White noise

In signal processing, white noise is characterized as a random signal exhibiting uniform intensity across various frequencies, resulting in a constant power spectral density. This terminology, or similar interpretations, finds application across numerous scientific and technical fields, such as physics, acoustical engineering, telecommunications, and statistical forecasting. It is important to note that white noise represents a statistical model for signals and their sources, rather than denoting a particular signal. The designation 'white noise' originates from white light; however, light perceived as white typically lacks a flat power spectral density across the visible spectrum.

Within a discrete-time framework, white noise manifests as a discrete signal where individual samples are considered a sequence of serially uncorrelated random variables, possessing a zero mean and finite variance; a singular instance of white noise is termed a random shock. Certain contexts further stipulate that these samples must be independent and exhibit an identical probability distribution, making independent and identically distributed random variables the most straightforward representation of white noise. Specifically, when each sample follows a normal distribution with a zero mean, the signal is then classified as additive white Gaussian noise.

Samples comprising a white noise signal can be ordered sequentially in time or distributed across one or multiple spatial dimensions. In the realm of digital image processing, pixels within a white noise image are commonly organized into a rectangular grid and are presumed to be independent random variables with a uniform probability distribution across a specified interval. This concept extends to signals disseminated over more intricate domains, including spherical or toroidal geometries.

An infinite-bandwidth white noise signal represents an entirely theoretical construct. In practical applications, the bandwidth of white noise is constrained by the noise generation mechanism, the transmission medium, and inherent limitations in observation capabilities. Consequently, random signals are classified as white noise if their spectrum appears flat across the frequency range pertinent to a given context. For an audio signal, the relevant spectrum encompasses the audible sound frequencies, typically ranging from 20 to 20,000 Hz. Such a signal is perceived by the human ear as a distinct hissing sound, akin to the /h/ phoneme produced during sustained aspiration. Conversely, the sh sound /ʃ/, as heard in the word ash, constitutes a colored noise due to its characteristic formant structure. Within the disciplines of music and acoustics, the designation white noise may be applied to any signal that produces a comparable hissing auditory effect.

Within the framework of phylogenetically based statistical methodologies, the term white noise may denote an absence of phylogenetic patterning in comparative datasets. In non-technical discourse, this term occasionally signifies 'random communication devoid of substantive content'.

Statistical Properties

Any value distribution is permissible, provided it possesses a zero DC component. Even a binary signal, restricted to values of 1 or -1, can exhibit white noise characteristics if its sequence is statistically uncorrelated. Naturally, noise characterized by a continuous distribution, such as a normal distribution, can also be classified as white.

A common misconception is that Gaussian noise, defined as noise with a Gaussian amplitude distribution (referencing the normal distribution), inherently implies white noise; however, neither property necessitates the other. Gaussianity pertains to the probability distribution of the signal's values, specifically the likelihood of the signal's amplitude falling within a given range. Conversely, the descriptor 'white' characterizes how the signal power is distributed, specifically independently, across time or frequencies.

A specific manifestation of white noise is the generalized mean-square derivative of either the Wiener process or Brownian motion.

The white noise measure represents a generalization applicable to random elements within infinite-dimensional spaces, including random fields.

Practical Applications

Music

White noise finds frequent application in the production of electronic music, either utilized directly or as an input to a filter for generating alternative noise signal types. Its use is widespread in audio synthesis, particularly for emulating percussive instruments like cymbals or snare drums, which inherently possess significant noise content within their frequency domain. A straightforward illustration of white noise is the static encountered from a non-existent radio station.

Electronics Engineering

White noise is also employed to determine the impulse response of electrical circuits, particularly for amplifiers and other audio apparatus. However, it is unsuitable for loudspeaker testing due to its excessive high-frequency spectral content. Conversely, pink noise, distinguished by its equal energy distribution across each octave, is utilized for evaluating transducers such as loudspeakers and microphones.

Computing

White noise serves as a foundational element for certain random number generators. For instance, Random.org employs a network of atmospheric antennas to produce random digit sequences derived from sources accurately representable as white noise.

