DCT Laboratory — Volume II, Chapter 11
Distributionally Robust Enterprise Optimization
Seed 26211 · Companion to Chapter 11 and AXIOM Lab 2.11 (module AXIOM-11) · Mirrored in DCT_V2_Ch11_Lab.xlsx
This laboratory reproduces Worked Example 11.1, How Much Should Meridian Trust Its Law?, with the book's declarations, and works the five steps of AXIOM Lab 2.11: read the records, price the family, try balls, set a radius from a record, and moments only. The law, the levers and the floor are Chapter 9's. The new declarations are the records behind the law: the treasury's two closures in twenty years and the engineering office's forty quarters of deviations. Every law the records allow is simulated on Chapter 9's 100,000 paths, drawn from Volume I, Chapter 8's stream with seed 8, and checked on the book's 100,000 fresh paths, seed 26209; the closures a higher rate adds come from seeds 1108 and 26208 (Online Appendix Remark 11.B.1), and the records of 1,000 paths from seeds 900000–900199 (Chapter 9's sample-size study). The lab's own seed, 26211, draws only the simulated records of Panel 1's coverage check. Every Meridian number below is the book's (Table 11.3, Figures 11.1–11.3, Online Appendix Example 11.B.2); the final cell checks them. Money is in $ million.
The declarations (Worked Example 11.1, Step 1; Online Appendix Remark 11.B.1)
Chapter 9's law, levers and floor. Planned liquidity follows the Chapter 1 plan, whose draw dips to $482 million in quarter 7 of the program's 12; the board's floor is $350 million. Execution and market deviations move it, with of spread $50 million a quarter; funding windows close at random, 0.10 a year, each removing a gamma-distributed amount (mean $120 million, standard deviation $50 million) that is never recovered. A standby line covers up to of each closure, for fees of over the program; a quarter of committed delay makes the draw 4 index points shallower, at $8 million. The board's appetite: at most a 5 percent chance of a breach of the floor in some quarter. Chapter 9 kept it with Volume I's package, a $170 million line and two-thirds of a quarter of delay, $10.44 million.
The records behind the law. The treasury's rate, 0.10 a year, is two closures in its twenty years of history. The engineering office's spread, $50 million a quarter, is the root-mean-square of 39 innovations from forty quarters of deviations (the persistence, 0.6, is taken as known). Neither record pins its number down. The records' family holds every law of Chapter 9's form with closures at most 0.315 a year and a spread at most $61.6 million, the records' one-sided 95 percent bounds (Proposition 11.2).
How the family's laws are simulated (Remark 11.B.1). Every law is simulated on Chapter 9's paths, coupled so that the declared law is exactly Chapter 9's. A rate adds the closures of an independent stream of rate , drawn by thinning one extra stream of rate with uniforms, so that a higher rate only adds closures; a rate keeps each of the paths' own closures with chance . The extra closures and the uniforms come from seed 1108 for the book's paths and seed 26208 for the fresh ones: each quarter draws the uniforms for the paths' own closures first, then the extra stream's counts, gamma sizes and uniforms. A spread million scales every deviation by , because the deviations are linear in their innovations.
Python Code
import math
import numpy as np
from scipy import optimize, stats
import matplotlib.pyplot as plt
plt.rcParams.update({"figure.dpi": 110, "axes.spines.top": False, "axes.spines.right": False})
SEED = 26211 # the lab's seed: it draws only the simulated records of Panel 1's coverage check
BOOK_SEED = 8 # Volume I, Chapter 8's stream: the book's 100,000 paths (Online Appendix 1.B.4)
FRESH_SEED = 26209 # the book's 100,000 fresh paths (Chapter 9's lab seed)
EXTRA_SEED = {BOOK_SEED: 1108, FRESH_SEED: 26208} # the family's extra closures and uniforms (Remark 11.B.1)
N, Q, YEARS = 100_000, 12, 3.0 # paths; quarters of the program; its years
# Chapter 9's law in USD billion, exactly as Volume I's engine draws it; every result is reported in $ million
FLOOR = 0.35 # the board's floor
A_E, SD_W = 0.6, 0.05 # E' = 0.6 E + w, w with spread $50m a quarter
LAM, J_MEAN, J_SD = 0.10, 0.12, 0.05 # closures a year; gamma sizes, mean $120m and sd $50m, never recovered
GAMMA = ((J_MEAN / J_SD) ** 2, J_SD ** 2 / J_MEAN) # the sizes' shape and scale
LAM_MAX = 0.40 # the extra stream's rate ceiling (Remark 11.B.1)
DEPTH0, DPQ = 28.0, 4.0 # the draw's depth (index points); a quarter of delay: 4 points shallower
FEE, VQ, ALPHA = 0.03, 8.0, 0.05 # line fees over the program; a quarter of delay ($m); the appetite
MAX_DELAY = DEPTH0 / DPQ # 7 quarters remove the draw
LINES = np.arange(0.0, 400.01, 5.0) # the line on a $5m grid up to $400m; the delay is continuous
def relax(x0, x_inf, k, t):
return x_inf + (x0 - x_inf) * np.exp(-k * t)
def trough(depth, tau, t):
return depth * (t / tau) * np.exp(1.0 - t / tau)
TQ = np.arange(1, Q + 1) / 4.0 # the twelve quarters, in years
BASE = 0.62 * relax(100.0, 112.0, 0.35, TQ) / 100.0 # planned liquidity without the draw (index 100 = $0.62bn)
SHAPE = 0.62 * trough(1.0, 2.0, TQ) / 100.0 # the draw per index point of depth, deepest at two years
LIFT = 1000.0 * DPQ * SHAPE # $m of liquidity a quarter of delay adds, quarter by quarter
MARGIN = 1000.0 * (BASE - DEPTH0 * SHAPE - FLOOR) # the plan's margin above the floor ($m)
def draws(seed, n=N):
# Volume I's stream in its order: each quarter the execution shocks, the closure counts, then the sizes path by path
rng = np.random.default_rng(seed)
W, events = np.empty((n, Q)), []
for k in range(Q):
W[:, k] = rng.normal(0.0, SD_W, n)
count = rng.poisson(LAM / 4.0, n)
events.append((np.repeat(np.arange(n), count), rng.gamma(*GAMMA, int(count.sum()))))
E, e = np.zeros((n, Q)), np.zeros(n)
for k in range(Q):
e = A_E * e + W[:, k]
E[:, k] = e
return E, events
class Paths:
# The paths of one law. A line l ($m) covers up to l of each closure; phi quarters of delay make the draw 28 - 4 phi
# deep. A path stays above the floor exactly when the draw is no deeper than its critical depth, so chances are
# exact counts and the least delay meeting a chance is a quantile of the critical depths.
def __init__(self, E, events, seed):
self.E, self.events, self.seed, self.n = E, events, seed, E.shape[0]
self._crit, self._sorted = {}, {}
@classmethod
def draw(cls, seed, n=N):
return cls(*draws(seed, n), seed)
def lost(self, line): # cumulative closures the line does not cover (USD bn), quarter by quarter
J, j = np.zeros((self.n, Q)), np.zeros(self.n)
for k, (path, size) in enumerate(self.events):
j = j + np.bincount(path, np.maximum(size - line / 1000.0, 0.0), self.n)
J[:, k] = j
return J
def crit(self, line): # each path's critical depth
key = round(float(line), 6)
if key not in self._crit:
self._crit[key] = np.min((BASE[None, :] + self.E - self.lost(line) - FLOOR) / SHAPE[None, :], axis=1)
return self._crit[key]
def p(self, line, phi): # the chance of a breach in some quarter
return float(np.mean(self.crit(line) < DEPTH0 - DPQ * phi))
def delay_for(self, line, p=ALPHA): # the least delay meeting the chance p (NaN: no delay suffices)
key = round(float(line), 6)
if key not in self._sorted:
self._sorted[key] = np.sort(self.crit(line))
d = min(DEPTH0, self._sorted[key][int(np.floor(p * self.n + 1e-9))])
return np.nan if d < 0 else (DEPTH0 - d) / DPQ
def cost(line, phi):
return FEE * line + VQ * phi
def cheapest(P, p=ALPHA, lines=LINES):
# Chapter 9's design: the cheapest committed package keeping the chance p on P's paths
best = None
for line in lines:
phi = P.delay_for(line, p)
if not np.isnan(phi) and (best is None or cost(line, phi) < best[2]):
best = (float(line), float(phi), float(cost(line, phi)))
return best
_EXTRA = {}
def extra_stream(P):
# Remark 11.B.1: each quarter, uniforms for the paths' own closures; then the extra stream at rate 0.40 - 0.10
# (counts, gamma sizes, uniforms); seed 1108 on the book's paths, 26208 on the fresh paths
if P.seed not in _EXTRA:
rng = np.random.default_rng(EXTRA_SEED[P.seed])
own, more = [], []
for k in range(Q):
path, size = P.events[k]
own.append(rng.random(len(path)))
count = rng.poisson((LAM_MAX - LAM) / 4.0, P.n)
jj = np.repeat(np.arange(P.n), count)
more.append((jj, rng.gamma(*GAMMA, len(jj)), rng.random(len(jj))))
_EXTRA[P.seed] = (own, more)
return _EXTRA[P.seed]
def law(P, lam=LAM, sigma=50.0):
# The law with closures at lam a year and a spread of sigma ($m a quarter), on P's paths (Remark 11.B.1)
c = sigma / 50.0
if abs(lam - LAM) < 1e-15:
return P if c == 1.0 else Paths(P.E * c, P.events, P.seed)
own, more = extra_stream(P)
events = []
for k, (path, size) in enumerate(P.events):
if lam < LAM: # thinning: keep each own closure with chance lam / 0.10
keep = own[k] < lam / LAM
events.append((path[keep], size[keep]))
else: # superposition: add the extra closures whose uniform is below the share
jj, sz, u = more[k]
add = u < (lam - LAM) / (LAM_MAX - LAM)
events.append((np.concatenate([path, jj[add]]), np.concatenate([size, sz[add]])))
return Paths(P.E * c, events, P.seed)
def robust10(D, S):
# Chapter 10, Proposition 10.1(ii): the cheapest package robust on deviations within D and one closure of at most S
phi = max(0.0, float(np.max((D - MARGIN) / LIFT)))
u = min(S, float(np.min(MARGIN - D + LIFT * phi)))
return S - u, phi
BOOK, FRESH = Paths.draw(BOOK_SEED), Paths.draw(FRESH_SEED)
PKG = cheapest(BOOK) # Chapter 9: the cheapest package keeping 5 percent under the law
PKG_P, PKG_F = BOOK.p(*PKG[:2]), FRESH.p(*PKG[:2])
D10, S10 = 200.0 - 150.0 * 0.38, 300.0 - 180.0 * 0.38 # Chapter 10's max-min cut, 0.38
CH10 = robust10(D10, S10)
plan = 1000.0 * (BASE - DEPTH0 * SHAPE)
print(f"planned liquidity dips to {plan.min():.0f} in quarter {plan.argmin() + 1}; a quarter of delay lifts it by "
f"{LIFT[plan.argmin()]:.1f} there")
print(f"Volume I's package (Chapter 9's design): line {PKG[0]:.0f}, delay {PKG[1]:.4f}, cost {PKG[2]:.2f}; "
f"under the law {100 * PKG_P:.2f}% (fresh {100 * PKG_F:.2f}%)")
print(f"Chapter 10's max-min package (deviations within {D10:.1f}, a closure within {S10:.1f}): line {CH10[0]:.1f}, "
f"delay {CH10[1]:.3f}, cost {cost(*CH10):.2f}; under the law {100 * BOOK.p(*CH10):.2f}%")planned liquidity dips to 482 in quarter 7; a quarter of delay lifts it by 24.6 there Volume I's package (Chapter 9's design): line 170, delay 0.6676, cost 10.44; under the law 5.00% (fresh 4.91%) Chapter 10's max-min package (deviations within 143.0, a closure within 231.6): line 231.6, delay 0.449, cost 10.54; under the law 5.64%
Panel 1 — Read the records (Proposition 11.2(i)–(iii); Worked Example 11.1, Step 1; Online Appendix Example 11.B.2: rate at most 0.3148 a year; spread from $42.27m to $61.60m; Volume I's package from 4.47% to 6.35% by rate and from 1.03% to 17.87% by spread; 15.79% at the corner)
AXIOM Lab 2.11, step 1. With events in years, the rate is at most with confidence (Garwood's bound, Proposition 11.2(i)); with Gaussian innovations of root-mean-square , the spread is at most and at least , each with confidence (part (ii)). Two closures in twenty years and 39 innovations of $50 million give the records' bounds at 95 percent each, and 90 percent together by Bonferroni (part (iii)); a record twice as long, at the same estimates, gives narrower ones. Volume I's package is then judged under the laws the records allow, one doubt at a time, and at the corner, both bounds at once: more frequent closures only remove liquidity and a wider spread raises the chance of a breach (checked across the family in Panel 2), so the family's worst law is its corner. Last, the lab's own draws (seed 26211): 100,000 records simulated under the declared law, to see how often the bounds hold the true rate and spread.
Python Code
def rate_bounds(n, years, conf=0.95):
# Proposition 11.2(i): Garwood's one-sided bounds for a Poisson rate from n events in `years` years
lo = stats.chi2.ppf(1.0 - conf, 2 * n) / (2.0 * years) if n > 0 else 0.0
return lo, stats.chi2.ppf(conf, 2 * n + 2) / (2.0 * years)
def spread_bounds(df, s=50.0, conf=0.95):
# Proposition 11.2(ii): one-sided bounds for a spread from df Gaussian innovations of root-mean-square s
return s * math.sqrt(df / stats.chi2.ppf(conf, df)), s * math.sqrt(df / stats.chi2.ppf(1.0 - conf, df))
LAM_L, LAM_U = rate_bounds(2, 20.0) # the treasury: two closures in twenty years
SIG_L, SIG_U = spread_bounds(39) # the engineering office: forty quarters, 39 innovations
LAM_U2 = rate_bounds(4, 40.0)[1] # twice the record: four closures in forty years
SIG_L2, SIG_U2 = spread_bounds(79) # eighty quarters, 79 innovations
print(f"rate: two closures in twenty years -> estimate {2 / 20:.2f}; one-sided 95% bounds {LAM_L:.4f} and {LAM_U:.4f} a year")
print(f"spread: 39 innovations of $50m -> one-sided 95% bounds {SIG_L:.2f} and {SIG_U:.2f}")
print(f"together (Bonferroni): at least {100 * (1 - 0.05 - 0.05):.0f}% confidence")
print(f"twice the record (four in forty years; 79 innovations): rate at most {LAM_U2:.4f}; spread at most {SIG_U2:.2f} "
f"(at least {SIG_L2:.2f})")
RATES = [0.0, LAM_L, 0.05, LAM, 0.15, 0.2, 0.25, LAM_U, 0.35]
SPREADS = [40.0, SIG_L, 45.0, 50.0, 55.0, 60.0, SIG_U, 65.0]
BY_RATE = [law(BOOK, lam).p(*PKG[:2]) for lam in RATES]
BY_SPREAD = [law(BOOK, LAM, s).p(*PKG[:2]) for s in SPREADS]
CORNER, CORNER_F = law(BOOK, LAM_U, SIG_U), law(FRESH, LAM_U, SIG_U) # the records' worst law, on both sets of paths
P_CORNER, P_CORNER_F = CORNER.p(*PKG[:2]), CORNER_F.p(*PKG[:2])
print("\nVolume I's package under the laws the records allow (spread $50m):")
print(" rate a year " + " ".join(f"{x:6.3f}" for x in RATES))
print(" breach (%) " + " ".join(f"{100 * x:6.2f}" for x in BY_RATE))
print("and at the declared rate, 0.10:")
print(" spread ($m) " + " ".join(f"{x:6.1f}" for x in SPREADS))
print(" breach (%) " + " ".join(f"{100 * x:6.2f}" for x in BY_SPREAD))
print(f"at the corner (rate {LAM_U:.4f}, spread {SIG_U:.1f}): {100 * P_CORNER:.2f}% (fresh {100 * P_CORNER_F:.2f}%), "
f"against {100 * PKG_P:.2f}% under the declared law")
# the lab's own draws: records simulated under the declared law (seed 26211)
rng = np.random.default_rng(SEED)
M = 100_000
n_closures = rng.poisson(LAM * 20.0, M)
rms = np.sqrt((rng.normal(0.0, 50.0, (M, 39)) ** 2).mean(axis=1))
rate_holds = stats.chi2.ppf(0.95, 2 * n_closures + 2) / 40.0 >= LAM
spread_holds = rms * math.sqrt(39 / stats.chi2.ppf(0.05, 39)) >= 50.0
COVER = dict(rate=float(rate_holds.mean()), spread=float(spread_holds.mean()), both=float((rate_holds & spread_holds).mean()))
print(f"\n{M:,} records simulated under the declared law (the lab's seed {SEED}): the rate's bound holds the true rate in "
f"{100 * COVER['rate']:.1f}%, the spread's in {100 * COVER['spread']:.1f}%, both in {100 * COVER['both']:.1f}%")
print(f" with no closure at all the rate's bound is {rate_bounds(0, 20.0)[1]:.3f}, above 0.10: Garwood's bound errs on the "
f"safe side; at a true rate of 0.20 it holds with chance "
f"{100 * stats.poisson.sf(0, 0.20 * 20):.1f}%, at 0.3148 with {100 * stats.poisson.sf(1, LAM_U * 20):.1f}%")rate: two closures in twenty years -> estimate 0.10; one-sided 95% bounds 0.0178 and 0.3148 a year spread: 39 innovations of $50m -> one-sided 95% bounds 42.27 and 61.60 together (Bonferroni): at least 90% confidence twice the record (four in forty years; 79 innovations): rate at most 0.2288; spread at most 57.60 (at least 44.28) Volume I's package under the laws the records allow (spread $50m): rate a year 0.000 0.018 0.050 0.100 0.150 0.200 0.250 0.315 0.350 breach (%) 4.47 4.57 4.73 5.00 5.27 5.53 5.80 6.16 6.35 and at the declared rate, 0.10: spread ($m) 40.0 42.3 45.0 50.0 55.0 60.0 61.6 65.0 breach (%) 1.03 1.59 2.51 5.00 8.58 12.92 14.43 17.87 at the corner (rate 0.3148, spread 61.6): 15.79% (fresh 15.83%), against 5.00% under the declared law 100,000 records simulated under the declared law (the lab's seed 26211): the rate's bound holds the true rate in 100.0%, the spread's in 95.0%, both in 95.0% with no closure at all the rate's bound is 0.150, above 0.10: Garwood's bound errs on the safe side; at a true rate of 0.20 it holds with chance 98.2%, at 0.3148 with 98.7%
Panel 2 — Price the family (Theorem 11.1(iv); Proposition 11.2(iii); Worked Example 11.1, Step 2; Figure 11.3(a): a $200m line and 2.25 quarters, $24.00m; $11.64m for the closures' doubt alone and $22.85m for the spread's; at most 5.0% on a grid of 49 laws; 0.9% under the declared law)
AXIOM Lab 2.11, step 2. For each declaration of doubt, the cheapest package that keeps 5 percent at its worst law is Chapter 9's design at that law: the line on a $5 million grid up to $400 million and the delay continuous, the least delay at each line being a quantile of the paths' critical depths. Four declarations: the closures' doubt alone (rate 0.3148, spread $50 million), the spread's doubt alone (rate 0.10, spread $61.6 million), the records' family (both bounds) and a record twice as long (rate 0.2288, spread $57.6 million). Each package is checked on the fresh paths at its law and judged under the declared law. The family's package is then checked across the family, on a grid of 49 laws (seven rates from 0 to 0.315, seven spreads from $42.3 to $61.6 million). By Theorem 11.1(iv), a family that holds the true law with 90 percent confidence makes its package's promise hold with 90 percent confidence.
Python Code
def family_case(lam, sigma):
B, F = law(BOOK, lam, sigma), law(FRESH, lam, sigma)
d = cheapest(B)
return dict(lam=lam, sigma=sigma, p_pkg=B.p(*PKG[:2]), p_pkg_fresh=F.p(*PKG[:2]), design=d, p_fresh=F.p(*d[:2]),
p_law=BOOK.p(*d[:2]), p_law_fresh=FRESH.p(*d[:2]))
CASES = {"rate": family_case(LAM_U, 50.0), "spread": family_case(LAM, SIG_U), "both": family_case(LAM_U, SIG_U),
"longer": family_case(LAM_U2, SIG_U2)}
NAMES = {"rate": "closures' doubt", "spread": "spread's doubt", "both": "the records' family", "longer": "twice the record"}
print(f"{'declaration':22s} {'rate':>6s} {'spread':>6s} | {'Vol I pkg':>9s} {'fresh':>6s} | {'line':>5s} {'delay':>6s} "
f"{'cost':>6s} {'fresh':>6s} | {'declared law':>12s} {'fresh':>6s}")
for key, c in CASES.items():
l_, ph, co = c["design"]
print(f"{NAMES[key]:22s} {c['lam']:6.4f} {c['sigma']:6.1f} | {100 * c['p_pkg']:8.2f}% {100 * c['p_pkg_fresh']:5.2f}% | "
f"{l_:5.0f} {ph:6.3f} {co:6.2f} {100 * c['p_fresh']:5.2f}% | {100 * c['p_law']:11.2f}% {100 * c['p_law_fresh']:5.2f}%")
FAM = CASES["both"]["design"]
GRID = [(float(lam), float(s), law(BOOK, lam, s).p(*FAM[:2])) for lam in np.linspace(0.0, LAM_U, 7)
for s in np.linspace(SIG_L, SIG_U, 7)]
worst_law = max(GRID, key=lambda g: g[2])
print(f"\nthe family's package on the grid of 49 laws: largest chance {100 * worst_law[2]:.2f}% at rate {worst_law[0]:.4f}, "
f"spread {worst_law[1]:.1f} (the corner)")
print(f"the spread is what costs: the closures' doubt adds {CASES['rate']['design'][2] - PKG[2]:.2f} to Volume I's "
f"{PKG[2]:.2f}; the spread's {CASES['spread']['design'][2] - PKG[2]:.2f}; both {FAM[2] - PKG[2]:.2f}")
sg = np.linspace(40.0, 65.0, 26)
curves = {(pk, lam): [100 * law(BOOK, lam, s).p(*pk[:2]) for s in sg] for pk in (PKG, FAM) for lam in (LAM, LAM_U)}
fig, ax = plt.subplots(figsize=(7.6, 3.6))
for pk, col, name in ((PKG, "#1F4E79", "Volume I's package"), (FAM, "#E65100", "the records' family's package")):
ax.plot(sg, curves[(pk, LAM)], color=col, lw=1.8, label=f"{name}, closures 0.10 a year")
ax.plot(sg, curves[(pk, LAM_U)], color=col, lw=1.8, ls="--", label=f"{name}, closures 0.315 a year")
ax.axhline(5.0, color="#999999", lw=0.8, ls=":")
for v, txt in ((50.0, "declared"), (SIG_U, "bound")):
ax.axvline(v, color="#999999", lw=0.8, ls="--")
ax.text(v + 0.3, 19.3, txt, fontsize=8, color="#616161", va="top")
ax.annotate(f"{100 * P_CORNER:.1f}% at the corner", xy=(SIG_U, 100 * P_CORNER), xytext=(52.5, 16.5), fontsize=8,
arrowprops=dict(arrowstyle="-", lw=0.6, color="#333333"))
ax.set(xlim=(40, 65), ylim=(0, 20), xlabel="spread of deviations ($ million a quarter)", ylabel="chance of a breach (%)",
title="Volume I's package breaches on 15.8% at the records' corner (seed 26211)")
ax.legend(frameon=False, fontsize=7.5, loc="upper left"); plt.tight_layout(); plt.show()declaration rate spread | Vol I pkg fresh | line delay cost fresh | declared law fresh closures' doubt 0.3148 50.0 | 6.16% 6.06% | 210 0.667 11.64 4.88% | 4.63% 4.53% spread's doubt 0.1000 61.6 | 14.43% 14.44% | 170 2.218 22.85 4.91% | 1.06% 1.03% the records' family 0.3148 61.6 | 15.79% 15.83% | 200 2.250 24.00 4.91% | 0.89% 0.86% twice the record 0.2288 57.6 | 11.60% 11.43% | 195 1.690 19.37 4.87% | 1.66% 1.62% the family's package on the grid of 49 laws: largest chance 5.00% at rate 0.3148, spread 61.6 (the corner) the spread is what costs: the closures' doubt adds 1.20 to Volume I's 10.44; the spread's 12.41; both 13.56

Panel 3 — Try balls: divergence (Theorem 11.2; Figure 11.1(a); Worked Example 11.1, Step 3: radii 0.439, 0.603 and 1.042; Volume I's package at 8.36% in the ball of 0.01; 2.5% under the law, a $175m line and 1.354 quarters, $16.08m; 3 in a million at 0.44; 2 in 100 billion at 1.04)
AXIOM Lab 2.11, step 3, first part. A Kullback–Leibler ball of radius around the declared law holds every law that reweights its futures with . The worst chance over the ball of an event of chance is the with (Theorem 11.2(ii)); so the ball keeps 5 percent exactly when the law keeps the smaller chance with , and the ball's package is Chapter 9's design at the appetite — no new paths are needed. A ball must hold the doubts it is meant for: by part (iii), closures at instead of over years lie at , twelve innovations with their spread scaled by at , and the two add. The levers' reach ends where the law's appetite falls below one path in 100,000.
Python Code
def kl(q, p):
# Theorem 11.2(ii): kl(q || p), the divergence of a chance q from a chance p
out = 0.0
if q > 0:
out += q * math.log(q / p)
if q < 1:
out += (1 - q) * math.log((1 - q) / (1 - p))
return out
def worst_chance(p, d):
# q*(p, d): the worst chance over the ball of radius d of an event of chance p under the law
if d <= 0:
return p
if kl(1.0 - 1e-15, p) <= d:
return 1.0
return optimize.brentq(lambda q: kl(q, p) - d, p, 1.0 - 1e-15, xtol=1e-15)
def appetite(d, a=ALPHA):
# alpha'(d): the chance the law must keep for every law in the ball to keep a
if d <= 0:
return a
return optimize.brentq(lambda p: kl(a, p) - d, 1e-300, a, xtol=1e-300, rtol=1e-14)
def kl_rate(lam, lam0=LAM, years=YEARS): # Theorem 11.2(iii): closures at lam instead of lam0 over the program
return years * (lam * math.log(lam / lam0) - lam + lam0)
def kl_spread(c, n=Q): # n Gaussian innovations whose spread is scaled by c
return n * (math.log(1.0 / c) + (c * c - 1.0) / 2.0)
D_RATE, D_SPREAD = kl_rate(LAM_U), kl_spread(SIG_U / 50.0)
RADII = {"closures' doubt": D_RATE, "spread's doubt": D_SPREAD, "both doubts": D_RATE + D_SPREAD}
print(f"closures at {LAM_U:.4f}: {YEARS:.0f} x ({LAM_U:.4f} log({LAM_U:.4f} / 0.10) - {LAM_U - LAM:.4f}) = {D_RATE:.3f}")
print(f"spread x {SIG_U / 50:.3f}: {Q} x (log(1 / {SIG_U / 50:.3f}) + ({SIG_U / 50:.3f}^2 - 1) / 2) = {D_SPREAD:.3f}")
for k, d in RADII.items():
print(f" {k:16s} radius {d:.3f}: to keep 5% over that ball, the law must keep {appetite(d):.3g}")
P_MIN = BOOK.p(400.0, MAX_DELAY)
D_LEVERS = kl(ALPHA, P_MIN)
print(f"a $400m line and 7 quarters of delay breach on {P_MIN * N:.0f} path of {N:,}: beyond the levers from "
f"kl(0.05 || {P_MIN:g}) = {D_LEVERS:.3f}")
KL_WORST = {d: worst_chance(PKG_P, d) for d in (0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1.0)}
print("\nworst chance of Volume I's package over the ball (it keeps 5.00% under the law):")
print(" d " + " ".join(f"{d:6g}" for d in KL_WORST))
print(" worst (%) " + " ".join(f"{100 * q:6.2f}" for q in KL_WORST.values()))
KL_D = (0.0, 0.001, 0.002, 0.005, 0.01, 0.02, 0.03, 0.05, 0.075, 0.1, 0.15)
KL_DESIGNS = {}
print("\nkeeping 5% over the ball: d, the law's appetite, line, delay, cost; under the law (fresh)")
for d in KL_D:
a = appetite(d)
b = cheapest(BOOK, p=a)
KL_DESIGNS[d] = dict(alpha=a, design=b, p=BOOK.p(*b[:2]), p_fresh=FRESH.p(*b[:2]))
print(f" {d:5g} {100 * a:5.2f}% {b[0]:4.0f} {b[1]:6.3f} {b[2]:6.2f} {100 * KL_DESIGNS[d]['p']:5.2f}% "
f"({100 * KL_DESIGNS[d]['p_fresh']:.2f}%)")
for d in (0.01, 0.03):
b = KL_DESIGNS[d]["design"]
print(f"the ball of {d}: its package at the records' corner breaches on {100 * CORNER.p(*b[:2]):.2f}% "
f"(fresh {100 * CORNER_F.p(*b[:2]):.2f}%)")
ds = np.exp(np.linspace(np.log(1e-3), np.log(1.5), 200))
fig, ax = plt.subplots(figsize=(7.6, 3.6))
ax.loglog(ds, [100 * worst_chance(PKG_P, d) for d in ds], color="#B71C1C", lw=1.8,
label="worst chance of Volume I's package over the ball")
ax.loglog(ds, [100 * appetite(d) for d in ds], color="#1F4E79", lw=1.8, label="chance the law must keep for the ball to keep 5%")
ax.axhline(5.0, color="#999999", lw=0.8, ls=":")
ax.axhline(1e-3, color="#999999", lw=0.8, ls=":")
ax.text(1.15e-3, 1.4e-3, "one path in 100,000: the levers' floor", fontsize=8, color="#616161")
for d, txt in zip(RADII.values(), ("closures", "spread", "both")):
ax.axvline(d, color="#E65100", lw=0.9, ls="--")
ax.text(d / 1.04, 1.5e-7, f"{txt} {d:.2f}", fontsize=7.5, color="#E65100", rotation=90, ha="right", va="bottom")
ax.set(xlim=(1e-3, 1.5), ylim=(1e-7, 300), xlabel="radius of the ball, d", ylabel="chance of a breach (%)",
title="Keeping 5% over a ball of 0.01 means 2.5% under the law (seed 26211)")
ax.set_yticks([1e-6, 1e-4, 1e-2, 1, 100], ["0.000001", "0.0001", "0.01", "1", "100"])
ax.set_xticks([1e-3, 1e-2, 1e-1, 1], ["0.001", "0.01", "0.1", "1"])
ax.legend(frameon=False, fontsize=8, loc="lower left"); plt.tight_layout(); plt.show()closures at 0.3148: 3 x (0.3148 log(0.3148 / 0.10) - 0.2148) = 0.439
spread x 1.232: 12 x (log(1 / 1.232) + (1.232^2 - 1) / 2) = 0.603
closures' doubt radius 0.439: to keep 5% over that ball, the law must keep 2.93e-06
spread's doubt radius 0.603: to keep 5% over that ball, the law must keep 1.09e-07
both doubts radius 1.042: to keep 5% over that ball, the law must keep 1.69e-11
a $400m line and 7 quarters of delay breach on 1 path of 100,000: beyond the levers from kl(0.05 || 1e-05) = 0.377
worst chance of Volume I's package over the ball (it keeps 5.00% under the law):
d 0.001 0.003 0.01 0.03 0.1 0.3 1
worst (%) 6.00 6.78 8.36 11.15 17.29 28.85 55.55
keeping 5% over the ball: d, the law's appetite, line, delay, cost; under the law (fresh)
0 5.00% 170 0.668 10.44 5.00% (4.91%)
0.001 4.08% 160 0.924 12.19 4.08% (4.05%)
0.002 3.74% 165 0.998 12.93 3.74% (3.68%)
0.005 3.11% 170 1.162 14.40 3.11% (3.09%)
0.01 2.50% 175 1.354 16.08 2.50% (2.51%)
0.02 1.78% 180 1.671 18.77 1.78% (1.75%)
0.03 1.34% 175 1.955 20.89 1.34% (1.30%)
0.05 0.81% 195 2.356 24.70 0.81% (0.79%)
0.075 0.46% 215 2.737 28.35 0.46% (0.46%)
0.1 0.27% 200 3.173 31.39 0.27% (0.30%)
0.15 0.10% 240 3.851 38.01 0.10% (0.11%)
the ball of 0.01: its package at the records' corner breaches on 10.25% (fresh 10.08%)
the ball of 0.03: its package at the records' corner breaches on 6.87% (fresh 6.71%)

Panel 4 — Try balls: Wasserstein (Theorem 11.3(ii)–(iii); Figure 11.2(a); Worked Example 11.1, Step 3: Volume I's package at 7.11% in the ball of $0.1m; a $165m line and 1.148 quarters, $14.13m, 3.23% under the law; the spread's doubt at $26.3m, asking $525m of the thinnest paths, against $121.9m within the levers' reach)
AXIOM Lab 2.11, step 3, second part. A Wasserstein ball moves futures instead of reweighting them. Let a path move by shifting its deviations, its closures fixed, at a cost equal to the largest shift in any quarter: a path's distance from a breach is then its thinnest margin above the floor over the twelve quarters. The worst chance over the ball of radius moves the paths nearest the floor first, a fractional knapsack with budget (Theorem 11.3(iii)); so the ball keeps 5 percent exactly when the thinnest 5 percent of paths average at least above the floor. A package for the ball needs, at each line, the least delay that gives them that average: margins only rise with the delay, so a bisection on the delay finds it (34 halvings of the 7 quarters, as the book's engine), evaluating at each step only the paths that can still be among the thinnest. Scaling every deviation by moves each path by times its largest deviation, so the ball that holds the spread's doubt has radius times the paths' average largest deviation (Online Appendix Example 11.B.2).
Python Code
def margin_parts(P, line):
# each path's margin above the floor in each quarter with no delay ($m)
return 1000.0 * (BASE[None, :] - DEPTH0 * SHAPE[None, :] + P.E - P.lost(line) - FLOOR)
def thinnest(A, phi):
# a path's distance from a breach: its thinnest margin over the quarters ($m)
return np.min(A + LIFT[None, :] * phi, axis=1)
def worst_chance_w(m, eps):
# Theorem 11.3(iii): the share of paths a budget of N eps can push below the floor, nearest first
d = np.sort(np.maximum(m, 0.0))
n = len(d)
cs = np.cumsum(d)
j = int(np.searchsorted(cs, n * eps, side="right"))
if j >= n:
return 1.0
frac = (n * eps - (cs[j - 1] if j > 0 else 0.0)) / d[j] if d[j] > 0 else 1.0
return (j + min(frac, 1.0)) / n
def thin_mean(m, a=ALPHA):
# the average margin (floored at zero) of the thinnest share a of the paths
return float(np.mean(np.sort(np.maximum(m, 0.0))[:int(round(a * len(m)))]))
def least_delays(A, radii, p=ALPHA, steps=34):
# For each radius: the least delay at which the K = p n thinnest paths' margins sum to at least n * radius,
# by bisection on [0, 7] quarters; NaN if 7 quarters do not suffice. After each halving, only the paths whose
# margin at the interval's low end is within the K-th thinnest margin at its high end can matter.
n = A.shape[0]
K = int(round(p * n))
need = n * np.asarray(radii, dtype=float)
def thin_sum(Ac, phis): # sum of the K thinnest (in increasing order), the K-th, all margins
d = np.maximum(np.min(Ac[None, :, :] + LIFT[None, None, :] * np.asarray(phis)[:, None, None], axis=2), 0.0)
part = np.partition(d, K - 1, axis=1)[:, :K]
return np.cumsum(np.sort(part, axis=1), axis=1)[:, -1], part.max(axis=1), d
(s0, s7), (_, v7), (d0, _) = thin_sum(A, [0.0, MAX_DELAY])
out = np.where(s7 >= need, np.where(s0 >= need, 0.0, MAX_DELAY), np.nan)
act = np.nonzero((s7 >= need) & (s0 < need))[0]
if len(act) == 0:
return out
need = need[act]
lo, hi, v_hi = np.zeros(len(act)), np.full(len(act), MAX_DELAY), np.full(len(act), v7)
m_lo = np.repeat(d0[None, :], len(act), axis=0)
keep = (m_lo <= v_hi[:, None]).any(axis=0)
Ac, m_lo = A[keep], m_lo[:, keep]
for _ in range(steps):
mid = 0.5 * (lo + hi)
u, back = np.unique(mid, return_inverse=True) # radii still sharing an interval share the work
s, v, d = (x[back] for x in thin_sum(Ac, u))
ok = s >= need
hi, v_hi = np.where(ok, mid, hi), np.where(ok, v, v_hi)
lo, m_lo = np.where(ok, lo, mid), np.where(ok[:, None], m_lo, d)
keep = (m_lo <= v_hi[:, None]).any(axis=0)
Ac, m_lo = Ac[keep], m_lo[:, keep]
out[act] = hi
return out
def design_w(P, radii, p=ALPHA, lines=np.arange(100.0, 300.01, 5.0), steps=34):
# the cheapest package keeping p over the Wasserstein ball of each radius
best = [None] * len(radii)
for line in lines:
for r, phi in enumerate(least_delays(margin_parts(P, line), radii, p, steps)):
if not np.isnan(phi) and (best[r] is None or cost(line, phi) < best[r][2]):
best[r] = (float(line), float(phi), float(cost(line, phi)))
return best
M_PKG = thinnest(margin_parts(BOOK, PKG[0]), PKG[1])
W_WORST = {e: worst_chance_w(M_PKG, e) for e in (0.01, 0.05, 0.1, 0.2, 0.5, 1.0)}
print(f"Volume I's package: {100 * PKG_P:.2f}% of the paths breach, and the next 1% sit just above the floor, "
f"{np.sort(M_PKG)[5000:6000].mean():.2f} on average")
print("worst chance over the ball ($m radius): " + " ".join(f"{e:g}: {100 * q:.2f}%" for e, q in W_WORST.items()))
W_RADII = (0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1.0)
W_DESIGNS = {}
print("\nkeeping 5% over the ball: radius, line, delay, cost; under the law (fresh); the thinnest 5% average (needed)")
for e, b in zip(W_RADII, design_w(BOOK, W_RADII)):
W_DESIGNS[e] = dict(design=b, p=BOOK.p(*b[:2]), p_fresh=FRESH.p(*b[:2]),
thin=thin_mean(thinnest(margin_parts(BOOK, b[0]), b[1])))
w = W_DESIGNS[e]
print(f" {e:4g} {b[0]:4.0f} {b[1]:6.3f} {b[2]:6.2f} {100 * w['p']:5.2f}% ({100 * w['p_fresh']:.2f}%) "
f"{w['thin']:6.2f} ({e / ALPHA:.1f})")
for e in (0.1, 1.0):
b = W_DESIGNS[e]["design"]
print(f"the ball of ${e:g}m: its package at the records' corner breaches on {100 * CORNER.p(*b[:2]):.2f}% "
f"(fresh {100 * CORNER_F.p(*b[:2]):.2f}%)")
W_MAXDEV = 1000.0 * float(np.mean(np.max(np.abs(BOOK.E), axis=1))) # the paths' average largest deviation ($m)
W_SPREAD = (SIG_U / 50.0 - 1.0) * W_MAXDEV
BEST_THIN = thin_mean(thinnest(margin_parts(BOOK, 400.0), MAX_DELAY))
print(f"\nthe spread's doubt lies at {SIG_U / 50.0 - 1.0:.3f} x {W_MAXDEV:.1f} = {W_SPREAD:.1f}: the thinnest 5% would need "
f"to average {W_SPREAD / ALPHA:.0f}; a $400m line and 7 quarters give {BEST_THIN:.1f}")
mw = np.sort(thinnest(margin_parts(BOOK, W_DESIGNS[0.1]["design"][0]), W_DESIGNS[0.1]["design"][1]))
x = 100.0 * (np.arange(N) + 0.5) / N
sel = slice(0, int(0.12 * N))
pushed = (x >= 5.0) & (x <= 100 * W_WORST[0.1])
mp = np.sort(M_PKG)
fig, ax = plt.subplots(figsize=(7.6, 3.6))
ax.axhline(0.0, color="#999999", lw=0.8)
ax.fill_between(x[pushed], 0.0, np.maximum(mp[pushed], 0.0), color="#B71C1C", alpha=0.3, lw=0,
label=f"pushed below the floor by a radius of \\$0.1m: 5.0% to {100 * W_WORST[0.1]:.1f}%")
ax.plot(x[sel], mp[sel], color="#1F4E79", lw=1.8, label="Volume I's package")
ax.plot(x[sel], mw[sel], color="#E65100", lw=1.8, label="the package keeping 5% over that ball: \\$14.13m")
ax.set(xlim=(0, 12), ylim=(-32, 32), xlabel="paths, thinnest first (%)", ylabel="thinnest margin ($ million)",
title="A ball of \\$0.1m lifts the chance from 5.0% to 7.1% (seed 26211)")
ax.legend(frameon=False, fontsize=8, loc="lower right"); plt.tight_layout(); plt.show()Volume I's package: 5.00% of the paths breach, and the next 1% sit just above the floor, 2.39 on average
worst chance over the ball ($m radius): 0.01: 5.64% 0.05: 6.47% 0.1: 7.11% 0.2: 8.05% 0.5: 10.00% 1: 12.32%
keeping 5% over the ball: radius, line, delay, cost; under the law (fresh); the thinnest 5% average (needed)
0.01 165 0.827 11.57 4.40% (4.32%) 0.20 (0.2)
0.02 165 0.887 12.04 4.17% (4.11%) 0.40 (0.4)
0.05 165 1.007 13.00 3.71% (3.65%) 1.00 (1.0)
0.1 165 1.148 14.13 3.23% (3.21%) 2.00 (2.0)
0.2 170 1.335 15.78 2.61% (2.61%) 4.00 (4.0)
0.5 175 1.761 19.34 1.67% (1.61%) 10.00 (10.0)
1 175 2.337 23.95 0.91% (0.89%) 20.00 (20.0)
the ball of $0.1m: its package at the records' corner breaches on 12.15% (fresh 12.05%)
the ball of $1m: its package at the records' corner breaches on 5.23% (fresh 5.15%)
the spread's doubt lies at 0.232 x 113.2 = 26.3: the thinnest 5% would need to average 525; a $400m line and 7 quarters give 121.9

Panel 5 — Set a radius from a record (Theorem 11.3(iv); Proposition 11.2(iv); Figure 11.2(b); Worked Example 11.1, Step 4: designed on 1,000 paths, 38.5% keep 5% at $10.19m; a one-point margin, 87.0% at $12.09m; a radius of $20,000, 86.5% at $11.94m; of $50,000, 97.5% at $12.94m)
AXIOM Lab 2.11, step 4. A ball fits the sampling error of a record. As in Chapter 9's sample-size study (Online Appendix 9.B.2), 200 records of 1,000 paths each (seeds 900000–900199) are drawn from the declared law. On each, a package is designed two ways, among Chapter 9's lines ($120–220 million on the $5 million grid): aimed at a lower target, (a margin), or kept at 5 percent over a Wasserstein ball around the record (a radius, the bisection run with 30 halvings, as the book's engine runs it). Each design's true chance is read on the fresh 100,000 paths. A design fitted to its own small record looks better than it is (Proposition 11.2(iv)); a radius, like a margin, buys reliability, and says what it protects against.
Python Code
CH9_LINES = np.arange(120.0, 220.01, 5.0) # Chapter 9's lines for the sample-size study
RECORDS = [Paths.draw(900000 + r, n=1000) for r in range(200)]
REC_MARGINS = (0.0, 0.005, 0.01, 0.015, 0.02)
REC_RADII = (0.01, 0.02, 0.03, 0.05, 0.1, 0.2)
rows_m = {mg: [] for mg in REC_MARGINS}
rows_w = {e: [] for e in REC_RADII}
for P in RECORDS:
for mg in REC_MARGINS:
b = cheapest(P, p=ALPHA - mg, lines=CH9_LINES)
rows_m[mg].append((FRESH.p(*b[:2]), b[2], b[0]))
for e, b in zip(REC_RADII, design_w(P, REC_RADII, lines=CH9_LINES, steps=30)):
rows_w[e].append((FRESH.p(*b[:2]), b[2], b[0]))
def summary(rows):
tp, cs, ln = (np.array(v) for v in zip(*rows))
return dict(meet=float(np.mean(tp <= ALPHA)), mean_true=float(tp.mean()), q90=float(np.quantile(tp, 0.9)),
mean_cost=float(cs.mean()), mean_line=float(ln.mean()), reps=len(tp))
REC_M = {mg: summary(r) for mg, r in rows_m.items()}
REC_W = {e: summary(r) for e, r in rows_w.items()}
print("designed on a record of 1,000 paths, judged on the fresh 100,000:")
print(" keep 5% mean true chance 90th percentile mean cost")
for mg, s in REC_M.items():
print(f" target lower by {100 * mg:3.1f} pt {100 * s['meet']:6.1f}% {100 * s['mean_true']:12.2f}% "
f"{100 * s['q90']:16.2f}% {s['mean_cost']:11.2f}")
for e, s in REC_W.items():
print(f" radius ${1e6 * e:7,.0f} {100 * s['meet']:6.1f}% {100 * s['mean_true']:12.2f}% "
f"{100 * s['q90']:16.2f}% {s['mean_cost']:11.2f}")
print(f"the book's 100,000 paths, Volume I's package: fresh {100 * PKG_F:.2f}% at {PKG[2]:.2f}")
fig, ax = plt.subplots(figsize=(7.6, 3.6))
sa = list(REC_M.values())
wa = [REC_M[0.0]] + list(REC_W.values())
ax.plot([s["mean_cost"] for s in sa], [100 * s["meet"] for s in sa], "o-", color="#1F4E79", lw=1.6, ms=4,
label="lower the target (Chapter 9)")
ax.plot([s["mean_cost"] for s in wa], [100 * s["meet"] for s in wa], "s-", color="#E65100", lw=1.6, ms=4,
label="a Wasserstein radius")
for mg in (0.005, 0.01, 0.015):
s = REC_M[mg]
ax.text(s["mean_cost"] + 0.12, 100 * s["meet"] - 4.0, f"{100 * mg:g} pt", fontsize=8, color="#1F4E79")
for e in (0.02, 0.05, 0.1):
s = REC_W[e]
ax.text(s["mean_cost"] - 0.12, 100 * s["meet"] + 3.0, f"\\${1000 * e:.0f}k", fontsize=8, color="#E65100", ha="right")
ax.annotate(f"no margin: {100 * REC_M[0.0]['meet']:.1f}%", xy=(REC_M[0.0]["mean_cost"], 100 * REC_M[0.0]["meet"]),
xytext=(11.2, 36), fontsize=8, arrowprops=dict(arrowstyle="-", lw=0.6, color="#333333"))
ax.set(xlim=(10, 16), ylim=(30, 104), xlabel="average cost of the designs ($ million)",
ylabel="designs keeping 5% (%)", title="On 1,000 paths a radius buys reliability like a margin (seed 26211)")
ax.legend(frameon=False, fontsize=8, loc="lower right"); plt.tight_layout(); plt.show()designed on a record of 1,000 paths, judged on the fresh 100,000:
keep 5% mean true chance 90th percentile mean cost
target lower by 0.0 pt 38.5% 5.21% 6.23% 10.19
target lower by 0.5 pt 70.0% 4.70% 5.77% 11.09
target lower by 1.0 pt 87.0% 4.21% 5.09% 12.09
target lower by 1.5 pt 98.5% 3.73% 4.52% 13.12
target lower by 2.0 pt 100.0% 3.22% 3.85% 14.33
radius $ 10,000 78.5% 4.51% 5.51% 11.44
radius $ 20,000 86.5% 4.26% 5.19% 11.94
radius $ 30,000 91.5% 4.08% 4.96% 12.33
radius $ 50,000 97.5% 3.80% 4.60% 12.94
radius $100,000 99.5% 3.32% 4.06% 14.07
radius $200,000 100.0% 2.70% 3.31% 15.73
the book's 100,000 paths, Volume I's package: fresh 4.91% at 10.44

Panel 6 — Moments only (Proposition 11.1(iii)–(iv); Section 11.4: quarter 7's margin, $147.5m on average with a spread of $62.9m, breaches with chance up to 15.4%; 4.36 spreads of margin; a $160m line and 6.06 quarters, $53.29m)
AXIOM Lab 2.11, step 5. The widest family a record supports declares only means and spreads. For a margin of mean and spread , the worst chance that it falls to the floor is Cantelli's , attained by a two-point law: the margin is exactly zero with that chance and otherwise (Proposition 11.1(iii)), the lump a closure produces. Every such law keeps 5 percent exactly when . With twelve quarterly margins, the worst chance that some margin falls to zero lies between the largest bound and their sum (part (iv)). A delay adds to each quarter's mean and leaves its spread alone, so the cheapest package keeping every quarter's bound at 5 percent takes, at each line, the delay .
Python Code
def moments(P, line, phi):
# the mean and spread of each quarter's margin above the floor ($m) under the package (line, phi)
m = 1000.0 * (BASE[None, :] - (DEPTH0 - DPQ * phi) * SHAPE[None, :] + P.E - P.lost(line) - FLOOR)
return m.mean(axis=0), m.std(axis=0)
def cantelli(mu, sd):
# Proposition 11.1(iii): the worst chance of a margin at or below zero, over laws with mean mu and spread sd
return np.where(mu > 0, sd ** 2 / (sd ** 2 + mu ** 2), 1.0)
MU, SD = moments(BOOK, *PKG[:2])
CB = cantelli(MU, SD)
k7 = int(np.argmax(CB))
print("Volume I's package, quarter by quarter:")
print(" quarter " + " ".join(f"{k:6d}" for k in range(1, Q + 1)))
print(" mean margin " + " ".join(f"{x:6.1f}" for x in MU))
print(" spread " + " ".join(f"{x:6.1f}" for x in SD))
print(" Cantelli (%) " + " ".join(f"{100 * x:6.1f}" for x in CB))
print(f"worst in quarter {k7 + 1}: {100 * CB[k7]:.1f}%, a two-point law: the margin is 0 with that chance, "
f"{MU[k7] + SD[k7] ** 2 / MU[k7]:.1f} otherwise; the bounds sum to {100 * CB.sum():.1f}%, so the union bound says nothing")
Z = math.sqrt((1 - ALPHA) / ALPHA)
Z_UNION = math.sqrt((1 - ALPHA / Q) / (ALPHA / Q))
best = None
for line in LINES:
mu0, sd = moments(BOOK, line, 0.0)
phi = max(0.0, float(np.max((Z * sd - mu0) / LIFT)))
if phi <= MAX_DELAY and (best is None or cost(line, phi) < best[2]):
best = (float(line), phi, float(cost(line, phi)))
MOM = best
MOM_MU0, MOM_SD = moments(BOOK, MOM[0], 0.0)
print(f"\nevery quarter's bound at 5% (mean at least {Z:.3f} spreads): line {MOM[0]:.0f}, delay {MOM[1]:.3f}, cost {MOM[2]:.2f}; "
f"under the law {100 * BOOK.p(*MOM[:2]):.3f}% (fresh {100 * FRESH.p(*MOM[:2]):.3f}%)")
print(f"the program's appetite by the union bound would need {Z_UNION:.2f} spreads in every quarter")
print(f"at the records' corner the moment package breaches on {100 * CORNER.p(*MOM[:2]):.2f}% "
f"(fresh {100 * CORNER_F.p(*MOM[:2]):.2f}%)")Volume I's package, quarter by quarter: quarter 1 2 3 4 5 6 7 8 9 10 11 12 mean margin 229.0 198.6 176.9 162.1 152.9 148.4 147.5 149.5 153.3 158.9 166.1 174.0 spread 50.1 58.4 61.2 62.3 62.7 62.9 62.9 62.9 63.1 63.1 63.3 63.5 Cantelli (%) 4.6 8.0 10.7 12.9 14.4 15.2 15.4 15.0 14.5 13.6 12.7 11.8 worst in quarter 7: 15.4%, a two-point law: the margin is 0 with that chance, 174.3 otherwise; the bounds sum to 148.6%, so the union bound says nothing every quarter's bound at 5% (mean at least 4.359 spreads): line 160, delay 6.062, cost 53.29; under the law 0.026% (fresh 0.027%) the program's appetite by the union bound would need 15.46 spreads in every quarter at the records' corner the moment package breaches on 0.35% (fresh 0.37%)
Panel 7 — Price the doubt (Worked Example 11.1, Step 5; Table 11.3; Figure 11.3(b): $24.00m against $10.44m, a price of ambiguity of $13.56m, mostly the spread's; a record twice as long worth $4.63m)
Worked Example 11.1, Step 5. Every declaration's package side by side: its cost, its chance of a breach under Chapter 9's law, and under the records' worst law (closures 0.315 a year, spread $61.6 million), on the book's 100,000 paths and on the fresh ones. The price of ambiguity is what keeping the appetite under every law the records allow costs beyond trusting the law; the value of a longer record is what a record twice as long, at the same estimates, would save. Each chance is a count of 100,000 paths, shown exactly, to three decimals; Table 11.3 rounds them to one.
