Van den Berg–Kesten–Reimer Inequality

• The van den Berg–Kesten–Reimer inequality is a fundamental probabilistic bound that quantifies the disjoint occurrence of events in product spaces using submultiplicative principles.
• Its functional and dual extensions apply to combinatorial optimization, random graphs, and spin systems by leveraging techniques from measure theory and cluster analysis.
• The inequality underpins practical methods in percolation theory, order statistics, and decision-tree algorithms, offering insight into joint probability bounds in complex stochastic models.

The van den Berg–Kesten–Reimer (BKR) inequality is a fundamental correlation bound in probability theory, combinatorics, and statistical mechanics, governing the probability that two or more events occur "disjointly" in a product probability space. It precisely quantifies the submultiplicativity of probabilities for events occurring for disjoint reasons, with extensions to functional settings, non-product measures, infinite product spaces, and dual inequalities. Its combinatorial and measure-theoretic structure underpins a wide array of applications, including percolation theory, order statistics, assignment problems, random graphs, spin systems, and the analysis of stochastic processes in the KPZ universality class.

1. Formal Statement and Definitions

Let $(S_i,\Sigma_i,P_i)$, $i=1,\dots,n$, be n probability spaces. Define their product space $S=\prod_{i=1}^n S_i$ with product $\sigma$-algebra $\Sigma=\bigotimes_{i=1}^n \Sigma_i$ and product measure $P=\bigotimes_{i=1}^n P_i$. For any $K \subset \{1,\dots,n\}$ and $x=(x_1,\dots,x_n)\in S$, the $K$-cylinder is

$$[x]_K = \{ y \in S : y_i = x_i \text{ for every } i \in K \}.$$

Given measurable events $A, B \subset S$, define their "disjoint-occurrence" (BKR-product): $$A \Box B = \left\{ x \in S : \exists\, K, L \subset \{1,\dots,n\},\, K \cap L = \emptyset,\; [x]_K \subset A,\; [x]_L \subset B \right\}.$$ The classical BKR inequality, proved by Reimer, states: $P(A \Box B) \leq P(A) P(B)$ for any product measure $P$ on a finite product of finite sets $S$ (Goldstein et al., 2015).

A functional extension holds: for non-negative measurable $f, g : S \to [0, \infty)$,

$$E\left[ \max_{K \cap L = \emptyset} \underline{f}_K(X)\, \underline{g}_L(X) \right] \leq E[f(X)]\, E[g(X)],$$

with $$\underline{f}_K(x) = \operatorname{ess\,inf}_{y \in [x]_K} f(y)$$. For indicator functions, the classical (event) BKR inequality is recovered.

2. Combinatorial and Probabilistic Structure

The essence of the BKR inequality lies in the property that witnessing both $A$ and $B$ using disjoint sets of coordinates is a stronger requirement than the independent satisfaction of $A$ and $B$. This "disjoint reasons" paradigm is crucial in combinatorial probability. The set $A \Box B$ collects all sample points such that $A$ and $B$ are both forced by coordinate-wise information on mutually exclusive sets.

The dual inequality, known as the Kahn–Saks–Smyth (KSS) inequality, formalizes an upper bound in the product of two independent copies: for $X, Y$ independent and identically distributed under $P$,

$$E\left[ \max_{K \cap L = \emptyset} \underline{f}_K(X)\, \underline{g}_L(Y) \right] \leq E[ f(X) g(X) ],$$

with the event version $(P_X \otimes P_Y)(A \Box B) \leq P(A \cap B)$ (Goldstein et al., 2015).

Reimer’s proof exploits a delicate combinatorial path-counting argument, extended to the functional setting via monotone-class techniques and essential infima over product cylinders.

3. Extensions to Non-Product Measures and Models

Classical BK-type inequalities require product measures; however, there are significant extensions:

• k-out-of-n Measures: For the uniform law on configurations with exactly $k$ ones ($P_{k,n}$), the inequality holds for all increasing $A, B \subset \{0,1\}^n$ (Berg et al., 2011, Berg et al., 2012):

$P_{k,n}(A \Box B) \leq P_{k,n}(A) P_{k,n}(B)$

• Weighted k-subset Measures: For non-negative weights, the result extends to conditional Poisson measures, with proofs involving symmetry, foldings, and random-cluster representations (Berg et al., 2012).
• Spin Systems: In the antiferromagnetic Curie–Weiss model ($J \leq 0$), the BK property holds for all increasing events. In the ferromagnetic Ising model, a "cluster-disjoint" variant of disjoint occurrence applies (Berg et al., 2012).
• Random Matchings: For the distribution of matchings in bipartite graphs (and appropriate maps of matchings to subsets of vertices), increasing events in the image satisfy the BK inequality (Mészáros, 2020).

