How Weak Can Boundary Enforcement Be?

弱施加邊界條件，權重究竟可以多弱？

For elliptic FOSLS, the sharp boundary weights are $h_F^{-1}$ on Dirichlet faces and 1 on Neumann faces. Their origin is the stability of the underlying PDE.

對橢圓型 FOSLS，正確的最低邊界權重是 Dirichlet 的 $h_F^{-1}$ 與 Neumann 的 1。這個結果來自原 PDE 的穩定性。

The Boundary-Weight Question 邊界權重問題

First-order system least-squares methods can impose Dirichlet and Neumann boundary conditions through residuals rather than essential constraints on the trial space. The real question is how strongly those residuals must be weighted to control the natural error.

一階系統最小二乘方法可以把 Dirichlet 與 Neumann 邊界條件直接寫成殘差，而不必把邊界數據強制放入試探空間。關鍵問題是：邊界殘差需要多強的權重，才能控制自然的誤差範數？

The trace spaces suggest an apparent pair of weights, $h_F^{-1}$ for Dirichlet data and $h_F$ for Neumann data. The second choice is too weak for mesh-uniform norm equivalence. The correct minimal weights are $h_F^{-1}$ and $1$ .

從自然 trace space 的尺度出發，容易想到 Dirichlet 使用 $h_F^{-1}$ 、Neumann 使用 $h_F$ 。但後者不足以得到 mesh-uniform norm equivalence。正確的最低權重分別是 $h_F^{-1}$ 和 $1$ 。

A First-Order System with Mixed Boundary Data 混合邊界條件的一階系統

Consider a general, possibly indefinite, second-order elliptic boundary-value problem in its flux formulation:

\begin{aligned} \boldsymbol\sigma+A\nabla u+\boldsymbol b\,u+\boldsymbol f_2&=0 &&\text{in }\Omega,\\ \nabla\!\cdot\boldsymbol\sigma+\boldsymbol\beta\cdot\nabla u+\mu u&=f_1 &&\text{in }\Omega,\\ u&=\phi &&\text{on }\Gamma_D,\\ \boldsymbol\sigma\cdot\boldsymbol n&=\psi &&\text{on }\Gamma_N. \end{aligned}

考慮具有一般低階項、標準弱形式可能不定的二階橢圓邊界值問題。引入通量後，兩個體積方程與 Dirichlet–Neumann 混合邊界條件組成上面的一階系統。

With $\phi\in H^{1/2}(\Gamma_D)$ and $\psi\in L^2(\Gamma_N)$ , both boundary residuals can be included directly in a least-squares functional on unconstrained finite element spaces.

對 $\phi\in H^{1/2}(\Gamma_D)$ 及 $\psi\in L^2(\Gamma_N)$ ，可以在不施加 essential boundary constraints 的有限元空間上直接建立最小二乘泛函。

The Least-Squares Functional 最小二乘泛函

The proposed functional measures two volume residuals and two boundary residuals:

\begin{aligned} \mathcal J_h(\boldsymbol\tau,v) ={}&\|A^{-1/2}(\boldsymbol\tau+A\nabla v+\boldsymbol b\,v+\boldsymbol f_2)\|_{0,\Omega}^{2}\\ &+\|\nabla\!\cdot\boldsymbol\tau+\boldsymbol\beta\cdot\nabla v+\mu v-f_1\|_{0,\Omega}^{2}\\ &+\sum_{F\subset\Gamma_D}h_F^{-1}\|v-\phi\|_{0,F}^{2}\\ &+\|\boldsymbol\tau\cdot\boldsymbol n-\psi\|_{0,\Gamma_N}^{2}. \end{aligned}

兩個體積殘差仍然使用標準的 $L^2$ 範數；邊界條件則通過 face-wise $L^2$ 殘差弱施加。Dirichlet 與 Neumann 的權重不是相同的尺度。

Dirichlet boundary Dirichlet 邊界

h_F^{-1}\|v-\phi\|_{0,F}^{2}

An inverse trace inequality controls the $H^{1/2}$ norm of the discrete trace.

以離散 trace 的 inverse inequality 控制 $H^{1/2}$ 範數。

Neumann boundary Neumann 邊界

\|\boldsymbol\tau\cdot\boldsymbol n-\psi\|_{0,\Gamma_N}^{2}

The unweighted $L^2$ norm controls the natural negative-order trace norm.

