Sparse Interpolation Annie Cuyt and Wen-shin Lee

advertisement
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Sparse Interpolation
Applications
References
Annie Cuyt
and
Wen-shin Lee
Canazei Workshop 7–12/9/2014
Universiteit Antwerpen
Department of Mathematics
and Computer Science
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
1/77
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Motivation
11001 0 0 1 1
Universiteit
Antwerpen
2/77
Sparse
Interpolation
Motivation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
interpolate
Applications
References
f (x) = α1 + α2 x 100
101
011110111110001
1100
1011
10100
▸
Newton/Lagrange interpolation: 101 samples
▸
only 4 unknowns: α1 , α2 , x 0 , x 100 !
▸
how to solve it from 4 samples?
11001 0 0 1 1
Universiteit
Antwerpen
3/77
Sparse
Interpolation
Motivation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
▸
exponential analysis
Applications
▸
generalized eigenvalue problems
References
▸
computer algebra
▸
orthogonal polynomials
▸
signal processing
▸
moment problems
▸
nonlinear approximation theory
▸
many applications . . .
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
4/77
Sparse
Interpolation
Motivation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
xs = s∆,
s = 0, 1, 2, . . .
Applications
References
n
k
∑ αi xs i = fs ,
n ≪ max(ki ),
ki ∈ N
i=1
n1
n2
∑ αi,1 cos(φi,1 xs ) + ∑ αi,2 sin(φi,2 xs ) = fs
i=1
i=1
n
∑ αi exp(φi xs ) = fs ,
φi ∈ C
i=1
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
5/77
Sparse
Interpolation
Motivation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
1. Univariate exponential sparse interpolation
(Exercise)
2. Multivariate polynomial sparse interpolation
(Exercise)
3. Connection with rational approximation theory
(Exercise)
4. Applications unlimited
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
6/77
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Basics: exponential
11001 0 0 1 1
Universiteit
Antwerpen
7/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Gaspard Riche de Prony [1795]
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
8/77
Sparse
Interpolation
Motivation
Basics: exponential
interpolation problem:
Basics:
exponential
n
Basics:
polynomial
Approximation
theory
∑ αi exp(φi xs ) = fs ,
s = 0, . . . , 2n − 1
i=1
Applications
References
2π
, ω = 2π/M
M
∣I(φi )∣ < M/2, Ωi = exp(φi ω),
xs = s
n
fs = ∑ αi Ωsi ,
s = 0, . . . , 2n − 1
i=1
101
0111101111100
1
0 010
0111
1101
0011001 0 0 1 1
Universiteit
Antwerpen
⎧
α1 + ⋅ ⋅ ⋅ + αn = f 0
⎪
⎪
⎪
⎪
⎪
⎪
⎪α1 Ω1 + ⋅ ⋅ ⋅ + αn Ωn = f1
⎨
⎪
⋮
⎪
⎪
⎪
⎪
⎪
2n−1
2n−1
⎪
= f2n−1
⎩α1 Ω1 + ⋅ ⋅ ⋅ + αn Ωn
e iω
ω=
2π
M
9/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
finding Ωi :
Basics:
polynomial
Approximation
theory
Applications
n
References
n
n−1
+ ⋅ ⋅ ⋅ + b1 z + b0
∏(z − Ωi ) = z + bn−1 z
i=1
n
0 = ∑ αi Ωsi (Ωni + bn−1 Ωn−1
+ ⋅ ⋅ ⋅ + b0 )
i
i=1
n
n−1
n
j=0
i=1
= ∑ αi Ωn+s
+ ∑ bj (∑ αi Ωj+s
i
i )
i=1
n−1
= fs+n + ∑ bj fs+j
j=0
101
0111101111100
1
0 010
0111
1101
0011001 0 0 1 1
Universiteit
Antwerpen
10/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
⎛ f0
⎜ ⋮
⎝fn−1
...
⋰
...
fn−1 ⎞ ⎛ b0 ⎞
⎛ fn ⎞
⋮ ⎟⎜ ⋮ ⎟ = −⎜ ⋮ ⎟
⎝f2n−1 ⎠
f2n−2 ⎠ ⎝bn−1 ⎠
Hankel matrix:
(r )
Hn
101
011110111110001
1100
1011
10100
⎛ fr
=⎜ ⋮
⎝fr +n−1
...
⋰
...
fr +n−1 ⎞
⋮ ⎟
fr +2n−2 ⎠
11001 0 0 1 1
Universiteit
Antwerpen
11/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Hadamard polynomial:
Approximation
theory
Applications
References
RRR f
RRR 0
R ⋮
(0)
Hn (z) = RRRRR
RRRfn−1
RRR 1
R
fn−1
⋮
f2n−2
z n−1
RRR
RRR
RRR
R
f2n−1 RRRR
R
z n RRRR
fn
⋮
(0)
n
∏(z − Ωi ) =
i=1
...
⋰
...
...
Hn (z)
(0)
∣Hn ∣
= z n + bn−1 z n−1 + ⋅ ⋅ ⋅ + b1 z + b0
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
12/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
formally orthogonal polynomial:
Basics:
polynomial
Approximation
theory
γ ∶ z s → fs ,
Applications
s = 0, 1, . . .
n
References
γ ∶ exp(φi xs ) = Ωsi → ∑ αi Ωsi = fs
γ∶
(0)
Hn (z)
zi
(0)
∣Hn ∣
i=1
→ 0,
i = 0, . . . , n − 1
(0)
Hn (z)
(0)
⊥γ z i ,
i = 0, . . . , n − 1
∣Hn ∣
[Henrici, 1974]
101
0111101111100
1
0 010
0111
1101
0011001 0 0 1 1
Universiteit
Antwerpen
13/77
Sparse
Interpolation
Basics: exponential
Motivation
(0)
Basics:
exponential
roots of
Basics:
polynomial
Hn (z)
(0)
from GEP:
∣Hn ∣
Approximation
theory
Applications
References
⎛ 1
⎜ Ω
(0)
Hn = ⎜ 1
⎜ ⋮
⎝Ωn−1
=
(1)
Hn
101
011110111110001
1100
1011
10100
...
Ω2
...
...
1
T
Vn D α Vn
⎛Ω 1
= VnT Dα ⎜
⎝
⋱
1 ⎞ ⎛ α1
Ωn ⎟ ⎜
⎟⎜
⋮ ⎟⎜
n−1 ⎠ ⎝
Ωn
⎞ ⎛1 Ω 1
⋱
⎟ ⎜ ⋮ Ω2
⎟⎜
⎟⎜
⋮
αn ⎠ ⎝1 Ωn
...
