model reduction in the loewner framework - some issues€¦ · kit, november 17 - 18 2015 thanos...
TRANSCRIPT
Model Reduction in the Loewner FrameworkSome issues
Thanos Antoulas
Rice University, Houston, and Jacobs University, Bremen
e-mail: [email protected]: www.ece.rice.edu/ aca
Workshop: MOR 4 MEMS
KIT, November 17 - 18 2015
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 1 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 2 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 3 / 44
The Loewner matrix
Given:row array (µj , vj), j = 1, · · · , k , and
column array (λi ,wi ), i = 1, · · · , q,
the associated Loewner matrix is:
L =
v1−w1
µ1−λ1· · · v1−wq
µ1−λq
.... . .
...vk−w1
µk−λ1· · · vk−wq
µk−λq
∈ Ck×q
If there is a known underlying function g, then wi = g(λi ), and vj = g(µj).
Main property
Let L be a p × k Loewner matrix. Then p, k ≥ deg g ⇒ rankL = deg g.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 4 / 44
Karel Lowner 1893 - 1968
• Born in Bohemia
• Studied in Prague under Georg Pick
• Emigrated to the US in 1939
• adapted his name to Charles Loewner
Seminal paper: Uber monotone Matrixfunctionen, Math. Zeitschrift (1934).
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 5 / 44
General framework – tangential interpolation
Given: • right data: (λi ; ri ,wi ), i = 1, · · · , k;
• left data: (µj ; `∗j , v∗j ), j = 1, · · · , q.
We assume for simplicity that all points are distinct.
Problem: Find rational p ×m matrices H(s), such that
H(λi )ri = wi `∗j H(µj ) = v∗j
Also define:
Λ =
λ1
. . .
λk
∈ Ck×k ,
R = [r1 r2, · · · rk ] ∈ Cm×k ,
W = [w1 w2 · · · wk ] ∈ Cp×k ,
Left data:
M =
µ1
. . .
µq
∈Cq×q, L =
`∗1...`∗q
∈Cq×p,V =
v∗1...
v∗q
∈ Cq×m
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 6 / 44
Tangential interpolation: the Loewner pencil
Recall data: H(λi )ri = wi , `∗j H(µj ) = v∗j .
The Loewner matrix L ∈ Cq×k is:
L =
v∗1 r1−`∗1 w1
µ1−λ1· · · v∗1 rk−`∗1 wk
µ1−λk
.... . .
...v∗q r1−`∗q w1
µq−λ1· · ·
v∗q rk−`∗q wk
µq−λk
if we are given H(s) = C(sE−A)−1B:
Ri = (λiE− A)−1Bri ⇒
R: generalized reachability matrix
O∗j = `∗j C(µjE− A)−1 ⇒
O: generalized observability matrix.
L can be factored as
⇒ L = −OER
The Loewner matrix L ∈ Cq×k is:
L =
v∗1 r1−`∗1 w1
µ1−λ1· · · v∗1 rk−`∗1 wk
µ1−λk
.... . .
...v∗q r1−`∗q w1
µq−λ1· · ·
v∗q rk−`∗q wk
µq−λk
L can be factored as
⇒ L = −OER
The shifted Loewner matrix L ∈ Cq×k is:
Lσ =
µ1v∗1 r1−`∗1 w1λ1
µ1−λ1· · · µ1v∗1 rk−`∗1 wkλk
µ1−λk
.... . .
...µqv∗q r1−`∗q w1λ1
µq−λ1· · ·
µqv∗q rk−`∗q wkλk
µq−λk
Lσ can be factored as
⇒ Lσ = −OAR
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 7 / 44
Structure of the Loewner pencil
Proposition. The controllability and observability matrices satisfy:
AR+ B R = ERΛ, OA + L C = MO E,
and hence:
Lσ = LΛ + VR Lσ = ML + LW
Furthermore L and Lσ satisfy:
ML− LΛ = VR− LW, MLσ − LσΛ = M VR− LWΛ.
Corollary. All systems (C, A, B), with E = I, satisfying given interpolation conditions, can beparametrized as follows:
A = Λ− BR, B =[bj], C =
[ηj], i , j = 1, · · · , k,
where B is free and Λ, R, C are given.
Corollary. Placement of poles and zeros.
N If ri = 0 (wi arbitrary) or `j = 0 (vj arbitrary) ⇒ λi or µj is a pole.
N If wi = 0 (ri arbitrary) or vj = 0 (`j arbitrary) ⇒ λi or µj is a zero.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 8 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 9 / 44
Construction of Interpolants (Models)
• If the pencil (Lσ , L) is regular, i.e.
Φ(s) = Lσ − s L, is invertible, then
E = −L, A = −Lσ , B = V, C = W
is a minimal interpolant of the data ⇒ H(s) = W Φ(s)−1 V
• If Φ(s) = Lσ − s L, is singular, let
Φ(s)# be a generalized inverse of Φ(s)
(Drazin or Moore-Penrose).
⇒ H(s) = W Φ(s)# V
• In the latter case, if the numerical rankL = k, compute the rank revealing SVD:
L = YΣX∗ ≈ YkΣkX∗k
Theorem. A realization [C,E,A,B], of an approximate interpolant is given as follows:
E = −Y∗kLXk , A = −Y∗kLσXk , B = Y∗k V, C = WXk .
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 10 / 44
A simple example Consider the system
x1(t) = x2(t),
x2(t) = −x1(t)− x2(t) + u(t), y(t) = x2(t) ⇒ H(s) =
s
s2 + s + 1.
We now wish to recover state equations equivalent to the ones above from measurements of thetransfer function.
Data: obtained by evaluating the transfer function at λ1 = 12
, λ2 = 1, as well as µ1 = − 12
,µ2 = −1. The corresponding values of H are collected in the matrices
W =(
27
13
), V =
(− 2
3−1
)T.
Furthermore with R = [1 1], L = RT , we construct the Loewner pencil:
L =
[ 2021
23
67
23
], Lσ =
[− 4
210
− 47− 1
3
].
Since the pencil (Lσ , L) is regular, and the rank of both matrices is two:
H(s) = WΦ(s)−1V =s
s2 + s + 1, where Φ(s) = Lσ − s L.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 11 / 44
Hence, the measurements above yield a minimal (descriptor) realization of the system in termsof the (state) variables ξ1, ξ2:
2021ξ1(t) + 2
3ξ1(t) = − 4
21ξ1(t) + 2
3u(t),
67ξ1(t) + 2
3ξ2(t) = − 4
7ξ1(t)− 1
3ξ2(t) + u(t),
y(t) = 27ξ1(t) + 1
3ξ2(t).
Question: what happens if we collect more data that necessary:
Λ = diag(
12
1 32
2), M = diag
(− 1
2−1 − 3
2−2
).
In this case, the associated measurements are
W =(
27
13
619
27
), V =
(− 2
3−1 − 6
7− 2
3
)T,
and with R = [1 1 1 1], L = RT , the Loewner pencil is:
L =
2021
23
2857
821
67
23
1019
37
47
1021
52133
1649
821
13
1657
521
, Lσ =
− 4
210 4
572
21
− 47
− 13
− 419
− 17
− 47− 8
21− 36
133− 10
49
− 1021
− 13
− 1457
− 421
.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 12 / 44
It turns out that we can choose arbitrary X,Y ∈ R4×2, such that YT X is nonsingular, e.g.
X =
−1 0
0 −10 0−2 1
, YT =
[0 1 0 −11 −1 −1 1
],
so that the projected quantities
W = WX =[− 6
7− 1
21
], L = YTLX =
[− 6
7− 1
7
1849
1147
],
Ls = YTLσX =
[0 1
21
− 4849
− 19147
], V = YT V =
[− 1
3
1121
],
constitute a minimal realization of H(s):
H(s) = W(Ls − sL
)−1V =
s
s2 + s + 1.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 13 / 44
There is another way to express the above relationship avoiding arbitrary projectors.
Basic ingredients: the Moore-Penrose generalized inverse and theDrazin generalized inverse.
The Moore-Penrose inverse of the (rectangular) matrix
M ∈ Rq×k , is denoted by MMP ∈ Rk×q , and satisfies:
(a) MMMPM = M, (b) MMPMMMP = MMP ,
(c)[
MMMP]T
= MMMP , (d)[
MMPM]T
= MMPM.
This generalized inverse always exists and is unique.
Given a square matrix M ∈ Rq×q , its index is the least nonneg-ative integer κ such that rank Mκ+1 = rank Mκ.
The Drazin inverse of M is the unique matrix MD satisfying:
(a) Mκ+1MD = Mκ, (b) MDMMD = MD ,
(c) MMD = MDM.
In the sequel we will be concerned with rectangular n ×m polynomial matrices which have anexplicit (rank revealing) factorization as follows:
M = X∆YT ,
where X, ∆, Y have dimension q × n, n × n, n × k, n ≤ q, k, and all have full rank k.
The Moore-Penrose generalized inverse is:
MMP = Y(YT Y)−1∆−1(XT X)−1XT.
If q=k and YTX is invertible, the Drazin generalized inverse is:
MD = X(YT X)−1∆−1(YT X)−1YT.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 14 / 44
Example (continued). The quantities needed are the generalized inverses of
Φ(s) = Lσ − s L =
− 20s21− 4
21− 2s
34
57− 28s
572
21− 8s
21
− 6s7− 4
7− 2s
3− 1
3− 10s
19− 4
19− 3s
7− 1
7
− 4s7− 4
7− 10s
21− 8
21− 52s
133− 36
133− 16s
49− 10
49
− 8s21− 10
21− s
3− 1
3− 16s
57− 14
57− 5s
21− 4
21
= X∆(s)YT
.
Let the common range of the columns of L, Lσ be spanned by the columns of X and thecommon range of the rows of the same matrices by the rows of Y; it follows that
X =
1 0
0 1
− 37
87
− 12
1
, Y =
[1 0 − 7
19− 1
2
0 1 2419
97
]⇒ det (YX) 6= 0.
Thus with ∆(s) = Φ(1 : 2, 1 : 2)(s) there holds Φ(s)MP = 180989667
1s2+s+1
·
·
−28 (11610185s + 7274073) 14 (3558666s − 5604037) 6076 (32301s − 391) 14 (15168851s + 1670036)
294 (225182s + 281171) (−147) (192415s − 19668) −2058 (29494s + 15609) −147 (417597s + 261503)
3724 (54617s + 48189) (−1862) (29046s − 17485) −26068 (5715s + 1523) −1862 (83663s + 30704)
98 (2527157s + 2123670) −49 (1250553s − 876439) −98 (1797669s + 409322) −49 (3777710s + 1247231])
,
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 15 / 44
and Φ(s)D = 14897369
1s2+s+1
·
·
−84 (234677s + 152881) 294 (10652s − 13755) 588 (19079s − 641) 42 (330545s + 29086)
126 (31956s + 42829) −147 (11885s + 4) −882 (4184s + 2255) −63 (67611s + 42841)
684 (19079s + 17063) −798 (4184s − 2171) −4788 (1885s + 441) −342 (31631s + 10550)
42 (330545s + 281368) −147 (22537s − 13751) −294 (31631s + 6124) −21 (533378s + 157609)
.In the rectangular case, where there are two less right measuremnents, i.e we only haveΛ = diag
[12, 1
], while M remains the same, the right values are W = W(:, 1 : 2) and hence
Φ(s) = Ls − s L =
− 20s21− 4
21− 2s
3
− 6s7− 4
7− 2s
3− 1
3
− 4s7− 4
7− 10s
21− 8
21
− 8s21− 10
21− s
3− 1
3
= X ∆(s) YT
,
has dimension 4× 2, where Y = Y(1 : 2, 1 : 2). In this case the Moore-Penrose inverse is
Φ(s)MP =1
737 (s2 + s + 1)
−4767s − 3402 18272
s − 20372
3087s + 294 3297s + 13652
5838s + 5250 −1596s + 903 −4326s − 1218 −4515s − 1722
.
⇒ W Φ(s)MP V = W Φ(s)MP V = W Φ(s)D V = H(s)
Thus, the Loewner framework allows the definition of rectangular and/or singular systems.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 16 / 44
Revisit: Construction of Interpolants
• If the pencil (Lσ , L) is regular, i.e.
Φ(s) = Lσ − s L, is invertible, then
E = −L, A = −Lσ , B = V, C = W
is a minimal interpolant of the data ⇒ H(s) = W Φ(s)−1 V
• If Φ(s) = Lσ − s L, is singular, let
Φ(s)# be a generalized inverse of Φ(s)
(Drazin or Moore-Penrose).
⇒ H(s) = W Φ(s)# V
• In the latter case, if the numerical rankL = k, compute the rank revealing SVD:
L = YΣX∗ ≈ YkΣkX∗k
Theorem. A realization [C,E,A,B], of an approximate interpolant is given as follows:
E = −Y∗kLXk , A = −Y∗kLσXk , B = Y∗k V, C = WXk .
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 17 / 44
Microstrip - data: 1001 S-parameter measurements between 10-18 GHz (CST)
Data frequency response ‖Si,j‖, i, j = 1, 2. Data two singular values.
Singular values of 1001× 1001 Loewner matrix Singular-value fit of model k = 72
S-parameter-error: ∈ [10−6, 10−4] Two singular values of model: ω ∈ [0, 10THz]
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 18 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 19 / 44
An Euler-Bernoulli beam 11
11/2/2015 https://upload.wikimedia.org/wikipedia/commons/e/e9/Cantilever_Beam.svg
https://upload.wikimedia.org/wikipedia/commons/e/e9/Cantilever_Beam.svg 1/1
Fixed End
Free End
BC
w(0, t) = 0, ∂w∂x
(0, t) = 0, E I ∂2w(L,t)
∂x2 + cd I∂3w(L,t)
∂x2∂t= 0,
−E I ∂3w(L,t)
∂x3 − cd I∂4w(L,t)
∂x3∂t= u(t), y(t) = ∂w(L,t)
∂t,
∂2w(x , t)
∂t2+
∂2
∂x2
[E I
∂2w(x , t)
∂x2+ cd I
∂3w(x , t)
∂x2∂t
]= 0,
where E , I , cd are constants. The transfer function is
G(s) =s N(s)
(E I + s cd I )m3(s)D(s)
where m(s) =
[−s2
E I + cd I s
] 14
,
N(s) = cosh(Lm(s)) sin(Lm(s))− sinh(Lm(s)) cos(Lm(s)) and
D(s) = 1 + cosh(Lm(s)) cos(Lm(s)) .
Parameter values: E = 69, GPa = 6.9 · 1010N/m2 - Young’s modulus elasticity constant, I = (1/12) · 7 · 8.53 · 10−11m4 -
moment of inertia, cd = 5 · 10−4 - damping constant, L = 0.7m, b = 7cm, h = 8.5mm - length, base, height of therectangular cross section.
1R. Curtain, K. Morris, Transfer Functions of Distributed Parameter Systems: A Tutorial, Automatica (2009).
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 20 / 44
Model Reduction of the beam
Poles: solutions ofs2 + cd I α4
k s + E I α4k = 0,
where αk are the roots in α, of the equation
1 + cosh(Lα) cos(Lα) = 0.
Reduction methods:
Modal truncation.
FEM followed by Loewner.
Loewner based on the transfer function.
Finally, parametric Loewner with damping as a parameter.
Order of reduced models: 32.
Parametric reduced order model 32, 4.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 21 / 44
Plots
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 22 / 44
Plots
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 23 / 44
Poles and Loewner parametric reduction
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 24 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 25 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 26 / 44
Bilinear systems
Bilinear systems with one input and one output are described by:
Σ : Ex(t) = Ax(t) + Nx(t)u(t) + Bu(t), y(t) = Cx(t),
where x ∈ Rn, u, y ∈ R. Σ is equivalent to the infinite set of systems of the form:
Ex1(t) = Ax1(t) + Bu(t),
Ex2(t) = Ax2(t) + Nx1(t)u(t),
Ex3(t) = Ax3(t) + Nx2(t)u(t), · · ·
The solution is given as x(t) =∑∞
i=0 xi (t).Such systems are equivalent in the frequency domainto an inifnite sequence of rational multivariate functions:
H1(s1) = C [s1E− A]−1 B
H2(s1, s2) = C [s1E− A]−1 N [s2E− A]−1 B
H3(s1, s2, s3) = C [s1E− A]−1 N [s2E− A]−1 N [s3E− A]−1 B, · · ·
...Consequence: model reduction can be performed by means of interpolatory methods.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 27 / 44
Recall: if the underlying transfer function is
H(s) = C(sE− A)−1B,
let O, R be the generalized controllability and the generalized observability matricesassociated with the data: {(µj , vj ), j = 1, · · · , q}, {(λi ,wi ), i = 1, · · · , k}.
O =
C(µ1E− A)−1
...C(µqE− A)−1
, R =[
(λ1E− A)−1B · · · (λkE− A)−1B]
⇒ L = −OER, Lσ = −OAR, V = CR, W = OB
Property generalized: factorization of L, Lσ, V, W
• Fact: if q = k = N ⇒ 2N moments of H are matched.
• the singular values of the Loewner pencil (Lσ ,L) provide a trade-off between accuracy of fitand complexity of the reduced model.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 28 / 44
Nested multi-tuples
The nested right multi-tuples and the nested left multi-tuples
λ ={λ(1),λ(2), · · · ,λ(`)
}, µ =
{µ(1), µ(2), · · · , µ(k)
},
composed of the right ith tuples and the left jth tuples:
λ(i) =
{λ(i)1 },
{λ(i)2 , λ
(i)1 },
...
{λ(i)mi−1, · · · , λ
(i)2 , λ
(i)1 },
{λ(i)mi, λ
(i)mi−1, · · · , λ
(i)2 , λ
(i)1 },
, µ(j) =
{µ(j)1 },
{µ(j)2 , µ
(j)1 },
...
{µ(j)pj−1, · · · , µ
(j)2 , µ
(j)1 },
{µ(j)pj , µ
(j)pj−1, · · · , µ
(j)2 , µ
(j)1 },
where λ(i)j , µ
(j)j ∈ C and m1 + · · · + mk = k, p1 + · · · + pk = k.
Notice: the nestedness property of these tuples is reflected in the fact that each row in λ(i)
(µ(j)) is contained in the subsequent ones.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 29 / 44
To λ(i) we associate:
R(i) = Φ(λ(i)mi
) N · · · N Φ(λ(i)2 ) N Φ(λ
(i)1 ) B, i = 1, · · · , k.
The matrixR =
[R(1), R(2), · · · , R(k)
]∈ Cn×k ,
is the generalized controllability matrix of the bilinear system Σ, associated with the rightmutli-tuple λ.Similarly, to the left tuples we associate the matrices
O(j) = C Φ(µ(j)1 ) N Φ(µ
(j)2 ) N · · · N Φ(µ
(j)pj ), i = 1, · · · , k,
and the generalized observability matrix:
O =
O(1)
...
O(k)
∈ Ck×n.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 30 / 44
Interpolation of bilinear systems
Lemma. Let Σ = (C,E,A,N,B) be a scalar bilinear system of order n. Assume that it isprojected to a kth order systems by means of
V = R and WT = O:
E = WT EV, A = WT AV, N = WT NV, B = WT B, C = CV,
The system Σ = (C, E, A, N, B), of order k, satisfies the 2k + k2 interpolation conditions:
Hj (µ1, · · · , µj ) = Hj (µ1, · · · , µj ), Hi (λi , · · · , λ1) = Hi (λj , · · · , λ1),
Hj+i (µ1, · · · , µj , λi , · · · , λ1) = Hj+i (µ1, · · · , µj , λi , · · · , λ1), i , j = 1, · · · , k.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 31 / 44
The generalized Loewner pencil
We define:L = O ER, Lσ = OAR
In addition we define the quantities
Ψ = ONR, V = OB, W = CR
Proposition. The following relationships hold:
L(k, `) =
Hk+`−1(µ1, · · · , µk , λ`−1, · · · , λ1)−Hk+`−1(µ1, · · · , µk−1, λ`, · · · , λ1)µk − λ`
Lσ(k, `) =
µkHk+`−1(µ1, · · · , µk , λ`−1, · · · , λ1)− λiHk+`−1(µ1, · · · , µk−1, λ`, · · · , λ1)µk − λ`
while V(k, 1) = Hk (µ1, · · · , µk−1, µk ),
W(1, `) = H`(λ`, λ`−1, · · · , λ1),
Ψ(k, `) = Hk+`(µ1, · · · , µk−1, µk , λ`, λ`−1, · · · , λ1).
