Trefftz methods approximate a differential-equation solution using functions that already satisfy the equation locally. The remaining work is to fit boundary conditions, connect neighbouring regions correctly and control numerical error. Their gold standard is a trustworthy global solution: satisfying the equation inside each region is valuable, but does not by itself establish accuracy, uniqueness or physical balance.
This guide develops that distinction through complete harmonic, wave and interface examples. You will see how a perfect sampled boundary fit can miss an entire solution component, why resonance can defeat an otherwise exact basis, and how conservation determines the connection between two materials. The calculations are teaching models with stated assumptions; practical performance depends on the equation, geometry, basis and coupling formulation.
- Start from the Gold Standard master and Education parent for dependable work and learning that transfers.
- A Trefftz basis removes the homogeneous differential-equation residual locally; boundary and interface errors still need testing.
- Read the existing Trefftz method specialist for the initial connection between basis functions and boundary fitting.
- Use the Elliptic Partial Differential Equations root for boundary conditions and well-posedness, and the Method of Fundamental Solutions root for an external-source construction.
- Explore How X Works when connecting mathematical models with physical mechanisms.
Choose your Trefftz methods route
Chapters 1–3: Understand the equation and boundary fit
Chapters 4–6: Solve waves and examine resonance
Chapters 7–9: Connect basis stability and interfaces
Chapters 10–12: Handle forcing, fitting and verification
Chapter index
Understand the equation and boundary fit · Chapters 1–3
1. Begin with the equation and the missing conditions
Solve waves and examine resonance · Chapters 4–6
4. Solve a wave problem from start to finish
Connect basis stability and interfaces · Chapters 7–9
7. Understand why similar waves create fragile coefficients
Handle forcing, fitting and verification · Chapters 10–12
10. Add forcing and variable coefficients honestly
Transfer and continue learning · Chapters 13–14
CHAPTER 1 OF 14 · Understand the equation and boundary fit
1. Begin with the equation and the missing conditions
Suppose a field u is required to satisfy Lu = 0 inside a region, where L is a linear differential operator. A Trefftz approximation has the form uₙ = c₁φ₁ + … + cₙφₙ, with Lφⱼ = 0 wherever each basis function is used. Linearity then gives Luₙ = 0 independently of the coefficient values. This is the central construction, not a conclusion that uₙ is already the desired answer.
Consider a steady temperature model with constant conductivity and no internal heat source. The interior equation is Laplace’s equation, Δu = 0. Many temperature fields satisfy it. The imposed temperature or heat flux at the boundary determines which field is physically relevant. A constant field, a linear gradient and a suitable harmonic quadratic can all satisfy the same interior equation while representing different experiments.
For time-harmonic waves in a homogeneous medium, a common scalar model is Δu + k²u = 0, with wavenumber k. Again, the equation admits many fields. Boundary conditions, incident waves, interface laws and, for an exterior problem, an appropriate radiation condition select the intended solution. Complex amplitudes describe phase as well as magnitude; a real-valued plot can hide phase errors.
State the whole mathematical job
Before selecting a basis, write the operator, region, coefficient values, forcing, boundary conditions and desired output. A pressure value, a boundary normal derivative and an integrated flux are different outputs. An approximation adequate for one may be inadequate for another.
The survey by Hiptmair, Moiola and Perugia describes several ways to formulate Trefftz approximations for Helmholtz problems, including least-squares and discontinuous Galerkin approaches. These share an equation-adapted space but impose the remaining conditions differently.
The useful question is therefore: which parts of this particular problem are satisfied by construction, and which parts must the numerical scheme still enforce? Keeping those two lists separate prevents an exact local property from becoming an unjustified claim about the complete calculation.
CHAPTER 2 OF 14 · Understand the equation and boundary fit
2. Build a harmonic basis by differentiation
In two dimensions, Δu means uₓₓ + uᵧᵧ. Five simple harmonic functions are 1, x, y, x² − y² and 2xy. For the fourth function, the two second derivatives are 2 and −2, which cancel. For the fifth, both pure second derivatives vanish. Any linear combination of these functions is harmonic throughout the plane.
Start instead with a general quadratic, a + bx + cy + dx² + exy + fy². Its Laplacian is 2d + 2f. Requiring harmonicity gives f = −d, leaving five independent coefficients. We have reduced the six-dimensional quadratic polynomial space to the five-dimensional subspace compatible with the equation. This is a concrete example of using known structure before solving for unknown values.
On the unit circle, x = cos θ and y = sin θ. The same five functions become 1, cos θ, sin θ, cos 2θ and sin 2θ. That boundary representation makes their distinct angular behaviours visible. It also suggests why sample placement matters: an observation pattern must distinguish the angular modes that the basis can represent.
What a larger space can and cannot promise
Adding higher-degree harmonic polynomials expands the range of representable harmonic fields. It does not automatically cure a poorly conditioned fit, insufficient boundary sampling or a singularity that limits smoothness. Approximation quality and algebraic stability are separate questions.
The harmonic-polynomial approximation study by Hiptmair, Moiola, Perugia and Schwab establishes exponential approximation results under specific domain and analytic-extension assumptions. Those assumptions matter. A general statement that every harmonic problem converges exponentially with a single global polynomial basis would go beyond the result.
For a first implementation, differentiate every proposed basis function independently and inspect the result. If a supposed Laplace basis includes x² + y², its Laplacian is 4, not zero. That function may be useful as a particular solution for a forced equation, but it does not belong in this homogeneous harmonic space.
CHAPTER 3 OF 14 · Understand the equation and boundary fit
3. Expose a boundary fit that cannot see the answer
Take the unit disk and prescribe the boundary value g(x,y) = xy. The exact interior solution is u = xy because its Laplacian vanishes and it agrees with the boundary data. In the five-function basis of Chapter 2, its coefficient on 2xy is one half and all other coefficients are zero.
Now sample the boundary only at the four cardinal points: (1,0), (0,1), (−1,0) and (0,−1). At every one of these points, xy = 0. The zero field therefore fits all four observations exactly. So does every multiple of xy. The observations cannot determine the coefficient of this mode.
At the diagonal boundary point (1/√2, 1/√2), however, the exact value is one half. The zero approximation has error one half there. At the interior point (1/2,1/2), its error is one quarter. Neither error appears in the four fitted residuals. Nor does checking the Laplace residual help: both the true and zero fields satisfy the interior equation exactly.
Repair the observation design
Use eight equally spaced boundary angles θ = jπ/4 for j = 0,…,7. The five traces 1, cos θ, sin θ, cos 2θ and sin 2θ are now distinguishable; their sampled columns are linearly independent. Exact-data least squares recovers the one-half coefficient on sin 2θ and therefore the exact field xy.
The repair is not simply “use more points”. Repeating the same four cardinal positions many times would leave the invisible mode invisible. Additional samples must contribute independent information. Inspect matrix rank and use validation points that do not repeat the fitting pattern.
For a practical diagnostic, rotate the validation angles away from the fitting angles and compare maximum boundary error. If a sampled fit reports zero while rotated points show a substantial discrepancy, investigate unresolved modes or insufficient sampling before changing the linear solver. This example fails at the information stage, even with exact arithmetic.
CHAPTER 4 OF 14 · Solve waves and examine resonance
4. Solve a wave problem from start to finish
Consider the one-dimensional Helmholtz equation u″ + k²u = 0 on 0 ≤ x ≤ L, with k positive. Its general solution is a cos(kx) + b sin(kx). Both basis functions satisfy the equation exactly. If u(0) = A, then a = A. Imposing u(L) = B gives b = [B − A cos(kL)]/sin(kL), provided the denominator is nonzero.
