Skip to main content
SpringerLink
Log in

Interpolatory model reduction of quadratic-bilinear dynamical systems with quadratic-bilinear outputs

  • Published:
Advances in Computational Mathematics Aims and scope Submit manuscript

Abstract

This paper extends interpolatory model reduction to systems with (up to) quadratic-bilinear dynamics and quadratic-bilinear outputs. These systems are referred to as QB-QB systems and arise in a number of applications, including fluid dynamics, optimal control, and uncertainty quantification. In the interpolatory approach, the reduced order models (ROMs) are based on a Petrov-Galerkin projection, and the projection matrices are constructed so that transfer function components of the ROM interpolate the corresponding transfer function components of the original system. To extend the approach to systems with QB outputs, this paper derives system transfer functions and sufficient conditions on the projection matrices that guarantee the aforementioned interpolation properties. Alternatively, if the system has linear dynamics and quadratic outputs, one can introduce auxiliary state variables to transform it into a system with QB dynamics and linear outputs to which known interpolatory model reduction can be applied. This transformation approach is compared with the proposed extension that directly treats quadratic outputs. The comparison shows that transformation hides the problem structure. Numerical examples illustrate that keeping the original QB-QB structure leads to ROMs with better approximation properties.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Log in via an institution

Price excludes VAT (USA)
Tax calculation will be finalised during checkout.

Instant access to the full article PDF.

Institutional subscriptions

Similar content being viewed by others

References

  1. Asif, M.M.A., Ahmad, M.I., Benner, P., Feng, L., Stykel, T.: Implicit higher-order moment matching technique for model reduction of quadratic-bilinear systems. J. Franklin Inst. 358(3), 2015–2038 (2021). https://doi.org/10.1016/j.jfranklin.2020.11.012

    Article  MathSciNet  Google Scholar 

  2. Benner, P., Breiten, T.: Two-sided projection methods for nonlinear model order reduction. SIAM J. Sci. Comput. 37(2), 239–260 (2015). https://doi.org/10.1137/14097255X

    Article  MathSciNet  Google Scholar 

  3. Gosea, I.V., Antoulas, A.C.: Data-driven model order reduction of quadratic-bilinear systems. Numer. Linear Algebra Appl. 25(6), 2200 (2018). https://doi.org/10.1002/nla.2200

    Article  MathSciNet  Google Scholar 

  4. Antoulas, A.C., Beattie, C.A., Gugercin, S.: Interpolatory model reduction. Comput. Sci. Eng., vol. 21. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA (2020). https://doi.org/10.1137/1.9781611976083

  5. Ahmad, M.I., Benner, P., Goyal, P., Heiland, J.: Moment-matching based model reduction for Navier-Stokes type quadratic-bilinear descriptor systems. ZAMM Z. Angew. Math. Mech. 97(10), 1252–1267 (2017). https://doi.org/10.1002/zamm.201500262

    Article  MathSciNet  Google Scholar 

  6. Benner, P., Goyal, P.: An iterative model reduction scheme for quadratic-bilinear descriptor systems with an application to Navier-Stokes equations. In: Keiper, W., Milde, A., Volkwein, S. (eds.) Reduced-order modeling (ROM) for simulation and optimization, pp. 1–19. Springer Verlag, ??? (2018). https://doi.org/10.1007/978-3-319-75319-5_1

  7. Diaz, A.N., Heinkenschloss, M.: Towards data-driven model reduction of the Navier-Stokes equations using the Loewner framework. In: King, R., Peitsch, D. (eds.) Active flow and combustion control 2021, pp. 225–239. Springer, Cham (2022). https://doi.org/10.1007/978-3-030-90727-3_14

  8. Antil, H., Heinkenschloss, M., Hoppe, R.H.W., Sorensen, D.C.: Domain decomposition and model reduction for the numerical solution of PDE constrained optimization problems with localized optimization variables. Comput. Vis. Sci. 13, 249–264 (2010). https://doi.org/10.1007/s00791-010-0142-4

