algorithm-space-saving
The groupant beside algorithms/src/space_saving/mod.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.
Facts
Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.
Components
| component | label | where |
|---|---|---|
| approximant | partial | forward |
| ast | partial | syntax.json |
| cost | partial | cost |
| descriptant | partial | descriptant |
| invariant | partial | invariant |
| tests | partial | tests |
Obernants
The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.
| id | type | label | where |
|---|---|---|---|
algorithm-space-saving-behavior-complement | behavior_complement | partial source model | behavior-complement |
algorithm-space-saving-reverse-reconstruction | reverse_reconstruction | partial source model | reverse-reconstruction |
algorithm-space-saving-estimate-mass | estimate_conservation | partial source model | estimate-mass |
algorithm-space-saving-minimum-envelope | additive_frequency_certificate | partial source model | minimum-envelope |
algorithm-space-saving-tie-order | deterministic_argmin | partial source model | tie-order |
algorithm-space-saving-atomic-count | checked_transition | partial source model | atomic-count |
Model text
model
Space-Saving deterministic frequency intervals
2026-10-30. Original std-only implementation. Literature: Metwally, Agrawal and El Abbadi (2005), Efficient computation of frequent and top-k elements in data streams.
domain
Finite streams; BTreeMap key to (estimate,error), capacity k, total N. Successful runtime/standard allocations assumed unless an explicit allocation error is returned. Configuration and numeric limits follow the forward model.
forward
Increment a retained estimate. Insert (1,0) if room. Otherwise replace lexicographically smallest (estimate,key) with (minimum+1,minimum). Retained bounds are [estimate-error,estimate]; absent bounds are [0,minimum] once full, [0,0] otherwise.
invariant
Sum estimates=N; for retained keys estimate-error<=f<=estimate. For untracked keys f<=minimum. Replacement errors are historical minima, bounded by floor(N/k); min estimates never decrease once full. Therefore interval width/N<=1/k=p/100 for positive N. Empty bounds exact. This does not bound relative error of small f.
cost
O(k) worst-case update and absent query; O(log k) retained query; O(k) memory
behavior-complement
Within universe of streams/configurations and observations, intervals omitting true frequency or with width>N/k are excluded. Replacement of a larger key at tied minimum contradicts the deterministic transition relation.
reverse-reconstruction
Retain every stream executing to the given map; query-only relaxation retains all frequency tables inside intervals summing to N, without uniquely recovering order or evicted identities. Permutations of a below-capacity stream are a concrete nonunique fiber.
estimate-mass
Every event increases the sum of estimates by one, including replacement m -> m+1; inherited error records uncertainty rather than additional count.
minimum-envelope
Once full the minimum is nondecreasing and <=floor(N/k). New retained error equals previous minimum, and absent frequency is bounded by current minimum.
tie-order
Eviction minimizes (estimate,key) using BTreeMap iteration and explicit key tie breaker. Replays are identical; ordering of stream events may change approximate states.
atomic-count
N+1 checked before mutation. Estimates and inherited error remain <=N, guaranteeing safe replacement +1. Allocation failure follows standard map behavior outside Result.
tests
Four focused tests per module cover independent references, boundary/error states, nonunique fibers, and an executable compiled mutation guard. See evidence.md for actual runs and restored-source controls.
descriptant
The adjacent source remains authoritative. This model separates exact stream counts, additive frequency intervals, discrete rank/value rectangles and rounded numerical moments; these error metrics cannot be interchanged. syntax.json is a source-bound partial syntax projection without macro expansion or semantic AST. Prospective commitments precede implementation in prospective.md. No standalone executable surrogate is attached: production already uses bounded summaries, and a faster surrogate with a separately meaningful assurance metric remains an explicit gap. Finite tests do not certify numerical behavior universally or speed.
evidence
Validation epoch 2026-10-30
Actual CPU checks: 16 new focused tests, four per algorithm, included in all 505 library tests. Twice debug and twice release passed, with identical normalized 505 named test outcomes. No timing or speedup claim. Strict algorithms Clippy all-targets passed. Focused rustfmt passed. Source-only rustc test binaries compiled successfully for each production mutant; all four mutants failed their named guard assertion, source bytes were restored and each four-test positive suite passed.
Source SHA256: f1acc5b7bfafc39bca27359c4cfdbea5810f35943c606eee73838ddc7dacc0f9. Syntax projection generated twice into fresh files and compared byte-for-byte; no semantic AST or proof.
Independent references: exact BTreeMap stream frequencies for 3000 deterministic events at all four capacities; exhaustive 625 length-four arrays across five histogram widths compared with sorted values; exhaustive 3125 length-five bounded integer arrays, every partition, compared with independent integer sum and squared-sum moments. Boundary cases include empty streams, invalid tiers, endpoints, atomic count overflow, nonfinite inputs, and moment overflow. Translation is tested on a fixed binary-exact fixture, not universally certified.
Compiled mutants: MG suppress cancellation count (cancellation_conservation: one cancellation epoch); Space-Saving erase inherited error (replacement_certificate_and_tie: replacement inherits one unit error); histogram include exclusive rank endpoint (bucket_boundary_mutation_guard: rank at boundary selects next bucket); moments use old residual (variance_mutation_guard: Welford residual yields population variance one).
External evidence is under $MOVA_DATA/runs/octopus-wave19-streams-2026-10-30. Hashes bind the actual retained logs:
- debug1.log: e12ddcad38ace3b71d1669d5227ac04a67b4b09d4df6fffa0866b721f9a141b4
- debug2.log: 9bb3140eafe99143e2b9cb53f01a23112f3839052e8b7b835e5cf6990e942e64
- release1.log: 776d8e4bf51d3bde2938e089601a7067084099d594fcf0e64cd99e8361cb1836
- release2.log: 154dfee217199167c25a28cb603f50cf51be5062c5b95de63c47769294a0b7b8
- clippy.log: 038083ac66185a114552e9a3f886ed4081785d8b5cdf03227be1447e9c9e7f9c
- normalized.txt: 9492a4f663676b851e9ae50322f7b0a0b4a2b3c0d8328257dfca18b04c1a9c2e
- space_saving-mutant-compile.log: e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
- space_saving-mutant.log: 0182744b9eb25069ca2c001744da327a6e774eec1c90bda6541e02995120d3d1
- space_saving-restored.log: a6d7303bd1e9eab70b37d829539fb1b4e769a4f400e3d3a40d9c383de76f30a2
- space_saving-mutation-hashes.txt: 5de039790c4a9523fc8ee4839f412a52c9fde378277d76e65fd6b8f8f0bb1b33
Scope: finite reference correspondence, deterministic outcomes and mutation sensitivity; no universal numerical certificate, deployment, live teacher/model/corpus jobs or benchmark claims. Allocator aborts and full Rust semantic analysis remain outside these tests.
prospective
Prospective model (before implementation)
Streams own four std-only algorithms. Misra-Gries decrements k counters when full, preserving estimate <= frequency <= estimate + cancellations, cancellations <= N/(k+1). Space-Saving replaces smallest estimate (key breaks ties) and stores inherited error; retained frequency in [estimate-error,estimate], absent frequency <= min estimate. k=100/p for p in {1,2,5,10}, hence additive absolute frequency error/N <= p/100 (zero stream exact). BTreeMap uses O(k) storage, O(k) worst update. Checked N overflow is atomic; allocator failure outside Result. Histogram stores fixed width buckets of discrete bounded u64 domain, capped at 1,000,000 buckets. Query by exact zero-based rank returns bucket value interval and occupied rank interval; no relative value or <=10% rank guarantee is declared. O(1) updates, O(b) query and memory. u128 bucket arithmetic handles MAX endpoints; allocation errors returned at construction. Moments stores count, mean, unnormalized squared deviations using Welford updates and Chan merge. O(1) storage/work; count<=2^53, reject nonfinite input and nonfinite intermediate results atomically. IEEE rounding means no universal relative error cap; underflow may lose subnormal variance, cancellation/large offsets require scoped evidence. Independent integer sums and sums-of-squares oracles on bounded integral data, translation fixtures, partition merging. No standalone executable surrogate: these are production approximations; faster surrogate and certified floating error remain gaps. All four: exhaustive bounded fixtures, adversarial edges, repeated debug/release equality, compiled mutant controls, source-bound groupants and partial syntax projections. No speed claims.
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/algorithms/src/space_saving/mod.rs.groupant/