What Are Signal Representations?

Signal representations are the mathematical forms in which signals are expressed, stored, or manipulated within a processing system. Different representations make different properties of a signal accessible: a time-domain waveform reveals amplitude over time, a Fourier spectrum reveals frequency content, and a wavelet decomposition reveals localized time-frequency structure. Choosing among these forms is not merely a mathematical convenience; the representation determines what information is readily extractable, what compression is achievable, and how efficiently a given algorithm can operate on the data.

The study of signal representations draws from harmonic analysis, linear algebra, and information theory. Mathematicians including Jean-Baptiste Joseph Fourier, who introduced trigonometric series decomposition in the early nineteenth century, and Harry Nyquist and Claude Shannon, whose sampling and information theorems established the limits of discrete representation, laid the theoretical groundwork that practitioners build on today.

Analog and Digital Representations

Analog representations express a signal as a continuous function of time or space, preserving amplitude variation without discretization. Digital representations, by contrast, describe the signal through a finite sequence of numbers obtained by sampling in time and quantizing in amplitude. The conversion from analog to digital is governed by the Nyquist-Shannon sampling theorem, which states that a bandlimited signal of maximum frequency W Hz can be exactly reconstructed from samples taken at a rate of at least 2W samples per second. The number of quantization bits sets the dynamic range of the digital representation: each additional bit of resolution extends the representable range by approximately 6 dB. The IEEE Transactions on Signal Processing has published decades of research on how analog-to-digital conversion, sampling strategies, and quantization noise shape the fidelity of digital representations.

Parametric and Nonparametric Representations

Parametric representations describe a signal through a compact set of model parameters rather than a sample-by-sample record. Autoregressive (AR) models, for example, express each sample as a linear combination of past samples plus noise, requiring only a small coefficient vector to characterize a signal that may span many thousands of samples. Linear predictive coding (LPC), a parametric method widely used in speech processing, extracts vocal tract resonances in the form of a few dozen predictor coefficients, enabling highly compressed transmission and synthesis. Nonparametric representations make no explicit structural assumptions and instead describe the signal directly through its sampled values, its empirical spectrum, or its empirical correlation function. The NIST Digital Library of Mathematical Functions provides reference definitions for the mathematical constructs, including orthogonal series, Legendre polynomials, and Bessel functions, that underlie many parametric and nonparametric signal models.

Sparse and Overcomplete Representations

Sparse representations describe a signal using only a small number of nonzero coefficients drawn from a large dictionary of basis functions. When the dictionary is overcomplete, containing more elements than the signal dimension, there are in principle many decompositions of any given signal, and sparsity serves as the principle that selects among them. Compressed sensing, developed in the mid-2000s by Candes, Romberg, Tao, and Donoho, proved that signals sparse in a known basis can be accurately recovered from a number of random measurements far below the Nyquist rate. An influential review of sparse and redundant representations in signal processing in IEEE Transactions on Signal Processing surveys the algorithms and theory that connect sparse representations to practical problems in image restoration, source separation, and coding.

Applications

Signal representations have applications across a wide range of disciplines, including:

• Audio and speech coding, where parametric representations enable compact transmission at low bit rates
• Image and video compression, where transform-domain representations underpin standards such as JPEG and H.265
• Wireless communications, where frequency-domain representations support OFDM modulation and channel estimation
• Machine learning, where signal representations serve as feature spaces for classification and regression
• Medical imaging, where sparse and wavelet representations improve reconstruction quality in MRI and CT