What Is Gaussian Noise?

Gaussian noise is a type of statistical noise in which the amplitude values of the noise signal follow a Gaussian (normal) probability distribution. At any given instant, the noise voltage or intensity sample is drawn from a distribution with a specified mean, typically zero, and a variance that determines the noise power. This model describes the thermal noise that arises from the random motion of electrons in resistors and amplifiers, and it provides the standard mathematical framework for analyzing noise in communication systems, sensors, and imaging devices.

The Gaussian model is favored in engineering analysis for both theoretical and practical reasons. On the theoretical side, the central limit theorem implies that noise formed by the superposition of many small independent random disturbances converges to Gaussian form. On the practical side, the Gaussian distribution is fully characterized by two parameters, mean and variance, which simplifies closed-form analysis of signal detection, estimation, and filtering.

Additive White Gaussian Noise

The most widely used noise model in communications is additive white Gaussian noise (AWGN), which combines two properties. "Additive" means the noise is superimposed directly on the signal rather than multiplying it or distorting it in a nonlinear way. "White" means the power spectral density of the noise is flat across all frequencies, so each frequency band contains equal noise power. Together, these properties allow a received signal to be written as the sum of the transmitted signal and an independent Gaussian random variable at each sample point.

AWGN forms the basis for the Shannon channel capacity formula and for receiver design criteria such as the matched filter, which is optimal when noise is additive and Gaussian. IEEE standards for physical layer performance, including those for Wi-Fi and digital video broadcasting, specify receiver sensitivity in terms of the signal-to-noise ratio at which a target bit error rate is achieved under AWGN conditions. Research on noise level estimation for digital images consistently models sensor noise as AWGN to establish benchmarks for denoising algorithms.

Image Denoising

When Gaussian noise corrupts a digital image, each pixel value is perturbed by a noise sample drawn from a Gaussian distribution. The standard deviation of this distribution, known as the noise level, controls the visible grain in the image. Estimating and removing this noise while preserving edges and textures is a central problem in image processing.

Classical denoising methods include Gaussian smoothing filters, which suppress noise by averaging neighboring pixels, and Wiener filters, which use the local signal-to-noise ratio to adaptively balance smoothing against detail preservation. More recent methods include the BM3D (block-matching and 3D collaborative filtering) algorithm, which exploits self-similar patterns across the image to achieve near-optimal denoising performance. A review published in Visual Computing for Industry, Biomedicine, and Art surveys the major families of denoising algorithms and their performance on standard Gaussian noise benchmarks. Deep learning approaches, including convolutional neural networks trained on noisy-clean image pairs, now achieve denoising performance that surpasses classical methods across a wide range of noise levels.

Applications

Gaussian noise has applications in a wide range of disciplines, including:

• Communication system design, where AWGN benchmarks define receiver sensitivity and coding gain targets
• Television broadcasting, where thermal noise once caused visible grain in analog reception, and where modern digital standards specify signal margin requirements under AWGN conditions
• Medical imaging, including MRI and CT, where noise reduction algorithms preserve diagnostic image quality
• Remote sensing and radar, where clutter and receiver noise are modeled as additive white Gaussian noise for detection threshold analysis
• Monte Carlo simulation and stochastic modeling, where Gaussian noise samples represent uncertain perturbations