File size: 12,839 Bytes
79bd9ac | 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 | """
Core library for reproducing the linear-regression theory of
"Why Self-Training Helps and Hurts: Denoising vs. Signal Forgetting" (arXiv 2602.14029).
Implements:
* Algorithm 1 (iterative self-training / self-distillation, ridgeless & ridge)
* Theorem 3.2 single-spike deterministic risk recursion (B*_t, V*_t)
* Theorem 3.6 multi-spike deterministic risk recursion
* iGCV estimator (eq. 11-12) and the survival / suppression factors
Notation matches the paper:
y = x'beta + eps, x ~ N(0, Sigma), eps ~ N(0, sigma^2) [noise only at t=0]
rho = p / n, tau = rho - 1 (ridgeless effective regularization)
Prediction risk R(bhat) = (bhat - beta)' Sigma (bhat - beta)
"""
import numpy as np
# ---------------------------------------------------------------------------
# Estimators & Algorithm 1
# ---------------------------------------------------------------------------
def ridgeless_fit(X, Y):
"""Minimum-norm least squares (X'X)^+ X'Y.
In the overparameterized regime (n < p, which is our entire setting) the
min-norm interpolator has the closed form X'(XX')^-1 Y, requiring only an
n x n solve -- far faster than SVD-based lstsq. Falls back to lstsq if the
Gram matrix is singular or n >= p.
"""
n, p = X.shape
if n < p:
G = X @ X.T # n x n
try:
z = np.linalg.solve(G, Y)
return X.T @ z
except np.linalg.LinAlgError:
pass
beta, *_ = np.linalg.lstsq(X, Y, rcond=None)
return beta
def project_rowspace(X, V):
"""Apply P = X^+ X (orthogonal projection onto row space of X) to columns of V.
P V = X'(XX')^-1 X V when n < p. V may be 1-D or 2-D."""
n, p = X.shape
if n < p:
G = X @ X.T
return X.T @ np.linalg.solve(G, X @ V)
# underparameterized: P = I
return V
def ridge_fit(X, Y, lam):
"""Ridge: (X'X + n*lam I)^-1 X'Y."""
n, p = X.shape
if lam == 0.0:
return ridgeless_fit(X, Y)
A = X.T @ X + n * lam * np.eye(p)
return np.linalg.solve(A, X.T @ Y)
class SpikedCov:
"""Structured spiked covariance Sigma = sum_j (s_j-1) u_j u_j' + I.
Provides O(n p) sampling and O(p) quadratic forms (no dense p x p matmul).
U columns are the orthonormal spike directions; `spikes` their strengths."""
def __init__(self, p, spikes, U):
self.p = p
self.spikes = np.asarray(spikes, float)
self.U = U # (p, k)
self.a = np.sqrt(self.spikes) - 1.0 # sqrt-eigval offset
def sample(self, m, rng):
Z = rng.standard_normal((m, self.p))
return Z + (Z @ self.U) * self.a @ self.U.T # X ~ N(0, Sigma)
def quad(self, d): # d' Sigma d
Ud = self.U.T @ d
return float(d @ d + ((self.spikes - 1.0) * Ud) @ Ud)
class DiagCov:
"""Diagonal covariance Sigma = diag(v). O(n p) sampling."""
def __init__(self, v):
self.v = np.asarray(v, float)
self.sq = np.sqrt(self.v)
self.p = len(self.v)
def sample(self, m, rng):
return rng.standard_normal((m, self.p)) * self.sq
def quad(self, d):
return float((d * self.v) @ d)
def iterative_self_train(Sigma_sqrt, beta, n, sigma, T, lam=0.0, rng=None,
return_betas=False, cov=None):
"""
Run Algorithm 1 for T iterations (t = 0 .. T).
t=0 : fit on noisy data (Y0 = X0 beta + eps).
t>=1: fresh X_t, noiseless pseudo-labels Y_t = X_t bhat_{t-1}, refit.
Sigma_sqrt : (p,p) symmetric square-root of the feature covariance Sigma.
Returns array of prediction risks R_t, shape (T+1,). Optionally the betas.
"""
if rng is None:
rng = np.random.default_rng()
p = beta.shape[0]
if cov is not None:
gen_X = lambda m: cov.sample(m, rng)
def risk(bhat):
return cov.quad(bhat - beta)
else:
Sigma = Sigma_sqrt @ Sigma_sqrt # Sigma_sqrt is symmetric
gen_X = lambda m: rng.standard_normal((m, p)) @ Sigma_sqrt
def risk(bhat):
d = bhat - beta
return float(d @ (Sigma @ d))
# t = 0 : noisy fit
X0 = gen_X(n)
eps = sigma * rng.standard_normal(n)
Y0 = X0 @ beta + eps
bhat = ridge_fit(X0, Y0, lam) if lam > 0 else ridgeless_fit(X0, Y0)
risks = [risk(bhat)]
betas = [bhat.copy()]
for t in range(1, T + 1):
Xt = gen_X(n)
Yt = Xt @ bhat # noiseless pseudo-labels
bhat = ridge_fit(Xt, Yt, lam) if lam > 0 else ridgeless_fit(Xt, Yt)
risks.append(risk(bhat))
betas.append(bhat.copy())
risks = np.array(risks)
return (risks, betas) if return_betas else risks
def simulate_risk(Sigma_sqrt, beta, n, sigma, T, lam=0.0, trials=10, seed=0, cov=None):
"""Monte-Carlo prediction risk R_t averaged over `trials`."""
