Direct Methods for Sparse

Matrices
Miroslav Tůma

Institute of Computer Science

Academy of Sciences of the Czech Republic
and
Technical University in Liberec

1
This is a lecture prepared for the SEMINAR ON NUMERICAL
ANALYSIS: Modelling and Simulation of Challenging Engineer-
ing Problems, held in Ostrava, February 7–11, 2005. Its purpose
is to serve as a first introduction into the field of direct methods
for sparse matrices. Therefore, it covers only the most classical
results of a part of the field, typically without citations. The
lecture is a first of the three parts which will be presented in
the future. The contents of subsequent parts is indicated in the
outline.

The presentation is partially supported by the project within the

National Program of Research “Information Society” under No.
1ET400300415

2
 Outline

1. Part I.
2. Sparse matrices, their graphs, data structures
3. Direct methods
4. Fill-in in SPD matrices
5. Preprocessing for SPD matrices
6. Algorithmic improvements for SPD decompositions
7. Sparse direct methods for SPD matrices
8. Sparse direct methods for nonsymmetric matrices

3
 Outline (continued)

1. Part II. (not covered here)

2. Fill-in in LU decomposition
3. Reorderings for LU decomposition
4. LU decompositions based on partial pivoting
5. LU decompositions based on full/relaxed pivoting
6. Part III. (not covered here)
7. Parallel sparse direct methods
8. Parallel SPD sparse direct methods
9. Parallel nonsymetric sparse direct methods
10. Sparse direct methods: sequential and parallel codes

4
1. Sparse matrices, their graphs, data structures

1.a) Concept of sparse matrices: introduction

Definition 1 Matrix A ∈ IRm×n is said to be sparse if it has

O(min{m, m}) entries.

50

100

150

200

250

300

350

400
0 50 100 150 200 250 300 350 400

5
 1.a) Concept of sparse matrices: other definitions

Definition 2 Matrix A ∈ IRm×n is said to be sparse if it has row

counts bounded by rmax << n or column counts bounded by
cmax << n.

Definition 3 Matrix A ∈ IRm×n is said to be sparse if its number

of nonzero entries is O(n1+γ ) for some γ < 1.

Definition 4 (pragmatic definition: J.H. Wilkinson) Matrix

A ∈ IRm×n is said to be sparse if we can exploit the fact that a
part of its entries is equal to zero.

6
1.a) Concept of sparse matrices: an example showing
importance of the small exponent γ for n = 104

γ n1+γ
0.1 25119
0.2 63096
0.3 158489
0.4 398107
0.5 1000000

7
 1.b) Matrices and their graphs: introduction

Matrices, or their structures (i.e., positions of nonzero

entries) can be conveniently expressed by graphs

Different graph models for different purposes

• undirected graph
• directed graph
• bipartite graph

8
1.b) Matrices and their graphs: undirected graphs

Definition 5 A simple undirected graph is an ordered pair of

sets (V, E) such that E = {{i, j}|i ∈ V, j ∈ V }. V is called the
vertex (node) set and E is called the edge set.

9
 1.b) Matrices and their graphs: directed graphs

Definition 6 A simple directed graph is an ordered pair of sets

(V, E) such that E = {(i, j)|i ∈ V, j ∈ V }. V is called the vertex
(node) set and E is called the edge (arc) set.

10
 1.b) Matrices and their graphs: bipartite graphs

Definition 7 A simple bipartite graph is an ordered pair of sets

(R, C, E) such that E = {{i, j}|i ∈ R, j ∈ C}. R is called the row
vertex set, C is called the column vertex set and E is called
the edge set.

