lit-matrix
The matrix is the single place where "who did what, under which parameters, with which numbers" lives. Writers copy from it; the falsifier attacks it; baselines are chosen from it.
Format (in LITERATURE.md)
## Related-work matrix
| work | venue/year | setting | technique | asymptotics | concrete numbers + source | code? | our delta |
|---|---|---|---|---|---|---|---|
| [@RS21] VOLE-PSI | EC 2021 | 2PC semi-honest, n=2^20 | VOLE + OKVS | O(n) comm | 51 MB, 1.3 s LAN (Tab. 3) | yes https://... | we ... |
Column rules (enforced by scripts/lit_matrix.py):
- work — bib key (
[@key]or\cite{key}) or ePrint id, plus a short human name. - venue/year — venue of the version you cite (ePrint-only is fine, say so).
- setting — everything needed to compare fairly: parameter set / security level / threat model / network model / ring dimension / field / number of parties / hardware.
- technique — ≤12 words; the mechanism, not the marketing.
- asymptotics — the cost measure the community uses (ops, mults, depth, rounds, key-switches, comm).
- concrete numbers + source — numbers with a pointer (Tab./Fig./Sec./p.) or
measured E<n>(an EVIDENCE row, if we re-ran it). Numbers from a paper and numbers we measured are never mixed in one cell without saying so. - code? —
yes <url>/no/unknown. - our delta — what we do that this row does not.
noneis an honest answer and triggers a PI decision.
Procedure
- Seed from
references/reading-lists/<area>.md, then expand witheprint-search(last 12 months first). - One row per result (a paper with two incomparable results gets two rows).
- Fill numbers only from the paper's tables/figures; if the paper's parameters differ from ours, note it in setting instead of "normalising" silently.
- Mark the 2–4 rows that are the real competitors (bold the work cell);
baseline-engineerpins those. - Lint:
python3 skills/lit-matrix/scripts/lit_matrix.py LITERATURE.md --bib refs.bib. python3 skills/lit-matrix/scripts/lit_matrix.py --templateprints an empty skeleton.
Reading the matrix for gaps
Sort by setting, then look for (a) settings with no row (application gap — weakest kind of novelty, must be
argued), (b) a technique used in one setting but not the neighbouring one (transfer), (c) a parameter every row
fixes to the same value (a "hidden fixed parameter" — hand to idea-miner), (d) asymptotics that nobody has a
lower bound for.