rng = np.random.default_rng(seed)
acc = np.zeros(T + 1)
sq = np.zeros(T + 1)
for _ in range(trials):
r = iterative_self_train(Sigma_sqrt, beta, n, sigma, T, lam=lam, rng=rng, cov=cov)
acc += r
sq += r ** 2
mean = acc / trials
std = np.sqrt(np.maximum(sq / trials - mean ** 2, 0.0))
return mean, std
# ---------------------------------------------------------------------------
# Theorem 3.2 -- single-spike deterministic recursion
# ---------------------------------------------------------------------------
def spiked_theory(s, rho, r2, sigma2, T):
"""
Deterministic risk R*_t = B*_t + V*_t for the single-spike model (Thm 3.2).
Sigma = (s-1) u1 u1' + I, beta = r u1 with r^2 = r2.
tau = rho - 1.
Returns dict with arrays B, V, R (length T+1) and 'survival' factor.
"""
tau = rho - 1.0
kappa = s / (s + tau) # contraction / survival factor
survival = kappa ** (np.arange(T + 1) + 1) # (s/(s+tau))^{t+1}
B = r2 * s * (1.0 - survival) ** 2 # eq (4)
V = np.zeros(T + 1)
V[0] = sigma2 / tau + (tau * s / (s + tau) ** 2) * r2
for t in range(1, T + 1):
# eq (5): V_t = V_{t-1}/(1+tau) + tau r^2 s^{2t+1}/(s+tau)^{2(t+1)}
V[t] = V[t - 1] / (1.0 + tau) + tau * r2 * s ** (2 * t + 1) / (s + tau) ** (2 * (t + 1))
return {"B": B, "V": V, "R": B + V, "tau": tau, "kappa": kappa,
"survival": survival}
def multi_spike_theory(spikes, r2s, rho, sigma2, T):
"""
Multi-spike deterministic recursion (Thm 3.6).
spikes : list of spike strengths s_1..s_k (each > 1)
r2s : list of signal powers r_j^2 along each spike direction.
Returns dict with B, V, R arrays and per-direction survival factors.
"""
spikes = np.asarray(spikes, float)
r2s = np.asarray(r2s, float)
tau = rho - 1.0
tt = np.arange(T + 1)
# eq (6): B_t = sum_j r_j^2 s_j (1 - (s_j/(s_j+tau))^{t+1})^2
B = np.zeros(T + 1)
survivals = {}
for j, (s, rj2) in enumerate(zip(spikes, r2s)):
surv = (s / (s + tau)) ** (tt + 1)
survivals[j] = surv
B += rj2 * s * (1.0 - surv) ** 2
# eq (7): V recursion
V = np.zeros(T + 1)
V[0] = sigma2 / tau + np.sum(tau * r2s * spikes / (spikes + tau) ** 2)
for t in range(1, T + 1):
inject = np.sum(tau * r2s * spikes ** (2 * t + 1) / (spikes + tau) ** (2 * (t + 1)))
V[t] = V[t - 1] / (1.0 + tau) + inject
return {"B": B, "V": V, "R": B + V, "tau": tau, "survivals": survivals,
"kappas": spikes / (spikes + tau)}
def general_diag_theory(eigs, beta, rho, sigma2, T, lam=0.0):
"""
General deterministic-equivalent recursion (Section 4, Thm 4.2 / eq 10) for a
diagonal feature covariance Sigma = diag(eigs), identical across iterations,
with aspect ratio rho = p/n_t fixed (so tau_t = tau constant). Returns the
deterministic prediction risk R*_t decomposed into systematic + stochastic.
Because Sigma is diagonal and constant, Q_t = Q = diag(q_i), q_i=lam_i/(lam_i+tau),
and every trace reduces to a 1-D sum over eigenvalues.