The key structural insight across these extensions is the decomposition into matchings, cluster structures, and the use of generalized random-cluster representations.

4. Infinite Product Spaces and Generalizations

The BKR inequality has been generalized to arbitrary product probability spaces, including infinite-dimensional cases (Arratia et al., 2015). For Borel sets $A, B \subset [0,1]^d$ (or more general Polish spaces) with product measure $m_d$,

$m_d(A \Box B) \leq m_d(A) m_d(B)$

with $A \Box B$ defined using finite disjoint sets of coordinates. Proper attention to measurability—analytic- and Borel-structure of cylinders, and completion of measures—is necessary to maintain validity in the infinite (even uncountable) setting. The construction extends directly to $r$-fold disjoint occurrence for collections $(A_1, \ldots, A_r)$: $$m_d\left(\Box_{i=1}^r A_i\right) \leq \prod_{i=1}^r m_d(A_i)$$

5. Functional, Dual, and Threshold Generalizations

The functional BKR inequality replaces indicators by arbitrary non-negative measurable functions, employing essential infima over cylinders and maximizing over disjoint coordinate sets. This considerably strengthens the event-level inequality, allowing for a range of probabilistic inequalities in order statistics, allocation, network routing, and random graphs (Goldstein et al., 2015).

Refinements include:

• "Almost Sure" Disjoint Occurrence: Uses the set $A \Box_{11} B$, where conditioning on coordinate sets forces $A$ and $B$ with probability $1$ (Goldstein et al., 2016). This set generally strictly contains $A \Box B$ and leads to a tighter upper bound in some cases.
• Lenient (s, t)-Box: For thresholds $s, t \in [0,1]$, $A \Box_{s,t} B$ consists of $x$ such that coordinate revelations guarantee $A, B$ with probability at least $s$ and $t$, respectively (Goldstein et al., 2016).

For more than two events, Baron–Kahn established that the maximal number of events which can be simultaneously realized disjointly is stochastically dominated by the sum of independent Bernoullis with matching marginals: $$P(X \geq r) \leq P\left(\operatorname{Binomial}(k, p) \geq r \right)$$ where $X$ is the maximal number of realized disjoint occurrences among $k$ monotone events $A_1, \dots, A_k$ (Baron et al., 2019).

6. Applications Across Probability and Combinatorics

The BKR and related inequalities have been foundational in:

• Order Statistics and Symmetric Functions: Bounds for moments and mixed products of order statistics of i.i.d. random variables; e.g., $$\mathbb{E}[X_{[1]} X_{[2]}] \leq (\mathbb{E}[X_{[1]}])^2$$ (Goldstein et al., 2015).
• Assignment and Allocation Problems: Upper bounds on joint completions when assignment tasks use disjoint subsets of resources, vital in combinatorial optimization (Goldstein et al., 2015).
• Random Networks and Graphs: Submultiplicativity (and Chernoff-type upper tail bounds) for counts of disjoint substructures, such as disjoint paths or packings, in percolation, random graphs, and subgraph counts (Goldstein et al., 2015, Baron et al., 2019).
• Statistical Mechanics Models: Control of correlation and coexistence probabilities in Ising-type and Curie–Weiss systems, including cluster structures (Berg et al., 2012).
• KPZ Line Ensembles and Last Passage Percolation: BK-type inequalities for disjoint polymers or lines, allowing sharp analysis of upper tail probabilities and scaling limits in the KPZ universality class. Integrability is crucial for their validity in positive-temperature polymer models (Ganguly et al., 19 Dec 2025).
• Decision Tree and Adaptive Algorithms: Decision-tree variants enable adaptivity and revealment arguments for improved bounds in percolation and connectivity estimates in networks (Gladkov, 2024).

7. Methodological Foundations and Outlook

Reimer’s proof of the BKR inequality in the Boolean cube, via combinatorial switching and involution arguments, laid the groundwork for subsequent functional and structural generalizations (Goldstein et al., 2015). The monotone class argument extends the result to arbitrary non-negative measurable functions, and the "folding" procedure enables extensions to non-product and Gibbs measures (Berg et al., 2012).

Measurability complexities in infinite-product spaces are addressed via descriptive set theory and approximation by finite-atom sigma-algebras (Arratia et al., 2015). Abstract random-cluster representations and folding techniques have unified generalizations across dependent measures and cluster-based models (Berg et al., 2012).

The BKR framework continues to inspire ongoing developments in negative dependence, stochastic domination, random structure analysis, and applications to mathematical physics. Its functional and dual forms remain central tools for bounding probabilities of complex composite events in product and dependent settings.