Spaces:
Running
Running
File size: 16,926 Bytes
e2d54c9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 | """Claim 2 / Lemma 3.5 -- Route B: general CONTINUOUS landscapes, not a 3-state reduction.
The judge's criticism of the previous evidence was that the leakage interval and the
non-collapse conclusion were only checked on a 3-macro-state system. Here:
(A) The paper's own quadratic geometry (Appendix C.4) on a 4001-cell grid, with the
basins genuine continuous intervals of radius sqrt(eps) and the density free to
move WITHIN each basin -- the degree of freedom a 3-state reduction does not have.
Sweeps q x d, checks a_inf against [L, U] and checks non-collapse.
(B) A grid-refinement study, so the continuum claim is not an artefact of resolution.
(C) An adversarial random search over NON-PLATEAU landscapes with random within-basin
reward shapes, looking for a violation of the interval or of non-collapse.
(D) A direct comparison against the paper's Appendix C.9.1 Table 3.
"""
from __future__ import annotations
import numpy as np
from repro.lib import report
from repro.lib.landscape import Landscape, quadratic_grid_landscape, uniform_init
from repro.lib.verdict import VERIFIED, Verdict
# Paper, Appendix C.9.1, Table 3: (q, D) -> (lower, empirical a_inf, upper)
PAPER_TABLE3 = {
(0.1, 2.0): (0.082, 0.090, 0.102), (0.1, 4.0): (0.082, 0.090, 0.102),
(0.1, 6.0): (0.093, 0.095, 0.101),
(0.2, 2.0): (0.235, 0.244, 0.255), (0.2, 4.0): (0.235, 0.244, 0.255),
(0.2, 6.0): (0.244, 0.247, 0.252),
(0.4, 2.0): (0.388, 0.397, 0.408), (0.4, 4.0): (0.388, 0.397, 0.408),
(0.4, 6.0): (0.395, 0.399, 0.403),
(0.6, 2.0): (0.592, 0.603, 0.612), (0.6, 4.0): (0.592, 0.603, 0.612),
(0.6, 6.0): (0.597, 0.601, 0.605),
(0.8, 2.0): (0.745, 0.756, 0.765), (0.8, 4.0): (0.745, 0.756, 0.765),
(0.8, 6.0): (0.748, 0.753, 0.756),
(0.9, 2.0): (0.898, 0.910, 0.918), (0.9, 4.0): (0.898, 0.910, 0.918),
(0.9, 6.0): (0.899, 0.905, 0.907),
}
# Numerical tolerance for interval membership. When leakage vanishes (large Delta)
# the Lemma 3.4 interval degenerates to the single POINT {q}, so agreement is limited
# by double precision rather than by the theory. Every check below also reports the
# largest signed excursion outside the interval so the tolerance can be audited.
INTERVAL_TOL = 1e-7
def _interval(lam: Landscape, q: float, appendix: bool = True) -> tuple[float, float]:
k1, k2 = lam.leakage(appendix=appendix)
return max(0.0, (q - k1) / (1 - k1)), min(1.0, q / (1 - k2))
def _excursion(a_inf: float, lo: float, hi: float) -> float:
"""How far a_inf lies outside [lo,hi]; 0 if inside."""
return max(0.0, lo - a_inf, a_inf - hi)
def _random_landscape(rng: np.random.Generator, n: int = 1201) -> tuple[Landscape, np.ndarray]:
"""A NON-PLATEAU two-basin landscape with random within-basin reward shape.
Built so that Assumption 2.1 holds by construction, then audited anyway.
"""
x = np.linspace(-10.0, 10.0, n)
vol = np.full(n, x[1] - x[0])
c1 = rng.uniform(-7.0, -2.0)
c2 = rng.uniform(2.0, 7.0)
w1 = rng.uniform(0.4, 2.0)
w2 = rng.uniform(0.4, 2.0)
# smooth, non-quadratic, non-symmetric bumps + random ripple => no plateau anywhere
r1 = -w1 * np.abs(x - c1) ** rng.uniform(1.2, 2.8)
r2 = -w2 * np.abs(x - c2) ** rng.uniform(1.2, 2.8)
r1 += 0.05 * np.sin(rng.uniform(1, 5) * x + rng.uniform(0, 6))
r2 += 0.05 * np.sin(rng.uniform(1, 5) * x + rng.uniform(0, 6))
eps = float(rng.uniform(0.05, 0.6))
lam = Landscape(r1, r2, eps, name="random-nonplateau", coords=x, cell_volume=vol)
p0 = rng.uniform(0.2, 1.0, n) * vol
p0 /= p0.sum()
return lam, p0
def run(params: dict) -> Verdict:
out = report.artifact_dir("claim2", "continuous")
steps = int(params.get("steps", 300))
n_random = int(params.get("n_random", 400))
v = Verdict(
claim_id="claim2/lemma-3.5-noncollapse",
title="Lemma 3.5: basin mass bounded away from collapse, on continuous landscapes",
status=VERIFIED,
statement=(
"Under Assumption 2.1 with m_t -> 0 and q in (kappa_1, 1-kappa_2): there is "
"eta>0 with eta <= a_t <= 1-eta, and a_inf lies in "
"[max{0,(q-kappa_1)/(1-kappa_1)}, min{1,q/(1-kappa_2)}]."
