def M_formula(power, tau = True):
r"""
Generate the formula for the conditional moments with second-order
corrections based on the relation with the ordinary Bell polynomials
.. math::
M_n(x^{\prime},\tau) \sim (n!)\tau D_n(x^{\prime}) + \frac{(n!)\tau^2}{2}
\sum_{m=1}^{n-1} D_m(x^{\prime}) D_{n-m}(x^{\prime})
Parameters
----------
power: int
Desired order of the formula.
Returns
-------
term: sympy.symbols
Expression up to given ``power``.
"""
init_sym = symbols('D:'+str(int(power+1)))[1:]
sym = ()
for i in range(1,power + 1):
sym += (factorial(i)*init_sym[i-1],)
if tau == True:
t = symbols('tau')
term = t*bell(power, 1, sym) + t**2*bell(power, 2, sym)
else:
term = bell(power, 1, sym) + bell(power, 2, sym)
return term
def test_bell():
assert [bell(n) for n in range(8)] == [1, 1, 2, 5, 15, 52, 203, 877]
assert bell(0, x) == 1
assert bell(1, x) == x
assert bell(2, x) == x**2 + x
assert bell(5, x) == x**5 + 10 * x**4 + 25 * x**3 + 15 * x**2 + x
X = symbols('x:6')
# X = (x0, x1, .. x5)
# at the same time: X[1] = x1, X[2] = x2 for standard readablity.
# but we must supply zero-based indexed object X[1:] = (x1, .. x5)
assert bell(6, 2,
X[1:]) == 6 * X[5] * X[1] + 15 * X[4] * X[2] + 10 * X[3]**2
assert bell(
6, 3,
X[1:]) == 15 * X[4] * X[1]**2 + 60 * X[3] * X[2] * X[1] + 15 * X[2]**3
X = (1, 10, 100, 1000, 10000)
assert bell(6, 2, X) == (6 + 15 + 10) * 10000
X = (1, 2, 3, 3, 5)
assert bell(6, 2, X) == 6 * 5 + 15 * 3 * 2 + 10 * 3**2
X = (1, 2, 3, 5)
assert bell(6, 3, X) == 15 * 5 + 60 * 3 * 2 + 15 * 2**3
def test_bell():
assert [bell(n) for n in range(8)] == [1, 1, 2, 5, 15, 52, 203, 877]
assert bell(0, x) == 1
assert bell(1, x) == x
assert bell(2, x) == x**2 + x
assert bell(5, x) == x**5 + 10*x**4 + 25*x**3 + 15*x**2 + x
X = symbols('x:6')
# X = (x0, x1, .. x5)
# at the same time: X[1] = x1, X[2] = x2 for standard readablity.
# but we must supply zero-based indexed object X[1:] = (x1, .. x5)
assert bell(6, 2, X[1:]) == 6*X[5]*X[1] + 15*X[4]*X[2] + 10*X[3]**2
assert bell(
6, 3, X[1:]) == 15*X[4]*X[1]**2 + 60*X[3]*X[2]*X[1] + 15*X[2]**3
X = (1, 10, 100, 1000, 10000)
assert bell(6, 2, X) == (6 + 15 + 10)*10000
X = (1, 2, 3, 3, 5)
assert bell(6, 2, X) == 6*5 + 15*3*2 + 10*3**2
X = (1, 2, 3, 5)
assert bell(6, 3, X) == 15*5 + 60*3*2 + 15*2**3
def test_bell():
assert [bell(n) for n in range(8)] == [1, 1, 2, 5, 15, 52, 203, 877]
assert bell(0, x) == 1
assert bell(1, x) == x
assert bell(2, x) == x**2 + x
assert bell(5, x) == x**5 + 10 * x**4 + 25 * x**3 + 15 * x**2 + x
def test_bell():
assert [bell(n) for n in range(8)] == [1, 1, 2, 5, 15, 52, 203, 877]
assert bell(0, x) == 1
assert bell(1, x) == x
assert bell(2, x) == x**2 + x
assert bell(5, x) == x**5 + 10*x**4 + 25*x**3 + 15*x**2 + x
def F_formula(power):
r"""
Generate the formula for the conditional moments with second-order
corrections based on the relation with the ordinary Bell polynomials
.. math::
D_n(x) &= \frac{1}{\tau (n!)} \bigg[ \hat{B}_{n,1}
\left(M_1(x,\tau),M_2(x,\tau),\ldots,M_{n}(x,\tau)\right) \\
&\qquad \left.-\frac{\tau}{2} \hat{B}_{n,2}
\left(M_1(x,\tau),M_2(x,\tau),\ldots,M_{n-1}(x,\tau)\right)\right].
