System Level Parameterization Overview

• System Level Parameterization (SLP) is a method using convex and affine parameterizations that systematically characterizes all internally stabilizing controllers.
• SLP underpins applications such as tube-based MPC, robust distributed control, and data-driven synthesis by shifting the focus to system-level responses.
• SLP supports scalable hardware/software co-design and pipeline modeling by enabling structured, local controller synthesis with quantifiable performance tradeoffs.

System Level Parameterization (SLP) refers to a set of convex parameterizations for the achievable closed-loop responses and structural design of controllers for linear dynamical systems. SLP provides an affine representation of all internally stabilizing controllers via maps from disturbances to internal system variables, shifting attention from controller search to system-level responses. SLP underpins recent advances in distributed control, robust model predictive control (MPC), direct data-driven synthesis, and large-scale optimization for hardware and software systems. The concept has also been extended to pipeline modeling and data-driven estimation within distinct domains.

1. Mathematical Foundation of System Level Parameterization

SLP is formulated for discrete-time linear time-invariant (LTI) systems of the form: $x[t+1] = A\,x[t] + B\,u[t] + w[t],$ where $x[t] \in \mathbb{R}^n$ is the state, $u[t] \in \mathbb{R}^m$ is the control input, and $w[t]$ is an exogenous disturbance. For internally stabilizing, causal state-feedback controllers $u = K(z)x$, SLP parameterizes all achievable closed-loop disturbance responses: $$\Phi_x(z) = (zI - A - BK(z))^{-1}, \quad \Phi_u(z) = K(z)(zI - A - BK(z))^{-1},$$ subject to the affine achievability constraint: $(zI - A)\,\Phi_x(z) + B\,\Phi_u(z) = I,$ with the stability requirement $$\Phi_x(z), \Phi_u(z) \in \frac{1}{z}\mathcal{RH}_\infty$$ (Anderson et al., 2019). The controller is recovered as $K(z) = \Phi_u(z)\Phi_x^{-1}(z)$.

For output feedback or block-structured systems, SLP extends to four-block parameterizations: $\begin{bmatrix}x\u\end{bmatrix} = \begin{bmatrix} \Phi_{xx} & \Phi_{xy} \ \Phi_{ux} & \Phi_{uy} \end{bmatrix} \begin{bmatrix}\delta_x \ \delta_y\end{bmatrix},$ with two affine constraints involving the plant's dynamics and output matrices (Zheng et al., 2019).

In the finite-horizon setting, the affine constraint becomes: $x[t] \in \mathbb{R}^n$0 where $x[t] \in \mathbb{R}^n$1 is the block down-shift operator (Sieber et al., 2021, Xue et al., 2020).

2. SLP in Robust and Tube-based Model Predictive Control

SLP is central in the modern treatment of robust and tube-based MPC. By representing the closed-loop system via affine operators, SLP allows both the nominal and error dynamics to be parameterized within a convex framework. Given horizon $x[t] \in \mathbb{R}^n$2, operators are typically further decomposed as: $x[t] \in \mathbb{R}^n$3

$x[t] \in \mathbb{R}^n$4

where $x[t] \in \mathbb{R}^n$5, $x[t] \in \mathbb{R}^n$6 handle the nominal trajectory, and $x[t] \in \mathbb{R}^n$7, $x[t] \in \mathbb{R}^n$8 describe the impulse response to disturbance (Sieber et al., 2021, Sieber et al., 2021).

Imposing a finite impulse response (FIR) constraint (i.e., $x[t] \in \mathbb{R}^n$9) restricts disturbance effects to a finite window and ensures all reachable error states are contained within a sequence of disturbance-reachable sets (“system-level tubes”): $u[t] \in \mathbb{R}^m$0 where $u[t] \in \mathbb{R}^m$1 is the disturbance set and $u[t] \in \mathbb{R}^m$2 denotes the Minkowski sum.

In SLP-based tube MPC, online optimization variables include both the nominal trajectory and SLP tube operators (Φ blocks), subject to affine, FIR, state/input constraint-tightening, and terminal set conditions. This enables concurrent optimization for minimal conservatism. Alternatively, tubes may be precomputed offline, reducing the online problem to standard nominal MPC but retaining less-conservative tube shapes and positively invariant terminal sets (Sieber et al., 2021).

Empirical results show that such SLP-based robust MPC (SLTMPC) achieves region-of-attraction enlargement and reduced closed-loop cost relative to precomputed tubes, with solve time and performance between full disturbance-feedback MPC and standard tube MPC (Sieber et al., 2021).