Tinnitus treatment

White noise is a frequently utilized synthetic noise source for sound masking in tinnitus therapy. Commercial white noise generators are marketed as devices to enhance privacy, facilitate sleep, and alleviate tinnitus symptoms. The Marpac Sleep-Mate, developed in 1962 by salesman Jim Buckwalter, represents the inaugural domestic white noise machine. A more straightforward and economical alternative involves tuning an AM radio to an unused frequency, generating "static" that functions as white noise. Nevertheless, white noise produced by a standard commercial radio receiver tuned to an unoccupied frequency is highly susceptible to contamination from spurious signals, including adjacent radio stations, harmonics from distant stations, interference from nearby electrical equipment, or atmospheric phenomena like solar flares and lightning.

Work environment

The impact of white noise on cognitive function presents conflicting findings. A limited 2007 study reported that ambient white noise stimulation enhanced cognitive performance in secondary students diagnosed with attention deficit hyperactivity disorder (ADHD), yet simultaneously diminished the performance of students without ADHD. Further research suggests its efficacy in elevating worker mood and productivity by obscuring ambient office noise, though it has been observed to impair cognitive performance in intricate card sorting exercises.

In a related investigation, an experiment involving sixty-six healthy participants explored the advantages of incorporating white noise into a learning environment. Participants were tasked with identifying various images amidst diverse background auditory stimuli. The cumulative findings indicated that white noise indeed confers benefits pertinent to learning. Specifically, the studies demonstrated a modest enhancement in participants' learning capabilities and recognition memory attributable to white noise.

Mathematical definitions

White noise vector

A random vector, defined as a random variable mapping to Rⁿ, is classified as a white noise vector or white random vector if each of its constituent components exhibits a probability distribution characterized by a zero mean and finite variance. Although conventional signal processing definitions mandate identical variances to ensure a perfectly flat power spectrum, more expansive statistical frameworks occasionally only stipulate finite variances and statistical independence among components, meaning their joint probability distribution must be the product of their individual distributions.

A prerequisite, though generally not a sufficient condition, for the statistical independence of two variables is their statistical uncorrelatedness, implying a zero covariance. Consequently, the covariance matrix R for the n components of a white noise vector w must be an n by n diagonal matrix, where each diagonal entry R_ii corresponds to the variance of component w_i. Furthermore, the correlation matrix must be the n by n identity matrix. Should the variances be identical, the covariance matrix simplifies to a scalar multiple of the identity matrix, expressed as R = σ<!-- σ --> §3738§ I {\displaystyle R=\sigma ^{2}I} .

A vector w is defined as a Gaussian white noise vector if, beyond the independence of its constituent variables, each variable adheres to a normal distribution with a zero mean and a uniform variance of σ<!-- σ --> §1314§ {\displaystyle \sigma ^{2}} . In this scenario, the joint distribution of w is a multivariate normal distribution, where the independence among variables confers spherical symmetry upon the distribution within an n-dimensional space. Consequently, any orthogonal transformation applied to this vector will also produce a Gaussian white random vector. Notably, under most discrete Fourier transform methodologies, such as FFT and Hartley transforms, the transform W of w will also be a Gaussian white noise vector. This means the n Fourier coefficients of w will be independent Gaussian variables, each characterized by a zero mean and an identical variance of σ<!-- σ --> §5051§ {\displaystyle \sigma ^{2}} .

The power spectrum P of a random vector w is formally defined as the expected value of the squared modulus of each coefficient from its Fourier transform W, represented by the equation P_i = E(|W_i|§1415§). Based on this definition, a Gaussian white noise vector exhibits a perfectly flat power spectrum, with P_i = σ§2223§ for all i.

If w is characterized as a white random vector but lacks the Gaussian property, its Fourier coefficients W_i will not be entirely independent. Nevertheless, for large values of n and under common probability distributions, these dependencies are typically very subtle, permitting the assumption that their pairwise correlations are zero.

The definition of white noise frequently incorporates the weaker condition of statistical uncorrelatedness, as opposed to statistical independence. However, several properties typically anticipated for white noise, such as a flat power spectrum, may not be applicable under this less stringent criterion. When operating under this weaker assumption, the more rigorous variant can be explicitly termed an independent white noise vector. Some authors alternatively employ the designations "strongly white" and "weakly white."

