Chapter map

1. 1. Nodes
2. 2. Kernel
3. 3. Matrix diagonal
4. 4. Nearest entries
5. 5. Outer entries
6. 6. Data
7. 7. Solve
8. 8. Node check
9. 9. New point
10. 10. Derivative
11. 11. PDE rows
12. 12. Conditioning
13. 13. Refinement
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. Nodes

A small example makes the mechanism visible. Use x=0,1,2.

A concrete situation

Three centres.

What the mathematics does not license

Scale coordinates explicitly. Evidence becomes stronger when the limitation is tested instead of hidden.

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. Kernel

This is where modelling choices begin to affect evidence. Use phi(r)=exp(-r squared), epsilon=1.

A concrete situation

Gaussian global basis.

What the mathematics does not license

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

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.

---

3. Matrix diagonal

The central mathematical idea is concrete. Evaluate r=0.

A concrete situation

Diagonal entries=1.

What the mathematics does not license

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

Try it, check it, transfer it

Record entries. 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. Nearest entries

Here the geometry becomes operational. Evaluate r=1.

A concrete situation

exp(-1)=.367879.

What the mathematics does not license

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

Try it, check it, transfer it

Record value. 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. Outer entries

This step deserves a careful pause. Evaluate r=2.

A concrete situation

exp(-4)=.018316.

What the mathematics does not license

Longer interaction is smaller. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record value. 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. Data

The useful question is not only how to calculate, but what the calculation assumes. Set f=[0,1,0].

A concrete situation

Symmetric target values.

What the mathematics does not license

Interpolation problem only. 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.

---

7. Solve

A small example makes the mechanism visible. Solve A alpha=f.

A concrete situation

alpha≈[-.492051,1.362031,-.492051].

What the mathematics does not license

Use a stable solver. Evidence becomes stronger when the limitation is tested instead of hidden.

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.

---

8. Node check

This is where modelling choices begin to affect evidence. Multiply A alpha.

A concrete situation

Returns [0,1,0] to roundoff.

What the mathematics does not license

Small residual is necessary only. Keeping this boundary clear makes the result easier to audit and transfer.

Try it, check it, transfer it

Record norm. 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. New point

The central mathematical idea is concrete. Evaluate at x=.5.

A concrete situation

Interpolant≈.625679.

What the mathematics does not license

All centres contribute. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record value. 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. Derivative

Here the geometry becomes operational. Differentiate Gaussian analytically.

A concrete situation

Second derivative=(4r squared-2)exp(-r squared).

What the mathematics does not license

Include epsilon factors when changed. Evidence becomes stronger when the limitation is tested instead of hidden.

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.

---

11. PDE rows

This step deserves a careful pause. Use operator rows inside and boundary rows outside.

A concrete situation

Algebra mirrors the boundary-value problem.

What the mathematics does not license

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

Try it, check it, transfer it

Record matrix. 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. Conditioning

The useful question is not only how to calculate, but what the calculation assumes. Vary epsilon.

A concrete situation

Accuracy and condition number change.

What the mathematics does not license

Flat limit needs stable algorithms. A trustworthy workflow records the choice before seeing the final test result.

Try it, check it, transfer it

Record both. 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. Refinement

A small example makes the mechanism visible. Add nodes and reduce fill distance.

A concrete situation

Compare solution between nodes.

What the mathematics does not license

Do not use residual alone. 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.

---

14. Conclusion

This is where modelling choices begin to affect evidence. State kernel, nodes, collocation and solver.

A concrete situation

Worked coefficients and value .625679.

What the mathematics does not license

Three-node teaching interpolation. Keeping this boundary clear makes the result easier to audit and transfer.

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

Radial basis function collocation approximates a field with basis functions that depend on distance from chosen centres. Because the nodes may be scattered, the method offers a direct route from geometry and data locations to interpolation, differentiation and partial differential equation residuals.

For a small interpolation example, use nodes x=0,1,2 and Gaussian basis φ(r)=exp(-r²) with shape parameter ε=1. The interpolation matrix has diagonal entries 1, nearest-neighbour entries exp(-1)≈0.367879 and outer entries exp(-4)≈0.018316.

Set nodal values f=[0,1,0]. Solving the symmetric 3×3 system gives coefficients approximately [-0.492051,1.362031,-0.492051]. Substitution returns the three nodal values to roundoff.

At x=0.5, the basis distances are 0.5,0.5 and 1.5. The interpolant is approximately 0.625679. That value is determined by all three centres because the Gaussian basis has global support.

