Discrete Quasi-Trefftz Spaces

• Discrete quasi-Trefftz spaces are finite-dimensional, PDE-adapted approximation spaces that enforce local residual vanishing to balance exactness with computational efficiency.
• They extend classical Trefftz methods to variable-coefficient regimes by approximating PDE satisfaction through controlled Taylor residual annihilation.
• Their integration into DG and conforming frameworks yields high-order convergence and reduced global system sizes across wave, diffusion, and Maxwell problems.

Discrete quasi-Trefftz spaces are finite-dimensional, equation-dependent approximation spaces whose functions satisfy the governing partial differential equation exactly only in special cases and otherwise satisfy it approximately in a controlled local sense, typically through vanishing Taylor coefficients of the residual or through a relaxed weak local constraint. They were developed to extend classical Trefftz discretizations beyond the regime of linear homogeneous PDEs with piecewise-constant coefficients and explicitly available local solutions, while retaining the central Trefftz objective: replacing full polynomial spaces by much smaller problem-adapted spaces without sacrificing approximation order (Hiptmair et al., 2015, Lederer et al., 2024, Imbert-Gerard, 24 May 2025).

1. Definition and relation to classical Trefftz spaces

In the classical Helmholtz setting, a discrete Trefftz space is a finite-dimensional subspace of functions that satisfy the PDE pointwise on each mesh element; for a mesh $\mathcal T_h$, the survey literature writes

$$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$

and a discrete Trefftz space is any finite-dimensional subspace assembled from local spaces $V_{p_K}(K)\subset T(\mathcal T_h)$ (Hiptmair et al., 2015). In that same survey, quasi-Trefftz is described as a typically larger space whose elements satisfy the PDE only approximately, for example up to higher-order terms in $h_K$ (Hiptmair et al., 2015).

Later formulations make that approximation notion precise. For a second-order scalar elliptic operator $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$, the unified framework defines on each element $K$ the local quasi-Trefftz space

$$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$

where $x_K$ is a chosen center and $A_K$ collects the scaled residual jets at that point (Lederer et al., 2024). The same structure appears in the general Taylor-based theory, where for an operator $\mathcal L$ of order $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$0 one defines

$$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$1

so the local quasi-Trefftz space is the kernel of “Taylor truncation composed with the differential operator” (Imbert-Gerard, 24 May 2025).

This formulation places exact Trefftz and quasi-Trefftz spaces in a strict inclusion chain whenever exact polynomial solutions exist. For the elliptic framework,

$$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$2

and $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$3 in general (Lederer et al., 2024). A common misconception is that quasi-Trefftz spaces are merely ad hoc enrichments; the cited constructions instead characterize them as structured kernels of local residual operators.

2. Local algebraic structure, basis construction, and dimensions

The local algebraic core is a constrained polynomial space. In the general scalar theory, polynomial spaces are decomposed by total degree,

$$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$4

and the quasi-Trefftz operator $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$5 is shown, under a minimal non-degeneracy assumption on the principal part of $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$6, to be surjective (Imbert-Gerard, 24 May 2025). This yields the dimension formula

$$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$7

which is the canonical dimension count for Taylor-based polynomial quasi-Trefftz spaces (Imbert-Gerard, 24 May 2025).

At the implementation level, several equivalent constructions recur across the literature. One may assemble the matrix of residual moments or truncated jets in a monomial basis and compute a nullspace basis by Gaussian elimination, SVD, or rank-revealing QR (Lederer et al., 2024, Imbert-Gérard et al., 2024, Imbert-Gerard, 24 May 2025). Alternatively, one may prescribe “Cauchy data” and recover the remaining coefficients by explicit recurrences ordered by homogeneous degree; this is the preferred construction in the space-time wave equation, the diffusion-advection-reaction equation, and several first-order or vector generalizations (Imbert-Gérard et al., 2020, Perinati, 2023, Imbert-Gérard et al., 10 Sep 2025, Imbert-Gérard, 29 Aug 2025).

Representative local dimensions appearing in the literature are as follows.

Setting	Local space	Dimension
Scalar operator of order $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$8 in $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$9 dimensions	$V_{p_K}(K)\subset T(\mathcal T_h)$0	$V_{p_K}(K)\subset T(\mathcal T_h)$1
Second-order scalar elliptic, $V_{p_K}(K)\subset T(\mathcal T_h)$2	$V_{p_K}(K)\subset T(\mathcal T_h)$3	$V_{p_K}(K)\subset T(\mathcal T_h)$4
Space-time wave equation	$V_{p_K}(K)\subset T(\mathcal T_h)$5	$V_{p_K}(K)\subset T(\mathcal T_h)$6
Second-order time-harmonic Maxwell, $V_{p_K}(K)\subset T(\mathcal T_h)$7	$V_{p_K}(K)\subset T(\mathcal T_h)$8	$V_{p_K}(K)\subset T(\mathcal T_h)$9
Embedded Trefftz on quadrilaterals	$h_K$0	$h_K$1

These counts reflect the same structural principle: quasi-Trefftz spaces eliminate precisely the polynomial directions that generate low-order residuals. In two-dimensional second-order scalar problems, this produces the particularly important linear-in-$h_K$2 count $h_K$3, in contrast with the quadratic growth of the full polynomial space (Lederer et al., 2024, Perinati, 2023).

3. Global discretization patterns

Most discrete quasi-Trefftz methods are formulated in discontinuous Galerkin form. In the unified framework one begins with the broken polynomial space

$h_K$4

then decomposes it as $h_K$5, where $h_K$6 is the discontinuous global space whose elementwise restrictions lie in the local quasi-Trefftz spaces (Lederer et al., 2024). The method enforces two conditions simultaneously: local residual matching in the $h_K$7-direction and the DG residual against all test functions in $h_K$8. The resulting $h_K$9 block system simplifies in most quasi-Trefftz methods because the upper-right block vanishes, so one first solves local particular problems and then a reduced global DG problem on $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$0 (Lederer et al., 2024).

For scalar elliptic problems, the global bilinear form is typically an interior-penalty formulation with upwind treatment of advection. The diffusion-advection-reaction quasi-Trefftz DG method uses broken trial and test spaces with a DG norm containing broken $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$1-seminorms, jump penalization, and upwind terms, and proves consistency, discrete coercivity, and boundedness (Perinati, 2023). A closely related general elliptic formulation employs SIPG/upwind couplings for local spaces $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$2 defined by vanishing residual jets at element centers (Imbert-Gérard et al., 2024).

The space-time wave equation provides a different global pattern. There, local quasi-Trefftz polynomial spaces on space-time elements are assembled into a DG space for the first-order acoustic variables $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$3, and the choice of numerical fluxes makes the coupling causal, so no global space-time solve is required and one may march in time via element patches, even in parallel (Imbert-Gérard et al., 2020).

A distinct line is the embedded Trefftz DG method. On each anisotropic quadrilateral element $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$4, one defines a tensor-product space $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$5, a smaller local test space $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$6, and the relaxed local Trefftz condition

$$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$7

with $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$8 (Gómez et al., 2 Jun 2026). The corresponding local embedded Trefftz space $$L v=-\nabla\!\cdot(\alpha\nabla v)+\beta\cdot\nabla v+\gamma v$$9 is not imposed by pointwise residual annihilation but by weak orthogonality to $K$0; the global method then imposes the DG bilinear form only on the product space $K$1, again yielding a reduced global system (Gómez et al., 2 Jun 2026).

Although quasi-Trefftz methods are predominantly discontinuous, a globally conforming variant now exists for Helmholtz. The Trefftz Continuous Galerkin method constructs a space $K$2 from edge-based and node-based functions that are exact local Helmholtz solutions on each cell and continuous across the global Cartesian cut mesh, leading to a conforming Trefftz discretization rather than a DG coupling (Galante et al., 1 Dec 2025).

4. Approximation theory, stability, and convergence

The defining analytical feature of quasi-Trefftz spaces is that they preserve the approximation order of the surrounding full polynomial space. In the general scalar Taylor-based theory, if $K$3 is an exact local solution of $K$4 with $K$5, then its Taylor polynomial $K$6 belongs to $K$7, and the best-approximation error is $K$8 in $K$9 and $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$0 in the gradient (Imbert-Gerard, 24 May 2025). For local spaces $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$1 associated with a linear differential operator of order $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$2, one has

$$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$3

so the same $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$4-rates as full $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$5-approximation are recovered (Imbert-Gérard et al., 2024).

The unified DG framework turns this local property into global quasi-optimality. Discrete stability and Céa-type arguments yield

$$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$6

and for coercive second-order problems one obtains

$$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$7

under the stated regularity assumptions (Lederer et al., 2024). The diffusion-advection-reaction analysis reaches the same conclusion: the quasi-Trefftz DG method is well-posed, consistent, stable, and high-order convergent, with local spaces smaller than full polynomial spaces of the same degree (Perinati, 2023).

For the space-time wave equation with piecewise-smooth coefficients, the exact solution’s Taylor polynomial of degree $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$8 belongs to the local quasi-Trefftz space $$V_h^{QT}(K):=\{v\in P^p(K):D^i(Lv)(x_K)=0\ \forall\,|i|\le p-2\}=\ker A_K,$$9, and the assembled DG method satisfies a quasi-optimal bound in the mesh-dependent energy norm with asymptotic order $x_K$0, together with $x_K$1-type behavior at final time under duality arguments (Imbert-Gérard et al., 2020). The first-order Helmholtz formulation has the analogous local statement: if $x_K$2 is a sufficiently smooth exact solution, then $x_K$3, and the local approximation error scales like $x_K$4 for the fields and $x_K$5 for their gradients (Imbert-Gérard et al., 10 Sep 2025).

A misconception often attached to quasi-Trefftz spaces is that approximate local PDE satisfaction necessarily degrades stability. The available analyses point in the opposite direction: local residual annihilation is designed so that the global constants remain those of the full DG method, while the main effect is a reduction of the global dimension (Lederer et al., 2024). This suggests that quasi-Trefftz spaces should be understood less as weakened Trefftz spaces than as dimension-reduced polynomial spaces with PDE-adapted null directions.

5. Wave-propagation specializations

Wave problems have supplied the main impetus for quasi-Trefftz developments because the exact Trefftz philosophy is especially attractive there and especially fragile under variable coefficients. In homogeneous Helmholtz media, plane waves $x_K$6, generalized harmonic polynomials, evanescent waves, fundamental solutions, and multipoles are exact local Trefftz functions (Hiptmair et al., 2015). As soon as the coefficients vary, “no closed-form family of exact local solutions is available,” which motivates replacing exact local solutions by high-order approximate ones (Fontana et al., 13 Aug 2025).

One major specialization is the generalized plane wave construction. On each cell $x_K$7, one chooses a center $x_K$8 and basis functions of the form

$x_K$9

and determines the higher-order coefficients by imposing that the first $A_K$0 terms of the Taylor expansion of the PDE residual vanish (Fontana et al., 13 Aug 2025). The same paper explicitly relates GPW bases to polynomial quasi-Trefftz bases, making the phase-based and polynomial constructions part of a single framework.

Another specialization is the Trefftz Continuous Galerkin method for the two-dimensional Helmholtz equation. On each Cartesian rectangle $A_K$1, one solves a homogeneous Helmholtz–Dirichlet problem with nonzero trace on one chosen edge and zero trace on the others, obtaining single-edge modes

$A_K$2

which are propagative for $A_K$3 and evanescent for $A_K$4 (Galante et al., 1 Dec 2025). These functions can also be written as linear combinations of four complex-direction plane waves, are transplanted to adjacent cells sharing an edge, and yield globally continuous edge and node spaces in $A_K$5 (Galante et al., 1 Dec 2025). The resulting method proves wavenumber-explicit best-approximation bounds, bounded coefficients, and spectral accuracy for analytic Helmholtz solutions, with exponential decay of the approximation error both at fixed frequency and along suitable high-frequency sequences (Galante et al., 1 Dec 2025).

Quasi-Trefftz ideas have also been extended to first-order and vector systems. For a first-order formulation of the Helmholtz equation, the local space

$A_K$6

is characterized equivalently by the scalar condition $A_K$7 together with $A_K$8, and has dimension $A_K$9 (Imbert-Gérard et al., 10 Sep 2025). For the second-order time-harmonic Maxwell equation with variable coefficients, the vector-valued space $\mathcal L$0 is defined through coupled curl-curl and divergence truncation constraints, its dimension is $\mathcal L$1, and its constructive theory relies on a Helmholtz decomposition of homogeneous vector polynomial fields into solenoidal, irrotational, and harmonic components (Imbert-Gérard, 29 Aug 2025).

6. Computational aspects, numerical behavior, and active directions

Computationally, discrete quasi-Trefftz spaces trade larger local basis-construction work for smaller global systems. For scalar polynomial spaces, the construction is entirely local and can be done from monomial matrices, nullspace extraction, or block-triangular recursions (Imbert-Gérard et al., 2024, Imbert-Gerard, 24 May 2025). For the first-order Helmholtz formulation, explicit recurrences cost $\mathcal L$2 operations per function and $\mathcal L$3 to build the full $\mathcal L$4 basis, whereas SVD-based algebraic approaches cost $\mathcal L$5 flops (Imbert-Gérard et al., 10 Sep 2025). In the space-time wave equation, the local basis is built by a single pass of dimension-$\mathcal L$6 linear recurrences and its conditioning remains moderate in the reported experiments (Imbert-Gérard et al., 2020).

Assembly may also benefit from the PDE-adapted basis. The Helmholtz survey emphasizes that Trefftz formulations often involve only face or boundary integrals and that plane-wave integrals can be evaluated in closed form on straight segments and polygonal faces (Hiptmair et al., 2015). The TCG Helmholtz construction sharpens this point: each local basis function expands into four complex-direction plane waves on each adjacent cell, and since the required cell and boundary integrals have closed-form expressions, all entries of the stiffness, mass, and boundary-term matrices can be computed exactly without quadrature for polygonal cells (Galante et al., 1 Dec 2025).

Conditioning remains a central issue. The Helmholtz survey records that plane-wave bases become nearly linearly dependent when $\mathcal L$7 is small or directions cluster, and reviews remedies such as limiting the number of plane waves per element, local orthonormalization by Gram–Schmidt or SVD, oversampling or least-squares formulations, hybrid bases, and directional adaptivity (Hiptmair et al., 2015). Quasi-Trefftz spaces inherit part of this challenge, but several constructions report favorable coefficient control. In the TCG Helmholtz method, edge and node expansion coefficients remain uniformly bounded in $\mathcal L$8 as $\mathcal L$9 at fixed $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$00, grow at most polynomially in $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$01 at fixed frequency, and remain modest numerically in the high-frequency regime when $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$02 (Galante et al., 1 Dec 2025).

The numerical record across applications is consistent. The diffusion-advection-reaction method shows $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$03-errors of order $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$04, $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$05-errors of order $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$06, and exponential $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$07-convergence for smooth tests while using a local dimension $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$08 in two dimensions rather than $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$09 (Perinati, 2023). The embedded Trefftz DG method on anisotropic quadrilateral meshes reproduces the convergence rates of the full DG method while reducing the size of the global system, with reported tests showing roughly half as many global unknowns or up to $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$10 fewer global unknowns for comparable accuracy (Gómez et al., 2 Jun 2026). The TCG Helmholtz method reports exponential convergence for analytic targets and robust high-frequency behavior, including $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$11–$$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$12 digits on complex polygonal domains with only a few DOFs per wavelength in least-squares Petrov-Galerkin solves (Galante et al., 1 Dec 2025).

Current research directions follow directly from these patterns. The 2015 Helmholtz survey identifies $$T(\mathcal T_h):=\{v\in H^1_{\rm pw}(\Omega):(-\Delta-k^2)v=0 \text{ pointwise in each }K\in\mathcal T_h\},$$13-robust discretizations, variable media, complex geometries, adaptive directional refinement, and improved preconditioning as central open problems for Trefftz-type methods (Hiptmair et al., 2015). Subsequent quasi-Trefftz work has already moved into general scalar operators, space-time hyperbolic problems, first-order systems, Maxwell equations, generalized plane waves, anisotropic meshes, and globally conforming Helmholtz constructions (Imbert-Gerard, 24 May 2025, Imbert-Gérard, 29 Aug 2025, Gómez et al., 2 Jun 2026, Galante et al., 1 Dec 2025). A plausible implication is that “discrete quasi-Trefftz spaces” now designate not a single method class, but a unifying local approximation principle: encode the PDE directly into the finite-dimensional space by annihilating low-order residual structure, then exploit that space in DG, embedded, least-squares, or conforming global couplings.