不加網格權重的 $L^2$ 範數即可控制自然的負階 trace norm。

PDE Stability Gives Coercivity 從 PDE 穩定性得到 Coercivity

The proof starts with the stability estimate of the original boundary-value problem. For the homogeneous residual system, it controls the potential and flux by the volume residuals together with the natural boundary trace norms:

\begin{aligned} \|v\|_{H^{1}(\Omega)}+\|\boldsymbol\tau\|_{H(\mathrm{div};\Omega)}\ \lesssim\ &\|R_1\|_{0,\Omega}+\|R_2\|_{0,\Omega}\\ &+\|v\|_{H^{1/2}(\Gamma_D)} +\|\boldsymbol\tau\cdot\boldsymbol n\|_{H^{-1/2}_{00}(\Gamma_N)}. \end{aligned}

證明直接使用原邊界值問題的 stability estimate：兩個體積殘差與自然的邊界 trace 範數共同控制勢變量及通量變量。對一般 indefinite elliptic problems，穩定性仍可從適當的唯一性假設得到。

The Dirichlet and Neumann traces are handled by two different inequalities:

\begin{aligned} \|g_h\|_{H^{1/2}(\Gamma_D)} &\lesssim\left(\sum_{F\subset\Gamma_D}h_F^{-1}\|g_h\|_{0,F}^{2}\right)^{1/2},\quad g_h\in W_h^D,\\ \|q\|_{H^{-1/2}_{00}(\Gamma_N)}&\lesssim\|q\|_{L^{2}(\Gamma_N)},\quad q\in L^{2}(\Gamma_N). \end{aligned}

Dirichlet 一側使用離散 trace space $W_h^D$ 上的 inverse inequality；Neumann 一側使用連續嵌入，並不需要離散 inverse estimate。這正好給出兩種不同的最低權重。

The coercivity comes from PDE stability. The boundary weights make that stability usable in the least-squares functional.

LSFEM coercivity 來自原 PDE 的穩定性；正確的邊界權重使這個穩定性估計可以直接控制 least-squares functional。

In particular, the homogeneous functional is equivalent to the natural mesh-dependent norm, with constants independent of the mesh, on pairs whose Dirichlet trace belongs to the discrete trace space. The interior scalar and flux fields need not themselves be discrete in this continuous stability argument.

因此，在 Dirichlet trace 屬於離散 trace space 的相應空間上，可以得到 mesh-uniform norm equivalence。這個證明的離散要求只出現在 Dirichlet trace；內部的勢變量和通量不必都是離散函數。

From Norm Equivalence to Error Equivalence 從範數等價到誤差等價

The norm equivalence controls homogeneous pairs. For an approximation to the exact solution, the Dirichlet error trace $\phi-v$ is not generally a discrete trace. The paper separates it into a data-approximation part and a discrete part:

\phi-v=(\phi-Q_h\phi)+(Q_h\phi-v).

真正的誤差 $\phi-v$ 未必屬於離散 trace space，因此不能直接套用 inverse inequality。把它分成純資料逼近誤差及離散 trace 差，才能得到對真解的 error equivalence。

\| (\boldsymbol\sigma-\boldsymbol\tau,u-v)\|_{h} \ \simeq\ \mathcal J_h(\boldsymbol\tau,v)^{1/2} +\|\phi-Q_h\phi\|_{H^{1/2}(\Gamma_D)}.

Here $Q_h$ is a suitable projection onto the Dirichlet finite element trace space, and the equivalence applies to every admissible pair—not just the least-squares minimizer. The extra term depends only on the Dirichlet datum and its approximation.

其中 $Q_h$ 是到 Dirichlet 有限元 trace space 的適當投影。這個等價關係適用於每一個 admissible pair，不要求它是 least-squares minimizer；額外項只依賴 Dirichlet 資料及其逼近。

Sharp Weights and Over-Penalization 最小權重與 Over-Penalization

Write the boundary weights more generally as $h_F^{-\alpha}$ on $\Gamma_D$ and $h_F^{-\beta}$ on $\Gamma_N$ . The paper identifies the exact stability thresholds:

\boxed{\ \alpha\ge 1,\qquad \beta\ge 0\ }

Below either threshold, mesh-uniform norm equivalence fails, already on quasi-uniform meshes. In particular, the tempting Neumann weight $h_F$ corresponds to $\beta=-1$ and is too weak.

只要 $\alpha<1$ 或 $\beta<0$ ，即使在 quasi-uniform meshes 上也無法得到 mesh-uniform norm equivalence。尤其 Neumann 使用 $h_F$ （即 $\beta=-1$ ）是不夠的。

Stronger boundary penalties remain stable. They can also preserve the optimal physical-norm approximation rate: a comparison function meets the projected boundary data, making the over-penalized residual vanish; for strongly over-penalized Dirichlet data the estimate retains a Dirichlet projection-approximation term.

