Abstract:We construct least squares formulations of PDEs with inhomogeneous essential boundary conditions, where boundary residuals are not measured in unpractical fractional Sobolev norms, but which formulations nevertheless are shown to yield a quasi-best approximations from the employed trial spaces. Dual norms do enter the least-squares functional, so that solving the least squares problem amounts to solving a saddle point or minimax problem. For finite element applications we construct uniformly stable finite element pairs, whereas for Machine Learning applications we employ adversarial networks.
What problem does this paper attempt to address?
This paper attempts to solve the key problems in the least - squares discretization of partial differential equations (PDEs) with non - homogeneous boundary conditions, that is, how to correctly impose essential boundary conditions, especially non - homogeneous boundary conditions. Specifically:
1. **Treatment of non - homogeneous boundary conditions**:
- In the traditional least - squares method, when it comes to non - homogeneous boundary conditions, it is usually necessary to use fractional - order Sobolev norms to measure the boundary residuals. However, these norms are difficult to evaluate in actual calculations.
- This paper proposes a new method. By introducing appropriate pre - conditioners and selecting suitable test spaces, it avoids the direct evaluation of fractional - order Sobolev norms, thus achieving a quasi - optimal least - squares approximation.
2. **Improved variational formula**:
- The paper proposes a modified variational formula, so that all the involved function spaces are in the integer - order Sobolev space or its dual space, rather than the fractional - order Sobolev space. This not only simplifies the calculation but also ensures the stability and accuracy of the numerical solution.
- For second - order elliptic partial differential equations and stationary Stokes equations, the paper shows how to construct uniformly stable finite - element space pairs and proves the quasi - optimality of these methods.
3. **Machine learning applications**:
- This method can also be applied to the solution of partial differential equations based on neural networks. By transforming the least - squares functional into the form of adversarial networks, the satisfaction of boundary conditions can be approximately enforced without explicitly imposing boundary conditions.
- This method has advantages in dealing with non - homogeneous boundary conditions, especially in high - dimensional problems, avoiding the complexity of directly constructing trial functions that satisfy boundary conditions.
### Formula summary
- **Alternative expressions of fractional - order Sobolev norms**:
\[
\| \cdot \|_{H^{\frac{1}{2}}(\partial\Omega)} \approx \sup_{0 \neq \vec{v} \in H(\text{div}; \Omega)} \frac{\left| \int_{\partial\Omega} \cdot \, \gamma_n \vec{v} \, ds \right|}{\|\vec{v}\|_{H(\text{div}; \Omega)}}
\]
\[
\| \cdot \|_{H^{-\frac{1}{2}}(\partial\Omega)} \approx \|\gamma' \cdot \|_{H_0^1(\Omega)'}
\]
- **Discretized form of the least - squares problem**:
\[
u_\delta := \arg\min_{w \in X_\delta} \frac{1}{2} \left( \sup_{0 \neq v \in Y_\delta^1} \frac{|(G_1 w - f_1)(v)|^2}{\|v\|^2_{Y_1}} + \|G_2 w - f_2\|^2_{Y_2'} \right)
\]
- **Discretized form of the saddle - point problem**:
\[
\begin{cases}
\langle \lambda_\delta, \mu \rangle_{Y_1} + (G_1 u_\delta)(\mu) = f_1(\mu) & (\mu \in Y_\delta^1), \\
(G_1 w)(\lambda_\delta) - \langle G_2 u_\delta, G_2 w \rangle_{Y_2'} = -\langle f_2, G_2 w \rangle_{Y_2'} & (w \in X_\delta).
\end{cases}
\]
Through these methods, the paper provides...