import itertools
import numpy as np
from ase.utils import pbc2pbc
max_it = 100000 # in practice this is not exceeded
class CycleChecker:
def __init__(self, d):
assert d in [2, 3]
# worst case is the hexagonal cell in 2D and the fcc cell in 3D
n = {2: 6, 3: 12}[d]
# max cycle length is total number of primtive cell descriptions
max_cycle_length = np.prod([n - i for i in range(d)]) * np.prod(d)
self.visited = np.zeros((max_cycle_length, 3 * d), dtype=int)
def add_site(self, H):
# flatten array for simplicity
H = H.ravel()
# check if site exists
found = (self.visited == H).all(axis=1).any()
# shift all visited sites down and place current site at the top
self.visited = np.roll(self.visited, 1, axis=0)
self.visited[0] = H
return found
def reduction_gauss(B, hu, hv):
"""Calculate a Gauss-reduced lattice basis (2D reduction)."""
cycle_checker = CycleChecker(d=2)
u = hu @ B
v = hv @ B
for it in range(max_it):
x = int(round(np.dot(u, v) / np.dot(u, u)))
hu, hv = hv - x * hu, hu
u = hu @ B
v = hv @ B
site = np.array([hu, hv])
if np.dot(u, u) >= np.dot(v, v) or cycle_checker.add_site(site):
return hv, hu
raise RuntimeError(f"Gaussian basis not found after {max_it} iterations")
def relevant_vectors_2D(u, v):
cs = np.array([e for e in itertools.product([-1, 0, 1], repeat=2)])
vs = cs @ [u, v]
indices = np.argsort(np.linalg.norm(vs, axis=1))[:7]
return vs[indices], cs[indices]
def closest_vector(t0, u, v):
t = t0
a = np.zeros(2, dtype=int)
rs, cs = relevant_vectors_2D(u, v)
dprev = float("inf")
for it in range(max_it):
ds = np.linalg.norm(rs + t, axis=1)
index = np.argmin(ds)
if index == 0 or ds[index] >= dprev:
return a
dprev = ds[index]
r = rs[index]
kopt = int(round(-np.dot(t, r) / np.dot(r, r)))
a += kopt * cs[index]
t = t0 + a[0] * u + a[1] * v
raise RuntimeError(f"Closest vector not found after {max_it} iterations")
def reduction_full(B):
"""Calculate a Minkowski-reduced lattice basis (3D reduction)."""
cycle_checker = CycleChecker(d=3)
H = np.eye(3, dtype=int)
norms = np.linalg.norm(B, axis=1)
for it in range(max_it):
# Sort vectors by norm
H = H[np.argsort(norms, kind='merge')]
# Gauss-reduce smallest two vectors
hw = H[2]
hu, hv = reduction_gauss(B, H[0], H[1])
H = np.array([hu, hv, hw])
R = H @ B
# Orthogonalize vectors using Gram-Schmidt
u, v, _ = R
X = u / np.linalg.norm(u)
Y = v - X * np.dot(v, X)
Y /= np.linalg.norm(Y)
# Find closest vector to last element of R
pu, pv, pw = R @ np.array([X, Y]).T
nb = closest_vector(pw, pu, pv)
# Update basis
H[2] = [nb[0], nb[1], 1] @ H
R = H @ B
norms = np.linalg.norm(R, axis=1)
if norms[2] >= norms[1] or cycle_checker.add_site(H):
return R, H
raise RuntimeError(f"Reduced basis not found after {max_it} iterations")
[docs]def minkowski_reduce(cell, pbc=True):
"""Calculate a Minkowski-reduced lattice basis. The reduced basis
has the shortest possible vector lengths and has
norm(a) <= norm(b) <= norm(c).
Implements the method described in:
Low-dimensional Lattice Basis Reduction Revisited
Nguyen, Phong Q. and Stehlé, Damien,
ACM Trans. Algorithms 5(4) 46:1--46:48, 2009
https://doi.org/10.1145/1597036.1597050
Parameters:
cell: array
The lattice basis to reduce (in row-vector format).
pbc: array, optional
The periodic boundary conditions of the cell (Default `True`).
If `pbc` is provided, only periodic cell vectors are reduced.
Returns:
rcell: array
The reduced lattice basis.
op: array
The unimodular matrix transformation (rcell = op @ cell).
"""
pbc = pbc2pbc(pbc)
dim = pbc.sum()
op = np.eye(3, dtype=int)
if dim == 2:
perm = np.argsort(pbc, kind='merge')[::-1] # stable sort
pcell = cell[perm][:, perm]
norms = np.linalg.norm(pcell, axis=1)
norms[2] = float("inf")
indices = np.argsort(norms)
op = op[indices]
hu, hv = reduction_gauss(pcell, op[0], op[1])
op[0] = hu
op[1] = hv
invperm = np.argsort(perm)
op = op[invperm][:, invperm]
elif dim == 3:
_, op = reduction_full(cell)
# maintain cell handedness
if dim == 3:
if np.sign(np.linalg.det(cell)) != np.sign(np.linalg.det(op @ cell)):
op = -op
elif dim == 2:
index = np.argmin(pbc)
_cell = cell.copy()
_cell[index] = (1, 1, 1)
_rcell = op @ cell
_rcell[index] = (1, 1, 1)
if np.sign(np.linalg.det(_cell)) != np.sign(np.linalg.det(_rcell)):
index = np.argmax(pbc)
op[index] *= -1
norms1 = np.sort(np.linalg.norm(cell, axis=1))
norms2 = np.sort(np.linalg.norm(op @ cell, axis=1))
if not (norms2 <= norms1 + 1E-12).all():
raise RuntimeError("Minkowski reduction failed")
return op @ cell, op