An illustrative example of a random vector that constitutes Gaussian white noise in the weak sense but not in the strong sense is x = [ x §1617§ , x §2627§ ] {\displaystyle x=[x_{1},x_{2}]} . In this construct, x §5051§ {\displaystyle x_{1}} is a normal random variable with a zero mean, and ${\displaystyle x_{2}}$ is either + x §9697§ {\displaystyle +x_{1}} or −<!-- − --> x §121122§ {\displaystyle -x_{1}} , each occurring with equal probability. These variables are uncorrelated and individually normally distributed; however, they are neither jointly normally distributed nor independent. If ${\displaystyle x}$ undergoes a 45-degree rotation, its two components will remain uncorrelated, but their distribution will no longer be normal.

In certain contexts, the definition can be broadened to permit each component of a white random vector ${\displaystyle w}$ to possess a non-zero expected value ${\displaystyle \mu }$. Particularly in image processing, where samples are generally constrained to positive values, ${\displaystyle \mu }$ is frequently set to half of the maximum sample value. Consequently, the Fourier coefficient W §6061§ {\displaystyle W_{0}} , which represents the zero-frequency component (effectively, the mean of the ${\displaystyle w_{i}}$), will similarly exhibit a non-zero expected value of ${\displaystyle \mu {\sqrt {n}}}$. In this scenario, the power spectrum ${\displaystyle P}$ will maintain flatness exclusively across non-zero frequencies.

Discrete-time white noise

A discrete-time stochastic process, denoted as ${\displaystyle W(n)}$, represents an extension of a random vector from a finite to an infinite number of components. This type of discrete-time stochastic process, specifically ${\displaystyle W(n)}$, is characterized as weak-sense white noise (frequently termed "white noise" within signal processing contexts) if two conditions are met: first, its mean is zero for all ${\displaystyle n}$, expressed as E ⁡<!-- ⁡ --> [ W ( n ) ] = §8788§ {\displaystyle \operatorname {E} [W(n)]=0} ; and second, its autocorrelation function, ${\displaystyle R_{W}(n)=\operatorname {E} [W(k+n)W(k)]}$, exhibits a non-zero value exclusively when n = §166167§ {\displaystyle n=0} . This latter condition is formally expressed as ${\displaystyle R_{W}(n)=\sigma ^{2}\delta (n)}$, where σ<!-- σ --> §235236§ {\displaystyle \sigma ^2} denotes the variance and δ<!-- δ --> ( n ) {\displaystyle \delta (n)} represents the Kronecker delta function.

For ${\displaystyle W(n)}$, the variables are merely required to be uncorrelated, not necessarily statistically independent. This distinction is crucial because the defining flat power spectral density of white noise is solely determined by the process's second-order moments, specifically its autocorrelation function. In contrast, a process demanding that its samples be independent and identically distributed (i.i.d.) is designated as strict-sense white noise. It is important to note that if these uncorrelated random variables also exhibit a jointly Gaussian distribution, their statistical independence is inherently assured.

Continuous-time white noise

To establish the concept of white noise within the framework of continuous-time signal theory, the traditional random vector must be superseded by a continuous-time random signal. This implies a random process capable of generating a function ${\displaystyle w}$, which is dependent on a real-valued parameter ${\displaystyle t}$.

A process is categorized as white noise in its most stringent definition if the value ${\displaystyle w(t)}$ at any given time ${\displaystyle t}$ constitutes a random variable that maintains statistical independence from its entire preceding history prior to ${\displaystyle t}$. A less restrictive definition mandates independence solely between the values w ( t §6869§ ) {\displaystyle w(t_{1})} and ${\displaystyle w(t_{2})}$ for every pair of distinct time points t §120121§ {\displaystyle t_{1}} and ${\displaystyle t_{2}}$. An even more relaxed definition merely stipulates that such pairs, specifically w ( t §168169§ ) {\displaystyle w(t_{1})} and ${\displaystyle w(t_{2})}$, must be uncorrelated. Consistent with the discrete case, some researchers employ the weaker definition for white noise, reserving the term "independent" to denote either of the more rigorous definitions. Conversely, other scholars differentiate these definitions using the descriptors "weakly white" and "strongly white."

Establishing a precise definition for these concepts presents significant challenges, primarily because quantities represented as finite sums in discrete systems must be substituted with integrals that may not always converge. Consequently, the ensemble of all potential realizations for a signal ${\displaystyle w}$ transitions from a finite-dimensional space ${\displaystyle \mathbb {R} ^{n}}$ to an infinite-dimensional function space. Furthermore, regardless of the specific definition employed, a white noise signal ${\displaystyle w}$ inherently exhibits essential discontinuity at every point. This characteristic necessitates the application of advanced mathematical frameworks even for fundamental operations on ${\displaystyle w}$, such as integration over a finite interval.

