Skip to content

Glossary

Jip Claassens edited this page Sep 15, 2026 · 1 revision

Glossary

In one sentence: every term a returning colleague meets on this wiki, in the Julia and GeoDMS code, in the sweep logs and on the deck — one to three lines each, with the code name where it differs and the page that owns the full explanation.

Where this sits: beside every page. The pipeline is GeoDMS → Arrow → Julia LP → λ-sweep → rounding → scenarios → deck (The pipeline in plain words); Maarten's compact formal pages are Lambda sweep method and Pharmacy service access. Everything here is as of 15 September 2026, commit e25416b.

How to read this page

Entries are alphabetical; Greek letters and symbols are filed under their spelled-out names (λ under L, w under W, ★ under B for baseline). Synonyms have a "see …" stub. An entry gives the plain meaning, then in code: the identifier or log column where it differs, then the owning page. Four conventions apply throughout:

  • x and y were swapped on 24 July 2026 (4cee238): since then y is facility openness and x the client share; logs, issue #47 and struct WarmStartState{TX, TY} from before that date use the opposite letters (sum_x, frac_x). → The facility location model explained
  • A star marks an LP optimum: y*_j is the solved value of y_j (y_relaxed[j] in solve_at_w!); Σ y* is the log's sum_y.
  • Ids are 0-based on disk and in Julia; facility ids are local to each _j.arrow, so the two sides match cells only by coordinates.
  • Seconds on disk, minutes in Julia; every count is in 1 km² cells, not pharmacies. The one-table version of all units is on Log lines and output files.

A

  • A-Res (weighted reservoir sampling) — Efraimidis–Spirakis (2006): k items drawn without replacement with probability ∝ weight; multistart's random seeds. In code: randomized_seed in lp_run.jl. → From LP fractions to real pharmacies
  • Adapted logit — the deck's name for the LOGISTIC travel-cost function: a plain logistic curve, not a log-logistic and not a choice model. → Travel cost functions
  • Aggregate / AGGREGATE — the EU-wide frontier: 41 disjoint areas summed at common w, gaps interpolated log-in-w; exact by separability. In code: aggregate_fd in doc/build_charts.py. → Metrics aggregation and ranking
  • Archetype (T1–T4) — see typology.
  • Arrow — Apache Arrow, the format of every GeoDMS ↔ Julia hand-over: <area>_od.arrow, _i.arrow, _j.arrow in; assignment.arrow, traveltime.arrow out. In code: GeoDMS gdalwrite.vect; Julia Arrow.jl. → Log lines and output files
  • assignment.arrow / traveltime.arrow — the driver's outputs per point: (id, open) per cell and (id, t_ij) in minutes per assigned client, under <side>\<area>\lambda_sweep\<FUNC>\<label>\ (label baseline, w=<float>, S1, S2). In code: write_assignment_arrow, write_traveltime_arrow. → Log lines and output files

B

  • Balanced improvement — see diagonal crossing.
  • Baseline ★ — today's network priced by the model: every client at its road-nearest existing pharmacy cell, unreachable ones at BIG; its count and travel are the S1 and S2 targets ("B" in the rectangle formulas). In code: baseline_metrics in lambda_sweep_simplex.jl; n_used → target_cells, total travel cost(c). → Scenarios S1 S2 S3
  • Basis — the simplex method's record of which variables are "on" at its current corner (vertex); HiGHS starts the next solve from the last one — a near step from a nearby optimum is cheap, a far basis (aborted solve, big jump in λ) is not. → Background theory
  • Basis restore — after a bisection detour, re-solving the bracket's upper grid point without rounding so the next grid point is again a near step. In code: resolve_for_basis! in lp_run.jl; log basis restored at w=…. → Code walkthrough lambda_sweep_simplex
  • BIG — the price of one unserved resident: c(120 min) = 120 person-minutes under LINEAR, 1.0 under LOGISTIC; charged alike in the LP, the rounding score and the baseline. In code: big_cost() in settings.jl; env BIG_TRAVELTIME_MIN. → Travel cost functions
  • Bind / binding — a constraint binds when it is actually limiting the solution (the choice-set radius in SE2 and ITF); a cap that does not bind changes nothing. → Scope rules and data gaps
  • Bisection (geometric) — pinning S1 or S2 inside a bracket by solving at √(lo·hi) and keeping the half that still encloses the target. In code: bisect!; env REFINE_TOL (0.002: one cell or 0.2 % of today's travel), REFINE_ITERS (8). → Scenarios S1 S2 S3
  • Bracket / bracketed — two consecutive solved grid points enclosing a target: sum_y around today's cell count (S1) or cost_c around today's travel (S2). In code: crosses, find_bracket; the CSV flags s1_bracketed/s2_bracketed only say the swept range spans the target. → Scenarios S1 S2 S3
  • Branch-and-bound — see MIP.

C

  • c(t) — the travel-cost function: the price of one resident making a trip of t minutes; LINEAR and LOGISTIC in production, so every result exists twice. In code: c(t) in settings.jl, t in minutes; env TRAVEL_FUNC. → Travel cost functions
  • Candidate (j) — a 1 km² cell the LP may open: at least 50 residents or a pharmacy today, in a country with pharmacy data. In code: NewPharmacyLocations in SourceData/Locations.dms; MinPopSizeForPharmacyLocation := 50. → GeoDMS OD matrix and choice set
  • Candidate subsampling / stride — keeping only every K-th candidate so a huge area fits a solve; a subsample can only do worse (FRI). See PROTECT_BASELINE. In code: env LOCATION_SELECTION_FACTOR (1). → Scope rules and data gaps
  • Cap ladder (A→D) — the abandoned June 2026 plan to answer S1/S2 with catchment caps, the λ-sweep only as fallback D; dropped 4 September 2026. In code: cap_scenario.jl, s1_cap.jl, s2_cap.jl. → Alternatives not pursued
  • Capacitated / capacity — a location model in which each open cell has a maximum load (the cap ladder's MAX_CAP); the λ model has none; adding one is a roadmap item. → Alternatives not pursued
  • Catchment — the residents whose road-nearest pharmacy cell is a given one (per cell, Lewis's "metric 6") or that number split equally over the cell's pharmacies ("metric 5"); descriptive only, not a model quantity. In code: cell_catchment, pharm_catchment in Analyses.dms. → Metrics aggregation and ranking
  • CELF — see lazy greedy.
  • Cell / location — a 1 km² grid cell, the unit of "facility" on both sides; "locations" on the deck means cells. In code: Pharmacies/within_StudyArea/uq_cells. → GeoDMS OD matrix and choice set
  • Cells vs raw counts — every model count is in cells; the raw pharmacy count is known to Julia only for the Netherlands (1 992 pharmacies in 1 617 cells), as an informational column. In code: RAW_PHARMACY_COUNTS; log columns n-cells, n-raw. → Scenarios S1 S2 S3
  • Certified gap / bound sandwich — travel_relax + λ·sum_y ≤ integer optimum ≤ travel_multi + λ·n_open, on total cost (the deck's "multi vs relax" compares travel only). Figure: images/explainer_bounds_and_rounding.png. → From LP fractions to real pharmacies
  • Choice set — the cells in a client's OD rows, the only ones the LP may assign it to: every candidate within the road time to the client's 5th-nearest existing pharmacy cell, capped at 120 min. In code: the limit and cut directives. → Scope rules and data gaps
  • Chord interpolation — reading the frontier at exactly today's count or travel by linear interpolation between the two bracketing solved points; the deck tables and chart markers since cb21c6c. In code: interp in build_deck.mjs; interp_y_at_x in frontier_metrics.py. → Scenarios S1 S2 S3
  • Cliff — see degenerate LP.
  • Client (i) / ModelClient — a populated 1 km² cell weighted by its residents; ModelClient restricts that to countries with a pharmacy record (scope Rule 0). In code: Client, ModelClient in Analyses.dms; client_rel in the OD. → GeoDMS OD matrix and choice set
  • Client weight — residents per client cell, the multiplier on c(t). In code: env CLIENT_WEIGHT (total_pop); wpop[k] per OD row, client_pop[i] per client. → Code walkthrough settings
  • Concave dent — a count p whose best travel lies above the line joining the best travel at p − 1 and p + 1; no λ selects it. Not a kink. → The lambda sweep and the Pareto frontier
  • Connector link — the synthetic 30 km/h link from a client or facility cell centre to the nearest road; its start node is the location's OD node. Not the clean-up's ConnectorLink, a road link bordering a junction-free section. In code: LinkSet_Calc/location_connectors in Templates.dms (around line 134). → GeoDMS OD matrix and choice set
  • Constant drops out — each OD row's coefficient is (c(t) − BIG) · pop and the constant BIG · Σpop is omitted, so HiGHS's printed objective is a large negative number — read travel_relax. In code: the @objective in build_lp_warmstart. → The facility location model explained
  • Convex hull — a rubber band stretched around a cloud of points; a sliding straight line touches only the hull, hence lower convex envelope. → The lambda sweep and the Pareto frontier
  • cost_c — the coverage-honest travel of the selected rounding (multistart by default), i.e. travel_multi; what S2 is bracketed and bisected on. In code: field cost_c of the solve_at_w! result; Inf on a light row. → Log lines and output files
  • CountryData — one side's OD in memory (N rows, M cells; per-row client, cell, minutes, weight). Four loaders build it; the sweep's load_from in lambda_sweep_simplex.jl is the only one applying the NUTS exclusion. → Code walkthrough settings
  • Coverage-consistent / coverage-honest — pricing a client with no reachable open (or existing) cell at BIG rather than dropping it; the log says the former of the baseline, the travel_of comment the latter of the rounding score. In code: travel_of in lp_run.jl. → The facility location model explained
  • Covering models — the location family that counts only whether a client is within a distance standard D (set covering, maximal covering). → Background theory
  • Crossover — pushing an interior-point solution to a corner (a basis) so y comes out mostly 0/1. In code: HiGHS run_crossover, on only when SOLVER=ipm. → Background theory
  • cut / limit — the two stopping rules of the OD search from each client: cut at 7 200 s (120 min); limit once five cells holding a pharmacy today are reached. In code: impedance_matrix_od64(…) in Analyses.dms; max_nr_facilities_per_client in ModelParameters.dms. → GeoDMS OD matrix and choice set