The second derivative of exp(-r²) with respect to x is (4r²-2)exp(-r²). Evaluating derivative formulas at collocation nodes turns the same basis into a differentiation matrix for a differential equation.

For a boundary-value problem, boundary rows enforce boundary data while interior rows enforce the differential equation. The resulting algebraic system therefore contains the model, geometry and boundary conditions in visibly different rows.

Did You Know? The shape parameter creates a famous trade-off: flatter Gaussian basis functions can be very accurate for smooth problems yet produce severely ill-conditioned matrices in straightforward arithmetic.

Stable algorithms such as RBF-QR or alternative bases can separate approximation quality from numerical instability in flat regimes. Simply increasing decimal precision may help diagnosis but does not replace a stable formulation.

Globally supported RBFs create dense matrices, so memory and direct-solve cost grow rapidly. Compactly supported RBFs, partition-of-unity methods, local RBF finite differences and iterative solvers offer different locality and sparsity trade-offs.

Polynomial augmentation may be required for conditionally positive-definite kernels and can improve reproduction properties. The accompanying side constraints are part of the system, not optional decoration.

Node placement matters. Closely clustered nodes can worsen conditioning, while large holes reduce local resolution. Fill distance, separation distance and boundary coverage provide more meaningful diagnostics than node count alone.

Collocation can be easy to write but still sensitive to boundary treatment and operator stability. A small interpolation residual does not prove that the PDE solution is accurate between nodes.

Students can build the 3×3 matrix, solve for coefficients, verify nodal interpolation, evaluate at a new point and then differentiate the basis. Repeating the exercise for several shape parameters reveals accuracy and conditioning together.

Applications include scattered-data interpolation, geoscience, surfaces, fluid flow, heat transfer and local high-order PDE stencils. Kernel choice should follow the mathematical problem and evidence, not fashion.

For parents and students, the transferable idea is that distance can organise a flexible coordinate system even when data points do not lie on a neat grid.

A responsible RBF report states kernel, shape parameter, centres, scaling, polynomial terms, collocation rows, boundary treatment, conditioning, linear solver, stabilisation, node refinement, error tests and validation.

---

Frequently asked questions

What is an RBF?

A function whose value depends on distance from a centre, often with a scale or shape parameter.

What is collocation?

Enforcing interpolation data, differential equations or boundary conditions at selected points.

Why are RBF matrices often dense?

Globally supported kernels connect every centre to every collocation point.

What is the shape-parameter problem?

Accuracy and conditioning can pull in opposite directions for smooth kernels such as Gaussians.

Does a tiny collocation residual prove an accurate PDE solution?

No. Error between nodes, boundaries, stability, conditioning, refinement and validation must be checked.

What must be reported?

Kernel, shape parameter, nodes, polynomial terms, collocation rows, boundaries, conditioning, solver, convergence and validation.

---

Primary source and fact-check note

E. J. Kansa, “Multiquadrics—A Scattered Data Approximation Scheme with Applications to Computational Fluid-Dynamics—II,” Computers & Mathematics with Applications 19, 1990. 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

• Isogeometric analysis
• Finite element method
• Vortex particle 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: what must be measured

Node ledger

Record centres, collocation points, scaling and fill distance.

Kernel ledger

State basis, shape parameter and polynomial augmentation.

System ledger

Separate interpolation, PDE and boundary rows.

Numerical ledger

Report conditioning, solver, residual, refinement and validation.

Worked evidence notebook: reconstruct every number

Evidence notebook 1: Nodes

Use x=0,1,2. Three centres. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Scale coordinates explicitly. Record nodes. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 2: Kernel

Use phi(r)=exp(-r squared), epsilon=1. Gaussian global basis. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. State shape parameter. Record formula. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 3: Matrix diagonal

Evaluate r=0. Diagonal entries=1. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Exact kernel value. Record entries. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 4: Nearest entries

Evaluate r=1. exp(-1)=.367879. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Symmetric matrix. Record value. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 5: Outer entries

Evaluate r=2. exp(-4)=.018316. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Longer interaction is smaller. Record value. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 6: Data

Set f=[0,1,0]. Symmetric target values. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Interpolation problem only. Record vector. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 7: Solve

Solve A alpha=f. alpha≈[-.492051,1.362031,-.492051]. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Use a stable solver. Record residual. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 8: Node check

Multiply A alpha. Returns [0,1,0] to roundoff. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Small residual is necessary only. Record norm. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 9: New point

Evaluate at x=.5. Interpolant≈.625679. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. All centres contribute. Record value. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 10: Derivative

