src/run_gauge_revision.py

"""Reproduce the v3 finite experiments. These are synthetic post-conversion
electrical models, not a calibrated molecular simulator or a hardware benchmark.
"""
import os
os.environ.setdefault('OPENBLAS_NUM_THREADS', '1')
os.environ.setdefault('OMP_NUM_THREADS', '1')
from pathlib import Path
import argparse
import json
import time
import numpy as np
from scipy.linalg import cho_factor, cho_solve
from scipy.sparse.linalg import spsolve
from scipy.integrate import quad
from contract_genome import capsule, compile_grid, seal_ready
from contract_growth import delivery_time
from contracts import laplacian, kron, conductance
from gauge_contracts import (RelativeEstimator, comparison_line_graph,
    grid_edges, ratio_observation, exact_feasibility, robust_feasibility,
    uniform_reserve_yield, projective_response_error,
    closure_bridge_is_needed)

ROOT = Path(__file__).resolve().parents[1]
OUT = ROOT/'results/gauge'
METHODS = ['projective', 'common_detector_gain', 'common_material_scale',
           'conventional_joint', 'fixed_representative', 'insufficient_reserve',
           'differential_bias', 'early_reference_release']
PARAMETERS = {'tau': .04, 'seal_log_bound': .005, 'diagnostic_log_radius': .006,
    'ratio_sigma': .012, 'delta': .005, 'horizon': 128,
    'increment_min': .007, 'increment_max': .012,
    'reserve': 1.05, 'initial_low': .65, 'initial_high': 1.35,
    'root_material_rate': 2., 'hop_time': .03, 'sample_time': .001,
    'material_per_conductance': .4, 'scaffold_per_node': .1}
PARAMETERS['numerical_log_budget'] = 1e-8


def prepare(n, dim):
    genome = json.loads((ROOT/f'genomes/gauge_{dim}d.json').read_text())
    if (genome['extent'] != n or genome['dimension'] != dim or
        genome['functional_equivalence'] != 'positive_common_conductance_scale' or
        genome['external_ports'] != 4 or genome['motif'] != 'parity_conductance'):
        raise ValueError('The requested grid and declared function capsule disagree')
    expected = {'log_tolerance': PARAMETERS['tau'],
                'seal_log_bound': PARAMETERS['seal_log_bound'],
                'reserve_capacity': PARAMETERS['reserve'],
                'ratio_log_radius': PARAMETERS['diagnostic_log_radius'],
                'inspection_horizon': PARAMETERS['horizon'],
                'increment_bounds': [PARAMETERS['increment_min'], PARAMETERS['increment_max']],
                'witness': 'stable_uncalibrated',
                'closure': 'preserve_comparisons_until_all_certified'}
    if any(genome[key] != value for key, value in expected.items()):
        raise ValueError('Update the recorded experiment parameters with the capsule')
    net = compile_grid(genome)
    count, edges = comparison_line_graph(net['edges'])
    estimator = RelativeEstimator(count, edges)
    N = len(net['coordinates'])
    ports = np.array([0, N-1, n-1, (n-1)*n**(dim-1)])
    D0 = kron(laplacian(N, net['edges'], net['target']), ports)
    return net, estimator, ports, D0


def voltage_solution(net, g):
    N = len(net['coordinates'])
    L = laplacian(N, net['edges'], g)
    fixed = np.r_[net['left'], net['right']]
    free = np.setdiff1d(np.arange(N), fixed)
    v = np.zeros(N); v[net['left']] = 1.
    v[free] = spsolve(L[free][:, free], -L[free][:, fixed]@v[fixed])
    return v


