Source code for sympy.core.evalf

Adaptive numerical evaluation of SymPy expressions, using mpmath
for mathematical functions.
from __future__ import print_function, division

import math

import sympy.mpmath.libmp as libmp
from sympy.mpmath import make_mpc, make_mpf, mp, mpc, mpf, nsum, quadts, quadosc
from sympy.mpmath import inf as mpmath_inf
from sympy.mpmath.libmp import (from_int, from_man_exp, from_rational, fhalf,
        fnan, fnone, fone, fzero, mpf_abs, mpf_add,
        mpf_atan, mpf_atan2, mpf_cmp, mpf_cos, mpf_e, mpf_exp, mpf_log, mpf_lt,
        mpf_mul, mpf_neg, mpf_pi, mpf_pow, mpf_pow_int, mpf_shift, mpf_sin,
        mpf_sqrt, normalize, round_nearest, to_int, to_str)
from sympy.mpmath.libmp import bitcount as mpmath_bitcount
from sympy.mpmath.libmp.backend import MPZ
from sympy.mpmath.libmp.libmpc import _infs_nan
from sympy.mpmath.libmp.libmpf import dps_to_prec
from sympy.mpmath.libmp.gammazeta import mpf_bernoulli

from .compatibility import SYMPY_INTS
from .sympify import sympify
from .core import C
from .singleton import S
from .containers import Tuple

LG10 = math.log(10, 2)
rnd = round_nearest

def bitcount(n):
    return mpmath_bitcount(int(n))

# Used in a few places as placeholder values to denote exponents and
# precision levels, e.g. of exact numbers. Must be careful to avoid
# passing these to mpmath functions or returning them in final results.
INF = float(mpmath_inf)
MINUS_INF = float(-mpmath_inf)

# ~= 100 digits. Real men set this to INF.

[docs]class PrecisionExhausted(ArithmeticError): pass #----------------------------------------------------------------------------# # # # Helper functions for arithmetic and complex parts # # # #----------------------------------------------------------------------------#
""" An mpf value tuple is a tuple of integers (sign, man, exp, bc) representing a floating-point number: [1, -1][sign]*man*2**exp where sign is 0 or 1 and bc should correspond to the number of bits used to represent the mantissa (man) in binary notation, e.g. >>> from sympy.core.evalf import bitcount >>> sign, man, exp, bc = 0, 5, 1, 3 >>> n = [1, -1][sign]*man*2**exp >>> n, bitcount(man) (10, 3) A temporary result is a tuple (re, im, re_acc, im_acc) where re and im are nonzero mpf value tuples representing approximate numbers, or None to denote exact zeros. re_acc, im_acc are integers denoting log2(e) where e is the estimated relative accuracy of the respective complex part, but may be anything if the corresponding complex part is None. """ def fastlog(x): """Fast approximation of log2(x) for an mpf value tuple x. Notes: Calculated as exponent + width of mantissa. This is an approximation for two reasons: 1) it gives the ceil(log2(abs(x))) value and 2) it is too high by 1 in the case that x is an exact power of 2. Although this is easy to remedy by testing to see if the odd mpf mantissa is 1 (indicating that one was dealing with an exact power of 2) that would decrease the speed and is not necessary as this is only being used as an approximation for the number of bits in x. The correct return value could be written as "x[2] + (x[3] if x[1] != 1 else 0)". Since mpf tuples always have an odd mantissa, no check is done to see if the mantissa is a multiple of 2 (in which case the result would be too large by 1). Examples ======== >>> from sympy import log >>> from sympy.core.evalf import fastlog, bitcount >>> s, m, e = 0, 5, 1 >>> bc = bitcount(m) >>> n = [1, -1][s]*m*2**e >>> n, (log(n)/log(2)).evalf(2), fastlog((s, m, e, bc)) (10, 3.3, 4) """ if not x or x == fzero: return MINUS_INF return x[2] + x[3] def pure_complex(v): """Return a and b if v matches a + I*b where b is not zero and a and b are Numbers, else None. >>> from sympy.core.evalf import pure_complex >>> from sympy import Tuple, I >>> a, b = Tuple(2, 3) >>> pure_complex(a) >>> pure_complex(a + b*I) (2, 3) >>> pure_complex(I) (0, 1) """ h, t = v.as_coeff_Add() c, i = t.as_coeff_Mul() if i is S.ImaginaryUnit: return h, c def scaled_zero(mag, sign=1): """Return an mpf representing a power of two with magnitude ``mag`` and -1 for precision. Or, if ``mag`` is a scaled_zero