ccevolve/tx_workspace/ac1_alphaevolve/solution_v12.py

"""
L-BFGS-B with JAX analytical gradients.
Many random restarts. Fast convergence.
"""
import numpy as np
from scipy.optimize import minimize
import jax
import jax.numpy as jnp


def compute_c1_numpy(f_values, n_points):
    dx = 0.5 / n_points
    f_nn = np.maximum(f_values, 0.0)
    autoconv = np.convolve(f_nn, f_nn, mode='full') * dx
    integral_sq = (np.sum(f_nn) * dx) ** 2
    if integral_sq < 1e-12:
        return 1e10
    return np.max(autoconv) / integral_sq


def make_objective(N, temp):
    dx = 0.5 / N

    @jax.jit
    def obj_and_grad(params):
        f = jnp.exp(jnp.clip(params, -8, 4))
        padded = jnp.zeros(2 * N)
        padded = padded.at[:N].set(f)
        fft_f = jnp.fft.rfft(padded)
        conv = jnp.fft.irfft(fft_f * fft_f, n=2 * N) * dx
        integral_sq = (jnp.sum(f) * dx) ** 2
        smooth_max = jax.nn.logsumexp(temp * conv) / temp
        return smooth_max / integral_sq

    value_and_grad = jax.jit(jax.value_and_grad(obj_and_grad))

    def scipy_wrapper(params_np):
        p = jnp.array(params_np)
        v, g = value_and_grad(p)
        return float(v), np.array(g, dtype=np.float64)

    return scipy_wrapper


def run():
    best_c1 = float('inf')
    best_f = None
    best_n = None

    for N in [1000, 2000, 4000]:
        dx = 0.5 / N
        print(f"\n=== N={N} ===")

        # Generate many initializations
        inits = []
        x = np.linspace(0, 1, N)

        for seed in range(30):
            np.random.seed(seed)
            if seed == 0:
                f = np.ones(N)
            elif seed == 1:
                f = 0.5 + 0.5 * np.cos(2*np.pi*x)  # U-shape
            elif seed == 2:
                f = np.exp(-10*(x-0.5)**2) + 0.1  # Gaussian
            elif seed == 3:
                f = np.exp(-5*(x-0.3)**2) + 0.05
            elif seed == 4:
                f = np.exp(-5*(x-0.7)**2) + 0.05
            elif seed == 5:
                f = 0.2 + 0.8 * np.cos(np.pi*x)**2  # centered bump
            elif seed == 6:
                f = 0.5 + 0.3*np.cos(2*np.pi*x) + 0.1*np.cos(4*np.pi*x)
            elif seed == 7:
                f = np.maximum(1 - 4*np.abs(x-0.5), 0) + 0.1  # triangle
            elif seed == 8:
                f = 0.3 + 0.7*x  # linear increasing
            elif seed == 9:
                f = 0.3 + 0.7*(1-x)  # linear decreasing
            elif seed == 10:
                f = 1 + 0.5*np.cos(6*np.pi*x)  # wavy
            elif seed == 11:
                f = np.where((x > 0.15) & (x < 0.85), 1.0, 0.3)  # wide box
            elif seed == 12:
                f = 0.5 + 0.3*np.cos(2*np.pi*x) + 0.15*np.cos(4*np.pi*x) + 0.05*np.cos(6*np.pi*x)
            else:
                f = np.abs(np.random.randn(N)) * 0.3 + 0.2

            inits.append(np.maximum(f, 0.01))

        # Run L-BFGS-B with increasing temperature
        for init_idx, init_f in enumerate(inits):
            params = np.log(np.maximum(init_f, 1e-6))

            # Progressive temperature increase
            for temp in [50.0, 200.0, 1000.0, 5000.0, 20000.0]:
                obj_fn = make_objective(N, temp)
                result = minimize(
                    obj_fn, params, method='L-BFGS-B', jac=True,
                    options={'maxiter': 3000, 'ftol': 1e-15, 'gtol': 1e-12},
                )
                params = result.x

            f_opt = np.exp(np.clip(params, -8, 4))
            c1 = compute_c1_numpy(f_opt, N)

            if c1 < best_c1:
                best_c1 = c1
                best_f = f_opt
                best_n = N
                print(f"  init={init_idx}: C1={c1:.10f} ***", flush=True)
            elif init_idx < 13:
                print(f"  init={init_idx}: C1={c1:.10f}", flush=True)

    print(f"\nFinal best C1: {best_c1:.10f}")
    return best_f, best_c1, best_c1, best_n