def _eval_evalf(self, prec=15, **options):
"""Evaluate the coordinates of the point.
This method will, where possible, create and return a new Point
where the coordinates are evaluated as floating point numbers to
the precision indicated (default=15).
Parameters
==========
prec : int
Returns
=======
point : Point
Examples
========
>>> from sympy import Point, Rational
>>> p1 = Point(Rational(1, 2), Rational(3, 2))
>>> p1
Point2D(1/2, 3/2)
>>> p1.evalf()
Point2D(0.5, 1.5)
"""
dps = prec_to_dps(prec)
coords = [x.evalf(n=dps, **options) for x in self.args]
return Point(*coords, evaluate=False)
def _eval_evalf(self, prec):
try:
_roots = self.poly.nroots(n=prec_to_dps(prec))
except (DomainError, PolynomialError):
return self
else:
return Add(*[self.fun(r) for r in _roots])
def _eval_evalf(self, prec):
try:
_roots = self.poly.nroots(n=prec_to_dps(prec))
except (DomainError, PolynomialError):
return self
else:
return Add(*[self.fun(r) for r in _roots])
def _eval_evalf(self, prec):
if not self.args:
return self
new_args = []
dps = prec_to_dps(prec)
for mat, frame in self.args:
new_args.append([mat.evalf(n=dps), frame])
return Vector(new_args)
def _eval_evalf(self, prec):
if not self.args:
return self
new_args = []
dps = prec_to_dps(prec)
for inlist in self.args:
new_inlist = list(inlist)
new_inlist[0] = inlist[0].evalf(n=dps)
new_args.append(tuple(new_inlist))
return Dyadic(new_args)
def _eigenvals_eigenvects_mpmath(M):
norm2 = lambda v: mp.sqrt(sum(i**2 for i in v))
v1 = None
prec = max([x._prec for x in M.atoms(Float)])
eps = 2**-prec
while prec < DEFAULT_MAXPREC:
with workprec(prec):
A = mp.matrix(M.evalf(n=prec_to_dps(prec)))
E, ER = mp.eig(A)
v2 = norm2([i for e in E for i in (mp.re(e), mp.im(e))])
if v1 is not None and mp.fabs(v1 - v2) < eps:
return E, ER
v1 = v2
prec *= 2
# we get here because the next step would have taken us
# past MAXPREC or because we never took a step; in case
# of the latter, we refuse to send back a solution since
# it would not have been verified; we also resist taking
# a small step to arrive exactly at MAXPREC since then
# the two calculations might be artificially close.
raise PrecisionExhausted
def _eval_evalf(self, prec):
"""Returns the floating point approximations (decimal numbers) of the quaternion.
Returns
=======
Quaternion
Floating point approximations of quaternion(self)
Examples
========
>>> from sympy import Quaternion
>>> from sympy import sqrt
>>> q = Quaternion(1/sqrt(1), 1/sqrt(2), 1/sqrt(3), 1/sqrt(4))
>>> q.evalf()
1.00000000000000
+ 0.707106781186547*i
+ 0.577350269189626*j
+ 0.500000000000000*k
"""
nprec = prec_to_dps(prec)
return Quaternion(*[arg.evalf(n=nprec) for arg in self.args])
def hypsum(expr, n, start, prec):
"""
Sum a rapidly convergent infinite hypergeometric series with
given general term, e.g. e = hypsum(1/factorial(n), n). The
quotient between successive terms must be a quotient of integer
polynomials.
"""
from sympy import Float, hypersimp, lambdify
if prec == float('inf'):
raise NotImplementedError('does not support inf prec')
if start:
expr = expr.subs(n, n + start)
hs = hypersimp(expr, n)
if hs is None:
raise NotImplementedError("a hypergeometric series is required")
num, den = hs.as_numer_denom()
func1 = lambdify(n, num)
func2 = lambdify(n, den)
h, g, p = check_convergence(num, den, n)
if h < 0:
raise ValueError("Sum diverges like (n!)^%i" % (-h))
term = expr.subs(n, 0)
if not term.is_Rational:
raise NotImplementedError(
"Non rational term functionality is not implemented.")
