Python Sympy Arbitrary Approximation to Arbitrary Sympy Expression?

Question

I find myself wanting to use approxmations provided as part of the mpmath package, but getting confused on exactly what they are supposed to be doing:

http://docs.sympy.org/dev/modules/mpmath/calculus/approximation.html

What exactly is the difference between a sympy expression and a sympy.mpmath expression ?

If I want a taylor approximation to a symbolic expression without understanding what mpmath package is doing I can do the following:

#Imports
import sympy
import sympy.parsing
import sympy.parsing.sympy_parser
import Library_TaylorApproximation

#Create a sympy expression to approximate
ExampleStringExpression = 'sin(x)'
ExampleSympyExpression = sympy.parsing.sympy_parser.parse_expr(ExampleStringExpression)


#Create a taylor expantion sympy expression around the point x=0
SympyTaylorApproximation = sympy.series( 
    ExampleSympyExpression,
    sympy.Symbol('x'),
    1, 
    4,
    ).removeO()

#Cast the sympy expressions to python functions which can be evaluated:
VariableNames = [str(var) for var in SympyTaylorApproximation.free_symbols]
PythonFunctionOriginal =  sympy.lambdify(VariableNames, ExampleSympyExpression)
PythonFunctionApproximation = sympy.lambdify(VariableNames, SympyTaylorApproximation)

#Evaluate the approximation and the original at a point:
print PythonFunctionOriginal(2)
print PythonFunctionApproximation(2)

#>>> 0.909297426826
#>>> 0.870987413961

However, if I try to do the same thing with mpmath based on the documentation:

TaylorCoefficients = sympy.mpmath.taylor(ExampleSympyExpression, 1, 4 )
print 'TaylorCoefficients', TaylorCoefficients

#>>> TypeError: 'sin' object is not callable

I can try to cram the python function in there (which is callable):

TaylorCoefficients = sympy.mpmath.taylor(PythonFunctionOriginal, 1, 4 )
print 'TaylorCoefficients', TaylorCoefficients

#>>> TaylorCoefficients [mpf('0.8414709848078965'), mpf('0.0'), mpf('0.0'), mpf('0.0'), mpf('-8.3694689805155739e+57')]

But the above does not make any sense, because I know that derivatives cannot be taken of a python function.

I can call the mpmath function sin :

TaylorCoefficients = sympy.mpmath.taylor(sympy.mpmath.sin, 1, 4 )
print 'TaylorCoefficients', TaylorCoefficients
#>>> TaylorCoefficients [mpf('0.8414709848078965'), mpf('0.54030230586813977'), mpf('-0.42073549240394825'), mpf('-0.090050384311356632'), mpf('0.035061291033662352')]

But then I cannot do manipulations on it the way I would want too -> like If I want

SinTimesCos = sympy.mpmath.sin*sympy.mpmath.cos
TaylorCoefficients = sympy.mpmath.taylor(SinTimesCos, 1, 4 )
print 'TaylorCoefficients', TaylorCoefficients
#>>> TypeError: unsupported operand type(s) for *: 'function' and 'function'

Exactly WHAT is an mpmath function ?

It is not a sympy expression, and it is also not a python function. How do I do manipulations on arbitrary expressions?

It would appear that I cannot take approximations of arbitrary sympy expressions in the documentation. http://docs.sympy.org/dev/modules/mpmath/calculus/approximation.html

How do I take arbitrary approximations ( Pade / Cheby Chev / Fourier ) to arbitrary sympy expressions?

EDIT:

So an example of what I am looking for is the following approximation:

#Start with a sympy expression of (a, b, x)
expressionString = 'cos(a*x)*sin(b*x)*(x**2)'
expressionSympy = sympy.parsing.sympy_parser.parse_expr(expressionString)

#Do not want to decide on value of `a or b` in advance.
#Do want approximation with respect to x:

wantedSympyExpression = SympyChebyChev( expressionSympy, sympy.Symbol('x') ) 

Result could either be a list of coefficient expressions that are functions of a , and b :

wantedSympyExpressionCoefficients = [ Coef0Expression(a,b), Coef1Expression(a,b), ... , CoefNExpression(a,b)]

OR the result could be the entire sympy expression itself (which is itself a function of a , b ):

wantedSympyExpression = Coef0Expression(a,b) + Coef1Expression(a,b) *(x**2) + ... + CoefNExpression(a,b) (x**N)

Note that a and b are not chosen in advance of performing the approximation.

Answer 1

mpmath functions are ordinary Python functions. They simply do their math in arbitrary-precision arithmetic.

But the above does not make any sense, because I know that derivatives cannot be taken of a python function.

You can't take the derivative symbolically , but you can compute an approximation of the derivative by evaluating the function several times and using numerical differentiation techniques. This is what sympy.mpmath.taylor does. Quoting the docs:

The coefficients are computed using high-order numerical differentiation. The function must be possible to evaluate to arbitrary precision.

Answer 2

If you have a SymPy expression and want to evaluate it to arbitrary precision, use evalf , like

sympy.sin(1).evalf(100)

You can use sin(x).evalf(100, subs={x:1}) to replace the x with 1 before evaluating. evalf uses mpmath under the hood, so this will give you the same result that mpmath would, but without having to use mpmath directly.

Answer 3

EDIT: Rereading my answer -> I thought I would fill in a few missing pieces as a service to someone someday actually using this. Below I labeled how I named my libraries, and what imports were required. I don't have the time to be a real contributor to sympy at the moment, but feel this functionality would certainly be used by other math/physics professors/students.

Note for space reasons the following two libraries are omitted, and I will throw a link to my repo at a future date.

import Library_SympyExpressionToPythonFunction

Creates a python callable function object with the same args ( number and names ) of the free variables in a sympy expression.

import Library_SympyExpressionToStringExpression

Literally just does str(SympyExpression)

#-------------------------------------------------------------------------------

Library_GenerateChebyShevPolynomial:::

#-------------------------------------------------------------------------------


import pprint
import Library_SympyExpressionToPythonFunction
import Library_SympyExpressionToStringExpression
import sympy
import sympy.core

def Main(
    ApproximationSymbol = sympy.Symbol('x'),
    ResultType = 'sympy',
    Kind= None,
    Order= None,
    ReturnAll = False,
    CheckArguments = True,
    PrintExtra = False,
    ):

    Result = None

    if (CheckArguments):
        ArgumentErrorMessage = ""

        if (len(ArgumentErrorMessage) > 0 ):
            if(PrintExtra):
                print "ArgumentErrorMessage:\n", ArgumentErrorMessage
            raise Exception(ArgumentErrorMessage)

    ChebyChevPolynomials = []
    ChebyChevPolynomials.append(sympy.sympify(1.))
    ChebyChevPolynomials.append(ApproximationSymbol)

    #Generate the polynomial with sympy:
    for Term in range(Order + 1)[2:]:
        Tn = ChebyChevPolynomials[Term - 1]
        Tnminus1 = ChebyChevPolynomials[Term - 2]
        Tnplus1 = 2*ApproximationSymbol*Tn - Tnminus1

        ChebyChevPolynomials.append(Tnplus1.simplify().expand().trigsimp())

    if(PrintExtra): print 'ChebyChevPolynomials'
    if(PrintExtra): pprint.pprint(ChebyChevPolynomials)


    if (ReturnAll):
        Result = []
        for SympyChebyChevPolynomial in ChebyChevPolynomials:
            if (ResultType == 'python'):
                Result.append(Library_SympyExpressionToPythonFunction.Main(SympyChebyChevPolynomial))
            elif (ResultType == 'string'):
                Result.append(Library_SympyExpressionToStringExpression.Main(SympyChebyChevPolynomial))
            else:
                Result.append(SympyChebyChevPolynomial)

    else:
        SympyExpression = ChebyChevPolynomials[Order] #the last one

        #If the result type is something other than sympy, we can cast it into that type here:
        if (ResultType == 'python'):
            Result = Library_SympyExpressionToPythonFunction.Main(SympyExpression)
        elif (ResultType == 'string'):
            Result = Library_SympyExpressionToStringExpression.Main(SympyExpression)
        else:
            Result = SympyExpression



    return Result 


#-------------------------------------------------------------------------------

Library_SympyChebyShevApproximationOneDimension

#-------------------------------------------------------------------------------


import numpy
import sympy
import sympy.mpmath
import pprint
import Library_SympyExpressionToPythonFunction
import Library_GenerateChebyShevPolynomial

def Main(
    SympyExpression= None,
    DomainMinimumPoint= None,
    DomainMaximumPoint= None,
    ApproximationOrder= None,
    CheckArguments = True,
    PrintExtra = False,
    ):

    #Tsymb = sympy.Symbol('t')
    Xsymb = sympy.Symbol('x')
    DomainStart = DomainMinimumPoint[0]
    print 'DomainStart', DomainStart
    DomainEnd = DomainMaximumPoint[0]
    print 'DomainEnd', DomainEnd

    #Transform the coefficients and the result to be on arbitrary inverval instead of from 0 to 1
    DomainWidth = DomainEnd - DomainStart
    DomainCenter = (DomainEnd - DomainStart) / 2.
    t = (Xsymb*(DomainWidth) + DomainStart + DomainEnd) / 2.
    x = (2.*Xsymb - DomainStart - DomainEnd) / (DomainWidth)
    SympyExpression = SympyExpression.subs(Xsymb, t)

    #GET THE COEFFICIENTS:
    Coefficients = []
    for CoefficientNumber in range(ApproximationOrder):
        if(PrintExtra): print 'CoefficientNumber', CoefficientNumber

        Coefficient = 0.0
        for k in range(1, ApproximationOrder + 1):
            if(PrintExtra): print '  k', k

            CoefficientFunctionPart = SympyExpression.subs(Xsymb, sympy.cos( sympy.pi*( float(k) - .5 )/ float(ApproximationOrder) )  )
            if(PrintExtra): print '  CoefficientFunctionPart', CoefficientFunctionPart

            CeofficientCosArg = float(CoefficientNumber)*( float(k) - .5 )/ float( ApproximationOrder)
            if(PrintExtra): print '  ',CoefficientNumber,'*','(',k,'-.5)/(', ApproximationOrder ,') == ', CeofficientCosArg

            CoefficientCosPart      =   sympy.cos( sympy.pi*CeofficientCosArg )
            if(PrintExtra): print '  CoefficientCosPart', CoefficientCosPart

            Coefficient += CoefficientFunctionPart*CoefficientCosPart

        if(PrintExtra): print 'Coefficient==', Coefficient

        Coefficient = (2./ApproximationOrder)*Coefficient.evalf(10)

        if(PrintExtra): print 'Coefficient==', Coefficient

        Coefficients.append(Coefficient)

    print '\n\nCoefficients'
    pprint.pprint( Coefficients )


    #GET THE POLYNOMIALS:
    ChebyShevPolynomials = Library_GenerateChebyShevPolynomial.Main(
        ResultType = 'sympy',
        Kind= 1,
        Order= ApproximationOrder-1,
        ReturnAll = True,
        )

    print '\nChebyShevPolynomials'
    pprint.pprint( ChebyShevPolynomials )


    Result = 0.0 -.5*(Coefficients[0])
    for Coefficient, ChebyShevPolynomial in zip(Coefficients, ChebyShevPolynomials):
        Result += Coefficient*ChebyShevPolynomial

    #Transform the coefficients and the result to be on arbitrary inverval instead of from 0 to 1
    Result = Result.subs(Xsymb, x)

    return Result

------------------------------------------------------------------------------

Example_SympyChebyShevApproximationOneDimension:

#------------------------------------------------------------------------------
import sympy
import sympy.mpmath
import matplotlib.pyplot as plt
import json
import pprint




import Library_GenerateBesselFunction
import Library_SympyChebyShevApproximationOneDimension
import Library_SympyExpressionToPythonFunction
import Library_GraphOneDimensionalFunction


ApproximationOrder = 10

#CREATE THE EXAMPLE EXRESSION:
Kind = 1
Order = 2
ExampleSympyExpression = sympy.sin(sympy.Symbol('x'))

"""
Library_GenerateBesselFunction.Main(
    ResultType =  'sympy',
    Kind =  Kind,
    Order =  Order,
    VariableNames = ['x'],
    ) 
"""
PythonOriginalFunction = Library_SympyExpressionToPythonFunction.Main( 
    ExampleSympyExpression ,
    FloatPrecision = 100,
    )

#CREATE THE NATIVE CHEBY APPROXIMATION

ChebyDomainMin = 5.
ChebyDomainMax = 10.
ChebyDomain = [ChebyDomainMin, ChebyDomainMax]
ChebyExpandedPolynomialCoefficients, ChebyError = sympy.mpmath.chebyfit(
    PythonOriginalFunction, 
    ChebyDomain, 
    ApproximationOrder, 
    error=True
    )
print 'ChebyExpandedPolynomialCoefficients'
pprint.pprint( ChebyExpandedPolynomialCoefficients )
def PythonChebyChevApproximation(Point):
    Result = sympy.mpmath.polyval(ChebyExpandedPolynomialCoefficients, Point)
    return Result


#CREATE THE GENERIC ONE DIMENSIONAL CHEBY APPROXIMATION:
SympyChebyApproximation = Library_SympyChebyShevApproximationOneDimension.Main(
    SympyExpression = ExampleSympyExpression*sympy.cos( sympy.Symbol('a') ),
    ApproximationSymbol = sympy.Symbol('x'),
    DomainMinimumPoint = [ChebyDomainMin],
    DomainMaximumPoint = [ChebyDomainMax],
    ApproximationOrder = ApproximationOrder
    )


print 'SympyChebyApproximation', SympyChebyApproximation

SympyChebyApproximation = SympyChebyApproximation.subs(sympy.Symbol('a'), 0.0)

print 'SympyChebyApproximation', SympyChebyApproximation

PythonCastedChebyChevApproximationGeneric = Library_SympyExpressionToPythonFunction.Main( 
    SympyChebyApproximation ,
    FloatPrecision = 100,
    )

print 'PythonCastedChebyChevApproximationGeneric(1)', PythonCastedChebyChevApproximationGeneric(1.)