def _eval_expand_basic(self, **hints):
summand = self.function.expand(**hints)
if summand.is_Add and summand.is_commutative:
return Add(*[self.func(i, *self.limits) for i in summand.args])
elif summand != self.function:
return self.func(summand, *self.limits)
return self
def _eval_nseries(self, x, n, logx):
expr = self.as_dummy()
symb = x
for l in expr.limits:
if x in l[1:]:
symb = l[0]
break
terms, order = expr.function.nseries(x=symb, n=n,
logx=logx).as_coeff_add(Order)
return integrate(terms, *expr.limits) + Add(*order) * x
def _eval_nseries(self, x, n, logx):
if len(self.args) == 1:
from diofant import O, Add, Integer, factorial
x = self.args[0]
o = O(x**n, x)
l = S.Zero
if n > 0:
l += Add(*[Integer(-k)**(k - 1)*x**k/factorial(k)
for k in range(1, n)])
return l + o
return super(LambertW, self)._eval_nseries(x, n=n, logx=logx)
def _eval_expand_log(self, deep=True, **hints):
from diofant import unpolarify, expand_log
from diofant.concrete import Sum, Product
force = hints.get('force', False)
if (len(self.args) == 2):
return expand_log(self.func(*self.args), deep=deep, force=force)
arg = self.args[0]
if arg.is_Integer:
# remove perfect powers
p = perfect_power(int(arg))
if p is not False:
return p[1]*self.func(p[0])
elif arg.is_Mul:
expr = []
nonpos = []
for x in arg.args:
if force or x.is_positive or x.is_polar:
a = self.func(x)
if isinstance(a, log):
expr.append(self.func(x)._eval_expand_log(**hints))
else:
expr.append(a)
elif x.is_negative:
a = self.func(-x)
expr.append(a)
nonpos.append(S.NegativeOne)
else:
nonpos.append(x)
return Add(*expr) + log(Mul(*nonpos))
elif arg.is_Pow:
if force or (arg.exp.is_extended_real and arg.base.is_positive) or \
arg.base.is_polar:
b = arg.base
e = arg.exp
a = self.func(b)
if isinstance(a, log):
return unpolarify(e) * a._eval_expand_log(**hints)
else:
return unpolarify(e) * a
elif isinstance(arg, Product):
if arg.function.is_positive:
return Sum(log(arg.function), *arg.limits)
return self.func(arg)
def evalf_log(expr, prec, options):
from diofant import Abs, Add, log
if len(expr.args) > 1:
expr = expr.doit()
return evalf(expr, prec, options)
arg = expr.args[0]
workprec = prec + 10
xre, xim, xacc, _ = evalf(arg, workprec, options)
if xim:
# XXX: use get_abs etc instead
re = evalf_log(log(Abs(arg, evaluate=False), evaluate=False), prec,
options)
im = mpf_atan2(xim, xre or fzero, prec)
return re[0], im, re[2], prec
imaginary_term = (mpf_cmp(xre, fzero) < 0)
re = mpf_log(mpf_abs(xre), prec, rnd)
size = fastlog(re)
if prec - size > workprec:
# We actually need to compute 1+x accurately, not x
arg = Add(S.NegativeOne, arg, evaluate=False)
xre, xim, _, _ = evalf_add(arg, prec, options)
prec2 = workprec - fastlog(xre)
# xre is now x - 1 so we add 1 back here to calculate x
re = mpf_log(mpf_abs(mpf_add(xre, fone, prec2)), prec, rnd)
re_acc = prec
if imaginary_term:
return re, mpf_pi(prec), re_acc, prec
else:
return re, None, re_acc, None
def eval(cls, p, q):
from diofant.core.add import Add
from diofant.core.mul import Mul
from diofant.core.singleton import S
from diofant.core.exprtools import gcd_terms
from diofant.polys.polytools import gcd
def doit(p, q):
"""Try to return p % q if both are numbers or +/-p is known
to be less than or equal q.
"""
if p.is_infinite or q.is_infinite:
return nan
if (p == q or p == -q
or p.is_Pow and p.exp.is_Integer and p.base == q
or p.is_integer and q == 1):
return S.Zero
if q.is_Number:
if p.is_Number:
return (p % q)
if q == 2:
if p.is_even:
return S.Zero
elif p.is_odd:
return S.One
# by ratio
r = p / q
try:
d = int(r)
except TypeError:
pass
else:
rv = p - d * q
if (rv * q).is_nonnegative:
return rv
elif (rv * q).is_nonpositive:
return rv + q
# by difference
d = p - q
if d.is_negative:
if q.is_negative:
return d
elif q.is_positive:
return p
rv = doit(p, q)
if rv is not None:
return rv
# denest
if p.func is cls:
# easy
qinner = p.args[1]
if qinner == q:
return p
# XXX other possibilities?
# extract gcd; any further simplification should be done by the user
G = gcd(p, q)
if G != 1:
p, q = [
gcd_terms(i / G, clear=False, fraction=False) for i in (p, q)
]
pwas, qwas = p, q
# simplify terms
# (x + y + 2) % x -> Mod(y + 2, x)
if p.is_Add:
args = []
for i in p.args:
a = cls(i, q)
if a.count(cls) > i.count(cls):
args.append(i)
else:
args.append(a)
if args != list(p.args):
p = Add(*args)
else:
# handle coefficients if they are not Rational
# since those are not handled by factor_terms
# e.g. Mod(.6*x, .3*y) -> 0.3*Mod(2*x, y)
cp, p = p.as_coeff_Mul()
cq, q = q.as_coeff_Mul()
ok = False
if not cp.is_Rational or not cq.is_Rational:
r = cp % cq
if r == 0:
G *= cq
p *= int(cp / cq)
ok = True
if not ok:
p = cp * p
q = cq * q
# simple -1 extraction
if p.could_extract_minus_sign() and q.could_extract_minus_sign():
G, p, q = [-i for i in (G, p, q)]
# check again to see if p and q can now be handled as numbers
rv = doit(p, q)
if rv is not None:
return rv * G
# put 1.0 from G on inside
if G.is_Float and G == 1:
p *= G
return cls(p, q, evaluate=False)
elif G.is_Mul and G.args[0].is_Float and G.args[0] == 1:
p = G.args[0] * p
G = Mul._from_args(G.args[1:])
return G * cls(p, q, evaluate=(p, q) != (pwas, qwas))
def test_core_add():
for c in (Add, Add(x, 4)):
check(c)
def _eval_integral(self,
f,
x,
meijerg=None,
risch=None,
conds='piecewise'):
"""
Calculate the anti-derivative to the function f(x).
