space-rha.comShort courseChapter 8 · Data analysis and interpretation

Chapter 8 · Lesson 1 of 4 · about 18 minutes

Counting events and bounding with zero

One ion can upset several cells, leave a persistent wrong state or stop a device that then produces thousands of bad outputs. Decide what the counter counts before you compute a cross section, because the system may care about corrupted words, interrupts or lost functions rather than raw bit errors.

What you need to know

Before you compute a cross section, decide what the counter counts. A single ion can flip one cell, several adjacent cells, or a whole word of the error-correction code. It can leave a persistent wrong state that is then read a thousand times, or stop a controller that produces nothing but errors until it is reset. Different parts of the system care about different counts. The scrubber cares about physical events and how many cells each one touched, the error-correction code about errors per protected word, and the operations team about interrupts and how they were recovered. The classification table of the run record keeps the timestamps, addresses, patterns, durations and recovery steps so that the same raw record can be counted each of these ways. Where the record cannot tell which case it was, the table says so rather than placing the event in the most convenient category. The corrupted output after a SEFI is not counted as upsets. A persistent error is one event however many times it is read, and a multi-cell upset is one event with a multiplicity.

Counts of independent events in a fixed exposure follow a Poisson distribution, and the cross section is the count divided by the live fluence (chapter 7). The familiar square-root uncertainty on the count holds only when the count is large. Garwood gave the exact limits on the mean μ for N observed counts in 1936: the lower limit is the μ for which the probability of N or more counts is α/2, and the upper limit is the μ for which the probability of N or fewer is α/2, where 1 − α is the confidence. For N = 3 at 95 % the interval runs from 0.62 to 8.77, that is from 21 % to 292 % of the estimate. For N = 100 it runs from 81 % to 122 %, close to the 1 ± 2/√N approximation. The figure shows the exact interval closing on the estimate as the count grows. Below about twenty counts the approximation is poor and the interval is asymmetric.

A run with no events still gives a number. With zero counts the one-sided upper limit at confidence 1 − α is −ln α, which at 95 % is 2.996, so the cross section is below about 3 divided by the live fluence. The bound holds only under the conditions of the run: the event definition, the detection chain, the live fluence, the bias and the temperature. It is not evidence that the mechanism is absent, and the three in the numerator cannot be argued down. Only more fluence tightens it. A run that stopped at a fixed count rather than a fixed fluence has a different likelihood and needs an interval computed for that stopping rule (chapter 7 lesson 1).

The counting interval is the smallest uncertainty in a cross section, and reporting it alone misleads the reader. Dosimetry (chapter 7 lesson 2), population transfer (chapter 6 lesson 1), unobserved modes and the choice of model (chapter 8 lesson 2) are generally larger, and lesson 8.4 puts them in an uncertainty budget together with the interval. A 95 % interval on a count does not give 95 % confidence in the engineering claim that uses it.

Figure 8.1. Exact 95 % limits on a Poisson mean, divided by the count. Below twenty counts the interval is wide and asymmetric; the normal approximation (dashed) is adequate only above about fifty.
Exact 95 % limits on a Poisson mean, relative to the count 1 10 100 0.1 1 10 events counted limit divided by the count the estimate itself upper limit / N lower limit / N dashed: the large-count approximation 1 ± 2/√N
Try it: counts to a cross section with exact limits

Enter a count and a live fluence; change the confidence; divide by the number of elements exercised for a per-bit value.

Worked example 8.1. Three events, and none

A U2 SEFI run at LET 21 delivered 1.0 × 10⁶ ions per cm² of live fluence and recorded 3 interrupts, each cleared by a reset. The cross section is 3 / 1.0 × 10⁶ = 3.0 × 10⁻⁶ cm² per device. The exact 95 % interval on the mean count is 0.62 to 8.77, so the cross section lies between 6.2 × 10⁻⁷ and 8.8 × 10⁻⁶ cm² at 95 %. The upper limit is nearly three times the estimate and the lower limit a fifth of it. The run at LET 3.5 delivered 1.0 × 10⁶ ions per cm² with no interrupt, so the cross section there is below 3.0 × 10⁻⁶ cm² at 95 % (one-sided). That is the point drawn as an open downward triangle in the plot of lesson 8.2; it is a bound, not a zero.

Count NLower 95 %Upper 95 %One-sided 95 % upper
00.0003.693.00
10.0255.574.74
20.2427.226.30
30.6198.777.75
51.62311.6710.51
104.79518.3916.96
2012.21730.8929.06

The table is in units of counts; dividing by the live fluence gives cross sections. Quote the one-sided upper bound for a clean run and the two-sided interval with an estimate.

The four teaching parts in this chapter

Four constructed parts run through chapters 6 to 12 so that each chapter's work can be followed on the same hardware: U1, a shared regulator that powers sensing, logic and the recovery controller; U2, a 1 Gbit recorder memory under SEC-DED error correction and periodic scrubbing; U3, a precision analog chain with an accuracy allocation; and U4, an optical interface that must keep its signal margin. Their response curves are stated in the build (constructed from a declared true curve with Poisson sampling), so that every number in these chapters can be traced; they are teaching cases, not commercial devices.

