linear algebra review 061904 -...
TRANSCRIPT
1 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Review Linear Algebra
Peter AvitabileMechanical Engineering DepartmentUniversity of Massachusetts Lowell
[ ] [ ][ ][ ]ADet
AintAdjoA 1=−
[ ]
=
x0..0.x0....x0....x0x...x
U
[ ] { } { }[ ] [ ] [ ] [ ]{ } [ ] { } [ ] [ ] [ ][ ] { }{ } [ ] [ ] [ ][ ]{ }n
Tnn
gnmmmm
ngT
mmnmnnngnmm
Tmmnmnnnm
nmnm
BVSUX
BUSVBAX
USVA
BXA
=
==
=
=[ ] { } { } { }[ ]
{ }{ }{ }
=
MO
L T3
T2
T1
3
2
1
321 vvv
ss
s
uuuA
[ K ]n
[ M ]n [ M ]a[ K ]a [ E ]a
[ ω ]2
Structural Dynamic Modeling Techniques & Modal Analysis Methods
2 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra
The analytical treatment of structural dynamic systems naturally results in algebraic equations that are best suited to be represented through the use of matrices
Some common matrix representations and linear algebra concepts are presented in this section
3 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra
Common analytical and experimental equations needing linear algebra techniques
[ ] [ ] [ ]ffyf GHG =
[ ]{ } [ ]{ } [ ]{ } { })t(FxKxCxM =++ &&& [ ] [ ][ ]{ } 0xMK =λ−
[ ] [ ][ ] 1ffyf GGH −=
( )[ ] ( ){ } ( ){ }sFsxsB = ( )[ ] ( )[ ] ( )[ ]( )[ ]sBdetsBAdjsHsB 1 ==−
( )[ ] [ ] [ ]TLSUsH
=
O
O
or
4 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Matrix Notation
A matrix [A] can be described using row,column as
[ ]
=
54535251
44434241
34333231
24232221
14131211
aaaaaaaaaaaaaaaaaaaa
A
( row , column )
[A]T -Transpose - interchange rows & columns[A]H - Hermitian - conjugate transpose
5 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Matrix Notation
Square
[ ]
=
5554535251
4544434241
3534333231
2524232221
1514131211
aaaaaaaaaaaaaaaaaaaaaaaaa
A
[ ]
=
5545352515
4544342414
3534332313
2524232212
1514131211
aaaaaaaaaaaaaaaaaaaaaaaaa
A
[ ]
=
55
44
33
22
11
aa
aa
a
A [ ]
=
55
4544
353433
25242322
1514131211
a0000aa000aaa00aaaa0aaaaa
A
Triangular Diagonal
Symmetric Vandermonde Toeplitz
[ ]
=
54321
65432
76543
87654
98765
aaaaaaaaaaaaaaaaaaaaaaaaa
A [ ]
=
244
233
222
211
aa1aa1aa1aa1
A
A matrix [A] can have some special forms
6 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
General Matrix
=
nm1n1n
ij
m12221
m11211
aaa
a
aaaaaa
]A[
L
MMM
L
L
L Column Vector
Row Vector
=
n
i
2
1
b
b
bb
}B{
M
M
mj21 ccccC LL=
7 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
AdditionScalar MultiplierMatrix Multiplication
ijijij bac]B[]A[]C[ +=⇒+=
ijij a*sb]A[*s]B[ =⇒=
{ }jiij bac]B][A[]C[ =⇒=
∑=⇒
=k
kjikij
kj
j2
j1
ik2i1iij bac
b
bb
aaac
ML
8 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Matrix Manipulation
A matrix [C] can be computed from [A] & [B] as
=
3231
2221
1211
5251
4241
3231
2221
1211
3534333231
2524232221
1514131211
cccccc
bbbbbbbbbb
aaaaaaaaaaaaaaa
5125412431232122112121 bababababac ++++=
∑=k
kjikij bac
9 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Multiplication Rules]A][B[]C[]B][A[ ≠=
]C][A[]B][A[])C[]B]([A[ +=+
])C][B]([A[]C])[B][A([ =
10 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Pre-Multiplication by a Diagonal Matrix
=
nm2n1nnn
im2i1iii
m2222122
m1121111
aaad
aaad
aaadaaad
]A[D
L
L
L
L
O
O
11 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Post-Multiplication by a Diagonal Matrix
=
ni
i2
i1
ii
2n
22
12
22
1n
21
11
11
a
aa
d
a
aa
d
a
aa
dD]A[MMM
O
O
12 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Transpose of a Matrix
==⇒
=ji
2212
2111
T
ij
2221
1211
a
aaaa
]A[]B[a
aaaa
]A[
13 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Transposition RulesTTT ]B[]A[])B[]A([ +=+
[ ][ ] ]A[ATT = TTT ]A[]B[])B][A([ =
( ) TTTT ]A[]B[]C[]C][B][A[ =
Symmetric Rules
( )TTT ]B][A[]B][A[;]B[]B[;]A[]A[ ≠==
[ ] [ ]TTT C]C[;]B][A[]B[]C[;A]A[ ===
14 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Inverse of a Matrix
Properties of an Inverse
[ ]ij)ji(
ijT1 M)1(cwhere]C[]A[Adj;
]Adet[]A[Adj]A[ +− −===
[ ][ ] ]A[A11 =
−−
111 ]A[]B[])B][A([ −−− =
15 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Simple Set of Equations
A common form of a set of equations is
Underdetermined # rows < # columnsmore unknowns than equations(optimization solution)
Determined # rows = # columnsequal number of rows and columns
Overdetermined # rows > # columnsmore equations than unknowns(least squares or generalized inverse solution)
[ ]{ } [ ]bxA =