Python Code
T113 = [("One law (Chapter 9)", PKG), ("A set, max-min choice (Chapter 10)", CH10 + (cost(*CH10),)),
("Wasserstein ball, $0.1 million", W_DESIGNS[0.1]["design"]), ("Divergence ball, d = 0.01", KL_DESIGNS[0.01]["design"]),
("Closures' doubt (rate <= 0.315)", CASES["rate"]["design"]), ("Spread's doubt (spread <= 61.6)", CASES["spread"]["design"]),
("The records' family (both)", CASES["both"]["design"]), ("Records twice as long", CASES["longer"]["design"]),
("Moments only (every quarter)", MOM)]
TABLE = []
print(f"{'declaration':36s} {'line':>5s} {'delay':>6s} {'cost':>6s} | {'law':>7s} {'fresh':>7s} | {'worst law':>9s} {'fresh':>7s}")
for name, (l_, ph, co) in T113:
row = dict(name=name, line=l_, delay=ph, cost=co, p_law=BOOK.p(l_, ph), p_law_fresh=FRESH.p(l_, ph),
p_worst=CORNER.p(l_, ph), p_worst_fresh=CORNER_F.p(l_, ph))
TABLE.append(row)
print(f"{name:36s} {l_:5.1f} {ph:6.3f} {co:6.2f} | {100 * row['p_law']:6.3f}% {100 * row['p_law_fresh']:6.3f}% | "
f"{100 * row['p_worst']:8.3f}% {100 * row['p_worst_fresh']:6.3f}%")
PRICE = CASES["both"]["design"][2] - PKG[2]
RECORD_VALUE = CASES["both"]["design"][2] - CASES["longer"]["design"][2]
gaps = [abs(r[a] - r[b]) for r in TABLE for a, b in (("p_law", "p_law_fresh"), ("p_worst", "p_worst_fresh"))]
SE_MAX = max(math.sqrt(r[k] * (1 - r[k]) / N) for r in TABLE for k in ("p_law", "p_worst"))
print(f"\nthe price of ambiguity: {CASES['both']['design'][2]:.2f} - {PKG[2]:.2f} = {PRICE:.2f}, of which the spread's doubt "
f"alone accounts for {CASES['spread']['design'][2] - PKG[2]:.2f}")
print(f"the value of a record twice as long: {CASES['both']['design'][2]:.2f} - {CASES['longer']['design'][2]:.2f} = "
f"{RECORD_VALUE:.2f}")
print(f"fresh paths within {100 * max(gaps):.2f} points of the book's; the largest standard error of a chance in the table: "
f"{100 * SE_MAX:.3f} points")declaration line delay cost | law fresh | worst law fresh One law (Chapter 9) 170.0 0.668 10.44 | 5.000% 4.907% | 15.789% 15.835% A set, max-min choice (Chapter 10) 231.6 0.449 10.54 | 5.637% 5.508% | 16.160% 16.046% Wasserstein ball, $0.1 million 165.0 1.148 14.13 | 3.233% 3.205% | 12.153% 12.051% Divergence ball, d = 0.01 175.0 1.354 16.08 | 2.498% 2.515% | 10.248% 10.075% Closures' doubt (rate <= 0.315) 210.0 0.667 11.64 | 4.631% 4.534% | 14.431% 14.391% Spread's doubt (spread <= 61.6) 170.0 2.218 22.85 | 1.063% 1.032% | 5.858% 5.751% The records' family (both) 200.0 2.250 24.00 | 0.894% 0.860% | 5.000% 4.905% Records twice as long 195.0 1.690 19.37 | 1.664% 1.622% | 7.665% 7.438% Moments only (every quarter) 160.0 6.062 53.29 | 0.026% 0.027% | 0.354% 0.369% the price of ambiguity: 24.00 - 10.44 = 13.56, of which the spread's doubt alone accounts for 12.41 the value of a record twice as long: 24.00 - 19.37 = 4.63 fresh paths within 0.23 points of the book's; the largest standard error of a chance in the table: 0.116 points
Validation — the book's numbers, and agreement with DCT_V2_Ch11_Lab.xlsx
Each checkpoint is a number printed in Chapter 11 (Sections 11.1–11.5, Worked Example 11.1, Table 11.3, Figures
11.1–11.3) or in Online Appendix Example 11.B.2, with a tolerance of half a unit in its last printed digit (counts that
fall exactly halfway, such as 7,665 paths of 100,000, pass either way); a few check what a proposition promises.
Chances are in percent. The divergence ball's package at the records' corner (10,248 paths of 100,000) is checked
against Online Appendix Example 11.B.2's two decimals, and the levers' floor, , against the Online Appendix's 0.377. The workbook recomputes the checkpoints marked workbook with Excel formulas
and shows PASS against the same engine values on its Checks sheet.
Python Code
pc = lambda x: 100 * x
h = lambda dp: 0.5 * 10.0 ** -dp + 1e-9 # half a unit in the last printed digit
OA_RATE = (4.47, 4.57, 4.73, 5.00, 5.27, 5.53, 5.80, 6.16, 6.35)
OA_SPREAD = (1.03, 1.59, 2.52, 5.00, 8.59, 12.92, 14.43, 17.87)
OA_CASES = {"rate": (6.16, 6.06, 210, 0.667, 11.64, 4.88, 4.63, 4.53), "spread": (14.43, 14.44, 170, 2.218, 22.85, 4.91, 1.06, 1.03),
"both": (15.79, 15.83, 200, 2.250, 24.00, 4.91, 0.89, 0.86), "longer": (11.60, 11.43, 195, 1.690, 19.37, 4.87, 1.66, 1.62)}
OA_KL_WORST = {0.001: 6.00, 0.003: 6.78, 0.01: 8.36, 0.03: 11.15, 0.1: 17.29, 0.3: 28.85, 1.0: 55.55}
OA_KL = {0.001: (4.08, 160, 0.924, 12.19, 4.05), 0.002: (3.74, 165, 0.998, 12.93, 3.68), 0.005: (3.11, 170, 1.162, 14.40, 3.09),
0.01: (2.50, 175, 1.354, 16.08, 2.51), 0.02: (1.78, 180, 1.671, 18.77, 1.75), 0.03: (1.34, 175, 1.955, 20.89, 1.30),
0.05: (0.81, 195, 2.356, 24.70, 0.79), 0.075: (0.46, 215, 2.737, 28.35, 0.46), 0.1: (0.27, 200, 3.173, 31.39, 0.30),
0.15: (0.10, 240, 3.851, 38.01, 0.11)}
OA_W_WORST = {0.01: 5.64, 0.05: 6.47, 0.1: 7.11, 0.2: 8.05, 0.5: 10.00, 1.0: 12.32}
OA_W = {0.01: (165, 0.827, 11.57, 4.40, 4.32), 0.02: (165, 0.887, 12.04, 4.17, 4.11), 0.05: (165, 1.007, 13.00, 3.71, 3.66),
0.1: (165, 1.148, 14.13, 3.23, 3.21), 0.2: (170, 1.335, 15.78, 2.61, 2.61), 0.5: (175, 1.761, 19.34, 1.67, 1.61),
1.0: (175, 2.337, 23.95, 0.91, 0.89)}
OA_REC_M = {0.0: (38.5, 5.21, 6.23, 10.19), 0.005: (70.0, 4.70, 5.77, 11.09), 0.01: (87.0, 4.21, 5.09, 12.09),
0.015: (98.5, 3.73, 4.52, 13.12), 0.02: (100.0, 3.22, 3.85, 14.33)}
OA_REC_W = {0.01: (78.5, 4.51, 5.51, 11.44), 0.02: (86.5, 4.26, 5.19, 11.94), 0.03: (91.5, 4.08, 4.96, 12.33),
0.05: (97.5, 3.80, 4.60, 12.94), 0.1: (99.5, 3.32, 4.06, 14.07), 0.2: (100.0, 2.70, 3.31, 15.73)}
OA_MU = (229.0, 198.6, 176.9, 162.1, 152.9, 148.4, 147.5, 149.5, 153.3, 158.9, 166.1, 174.0)
OA_SD = (50.1, 58.4, 61.2, 62.3, 62.7, 62.9, 62.9, 62.9, 63.1, 63.1, 63.3, 63.5)
OA_CB = (4.6, 8.0, 10.7, 12.9, 14.4, 15.2, 15.4, 15.0, 14.5, 13.6, 12.7, 11.8)
OA_CORNER = {"Chapter 10's max-min package": (CH10, 16.16, 16.05), "the closures' doubt's package": (CASES["rate"]["design"], 14.43, 14.39),
"the spread's doubt's package": (CASES["spread"]["design"], 5.86, 5.75),
"twice the record's package": (CASES["longer"]["design"], 7.67, 7.44),
"the divergence ball's package, d = 0.01": (KL_DESIGNS[0.01]["design"], 10.25, 10.08),
"the divergence ball's package, d = 0.03": (KL_DESIGNS[0.03]["design"], 6.87, 6.71),
"the Wasserstein ball's package, $0.1m": (W_DESIGNS[0.1]["design"], 12.15, 12.05),
"the Wasserstein ball's package, $1m": (W_DESIGNS[1.0]["design"], 5.23, 5.15), "the moment package": (MOM, 0.35, 0.37)}
checks = [
# Proposition 11.2; Section 11.1; Worked Example 11.1, Step 1 (Panel 1)
("records: rate, upper 95% bound (workbook)", LAM_U, 0.3148, h(4)),
("records: rate, lower 95% bound (workbook)", LAM_L, 0.0178, h(4)),
("records: rate bound, Section 11.1 (workbook)", LAM_U, 0.315, h(3)),
("records: spread, upper 95% bound (workbook)", SIG_U, 61.60, h(2)),
("records: spread, lower 95% bound (workbook)", SIG_L, 42.27, h(2)),
("twice the record: rate bound (workbook)", LAM_U2, 0.2288, h(4)),
("twice the record: spread, upper bound (workbook)", SIG_U2, 57.60, h(2)),
("twice the record: spread, lower bound (workbook)", SIG_L2, 44.28, h(2)),
("the records' family: confidence together, % (workbook)", 100 * (1 - 0.05 - 0.05), 90, 1e-9),
("Volume I's package: line", PKG[0], 170, 0),
("Volume I's package: delay", PKG[1], 0.668, h(3)),
("Volume I's package: cost (workbook)", PKG[2], 10.44, h(2)),
("Volume I's package: chance under the law", pc(PKG_P), 5.00, h(2)),
("Volume I's package: chance on the fresh paths", pc(PKG_F), 4.91, h(2)),
] + [(f"Volume I's package at rate {lam:.3f}", pc(p_), b_, h(2)) for lam, p_, b_ in zip(RATES, BY_RATE, OA_RATE)] + [
(f"Volume I's package at spread {s:.1f}", pc(p_), b_, h(2)) for s, p_, b_ in zip(SPREADS, BY_SPREAD, OA_SPREAD)] + [
("Step 1: across the closures' range, from (%)", pc(min(BY_RATE[:8])), 4.5, h(1)),
("Step 1: across the closures' range, to (%)", pc(max(BY_RATE[:8])), 6.2, h(1)),
("Step 1: across the spread's range, from (%)", pc(BY_SPREAD[1]), 1.6, h(1)),
("Step 1: across the spread's range, to (%)", pc(BY_SPREAD[6]), 14.4, h(1)),
("Volume I's package at the records' corner", pc(P_CORNER), 15.79, h(2)),
("Volume I's package at the corner, fresh paths", pc(P_CORNER_F), 15.83, h(2)),
("Section 11.1: at the corner, %", pc(P_CORNER), 15.8, h(1)),
("Proposition 11.2(ii): the spread's bound holds, % of simulated records", pc(COVER["spread"]), 95.0, 0.25),
("Proposition 11.2(iii): both bounds hold in at least 90% of records", int(COVER["both"] >= 0.90), 1, 0),
]
for key, (pp, ppf, l_, ph, co, pf, pl, plf) in OA_CASES.items():
c = CASES[key]
checks += [(f"{NAMES[key]}: Volume I's package at its law", pc(c["p_pkg"]), pp, h(2)),
(f"{NAMES[key]}: Volume I's package, fresh", pc(c["p_pkg_fresh"]), ppf, h(2)),
(f"{NAMES[key]}: line", c["design"][0], l_, 0),
(f"{NAMES[key]}: delay", c["design"][1], ph, h(3)),
(f"{NAMES[key]}: cost (workbook)", c["design"][2], co, h(2)),
(f"{NAMES[key]}: fresh paths at its law", pc(c["p_fresh"]), pf, h(2)),
(f"{NAMES[key]}: under the declared law", pc(c["p_law"]), pl, h(2)),
(f"{NAMES[key]}: under the declared law, fresh", pc(c["p_law_fresh"]), plf, h(2))]
checks += [
# Theorem 11.1(iv); Worked Example 11.1, Step 2 (Panel 2)
("the family's package: largest chance on the 49-law grid (workbook)", pc(worst_law[2]), 5.00, h(2)),
("the family's package: the grid's worst law is the corner", int(worst_law[0] == GRID[-1][0] and worst_law[1] == GRID[-1][1]), 1, 0),
("Step 2: on fresh paths at the corner (%)", pc(CASES["both"]["p_fresh"]), 4.9, h(1)),
("Step 2: under the declared law (%)", pc(CASES["both"]["p_law"]), 0.9, h(1)),
# Theorem 11.2; Figure 11.1; Section 11.2; Step 3 (Panel 3)
("divergence: radius of the closures' doubt (workbook)", RADII["closures' doubt"], 0.439, h(3)),
("divergence: radius of the spread's doubt (workbook)", RADII["spread's doubt"], 0.603, h(3)),
("divergence: radius of both (workbook)", RADII["both doubts"], 1.042, h(3)),
("divergence: the law's appetite at 0.439, per million (workbook)", 1e6 * appetite(RADII["closures' doubt"]), 2.93, h(2)),
("divergence: the law's appetite at 0.603, per 10 million (workbook)", 1e7 * appetite(RADII["spread's doubt"]), 1.09, h(2)),
("divergence: the law's appetite at 1.042, per 100 billion (workbook)", 1e11 * appetite(RADII["both doubts"]), 1.69, h(2)),
("Section 11.2: 3 in a million", round(1e6 * appetite(RADII["closures' doubt"])), 3, 0),
("Step 3: 2 in 100 billion", round(1e11 * appetite(RADII["both doubts"])), 2, 0),
