Sampling Zeros

Controlling the physical world, from the flight of an aircraft to the motion of a robotic arm, presents a fundamental challenge for modern technology. While these systems operate in continuous time, the digital computers that command them think in discrete steps. This gap between the continuous and the discrete necessitates a translation, an approximation that, while essential, is not without consequence. This very act of digital observation and control introduces subtle yet powerful artifacts into a system's behavior, creating hidden dynamics that can limit performance or even lead to instability.

This article delves into one of the most profound of these artifacts: the sampling zero. We will explore how these mathematical phantoms are born from the interface between the digital controller and the physical plant. In the following sections, you will gain a comprehensive understanding of this critical concept. The chapter on ​​Principles and Mechanisms​​ will demystify how discretization, particularly through the common Zero-Order Hold, creates sampling zeros and how a system's intrinsic properties can lead these zeros to become unstable. Subsequently, the chapter on ​​Applications and Interdisciplinary Connections​​ will reveal the tangible, real-world impact of these zeros, from imposing fundamental limits on control performance to necessitating a complete shift in design philosophy toward direct digital methods.

Imagine you are watching a graceful ballet dancer. Her movements are fluid, continuous, a symphony of motion. Now, imagine trying to describe this dance using only a sequence of still photographs taken once every second. You capture her pose at the beginning of the second, in the middle, and at the end. But what happens in between the clicks of the shutter? You lose the subtle accelerations, the flowing transitions, the very essence of the dance. You are left with a series of snapshots, a discrete representation of a continuous reality.

This is the fundamental challenge at the heart of digital control. Our computers, the brains of modern robotics, aerospace vehicles, and automated factories, live in a world of discrete time steps. They think in terms of now, next step, step after next. But the physical world they are tasked with controlling—the flight of an aircraft, the temperature of a chemical reactor, the motion of a robot arm—is continuous. It flows like time itself. To bridge this gap, we must perform an act of translation, and it is in the subtleties of this translation that we discover a fascinating and sometimes treacherous new landscape.

How does a digital controller, which calculates a new command only at discrete ticks of its internal clock (say, every $T$ seconds), command a physical system that needs continuous guidance? The simplest, most common strategy is what we call a ​​Zero-Order Hold (ZOH)​​. It’s a wonderfully straightforward, if somewhat brutish, approach: the controller calculates a command, say "apply 5 volts to the motor," and the ZOH circuit holds that command constant for the entire interval of $T$ seconds, until the controller has a new thought.

Think of it as giving instructions to a sculptor: "For the next minute, push on the clay with a force of 10 newtons. ... OK, now for the next minute, push with 12 newtons." The input to the system isn't a smooth, varying force but a series of steps—a staircase. This staircase is a "lie," an approximation of the ideal, smooth command we might wish to provide. It is a necessary lie, born from the discrete nature of our digital minds, but as with all approximations, it comes with consequences. The act of sampling the world and holding our commands constant introduces its own dynamics, artifacts of our digital interface that are not present in the original, continuous system. The most profound of these artifacts are called ​​sampling zeros​​.

When we analyze a system, we often talk about its poles and zeros. Poles are like the system's natural resonances, its preferred frequencies of vibration. The mapping of continuous-time poles to discrete-time is beautifully simple and intuitive: a pole at location $s=p$ in the continuous world maps directly to a pole at $z = \exp(pT)$ in the discrete world. A stable pole in the left-half of the complex plane ($\mathrm{Re}(p) 0$) maps to a stable pole inside the unit circle ($|z| 1$). It's an elegant correspondence.

One might naively assume that zeros—which influence how a system responds to different input frequencies—would map just as cleanly. But this is not the case. The "lie" of the ZOH gets in the way. The process of holding the input constant fundamentally alters the system's zero structure. In fact, even if a continuous system has no zeros at all, the process of ZOH discretization will almost always create them out of thin air.

Let's consider a simple, well-behaved system, like a car whose acceleration is proportional to how far you press the gas pedal. This system has a "lag" between the input (pedal press) and the output (position), which we call its ​​relative degree​​. For a simple second-order system like a mass on a spring, the relative degree is two. The continuous transfer function might look something like $G(s) = \frac{1}{(s+1)(s+2)}$, which has two poles and no finite zeros. It’s perfectly well-behaved.

