📈 economics

QUBO-Based Optimization of Social Indicator Configurations for Working-Age Population Growth

This paper presents a novel QUBO-based framework using Japanese municipal data to model and optimize social indicator configurations for working-age population growth, demonstrating that quantum and classical annealing methods can identify optimal policy scenarios while capturing nonlinear indicator interactions in an interpretable manner.

Original authors: Hayate Wada, Seiya Miyamoto, Tatsumasa Ogawa, Kazuki Uehara, Masaru Hitomi, Masayuki Ohzeki

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Hayate Wada, Seiya Miyamoto, Tatsumasa Ogawa, Kazuki Uehara, Masaru Hitomi, Masayuki Ohzeki

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Puzzle of People and Places

Imagine you are trying to bake the perfect cake, but instead of flour and sugar, your ingredients are things like "how many people live in a town," "how old they are," and "what kind of jobs they have." In the world of science, this is a bit like studying how different parts of a society fit together to make a community thrive. For decades, scientists have known that when a town loses its working-age population (people roughly between 15 and 64 years old), it faces a tough future. It's like a sports team losing its star players; the whole team struggles to keep playing.

But here is the tricky part: you can't just fix one thing. You can't simply say, "If we just lower the average age, everything will be great!" or "If we just add more foreign residents, the problem is solved!" The real world is messy. These factors mix together in complicated ways, like a recipe where adding a pinch of salt changes how the sugar tastes. To figure out the best combination, researchers use a special kind of math called optimization. Think of it as a super-smart search engine that doesn't just look for one answer, but explores millions of possible "recipes" to find the one that works best. Recently, a new type of computer technology called quantum annealing has been introduced to help solve these massive puzzles faster, acting like a magical compass that can jump over hills and valleys in a landscape of possibilities to find the lowest point (or the highest peak, depending on what you want).

The Paper's Big Idea: Finding the Perfect Social Recipe

In this study, a team of researchers from Tohoku University and other institutions decided to tackle the problem of shrinking working-age populations in Japan. They wanted to know: If we could magically tweak a town's social statistics, what combination of changes would lead to the most growth in the working-age population?

Instead of just guessing, they built a digital "surrogate model." Imagine this as a crystal ball that isn't magic, but is trained on real data. They took data from 1,855 Japanese towns, looking at ten different social indicators from the 2010 census (like population density, average age, and the ratio of different types of workers) and seeing how those related to population growth between 2010 and 2020. They turned these continuous numbers into a simple "one-hot" code, like turning a dial with four settings for each indicator.

The result was a quadratic equation—a mathematical formula that captures not just how each indicator works on its own, but how they interact with each other. It's like realizing that "low average age" is good, but "low average age" plus "a moderate number of primary industry workers" is even better. The model was surprisingly accurate, predicting population growth with a correlation coefficient of 0.84 and an average coefficient of determination ( $R^2$ ) of 0.76. This means the mathematical recipe they created matched the real-world outcomes quite well, though the authors are careful to note this is a tool for exploring possibilities, not a crystal ball that guarantees the future.

The Optimization Hunt: Quantum vs. Classical

Once they had their recipe, the team needed to find the absolute best combination of settings. This is where the "QUBO" (Quadratic Unconstrained Binary Optimization) framework comes in. They turned their growth model into a puzzle where the goal is to find the specific mix of the ten indicators that maximizes the predicted growth.

To solve this, they used three different "searchers":

Gurobi: A powerful, exact computer solver that acts like a meticulous librarian who checks every single book to find the perfect one.
Simulated Annealing: A classical computer method that mimics the cooling of metal, slowly settling into a good solution.
Quantum Annealing: A method using a D-Wave Advantage2 system, which uses quantum physics to explore the solution space in a unique way.

All three methods agreed on the exact same best configuration! The optimal "recipe" for growth involved:

The highest level of population density.
The lowest level of average age.
The highest level of sex ratio (males per 100 females).
The highest level of foreign residents.
The highest level of single-person households.
The second level of three-generation households.
The lowest level of primary industry workers.
The lowest level of secondary industry workers.
The highest level of tertiary industry workers.
The highest level of daytime-nighttime population ratio.

The predicted growth rate for this perfect configuration was approximately 6.35. However, the researchers didn't just stop at the "perfect" answer. They noticed that the annealing methods (both quantum and simulated) also found many other good, feasible configurations that were slightly less than perfect but still very promising. This is crucial because, in the real world, a town might not be able to instantly become the "perfect" configuration due to budget or geography. Having a list of "good enough" alternatives is often more useful than just one theoretical ideal.

The Twist: It's All About the Mix

To see how this model works in practice, the team picked three real towns (anonymized as A, B, and C) that had similar population growth rates. They then played a "what if" game: What happens if we just change one thing in these towns?

They found something fascinating. Changing a single indicator didn't always help. For example, in some towns, moving the "daytime-nighttime population ratio" closer to the "optimal" level actually lowered the predicted growth. This proved that the indicators are deeply connected. You can't just fix one part of the puzzle in isolation; the effect of changing one thing depends entirely on what the other things are doing.

The study also highlighted a specific, strong interaction: the combination of a low average age and a moderate ratio of primary industry workers was a powerful driver for growth. This suggests that the magic isn't just in having young people or farmers, but in having the right mix of them.

What This Means (and What It Doesn't)

The authors are very clear about what their work is and isn't. They state that this is a methodological demonstration, not a causal proof. In other words, they aren't saying, "If you do X, Y will definitely happen." Instead, they are saying, "Our model suggests that these combinations are associated with growth." Real-world policy is complicated, and actually changing a town's demographics involves laws, economics, and human behavior that this math model doesn't capture.

However, the study successfully shows that we can take complex social data, turn it into an interpretable math model, and use advanced optimization tools (including quantum computing) to explore different scenarios. It provides a transparent way to ask, "What if we tried this combination?" and get a data-driven answer. While the quantum computer didn't necessarily beat the classical one in this specific, small-scale test, the fact that they all found the same answer gives confidence that the method works. The real value lies in having a tool that can generate a list of feasible, optimized scenarios for policymakers to discuss, rather than just staring at a single number and hoping for the best.

Technical Summary: QUBO-Based Optimization of Social Indicator Configurations for Working-Age Population Growth

Problem Statement
The decline of the working-age population poses a critical challenge for regional sustainability, particularly in aging societies like Japan. While demographic changes are known to correlate with multiple socioeconomic indicators, translating these associations into interpretable, optimization-ready models remains difficult. Existing literature often lacks a framework that can simultaneously handle the non-linear interactions between indicators and the combinatorial nature of policy constraints (where improving one indicator may require complementary changes in others). Furthermore, identifying the optimal joint configuration of discretized policy levels constitutes a high-dimensional discrete search problem that grows exponentially with the number of indicators.

Methodology
The authors propose a methodological framework that integrates quadratic surrogate modeling with Quadratic Unconstrained Binary Optimization (QUBO) to explore social indicator configurations associated with working-age population growth. The approach consists of three main stages:

Data Preprocessing and Encoding:
- Data Source: The study utilizes 2010–2020 municipal data from Japan, covering 1,855 municipalities (excluding merged/dissolved entities and those affected by the Great East Japan Earthquake).
- Variables: Ten social indicators from the 2010 Population Census (e.g., population density, average age, sex ratio, industry worker ratios) serve as explanatory variables. The target variable is the working-age population (15–64) growth rate from 2010 to 2020.
- Discretization and One-Hot Encoding: Continuous indicator values are discretized into four levels based on quartiles. These levels are then encoded into binary variables using a one-hot scheme ( $x_{in} \in \{0,1\}$ ), where exactly one level per indicator is active. This transforms the problem into a binary vector space $\vec{x} \in \{0,1\}^{40}$ .
Quadratic Surrogate Model Construction:
- A quadratic function $G(\vec{x})$ is constructed to regress the growth rate on the binary-encoded indicators. The model includes linear terms (individual indicator contributions) and quadratic interaction terms (pairwise associations between indicator levels).
- Parameter Estimation: Model parameters are estimated using Maximum A Posteriori (MAP) estimation with an $L_2$ regularization penalty (equivalent to regularized least squares). The resulting model captures both individual effects and pairwise compatibilities between social conditions.
QUBO Formulation and Optimization:
- The surrogate model is converted into a QUBO cost function $H(\vec{x}) = -G(\vec{x}) + \alpha \sum_{i} (\sum_{n} x_{in} - 1)^2$ . The first term represents the objective (maximizing growth), and the second term enforces the one-hot constraint (ensuring only one level per indicator is active) via a penalty coefficient $\alpha$ .
- Optimization Solvers: The QUBO problem is solved using three approaches:
  - Gurobi: Used as an exact reference optimizer.
  - Simulated Annealing (SA): Implemented via the Neal sampler.
  - Quantum Annealing (QA): Executed on the D-Wave Advantage2 system.
- Sensitivity Analysis: A "single-indicator flip" analysis is performed on selected municipalities to evaluate how predicted growth rates change when a single indicator is modified while others remain fixed.

Key Results

Predictive Performance: The quadratic surrogate model achieved strong predictive performance on test sets, with a correlation coefficient of 0.84, an average coefficient of determination ( $R^2$ ) of 0.76, and a mean squared error of 33.36 across 20 train-test splits.
Interpretability: The coefficient matrix revealed interpretable structures. The largest positive diagonal contribution ( $Q_{5,5} = 1.75$ ) corresponded to the lowest quartile of average age, aligning with the expectation that younger demographics support working-age growth. Significant off-diagonal interactions were identified, such as a strong positive interaction ( $Q_{5,26} = 1.30$ ) between the lowest average-age quartile and the second quartile of the primary-industry worker ratio.
Optimization Consistency: All three solvers (Gurobi, SA, and QA) identified the same optimal feasible configuration with a predicted growth rate of $G_{opt} \approx 6.35$ . The optimal configuration included high population density, low average age, high foreign resident ratios, and specific industry distributions.
Sampling and Suboptimal Solutions: While SA generated 100% feasible samples, QA produced 39.4% infeasible samples due to embedding constraints on the D-Wave hardware. However, both annealing methods successfully sampled feasible suboptimal configurations with varying growth rates, providing a landscape of candidate scenarios beyond the single global optimum.
Local Sensitivity: Single-indicator flip analyses demonstrated that modifying one indicator does not always improve the predicted growth rate; the effect depends on the configuration of other indicators. For instance, moving an indicator toward its "optimal" level in isolation sometimes decreased the predicted growth, highlighting the necessity of joint optimization.

Significance and Claims
The paper positions this work as a methodological demonstration rather than a causal evaluation of policy interventions. Its primary contributions are:

Framework Integration: It successfully integrates subnational demographic policy relevance, multifactor quantitative modeling, and QUBO-based combinatorial optimization.
Interpretability: It provides a transparent, interpretable quadratic objective function that links municipal social statistics to optimization-ready models, explicitly visualizing both individual contributions and pairwise interactions.
Scenario Generation: The framework serves as an exploratory tool for policy discussion. By generating feasible candidate configurations (including suboptimal ones), it offers a set of realistic target states that may be more achievable in practice than a single global optimum, accounting for budgetary and institutional constraints.
Validation of QUBO: The study validates the consistency of QUBO formulations for this class of problems, showing that annealing-based methods can recover exact optima and explore the solution landscape effectively, even if computational advantages over exact solvers are not claimed for this specific problem size.

The authors emphasize that the results represent associations learned from data, not causal estimates, and that actual policy implementation would require further institutional and causal analysis.