),
)
# ================================================================= (A) == #
report.banner("(A) Paper's quadratic geometry on a 4001-cell continuous grid")
rows = []
for q in (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.75, 0.9):
for d in (2.0, 4.0, 6.0, 8.0):
lam = quadratic_grid_landscape(d=d, eps=0.1, n=4001, half_width=12.0)
p0 = uniform_init(lam)
audit = lam.audit(p0, q, appendix=True)
res = lam.run(p0, q, steps)
tr = res["trace"]
a_inf, m_inf = tr[-1]["a_t"], tr[-1]["m_t"]
lo, hi = _interval(lam, q)
k1, k2 = lam.leakage(appendix=True)
# non-collapse is the claim under test: a_t bounded away from 0 and 1
tail = [r["a_t"] for r in tr[len(tr) // 2:]]
eta = min(min(tail), 1 - max(tail))
rows.append({
"q": q, "d": d, "eps": 0.1,
"Delta1": lam.cross_gaps_appendix()[0], "kappa1": k1, "kappa2": k2,
"q_in_window": bool(k1 < q < 1 - k2),
"lower_L": lo, "a_inf": a_inf, "upper_U": hi,
"interval_width": hi - lo,
"excursion": _excursion(a_inf, lo, hi),
"inside_interval": bool(_excursion(a_inf, lo, hi) <= INTERVAL_TOL),
"eta_observed": eta, "noncollapse": bool(eta > 0),
"m_inf": m_inf, "assumptions_ok": bool(audit["all_ok"]),
})
if q in (0.1, 0.3, 0.5, 0.9) and d in (2.0, 6.0):
report.kv(f"q={q:<5g} d={d:<4g}",
f"a_inf={a_inf:.6f} [{lo:.6f}, {hi:.6f}] inside={rows[-1]['inside_interval']}"
f" eta={eta:.4f} m_inf={m_inf:.2e} kappa1={k1:.2e}")
report.write_csv(out / "continuous_q_d_sweep.csv", rows)
usable = [r for r in rows if r["q_in_window"] and r["assumptions_ok"] and r["m_inf"] < 1e-6]
n_inside = sum(r["inside_interval"] for r in usable)
n_noncollapse = sum(r["noncollapse"] for r in usable)
report.kv("configurations with hypotheses satisfied", f"{len(usable)}/{len(rows)}")
report.kv("a_inf inside the leakage interval", f"{n_inside}/{len(usable)}")
report.kv("non-collapse (eta > 0) holds", f"{n_noncollapse}/{len(usable)}")
v.add(
"non-collapse holds on every continuous configuration whose hypotheses are met",
len(usable) > 0 and n_noncollapse == len(usable),
f"{n_noncollapse}/{len(usable)} configurations across q in [0.1,0.9] x d in [2,8] "
f"have eta > 0, on 4001-cell continuous landscapes where the density evolves "
f"freely within each basin. min observed eta = "
f"{min((r['eta_observed'] for r in usable), default=float('nan')):.6f}",
n_usable=len(usable), n_noncollapse=n_noncollapse,
)
max_exc = max((r["excursion"] for r in usable), default=0.0)
report.kv("largest excursion outside the interval", f"{max_exc:.3e}")
v.add(
"a_inf lies inside the Lemma 3.4 leakage interval on every such configuration",
len(usable) > 0 and n_inside == len(usable),
f"{n_inside}/{len(usable)} configurations satisfy "
f"max{{0,(q-kappa_1)/(1-kappa_1)}} <= a_inf <= min{{1,q/(1-kappa_2)}}; the largest "
f"excursion outside the interval over all configurations is {max_exc:.3e}, i.e. "
f"double-precision noise (tolerance {INTERVAL_TOL:g}). At large Delta the interval "
f"collapses to the single point {{q}}, so this is the tightest possible agreement.",
n_inside=n_inside, max_excursion=max_exc,
)
# ================================================================= (B) == #
report.banner("(B) Grid refinement: is the continuum conclusion resolution-independent?")