...
Ωn−1
1 ⎞
⋮ ⎟
⎟
⋮ ⎟
n−1 ⎠
Ωn
⎞
⎟ Vn
Ωn ⎠
11001 0 0 1 1
Universiteit
Antwerpen
14/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
(1)
References
det (Hn
⎛
⎛Ω 1 − λ
(0)
− λHn ) = det ⎜VnT Dα ⎜
⎝
⎝
= 0 for λ = Ωi ,
⋱
⎞ ⎞
⎟ Vn ⎟
Ωn − λ⎠ ⎠
i = 1, . . . , n
[Hua and Sarkar, 1990]
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
15/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
finding φi :
Applications
References
exp(φi ) = exp(R(φi ))e iI(φi )
M
∣I(φi )∣ <
∶
2
arg(Ωi ) = arg(exp(φi ω))
2π
= I(φi )
∈ ]−π, π[
M
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
16/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
finding αi :
Applications
References
n
s+j
∑ αi Ωi
= fs+j ,
s = 0, . . . , n − 1,
0≤j ≤n
i=1
j
⎛ Ω1
⎜ ⋮
⎝Ωj+n−1
1
...
...
Ωjn ⎞ ⎛α1 ⎞ ⎛ fj ⎞
⋮ ⎟⎜ ⋮ ⎟ = ⎜ ⋮ ⎟
Ωj+n−1 ⎠ ⎝αn ⎠ ⎝fj+n−1 ⎠
n
remaining interpolation conditions are linearly dependent
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
17/77
Sparse
Interpolation
Basics: exponential
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
finding n:
Applications
References
(r )
N < n ∶ ∣HN ∣ ≡/ 0,
(r )
N = n ∶ ∣HN ∣ ≠ 0
(r )
N > n ∶ ∣HN ∣ ≡ 0,
101
011110111110001
1100
1011
10100
r = 0, 1, . . .
if Ωi ≠ Ωj for i ≠ j
[Kaltofen and Lee, 2003]
r = 0, 1, . . .
11001 0 0 1 1
Universiteit
Antwerpen
18/77
Sparse
Interpolation
Example: exponential
Motivation
Basics:
exponential
Basics:
polynomial
4
Approximation
theory
φ(x) = ∑ αi exp(φi x)
i=1
Applications
References
α1 = 1
φ1 = 0
α2 = 2.4
φ2 = −5 + 19.97i
α3 = −2.1
φ3 = 3 + 45i
α4 = 0.2
φ4 = 5.3i
2π
evaluate at xs = s 100
,
M = 100,
∣I(φi )∣ < 50
sequence f0 , . . . , f7 , . . . is linearly generated
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
19/77
Sparse
Interpolation
Example: exponential
Motivation
102
Basics:
exponential
Basics:
polynomial
10−4
Approximation
theory
Applications
References
10−10
10−16
2
4
(0)
Figure: HN
̃j − Ωj ∣
∣Ω
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
∣Ωj ∣
≤ 2 × 10−12 ,
6
singular, N = 6
∣φ̃j − φj ∣
≤ 2 × 10−12
∣φj ∣
20/77
Sparse
Interpolation
Exercise: exponential
Motivation
Basics:
exponential
φ(x) = α1 exp(φ1 x) + α2 exp(φ2 x)
Basics:
polynomial
Approximation
theory
Applications
4 unknowns φ1 , φ2 , α1 , α2
identify φ(x) from
References
φ(0) = α1 + α2
φ′ (0) = α1 φ1 + α2 φ2
φ′′ (0) = α1 φ21 + α2 φ22
φ′′′ (0) = α1 φ31 + α2 φ32
solution:
(0)
∣H3 ∣
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
RRR φ
RR 0
= RRRR φ′0
RRR ′′
RRφ0
φ′0
φ′′0
φ′′′
0
φ′′0 RRRR
RRR
φ′′′
0 RR = 0
RRR
φ′v
0 RR
symbolically
n=2
21/77
> ###########################
## Exercise: exponential ##
###########################
> restart:
> with(LinearAlgebra):
> Phi[0] := alpha[1] * exp(phi[1]*x) + alpha[2] * exp(phi[2]*x);
f x
1
F0 := a 1 e
> for i from 1 to 4 do
Phi[i] := diff(Phi[i-1], x);
od;
f x
1
F1 := a 1 f1 e
2 f1 x
F2 := a 1 f1 e
F3 :=
4 f1 x
(1)
f x
2
C a 2 f2 e
2 f2 x
C a 2 f2 e
3 f1 x
a 1 f1 e C a 2
F4 := a 1 f1 e
> for i from 0 to 4 do
f[i] := simplify(subs(x=0, Phi[i]));
od;
f x
2
C a2 e
3 f2 x
f2 e
4 f2 x
C a 2 f2 e
(2)
f0 := a 1 C a 2
f1 := a 1 f1 C a 2 f2
2
2
f2 := a 1 f1 C a 2 f2
f3 :=
3
a 1 f1 C a 2
f2
f4 :=
4
a 1 f1 C a 2
3
f2
4
(3)
> H[3] := HankelMatrix([f[0],f[1],f[2],f[3],f[4]],3);
(4)
a1 C a2
H3 :=
2
2
3
3
a 1 f1 C a 2 f2 a 1 f1 C a 2 f2
2
2
a 1 f1 C a 2 f2 a 1 f1 C a 2 f2 a 1 f1 C a 2 f2
2
a 1 f1 C a 2
2
f2
> Determinant(H[3]);
3
a 1 f1 C a 2
3
f2
4
a 1 f1 C a 2
0
(4)
4
f2
(5)
> H[2] := HankelMatrix([f[0],f[1],f[2]],2);
H2 :=
a1 C a2
a 1 f1 C a 2 f2
2
2
a 1 f1 C a 2 f2 a 1 f1 C a 2 f2
(6)
> F2 := Matrix([[f[2]],[f[3]]]);
2
F2 :=
2
a 1 f1 C a 2 f2
3
a 1 f1 C a 2
3
f2
(7)
