scripts/exp00_selfcheck.py

#!/usr/bin/env python3
"""exp00 self-check: builds/exercises core/{klinf,popquant,seqtest,par}.py and
writes results/exp00_selfcheck.json per specs/exp00_core.md.

    python3 scripts/exp00_selfcheck.py [--toy]

No experiment-specific logic lives here; this only validates the shared core.
"""
import argparse
import json
import os
import sys
import time

import numpy as np
from scipy.optimize import minimize_scalar

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

from core import par
from core.klinf import klinf_emp, klinf_emp_generic, klinf_path
from core.popquant import pop, beta_phi_closed_form
from core.seqtest import run_path

RESULTS_DIR = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), "results")


def log(msg):
    print(f"[exp00] {msg}", file=sys.stderr, flush=True)


# --------------------------------------------------------------------- primal
def _primal_maximize(x, m_o, T_max=40.0, xatol=1e-13):
    """Independent maximizer of F(lam)=mean(log(1-lam(x-m_o))) using a
    log-distance-from-boundary substitution so the search stays well
    conditioned right up to the singular endpoint. Cross-checks the dual
    (Brent-on-Psi) solver in klinf_emp without sharing any code path."""
    xbar = x.mean()
    singular_at_hi = m_o > xbar
    if singular_at_hi:
        lo, hi = 0.0, 1.0 / (1.0 - m_o)
    else:
        lo, hi = -1.0 / m_o, 0.0
    span = hi - lo

    def lam_of_t(t):
        if singular_at_hi:
            return hi - np.exp(-t) * span
        return lo + np.exp(-t) * span

    def negF(t):
        lam = lam_of_t(t)
        denom = 1.0 - lam * (x - m_o)
        if np.any(denom <= 0):
            return 1e18
        return -float(np.mean(np.log(denom)))

    res = minimize_scalar(negF, bounds=(0.0, T_max), method="bounded", options=dict(xatol=xatol))
    return -res.fun


# ---------------------------------------------------------------- trial funcs
def _dual_vs_primal_trial(seed):
    rng = np.random.default_rng(seed)
    n = int(rng.integers(5, 80))
    a, b = rng.uniform(0.3, 6), rng.uniform(0.3, 6)
    x = rng.beta(a, b, size=n)
    m_o = rng.uniform(0.02, 0.98)
    if x.mean() == m_o:
        return 0.0
    v_dual, _, _ = klinf_emp(x, m_o, use_fastpath=False)
    v_primal = _primal_maximize(x, m_o)
    return abs(v_dual - v_primal)


def _fastpath_trial(seed):
    rng = np.random.default_rng(seed)
    p = rng.uniform(0.0, 1.0)
    n = int(rng.integers(5, 300))
    m_o = rng.uniform(0.02, 0.98)
    x = (rng.uniform(size=n) < p).astype(float)
    if x.mean() == m_o:
        return 0.0
    v_fast, lam_fast, _ = klinf_emp(x, m_o, use_fastpath=True)
    v_gen, lam_gen, _ = klinf_emp_generic(x, m_o)
    return max(abs(v_fast - v_gen), abs(lam_fast - lam_gen))


def _path_trial(seed, n):
    rng = np.random.default_rng(seed)
    a, b = rng.uniform(0.5, 5), rng.uniform(0.5, 5)
    x = rng.beta(a, b, size=n)
    m_o = rng.uniform(0.1, 0.9)
    path = klinf_path(x, m_o)
    pointwise = np.array([klinf_emp(x[:k], m_o)[0] for k in range(1, n + 1)])
    return float(np.max(np.abs(path - pointwise)))


