Portfolio optimization represents a fundamental computational problem in financial engineering and quantitative investment management, requiring the careful allocation of finite capital across candidate assets to achieve a balance between expected return and associated market risk. As the number of available financial instruments expands within increasingly complex global markets, constructing an optimal asset mix evolves from a conventional financial decision into an intricate, high-dimensional numerical optimization challenge. Classical mean-variance frameworks provide an analytical foundation by modeling the relationship between risk and expected returns through variance minimization under prescribed target returns. However, practical deployment introduces substantial operational hurdles that classical formulations often neglect, including real-world investment restrictions, non-negative asset weighting bounds, the complete prohibition of short-selling leverage, and background uncertainties such as income fluctuations or broader economic exposures. To evaluate asset allocations systematically under these realistic market dynamics, researchers frequently adopt the Sharpe ratio model, which measures excess portfolio return per unit of standard deviation risk. In a recent study published in Discover Computing, researchers Yuqing Yang, Yixuan Li, Mingyang Yu, and Shengkun Xu introduced a novel computational framework termed Quantum-Inspired Differential Evolution with Elite Perturbation to address the computational tractability and numerical hurdles of constrained portfolio optimization.
From a mathematical perspective, Sharpe ratio maximization across high-dimensional asset pools is difficult because of non-convex objective landscapes, extensive search spaces, and multiple coupled operational constraints. Conventional deterministic optimization algorithms frequently demand strict mathematical assumptions regarding the continuity, convexity, or differentiability of the underlying objective functions and constraints, severely restricting their utility in realistic financial settings. Conversely, stochastic metaheuristic algorithms have gained widespread adoption because of their derivative-free mechanics and capacity to traverse multimodal, high-dimensional search spaces without relying on analytical derivatives. Standard Differential Evolution has proven effective across diverse engineering and scientific domains because of its simple structure, limited control parameters, and strong global exploration capabilities. Nevertheless, when applied to complex financial portfolios with tens or hundreds of candidate stocks, standard Differential Evolution exhibits significant vulnerabilities, including premature stagnation in local optima, diminished search efficiency, and an inability to maintain feasibility when equality and bound constraints limit the valid search space. The research team tackled these limitations by combining a penalty-based unconstrained reformulation with three coordinated algorithmic enhancement strategies.
To simplify constraint handling while preserving the financial interpretation of risk-return optimization, the researchers reformulated the constrained Sharpe ratio model into an unconstrained single-objective problem. Following established penalty-function approaches, the equality constraint ensuring that portfolio weights sum to exactly one, alongside inequality constraints confining each individual asset weight between zero and one, were embedded directly into the objective function via quadratic penalty terms. In this formulation, violations of the full-investment condition and asset-weight bounds are squared, weighted by large penalty coefficients, and subtracted from the primary Sharpe ratio objective. In their baseline experimental configuration, the primary Sharpe ratio coefficient was fixed at one, while the three constraint penalty coefficients were collectively set to one million. Under this formulation, any candidate portfolio allocating negative capital or exceeding available funds suffers severe objective-value penalties, driving the population toward feasible asset allocations. This transformation streamlines numerical updates, enabling evolutionary operators to explore candidate weights freely while maintaining rigorous enforcement of budgetary and short-selling boundaries.
The newly proposed algorithm enhances standard Differential Evolution through three complementary components designed to operate across distinct phases of the optimization cycle. First, to mitigate the clustering and uneven spatial distribution characteristic of standard pseudorandom initialization, the team integrated a good point set initialization strategy paired with chaotic reverse learning. Developed theoretically from the number-theoretic methods of Hua Luogeng, the good point set selects deterministic, representative points distributed uniformly within a multi-dimensional unit hypercube, ensuring broad initial coverage of the decision space. To further enrich population diversity, a Logistic chaotic map with a chaotic coefficient fixed at four is recursively applied to generate opposite candidate solutions through reverse learning. By evaluating both the good point set population and its chaotic reverse counterparts, the algorithm selects the fittest candidate solutions to constitute the initial population. This coordinated initialization provides a structured starting distribution, reducing early search bias and establishing a well-distributed baseline for subsequent evolutionary updates.