> B:=LinearSolve(H[2],-F2);
B :=
> B_poly := z^2 + B(2)*z + B(1);
> z_root:=solve(B_poly=0, z);
f1 f2
Kf1 K f2
(8)
B_poly := z2 C Kf1 K f2 z C f1 f2
(9)
z_root := f2 , f1
(10)
> V := Transpose(VandermondeMatrix([z_root[1],z_root[2]],2,2));
1 1
V :=
f2 f1
(11)
> F0 := Matrix([[f[0]],[f[1]]]);
F0 :=
a1 C a2
a 1 f1 C a 2 f2
(12)
> A:=LinearSolve(V,F0);
A :=
> alpha[1] := Pi;
alpha[2] := 5;
phi[1] := 2;
phi[2] := -1/7;
a2
a1
(13)
a 1 := p
a 2 := 5
f1 := 2
1
7
f2 := K
> for i from 0 to 4 do
f[i] := simplify(subs(x=0, Phi[i]));
od;
(14)
f0 := p C 5
5
7
5
f2 := 4 p C
49
5
f3 := 8 p K
343
5
f4 := 16 p C
2401
f1 := 2 p K
(15)
> H[3] := HankelMatrix([f[0],f[1],f[2],f[3],f[4]],3);
5
pC5
2pK
7
H3 :=
2pK
5
7
4pC
5
5
4pC
8pK
49
343
> Determinant(H[3]);
5
49
8pK
5
343
5
49
4pC
(16)
5
16 p C
2401
0
(17)
> H[2] := HankelMatrix([f[0],f[1],f[2]],2);
2pK
5
7
4pC
5
49
pC5
H2 :=
2pK
5
7
(18)
> F2 := Matrix([[f[2]],[f[3]]]);
4pC
5
49
8pK
5
343
F2 :=
(19)
> B:=LinearSolve(H[2],-F2);
2
7
K
B :=
13
7
(20)
K
> B_poly := z^2 + B(2)*z + B(1);
B_poly := z2 K
13
2
zK
7
7
(21)
> z_root:=solve(B_poly=0,z);
(22)
1
7
z_root := 2, K
(22)
> V := Transpose(VandermondeMatrix([z_root[1],z_root[2]],2,2));
1 1
V :=
(23)
1
7
2 K
> F0 := Matrix([[f[0]],[f[1]]]);
pC5
F0 :=
2pK
5
7
(24)
> A:=LinearSolve(V,F0);
A :=
p
5
(25)
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Basics: polynomial
11001 0 0 1 1
Universiteit
Antwerpen
22/77
Sparse
Interpolation
Basics: polynomial
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Michael Ben-Or and Prasoon Tiwari [1988]
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
23/77
Sparse
Interpolation
Motivation
Basics: polynomial
interpolation problem:
Basics:
exponential
Basics:
polynomial
∑
Approximation
theory
(k1 ,...,kd )∈K
αk1 ,...,kd x1k1 ⋯xdkd ,
#K = n
Applications
References
evaluate at
(x1 , . . . , xd ) = (ω1s , . . . , ωds )
ωj = exp(2πi/pj ),
p(ω1s , . . . , ωds ) = fs ,
(i)
Ωi =
k
ω1 1
pj > ∂j p(x1 , . . . , xd ),
0 ≤ s ≤ 2n − 1
(i)
k
⋯ωdd
,
pj mutually prime
e iωj
i = 1, . . . , n
ωj =
101
011110111110001
1100
1011
10100
2π
pj
11001 0 0 1 1
Universiteit
Antwerpen
[Giesbrecht, Labahn, and Lee, 2006]
24/77
Sparse
Interpolation
Basics: polynomial
Motivation
Basics:
exponential
finding Ωi :
Basics:
polynomial
Approximation
theory
Applications
n
References
n
n−1
+ ⋅ ⋅ ⋅ + b1 z + b0
∏(z − Ωi ) = z + bn−1 z
i=1
n
0 = ∑ αi Ωsi (Ωni + bn−1 Ωn−1
+ ⋅ ⋅ ⋅ + b0 )
i
i=1
n
n−1
n
j=0
i=1
= ∑ αi Ωn+s
+ ∑ bj (∑ αi Ωj+s
i
i )
i=1
n−1
= fs+n + ∑ bj fs+j
j=0
101
0111101111100
1
0 010
0111
1101
0011001 0 0 1 1
Universiteit
Antwerpen
25/77
Sparse
Interpolation
Basics: polynomial
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
⎛ f0
⎜ ⋮
⎝fn−1
...
⋰
...
fn−1 ⎞ ⎛ b0 ⎞
⎛ fn ⎞
⋮ ⎟⎜ ⋮ ⎟ = −⎜ ⋮ ⎟
⎝f2n−1 ⎠
f2n−2 ⎠ ⎝bn−1 ⎠
Ωi : zeros of formally orthogonal Hadamard polynomial
n
∏(z − Ωi ) =
i=1
101
011110111110001
1100
1011
10100
(0)
Hn (z)
(0)
∣Hn ∣
11001 0 0 1 1
Universiteit
Antwerpen
26/77
Sparse
Interpolation
Basics: polynomial
Motivation
(i)
Basics:
exponential
(i)
finding (k1 , . . . , kd ):
Basics:
polynomial
Approximation
theory
Applications
reverse Chinese remainder theorem
References
m = p1 ⋯pd
Ωi = ω k(i) ,
(i)
k(i ) = k1
(i) m
kj
pj
(i)
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
kj
ω = exp
⎛ 2πi ⎞
,
⎝ ∏dj=1 pj ⎠
m
(i) m
+ ⋅ ⋅ ⋅ + kd
p1
pd
mod pj = k(i ) mod pj ,
< pj ,
gcd(pj , m/pj ) = 1,
j = 1, . . . , d
27/77
Sparse
Interpolation
Basics: polynomial
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
finding αi = αk (i) ,...,k (i) :
1
d
Applications
References
n
s+j
∑ αi Ωi
= fs+j ,
s = 0, . . . , n − 1,
0≤j ≤n
i=1
j
⎛ Ω1
⎜ ⋮
⎝Ωj+n−1
1
...
...
