File size: 15,043 Bytes
9f8d17a | 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 | """Sections D and E: structural-precondition tests on real solver output and
measured pseudo-dimension of real tuning classes against the theorem bounds."""
import itertools, json, math, sys, time
import numpy as np
from scipy import linalg
from evidence import stated_bound, measure_pdim_pool
RNG = np.random.default_rng(4242)
# --------------------------------------------------------------- ElasticNet
def enet_solve(A, b, a1, a2, iters=6000):
"""Cyclic coordinate descent for min ||A th - b||^2 + a1|th|_1 + a2|th|^2."""
n, d = A.shape
th = np.zeros(d)
col = (A ** 2).sum(0)
r = -b.copy()
for _ in range(iters):
mx = 0.0
for k in range(d):
r -= A[:, k] * th[k]
rho = -2.0 * A[:, k] @ r
den = 2.0 * col[k] + 2.0 * a2
new = np.sign(rho) * max(abs(rho) - a1, 0.0) / den
mx = max(mx, abs(new - th[k]))
th[k] = new
r += A[:, k] * th[k]
if mx < 1e-14:
break
return th
def cramer_residual(A, b, a1, a2, th, tol=1e-9):
"""On the active set S with signs s, stationarity says
(2 A_S'A_S + 2 a2 I) th_S = 2 A_S'b - a1 s, whose solution is a rational
function of (a1, a2) by Cramer's rule. Return the residual norm."""
S = np.where(np.abs(th) > tol)[0]
if S.size == 0:
return 0.0, 0
As = A[:, S]
G = 2.0 * As.T @ As + 2.0 * a2 * np.eye(S.size)
rhs = 2.0 * As.T @ b - a1 * np.sign(th[S])
return float(np.linalg.norm(G @ th[S] - rhs)), int(S.size)
def group_lasso_solve(A, b, alphas, groups, iters=40000):
"""Proximal gradient for min ||A th - b||^2 + sum_i alpha_i ||th_{G_i}||_2."""
d = A.shape[1]
th = np.zeros(d)
Lc = 2.0 * np.linalg.norm(A, 2) ** 2
for _ in range(iters):
g = 2.0 * A.T @ (A @ th - b)
z = th - g / Lc
new = z.copy()
for gi, G in enumerate(groups):
nz = np.linalg.norm(z[G])
sc = max(0.0, 1.0 - alphas[gi] / (Lc * max(nz, 1e-300)))
new[G] = z[G] * sc
if np.max(np.abs(new - th)) < 1e-15:
th = new
break
th = new
return th
def section_D():
out = {}
# ---- D1 Assumption 7.1: is the ElasticNet path piecewise rational?
d, n = 6, 25
A = RNG.standard_normal((n, d)); b = RNG.standard_normal(n)
res, sizes = [], []
grid = [(a1, a2) for a1 in np.linspace(0.05, 20.0, 40)
for a2 in np.linspace(0.05, 5.0, 40)]
for a1, a2 in grid:
th = enet_solve(A, b, a1, a2)
r, sz = cramer_residual(A, b, a1, a2, th)
res.append(r); sizes.append(sz)
out["enet"] = dict(solves=len(grid), max_cramer_residual=float(np.max(res)),
mean_cramer_residual=float(np.mean(res)),
active_sizes=[int(min(sizes)), int(max(sizes))])
# rational vs equal-capacity polynomial fit inside one active-set region
a1f = 3.0
pts = []
for a2 in np.linspace(0.20, 3.00, 60):
th = enet_solve(A, b, a1f, a2)
pts.append((a2, th))
S0 = tuple(np.where(np.abs(pts[0][1]) > 1e-9)[0])
same = [(a, t) for a, t in pts if tuple(np.where(np.abs(t) > 1e-9)[0]) == S0]
xa = np.array([a for a, _ in same]); ya = np.array([t[S0[0]] for _, t in same])