"""
eigs = np.asarray(eigs, float)
beta = np.asarray(beta, float)
p = len(eigs)
# solve fixed point (8): 1/rho = (1/p) sum_i lam_i/(lam_i+tau) + lam/tau
def fp(tau):
return (np.mean(eigs / (eigs + tau)) + lam / tau) - 1.0 / rho
lo, hi = 1e-8, 1e8
for _ in range(200):
mid = np.sqrt(lo * hi)
if fp(mid) > 0: # decreasing in tau
lo = mid
else:
hi = mid
tau = np.sqrt(lo * hi)
q = eigs / (eigs + tau) # Q diagonal
L = lam / tau + (tau / p) * np.sum(eigs / (eigs + tau) ** 2)
# deterministic effective noise D^2_t
D = np.zeros(T + 1)
D[0] = (sigma2 + tau ** 2 * np.sum(beta ** 2 * eigs / (eigs + tau) ** 2)) / L
for t in range(1, T + 1):
term1 = tau ** 2 * np.sum(beta ** 2 * q ** (2 * t + 1) / (eigs + tau))
term2 = 0.0
for h in range(t):
trace = np.sum(q ** (2 * (t - h)) / (eigs + tau) ** 2)
term2 += (D[h] / p) * trace
D[t] = (term1 + tau ** 2 * term2) / L
# deterministic risk R*_t (eq 10)
B = np.zeros(T + 1) # systematic
V = np.zeros(T + 1) # stochastic
for t in range(T + 1):
B[t] = np.sum(eigs * (q ** (t + 1) - 1.0) ** 2 * beta ** 2)
acc = 0.0
for h in range(t + 1):
trace = np.sum(q ** (2 * (t - h)) * eigs ** 2 / (eigs + tau) ** 2)
acc += (D[h] / p) * trace
V[t] = acc
return {"B": B, "V": V, "R": B + V, "tau": tau, "D": D}
def build_spiked_covariance(p, spikes, dirs=None):
"""
Sigma = sum_j (s_j - 1) u_j u_j' + I_p. Returns (Sigma_sqrt, U) with U the
spike eigenvectors (columns). dirs: optional (p,k) orthonormal directions;
default = first k canonical basis vectors.
"""
spikes = np.asarray(spikes, float)
k = len(spikes)
if dirs is None:
U = np.zeros((p, k))
for j in range(k):
U[j, j] = 1.0
else:
U = dirs
# eigen-decomposition is trivial: Sigma_sqrt = I + sum_j (sqrt(s_j)-1) u_j u_j'
Sigma_sqrt = np.eye(p)
for j in range(k):
uj = U[:, j]
Sigma_sqrt += (np.sqrt(spikes[j]) - 1.0) * np.outer(uj, uj)
return Sigma_sqrt, U
# ---------------------------------------------------------------------------
# iGCV estimator (Section 4.2, eq. 11-12)
# ---------------------------------------------------------------------------
def igcv_trajectory(Sigma_sqrt, beta, n, sigma, T, lam=0.0, rng=None, cov=None):
"""
One trial: returns (true_risk[t], igcv_est[t]) for t=0..T.
iGCV (eq 12) estimates R(bhat_t) + sigma^2 using ONLY the initial noisy
dataset D0 and the cumulative projection A_t = P_t...P_1.
We report igcv_est - sigma^2 as the estimate of R(bhat_t).
"""
if rng is None:
rng = np.random.default_rng()
p = beta.shape[0]
if cov is not None:
gen_X = lambda m: cov.sample(m, rng)
risk = lambda bhat: cov.quad(bhat - beta)
else:
Sigma = Sigma_sqrt @ Sigma_sqrt
gen_X = lambda m: rng.standard_normal((m, p)) @ Sigma_sqrt
def risk(bhat):
d = bhat - beta
return float(d @ (Sigma @ d))
# Initial ridge fit on the noisy data D0. lam>0 keeps the GCV correction
# well-conditioned (the interpolating ridgeless fit has zero residuals, so
# the leave-one-out correction 1 - tr(H)/n0 degenerates; the paper uses the
# ridge / pseudoinverse-continuity profile, Hastie et al. 2022; Patil 2021).
# The self-training iterations t>=1 are always ridgeless row-space projections.
lam0 = lam if lam > 0 else 1e-3
X0 = gen_X(n)
eps = sigma * rng.standard_normal(n)
Y0 = X0 @ beta + eps
bhat0 = ridge_fit(X0, Y0, lam0)
# smoother H = X0 (X0'X0/n + lam0 I)^-1 X0' / n0 (n0 x n0); X0 @ C = H
G = X0.T @ X0 / n
Ginv = np.linalg.inv(G + lam0 * np.eye(p))
C = Ginv @ X0.T / n # p x n0
H = X0 @ C # n0 x n0
denom = 1.0 - np.trace(H) / n
resid0 = Y0 - X0 @ bhat0 # y_i - x_i' bhat0 (nonzero for lam0>0)
# Trajectory: bhat_t = P_t ... P_1 bhat0 = A_t bhat0.
A_bhat0 = bhat0.copy()
A_C = C.copy()
true_risk, igcv = [], []
for t in range(0, T + 1):
if t >= 1:
Xt = gen_X(n)
A_C = project_rowspace(Xt, A_C) # P_t A_{t-1} C
A_bhat0 = project_rowspace(Xt, A_bhat0) # P_t A_{t-1} bhat0
true_risk.append(risk(A_bhat0))
# leverage multiplier M_t (eq 11) and corrected residual iGCV (eq 12)
Mt = (np.trace(X0 @ A_C) / n) / denom
corr = (Y0 - X0 @ A_bhat0) + resid0 * Mt
igcv.append(float(np.mean(corr ** 2)) - sigma ** 2)
return np.array(true_risk), np.array(igcv)
|