b3EvCd8sYE / code /tools /figures.py
DineshAI's picture
Publish the round-3 verification: all six claims with raw data, checkers, controls and a visibility matrix
59e1897 verified
Raw
History Blame Contribute Delete
17.2 kB
"""Figures for the reproduction report, drawn from the raw artifact CSVs.
Every figure reads the same CSV files that the pages link to, so a figure can
never disagree with the numbers printed beside it. Both an SVG (for the
Hugging Face logbook, which takes text uploads only) and a PNG (for the GitHub
report) are written for each figure.
"""
from __future__ import annotations
import csv
import os
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
# A restrained, colour-blind-safe qualitative ramp; one hue per curve.
PALETTE = ["#3B6FD4", "#D9822B", "#2E9E7E", "#B5478F", "#7A6BD1", "#B0543C"]
GRID = dict(color="#D6D9DE", linewidth=0.7)
def _style(ax, xlabel, ylabel, title=None):
ax.set_facecolor("white")
ax.grid(True, **GRID)
ax.set_axisbelow(True)
for side in ("top", "right"):
ax.spines[side].set_visible(False)
for side in ("left", "bottom"):
ax.spines[side].set_color("#8A8F98")
ax.tick_params(colors="#41464F", labelsize=9)
ax.set_xlabel(xlabel, fontsize=10, color="#24272C")
ax.set_ylabel(ylabel, fontsize=10, color="#24272C")
if title:
ax.set_title(title, fontsize=11, color="#14171A", pad=8)
def _read(path):
with open(path) as fh:
return list(csv.DictReader(fh))
def _save(fig, out_dir, name):
os.makedirs(out_dir, exist_ok=True)
svg = os.path.join(out_dir, f"{name}.svg")
png = os.path.join(out_dir, f"{name}.png")
fig.savefig(svg, format="svg", bbox_inches="tight", facecolor="white")
fig.savefig(png, format="png", dpi=160, bbox_inches="tight", facecolor="white")
plt.close(fig)
return svg, png
# --------------------------------------------------------------------------
def figure2(art: str, out: str):
"""Headline: fraction of positive compatibility scores vs injected error."""
rows = _read(os.path.join(art, "claim4", "figure2_synthetic.csv"))
panels = ["m", "n", "p"]
labels = {"m": "hidden variables $m$", "n": "observed variables $n$",
"p": "edge density $p$"}
fig, axes = plt.subplots(1, 3, figsize=(12.5, 3.9), sharey=True)
for ax, panel in zip(axes, panels):
sub = [r for r in rows if r["panel"] == panel]
values = sorted({float(r["value"]) for r in sub})
for k, val in enumerate(values):
cur = sorted([r for r in sub if float(r["value"]) == val],
key=lambda r: float(r["sigma"]))
x = [float(r["sigma"]) for r in cur]
y = [float(r["frac_positive"]) for r in cur]
lo = [float(r["frac_ci_lo"]) for r in cur]
hi = [float(r["frac_ci_hi"]) for r in cur]
c = PALETTE[k % len(PALETTE)]
ax.plot(x, y, "-o", color=c, markersize=3.6, linewidth=1.7,
label=f"{panel} = {val:g}")
ax.fill_between(x, lo, hi, color=c, alpha=0.15, linewidth=0)
_style(ax, r"injected error variance $\sigma$", "", labels[panel])
ax.legend(frameon=False, fontsize=8.5)
ax.set_ylim(-0.03, 1.05)
axes[0].set_ylabel("fraction of lists with\npositive compatibility score",
fontsize=10, color="#24272C")
fig.suptitle("Reproduction of Figure 2 — the fraction of positive scores "
"strictly decreases as statements degrade",
fontsize=12.5, color="#14171A", y=1.04)
return _save(fig, out, "figure2_fraction_positive")
def figure5(art: str, out: str):
"""Mean heuristic incompatibility vs number of injected graph errors."""
rows = _read(os.path.join(art, "claim6", "figure5_monotonicity.csv"))
panels = ["m", "n", "p"]
labels = {"m": "hidden variables $m$", "n": "observed variables $n$",
"p": "edge density $p$"}
fig, axes = plt.subplots(1, 3, figsize=(12.5, 3.9))
for ax, panel in zip(axes, panels):
sub = [r for r in rows if r["panel"] == panel]
values = sorted({float(r["value"]) for r in sub})
for k, val in enumerate(values):
cur = sorted([r for r in sub if float(r["value"]) == val],
key=lambda r: int(r["n_errors"]))
x = [int(r["n_errors"]) for r in cur]
y = [float(r["mean_score"]) for r in cur]
lo = [float(r["ci_lo"]) for r in cur]
hi = [float(r["ci_hi"]) for r in cur]
c = PALETTE[k % len(PALETTE)]
ax.plot(x, y, "-o", color=c, markersize=3.6, linewidth=1.7,
label=f"{panel} = {val:g}")
ax.fill_between(x, lo, hi, color=c, alpha=0.15, linewidth=0)
ax.plot(x, x, "--", color="#8A8F98", linewidth=1.2,
label="true error count")
_style(ax, "injected errors", "", labels[panel])
ax.legend(frameon=False, fontsize=8.5)
axes[0].set_ylabel("mean heuristic\nincompatibility $c(G)$", fontsize=10,
color="#24272C")
fig.suptitle("Reproduction of Figure 5 — incompatibility increases "
"monotonically with injected errors",
fontsize=12.5, color="#14171A", y=1.04)
return _save(fig, out, "figure5_monotonicity")
def sample_complexity(art: str, out: str):
"""Measured minimum sample size against the theorem's predicted scaling."""
rows = _read(os.path.join(art, "claim3", "min_sample_complexity.csv"))
fig, axes = plt.subplots(1, 2, figsize=(11.5, 4.1))
ax = axes[0]
sweeps = sorted({r["sweep"] for r in rows})
for k, sw in enumerate(sweeps):
sub = sorted([r for r in rows if r["sweep"] == sw],
key=lambda r: float(r["sweep_value"]))
pred = [float(r["n"]) ** 4 * (1 + float(r["a"]) + float(r["b"])) ** 4
* float(r["V"]) ** 4 / float(r["eps"]) ** 2
* np.log(float(r["n"]) / float(r["delta"])) for r in sub]
meas = [float(r["N_star_median"]) for r in sub]
ax.loglog(pred, meas, "o", color=PALETTE[k % len(PALETTE)],
markersize=6, label=sw, alpha=0.85)
lim = ax.get_xlim()
xs = np.geomspace(max(lim[0], 1e-3), lim[1], 50)
for c_, style in ((1e-3, "-"), (1e-4, "--"), (1e-5, ":")):
ax.loglog(xs, c_ * xs, style, color="#8A8F98", linewidth=1.1,
label=f"$N = {c_:g}\\times$ bound")
_style(ax, r"theorem bound $n^4(1{+}a{+}b)^4V^4\varepsilon^{-2}\log(n/\delta)$",
"measured minimum $N^*$",
"Measured sample complexity vs the theorem's bound")
ax.legend(frameon=False, fontsize=8, ncol=2)
ax = axes[1]
import json
with open(os.path.join(art, "claim3", "rates.json")) as fh:
rates = json.load(fh)
jf = rates["joint_fit"]
names = list(jf)
xs = np.arange(len(names))
vals = [jf[k]["exponent"] for k in names]
errs = [1.96 * jf[k]["stderr"] for k in names]
caps = [jf[k]["cap"] for k in names]
ax.bar(xs - 0.19, vals, 0.36, yerr=errs, capsize=3, color=PALETTE[0],
label="measured exponent (95% CI)")
ax.bar(xs + 0.19, caps, 0.36, color="#C9CDD3", label="theorem's exponent")
ax.set_xticks(xs)
ax.set_xticklabels(["$n$", "$1{+}a{+}b$", "$V$", r"$1/\varepsilon$",
r"$\log(n/\delta)$"], fontsize=9)
ax.axhline(0, color="#8A8F98", linewidth=0.8)
_style(ax, "", "exponent in $\\log N^*$",
"Joint-fit exponents never exceed the theorem's")
ax.legend(frameon=False, fontsize=8.5)
return _save(fig, out, "sample_complexity")
def expected_compatibility(art: str, out: str):
"""Theorem 2.9: E[comp] > 0 across dimensions and distribution families."""