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 32 / 44
Example. Given the SISO bilinear system (C,E,A,N,B), where A is n × n, consider the
ordered tuples of left and right interpolation points:
[{µ1}{µ1, µ2}
], [{λ1}, {λ2, λ1}].
The associated generalized observability and controllability matrices are
O =
[C(µ1E− A)−1
C(µ1E− A)−1N(µ2E− A)−1
], R =
[(λ1E− A)−1B, (λ2E− A)−1N(λ1E− A)−1B
].
Then:
L =
[H1(µ1)−H1(λ1)
µ1−λ1
H2(µ1,λ1)−H2(λ2,λ1)µ1−λ2
H2(µ1,µ2)−H2(µ1,λ1)µ2−λ1
H3(µ1,µ2,λ1)−H3(µ1,λ2,λ1)µ2−λ2
]= −OER,
Lσ =
[µ1H1(µ1)−λ1H1(λ1)
µ1−λ1
µ1H2(µ1,λ1)−λ2H2(λ2,λ1)µ1−λ2
µ2H2(µ1,µ2)−λ1H2(µ1,λ1)µ2−λ1
µ2H3(µ1,µ2,λ1)−λ2H3(µ1,λ2,λ1)µ2−λ2
]= −OAR,
Ψ =
[H2(µ1, µ2) H3(µ1, λ2, λ1)
H3(µ1, µ2, λ1) H4(µ1, µ2, λ2, λ1)
]= ONR,
V =
[H1(µ1)
H2(µ1, µ2)
]= OB,
W =[
H1(λ1) H2(λ2, λ1)]
= CR.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 33 / 44
It readily follows that given the bilinear system (C, E, A, N, B) a reduced bilinear system oforder two, can be obtained without computation (matrix factorizations or solves) as:
E = OER, A = OAR, N = ONR, B = OB, C = CR.
• The following measurements are needed to construct the reduced second order system:
Linear Bilinear
H(µ1) H1(µ1)
H(λ1) H1(λ1)
H(µ2) H2(µ1, µ2)
H(λ2) H2(λ2, λ1)
H2(µ1, λ1)
H3(µ1, µ2, λ1)
H3(µ1, λ2, λ1)
H4(µ1, µ2, λ2, λ1)
2n moments n2 + 2n moments
matched matched
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 34 / 44
Construction of interpolants
Lemma. If k = `, and (Lσ , L), is a regular pencil, then
E = −L, A = −Lσ , N = Ψ, B = V, C = W,
is a minimal realization of an interpolant of the data, i.e., the rational functions:
Hk (s1, · · · ., sk ) = W(Lσ − s1L)−1Ψ · · ·Ψ(Lσ − skL)−1V, k ≥ 1,
interpolate the data.
Theorem. In the case of redundant data, the pencil (Lσ , L) is singular, and we constructX,Y ∈ Rρ×k as before. The quintuple (C,E,A,N,B) given by:
E = −Y∗LX, A = −Y∗LσX, N = Y∗ΨX, B = Y∗V, C = WX,
is the realization of an (approximate) interpolant of the data.
Remark. As in the linear case, if we have more data than necessary, we can either consider(W,−L,−Lσ , Ψ,V) as an exact but singular model of the data or
(WX,−Y∗LX,−Y∗LσX, Y∗ΨX, Y∗V),
as an approximate (nonsingular) model of the data.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 35 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 36 / 44
Numerical experiment: Bilinear controlled heat transfer system
Consider the heat equation∂x
∂t= ∆x on [0, 1]× [0, 1]
with the mixed Dirichlet and Robin boundary conditions:
n · ∇x = u1(x − 1), on Γ1 = {0}×c0, 1b,
n · ∇x = u2(x − 1), on Γ2 =c0, 1b×{0},
x = 0, on Γ3 = {1} × [0, 1] and Γ4 = [0, 1]× {1}.
