TL;DR
This paper moves data-driven control from "trial and error" to "guaranteed by design." The authors provide explicit mathematical formulas to create input sequences that are guaranteed to be "persistently exciting" (PE)—meaning they contain enough information to identify or control a system—for linear, Hammerstein, and a broad class of nonlinear (SISO flat) systems.
What Problem Does This Paper Address?
In data-driven control, we skip the step of building a complex mathematical model and instead design controllers directly from experimental data. However, for this to work, the data must be "rich" enough to capture the system's dynamics. This property is called Persistence of Excitation (PE).
For linear systems, PE is well-understood: if you shake the system hard enough with random noise, you usually get good data. But for nonlinear systems, the requirements are much stricter. Currently, most researchers perform an experiment, collect data, and then check after the fact (a posteriori) if the data was good enough. If it wasn't, they have to start over. There is a significant lack of guidelines on how to design inputs before the experiment (a priori) to guarantee that the resulting data will satisfy the necessary rank conditions for nonlinear control.
Methods
The authors propose structured, sparse input sequences (often just a series of timed pulses) that guarantee PE for three categories of systems:
- Linear Time-Invariant (LTI) Systems: They prove that a simple sequence of discrete impulses—giving a pulse to each input channel every $L$ steps—is sufficient to satisfy PE conditions. This is much more "sample-efficient" and less invasive than continuous random noise.
- Hammerstein Systems: These are systems where a nonlinear function acts on the input before it enters a linear system. The authors show that by choosing specific pulse magnitudes ($\lambda_j$) that make a "basis function matrix" invertible, they can guarantee PE for the entire system.
- SISO Flat Nonlinear Systems: These are complex nonlinear systems that can be "straightened out" into a linear-like form through coordinate changes. The authors use monomial basis functions and a "mosaic Hankel matrix" approach. They design a sequence of inputs that systematically "probes" the system's reachability space to ensure the data matrix has full rank.
Datasets
As a theoretical paper in control theory, it does not use large-scale public datasets (like ImageNet). Instead, the authors evaluate their methods using:
- Numerical Simulations: Specifically, a second-order Single-Input Single-Output (SISO) flat nonlinear system.
- Comparative Baselines: They compare their designed input against standard random input sampling (uniform distribution).
Key Insights and Ideas
- Structure over Randomness: While random noise can work, structured pulses are mathematically guaranteed to work. This is particularly useful in systems with bandwidth constraints or where "frequent" inputs are undesirable (e.g., medical bolus delivery).
- The Power of Sparsity: You don't need to excite the system at every time step. For an LTI system, a pulse every $L$ steps is enough to satisfy the rank conditions required by the "Fundamental Lemma" of data-driven control.
- Collective PE: For nonlinear systems, the authors utilize "collective persistence of excitation," which combines data from multiple short experiments rather than one long one. This helps prevent unstable nonlinear systems from "blowing up" during a long identification phase.
- Inverting the Basis: The core trick for Hammerstein and flat systems is ensuring that the chosen input magnitudes, when passed through the system's basis functions (like $u, u^2, u^3$), result in a matrix that is invertible.
Key Results
The authors tested their "designed" input against random inputs on an unstable nonlinear system:
- Higher Information Content: The designed input (Controller iii) achieved a significantly higher "level" of PE. Quantitatively, the minimum singular value ($\sigma_{min}$) of the data matrix—a measure of data richness—was 0.0399 for the designed input, compared to only 0.0083 for the random input.
- Better Control Performance: Because the data was richer, the resulting controller was more accurate. The average stabilization error for the first state ($x_1$) was 0.0553 with the designed input, outperforming the random input's error of 0.0620.
- Robustness: The designed input approach resulted in fewer "unstable" experiments compared to random sampling, suggesting that structured excitation is safer for sensitive nonlinear systems.
Limitations and Future Work
- Knowledge of Basis Functions: The method for nonlinear systems assumes the user can choose a set of basis functions (like polynomials or sinusoids) that can at least approximately represent the system's nonlinearities.
- Local Reachability: The results for nonlinear systems currently rely on the system being "locally reachable" at the origin.
- SISO Focus: While the LTI and Hammerstein results are general, the specific design for flat nonlinear systems is currently tailored for Single-Input Single-Output (SISO) cases. Extending this to Multi-Input Multi-Output (MIMO) nonlinear systems is a logical next step.
- Noise Sensitivity: The paper focuses on the "rank" of the matrices (the deterministic case). Future work could explore how these sparse pulse designs perform in the presence of heavy measurement noise.
Summary updated