Low-dimensional LPV approximations for nonlinear control

Blacksburg – May 2023

Jan Heiland & Peter Benner & Steffen Werner (MPI Magdeburg)


\[\dot x = f(x) + Bu\]

Control of an inverted pendulum

  • 9 degrees of freedom
  • but nonlinear controller.

Stabilization of a laminar flow

  • 50’000 degrees of freedom
  • but linear regulator.

Control of Nonlinear & Large-Scale Systems

A general approach would include

  • powerful backends (linear algebra / optimization)
  • exploitation of general structures
  • model order reduction
  • data-driven surrogate models
  • all of it?!

LPV Representation

\[\begin{align} \dot x -Bu & = f(x) \\ & \approx [A_0+\rho_1(x)A_1+ \dotsm + \rho_r(x) A_r]\, x \end{align}\]

The linear parameter varying (LPV) representation/approximation \[ \dot x \approx \bigl [\Sigma \,\rho_i(x)A_i \bigr]\, x + Bu \] for nonlinear controller comes with

  • a general structure (linear but parameter-varying)

and extensive theory on

  • LPV controller design


In this talk, we will consider LPV series expansions of control laws.

LPV system approaches

For linear parameter-varying systems \[ \dot x = A(\rho(x))\,x + Bu \] there exist established methods that provide control laws based one

  • robustness against parameter variations (Peaucelle and Arzelier 2001)
  • adaption with the parameter, i.e. gain scheduling, (Apkarian, Gahinet, and Becker 1995)

A major issue: require solutions of coupled LMI systems.

SDRE series expansion

Consider the optimal regulator control problem

\[ \int_0^\infty \|y\|^2 + \alpha \|u\|^2\, \mathsf{d}s \to \min_{(y, u)} \] subject to \[ \dot x = A(\rho(x))\,x+Bu, \quad y=Cx. \]

Theorem (Beeler, Tran, and Banks 2000)

If there exists \(\Pi\) as a function of \(x\) such that \[ \begin{aligned} & \dot{\Pi}(x)+\bigl[\frac{\partial(A(\rho(x)))}{\partial x}\bigr]^T \Pi(x)\\ & \quad+\Pi(x) A(\rho(x))+A^T(\rho(x)) \Pi(x)-\frac{1}{\alpha} \Pi(x) BB^T \Pi(x)=-C^TC . \end{aligned} \]

Then \[u=-\frac{1}{\alpha}B^T\Pi(x)\,x\] is an optimal feedback for the control problem.

In Praxis, parts of the HJB are discarded and we use \(\Pi(x)\) that solely solves the state-dependent Riccati equation (SDRE) \[ \Pi(x) A(\rho(x))+A^T(\rho(x)) \Pi(x)-\frac{1}{\alpha} \Pi(x) BB^T\Pi(x)=-C^TC, \] and the SDRE feedback \[ u=-\frac{1}{\alpha}B^T\Pi(x)\,x. \]

  • numerous application examples and
  • proofs of performance (Banks, Lewis, and Tran 2007)
  • also beyond smallness conditions (Benner and Heiland 2018)
  • Although the SDRE is an approximation already,

  • the repeated solve of the Riccati equation is not feasible.

  • However, for affine LPV systems, a series expansion

  • enables an efficient approximation at runtime.

The series expansion

We note that \(\Pi\) depends on \(x\) through \(A(\rho(x))\).

Thus, we can consider \(\Pi\) as a function in \(\rho\) and its corresponding multivariate Taylor expansion up to order \(K\) \[\begin{equation} \label{eq:taylor-expansion-P} \Pi (\rho) \approx \Pi (0) + \sum_{1\leq |\beta| \leq K} \rho^{(\beta)}P_{\beta}, \end{equation}\] where

  • \(\beta=(\beta_1, \dotsc, \beta_r)\in \mathbb N^r\) is a multiindex and the
  • \(P_{\beta}\in \mathbb R^{n\times n}\) are constant matrices.


If \(A(\rho)\) is affine, i.e. \(A(\rho) = A_0 + \sum_{k=1}^r \rho_k A_k\).

Then the coefficients of the first order Taylor approximation \[ \Pi (\rho) \approx \Pi(0) + \sum_{|\beta| = 1} \rho^{(\beta)}P_{\beta} =: P_0 + \sum_{k=1}^r \rho_k L_k. \] are the solutions to

  • \(A_{0}^{T} P_{0} + P_{0} A_{0} - P_{0} B B^{T} P_{0} = -C^{T} C\),

and, for \(k=1,\dotsc,r\),

  • \((A_{0} - B B^{T} P_{0})^{T} L_{k} + L_{k} ( A_{0} - B B^{T} P_{0} )= -(A_{k}^{T} P_{0} + P_{0} A_{k})\).


Insert the Taylor expansion of \(\Pi\) and the LPV representation of \(A\) into the SDRE and match the coefficients.


The corresponding nonlinear feedback is realized as \[ u = -\frac{1}{\alpha}B^T[P_0 + \sum_{k=1}^r \rho_k(x) L_k]\,x. \]

Cp., e.g., (Beeler, Tran, and Banks 2000) and (Alla, Kalise, and Simoncini 2023).

Intermediate Summary

A representation/approximation of the nonlinear system via \[ \dot x = [A_0 + \sum_{k=1}^r \rho_k(x) A_k]\, x + Bu \] enables the nonlinear feedback design through truncated expansions of the SDRE.

How to Design an LPV approximation

A general procedure

If \(f(0)=0\) and under mild conditions, the flow \(f\) can be factorized \[ f( x) = [A(x)]\,x \] with some \(A\colon \mathbb R^{n} \to \mathbb R^{n\times n}\).

  1. If \(f\) has a strongly continuous Jacobian \(\partial f\), then \[ f(x) = [\int_0^1 \partial f(sx)\mathsf{d} s]\, x \]
  2. The trivial choice of \[ f(x) = [\frac{1}{x^Tx}f(x)x^T]\,x \] doesn’t work well – neither do the improvements (Lin, Vandewalle, and Liang 2015).

For the factorization \(f(x)=A(x)\,x\), one can say that

  1. it is not unique
  2. it can be a design parameter
  3. often, it is indicated by the structure.

… like in the advective term in the Navier-Stokes equations: \[ (v\cdot \nabla)v = \mathcal A_s(v)\,v \] with \(s\in[0,1]\) and the linear operator \(\mathcal A_s(v)\) defined via \[\mathcal A_s(v)\,w := s\,(v\cdot \nabla)w + (1-s)\, (w\cdot \nabla)v.\]

Now, we have an state-dependent coefficient representation

\[ f(x) = A(x)\,x.\]

How to obtain an LPV representation/approximation?

\(\dot x = A(x)\,x + Bu\)

  • Trivially, this is an LPV representation \[ \dot x = A(\rho(x))\, x + Bu \] with \(\rho(x) = x\).

  • Take any model order reduction scheme that compresses (via \(\mathcal P\)) the state and lifts it back (via \(\mathcal L\)) so that \[ \tilde x = \mathcal L(\hat x) = \mathcal L (\mathcal P(x)) \approx x \]

  • Then \(\rho = \mathcal P(x)\) gives a low-dimensional LPV approximation by means of \[ A(x)\,x \approx A(\tilde x)\, x = A(\mathcal L \rho (x))\,x. \]


  • If \(x\mapsto A(x)\) itself is affine linear

  • and \(\mathcal L\) is linear,

  • then \[ \dot x \approx A(\mathcal L \rho(x))\,x + Bu = [A_0 + \sum_{i=1}^r \rho_i(x) A_i]\, x + Bu \] is affine with

    • \(\rho_i(x)\) being the components of \(\rho(x)\in \mathbb R^r\)
    • and constant matrices \(A_0\), \(A_1\), …, \(A_r \in \mathbb R^{n\times n}\).

Intermediate Summary

  • Generally, a nonlinear \(f\) can be factorized as \(f(x) = A(x)\,x\).

  • Model order reduction provides a low dimensional LPV representation \(A(x)\,x\approx A(\mathcal \rho(x))\,x\).

  • The needed affine-linearity in \(\rho\) follows from system’s structure (or from another layer of approximation (see, e.g, (Koelewijn and Tóth 2020)).

Numerical Realization

The Navier-Stokes equations

\[ \dot v + (v\cdot \nabla) v- \frac{1}{\mathsf{Re}}\Delta v + \nabla p= f, \]

\[ \nabla \cdot v = 0. \]

Control Problem:

  • use two small outlets for fluid at the cylinder boundary
  • to stabilize the unstable steady state
  • with a few point observations in the wake.

Simulation model:

  • we use finite elements to obtain
  • the dynamical model of type

\(\dot x = Ax + N(x,x) + Bu, \quad y = Cx\)

  • with \(N\) being bilinear in \(x\)
  • and a state dimension of about \(n=50'000\).

The Algorithm

Nonlinear controller design for \[ \dot x = f(x) + Bu \] by LPV approximations and truncated SDRE expansions.

  1. Compute an affine LPV approximative model with \[f(x)\approx A_0x + \sum_{k=1}^r \rho_k(x)A_kx.\]

  2. Solve one Riccati and \(r\) Lyapunov equations for \(P_0\) and the \(L_k\)s.

  3. Close the loop with \(u = -\frac{1}{\alpha}B^T[P_0x + \sum_{k=1}^r \rho_k(x) L_kx ].\)

Step-1 – Compute the LPV Approximation

We use POD coordinates with the matrix \(V\in \mathbb R^{n\times r}\) of POD modes \(v_k\)

  • \(\rho(x) = V^T x\),

  • \(\tilde x = V\rho(x)=\sum_{k=1}^r\rho_i(x)v_k.\)

Then: \[N(x,x)\approx N(\tilde x, x) = N(\sum_{k=1}^r\rho_i(x)v_k, x) = \sum_{k=1}^r\rho_i(x) N(v_k, x) \] which is readily realized as \[ [\sum_{k=1}^r\rho_i(x) A_k]\,x.\]

Step-2 – Compute \(P_0\) and the \(L_k\)s

This requires solutions of large-scale (\(n=50'000\)) matrix equations

  1. Riccati – nonlinear but fairly standard
  2. Lyapunovs – linear but indefinite.

We use state-of-the-art low-rank ADI iterations (ask Steffen for details).

Step-3 – Close the Loop

  • Setup: Start from the steady-state
  • Goal: Stabilize the steady-state

Comparison of feedback designs

  • LQR – plain LQR controller
  • xSDRE-r – truncated (at r) SDRE feedback

Parameters of the Control Setup

We check the performance with respect to two parameters

  • \(\alpha\) … the regularization parameter that penalizes the control

  • \(t_{\mathsf c} > 0\) … time before the controller is activated

  • The parameter \(t_c\) describes the domain of attraction.

  • For r=0 the xSDRE-r feedback recovers the LQR feedback.

Norm plot of the feedback signals.

  • LQR fails to stabilize
  • increasing r means better performance
  • stability achieved at r=10

Less regularization

  • less smooth feedback actions
  • again LQR fails
  • xSDRE can achieve stability
  • stability achieved for certain r

The Full Picture

Conclusion for the Numerical Results

  • Measurable and reliable improvements with respect to \(\alpha\)

    • more performant feedback action at higher regularization
  • no measurable performance gain with respect to \(t_{\mathsf c}\)

    • no extension of the domain of attraction
  • still much space for improvement

    • find better bases for the parametrization?
    • increase the r?
    • second order truncation of the SDRE?


… and Outlook

  • General approach to model structure reduction by low-dimensional affine LPV systems.

\[f(x) \quad \to\quad A(x)\,x\quad \to\quad \tilde A(\rho(x))\,x\quad \to\quad [A_0 + \sum_{k=1}^r\rho_k(x)A_k]\,x\]

  • Proof of concept for nonlinear controller design with POD and truncated SDRE (Heiland and Werner 2023).

  • General and performant but still heuristic approach.

  • Detailed roadmap for developing the LPV (systems) theory is available.

  • PhD student wanted!

Thank You!

References -- LPV Systems

Apkarian, Pierre, Pascal Gahinet, and Greg Becker. 1995. “Self-Scheduled \(H_\infty\) Control of Linear Parameter-Varying Systems: A Design Example.” Autom. 31 (9): 1251–61. https://doi.org/10.1016/0005-1098(95)00038-X.
Peaucelle, D., and D. Arzelier. 2001. “Robust Performance Analysis with LMI-Based Methods for Real Parametric Uncertainty via Parameter-Dependent Lyapunov Functions.” IEEE Trans. Automat. Control 46 (4): 624–30. https://doi.org/10.1109/9.917664.
Koelewijn, Patrick J. W., and Roland Tóth. 2020. “Scheduling Dimension Reduction of LPV Models - A Deep Neural Network Approach.” In 2020 American Control Conference, 1111–17. IEEE. https://doi.org/10.23919/ACC45564.2020.9147310.


Banks, H. T., B. M. Lewis, and H. T. Tran. 2007. “Nonlinear Feedback Controllers and Compensators: A State-Dependent Riccati Equation Approach.” Comput. Optim. Appl. 37 (2): 177–218. https://doi.org/10.1007/s10589-007-9015-2.
Beeler, S. C., H. T. Tran, and H. T. Banks. 2000. “Feedback Control Methodologies for Nonlinear Systems.” J. Optim. Theory Appl. 107 (1): 1–33. https://doi.org/10.1023/A:1004607114958.
Benner, Peter, and Jan Heiland. 2018. “Exponential Stability and Stabilization of Extended Linearizations via Continuous Updates of Riccati Based Feedback.” Internat. J. Robust and Nonlinear Cont. 28 (4): 1218–32. https://doi.org/10.1002/rnc.3949.

References--SDRE ctd

Alla, A., D. Kalise, and V. Simoncini. 2023. “State-Dependent Riccati Equation Feedback Stabilization for Nonlinear PDEs.” Adv. Comput. Math. 49 (1): 9. https://doi.org/10.1007/s10444-022-09998-4.
Lin, Li-Gang, Joos Vandewalle, and Yew-Wen Liang. 2015. “Analytical Representation of the State-Dependent Coefficients in the SDRE/SDDRE Scheme for Multivariable Systems.” Autom. 59: 106–11. https://doi.org/10.1016/j.automatica.2015.06.015.

References--LPV/SDRE Approximation

Heiland, Jan, Peter Benner, and Rezvan Bahmani. 2022. “Convolutional Neural Networks for Very Low-Dimensional LPV Approximations of Incompressible Navier-Stokes Equations.” Frontiers in Applied Mathematics and Statistics 8. https://doi.org/10.3389/fams.2022.879140.
Heiland, Jan, and Steffen W. R. Werner. 2023. “Low-Complexity Linear Parameter-Varying Approximations of Incompressible Navier-Stokes Equations for Truncated State-Dependent Riccati Feedback.” arxiv.