# inside a fixed active set S the exact solution is
# th_S(a2) = (G0 + 2 a2 I)^{-1} r , a ratio of a degree-(|S|-1) numerator
# to a degree-|S| denominator in a2. Fit that form, and an
# equal-parameter-count polynomial, and compare residuals.
nS = len(S0)
xs_ = (xa - xa.mean()) / xa.std()
Mr = np.column_stack([xs_ ** k for k in range(nS)] +
[-(xs_ ** k) * ya for k in range(1, nS + 1)])
cr, *_ = np.linalg.lstsq(Mr, ya, rcond=None)
num = sum(cr[k] * xs_ ** k for k in range(nS))
den = 1.0 + sum(cr[nS + k - 1] * xs_ ** k for k in range(1, nS + 1))
fit_r = num / den
npar = Mr.shape[1]
Mp = np.column_stack([xs_ ** k for k in range(npar)])
cp, *_ = np.linalg.lstsq(Mp, ya, rcond=None)
fit_p = Mp @ cp
out["enet_fit"] = dict(region_points=len(same),
rational_rel_residual=float(np.linalg.norm(fit_r - ya) /
max(np.linalg.norm(ya), 1e-30)),
polynomial_rel_residual=float(np.linalg.norm(fit_p - ya) /
max(np.linalg.norm(ya), 1e-30)),
active_set_size=nS, free_parameters_each=int(npar))
# control: group LASSO violates the same stationarity identity
groups = [np.arange(0, 3), np.arange(3, 6)]
viol = []
for a in np.linspace(0.5, 3.0, 12):
th = group_lasso_solve(A, b, np.array([a, a]), groups)
S = np.where(np.abs(th) > 1e-9)[0]
if S.size == 0:
continue
As = A[:, S]
G = 2.0 * As.T @ As
rhs = 2.0 * As.T @ b - a * np.sign(th[S])
viol.append(float(np.linalg.norm(G @ th[S] - rhs)))
out["group_lasso_control"] = dict(n=len(viol), max_violation=float(np.max(viol)),
min_violation=float(np.min(viol)))
# ---- D2 Theorem 8.1 premise: ||v||_2 is not piecewise polynomial
NS = 6000
v = RNG.standard_normal((NS, 4))
nrm = np.linalg.norm(v, axis=1)
sq = nrm ** 2
feats = []
for deg in range(9):
for e in itertools.combinations_with_replacement(range(4), deg):
c = np.ones(NS)
for i in e:
c = c * v[:, i]
feats.append(c)
F = np.column_stack(feats)
def relres(y):
c, *_ = np.linalg.lstsq(F, y, rcond=None)
return float(np.linalg.norm(F @ c - y) / np.linalg.norm(y))
out["nonpoly"] = dict(deg8_residual_norm=relres(nrm),
deg8_residual_normsq=relres(sq),
n_features=F.shape[1])
# nu-encoding on real proximal-gradient solves
enc = []
for _ in range(48):
Ai = RNG.standard_normal((30, 8)); bi = RNG.standard_normal(30)
gs = [np.arange(0, 4), np.arange(4, 8)]
al = RNG.uniform(0.3, 3.0, 2)
th = group_lasso_solve(Ai, bi, al, gs)
nu = np.array([np.linalg.norm(th[G]) for G in gs])
e1 = max(abs(nu[i] ** 2 - float(th[gs[i]] @ th[gs[i]])) for i in range(2))
# KKT residual of the group-LASSO stationarity condition
gsm = 2.0 * Ai.T @ (Ai @ th - bi)
kkt = 0.0
for i, G in enumerate(gs):
if nu[i] > 1e-10:
kkt = max(kkt, float(np.linalg.norm(gsm[G] + al[i] * th[G] / nu[i])))
else:
kkt = max(kkt, max(0.0, float(np.linalg.norm(gsm[G])) - al[i]))
enc.append((e1, kkt, float(nu.min())))
out["nu_encoding"] = dict(solves=len(enc),
max_nu_sq_error=float(max(e[0] for e in enc)),
max_kkt_residual=float(max(e[1] for e in enc)),
min_nu=float(min(e[2] for e in enc)))
# ---- D3 Proposition G.1 for weighted fused LASSO
def fused(y, w):
d = y.size
D = np.zeros((d - 1, d))
for i in range(d - 1):
D[i, i] = -1.0; D[i, i + 1] = 1.0
H = D @ D.T
u = np.zeros(d - 1)
Lc = np.linalg.norm(H, 2)
for _ in range(30000):
