Trade-off invariance for weighted scalarizations in multi-objective optimization

Technical Summary: Trade-off Invariance for Weighted Scalarizations in Multi-Objective Optimization

Problem Statement
The paper addresses a fundamental question in multi-objective optimization regarding the stability of solutions obtained via weighted-sum scalarization. Consider an abstract multi-objective minimization problem defined by a vector-valued map $f: U \to (\mathbb{R} \cup \{+\infty\})^N$, where $U$ is an arbitrary nonempty set. The problem is to minimize $f(u) = (f_1(u), \dots, f_N(u))$.

The standard approach to solving such problems is weighted-sum scalarization, which replaces the vector problem with a family of scalar problems:
$$H_\lambda(u) := \sum_{i=1}^N \lambda_i f_i(u),$$
where $\lambda \in (0, +\infty)^N$ is a weight vector. The authors investigate whether the "trade-off" selected by a specific weight vector $\lambda$ is intrinsic to the weight itself or if it depends on the specific minimizer or minimizing sequence chosen. Specifically, they ask:

1. If the set of minimizers $M_\lambda$ contains multiple elements, do they yield distinct objective vectors $f(u)$?
2. If minimizers do not exist, do different minimizing sequences converge to distinct limiting objective vectors?

The setting is deliberately abstract: no assumptions are made regarding the topology, convexity, compactness, or lower semi-continuity of $U$ or the functionals $f_i$.

Methodology
The authors employ a purely variational approach, leveraging the Trade-off Invariance Principle recently introduced in [3] for regularized functionals of the form $F + \sum \alpha_j G_j$. The methodology involves three key steps:

1. Parametrization via the Simplex: Since only the relative proportions of weights matter, the parameter space is restricted to the open simplex $\Delta^\circ_N = \{ \lambda \in (0, +\infty)^N : \sum \lambda_i = 1 \}$.
2. Reduction to Anchored Form: The authors demonstrate that any symmetric weighted-sum scalarization can be rewritten in the "anchored" form treated in [3] by selecting one objective function as a reference. Specifically, for a fixed $k$, $H_\lambda$ is equivalent (up to a positive scaling factor) to $\tilde{H}^{(k)}_\alpha = f_k + \sum_{i \neq k} \alpha_i f_i$, where $\alpha_i = \lambda_i / \lambda_k$.
3. Measure-Theoretic Transfer: Using the smoothness and local Lipschitz continuity of the coordinate transformation between the anchored parameters $\alpha \in (0, +\infty)^{N-1}$ and the simplex $\Delta^\circ_N$, the authors prove that negligible (null-measure) exceptional sets in the parameter space of [3] map to negligible sets on the simplex. This allows the transfer of generic uniqueness results from the anchored framework to the symmetric weighted-sum framework.

Key Contributions and Results

• Generic Uniqueness of Minimizers (Theorem 3.1):
  The paper proves that for almost every weight vector $\lambda \in \Delta^\circ_N$ (with respect to the $(N-1)$-dimensional Lebesgue measure), the set of objective vectors corresponding to the minimizers of $H_\lambda$ contains at most one point.

  • Implication: Even if the scalarized problem has multiple distinct minimizers in the decision space $U$ (i.e., $|M_\lambda| > 1$), they all map to the exact same point in the objective space. The trade-off is unique for almost all weights.

• Generic Uniqueness of Limiting Objective Vectors (Theorem 3.2):
  Extending the result to approximate minimization, the authors show that for almost every $\lambda \in \Delta^\circ_N$ (provided the infimum is not $-\infty$), all minimizing sequences for $H_\lambda$ determine the same limiting objective vector $y_\lambda$.

  • Implication: The asymptotic trade-off is also intrinsic to the weight vector and independent of the specific sequence of iterates chosen.

• Bi-Objective Case (Corollary 3.3):
  In the specific case where $N=2$, the exceptional set of weights where uniqueness fails is at most countable. This reduces the genericity statement to the one-parameter Trade-off Invariance Principle established in [3].

• Geometric Interpretation:
  The results are interpreted geometrically in the attainable objective set $Y = f(U)$.

  • Exact Level: For almost every $\lambda$, the supporting hyperplane defined by $\lambda$ exposes at most one point in $Y$. If the set of exposed points is non-empty, it is a singleton.
  • Asymptotic Level: The authors define "asymptotically exposed objective vectors" as limit points of sequences in $Y$ that minimize the linear functional asymptotically. They prove that for almost every $\lambda$, the set of such vectors contains at most one point. Furthermore, these asymptotically exposed points are nondominated.

Significance and Claims
The paper claims to establish a generic uniqueness property in the objective space for weighted-sum scalarizations. The significance lies in the fact that this result holds under minimal assumptions (no convexity, compactness, or continuity required).

The authors assert that the "trade-off" selected by a weighted-sum scalarization is an intrinsic property of the weight vector for almost all weights, rather than an artifact of the specific solution path or the existence of multiple minimizers. This provides a rigorous theoretical justification for the stability of trade-offs in multi-objective optimization, confirming that the scalarization method selects a unique objective-space vector (or a unique asymptotic limit) generically, even in highly non-convex or non-smooth settings.

The work does not propose new algorithms or experimental applications; rather, it provides a foundational variational analysis that clarifies the relationship between the parameter space of weights and the geometry of the attainable objective set.