Principle Component Analysis and Model Reduction for ...sorensen/Talks/SymmSVD/VAtech.pdf · Principle Component Analysis and Model Reduction for ... I Brief Intro to Model Reduction

Principle Component Analysis

and

Model Reduction for Dynamical Systems

D.C. Sorensen

Virginia Tech 12 Nov 2004

Protein Substate Modeling and Identification

PCA Dimension Reduction Using the SVD• Tod Romo • George Phillips

T.D. Romo, J.B. Clarage, D.C. Sorensen and G.N. Phillips, Jr., Automatic Identification of

Discrete Substates in Proteins: Singular Value Decomposition Analysis of Time Averaged

Crystallographic Refinements, Proteins: Structure, Function, and Genetics22,311-321,(1995).

Outline

I Brief Intro to Model Reduction for Dynamical Systems

I Reduced Basis Trajectory Time Integration for MD

I The Symmetric SVD: Reduced Dimension MD Simulation

D.C. Sorensen 3

LTI Systems and Model Reduction

Time Domain

x = Ax + Bu

y = Cx

A ∈ Rn×n, B ∈ Rn×m, C ∈ Rp×n, n >> m, p

Frequency Domain

sx = Ax + Bu

y = Cx

Transfer Function

H(s) ≡ C(sI− A)−1B, y(s) = H(s)u(s)

D.C. Sorensen 4

Model Reduction

Construct a new system {A, B, C} with LOW dimension k << n

˙x = Ax + Bu

y = Cx

Goal: Preserve system response

y should approximate y

Projection: x(t) = V ˆx(t) and V ˙x = AVx + Bu

D.C. Sorensen 5

Model Reduction by (Krylov) Projection

Approximate x ∈ SV = Range(V) k-diml. subspacei.e. Put x = Vx, and then force

WT [V ˙x − (AVx + Bu)] = 0

y = CVx

If WTV = Ik , then the k dimensional reduced model is

˙x = Ax + Bu

y = Cx

where A = WTAV, B = WTB, C = CV.

D.C. Sorensen 6

Moment Matching ↔ Krylov Subspace Projection

Pade via Lanczos (PVL)

Freund, Feldmann

Bai

Multipoint Rational Interpolation

Grimme

Gallivan, Grimme, Van Dooren

Gugercin, Antoulas, Beattie

D.C. Sorensen 7

Gramian Based Model Reduction

Proper Orthogonal Decomposition (POD)Principle Component Analysis (PCA)

x(t) = f(x(t),u(t)), y = g(x(t),u(t))

The gramian

P =

∫ ∞

ox(τ)x(τ)Tdτ

Eigenvectors of P

P = VS2VT

Orthogonal Basisx(t) = VSw(t)

D.C. Sorensen 8

PCA or POD Reduced Basis

Low Rank Approximation

x ≈ Vk xk(t)

Galerkin condition – Global Basis

˙xk = VTk f(Vk xk(t),u(t))

Global Approximation Error (H2 bound for LTI)

‖x− Vk xk‖2 ≈ σk+1

Snapshot Approximation to P

P ≈ 1

m

m∑j=1

x(tj)x(tj)T

D.C. Sorensen 9

SVD of Snapshot Trajectory (Conformations)

X = [x(t1), x(t2), . . . , x(tm)]

SVD of X:

X = VSWT ≈ VkSkWTk

where

VTV = WTW = In S = diag(σ1, σ2, · · · , σn)

with σ1 ≥ σ2 ≥ · · · ≥ σn.

D.C. Sorensen 10

SVD Compression

m

k ( m + n)

m x n

v.s.