Set k = 1, L = π/2, A = 2 and B = 3. Then cos(kL) = 0 and sin(kL) = 1, so the complete solution is u(x) = 2 cos x + 3 sin x. Direct substitution gives u″ = −2 cos x − 3 sin x, hence u″ + u = 0. The endpoint values are exactly 2 and 3.
At the midpoint x = π/4, the answer is 5/√2, approximately 3.535534. This is already larger than either endpoint value. The maximum occurs where u′ = −2 sin x + 3 cos x = 0, or tan x = 3/2. At that interior point the value is √13, approximately 3.605551.
Do not transfer the wrong error principle
For harmonic functions, a maximum principle can control interior values through boundary values under the appropriate assumptions. This positive-k Helmholtz example shows why one cannot carry that statement over unchanged. Oscillation permits an interior peak greater than both endpoint values.
The calculation also shows the attraction of an adapted basis: two coefficients describe the complete solution here. That success follows from the exact solution space of this simple ordinary differential equation. It does not imply that a complicated two-dimensional scattering problem requires only two functions.
Use this case as a unit-level mathematical check for signs, derivative conventions and endpoint assembly. If software cannot reproduce this solution away from resonance, adding directions, elements or more elaborate geometry will make diagnosis harder rather than repair the underlying mistake.
CHAPTER 5 OF 14 · Solve waves and examine resonance
5. Separate resonance from numerical ill-conditioning
Return to b = [B − A cos(kL)]/sin(kL). When kL = mπ for a positive integer m, the denominator vanishes. The endpoint equations are compatible only if B = (−1)ᵐA. If that condition holds, b is arbitrary: infinitely many fields satisfy the same boundary data. If it fails, there is no solution to these imposed endpoint conditions.
For example, with k = 1, L = π and A = B = 0, every field b sin x satisfies the equation and both boundary conditions. The zero boundary residual cannot distinguish a zero field from a large standing wave. This is nonuniqueness in the mathematical problem, not an error caused by choosing too few basis functions.
A near-resonant sensitivity calculation
Let k = 1, L = π − 0.01, A = 0 and B = 0.001. Then b = 0.001/sin(0.01), approximately 0.100001667. At x = L/2, the field is approximately 0.100000417. A boundary value of one thousandth produces an interior value of about one tenth in this example.
The derivative of b with respect to B is 1/sin(kL), approximately 100.001667. This quantifies the sensitivity: a small perturbation of the endpoint data can create a much larger coefficient change. The differential-equation residual remains zero for every value of b.
A more stable basis or higher-precision arithmetic may reduce computational error, but cannot remove sensitivity inherent in the boundary-value problem. Conversely, an awkward basis can introduce additional numerical ill-conditioning even when the underlying problem is well posed. Diagnose both levels.
In a physical model, damping, impedance conditions or other boundary physics may change the resonance behaviour. Include them because they describe the actual system, not merely to make a matrix invertible. If the requested model is singular, report that fact and identify the missing compatibility or uniqueness condition. An arbitrary minimum-norm coefficient choice is not a substitute for a physical selection rule.
CHAPTER 6 OF 14 · Solve waves and examine resonance
6. Choose plane waves that satisfy the intended operator
In several dimensions, a standard Helmholtz basis function is φ(x) = exp(ik d·x), where d is a unit direction vector. Differentiating twice gives Δφ = −k²|d|²φ. Therefore Δφ + k²φ = k²(1 − |d|²)φ, which vanishes when |d| = 1. The real and imaginary parts give cosine and sine solutions.
This calculation checks two implementation details immediately. Direction vectors need the intended normalisation, and the wavenumber used in the basis must match the operator. Using a unit vector with the wrong k still produces a smooth wave, but not a solution of the specified equation.
Direction error becomes phase error
Suppose the desired wave travels along d₀ but the basis contains a nearby direction d. Relative to an origin with |x| ≤ R, the phase difference is k(d − d₀)·x, whose magnitude is at most kR|d − d₀|. The corresponding complex exponential difference is bounded by that phase difference in magnitude, using |exp(iα) − exp(iβ)| ≤ |α − β|.
If the angle between the two unit directions is δ, their separation is 2 sin(δ/2). At kR = 20 and δ = 0.05 radians, the bound is approximately 0.999896. Reducing δ to 0.005 gives approximately 0.100000. These are upper bounds for one mismatched wave, not predicted errors for an optimised combination of many waves.
The lesson is that a small angular mismatch can matter when many wavelengths fit across a region. Refining a direction set and subdividing the domain address different parts of that challenge. A combination of directions can also approximate a field more effectively than the nearest single wave.
Evanescent behaviour and geometric singularities may require different local functions or refinement. Do not assume that a large collection of equally spaced propagating directions resolves every wave feature. Inspect which physical and mathematical behaviours the chosen space actually contains.
CHAPTER 7 OF 14 · Connect basis stability and interfaces
7. Understand why similar waves create fragile coefficients
Even distinct plane waves can look almost identical on a sufficiently small region. Examine φ₊ = exp(ikx) and φ₋ = exp(−ikx) on −h/2 ≤ x ≤ h/2. Use the normalised inner product obtained by integrating a function times the complex conjugate of the other and dividing by h. Each wave has squared norm one.
Their cross inner product is ρ = sin(kh)/(kh). The two-by-two Gram matrix has diagonal entries one and off-diagonal entries ρ. Its eigenvalues are 1 + ρ and 1 − ρ. When kh is small and positive, ρ is close to one, so the second eigenvalue is small.
Using sin z/z = 1 − z²/6 + higher-order terms, the Gram-matrix condition number behaves like 12/(kh)². The condition number of the corresponding basis-to-function map is its square root, approximately √12/(kh). Confusing these two quantities overstates or understates the numerical sensitivity being reported.
At kh = 0.1, ρ is approximately 0.998334166. The Gram condition number is approximately 1199.6 and the basis-map condition number approximately 34.635. This is a continuous two-function example; a sampled boundary matrix has its own conditioning and must be examined separately.
Repair the representation without changing the physics
Adding and subtracting these waves produces cosine and sine components. Rescaling the small sine component can make the local representation more balanced. Orthogonalisation can similarly expose independent information more clearly. These transformations preserve the span when performed nonsingularly, although finite-precision implementation still matters.
A large coefficient norm is therefore a warning to interpret, not an automatic proof that the field is wrong. Cancellation may produce an accurate field from large coefficients, but can make evaluation sensitive to rounding and data perturbations. Check singular values, field stability and independent residuals together.
Reducing redundant directions may improve robustness while sacrificing little useful approximation power. The decision should follow measured rank and error, rather than a rule that more basis functions must always be better.
CHAPTER 8 OF 14 · Connect basis stability and interfaces
8. Connect local solutions with the correct interface law
Consider steady one-dimensional conduction on 0 ≤ x ≤ 1. Conductivity is κ₁ = 1 on the left half and κ₂ = 4 on the right half. Assume no internal sources, perfect contact at x = 1/2, and endpoint temperatures u(0) = 0 and u(1) = 1. The equation is −(κu′)′ = 0.
Within each constant-conductivity region, affine functions satisfy the equation exactly. Let the unknown interface temperature be a. The left solution is u₁ = 2ax. The right solution, joining a to 1, has slope 2(1 − a). Temperature continuity alone allows every value of a.
Flux continuity supplies the missing equation. Since the physical heat flux is q = −κu′, it requires κ₁u₁′ = κ₂u₂′. Thus 2a = 8(1 − a), giving a = 0.8. The complete solution is u₁ = 1.6x and u₂ = 0.6 + 0.4x. Both give 0.8 at the interface, and the physical flux is −1.6 on both sides.
The same result follows from the total resistance: 0.5/1 + 0.5/4 = 0.625. The unit temperature difference divided by that resistance gives κu′ = 1.6. This independent calculation checks the interface assembly.
