�
��g` � � � d Z ddlmZ ddlmZmZmZmZmZ ddl m
Z
ddlmZ ddl
mZmZ ddlmZmZ ddlmZ dd lmZmZmZ dd
lmZ ddlmZ ddlmZ dd
l!m"Z" e"d� � Z#d� Z$ G d� de%� � Z&e d� � � Z'd� Z(d� Z)e e
e#d� � � � � � � Ze e#e
d� � � � � � � Z*d� Z+d� Z,e e#d d�� � � � Z-e e#e
d� � � � � � � Z.d� Z/e e#d� � � � � Z0d!d�Z1dS )"aS
Limits
======
Implemented according to the PhD thesis
https://www.cybertester.com/data/gruntz.pdf, which contains very thorough
descriptions of the algorithm including many examples. We summarize here
the gist of it.
All functions are sorted according to how rapidly varying they are at
infinity using the following rules. Any two functions f and g can be
compared using the properties of L:
L=lim log|f(x)| / log|g(x)| (for x -> oo)
We define >, < ~ according to::
1. f > g .... L=+-oo
we say that:
- f is greater than any power of g
- f is more rapidly varying than g
- f goes to infinity/zero faster than g
2. f < g .... L=0
we say that:
- f is lower than any power of g
3. f ~ g .... L!=0, +-oo
we say that:
- both f and g are bounded from above and below by suitable integral
powers of the other
Examples
========
::
2 < x < exp(x) < exp(x**2) < exp(exp(x))
2 ~ 3 ~ -5
x ~ x**2 ~ x**3 ~ 1/x ~ x**m ~ -x
exp(x) ~ exp(-x) ~ exp(2x) ~ exp(x)**2 ~ exp(x+exp(-x))
f ~ 1/f
So we can divide all the functions into comparability classes (x and x^2
belong to one class, exp(x) and exp(-x) belong to some other class). In
principle, we could compare any two functions, but in our algorithm, we
do not compare anything below the class 2~3~-5 (for example log(x) is
below this), so we set 2~3~-5 as the lowest comparability class.
Given the function f, we find the list of most rapidly varying (mrv set)
subexpressions of it. This list belongs to the same comparability class.
Let's say it is {exp(x), exp(2x)}. Using the rule f ~ 1/f we find an
element "w" (either from the list or a new one) from the same
comparability class which goes to zero at infinity. In our example we
set w=exp(-x) (but we could also set w=exp(-2x) or w=exp(-3x) ...). We
rewrite the mrv set using w, in our case {1/w, 1/w^2}, and substitute it
into f. Then we expand f into a series in w::
f = c0*w^e0 + c1*w^e1 + ... + O(w^en), where e0<e1<...<en, c0!=0
but for x->oo, lim f = lim c0*w^e0, because all the other terms go to zero,
because w goes to zero faster than the ci and ei. So::
for e0>0, lim f = 0
for e0<0, lim f = +-oo (the sign depends on the sign of c0)
for e0=0, lim f = lim c0
We need to recursively compute limits at several places of the algorithm, but
as is shown in the PhD thesis, it always finishes.
Important functions from the implementation:
compare(a, b, x) compares "a" and "b" by computing the limit L.
mrv(e, x) returns list of most rapidly varying (mrv) subexpressions of "e"
rewrite(e, Omega, x, wsym) rewrites "e" in terms of w
leadterm(f, x) returns the lowest power term in the series of f
mrv_leadterm(e, x) returns the lead term (c0, e0) for e
limitinf(e, x) computes lim e (for x->oo)
limit(e, z, z0) computes any limit by converting it to the case x->oo
All the functions are really simple and straightforward except
rewrite(), which is the most difficult/complex part of the algorithm.
When the algorithm fails, the bugs are usually in the series expansion
(i.e. in SymPy) or in rewrite.
This code is almost exact rewrite of the Maple code inside the Gruntz
thesis.
