Set-Based Scalarization Techniques

• Set-based scalarization techniques are methods that extend classical scalarization by optimizing over sets, targeting entire Pareto fronts and nonconvex boundaries.
• They employ advanced mathematical tools, including Tchebycheff and Gerstewitz scalarizations, to handle set-valued objectives and robust optimization.
• Applications span many-objective, set-valued, and lattice optimization, delivering enhanced coverage, computational efficiency, and strong theoretical guarantees.

Set-based scalarization techniques are a family of methods in vector, multiobjective, and set-valued optimization where the scalarization is constructed around sets of solutions, set-valued objectives, or nonconvex feasible regions, as opposed to classical pointwise or single-solution scalarizations. These approaches generalize traditional scalarization, permitting the direct manipulation, optimization, and characterization of sets (e.g., Pareto sets, nonconvex boundaries, robust solution sets), providing powerful machinery for both theory and computation in the multiobjective and set-optimization regimes.

1. Foundations and Motivation

Set-based scalarization techniques extend scalarization beyond the classical reduction of vector or multiobjective problems to single-objective optimization, to settings where the object of interest is itself a set—typically arising either as a solution set (e.g., sets of weakly efficient points), the image of a set-valued objective, the boundary of a feasible region, or the robust/uncertain extension under adversarial perturbations. The principal motivations are:

• Direct optimization over solution sets to efficiently cover complex Pareto fronts without exponential enumeration.
• Scalarization within or over set-valued objectives, enabling the application of variational, duality, and subdifferential tools to set-minimization, robust, and variational problems.
• Characterization and computation of nonconvex or implicitly-defined boundaries via parameterized families of scalar optimization problems, leveraging dual representations and variational geometry.
• Robust set optimization under uncertainty, where performance and feasibility are evaluated uniformly or with respect to the “worst-case” over sets of scenarios.

The set-based perspective is essential for scalability, coverage guarantees, and maintaining a direct connection to intrinsic set orderings and generalized solution concepts (e.g., infimizers, lattice minimizers) (Hamel et al., 2020, Ma et al., 10 Oct 2025, Fakhar et al., 2023).

2. Set-Based Scalarization in Many-Objective and Pareto Set Coverage

Classical scalarizations (e.g., weighted sum, Tchebycheff) focus on generating individual efficient points, each run requiring a new choice of weights to scan the Pareto front. Set-based scalarization for many-objective optimization directly targets a whole set of solutions, optimizing to ensure global coverage or complementary addressing of objectives.

A canonical example is the Tchebycheff Set Scalarization (Lin et al., 2024), which for $m\gg1$ objectives, seeks a set $X_K=\{x^{(1)},\dots,x^{(K)}\}$ ($K\ll m$) to “cover” all objectives via

$$g^{\mathrm{TCH-Set}}(X_K\mid \lambda) = \max_{1\leq i\leq m} \Big\{\lambda_i \left(\min_{1\leq k\leq K} f_i(x^{(k)}) - z^*_i\right)\Big\},$$

where each objective’s best coverage over the $K$ solutions is penalized, minimizing the worst offense. The method admits a smooth surrogate using log-sum-exp, facilitating efficient gradient-based joint optimization of all $K$ points. Minimizers are sets of weakly Pareto-optimal (often strictly Pareto-optimal) points, and the approach consistently outperforms classical scalarizations, MosT, and SoM in high-dimensional benchmarks, both in worst and mean coverage metrics (Lin et al., 2024).

Another paradigm is the Target Point-based Tchebycheff Distance (TPTD) (Nagakane et al., 1 May 2025), in which instead of fixed weights, variable target points $t$ on a hyperplane adapted to the Pareto front shape are used: $$s^{\mathrm{TPTD}}(x|f', t) = \max_{i=1}^m |f'_i(x) - t_i|.$$ Here, multiple optimization runs on scalarized subproblems (each defined by a distinct $t$) ensure uniform, shape-adaptive coverage even in highly nonconvex or “inverted” Pareto fronts, with each solution corresponding to the weakly Pareto-optimal point closest (in Chebyshev distance) to its target. This adaptive placement of anchors vastly improves both coverage and computational efficiency in many-objective scenarios, compared to fixed-weight schemes.

3. Set-Based Scalarization in Set-Valued and Lattice Optimization

In the setting of set-valued mappings and set optimization, the object of analysis is a function $f : X \to \mathcal L$ with values in a complete lattice of sets. The solution concept is inherently set-valued: one seeks infimizer sets $M$ such that the lattice infimum over $M$ coincides with the infimum over $X$, while also maintaining minimality (in the lattice sense).

The inf-translation principle (Hamel et al., 2020) provides a foundational tool: for $M\subset X$,

$$f(\cdot; M): X \to \mathcal L, \qquad f(x; M) = \inf_{y\in M} f(x+y).$$

This enables the reduction of set-minimization problems to scalarized problems at the origin, allowing the application of support-type scalarizations $\phi_{z^*}(x) = \inf\{z^*(z): z\in f(x)\}$ over dual cone directions $z^*$. Necessary and sufficient optimality and variational conditions follow, crucially facilitating both exact characterization and algorithmic realization of set-valued minimization.

Further, scalarization of set-valued objectives in robust optimization and uncertain settings leads to more refined notions of robust optimality, such as $L, U, S$-ordered robust solutions, with precise characterization via Gerstewitz-type functionals and sup/inf formulations over set-unions (Digar et al., 2 Nov 2025).

