File size: 10,168 Bytes
2facf1f | 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 | """
Mathematical constructions for low C1.
Key insight: we need functions where the sumset A+A is uniformly distributed.
This is a problem in additive combinatorics.
Constructions to try:
1. Quadratic residue indicator functions
2. Functions based on polynomial evaluation over finite fields
3. B-spline wavelets
4. Optimized step functions via DE at small K
5. Functions that equalize the autoconvolution
"""
import numpy as np
from scipy.optimize import minimize, differential_evolution
from scipy.interpolate import interp1d
import time
def compute_c1_fft(f_values, dx):
f = np.maximum(f_values, 0.0)
N = len(f)
M = 2 * N
fft_f = np.fft.rfft(f, n=M)
conv = np.fft.irfft(fft_f * fft_f, n=M) * dx
integral_sq = (np.sum(f) * dx) ** 2
if integral_sq < 1e-20:
return 1e10
return float(np.max(conv) / integral_sq)
def compute_c1_smooth_and_grad(f, N, dx, alpha=200.0):
M = 2 * N
fft_f = np.fft.rfft(f, n=M)
conv = np.fft.irfft(fft_f * fft_f, n=M) * dx
integral = np.sum(f) * dx
if integral < 1e-15:
return 1e10, np.zeros(N)
integral_sq = integral ** 2
max_val = np.max(conv)
shifted = conv - max_val
mask = shifted > -50.0 / alpha
weights = np.zeros_like(conv)
weights[mask] = np.exp(alpha * shifted[mask])
sum_w = np.sum(weights)
if sum_w < 1e-30:
weights[np.argmax(conv)] = 1.0
sum_w = 1.0
smooth_max = max_val + np.log(sum_w) / alpha
softmax_w = weights / sum_w
c1 = smooth_max / integral_sq
fft_sw = np.fft.rfft(softmax_w, n=M)
fft_fp = np.fft.rfft(f, n=M)
corr = np.fft.irfft(fft_sw * np.conj(fft_fp), n=M)
grad_f = 2.0 * corr[:N] * dx / integral_sq - 2.0 * smooth_max * dx / (integral**3)
return c1, grad_f
def opt_pipeline(f_init, N, dx, maxiter=2000):
"""Full optimization pipeline."""
params = np.sqrt(np.maximum(f_init, 0.0) + 1e-12)
for alpha in [0.5, 5.0, 50.0, 500.0, 5000.0, 50000.0]:
def obj(p, a=alpha):
f = p ** 2
c1, g = compute_c1_smooth_and_grad(f, N, dx, a)
return c1, g * 2 * p
result = minimize(obj, params, jac=True, method='L-BFGS-B',
options={'maxiter': maxiter, 'ftol': 1e-16, 'gtol': 1e-15})
params = result.x
f_out = params ** 2
return f_out, compute_c1_fft(f_out, dx)
def upscale(f, N_new):
N_old = len(f)
x_old = np.linspace(0, 1, N_old)
x_new = np.linspace(0, 1, N_new)
interp = interp1d(x_old, f, kind='linear', fill_value=0.0, bounds_error=False)
return np.maximum(interp(x_new), 0.0)
# =====================================================
# Construction 1: Differential evolution on step heights
# =====================================================
print("=== DE on step functions ===")
t0 = time.time()
best_results = []
for K in [10, 15, 20, 25, 30, 40, 50, 75, 100]:
N_k = K * 10 # Use higher resolution per step
dx_k = 0.5 / N_k
step_size = N_k // K
def obj_de(heights):
f = np.repeat(np.maximum(heights, 0.0), step_size)[:N_k]
return compute_c1_fft(f, dx_k)
bounds = [(0.0, 5.0)] * K
result = differential_evolution(obj_de, bounds, maxiter=500, seed=42,
popsize=20, tol=1e-12)
c1_raw = result.fun
# Refine with gradient optimization
f_step = np.repeat(np.maximum(result.x, 0.0), step_size)[:N_k]
f_opt, c1_opt = opt_pipeline(f_step, N_k, dx_k, maxiter=1000)
best_results.append((K, c1_opt, f_opt, N_k, dx_k))
elapsed = time.time() - t0
print(f" K={K:3d}: DE C1={c1_raw:.8f}, opt C1={c1_opt:.8f} ({elapsed:.0f}s)")
# =====================================================
# Construction 2: Functions from finite field arithmetic
# =====================================================
print("\n=== Finite field constructions ===")
# For prime p, define f(x) = indicator of {x^2 mod p : x in [0,p)} scaled to [0, 0.5]
for p in [31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97, 101, 127, 131]:
N_p = p * 10
dx_p = 0.5 / N_p
# Quadratic residues
qr = set()
for a in range(p):
qr.add((a * a) % p)
f = np.zeros(N_p)
for r in qr:
block_start = int(r * N_p / p)
block_end = int((r + 1) * N_p / p)
f[block_start:block_end] = 1.0
f_opt, c1 = opt_pipeline(f, N_p, dx_p, maxiter=1000)
if c1 < 1.515:
best_results.append(('QR', c1, f_opt, N_p, dx_p))
print(f" p={p}: C1 = {c1:.8f} ***")
# =====================================================
# Construction 3: Autoconvolution equalization
# Iteratively adjust f to make autoconvolution more uniform
# =====================================================
print("\n=== Autoconvolution equalization ===")
