Basis Update & Galerkin Methods

• Basis update and Galerkin methods are advanced projection techniques that adjust the approximation space to retain stability and computational properties.
• They utilize mechanisms such as state-space transformations, correction steps, recombined bases, and test-space modifications to enhance numerical modeling.
• These methods are applied in stochastic Galerkin settings, finite element recombination, and dynamical low-rank approximations to improve accuracy and efficiency.

Basis update and Galerkin method denote a family of projection procedures in which the approximation space, the state variables, or the test space is altered so that the projected problem retains computational or structural properties of the original one. In the literature considered here, this alteration appears as a parameter-dependent state-space transformation for stochastic Galerkin ODEs, a correction from an auxiliary superspace, a recombined spectral or finite-element basis, a Petrov–Galerkin test-space modification, an adaptive reduced basis, or an augmentation–truncation cycle in dynamical low-rank time integration (Pulch et al., 2017, Podvigina, 2023, Parish et al., 2018, Einkemmer et al., 2023).

1. Core projected formulations

The common starting point is a Galerkin projection onto finite-dimensional trial and test spaces. In one standard formulation, the unknown $v$ belongs to a finite-dimensional space $V$ and is defined by

$P(Av-f)=0, \qquad v\in V,$

with $P$ the orthogonal projection onto $V$. In stochastic Galerkin settings, the parameter dependence is expanded in an orthonormal basis $(\Phi_i)$, the state is truncated to $m$ modes, and the coefficient vector satisfies a deterministic enlarged system

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$

In reduced-basis stochastic Galerkin methods, the approximation space is a tensor product $W_{hp}=V_h\otimes S_p$, later compressed by replacing the physical space $V_h$ with a reduced space $V$0 (Podvigina, 2023, Pulch et al., 2017, Wang et al., 2022, Siena et al., 2022).

What changes across these formulations is not the existence of a projection, but the object being updated before or during projection. The update may target the state representation, the basis itself, the admissible stochastic multi-index set, the test space, or the low-rank factors used to parametrize the solution manifold.

Update mode	Mechanism	Representative setting
State-space transformation	$V$1 before projection	Stability preservation
Correction from a larger space	Solve in $V$2, correct into $V$3	Boundary-constrained Galerkin
Recombined basis	Local stencil or derivative-interpolatory redesign	Banded spectral and ADER predictors
Test-space modification	Petrov–Galerkin, SUPG, collocation constraints	Stabilization and closure
Adaptive basis selection	Greedy, ANOVA, reduced basis, DNN coefficients	High-dimensional parametric problems
Basis augmentation and truncation	Enlarged Galerkin step followed by compression	Dynamical low-rank integrators

2. Stability-preserving basis updates in stochastic Galerkin projection

A central stability issue arises when a parameter-dependent linear ODE

$V$4

is asymptotically stable for almost every parameter realization, but its Galerkin matrix $V$5 is not. The instability is especially associated with non-normal matrices: Galerkin projection mixes parameter-dependent modes through the basis functions, and the coupled block matrix can acquire eigenvalues with nonnegative real part even when every $V$6 has $V$7. A sufficient condition remains classical: if the symmetric part $V$8 is negative definite, then $V$9 is stable. The difficulty is that this property need not survive standard stochastic projection (Pulch et al., 2017).

The stability-preserving remedy is a parameter-dependent similarity transformation

$P(Av-f)=0, \qquad v\in V,$0

chosen so that $P(Av-f)=0, \qquad v\in V,$1. For each parameter value, one solves the Lyapunov equation

$P(Av-f)=0, \qquad v\in V,$2

with $P(Av-f)=0, \qquad v\in V,$3. The unique symmetric positive definite solution is factorized as $P(Av-f)=0, \qquad v\in V,$4, and the transformation is set to $P(Av-f)=0, \qquad v\in V,$5. Then

$P(Av-f)=0, \qquad v\in V,$6

Because the transformed symmetric part is negative definite pointwise in parameter space, the projected matrix $P(Av-f)=0, \qquad v\in V,$7 also has negative definite symmetric part, and the Galerkin-projected transformed system is asymptotically stable.

The same logic extends to nonlinear autonomous systems

$P(Av-f)=0, \qquad v\in V,$8

near asymptotically stable stationary solutions $P(Av-f)=0, \qquad v\in V,$9. One shifts the dynamics by the stationary family,

$P$0

so that the equilibrium is moved to the origin, and then applies the Lyapunov-based transformation to the Jacobian

$P$1

The transformed nonlinear system

$P$2

has a Jacobian whose symmetric part is negative definite at the equilibrium, and the stochastic Galerkin projection preserves asymptotic stability of the stationary solution.

The analysis also includes a perturbation criterion for numerically assembled Galerkin matrices: if the quadrature error satisfies

$P$3

the approximate projected matrix remains stable. The numerical examples include a stable $P$4 polynomial matrix whose standard Galerkin projections are unstable for all tested polynomial degrees, and a two-dimensional nonlinear system with equilibrium

$P$5

whose standard projected equilibria are unstable but become asymptotically stable after the basis update.

3. Correction, recombination, and fixed basis design

In another usage, basis update means solving an easier problem in a larger space and then correcting the result to recover the exact Galerkin solution in the original space. For physically realistic boundary conditions, the target problem is

$P$6

with $P$7 built from basis functions satisfying the boundary conditions and therefore often non-orthogonal. If a larger space $P$8 makes the intermediate solve cheaper, one first computes $P$9 from

$V$0

then corrects it by a complement-based update. If $V$1 form an orthonormal basis of $V$2 and each $V$3 with $V$4 and $V$5, then the exact Galerkin solution is

$V$6

The preprocessing is done once, and later solves require only inner products and low-dimensional vector corrections (Podvigina, 2023).

A different line of work redesigns the basis so that the Galerkin matrix is sparse or banded from the outset. For variable-coefficient ODEs with linear constraints, a new Petrov–Galerkin spectral method constructs trial functions as local recombinations of consecutive orthogonal polynomials,

$V$7

with the coefficients $V$8 chosen so that each $V$9 satisfies the homogeneous constraints. The resulting discrete matrix takes the form

$(\Phi_i)$0

and is strictly banded for smooth variable coefficients, allowing linear-complexity construction and solution. The same framework is presented as encompassing Mortensen’s Galerkin method and earlier banded Galerkin spectral methods (Qin et al., 17 Feb 2025).

The ADER-FV literature gives another recombination mechanism through the Montecinos–Balsara polynomial basis. There, the spatial basis simultaneously interpolates values and first derivatives at Gauss points: $(\Phi_i)$1 The corresponding coefficient-space differentiation matrix $(\Phi_i)$2 is nilpotent, since $(\Phi_i)$3, and therefore the Galerkin predictor matrices $(\Phi_i)$4 have only zero eigenvalues. This is the basis for fast convergence of the local nonlinear predictor iteration in multidimensional, non-conservative ADER-FV systems (Jackson, 2017).

Basis redesign may also be geometric rather than algebraic. On polygonal Voronoi meshes, a discontinuous Galerkin method replaces a modal polynomial basis with an agglomerated finite element basis built on a subgrid of triangles inside each polygon. The local space-time ADER predictor is still obtained from a Galerkin formulation, but mass, flux, and stiffness matrices can be precomputed on a reference simplex, enabling a quadrature-free implementation (Boscheri et al., 2022). By contrast, the Gray–Scott SOBGFEM method uses a spectral orthogonal basis in the finite-element space that is constructed once from a self-adjoint positive definite operator and then kept fixed; only the modal coefficients evolve. The paper explicitly states that there is no iterative basis update in time (Ngondiep, 13 Apr 2026).

4. Petrov–Galerkin test-space updates and stabilization

A major branch of basis-update methodology modifies the test space rather than the trial space. For the radial Dirac eigenvalue problem, standard Galerkin finite elements with piecewise linear trial functions produce spurious eigenvalues because the operator is convection-dominated and lacks a stabilizing second-order term. The streamline upwind Petrov–Galerkin construction keeps the trial basis $(\Phi_i)$5 but enriches the test functions to

$(\Phi_i)$6

which introduces diffusion-like terms into the weak form. The stabilization parameter is derived elementwise as

$(\Phi_i)$7

and the resulting generalized eigenvalue problem removes the two kinds of spectrum pollution reported in the standard discretization (Almanasreh, 2017).

The Adjoint Petrov–Galerkin method for nonlinear model reduction updates the test basis dynamically. Starting from a resolved–unresolved decomposition and a Mori–Zwanzig finite-memory approximation, the reduced model becomes

$(\Phi_i)$8

Equivalently, it is a Petrov–Galerkin method with nonlinear, time-varying test basis

$(\Phi_i)$9

Galerkin is recovered when $m$0. The formulation is explicitly related to adjoint stabilization and compared with LSPG, whose test basis is also Jacobian-dependent but fully discrete (Parish et al., 2018).

In multiscale elliptic problems, the discontinuous Petrov–Galerkin update is implemented through a transfer operator

$m$1

which maps oversampling multiscale trial functions to corresponding coarse linear functions. The bilinear form then mixes multiscale trial functions with projected test functions. This construction yields the multiscale discontinuous Petrov–Galerkin method, presented as eliminating resonance error and reducing assembly cost relative to the corresponding multiscale DG formulation (Fei et al., 2017).

A related but distinct test-space modification appears in Galerkin–collocation time discretization for the wave equation. The cGP-C$m$2(k) scheme retains the degree-$m$3 polynomial trial space in time but adds endpoint collocation conditions for $m$4, reduces the test space to $m$5, and uses a Hermite-type quadrature exact up to degree $m$6. The resulting discrete solution is globally $m$7 in time, and a local post-processing produces the $m$8 family (Anselmann et al., 2019).

