File size: 14,355 Bytes
1c49ecf a9ed094 1c49ecf a9ed094 1c49ecf a9ed094 1c49ecf | 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 | """
rmt_diffusion.py
================
Core library for reproducing the LINEAR-THEORY claims of:
"A Random Matrix Theory Perspective on the Consistency of Diffusion Models"
Binxu Wang, Jacob A. Zavatone-Veth, Cengiz Pehlevan (ICML 2026, arXiv:2602.02908)
The paper's central claims about linear diffusion models reduce to random-matrix /
linear-algebra statements about the empirical covariance of a finite dataset. This
module implements:
* solve_kappa -- the self-consistent renormalized-noise map kappa(lambda) (Eq. 4)
* denoiser_matrix -- the optimal linear denoiser Sigma_hat (Sigma_hat + s^2 I)^-1 (Eq. 2)
* deterministic-equivalence predictions for the denoiser EXPECTATION (Result 4.1)
and VARIANCE (Result 4.2), plus the sampling-map overshrinkage (Result 5.1)
* Monte-Carlo estimators over dataset realizations to VALIDATE those predictions
Everything runs on CPU in minutes. We work in the eigenbasis of the population
covariance Sigma = diag(eigs) (WLOG) and set the population mean mu = 0, exactly the
simplification the paper adopts (mu_hat = mu) to isolate finite-sample covariance effects.
"""
from __future__ import annotations
import numpy as np
# np.trapz was renamed to np.trapezoid in NumPy 2.0 (and removed later); support both.
_trapz = getattr(np, "trapezoid", None) or np.trapz
# --------------------------------------------------------------------------------------
# Population covariance
# --------------------------------------------------------------------------------------
def power_law_spectrum(d: int, alpha: float = 1.0, floor: float = 1e-3,
normalize: bool = True) -> np.ndarray:
"""Population eigenvalues lambda_k = k^{-alpha}, k = 1..d.
Natural-image covariances have an approximately power-law spectrum (Ruderman 1994),
which is the regime the paper studies. `floor` keeps the smallest eigenvalues from
underflowing; `normalize` sets the top eigenvalue to 1.
"""
k = np.arange(1, d + 1, dtype=float)
eigs = k ** (-alpha) + floor
if normalize:
eigs = eigs / eigs[0]
return eigs # descending
def sample_empirical_cov(eigs: np.ndarray, n: int, rng: np.random.Generator) -> np.ndarray:
"""Draw n samples x_i ~ N(0, diag(eigs)) and return the empirical covariance
Sigma_hat = (1/n) sum_i x_i x_i^T (a d x d Wishart-type matrix)."""
d = eigs.shape[0]
z = rng.standard_normal((n, d))
x = z * np.sqrt(eigs)[None, :] # rows ~ N(0, diag(eigs))
return (x.T @ x) / n
# --------------------------------------------------------------------------------------
# Renormalized noise scale kappa(lambda) (Silverstein / Marchenko-Pastur; Eq. 4)
# --------------------------------------------------------------------------------------
def _normalized_trace_resolvent(eigs: np.ndarray, kappa: float) -> float:
""" (1/d) * sum_k lambda_k / (lambda_k + kappa) = tr[Sigma (Sigma + kappa I)^-1]. """
return float(np.mean(eigs / (eigs + kappa)))
def solve_kappa(lam: float, eigs: np.ndarray, gamma: float,
tol: float = 1e-12, max_iter: int = 200) -> float:
"""Solve the self-consistent equation (Eq. 4):
kappa - lam = gamma * kappa * tr[Sigma (Sigma + kappa I)^-1]
for the unique kappa > 0, by bisection. gamma = d / n is the aspect ratio.
Returns kappa >= lam (finite data renormalize the noise scale UP).
"""
if lam <= 0:
return 0.0
def h(kappa: float) -> float:
return kappa - lam - gamma * kappa * _normalized_trace_resolvent(eigs, kappa)
lo = lam # h(lam) = -gamma*lam*g(lam) <= 0
hi = lam + max(lam, 1.0)
# expand upper bracket until h(hi) > 0
it = 0
while h(hi) < 0 and it < 100:
hi *= 2.0
it += 1
for _ in range(max_iter):
mid = 0.5 * (lo + hi)
hm = h(mid)
if abs(hm) < tol or (hi - lo) < tol * max(1.0, mid):
return mid
if hm < 0:
lo = mid
else:
hi = mid
return 0.5 * (lo + hi)
# --------------------------------------------------------------------------------------
# Degrees-of-freedom functions (Eq. 5, UNNORMALIZED trace Tr)
# --------------------------------------------------------------------------------------
def df1(eigs: np.ndarray, lam: float) -> float:
return float(np.sum(eigs / (eigs + lam)))
def df2(eigs: np.ndarray, lam: float) -> float:
return float(np.sum(eigs ** 2 / (eigs + lam) ** 2))
# --------------------------------------------------------------------------------------
# Linear denoiser (Eq. 2, mu = 0): D*(x; s) = Sigma_hat (Sigma_hat + s^2 I)^-1 x
# --------------------------------------------------------------------------------------
def denoiser_matrix(Sigma: np.ndarray, sigma2: float) -> np.ndarray:
"""Return the linear-denoiser matrix M = C (C + sigma2 I)^-1 for covariance C."""
d = Sigma.shape[0]
return Sigma @ np.linalg.solve(Sigma + sigma2 * np.eye(d), np.eye(d))
def population_denoiser_diag(eigs: np.ndarray, ridge: float) -> np.ndarray:
"""Diagonal (in population eigenbasis) of the population denoiser with ridge penalty:
lambda_k / (lambda_k + ridge). With ridge = kappa(sigma2) this is Result 4.1."""
return eigs / (eigs + ridge)
# --------------------------------------------------------------------------------------
# Deterministic-equivalence predictions
# --------------------------------------------------------------------------------------
def predict_shrinkage_along_pc(eigs: np.ndarray, sigma2: float, gamma: float):
"""Result 4.1 / Fig 2C: expected shrinkage of the empirical denoiser along population
PC u_k is lambda_k/(lambda_k + kappa(sigma2)) (renormalized), which OVER-shrinks
relative to the naive population value lambda_k/(lambda_k + sigma2)."""
kappa = solve_kappa(sigma2, eigs, gamma)
renorm = eigs / (eigs + kappa) # DE prediction (what finite data give)
naive = eigs / (eigs + sigma2) # infinite-data population denoiser
return kappa, renorm, naive
def predict_denoiser_variance_along_pc(eigs: np.ndarray, sigma2: float, n: int,
x_vec: np.ndarray):
"""Result 4.2: Var over dataset realizations of u_k^T D*_hat(x; sigma) , per PC k.
Var ~ [ kappa^2 / (n - df2(kappa)) ] * chi(lambda_k, kappa) * calD(x, kappa)
with chi(lambda, kappa) = lambda/(lambda+kappa)^2 (anisotropy; bell-shaped, peak at
lambda = kappa, peak value 1/(4 kappa)) and calD(x,kappa) = sum_k lambda_k x_k^2/(lambda_k+kappa)^2
(inhomogeneity). Returns (kappa, per-k variance prediction, chi, peak_value).
"""
d = eigs.shape[0]
gamma = d / n
kappa = solve_kappa(sigma2, eigs, gamma)
chi = eigs / (eigs + kappa) ** 2 # anisotropy per PC
inhom = float(np.sum(eigs * x_vec ** 2 / (eigs + kappa) ** 2)) # calD(x, kappa)
prefactor = kappa ** 2 / (n - df2(eigs, kappa))
var_pred = prefactor * chi * inhom
peak_value = 1.0 / (4.0 * kappa) # max of chi at lambda = kappa
return kappa, var_pred, chi, inhom, peak_value
def predict_sqrt_cov_scaling(eigs: np.ndarray, n: int):
"""Result 5.1 / Fig 4A: the sampling map contains Sigma_hat^{1/2}. Its expected scaling
along population eigenmode u_k, E[u_k^T Sigma_hat^{1/2} u_k], OVER-shrinks relative to
the ideal sqrt(lambda_k), most severely for low eigenmodes and small n.