Differentiate Gaussian analytically. Second derivative=(4r squared-2)exp(-r squared). Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Include epsilon factors when changed. Record formula. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 11: PDE rows

Use operator rows inside and boundary rows outside. Algebra mirrors the boundary-value problem. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. State row ownership. Record matrix. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 12: Conditioning

Vary epsilon. Accuracy and condition number change. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Flat limit needs stable algorithms. Record both. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 13: Refinement

Add nodes and reduce fill distance. Compare solution between nodes. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Do not use residual alone. Record error. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Evidence notebook 14: Conclusion

State kernel, nodes, collocation and solver. Worked coefficients and value .625679. Keep every unit, direction, coordinate, matrix entry, sign convention and rounding decision visible. Three-node teaching interpolation. Write narrow claim. Then change one input, recompute the affected mechanism and explain whether the difference comes from geometry, constitutive choice, kernel support, conditioning, boundary data, time integration or numerical precision. Preserve the calculation so another reader can reproduce the narrow conclusion.

Misconceptions, failure modes and recovery

RBF means Gaussian only

Many positive-definite and conditionally positive-definite kernels exist. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

Small residual means small error

Collocation can miss error between nodes. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

Flatter is always better

Ill conditioning can destroy straightforward calculations. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

Dense matrices scale cheaply

Memory and direct solve cost grow rapidly. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

More nodes always help

Clustering and scaling can worsen conditioning. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

Boundary rows are ordinary data

They encode different mathematical constraints. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

One shape parameter fits all

Kernel scale interacts with geometry and operator. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

Smooth contours prove accuracy

Refinement and independent validation are still needed. Build the smallest counterexample, restore the missing condition and repeat the calculation before making a broader claim.

A responsible eight-step decision path

1. Scale geometry

Normalise coordinates and select centres.

2. Choose kernel

State basis, shape parameter and polynomials.

3. Build rows

Separate data, PDE and boundary equations.

4. Solve stably

Monitor conditioning, residual and precision.

5. Evaluate

Interpolate or differentiate at target points.

6. Challenge nodes

Test clustering, holes and boundary coverage.

7. Refine

Reduce fill distance and compare stable algorithms.

8. Validate

Measure error away from collocation nodes.

Learner lab: build transferable understanding

Matrix build

Evaluate a Gaussian matrix by hand. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Coefficient solve

Reproduce the three-node example. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

New point

Evaluate at several locations between nodes. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Shape sweep

Track error and condition number against epsilon. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Derivative

Verify the analytic second derivative numerically. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Boundary rows

Construct a Poisson collocation system. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Node quality

Compare uniform and clustered centres. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Local method

Build a small RBF-FD stencil. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Student explanation

Explain a coordinate system made from distance. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Parent explanation

Relate scattered measurements to a smooth map. State a prediction, show the arithmetic, record a diagnostic, compare a baseline and finish with one justified conclusion plus one limit.

Deeper transfer: from classroom mechanism to responsible use

Positive definiteness

For distinct centres, some kernels generate positive-definite interpolation matrices. Others are conditionally positive definite and require polynomial augmentation plus side constraints. The classification determines which algebraic system is mathematically eligible.

Scaling the coordinates

If one coordinate is measured in kilometres and another in metres, raw Euclidean distance gives an arbitrary geometry. Nondimensionalisation or anisotropic scaling should reflect the problem before a shape parameter is tuned.

Global versus local accuracy

Global RBF collocation uses every centre and can deliver high accuracy for smooth fields, but dense algebra limits scale. Local RBF-FD builds sparse differentiation stencils; its stability then depends on stencil geometry, polynomial degree and weights.

Boundary geometry

Scattered nodes can follow curved boundaries without a structured mesh, but normals and boundary operators must still be accurate. Poorly spaced boundary nodes may dominate error even when interior fill distance is small.

Condition number is not error

A large condition number warns that input or roundoff perturbations may be amplified. It does not directly state approximation error. Report conditioning, solver residual and independent solution error as different quantities.

Leave-one-out checks

For data interpolation, remove one observation, rebuild or update the model and predict the omitted value. This tests transfer between nodes. It is not a substitute for PDE convergence when the target is a differential equation.

Manufactured solutions

Choose a smooth exact field, apply the differential operator to derive a source and impose matching boundaries. The numerical error then exposes collocation, differentiation and boundary implementation before a real application is attempted.

Career connection

Geoscientists, graphics specialists, engineers and numerical analysts use radial bases for scattered measurements and differential equations. Students practise distance geometry, matrix methods, calculus, conditioning, algorithm design and honest uncertainty reporting.