Human population dynamics as a Bayesian inverse transport problem

Technical Summary: Human Population Dynamics as a Bayesian Inverse Transport Problem

Problem Statement
Many dynamical systems in physics, biology, and engineering involve non-equilibrium transport processes where the governing conservation laws are known, but the underlying constitutive relations (e.g., fertility and mortality rates) are latent, time-varying, and difficult to infer. In demography, conventional approaches face three primary limitations: (i) data-driven machine learning models often treat age-specific vital rates as independent parameters, failing to capture cross-cohort correlations; (ii) they frequently violate fundamental mass conservation laws; and (iii) they rely on empirical projections lacking physical interpretability. Furthermore, traditional cohort-component methods often lack native uncertainty quantification, while standard time-series forecasts treat cohort trajectories as independent statistical trends, ignoring the structural constraints of age-structured transport.

Methodology
The authors propose a unified Bayesian inverse transport framework that embeds Bayesian Neural Networks (BNNs) directly within exact partial differential equations (PDEs) governing age-time population transport.

1. Governing Equations: The system is modeled using the first-order transport PDE for age-structured populations:
   $$\partial_t \rho + \partial_a \rho = -\mu(a, t)\rho$$
   where $\rho(a, t)$ is the age distribution density and $\mu(a, t)$ is the mortality hazard rate. The system is closed by a birth boundary condition at $a=0$ driven by a fertility kernel $f(a, t)$, and a terminal age boundary at $A_{max}$.
2. Neural Parameterization: Instead of prescribing functional forms for fertility and mortality, the authors parameterize these constitutive laws using BNNs.
   • Fertility: Modeled as a Gaussian density scaled by a latent Total Fertility Rate (TFR), with parameters (mean, variance, TFR) output by a Multi-Layer Perceptron (MLP).
   • Mortality: For countries with complete data (Japan, South Korea), survival probabilities are modeled directly via a BNN. For data-sparse scenarios (China), survival is modeled as a perturbation around a standard baseline hazard function to ensure numerical stability.
3. Bayesian Inference: The problem is formulated as inferring the posterior distribution over BNN weights ($W$) and the initial population profile ($\rho_0$) given sparse, noisy observations ($Y$). The likelihood is constructed from heterogeneous demographic data (total population, births, deaths, and census age brackets).
   • Sampling: The authors employ the No-U-Turn Sampler (NUTS), a Hamiltonian Monte Carlo method, to sample the high-dimensional posterior.
   • Implementation: The framework utilizes JAX for high-performance compilation and automatic differentiation, with the transport equations compiled into a single XLA loop to accelerate gradient evaluation.
4. Physical Constraints: By embedding the BNNs within the exact transport PDE, every posterior sample automatically satisfies demographic mass conservation and cohort advection, ensuring physically consistent age profiles even under sparse observations.

Key Contributions

• Unified Framework: The work unifies historical reconstruction, forecasting, and uncertainty quantification into a single framework where missing data are recovered via Bayesian inference rather than interpolation.
• Mass Conservation: Unlike standard deep learning approaches, this framework enforces demographic consistency as a hard physical constraint, preventing violations of mass conservation.
• Uncertainty Propagation: The method guarantees that uncertainty in fertility at time $t$ propagates downstream along the characteristics of the advection equation, preserving structural cross-cohort correlations over time.
• New Metrics: The authors introduce information-theoretic metrics to quantify structural demographic transitions:
  • Total Demographic Entropy ($S_{total}$): Combines Shannon entropy with a "phase volume" potential to distinguish between young and old populations, identifying structural contraction.
  • KL Divergence from Lotka Equilibrium ($D_{KL}$): Measures the distance of the current age structure from the instantaneous stable attractor, serving as a physical measure of the demographic driving force.

Results
The framework was applied to three East Asian economies—China, Japan, and South Korea—characterized by rapid fertility collapse.

• Reconstruction: The model successfully reconstructed historical age-structured grids for all three countries, including China, despite the absence of annual age-structured data (relying only on total population, births, deaths, and decennial census brackets). The reconstructed distributions matched independent census observations with high precision.
• Latent Inference: The model inferred latent fertility trajectories without TFR as a training input, accurately capturing the timing and depth of South Korea's fertility crash and China's sub-replacement shift.
• Projections (Status Quo Scenario):
  • China: The working-age population (15-64) is projected to contract by ~25% by 2050. The Potential Support Ratio (PSR) is expected to approach 2.0 by 2049.
  • South Korea: Exhibits the most dramatic structural transformation, with a projected PSR of 1.34 by 2050 and an Elderly Dependency Ratio (EDR) exceeding 75%.
  • Japan: Shows a stable but heavily aged pillar structure, with a PSR already below 2.0.
• Entropy and Non-Equilibrium Dynamics: The analysis reveals that structural entropy contraction acts as a leading indicator of demographic decline, preceding peaks in working-age and total population by decades. South Korea exhibits an unprecedented, near-vertical drop in entropy starting in the 2000s, indicating a rapid transition to a low-entropy, hyper-aged state.

Significance and Claims
The paper claims that this framework provides a generalizable foundation for observing and forecasting non-equilibrium boundary dynamics across various fields. By treating transport equations as exact physical constraints and using BNNs to infer latent constitutive laws, the method offers a mathematically rigorous way to handle sparse data and propagate uncertainty.

The authors assert that the model does not merely predict population counts but resolves the full demographic pipeline, providing credible intervals for age-specific quantities critical for policy (e.g., labor supply, pension systems, education planning). The introduction of "demographic entropy" and "Lotka divergence" offers a new physical language to describe the structural rigidity and non-equilibrium nature of aging societies. The authors suggest this approach is applicable beyond human demographics to any system characterized by advective transport, time-varying boundary generation, and internal sinks, such as cellular flows, polymer degradation, or viral load transport.