Datasets:
Formats:
json
Languages:
English
Size:
< 1K
ArXiv:
Tags:
mathematics
optimization
gradient-descent
convex-optimization
learning-theory
recursive-self-improvement
File size: 7,340 Bytes
868c741 | 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 | #!/usr/bin/env python3
"""Independent exact checks for the adjacent-concavity negative direction.
The all-parameter conclusion rests on the written inequalities and analytic
proof. This file checks the endpoint arithmetic, polynomial identities by a
degree-certified grid, and a rational three-dimensional repaired transcript.
"""
from fractions import Fraction as Q
from itertools import product
import json
import argparse
from pathlib import Path
def dot(x, y):
return sum((a * b for a, b in zip(x, y)), Q(0))
def sub(x, y):
return tuple(a - b for a, b in zip(x, y))
def jet(t, A, B):
u = [Q(0), -A, 1 - A, 1 - A - B]
v = [
Q(0),
(-3*A*A*t - 3*A*A - A*t + 4*A - 2*B + 2*t - 1)/(6*t),
(-A*A*t - A*A + A*t + 2*A - 1)/(2*t),
Q(-10),
]
w = [
Q(0),
A*A*t/2 + A*t/6 - 2*A/3 + B/3 - t/3 + Q(1,6),
A*A*t/2 + A*A + A*t/6 - 5*A/3 + B/3 - t/3 + Q(2,3),
A*A*t/2 + 2*A*A + 7*A*t/6 - 11*A/3 + B*B/2 + B/3 - 4*t/3 + Q(5,3),
]
U = [sum(u[:i], Q(0)) for i in range(4)]
V = [sum(v[:i], Q(0)) for i in range(4)]
J, K = {}, {}
for i, j in product(range(4), repeat=2):
if i == j:
continue
J[i,j] = w[i]-w[j]+t*((i-j)*v[j]+V[i]-V[j]+u[j]*(U[i]-U[j]))-(u[i]-u[j])**2/2
K[i,j] = t*v[j]*(V[i]-V[j])-(v[i]-v[j])**2/2
return u, v, w, U, V, J, K
def leading_formulas(t, A, B):
return {
(0,1): Q(0), (0,2): Q(0), (2,0): Q(0), (3,1): Q(0), (3,2): Q(0),
(1,0): (3*A*A*(t-1)+A*(t-4)+2*B-2*t+1)/6,
(1,2): (2*A*(t-1)+B-t-1)/3,
(2,1): (2*t-1-A*(t-1)-2*B)/3,
(3,0): (A-1)*(2*t-1-A*(t-1)-2*B)/2,
(0,3): -(3*A*A*(t+1)+4*A*B*t+2*A*B-3*A*t-6*A+2*B*B-2*B*t-2*B-60*t+3)/2,
(1,3): -(3*A*A*(t+1)+6*A*B*t-5*A*t-4*A+3*B*B-3*B*t-4*B-59*t+4)/3,
(2,3): -(A*A*(t+1)+2*A*B*t-A*t-2*A+2*B*B-2*B*t-20*t+1)/2,
}
def kappa(t, A, B):
return (
-(3*A*A*t - 6*A*A + A*t + 2*A + 2*B - 2*t - 2)/6,
(2*A*(t-1) + B*B - 2*t + 1)/2,
)
def polynomial_identities():
# Every J expression has individual degrees <= (1,2,2) in (t,A,B)
# after t cancels the denominators of v. This tensor grid determines
# those polynomial identities exactly.
count = 0
for t, A, B in product(map(Q,[1,2]), map(Q,[0,1,2]), map(Q,[0,1,2])):
u, v, w, U, V, J, K = jet(t,A,B)
assert J == leading_formulas(t,A,B)
k0, k1 = kappa(t,A,B)
assert w[0]-2*w[1]+w[2] == k0
assert w[1]-2*w[2]+w[3] == k1
C = 3*A*A*(t-2)+A*(7*t-4)-8*t+1
assert k0-k1 == -(3*B*B+2*B+C)/6
for r in map(Q,[0,1]):
m00 = r*(1-t/2)
m01 = r*(t+2)/12-(t-1)/2
m11 = r*(t+1)/3-t+Q(1,2)
m12 = r/6
m22 = Q(1,2)
form = m00*A*A-2*m01*A+m11-2*m12*B+m22*B*B
assert form == r*k0+k1
count += 1
return count
def endpoint_checks():
alo, ahi = Q(13517,10000), Q(6759,5000)
hlo, hhi = Q(1863,1000), Q(233,125)
blo, bhi = Q(5,8), Q(157,250)
def C(A,h):
return 3*A*A*(h-2)+A*(7*h-4)-8*h+1
assert 6*ahi*(hlo-2)+7*hlo-4 > 0
assert 3*alo*alo+7*alo-8 > 0
cmin, cmax = C(alo,hlo), C(ahi,hhi)
assert cmin == -Q(243421547779,100000000000)
assert cmax == -Q(7582743131,3125000000)
assert 3*blo*blo+2*blo+cmax == -Q(3595939,781250000)
assert 3*bhi*bhi+2*bhi+cmin == Q(493652221,100000000000)
slope = ahi-1-blo*(3*alo*alo+7*alo-8)/(6*bhi+2)
assert slope == -Q(369648683,922880000) < -Q(3,8)