Ωjn ⎞ ⎛α1 ⎞ ⎛ fj ⎞
⋮ ⎟⎜ ⋮ ⎟ = ⎜ ⋮ ⎟
j+n−1 ⎠ ⎝α ⎠ ⎝f
⎠
Ω
n
j+n−1
n
remaining interpolation conditions are linearly dependent
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
28/77
Sparse
Interpolation
Basics: polynomial
Motivation
Basics:
exponential
Basics:
polynomial
finding n:
Approximation
theory
Applications
References
floating-point arithmetic:
(r )
∣HN ∣ ≡ 0,
N > n,
r = 0, 1, . . .
exact arithmetic:
increase n till
δs ∶= fs + bn−1 fs−1 + ⋅ ⋅ ⋅ + b0 fs−n
equals 0, s ≥ 2n
[Massey, 1969]
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
29/77
Sparse
Interpolation
Example: polynomial
Motivation
Basics:
exponential
p(x, y ) = x 5 y + 2.2x 4 y 4 − 0.5xy 11 + 0.1xy 12
Basics:
polynomial
Approximation
theory
Applications
p1 = 6,
p2 = 13,
ω1 = exp(2πi/6),
ω2 = exp(2πi/13),
References
p(ω1s , ω2s ),
s = 0, . . . , 7
(floating-point)
s = 0, . . . , 7
(exact arithmetic)
or
p(p1s , p2s ),
sequence f0 , f1 , . . . , f7 , . . . is linearly generated and δ8 = 0
Hadamard polynomial:
z 4 + (−3.67 + 0.0799i)z 3 + (5.35 − 0.216i)z 2
(−3.67 + 0.216i)z + (0.997 − 0.0805i)
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
30/77
Sparse
Interpolation
Basics: polynomial
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
m = 78,
Ω1 = ω
Applications
ω = exp(2πi/78),
k(1)
= ω 71 = ω 5×13+1×6
Ω2 = ω k(2) = ω 76 = ω 4×13+4×6
References
Ω3 = ω k(3) = ω 79 = ω 1×13+11×6
Ω4 = ω k(4) = ω 85 = ω 1×13+12×6
K = {(5, 1), (4, 4), (1, 11), (1, 12)} ⇒ terms x 5 y , x 4 y 4 , xy 11 , xy 12
Vandermonde system
4
s+j
∑ αi Ωi
= fs+j ,
s = 0, . . . , n − 1,
0≤j ≤n
i=1
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
31/77
Sparse
Interpolation
Exercise: polynomial
Motivation
Basics:
exponential
Basics:
polynomial
p(x, y ) = (x − 3)5 (y + 5) + 2.2(x − 3)4 (y + 5)4
Approximation
theory
− 0.5(x − 3)(y + 5)11 + 0.1(x − 3)(y + 5)12
Applications
= u 5 v + 2.2u 4 v 4 − 0.5uv 11 + 0.1uv 12 ,
References
u = x − 3, v = y + 5
p1 = 6,
p2 = 13,
ω1 = exp(2πi/6),
ω2 = exp(2πi/13),
u=
ω1s ,
p(ω1s + 3, ω2s − 5),
101
011110111110001
1100
1011
10100
v = ω2s
s = 0, . . . , 7
11001 0 0 1 1
Universiteit
Antwerpen
32/77
> ##########################
## Exercise: polynomial ##
##########################
> restart:
> with(LinearAlgebra):
> read "./ExpLog.mm":
read "./RevChRem.mm":
> Digits:= 16:
> poly_uv := u^5*v + 2.2*u^4*v^4-0.5*u*v^11+0.1*u*v^12;
5
4 4
11
12
poly_uv := u v C 2.2 u v K 0.5 u v C 0.1 u v
> poly_xy := subs(u=x-3,v=y+5, poly_uv);
poly_xy := x K 3 5 y C 5 C 2.2 x K 3
4
yC5
4
K 0.5 x K 3
yC5
11
(1)
C 0.1 x K 3
yC5
12
(2)
> expand(poly_xy);
K1.464750 105 x K 1.45595805 107 y C 1.101600 105 C 72900.0 x2 K 3.21998325 107 y2 C x5 y C 1085.0 x4 y K 13110.0 x3 y C 59130.0 x2 y
6
4 4
4 3
3 4
4 2
3 3
2 4
3 2
(3)
2 3
C 4.7644175 10 x y C 2.2 x y C 44.0 x y K 26.4 x y C 330.0 x y K 528.0 x y C 118.8 x y K 3960.0 x y C 2376.0 x y
C 6.4450749 106 x y4 C 17820.0 x2 y2 C 1.07374355 107 x y3 C 1.07065475 107 x y2 C 5.5 x y11 C 137.5 x y10 C 2062.5 x y9
C 20625.0 x y8 C 1.443750 105 x y7 C 7.218750 105 x y6 C 2.5781250 106 x y5 C 0.1 x y12 C 5 x5 C 1300.0 x4 K 16050.0 x3
K 1.93357593 107 y4 K 3.22229985 107 y3 K 16.5 y11 K 412.5 y10 K 6187.5 y9 K 61875.0 y8 K 4.331250 105 y7 K 2.1656250 106 y6
K 7.7343750 106 y5 K 0.3 y12
> p1 := 6;
p2 := 13;
> w1 := evalf(exp(2*Pi*I/p1));
w2 := evalf(exp(2*Pi*I/p2));
p1 := 6
p2 := 13
(4)
w1 := 0.5000000000000000 C 0.8660254037844385 I
w2 := 0.8854560256532100 C 0.4647231720437684 I
(5)
> for i from 0 to 7 do
f[i] := evalf(subs(x=w1^i+3, y=w2^i-5, poly_xy)):
od;
f0 := 2.8 C 0. I
f1 := 2.602841017834476 K 0.8741573457604577 I
f2 := 2.064816858508122 K 1.589980617941464 I
f3 := 1.329941999443174 K 2.035486797328322 I
f4 := 0.5888159453592064 K 2.177111443869176 I
f5 := 0.02061376974983473 K 2.067629883507751 I
f6 := K0.2610215691867326 K 1.827541993129119 I
f7 := K0.2414681965457515 K 1.605741666597298 I
(6)
> H0[4] := HankelMatrix([f[0],f[1],f[2],f[3],f[4],f[5],f[6]],4);
H04 := 2.8 C 0. I, 2.602841017834476 K 0.8741573457604577 I, 2.064816858508122 K 1.589980617941464 I, 1.329941999443174
(7)
K 2.035486797328322 I ,
2.602841017834476 K 0.8741573457604577 I, 2.064816858508122 K 1.589980617941464 I, 1.329941999443174
K 2.035486797328322 I, 0.5888159453592064 K 2.177111443869176 I ,
2.064816858508122 K 1.589980617941464 I, 1.329941999443174 K 2.035486797328322 I, 0.5888159453592064
K 2.177111443869176 I, 0.02061376974983473 K 2.067629883507751 I ,
1.329941999443174 K 2.035486797328322 I, 0.5888159453592064 K 2.177111443869176 I, 0.02061376974983473
K 2.067629883507751 I, K0.2610215691867326 K 1.827541993129119 I
> H1[4] := HankelMatrix([f[1],f[2],f[3],f[4],f[5],f[6],f[7]],4);
H14 := 2.602841017834476 K 0.8741573457604577 I, 2.064816858508122 K 1.589980617941464 I, 1.329941999443174
(8)
K 2.035486797328322 I, 0.5888159453592064 K 2.177111443869176 I ,
2.064816858508122 K 1.589980617941464 I, 1.329941999443174 K 2.035486797328322 I, 0.5888159453592064
K 2.177111443869176 I, 0.02061376974983473 K 2.067629883507751 I ,