rows = _read(os.path.join(art, "claim2", "expected_compatibility.csv"))
gating = [r for r in rows if r["gating"] == "True"]
fams = sorted({r["coef_family"] for r in gating})
dims = sorted({int(r["n"]) for r in gating})
fig, ax = plt.subplots(figsize=(9.5, 4.3))
for k, fam in enumerate(fams):
for noise, marker in (("diag_exponential", "o"), ("wishart_dense", "s")):
sub = sorted([r for r in gating if r["coef_family"] == fam
and r["noise_family"] == noise],
key=lambda r: int(r["n"]))
if not sub:
continue
x = [int(r["n"]) for r in sub]
y = [float(r["mean_comp"]) for r in sub]
lo = [float(r["comp_ci_lo"]) for r in sub]
ax.plot(x, y, marker=marker, linestyle="-", color=PALETTE[k % len(PALETTE)],
markersize=4.5, linewidth=1.4, alpha=0.9,
label=f"{fam} / {noise.split('_')[0]}")
ax.fill_between(x, lo, [float(r["comp_ci_hi"]) for r in sub],
color=PALETTE[k % len(PALETTE)], alpha=0.10, linewidth=0)
ax.axhline(0, color="#B0543C", linewidth=1.3, linestyle="--",
label="zero (theorem: strictly above)")
ax.set_yscale("symlog", linthresh=1e-2)
ax.set_xticks(dims)
_style(ax, "number of variables $n$", r"$\mathbb{E}[\mathrm{comp}(\Sigma_X, A)]$",
"Theorem 2.9 — expected compatibility of TRUE statements is positive\n"
"(Bonferroni-corrected 95% intervals, every gating configuration)")
ax.legend(frameon=False, fontsize=7.5, ncol=2, loc="upper left")
return _save(fig, out, "expected_compatibility")
def llm_scores(art: str, out: str):
"""LLM statement quality against model capacity (Figures 4, 6 and 7)."""
lin = _read(os.path.join(art, "claim4", "figure4_llm_extended_ladder.csv"))
gra = _read(os.path.join(art, "claim6", "figures67_llm_extended_ladder.csv"))
fig, axes = plt.subplots(1, 2, figsize=(12.0, 4.3))
ax = axes[0]
x = [float(r["params_b"]) for r in lin]
y = [float(r["median_comp"]) for r in lin]
ax.semilogx(x, y, "o", color=PALETTE[0], markersize=7)
for r in lin:
ax.annotate(r["paper_name"], (float(r["params_b"]),
float(r["median_comp"])),
fontsize=6.5, color="#41464F",
xytext=(3, 4), textcoords="offset points")
ax.axhline(0, color="#B0543C", linewidth=1.2, linestyle="--")
if len(x) > 2:
b = np.polyfit(np.log(x), y, 1)
xs = np.geomspace(min(x), max(x), 40)
ax.plot(xs, np.polyval(b, np.log(xs)), "-", color="#8A8F98", linewidth=1.2)
_style(ax, "total parameters (B, log scale)",
"median compatibility score",
"Figure 4 — linear statements\n(above the dashed line = not falsified)")
ax = axes[1]
sub = [r for r in gra if r["mean_incomp_below_cap"] not in ("", "nan")]
x = [float(r["params_b"]) for r in sub]
y = [float(r["mean_incomp_below_cap"]) for r in sub]
ax.semilogx(x, y, "s", color=PALETTE[2], markersize=7)
for r in sub:
ax.annotate(r["paper_name"], (float(r["params_b"]),
float(r["mean_incomp_below_cap"])),
fontsize=6.5, color="#41464F",
xytext=(3, 4), textcoords="offset points")
if len(x) > 2:
b = np.polyfit(np.log(x), y, 1)
xs = np.geomspace(min(x), max(x), 40)
ax.plot(xs, np.polyval(b, np.log(xs)), "-", color="#8A8F98", linewidth=1.2)
_style(ax, "total parameters (B, log scale)",
"mean incompatibility $c(G)$",
"Figure 7 — graphical statements, density $\\leq 2/3$\n"
"(lower is better)")
return _save(fig, out, "llm_scores")
def lemma37(art: str, out: str):
"""How tight the heuristic is against the exact incompatibility score."""
rows = _read(os.path.join(art, "claim6", "lemma37_exhaustive.csv"))
fig, ax = plt.subplots(figsize=(8.0, 4.0))
labels = [f"n={r['n']}\n{int(r['n_graphs']):,} graphs" for r in rows]
xs = np.arange(len(rows))
exact = [float(r["frac_exact"]) for r in rows]
ax.bar(xs, exact, 0.5, color=PALETTE[2], label="heuristic exactly equals incomp(G)")
ax.bar(xs, [1 - e for e in exact], 0.5, bottom=exact, color="#C9CDD3",
label="heuristic strictly over-estimates")
for i, r in enumerate(rows):
ax.text(i, 1.02, f"0 violations\nmean gap {float(r['mean_gap']):.2f}",
ha="center", fontsize=8, color="#41464F")
ax.set_xticks(xs)
ax.set_xticklabels(labels, fontsize=8.5)
ax.set_ylim(0, 1.22)
_style(ax, "", "fraction of statement graphs",
"Lemma 3.7 — c(G) never falls below incomp(G), and is zero exactly "
"when incomp(G) is")
ax.legend(frameon=False, fontsize=8.5, loc="lower right")
return _save(fig, out, "lemma37")
def derivation(art: str, out: str):
"""Theorem 2.10 reconstructed: sensitivity x concentration, and the rate.
