Chapter map

1. 1. Stencil
2. 2. Constant condition
3. 3. Linear condition
4. 4. Quadratic condition
5. 5. Solve
6. 6. Constant test
7. 7. Linear test
8. 8. Quadratic test
9. 9. Quartic test
10. 10. RBF system
11. 11. Scaling
12. 12. Global row
13. 13. Boundary
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. Stencil

The central mathematical idea is concrete. Use nodes -1,0,1.

A concrete situation

Symmetric local cloud.

What the mathematics does not license

Second derivative at zero. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record nodes. 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. Constant condition

Here the geometry becomes operational. Require sum weights=0.

A concrete situation

Second derivative of one is zero.

What the mathematics does not license

Polynomial augmentation. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record equation. 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. Linear condition

This step deserves a careful pause. Require -w1+w3=0.

A concrete situation

Second derivative of x is zero.

What the mathematics does not license

Symmetry emerges. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record equation. 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. Quadratic condition

The useful question is not only how to calculate, but what the calculation assumes. Require w1+w3=2.

A concrete situation

Second derivative of x squared is two.

What the mathematics does not license

Physical spacing is one. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record equation. 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. Solve

A small example makes the mechanism visible. Combine conditions.

A concrete situation

Weights=[1,-2,1].

What the mathematics does not license

Check all moments. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record vector. 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. Constant test

This is where modelling choices begin to affect evidence. Apply weights to ones.

A concrete situation

Result=0.

What the mathematics does not license

Exact. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record check. 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. Linear test

The central mathematical idea is concrete. Apply to node coordinates.

A concrete situation

Result=0.

What the mathematics does not license

Exact. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record check. 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. Quadratic test

Here the geometry becomes operational. Apply to squares.

A concrete situation

Result=2.

What the mathematics does not license

Exact. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record check. 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. Quartic test

This step deserves a careful pause. Apply to fourth powers.

A concrete situation

Result=2.

What the mathematics does not license

Exact value at zero is zero. Evidence becomes stronger when the limitation is tested instead of hidden.

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.

---

10. RBF system

The useful question is not only how to calculate, but what the calculation assumes. Add kernel equations.

A concrete situation

Local augmented matrix.

What the mathematics does not license

State epsilon. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record conditioning. 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. Scaling

A small example makes the mechanism visible. Normalise stencil radius.

A concrete situation

Weights rescale by length squared.

What the mathematics does not license

Keep units. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record factor. 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. Global row

This is where modelling choices begin to affect evidence. Insert weights into sparse operator.

A concrete situation

One centre per row.

What the mathematics does not license

Check ownership. Evidence becomes stronger when the limitation is tested instead of hidden.

Try it, check it, transfer it

Record pattern. 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. Boundary

The central mathematical idea is concrete. Replace operator row as required.

A concrete situation

Use value or normal derivative.

What the mathematics does not license

State normals. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record residual. 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

Here the geometry becomes operational. State kernel, polynomials, stencil and validation.

A concrete situation

Worked polynomial weights [1,-2,1].

What the mathematics does not license

Three-node teaching stencil. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Write narrow claim. 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 radial basis function finite-difference method computes local differentiation weights from radial basis interpolation on each stencil. It keeps the geometric flexibility of scattered nodes while assembling sparse operators like a finite-difference method.

Use the one-dimensional stencil x=-1,0,1 and approximate the second derivative at zero. Polynomial augmentation through degree two requires weights that differentiate 1, x and x squared exactly.

The constant condition is w1+w2+w3=0. The linear condition is -w1+w3=0. The quadratic condition is w1+w3=2 because the second derivative of x squared is 2.

Solving gives weights [1,-2,1]. Applied to f=x squared, the result is 1+1=2, exactly matching f double prime(0)=2.

Applied to f=x to the fourth power, the same stencil gives 2 while the exact second derivative at zero is 0. The mismatch displays local truncation error beyond the reproduced polynomial space.

An RBF-FD system also includes radial basis equations. For a Gaussian phi(r)=exp(-(epsilon r) squared), the shape parameter affects the weights and conditioning unless polynomial constraints dominate this tiny example.

Did You Know? Polynomial augmentation is now central in many practical RBF-FD methods because it stabilises algebraic convergence and guarantees exact differentiation of the included polynomials.

Each centre chooses a local stencil and solves a small augmented system for the target operator. Those weights fill one sparse row of the global matrix.

