X-Git-Url: http://gitweb.michael.orlitzky.com/?a=blobdiff_plain;f=mjo%2Feja%2Feja_algebra.py;h=1fc3618005eb0ac8be12e0e3e09849ab6f983819;hb=ad25c5b8995a1cacefbf4d677316b9e7069521ff;hp=ff2b5d7a9c52ff5c13df7d7aa3682899b9a30559;hpb=cd164e717e7224ec1af16d31152555f0c7cf49cf;p=sage.d.git diff --git a/mjo/eja/eja_algebra.py b/mjo/eja/eja_algebra.py index ff2b5d7..1fc3618 100644 --- a/mjo/eja/eja_algebra.py +++ b/mjo/eja/eja_algebra.py @@ -54,15 +54,18 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): def __init__(self, field, mult_table, - rank, prefix='e', category=None, natural_basis=None, - check=True): + check_field=True, + check_axioms=True): """ SETUP:: - sage: from mjo.eja.eja_algebra import (JordanSpinEJA, random_eja) + sage: from mjo.eja.eja_algebra import ( + ....: FiniteDimensionalEuclideanJordanAlgebra, + ....: JordanSpinEJA, + ....: random_eja) EXAMPLES: @@ -76,22 +79,34 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): TESTS: - The ``field`` we're given must be real:: + The ``field`` we're given must be real with ``check_field=True``:: sage: JordanSpinEJA(2,QQbar) Traceback (most recent call last): ... - ValueError: field is not real + ValueError: scalar field is not real + + The multiplication table must be square with ``check_axioms=True``:: + + sage: FiniteDimensionalEuclideanJordanAlgebra(QQ,((),())) + Traceback (most recent call last): + ... + ValueError: multiplication table is not square """ - if check: + if check_field: if not field.is_subring(RR): # Note: this does return true for the real algebraic - # field, and any quadratic field where we've specified - # a real embedding. - raise ValueError('field is not real') + # field, the rationals, and any quadratic field where + # we've specified a real embedding. + raise ValueError("scalar field is not real") + + # The multiplication table had better be square + n = len(mult_table) + if check_axioms: + if not all( len(l) == n for l in mult_table ): + raise ValueError("multiplication table is not square") - self._rank = rank self._natural_basis = natural_basis if category is None: @@ -100,7 +115,7 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): fda = super(FiniteDimensionalEuclideanJordanAlgebra, self) fda.__init__(field, - range(len(mult_table)), + range(n), prefix=prefix, category=category) self.print_options(bracket='') @@ -116,6 +131,13 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): for ls in mult_table ] + if check_axioms: + if not self._is_commutative(): + raise ValueError("algebra is not commutative") + if not self._is_jordanian(): + raise ValueError("Jordan identity does not hold") + if not self._inner_product_is_associative(): + raise ValueError("inner product is not associative") def _element_constructor_(self, elt): """ @@ -194,6 +216,24 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): coords = W.coordinate_vector(_mat2vec(elt)) return self.from_vector(coords) + @staticmethod + def _max_test_case_size(): + """ + Return an integer "size" that is an upper bound on the size of + this algebra when it is used in a random test + case. Unfortunately, the term "size" is quite vague -- when + dealing with `R^n` under either the Hadamard or Jordan spin + product, the "size" refers to the dimension `n`. When dealing + with a matrix algebra (real symmetric or complex/quaternion + Hermitian), it refers to the size of the matrix, which is + far less than the dimension of the underlying vector space. + + We default to five in this class, which is safe in `R^n`. The + matrix algebra subclasses (or any class where the "size" is + interpreted to be far less than the dimension) should override + with a smaller number. + """ + return 5 def _repr_(self): """ @@ -219,168 +259,74 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): def product_on_basis(self, i, j): return self._multiplication_table[i][j] - def _a_regular_element(self): - """ - Guess a regular element. Needed to compute the basis for our - characteristic polynomial coefficients. - - SETUP:: - - sage: from mjo.eja.eja_algebra import random_eja - - TESTS: - - Ensure that this hacky method succeeds for every algebra that we - know how to construct:: - - sage: set_random_seed() - sage: J = random_eja() - sage: J._a_regular_element().is_regular() - True + def _is_commutative(self): + r""" + Whether or not this algebra's multiplication table is commutative. + This method should of course always return ``True``, unless + this algebra was constructed with ``check_axioms=False`` and + passed an invalid multiplication table. """ - gs = self.gens() - z = self.sum( (i+1)*gs[i] for i in range(len(gs)) ) - if not z.is_regular(): - raise ValueError("don't know a regular element") - return z - - - @cached_method - def _charpoly_basis_space(self): - """ - Return the vector space spanned by the basis used in our - characteristic polynomial coefficients. This is used not only to - compute those coefficients, but also any time we need to - evaluate the coefficients (like when we compute the trace or - determinant). - """ - z = self._a_regular_element() - # Don't use the parent vector space directly here in case this - # happens to be a