D

  • Deck — the results deck doc/lambda_sweep5.pptx (74 slides at e25416b; "pNN" on this wiki is a slide number), generated by build_deck.mjs and merged by merge_deck.ps1 with a few hand-authored slides. → Deck and charts pipeline
  • deck_data.json — the parsed logs: per area and function the baseline, every Combined-table row and the snapped scen summary; the single input of every chart, table and CSV. In code: doc/build_deck_data.py. → Log lines and output files
  • DECIDE item — the needs the group status in doc/deck_changes_from_review.md (beside DONE, PROPOSE, ANSWER, DROPPED) and on the roadmap (p63); the one open DECIDE is Bernhard Nöbauer's logistic rescaling. → Travel cost functions, Decision log
  • Degenerate LP — many bases represent the same corner, so simplex pivots without improving — the "cliff" at the top of the grid under hard coverage. Reviewer Martijn Brons used the word differently for LOGISTIC: many λ mapping to one solution. → Background theory
  • DEGURBA — Eurostat's degree-of-urbanisation grid; coloured the ranked lists until 17 July 2026, nothing reads it at e25416b. In code: doc/region_typology.jl (orphaned). → Metrics aggregation and ranking
  • Descriptives — Lewis's six indicators of today's network per area, computed in GeoDMS on road-nearest catchments; deck p3–6. In code: /Analyses/Pharmacies/Descriptives/Table; collect_descriptives.jl → doc/pharmacy_descriptives.csv. → Metrics aggregation and ranking
  • Diagonal crossing / balanced improvement / λ at the crossing — where the line from ★ to the rectangle's far corner C meets the frontier; λ_cross is interpolated linearly in w (why is not recorded). In code: diagonal_cross, interp_w_at_x in doc/frontier_metrics.py; key lambda_cross. → Metrics aggregation and ranking
  • Dominated — a (count, travel) point another beats on both; pareto_frontier drops such rows and keeps the Pareto envelope. → Metrics aggregation and ranking
  • Dual simplex — HiGHS's default simplex variant; cheap on a warm re-solve because a small change in λ flips few reduced costs and so needs few pivots (steps to a neighbouring corner). → Background theory

E

  • Empty pharmacy / #empty — a pharmacy (or cell) that is nobody's road-nearest, so its catchment is 0; counted separately. In code: pharm_n_empty, cell_n_empty. → Metrics aggregation and ranking
  • Envelope — see Pareto frontier and lower convex envelope.
  • Era (school era / pharmacy era) — the school era (network model from 2023; school optimisation February–May 2026: run_scenario, hard coverage) and the pharmacy era (from 24 May 2026: the λ-sweep). → Decision log, Service allocation procedure
  • Existing side / New side — the two GeoDMS runs per area: Existing (destinations = today's pharmacy cells; gives the baseline) and New (destinations = the candidates; what the LP optimises). In code: ProjName ExistingPharmacies / NewPharmacies (Analyses.dms); Julia EXISTING_PATH, NEW_PATH. → GeoDMS OD matrix and choice set

F

  • fac_€ — log column: the placeholder facility bill at the fractional count, sum_y × 100 000. Not the facility term of the objective (λ × n_open). → Log lines and output files
  • Facility (j) — the code's word for an existing pharmacy cell or a candidate cell, depending on the side. In code: facility_rel in the OD, id in _j.arrow (0-based, per-file local). → GeoDMS OD matrix and choice set
  • FACILITY_MIN_COSTS (the €100 000 placeholder) — the notional fixed cost of one location that converts w to λ; never calibrated — only € labels depend on it. In code: settings.jl line 17; hard-coded again in doc/ (FMIN, FMC). → The lambda sweep and the Pareto frontier
  • facility_cost / FACILITY_CLIENT_COSTS / FACILITY_FUNC — the school-era per-facility cost family (a + b·q, floor, concave power law); dead code — a size-dependent term cancels when everyone is served, leaving λ per location. → Alternatives not pursued
  • Far basis — see basis.
  • Fast interchange — Resende & Werneck (2007): the exact travel change of a swap as gain[f] − loss[r] + extra[f][r]. In code: swap_round! in lp_run.jl. → From LP fractions to real pharmacies
  • Fine sweep — a retry after the grid loop of timed-out grid points inside the S1/S2 brackets; with 1,2,5 multipliers nothing lies between consecutive grid points, so it is usually empty (vestigial). In code: log Fine λ sweep over w in […]. → Code walkthrough lambda_sweep_simplex
  • Formalisation (HIGH / MEDIUM / LOW) — see typology.
  • frac_y — the number of candidates with 1e-6 < y* < 1 − 1e-6 at a sweep point. Called frac_x before 24 July 2026. In code: n_fractional_y; frac in deck_data.json. → The lambda sweep and the Pareto frontier
  • Fractional pool / must-open — the rounding's two classes: must-open cells (y* ≥ 1 − 1e-6, always open) and the pool (1e-6 < y* < 1 − 1e-6), the only cells a seed or a swap may open or close. → From LP fractions to real pharmacies
  • Frontier — see Pareto frontier.
  • Full-coverage floor — under hard coverage, the smallest number of locations that reaches every client; in several areas above today's count (NL: 1 899 vs 1 615). Soft coverage removed it. → The facility location model explained

G

  • Greedy — see lazy greedy; log column travel_greedy.

H

  • Hard coverage — the school-era rule that every client must be fully assigned (Σ_j x_ij = 1); still in run_scenario; replaced by soft coverage on 16 July 2026. → The facility location model explained
  • HiGHS — the open-source LP solver the Julia side drives (1.13.1 in every sweep log); its iteration output is interleaved in the sweep logs. → Code walkthrough lp_run

I

  • Impedance — a link's travel time in seconds (LengthKm / KPH plus 0–6 s per link end for junctions); t_ij is its sum along the fastest path, floored at 1 s. In code: Templates.dms around line 223. → GeoDMS OD matrix and choice set
  • Improvement rectangle — the box with corners ★ (B), S1, S2 and C (S2's count, S1's travel); rect_area is the raw product, rect_rel the product of the two shares of today's values (the ranking metric). In code: metrics_for in doc/frontier_metrics.py. → Metrics aggregation and ranking
  • Integer optimum — the exact answer with y ∈ {0, 1}, which nobody computes at this size; sandwiched between the LP bound and the multistart set (certified gap). → The facility location model explained
  • Integrality / nearly integral / integrality gap — a relaxation is integral when its optimal y are all 0 or 1; the strong formulation is nearly so (small frac_y). The integrality gap is the distance between LP and integer optimum. → Background theory
  • Interchange — see swap.
  • Interpolated S1/S2 — see chord interpolation and snapped S1/S2.
  • IP (integer programme) — the honest problem with y ∈ {0, 1}; see MIP, LP relaxation. → Background theory
  • IPM (interior-point method) — an LP method moving through the inside of the feasible region rather than corner to corner; immune to a far basis, but full price per solve and needs crossover. July 2026's solver, now an escape hatch. In code: env SOLVER=ipm (default simplex). → Background theory

J

  • JuMP — the Julia modelling layer that builds the LP and hands it to HiGHS. → Code walkthrough lp_run
  • Junction-free section — a chain of degree-2 road nodes collapsed into one link by CreateMoreEfficientNetwork; a closed one with no ConnectorLink crashed alloc on Norway, SE2 and SE3 (#50) and is now dropped. In code: Templates.dms: NrConnectorLinks (around line 522), ToBeOmitted (608). → Scope rules and data gaps

K

  • Kink / bump — a step in a charted LP bound where rows of two different LPs meet (issue #54); a provenance artefact, not a concave dent. Lewis's "kinks" at ~5 and ~45 min are the logistic's bends (Travel cost functions). → Branches data and environment

