Locality Of Reference¶
Core Idea¶
Locality of reference is the empirical structural regularity that accesses, events, or interactions are not spread uniformly across their possible targets but cluster: recently-used items are disproportionately likely to be used again soon (temporal locality), and items near a used one are disproportionately likely to be used next (spatial locality). The concept emerged in computer architecture from the observation, formalized by Denning (1968) in the working-set model, that a running program touches only a small, slowly-drifting subset of its address space at any instant rather than ranging freely over all of memory.[1] The pattern is fundamentally a statement about the distribution of access and not about any single access: usage is autocorrelated in time and space, so that knowing what was just referenced sharply narrows the prediction of what will be referenced next.[2]
What makes locality a prime rather than a mere observation about computers is that the same distributional shape recurs wherever a population of agents draws repeatedly from a space of targets. The reason a fast-but-small local store can stand in for a slow-but-large source is never that the local store is complete; it is that the access distribution is concentrated, so a tiny resident subset captures the overwhelming majority of demand. Locality names the precondition that makes any "keep the hot subset close" strategy worth attempting, and equally names its absence as the diagnostic that such a strategy will fail.[2]
How would you explain it like I'm…
Stuff Near Stuff Gets Used Together
Recently-Used and Nearby Stuff Repeats
Access Clustering in Time and Space
Structural Signature¶
Locality of reference encodes a structural pattern: a stream of accesses over a target space → autocorrelation in time and space → concentration of demand into a small, drifting working set. It separates two regimes (uniform/random access versus clustered/autocorrelated access) and names the property of the access stream that distinguishes them. The signature is purely distributional: it says nothing about what is accessed or why, only that the empirical histogram of references is peaked and that the peaks move slowly and predictably.
Recurring features:
- Accesses cluster rather than spreading uniformly across targets
- Recently-used items are likely to be used again soon
- Items near a used one are likely to be referenced next
- A small working set captures most of the demand
- The access distribution is concentrated, not flat
- Usage is autocorrelated in time and in space
- The hot subset drifts slowly and predictably over time
The structural insight is robust across substrate: a CPU touching cache lines, a city whose interactions decay with distance, an epidemic spreading through households, a library whose circulation concentrates on a handful of titles, and a word stream where reuse is bursty all exhibit the same peaked-and-autocorrelated access shape.[2] The strength of locality is a measurable quantity, not a binary: a workload can have strong, weak, or no temporal locality, and the same stream can be highly local in space but not in time, or vice versa, which is why the two axes are treated as separable dimensions of a single distributional pattern.[3]
What It Is Not¶
Locality of reference is not a guarantee about any individual access. It makes no claim that the next reference will fall in the hot subset, only that it is far more likely to than a uniform draw would predict. A locality argument is therefore probabilistic and aggregate; betting that a particular access will hit the working set is a misuse, and locality offers no comfort against worst-case access patterns crafted to defeat it.[3]
It is also not the same as the existence of a hot subset in a fixed snapshot. Locality requires that the concentration persist and drift slowly over the access stream, not merely that some items are popular at one instant. A distribution that is concentrated but whose peaks jump erratically from moment to moment has high skew but poor temporal locality, because nothing learned from the recent past predicts the near future. The drift must be slow enough that the working set remains a useful predictor across the timescale at which a local store is maintained.
Locality is not a causal mechanism and does not explain why accesses cluster. Clustering may arise from program loops, from physical adjacency, from social network structure, from contagion dynamics, or from human attention; locality is agnostic to all of these. It describes the shape of the resulting access distribution, not its generator. Treating locality as if it were itself the cause ("the cache works because of locality") collapses a property of the workload into an explanation, when the real explanatory work is done by whatever process produced the autocorrelation.
Finally, locality does not assert that the whole space is small or that most of it is unused permanently. The full address space, the full set of geographic targets, or the full catalog may be enormous and may all eventually be touched. Locality claims only that at any given window the active subset is small relative to the whole, and that windows are correlated. Over a long enough horizon, demand may spread broadly even while remaining locally concentrated at every instant.
