The Limitations of Linear Scalarization

40 concepts43 papers1 proposal

When designing an autonomous system, training a machine learning model, or engineering a complex physical structure, we rarely have a single goal. We want a robot that is both fast and safe; a recommendation engine that is both accurate and diverse; a fusion reactor that maximizes plasma confinement while minimizing engineering complexity.

The most common way to handle these conflicting goals is linear scalarization (or the weighted-sum approach). We assign a numerical weight to each objective, multiply them together, and add them up to form a single scalar score:

$$F(x) = w_1 f_1(x) + w_2 f_2(x) + \dots + w_n f_n(x)$$

This reduces a complex, multi-dimensional problem into a familiar single-objective optimization problem. It allows us to use standard gradient descent, genetic algorithms, or reinforcement learning frameworks straight out of the box.

However, collapsing multi-dimensional trade-offs into a single scalar value introduces fundamental limitations. Across fields as diverse as robotics, deep learning, and clinical medicine, researchers are finding that linear scalarization obscures the true problem landscape, introduces severe optimization pathologies, and forces arbitrary design decisions.

The A Priori Guessing Game

The most immediate practical flaw of linear scalarization is that it requires practitioners to choose the weights before the optimization process begins. This is known as a priori weight selection.

To pick an optimal weight vector, you must already understand the exact exchange rate between your objectives. For example, in multi-UAV path planning, how many meters of flight distance are you willing to sacrifice to reduce the threat level by 5%? In a power grid restoration scenario, how do you weigh the amount of power restored against the physical effort of flipping manual switches?

As Jun Inagaki et al. observe in A GA-based Multi-Objective Optimization for Service Restoration in Power Distribution Systems, this forces operators to make highly subjective, arbitrary guesses without understanding the underlying trade-offs. If the guess is wrong, the entire optimization run must be discarded, leading to an inefficient trial-and-error loop.

Howard Coffin notes in A Local Optimization Framework for Multi-Objective Ergodic Search that the ideal trade-off is fundamentally unknown until the set of possible optimal solutions—the Pareto front—is explicitly explored. By choosing a fixed weight vector upfront, designers are effectively flying blind, optimizing for a preconceived compromise that may not even be physically or computationally viable.

The Geometry of Blind Spots

Even if you are willing to guess dozens of different weight combinations, linear scalarization suffers from a deeper mathematical limitation: it is geometrically incapable of capturing the full landscape of optimal compromises.

Imagine a trade-off curve between two objectives where the frontier dips inward—a non-convex Pareto front. Geometrically, a linear weighted sum acts like a flat blade pressing against this objective space. As you vary the weights, the blade rotates, but it can only ever touch the outermost convex hull of the frontier. It will cleanly skip right over any solutions resting inside the non-convex "dents" or pockets.

Objective 2
   ^
   |   * (Convex point found by scalarization)
   |    \
   |     \   o  <-- Non-convex compromise (Invisible to linear scalarization)
   |      \ /
   |       * (Convex point found by scalarization)
   +--------------------------------------------> Objective 1

No matter what combination of positive weights you choose, a linear scalarization will never settle on these non-convex compromise solutions. As Anabel Yong et al. demonstrate in A study of EHVI vs fixed scalarization for molecule design, fixed scalarization typically collapses the search to a single point on the frontier, severely limiting the variety of candidate designs available for further testing.

By hiding the Pareto frontier, as argued by Joost van Doorn et al. in Balancing Relevance Criteria through Multi-Objective Optimization, scalarization deprives decision-makers of the ability to choose a model that specifically favors one objective over another when operational contexts change.

Dynamic Pathologies and Runaway Gradients

Linear scalarization does not just limit what solutions you can find; it also actively disrupts the dynamics of the optimization process itself.

When multiple objectives are combined into a single score, the optimizer sees only the aggregate gradient. If one objective has a steeper gradient or is significantly easier to satisfy early in the search, the optimizer will aggressively exploit that "easy" objective. It will run down that gradient, often sacrificing performance on the harder objectives to maximize the total scalar score.

This causes several distinct pathologies:

Utilitarian Exploitation vs. Fairness

At its core, linear scalarization is a utilitarian mathematical framework: it maximizes the total aggregate performance. It does not care if that total is achieved by balancing all goals beautifully, or by completely maximizing one goal while reducing another to zero.

In many real-world applications, this utilitarian bias is unacceptable. Consider an algorithm managing a mobile edge computing network or a power dispatch system. If it uses a weighted sum to balance energy efficiency and latency, it might find a solution that achieves a spectacular aggregate score by offering near-zero latency to a few users while completely starving others of network resources.

To combat this, Giseung Park et al. propose a shift in The Max-Min Formulation of Multi-Objective Reinforcement Learning: From Theory to a Model-Free Algorithm. By replacing the weighted sum with a max-min formulation, the optimization objective shifts from maximizing the average to maximizing the worst-performing metric. This guarantees a baseline level of fairness and performance across all objectives, ensuring no single goal is neglected for the sake of an aggregate score.

Preserving the Multi-Dimensional Frontier

To overcome these limitations, modern research is moving away from scalarization entirely, treating multi-objective problems as fundamentally multi-dimensional throughout the entire optimization process.

Instead of searching for a single point, algorithms like Pareto Multi-Task Learning, developed by Xi Lin et al. in Pareto Multi-Task Learning, decompose the objective space into distinct sub-regions. This allows the model to map out a diverse set of Pareto-optimal solutions simultaneously, giving practitioners a comprehensive menu of high-quality models to choose from.

Objective 2
   ^
   |   *  *  *  *  *  <-- True Pareto front mapped simultaneously 
   |                      without collapsing into a single scalar
   |
   +--------------------------------------------> Objective 1

Similarly, evolutionary frameworks like MOVE, introduced by Jackson Dean et al. in Many-objective Optimization via Voting for Elites, maintain specialized "elites" that focus on different subsets of objectives. This allows solutions to dynamically switch goals during the search, discovering intermediate stepping stones that a rigid, scalarized reward function would completely miss.

In safety-critical systems where trade-offs cannot be compromised, researchers are abandoning numerical scores altogether. In Posetal Games: Efficiency, Existence, and Refinement of Equilibria in Games With Prioritized Metrics, Alessandro Zanardi et al. replace scalar costs with partially ordered sets (posets). This allows an autonomous vehicle to navigate using strict, un-weighted qualitative hierarchies—such as "safety first, then traffic rules, then passenger comfort"—ensuring that comfort can never mathematically cannibalize safety, no matter how high the comfort score might be.

Key Foundational Work

Selection of Linguistic Classfication Rules Using Two-objective Genetic Algorithms (1995) — contributed 1 concept elements to this direction.
Active harmony: towards automated performance tuning (2002) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_scalarization_limitations_14513f5b,
  title = {scalarization limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scalarization-limitations/rdr_14513f5b14364eab83f7b2428437114f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_scalarization_limitations_14513f5b_v1,
  title = {scalarization limitations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scalarization-limitations/rdr_14513f5b14364eab83f7b2428437114f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake