曲面偏微分方程：参数化有限元方法

曲面偏微分方程：参数化有限元方法

前面介绍的$\mathbf{P}$和$\mathbf{P}\mathbf{d}$，及其对应的同参等值延伸及投影自然延伸，对于曲面上有限元方法的理论分析起着至关重要的作用。

利普希茨参数曲面上的FEM

利普希茨参数化曲面

因为同参映射$\mathbf{P}$的双利普希茨性，存在一个常数$L$，使得
$$\begin{array}{ll}\text{ (64) }& L^{-1}\left|{\mathrm{x}}_1-{\mathrm{x}}_2\right|\le \left|{\tilde{\mathrm{x}}}_1-{\tilde{\mathrm{x}}}_2\right|\le L\left|{\mathrm{x}}_1-{\mathrm{x}}_2\right|,{\tilde{\mathrm{x}}}_i=\mathrm{P}\left({\mathrm{x}}_i\right),i=1,2\end{array}$$

定义$T$周围区域macro patches如下，这里$T$表示的是$\gamma$的一次逼近曲面$\Gamma$的每个分片。
( 65 ) ω T = ∪ { T ′ : T ′ ∩ T ≠ ∅ } , ω ~ T = P ( ω T ) (65) \quad \omega_{T}=\cup\left\{T^{\prime}: T^{\prime} \cap T \neq \emptyset\right\}, \quad \tilde{\omega}_{T}=\mathbf{P}\left(\omega_{T}\right) (65)ωT​=∪{T′:T′∩T​=∅},ω~T​=P(ωT​)

定义形正则化常数，这里$h_T:=|T{|}^{\frac{1}{n}}$，从表达式可以看出来，这个值越大，正则性越差。
$$(66)\sigma :=\underset{\mathcal{T}}{\sup }\underset{T\in \mathcal{T}}{\max }\frac{\mathrm{diam}(T)}{h_T}$$

我们用$\hat{T}$来表示参考单元单纯形。用$T={\mathbf{X}}_T(\hat{T})$来表示逼近曲面的参数化。从(66)可以看得出来下式。这个很重要，要怎么理解呢？$D{\mathbf{X}}_T$就是一个线性变换，它的算子至于映射到$T$上，模和$\mathrm{diam}(T)$同阶，那么$D{\mathbf{X}}_T$就是$h_T$阶的。
$$\begin{array}{ll}\text{ (67) }& h_T|\mathrm{w}|\lesssim \left|D{\mathrm{X}}_T\mathrm{w}\right|\lesssim h_T|\mathrm{w}|,\forall \mathrm{w}\in {\mathbb{R}}^n\end{array}$$

因为投影的双利普希茨性，我们也知道$
D \chi_{T}$和$h_{T}$是同阶的。
$$(68)h_T|\mathrm{w}|\lesssim \left|D{\chi }_T(\mathrm{y})\mathrm{w}\right|\lesssim h_T|\mathrm{w}|\forall \mathrm{w}\in {\mathbb{R}}^n,\mathrm{y}\in \hat{T}$$

我们依然用tilde符号来表示在曲面$\gamma$上，hat符号表示在参数域是（参考单元）上。$\Gamma$的参数化表示为${\mathbf{X}}_T$，$\gamma$的参数化表示为${\chi }_T=\mathbf{P}\circ {\mathbf{X}}_T$，那么在macro patches上，有，
$$\begin{array}{ll}\text{ (69) }& L^{-1}{h}_T\left|{\mathbf{y}}_1-{\mathbf{y}}_2\right|\le \left|{\tilde{\mathbf{x}}}_1-{\tilde{\mathbf{x}}}_2\right|\le L{h}_T\left|{\mathbf{y}}_1-{\mathbf{y}}_2\right|\end{array}$$

各个区域上的函数值定义如下：
$$\begin{array}{ll}\text{ (70) }{\hat{v}}_T(\mathbf{y}):={\tilde{v}}_T\left({\chi }_T(\mathbf{y})\right)\forall \hat{\mathbf{x}}\in \hat{T}& \text{ and }{v}_T(\mathbf{x}):={\tilde{v}}_T(\mathbf{P}(\mathbf{x}))\forall \mathbf{x}\in T\end{array}$$

多边形曲面上的微分几何

同以前的方式，可以定义第一标准型和面积元，
$$\begin{array}{ll}\text{ (71) }& {\mathrm{g}}_T:={\left(D{\mathrm{X}}_T\right)}^{t}D{\mathrm{X}}_T,q_T:=\sqrt{\det {\mathrm{g}}_T},\forall T\in \mathcal{T}\end{array}$$

且满足，
$$\begin{array}{ll}\text{ (72) }& \text{ eigen }\left({\mathrm{g}}_T\right)\approx h_T^2,q_T\approx h_T^n,\forall T\in \mathcal{T}\end{array}$$

因为$D{\mathbf{X}}_T$和$D{\chi }_T$的同阶性，所以稳定化常数其实和网格尺寸是没有关系的，即
$$(73)S_{\chi }\approx 1$$

