Find the question you need
Why the mean is incomplete · Variance proof · Exact tails · Chebyshev bound · Full index
Two exact probability models share mean zero and variance one but differ in chance of a deviation of three or more. The full event probability matters more than a moment summary when a limit must be respected.
Chapter index
Advanced worked companion: events and bounds
Chapters 1–3
1. One mean and standard deviation, two very different tails · 2. Declare the events and their denominators · 3. Compute the expectations
Chapters 4–6
4. Compute and independently verify variance · 5. Calculate the exact extreme-event probabilities · 6. Compare higher-order moments
Chapters 7–9
7. Construct an entire family of matching moments · 8. Use a variance bound without treating it as a forecast · 9. Why a normal-distribution memory rule is not enough
Chapters 10–12
10. Multiple draws need a dependence assumption · 11. An extreme-risk statistic is not the same as expected downside · 12. Audit and connected reading
Imagine two fictional measurement processes with population mean zero and standard deviation one. That agreement seems strong until the decision asks how often the measurement crosses an absolute threshold of three units. One constructed process can never cross it; another does so on four percent of draws.
Let X be either −1 or +1, each with probability 0.5. Let Y equal zero with probability 0.96 and equal −5 or +5 with probability 0.02 each. These are original mathematical models, not empirical measurements or predictions for a real safety or financial problem.
The example establishes an information limit: mean and variance summarise the centre and dispersion, but do not uniquely identify extreme-event probabilities. The entire distribution or an appropriate bound is needed to evaluate a specified tail event.
For X, the probabilities of −1 and +1 sum to one. For Y, 0.02+0.96+0.02=1. Each variable is a single draw from its fully stated discrete probability distribution.
Define a two-sided extreme as |Z|≥3. The inclusive threshold matters for outcomes exactly equal to three. A separate negative-loss event Z≤−3 counts only values in the lower tail, not both signs.
A probability about one draw is different from the chance that at least one event occurs over a series of draws. The latter depends on assumptions about dependence between observations, which we will state when needed.
E[X]=0.5(−1)+0.5(+1)=0. The two possible values are opposites with equal probability, so they cancel in the expectation, though no single draw of X is zero.
E[Y]=0.02(−5)+0.96(0)+0.02(+5)=0. The probability-weighted contributions of the two rare extreme values cancel in exactly the same way.
Identical expectations do not imply identical supports, likely observations or negative-outcome risks. The mean is useful for a mean question but not enough to describe the shape of a distribution.
With mean zero, variance equals E[Z²]. For X, E[X²]=0.5×1+0.5×1=1. Its standard deviation is exactly one.
For Y, E[Y²]=0.02×25+0.96×0+0.02×25=1. It has the same variance and standard deviation despite placing most probability mass at zero.
This equality is exact under the stipulated population models. A finite sample would generally have estimated moments that fluctuate, so the model should not be confused with observed sample data.
X can only take magnitudes of one. Therefore P(|X|≥3)=0 and P(X≤−3)=0. A three-unit magnitude is outside its possible support.
Y has two five-unit outcomes. P(|Y|≥3)=0.02+0.02=0.04, while P(Y≤−3)=0.02. The two-sided and one-sided chances differ because they refer to different events.
The precise conclusion is zero versus four percent for the two-sided threshold, even though the means and variances match. A dashboard containing only mean and SD cannot reconstruct this difference.
E[X⁴]=1, because every possible fourth power is one. For Y, E[Y⁴]=0.04×5⁴=25, because the zero state contributes nothing.
Each variance is one, so the non-excess kurtosis values are one for X and twenty-five for Y. Excess kurtosis is obtained by subtracting three, yielding −2 and 22.
Fourth moments amplify contributions from large magnitudes, but they are themselves summaries. When the practical question is the probability of crossing an explicit threshold, the exact event probability remains the more direct measure.
For a≥1, define Z_a as zero with probability 1−1/a² and as +a and −a with probability 1/(2a²) each. Its mean is zero by symmetry.
Its variance is a²×[1/(2a²)+1/(2a²)]=1. Therefore the parameter a can move nonzero events farther from zero while preserving the same mean and variance.
At a=3 the chance of |Z_a|≥3 equals 1/9; at a=5 it equals 1/25=0.04. This family demonstrates mathematically why second moments do not determine a tail event.
Chebyshev’s inequality states that a finite-variance distribution satisfies P(|Z−μ|≥kσ)≤1/k² for k>0. With μ=0, σ=1 and k=3, the bound is 1/9, about 11.111 percent.
X’s exact zero percent and Y’s four percent both satisfy the bound. The family Z_3 puts precisely 1/9 probability at magnitudes three, attaining the inclusive bound.
The Book of Statistical Proofs derives this inequality. The bound is a worst-case guarantee across a broad family, not an estimate that eleven percent of real observations actually exceed three standard deviations.
A bell-shaped empirical rule is useful only under suitable distributional assumptions. Our X and Y are finite discrete distributions, not normal laws, so a normal-tail approximation is unnecessary and potentially misleading.
Because every model outcome and probability is specified, direct summation gives exact tail probabilities. This is stronger evidence within the invented model than invoking an approximate percentage from an unrelated distribution.
When analysing real observations, first ask whether the assumed shape is justified, whether rare observations were captured and whether the quantity is stationary across the chosen period.
Suppose ten independent observations are taken from Y. Each has four-percent chance of |Y|≥3. The probability all ten avoid an extreme is 0.96¹⁰≈0.664833.
Thus the chance at least one ten-draw extreme occurs is 1−0.96¹⁰≈0.335167, or about 33.52 percent. For X the corresponding chance is zero.
This calculation is conditional on independence. If extremes cluster together through a hidden common cause, the ten-draw chance can change even if each observation still has the same four-percent marginal event probability.
CHAPTER 11 OF 12
11. An extreme-risk statistic is not the same as expected downside
X has a negative outcome with probability one-half, while Y has a negative outcome with probability only two percent. At a negative threshold of −3, however, X has zero chance and Y two percent.
The expected magnitude of a negative part, E[max(−Z,0)], is 0.5 for X but 0.1 for Y. That ranking differs from the three-unit extreme-event ranking.
Neither variable is universally less risky for every objective. Choose a metric that reflects the actual event and severity of concern, and do not let the equal variance conceal the distribution of consequences.
Record the variable’s units, the event definition and inclusive threshold, the probability model, the evidence for its parameters and whether multiple observations are independent. Label exact model results separately from estimated rates or mathematical upper bounds.
Practice constructing Z_4 and verifying that it has mean zero, variance one and P(|Z_4|≥3)=1/16. Explain why raising the threshold to five makes that particular event impossible.
Continue with The Gold Standard of Probability, Statistics, Chebyshev’s inequality and the Gold Standard hub. All distributions here are original teaching examples.
Advanced worked companion: moments, tail bounds and decision events
1. Check every possible outcome before invoking a theorem
The support of X consists only of −1 and +1, with probability one-half each. Therefore no magnitude above one is possible within this fully specified model. The event |X|≥3 is empty, and its probability is exactly zero, not merely small. The support of Y instead contains −5 and +5 with a total probability of four percent, which means the identical threshold is reached on those two rare outcomes.
This event-based audit is preferable to inferring a tail percentage from a familiar standard deviation. A complete distribution permits a direct count of eligible probability mass. The event sign also matters: Y≤−3 has probability two percent, not four, because only its negative five-unit state qualifies. Whenever a performance report states ‘extreme chance’, it must tell the reader whether the extreme event includes positive, negative or both directions.
2. How X and Y accumulate identical variance by different means
For X every draw has squared deviation one from the zero mean. Each possible outcome contributes one to the squared magnitude, and average squared magnitude is also one. In Y, ninety-six percent of draws are exactly zero, hence contribute no squared deviation at all. Its nonzero four percent, however, each have squared deviation twenty-five. The weighted sum is 0.04×25=1.
The same second moment has been assembled from two different probability arrangements. This demonstrates that equal variance does not imply equal event frequency, common absolute deviation or common maximum possible value. If a physical threshold trips at three units, Y can activate it and X cannot. If a process instead reacts to any departure from exactly zero, X activates on every draw and Y on only four percent. A metric must reflect the actual criterion.
3. Fourth moments reveal another dimension of shape
X’s fourth moment E[X⁴] is one because ±1 each have fourth power one. Y’s fourth moment is 0.04×5⁴=25. Both have variance one, so their fourth central moment divided by the squared variance—the non-excess kurtosis—is one and twenty-five, respectively. Subtract three for the excess values −2 and 22 under the standard definition.
The fourth-power statistic gives much greater numerical weight to large deviations than variance does, but it remains a summary. Different distributions can share several moments and still have different threshold-event probabilities. In a task where the probability of exceeding a defined engineering limit matters, a full supported distribution or a valid tail bound is usually more interpretable than using kurtosis alone as a verdict.
4. Compare all thresholds, not only the convenient one
For any positive threshold t no greater than one, X satisfies P(|X|≥t)=1 because it always has magnitude one. If t exceeds one, this probability becomes zero. Y behaves differently: for every t strictly above zero and no greater than five, the exceedance probability is 0.04. When t>5 it becomes zero as well.
Thus at t=0.5, X exceeds the threshold on every draw while Y does so only four percent of the time. At t=3, X never exceeds while Y retains four percent. At t=6, neither can exceed. Neither variable can be declared universally more extreme under every threshold rule. The answer is a property of the specified event and distribution, rather than a permanent label attached to the mean or variance.
5. Derive a whole family of equal-variance models
For any a≥1, construct Z_a that equals zero with probability 1−1/a², equals +a with probability 1/(2a²), and equals −a with the same small probability. Its expected value is zero by symmetry. Its variance is a² times the total probability of a nonzero draw, which equals a²×1/a²=1.
This family links X and Y directly. At a=1, the zero state has probability zero and Z_a is just X. At a=5, the zero state has probability 24/25=0.96, with probability 1/50=0.02 at each of ±5, giving Y. At a=10, nonzero events occur only one percent of the time but reach magnitude ten. The unchanged second moment cannot specify event size or rarity separately.
6. Find the maximum three-sigma tail from a variance bound
Chebyshev’s inequality applies to any finite-variance random variable. If mean is μ and standard deviation is σ, then P(|Z−μ|≥kσ)≤1/k² for positive k. With μ=0, σ=1 and k=3, that universal upper bound is 1/9, or approximately 11.111 percent.
Take Z_3 from the constructed family. It is zero with probability 8/9 and equals each of +3 and −3 with probability 1/18. It still has variance one, and because the threshold is inclusive, its three-unit exceedance probability is exactly 1/9. The upper bound is therefore attainable; it is not an arbitrary loose warning. The Book of Statistical Proofs gives the theorem’s derivation.
7. A universal bound is not an estimated event rate
The bound of 1/9 says that any model meeting the first-two-moment assumptions has at most that much probability outside the specified three-standard-deviation threshold. It does not say that all such distributions have that probability. X has zero, Y has four percent, and Z_3 has the full 11.111 percent. These different values all satisfy the exact same variance constraint.
A field report that quotes the Chebyshev bound should distinguish ‘at most’ from ‘expected to occur’. The upper bound can be useful when little distributional information is available, but it may be much less informative than a justified empirical or mechanistic distribution. A numerical bound cannot identify whether a real process resembles X, Y or another member of an admissible family.
8. Do not import a normal-tail rule without evidence
The familiar normal empirical rule is an approximation for a particular bell-shaped family of distributions. It cannot be treated as a universal theorem about any mean-and-variance pair. X is a two-point law; Y is a three-point law, with a large probability spike at zero and symmetric distant extremes. Neither follows a normal density.
When a complete probability table is known, there is no reason to replace exact event summation with an unrelated normal-tail estimate. For real observed data, a fitted parametric distribution may be helpful but needs goodness-of-fit checks, adequate tail observations and attention to censoring. The LibreTexts distinction between the empirical rule and Chebyshev’s theorem explains their different assumptions.
9. A series of observations asks a different probability question
For one draw of Y, P(|Y|≥3)=0.04. Suppose ten independent draws are observed and a monitoring rule triggers whenever at least one of them crosses the threshold. The probability all ten avoid the event is 0.96¹⁰≈0.664833. Its complement is approximately 33.517 percent.
This chance is far larger than four percent because the observer has ten opportunities for an event. But it depends on independence. Correlated failures, shared environmental shocks or regime changes can make events cluster, changing the chance of at least one exceedance in a given window. A report that moves from ‘per sample’ to ‘per day’ or ‘per batch’ must state the number of observations and a defensible dependence model.
10. Perfectly dependent extremes have a different window probability
Consider a separate hypothetical ten-observation sequence controlled by one binary common shock. With probability 0.04, the entire sequence is in an extreme state; otherwise no observation is extreme. Each position still has a four-percent marginal chance of an extreme, like a single Y draw.
However, the chance of at least one extreme anywhere in that ten-observation sequence is now exactly four percent, not 33.52 percent. One-draw marginal rates alone therefore do not determine a window-level event probability. The independence calculation and common-shock calculation describe different joint processes with matching one-time event probabilities.
Neither example is a prediction of a particular monitoring network. They establish the need for sequence or dependency information when the operational decision concerns multiple observations rather than one isolated draw.
11. Finite-sample observations and rare-event absence
In a hundred independent Y draws, expected extreme count is four. Actual counts vary around that expectation according to the underlying Bernoulli indicators. The chance of observing no three-unit extremes in one hundred draws is 0.96¹⁰⁰≈0.016870, which is small but nonzero.
This means a finite log with no observed extreme event cannot, on its own, prove the population event probability is zero. X’s zero extreme probability was established by the modelled support, not by failing to see a rare event in a sample. In real applications the population distribution is usually not known exactly, and sampling design, recording completeness and tail uncertainty can be the main challenges.
