Tchebycheff Scalarization for Non-Convex Pareto Fronts

This direction is merged into another direction: linear scalarization limitations. Its last version stays readable and citable here.

Version 1 of 126 papers38 conceptsfrozen 2026-10-04

In most real-world machine learning and robotics problems, we are rarely optimizing for a single goal. A self-driving car must balance safety against speed; a federated learning model must balance global accuracy against local fairness. These conflicting objectives create a Pareto front—a set of optimal solutions where you cannot improve one objective without degrading another.

The standard way to navigate this trade-off is linear scalarization: assigning a weight to each objective and minimizing their weighted sum. However, as Susan Wei and Marc Niethammer point out in The Fairness-Accuracy Pareto Front, this approach has a fundamental geometric flaw: it is mathematically incapable of reaching optimal solutions located in non-convex (or "dented") regions of the Pareto front.

The Convexity Trap

To understand why linear weighting fails, imagine the Pareto front as a physical boundary in objective space. Linear scalarization acts like a flat board being pushed against this boundary. If the boundary is convex (bowed outward), the board can touch every point by changing its angle (the weights).

But if the boundary is non-convex (bowed inward), the board will "jump" across the gap, touching only the edges of the dent and missing all the optimal trade-offs inside it. In fields like shared autonomy, Jie Fu 0002 notes in Pareto efficiency in synthesizing shared autonomy policies with temporal logic constraints that these missed points often represent the most desirable compromises between human effort and robot performance.

The Tchebycheff Alternative

Instead of a weighted sum, researchers are increasingly turning to Tchebycheff (or Chebyshev) scalarization. Rather than pushing a flat board against the front, Tchebycheff scalarization minimizes the maximum weighted distance to an "ideal" (but unreachable) reference point.

Geometrically, this replaces the flat "board" of linear weighting with an L-shaped contour. Because this corner can be tucked into any nook or cranny of the objective space, it can "touch" any point on the Pareto front, regardless of whether the region is convex, concave, or jagged.

As Daqing Yi, Michael A. Goodrich, and Kevin D. Seppi demonstrate with MORRF*, this approach produces a much more diverse and uniform set of solutions. It ensures that the resulting paths or models represent the full spectrum of possible trade-offs rather than clustering around a few easy-to-find convex points.

Making it Differentiable

Despite its power, the classic Tchebycheff function has a major drawback for modern AI: the max operator is non-differentiable, making it difficult to use with standard gradient descent.

To solve this, Xi Lin and colleagues introduced Smooth Tchebycheff Scalarization for Multi-Objective Optimization. By applying a smoothing operator to the Tchebycheff formula, they created a differentiable objective that retains the ability to find non-convex solutions while remaining compatible with standard optimizers. This "smooth" version has proven highly effective in:

  • Model Merging: Weiyu Chen and James T. Kwok use it in Pareto Merging to merge multiple task-specific models into a single flexible model that can navigate task priorities at runtime.
  • Federated Learning: In PraFFL, Ming Tang and Rongguang Ye employ Tchebycheff scalarization to allow clients to dynamically adjust their preference for fairness versus accuracy without retraining the entire network.
  • Large Language Model Alignment: Zhiyao Zhang and Jia Liu apply it to align LLMs across conflicting human preferences in A Tale of Two Problems.

Learning the Whole Front at Once

The ultimate goal of this research direction is to move away from training a new model for every specific weight. By combining Tchebycheff objectives with Hypernetworks, researchers like Aviv Navon and Gal Chechik in Learning the Pareto Front with Hypernetworks have developed systems that learn the entire Pareto front in a single training run.

In these architectures, a user inputs their desired preference vector, and the hypernetwork instantly generates the parameters for a model that sits exactly on that point of the trade-off curve. Because they use Tchebycheff-based logic, these hypernetworks can accurately map even the most complex, non-convex frontiers, providing a "continuous" menu of optimal choices for end-users.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_tchebycheff_scalarization_for_non_convex_b10675e3,
  title = {Tchebycheff Scalarization for Non-Convex Pareto Fronts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/tchebycheff-scalarization-for-non-convex-pareto-fronts/rdr_b10675e320af401baf6600f70cb3caf3},
  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_tchebycheff_scalarization_for_non_convex_b10675e3_v1,
  title = {Tchebycheff Scalarization for Non-Convex Pareto Fronts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/tchebycheff-scalarization-for-non-convex-pareto-fronts/rdr_b10675e320af401baf6600f70cb3caf3/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}