Matrix RationalH2 Approximation:
a State-Space Approach using Schur Parameters
J.P. Marmorat, M. Olivi, B. Hanzon, R. Peeters
Sophia–Antipolis, France
1
Introduction
We present a method to compute astablerationalL2-approximation of
specified ordern to a givenmultivariabletransfer function.
• it works formultivariablesystems,
• it uses a niceparametrization of stable allpass systems, which
– takes into account thestability constraint
– ensures identifiability
– is well-conditionned
• it uses arecursive search on the degreewhich improves the chances
to reach the global minimum.
2
TheL2-criterion in state-space form
• F (z) = C(zIN −A)−1B +D, m× p given transfer function
• H(z) = γ(zIn −A)−1B +D, approximant at ordern
(A,B) input-normal pair:AA∗ + BB∗ = I.
L2-norm of the errorF −H:
J(A,B) = ‖F‖2 − Tr (γγ∗) ,
where
γ = CW,
andW solution to the Lyapunov equation:
AWA∗ + BB∗ = W.
3
Optimization set: stable-allpass systems
The two following sets (equivalence classes) are diffeomorphic:
• input-normal pairs(A,B)
• stable allpass functionsG(z) = D + C(zIn −A)−1B
up to a left constant unitary factor D C
B A
unitary matrix.
Alpay, Baratchart, Gombani[1];
4
Parametrization issue
Desirable properties:
• ensuresidentifiability
• a small perturbation of the parameterspreserves the stability and the
orderof the system?
• allows for the use ofdifferential tools.
→ Differentiable manifold.
Atlas of charts or overlapping canonical forms :
a collection oflocal parametrizationswith compatibility conditions
(changes of charts are smooth).
5
Atlases of charts
Two families can be found in the litterature
1. from a tangential Schur algorithm:
Gn(1/w̄)u = v, ‖v‖ < 1, GnLFT=⇒ Gn−1
Gn, . . . , Gk(wk,uk,vk)
=⇒ Gk−1, . . . , G0.
Alpay, Baratchart, Gombani[1]; Fulcheri, Olivi [2]
2. from state-space representations
Hanzon, Ober[3]
A parametrization that combines the two approaches:
Peeters, Hanzon, Olivi[4]
6
Encoding stable-allpass systems
A p× p stable allpass systems of degreen is encoded by
• w1, w2, . . . , wn points of the unit circle,
• u1, u2, . . . , un unit complexp-vectors,
• v1, v2, . . . , vn complexp-vectors,‖vi‖ < 1.
Thewi’s and theui’s definethe chartwhile thevi’s arethe Schur
parametersof the system in the chart.In a given chart, a system is
perfectly determined by its Schur parameters (identifiability).
7
Encoding stable-allpass systems (2)
The stable allpass system encoded in that way hasunitary realization
matrix Dn Cn
Bn An
computed by induction
Dk Ck
Bk Ak
=
Vk 0
0 Ik−1
1 0 0
0 Dk−1 Ck−1
0 Bk−1 Ak−1
U∗
k 0
0 Ik−1
,
whereAk is k × k, Dk is p× p, andD0 = Ip.
→ very nice numerical behavior
8
Encoding stable-allpass systems (3)
Uk andVk are the(p + 1)× (p + 1) unitary matrices:
Uk =
ξkuk Ip − (1 + ηkwk)uku∗k
ηkwk ξku∗k
Vk =
ξkvk Ip − (1− ηk) vkv∗k‖vk‖2
ηk − ξkv∗k
ξk =
√1− |wk|2√
1− |wk|2‖vk‖2, ηk =
√1− ‖vk‖2√
1− |wk|2‖vk‖2
9
Main steps of the algorithm
• finding an adapted chart:
realization in Schur form (A lower triangular)
A =
wn 0 · · · 0...
......
...
∗ · · · w1 0
, B =
√1− |wn|2 u∗n
...
→ v1 = v2 = . . . vn = 0 (Potapov factorization)
• optimization over the manifold
• recursive search on the degree (optional):
minimum of degreek → initial point of degreek + 1 (error
preserved)
10
The RARL2 software
This software computes astablerationalL2-approximation of specified
ordern to amultivariabletransfer function given in one of the following
forms:
• a realization
• a finite number ofFourier coefficients
• somepointwise valueson the unit circle.
It has been implemented using standard MATLAB subroutines. The
optimizer of the toolkit OPTIM is used to find a local minimum, given by
a realization, of the nonlinearL2-criterion.
http://www-sop.inria.fr/miaou/Martine.Olivi/me.html
11
Automobile gas turbine
Hung, MacFarlane[6]; Glover[5]; Yan, Lam[7]
12
Nyquist diagrams
2× 2; order12.
13
Approximants: order 1
14
Approximants: order 2
15
Approximants: order 3
16
Approximants: order 4
17
Approximants: order 5
18
Approximants: order 6
19
Approximants: order 7
20
Approximants: order 8
21
Hyperfrequency Filter
The problem:find a8th ordermodel of aMIMO (2× 2) hyperfrequency
filter, from experimental pointwise values in some range of frequencies
provided by the CNES (French space agency).
First stage (interpolation/completion):compute astable matrix transfer
function of high orderwhich approximates these data, given by a great
number(800)of Fourier coefficients.
PRESTO-HF: software byF. Seyfert;
HYPERION: software byJ. Grimm.
22
Data and approximant at order 8
Nyquist diagrams
23
Data and approximant at order 8
Bode diagrams
24
References
[1] D. ALPAY, L. BARATCHART, AND A. GOMBANI , On the differentialstructure of matrix-valued rational inner functions, Operator Theory :Advances and Applications, 73 (1994), pp. 30–66.
[2] P. FULCHERI AND M. OLIVI , Matrix RationalH2-Approximation: aGradient Algorithm Based on Schur Analysis, SIAM J. Contr. and Opt.,36 (1998), pp. 2103–2127.
[3] B. HANZON AND R.J. OBER, Overlapping block-balanced canonicalforms for various classes of linear systems, Linear Algebra and its Ap-plications, 281 (1998), pp. 171-225.
[4] R.L.M. PEETERS, B. HANZON AND M. OLIVI , Balanced realizationsof discrete-time stable all-pass systems and the tangential Schur algo-rithm, in Proceedings of the ECC99, Karlsruhe, Germany, August 31-September 3.
24-1
[5] K. GLOVER, All optimal Hankel-norm approximations of linear multi-variable systems and theirL∞-error bounds, Int. J. Control 39 (1984),pp. 1115-1193.
[6] Y.S. HUNG AND A.G.J. MACFARLANE, Multivariable feedback: aquasi-classical approach, Lect. Notes in Control Sci. 40, Springer- Verlag,1982.
[7] W.-Y. YAN AND J. LAM , An approximate approach toH2 optimal modelreduction, IEEE Trans. Autom. Control 44, No.7 (1999), pp 1341-1358.
24-2