a tale of two eigenvalue problems - cornell universitybindel/present/2013-11-brownbag.pdf · 2014....
TRANSCRIPT
A Tale of Two Eigenvalue ProblemsMusic of the Microspheres + Hearing the Shape of a Graph
David Bindel
Department of Computer ScienceCornell University
Brown Bag Talk, 26 Nov 2013
Brown bag 1 / 46
The Computational Science & Engineering Picture
Application
Analysis Computation
MEMSSmart gridsNetworks
Linear algebraApproximation theorySymmetry + structure
HPC / cloudSimulatorsLittle languages
Brown bag 2 / 46
The purpose of computing is insight, not numbers.Richard Hamming
Brown bag 3 / 46
G. H. Bryan (1864–1928)
Fellow of the Royal Society (1895)Stability in Aviation (1911)Thermodynamics, hydrodynamics
Bryan was a friendly, kindly, veryeccentric individual...(Obituary Notices of the FRS)
... if he sometimes seemed a colossal buffoon, he himself didnot help matters by proclaiming that he did his best workunder the influence of alcohol.(Williams, J.G., The University College of North Wales, 1884–1927)
Brown bag 4 / 46
Bryan’s Experiment
“On the beats in the vibrations of a revolving cylinder or bell”by G. H. Bryan, 1890
Brown bag 5 / 46
Bryan’s Experiment Today
Brown bag 6 / 46
The Beat Goes On
0.00 6.48 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 1.10 1.20 1.30 1.40 1.50 1.60 1.70 1.80 1.90 2.00 2.10 2.20 2.30 2.40 2.50 2.60 2.70 2.80 2.90 3.00 3.10 3.20 3.30 3.40 3.50 3.60 3.70 3.80 3.90 4.00 4.10 4.20 4.30 4.40 4.50 4.60 4.70 4.80 4.90 5.00 5.10 5.20 5.30 5.40 5.50 5.60 5.70 5.80 5.90 6.00 6.10 6.20 6.30 6.40
Audio Track
0.00 8.36 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00 1.10 1.20 1.30 1.40 1.50 1.60 1.70 1.80 1.90 2.00 2.10 2.20 2.30 2.40 2.50 2.60 2.70 2.80 2.90 3.00 3.10 3.20 3.30 3.40 3.50 3.60 3.70 3.80 3.90 4.00 4.10 4.20 4.30 4.40 4.50 4.60 4.70 4.80 4.90 5.00 5.10 5.20 5.30 5.40 5.50 5.60 5.70 5.80 5.90 6.00 6.10 6.20 6.30 6.40 6.50 6.60 6.70 6.80 6.90 7.00 7.10 7.20 7.30 7.40 7.50 7.60 7.70 7.80 7.90 8.00 8.10 8.20
Audio Track
Brown bag 7 / 46
The Beat Goes On
99.4
99.6
99.8
100
100.2
100.4
100.6
0 0.2 0.4 0.6 0.8 1
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
0 2 4 6 8 10 12 14
Free vibrations in a rotating frame (simplified):
¨
q+ 2�⌦J ˙
q+ !20q = 0, J ⌘
0 �1
1 0
�
Eigenvalue problem:��!2
I+ 2i!�⌦J+ !20
�q = 0.
Solutions: ! ⇡ ⌦0 ± �⌦. =) beating / ⌦!
Brown bag 8 / 46
A General Picture
=q1(t) +q2(t)
q2
q1
q2
q1
q1(t)q2(t)
�⇡
cos(��⌦t) � sin(��⌦t)sin(��⌦t) cos(��⌦t)
� q01(t)q02(t)
�.
Brown bag 9 / 46
Foucault in Solid State
Brown bag 10 / 46
A Small Application
Northrup-Grummond HRG(developed c. 1965–early 1990s)
Brown bag 11 / 46
A Smaller Application (Cornell)
Brown bag 12 / 46
A Smaller Application (UMich, GA Tech, Irvine)
Brown bag 13 / 46
Uncritical FEA: Fail!
Brown bag 14 / 46
The Perturbation Picture
Perturbations split degenerate modes:Coriolis forces (good)Imperfect fab (bad, but physical)Discretization error (non-physical)
Brown bag 15 / 46
Perfect Geometry
Free vibration problem in weak form
8w, b(w,a) + a(w,u) = 0.
wherea =
¨
u+ 2⌦⇥ ˙
u+⌦⇥ (⌦⇥ x) +
˙
⌦⇥ x
Expand u (and w) in Fourier series:
u(r, z, ✓) =
1X
m=0
�
c
m
(✓)uc
m
(r, z) +
1X
m=0
�
s
m
(✓)us
m
(r, z)
Discretize by finite elements:
M
¨
u
h
+C
˙
u
h
+K
¨
u
h
= 0
where C comes from Coriolis term (2b(w,⌦⇥ ˙
u)).Brown bag 16 / 46
Block Diagonal Structure
K =
2
666666666664
K
cc
0
K
ss
1
K
cc
1
K
ss
2
K
cc
2. . .
K
ss
M
K
cc
M
3
777777777775
Mass has same structure.
Brown bag 17 / 46
Block Structure of Finite Element Matrix
Discretize 2b(w,⌦⇥ ˙
u):
C =
2
6664
C00 C01
C10 C11 C12
C21 C22 C23. . . . . . . . .
3
7775
Off-diagonal blocks come from cross-axis sensitivity:
⌦ = ⌦
z
e
z
+⌦
r✓
=
2
40
0
⌦
z
3
5+
2
4cos(✓)⌦
x
� sin(✓)⌦y
sin(✓)⌦x
+ cos(✓)⌦y
0
3
5 .
Neglect cross-axis effects (O(⌦
2/!20), like centrifugal effect).
Brown bag 18 / 46
Analysis in Ideal Case
Only need to mesh a 2D cross-section!Compute an operating mode u
c
for the non-rotating geometry.Compute � = b(u
c
, ez
⇥ u
s
).Model: motion is approximately q1uc
+ q2us
, and
¨
q+ 2�⌦J ˙
q+ !20q = 0,
Brown bag 19 / 46
Representing Imperfection
R 7! R+ ✏ (R)
B0 B
Write Fourier series for , too!
Brown bag 20 / 46
Typical Fabrication Imperfections
p = 0 p = 1 p = 2 p = 4
Brown bag 21 / 46
Block matrix structure
Ex: p = 2
K =
2
666666666664
K0 ✏ ✏2 ✏3
K1 ✏ ✏2
✏ K2 ✏ ✏2
✏ K3 ✏✏2 ✏ K4 ✏
✏2 ✏ K5
✏3 ✏2 ✏ K6. . .
3
777777777775
Brown bag 22 / 46
Qualitative Information
Operating wave number m, perturbation number p:
p = 2m frequencies split by O(✏)
kp = 2m frequencies split at most O(✏2)
p 6 |2m frequencies change at O(✏2), no splitp = 1
p = 2m± 1 O(✏) cross-axis coupling.
Note:m = 2 affected at first order by p = 0 and p = 4
(and O(✏2) split from p = 1 and p = 2).m = 3 affected at first order by p = 0 and p = 6
(and O(✏2) split from p = 1 and p = 3).
Brown bag 23 / 46
Analyzing Imperfect Rings
Brown bag 24 / 46
Analyzing Imperfect Rings
q2
q1
q2
q1
q2
q1
q2
q1
q2
q1
q2
q1
Brown bag 25 / 46
Read All About It!
Yilmaz and Bindel“Effects of Imperfections on Solid-Wave Gyroscope Dynamics”Proceedings of IEEE Sensors 2013, Nov 3–6.
Thanks to DARPA MRIG + Sunil Bhave and Laura Fegely.
Brown bag 26 / 46
Can One Hear the Shape of a Drum?
�r2u = �u on ⌦
u = 0 on @⌦
Brown bag 27 / 46
What Do You Hear?
