Allen’s Interval Algebra Overview

Updated 3 March 2026

Allen’s Interval Algebra is a framework for qualitative reasoning that defines 13 unique relationships between temporal intervals.
It supports complex constraint networks through methods like path-consistency and backtracking to ensure consistent temporal assignments.
Recent advancements using dynamic programming and sublinear partitioning improve algorithmic bounds, impacting scheduling, planning, and temporal databases.

Allen’s Interval Algebra (IA) is a foundational formalism for qualitative reasoning about temporal intervals and their relationships. Defined originally by J.F. Allen in 1983, IA encodes temporal information using a fixed, jointly exhaustive, pairwise disjoint set of atomic relations over closed intervals on a linear order such as the real line or rationals. Its expressiveness and combinatorial complexity make it a central object in artificial intelligence, constraint reasoning, theoretical computer science, and the study of interval temporal logics.

1. Formal Structure and Relations

Allen’s Interval Algebra considers the universe of all (closed) intervals $I = [l, r]$ on a linear order (typically $\mathbb{R}$ or $\mathbb{Q}$ ), with endpoints $l < r$ . For any two intervals $I_i = [l_i, r_i]$ and $I_j = [l_j, r_j]$ , IA defines 13 atomic relations, ensuring that exactly one holds for each pair (JEPD):

Abbr.	Name	Defining Conditions
p	precedes	$r_i < l_j$
pi	preceded-by	$r_j < l_i$
m	meets	$r_i = l_j$
mi	met-by	$r_j = l_i$
o	overlaps	$l_i < l_j < r_i < r_j$
oi	overlapped-by	$l_j < l_i < r_j < r_i$
s	starts	$l_i = l_j, r_i < r_j$
si	started-by	$l_i = l_j, r_j < r_i$
d	during	$l_j < l_i < r_i < r_j$
di	contains	$l_i < l_j < r_j < r_i$
f	finishes	$l_i > l_j, r_i = r_j$
fi	finished-by	$l_j > l_i, r_i = r_j$
eq	equals	$l_i = l_j, r_i = r_j$

Every pair of intervals is related by exactly one of these 13 relations. Each relation is closed under composition defined by a 13×13 table, capturing the possible relations that can arise from chaining two atomic relations.

2. Constraint Networks and Reasoning Tasks

A constraint network in IA consists of variables, each ranging over intervals, and binary constraints specifying possible relations (possibly disjunctions) between pairs. The main reasoning task is the consistency problem: given a network, does there exist an assignment to variables such that all constraints are satisfied?

Algorithmic primitives in IA include:

Path-consistency enforcement: Iteratively refine constraints via weak composition, i.e., for variables $i,j,k$ ,

$C(i,k) := C(i,k) \cap (C(i,j) \circledast C(j,k)),$

with $\circledast$ the weak composition operator. For certain tractable subclasses (e.g., ORD-Horn networks), path consistency suffices for global consistency (Janhunen et al., 2019).

Backtracking and branching: In general, since path-consistency is incomplete for arbitrary IA networks, global search and backtracking are required to resolve ambiguity in disjunctive constraints.
Advanced queries: Modern approaches handle backbone computation (relations holding in all solutions) and “minimum labeling” (relations permitted in some solution), which are more computationally demanding than consistency checking.

3. Algorithmic Complexity and Recent Advances

The general consistency problem for IA is NP-complete. The search space grows rapidly—enumerating all total orders of endpoints yields a naïve upper bound of $O^*((2n)^{2n})$ for $n$ variables (Eriksson et al., 2023). Recent research has driven fine-grained improvements:

Previous best: $O^*((1.0615n)^n)$ , based on more efficient enumeration and data structures.
Dynamic programming with sublinear partitioning: A new technique encodes only a sublinear number of “holes” (blocks) to insert interval endpoints, maintaining the feasibility of constraint recursive extensions. This yields an upper bound:

$O^*\left( \left( c n / \ln n \right)^n \right)$

which, for any constant $c'$ , asymptotically outperforms $(c' n)^n$ (Eriksson et al., 2023).