Parameters
----------
power: int
Desired order of the formula.
Returns
-------
term: sympy.symbols
Expression up to given ``power``.
"""
init_sym = symbols('D:'+str(int(power+1)))[1:]
sym = ()
for i in range(1,power + 1):
sym += (factorial(i)*init_sym[i-1],)
term = (symbols('M'+str(int(power))) - bell(power, 2, sym))/factorial(power)
return term
def get_n_der_vecs(dk_f, gx, N):
"""
This function applies Faa di Bruno's formula to compute the derivatives of order from 1 to N
given the calling elementary operator has dk_f as its kth order derivative function.
INPUTS
=======
dk_f(val, k): A lambda function of the kth order derivative of f at the point val
gx: Potentially an AutoDiff object
N: highest derivative order of gx
RETURNS
=======
a list of high-order derivatives up until gx.N
"""
# Create symbols and symbol-value mapping for eval() in the loop
dxs = symbols('q:%d' % N)
dx_mapping = {str(dxs[i]): gx.der[i] for i in range(N)}
# Use Faa di Bruno's formula
der_new = []
for n in range(1, N + 1):
nth_der = 0
for k in range(1, n + 1):
# The first n-k+1 derivatives
t = n - k + 1
vars = dxs[:t]
# bell polynomial as python function str
bell_nk_str = str(bell(n, k, vars))
# evaluate the bell polynomial using the symbol-value mapping
val_bell_nk = eval(bell_nk_str, dx_mapping)
nth_der += dk_f(gx.val, k) * val_bell_nk
der_new.append(nth_der)
return der_new
def bellyn(self, nmax):
Lbell = [[0] * nmax for i in range(nmax)]
Lbell[0][0] = 0
for n in range(1, nmax):
for k in range(1, n + 1):
Lbell[n][k] = bell(n, k, self.kminus)
return Lbell
def kappac(self, n):
fn = self.fn
r = self.r
z = self.z
y = self.y
W = self.W
Q = self.Q
S = self.S
NBc = self.NBc
NBbc = self.NBbc
j = self.j
L1 = [fn[0] * bell(n, 1, (r))]
for k in range(2, n + 1):
L1.extend([fn[k - 1] * bell(n, k, r)])
L2 = Sum(Indexed('y', j), (j, 0, n - 1))
L22 = lambdify(y, L2)
L3 = L22(L1)
L4 = L3.subs({W(z): Q(z) * S(z) - NBc * NBbc})
L5 = L4.subs({S(z): NBc + NBbc, Q(z): z * z - NBc * NBbc})
return L5
def _derivative_transformation_matrix(deriv_func_list: list, point: float,
order: int):
r"""Compute transformation from one variable to another.
Suppose there is a function :math:`f(x)` and a :math:`N`-th differentiable
transformation :math:`g(x) = r` from variable :math:`x` to variable :math:`r`.
This function returns a matrix that converts the derivatives
:math:`[\frac{df}{dx}, \cdots, \frac{d^N f}{dx^N}]^T` by matrix multiplication
to the derivatives of the new transformation
:math:`[\frac{df}{dr}, \cdots, \frac{d^Nf}{dr^N}]^T`.
