What Are Stochastic Processes?

Stochastic processes are mathematical models for systems that evolve over time in a manner governed by probability. Unlike deterministic systems, where the future state is fully specified by the current state and the governing equations, stochastic processes incorporate randomness as a fundamental feature rather than an approximation. They are used to model phenomena ranging from the thermal noise in electronic circuits to the price movements of financial assets, from the arrival of packets at a network router to the dynamics of gene expression within a cell.

A stochastic process is formally a collection of random variables indexed by time (or another parameter such as space), all defined on a common probability space. The character of a stochastic process is determined by the joint distribution of these random variables and by structural assumptions about how the distribution evolves with the index. Different classes of structural assumption lead to the major families of stochastic processes studied in probability theory and engineering.

Markov Processes

A Markov process satisfies the Markov property: the future evolution of the process, given the present state, is independent of the past. This memorylessness property makes Markov processes analytically tractable and computationally manageable. Discrete-state, discrete-time Markov chains are described by transition probability matrices, while continuous-time Markov chains use transition rate matrices and are governed by the Chapman-Kolmogorov equations. Continuous-state Markov processes include diffusions described by stochastic differential equations. Hidden Markov models, which add an observation layer above the Markov chain, underpin speech recognition, bioinformatics sequence analysis, and many other pattern recognition applications. NIST Digital Library of Mathematical Functions coverage of Markov processes documents the mathematical framework underlying these models across applied mathematics.

Gaussian Processes

A Gaussian process is a stochastic process in which any finite collection of variables has a joint Gaussian distribution. A Gaussian process is completely characterized by its mean function and covariance (kernel) function. Gaussian processes are widely used in spatial statistics (kriging) and in machine learning as non-parametric probabilistic models that provide calibrated uncertainty estimates alongside predictions. The choice of kernel encodes assumptions about smoothness, periodicity, and length scale of the underlying function. Research on Gaussian process methods published through the Journal of Machine Learning Research and related venues has produced scalable approximations that extend their applicability to large datasets.

Poisson Processes

A Poisson process models the occurrence of events that happen independently and at a constant average rate. The number of events in any time interval follows a Poisson distribution, and the inter-event times follow an exponential distribution. The Poisson process is the foundational model for arrival processes in queueing theory, for photon arrivals in optical receivers, for radioactive decay events, and for packet arrivals in communication networks. Generalizations including the nonhomogeneous Poisson process (with time-varying rate) and the compound Poisson process (with random event magnitudes) extend the basic model to a wider range of physical systems.

Wiener Process and Stationary Processes

The Wiener process, also called Brownian motion, is the continuous-time limit of a random walk: it has independent and normally distributed increments, starts at zero, and has almost surely continuous sample paths. It serves as the driving noise in stochastic differential equations and underpins the Black-Scholes model in finance. Stationary processes are those whose statistical properties do not change with time shifts, meaning the distribution of any finite collection of variables depends only on the time differences between them. IEEE Transactions on Signal Processing publications on stationary process estimation address spectral estimation and filtering problems where stationarity is the key assumption enabling frequency-domain analysis.

Applications

• Communication systems use stochastic channel models to characterize fading, interference, and noise in wireless links.
• Queueing theory uses Markov chains and Poisson arrival models to design and size network infrastructure and service systems.
• Kalman filtering applies Gaussian process models to optimally estimate the state of dynamical systems from noisy sensor measurements.
• Financial engineering uses stochastic differential equations to price options and manage risk in derivative securities portfolios.
• Neuroscience research models neural spike trains as point processes to analyze coding properties of sensory neurons.