Persistency of Excitation

In the quest to understand and control the world around us, from complex machinery to biological cells, we often face 'black box' systems whose internal rules are unknown. How can we reliably deduce these rules simply by observing how the system responds to external stimuli? This fundamental challenge of system identification hinges on a powerful, elegant concept: ​​Persistency of Excitation (PE)​​. Without it, our efforts to model a system can lead to ambiguous results, false confidence, or even catastrophic failure. This article delves into this cornerstone principle, addressing the critical knowledge gap between simply collecting data and collecting the right data.

The journey is structured across two key chapters. In ​​Principles and Mechanisms​​, we will explore the core theory of Persistency of Excitation, answering what it means for a signal to be 'rich' enough and uncovering the severe consequences of its absence in adaptive systems. Then, in ​​Applications and Interdisciplinary Connections​​, we will transition from theory to practice, discovering how engineers and scientists purposefully design exciting signals and how the PE principle underpins a vast range of modern technologies, from active noise control and data-driven control to the cutting edge of synthetic biology.

Imagine you are a detective faced with a mysterious black box. You can't open it, but you want to understand its inner workings—the rules that govern its behavior. Your only tool is to send signals into the box (inputs) and observe what signals come out (outputs). This is the fundamental challenge of system identification: to deduce the hidden parameters of a system from its external behavior. But what kind of input signals should you use? Should you send a simple, constant signal? A single, sharp poke? A gentle, oscillating wave? As we will see, the choice of the "question" you ask the system is everything. The art and science of asking the right questions is the essence of ​​Persistency of Excitation​​.

Let's begin with a simple case. Suppose our black box implements a basic Finite Impulse Response (FIR) model. Its output at any time is just a weighted sum of its a few most recent inputs:

where the weights $\theta_1, \dots, \theta_n$ are the secret parameters we want to find. For each moment in time, we get one such equation. If we collect data over many moments, we get a large system of linear equations. The uniqueness of our solution for the parameter vector $\theta = [\theta_1, \dots, \theta_n]^\top$ depends entirely on the input values we fed into the system.

Think of it this way: if you want to determine the orientation of a flat plane in three-dimensional space, you need to measure its position at three points. But what if you choose three points that all lie on the same straight line? You'll be able to identify that line, but the plane could be tilted in any direction around it. You have infinite possible solutions. To uniquely define the plane, you must choose three points that are not collinear. You need to probe the space in a sufficiently "rich" way.

The exact same principle applies to our black box. If our input signal $u(k)$ is too simple or repetitive, it might not create enough independent "probes" to distinguish between different possible sets of parameters. For example, if we use a constant input $u(k) = c$, then all the input terms in our equation become the same, and we can only ever hope to learn their sum, $\sum \theta_i$, not the individual values. The input must be sufficiently "wiggly" or "varied" to ensure that the equations it generates are linearly independent. This need for a sufficiently rich input signal is called ​​Persistency of Excitation (PE)​​.

So, what does it mean mathematically for an input to be "rich" enough? The concept of persistency of excitation can be defined in several equivalent ways, each offering a unique insight.

​​The Time-Domain View:​​ A signal $u(t)$ is said to be persistently exciting of order $n$ if no linear combination of $n$ of its consecutive past values can be identically zero. In other words, there is no set of non-zero coefficients $\alpha_1, \dots, \alpha_n$ such that $\sum_{i=1}^n \alpha_i u(t-i+1) = 0$ for all time $t$. This means that the vectors formed by sliding a window of length $n$ along the signal, like $\varphi(t) = [u(t), \dots, u(t-n+1)]^\top$, are always linearly independent over any sufficiently long time interval. Formally, their Gramian matrix, $\sum \varphi(t)\varphi(t)^\top$, must be positive definite and bounded away from singularity.

​​The Frequency-Domain View:​​ Perhaps the most intuitive view is through the lens of frequency. Imagine trying to characterize an audio equalizer by only playing a single musical note (a pure sine wave) through it. You can measure how the equalizer affects that specific frequency, but you will learn absolutely nothing about how it affects any other frequency. To fully characterize the equalizer, you need to play a signal with a rich spectrum, like white noise or a frequency sweep. The same is true for our black box. To identify a system with $n$ parameters, the input signal must contain at least $\lceil n/2 \rceil$ distinct frequency components. A signal whose power spectral density is strictly positive over a band of frequencies is a good candidate for being persistently exciting.