PartFunctionThis chapter's step
U1 · shared powerA shared regulator powers sensing, logic and the recovery controller.Classify each signature and its recovery before pooling the counts. Keep destructive outcomes and the exposure survived without one as separate records.
U2 · recorder memoryA memory stores science data under ECC and periodic scrubbing.Separate the bits flipped by one event from repeated reads of one persistent error, and state the per-bit and per-device denominators.
U3 · precision analogA precision analog chain must maintain an application accuracy and timing allocation.Keep the censored parts and every parameter trajectory, including the unfavorable ones. A tight confidence interval on the mean says little about the weak tail of the lot.
U4 · optical interfaceAn optical interface must deliver enough signal margin throughout the mission.Follow each part's endpoint through to its consequence in the circuit. Check any NIEL-based transfer against the response and annealing behavior that apply to the part.

Exercise

For every cross section you report, write down how an event was defined, which classification rules were applied and which cases the record could not resolve.

Give the count, the live fluence, the stopping rule and the confidence convention with each estimate or bound, so that a reader can recompute the interval.

Going deeper

When the classification is uncertain

The clustering window and the readout cadence can merge two independent strikes into one event or split one physical event into two. Whether several physical upsets become one uncorrectable word depends on the physical layout, the memory interleaving and the state of the device at the time. When the raw record cannot say which it was, analyze the record under each plausible classification and keep the unresolved cases in the report.

Report the event-level sensitivity and the consequence for the application as separate results, each with its own denominator and the limits of what it can say.

What the interval depends on

A counting interval holds for the classification, the detection chain, the fluence and the Poisson process assumed behind them. It does not include modes the test missed, the transfer to the flight population, energy dependence or a wrong model. A run that stopped after a set number of events needs an interval computed for that sampling plan; the fixed-exposure formula does not apply to it.

The statistical confidence on the count is one thing, and the probability that the whole engineering claim is right is another. Keep the two apart in the report.

The common mistake

Two events that look the same at the interface do not necessarily come from the same physical mechanism.

A 95% interval on the count is not 95% confidence in the mission claim built on it.

Takeaway

  • Counting the wrong thing distorts the rate even when the dosimetry is perfect.
  • A finite exposure can bound a cross section but cannot show that it is zero.

Check yourself

A memory test reports 1,000 bit errors. Why is the Poisson interval on 1,000 not yet the uncertainty of the event cross section?

Because the 1,000 bit errors may come from far fewer physical events (multi-cell upsets, persistent errors read repeatedly, the corruption after a SEFI). The Poisson count is the number of independent events, and the classification decides what that number is.

Two runs at the same LET gave 0 events in 10⁶ and 2 events in 10⁷. Can they be pooled?

If the conditions were the same, yes. The pooled result is 2 events in 1.1 × 10⁷, and the Poisson interval is computed on 2. Pooling runs taken at different bias, temperature or mode assumes a model, and that assumption has to be stated and justified.

Why does a clean run at 10⁶ ions per cm² not support a cross-section bound of 10⁻⁷ cm²?

The 95 % one-sided bound from zero counts is 3.00 / 10⁶ = 3.0 × 10⁻⁶ cm². To bound at 10⁻⁷ the fluence must reach 3.0 × 10⁷ ions per cm².

A clean run with no events is written up as showing that the part is "immune" to SEFI. Why does the course insist on a bound with its fluence instead, and what does the bound depend on?

A finite exposure can bound a cross section but cannot show that it is zero. With zero counts the one-sided 95 % upper limit is 2.996 counts, so the cross section is below about 3 divided by the live fluence, which for the LET 3.5 run of worked example 8.1 is 3.0 × 10⁻⁶ cm². The bound holds only under the conditions of the run (the event definition, the detection chain, the live fluence, the bias and the temperature), and it is not evidence that the mechanism is absent. The three in the numerator cannot be argued down; only more fluence tightens the bound, and the report quotes the bound, the fluence and the conditions together.

A report quotes the 95 % interval on a three-event cross section as the estimate times 1 ± 2/√N. What is wrong with that interval, and above what count does the approximation become adequate?

The exact 95 % interval for N = 3 runs from 21 % to 292 % of the estimate, so the upper limit is nearly three times the estimate and the lower limit about a fifth of it, and a symmetric interval centered on the estimate misstates both ends. Below about twenty counts the approximation is poor and the interval is asymmetric, and the figure shows the normal approximation adequate only above about fifty. At N = 100 the exact interval, 81 % to 122 %, is close to 1 ± 2/√N; for small counts the exact Garwood limits are quoted with their asymmetry.

A test procedure stops each run after ten events to save beam time. Can the fixed-exposure Poisson interval be applied to those runs?

Generally not as it stands. The Garwood interval assumes a fixed exposure in which the count is the random quantity; a run stopped at a set count makes the fluence the random quantity, which gives a different likelihood, and the interval has to be computed for that sampling plan (chapter 7 lesson 1). This is why the count, the live fluence, the stopping rule and the confidence convention are reported with each estimate or bound, so that a reader can recompute the interval.

Further reading

  • F. Garwood, "Fiducial limits for the Poisson distribution," Biometrika, vol. 28, no. 3/4, pp. 437-442, 1936.
  • G. J. Feldman and R. D. Cousins, "Unified approach to the classical statistical analysis of small signals," Phys. Rev. D, vol. 57, pp. 3873-3889, 1998.
  • R. Ladbury, "Statistical properties of SEE rate calculation in the limits of large and small event counts," IEEE Trans. Nucl. Sci., vol. 54, no. 6, pp. 2113-2119, 2007.
  • JEDEC JESD57A, "Test procedures for the measurement of single-event effects in semiconductor devices from heavy-ion irradiation," 2017.
  • NASA, "Avionics Radiation Hardness Assurance (RHA) Guidelines," NASA/TM-20210018053, 2021. NTRS 20210018053.