Chapter map

1. 1. Function
2. 2. Spacing
3. 3. Values
4. 4. First stencil
5. 5. First error
6. 6. Second numerator
7. 7. Second stencil
8. 8. Halve h
9. 9. Boundary
10. 10. Noise
11. 11. Stability
12. 12. Grid study
13. 13. Units
14. 14. Conclusion

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

---

1. Function

This is where modelling choices begin to affect evidence. Use f(x)=x^3 at x=1.

A concrete situation

Exact f prime=3 and f double prime=6.

What the mathematics does not license

Keep units. 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.

---

2. Spacing

The central mathematical idea is concrete. Set h=0.1.

A concrete situation

Neighbours are .9 and 1.1.

What the mathematics does not license

Uniform grid. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record 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. Values

Here the geometry becomes operational. Evaluate f.

A concrete situation

f(1.1)=1.331,f(.9)=.729.

What the mathematics does not license

Keep decimals exact here. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record values. 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. First stencil

This step deserves a careful pause. Subtract and divide by .2.

A concrete situation

Estimate=3.01.

What the mathematics does not license

Centred formula. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record result. 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. First error

The useful question is not only how to calculate, but what the calculation assumes. Compare with 3.

A concrete situation

Absolute error=.01.

What the mathematics does not license

Do not call exact. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record error. 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. Second numerator

A small example makes the mechanism visible. Compute 1.331-2+.729.

A concrete situation

Numerator=.06.

What the mathematics does not license

Check signs. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record numerator. 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. Second stencil

This is where modelling choices begin to affect evidence. Divide by .01.

A concrete situation

Estimate=6.

What the mathematics does not license

Exact for this cubic. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record result. 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. Halve h

The central mathematical idea is concrete. Repeat with .05.

A concrete situation

First-derivative error quarters.

What the mathematics does not license

Expected second order. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record ratio. 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. Boundary

Here the geometry becomes operational. Try x=0 without left data.

A concrete situation

Centred stencil unavailable.

What the mathematics does not license

Choose one-sided rule. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record 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.

---

10. Noise

This step deserves a careful pause. Perturb values slightly.

A concrete situation

Difference can amplify noise.

What the mathematics does not license

Report data quality. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record sensitivity. 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. Stability

The useful question is not only how to calculate, but what the calculation assumes. Embed in a time equation.

A concrete situation

Step restrictions appear.

What the mathematics does not license

Check eligible CFL. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Record 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.

---

12. Grid study

A small example makes the mechanism visible. Use h,h/2,h/4.

A concrete situation

Observed order can be estimated.

What the mathematics does not license

Stay in asymptotic range. The same reasoning helps students move from an exercise to a real dataset.

Try it, check it, transfer it

Record errors. 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. Units

This is where modelling choices begin to affect evidence. Track x units.

A concrete situation

First derivative divides by x, second by x squared.

What the mathematics does not license

Check dimensions. That distinction matters because a correct formula can still answer the wrong operational question.

Try it, check it, transfer it

Record units. 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. Conclusion

The central mathematical idea is concrete. State grid,stencil,boundaries,stability and convergence.

A concrete situation

Worked first error is .01.

What the mathematics does not license

Keep polynomial scope. The habit of checking assumptions is one of the practical benefits of learning mathematics.

Try it, check it, transfer it

Write 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 finite difference method replaces derivatives by weighted differences of function values on a grid. Taylor expansion explains the weights, the formal order of accuracy and the terms that are neglected.

For f(x)=x cubed at x=1 with h=0.1, the centred first derivative is [f(1.1)-f(0.9)]/(0.2). The values are 1.331 and 0.729, so the estimate is 3.01.

The exact derivative is 3. The absolute error 0.01 equals h squared for this cubic example, making the second-order truncation behaviour visible without a long calculation.

The centred second derivative [f(1.1)-2f(1)+f(0.9)]/h squared equals 6 exactly here. Exactness for this polynomial and stencil does not mean every smooth function is recovered exactly.

Forward, backward and centred differences use different information. Boundaries may force one-sided formulas, while interior points can often use symmetric stencils.

Fornberg’s algorithm generates weights for arbitrary derivative orders and irregularly spaced nodes. A formula valid on a uniform grid cannot simply be copied onto a stretched grid.

Did You Know? Making h smaller reduces truncation error only until rounding, noisy data or subtraction cancellation becomes important. There is often an accuracy balance rather than a smallest-is-best rule.

For a differential equation, consistency of each stencil is not enough. Stability and boundary treatment determine whether local errors remain controlled as the computation advances.