Broad Use¶
Computer science: Memory, cache, and disk accesses cluster, which is the reason caches, prefetching, paging, and the entire memory hierarchy work at all; the working-set model and the design of replacement policies (LRU and its approximations) are direct exploitations of temporal and spatial locality.[3]
Geography (non-obvious): Tobler's first law of geography — "everything is related to everything else, but near things are more related than distant things" — is spatial locality of interaction, and it underwrites spatial autocorrelation statistics, distance-decay models, and the design of spatial indexes.[4]
Epidemiology: Contacts and infections cluster spatially and temporally (households, neighborhoods, recent exposure windows), so that case locations and times are autocorrelated rather than uniformly scattered; this is precisely why ring vaccination, local containment, and contact tracing are effective interventions.[5]
Library and information science: A small set of recently and frequently consulted works dominates circulation, and reserve collections, reshelving heuristics, and open-stack layout all exploit the fact that the next request is disproportionately likely to be for a recently-requested or topically-adjacent item.[6]
Linguistics and text: Word reuse is bursty — a word just used is disproportionately likely to recur soon, far more than its overall frequency would predict — which is the temporal-locality property that cache-based and adaptive language models exploit to outperform static frequency models.[7]
Clarity¶
Locality names why a fast-but-small local store can substitute for a slow-but-large source: not because the local store is complete, but because the access distribution is concentrated and slowly-drifting. This is a clarifying move because it separates two questions that are easily conflated. The first is whether a resource is large; the second is whether its use is concentrated. A petabyte store with strong locality can be served well by a megabyte of cache, while a megabyte store with uniform random access cannot be meaningfully accelerated at all. Locality redirects attention from the size of the space to the shape of the demand on it.[2]
It also lets practitioners distinguish, before building anything, workloads with exploitable structure from genuinely random-access ones where no caching, prefetching, or local-intervention strategy can help. This converts an architectural decision into an empirical measurement: profile the access stream, estimate the hit rate a given working-set size would achieve, and only then commit to a hierarchy. Where locality is strong the engineering payoff is large and cheap; where it is absent the same engineering is wasted effort, and locality is the concept that tells the two apart in advance.
Manages Complexity¶
By asserting that the working set at any moment is small relative to the whole address space, locality justifies maintaining and reasoning about only a tiny active subset, collapsing a vast space of possible accesses into a manageable hot region. The combinatorial space of "everything that could be referenced" is replaced by the far smaller and slowly-changing space of "what is being referenced now," and correctness or performance arguments can be made against the working set rather than the entire universe of targets.[1]
This complexity reduction is what makes hierarchical design tractable across domains. A memory system need not make all of DRAM fast; it makes the working set fast. A public-health response need not surveil an entire population uniformly; it concentrates testing and tracing on the recent-and-near cluster. A library need not keep the whole catalog within arm's reach; it keeps the reserve shelf stocked. In each case locality lets a designer ignore the long cold tail of the distribution for the purposes of the active decision, dramatically shrinking what must be modeled, stored, or acted upon at any instant, while accepting a bounded miss rate on the rare excursions into the tail.
Abstract Reasoning¶
Recognizing locality supports the inference "if accesses cluster, then keeping the recent-or-near subset close yields a high hit rate" — and, just as importantly, the contrapositive warning that when locality is absent (uniform random access) caching, prefetching, and local intervention will fail no matter how well-engineered. It turns the design question "should we cache, prefetch, or localize?" from a matter of taste into an empirical question about the concentration and drift of the access distribution.[2]
The reasoning is also generative across the temporal and spatial axes independently. If a stream has strong temporal but weak spatial locality, the abstract response is to keep recently-used items resident but not to prefetch neighbors; if spatial locality is strong but temporal is weak, the response inverts to aggressive prefetching of adjacent items with rapid eviction. Locality thereby licenses a small family of structurally-distinct strategies and tells a reasoner which one a given workload calls for, by reading off which axis of autocorrelation is present. This same axis-decomposition transfers: an epidemiologist asks whether clustering is mainly temporal (recent-exposure windows) or spatial (geographic hotspots) and selects between time-bounded quarantine and place-bounded cordon accordingly.
Knowledge Transfer¶
The CPU-cache insight transfers directly to epidemiology: both treat a clustered access-or-contact distribution as the lever. A cache keeps the hot working set local and serves most requests from it; contact tracing keeps containment local to the recent-and-near cluster and intercepts most onward transmission there. The structural identity is exact — a concentrated, slowly-drifting distribution over targets — even though one substrate is silicon and the other is a population, and the transfer is explicit in the literature rather than merely analogical.[5] A practitioner who has internalized why LRU eviction works on memory pages already understands why recent-contact prioritization works in an outbreak: in both, recency is a strong predictor of near-future relevance because the underlying access stream is temporally autocorrelated.
