Signal Representation

The art of describing the world is central to science and engineering. Just as a painting can be described speck by speck or by its dominant colors, a signal—be it sound, light, or voltage—can be represented in different languages that reveal distinct aspects of its nature. Signal representation is the study of these languages, providing the tools to transform complex, raw data into a form that is structured, insightful, and computationally simple. The core challenge this addresses is how to choose the right "lens" to view a signal, unlocking hidden patterns that are invisible in its original form. This article will guide you through this foundational concept in signal processing.

First, in "Principles and Mechanisms," we will explore the fundamental building blocks of signals. We will contrast the "atomic" view of a signal built from instantaneous impulses with the "harmonic" view of a signal composed of pure oscillations. We will see how complex numbers elegantly unify these concepts and learn about the profound trade-offs between time and frequency, as captured by the Nyquist-Shannon sampling theorem. Then, in "Applications and Interdisciplinary Connections," we will witness these theories in action. We will discover how choices in representation enable everything from high-fidelity digital audio and modern wireless communications to powerful data compression techniques, revealing how abstract mathematics forms the bedrock of our digital world.

Imagine you want to describe a complex painting. You could go about it in two ways. You could create a fantastically detailed map, listing the exact color and location of every single speck of paint on the canvas. Or, you could describe it in broad strokes, noting that it's made of 50% sky blue, 30% grass green, and 20% sunshine yellow, with various shades and textures. Both descriptions can be complete, yet they tell very different stories about the painting.

The world of signals—be it the sound of a violin, the voltage in a circuit, or the radio waves carrying your Wi-Fi signal—can be described in similarly different, yet equally powerful, ways. The art of signal representation is about choosing the right language to reveal a signal's hidden structure and make complex problems astonishingly simple.

Let's start with the first approach: describing the painting speck by speck. What is the most fundamental, indivisible "speck" of a signal? In the continuous world, it's an idea of pure genius: the ​​Dirac delta function​​, or ​​impulse​​, denoted by $\delta(t)$. You can think of it as an infinitely tall, infinitely narrow spike right at $t=0$, whose total area is exactly one. It’s a strange beast, but its one magical property is all that matters. It has the power to sift through a signal and pick out its value at a single point in time. This is called the ​​sifting property​​:

$\int_{-\infty}^{\infty} x(\tau) \delta(t - \tau) \, d\tau = x(t)$

What does this equation truly say? It reveals that any well-behaved signal, $x(t)$, can be thought of as a seamless sum (an integral) of an infinite number of impulses. Each impulse, $\delta(t - \tau)$, is located at a specific time $\tau$ and is weighted by the signal's value at that very instant, $x(\tau)$. So, to build a ramp signal $r(t) = t \cdot u(t)$ (where $u(t)$ is the step function that is zero for $t \lt 0$ and one otherwise), we simply need to tell the integral to use the value of the ramp itself as the weighting function. The integral then dutifully assembles the signal, instant by instant, from these weighted impulses. Even a simple operation like shifting the signal by $t_0$ to get $x(t-t_0)$ translates directly into using a shifted weighting function, $x(\tau - t_0)$, inside the integral. This representation is powerful because it's universal; it can describe any signal by its point-by-point construction.

In the digital world, this "atomic" view is even more intuitive. A discrete-time signal is just a sequence of numbers, a list of values at integer time steps. Here, the unit impulse $\delta[n]$ is simply a 1 at $n=0$ and 0 everywhere else. Any discrete signal $x[n]$ can be written as a sum of scaled and shifted impulses. For a sparse signal that is non-zero only at a few points, say with values $A_k$ at times $n_k$, its representation is simply:

$x[n]=\sum_{k=1}^{M} A_{k}\,\delta[n-n_{k}]$

This isn't just a notational trick. If you want to calculate the total energy of this signal, which is the sum of its squared values, this representation makes it trivial. The impulses are "orthogonal"—a shifted impulse $\delta[n-n_k]$ has no overlap with $\delta[n-n_\ell]$ if $k \neq \ell$. This means all the cross-terms vanish when you square the sum, and the total energy beautifully simplifies to just the sum of the squares of the individual measurements, $\sum_{k=1}^{M} A_{k}^{2}$.

Now for the second approach: describing the painting by its constituent colors. In the world of signals, the primary "colors" are sinusoids—pure, endless oscillations like the hum of a tuning fork. Many signals we encounter, especially in electronics and physics, are either pure sinusoids or can be built from them.

