Solve with Your Preferred Solver. Estimate with Least Squares.

One Error Estimator, Independent of Your Solver 不依賴求解器的後驗誤差估計

The solver and the a posteriori error estimator are independent modules. Compute the approximation with the numerical method you prefer. Keep that approximation fixed; use least squares to recover an auxiliary flux and evaluate the first-order residuals. The resulting error estimation rests on PDE stability and least-squares norm equivalence, not on which solver produced the approximation.

求解器與後驗誤差估計器是兩個獨立的部分。你用自己選擇的方法計算近似解，然後固定這個解，通過 least-squares recovery 得到輔助通量，再計算一階系統的殘差。誤差估計來自 PDE stability 與 LS norm equivalence，並不依賴原近似解是由哪一種 solver 算出來的。

The General Elliptic Model 一般二階橢圓模型

For the general, possibly nonsymmetric or indefinite, second-order elliptic equation, the paper considers

with the flux $\boldsymbol\sigma=\boldsymbol f_2-A\nabla u$. For illustration, take the diffusion-flux form of the least-squares functional:

對一般非對稱、甚至 indefinite 的二階橢圓問題，先引入具有物理意義的 flux，再由 first-order system 的兩個殘差構造 least-squares functional。上式展示 diffusion-flux 形式；論文也處理以 $\gamma\in[0,1]$ 統一的 lower-order-term splitting。

The associated norm equivalence is the essential ingredient. For any conforming pair $(\boldsymbol\tau,v)\in H(\mathrm{div};\Omega)\times H_0^1(\Omega)$, the homogeneous functional controls the natural product norm.

Your Solver First, Least-Squares Estimation Second 原方法負責求解，LS 負責誤差估計

For the conforming Galerkin FEM studied in the paper, let $u_h^c$ be the scalar approximation returned by that solver. Keep the Galerkin solution unchanged and hold $u_h^c$ fixed and solve only for a Raviart–Thomas flux:

以論文分析的 conforming Galerkin FEM 為具體例子，先用原 solver 算出 $u_h^c$。保持這個解不變，只對 $H(\mathrm{div})$-conforming flux 作 least-squares 極小化。這一步是獨立的 recovery solve，不需要重新求解 scalar PDE。

The recovery is a linear, symmetric positive definite problem for the flux. Its exact counterpart contains the physical flux, and the recovery error satisfies a best-approximation bound with the scalar Galerkin error as a perturbation:

Recovered flux 的誤差由 flux 空間的最佳逼近及原方法的 scalar error 控制。Recovery 本身是線性的 SPD 系統；它並非把原本的 Galerkin 方法改寫成 coupled LSFEM。

Why Reliability and Efficiency Follow from the LS Norm 為什麼 Reliability 與 Efficiency 直接成立

Write the exact solution as $(\boldsymbol\sigma,u)$. The residual identity and first-order norm equivalence give, for every conforming pair,

In particular, the non-intrusive estimator is $\eta_h=\mathcal L(\boldsymbol\sigma_h^r,u_h^c)^{1/2}$. Its elementwise contributions are simply the two squared first-order residuals. The same norm equivalence gives reliability and efficiency for the combined flux–potential pair, without using Galerkin orthogonality or least-squares optimality.

把 Galerkin approximation 和 recovered flux 合在一起後，least-squares functional 就是整個 flux–potential pair 的誤差估計器。兩側估計來自 first-order system 的 norm equivalence，不需要利用原 Galerkin 方法的正交性，也不需要這一對變量是完整 LSFEM 的 minimizer。

The Combined Two-Step Interpretation 把兩步看作同一個問題

For the Galerkin example analyzed in the paper, the original variational equation and the auxiliary flux-recovery equation are combined into a single two-field problem. Its block structure is triangular: the scalar equation is solved first, and the recovery equation uses that scalar result. The combined formulation has precisely the same discrete solution as the original sequential algorithm.

對論文具體分析的 Galerkin 例子，原來的 Galerkin equation 與 flux-recovery equation 可以寫成一個 combined two-step problem。它具有先 scalar、後 flux 的三角結構；這是分析上的統一表示，並沒有改變實際求解順序。

This viewpoint supplies a direct framework for a priori analysis and, crucially, for analyzing the adaptive algorithm driven by the least-squares estimator.

這個角度的用處，是把 two-step procedure 當作完整的數值逼近來研究 a priori error 與 adaptive convergence，而不必為 scalar-only estimator 另外建構一套 indirect efficiency argument。

Adaptive Refinement: Plain Convergence 自適應加密的 Plain Convergence

The adaptive algorithm repeats solve, recover, estimate, mark, and refine. Under the paper’s marking and mesh-refinement assumptions, the elliptic problem has the convergence result

This is plain convergence: both approximations converge and the least-squares estimator tends to zero. The theorem does not assert an optimal adaptive convergence rate. For the general indefinite linear problem, the analysis also uses the appropriate initial-mesh stability condition.

論文證明的是 plain convergence：scalar 與 flux 的誤差都趨於零，estimator 也趨於零。這不是對 adaptive optimal rate 的宣稱。對一般 indefinite 線性問題，分析中也使用相應的初始網格離散穩定性條件。

Nonlinear Problems: Keep the Nonlinear Solve, Recover Linearly 非線性問題：原方程照常求解，Recovery 仍是線性的

The paper studies the monotone nonlinear model

with $\boldsymbol\sigma=-\nabla u$. Its nonlinear least-squares functional is

First solve the nonlinear variational problem for $u_h^{nl}$. Then freeze $u_h^{nl}$ and minimize $\mathcal N(\boldsymbol\tau_h,u_h^{nl})$ only over the flux. The cubic term is now known, so the recovery is again a linear $H(\mathrm{div})$ problem.

先用標準非線性 Galerkin 方法得到 $u_h^{nl}$；隨後把它固定。此時 $(u_h^{nl})^3$ 已經是已知數據，flux recovery 仍然只需解線性系統。因而不必解 nonlinear least-squares minimization problem。

Strong monotonicity gives reliability; the paper establishes local efficiency and plain convergence for the corresponding adaptive algorithm as well. The full $H(\mathrm{div})$ reliability constant in the nonlinear analysis depends on the data and the scalar norm.

利用 monotonicity 可以得到 error bound；論文進一步證明 local efficiency 及 adaptive plain convergence。非線性情形下完整 $H(\mathrm{div})$ 誤差界中的常數依賴數據及 scalar approximation 的範數。

The Solver Is Yours. The Estimator Is Least Squares. 求解方法由你決定，後驗估計交給 LS

The auxiliary flux problem is linear and admits efficient H(div) solvers. The a posteriori estimator remains the same least-squares functional regardless of the original PDE solver.

輔助通量問題是線性的，也可採用高效的 H(div) 求解器。不論原方程用哪種 solver，後驗誤差估計都由同一個 least-squares functional 完成。

Core Paper 核心論文