    Article  MathSciNet  Google Scholar 

  9. Pulch, R., Narayan, A.: Balanced truncation for model order reduction of linear dynamical systems with quadratic outputs. SIAM J. Sci. Comput. 41(4), 2270–2295 (2019). https://doi.org/10.1137/17M1148797

    Article  MathSciNet  Google Scholar 

  10. Van Beeumen, R., Meerbergen, K.: Model reduction by balanced truncation of linear systems with a quadratic output. In: International Conference on Numerical Analysis and Applied Mathematics (ICNAAM) 2010, AIP Conference Proceedings, vol. 1281, pp. 2033–2036. American Institute of Physics, ??? (2010). https://doi.org/10.1063/1.3498345

  11. Benner, P., Goyal, P., Duff, I.P.: Gramians, energy functionals, and balanced truncation for linear dynamical systems with quadratic outputs. IEEE Trans. Autom. Control 67(2), 886–893 (2022). https://doi.org/10.1109/TAC.2021.3086319

    Article  MathSciNet  Google Scholar 

  12. Benner, P., Goyal, P.: Balanced truncation model order reduction for quadratic-bilinear control systems (2017). arXiv:1705.00160v1https://doi.org/10.48550/arXiv.1705.00160

  13. Van Beeumen, R., Van Nimmen, K., Lombaert, G., Meerbergen, K.: Model reduction for dynamical systems with quadratic output. Internat. J. Numer. Methods Engrg. 91(3), 229–248 (2012). https://doi.org/10.1002/nme.4255

    Article  Google Scholar 

  14. Gosea, I.V., Antoulas, A.C.: A two-sided iterative framework for model reduction of linear systems with quadratic output. In: 2019 IEEE 58th Conference on Decision and Control (CDC), pp. 7812–7817 (2019). https://doi.org/10.1109/CDC40024.2019.9030025

  15. Gosea, I.V., Gugercin, S.: Data-driven modeling of linear dynamical systems with quadratic output in the AAA framework. J. Sci. Comput. 91(1), 16–28 (2022). https://doi.org/10.1007/s10915-022-01771-5

    Article  MathSciNet  Google Scholar 

  16. Holicki, T., Nicodemus, J., Schwerdtner, P., Unger, B.: Energy matching in reduced passive and port-Hamiltonian systems (2023). arXiv:2309.05778v1https://doi.org/10.48550/arXiv.2309.05778

  17. Benner, P., Feng, L.: Model order reduction based on moment-matching. In: Benner, P., Grivet-Talocia, S., Quarteroni, A., Rozza, G., Schilders, W., Silveira, L.M. (eds.) Model Order Reduction. xVolume 1: System- and Data-Driven Methods and Algorithms, pp. 57–96. Walter de Gruyter & Co., Berlin, ??? (2021). https://doi.org/10.1515/9783110498967-003

  18. Rugh, W.J.: Nonlinear system theory. The Volterra/Wiener approach. Johns Hopkins series in information sciences and systems. Johns Hopkins University Press, Baltimore, Md., ??? (1981). PDF version with corrections (2002). https://sites.google.com/site/wilsonjrugh Accessed 13-Feb.-2018

  19. Brockett, R.W.: Volterra series and geometric control theory. Automatica 12(2), 167–176 (1976). https://doi.org/10.1016/0005-1098(76)90080-7

    Article  MathSciNet  Google Scholar 

  20. Bullo, F.: Series expansions for analytic systems linear in control. Automatica J. IFAC 38(8), 1425–1432 (2002). https://doi.org/10.1016/S0005-1098(02)00042-0

    Article  MathSciNet  Google Scholar 

  21. Hélie, T., Laroche, B.: Computation of convergence bounds for Volterra series of linear-analytic single-input systems. IEEE Trans. Automat. Control 56(9), 2062–2072 (2011). https://doi.org/10.1109/TAC.2010.2091435

    Article  MathSciNet  Google Scholar 

  22. Gugercin, S., Antoulas, A.C., Beattie, C.: \({\cal{H}}_2\) model reduction for large-scale linear dynamical systems. SIAM J. Matrix Anal. Appl. 30(2), 609–638 (2008). https://doi.org/10.1137/060666123