Stability-oriented basis construction for Dirac equations also appears in a non-Petrov setting through kinetic and atomic balance. In the prolate-spheroidal B-spline Galerkin formulation, the small spinor component is generated from the large component by

$m$9

so that the basis itself encodes the physical large/small-component coupling and suppresses spurious states (Fillion-Gourdeau et al., 2015).

5. Adaptive and reduced bases within Galerkin frameworks

Adaptive basis update is central when the approximation space is too large to be fixed a priori. In adaptive ANOVA stochastic Galerkin methods, the solution is decomposed into ANOVA components $$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$0, and each component is expanded only on the multi-index set supported on $$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$1. The stochastic basis is then updated hierarchically by a variance indicator

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$2

retaining only terms with $$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$3. Higher-order subsets are allowed to enter only when all lower-order subsets are already active and significant. If $$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$4 is sufficiently small, the method becomes equivalent to the standard full stochastic Galerkin method (Wang et al., 2023).

Reduced-basis stochastic Galerkin methods compress the physical space while keeping the stochastic polynomial basis fixed. The reduced space

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$5

is constructed from snapshot solutions, and the Galerkin system is projected by

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$6

A secant method is then used to identify the number of reduced basis functions from the relative residual

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$7

The stated purpose is to reduce the cost of matrix-vector manipulation in the coupled stochastic Galerkin system (Wang et al., 2022).

The POD-Greedy-Galerkin reduced-basis method updates the basis by selecting the parameter with maximum estimated error,

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$8

solving the full model there, and enriching the reduced space with the new snapshot. POD provides an alternative global compression through the eigenproblem for the correlation matrix

$$\dot{\hat v}(t)=\hat A\,\hat v(t), \qquad \hat A_{ij}=\mathbb{E}[A\Phi_i\Phi_j].$$9

The reduced approximation then solves the Galerkin problem only on $W_{hp}=V_h\otimes S_p$0, and the residual-based estimator certifies the error (Siena et al., 2022).

Deep adaptive basis Galerkin methods reinterpret basis update in a nonlinear-parametric way. The temporal dependence is represented spectrally,

$W_{hp}=V_h\otimes S_p$1

while each spatial coefficient $W_{hp}=V_h\otimes S_p$2 is a trainable neural network in a function class $W_{hp}=V_h\otimes S_p$3. The Galerkin projection in time yields a coupled system for the unknown spatial networks, and the a posteriori error is bounded by the minimal loss function together with an $W_{hp}=V_h\otimes S_p$4 term. The method is explicitly introduced for oscillatory high-dimensional evolution equations (Gu et al., 2021).

6. Basis-update and Galerkin integrators in dynamical low-rank approximation

In dynamical low-rank approximation, basis update and Galerkin are combined inside the time integrator itself. For kinetic equations

$W_{hp}=V_h\otimes S_p$5

the low-rank manifold is

$W_{hp}=V_h\otimes S_p$6

with orthonormal basis functions $W_{hp}=V_h\otimes S_p$7, $W_{hp}=V_h\otimes S_p$8, and coefficient matrix $W_{hp}=V_h\otimes S_p$9. The basis-update and Galerkin (BUG) integrator avoids explicit use of $V_h$0 by evolving the compressed factors

$V_h$1

augmenting the basis with the old basis, performing a Galerkin $V_h$2-step in the enlarged space, and then truncating back to low rank. A key structural feature is that the old basis is contained in the enlarged basis, so the initial condition for the $V_h$3-update is represented exactly (Einkemmer et al., 2023).

The conservative augmented BUG extension preserves local conservation laws when two conditions are met: the $V_h$4-step is discretized by forward Euler, and the truncation is conservative. The discrete substeps are

$V_h$5

followed by a Galerkin update of the enlarged coefficient matrix and a conservative truncation. The paper emphasizes that this requires only minor modifications of existing implementations and avoids changing the underlying low-rank evolution equations (Einkemmer et al., 2023).

The same philosophy has been extended from matrices to Tucker tensors and general tree tensor networks. For Tucker tensors, all mode factors $V_h$6 are updated in parallel, and the core is advanced by a Galerkin ODE. For a general tree tensor network, every leaf basis and every connecting tensor is evolved in parallel over a time step, after which the representation is augmented and truncated recursively. The error bound is stated as

$V_h$7

with constants independent of the singular values of matricizations of the connecting tensors. This independence is the central robustness claim, since small singular values are precisely what destabilize conventional low-rank time integrators (Ceruti et al., 2024).

A recurrent misconception is that basis update always means changing a polynomial or finite-element basis in the classical sense. The literature surveyed here shows a broader picture. It may instead mean a Lyapunov-based state transformation, a low-dimensional correction, a Petrov test-space modification, a greedy or ANOVA activation rule, or an augmentation–truncation cycle on a low-rank manifold. The unifying feature is that Galerkin projection is retained, but the space on which it acts is deliberately modified so that the reduced or discretized model inherits stability, sparsity, conservation, or efficiency properties that a fixed projection may fail to preserve.