However, working directly with trigonometric functions like sines and cosines is notoriously clumsy. Adding two shifted cosine waves requires wrestling with cumbersome trig identities. There must be a better way. And there is, thanks to one of the most beautiful equations in all of mathematics, Euler's formula:

$e^{j\theta} = \cos(\theta) + j\sin(\theta)$

This is the Rosetta Stone connecting oscillations and complex numbers. It tells us that the cosine function is simply the real part of a rotating vector, $e^{j\theta}$, in the complex plane. A signal like $x(t) = A \cos(\omega t + \phi)$ can be seen as the "shadow" of a vector of length $A$ rotating with angular velocity $\omega$, starting at an angle $\phi$. To simplify our lives, we can forget the rotation and the shadow-casting for a moment and just capture the two essential pieces of information—amplitude $A$ and starting phase $\phi$—in a single complex number. This static snapshot is the ​​phasor​​: $X = A e^{j\phi}$.

The beauty of this is that it turns calculus into algebra. Let's see it in action. To convert a time-domain signal like $v(t) = 170\sin(120\pi t + \pi/6)$ into a phasor, we first use the identity $\sin(x) = \cos(x - \pi/2)$ to put it into our standard cosine form. This yields $v(t) = 170\cos(120\pi t - \pi/3)$. From here, we can simply read off the amplitude and phase to get the phasor $\mathbf{V} = 170\exp(-j\pi/3)$.

The real power of phasors shines when signals interact. If you add two signals, like $x(t) = A \cos(\omega_0 t)$ and $y(t) = A \sin(\omega_0 t)$, you simply add their phasors as if they were vectors. The phasor for $x(t)$ is $A$, and the phasor for $y(t)$ is $A e^{-j\pi/2}$ (or $-jA$). Their sum is $A(1-j)$, which in polar form is $\sqrt{2}A e^{-j\pi/4}$. This immediately tells us the resulting signal is $\sqrt{2}A \cos(\omega_0 t - \pi/4)$—a result that is much more tedious to derive using trigonometry.

Furthermore, when a sinusoidal signal passes through a linear time-invariant (LTI) system, like an electrical circuit, the effect of the system in steady-state is just to multiply the input phasor by a complex number—the system's frequency response. If a system doubles the amplitude and adds a phase lead of $2\pi/5$, its effect is captured by the phasor $2e^{j2\pi/5}$. Finding the output signal becomes a simple multiplication of complex numbers, a vastly simpler task than solving differential equations.

Phasors are wonderful for pure tones, but what about more complex sounds, like a musical chord or a square wave? The French mathematician Jean-Baptiste Joseph Fourier showed that any periodic signal—any signal that repeats itself over and over—can be represented as a sum of sinusoids. This sum, the ​​Fourier Series​​, is like a recipe for the signal. It consists of a constant (DC) component and a series of sinusoids whose frequencies are integer multiples of the signal's fundamental frequency.

Using the complex exponential form, the recipe looks elegant: $x(t)=\sum_{k=-\infty}^{\infty} a_{k}\exp(j k \omega_{0} t)$ The complex numbers $a_k$ are the ​​Fourier coefficients​​. They are the heart of the representation, telling us the amplitude and phase of each harmonic component. For example, if we are told that a real signal's only non-zero coefficients are $a_0 = 1$, $a_2 = 2$, and $a_{-2} = 2$, we can immediately reconstruct the signal. The $a_0$ term is the DC offset. The $a_2$ and $a_{-2}$ terms combine to form a single cosine wave at twice the fundamental frequency. The signal is simply $x(t) = 1 + 4\cos(2\omega_0 t)$. The entire complexity of the original waveform is perfectly captured in a short list of numbers.

How do we apply these ideas to modern communications, like a radio signal carrying a voice message? Here, we have a low-frequency information signal (the voice) riding on a high-frequency carrier wave. It would be inefficient to analyze the fast carrier oscillations every time we just want to know about the slower message.

The solution is to use the ​​analytic signal​​, $z_x(t)$. The real-world signal we transmit, $x(t)$, is just the real part of this more abstract complex signal. The analytic signal itself is factored into two parts: a fast-spinning carrier $e^{j\omega_c t}$ and a slow-moving ​​complex envelope​​ $\tilde{x}(t)$.