1.329941999443174 K 2.035486797328322 I, 0.5888159453592064 K 2.177111443869176 I, 0.02061376974983473
K 2.067629883507751 I, K0.2610215691867326 K 1.827541993129119 I ,
0.5888159453592064 K 2.177111443869176 I, 0.02061376974983473 K 2.067629883507751 I, K0.2610215691867326
K 1.827541993129119 I, K0.2414681965457515 K 1.605741666597298 I
> eigenvalue_list := Eigenvalues(H1[4],H0[4], output=list);
eigenvalue_list := 0.8451900855580868 C 0.5344658261253400 I, 0.8451900855413997 K 0.5344658261265820 I, 0.9870502626364418
K 0.1604112808397337 I, 0.9967573080994497 C 0.08046656865700708 I
> total_exp_list := []:
for i from 1 to 4 do
total_exp_list := [op(total_exp_list), round(ExpLog(eigenvalue_list[i],6*13))]:
od:
> total_exp_list;
7, 71, 76, 1
(9)
(10)
> exp_list:=[]:
for i from 1 to 4 do
exp_list := [op(exp_list), [RevChRem(total_exp_list[i], 6, 6*13), RevChRem(total_exp_list[i], 13,
6*13)]]:
od:
> exp_list;
1, 12 , 5, 1 , 4, 4 , 1, 11
(11)
> V := Transpose(VandermondeMatrix(eigenvalue_list));
V := 1. C 0. I, 1. C 0. I, 1. C 0. I, 1. C 0. I ,
0.8451900855580868 C 0.5344658261253400 I, 0.8451900855413997 K 0.5344658261265820 I, 0.9870502626364418
K 0.1604112808397337 I, 0.9967573080994497 C 0.08046656865700708 I ,
0.4286925614298439 C 0.9034504346214993 I, 0.4286925614003088 K 0.9034504346057614 I, 0.9485364419500248
K 0.3166679937654143 I, 0.9870502625782285 C 0.1604112807331159 I ,
K0.1205366802302719 C 0.9927088741336253 I, K0.1205366802550991 K 0.9927088740899947 I, 0.8854560256661491
K 0.4647231719910745 I, 0.9709418173518603 C 0.2393156640939936 I
> F := Matrix([[f[0]],[f[1]],[f[2]],[f[3]]]);
2.8 C 0. I
F :=
2.602841017834476 K 0.8741573457604577 I
2.064816858508122 K 1.589980617941464 I
(12)
(13)
1.329941999443174 K 2.035486797328322 I
> A := LinearSolve(V, F);
(14)
-11
0.1000000000041224 C 1.216000331768808 10
-11
A :=
1.000000000007150 C 2.565636441003167 10
I
I
(14)
2.200000000261579 K 7.965585370092284 10-11 I
-11
K0.5000000002728518 C 4.183975685423307 10
I
> var_list := [u,v]:
p := 0:
for i from 1 to 4 do
term := A[i,1]:
for j from 1 to 2 do
term := term*var_list[j]^exp_list[i,j]:
od:
p := p + term:
od:
> p;
0.1000000000041224 C 1.216000331768808 10-11 I u v12 C 1.000000000007150 C 2.565636441003167 10-11 I u5 v
(15)
C 2.200000000261579 K 7.965585370092284 10-11 I u4 v4 C K0.5000000002728518 C 4.183975685423307 10-11 I u v11
> poly_uv;
u5 v C 2.2 u4 v4 K 0.5 u v11 C 0.1 u v12
(16)
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Approximation theory
11001 0 0 1 1
Universiteit
Antwerpen
33/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Henri Padé [1892] and Christian Pommerenke [1973]
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
34/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
n
Basics:
polynomial
Approximation
theory
fs = ∑ αi exp(φi xs ),
s = 0, 1, . . . , 2n − 1
i=1
Applications
∞
References
f (z) = ∑ fj z j ,
fj = 0,
j <0
j=0
m
p(z) = ∑ ai z i ,
i=0
n
q(z) = ∑ bi z i
i=0
⎛ ∞ j⎞
i
∑ fj z q(z) − p(z) = ∑ ci z
⎝j=0
⎠
i≥m+n+1
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
35/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
⎧
f0 b0 = a0
⎪
⎪
⎪
⎪
⎪
⎪
⎪f1 b0 + f0 b1 = a1
⎨
⎪
⋮
⎪
⎪
⎪
⎪
⎪
⎪
f
b
⎩ m 0 + ⋅ ⋅ ⋅ + fm−n bn = am
b0 = 1
⎧
⎪
fm+1 b0 + ⋅ ⋅ ⋅ + fm−n+1 bn = 0
⎪
⎪
⎪
⎪
⎨ ⋮
⎪
⎪
⎪
⎪
⎪
⎩fm+n b0 + ⋅ ⋅ ⋅ + fm bn = 0
H(n)
(m+1−n)
⎛b1 ⎞
⎛fm+1 ⎞
⎜ ⋮ ⎟ = −⎜ ⋮ ⎟
⎝bn ⎠
⎝fm+n ⎠
[m/n](z) ∶= p(z)/q(z)
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
36/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
n
Approximation
theory
Applications
fs = ∑ αi exp(φi xs )
i=1
n
= ∑ αi Ωsi
References
i=1
∞
f (z) = ∑ fj z j
j=0
n
αi
1
−
zΩi
i=1
=∑
n
= Laplace transform of ∑ αi exp(φi x)
i=1
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
37/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
[n − 1/n](z) = p(z)/q(z)
Applications
References
n
q(z) = ∏(1 − zΩi )
i=1
= zn
(0)
Hn (1/z)
(0)
∣Hn ∣
= b0 z n + b1 z n−1 + ⋅ ⋅ ⋅ + bn−1 z + 1
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
38/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
(m+1−r )
Approximation
theory
→ Hr
Applications
regular, r ≥ n
∞
References
j
∑ fj x from rational function
j=0
(r )
Hν singular, ν > n, r > m + 1 − ν
↓
(r )
Hn regular, r ≥ m + 1 − n
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
39/77
Sparse
Interpolation
Approximation theory
Motivation
f (z) + ε(z) meromorphic with poles in 0 ≤ ∣z∣ < R of total
multiplicity n [de Montessus de Ballore, 1905]
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
⇓
[m/n](z) → f (z) + ε(z) uniformly on compact sets excluding
poles, with poles of f (z) + ε(z) attracting poles of [m/n](z)
according to their multiplicity
11001 0 0 1 1
Universiteit
Antwerpen
40/77
Sparse
Interpolation
Motivation
Approximation theory
f (z) + ε(z) analytic except for a countable number of poles
[Nuttall, 1970] and essential singularities [Pommerenke, 1973]
Basics:
exponential
Basics:
polynomial
⇓
Approximation
theory
[m − 1/m](z) → f (z) + ε(z) in measure on compact sets, i.e.
