Space-Filling-Curve Layouts

• Space-filling-curve layouts are recursive, locality-preserving mappings from a one-dimensional interval to multi-dimensional domains, used to linearize data while maintaining spatial relationships.
• They underpin numerous applications such as scientific computing, database indexing, parallel load balancing, and matrix algorithms, demonstrated by Hilbert, Morton, and Peano curves.
• Choosing an SFC layout involves balancing locality, bounding-box quality, and computational complexity, with optimal performance found in Hilbert and balanced Peano variants.

A space-filling curve (SFC) layout is a recursive, locality-preserving mapping from a one-dimensional parameter space—typically the unit interval [0,1]—onto a d-dimensional domain such as a square, cube, or simplex, with the property that the curve is surjective and continuous in the parameter but generally not invertible. SFC layouts are foundational to spatial data structures, scientific computing, visualization, matrix algorithms, parallel load balancing, database indexing, and high-dimensional statistics due to their ability to linearize multidimensional data while approximately retaining neighborhood relationships. The choice and analysis of an SFC layout critically impact locality, cache efficiency, bounding-box clustering, update/query complexity, and adaptation to data distribution.

1. Mathematical Foundations of Space-Filling Curves

A space-filling curve is a map $F:[0,1]\to[0,1]^d$ that is continuous and surjective. Early constructions include Peano’s base-3 subdivision, Lebesgue’s bit-interleaved “Z-order,” and Hilbert’s recursive dyadic mapping, all of which are now canonical. In multidimensional discrete settings, SFCs operationalize as bijections between $\{0,\ldots,N-1\}^d$ and $\{0,\ldots,N^d-1\}$ determined by recursive serpentine or bit-interleaving (Morton) rules, Gray-coding, or more recent substitution systems such as the Fibonacci SFC (Kanungo, 2024, Ozkaraca, 2024, Jaffer, 2014).

A representative formalism for constructing such layouts involves defining a finite set of states and a Hamiltonian path $H$ over the base grid $(\mathbb{Z}\bmod s)^d$. Let $H(t):\{0,\ldots,s^d-1\}\to\{0,\ldots,s-1\}^d$ denote the path, with a corresponding reflection/permutation operation $A(V,t,w)$ to map subcells recursively (Jaffer, 2014). For bit-based SFCs (Hilbert, Morton), the layout is specified by forming a $d\cdot\omega$-bit index from the coordinates: for 2D Morton,

$$\pi(x,y) = \sum_{i=0}^{\omega-1} \left(x_i2^{2i} + y_i2^{2i+1}\right).$$

Hilbert indices use recursive quadrant orientation (Gray-code rotations) at each level (Kanungo, 2024, Abedi et al., 2 Dec 2025).

2. Locality, Clustering, and Bounding-Box Quality

The principal analytic desideratum for an SFC layout is locality: neighboring points in $\mathbb{R}^d$ are mapped to nearby values in the one-dimensional SFC parameter, and curve segments map to spatially coherent subsets. Locality is quantified via worst-case $\{0,\ldots,N-1\}^d$0-separation over curve segments ($\{0,\ldots,N-1\}^d$1 in (0806.4787)), and bounding-box quality is measured by the ratio $\{0,\ldots,N-1\}^d$2.

Empirically and theoretically (0806.4787, Jaffer, 2014):

• Hilbert curves achieve optimal or near-optimal locality: worst-case $\{0,\ldots,N-1\}^d$3 locality is 6, and bounding-box area ratio is 2.40.
• Morton/Z-order is the simplest to compute but has unbounded worst-case locality.
• Balanced Peano variants can achieve $\{0,\ldots,N-1\}^d$4 but trade off with higher locality ratios.
• Sierpinski–Knopp has provably optimal $\{0,\ldots,N-1\}^d$5 and $\{0,\ldots,N-1\}^d$6 locality ($\{0,\ldots,N-1\}^d$7) but $\{0,\ldots,N-1\}^d$8.

Best all-rounders for R-tree spatial layouts are Hilbert and balanced Peano (GP) variants; Z-order should be avoided if worst-case clustering matters (0806.4787).

3. Algorithms and Complexity for Index Conversion and Navigation

Efficient index conversion—mapping from multidimensional coordinates to SFC rank and back—is fundamental for performance-critical applications such as cache-oblivious matrix algorithms and spatial indexing (Salem et al., 2016).

Curve	Encode/Decode Complexity	Lookup Tables	Memory Footprint
Morton (Z)	$\{0,\ldots,N-1\}^d$9 bit ops	None	Minimal
Hilbert	$\{0,\ldots,N^d-1\}$0 bit ops	$\{0,\ldots,N^d-1\}$1 (2D)	Small (tens of ints)
Peano	$\{0,\ldots,N^d-1\}$2	$\{0,\ldots,N^d-1\}$3	Small
Morton-Hybrid	$\{0,\ldots,N^d-1\}$4	None	Small

Morton and Morton-hybrid permit constant-time conversion via bit-dilation; Hilbert and Peano incur logarithmic-time table-driven state machine transitions, with Peano’s ternary arithmetic imposing higher per-conversion cost (Salem et al., 2016). For non-cuboidal meshes, variants such as the tetrahedral Morton index are defined by interleaving anchor coordinates with ancestor type codes, yielding constant-time navigation and optimal cache/block locality for adaptive mesh refinement routines (Burstedde et al., 2015).

Neighbor finding on SFC layouts is handled via three precomputed tables for regularity-preserving SFCs (Morton, Hilbert, Peano, Sierpinski), achieving average-case $\{0,\ldots,N^d-1\}$5 per query (Holzmüller, 2017).

4. SFCs in Data Structures, Indexing, and Query Optimization

SFC layouts underpin one-dimensional linearizations for multi-dimensional spatial data structures. Sorting points by Morton or Hilbert code and storing them contiguously is a simple yet highly performant approach for range and distance queries (Berg et al., 27 Aug 2025). For a distance query:

• Scatter $\{0,\ldots,N^d-1\}$6 by SFC keys, sort.
• To answer a ball query $\{0,\ldots,N^d-1\}$7, select up to $\{0,\ldots,N^d-1\}$8 curve intervals covering an outer cube of side $\{0,\ldots,N^d-1\}$9 intersecting $H$0, binary search to the corresponding ranges, and scan for true hits.

This method achieves query latency within a small factor of highly tuned trees; its contiguous layout ensures minimal cache misses and maximum scan efficiency. Insertion and update operations in dynamic settings are faster than R-trees, range trees, or kd-trees, as shown by empirical comparison (Berg et al., 27 Aug 2025). Hilbert's improved locality is largely moot for pure scan-based implementations, so Morton/Z-order is typically preferred for minimal key computation overhead.

In database and spatial-indexing systems, SFC-based layouts enable spatial queries and R-tree bounding-box hierarchies with tight clustering: Hilbert and balanced Peano minimize bounding-box area and average bounding-box size (0806.4787). Large-scale “scaled Gray-Hilbert” indices adaptively refine the SFC-tree branches only in dense regions, saving up to 98% of indexing nodes versus a static tree in high-dimensional datasets (Jahn et al., 2019).

Efficient workload-sensitive SFC selection for range/workload queries can be achieved via constant-time cost models—enabling reinforcement learning–based or Bayesian search for the optimal bit-merging curve, with large performance gains over fixed-order SFCs (Liu et al., 2023).

5. Adaptivity, Data-Driven Extensions, and Alternative Constructions

Classical SFC layouts are independent of data; however, data-driven SFCs explicitly optimize the order for feature preservation. Flexible MST-Hamiltonian approaches create SFCs that mix data similarity and spatial locality, outperforming Hilbert/Morton in value autocorrelation and feature coherency in visualization and analysis of ensemble or multiscale datasets (Zhou et al., 2020). For adaptive or quadtree/octree hierarchical data, multiscale SFC construction recursively refines cells and preserves local ordering, achieving both value and locality objectives.

Alternative SFCs—including the Fibonacci SFC, which uses inflation rules governed by the golden ratio and admits two rectangle aspect ratios—offer advantages in situations where the data or computational mesh aligns with non-dyadic tilings, e.g., quasicrystals, golden-ratio block hierarchies, or multigrid methods with non-binary refinement (Ozkaraca, 2024).

Simplicial/tetrahedral SFCs extend the Morton logic to arbitrary mesh types, using additional type information to enable contiguous memory layouts, constant-time tree navigation, and minimal storage per element, scalable to hundreds of billions of elements on exascale high-performance computing platforms (Burstedde et al., 2015).

6. Applications: Scientific Computing, Parallelization, and Indexing

SFC layouts are foundational to parallel scientific computing, cache-oblivious matrix algorithms, and high-performance data-intensive systems:

• In parallel load balancing for reservoir and PDE simulations, Hilbert SFCs provide partitions with low boundary surface area, minimal inter-rank neighbor cardinality, and linear scalability (billions of cells, 100k+ ranks), outperforming graph partitioners such as ParMETIS in implementation simplicity, albeit with slightly higher cut volume but better neighbor reduction (Liu et al., 2017).
• In tensor network simulations, e.g., 2D Hubbard model computation, Hilbert SFC mapping of the lattice to a 1D chain minimizes the effective range of mapped nearest-neighbor couplings, drastically reducing matrix product state bond dimension, energy truncation errors, and computational wall time relative to naive row-snake mappings. The Hilbert mapping's near-optimal average 1D separation for 2D nearest neighbors (constant versus $H$1 for snake) enables DMRG-like methods to scale to much larger system sizes (Abedi et al., 2 Dec 2025).
• Google’s S2 system for spherical spatial indexing applies face-by-face Hilbert SFC subdivisions, achieving contiguous linearization with minimal encoding (64 bits per cell), high query throughput, and low page-fault rates (Kanungo, 2024).

7. Open Directions and Practical Considerations

SFC layout selection involves a nuanced trade-off among locality, bounding-box tightness, computational speed of key computation and neighbor querying, and adaptability to data distribution. Theoretical lower bounds establish that, for recursive 2D SFCs, the bounding-box area ratio $H$2, and $H$3-locality $H$4, with balanced GP or Hilbert offering optimal or near-optimal performance (0806.4787).

Adaptive and workload-aware layouts enable further gains over static SFCs, both in space and query-time efficiency (Jahn et al., 2019, Liu et al., 2023, Zhou et al., 2020). For domain-specific layouts (triangular or tetrahedral meshes, golden-ratio refinements) bespoke SFCs deliver the same locality and efficiency guarantees as classical cuboidal SFCs. In high-performance applications, careful implementation (minimal lookup tables, bit manipulation, contiguous memory blocks) is crucial for exploiting the theoretical advantages in practice.

Space-filling-curve layouts thus remain indispensable tools for organizing multi-dimensional data, enabling performance and scalability far beyond direct coordinate-based orderings. Continual innovations—data-driven ordering, adaptive decomposition, and new substitution-based SFCs—expand their reach across scientific, computational, and data-intensive domains (Kanungo, 2024, Abedi et al., 2 Dec 2025, Liu et al., 2023, Ozkaraca, 2024, Liu et al., 2017, Zhou et al., 2020, Jahn et al., 2019, Salem et al., 2016, Burstedde et al., 2015, 0806.4787, Berg et al., 27 Aug 2025, Holzmüller, 2017, Jaffer, 2014).