Multiple Linear Goal Functionals

• Multiple linear goal functionals are linear mappings that quantify distinct performance objectives in settings such as PDE-constrained optimization, reinforcement learning, and simplex methods.
• They employ adjoint problems and dual-weighted residual error estimators to guide adaptive mesh refinement, ensuring optimal convergence and computational efficiency.
• Efficient solution characterization via LP-based tests enables robust detection of Pareto frontiers and practical scalarization in multi-objective frameworks.

Multiple linear goal functionals refer to a collection of linear mappings—often termed “quantities of interest” or objectives—defined on the solution space of an underlying problem. In PDE-constrained optimization, finite element analysis, multi-objective reinforcement learning, and vector optimization over the simplex, such functionals formalize the simultaneous assessment or maximization of several performance criteria, outputs, or objective values. The mathematical and algorithmic treatment of multiple linear goals comprises their formal definition, efficient solution characterization, error estimation, and adaptive computational frameworks for controlling goal error or for efficiently identifying Pareto frontiers.

1. Mathematical Formulation and Classes of Multiple Linear Goal Functionals

Let the solution variable $u$ (or $x$ for finite-dimensional vector problems) reside in a function or vector space, such as $H^1_0(\Omega)$ in PDE theory or $\Delta \subset \mathbb{R}^n$ in the probability simplex. Given $N$ linear functionals $G_j$, they take the form: $$G_j(u) = \int_\Omega (g_j\,u + \mathbf{g}_j \cdot \nabla u)\,dx, \qquad \text{or} \qquad f_k(x) = c_k^T x$$ where the coefficients $(g_j, \mathbf{g}_j)$ or $c_k$ specify the particular quantities of interest (Becker et al., 5 Jan 2026, Mifrani, 2024).

These functionals are aggregated either via a vector-valued mapping $F(u) = (G_1(u),...,G_N(u))$ for Pareto analysis, or through scalarization (linear or non-linear) in optimization ($f(V) = w^\top V$ for policy gradient objectives (Bai et al., 2021)). In vector optimization over the simplex, feasible $x \in \Delta$ is classified into deterministic (vertex/extreme), partially randomized (face-interior), or randomized (interior) points, each admitting distinct efficiency characterizations for the multi-objective maximizing problem (Mifrani, 2024).

2. Duality, Adjoint Problems, and Combined Error Representation

Each goal functional induces an adjoint (dual) problem central for both optimality and adaptive error estimation: $a(v, z_j^\star) = G_j(v) \qquad \forall v$ for the Hilbert-space setting or, in linear programming over the simplex, via Evans–Steuer weight-sum characterizations: efficient $x^*$ solves a single-objective LP for some $\lambda > 0$: $\max_{x \in \Delta} \, (\lambda^T C)x$ where $C$ collects the $c_k$ (Becker et al., 5 Jan 2026, Mifrani, 2024). In discretized PDE contexts, Galerkin orthogonality and dual-weighted residual (DWR) theory yield the error split: $$G_j(u^\star) - G_j(u_H) = a(u^\star - u_H, z_j^\star - z_{j,H})$$ leading to total error bounds: $$\sum_{j=1}^N |G_j(u^\star) - G_j(u_H)| \leq \|u^\star - u_H\|_a \sum_{j=1}^N \|z_j^\star - z_{j,H}\|_a$$ and, for adaptive methods, combined a posteriori estimators constructed from weighted sums of dual solutions (Becker et al., 5 Jan 2026, Endtmayer et al., 2018).

3. Adaptive Algorithms for Simultaneous Goal Control

In PDE and finite element settings, adaptive refinement seeks to simultaneously control the error in all $N$ linear goals; rigorous frameworks such as the multigoal-oriented adaptive finite element method (NGO-AFEM) implement this for symmetric elliptic PDEs:

• On each mesh, solve the primal system and one dual (cycling through $j = 1,...,N$ over $N$ steps).
• Use residual-based error estimators for primal and active dual; mark elements for refinement according to Dörfler marking, irregular marking, and cardinality control to optimize refinement locally for all goals.
• Only two linear systems are solved per refinement level: primal and single dual (Becker et al., 5 Jan 2026).