因为(72)我们知道$q$和$q_{\Gamma }$是同阶的，即
$$C_1\le \frac{q}{q_{\Gamma }}\le C_2$$

回忆曲面梯度和拉普拉斯算子的定义，有，
$$\begin{array}{ll}\text{ (75) }& \nabla \hat{v}=(D\mathrm{X}{)}^t{\nabla }_{\mathrm{T}}v,{\nabla }_{\mathrm{T}}v=(D\mathrm{X}){\mathrm{g}}_{\mathrm{\Gamma }}^{-1}\nabla \hat{v}\end{array}$$
$$\begin{array}{ll}\text{ (76) }& {\Delta }_{\Gamma }v=\frac{1}{q_{\Gamma }}\mathrm{div}\left(q_{\Gamma }{\mathbf{g}}_{\Gamma }^{-1}\nabla \hat{v}\right)\end{array}$$

使用分布积分公式，很容易看到：
$$\begin{array}{rl}{\int }_{\Gamma }{\nabla }_{\Gamma }v\cdot {\nabla }_{\Gamma }w& =\sum _{T\in \mathcal{T}}-{\int }_{T}w{\Delta }_{\Gamma }v+{\int }_{\partial T}w{\nabla }_{\Gamma }v\cdot {\mathbf{\mu }}_T\\ & =\sum _{T\in \mathcal{T}}-{\int }_{T}w{\Delta }_{\Gamma }v+\sum _{S\in \mathcal{S}}{\int }_{S}w\left[{\nabla }_{\Gamma }v\right]\end{array}$$
这里的跳跃，就定义为：
$$(77)\left[{\nabla }_{\Gamma }v\right]:={\nabla }_{\Gamma }{v}_{+}\cdot {\mu }_{+}+{\nabla }_{\Gamma }{v}_{-}\cdot {\mu }_{-}$$

参数化有限元方法

所谓的参数化有限元方法就是，寻找$U:=U_{\mathcal{T}}\in {\mathbb{V}}_{\#}(\mathcal{T})$，使得，
$$\begin{array}{ll}\text{ (78) }& {\int }_{\Gamma }{\nabla }_{\Gamma }U\cdot {\nabla }_{\Gamma }V={\int }_{\Gamma }FV\forall V\in {\mathbb{V}}_{\#}(\mathcal{T})\end{array}$$
这里的$F\in L_{2,\#}(\Gamma )$。这里定义了分片线性连续空间，及其对应的零均值空间：
$$\mathbb{V}(\mathcal{T}):=\left\{V\in C^0(\Gamma )|V{|}_T=\hat{V}\circ {\mathrm{X}}^{-1}\text{ for some }\hat{V}\in \mathcal{P},T\in \mathcal{T}\right\}$$
$${\mathbb{V}}_{\#}(\mathcal{T}):=\mathbb{V}(\mathcal{T})\cap L_{2,\#}(\Gamma )$$
$\mathcal{P}$是线性多项式空间。

因为$F\in L_{2,\#}(\Gamma )$，那么，我们也有：
$$\text{ (79) }{\int }_{\Gamma }{\nabla }_{\Gamma }U\cdot {\nabla }_{\Gamma }V={\int }_{\Gamma }FV\forall V\in \mathbb{V}(\mathcal{T})$$

介绍一个引理：

证明：用一下误差矩阵的定义以及(78)，再用一下$\gamma$上的弱形式，并放到$\Gamma$上，我们有，
$${\int }_{\Gamma }{\nabla }_{\Gamma }(u-U)\cdot {\nabla }_{\Gamma }V={\int }_{\gamma }{\nabla }_{\gamma }\tilde{u}\cdot {\nabla }_{\gamma }\tilde{V}+{\int }_{\Gamma }{\nabla }_{\Gamma }u\cdot {\mathbf{E}}_{\Gamma }{\nabla }_{\Gamma }V-{\int }_{\Gamma }FV$$

几何一致性

$\Gamma$上的一致庞加莱估计

回忆一下庞加莱不等式：
$$\begin{array}{ll}\text{ (82) }& \Vert v{\Vert }_{L_2(\Gamma )}\lesssim \Vert \nabla v{\Vert }_{L_2(\Gamma )}\forall v\in H_{\#}^1(\Gamma )\end{array}$$

几何估计量

定义一个几何元指示子，反应的是曲面及其逼近的导数之间的距离：
$$\text{ (83) }{\lambda }_T:={\Vert D\left(\mathbf{P}-{\mathcal{I}}_{\mathcal{T}}\mathbf{P}\right)\Vert }_{L_{\infty }(T)}=\Vert D\mathbf{P}-\mathbf{I}{\Vert }_{L_{\infty }(T)}\forall T\in \mathcal{T}$$

相应的几何估计量就定义为：
$$\begin{array}{ll}\text{ (84) }& {\lambda }_{\mathcal{T}}(\Gamma ):={\max }_{T\in \mathcal{T}}{\lambda }_T\end{array}$$

