heurisch.py

from __future__ import annotations

from itertools import permutations
from functools import reduce

from sympy.core.add import Add
from sympy.core.basic import Basic
from sympy.core.mul import Mul
from sympy.core.symbol import Wild, Dummy, Symbol
from sympy.core.basic import sympify
from sympy.core.numbers import Rational, pi, I
from sympy.core.relational import Eq, Ne
from sympy.core.singleton import S
from sympy.core.sorting import ordered
from sympy.core.traversal import iterfreeargs

from sympy.functions import exp, sin, cos, tan, cot, asin, atan
from sympy.functions import log, sinh, cosh, tanh, coth, asinh
from sympy.functions import sqrt, erf, erfi, li, Ei
from sympy.functions import besselj, bessely, besseli, besselk
from sympy.functions import hankel1, hankel2, jn, yn
from sympy.functions.elementary.complexes import Abs, re, im, sign, arg
from sympy.functions.elementary.exponential import LambertW
from sympy.functions.elementary.integers import floor, ceiling
from sympy.functions.elementary.piecewise import Piecewise
from sympy.functions.special.delta_functions import Heaviside, DiracDelta

from sympy.simplify.radsimp import collect

from sympy.logic.boolalg import And, Or
from sympy.utilities.iterables import uniq

from sympy.polys import quo, gcd, lcm, factor_list, cancel, PolynomialError
from sympy.polys.monomials import itermonomials
from sympy.polys.polyroots import root_factors

from sympy.polys.rings import PolyRing
from sympy.polys.solvers import solve_lin_sys
from sympy.polys.constructor import construct_domain

from sympy.integrals.integrals import integrate


def components(f, x):
    """
    Returns a set of all functional components of the given expression
    which includes symbols, function applications and compositions and
    non-integer powers. Fractional powers are collected with
    minimal, positive exponents.

    Examples
    ========

    >>> from sympy import cos, sin
    >>> from sympy.abc import x
    >>> from sympy.integrals.heurisch import components

    >>> components(sin(x)*cos(x)**2, x)
    {x, sin(x), cos(x)}

    See Also
    ========

    heurisch
    """
    result = set()

    if f.has_free(x):
        if f.is_symbol and f.is_commutative:
            result.add(f)
        elif f.is_Function or f.is_Derivative:
            for g in f.args:
                result |= components(g, x)

            result.add(f)
        elif f.is_Pow:
            result |= components(f.base, x)

            if not f.exp.is_Integer:
                if f.exp.is_Rational:
                    result.add(f.base**Rational(1, f.exp.q))
                else:
                    result |= components(f.exp, x) | {f}
        else:
            for g in f.args:
                result |= components(g, x)

    return result

# name -> [] of symbols
_symbols_cache: dict[str, list[Dummy]] = {}


# NB @cacheit is not convenient here
def _symbols(name, n):
    """get vector of symbols local to this module"""
    try:
        lsyms = _symbols_cache[name]
    except KeyError:
        lsyms = []
        _symbols_cache[name] = lsyms

    while len(lsyms) < n:
        lsyms.append( Dummy('%s%i' % (name, len(lsyms))) )

    return lsyms[:n]


def heurisch_wrapper(f, x, rewrite=False, hints=None, mappings=None, retries=3,
                     degree_offset=0, unnecessary_permutations=None,
                     _try_heurisch=None):
    """
    A wrapper around the heurisch integration algorithm.

    Explanation
    ===========

    This method takes the result from heurisch and checks for poles in the
    denominator. For each of these poles, the integral is reevaluated, and
    the final integration result is given in terms of a Piecewise.