def run_one(prepared, method, seed):
    net, estimator, ports, D0 = prepared
    par = PARAMETERS
    R, N = len(net['target']), len(net['coordinates'])
    rng = np.random.default_rng(seed)
    rng_noise = np.random.default_rng(seed+11000000)
    rng_action = np.random.default_rng(seed+22000000)
    rng_gain = np.random.default_rng(seed+33000000)
    x0 = rng.uniform(par['initial_low'], par['initial_high'], R)
    x = x0.copy()
    capacity = .4 if method == 'insufficient_reserve' else par['reserve']
    physical_scale = .6 if method == 'common_material_scale' else 1.
    target_aug = np.r_[net['target'], 1.]
    samples = estimator.sample_budget(par['ratio_sigma'],
        par['diagnostic_log_radius']-par['numerical_log_budget'], par['horizon'], par['delta'])
    radius = estimator.radius(par['ratio_sigma'], samples,
        par['horizon'], par['delta'])+par['numerical_log_budget']
    bias = np.zeros(R+1)
    if method == 'differential_bias':
        positions = net['coordinates'][net['owner'], 0]
        bias[:-1] = .30*(positions/positions.max()-.5)
    accepted = np.zeros(R, bool)
    added = np.zeros(R)
    observations = 0; ratio_samples = 0; largest_error = 0.
    cumulative_cycle_energy = 0.; deposition_operations = 0
    maximum_numerical_bound = 0.
    ledger_time = 0.; messages = 0
    colors = int(estimator.colors.max())+1
    trace = []
    # The geometry is the inherited seed-grown lattice. Initial and repair
    # material are paid using the same capacity reservation model as v2.
    initial_mass = par['scaffold_per_node']*N
    initial_mass += par['material_per_conductance']*float(net['target']@x0)
    demand = np.full(N, par['scaffold_per_node'])
    np.add.at(demand, net['owner'],
              par['material_per_conductance']*net['target']*x0)
    growth_time = 0.
    for layer in range(int(net['depth'].max())+1):
        service, _ = delivery_time(np.where(net['depth']==layer, demand, 0.),
            net['parent'], par['root_material_rate'], par['hop_time'],
            np.minimum(net['depth'], layer))
        growth_time += max(1., service)
    ledger_time += growth_time

    def inspect():
        nonlocal observations, ratio_samples, largest_error
        nonlocal cumulative_cycle_energy, ledger_time
        nonlocal maximum_numerical_bound
        gains = np.ones(len(estimator.edges))
        if method == 'common_detector_gain':
            gains = np.exp(rng_gain.uniform(-.7, .7, len(gains)))
        noise = rng_noise.normal(0, par['ratio_sigma']/np.sqrt(samples),
                                len(estimator.edges))
        y = ratio_observation(np.r_[x, 1.], target_aug, estimator.edges,
            gains, noise, physical_scale=physical_scale, node_bias=bias)
        z = estimator.estimate(y)
        maximum_numerical_bound = max(maximum_numerical_bound,estimator.last_numerical_bound)
        if estimator.last_numerical_bound > par['numerical_log_budget']:
            raise FloatingPointError('The numerical inference budget is not certified')
        largest_error = max(largest_error,
                           float(np.max(np.abs(z[:-1]-np.log(x)))))
        residual = y-estimator.B@z
        cumulative_cycle_energy += float(residual@residual)
        observations += 1; ratio_samples += samples*len(estimator.edges)
        ledger_time += colors*samples*par['sample_time']
        return z[:-1]

    estimate = inspect()
    cert = robust_feasibility(np.exp(estimate-radius), np.exp(estimate+radius),
        capacity, par['tau'], par['seal_log_bound'], radius,
        par['increment_max'])
    kappa = cert.get('scale_lower', 1.)
    reason = cert['reason']
    if method == 'fixed_representative':
        kappa = 1.
        cert['feasible'] = bool(cert.get('scale_lower', 2.) <= 1 <=
                                cert.get('scale_upper', 0.))
        if not cert['feasible']:
            reason = 'fixed_scale_not_reachable'
    steps = 0
    if cert['feasible']:
        half_band = par['tau']-par['seal_log_bound']
        threshold_low = np.log(kappa)-half_band+radius
        threshold_high = np.log(kappa)+half_band-radius
        for turn in range(par['horizon']):
            if turn:
                if method == 'early_reference_release' and turn == 3:
                    ref_edges = np.flatnonzero(
                        np.any(estimator.edges==estimator.reference, axis=1))
                    assert len(ref_edges)==1
                    assert closure_bridge_is_needed(estimator.nodes,
                                estimator.edges, ref_edges[0])
                    reason = 'reference_access_lost'
                    break
                estimate = inspect()
            accepted |= (estimate >= threshold_low) & (estimate <= threshold_high)
            if accepted.all():
                reason = 'complete'
                break
            need = (~accepted) & (estimate < threshold_low)
            if np.any((~accepted) & (estimate > threshold_high)):
                reason = 'unrepairable_high_state'
                break
            delta = rng_action.uniform(par['increment_min'],
                                       par['increment_max'], R)
            delta[~need] = 0.
            if np.any(added+delta > capacity+1e-12):
                reason = 'reserve_exhausted'
                break
            demand = np.zeros(N)
            np.add.at(demand, net['owner'],
                      par['material_per_conductance']*net['target']*delta)
            service, _ = delivery_time(demand, net['parent'],
                par['root_material_rate'], par['hop_time'], net['depth'])
            ledger_time += max(1., service)
            x += delta; added += delta
            steps += 1; deposition_operations += int(need.sum())
            # Max/min reductions use the same service tree; their latency is
            # not hidden inside the following physical-time subtotal.
            messages += 2*(N-1)
            trace.append([observations, int(accepted.sum()), float(added.sum())])
    complete = bool(accepted.all())
    sealed = np.zeros(N, bool)
    closure_rounds = 0
    final_x = x.copy()
    if complete:
        # The calibration backbone is retained until ALL certificates exist.
        # Only then may the ordinary local postorder closure proceed.
        for _ in range(N+1):
            sealed = seal_ready(np.ones(N, bool), sealed, net['children'])
            closure_rounds += 1
            if sealed.all():
                break
        assert sealed.all()
        final_x *= np.exp(rng_action.uniform(-par['seal_log_bound'],
                                             par['seal_log_bound'], R))
        ledger_time += closure_rounds*par['hop_time']
    final_g = physical_scale*net['target']*final_x
    D = kron(laplacian(N, net['edges'], final_g), ports)
    projective_error = projective_response_error(D, D0)
    local_contract_error = float(np.max(np.abs(np.log(final_x/kappa))))
    v0 = voltage_solution(net, net['target'])
    vf = voltage_solution(net, final_g)
    final_G = conductance(N, net['edges'], final_g, net['left'], net['right'])
    target_G = conductance(N, net['edges'], net['target'], net['left'], net['right'])
    outcome = {'seed': int(seed), 'method': method, 'nodes': N, 'modules': R,
        'dimension': net['coordinates'].shape[1], 'complete': complete,
        'termination_reason': reason, 'projective_error': projective_error,
        'projective_function_pass': bool(complete and projective_error<=par['tau']),
        'local_certificate_valid': bool(complete and local_contract_error<=par['tau']+1e-10),
        'false_local_certificate': bool(complete and local_contract_error>par['tau']+1e-10),
        'local_log_error': local_contract_error,
        'voltage_max_absolute_error': float(np.max(np.abs(vf-v0))),
        'absolute_conductance_ratio': float(final_G/target_G),
        'physical_common_scale': physical_scale, 'chosen_relative_scale': float(kappa),
        'samples_per_pair': samples, 'pair_edges': len(estimator.edges),
        'probe_color_slots': colors, 'rho_max': estimator.rho_max,
        'diagnostic_radius': radius, 'inspections': observations,
        'maximum_numerical_log_error_bound': maximum_numerical_bound,
        'ratio_samples': ratio_samples, 'maximum_actual_log_estimation_error': largest_error,
        'mean_cycle_residual_energy': cumulative_cycle_energy/observations,
        'repair_steps': steps, 'deposition_operations': deposition_operations,
        'additional_material_proxy': float(par['material_per_conductance']*net['target']@added),
        'initial_material_proxy': float(initial_mass), 'reserve_capacity': float(capacity),
        'max_reserve_used': float(added.max()), 'growth_time': growth_time,
        'service_and_acquisition_time_subtotal': ledger_time,
        'inference_latency_excluded_from_subtotal': True,
        'service_reduction_messages': messages, 'closure_rounds': closure_rounds,
        'guard_certificate': cert}
    snapshot = {'initial': x0, 'final': final_x, 'target': net['target'],
                'coordinates': net['coordinates'], 'edges': net['edges'],
                'trace': np.array(trace), 'voltage': vf, 'target_voltage': v0,
                'comparison_edges': estimator.edges}
    return outcome, snapshot


def phase_experiment():
    rng = np.random.default_rng(39060919)
    low, high, tau = .65, 1.35, .04
    rows = []
    for size in [4, 16, 64, 256, 1024]:
        x = rng.uniform(low, high, (5000, size))
        minimum, maximum = x.min(axis=1), x.max(axis=1)
        for cap in np.linspace(.4, .7, 31):
            good = maximum <= np.exp(2*tau)*(minimum+cap)
            probability = uniform_reserve_yield(size, low, high, float(cap), tau)
            rows.append({'modules': size, 'capacity': float(cap), 'trials': 5000,
                         'passes': int(good.sum()), 'exact_probability': probability})
    return {'low': low, 'high': high, 'tau': tau,
            'critical_capacity': float(high*np.exp(-2*tau)-low), 'rows': rows,
            'sampling_note': 'Within each size, the same 5000 arrays are reused across capacities.'}


def theorem_experiments():
    rng = np.random.default_rng(39160919)
    maximum_quadrature_error = 0.
    for _ in range(100):
        R = int(rng.integers(2, 60)); low = float(rng.uniform(.1, .8))
        high = low+float(rng.uniform(.2, 1.5)); tau = float(rng.uniform(.003, .12))
        c = float(rng.uniform(0, high*np.exp(-2*tau)-low)) if high*np.exp(-2*tau)>low else 0.
        q = np.exp(2*tau); width = high-low
        point = high/q-c
        f = lambda t: R/width*(max(0., min(high, q*(t+c))-t)/width)**(R-1)
        integration = quad(f, low, high, points=[point] if low<point<high else [],
                           epsabs=1e-11)[0]
        maximum_quadrature_error = max(maximum_quadrature_error,
            abs(integration-uniform_reserve_yield(R,low,high,c,tau)))
    # Positive-scale freedom is nontrivial only for >=3 external ports.
    from contracts import relative_spectrum
    max_composition_excess = -1e9
    for _ in range(200):
        coords, edges = grid_edges(4, 3)
        target = rng.uniform(.5, 1.5, len(edges))
        tau = .04; common = float(np.exp(rng.uniform(-2,2)))
        actual = common*target*np.exp(rng.uniform(-tau,tau,len(edges)))
        D0 = kron(laplacian(len(coords),edges,target),[0,3,48,63])
        D = kron(laplacian(len(coords),edges,actual),[0,3,48,63])
        max_composition_excess = max(max_composition_excess,
                                   projective_response_error(D,D0)-tau)
    return {'quadrature_trials': 100, 'max_phase_formula_error': maximum_quadrature_error,
            'composition_trials': 200, 'maximum_projective_bound_excess': max_composition_excess}


def calibration_experiments():
    records = []
    for dim, sizes in [(1,[8,16,32,64,128]),(2,[4,8,12,16]),(3,[3,4,6,8])]:
        for n in sizes:
            coords, edges = grid_edges(n,dim)
            est = RelativeEstimator(len(coords),edges,reference=0)
            R = len(coords)-1
            records.append({'dimension': dim,'side':n,'nodes':len(coords),
                'rho_max':est.rho_max,'radius_at_64_samples':
                est.radius(.012,64,128,.005),
                'samples_for_radius_006':est.sample_budget(.012,.006,128,.005)})
    local_checks = []
    for n,dim in [(4,2),(4,3),(6,3)]:
        coords, edges = grid_edges(n,dim)
        est=RelativeEstimator(len(coords),edges,reference=0)
        rng=np.random.default_rng(39260919+dim*100+n)
        truth=rng.normal(0,.2,len(coords)); truth-=truth[0]
        y=est.B@truth+rng.normal(0,.002,len(edges))
        exact=est.estimate(y)
        local,info=est.local_estimate(y,tolerance=1e-5)
        info.update({'side':n,'dimension':dim,'nodes':len(coords),
                     'actual_error':float(np.max(np.abs(local-exact)))})
        local_checks.append(info)
    coords,edges=grid_edges(5,3); est=RelativeEstimator(len(coords),edges,reference=0)
    rng=np.random.default_rng(39360919)
    noise=rng.normal(size=(len(edges),6000))
    errors=cho_solve(est.factor,est.Bg.T@noise)
    empirical=errors.var(axis=1,ddof=1)
    covariance_error=float(np.max(np.abs(empirical/est.rho-1)))
    truth=rng.normal(0,.2,len(coords));truth-=truth[0]
    invisible=.15*coords[:,0]/4
    y=est.B@(truth+invisible)
    fitted=est.estimate(y)
    return {'scaling':records,'local_message_solver':local_checks,
        'covariance_trials':6000,'covariance_max_relative_sampling_error':covariance_error,
        'gradient_bias_example':{'cycle_residual_norm':float(np.linalg.norm(y-est.B@fitted)),
                                'maximum_undetected_bias':float(np.max(np.abs(fitted-truth)))}}


def main():
    parser=argparse.ArgumentParser()
    parser.add_argument('--reps',type=int,default=32)
    parser.add_argument('--skip-theory',action='store_true')
    args=parser.parse_args()
    OUT.mkdir(parents=True,exist_ok=True)
    start=time.monotonic()
    rows=[]
    for n,dim in [(8,2),(5,3)]:
        prepared=prepare(n,dim)
        for method in METHODS:
            for rep in range(args.reps):
                row,snapshot=run_one(prepared,method,39460919+1000*dim+rep)
                rows.append(row)
                if rep==0: np.savez_compressed(OUT/f'{dim}d_{method}.npz',**snapshot)
            recent=rows[-args.reps:]
            print(dim,method,sum(x['complete'] for x in recent),
                  sum(x['projective_function_pass'] for x in recent),flush=True)
    summary=[]
    for dim in [2,3]:
        for method in METHODS:
            subset=[r for r in rows if r['dimension']==dim and r['method']==method]
            summary.append({'dimension':dim,'method':method,'runs':len(subset),
                'completed':sum(r['complete'] for r in subset),
                'functional':sum(r['projective_function_pass'] for r in subset),
                'false_certificates':sum(r['false_local_certificate'] for r in subset),
                **{key+'_mean':float(np.mean([r[key] for r in subset])) for key in [
                    'projective_error','voltage_max_absolute_error','absolute_conductance_ratio',
                    'additional_material_proxy','inspections','ratio_samples',
                    'service_and_acquisition_time_subtotal','mean_cycle_residual_energy']}})
    (OUT/'runs.json').write_text(json.dumps(rows,indent=2)+'\n')
    (OUT/'summary.json').write_text(json.dumps(summary,indent=2)+'\n')
    (OUT/'parameters.json').write_text(json.dumps(PARAMETERS,indent=2)+'\n')
    if not args.skip_theory:
        for name,result in [('phase',phase_experiment()),('theorem_checks',theorem_experiments()),
                            ('calibration',calibration_experiments())]:
            (OUT/(name+'.json')).write_text(json.dumps(result,indent=2)+'\n')
    (OUT/'execution.json').write_text(json.dumps({'wall_seconds':time.monotonic()-start,
        'runs':len(rows),'no_laboratory_measurements':True},indent=2)+'\n')
    print('Completed',len(rows),'runs in',round(time.monotonic()-start,2),'seconds')


if __name__=='__main__':
    main()