Debugging
---------
Because the gruntz algorithm is highly recursive, it's difficult to
figure out what went wrong inside a debugger. Instead, turn on nice
debug prints by defining the environment variable SYMPY_DEBUG. For
example:
[user@localhost]: SYMPY_DEBUG=True ./bin/isympy
In [1]: limit(sin(x)/x, x, 0)
limitinf(_x*sin(1/_x), _x) = 1
+-mrv_leadterm(_x*sin(1/_x), _x) = (1, 0)
| +-mrv(_x*sin(1/_x), _x) = set([_x])
| | +-mrv(_x, _x) = set([_x])
| | +-mrv(sin(1/_x), _x) = set([_x])
| | +-mrv(1/_x, _x) = set([_x])
| | +-mrv(_x, _x) = set([_x])
| +-mrv_leadterm(exp(_x)*sin(exp(-_x)), _x, set([exp(_x)])) = (1, 0)
| +-rewrite(exp(_x)*sin(exp(-_x)), set([exp(_x)]), _x, _w) = (1/_w*sin(_w), -_x)
| +-sign(_x, _x) = 1
| +-mrv_leadterm(1, _x) = (1, 0)
+-sign(0, _x) = 0
+-limitinf(1, _x) = 1
And check manually which line is wrong. Then go to the source code and
debug this function to figure out the exact problem.
� )�reduce)�Basic�S�Mul� PoleError�
expand_mul)�cacheit)�ilcm)�I�oo)�Dummy�Wild)� bottom_up)�log�exp�sign)�Order)�SymPyDeprecationWarning)�debug_decorator)�timethis�gruntzc � � t | � � t |� � }}t | t � � r8t | t � � s| j r| j t j k r| j }t |t � � r8t |t � � s|j r|j t j k r|j }t ||z |� � }|dk rdS |j rdS dS )z/Returns "<" if a<b, "=" for a == b, ">" for a>br �<�>�=)
r �
isinstancer r �is_Pow�baser �Exp1�limitinf�is_infinite)�a�b�x�la�lb�cs �c/home/asafur/pinokio/api/open-webui.git/app/env/lib/python3.11/site-packages/sympy/series/gruntz.py�comparer) � s� � � ��V�V�S��V�V��B��!�U��� ��A�s�!3�!3� ��� �Q�V�q�v�EU�EU�
�U���!�U��� ��A�s�!3�!3� ��� �Q�V�q�v�EU�EU�
�U����B�����A��A�v�v��s�
�� ��s��s� c �H � � e Zd ZdZd� Z� fd�Zd� Zd� Zd� Zd
d�Z d � Z
� xZS )�SubsSeta~
Stores (expr, dummy) pairs, and how to rewrite expr-s.
Explanation
===========
The gruntz algorithm needs to rewrite certain expressions in term of a new
variable w. We cannot use subs, because it is just too smart for us. For
example::
> Omega=[exp(exp(_p - exp(-_p))/(1 - 1/_p)), exp(exp(_p))]
> O2=[exp(-exp(_p) + exp(-exp(-_p))*exp(_p)/(1 - 1/_p))/_w, 1/_w]
> e = exp(exp(_p - exp(-_p))/(1 - 1/_p)) - exp(exp(_p))
> e.subs(Omega[0],O2[0]).subs(Omega[1],O2[1])
-1/w + exp(exp(p)*exp(-exp(-p))/(1 - 1/p))
is really not what we want!
So we do it the hard way and keep track of all the things we potentially
want to substitute by dummy variables. Consider the expression::
exp(x - exp(-x)) + exp(x) + x.
The mrv set is {exp(x), exp(-x), exp(x - exp(-x))}.
We introduce corresponding dummy variables d1, d2, d3 and rewrite::
d3 + d1 + x.
This class first of all keeps track of the mapping expr->variable, i.e.
will at this stage be a dictionary::
{exp(x): d1, exp(-x): d2, exp(x - exp(-x)): d3}.
[It turns out to be more convenient this way round.]
But sometimes expressions in the mrv set have other expressions from the
mrv set as subexpressions, and we need to keep track of that as well. In
this case, d3 is really exp(x - d2), so rewrites at this stage is::
{d3: exp(x-d2)}.
The function rewrite uses all this information to correctly rewrite our
expression in terms of w. In this case w can be chosen to be exp(-x),
i.e. d2. The correct rewriting then is::
exp(-w)/w + 1/w + x.
c � � i | _ d S �N)�rewrites��selfs r( �__init__zSubsSet.__init__� s
� ���
�
�
r* c �~ �� t � � � � � dz | j � � � z S )Nz, )�super�__repr__r/ )r1 � __class__s �r( r5 zSubsSet.__repr__� s1 �� ��w�w���!�!�D�(�4�=�+A�+A�+C�+C�C�Cr* c �b � || vrt � � | |<