It is a matrix with (i,j)-th entries:
.. math::
m_{i, j} = \begin{cases}
B_{i,j}\bigg(\frac{dg}{dx}, \cdots, \frac{d^{k-j+1} g}{dx^{k-j+1}} \bigg) &
i \leq j \\
0 & \text{else}
\end{cases}
Parameters
----------
deriv_func_list : list[K]
List of size :math:`K` callable functions representing the derivatives of
:math:`g(x)`, i.e. :math:`[\frac{dg}{dx}, \cdots, \frac{d^K g}{dx^K}]`.
point : float
The point in :math:`r`, in which the derivatives are being calculated at.
order : int
The number of derivatives from 1 to `order` that are going to be transformed.
Returns
-------
ndarray((order, order))
Returns the matrix that converts derivatives of one domain to another.
"""
numb_derivs = len(deriv_func_list)
if order > numb_derivs:
raise ValueError("TODO")
# Calculate derivatives of transformation evaluated at the point
derivs_at_pt = np.array([dev(point) for dev in deriv_func_list],
dtype=np.float64)
deriv_transf = np.zeros((order, order))
for i in range(0, order):
for j in range(0, i + 1):
deriv_transf[i, j] = float(bell(i + 1, j + 1, derivs_at_pt))
return deriv_transf
def chain_rule(ad, new_val, der, der2, higher_der=None):
"""
Applies chain rule to returns a new AD object with correct value and derivatives.
Parameters:
ad (AD): An AD object
new_val (float): Value of the new AD object
der (float): Derivative of the outer function in chain rule
Returns:
new_ad (AD): a new AD object with correct value and derivatives
Example:
>>> import AD as AD
>>> x = AD.AD(2,order=5)
>>> newad = 5*x**3+2
>>> higherde = np.array([60,60,30,0,0])
>>> chain_rule(x, 42, 60, 60, higher_der=higherde)
AD(value: [42], derivatives: [60.])
"""
new_der = der * ad.der
new_der2 = der * ad.der2 + der2 * np.matmul(
np.array([ad.der]).T, np.array([ad.der]))
if ad.higher is None:
new_ad = AD.AD(new_val, tag=ad.tag, der=new_der, der2=new_der2)
else:
new_higher_der = np.array([0.0] * len(ad.higher))
new_higher_der[0] = new_der
new_higher_der[1] = new_der2
for i in range(2, len(ad.higher)):
n = i + 1
sum = 0
for k in range(1, n + 1):
sum += higher_der[k - 1] * sp.bell(n, k,
ad.higher[0:n - k + 1])
new_higher_der[i] = sum
new_ad = AD.AD(new_val,
tag=ad.tag,
der=new_der,
der2=new_der2,
order=len(ad.higher))
new_ad.higher = new_higher_der
return new_ad
def grand_potential_derivative(self, n_elec, order=1):
r"""
Evaluate the :math:`n^{\text{th}}`-order derivative of grand potential at the given n_elec.
This returns the :math:`n^{\text{th}}`-order derivative of grand potential model w.r.t.
to the chemical potential, at fixed external potential, evaluated for the specified
number of electrons :math:`N_{\text{elec}}`.
That is,
.. math::
\left. \left(\frac{\partial^n \Omega}{\partial \mu^n}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}} =
\left. \left(\frac{\partial^{n-1}}{\partial \mu^{n-1}}
\frac{\partial \Omega}{\partial \mu}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}} =
- \left. \left(\frac{\partial^{n-1} N}{\partial \mu^{n-1}}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}}
\quad n = 1, 2, \dots
These derivatives can be computed using the derivative of energy model w.r.t. number of
electrons, at fixed external potential, evaluated at :math:`N_{\text{elec}}`.
More specifically,
.. math::
\left. \left(\frac{\partial \Omega}{\partial \mu}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}} &= - N_{\text{elec}} \\
\left. \left(\frac{\partial^2 \Omega}{\partial \mu^2}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}} &= -\frac{1}{\eta^{(1)}}
where :math:`\eta^{(n)}` denotes the :math:`(n+1)^{\text{th}}`-order derivative of energy
w.r.t. number of electrons evaluated at :math:`N_{\text{elec}}`, i.e.
.. math::
\eta^{(n)} = \left. \left(\frac{\partial^{n+1} E}{\partial N^{n+1}}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}}
\qquad n = 1, 2, \dots
To compute higher-order derivatives, Faa di Bruno formula which generalizes the chain rule
to higher derivatives can be used. i.e. for :math:`n \geq 2`,
.. math::
\left. \left(\frac{\partial^n \Omega}{\partial \mu^n}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}} =
\frac{-\displaystyle\sum_{k=1}^{n-2} \left.\left(\frac{\partial^k \Omega}{\partial \mu^k}
\right)_{v(\mathbf{r})} \right|_{N = N_{\text{elec}}}
\cdot B_{n-1,k} \left(\eta^{(1)}, \eta^{(2)}, \dots, \eta^{(n-k)} \right)}
{B_{n-1,n-1} \left(\eta^{(1)}\right)}
where :math:`B_{n-1,k} \left(x_1, x_2, \dots, x_{n-k}\right)` denotes the Bell polynomials.
Parameters
----------
n_elec : float
Number of electrons, :math:`N_{\text{elec}}`.
order : int, default=1
The order of derivative denoted by :math:`n` in the formula.
"""
if n_elec is not None and n_elec < 0.0:
raise ValueError(
'Number of electrons cannot be negativ! #elec={0}'.format(
n_elec))
if not (isinstance(order, int) and order > 0):
raise ValueError(
'Argument order should be an integer greater than or equal to 1.'
)
if n_elec is None:
deriv = None
elif order == 1:
# 1st order derivative is minus number of electrons
deriv = -n_elec
elif order == 2:
# 2nd order derivative is inverse hardness
hardness = self.energy_derivative(n_elec, order=2)
if hardness is not None and hardness != 0.0:
deriv = -1.0 / hardness
else:
deriv = None
else:
# higher-order derivatives are compute with Faa Di Bruno formula
# list of hyper-hardneses (derivatives of energy w.r.t. N)
e_deriv = [
self.energy_derivative(n_elec, i + 1)
for i in xrange(1, order)
]
g_deriv = [
self.grand_potential_derivative(n_elec, k + 1)
for k in xrange(1, order - 1)
]
if any([item is None for item in e_deriv]) or any(
[item is None for item in g_deriv]):
deriv = None
else:
deriv = 0
for k in xrange(1, order - 1):
deriv -= g_deriv[k - 1] * sp.bell(order - 1, k,
e_deriv[:order - k])
deriv /= sp.bell(order - 1, order - 1, [e_deriv[0]])
return deriv
def count(smilesdata,
output='lib_data.txt',
timeout=5,
symmetry=True,
sample_size=8000):
'''
Count...?
'''
x = []
compression = []
#counter = 0
line_count = 0
#sample_size=int(sample_size)
for file_lines in open(smilesdata).readlines():
line_count += 1
if line_count <= sample_size:
sample_idx = np.arange(line_count)
else:
sample_idx = np.random.choice(line_count, sample_size, replace=False)
f = open(output, "w")
f.write(
'#SMILES heavy_atoms bond_number bell_number naive_count starsbars symmetry_count MOG_nodes \n'
)
with open(smilesdata) as infile:
for counter, line in enumerate(infile):
if counter not in sample_idx:
continue
smiles = line.split()[1]
#print(smiles)
try:
G = smiles2graph(smiles)
except:
continue
mol = rdkit.Chem.MolFromSmiles(smiles)
heavy_atoms = mol.GetNumHeavyAtoms()
atoms = len(G)
edges = G.number_of_edges()
LG = chem_line_graph(G)
bond_classes = equiv_classes(LG)
atom_classes = equiv_classes(G,
node_key='atom_type',
edge_key='bond')
bellnum = bell(atoms) - 1
naive_count = (2**edges) - 1
product = 1
for i in range(len(bond_classes)):
product *= (len(bond_classes[i]) + 1)
stars_bars = product - 1
symmetric_counts = (2**(len(bond_classes))) - 1
#print(G.number_of_edges(),len(atom_classes))
if (edges == 0) or (len(atom_classes) == 1):
continue
try:
mog = MOG(smiles, symmetry, timeout)
f.write(
'{smiles} {h_atoms} {bond_number} {bell} {naive} {starsbars} {symmetric} {mog_nodes} \n'
.format(smiles=smiles,
h_atoms=heavy_atoms,
bond_number=edges,
bell=bellnum,
naive=naive_count,
starsbars=stars_bars,
symmetric=symmetric_counts,
mog_nodes=len(mog.graph)))
except MOG.TimeoutError:
f.write(
'{smiles} {h_atoms} {bond_number} {bell} {naive} {starsbars} {symmetric} {mog_nodes} \n'
.format(smiles=smiles,
h_atoms=heavy_atoms,
bond_number=edges,
bell=bellnum,
naive=naive_count,
starsbars=stars_bars,
symmetric=symmetric_counts,
mog_nodes='NA'))
continue
f.close()
def _transform_ode_from_derivs(coeffs: Union[list, np.ndarray],
deriv_transformation: list, x: np.ndarray):
r"""
Transform the coefficients of ODE from one variable to another evaluated on mesh points.