Storage

Advantage of SVD Compression

k

n

0 20 40 60 80 100 120 140 160 180 2000

10

20

30

40

50

60

70

80

90

100SVD of Clown

D.C. Sorensen 11

Image Compression - Feature Detection

original rank = 10

rank = 30 rank = 50

POD in CFD

Extensive Literature

Karhunen-Loeve, L. Sirovich

Burns, King

Kunisch and Volkwein

Many, many others

Incorporating Observations – Balancing

Lall, Marsden and Glavaski

K. Willcox and J. Peraire

D.C. Sorensen 13

POD for LTI systems

Impulse Response: H(t) = C(tI− A)−1B, t ≥ 0

Input to State Map: x(t) = eAtB

Controllability Gramian:

P =

∫ ∞

ox(τ)x(τ)Tdτ =

∫ ∞

oeAτBBT eAT τdτ

State to Output Map: y(t) = CeAtx(0)

Observability Gramian:

Q =

∫ ∞

oeAT τCTCeAτdτ

D.C. Sorensen 14

Balanced Reduction (Moore 81)

Lyapunov Equations for system Gramians

AP + PAT

+ BBT

= 0 ATQ+QA + C

TC = 0

With P = Q = S : Want Gramians Diagonal and Equal

States Difficult to Reach are also Difficult to Observe

Reduced Model Ak = WTk AVk , Bk = WT

k B , Ck = CkVk

I PVk = WkSk QWk = VkSk

I Reduced Model Gramians Pk = Sk and Qk = Sk .

D.C. Sorensen 15

Hankel Norm Error estimate (Glover 84)

Why Balanced Realization?

I Hankel singular values =√

λ(PQ)

I Model reduction H∞ error (Glover)

‖y − y‖2 ≤ 2× (sum neglected singular values)‖u‖2I Extends to MIMO

I Preserves Stability

Key Challenge

I Approximately solve large scale Lyapunov Equationsin Low Rank Factored Form

D.C. Sorensen 16

CD Player Impluse Response

n = 120 k = 11 , tol = 5e-3

Frequency (rad/sec)

Sin

gula

r V

alue

s (d

B)

Sigma Plot: n = 120 CDplayer

100

101

102

103

104

105

−150

−100

−50

0

50

Frequency (rad/sec)

Sin

gula

r V

alue

s (d

B)

Error plot

100

101

102

103

104

105

−90

−80

−70

−60

−50

−40

−30

−20

CD Player Impluse Response

k = 17 , tol = 5e-4

Frequency (rad/sec)

Sin

gula

r V

alue

s (d

B)

Sigma Plot: n = 120 CDplayer

100

101

102

103

104

105

−150

−100

−50

0

50

Frequency (rad/sec)

Sin

gula

r V

alue

s (d

B)

Error plot

100

101

102

103

104

105

−80

−70

−60

−50

−40

−30

CD Player Impluse Response

k = 31 , tol = 5e-5

Frequency (rad/sec)

Sin

gula

r V

alue

s (d

B)

Sigma Plot: n = 120 CDplayer

100

101

102

103

104

105

−150

−100

−50

0

50

Frequency (rad/sec)

Sin

gula

r V

alue

s (d

B)

Error plot

100

101

102

103

104

105

−100

−90

−80

−70

−60

−50

−40

CD Player - Hankel Singular Values

0 20 40 60 80 100 12010

−14

10−12

10−10

10−8

10−6

10−4

10−2

100

102

Hankel Singular Values

D.C. Sorensen 20

Reduction of Second Order Systems

Mx + Gx + Kx = Buy(t) = Cx(t)

→ M¨x + G ˙x + Kx = Bu

y(t) = Cx(t)

where M = VTMV, etc. with VTV = I.

Key Point: Preserve Second Order Form– DO NOT convert to First Order Sys.

Keeps Physical Meaning - can be built

D.C. Sorensen 21

Applications

Mechanical Systems Electrical Systems MEMS devices

e.g. Building Model N = 26394, k = 200 (ROM)

0 5 10 15 20 25 300

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5Frequency Response α = .67 , β = .0032

Frequency ω

Res

pons

e |G

( jω

)|

reducedoriginal

D.C. Sorensen 22

Error Bound for Second Order Systems

A. Antoulas, C. Teng

Controllability Gramian - Impulse Response

P :=

∫ ∞

0x(t)x(t)∗dt.

Reduce with Dominant Eigenspace P: PV1 = V1S1

Bounded H2 norm of error system E = Σ− Σ

‖E‖2H2≤ Cotr{S2}

Key: Expression for P in frequency domain.

D.C. Sorensen 23

PCA Model Reduction for Molecular Dynamics

• Rachel Vincent • Monte Pettitt

D.C. Sorensen 24

Classical Equations of Motion

Molecular dynamics (MD) simulation is a computational tool usedto study a molecular system as it evolves through time. Newton’ssecond law of motion governs atomic motion in MD:

Mr(t) = −∇V(r(t)).