S
A B
Size of bottlenecks (Cheeger inequality)
h 2
p�2
Brown bag 28 / 46
What Do You Hear?
Volume (Weyl law)
lim
x!1
N(x)
xd/2= (2⇡)�d!
d
vol(⌦), N(x) = {# eigenvalues x}
Brown bag 29 / 46
Can One Hear the Shape of a Graph?
From eigenvalues of adjacency, Laplacian, normalized Laplacian?
Brown bag 30 / 46
What Do You Hear?
Size of separators (Cheeger inequality)
Brown bag 31 / 46
What Do You Hear?
What information hides in the eigenvalue distribution?1 Discretizations of Laplacian: something like Weyl’s law2 Sparse random graphs: Wigner semicircular distribution3 “Real” networks: less well understood
But computing all eigenvalues seems expensive!
Brown bag 32 / 46
Exploring Spectral Densities
Kernel polynomial method (see Weisse, Reviews of Modern Physics)Think of spectral distribution as a generalized function
Z 1
�1µ(x)f(x) dx =
1
N
NX
k=1
f(�k
)
Write f(x) =P1
j=1 cjTj
(x) and µ(x) =P1
j=1 dj�j
(x), whereR 1�1 �j
(x)Tk
(x) dx = �jk
Estimate dj
=
1N
tr(Tj
(A)) =
1N
E[zTTj
(A)z], z a random probevectorTruncate series for µ(x) and filter (avoid Gibbs)
Much cheaper than computing all eigenvalues!
Brown bag 33 / 46
Exploring Spectral Densities (with David Gleich)
Consider spectrum of normalized Laplacian (random walk matrix)Approximate via KPM and compare to full eigencomputation
Things we knowEigenvalues in [�1, 1]; nonsymmetric in generalStability: change d edges, have
�j�d
ˆ�j
�j+d
kth moment = probability of return after k-step random walkEigenvalue cluster near 1 ⇠ well-separated clustersEigenvalue cluster near 0 ⇠ triangles connected by one node
What else can we “hear”?
Brown bag 34 / 46
Erdos
−1 −0.5 0 0.5 1 1.50
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
20
40
60
80
100
120
140
160
180
200
Brown bag 35 / 46
Internet topology
−1 −0.5 0 0.5 1 1.50
2000
4000
6000
8000
10000
12000
14000
16000
18000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
Brown bag 36 / 46
Marvel characters
−0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1−200
0
200
400
600
800
1000
1200
Brown bag 37 / 46
Marvel comics
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
1000
2000
3000
4000
5000
6000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
Brown bag 38 / 46
PGP
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
500
1000
1500
2000
2500
3000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
Brown bag 39 / 46
Yeast
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
100
200
300
400
500
600
700
800
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
10
20
30
40
50
60
70
80
90
100
Brown bag 40 / 46
Enron emails (SNAP)
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
2000
4000
6000
8000
10000
12000
14000
Brown bag 41 / 46
Reuters911 (Pajek)
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
500
1000
1500
2000
2500
3000
3500
Brown bag 42 / 46
US power grid (Pajek)
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
100
200
300
400
500
600
700
Brown bag 43 / 46
DBLP 2010 (LAW)
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6x 10
4
N = 326186, nnz = 1615400, 80 s (1000 moments, 10 probes)
Brown bag 44 / 46
Hollywood 2009 (LAW)
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
2
2.5
3
3.5
4x 10
5
N = 1139905, nnz = 113891327, 2093 s (1000 moments, 10 probes)
Brown bag 45 / 46
What Do You Hear?
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
20
40
60
80
100
120
140
160
180
200
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
−0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1−200
0
200
400
600
800
1000
1200
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
50
100
150
200
250
300
350
400
450
500
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
10
20
30
40
50
60
70
80
90
100
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
2000
4000
6000
8000
10000
12000
14000
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
500
1000
1500
2000
2500
3000
3500
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
100
200
300
400
500
600
700
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
1
2
3
4
5
6x 10
4
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 10
0.5
1
1.5
2
2.5
3
3.5
4x 10
5
Brown bag 46 / 46