Loading field map
A stochastic process is a mathematical object used to describe systems that evolve over time in a way that involves randomness. Where a classical function assigns a definite value to each point in time, a stochastic process assigns a random variable to each point in time, and the collection of these random variables is studied jointly. The term stochastic comes from the Greek stokhastikos, meaning "capable of guessing" or "aiming at a target," and it signals that the process is governed by probabilistic laws rather than deterministic ones.
The central questions of the field concern how such random evolutions behave. Given the rules that govern a process, what can be said about its future? Does it settle into a steady state, drift away, or oscillate forever? How do the dependencies between different times shape the long-run behavior? And, conversely, given observations of a process, what can be inferred about the rules that generated it? These questions make stochastic processes the mathematical backbone of fields as diverse as statistical physics, finance, population genetics, queuing theory, and signal processing.
To define a stochastic process formally, one starts with a probability space—a set of possible outcomes, a collection of events, and a probability measure assigning likelihoods to those events. A stochastic process is then a family of random variables \(X_t\), indexed by a parameter \(t\) that usually represents time. The index set can be discrete, as in \(t = 0, 1, 2, \dots\), or continuous, as in \(t \geq 0\). The values the process takes, called its state space, can also be discrete (like the number of customers in a queue) or continuous (like the position of a particle).
A crucial distinction is between the law of the process—the full probability distribution over all possible paths—and the individual paths themselves. A path is a single realization of the process, a function of time obtained by picking one outcome from the probability space. The law assigns probabilities to collections of paths, and most of the field's work consists of deriving properties of the law from the local rules that define it.
The most influential organizing idea in stochastic processes is the Markov property, named after the Russian mathematician Andrey Markov, who studied it in the early twentieth century. A process has the Markov property if, given the present state, the future is independent of the past. In other words, the process has no memory beyond its current position: how it arrived at the present state carries no information about where it will go next. This assumption is not always realistic, but it is remarkably often a good approximation, and it makes analysis tractable.
Markov processes come in several major varieties. A Markov chain has a discrete state space and evolves in discrete time; the classic example is a random walk on the integers, where at each step the process moves up or down with fixed probabilities. Markov chains are studied through their transition probabilities—the probabilities of moving from one state to another in a single step—and through the matrix of these probabilities. A central result is the classification of states as recurrent (the process returns to them infinitely often) or transient (it eventually leaves them forever), and the existence of stationary distributions, long-run proportions of time spent in each state that are invariant under the transition law.
When time is continuous but the state space remains discrete, one obtains continuous-time Markov chains, which are often described by transition rates rather than probabilities. These appear naturally in queuing theory, where customers arrive and are served at random times, and in chemical reaction networks, where molecules collide and react at random rates.
When the state space is continuous, the most important Markov process is Brownian motion, also called the Wiener process. Originally proposed to describe the erratic movement of pollen grains suspended in water, Brownian motion is characterized by independent, normally distributed increments: the displacement over any time interval is Gaussian, and displacements over non-overlapping intervals are independent. It has the striking property that its paths are continuous but nowhere differentiable—they are too jagged to have a well-defined velocity at any instant. This apparent paradox was resolved by Norbert Wiener in the 1920s, who constructed the process rigorously and showed that, although each path is continuous, the set of paths is so large that differentiability fails almost surely.
Brownian motion is the building block for diffusion processes, which solve stochastic differential equations. A stochastic differential equation (SDE) is an equation of the form
\[ dXt = b(Xt)\,dt + \sigma(Xt)\,dBt, \]
where \(Bt\) is Brownian motion. The term \(b(Xt)\,dt\) represents a deterministic drift, while \(\sigma(Xt)\,dBt\) represents random fluctuations whose size may depend on the current state. The theory of SDEs, developed by Kiyosi Itô and later refined by others, provides a calculus for these objects. The Itô integral, which defines what it means to integrate against Brownian motion, is subtle because Brownian motion is too irregular for ordinary Riemann–Stieltjes integration. The resulting Itô calculus has its own chain rule, the Itô formula, which differs from the classical chain rule by an extra term arising from the quadratic variation of Brownian motion.
The Markov property is a strong assumption, and a substantial part of the field concerns processes that violate it. Long-memory processes exhibit correlations between distant times that decay slowly, so that the past continues to influence the future over very long horizons. These arise in hydrology, internet traffic, and some financial time series. The fractional Brownian motion, a generalization of Brownian motion with correlated increments, is a standard model here, though its analysis is considerably harder because the Itô calculus does not apply directly.