The main algorithmic outcomes are:

• R-linear convergence of the estimator product (primal error times sum of dual errors).
• Optimal algebraic rates: $(\dim X_\ell)^{-(s+t)}$ decay for combined goal errors given solution regularity (nonlinear approximation class).
• Mesh sequences generated suffice to resolve all singularities corresponding to any dual goal. This avoids the need for separate meshes per goal.

4. Computational Characterization of Efficiency and Pareto Solutions

In vector/matrix linear programming over the simplex, efficient solutions (“Pareto optimal”) are precisely characterized:

• Weight-sum lemma: $x^*$ is efficient iff there exists $\lambda > 0$ such that $x^*$ solves $\max (\lambda^T C)x$ over $\Delta$.
• Efficient points are classified directly by the support of $x^*$ (full, face, extreme) and the index set $J^*(\lambda^T C)$ of maximizers of the weighted score.
• LP-based procedures (T⁰, T¹, T²) test efficiency in polynomial time by examining the maximality and support of candidate points; no need for vertex enumeration (Mifrani, 2024).

Point Class	Efficiency Condition (via λ, support)	LP Test
Interior (randomized)	All $(\lambda^T C)_j$ equal	T⁰
Face-interior (partial)	Maximal indices $J'$ match support	T¹
Vertex (deterministic)	Unique maximum at $j$	T²

This enables full recovery of efficient frontiers for multiple linear objectives.

5. Sample Complexity, Convergence, and Practical Guidelines

In multi-objective reinforcement learning, simultaneous maximization of linearly combined value functionals $V_i(\theta)$ is tractable via policy-gradient methods. For linear scalarizations $f(V) = w^\top V$:

• The update reduces to a standard policy-gradient (“REINFORCE”) on scalar reward $\bar{r}(s,a) = \sum_i w_i r_i(s,a)$.
• Sample complexity to achieve $\epsilon$-optimality is $$\mathcal{O}(M^4 \sigma^2 / (1 - \gamma)^8 \epsilon^4)$$, with $M$ objectives and trajectory variance $\sigma^2$ (Bai et al., 2021).
• Weight selection $w$ encodes trade-off; normalization $\sum_i w_i \approx 1$ is advised for conditioning and convergence efficiency.

For adaptivity in FEM or DWR, weights can reflect relative tolerances or importance per goal, affecting marking selection and local mesh refinement (Endtmayer et al., 2018).

6. Numerical Evidence and Implementation Considerations

Numerical experiments in elliptic PDE settings (unit square, Z-shaped domains) confirm theoretically predicted convergence rates for simultaneous control of $N$ linear goals:

• The multigoal estimator product decays as $\mathcal{O}(\# \mathrm{DoF}^{-(s+t)})$ in the adaptive algorithm (Becker et al., 5 Jan 2026).
• Individual dual error estimators exhibit staircase behavior (updated only when active) but overall decay optimally.

Variants that omit dual cycling or irregular marking lose optimality, highlighting the necessity of specialized adaptive marking schemes for multigoal scenarios. Per iteration cost is invariant in $N$, as only two sparse linear solves are required, affording scalability to high $N$.

Efficient solution characterization for simplex problems is actionable via the provided LP tests, demonstrated with fully worked examples (e.g., $n = m = 3$ matrices recovering edge-frontier efficiency) (Mifrani, 2024).

7. Connections and Generalizations

Multiple linear goal functionals are foundational tools in goal-oriented adaptive mesh refinement, multi-objective optimization, and vector optimization. The combination of adjoint-based error control, efficient frontier characterization, and adaptive computational strategies enables robust, scalable management of multiple objectives in mathematically rigorous frameworks.

This topic connects directly to:

• Pareto optimization,
• Dual-weighted residual theory,
• Multi-objective RL scalarization,
• Multi-goal adaptive FEM,
• Efficient solution characterization for linear programming.

Current research verifies optimality and computational efficiency in these frameworks, providing practical and theoretic guarantees for simultaneous control or maximization of multiple linear goals.