Stencil size must be large enough for the polynomial basis and node geometry. Oversized stencils cost more and can complicate stability; undersized or clustered stencils can be singular.

Scaling coordinates within each stencil improves numerical conditioning and makes shape parameters interpretable. Operator weights must be rescaled back to physical units.

For time-dependent PDEs, the spatial RBF-FD matrix couples to an eligible time integrator. A stable local interpolation does not guarantee a stable global evolution step.

Boundary rows approximate value, normal derivative or mixed conditions. Normals and stencil geometry near corners need special care.

Students can recover [1,-2,1] from moment conditions, then move one node, rebuild the augmented RBF system and compare polynomial exactness and conditioning.

Applications include atmospheric flow, wave equations, geoscience and PDEs on surfaces. Local radial stencils offer sparse high-order flexibility when node and boundary quality are controlled.

A responsible RBF-FD report states kernel, shape parameter, polynomial degree, nodes, stencils, scaling, operator weights, boundary rows, conditioning, solver, time step, convergence and validation.

For parents and students, the transferable lesson is that local geometry can be turned into a custom derivative formula while exact polynomial tests reveal what that formula can and cannot promise.

---

Frequently asked questions

What is RBF-FD?

A local method that uses radial basis interpolation to generate finite-difference-like operator weights.

Why add polynomials?

They guarantee exact differentiation of included polynomial fields and improve convergence behaviour.

Is the global matrix dense?

Usually no; local stencils create sparse rows.

What controls stencil quality?

Node geometry, size, kernel, shape parameter, polynomial degree and scaling.

Does local interpolation stability guarantee time stability?

No. The global operator and time integrator must also be analysed.

What must be reported?

Kernel, epsilon, polynomials, nodes, stencils, scaling, boundaries, solver, convergence and validation.

---

Primary source and fact-check note

G. B. Wright and B. Fornberg, “Scattered Node Compact Finite Difference-Type Formulas Generated from Radial Basis Functions,” Journal of Computational Physics 212, 2006. 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

• Radial basis function collocation
• Generalized finite difference method
• Finite difference method
• Computational mathematics hub
• 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: approximation, operator and evidence

Stencil audit

Record centres, neighbours, scale and polynomial degree.

Kernel audit

State radial basis, shape parameter and augmented system.

Operator audit

Test exact polynomial derivatives and sparse row ownership.

Evolution audit

Check boundaries, spectrum, time step and validation.

Worked evidence notebook: reconstruct every number

Evidence notebook 1: Stencil

Use nodes -1,0,1. Symmetric local cloud. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Second derivative at zero. Record nodes. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 2: Constant condition

Require sum weights=0. Second derivative of one is zero. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Polynomial augmentation. Record equation. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 3: Linear condition

Require -w1+w3=0. Second derivative of x is zero. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Symmetry emerges. Record equation. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 4: Quadratic condition

Require w1+w3=2. Second derivative of x squared is two. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Physical spacing is one. Record equation. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 5: Solve

Combine conditions. Weights=[1,-2,1]. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Check all moments. Record vector. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 6: Constant test

Apply weights to ones. Result=0. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Exact. Record check. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 7: Linear test

Apply to node coordinates. Result=0. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Exact. Record check. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 8: Quadratic test

Apply to squares. Result=2. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Exact. Record check. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 9: Quartic test

Apply to fourth powers. Result=2. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Exact value at zero is zero. Record error. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 10: RBF system

Add kernel equations. Local augmented matrix. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. State epsilon. Record conditioning. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 11: Scaling

Normalise stencil radius. Weights rescale by length squared. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Keep units. Record factor. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 12: Global row

Insert weights into sparse operator. One centre per row. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Check ownership. Record pattern. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 13: Boundary

Replace operator row as required. Use value or normal derivative. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. State normals. Record residual. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Evidence notebook 14: Conclusion

State kernel, polynomials, stencil and validation. Worked polynomial weights [1,-2,1]. Keep units, coordinates, matrix entries, signs, basis order and rounding visible. Three-node teaching stencil. Write narrow claim. Change one input, recompute the affected mechanism and identify whether the difference comes from node geometry, support, integration, boundary data, conditioning, time stepping or physical modelling. Preserve enough evidence for another reader to reproduce the narrow claim.

