thedixitjain ITCS Experiments
ITCS is a pure-theory venue: the primary — usually the only — evidence for a claim is a proof. There is no experiments requirement, no benchmark, no leaderboard, and no reviewer expectation of an empirical section. Bringing an ML-conference or SE reflex here (a table of numbers "showing" the method works) misreads the venue: at ITCS a theorem is proved, not measured. This skill is about matching evidence to a theory claim and about the rare case where a small computation genuinely helps.
Proofs are the evidence
- Every central claim is settled by a complete proof, not by examples. A pattern that "holds in all cases we tried" is a conjecture, not a theorem — label it as such and prove or drop it.
- Match the claim shape to the argument shape: an upper bound needs a construction + analysis; a lower bound needs an adversary/reduction; a separation needs a witness object; an impossibility needs a contradiction from the assumption. Reviewers check that the kind of argument fits the kind of claim.
- A new model needs an anchoring result (see
itcs-writing-style) — a separation or a surprising possibility that proves the model is neither empty nor everything. That anchoring result is the "experiment" an ITCS PC wants: evidence the model is alive.
Worked examples and separations as evidence
Non-proof evidence that is welcome, because it makes the mathematics concrete:
- Worked examples that instantiate a definition on a small case and show the intended behavior — invaluable for a new model, and cheap insurance against the "is this trivial?" objection.
- Explicit separating objects — a small graph, code, distribution, or gadget that witnesses a gap between two settings. If finite, include the object so a reviewer verifies the separation directly.
- Tight-example constructions showing an analysis cannot be improved — the theory analogue of an ablation, demonstrating the bound is not loose by accident.
The rare, well-scoped computation
Some ITCS papers include a small computational component: a computer search that found a gadget, a SAT/SMT solve certifying a finite separation, a numerically evaluated construction. When one genuinely helps, scope it tightly:
- It supports a proved claim; it is never the claim. "A search over all graphs on <= 12 vertices found the gadget of Lemma 4, whose properties we then prove" is legitimate. "Our method achieves 92% on a benchmark" is a category error at ITCS.
- Make it checkable without rerunning. State the exact search space, the tool and version, and — crucially — include the finite object the search produced (the graph, certificate, code) so verification is a static check, not a re-computation. A reviewer should be able to confirm the object has the claimed property by hand or with a one-line check.
- Report it honestly. If a construction is only verified numerically (not proved), say so and mark exactly which claims rest on computation versus proof.
- Keep it off the anonymity leak surface. A linked repository under a personal GitHub is a
lightweight-double-blind slip; fold the object into an appendix or host it neutrally (see
itcs-submission).
What NOT to import from empirical venues
| Empirical-venue habit | Why it misfires at ITCS |
|---|---|
| A benchmark table as the main result | ITCS proves; it does not measure. A table cannot establish a theorem |
| "Outperforms baselines by X%" framing | There are no baselines to beat; the contribution is an idea/proof |
| Runtime plots to argue efficiency | State and prove the asymptotic bound instead |
| An artifact/reproducibility package of code | No artifact track exists; the "artifact" is the proof (see itcs-artifact-evaluation) |
| Statistical significance / error bars | Irrelevant to a deterministic mathematical claim |
Decision procedure
[Claim] is it a theorem (prove it) or a pattern (label as conjecture, or prove/drop)?
[Argument fit] upper=construction+analysis / lower=adversary / separation=witness / impossibility=contradiction
[Alive] new model? -> anchoring separation or surprising-possibility result present?
[Compute?] does a small search/solve genuinely help a proved claim? if not, omit it
[Checkable] if compute used: search space + tool/version stated, finite object included?
[Honesty] each claim tagged proved vs. numerically-verified; anonymity leak surface clean?
Output format
[ITCS evidence status] proof-complete / gaps / mis-imported-empirics
[Central claims] each has a complete proof of matching shape? yes/no + list gaps
[Model alive] anchoring result present for any new model? yes/no
[Computation] present? if so: supports-a-proof only? checkable object included?
[Anonymity] no personal-repo leak from any computational content? yes/no
[Fix queue] <ordered edits>
Source: brycewang-stanford/Awesome-Journal-Skills → ITCS-Skills/skills/itcs-experiments/SKILL.md