The Courant-Friedrichs-Lewy condition links time step, grid size and propagation speed for many explicit schemes. It is a method-dependent stability restriction rather than a law that every discretisation shares.

Grid resolution must capture the physical length and time scales of interest. A stable answer on a coarse grid can still be a poor approximation of the continuous problem.

Students can halve h repeatedly, compute error ratios and estimate observed order. This convergence study checks the implemented formula more honestly than one apparently plausible output.

Non-smooth functions can lose the expected high order because Taylor assumptions fail near corners or discontinuities. Adaptive meshes or shock-capturing ideas address different difficulties.

Engineering uses include heat transfer, waves, finance, image processing and fluid models. Numerical agreement does not validate material properties, loads or boundary conditions.

Units matter: a derivative divides by length or time, while a second derivative divides by its square. A missing scale factor can produce a numerically neat but dimensionally wrong answer.

For parents and students, the transferable lesson is that change can be estimated from nearby observations when the spacing, symmetry and error are made explicit.

A responsible report states the equation, grid, stencil, order, boundary treatment, time integrator, stability restriction, precision, convergence study and comparison with known or measured behaviour.

---

Frequently asked questions

What is a finite difference?

A weighted combination of sampled values that approximates a derivative.

Why use a centred stencil?

Symmetry often cancels lower-order Taylor terms.

Does smaller h always improve accuracy?

No. Rounding, cancellation and noise can eventually dominate.

What else matters for PDEs?

Stability, boundary treatment, time integration and convergence.

Can irregular grids be used?

Yes, with weights derived for their actual node positions.

What must be reported?

Equation, grid, stencil, order, boundaries, stability, precision, convergence and validation.

---

Primary source and fact-check note

Bengt Fornberg, “Generation of Finite Difference Formulas on Arbitrarily Spaced Grids,” Mathematics of Computation 51, 1988. 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

• GMRES
• Conjugate gradient method
• LSQR
• How Mathematics Works
• Study and learning methods

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.

Method audit: equations, discretisation and evidence

Equation audit

State the continuous equation and eligible discretisation.

Grid audit

Record nodes, cells, elements or modes with units.

Error audit

Separate truncation, iteration, rounding and modelling error.

Validation audit

Check convergence and independent physical evidence.

Mechanism workshop: calculate, challenge and transfer

Manual reconstruction

Build the smallest grid, mesh, cell balance or Fourier sample and label every unit.

Taylor or variational map

Connect each discrete coefficient to the continuous equation.

Counterexample

Remove one stability, smoothness or conservation condition and locate failure.

Sensitivity

Change spacing, element size, time step or mode count.

Independent verification

Compare with an exact polynomial, manufactured solution or conservation balance.

Precision experiment

Round data and measure cancellation or residual drift.

Convergence study

Refine systematically and calculate observed error rates.

Engineering transfer

Map the mechanism to heat, structures, flow or signals.

Student explanation

Explain the discrete representation in plain language.

Parent explanation

Describe why simulation still needs checking against reality.

Evidence notebook: reconstruct every step

Evidence notebook 1: Function

Step 1 isolates function. Use f(x)=x^3 at x=1. The checked result is: Exact f prime=3 and f double prime=6. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Keep units. Record target. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 2: Spacing

Step 2 isolates spacing. Set h=0.1. The checked result is: Neighbours are .9 and 1.1. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Uniform grid. Record h. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 3: Values

Step 3 isolates values. Evaluate f. The checked result is: f(1.1)=1.331,f(.9)=.729. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Keep decimals exact here. Record values. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 4: First stencil

Step 4 isolates first stencil. Subtract and divide by .2. The checked result is: Estimate=3.01. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Centred formula. Record result. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 5: First error

Step 5 isolates first error. Compare with 3. The checked result is: Absolute error=.01. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Do not call exact. Record error. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 6: Second numerator

Step 6 isolates second numerator. Compute 1.331-2+.729. The checked result is: Numerator=.06. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Check signs. Record numerator. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 7: Second stencil

Step 7 isolates second stencil. Divide by .01. The checked result is: Estimate=6. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Exact for this cubic. Record result. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 8: Halve h

Step 8 isolates halve h. Repeat with .05. The checked result is: First-derivative error quarters. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Expected second order. Record ratio. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 9: Boundary

Step 9 isolates boundary. Try x=0 without left data. The checked result is: Centred stencil unavailable. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Choose one-sided rule. Record formula. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 10: Noise

Step 10 isolates noise. Perturb values slightly. The checked result is: Difference can amplify noise. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Report data quality. Record sensitivity. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 11: Stability