("the levers' floor: paths that breach at a $400m line and 7 quarters", P_MIN * N, 1, 0),
("the levers' floor: kl(0.05 || 1e-5), Online Appendix 11.B.2 (workbook)", D_LEVERS, 0.377, h(3)),
] + [(f"divergence: worst chance of Volume I's package at d = {d:g} (workbook)", pc(KL_WORST[d]), b_, h(2))
for d, b_ in OA_KL_WORST.items()]
for d, (a_, l_, ph, co, pf) in OA_KL.items():
k_ = KL_DESIGNS[d]
checks += [(f"divergence d = {d:g}: the law's appetite, % (workbook)", pc(k_["alpha"]), a_, h(2)),
(f"divergence d = {d:g}: line", k_["design"][0], l_, 0),
(f"divergence d = {d:g}: delay", k_["design"][1], ph, h(3)),
(f"divergence d = {d:g}: cost (workbook)", k_["design"][2], co, h(2)),
(f"divergence d = {d:g}: fresh paths", pc(k_["p_fresh"]), pf, h(2))]
checks += [
("Table 11.3: divergence ball d = 0.01, under the law (%)", pc(KL_DESIGNS[0.01]["p"]), 2.5, h(1)),
# Theorem 11.3; Figure 11.2(a); Section 11.3; Step 3 (Panel 4)
] + [(f"Wasserstein: worst chance of Volume I's package at ${e:g}m", pc(W_WORST[e]), b_, h(2)) for e, b_ in OA_W_WORST.items()]
for e, (l_, ph, co, pl, pf) in OA_W.items():
w_ = W_DESIGNS[e]
checks += [(f"Wasserstein ${e:g}m: line", w_["design"][0], l_, 0),
(f"Wasserstein ${e:g}m: delay", w_["design"][1], ph, h(3)),
(f"Wasserstein ${e:g}m: cost (workbook)", w_["design"][2], co, h(2)),
(f"Wasserstein ${e:g}m: under the law", pc(w_["p"]), pl, h(2)),
(f"Wasserstein ${e:g}m: fresh paths", pc(w_["p_fresh"]), pf, h(2)),
(f"Wasserstein ${e:g}m: the thinnest 5% average radius / 0.05 (workbook)", w_["thin"], e / ALPHA, 1e-6)]
checks += [
("Section 11.3: thinnest 5% must average, $m (workbook)", W_DESIGNS[0.1]["thin"], 2, h(0)),
("Wasserstein: the spread's excess (workbook)", SIG_U / 50.0 - 1.0, 0.232, h(3)),
("Wasserstein: the paths' average largest deviation, $m", W_MAXDEV, 113.2, h(1)),
("Wasserstein: radius of the spread's doubt, $m (workbook)", W_SPREAD, 26.3, h(1)),
("Wasserstein: the thinnest 5% would need, $m (workbook)", W_SPREAD / ALPHA, 525, h(0)),
("Wasserstein: within the levers' reach, $m", BEST_THIN, 121.9, h(1)),
("Section 11.3: the levers reach, $m", BEST_THIN, 122, h(0)),
]
for mg, (mt, mtrue, q9, mc) in OA_REC_M.items():
s = REC_M[mg]
checks += [(f"records of 1,000, target lower by {100 * mg:g} pt: keep 5%", pc(s["meet"]), mt, h(1)),
(f"records of 1,000, target lower by {100 * mg:g} pt: mean true chance", pc(s["mean_true"]), mtrue, h(2)),
(f"records of 1,000, target lower by {100 * mg:g} pt: 90th percentile", pc(s["q90"]), q9, h(2)),
(f"records of 1,000, target lower by {100 * mg:g} pt: mean cost", s["mean_cost"], mc, h(2))]
for e, (mt, mtrue, q9, mc) in OA_REC_W.items():
s = REC_W[e]
checks += [(f"records of 1,000, radius ${1e6 * e:,.0f}: keep 5%", pc(s["meet"]), mt, h(1)),
(f"records of 1,000, radius ${1e6 * e:,.0f}: mean true chance", pc(s["mean_true"]), mtrue, h(2)),
(f"records of 1,000, radius ${1e6 * e:,.0f}: 90th percentile", pc(s["q90"]), q9, h(2)),
(f"records of 1,000, radius ${1e6 * e:,.0f}: mean cost", s["mean_cost"], mc, h(2))]
checks += [(f"moments: mean margin, quarter {k + 1}", MU[k], OA_MU[k], h(1)) for k in range(Q)]
checks += [(f"moments: spread, quarter {k + 1}", SD[k], OA_SD[k], h(1)) for k in range(Q)]
checks += [(f"moments: Cantelli bound, quarter {k + 1} (workbook)", pc(CB[k]), OA_CB[k], h(1)) for k in range(Q)]
checks += [
("moments: the worst quarter (workbook)", k7 + 1, 7, 0),
("moments: the bounds' sum, % (workbook)", pc(CB.sum()), 148.6, h(1)),
("moments: spreads of margin at 5% (workbook)", Z, 4.359, h(3)),
("moments: spreads of margin, Section 11.4 (workbook)", Z, 4.36, h(2)),
("moments: the union bound would need, spreads (workbook)", Z_UNION, 15.46, h(2)),
("moments: line (workbook)", MOM[0], 160, 0),
("moments: delay (workbook)", MOM[1], 6.062, h(3)),
("moments: cost (workbook)", MOM[2], 53.29, h(2)),
("moments: chance under the law, %", pc(BOOK.p(*MOM[:2])), 0.026, h(3)),
]
for name, (pk, pb, pf) in OA_CORNER.items():
checks += [(f"at the corner: {name}", pc(CORNER.p(*pk[:2])), pb, h(2)),
(f"at the corner: {name}, fresh", pc(CORNER_F.p(*pk[:2])), pf, h(2))]
checks += [
("Chapter 10's max-min package: line", CH10[0], 231.6, h(1)),
("Chapter 10's max-min package: delay", CH10[1], 0.449, h(3)),
("Chapter 10's max-min package: cost (workbook)", cost(*CH10), 10.54, h(2)),
("Chapter 10's max-min package: chance under the law", pc(BOOK.p(*CH10)), 5.64, h(2)),
# Table 11.3 and Step 5 (Panel 7)
("price of ambiguity (workbook)", PRICE, 13.56, h(2)),
("value of a record twice as long (workbook)", RECORD_VALUE, 4.63, h(2)),
("Table 11.3: chances on fresh paths within 0.3 points", int(max(gaps) <= 0.003), 1, 0),
("Section 11.5: every chance in Table 11.3 has a standard error below 0.12 points (workbook)", int(SE_MAX < 0.0012), 1, 0),
]
T113_BOOK = {"One law (Chapter 9)": (170, 0.67, 10.44, 5.0, 15.8), "A set, max-min choice (Chapter 10)": (232, 0.45, 10.54, 5.6, 16.2),
"Wasserstein ball, $0.1 million": (165, 1.15, 14.13, 3.2, 12.2), "Divergence ball, d = 0.01": (175, 1.35, 16.08, 2.5, None),
"Closures' doubt (rate <= 0.315)": (210, 0.67, 11.64, 4.6, 14.4), "Spread's doubt (spread <= 61.6)": (170, 2.22, 22.85, 1.1, 5.9),
"The records' family (both)": (200, 2.25, 24.00, 0.9, None), "Records twice as long": (195, 1.69, 19.37, 1.7, 7.7),
"Moments only (every quarter)": (160, 6.06, 53.29, 0.03, 0.4)}
for r in TABLE:
l_, ph, co, pl, pw = T113_BOOK[r["name"]]
checks += [(f"Table 11.3, {r['name']}: line", r["line"], l_, h(0)), (f"Table 11.3, {r['name']}: delay", r["delay"], ph, h(2)),
(f"Table 11.3, {r['name']}: cost (workbook)", r["cost"], co, h(2)),
(f"Table 11.3, {r['name']}: under the law", pc(r["p_law"]), pl, h(2 if pl < 0.1 else 1))]
if pw is not None:
checks.append((f"Table 11.3, {r['name']}: under the worst law", pc(r["p_worst"]), pw, h(1)))
checks += [("Table 11.3, the records' family: under the worst law (the grid's corner)", pc(worst_law[2]), 5.0, h(1))]
ok = 0
for name, got, want, tol in checks:
status = "PASS" if abs(got - want) <= tol else "FAIL"; ok += status == "PASS"
print(f"{name:78s} {got:12.4f} book {want:10.4f} {status}")
print(f"\n{ok} of {len(checks)} checkpoints PASS")records: rate, upper 95% bound (workbook) 0.3148 book 0.3148 PASS records: rate, lower 95% bound (workbook) 0.0178 book 0.0178 PASS records: rate bound, Section 11.1 (workbook) 0.3148 book 0.3150 PASS records: spread, upper 95% bound (workbook) 61.5991 book 61.6000 PASS records: spread, lower 95% bound (workbook) 42.2685 book 42.2700 PASS twice the record: rate bound (workbook) 0.2288 book 0.2288 PASS twice the record: spread, upper bound (workbook) 57.6028 book 57.6000 PASS twice the record: spread, lower bound (workbook) 44.2756 book 44.2800 PASS the records' family: confidence together, % (workbook) 90.0000 book 90.0000 PASS Volume I's package: line 170.0000 book 170.0000 PASS Volume I's package: delay 0.6676 book 0.6680 PASS Volume I's package: cost (workbook) 10.4406 book 10.4400 PASS Volume I's package: chance under the law 5.0000 book 5.0000 PASS Volume I's package: chance on the fresh paths 4.9070 book 4.9100 PASS Volume I's package at rate 0.000 4.4710 book 4.4700 PASS Volume I's package at rate 0.018 4.5670 book 4.5700 PASS Volume I's package at rate 0.050 4.7320 book 4.7300 PASS Volume I's package at rate 0.100 5.0000 book 5.0000 PASS Volume I's package at rate 0.150 5.2710 book 5.2700 PASS Volume I's package at rate 0.200 5.5290 book 5.5300 PASS Volume I's package at rate 0.250 5.7960 book 5.8000 PASS Volume I's package at rate 0.315 6.1600 book 6.1600 PASS Volume I's package at rate 0.350 6.3530 book 6.3500 PASS Volume I's package at spread 40.0 1.0340 book 1.0300 PASS Volume I's package at spread 42.3 1.5910 book 1.5900 PASS Volume I's package at spread 45.0 2.5150 book 2.5200 PASS Volume I's package at spread 50.0 5.0000 book 5.0000 PASS Volume I's package at spread 55.0 8.5850 book 8.5900 PASS Volume I's package at spread 60.0 12.9170 book 12.9200 PASS Volume I's package at spread 61.6 14.4270 book 14.4300 PASS Volume I's package at spread 65.0 17.8650 book 17.8700 PASS Step 1: across the closures' range, from (%) 4.4710 book 4.5000 PASS Step 1: across the closures' range, to (%) 6.1600 book 6.2000 PASS Step 1: across the spread's range, from (%) 1.5910 book 1.6000 PASS Step 1: across the spread's range, to (%) 14.4270 book 14.4000 PASS Volume I's package at the records' corner 15.7890 book 15.7900 PASS Volume I's package at the corner, fresh paths 15.8350 book 15.8300 PASS Section 11.1: at the corner, % 15.7890 book 15.8000 PASS Proposition 11.2(ii): the spread's bound holds, % of simulated records 94.9860 book 95.0000 PASS Proposition 11.2(iii): both bounds hold in at least 90% of records 1.0000 book 1.0000 PASS closures' doubt: Volume I's package at its law 6.1600 book 6.1600 PASS closures' doubt: Volume I's package, fresh 6.0640 book 6.0600 PASS closures' doubt: line 210.0000 book 210.0000 PASS closures' doubt: delay 0.6672 book 0.6670 PASS closures' doubt: cost (workbook) 11.6373 book 11.6400 PASS closures' doubt: fresh paths at its law 4.8830 book 4.8800 PASS closures' doubt: under the declared law 4.6310 book 4.6300 PASS closures' doubt: under the declared law, fresh 4.5340 book 4.5300 PASS spread's doubt: Volume I's package at its law 14.4270 book 14.4300 PASS spread's doubt: Volume I's package, fresh 14.4430 book 14.4400 PASS spread's doubt: line 170.0000 book 170.0000 PASS spread's doubt: delay 2.2184 book 2.2180 PASS spread's doubt: cost (workbook) 22.8469 book 22.8500 PASS spread's doubt: fresh paths at its law 4.9140 book 4.9100 PASS spread's doubt: under the declared law 1.0630 book 1.0600 PASS spread's doubt: under the declared law, fresh 1.0320 book 1.0300 PASS the records' family: Volume I's package at its law 15.7890 book 15.7900 PASS the records' family: Volume I's package, fresh 15.8350 book 15.8300 PASS the records' family: line 200.0000 book 200.0000 PASS the records' family: delay 2.2504 book 2.2500 PASS the records' family: cost (workbook) 24.0032 book 24.0000 PASS the records' family: fresh paths at its law 4.9050 book 4.9100 PASS the records' family: under the declared law 0.8940 book 0.8900 PASS the records' family: under the declared law, fresh 0.8600 book 0.8600 PASS twice the record: Volume I's package at its law 11.5980 book 11.6000 PASS twice the record: Volume I's package, fresh 11.4290 book 11.4300 PASS twice the