The same insight transfers to recommender and search systems, which prefetch and pre-rank the predictably-near-next items, and to library science, where reserve collections are physical caches of the temporally-hot subset. The vocabulary of working sets, hit rates, and prefetching, developed in computer architecture, gives practitioners in geography, public health, and information science a precise language for what they were already doing intuitively, and the distance-decay intuition of the geographer reciprocally sharpens the systems engineer's sense of spatial locality. The transfer is grounded in shared distributional structure, not in surface resemblance, which is why insights move in both directions.
Examples¶
Formal/abstract¶
Memory hierarchy and the working-set model: A program executing a nested loop over an array references the same instruction addresses repeatedly (strong temporal locality) and walks through contiguous data addresses (strong spatial locality). Denning's working-set model formalizes this by defining the working set W(t, τ) as the set of pages referenced in the window (t − τ, t]; the key empirical claim is that |W| is small and stable relative to the full address space for most programs. A cache or page allocator that keeps exactly the working set resident achieves a hit rate far above what the cache's fractional size would suggest under uniform access. When the program transitions between phases, the working set drifts — old pages cool, new ones heat — but it does so slowly enough that an LRU policy, evicting the least-recently-used page, tracks the drift well. Mapped back: This is the canonical structural picture of locality. The access stream is autocorrelated in both time and space; the demand collapses onto a small, slowly-moving working set; and a local store sized to that working set, not to the whole space, captures most accesses. Strip away the words "page" and "cache" and what remains is a peaked, drifting distribution over targets — the prime itself.
Spatial autocorrelation and Tobler's first law: Consider any field measured over geographic space — soil pH, household income, disease incidence. Tobler's first law asserts that values at nearby locations are more similar than values at distant ones, which is spatial locality stated as autocorrelation. Formally, Moran's I or a semivariogram quantifies how correlation decays with distance; positive spatial autocorrelation is exactly the statement that the "access distribution" of similar values is spatially concentrated. A spatial interpolation that predicts an unmeasured point from its nearest neighbors works for precisely the reason a cache works: the relevant information is locally concentrated, so a small neighborhood stands in for the whole field. Mapped back: The target space here is geographic locations rather than memory addresses, but the structure is identical — autocorrelation makes the near-and-recent (here, near-in-space) disproportionately predictive of an unknown reference, so a small local subset suffices for most of the inferential work, and the absence of autocorrelation (a spatially random field) would defeat any neighborhood-based method.
Applied/industry¶
Outbreak containment via contact tracing: During an infectious-disease outbreak, new cases are not scattered uniformly across the population; they cluster in households, workplaces, and neighborhoods, and they cluster in time within the incubation-and-infectious window following an exposure. A contact-tracing program exploits this by treating each confirmed case as a recently-accessed item and prioritizing its spatial-and-temporal neighbors (household members, recent close contacts) for testing and isolation. The "working set" is the set of recently-exposed, near-in-network individuals; it is tiny relative to the whole population, and it drifts as the outbreak moves. Ring vaccination generalizes the same move spatially. The strategy succeeds exactly to the degree that transmission is autocorrelated; in a hypothetical pathogen that infected uniformly at random across the population, tracing the near-and-recent would offer no advantage. Mapped back: Substitute "individual" for "memory page" and "transmission" for "memory reference," and the epidemiological response is a caching strategy: keep the hot (recently-and-nearly-exposed) working set under close watch, accept that the cold tail of the population is unlikely to be the next case, and let the watched set drift with the access stream. The decision to trace at all is a bet on locality, and it is correct only when the contact distribution is in fact concentrated.
Library reserve collections and circulation management: An academic library cannot keep its entire catalog at the circulation desk, nor does it need to. Circulation data show that a small fraction of titles — current course readings, recent acquisitions, topically-trending works — accounts for the majority of checkouts, and that a title checked out recently is disproportionately likely to be requested again soon. The library responds by maintaining a reserve collection: a physically-local, fast-access cache of the temporally-hot working set, with short loan periods that force rapid eviction so the reserve shelf tracks the drift of demand. Reshelving and open-stack adjacency exploit spatial locality, since a patron who pulls one book on a topic is likely to want a neighbor on the same shelf. Mapped back: The reserve shelf is an LRU cache over the catalog, sized to the working set of currently-hot titles; the short loan period is an eviction policy that keeps it tracking demand drift; and shelf adjacency is a prefetch of spatially-local items. The whole apparatus is justified by, and only by, the concentration and slow drift of the circulation distribution — the same distributional property that justifies a CPU cache.
