File size: 9,736 Bytes
ca4f4a8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""CPU-only scope expansion for the generalized-convex audit.

This file deliberately performs new work rather than copying the original
eight-point result.  The finite-transform, leanness, auction, and transport
checks are all evaluated again on larger or denser exact grids.  The script
writes only the requested JSON report; it does not modify the packaged
artifacts used by the original reproduction.
"""

from __future__ import annotations

import argparse
import itertools
import json
import re
import sys
from fractions import Fraction
from pathlib import Path


def finite_transform(x: Fraction, ys: list[Fraction]) -> Fraction:
    return max(x * y - y * y / 2 for y in ys)


def transform_scope() -> dict:
    rows = []
    for k in (128, 256, 512):
        h = Fraction(1, k)
        ys = [Fraction(j, k) for j in range(-k, k + 1)]
        errors = []
        gradient_errors = []
        gradient_checks = 0
        for left, right in zip(ys, ys[1:]):
            midpoint = (left + right) / 2
            x = midpoint
            errors.append(x * x / 2 - finite_transform(x, ys))
            for offset in (Fraction(-1, 1000), Fraction(-1, 4), Fraction(1, 4), Fraction(1, 1000)):
                sample = midpoint + offset * h
                values = [sample * y - y * y / 2 for y in ys]
                best = max(range(len(ys)), key=values.__getitem__)
                assert values.count(values[best]) == 1
                gradient_errors.append(abs(ys[best] - sample))
                gradient_checks += 1
        maximum = max(errors)
        observed_gradient = max(gradient_errors)
        assert min(errors) >= 0 and maximum == h * h / 8
        assert observed_gradient == Fraction(499, 1000) * h
        rows.append(
            {
                "k": k,
                "finite_atoms": len(ys),
                "cells_exhausted": len(errors),
                "exact_uniform_error": f"{maximum.numerator}/{maximum.denominator}",
                "uniform_error": float(maximum),
                "gradient_points_checked": gradient_checks,
                "exact_gradient_error": f"{observed_gradient.numerator}/{observed_gradient.denominator}",
                "gradient_error": float(observed_gradient),
            }
        )
    return {
        "new_k_values": [128, 256, 512],
        "rows": rows,
        "claim1_scope": "all cells of three grids through 1,025 finite atoms",
        "claim2_scope": "four non-tie points in every cell of the same three grids",
    }


def lean_scope() -> dict:
    ys = [Fraction(j, 4) for j in range(-8, 9)]
    a_values = [Fraction(1, 4), Fraction(1, 2), Fraction(1), Fraction(3, 2), Fraction(2), Fraction(4)]
    b_values = [Fraction(-2), Fraction(-1), Fraction(-1, 2), Fraction(0), Fraction(1, 2), Fraction(1), Fraction(2)]
    parameters = [(a, b, (a + b) / 17) for a in a_values for b in b_values]
    pair_count = 0
    combination_count = 0
    inequality_count = 0
    for p, q in itertools.combinations(parameters, 2):
        pair_count += 1
        for lam in (Fraction(1, 5), Fraction(2, 5), Fraction(3, 5), Fraction(4, 5)):
            a = (1 - lam) * p[0] + lam * q[0]
            b = (1 - lam) * p[1] + lam * q[1]
            c = (1 - lam) * p[2] + lam * q[2]
            assert a > 0
            for yi in ys:
                x = 2 * a * yi + b
                chosen = x * yi - (a * yi * yi + b * yi + c)
                for yj in ys:
                    competitor = x * yj - (a * yj * yj + b * yj + c)
                    assert chosen - competitor >= 0
                    inequality_count += 1
            combination_count += 1
    return {
        "a_values": [str(x) for x in a_values],
        "b_values": [str(x) for x in b_values],
        "y_grid": [str(x) for x in ys],
        "parameterizations": len(parameters),
        "unordered_pairs": pair_count,
        "convex_combinations_checked": combination_count,
        "exact_activation_inequalities": inequality_count,
    }