The second structural enhancement is a quantum-inspired evolution strategy based on population information exchange, formulated to strengthen global exploration during intermediate iterations. Drawing mathematical inspiration from the physical delta-potential-well model in quantum mechanics, the search process represents candidate position updates through a probability distribution governed by the particle wave function. Rather than simulating physical quantum hardware, this probabilistic update mechanism executes on classical computing architectures to generate exploratory candidate steps using Monte Carlo sampling. An adaptive parameter dynamically scales the step size across iterations based on the distance between each individual and the population’s current best solution, while population information exchange factors balance individual responses to both the best and worst performing members. By integrating delta-potential-well sampling with centroid tracking, the algorithm maintains widespread exploration of promising search areas and reduces premature convergence across complex multidimensional landscapes.
The third core mechanism introduced in the algorithm is a dynamic elite pool integrated with a Cauchy-Gaussian hybrid perturbation operator, formulated to prevent late-stage stagnation around sub-optimal local attractors. In conventional Differential Evolution, candidate solutions tend to cluster tightly around the best-performing individual in later iterations, often causing the algorithm to stall if that solution occupies a local peak. To preserve late-stage diversity, the algorithm maintains an elite pool containing the three highest-ranking candidate solutions alongside their computed arithmetic mean position, providing four reference positions. At each iteration, reference positions are sampled randomly from this pool to guide updates rather than exclusively following the single best candidate. Furthermore, candidate elite solutions undergo perturbation through an adaptive combination of Cauchy and Gaussian random distributions. The long-tailed Cauchy distribution introduces substantial step perturbations that facilitate escape from local optima, while the concentrated Gaussian distribution executes refined local searches around promising candidate vectors, smoothly shifting emphasis from global perturbation toward fine-grained exploitation as iterations advance.
Prior to financial testing, the researchers evaluated the numerical search capabilities of their algorithm on standardized synthetic optimization benchmarks, specifically the CEC 2020 suite at ten and twenty dimensions and the CEC 2022 suite at twenty dimensions. Across these tests, the algorithm was benchmarked against seven competing metaheuristic methods: standard Differential Evolution, Advanced Differential Evolution, the related quantum-enhanced Dung Beetle Optimizer, and four cross-family swarm intelligence algorithms comprising Grey Wolf Optimizer, Whale Optimization Algorithm, African Vultures Optimization Algorithm, and Dung Beetle Optimizer. On the CEC 2020 suite, the proposed method achieved the best mean performance on eight out of ten functions at both ten and twenty dimensions, securing the lowest Friedman average ranks of 2.21 and 2.24, respectively. On the 20-dimensional CEC 2022 benchmark, it achieved the best results on eleven of twelve test functions with an overall Friedman rank of 1.17. Non-parametric Wilcoxon rank-sum testing confirmed that these performance advantages were statistically significant across the majority of benchmark landscapes, demonstrating robust run-to-run consistency.
To investigate the individual contributions of each enhancement, the authors conducted ablation experiments on the CEC 2020 benchmark suite by evaluating variants featuring only isolated components alongside leave-one-out configurations. These variants were compared against standard Differential Evolution and the complete algorithm. The results revealed that while each component contributed positively to overall solution accuracy, omitting the quantum-inspired population information exchange strategy produced the most severe performance deterioration, causing the mean ranking on the 20-dimensional CEC 2020 functions to drop from 2.700 to 6.100. Omitting the good point set initialization also substantially worsened performance across higher dimensions, whereas removing the elite perturbation pool produced a smaller relative decline in overall ranking. The ablation experiments confirmed that the superior performance of the full algorithm stems from the coordinated interaction of structured initialization, quantum-inspired exploration, and elite-guided perturbation rather than any isolated mechanism operating independently.
