Cohen's Class

In the world of signal processing, obtaining a clear picture of a signal's energy in both time and frequency simultaneously is a fundamental but elusive goal. While the Fourier transform reveals a signal's frequency content, it obscures its temporal evolution. This gap in understanding limits our ability to analyze dynamic signals like human speech, radar echoes, or biological rhythms. The central problem is how to design a representation that offers sharp detail without introducing misleading artifacts.

This article delves into Cohen's Class, a powerful and elegant theoretical framework that unifies a vast family of time-frequency distributions. It provides a master recipe for understanding and designing these tools. The following chapters will first uncover the core principles and mechanisms of this class, exploring the ideal but problematic Wigner-Ville Distribution and its ghostly "cross-terms". We will then see how these tools are applied and connected across disciplines, revealing how engineers and scientists deliberately choose a representation to tame these artifacts and extract meaningful information from complex, changing signals.

Imagine you want to create the ultimate musical score, one that shows not just which notes are played, but precisely when they are played and how their pitch might slide and change over time. The Fourier transform gives us the notes (the frequencies), but it scrambles all the timing information. The original time-domain signal gives us the timing, but the notes are all mixed up. The quest for a perfect time-frequency representation is a central story in signal processing, a story of profound beauty, frustrating paradoxes, and elegant compromises. At the heart of this story lies a family of mathematical tools known as ​​Cohen's Class​​.

Let's start with the most ambitious and, in a way, the most beautiful member of this family: the ​​Wigner-Ville Distribution (WVD)​​. If we were to design a time-frequency representation from pure intuition, we might want it to have some basic properties. If a signal is a single, instantaneous "clap" at time zero—a Dirac delta function, $x(t) = \delta(t)$—we'd want our representation to show energy at that exact moment, $t=0$, spread out over all frequencies. And that's exactly what the WVD does; it gives $W_x(t,f) = \delta(t)$. Conversely, if the signal is a pure, eternal musical note—a complex exponential, $x(t) = \exp(j2\pi f_0 t)$—we'd want to see energy only at that frequency, $f_0$, for all time. The WVD delivers again, yielding $W_x(t,f) = \delta(f-f_0)$. It perfectly localizes these fundamental building blocks of signals.

This remarkable ability isn't limited to static events. Consider a signal whose frequency is constantly changing, like the sound of a bird's chirp or the Doppler shift from a passing ambulance. A simple model for this is a linear chirp, $x(t) = \exp(j\pi\alpha t^2)$, whose instantaneous frequency is $f(t) = \alpha t$. The WVD performs a miracle: its representation is $W_x(t,f) = \delta(f-\alpha t)$, an infinitely sharp line that perfectly traces the signal's changing frequency across the time-frequency plane. In these ideal cases, the WVD is not just good; it's perfect. It seems to have surmounted the famous uncertainty principle, which limits the simultaneous resolution of linear methods. But how? The answer lies in its structure, and this is where our story takes a fascinating turn.

The WVD's power comes from a property called ​​bilinearity​​. Unlike the linear Fourier transform where the transform of a sum is the sum of the transforms, the WVD of a sum of two signals, $s(t) = s_1(t) + s_2(t)$, is not just the sum of their individual WVDs. Instead, we get:

The first two terms, $W_{s_1}$ and $W_{s_2}$, are the "auto-terms" we expect. They represent the energy of each component. But the third term, the ​​cross-term​​, is something entirely new—a "ghost" in the machine generated by the interaction between the two components.

So, where do these ghosts appear? Astonishingly, they tend to appear halfway between the two true components. If you have two signals centered at $(t_1, f_1)$ and $(t_2, f_2)$, the cross-term will be located near their midpoint, $(\frac{t_1+t_2}{2}, \frac{f_1+f_2}{2})$. This means the WVD can show significant energy in a time-frequency region where neither signal actually exists! To make matters worse, this ghost term isn't a simple blob of energy; it oscillates wildly. A crucial insight is that the rate of this oscillation is directly proportional to the separation between the components. If two tones are separated by $\Delta f$ in frequency, their cross-term oscillates in time with a period of $1/|\Delta f|$. If two impulses are separated by $\Delta t$ in time, their cross-term oscillates in frequency with a period of $1/|\Delta t|$. The further apart the signals, the more frenetically their ghost-term oscillates.