The lower bound under ETH is only $2^{o(n)}$ , leaving the exponential constant gap open.

Algorithmic Method	Runtime Bound
Naïve Enumeration	$O^*((2n)^{2n})$
DP/Refinements (Eriksson, 2021)	$O^*((1.0615n)^n)$
Sublinear Partitioning (recent)	$O^*\left((cn/\ln n)^n\right)$

A plausible implication is that further progress on single-exponential IA algorithms likely requires significant new combinatorial insights.

4. Constraint Satisfaction and Parameterized Complexity

For parameterized reasoning, MinCSP problems over IA ask whether up to $k$ constraints can be deleted to restore satisfiability. A complete dichotomy is known (Dabrowski et al., 2023):

MinCSP $(\Gamma)$ is in FPT iff $\Gamma \subseteq \{m, p\}$ or $\Gamma \subseteq \{r_1, r_2, eq\}$ for $r_1 \in \{s, f\}, r_2 \in \{p, o, d\}$ .
Otherwise, W[1]-hardness holds.
The tractable cases correspond to reductions to directed feedback arc set and related cut problems.

Approximation results show that MinCSP $(A)$ has a factor-2 fpt-approximation in time $O^*(2^{O(k^3)})$ and a factor-4 fpt-approximation in $O^*(2^{O(k)})$ , via endpoint separation (Dabrowski et al., 2023).

Halpern–Shoham’s modal interval logic (HS) formalizes IA as modal operators over the 13 Allen relations. Many fragments of HS have undecidable satisfiability. However, certain expressive fragments are maximal decidable:

The fragment $\mathcal L(A,B,\overline B,\overline L)$ (meets, begins, begun-by, before) is decidable over strongly-discrete orders (finite, $\mathbb N$ , $\mathbb Z$ ). Its satisfiability problem is EXPSPACE-complete—achieved by bounding the model size using “compass” structures and contraction (Bresolin et al., 2010).
Any further addition of Allen relations leads to undecidability except on finite domains.
This locates a sharp frontier between the decidable and undecidable interval temporal logics for interval-based CSPs.

6. Encodings and Practical Implementations

Modern encodings of IA include placement into Answer-Set Programming with Difference Logic (ASP(DL)), as exemplified in Clingo-DL (Janhunen et al., 2019). Intervals map to integer variables, and each of the 13 relations corresponds to a small system of difference constraints (e.g., $r_i < l_j$ for precedes). This approach:

Supports disjunctive constraints and leverages difference-logic propagation to enforce both satisfaction and path-consistency.
Enables advanced queries (backbones, minimum labeling) not efficiently addressable by native IA solvers.
Is competitive for networks up to $n=100$ intervals and moderate density, though specialized QSTR tools remain faster for pure satisfiability tasks.

7. Applications and Extensions

Allen’s Interval Algebra is widely adopted in:

Temporal planning and scheduling: Formulating action and fluent concurrency in domains including robotics and autonomous systems, often abstracted further in constraint programming formalisms (Ramirez et al., 2024).
Temporal databases, natural language semantics, molecular biology: As a general model for qualitative temporal relations.
Qualitative spatial reasoning: Sublinear partitioning methods (originally for IA) extend to calculi such as the Cardinal Direction Point Algebra, RCC-8, and oriented point relation algebra (Eriksson et al., 2023).

The algebra’s tractable fragments, decidability boundaries, and parameterized complexity discoveries guide both the design of efficient temporal reasoning engines and the practical handling of intractability in qualitative time. Open questions include the existence of truly single-exponential algorithms for full IA, extension of sublinear partitioning to broader calculi, and further refinement of the modal logic decidability lattice.

References: (Eriksson et al., 2023, Dabrowski et al., 2023, Janhunen et al., 2019, Bresolin et al., 2010, Ramirez et al., 2024)