WHY MATHEMATICS?

Why Mathematics? | Local Polynomial Quantile Regression: Kernel Weights, Bandwidth, Weighted Medians, Boundary Bias and Local Slopes

How nearby observations build a conditional percentile at one predictor value, why bandwidth governs bias and variance, and why sparse regions and curve crossing need care.

Why mathematics is important becomes easier to see when a real decision depends on it. This guide follows the mechanism from definitions and dimensions to a worked calculation, validation, practical learning and responsible use. It keeps factual evidence separate from interpretation and treats every illustrative scenario as an example rather than a promise of results.

Choose a reading route

Choose a target point, bandwidth and kernel
Centre predictors and minimise weighted check loss locally
Read the local level, slope and effective sample size
Validate bandwidth, boundaries, crossings and uncertainty

Chapter map

1. Target point
2. Bandwidth
3. Kernel
4. Responses
5. Local constant
6. Weighted check loss
7. Weighted median
8. Estimate
9. Outlier influence
10. Local linear
11. Boundary bias
12. Effective sample size
13. Bandwidth path
14. Uncertainty
15. Crossing
16. Claim

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

1. Target point

Here the geometry becomes operational. Choose x0 within observed support.

A concrete situation

Use x0=1.

What the mathematics does not license

Extrapolation has no local neighbours. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Mark support. 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. Bandwidth

This step deserves a careful pause. Define the neighbourhood scale h.

A concrete situation

Include three nearby observations.

What the mathematics does not license

Bandwidth controls bias and variance. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Report h. 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. Kernel

The useful question is not only how to calculate, but what the calculation assumes. Assign nonnegative proximity weights.

A concrete situation

Use .25,.50,.25.

What the mathematics does not license

Weights should match distances and kernel rule. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Show formula. 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. Responses

A small example makes the mechanism visible. Use y=1,2,10.

A concrete situation

Ten is locally extreme.

What the mathematics does not license

Do not delete it silently. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Keep ID. 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.

Did You Know? In this chapter, the most important number is useful only together with its construction rule. Changing scaling, sample membership or validation design can change the numerical answer without changing the arithmetic.

5. Local constant

This is where modelling choices begin to affect evidence. Set fitted quantile to b0.

A concrete situation

One parameter summarises the neighbourhood.

What the mathematics does not license

It cannot represent local slope. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Compare linear. 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. Weighted check loss

The central mathematical idea is concrete. Multiply each tau loss by its kernel weight.

A concrete situation

Use tau=.5.

What the mathematics does not license

Forgetting weights changes the estimator. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Audit products. 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. Weighted median

Here the geometry becomes operational. Order y and cumulate weights.

A concrete situation

Cumulative weights are .25,.75,1.

What the mathematics does not license

Ties can create solution intervals. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

State convention. 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. Estimate

This step deserves a careful pause. Choose b0=2.

A concrete situation

Weighted median is two.

What the mathematics does not license

The estimate is local to x0. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Do not globalise. 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. Outlier influence

The useful question is not only how to calculate, but what the calculation assumes. Raise ten farther.

A concrete situation

The weighted median remains two.

What the mathematics does not license

Its scientific meaning may still matter. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Preserve provenance. 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. Local linear

A small example makes the mechanism visible. Add b1(x-x0).

A concrete situation

Estimate level and slope together.

What the mathematics does not license

Poor local x spread makes slope unstable. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Check condition. 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. Boundary bias

This is where modelling choices begin to affect evidence. Compare local constant and local linear near support edge.

A concrete situation

Linear terms can reduce first-order bias.

What the mathematics does not license

Variance may increase. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Report trade-off. 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. Effective sample size

The central mathematical idea is concrete. Summarise weight concentration.

A concrete situation

Three nonzero weights are few.

What the mathematics does not license

Nominal n overstates local information. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Plot n_eff. 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. Bandwidth path

Here the geometry becomes operational. Fit smaller and larger h.

A concrete situation

Watch peaks smooth or disappear.

What the mathematics does not license

After-the-fact choice overfits. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Cross-validate. 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. Uncertainty

This step deserves a careful pause. Bootstrap whole smoothing choice when justified.

A concrete situation

Repeat local fitting at every x0.

What the mathematics does not license

Fixed-h uncertainty omits tuning. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Mirror pipeline. 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. Crossing

The useful question is not only how to calculate, but what the calculation assumes. Compare local fits across tau.

A concrete situation

Sparse regions can reverse order.

What the mathematics does not license

Locality does not guarantee coherence. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Check grid. 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

A small example makes the mechanism visible. State x0,h,kernel,tau and support.

A concrete situation

Teaching local median is two.

What the mathematics does not license

Do not promise tail coverage. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Validate level. 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

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

Local polynomial quantile regression replaces one global curve with many small weighted quantile fits. Around a target x0, it approximates the conditional quantile by b0+b1(x-x0)+… and minimises kernel-weighted check loss. Nearby observations receive more weight, distant observations receive little or none, and b0 estimates the quantile at x0. Higher polynomial terms can estimate curvature or derivatives locally.

At x0=1, suppose three nearby responses are 1,2,10 with kernel weights .25,.50,.25. A local-constant median fit chooses the weighted median two: the cumulative weight reaches .75 there, and half the total weight lies on each side under the usual weighted-median condition. The extreme response ten still contributes, yet it cannot move the weighted median past two because its weight is only .25.

A local-linear fit also estimates a slope from centred predictor values. It often reduces boundary bias compared with local constants, but it can become unstable when the local design has little spread. Bandwidth is therefore central. A very small bandwidth leaves too few effective observations; a very large one erases local structure and approaches a global fit. Kernel shape usually matters less than bandwidth, but both must be reported.

Local quantile curves can cross across tau, become noisy in sparse regions and extrapolate poorly beyond observed support. Density near the target quantile affects uncertainty, while repeated predictors and clustered observations affect the effective information. Validate on held-out data using level-specific check loss and calibration, show effective sample size across x, and keep smoothing choices inside the uncertainty pipeline.

Frequently asked questions

What makes the fit local?

Kernel weights shrink with distance from the target predictor value.

What does bandwidth do?

It controls the size of the neighbourhood and the bias-variance trade-off.

What is a local constant median?

A kernel-weighted median of nearby responses.

Why use a local linear term?

It estimates local slope and can reduce boundary bias.

Can local quantile curves cross?

Yes, especially in sparse regions or with different smoothing choices.

What must be reported?

Target grid, bandwidth, kernel, polynomial order, effective sample size, support and validation.

Primary source and fact-check note

Probal Chaudhuri, “Nonparametric Estimates of Regression Quantiles and Their Local Bahadur Representation,” The Annals of Statistics 19(2), 760–777 (1991), DOI 10.1214/aos/1176348119. 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

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: topic-specific validation and responsible use

Design audit

Freeze sampling units, response coding, boundary rules, predictor support, dependence and missingness.

Mechanism audit

Write the censoring map, kernel weights, check losses, roughness term or rank scores row by row.

Tuning audit

Record thresholds, tau, bandwidth, kernel, polynomial order, lambda, solver starts and tolerances.

Geometry audit

Inspect active uncensored sets, knots, local design spread, score projections and predictor leverage.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Coherence audit

Check boundary frequencies, quantile order, curve crossing, support and effective sample size.

Uncertainty audit

Repeat fitting, tuning and repair in the covariance or bootstrap procedure.

Validation audit

Use untouched data to examine level-specific loss, calibration, boundary behaviour and subgroup relevance.

Final claim

Translate the result into original units without causal, coverage or censoring overclaims.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Mechanism workshop: calculate, challenge and transfer

Workshop 1: Target point

Begin with this mechanism: Choose x0 within observed support. Recreate the concrete case: Use x0=1. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Extrapolation has no local neighbours. Complete the transfer step: Mark support. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Bandwidth

Begin with this mechanism: Define the neighbourhood scale h. Recreate the concrete case: Include three nearby observations. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Bandwidth controls bias and variance. Complete the transfer step: Report h. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Kernel

Begin with this mechanism: Assign nonnegative proximity weights. Recreate the concrete case: Use .25,.50,.25. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Weights should match distances and kernel rule. Complete the transfer step: Show formula. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Responses

Begin with this mechanism: Use y=1,2,10. Recreate the concrete case: Ten is locally extreme. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not delete it silently. Complete the transfer step: Keep ID. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 5: Local constant

Begin with this mechanism: Set fitted quantile to b0. Recreate the concrete case: One parameter summarises the neighbourhood. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: It cannot represent local slope. Complete the transfer step: Compare linear. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Weighted check loss

Begin with this mechanism: Multiply each tau loss by its kernel weight. Recreate the concrete case: Use tau=.5. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Forgetting weights changes the estimator. Complete the transfer step: Audit products. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Weighted median

Begin with this mechanism: Order y and cumulate weights. Recreate the concrete case: Cumulative weights are .25,.75,1. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Ties can create solution intervals. Complete the transfer step: State convention. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Estimate

Begin with this mechanism: Choose b0=2. Recreate the concrete case: Weighted median is two. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The estimate is local to x0. Complete the transfer step: Do not globalise. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 9: Outlier influence

Begin with this mechanism: Raise ten farther. Recreate the concrete case: The weighted median remains two. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Its scientific meaning may still matter. Complete the transfer step: Preserve provenance. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Local linear

Begin with this mechanism: Add b1(x-x0). Recreate the concrete case: Estimate level and slope together. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Poor local x spread makes slope unstable. Complete the transfer step: Check condition. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Boundary bias

Begin with this mechanism: Compare local constant and local linear near support edge. Recreate the concrete case: Linear terms can reduce first-order bias. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Variance may increase. Complete the transfer step: Report trade-off. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Effective sample size

Begin with this mechanism: Summarise weight concentration. Recreate the concrete case: Three nonzero weights are few. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Nominal n overstates local information. Complete the transfer step: Plot n_eff. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Did You Know? A sophisticated adjustment can be algebraically correct and still answer the wrong scientific question when samples, labels, batches or validation boundaries are wrong.

Workshop 13: Bandwidth path

Begin with this mechanism: Fit smaller and larger h. Recreate the concrete case: Watch peaks smooth or disappear. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: After-the-fact choice overfits. Complete the transfer step: Cross-validate. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Uncertainty

Begin with this mechanism: Bootstrap whole smoothing choice when justified. Recreate the concrete case: Repeat local fitting at every x0. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Fixed-h uncertainty omits tuning. Complete the transfer step: Mirror pipeline. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Crossing

Begin with this mechanism: Compare local fits across tau. Recreate the concrete case: Sparse regions can reverse order. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Locality does not guarantee coherence. Complete the transfer step: Check grid. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this mechanism: State x0,h,kernel,tau and support. Recreate the concrete case: Teaching local median is two. Write target predictor values, distances, bandwidths, kernel weights, responses, centered predictors, weighted losses, local coefficients, effective sample sizes and original measurement units with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not promise tail coverage. Complete the transfer step: Validate level. Change exactly one observation, censoring boundary, latent interpretation, active set, quantile level, kernel, bandwidth, polynomial order, roughness definition, lambda, knot, residual sign, dual score, scale covariate, solver start, support boundary, bootstrap rule or validation sample. Explain which mathematical object changes, which estimator is selected, which conclusion survives and which claim must weaken. Keep design, mechanism, tuning, geometry, coherence, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with the complete table and once to a parent with the real decision and original units. Add a counterexample involving structural zeros, truncation, repeated predictors, sparse support, dependence, a real minority group, high leverage or measurement error. State what extra graph or measurement would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.