the empirical cross gramian for parametrized nonlinear systems
TRANSCRIPT
The Empirical Cross Gramian for ParametrizedNonlinear Systems
Christian Himpe ([email protected])Mario Ohlberger ([email protected])
WWU MünsterInstitute for Computational and Applied Mathematics
MathMod 201519.02.2015
Networks
The brain:A fairly complex network.
Excited by external sensory stimuli.Information propagation due toconnectivity between regions.Usually, no direct measurement ofneuronal activity.
Modelling
Adjacency Matrix:
A =
a11 a12 . . . a1Na21 a22 . . ....
.... . .
aN1 aNN
(aij = Connection from j-th node to i-th node.)
Controlled External InputSmall Number of (Output) SensorsParametrized Connectivity (Components of A)
→ (Parametrized) Control System
Model
(Dynamical) Input-Output System
Linear (Time-Invariant) System:x(t) = Ax(t) + Bu(t)
y(t) = Cx(t)
(x(t) ∈ RN , u(t) ∈ RM , y(t) ∈ RO ,A ∈ RN×N ,B ∈ RN×M ,C ∈ RO×N )
General (Possibly Nonlinear) System:x(t) = f (x(t), u(t))
y(t) = g(x(t), u(t))
(f : R×RN ×RM → RN , g : R×RN ×RM → R
O )
Model Order Reduction
Input-Output Mapping:u ∈ LM
2 [0,∞)→ RN → y ∈ LO
2 [0,∞)
(N � 1, dim(u(t))� N, dim(y(t))� N)
Linear Reduced Order Model:xr (t) = Arxr (t) + Bru(t)
yr (t) = Crxr (t)
(dim(xr (t))� dim(x(t)) and ‖y − yr‖ � 1 in a suitable norm.)
General Reduced Order Model:xr (t) = fr (xr (t), u(t))
yr (t) = gr (xr (t), u(t))
(dim(xr (t))� dim(x(t)) and ‖y − yr‖ � 1 in a suitable norm.)
Projection-Based MOR
Petrov-Galerkin Projection:U ∈ RN×n,V ∈ Rn×N ,VU = 1
(V = UT Galerkin projection)
Linear ROM:xr (t) = VAUxr (t) + VBu(t)
yr (t) = CUxr (t)
(Generally fast.)
General ROM:xr (t) = Vf (Uxr (t), u(t))
yr (t) = g(Uxr (t), u(t))
(Generally slow.)
Hankel Singular Values i.e. [Antoulas’05]
Impulse Response:h(t) := CeAtB
(x0 = 0 and u(t) = δ(t))
Convolution Operator:
S(u)(t) = (h ∗ u)(t) =
∫ ∞0
CeA(t−τ)Bu(τ)dτ = y(t)
(S has infinite spectrum)
Hankel Operator:
H(u)(t) := (S ◦ F )(u)(t)
∫ 0
−∞CeA(t+τ)Bu(τ)dτ
(Time-flip operator F : [0,∞)→ (−∞, 0]; H has finite spectrum)
Cross Gramian [Fernando & Nicholson’83]
Factorization of Hankel Operator:H = OC
(With controllability and observability operator factors)
Cross Gramian:
WX := CO =
∫ ∞0
eAtBCeAtdt
(For square systems: #inputs = #outputs)
Core Property of the Cross Gramian:OC = (OC)∗ ⇒ |λi (WX )| = σi (H)
(For symmetric systems)
Direct Truncation / Approximate Balancing
Singular Value Decomposition of Cross Gramian:
WXSVD= UDV
(Truncated SVD in production code.)
Truncation of singular vectors yields Galerkin projectionU =
(U1 U2
)V1 = UT
1
(Two-sided Petrov-Galerkin is not guaranteed to be stable.)
Cross-Gramian-Based ROM:xr (t) = UT
1 AU1xr (t) + UT1 Bu(t)
yr (t) = CU1xr (t)
(Initial condition is projected too: x0,r = UT1 x0. Equivalent to BT for state-space symmetric systems.)
Empirical Gramians [Lall et al’99]
Snapshots:X =
(x0 . . . xT
), Y =
(y0 . . . yT
)(Discrete state and output trajectories.)
Perturbation Sets:QU = {ei=1...M} × {Rj ∈ RM×M |RjRT
j = 1} × {qk ∈ R}QX = {ei=1...N} × {Rl ∈ RN×N |RlRT
l = 1} × {qm ∈ R}(Perturbing (impulse) input and initial state.)
Empirical Controllability and Observability Gramian:
WC =
|QU |∑q=1
XqXTq , WO =
|QX \{ei}|∑q=1
(Y...q,1, . . . ,Y
...q,N)T (Y
...q,1, . . . ,Y
...1,N)
(WC = WC and WO = WO for linear systems.)
Empirical Cross Gramians [Streif et al’06, H. & Ohlberger’14]
Cross Gramian:
WX =
∫ ∞0
(eAtB)︸ ︷︷ ︸u→x
(CeAt)︸ ︷︷ ︸x→y
dt
(The (inner) product of controllability and observability operator.)