用一下复合函数求导，我们有：
( 85 ) max ⁡ y ∈ T ˉ ∣ D ( χ T − X T ) ( y ) ∣ min ⁡ { ∣ D − χ T ( y ) ∣ , ∣ D − X T ( y ) ∣ } ≤ S χ λ T ∀ T ∈ T (85) \quad \max _{\mathbf{y} \in \bar{T}} \frac{\left|D\left(\chi_{T}-\mathbf{X}_{T}\right)(\mathbf{y})\right|}{\min \left\{\left|D^{-} \chi_{T}(\mathbf{y})\right|,\left|D^{-} \mathbf{X}_{T}(\mathbf{y})\right|\right\}} \leq S_{\chi} \lambda_{T} \quad \forall T \in \mathcal{T} (85)y∈Tˉmax​min{∣D−χT​(y)∣,∣D−XT​(y)∣}∣D(χT​−XT​)(y)∣​≤Sχ​λT​∀T∈T

另外定义两个量，他们是$h_T$的高阶无穷小量，
$$\begin{array}{ll}\text{ (86) }& {\beta }_T:={\Vert \mathbf{P}-{\mathcal{I}}_{\mathcal{T}}\mathbf{P}\Vert }_{L_{\infty }(T)},{\beta }_{\mathcal{T}}(\Gamma ):={\max }_{T\in \mathcal{T}}{\beta }_T\end{array}$$
$$(87){\mu }_T:={\beta }_T+{\lambda }_T^2,{\mu }_{\mathcal{T}}(\Gamma ):=\underset{T\in \mathcal{T}}{\max }{\mu }_T$$

C1曲面的几何一致性误差

给一个推论，
$$\begin{array}{l}\text{ Corollary }32\text{ (geometric consistency errors for }{C}^{1,\alpha }\text{ surfaces). If }\mathrm{X}\text{ and }\chi \text{ satisfy }\\ (67)\text{ and }(68),\text{ then for all }T\in \mathcal{T}\text{ we have }\\ \begin{array}{ll}\text{ (88) }{\Vert 1-q^{-1}{q}_{\Gamma }\Vert }_{L_{\infty }(\hat{T})},{\Vert \mathbf{I}-{\mathrm{grg}}^{-1}\Vert }_{L_{\infty }(\hat{T})},{\Vert {\nu }_{\Gamma }-\nu \Vert }_{L_{\infty }(T)}\lesssim {\lambda }_T& \\ \text{ where the hidden constants depend on }{S}_{\chi }\approx 1\text{ defined in }(38).\text{ Moreover, }& \\ \text{ (89) }& \Vert \mathbf{E}{\Vert }_{L_{\infty }(\hat{T})}+{\Vert {\mathbf{E}}_{\Gamma }\Vert }_{L_{\infty }(\hat{T})}\lesssim {\lambda }_T\forall T\in \mathcal{T}\end{array}\end{array}$$
这个证明和自然而然。用一下引理16和24。(88)的表达通通小于${\lambda }_{\infty }$，用(85)放一下即可。$E$的估计用(47)。

C2曲面的几何一致性误差

首先要让曲面和逼近曲面足够近，使得，
$$\begin{array}{ll}\text{ (90) }& {\beta }_{\mathcal{T}}(\Gamma )<\frac{1}{2K_{\infty }}\Rightarrow \Gamma \subset \mathcal{N}\end{array}$$
其次，要让两个不同的投影走一个来回还是属于一个macro patches。
$$(91){\mathrm{P}}_d\circ {\mathrm{P}}^{-1}(\tilde{T})\subset {\tilde{\omega }}_T\forall T\in \mathcal{T}$$
那么走一个来回之后，曲面上的两点的距离满足，
$$(92)\left|\tilde{\mathbf{x}}-{\mathbf{P}}_d\circ {\mathbf{P}}^{-1}(\tilde{\mathbf{x}})\right|=\left|\mathbf{P}(\mathbf{x})-{\mathbf{P}}_d(\mathbf{x})\right|\le 2{\beta }_T\forall \mathbf{x}\in T$$

那么，有这么一个推论，
$$\begin{array}{l}\text{ Corollary }33\text{ (geometric consistency errors for }{C}^2\text{ surfaces). If }(90)\text{ and }(74)\\ \text{ hold, then so do the following estimates for all }T\in \mathcal{T}\\ \text{ (93) }\Vert d{\Vert }_{L_{\infty }(T)}\lesssim {\beta }_T,{\Vert \nu -{\nu }_{\Gamma }\Vert }_{L_{\infty }(T)}\lesssim {\lambda }_T,{\Vert 1-q^{-1}{q}_{\Gamma }\Vert }_{L_{\infty }(T)}\lesssim {\mu }_T\\ \text{ where all the geometric quantities are defined using the parametrizations }\chi ={\mathbf{P}}_d\circ \mathbf{X}\\ \text{ and }\mathbf{X}.\text{ Moreover, }\\ \begin{array}{llll}\text{ (94) }& \Vert \mathbf{E}{\Vert }_{L_{\infty }(T)},{\Vert {\mathbf{E}}_{\Gamma }\Vert }_{L_{\infty }(T)}& \lesssim {\mu }_T& \forall T\in \mathcal{T}\end{array}\end{array}$$
证明只需用到C2曲面的相应估计式，以及推论已有结果即可。
由上可以得到一些有用的结果，利用$\tilde{\omega }$的利普希茨性质，可以有，
$$\left|\tilde{w}(\tilde{\mathbf{x}})-\tilde{w}\left({\mathbf{P}}_d\circ {\mathbf{P}}^{-1}(\tilde{\mathbf{x}})\right)\right|\le {\Vert {\nabla }_{\gamma }\tilde{w}\Vert }_{L_{\infty }\left({\tilde{\omega }}_T\right)}\left|\tilde{\mathbf{x}}-{\mathbf{P}}_d\circ {\mathbf{P}}^{-1}(\tilde{\mathbf{x}})\right|\le 2{\Vert {\nabla }_{\gamma }\tilde{w}\Vert }_{L_{\infty }\left({\tilde{\omega }}_T\right)}{\beta }_T$$