The following algorithms are applied (roughly in this order):
1. Simple heuristics (based on pattern matching and integral table):
- most frequently used functions (e.g. polynomials, products of trig functions)
2. Integration of rational functions:
- A complete algorithm for integrating rational functions is
implemented (the Lazard-Rioboo-Trager algorithm). The algorithm
also uses the partial fraction decomposition algorithm
implemented in apart() as a preprocessor to make this process
faster. Note that the integral of a rational function is always
elementary, but in general, it may include a RootSum.
3. Full Risch algorithm:
- The Risch algorithm is a complete decision
procedure for integrating elementary functions, which means that
given any elementary function, it will either compute an
elementary antiderivative, or else prove that none exists.
Currently, part of transcendental case is implemented, meaning
elementary integrals containing exponentials, logarithms, and
(soon!) trigonometric functions can be computed. The algebraic
case, e.g., functions containing roots, is much more difficult
and is not implemented yet.
- If the routine fails (because the integrand is not elementary, or
because a case is not implemented yet), it continues on to the
next algorithms below. If the routine proves that the integrals
is nonelementary, it still moves on to the algorithms below,
because we might be able to find a closed-form solution in terms
of special functions. If risch=True, however, it will stop here.
4. The Meijer G-Function algorithm:
- This algorithm works by first rewriting the integrand in terms of
very general Meijer G-Function (meijerg in Diofant), integrating
it, and then rewriting the result back, if possible. This
algorithm is particularly powerful for definite integrals (which
is actually part of a different method of Integral), since it can
compute closed-form solutions of definite integrals even when no
closed-form indefinite integral exists. But it also is capable
of computing many indefinite integrals as well.
- Another advantage of this method is that it can use some results
about the Meijer G-Function to give a result in terms of a
Piecewise expression, which allows to express conditionally
convergent integrals.
- Setting meijerg=True will cause integrate() to use only this
method.
5. The Heuristic Risch algorithm:
- This is a heuristic version of the Risch algorithm, meaning that
it is not deterministic. This is tried as a last resort because
it can be very slow. It is still used because not enough of the
full Risch algorithm is implemented, so that there are still some
integrals that can only be computed using this method. The goal
is to implement enough of the Risch and Meijer G-function methods
so that this can be deleted.
"""
from diofant.integrals.deltafunctions import deltaintegrate
from diofant.integrals.heurisch import heurisch, heurisch_wrapper
from diofant.integrals.rationaltools import ratint
from diofant.integrals.risch import risch_integrate
if risch:
try:
return risch_integrate(f, x, conds=conds)
except NotImplementedError:
return
# if it is a poly(x) then let the polynomial integrate itself (fast)
#
# It is important to make this check first, otherwise the other code
# will return a diofant expression instead of a Polynomial.
#
# see Polynomial for details.
if isinstance(f, Poly) and not meijerg:
return f.integrate(x)
# Piecewise antiderivatives need to call special integrate.
if f.func is Piecewise:
return f._eval_integral(x)
# let's cut it short if `f` does not depend on `x`
if not f.has(x):
return f * x
# try to convert to poly(x) and then integrate if successful (fast)
poly = f.as_poly(x)
if poly is not None and not meijerg:
return poly.integrate().as_expr()
if risch is not False:
try:
result, i = risch_integrate(f,
x,
separate_integral=True,
conds=conds)
except NotImplementedError:
pass
else:
if i:
# There was a nonelementary integral. Try integrating it.
return result + i.doit(risch=False)
else:
return result
# since Integral(f=g1+g2+...) == Integral(g1) + Integral(g2) + ...
# we are going to handle Add terms separately,
# if `f` is not Add -- we only have one term
# Note that in general, this is a bad idea, because Integral(g1) +
# Integral(g2) might not be computable, even if Integral(g1 + g2) is.
# For example, Integral(x**x + x**x*log(x)). But many heuristics only
# work term-wise. So we compute this step last, after trying
# risch_integrate. We also try risch_integrate again in this loop,
# because maybe the integral is a sum of an elementary part and a
# nonelementary part (like erf(x) + exp(x)). risch_integrate() is
# quite fast, so this is acceptable.
parts = []
args = Add.make_args(f)
for g in args:
coeff, g = g.as_independent(x)
# g(x) = const
if g is S.One and not meijerg:
parts.append(coeff * x)
continue
# g(x) = expr + O(x**n)
order_term = g.getO()
if order_term is not None:
h = self._eval_integral(g.removeO(), x)
if h is not None:
parts.append(coeff *
(h + self.func(order_term, *self.limits)))
continue
# NOTE: if there is O(x**n) and we fail to integrate then there is
# no point in trying other methods because they will fail anyway.
return
# c
# g(x) = (a*x+b)
if g.is_Pow and not g.exp.has(x) and not meijerg:
a = Wild('a', exclude=[x])
b = Wild('b', exclude=[x])
M = g.base.match(a * x + b)
if M is not None:
if g.exp == -1:
h = log(g.base)
elif conds != 'piecewise':
h = g.base**(g.exp + 1) / (g.exp + 1)
else:
h1 = log(g.base)
h2 = g.base**(g.exp + 1) / (g.exp + 1)
h = Piecewise((h1, Eq(g.exp, -1)), (h2, True))
parts.append(coeff * h / M[a])
continue
# poly(x)
# g(x) = -------
# poly(x)
if g.is_rational_function(x) and not meijerg:
parts.append(coeff * ratint(g, x))
continue
if not meijerg:
# g(x) = Mul(trig)
h = trigintegrate(g, x, conds=conds)
if h is not None:
parts.append(coeff * h)
continue
# g(x) has at least a DiracDelta term
h = deltaintegrate(g, x)
if h is not None:
parts.append(coeff * h)