for N_eq in [200, 500]:
dx_eq = 0.5 / N_eq
# Start from various initializations
for seed in range(20):
np.random.seed(seed + 3000)
f = np.random.exponential(1.0, N_eq) + 0.1
# Iterative equalization
for iteration in range(50):
f_nn = np.maximum(f, 0.0)
M = 2 * N_eq
fft_f = np.fft.rfft(f_nn, n=M)
conv = np.fft.irfft(fft_f * fft_f, n=M) * dx_eq
# Identify high points in autoconvolution
max_val = np.max(conv)
threshold = max_val * 0.99
high_mask = conv > threshold
# The autoconvolution at position t is sum_j f[j]*f[t-j]*dx
# To reduce it, we need to reduce f at positions that contribute
# This is related to the gradient
# Simple approach: multiply f by a correction factor
# based on how much each position contributes to the max
power_spectrum = np.abs(fft_f) ** 2
# Reduce the highest frequency components
cutoff = len(power_spectrum) // 2
damping = np.ones(len(power_spectrum))
top_freqs = np.argsort(power_spectrum[1:cutoff])[::-1][:10] + 1
damping[top_freqs] *= 0.95
fft_mod = fft_f * damping
f = np.fft.irfft(fft_mod, n=M)[:N_eq]
f = np.maximum(f, 0.0) + 0.001
f_opt, c1 = opt_pipeline(f, N_eq, dx_eq, maxiter=1000)
if c1 < 1.515:
best_results.append(('EQ', c1, f_opt, N_eq, dx_eq))
print(f" N={N_eq}, seed={seed}: C1 = {c1:.8f}")
# =====================================================
# Construction 4: Optimized B-spline basis
# =====================================================
print("\n=== B-spline basis ===")
for n_knots in [5, 10, 15, 20, 30]:
N_bs = 500
dx_bs = 0.5 / N_bs
x = np.linspace(0, 0.5, N_bs, endpoint=False)
# B-spline basis with n_knots uniformly spaced knots
knots = np.linspace(0, 0.5, n_knots + 2)
basis = np.zeros((N_bs, n_knots))
for i in range(n_knots):
# Triangular basis function
center = knots[i + 1]
width = knots[i + 2] - knots[i]
basis[:, i] = np.maximum(1 - np.abs(x - center) / (width/2), 0)
def obj_bs(coeffs):
f = np.maximum(basis @ coeffs, 0.0)
return compute_c1_fft(f, dx_bs)
for seed in range(20):
np.random.seed(seed + 4000)
c0 = np.random.exponential(1.0, n_knots)
result = minimize(obj_bs, c0, method='Nelder-Mead',
options={'maxiter': 10000, 'xatol': 1e-10, 'fatol': 1e-10})
c1 = result.fun
if c1 < 1.515:
f_bs = np.maximum(basis @ result.x, 0.0)
f_opt, c1_opt = opt_pipeline(f_bs, N_bs, dx_bs, maxiter=1000)
best_results.append(('BS', c1_opt, f_opt, N_bs, dx_bs))
print(f" n_knots={n_knots}, seed={seed}: C1 = {c1:.4f} → {c1_opt:.8f}")
# =====================================================
# Sort all results and upscale the best to N=5000
# =====================================================
# Filter to tuples with the right structure
valid = [(r[1], r[2], r[3], r[4]) for r in best_results if isinstance(r[0], (int, str))]
valid.sort(key=lambda x: x[0])
print(f"\n=== Top 5 results ===")
for i, (c1, f, N_r, dx_r) in enumerate(valid[:5]):
print(f" #{i+1}: C1 = {c1:.10f} at N={N_r}")
if valid:
best_c1, best_f, best_N, best_dx = valid[0]
# Upscale to N=5000 and refine
print(f"\nUpscaling best (C1={best_c1:.10f}) to N=5000...")
N_final = 5000
dx_final = 0.5 / N_final
f_up = upscale(best_f, N_final)
f_final, c1_final = opt_pipeline(f_up, N_final, dx_final, maxiter=5000)
# Alpha cycling
params = np.sqrt(np.maximum(f_final, 0.0) + 1e-12)
for cycle in range(5):
for alpha in [0.5, 2.0, 10.0]:
def obj(p, a=alpha):
f = p ** 2
c1, g = compute_c1_smooth_and_grad(f, N_final, dx_final, a)
return c1, g * 2 * p
result = minimize(obj, params, jac=True, method='L-BFGS-B',
options={'maxiter': 500, 'ftol': 1e-16, 'gtol': 1e-15})
params = result.x
for alpha in [100.0, 1000.0, 10000.0, 100000.0]:
def obj(p, a=alpha):
f = p ** 2
c1, g = compute_c1_smooth_and_grad(f, N_final, dx_final, a)
return c1, g * 2 * p
result = minimize(obj, params, jac=True, method='L-BFGS-B',
options={'maxiter': 2000, 'ftol': 1e-16, 'gtol': 1e-15})
params = result.x
f_out = params ** 2
c1 = compute_c1_fft(f_out, dx_final)
if c1 < c1_final:
c1_final = c1
f_final = f_out.copy()
print(f"Final at N=5000: C1 = {c1_final:.10f}")
print(f"Score: {1.5052939684401607 / c1_final:.10f}")
# Compare with existing best
try:
f_existing = np.load('/workspace/best_f_5000.npy')
c1_existing = compute_c1_fft(f_existing, 0.5/len(f_existing))
print(f"Existing best: C1 = {c1_existing:.10f}")
if c1_final < c1_existing:
np.save('/workspace/best_f_5000_math.npy', f_final)
print("NEW BEST! Saved to best_f_5000_math.npy")
except:
pass
print(f"\nTotal time: {time.time()-t0:.0f}s")
|