Chapter map

1. 1. Inputs
2. 2. Statistic
3. 3. Single variance
4. 4. Mean variance
5. 5. Standard deviation
6. 6. Independent copy
7. 7. Replaced statistic
8. 8. Difference
9. 9. Squared input change
10. 10. Squared output change
11. 11. Sum
12. 12. Half factor
13. 13. Equality check
14. 14. McDiarmid comparison
15. 15. Nonlinear transfer
16. 16. Claim

Visit the How Mathematics Works hub for the wider series and its connections to education, reasoning, science, computing, design and careers.

---

1. Inputs

This is where modelling choices begin to affect evidence. Use 100 independent Bernoulli(.2) variables.

A concrete situation

Each X_i is zero or one.

What the mathematics does not license

Independence is essential. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record model. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

2. Statistic

The central mathematical idea is concrete. Let Z be their sample mean.

A concrete situation

Z is a function of all coordinates.

What the mathematics does not license

The method also covers nonlinear functions. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record Z. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

3. Single variance

Here the geometry becomes operational. Compute .2(.8).

A concrete situation

Var(X_i)=.16.

What the mathematics does not license

Use the population probability. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record variance. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

4. Mean variance

This step deserves a careful pause. Divide .16 by 100.

A concrete situation

Var(Z)=.0016.

What the mathematics does not license

This exact calculation is a benchmark. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record target. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

5. Standard deviation

The useful question is not only how to calculate, but what the calculation assumes. Take the square root.

A concrete situation

SD(Z)=.04.

What the mathematics does not license

Variance and standard deviation have different units. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record SD. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

6. Independent copy

A small example makes the mechanism visible. Create X_i prime independently with the same law.

A concrete situation

Only coordinate i is replaced.

What the mathematics does not license

Do not reuse the original value. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record copy. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

7. Replaced statistic

This is where modelling choices begin to affect evidence. Define Z_i prime with X_i prime in position i.

A concrete situation

Other coordinates stay fixed.

What the mathematics does not license

Replacement is coordinate-specific. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record statistic. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

8. Difference

The central mathematical idea is concrete. Subtract the two means.

A concrete situation

Z-Z_i prime=(X_i-X_i prime)/100.

What the mathematics does not license

Keep the denominator inside the square. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record difference. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

9. Squared input change

Here the geometry becomes operational. Use E[(X_i-X_i prime)^2]=2Var(X_i).

A concrete situation

Value is .32.

What the mathematics does not license

This identity uses iid copies. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record expectation. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

10. Squared output change

This step deserves a careful pause. Divide .32 by 10000.

A concrete situation

Value is .000032.

What the mathematics does not license

Square one hundred. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record coordinate term. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

11. Sum

The useful question is not only how to calculate, but what the calculation assumes. Add one hundred identical terms.

A concrete situation

Total is .0032.

What the mathematics does not license

For unequal coordinates, sum separately. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record sum. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

12. Half factor

A small example makes the mechanism visible. Multiply by one half.

A concrete situation

Efron–Stein bound is .0016.

What the mathematics does not license

Use the standard replace-one form. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record bound. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

13. Equality check

This is where modelling choices begin to affect evidence. Compare with exact Var(Z).

A concrete situation

Both equal .0016 here.

What the mathematics does not license

Equality is not universal. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record comparison. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

14. McDiarmid comparison

The central mathematical idea is concrete. Compute worst-case replacement effect .01.

A concrete situation

That route targets tails rather than variance.

What the mathematics does not license

Expected squared and worst-case changes differ. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Explain distinction. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

15. Nonlinear transfer

Here the geometry becomes operational. Replace the mean with a stable statistic.

A concrete situation

The same coordinate method applies.

What the mathematics does not license

The bound may be loose or hard to evaluate. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Test statistic. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

16. Claim

This step deserves a careful pause. State independence, replacement construction and half-sum formula.

A concrete situation

The teaching variance bound is .0016.

What the mathematics does not license

Do not infer a tail probability without another step. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Bound conclusion. Then explain the result to someone who has not seen the formula. If the explanation needs a hidden assumption, add that assumption to the model record.

---

A seven-day practice lab

1. Day 1: define the decision, observation unit, X variables and Y or reference condition.
2. Day 2: reproduce the smallest worked example and check every dimension, unit and sign.
3. Day 3: build a transparent baseline and write what it assumes.
4. Day 4: fit the main method inside a training-only pipeline; record every parameter.
5. Day 5: challenge one assumption with a deliberately difficult dataset or operating regime.
6. Day 6: compare validation metrics with the practical workload created by errors.
7. Day 7: present one supported conclusion, one unresolved uncertainty and one safe next experiment.

Questions for a student, parent or mentor

• Which part of the model is a mathematical definition, and which part is a choice?
• What small calculation would catch a sign, scale or dimension error?
• What evidence comes from training, validation and the untouched test?
• What benign condition could look like a fault or prediction failure?
• How would the answer change for another response, asset, school project or time period?
• What should a human review before acting on the score?

A method-specific audit before deployment

The Efron–Stein inequality bounds the variance of a function of independent inputs by measuring how much the output changes when one coordinate is replaced by an independent copy. It converts local replace-one sensitivity into a global variance statement.

Take the mean of one hundred independent Bernoulli variables with success probability .2. The variance of one variable is .2(.8)=.16, so the exact variance of the sample mean is .16/100=.0016 and its standard deviation is .04.

For coordinate i, replace X_i by an independent copy X_i prime. The mean changes by (X_i-X_i prime)/100. Because E[(X_i-X_i prime) squared]=2 Var(X_i)=.32, the expected squared replace-one change is .32/10000=.000032.

Summing over one hundred coordinates gives .0032. Multiplying by one half in the standard Efron–Stein form gives .0016, exactly matching the variance of the mean in this example. Equality is special here; for nonlinear statistics the inequality can be loose.

This result differs from McDiarmid’s inequality. Efron–Stein controls variance through expected squared replacement effects, while McDiarmid controls tails through worst-case bounded differences. Students should record which quantity they need, how replacement samples are defined and whether the original inputs are independent before choosing a theorem.

---

Frequently asked questions

What is the Efron–Stein inequality?

A variance bound based on replacing one independent input at a time with an independent copy.

Why use an independent copy?

It measures coordinate influence while preserving the input distribution.

What does the half factor do?

It corrects the double contribution in the expected squared difference between iid copies.

Does it give a tail probability?

Not directly; it bounds variance, which can feed another inequality if appropriate.

How does it differ from McDiarmid?

Efron–Stein uses expected squared changes for variance; McDiarmid uses worst-case changes for tails.

What must be reported?

Input independence, statistic, copy construction, replacement effect and exact inequality form.

---

Primary source and fact-check note

Bradley Efron and Charles Stein, “The Jackknife Estimate of Variance,” The Annals of Statistics 9(3), 1981. Definitions and core claims in this article were checked against this primary publication. Calculations shown here were independently recomputed. Examples are educational illustrations; they are not reported industrial performance results.

Useful next reading

• McDiarmid’s inequality
• Algorithmic stability
• Empirical Bernstein bounds
• How Mathematics Works
• eduKateSingapore learning routes

Keep the options open. Mathematics education supports clearer questions, transferable problem-solving skills and informed choices across science, engineering, computing, AI, statistics, design and many other careers. It does not guarantee intelligence, income, admission or employment; those outcomes depend on far more than one subject or model.

Extended laboratory: variance, convexity, conditioning and replacement

Copy audit

Construct an iid copy independent of the full original sample.

Coordinate audit

Replace exactly one coordinate at a time.

Scale audit

Square all denominators in the output difference.

Target audit

Call the result a variance bound, not automatically a tail bound.

Mechanism workshop: calculate, challenge and transfer

Manual reconstruction

Rebuild the smallest example for replace-one variance control. Label every random variable, unit and theorem convention.

Assumption map

Place each assumption beside the exact algebraic step it permits.

Counterexample

Change one condition and identify the first invalid inference.

Sensitivity

Vary the threshold, cap, variance or coordinate effect while holding the comparison protocol fixed.

Numerical verification

Recompute logs, squares, exponents and bounds with an independent script before rounding.

Student transfer

Apply the mechanism to a small reliability, learning, network or sequential experiment.

Parent explanation

Explain why a conservative bound supports a decision without promising an outcome.

Evidence ladder

Separate definition, identity, theorem, simulation and practical interpretation.