    Article  MathSciNet  Google Scholar 

  23. Saak, J., Köhler, M., Benner, P.: M-M.E.S.S.-2.2 – The matrix equations sparse solvers library (2022). see also:https://www.mpi-magdeburg.mpg.de/projects/messhttps://doi.org/10.5281/zenodo.5938237

  24. Benner, P., Goyal, P., Gugercin, S.: \({\cal{H} }_2\)-Quasi-optimal model order reduction for quadratic-bilinear control systems. SIAM J. Matrix Anal. Appl. 39(2), 983–1032 (2018). https://doi.org/10.1137/16M1098280

    Article  MathSciNet  Google Scholar 

Download references

Acknowledgements

We thank the reviewers for their comments, which have led to improvements in the presentation.

Funding

This research was supported in part by AFOSR Grant FA9550-22-1-0004, NSF grants CCF-1816219 and DMS-1819144, and by a 2021 National Defense Science and Engineering Graduate (NDSEG) Fellowship Award.

Author information

Authors and Affiliations

  1. Department of Computational Applied Mathematics and Operations Research, Rice University, 6100 Main St.-MS 134, Houston, 77005, TX, USA

    Alejandro N. Diaz, Matthias Heinkenschloss & Ion Victor Gosea

  2. Max Planck Institute for Dynamics of Complex Technical Systems Magdeburg, Sandtorstrasse 1, Magdeburg, 39106, Germany

    Ion Victor Gosea & Athanasios C. Antoulas

  3. Department of Electrical an Computer Engineering, Rice University, 6100 Main St.-MS 380, Houston, 77005, TX, USA

    Athanasios C. Antoulas

Authors
  1. Alejandro N. Diaz
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Matthias Heinkenschloss
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Ion Victor Gosea
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Athanasios C. Antoulas
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to Alejandro N. Diaz.

Ethics declarations

Conflict of interest

The authors declare no competing interests.

Additional information

Communicated by: Tobias Breiten

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Appendices

Appendix

A   Proof of Theorem 1

We prove Theorem 1. First observe that

$$\begin{aligned} \widehat{\varvec{\Phi }}(s)^{-1} = s\widehat{\textbf{E}}-\widehat{\textbf{A}}= \textbf{W}^*(s\textbf{E}-\textbf{A})\textbf{V}= \textbf{W}^*{\varvec{\Phi }}(s)^{-1}\textbf{V}. \end{aligned}$$

By Eq. 13a, there exist vectors \(\textbf{v}_{j, 1} \in \mathbb {C}^r\) such that

$$\begin{aligned} \textbf{V}\textbf{v}_{j, 1} = {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j, \qquad j=1, \dots , \ell . \end{aligned}$$
(1)

Multiplying Eq. 44 by \(\textbf{W}^*{\varvec{\Phi }}(\sigma _j)^{-1}\) yields \(\textbf{W}^*{\varvec{\Phi }}(\sigma _j)^{-1}\textbf{V}\textbf{v}_{j, 1} =\widehat{\varvec{\Phi }}(\sigma _j)^{-1}\textbf{v}_{j, 1} = \textbf{W}^*\textbf{B}\textbf{b}_j = \widehat{\textbf{B}}\textbf{b}_j\). Hence

$$\begin{aligned} \textbf{v}_{j, 1}=\widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j. \end{aligned}$$
(2a)

By equation Eq. 45 and the definition Eq. 44 of \(\textbf{v}_{j,1}\),

$$\begin{aligned} \widehat{\textbf{H}}_1(\sigma _j)\textbf{b}_j&= \widehat{\textbf{C}}\widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j = \textbf{C}\textbf{V}\textbf{v}_{j, 1} =\textbf{C}{\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j =\textbf{H}_1(\sigma _j)\textbf{b}_j, \end{aligned}$$

which is 14a.

By Eq. 13b, there exist vectors \(\textbf{v}_{j, 2}\in \mathbb {C}^r\) such that

$$\begin{aligned} \textbf{V}\textbf{v}_{j, 2} = {\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j)+\textbf{N}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes \textbf{b}_j) \right] , \end{aligned}$$
(2b)

\(j=1, \dots , \ell \).