Certain introductory engineering texts conceptualize this as a heuristic rather than a rigorous mathematical definition, stipulating that each value ${\displaystyle w(t)}$ must be a real-valued random variable possessing an expectation ${\displaystyle \mu }$ and a finite variance ${\displaystyle \sigma ^{2}}$. Consequently, the covariance E ( w ( t §8283§ ) ) ⋅ w ( t §99100§) ) ) {\displaystyle \mathrm {E} (w(t_{1})\cdot w(t_{2}))} between values at two distinct time points, t §125126§ {\displaystyle t_{1}} and ${\displaystyle t_{2}}$, is precisely defined: it evaluates to zero when the times differ, and to ${\displaystyle \sigma ^{2}}$ when they are identical. Nevertheless, under this specific definition, the integral

${\displaystyle W_{[a,a+r]}=\int _{a}^{a+r}w(t)\,dt}$

Consequently, this integral, when evaluated over any interval possessing a positive width ${\displaystyle r}$, would merely equate to the product of the width and the expectation: ${\displaystyle r\mu }$. Although a zero expectation is not intrinsically problematic, a finite pointwise variance would result in a zero variance for this integral, thereby indicating that the signal lacks any measurable energy. This characteristic fundamentally disqualifies the concept as an appropriate model for white noise signals, whether considered from a physical or mathematical perspective. A genuine white noise process necessitates a flat power spectral density, which inherently demands an infinite, rather than finite, pointwise variance.

Consequently, the majority of researchers characterize the signal ${\displaystyle w}$ through an indirect method, which involves assigning random values to the integrals of ${\displaystyle w(t)}$ and ${\displaystyle |w(t)|^{2}}$ across each interval ${\displaystyle [a,a+r]}$.

Within this framework, the integral ${\displaystyle W_{I}}$ of ${\displaystyle w(t)}$ over a specified interval ${\displaystyle I=[a,b]}$ is typically defined as a real random variable characterized by a normal distribution, a zero mean, and a variance of ${\displaystyle (b-a)\sigma ^{2}}$. Furthermore, the covariance ${\displaystyle \mathrm {E} (W_{I}\cdot W_{J})}$ between the integrals WI{\displaystyle W_I} and WJ{\displaystyle W_J} is given by ${\displaystyle r\sigma ^{2}}$, where ${\displaystyle r}$ represents the length of the intersection ${\displaystyle I\cap J}$ of the two intervals ${\displaystyle I,J}$. This configuration is referred to as a Gaussian white noise signal or process.

Within the mathematical discipline of white noise analysis, a Gaussian white noise, denoted as ${\displaystyle w}$, is formally characterized as a stochastic tempered distribution. This implies it is a random variable whose values reside within the space ${\displaystyle {\mathcal {S}}'(\mathbb {R} )}$ of tempered distributions. Similar to the methodology employed for finite-dimensional random vectors, a probability law on the infinite-dimensional space ${\displaystyle {\mathcal {S}}'(\mathbb {R} )}$ can be established through its characteristic function. The existence and uniqueness of this definition are ensured by an extension of the Bochner–Minlos theorem, specifically known as the Bochner–Minlos–Sazanov theorem.

• In an analogous manner to the multivariate normal distribution, represented as ${\displaystyle X\sim {\mathcal {N}}_{n}(\mu ,\Sigma )}$, the characteristic function is defined as follows:

∀<!-- ∀ --> k ∈<!-- ∈ --> R n : E ( e i ⟨<!-- ⟨ --> k , X ⟩<!-- ⟩ --> ) = e i ⟨<!-- ⟨ --> k , μ<!-- μ --> ⟩<!-- ⟩ --> −<!-- − --> §9091§ §9293§ ⟨ k , Σ k ⟩ , {\displaystyle \forall k\in \mathbb {R} ^{n}:\quad \mathrm {E} (\mathrm {e} ^{\mathrm {i} \langle k,X\rangle })=\mathrm {e} ^{\mathrm {i} \langle k,\mu \rangle -{\frac {1}{2}}\langle k,\Sigma k\rangle },}

