Center-Outward q-Dominance: A Sample-Computable Proxy for Strong Stochastic Dominance in Multi-Objective Optimisation¶
Conference: AAAI 2026 arXiv: 2511.12545 Code: None Area: Other Keywords: Stochastic multi-objective optimisation, stochastic dominance, optimal transport, hyperparameter tuning, center-outward quantiles
TL;DR¶
Building on the center-outward distribution function from optimal transport theory, this paper proposes the q-dominance relation as a computable approximation of strong first-order stochastic dominance (strong FSD). It proves that q-dominance over the full quantile range implies strong FSD, derives explicit sample-size thresholds for Type I error control, and validates practical utility in hyperparameter tuning ranking and noisy multi-objective optimisation.
Background & Motivation¶
Problem Definition¶
Stochastic Multi-Objective Optimisation Problem (SMOOP): When objective functions involve randomness (e.g., repeated runs of an algorithm in hyperparameter optimisation), each candidate solution corresponds to the distribution of a multivariate random vector. Comparing solutions requires ranking distributions rather than deterministic points.
Limitations of Prior Work¶
Scalarisation: Converts multi-objective problems into single-objective ones, discarding distributional structure and depending on predefined scalarisation functions.
Moment surrogates (mean/variance/CVaR): Replace stochastic objectives with summary statistics, ignoring dependence structure.
Weak FSD: Implemented via CDF comparisons, but overly permissive in dimension \(d>1\) — it may assert \(\mathbf{X} \succeq_w \mathbf{Y}\) even when \(\mathbf{X}\) is inferior to \(\mathbf{Y}\) under certain monotone utility functions.
Rioux et al. (2025) OT approach: Computes one optimal coupling per distribution pair, lacks a globally consistent reference frame for multi-distribution comparisons, and incurs computational cost that grows quadratically in the number of distribution pairs.
Gap Between Weak and Strong FSD¶
The paper presents a clear counterexample: in two dimensions, \(\mathbf{X}=\{(0,1),(1,0)\}\) weakly dominates \(\mathbf{Y}=\{(0,0),(1,1)\}\) in the CDF sense, but does not strongly dominate it — under utility \(u(x_1,x_2)=\exp(x_1+x_2)\), \(\mathbb{E}[u(\mathbf{X})] < \mathbb{E}[u(\mathbf{Y})]\).
Core Problem¶
A computable, globally consistent approximation of strong FSD is needed, with the following properties: no hyperparameters, sample-size guarantees, and a natural hierarchy of relaxations.
Method¶
Overall Architecture¶
Core Idea: Each distribution is mapped to the uniform distribution \(U_d\) on the unit ball via its center-outward distribution function, establishing a unified reference frame. Sample points sharing the same center-outward rank and sign are naturally paired, enabling quantile-by-quantile comparison.
Key Designs¶
1. Center-Outward Distribution/Quantile Functions¶
Based on McCann's theorem and the theory of Hallin (2017, 2021):
- Center-outward quantile function \(\mathbf{Q}^\pm\): The optimal quadratic transport map from \(U_d\) to target distribution \(P\), satisfying \(\mathbf{Q}^\pm \# U_d = P\).
- Center-outward distribution function \(\mathbf{F}^\pm\): Its inverse, mapping \(P\) to \(U_d\), satisfying \(\mathbf{F}^\pm \# P = U_d\).
Key properties: \(\|\mathbf{F}^\pm(\mathbf{Z})\|\) follows a uniform distribution on \([0,1]\) (rank); \(\mathbf{F}^\pm(\mathbf{Z})/\|\mathbf{F}^\pm(\mathbf{Z})\|\) is uniformly distributed on the unit sphere (sign); and the two are independent.
Empirical estimation: Constructed via an augmented grid on the unit ball (\(n_R\) concentric hyperspheres \(\times\) \(n_S\) unit vectors), with optimal assignment solved using the Hungarian algorithm (\(O(n^3)\)).
2. Definition of q-Dominance¶
For \(q \in [0,1)\), \(P_1 \succeq_q P_2\) if and only if for all \(\mathbf{y} \in \mathbb{C}_{P_2}(q)\) (the \(q\)-quantile region of \(P_2\)):
Intuition: \(\mathbf{y}\) is mapped to the unit ball via \(P_2\)'s distribution function, then mapped back to data space via \(P_1\)'s quantile function; the resulting point \(\mathbf{x}\) is componentwise no less than \(\mathbf{y}\).
Relationship to strong FSD (Theorem 5): If \(P_1 \succeq_q P_2\) holds for all \(q \in [0,1)\), then \(P_1\) strongly first-order stochastically dominates \(P_2\).
3. Finite-Sample Testing and Sample-Size Thresholds¶
Theorem 6 establishes: under bi-Lipschitz conditions, for any confidence level \(0<\delta<1\), there exists an explicit threshold \(n^*(\delta)\) such that when the sample size \(n \geq n^*(\delta)\), empirical q-dominance correctly reflects theoretical q-dominance with probability at least \(1-\delta\).
The expression for \(n^*(\delta)\), while complex (involving dimension \(d\), Lipschitz constants, the minimum gap \(\Delta\) of quantile functions, etc.), is fully explicit — no Monte Carlo or bootstrap estimation is required.
Loss & Training¶
This paper does not involve neural network training. The core algorithmic pipeline is:
- Compute the empirical center-outward quantile map for each distribution (computed once).
- Perform pairwise q-dominance tests on the shared augmented grid.
- Construct non-dominated fronts based on q-dominance using a sorting algorithm (Algorithm 1).
Key Experimental Results¶
Main Results: Multi-Objective Hyperparameter Optimisation Ranking¶
Seven HPO methods (Random Search, ParEGO, SMS-EGO, EHVI, MEGO, MIES, etc.) are compared on the YAHPO-MO benchmark:
| Method | Avg. Rank (q-dominance) | Avg. Rank (HVI) | Key Difference |
|---|---|---|---|
| MIES | Best (~2.0) | Moderate | Best under q-dominance |
| ParEGO | Moderate (~3.1) | Best | Best under HVI |
| Random Search | Worst | Worst | Consistent across both |
| Analysis Dimension | q-dominance Ranking | HVI Ranking |
|---|---|---|
| Method discriminability | More compact rankings | Larger spread |
| Methods not significantly better than RS | Only EHVI (1 method) | EHVI + SMS-EGO + MIES (3 methods) |
| Information retention | Retains full distributional information | Uses expected values only |
Ablation Study: Noise-Augmented ZDT Benchmark¶
| ZDT Problem | NSGA-II + q-dom | NSGA-II mean | NSGA-II single | Notes |
|---|---|---|---|---|
| ZDT1 | Fastest convergence | Slow | Slowest | Significant advantage of q-dominance |
| ZDT2 | Fastest convergence | Slow | Slowest | Same as above |
| ZDT3 | Fastest convergence | Slow | Slowest | Same as above |
| ZDT4 | All three similar | Similar | Similar | ZDT4 sensitive to noise |
| ZDT6 | Fastest convergence | Slow | Slowest | Significant advantage of q-dominance |
Experimental setup: \(\sigma=0.1\), population size 20, 200 generations, 64 samples per solution (\(n_R=n_S=8\)).
Key Findings¶
- MIES vs. ParEGO case study: On lcbench 167152, MIES yields better outcomes with high probability but carries slightly higher risk; HVI (expected value) favours ParEGO due to its lower variance. q-dominance correctly identifies MIES's distributional advantage.
- Under q-dominance ranking, meaningful rankings can still be produced when HVI fails to distinguish methods (i.e., when expected values are close).
- Replacing mean-based selection with q-dominance selection in NSGA-II significantly improves convergence on 5 out of 6 ZDT problems.
Highlights & Insights¶
- Globally consistent reference frame: Each distribution requires only one mapping to the uniform distribution; all comparisons share the same reference — avoiding the inconsistency across different couplings in the Rioux et al. approach.
- Natural hierarchy of relaxations: Different values of \(q\) correspond to dominance relations of varying strictness, and these relaxations are obtained at zero additional cost after the mapping is computed once.
- No hyperparameters: Unlike the method of Rioux et al., which requires tuning regularisation parameter \(\lambda\) and function \(h\).
- Explicit sample-size thresholds: Provides actionable finite-sample guarantees.
Limitations & Future Work¶
- \(O(n^3)\) complexity of the Hungarian algorithm limits scalability to large-sample settings.
- The expression for \(n^*(\delta)\) depends on Lipschitz constants and \(\Delta\), which are difficult to estimate accurately in practice.
- For dimension \(d \geq 5\), the available range of \(\theta\) narrows, potentially limiting practical utility.
- Experimental scope is limited (comparison with only a few baselines); the authors explicitly note that large-scale benchmarking is left for future work.
- The theoretical relationship with Rioux et al. (2025) has not been fully clarified.
Related Work & Insights¶
- The center-outward quantile theory of Hallin (2021) provides the core mathematical tool for this work.
- The entropic OT approach of Rioux et al. (2025) is used for approximate stochastic dominance testing but lacks global consistency.
- The selection mechanism in NSGA-II can be directly replaced by q-dominance, offering broad applicability.
- The concept of q-dominance may generalise to other settings requiring distribution comparison, such as robust optimisation and Bayesian optimisation.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ (The combination of center-outward quantiles and stochastic dominance represents a genuinely novel theoretical contribution.)
- Experimental Thoroughness: ⭐⭐⭐ (Limited experimental scale, but sufficient to validate the theory.)
- Writing Quality: ⭐⭐⭐⭐⭐ (Mathematically rigorous, clearly structured, with smooth theorem–proof–application flow.)
- Value: ⭐⭐⭐⭐ (Provides a theoretically more grounded foundation for stochastic multi-objective optimisation.)