g = H @ u - D @ y
un = np.clip(u - g / Lc, -w, w)
if np.max(np.abs(un - u)) < 1e-15:
u = un; break
u = un
th = y - D.T @ u
prim = 0.5 * float((th - y) @ (th - y)) + float(w @ np.abs(D @ th))
dual = -0.5 * float(u @ H @ u) + float(u @ (D @ y))
return th, u, prim, dual, H
gaps, eigs, regions = [], [], []
for d in [4, 6, 8, 10, 12]:
y = RNG.standard_normal(d)
seen = set()
for _ in range(400):
w = RNG.uniform(0.05, 1.5, d - 1)
th, u, prim, dual, H = fused(y, w)
gaps.append(abs(prim - dual))
seen.add(tuple(np.sign(np.round(np.diff(th), 9)).astype(int)))
eigs.append(float(np.min(np.linalg.eigvalsh(H))))
regions.append(dict(d=d, regions=len(seen), cap=3 ** (d - 1)))
out["fused"] = dict(max_duality_gap=float(np.max(gaps)),
min_dual_hessian_eig=float(np.min(eigs)),
regions=regions,
all_under_cap=all(r["regions"] < r["cap"] for r in regions))
# piecewise-affine check: within one region theta*(w) is affine in w
y = RNG.standard_normal(8)
base = RNG.uniform(0.4, 0.6, 7)
th0, u0, *_ = fused(y, base)
sgn0 = tuple(np.sign(np.round(np.diff(th0), 9)).astype(int))
W, T = [], []
for _ in range(120):
w = base + RNG.uniform(-0.02, 0.02, 7)
th, u, *_ = fused(y, w)
if tuple(np.sign(np.round(np.diff(th), 9)).astype(int)) == sgn0:
W.append(np.concatenate([[1.0], w])); T.append(th)
W = np.array(W); T = np.array(T)
c, *_ = np.linalg.lstsq(W, T, rcond=None)
inreg = float(np.linalg.norm(W @ c - T) / max(np.linalg.norm(T), 1e-30))
W2, T2 = [], []
for _ in range(160):
w = RNG.uniform(0.05, 1.5, 7)
th, u, *_ = fused(y, w)
W2.append(np.concatenate([[1.0], w])); T2.append(th)
W2 = np.array(W2); T2 = np.array(T2)
c2, *_ = np.linalg.lstsq(W2, T2, rcond=None)
across = float(np.linalg.norm(W2 @ c2 - T2) / max(np.linalg.norm(T2), 1e-30))
out["fused_affine"] = dict(in_region_points=len(T), in_region_rel_residual=inreg,
across_region_points=len(T2),
across_region_rel_residual=across,
separation=across / max(inreg, 1e-300))
# rank-deficient control breaks Prop G.1's precondition
Dbad = np.zeros((3, 4)); Dbad[0] = [-1, 1, 0, 0]; Dbad[1] = [-1, 1, 0, 0]; Dbad[2] = [0, 0, -1, 1]
Hb = Dbad @ Dbad.T
out["fused_control"] = dict(min_eig_rank_deficient=float(np.min(np.linalg.eigvalsh(Hb))),
rank=int(np.linalg.matrix_rank(Dbad)), rows=3)
return out
# --------------------------------------------------- E: real tuning classes
def make_instance(n, d, p, seed):
r = np.random.default_rng(seed)
A = r.standard_normal((n, d)); b = r.standard_normal(n)
Ap = r.standard_normal((n, d)); bp = r.standard_normal(n)
groups = np.array_split(np.arange(d), p)
return A, b, Ap, bp, groups
def ridge_theta(A, b, alpha_vec):
G = A.T @ A + np.diag(alpha_vec)
return np.linalg.solve(G, A.T @ b)
def section_E():
out = {}
# ---- f != g check on a real bi-level ridge instance
A, b, Ap, bp, groups = make_instance(40, 8, 4, 11)
al = np.array([0.7, 1.3, 0.2, 2.1])
av = np.zeros(8)
for gi, G in enumerate(groups):
av[G] = al[gi]
th = ridge_theta(A, b, av)
gf = 2.0 * (A.T @ (A @ th - b) + av * th)
gg = 2.0 * Ap.T @ (Ap @ th - bp)
out["bilevel_fneqg"] = dict(train_stationarity=float(np.linalg.norm(gf)),
val_gradient_at_same_point=float(np.linalg.norm(gg)))
# ---- certified pseudo-dimension lower bounds on real tuning classes
def pdim_of(loss, xs, ts, alphas, kmax):
