File size: 13,553 Bytes
19591f3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 | #!/usr/bin/env python3
"""Independent exact-arithmetic checks of constructions and finite certificates.
Run: python verify.py
These tests validate code and finite instances; they are not a proof assistant
formalization of the universal theorems in the manuscript.
"""
from fractions import Fraction as F
from itertools import product
from pathlib import Path
from collections import Counter
import json
import random
import subprocess
import sys
import tempfile
import time
import selection
HERE = Path(__file__).resolve().parent
COUNTS = Counter()
RNG = random.Random(61492026)
def sq_distance(a,b):
return sum((F(x)-F(y))**2 for x,y in zip(a,b))
def reference_moments(points,weights=None):
p = [F(1,len(points))]*len(points) if weights is None else list(map(F,weights))
mu = [sum(w*F(x[j]) for x,w in zip(points,p)) for j in range(len(points[0]))]
# Pairwise-distance identity is independent of the selector's central-moment code.
var = sum(p[i]*p[j]*sq_distance(points[i],points[j])
for i in range(len(points)) for j in range(i))/1
return mu,var
def count_error(points,weights,counts,n):
assert all(isinstance(i,int) and 0 <= i < len(points) for i in counts)
assert all(isinstance(v,int) and not isinstance(v,bool) and v > 0 for v in counts.values())
assert sum(counts.values()) == n
mu,var = reference_moments(points,weights)
selected = [sum(v*F(points[i][j]) for i,v in counts.items())/n
for j in range(len(mu))]
return sq_distance(mu,selected),mu,var,selected
def reference_coefficient(rank,n):
if rank == 0:
return F(0)
if n == 1:
return F(1)
if rank <= 2:
return F(1,2*n-1)
a = max(F(1,3),F(rank-1,2*rank))
if n <= rank:
return (F(n-2)+4*a)/(n*n)
b = (F(rank-2)+4*a)/(rank*rank)
return rank*b/(n*(1+b)-rank*b)
def check_output(points,n):
out = selection.select((row for row in points),n)
error,mu,var,h = count_error(points,None,out['counts'],n)
assert [F(x) for x in out['mean']] == mu
assert [F(x) for x in out['selected_mean']] == h
assert F(out['variance']) == var
assert F(out['squared_mean_error']) == error
rank = out['resident_affine_rank']
assert error <= reference_coefficient(rank,n)*var
assert F(out['guaranteed_upper_coefficient']) == reference_coefficient(rank,n)
assert F(out['full_data_loss_ratio']) == (1+error/var if var else 1)
indices = out['compressed_indices']
compressed = [[F(v) for v in row] for row in out['compressed_points']]
weights = [F(v) for v in out['compressed_weights']]
assert all(w>0 for w in weights) and sum(weights)==1
assert compressed == [list(map(F,points[i])) for i in indices]
cm,cv = reference_moments(compressed,weights)
assert cm == mu and cv <= var
assert cv == F(out['compressed_variance'])
assert len(compressed) <= len(points[0])+1
# The chosen output is at least as good as either computed certificate.
assert error <= F(out['residual_rounding_bound'])
COUNTS['full_dataset_outputs'] += 1
COUNTS['selected_method_'+out['selected_method']] += 1
for entry in out['construction_trace']:
COUNTS['trace_'+entry['step']] += 1
return out
def independent_rule_optimum(points,n,weights,rules):
m=len(points)
lower=rules.get('lower',[0]*m)
upper=rules.get('upper',[n]*m)
best=None
feasible=0
states=0
# Cartesian tuples, unlike the implementation's recursive compositions.
for counts in product(range(n+1),repeat=m):
if sum(counts)!=n:
continue
states+=1
if any(c<lo or c>hi for c,lo,hi in zip(counts,lower,upper)):
continue
ok=True
for row in rules.get('linear',[]):
lhs=sum(F(str(a))*c for a,c in zip(row['a'],counts))
rhs=F(str(row['rhs']))
if not {'<=':lhs<=rhs,'>=':lhs>=rhs,'==':lhs==rhs}[row['sense']]:
ok=False
break
if not ok:
continue
feasible+=1
e,*_=count_error(points,weights,{i:c for i,c in enumerate(counts) if c},n)
if best is None or (e,counts)<best:
best=(e,counts)
return best,feasible,states
def main():
started=time.monotonic()
for rank in range(13):
for n in range(1,81):
c=selection.coefficient(rank,n)
assert c==reference_coefficient(rank,n) and 0<=c<=F(1,n)
if rank>=3 and n>rank:
prev=selection.coefficient(rank,n-1)
assert c==F(n-1)*prev/(n+prev)
COUNTS['coefficient_identities']+=1
for d in range(1,9):
for case in range(3):
points=[[F(RNG.randint(-9,9),RNG.randint(1,5)) for _ in range(d)]
for _ in range(d+5)]
for n in sorted({1,2,3,d,d+1,d+4}):
check_output(points,n)
degenerate=[
[[0]*8 for _ in range(7)],
[[F(7,11),-2,3] for _ in range(4)],
[[i,2*i+1,-3*i,4*i-2,0,1,2*i,7] for i in range(-3,4)],
[[i,j,i+j,2*i-j,0,4,3*i,5*j] for i,j in [(0,0),(1,0),(0,1),(2,-1),(-3,2),(4,4)]],
[[2**80+F(i,17),-2**75+F(i*i,29)] for i in range(7)],
]
for points in degenerate:
for n in (1,2,3,7,23):
check_output(points,n)
COUNTS['degenerate_or_large_exact_inputs']+=1
# A deliberately hard residual case must use the direct fallback.
fallback=[[F(0),F(0)] for _ in range(18)]+[[F(1,10),F(1)],[F(1,10),F(-1)]]
result=check_output(fallback,10)
assert result['selected_method']=='pair-padding-peeling'
assert F(result['squared_mean_error'])==F(1,10000)
COUNTS['forced_direct_fallback_cases']+=1
# A billion requested copies must stay compact and avoid proportional loops.
result=check_output([[0,0],[1,0],[0,1]],10**9)
assert result['selected_method']=='floor-residual-atomic-selection'
assert len(result['counts'])<=3 and sum(result['counts'].values())==10**9
COUNTS['billion_budget_compact_fastpath_cases']+=1
for m in range(2,10):
for case in range(6):
d=min(6,m)
points=[[F(RNG.randint(-7,7),RNG.randint(1,4)) for _ in range(d)] for _ in range(m)]
raw=[RNG.randint(1,17) for _ in range(m)]
if case==0:
raw[0]=10**8
if case==1:
raw[-1]=0
weights=[F(v,sum(raw)) for v in raw]
active=sum(w>0 for w in weights)
for n in sorted({1,2,max(3,active-1),active+2}):
counts,trace=selection.select_atomic(points,weights,n)
error,_,var,_=count_error(points,weights,counts,n)
assert error<=reference_coefficient(active-1,n)*var
COUNTS['weighted_atomic_outputs']+=1
rcounts,bound=selection.residual_candidate(points,weights,n)
re,_,_,_=count_error(points,weights,rcounts,n)
assert re<=bound
residual=[n*p-int(n*p) for p in weights]
k=sum(residual)
assert k.denominator==1 and 0<=k<=active-1
if not k:
assert re==bound==0
COUNTS['fractional_residual_certificates']+=1
rational_points=[[F(0),F(0)],[F(1,3),F(2,7)],[F(-5,11),F(3,2)]]
rational_weights=[F(1,6),F(1,3),F(1,2)]
for n in (6,12,24,60,126):
counts,bound=selection.residual_candidate(rational_points,rational_weights,n)
error,*_=count_error(rational_points,rational_weights,counts,n)
assert error==bound==0
COUNTS['exact_denominator_budgets']+=1
for n in range(1,31):
points=[[0],[1]]
weights=[1-F(1,2*n),F(1,2*n)]
counts,_=selection.select_atomic(points,weights,n)
error,_,var,_=count_error(points,weights,counts,n)
assert error/var==F(1,2*n-1)
COUNTS['finite_line_sharpness_budgets']+=1
for case in range(32):
m=2+case%4
n=1+case%5
points=[[F(RNG.randint(-4,4),RNG.randint(1,3)) for _ in range(1+case%3)] for _ in range(m)]
raw=[RNG.randint(1,8) for _ in range(m)]
weights=[F(a,sum(raw)) for a in raw]
mode=case%8
if mode==0:
rules={}
elif mode==1:
rules={'upper':[1]*m}
elif mode==2:
rules={'lower':[1]+[0]*(m-1)}
elif mode==3:
rules={'linear':[{'a':[i%2 for i in range(m)],'sense':'==','rhs':n//2}]}
elif mode==4:
rules={'linear':[{'a':[i+1 for i in range(m)],'sense':'<=','rhs':n+2}]}
elif mode==5:
rules={'linear':[{'a':['1/2']+[0]*(m-1),'sense':'>=','rhs':'1/2'}],
'upper':[2]*m}
elif mode==6:
rules={'lower':[n+1]+[0]*(m-1)}
else:
rules={'upper':[n]*(m-1)+[0],
'linear':[{'a':[i-2 for i in range(m)],'sense':'>=','rhs':-n}]}
actual=selection.exact_rule_selection(points,n,weights,rules)
best,feasible,states=independent_rule_optimum(points,n,weights,rules)
assert actual['enumerated_states']==states and actual['feasible_states']==feasible
if best is None:
assert actual['status']=='INFEASIBLE'
COUNTS['infeasible_rule_instances']+=1
else:
counts=tuple(actual['counts'].get(i,0) for i in range(m))
assert actual['status']=='OPTIMAL'
assert (F(actual['squared_mean_error']),counts)==best
COUNTS['optimal_rule_instances']+=1
COUNTS['independent_rule_comparisons']+=1
# Exact rational Gram witness: a^2=5, b^2=9, tetrahedral inner products -1/3.
p=[F(2,5)]+[F(3,20)]*4
gram=[[F(0)]*5 for _ in range(5)]
for i in range(1,5):
for j in range(1,5):
gram[i][j]=F(14 if i==j else 2)
mean_norm=sum(p[i]*p[j]*gram[i][j] for i in range(5) for j in range(5))
variance=sum(p[i]*gram[i][i] for i in range(5))-mean_norm
values=[]
for counts in product(range(4),repeat=5):
if sum(counts)!=3:
continue
delta=[F(c,3)-w for c,w in zip(counts,p)]
error=sum(delta[i]*delta[j]*gram[i][j] for i in range(5) for j in range(5))
values.append(error)
assert variance==F(33,5) and min(values)==F(61,45)
assert min(values)/variance==F(61,297)>F(1,5)
COUNTS['exact_rare_simplex_gram_witnesses']+=1
tetra=[[0,0,0,0],[4,3,3,3],[4,3,-3,-3],[4,-3,3,-3],[4,-3,-3,3]]
best,_,_=independent_rule_optimum(tetra,3,p,{})
assert best[0]==F(916,225)
assert reference_moments(tetra,p)[1]==F(501,25)
assert best[0]/F(501,25)==F(916,4509)>F(1,5)
finite=tetra[:1]*8+sum(([row]*3 for row in tetra[1:]),[])
check_output(finite,3)
COUNTS['rational_coordinate_rare_simplex_witnesses']+=1
# Exact polynomial check that B_(4,n) exceeds 1/(2n-1).
for n in range(2,101):
k=4
a=2*n*(k-2)-(k-3)
delta=(k-3)*(2*n*(n-2)*(k-2)+(k-3))
c=F(1,2*n-1)
assert c<F(a,k*n*n)
assert (k*n*n*c)**2-2*a*k*n*n*c+a*a-delta>0
COUNTS['dimension_four_strict_separation_checks']+=1
with tempfile.TemporaryDirectory() as directory:
inp=Path(directory)/'input.json'
inp.write_text('{"points":[[0.1000000000000000000000001,0.2],[0.3,-0.4],[0.5,0.6]],"n":7}')
completed=subprocess.run([sys.executable,str(HERE/'selection.py'),str(inp)],
check=True,text=True,capture_output=True)
out=json.loads(completed.stdout)
assert F(out['mean'][0])==F('0.9000000000000000000000001')/3
COUNTS['exact_decimal_cli_checks']+=1
inp.write_text('{"points":[[0],[1],[2]],"weights":["1/2","1/3","1/6"],"n":3,"rules":{"upper":[1,1,1]}}')
completed=subprocess.run([sys.executable,str(HERE/'selection.py'),str(inp),'--exact'],
check=True,text=True,capture_output=True)
out=json.loads(completed.stdout)
assert out['status']=='OPTIMAL' and out['counts']=={'0':1,'1':1,'2':1}
COUNTS['weighted_constrained_cli_checks']+=1
bad_calls=[
lambda:selection.select([],2),
lambda:selection.select([[1],[2,3]],2),
lambda:selection.select([[0],[1]],True),
lambda:selection.select([[0],[1]],0),
lambda:selection.select_atomic([[0],[1]],[F(1,3),F(1,3)],2),
lambda:selection.select_atomic([[0],[1]],[F(-1),F(2)],2),
lambda:selection.exact_rule_selection([[0],[1]],3,rules={'linear':[{'a':[1,2],'sense':'!=','rhs':0}]}),
lambda:selection.exact_rule_selection([[0],[1],[2]],20,max_states=5),
]
for call in bad_calls:
try:
call()
except ValueError:
COUNTS['invalid_input_rejections']+=1
else:
raise AssertionError('Invalid input was accepted')
result={'status':'PASS','arithmetic':'fractions.Fraction; exact comparisons',
'seed':61492026,'counts':dict(sorted(COUNTS.items())),
'wall_seconds':round(time.monotonic()-started,3),
'independence':'Original-data variance uses pairwise distances; rule oracle uses Cartesian count tuples; rare-simplex certificate uses a rational Gram matrix.',
'scope':'Implementation and finite certificates only; universal claims rely on the manuscript proofs. No claim of proof-assistant formalization.'}
(HERE/'verification_results.json').write_text(json.dumps(result,indent=2)+'\n')
print(json.dumps(result,indent=2))
if __name__=='__main__':
main()
|