A value can move arbitrarily far from the other observations while its z-score, calculated using that same sample’s mean and standard deviation, stays below a fixed ceiling. The observation helps set the scale against which it is being judged.
The short answer is that, for n observations and a positive sample standard deviation using denominator n − 1, every same-sample score zᵢ = (xᵢ − mean)/s satisfies |zᵢ| ≤ (n − 1)/√n. With four observations the ceiling is only 1.5. Increasing one raw value cannot make that score exceed three.
This article derives the limit, constructs examples that attain or approach it, and compares a fixed reference, a different variance convention and a leave-one-out calculation. All numerical records are fictional. The Z-Score subject guide supplies the wider interpretation of reference populations, normality and responsible investigation.
eduKateSG · Z-Score
Find your next learning step
Derive the ceiling, change the reference and test which thresholds can actually trigger.
ROUTE 1 · CHAPTERS 1–5
Define and prove the ceiling
Reconstruct a four-value example and its sharp general bound.
ROUTE 2 · CHAPTERS 6–10
Change the threshold and scale
Check feasible cutoffs, saturation and variance denominators.
ROUTE 3 · CHAPTERS 11–15
Separate internal and external references
Derive leave-one-out identities and their limits.
ROUTE 4 · CHAPTERS 16–21
Work masking and evidence boundaries
Change group sizes, count possible flags and preserve model assumptions.
ROUTE 5 · CHAPTERS 22–29
Transfer and independently verify
Test new samples, sample growth, geometric proof and arithmetic checks.
Chapter contents
Define and prove the ceiling · Chapters 1–5
1. Name the score before using a threshold
2. Work a four-value example that reaches the ceiling
3. Translate the problem into deviations
Change the threshold and scale · Chapters 6–10
6. Read small-sample ceilings with the comparison sign
7. Derive feasibility for any proposed cutoff
8. Increase one value and observe saturation
Separate internal and external references · Chapters 11–15
11. Compare an external reference that does not move
12. Remove the selected observation from the centre
13. Derive the leave-one-out variance
Work masking and evidence boundaries · Chapters 16–21
16. Explain why several separated values can mask one another
17. Work the group-size change numerically
18. Check how many large scores are even possible
19. Use invariance as an implementation check
Transfer and independently verify · Chapters 22–29
22. Try a changed sample and denominator
23. Build a repair, stabilisation and extension route
24. Add reference observations without changing the selected value
25. Recover the bound from geometric constraints
26. Identify what standardized scores cannot reconstruct
27. Check arithmetic before interpreting a boundary breach
CHAPTER 1 OF 29 · Define and prove the ceiling
1. Name the score before using a threshold
Let n be at least two. We have real observations x₁ through xₙ, their arithmetic mean, and sample variance s² = Σ(xⱼ − mean)²/(n − 1). Assume s is positive. The score being studied is zᵢ = (xᵢ − mean)/s for an observation included in those calculations.
This is different from comparing a new observation with a fixed external mean and standard deviation. It is also different from a z test of a sample mean, whose denominator is a standard error, and from some regression residual conventions that include a leverage adjustment.
The formulas can look similar while having different mathematical constraints. Write the centre, scale, denominator and inclusion rule explicitly before deciding that a threshold of two or three means anything for the statistic at hand.
No probability distribution is needed for our algebraic bound. The observations may be skewed, discrete or deliberately constructed. A later statistical inference about how unusual a value is requires further assumptions; the deterministic constraint does not supply them.
CHAPTER 2 OF 29 · Define and prove the ceiling
2. Work a four-value example that reaches the ceiling
Take [7,7,7,31]. The mean is 13. The deviations are [−6,−6,−6,18], and their squared sum is 3 × 36 + 324 = 432. Dividing by n − 1 = 3 gives sample variance 144 and sample standard deviation 12.
The z-score of 31 is 18/12 = 1.5. Each seven has score −6/12 = −0.5. The top observation is twenty-four raw units above the other three, yet its same-sample score is only one and a half.
A rule that flags only |z| > 3 returns no flag. That output is an exact consequence of the specified rule. It does not establish that 31 was recorded correctly, belongs to the intended population or should be kept without investigation.
The example is intentionally extreme and tidy. Its purpose is to expose a structural constraint, not to establish a recommended outlier procedure. We will now derive why 1.5 is the largest possible absolute score for any nonconstant four-observation sample under this convention.
