File size: 13,018 Bytes
bd2c7f9 | 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 | #!/usr/bin/env python3
"""Correctness tests for the paper's Section 5 / Appendix C application models.
Every transcription and every closed form used by run_apub_applications.py is
pinned here against an independent computation, so a silent transcription bug
cannot pass as a result. Run: python3 code/test_apub_paper_models.py
"""
import json
import sys
import numpy as np
from scipy import stats
sys.path.insert(0, __file__.rsplit("/", 1)[0])
import apub_paper_models as P
rng = np.random.default_rng(20260802)
RESULTS = []
def check(name, ok, detail):
RESULTS.append({"test": name, "pass": bool(ok), "detail": detail})
print(f"[{'PASS' if ok else 'FAIL'}] {name}: {detail}")
return ok
# ---------------------------------------------------------------------------
# T1 positive-stable draw has Laplace transform exp(-t^a)
# ---------------------------------------------------------------------------
def t1():
worst = 0.0
for lam in (2.0, 5.0):
a = 1.0 / lam
s = P.positive_stable(a, size=400000, rng=rng)
for t in (0.3, 1.0, 3.0):
emp = np.mean(np.exp(-t * s))
worst = max(worst, abs(emp - np.exp(-t ** a)))
return check("positive_stable_laplace_transform", worst < 5e-3,
f"max |E[exp(-tS)] - exp(-t^a)| = {worst:.2e} over lam in "
f"{{2,5}}, t in {{0.3,1,3}} (400k draws)")
# ---------------------------------------------------------------------------
# T2 Gumbel copula: uniform marginals and Kendall tau = 1 - 1/lambda
# ---------------------------------------------------------------------------
def t2():
ok, det = True, []
for lam in (2.0, 5.0):
u = P.gumbel_copula_uniforms(60000, 4, lam, rng)
ks = max(stats.kstest(u[:, k], "uniform").statistic for k in range(4))
tau = stats.kendalltau(u[:2000, 0], u[:2000, 1]).statistic
target = 1.0 - 1.0 / lam
ok &= (ks < 0.01) and (abs(tau - target) < 0.03)
det.append(f"lam={lam}: max KS={ks:.4f}, tau={tau:.3f} (theory {target:.3f})")
return check("gumbel_copula_marginals_and_kendall_tau", ok, "; ".join(det))
# ---------------------------------------------------------------------------
# T3 closed-form product-mix recourse == recourse LP
# ---------------------------------------------------------------------------
def t3():
xi = P.sample_product_mix_xi(40, rng)
worst = 0.0
for _ in range(6):
x = rng.uniform(0, 12, size=20)
cf, feas = P.recourse_pm_closed_form(x, xi)
assert feas.all(), "closed form only claims optimality where sum_j y_j <= h2"
for n in range(0, 40, 7):
lp = P.recourse_pm_lp(x, xi["q"][n], xi["w"][n], xi["h1"][n], xi["h2"][n])
worst = max(worst, abs(cf[n] - lp) / max(1.0, abs(lp)))
return check("recourse_closed_form_equals_LP", worst < 1e-8,
f"max relative gap over 6 x-vectors x 6 scenarios = {worst:.2e}")
# ---------------------------------------------------------------------------
# T4 APUB(alpha) is the CVaR of the bootstrap means and its t-formulation
# agrees with the sorted-tail formula
# ---------------------------------------------------------------------------
def t4():
theta = rng.gamma(2.0, 3.0, size=40)
V = P.bootstrap_multiplicities(40, 4000, rng)
zm = (V @ theta) / 40
worst = 0.0
for alpha in (0.05, 0.2, 0.5, 1.0):
direct = P.apub_from_costs(theta, V, alpha)
grid = np.linspace(zm.min() - 1, zm.max() + 1, 200001)
var = np.quantile(zm, 1 - alpha) if alpha < 1 else zm.min() - 1
cand = np.unique(np.concatenate([grid[::200], [var]]))
vals = cand + (np.maximum(zm[None, :] - cand[:, None], 0).mean(axis=1)) / alpha
worst = max(worst, abs(direct - vals.min()) / max(1.0, abs(direct)))
return check("apub_equals_inf_t_formulation", worst < 1e-3,
f"max relative gap between sorted-tail CVaR and "
f"inf_t {{t + E[(Z-t)_+]/alpha}} = {worst:.2e}")
# ---------------------------------------------------------------------------
# T5 APUB monotone decreasing in alpha, and alpha = 1 reduces to SAA
# (Remark rem:APUB-DEF-OPT), on the real 20x8 model
# ---------------------------------------------------------------------------
def t5():
xi = P.sample_product_mix_xi(25, rng)
V = P.bootstrap_multiplicities(25, 400, rng)
objs = {}
for alpha in (0.05, 0.2, 0.5, 1.0):
objs[alpha] = P.solve_pm_random_recourse(xi, alpha, V, x_ub=5000.0)["obj"]
saa = P.solve_pm_random_recourse(xi, 1.0, None, x_ub=5000.0)["obj"]
mono = all(objs[a] >= objs[b] - 1e-6
for a, b in zip([0.05, 0.2, 0.5], [0.2, 0.5, 1.0]))
# Remark rem:APUB-DEF-OPT holds for the EXACT bootstrap law; with a finite
# M the alpha = 1 model is the Monte-Carlo bootstrap approximation of SAA,
# so the right check is that the gap contracts as M grows.
gaps = []
for M in (200, 1600, 12800):
Vm = P.bootstrap_multiplicities(25, M, rng)
o = P.solve_pm_random_recourse(xi, 1.0, Vm, x_ub=5000.0)["obj"]
gaps.append(abs(o - saa) / abs(saa))
# At alpha = 1 the model reduces to a bootstrap-REWEIGHTED SAA with weights
# w_n = (1/M) sum_m V_mn (mean 1, sd ~ M^{-1/2}), so the gap to plain SAA is
# O(M^{-1/2}), not zero at finite M.
contracts = all(gaps[i] > gaps[i + 1] for i in range(len(gaps) - 1)) \
and gaps[-1] < 1e-2
return check("apub_sp_monotone_in_alpha_and_alpha1_converges_to_SAA",
mono and contracts,
f"objs by alpha = { {k: round(v, 4) for k, v in objs.items()} } "
f"(monotone non-increasing in alpha: {mono}); SAA = {saa:.4f}; "
f"alpha=1 relative gap to SAA at M=(200,1600,12800) = "
f"{[f'{g:.2e}' for g in gaps]}")
# ---------------------------------------------------------------------------
# T6 the APUB-SP LP optimum equals the APUB of the recourse costs evaluated
# independently at the returned x (i.e. the LP really optimises the paper's
# objective, not a relaxation of it)
# ---------------------------------------------------------------------------
def t6():
xi = P.sample_product_mix_xi(25, rng)
V = P.bootstrap_multiplicities(25, 400, rng)
det = []
ok = True
for alpha in (0.1, 0.5):
sol = P.solve_pm_random_recourse(xi, alpha, V, x_ub=5000.0)
cf, _ = P.recourse_pm_closed_form(sol["x"], xi)
rebuilt = float(P.PM_C @ sol["x"]) + P.apub_from_costs(cf, V, alpha)
gap = abs(rebuilt - sol["obj"]) / max(1.0, abs(sol["obj"]))
ok &= gap < 1e-6
det.append(f"alpha={alpha}: LP obj={sol['obj']:.4f}, "
f"independent rebuild={rebuilt:.4f}, rel gap={gap:.2e}")
return check("apub_sp_LP_objective_matches_independent_rebuild", ok, "; ".join(det))
# ---------------------------------------------------------------------------
# T7 fixed-recourse closed form == the W y = h - T x recourse LP
# ---------------------------------------------------------------------------
def t7():
from scipy.optimize import linprog
gam = P.sample_fixed_recourse_gamma(30, rng)
worst = 0.0
for _ in range(8):
x = rng.uniform(0, 200, size=4)
cf = P.fr_recourse(x, gam)
Tg = P.FR_T_BASE - gam[:, :, None] * 0.25 # (n,2,4)
for n in range(0, 30, 5):
rhs = 500.0 * gam[n] - Tg[n] @ x
res = linprog(P.FR_QCOST, A_eq=P.FR_W, b_eq=rhs,
bounds=[(0, None)] * 4, method="highs")
worst = max(worst, abs(cf[n] - res.fun) / max(1.0, abs(res.fun)))
return check("fixed_recourse_closed_form_equals_LP", worst < 1e-8,
f"max relative gap over 8 x-vectors x 6 scenarios = {worst:.2e}")
# ---------------------------------------------------------------------------
# T8 the DRO Lipschitz moduli are the true ones (finite-difference check)
# ---------------------------------------------------------------------------
def t8():
ok, det = True, []
for _ in range(5):
x = rng.uniform(0, 200, size=4)
g = rng.uniform(0.5, 20.0, size=(4000, 2))
f = P.fr_recourse(x, g)
pert = g + rng.uniform(-1e-3, 1e-3, size=g.shape)
num = np.abs(P.fr_recourse(x, pert) - f) / np.abs(pert - g).sum(axis=1)
ok &= num.max() <= P.fr_lipschitz(x) * (1 + 1e-6)
det.append(f"emp={num.max():.2f} <= analytic={P.fr_lipschitz(x):.2f}")
# newsvendor: modulus is max(h, b) and is x-independent
for _ in range(3):
x = rng.uniform(20, 90, size=10)
d = rng.uniform(20, 90, size=(4000, 10))
pert = d + rng.uniform(-1e-3, 1e-3, size=d.shape)
num = np.abs(P.nv_cost(x, pert) - P.nv_cost(x, d)) / \
np.abs(pert - d).sum(axis=1)
ok &= num.max() <= max(P.NV_H, P.NV_B) * (1 + 1e-6)
det.append(f"newsvendor modulus max(h,b)={max(P.NV_H, P.NV_B)} respected")
return check("wasserstein_lipschitz_moduli_are_exact", ok, "; ".join(det))
# ---------------------------------------------------------------------------
# T9 newsvendor SAA LP optimum equals the direct empirical-average minimiser
# (checked against the closed-form critical quantile, product by product)
# ---------------------------------------------------------------------------
def t9():
xi = P.sample_newsvendor(400, rng, case=1)
sol = P.solve_nv(xi, alpha=1.0, V=None)