tuple, then just remove the sign from within the list that it was initially wrapped in. Examples ======== >>> from sympy.core.evalf import scaled_zero >>> from sympy import Float >>> z, p = scaled_zero(100) >>> z, p (([0], 1, 100, 1), -1) >>> ok = scaled_zero(z) >>> ok (0, 1, 100, 1) >>> Float(ok) 1.26765060022823e+30 >>> Float(ok, p) 0.e+30 >>> ok, p = scaled_zero(100, -1) >>> Float(scaled_zero(ok), p) -0.e+30 """ if type(mag) is tuple and len(mag) == 4 and iszero(mag, scaled=True): return (mag[0][0],) + mag[1:] elif isinstance(mag, SYMPY_INTS): if sign not in [-1, 1]: raise ValueError('sign must be +/-1') rv, p = mpf_shift(fone, mag), -1 s = 0 if sign == 1 else 1 rv = ([s],) + rv[1:] return rv, p else: raise ValueError('scaled zero expects int or scaled_zero tuple.') def iszero(mpf, scaled=False): if not scaled: return not mpf or not mpf[1] and not mpf[-1] return mpf and type(mpf[0]) is list and mpf[1] == mpf[-1] == 1 def complex_accuracy(result): """ Returns relative accuracy of a complex number with given accuracies for the real and imaginary parts. The relative accuracy is defined in the complex norm sense as ||z|+|error|| / |z| where error is equal to (real absolute error) + (imag absolute error)*i. The full expression for the (logarithmic) error can be approximated easily by using the max norm to approximate the complex norm. In the worst case (re and im equal), this is wrong by a factor sqrt(2), or by log2(sqrt(2)) = 0.5 bit. """ re, im, re_acc, im_acc = result if not im: if not re: return INF return re_acc if not re: return im_acc re_size = fastlog(re) im_size = fastlog(im) absolute_error = max(re_size - re_acc, im_size - im_acc) relative_error = absolute_error - max(re_size, im_size) return -relative_error def get_abs(expr, prec, options): re, im, re_acc, im_acc = evalf(expr, prec + 2, options) if not re: re, re_acc, im, im_acc = im, im_acc, re, re_acc if im: return libmp.mpc_abs((re, im), prec), None, re_acc, None elif re: return mpf_abs(re), None, re_acc, None else: return None, None, None, None def get_complex_part(expr, no, prec, options): """no = 0 for real part, no = 1 for imaginary part""" workprec = prec i = 0 while 1: res = evalf(expr, workprec, options) value, accuracy = res[no::2] # XXX is the last one correct? Consider re((1+I)**2).n() if (not value) or accuracy >= prec or -value[2] > prec: return value, None, accuracy, None workprec += max(30, 2**i) i += 1 def evalf_abs(expr, prec, options): return get_abs(expr.args[0], prec, options) def evalf_re(expr, prec, options): return get_complex_part(expr.args[0], 0, prec, options) def evalf_im(expr, prec, options): return get_complex_part(expr.args[0], 1, prec, options) def finalize_complex(re, im, prec): if re == fzero and im == fzero: raise ValueError("got complex zero with unknown accuracy") elif re == fzero: return None, im, None, prec elif im == fzero: return re, None, prec, None size_re = fastlog(re) size_im = fastlog(im) if size_re > size_im: re_acc = prec im_acc = prec + min(-(size_re - size_im), 0) else: im_acc = prec re_acc = prec + min(-(size_im - size_re), 0) return re, im, re_acc, im_acc def chop_parts(value, prec): """ Chop off tiny real or complex parts. """ re, im, re_acc, im_acc = value # Method 1: chop based on absolute value if re and re not in _infs_nan and (fastlog(re) < -prec + 4): re, re_acc = None, None if im and im not in _infs_nan and (fastlog(im) < -prec + 4): im, im_acc = None, None # Method 2: chop if inaccurate and relatively small if re and im: delta = fastlog(re) - fastlog(im) if re_acc < 2 and (delta - re_acc <= -prec + 4): re, re_acc = None, None if im_acc < 2 and (delta - im_acc >= prec - 4): im, im_acc = None, None return re, im, re_acc, im_acc def check_target(expr, result, prec): a = complex_accuracy(result) if a < prec: raise PrecisionExhausted("Failed to distinguish the expression: \n\n%s\n\n" "from zero. Try simplifying the input, using chop=True, or providing " "a higher maxn for evalf" % (expr)) def get_integer_part(expr, no, options, return_ints=False): """ With no = 1, computes ceiling(expr) With no = -1, computes floor(expr) Note: this function either gives the