Next, notice that by the definitions Eqs. 44, 46 of \(\textbf{v}_{j, 1}\) and \(\textbf{v}_{j, 2}\) and Eq. 45,

$$\begin{aligned}&\textbf{W}^*{\varvec{\Phi }}(2\sigma _j)^{-1}\textbf{V}\textbf{v}_{j, 2} =\widehat{\varvec{\Phi }}(2\sigma _j)^{-1}\textbf{v}_{j, 2} \\&=\textbf{W}^*{\varvec{\Phi }}(2\sigma _j)^{-1}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j)+\textbf{N}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes \textbf{b}_j) \right] \\&=\textbf{W}^*\textbf{Q}(\textbf{V}\otimes \textbf{V})(\textbf{v}_{j, 1}\otimes \textbf{v}_{j, 1})+ \textbf{W}^*\textbf{N}(\textbf{V}\otimes \textbf{I}_m) (\textbf{v}_{j, 1}\otimes \textbf{b}_j) \\&= \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\otimes \textbf{b}_j\right) , \qquad j=1, \dots , \ell . \end{aligned}$$

Hence

$$\begin{aligned} \textbf{v}_{j, 2} = \widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j \otimes \textbf{b}_j\right) \right] , \end{aligned}$$
(2c)

\(j=1, \dots , \ell \).

Using Eqs. 47 and 45 and then the definitions Eqs. 44, 46 of \(\textbf{v}_{j, 1}\) and \(\textbf{v}_{j, 2}\) it follows that

$$\begin{aligned} \widehat{\textbf{H}}_{2}(\sigma _j, \sigma _j)(\textbf{b}_j\otimes \textbf{b}_j)&=\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j \otimes \textbf{b}_j\right) \right] \\&\qquad + \widehat{\textbf{K}}(\widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j) + \widehat{\textbf{J}}(\widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\otimes \textbf{b}_j)\\&= \textbf{C}\textbf{V}\textbf{v}_{j, 2} + \textbf{K}(\textbf{V}\otimes \textbf{V})(\textbf{v}_{j, 1}\otimes \textbf{v}_{j, 1}) + \textbf{J}(\textbf{V}\otimes \textbf{I}_m)(\textbf{v}_{j, 1}\otimes \textbf{b}_j)\\&= \textbf{C}\textbf{V}\textbf{v}_{j, 2} + \textbf{K}(\textbf{V}\textbf{v}_{j, 1} \otimes \textbf{V}\textbf{v}_{j, 1}) + \textbf{J}(\textbf{V}\textbf{v}_{j, 1}\otimes \textbf{b}_j)\\&= \textbf{C}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j)+\textbf{N}\left( {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes \textbf{b}_j\right) \right] \\&\qquad + \textbf{K}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j) + \textbf{J}({\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j\otimes \textbf{b}_j)\\&= \textbf{H}_{2}(\sigma _j, \sigma _j)(\textbf{b}_j\otimes \textbf{b}_j), \qquad j=1, \dots , \ell , \end{aligned}$$

which is Eq. 14b.

By Eq. 13c, there exist vectors \(\textbf{w}_{j, 1} \in \mathbb {C}^r\) such that

$$\begin{aligned} \textbf{w}_{j, 1}^*\textbf{W}^* = \textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j), \qquad j=1, \dots , \ell . \end{aligned}$$
(3a)

Multiplying Eq. 48 by \({\varvec{\Phi }}(2\sigma _j)^{-1}\textbf{V}\) yields \(\textbf{w}_{j, 1}^*\textbf{W}^*{\varvec{\Phi }}(2\sigma _j)^{-1}\textbf{V}= \textbf{w}_{j, 1}^*\widehat{\varvec{\Phi }}(2\sigma _j)^{-1} = \textbf{c}_j^*\textbf{C}\textbf{V}= \textbf{c}_j^*\widehat{\textbf{C}}\). Hence,

$$\begin{aligned} \textbf{w}_{j, 1}^* = \textbf{c}_j^*\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j), \qquad j=1, \dots , \ell . \end{aligned}$$
(3b)

Using Eqs. 49 and 48 gives

$$\begin{aligned} \textbf{c}_j^*\widehat{\textbf{H}}_1(2\sigma _j)&= \textbf{c}_j^*\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j) \widehat{\textbf{B}}= \textbf{w}_{j, 1}^* \textbf{W}^* \textbf{B}= \textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\textbf{B}= \textbf{c}_j^*\textbf{H}_1(2\sigma _j), \qquad j=1, \dots , \ell , \end{aligned}$$

which is Eq. 14c.