​​The Statistical View:​​ If the input is a random signal, like noise, we can look at its statistical properties. The condition for persistency of excitation of order $n$ translates to a condition on the signal's autocorrelation function. Specifically, the $n \times n$ symmetric ​​Toeplitz matrix​​ formed from the autocorrelation lags, $R_u^{(n)}$, must be strictly positive definite. This guarantees that, on average, no value of the signal is perfectly predictable as a linear combination of its $n-1$ predecessors.

The plot thickens when we consider systems with feedback, where the output depends not only on past inputs but also on its own past values. A common example is the AutoRegressive with eXogenous input (ARX) model:

Here, $y(t)$ is related to past values like $y(t-1), \dots, y(t-n_a)$ and past inputs $u(t-n_k-1), \dots, u(t-n_k-n_b)$. It might seem that the feedback loop would automatically make the internal signals "wiggly" and complex, relaxing the need for a rich external input.

But this intuition is misleading. A careful analysis reveals that the ultimate source of excitation is still the external input $u(t)$. The feedback loop can propagate and color this excitation, but it cannot create it out of thin air. To uniquely identify all $n_a$ parameters of the autoregressive part ($A$) and all $n_b$ parameters of the input part ($B$), the input signal $u(t)$ must be persistently exciting of order ​​$n_a + n_b$​​—the total number of unknown parameters to be identified. The input must be rich enough to independently stimulate all the system's internal pathways so we can tell them apart. It's also worth noting that a pure time delay in the input, represented by the term $n_k$, does not change the required order of excitation; it simply shifts when the input's influence is felt.

What happens if we fail to provide a persistently exciting input? The consequences can range from misleading to catastrophic.

​​The Silent Regressor and False Confidence:​​ Consider using an input that decays to zero, such as $u(t) = \exp(-t)$. This signal is not persistently exciting because its "energy" is finite. An adaptive algorithm trying to learn the system's parameter $\theta$ might find that the prediction error, $e(t) = (\hat{\theta}(t) - \theta)u(t)$, goes to zero. We might be tempted to declare success. However, the error is only zero because the input $u(t)$ has vanished! The parameter error, $\hat{\theta}(t) - \theta$, does not go to zero; it simply freezes at whatever incorrect value it had when the input died out. We have achieved a false sense of confidence by listening to silence.

​​Stabilization Without Learning:​​ This same phenomenon is a classic issue in adaptive control. An adaptive controller might succeed in its primary goal of stabilizing a system, for instance, driving the state $x(t)$ to zero. But the adaptation law that updates the parameter estimates often depends on the state, e.g., $\dot{\hat{\theta}}(t) = \gamma x(t)^2$. As $x(t)$ goes to zero, the adaptation simply stops. LaSalle's Invariance Principle confirms this: the state $x(t)$ converges to zero, but the parameter error $\tilde{\theta}(t)$ converges to an arbitrary constant. The system is controlled, but it has not been learned. This is a crucial distinction: control does not imply identification.

​​Catastrophic Parameter Drift:​​ In a worst-case scenario, the lack of excitation can lead to utter disaster. Imagine a "self-tuning regulator" that adjusts its control law based on its current best guess of the system parameters. If the system becomes stable and the output signal fades away, PE is lost. Now, suppose the estimation algorithm includes a "leakage" term, a common feature designed to prevent parameters from drifting aimlessly. In the absence of new, exciting information from the input, this leakage can cause the parameter estimates to drift—for example, towards zero. If the control law happens to have one of these drifting parameters, say $\hat{b}(t)$, in the denominator of a gain term, $u(t) = -\frac{\hat{a}(t)-\alpha}{\hat{b}(t)}y(t)$, the result is catastrophic. As $\hat{b}(t) \to 0$, the control gain explodes, and the controller, now blind and ignorant, actively drives the stable system into violent instability.

The principle of persistency of excitation is a cornerstone of modern control and signal processing, with profound implications.

​​Fuel for Adaptive Algorithms:​​ In practical algorithms such as Recursive Least Squares (RLS) with a forgetting factor $\lambda < 1$, PE is the very "fuel" that keeps the estimation process running. It ensures that the algorithm's covariance matrix remains well-conditioned—bounded from above and below—allowing it to continuously learn and adapt. Without PE, this matrix can either become singular (if the input is not rich enough) or shrink to zero (if no forgetting is used), effectively halting the learning process.

​​The Limits of Sight:​​ While PE is powerful, it is not magic. It can only help us learn what is, in principle, knowable from input-output observations. If a system has internal dynamics that are completely disconnected from the output—what are known as ​​unobservable modes​​—then no input signal, no matter how rich, can ever reveal their properties. PE guarantees that we can find a perfect, minimal model of the system's observable input-output behavior, but it cannot grant us sight into hidden, causally disconnected parts of the system.

