File size: 6,337 Bytes
b330840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Fresh CPU audit of the Semi-knockoffs Section 5.1 comparison.

This is an independent execution, not a read of the authors' CSV outputs.  It
uses the paper's adjacent-support design (Sigma_ij=0.6**abs(i-j), first
quarter of beta nonzero, gradient boosting) and a separately implemented HRT
holdout.  A second exact construction exercises the masked-correlation
five-permutation assertion.
"""
import json
import sys
from pathlib import Path

import numpy as np
from sklearn.base import clone
from sklearn.ensemble import GradientBoostingRegressor
from scipy.stats import wilcoxon

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from skocore import ar1_design


ALPHA = 0.05
OUT = {"alpha": ALPHA, "design": "independent_cpu_claim5_scope_audit"}


def source_design(n, p, seed):
    rng = np.random.default_rng(seed)
    # ar1_design is the exact Gaussian construction for Sigma_ij=rho^|i-j|.
    X = ar1_design(n, p, 0.6, rng)
    beta = np.zeros(p)
    beta[: p // 4] = rng.uniform(1.0, 2.0, size=p // 4)
    y = np.einsum("ij,j->i", X, beta) + rng.normal(size=n)
    return X, y, rng


def ridge_predict(Z_train, target, Z_eval, alpha=1.0):
    """Closed-form finite Ridge prediction using einsum, avoiding BLAS matmul."""
    gram = np.einsum("ij,ik->jk", Z_train, Z_train)
    rhs = np.einsum("ij,i->j", Z_train, target)
    coef = np.linalg.solve(gram + alpha * np.eye(Z_train.shape[1]), rhs)
    return np.einsum("ij,j->i", Z_eval, coef)


def semi_pvalue(X, y, j, model, rng, seed, n_perm=1):
    """Standalone Algorithm-1 p-value with closed-form Ridge nuisances."""
    Xmj = np.delete(X, j, axis=1)
    xj = X[:, j]
    nu = ridge_predict(Xmj, xj, Xmj)
    rho = ridge_predict(np.column_stack([Xmj, y]), xj,
                        np.column_stack([Xmj, y]))
    e1 = xj - nu
    e2 = xj - rho
    diffs = []
    for _ in range(n_perm):
        x1 = X.copy()
        x2 = X.copy()
        x1[:, j] = nu + e1[rng.permutation(len(y))]
        x2[:, j] = rho + e2[rng.permutation(len(y))]
        diffs.append((model.predict(x1) - y) ** 2 -
                     (model.predict(x2) - y) ** 2)
    d = np.concatenate(diffs)
    if np.allclose(d, 0.0):
        return 1.0
    return float(wilcoxon(d, alternative="greater", zero_method="zsplit").pvalue)


def hrt_pvalue(X, y, j, seed, permutations=100):
    """One HRT p-value with one train/test split and a conditional ridge draw."""
    rng = np.random.default_rng(seed)
    order = rng.permutation(len(y))
    n_train = len(y) // 2
    train, test = order[:n_train], order[n_train:]
    model = GradientBoostingRegressor(random_state=seed).fit(X[train], y[train])
    xm = np.delete(X, j, axis=1)
    mu = ridge_predict(xm[train], X[train, j], xm[test])
    residual = X[train, j] - ridge_predict(xm[train], X[train, j], xm[train])
    observed = np.mean((model.predict(X[test]) - y[test]) ** 2)
    count = 0
    for _ in range(permutations):
        xp = X[test].copy()
        xp[:, j] = mu + rng.choice(residual, size=len(test), replace=True)
        count += float(np.mean((model.predict(xp) - y[test]) ** 2) <= observed)
    return (1.0 + count) / (1.0 + permutations)


def adjacent_grid(ns=(200, 300, 400), reps=15):
    rows = []
    for n in ns:
        sko_alt = []
        sko_null = []
        hrt_alt = []
        hrt_null = []
        for r in range(reps):
            seed = 18000 + 100 * n + r
            X, y, rng = source_design(n, 50, seed)
            model = GradientBoostingRegressor(random_state=seed).fit(X, y)
            # j=0 is in the first-quarter adjacent support; j=30 is null.
            sko_alt.append(semi_pvalue(X, y, 0, model, rng, seed=seed) <= ALPHA)
            sko_null.append(semi_pvalue(X, y, 30, model, rng, seed=seed + 1) <= ALPHA)
            hrt_alt.append(hrt_pvalue(X, y, 0, seed + 2) <= ALPHA)
            hrt_null.append(hrt_pvalue(X, y, 30, seed + 3) <= ALPHA)
        row = {
            "n": n,
            "reps": reps,
            "sko_power": float(np.mean(sko_alt)),
            "sko_type_I": float(np.mean(sko_null)),
            "hrt_power": float(np.mean(hrt_alt)),
            "hrt_type_I": float(np.mean(hrt_null)),
        }
        row["power_gap"] = row["sko_power"] - row["hrt_power"]
        rows.append(row)
        print("adjacent", row, flush=True)
    return rows


def masked_grid(reps=15, amplitudes=(0.25, 0.5, 1.0)):
    rows = []
    for amplitude in amplitudes:
        cells = []
        for r in range(reps):
            seed = 24000 + 100 * int(amplitude * 100) + r
            rng = np.random.default_rng(seed)
            X = ar1_design(300, 50, 0.6, rng)
            signal = 25
            null = signal - 1
            y = amplitude * X[:, signal] + 0.5 * rng.normal(size=300)
            # Replace the adjacent coordinate by a correlated but label-null copy.
            X[:, null] = X[:, signal] + 0.5 * rng.normal(size=300)
            model = GradientBoostingRegressor(random_state=seed).fit(X, y)
            one = semi_pvalue(X, y, signal, model, rng, seed=seed, n_perm=1) <= ALPHA
            five = semi_pvalue(X, y, signal, model, rng, seed=seed + 10,
                               n_perm=5) <= ALPHA
            null_p = semi_pvalue(X, y, null, model, rng, seed=seed + 30,
                                 n_perm=5)
            cells.append({
                "one_reject": bool(one),
                "five_reject": bool(five),
                "five_null_reject": bool(null_p <= ALPHA),
            })
        row = {
            "signal_amplitude": amplitude,
            "reps": reps,
            "single_power": float(np.mean([r["one_reject"] for r in cells])),
            "five_power": float(np.mean([r["five_reject"] for r in cells])),
            "five_null_type_I": float(np.mean([r["five_null_reject"] for r in cells])),
        }
        row["power_gain"] = row["five_power"] - row["single_power"]
        rows.append(row)
        print("masked", row, flush=True)
    out = {"reps": reps, "amplitudes": list(amplitudes), "cells": rows}
    print("masked", out, flush=True)
    return out


def main():
    OUT["adjacent"] = adjacent_grid()
    OUT["masked"] = masked_grid()
    out = Path("outputs/claim5_scope_audit.json")
    out.write_text(json.dumps(OUT, indent=2) + "\n")
    print("saved", out, flush=True)


if __name__ == "__main__":
    main()