vimarsh's picture
Add reproduction scripts
1f48ccf verified
Raw
History Blame Contribute Delete
10.5 kB
"""Core simulation library for reproducing arXiv:2602.02431 (ICML 2026 #26332).
Single-index model: x_i ~ N(0, I_d), y_i = sigma(<x_i, theta*>), ||theta*|| = 1.
Activations
-----------
* ``quad`` sigma(z) = z^2 (paper Sec. 3.1)
* ``trunc`` sigma(z) = min(z^2, M) (paper eq. 4.3, hard truncation)
* ``smooth`` sigma(z) = int_0^{z^2} phi(u) du (paper eq. 3.10, smooth truncation)
Algorithms
----------
* ``spherical_flow`` full-batch spherical GD on the correlation loss
L(theta) = -(1/n) sum_i y_i sigma(<x_i, theta>)
theta <- normalize(theta + eta (I - theta theta^T) A(theta) theta)
with A(theta) = (2/n) sum_i y_i phi(<x_i,theta>^2) x_i x_i^T.
* ``online_sgd`` one-pass spherical SGD on the same loss (each sample used once).
* ``squared_gd`` full-batch Euclidean GD on the squared loss (paper Sec. 4).
"""
from __future__ import annotations
import math
from dataclasses import dataclass
import torch
# --------------------------------------------------------------------------- #
# activations
# --------------------------------------------------------------------------- #
def _bump(u: torch.Tensor) -> torch.Tensor:
"""gamma(u) = exp(-1/u) for u > 0, else 0."""
out = torch.zeros_like(u)
pos = u > 0
out[pos] = torch.exp(-1.0 / u[pos])
return out
def phi_smooth(u: torch.Tensor, M: float) -> torch.Tensor:
"""C^inf cutoff: phi = 1 for |u| <= M, 0 for |u| >= 2M (paper Sec. 3.2)."""
t = (u.abs() - M) / M
g0, g1 = _bump(t), _bump(1.0 - t)
S = torch.where(g0 + g1 > 0, g0 / (g0 + g1 + 1e-300), torch.zeros_like(t))
return (1.0 - S).clamp_(0.0, 1.0)
_SMOOTH_TABLE: dict[tuple[float, str, str], tuple[torch.Tensor, torch.Tensor]] = {}
def _smooth_sigma_table(M: float, device, dtype, npts: int = 200_001):
key = (M, str(device), str(dtype))
if key not in _SMOOTH_TABLE:
u = torch.linspace(0.0, 2.0 * M, npts, device=device, dtype=dtype)
f = phi_smooth(u, M)
du = u[1] - u[0]
cum = torch.cumsum((f[1:] + f[:-1]) * 0.5 * du, dim=0)
cum = torch.cat([torch.zeros(1, device=device, dtype=dtype), cum])
_SMOOTH_TABLE[key] = (u, cum)
return _SMOOTH_TABLE[key]
def sigma(z: torch.Tensor, act: str, M: float) -> torch.Tensor:
if act == "quad":
return z * z
if act == "trunc":
return torch.clamp(z * z, max=M)
if act == "smooth":
u, cum = _smooth_sigma_table(M, z.device, z.dtype)
w = torch.clamp(z * z, max=2.0 * M)
idx = torch.clamp(
torch.searchsorted(u, w.reshape(-1).contiguous()), 1, u.numel() - 1
)
u0, u1 = u[idx - 1], u[idx]
c0, c1 = cum[idx - 1], cum[idx]
frac = (w.reshape(-1) - u0) / (u1 - u0)
return (c0 + frac * (c1 - c0)).reshape(z.shape)
raise ValueError(act)
def phi(w: torch.Tensor, act: str, M: float) -> torch.Tensor:
"""phi(u) with sigma'(z) = 2 z phi(z^2); argument ``w`` is z^2."""
if act == "quad":
return torch.ones_like(w)
if act == "trunc":
return (w < M).to(w.dtype)
if act == "smooth":
return phi_smooth(w, M)
raise ValueError(act)
def sigma_prime(z: torch.Tensor, act: str, M: float) -> torch.Tensor:
return 2.0 * z * phi(z * z, act, M)
# --------------------------------------------------------------------------- #
# data
# --------------------------------------------------------------------------- #
@dataclass
class Data:
X: torch.Tensor
y: torch.Tensor
theta_star: torch.Tensor
def make_data(d: int, n: int, seed: int, act: str, M: float, device, dtype) -> Data:
g = torch.Generator(device=device).manual_seed(seed)
theta_star = torch.randn(d, generator=g, device=device, dtype=dtype)
theta_star /= theta_star.norm()
X = torch.randn(n, d, generator=g, device=device, dtype=dtype)
y = sigma(X @ theta_star, act, M)
return Data(X, y, theta_star)
def rand_sphere(d: int, seed: int, device, dtype) -> torch.Tensor:
g = torch.Generator(device=device).manual_seed(seed)
v = torch.randn(d, generator=g, device=device, dtype=dtype)
return v / v.norm()
# --------------------------------------------------------------------------- #
# full-batch spherical gradient descent on the correlation loss
# --------------------------------------------------------------------------- #
def a_star(data: Data) -> torch.Tensor:
"""A* = (2/n) sum_i y_i x_i x_i^T (paper eq. 3.3)."""
n = data.X.shape[0]
return (2.0 / n) * (data.X.T @ (data.y[:, None] * data.X))
def _Atheta_matvec(data: Data, theta: torch.Tensor, act: str, M: float) -> torch.Tensor:
"""A(theta) @ theta without forming A(theta) (paper eq. 3.11)."""
z = data.X @ theta
w = data.y * phi(z * z, act, M) * z
return (2.0 / data.X.shape[0]) * (data.X.T @ w)
def spherical_flow(
data: Data,
theta0: torch.Tensor,
act: str,
M: float,
eta: float = 0.1,
T: int = 1000,
tol: float = 1e-12,
check_every: int = 50,
use_matrix: bool | None = None,
record_every: int = 0,
):
"""Full-batch spherical GD on the correlation loss (paper eq. 3.4 / 3.12).