​​Exciting the Nonlinear World:​​ Remarkably, the concept of PE extends naturally to the identification of nonlinear systems. For a model described by a Volterra series, the regressor vector consists of monomials of the input (e.g., $u(t-1)$, $u(t-k)^2$, $u(t)u(t-j)$). To ensure these regressors are linearly independent, the input signal must be rich in a more profound sense: its ​​higher-order moments​​ must be non-degenerate. A signal that is merely PE for a linear system might not be sufficient to identify quadratic or cubic nonlinearities. And here lies a final, beautiful insight: simple zero-mean Gaussian white noise, a signal often considered "unstructured," is an excellent choice for exciting nonlinear systems. While its higher-order cumulants are zero, its higher-order moments are decidedly not, providing exactly the statistical richness needed to probe and identify complex nonlinear dynamics. This demonstrates the deep unity of the principle: to get unambiguous answers, one must always ask rich questions.

In the last chapter, we uncovered a delightful principle—Persistency of Excitation. We saw that to learn about an unknown system, to truly understand its inner workings, we can't just sit and watch it passively. We have to "poke" it, to "excite" it, in just the right way. Persistency of Excitation is the physicist's and engineer's precise language for "asking good questions." A system that is persistently excited is one that is being asked a rich and varied enough stream of questions that it is forced to reveal all of its secrets.

Now, this might sound like an abstract idea, a bit of mathematical philosophy. But the truth is, this principle is the silent, unsung hero behind an astonishing array of modern technologies and scientific inquiries. It's the difference between a self-driving car that learns and one that doesn't, between a clear cell phone call and one drowned in noise, and even between a successful gene therapy and a failed one. So, let's take a journey and see where this simple, beautiful idea shows up in the world.

Suppose you have a black box—a physical system—and you want to find a model for it. This is the classic problem of system identification. The most direct way to apply our principle is to design the input signal, the "interrogation," that we send into the box. What kind of signal should we use?

A wonderfully effective approach is to send in something that looks like random noise! A ​​Pseudo-Random Binary Sequence (PRBS)​​, for instance, is a signal that hops unpredictably between two values, say $+A$ and $-A$. While it seems random, it's actually a carefully constructed deterministic sequence. Why is this so effective? Because its randomness contains a rich mix of frequencies, it "shakes up" the system in many different ways at once, preventing it from hiding any of its dynamical modes. When we design such an input, we are practicing the art of excitation. We must ensure the signal is rich enough to identify all the parameters we care about, but also respect real-world limits, like the maximum power an actuator can handle.

But we can be more delicate than just sending in noise. If we think of identifying $p$ unknown parameters in a model, it feels a bit like solving for $p$ variables in a system of equations. This suggests a more tailored approach. Imagine instead of random noise, we send in a beautiful, harmonious chord—a ​​multisine signal​​, which is a sum of several pure sine waves at different frequencies. Each sine wave, with its corresponding cosine, can be thought of as a "question pair" we are asking the system. A remarkable result shows that to identify $p$ unknown parameters in a common linear model, you need a minimum of $K_{\min} = \lceil (n_a+n_b)/2 \rceil$ sinusoids, where $p=n_a+n_b$ is the number of parameters. There's a beautiful economy to it: the number of questions we must ask is directly related to the number of things we want to know.

Designing the perfect input is one thing, but often the world doesn't give us that luxury. PE is just as crucial for making sense of signals in more complex, real-world scenarios.

Imagine you're in a room with a loud, droning hum from a machine. You want to build an ​​Active Noise Control (ANC)​​ system to cancel it. The idea is to play an "anti-hum" through a speaker that perfectly cancels the machine's noise at your ear. To do this, the system's adaptive filter needs to know the "secondary path"—the acoustic path from its speaker to your ear. It learns this by listening to its own output. But here's the catch: the only signal it ever plays is the anti-hum, a single, pure tone. The regressor signal used for learning is therefore just a sinusoid. As we've seen, a single sinusoid is not persistently exciting for a system with more than two parameters. The controller becomes an expert at modeling the room's acoustics at that one single frequency, but it remains utterly ignorant of how any other sound travels. It's a classic case ofPE failure.

So what's the clever solution? While the main controller is busy shouting the anti-hum, we have it simultaneously whisper a very quiet, broadband probe signal—a little bit of white noise, for example. This injected dither is too quiet to be annoying, but it's rich enough to persistently excite the acoustic path, allowing the system to build a full, accurate model and perform robustly. It's a beautiful engineering trick that balances the immediate goal (cancellation) with the long-term need for learning.

