Spaces:
Running
Running
File size: 17,474 Bytes
5338e3e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 | """Direct numerical contract for the resolution limit in Theorem 4.2.
This module uses only deterministic quadrature. The discrete measures are
positive midpoint rules that weakly converge to a positive continuous measure
on [0, 1]. The continuous reference is computed independently with
Gauss--Legendre quadrature at two orders.
"""
from __future__ import annotations
import csv
import hashlib
import json
import os
import platform
import subprocess
import sys
import time
from pathlib import Path
import numpy as np
from numpy.polynomial.legendre import leggauss
from scipy.spatial.distance import cdist
ROOT = Path(__file__).resolve().parents[2]
ARTIFACT_DIR = ROOT / ".openresearch" / "artifacts" / "claim_2"
FIXED_COMMAND = "uv run python repro/src/verify.py"
def _kernel(left: np.ndarray, right: np.ndarray, name: str, sigma: float) -> np.ndarray:
distances = cdist(left[:, None], right[:, None], metric="euclidean")
if name == "gaussian":
return np.exp(-(distances**2) / (2.0 * sigma**2))
if name == "exponential":
return np.exp(-distances / sigma)
raise ValueError(f"unsupported kernel: {name}")
def _density(points: np.ndarray, name: str) -> np.ndarray:
if name == "uniform":
return np.ones_like(points)
if name == "affine":
# Strictly positive on the closed domain and integrates to one.
return 0.5 + points
if name == "oscillatory":
# Strict lower bound 0.55; integral is one.
return 1.0 + 0.45 * np.cos(2.0 * np.pi * points)
raise ValueError(f"unsupported density: {name}")
def _signal(points: np.ndarray, name: str) -> np.ndarray:
if name == "constant":
return np.ones_like(points)
if name == "linear":
return points
if name == "quadratic":
return points**2
if name == "sin_2pi":
return np.sin(2.0 * np.pi * points)
if name == "cos_3pi":
return np.cos(3.0 * np.pi * points)
raise ValueError(f"unsupported signal: {name}")
def _sinkhorn_scaling(
points: np.ndarray,
weights: np.ndarray,
kernel_name: str,
sigma: float,
tolerance: float = 2e-14,
max_iterations: int = 2_000,
) -> tuple[np.ndarray, int, float]:
kernel = _kernel(points, points, kernel_name, sigma)
scaling = np.ones(points.size, dtype=np.float64)
residual = float("inf")
for iteration in range(1, max_iterations + 1):
denominator = kernel @ (weights * scaling)
scaling = np.sqrt(scaling / np.maximum(denominator, 1e-300))
residual = float(
np.max(np.abs(scaling * (kernel @ (weights * scaling)) - 1.0))
)
if residual < tolerance:
return scaling, iteration, residual
raise RuntimeError(
f"Sinkhorn failed: kernel={kernel_name} n={points.size} residual={residual}"
)
def _operator_on_probes(
nodes: np.ndarray,
weights: np.ndarray,
scaling: np.ndarray,
probes: np.ndarray,
signal_name: str,
kernel_name: str,
sigma: float,
) -> tuple[np.ndarray, float]:
cross_kernel = _kernel(probes, nodes, kernel_name, sigma)
probe_denominator = cross_kernel @ (weights * scaling)
probe_scaling = 1.0 / np.maximum(probe_denominator, 1e-300)
values = probe_scaling * (
cross_kernel @ (weights * scaling * _signal(nodes, signal_name))
)
constant_residual = float(
np.max(
np.abs(
probe_scaling * (cross_kernel @ (weights * scaling))
- np.ones(probes.size)
)
)
)
return values, constant_residual
def _midpoint_measure(n: int, density_name: str) -> tuple[np.ndarray, np.ndarray]:
nodes = (np.arange(n, dtype=np.float64) + 0.5) / n
weights = _density(nodes, density_name) / n
# Normalization removes only finite quadrature error, preserves positivity,
# and does not affect weak convergence.
weights /= weights.sum()
return nodes, weights
def _quantile_measure(n: int, density_name: str) -> tuple[np.ndarray, np.ndarray]:
probabilities = (np.arange(n, dtype=np.float64) + 0.5) / n
if density_name == "uniform":
nodes = probabilities
elif density_name == "affine":
# Invert F(x) = (x + x^2) / 2.
nodes = 0.5 * (-1.0 + np.sqrt(1.0 + 8.0 * probabilities))
elif density_name == "oscillatory":