11
 1.b) Matrices and their graphs: relation matrix → graph

Definition 8
{x, y} ∈ E or (x, y) ∈ E ⇔ vertices x and y are adjacent

Adj(x) = {y|y and x are adjacent }

Structure of a nonsymmetric matrix and its graph

 
∗ ∗
 
∗ ∗ ∗
 
 ∗ ∗ 
 
 ∗ ∗ 
 

 ∗ ∗ 
∗ ∗ ∗

12
1.b) Matrices and their graphs: relation matrix → graph

Structure of a symmetric matrix and its graph

 
∗ ∗ ∗
 
∗ ∗ ∗ 
 
 ∗ ∗ ∗
 
∗ ∗ 
 

 ∗ ∗

∗ ∗ ∗

13
1.c) Data structures for sparse matrices: sparse vectors

³ ´T
v = 3.1 0 2.8 0 0 5 4.3

AV 3.1 2.8 5 4.3

JV 1 3 6 7

14
 1.c) Data structures for sparse matrices: static data
structures

• static: difficult/costly entry insertion, deletion

• CSR (Compressed Sparse by Rows) format: stores matrix
rows as sparse vectors one after another
• CSC: analogical format by columns
• connection tables: 2D array with n rows and m colums
where m denotes maximum count of a row of the stored
matrix
• A lot of other general / specialized formats

15
1.c) Data structures for sparse matrices: CSR format
example

 T
3.1 2.8
 5 4.3
 
 
 2 6 1 
1

A 3.1 2.8 5 4.3 2 6 1 1

JA 1 3 3 4 1 2 4 1

IA 1 3 5 8 9
16
1.c) Data structures for sparse matrices: dynamic data
structures

• dynamic: easy entry insertion, deletion

• linked list - based format: stores matrix rows/columns as
items connected by pointers
• linked lists can be cyclic, one-way, two-way
• rows/columns embedded into a larger array: emulated dy-
namic behavior

2 6 1
17
1.c) Data structures for sparse matrices: static versus
dynamic data structures

• dynamic data structures:

• – more flexible but this flexibility might not be needed
• – difficult to vectorize
• – difficult to keep spatial locality
• – used preferably for storing vectors
• static data structures:
• – we need to avoid ad-hoc insertions/deletions
• – much simpler to vectorize
• – efficient access to rows/columns

18
 2. Direct methods

2.a) Dense direct methods: introduction

• methods based on solving Ax = b by a matrix decomposition

– variant of Gaussian elimination; typical decompositions:
• – A = LLT , A = LDLT (Cholesky decomposition, LDLT de-
composition for SPD matrices)
• – A = LU (LU decomposition for general nonsymmetric ma-
trices)
• – A = LBLT (symmetric indefinite / diagonal pivoting de-
composition for A symmetric indefinite)

three steps of a (basic!) direct method:

1) A → LU , 2) y from Ly = b, 3) x from U x = y
19
 2.a) Dense direct methods: elimination versus
decomposition

• Householder (end of 1950’s, beginning of 1960’s): expressing

Gaussian elimination as a decomposition
• Various reformulations of the same decomposition: different
properties in
• – sparse implementations
• – vector processing
• – parallel implementations
• We will show six basic algorithms but there are others (bor-
dering, Dongarra-Eisenstat)

20
Algorithm 1 ikj lu decomposition (delayed row dense algorithm)

l = In
u = On
u11:n = a1,1:n
for i=2:n

i

for k=1:i-1
lik = aik /akk
for j=k+1:n
aij = aij − lik ∗ akj
end
end
uii:n = aii:n
21
end
Algorithm 2 ijk lu decomposition (dot product - based row
dense algorithm)
l = In , u = On , u11:n = a11:n
for i=2:n
for j=2:i
lij−1 = aij−1 /aj−1j−1
for k=1:j-1
aij = aij − lik ∗ akj

i

end
end
for j=i+1:n
for k=1:i-1
aij = aij − lik ∗ akj
end
end
ui,i:n = ai,i:n 22

end
Algorithm 3 jki lu decomposition (delayed column dense algo-
rithm)

l = In, u = On, u11 = a11

j
for j=2:n

for s=j:n

lsj−1 = asj−1/aj−1j−1

end

for k=1:j-1

for i=k+1:n

aij = aij − lik ∗ akj

end
end
u1:jj = a1:jj 23

end
Algorithm 4 jik lu decomposition (dot product - based column
dense algorithm)
l = In , u11 = a11
for j=2:n
for s=j:n
lsj−1 = asj−1/aj−1j−1
end

j

for i=2:j

for k=1:i-1

aij = aij − lik ∗ akj

end

end

for i=j+1:n

for k=1:j-1

aij = aij − lik ∗ akj

end
end 24
u1:jj = a1:jj
Algorithm 5 kij lu decomposition (row oriented submatrix dense
algorithm)

l = In
u = On
for k=1:n-1
k

for i=k+1:n

lik = aik /akk

for j=k+1:n

aij = aij − lik ∗ akj

end

end
ukk:n = akk:n
end 25

unn = ann
Algorithm 6 kji lu decomposition (column oriented submatrix
dense algorithm)

l = In , u = On
for k=1:n-1
for s=k+1:n
lsk = as,k /ak,k
k

end

for j=k+1:n

for i=k+1:n

aij = aij − lik ∗ akj

end

end
ukk:n = akk:n
end 26

unn = ann
 2.b) Sparse direct methods: existence of fill-in

• Not all the algorithms equally desirable when A is sparse

• The problem: sparsity structure of L+U (L+LT , L+B+LT )
do not need to be the same as the sparsity structure of A:
new nonzeros (fill-in) may arise

0 0

50 50

100 100

150 150

200 200

250 250

300 300

350 350

400 400
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400

27
 2.b) Sparse direct methods: fill-in

• Arrow matrix
   
∗ ∗ ∗ ∗ ∗ ∗ ∗
   
∗ ∗   ∗ ∗
   
∗ ∗   ∗ ∗
   
∗ ∗   ∗ ∗
   
∗ ∗ ∗ ∗ ∗ ∗ ∗

• How to describe the fill-in

• How to avoid it

28
 2.b) Sparse direct methods: fill-in description

Definition 9 Sequence of elimination matrices: A(0) ≡ A, A(1),

A(2), . . ., A(n): computed entries from factors replace original
(zero and nonzero) entries of A.

• Local description of fill-in using the matrix structure (entries

of elimination matrices denoted with superscripts in paren-
theses)
• Note that we use the non-cancellation assumption

Lemma 1 (fill-in lemma) Let i > j, k < n. Then

(k) (k−1) (k−1) (k−1)
aij 6= 0 ⇔ aij 6= 0 or (aik 6= 0 ∧ akj 6= 0)

29
2.b) Sparse direct methods: fill-in illustration for a
symmetric matrix

30
 2.b) Sparse direct methods: fill-in path theorem

• Simple global description via the graph model (follows from

repeated use of fill-in lemma)

(k)
Theorem 1 (fill-in path theorem) Let i > j, k < n. Then aij 6=
0 ⇔ ∃ a path xi, xp1 , . . . , xpt , xj in G(A) such that (∀l ∈ t̂)(pl < k).

31
2.b) Sparse direct methods: path theorem illustration for
a symmetric matrix

k *

l * *

j * *

i * * * *

path: i−k, k−l, l−j

32
 3. Fill-in in SPD matrices

3.a) Fill-in description: why do we restrict to the SPD

case?

• SPD case enables to separate structural properties of ma-

trices from their numerical properties.
• SPD case is simpler and more transparent
• solving sparse SPD systems is very important
• it was historically the first case with a nontrivial insight into
the mechanism of the (Cholesky) decomposition (but not the
first studied case)
• SPD case enables in many aspects smooth transfer to the
general nonsymmetric case

33
 3.b) Elimination tree: introduction

• Transparent global description: based on the concept of elim-

ination tree (for symmetric matrices) or elimination di-
rected acyclic graph (nonsymmetric matrices)

Definition 10 Elimination tree T = (V, E) of a symmetric ma-

trix is a rooted tree with V = {x1, . . . , xn}, E = {(xi, xj )|xj =
min{k|(k > i) ∧ (lik 6= 0)}}

