4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174 | def quadtree_laplacian(C,W,CH,D,A):
"""Finite difference Laplacian matrix on a quadtree
Builds a finite difference laplacian on a quadtree following a centered finite difference scheme, with the adjacency as suggested by Bickel et al. "Adaptative Simulation of Electrical Discharges".
Parameters
----------
C : numpy double array
Matrix of cell centers
W : numpy double array
Vector of cell half widths
CH : numpy int array
Matrix of child indeces (-1 if leaf node)
D : numpy int array
Vector of tree depths
A : scipy sparse.csr_matrix
Sparse node adjacency matrix, where a value of a in the (i,j) entry means that node j is to the a-th direction of i (a=1: left; a=2: right; a=3: bottom; a=4: top).
Returns
-------
L : scipy sparse.csr_matrix
sparse Laplacian matrix
stored_at : numpy double array
Matrix of child cell centers, where the values of L are stored
See also
--------
quadtree_gradient, initialize_quadtree.
Notes
-----
This code is *purposefully* not optimized beyond asymptotics for simplicity in understanding its functionality and translating it to other programming languages beyond prototyping.
Examples
--------
```python
# Create a random point cloud
P = 2*np.random.rand(100,2) - 1
# Initialize the quadtree
C,W,CH,PAR,D,A = gpytoolbox.initialize_quadtree(P,graded=True,max_depth=8)
# Get the Laplacian matrix
G, stored_at = gpytoolbox.quadtree_laplacian(C,W,CH,D,A)
```
"""
# Builds a finite difference laplacian on a quadtree following a centered
# finite difference scheme, with the adjacency as suggested by
# Bickel et al. "Adaptative Simulation of Electrical
# Discharges". This code is *purposefully* not optimized beyond
# asymptotics for simplicity in understanding its functionality and
# translating it to other programming languages beyond prototyping.
#
# G = quadtree_laplacian(C,W,CH,D,A)
# G,stored_at = quadtree_laplacian(C,W,CH,D,A)
#
# Inputs:
# C #nodes by 3 matrix of cell centers
# W #nodes vector of cell widths (**not** half widths)
# CH #nodes by 4 matrix of child indeces (-1 if leaf node)
# D #nodes vector of tree depths
# A #nodes by #nodes sparse adjacency matrix, where a value of a in the
# (i,j) entry means that node j is to the a-th direction of i
# (a=1: left a=2: right a=3: bottom a=4: top).
#
# Outputs:
# G #num_children by #num_children sparse laplacian matrix
# stored_at #num_children by 3 matrix of child cell centers, where the
# values of L are stored
dim = C.shape[1]
num_faces_per_cell = 2*dim
# We will store Laplacian values at
# child cell indeces
children = np.nonzero(CH[:,1]==-1)[0]
# map from all cells to children
cell_to_children = -np.ones(W.shape[0],dtype=int)
cell_to_children[children] = np.linspace(0,children.shape[0]-1,children.shape[0],dtype=int)
# Vectors for constructing the Laplacian
I = []
J = []
vals = []
for i in range(children.shape[0]):
new_I = []
new_J = []
new_vals = []
l = [0,0,0,0,0]
new_dirs = []
child = children[i]
d = D[child]
num_dirs = 0
# Let's build d u(child)/dx^2 ~ u(child+W(child)*[1,0])/hr(hl+hr) -
# 2u(child)/hlhr + u(child-W(child)*[1,0])/hr(hl+hr)
# So, let's look for the value to the j direction. To do this, we seek the
# lowest-depth neighbor to the j direction. As a reminder the octree
# adjacency convention is i->j (1:left-2:right-3:bottom-4:top)
for j in range(1,num_faces_per_cell+1):
j_neighbors = (A[child,:]==j).nonzero()[1]
if len(j_neighbors)>0:
depths_j_neighbors = D[j_neighbors]
max_depth_j_neighbor = np.argmax(depths_j_neighbors)
max_depth_j = depths_j_neighbors[max_depth_j_neighbor]
max_depth_j_neighbor = j_neighbors[max_depth_j_neighbor]
# There are two options:
# One: the leaf node to our j direction has lower or equal depth to
# us
if max_depth_j<=d:
l[j] = (W[child] + W[max_depth_j_neighbor])/2.0
# then it's easy, just add this node
new_I.append(i)
# THIS HAS TO BE A CHILD !
assert(cell_to_children[max_depth_j_neighbor]>=0)
new_J.append(cell_to_children[max_depth_j_neighbor])
new_vals.append(-1.0)
new_dirs.append(j)
else:
# In this case, assuming the grid is graded, there should
# be two j-neighbors at depth d+1
nn = j_neighbors[D[j_neighbors]==(d+1)]
assert len(nn)==2, "Are you sure you are inputting a graded quadtree?"
assert all(CH[nn,1]==-1)
# Then we simply average both
l[j] = (W[child] + W[nn[1]])/2.0
new_I.append(i)
new_I.append(i)
new_J.append(cell_to_children[nn[0]])
new_J.append(cell_to_children[nn[1]])
new_vals.append(-0.5)
new_vals.append(-0.5)
new_dirs.append(j)
new_dirs.append(j)
num_dirs = num_dirs + 1
# At this point, we have to divide by the edge-lengths and add sign
for s in range(len(new_dirs)):
if new_dirs[s]==1:
if l[2]==0:
new_vals[s] = new_vals[s]/(l[1]*(l[1]+l[1]))
else:
new_vals[s] = new_vals[s]/(l[1]*(l[1]+l[2]))
elif new_dirs[s]==2:
if l[1]==0:
new_vals[s] = new_vals[s]/(l[2]*(l[2]+l[2]))
else:
new_vals[s] = new_vals[s]/(l[2]*(l[1]+l[2]))
elif new_dirs[s]==3:
if l[4]==0:
new_vals[s] = new_vals[s]/(l[3]*(l[3]+l[3]))
else:
new_vals[s] = new_vals[s]/(l[3]*(l[3]+l[4]))
elif new_dirs[s]==4:
if l[3]==0:
new_vals[s] = new_vals[s]/(l[4]*(l[4]+l[4]))
else:
new_vals[s] = new_vals[s]/(l[4]*(l[3]+l[4]))
# And add them to the big sparse Laplacian construction vectors
I.extend(new_I)
J.extend(new_J)
vals.extend(new_vals)
# THE LAPLACIAN IS NEGATIVE SEMI DEFINITE!
L = -2*csr_matrix((vals,(I,J)),(children.shape[0],children.shape[0]))
L = L - diags(np.array(L.sum(axis=1)).squeeze(),0)
stored_at = C[children,:]
return L, stored_at
|