Multi-Resolution Search¶
Method — instantiates Coarse-to-Fine Search
Implements the archetype by scanning at multiple levels of resolution and escalating detail only where the lower-resolution pass indicates value, uncertainty, or risk.
Multi-Resolution Search keeps not one representation of the space but a stack of them — each level finer than the one below — and moves a candidate window up the stack only when the coarser view says the added detail could change the answer. Its defining move is the explicit ladder of resolutions and the transition rule that governs each rung: a window is promoted to the next finer level only when its coarse score, widened by an uncertainty margin, clears a bar. Detail is never spread evenly across the space; it is spent rung by rung, precisely where the previous rung could not yet decide. This is what separates it from a sibling that runs a single cheap screen and a single deep pass — here the escalation itself is graduated.
Example¶
A search-and-rescue team has eight hours of daylight and a mountain range to cover for a lost hiker last seen in an orange tent. They fly the whole range once and build the coarsest layer: heavily downsampled aerial imagery in which the tent is at most a smudge. A cheap color-and-brightness detector flags every tile with an orange-ish, tent-sized blob — a few hundred of the thousands scanned. Those tiles, and only those, are pulled at the next resolution up, where a sharper detector rejects road cones, autumn foliage, and a kayak on a truck. The dozen survivors of that pass are examined at full native resolution by an analyst.
The point is not that any single detector is good — each is deliberately crude for its level. The point is that the coarse layer never has to be right, only inclusive: its job is to hand a short, uncertainty-padded list up the ladder, so that full-resolution attention lands on twelve tiles instead of ten thousand. The hiker's tent turns out to sit in a tile the coarse pass scored low but the uncertainty margin kept alive.
How it works¶
- Build the pyramid. Represent the space at several resolutions at once, each a fixed factor coarser than the next. The coarsest level must be cheap enough to scan in full.
- Scan the coarsest level exhaustively, scoring every region with a detector matched to that resolution.
- Apply the transition rule per rung. Promote a region to the next finer level when its score plus its uncertainty margin crosses the promotion threshold — not when the score alone does. Everything else is deferred, not deleted.
- Refine only inside promoted regions, re-scoring at the finer resolution with a sharper (and costlier) detector, and repeat up the ladder to native resolution.
The escalation is monotone in cost and selective in target: each rung sees fewer regions than the one below and spends more per region.
Tuning parameters¶
- Number of levels — more rungs give a gentler cost ramp and finer control over where detail lands, but each transition leaks a few true positives; too many rungs multiplies the leak.
- Downsampling factor between levels — a large factor makes the coarse pass cheaper but blurs small targets toward invisibility; a small factor is safer but slower.
- Promotion threshold — how good a coarse score must be to earn the next rung. Loose thresholds protect recall and flood the next level; tight thresholds throttle cost and drop marginal true positives.
- Uncertainty margin width — how far below the threshold a region can score and still be promoted. This is the recall dial: widen it where a miss is costly, narrow it where fine-stage capacity is scarce.
When it helps, and when it misleads¶
Its strength is enormous efficiency without committing to a single crude verdict: expensive resolution is concentrated where cheaper resolution was genuinely undecided, and the ladder gives many chances to catch a target that one screen would have discarded. It is the natural mechanism when the space is huge, roughly continuous, and cheaply summarizable at low fidelity[1] — the same logic as an image pyramid in computer vision.
Its characteristic failure is the small target that falls between coarse samples and never registers at the lowest resolution — an aliasing-style miss the finer levels never get a chance to correct, because a region absent from the coarse list is never promoted. Related is over-trusting a tidy coarse score as if it were a verdict, which is exactly what the uncertainty margin exists to prevent; set that margin to zero and the method degenerates into a single hard screen wearing a ladder as costume. The guarding discipline is to size the downsampling factor and the margin against the smallest target worth finding, and to periodically full-resolution-audit a sample of unpromoted regions to measure what the coarse pass is silently losing.
How it implements the components¶
coarse_representation— the pyramid's lower levels are the coarse representations; the mechanism's premise is that several coexist at once.resolution_transition_rule— the per-rung promotion condition (score + margin ≥ threshold) is the transition rule made explicit and graduated across levels.refinement_step— re-scoring a promoted region with the sharper detector of the next level is the targeted refinement.uncertainty_margin— the width below threshold within which low-scoring regions are still promoted keeps borderline candidates alive up the ladder.
It does not build a durable audit trail of exclusions (coverage_record) or a systematic re-opening of rejected regions (false_negative_check, backtracking_path) — those are the province of Search Tree Pruning with Refinement and Diagnostic Narrowing.
Related¶
- Instantiates: Coarse-to-Fine Search — supplies the multi-level scan-and-escalate engine at the archetype's core.
- Sibling mechanisms: Coarse Grid Search · Funnel Process · Diagnostic Narrowing · Portfolio Screening · Progressive Candidate Review · Design Downselection · Search Tree Pruning with Refinement
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: The mechanism constructs a resolution pyramid, scores regions, and recursively refines only those whose value plus uncertainty crosses a promotion threshold.
Nearest alternative: Decision, Gate & Allocation — Promotion gates allocate search effort, but they are internal steps in the defining multiscale search analysis.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Computer Science & Software Engineering
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Coarse-to-fine search over image pyramids and multiresolution representations is a canonical computer-vision algorithmic pattern.
Related originating lineages:
- Data Science & Analytics — Search and detection pipelines apply the method to large data spaces.
- Mathematics — Multiscale approximation and wavelet ideas provide formal foundations.
- Operations Research — Hierarchical optimization and branch-and-bound contribute selective refinement.
Review resolution: Both independent reviews agree on primary origin computer_science; reconciliation resolves secondary fields (alternate_origin_disagreement, domain_reach_disagreement). Alternate origins retained (mathematics, operations_research, data_science) are the union of reviewer-supported formative lineages with explicit rationales, not a list of later application domains. Present-day breadth is represented separately as domain_reach=multi_domain; origin_mode=single_lineage records the historical relationship among lineages. Confidence is conservatively reconciled to high, and encyclopedia_synthesis=false preserves either reviewer's finding that the encyclopedia generalized the mechanism.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The method only works when coarse scores are predictive of fine scores — the correspondence between adjacent rungs is an invariant, not a convenience. When a downsampling factor breaks that correspondence (the coarse detector responds to something the fine detector ignores, or vice versa), the ladder promotes the wrong regions and the efficiency gain becomes an efficiency of error.
References¶
[1] Burt, P. J., & Adelson, E. H. "The Laplacian Pyramid as a Compact Image Code". IEEE Transactions on Communications 31(4), 532–540 (1983). Uses the low-fidelity summarization logic of an image pyramid in computer vision. registry ↩