    Examples
    ========

    >>> from sympy import cos, symbols
    >>> from sympy.integrals.heurisch import heurisch, heurisch_wrapper
    >>> n, x = symbols('n x')
    >>> heurisch(cos(n*x), x)
    sin(n*x)/n
    >>> heurisch_wrapper(cos(n*x), x)
    Piecewise((sin(n*x)/n, Ne(n, 0)), (x, True))

    See Also
    ========

    heurisch
    """
    from sympy.solvers.solvers import solve, denoms
    f = sympify(f)
    if not f.has_free(x):
        return f*x

    res = heurisch(f, x, rewrite, hints, mappings, retries, degree_offset,
                   unnecessary_permutations, _try_heurisch)
    if not isinstance(res, Basic):
        return res

    # We consider each denominator in the expression, and try to find
    # cases where one or more symbolic denominator might be zero. The
    # conditions for these cases are stored in the list slns.
    #
    # Since denoms returns a set we use ordered. This is important because the
    # ordering of slns determines the order of the resulting Piecewise so we
    # need a deterministic order here to make the output deterministic.
    slns = []
    for d in ordered(denoms(res)):
        try:
            slns += solve([d], dict=True, exclude=(x,))
        except NotImplementedError:
            pass
    if not slns:
        return res
    slns = list(uniq(slns))
    # Remove the solutions corresponding to poles in the original expression.
    slns0 = []
    for d in denoms(f):
        try:
            slns0 += solve([d], dict=True, exclude=(x,))
        except NotImplementedError:
            pass
    slns = [s for s in slns if s not in slns0]
    if not slns:
        return res
    if len(slns) > 1:
        eqs = []
        for sub_dict in slns:
            eqs.extend([Eq(key, value) for key, value in sub_dict.items()])
        slns = solve(eqs, dict=True, exclude=(x,)) + slns
    # For each case listed in the list slns, we reevaluate the integral.
    pairs = []
    for sub_dict in slns:
        expr = heurisch(f.subs(sub_dict), x, rewrite, hints, mappings, retries,
                        degree_offset, unnecessary_permutations,
                        _try_heurisch)
        cond = And(*[Eq(key, value) for key, value in sub_dict.items()])
        generic = Or(*[Ne(key, value) for key, value in sub_dict.items()])
        if expr is None:
            expr = integrate(f.subs(sub_dict),x)
        pairs.append((expr, cond))
    # If there is one condition, put the generic case first. Otherwise,
    # doing so may lead to longer Piecewise formulas
    if len(pairs) == 1:
        pairs = [(heurisch(f, x, rewrite, hints, mappings, retries,
                              degree_offset, unnecessary_permutations,
                              _try_heurisch),
                              generic),
                 (pairs[0][0], True)]
    else:
        pairs.append((heurisch(f, x, rewrite, hints, mappings, retries,
                              degree_offset, unnecessary_permutations,
                              _try_heurisch),
                              True))
    return Piecewise(*pairs)

class BesselTable:
    """
    Derivatives of Bessel functions of orders n and n-1
    in terms of each other.

    See the docstring of DiffCache.
    """

    def __init__(self):
        self.table = {}
        self.n = Dummy('n')
        self.z = Dummy('z')
        self._create_table()

    def _create_table(t):
        table, n, z = t.table, t.n, t.z
        for f in (besselj, bessely, hankel1, hankel2):
            table[f] = (f(n-1, z) - n*f(n, z)/z,
                        (n-1)*f(n-1, z)/z - f(n, z))

        f = besseli
        table[f] = (f(n-1, z) - n*f(n, z)/z,
                    (n-1)*f(n-1, z)/z + f(n, z))
        f = besselk
        table[f] = (-f(n-1, z) - n*f(n, z)/z,
                    (n-1)*f(n-1, z)/z - f(n, z))

        for f in (jn, yn):
            table[f] = (f(n-1, z) - (n+1)*f(n, z)/z,
                        (n-1)*f(n-1, z)/z - f(n, z))

    def diffs(t, f, n, z):
        if f in t.table:
            diff0, diff1 = t.table[f]
            repl = [(t.n, n), (t.z, z)]
            return (diff0.subs(repl), diff1.subs(repl))

    def has(t, f):
        return f in t.table

_bessel_table = None

class DiffCache:
    """
    Store for derivatives of expressions.