I r(t) = vector of atomic coordinates at time t= [x1t y1t z1t · · · xNat yNat zNat ]

T

I M = diagonal matrix of atomic masses

I V(r(t)) = potential energy function

D.C. Sorensen 25

Time Step Barrier ∝ fs

Example: DHFR (dihydrofolate reductase), 23,558 atoms

To realize a microsecond simulation with a time step of2 fs would require about 13 months of simulation

time when utilizing 126 processors.

Time with respect to simulation using the NAMD program (NotAnother Molecular Dynamics program) on an Origin 2000R10000/250.

D.C. Sorensen 26

PCA Reduced Basis Simulation

MVy(t) = −∇V(Vy(t)) → My(t) = −VT∇V(Vy(t))

1. Initial Basis V: truncated SVD of short traditional MDtrajectory using ARPACK.

2. Approximate the reduced basis potential energy with RadialBasis Fit.

3. Update reduced basis positions y and velocities y in kdimensions using the approximate potential.

4. Reconstruct 3ND trajectory r = Vy.

5. Update and truncate reduced basis and perform full spacecorrection as needed.

D.C. Sorensen 27

Remarks

I Butane (n = 42):80% to 90% of the total motion with 10 - 15 LSVs (24% - 36%

DOF)

Reduced Simulation times order of secondsTraditional MD simulation took several minutes.

I BPTI(n = 2700):80% to 90% of the total motion with 300 - 500 LSVs (11% - 19%

DOF)

Reduced Simulation times order of minutesTraditional MD simulation took several hours.

D.C. Sorensen 28

Symmetry Preserving SVD (Mili Shah)

Collaboration with the Physical and Biological Computing Group

I Lydia Kavraki

I Mark Moll

I David Schwarz

I Amarda Shehua

I Allison Heath

D.C. Sorensen 29

Symmetry in HIV-1 protease

Backbone representation of HIV-1 protease (from M. Moll)

bound to an inhibitor (shown in orange)Uses PCA dimension reduction of Molecular Dynamics Simulations

Symmetry across a plane should be present

D.C. Sorensen 30

Animation: Symmetric SVD Approximation

−1.5 −1 −0.5 0 0.5 1 1.5−1.5

−1

−0.5

0

0.5

1

1.5Comparison of Regular and Symmetric Preserving SVD

−1.5 −1 −0.5 0 0.5 1 1.5−1.5

−1

−0.5

0

0.5

1

1.5Comparison of Regular and Symmetric Preserving SVD

click below figures for movies

D.C. Sorensen 31

Finding the Plane of Symmetry

Suppose

X = [x1, x2, . . . , xn] and Y = [y1, y2, . . . , yn]

are two sets of points symmetric across a plane

Exact symmetry condition:

Y = (I− 2wwT )X,

where w is the normal to the (hyper-) Plane of Symmetry

H = {x : wTx = 0}

Remark: In Numerical Linear Algebra (I− 2wwT ) is aHouseholder Transformation or Elementary Reflector

D.C. Sorensen 32

Best Approximate Plane of Symmetry

Symmetry condition with Noise:

Y = (I− 2wowTo )X + E,

Problem: Compute a unit vector w that gives the bestApproximate Plane of Symmetry

minw‖Y − (I− 2wwT )X‖F ,

Solution:

(XYT + YXT )v = vλmin, w = v

gives the normal w to the best approximate plane of symmetry

D.C. Sorensen 33

Supressing Outlier Effects

Iteratively determine diagonal weighting matrix Dw

The i-th diagonal of Dw is 1/discrepancy,

discrepancy = ‖yi − (I− 2wwT )xi‖Problem: Compute a unit vector w that gives the best

Weighted Approximate Plane of Symmetry

minv‖[Y − (I− 2vvT )X]Dw‖F ,

Solution:

(XD2wYT + YD2

wXT )v = vλmin, w← v

gives the normal w to the best weighted approximate plane ofsymmetry

D.C. Sorensen 34

Finding Normal to “Best” Plane of Symmetry

−1.5 −1 −0.5 0 0.5 1

−1

−0.5

0

0.5

1

click below for movie

D.C. Sorensen 35

The Symmetric SVD Approximation

If WX2 = X1 + E where W = blockdiag(I− 2wwT )

minWX2=X1

∥∥∥∥(X1

X2

)−

(X1

X2

)∥∥∥∥2

F

and(

X1

X2

)= USVT

Solved by:

U =1√2

(U1

U2

), S =

√2S1, V = V1. and U2 = WU1,

with

U1S1VT1 = 1

2 (X1 + WX2)

D.C. Sorensen 36

Symmetric Major Modes: HIV-1 protease

I Major mode regular SVD is red

I Major mode SYMMETRIC SVD is blue

I 3120 atoms (3*3120=9360 degrees of freedom)

I MD trajectory consisted of 10000 conformations (NAMD)

I SVD and SymSVD used P ARPACK on a Linux cluster

I dual-processor nodes; 1600MHz AMD Athlon processors, 1GBRAM per node. 1GB/s Ethernet connection . 12 Processors= 6 nodes.

I First 10 standard singular vectors: 88 secs.

I First 10 symmetric singular vectors: 131 secs.

D.C. Sorensen 37

Animation: Symmetric SVD on HIV1 Protease

click for movie

Red = Unsymmetric Blue = Symmetric

First SVD mode – Symmetric vs. Unsymmetric

D.C. Sorensen 38

Rotational Symmetry

−1.5 −1 −0.5 0 0.5 1 1.5−1.5

−1

−0.5

0

0.5

1

1.5Perfect Rotational Symmetry No Noise

Xj = WXj−1, j = 1 : k − 1, where W = I−QGQT

Ip − G is a rotation Xk = X0.

D.C. Sorensen 39

Finding the Axis of Rotation

q is an axis of rotation iff QTq = 0

qTW = qT (I−QGQT ) = qT ⇒ qTX0 = qTXj

Let M = (k − 1)X0 −∑k−1

j=1 Xj

min‖q‖=1

‖MTq‖ (= 0 if exact symmetry holds)

Good for noisy data

(for another condition see Minovic,Ishikawa and Kato )

D.C. Sorensen 40

Best Rotationally Symmetric Approximation

If Wk−jXj = X0 + Ej , j = 1 : k − 1

minbXj+1=WbXj

∥∥∥∥∥∥ X0

:Xk−1

− X0

:

Xk−1

∥∥∥∥∥∥2

F

=1

k

k−1∑j=0

‖Ej‖2F ,

[XT

0 . . . XTk−1

]T= USVT with U =

1√k

[UT

0 . . .UTk−1

]T

U0S0VT0 =

1

k(X0 + Wk−1X1 + Wk−2X2 + ... + WXk−1).

S =√

kS0 V = V0 Uj = WjU0, j = 0, 1, 2, ..., k − 1

D.C. Sorensen 41

Animation: Rotationally Symmetric SVDApproximation

−1.5 −1 −0.5 0 0.5 1 1.5−1.5

−1

−0.5

0

0.5

1

1.5Comparison of Symmetric and Regular SVD

−1.5 −1 −0.5 0 0.5 1 1.5−1.5

−1

−0.5

0

0.5

1

1.5Comparison of Symmetric and Regular SVD

click below figures for movies

D.C. Sorensen 42

Animation: Rotationally Symmetric SVD on HIV1

click for movie

Red = Unsymmetric Blue = Symmetric

Second SVD mode – Rotationally Symmetric vs. Unsymmetric

D.C. Sorensen 43

Potential for Symmetric SVD

I Obtain a Symmetric PCA reduced dimension approximatetrajectory

I Test Hypothesis of Symmetry in an Unknown Protein

I Locate Symmeteric Sub-Structures

Things to Do:

I Improve convergence rate for finding w

I Give a complete analysis of convergence

I Give a complete analysis of discrepancy weighting

I Extend to more complex symmetries

I Find New Applications

D.C. Sorensen 44

Contact Info.

e-mail: [email protected]

web page: www.caam.rice.edu/ ˜ sorensen/

ARPACK: www.caam.rice.edu/software/ARPACK/