continue
# Try risch again.
if risch is not False:
try:
h, i = risch_integrate(g,
x,
separate_integral=True,
conds=conds)
except NotImplementedError:
h = None
else:
if i:
h = h + i.doit(risch=False)
parts.append(coeff * h)
continue
# fall back to heurisch
try:
if conds == 'piecewise':
h = heurisch_wrapper(g, x, hints=[])
else:
h = heurisch(g, x, hints=[])
except PolynomialError:
# XXX: this exception means there is a bug in the
# implementation of heuristic Risch integration
# algorithm.
h = None
else:
h = None
if meijerg is not False and h is None:
# rewrite using G functions
try:
h = meijerint_indefinite(g, x)
except NotImplementedError:
from diofant.integrals.meijerint import _debug
_debug('NotImplementedError from meijerint_definite')
res = None
if h is not None:
parts.append(coeff * h)
continue
# if we failed maybe it was because we had
# a product that could have been expanded,
# so let's try an expansion of the whole
# thing before giving up; we don't try this
# at the outset because there are things
# that cannot be solved unless they are
# NOT expanded e.g., x**x*(1+log(x)). There
# should probably be a checker somewhere in this
# routine to look for such cases and try to do
# collection on the expressions if they are already
# in an expanded form
if not h and len(args) == 1:
f = f.expand(mul=True, deep=False)
if f.is_Add:
# Note: risch will be identical on the expanded
# expression, but maybe it will be able to pick out parts,
# like x*(exp(x) + erf(x)).
return self._eval_integral(f,
x,
meijerg=meijerg,
risch=risch,
conds=conds)
if h is not None:
parts.append(coeff * h)
else:
return
return Add(*parts)
def rsolve_poly(coeffs, f, n, **hints):
"""
Given linear recurrence operator `\operatorname{L}` of order
`k` with polynomial coefficients and inhomogeneous equation
`\operatorname{L} y = f`, where `f` is a polynomial, we seek for
all polynomial solutions over field `K` of characteristic zero.
The algorithm performs two basic steps:
(1) Compute degree `N` of the general polynomial solution.
(2) Find all polynomials of degree `N` or less
of `\operatorname{L} y = f`.
There are two methods for computing the polynomial solutions.
If the degree bound is relatively small, i.e. it's smaller than
or equal to the order of the recurrence, then naive method of
undetermined coefficients is being used. This gives system
of algebraic equations with `N+1` unknowns.
In the other case, the algorithm performs transformation of the
initial equation to an equivalent one, for which the system of
algebraic equations has only `r` indeterminates. This method is
quite sophisticated (in comparison with the naive one) and was
invented together by Abramov, Bronstein and Petkovšek.
It is possible to generalize the algorithm implemented here to
the case of linear q-difference and differential equations.
Lets say that we would like to compute `m`-th Bernoulli polynomial
up to a constant. For this we can use `b(n+1) - b(n) = m n^{m-1}`
recurrence, which has solution `b(n) = B_m + C`. For example:
>>> from diofant import Symbol, rsolve_poly
>>> n = Symbol('n', integer=True)
>>> rsolve_poly([-1, 1], 4*n**3, n)
C0 + n**4 - 2*n**3 + n**2
References
==========
.. [1] S. A. Abramov, M. Bronstein and M. Petkovšek, On polynomial
solutions of linear operator equations, in: T. Levelt, ed.,
Proc. ISSAC '95, ACM Press, New York, 1995, 290-296.
.. [2] M. Petkovšek, Hypergeometric solutions of linear recurrences
with polynomial coefficients, J. Symbolic Computation,
14 (1992), 243-264.
.. [3] M. Petkovšek, H. S. Wilf, D. Zeilberger, A = B, 1996.
"""
f = sympify(f)
if not f.is_polynomial(n):
return
homogeneous = f.is_zero
r = len(coeffs) - 1
coeffs = [Poly(coeff, n) for coeff in coeffs]
g = gcd_list(coeffs + [f], n, polys=True)
if not g.is_ground:
coeffs = [quo(c, g, n, polys=False) for c in coeffs]
f = quo(f, g, n, polys=False)
polys = [Poly(0, n)] * (r + 1)
terms = [(S.Zero, S.NegativeInfinity)] * (r + 1)
for i in range(0, r + 1):
for j in range(i, r + 1):
polys[i] += coeffs[j] * binomial(j, i)
if not polys[i].is_zero:
(exp, ), coeff = polys[i].LT()
terms[i] = (coeff, exp)
d = b = terms[0][1]
for i in range(1, r + 1):
if terms[i][1] > d:
d = terms[i][1]
if terms[i][1] - i > b:
b = terms[i][1] - i
d, b = int(d), int(b)
x = Dummy('x')
degree_poly = S.Zero
for i in range(0, r + 1):
if terms[i][1] - i == b:
degree_poly += terms[i][0] * FallingFactorial(x, i)
nni_roots = list(
roots(degree_poly, x, filter='Z', predicate=lambda r: r >= 0).keys())
if nni_roots:
N = [max(nni_roots)]
else:
N = []
if homogeneous:
N += [-b - 1]
else:
N += [f.as_poly(n).degree() - b, -b - 1]
N = int(max(N))
if N < 0:
if homogeneous:
if hints.get('symbols', False):
return S.Zero, []
else:
return S.Zero
else:
return
if N <= r:
C = []
y = E = S.Zero
for i in range(0, N + 1):
C.append(Symbol('C' + str(i)))
y += C[i] * n**i
for i in range(0, r + 1):
E += coeffs[i].as_expr() * y.subs(n, n + i)
solutions = solve_undetermined_coeffs(E - f, C, n)
if solutions is not None:
C = [c for c in C if (c not in solutions)]
result = y.subs(solutions)
else:
return # TBD
else:
A = r
U = N + A + b + 1
nni_roots = list(
roots(polys[r], filter='Z', predicate=lambda r: r >= 0).keys())
if nni_roots != []:
a = max(nni_roots) + 1
else:
a = S.Zero
def _zero_vector(k):
return [S.Zero] * k
def _one_vector(k):
return [S.One] * k
def _delta(p, k):
B = S.One
D = p.subs(n, a + k)
for i in range(1, k + 1):
B *= -Rational(k - i + 1, i)
D += B * p.subs(n, a + k - i)
return D
alpha = {}
for i in range(-A, d + 1):
I = _one_vector(d + 1)
for k in range(1, d + 1):
I[k] = I[k - 1] * (x + i - k + 1) / k
alpha[i] = S.Zero
for j in range(0, A + 1):
for k in range(0, d + 1):
B = binomial(k, i + j)
D = _delta(polys[j].as_expr(), k)
alpha[i] += I[k] * B * D
V = Matrix(U, A, lambda i, j: int(i == j))
if homogeneous:
for i in range(A, U):
v = _zero_vector(A)
for k in range(1, A + b + 1):
if i - k < 0:
break
B = alpha[k - A].subs(x, i - k)
for j in range(0, A):
v[j] += B * V[i - k, j]
denom = alpha[-A].subs(x, i)
for j in range(0, A):