CHAPTER 3 OF 29 · Define and prove the ceiling
3. Translate the problem into deviations
Write yⱼ = xⱼ − mean and Q = Σyⱼ². By the definition of the mean, Σyⱼ = 0. The sample variance is Q/(n − 1), and the squared score is zᵢ² = (n − 1)yᵢ²/Q.
Fix one observation i. All the other deviations must sum to −yᵢ. A large positive deviation therefore requires negative deviations elsewhere, or a mixture whose total supplies that balance.
Those other deviations also contribute to Q. We cannot make yᵢ large while keeping the denominator’s squared-deviation total arbitrarily small. The same value affects both numerator and scale.
This is the central mechanism. Merely observing that Q contains yᵢ² gives the looser bound |zᵢ| ≤ √(n − 1). The zero-sum requirement forces additional contribution from the other observations and yields the sharper ceiling we need.
CHAPTER 4 OF 29 · Define and prove the ceiling
4. Find the smallest balancing squared sum
Among n − 1 numbers whose sum is −yᵢ, their sum of squares is at least yᵢ²/(n − 1). One proof uses the nonnegativity of squared differences from their own mean, which is −yᵢ/(n − 1).
Expand Σⱼ≠ᵢ[yⱼ + yᵢ/(n − 1)]² ≥ 0. Because the other deviations sum to −yᵢ, the cross terms simplify, leaving Σⱼ≠ᵢyⱼ² − yᵢ²/(n − 1) ≥ 0.
Equality holds precisely when all the other deviations equal −yᵢ/(n − 1). In raw observations, that means all observations except the selected one are identical. Our [7,7,7,31] example has exactly that structure.
This proof explains both the inequality and its equality case. It is stronger than quoting a ceiling without knowing when it can be reached. It also gives a useful diagnostic: variation among the remaining observations makes the selected score strictly smaller than the ceiling when its deviation is nonzero.
CHAPTER 5 OF 29 · Define and prove the ceiling
5. Complete the bound and check its scope
Add yᵢ² to the balancing inequality. We obtain Q ≥ yᵢ² + yᵢ²/(n − 1) = nyᵢ²/(n − 1). Substitute this into zᵢ² = (n − 1)yᵢ²/Q to get zᵢ² ≤ (n − 1)²/n.
Taking square roots gives |zᵢ| ≤ (n − 1)/√n. If yᵢ = 0, the score is zero and the inequality is immediate. The argument therefore covers every observation in every nonconstant real sample under the declared variance convention.
The ceiling is sharp because the equality construction exists. Choose one value different from n − 1 identical values. The selected value reaches the positive or negative ceiling according to which side it occupies.
This is not a normal-distribution result. No random sampling, independence or tail model entered the derivation. It is an arithmetic relationship created by centring and scaling a finite collection with its own statistics.
CHAPTER 6 OF 29 · Change the threshold and scale
6. Read small-sample ceilings with the comparison sign
| Sample size n | Largest possible absolute same-sample score |
|---|---|
| 2 | 1/√2, about 0.707 |
| 3 | 2/√3, about 1.155 |
| 4 | 1.5 |
| 5 | 4/√5, about 1.789 |
| 10 | 9/√10, about 2.846 |
| 11 | 10/√11, about 3.015 |
For n ≤ 10, a strict |z| > 3 rule cannot flag any value under our same-sample convention. This remains true if one observation is replaced by a number a million times larger. The reference mean and spread change along with it.
At n = 11, exceeding three becomes mathematically possible. It does not become automatically probable, appropriate as a decision threshold or sufficient evidence of a recording error.
Keep strict and non-strict comparisons distinct. A ceiling equal to a chosen threshold allows equality but not strict exceedance. The sample convention here does not put a ceiling exactly at three for an integer n, but other denominator conventions can create that boundary.
With only two distinct observations, both absolute scores are exactly 1/√2, whatever their nonzero separation. Their deviations are plus and minus half the raw difference, and their sample standard deviation is the absolute difference divided by √2. This limiting small-sample case is a useful diagnostic exercise: a gap of one unit and a gap of one million units produce the same two scores. The original values and a justified external reference remain necessary for judging the practical size of the gap.
CHAPTER 7 OF 29 · Change the threshold and scale
7. Derive feasibility for any proposed cutoff
For a positive threshold c, strict exceedance is possible only if (n − 1)/√n > c. Squaring gives (n − 1)² > nc². Because the bound is attained by a valid sample, this is also sufficient for the existence of some sample with |z| > c.
For c = 2, n = 5 gives a ceiling below two, while n = 6 gives 5/√6, approximately 2.041. Six is the first sample size that can produce a strict score above two under this definition.
For c = 3, the inequality is n² − 11n + 1 > 0. The relevant positive crossing lies between ten and eleven, agreeing with the table.
This calculation is a feasibility check, not a method for selecting a statistically justified cutoff. It asks whether the rule can ever trigger. A rule can be mathematically feasible and still have unsuitable false-alarm behaviour, weak detection or poor relevance to the measurement problem.
CHAPTER 8 OF 29 · Change the threshold and scale
8. Increase one value and observe saturation
Now use [−1,0,1,L] with L positive. The mean is L/4. The sum of squared deviations is 2 + 3L²/4, so sample variance is 2/3 + L²/4.
The score of L is (3L/4)/√(L²/4 + 2/3). Divide numerator and denominator by L to see the limit as L increases: (3/4)/√(1/4 + 2/(3L²)) tends to 1.5.
The raw distance from the other three observations can grow without bound, while the score approaches a finite value from below. The remaining three values differ, so the exact equality case is not reached for finite positive L.
This example separates two changes that are easily conflated. The measurement becomes more separated in raw units, but it also enlarges the scale estimated from the sample. A stable-looking z-score can coexist with an increasingly extreme raw record under a fixed domain interpretation.
CHAPTER 9 OF 29 · Change the threshold and scale
9. Check a finite value without relying on the limit
Set L = 10. The data are [−1,0,1,10], with mean 2.5 and deviations [−3.5,−2.5,−1.5,7.5]. Their squares sum to 77, giving s² = 77/3.
The score of ten is 7.5/√(77/3), approximately 1.4804. Its squared value is exactly 675/308. This is below the ceiling squared, 9/4, as the proof requires.
Set L = 100 instead. The mean is 25, the sample variance is 2500 + 2/3, and the selected score is 75/√(2500 + 2/3), approximately 1.4998. The much larger raw value still does not cross two or three.
A useful independent check is to calculate Q using Σxⱼ² − n mean² rather than summing the displayed deviations. The two routes should agree. Preserve exact fractions until the final display so that rounding does not obscure the approach to the bound.
CHAPTER 10 OF 29 · Change the threshold and scale
10. Change the variance denominator explicitly
Some software uses the within-record variance Q/n instead of Q/(n − 1). Let s₀ = √(Q/n), and define z₀ᵢ = yᵢ/s₀. Then z₀ᵢ = √[n/(n − 1)]zᵢ.
Its ceiling is therefore √(n − 1), rather than (n − 1)/√n. In [7,7,7,31], s₀ = √108 = 6√3, and the selected score is 18/(6√3) = √3, approximately 1.732.
Both scores describe the same raw observations. Neither is an arithmetic mistake if its convention is stated correctly. They answer differently scaled within-record questions.
For n = 10, the Q/n convention permits a ceiling of exactly three, attained when nine observations are equal and the tenth differs. A strict score above three remains impossible, while a rule using |z₀| ≥ 3 can trigger at equality. This illustrates why both the denominator and the comparison sign belong in a reproducible rule.
CHAPTER 11 OF 29 · Separate internal and external references
11. Compare an external reference that does not move
Suppose an independently specified reference has mean seven and standard deviation two. Comparing the value 31 with that reference gives (31 − 7)/2 = 12. The same raw value had same-sample score 1.5 in [7,7,7,31].
There is no contradiction. The external reference remains fixed when 31 changes. The within-sample reference incorporates it and expands its scale. The formula’s symbols resemble each other, but the reference construction differs.
This does not prove the external score is the right one for the application. The proposed reference may be poorly matched, outdated or uncertain. A large standardized distance against an inappropriate comparison group can be misleading.
The practical question is which population or baseline answers the intended task. Preserve the raw value and reference documentation. Do not replace a same-sample score with a fixed-reference score merely because it produces a more dramatic number, or choose the reference after seeing which result is preferred.
CHAPTER 12 OF 29 · Separate internal and external references
12. Remove the selected observation from the centre
For n greater than two, write mean₍ᵢ₎ for the mean of all observations except i. Since their sum is n mean − xᵢ, we have mean₍ᵢ₎ = (n mean − xᵢ)/(n − 1).
Subtract this from xᵢ. The leave-one-out difference is xᵢ − mean₍ᵢ₎ = n(xᵢ − mean)/(n − 1) = nyᵢ/(n − 1). Removing the selected value changes the comparison centre as well as the scale.
In [−1,0,1,10], leaving out ten gives mean zero. The new difference is ten, compared with the original deviation 7.5. The multiplier n/(n − 1) is 4/3, as the identity predicts.
Do not describe a calculation as leave-one-out if it retains the original mean but only changes the variance. That hybrid may be useful under a separately defined method, but it is not the statistic we are now deriving. Each stage should identify which observations contributed to its reference.
CHAPTER 13 OF 29 · Separate internal and external references
13. Derive the leave-one-out variance
Let Q₍ᵢ₎ be the sum of squared deviations of the remaining observations about their own mean. A direct expansion gives Q₍ᵢ₎ = Q − [n/(n − 1)]yᵢ². The leave-one-out sample variance is s₍ᵢ₎² = Q₍ᵢ₎/(n − 2).
Check our example: Q is 77, yᵢ² is 56.25, and multiplying by 4/3 gives 75. Thus Q₍ᵢ₎ = 2. Dividing by two gives leave-one-out variance one, consistent with the remaining values [−1,0,1].
The formula also explains the ceiling. Nonnegative Q₍ᵢ₎ requires Q ≥ nyᵢ²/(n − 1), exactly the inequality used earlier. Equality means the remaining observations have zero spread.
This is a useful independent connection, but it does not remove the zero-denominator problem. If the remaining observations are identical, their sample standard deviation is zero. A procedure that divides by it is undefined unless it explicitly supplies another convention; we should not silently report a finite score or invent a probability.
CHAPTER 14 OF 29 · Separate internal and external references
14. Define a predictive leave-one-out distance
When s₍ᵢ₎ is positive, define Tᵢ = [xᵢ − mean₍ᵢ₎]/[s₍ᵢ₎√(1 + 1/(n − 1))]. The square-root factor recognizes that the difference compares one observation with an estimated mean from the other n − 1 observations under a common-variance independent-sampling model.
The formula is stated here as a particular distance statistic. Assigning it a sampling distribution, p-value or simultaneous decision rule requires a justified model. Merely calculating it does not establish those assumptions.
For [−1,0,1,L], excluding L leaves mean zero and sample standard deviation one. Therefore T for L is L√3/2. At L = 10, it is approximately 8.660, while the same-sample score is only about 1.4804.
As L increases, this leave-one-out distance is unbounded. Its reference does not absorb the selected observation’s size. The two statistics have different sensitivity because they construct their denominators differently, not because one arithmetic routine discovered information absent from the raw record.
CHAPTER 15 OF 29 · Separate internal and external references
15. Relate the two statistics without confusing their names
Substitute the leave-one-out mean and variance identities into Tᵢ². In terms of the same-sample score, the result is Tᵢ² = n(n − 2)zᵢ²/[(n − 1)² − nzᵢ²], provided the leave-one-out scale is positive.
The denominator approaches zero as zᵢ² approaches its ceiling (n − 1)²/n. Thus a small change near that ceiling can correspond to a large change in the leave-one-out distance.
For n = 4 and zᵢ² = 675/308, the formula gives Tᵢ² = 75, matching (10√3/2)². This checks the relation using a separately computed leave-one-out sample.
Terminology varies across software and regression texts. Internally or externally studentized residuals may include factors absent from the plain score (xᵢ − mean)/s. Preserve the actual equation beside the name. Our finite maximum applies to the explicitly defined same-sample score, and transformed residual conventions need their own corresponding bounds.
CHAPTER 16 OF 29 · Work masking and evidence boundaries
16. Explain why several separated values can mask one another
Construct a sample with k identical high values b and n − k identical low values a, where 1 ≤ k < n and b > a. Let Δ = b − a. The mean is a + kΔ/n.
Each high value has deviation (n − k)Δ/n, while each low value has deviation −kΔ/n. Summing their squares gives Q = k(n − k)Δ²/n.
The squared same-sample score for a high value is therefore (n − k)(n − 1)/(nk). The separation Δ cancels. Widening the gap between the two groups does not change their scores in this two-level construction.
This is not a claim that the high group consists of erroneous observations. It shows how several similarly separated values jointly change the reference mean and spread. The numerical mechanism can contribute to masking, while the scientific meaning of either group still depends on the measurement context and population model.
CHAPTER 17 OF 29 · Work masking and evidence boundaries
17. Work the group-size change numerically
Take n = 8. With one high value and seven identical low values, the high score reaches 7/√8, approximately 2.475. With two identical high values and six low values, each high squared score becomes 21/8, giving a score about 1.620.
For an original example, let the low values be four and the two high values be twenty-four. The mean is nine. Each high deviation is fifteen; each low deviation is minus five. Q = 2 × 225 + 6 × 25 = 600, so s² = 600/7.
Each high score is 15/√(600/7), whose square is 225 × 7/600 = 21/8. Replacing twenty-four by two hundred and four leaves the group scores unchanged because the same two-level geometry remains.
If four of the eight observations occupy each level, every absolute score is √(7/8), about 0.935. Large between-group separation is not automatically a large same-sample standardized distance when both groups substantially determine the scale.
CHAPTER 18 OF 29 · Work masking and evidence boundaries
18. Check how many large scores are even possible
The same-sample scores satisfy Σzᵢ = 0 and Σzᵢ² = n − 1. These follow immediately from centring by the mean and dividing by the sample standard deviation.
If k scores strictly exceed a positive threshold c in absolute value, their squared contributions exceed kc². Therefore kc² < n − 1. This is a necessary count constraint, not a complete characterization of all possible score patterns.
For n = 9 and c = 2, two strict exceedances would contribute more than eight, but the entire squared-score budget is eight. At most one score can strictly exceed two in absolute value.
The zero-sum condition can impose further restrictions, so the squared-budget argument should not be advertised as always sharp. Its use is diagnostic: an output listing many impossible same-sample scores may reveal a different denominator, separate group references, excluded observations or a calculation error. Inspect the construction before assuming the data violate mathematics.
CHAPTER 19 OF 29 · Work masking and evidence boundaries
19. Use invariance as an implementation check
Add the same constant to every observation. Each value and the mean shift equally, while deviations and sample spread remain unchanged. All same-sample z-scores should therefore remain identical.
Multiply all values by a positive constant. Deviations and spread scale by the same factor, so the scores again remain unchanged. Multiplication by a negative constant reverses every score’s sign but preserves its magnitude.
These properties explain why changing the physical unit consistently cannot defeat the finite ceiling. Converting a whole dataset from one unit to another does not create more standardized separation. Changing only one record’s unit without converting the others changes the data and may create an error.
For a concrete check, transform [7,7,7,31] to [19,19,19,67] using twice each value plus five. The score list remains [−0.5,−0.5,−0.5,1.5]. This gives an independent test of centring, scaling and denominator consistency in a software implementation.
CHAPTER 20 OF 29 · Work masking and evidence boundaries
20. Keep normal tails separate from finite-sample geometry
A same-sample standardized value does not become an independent standard normal variable merely because it is labelled z. Its finite support already shows a difference: with n = 4 it cannot exceed 1.5 in magnitude, while a standard normal model assigns nonzero probability beyond that point.
The scores within one sample are also linked by their zero sum and fixed squared sum. They cannot be treated as independent draws from a distribution that ignores those constraints.
This does not make normal models useless. A suitable normal model can support formal procedures with the correct statistic and reference distribution. The point is to derive or verify that procedure rather than borrowing a normal tail area from an individual fixed-reference formula.
NIST’s outlier-detection guidance explicitly notes the maximum same-sample z-score and distinguishes potential-outlier investigation from formal identification. Our proofs and numerical constructions explain the arithmetic mechanism; they do not supply a universal deletion rule.
CHAPTER 21 OF 29 · Work masking and evidence boundaries
21. Handle zero spread and uncertain records honestly
If every observation is identical, Q and s are zero. The original same-sample score is undefined, including for an observation equal to that common value. Do not assign zero merely because its numerator is also zero; the formula is a division by zero.
If a leave-one-out sample has zero spread, its distance statistic is likewise undefined. This happens in the exact ceiling example when the selected value is removed. The unbounded limit in a nearby varying-reference example should not be confused with a defined infinite score at the zero-spread endpoint.
A conspicuous raw observation may warrant checking the source record, instrument, unit, timestamp or population membership even when a convenient same-sample rule cannot flag it. The right investigation follows the data-generating process.
Conversely, a large external-reference or leave-one-out distance does not prove a mistake. Genuine rare values, mixtures and model mismatch remain possible. Preserve the original record and document any correction or exclusion rather than allowing a standardized number to make that decision silently.
CHAPTER 22 OF 29 · Transfer and independently verify
22. Try a changed sample and denominator
Use [3,3,3,3,18]. The mean is six. The deviations are four copies of minus three and one value twelve. Their squared sum is 180, so sample variance is 45 and sample standard deviation is 3√5.
The selected score is 12/(3√5) = 4/√5, approximately 1.789. It reaches the n = 5 ceiling because all the other values are identical. Under the Q/n convention, the standard deviation is six and the selected score is exactly two.
Now ask whether a strict threshold above two can flag that value. It cannot under either convention in this sample. A non-strict threshold of at least two can flag it under Q/n but not under the sample-standard-deviation convention.
Finally compare eighteen with an explicitly supplied external mean three and standard deviation two. Its distance is 7.5. Explain why all three calculations can be correct and why none determines whether eighteen is an error without a justified reference and domain evidence.
CHAPTER 23 OF 29 · Transfer and independently verify
23. Build a repair, stabilisation and extension route
For repair, work the four-value example by hand and retain raw units beside deviations. If the learner treats standard deviation as a fixed ruler while changing one observation, recalculate the ruler explicitly.
For stabilisation, prove the zero-sum relationship and the balancing squared-sum inequality. Ask when equality holds. Then change the variance denominator and require the learner to state the new ceiling rather than reuse the old one.
For extension, derive the leave-one-out variance identity and the two-level group formula. Mix questions about feasibility, count constraints, fixed references and undefined cases. The learner should choose the relevant construction before calculating.
A useful evidence record includes the full raw sample, inclusion rule, n, mean, Q, variance denominator, score equation and threshold comparison. These details let another person reproduce a surprising result. A displayed score without them can conceal the very distinction that explains why an extreme raw value was not flagged.
CHAPTER 24 OF 29 · Transfer and independently verify
24. Add reference observations without changing the selected value
Return to the raw value 31 and a baseline value seven. With three baseline observations, the four-value sample gives the selected score 1.5. Add six more sevens. The ten-value sample now gives 9/√10, approximately 2.846, although 31 itself has not changed.
Add one further seven, giving ten sevens and one 31. The eleven-value sample has mean 101/11 and sample standard deviation 24/√11. The selected deviation is 240/11, so its score is 10/√11, approximately 3.015. A strict threshold above three now triggers.
The selected measurement, the raw gap and the arithmetic definition stayed fixed. The reference collection changed. Reporting the new flag as evidence that the measurement suddenly became more extreme in raw units would therefore miss the cause.
This is a constructed sensitivity exercise, not a recommendation to duplicate baseline rows until a flag appears. Repeated copies of one observation do not create new independent evidence. A real growing reference archive should record actual distinct observations, sampling conditions and dependencies. Its numerical score can change with sample size while the credibility of an inference still depends on the data collection.
CHAPTER 25 OF 29 · Transfer and independently verify
25. Recover the bound from geometric constraints
View the n scores as a vector z. It lies in the hyperplane whose coordinates sum to zero, and its squared Euclidean length is n − 1. We can derive the coordinate ceiling directly from those two properties.
Let aᵢ be the coordinate unit vector, with one in position i and zero elsewhere. Subtract the constant vector having every entry 1/n to obtain pᵢ. This vector lies in the zero-sum hyperplane and has squared length (n − 1)/n.
Because z has zero coordinate sum, its inner product with pᵢ equals zᵢ. The Cauchy–Schwarz inequality gives |zᵢ| ≤ length(z) × length(pᵢ) = √(n − 1)√[(n − 1)/n]. This is the same ceiling (n − 1)/√n.