下面介绍一个命题，
$$\begin{array}{l}\text{ Proposition 34 (mismatch between }\mathbf{P}\text{ and }{\mathbf{P}}_d\text{ ). Assume that (67) as well as the }\\ \text{ assumptions (74), (90) and (91) hold. Then there exists }{\lambda }_{\ast }>0\text{ such for }\tilde{w}\in H^1(\gamma )\\ \text{ and }T\in \mathcal{T}\text{ we have }\\ {\Vert \tilde{w}-\tilde{w}\circ {\mathbf{P}}_d\circ {\mathbf{P}}^{-1}\Vert }_{L_2(\tilde{T})}\lesssim {\beta }_T\Vert \tilde{w}{\Vert }_{H^1\left({\tilde{\omega }}_T\right)}\\ \text{ provided }{\lambda }_T\le {\lambda }_{\ast }\text{ and }{\tilde{\omega }}_T\text{ is a patch in }\gamma \text{ around }\tilde{T}.\end{array}$$
这个命题的证明分为3步：

• reduce到${\mathbb{R}}^n$上
• 光滑化
• 估计各项
• 找$ϵ$的界

抓关键步骤，简言之，就是先把要证的东西做个参数域的转换，
$$\Vert \tilde{w}-\tilde{w}\circ \psi {\Vert }_{L_2(\tilde{T})}\lesssim h_T^{n/2}\Vert \hat{w}-\hat{w}\circ \hat{\psi }{\Vert }_{L_2(\hat{T})}$$

再把转换出来的东西拆成几部分：
$$\Vert \hat{w}-\hat{w}\circ \hat{\psi }{\Vert }_{L_2(\hat{T})}\lesssim {\Vert \hat{w}-{\hat{w}}_{\epsilon }\Vert }_{L_2(\hat{T})}+{\Vert {\hat{w}}_{\epsilon }-{\hat{w}}_{\epsilon }\circ \hat{\psi }\Vert }_{L_2(\hat{T})}+{\Vert {\hat{w}}_{\epsilon }\circ \hat{\psi }-\hat{w}\circ \hat{\psi }\Vert }_{L_2(\hat{T})}$$

估计第一部分……
$${\Vert \hat{w}-{\hat{w}}_{\epsilon }\Vert }_{L_2(\hat{T})}\lesssim \epsilon |\hat{w}{|}_{H^1\left({\mathbb{R}}^n\right)}\lesssim \epsilon |\hat{w}{|}_{H^1\left({\hat{\omega }}_T\right)}$$

估计第三部分……
$${\Vert \left({\hat{w}}_{\epsilon }-\hat{w}\right)\circ \hat{\psi }\Vert }_{L_2(\hat{T})}\lesssim {\Vert {\hat{w}}_{\epsilon }-\hat{w}\Vert }_{L_2\left({\hat{\omega }}_T\right)}\lesssim \epsilon |\hat{w}{|}_{H^1\left({\hat{\omega }}_T\right)}$$

估计第二部分……
$$\begin{array}{rl}{\Vert {\hat{w}}_{\epsilon }-{\hat{w}}_{\epsilon }\circ \hat{\psi }\Vert }_{L_2(\hat{T})}^2& \lesssim {\epsilon }^n\sum _i{\Vert {\hat{w}}_{\epsilon }-{\hat{w}}_{\epsilon }\circ \hat{\psi }\Vert }_{L_{\infty }\left(B\left({\mathbf{y}}_i,\epsilon \right)\cap \hat{T}\right)}\\ & \lesssim {\epsilon }^2\sum _i|\hat{w}{|}_{H^1}^2\left(B\left({\mathbf{y}}_i,3\epsilon \right)\right)\lesssim {\epsilon }^2|\hat{w}{|}_{H^1\left({\mathbb{R}}^n\right)}^2\lesssim {\epsilon }^2|\hat{w}{|}_{H^1\left({\hat{\omega }}_T\right)}^2\end{array}$$

估计$ϵ$……
$$\epsilon \le 2L{h}_T^{-1}{\beta }_T$$

把所有东西合起来，完事。
$$\begin{array}{rl}\Vert \tilde{w}-\tilde{w}\circ \psi {\Vert }_{L_2(\tilde{T})}^2& \lesssim h_T^n\Vert \hat{w}-\hat{w}\circ \hat{\psi }{\Vert }_{L_2(\hat{T})}^2\lesssim h_T^n{\epsilon }^2|\hat{w}{|}_{H^1\left({\hat{\omega }}_T\right)}^2\\ & \lesssim h_T^n{h}_T^{-2}{\beta }_T^2{h}_T^{2-n}|\tilde{w}{|}_{H^1\left({\tilde{w}}_T\right)}^2={\beta }_T^2|\tilde{w}{|}_{H^1\left({\tilde{\omega }}_T\right)}^2\end{array}$$