subalgebra. In that case, we would be e.g. - # two-dimensional but span_of_basis() would expect three - # coordinates. - V = VectorSpace(self.base_ring(), self.vector_space().dimension()) - basis = [ (z**k).to_vector() for k in range(self.rank()) ] - V1 = V.span_of_basis( basis ) - b = (V1.basis() + V1.complement().basis()) - return V.span_of_basis(b) - + return all( self.product_on_basis(i,j) == self.product_on_basis(i,j) + for i in range(self.dimension()) + for j in range(self.dimension()) ) + def _is_jordanian(self): + r""" + Whether or not this algebra's multiplication table respects the + Jordan identity `(x^{2})(xy) = x(x^{2}y)`. + + We only check one arrangement of `x` and `y`, so for a + ``True`` result to be truly true, you should also check + :meth:`_is_commutative`. This method should of course always + return ``True``, unless this algebra was constructed with + ``check_axioms=False`` and passed an invalid multiplication table. + """ + return all( (self.monomial(i)**2)*(self.monomial(i)*self.monomial(j)) + == + (self.monomial(i))*((self.monomial(i)**2)*self.monomial(j)) + for i in range(self.dimension()) + for j in range(self.dimension()) ) + + def _inner_product_is_associative(self): + r""" + Return whether or not this algebra's inner product `B` is + associative; that is, whether or not `B(xy,z) = B(x,yz)`. - @cached_method - def _charpoly_coeff(self, i): - """ - Return the coefficient polynomial "a_{i}" of this algebra's - general characteristic polynomial. - - Having this be a separate cached method lets us compute and - store the trace/determinant (a_{r-1} and a_{0} respectively) - separate from the entire characteristic polynomial. - """ - (A_of_x, x, xr, detA) = self._charpoly_matrix_system() - R = A_of_x.base_ring() - - if i == self.rank(): - return R.one() - if i > self.rank(): - # Guaranteed by theory - return R.zero() - - # Danger: the in-place modification is done for performance - # reasons (reconstructing a matrix with huge polynomial - # entries is slow), but I don't know how cached_method works, - # so it's highly possible that we're modifying some global - # list variable by reference, here. In other words, you - # probably shouldn't call this method twice on the same - # algebra, at the same time, in two threads - Ai_orig = A_of_x.column(i) - A_of_x.set_column(i,xr) - numerator = A_of_x.det() - A_of_x.set_column(i,Ai_orig) - - # We're relying on the theory here to ensure that each a_i is - # indeed back in R, and the added negative signs are to make - # the whole charpoly expression sum to zero. - return R(-numerator/detA) + This method should of course always return ``True``, unless + this algebra was constructed with ``check_axioms=False`` and + passed an invalid multiplication table. + """ + # Used to check whether or not something is zero in an inexact + # ring. This number is sufficient to allow the construction of + # QuaternionHermitianEJA(2, RDF) with check_axioms=True. + epsilon = 1e-16 - @cached_method - def _charpoly_matrix_system(self): - """ - Compute the matrix whose entries A_ij are polynomials in - X1,...,XN, the vector ``x`` of variables X1,...,XN, the vector - corresponding to `x^r` and the determinent of the matrix A = - [A_ij]. In other words, all of the fixed (cachable) data needed - to compute the coefficients of the characteristic polynomial. - """ - r = self.rank() - n = self.dimension() + for i in range(self.dimension()): + for j in range(self.dimension()): + for k in range(self.dimension()): + x = self.monomial(i) + y = self.monomial(j) + z = self.monomial(k) + diff = (x*y).inner_product(z) - x.inner_product(y*z) - # Turn my vector space into a module so that "vectors" can - # have multivatiate polynomial entries. - names = tuple('X' + str(i) for i in range(1,n+1)) - R = PolynomialRing(self.base_ring(), names) - - # Using change_ring() on the parent's vector space doesn't work - # here because, in a subalgebra, that vector space has a basis - # and change_ring() tries to bring the basis along with it. And - # that doesn't work unless the new ring is a PID, which it usually - # won't be. - V = FreeModule(R,n) - - # Now let x = (X1,X2,...,Xn) be the vector whose entries are - # indeterminates... - x = V(names) - - # And figure out the "left multiplication by x" matrix in - # that setting. - lmbx_cols = [] - monomial_matrices = [ self.monomial(i).operator().matrix() - for i in range(n) ] # don't recompute these! - for k in range(n): - ek = self.monomial(k).to_vector() - lmbx_cols.append( - sum( x[i]*(monomial_matrices[i]*ek) - for i in range(n) ) ) - Lx = matrix.column(R, lmbx_cols) - - # Now we can compute powers of x "symbolically" - x_powers = [self.one().to_vector(), x] - for d in range(2, r+1): - x_powers.append( Lx*(x_powers[-1]) ) - - idmat = matrix.identity(R, n) - - W = self._charpoly_basis_space() - W = W.change_ring(R.fraction_field()) - - # Starting with the standard coordinates x = (X1,X2,...,Xn) - # and then converting the entries to W-coordinates allows us - # to pass in the standard coordinates to the charpoly and get - # back the right answer. Specifically, with x = (X1,X2,...,Xn), - # we have - # - # W.coordinates(x^2) eval'd