WHY MATHEMATICS?

Why Mathematics? | Hyvärinen Score for Unnormalised Models: Log-Density Gradients, Laplacians, Normalising Constants and Boundary Conditions

How derivatives remove an unknown partition function, with a standard-normal calculation, assumptions, numerical checks and learning steps.

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

Differentiate log density with respect to the observation
Combine the Laplacian and squared gradient under a stated convention
See why multiplicative normalising constants disappear
Check smoothness, boundaries and computation before interpreting results

Chapter map

1. Model shape
2. Log density
3. Gradient
4. Laplacian
5. Convention
6. Observation
7. Squared gradient
8. Half term
9. Point score
10. Scale variant
11. Constant cancellation
12. Boundary term
13. Smoothness
14. Computation
15. Companion evidence
16. Claim

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

1. Model shape

A small example makes the mechanism visible. Write q(y)=c exp(-y squared/2).

A concrete situation

Leave c unknown but positive.

What the mathematics does not license

The example is a shape on the real line. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

State domain. 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. Log density

This is where modelling choices begin to affect evidence. Take log q=log c-y squared/2.

A concrete situation

Separate constant and variable terms.

What the mathematics does not license

A zero or negative q is invalid. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Check positivity. 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. Gradient

The central mathematical idea is concrete. Differentiate once.

A concrete situation

The derivative is -y.

What the mathematics does not license

Differentiate with respect to observation, not parameter. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Label variable. 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. Laplacian

Here the geometry becomes operational. Differentiate again.

A concrete situation

In one dimension it is -1.

What the mathematics does not license

In several dimensions sum coordinate second derivatives. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

State dimension. 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. Convention

This step deserves a careful pause. Use Delta log q plus half gradient norm squared.

A concrete situation

Lower is better.

What the mathematics does not license

Some sources multiply the score by two. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

State scaling. 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. Observation

The useful question is not only how to calculate, but what the calculation assumes. Set y=.5.

A concrete situation

Gradient is -.5.

What the mathematics does not license

One observation is illustrative. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Label case. 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. Squared gradient

A small example makes the mechanism visible. Square -.5.

A concrete situation

Result is .25.

What the mathematics does not license

The norm removes direction but not magnitude. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Show arithmetic. 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. Half term

This is where modelling choices begin to affect evidence. Take half of .25.

A concrete situation

Result is .125.

What the mathematics does not license

Do not omit the convention factor. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Audit 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.

9. Point score

The central mathematical idea is concrete. Add -1 and .125.

A concrete situation

Score is -.875.

What the mathematics does not license

A negative value is allowed. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Explain orientation. 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. Scale variant

Here the geometry becomes operational. Multiply by two if using alternate convention.

A concrete situation

Score becomes -1.75.

What the mathematics does not license

Only consistent scaling preserves comparisons. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Name 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.

11. Constant cancellation

This step deserves a careful pause. Differentiate log c.

A concrete situation

It becomes zero.

What the mathematics does not license

Parameters inside c may still matter through optimisation context. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Explain scope. 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. Boundary term

The useful question is not only how to calculate, but what the calculation assumes. Check decay at domain boundary.

A concrete situation

Make integration by parts valid.

What the mathematics does not license

Bounded supports need modified treatment. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

State 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.

13. Smoothness

A small example makes the mechanism visible. Check first and second derivatives exist.

A concrete situation

Use analytic and automatic checks.

What the mathematics does not license

Nonsmooth models break the simple formula. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Test gradients. 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. Computation

This is where modelling choices begin to affect evidence. Compare symbolic, finite-difference and automatic derivatives.

A concrete situation

Use tolerances.

What the mathematics does not license

Numerical noise can dominate second derivatives. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Report method. 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. Companion evidence

The central mathematical idea is concrete. Evaluate held-out fit with another proper score.

A concrete situation

Check samples and calibration.