Structural Tensions¶
T1: Locality is a property of the workload, not the system, yet systems are built to assume it. A cache, a tracing program, or a reserve shelf embodies a bet that the access stream is concentrated and slowly-drifting. But the stream is generated by something outside the system's control — a program, a population, a readership — and can change character without warning. The tension is that the engineering artifact is fixed-purpose while the property it relies on is contingent and external, so a workload shift can silently convert a well-tuned local store into dead weight that adds latency without adding hits.
T2: Temporal and spatial locality can pull a design in opposite directions. A stream with strong temporal but weak spatial locality rewards keeping recently-used items resident and punishes prefetching neighbors; a stream with strong spatial but weak temporal locality rewards aggressive neighbor prefetch and rapid eviction. A system that assumes both, or that optimizes for the wrong axis, can perform worse than a naive design: prefetching neighbors in a temporally-local-only workload pollutes the cache with items that will never be touched. The single name "locality" conceals two separable and sometimes conflicting dimensions.
T3: Exploiting locality can destroy the locality it depends on. Aggressive prefetching, batching, or reorganization aimed at exploiting an observed access pattern can change the pattern itself. A recommender that prefetches and surfaces the predictably-near-next items reshapes user behavior, potentially flattening or concentrating the very distribution it measured. In epidemiology, a containment that succeeds in breaking local transmission alters the future contact distribution, so the locality estimated from past data no longer holds. The act of measuring-and-exploiting is not neutral with respect to the thing measured.
T4: A concentrated distribution is not the same as a locally-predictable one. High skew (a few very popular items) is easily mistaken for locality, but locality additionally requires that concentration persist and drift slowly enough to be predicted from recent history. A workload whose hot set is small but jumps erratically has strong skew and weak temporal locality; a frequency-based cache will do well on it while an LRU cache will thrash. Conflating concentration with predictability leads to selecting the wrong exploitation strategy and misattributing its failure.
T5: Optimizing for the common case silently penalizes the cold tail. Every locality-exploiting design accepts a bounded miss rate on rare excursions into the unloved tail of the distribution. Usually this is the correct trade. But when tail accesses are disproportionately important — a rarely-touched safety record, an isolated outbreak in a sparsely-connected community, an obscure but critical reference work — the locality strategy systematically under-serves exactly the cases where a miss is most costly. The distribution's shape and the value of its regions need not align, and locality reasoning attends only to the former.
T6: Locality justifies a local store but cannot bound the cost of a miss. The hit-rate argument says most accesses are cheap; it is silent on what happens on the misses. A system can have an excellent average hit rate and still be unacceptable if the penalty for the residual misses is catastrophic or highly variable. A memory hierarchy with a 99% hit rate may still be dominated by the latency of the 1% that reach disk; a tracing program with high local coverage may be undone by a single missed long-range transmission that seeds a new cluster. Locality reasons about frequency, not about the tail risk that frequency leaves untouched.
Structural–Framed Character¶
Locality Of Reference sits at the structural end of the structural–framed spectrum: it is an empirical regularity, the same wherever it appears, that accesses or events cluster rather than spread uniformly — recently-used items are disproportionately likely to be used again soon, and items near a used one are disproportionately likely to be used next.
The concept emerged in computer architecture as a statistical observation, formalized in the working-set model, but it carries no normative weight and needs no reference to human practice — it is just an autocorrelation structure over a space of targets. Applying it recognizes a clustering already present in the data rather than importing a perspective: the same regularity is Tobler's first law of geography (near things are more related than distant ones) and the spatial-temporal clustering of cases in epidemiology. On every diagnostic, it reads structural.
Substrate Independence¶
Locality Of Reference is a highly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. At root it is a substrate-agnostic statement about an access distribution: accesses cluster in time and space, and usage is autocorrelated rather than uniform, a claim that owes nothing to any one medium. It transfers cleanly across computational caching and prefetching, geographic regularity (Tobler's first law), biological and epidemiological infection clustering, and social recommender prefetching — the cache-to-contact-tracing crossing is made explicit. What holds it below the top is that it lives more as a statistical-distribution pattern than as a formal or cognitive one, which keeps its breadth strong but not universal.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 5 / 5
- Transfer evidence — 4 / 5
Relationships to Other Primes¶
Parents (2) — more general patterns this builds on
-
Locality Of Reference is a kind of Recurrence
Locality of reference is a specialization of recurrence: it asserts that access events reappear across time (temporal locality: recently-used items return soon) and space (spatial locality: items near a used one are used next). It inherits recurrence's structural commitment that patterns reappear with predictable spacing, particularized to the access-distribution domain. The working-set phenomenon is precisely the autocorrelated-recurrence signature in which past usage predicts future usage at non-uniform rates.