bounds = [
(3*alo*alo*(hlo-1)+ahi*(hlo-4)+2*blo-2*hhi+1)/6,
(2*alo*(hlo-1)+blo-hhi-1)/3,
(2*hlo-1-ahi*(hhi-1)-2*bhi)/3,
(alo-1)*(2*hlo-1-ahi*(hhi-1)-2*bhi)/2,
(60*hlo-(3*ahi*ahi*(hhi+1)+4*ahi*bhi*hhi+2*ahi*bhi+2*bhi*bhi+3))/2,
(59*hlo-(3*ahi*ahi*(hhi+1)+6*ahi*bhi*hhi+3*bhi*bhi+4))/3,
(20*hlo-(ahi*ahi*(hhi+1)+2*ahi*bhi*hhi+2*bhi*bhi+1))/2,
]
claimed = [Q(3,50), Q(3,100), Q(1,10), Q(1,20), Q(42), Q(26), Q(13)]
assert all(a>b for a,b in zip(bounds,claimed))
return {"slope_upper":str(slope),"leading_slack_lower_bounds":[str(x) for x in bounds]}
def rational_repair():
t, A, B = Q(233,125), Q(13517,10000), Q(313,500)
u,v,w,U,V,J,K = jet(t,A,B)
assert min(J.values()) == 0
k0,k1 = kappa(t,A,B)
assert max(k0,k1)<0
rho=1-t/2
z=[2*rho**i for i in range(4)]
b=[rho**i for i in range(4)]
aa=[rho**(2*i) for i in range(4)]
P={ij:(rho**ij[0]-rho**ij[1])**2/2 for ij in J}
# Both square-root scales are rational in this concrete audit case.
c=Q(10**8)
e=Q(1,2*10**6)
e2=e*e
assert c>max(abs(K[ij])/P[ij] for ij in J)
q0=(1-rho*rho)**2
assert c*e2*q0<=-max(k0,k1)/2
# Scaled coordinates are represented without square roots:
# x=(x0, sqrt(e2)*x1, sqrt(c)*e2*x2), likewise for g.
X=[(-t*i-t*e2*V[i],-t*U[i],z[i]) for i in range(4)]
G=[(1+e2*v[i],u[i],b[i]) for i in range(4)]
f=[-t*i+e2*w[i]+c*e2*e2*aa[i] for i in range(4)]
weights=(Q(1),e2,c*e2*e2)
def wdot(x,y):
return sum((ww*xx*yy for ww,xx,yy in zip(weights,x,y)),Q(0))
for i in range(3):
assert all(X[i+1][j]==X[i][j]-t*G[i][j] for j in range(3))
residuals={}
for i,j in J:
dx,dg=sub(X[i],X[j]),sub(G[i],G[j])
I=f[i]-f[j]-wdot(G[j],dx)-wdot(dg,dg)/2
assert I==e2*J[i,j]+e2*e2*(K[i,j]+c*P[i,j])>0
residuals[i,j]=I
curves=[f[i]-2*f[i+1]+f[i+2] for i in range(2)]
assert max(curves)<=e2*max(k0,k1)/2<0
amin=min(residuals.values())
bmax=max(wdot(sub(sub(X[i],X[j]),sub(G[i],G[j])),
sub(sub(X[i],X[j]),sub(G[i],G[j]))) for i,j in J)
mu=amin/(amin+bmax+1)
assert 0<mu<1
assert all(I>mu*bmax/(2*(1-mu)) for I in residuals.values())
root_c=Q(10**4)
actual_x=[(row[0],e*row[1],root_c*e2*row[2]) for row in X]
actual_g=[(row[0],e*row[1],root_c*e2*row[2]) for row in G]
assert root_c*root_c==c
M=1+max(wdot(row,row) for row in G)
delta=-max(curves)
C=16*M*t+6+27*t*t
sigma=min(Q(1),delta/(2*C))
assert C*sigma<=delta/2
return {
"step":str(t),"A":str(A),"B":str(B),
"rational_points":[[str(z)for z in row]for row in actual_x],
"rational_gradients":[[str(z)for z in row]for row in actual_g],
"rational_values":[str(z)for z in f],
"exact_interpolation_slacks":{str(ij):str(v)for ij,v in residuals.items()},
"exact_repaired_curvatures":[str(v)for v in curves],
"mu":str(mu),"analytic_sigma":str(sigma),
"leading_curvatures":[str(k0),str(k1)],
"repaired_curvatures_display":[float(v) for v in curves],
"strong_convexity_display":float(mu),
"strict_interpolation_pairs":len(residuals),
"all_exact_checks_passed":True,
}
if __name__=="__main__":
parser=argparse.ArgumentParser()
parser.add_argument("--output",type=Path)
args=parser.parse_args()
result={
"polynomial_identity_grid_cases":polynomial_identities(),
"endpoint_checks":endpoint_checks(),
"rational_repair":rational_repair(),
"all_exact_checks_passed":True,
}
out=json.dumps(result,indent=2)+"\n"
if args.output:
args.output.parent.mkdir(parents=True,exist_ok=True)
args.output.write_text(out)
print(out)
|