CUSUM Control Charts and Average Run Length: the Fastest Alarm within a False-Alarm Budget for Latency Drift and Silent Conversion Drops: chapter answers
People who run online services commonly believe that one well-chosen fixed red line each for P99 latency and conversion is enough. This book overturns that belief with average run lengths: with false alarms kept equally rare, a fixed threshold reacts very slowly to small persistent drifts, while a cumulative sum (CUSUM) chart cuts the detection delay to a fraction. The project blueprint attributed the claim to papers by Page of 1954 and Moustakides of 1986; after checking, the book corrects the blueprint: Page proposed the CUSUM, the worst-case delay criterion and asymptotic optimality come f…
DOI pending · monograph page goes live after publication · 5DL-B-305 · 中文版
- Chapter 1 · The Sluggish Red Line: at Equal False-Alarm Rates the CUSUM Is Several Times Faster
Chapter 1 shows that, with the same budget of one false alarm in 370 days, a drift of half a standard deviation is found by a fixed threshold after 88.9 days on average and by a CUSUM after 27.7 days, a ratio of 3.21; f… - Chapter 2 · One Line of Recursion: the CUSUM Is a Running Log-Likelihood Ratio
To keep the same false-alarm budget with two sides, the limit must rise from 4.096 to 4.775, since now two sums can each raise a false alarm. The price is paid in detection speed. For a shift of half a standard deviatio… - Chapter 3 · The Markov Chain of Brook and Evans: the Average Run Length as a System of Linear Equations
Chapter 3 solves the Markov chain of Brook and Evans by hand-written Gaussian elimination and shows that a CUSUM with one false alarm a year on average still raises at least one within a year with probability 0.628. - Chapter 4 · Siegmund's Approximation and Lorden's Bound: Detection Delay Grows Only with the Logarithm of the False-Alarm Interval
The information per reading is 0.125. With each tenfold increase in the false-alarm interval, the CUSUM's delay rises by 16.2, then 18.1, then 18.4 days, approaching the slope of the bound, the natural logarithm of ten… - Chapter 5 · Design Tables for the Reference Value and the Limit: the Cost of a Wrong Shift and a Wrong Baseline
Now the mean. With a baseline of 14 days, the standard error of the mean is 0.267 standard deviations. If the mean happens to be estimated one standard error too low, the false-alarm interval shrinks to 74.9 days, about… - Chapter 6 · The Boundaries of Optimality: Autocorrelation, Residual Charts and Large Jumps
In the combined rule the CUSUM limit rises from 4.096 to 4.148, and the false-alarm interval is still 370.0 days. The cost falls on small drifts and is small: for half a standard deviation the combined rule needs 28.0 d… - Chapter 7 · Eastern and Western Sources: the Dike of a Thousand Zhang and Shewhart's Three Sigma
Chapter 7 examines the Han Feizi's dike that collapses through ant holes and Shewhart's control chart of 1931. - Chapter 8 · A Daily CUSUM at the Toll Gate: P99 Latency, Silent Conversion Drops and a Falsification Protocol
Chapter 8 sets out a daily CUSUM routine for a toll gate's P99 latency and conversion: when conversion falls by a tenth, the CUSUM alarms after 13.7 days on average and a fixed count rule after 52.8; it also presents th…
Illustrative model calculation, not a live quote