-
Locality Of Reference presupposes Heavy-Tailed Distributions
Locality of reference presupposes heavy-tailed distributions because the empirical regularity that a small slowly-drifting working set absorbs the vast majority of accesses IS a heavy-tailed access-frequency distribution: a few addresses account for most references while the bulk receive almost none. Without heavy tails' signature of a few extremes dominating sums and averages, locality reduces to uniform access. The working-set phenomenon, page-fault rates, and cache-hit predictions all rest on the same skewed distribution heavy-tailed analysis describes.
Children (1) — more specific cases that build on this
-
Caching presupposes Locality Of Reference
Caching presupposes locality of reference because the economic case for maintaining a small fast tier alongside a large slow tier depends entirely on workloads in which accesses cluster: recently-used items are likely to be reused soon, and items near a used one are likely to be touched next. Locality supplies the empirical distribution-of-access regularity that makes a cache pay off; without it, every access is equally likely to miss and the cache provides no average-case advantage over going straight to the source.
Path to root: Locality Of Reference → Recurrence
Neighborhood in Abstraction Space¶
Locality Of Reference sits among the more crowded primes in the catalog (35th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Partition, Contrast & Structural Difference (24 primes)
Nearest neighbors
- Interleaving — 0.81
- Exaptation — 0.80
- Bias — 0.80
- Symbolic Representation — 0.80
- Recurrence — 0.79
Computed from structural-signature embeddings · 2026-05-29
Not to Be Confused With¶
Locality of reference must be distinguished from Frame of Reference, the prime to which it is closest in surface vocabulary (both contain the word "reference," and embedding models score them as near neighbors). The resemblance is an artifact of shared lexicon, not of shared structure. Frame of reference is a chosen coordinate system or vantage point relative to which positions, motions, or meanings are described — the Newtonian inertial frame, the observer's perspective, the interpretive context that fixes what counts as "at rest" or "moving." It is fundamentally about the relativity of description to an observer's standpoint. Locality of reference, by contrast, says nothing about coordinate systems or vantage points; the "reference" in its name is an access — a memory reference, a contact, a checkout, a citation — and the prime is a claim about the statistical distribution of such accesses, namely that they cluster in time and space. One concept concerns the relativity of measurement to a frame; the other concerns the concentration of a stream of events. They share no structural machinery: frame of reference is invariant-and-transformation reasoning, while locality of reference is autocorrelation-and-distribution reasoning. A workload can have strong locality regardless of which coordinate frame one uses to describe its accesses, and a frame can be chosen regardless of whether the events described in it cluster at all. The high embedding similarity is precisely the kind of false neighbor that distributional-vocabulary screening must filter out.
Locality of reference is also not Caching, with which it is most often conflated in practice. Caching is the technique — the mechanism of maintaining a small fast store that holds a subset of a larger slow store and serving requests from it when possible. Locality is the empirical precondition that makes caching worthwhile: caches yield high hit rates only when the access distribution is concentrated and slowly-drifting. The two are related as a mechanism is to the property of the world that justifies it. One can implement a cache against a uniformly-random workload — it will simply achieve a hit rate equal to its fractional size and provide no speedup, because the locality it presupposes is absent. Conversely, locality exists in a workload whether or not anyone builds a cache to exploit it; Tobler's first law held over geographic data long before spatial indexes existed. Confusing the two leads to the circular explanation "the cache works because of locality" being mistaken for a mechanism, when in fact locality is a measurable property of the access stream and caching is one of several techniques (alongside prefetching, paging, and local intervention) that exploit it. The relationship is many-to-one: many techniques rest on the single underlying property.
