Loading field map
How can we describe a system that changes unpredictably over time? A stock price, a jittering pollen grain in water, the queue length at a busy checkout counter—each evolves under the influence of chance. Ordinary calculus, built for smooth and deterministic paths, cannot handle such irregular motion. The history of stochastic processes is the story of building mathematical frameworks that can describe, analyze, and predict random evolution, each framework addressing a limitation of its predecessors while preserving what worked.
The first major framework emerged from a simple but powerful constraint. In 1906, Andrey Markov introduced what we now call Markov Chains: sequences of random variables where the future depends on the present alone, not on the full past. This Markov property turned an intractable history-dependent process into a manageable sequence of one-step transitions. Markov chains work on discrete state spaces—like "rainy" or "sunny"—and evolve in discrete time steps. They remain a workhorse today for modeling queues, text generation, and population genetics.
Markov Processes extended the same idea to continuous time. Instead of jumping at fixed intervals, the system can change at any moment. The state space can also be continuous, as in the classic example of radioactive decay, where each atom's survival time is random. The Markov property remains the same: given the present, the future is independent of the past. This framework absorbed the discrete-time chain as a special case and opened the door to modeling real-world phenomena that unfold continuously, such as chemical reactions or epidemic spread. Both frameworks remain active; Markov chains are preferred when the problem is naturally discrete or when computational simplicity matters, while Markov processes handle continuous-time dynamics.
A major turning point came in the 1920s with the Wiener Process, often called Brownian motion. Norbert Wiener constructed a rigorous mathematical model for the erratic path of a particle suspended in fluid. The Wiener process is a continuous-time Markov process with continuous but nowhere-differentiable paths—it wiggles so violently that its velocity is undefined at every point. This was both a breakthrough and a puzzle. It provided a universal model for noise in physics, engineering, and finance, but it also exposed a deep problem: how can we integrate or differentiate functions of a path that has no smoothness? The Wiener process became the canonical example of a stochastic process that is continuous yet infinitely irregular, setting the stage for the next frameworks.
While Markovian models focus on the conditional independence of the future, another line of inquiry asked about stability over time. Stationary Processes, developed from the 1930s onward, describe systems whose statistical properties—mean, variance, correlations—do not change when you shift the time origin. A stationary process need not be Markovian; its defining feature is that its joint distributions are invariant under time translation. This framework proved essential for time-series analysis, signal processing, and econometrics. It coexists with Markovian frameworks: a stationary process can also be Markovian (e.g., an autoregressive process of order one), but the two frameworks emphasize different aspects. Stationary processes are best at capturing long-run regularities, while Markov processes excel at modeling short-term transitions.
In the 1940s, Joseph Doob and others developed Martingale Theory, which introduced a different kind of constraint. A martingale is a stochastic process whose future expected value, given the present, equals the present value. It models a fair game: no matter what has happened, the expected future wealth is the current wealth. Martingales are not necessarily Markovian, and they are not necessarily stationary. Their power lies in their convergence properties and in the optional stopping theorem, which describes when a fair game remains fair even if you stop at a random time. Martingale theory provided a unifying tool for analyzing stochastic processes of all kinds. It became the backbone of modern probability, used to prove laws of large numbers, central limit theorems, and to price financial derivatives. It did not replace Markov processes or stationary processes; instead, it offered a new lens that could be applied within those frameworks.
The Wiener process had no derivative, yet physicists and engineers needed to write differential equations driven by noise. Stochastic Calculus, developed by Kiyoshi Itô in the 1940s, solved this by redefining integration. The Itô integral integrates a process against the Wiener process, using a forward-looking discretization that respects the non-anticipating nature of information flow. This led to stochastic differential equations (SDEs), where the change in a process is the sum of a drift term and a diffusion term driven by Brownian motion. The Itô calculus gave a rigorous way to model everything from stock prices to particle diffusion. A rival formulation, the Stratonovich integral, uses a midpoint discretization and obeys the ordinary chain rule, making it more natural for some physical applications. Both coexist today; the choice depends on whether one wants martingale properties (Itô) or geometric intuition (Stratonovich). Stochastic calculus transformed the Wiener process from a mathematical curiosity into a practical tool.
By the 1960s, it became clear that stochastic calculus could be extended beyond the Wiener process. Semimartingale Theory identified the largest class of processes for which a sensible stochastic integral can be defined. A semimartingale is any process that can be decomposed into a local martingale plus a process of finite variation. This class includes Markov processes, Wiener processes, and many others. The key insight is that the Itô calculus works for any semimartingale as integrator, not just Brownian motion. Semimartingale theory absorbed earlier frameworks by providing a unified setting: if you can write a process as a semimartingale, you can integrate with respect to it. This framework is now the standard foundation for stochastic calculus in mathematical finance and general probability theory. It narrowed the focus from arbitrary processes to those with a tractable decomposition, and it remains the dominant framework for defining stochastic integrals.
From the 1970s onward, Stochastic Analysis broadened the scope further. While semimartingale theory unified the integrator, stochastic analysis asks what happens when the integrand itself is more general—for example, when it depends on the entire path of the process, not just the current value. This led to the Malliavin calculus, which extends differential calculus to functionals of stochastic processes, and to rough path theory, which handles integration when the driving signal is even rougher than Brownian motion. Stochastic analysis also encompasses the general theory of processes, including predictable and optional sigma-algebras, which provide the measure-theoretic infrastructure for stopping times and filtrations. This framework does not replace semimartingale theory; it builds on it, adding tools for sensitivity analysis (Malliavin's calculus) and for paths with Hölder regularity less than 1/2 (rough paths). Stochastic analysis is the current frontier, used in mathematical finance for hedging, in physics for stochastic partial differential equations, and in statistics for inference on diffusion processes.
Today, the leading frameworks coexist with a clear division of labor. Markov Chains and Markov Processes remain essential for modeling systems where the Markov property holds approximately, from Google's PageRank algorithm to epidemiology. Wiener Process is the canonical noise model, still taught as the entry point to continuous-time stochastic processes. Stationary Processes dominate time-series analysis and signal processing. Martingale Theory provides the fundamental convergence tools used across all of probability. Stochastic Calculus (Itô and Stratonovich) is the standard language for SDEs. Semimartingale Theory is the accepted foundation for defining stochastic integrals in general. Stochastic Analysis pushes the boundaries with Malliavin calculus and rough paths.
What these frameworks agree on is the centrality of measure-theoretic probability: all modern stochastic processes are defined on a probability space with a filtration, and all rely on the Kolmogorov axioms. They disagree on what constraints are most useful. Markovian frameworks insist on the conditional independence of the future given the present; martingale theory insists on the fairness condition; stationary processes insist on time-invariance; semimartingale theory insists on the decomposition into martingale plus finite variation. Each constraint highlights a different aspect of random evolution, and each has generated a rich body of results. The field today is pluralistic: a researcher chooses the framework that best captures the essential structure of the phenomenon under study, and often combines several frameworks in a single analysis.