V[i, j] = -v[j] / denom
else:
G = _zero_vector(U)
for i in range(A, U):
v = _zero_vector(A)
g = S.Zero
for k in range(1, A + b + 1):
if i - k < 0:
break
B = alpha[k - A].subs(x, i - k)
for j in range(0, A):
v[j] += B * V[i - k, j]
g += B * G[i - k]
denom = alpha[-A].subs(x, i)
for j in range(0, A):
V[i, j] = -v[j] / denom
G[i] = (_delta(f, i - A) - g) / denom
P, Q = _one_vector(U), _zero_vector(A)
for i in range(1, U):
P[i] = (P[i - 1] * (n - a - i + 1) / i).expand()
for i in range(0, A):
Q[i] = Add(*[(v * p).expand() for v, p in zip(V[:, i], P)])
if not homogeneous:
h = Add(*[(g * p).expand() for g, p in zip(G, P)])
C = [Symbol('C' + str(i)) for i in range(0, A)]
def g(i):
return Add(*[c * _delta(q, i) for c, q in zip(C, Q)])
if homogeneous:
E = [g(i) for i in range(N + 1, U)]
else:
E = [g(i) + _delta(h, i) for i in range(N + 1, U)]
if E != []:
solutions = solve(E, *C)
if not solutions:
if homogeneous:
if hints.get('symbols', False):
return S.Zero, []
else:
return S.Zero
else:
return
else:
solutions = {}
if homogeneous:
result = S.Zero
else:
result = h
for c, q in list(zip(C, Q)):
if c in solutions:
s = solutions[c] * q
C.remove(c)
else:
s = c * q
result += s.expand()
if hints.get('symbols', False):
return result, C
else:
return result
def rsolve_hyper(coeffs, f, n, **hints):
"""
Given linear recurrence operator `\operatorname{L}` of order `k`
with polynomial coefficients and inhomogeneous equation
`\operatorname{L} y = f` we seek for all hypergeometric solutions
over field `K` of characteristic zero.
The inhomogeneous part can be either hypergeometric or a sum
of a fixed number of pairwise dissimilar hypergeometric terms.
The algorithm performs three basic steps:
(1) Group together similar hypergeometric terms in the
inhomogeneous part of `\operatorname{L} y = f`, and find
particular solution using Abramov's algorithm.
(2) Compute generating set of `\operatorname{L}` and find basis
in it, so that all solutions are linearly independent.
(3) Form final solution with the number of arbitrary
constants equal to dimension of basis of `\operatorname{L}`.
Term `a(n)` is hypergeometric if it is annihilated by first order
linear difference equations with polynomial coefficients or, in
simpler words, if consecutive term ratio is a rational function.
The output of this procedure is a linear combination of fixed
number of hypergeometric terms. However the underlying method
can generate larger class of solutions - D'Alembertian terms.
Note also that this method not only computes the kernel of the
inhomogeneous equation, but also reduces in to a basis so that
solutions generated by this procedure are linearly independent
Examples
========
>>> from diofant.solvers import rsolve_hyper
>>> from diofant.abc import x
>>> rsolve_hyper([-1, -1, 1], 0, x)
C0*(1/2 + sqrt(5)/2)**x + C1*(-sqrt(5)/2 + 1/2)**x
>>> rsolve_hyper([-1, 1], 1 + x, x)
C0 + x*(x + 1)/2
References
==========
.. [1] M. Petkovšek, Hypergeometric solutions of linear recurrences
with polynomial coefficients, J. Symbolic Computation,
14 (1992), 243-264.
.. [2] M. Petkovšek, H. S. Wilf, D. Zeilberger, A = B, 1996.
"""
coeffs = list(map(sympify, coeffs))
f = sympify(f)
r, kernel, symbols = len(coeffs) - 1, [], set()
if not f.is_zero:
if f.is_Add:
similar = {}
for g in f.expand().args:
if not g.is_hypergeometric(n):
return
for h in similar.keys():
if hypersimilar(g, h, n):
similar[h] += g
break
else:
similar[g] = S.Zero
inhomogeneous = []
for g, h in similar.items():
inhomogeneous.append(g + h)
elif f.is_hypergeometric(n):
inhomogeneous = [f]
else:
return
for i, g in enumerate(inhomogeneous):
coeff, polys = S.One, coeffs[:]
denoms = [S.One] * (r + 1)
s = hypersimp(g, n)
for j in range(1, r + 1):
coeff *= s.subs(n, n + j - 1)
p, q = coeff.as_numer_denom()
polys[j] *= p
denoms[j] = q
for j in range(0, r + 1):
polys[j] *= Mul(*(denoms[:j] + denoms[j + 1:]))
R = rsolve_ratio(polys, Mul(*denoms), n, symbols=True)
if R is not None:
R, syms = R
if syms:
R = R.subs(zip(syms, [0] * len(syms)))
if R:
inhomogeneous[i] *= R
else:
return
result = Add(*inhomogeneous)
result = simplify(result)
else:
result = S.Zero
Z = Dummy('Z')
p, q = coeffs[0], coeffs[r].subs(n, n - r + 1)
p_factors = [z for z in roots(p, n).keys()]
q_factors = [z for z in roots(q, n).keys()]
factors = [(S.One, S.One)]
for p in p_factors:
for q in q_factors:
if p.is_integer and q.is_integer and p <= q:
continue
else:
factors += [(n - p, n - q)]
p = [(n - p, S.One) for p in p_factors]
q = [(S.One, n - q) for q in q_factors]
factors = p + factors + q
for A, B in factors:
polys, degrees = [], []
D = A * B.subs(n, n + r - 1)
for i in range(0, r + 1):
a = Mul(*[A.subs(n, n + j) for j in range(0, i)])
b = Mul(*[B.subs(n, n + j) for j in range(i, r)])
poly = quo(coeffs[i] * a * b, D, n)
polys.append(poly.as_poly(n))
if not poly.is_zero:
degrees.append(polys[i].degree())
d, poly = max(degrees), S.Zero
for i in range(0, r + 1):
coeff = polys[i].nth(d)
if coeff is not S.Zero:
poly += coeff * Z**i
for z in roots(poly, Z).keys():
if z.is_zero:
continue
(C, s) = rsolve_poly([polys[i] * z**i for i in range(r + 1)],
0,
n,
symbols=True)
if C is not None and C is not S.Zero:
symbols |= set(s)
ratio = z * A * C.subs(n, n + 1) / B / C
ratio = simplify(ratio)
skip = max([-1] + [
v for v in roots(Mul(*ratio.as_numer_denom()), n).keys()
if v.is_Integer
]) + 1
K = product(ratio, (n, skip, n - 1))
if K.has(factorial, FallingFactorial, RisingFactorial):
K = simplify(K)
if casoratian(kernel + [K], n, zero=False) != 0:
kernel.append(K)
kernel.sort(key=default_sort_key)
sk = list(zip(numbered_symbols('C'), kernel))
for C, ker in sk:
result += C * ker
if hints.get('symbols', False):
symbols |= {s for s, k in sk}
return result, list(symbols)
else:
return result
def g(i):
return Add(*[c * _delta(q, i) for c, q in zip(C, Q)])
def _integrate(field=None):
irreducibles = set()
for poly in reducibles:
for z in poly.free_symbols:
if z in V:
break # should this be: `irreducibles |= \
else: # set(root_factors(poly, z, filter=field))`
continue # and the line below deleted?
