Why are maximal inequalities so valued?

Many arguments need to control the largest value a random process ever takes, not just its value at a fixed time; Doob's maximal inequalities provide exactly this control over the entire trajectory using only information about the endpoint.

What does the Azuma-Hoeffding inequality add over Chebyshev's?

Chebyshev gives only polynomially decaying tail bounds from the variance, whereas Azuma-Hoeffding gives exponentially decaying, Gaussian-type bounds for martingales with bounded increments, which is far sharper for rare large deviations.

Martingale Inequalities

Martingale inequalities bound how large a martingale can grow over its whole history in terms of its final value, turning control of an endpoint into control of an entire random trajectory.

Definition

Martingale inequalities are bounds that control the running maximum or the fluctuations of a martingale or submartingale, typically in terms of its terminal value, its increments, or its quadratic variation.

Scope

The topic covers Doob's maximal inequality bounding the probability that a submartingale ever exceeds a level, Doob's Lp inequality bounding the maximum in the p-th mean for p greater than one, the Azuma-Hoeffding inequality giving exponential concentration for martingales with bounded increments, and the Burkholder-Davis-Gundy inequalities relating the maximum of a martingale to its quadratic variation.

Core questions

How can the probability that a martingale ever crosses a high level be bounded?
How is the largest value of a martingale controlled in the p-th mean?
When do martingales with bounded increments concentrate exponentially around their mean?
How is the size of a martingale related to its accumulated quadratic variation?

Key concepts

Doob's maximal inequality
Doob's Lp inequality
Azuma-Hoeffding concentration
quadratic variation
Burkholder-Davis-Gundy inequalities

Key theories

Doob's maximal and Lp inequalities: The probability that a non-negative submartingale ever exceeds a level is bounded by its terminal mean divided by that level, and for p greater than one the p-th mean of the running maximum is controlled by a constant times the p-th mean of the terminal value, extending Markov's inequality to whole trajectories.
Azuma-Hoeffding inequality: A martingale whose successive increments are bounded deviates from its starting value by a given amount only with probability decaying like a Gaussian tail, providing sharp concentration bounds for sums with limited dependence.
Burkholder-Davis-Gundy inequalities: For each exponent the p-th mean of the maximum of a martingale is comparable, up to universal constants, to the p-th mean of the square root of its quadratic variation, linking the size of a martingale to its accumulated variability and underpinning stochastic integration.

Clinical relevance

Martingale inequalities are central to modern probabilistic analysis: Azuma-Hoeffding concentration bounds the deviations of complex random quantities in the analysis of algorithms and machine learning, Doob's inequalities control suprema in the convergence of stochastic processes, and the Burkholder-Davis-Gundy inequalities are essential to the construction and estimates of stochastic integrals.

History

Doob's maximal inequalities were part of his foundational martingale theory; Hoeffding's concentration bounds for sums were extended to martingales by Azuma in 1967, and Burkholder, Davis, and Gundy established the equivalence of martingale maxima and quadratic variation in the 1970s, a cornerstone of stochastic analysis.

Key figures

Joseph L. Doob
Kazuoki Azuma
Wassily Hoeffding
Donald Burkholder

Seminal works

doob1953

Frequently asked questions

Why are maximal inequalities so valued?: Many arguments need to control the largest value a random process ever takes, not just its value at a fixed time; Doob's maximal inequalities provide exactly this control over the entire trajectory using only information about the endpoint.
What does the Azuma-Hoeffding inequality add over Chebyshev's?: Chebyshev gives only polynomially decaying tail bounds from the variance, whereas Azuma-Hoeffding gives exponentially decaying, Gaussian-type bounds for martingales with bounded increments, which is far sharper for rare large deviations.