Another important departure from the Markov framework is the hidden Markov model, where the observed process is a noisy function of an underlying Markov chain that is not directly visible. The underlying chain evolves according to Markov rules, but the observer sees only emissions from it. The central computational problem is filtering: given the observations up to the present, what is the distribution of the hidden state? This problem has a complete solution in the linear-Gaussian case, where it leads to the Kalman filter, and it is approached through sequential Monte Carlo methods, or particle filters, in more general settings.
Not all stochastic processes evolve continuously in time. Point processes describe the random locations of events in time or space. The simplest is the Poisson process, in which events occur independently at a constant rate; the number of events in any interval has a Poisson distribution, and the waiting times between events are independent and exponentially distributed. The Poisson process is the building block for more general renewal processes, where the inter-event times are independent but not necessarily exponential, and for Cox processes, where the rate itself is random.
Point processes extend naturally to spatial settings, where they describe the random placement of points in a plane or higher-dimensional space. These are central to spatial statistics, where they model the locations of trees in a forest, epicenters of earthquakes, or stars in a galaxy. The theory of random measures generalizes point processes further, allowing the random object to be a measure rather than a discrete set of points.
A martingale is a stochastic process that, on average, stays constant: the expected value of the future, given the present, is the present value. Martingales formalize the idea of a fair game, where no betting strategy can improve the expected outcome. They are not a separate class of processes so much as a property that many processes possess or can be transformed to possess. The optional stopping theorem states that, under mild conditions, a martingale stopped at a random time has the same expectation as its initial value; this simple fact underlies the impossibility of profitable betting systems and the pricing of financial derivatives.
The theory of martingales, developed by Joseph Doob in the mid-twentieth century, provides powerful convergence theorems: under broad conditions, a martingale that is bounded in a suitable sense converges almost surely to a limit. This gives a general method for proving that processes stabilize, and it is used throughout the field. The Doob–Meyer decomposition shows that a large class of processes can be split into a martingale part and a predictable, drift-like part, which is the basis for the modern theory of stochastic integration.
The development of stochastic processes has been driven by a constant exchange between mathematical theory and concrete problems. The theory of Markov chains grew partly from questions in statistical physics about the behavior of gases, and it later became central to computer science through Markov chain Monte Carlo methods, which use carefully constructed Markov chains to sample from complicated probability distributions. Brownian motion was studied by physicists before Wiener gave it a rigorous mathematical foundation, and it later became the basis for the Black–Scholes model in finance, which prices options by assuming that asset prices follow a geometric Brownian motion. The limitations of that model—its assumption of constant volatility and normally distributed returns—have in turn motivated the study of processes with jumps, such as Lévy processes, and of stochastic volatility models.
A recurring theme is the tension between tractability and realism. The Markov property and Brownian motion are mathematically convenient but often unrealistic; more general processes are more faithful to observed phenomena but harder to analyze. Much of the field's progress consists of finding intermediate models that capture essential features while remaining amenable to rigorous treatment.
Contemporary stochastic processes is a mature field with a well-developed toolkit. The Itô calculus has been extended to semimartingales, the largest class of processes for which a satisfactory integration theory exists. Rough path theory, developed by Terry Lyons, provides an alternative approach to integration against very irregular paths, including fractional Brownian motion, by encoding information about the path's higher-order increments. Stochastic partial differential equations extend the theory to fields that evolve randomly in both time and space, modeling phenomena like the random growth of interfaces or the fluctuations of a randomly stirred fluid.
The field also maintains a strong connection to statistics. Statistical inference for stochastic processes addresses the problem of estimating the parameters of a process from observed data, which requires new tools because the observations are dependent. Sequential analysis and optimal stopping study when to stop observing a process to make the best decision, with applications in clinical trials and finance.
Computational methods have become indispensable. Exact simulation of Brownian motion and related processes is possible in principle, but many models require approximation. Euler–Maruyama schemes approximate solutions of SDEs by discretizing time, and their convergence properties are well understood. Particle filters and Markov chain Monte Carlo methods make it possible to perform inference for models that are analytically intractable.
Throughout its history, the field has been characterized by a distinctive combination of probabilistic intuition and analytic rigor. The objects it studies are concrete enough to model real phenomena, yet subtle enough to require deep mathematics. The result is a discipline that is simultaneously a branch of pure mathematics, a toolbox for applied science, and a language for describing randomness in the world.