Given a :math:`K`-th differentiable transformation function :math:`g(x)`
and a linear ODE of :math:`K`-th order on independent variable :math:`x`
.. math::
\sum_{k=1}^{K} a_k(x) \frac{d^k y(x)}{d x^k}.
This transforms it into a new coordinate system :math:`g(x) =: r`
.. math::
\sum_{j=1}^K b_j(r) \frac{d^j y(r)}{d r^j},
where :math:`b_j(r) = \sum_{k=1}^K a_k(g^{-1}(r)) B_{k, j}({g^{-1}}^\prime(g^{-1}(r)),
\cdots, {g^{-1}}^{k - j + 1}(g^{-1}(r)))`.
Parameters
----------
coeffs : list[callable or number] or ndarray
Coefficients :math:`a_k` of each term :math:`\frac{d^k y(x}{d x^k}`
ordered from 0 to K. Either a list of callable functions :math:`a_k(x)` that depends
on :math:`x` or array of constants :math:`\{a_k\}_{k=0}^{K}`.
deriv_transformation : list[callable]
List of functions for compute transformation derivatives from 1 to :math:`K`.
x : ndarray
Points from the original domain that is to be transformed.
Returns
-------
ndarray((K+1, N))
Coefficients :math:`b_j(r)` of the new ODE with respect to transformed variable :math:`r`.
"""
# `derivs` has shape (K, N), calculate d^j g/ dr^j
derivs = np.array([dev(x) for dev in deriv_transformation],
dtype=np.float64)
total = len(coeffs) # Should be K+1, K is the order of the ODE
# coeff_a_mtr has shape (K+1, N)
coeff_a_mtr = _evaluate_coeffs_on_points(x, coeffs) # a_k(x)
coeff_b = np.zeros((total, x.size), dtype=float)
# The first term doesn't contain a derivative so no transformation is required:
coeff_b[0] += coeff_a_mtr[0]
# construct 1 - 3 directly with vectorization (faster)
if total > 1:
coeff_b[1] += coeff_a_mtr[1] * derivs[0]
if total > 2:
coeff_b[1] += coeff_a_mtr[2] * derivs[1]
coeff_b[2] += coeff_a_mtr[2] * derivs[0]**2
if total > 3:
coeff_b[1] += coeff_a_mtr[3] * derivs[2]
coeff_b[2] += coeff_a_mtr[3] * 3 * derivs[0] * derivs[1]
coeff_b[3] += coeff_a_mtr[3] * derivs[0]**3
# construct 4th order and onwards without vectorization (slower)
# formula is: d^k f / dr^k = \sum_{j=1}^{k} Bell_{k, j}(...) (d^jf / dx^j)
if total > 4:
# Go Through each Pt
for i_pt in range(len(x)):
# Go through each order from 4 to K + 1
for j in range(4, total):
# Go through the sum to calculate Bell's polynomial
for k in range(j, total):
all_derivs_at_pt = derivs[:, i_pt]
coeff_b[j, i_pt] += (float(bell(k, j, all_derivs_at_pt)) *
coeff_a_mtr[k, i_pt])
return coeff_b