Datasets:
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
mathematics
optimization
gradient-descent
convex-optimization
learning-theory
recursive-self-improvement
File size: 13,636 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 | #!/usr/bin/env python3
"""Auditable comparison gate and resource ledger from Chapter 9.
All externally supplied numbers must be integers, decimal/rational strings,
or fractions.Fraction values. Float inputs are deliberately rejected.
The statistical radius uses a rational upper bound for log followed by
an upward rational-grid square root, so rounding cannot make it smaller
than the manuscript's Hoeffding radius.
This module does not establish that an evaluation data source is fresh,
that its conditional mean is the declared risk difference, or that a
claimed numerical-error bound is true. Those are the explicit modeling
obligations in the manuscript. It also does not generate candidates or
claim an empirical RSI speedup.
"""
from __future__ import annotations
import argparse
import json
from dataclasses import dataclass
from fractions import Fraction
from math import isqrt
from pathlib import Path
from typing import Iterable, Optional
F = Fraction
def exact(value) -> F:
if isinstance(value, float):
raise TypeError("Use an exact decimal string or Fraction, not a float.")
if isinstance(value, (int, str, F)):
return F(value)
raise TypeError("Expected int, str, or fractions.Fraction.")
def log_series_interval(w: F, terms: int = 24):
"""Bounds for log((1+w)/(1-w)), for 0<=w<=1/3.
log((1+w)/(1-w)) = 2 sum_{j>=0} w^(2j+1)/(2j+1).
Its positive omitted tail is at most
2*w^(2N+1)/((2N+1)*(1-w*w)).
"""
w = exact(w)
if not 0 <= w <= F(1, 3) or terms < 1:
raise ValueError("Require 0<=w<=1/3 and terms>=1.")
total = F(0)
power = w
for j in range(terms):
total += 2 * power / (2 * j + 1)
power *= w * w
tail = 2 * power / ((2 * terms + 1) * (1 - w * w))
return total, total + tail
def rational_log_upper(x, terms: int = 24) -> F:
"""A rigorous rational upper bound for log(x), x>=1."""
x = exact(x)
if x < 1:
raise ValueError("This function is defined here only for x>=1.")
k = 0
y = x
while y >= 2:
y /= 2
k += 1
_, log_two_upper = log_series_interval(F(1, 3), terms)
_, log_y_upper = log_series_interval((y - 1) / (y + 1), terms)
return k * log_two_upper + log_y_upper
def rational_sqrt_upper(x, scale: int = 10**12) -> F:
"""Least multiple of 1/scale whose square is at least exact x."""
x = exact(x)
if x < 0 or scale < 1:
raise ValueError("Require x>=0 and a positive integer grid scale.")
squared_grid = x * scale * scale
root = isqrt(squared_grid.numerator // squared_grid.denominator)
if root * root * squared_grid.denominator < squared_grid.numerator:
root += 1
result = F(root, scale)
assert result * result >= x
if root:
assert F(root - 1, scale) ** 2 < x
return result
def anytime_radius(trial: int, n: int, delta, *, terms: int = 24,
scale: int = 10**12) -> F:
"""Rational upper bound for sqrt(2/n log(2t(t+1)n(n+1)/delta))."""
delta = exact(delta)
if trial < 1 or n < 1 or not 0 < delta < 1:
raise ValueError("Require trial,n>=1 and 0<delta<1.")
log_argument = F(2 * trial * (trial + 1) * n * (n + 1)) / delta
log_bound = rational_log_upper(log_argument, terms)
return rational_sqrt_upper(2 * log_bound / n, scale)
@dataclass(frozen=True)
class ComparisonCertificate:
trial: int
n: int
mean: F
statistical_radius: F
measurement_radius: F
delta: F
@property
def radius(self):
return self.statistical_radius + self.measurement_radius
@property
def lower(self):
return self.mean - self.radius
@property
def upper(self):
return self.mean + self.radius
def to_dict(self):
return {"trial": self.trial, "n": self.n,
"mean": str(self.mean), "statistical_radius": str(self.statistical_radius),
"measurement_radius": str(self.measurement_radius),
"lower": str(self.lower), "upper": str(self.upper),
"whole_run_delta": str(self.delta)}
def compare(values: Iterable, *, trial: int, delta,
measurement_errors: Optional[Iterable] = None) -> ComparisonCertificate:
"""Certify a single frozen paired comparison at any collected sample size.
values[j] is the recorded incumbent-minus-candidate loss on paired task j.
measurement_errors[j], if supplied, is a true bound on its recording error.
Repeated calls may use longer prefixes of the same frozen comparison.
Use the same whole-run delta and a unique positive trial index per proposal.
"""
values = [exact(x) for x in values]
if not values:
raise ValueError("At least one paired observation is required.")
if measurement_errors is None:
errors = [F(0)] * len(values)
else:
errors = [exact(x) for x in measurement_errors]
if len(errors) != len(values):
raise ValueError("Each observation needs its own error bound.")
for value, error in zip(values, errors):
if error < 0 or abs(value) > 1 + error:
raise ValueError("Error interval must be compatible with a true value in [-1,1].")
n = len(values)
return ComparisonCertificate(trial, n, sum(values, F(0)) / n,
anytime_radius(trial, n, delta),
sum(errors, F(0)) / n, exact(delta))
def should_accept(certificate: ComparisonCertificate, margin=0) -> bool:
margin = exact(margin)
if margin < 0:
raise ValueError("The required improvement must be nonnegative.")
return certificate.lower >= margin
def comparison_action(certificate: ComparisonCertificate, margin=0,
*, force_decision: bool = False) -> str:
"""Return accept, reject, or continue for one frozen comparison.
A resource cap can force a rejection when acceptance is not certified.
This still obeys the manuscript's finite oracle inequality, with the
confidence radius at the actual stopping sample size.
"""
margin = exact(margin)
if margin < 0:
raise ValueError("The required improvement must be nonnegative.")
if certificate.lower >= margin:
return "accept"
if certificate.upper < margin or force_decision:
return "reject"
return "continue"
class CreditLedger:
"""Exact cost reservation and settlement; rejected trials are charged.
Credit is in risk-improvement units, cost_weight converts cost units.
A caller must reserve before undertaking a trial and honor the cap.
An accepted gain must be a valid nonnegative lower certificate.
The ledger verifies arithmetic and caps, not the certificate's provenance.
"""
def __init__(self, initial_credit, cost_weight):
self.initial_credit = exact(initial_credit)
self.cost_weight = exact(cost_weight)
if self.initial_credit < 0 or self.cost_weight <= 0:
raise ValueError("Require nonnegative credit and positive cost weight.")
self.credit = self.initial_credit
self.total_cost = F(0)
self.total_certified_gain = F(0)
self.completed_trials = 0
self._reserved_cap = None
def _check_accounting(self):
escrow_cost = self._reserved_cap if self._reserved_cap is not None else F(0)
assert self.credit >= 0
assert self.credit + self.cost_weight * escrow_cost == (
self.initial_credit - self.cost_weight * self.total_cost
+ self.total_certified_gain)
def reserve(self, cost_cap):
cap = exact(cost_cap)
if self._reserved_cap is not None:
raise RuntimeError("Settle the active reservation before starting another.")
if cap < 0 or self.cost_weight * cap > self.credit:
raise ValueError("The full cost cap must fit the available credit.")
self.credit -= self.cost_weight * cap
self._reserved_cap = cap
self._check_accounting()
def settle(self, actual_cost, *, accepted=False, certified_lower_gain=0):
cost = exact(actual_cost)
gain = exact(certified_lower_gain) if accepted else F(0)
if self._reserved_cap is None:
raise RuntimeError("Reserve a cost cap before settling a trial.")
if cost < 0 or cost > self._reserved_cap:
raise ValueError("Actual cost must stay within the reserved cap.")
if accepted and not 0 <= gain <= 1:
raise ValueError("Accepted risk-improvement certificate must lie in [0,1].")
self.credit += self.cost_weight * (self._reserved_cap - cost) + gain
self.total_cost += cost
self.total_certified_gain += gain
self.completed_trials += 1
self._reserved_cap = None
self._check_accounting()
def to_dict(self):
return {"initial_credit": str(self.initial_credit),
"cost_weight": str(self.cost_weight),
"available_credit": str(self.credit),
"reserved_cost_cap": (str(self._reserved_cap)
if self._reserved_cap is not None else None),
"total_cost": str(self.total_cost),
"total_certified_gain": str(self.total_certified_gain),
"completed_trials": self.completed_trials}
def self_test():
assertions = 0
for w in [F(0), F(1, 10), F(1, 3)]:
lo, hi = log_series_interval(w, 24)
lo_more, hi_more = log_series_interval(w, 25)
assert lo <= lo_more <= hi_more <= hi
assertions += 1
lo2, hi2 = log_series_interval(F(1, 3), 24)
assert F(2, 3) < lo2 < hi2 < F(7, 10)
assert rational_log_upper(F(1)) == 0
assert rational_log_upper(F(8)) == 3 * hi2
assertions += 3
for x in [F(0), F(1), F(2), F(7, 9), F(137, 17)]:
b = rational_sqrt_upper(x, scale=10**6)
assert b * b >= x
if b:
assert (b - F(1, 10**6)) ** 2 < x
assertions += 1
for T in [1, 2, 7, 50]:
total = sum((F(1, t * (t + 1)) for t in range(1, T + 1)), F(0))
assert total == F(T, T + 1) < 1
assertions += 1
cert = compare(["3/4"] * 1000, trial=1, delta="1/20")
assert should_accept(cert, "1/2")
cert_error = compare(["3/4"] * 1000, trial=1, delta="1/20",
measurement_errors=["1/100"] * 1000)
assert cert_error.lower == cert.lower - F(1, 100)
assert cert_error.upper == cert.upper + F(1, 100)
negative = compare(["-1/2"] * 1000, trial=2, delta="1/20")
assert not should_accept(negative)
assertions += 4
ambiguous = compare([0] * 1000, trial=3, delta="1/20")
assert comparison_action(cert, "1/2") == "accept"
assert comparison_action(negative) == "reject"
assert comparison_action(ambiguous) == "continue"
assert comparison_action(ambiguous, force_decision=True) == "reject"
assertions += 4
ledger = CreditLedger("1/10", "1/1000")
ledger.reserve(50)
ledger.settle(40, accepted=False)
assert ledger.credit == F(3, 50)
ledger.reserve(50)
ledger.settle(50, accepted=True, certified_lower_gain=cert.lower)
assert ledger.credit == F(1, 100) + cert.lower
ledger.reserve(60)
ledger.settle(30, accepted=False)
assert ledger.total_cost == 120
assert ledger.cost_weight * ledger.total_cost <= (
ledger.initial_credit + ledger.total_certified_gain)
assertions += 4
# Pathwise oracle inequality under declared valid intervals.
incumbent = F(9, 10)
initial = incumbent
candidates = [F(7, 10), F(4, 5), F(1, 5), F(1, 4)]
radius, margin = F(1, 20), F(1, 50)
offsets = [-radius / 2, radius / 2, -radius / 2, radius / 2]
accepted_margins = F(0)
for candidate, offset in zip(candidates, offsets):
delta = incumbent - candidate
lower = delta + offset - radius
upper = delta + offset + radius
assert lower <= delta <= upper
if lower >= margin:
incumbent = candidate
accepted_margins += margin
assert incumbent <= initial - accepted_margins
assertions += 2
assert incumbent == F(1, 5)
assert incumbent <= min(candidates) + margin + 2 * radius
assertions += 2
# API guard checks must fail before any incorrect ledger mutation.
try:
exact(0.1)
raise AssertionError("Float guard did not trigger")
except TypeError:
assertions += 1
before = ledger.to_dict()
try:
ledger.reserve(10**10)
raise AssertionError("Unaffordable reservation was accepted")
except ValueError:
assert ledger.to_dict() == before
assertions += 1
return {"all_checks_passed": True, "reported_assertion_groups": assertions,
"arithmetic": "exact rational log-series bound and upward square-root grid",
"acceptance_example": cert.to_dict(),
"ledger_example": ledger.to_dict(),
"scope": "Arithmetic and accounting audits; probability and data contracts are proved/stated in Chapter 9."}
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--self-test", action="store_true")
parser.add_argument("--output", type=Path)
args = parser.parse_args()
if not args.self_test:
parser.print_help()
return
report = self_test()
if args.output:
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|