16 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Simple Set of Equations
3zy2z1y2x
1yx2
=+−=−+−
=−
=
−−−
−
321
zyx
110121
012
This set of equations has a unique solution
whereas this set of equations does not
2y2x42z1y2x
1yx2
=−=−+−
=−
=
−−−
−
221
zyx
024121
012
17 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
A matrix [A] can be decomposed and written as
Where [L] and [U] are the lower and upper diagonal matrices that make up the matrix [A]
[ ] [ ] [ ]ULA =
[ ] [ ]
=
=
x0000xx000xxx00xxxx0xxxxx
U
xxxxx0xxxx00xxx000xx0000x
L
18 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
Once the matrix [A] is written in this form then the solution for {x} can easily be obtained as
[ ]{ } [ ] [ ]BLXU 1−=
[ ] [ ] [ ]ULA =
Applications for static decomposition and inverse of a matrix are plentiful. Common methods are
Gaussian elimination Crout reductionGauss-Doolittle reduction Cholesky reduction
19 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
The individual terms of the decomposition using a process such as Crout gives
rjirijijjj
jirijijriiriiii ulau;
uula
l;ulau ∑∑∑ −=−
=−=
20 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
The simple 3x3 stiffness matrix can be decomposed to be
−
−
−−=
−−−
−
0.115.1
012
1667.00.015.0
1
11121
12
21 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
Each of the factors are given by
[ ] [ ][ ]
( )
−−−−
=
∑∑∑∑
133133221331321131
1331231221221121
131211
aaaaaaaaaaaaaaaaa
aaa
ULA
22 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
Each of the equations are processed in the order shown below where the first row is retained followed by the decomposition of the first column followed by the decomposition of the 2nd row starting from the 2-2 position and so on until the entire matrix is decomposed.
652431
aaa 131211
23 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Static Decomposition
⇓↓⇓↓⇓↓⇓↓
⇒⇒⇒⇒⇒⇓↓→→→→→→↓
⇓↓⇓↓⇓↓⇓↓⇓↓
→→→→→→↓
↓↓↓↓↓
→→→→→→↓
↓↓↓↓↓↓
nn14131211nn14131211
nn14131211nn14131211
aaaaa
then
aaaaa
aaaaa
then
aaaaa
24 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Eigenvalue Problems
Many problems require that two matrices [A] & [B] need to be reduced
Applications for solution of eigenproblems are plentiful. Common methods are
Jacobi Givens HouseholderSubspace Iteration Lanczos
[ ]{ } [ ]{ } { })t(QxBxA =+&& [ ] [ ][ ]{ } 0xAB =λ−
25 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Define an Eigenproblem
Generalized Inverse
[ ] [ ][ ]{ } { } { }i2i x;0XBA ω⇒=λ−
{ } [ ]{ } { } [ ] { }xUppUx g=⇒=
[ ] [ ] [ ]( ) [ ]T1Tg UUUU−
=
26 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Rules & Definitions
Moore-Penrose Conditions for Generalized Inverse
[ ][ ] [ ] [ ]UUUU g =
[ ] [ ][ ] [ ]ggg UUUU =
[ ] [ ]( ) [ ] [ ]UUUU gTg =
[ ][ ]( ) [ ][ ]gTg UUUU =
27 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Singular Valued Decomposition
[ ] [ ][ ][ ]TVSUA =
Any matrix can be decomposed using SVD
[U] - matrix containing left hand eigenvectors[S] - diagonal matrix of singular values[V] - matrix containing right hand eigenvectors
28 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Singular Valued Decomposition
SVD allows this equation to be written as
which implies that the matrix [A] can be written in terms of linearly independent pieces which form the matrix [A]
[ ] { } { } { }[ ]
{ }{ }{ }
=
MO
L T3
T2
T1
3
2
1
321 vvv
ss
s
uuuA
[ ] { } { } { } { } { } { } L+++= T333
T222
T111 vsuvsuvsuA
29 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Singular Valued Decomposition
Assume a vector and singular value to be
1sand321
u 11 =
=
Then the matrix [A1] can be formed to be
[ ] { } { } [ ]{ }
=
==
963642321
3211321
usuA T1111
The size of matrix [A1] is (3x3) but its rank is 1There is only one linearly independent
piece of information in the matrix
30 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Singular Valued Decomposition
Consider another vector and singular value to be
1sand1
11
u 22 =
−=
Then the matrix [A2] can be formed to be
[ ] { } { } [ ]{ }
−−−−
=−
−==
111111111
11111
11
usuA T2222
The size and rank are the same as previous caseClearly the rows and columns
are linearly related
31 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Singular Valued Decomposition
Now consider a general matrix [A3] to be
The characteristics of this matrix are not obvious at first glance.