下面有另外一个命题，
$$\begin{array}{l}\text{ Proposition 35 (Lipschitz perturbation). Let }{\Omega }_1,{\Omega }_2\subset \subset \Omega \subset {\mathbb{R}}^{n+1}\text{ be Lipschitz}\\ \text{ bounded domains and }\mathbf{L}:{\Omega }_1\to {\Omega }_2\text{ be a bi-Lipschitz isomorphism. If }\\ r:={\max }_{\mathbf{x}\in {\Omega }_1}|\mathbf{L}(\mathbf{x})-\mathbf{x}|\\ \text{ is sufficiently small so that }\left({\Omega }_1\cup {\Omega }_2\right)+B(0,r)\subset \Omega \text{ then for all }g\in H^1(\Omega )\text{ we have }\\ \Vert g-g\circ \mathbf{L}{\Vert }_{L^2\left({\Omega }_1\right)}\lesssim r\Vert g{\Vert }_{H^1(\Omega )}\end{array}$$

先验误差分析

C2曲面的先验误差估计

$$\begin{array}{l}\text{ Lemma }36\text{ (approximability in }{H}^1(\Gamma )).\text{ Let }\gamma \text{ be a surface of class }{C}^2\text{ and }\tilde{u}\in \\ H^2(\gamma ).\text{ Let }{K}_{\infty }\text{ be defined in }(30)\text{ and }{\beta }_{\mathcal{T}}(\Gamma )\text{ be given in }(86).\text{ Then we have }\\ \begin{array}{lll}\text{ (95) }& {\inf }_{V\in \mathrm{V}(\mathcal{T})}{\Vert {\nabla }_{\Gamma }\left(\tilde{u}\circ {\mathbf{P}}_d-V\right)\Vert }_{L_2(\Gamma )}& \lesssim h_{\mathcal{T}}|\tilde{u}{|}_{H^2(\gamma )}+{\beta }_{\mathcal{T}}(\Gamma )K_{\infty }{\Vert {\nabla }_{\gamma }\tilde{u}\Vert }_{L_2(\gamma )}\end{array}\end{array}$$
这个引理告诉我们的是，解在$H^1$空间中逼近的下界是可控的。
它的证明用到整体下界小等于分片下界和：
$$(96)\underset{V\in \mathbf{V}(\mathcal{T})}{\inf }{\Vert {\nabla }_{\Gamma }\left(\tilde{u}\circ {\mathbf{P}}_d-V\right)\Vert }_{L_2(\Gamma )}^2\lesssim \sum _{T\in \mathcal{T}}\underset{V_T\in \mathbf{V}(T)}{\inf }{\Vert {\nabla }_{\Gamma }\left(\tilde{u}\circ {\mathbf{P}}_d-V_T\right)\Vert }_{L_2(T)}^2$$

下面是$H^1$先验误差估计：
$$\begin{array}{l}\text{ Theorem }37(H^1\text{ a-priori error estimate for }{C}^2\text{ surfaces). Let }\gamma \text{ be of class }{C}^2\\ \tilde{f}\in L_{2,\#}(\gamma )\text{ and }\tilde{u}\in H^2(\gamma )\text{ be the solution of }(18).\text{ Let }U\in {\mathbb{V}}_{\#}(\mathcal{T})\text{ be the solution }\\ \text{ to }(78)\text{ with }F=\tilde{f}\circ \mathbf{P}\frac{q}{q_{\mathrm{r}}}\text{ defined via the lift }\mathbf{P}.\text{ If the geometric assumptions }(69),\\ \text{ (90), and (91) are valid, then }\\ {\Vert {\nabla }_{\Gamma }(\tilde{u}\circ \mathbf{P}-U)\Vert }_{L_2(\Gamma )}\lesssim (h_{\mathcal{T}}+\lambda \tau (\Gamma ))\Vert \tilde{f}{\Vert }_{L_2(\gamma )}\lesssim h\tau \Vert \tilde{f}{\Vert }_{L_2(\gamma )}\\ \text{ as well as }\\ {\Vert {\nabla }_{\Gamma }\left(\tilde{u}\circ {\mathbf{P}}_d-U\right)\Vert }_{L_2(\Gamma )}\lesssim \left(h_{\mathcal{T}}+{\mu }_{\mathcal{T}}(\Gamma )\right)\Vert \tilde{f}{\Vert }_{L_2(\gamma )}\lesssim h_{\mathcal{T}}\Vert \tilde{f}{\Vert }_{L_{2(\gamma )}}\end{array}$$
这个估计表明了参数化方法在$H^1$空间中的一阶收敛性。

