Galerkin Approximations in PDE Analysis

• Galerkin approximations are numerical methods that discretize PDEs by projecting them onto finite-dimensional subspaces via weak variational formulations.
• They ensure systematic convergence and efficiently blend finite element, spectral, and collocation techniques with robust error control.
• Recent advancements incorporate higher temporal regularity and effective post-processing, enhancing accuracy for wave propagation and complex simulations.

Galerkin approximations are fundamental methods for the discretization and numerical analysis of partial differential equations (PDEs) and evolution problems. They are based on expanding the exact infinite-dimensional solution in a finite-dimensional subspace, projecting the governing equation onto that subspace via a variational (weak) formulation. The approach achieves systematic convergence as the approximation space is enriched and is compatible with both classical finite element and spectral methods, as well as their time-discretization and hybridizations. Recent research extends the framework to guarantee higher temporal regularity, lower algebraic complexity, and enables efficient post-processing for accuracy enhancement. This article provides a detailed exposition of the mathematical foundation, discrete realization, regularity mechanisms, error theory, post-processing strategies, and representative computational results of Galerkin (and Galerkin–collocation) approximations, especially for hyperbolic and wave propagation problems (Anselmann et al., 2019).

1. Mathematical Formulation and Discretization

Consider the prototypical second-order wave equation on a spatial domain $\Omega\subset\mathbb{R}^d$ ($d=2,3$), for $t\in[0,T]$,

$$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$

The Galerkin approximation first reformulates the problem in evolution-variational form: $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$ where $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$, $H = L^2(\Omega)$, $V = H^1_0(\Omega)$. The finite element space $V_h\subset H^1_0(\Omega)$ has degree $r$ and mesh size $d=2,3$0.

Time discretization: The time interval $d=2,3$1 is partitioned into subintervals $d=2,3$2. On each $d=2,3$3, the trial space is $d=2,3$4 $d=2,3$5-valued polynomials of degree $d=2,3$6. Globally, define:

• Continuous in time: $d=2,3$7,
• Test space: $d=2,3$8.

Standard Galerkin method ("cGP($d=2,3$9)"): Seek $t\in[0,T]$0 with $t\in[0,T]$1 such that, for all $t\in[0,T]$2,

$t\in[0,T]$3

Galerkin–collocation approach (cGP–C$t\in[0,T]$4($t\in[0,T]$5)):

• Reduce the test space to $t\in[0,T]$6 and impose endpoint collocation:
  • Enforce both the residual at the endpoints and variational identity in interior points using a Hermite-type quadrature.
• The scheme enforces $t\in[0,T]$7.

2. Connection to Collocation and Quadrature

A distinctive feature of the Galerkin–collocation approach is its direct link between high-order variational discretizations and classical collocation:

• Hermite quadrature (Q$t\in[0,T]$8): Uses values (and derivatives) at interval endpoints plus function values at $t\in[0,T]$9 interior points determined by Jacobi polynomials; exact on polynomials of degree $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$0.
• Gauss–Lobatto quadrature (Q$$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$1): Uses $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$2 points including endpoints; also exact on degree $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$3 polynomials.

Equivalence: On each interval, the resulting method is equivalent to collocation at the Gauss–Lobatto nodes, with Hermite interpolation of $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$4. The time-continuity and collocation constraints produce discrete solutions that are continuously differentiable (or twice differentiable in the $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$5 variant) in time.

3. Construction of Regularity: $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$6 and $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$7 Discrete Solutions

$$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$8-continuity: The Hermite quadrature in the test formulation encodes both function and derivative matching at time nodes. Because the trial space is $$\partial_{tt} u - \Delta u = f \quad\text{in } \Omega\times(0,T], \qquad u|_{\partial\Omega} = 0, \quad u(\cdot,0) = u_0,\; \partial_t u(\cdot,0)=u_1.$$9 and both the endpoint collocation and Hermite quadrature enforce matching of $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$0 at the nodes, the discrete solution is $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$1 in time.

$$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$2-continuity ("cGP–C$$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$3($$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$4)"): By imposing both first and second time derivatives of the residual to vanish at interval endpoints and using higher-degree Hermite-type quadrature, one constructs discrete solutions $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$5.

4. Error Estimates and Convergence Rates

Assuming sufficient regularity, $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$6, the Galerkin–collocation scheme achieves the following error bounds: $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$7 uniformly for $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$8. The error splits into interpolation and discrete contributions, with discrete stability (energy estimates) guaranteed for the quadrature-induced bilinear forms. Both $$\partial_t U + A U = F, \quad U(0)=(u_0,\,u_1), \qquad A = \begin{pmatrix}0 & -I \ -\Delta & 0\end{pmatrix}.$$9 in time and $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$0-integrated norms satisfy these asymptotic rates (Anselmann et al., 2019).

5. Post-Processing for Higher Accuracy ($U = (u^0, u^1)^T = (u,\ \partial_t u)^T$1-Lifting)

An efficient lifting procedure can post-process the $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$2 time-continuous solution $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$3 into a $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$4 solution of higher temporal accuracy:

• Construct piecewise Hermite-basis polynomials $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$5 of degree $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$6 vanishing at endpoints except for derivative constraints.
• For each subinterval, define $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$7 to match the second-derivative residual, and set

$U = (u^0, u^1)^T = (u,\ \partial_t u)^T$8

• This $U = (u^0, u^1)^T = (u,\ \partial_t u)^T$9 solves the $H = L^2(\Omega)$0 (cGP–C$H = L^2(\Omega)$1($H = L^2(\Omega)$2)) scheme, attaining one additional order in temporal accuracy for negligible additional cost.

6. Computational Results and Applications

Numerical experiments confirm the theoretical convergence rates for both $H = L^2(\Omega)$3 and post-processed $H = L^2(\Omega)$4 schemes. For instance:

• On the unit square, $H = L^2(\Omega)$5, the $H = L^2(\Omega)$6 scheme attains $H = L^2(\Omega)$7 in $H = L^2(\Omega)$8-in-time energy norms, matching theoretical predictions.
• With $H = L^2(\Omega)$9, after post-processing (lifting), $V = H^1_0(\Omega)$0 convergence in $V = H^1_0(\Omega)$1 and energy norms is observed, demonstrating the effectiveness of the post-processing step.

The framework is particularly attractive for wave propagation and coupled multiphysics problems requiring high temporal regularity, such as fluid-structure interaction or poroelasticity (Anselmann et al., 2019).

7. Summary and Significance

Galerkin (and Galerkin–collocation) approximation schemes provide a rigorous variational foundation for high-order discretizations of PDEs, enabling optimal convergence rates, stability, and energy-conserving properties. The collocation-inspired approach reduces computational complexity, achieving collocation efficiency with Galerkin accuracy. Regularity and accuracy can be systematically enhanced via quadrature and post-processing techniques. The resulting schemes are robust and highly accurate for complex, temporally regular problems in scientific computing and engineering (Anselmann et al., 2019).