Applications
References
101
011110111110001
1100
1011
10100
Λ2 ({z ∶ ∣f (z) + ε(z) − [m − 1/m](z)∣ ≥ τ }) → 0
11001 0 0 1 1
Universiteit
Antwerpen
41/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
mathematical (noise free):
Applications
(0)
1. build Hν ,
References
(0)
2. Hν
ν = 0, 1, 2, . . .
= UΣV T singular value decomposition
⎛σ1
3. Σ = ⎜
⎝
⋱
⎞
⎟,
σν ⎠
σ1 ≥ σ2 ≥ ⋅ ⋅ ⋅ ≥ σn > σn+1 = ⋅ ⋅ ⋅ = σν = 0
4. find Ωi , φi , αi , i = 1, . . . , n
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
42/77
Sparse
Interpolation
Approximation theory
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
numerical (with noise):
1. take ν large enough so that noise is clearly separated from n
References
(1)
(0)
2. solve Hν vi = λi Hν vi ,
i = 1, . . . , ν,
λi = Ωi ,
i = 1, . . . , n
3. find φi
n
4. solve ∑ αi exp(φi xj ) = fj ,
0 ≤ j ≤ 2ν − 1
i=1
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
43/77
Sparse
Interpolation
Example: noise
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
φ1 = 0,
α1 = 1,
φ2 = −0.2 + 39.5i,
α2 = 2,
φ3 = −0.5 + 40i,
α3 = 4,
φ4 = −1,
α4 = 8,
References
∥ε(z)∥∞ = 10−2 ,
101
011110111110001
1100
1011
10100
2π
,
100
M = 100
xs = s
uniform random noise
11001 0 0 1 1
Universiteit
Antwerpen
44/77
Sparse
Interpolation
Example: noise
Motivation
Basics:
exponential
102
Basics:
polynomial
Approximation
theory
101
Applications
References
100
10−1
10−2
10−3
1
2
3
4
(0)
Figure: Singular values Hν
101
011110111110001
1100
1011
10100
5
6
with n = 4, ν = 6
11001 0 0 1 1
Universiteit
Antwerpen
45/77
Sparse
Interpolation
Example: noise
Motivation
Basics:
exponential
103
Basics:
polynomial
Approximation
theory
102
Applications
101
References
100
10−1
10−2
10−3
10
20
30
(0)
Figure: Singular values Hν
101
011110111110001
1100
1011
10100
40
50
with n = 4, ν = 50
11001 0 0 1 1
Universiteit
Antwerpen
46/77
Sparse
Interpolation
Exercise: approximation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
φ1 = 1.5i,
α1 = 10−3 ,
φ2 = 12.7i,
α2 = 2,
φ3 = −0.1 + 40i,
α3 = 4,
φ4 = −0.3 + 25.2i,
α4 = 8,
∥ε(z)∥∞ = 2 × 10−3 ,
101
011110111110001
1100
1011
10100
2π
,
100
M = 100
xs = s
uniform random noise
11001 0 0 1 1
Universiteit
Antwerpen
47/77
format long;
phi = [1.5*1i, 12.7*1i, -0.1+40*1i, -0.3+25.2*1i];
alpha = [10^(-3), 2, 4, 8];
eps = 2*10^(-3);
M = 100;
plot_signal
pause
plot_fft
pause
% synthesized input data with added noise
N = input('Enter the dimension for SVD: ');
% (10, 6, 3), (100, 100, 4)
randn('seed',0);
omega = 2*pi/M*(0:2*N-1);
f = syn_exp(alpha, phi, omega);
v = randn(size(f))+randn(size(f))*1i;
vv = v/norm(v,Inf);
f = f + eps*vv;
% form Hankel matrices H0 and H1 from y sequence
[H0,H1] = mat_ge(f);
plot_svd
pause
% reconstruct the parameters via generalized eigenvalues
n = input('Size of the model: ');
% compute the generalized eigenvalues and form the Vandermonde system
E = eig(H1(1:n,1:n),H0(1:n,1:n));
V = rot90(vander(E));
% amplitudes
A = V\f(1:n).';
% frequencies and damping factors
alpha_rec = A;
phi_rec = log(E)*M/(2*pi);
pause
% extract the non-zero terms
extract
% plot computed parameters
plot_recontructed_parameters
15
3.5
3
10
2.5
5
2
0
1.5
−5
1
−10
−15
0.5
0
pi
pi/2
3pi/2
2pi
102
0
−50
−25
0
25
50
−25
0
25
50
−25
0
25
50
8
6
101
4
10
2
0
0
−50
10−1
0
−0.2
10−2
−0.4
10−3
2
4
6
8
10
−50
15
3.5
3
10
2.5
5
2
0
1.5
−5
1
−10
−15
0.5
0
pi
pi/2
3pi/2
2pi
103
0
−50
−25
0
25
50
−25
0
25
50
−25
0
25
50
8
102
6
4
101
2
100
0
−50
10−1
0
10−2
−0.2
10−3
−0.4
10−4
20
40
60
80
100
−50
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Applications
11001 0 0 1 1
Universiteit
Antwerpen
48/77
Sparse
Interpolation
Applications
Motivation
Basics:
exponential
Exponential analysis in physical phenomena:
Basics:
polynomial
▸
power system transient detection
Approximation
theory
▸
motor fault diagnosis
Applications
▸
drug clearance / glucose tolerance
References
▸
magnetic resonance / infrared spectroscopy
▸
vibration analysis
▸
seismic data analysis
▸
music signal processing
▸
corrosion rate / crack initiation
▸
odour recognition with electronic nose
▸
typed keystroke recognition
▸
liquid (explosive) identification
▸
...
