matter-embryogenesis / tests /test_gauge.py
PureOne's picture
Release Matter Embryogenesis v3.0.0: theory, code, data and audit
52fc221 verified
Raw
History Blame
6.77 kB
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]/'src'))
import unittest
import numpy as np
from scipy.optimize import linprog
from scipy.linalg import cho_solve
from gauge_contracts import (RelativeEstimator, grid_edges, incidence,
comparison_line_graph, exact_feasibility, robust_feasibility,
uniform_reserve_yield, ratio_observation, projective_response_error,
closure_bridge_is_needed)
from contracts import laplacian, kron
class GaugeTheoryTests(unittest.TestCase):
def test_exact_reachability_and_minimum_material_against_linear_program(self):
rng=np.random.default_rng(811)
for _ in range(80):
count=int(rng.integers(2,15))
x=rng.uniform(.2,2.,count);c=rng.uniform(0,1.5,count)
tau=float(rng.uniform(.005,.15));cost=rng.uniform(.2,3.,count)
A=np.zeros((2*count,count+1))
A[:count,:count]=np.eye(count);A[:count,-1]=-np.exp(tau)
A[count:,:count]=-np.eye(count);A[count:,-1]=np.exp(-tau)
lp=linprog(np.r_[cost,0.],A_ub=A,b_ub=np.zeros(2*count),
bounds=list(zip(x,x+c))+[(0,None)],method='highs')
lo,hi,y=exact_feasibility(x,c,tau)
self.assertEqual(lp.success,y is not None)
if y is not None:
self.assertAlmostEqual(float(cost@y),lp.fun,places=8)
self.assertTrue(np.all(y>=x-1e-12))
self.assertTrue(np.all(y<=x+c+1e-12))
def test_robust_finite_repair_under_adversarial_endpoints(self):
rng=np.random.default_rng(812)
for trial in range(500):
x=rng.uniform(.65,1.35,8);initial=x.copy()
eps=.006; tau=.04;beta=.005;hlo=.007;hhi=.012;cap=1.05
z=np.log(x)+rng.choice([-eps,eps],len(x))
cert=robust_feasibility(np.exp(z-eps),np.exp(z+eps),
cap,tau,beta,eps,hhi)
self.assertTrue(cert['feasible'])
k=cert['scale_lower'];a=cert['a'];b=cert['b']
accepted=np.zeros(len(x),bool)
bound=1+int(np.ceil(np.max(np.maximum(0,a*k-initial))/hlo))
for turn in range(bound+1):
# Force delay at the lower threshold, then alternate endpoints.
err=np.where(np.arange(len(x))%2==turn%2,eps,-eps)
measured=np.log(x)+err
# Exact endpoint equality can round by about 2e-16 in log.
accepted |= (measured>=np.log(k)-(tau-beta)+eps-1e-13) & \
(measured<=np.log(k)+(tau-beta)-eps+1e-13)
if accepted.all():break
need=~accepted
self.assertTrue(np.all(x[need]<a*k+1e-12))
x[need]+=rng.choice([hlo,hhi],int(need.sum()))
self.assertTrue(np.all(x<=b*k+1e-12))
self.assertTrue(accepted.all())
self.assertTrue(np.all(x-initial<=cap+1e-12))
final=x*np.exp(rng.choice([-beta,beta],len(x)))
self.assertLessEqual(np.max(np.abs(np.log(final/k))),tau+1e-12)
def test_no_guard_band_is_honest_failure(self):
result=robust_feasibility(np.array([1.]),np.array([1.]),1.,.02,.01,.01,.01)
self.assertFalse(result['feasible'])
def test_uniform_zero_tolerance_reduces_to_range_distribution(self):
for n in [2,3,10]:
for c in [.01,.3,.7]:
r=c/2
expected=n*r**(n-1)-(n-1)*r**n
self.assertAlmostEqual(uniform_reserve_yield(n,1,3,c,0),expected,places=14)
def test_uniform_threshold_all_sizes(self):
low,high,tau=.65,1.35,.04
critical=high*np.exp(-2*tau)-low
for n in [1,10,1000,1000000]:
self.assertEqual(uniform_reserve_yield(n,low,high,critical,tau),1.)
self.assertLess(uniform_reserve_yield(1000,low,high,critical-.01,tau),1e-5)
def test_common_gain_and_material_scale_cancel(self):
rng=np.random.default_rng(813)
coords,edges=grid_edges(4,3)
x=rng.uniform(.3,2,len(coords));target=rng.uniform(.2,3,len(coords))
expected=incidence(len(coords),edges)@np.log(x)
actual=ratio_observation(x,target,edges,
np.exp(rng.uniform(-3,3,len(edges))),np.zeros(len(edges)),
physical_scale=.017)
np.testing.assert_allclose(actual,expected,atol=2e-15)
def test_estimator_covariance_identity(self):
coords,edges=grid_edges(3,3);est=RelativeEstimator(len(coords),edges,reference=0)
A=cho_solve(est.factor,est.Bg.T.toarray())
np.testing.assert_allclose(A@A.T,est.covariance_unit,atol=2e-14)
def test_gradient_bias_is_exactly_invisible_to_cycles(self):
coords,edges=grid_edges(4,3);est=RelativeEstimator(len(coords),edges,reference=0)
truth=np.sin(np.arange(len(coords)));truth-=truth[0]
bias=.3*coords[:,0]/3
y=est.B@(truth+bias)
fit=est.estimate(y)
self.assertLess(np.linalg.norm(y-est.B@fit),1e-12)
self.assertAlmostEqual(np.max(np.abs(fit-truth)),.3,places=12)
def test_local_messages_have_valid_residual_error_bound(self):
coords,edges=grid_edges(4,2);est=RelativeEstimator(len(coords),edges,reference=0)
y=np.cos(np.arange(len(edges)))
local,info=est.local_estimate(y,tolerance=1e-6)
error=np.max(np.abs(local-est.estimate(y)))
self.assertLessEqual(error,info['certified_numerical_radius']+1e-12)
self.assertLessEqual(info['certified_numerical_radius'],1e-6)
def test_reference_bridge_and_disconnection(self):
coords,edges=grid_edges(3,2)
nodes,pairs=comparison_line_graph(edges)
ref_edge=int(np.flatnonzero(np.any(pairs==nodes-1,axis=1))[0])
self.assertTrue(closure_bridge_is_needed(nodes,pairs,ref_edge))
with self.assertRaises(ValueError):
RelativeEstimator(nodes,np.delete(pairs,ref_edge,axis=0))
def test_probe_schedule_uses_disjoint_modules(self):
coords,edges=grid_edges(4,3)
nodes,pairs=comparison_line_graph(edges);est=RelativeEstimator(nodes,pairs)
for color in np.unique(est.colors):
touched=pairs[est.colors==color].ravel()
self.assertEqual(len(touched),len(set(touched)))
def test_full_four_port_operator_is_scale_invariant(self):
coords,edges=grid_edges(4,3)
g=1.+np.arange(len(edges))%3
D=kron(laplacian(len(coords),edges,g),[0,3,48,63])
self.assertLess(projective_response_error(.006*D,D),1e-12)
perturbed=g.copy();perturbed[:10]*=.2
altered=kron(laplacian(len(coords),edges,perturbed),[0,3,48,63])
self.assertGreater(projective_response_error(altered,D),.04)
if __name__=='__main__':
unittest.main()