# Direct summation if geometric or faster
if h > 0 or (h == 0 and abs(g) > 1):
term = (MPZ(term.p) << prec) // term.q
s = term
k = 1
while abs(term) > 5:
term *= MPZ(func1(k - 1))
term //= MPZ(func2(k - 1))
s += term
k += 1
return from_man_exp(s, -prec)
else:
alt = g < 0
if abs(g) < 1:
raise ValueError("Sum diverges like (%i)^n" % abs(1 / g))
if p < 1 or (p == 1 and not alt):
raise ValueError("Sum diverges like n^%i" % (-p))
# We have polynomial convergence: use Richardson extrapolation
vold = None
ndig = prec_to_dps(prec)
while True:
# Need to use at least quad precision because a lot of cancellation
# might occur in the extrapolation process; we check the answer to
# make sure that the desired precision has been reached, too.
prec2 = 4 * prec
term0 = (MPZ(term.p) << prec2) // term.q
def summand(k, _term=[term0]):
if k:
k = int(k)
_term[0] *= MPZ(func1(k - 1))
_term[0] //= MPZ(func2(k - 1))
return make_mpf(from_man_exp(_term[0], -prec2))
with workprec(prec):
v = nsum(summand, [0, mpmath_inf], method='richardson')
vf = Float(v, ndig)
if vold is not None and vold == vf:
break
prec += prec # double precision each time
vold = vf
return v._mpf_
def _jordan_form(M, calc_transform=True, *, chop=False):
"""Return $(P, J)$ where $J$ is a Jordan block
matrix and $P$ is a matrix such that $M = P J P^{-1}$
Parameters
==========
calc_transform : bool
If ``False``, then only $J$ is returned.
chop : bool
All matrices are converted to exact types when computing
eigenvalues and eigenvectors. As a result, there may be
approximation errors. If ``chop==True``, these errors
will be truncated.
Examples
========
>>> from sympy.matrices import Matrix
>>> M = Matrix([[ 6, 5, -2, -3], [-3, -1, 3, 3], [ 2, 1, -2, -3], [-1, 1, 5, 5]])
>>> P, J = M.jordan_form()
>>> J
Matrix([
[2, 1, 0, 0],
[0, 2, 0, 0],
[0, 0, 2, 1],
[0, 0, 0, 2]])
See Also
========
jordan_block
"""
if not M.is_square:
raise NonSquareMatrixError("Only square matrices have Jordan forms")
mat = M
has_floats = M.has(Float)
if has_floats:
try:
max_prec = max(term._prec for term in M.values()
if isinstance(term, Float))
except ValueError:
# if no term in the matrix is explicitly a Float calling max()
# will throw a error so setting max_prec to default value of 53
max_prec = 53
# setting minimum max_dps to 15 to prevent loss of precision in
# matrix containing non evaluated expressions
max_dps = max(prec_to_dps(max_prec), 15)
def restore_floats(*args):
"""If ``has_floats`` is `True`, cast all ``args`` as
matrices of floats."""
if has_floats:
args = [m.evalf(n=max_dps, chop=chop) for m in args]
if len(args) == 1:
return args[0]
return args
# cache calculations for some speedup
mat_cache = {}
def eig_mat(val, pow):
"""Cache computations of ``(M - val*I)**pow`` for quick
retrieval"""
if (val, pow) in mat_cache:
return mat_cache[(val, pow)]
if (val, pow - 1) in mat_cache:
mat_cache[(val, pow)] = mat_cache[(val, pow - 1)].multiply(
mat_cache[(val, 1)], dotprodsimp=None)
else:
mat_cache[(val, pow)] = (mat - val * M.eye(M.rows)).pow(pow)
return mat_cache[(val, pow)]
# helper functions
def nullity_chain(val, algebraic_multiplicity):
"""Calculate the sequence [0, nullity(E), nullity(E**2), ...]
until it is constant where ``E = M - val*I``"""
# mat.rank() is faster than computing the null space,
# so use the rank-nullity theorem
cols = M.cols
ret = [0]
nullity = cols - eig_mat(val, 1).rank()
i = 2
while nullity != ret[-1]:
ret.append(nullity)
if nullity == algebraic_multiplicity:
break
nullity = cols - eig_mat(val, i).rank()
i += 1
# Due to issues like #7146 and #15872, SymPy sometimes
# gives the wrong rank. In this case, raise an error
# instead of returning an incorrect matrix
if nullity < ret[-1] or nullity > algebraic_multiplicity:
raise MatrixError("SymPy had encountered an inconsistent "
"result while computing Jordan block: "
"{}".format(M))
return ret
def blocks_from_nullity_chain(d):
"""Return a list of the size of each Jordan block.
If d_n is the nullity of E**n, then the number
of Jordan blocks of size n is
2*d_n - d_(n-1) - d_(n+1)"""
# d[0] is always the number of columns, so skip past it
mid = [2 * d[n] - d[n - 1] - d[n + 1] for n in range(1, len(d) - 1)]
# d is assumed to plateau with "d[ len(d) ] == d[-1]", so
# 2*d_n - d_(n-1) - d_(n+1) == d_n - d_(n-1)
end = [d[-1] - d[-2]] if len(d) > 1 else [d[0]]
return mid + end
def pick_vec(small_basis, big_basis):
"""Picks a vector from big_basis that isn't in
the subspace spanned by small_basis"""
if len(small_basis) == 0:
return big_basis[0]
for v in big_basis:
_, pivots = M.hstack(*(small_basis +
[v])).echelon_form(with_pivots=True)
if pivots[-1] == len(small_basis):
return v
# roots doesn't like Floats, so replace them with Rationals
if has_floats:
mat = mat.applyfunc(lambda x: nsimplify(x, rational=True))
# first calculate the jordan block structure
eigs = mat.eigenvals()
# Make sure that we have all roots in radical form
for x in eigs:
if x.has(CRootOf):
raise MatrixError(
"Jordan normal form is not implemented if the matrix have "
"eigenvalues in CRootOf form")
# most matrices have distinct eigenvalues
# and so are diagonalizable. In this case, don't
# do extra work!
if len(eigs.keys()) == mat.cols:
blocks = list(sorted(eigs.keys(), key=default_sort_key))
jordan_mat = mat.diag(*blocks)
if not calc_transform:
return restore_floats(jordan_mat)
jordan_basis = [eig_mat(eig, 1).nullspace()[0] for eig in blocks]
basis_mat = mat.hstack(*jordan_basis)
return restore_floats(basis_mat, jordan_mat)
block_structure = []
for eig in sorted(eigs.keys(), key=default_sort_key):
algebraic_multiplicity = eigs[eig]
chain = nullity_chain(eig, algebraic_multiplicity)
block_sizes = blocks_from_nullity_chain(chain)
# if block_sizes = = [a, b, c, ...], then the number of
# Jordan blocks of size 1 is a, of size 2 is b, etc.
# create an array that has (eig, block_size) with one
# entry for each block
size_nums = [(i + 1, num) for i, num in enumerate(block_sizes)]
# we expect larger Jordan blocks to come earlier
size_nums.reverse()
block_structure.extend(
(eig, size) for size, num in size_nums for _ in range(num))
jordan_form_size = sum(size for eig, size in block_structure)
if jordan_form_size != M.rows:
raise MatrixError("SymPy had encountered an inconsistent result while "
"computing Jordan block. : {}".format(M))
blocks = (mat.jordan_block(size=size, eigenvalue=eig)
for eig, size in block_structure)
jordan_mat = mat.diag(*blocks)
if not calc_transform:
return restore_floats(jordan_mat)