# |
# V
irreducibles |= set(root_factors(poly, z, filter=field))
log_coeffs, log_part = [], []
B = _symbols('B', len(irreducibles))
# Note: the ordering matters here
for poly, b in reversed(list(ordered(zip(irreducibles, B)))):
if poly.has(*V):
poly_coeffs.append(b)
log_part.append(b * log(poly))
# TODO: Currently it's better to use symbolic expressions here instead
# of rational functions, because it's simpler and FracElement doesn't
# give big speed improvement yet. This is because cancelation is slow
# due to slow polynomial GCD algorithms. If this gets improved then
# revise this code.
candidate = poly_part / poly_denom + Add(*log_part)
h = F - _derivation(candidate) / denom
raw_numer = h.as_numer_denom()[0]
# Rewrite raw_numer as a polynomial in K[coeffs][V] where K is a field
# that we have to determine. We can't use simply atoms() because log(3),
# sqrt(y) and similar expressions can appear, leading to non-trivial
# domains.
syms = set(poly_coeffs) | set(V)
non_syms = set()
def find_non_syms(expr):
if expr.is_Integer or expr.is_Rational:
pass # ignore trivial numbers
elif expr in syms:
pass # ignore variables
elif not expr.has(*syms):
non_syms.add(expr)
elif expr.is_Add or expr.is_Mul or expr.is_Pow:
list(map(find_non_syms, expr.args))
else:
# TODO: Non-polynomial expression. This should have been
# filtered out at an earlier stage.
raise PolynomialError
try:
find_non_syms(raw_numer)
except PolynomialError:
return
else:
ground, _ = construct_domain(non_syms, field=True)
coeff_ring = PolyRing(poly_coeffs, ground)
ring = PolyRing(V, coeff_ring)
numer = ring.from_expr(raw_numer)
solution = solve_lin_sys(numer.coeffs(), coeff_ring)
if solution is None:
return
else:
solution = [(coeff_ring.symbols[coeff_ring.index(k)], v.as_expr())
for k, v in solution.items()]
return candidate.subs(solution).subs(
list(zip(poly_coeffs, [S.Zero] * len(poly_coeffs))))
def _derivation(h):
return Add(*[d * h.diff(v) for d, v in zip(numers, V)])
def heurisch(f,
x,
rewrite=False,
hints=None,
mappings=None,
retries=3,
degree_offset=0,
unnecessary_permutations=None):
"""
Compute indefinite integral using heuristic Risch algorithm.
This is a heuristic approach to indefinite integration in finite
terms using the extended heuristic (parallel) Risch algorithm, based
on Manuel Bronstein's "Poor Man's Integrator" [1]_.
The algorithm supports various classes of functions including
transcendental elementary or special functions like Airy,
Bessel, Whittaker and Lambert.
Note that this algorithm is not a decision procedure. If it isn't
able to compute the antiderivative for a given function, then this is
not a proof that such a functions does not exist. One should use
recursive Risch algorithm in such case. It's an open question if
this algorithm can be made a full decision procedure.
This is an internal integrator procedure. You should use toplevel
'integrate' function in most cases, as this procedure needs some
preprocessing steps and otherwise may fail.
Parameters
==========
heurisch(f, x, rewrite=False, hints=None)
f : Expr
expression
x : Symbol
variable
rewrite : Boolean, optional
force rewrite 'f' in terms of 'tan' and 'tanh', default False.
hints : None or list
a list of functions that may appear in anti-derivate. If
None (default) - no suggestions at all, if empty list - try
to figure out.
Examples
========
>>> from diofant import tan
>>> from diofant.integrals.heurisch import heurisch
>>> from diofant.abc import x, y
>>> heurisch(y*tan(x), x)
y*log(tan(x)**2 + 1)/2
References
==========
.. [1] Manuel Bronstein's "Poor Man's Integrator",
http://www-sop.inria.fr/cafe/Manuel.Bronstein/pmint/index.html
.. [2] K. Geddes, L. Stefanus, On the Risch-Norman Integration
Method and its Implementation in Maple, Proceedings of
ISSAC'89, ACM Press, 212-217.
.. [3] J. H. Davenport, On the Parallel Risch Algorithm (I),
Proceedings of EUROCAM'82, LNCS 144, Springer, 144-157.
.. [4] J. H. Davenport, On the Parallel Risch Algorithm (III):
Use of Tangents, SIGSAM Bulletin 16 (1982), 3-6.
.. [5] J. H. Davenport, B. M. Trager, On the Parallel Risch
Algorithm (II), ACM Transactions on Mathematical
Software 11 (1985), 356-362.
See Also
========
diofant.integrals.integrals.Integral.doit
diofant.integrals.integrals.Integral
diofant.integrals.heurisch.components
"""
f = sympify(f)
if x not in f.free_symbols:
return f * x
if not f.is_Add:
indep, f = f.as_independent(x)
else:
indep = S.One
rewritables = {
(sin, cos, cot): tan,
(sinh, cosh, coth): tanh,
}
if rewrite:
for candidates, rule in rewritables.items():
f = f.rewrite(candidates, rule)
else:
for candidates in rewritables.keys():
if f.has(*candidates):
break
else:
rewrite = True
terms = components(f, x)
if hints is not None:
if not hints:
a = Wild('a', exclude=[x])
b = Wild('b', exclude=[x])
c = Wild('c', exclude=[x])
for g in set(terms): # using copy of terms
if g.is_Function:
if g.func is li:
M = g.args[0].match(a * x**b)
if M is not None:
terms.add(
x *
(li(M[a] * x**M[b]) -
(M[a] * x**M[b])**(-1 / M[b]) * Ei(
(M[b] + 1) * log(M[a] * x**M[b]) / M[b])))
# terms.add( x*(li(M[a]*x**M[b]) - (x**M[b])**(-1/M[b])*Ei((M[b]+1)*log(M[a]*x**M[b])/M[b])) )
# terms.add( x*(li(M[a]*x**M[b]) - x*Ei((M[b]+1)*log(M[a]*x**M[b])/M[b])) )
# terms.add( li(M[a]*x**M[b]) - Ei((M[b]+1)*log(M[a]*x**M[b])/M[b]) )
elif g.is_Pow:
if g.base is S.Exp1:
M = g.exp.match(a * x**2)
if M is not None:
if M[a].is_positive:
terms.add(erfi(sqrt(M[a]) * x))
else: # M[a].is_negative or unknown
terms.add(erf(sqrt(-M[a]) * x))
M = g.exp.match(a * x**2 + b * x + c)
if M is not None:
if M[a].is_positive:
terms.add(
sqrt(pi / 4 * (-M[a])) *
exp(M[c] - M[b]**2 / (4 * M[a])) * erfi(
sqrt(M[a]) * x + M[b] /
(2 * sqrt(M[a]))))
elif M[a].is_negative:
terms.add(
sqrt(pi / 4 * (-M[a])) *
exp(M[c] - M[b]**2 / (4 * M[a])) * erf(
sqrt(-M[a]) * x - M[b] /
(2 * sqrt(-M[a]))))
M = g.exp.match(a * log(x)**2)
if M is not None:
if M[a].is_positive:
terms.add(
erfi(
sqrt(M[a]) * log(x) + 1 /
(2 * sqrt(M[a]))))
if M[a].is_negative:
terms.add(
erf(
sqrt(-M[a]) * log(x) - 1 /
(2 * sqrt(-M[a]))))
elif g.exp.is_Rational and g.exp.q == 2:
M = g.base.match(a * x**2 + b)
if M is not None and M[b].is_positive:
if M[a].is_positive:
terms.add(asinh(sqrt(M[a] / M[b]) * x))
elif M[a].is_negative:
terms.add(asin(sqrt(-M[a] / M[b]) * x))
M = g.base.match(a * x**2 - b)
if M is not None and M[b].is_positive:
if M[a].is_positive:
terms.add(acosh(sqrt(M[a] / M[b]) * x))
elif M[a].is_negative:
terms.add((-M[b] / 2 * sqrt(-M[a]) * atan(
sqrt(-M[a]) * x / sqrt(M[a] * x**2 - M[b]))
))
else:
terms |= set(hints)
for g in set(terms): # using copy of terms
terms |= components(cancel(g.diff(x)), x)
# TODO: caching is significant factor for why permutations work at all. Change this.
V = _symbols('x', len(terms))
# sort mapping expressions from largest to smallest (last is always x).
mapping = list(
reversed(
list(
zip(*ordered( #
[(a[0].as_independent(x)[1], a)
for a in zip(terms, V)])))[1])) #
rev_mapping = {v: k for k, v in mapping} #
if mappings is None: #
# optimizing the number of permutations of mapping #
assert mapping[-1][0] == x # if not, find it and correct this comment
unnecessary_permutations = [mapping.pop(-1)]
mappings = permutations(mapping)
else:
unnecessary_permutations = unnecessary_permutations or []
def _substitute(expr):
return expr.subs(mapping)
for mapping in mappings:
mapping = list(mapping)
mapping = mapping + unnecessary_permutations
diffs = [_substitute(cancel(g.diff(x))) for g in terms]
denoms = [g.as_numer_denom()[1] for g in diffs]
if all(h.is_polynomial(*V)
for h in denoms) and _substitute(f).is_rational_function(*V):
denom = reduce(lambda p, q: lcm(p, q, *V), denoms)
break
else:
if not rewrite:
result = heurisch(
f,
x,
rewrite=True,
hints=hints,
unnecessary_permutations=unnecessary_permutations)
if result is not None:
return indep * result
return
numers = [cancel(denom * g) for g in diffs]
def _derivation(h):
return Add(*[d * h.diff(v) for d, v in zip(numers, V)])
def _deflation(p):
for y in V:
if not p.has(y):
continue
if _derivation(p) is not S.Zero:
c, q = p.as_poly(y).primitive()
return _deflation(c) * gcd(q, q.diff(y)).as_expr()
else:
return p
def _splitter(p):
for y in V:
if not p.has(y):
continue
if _derivation(y) is not S.Zero:
c, q = p.as_poly(y).primitive()
q = q.as_expr()
h = gcd(q, _derivation(q), y)
s = quo(h, gcd(q, q.diff(y), y), y)
c_split = _splitter(c)
if s.as_poly(y).degree() == 0:
return c_split[0], q * c_split[1]
q_split = _splitter(cancel(q / s))
return c_split[0] * q_split[0] * s, c_split[1] * q_split[1]
else:
return S.One, p
special = {}
for term in terms:
if term.is_Function:
if term.func is tan:
special[1 + _substitute(term)**2] = False
elif term.func is tanh:
special[1 + _substitute(term)] = False
special[1 - _substitute(term)] = False
elif term.func is LambertW:
special[_substitute(term)] = True
F = _substitute(f)
P, Q = F.as_numer_denom()
u_split = _splitter(denom)
v_split = _splitter(Q)
polys = set(list(v_split) + [u_split[0]] + list(special.keys()))
s = u_split[0] * Mul(*[k for k, v in special.items() if v])
polified = [p.as_poly(*V) for p in [s, P, Q]]
if None in polified:
return
# --- definitions for _integrate ---
a, b, c = [p.total_degree() for p in polified]
poly_denom = (s * v_split[0] * _deflation(v_split[1])).as_expr()
def _exponent(g):
if g.is_Pow:
if g.exp.is_Rational and g.exp.q != 1:
if g.exp.p > 0:
return g.exp.p + g.exp.q - 1
else:
return abs(g.exp.p + g.exp.q)
else:
return 1
elif not g.is_Atom and g.args:
return max([_exponent(h) for h in g.args])
else:
return 1
A, B = _exponent(f), a + max(b, c)
if A > 1 and B > 1:
monoms = itermonomials(V, A + B - 1 + degree_offset)
else:
monoms = itermonomials(V, A + B + degree_offset)
poly_coeffs = _symbols('A', len(monoms))
poly_part = Add(*[
poly_coeffs[i] * monomial for i, monomial in enumerate(ordered(monoms))
])
reducibles = set()
for poly in polys:
if poly.has(*V):
try:
factorization = factor(poly, greedy=True)
except PolynomialError:
factorization = poly
factorization = poly
if factorization.is_Mul:
reducibles |= set(factorization.args)
else:
reducibles.add(factorization)
def _integrate(field=None):
irreducibles = set()
for poly in reducibles:
for z in poly.free_symbols:
if z in V:
break # should this be: `irreducibles |= \
else: # set(root_factors(poly, z, filter=field))`
continue # and the line below deleted?