Singular valued decomposition can be used to determine the characteristics of this matrix
[ ] [ ] [ ]213 AA1052553232
A +=
=
32 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Singular Valued Decomposition
The SVD of matrix [A3] is
or
[ ]{ }{ }{ }
−
−
=
000111321
01
1
000
111
321
A
These are the independent quantities that make up the matrix which has a rank of 2
[ ] { } { } { }TTT 0000000
11111
11
3211321
A
+−
−+
=
33 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
The basic solid mechanics formulations as well as the individual elements used to generate a finite element model are described by matrices
L
E, I
F F
θ i
i j
θ j
ν i ν j
{ } [ ]{ }
γγγεεε
=
τττσσσ
⇒ε=σ
yz
xz
xy
z
y
x
666564636261
565554535251
464544434241
363534333231
262524232221
161514131211
yz
xz
xy
z
y
x
CCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC
C
[ ]
−−−−
−−
=
L4L6L2L6L612L612
L2L6L4L6L612L612
LEIk
2
22
3
[ ]
−−
−−−−
ρ=
22
22
L4L22L3L13L22156L1354L3L13L4L22L1354L22156
420ALm
34 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Finite element model development uses individual elements that are assembled into system matrices
35 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Structural system equations - coupled
Eigensolution - eigenvalues & eigenvectors
[ ]{ } [ ]{ } [ ]{ } { })t(FxKxCxM =++ &&&
[ ] [ ][ ]{ } 0xMK =λ−
{ } { } { } [ ] { }FUp\
K\
p\
C\
p\
M\
T=
+
+
&&&
Modal space representation of equations - uncoupled
36 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Multiple Input Multiple Output Data Reduction
FREQUENCY RESPONSE FUNCTIONS FORCE
[H] [Gxx][Gyx]
RESPONSE
=
=
(MEASURED) (UNKNOWN) (MEASURED)
[ ] [ ] [ ]xxyx GHG = [ ] [ ][ ] 1xxyx GGH −=
Matrix inversion can only be performed if the matrix [Gxx] has linearly independent inputs
37 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Principal Component Analysis using SVD
[Gxx]
SVD of the input excitation matrix identifies the rank of the matrix - that is an indication of how many linearly independent inputs exist
[ ] { } { } { }[ ]
{ }{ }{ }
=
MO
L T
T2
T1
2
1
21xx 0vv
0s
s
0uuG
38 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
SVD of Multiple Reference FRF Data
SVD of the [H] matrix gives an indication of how many modes exist in the data
[ ] { } { } { }[ ]
{ }{ }{ }
=
MO
L T3
T2
T1
3
2
1
321 vvv
ss
s
uuuH
FREQUENCY RESPONSE FUNCTIONS
[H]
0 50 100 150 200 250 300 350 400 450 500Frequency (Hz)
39 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Least Squares or Generalized Inverse for Modal Parameter Estimation Techniques
Least squares error minimization of measured data to an analytical function
( )[ ] [ ]( )
[ ]( )*
k
*k
j
ik k
k
ssA
ssAsH
−+
−= ∑
=
40 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Extended analysis and evaluation of systems
[ ][ ] [ ][ ][ ]2I `UMUK ω=
[ ][ ][ ] [ ] [ ][ ][ ]2I
TI
T `UMUUKU ω=
[ ] [ ] [ ] [ ][ ][ ][ ][ ] [ ][ ] [ ][ ][ ] [ ][ ]TI
TSI
TS
S2T
SI
MUUKMUUK
VK`VKK
−−
+ω+=
[ ] [ ] [ ] [ ][ ] [ ][ ][ ][ ] [ ][ ][ ][ ]TSSS
2TSI VUKVUKVK`VKK −−+ω+=
generally require matrix manipulation of some type
41 Dr. Peter AvitabileModal Analysis & Controls Laboratory22.515 - Linear Algebra Concepts
Linear Algebra Applications
Many other applications exist
Correlation Model UpdatingAdvanced Data Manipulation
Operating Data Rotating EquipmentNonlinearities Modal Parameter Estimation
and the list goes on and on