在$L^2$空间中，我们也有相应的估计：
$$\begin{array}{l}\text{ Theorem 38 ( }{L}_2\text{ a-priori error estimate for }{C}^2\text{ surfaces). Let }\gamma \text{ be of class }{C}^2\text{ and }\\ \text{ be described by a generic lift }\mathbf{P}\text{ of class }{C}^2.\text{ Let the geometric conditions }(69),\text{ , }\\ \text{ and }(91)\text{ be satisfied. Let }\tilde{u}\in H_{\#}^1(\gamma )\text{ solve }(19)\text{ and }U\in {\mathbb{V}}_{\#}(\mathcal{T})\text{ solve }(78)\text{ with }\\ F=\tilde{f}\circ \mathbf{P}\frac{q}{q_{\mathrm{T}}}.\text{ Then }\\ \begin{array}{ll}\text{ (97) }& \Vert \tilde{u}\circ \mathbf{P}-U{\Vert }_{L_2(\Gamma )}\lesssim h_{\mathcal{T}}^2\Vert \tilde{f}{\Vert }_{L_2(\gamma )}\end{array}\end{array}$$
它的证明，其实就是用到所谓的对偶技巧。
构造辅助问题：
$${\int }_{\gamma }{\nabla }_{\gamma }\tilde{z}\cdot {\nabla }_{\gamma }\tilde{w}={\int }_{\gamma }\left(\tilde{u}-{\tilde{U}}_{\#}\right)\tilde{w}\forall \tilde{w}\in H_{\#}^1(\gamma )$$
它的有限元逼近是：
$${\int }_{\Gamma }{\nabla }_{\Gamma }Z\cdot {\nabla }_{\Gamma }W={\int }_{\Gamma }\left(u_{\#}-U\right)W,\forall W\in \mathbb{V}(\mathcal{T})$$
然后想办法将精确解和数值解的零中值化的差值表达成如下：
$$\begin{array}{rl}{\Vert \tilde{u}-{\tilde{U}}_{\#}\Vert }_{L_2(\gamma )}^2& ={\int }_{\gamma }{\nabla }_{\gamma }(\tilde{u}-\tilde{U})\cdot {\nabla }_{\gamma }(\tilde{z}-\tilde{Z})\\ & +{\int }_{\gamma }\tilde{f}\left(Z\circ {\mathbf{P}}_d^{-1}-Z\circ {\mathbf{P}}^{-1}\right)\\ & +{\int }_{\gamma }{\nabla }_{\gamma }\tilde{U}\cdot \mathbf{E}{\nabla }_{\gamma }\tilde{Z}\end{array}$$
之后，分别估计每一项的界，就OK了。
上面的结论是对于$C^2$曲面的，我们相信，对于$C^3$曲面，取$\mathbf{P}={\mathbf{P}}_d$，我们有更好的结论如下：
$$\begin{array}{ll}\text{ (99) }& \Vert \tilde{u}\circ \mathbf{P}-U{\Vert }_{L_2(\Gamma )}\lesssim h_{\mathcal{T}}^2|d{|}_{W_{\infty }^2(\mathcal{N})}\Vert \tilde{f}{\Vert }_{L_2(\gamma )}\end{array}$$

C1曲面的先验误差估计

对于C1曲面，我们同样有$H^1$空间中的逼近性质：
$$\begin{array}{l}\text{ Lemma }39\text{ (approximability in }{H}^1(\Gamma )).\text{ Let }\gamma \text{ be a surface of class }{C}^{1,\alpha }\text{ and }\\ \tilde{u}\in H^{1+s}(\gamma ),\text{ where }0<s<\alpha <1\text{ or }0<s\le \alpha =1.\text{ Then we have }\\ \begin{array}{ll}\text{ (100) }& {\inf }_{V\in \mathbb{V}(\mathcal{T})}{\Vert {\nabla }_{\Gamma }(\tilde{u}\circ \mathbf{P}-V)\Vert }_{L_2(\Gamma )}\lesssim h_{\mathcal{T}}^s|\tilde{u}{|}_{H^{1+s}(\gamma )}\end{array}\end{array}$$

那么C1曲面的$H^1$先验误差估计就变成了：
$$\begin{array}{l}\text{ Theorem }40(H^1\text{ a-priori error estimate for }{C}^{1,\alpha }\text{ surfaces). Let }\gamma \text{ be of class }{C}^{1,\alpha },\\ 0<\alpha \le 1,\text{ and assume that the geometric assumptions }(69),(90),\text{ and }(91)\text{ are }\\ \text{ valid. Let }\hat{f}\in L_{2,\#}(\gamma )\text{ and }\tilde{u}\in H^{1+s}(\gamma )\text{ be the solution of }(18)\text{ and satisfy }\end{array}$$
$$\begin{array}{l}\Vert \tilde{u}{\Vert }_{H^{1+s}(\gamma )}\lesssim \Vert \tilde{f}{\Vert }_{L_2(\gamma )}\\ \text{ provided }0<s<\alpha <1\text{ or }0<s\le \alpha =1.\text{ If }U\in {\mathbb{V}}_{\#}(\mathcal{T})\text{ is the solution to }(78)\\ \text{ with }F=\tilde{f}\circ \mathbf{P}\frac{q}{q_{\mathrm{r}}}\text{ defined via the lift }\mathbf{P},\text{ then }\\ {\Vert {\nabla }_{\Gamma }(\tilde{u}\circ \mathbf{P}-U)\Vert }_{L_2(\Gamma )}\lesssim h_{\mathcal{T}}^s\Vert \tilde{u}{\Vert }_{H^{1+s}(\gamma )}+{\lambda }_{\mathcal{T}}(\Gamma )\Vert \tilde{f}{\Vert }_{L_2(\gamma )}\lesssim h_{\mathcal{T}}^s\Vert \tilde{f}{\Vert }_{L_2(\gamma )}\end{array}$$