# |
# V
irreducibles |= set(root_factors(poly, z, filter=field))
log_coeffs, log_part = [], []
B = _symbols('B', len(irreducibles))
# Note: the ordering matters here
for poly, b in reversed(list(ordered(zip(irreducibles, B)))):
if poly.has(*V):
poly_coeffs.append(b)
log_part.append(b * log(poly))
# TODO: Currently it's better to use symbolic expressions here instead
# of rational functions, because it's simpler and FracElement doesn't
# give big speed improvement yet. This is because cancelation is slow
# due to slow polynomial GCD algorithms. If this gets improved then
# revise this code.
candidate = poly_part / poly_denom + Add(*log_part)
h = F - _derivation(candidate) / denom
raw_numer = h.as_numer_denom()[0]
# Rewrite raw_numer as a polynomial in K[coeffs][V] where K is a field
# that we have to determine. We can't use simply atoms() because log(3),
# sqrt(y) and similar expressions can appear, leading to non-trivial
# domains.
syms = set(poly_coeffs) | set(V)
non_syms = set()
def find_non_syms(expr):
if expr.is_Integer or expr.is_Rational:
pass # ignore trivial numbers
elif expr in syms:
pass # ignore variables
elif not expr.has(*syms):
non_syms.add(expr)
elif expr.is_Add or expr.is_Mul or expr.is_Pow:
list(map(find_non_syms, expr.args))
else:
# TODO: Non-polynomial expression. This should have been
# filtered out at an earlier stage.
raise PolynomialError
try:
find_non_syms(raw_numer)
except PolynomialError:
return
else:
ground, _ = construct_domain(non_syms, field=True)
coeff_ring = PolyRing(poly_coeffs, ground)
ring = PolyRing(V, coeff_ring)
numer = ring.from_expr(raw_numer)
solution = solve_lin_sys(numer.coeffs(), coeff_ring)
if solution is None:
return
else:
solution = [(coeff_ring.symbols[coeff_ring.index(k)], v.as_expr())
for k, v in solution.items()]
return candidate.subs(solution).subs(
list(zip(poly_coeffs, [S.Zero] * len(poly_coeffs))))
if not (F.free_symbols - set(V)):
solution = _integrate('Q')
if solution is None:
solution = _integrate()
else:
solution = _integrate()
if solution is not None:
antideriv = solution.subs(rev_mapping)
antideriv = cancel(antideriv).expand(force=True)
if antideriv.is_Add:
antideriv = antideriv.as_independent(x)[1]
return indep * antideriv
else:
if retries >= 0:
result = heurisch(
f,
x,
mappings=mappings,
rewrite=rewrite,
hints=hints,
retries=retries - 1,
unnecessary_permutations=unnecessary_permutations)
if result is not None:
return indep * result
return
def dot(self, p2):
"""Return dot product of self with another Point."""
p2 = Point(p2)
return Add(*[a*b for a, b in zip(self, p2)])
def doit(self, **hints):
"""
Perform the integration using any hints given.
Examples
========
>>> from diofant import Integral
>>> from diofant.abc import x, i
>>> Integral(x**i, (i, 1, 3)).doit()
Piecewise((2, Eq(log(x), 0)), (x**3/log(x) - x/log(x), true))
See Also
========
diofant.integrals.trigonometry.trigintegrate
diofant.integrals.heurisch.heurisch
diofant.integrals.rationaltools.ratint
diofant.integrals.integrals.Integral.as_sum : Approximate the integral using a sum
"""
if not hints.get('integrals', True):
return self
deep = hints.get('deep', True)
meijerg = hints.get('meijerg', None)
conds = hints.get('conds', 'piecewise')
risch = hints.get('risch', None)
if conds not in ['separate', 'piecewise', 'none']:
raise ValueError('conds must be one of "separate", "piecewise", '
'"none", got: %s' % conds)
if risch and any(len(xab) > 1 for xab in self.limits):
raise ValueError(
'risch=True is only allowed for indefinite integrals.')
# check for the trivial zero
if self.is_zero:
return S.Zero
# now compute and check the function
function = self.function
if isinstance(function, MatrixBase):
return function.applyfunc(
lambda f: self.func(f, self.limits).doit(**hints))
if deep:
function = function.doit(**hints)
if function.is_zero:
return S.Zero
# There is no trivial answer, so continue
undone_limits = []
# ulj = free symbols of any undone limits' upper and lower limits
ulj = set()
for xab in self.limits:
# compute uli, the free symbols in the
# Upper and Lower limits of limit I
if len(xab) == 1:
uli = set(xab[:1])
elif len(xab) == 2:
uli = xab[1].free_symbols
elif len(xab) == 3:
uli = xab[1].free_symbols.union(xab[2].free_symbols)
# this integral can be done as long as there is no blocking
# limit that has been undone. An undone limit is blocking if
# it contains an integration variable that is in this limit's
# upper or lower free symbols or vice versa
if xab[0] in ulj or any(v[0] in uli for v in undone_limits):
undone_limits.append(xab)
ulj.update(uli)
function = self.func(*([function] + [xab]))
factored_function = function.factor()
if not isinstance(factored_function, Integral):
function = factored_function
continue
# There are a number of tradeoffs in using the Meijer G method.
# It can sometimes be a lot faster than other methods, and
# sometimes slower. And there are certain types of integrals for
# which it is more likely to work than others.
# These heuristics are incorporated in deciding what integration
# methods to try, in what order.
# See the integrate() docstring for details.
def try_meijerg(function, xab):
ret = None
if len(xab) == 3 and meijerg is not False:
x, a, b = xab
try:
res = meijerint_definite(function, x, a, b)
except NotImplementedError:
from diofant.integrals.meijerint import _debug
_debug('NotImplementedError from meijerint_definite')
res = None
if res is not None:
f, cond = res
if conds == 'piecewise':
ret = Piecewise((f, cond),
(self.func(function,
(x, a, b)), True))
elif conds == 'separate':
if len(self.limits) != 1:
raise ValueError(
'conds=separate not supported in '
'multiple integrals')
ret = f, cond
else:
ret = f
return ret
meijerg1 = meijerg
if len(xab) == 3 and xab[1].is_extended_real and xab[2].is_extended_real \
and not function.is_Poly and \
(xab[1].has(oo, -oo) or xab[2].has(oo, -oo)):
ret = try_meijerg(function, xab)
if ret is not None:
function = ret
continue
else:
meijerg1 = False
# If the special meijerg code did not succeed in finding a definite
# integral, then the code using meijerint_indefinite will not either
# (it might find an antiderivative, but the answer is likely to be
# nonsensical).
# Thus if we are requested to only use Meijer G-function methods,
# we give up at this stage. Otherwise we just disable G-function
# methods.
if meijerg1 is False and meijerg is True:
antideriv = None
else:
antideriv = self._eval_integral(function,
xab[0],
meijerg=meijerg1,
risch=risch,
conds=conds)
if antideriv is None and meijerg1 is True:
ret = try_meijerg(function, xab)
if ret is not None:
function = ret
continue
if antideriv is None:
undone_limits.append(xab)
function = self.func(*([function] + [xab])).factor()
factored_function = function.factor()
if not isinstance(factored_function, Integral):
function = factored_function
continue
else:
if len(xab) == 1:
function = antideriv
else:
if len(xab) == 3:
x, a, b = xab
elif len(xab) == 2:
x, b = xab
a = None
else:
raise NotImplementedError
if deep:
if isinstance(a, Basic):
a = a.doit(**hints)
if isinstance(b, Basic):
b = b.doit(**hints)
if antideriv.is_Poly:
gens = list(antideriv.gens)
gens.remove(x)
antideriv = antideriv.as_expr()
function = antideriv._eval_interval(x, a, b)
function = Poly(function, *gens)
elif (isinstance(antideriv, Add) and any(
isinstance(t, Integral) for t in antideriv.args)):
function = Add(*[
i._eval_interval(x, a, b)
for i in Add.make_args(antideriv)
])
else:
try:
function = antideriv._eval_interval(x, a, b)
except NotImplementedError:
# This can happen if _eval_interval depends in a
# complicated way on limits that cannot be computed
undone_limits.append(xab)
function = self.func(*([function] + [xab]))
factored_function = function.factor()
if not isinstance(factored_function, Integral):
function = factored_function
return function
def test_equality(r, alg="Greedy"):
return r == Add(*[Rational(1, i) for i in egyptian_fraction(r, alg)])
def add_terms(terms, prec, target_prec):
"""
Helper for evalf_add. Adds a list of (mpfval, accuracy) terms.
Returns
-------
- None, None if there are no non-zero terms;
- terms[0] if there is only 1 term;
- scaled_zero if the sum of the terms produces a zero by cancellation
e.g. mpfs representing 1 and -1 would produce a scaled zero which need
special handling since they are not actually zero and they are purposely
malformed to ensure that they can't be used in anything but accuracy
calculations;
- a tuple that is scaled to target_prec that corresponds to the
sum of the terms.
The returned mpf tuple will be normalized to target_prec; the input
prec is used to define the working precision.
XXX explain why this is needed and why one can't just loop using mpf_add
"""
terms = [t for t in terms if not iszero(t)]
if not terms:
return None, None
elif len(terms) == 1:
return terms[0]
# see if any argument is NaN or oo and thus warrants a special return
special = []
from diofant.core.numbers import Float
for t in terms:
arg = Float._new(t[0], 1)
if arg is S.NaN or arg.is_infinite:
special.append(arg)
if special:
from diofant.core.add import Add
rv = evalf(Add(*special), prec + 4, {})
return rv[0], rv[2]
working_prec = 2 * prec
sum_man, sum_exp, absolute_error = 0, 0, MINUS_INF
for x, accuracy in terms:
sign, man, exp, bc = x
if sign:
man = -man
absolute_error = max(absolute_error, bc + exp - accuracy)
delta = exp - sum_exp
if exp >= sum_exp:
# x much larger than existing sum?
# first: quick test
if ((delta > working_prec)
and ((not sum_man)
or delta - bitcount(abs(sum_man)) > working_prec)):
sum_man = man
sum_exp = exp
else:
sum_man += (man << delta)
else:
delta = -delta
# x much smaller than existing sum?
if delta - bc > working_prec:
if not sum_man:
sum_man, sum_exp = man, exp
else:
sum_man = (sum_man << delta) + man
sum_exp = exp
if not sum_man:
return scaled_zero(absolute_error)
if sum_man < 0:
sum_sign = 1
sum_man = -sum_man
else:
sum_sign = 0
sum_bc = bitcount(sum_man)
sum_accuracy = sum_exp + sum_bc - absolute_error
r = normalize(sum_sign, sum_man, sum_exp, sum_bc, target_prec,
rnd), sum_accuracy
return r