Following numerical validation, the team applied the algorithms to practical in-sample portfolio selection tasks constructed from constituent equities of China’s CSI 300 Index. Using daily closing price data collected between October 8, 2024, and November 13, 2024, four distinct portfolio optimization instances were established by selecting the top 20, 40, 60, and 100 stocks based on their average daily returns across that observation window. Each algorithm was executed across 30 independent runs with a population size of 50 for 100 iterations. In these financial experiments, the proposed algorithm achieved the highest mean penalized objective function value across all four asset instances. It attained the highest mean Sharpe ratio on the 40-stock and 100-stock portfolios (16.8050 and 11.4730, respectively), and the second-highest mean Sharpe ratio on the 20-stock and 60-stock portfolios (21.6827 and 13.6588, respectively), consistently maintaining the full-investment constraint sum close to exactly one with negligible standard deviations.
A critical observation emerging from the portfolio experiments was the sharp performance divergence between the proposed method and standard Differential Evolution as dimensionality scaled upwards. While standard Differential Evolution demonstrated competitive results on the 20-asset and 40-asset instances, its constraint satisfaction deteriorated severely in higher dimensions. On the 60-stock problem, standard Differential Evolution yielded an average weight sum of 1.0627, and on the 100-stock instance, its average weight sum ballooned to 3.7457, representing a massive violation of the full-investment condition. Because standard mutation and crossover operations alter individual components independently within box bounds, the algorithm struggled to locate candidate solutions satisfying the strict summation constraint within 100 iterations without specialized exploration mechanisms. In contrast, the proposed quantum-inspired framework maintained near-perfect constraint enforcement across all dimensions, recording an average weight sum of 0.9999 on the 100-asset portfolio while delivering superior risk-adjusted returns.
To assess the impact of penalty scaling, the authors performed an exploratory sensitivity analysis on the 20-asset instance across three penalty coefficient values: 100,000, 1,000,000, and 10,000,000. The findings demonstrated a clear operational trade-off: lower penalty weights permitted slightly higher unpenalized Sharpe ratios but produced wider variations in constraint compliance, whereas higher penalties enforced rigid constraint adherence at the expense of lower objective values. The chosen baseline coefficient of 1,000,000 successfully balanced near-exact constraint satisfaction with superior objective performance. However, the authors explicitly emphasized that all comparative findings were obtained under an equal-population and equal-iteration experimental protocol rather than equal objective-function evaluation budgets. Per iteration, the proposed algorithm requires between 202 and 203 objective evaluations for portfolio tasks compared to 101 for standard Differential Evolution and 51 for several swarm baselines. Furthermore, the financial evaluations were limited to an in-sample five-week period over pre-selected high-return equities, meaning the findings do not demonstrate out-of-sample trading performance or generalize to dynamic market conditions.
Subject of Research: Quantum-inspired differential evolution algorithm for constrained high-dimensional portfolio optimization
Article Title: Quantum-inspired differential evolution with Elite Perturbation for high-dimensional portfolio optimization
Article References: Yang, Y., Li, Y., Yu, M., & Xu, S. (2026). Quantum-inspired differential evolution with Elite Perturbation for high-dimensional portfolio optimization. Discover Computing, 29(1), Article 872. https://doi.org/10.1007/s10791-026-10814-6
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10814-6
Keywords: Differential Evolution, Portfolio Optimization, Quantum-Inspired Computing, Sharpe Ratio, Metaheuristic Algorithms, Modern Portfolio Theory, Good Point Set, Reverse Learning, Logistic Chaotic Map, Delta-Potential-Well Model, Cauchy-Gaussian Perturbation, Elite Pool Strategy, CEC 2020 Benchmark, CEC 2022 Benchmark, CSI 300 Index, Penalty Formulation, Constraint Handling
News Source: Gavin Prescott. (November 12, 2026). Quantum-Inspired Differential Evolution Enhances High-Dimensional Portfolio Optimization Under Complex Constraints. Scienmag.