A sound estimate should show its observation count, time window and uncertainty rather than report a point percentage with unsupported decimal precision.
12. Frequent small losses versus rare larger losses
Define a negative event as Z<0. X has a negative outcome on fifty percent of draws, with loss magnitude one. Y has a negative outcome on only two percent, but its negative magnitude is five. By this event definition, X is more frequently negative even though Y has the more severe negative state.
Define the larger-loss event as Z≤−3 and the ranking reverses: X has probability zero while Y has two percent. The expected magnitude of the negative part is E[max(−X,0)]=0.5 and E[max(−Y,0)]=0.1. This third measure ranks them differently again.
A decision maker should not describe one model as unconditionally safer from all possible perspectives. Loss frequency, magnitude, threshold violations and expected downside cost are different quantities.
13. Changing the threshold can change a decision
Suppose a fictional device has an absolute limit of 2 units. X cannot exceed it because |X| always equals one. Y can exceed it on its four-percent nonzero states. An operator with a strict no-trip rule might therefore prefer the X-type distribution under the assumed model.
Suppose instead a process is sensitive to any nonzero movement and such movement incurs a fixed handling cost. X triggers every time, while Y triggers only four percent of the time. Under that alternative event cost, Y can look preferable. The mean and variance remain identical in both comparisons.
Neither hypothetical cost function establishes a real-world preference. The exercise shows why the operational threshold and consequence model belong in the analysis before a statistic is selected.
14. Estimate uncertainty honestly when fitting a tail model
A tail probability obtained from observations may change with sample selection, instrument resolution, data cleaning, clustering, seasonality and the choice of threshold. Rare-event data can be especially sensitive to a handful of extreme observations or records that were censored during a measurement failure.
A model that assumes a stationary independent distribution must be tested where possible. If the generating process shifts or observations share common causes, multiplying a single estimated event rate by an assumed number of independent trials can yield misleading results. A sensitivity report should consider plausible alternative assumptions.
The exact discrete models here are useful because they allow clear proofs. They are not adequate substitutes for a real safety analysis, probability calibration study or physical failure investigation.
15. An auditable table makes the hidden difference visible
A practical table should include each possible value, its probability, whether it crosses the chosen threshold, and its contribution to mean and variance. For X the two possible squared deviations each contribute 0.5 to the expected square. For Y the two rare extremes contribute 0.5 each while the common zero contributes nothing.
The same table can be used to calculate the fourth moment, one-sided shortfall and event chance at multiple thresholds. It prevents arithmetic shortcuts from concealing that the distributions are different even when mean and variance agree.
The correct reporting sequence is to define the variable and threshold, verify the probability mass sums to one, compute requested metrics, and state whether calculations are theoretical or empirical. These steps are more trustworthy than a headline ‘standard deviation one’ on its own.
16. Practice with the four-unit example
Choose a=4 in the Z_a family. The probabilities are 1/32 at −4, 15/16 at zero, and 1/32 at +4. Mean is zero and variance is 2×(1/32)×16=1. At threshold three, the two nonzero outcomes qualify and the total exceedance probability is 1/16=6.25 percent.
Now change the threshold to five. Neither ±4 outcome qualifies, so the probability becomes zero. The model, mean and variance are unchanged; only the event definition changes. As another check, compute the fourth moment: 2×(1/32)×4⁴=16, higher than X’s one and lower than Y’s twenty-five.
This series of calculations gives readers a complete, reproducible exercise without inventing missing observed data.
17. Final methodological audit and connected reading
The durable conclusion is exact: equal mean and variance do not uniquely identify extreme-event risk. X and Y share μ=0 and σ²=1, yet the two-sided chance of magnitude at least three is zero versus four percent. Chebyshev supplies a valid broad upper bound, but it does not identify the actual rate in either model.
For evidence beyond classroom probabilities, report the observation population, how often the threshold was tested, whether events are temporally dependent, and uncertainty in any estimated distribution. A modelled support bound is different from an unobserved tail in finite samples. The Chebyshev proof and The Gold Standard of Probability provide foundation.
Return to the Gold Standard hub and How X Works for the broader lesson: a summary statistic is valid for the question it was designed to answer, not every downstream decision.
| Property | X | Y |
|---|---|---|
| Values | −1, +1 | −5, 0, +5 |
| Probabilities | 0.5, 0.5 | 0.02, 0.96, 0.02 |
| Mean | 0 | 0 |
| Variance | 1 | 1 |
| Standard deviation | 1 | 1 |
| P(|Z|≥3) | 0% | 4% |
| P(Z≤−3) | 0% | 2% |
| Fourth moment | 1 | 25 |