后验误差分析

后验误差估计依赖于数值解$U$和数据，但是不用到精确解$\tilde{u}$。

要说明后验误差估计，我们可以先了解一下scott-zhang插值：
$${\mathcal{I}}_{\mathcal{T}}^{\mathrm{s}\mathrm{z}}:H^1(\Gamma )\to \mathbb{V}(\mathcal{T})$$
及其性质，
$$(101){\Vert v-{\mathcal{I}}_{\mathcal{T}}^{\mathrm{s}\mathrm{z}}v\Vert }_{L^2(T)}\lesssim h_T{\Vert {\nabla }_{\Gamma }v\Vert }_{L^2\left({\omega }_T\right)},{\Vert {\nabla }_{\Gamma }{\mathcal{I}}_{\mathcal{T}}^{\mathrm{s}\mathrm{z}}v\Vert }_{L^2(T)}\lesssim {\Vert {\nabla }_{\Gamma }v\Vert }_{L^2\left({\omega }_T\right)}$$

我们还需要定义两个函数量“内部”和“跳跃残差”：
$$\begin{array}{rl}{R}_T(V)& :={F|}_T+{{\Delta }_{\Gamma }V|}_T\forall T\in \mathcal{T}\\ J_S(V)& :={{\nabla }_{\Gamma }{V}^{+}|}_S\cdot {\mathbf{\mu }}_S^{+}+{{\nabla }_{\Gamma }{V}^{-}|}_S\cdot {\mathbf{\mu }}_S^{-}\forall S\in S_{\mathcal{T}}\end{array}$$
以及元指示子和误差估计子：
$${\eta }_{\mathcal{T}}(V,T{)}^2:=h_T^2{\Vert R_T(V)\Vert }_{L^2(T)}^2+h_T{\Vert J_{\partial T}(V)\Vert }_{L^2(\partial T)}^2\forall T\in \mathcal{T}$$
$${\eta }_{\mathcal{T}}(V{)}^2:=\sum _{T\in \mathcal{T}}\eta \mathcal{T}(V,T{)}^2$$

那么，我们首先有C1曲面的后验误差上界（H1空间）：
$$\begin{array}{l}\text{ Theorem }41\text{ (a-posteriori upper bound for }{C}^{1,\alpha }\text{ surfaces). Let }\gamma \text{ be of class }{C}^{1,\alpha },\\ \text{ be parametrized by }\chi =\mathbf{P}\circ \mathbf{X}\text{ and satisfy the geometric assumption (69). Let }\\ \tilde{u}\in H^1(\gamma )\text{ be the solution to }(18)\text{ and }U\in {\mathbb{V}}_{\#}(\mathcal{T})\text{ be the solution to (78) with }\\ F=\tilde{f}\circ \mathbf{P}\frac{q}{q_{\mathrm{r}}}\in L_{2,\#}(\Gamma ).\text{ Then, for }\tilde{U}:=U\circ {\mathbf{P}}^{-1}:\gamma \to \mathbb{R}\text{ we have }\\ {\Vert {\nabla }_{\gamma }(\tilde{u}-\tilde{U})\Vert }_{L^2(\gamma )}^2\lesssim {\eta }_{\mathcal{T}}(U{)}^2+{\lambda }_{\mathcal{T}}^2(\Gamma )\Vert \tilde{f}{\Vert }_{L_2(\gamma )}^2\end{array}$$
证明使用一个非常常用的套路，将误差所构成的双线性型拆成三部分：
$$(102){\int }_{\gamma }{\nabla }_{\gamma }(\tilde{u}-\tilde{U})\cdot {\nabla }_{\gamma }\tilde{v}=I_1+I_2+I_3$$
$$\begin{array}{rl}{I}_1& =-{\int }_{\Gamma }{\nabla }_{\Gamma }U\cdot {\nabla }_{\Gamma }(v-V)+{\int }_{\Gamma }F(v-V)\\ I_2& ={\int }_{\gamma }{\nabla }_{\gamma }\tilde{U}\cdot \mathbf{E}{\nabla }_{\gamma }\tilde{v}\\ I_3& ={\int }_{\gamma }\tilde{f}\tilde{v}-{\int }_{\Gamma }Fv\end{array}$$
易知，
$$(103)I_1=\sum _{T\in \mathcal{T}}{\int }_T{R}_T(U)(v-V)+\sum _{S\in \mathcal{S}}{\int }_S{J}_S(U)(v-V)$$
$$(104)I_1\lesssim {\eta }_{\mathcal{T}}(U){\Vert {\nabla }_{\Gamma }v\Vert }_{L^2(\Gamma )}\lesssim {\eta }_{\mathcal{T}}(U){\Vert {\nabla }_{\gamma }\tilde{v}\Vert }_{L^2(\gamma )}$$

