File size: 6,533 Bytes
35f2be5 | 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 | """Exactness checks for the SPDE formulation -- the machinery behind Claim 2.
These are *analytic* comparisons (not Monte Carlo): the SPDE propagation is
linear-Gaussian, so the joint law it induces over the field at future times can
be written in closed form and compared to the dense GP built from k_tHelm.
"""
import jax
import jax.numpy as jnp
import numpy as np
jax.config.update("jax_enable_x64", True)
from ballast.gp import posterior_ext_state
from ballast.kernels import HelmParams, k_helm_mat, k_thelm_mat
from ballast.spde import make_ops, temporal_matrices
P = HelmParams(
phi_ls=0.8, phi_var=0.5, psi_ls=0.5, psi_var=0.5, time_ls=2.5, time_var=1.0
)
def _grid(n=3):
xs = jnp.linspace(-1.0, 1.0, n)
X, Y = jnp.meshgrid(xs, xs, indexing="ij")
return jnp.stack([X.reshape(-1), Y.reshape(-1)], -1)
def _full_matrices(R, p, dt):
"""Explicit Phi_full = I_{2N} (x) Phi and Q_full = K_space (x) Q for the
flattened state layout (i, c, b) -> i*4 + c*2 + b."""
phi, q, pinf = temporal_matrices(p.time_ls, p.time_var, dt)
Ks = k_helm_mat(R, R, p)
n2 = Ks.shape[0]
return (
jnp.kron(jnp.eye(n2), phi),
jnp.kron(Ks, q),
jnp.kron(Ks, pinf),
)
def _spde_joint_cov(R, p, dt, n_steps, Sigma0, mu0):
"""Joint mean/cov of the f-component at steps 0..n_steps under the SPDE.
Propagates the Gaussian analytically:
mu_{k+1} = Phi_full mu_k
Var_{k+1} = Phi_full Var_k Phi_full^T + Q_full
Cov(X_j,X_i) = Phi_full^{j-i} Var_i (j > i)
"""
Phi_f, Q_f, _ = _full_matrices(R, p, dt)
d = Phi_f.shape[0]
mus = [mu0]
vars_ = [Sigma0]
for _ in range(n_steps):
mus.append(Phi_f @ mus[-1])
vars_.append(Phi_f @ vars_[-1] @ Phi_f.T + Q_f)
nT = n_steps + 1
C = jnp.zeros((nT * d, nT * d))
for i in range(nT):
for j in range(nT):
if j >= i:
blk = jnp.linalg.matrix_power(Phi_f, j - i) @ vars_[i]
C = C.at[j * d : (j + 1) * d, i * d : (i + 1) * d].set(blk)
else:
blk = vars_[j] @ jnp.linalg.matrix_power(Phi_f, i - j).T
C = C.at[j * d : (j + 1) * d, i * d : (i + 1) * d].set(blk)
mu = jnp.concatenate(mus)
# select the f-component (b = 0) of every (location, velocity component)
sel = jnp.arange(d).reshape(-1, 2)[:, 0]
sel_all = jnp.concatenate([sel + k * d for k in range(nT)])
return mu[sel_all], C[jnp.ix_(sel_all, sel_all)]
def test_prior_matches_dense_gp():
"""SPDE prior started at equilibrium reproduces k_tHelm exactly."""
R = _grid(3)
dt, n_steps = 0.37, 4
_, _, Pinf_full = _full_matrices(R, P, dt)
d = Pinf_full.shape[0]
mu, C = _spde_joint_cov(R, P, dt, n_steps, Pinf_full, jnp.zeros(d))
times = jnp.arange(n_steps + 1) * dt
Rr = jnp.tile(R, (n_steps + 1, 1))
tr = jnp.repeat(times, R.shape[0])
K = k_thelm_mat(Rr, tr, Rr, tr, P)
np.testing.assert_allclose(mu, 0.0, atol=1e-12)
np.testing.assert_allclose(C, K, rtol=1e-8, atol=1e-10)
def test_posterior_sampling_is_exact_with_nongridded_observations():
"""The Sec. 4.1 scheme is exact: extended-state posterior at t_m + SPDE
propagation == dense GP posterior at future times, even though the
observations sit at non-gridded (Lagrangian) locations.
This is the core correctness claim of BALLAST's sampler, and the reason it
can avoid filtering over the observation locations.
"""
key = jax.random.PRNGKey(0)
R = _grid(3)
sigma = 0.1
dt, n_steps, t_m = 0.37, 4, 1.5
# observations at random NON-grid locations, all strictly before t_m
k1, k2, k3 = jax.random.split(key, 3)
S_obs = jax.random.uniform(k1, (7, 2), minval=-1.4, maxval=1.4)
t_obs = jax.random.uniform(k2, (7,), minval=0.0, maxval=t_m)
y_obs = jax.random.normal(k3, (7, 2))
# --- SPDE route: posterior of the extended state at t_m, then propagate
mean, chol = posterior_ext_state(S_obs, t_obs, y_obs, R, t_m, P, sigma, jitter=0.0)
Sigma0 = chol @ chol.T
mu_s, C_s = _spde_joint_cov(R, P, dt, n_steps, Sigma0, mean)
# --- dense route: GP posterior directly at R x future times
times = t_m + jnp.arange(n_steps + 1) * dt
Rr = jnp.tile(R, (n_steps + 1, 1))
tr = jnp.repeat(times, R.shape[0])
K_oo = k_thelm_mat(S_obs, t_obs, S_obs, t_obs, P) + sigma**2 * jnp.eye(14)
K_ot = k_thelm_mat(S_obs, t_obs, Rr, tr, P)
K_tt = k_thelm_mat(Rr, tr, Rr, tr, P)
sol = jnp.linalg.solve(K_oo, K_ot)
mu_d = sol.T @ y_obs.reshape(-1)
C_d = K_tt - K_ot.T @ sol
np.testing.assert_allclose(mu_s, mu_d, rtol=1e-6, atol=1e-9)
np.testing.assert_allclose(C_s, C_d, rtol=1e-6, atol=1e-9)
def test_sampler_empirical_moments():
"""End-to-end: the actual sampling code path reproduces the analytic posterior."""
from ballast.gp import sample_ext_state
from ballast.spde import propagate
key = jax.random.PRNGKey(2)
R = _grid(2)
sigma, dt, n_steps, t_m = 0.1, 0.25, 3, 1.0
ops = make_ops(R, P, dt, jitter=0.0)
k1, k2, k3 = jax.random.split(key, 3)
S_obs = jax.random.uniform(k1, (5, 2), minval=-1.0, maxval=1.0)
t_obs = jax.random.uniform(k2, (5,), minval=0.0, maxval=t_m)
y_obs = jax.random.normal(k3, (5, 2))
mean, chol = posterior_ext_state(S_obs, t_obs, y_obs, R, t_m, P, sigma, jitter=1e-12)
n = 40000
keys = jax.random.split(jax.random.PRNGKey(7), n)
def one(k):
ka, kb = jax.random.split(k)
X0 = sample_ext_state(ka, mean, chol, R.shape[0])
return propagate(X0, kb, ops, n_steps)[-1].reshape(-1)
draws = jax.vmap(one)(keys)
emp_mu = draws.mean(0)
emp_cov = jnp.cov(draws.T)
_, C_an = _spde_joint_cov(
R, P, dt, n_steps, chol @ chol.T, mean
)
mu_an, _ = _spde_joint_cov(R, P, dt, n_steps, chol @ chol.T, mean)
d = 2 * R.shape[0]
mu_last = mu_an[-d:]
C_last = C_an[-d:, -d:]
# Statistical tolerances: fixed rtol/atol are meaningless here because the
# small off-diagonal covariance entries are dominated by Monte Carlo noise.
se_mu = jnp.sqrt(jnp.diag(C_last) / n)
assert jnp.all(jnp.abs(emp_mu - mu_last) < 5 * se_mu), "sample mean off"
# SE of an empirical covariance entry: sqrt((C_ii C_jj + C_ij^2)/n)
d_ = jnp.diag(C_last)
se_cov = jnp.sqrt((d_[:, None] * d_[None, :] + C_last**2) / n)
z = jnp.abs(emp_cov - C_last) / se_cov
assert z.max() < 5.0, f"max z-score {z.max():.2f} between empirical and analytic cov"
|