def compute_quadratic_function_biggest_stable_learning_rate(
f, threshold, *func_vars):
"""Computes the biggest stable (steepest descent) learning rate for a quadratic function.
Args:
f (function): quadratic function.
threshold (float): maximum distance from the best learning rate.
*func_vars (sympy.Symbol): a list of function variables. the derivative of the function
f will be computed with respect to these parameters.
Returns:
alpha (float): the biggest stable learning rate.
Examples:
>>> def f1(x1, x2): return x1 ** 2 + 25 * x2 ** 2
>>> x1, x2 = symbols('x1, x2')
>>> compute_quadratic_function_biggest_stable_learning_rate(f1, 0.001, x1, x2)
0.039
"""
gradient = derive_by_array(f(*func_vars), func_vars)
hessian = derive_by_array(gradient, func_vars).tomatrix()
S = mp.svd(hessian, compute_uv=False)
alpha = 2. / float(max(S)) - threshold
return alpha
def _mp_svd(A, full_matrices=True):
"""Convenience wrapper for mpmath high-precision SVD."""
assert mpmath, 'mpmath package is not installed'
AA = mp.matrix(A.tolist())
U, Sigma, VT = mp.svd(AA, full_matrices=full_matrices)
return np.array(U.tolist()), np.array(Sigma.tolist()).ravel(), np.array(
VT.tolist())
def preserves_hermitian_form(SL2C_matrices):
"""
>>> CC = ComplexField(100)
>>> A = matrix(CC, [[1, 1], [1, 2]]);
>>> B = matrix(CC, [[0, 1], [-1, 0]])
>>> C = matrix(CC, [[CC('I'),0], [0, -CC('I')]])
>>> ans, sig, form = preserves_hermitian_form([A, B])
>>> ans
True
>>> sig
'indefinite'
>>> form.change_ring(ComplexField(10))
[ 0.00 -1.0*I]
[ 1.0*I 0.00]
>>> preserves_hermitian_form([A, B, C])
(False, None, None)
>>> ans, sig, form = preserves_hermitian_form([B, C])
>>> sig
'definite'
"""
M = block_matrix(len(SL2C_matrices), 1, [
left_mult_by_adjoint(A) - right_mult_by_inverse(A)
for A in SL2C_matrices
])
CC = M.base_ring()
mp.prec = CC.prec()
RR = RealField(CC.prec())
epsilon = RR(2)**(-int(0.8 * mp.prec))
U, S, V = mp.svd(sage_matrix_to_mpmath(M))
S = list(mp.chop(S, epsilon))
if mp.zero not in S:
return False, None, None
elif S.count(mp.zero) > 1:
for i, A in enumerate(SL2C_matrices):
for B in SL2C_matrices[i + 1:]:
assert (A * B - B * A).norm() < epsilon
sig = 'indefinite' if any(abs(A.trace()) > 2
for A in SL2C_matrices) else 'both'
return True, sig, None
else:
in_kernel = list(mp.chop(V.H.column(S.index(mp.zero))))
J = mp.matrix([in_kernel[:2], in_kernel[2:]])
iJ = mp.mpc(imag=1) * J
J1, J2 = J + J.H, iJ + iJ.H
J = J1 if mp.norm(J1) >= mp.norm(J2) else J2
J = (1 / mp.sqrt(abs(mp.det(J)))) * J
J = mpmath_matrix_to_sage(J)
assert all((A.conjugate_transpose() * J * A - J).norm() < epsilon
for A in SL2C_matrices)
sig = 'definite' if J.det() > 0 else 'indefinite'
return True, sig, J
def apply(self, m, evaluation):
'SingularValueDecomposition[m_]'
if not any(leaf.is_inexact() for row in m.leaves for leaf in row.leaves):
# symbolic argument (not implemented)
evaluation.message('SingularValueDecomposition', 'nosymb')
matrix = to_mpmath_matrix(m)
U, S, V = mp.svd(matrix)
S = mp.diag(S)
U_list = Expression('List', *U.tolist())
S_list = Expression('List', *S.tolist())
V_list = Expression('List', *V.tolist())
return Expression('List', *[U_list, S_list, V_list])
def apply(self, m, evaluation):
'SingularValueDecomposition[m_?MatrixQ]'
if not any(leaf.is_inexact() for row in m.leaves for leaf in row.leaves):
# symbolic argument (not implemented)
evaluation.message('SingularValueDecomposition', 'nosymb')
matrix = to_mpmath_matrix(m)
U, S, V = mp.svd(matrix)
S = mp.diag(S)
U_list = Expression('List', *U.tolist())
S_list = Expression('List', *S.tolist())
V_list = Expression('List', *V.tolist())
return Expression('List', *[U_list, S_list, V_list])
def apply(self, m, evaluation):
"SingularValueDecomposition[m_]"
matrix = to_mpmath_matrix(m)
if matrix is None:
return evaluation.message("SingularValueDecomposition", "matrix",
m, 1)
if not any(leaf.is_inexact() for row in m.leaves
for leaf in row.leaves):
# symbolic argument (not implemented)
evaluation.message("SingularValueDecomposition", "nosymb")
U, S, V = mp.svd(matrix)
S = mp.diag(S)
U_list = Expression("List", *U.tolist())
S_list = Expression("List", *S.tolist())
V_list = Expression("List", *V.tolist())
return Expression("List", *[U_list, S_list, V_list])
def svd_func(_matrix):
U, S, V = mp.svd(_matrix)
smallest_eig_values = (S[len(S)-1], S[len(S)-2])
ss_raw = V[len(V)-1, :]
ss = ss_raw / fsum(ss_raw)
return ss.T, smallest_eig_values