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
49/77
Sparse
Interpolation
Transients
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Transient detection and characterization
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
50/77
Sparse
Interpolation
Transients
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
short lived high frequency signal:
References
101
011110111110001
1100
1011
10100
▸
speech processing
▸
turbulent flow
▸
power lines
▸
...
11001 0 0 1 1
Universiteit
Antwerpen
51/77
Sparse
Interpolation
Transients
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
▸
model with φi = 120πi,
n
Applications
∑ αi cos(120πx + γi )1[Ai ,Zi [
References
i=1
101
011110111110001
1100
1011
10100
▸
n = 3, αi = 1, γ1,3 = −π/2, γ2 = 3π/4
▸
[A1 , Z1 [ = [0, 0.0308[
[A2 , Z2 [ = [0.0308, 0.0625[
[A3 , Z3 [ = [0.0625, 0.1058[
▸
M = 1200
▸
uniformly distributed noise in [−0.05, 0.05]
11001 0 0 1 1
Universiteit
Antwerpen
52/77
Sparse
Interpolation
Transients
Motivation
Basics:
exponential
1
Basics:
polynomial
Approximation
theory
Applications
References
0
−1
0
37
75
127
Figure: Given transient signal
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
53/77
Sparse
Interpolation
Transients
Motivation
Basics:
exponential
Basics:
polynomial
▸
at each instance: 2 exponential terms
Approximation
theory
▸
characteristics of terms change
Applications
▸
inspect rank of
References
(1)
H4
⎛f 1
⎜f
=⎜2
⎜f 3
⎝f 4
f2
f3
f4
f5
f3
f4
f5
f6
f4 ⎞
f5 ⎟
⎟
f6 ⎟
f7 ⎠
[A1 , Z1 [ = [0/M, 37/M[,
[A2 , Z2 [ = [37/M, 75/M[,
[A3 , Z3 [ = [75/M, 127/M[
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
54/77
Sparse
Interpolation
Transients
Motivation
Basics:
exponential
4
Basics:
polynomial
Approximation
theory
3
singular values
Applications
References
2
1
0
31 37
69 75
(r )
Figure: Numerical rank of H4
101
011110111110001
1100
1011
10100
127
evolving over time x
11001 0 0 1 1
Universiteit
Antwerpen
55/77
Sparse
Interpolation
Audio signals
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Reconstructing undersampled audio signals
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
56/77
Sparse
Interpolation
Audio signals
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
song containing 29 notes of 0.25 seconds each:
▸
M = 44100 (Hz)
▸
11025 samples per note, 319725 in total
▸
16.35 ≤ φi ≤ 4978.03, i = 1, . . . , 100
▸
complex exponential model
11001 0 0 1 1
Universiteit
Antwerpen
57/77
Sparse
Interpolation
Audio signals
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
0.25
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
−0.25
0
0.05
0.1
0.15
0.2
0.25
Figure: Sampled signal produced by 1 note
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
58/77
Sparse
Interpolation
Audio signals
compressive sensing (optimisation, probabilistic)
Motivation
Basics:
exponential
−5
4
−4
x 10
4
2
Basics:
polynomial
Approximation
theory
0
0
−2
−2
−4
−4
References
1
2
3
4
5
6
7
1
2
3
sec
5
6
7
4
5
6
7
4
5
6
7
4
5
6
7
−4
x 10
2
0.5
x 10
1
0
0
−0.5
−1
−1
4
sec
−5
1
Applications
x 10
2
1
2
3
4
5
6
7
−2
1
2
3
sec
sec
Figure: 4 runs with 1229 samples
0.4
0.2
0.2
0.1
0
0
−0.1
−0.2
−0.2
1
2
3
4
5
6
7
−0.4
1
2
3
sec
sec
0.5
1
0.5
0
0
−0.5
−1
−0.5
1
2
3
4
sec
101
011110111110001
1100
1011
10100
5
6
7
−1
1
2
3
sec
11001 0 0 1 1
Universiteit
Antwerpen
Figure: 4 runs with 456 samples
59/77
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Audio signals
0.25
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
−0.25
0
0.05
0.1
0.15
0.2
0.25
Figure: Full set of samples per note
0.25
0.2
0.15
0.1
0.05
0
−0.05
−0.1
−0.15
−0.2
−0.25
0
101
0111101111100
1
0 010
0111
1101
0.05
0.1
0.15
0.2
0.25
0011001 0 0 1 1
Universiteit
Antwerpen
Figure: Sparse interpolation with 7 samples per note
60/77
Sparse
Interpolation
MCSA
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Preventive diagnosis of a broken rotor bar
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
61/77
Sparse
Interpolation
MCSA
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
3-phase induction motors:
▸
consume 40 − 50% of all
electricity in industrialized
countries
▸
rotor made up metal bars
▸
stator current signal analysed
▸
broken bar(s) characterized
by sideband frequencies
▸
difficult to diagnoze under
low or no load
Applications
References
101
011110111110001
1100
1011
10100
Figure: Stator and rotor
11001 0 0 1 1
Universiteit
Antwerpen
62/77
Sparse
Interpolation
MCSA
Motivation
Basics:
exponential
Service Factor
is about once
Basics:
processes,
this
polynomial
1.15
(a)
(a)
Approximation
theory
Applications
References
(b)
Fig. 8. Medium-voltage induction motor with one broken rotor bar. (a)
The medium-voltage test motor. (b) The rotor cage with one rotor bar
removed.
(a)
monitoring interval can
be changed
to as fast
as onceFFT
every
Figure:
Stator
current
120 seconds when appropriate. This helps safeguard the
motor against potential failures on its rotor cage.
V.
EXPERIMENTAL RESULTS
To test the reliability of the fault severity evaluation
algorithm on large medium-voltage motors, a 500-hp 4.16kV
induction motor is tested under both healthy and faulty
Universiteit
conditions. Experimental setup and test results are reported
Antwerpen
in this section.
101
011110111110001
1100
1011
10100
11001 0 0 1 1
(b)
Fig. 9. Stator
current spectrum.