at (standard z-coords) - # = - # W-coords of (z^2) - # = - # W-coords of (standard coords of x^2 eval'd at std-coords of z) - # - # We want the middle equivalent thing in our matrix, but use - # the first equivalent thing instead so that we can pass in - # standard coordinates. - x_powers = [ W.coordinate_vector(xp) for xp in x_powers ] - l2 = [idmat.column(k-1) for k in range(r+1, n+1)] - A_of_x = matrix.column(R, n, (x_powers[:r] + l2)) - return (A_of_x, x, x_powers[r], A_of_x.det()) + if self.base_ring().is_exact(): + if diff != 0: + return False + else: + if diff.abs() > epsilon: + return False + return True @cached_method - def characteristic_polynomial(self): + def characteristic_polynomial_of(self): """ - Return a characteristic polynomial that works for all elements - of this algebra. + Return the algebra's "characteristic polynomial of" function, + which is itself a multivariate polynomial that, when evaluated + at the coordinates of some algebra element, returns that + element's characteristic polynomial. The resulting polynomial has `n+1` variables, where `n` is the dimension of this algebra. The first `n` variables correspond to @@ -400,7 +346,7 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): Alizadeh, Example 11.11:: sage: J = JordanSpinEJA(3) - sage: p = J.characteristic_polynomial(); p + sage: p = J.characteristic_polynomial_of(); p X1^2 - X2^2 - X3^2 + (-2*t)*X1 + t^2 sage: xvec = J.one().to_vector() sage: p(*xvec) @@ -413,27 +359,28 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): any argument:: sage: J = TrivialEJA() - sage: J.characteristic_polynomial() + sage: J.characteristic_polynomial_of() 1 """ r = self.rank() n = self.dimension() - # The list of coefficient polynomials a_0, a_1, a_2, ..., a_n. - a = [ self._charpoly_coeff(i) for i in range(r+1) ] + # The list of coefficient polynomials a_0, a_1, a_2, ..., a_(r-1). + a = self._charpoly_coefficients() # We go to a bit of trouble here to reorder the # indeterminates, so that it's easier to evaluate the # characteristic polynomial at x's coordinates and get back # something in terms of t, which is what we want. - R = a[0].parent() S = PolynomialRing(self.base_ring(),'t') t = S.gen(0) - S = PolynomialRing(S, R.variable_names()) - t = S(t) + if r > 0: + R = a[0].parent() + S = PolynomialRing(S, R.variable_names()) + t = S(t) - return sum( a[k]*(t**k) for k in range(len(a)) ) + return (t**r + sum( a[k]*(t**k) for k in range(r) )) def inner_product(self, x, y): @@ -578,8 +525,15 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): """ Return the matrix space in which this algebra's natural basis elements live. + + Generally this will be an `n`-by-`1` column-vector space, + except when the algebra is trivial. There it's `n`-by-`n` + (where `n` is zero), to ensure that two elements of the + natural basis space (empty matrices) can be multiplied. """ - if self._natural_basis is None or len(self._natural_basis) == 0: + if self.is_trivial(): + return MatrixSpace(self.base_ring(), 0) + elif self._natural_basis is None or len(self._natural_basis) == 0: return MatrixSpace(self.base_ring(), self.dimension(), 1) else: return self._natural_basis[0].matrix_space() @@ -712,6 +666,25 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): Vector space of degree 6 and dimension 2... sage: J1 Euclidean Jordan algebra of dimension 3... + sage: J0.one().natural_representation() + [0 0 0] + [0 0 0] + [0 0 1] + sage: orig_df = AA.options.display_format + sage: AA.options.display_format = 'radical' + sage: J.from_vector(J5.basis()[0]).natural_representation() + [ 0 0 1/2*sqrt(2)] + [ 0 0 0] + [1/2*sqrt(2) 0 0] + sage: J.from_vector(J5.basis()[1]).natural_representation() + [ 0 0 0] + [ 0 0 1/2*sqrt(2)] + [ 0 1/2*sqrt(2) 0] + sage: AA.options.display_format = orig_df + sage: J1.one().natural_representation() + [1 0 0] + [0 1 0] + [0 0 0] TESTS: @@ -726,9 +699,10 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): sage: J1.superalgebra() == J and J1.dimension() == J.dimension() True - The identity elements in the two subalgebras are the - projections onto their respective subspaces of the - superalgebra's identity element:: + The decomposition is into eigenspaces, and its components are + therefore necessarily orthogonal. Moreover, the identity + elements in the two subalgebras are the projections onto their + respective subspaces of the superalgebra's identity element:: sage: set_random_seed() sage: J = random_eja() @@ -738,6 +712,16 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): ....: x = J.random_element() sage: c = x.subalgebra_idempotent() sage: J0,J5,J1 = J.peirce_decomposition(c) + sage: ipsum = 0 + sage: for (w,y,z) in zip(J0.basis(), J5.basis(), J1.basis()): + ....: w = w.superalgebra_element() + ....: y = J.from_vector(y) + ....: z = z.superalgebra_element() + ....: ipsum += w.inner_product(y).abs() + ....: ipsum += w.inner_product(z).abs() + ....: ipsum += y.inner_product(z).abs() + sage: ipsum + 0 sage: J1(c) == J1.one() True sage: J0(J.one() - c) == J0.one() @@ -762,7 +746,9 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): J5 = eigspace else: gens = tuple( self.from_vector(b) for