def _ht_violation_count(seed, path_len):
    """eq (11)'s bound KL_inf <= 1/(1-m_o) is a fact about the '+' branch
    (m_o >= running mean): pointwise, log(1-lam(x-m_o)) <= log(1+lam*m_o) <=
    log(1/(1-m_o)) < 1/(1-m_o) for x in [0,1], lam in [0,lam_bar]. The
    mirrored '-' branch (m_o < running mean) obeys the mirrored bound
    1/m_o by the same argument. We check each step against whichever bound
    matches its branch."""
    rng = np.random.default_rng(seed)
    dist_kind = rng.choice(["beta", "bernoulli"])
    m_o = rng.uniform(0.05, 0.95)
    if dist_kind == "beta":
        a, b = rng.uniform(0.5, 5), rng.uniform(0.5, 5)
        x = rng.beta(a, b, size=path_len)
    else:
        p = rng.uniform(0.05, 0.95)
        x = (rng.uniform(size=path_len) < p).astype(float)
    path = klinf_path(x, m_o)
    running_mean = np.cumsum(x) / np.arange(1, path_len + 1)
    branch_plus = m_o >= running_mean
    bound = np.where(branch_plus, 1.0 / (1.0 - m_o), 1.0 / m_o)
    return int(np.sum(path > bound + 1e-9))


def _beta_phi_trial(seed):
    rng = np.random.default_rng(seed)
    a, b = rng.uniform(1.05, 8), rng.uniform(1.05, 8)
    mean = a / (a + b)
    m_o = rng.uniform(mean, 0.995)  # ensure '+' branch (m_o >= mean)
    r = pop(("beta", a, b), m_o)
    cf = beta_phi_closed_form(a, b, m_o)
    if r["phi"] is None or not np.isfinite(cf) or cf == 0:
        return 0.0
    return abs(r["phi"] - cf) / abs(cf)


# ------------------------------------------------------------- determinism
def _determinism_replicate(rng, r):
    n = 50 + r
    x = rng.beta(2.0, 3.0, size=n)
    m_o = 0.3 + 0.05 * (r % 5)
    val, lam, case = klinf_emp(x, m_o)
    return (val, lam, case)


def _check_determinism(n_reps, config_id=777):
    r1 = par.make_parallel(n_jobs=1)(
        __import__("joblib").delayed(_determinism_replicate)(par.rng_for(config_id, r), r)
        for r in range(n_reps)
    )
    r2 = par.make_parallel(n_jobs=min(2, os.cpu_count() or 1))(
        __import__("joblib").delayed(_determinism_replicate)(par.rng_for(config_id, r), r)
        for r in range(n_reps)
    )
    return r1 == r2


# --------------------------------------------------------------------- main
def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--toy", action="store_true")
    args = ap.parse_args()
    toy = args.toy

    n_jobs = par.n_cores()
    log(f"starting exp00 self-check (toy={toy}), JOB_CORES-derived n_jobs={n_jobs}")

    if toy:
        n_dual_primal, n_fastpath, n_path, n_ht, n_beta_phi, n_det = 20, 60, 6, 8, 12, 8
        path_n, ht_path_len = 60, 300
    else:
        n_dual_primal, n_fastpath, n_path, n_ht, n_beta_phi, n_det = 300, 1500, 30, 60, 150, 16
        path_n, ht_path_len = 400, 3000

    t0 = time.time()
    parallel = par.make_parallel(n_jobs=n_jobs)
    joblib_delayed = __import__("joblib").delayed

    log(f"running {n_dual_primal} dual-vs-primal trials")
    dual_vs_primal_errs = parallel(joblib_delayed(_dual_vs_primal_trial)(i) for i in range(n_dual_primal))
    dual_vs_primal_max_abs_err = float(max(dual_vs_primal_errs))
    log(f"  max abs err = {dual_vs_primal_max_abs_err:.3e}")

    log(f"running {n_fastpath} Bernoulli fast-path trials")
    fastpath_errs = parallel(joblib_delayed(_fastpath_trial)(i + 10_000) for i in range(n_fastpath))
    bernoulli_fastpath_max_abs_err = float(max(fastpath_errs))
    log(f"  max abs err = {bernoulli_fastpath_max_abs_err:.3e}")

