Persistence of Excitation in Data-Driven Control

289 concepts185 papers1 proposal

Imagine trying to figure out how a car handles without ever turning the steering wheel more than a fraction of a degree, or trying to map out its suspension while driving exclusively on a perfectly smooth, flat highway. You might collect millions of data points of the car moving forward at a constant speed, but that data is fundamentally uninformative about how the car behaves during a sharp turn or over a pothole.

In modern control theory, bypassing physics-based mathematical models to design controllers directly from operational data has become a powerful paradigm. However, as Julian Berberich et al. emphasize in Robust data-driven state-feedback design, this data-driven synthesis relies on a foundational requirement: the input signals used to collect the data must be persistently exciting (PE). Intuitively, persistence of excitation means the input signal must be rich and varied enough to shake the system out of its quiet states, forcing it to reveal all of its internal dynamics and hidden modes.


The Geometry of Informativity

Why exactly does data-driven control require this richness? The answer lies in how we reconstruct system trajectories without an explicit model. A cornerstone of this field is Willems' Fundamental Lemma, which proves that any valid trajectory of a linear time-invariant system can be represented as a linear combination of previously collected data—provided that the historical input sequence is persistently exciting.

Mathematically, this translates to a rank condition on data matrices (such as Hankel matrices). If an input signal lacks sufficient variation, the columns of these matrices become linearly dependent, creating a flat optimization landscape where parameters or control laws cannot be uniquely identified. As the authors of On the Synthesis of Bellman Inequalities for Data-Driven Optimal Control point out, the informativity of a dataset—its level of persistent excitation—is far more critical than the total volume of data collected. A short, highly excited trajectory can fully characterize a system, whereas a massive dataset of redundant, quiet behavior fails completely.

Low Excitation (Constant Input)  --> Redundant Data Columns --> Low Matrix Rank --> Unidentifiable Dynamics
High Excitation (Varied Input)   --> Diverse Data Columns   --> Full Matrix Rank --> Fully Characterized System

The exact requirement for how "rich" a signal needs to be depends heavily on the system's underlying structure. For instance, Julian Berberich et al. demonstrated in A quantitative and constructive proof of Willems' Fundamental Lemma and its implications that the required order of persistence of excitation is directly influenced by the number of system zeros. When moving from purely linear systems to affine systems (which include a constant bias term), Andrea Martinelli et al. showed in Data-Driven Optimal Control of Affine Systems: A Linear Programming Perspective that an extra degree of freedom in the excitation order is strictly necessary to mathematically isolate the bias without a prior model.


The Paradox of Control Success

One of the most fascinating challenges of the persistence of excitation requirement is what happens when a data-driven controller actually succeeds.

Suppose you deploy a data-driven predictive controller to stabilize a system at a specific target setpoint. As the controller does its job, the system smoothly approaches the setpoint and settles into a steady state. The input variance drops to near zero.

This creates a fundamental paradox. As highlighted by Julian Berberich et al. in Data-driven model predictive control: closed-loop guarantees and experimental results and Linear tracking MPC for nonlinear systems Part II: The data-driven case, reaching a steady state kills the persistence of excitation. Without input variation, the data buffer becomes redundant, the local linear approximations degrade, and the controller loses its predictive power. This loss of informativity can cause the system to drift or break out into undesirable oscillations just to re-excite its own dynamics.

[Controller Succeeds] ──> [System Stabilizes] ──> [Input Variance Drops]
         ▲                                                 │
         │ (System oscillates to regain PE)                ▼
[Predictive Power Lost] <── [Data Buffer Degrades] <── [PE is Killed]

To combat this, online learning algorithms often have to inject artificial "probing noise" into the control signal. However, Feiran Zhao et al. note in Data-Enabled Policy Optimization for Direct Adaptive Learning of the LQR that this introduces a painful trade-off: the added noise degrades immediate tracking performance and efficiency in order to gather the information necessary for long-term stability.


Real-World Constraints and Safety Limits

In a simulation, injecting high-frequency probing noise to satisfy PE is trivial. In real-world physical infrastructure, it can be dangerous or prohibited.

Consider controlling electrical power grids. As Linbin Huang et al. explain in Data-Enabled Predictive Control for Grid-Connected Power Converters, data-driven predictive control can stabilize weak grids, but injecting the necessary level of noise to achieve persistent excitation is strictly restricted by utility regulations and safety requirements. Similarly, in medical applications like upper-extremity rehabilitation robots, Borna Ghannadi et al. note in Adaptive Control of an Upper Extremity Rehabilitation Robot with Backlash that high-frequency, jerky inputs violate user comfort and safety, yet avoiding them risks parameter drift over long durations.

Furthermore, collecting pure open-loop excitation data is often an impossible luxury. Industrial systems must remain under continuous feedback control for stability and operational continuity. This has forced researchers like Yibo Wang et al. in Data-Driven Predictive Control Using Closed-Loop Data: An Instrumental Variable Approach to develop specialized identification approaches that can extract system dynamics from closed-loop data, where the excitation is inherently constrained by the existing feedback loop.


The Nonlinear Frontier

While PE for linear systems can be verified via straightforward rank tests on inputs, verifying data richness becomes significantly more complex as we move to nonlinear regimes.

For nonlinear systems, the richness of the data depends heavily on the specific states visited during data collection. As Zhenyi Yuan et al. observe in Data-Driven Optimal Control of Bilinear Systems, simple offline input selection is no longer enough to guarantee that a trajectory will be representative of the entire state space. In flat nonlinear systems, Mohammad Alsalti et al. leverage geometric transformations in Data-Based System Analysis and Control of Flat Nonlinear Systems to map data back to linear spaces, but verifying that the original data satisfies PE remains a major hurdle.

Even in model-free reinforcement learning and deep learning representations for control, the ghost of PE remains. In Finite-time System Identification and Adaptive Control in Autoregressive Exogenous Systems, Sahin Lale demonstrates that finite-time regret bounds and control stability are fundamentally constrained by how effectively the system's inputs excite its state space. Ultimately, whether a controller uses a classical linear buffer or a deep neural network, it cannot learn what it has not seen; the system must be shaken before it can be controlled.

Key Foundational Work

Basic adaptive control theory and its application to robot control-II. (1985) — contributed 3 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 201 papers
  2. Version 1 · 2026-10-04 · 163 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_persistent_excitation_requirement_0b79187c,
  title = {persistent excitation requirement},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/persistent-excitation-requirement/rdr_0b79187c88c24573a2c3f26cd16c27b2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_persistent_excitation_requirement_0b79187c_v2,
  title = {persistent excitation requirement},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/persistent-excitation-requirement/rdr_0b79187c88c24573a2c3f26cd16c27b2/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake