Bilinear Transformation

In the transition from analog circuits to digital processors, a fundamental challenge emerged: how can we leverage decades of established analog filter and control system design in the new digital realm? A simple translation is insufficient, as the continuous nature of analog signals and the discrete steps of digital computation operate under different mathematical rules. The bilinear transform provides a powerful and elegant bridge between these two worlds, but it is far more than a simple substitution. It is a profound geometric mapping with critical consequences that must be understood to be harnessed effectively.

This article explores the theory and application of the bilinear transform. In the first chapter, ​​Principles and Mechanisms​​, we will dissect the transform's geometric foundations as a Möbius transformation, revealing how it masterfully maps the continuous stability region into its discrete counterpart. We will also confront its most famous side effect—frequency warping—and discover how this apparent distortion is key to its alias-free performance. Following this, the chapter on ​​Applications and Interdisciplinary Connections​​ will showcase how these principles are put into practice. We will see how engineers use prewarping to design high-precision digital filters and how the transform provides a language for digital control systems, ultimately unifying the analysis of stability across both continuous and discrete domains.

After our initial glimpse into the world of the bilinear transform, it's time to roll up our sleeves and look under the hood. What is this transformation, really? Is it just a clever algebraic trick, or is there a deeper, more beautiful structure at play? As we'll see, its power lies in a profound geometric rearrangement of the world of signals and systems, a rearrangement with fascinating and crucial consequences.

At its heart, the bilinear transform is a specific type of ​​Möbius transformation​​, a function from the family $w = \frac{\alpha z + \beta}{\gamma z + \delta}$ that has long fascinated mathematicians. At first glance, this fraction might seem opaque. But we can demystify it by breaking it down into a sequence of simpler, more intuitive actions. Any such transformation can be thought of as a combination of just four fundamental operations: translation (shifting the plane), scaling and rotation (stretching and turning), and the most magical ingredient of all, ​​inversion​​ (turning the plane inside out via $z \mapsto 1/z$). It's like a geometric dance: you slide the floor, flip it over, stretch and turn it, and slide it again.

The true beauty of this dance is what it does to shapes. Möbius transformations have a remarkable property: they map ​​generalized circles​​ to other generalized circles. Now, what's a "generalized circle"? It's simply a term that includes both ordinary circles and straight lines (which you can think of as circles with an infinite radius). This means that if you take any circle or any straight line in the complex plane and apply a bilinear transform to all of its points, the resulting shape will always be another circle or another straight line.

Let's see this in action. Imagine a set of poles in a continuous-time system that all share the same exponential decay rate. These poles lie on a vertical line in the complex $s$-plane, say $\text{Re}(s) = -\sigma_0$. When we apply the bilinear transform, this straight line is bent and curved in just the right way to become a perfect circle in the $z$-plane. Similarly, if we take the poles of a classic Butterworth filter, which lie on a semicircle, the transformation maps them onto another circular arc. This isn't a coincidence; it's a deep geometric truth. The transformation doesn't just shuffle points around; it preserves a fundamental kind of shape.

Now let's turn to the specific form of the transformation that is so vital in engineering, which connects the continuous-time $s$-plane to the discrete-time $z$-plane:

Here, $T$ is the sampling period, a bridge between the two worlds. What grand geometric feat does this particular transformation accomplish?

It performs nothing less than a complete reorganization of the plane. First, it takes the entire infinite imaginary axis of the $s$-plane (where $s=j\Omega$, the home of pure oscillations in continuous time) and maps it perfectly onto the unit circle of the $z$-plane (where $z=e^{j\omega}$, the home of pure oscillations in discrete time).

But the true crown jewel of this transformation is what it does with the rest of the plane. In control theory and signal processing, stability is paramount. For a continuous-time system, stability means all its poles must lie in the ​​left half-plane​​ (LHP), where $\text{Re}(s) 0$. For a discrete-time system, stability means all its poles must lie inside the ​​unit disk​​, where $|z| 1$. The bilinear transform performs the remarkable feat of mapping the entire infinite LHP of the $s$-plane precisely into the finite area of the open unit disk in the $z$-plane.

Let's make this concrete. Suppose we have a stable analog filter with a pole at $s_p = -1 + j$. Using the transform with a sampling period of $T=2$ for simplicity, we find the new pole location:

The magnitude of this new pole is $|z_p| = \sqrt{(-1/5)^2 + (2/5)^2} = \sqrt{5/25} = 1/\sqrt{5}$, which is less than 1. The pole that was in the stable LHP has landed safely inside the stable unit disk. This works every time. A stable analog filter is guaranteed to become a stable digital filter. In the same way, the unstable right half-plane ($\text{Re}(s) > 0$) is mapped to the exterior of the unit disk ($|z| > 1$), so instability is also faithfully preserved. This one-to-one mapping of stability regions is the primary reason for its widespread use. Furthermore, because the transform's coefficients are real, it preserves the conjugate symmetry of poles and zeros, ensuring a real-world system remains real.

So, the transformation gives us a perfect mapping of stability regions. It seems like the ideal dictionary for translating from analog to digital. But there's a fascinating and crucial twist in the tale. The translation isn't entirely literal. While the locations of stability are preserved, the scale of frequency is not.

When we look closely at how the continuous frequency axis ($s=j\Omega$) maps to the discrete frequency circle ($z=e^{j\omega}$), we find this relationship:

This equation, known as the ​​frequency warping​​ relation, is one of the most important consequences of the bilinear transform. What does it mean? It means the relationship between the analog frequency $\Omega$ and the digital frequency $\omega$ is fundamentally nonlinear.

Imagine the infinite analog frequency axis as an infinitely long, stretchy rubber band. The bilinear transform takes this rubber band and squashes it to fit onto the circumference of the unit circle, which represents the digital frequency range from $\omega=-\pi$ to $\omega=\pi$. Near the center of the band (low frequencies, $\Omega \approx 0$), the relationship is almost linear; the band is barely distorted. But as you move toward higher analog frequencies, the band must be compressed more and more dramatically. The entire infinite expanse of high frequencies is squeezed together, getting denser and denser as it approaches the single point $z=-1$ (or $\omega = \pm \pi$) on the unit circle. This point, $z=-1$, becomes the destination for all of infinite frequency. This nonlinear compression is the essence of frequency warping.

This warping might sound like a terrible problem, a flaw in the transformation. But it is also, in a way, its greatest strength. To understand why, we must meet the arch-nemesis of digital signal processing: ​​aliasing​​.

Consider an alternative, more "obvious" way to make a digital filter: simply take samples of the analog filter's impulse response. This method, called "impulse invariance," seems direct, but it has a dire consequence in the frequency domain. The sampling process causes the analog filter's frequency spectrum to be replicated at intervals, creating an infinite sum of overlapping copies. High frequencies from one copy can "fold over" into the low-frequency range of another, masquerading as frequencies they are not. This is aliasing, and it can catastrophically distort your signal.

The bilinear transform, by its very nature, avoids this completely. The frequency warping is a one-to-one mapping. Every single frequency on the infinite analog axis, from zero to infinity, is given its own unique spot on the digital unit circle. There is no folding, no overlapping, and therefore, ​​no aliasing​​. This is a tremendous advantage.

The price we pay for this gift of no aliasing is the warping itself. The magnitude response of our filter, $|H(j\Omega)|$, is mapped perfectly point-for-point to the digital domain. But because the frequency axis underneath it is stretched and squeezed, the shape of the response is distorted. A passband that is 100 Hz wide in the analog domain might become 80 Hz wide in the digital domain, while another 100 Hz-wide stopband at a higher frequency might be compressed into a mere 20 Hz.

So, we have a transformation that preserves stability and eliminates aliasing, but distorts the frequency axis. How do we design a filter that has the exact cutoff frequency we want? We must be clever. We must anticipate the distortion. This clever trick is called ​​prewarping​​.

The logic is simple but powerful. If you want your final digital filter to have a critical feature, like a passband edge at a specific digital frequency $\omega_p$, you cannot simply design the analog prototype to have its edge at the naively scaled frequency $\Omega = \omega_p / T$. The warping will shift it.

Instead, you use the warping formula in reverse to find the analog frequency $\Omega_p$ that will be warped into your desired digital frequency $\omega_p$:

You then design your analog filter using this "prewarped" frequency as its specification. When you apply the bilinear transform, the inevitable frequency warping will bend $\Omega_p$ right back to your target, $\omega_p$. It's like an archer aiming ahead of a moving target to score a direct hit.

This technique is incredibly effective. It doesn't eliminate the warping—the frequency axis is still compressed—but it compensates for it at the most critical frequencies. Even for sophisticated designs like elliptic filters, which have carefully optimized equiripple behavior, this method works beautifully. The magnitude values of the ripples are preserved exactly, and prewarping ensures the band edges land precisely where they need to be, guaranteeing the final digital filter meets its specifications. By understanding the principle of warping, we can turn it from a problem into a predictable feature, allowing us to harness the full, alias-free power of the bilinear transform.

We have seen that the bilinear transform is a remarkable mathematical bridge connecting the continuous world of analog systems to the discrete world of digital computers. It is our looking-glass, a mapping between the complex $s$-plane and the $z$-plane. But what is this bridge for? What can we build with it? To simply say it converts continuous systems to discrete ones is like saying a telescope is just a tube with glass. The real magic lies in what it allows us to see and do. In this chapter, we will explore the profound and practical applications of this transform, journeying from the design of digital audio filters to the fundamental analysis of a robot's stability.

For decades before the digital revolution, engineers and physicists became masters at designing analog filters—circuits built from resistors, capacitors, and inductors that could shape signals with incredible precision. They created a "hall of fame" of filter designs: the smooth and elegant Butterworth filter, the sharp-edged Chebyshev filter, and many others. This was a rich and powerful body of knowledge. When digital processors arrived, a natural question arose: must we reinvent everything from scratch?

The answer, thanks to the bilinear transform, is a resounding no. We can stand on the shoulders of these analog giants. The core idea is to take a proven analog filter design, with its transfer function $H(s)$, and use the bilinear transform to convert it into a digital filter $H(z)$. But this is not a simple, brute-force substitution. A naive approach leads to a curious and fascinating problem: ​​frequency warping​​.

Imagine looking at your reflection in a funhouse mirror. Your head might look squashed and your legs stretched. The bilinear transform acts as a kind of mathematical funhouse mirror for the frequency axis. It takes the infinite frequency axis of the analog world (the imaginary axis in the $s$-plane) and non-linearly compresses it onto the finite boundary of the unit circle in the $z$-plane. The relationship we discovered earlier, $\Omega = \frac{2}{T} \tan(\frac{\omega}{2})$, tells the whole story. An analog frequency $\Omega$ does not map to an equal digital frequency $\omega$. This warping means that if we take a carefully designed analog filter with a cutoff frequency of, say, 1000 Hz, its digital counterpart obtained by a simple transformation will not have its cutoff at the corresponding digital frequency. Its critical features will be shifted.

So, what does a clever engineer do? They don't try to fix the mirror; they change what they put in front of it! This ingenious technique is called ​​frequency pre-warping​​. The design process becomes a beautiful three-step dance:

1. Start with the desired digital filter specifications, for example, a cutoff frequency of $\omega_d$.
2. Use the inverse of the warping formula to calculate the "pre-warped" analog frequency $\Omega_p$ that will be warped to your desired $\omega_d$ by the transform. This is like figuring out what distorted shape you need to hold up to the funhouse mirror to make your reflection look normal.
3. Design a standard analog filter (like a Butterworth or Chebyshev) using this pre-warped frequency $\Omega_p$, and then apply the bilinear transform.

The result is extraordinary. The distortion of the pre-warped analog filter perfectly cancels the distortion of the bilinear transform, and the final digital filter has its critical frequencies located exactly where you want them. This procedure is so precise that the accuracy of this frequency mapping is independent of the initial analog prototype's characteristics, a testament to the mathematical certainty of the method. This elegant combination of foresight and transformation allows us to translate the entire library of classical analog filter designs into the digital domain, creating everything from high-fidelity audio equalizers to noise-reduction filters in medical imaging equipment.

Beyond shaping signals, the bilinear transform provides a language for commanding digital systems to perform actions that are inherently continuous. Consider one of the most fundamental operations in calculus and control theory: integration. An integrator is a key component in the ubiquitous PID (Proportional-Integral-Derivative) controller, the workhorse of industrial automation. In the analog world, its transfer function is simply $C(s) = \frac{1}{s}$.