exact result or signals failure. """ # The expression is likely less than 2^30 or so assumed_size = 30 ire, iim, ire_acc, iim_acc = evalf(expr, assumed_size, options) # We now know the size, so we can calculate how much extra precision # (if any) is needed to get within the nearest integer if ire and iim: gap = max(fastlog(ire) - ire_acc, fastlog(iim) - iim_acc) elif ire: gap = fastlog(ire) - ire_acc elif iim: gap = fastlog(iim) - iim_acc else: # ... or maybe the expression was exactly zero return None, None, None, None margin = 10 if gap >= -margin: ire, iim, ire_acc, iim_acc = \ evalf(expr, margin + assumed_size + gap, options) # We can now easily find the nearest integer, but to find floor/ceil, we # must also calculate whether the difference to the nearest integer is # positive or negative (which may fail if very close). def calc_part(expr, nexpr): nint = int(to_int(nexpr, rnd)) expr = C.Add(expr, -nint, evaluate=False) x, _, x_acc, _ = evalf(expr, 10, options) try: check_target(expr, (x, None, x_acc, None), 3) except PrecisionExhausted: if not expr.equals(0): raise PrecisionExhausted x = fzero nint += int(no*(mpf_cmp(x or fzero, fzero) == no)) nint = from_int(nint) return nint, fastlog(nint) + 10 re, im, re_acc, im_acc = None, None, None, None if ire: re, re_acc = calc_part(C.re(expr, evaluate=False), ire) if iim: im, im_acc = calc_part(C.im(expr, evaluate=False), iim) if return_ints: return int(to_int(re or fzero)), int(to_int(im or fzero)) return re, im, re_acc, im_acc def evalf_ceiling(expr, prec, options): return get_integer_part(expr.args[0], 1, options) def evalf_floor(expr, prec, options): return get_integer_part(expr.args[0], -1, options) #----------------------------------------------------------------------------# # # # Arithmetic operations # # # #----------------------------------------------------------------------------# 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 """ from sympy.core.core import C 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 = [] for t in terms: arg = C.Float._new(t[0], 1) if arg is S.NaN or arg.is_unbounded: special.append(arg) if special: from sympy.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 #print "returning", to_str(r[0],50), r[1] return r def evalf_add(v, prec, options): res = pure_complex(v) if res: h, c = res re, _, re_acc, _ = evalf(h, prec, options) im, _, im_acc, _ = evalf(c, prec, options) return re, im, re_acc, im_acc oldmaxprec = options.get('maxprec', DEFAULT_MAXPREC) i = 0 target_prec = prec while 1: options['maxprec'] = min(oldmaxprec, 2*prec) terms = [evalf(arg, prec + 10, options) for arg in v.args] re, re_acc = add_terms( [a[0::2] for a in terms if a[0]], prec, target_prec) im, im_acc = add_terms( [a[1::2] for a in terms if a[1]], prec, target_prec) acc = complex_accuracy((re, im, re_acc, im_acc)) if acc >= target_prec: if options.get('verbose'): print("ADD: wanted", target_prec, "accurate bits, got", re_acc, im_acc) break else: if (prec - target_prec) > options['maxprec']: break prec = prec + max(10 + 2**i, target_prec - acc) i += 1 if options.get('verbose'): print("ADD: restarting with prec", prec) options['maxprec'] = oldmaxprec if iszero(re, scaled=True): re = scaled_zero(re) if iszero(im, scaled=True): im = scaled_zero(im) return re, im, re_acc, im_acc def evalf_mul(v, prec, options): from sympy.core.core import C res = pure_complex(v) if res: # the only pure complex that is a mul is h*I _, h = res im, _, im_acc, _ = evalf(h, prec, options) return None, im, None, im_acc args = list(v.args) # see if any argument is NaN or oo and thus warrants a special return special = [] for arg in args: arg = evalf(arg, prec, options) if arg[0] is None: continue arg = C.Float._new(arg[0], 1) if arg is S.NaN or arg.is_unbounded: special.append(arg) if special: from sympy.core.mul import Mul special = Mul(*special) return evalf(special, prec + 4, {}) # With guard digits, multiplication in the real case does not destroy # accuracy. This is also true in the complex case when considering the # total accuracy; however accuracy for the real or imaginary parts # separately may be lower. acc = prec # XXX: big overestimate working_prec = prec + len(args) + 5 # Empty product is 1 start = man, exp, bc = MPZ(1), 0, 1 # First, we multiply all pure real or pure imaginary numbers. # direction tells us that the result should be multiplied by # I**direction; all other numbers get put into complex_factors # to be multiplied out after the first phase. last = len(args) direction = 0 args.append(S.One) complex_factors = [] for i, arg in enumerate(args): if i != last and pure_complex(arg): args[-1] = (args[-1]*arg).expand() continue elif i == last and arg is S.One: continue re, im, re_acc, im_acc = evalf(arg, working_prec, options) if re and im: complex_factors.append((re, im, re_acc, im_acc)) continue elif re: (s, m, e, b), w_acc = re, re_acc elif im: (s, m, e, b), w_acc = im, im_acc direction += 1 else: return None, None, None, None direction += 2*s man *= m exp += e bc += b if bc > 3*working_prec: man >>= working_prec exp += working_prec acc = min(acc, w_acc) sign = (direction & 2) >> 1 if not complex_factors: v = normalize(sign, man, exp, bitcount(man), prec, rnd) # multiply by i if direction & 1: return None, v, None, acc else: return v, None, acc, None else: # initialize with the first term if (man, exp, bc) != start: # there was a real part; give it an imaginary part re, im = (sign, man, exp, bitcount(man)), (0, MPZ(0), 0, 0) i0 = 0 else: # there is no real part to start (other than the starting 1) wre, wim, wre_acc, wim_acc = complex_factors[0] acc = min(acc, complex_accuracy((wre, wim, wre_acc, wim_acc))) re = wre im = wim i0 = 1 for wre, wim, wre_acc, wim_acc in complex_factors[i0:]: # acc is the overall accuracy of the product; we aren't # computing exact accuracies of the product. acc = min(acc, complex_accuracy((wre, wim, wre_acc, wim_acc))) use_prec = working_prec A = mpf_mul(re, wre, use_prec) B = mpf_mul(mpf_neg(im), wim, use_prec) C = mpf_mul(re, wim, use_prec) D = mpf_mul(im, wre, use_prec) re = mpf_add(A, B, use_prec) im = mpf_add(C, D, use_prec) if options.get('verbose'): print("MUL: wanted", prec, "accurate bits, got", acc) # multiply by I if direction & 1: re, im = mpf_neg(im), re return re, im, acc, acc def evalf_pow(v, prec, options): target_prec = prec base, exp = v.args # We handle x**n separately. This has two purposes: 1) it is much # faster, because we avoid calling evalf on the exponent, and 2) it # allows better handling of real/imaginary parts that are exactly zero if exp.is_Integer: p = exp.p # Exact if not p: return fone, None, prec, None # Exponentiation by p magnifies relative error by |p|, so the # base must be evaluated with increased precision if p is large prec += int(math.log(abs(p), 2)) re, im, re_acc, im_acc = evalf(base, prec + 5, options) # Real to integer power if re and not im: return mpf_pow_int(re, p, target_prec), None, target_prec, None # (x*I)**n = I**n * x**n if im and not re: z = mpf_pow_int(im, p, target_prec) case = p % 4 if case == 0: return z, None, target_prec, None if case == 1: return None, z, None, target_prec if case == 2: return mpf_neg(z), None, target_prec, None if case == 3: return None, mpf_neg(z), None, target_prec # Zero raised to an integer power if not re: return None, None, None, None # General complex number to arbitrary integer power re, im = libmp.mpc_pow_int((re, im), p, prec) # Assumes full accuracy in input return finalize_complex(re, im, target_prec) # Pure square root if exp is S.Half: xre, xim, _, _ = evalf(base, prec + 5, options) # General complex square root if xim: re, im = libmp.mpc_sqrt((xre or fzero, xim), prec) return finalize_complex(re, im, prec) if not xre: return None, None, None, None # Square root of a negative real number if mpf_lt(xre, fzero): return None, mpf_sqrt(mpf_neg(xre), prec), None, prec # Positive square root return mpf_sqrt(xre, prec), None, prec, None # We first evaluate the exponent to find its magnitude # This determines the working precision that must be used prec += 10 yre, yim, _, _ = evalf(exp, prec, options) # Special cases: x**0 if not (yre or yim): return fone, None, prec, None ysize = fastlog(yre) # Restart if too big # XXX: prec + ysize might exceed maxprec if