b in eigspace.basis() ) - subalg = FiniteDimensionalEuclideanJordanSubalgebra(self, gens) + subalg = FiniteDimensionalEuclideanJordanSubalgebra(self, + gens, + check_axioms=False) if eigval == 0: J0 = subalg elif eigval == 1: @@ -773,10 +759,61 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): return (J0, J5, J1) - def random_elements(self, count): + def random_element(self, thorough=False): + r""" + Return a random element of this algebra. + + Our algebra superclass method only returns a linear + combination of at most two basis elements. We instead + want the vector space "random element" method that + returns a more diverse selection. + + INPUT: + + - ``thorough`` -- (boolean; default False) whether or not we + should generate irrational coefficients for the random + element when our base ring is irrational; this slows the + algebra operations to a crawl, but any truly random method + should include them + + """ + # For a general base ring... maybe we can trust this to do the + # right thing? Unlikely, but. + V = self.vector_space() + v = V.random_element() + + if self.base_ring() is AA: + # The "random element" method of the algebraic reals is + # stupid at the moment, and only returns integers between + # -2 and 2, inclusive: + # + # https://trac.sagemath.org/ticket/30875 + # + # Instead, we implement our own "random vector" method, + # and then coerce that into the algebra. We use the vector + # space degree here instead of the dimension because a + # subalgebra could (for example) be spanned by only two + # vectors, each with five coordinates. We need to + # generate all five coordinates. + if thorough: + v *= QQbar.random_element().real() + else: + v *= QQ.random_element() + + return self.from_vector(V.coordinate_vector(v)) + + def random_elements(self, count, thorough=False): """ Return ``count`` random elements as a tuple. + INPUT: + + - ``thorough`` -- (boolean; default False) whether or not we + should generate irrational coefficients for the random + elements when our base ring is irrational; this slows the + algebra operations to a crawl, but any truly random method + should include them + SETUP:: sage: from mjo.eja.eja_algebra import JordanSpinEJA @@ -791,25 +828,85 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): True """ - return tuple( self.random_element() for idx in range(count) ) + return tuple( self.random_element(thorough) + for idx in range(count) ) + + @classmethod + def random_instance(cls, field=AA, **kwargs): + """ + Return a random instance of this type of algebra. + + Beware, this will crash for "most instances" because the + constructor below looks wrong. + """ + if cls is TrivialEJA: + # The TrivialEJA class doesn't take an "n" argument because + # there's only one. + return cls(field) + + n = ZZ.random_element(cls._max_test_case_size() + 1) + return cls(n, field, **kwargs) @cached_method - def rank(self): + def _charpoly_coefficients(self): + r""" + The `r` polynomial coefficients of the "characteristic polynomial + of" function. """ - Return the rank of this EJA. + n = self.dimension() + var_names = [ "X" + str(z) for z in range(1,n+1) ] + R = PolynomialRing(self.base_ring(), var_names) + vars = R.gens() + F = R.fraction_field() - ALGORITHM: + def L_x_i_j(i,j): + # From a result in my book, these are the entries of the + # basis representation of L_x. + return sum( vars[k]*self.monomial(k).operator().matrix()[i,j] + for k in range(n) ) - We first compute the polynomial "column matrices" `p_{k}` that - evaluate to `x^k` on the coordinates of `x`. Then, we begin - adding them to a matrix one at a time, and trying to solve the - system that makes `p_{0}`,`p_{1}`,..., `p_{s-1}` add up to - `p_{s}`. This will succeed only when `s` is the rank of the - algebra, as proven in a recent draft paper of mine. + L_x = matrix(F, n, n, L_x_i_j) + + r = None + if self.rank.is_in_cache(): + r = self.rank() + # There's no need to pad the system with redundant + # columns if we *know* they'll be redundant. + n = r + + # Compute an extra power in case the rank is equal to + # the dimension (otherwise, we would stop at x^(r-1)). + x_powers = [ (L_x**k)*self.one().to_vector() + for k in range(n+1) ] + A = matrix.column(F, x_powers[:n]) + AE = A.extended_echelon_form() + E = AE[:,n:] + A_rref = AE[:,:n] + if r is None: + r = A_rref.rank() + b = x_powers[r] + + # The theory says that only the first "r" coefficients are + # nonzero, and they actually live in the original polynomial + # ring and not the fraction field. We negate them because + # in the actual characteristic polynomial, they get moved + # to the other side where x^r lives. + return -A_rref.solve_right(E*b).change_ring(R)[:r] + + @cached_method + def rank(self): + r""" + Return the rank of this EJA. + + This is a cached method because we know the rank a priori for + all of the algebras we can construct. Thus we can avoid the + expensive ``_charpoly_coefficients()`` call unless we truly + need to compute the whole characteristic polynomial. SETUP:: - sage: from mjo.eja.eja_algebra import (JordanSpinEJA, + sage: from mjo.eja.eja_algebra import (HadamardEJA, + ....: JordanSpinEJA, ....: RealSymmetricEJA, ....: ComplexHermitianEJA, ....: QuaternionHermitianEJA, @@ -885,45 +982,8 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): sage: J.rank.clear_cache() sage: J.rank() 2 - """ - n = self.dimension() - if n == 0: - return 0 - elif n == 1: - return 1 - - var_names = [ "X" + str(z) for z in range(1,n+1) ] - R = PolynomialRing(self.base_ring(), var_names) - vars = R.gens() - - def L_x_i_j(i,j): - # From a result in my book, these are the entries of the - # basis representation of L_x. - return sum( vars[k]*self.monomial(k).operator().matrix()[i,j] - for k in range(n) ) - - L_x = matrix(R, n, n, L_x_i_j) - x_powers = [ vars[k]*(L_x**k)*self.one().to_vector() - for k in range(n) ] - - # Can assume n >= 2 - M = matrix([x_powers[0]]) - old_rank = 1 - - for d in range(1,n): - M = matrix(M.rows() + [x_powers[d]]) - M.echelonize() - # TODO: we've basically solved the system here. - # We should save the echelonized matrix somehow - # so that it can be reused in the charpoly method. - new_rank = M.rank() - if new_rank == old_rank: - return new_rank - else: - old_rank = new_rank - - return n + return len(self._charpoly_coefficients()) def vector_space(self): @@ -947,134 +1007,8 @@ class FiniteDimensionalEuclideanJordanAlgebra(CombinatorialFreeModule): Element = FiniteDimensionalEuclideanJordanAlgebraElement -class KnownRankEJA(object): - """ - A class for algebras that we actually know we can construct. The - main issue is that, for most of our methods to make sense, we need - to know the rank of our algebra. Thus we can't simply generate a - "random" algebra, or even check that a given basis and product - satisfy the axioms; because even if everything looks OK, we wouldn't - know the rank we need to actuallty build the thing. - - Not really a subclass of FDEJA because doing that causes method - resolution errors, e.g. - - TypeError: Error when calling the metaclass bases - Cannot create a consistent method resolution - order (MRO) for bases FiniteDimensionalEuclideanJordanAlgebra, - KnownRankEJA - - """ - @staticmethod - def _max_test_case_size(): - """ - Return an integer "size" that is an upper bound on the size of - this algebra when it is used in a random test - case. Unfortunately, the term "size" is quite vague -- when - dealing with `R^n` under either the Hadamard or Jordan spin - product, the "size" refers to the dimension `n`. When dealing - with a matrix algebra (real symmetric or complex/quaternion - Hermitian), it refers to the size of the matrix, which is - far less than the dimension of the underlying vector space. - - We default to five in this class, which is safe in `R^n`. The - matrix algebra subclasses (or any class where the "size" is - interpreted to be far less than the dimension) should override - with a smaller number. - """ - return 5 - - @classmethod - def random_instance(cls, field=AA, **kwargs): - """ - Return a random instance of this type of algebra. - - Beware, this will crash for "most instances" because the - constructor below looks wrong. - """ - if cls is TrivialEJA: - # The TrivialEJA class doesn't take an "n" argument because - # there's only one. - return cls(field) - - n = ZZ.random_element(cls._max_test_case_size()) + 1 - return cls(n, field, **kwargs) - - -class HadamardEJA(FiniteDimensionalEuclideanJordanAlgebra, KnownRankEJA): - """ - Return the Euclidean Jordan Algebra corresponding to the set - `R^n` under the Hadamard product. - - Note: this is nothing more than the Cartesian product of ``n`` - copies of the spin algebra. Once Cartesian product algebras - are implemented, this can go. - - SETUP:: - - sage: from mjo.eja.eja_algebra import HadamardEJA - - EXAMPLES: - - This multiplication table can be verified by hand:: - - sage: J = HadamardEJA(3) - sage: e0,e1,e2 = J.gens() - sage: e0*e0 - e0 - sage: e0*e1 - 0 - sage: e0*e2 - 0 - sage: e1*e1 - e1 - sage: e1*e2 - 0 - sage: e2*e2 - e2 - - TESTS: - - We can change the generator prefix:: - - sage: HadamardEJA(3, prefix='r').gens() - (r0, r1, r2) - """ - def __init__(self, n, field=AA, **kwargs): - V = VectorSpace(field, n) - mult_table = [ [ V.gen(i)*(i == j) for j in range(n) ] - for i in range(n) ] - - fdeja = super(HadamardEJA, self) - return fdeja.__init__(field, mult_table, rank=n, **kwargs) - - def inner_product(self, x, y): - """ - Faster to reimplement than to use natural representations. - - SETUP:: - - sage: from mjo.eja.eja_algebra import HadamardEJA - - TESTS: - - Ensure that this is the usual inner product for the algebras - over `R^n`:: - - sage: set_random_seed() - sage: J = HadamardEJA.random_instance() - sage: x,y = J.random_elements(2) - sage: X = x.natural_representation() - sage: Y = y.natural_representation() - sage: x.inner_product(y) == J.natural_inner_product(X,Y) - True - - """ - return x.to_vector().inner_product(y.to_vector()) - - -def random_eja(field=AA, nontrivial=False): +def random_eja(field=AA): """ Return a "random" finite-dimensional Euclidean Jordan Algebra. @@ -1088,15 +1022,53 @@ def random_eja(field=AA, nontrivial=False): Euclidean Jordan algebra of dimension... """ - eja_classes = KnownRankEJA.