A convenient deterministic-equivalence-style prediction (Balakrishnan integral of the
kappa map) for the per-mode scaling is:
s_k ~ (2/pi) * integral_0^inf lambda_k / (lambda_k + kappa(u^2)) du
which we evaluate numerically. Returns (ideal sqrt(lambda_k), predicted s_k)."""
ideal = np.sqrt(eigs)
gamma = eigs.shape[0] / n
# integrate over u on a log-spaced grid; integrand decays like lambda_k/u^2 for large u
u = np.concatenate([np.linspace(1e-4, 5.0, 4000), np.linspace(5.0, 200.0, 4000)])
kap = np.array([solve_kappa(uu ** 2, eigs, gamma) for uu in u])
pred = np.empty_like(eigs)
for k, lk in enumerate(eigs):
integrand = lk / (lk + kap)
pred[k] = (2.0 / np.pi) * _trapz(integrand, u)
return ideal, pred
# --------------------------------------------------------------------------------------
# Monte-Carlo estimators (ground truth to validate the DE predictions above)
# --------------------------------------------------------------------------------------
def mc_trace_resolvent(eigs: np.ndarray, lam: float, n: int, R: int,
rng: np.random.Generator) -> float:
"""MC estimate of (1/d) Tr[Sigma_hat (Sigma_hat + lam I)^-1], averaged over R draws.
Validates the deterministic equivalence ~ (1/d) Tr[Sigma (Sigma + kappa(lam) I)^-1]."""
d = eigs.shape[0]
vals = np.empty(R)
I = np.eye(d)
for r in range(R):
C = sample_empirical_cov(eigs, n, rng)
M = C @ np.linalg.solve(C + lam * I, I)
vals[r] = np.trace(M) / d
return float(vals.mean())
def mc_denoiser_stats(eigs: np.ndarray, sigma2: float, n: int, R: int,
x_vec: np.ndarray, rng: np.random.Generator):
"""Monte-Carlo mean and variance, over R dataset realizations, of the per-PC denoiser
response u_k^T D*_hat(x; sigma) (with population PCs = coordinate axes here).
Returns (mean_k, var_k) arrays of length d."""
d = eigs.shape[0]
I = np.eye(d)
resp = np.empty((R, d))
for r in range(R):
C = sample_empirical_cov(eigs, n, rng)
M = C @ np.linalg.solve(C + sigma2 * I, I) # denoiser matrix
resp[r] = M @ x_vec # response vector; u_k^T (.) = coord k
return resp.mean(axis=0), resp.var(axis=0, ddof=1)
def mc_sqrt_cov_scaling(eigs: np.ndarray, n: int, R: int,
rng: np.random.Generator) -> np.ndarray:
"""MC estimate of E[u_k^T Sigma_hat^{1/2} u_k] per population eigenmode k."""
d = eigs.shape[0]
acc = np.zeros(d)
for r in range(R):
C = sample_empirical_cov(eigs, n, rng)
w, V = np.linalg.eigh(C)
w = np.clip(w, 0.0, None)
C_half = (V * np.sqrt(w)) @ V.T
acc += np.diag(C_half) # u_k = e_k in population eigenbasis
return acc / R
def sqrt_cov(C: np.ndarray) -> np.ndarray:
"""Symmetric PSD square root of C (the linear generative / sampling map)."""
w, V = np.linalg.eigh(C)
w = np.clip(w, 0.0, None)
return (V * np.sqrt(w)) @ V.T
def mc_sqrtmap_variance(eigs: np.ndarray, n: int, R: int,
rng: np.random.Generator) -> np.ndarray:
"""Result 5.2: per-mode VARIANCE, across dataset realizations, of the sampling-map
diagonal u_k^T Sigma_hat^{1/2} u_k. For a linear/Gaussian score model the
probability-flow ODE integrates in closed form to the map x = Sigma_hat^{1/2} z, so
this is the variance of the FULL generative trajectory (not a one-step denoise).