• Note that it is defined for the structure of L

• Root of the elimination tree: vertex n
• If need we denote vertices only by their indices
• Edges in T connect vertices (i, j) such that i < j

34
 3.b) Elimination tree: illustration

   
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗
   
∗ ∗ ∗ ∗ ∗ ∗ f ∗ ∗
   
 ∗ ∗ ∗  ∗ ∗ ∗
   
 ∗ ∗ ∗ ∗  ∗ ∗ ∗ ∗
   
   
∗ ∗ ∗ ∗ ∗ f ∗ ∗ f ∗
   
 ∗ ∗ ∗ ∗  ∗ ∗ f ∗ f ∗
   
 ∗ ∗ ∗  ∗ f ∗ ∗
   
∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗

35
 3.b) Elimination tree: illustration (II.)

6
5 4
2 3

terminology: parent, child, ancestor, descendant

36
3.b) Elimination tree: necessary condition for an entry of
L to be (structurally!) nonzero

Lemma 2 If lji 6= 0 then xj is an ancestor of xi in the elimination

tree

37
 3.b) Elimination tree: natural source of parallelism

Lemma 3 Let T [xi] and T [xj ] be disjoint subtrees of the elimi-

nation tree T . Then lrs = 0 for all xr ∈ T [xi] and xs ∈ T [xj ].

x_i x_j
T[x_i]

T[x_j]

38
3.b) Elimination tree: full characterization of entries of L

Lemma 4 For j > i we have lji 6= 0 if and only if xi is an ancestor

of some xk in the elimination tree for which ajk 6= 0.

k*

j * *

i * *

i * *

j* * * * *
initiating edge
k
edges in G(L)
edge (j,i) in G(L)

39
3.b) Elimination tree: row structure of L is given by a row
subtree of the elimination tree

i
i

k k’ k’’
k k’ k’’
k’’’
k’’’

40
3.c) Computation of row (and column) counts: algorithm

initialize all colcounts to 1

for i = 1 to n do
rowcount(i) = 1
mark(xi) = i
for k such that k < i ∧ aik 6= 0 do
i
j=k
while mark(xj ) 6= i do
rowcount(i) = rowcount(i) + 1
colcount(j) = colcount(j) + 1 k k’ k’’
mark(xj ) = i
k’’’
j = parent(j)
end while
end k
41
end i
3.c) Computation of row (and column) counts: illustration

• computational complexity of evaluation row and column counts:

O(|L|)
• there exist algorithms with the complexity O(|A|, α(|A|, n))
based on decomposition of the row subtrees on independent
subtrees

k k’ k’’

k’’’

42
 3.d) Column structure of L: introduction

Lemma 5 Column j is updated in the decomposition by columns

i such that li,j 6= 0.

Lemma 6
S
Struct(L∗j ) = Struct(A∗j ) ∪ i,lij 6=0 Struct(L∗i) \ {1, . . . , j − 1}.

****
*
*
*
*
*
*
*

43
 3.d) Column structure of L: an auxiliary result

Lemma 7 Struct(L∗j ) {j} ⊆ Struct(L∗parent(j))

44
 3.d) Column structure of L: final formula

Consequently:

[
Struct(L∗j ) = Struct(A∗j ) ∪ Struct(L∗i) \ {1, . . . , j − 1}.
i,j=parent(i)

This fact directly implies an algorithm to compute

structures of columns

45
 3.d) Column structure of L: algorithm

for j = 1 to n do
listxj = ∅
end j
for j = 1 to n do
col(j) = adj(xj ) \ {x1, . . . , xj−1}
for xk ∈ listxj do
col(j) = col(j) ∪ col(k) \ {xj }
end xk
if col(j) 6= 0 then
p = min{i | xi ∈ col(j)}
listxp = listxp ∪ {xj }
end if
end j
end i