__subclasses__() - if nontrivial: - eja_classes.remove(TrivialEJA) - classname = choice(eja_classes) + classname = choice([TrivialEJA, + HadamardEJA, + JordanSpinEJA, + RealSymmetricEJA, + ComplexHermitianEJA, + QuaternionHermitianEJA]) return classname.random_instance(field=field) +class RationalBasisEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): + r""" + Algebras whose basis consists of vectors with rational + entries. Equivalently, algebras whose multiplication tables + contain only rational coefficients. + + When an EJA has a basis that can be made rational, we can speed up + the computation of its characteristic polynomial by doing it over + ``QQ``. All of the named EJA constructors that we provide fall + into this category. + """ + @cached_method + def _charpoly_coefficients(self): + r""" + Override the parent method with something that tries to compute + over a faster (non-extension) field. + """ + if self.base_ring() is QQ: + # There's no need to construct *another* algebra over the + # rationals if this one is already over the rationals. + superclass = super(RationalBasisEuclideanJordanAlgebra, self) + return superclass._charpoly_coefficients() + + mult_table = tuple( + map(lambda x: x.to_vector(), ls) + for ls in self._multiplication_table + ) + + # Do the computation over the rationals. The answer will be + # the same, because our basis coordinates are (essentially) + # rational. + J = FiniteDimensionalEuclideanJordanAlgebra(QQ, + mult_table, + check_field=False, + check_axioms=False) + return J._charpoly_coefficients() class MatrixEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): @@ -1106,20 +1078,20 @@ class MatrixEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): # field can have dimension 4 (quaternions) too. return 2 - def __init__(self, field, basis, rank, normalize_basis=True, **kwargs): + def __init__(self, field, basis, normalize_basis=True, **kwargs): """ Compared to the superclass constructor, we take a basis instead of a multiplication table because the latter can be computed in terms of the former when the product is known (like it is here). """ - # Used in this class's fast _charpoly_coeff() override. + # Used in this class's fast _charpoly_coefficients() override. self._basis_normalizers = None # We're going to loop through this a few times, so now's a good # time to ensure that it isn't a generator expression. basis = tuple(basis) - if rank > 1 and normalize_basis: + if len(basis) > 1 and normalize_basis: # We'll need sqrt(2) to normalize the basis, and this # winds up in the multiplication table, so the whole # algebra needs to be over the field extension. @@ -1135,15 +1107,14 @@ class MatrixEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): Qs = self.multiplication_table_from_matrix_basis(basis) - fdeja = super(MatrixEuclideanJordanAlgebra, self) - return fdeja.__init__(field, - Qs, - rank=rank, - natural_basis=basis, - **kwargs) + super(MatrixEuclideanJordanAlgebra, self).__init__(field, + Qs, + natural_basis=basis, + **kwargs) - def _rank_computation(self): + @cached_method + def _charpoly_coefficients(self): r""" Override the parent method with something that tries to compute over a faster (non-extension) field. @@ -1151,52 +1122,41 @@ class MatrixEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): if self._basis_normalizers is None: # We didn't normalize, so assume that the basis we started # with had entries in a nice field. - return super(MatrixEuclideanJordanAlgebra, self)._rank_computation() + return super(MatrixEuclideanJordanAlgebra, self)._charpoly_coefficients() else: basis = ( (b/n) for (b,n) in zip(self.natural_basis(), self._basis_normalizers) ) # Do this over the rationals and convert back at the end. # Only works because we know the entries of the basis are - # integers. + # integers. The argument ``check_axioms=False`` is required + # because the trace inner-product method for this + # class is a stub and can't actually be checked. J = MatrixEuclideanJordanAlgebra(QQ, basis, - self.rank(), - normalize_basis=False) - return J._rank_computation() - - @cached_method - def _charpoly_coeff(self, i): - """ - Override the parent method with something that tries to compute - over a faster (non-extension) field. - """ - if self._basis_normalizers is None: - # We didn't normalize, so assume that the basis we started - # with had entries in a nice field. - return super(MatrixEuclideanJordanAlgebra, self)._charpoly_coeff(i) - else: - basis = ( (b/n) for (b,n) in zip(self.natural_basis(), - self._basis_normalizers) ) - - # Do this over the rationals and convert back at the end. - J = MatrixEuclideanJordanAlgebra(QQ, - basis, - self.rank(), - normalize_basis=False) - (_,x,_,_) = J._charpoly_matrix_system() - p = J._charpoly_coeff(i) - # p might be missing some vars, have to substitute "optionally" - pairs = zip(x.base_ring().gens(), self._basis_normalizers) - substitutions = { v: v*c for (v,c) in pairs } - result = p.subs(substitutions) - - # The result of "subs" can be either a coefficient-ring - # element or a polynomial. Gotta handle both cases. - if result in QQ: - return self.base_ring()(result) - else: - return