'll tried to get it reviewed ASAP and have it the next thing in.

[mwhansen]

comment:7

Here's my review.

There are a number of things which break old code -- they should be deprecated first.

- exp(2,prec=100), gamma(pi,prec=100), etc.

- sage: Q.<i> = NumberField(x^2+1) 
  sage: gamma(i) 
  sage: gamma(QQbar(I))

Conversion of polylog to maxima is broken:

sage: polylog(2, x)._maxima_init_()
'polylog(2,x)'

instead of 'li[2](x)'.

Some doctests are missing:

sage/interfaces/maxima.py: _symbolic_
sage/rings/number_field/number_field_element.pyx: _mpfr_, __complex__

Why do you have to use

f = CallableConvertMap(RR, RR, lambda x: x.exp(), parent_as_first_arg=False) 

instead of

f = CallableConvertMap(RR, RR, exp, parent_as_first_arg=False) 

, which is more natural?

In expression.pyx, some things are missing from the _convert docstring. Also, f._convert(int) gives -0.989992496600445*sqrt(2) which seems unexpected. Maybe the docstring can be clarified further?

Finally, there are some numerical issues it seems with evaluations: complex(I) gives 0.99999999999999967j instead of 1j. I'm not sure where the discrepancy is occurring.

[mwhansen]

Reviewer: Mike Hansen