Returns ``(theta, steps_run, trace)`` where ``trace`` is a list of
``(step, squared_overlap)`` when ``record_every > 0``.
"""
if use_matrix is None:
use_matrix = act == "quad"
A = a_star(data) if use_matrix else None
theta = theta0.clone()
ts = data.theta_star
trace = []
prev_ray = None
prev_ov = None
steps = T
for t in range(T):
Ath = (A @ theta) if use_matrix else _Atheta_matvec(data, theta, act, M)
ray = theta @ Ath
grad = Ath - ray * theta # (I - theta theta^T) A(theta) theta
theta = theta + eta * grad
theta = theta / theta.norm()
if record_every and (t % record_every == 0 or t == T - 1):
trace.append((t + 1, float((theta @ ts) ** 2)))
if (t + 1) % check_every == 0:
# converged: Rayleigh quotient (= -loss) and overlap both stationary
ray, ov = float(ray), float((theta @ ts) ** 2)
if (
prev_ray is not None
and abs(ray - prev_ray) <= tol * max(abs(ray), 1e-30)
and abs(ov - prev_ov) <= tol
):
steps = t + 1
break
prev_ray, prev_ov = ray, ov
return theta, steps, trace
# --------------------------------------------------------------------------- #
# one-pass (online) spherical SGD on the correlation loss
# --------------------------------------------------------------------------- #
def online_sgd(
d: int,
n: int,
seeds: int,
act: str,
M: float,
eta: float,
seed0: int,
device,
dtype,
checkpoints: list[int],
chunk: int = 2048,
):
"""One-pass spherical SGD, vectorised over ``seeds`` independent replicas.
theta <- normalize(theta + eta (I - theta theta^T) y_t sigma'(<x_t,theta>) x_t)
Returns dict ``{n_used: mean squared overlap}`` measured at ``checkpoints``.
"""
g = torch.Generator(device=device).manual_seed(seed0)
ts = torch.randn(seeds, d, generator=g, device=device, dtype=dtype)
ts /= ts.norm(dim=1, keepdim=True)
th = torch.randn(seeds, d, generator=g, device=device, dtype=dtype)
th /= th.norm(dim=1, keepdim=True)
out: dict[int, float] = {}
cps = sorted(checkpoints)
ci = 0
done = 0
while done < n:
m = min(chunk, n - done)
Xc = torch.randn(seeds, m, d, generator=g, device=device, dtype=dtype)
for j in range(m):
x = Xc[:, j, :] # (S, d)
zstar = (x * ts).sum(1)
y = sigma(zstar, act, M)
z = (x * th).sum(1)
coef = y * sigma_prime(z, act, M) # (S,)
gvec = coef[:, None] * x
gvec = gvec - (gvec * th).sum(1, keepdim=True) * th
th = th + eta * gvec
th = th / th.norm(dim=1, keepdim=True)
done += 1
while ci < len(cps) and done == cps[ci]:
out[done] = float(((th * ts).sum(1) ** 2).mean())
ci += 1
del Xc
return out
# --------------------------------------------------------------------------- #
# full-batch Euclidean GD on the squared loss (paper Sec. 4)
# --------------------------------------------------------------------------- #
def squared_gd(
data: Data,
theta0: torch.Tensor,
act: str,
M: float,
eta: float,
T: int,
record_every: int = 1,
stop_err: float | None = None,
):
"""theta_{t+1} = theta_t - eta * (1/n) sum_i (sigma(<x_i,th>) - y_i) sigma'(<x_i,th>) x_i.
Returns a dict of trajectory arrays (step, sq_overlap, norm, dist2, loss).
"""
X, y, ts = data.X, data.y, data.theta_star
n = X.shape[0]
theta = theta0.clone()
rec = {"step": [], "sq_overlap": [], "norm": [], "dist2": [], "loss": []}
def _record(t):
nr = float(theta.norm())
ov = float((theta @ ts) ** 2) / max(nr * nr, 1e-300)
d2 = min(
float(((theta - ts) ** 2).sum()), float(((theta + ts) ** 2).sum())
)
z = X @ theta
loss = float((0.5 / n) * ((sigma(z, act, M) - y) ** 2).sum())
rec["step"].append(t)
rec["sq_overlap"].append(ov)
rec["norm"].append(nr)
rec["dist2"].append(d2)
rec["loss"].append(loss)
return d2
_record(0)
for t in range(1, T + 1):
z = X @ theta
resid = (sigma(z, act, M) - y) * sigma_prime(z, act, M)
grad = (X.T @ resid) / n
theta = theta - eta * grad
if record_every and (t % record_every == 0 or t == T):
d2 = _record(t)
if stop_err is not None and d2 < stop_err:
break
return rec
# --------------------------------------------------------------------------- #
# helpers
# --------------------------------------------------------------------------- #
def top2_eig(A: torch.Tensor):
"""Top-two eigenvalues and top eigenvector of a symmetric matrix."""
A = 0.5 * (A + A.T)
evals, evecs = torch.linalg.eigh(A.double())
return float(evals[-1]), float(evals[-2]), evecs[:, -1].to(A.dtype)
def log2_steps(d: int, mult: float = 1000.0) -> int:
return int(mult * math.log(d) ** 2)