ysize > 5: prec += ysize yre, yim, _, _ = evalf(exp, prec, options) # Pure exponential function; no need to evalf the base if base is S.Exp1: if yim: re, im = libmp.mpc_exp((yre or fzero, yim), prec) return finalize_complex(re, im, target_prec) return mpf_exp(yre, target_prec), None, target_prec, None xre, xim, _, _ = evalf(base, prec + 5, options) # 0**y if not (xre or xim): return None, None, None, None # (real ** complex) or (complex ** complex) if yim: re, im = libmp.mpc_pow( (xre or fzero, xim or fzero), (yre or fzero, yim), target_prec) return finalize_complex(re, im, target_prec) # complex ** real if xim: re, im = libmp.mpc_pow_mpf((xre or fzero, xim), yre, target_prec) return finalize_complex(re, im, target_prec) # negative ** real elif mpf_lt(xre, fzero): re, im = libmp.mpc_pow_mpf((xre, fzero), yre, target_prec) return finalize_complex(re, im, target_prec) # positive ** real else: return mpf_pow(xre, yre, target_prec), None, target_prec, None #----------------------------------------------------------------------------# # # # Special functions # # # #----------------------------------------------------------------------------# def evalf_trig(v, prec, options): """ This function handles sin and cos of complex arguments. TODO: should also handle tan of complex arguments. """ if v.func is C.cos: func = mpf_cos elif v.func is C.sin: func = mpf_sin else: raise NotImplementedError arg = v.args[0] # 20 extra bits is possibly overkill. It does make the need # to restart very unlikely xprec = prec + 20 re, im, re_acc, im_acc = evalf(arg, xprec, options) if im: if 'subs' in options: v = v.subs(options['subs']) return evalf(v._eval_evalf(prec), prec, options) if not re: if v.func is C.cos: return fone, None, prec, None elif v.func is C.sin: return None, None, None, None else: raise NotImplementedError # For trigonometric functions, we are interested in the # fixed-point (absolute) accuracy of the argument. xsize = fastlog(re) # Magnitude <= 1.0. OK to compute directly, because there is no # danger of hitting the first root of cos (with sin, magnitude # <= 2.0 would actually be ok) if xsize < 1: return func(re, prec, rnd), None, prec, None # Very large if xsize >= 10: xprec = prec + xsize re, im, re_acc, im_acc = evalf(arg, xprec, options) # Need to repeat in case the argument is very close to a # multiple of pi (or pi/2), hitting close to a root while 1: y = func(re, prec, rnd) ysize = fastlog(y) gap = -ysize accuracy = (xprec - xsize) - gap if accuracy < prec: if options.get('verbose'): print("SIN/COS", accuracy, "wanted", prec, "gap", gap) print(to_str(y, 10)) if xprec > options.get('maxprec', DEFAULT_MAXPREC): return y, None, accuracy, None xprec += gap re, im, re_acc, im_acc = evalf(arg, xprec, options) continue else: return y, None, prec, None def evalf_log(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( C.log(C.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 = C.Add(S.NegativeOne, arg, evaluate=False) xre, xim, _, _ = evalf_add(arg, prec, options) prec2 = workprec - fastlog(xre) re = mpf_log(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 evalf_atan(v, prec, options): arg = v.args[0] xre, xim, reacc, imacc = evalf(arg, prec + 5, options) if xre is xim is None: return (None,)*4 if xim: raise NotImplementedError return mpf_atan(xre, prec, rnd), None, prec, None def evalf_subs(prec, subs): """ Change all Float entries in `subs` to have precision prec. """ newsubs = {} for a, b in subs.items(): b = S(b) if b.is_Float: b = b._eval_evalf(prec) newsubs[a] = b return newsubs def evalf_piecewise(expr, prec, options): if 'subs' in options: expr = expr.subs(evalf_subs(prec, options['subs'])) newopts = options.copy() del newopts['subs'] if hasattr(expr, 'func'): return evalf(expr, prec, newopts) if type(expr) == float: return evalf(C.Float(expr), prec, newopts) if type(expr) == int: return evalf(C.Integer(expr), prec, newopts) # We still have undefined symbols raise NotImplementedError def evalf_bernoulli(expr, prec, options): arg = expr.args[0] if not arg.is_Integer: raise ValueError("Bernoulli number index must be an integer") n = int(arg) b = mpf_bernoulli(n, prec, rnd) if b == fzero: return None, None, None, None return b, None, prec, None #----------------------------------------------------------------------------# # # # High-level operations # # # #----------------------------------------------------------------------------# def as_mpmath(x, prec, options): x = sympify(x) if isinstance(x, C.Zero): return mpf(0) if isinstance(x, C.Infinity): return mpf('inf') if isinstance(x, C.NegativeInfinity): return mpf('-inf') # XXX re, im, _, _ = evalf(x, prec, options) if im: return mpc(re or fzero, im) return mpf(re) def do_integral(expr, prec, options): func = expr.args[0] x, xlow, xhigh = expr.args[1] orig = mp.prec oldmaxprec = options.get('maxprec', DEFAULT_MAXPREC) options['maxprec'] = min(oldmaxprec, 2*prec) try: mp.prec = prec + 5 xlow = as_mpmath(xlow, prec + 15, options) xhigh = as_mpmath(xhigh, prec + 15, options) # Integration is like summation, and we can phone home from # the integrand function to update accuracy summation style # Note that this accuracy is inaccurate, since it fails # to account for the variable quadrature weights, # but it is better than nothing have_part = [False, False] max_real_term = [MINUS_INF] max_imag_term = [MINUS_INF] def f(t): re, im, re_acc, im_acc = evalf(func, mp.prec, {'subs': {x: t}}) have_part[0] = re or have_part[0] have_part[1] = im or have_part[1] max_real_term[0] = max(max_real_term[0], fastlog(re)) max_imag_term[0] = max(max_imag_term[0], fastlog(im)) if im: return mpc(re or fzero, im) return mpf(re or fzero) if options.get('quad') == 'osc': A = C.Wild('A', exclude=[x]) B = C.Wild('B', exclude=[x]) D = C.Wild('D') m = func.match(C.cos(A*x + B)*D) if not m: m = func.match(C.sin(A*x + B)*D) if not m: raise ValueError("An integrand of the form sin(A*x+B)*f(x) " "or cos(A*x+B)*f(x) is required for oscillatory quadrature") period = as_mpmath(2*S.Pi/m[A], prec + 15, options) result = quadosc(f, [xlow, xhigh], period=period) # XXX: quadosc does not do error detection yet quadrature_error = MINUS_INF else: result, quadrature_error = quadts(f, [xlow, xhigh], error=1) quadrature_error = fastlog(quadrature_error._mpf_) finally: options['maxprec'] = oldmaxprec mp.prec = orig if have_part[0]: re = result.real._mpf_ if re == fzero: re, re_acc = scaled_zero( min(-prec, -max_real_term[0], -quadrature_error)) re = scaled_zero(re) # handled ok in evalf_integral else: re_acc = -max(max_real_term[0] - fastlog(re) - prec, quadrature_error) else: re, re_acc = None, None if have_part[1]: im = result.imag._mpf_ if im == fzero: im, im_acc = scaled_zero( min(-prec, -max_imag_term[0], -quadrature_error)) im = scaled_zero(im) # handled ok in evalf_integral else: im_acc = -max(max_imag_term[0] - fastlog(im) - prec, quadrature_error) else: im, im_acc = None, None result = re, im, re_acc, im_acc return result def evalf_integral(expr, prec, options): limits = expr.limits if len(limits) != 1 or len(limits[0]) != 3: raise NotImplementedError workprec = prec i = 0 maxprec = options.get('maxprec', INF) while 1: result = do_integral(expr, workprec, options) # if a scaled_zero comes back accuracy will compute to -1 # which will cause workprec to increment by 1 accuracy = complex_accuracy(result) if accuracy >= prec or workprec >= maxprec: return result workprec += prec - max(-2**i, accuracy) i += 1 def check_convergence(numer, denom, n): """ Returns (h, g, p) where -- h is: > 0 for convergence of rate 1/factorial(n)**h < 0 for divergence of rate factorial(n)**(-h) = 0 for geometric or polynomial convergence or divergence -- abs(g) is: > 1 for geometric convergence of rate 1/h**n < 1 for geometric divergence of rate h**n = 1 for polynomial convergence or divergence (g < 0 indicates an alternating series) -- p is: > 1 for polynomial convergence of rate 1/n**h <= 1 for polynomial divergence of rate n**(-h) """ npol = C.Poly(numer, n) dpol = C.Poly(denom, n) p = npol.degree() q = dpol.degree() rate = q - p if rate: return rate, None, None constant = dpol.LC() / npol.LC() if abs(constant) != 1: return rate, constant, None if npol.degree() == dpol.degree() == 0: return rate, constant, 0 pc = npol.all_coeffs()[1] qc = dpol.all_coeffs()[1] return rate, constant, (qc - pc)/dpol.LC() 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 hypersimp, lambdify 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)) # Direct summation if geometric or faster if h > 0 or (h == 0 and abs(g) > 1): term = expr.subs(n, 0) 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 # Need to use at least quad precision because a lot of cancellation # might occur in the extrapolation process prec2 = 4*prec term = expr.subs(n, 0) term = (MPZ(term.p) << prec2) // term.q def summand(k, _term=[term]): 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)) orig = mp.prec try: mp.prec = prec v = nsum(summand, [0, mpmath_inf], method='richardson') finally: mp.prec = orig return v._mpf_ def evalf_sum(expr, prec, options): if 'subs' in options: expr = expr.subs(options['subs']) func = expr.function limits = expr.limits if len(limits) != 1 or len(limits[0]) != 3: raise NotImplementedError if func is S.Zero: return mpf(0), None, None, None prec2 = prec + 10 try: n, a, b = limits[0] if b != S.Infinity or a != int(a): raise NotImplementedError # Use fast hypergeometric summation if possible v = hypsum(func, n, int(a), prec2) delta = prec - fastlog(v) if fastlog(v) < -10: v = hypsum(func, n, int(a), delta) return v, None, min(prec, delta), None except NotImplementedError: # Euler-Maclaurin summation for general series eps = C.Float(2.0)**(-prec) for i in range(1, 5): m = n = 2**i * prec s, err = expr.euler_maclaurin(m=m, n=n, eps=eps, eval_integral=False) err = err.evalf() if err <= eps: break err = fastlog(evalf(abs(err), 20, options)[0]) re, im, re_acc, im_acc = evalf(s, prec2, options) if re_acc is None: re_acc = -err if im_acc is None: im_acc = -err return re, im, re_acc, im_acc #----------------------------------------------------------------------------# # # # Symbolic interface # # # #----------------------------------------------------------------------------# def evalf_symbol(x, prec, options): val = options['subs'][x] if isinstance(val, mpf): if not val: return None, None, None, None return val._mpf_, None, prec, None else: if not '_cache' in options: options['_cache'] = {} cache = options['_cache'] cached, cached_prec = cache.get(x.name, (None, MINUS_INF)) if cached_prec >= prec: return cached v = evalf(sympify(val), prec, options) cache[x.name] = (v, prec) return v evalf_table = None def _create_evalf_table(): global evalf_table evalf_table = { C.Symbol: evalf_symbol, C.Dummy: evalf_symbol, C.Float: lambda x, prec, options: (x._mpf_, None, prec, None), C.Rational: lambda x, prec, options: (from_rational(x.p, x.q, prec), None, prec, None), C.Integer: lambda x, prec, options: (from_int(x.p, prec), None, prec, None), C.Zero: lambda x, prec, options: (None, None, prec, None), C.One: lambda x, prec, options: (fone, None, prec, None), C.Half: lambda x, prec, options: (fhalf, None, prec, None), C.Pi: lambda x, prec, options: (mpf_pi(prec), None, prec, None), C.Exp1: lambda x, prec, options: (mpf_e(prec), None, prec, None), C.ImaginaryUnit: lambda x, prec, options: (None, fone, None, prec), C.NegativeOne: lambda x, prec, options: (fnone, None, prec, None), C.NaN : lambda x, prec, options: (fnan, None, prec, None), C.exp: lambda x, prec, options: evalf_pow(C.Pow(S.Exp1, x.args[0], evaluate=False), prec, options), C.cos: evalf_trig, C.sin: evalf_trig, C.Add: evalf_add, C.Mul: evalf_mul, C.Pow: evalf_pow, C.log: evalf_log, C.atan: evalf_atan, C.Abs: evalf_abs, C.re: evalf_re, C.im: evalf_im, C.floor: evalf_floor, C.ceiling: evalf_ceiling, C.Integral: evalf_integral, C.Sum: evalf_sum, C.Piecewise: evalf_piecewise, C.bernoulli: evalf_bernoulli, } def evalf(x, prec, options): from sympy import re as re_, im as im_ try: rf = evalf_table[x.func] r = rf(x, prec, options) except KeyError: try: # Fall back to ordinary evalf if possible if 'subs' in options: x = x.subs(evalf_subs(prec, options['subs'])) re, im = x._eval_evalf(prec).as_real_imag() if re.has(re_) or im.has(im_): raise NotImplementedError if re == 0: re = None reprec = None else: re = re._to_mpmath(prec, allow_ints=False)._mpf_ reprec = prec if im == 0: im = None imprec = None else: im = im._to_mpmath(prec, allow_ints=False)._mpf_ imprec = prec r = re, im, reprec, imprec except AttributeError: raise NotImplementedError if options.get("verbose"): print("### input", x) print("### output", to_str(r[0] or fzero, 50)) print("### raw", r ) # r[0], r[2] print() chop = options.get('chop', False) if chop: if chop is True: chop_prec = prec else: # convert (approximately) from given tolerance; # the formula here will will make 1e-i rounds to 0 for # i in the range +/-27 while 2e-i will not be chopped chop_prec = int(round(-3.321*math.log10(chop) + 2.5)) if chop_prec == 3: chop_prec -= 1 r = chop_parts(r, chop_prec) if options.get("strict"): check_target(x, r, prec) return r class EvalfMixin(object): """Mixin class adding evalf capabililty.""" __slots__ = [] def evalf(self, n=15, subs=None, maxn=100, chop=False, strict=False, quad=None, verbose=False): """ Evaluate the given formula to an accuracy of n digits. Optional keyword arguments: subs=<dict> Substitute numerical values for symbols, e.g. subs={x:3, y:1+pi}. maxn=<integer> Allow a maximum temporary working precision of maxn digits (default=100) chop=<bool> Replace tiny real or imaginary parts in subresults by exact zeros (default=False) strict=<bool> Raise PrecisionExhausted if any subresult fails to evaluate to full accuracy, given the available maxprec (default=False) quad=<str> Choose algorithm for numerical quadrature. By default, tanh-sinh quadrature is used. For oscillatory integrals on an infinite interval, try quad='osc'. verbose=<bool> Print debug information (default=False) """ # for sake of sage that doesn't like evalf(1) if n == 1 and isinstance(self, C.Number): from sympy.core.expr import _mag rv = self.evalf(2, subs, maxn, chop, strict, quad, verbose) m = _mag(rv) rv = rv.round(1 - m) return rv if not evalf_table: _create_evalf_table() prec = dps_to_prec(n) options = {'maxprec': max(prec, int(maxn*LG10)), 'chop': chop, 'strict': strict, 'verbose': verbose} if subs is not None: options['subs'] = subs if quad is not None: options['quad'] = quad try: result = evalf(self, prec + 4, options) except NotImplementedError: # Fall back to the ordinary evalf v = self._eval_evalf(prec) if v is None: return self try: # If the result is numerical, normalize it result = evalf(v, prec, options) except NotImplementedError: # Probably contains symbols or unknown functions return v re, im, re_acc, im_acc = result if re: p = max(min(prec, re_acc), 1) #re = mpf_pos(re, p, rnd) re = C.Float._new(re, p) else: re = S.Zero if im: p = max(min(prec, im_acc), 1) #im = mpf_pos(im, p, rnd) im = C.Float._new(im, p) return re + im*S.ImaginaryUnit else: return re n = evalf def _evalf(self, prec): """Helper for evalf. Does the same thing but takes binary precision""" r = self._eval_evalf(prec) if r is None: r = self return r def _eval_evalf(self, prec): return def _to_mpmath(self, prec, allow_ints=True): # mpmath functions accept ints as input errmsg = "cannot convert to mpmath number" if allow_ints and self.is_Integer: return self.p if hasattr(self, '_as_mpf_val'): return make_mpf(self._as_mpf_val(prec)) try: re, im, _, _ = evalf(self, prec, {}) if im: if not re: re = fzero return make_mpc((re, im)) elif re: return make_mpf(re) else: return make_mpf(fzero) except NotImplementedError: v = self._eval_evalf(prec) if v is None: raise ValueError(errmsg) if v.is_Float: return make_mpf(v._mpf_) # Number + Number*I is also fine re, im = v.as_real_imag() if allow_ints and re.is_Integer: re = from_int(re.p) elif re.is_Float: re = re._mpf_ else: raise ValueError(errmsg) if allow_ints and im.is_Integer: im = from_int(im.p) elif im.is_Float: im = im._mpf_ else: raise ValueError(errmsg) return make_mpc((re, im))
[docs]def N(x, n=15, **options): """ Calls x.evalf(n, \*\*options). Both .n() and N() are equivalent to .evalf(); use the one that you like better. See also the docstring of .evalf() for information on the options. Examples ======== >>> from sympy import Sum, oo, N >>> from sympy.abc import k >>> Sum(1/k**k, (k, 1, oo)) Sum(k**(-k), (k, 1, oo)) >>> N(_, 4) 1.291 """ return sympify(x).evalf(n, **options)