But when we discretize this system using a ZOH, something remarkable happens. A finite zero appears in the discrete-time model. For the system above with a sampling time of $T=0.5$, a single sampling zero appears at $z \approx -0.6065$. Where did it come from? It is a ghost in the machine, a mathematical echo of the staircase input. The ZOH holds the input constant, letting the system "coast" for a full sampling period. This coasting, followed by a sudden change, creates an inter-sample ripple. The discrete model, which only sees the snapshots at times $kT$, tries to make sense of this behavior and invents a zero to explain the relationship between the input it gives and the output it sees. For a system with relative degree $r=2$, the ZOH introduces exactly $r-1=1$ sampling zero, and as the sampling period $T$ gets very small, this zero always scurries towards the location $z=-1$ on the edge of the unit circle. A value of $z=-1$ corresponds to a signal that flips its sign at every sample—a high-frequency oscillation, a direct signature of the "jerkiness" introduced by the hold.

For a system with a relative degree of two, this new sampling zero is a curious ripple, but it stays within the bounds of stability. But what happens if the system has more intrinsic lag? What if its relative degree is three or higher? This would be like a large, heavy ship where turning the rudder (input) takes a very long time to affect the ship's heading (output).

Let's take the purest example of a system with a relative degree of three: a triple integrator, with transfer function $G(s) = 1/s^3$. This system is perfectly minimum-phase in continuous time (it has no zeros). Yet, when we apply the ZOH discretization, a truly astonishing thing occurs. The process creates not one, but $r-1=2$ sampling zeros. Their locations are found by solving the simple quadratic equation $z^2 + 4z + 1 = 0$. The roots are $z = -2 \pm \sqrt{3}$.

Let's look at these numbers. One zero is $z \approx -0.268$, which is small and harmless, safely inside the unit circle. But the other is $z \approx -3.732$. This zero is outside the unit circle. This is a ​​non-minimum phase​​ zero, and it is a control engineer's nightmare.

A non-minimum phase zero signifies an initial "wrong-way" response. Imagine pushing a child on a swing. You expect that a forward push will make the swing go forward. But a non-minimum phase system is one where a forward push initially causes the swing to move slightly backward before it moves forward. Attempting to control such a system is like trying to balance a broomstick with a floppy, delayed connection to your hand. It is fundamentally difficult and places severe limitations on the performance you can achieve.

This is a profound and unsettling discovery. The simple, innocent act of sampling a perfectly well-behaved continuous system can, if the system's intrinsic lag (relative degree) is large enough ($r \ge 3$), inevitably create a pathological, unstable feature in its discrete model. This isn't an artifact of a poor calculation or a strange special case; it is a fundamental consequence of the interaction between the ZOH and the system's dynamics. The staircase input, held constant for too long relative to the system's slow response time, causes an overshoot and undershoot between samples that, from the perspective of the discrete controller, looks like a non-causal, wrong-way effect.

Is this a fundamental law of nature we must resign ourselves to? Is every high-relative-degree system doomed to poor performance when controlled digitally? The answer, beautifully, is no. The problem wasn't sampling itself, but the crudeness of our "lie"—the Zero-Order Hold. What if we use a more intelligent lie?

Instead of holding the input constant and creating a staircase, what if we draw a straight line between the command value at one sample and the next? This is called a ​​First-Order Hold (FOH)​​. It generates a series of ramps, a much better approximation of a smooth, continuous signal.

This seemingly small change has dramatic consequences. A First-Order Hold introduces only $r-2$ sampling zeros, one fewer than a ZOH. Let's return to our troublesome plant with relative degree $r=3$. The ZOH created two sampling zeros, one of which was unstable. The FOH, in contrast, creates only $3-2=1$ sampling zero. And it turns out this lone zero is stable, remaining safely inside the unit circle! By choosing a more sophisticated way to fill in the gaps between samples, we have tamed the non-minimum phase beast.

This reveals the true artistry in control engineering. The "pathologies" we encounter are often not unbreakable laws of physics, but consequences of the tools we choose to use. The appearance of sampling zeros, and particularly their journey into the unstable region outside the unit circle, is a beautiful story of the intricate dance between the continuous physical world and our discrete digital approximation of it. Understanding this dance allows us to not only anticipate these challenges but to elegantly design our way around them, turning a potential disaster into a triumph of engineering insight.