$z_x(t) = \tilde{x}(t)e^{j\omega_c t}$

The magic is that all the useful information is now contained in the complex envelope. This envelope $\tilde{x}(t)$ is itself a complex signal, which can be broken down into its real and imaginary parts: $\tilde{x}(t) = x_i(t) + jx_q(t)$. These are called the ​​in-phase ($x_i$)​​ and ​​quadrature ($x_q$)​​ components. They represent two independent channels of information that can be carried on a single carrier wave. For a given analytic signal, we can identify these components by simply isolating the complex envelope and finding its real and imaginary parts. This very principle is the engine behind modern high-speed data transmission in Wi-Fi, 4G, and 5G.

Finally, we arrive at one of the deepest and most practical questions in all of signal processing: can we perfectly capture a continuous, analog signal using a finite number of discrete samples? The ​​Nyquist-Shannon sampling theorem​​ gives us the answer, and it's a conditional "yes". It states that perfect reconstruction is possible, but only if two conditions are met: the signal must be ​​band-limited​​ (its frequency content must be zero above some maximum frequency), and the sampling rate must be at least twice this maximum frequency.

This brings us to a beautiful paradox. Consider one of the simplest-looking signals imaginable: a perfect rectangular pulse. It's on for a bit, then it's off. Nothing could seem more straightforward. Yet, if we sample this signal, even at an absurdly high rate, and try to reconstruct it with a perfect filter, we will fail. The reconstructed signal will always have ripples and overshoots near the sharp edges.

Why? Because the rectangular pulse is not band-limited. To create those perfectly sharp, instantaneous edges in the time domain, you need an infinite range of frequencies in the frequency domain. The Fourier transform of a rectangle is a sinc function, which ripples out to infinity. When you sample this signal, no matter how fast, the infinite tails of its frequency spectrum get "folded" back into the main part of the spectrum. This spectral overlap, called ​​aliasing​​, is an incurable corruption. The high frequencies, which are essential for creating the sharp edges, masquerade as low frequencies, and the original information is lost forever.

This reveals a profound truth about the nature of signals, a kind of uncertainty principle. A signal cannot be perfectly confined in both the time domain and the frequency domain simultaneously. The sharper and more localized a signal is in time, the more spread out and infinite it must be in frequency. This trade-off is not a limitation of our technology; it is a fundamental property of the universe, woven into the very fabric of how we describe change and information.

In our journey so far, we have learned the basic grammar for the language of signals. We have seen how to describe them, manipulate them, and view them in different domains. But this is like learning the rules of chess; the real excitement comes from seeing the game played by masters. Now, we shall explore the poetry of signal representation. We will see how these mathematical ideas are not mere abstractions but are the very blueprints for our digital world, the tools we use to capture, transmit, and understand the universe of information around us. This is where the mathematics breathes, where abstract concepts become tangible realities like crystal-clear audio, instant global communication, and sharp medical images.

Let's start with the most fundamental challenge of the digital age: how do you take a piece of the real world, like the rich, continuous pressure wave of a sound, and store it in the rigidly discrete world of a computer? The world is continuous; a computer's memory is a grid of finite bits. To get from one to the other, we must draw a map, and every map involves choices and compromises.

Imagine you are designing a high-fidelity audio system. The signal, once captured by a microphone and converted to a voltage, is normalized to a convenient range, say between -1.0 and 1.0. You have 16 bits to store each sample. How do you use them? This is not just a technical question; it's a question of resource allocation. A fixed-point representation, known as a $Q_{m.n}$ format, forces us to decide: how many bits, $m$, should we reserve for the integer part of the number (for the large-scale values), and how many, $n$, for the fractional part (for the fine details)? For a high-fidelity audio signal that lives entirely within $[-1, 1)$, there are no large integer values to worry about. The action is all in the subtle nuances, the delicate variations close to zero. The wisest choice, then, is to devote as many bits as possible to precision. We choose a format like $Q_{1.15}$, using only one bit for the sign and the (zero) integer part, and a luxurious 15 bits for the fractional part. This allows us to represent the signal with maximum fidelity, sacrificing a dynamic range we never needed in the first place.

But this act of mapping from the continuous to the discrete has consequences. What happens to a value that falls between the cracks of our $2^{15}$ fractional steps? It must be rounded to the nearest representable value. And what if a sudden, loud sound spike exceeds our $[-1, 1)$ range? To prevent the digital representation from nonsensically "wrapping around" due to the nature of two's complement arithmetic—which would turn a loud clap into a deafening digital buzz—the hardware must enforce saturation. The value is simply clipped to the maximum or minimum representable value. The process of converting a real number $x$ to its final integer code involves scaling, rounding, and saturating, a careful three-step dance to tame the infinite continuum into a finite set of bits without creating ugly artifacts. This is the invisible, foundational engineering that underpins every digital song you've ever heard.

