Space-Filling Design¶
Method — instantiates Dense-Subset Coverage Design
Places design points across a multidimensional domain to reduce large uncovered regions.
A space-filling design chooses a modest number of points so they spread as evenly as possible across a whole multidimensional domain, leaving no large empty pocket. Its defining commitment is global uniformity rather than a per-point certificate: it does not promise "no point is farther than ε from an anchor" the way an epsilon-net does — it optimizes an aggregate spread criterion (maximize the minimum inter-point distance, or minimize the discrepancy between the point cloud and a perfectly uniform one) so that, on the whole, the domain is sampled without clumps or voids. Its second signature is that a good space-filling design can be built to be extensible: draw the points from a sequential low-discrepancy sequence and the first ten, the first hundred, and the first thousand are each well-spread on their own, so the cover can be deepened layer by layer without discarding the earlier work.
Example¶
An engineering team is calibrating a digital twin of a wind-turbine gearbox. The twin's behavior depends on four continuous inputs — torque load, oil temperature, gear-mesh ratio, and rotor speed — and each simulated run takes most of a day on a cluster, so they can afford roughly sixty runs before a design review. Testing each input at a few fixed levels would pile runs onto a coarse lattice and still leave whole diagonal swaths of the operating envelope unsampled. Instead they lay down a sixty-point space-filling design: a Latin-hypercube layout, optimized so the closest pair of points is as far apart as possible, that scatters the runs across the four-dimensional box with no two conditions nearly identical and no large region untouched. When the review asks for more resolution near high-load operation, they extend the same sequence with sixty more points that interleave with the first sixty rather than replacing them — the cover simply gets denser everywhere, fastest where they steer it.
How it works¶
- Pick a spread criterion. Maximin distance pushes points apart so the largest void shrinks; low-discrepancy (quasi-random) construction makes the cloud statistically indistinguishable from uniform; Latin-hypercube construction additionally guarantees even spacing in every single-axis projection. The criterion is the design's notion of "close," standing in for the distance relation.
- Generate and optimize. Produce a candidate set and improve it — swap points, run a maximin optimizer, or advance a Sobol/Halton sequence — until the criterion stops improving materially.
- Layer it. Because a sequential sequence is well-spread at every prefix length, the same construction yields nested resolutions: a coarse cover you can thicken on demand without re-solving from scratch.
Tuning parameters¶
- Point budget — how many anchors you can afford. More points shrink the largest gap but cost proportionally more to evaluate and maintain; the budget is usually fixed first and everything else optimized within it.
- Spread criterion — maximin vs. low-discrepancy vs. projection-balanced (Latin hypercube). Maximin protects against the single worst void; projection balance protects against a factor being under-varied; they occasionally pull against each other.
- Sequential vs. one-shot — a nested sequence buys cheap refinement later at the price of slightly worse spread at any fixed count than a design optimized for that exact count.
- Active dimensions — which inputs to fill. Dropping or collapsing inputs that don't move the outcome keeps the design feasible; spending points on inert dimensions wastes the budget.
- Boundary handling — whether to force points onto edges and corners, which pure interior-spread designs tend to starve.
When it helps, and when it misleads¶
Its strength is economy of exploration: with a fixed, affordable number of expensive evaluations it leaves no large region of the domain unseen, which is exactly what you want before you know where the interesting behavior lives. The nested-layer property makes it forgiving — you can start cheap and deepen only where a first pass warrants it.
It misleads when "evenly spread in the coordinates" is not "evenly spread in what matters." A design that is beautifully uniform in raw inputs can be badly clumped after a nonlinear transform[1], so uniformity measured by star discrepancy on the input box can coexist with gaping holes in the response. The classic misuse is trusting a low-discrepancy count as if it were an epsilon-net's guarantee: space-filling optimizes the average void, not the worst one, so a single stubborn empty pocket can survive a design that scores well overall — and in high dimensions the point budget needed for genuine fill grows faster than any budget can follow. The guarding discipline is to reduce dimension to the inputs that actually drive the outcome, choose the spread metric in the space where behavior varies, and hand the worst-case-gap question to a mechanism built to answer it.
How it implements the components¶
dense_reference_subset— the design is the anchor set: a spread cloud of design points chosen to leave no large void.distance_or_neighborhood_relation— the spread criterion (maximin distance, discrepancy) supplies the operative notion of nearness the points are optimized against.multi_resolution_anchor_layers— sequential low-discrepancy construction yields nested covers, each prefix well-spread, so resolution can be added in layers.
It does not measure and display where the residual gaps fall (that is Coverage Heatmap), deliberately densify thresholds and transitions (that is Boundary-Value Test Suite), or convert its spread into a hard covering-radius guarantee (that is Epsilon-Net or Covering Grid).
Related¶
- Instantiates: Dense-Subset Coverage Design — it supplies the evenly-spread anchor set the rest of the design assesses, routes over, and maintains.
- Sibling mechanisms: Epsilon-Net or Covering Grid · Boundary-Value Test Suite · Coverage Heatmap · Nearest-Neighbor Assignment Rule · Adaptive Refinement Loop · Sensor or Service Radius Map · Anchor Case Library
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Space-Filling Design operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it places design points across a multidimensional domain to reduce large uncovered regions.
Independent corroboration: The frozen evidence defines Space-Filling Design as 'Places design points across a multidimensional domain to reduce large uncovered regions', so its operative form is Analysis, Modeling & Optimization.
Nearest alternative: Representation, Specification & Plan — Space-Filling Design includes features of a static representation, map, specification, schema, or prospective plan that externalizes information, but its defining operation is an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Distributing experimental points to cover a multidimensional region with minimal gaps is canonical space-filling design of experiments.
Related originating lineages:
- Data Science & Analytics — Active sampling benefits from broad domain coverage.
- Engineering & Design — Computer experiments use these designs for surrogate modeling.
- Mathematics — Discrepancy, distance, and packing criteria formalize coverage.
- Operations Research — Operations research, optimization, and queueing analysis supplies a parallel or contributing lineage for the mechanism's defining operation: places design points across a multidimensional domain to reduce large uncovered regions.
Review resolution: The blind reviewers agree that statistics_experimental_design is the primary origin and differ only on alternate origin disagreement, domain reach disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain single_lineage because the combined evidence shows one traceable formative lineage. The broader reach of multi_domain records portability separately from historical provenance; encyclopedia_synthesis=false preserves the affirmative synthesis judgment where either reviewer identified one.
Review outcome: Reconciled after independent review; high confidence.
References¶
[1] Li, Y., Kang, L., and Hickernell, F. J. "Is a Transformed Low Discrepancy Design Also Low Discrepancy?". In Contemporary Experimental Design, Multivariate Analysis and Data Mining. Springer (2020). Shows that transforming a low-discrepancy uniform design can produce a design with large discrepancy for the target distribution. registry ↩