    Explanation
    ===========

    The standard form of the derivative of a Bessel function of order n
    contains two Bessel functions of orders n-1 and n+1, respectively.
    Such forms cannot be used in parallel Risch algorithm, because
    there is a linear recurrence relation between the three functions
    while the algorithm expects that functions and derivatives are
    represented in terms of algebraically independent transcendentals.

    The solution is to take two of the functions, e.g., those of orders
    n and n-1, and to express the derivatives in terms of the pair.
    To guarantee that the proper form is used the two derivatives are
    cached as soon as one is encountered.

    Derivatives of other functions are also cached at no extra cost.
    All derivatives are with respect to the same variable `x`.
    """

    def __init__(self, x):
        self.cache = {}
        self.x = x

        global _bessel_table
        if not _bessel_table:
            _bessel_table = BesselTable()

    def get_diff(self, f):
        cache = self.cache

        if f in cache:
            pass
        elif (not hasattr(f, 'func') or
            not _bessel_table.has(f.func)):
            cache[f] = cancel(f.diff(self.x))
        else:
            n, z = f.args
            d0, d1 = _bessel_table.diffs(f.func, n, z)
            dz = self.get_diff(z)
            cache[f] = d0*dz
            cache[f.func(n-1, z)] = d1*dz

        return cache[f]

def heurisch(f, x, rewrite=False, hints=None, mappings=None, retries=3,
             degree_offset=0, unnecessary_permutations=None,
             _try_heurisch=None):
    """
    Compute indefinite integral using heuristic Risch algorithm.

    Explanation
    ===========

    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".

    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 top level
    'integrate' function in most cases, as this procedure needs some
    preprocessing steps and otherwise may fail.

    Specification
    =============

     heurisch(f, x, rewrite=False, hints=None)

       where
         f : expression
         x : symbol

         rewrite -> force rewrite 'f' in terms of 'tan' and 'tanh'
         hints   -> a list of functions that may appear in anti-derivate

          - hints = None          --> no suggestions at all
          - hints = [ ]           --> try to figure out
          - hints = [f1, ..., fn] --> we know better

    Examples
    ========

    >>> from sympy import tan
    >>> from sympy.integrals.heurisch import heurisch
    >>> from sympy.abc import x, y

    >>> heurisch(y*tan(x), x)
    y*log(tan(x)**2 + 1)/2

    See Manuel Bronstein's "Poor Man's Integrator":

    References
    ==========

    .. [1] https://www-sop.inria.fr/cafe/Manuel.Bronstein/pmint/index.html

    For more information on the implemented algorithm refer to:

    .. [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
    ========

    sympy.integrals.integrals.Integral.doit
    sympy.integrals.integrals.Integral
    sympy.integrals.heurisch.components
    """
    f = sympify(f)

    # There are some functions that Heurisch cannot currently handle,
    # so do not even try.
    # Set _try_heurisch=True to skip this check
    if _try_heurisch is not True:
        if f.has(Abs, re, im, sign, Heaviside, DiracDelta, floor, ceiling, arg):
            return

    if not f.has_free(x):
        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)
    dcache = DiffCache(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 isinstance(g, 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 isinstance(g, exp):
                        M = g.args[0].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.args[0].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.args[0].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.is_Pow:
                    if 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:
                                dF = 1/sqrt(M[a]*x**2 - M[b])
                                F = log(2*sqrt(M[a])*sqrt(M[a]*x**2 - M[b]) + 2*M[a]*x)/sqrt(M[a])
                                dcache.cache[F] = dF  # hack: F.diff(x) doesn't automatically simplify to f
                                terms.add(F)
                            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(dcache.get_diff(g), x)

