Definition and Finite-Dimensional Distributions

Definition:

Gaussian Process

A real-valued stochastic process {X(t):tโˆˆT}\{X(t) : t \in \mathcal{T}\} is a Gaussian process if for every Nโ‰ฅ1N \geq 1 and every choice of time instants t1,t2,โ€ฆ,tNโˆˆTt_1, t_2, \ldots, t_N \in \mathcal{T}, the random vector X=[X(t1)X(t2)โ‹ฎX(tN)]\mathbf{X} = \begin{bmatrix} X(t_1) \\ X(t_2) \\ \vdots \\ X(t_N) \end{bmatrix} has a joint Gaussian distribution N(ฮผ,ฮฃ)\mathcal{N}(\boldsymbol{\mu}, \boldsymbol{\Sigma}), where [ฮผ]k=E[X(tk)][\boldsymbol{\mu}]_k = \mathbb{E}[X(t_k)] and [ฮฃ]k,โ„“=Cov(X(tk),X(tโ„“))[\boldsymbol{\Sigma}]_{k,\ell} = \text{Cov}(X(t_k), X(t_\ell)).

Equivalently, every finite linear combination โˆ‘k=1NakX(tk)\sum_{k=1}^N a_k X(t_k) is a Gaussian random variable for all a1,โ€ฆ,aNโˆˆRa_1, \ldots, a_N \in \mathbb{R}.

The equivalence follows because โˆ‘kakX(tk)=aTX\sum_k a_k X(t_k) = \mathbf{a}^\mathsf{T} \mathbf{X}, and a random vector is jointly Gaussian if and only if every linear combination of its components is Gaussian.

Complete Characterization by First and Second Moments

A Gaussian distribution is uniquely determined by its mean and covariance. Therefore, a Gaussian process {X(t)}\{X(t)\} is completely characterized by two functions:

  1. Mean function: ฮผX(t)=E[X(t)]\mu_X(t) = \mathbb{E}[X(t)]
  2. Covariance function: CX(t1,t2)=Cov(X(t1),X(t2))=E[(X(t1)โˆ’ฮผX(t1))(X(t2)โˆ’ฮผX(t2))]C_X(t_1, t_2) = \text{Cov}(X(t_1), X(t_2)) = \mathbb{E}[(X(t_1) - \mu_X(t_1))(X(t_2) - \mu_X(t_2))]

This is a remarkable simplification: while a general process requires the full hierarchy of finite-dimensional distributions, for a Gaussian process the mean and covariance functions tell us everything.

Theorem: Closure of Gaussian Processes Under Linear Operations

Let {X(t)}\{X(t)\} be a Gaussian process and let Y=โˆซabg(t)X(t)โ€‰dtY = \int_a^b g(t) X(t)\,dt for a deterministic function g(t)g(t) (where the integral exists in mean square). Then YY is a Gaussian random variable.

More generally, if {X(t)}\{X(t)\} is Gaussian and Y(s)=โˆซg(s,t)X(t)โ€‰dtY(s) = \int g(s,t) X(t)\,dt, then {Y(s)}\{Y(s)\} is also a Gaussian process.

Gaussian distributions are closed under linear operations (sums, integrals, limits in mean square). Since an integral is a limit of finite sums, and each finite sum of Gaussian RVs is Gaussian, the limit inherits Gaussianity.

Definition:

Properties of the Covariance Function

A function C:Tร—Tโ†’RC : \mathcal{T} \times \mathcal{T} \to \mathbb{R} is a valid covariance function if and only if it satisfies:

  1. Symmetry: C(t1,t2)=C(t2,t1)C(t_1, t_2) = C(t_2, t_1)
  2. Positive semi-definiteness: For all NN, all t1,โ€ฆ,tNt_1, \ldots, t_N, and all a1,โ€ฆ,aNโˆˆRa_1, \ldots, a_N \in \mathbb{R}, โˆ‘k=1Nโˆ‘โ„“=1Nakaโ„“โ€‰C(tk,tโ„“)โ‰ฅ0\sum_{k=1}^N \sum_{\ell=1}^N a_k a_\ell\, C(t_k, t_\ell) \geq 0

Any function satisfying these two conditions defines a unique Gaussian process (up to the choice of mean function), by the Kolmogorov extension theorem.

Condition 2 is equivalent to requiring that every finite-dimensional covariance matrix ฮฃ\boldsymbol{\Sigma} be positive semi-definite.

Example: Is X(t)=Acosโก(2ฯ€f0t+ฮ˜)X(t) = A\cos(2\pi f_0 t + \Theta) a Gaussian Process?

Let AโˆผN(0,ฯƒ2)A \sim \mathcal{N}(0, \sigma^2) and ฮ˜\Theta uniform on [0,2ฯ€)[0, 2\pi), with AA and ฮ˜\Theta independent. Determine whether X(t)=Acosโก(2ฯ€f0t+ฮ˜)X(t) = A\cos(2\pi f_0 t + \Theta) is a Gaussian process.