Finally, locality of reference is distinct from Buffering, with which it is sometimes grouped because both involve an intermediate store between a fast and a slow component. Buffering absorbs a rate or timing mismatch between a producer and a consumer over time: a buffer smooths bursty input into steady output, or decouples a fast writer from a slow reader, by temporarily holding data in transit. Its justification is a mismatch in throughput or timing, and it works even when accesses are uniformly random, because its job is to decouple rates rather than to predict which items will be reused. Locality makes no claim about rate mismatch at all; it asserts that which targets are accessed is clustered, a claim about the spatial-temporal distribution of accesses rather than about the speed at which they arrive. A buffer with no reuse — every item passes through exactly once, in order — exhibits no locality, yet buffering is doing useful work; a cache exploiting strong locality may serve a workload with no rate mismatch at all. The structural distinction is between absorbing a temporal rate mismatch (buffering) and exploiting a concentrated access distribution (locality), and a system can need either, both, or neither independently.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.
Notes¶
Locality is conventionally decomposed into two axes — temporal and spatial — but these are not exhaustive of the structures that travel under the name. Some workloads exhibit branch or sequential locality (the next instruction is usually the textually-next one), and others exhibit associative or semantic locality (the next reference is topically rather than physically adjacent, as in a citation network or a topical shelf). Each is a different generator of the same underlying distributional property: a peaked, slowly-drifting histogram over targets. The two-axis decomposition is a useful default but should not be mistaken for a complete taxonomy.
The strength of locality is workload-specific and time-varying, which is why measurement, not assumption, is the disciplined stance. A program has different locality in different phases; a population has different contact locality during a holiday than during a workweek; a library's circulation locality spikes at the start of a semester. Designs that assume a fixed locality profile tend to fail at exactly the transitions where the profile shifts, and adaptive policies (which re-estimate the working set continuously) exist precisely to track this non-stationarity.
There is a deep connection between locality and the heavy-tailed frequency distributions (Zipf's law, the 80/20 rule) that recur across the same domains. Skew and locality are related but, as the structural tensions note, not identical: skew is a property of the marginal frequency distribution, while locality additionally requires temporal autocorrelation — predictability of the near future from the recent past. A workload can be Zipfian in aggregate yet have weak temporal locality if its popularity ranking churns rapidly. Keeping the two notions distinct is essential to choosing between frequency-based and recency-based exploitation strategies.
References¶
[1] Denning, P. J. (1968). The working set model for program behavior. Communications of the ACM, 11(5), 323–333. Introduces the working-set model: a running program references only a small, slowly-drifting subset of its address space at any instant, so the working set rather than the whole address space is the unit of reasoning and management. ↩
[2] Denning, S. (2005). The Leader's Guide to Storytelling: Mastering the Art and Discipline of Business Narrative. Jossey-Bass. Practitioner catalog of eight narrative patterns in organizational settings — motivating action, building trust in the leader, building trust in the company (branding), transmitting values, getting collaboration, sharing tacit knowledge, taming the grapevine, and creating shared vision — each tied to the use of story to deliver effects that propositional analysis cannot. ↩
[3] Hennessy, J. L., & Patterson, D. A. (2017). Computer Architecture: A Quantitative Approach (6th ed.). Morgan Kaufmann. Canonical text on the principle of locality: treats temporal and spatial locality as separable, measurable dimensions that the memory hierarchy, paging, prefetching, and LRU-family replacement policies exploit, while noting locality is a probabilistic/aggregate property with no guarantee on individual or adversarial accesses. ↩
[4] Tobler, W. R. (1970). A computer movie simulating urban growth in the Detroit region. Economic Geography, 46(sup1), 234–240. Source of Tobler's first law of geography ("near things are more related than distant things")—spatial locality of interaction—underwriting spatial autocorrelation statistics and distance-decay models. ↩
[5] Knox, E. G. (1964). The detection of space-time interactions. Journal of the Royal Statistical Society. Series C (Applied Statistics), 13(1), 25–30. Seminal test for spatial-and-temporal clustering of disease cases; formalizes that infections are autocorrelated in space and time, the structural basis for ring vaccination, local containment, and contact tracing and for the explicit cache-to-contact-tracing transfer. ↩
[6] Trueswell, R. W. (1969). Some behavioral patterns of library users: The 80/20 rule. Wilson Library Bulletin, 43(5), 458–461. Circulation study showing demand concentrates on a small recently-and-frequently-consulted subset of holdings (the 80/20 rule), the empirical basis for reserve collections and reshelving heuristics as physical caches. ↩
[7] Kuhn, R., & De Mori, R. (1990). A cache-based natural language model for speech recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(6), 570–583. Introduces the cache language model exploiting bursty word reuse (temporal locality)—a recently-used word is disproportionately likely to recur—outperforming static frequency (n-gram) models. ↩