We have journeyed through the somewhat abstract world of sampling zeros, seeing how the simple, seemingly innocent act of observing a continuous process at discrete moments in time can conjure these curious mathematical entities. But are they mere ghosts in the machine, phantoms of our equations? Or do they have real, tangible effects on the world of engineering and science? The answer, perhaps surprisingly, is that they are very real indeed. These zeros represent fundamental truths about the interface between the digital and the physical, and understanding them is not just an academic exercise—it is essential for building things that work, from the robots in our factories to the aircraft in our skies.

Let us now explore the practical consequences of sampling zeros, to see where they leave their fingerprints and how they shape our world.

Perhaps the most dramatic consequence of sampling occurs when our choice of sampling time, $T$, conspires with the natural rhythm of the system we are trying to control, leading to a complete loss of authority. Imagine trying to control a child on a swing. If you only look at the swing at the very moment it reaches its peak height, it will always appear motionless to you. If you try to give it a push only at that precise instant, your pushes might not be very effective at changing the amplitude of the swing. You have become "blind" to the system's motion.

This is not just an analogy; it is a precise description of what can happen in a real control system. Consider a simple harmonic oscillator, the mathematical model for everything from a pendulum to an electrical circuit. This system has a natural frequency of oscillation, $\omega$. If we sample this system with a period $T$ that is exactly half the natural period of oscillation ($T = \pi/\omega$), we lose controllability. The discrete-time model we build from our samples will have certain states that our control input simply cannot influence. At these "pathological" sampling rates, the discretization process creates a sampling zero that lands precisely on top of a system pole in the $z$-domain, effectively hiding that dynamic mode from the controller.

The situation can be even more stark. If we choose a sampling period equal to the full natural period of the system ($T = 2\pi/\omega$), the input's effect over one period can average out to exactly zero. The discrete-time input matrix, $B_d$, can become a vector of zeros, meaning our control input has literally no effect on the state at the sampling instants. We have essentially unplugged our controller, not by pulling a wire, but by choosing the wrong frequency at which to look.

Losing control entirely is a catastrophic failure. But more often, sampling zeros impose limits that are more subtle, yet just as fundamental. They don't always cause a complete breakdown, but they place a hard ceiling on the best possible performance we can ever hope to achieve. This is a deep and profound idea in modern control theory.

In a feedback control system, we often have competing objectives. We want our system to track desired commands accurately, which means we want the "sensitivity function," $S(z)$, to be small at low frequencies. We also want it to be insensitive to measurement noise, which often means we want the "complementary sensitivity function," $T(z)$, to be small at high frequencies.

Now, enter the sampling zero. It turns out that for a vast class of physical systems (those with a "relative degree" of two, like a simple mass-spring-damper), the process of sampling with a zero-order hold inevitably creates a non-minimum phase sampling zero that, for fast sampling, is located very close to $z=-1$ on the $z$-plane. This location corresponds to the highest possible frequency in a discrete system—the Nyquist frequency.

A non-minimum phase zero at $z_0$ acts like an unremovable knot in the system dynamics. For any stabilizing controller, it forces an "interpolation constraint." In this case, the zero near $z=-1$ forces the complementary sensitivity to be zero at that point: $T(-1) = 0$. So, we have two constraints: we want good tracking, so $T(z) \approx 1$ for $z$ near $1$ (low frequencies), and we are forced to have $T(z)=0$ for $z$ near $-1$ (high frequencies).

Here is the rub, known as the "waterbed effect." A mathematical theorem, related to the Poisson integral, tells us that the logarithm of the magnitude of $T(z)$ must have a certain average value over the unit circle. If you push the magnitude down in some frequency ranges (like at high frequencies), it must pop up somewhere else. Like pushing down on a waterbed, depressing it in one spot causes it to bulge in another. The result is an unavoidable peak in the magnitude of $|T(e^{j\omega})|$ at intermediate frequencies. This peak can lead to poor robustness and amplification of noise. It is a fundamental performance trade-off imposed not by our design choices, but by the very act of sampling.

This limitation can be seen in a very concrete way in the context of advanced control techniques like $H_\infty$ control. The sampling zero at $z=-1$ forces the sensitivity function to be exactly one, $S(-1)=1$. If our performance specification demands small sensitivity at high frequencies, we run into a direct contradiction. This establishes a hard, quantifiable lower bound on the best possible performance we can achieve, a bound determined by the sampling process itself.

How do these theoretical limits affect the working engineer? An engineer's primary window into a system's frequency behavior is the Bode plot. A sampling zero near $z=-1$ leaves a distinct fingerprint on this plot: it creates a deep magnitude "notch" and a very large, destabilizing phase contribution as the frequency approaches Nyquist.

This realization has forced a revolution in how digital controllers are designed. The old, naive approach was to design a good controller in the continuous-time world and then, as an afterthought, use a mathematical tool like the bilinear transform to "discretize" it for implementation on a computer. But this "design then discretize" philosophy is dangerous. The performance and robustness guarantees of the continuous-time design do not necessarily survive the discretization process, precisely because this process introduces new dynamics like sampling zeros that were not accounted for.

The modern, rigorous approach is "direct digital design." We start by creating an exact discrete-time model of our plant, including the effects of the sampler and hold. This model contains the sampling zeros from the very beginning. We then perform our control design directly in the discrete-time domain, working with, and around, the limitations imposed by the sampling zeros. This has led to the development of sophisticated techniques, from choosing weighting functions in $H_\infty$ loop-shaping that respect the Nyquist limit, to methods like "prewarping" in the bilinear transform that aim to preserve critical features like the crossover frequency. We must confront the ghost in the machine head-on, rather than pretending it isn't there.

What happens if our system is more complex? For systems with a higher relative degree (for example, relative degree three), the situation becomes even more dire. The zero-order hold discretization can create not just a non-minimum phase zero on the unit circle, but one that is unstable—located outside the unit circle. An unstable zero represents an inherent tendency for the system to respond in the "wrong" direction initially, and this dynamic is impossible to remove with standard feedback. When designing an optimal controller like an LQR, this unstable sampling zero means that the performance of the digital controller is fundamentally and permanently worse than what could be achieved with an ideal continuous-time controller, no matter how fast you sample. The act of sampling has actively degraded the system's intrinsic properties.

The influence of sampling zeros extends far beyond the realm of simple linear control systems.

​​Hardware and Uncertainty:​​ Where does the sampling period $T$ come from? It's dictated by a physical device, typically a quartz crystal clock. And no physical device is perfect. Manufacturing tolerances mean that the actual sampling period might lie in a range, say $T \in [T_{nom} \pm \delta T]$. This uncertainty in a physical parameter translates directly into uncertainty in the location of the sampling zero. If we are to design a controller that is robust—that works reliably on every unit coming off the assembly line—we must account for this range of possible zero locations. This provides a beautiful and direct link between hardware limitations and the abstract models of robust control.

​​A Universe of Non-Minimum Phase Behavior:​​ It's also important to realize that sampling is not the only source of non-minimum phase behavior. Many physical systems possess this property inherently. A common example is a system with a significant time delay. Approximating this delay in our models, for instance with a Padé approximation, often introduces a non-minimum phase zero in the continuous-time model itself. When we then sample this system, the mapping $z = \exp(sT)$ faithfully translates the continuous-time right-half-plane zero into a discrete-time zero outside the unit circle. The "original sin" of the continuous system is preserved and passed into the digital domain.

​​The Nonlinear Frontier:​​ Even when we venture into the complex and fascinating world of nonlinear systems, the core ideas remain relevant. The concepts of relative degree and zero dynamics can be extended to nonlinear systems. When we sample a nonlinear system, the resulting discrete-time model also exhibits phenomena analogous to sampling zeros. Under certain conditions, the "sampled zero dynamics" can be seen as a numerical approximation of the underlying continuous-time zero dynamics. Furthermore, the linearization of the sampled nonlinear system will exhibit the same sampling zeros as a linear system with the same relative degree. This shows the remarkable unity of the concept, providing a bridge between our understanding of linear and nonlinear sampled-data systems.

The story of sampling zeros is, in a way, a lesson in both humility and ingenuity. It teaches us that the act of measurement is not a passive observation. By choosing to look at the world through the discrete lens of a digital computer, we actively change the properties of the system we are trying to control. We introduce new dynamics, new behaviors, and new limitations that are not present in the underlying continuous reality.

But this is not a story of defeat. These fundamental limits have spurred decades of creativity and ingenuity in the fields of control, signal processing, and systems theory. They have forced us to move beyond simple models and develop the sophisticated tools of modern digital and robust control. The ghost in the machine, it turns out, has been a formidable, and invaluable, teacher. It reminds us that to successfully command the physical world with our digital tools, we must first listen carefully to the subtle, and sometimes inconvenient, truths it has to tell us.