46
 3.d) Column structure of L: symbolic factorization

• array list stores children of a node

• the fact that parent of a node has a higher label than the
node induce the correctness of the algorithm
• the algorithm for finding structures of columns also called
symbolic factorization
• the derived descriptions used for:

– to allocate space for L

– to store and manage L in static data structures

• needed elimination tree

47
 3.e) Elimination tree construction: algorithm
(complexity: O(|A|, α(|A|, n)))

for i = 1 to n do
parent(i) = 0
for k such that xk ∈ adj(xi) ∧ k < i do
j=k
i
while (parent(j) 6= 0 ∧ parent(j) 6= i) do
r = parent(j)
end while
k k’ k’’
if parent(j) = 0 then parent(j) = i
end k k’’’
end i

48
 4. Preprocessing for SPD matrices

4.a) Preprocessing: the problem of reordering to minimize

fill-in

• Arrow matrix (again)

   
∗ ∗ ∗ ∗ ∗ ∗ ∗
   
∗ ∗   ∗ ∗
   
∗ ∗   ∗ ∗
   
∗ ∗   ∗ ∗
   
∗ ∗ ∗ ∗ ∗ ∗ ∗

Find efficient reorderings to minimize fill-in

49
 4.b) Solving the reordered system: overview

Factorize
P T AP = LLT ,

Compute y from
Ly = P T b,

Compute x from
LT P T x = y.

50
 4.c) Static reorderings: local and global reorderings

Static reorderings

• static differs them from dynamic reordering strategies (piv-

oting)
• two basic types

– local reorderings: based on local greedy criterion

– global reorderings: taking into account the whole graph /
matrix

51
4.d) Local reorderings: minimum degree (MD): the basic
algorithm

G = G(A)
for i = 1 to n do
find v such that degG(v) = minv∈V degG(v)
G = Gv
end i
The order of found vertices induces their new renumbering

• deg(v) = |Adj(v)|; graph G as a superscript determines the

current graph

52
4.d) Local reorderings: minimum degree (MD): an
example

G G_v

v v

53
 4.d) Local reorderings: minimum degree (MD):
indistinguishability

Definition 11 u a v are called indistinguishable if

AdjG(u) ∪ {u} = AdjG(v) ∪ {v}. (1)

Lemma 8 If u and v are indistinguishable in G and y ∈ V , y 6=

u, v. Then u and v are indistinguishable also in Gy .

Corollary 1 Let u and v be indistinguishable in G, y ≡ u has

minimum degree in G. Then v has minimum degree in Gy .

54
4.d) Local reorderings: minimum degree (MD):
indistinguishability (example)

G G_v

u v u v

55
 4.d) Local reorderings: minimum degree (MD):
dominance

Definition 12 Vertex v is called dominated by u if

AdjG(u){u} ⊆ AdjG(v) ∪ {v}. (2)

Lemma 9 If v is dominated by u in G, and y 6= u, v has minimum

degree in G. Then v is dominated by u also in Gy .

Corollary: Graph degrees do not need to be recomputed

for dominated vertices (using following relations)

v 6∈ AdjG(y) ⇒ AdjGy (v) = AdjG(v) (3)

v ∈ AdjG(y) ⇒ AdjGy (v) = (AdjG(y) ∪ AdjG(v)) − {y} (4)

56
 4.d) Local reorderings: minimum degree (MD):
implementation

• graph is represented in a clique representation {K1, . . . , Kq }

• – clique: a complete subgraph

57
 4.d) Local reorderings: minimum degree (MD):
implementation (II.)

• cliques are being created during the factorization

• they answer the main related question: how to store elimi-

nation graphs with their new edges

Lemma 10
t
X
|K| < |Ksi |
i=1
for merging t cliques Ksi into K.