result.change_ring(self.base_ring()) + normalize_basis=False, + check_field=False, + check_axioms=False) + a = J._charpoly_coefficients() + + # Unfortunately, changing the basis does change the + # coefficients of the characteristic polynomial, but since + # these are really the coefficients of the "characteristic + # polynomial of" function, everything is still nice and + # unevaluated. It's therefore "obvious" how scaling the + # basis affects the coordinate variables X1, X2, et + # cetera. Scaling the first basis vector up by "n" adds a + # factor of 1/n into every "X1" term, for example. So here + # we simply undo the basis_normalizer scaling that we + # performed earlier. + # + # The a[0] access here is safe because trivial algebras + # won't have any basis normalizers and therefore won't + # make it to this "else" branch. + XS = a[0].parent().gens() + subs_dict = { XS[i]: self._basis_normalizers[i]*XS[i] + for i in range(len(XS)) } + return tuple( a_i.subs(subs_dict) for a_i in a ) @staticmethod @@ -1214,6 +1174,9 @@ class MatrixEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): # is supposed to hold the entire long vector, and the subspace W # of V will be spanned by the vectors that arise from symmetric # matrices. Thus for S^2, dim(V) == 4 and dim(W) == 3. + if len(basis) == 0: + return [] + field = basis[0].base_ring() dimension = basis[0].nrows() @@ -1260,16 +1223,11 @@ class MatrixEuclideanJordanAlgebra(FiniteDimensionalEuclideanJordanAlgebra): Yu = cls.real_unembed(Y) tr = (Xu*Yu).trace() - if tr in RLF: - # It's real already. - return tr - - # Otherwise, try the thing that works for complex numbers; and - # if that doesn't work, the thing that works for quaternions. try: - return tr.vector()[0] # real part, imag part is index 1 + # Works in QQ, AA, RDF, et cetera. + return tr.real() except AttributeError: - # A quaternions doesn't have a vector() method, but does + # A quaternion doesn't have a real() method, but does # have coefficient_tuple() method that returns the # coefficients of 1, i, j, and k -- in that order. return tr.coefficient_tuple()[0] @@ -1293,7 +1251,7 @@ class RealMatrixEuclideanJordanAlgebra(MatrixEuclideanJordanAlgebra): return M -class RealSymmetricEJA(RealMatrixEuclideanJordanAlgebra, KnownRankEJA): +class RealSymmetricEJA(RealMatrixEuclideanJordanAlgebra): """ The rank-n simple EJA consisting of real symmetric n-by-n matrices, the usual symmetric Jordan product, and the trace inner @@ -1371,6 +1329,11 @@ class RealSymmetricEJA(RealMatrixEuclideanJordanAlgebra, KnownRankEJA): sage: x.operator().matrix().is_symmetric() True + We can construct the (trivial) algebra of rank zero:: + + sage: RealSymmetricEJA(0) + Euclidean Jordan algebra of dimension 0 over Algebraic Real Field + """ @classmethod def _denormalized_basis(cls, n, field): @@ -1411,7 +1374,11 @@ class RealSymmetricEJA(RealMatrixEuclideanJordanAlgebra, KnownRankEJA): def __init__(self, n, field=AA, **kwargs): basis = self._denormalized_basis(n, field) - super(RealSymmetricEJA, self).__init__(field, basis, n, **kwargs) + super(RealSymmetricEJA, self).__init__(field, + basis, + check_axioms=False, + **kwargs) + self.rank.set_cache(n) class ComplexMatrixEuclideanJordanAlgebra(MatrixEuclideanJordanAlgebra): @@ -1573,7 +1540,7 @@ class ComplexMatrixEuclideanJordanAlgebra(MatrixEuclideanJordanAlgebra): return RealMatrixEuclideanJordanAlgebra.natural_inner_product(X,Y)/2 -class ComplexHermitianEJA(ComplexMatrixEuclideanJordanAlgebra, KnownRankEJA): +class ComplexHermitianEJA(ComplexMatrixEuclideanJordanAlgebra): """ The rank-n simple EJA consisting of complex Hermitian n-by-n matrices over the real numbers, the usual symmetric Jordan product, @@ -1643,6 +1610,11 @@ class ComplexHermitianEJA(ComplexMatrixEuclideanJordanAlgebra, KnownRankEJA): sage: x.operator().matrix().is_symmetric() True + We can construct the (trivial) algebra of rank zero:: + + sage: ComplexHermitianEJA(0) + Euclidean Jordan algebra of dimension 0 over Algebraic Real Field + """ @classmethod @@ -1701,7 +1673,11 @@ class ComplexHermitianEJA(ComplexMatrixEuclideanJordanAlgebra, KnownRankEJA): def __init__(self, n, field=AA, **kwargs): basis = self._denormalized_basis(n,field) - super(ComplexHermitianEJA,self).__init__(field, basis, n, **kwargs) + super(ComplexHermitianEJA,self).__init__(field, + basis, + check_axioms=False, + **kwargs) + self.rank.set_cache(n) class QuaternionMatrixEuclideanJordanAlgebra(MatrixEuclideanJordanAlgebra): @@ -1867,8 +1843,7 @@ class QuaternionMatrixEuclideanJordanAlgebra(MatrixEuclideanJordanAlgebra): return RealMatrixEuclideanJordanAlgebra.natural_inner_product(X,Y)/4 -class QuaternionHermitianEJA(QuaternionMatrixEuclideanJordanAlgebra, - KnownRankEJA): +class QuaternionHermitianEJA(QuaternionMatrixEuclideanJordanAlgebra): """ The rank-n simple EJA consisting of self-adjoint n-by-n quaternion matrices, the usual symmetric Jordan product, and the @@ -1938,6 +1913,11 @@ class QuaternionHermitianEJA(QuaternionMatrixEuclideanJordanAlgebra, sage: x.operator().matrix().is_symmetric() True + We can construct the (trivial) algebra of rank zero:: + + sage: QuaternionHermitianEJA(0) + Euclidean Jordan algebra of dimension 0 over Algebraic Real Field + """ @classmethod def _denormalized_basis(cls, n, field): @@ -1997,10 +1977,90 @@ class QuaternionHermitianEJA(QuaternionMatrixEuclideanJordanAlgebra, def __init__(self, n, field=AA, **kwargs): basis = self._denormalized_basis(n,field) - super(QuaternionHermitianEJA,self).