What the mathematics does not license

Score matching objective alone is not deployment validation. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Add checks. 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

Here the geometry becomes operational. Name domain, convention and boundary assumptions.

A concrete situation

Teaching score is -.875.

What the mathematics does not license

Do not imply every unnormalised model is easy. Evidence becomes stronger when the limitation is tested instead of hidden.

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

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

Many statistical and AI models specify a density only up to an unknown normalising constant. Hyvarinen score matching avoids evaluating that constant by differentiating the log density with respect to the observation. Under one common loss convention, the pointwise score is the Laplacian of log q at y plus one half of the squared norm of its gradient. Multiplying q by a positive constant adds a constant to log q, and derivatives remove it.

For the standard normal shape q(y)=c exp(-y squared/2), log q equals log c minus y squared/2. Its first derivative is -y and its second derivative is -1. The one-dimensional Hyvarinen score is therefore -1+y squared/2. At y=.5 it equals -.875. A convention multiplying the whole score by two gives -1.75; rankings are unchanged when the same positive scaling is used throughout.

The cancellation is powerful but conditional. The density must be differentiable enough, the relevant expectations finite and integration-by-parts boundary terms controlled. The method is naturally suited to continuous variables; discrete or bounded-support settings need modified forms rather than blindly reusing the real-line formula. A low pointwise score may be negative and does not behave like a probability or a distance.

A responsible study states the domain, derivative convention, boundary conditions, automatic-differentiation method, parameterisation, smoothness assumptions and numerical checks. It compares analytic and automatic gradients on simple cases and uses a held-out evaluation or another proper score to assess predictive usefulness. Students can see how calculus removes an intractable partition function while also learning why derivatives, limits and support assumptions are substantive mathematics.

Frequently asked questions

What problem does the Hyvärinen score solve?

It can evaluate smooth density shapes without computing a multiplicative normalising constant.

Why does the constant vanish?

Observation derivatives of its log are zero.

What is the Laplacian?

The sum of second partial derivatives across coordinates.

Can the score be negative?

Yes; orientation and differences matter under a fixed convention.

Does it work on every support?

No. Smoothness and boundary conditions matter, and modified forms may be needed.

What must be reported?

Domain, convention, derivatives, boundary assumptions, numerical method, sample design and companion validation.

Primary source and fact-check note

Aapo Hyvärinen, “Estimation of Non-Normalized Statistical Models by Score Matching,” Journal of Machine Learning Research 6(24), 695–709 (2005). 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: score-specific validation and responsible use

Target audit

Freeze whether the task scores one realised case, estimates an unnormalised density or compares two sample distributions; keep these estimands separate.

Formula audit

Write every logarithm, derivative, Laplacian, expectation, kernel term, self-pair, cross-pair, constant and denominator explicitly.

Computation audit

Reproduce the teaching arithmetic independently and test code on identical distributions, a shifted point, a zero probability and a known normal shape.

Support audit

Check event categories, continuous units, density support, smoothness, boundary conditions, sample dependence and feature scaling.

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.

Tuning audit

Predeclare clipping, derivative convention, kernel family, bandwidth, estimator and preprocessing; do not tune them to obtain a preferred ranking.

Companion audit

Add calibration, marginal or distance-based evidence so one local, derivative or kernel total cannot stand for every forecast property.

Uncertainty audit

Use matched cases, valid permutations or dependence-aware resampling according to the estimand; report finite-sample limits.

Final claim

Describe the exact probability, density-shape or distributional feature supported by the analysis and stop before causal or universal claims.