Misconceptions, failure modes and recovery

RBF-FD is global collocation

It uses local stencils. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

Kernel alone guarantees convergence

Polynomial augmentation and node quality matter. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

One epsilon fits every stencil

Scaling and geometry vary. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

Local systems are never singular

Clusters can destroy rank. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

Sparse means cheap automatically

Stencil solves and conditioning cost time. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

Polynomial exactness covers all functions

Higher-order error remains. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

Stable spatial rows guarantee stable time steps

The full spectrum and integrator matter. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

Surface plots validate PDEs

Manufactured solutions and refinement remain necessary. Build a minimal counterexample, restore the missing condition and recalculate before making a broader claim.

A responsible eight-step decision path

1. Choose nodes and stencils

State the variables, units and geometry.

2. Select kernel and polynomials

Write the local algebra and check rank.

3. Scale local coordinates

Verify the exact fields promised by the basis.

4. Solve operator weights

Connect local rules to the PDE operator.

5. Test polynomial exactness

Measure residual, conditioning and conservation.

6. Build boundary rows

Challenge difficult boundaries and irregular nodes.

7. Assemble and analyse

Repeat with systematic resolution changes.

8. Validate physics

Compare with independent evidence and state limits.

Learner lab: build transferable understanding

Polynomial constraints

Recover [1,-2,1]. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

RBF system

Add Gaussian equations. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Shape sweep

Vary epsilon and conditioning. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Irregular stencil

Move the right node. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Scaling

Normalise radius and restore units. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Sparse matrix

Assemble several local rows. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Boundary

Construct value and derivative rows. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Time step

Inspect the operator spectrum. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Student explanation

Explain a custom local stencil. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Parent explanation

Relate distance patterns to derivative rules. State a prediction, show the arithmetic, record one diagnostic, compare a baseline and finish with one supported conclusion plus one limit.

Deeper transfer: from local approximation to trustworthy simulation

Augmented system

The local saddle-point system couples RBF coefficients with polynomial constraints. Stable factorisation and consistent ordering matter, especially for high-degree polynomials.

Polyharmonic splines

Polyharmonic spline kernels avoid a shape parameter and are often paired with polynomials. Their accuracy depends strongly on polynomial degree and stencil quality.

Surface PDEs

For nodes on a surface, project or formulate derivatives in the tangent geometry. Normal estimation and curvature become part of the error budget.

Hyperviscosity

Advection-dominated RBF-FD schemes may add stabilising high-order dissipation. Its coefficient changes the model numerically and should be justified by spectra and convergence.

Stencil selection

Nearest-neighbour stencils are simple, but geometric quality filters or overlap control may improve robustness. Rejected stencils need a deterministic recovery rule.

Precomputation

Static nodes allow local weights to be computed once. Moving nodes require repeated neighbour search and solves, changing the cost profile.

Verification ladder

Begin with polynomial exactness, then manufactured PDEs, grid refinement, boundary tests and finally physical validation. Each rung answers a different question.

Career connection

RBF-FD combines radial geometry, polynomial approximation, sparse matrices and time integration. It is a rich bridge from school algebra and calculus to modern simulation.

Student and parent guidance: learning the method without losing the meaning

A student’s first route

Recover the familiar [1,-2,1] formula from polynomial constraints, then see how radial basis equations generalise the same idea to irregular nodes.

A parent’s useful question

Ask the learner what the stencil promises exactly and what its quartic counterexample reveals.

A small coding exercise

Build one augmented local system, vary a node position and monitor weights, polynomial defects and conditioning.

An engineering exercise

Assemble a sparse Laplacian from overlapping stencils and compare several polynomial degrees on a manufactured solution.

A verification habit

Check every local augmented solve, global boundary row, operator spectrum and refinement rate before trusting a time-dependent result.

A career pathway

RBF-FD develops practical command of kernels, polynomials, sparse linear algebra and PDE time integration for modern numerical modelling.

Scope, evidence and honest limits

What the worked example proves

Polynomial constraints recover one classic second-derivative stencil. A genuine RBF-FD method also solves the augmented radial system.

What a software result must add

A complete result adds kernel choice, shape or spline order, scaling, stencil selection, boundary rows, spectrum and time-step evidence.

What to compare

Use polynomial tests, manufactured PDEs and node-refinement studies, then compare with independent physical measurements where relevant.

What to preserve

Save stencils, augmented matrices, local weights, condition estimates, sparse operators, residuals and convergence plots.

A final communication check

Explain the method once to a student and once to a parent. Keep the mechanism, arithmetic, evidence and limit intact while changing only the vocabulary. If the claim becomes stronger during simplification, correct it.