4. Variational, Duality, and Nonconvex Boundary Characterization

Set-based scalarization techniques are essential in variational analysis, generalized convex duality, and the variational geometry of boundaries and feasible regions:

• Gerstewitz and Uniform-Sublevel Scalarizations: Gerstewitz functionals,

$$\varphi_{A, k}(y) = \inf\{t \in \mathbb{R} : y \in A + t k \},$$

reveal tight characterizations of efficient, weakly efficient, and minimal solutions in vector optimization under arbitrary domination structures and for both convex and nonconvex problems. They admit explicit geometric interpretations, are robust under parameter perturbation, and their minimizers coincide with generalized Pascoletti–Serafini scalarizations (Weidner, 2017, Weidner, 2016, Weidner, 2016).

• Nonconvex Boundary Representation: For domains $D$ with possibly nonconvex structure, boundary characterization can be obtained via a Pascoletti–Serafini-type scalarization utilizing local partial orders induced by parameterized spherical cones (Ma et al., 10 Oct 2025). The boundary is then reconstructed as the collection of maximizers of scalarized subproblems indexed by reference points and directions:

$$f \in \partial D \quad \Longleftrightarrow \quad f \in \arg\max_{g\in D} H^\eta_{a,b}(g),$$

with $H^\eta_{a,b}(g)$ a closed-form function parameterizing the cone geometry. This dual representation generalizes to infinite-dimensional function spaces and underlies algorithms for tracing, approximating, and optimizing over nonconvex sets.

• Hamilton–Jacobi–Bellman (HJB) Equations in Set-Valued Control: Intrinsic scalarization—driven by the geometry of the dynamic set's boundary (via normal vectors)—allows the reduction of multivariate, time-inconsistent dynamic programming problems to time-consistent, path-dependent scalarizations (İşeri et al., 2023). The set-valued HJB equations yield classical solutions via a set-valued Itô formula; moving scalarizations are constructed from the normal flow on the evolving boundary.

5. Nonlinear and Generalized Scalarization Functionals

Set-based scalarization necessitates broadening the class of admissible scalarizers. This includes:

• Oriented Distance Scalarizations: Generalizations of the Hiriart–Urruty function and Minkowski functionals, defined as

$\Delta_{-K}(y) = d_{-K}(y) - d_{(-K)^c}(y),$

are used to compare, order, and provide descent directions in trust-region and line-search algorithms for set or vector-valued mappings, preserving Lipschitz continuity and strict monotonicity relative to ordering cones (Ghosh et al., 17 Sep 2025, Bouza et al., 2021).

• Support-Functional, Infimal, and Quasidifferentiable Scalarizations: Unified frameworks (e.g., by Drummond–Svaiter, Gerstewitz, subdifferential characterization) consider scalarizers as support functions of dual generators, or as differences of such supports over pairs of convex compact sets, admitting quasidifferentiable (nonconvex but still positively homogeneous) representatives. Their mutual inclusions, translation property, and dual cone representations are clarified; these grant flexibility for algorithmic and theoretical developments beyond convexity (Bouza et al., 2021).

6. Applications and Empirical Performance

The set-based scalarization principle yields powerful practical methods:

• Efficient Pareto coverage for large-scale many-objective optimization with $m\gg 10^2$ (Lin et al., 2024, Nagakane et al., 1 May 2025).
• Robust and adaptive coverage of highly nonconvex, variable-dependent or pathological Pareto fronts—outperforming traditional MOEAs, as measured by hypervolume and computational speedup factors up to $474\times$ (Nagakane et al., 1 May 2025).
• Nonmonotone trust-region and line-search optimization for set-valued objectives, exhibiting superior convergence and robustness relative to classical monotone approaches (Ghosh et al., 17 Sep 2025).
• Algorithmic frameworks for multi-objective branch-and-bound, utilizing set-based scalarization to prioritize Pareto gap closure and prune search space adaptively (Bauß et al., 2023).
• New existence, optimality, and duality claims for set-valued (possibly noncontinuous and noncompact) optimization problems, using scalarizations that naturally integrate co-level set geometry and variational properties (Fakhar et al., 2023, Schrage, 2010).

Typically, set-based scalarization admits efficient parallelization and decouples optimization over large or complex solution sets into tractable single-objective (or low-dimensional) scalar subproblems, suitable for black-box or gradient-based solvers.

7. Theoretical Guarantees and Future Directions

Set-based scalarization techniques confer essential theoretical properties:

• Pareto-optimality or Pareto-stationarity of resulting solution sets under mild conditions (even for structured smooth approximations).
• Uniform convergence, error bounds, and continuity with respect to scalarization parameters (e.g., smoothing, target points).
• Direct links between the minimizers (or maximizers) of set-based scalarizations and intrinsic geometric objects: efficient solutions, boundary points, infimizers, and robust solution sets.
• Dual representations and strong/weak duality theorems in the set-valued framework, allowing transference of optimality and saddle-point conditions from scalar to set (or vector) scales (Hamel et al., 2020, Schrage, 2010).

Ongoing work targets further unification of generalized scalarizations; extension to nonconvex and infinite-dimensional settings; efficient sampling or parameterization of solution representations; invariance and adaptivity in robustness and uncertainty; and scalable, high-performance algorithms for large-scale, set-based, multiobjective computational tasks. The set-based scalarization paradigm provides a universal template for translating between abstract set-valued orderings and single-valued optimization landscapes, with theoretical rigor and broad algorithmic applicability.