old wiki Sympy Research

Alexey U. Gudchenko edited this page Dec 2, 2011 · 4 revisions
Clone this wiki locally

old wiki Sympy Research

old wiki


This page collects documentation notes for the new trunk, which merged new core developed by Pearu with !SymPy. More information can be found in [http://code.google.com/p/sympy/issues/detail?id=144 Issue 144]


simply checkout the trunk, most things should be working the same way as before.

Mathematical constants

The following mathematical constants are defined in sympy:

  • I --- imaginary unit sqrt(-1)
  • oo --- positive infinity
  • nan --- undefined value
  • pi --- pi = 3.14..
  • E, exp(1) --- Euler number E = 2.718.., exp(1)

Use .evalf() method to get floating point approximation with the current precision Basic.set_precision().


>>> Basic.set_precision()
>>> pi.evalf(), E.evalf()
(3.141592653589793238462643383, 2.718281828459045235360287471)
>>> Basic.set_precision(40)
>>> pi.evalf(), E.evalf()
>>> I,I**2,oo,nan
(I, -1, oo, nan)

Assigning properties to symbols and expressions

Construction call

(, property1=value1, property2, ..)

Method call

<expr>.assume(property1=value1, property2=value2, ..)


  • property1, ... --- names of mathematical properties
  • value1, ... --- True or False or None


  • The following predefined mathematical propeties are supported (opposite properties are given in parenthesis):

    1. positive (nonpositive), negative (nonnegative) --- expr can have only positive or negative real values, respectively
    2. complex (noncomplex) --- expr has values from the set of complex numbers
    3. real (nonreal) --- expr has values from the set of real numbers
    4. rational (irrational) --- expr has values from the set of rational numbers
    5. integer (noninteger) --- expr has values from the set of integers
    6. odd (even) --- expr value is an odd integer
    7. prime (composite) --- expr value is a prime
    8. zero (nonzero) --- expr value is zero
    9. bounded (unbounded) --- expr absolute value is bounded
    10. finite (infinitesimal) --- expr value is nonzero and finite
    11. commutative (noncimmutative) --- expr value commutes with other values with respect to multiplication
    12. homogeneous (inhomogeneous) --- expr value is a homogeneous function with respect to its arguments
    13. comparable --- expr value can be compared with numbers (expr.evalf() returns Number object).
    14. pi is alias to integer, positive properties
    15. nni is alias to integer, 'nonnegative' properties
  • If property value is False, the opposite property is set. For example, setting positive=False is equivalent to nonpositive=True.

  • Assigned properties can be unset using property value None.

  • Unsetting a property will unset related properties. For example, unsetting real will also unset positive property, for instance.

  • expr properties can be requested via .is_property attributes. If no information is known for a given property then .is_property attribute is None.

  • expr properties can be set using property=value keyword arguments in the corresponding class constructor.


From a set of properties on can derive more properties. For example, if an expression has property positive then it follows that the expression has also a property being real (as the set of positive numbers is a subset of real numbers) as well as being complex (as the set of real numbers is a subset of complex numbers), etc. In general, a property, say P, may be a union of subproperties, say P^1^, P^2^, etc. For example, if P is real then P^1^ and P^2^ can be positive and nonpositive, respectively. Each property has opposite property, e.g. the opposite property of odd is even.

The following rules apply for derived properties:

  • if x sets P^i^ to True or False, then P is True
  • if x sets P^i^ to None, then P keeps its value.
  • if x sets P to True, P^i^ is not modified (it's probably None)
  • if x sets P to False or None, then P^i^ is set to None.
  • if x sets P^1^ to False or True or None, then its opposite property _P_^2^ is set to True or False or None, respectively.


>>> x = Symbol('x')
>>> x.is_real
>>> x.assume(positive=True)
>>> x.is_real
>>> x.is_negative
>>> exp(x).is_positive
>>> log(x).is_negative
>>> x.assume(infinitesimal=True)
>>> log(x).is_negative
>>> y = Symbol('y', even=True)
>>> y.is_integer
>>> y.is_odd

O – the domain of order terms (Landau symbols)

O(f, x) represents the Landau symbol O(f, x -> 0).


O(f, x, y, ..)


  • f --- an arithmetical expression representing a function in x, y, etc.
  • x, y --- variables, assumed to be infinitesimal and positive symbols

Return value

an O or Zero instance.


  • According to the mathematical definition, for a function f with variables x,y, .., the Landau symbol g = O(f, x, y, ..) is a representative of a class of functions having the following property: there exists a constant M and a neighborhood of the limit point (0, 0, ..) such that |g| <= M|f| for all values (x,y,..) in that neighborhood. Note that in general O(f(x,y),x,y) + O(g(x,y),x,y) is not the same as O(f(x,y)+g(x,y), x,y).

  • If variables x, y, .. are not specified then all symbols in f are considered as order variables. Otherwise all other symbols in f except x, y, .. are considered as parameters.

  • The symbols in O instance can be accessed via .symbols attribute.

  • The expression in O instance can be accessed via .expr attribute.

  • The inclusion relation between O(f, x) and O(g, x) is determined by the result of limit(f/g, x, 0). If the result is unbounded (e.g. oo) then O(g, x) is a subset of O(f, x) and O(f,x)+O(g,x) results O(f,x).

  • Automatic simplification is applied to O expression. See examples.

  • Assumptions infinitesimal=True, positive=True are set to O symbols.

  • Using O in the argument expression of a function leads to either Add or Order instance after applying the rule: f(g(x) + O(h(x),x)) -> f(g(x)) + O(h(x) f'(g(x)), x) provided that f'(g(x)) is not zero.


>>> from sympy.core import *
>>> x,y=Symbol('x'), Symbol('y')
>>> O(x,x),O(3*x,x),O(x+x**2),O(x+1/x),O(x+1/x+exp(2/x))
(O(x, x), O(x, x), O(x, x), O(1 / x, x), O(exp(2 / x), x))
>>> O(x+3*y), O(x+x*y), O(3*x*y**2-x**2*y+x**2*y**2), O(y**2*x**2+x*y**3,y)
(O(x + y, x, y), O(x, x), O(x * y ** 2 + y * x ** 2, x, y), O(y ** 2, y))
>>> O(x*y).expr, O(x*y).symbols
(x * y, (x, y))
>>> O(0,x), O(2,x)
(0, O(1))
>>> O(x) + O(y), O(x+y)
(O(x, x) + O(y, y), O(x + y, x, y))
>>> exp(1+x+O(x**3))
exp(1 + x) + O(x ** 3, x)
>>> cos(O(x))
1 + O(x ** 2, x)


To define a new function in !SymPy, a three step process is needed. First of all we need to create the function with its name, the number of arguments, default properties, series handling and, especially, evaluation procedure. As an example consider the following code:

class Im(DefinedFunction):
    nofargs = 1

    is_real = True

    def _eval_apply(self, arg):
        arg = Basic.sympify(arg)

        if isinstance(arg, Basic.NaN):
            return S.NaN
        elif arg.is_real:
            return S.Zero

This way new function called Im has been defined (which stands here for the simplified version of an imaginary part function, which is implemented in sympy/functions/elementary/complexes.py). Later this function will be identified via this name in lower case form.

The number of obligatory arguments is specified by setting nofargs class member. Additional arguments do not account here, but they will be visible in _eval_apply method. For example logarithm has fnoargs=1 for primary expression but you can also pass its base as an additional parameter. However logarithm function will rewrite itself to form with exactly one parameter (using the base change formula).

If function has properties which are valid for all arguments, you can set these using assumptions mechanism. In case of imaginary part, I've set is_real=True as this is what always holds.

Now the most important part of our function is evaluation procedure, _eval_apply, which will have exactly nofargs arguments without default value, plus additional with default value set to None. It's recommended, as the first step in this procedure, to check if all arguments are !SymPy compatible using Basic.sympify(). This step is a cheap if those arguments are already derived from the Basic class.

Next step is to analyze the arguments. First we needed to get rid of the optional ones. We must check for them and then rewrite the function so that only obligatory arguments will be present. If this is done, the next step is to put function into a normal form. What this term means depends on a concrete function and its implementation. However, the general rule is to perform non-expensive rewriting to function's arguments, which means you must avoid using complex procedures like polynomial algorithms, expand(), combine() functions and any general simplification procedures like simplify(). What you should use are all representation functions (those which begin with as_, also functions like has(), atoms() and eventually match().

If any reasonable rewrite is possible, then the function should return new value if function invocation is no more needed or self with new arguments otherwise. If no rewrite is possible for some classes of arguments then _eval_apply should return nothing ie. no return at all, or stand alone return statement or return None. In all three cases the super class, which is DefinedFunction, will fall back to the original function with arguments left untouched, for efficiency reasons.

If you follow the above steps you will get a !SymPy compatible function. However its capabilities are quite limited. What you can improve is to define its derivative and inverse, add floating point evaluation, power series expansion, non-default handling of value substitution etc.

It is important to notice that thte function defined is just an ordinary Python object so it is created with following syntax: function_name() with no arguments eg. Im(). One must not mix this with application of arguments to a function, which is done with calling an instance of defined function class ie. Im()(x) where x is a symbol. This way instance of Apply class is created which will handle 'runtime' properties of this function. Try in isympy shell the following:

>>> type(Basic.Log())
<class 'sympy.core.defined_function.Log'>

>>> type(Basic.Log()(x))
<class 'sympy.core.defined_function.ApplyLog'>

For usage simplicity Basic class provides singleton attribute, which is a mapping of simple names to instances of defined functions. For example you can write just log for the logarithm function rather than Log(). However notice that this is available only on run time and in library code you are obliged to use the full form.

When you issue log(x) in isympy shell or Log()(x) in library code, new instance of Apply class is created with first argument being the logarithm function and obligatory and also additional parameters.

Having only an instance of defined function class we are able to work with the function itself eg. we can compute function derivative. However the result might be interesting:

>>> log.fdiff()
lambda _x: 1 / _x


>>> sin.fdiff()

In the first example we got anonymous function (or just lambda for functional programmers). In the other we got another defined function. In both cases the result is unevaluated function (notice the f in front of diff). To compute the well known derivative we need function application rather than definition, eg.

>>> log.fdiff()(x)
1 / x

>>> log(y).diff(y)
1 / y


>>> sin(2*x).diff(x)

By default instance of Apply is created. You can change this behavior by overriding this class and implementing new specific one. Going back to the very first example, we can define the following:

class ApplyIm(Apply):

    def _eval_is_real(self):
        return True

    def _eval_complex_expand(self):
        return self.func(self.args[0].as_real_imag()[1])

When implementing such class we care about the runtime behavior of our defined function. Previously we stated that is_real=True. However now we have to restate the same in _eval form.

Notice also _eval_complex_expand method. Finally we have access to function's arguments and its name, which will be im in this case. What can be implemented more here are all functions with names beginning with _eval_* (with exceptions like _eval_apply and _eval_apply_evalf). You can also implement representations functions here (those with begin with as_) and tostr() method.

When you are finished with implementation, remember to add both new classes to ordering_of_classes list in core/basic_methods.py in appropriate places. This will help !SymPy properly display new functions in combination with all other objects.

The last step is to write tests for all functionalities you have implemented. Especially great care must be taken in testing evaluation procedure. For this there is separate test suit in core/tests/test_eval_apply.py. One general rule applies, all cases must be covered with at least one test. This way when we modify this function's logic or we change something very different we will be sure that with those modifications functions are evaluated properly. Of course the same applies to other code too.

Currently, there are all elementary functions implemented (exponential, trigonometric, hyperbolic and their inverses), complex components (real and imaginary parts, conjugate, argument and absolute value) and integer functions like floor and ceiling. All of these are located in core/defined_functions.py. There is also a preliminary support for some special functions ie. complete and incomplete gamma and zeta, and combinatorial functions, in specfun module. In most cases symbolic evaluation is done, however improvements are needed to power series expansions (especially handling of singularities) and assumptions.

Developers notes

Below some design decisions in the sympy.core are explained.

sympy.core uses a symbolic model where all symbolic objects are immutable, they are constructed from classes that are derived from Basic class. For example, sympy.core defines the following classes for elementary symbolic manipulations: Symbol, Number, Add, Mul, Pow, Function, Apply. There is a number of other classes derived from Basic or classes listed above for adding more features to the used symbolic model. For example, Symbol is a base class to Wild, Temporary and Number is a base class to Real, Rational, and Integer classes.

The implementation of the symbolic model uses many advanced features from the Python language such as newstyle classes, metaclasses, properties, decorators, etc. The goal is to keep code minimal, readable, and to get maximum performance using pure Python code while keeping in mind that in future some computationally intensive code should be implemented in a C/C++ extension module.

For efficiency, the following is taken into account:

  • In pure Python constructing an object is an expensive operations. This is especially true for symbolic objects that may need to evaluate their arguments to some normal form. For example, constructing a rational number involves finding gcd of its numerator and denominator. In constructing Add and Mul instances some elementary simplifications (evaluations) are carried out. Therefore, many frequently used symbolic notions like numbers are created only once (via singleton or caching technique) and computed results that might be needed in further operations are cached.

  • Also, the result of an evaluation may be different from the original constructor. For example, Add(x, x) will result in Mul(2,x). In this situation, sympy.core will never create an intermediate object representing x+x but carries out evaluation and then creates a final object representing 2*x.

  • Comparing symbolic objects for an equality can be very expensive as one may need to walk through large trees of symbolic objects. Therefore, the code must be optimized for elementary comparisons such as if a symbolic object is an instance of some special number, etc. So, a codelet like obj == 1 should be replaced with obj is S.One or isinstance(obj, Basic.One) or obj.is_one, that involve only single operations, for efficency. Note that executing obj==1 involves a number of operations, it is equivalent to the following code: Equality(obj, Basic.sympify(1)).__nonzero__() where the __nonzero__ code contains comparison code for symbolic objects.

  • In sympy.core the comparison of symbolic objects takes advantage of the fact that all symbolic objects are strictly ordered (not always in mathematical sense, of course). This is achived by sorting the content of otherwise commutative operations such as Add and Mul. The order of symbolic objects is first defined by the class name in the ordering_of_classes list in basic_methods.py file, and then by the order of tuple representation of symbolic objects returned by the ._hashable_content() methods. See the implementation of Basic.compare for more details.

For extending the symbolic model, the following is taken into account:

  • Different symbolic classes may need to have access to each other features while the classes are implemented in different modules or even in different packages. Therefore one must be careful with importing modules as in the given situation cyclic imports are introduced. sympy.core resolves this issue by making all Basic subclasses to be attributes of the Basic class at the moment of defining the subclasses. As a result, one only needs to import Basic from sympy.core.basic module to get access to all classes that are derived from the Basic class.