Quantum Multi-Objective Optimization

• Quantum multi-objective optimization is the integration of quantum algorithms with classical Pareto-efficient methods to compute non-dominated solutions across conflicting objectives.
• It leverages techniques such as quantum annealing, variational circuits, and quantum-inspired metaheuristics to explore high-dimensional, nonconvex optimization landscapes.
• Recent approaches utilize advanced scalarization methods and hypervolume optimization to enhance Pareto front recovery while addressing hardware and computational constraints.

Quantum Multi-Objective Optimization is the study and implementation of algorithms on quantum and quantum-inspired systems to efficiently compute the Pareto set or Pareto front for combinatorial optimization problems with multiple competing objectives. This field blends quantum algorithms—based on quantum annealing, variational circuits, amplitude amplification, or quantum-inspired heuristics—with multi-objective optimization principles from classical operations research, aiming to exploit quantum advantages in exploring high-dimensional, highly nonconvex, and conflicting objective landscapes.

1. Mathematical Foundations and Pareto Front Computation

A quantum multi-objective optimization problem is formalized over a discrete configuration space $\mathcal X$ (often $\{0,1\}^N$ for QUBO problems), with $m$ objectives $f_i:\mathcal X\to\mathbb R,~i=1,\dots,m$. Pareto dominance underpins the formalism: $x\in\mathcal X$ Pareto-dominates $y$ if $f_i(x)\le f_i(y)~\forall i$ and $f_j(x)<f_j(y)$ for at least one $j$. The Pareto front comprises all non-dominated solutions.

Quantum approaches operationalize scalarization (e.g., weighted-sum, Tchebycheff, min–max) to convert the vector-valued objective into Hamiltonians whose ground states correspond to Pareto-optimal configurations. For instance, the weighted sum $F_{\mathbf w}(x) = \sum_{i=1}^m w_i f_i(x)$ is encoded as a diagonal Hamiltonian $H(\mathbf w)$ in the computational basis, and its ground state yields a Pareto-optimal solution for the chosen weights (Baran et al., 2016, King, 3 Nov 2025). However, scalarization is provably limited to the convex hull of the front, missing unsupported (non-convex) Pareto-optima (Baran et al., 2016, Takahashi et al., 4 Jan 2026, Plehn et al., 12 Dec 2025).

Recent work exploits intermediate quantum states during annealing or variational evolution to access non-convex regions, as well as alternative scalarization techniques (e.g., Tchebycheff, min–max via $p$-norm approximations) to enhance front coverage and diversity (Takahashi et al., 4 Jan 2026, Plehn et al., 12 Dec 2025, Egginger et al., 15 Oct 2025).

2. Quantum Algorithmic Paradigms

Quantum multi-objective optimization leverages several algorithmic classes:

• Quantum Annealing (QA) and Adiabatic Quantum Computation (AQC):

The problem Hamiltonian $H(\mathbf w)$, built as a convex combination of objective Hamiltonians, is reached via adiabatic interpolation from a simple driver (typically a transverse field). Measurement at the end (or during intermediate states) samples low-energy configurations. Quantum annealing enables high-throughput parallel sampling and has demonstrated rapid convergence to optimal Pareto fronts in moderate-scale problems, outperforming both classical solvers and circuit-based QAOA implementations (King, 3 Nov 2025, Plehn et al., 12 Dec 2025, Takahashi et al., 4 Jan 2026). Notably, mid-anneal measurements or quench protocols sample superpositions containing low-energy excited states, which contribute to non-supported Pareto points (beyond weighted-sum scalarized optima) (Takahashi et al., 4 Jan 2026).

• Quantum Approximate Optimization Algorithm (QAOA) and Variational Circuits:

QAOA and variational paradigms for multi-objective optimization proceed either by scalarization (QAOA with weighted-sum Hamiltonians), or by entangling all objectives in the variational ansatz (QMOO, see Section 5). Parameter transfer and Pareto-archive enhancements boost coverage and convergence (Kotil et al., 28 Mar 2025, Ekstrom et al., 2023, Ekstrøm et al., 11 Feb 2026). Recent QMOO ansätze use layered application of cost-phase gates for each objective followed by universal mixers, directly targeting Pareto-superpositions and optimizing hypervolume (Ekstrom et al., 2023, Ekstrøm et al., 11 Feb 2026).

• Quantum-Inspired Metaheuristics:

Quantum-behaved particle swarm optimization (QPSO) and swarm-based heuristics utilize quantum-inspired update rules—such as sampling from delta-potential-well distributions or quantum-inspired chaotic/Levy steps—to traverse multi-objective landscapes where gradient evaluation is impractical. These algorithms maintain external Pareto archives and leverage crowding distance to maintain diversity (John et al., 2019, Hassani et al., 2016, Yu et al., 6 Feb 2025).

3. Hamiltonian Construction and Scalarization Techniques

The core mapping problem is to encode multi-objective criteria and constraints into a quantum-compatible Hamiltonian:

• Weighted-Sum and Convex Scalarizations:

The most natural approach is to form $H(\mathbf w) = \sum_{i=1}^m w_i H_i$, with $H_i$ the diagonal (or $k$-local) cost Hamiltonian for objective $f_i$. This maps directly to both QA and QAOA, but only produces supported Pareto points (Baran et al., 2016, King, 3 Nov 2025).

• p-Norm Min–Max (MOQA) Approximations:

The Multi-Objective Quantum Approximation (MOQA) framework encodes the (hard) min–max objective $h_{\max}(x) = \max_i f_i(x)$ as a $p$-norm approximation: $$h_{(p)}(x) = \left(\sum_{i=1}^m f_i(x)^p\right)^{1/p}$$, converging to $\max_i f_i(x)$ as $p\to\infty$. The corresponding Hamiltonian $H_{(p)}=\sum_{i=1}^m H_i^p$ is solved via standard QA/QAOA. Spectral-gap-based theorems provide guarantees on correctness for sufficiently large $p$ (Egginger et al., 15 Oct 2025, Egginger et al., 15 Oct 2025).

• Tchebycheff and Non-Convex Scalarization:

Data-driven approaches (DDTS) employ Tchebycheff scalarization $T_{\mathbf w}(x) = \max_{i} w_i |f_i(x)-z^*_i|$, allowing the recovery of non-convex front regions. These approaches preprocess training data, sample weights, and train QUBO surrogates accordingly (Plehn et al., 12 Dec 2025).

• Equality/Inequality Constraints as Objectives:

Constraints can be promoted to objectives or penalized via auxiliary Hamiltonians (penalty terms or slack variables). MOQA and related methods encode all constraint violations as additional objective Hamiltonians, allowing their inclusion in the $p$-norm or weighted-sum scalarization, thus unifying constraint satisfaction and Pareto optimization (Egginger et al., 15 Oct 2025, Egginger et al., 15 Oct 2025, Ayodele et al., 2022).

4. Quantum Hardware Implementations and Benchmarking

Quantum multi-objective optimization methods have been implemented and tested across different quantum hardware modalities and classical-quantum hybrid pipelines:

• Quantum Annealing Devices (D-Wave):

Employed for direct Ising Hamiltonian optimization, QA achieves high-throughput, parallel sampling for thousands of weight vectors in a single QPU call, which—in benchmarks—led to full Pareto-front recovery in seconds for three- and four-objective Max-Cut problems on 42 nodes, outperforming IBM QAOA and classical solvers by over 10–1000$\times$ in wall-clock time (King, 3 Nov 2025).

• Gate-Model Quantum Computers (IBM Quantum, IonQ):

QAOA-based workflows for multi-objective routing and MCOPs utilize parameter transfer from pre-computed training problems to larger graphs, with systematic sampling of weight vectors to approximate the supported front. Pareto-optimal points in both feasible and unsupported (concave) front segments have been reconstructed on 8- and 11-qubit IonQ Harmony and 42-qubit Falcon hardware (Chiew et al., 2023, Kotil et al., 28 Mar 2025). Variational approaches (QMOO) further utilize layered ansätze and hypervolume-driven optimization for front approximation (Ekstrom et al., 2023, Ekstrøm et al., 11 Feb 2026).

• Quantum-Classical Hybrids and Metaheuristics:

Job-shop scheduling and industrial multi-objective optimization tasks have been tackled by hybrid decomposition—allocating bin-packing to quantum annealers (as QUBO), followed by MILP-based task-scheduling—to more effectively sweep the tradeoff space and approximate the Pareto frontier (Sawamura et al., 5 Nov 2025). Swarm-inspired quantum metaheuristics maintain external Pareto archives, inject diversity through chaos or wavelike updates, and are applicable where direct gradient-based quantum algorithms are unavailable (John et al., 2019, Hassani et al., 2016, Yu et al., 6 Feb 2025).

Method	Hardware Implementation	Front Coverage
Weighted-sum scalar QAOA/QA	Both gate-based & annealing	Supported (convex)
Tchebycheff scalarization	Classical preprocessing, QA	Non-convex regions
MOQA ($p$-norm)	QAOA, QA (requires $p$-locality)	Min–max (discrete PF)
Variational QMOO (multi-H)	Gate-based (NISQ/qudit)	Full superposition

5. Variational Quantum Multi-Objective Algorithms and Hypervolume Optimization

The Variational Quantum Multi-Objective Optimization (QMOO) framework implements an ansatz that applies phase gates of all objective Hamiltonians in separate layers interleaved with generalized mixers, enabling the quantum state to be a superposition of (ideally) all (or weighted) Pareto-optimal solutions (Ekstrom et al., 2023, Ekstrøm et al., 11 Feb 2026). The algorithm iterates as follows:

1. Parameter Initialization: Start with a hardware-friendly initial state (often uniform superposition), and randomly initialize variational parameters.
2. Quantum Circuit Execution: Execute the multi-layer variational circuit; each layer applies a phase separator $e^{-i\gamma_k H_k}$ for objective $k$ alongside universal mixers.
3. Measurement and Candidate Selection: Perform multiple shots; select the top $N_{\rm most\,prob}$ distinct bitstrings by frequency or via Pareto-substitution.
4. Archiving and Pareto Filtering: Maintain an archive of non-dominated solutions gathered across iterations. Optionally, apply dominated-solution substitution to accelerate early hypervolume ascent (Ekstrøm et al., 11 Feb 2026).
5. Objective Evaluation and Hypervolume Computation: Evaluate the objective vectors for the candidates. Hypervolume with respect to a reference point is computed to objectively assess front coverage.
6. Classical Parameter Optimization: Classical optimizers (e.g., Powell, L-BFGS-B, evolutionary strategies) update circuit parameters to maximize the instantaneous hypervolume.

Empirical performance on benchmark RMNK-landscapes, bi- and multi-objective combinatorial problems confirms that tuned QMOO—with substitution and archiving—is competitive with well-established classical EMO algorithms (NSGA-II/III) for small to moderate problem sizes, and in high-epistasis (rugged) regimes can demonstrate resilience beyond purely classical methods (Ekstrom et al., 2023, Ekstrøm et al., 11 Feb 2026).

6. Practical Applications and Domain-Specific Adaptations

Quantum multi-objective optimization has been applied to diverse domains:

• Wireless Multihop Routing:

Pareto-optimal routing under BER, energy, and delay objectives in large networks is accelerated by quantum dynamic programming and amplitude-amplification, with back-tracing mechanisms ensuring nearly complete Pareto-front recovery at near-polynomial quantum complexity (Alanis et al., 2018).

• Robust Quantum Gate Design:

Multi-objective Pareto optimization (e.g., via evolutionary search and entropy weighting) provides hardware-robust pulse schemes for nonadiabatic holonomic quantum gates, balancing sensitivity to amplitude/phase errors and decoherence (Zhang et al., 24 Apr 2025).

• Quantum Control, Estimation, and Materials Design:

Multi-objective metaheuristics address the tradeoff between fidelity, energy, smoothness in control pulses (Yu et al., 6 Feb 2025), estimation precision versus probe state deformation (Gong et al., 2017), and competing materials properties, facilitated by data-driven QUBO surrogates and constraint-guided encodings (Plehn et al., 12 Dec 2025).

• Production, Scheduling, and Portfolio Optimization:

Hybrid QA-classical pipelines accelerate multi-objective scheduling; iterative scalarization and non-scalarization methods enhance diversity and hypervolume in resource allocation and finance (Sawamura et al., 5 Nov 2025, Ayodele et al., 2022, Ayodele et al., 2022).

7. Limitations, Open Problems, and Future Perspectives

While quantum multi-objective optimization demonstrates promising scalability, flexibility, and benchmarking superiority in some instances, there remain key challenges:

• Unsupported Pareto Solutions:

Weighted-sum scalarization is inherently limited; integrating or hybridizing Tchebycheff, min–max, and intermediate-state readouts is necessary for comprehensive front discovery (Baran et al., 2016, Takahashi et al., 4 Jan 2026, Plehn et al., 12 Dec 2025).

• Hamiltonian Complexity:

MOQA and $p$-norm approaches induce higher locality and exponential Pauli-term growth for large $p$, although practical instances require moderate $p$ with acceptable resource overhead (Egginger et al., 15 Oct 2025, Egginger et al., 15 Oct 2025).

• Parameter Optimization and Trainability:

Deep variational circuits introduce barren plateau risks; efficient layer-wise growth, warm starts, and classical archiving strategies partially mitigate this (Ekstrom et al., 2023, Ekstrøm et al., 11 Feb 2026).

• Constraint Integration and Scaling:

Systematic handling of equality and inequality constraints within Hamiltonian encodings remains complex, and, for penalized constraints, parameter tuning is nontrivial (Egginger et al., 15 Oct 2025, Egginger et al., 15 Oct 2025, Ayodele et al., 2022).

• Hardware and Algorithmic Limitations:

Hardware implementations (QA/QAOA) are limited by noise, shot count, qubit connectivity, and analog programming constraints. Efficient mid-circuit measurements, direct support for high-order interactions, and improved mixers would expand applicability (King, 3 Nov 2025, Takahashi et al., 4 Jan 2026, Ekstrom et al., 2023).

• Benchmarking and Evidence for Advantage:

On small to moderate sizes ($N\lesssim24$), classical EMO and quantum approaches are often comparable. QMOO appears robust to rugged landscapes where classical algorithms degrade, but evidence for quantum advantage at scale awaits larger, less classical-friendly instances and further hardware advances (Ekstrøm et al., 11 Feb 2026, Plehn et al., 12 Dec 2025).

Ongoing research targets integration of adaptive scalarization, hardware-native $p$-local couplers, systematic theoretical analysis of front-sampling properties, and expansion to black-box and non-convex objectives via data-driven quantum surrogates.

---

References

• (Alanis et al., 2018) Quantum-aided Multi-Objective Routing Optimization Using Back-Tracing-Aided Dynamic Programming
• (Baran et al., 2016) Multiobjective Optimization in a Quantum Adiabatic Computer
• (King, 3 Nov 2025) Multi-objective optimization by quantum annealing
• (Takahashi et al., 4 Jan 2026) Utilizing intermediate states in quantum annealing for multi-objective optimization
• (Ekstrom et al., 2023) Variational Quantum Multi-Objective Optimization
• (Ekstrøm et al., 11 Feb 2026) Improving Quantum Multi-Objective Optimization with Archiving and Substitution
• (Egginger et al., 15 Oct 2025, Egginger et al., 15 Oct 2025) A Rigorous Quantum Framework for Inequality-Constrained and Multi-Objective Binary Optimization
• (Plehn et al., 12 Dec 2025) Progress on Data-Driven, Multi-Objective Quantum Optimization
• (Kotil et al., 28 Mar 2025) Quantum Approximate Multi-Objective Optimization
• (Ayodele et al., 2022) Multi-objective QUBO Solver: Bi-objective Quadratic Assignment
• (Ayodele et al., 2022) A Study of Scalarisation Techniques for Multi-Objective QUBO Solving
• (Chiew et al., 2023) Multi-Objective Optimization and Network Routing with Near-Term Quantum Computers
• (Cheng et al., 2024) Variational quantum simulation of ground states and thermal states for lattice gauge theory with multi-objective optimization
• (Yu et al., 6 Feb 2025) Multi-Objective Mobile Damped Wave Algorithm (MOMDWA): A Novel Approach For Quantum System Control
• (Zhang et al., 24 Apr 2025) Multiobjective Optimization for Robust Holonomic Quantum Gates
• (Gong et al., 2017) Multi-objective Optimization in Quantum Parameter Estimation
• (Hassani et al., 2016) Multi-Objective Design of State Feedback Controllers Using Reinforced Quantum-Behaved Particle Swarm Optimization
• (John et al., 2019) Chaotic Quantum Behaved Particle Swarm Optimization for Multiobjective Optimization in Habitability Studies