L

  • λ (lambda) — the price the model charges for one open location, in the travel bill's unit (person-minutes under LINEAR, dimensionless under LOGISTIC), labelled in placeholder euros; the slope of the supporting line. In code: λ = w * FACILITY_MIN_COSTS in solve_at_w!; log column λ (€). → The lambda sweep and the Pareto frontier
  • λ-axis chart — the sweep read along λ (log x-axis): the two travel bounds and today's travel on the left axis, n_open/sum_y and today's count on the right. Deck p65–66; images/lambda_axis_*.png. In code: doc/lambda_axis_chart.py. → The lambda sweep and the Pareto frontier
  • Lagrangian relaxation — moving a constraint (here "exactly p facilities") into the objective with a multiplier (λ) instead of imposing it. → Background theory
  • Landbody — one polygon of a country's multipolygon (the mainland, Sicilia, each Wadden island); since 79cb58b the road network keeps the largest strongly connected component per landbody (ITG: 983 169 → 591 unreachable residents). In code: Country_Split in RegionalUnits.dms; a seldomain in Templates.dms. → Scope rules and data gaps
  • Lazy greedy (CELF) — Leskovec et al. (2007): because the travel saving is submodular, only the top-ranked cell needs recomputing. Multistart's second seed; prices an unreached client at Inf. In code: greedy_round in lp_run.jl. → From LP fractions to real pharmacies
  • Lexicographic — solving two objectives in strict order rather than adding them with a weight; the cap ladder's S2. In code: cap_scenario.jl. → Alternatives not pursued
  • Light row / LP-only row — a grid point solved without rounding during a refine-only walk; prints - and (LP only), writes no Arrow files, cost_c = Inf. In code: solve_at_w!(…; rounding=false). → Scenarios S1 S2 S3
  • LINEAR / LOGISTIC — the two production travel-cost functions: c(t) = t (BIG 120) and c(t) = 1 / (1 + e^{−(t − 25)/10}) (midpoint 25 min, scale 10 min; BIG 1.0). Figure: images/explainer_travel_cost_functions.png. In code: env TRAVEL_FUNC, LOGISTIC_MIDPOINT, LOGISTIC_SCALE. → Travel cost functions
  • Log-in-w interpolation — any value read between two solved grid points is interpolated in ln w, because the grid is geometric; the one exception is lambda_cross. In code: interp(…, "w") in build_deck.mjs; _interp_at in build_charts.py. → Metrics aggregation and ranking
  • Lower bound / upper bound — the LP's travel (travel_relax, grey dashed) can never be beaten by any real network at that λ; the multistart set's (travel_multi, solid blue) is achievable. See certified gap. → From LP fractions to real pharmacies
  • Lower convex envelope — the highest curve with no bumps that fits under the points (p, best travel with p sites): the only points a priced sweep can land on. → Background theory
  • LP / LP relaxation — a linear programme; here y ∈ {0, 1} relaxed to 0 ≤ y ≤ 1, so a cell may be half open; its optimum is a lower bound on the integer one. In code: build_lp_warmstart in lp_run.jl. → The facility location model explained

M

  • Matheuristic — a heuristic that uses a mathematical-programming solution as raw material; the rounding is one. → Background theory
  • mean_t — population-weighted mean travel minutes. Not coverage-consistent (a sweep row counts stranded residents at 0 min, the baseline at 120 min); compare cost_c instead. → Log lines and output files
  • MIP / branch-and-bound / exact solve — the integer programme solved exactly by splitting on a fractional y and re-solving LPs; out of reach per sweep point at this size. An exact MIP to pin S1/S2 is a roadmap item (doc/todo.md B5). → Background theory
  • .mmd / purge — GeoDMS's memory-mapped storage: the cached Linkset_* and FinalSet_* directories under %LocalDataProjDir%; the rebuild runners delete ("purge") them first. In code: run_rebuild_all.ps1. → Running a sweep end to end
  • Movers — the areas whose OD changed after the landbody and #50 fixes — ITG, PL8, Portugal, SE2, SE3 — re-swept on 2–3 September 2026. → Branches data and environment
  • Multistart — the selected rounding: ten starting sets (seeds: top-p, greedy, eight y*-weighted random draws), each polished by swaps, the lowest coverage-honest travel kept. In code: multistart_round in lp_run.jl; env MS_RESTARTS (10), MS_ROUNDS (12), MS_SEED (20240601). → From LP fractions to real pharmacies
  • "multi vs relax" — the deck column (travel_multi − travel_relax) / travel_relax; can be negative because n_open may exceed sum_y. → From LP fractions to real pharmacies
  • Must-open — see fractional pool / must-open.

N

  • n_open / p / p_open — the integer count every rounding must open: sum_y rounded, clamped to [1, M] (issue #47's "apples-to-apples"). In code: p_open = clamp(round(Int, sum_y), 1, length(facilities)). → From LP fractions to real pharmacies
  • n_used — see baseline.
  • Near step — see basis.
  • network0 / network1 / network2 / alloc — the GeoDMS batch steps per side: cut the per-area street subset; build the initial link set; clean it; export the three Arrow files. In code: STEPS in run_pharmacy_pipeline.bat, run_new_pharmacy_pipeline.bat. → Running a sweep end to end
  • NUTS column — the NUTS-3 code of the polygon containing a cell centre, empty outside every NUTS-3 polygon; NUTS-2 and NUTS-1 are its first 4 and 3 characters. In code: column NUTS in _i.arrow and _j.arrow. → GeoDMS OD matrix and choice set
  • NUTS exclusion / region exclusion — scope Rule 2 (#49): a NUTS region with more than 50 % of its inhabited cells absent from the existing-side OD leaves the model entirely; fires only for PT200 Açores and PT300 Madeira. In code: excluded_nuts_regions in settings.jl; env NUTS_EXCLUDE_SHARE (0.5). → Scope rules and data gaps