# Invert F(x) = x + 0.45 sin(2 pi x)/(2 pi) by bisection.
lower = np.zeros(n, dtype=np.float64)
upper = np.ones(n, dtype=np.float64)
for _ in range(60):
middle = 0.5 * (lower + upper)
cdf = middle + 0.45 * np.sin(2.0 * np.pi * middle) / (
2.0 * np.pi
)
lower = np.where(cdf < probabilities, middle, lower)
upper = np.where(cdf >= probabilities, middle, upper)
nodes = 0.5 * (lower + upper)
else:
raise ValueError(f"unsupported density: {density_name}")
return nodes, np.full(n, 1.0 / n, dtype=np.float64)
def _discrete_measure(
n: int, density_name: str, discretization: str
) -> tuple[np.ndarray, np.ndarray]:
if discretization == "midpoint":
return _midpoint_measure(n, density_name)
if discretization == "quantile":
return _quantile_measure(n, density_name)
raise ValueError(f"unsupported discretization: {discretization}")
def _gauss_measure(order: int, density_name: str) -> tuple[np.ndarray, np.ndarray]:
canonical_nodes, canonical_weights = leggauss(order)
nodes = 0.5 * (canonical_nodes + 1.0)
weights = 0.5 * canonical_weights * _density(nodes, density_name)
weights /= weights.sum()
return nodes, weights
def _fit_slope(resolutions: list[int], errors: list[float]) -> float:
safe = np.maximum(np.asarray(errors, dtype=np.float64), 1e-18)
return float(np.polyfit(np.log(np.asarray(resolutions)), np.log(safe), 1)[0])
def _reference_outputs(
order: int,
density_name: str,
kernel_name: str,
sigma: float,
signals: list[str],
probes: np.ndarray,
) -> tuple[dict[str, np.ndarray], dict[str, float], int, float]:
nodes, weights = _gauss_measure(order, density_name)
scaling, iterations, residual = _sinkhorn_scaling(
nodes, weights, kernel_name, sigma
)
values: dict[str, np.ndarray] = {}
constant_residuals: dict[str, float] = {}
for signal_name in signals:
values[signal_name], constant_residuals[signal_name] = _operator_on_probes(
nodes,
weights,
scaling,
probes,
signal_name,
kernel_name,
sigma,
)
return values, constant_residuals, iterations, residual
def _git_sha() -> str:
result = subprocess.run(
["git", "rev-parse", "HEAD"],
cwd=ROOT,
check=True,
capture_output=True,
text=True,
)
return result.stdout.strip()
def _sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for block in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(block)
return digest.hexdigest()
def run_resolution_contract(config_path: Path | None = None) -> dict:
start = time.perf_counter()
if config_path is None:
config_path = ROOT / "repro" / "config.json"
config = json.loads(config_path.read_text(encoding="utf-8"))
probes = np.linspace(0.0, 1.0, int(config["probe_count"]))
resolutions = [int(value) for value in config["resolutions"]]
reference_orders = [int(value) for value in config["reference_orders"]]
signals = [str(value) for value in config["signals"]]
cases: list[dict] = []
for density_name in config["densities"]:
for discretization in config["discretizations"]:
for kernel_spec in config["kernels"]:
kernel_name = str(kernel_spec["name"])
sigma = float(kernel_spec["sigma"])
coarse_reference, coarse_constant, coarse_iterations, coarse_residual = (
_reference_outputs(
reference_orders[0],
density_name,
kernel_name,
sigma,
signals,
probes,
)
)
fine_reference, fine_constant, fine_iterations, fine_residual = (
_reference_outputs(
reference_orders[1],
density_name,
kernel_name,
sigma,
signals,
probes,
)
)
reference_stability = {
signal_name: float(
np.max(
np.abs(
coarse_reference[signal_name]
- fine_reference[signal_name]
)
)
)
for signal_name in signals
}
error_curves = {signal_name: [] for signal_name in signals}
constant_residuals: list[float] = []
scaling_records: list[dict] = []
for n in resolutions:
nodes, weights = _discrete_measure(
n, density_name, discretization
)
scaling, iterations, residual = _sinkhorn_scaling(
nodes, weights, kernel_name, sigma
)
scaling_records.append(
{
"n": n,
"iterations": iterations,
"node_residual": residual,
"min_weight": float(weights.min()),
"min_scaling": float(scaling.min()),
}
)
for signal_name in signals:
values, constant_residual = _operator_on_probes(
nodes,
weights,
scaling,
probes,
signal_name,
kernel_name,
sigma,
)
error_curves[signal_name].append(
float(
np.max(
np.abs(
values - fine_reference[signal_name]
)
)
)
)
constant_residuals.append(constant_residual)
signal_records = []
for signal_name in signals:
errors = error_curves[signal_name]
signal_records.append(
{
"signal": signal_name,
"uniform_errors": errors,
"final_error": errors[-1],
"reduction_ratio": errors[-1]
/ max(errors[0], 1e-300),
"loglog_slope": _fit_slope(resolutions, errors),
"reference_stability": reference_stability[
signal_name
],
}
)
cases.append(
{
"density": density_name,
"discretization": discretization,
"kernel": kernel_name,
"sigma": sigma,
"reference": {
"orders": reference_orders,
"coarse_iterations": coarse_iterations,
"fine_iterations": fine_iterations,
"coarse_node_residual": coarse_residual,
"fine_node_residual": fine_residual,
"max_constant_residual": max(
list(coarse_constant.values())
+ list(fine_constant.values())
),
},
"scalings": scaling_records,
"max_probe_constant_residual": max(
constant_residuals
),
"signals": signal_records,
}
)
