What does Differential Privacy for Business Statistics conclude, in numbers, about “How Large Should Epsilon Be: the Dispute over the 2020 United States Census and the Price of Small Numbers”?

Answer: Chapter 6 reads the choice of epsilon through a posterior ceiling and a table of relative errors, and sets out the dispute over the total budget of 19.61 for the redistricting data of the 2020 United States census: at epsilon 1 a count of 3 has a relative half-width of 0.999.

Scope of the numbers: Chapter 3 deals with sensitivity: revenue clipped per person has a root mean square error of 6545 yuan in the illustrative data, 4.09 % of revenue, and a customer profile spends the budget only once by parallel composition. Marked boxes separate published evidence, runnable Python calculations (20 programs), classical sources and the author's hypotheses.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 19.61

largest posterior after an eps-DP release, by prior belief
epse**epsp0 0.001p0 0.01p0 0.1p0 0.5
0.101.10.00110.01100.10940.5250
0.501.60.00160.01640.15480.6225
1.002.70.00270.02670.23200.7311
2.007.40.00730.06950.45090.8808
2.4711.80.01170.10670.56780.9220
5.00148.40.12930.59990.94280.9933
8.002981.00.74900.96790.99700.9997
12.20198789.20.99500.99951.00001.0000
17.1427784809.31.00001.00001.00001.0000
19.61328484430.81.00001.00001.00001.0000
eps at which prior 0.001 can reach even odds: 6.91
eps at which prior 0.01 can reach even odds: 4.60
Show the full program
# Chapter 6. p61: what a given eps allows an attacker to believe.
# Worst case: the attacker knows everyone else and holds prior belief
# P0 that a fact about the target is true. After any eps-DP release the
# posterior is at most P0 * e**eps / (1 - P0 + P0 * e**eps). The rows
# include the totals the U.S. Census Bureau announced for its 2020
# redistricting data (12.2 in the April 2021 demonstration; 17.14 for
# persons, 2.47 for housing units and 19.61 in all, June 2021). This is
# an upper bound for the strongest attacker, not a measured attack rate.
import math

ROWS = [0.1, 0.5, 1.0, 2.0, 2.47, 5.0, 8.0, 12.2, 17.14, 19.61]
PRIORS = [0.001, 0.01, 0.1, 0.5]


def posterior(p0, eps):
    up = p0 * math.exp(eps)
    return up / (1.0 - p0 + up)


print("largest posterior after an eps-DP release, by prior belief")
print("%7s %13s" % ("eps", "e**eps") + "".join(
    "%10s" % ("p0 %g" % p) for p in PRIORS))
for e in ROWS:
    print("%7.2f %13.1f" % (e, math.exp(e)) + "".join(
        "%10.4f" % posterior(p, e) for p in PRIORS))
for p in PRIORS[:2]:
    need = math.log((1.0 - p) / p)
    print("eps at which prior %g can reach even odds: %.2f" % (p, need))

Single source and citation

Fan, Y. (2026). Differential Privacy, the Laplace Mechanism and Privacy-Budget Accounting: Publishing Business Statistics without Exposing Any Individual (Version 1.0.0), Chapter 6: How Large Should Epsilon Be: the Dispute over the 2020 United States Census and the Price of Small Numbers. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-296 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026296,
  author    = {Fan, Yafu},
  title     = {Differential Privacy, the Laplace Mechanism and Privacy-Budget Accounting: Publishing Business Statistics without Exposing Any Individual},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book