Limit statement

Write one supported conclusion, one overclaim and one repaired conclusion.

Next measurement

Choose an exact calculation, simulation or held-out check and explain the new evidence it adds.

Evidence notebook: reconstruct every mathematical step

Evidence notebook 1: Inputs

Step 1 isolates inputs. Use 100 independent Bernoulli(.2) variables. The reproducible result is: Each X_i is zero or one. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Independence is essential. Record model. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 2: Statistic

Step 2 isolates statistic. Let Z be their sample mean. The reproducible result is: Z is a function of all coordinates. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: The method also covers nonlinear functions. Record Z. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 3: Single variance

Step 3 isolates single variance. Compute .2(.8). The reproducible result is: Var(X_i)=.16. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Use the population probability. Record variance. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 4: Mean variance

Step 4 isolates mean variance. Divide .16 by 100. The reproducible result is: Var(Z)=.0016. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: This exact calculation is a benchmark. Record target. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 5: Standard deviation

Step 5 isolates standard deviation. Take the square root. The reproducible result is: SD(Z)=.04. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Variance and standard deviation have different units. Record SD. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 6: Independent copy

Step 6 isolates independent copy. Create X_i prime independently with the same law. The reproducible result is: Only coordinate i is replaced. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Do not reuse the original value. Record copy. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 7: Replaced statistic

Step 7 isolates replaced statistic. Define Z_i prime with X_i prime in position i. The reproducible result is: Other coordinates stay fixed. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Replacement is coordinate-specific. Record statistic. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 8: Difference

Step 8 isolates difference. Subtract the two means. The reproducible result is: Z-Z_i prime=(X_i-X_i prime)/100. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Keep the denominator inside the square. Record difference. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 9: Squared input change

Step 9 isolates squared input change. Use E[(X_i-X_i prime)^2]=2Var(X_i). The reproducible result is: Value is .32. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: This identity uses iid copies. Record expectation. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 10: Squared output change

Step 10 isolates squared output change. Divide .32 by 10000. The reproducible result is: Value is .000032. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Square one hundred. Record coordinate term. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 11: Sum

Step 11 isolates sum. Add one hundred identical terms. The reproducible result is: Total is .0032. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: For unequal coordinates, sum separately. Record sum. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 12: Half factor

Step 12 isolates half factor. Multiply by one half. The reproducible result is: Efron–Stein bound is .0016. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Use the standard replace-one form. Record bound. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 13: Equality check

Step 13 isolates equality check. Compare with exact Var(Z). The reproducible result is: Both equal .0016 here. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Equality is not universal. Record comparison. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 14: McDiarmid comparison

Step 14 isolates mcdiarmid comparison. Compute worst-case replacement effect .01. The reproducible result is: That route targets tails rather than variance. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Expected squared and worst-case changes differ. Explain distinction. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 15: Nonlinear transfer

Step 15 isolates nonlinear transfer. Replace the mean with a stable statistic. The reproducible result is: The same coordinate method applies. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: The bound may be loose or hard to evaluate. Test statistic. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Evidence notebook 16: Claim

Step 16 isolates claim. State independence, replacement construction and half-sum formula. The reproducible result is: The teaching variance bound is .0016. This is evidence about expected coordinate replacement. Preserve the distribution, scale, probability and theorem convention beside every value. The safeguard is: Do not infer a tail probability without another step. Bound conclusion. Change one input while keeping every unrelated choice fixed. Recalculate each variance, replacement effect, correction, exponent or bound that should move. Explain whether the difference came from independence, conditioning, scale, approximation or rounding. Ask another reader to reproduce the arithmetic from the recorded inputs. Finish with the narrow conclusion and the evidence still missing.

Failure cases and responsible recovery

Reused copy

The replacement must be an independent copy. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Multiple changes

Replace one coordinate while fixing all others. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Missing half

The standard replace-one formula includes one half. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Unsquared denominator

Output differences must be squared completely. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Dependence omission

The input coordinates are assumed independent. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Tail conflation

A variance bound is not itself a tail probability. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Equality assumption

The inequality is not always tight. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.

Stability conflation

Expected squared influence differs from worst-case algorithmic stability. Build a counterexample, redo the calculation under an eligible theorem and state the missing evidence.