O

  • OD matrix / OD row — the sparse origin–destination table: one row per (client cell, facility cell) pair inside the client's choice set; the log's N is its row count, M the cells kept. In code: <area>_od.arrow (client_rel, facility_rel, t_ij in float32 seconds, d_ij). → GeoDMS OD matrix and choice set
  • OD node — see connector link.
  • Openness — a GeoDMS map for reading a scenario back: per cell, the inverse-square-distance potential (within 10 km) of the open cells divided by that of all. In code: Openness in ReadSweepResults_T. → GeoDMS OD matrix and choice set
  • Orchestrator / worker / tag — the PowerShell runners (run_resweep_batch.ps1, run_recalc_batch.ps1, run_new_country_sweeps.ps1; run_rebuild_all.ps1, run_realloc_all.ps1); a worker is one process working through a list of areas, a tag the suffix naming its progress log and CSV (q1, all, nuts). → Running a sweep end to end

P

  • p-median — the location problem with exactly p facilities and minimum weighted travel (Hakimi 1964; ReVelle & Swain 1970); S1 is a soft-coverage instance of it. → Background theory
  • Pareto frontier / frontier — the lower-left outline of all achievable (count, travel) points; depends on geography and c(t), not on the € placeholder. Pareto envelope is the doc/ name for the staircase left after dropping dominated rows. In code: pareto_frontier in doc/frontier_metrics.py. → The lambda sweep and the Pareto frontier
  • Phantom fractions — thousands of tiny positive y left by IPM + crossover when the optimum is not unique; why the count is round(sum_y), not "cells with y > 1e-6". → Alternatives not pursued
  • Pinned — a scenario whose bisection has finished: within tolerance or after REFINE_ITERS solves — bisect! prints "pinned" either way; read the (target …) residual on the S1 pinned at w=… line. → Scenarios S1 S2 S3
  • Pivot — see dual simplex.
  • Plateau detection / upward extension — a removed mechanism that multiplied w by 2.5 above the grid until sum_y fell through today's count or stopped moving (the plateau); added 206923c, removed f93aea1 as intractable under hard coverage. Lambda sweep method §2 and README.md still describe it. → The lambda sweep and the Pareto frontier
  • Point cloud / ranked lists — the analysis charts: a log–log scatter of every area's ★, S1, S2 and crossing (p67–68) and bars sorted by rect_rel, rect_area or lambda_cross (p69–74). In code: doc/frontier_charts.py; env METRIC_SUFFIX (the deck uses interp3). → Metrics aggregation and ranking
  • Pool — see fractional pool / must-open.
  • Presolve — HiGHS's model reduction before solving; skipped on warm re-solves ("Solving LP with useful basis so presolve not used"). → Code walkthrough lp_run
  • PROTECT_BASELINE — with a candidate subsample, keep every candidate that coincides with an existing pharmacy cell and stride only the rest, so the frontier can still reproduce today's network. In code: env PROTECT_BASELINE (1). → Scope rules and data gaps

Q

  • QUADRATIC — the school-era travel function 0.05·t² + 0.5·t (BIG 780), still the default TRAVEL_FUNC in settings.jl; nothing downstream parses a QUADRATIC sweep. Always set TRAVEL_FUNC. → Travel cost functions

R

  • Reduced cost — see dual simplex.
  • Refine-only — a rerun that only re-pins S1 and S2 on an already-swept area: walk the grid LP-only, bisect each bracket as it closes, stop. Honest only if sweep and refine used the same OD and candidate set (#54, kink). In code: env SWEEP_STOP_AFTER_REFINE=1; logs\refine_<area>_<FUNC>.log. → Scenarios S1 S2 S3
  • Rounding — turning the LP's fractional y* into a real set of exactly n_open open cells; top-p, greedy and multistart are always computed, env ROUNDING selects which is the result (any other value silently means top-p). → From LP fractions to real pharmacies
  • RSSV — random-sampling spatial-voting candidate reduction (Albuquerque, Figueiredo & Genre-Grandpierre); a roadmap item to replace the stride subsample, not implemented. → Decision log

S

  • S1 / S2 / S3 — the three scenarios read off the frontier against ★: S1 the point at today's count (less travel), S2 the point at today's travel (fewer locations), S3 the whole frontier. Defined by count and travel, never by a λ; three constructions coexist (Pitfalls). In code: target_cells, base.cost_c. → Scenarios S1 S2 S3
  • SCC (strongly connected component) — a set of road nodes that can all reach each other respecting one-way links. In code: strongly_connected_components on LinkSet_digraph in Templates.dms. → Scope rules and data gaps
  • Scope rules (Rule 0 / 1 / 2) — who is in the model before anything is optimised: Rule 0 country (ModelClient), Rule 1 landbody, Rule 2 NUTS region exclusion. The choice-set radius is a limit, not a rule. → Scope rules and data gaps
  • Seed — see multistart: a starting set for the swap polish, not the random-number seed MS_SEED.
  • seldomain — the Templates.dms word for a landbody as a selection domain (Check_Connectiveness_T_seldomain, IsConnected_seldomain, seldomain_rel). Say landbody in prose. → GeoDMS OD matrix and choice set
  • Separability — at one common λ the sum of the per-area optima is the optimum of the combined problem; what makes the aggregate exact and why every area is swept on the same w-grid. → Metrics aggregation and ranking
  • Side — see Existing side / New side.
  • Simplex — the LP method that walks from corner (vertex) to corner of the feasible region, one pivot at a time. → Background theory
  • Snapped S1/S2 / scen — the S1/S2 of the log's scenario summary: the rounded row whose sum_y (S1) or cost_c (S2) is nearest today's value — coarse before #52 (five LINEAR areas had S1 = S2); what the S1/S2 folders and exported location files use. In code: closest; scen in deck_data.json. → Scenarios S1 S2 S3
  • Soft coverage — the coverage constraint Σ_j x_ij ≤ 1: a client may be left wholly or partly unserved at BIG per resident; adopted 4ec47a9. Figure: images/explainer_soft_coverage.png. → The facility location model explained
  • Stranded / unreachable / unserved — a client cell with no open cell in its choice set (after rounding: stranded) or no existing pharmacy in its OD (baseline: unreachable), priced at BIG per resident. In code: n_stranded, stranded_pop; st in deck_data.json; log lines unreachable clients: … and stranded : topp a greedy b multi c. → Scope rules and data gaps
  • Strong (disaggregated) formulation — the linking constraint once per OD pair, x_ij ≤ y_j, rather than once per facility; more rows, but a nearly integral relaxation. → The facility location model explained
  • Study area / area — one modelling unit: a country by English name or a NUTS-1 code (FRI, ITG, PL8, SE2 …); France, Italy, Sweden and Poland are swept by NUTS-1 parts (42 with logs at e25416b, 41 aggregated, 44 in ORDER). In code: env STUDY_AREA (GeoDMS), COUNTRIES (Julia). → GeoDMS OD matrix and choice set
  • Submodular — a set function with diminishing returns; the travel saving of an open set is one, which licenses the greedy seed and CELF. → Background theory
  • sum_y — Σ_j y*_j, the LP's fractional count at a sweep point; monotone decreasing in w, hence the S1 bracketing variable and the frontier chart's x-axis. Called sum_x before 24 July 2026. In code: column 2 of the log row. → The lambda sweep and the Pareto frontier
  • Supporting line / tangency — in the (count, travel) plane, travel + λ·count is constant along a line of slope −λ; sliding it down until it last touches the achievable networks gives that λ's optimum. Figure: images/explainer_lambda_tangency.png; doc/tangent_chart.py (deck p10). → The lambda sweep and the Pareto frontier
  • Swap / vertex substitution / interchange — the classic p-median local search: close one open pool cell, open one closed pool cell, if travel falls; must-opens never closed. In code: swap_round! in lp_run.jl. → From LP fractions to real pharmacies
  • Sweep / λ-sweep — solving the same LP for a ladder of w values, warm-started, rounding each point, and reading S1/S2 off the result. In code: analyze_country in lambda_sweep_simplex.jl. → The lambda sweep and the Pareto frontier

T

  • t_ij / d_ij — travel time and distance from client i to facility j: t_ij float32 seconds on disk (floor 1 s), minutes in Julia and in traveltime.arrow; d_ij km along the fastest route. → GeoDMS OD matrix and choice set
  • Tail / tail stop — the grid points above the S2 bracket (S3 material); once both scenarios are pinned the first time-out ends the sweep. In code: env SWEEP_STOP_AFTER_TAIL_FAIL (1). → Scenarios S1 S2 S3
  • Tangency — see supporting line.
  • target_cells — see baseline.
  • Time-out — a solve that hits the per-solve limit; the point is skipped (termination_status=TIME_LIMIT) and HiGHS is left on an aborted basis that poisons later warm starts. In code: LP_TIME_LIMIT (env, default 3 600 s) in lp_run.jl. → Code walkthrough lp_run
  • Top-p — the simplest rounding: open the p candidates with the largest y*; blind to geography; multistart's first seed. In code: topp_set; log travel_topp. → From LP fractions to real pharmacies
  • Travel cost / total travel cost — Σ_i pop_i · c(t_i) over all clients, stranded ones at BIG; the y-axis of the frontier. In code: log line total travel cost(c) (baseline) and the travel_* columns. → Travel cost functions
  • travel_relax — the LP's coverage-honest travel: served travel plus BIG times each client's unserved share — the certified lower bound, not HiGHS's printed objective. In code: relax in deck_data.json. → The lambda sweep and the Pareto frontier
  • travel_topp / travel_greedy / travel_multi — the coverage-honest travel of the three rounded sets at a point; travel_multi ≤ min(travel_topp, travel_greedy) always. In code: topp, greedy, multi in deck_data.json. → Log lines and output files
  • Typology (T1–T4, HIGH / MEDIUM / LOW) — two country attributes from the JRC pharmacy-regulation deck: the archetype (T1–T4) and the formalisation level, which colours the charts. In code: doc/policy_typology.py → policy_typology.csv. → Metrics aggregation and ranking