This oscillatory nature means the cross-terms can take on negative values. This is a fatal blow to the idea of interpreting the WVD as a true energy distribution, as energy can't be negative. These artifacts, born from the same bilinearity that gives the WVD its superb resolution, make the raw WVD almost unusable for analyzing complex signals with multiple components.

It turns out the WVD is not an isolated curiosity but the patriarch of a vast family of time-frequency distributions, all united under the banner of ​​Cohen's Class​​. This framework provides a recipe book for cooking up any time-frequency distribution you can imagine, and in doing so, it reveals how to manage the trade-off between resolution and those pesky cross-terms.

The secret ingredient is the ​​Ambiguity Function​​, $A_x(\tau, \nu)$. It's a bit of a strange beast, but intuitively, it measures the similarity between a signal and a time-shifted and frequency-shifted version of itself. The WVD is simply the two-dimensional Fourier transform of the ambiguity function.

Cohen's great insight was that you could generate other distributions by inserting a "weighting function," or ​​kernel​​, $\Phi(\tau, \nu)$, into this relationship:

The WVD corresponds to the simplest possible kernel: $\Phi(\tau, \nu) = 1$. Every other choice of kernel gives a different distribution with different properties. The kernel acts like a filter in the "ambiguity domain." The auto-terms of a signal tend to have their ambiguity function content clustered around the origin $(\tau=0, \nu=0)$. The cross-terms, however, are located far away from the origin. This gives us a brilliant strategy: if we design a kernel $\Phi(\tau, \nu)$ that is large at the origin but small everywhere else (a 2D low-pass filter), we can suppress the cross-terms!

This brings us to the most popular time-frequency tool of all: the ​​spectrogram​​. A spectrogram is built from the ​​Short-Time Fourier Transform (STFT)​​, which looks at the signal through a small time window. It is indeed a member of Cohen's class. And what is its kernel? It's the ambiguity function of the window itself, $A_g(\tau, \nu)$.

Since any well-behaved window function is concentrated in time and frequency, its ambiguity function will be concentrated around the origin. It's a natural low-pass filter! This is the fundamental reason why spectrograms are mostly free of the wild, non-local cross-terms that plague the WVD. The interference that does remain is confined only to regions where the individual signal components actually overlap in the time-frequency plane.

But there's no free lunch. The kernel that filters out the cross-terms also acts on the auto-terms. In the time-frequency domain, this multiplication by a kernel corresponds to a convolution, or a "smoothing". The spectrogram is, in fact, the WVD of the signal smoothed by the WVD of the window function. This smoothing blurs the auto-terms, degrading the perfect resolution we admired in the WVD. The width of the window dictates the nature of this blur: a short window gives good time resolution but poor frequency resolution, while a long window gives good frequency resolution but poor time resolution. This is the ​​Heisenberg-Gabor uncertainty principle​​ rearing its head, imposed by the choice of window.

This reveals the central trade-off of time-frequency analysis: ​​resolution versus cross-term suppression​​. The WVD chooses perfect resolution at the cost of nightmarish cross-terms. The spectrogram chooses to suppress cross-terms at the cost of resolution. The vast majority of other distributions in Cohen's class, like the Choi-Williams or Born-Jordan distributions, are simply different attempts to find a "sweet spot" in this fundamental trade-off, using cleverly designed kernels to walk the tightrope between clarity and artifacts. Even the discretization of these methods introduces its own set of challenges, such as aliasing that can cause the frequency axis to unexpectedly repeat itself.

So far, we have talked about clean, deterministic signals. What about the chaotic reality of random noise? Here, the framework reveals another layer of elegance. If our signal is a deterministic component plus random, stationary noise, $x(t)=s(t)+w(t)$, the cross-terms between the signal and noise average out to zero. The expected or average WVD beautifully separates into two parts: the WVD of the deterministic signal, plus the power spectral density (PSD) of the noise. This provides a powerful link between the time-frequency world and the statistical world of the Wiener-Khinchin theorem.

However, in any single measurement, the noise doesn't just add a background floor; it also creates a fluctuating, variance-filled mess that can obscure the signal. The smoothing inherent in distributions like the spectrogram not only suppresses deterministic cross-terms but also averages out these noisy fluctuations, reducing the variance and yielding a more stable picture. This is yet another practical reason why, despite the WVD's theoretical perfection, we so often turn to its smoothed-out cousins in the real world. The quest for the perfect picture continues, but armed with the insights of Cohen's class, we can now choose our tools wisely, understanding the inherent beauty, the unavoidable ghosts, and the elegant compromises that define the art of seeing signals in time and frequency.

Now that we’ve tinkered with the beautiful machinery of Cohen’s class, you might be wondering, "What is all this for?" It's a fair question. Is this just a gallery of mathematical curiosities, or is it a workshop full of tools for real-world discovery? The answer, I hope to convince you, is emphatically the latter. The unified framework of Cohen’s class isn't just elegant; it's immensely practical. It's something of a Swiss Army knife for anyone who wants to understand a signal that changes in time. From radar engineers and communication experts to biologists and astrophysicists, the ability to see a signal's frequency content evolve is a kind of superpower.

In this chapter, we're going on a safari to see these tools in their natural habitats. We will see how choosing the right kernel function is an art, a principled design choice that allows us to peer through the fog of complex data and uncover the hidden truths within.

Imagine you are looking at a serene lake at night, and you see the reflection of two bright lanterns. Besides the two clear reflections, you might also see shimmering, ghostly patterns of light on the water's surface between them. These patterns aren't cast by a third lantern; they are interference patterns created by the waves from the two real sources.

The simplest and perhaps most fundamental time-frequency representation, the Wigner-Ville Distribution (WVD), suffers from a similar problem. While it gives an incredibly sharp view of a signal's energy, when a signal has multiple components—like two radio broadcasts, two notes from a flute, or two stars in a binary system—the WVD produces "cross-terms" or "ghosts." These are phantom features in the time-frequency plane that don't correspond to any real component but arise from the mathematical interference between the true ones.

This is where the power of Cohen’s class shines. It tells us we don't have to live with these ghosts. We can design a kernel to suppress them. A wonderful example is the Choi-Williams Distribution (CWD). In the Ambiguity Function domain—the "design space" for our time-frequency tools—the CWD kernel acts like a specialized filter. Its shape is a function like $\Phi_{CW}(\nu, \tau) = \exp(-(\nu \tau)^2 / \sigma)$. What does this mean intuitively? Auto-terms of a signal live near the axes of the ambiguity plane, while cross-terms tend to pop up far from both axes. The Choi-Williams kernel is designed to be large (equal to 1) along the axes, preserving the true signal components, but to fall off rapidly as we move away from them, effectively dimming or erasing the ghostly cross-terms.

Of course, there is no free lunch in physics or in signal processing. This filtering comes at a price. By smoothing out the ambiguity function to kill the cross-terms, we might also slightly blur the auto-terms, reducing our resolution. This introduces the central engineering dilemma of time-frequency analysis: the trade-off between cross-term suppression and auto-term resolution. Cohen’s class allows us to manage this trade-off explicitly. We can pose a formal design problem: suppose we need to reduce a cross-term at a certain location in the ambiguity plane by a factor of $\varepsilon$, but we cannot tolerate broadening our true signal component by more than a certain amount. The framework allows us to find the optimal kernel parameter that walks this tightrope, achieving just enough suppression without sacrificing too much resolution.

One of the most profound insights from Cohen's class is that many seemingly different time-frequency methods are, in fact, close relatives. Two of the most famous representations are the Wigner-Ville Distribution and the Spectrogram. The WVD, as we've seen, provides perfect resolution for certain signals but is plagued by cross-terms. The Spectrogram, which you might have encountered in audio software, is guaranteed to be non-negative (no weird negative energy values!) and has no cross-terms, but it pays for this pleasantness with fundamentally limited resolution, governed by Heisenberg's uncertainty principle.

For decades, these two were seen as distinct choices. But Cohen's class reveals them to be two ends of a continuous spectrum. We can design a family of kernels, parameterized by a single knob $\lambda \in [0,1]$, that smoothly interpolates between the WVD (at $\lambda=0$) and the Spectrogram (at $\lambda=1$). Turning this knob is like adjusting the focus on a strange new microscope. At one end, the image is perfectly sharp but full of interference artifacts. At the other end, the artifacts are gone, but the image is blurry. The magic is that we can choose any setting in between, tailoring our view to the specific signal and the specific question we are asking.