Empirical Cross Gramian:
WX :=1
|QU ||QX |
|QU |∑h=1
|QX |∑i=1
∫ ∞0
Ψhij(t)dt ∈ RN×N ,
Ψhijk,i (t) = (xhj
k (t)− xk)(y ijh (t)− yh)T ∈ R,
(WX = WX for linear systems.)
parametric Model Order Reduction
Parametrized General System:x(t) = f (x(t), u(t), θ)
y(t) = g(x(t), u(t), θ)
(θ ∈ RP = Θ)
Reduced Parametrized General System:xr (t) = fr (xr (t), u(t), θ)
yr (t) = gr (xr (t), u(t), θ)
(dim(xr (t))� dim(x(t)) and ‖y − yr‖ � 1 for all valid θ in a suitable norm.)
Linear Parametrization [Sun & Hahn’06]
Linear Parametric System:
x = Ax + Bu + Fθ → x = Ax +(B F
)( uuθ
)y = Cx
(F ∈ RN×P ; treat parameters as additional inputs.)
Parametric Controllability Gramian:
WC = WC ,0 +P∑
i=1
WC ,i ,
WC ,0 =
∫ ∞0
eAtBBT eAT tdt,
WC ,i>0 =
∫ ∞0
eAtFiFTi eAT tdt
Parametrized Empirical Cross Gramian
Discretized Parameter Space:QΘ =
(θ1 . . . θJ
)(Around a nominal parameter.)
Extended Perturbation Set:QU × QX × QΘ
(As opposed to the linear case where (QU ∪ QΘ)× QX .)
Parametrized Empirical Cross Gramian:
WX :=1
|QU ||QX ||QΘ|
|QU |∑h=1
|QX |∑i=1
|QΘ|∑j=1
∫ ∞0
Ψhij(t)dt ∈ RN×N ,
Ψhijk,i (t) = (xhj
k (t)− xk)(y ijh (t)− yh) ∈ R,
(Parameters are treated like inputs or states and are perturbed.)
Nonlinear Symmetry [Ionescu et al’09]
Linear Symmetry:OC = (OC)∗ ⇔ CeAB = (CeAB)∗
⇔ CA−1B = (CA−1B)∗
⇔ CAkB = (CAkB)∗
(Symmetry property for impulse response, system gain (transfer function) and Markov parameter.)
Nonlinear Symmetry:ImO = Im C+ ⇔ ∃WX : O = C+(WX ) = C+CO
⇔ ∃W +X : C+ = O(W +
X ) = OO+C+
(All SISO systems are symmetric!)
(Nonlinear) RC Ladder Benchmark [Chen’99]
Benchmark Problem [MORwiki http://modelreduction.org]Nonlinear SISO System (!)Variable State DimensionAlso used, for example, in [Condon & Ivanov’04]
(Source: C. Himpe “Nonlinear RC Ladder”, MORwiki, 2014)
(Resistor-capacitor cascade with nonlinear resistors (diodes).)
(Nonlinear) RC Ladder Benchmark II
x(t) =
−g(x1(t))− g(x1(t)− x2(t)) + u(t)g(x1(t)− x2(t))− g(x2(t)− x3(t))
...g(xk−1(t)− xk(t))− g(xk(t)− xk+1(t))
...g(xN−1(t)− xN(t))
,
y(t) = x1(t),
with:g(xi ) = exp(10xi ) + xi − 1
⇒ gθ(xi ) = exp(10xi ) + θixi − 1(dim(x(t)) = 1000⇒ θ ∈ R1000, For experiments: θ ∈ U 1
2 , 32
)
emgr - Empirical Gramian Framework (Version: 2.9, 01.2015)
Gramians:Empirical Controllability GramianEmpirical Observability GramianEmpirical Cross GramianEmpirical Linear Cross GramianEmpirical Sensitivity GramianEmpirical Identifiability GramianEmpirical Joint Gramian
Features:Uniform InterfaceCompatible with MATLAB & OCTAVE (& FREEMAT)Vectorized & ParallelizableOpen-Source licensed
More info at: http://gramian.de
State Reduction
10 20 30 40 50 60 70 80 90 100
10−15
10−10
10−5
100
Reduced States
Re
lative
Ou
tpu
t E
rro
r
Balanced Truncation (L2)
Balanced Truncation (L∞)
Cross Gramian (L2)
Cross Gramian (L∞)
0 50 100 150
WX
BT
Offline Time [s]
Parametrized State Reduction (Offline:3, Online:100 Samples)
10 20 30 40 50 60 70 80 90 100
10−15
10−10
10−5
100
Reduced States
Re
lative
Ou
tpu
t E
rro
r
Balanced Truncation (L2)
Balanced Truncation (L∞)
Cross Gramian (L2)
Cross Gramian (L∞)
0 50 100 150 200 250 300 350 400 450 500
WX
BT
Offline Time [s]
Outlook
Application to neuronal network models,in the context of neuroimaging data inversion (ongoing)
Non-Symmetric Cross Gramian (Preprint: http://arxiv.org/pdf/1501.05519)
Really Large SVDs(Distributed Memory) Parallelization
tl;dl
Cross-gramian-based model reduction,for parametrized nonlinear systems,using the empirical cross gramian,averaged over a discretized parameter space,around a steady state.
http://gramian.de
Thanks!
Get the Companion Code: http://j.mp/mathmod15