1. 1 Introduction to Model Order Reduction Luca Daniel Massachusetts Institute of Technology luca@mit.edu http://onigo.mit.edu/~dluca/2006PisaMOR www.rle.mit.edu/cpg

2. 2 Model Order Reduction of Linear Systems via Modal Analysis via Rational function fitting (point matching) via Quasi Convex Optimization via Pade approximation (AWE) Projection Framework SVD, PCA, LVD, POD Krylov Subspace Moment Matching Projection Methods  Arnoldi  PVL  PRIMA Truncated Balance Realization (TBR)  Positive Real TBR Distributed Systems (with Frequency Dependent Matrices)

3. Distributed Linear Systems Examples:  ODE’s with delays (e.g. full-wave integral equation solvers)  frequency-dependent basis functions  frequency dependent discretizations  solvers using layered-media Green functions (e.g. for handling substrate or dielectrics) NOTE: Distributed systems may have infinite order (e.g. delay)!!

4. Polynomial interpolation [Phillips96] Polynomial approximation e.g. Taylor expansion, or a polynomial interpolation for A(s) Convert to non-distributed model reduction problem Performance: Fast and accurate in the frequency band of interest Problem: Can not be used in a time domain circuit simulator because does not guarantee stability and passivity

5. Passivity condition on transfer function For systems with immittance matrix representation, passivity is equivalent to positive-realness of the transfer function (no unstable poles) (no negative resistors) (impulse response is real) It means its real part is a positive for any frequency. Note: it is a global property!!!! FOR ANY FREQUENCY

6. 6 Positive real transfer function in the complex plane for different frequencies Passive region Active region original system

7. 7 Why does polynomial interpolation fail when applied to the Laplace parameter ‘s’? original system Passive region Active region Although accurate in the frequency band of interest Polynomial interpolation is unlikely to preserve GLOBAL properties such as positive realness because it is GLOBALLY not well-behaved

8. Observation: practical systems have some loss at any frequency Most systems are non-ideal i.e. contain some small loss at any frequency i.e. can be described by strictly positive real functions original system Passive region Active region

9. Using global uniformly convergent interpolants If A(s) is strictly positive real, a GLOBALLY and UNIFORMLY convergent interpolant will eventually get close enough (for a large enough order M of the interpolant) and be positive-real as well. original system reduced system Passive region Active region Proof: just choose accuracy of interpolation smaller than minimum distance from imaginary axis

10. A good example of uniformly convergent interpolants: the Laguerre basis functions Consider the family of basis functions: They form a complete, rational, orthonormal basis over the imaginary axis which gives a uniformly convergent interpolant no poles in RHP (stable) (real time-domain representation)

11. 11 Calculation of interpolation coefficients Note: it is a bilinear transform that maps the Laguerre basis to Fourier series on the unit circle. Re{s} Im{s} Re{z} Im{z} Hence in practice one can use FFT to calculate the interpolation coefficients: very efficient! Note: FFT coefficients typically drop quickly and the series can be truncated to the first few M coefficients because field solver matrices A(s) are often smooth.

12. 12 An implementation example: Two wires on a MCM package [D. DAC02] Discretize Maxwell equations in integral form using PEEC NOTE: system matrices are frequency dependent because the substrate is handled by layered Green functions package s x=- A (s) x+b u

13. Step 2: Interpolation (example) FFT coefficients of A (s) 0102030405060 10 -4 10 -5 10 -6 10 -7 10 -8

14. Step 2: Interpolation (example) reconstructed from first 5 out of 64 FFT coefficients and compared to original A (s) 0102030405060 5 4 3 2 1 0 nH Real part Imaginary part

15. Reduction procedure [D. and Phillips DAC01] Matrix sizes System order Start from original system described by causal, strictly positive-real matrices ~ 3,000infinite 1) Evaluate and squash them at uniformly spaced points on the unit circle using e.g. POD with congruence transformation which preserves positive realness U T A(z k )U, k=1,2,...,64 b r =U T b ~ 6~ 6 x 64

16. Reduction procedure [D. and Phillips DAC01] Matrix sizes System order 3) Calculate first few (e.g 5) FFT coef of the reduced system matrix 6 6 x 5 4) Introduce extended state and realize a single matrix discrete time system 6 x 5 5) Transform to continuous time 6 x 5

17. Step 3: Realization (Multichip Module MCM example) Real part of frequency response l Inductive part of frequency response 5 4 3 2 nH frequencyfrequency 5 4 3 2 x10 4 Ohm 10 6 10 7 10 8 10 9 10 10 6 10 7 10 8 10 910 original system 3000 distrib reduced system 30 original system reduced system

18. Open issues for distributed systems Guaranteeing positive realness relies on accuracy of the uniform interpolant. Hence if the matrices are NOT smooth, we might need a large order of the interpolant.  working on internal matrices might give smoother matrices Laguerre basis functions are efficient since they use FFT. However equally spaced points on the unit circle correspond to non-equally spaced points on the imaginary axis accumulating around a reference center frequency.

19. Model Order Reduction of Linear Systems via Modal Analysis via Rational function fitting (point matching) via Quasi Convex Optimization via Pade approximation (AWE) Projection Framework SVD, PCA, LVD, POD Krylov Subspace Moment Matching Projection Methods  Arnoldi  PVL  PRIMA Truncated Balance Realization (TBR)  Positive Real TBR Laguerre interpolation for Distributed Systems

20. Model Order Reduction of Linear Systems via Modal Analysis via Rational function fitting (point matching) via Quasi Convex Optimization (use this for LARGE and passive systems, or for model construction from measurements, or for distributed systems) via Pade approximation (AWE) Projection Framework SVD, PCA, LVD, POD Krylov Subspace Moment Matching Projection Methods  Arnoldi  PVL (use this for HUGE systems if passivity is not an issue)  PRIMA (use this for HUGE and passive systems) Truncated Balance Realization (TBR) (use this for SMALL systems)  Positive Real TBR (use this for SMALL and passive systems) Laguerre interpolation for Distributed Systems (use this for LARGE systems with frequency dependent matrices, e.g. delays)