Returns the per-mode variance (length d)."""
d = eigs.shape[0]
vals = np.empty((R, d))
for r in range(R):
C = sample_empirical_cov(eigs, n, rng)
vals[r] = np.diag(sqrt_cov(C))
return vals.var(axis=0, ddof=1)
def predict_sqrtmap_variance(eigs: np.ndarray, n: int) -> np.ndarray:
"""Leading-order deterministic-equivalence prediction for Result 5.2. With u_k = e_k,
u_k^T Sigma_hat u_k = (1/n) sum_i (z_ik^2) lambda_k has variance 2 lambda_k^2 / n
exactly; the delta method through g(t)=sqrt(t) (g'=1/(2 sqrt(lambda_k))) gives
Var[u_k^T Sigma_hat^{1/2} u_k] ~ lambda_k / (2 n)
i.e. anisotropic (proportional to lambda_k) and decaying as 1/n. The residual vs MC is
the finite-sample coupling of off-diagonal Sigma_hat entries into the matrix sqrt."""
return eigs / (2.0 * n)
def mc_split_consistency(eigs: np.ndarray, n: int, n_seeds: int,
rng: np.random.Generator):
"""Fig 1 (linear model): two NON-OVERLAPPING data splits A, B (n samples each, disjoint)
each define a linear diffusion sampler x = Sigma_hat^{1/2} z. Generate samples from the
SAME seeds z under both splits and measure cross-split agreement:
* mean cosine similarity cos(x_A, x_B) -> 1 as n grows
* mean relative squared deviation ||x_A-x_B||^2 / (||x_A|| ||x_B||)
The deviation is set by the sampling-map variance (Result 5.2), so it decays ~ 1/n:
finite datasets that never share a sample still generate the same picture from a seed,
and they agree better with more data. Returns (mean_cosine, mean_rel_sq_deviation)."""
d = eigs.shape[0]
HA = sqrt_cov(sample_empirical_cov(eigs, n, rng)) # split A sampler
HB = sqrt_cov(sample_empirical_cov(eigs, n, rng)) # split B sampler (disjoint draw)
Z = rng.standard_normal((n_seeds, d))
XA = Z @ HA.T
XB = Z @ HB.T
nA = np.linalg.norm(XA, axis=1)
nB = np.linalg.norm(XB, axis=1)
cos = np.sum(XA * XB, axis=1) / (nA * nB)
rel_sq = np.sum((XA - XB) ** 2, axis=1) / (nA * nB)
return float(cos.mean()), float(rel_sq.mean())
def mc_total_denoiser_variance(eigs: np.ndarray, sigma2: float, n: int, R: int,
rng: np.random.Generator) -> float:
"""Global scaling (Fig 3D): total variance of the denoiser matrix entries across
realizations, Sum_{ij} Var[M_ij], which the theory predicts decays ~ 1/n at large n."""
d = eigs.shape[0]
I = np.eye(d)
mats = np.empty((R, d, d))
for r in range(R):
C = sample_empirical_cov(eigs, n, rng)
mats[r] = C @ np.linalg.solve(C + sigma2 * I, I)
return float(mats.var(axis=0, ddof=1).sum())
# --------------------------------------------------------------------------------------
# Small self-test
# --------------------------------------------------------------------------------------
if __name__ == "__main__":
rng = np.random.default_rng(0)
eigs = power_law_spectrum(50, alpha=1.0)
lam = 0.1
gamma = 50 / 500
k = solve_kappa(lam, eigs, gamma)
print(f"kappa({lam}) = {k:.5f} (>= lam: {k >= lam})")
de = _normalized_trace_resolvent(eigs, k)
mc = mc_trace_resolvent(eigs, lam, n=500, R=200, rng=rng)
print(f"trace resolvent DE={de:.5f} MC={mc:.5f} rel.err={abs(de-mc)/mc:.3%}")
|