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: Model shape

Begin with this mechanism: Write q(y)=c exp(-y squared/2). Recreate the concrete case: Leave c unknown but positive. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The example is a shape on the real line. Complete the transfer step: State domain. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 2: Log density

Begin with this mechanism: Take log q=log c-y squared/2. Recreate the concrete case: Separate constant and variable terms. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A zero or negative q is invalid. Complete the transfer step: Check positivity. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 3: Gradient

Begin with this mechanism: Differentiate once. Recreate the concrete case: The derivative is -y. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Differentiate with respect to observation, not parameter. Complete the transfer step: Label variable. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 4: Laplacian

Begin with this mechanism: Differentiate again. Recreate the concrete case: In one dimension it is -1. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: In several dimensions sum coordinate second derivatives. Complete the transfer step: State dimension. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that 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: Convention

Begin with this mechanism: Use Delta log q plus half gradient norm squared. Recreate the concrete case: Lower is better. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Some sources multiply the score by two. Complete the transfer step: State scaling. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 6: Observation

Begin with this mechanism: Set y=.5. Recreate the concrete case: Gradient is -.5. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: One observation is illustrative. Complete the transfer step: Label case. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 7: Squared gradient

Begin with this mechanism: Square -.5. Recreate the concrete case: Result is .25. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: The norm removes direction but not magnitude. Complete the transfer step: Show arithmetic. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 8: Half term

Begin with this mechanism: Take half of .25. Recreate the concrete case: Result is .125. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not omit the convention factor. Complete the transfer step: Audit formula. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that 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: Point score

Begin with this mechanism: Add -1 and .125. Recreate the concrete case: Score is -.875. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: A negative value is allowed. Complete the transfer step: Explain orientation. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 10: Scale variant

Begin with this mechanism: Multiply by two if using alternate convention. Recreate the concrete case: Score becomes -1.75. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Only consistent scaling preserves comparisons. Complete the transfer step: Name convention. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 11: Constant cancellation

Begin with this mechanism: Differentiate log c. Recreate the concrete case: It becomes zero. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Parameters inside c may still matter through optimisation context. Complete the transfer step: Explain scope. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 12: Boundary term

Begin with this mechanism: Check decay at domain boundary. Recreate the concrete case: Make integration by parts valid. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Bounded supports need modified treatment. Complete the transfer step: State condition. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that 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: Smoothness

Begin with this mechanism: Check first and second derivatives exist. Recreate the concrete case: Use analytic and automatic checks. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Nonsmooth models break the simple formula. Complete the transfer step: Test gradients. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 14: Computation

Begin with this mechanism: Compare symbolic, finite-difference and automatic derivatives. Recreate the concrete case: Use tolerances. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Numerical noise can dominate second derivatives. Complete the transfer step: Report method. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 15: Companion evidence

Begin with this mechanism: Evaluate held-out fit with another proper score. Recreate the concrete case: Check samples and calibration. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Score matching objective alone is not deployment validation. Complete the transfer step: Add checks. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.

Workshop 16: Claim

Begin with this mechanism: Name domain, convention and boundary assumptions. Recreate the concrete case: Teaching score is -.875. Write observation units, log-density gradients, second derivatives, Laplacian terms, convention factors, point scores and boundary assumptions with units, signs, identifiers and assumptions. Calculate the smallest non-trivial example by hand and reproduce it independently. Challenge this limitation: Do not imply every unnormalised model is easy. Complete the transfer step: Bound conclusion. Change exactly one probability, support point, log base, density unit, clipping floor, observation, derivative, boundary condition, kernel, bandwidth, member, sample size, self-pair rule, feature scale, permutation scheme or dependence assumption. Explain which mathematical object changes, which score or discrepancy term moves, which ranking survives and which claim must weaken. Keep target definition, calculation, aggregation, uncertainty and interpretation on separate notebook lines. Explain the result once to a student with full arithmetic and once to a parent with the real decision and original units. Add a counterexample involving zero support, an unreported unit change, a nonsmooth density, a failed boundary condition, a noncharacteristic kernel, bandwidth collapse, duplicate samples, invalid permutations or a statistically detectable but practically tiny difference. State the extra evidence that would separate the explanations. Finish with one supported sentence, one unsupported sentence and one useful next measurement.