之后分别估计每一部分的界即可。

对于C1下界，有如下估计：
$$\begin{array}{l}\text{ Theorem }42\text{ (a-posteriori lower bound for }{C}^{1,\alpha }\text{ surfaces). Under the same condi- }\\ \text{ tions of Theorem }41\text{ (a-posteriori upper bound for }{C}^{1,\alpha }\text{ surfaces), we have }\\ {\eta }_{\mathcal{T}}(U,T{)}^2\lesssim {\Vert {\nabla }_{\gamma }(\tilde{u}-\tilde{U})\Vert }_{L^2\left({\tilde{\omega }}_T\right)}^2+{\mathrm{osc}}_{\mathcal{T}}{\left(F,{\omega }_T\right)}^2+{\lambda }_{\mathcal{T}}^2\left({\omega }_T\right)\end{array}$$

下面介绍C2曲面的后验误差上界，为此，介绍常量“数据震荡”如下：
$$\mathrm{osc}\mathcal{T}(F,T{)}^2:=h_T^2\Vert F-\bar{F}{\Vert }_{L^2(T)}^2,\mathrm{osc}\mathcal{T}(F{)}^2:=\sum _{T\in \mathcal{T}}\mathrm{osc}\mathcal{T}(F,T{)}^2$$

那么，我们有如下定理（C2曲面的后验误差上界）：
$$\begin{array}{l}\text{ Theorem }43\text{ (a-posteriori upper bound for }{C}^2\text{ surfaces). Let }\gamma \text{ be of class }{C}^2\text{ and }\\ (67),(74),(90),\text{ and (91) hold. Let }\tilde{u}\text{ be the solution of }(18)\text{ with }\tilde{f}\in L_{2,\#}(\gamma )\text{ and }\\ U\in \mathbb{V}(\mathcal{T})\text{ be the solution to }(78)\text{ with }F=\tilde{f}\circ \mathrm{P}\frac{q}{q_{\mathrm{r}}},\text{ where }q\text{ corresponds to the }\\ \text{ parametrization }\chi =\mathrm{P}\circ \mathrm{X}\text{ of }\gamma \text{ . Then }\\ {\Vert {\nabla }_{\gamma }\left(\tilde{u}-U\circ {\mathbf{P}}_d^{-1}\right)\Vert }_{L_2(\gamma )}^2\lesssim {\eta }_{\mathcal{T}}(U{)}^2+{\mu }_{\mathcal{T}}^2(\Gamma )\Vert \tilde{f}{\Vert }_{L_2(\gamma )}^2\end{array}$$
证明思路：
和前面是类似的拆分：
$$\begin{array}{l}(105){\int }_{\gamma }{\nabla }_{\gamma }(\tilde{u}-\tilde{U})\cdot {\nabla }_{\gamma }\tilde{v}=I_1+I_2+I_3\\ \text{ with }\\ \begin{array}{rl}{I}_1& =-{\int }_{\Gamma }{\nabla }_{\Gamma }U\cdot {\nabla }_{\Gamma }(v-V)+{\int }_{\Gamma }F(v-V)\\ I_2& ={\int }_{\gamma }{\nabla }_{\gamma }\tilde{U}\cdot \mathbf{E}{\nabla }_{\gamma }\tilde{v}\\ I_3& ={\int }_{\gamma }\tilde{f}\tilde{v}-{\int }_{\Gamma }Fv\end{array}\end{array}$$
只不过这时候，$I_3$不再等于0，而是：
$$(106)I_3={\int }_{\gamma }\tilde{f}\left(\tilde{v}-\tilde{v}\circ {\mathbf{P}}_d\circ {\mathbf{P}}^{-1}\right)$$
估计一下，有：
$$I_3\lesssim {\beta }_{\mathcal{T}}(\Gamma )\Vert \tilde{f}{\Vert }_{L_2(\gamma )}{\Vert {\nabla }_{\gamma }\tilde{v}\Vert }_{L_2(\gamma )}$$

C2曲面的后验误差下界如下所述：
$$\begin{array}{l}\text{ Theorem }44\text{ (a-posteriori lower bound for }{C}^2\text{ surfaces). Under the same condi- }\\ \text{ tions as Theorem }43\text{ (a-posteriori upper bound for }{C}^2\text{ surfaces), we have }\\ {\eta }_{\mathcal{T}}(U,T{)}^2\lesssim {\Vert {\nabla }_{\gamma }(\tilde{u}-\tilde{U})\Vert }_{L^2\left({\tilde{\omega }}_T\right)}^2+{\mathrm{osc}}_{\mathcal{T}}{\left(F,{\omega }_T\right)}^2+\mu \mathcal{T}{\left({\omega }_T\right)}^2\\ \text{ where }{\mu }_{\mathcal{T}}\left({\omega }_T\right)={\max }_{T^{\prime }\subset {\omega }_T}{\mu }_{T^{\prime }}\end{array}$$