Once a signal is safely inside a machine, how do we think about it? One of the most powerful ideas in all of science is to understand a complex thing by breaking it down into its simplest possible components. What is the simplest component, the fundamental "atom," of a discrete-time signal? It is a single, instantaneous blip of value 1 at a single point in time, and zero everywhere else: the unit impulse, $\delta[n]$.

A beautiful and profound property, sometimes called the sifting property, tells us that any signal $x[n]$ can be seen as a grand procession of these impulses, each occurring at a different time $k$ and scaled by the signal's value at that moment, $x[k]$. Mathematically, this is written as: $x[n] = \sum_{k=-\infty}^{\infty} x[k] \delta[n-k]$ This might look like a mere tautology, but its power is immense. It means that if we want to understand how any linear, time-invariant system (like a filter or an amplifier) will behave, we don't need to test it with every possible input signal. We only need to see how it responds to one signal: a single unit impulse. Its response to that one impulse, called the "impulse response," becomes a complete fingerprint of the system. To find the output for any input signal $x[n]$, we simply add up the scaled and shifted copies of the impulse response, a process known as convolution.

This representational shift from viewing a signal as a list of numbers to a sum of impulses makes complex operations startlingly clear. Consider the process of upsampling by a factor $L$, where we insert $L-1$ zeros between each sample of our original signal. What does this look like in the language of impulses? If our original signal was a parade of impulses, the upsampled signal is simply the same parade, but now the impulses are spaced $L$ times further apart. The representation changes from $\sum_k x[k] \delta[n-k]$ to $\sum_k x[k] \delta[n-kL]$. The underlying structure of the operation is laid bare by the representation.

For centuries, we have been taught that imaginary numbers are somehow less "real" than real ones. Yet in signal processing, we often find that the most insightful way to look at a very real physical phenomenon, like an oscillating wave, is to represent it in the complex plane. A real-world oscillation, like a simple cosine wave, has both an amplitude (how big is the oscillation?) and a phase (where are we in the cycle?). A single real number struggles to hold both pieces of information simultaneously.

The solution is an act of brilliant imagination: for any real signal $s(t)$, we construct a unique "shadow" signal, $\hat{s}(t)$ (its Hilbert transform), and define a new complex signal, the analytic signal, as $s_a(t) = s(t) + j\hat{s}(t)$. When we do this, the unruly world of trigonometric identities melts away into the clean, beautiful geometry of the complex plane. Our signal is no longer just a wave bobbing up and down; it becomes a vector rotating in the complex plane, which we can write in polar form as $A_e(t) e^{j\phi(t)}$. The magnitude of this vector, $A_e(t)$, is the instantaneous amplitude or envelope of the signal, and its angle, $\phi(t)$, is its instantaneous phase.

This representation is incredibly powerful. Consider the challenge of demodulating an AM radio signal, which consists of a high-frequency carrier wave whose amplitude is modulated by a lower-frequency audio signal. In the analytic signal representation, this complicated signal becomes a simple picture: a vector spinning very fast (at the carrier frequency), whose length is slowly changing according to the audio message. To recover the audio, we don't need complicated electronic filters; we just need to ask at every moment, "How long is the vector?" The demodulated signal is simply the magnitude, $y(t) = |s_a(t)|$.

This same technique solves another deep question: what is the "frequency" of a signal whose frequency is constantly changing? Think of a "chirp" signal, common in radar and sonar systems, whose pitch slides up or down over time. The analytic signal provides a perfect, unambiguous answer. The instantaneous frequency, $\omega(t)$, is simply how fast the representative vector is spinning at any given moment: $\omega(t) = \frac{d\phi(t)}{dt}$. A deeply intuitive physical concept is given a precise, elegant, and computable mathematical definition, all thanks to the courage of embracing complex numbers to describe real phenomena.

We have seen that a change in representation can reveal hidden structures. But this raises a crucial question: is there one "best" representation? The answer is no. The world of signals is like a rich, detailed tapestry. You can look at it from afar to see the whole picture, or you can get up close to examine the individual threads. But, as Heisenberg's uncertainty principle teaches us in a different context, you cannot do both perfectly at the same time. The more precisely you resolve a signal in time (asking "when" did something happen), the less precisely you can resolve it in frequency (asking "what" notes were played), and vice-versa.