The heat transfer coefficients u1 and u2 and the lower boundaries Γ1 and Γ2 are the inputvariables (spraying-intensities of a cooling- fluid acting on these boundaries). By a finitedifference discretization of the Poisson equation on an equidistant k × k mesh (with meshsizeh = 1
k+1) with nodes xij , we obtain the Poisson matrix:
P = I⊗ Tk + Tk ⊗ I, Tk = tridiag [1, −2, 1].
The dynamics of the heat flow yield a bilinear system described by
x = Ax + u1N1x + u2N2x + Bu,
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 37 / 44
where the system matrices are:
A =1
h2(I⊗ Tk + Tk ⊗ I + E1 ⊗ I + I⊗ Ek ), Ej = eje
Tj ,
N1 =1
hE1 ⊗ I, N2 =
1
hI⊗ Ek , b1 =
1
hE1 ⊗ e, b2 =
1
he⊗ Ek , e = [1, 1, · · · , 1]T ,
C =1
k2(e⊗ e)T .
ΣB is the 2500th order system obtained by discretiz-ing the heat equation.
1 ΣB is reduced using Loewner to obtain Σ1 oforder 28.
2 ΣB is reduced using BIRKA to obtain Σ2 oforder 28.
Singular values ofthe Loewner pencil
The first step is to collect samples from generalized bilinear transfer functions up to order 2, andplot the singular values of the ensuing Loewner pencil. We notice that σ1 = 1, σ28 ≈ 10−15
⇒ k = 28. Next, we compare the time-domain output of Σ0, and the outputs the reducedsystems when the input signals are:
u1(t) = sin(4t)exp(−t/2) + (0.5)exp(−t/2), u2(t) =1
2cos(πt + 1).
The tangential directions are chosen randomly.
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 38 / 44
Time domain simulation - Outputs and Approximation errors
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 39 / 44
Outline
1 The Loewner frameworkIntroduction
Construction of interpolants and generalized inverses
2 Data-driven model reduction: The example of a clamped beam
3 The Loewner framework for bilinear systemsIntroduction
A numerical experiment
4 Challenges and Concluding remarks
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 40 / 44
Challenge: Earthquake prevention for high-rise buildings
Taipei 101 damper
Taipei 101: 508m Damper between 87-91 floors 730 ton damper
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 41 / 44
Challenge: Damping of lightweight bridges
Lamot footbridge damper optimization
Lamot bridge finite elementmodel (n = 25,962)The goal is to determine theoptimal stiffness and dampingcoefficient of four bridgedampers (=8 parameters).
K. Meerbergen (KU Leuven) MODRED 2013 December 11–13, 2013 6 / 53
The Lamot footbridge Parametrized freq. response
• Lamot lightweight bridge in Mechelen: finite element model n = 25, 962.
• The goal: determine optimal stiffness and damping coefficients of the
four bridge dampers (8 parameters).
• Example due to Karl Meerbergen (KU Leuven).
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 42 / 44
Conclusions
Given data (`i , vi ), (λj ,wj ), construct the Loewner pencil (Lσ ,L).
The quadruple (W,L,Lσ ,V), where the pencil (Lσ ,L) may be singular,
is a natural model of the data. The construction involves no computation.
(Lσ ,L) and the underlying (A,E) have the same non-trivial eigenvalues.
The projection to a minimal realization can be chosen arbitrarily.
New: The Moore-Penrose or Drazin inverses of Φ(s) = Lσ − s L, satisfy:
H(s) = W Φ(s)# V
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 43 / 44
References
Overview article:
A.C. Antoulas, S. Lefteriu and A.C. Ionita,
A tutorial introduction to the Loewner framework for model reduction,
in Model Reduction and Approximation for Complex Systems, edited by P. Benner,
A. Cohen, M. Ohlberger, and K. Willcox, Birkhauser Verlag, ISNM Series (2015).
Just accepted:
A.C. Antoulas,
The Loewner Framework and Transfer Functions of Singular and Rectangular Systems,
Applied Mathematics Letters (2016).
Thanos Antoulas (Rice U. & Jacobs U.) The Loewner Framework 44 / 44