Periortree: Periodic AABB for Spatial Indexing

• Periortree is an extension of the R-Tree data structure that supports periodic boundary conditions using PBC-aware AABB routines to prevent replication.
• It adapts traditional axis-aligned bounding box operations with minimal-image computations, ensuring accurate retrievals across periodic boundaries.
• The method achieves smaller bounding volumes, reduces node visits by up to 40% in simulations, and readily integrates into existing spatial indexing systems.

Periortree is an extension of the classical R-Tree data structure explicitly designed for spatial data indexing under periodic boundary conditions (PBCs). The method enables efficient insertion, splitting, and querying of spatial objects in periodic domains without object or query replication. By adapting axis-aligned bounding box (AABB) routines to the minimal-image convention and allowing AABBs to wrap around boundaries, Periortree achieves significantly tighter bounds and improved search performance, making it especially suitable for scientific simulations involving PBCs, such as molecular dynamics (Niina, 2017).

1. Foundations and Motivation

Standard R-Trees store spatial objects in a height-balanced tree where each node is associated with an AABB, usually assumed to exist within an infinite or hard-wall (non-periodic) domain. Scientific applications frequently use domains with periodic boundary conditions, where, for example, a particle leaving one face of a rectangular cell reenters from the opposite face. Existing adaptations to support PBC in R-Trees face severe limitations: "copy-the-world" methods replicate every object $3^D$ times (where $D$ is the domain dimension), causing exponential storage blow-up, while "copy-the-query" methods replicate the query when it crosses boundaries, missing finite-size objects that straddle a boundary.

Periortree addresses these drawbacks by allowing AABBs to straddle periodic boundaries, using a center–radius form and minimal-image computations. This reduces bounding box volumes and obviates the need for object/query replication, providing an efficient and storage-optimal solution for periodic spatial indexing.

2. Mathematical Framework for Periodic AABBs

Periortree operates in $D$-dimensional periodic rectangular domains with side lengths $L=(L_1,\dots,L_D)$. Spatial coordinates are folded into the unit cell:

$x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$

AABBs are represented as $(c, r)$, with $c$ (center) and $r$ (radius) as $D$-vectors:

$$R = (c, r), \quad c \in \mathbb{R}^D, \; r \in \mathbb{R}^D_{\ge 0},$$

covering all $D$0 for which $D$1 (under minimal image) $D$2.

Minimal-Image and Folding Operations

• Minimal-image vector reduction:

$D$3

• Folding a point into the unit cell:

$D$4

Periodic AABB Operations

• Intersection test of $D$5 and $D$6:

$D$7

• Containment test:

$D$8

• Expanding $D$9 to contain $D$0:
  1. Compute shortest periodic offset $D$1.
  2. Unwrap $D$2 into $D$3's frame, calculate union's min/max per coordinate.
  3. The new center and radius:

  $D$4

where $D$5 and $D$6 are the minima and maxima of lower and upper faces in the unwrapped frame.

This guarantees the smallest PBC-aware AABB encapsulating both boxes.

3. Core Algorithms

Periortree functions by replacing AABB routines in any R-Tree with their periodic versions.

Insertion and Splitting

The main insertion routine involves:

• Traversing the tree to find the best leaf, using the minimal increment in periodic volume upon expansion.
• Inserting the object AABB.
• If node overflow occurs, apply Guttman’s quadratic split, now using PBC-aware bounding box expansions.

Pseudocode for insertion: $(c, r)$3

Range and Nearest-Neighbor Queries

• RangeQuery proceeds identically to standard R-Tree, except every intersection test uses the periodic minimal-image formulas: $(c, r)$4
• NearestNeighborQuery uses best-first search, with distances computed via periodic minimal-image:

$D$7

with all distances and box operations performed under PBC.

No modifications to query logic, tie breaking, or split heuristics are otherwise needed—only the bounding box operations are replaced.

4. Periodic versus Standard Bounding Strategies

A key distinction of Periortree is the treatment of bounding volumes:

Strategy	Box Expansion	Typical Volume	Node Visits
Standard R-Tree	Non-periodic	$D$8 (when objects straddle boundary)	Many
Periortree (PBC-aware)	Minimal-image	$D$9 (tight)	Fewer

Standard AABB expansion determines the maximum span ignoring periodicity:

$L=(L_1,\dots,L_D)$0

whereas the Periortree always selects the minimal periodic span:

$L=(L_1,\dots,L_D)$1

Periortree thus often yields much smaller bounding boxes, especially when objects are near opposite faces, resulting in deeper node fills and substantially fewer node accesses during search.

5. Computational Complexity

Let $L=(L_1,\dots,L_D)$2 be the number of stored objects, $L=(L_1,\dots,L_D)$3 the node capacity, and $L=(L_1,\dots,L_D)$4 the space dimension. The complexity characteristics are:

• Time Complexity
  • Insertion: $L=(L_1,\dots,L_D)$5, due to $L=(L_1,\dots,L_D)$6 node adjustments, each requiring up to $L=(L_1,\dots,L_D)$7 AABB computations, each $L=(L_1,\dots,L_D)$8 for minimal-image math.
  • Range query: $L=(L_1,\dots,L_D)$9 worst-case, but typically $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$0 if bounding boxes are tight; periodic overhead is only $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$1 per node.
  • Nearest-neighbor: $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$2 on average, with every distance using minimal-image convention.
• Space Complexity
  • Standard R-Tree: $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$3 nodes, $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$4 entries per node.
  • Periortree: $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$5, no $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$6 replication. Each entry stores (center, radius) $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$7-vectors in lieu of (min, max).

Periortree’s storage remains on par with the standard R-Tree, sharply contrasting with naive $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$8 replication, which requires $x_i \mapsto x_i \bmod L_i, \quad i=1,\ldots, D.$9 entries. On average, Periortree reduces node visits per query due to smaller, tighter indexes.

6. Experimental Observations

Benchmarking on point sets and molecular dynamics workloads reveals several empirical outcomes (Niina, 2017):

• Bounding Box Volume: When two points are near opposite faces, standard R-Tree boxes expand to cover nearly the whole cell ($(c, r)$0); Periortree boxes only span the periodic gap with volume $(c, r)$1 (Figures 3A, 3B).
• Query Correctness: Range queries crossing periodic boundaries retrieve the exact same objects as queries near boundaries inside the unit cell, with no missing objects (Figure 3C).
• Performance:
  • Memory usage remains equal to that of standard R-Tree.
  • Node visits are reduced by 10–40% in typical 3D molecular-dynamics workloads, which corresponds to observable speedups in neighbor-list construction.
  • Outperforms "copy-the-world" and "copy-the-query" approaches in both speed and accuracy, as it completely avoids both $(c, r)$2 blow-up and missed finite-size objects.

The implementation, as well as further documentation and correctness proofs, are available via the GitHub repository provided by the author.

7. Implementation Aspects and Applicability

Periortree is readily implementable in any R-Tree–derived spatial indexing library by substituting the standard AABB operations with their PBC-aware analogues. No change is required to core splitting algorithms or query logic; all modifications are strictly localized to bounding box management routines. The method is not limited to R-Trees but extends to any spatial data structure relying on axis-aligned bounding boxes in periodic domains. The simplicity of the required changes facilitates adoption in simulation codes and spatial database applications that require rigorous support for periodicity (Niina, 2017).