How do we instruct a microprocessor, which thinks in discrete steps, to "integrate"? The bilinear transform gives a direct and elegant answer. By applying the substitution $s = \frac{2}{T}\frac{z-1}{z+1}$, the abstract analog integrator $C(s) = \frac{1}{s}$ is reborn as a concrete digital algorithm, a pulse transfer function $D(z)$ that can be implemented in a few lines of code. Suddenly, a digital brain can be taught to think in calculus.

However, this translation is not without its subtleties. The digital mimic is not a perfect twin of its analog ancestor. Just as frequency is warped, other performance characteristics are altered. For instance, a key metric in control systems is the steady-state error, which tells us how accurately a system can follow a command. This is quantified by error constants, such as the static acceleration error constant, $K_a$. When an analog controller is converted to its digital form via the bilinear transform, the value of this constant changes in a predictable way. Similarly, the "quality factor" or $Q$ of a filter, which measures its sharpness or selectivity, is also transformed.

This is not a flaw, but a crucial insight. It reminds us that we are creating an equivalent, not an identical, system. The beauty of the transform is that it allows us not only to perform the translation but also to precisely calculate the consequences, enabling engineers to design digital control systems that meet exacting performance specifications in the real world.

Perhaps the most profound application of the bilinear transform is not in design, but in analysis. Consider this question: is a given digital system stable? Will a robot arm settle smoothly, or will it oscillate out of control? Will a digital pilot-assist system correct for turbulence, or will it amplify it catastrophically?

For a discrete-time system, the mathematical condition for stability is that all the poles of its transfer function $H(z)$ must lie inside the unit circle in the $z$-plane. Checking this condition can be algebraically daunting for a high-order system.

Now, let's look at the analog world. The condition for stability there is that all poles of $H(s)$ must lie in the left half of the $s$-plane. Over the years, mathematicians have developed powerful and straightforward tools for checking this, most famously the Routh-Hurwitz stability criterion, which requires only simple arithmetic on the coefficients of the system's characteristic polynomial.

Here is where the bilinear transform becomes a veritable Rosetta Stone. We know that it maps the entire left half of the $s$-plane to the interior of the unit circle in the $z$-plane. This is a two-way street: it also maps the interior of the unit circle back to the left-half plane. This means we can take a difficult stability problem from the discrete $z$-domain, apply the bilinear transform to it, and translate it into an equivalent problem in the continuous $s$-domain. Once there, we can use the simple and powerful Routh-Hurwitz test to find the answer. The transform allows us to choose our battlefield, moving the problem to a domain where we have superior tools.

Why does this "Rosetta Stone" work so perfectly? The reason is deep and beautiful. The bilinear transform is not just an arbitrary algebraic substitution; it is a ​​conformal map​​. In geometry, this means it is a transformation that preserves angles locally. For our purposes, its most important property is that it maps the boundary of a region to the boundary of the transformed region, and the interior to the interior.

This is the secret behind its power. It provides a mathematically rigorous one-to-one mapping between the stability region of the continuous world (the left-half plane) and the stability region of the discrete world (the unit circle). Therefore, a stable analog system, with all its poles in the left-half $s$-plane, is guaranteed to transform into a stable digital system, with all its poles inside the unit $z$-circle. Stability is an intrinsic property that is preserved by the transformation.

This insight scales to even more abstract concepts in control theory. The root locus, a powerful graphical method for analyzing stability, shows how a system's poles move as a gain parameter is varied. Under a bilinear mapping, the geometric character of the locus changes—straight-line asymptotes can transform into circles—but its fundamental topology and its relationship to the stability boundary are preserved. This explains why the remarkable robustness guarantees of advanced control techniques like the Linear Quadratic Regulator (LQR) are maintained when a continuous design is implemented digitally using the transform.

Ultimately, the bilinear transform is more than a clever tool for engineers. It is a window into the unity of the principles governing dynamic systems. It shows us that the concepts of filtering, control, and stability are not fundamentally different in the analog and digital realms. The language may change from $s$ to $z$, and the landscape may warp from a half-plane to a circle, but the underlying laws of cause and effect, of stability and response, remain beautifully and reassuringly consistent.