# This deliberately violates resolution increase: the same n=32 result is
# repeated at every nominal t. It must not satisfy the convergence gates.
negative_cases = []
for case in cases:
negative_signals = []
for record in case["signals"]:
repeated = [record["uniform_errors"][0]] * len(resolutions)
negative_signals.append(
{
"signal": record["signal"],
"uniform_errors": repeated,
"final_error": repeated[-1],
"reduction_ratio": 1.0,
"loglog_slope": _fit_slope(resolutions, repeated),
"reference_stability": record["reference_stability"],
}
)
negative_cases.append(
{
"density": case["density"],
"discretization": case["discretization"],
"kernel": case["kernel"],
"sigma": case["sigma"],
"reference": case["reference"],
"max_probe_constant_residual": case[
"max_probe_constant_residual"
],
"signals": negative_signals,
}
)
result = {
"schema_version": 1,
"claim": "Theorem 4.2 resolution convergence",
"variant": config["resolution_variant"],
"domain": "[0,1]",
"resolutions": resolutions,
"probe_count": int(config["probe_count"]),
"fixed_command": FIXED_COMMAND,
"seed": int(config["seed"]),
"cases": cases,
"negative_control": {
"name": "fixed_resolution_relabelled_as_increasing",
"expected_to_pass": False,
"cases": negative_cases,
},
"runtime_seconds": time.perf_counter() - start,
}
return result
def write_artifacts(result: dict) -> None:
ARTIFACT_DIR.mkdir(parents=True, exist_ok=True)
raw_json = ARTIFACT_DIR / "raw_results.json"
raw_json.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
with (ARTIFACT_DIR / "raw_results.csv").open(
"w", newline="", encoding="utf-8"
) as handle:
writer = csv.DictWriter(
handle,
fieldnames=[
"variant",
"density",
"discretization",
"kernel",
"sigma",
"signal",
"n",
"uniform_error",
],
)
writer.writeheader()
for case in result["cases"]:
for record in case["signals"]:
for n, error in zip(result["resolutions"], record["uniform_errors"]):
writer.writerow(
{
"variant": result["variant"],
"density": case["density"],
"discretization": case["discretization"],
"kernel": case["kernel"],
"sigma": case["sigma"],
"signal": record["signal"],
"n": n,
"uniform_error": f"{error:.17g}",
}
)
environment = {
"git_sha": _git_sha(),
"fixed_command": FIXED_COMMAND,
"python": sys.version,
"platform": platform.platform(),
"processor": platform.processor(),
"logical_cpu_count": os.cpu_count(),
"numpy": np.__version__,
"config_sha256": _sha256(ROOT / "repro" / "config.json"),
"lock_sha256": _sha256(ROOT / "uv.lock"),
"seed": result["seed"],
"runtime_seconds": result["runtime_seconds"],
}
(ARTIFACT_DIR / "environment.json").write_text(
json.dumps(environment, indent=2) + "\n", encoding="utf-8"
)
def print_summary(result: dict) -> None:
print("C2_RESOLUTION_VARIANT=" + str(result["variant"]))
print("C2_FIXED_COMMAND=" + FIXED_COMMAND)
print("C2_RESOLUTIONS=" + json.dumps(result["resolutions"]))
for case in result["cases"]:
for record in case["signals"]:
print(
"C2_RESULT "
f"density={case['density']} discretization={case['discretization']} "
f"kernel={case['kernel']} "
f"sigma={case['sigma']:.6g} signal={record['signal']} "
f"errors={json.dumps(record['uniform_errors'])} "
f"ratio={record['reduction_ratio']:.8g} "
f"slope={record['loglog_slope']:.8g} "
f"ref_stability={record['reference_stability']:.8g}"
)
print(f"C2_RUNTIME_SECONDS={result['runtime_seconds']:.6f}")
|