# For each generalized eigenspace, calculate a basis.
# We start by looking for a vector in null( (A - eig*I)**n )
# which isn't in null( (A - eig*I)**(n-1) ) where n is
# the size of the Jordan block
#
# Ideally we'd just loop through block_structure and
# compute each generalized eigenspace. However, this
# causes a lot of unneeded computation. Instead, we
# go through the eigenvalues separately, since we know
# their generalized eigenspaces must have bases that
# are linearly independent.
jordan_basis = []
for eig in sorted(eigs.keys(), key=default_sort_key):
eig_basis = []
for block_eig, size in block_structure:
if block_eig != eig:
continue
null_big = (eig_mat(eig, size)).nullspace()
null_small = (eig_mat(eig, size - 1)).nullspace()
# we want to pick something that is in the big basis
# and not the small, but also something that is independent
# of any other generalized eigenvectors from a different
# generalized eigenspace sharing the same eigenvalue.
vec = pick_vec(null_small + eig_basis, null_big)
new_vecs = [
eig_mat(eig, i).multiply(vec, dotprodsimp=None)
for i in range(size)
]
eig_basis.extend(new_vecs)
jordan_basis.extend(reversed(new_vecs))
basis_mat = mat.hstack(*jordan_basis)
return restore_floats(basis_mat, jordan_mat)
def _eval_evalf(self, prec=15, **options):
pt, r = self.args
dps = prec_to_dps(prec)
pt = pt.evalf(n=dps, **options)
r = r.evalf(n=dps, **options)
return self.func(pt, r, evaluate=False)
def _eval_evalf(self, prec=15, **options):
f, (t, a, b) = self.args
dps = prec_to_dps(prec)
f = tuple([i.evalf(n=dps, **options) for i in f])
a, b = [i.evalf(n=dps, **options) for i in (a, b)]
return self.func(f, (t, a, b))
def _eval_evalf(self, prec, **kwargs):
"""Evaluate this complex root to the given precision."""
# all kwargs are ignored
return self.eval_rational(n=prec_to_dps(prec))._evalf(prec)
def _eval_evalf(self, prec, **kwargs):
"""Evaluate this complex root to the given precision."""
# all kwargs are ignored
return self.eval_rational(n=prec_to_dps(prec))._evalf(prec)
def _eval_evalf(self, prec):
z = self.args[0]
nprec = prec_to_dps(prec)
return polygamma(1, z).evalf(n=nprec)
def eval(cls, arg):
from .complexes import im
if arg.is_integer:
return arg
if isinstance(arg, cls):
return arg
if arg.is_imaginary or (I*arg).is_extended_real:
i = im(arg)
if not i.has(I):
return cls(i)*I
return cls(arg, evaluate=False)
v = cls._eval_number(arg)
if v is not None:
return v
# Integral, numerical, symbolic part
ipart = npart = spart = Integer(0)
# Extract integral (or complex integral) terms
terms = Add.make_args(arg)
for t in terms:
if t.is_integer or (t.is_imaginary and im(t).is_integer):
ipart += t
elif t.free_symbols:
spart += t
else:
npart += t
if not (npart or spart):
return ipart
# Evaluate npart numerically if independent of spart
if npart and (not spart or npart.is_extended_real and
(spart.is_imaginary or (I*spart).is_extended_real) or
npart.is_imaginary and spart.is_extended_real):
npart_int = None
try:
from ...core.evalf import DEFAULT_MAXPREC as TARGET
prec = 10
r, i = Integer(0), Integer(0)
npart_re, npart_im = npart.as_real_imag()
while prec < TARGET:
dps = prec_to_dps(prec)
r, i = npart_re.evalf(dps), npart_im.evalf(dps)
if ((not r or int(2**prec*abs(r)) > 2**prec*abs(int(r))) and
(not i or int(2**prec*abs(i)) > 2**prec*abs(int(i)))):
npart_int = cls(r) + cls(i)*I
break
prec += 10
else:
raise PrecisionExhausted
except PrecisionExhausted:
npart_int = cls(r) + cls(i)*I
if not npart.equals(npart_int):
npart_int = None
if npart_int is not None:
ipart += npart_int
npart = Integer(0)
spart += npart
if not spart:
return ipart
elif spart.is_imaginary or (I*spart).is_extended_real:
return ipart + cls(im(spart), evaluate=False)*I
else:
return ipart + cls(spart, evaluate=False)
def hypsum(expr, n, start, prec):
"""
Sum a rapidly convergent infinite hypergeometric series with
given general term, e.g. e = hypsum(1/factorial(n), n). The
quotient between successive terms must be a quotient of integer
polynomials.
"""
from sympy import Float, hypersimp, lambdify
if prec == float('inf'):
raise NotImplementedError('does not support inf prec')
if start:
expr = expr.subs(n, n + start)
hs = hypersimp(expr, n)
if hs is None:
raise NotImplementedError("a hypergeometric series is required")
num, den = hs.as_numer_denom()
func1 = lambdify(n, num)
func2 = lambdify(n, den)
h, g, p = check_convergence(num, den, n)
if h < 0:
raise ValueError("Sum diverges like (n!)^%i" % (-h))
term = expr.subs(n, 0)
if not term.is_Rational:
raise NotImplementedError("Non rational term functionality is not implemented.")
# Direct summation if geometric or faster
if h > 0 or (h == 0 and abs(g) > 1):
term = (MPZ(term.p) << prec) // term.q
s = term
k = 1
while abs(term) > 5:
term *= MPZ(func1(k - 1))
term //= MPZ(func2(k - 1))
s += term
k += 1
return from_man_exp(s, -prec)
else:
alt = g < 0
if abs(g) < 1:
raise ValueError("Sum diverges like (%i)^n" % abs(1/g))
if p < 1 or (p == 1 and not alt):
raise ValueError("Sum diverges like n^%i" % (-p))
# We have polynomial convergence: use Richardson extrapolation
vold = None
ndig = prec_to_dps(prec)
while True:
# Need to use at least quad precision because a lot of cancellation
# might occur in the extrapolation process; we check the answer to
# make sure that the desired precision has been reached, too.
prec2 = 4*prec
term0 = (MPZ(term.p) << prec2) // term.q
def summand(k, _term=[term0]):
if k:
k = int(k)
_term[0] *= MPZ(func1(k - 1))
_term[0] //= MPZ(func2(k - 1))
return make_mpf(from_man_exp(_term[0], -prec2))
with workprec(prec):
v = nsum(summand, [0, mpmath_inf], method='richardson')
vf = Float(v, ndig)
if vold is not None and vold == vf:
break
prec += prec # double precision each time
vold = vf
return v._mpf_
def _eval_evalf(self, prec=15, **options):
pt, tup = self.args
dps = prec_to_dps(prec)
pt = pt.evalf(n=dps, **options)
tup = tuple([i.evalf(n=dps, **options) for i in tup])
return self.func(pt, normal_vector=tup, evaluate=False)
def eval(cls, arg, base=None):
from .complexes import unpolarify
if base is not None:
if base == 1:
if arg == 1:
return nan
else:
return zoo
try:
# handle extraction of powers of the base now
# or else expand_log in Mul would have to handle this
n = multiplicity(base, arg)
if n:
den = base**n
if den.is_Integer:
return n + log(arg // den) / log(base)
else:
return n + log(arg / den) / log(base)
except ValueError:
pass
if base is not E:
if arg.is_Float:
dps = prec_to_dps(arg._prec + 4)
return cls(arg) / cls(base).evalf(dps)
else:
return cls(arg) / cls(base)
else:
return cls(arg)
if arg.is_Number:
if arg == 0:
return zoo
elif arg == 1:
return Integer(0)
elif arg in (oo, -oo):
return oo
elif arg.is_Rational:
if arg.denominator != 1:
return cls(arg.numerator) - cls(arg.denominator)
if arg.is_Exp and arg.exp.is_extended_real:
return arg.exp
elif isinstance(arg, exp_polar):
return unpolarify(arg.exp)
if arg.is_number:
if arg.is_negative:
return pi * I + cls(-arg)
elif arg is zoo:
return zoo
elif arg is E:
return Integer(1)
# don't autoexpand Pow or Mul (see the issue sympy/sympy#3351):
if not arg.is_Add:
coeff = arg.as_coefficient(I)
if coeff is not None:
if coeff in (oo, -oo):
return oo
elif coeff.is_Rational:
if coeff.is_nonnegative:
return +pi * I / 2 + cls(+coeff)
else:
return -pi * I / 2 + cls(-coeff)