• The white noise, denoted as ${\displaystyle w:\Omega \to {\mathcal {S}}'(\mathbb {R} )}$, is required to fulfill the following condition:

∀<!-- ∀ --> φ<!-- φ --> ∈<!-- ∈ --> S ( R ) : E ( e i ⟨<!-- ⟨ --> w , φ<!-- φ --> ⟩<!-- ⟩ --> ) = e −<!-- − --> §7980§ §8182§ ‖ φ ‖ §9697§ §100101§ , {\displaystyle \forall \varphi \in {\mathcal {S}}(\mathbb {R} ):\quad \mathrm {E} (\mathrm {e} ^{\mathrm {i} \langle w,\varphi \rangle })=\mathrm {e} ^{-{\frac {1}{2}}\|\varphi \|_{2}^{2}},}

• The expression ${\displaystyle \langle w,\varphi \rangle }$ represents the canonical pairing between the tempered distribution ${\displaystyle w(\omega )}$ and the Schwartz function ${\displaystyle \varphi }$. In this context, ${\displaystyle \varphi }$ is treated as a constant linear functional on ${\displaystyle {\mathcal {S}}'(\mathbb {R} )}$, analogous to ${\displaystyle k}$ previously discussed.
• Furthermore, the expression ${\displaystyle \|\varphi \|_{2}^{2}=\int _{\mathbb {R} }\vert \varphi (x)\vert ^{2}\,\mathrm {d} x}$ is defined.

Mathematical applications

Time series analysis and regression

In statistics and econometrics, it is frequently posited that an observed data series results from the summation of values produced by a deterministic linear process, which is contingent upon specific independent (explanatory) variables, and a sequence of random noise values. Subsequently, regression analysis is employed to deduce the model process parameters from the empirical data, for instance, through ordinary least squares. This also involves testing the null hypothesis that each parameter is zero against the alternative hypothesis of a non-zero parameter. Typically, hypothesis testing postulates that the noise values are mutually uncorrelated, possess a zero mean, and adhere to an identical Gaussian probability distribution, implying that the noise is Gaussian white rather than merely white. Should a non-zero correlation exist among the noise values underpinning distinct observations, the estimated model parameters remain unbiased; however, their uncertainty estimates, such as confidence intervals, will exhibit bias (i.e., they will not be accurate on average). This phenomenon also occurs if the noise is heteroskedastic, meaning it displays varying variances across different data points.

Conversely, within the domain of time series analysis, a specialized area of regression analysis, explanatory variables are frequently limited to the historical values of the dependent variable being modeled. Under these circumstances, the noise process is commonly conceptualized as a moving average process, where the dependent variable's current value is determined by both present and preceding values of a sequential white noise process.

Random vector transformations

Through an appropriate linear transformation, termed a coloring transformation, a white random vector can be utilized to generate a non-white random vector (i.e., a collection of random variables) characterized by elements possessing a predefined covariance matrix. Conversely, a random vector with an established covariance matrix can be converted into a white random vector via an appropriate whitening transformation.

These two principles are fundamental in various applications, including channel estimation and channel equalization within communication and audio systems. Furthermore, these concepts find utility in data compression methodologies.

Generation

Digital generation of white noise can be accomplished using a digital signal processor, microprocessor, or microcontroller. This process generally involves supplying a suitable sequence of random numbers to a digital-to-analog converter. The resultant white noise's fidelity is directly contingent upon the efficacy of the underlying algorithm.

Colloquial Applications

The term "white noise" is occasionally employed colloquially to characterize an ambient soundscape that produces an indistinct or continuous auditory environment. Illustrative instances encompass:

• The collective vocalizations from numerous conversations occurring within the acoustic confines of an enclosed area.
• A persistent, non-intrusive auditory input utilized to obscure extraneous background disturbances and facilitate states of relaxation or slumber.
• The redundant and obfuscatory rhetoric frequently employed by political figures to conceal specific points they wish to remain unacknowledged.
• Musical compositions characterized by disagreeable, harsh, dissonant, or discordant qualities, devoid of discernible melodic structure.

Furthermore, the term finds metaphorical application, exemplified by Don DeLillo's 1985 novel, White Noise. This literary work investigates the convergent manifestations of contemporary culture that impede an individual's capacity to actualize their personal ideas and identity.

References

References