更強的邊界 penalty 仍可保持穩定。在適當的 projected-data comparison 下，被加大的邊界 penalty 乘上的殘差為零，因此仍能得到最佳階 physical-norm 估計；Dirichlet 強加權時，估計中保留相應的邊界投影逼近項。

Why the Optimal Rate Survives 為什麼仍能得到最佳階收斂

The key step for the Neumann boundary is an $L^2$ -projection identity. If $\Pi_k\psi$ denotes the facewise projection of $\psi$ onto the Raviart–Thomas normal-trace space, then for every discrete flux

\|\boldsymbol\tau_h\cdot\boldsymbol n-\psi\|_{0,\Gamma_N}^{2} =\|\boldsymbol\tau_h\cdot\boldsymbol n-\Pi_k\psi\|_{0,\Gamma_N}^{2} +\|\psi-\Pi_k\psi\|_{0,\Gamma_N}^{2}.

The second term is independent of the unknown, so projection does not change the minimizer. Moreover, the Raviart–Thomas comparison flux has normal trace exactly $\Pi_k\psi$ ; its projected Neumann residual is zero.

第二項與離散未知量無關，因此把 Neumann 數據投影後不會改變 minimizer。Raviart–Thomas comparison flux 的法向跡正好是 $\Pi_k\psi$ ，所以 projected Neumann residual 完全為零。

\|u-u_h\|_{H^1(\Omega)}+\|\boldsymbol\sigma-\boldsymbol\sigma_h\|_{H(\mathrm{div};\Omega)} \lesssim h_{\max}^{s}\bigl(|u|_{H^{1+s}}+|\boldsymbol\sigma|_{H^{s}}+|\nabla\!\cdot\boldsymbol\sigma|_{H^{s}}\bigr),\quad 0<s\le k+1.

因此，在論文所要求的 regularity 下，LSFEM 解在自然的 $H(\mathrm{div})\times H^1$ physical norm 中仍具有最佳階 a priori estimate，而無須附加 Neumann data oscillation。

Non-Intrusive Flux Recovery 不侵入原方法的通量恢復

The error equivalence is not tied to a least-squares solution. Start with a scalar approximation $\widehat u_h$ produced by another method—even if it does not satisfy the exact Dirichlet data—and recover a flux by fixing that scalar and minimizing the same functional:

\widehat{\boldsymbol\sigma}_h =\operatorname*{arg\,min}_{\boldsymbol\tau_h\in\Sigma_h} \mathcal J_h(\boldsymbol\tau_h,\widehat u_h).

對任何其他方法得到的 scalar approximation $\widehat u_h$ ，即使其 Dirichlet 邊界值與精確數據不完全一致，也可以固定 scalar、只對 flux 極小化同一泛函。

The recovered flux–potential pair inherits the a posteriori error equivalence. The Dirichlet residual directly records the computable boundary mismatch, and the functional decomposes into local volume and boundary indicators.

恢復出的 flux–potential pair 具有同樣的 a posteriori error equivalence。Dirichlet 殘差直接度量可計算的邊界違反量，而整個 functional 仍可逐單元、逐邊界面局部化。

The Fully Discontinuous Extension 完全非協調的延伸

The same framework extends to fully discontinuous flux and scalar trial spaces. A pure-penalty formulation uses interior jumps of the scalar and the normal flux, both with $h_F^{-1}$ weights:

\sum_{F\in\mathcal E_h^{\mathrm{int}}} h_F^{-1}\left(\|[\![v_h]\!]\|_{0,F}^{2} +\|[\![\boldsymbol\tau_h\cdot\boldsymbol n]\!]\|_{0,F}^{2}\right).

對 fully discontinuous 離散，內部 scalar jump 與 normal-flux jump 都使用 $h_F^{-1}$ 權重，不另外加入 moment constraints。

The argument proceeds from a broken discrete pair to conforming companions and then to the exact solution. It yields error equivalence without requiring the exact flux to have square-integrable normal traces on every interior face.

證明路線是 broken discrete pair → conforming companions → exact solution。由 jump 控制 companion correction，再利用前面的 exact-error equivalence；不要求精確 flux 在每一條 interior face 都具有 $L^2$ normal trace。

Paper 論文

Least-Squares Finite Element Methods with Weakly Imposed Boundary Conditions: Norm and Error Equivalences, Over-Penalization, and Fully Discontinuous Discretizations

Shun Zhang · arXiv:2610.17678 [math.NA] · 2026

arXiv · PDF