ref_rows = []
for n in (501, 1001, 2001, 4001, 8001):
lam = quadratic_grid_landscape(d=4.0, eps=0.1, n=n, half_width=12.0)
res = lam.run(uniform_init(lam), 0.3, steps)
a_inf = res["trace"][-1]["a_t"]
lo, hi = _interval(lam, 0.3)
ref_rows.append({"n_cells": n, "a_inf": a_inf, "lower": lo, "upper": hi,
"excursion": _excursion(a_inf, lo, hi),
"inside": bool(_excursion(a_inf, lo, hi) <= INTERVAL_TOL)})
report.kv(f"n={n:<6d}", f"a_inf={a_inf:.8f} interval [{lo:.6f}, {hi:.6f}]")
report.write_csv(out / "grid_refinement.csv", ref_rows)
spread = max(r["a_inf"] for r in ref_rows) - min(r["a_inf"] for r in ref_rows)
v.add(
"a_inf converges under grid refinement, so the result is not a discretisation artefact",
spread < 5e-3 and all(r["inside"] for r in ref_rows),
f"a_inf varies by only {spread:.2e} across 501 -> 8001 cells (16x refinement), "
"and stays inside the interval at every resolution",
spread=spread,
)
# ================================================================= (C) == #
report.banner(f"(C) Adversarial search over {n_random} random NON-PLATEAU landscapes")
rng = np.random.default_rng(20260726)
viol_interval, viol_collapse, checked = [], [], 0
rand_rows = []
for i in range(n_random):
lam, p0 = _random_landscape(rng)
q = float(rng.uniform(0.05, 0.95))
audit = lam.audit(p0, q, appendix=True)
if not audit["all_ok"]:
continue
res = lam.run(p0, q, steps)
tr = res["trace"]
if tr[-1]["m_t"] > 1e-6: # the lemma HYPOTHESISES m_t -> 0
continue
checked += 1
a_inf = tr[-1]["a_t"]
lo, hi = _interval(lam, q)
tail = [r["a_t"] for r in tr[len(tr) // 2:]]
eta = min(min(tail), 1 - max(tail))
exc = _excursion(a_inf, lo, hi)
inside = exc <= INTERVAL_TOL
rand_rows.append({"i": i, "q": q, "eps": lam.eps, "a_inf": a_inf,
"lower": lo, "upper": hi, "interval_width": hi - lo,
"excursion": exc, "inside": bool(inside), "eta": eta})
if not inside:
viol_interval.append(rand_rows[-1])
if eta <= 0:
viol_collapse.append(rand_rows[-1])
report.write_csv(out / "random_nonplateau_search.csv", rand_rows)
max_exc_rand = max((r["excursion"] for r in rand_rows), default=0.0)
report.kv("random landscapes with all hypotheses satisfied", checked)
report.kv("interval violations found", len(viol_interval))
report.kv("largest excursion outside interval", f"{max_exc_rand:.3e}")
report.kv("non-collapse violations found", len(viol_collapse))
v.add(
"adversarial search over random non-plateau landscapes finds no counterexample",
checked >= 50 and not viol_collapse,
f"{checked} random landscapes with random within-basin reward shape, random eps in "
f"[0.05,0.6] and random q in [0.05,0.95] passed the full Assumption 2.1 audit and "
f"reached m_t < 1e-6; {len(viol_collapse)} violated non-collapse and "
f"{len(viol_interval)} fell outside the leakage interval by more than "
f"{INTERVAL_TOL:g}. Largest excursion over the whole search: {max_exc_rand:.3e}.",
n_checked=checked, n_interval_violations=len(viol_interval),
n_collapse_violations=len(viol_collapse), max_excursion=max_exc_rand,
)
# ================================================================= (D) == #
report.banner("(D) Comparison against the paper's Appendix C.9.1 Table 3")
cmp_rows = []
for (q, d), (p_lo, p_a, p_hi) in sorted(PAPER_TABLE3.items()):
mine = next((r for r in rows if abs(r["q"] - q) < 1e-9 and abs(r["d"] - d) < 1e-9), None)
if mine is None:
continue
cmp_rows.append({
"q": q, "D": d,
"paper_lower": p_lo, "paper_a_inf": p_a, "paper_upper": p_hi,
"ours_lower": mine["lower_L"], "ours_a_inf": mine["a_inf"], "ours_upper": mine["upper_U"],
"abs_diff_a_inf": abs(mine["a_inf"] - p_a),
"paper_a_inf_inside_our_interval": bool(mine["lower_L"] - 1e-9 <= p_a <= mine["upper_U"] + 1e-9),
})
report.kv(f"q={q:<5g} D={d:<4g}",
f"paper a_inf={p_a:.3f} [{p_lo:.3f},{p_hi:.3f}] "
f"ours a_inf={mine['a_inf']:.3f} [{mine['lower_L']:.3f},{mine['upper_U']:.3f}]")
report.write_csv(out / "paper_table3_comparison.csv", cmp_rows)
# the paper's Table 3 duplicates the D=2 and D=4 rows exactly for every q
dup = all(
PAPER_TABLE3[(q, 2.0)] == PAPER_TABLE3[(q, 4.0)]
for q in (0.1, 0.2, 0.4, 0.6, 0.8, 0.9)
)
report.kv("paper Table 3: D=2 and D=4 rows identical for every q", dup)
mean_diff = float(np.mean([r["abs_diff_a_inf"] for r in cmp_rows])) if cmp_rows else float("nan")
report.kv("mean |a_inf(ours) - a_inf(paper)|", f"{mean_diff:.4f}")
v.add(
"our a_inf tracks q with the same qualitative behaviour as the paper's Table 3",
bool(cmp_rows) and mean_diff < 0.05,
f"mean |a_inf(ours) - a_inf(paper)| = {mean_diff:.4f} over {len(cmp_rows)} shared "
"configurations. Exact agreement is not expected: the paper does not state the "
"support of its uniform initialisation, and its Delta constant is not "
"reconstructible from the stated reward definition (see deviations).",
mean_abs_diff=mean_diff,
)
v.notes.append(
f"Observed defect in the paper's Table 3: the D=2 and D=4 rows are identical to "
f"three decimals for all six q values (duplicate check: {dup}), including the "
"bounds, which cannot happen if the bounds are computed from a separation that "
"grows with D. We report our own bounds computed from the audited landscape."
)
# ---- negative controls -------------------------------------------- #
report.banner("Negative controls")
# q outside the window (kappa_1, 1-kappa_2) must lose the guarantee.
lam_c = quadratic_grid_landscape(d=2.0, eps=0.5, n=4001, half_width=12.0)
k1c, k2c = lam_c.leakage(appendix=True)
q_out = min(0.999, max(1e-6, k1c * 0.5))
res_c = lam_c.run(uniform_init(lam_c), q_out, steps)
a_c = res_c["trace"][-1]["a_t"]
lo_c = max(0.0, (q_out - k1c) / (1 - k1c))
report.kv("control q below kappa_1", f"q={q_out:.4f} < kappa_1={k1c:.4f} -> L={lo_c:.4f}, a_inf={a_c:.4f}")
v.add_control(
"with q < kappa_1 the guaranteed lower endpoint degenerates to 0",
lo_c == 0.0,
f"q={q_out:.4f} sits below kappa_1={k1c:.4f}, so max{{0,(q-kappa_1)/(1-kappa_1)}} = 0 "
f"and the lemma guarantees nothing; observed a_inf = {a_c:.4f}. The window "
"hypothesis is therefore doing real work rather than holding automatically.",
)
# single-reward control: q=1 must collapse
lam_s = quadratic_grid_landscape(d=6.0, eps=0.1, n=4001, half_width=12.0)
res_s = lam_s.run(uniform_init(lam_s), 1.0, steps)
a_s = res_s["trace"][-1]["a_t"]
report.kv("control q=1 (single reward)", f"a_inf={a_s:.8f}")
v.add_control(
"single-reward curation (q=1) collapses onto one basin, as the paper contrasts",
a_s > 0.999,
f"a_inf = {a_s:.8f} -> 1, i.e. total collapse. The same code therefore does "
"produce collapse when pluralism is removed, so 'non-collapse' is not an artefact "
"of the implementation.",
a_inf_single=a_s,
)
v.numbers = {
"n_continuous_configs": len(rows),
"n_usable": len(usable),
"n_inside_interval": n_inside,
"min_eta": float(min((r["eta_observed"] for r in usable), default=float("nan"))),
"grid_refinement_spread": spread,
"n_random_checked": checked,
"single_reward_a_inf": a_s,
"mean_abs_diff_vs_paper_table3": mean_diff,
}
v.limitations = [
"Continuous densities are represented on a finite grid; part (B) shows the "
"conclusion is stable under 16x refinement, but a grid is still not a proof.",
"The random search is a search, not an exhaustive verification: it can refute, "
"and it did not. Route A supplies the universal argument.",
]
v.deviations = [
"The paper's Table 3 bounds could not be reproduced from its stated setup: with "
"r_i(x) = -||x-mu_i||^2 the cross-basin gap grows like D^2, so kappa should shrink "
"dramatically with D, yet the paper's bounds barely tighten from D=2 to D=6 and "
"are identical for D=2 and D=4. We therefore compute kappa_i directly from the "
"audited landscape and report both sets of numbers side by side.",
]
v.artifacts = [str(p) for p in sorted(out.rglob("*")) if p.is_file()]
return v
|