3. SLP in Data-Driven Control and Estimation

Recent developments generalize SLP to direct data-driven synthesis. For an unknown system with trajectories $u[t] \in \mathbb{R}^m$3 under bounded noise, the convex set of all dynamics consistent with the data is characterized by a quadratic matrix inequality (QMI): $u[t] \in \mathbb{R}^m$4 for suitably constructed $u[t] \in \mathbb{R}^m$5, $u[t] \in \mathbb{R}^m$6, and reference parameters $u[t] \in \mathbb{R}^m$7 (Brändle et al., 2024). This provides an exact and convex description of all plants compatible with the data, allowing for robust estimator or controller synthesis via LMI techniques and guaranteeing worst-case $u[t] \in \mathbb{R}^m$8 performance over all consistent models.

In fully data-driven SLP for controller synthesis, closed-loop system responses are constructed directly from input–state Hankel matrices of persistently exciting trajectories, replacing the need for explicit model identification. Under noise-free conditions, this yields exact equivalency with model-based synthesis; with process noise, quasi-convex performance optimization over robustified SLP constraints yields bounds on suboptimality scaling with noise and sample size (Xue et al., 2020).

The dual of SLP has also been introduced for closed-loop identification: estimation of the underlying plant from closed-loop data amounts to solving the SLP affine constraints with the controller treated as fixed, again admitting convex optimization and bypassing the need for nominal-plant knowledge (Srivastava et al., 2023).

4. SLP in Structured, Localized, and Large-Scale Control

SLP enables direct enforcement of locality, sparsity, and separable structure in distributed feedback. By constraining the supports of the impulse responses $u[t] \in \mathbb{R}^m$9 (e.g., imposing spatial or communication radii in large-scale systems), SLP allows the decentralized or partially distributed synthesis of controllers via convex programs (Anderson et al., 2019, Du et al., 2024).

In continuous time, SLP extends via Laplace-domain affine constraints: $w[t]$0 with rational, strictly proper $w[t]$1. Residualization onto preselected stable poles allows the SLP constraints and performance objectives to be cast as finite-dimensional convex programs, with further decomposition for parallel or local solution in graph-structured systems (Du et al., 2024).

This property uniquely enables practically scalable synthesis for large-scale distributed networked systems, as the synthesis complexity can be reduced to per-node or per-region subproblems, validated in grid control and power-system examples.

5. SLP in Pipeline Modeling and Hardware-Software Design

In system-level hardware modeling, SLP denotes decomposition of pipeline models into orthogonal, composable policy classes—Function, Communication, Timing, and Process. Policies are implemented as C++ templates in SystemC, enabling compile-time binding and operator-overloaded embedded DSLs for rapid pipeline topology specification. Performance modeling is parameterized via explicit formulas for stage delay, communication overhead, and throughput: $w[t]$2 where $w[t]$3 is the minimum initiation interval and $w[t]$4 the maximum service time (0801.2201).

The policies enable extensive, systematic design-space exploration in hardware/software co-design without rewriting code, with the DSL yielding automatic router/channel instantiation.

6. Numerical Robustness, FIR Truncation, and Comparison to Classical Parameterizations

SLP parameters are determined by affine constraints, so practical computation (especially with FIR truncation and floating-point arithmetic) may result in small infeasibilities. Analyses show that for state-feedback or open-loop stable plants, the SLP-based controller recovery is robust to such constraint perturbations (Zheng et al., 2019). The four-block SLP (full output-feedback), however, may require case-by-case analysis for robustness. Notably, SLP compares directly with the Youla and input–output parameterizations: SLP achieves full convexity and system-level transparency, and admits direct constraints on locality and structure not supported in Youla-based (quadratically-invariant) settings (Zheng et al., 2019).

Under FIR constraints, the fraction of achievable stabilizing controllers is ordered: $w[t]$5 with SLP being most conservative under a fixed horizon, but most natural for state-feedback and locality.

7. Applications and Advantages

SLP is foundational in:

• Robust and tube-based MPC with adjustable conservatism, enabling both offline and online optimization of tubes (Sieber et al., 2021, Sieber et al., 2021).
• Data-driven control and estimation with rigorous end-to-end sample complexity and guaranteed robustness (Brändle et al., 2024, Xue et al., 2020).
• Large-scale distributed and localized control, supporting explicit performance-locality tradeoffs (Anderson et al., 2019, Du et al., 2024).
• Hardware/Software co-design via generalized policy classes in system pipeline modeling (0801.2201).

SLP unifies and generalizes classical control synthesis architectures, provides direct convex representations for structural and robustness constraints, and underpins scalable, efficient algorithmic implementations across application domains. Its system-centric focus enables transparent performance assessment under model uncertainty and supports direct quantification of learning–robustness tradeoff.

References: (Anderson et al., 2019, Sieber et al., 2021, Sieber et al., 2021, Zheng et al., 2019, Brändle et al., 2024, Xue et al., 2020, Srivastava et al., 2023, Du et al., 2024, 0801.2201).