    # XXX: The commented line below makes heurisch more deterministic wrt
    # PYTHONHASHSEED and the iteration order of sets. There are other places
    # where sets are iterated over but this one is possibly the most important.
    # Theoretically the order here should not matter but different orderings
    # can expose potential bugs in the different code paths so potentially it
    # is better to keep the non-determinism.
    #
    # terms = list(ordered(terms))

    # 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(dcache.get_diff(g)) 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 None

    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()

        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])

        return (S.One, p)

    special = {}

    for term in terms:
        if term.is_Function:
            if isinstance(term, tan):
                special[1 + _substitute(term)**2] = False
            elif isinstance(term, tanh):
                special[1 + _substitute(term)] = False
                special[1 - _substitute(term)] = False
            elif isinstance(term, 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 None

    #--- 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 = tuple(ordered(itermonomials(V, A + B - 1 + degree_offset)))
    else:
        monoms = tuple(ordered(itermonomials(V, A + B + degree_offset)))

    poly_coeffs = _symbols('A', len(monoms))

    poly_part = Add(*[ poly_coeffs[i]*monomial
        for i, monomial in enumerate(monoms) ])

    reducibles = set()

    for poly in ordered(polys):
        coeff, factors = factor_list(poly, *V)
        reducibles.add(coeff)
        for fact, mul in factors:
            reducibles.add(fact)

    def _integrate(field=None):
        atans = set()
        pairs = set()

        if field == 'Q':
            irreducibles = set(reducibles)
        else:
            setV = set(V)
            irreducibles = set()
            for poly in ordered(reducibles):
                zV = setV & set(iterfreeargs(poly))
                for z in ordered(zV):
                    s = set(root_factors(poly, z, filter=field))
                    irreducibles |= s
                    break

        log_part, atan_part = [], []

        for poly in ordered(irreducibles):
            m = collect(poly, I, evaluate=False)
            y = m.get(I, S.Zero)
            if y:
                x = m.get(S.One, S.Zero)
                if x.has(I) or y.has(I):
                    continue  # nontrivial x + I*y
                pairs.add((x, y))
                irreducibles.remove(poly)

        while pairs:
            x, y = pairs.pop()
            if (x, -y) in pairs:
                pairs.remove((x, -y))
                # Choosing b with no minus sign
                if y.could_extract_minus_sign():
                    y = -y
                irreducibles.add(x*x + y*y)
                atans.add(atan(x/y))
            else:
                irreducibles.add(x + I*y)


        B = _symbols('B', len(irreducibles))
        C = _symbols('C', len(atans))

        # Note: the ordering matters here
        for poly, b in reversed(list(zip(ordered(irreducibles), B))):
            if poly.has(*V):
                poly_coeffs.append(b)
                log_part.append(b * log(poly))

        for poly, c in reversed(list(zip(ordered(atans), C))):
            if poly.has(*V):
                poly_coeffs.append(c)
                atan_part.append(c * 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 cancellation is slow
        # due to slow polynomial GCD algorithms. If this gets improved then
        # revise this code.
        candidate = poly_part/poly_denom + Add(*log_part) + Add(*atan_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_free(*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 None
        else:
            ground, _ = construct_domain(non_syms, field=True)

        coeff_ring = PolyRing(poly_coeffs, ground)
        ring = PolyRing(V, coeff_ring)
        try:
            numer = ring.from_expr(raw_numer)
        except ValueError:
            raise PolynomialError
        solution = solve_lin_sys(numer.coeffs(), coeff_ring, _raw=False)

        if solution is None:
            return None
        else:
            return candidate.xreplace(solution).xreplace(
                dict(zip(poly_coeffs, [S.Zero]*len(poly_coeffs))))

    if all(isinstance(_, Symbol) for _ in V):
        more_free = F.free_symbols - set(V)
    else:
        Fd = F.as_dummy()
        more_free = Fd.xreplace(dict(zip(V, (Dummy() for _ in V)))
            ).free_symbols & Fd.free_symbols
    if not more_free:
        # all free generators are identified in 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()

        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 None