By Eq. 13d, there exist vectors \(\textbf{w}_{j, 2} \in \mathbb {C}^r\) such that

$$\begin{aligned} \textbf{w}_{j, 2}^*\textbf{W}^*&= \textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\Big [\textbf{Q}({\varvec{\Phi }}(\sigma _j)\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j ) + \textbf{N}({\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j\Big ] \nonumber \\&\quad + \textbf{c}_j^*\textbf{K}({\varvec{\Phi }}(\sigma _j) \otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j) + \textbf{c}_j^*\textbf{J}({\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j). \end{aligned}$$
(3c)

In the next equalities we use the identity \(\textbf{M}= \textbf{M}\otimes 1\) for any matrix \(\textbf{M}\) to apply the Kronecker product property Eq. 1. Using Eq. 50, the identity \(\textbf{M}= \textbf{M}\otimes 1\), Eqs. 44, 48, 49, 45 gives

$$\begin{aligned} \textbf{w}_{j, 2}^*\widehat{\varvec{\Phi }}(\sigma _j)^{-1}&= \textbf{w}_{j, 2}^*\textbf{W}^* {\varvec{\Phi }}(\sigma _j)^{-1}\textbf{V}\\&=\Big (\textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j) \otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j) + \textbf{N}({\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j\right] \\&\quad + \textbf{c}_j^*\textbf{K}({\varvec{\Phi }}(\sigma _j) \otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j) + \textbf{c}_j^*\textbf{J}({\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j)\Big )\left( {\varvec{\Phi }}(\sigma _j)^{-1}\textbf{V}\otimes 1\right) \\&= \textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}(\textbf{V}\otimes \textbf{V}\textbf{v}_{j,1}) + \textbf{N}(\textbf{V}\otimes \textbf{b}_j)\right] + \textbf{c}_j^*\textbf{K}(\textbf{V}\otimes \textbf{V}\textbf{v}_{j,1}) + \textbf{c}_j^*\textbf{J}(\textbf{V}\otimes \textbf{b}_j) \\&= \textbf{w}_{j, 1}^*\textbf{W}^*\left[ \textbf{Q}(\textbf{V}\otimes \textbf{V})(\textbf{I}_r \otimes \textbf{v}_{j,1}) + \textbf{N}(\textbf{V}\otimes \textbf{I}_m)(\textbf{I}_r \otimes \textbf{b}_j)\right] \\&\quad + \textbf{c}_j^*\textbf{K}(\textbf{V}\otimes \textbf{V})(\textbf{I}_r\otimes \textbf{v}_{j, 1}) + \textbf{c}_j^*\textbf{J}(\textbf{V}\otimes \textbf{I}_m)(\textbf{I}_r \otimes \textbf{b}_j) \\&= \textbf{w}_{j, 1}^*\left[ \textbf{W}^*\textbf{Q}(\textbf{V}\otimes \textbf{V})(\textbf{I}_r \otimes \textbf{v}_{j,1}) + \textbf{W}^*\textbf{N}(\textbf{V}\otimes \textbf{I}_m)(\textbf{I}_r \otimes \textbf{b}_j)\right] \\&\quad + \textbf{c}_j^*\textbf{K}(\textbf{V}\otimes \textbf{V})(\textbf{I}_r \otimes \textbf{v}_{j,1}) + \textbf{c}_j^*\textbf{J}(\textbf{V}\otimes \textbf{I}_r)(\textbf{I}_r \otimes \textbf{b}_j) \\&= \textbf{c}_j^*\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \textbf{I}_r \otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}(\textbf{I}_r \otimes \textbf{b}_j)\right] \\&\quad + \textbf{c}_j^*\widehat{\textbf{K}}\left( \textbf{I}_r \otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \textbf{c}_j^*\widehat{\textbf{J}}(\textbf{I}_r \otimes \textbf{b}_j). \end{aligned}$$

Multiplying on the right by \(\widehat{\varvec{\Phi }}(\sigma _j)=\widehat{\varvec{\Phi }}(\sigma _j)\otimes 1\), we conclude that

$$\begin{aligned} \textbf{w}_{j, 2}^*&=\textbf{c}_j^*\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j\right) \right] \\&\quad + \textbf{c}_j^*\widehat{\textbf{K}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \textbf{c}_j^*\widehat{\textbf{J}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j\right) . \nonumber \end{aligned}$$
(3d)

Using Eqs. 51 and 50 gives

$$\begin{aligned} \textbf{c}_j^*\widehat{\textbf{H}}_2(\sigma _j, \sigma _j)(\textbf{I}_m \otimes \textbf{b}_j)&=\textbf{c}_j^*\Big [\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \textbf{I}_m\right) \right] \\&\qquad \quad + \widehat{\textbf{K}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\right) + \widehat{\textbf{J}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \textbf{I}_m\right) \Big ](\textbf{I}_m \otimes \textbf{b}_j)\\&=\textbf{c}_j^*\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \textbf{b}_j\right) \right] \\&\qquad \quad + \textbf{c}_j^*\widehat{\textbf{K}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \textbf{c}_j^*\widehat{\textbf{J}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\otimes \textbf{b}_j\right) \\&=\Big [\textbf{c}_j^*\widehat{\textbf{C}}\widehat{\varvec{\Phi }}(2\sigma _j)\left[ \widehat{\textbf{Q}}\left( \widehat{\varvec{\Phi }}(\sigma _j) \otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \widehat{\textbf{N}}\left( \widehat{\varvec{\Phi }}(\sigma _j) \otimes \textbf{b}_j\right) \right] \\&\qquad \quad + \textbf{c}_j^*\widehat{\textbf{K}}\left( \widehat{\varvec{\Phi }}(\sigma _j)\otimes \widehat{\varvec{\Phi }}(\sigma _j)\widehat{\textbf{B}}\textbf{b}_j\right) + \textbf{c}_j^*\widehat{\textbf{J}}\left( \widehat{\varvec{\Phi }}(\sigma _j) \otimes \textbf{b}_j\right) \Big ]\left( \widehat{\textbf{B}}\otimes 1\right) \\&= \textbf{w}_{j, 2}^* \textbf{W}^*\textbf{B}\\&= \Big [\textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j)\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j ) + \textbf{N}({\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j\right] \\&\qquad + \textbf{c}_j^*\textbf{K}({\varvec{\Phi }}(\sigma _j) \otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j) + \textbf{c}_j^*\textbf{J}({\varvec{\Phi }}(\sigma _j)\otimes \textbf{b}_j)\Big ]\left( \textbf{B}\otimes 1\right) \\&=\textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j ) + \textbf{N}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes \textbf{b}_j\right] \\&\qquad + \textbf{c}_j^*\textbf{K}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}\textbf{b}_j) + \textbf{c}_j^*\textbf{J}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes \textbf{b}_j)\\&= \Big [\textbf{c}_j^*\textbf{C}{\varvec{\Phi }}(2\sigma _j)\left[ \textbf{Q}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}) + \textbf{N}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes \textbf{I}_m )\right] \\&\qquad + \textbf{c}_j^*\textbf{K}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes {\varvec{\Phi }}(\sigma _j)\textbf{B}) + \textbf{c}_j^*\textbf{J}({\varvec{\Phi }}(\sigma _j)\textbf{B}\otimes \textbf{I}_m)\Big ](\textbf{I}_m \otimes \textbf{b}_j) \\&= \textbf{c}_j^*\textbf{H}_2(\sigma _j, \sigma _j)(\textbf{I}_m \otimes \textbf{b}_j), \qquad j=1, \dots , \ell , \end{aligned}$$

which is 14d.

B   System transformation

The QB-QB system Eq. 2 can be transformed into an equivalent system with linear output, but dynamics with higher nonlinearity. Specifically, we introduce the auxiliary variable

$$\begin{aligned} \textbf{z}(t) = \textbf{K}(\textbf{x}(t) \otimes \textbf{x}(t)) + \textbf{J}(\textbf{x}(t) \otimes \textbf{u}(t)). \end{aligned}$$
(3e)

To simplify the following presentation, we again assume that \(\textbf{K}\) is symmetric Eq. 17. Assuming that the input \(\textbf{u}\) is differentiable (if \( \textbf{J}\not = 0\)) and using the symmetry of \(\textbf{K}\) we obtain

$$\begin{aligned} \frac{d}{dt}\textbf{z}(t) =&2 \textbf{K}\big ( \textbf{x}(t) \otimes \frac{d}{dt}\textbf{x}(t) \big ) + \textbf{J}\big ( \frac{d}{dt}\textbf{x}(t) \otimes \textbf{u}(t) \big ) + \textbf{J}\big ( \textbf{x}(t) \otimes \frac{d}{dt}\textbf{u}(t) \big ). \end{aligned}$$

The derivative of \(\textbf{x}(t)\) is replaced using Eq. 2a. The QB-QB system Eq. 2 can be equivalently written as

$$\begin{aligned} \textbf{E}\frac{d}{dt}\textbf{x}(t) =&\textbf{A}\textbf{x}(t) + \textbf{Q}(\textbf{x}(t) \otimes \textbf{x}(t)) + \textbf{N}(\textbf{x}(t) \otimes \textbf{u}(t)) + \textbf{B}\textbf{u}(t), \quad t \in (0, T),\end{aligned}$$
(4a)
$$\begin{aligned} \frac{d}{dt}\textbf{z}(t) =&2 \textbf{K}\Big (\textbf{x}(t) \otimes \textbf{E}^{-1} \big ( \textbf{A}\textbf{x}(t) + \textbf{Q}(\textbf{x}(t) \otimes \textbf{x}(t)) + \textbf{N}(\textbf{x}(t) \otimes \textbf{u}(t)) + \textbf{B}\textbf{u}(t) \big ) \Big ) \nonumber \\&+ \textbf{J}\Big ( \textbf{E}^{-1} \big ( \textbf{A}\textbf{x}(t) + \textbf{Q}(\textbf{x}(t) \otimes \textbf{x}(t)) \ + \textbf{N}(\textbf{x}(t) \otimes \textbf{u}(t)) + \textbf{B}\textbf{u}(t) \big ) \otimes \textbf{u}(t) \Big ) \nonumber \\&+ \textbf{J}\big ( \textbf{x}(t) \otimes \frac{d}{dt}\textbf{u}(t) \big ), \hspace{32ex} t \in (0, T), \end{aligned}$$
(4b)
$$\begin{aligned} \textbf{x}(0) =&\textbf{0}, \quad \textbf{z}(0) = \textbf{0},\end{aligned}$$
(5)
$$\begin{aligned} \textbf{y}(t) =&\textbf{C}\textbf{x}(t) + \textbf{z}(t) + \textbf{D}\textbf{u}(t), \hspace{30ex} t \in (0, T), \end{aligned}$$
(6a)

One could rearrange the right hand side in Eq. 53b using Kronecker product properties like Eq. 1. While the transformed system Eq. 53 has a linear output, the price one pays are nonlinear terms of higher order (up to cubic) in the dynamics and the presence (if \( \textbf{J}\not = 0\)) of derivatives of the inputs. In general these additional terms cannot be dealt with easily by current system or interpolation-based model reduction techniques. However, if the original dynamics are linear and the output is quadratic, then the resulting transformed system is a QB system, to which in principle current model reduction techniques can be applied.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Diaz, A.N., Heinkenschloss, M., Gosea, I.V. et al. Interpolatory model reduction of quadratic-bilinear dynamical systems with quadratic-bilinear outputs. Adv Comput Math 49, 95 (2023). https://doi.org/10.1007/s10444-023-10096-2

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1007/s10444-023-10096-2

Keywords

Search

Navigation