Step 11 isolates stability. Embed in a time equation. The checked result is: Step restrictions appear. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Check eligible CFL. Record condition. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 12: Grid study

Step 12 isolates grid study. Use h,h/2,h/4. The checked result is: Observed order can be estimated. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Stay in asymptotic range. Record errors. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 13: Units

Step 13 isolates units. Track x units. The checked result is: First derivative divides by x, second by x squared. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Check dimensions. Record units. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Evidence notebook 14: Conclusion

Step 14 isolates conclusion. State grid,stencil,boundaries,stability and convergence. The checked result is: Worked first error is .01. Preserve dimensions, units, grids, basis conventions, flux orientation, boundary conditions and precision beside every number. The safeguard is: Keep polynomial scope. Write conclusion. Change one input, recompute the local formula and global check, and explain whether the difference comes from truncation, stability, mesh geometry, conservation, aliasing, rounding or a changed model. Ask a second reader to reproduce the arithmetic. Finish with the narrow conclusion and missing evidence.

Failure cases and responsible recovery

Formula transplant

Weights depend on node positions. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Smaller-is-always-better

Roundoff and noise can dominate. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Local-order proof

Stability is also needed. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Boundary silence

One-sided treatment changes error. Build a small counterexample, restore the missing condition and recalculate before making the claim.

CFL omission

Explicit time steps may be unstable. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Single-grid trust

Run convergence studies. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Units omission

Spacing powers matter. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Model leap

Numerical accuracy is not physical validity. Build a small counterexample, restore the missing condition and recalculate before making the claim.

Verified case study: fourteen steps from equation to decision

1. Define the equation

Write the continuous quantity, domain, units and boundary or initial conditions.

2. Choose representation

Explain why the grid, mesh, control volumes or global basis fits the geometry and regularity.

3. Derive one local rule

Connect a stencil, element integral, face flux or mode coefficient to the governing mathematics.

4. Calculate by hand

Use the smallest nontrivial example and keep exact values where possible.

5. Check invariants

Verify units, symmetry, conservation, boundary values or orthogonality as eligible.

6. Estimate error

Separate discretisation, iteration, roundoff, data and modelling contributions.

7. Refine

Halve spacing, improve elements or increase modes while controlling other parameters.

8. Test a failure

Use a bad time step, distorted element, inconsistent face flux or aliased wave.

9. Compare baseline

Check against an exact, manufactured, direct or experimental reference.

10. Count cost

Record degrees of freedom, memory, matrix or transform operations and elapsed time.

11. Validate

Compare application outputs with independent observations or accepted benchmarks.

12. Communicate

State what the simulation supports and what it cannot establish.

13. Preserve evidence

Save inputs, mesh or nodes, code, precision, parameters and logs.

14. Decide next step

State whether to refine, change method, gather data or seek expert review.

Learner lab: eight checks for transferable understanding

Lab question 1: representation

Draw the grid, elements, control volumes or basis waves before calculating. Mark where each unknown lives and how neighbouring or global pieces communicate. Then explain why the drawing matches the algebra.

Lab question 2: exact test

Choose a polynomial, manufactured solution, uniform state or resolved Fourier mode for which the expected answer is known. Calculate the discrete result and identify exactly which property makes the test pass.

Lab question 3: boundary test

Change one boundary from prescribed value to prescribed flux, periodic closure or another eligible condition. Re-derive the affected row or face balance instead of editing a final number.

Lab question 4: refinement

Create a table of resolution, error, cost and observed order. Keep time-step and solver tolerances consistent with the spatial refinement so one hidden error does not mask another.

Lab question 5: conservation and energy

Identify a quantity the continuous model conserves, dissipates or bounds. Measure the discrete counterpart and explain whether any change is physical, numerical or caused by external boundaries.

Lab question 6: implementation audit

Test indexing, orientation, matrix assembly, transform normalization and units with tiny arrays. A small transparent case catches mistakes that a large colourful output can conceal.

Lab question 7: interpretation

Translate the computed error into the units of temperature, displacement, mass, velocity or signal amplitude. Decide whether the difference is meaningful for the real decision rather than only numerically small.

Lab question 8: communication

Prepare a one-minute student explanation and a technical verification note. Both must name the method, the strongest evidence, the main limitation and the next check without guaranteeing a real-world outcome.

Lab question 9: competing method

Name one alternative discretisation and compare its natural strengths, data needs, boundary handling and error checks. Choose using the problem structure rather than fashion or a single benchmark.

Lab question 10: final reproduction

Give another learner only the equation, parameters and audit checklist. Compare independent results, resolve every discrepancy and preserve the corrected calculation as the reproducible class example.