This is not just an academic exercise. In a real-world scenario, an engineer might be faced with a signal containing two components that are close together in both time and frequency. A spectrogram might fail entirely, its inherent blurring merging the two components into a single, unresolvable blob. A smoothed WVD, like the Choi-Williams distribution, can be tuned to provide just enough cross-term suppression while retaining the high resolution needed to see the two components as distinct entities. The "best" tool is not absolute; it depends on the task at hand.

Many signals in nature and technology don't sit at a constant frequency. Think of a bird's chirp, the Doppler-shifted signal from a speeding car, or the way data is encoded in an FM radio broadcast. These are all examples of modulated signals, whose frequency or amplitude changes over time. A primary goal of time-frequency analysis is to track these changes, to extract the signal's instantaneous frequency (IF).

Here again, Cohen's class provides the perfect tool. If we first transform our real-valued signal $x(t) = a(t)\cos(\phi(t))$ into its "analytic" counterpart, $x_a(t) = a(t)\exp(j\phi(t))$, something wonderful happens. The Wigner-Ville distribution of this analytic signal, $W_{x_a}(t,f)$, becomes remarkably simple for many signals of interest. To a very good approximation, it traces a sharp ridge in the time-frequency plane right along the path of the instantaneous frequency, $f = f_i(t) = \frac{1}{2\pi}\phi'(t)$. All the signal's energy is concentrated along this "tune." This makes the WVD an exceptional tool for IF estimation. Using the analytic signal also has the tidy benefit of eliminating the distracting "mirror image" of the frequency track at negative frequencies.

For a signal whose frequency changes linearly with time—a linear "chirp"—this works perfectly. The WVD ridge is a straight line that exactly matches the IF. But what if the frequency changes in a more complex, non-linear way? In this case, even the WVD, when smoothed by a simple kernel to handle noise or multiple components, can be "fooled." The smoothing process can introduce a bias, causing the estimated ridge to deviate slightly from the true IF, typically underestimating its curvature. The choice of kernel becomes even more crucial and subtle, leading to advanced designs that can adapt to the signal's local structure.

The reach of Cohen's class extends far beyond visualizing clean, deterministic signals. It provides a robust foundation for tackling problems at the intersection of signal processing and statistics.

Real-world signals are always corrupted by noise. A radio signal has static; a radar echo is buried in clutter. A key question is: how does noise affect our time-frequency analysis, and can we design distributions that are robust to it? Using the Cohen's class framework, we can perform a statistical analysis of an IF estimator's performance in the presence of noise. We can calculate its mean-squared error and then optimize the kernel's shape to minimize this error. This analysis reveals a striking result: the "pure" WVD, which is so perfect for clean chirps, is statistically a very poor choice in noise, exhibiting huge variance. A carefully smoothed distribution, such as a Spectrogram or a Smoothed Pseudo Wigner-Ville Distribution (SPWVD), provides a much more stable and reliable estimate by optimally balancing bias and variance.

Another fascinating application lies in the analysis of cyclostationary signals. Unlike the "stationary" noise from a resistor, which looks statistically the same at all times, many man-made and biological signals have statistical properties (like their autocorrelation) that repeat periodically. This "hidden rhythm" is a feature of everything from digitally modulated communication signals to the firing patterns of neurons. The tools of Cohen's class provide a powerful pathway to this domain. By performing a further Fourier analysis on a time-frequency distribution, we can estimate a quantity called the spectral correlation function. This function is the cornerstone of cyclostationary signal processing, and TFDs offer a practical and insightful way to compute it, connecting two major fields of signal analysis.

Our journey has shown that Cohen’s class is much more than a catalog of distributions. It is a language, a unified philosophy for understanding time-varying signals. It teaches us that the choice of a time-frequency representation is not arbitrary; it is a deliberate act of engineering design. The kernel, $\Phi(\tau, \nu)$, is not just a mathematical formula. It is the embodiment of our assumptions about the signal. It is an epistemic constraint—a statement about what features we wish to see and what we are willing to discard as interference. Choosing a kernel is like building a specific lens to look at the world: one lens for separating the stars in a dense cluster, another for tracking a fast-moving comet against a noisy background. By providing the principles to design these lenses, Cohen's class gives us a deeper, clearer, and more powerful way to listen to the symphony of the universe.