__init__(field, basis, n, **kwargs) + super(QuaternionHermitianEJA,self).__init__(field, + basis, + check_axioms=False, + **kwargs) + self.rank.set_cache(n) + + +class HadamardEJA(RationalBasisEuclideanJordanAlgebra): + """ + Return the Euclidean Jordan Algebra corresponding to the set + `R^n` under the Hadamard product. + + Note: this is nothing more than the Cartesian product of ``n`` + copies of the spin algebra. Once Cartesian product algebras + are implemented, this can go. + + SETUP:: + + sage: from mjo.eja.eja_algebra import HadamardEJA + + EXAMPLES: + + This multiplication table can be verified by hand:: + + sage: J = HadamardEJA(3) + sage: e0,e1,e2 = J.gens() + sage: e0*e0 + e0 + sage: e0*e1 + 0 + sage: e0*e2 + 0 + sage: e1*e1 + e1 + sage: e1*e2 + 0 + sage: e2*e2 + e2 + + TESTS: + + We can change the generator prefix:: + + sage: HadamardEJA(3, prefix='r').gens() + (r0, r1, r2) + + """ + def __init__(self, n, field=AA, **kwargs): + V = VectorSpace(field, n) + mult_table = [ [ V.gen(i)*(i == j) for j in range(n) ] + for i in range(n) ] + + super(HadamardEJA, self).__init__(field, + mult_table, + check_axioms=False, + **kwargs) + self.rank.set_cache(n) + + def inner_product(self, x, y): + """ + Faster to reimplement than to use natural representations. + + SETUP:: + + sage: from mjo.eja.eja_algebra import HadamardEJA + + TESTS: + Ensure that this is the usual inner product for the algebras + over `R^n`:: -class BilinearFormEJA(FiniteDimensionalEuclideanJordanAlgebra, KnownRankEJA): + sage: set_random_seed() + sage: J = HadamardEJA.random_instance() + sage: x,y = J.random_elements(2) + sage: X = x.natural_representation() + sage: Y = y.natural_representation() + sage: x.inner_product(y) == J.natural_inner_product(X,Y) + True + + """ + return x.to_vector().inner_product(y.to_vector()) + + +class BilinearFormEJA(RationalBasisEuclideanJordanAlgebra): r""" The rank-2 simple EJA consisting of real vectors ``x=(x0, x_bar)`` with the half-trace inner product and jordan product ``x*y = @@ -2079,8 +2139,11 @@ class BilinearFormEJA(FiniteDimensionalEuclideanJordanAlgebra, KnownRankEJA): # The rank of this algebra is two, unless we're in a # one-dimensional ambient space (because the rank is bounded # by the ambient dimension). - fdeja = super(BilinearFormEJA, self) - return fdeja.__init__(field, mult_table, rank=min(n,2), **kwargs) + super(BilinearFormEJA, self).__init__(field, + mult_table, + check_axioms=False, + **kwargs) + self.rank.set_cache(min(n,2)) def inner_product(self, x, y): r""" @@ -2171,10 +2234,10 @@ class JordanSpinEJA(BilinearFormEJA): def __init__(self, n, field=AA, **kwargs): # This is a special case of the BilinearFormEJA with the identity # matrix as its bilinear form. - return super(JordanSpinEJA, self).__init__(n, field, **kwargs) + super(JordanSpinEJA, self).__init__(n, field, **kwargs) -class TrivialEJA(FiniteDimensionalEuclideanJordanAlgebra, KnownRankEJA): +class TrivialEJA(FiniteDimensionalEuclideanJordanAlgebra): """ The trivial Euclidean Jordan algebra consisting of only a zero element. @@ -2205,7 +2268,59 @@ class TrivialEJA(FiniteDimensionalEuclideanJordanAlgebra, KnownRankEJA): """ def __init__(self, field=AA, **kwargs): mult_table = [] - fdeja = super(TrivialEJA, self) + super(TrivialEJA, self).__init__(field, + mult_table, + check_axioms=False, + **kwargs) # The rank is zero using my definition, namely the dimension of the # largest subalgebra generated by any element. - return fdeja.__init__(field, mult_table, rank=0, **kwargs) + self.rank.set_cache(0) + + +class DirectSumEJA(FiniteDimensionalEuclideanJordanAlgebra): + r""" + The external (orthogonal) direct sum of two other Euclidean Jordan + algebras. Essentially the Cartesian product of its two factors. + Every Euclidean Jordan algebra decomposes into an orthogonal + direct sum of simple Euclidean Jordan algebras, so no generality + is lost by providing only this construction. + + SETUP:: + + sage: from mjo.eja.eja_algebra import (HadamardEJA, + ....: RealSymmetricEJA, + ....: DirectSumEJA) + + EXAMPLES:: + + sage: J1 = HadamardEJA(2) + sage: J2 = RealSymmetricEJA(3) + sage: J = DirectSumEJA(J1,J2) + sage: J.dimension() + 8 + sage: J.rank() + 5 + + """ + def __init__(self, J1, J2, field=AA, **kwargs): + n1 = J1.dimension() + n2 = J2.dimension() + n = n1+n2 + V = VectorSpace(field, n) + mult_table = [ [ V.zero() for j in range(n) ] + for i in range(n) ] + for i in range(n1): + for j in range(n1): + p = (J1.monomial(i)*J1.monomial(j)).to_vector() + mult_table[i][j] = V(p.list() + [field.zero()]*n2) + + for i in range(n2): + for j in range(n2): + p = (J2.monomial(i)*J2.monomial(j)).to_vector() + mult_table[n1+i][n1+j] = V([field.zero()]*n1 + p.list()) + + super(DirectSumEJA, self).__init__(field, + mult_table, + check_axioms=False, + **kwargs) + self.rank.set_cache(J1.rank() + J2.rank())