File size: 7,293 Bytes
5e4510c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Disable progress bar for cleaner output logs
import os

os.environ["TQDM_DISABLE"] = "1"

# EVOLVE-BLOCK-START
import jax
import jax.numpy as jnp
import optax
import numpy as np
from dataclasses import dataclass
from scipy.special import hermite
import tqdm


@dataclass
class Hyperparameters:
    learning_rate: float = 0.001
    num_steps: int = 100000
    num_restarts: int = 20
    num_hermite_coeffs: int = 4  # uses H0, H4, H8, H12


class UncertaintyOptimizer:
    """
    Finds coefficients for a generalized Hermite polynomial P(x) that minimize
    the largest positive root, providing an upper bound for C4.
    """

    def __init__(self, hypers: Hyperparameters):
        self.hypers = hypers
        self.degrees = [4 * k for k in range(hypers.num_hermite_coeffs)]
        max_degree = self.degrees[-1]
        hermite_polys = [hermite(d) for d in self.degrees]

        basis = []
        for poly in hermite_polys:
            pad_amount = max_degree - poly.order
            basis.append(jnp.array(np.pad(poly.coef, (pad_amount, 0))))
        self.hermite_basis = jnp.stack(basis)

        self.H_vals_at_zero = jnp.array([p(0) for p in hermite_polys])
        self.x_grid = jnp.linspace(0.0, 10.0, 3000)
        self.optimizer = optax.adam(self.hypers.learning_rate)

    @staticmethod
    def _objective_fn(params: jnp.ndarray, hermite_basis, H_vals_at_zero, x_grid):
        """Penalize negative values of P(x) on [0, Xmax]; mild weighting to emphasize larger x."""
        c_others, log_c_last = params[:-1], params[-1]
        c_last = jnp.exp(log_c_last)

        # Enforce P(0) = 0
        c0 = (
            -(jnp.sum(c_others * H_vals_at_zero[1:-1]) + c_last * H_vals_at_zero[-1])
            / H_vals_at_zero[0]
        )
        hermite_coeffs = jnp.concatenate([jnp.array([c0]), c_others, jnp.array([c_last])])

        poly_coeffs_std = jnp.sum(hermite_coeffs[:, None] * hermite_basis, axis=0)
        p_values = jnp.polyval(poly_coeffs_std, x_grid)

        # Slightly increasing weight toward the right end of the interval
        weights = 1.0 + (x_grid / (x_grid[-1] + 1e-12))
        loss = jnp.sum(weights * jax.nn.relu(-p_values))
        return loss

    @staticmethod
    def train_step(
        params: jnp.ndarray,
        opt_state: optax.OptState,
        optimizer,
        hermite_basis,
        H_vals_at_zero,
        x_grid,
    ):
        loss, grads = jax.value_and_grad(UncertaintyOptimizer._objective_fn)(
            params, hermite_basis, H_vals_at_zero, x_grid
        )
        updates, opt_state = optimizer.update(grads, opt_state, params)
        params = optax.apply_updates(params, updates)
        return params, opt_state, loss


def run_single_trial(optimizer: UncertaintyOptimizer, key: jax.random.PRNGKey):
    """Runs one full optimization from a near-good starting point with small noise."""
    num_params_to_opt = optimizer.hypers.num_hermite_coeffs - 1  # = 3 when using H0,H4,H8,H12
    assert num_params_to_opt == 3, "This initialization assumes num_hermite_coeffs == 4."

    base_c1 = -0.01158510802599293
    base_c2 = -8.921606035407065e-05
    base_log_c_last = np.log(1e-6)

    base = jnp.array([base_c1, base_c2, base_log_c_last], dtype=jnp.float32)
    noise = jax.random.normal(key, (num_params_to_opt,)) * 1e-3
    params = base + noise

    opt_state = optimizer.optimizer.init(params)
    jit_train_step = jax.jit(UncertaintyOptimizer.train_step, static_argnums=(2,))

    for _ in range(optimizer.hypers.num_steps):
        params, opt_state, _ = jit_train_step(
            params,
            opt_state,
            optimizer.optimizer,
            optimizer.hermite_basis,
            optimizer.H_vals_at_zero,
            optimizer.x_grid,
        )
    return params


def _build_P_from_hermite_coeffs(hermite_coeffs: np.ndarray, degrees: list[int]) -> np.poly1d:
    """Build monomial-basis polynomial P from Hermite-basis coefficients."""
    max_degree = degrees[-1]
    hermite_polys = [hermite(d) for d in degrees]

    P_poly_coeffs = np.zeros(max_degree + 1)
    for i, c in enumerate(hermite_coeffs):
        poly = hermite_polys[i]
        pad_amount = max_degree - poly.order
        P_poly_coeffs[pad_amount:] += c * poly.coef

    if P_poly_coeffs[0] < 0:
        P_poly_coeffs = -P_poly_coeffs
        hermite_coeffs[:] = -hermite_coeffs
    return np.poly1d(P_poly_coeffs)


def _c4_from_hermite_coeffs(hermite_coeffs: np.ndarray, num_hermite_coeffs: int):
    """Compute r_max and C4 from full Hermite coefficient vector (Google-style)."""
    degrees = [4 * k for k in range(num_hermite_coeffs)]
    P = _build_P_from_hermite_coeffs(hermite_coeffs.copy(), degrees)

    # Divide by x^2
    Q, R = np.polydiv(P, np.poly1d([1.0, 0.0, 0.0]))
    if np.max(np.abs(R.c)) > 1e-10:
        return None, None

    roots = Q.r
    real_pos = roots[(np.isreal(roots)) & (roots.real > 0)].real
    if real_pos.size == 0:
        return None, None

    # Tiny sign-change check around candidates
    r_candidates = np.sort(real_pos)
    r_max = None
    for r in r_candidates:
        eps = 1e-10 * max(1.0, abs(r))
        left = np.polyval(Q, r - eps)
        right = np.polyval(Q, r + eps)
        if left * right < 0:
            r_max = float(r)
    if r_max is None:
        r_max = float(r_candidates[-1])

    c4 = (r_max**2) / (2 * np.pi)
    return c4, r_max


def get_c4_from_params(params: np.ndarray, hypers: Hyperparameters):
    """Calculates the precise C4 bound from a final set of parameters."""
    c_others, log_c_last = params[:-1], params[-1]
    c_last = np.exp(log_c_last)

    degrees = [4 * k for k in range(hypers.num_hermite_coeffs)]
    hermite_polys = [hermite(d) for d in degrees]
    H_vals_at_zero = np.array([p(0) for p in hermite_polys])

    # Enforce P(0) = 0
    c0 = (
        -(np.sum(c_others * H_vals_at_zero[1:-1]) + c_last * H_vals_at_zero[-1]) / H_vals_at_zero[0]
    )
    hermite_coeffs = np.concatenate([[c0], np.array(c_others), [c_last]])

    c4, rmax = _c4_from_hermite_coeffs(hermite_coeffs, hypers.num_hermite_coeffs)
    if c4 is None:
        return None, None, None
    return hermite_coeffs, c4, rmax


def run():
    hypers = Hyperparameters()
    optimizer = UncertaintyOptimizer(hypers)
    main_key = jax.random.PRNGKey(42)
    best_c4_bound = float("inf")
    best_coeffs, best_r_max = None, None

    print(f"Running {hypers.num_restarts} trials to find the best C4 upper bound...")
    for _ in tqdm.tqdm(range(hypers.num_restarts), desc="Searching", disable=True):
        main_key, restart_key = jax.random.split(main_key)
        final_params = run_single_trial(optimizer, restart_key)

        coeffs, c4_bound, r_max = get_c4_from_params(np.array(final_params), hypers)

        if c4_bound is not None and c4_bound < best_c4_bound:
            best_c4_bound = c4_bound
            best_coeffs = coeffs
            best_r_max = r_max

    if best_coeffs is None:
        raise RuntimeError("Failed to find a valid solution in any restart.")

    print("\nSearch complete.")
    print(f"Best Hermite coeffs: {best_coeffs}")
    print(f"Best largest positive root r_max: {best_r_max:.8f}")
    print(f"Resulting best C4 upper bound: {best_c4_bound:.8f}")

    return best_coeffs, best_c4_bound, best_r_max


# EVOLVE-BLOCK-END