The art of signal representation is the art of choosing the right lens for the job, a lens that manages this time-frequency trade-off in a way that is useful for your specific question. In ​​Problem 1765715​​, we compare two different ways of creating a time-frequency "map" of a signal. The spectrogram, which is based on the linear Short-Time Fourier Transform (STFT), is a cautious approach. It provides a useful, if slightly blurry, picture, and because of its mathematical linearity, any interference "artifacts" it creates are confined to regions where the signal components actually overlap in time and frequency. In contrast, the Wigner-Ville Distribution (WVD) is far more ambitious. It is a bilinear representation that promises perfect resolution, but it pays a heavy price. Its non-local nature creates "ghost" artifacts or "cross-terms" that appear in the middle of nowhere, halfway in time and frequency between real signal components. Understanding the mathematical nature of your representation is key to interpreting the picture it produces.

This idea of choosing the right lens is most apparent when we talk about choosing a basis—a set of fundamental shapes, or "atoms," out of which we build our signal. ​​Problem 2449853​​ stages a classic duel between two of the most important bases. In one corner, we have the Discrete Fourier Transform (DFT) basis, whose atoms are the eternal, perfectly smooth sine and cosine waves. In the other, we have the Haar wavelet basis, whose atoms are short, choppy, localized block-like wiggles. Which is better? It depends entirely on the signal you are trying to describe. A pure musical tone, being a sine wave itself, can be described with just one or two Fourier atoms but requires a whole committee of wavelets to approximate. Conversely, a signal with a sharp, sudden change—a click in an audio track or a sharp edge in an image—is perfectly captured by a single, localized wavelet atom but requires a near-infinite sum of smooth sine waves (the Gibbs phenomenon) to describe in the Fourier world.

The goal is to find a representation in which the signal is sparse—that is, describable with the fewest possible non-zero coefficients. This principle of sparsity is the engine behind all modern compression, from MP3s to JPEGs. The different philosophies of Fourier and wavelets manifest in practical tools like filter banks. DFT-based filter banks chop the frequency spectrum into uniform, equal-width channels, ideal for applications like telecommunications. Wavelet-based filter banks use a logarithmic, octave-band structure, providing high time resolution for high-frequency events and high frequency resolution for low-frequency phenomena. This multiresolution analysis is perfectly suited for natural signals like images and sounds.

And what if no off-the-shelf basis is quite right? Advanced signal processing shows us we can even engineer our own representations. We can create overcomplete dictionaries—a vocabulary with more "words" than strictly necessary—to find even sparser ways to describe a signal. For instance, to combat the fact that a standard wavelet basis is not shift-invariant, we can create a dictionary containing wavelets at every possible shift, ensuring we always have an atom that lines up perfectly with a feature, no matter where it occurs.

Finally, let us zoom out from single signals to entire systems of interacting components. Think of a complex control system, an electrical circuit, or even an economic model. We can represent such systems as a signal flow graph, a drawing where signals are nodes and the processes that transform them are directed, weighted arrows connecting them. A dense web of algebraic equations becomes an intuitive picture, a network that we can analyze visually.

This graphical representation allows us to see the structure of feedback loops and pathways. But it also hides a deep connection to physical reality. For a system to be causal and well-posed—for it to be a valid description of a real-world process—it cannot contain instantaneous feedback loops where an output can determine its own value in the exact same instant. This would be like an effect preceding its cause. In the language of signal flow graphs, this physical constraint translates into a clean, crisp condition on the system's gain matrix, $G(s)$: the matrix $I - G(\infty)$ must be invertible. If this condition fails, it means the graph contains an algebraic loop that cannot be resolved in time. This is a breathtaking example of how an abstract mathematical representation, born of graph theory and linear algebra, must still bow to the fundamental laws of causality and the unidirectional arrow of time.

Our tour is complete. We have seen that "signal representation" is not a single subject but a vast and creative art form. It is the art of choosing the right language to describe a piece of the world, whether it's to store music on a chip, decode a radio wave from across the planet, compress a photograph, or ensure a robot arm moves just right. The inherent beauty lies not in finding a single, ultimate truth, but in mastering a diverse palette of descriptions and having the wisdom to choose the one that makes the complex simple and the invisible clear.