(a) Healthy
(b) Motor
with one
spectra:
healthy
and with
1 motor.
broken
bar
broken bar.
To emulate a broken rotor bar failure, a small part of the
rotor bar is completely removed. To avoid the subsequent
mechanical unbalance and resultant torque pulsation, the
rotor bar is then glued to a close rotor bar. Fig. 8 shows the
field photos of the medium-voltage motor with the created
one broken rotor bar failure.
B.
Experimental Results
63/77
Sparse
Interpolation
MCSA
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: 10% load, 16dB noise, ν = 400
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
64/77
Sparse
Interpolation
Bio-electrical
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Sparse EEG approximation
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
65/77
Sparse
Interpolation
Bio-electrical
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Bio-electrical signals:
▸
electrical activity of cells and tissues
▸
clinical studies of health status
▸
ECG, EEG, EMG, EOG, . . .
▸
sparse model (n = 8) is approximate
11001 0 0 1 1
Universiteit
Antwerpen
66/77
Sparse
Interpolation
Bio-electrical
Motivation
Basics:
exponential
40
original
reconstructed
Basics:
polynomial
Approximation
theory
Applications
30
20
References
10
0
−10
−20
0
1
2
3
4
time (s)
5
6
7
8
Figure: Reconstruction of 8 second [1 − 20] Hz bandpass filtered EEG
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
67/77
Sparse
Interpolation
Bio-electrical
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Sparse EOG approximation
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
68/77
Sparse
Interpolation
Bio-electrical
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Polysomnogram:
▸
12 channels
▸
22 wire attachments to patient
▸
heart rate, leg measurement, airflow (chest, abdomen), chin
muscle, EEG, EOG, . . .
11001 0 0 1 1
Universiteit
Antwerpen
69/77
Sparse
Interpolation
Bio-electrical
Motivation
Basics:
exponential
150
original
reconstructed
Basics:
polynomial
Approximation
theory
100
Applications
References
50
0
−50
−100
0
1
2
3
4
time (s)
5
6
7
8
Figure: Reconstruction of 8 second EOG (CC = 99.2%)
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
70/77
Sparse
Interpolation
Spectroscopy
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
Figure: Spectral analysis of FID
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
71/77
Sparse
Interpolation
Spectroscopy
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
Magnetic resonance spectroscopy:
▸
physical and chemical properties of molecules
▸
a.o. concentration of metabolites in the brain
▸
frequencies clustered → high frequence resolution
▸
free induction decay → time constraint
▸
Fourier methods need additional tools
11001 0 0 1 1
Universiteit
Antwerpen
72/77
Sparse
Interpolation
Spectroscopy
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
φ(x) = 5 × 10−2 + 2e (−0.97+i79.94π)x + 4e (−1+i80π)x + 8e −1.1x + ε(x)
∥ε(x)∥∞ = 10−3 ,
circular Gaussian noise
Figure: The real (left) and imaginary (right) part of φ(x)
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
73/77
Sparse
Interpolation
Spectroscopy
Motivation
Basics:
exponential
Basics:
polynomial
6
Approximation
theory
Applications
4
References
2
0
-100 pi
-50 pi
0
50 pi
100 pi
Figure: Real (black) and imaginary (red) parts of FFT
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
74/77
Sparse
Interpolation
Spectroscopy
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
8
6
4
2
0
-100 pi
-50 pi
0
50 pi
100 pi
-50 pi
0
50 pi
100 pi
0
−0.5
−1
-100 pi
Figure: Amplitudes (top) and damping factors (bottom) of φ(x)
101
0111101111100
1
0 010
0111
1101
0011001 0 0 1 1
Universiteit
Antwerpen
75/77
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
References
11001 0 0 1 1
Universiteit
Antwerpen
76/77
Sparse
Interpolation
Motivation
Basics:
exponential
Basics:
polynomial
Approximation
theory
Applications
References
101
011110111110001
1100
1011
10100
11001 0 0 1 1
Universiteit
Antwerpen
References
M. Ben-Or and P. Tiwari. A deterministic algorithm for sparse multivariate polynomial
interpolation. In STOC ’88: Proceedings of the twentieth annual ACM symposium on
Theory of computing, pages 301–309, New York, NY, USA, 1988. ACM.
http://dx.doi.org/10.1145/62212.62241.
M. de Montessus de Ballore. Sur les fractions continues algébriques. Rend. Circ. Mat.
Palermo, 19:185–257, 1905. http://dx.doi.org/10.1007/BF03014011.
R. de Prony. Essai expérimental et analytique sur les lois de la dilatabilité des fluides
élastiques et sur celles de la force expansive de la vapeur de l’eau et de la vapeur de
l’alkool, à différentes températures. J. Ec. Poly., 1:24–76, 1795.
M. Giesbrecht, G. Labahn, and W. Lee. Symbolic-numeric sparse interpolation of
multivariate polynomials. In ISSAC’06, 2006.
http://dx.doi.org/10.1145/1145768.1145792. 9-12 July.
P. Henrici. Applied and computational complex analysis I. John Wiley & Sons, New York,
1974.
Y. Hua and T. K. Sarkar. Matrix pencil method for estimating parameters of exponentially
damped/undamped sinusoids in noise. IEEE Trans. Acoust. Speech Signal Process., 38:
814–824, 1990. http://dx.doi.org/10.1109/29.56027.
E. Kaltofen and W.-s. Lee. Early termination in sparse interpolation algorithms. J. Symbolic
Comput., 36(3-4):365–400, 2003. http://dx.doi.org/10.1016/S0747-7171(03)00088-9.
International Symposium on Symbolic and Algebraic Computation (ISSAC’2002) (Lille).
J. L. Massey. Shift-register synthesis and BCH decoding. IEEE Trans. Inform. Theory, 15
(1):122–127, 1969. http://dx.doi.org/10.1109/TIT.1969.1054260.
J. Nuttall. The convergence of Padé approximants of meromorphic functions. J. Math.
Anal. Appl., 31:147–153, 1970. http://dx.doi.org/10.1016/0022-247X(70)90126-5.
H. Padé. Sur la représentation approchée d’une fonction par des fractions rationnelles. PhD
thesis, Faculté des sciences de Paris, 1892.
C. Pommerenke. Padé approximants and convergence in capacity. J. Math. Anal. Appl., 41:
77/77
775–780, 1973. http://dx.doi.org/10.1016/0022-247X(73)90248-5.
Download