Information Technology Reference
In-Depth Information
M
2,1
M
−
1
1,1
M
1,2
L
2
D
2
MULT
[
n/
2
]
Tr a n s p o s e
M
2,1
(
L
−
1
)
T
D
−
1
LDL
T
[
n/
2]
1
SCALE
[
n/
2
]
M
2,1
M
−
1
S
=
M
2,2
−
1,1
M
1,2
M
2,1
(
L
−
1
)
T
SUB
[
n/
2
]
MULT
[
n/
2
]
M
2,2
Tr a n s p o s e
M
2,1
L
−
1
1
TRI INV
[
n/
2
]
L
1
D
1
LDL
T
[
n/
2]
M
1,1
Figure 6.6
An algebraic circuit to produce the
LDL
T
factorization of an SPD matrix.
1,1
M
1,2
can be formed by multiplying
M
2,1
L
−
1
T
D
−
1
M
2,1
M
−
1
•
S
=
M
2,2
−
1
by the
transpose of
M
2,1
L
−
1
T
using MULT
[
n/
2
]
and subtracting the result from
M
2,2
by the
subtraction operator SUB
[
n/
2
]
.
The
LDL
T
•
factorization of the
n/
2
×
n/
2matrix
S
is computed using the procedure
LDL
T
[
n/
2] to produce the
n/
2
×
n/
2 triangular and diagonal matrices
L
2
and
D
2
,re-
spectively.
Let's now determine the size and depth of circuits to implement the algorithm for LDL
T
[
n
].
Let
f
(
n
)
n
2
(
n
2
+
n
)
/
2
be the function defined by the
LDL
T
factorization of an
n
R
→R
×
n
LDL
T
:
SPD matrix,
f
(
n
)
(
n
2
+
n
)
/
2
(
n
2
+
n
)
/
2
be the inversion of an
n
R
→R
×
n
lower triangular
tri inv
:
matrix,
f
(
n
)
n
2
+
n
n
2
be the computation of
N
(
D
−
1
)
for an
n
scale
:
R
→R
×
n
matrix
N
and
adiagonalmatrix
D
,
f
(
n
)
2
n
2
n
2
be the multiplication of two
n
mult
:
R
→R
×
n
matrices, and
f
(
n
)
2
n
2
n
2
sub
:
R
→R
the subtraction of two
n
×
n
matrices. Since a transposition can be done
Search WWH ::
Custom Search