A continuous field can still violate conservation
The simple field u = x satisfies both endpoint values, is continuous at the interface and satisfies each local differential equation. Yet κu′ is 1 on the left and 4 on the right. It does not satisfy the no-source interface balance. Local exactness plus continuity is insufficient.
When neighbouring elements use outward normals, those normals point in opposite directions at their shared interface. Conservation is expressed as the sum of outward fluxes being zero. Accidentally using the same normal direction for both elements can reverse the coupling condition.
Different interface physics changes the law: imperfect contact may permit a temperature jump related to flux. State that model explicitly. A generic penalty on continuity should not silently replace a prescribed physical interface condition.
CHAPTER 9 OF 14 · Connect basis stability and interfaces
9. See what boundary integration removes and retains
For a smooth region K and a sufficiently regular Helmholtz solution u, integration by parts gives a useful identity. The volume integral of ∇u·∇v̄ − k²uv̄ equals the boundary integral of (∂u/∂n)v̄, because Δu + k²u = 0. Here v̄ is the complex conjugate of a test function and n is the outward unit normal.
The cancellation can be checked directly: integration by parts replaces the gradient product by the boundary term minus the volume integral of (Δu)v̄. Substituting Δu = −k²u cancels the remaining volume term. Exact satisfaction of the local equation has removed this particular interior contribution.
That identity explains why suitable Trefftz formulations can concentrate substantial work on element boundaries. It does not specify a complete numerical method. One must still choose test spaces, boundary terms, interface coupling and any stabilisation, then establish the properties needed for the chosen problem.
Boundaries still require numerical care
The geometric boundary must be represented accurately. Normals must correspond to that geometry. Oscillatory traces need sufficiently accurate integration; a volume-free expression is not an integration-free expression. Under-resolving a boundary integral can spoil a calculation even when all basis functions satisfy the equation analytically.
For a multi-region problem, boundary contributions from neighbouring regions are linked through the physical transmission conditions. One cannot solve every local equation independently and expect the assembled field to conserve flux automatically. Chapter 8 gives a small example of exactly what would be missing.
For an exterior wave problem, a finite computational boundary also requires a justified treatment of outgoing behaviour. A basis that satisfies Helmholtz locally does not automatically impose the correct far-field condition.
The Method of Fundamental Solutions root provides a complementary construction using singularities outside the physical region. Its source placement issues differ from plane-wave direction selection, although both approaches exploit functions satisfying a homogeneous equation in the region of interest. Compare the actual approximation and enforcement mechanisms rather than grouping every boundary-based calculation under one interchangeable recipe.
CHAPTER 10 OF 14 · Handle forcing, fitting and verification
10. Add forcing and variable coefficients honestly
If Lu = f with f nonzero, a combination of homogeneous Trefftz functions alone cannot satisfy the full equation. A standard conceptual split is u = uₚ + uₕ, where Luₚ = f and Luₕ = 0. The boundary data for uₕ must subtract the boundary contribution of uₚ.
For a complete example, solve u″ + u = 1 on 0 ≤ x ≤ π/2, with zero values at both endpoints. Choose uₚ = 1. Then uₕ must have value −1 at both ends. The result is u = 1 − cos x − sin x. Its second derivative is cos x + sin x, so u″ + u = 1; both endpoints give zero. At the midpoint, u = 1 − √2, approximately −0.414214.
If an approximate particular solution is used, its forcing residual remains part of the total error. A small boundary mismatch cannot certify that the interior forcing was represented correctly. The existing approximate particular solutions specialist develops this connection between forcing and radial-basis constructions.
A constant-medium wave is not exact in a variable medium
For Δu + k(x)²u = 0, take φ = exp(ik₀d·x) with unit d. Its residual is [k(x)² − k₀²]φ. Freezing k at one point does not make that residual vanish elsewhere. Smaller regions may reduce the mismatch, but the method must account for it.
The Generalized Plane Waves study by Imbert-Gérard and Monk constructs approximate local solutions and modifies the formulation for spatially varying coefficients. It explicitly distinguishes those functions from exact constant-medium plane waves.
This is a useful boundary on terminology. An approximate or quasi-Trefftz construction can be valuable, but its local residual must be analysed or measured. Retaining the name while silently retaining the zero-residual assumption would hide an error source that the original exact construction did not have.
CHAPTER 11 OF 14 · Handle forcing, fitting and verification
11. Solve the boundary fit with meaningful weights
In a collocation or discrete least-squares approach, evaluate boundary operators applied to basis functions to form a matrix A. The coefficients c are chosen to reduce Ac − b, where b contains the prescribed observations. Rows might represent field values, normal derivatives or interface jumps. Their meaning determines how they should be scaled.
If u is measured in temperature units and a normal derivative in temperature per metre, adding their squared numerical residuals without scaling is dimensionally arbitrary. One possible dimensionally consistent derivative residual multiplies by a stated reference length. Physical flux rows may also require conductivity and reference scales. The weights express a norm or modelling choice; document that choice.
Boundary sampling density matters too. An unweighted sum gives more influence to a region containing more sample points. If the intended objective approximates a boundary integral, appropriate quadrature weights help represent boundary measure. For a least-squares matrix, square roots of those positive weights multiply the rows and data.
Use a solver appropriate to the rank
The LAPACK guide to linear least squares distinguishes full-rank QR/LQ solvers from methods that allow rank deficiency, including singular-value decomposition. This distinction matters for the invisible mode in Chapter 3 and the nearly dependent waves in Chapter 7.
Avoid treating the normal equations as a harmless rearrangement in a sensitive problem: forming A* A squares the spectral condition number for a full-column-rank matrix. QR or a suitable rank-revealing method avoids that particular amplification mechanism. Solver selection does not repair missing physical conditions or an unsuitable basis.
If singular values are truncated, record the threshold and examine how field values and validation residuals change across reasonable thresholds. Truncation trades information against stability. A smaller coefficient norm is not, by itself, evidence of a more accurate physical solution.
Keep residuals in both the solver’s weighted norm and understandable physical units. A single dimensionless score can conceal a large error in the output the reader actually needs.
CHAPTER 12 OF 14 · Handle forcing, fitting and verification
12. Validate the field through independent questions
A useful verification report separates the local equation, fitted conditions, independent boundary checks, interface balance and output error. These answer different questions. In an exact Trefftz space, a tiny interior differential residual is expected by construction and gives little information about whether the boundary data selected the right field.
Start with a manufactured solution such as xy on the disk or the two-material conduction example. Derive its boundary and interface data, then compare the computed field with the known answer at points excluded from fitting. Include derivatives or fluxes if those are the intended outputs. Report an absolute error and a relative error only where the reference scale makes that ratio meaningful.
Change one source of error at a time
Increase boundary sampling while keeping the basis fixed. Then enlarge the basis while keeping validation sufficiently resolved. In a piecewise method, refine the regions separately from increasing the local basis size. Tighten quadrature separately from solver tolerances. This sequence helps identify which change actually improves the result.
For oscillatory problems, report k together with element or region size h; kh describes local oscillation more meaningfully than h alone. Comparing two runs at different frequencies without recording this change can confuse a harder physical problem with a worse numerical method.
If a fit is accurate only at its training points, inspect sampling and rank. If local residuals are nonzero unexpectedly, inspect derivatives, coefficients and coordinate transformations. If interface fluxes disagree, inspect normals and coupling laws. If errors grow near a resonant parameter, investigate stability of the continuous problem as well as the matrix.
For an unknown solution, agreement between independently refined formulations, balance checks and residuals provides evidence, not automatic proof of every desired tolerance. A rigorous error bound requires its stated assumptions and stability constants. Near resonance, those constants may be unfavourable.
Finally, test the quantity that informs the decision. A small average field error may coexist with an inaccurate peak or normal derivative. Verification should follow the use of the calculation rather than the convenience of a single summary statistic.
CHAPTER 13 OF 14 · Transfer and continue learning
13. Transfer the method to a changed scale and material
Try these changes before reading the answers. First, a wave model uses x in metres, k = 2 inverse metres and L = 3 metres. Rewrite the same physical problem using centimetres. Second, keep the conduction problem of Chapter 8 but change the right conductivity from 4 to 9. Predict the new interface temperature and flux. Third, ask whether the plane-wave basis remains exact when k varies smoothly in space.
Worked answers
Write x꜀ = 100x. The same phase kx becomes (k/100)x꜀, so the new numerical wavenumber is 0.02 inverse centimetres and the length is 300 centimetres. Their product remains kL = 6. Keeping the numerical wavenumber at 2 while changing the coordinate units would change the physical wavelength by a factor of one hundred.
More generally, with ξ = x/L, the second derivative transforms as d²u/dx² = L⁻²d²u/dξ². Multiplying the Helmholtz equation by L² gives uξξ + (kL)²u = 0. This explains why the dimensionless product controls the rescaled problem. Derivative boundary data must transform consistently as well.
For the changed conductivity, total resistance is 0.5/1 + 0.5/9 = 5/9. Thus κu′ = 1.8 and the interface temperature is 0.9. The left slope is 1.8 and the right slope is 0.2, so the fields are 1.8x and 0.8 + 0.2x. The physical heat flux is −1.8. The higher-conductivity half supports a smaller temperature gradient for the same flux.
For varying k, the residual formula in Chapter 10 shows that the original constant-medium plane wave is generally no longer exact. You must change the local construction, retain the residual in the formulation or adopt another approximation strategy.
These changes test understanding beyond reproducing a familiar coefficient solve. The Education parent treats this ability to recognise what survives a changed context as a central learning outcome. Here, the invariant phase, transformed derivatives and conserved flux provide concrete evidence of that transfer.
CHAPTER 14 OF 14 · Transfer and continue learning
14. Questions, learning routes and primary reading
Are Trefftz methods always meshless?
No. A global construction may avoid a volume mesh, while a piecewise construction uses elements or subdomains with interface coupling. The defining idea is the local differential-equation property of the approximation functions, not a universal choice of geometric data structure.
Does an exact basis make the numerical solution exact?
Only if the required solution is representable and the remaining conditions are imposed and solved adequately. The disk example shows that even a representable exact solution can be missed by observations that cannot distinguish its mode. Rounding, geometry and integration introduce further practical issues.
Are all radial basis functions Trefftz functions?
No. A radial shape alone does not imply satisfaction of the chosen differential equation. Fundamental solutions away from their singularities can provide one relevant family. Other radial-basis methods approximate derivatives or construct particular solutions through different mechanisms. Check the operator applied to the actual function.
How should a beginner start?
First differentiate the five harmonic polynomials by hand. Next reproduce the cardinal-point failure and its eight-point repair. Then solve the one-dimensional wave problem and perturb its endpoint near resonance. Finally implement the two-material interface example. These exercises develop distinct skills: basis verification, information design, sensitivity analysis and conservation.
For practical wave modelling, add complex arithmetic, boundary normals, least-squares rank diagnostics and independent validation before attempting complicated scattering geometry. A small reproducible case with a known answer is a better diagnostic foundation than an impressive image without an error measure.
Which reading supports the next step?
Use the Helmholtz Trefftz survey to compare formulation families and local spaces. Read the harmonic-polynomial approximation paper to see how convergence statements depend on regularity and geometry. Follow the Generalized Plane Waves paper when coefficients vary, and the LAPACK least-squares guide when choosing the algebraic solve. The worked numerical cases in this guide are original teaching calculations rather than reported performance results from those papers.
Return to the Trefftz specialist for a compact conceptual explanation, then compare the fundamental-solutions root and elliptic-equations root. The next useful achievement is to explain which equation your basis satisfies, which conditions remain to be enforced, and which independent evidence shows that the resulting field is dependable.