What does Private-Channel Rate Shaping conclude, in numbers, about “Slowing Down on Feedback: Additive Increase, Multiplicative Decrease and Wald's Sequential Test”?

Answer: A small sender with 300 messages a day, whose true rate holds firmly at 0.10%, sees a single day read above 0.30% with probability 0.2593, about one false alarm every four days; pooling a week brings this down to 0.005834. At 1000 a day the daily false-alarm chance is 0.0802 and the weekly one 0.000014. At 5000 a day the daily chance is only 0.0002. The last column gives the reason: the 0.30% line lies only 1.10 standard errors above the true value at 300 a day and 8.95 at 20000 a day.

Scope of the numbers: An illustrative plan averaging 10.0 messages a minute still has 420 refused when sent in batches, and none once a token bucket of depth 20 shapes it. Parameters are illustrative, platform figures carry the document and the date read, and all numbers come from the book's programs, which read and write no file and use no network.

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: 0.000014, 0.0002, 0.005834, 0.0802, 0.10, 0.2593

true rate 0.10%; alarm line 0.30%
per dayP(day >= line)P(week >= line)z of line, day
3000.25930.0058341.10
10000.08020.0000142.00
20000.01650.0000002.83
50000.00020.0000004.47
200000.00000.0000008.95
Show the full program
# Chapter 5. p53: a daily spam-rate reading is noisy at small volume.
# Gmail's sender guidelines (read on 2026-11-13) ask senders to keep the
# spam rate in Postmaster Tools below 0.10% and never to reach 0.30%.
# Suppose a sender's true complaint rate sits exactly at 0.10%. With v
# messages a day, the daily count is Binomial(v, 0.001). The program
# gives the exact chance that one day reads 0.30% or more, and the same
# for a seven-day pooled reading (7 v messages). The ratio column shows
# the 0.30% line in standard errors above the true rate. (Same guide:
# marketing mail must offer one-click unsubscribe, RFC 8058.)
import math

TRUE, LINE = 0.001, 0.003


def tail(n, p, k0):
    """P(K >= k0), K ~ Binomial(n, p), by the log-pmf."""
    terms = []
    for k in range(k0, n + 1):
        lp = (math.lgamma(n + 1) - math.lgamma(k + 1)
              - math.lgamma(n - k + 1) + k * math.log(p)
              + (n - k) * math.log(1 - p))
        if lp < -60 and k > n * p:
            break
        terms.append(math.exp(lp))
    return math.fsum(terms)


print("true rate 0.10%; alarm line 0.30%")
print("per day   P(day >= line)  P(week >= line)  z of line, day")
for v in (300, 1000, 2000, 5000, 20000):
    k_day = math.ceil(LINE * v - 1e-9)
    k_wk = math.ceil(LINE * 7 * v - 1e-9)
    z = (LINE - TRUE) / math.sqrt(TRUE * (1 - TRUE) / v)
    print("%7d   %14.4f  %15.6f  %14.2f"
          % (v, tail(v, TRUE, k_day), tail(7 * v, TRUE, k_wk), z))

Single source and citation

Fan, Y. (2026). Rate Control for Private Messaging Channels without Account Bans: Token-Bucket Shaping, Platform-Published Limits and Consent-Based Reach Throughput (Version 1.0.0), Chapter 5: Slowing Down on Feedback: Additive Increase, Multiplicative Decrease and Wald's Sequential Test. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-271 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026271,
  author    = {Fan, Yafu},
  title     = {Rate Control for Private Messaging Channels without Account Bans: Token-Bucket Shaping, Platform-Published Limits and Consent-Based Reach Throughput},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book