Three panels, one per ingredient of the derivation. Together they are the
whole argument: (a) how far comp can move per unit error in Sigma, (b) how
fast that error shrinks with N, (c) how fast the failure probability decays
-- which is what the theorem's log(n/delta) factor encodes.
"""
sens = _read(os.path.join(art, "claim3", "sensitivity.csv"))
conc = _read(os.path.join(art, "claim3", "concentration.csv"))
rate = _read(os.path.join(art, "claim3", "deviation_rate.csv"))
fig, axes = plt.subplots(1, 3, figsize=(15.5, 4.4))
# (a) sensitivity L against each factor, with the derivation's prediction.
ax = axes[0]
# The "n" curve uses the deconfounded sweep, which shrinks the coefficient
# scale as n grows so that (1 + a + b) stays put; the plain n sweep moves
# both factors at once and its marginal slope is not an exponent in n.
for c, (sweep, key, lab, pred) in enumerate(
(("n_fixed_ab", "n", "n (with 1 + a + b held fixed)", 2.0),
("ab", "ab", "1 + a + b", 2.0), ("V", "V", "V", 1.0))):
rs = [r for r in sens if r["sweep"] == sweep]
vals = sorted({float(r["value"]) for r in rs})
xs = [float(np.median([float(r[key]) for r in rs
if float(r["value"]) == v])) for v in vals]
ys = [float(np.median([float(r["L"]) for r in rs
if float(r["value"]) == v])) for v in vals]
ax.plot(xs, ys, "o-", color=PALETTE[c], lw=1.8, ms=5,
label=f"{lab}\n measured {np.polyfit(np.log(xs), np.log(ys), 1)[0]:+.2f}, "
f"derivation predicts {pred:g}")
ax.set_xscale("log")
ax.set_yscale("log")
_style(ax, "factor value (log)", "sensitivity L = ||grad comp||₁ (log)",
"(a) how far comp moves per unit error in Σ")
ax.legend(fontsize=8, frameon=False)
# (b) t*sqrt(N) flat across N <=> the error falls exactly as 1/sqrt(N).
ax = axes[1]
deltas = sorted({float(r["delta"]) for r in conc}, reverse=True)
for c, d in enumerate(deltas):
rs = sorted((r for r in conc if float(r["delta"]) == d),
key=lambda r: int(r["N"]))
ax.plot([int(r["N"]) for r in rs], [float(r["t_sqrtN"]) for r in rs],
"o-", color=PALETTE[c % len(PALETTE)], lw=1.6, ms=5,
label=f"δ = {d:g}")
ax.set_xscale("log")
_style(ax, "samples N (log)", "t(N, δ) · √N",
"(b) horizontal ⇒ error ∝ 1/√N ⇒ exponent 2 in 1/ε")
ax.legend(fontsize=8, frameon=False, ncol=2)
# (c) -log P linear in N <=> N*(delta) affine in log(1/delta).
ax = axes[2]
for law, colour, style in (("gaussian", PALETTE[0], "o-"),
("student_t3", PALETTE[5], "s--")):
models = sorted({r["model"] for r in rate if r["law"] == law})
for mi in models:
rs = sorted((r for r in rate
if r["law"] == law and r["model"] == mi),
key=lambda r: int(r["N"]))
xs = [int(r["N"]) for r in rs if float(r["p"]) > 0]
ys = [-np.log(float(r["p"])) for r in rs if float(r["p"]) > 0]
ax.plot(xs, ys, style, color=colour, lw=1.5, ms=3.5, alpha=0.85,
label=("Gaussian (theorem's hypothesis)"
if law == "gaussian" else
"t(3) — no exponential moment [control]")
if mi == models[0] else None)
_style(ax, "samples N", "−log P(|comp̂ − comp| > ε)",
"(c) straight ⇒ N*(δ) affine in log(1/δ) ⇒ exponent 1")
ax.legend(fontsize=8, frameon=False, loc="upper left")
fig.suptitle("Theorem 2.10, reconstructed from its two ingredients — "
"the theorem's formula is never used", fontsize=12,
color="#14171A", y=1.03)
fig.tight_layout()
return _save(fig, out, "derivation")
ALL = [figure2, derivation, figure5, sample_complexity, expected_compatibility,
llm_scores, lemma37]
def build(art: str, out: str) -> list[str]:
made = []
for fn in ALL:
try:
svg, png = fn(art, out)
made.append(os.path.basename(png))
print(f" figure {os.path.basename(png)}")
except Exception as exc: # a missing artifact
print(f" SKIPPED {fn.__name__}: {exc!r}")
return made
if __name__ == "__main__":
import sys
build(sys.argv[1], sys.argv[2])