Equality requires z to be proportional to pᵢ. Its selected coordinate differs from n − 1 equal remaining coordinates, reproducing the earlier equality construction. This second proof checks the result from a different representation and explains why the normalized scores cannot vary independently. It introduces no distributional assumption and does not turn the geometry into a probability model.
CHAPTER 26 OF 29 · Transfer and independently verify
26. Identify what standardized scores cannot reconstruct
Suppose a nonzero score vector has coordinate sum zero and squared sum n − 1. Choose any real centre μ and any positive scale c, and construct observations xᵢ = μ + czᵢ. Their sample mean is μ and their sample standard deviation is c, so their scores are exactly the original zᵢ.
There are therefore infinitely many raw datasets with the same score vector. Standardization removes the location and positive scale unless those parameters are retained separately.
For example, [7,7,7,31] and [100,100,100,102] both produce [−0.5,−0.5,−0.5,1.5]. The selected gaps from the repeated baseline are twenty-four and two raw units. Equal same-sample scores do not establish equal physical significance.
This information loss is deliberate when a relative-scale comparison answers the question. It becomes a problem if a downstream reader needs an absolute tolerance, quantity or measurement limit. Preserve raw values and units beside the standardized record. A perfectly calculated score cannot recover a meaningful physical threshold from parameters that were discarded.
CHAPTER 27 OF 29 · Transfer and independently verify
27. Check arithmetic before interpreting a boundary breach
For exact real arithmetic, no valid nonconstant sample can exceed the proved ceiling. A software output that does so should prompt a reconstruction of the formula, reference and arithmetic, rather than a conclusion that the theorem has an exceptional dataset.
First check whether the selected observation was included in both mean and spread. Then inspect the variance denominator, missing-value policy, weights and grouping. A score based on a different reference is not constrained by this particular bound.
Next compare Q computed from centred deviations with the algebraically equivalent expression Σxⱼ² − n mean². The second expression subtracts potentially large quantities to obtain the remaining spread; finite-precision arithmetic can make such a subtraction sensitive. A negative computed squared spread cannot be a true sum of real squares.
Keep sufficient precision and use a numerically appropriate implementation. A tiny displayed overshoot near equality can reflect rounding, while a large discrepancy may reveal a different statistic or a substantive calculation problem. Do not automatically clip every output to the ceiling: that could hide the cause. Investigate, correct the method and retain the original evidence of what failed.
CHAPTER 28 OF 29 · Transfer and independently verify
28. Questions about ceilings and outlier claims
Can I make the same-sample score larger by changing units?
A consistent positive affine unit conversion leaves the scores unchanged. A larger raw scale changes the sample spread in proportion. Changing one observation alone changes the dataset and may change the score, but it cannot exceed the finite bound under the same convention.
Does no score above three mean the sample is clean?
No. With ten or fewer observations, that strict rule cannot trigger at all under the specified same-sample formula. Recording errors, meaningful rare events and unsuitable population mixtures require investigation beyond that output.
Is leave-one-out always the best replacement?
No. It answers a different reference question and can be unstable or undefined when the remaining sample has little or no spread. Formal inference, multiple comparisons and dependent observations need a suitable model and procedure.
Does the bound assume a bell-shaped distribution?
No. It follows from the sample mean, positive sample spread and finite real observations. Distributional assumptions enter when translating a statistic into a probability or a decision rule, not in this algebraic ceiling.
CHAPTER 29 OF 29 · Transfer and independently verify
29. Follow the reference and evidence routes
Return to Z-Score for reference-population choice, modified scores, normality and the distinction between an observation score and a test of a mean. Its general warning about masking is developed here through a sharp bound and explicit group-size calculations.
The primary technical reference is the NIST Engineering Statistics Handbook: Detection of Outliers, checked on 25 October 2026. It supplies the established maximum-score warning and the distinction between labeling, accommodation and formal identification. All example samples, algebraic working, denominator comparisons and transfer exercises here are independently constructed teaching material.
Connect the reasoning with Learning That Transfers into Life and the How X Works hub. The next useful step is to reconstruct a score from its raw reference, test whether its rule is mathematically capable of triggering, and keep that feasibility result separate from the evidence needed to judge an observation.