return measure_pdim_pool(loss, xs, ts, alphas, nmax=kmax)
rows = []
for (p, d) in [(2, 8), (3, 12), (4, 16), (6, 24), (8, 32)]:
A, b, Ap, bp, groups = make_instance(60, d, p, 100 + p * 7 + d)
insts = [make_instance(60, d, p, 500 + p * 31 + d * 5 + k) for k in range(p + 4)]
def loss_bi(alpha, inst):
Ai, bi, Api, bpi, gi = inst
av = np.zeros(d)
for k, G in enumerate(gi):
av[G] = alpha[k]
t = ridge_theta(Ai, bi, av)
return float(np.sum((Api @ t - bpi) ** 2))
def loss_single(alpha, inst):
Ai, bi, Api, bpi, gi = inst
av = np.zeros(d)
for k, G in enumerate(gi):
av[G] = alpha[k]
t = ridge_theta(Ai, bi, av)
return float(np.sum((Ai @ t - bi) ** 2) + av @ (t * t))
pool = np.exp(RNG.uniform(math.log(1e-3), math.log(1e3), (40000, p)))
ts_bi = [np.median([loss_bi(a, inst) for a in pool[:400]]) for inst in insts]
ts_si = [np.median([loss_single(a, inst) for a in pool[:400]]) for inst in insts]
kb, _ = pdim_of(loss_bi, insts, ts_bi, pool, min(p + 3, 12))
ks, _ = pdim_of(loss_single, insts, ts_si, pool, min(p + 3, 12))
Lb = np.array([[loss_bi(a, x) for x in insts] for a in pool[:8000]])
Sb = (Lb >= np.array(ts_bi)[None, :]).astype(np.int8)
Ls = np.array([[loss_single(a, x) for x in insts] for a in pool[:8000]])
Ss = (Ls >= np.array(ts_si)[None, :]).astype(np.int8)
cells_bi = len(set(map(tuple, Sb.tolist())))
cells_si = len(set(map(tuple, Ss.tolist())))
Mtot, Dtot = 6 * d + 64, 4
ub61 = stated_bound(p, (d, d), Mtot, Dtot)
ub51 = stated_bound(p, (d,), 3 * d + 64 + d, 4)
rows.append(dict(p=p, d=d, pdim_lb_bilevel=kb, pdim_lb_single=ks,
sign_cells_bilevel=cells_bi, sign_cells_single=cells_si,
cell_ratio=cells_bi / max(cells_si, 1),
thm61_bound=ub61, thm51_bound=ub51,
lb_under_bound=bool(kb <= ub61)))
print("E p=%d d=%d pdim_lb bi=%d single=%d cells %d/%d bound61=%.1f"
% (p, d, kb, ks, cells_bi, cells_si, ub61), flush=True)
out["pdim_rows"] = rows
# ---- precondition control: a non-semi-algebraic (sinusoidal) class
xs = [np.array([1.0])] * 12
pool1 = RNG.uniform(0.0, 200.0, (200000, 1))
def sinloss(a, x):
return float(np.sin(a[0] * (1.0 + 0.37 * x[0])))
freqs = [np.array([1.0 + 0.31 * k]) for k in range(12)]
def sinloss2(a, x):
return float(np.sin(a[0] * x[0]))
k_sin, _ = measure_pdim_pool(sinloss2, freqs, [0.0] * 12, pool1, nmax=12)
def linloss(a, x):
return float(a[0] * x[0])
k_lin, _ = measure_pdim_pool(linloss, freqs, [0.0] * 12, pool1, nmax=12)
out["precondition_control"] = dict(
p=1, sinusoidal_pdim_lb=k_sin, affine_pdim_lb=k_lin,
note="a single-parameter semi-algebraic class has Pdim 1; sin(alpha x) is not "
"semi-algebraic and shatters far more points, so Theorem 4.1's polynomial-FOL "
"hypothesis is doing real work")
return out
if __name__ == "__main__":
which = sys.argv[1] if len(sys.argv) > 1 else "de"
out = {}
if "d" in which:
t = time.time(); out["D"] = section_D(); out["D"]["secs"] = round(time.time() - t, 1)
print("D", json.dumps(out["D"])[:1500], flush=True)
if "e" in which:
t = time.time(); out["E"] = section_E(); out["E"]["secs"] = round(time.time() - t, 1)
print("E", json.dumps(out["E"])[:1200], flush=True)
with open("evidence_%s.json" % which, "w") as f:
json.dump(out, f, indent=1)
|