58
4.d) Local reorderings: minimum degree (MD): multiple
elimination

59
 4.d) Local reorderings: minimum degree (MD): MMD
algorithm

G = G(A)
while G 6= ∅
find all vj , j = 1, . . . , s such that
degG(vj ) = minv∈V (G)degG(v) and adj(vj ) ∩ adj(vk ) pro j 6= k
for j = 1 to s do
G = Gvj
end for
end while
The order of found vertices induces their new renumbering

60
 4.d) Local reorderings: other family members

• MD, MMD with the improvements (cliques, indistinguisha-

bility, dominance, improved clique arithmetic like clique ab-
sorbtions)
• more drastic changes: approximate minimum degree algo-
rithm
• approximate minimum fill algorithms
• in general: local fill-in minimization procedures typically suffer
from lack of tie-breaking strategies – multiple elimination
can be considered as such strategy

61
4.d) Local reorderings: illustration of minimum fill-in
reordering

Degree:4, Fill−in:1 Degree:3, Fill−in:3

62
 4.e) Global reorderings: nested dissection

Find separator
Reorder the matrix numbering nodes in the separator last
Do it recursively

C_1 C_2

Vertex separator S
63
4.e) Global reorderings: nested dissection after one level
of recursion

C_1 C_2 S

C_1

C_2

64
4.d) Global reorderings: nested dissection with more levels

1 7 4 43 22 28 25
3 8 6 44 24 29 27
2 9 5 45 23 30 36
19 20 21 46 40 41 42
10 16 13 47 31 37 34
12 17 15 48 33 38 36

11 18 14 49 32 39 35

65
4.e) Global reorderings: nested dissection with more
levels: elimination tree

49
48
47
46
45
44
43
21 42
20 41
19 40
9 18 30 39
8 17 29 38
7 16 28 37

3 6 12 15 24 27 33 36
1 2 4 5 10 11 13 14 22 23 25 26 31 32 34 35

66
 4.f) Static reorderings: a preliminary summary

• the most useful strategy: combining local and global re-

orderings
• modern nested dissections are based on graph partitioners:
partition a graph such that

– components have very similar sizes

– separator is small
– can be correctly formulated and solved for a general graph
– theoretical estimates for fill-in and number of operations

• modern local reorderings: used after a few steps of an in-

complete nested dissection

67
4.f) Static reorderings: classical schemes based on
pushing nonzeros towards the diagonal

** **
**** ****
***** *****
****** ******
***** *****
***** *****
****** ******
*** ***
*** ***
6 6
Band Profile

**
Moving ****
window - * *****
*****
*****
*****
******
***
***
Frontal method - dynamic band
68
 4.f) Static reorderings: classical schemes based on
pushing nonzeros towards the diagonal: an important
reason for their use

Band(L + LT ) = Band(A)

P rof ile(L + LT ) = P rof ile(A)

69
 4.f) Static reorderings: classical schemes based on
pushing nonzeros towards the diagonal: pros and cons

• +: simple data structure – locality, regularity

• –: structural zeros inside
• +: easy to vectorize
• –: short vectors
• +: easy to use out-of-core
• –: the other schemes are typically more efficient and this is
more important

Evaluation: for general case – more or less historical value

only; can be important for special matrices, reorderings in
iterative methods
70
 4.f) Static reorderings: an example of comparison

Example (Liu): 3D finite element discretization of the part of

the automobile chassis =⇒ linear system with a matrix of
dimension 44609. Memory size for the frontal ( = dynamic
band) solver: 52.2 MB; memory size for the general sparse
solver: 5.2MB!

71
 5. Algorithmic improvements for SPD
decompositions

5.a) Algorithmic improvements: introduction

• blocks and supernodes: less tolerance to memory latencies

and increase of efective memory bandwidth
• Reorderings based on the elimination tree

72
 5.b) Supernodes and blocks: supernodes

Definition 13 Let s, t ∈ Mn such that s + t − 1 ≤ n. Then the

columns with indices {s, s+1, . . . , s+t−1} form a supernode if this
the columns satisfy Struct(L∗s) = Struct(L∗s+t−1) ∪ {s, . . . , s + t −
2}, and the sequence is maximal.

s *
**
s−t+1 ***
****
****
****
****
****
73
5.b) Supernodes and blocks: blocks

74
 5.b) Supernodes and blocks: some notes

• enormous influence on the efficiency

• different definitions of supernodes and blocks
• blocks found in G(A), supernodes are found in G(L)
• blocks are induced by the application (degrees of freedom in
grid nodes) or efficient algorithms for finding blocks
• efficient algorithms to find supernodes
• complexity: O(|A|)

75
 5.c) Reorderings based on the elimination tree:
topological reorderings

Definition 14 Topological reorderings of the elimination tree

are such that each node has smaller index than its parent.

8 8

7 7

6 6
5 4 5 2
2 3 4 1

1 3

Tree with two different topological reorderings

76
 5.c) Reorderings based on the elimination tree:
postorderings

Definition 15 Postorderings are topological reorderings where

labels in each rooted subtree form an interval.

8 8

7 7

6 6
5 4 4 5
2 3 3 1

1 2
Postordered tree

77
 5.c) Reorderings based on the elimination tree:
postorderings

• Postorderings efficiently use memory hierarchies

• Postorderings are very useful in paging environments
• They are crucial for multifrontal methods
• Even some postorderings are better than the other: trans-
parent description for multifrontal method, see the example

78
5.c) Reorderings based on the elimination tree:
postorderings: example

9 9
8 7
1 8
7 5
2 6
6 3 4
3
5 4 1 2

79
 6. Sparse direct methods for SPD matrices

6.a) Algorithmic strategies: introduction

• Some algorithms strongly modified with respect to their

simple dense counterparts because of special data structures
• different symbolic steps for different algorithms
• different amount of overhead
• of course, algorithms provide the same results in exact arith-
metic

80
 6.b) Work and memory

• the same number of operations if no additional operations

performed

• µ(L): number of the arithmetic operations

• η(L) ≡ |L|, η(L∗i): size of the i-th column of L, etc.

n
X n
X
|L| = η(L∗i) = η(Li∗)
i=1 i=1
n
X
µ(L) = [η(L∗j ) − 1][η(L∗j ) + 2]/2
j=1
81
 6.c) Methods: rough classification

• Columnwise (left-looking) algorithm

• – columns are updated by a linear combination of previous
columns
• Standard submatrix (right-looking) algorithm
• – non-delayed outer-product updates of the remaining sub-
matrix
• Multifrontal algorithm
• – specific efficient delayed outer-product updates
• Each of the algorithms represent a whole class of approaches
• Supernodes, blocks and other enhancements extremely im-
portant

82
 6.d) Methods: sparse SPD columnwise decomposition

Preprocessing
– prepares the matrix so that the fill-in would be as small as
possible
Symbolic factorization
– determines structures of columns of L. Consequently, L can
be allocated and used for the actual decomposition
– due to a lot of enhancements the boundary between the first
two steps is somewhat blurred
Numeric factorization
– the actual decomposition to obtain numerical values of the
factor L

83
6.d) Methods: sparse SPD columnwise decomposition:
numeric decomposition

* * * *
* *
*
*
* *

84
6.e) Methods: sparse SPD multifrontal decomposition

• Right-looking method with delayed updates

• The updates are pushed into a stack and popped up only
when needed
• Postorder guarantees having needed updates on the top of
the stack
• some steps of the left-looking SPD modified

85
6.e) Methods: sparse SPD multifrontal decomposition:
illustration

−1
1) = stack
−1
= stack
2)

3) −1 = stack

4) + + +

86
 7. Sparse direct methods for nonsymmetric
systems

7.a) SPD decompositions versus nonsymmetric

decompositions

• LU factorization instead of Cholesky

• nonsymmetric (row/column) permutations needed for both
sparsity preservation and maintaining numerical stability
• consequently: dynamic reorderings (pivoting) needed
• a different model for fill-in analysis: generalizing the elimina-
tion tree

87