Datasets:
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
mathematics
optimization
gradient-descent
convex-optimization
learning-theory
recursive-self-improvement
File size: 7,843 Bytes
868c741 | 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 | #!/usr/bin/env python3
"""Exact audits for the information-limit constructions in Chapter 8.
Only Python's standard library is required. These finite audits check
algebraic implementations and stated examples. The universal claims and
analytic Gaussian envelope are proved in the manuscript; finite checks
are not a replacement for those proofs.
"""
from __future__ import annotations
import argparse
import json
from fractions import Fraction as F
from math import factorial
from pathlib import Path
def trim(p):
p = list(p)
while len(p) > 1 and p[-1] == 0:
p.pop()
return p
def add(p, q):
out = [F(0)] * max(len(p), len(q))
for i, c in enumerate(p):
out[i] += c
for i, c in enumerate(q):
out[i] += c
return trim(out)
def scale(p, c):
return trim([c * x for x in p])
def shift(p, n):
return [F(0)] * n + list(p)
def mul(p, q):
out = [F(0)] * (len(p) + len(q) - 1)
for i, a in enumerate(p):
for j, b in enumerate(q):
out[i + j] += a * b
return trim(out)
def derivative(p):
return trim([F(i) * p[i] for i in range(1, len(p))] or [F(0)])
def evaluate(p, x):
value = F(0)
for c in reversed(p):
value = value * x + c
return value
def gaussian_derivative_polynomial(p):
"""Return Q such that (exp(-x^2) P)' = exp(-x^2) Q."""
return add(derivative(p), scale(shift(p, 1), -2))
def piecewise_value(x, a, mu, v, offset=F(0)):
curvature = a if x <= 0 else mu
return offset + v * x + curvature * x * x / 2
def piecewise_gradient(x, a, mu, v):
return v + (a if x <= 0 else mu) * x
def audit_transcripts():
cases = []
total_queried_points = 0
total_future_values = 0
for mu, L, eta, v in [
(F(1, 16), F(1), F(1), F(1, 4)),
(F(1, 8), F(1), F(3, 2), F(2, 3)),
(F(1, 5), F(3, 2), F(1), F(3, 7)),
]:
assert 0 < eta * mu < 1 and 0 < eta * L < 2
for t in [0, 1, 2, 5, 9, 17]:
q = 1 - eta * mu
points = [v / mu * (q ** (i - t) - 1) for i in range(t + 1)]
assert points[-1] == 0 and all(x >= 0 for x in points)
transcripts = []
for a in [mu, (mu + L) / 2, L]:
data = [(x, piecewise_value(x, a, mu, v),
piecewise_gradient(x, a, mu, v)) for x in points]
transcripts.append(data)
for i in range(t):
assert data[i][0] - eta * data[i][2] == data[i + 1][0]
optimum = -v / a
assert piecewise_gradient(optimum, a, mu, v) == 0
remaining = piecewise_value(F(0), a, mu, v) - piecewise_value(
optimum, a, mu, v)
assert remaining == v * v / (2 * a)
x = F(0)
for n in range(1, 15):
x -= eta * piecewise_gradient(x, a, mu, v)
closed = v / a * ((1 - eta * a) ** n - 1)
assert x == closed and x <= 0
gain = piecewise_value(F(0), a, mu, v) - piecewise_value(
x, a, mu, v)
assert gain == v * v / (2 * a) * (1 - (1 - eta * a) ** (2 * n))
total_future_values += 1
assert transcripts[0] == transcripts[1] == transcripts[2]
for n in range(1, 15):
gains = [v * v / (2 * a) * (1 - (1 - eta * a) ** (2 * n))
for a in [mu, (mu + L) / 2, L]]
assert gains[0] > gains[1] > gains[2]
total_queried_points += t + 1
cases.append({"mu": str(mu), "L": str(L), "eta": str(eta),
"v": str(v), "t": t, "all_exact_checks_passed": True})
return {"cases": cases, "distinct_common_prefix_points": total_queried_points,
"exact_future_values_checked": total_future_values}
def audit_analytic_jets():
reports = []
jet_equalities = 0
mu, L, a, eta, x0 = F(1, 4), F(1), F(1, 2), F(1), F(1)
q = 1 - eta * a
for t in range(6):
for r in range(1, 4):
points = [q ** i * x0 for i in range(t + 1)]
p = [F(1)]
for x in points:
for _ in range(2 * r + 2):
p = mul(p, [-x, F(1)])
first = gaussian_derivative_polynomial(p)
second = gaussian_derivative_polynomial(first)
direct = add(add(derivative(derivative(p)),
scale(shift(derivative(p), 1), -4)),
add(scale(shift(p, 2), 4), scale(p, -2)))
assert second == direct
M = sum((abs(c) * factorial((j + 1) // 2)
for j, c in enumerate(second)), F(0))
assert M > 0
epsilon = min(a - mu, L - a) / (2 * M)
assert mu < a - epsilon * M < a + epsilon * M < L
d = p
for order in range(2 * r + 2):
for x in points:
assert evaluate(d, x) == 0
jet_equalities += 1
d = gaussian_derivative_polynomial(d)
next_x = q ** (t + 1) * x0
assert next_x not in points
assert evaluate(p, next_x) > 0
# exp(-next_x**2)>0; the rational prefactor below is positive.
separation_prefactor = 2 * epsilon * evaluate(p, next_x)
assert separation_prefactor > 0
reports.append({"t": t, "r": r, "polynomial_degree": len(p) - 1,
"M_numerator_bits": M.numerator.bit_length(),
"M_denominator_bits": M.denominator.bit_length(),
"epsilon_positive": epsilon > 0,
"next_loss_separation_positive": separation_prefactor > 0,
"all_exact_checks_passed": True})
return {"cases": reports, "exact_zero_jet_equalities_checked": jet_equalities,
"envelope": "M=sum_j abs(q_j)*ceil(j/2)!; global bound proved in text"}
def audit_curvature_budget():
checked = 0
for m in range(1, 65):
D0 = F(1, m + 3)
increments = [F(j + 1, (m + 3) ** 3) for j in range(m)]
decrements = [D0]
for increment in increments:
decrements.append(decrements[-1] + increment)
weighted = (m + 1) * D0 + sum(
((m - j) * value for j, value in enumerate(increments)), F(0))
assert sum(decrements, F(0)) == weighted
checked += 1
return {"weighted_curvature_identities_checked": checked}
def run_audit():
return {"arithmetic": "fractions.Fraction; exact integer polynomial operations",
"all_checks_passed": True,
"transcripts": audit_transcripts(),
"analytic_finite_jets": audit_analytic_jets(),
"curvature_budget": audit_curvature_budget(),
"scope": "Finite implementation audits; universal claims rest on manuscript proofs."}
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--output", type=Path)
args = parser.parse_args()
report = run_audit()
encoded = json.dumps(report, indent=2) + "\n"
if args.output:
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(encoded, encoding="utf-8")
print(json.dumps({"all_checks_passed": report["all_checks_passed"],
"transcript_cases": len(report["transcripts"]["cases"]),
"analytic_cases": len(report["analytic_finite_jets"]["cases"]),
"exact_future_values": report["transcripts"]["exact_future_values_checked"],
"exact_zero_jets": report["analytic_finite_jets"]["exact_zero_jet_equalities_checked"]},
indent=2))
if __name__ == "__main__":
main()
|