Another conundrum arises when a system is operating in a ​​closed-loop​​. Imagine trying to identify the dynamics of a self-driving car while its stability control system is active. The controller's very job is to counteract disturbances and keep the car moving smoothly. In doing so, it suppresses the very variations in the input signal that you need to identify the car's dynamics! The feedback makes the system "lazy" and uninformative. Simply checking if the car's steering input is PE is not enough, because that input is now a slave to the feedback loop and correlated with any disturbances. To break this cycle and restore identifiability, we must introduce an external excitation through the reference signal (the car's target trajectory), a signal that is independent of the system's noise and forces the system to reveal its true nature under command.

The importance of PE extends far beyond just getting accurate parameter estimates. It has profound consequences for how we design controllers and even for how we think about control itself.

Think about ​​robust control​​. We identify a model from data, but because of noise and finite experiment time, we know our estimate $\hat{\theta}$ isn't perfect. We instead define an "uncertainty set" $\Theta$, a small region in the parameter space where the true parameter $\theta_{\star}$ likely resides. The robust controller must work for every possible plant in this set. The size and shape of this uncertainty set depend directly on the quality of our identification experiment. If our input signal was poorly exciting, some directions in the parameter space are poorly determined, and the uncertainty ellipsoid $\Theta$ will be long and bloated in those directions. A controller designed for this large uncertainty must be very conservative—like driving slowly in a thick fog. But if we use a strongly persistently exciting input, we can shrink this uncertainty set considerably. This gives us higher confidence in our model and allows us to design a more aggressive, higher-performance controller. Thus, better PE directly leads to less conservative, better-performing robust control systems.

Even more profoundly, PE provides the foundation for the entire field of ​​data-driven control​​. A stunning result, known as a ​​Willems' Fundamental Lemma​​, tells us something that feels like magic: a single, finite-length data trajectory from an experiment can, by itself, contain enough information to generate every possible behavior of the system over a certain time horizon. What's the catch? What makes an experiment "good enough" for this magic to work? You guessed it. The input signal used in that one experiment must be persistently exciting of a sufficiently high order ($L+n$, to be precise, where $n$ is the system order and $L$ is the trajectory length). This elevates PE from a condition for estimation to a fundamental prerequisite for representing a system's behavior purely from data. Modern methods like ​​Data-enabled Predictive Control (DeePC)​​ are built directly on this principle, allowing us to design controllers from raw data without ever explicitly writing down a state-space model, all thanks to the power of a single, well-excited data log.

So far, we have mostly talked about linear systems. But what about the messy, nonlinear world we actually live in? The principle of PE lives on, though it becomes more subtle and even more profound.

Consider the ​​Extended Kalman Filter (EKF)​​, a workhorse algorithm for tracking everything from spacecraft to your phone's orientation. For a nonlinear system, whether you're getting good information depends not just on the inputs, but on the state of the system itself. You might fly into a region of the state space where your sensors are temporarily "blind" to certain state changes. The PE concept generalizes to a condition called "uniform complete observability." If this condition holds, it guarantees that the filter's uncertainty about the state remains bounded and the estimates will converge (at least locally). If it fails—if the system remains unobservable for too long—the filter's uncertainty can grow without bound. The filter "gets lost," and its estimates can diverge catastrophically, even if the nonlinearities are small. PE, in this guise, is a condition for the stability of learning itself.

Let's end our journey in an perhaps unexpected place: a ​​synthetic biology​​ lab. Scientists are engineering new gene circuits inside living cells, creating biological machines to fight disease or produce chemicals. To understand and control these circuits, they must estimate their unknown biochemical parameters. They face the exact same problem! They must design an input (say, the concentration of an inducer chemical over time) to perturb the cell in an informative way. The idea of PE is directly applicable: the input must be designed to ensure that the sensitivity of the cell's output (e.g., protein fluorescence) to each parameter is sufficiently rich and linearly independent from the others. This is necessary to get a well-conditioned estimation problem. But here, in the heart of nonlinear biology, we are also reminded of the principle's limits. Even with a perfect PE input, the inherent non-convexity of the problem means we might find a wrong answer, and the sheer complexity of a living cell means our simple model is never the full story.

And so we see that Persistency of Excitation is far more than a dry mathematical requirement. It is a deep and unifying principle of scientific inquiry. It teaches us that to learn, we must interact; to understand, we must ask questions. Whether we are trying to identify an electronic circuit, control an aircraft, cancel a sound, or decipher the code of life, the challenge remains the same: we must design our experiments to be rich enough to make the silent system speak its truths.