# The multi-product newsvendor separates by product; the marginal of
# p x + h(x-d)_+ + b(d-x)_+ is p + hF(x) - b(1-F(x)), so the empirical
# minimiser is the (b - p)/(h + b) empirical quantile of that column.
qlev = (P.NV_B - P.NV_P) / (P.NV_H + P.NV_B)
closed = np.quantile(xi, qlev, axis=0, method="inverted_cdf")
obj_closed = float(P.nv_cost(closed, xi).mean())
gap = abs(obj_closed - sol["obj"]) / max(1.0, abs(sol["obj"]))
return check("newsvendor_SAA_LP_equals_empirical_quantile_solution",
gap < 1e-9,
f"LP obj={sol['obj']:.6f}, closed-form quantile obj="
f"{obj_closed:.6f}, rel gap={gap:.2e}, "
f"max |x_LP - x_quantile| = {np.abs(sol['x'] - closed).max():.3f}")
# ---------------------------------------------------------------------------
# T10 newsvendor APUB-SP LP objective matches an independent rebuild, and the
# DRO shift is exactly eps*max(h,b) with an unchanged solution
# ---------------------------------------------------------------------------
def t10():
xi = P.sample_newsvendor(30, rng, case=1)
V = P.bootstrap_multiplicities(30, 800, rng)
sol = P.solve_nv(xi, alpha=0.1, V=V)
theta = P.nv_cost(sol["x"], xi)
rebuilt = P.apub_from_costs(theta, V, 0.1)
gap = abs(rebuilt - sol["obj"]) / max(1.0, abs(sol["obj"]))
saa = P.solve_nv(xi, alpha=1.0, V=None)
dro = P.solve_nv(xi, alpha=1.0, V=None, dro_eps=0.7)
shift_ok = abs((dro["obj"] - saa["obj"]) - 0.7 * max(P.NV_H, P.NV_B)) < 1e-6
same_x = np.abs(dro["x"] - saa["x"]).max() < 1e-6
return check("newsvendor_apub_rebuild_and_dro_constant_shift",
gap < 1e-6 and shift_ok and same_x,
f"APUB LP obj={sol['obj']:.6f} vs rebuild={rebuilt:.6f} "
f"(rel gap {gap:.2e}); DRO-SAA shift="
f"{dro['obj'] - saa['obj']:.6f} vs eps*max(h,b)="
f"{0.7 * max(P.NV_H, P.NV_B):.6f}; x unchanged={same_x}")
# ---------------------------------------------------------------------------
# T11 fixed-recourse DRO solution DOES move with eps (decision-dependent
# Lipschitz modulus), unlike the newsvendor
# ---------------------------------------------------------------------------
def t11():
gam = P.sample_fixed_recourse_gamma(120, rng)
saa = P.solve_fr(gam, 1.0, None)
moves = []
eps_grid = (0.05, 0.2, 1.0)
for eps in eps_grid:
dro = P.solve_fr(gam, 1.0, None, dro_eps=eps)
moves.append(float(np.linalg.norm(dro["x"] - saa["x"])))
nondec = all(moves[i] <= moves[i + 1] + 1e-6 for i in range(len(moves) - 1))
ok = moves[-1] > 1e-6 and nondec
return check("fixed_recourse_DRO_solution_is_decision_dependent", ok,
f"||x_DRO - x_SAA||_2 at eps={eps_grid} = "
f"{[round(m, 3) for m in moves]} (non-decreasing and eventually "
f"positive => the Wasserstein modulus really depends on x; "
f"contrast the newsvendor, where it cannot move at all)")
def main():
fns = [t1, t2, t3, t4, t5, t6, t7, t8, t9, t10, t11]
allok = True
for f in fns:
allok &= bool(f())
out = {"all_passed": bool(allok), "tests": RESULTS}
with open(__file__.rsplit("/", 2)[0] + "/results/paper_model_unit_tests.json", "w") as fh:
json.dump(out, fh, indent=1)
print("\nALL PASSED" if allok else "\nSOME FAILED")
return 0 if allok else 1
if __name__ == "__main__":
sys.exit(main())
|