U

  • UFL (uncapacitated facility location) — minimise assignment cost plus a fixed cost per open site, no capacity limits; this model is a UFL with the same λ everywhere and an outside option (stranding at BIG). → Background theory
  • Union w-grid — see aggregate.
  • Units — seconds on disk, minutes in Julia; travel cost in person-minutes (LINEAR) or dimensionless (LOGISTIC; the two are never comparable); λ in the travel unit, labelled €; counts in cells; population in residents (total_pop). → Log lines and output files
  • Unserved — see stranded.

V

  • Vertex — see basis.
  • Vertex substitution — see swap.

W

  • w — the dimensionless price knob the sweep turns; λ = w · €100 000. Grid values print exactly (0.2), bisection points rounded (0.34087) while their Arrow folder carries the full float. In code: first column of the log row; folder label w=<w>. → The lambda sweep and the Pareto frontier
  • w-grid (1-2-5 per decade, common grid) — the fixed ladder every area is swept on: 1, 2, 5 × 10^d from 1e-4 to w_max (5.0 LINEAR, 0.5 LOGISTIC); common so the aggregate can sum at equal λ. In code: ws_common; env SWEEP_WMAX, SWEEP_MULTS. → The lambda sweep and the Pareto frontier
  • Warm start — starting the next solve from the previous optimal basis; cheap between neighbouring λ because only the M y-coefficients changed. Why the driver is sequential. In code: set_objective_coefficient in solve_at_w!. → The lambda sweep and the Pareto frontier
  • WarmStartState — the struct that keeps the live JuMP/HiGHS model, its variables and the CountryData together across the sweep. In code: struct WarmStartState{TX, TY} in lp_run.jl (type-parameter names predate the x/y swap). → Code walkthrough lp_run

X

  • x_ij / y_j — the LP's two kinds of variable: x_ij, the share of client i's residents served by facility j (one per OD row), and y_j, how open facility j is (one per candidate, in [0, 1]); swapped in everything older than 24 July 2026. → The facility location model explained

In the code

Where the names above cluster:

Kind of term Open
Prices, cost functions, BIG, loaders, CountryData, region exclusion settings.jl — Code walkthrough settings
LP variables, objective, warm start, solve_at_w!, the three roundings lp_run.jl — Code walkthrough lp_run
Grid, baseline, brackets, bisection, light rows, closest, Arrow writers, log columns lambda_sweep_simplex.jl — Code walkthrough lambda_sweep_simplex
Clients, candidates, the OD call, landbody rule, exports, read-back, descriptives cfg/main/Analyses.dms, Templates.dms, SourceData/Locations.dms, ModelParameters.dms — GeoDMS OD matrix and choice set
Aggregate, rectangle, crossing, typology, deck doc/build_charts.py, frontier_metrics.py, frontier_charts.py, build_deck.mjs — Deck and charts pipeline
Runners, steps, tags, .mmd purge run_*.ps1, run_*.bat — Running a sweep end to end

Permalink to the documented commit: https://github.com/ObjectVision/NetworkModel_EU/blob/e25416b/.

Pitfalls and open questions

  • Three S1s. "S1" on a region-slide table (snapped row), in the λ tables (chord) and in an exported location file (bisected/snapped solve) are three different numbers; Scenarios S1 S2 S3 has the matrix. The same goes for S2.
  • Two meanings of "stranded" (baseline unreachable vs rounding-stranded), two of "degeneracy" (simplex vs the reviewer's), two of "connector link" (location-to-road vs the clean-up's ConnectorLink) and two of "kink" (the #54 artefact vs the logistic's bends) — see the entries.
  • x/y, sum_x/frac_x in anything dated before 24 July 2026 — swap mentally.
  • "Pinned" is not "within tolerance"; "bracketed" in a CSV flag is a weaker statement than a closed bracket in the driver.
  • fac_€ is not the facility term of the objective, and mean_t is not coverage-consistent.
  • Terms whose rationale is unrecorded (why 5 nearest, 120 min, 1-2-5 spacing, MS_* values, REFINE_TOL 0.2 %, EXCLUDE_SHARE 0.5, the €100 000 placeholder, lambda_cross linear in w): the entries say so rather than invent one; the open list is on Decision log.

See also

Clone this wiki locally