Example: Gaussian Process from Random Fourier Coefficients

Let AkโˆผN(0,ฯƒk2)A_k \sim \mathcal{N}(0, \sigma_k^2) for k=1,โ€ฆ,Kk = 1, \ldots, K be independent Gaussian RVs. Show that X(t)=โˆ‘k=1KAkฯ•k(t)X(t) = \sum_{k=1}^K A_k \phi_k(t), where ฯ•k(t)\phi_k(t) are deterministic functions, is a Gaussian process. Compute its mean and covariance.

Theorem: WSS Gaussian Process Is Strictly Stationary

If {X(t)}\{X(t)\} is a Gaussian process that is wide-sense stationary (WSS) โ€” i.e., ฮผX(t)=ฮผ\mu_X(t) = \mu is constant and rxx(ฯ„)=E[X(t+ฯ„)Xโˆ—(t)]r_{xx}(\tau) = \mathbb{E}[X(t+\tau)X^*(t)] depends only on ฯ„\tau โ€” then {X(t)}\{X(t)\} is strictly stationary.

A Gaussian distribution is completely determined by its mean and covariance. If these are time-shift invariant (WSS), then all finite-dimensional distributions are time-shift invariant, which is the definition of strict stationarity. For non-Gaussian processes, WSS does not imply strict stationarity because higher-order moments could still depend on absolute time.

,

Common Mistake: WSS Does Not Imply Strict Stationarity in General

Mistake:

Assuming that any WSS process is strictly stationary.

Correction:

This implication holds only for Gaussian processes. For a general process, WSS constrains only the first two moments; higher-order statistics (skewness, kurtosis, etc.) could still vary with time. A standard counterexample: let X[n]X[n] be i.i.d. Bernoulli(12\frac{1}{2}) taking values ยฑ1\pm 1. This is WSS (constant mean 0, rxx[m]=ฮด[m]r_{xx}[m] = \delta[m]) and also strictly stationary. But replace X[0]X[0] with a non-symmetric distribution keeping the same mean and variance โ€” the process is still WSS but no longer strictly stationary.

Sample Paths of a Gaussian Process

Draw sample paths from a zero-mean GP with different covariance kernels. Observe how the kernel controls smoothness, correlation length, and sample variability.

Parameters
1
1
5

Quick Check

A Gaussian process is completely characterized by which of the following?

Its mean function only

Its mean function and covariance function

All of its finite-dimensional PDFs

Its autocorrelation function and PSD

Gaussian Process

A stochastic process {X(t)}\{X(t)\} such that every finite-dimensional marginal [X(t1),โ€ฆ,X(tN)][X(t_1), \ldots, X(t_N)] is jointly Gaussian. Completely determined by its mean and covariance functions.

Related: Wide-Sense Stationary (WSS), Covariance Function (Kernel)

Covariance Function (Kernel)

A function C(t1,t2)=Cov(X(t1),X(t2))C(t_1, t_2) = \text{Cov}(X(t_1), X(t_2)) that must be symmetric and positive semi-definite. For a WSS process, C(t1,t2)=C(t1โˆ’t2)C(t_1, t_2) = C(t_1 - t_2) depends only on the time lag.

Related: Gaussian Process, Positive Semi-Definite (Function)

Positive Semi-Definite (Function)

A function C(ฯ„)C(\tau) is positive semi-definite if for all NN, all t1,โ€ฆ,tNt_1, \ldots, t_N, and all a1,โ€ฆ,aNa_1, \ldots, a_N, we have โˆ‘k,โ„“akaโ„“C(tkโˆ’tโ„“)โ‰ฅ0\sum_{k,\ell} a_k a_\ell C(t_k - t_\ell) \geq 0. Equivalently, its Fourier transform (the PSD) is nonnegative.

Related: Covariance Function (Kernel)

Historical Note: Kolmogorov and the Foundation of Stochastic Processes

1930s

Andrey Kolmogorov's 1933 monograph Grundbegriffe der Wahrscheinlichkeitsrechnung (Foundations of Probability Theory) placed probability on a rigorous measure-theoretic footing. His extension theorem guarantees the existence of a stochastic process with any consistent family of finite-dimensional distributions. For Gaussian processes, this means that specifying a mean and a positive semi-definite covariance function is sufficient to construct the process โ€” a fact we use freely throughout this chapter.

Key Takeaway

A Gaussian process is the simplest infinite-dimensional generalization of the multivariate Gaussian distribution: two functions โ€” the mean ฮผX(t)\mu_X(t) and the covariance CX(t1,t2)C_X(t_1, t_2) โ€” completely determine all statistical properties. This is why Gaussian models dominate engineering: they are rich enough to model complex phenomena yet tractable enough for closed-form analysis.