    log(f"running {n_path} klinf_path-vs-pointwise trials (n={path_n})")
    path_errs = parallel(joblib_delayed(_path_trial)(i + 20_000, path_n) for i in range(n_path))
    klinf_path_vs_pointwise_max_abs_err = float(max(path_errs))
    log(f"  max abs err = {klinf_path_vs_pointwise_max_abs_err:.3e}")

    log(f"running {n_ht} Honda-Takemura bound checks (path_len={ht_path_len})")
    ht_counts = parallel(joblib_delayed(_ht_violation_count)(i + 30_000, ht_path_len) for i in range(n_ht))
    ht_bound_violations = int(sum(ht_counts))
    log(f"  violations = {ht_bound_violations}")

    log(f"running {n_beta_phi} Beta Phi closed-form checks")
    beta_phi_errs = parallel(joblib_delayed(_beta_phi_trial)(i + 40_000) for i in range(n_beta_phi))
    beta_phi_closed_form_max_rel_err = float(max(beta_phi_errs))
    log(f"  max rel err = {beta_phi_closed_form_max_rel_err:.3e}")

    log(f"checking determinism across JOB_CORES (1 vs 2) on {n_det} replicates")
    determinism_jobcores_1_vs_2 = bool(_check_determinism(n_det))
    log(f"  bit-identical = {determinism_jobcores_1_vs_2}")

    log("smoke-testing run_path (raises on any internal invariant violation)")
    def _run_path_smoke(seed):
        rng = np.random.default_rng(seed)
        sampler = lambda rng_, size: rng_.beta(2, 3, size=size)
        run_path(rng, sampler, 0.5, [0.05, 0.1], "theory", n_max=2000 if not toy else 500)
        return True
    n_smoke = 6 if toy else 20
    smoke_results = parallel(joblib_delayed(_run_path_smoke)(i + 50_000) for i in range(n_smoke))
    assert all(smoke_results), "run_path invariant assertions failed on smoke paths"
    log(f"  {n_smoke} paths ran clean")

    log("computing canonical case labels")
    cases = {
        "beta_3_2_mo_0.7": pop(("beta", 3, 2), 0.7)["case"],
        "bern_0.6_mo_0.7": pop(("bernoulli", 0.6), 0.7)["case"],
        "bern_0.6_mo_0.2": pop(("bernoulli", 0.6), 0.2)["case"],
    }
    log(f"  cases = {cases}")

    out = {
        "dual_vs_primal_max_abs_err": dual_vs_primal_max_abs_err,
        "bernoulli_fastpath_max_abs_err": bernoulli_fastpath_max_abs_err,
        "klinf_path_vs_pointwise_max_abs_err": klinf_path_vs_pointwise_max_abs_err,
        "ht_bound_violations": ht_bound_violations,
        "beta_phi_closed_form_max_rel_err": beta_phi_closed_form_max_rel_err,
        "determinism_jobcores_1_vs_2": determinism_jobcores_1_vs_2,
        "cases": cases,
    }

    os.makedirs(RESULTS_DIR, exist_ok=True)
    out_path = os.path.join(RESULTS_DIR, "exp00_selfcheck.json")
    with open(out_path, "w") as f:
        json.dump(out, f, indent=2)

    elapsed = time.time() - t0
    print(f"exp00 self-check complete in {elapsed:.1f}s -> {out_path}")
    print(f"  dual_vs_primal_max_abs_err          = {dual_vs_primal_max_abs_err:.3e}")
    print(f"  bernoulli_fastpath_max_abs_err       = {bernoulli_fastpath_max_abs_err:.3e}")
    print(f"  klinf_path_vs_pointwise_max_abs_err  = {klinf_path_vs_pointwise_max_abs_err:.3e}")
    print(f"  ht_bound_violations                  = {ht_bound_violations}")
    print(f"  beta_phi_closed_form_max_rel_err     = {beta_phi_closed_form_max_rel_err:.3e}")
    print(f"  determinism_jobcores_1_vs_2           = {determinism_jobcores_1_vs_2}")
    print(f"  cases                                 = {cases}")


if __name__ == "__main__":
    main()