Traditional temporal reasoning frameworks operate under a fundamental assumption: time intervals are convex. A convex interval is a single, continuous, gap-free block of time. For example, being "at lunch from 12:00 PM to 1:00 PM" is a convex interval.
However, real-world events rarely fit into neat, continuous boxes. Consider a recurring event like "a class that meets every Tuesday" or an interrupted process like "working on a research paper, paused for a coffee break." As Ladkin noted in Primitives and Units for Time Specification, treating these fragmented occurrences as a single massive interval introduces logical falsehoods (e.g., implying the class meets on Wednesdays too). Conversely, treating them as entirely unrelated intervals destroys their semantic unity. Early formal proofs of completeness in temporal logic, such as the one demonstrated by Ladkin in The Completeness of a Natural System for Reasoning with Time Intervals, were explicitly restricted to convex intervals, leaving a significant gap in an AI's ability to handle real-world scheduling and planning.
To model the real world accurately, temporal reasoning had to evolve to handle non-convex intervals—periods of time that contain internal gaps.
Decomposing Gaps into Maximal Sub-Intervals
The core challenge of non-convex temporal reasoning is maintaining mathematical tractability. If time can break and resume arbitrarily, how do we prevent the relationships between intervals from becoming too complex to compute?
The foundational insight is to treat a non-convex interval not as an arbitrary set of points, but as a structured collection of its continuous parts. Early work by Kandrashina in Representation of Temporal Knowledge introduced the concept of a "chain"—an ordered sequence of time intervals that behaves as a single logical entity. This allowed systems to track a discontinuous state without losing its cohesive identity.
Ladkin formalized this mechanics in Time Representation: A Taxonomy of Internal Relations by representing discontinuous time periods as a union of maximal continuous sub-intervals. By decomposing a gappy interval into its constituent convex parts, researchers realized they could reuse classical interval logic on the sub-components while maintaining a single overarching logical object.
This concept eventually matured into the definition of n-intervals. As explored by Ligozat in Weak Representations of Interval Algebras and On Generalized Interval Calculi, an n-interval is an ordered sequence of convex sub-intervals. This abstraction accommodates an arbitrary or even unspecified number of occurrences and internal gaps, providing a robust framework for common-sense temporal reasoning.
Composition and Calendar Logic
Building a representation is only useful if a system can construct and manipulate these non-convex structures from natural language or domain rules.
Leban et al. addressed this in A Representation for Collections of Temporal Intervals by introducing compositional operators like slicing and dicing. These operators allow a system to take a broad, continuous interval (like "the year 2024") and break it down into non-convex collections (like "the first Monday of every month") by systematically carving out the gaps.
This becomes highly practical when dealing with real-world calendars. As Hafner demonstrated in Semantics of Temporal Queries and Temporal Data, standard continuous timelines fail to handle domain-specific temporal structures like business weeks or holiday schedules. Non-convex logic allows a system to explicitly exclude non-working days from duration calculations and relative date references, treating the "business calendar" as a non-contiguous sequence of valid time blocks rather than a single unbroken line.
Maintaining Computational Tractability
When moving from simple convex intervals to complex networks of n-intervals, a critical question arises: does the computational cost explode? In temporal databases, constraint satisfaction algorithms—specifically path consistency—are used to ensure that a set of temporal assertions contains no contradictions.
The breakthrough for non-convex intervals came from showing that generalized interval relations could still be computed efficiently. In Path Consistency in a Network of Non-Convex Intervals, Morris et al. proved that path-consistency algorithms can determine the consistency of an n-interval network in $O(k^3 m^3)$ time, where $k$ is the number of intervals and $m$ is the number of components. A parallel formulation in Path Consistency in a Network of Non-convex similarly demonstrated that constraint satisfaction networks remain a viable and efficient tool for managing non-convex temporal knowledge bases.
By mapping complex, first-order constraints into simpler underlying structures—such as dense linear orders, as shown by Ladkin in Satisfying First-Order Constraints About Time Intervals—the field proved that allowing gaps in time does not require sacrificing mathematical rigor or algorithmic efficiency. The logic works because the relationships between non-convex intervals ultimately reduce to a bounded combination of relationships between their convex parts.
Key Foundational Work
Representation of Temporal Knowledge (1983) — contributed 3 concept elements to this direction.
Semantics of Temporal Queries and Temporal Data (1985) — contributed 1 concept elements to this direction.