def transport_scope() -> dict:
    rows = []
    for n in (8, 9, 10):
        denominator = n * (n - 1) ** 2
        best_numerator = -1
        best_perm = None
        permutations_checked = 0
        for perm in itertools.permutations(range(n)):
            numerator = sum(i * perm[i] for i in range(n))
            if numerator > best_numerator:
                best_numerator = numerator
                best_perm = perm
            permutations_checked += 1
        identity = tuple(range(n))
        identity_numerator = sum(i * i for i in range(n))
        reverse = tuple(reversed(identity))
        reverse_numerator = sum(i * reverse[i] for i in range(n))
        dual_slacks = [(i - j) ** 2 for i in range(n) for j in range(n)]
        assert best_perm == identity
        assert best_numerator == identity_numerator
        assert min(dual_slacks) == 0
        rows.append(
            {
                "grid_points": n,
                "permutations_exhausted": permutations_checked,
                "optimal_permutation": list(best_perm),
                "primal_numerator": best_numerator,
                "dual_numerator": identity_numerator,
                "exact_primal_dual_gap": "0/1",
                "dual_constraints_checked": len(dual_slacks),
                "minimum_dual_slack_numerator": min(dual_slacks),
                "reverse_numerator": reverse_numerator,
                "reverse_gap_numerator": identity_numerator - reverse_numerator,
                "normalization_denominator": denominator,
            }
        )
    return {
        "surplus": "Phi(x,y)=xy",
        "grid_sizes": [8, 9, 10],
        "new_larger_grids": [9, 10],
        "rows": rows,
        "interpretation": "The identity map remains exactly optimal after exhaustive enumeration through 10 points; dual equality and zero minimum slack hold on every grid.",
    }


def auction_scope(author_root: Path) -> dict:
    sys.path.insert(0, str(author_root))
    import torch
    from mech_design.mechanism import Mechanism

    def dot_kernel(x, y):
        return (x * y).sum(dim=-1)

    model = Mechanism(
        npoints=1,
        kernel=dot_kernel,
        y_dim=1,
        temp=1.0,
        is_Y_parameter=False,
        is_there_default=True,
        y_min=0.0,
        y_max=1.0,
    )
    with torch.no_grad():
        model.Y_rest_raw.fill_(1.0)
        model.intercept_rest.fill_(0.5)
    rows = []
    with torch.no_grad():
        for points in (16001, 32001):
            xs = torch.linspace(0.0, 1.0, points)[:, None]
            choices, _ = model.forward(xs, selection_mode="hard")
            allocation = choices[:, 0]
            below = allocation[xs[:, 0] < 0.499]
            above = allocation[xs[:, 0] > 0.501]
            assert float(below.max()) == 0.0 and float(above.min()) == 1.0
            rows.append(
                {
                    "grid_points": points,
                    "max_allocation_below_0.499": float(below.max()),
                    "min_allocation_above_0.501": float(above.min()),
                    "below_points": int(below.numel()),
                    "above_points": int(above.numel()),
                }
            )
    return {
        "author_commit": "85a5da444a146ea28945173e22e3d130163dae42",
        "fresh_grid_rows": rows,
        "interpretation": "The released Mechanism preserves the exact threshold on two additional dense CPU grids, independently of the original 4,001-point run.",
    }


def table_scope(source_root: Path) -> dict:
    """Re-read every registered table row and recompute the printed gaps."""
    text = (source_root / "src/sections/_VII_experiments.tex").read_text(encoding="utf-8")
    pattern = re.compile(r"^(1|2|5|10|20)\s*&\s*([0-9.]+)\s*&\s*(---|[0-9.]+)", re.MULTILINE)
    rows = []
    for n, learned, benchmark in pattern.findall(text):
        gap = None if benchmark == "---" else abs(Fraction(learned) - Fraction(benchmark))
        rows.append(
            {
                "n": int(n),
                "learned": learned,
                "straight_jacket": benchmark,
                "exact_gap": None if gap is None else f"{gap.numerator}/{gap.denominator}",
                "within_printed_0.001": None if gap is None else gap <= Fraction(1, 1000),
            }
        )
    assert [row["n"] for row in rows] == [1, 2, 5, 10, 20]
    assert [row["within_printed_0.001"] for row in rows if row["within_printed_0.001"] is not None] == [True, True, True, True]
    return {
        "rows_reparsed": rows,
        "comparable_rows": 4,
        "exact_match_rows": [row["n"] for row in rows if row["exact_gap"] == "0/1"],
        "nonzero_exact_gap_rows": [row["n"] for row in rows if row["exact_gap"] not in (None, "0/1")],
        "scope": "all five registered n rows, with exact Fraction arithmetic and a separate 0.001 printed-precision check",
    }


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--author-root", type=Path, required=True)
    parser.add_argument("--source-root", type=Path, required=True)
    parser.add_argument("--output", type=Path, required=True)
    args = parser.parse_args()
    report = {
        "transform": transform_scope(),
        "lean": lean_scope(),
        "transport": transport_scope(),
        "auction": auction_scope(args.author_root),
        "table": table_scope(args.source_root),
    }
    args.output.write_text(json.dumps(report, indent=2, sort_keys=True) + "\n", encoding="utf-8")
    print(json.dumps({"transform_rows": len(report["transform"]["rows"]), "transport_rows": len(report["transport"]["rows"]), "auction_rows": len(report["auction"]["fresh_grid_rows"])}, sort_keys=True))


if __name__ == "__main__":
    main()