record: line 195.0000 book 195.0000 PASS twice the record: delay 1.6901 book 1.6900 PASS twice the record: cost (workbook) 19.3707 book 19.3700 PASS twice the record: fresh paths at its law 4.8710 book 4.8700 PASS twice the record: under the declared law 1.6640 book 1.6600 PASS twice the record: under the declared law, fresh 1.6220 book 1.6200 PASS the family's package: largest chance on the 49-law grid (workbook) 5.0000 book 5.0000 PASS the family's package: the grid's worst law is the corner 1.0000 book 1.0000 PASS Step 2: on fresh paths at the corner (%) 4.9050 book 4.9000 PASS Step 2: under the declared law (%) 0.8940 book 0.9000 PASS divergence: radius of the closures' doubt (workbook) 0.4386 book 0.4390 PASS divergence: radius of the spread's doubt (workbook) 0.6032 book 0.6030 PASS divergence: radius of both (workbook) 1.0418 book 1.0420 PASS divergence: the law's appetite at 0.439, per million (workbook) 2.9266 book 2.9300 PASS divergence: the law's appetite at 0.603, per 10 million (workbook) 1.0876 book 1.0900 PASS divergence: the law's appetite at 1.042, per 100 billion (workbook) 1.6869 book 1.6900 PASS Section 11.2: 3 in a million 3.0000 book 3.0000 PASS Step 3: 2 in 100 billion 2.0000 book 2.0000 PASS the levers' floor: paths that breach at a $400m line and 7 quarters 1.0000 book 1.0000 PASS the levers' floor: kl(0.05 || 1e-5), Online Appendix 11.B.2 (workbook) 0.3771 book 0.3770 PASS divergence: worst chance of Volume I's package at d = 0.001 (workbook) 6.0041 book 6.0000 PASS divergence: worst chance of Volume I's package at d = 0.003 (workbook) 6.7752 book 6.7800 PASS divergence: worst chance of Volume I's package at d = 0.01 (workbook) 8.3648 book 8.3600 PASS divergence: worst chance of Volume I's package at d = 0.03 (workbook) 11.1548 book 11.1500 PASS divergence: worst chance of Volume I's package at d = 0.1 (workbook) 17.2949 book 17.2900 PASS divergence: worst chance of Volume I's package at d = 0.3 (workbook) 28.8519 book 28.8500 PASS divergence: worst chance of Volume I's package at d = 1 (workbook) 55.5514 book 55.5500 PASS divergence d = 0.001: the law's appetite, % (workbook) 4.0848 book 4.0800 PASS divergence d = 0.001: line 160.0000 book 160.0000 PASS divergence d = 0.001: delay 0.9243 book 0.9240 PASS divergence d = 0.001: cost (workbook) 12.1942 book 12.1900 PASS divergence d = 0.001: fresh paths 4.0490 book 4.0500 PASS divergence d = 0.002: the law's appetite, % (workbook) 3.7401 book 3.7400 PASS divergence d = 0.002: line 165.0000 book 165.0000 PASS divergence d = 0.002: delay 0.9981 book 0.9980 PASS divergence d = 0.002: cost (workbook) 12.9347 book 12.9300 PASS divergence d = 0.002: fresh paths 3.6800 book 3.6800 PASS divergence d = 0.005: the law's appetite, % (workbook) 3.1142 book 3.1100 PASS divergence d = 0.005: line 170.0000 book 170.0000 PASS divergence d = 0.005: delay 1.1622 book 1.1620 PASS divergence d = 0.005: cost (workbook) 14.3973 book 14.4000 PASS divergence d = 0.005: fresh paths 3.0930 book 3.0900 PASS divergence d = 0.01: the law's appetite, % (workbook) 2.4981 book 2.5000 PASS divergence d = 0.01: line 175.0000 book 175.0000 PASS divergence d = 0.01: delay 1.3539 book 1.3540 PASS divergence d = 0.01: cost (workbook) 16.0812 book 16.0800 PASS divergence d = 0.01: fresh paths 2.5150 book 2.5100 PASS divergence d = 0.02: the law's appetite, % (workbook) 1.7786 book 1.7800 PASS divergence d = 0.02: line 180.0000 book 180.0000 PASS divergence d = 0.02: delay 1.6708 book 1.6710 PASS divergence d = 0.02: cost (workbook) 18.7665 book 18.7700 PASS divergence d = 0.02: fresh paths 1.7450 book 1.7500 PASS divergence d = 0.03: the law's appetite, % (workbook) 1.3373 book 1.3400 PASS divergence d = 0.03: line 175.0000 book 175.0000 PASS divergence d = 0.03: delay 1.9552 book 1.9550 PASS divergence d = 0.03: cost (workbook) 20.8919 book 20.8900 PASS divergence d = 0.03: fresh paths 1.3020 book 1.3000 PASS divergence d = 0.05: the law's appetite, % (workbook) 0.8101 book 0.8100 PASS divergence d = 0.05: line 195.0000 book 195.0000 PASS divergence d = 0.05: delay 2.3563 book 2.3560 PASS divergence d = 0.05: cost (workbook) 24.7001 book 24.7000 PASS divergence d = 0.05: fresh paths 0.7920 book 0.7900 PASS divergence d = 0.075: the law's appetite, % (workbook) 0.4595 book 0.4600 PASS divergence d = 0.075: line 215.0000 book 215.0000 PASS divergence d = 0.075: delay 2.7373 book 2.7370 PASS divergence d = 0.075: cost (workbook) 28.3482 book 28.3500 PASS divergence d = 0.075: fresh paths 0.4570 book 0.4600 PASS divergence d = 0.1: the law's appetite, % (workbook) 0.2687 book 0.2700 PASS divergence d = 0.1: line 200.0000 book 200.0000 PASS divergence d = 0.1: delay 3.1732 book 3.1730 PASS divergence d = 0.1: cost (workbook) 31.3854 book 31.3900 PASS divergence d = 0.1: fresh paths 0.2990 book 0.3000 PASS divergence d = 0.15: the law's appetite, % (workbook) 0.0957 book 0.1000 PASS divergence d = 0.15: line 240.0000 book 240.0000 PASS divergence d = 0.15: delay 3.8507 book 3.8510 PASS divergence d = 0.15: cost (workbook) 38.0059 book 38.0100 PASS divergence d = 0.15: fresh paths 0.1060 book 0.1100 PASS Table 11.3: divergence ball d = 0.01, under the law (%) 2.4980 book 2.5000 PASS Wasserstein: worst chance of Volume I's package at $0.01m 5.6426 book 5.6400 PASS Wasserstein: worst chance of Volume I's package at $0.05m 6.4693 book 6.4700 PASS Wasserstein: worst chance of Volume I's package at $0.1m 7.1132 book 7.1100 PASS Wasserstein: worst chance of Volume I's package at $0.2m 8.0529 book 8.0500 PASS Wasserstein: worst chance of Volume I's package at $0.5m 10.0006 book 10.0000 PASS Wasserstein: worst chance of Volume I's package at $1m 12.3200 book 12.3200 PASS Wasserstein $0.01m: line 165.0000 book 165.0000 PASS Wasserstein $0.01m: delay 0.8270 book 0.8270 PASS Wasserstein $0.01m: cost (workbook) 11.5657 book 11.5700 PASS Wasserstein $0.01m: under the law 4.4000 book 4.4000 PASS Wasserstein $0.01m: fresh paths 4.3170 book 4.3200 PASS Wasserstein $0.01m: the thinnest 5% average radius / 0.05 (workbook) 0.2000 book 0.2000 PASS Wasserstein $0.02m: line 165.0000 book 165.0000 PASS Wasserstein $0.02m: delay 0.8865 book 0.8870 PASS Wasserstein $0.02m: cost (workbook) 12.0422 book 12.0400 PASS Wasserstein $0.02m: under the law 4.1670 book 4.1700 PASS Wasserstein $0.02m: fresh paths 4.1070 book 4.1100 PASS Wasserstein $0.02m: the thinnest 5% average radius / 0.05 (workbook) 0.4000 book 0.4000 PASS Wasserstein $0.05m: line 165.0000 book 165.0000 PASS Wasserstein $0.05m: delay 1.0068 book 1.0070 PASS Wasserstein $0.05m: cost (workbook) 13.0048 book 13.0000 PASS Wasserstein $0.05m: under the law 3.7090 book 3.7100 PASS Wasserstein $0.05m: fresh paths 3.6550 book 3.6600 PASS Wasserstein $0.05m: the thinnest 5% average radius / 0.05 (workbook) 1.0000 book 1.0000 PASS Wasserstein $0.1m: line 165.0000 book 165.0000 PASS Wasserstein $0.1m: delay 1.1480 book 1.1480 PASS Wasserstein $0.1m: cost (workbook) 14.1344 book 14.1300 PASS Wasserstein $0.1m: under the law 3.2330 book 3.2300 PASS Wasserstein $0.1m: fresh paths 3.2050 book 3.2100 PASS Wasserstein $0.1m: the thinnest 5% average radius / 0.05 (workbook) 2.0000 book 2.0000 PASS Wasserstein $0.2m: line 170.0000 book 170.0000 PASS Wasserstein $0.2m: delay 1.3351 book 1.3350 PASS Wasserstein $0.2m: cost (workbook) 15.7811 book 15.7800 PASS Wasserstein $0.2m: under the law 2.6130 book 2.6100 PASS Wasserstein $0.2m: fresh paths 2.6070 book 2.6100 PASS Wasserstein $0.2m: the thinnest 5% average radius / 0.05 (workbook) 4.0000 book 4.0000 PASS Wasserstein $0.5m: line 175.0000 book 175.0000 PASS Wasserstein $0.5m: delay 1.7607 book 1.7610 PASS Wasserstein $0.5m: cost (workbook) 19.3360 book 19.3400 PASS Wasserstein $0.5m: under the law 1.6700 book 1.6700 PASS Wasserstein $0.5m: fresh paths 1.6130 book 1.6100 PASS Wasserstein $0.5m: the thinnest 5% average radius / 0.05 (workbook) 10.0000 book 10.0000 PASS Wasserstein $1m: line 175.0000 book 175.0000 PASS Wasserstein $1m: delay 2.3370 book 2.3370 PASS Wasserstein $1m: cost (workbook) 23.9463 book 23.9500 PASS Wasserstein $1m: under the law 0.9120 book 0.9100 PASS Wasserstein $1m: fresh paths 0.8890 book 0.8900 PASS Wasserstein $1m: the thinnest 5% average radius / 0.05 (workbook) 20.0000 book 20.0000 PASS Section 11.3: thinnest 5% must average, $m (workbook) 2.0000 book 2.0000 PASS Wasserstein: the spread's excess (workbook) 0.2320 book 0.2320 PASS Wasserstein: the paths' average largest deviation, $m 113.2319 book 113.2000 PASS Wasserstein: radius of the spread's doubt, $m (workbook) 26.2679 book 26.3000 PASS Wasserstein: the thinnest 5% would need, $m (workbook) 525.3575 book 525.0000 PASS Wasserstein: within the levers' reach, $m 121.9356 book 121.9000 PASS Section 11.3: the levers reach, $m 121.9356 book 122.0000 PASS records of 1,000, target lower by 0 pt: keep 5% 38.5000 book 38.5000 PASS records of 1,000, target lower by 0 pt: mean true chance 5.2093 book 5.2100 PASS records of 1,000, target lower by 0 pt: 90th percentile 6.2296 book 6.2300 PASS records of 1,000, target lower by 0 pt: mean cost 10.1850 book 10.1900 PASS records of 1,000, target lower by 0.5 pt: keep 5% 70.0000 book 70.0000 PASS records of 1,000, target lower by 0.5 pt: mean true chance 4.7018 book 4.7000 PASS records of 1,000, target lower by 0.5 pt: 90th percentile 5.7738 book 5.7700 PASS records of 1,000, target lower by 0.5 pt: mean cost 11.0893 book 11.0900 PASS records of 1,000, target lower by 1 pt: keep 5% 87.0000 book 87.0000 PASS records of 1,000, target lower by 1 pt: mean true chance 4.2102 book 4.2100 PASS records of 1,000, target lower by 1 pt: 90th percentile 5.0855 book 5.0900 PASS records of 1,000, target lower by 1 pt: mean cost 12.0864 book 12.0900 PASS records of 1,000, target lower by 1.5 pt: keep 5% 98.5000 book 98.5000 PASS records of 1,000, target lower by 1.5 pt: mean true chance 3.7277 book 3.7300 PASS records of 1,000, target lower by 1.5 pt: 90th percentile 4.5212 book 4.5200 PASS records of 1,000, target lower by 1.5 pt: mean cost 13.1223 book 13.1200 PASS records of 1,000, target lower by 2 pt: keep 5% 100.0000 book 100.0000 PASS records of 1,000, target lower by 2 pt: mean true chance 3.2249 book 3.2200 PASS records of 1,000, target lower by 2 pt: 90th percentile 3.8511 book 3.8500 PASS records of 1,000, target lower by 2 pt: mean cost 14.3276 book 14.3300 PASS records of 1,000, radius $10,000: keep 5% 78.5000 book 78.5000 PASS records of 1,000, radius $10,000: mean true chance 4.5059 book 4.5100 PASS records of 1,000, radius $10,000: 90th percentile 5.5135 book 5.5100 PASS records of 1,000, radius $10,000: mean cost 11.4423 book 11.4400 PASS records of 1,000, radius $20,000: keep 5% 86.5000 book 86.5000 PASS records of 1,000, radius $20,000: mean true chance 4.2566 book 4.2600 PASS records of 1,000, radius $20,000: 90th percentile 5.1903 book 5.1900 PASS records of 1,000, radius $20,000: mean cost 11.9416 book 11.9400 PASS records of 1,000, radius $30,000: keep 5% 91.5000 book 91.5000 PASS records of 1,000, radius $30,000: mean true chance 4.0799 book 4.0800 PASS records of 1,000, radius $30,000: 90th percentile 4.9599 book 4.9600 PASS records of 1,000, radius $30,000: mean cost 12.3273 book 12.3300 PASS records of 1,000, radius $50,000: keep 5% 97.5000 book 97.5000 PASS records of 1,000, radius $50,000: mean true chance 3.8000 book 3.8000 PASS records of 1,000, radius $50,000: 90th percentile 4.6046 book 4.6000 PASS records of 1,000, radius $50,000: mean cost 12.9388 book 12.9400 PASS records of 1,000, radius $100,000: keep 5% 99.5000 book 99.5000 PASS records of 1,000, radius $100,000: mean true chance 3.3230 book 3.3200 PASS records of 1,000, radius $100,000: 90th percentile 4.0581 book 4.0600 PASS records of 1,000, radius $100,000: mean cost 14.0668 book 14.0700 PASS records of 1,000, radius $200,000: keep 5% 100.0000 book 100.0000 PASS records of 1,000, radius $200,000: mean true chance 2.7006 book 2.7000 PASS records of 1,000, radius $200,000: 90th percentile 3.3121 book 3.3100 PASS records of 1,000, radius $200,000: mean cost 15.7320 book 15.7300 PASS moments: mean margin, quarter 1 229.0124 book 229.0000 PASS moments: mean margin, quarter 2 198.6151 book 198.6000 PASS moments: mean margin, quarter 3 176.9108 book 176.9000 PASS moments: mean margin, quarter 4 162.0724 book 162.1000 PASS moments: mean margin, quarter 5 152.8724 book 152.9000 PASS moments: mean margin, quarter 6 148.4060 book 148.4000 PASS moments: mean margin, quarter 7 147.4772 book 147.5000 PASS moments: mean margin, quarter 8 149.4924 book 149.5000 PASS moments: mean margin, quarter 9 153.2529 book 153.3000 PASS moments: mean margin, quarter 10 158.9373 book 158.9000 PASS moments: mean margin, quarter 11 166.1260 book 166.1000 PASS moments: mean margin, quarter 12 173.9834 book 174.0000 PASS moments: spread, quarter 1 50.1050 book 50.1000 PASS moments: spread, quarter 2 58.3755 book 58.4000 PASS moments: spread, quarter 3 61.1567 book 61.2000 PASS moments: spread, quarter 4 62.3368 book 62.3000 PASS moments: spread, quarter 5 62.6790 book 62.7000 PASS moments: spread, quarter 6 62.8537 book 62.9000 PASS moments: spread, quarter 7 62.8692 book 62.9000 PASS moments: spread, quarter 8 62.9152 book 62.9000 PASS moments: spread, quarter 9 63.0871 book 63.1000 PASS moments: spread, quarter 10 63.0684 book 63.1000 PASS moments: spread, quarter 11 63.3064 book 63.3000 PASS moments: spread, quarter 12 63.4894 book 63.5000 PASS moments: Cantelli bound, quarter 1 (workbook) 4.5681 book 4.6000 PASS moments: Cantelli bound, quarter 2 (workbook) 7.9516 book 8.0000 PASS moments: Cantelli bound, quarter 3 (workbook) 10.6747 book 10.7000 PASS moments: Cantelli bound, quarter 4 (workbook) 12.8871 book 12.9000 PASS moments: Cantelli bound, quarter 5 (workbook) 14.3914 book 14.4000 PASS moments: Cantelli bound, quarter 6 (workbook) 15.2092 book 15.2000 PASS moments: Cantelli bound, quarter 7 (workbook) 15.3783 book 15.4000 PASS moments: Cantelli bound, quarter 8 (workbook) 15.0470 book 15.0000 PASS moments: Cantelli bound, quarter 9 (workbook) 14.4904 book 14.5000 PASS moments: Cantelli bound, quarter 10 (workbook) 13.6040 book 13.6000 PASS moments: Cantelli bound, quarter 11 (workbook) 12.6804 book 12.7000 PASS moments: Cantelli bound, quarter 12 (workbook) 11.7515 book 11.8000 PASS moments: the worst quarter (workbook) 7.0000 book 7.0000 PASS moments: the bounds' sum, % (workbook) 148.6337 book 148.6000 PASS moments: spreads of margin at 5% (workbook) 4.3589 book 4.3590 PASS moments: spreads of margin, Section 11.4 (workbook) 4.3589 book 4.3600 PASS moments: the union bound would need, spreads (workbook) 15.4596 book 15.4600 PASS moments: line (workbook) 160.0000 book 160.0000 PASS moments: delay (workbook) 6.0618 book 6.0620 PASS moments: cost (workbook) 53.2947 book 53.2900 PASS moments: chance under the law, % 0.0260 book 0.0260 PASS at the corner: Chapter 10's max-min package 16.1600 book 16.1600 PASS at the corner: Chapter 10's max-min package, fresh 16.0460 book 16.0500 PASS at the corner: the closures' doubt's package 14.4310 book 14.4300 PASS at the corner: the closures' doubt's package, fresh 14.3910 book 14.3900 PASS at the corner: the spread's doubt's package 5.8580 book 5.8600 PASS at the corner: the spread's doubt's package, fresh 5.7510 book 5.7500 PASS at the corner: twice the record's package 7.6650 book 7.6700 PASS at the corner: twice the record's package, fresh 7.4380 book 7.4400 PASS at the corner: the divergence ball's package, d = 0.01 10.2480 book 10.2500 PASS at the corner: the divergence ball's package, d = 0.01, fresh 10.0750 book 10.0800 PASS at the corner: the divergence ball's package, d = 0.03 6.8690 book 6.8700 PASS at the corner: the divergence ball's package, d = 0.03, fresh 6.7100 book 6.7100 PASS at the corner: the Wasserstein ball's package, $0.1m 12.1530 book 12.1500 PASS at the corner: the Wasserstein ball's package, $0.1m, fresh 12.0510 book 12.0500 PASS at the corner: the Wasserstein ball's package, $1m 5.2280 book 5.2300 PASS at the corner: the Wasserstein ball's package, $1m, fresh 5.1540 book 5.1500 PASS at the corner: the moment package 0.3540 book 0.3500 PASS at the corner: the moment package, fresh 0.3690 book 0.3700 PASS Chapter 10's max-min package: line 231.6000 book 231.6000 PASS Chapter 10's max-min package: delay 0.4494 book 0.4490 PASS Chapter 10's max-min package: cost (workbook) 10.5429 book 10.5400 PASS Chapter 10's max-min package: chance under the law 5.6370 book 5.6400 PASS price of ambiguity (workbook) 13.5626 book 13.5600 PASS value of a record twice as long (workbook) 4.6326 book 4.6300 PASS Table 11.3: chances on fresh paths within 0.3 points 1.0000 book 1.0000 PASS Section 11.5: every chance in Table 11.3 has a standard error below 0.12 points (workbook) 1.0000 book 1.0000 PASS Table 11.3, One law (Chapter 9): line 170.0000 book 170.0000 PASS Table 11.3, One law (Chapter 9): delay 0.6676 book 0.6700 PASS Table 11.3, One law (Chapter 9): cost (workbook) 10.4406 book 10.4400 PASS Table 11.3, One law (Chapter 9): under the law 5.0000 book 5.0000 PASS Table 11.3, One law (Chapter 9): under the worst law 15.7890 book 15.8000 PASS Table 11.3, A set, max-min choice (Chapter 10): line 231.6000 book 232.0000 PASS Table 11.3, A set, max-min choice (Chapter 10): delay 0.4494 book 0.4500 PASS Table 11.3, A set, max-min choice (Chapter 10): cost (workbook) 10.5429 book 10.5400 PASS Table 11.3, A set, max-min choice (Chapter 10): under the law 5.6370 book 5.6000 PASS Table 11.3, A set, max-min choice (Chapter 10): under the worst law 16.1600 book 16.2000 PASS Table 11.3, Wasserstein ball, $0.1 million: line 165.0000 book 165.0000 PASS Table 11.3, Wasserstein ball, $0.1 million: delay 1.1480 book 1.1500 PASS Table 11.3, Wasserstein ball, $0.1 million: cost (workbook) 14.1344 book 14.1300 PASS Table 11.3, Wasserstein ball, $0.1 million: under the law 3.2330 book 3.2000 PASS Table 11.3, Wasserstein ball, $0.1 million: under the worst law 12.1530 book 12.2000 PASS Table 11.3, Divergence ball, d = 0.01: line 175.0000 book 175.0000 PASS Table 11.3, Divergence ball, d = 0.01: delay 1.3539 book 1.3500 PASS Table 11.3, Divergence ball, d = 0.01: cost (workbook) 16.0812 book 16.0800 PASS Table 11.3, Divergence ball, d = 0.01: under the law 2.4980 book 2.5000 PASS Table 11.3, Closures' doubt (rate <= 0.315): line 210.0000 book 210.0000 PASS Table 11.3, Closures' doubt (rate <= 0.315): delay 0.6672 book 0.6700 PASS Table 11.3, Closures' doubt (rate <= 0.315): cost (workbook) 11.6373 book 11.6400 PASS Table 11.3, Closures' doubt (rate <= 0.315): under the law 4.6310 book 4.6000 PASS Table 11.3, Closures' doubt (rate <= 0.315): under the worst law 14.4310 book 14.4000 PASS Table 11.3, Spread's doubt (spread <= 61.6): line 170.0000 book 170.0000 PASS Table 11.3, Spread's doubt (spread <= 61.6): delay 2.2184 book 2.2200 PASS Table 11.3, Spread's doubt (spread <= 61.6): cost (workbook) 22.8469 book 22.8500 PASS Table 11.3, Spread's doubt (spread <= 61.6): under the law 1.0630 book 1.1000 PASS Table 11.3, Spread's doubt (spread <= 61.6): under the worst law 5.8580 book 5.9000 PASS Table 11.3, The records' family (both): line 200.0000 book 200.0000 PASS Table 11.3, The records' family (both): delay 2.2504 book 2.2500 PASS Table 11.3, The records' family (both): cost (workbook) 24.0032 book 24.0000 PASS Table 11.3, The records' family (both): under the law 0.8940 book 0.9000 PASS Table 11.3, Records twice as long: line 195.0000 book 195.0000 PASS Table 11.3, Records twice as long: delay 1.6901 book 1.6900 PASS Table 11.3, Records twice as long: cost (workbook) 19.3707 book 19.3700 PASS Table 11.3, Records twice as long: under the law 1.6640 book 1.7000 PASS Table 11.3, Records twice as long: under the worst law 7.6650 book 7.7000 PASS Table 11.3, Moments only (every quarter): line 160.0000 book 160.0000 PASS Table 11.3, Moments only (every quarter): delay 6.0618 book 6.0600 PASS Table 11.3, Moments only (every quarter): cost (workbook) 53.2947 book 53.2900 PASS Table 11.3, Moments only (every quarter): under the law 0.0260 book 0.0300 PASS Table 11.3, Moments only (every quarter): under the worst law 0.3540 book 0.4000 PASS Table 11.3, the records' family: under the worst law (the grid's corner) 5.0000 book 5.0000 PASS 358 of 358 checkpoints PASS
Next. Chapter 12, Multi-Objective Enterprise Optimization, opens Part IV. Table 11.3 already sets two objectives side by side, a package's cost and its protection under the records' worst law; Chapter 12 treats such trade-offs directly, with Pareto frontiers and the prices along them. AXIOM-11 runs the same models interactively (Lab 2.11); this notebook and the workbook are its reproducible record.


