Skip to content

[P2 long-horizon probe] Location without shape: explicit toy CDF families #624

Description

@LightChainr

Probe MID — Location without shape: explicit toy families

Assign to: ordinary (adequate) reasoning, ordinary CPU. Analysis + numpy/mpmath. No percolation engine, no arXiv, no N=725, no #612 rescoring.

This is a construction probe. The deliverable is families you invent, not a match to a repository JSON.


Standing

Frontier: claude/matching-one-workspace-pwr5pv @ 8b5f9d1a.
Theorem L as a cited fact: notes/p613-quantile-convergence-20260907.md — location, not rate. You do not re-prove it.


North-star

Exhibit explicit sequences of CDFs F_N (or quantile functions Q_N) on [0,1] such that

Q_N(u) → p_*     uniformly on every compact [ε, 1−ε] ⊂ (0,1)

for some p_* ∈ (0,1), while the Aff(1)-invariant shape

Z_N(u) = (Q_N(u) − Q_N(a)) / (Q_N(b) − Q_N(a))

(with declared anchors, default a=0.2, b=0.8) does one of the following, by construction:

target what you must exhibit
T1 Z_N fails to converge (oscillation, two accumulation points, or N-dependent anchors that still oscillate after a fixed-anchor definition)
T2 Z_N converges, and the limit Z_∞ can be prescribed: for any continuous strictly increasing ζ with ζ(a)=0, ζ(b)=1, there is a family with that Z_∞
T3 Z_N converges, but the limit depends on the anchors (a,b) in a way that is not a reparametrisation (then “the” shape is not well-posed even for toys)

If T2 is achieved, W1 of #622 becomes a theorem about location statements in general: Theorem L cannot pin percolation’s shape. That is the point of this probe.

If you believe T2 is false for monotone CDFs with a single crossing of every u∈(ε,1−ε), prove that restriction and still do T1.


Constraints on the toys (so they are fair analogues)

Each family must be:

  • a CDF in p for each N: F_N(0)=0, F_N(1)=1, nondecreasing (cadlag allowed; say so);
  • strictly increasing on a neighbourhood of p_* for large N, so Q_N(u) is a singleton for u∈[ε,1−ε];
  • not a mixture of two distant jumps that “cheats” location by parking mass at 0 and 1 only — at least one of the families must have a window of width w_N → 0 about p_* that carries all interior quantiles (this is what Theorem L actually looks like);
  • parametrised in closed form, or by a 20-line sampler of a named distribution. No black-box neural CDF.

Write the two Aff(1) actions and say which one your Z quotients:

on p:   p ↦ αp+β
on Q:   Q(u) ↦ α Q(u)+β

Z as written quotients the Q-action. One family should also be inspected under a warp of the p-axis, to show the two quotients disagree (a numerical example is enough; the theorem is #622 W5).


Directions (all three targets attempted; T2 is the prize)

D1 — Logistic / probit window (baseline that does pin a shape)

F_N(p) = σ( (p − p_*) / w_N ),    w_N → 0,   σ a fixed logistic or Gaussian CDF.

Compute Z_N. It will converge to a fixed logistic/probit shape independent of w_N. This is the family people silently imagine when they say “sharp threshold ⇒ universal shape”. Record it as the non-example: location plus a fixed window profile pins Z. Theorem L does not give you a fixed profile.

D2 — Prescribed shape (T2)

Replace σ by a sequence σ_N, or by a mixture of two window profiles whose weights oscillate, or by an N-dependent skew (e.g. GEV / skew-logistic with a parameter γ_N).

Goal: pick any target ζ (give three: symmetric logistic; a strongly skew one; a piecewise-linear “kink”), build F_N with window w_N = N^{-1} or N^{-3/4} (the exponent is free in a toy), prove Q_N(u)→p_* uniformly on [ε,1−ε], and prove Z_N → ζ.

If a monotone one-parameter exponential family cannot do T2, say so and use a two-parameter window (location + skew), still with width → 0.

D3 — Oscillation (T1)

Two profiles σ and τ with different Z_∞, switched on even/odd N, or a skew parameter γ_N = sin(log N). Prove location still holds. Then Z_N has no limit. This kills “sharp ⇒ shape converges”.

D4 — Anchor pathology (T3)

Using one D2 family, recompute Z with anchors (0.1,0.9) vs (0.3,0.7) vs (0.2,0.8). If the three Z_∞ differ by more than reparametrisation of u (make that precise: they do not lie on one orbit under increasing maps that fix {a,b} — or they do). Either “anchors are gauge” or “the invariant is not well-posed”. One page.

D5 — What extra input would pin ζ in percolation

After T1/T2, one page, no literature search. Named extra inputs already in the #613/#606 notes (RSW, four-arm / F1, self-duality). For each: would it pin σ (the window profile), or only the width? Write Y/N/unknown. Unknown is allowed. Do not quote new papers.


Compute

Grid u = 0.05(0.05)0.95, N in {2^k} over at least two decades of w_N. JSON of Q_N, Z_N for each family. Plots optional as committed png under results/, not required.

No Monte Carlo of percolation. No L=3 enumerator.


Stop rules


Deliverables

notes/probe-location-without-shape-YYYYMMDD.md
  formulas, proofs of location, T1/T2/T3 verdict, D5 table
scripts/probe/toy_cdf_families.py
results/probe-location-without-shape/latest.json

Interface

#622 uses this as the W1 laboratory. #618 may cite T2 as “location does not imply a common ω for the 9-vector” — only if you actually proved T2. #620 is not involved.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    priority:P2Deferred research or on-demand support; no default new compute allocation.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions