Structure and dynamics jointly stabilize the international trade hypergraph

TL;DR

This work addresses how adverse fluctuations propagate in international trade by representing trades as a hypergraph (ITH) and identifying collapsed trades via $g_h(t) = \log \frac{w_h(t)}{w_h(t-1)} \le -1$. It introduces an SIR-like contagion on the ITH with an inhomogeneous infection rate $\beta_{h'h} = \beta A_{h'h} \frac{w_h'^{\alpha}}{\langle w^{\alpha} \rangle}$ and derives a critical threshold $\lambda_c = \frac{\langle k^2 \rangle \langle w^{\alpha} \rangle}{\langle k^2 w^{\alpha} \rangle}$ that governs global collapse, showing that negative $\alpha$ (weight-decay of infection) suppresses spread. Empirically, collapsed trades cluster and their incidence decays algebraically with trade volume, $c(w/\langle w\rangle) \sim (w/\langle w\rangle)^{-\zeta}$, with $\zeta \in [0.24,0.29]$ (2009: $\zeta \approx 0.17$). A positive degree–weight correlation further stabilizes the system by increasing $\lambda_c$ and reducing outbreaks, an effect that persisted through the 2008–2009 crisis; when dynamical correlations weakened, broader collapse occurred.Collectively, these results reveal a joint stabilization mechanism in a higher-order economic network and offer insights into managing systemic risk in complex adaptive systems.

Abstract

To understand how fluctuations arise and are distributed in international trade, a question crucial for economic risk assessment and policymaking, we analyze strong adverse fluctuations-collapsed trades-defined as individual trades with sharp annual volume declines. Adopting a hypergraph framework for a fine-scale trade-centric representation of international trade, we find that collapsed trades (hyperedges) are clustered and their occurrence decays algebraically with trade volume (weight), which suggests inhomogeneous, epidemic-like spreading of collapse in the international trade hypergraph. Modeling collapse propagation as a contagion process and analyzing its dynamics, we show that a positive degree-weight correlation and a volume-decaying collapse rate synergistically suppress the onset of global collective collapse. Notably, the degree-weight correlation persisted but the volume-decay of the collapse rate weakened during the 2008-2009 global economic recession, resulting in a broader collapse spread. Our study shows how the interplay between structure and dynamics stabilizes complex systems.

Figures (5)

• Figure 1: Collapsed trades in ITH. (a) Subhypergraph of ITH in 2009. Hyperedge colors indicate whether the trade is normal or collapsed. Circular vertices contain representative images, with boundary colors denoting vertex types; green for product categories and black for countries. Arrows represent the direction of product flow. (b) Proportions of collapsed neighbors around a collapsed ($\tilde{c}_{\text{collapsed}}$) and normal ($\tilde{c}_{\text{normal}}$) hyperedge in empirical data for each year. Points and error bars indicate mean and standard deviation. (c) Proportions of collapsed hyperedges, $c(w/\langle w\rangle)$, among hyperedges of given normalized weight for empirical data and for the Gaussian assumption, respectively. Inset: Power-law decay exponents $\zeta$ in Eq. \ref{['eq:zeta']} estimated from the empirical data by fitting over the range $w / \langle w\rangle=10^{-8} \times 2^{19}$ to $10^{-8} \times 2^{33}$. Error bars indicate standard error.
• Figure 2: Collapse propagation model. (a) Proportion of collapsed hyperedges in the final state $C$ as a function of $\lambda$ in the collapse propagation model on the empirical ITH with $\alpha=0$ (gray) and $\alpha=\alpha_{*}$ (colored) for selected years $t$. Points represent the Monte Carlo simulation results, and lines represent the mean-field solutions. (Inset) Schematic illustrations of the transition of a hyperedge $h'$ from $S$ state to $I$ state, due to infection from $h$ with rate $\beta_{h'h}$, and the transition from $I$ state to $C$ state with rate $\gamma$. (b) $\alpha_{*}$, (c) $\lambda_{0}$ and $\lambda_{*}$ estimated from Monte Carlo simulation results. (d, e) Mean proportions of collapsed neighbors around (d) a collapsed ($\tilde{c}_{\text{collapsed}}$) and (e) a normal ($\tilde{c}_{\text{normal}}$) hyperedge in Monte Carlo simulation results and empirical data for each year, respectively. (f) Proportion of collapsed hyperedges, $c(w/\langle w\rangle)$, among the hyperedges of given normalized weight $w/\langle w \rangle$ from Monte Carlo simulation results and empirical data, respectively, for selected years $t$. Points are the Monte Carlo simulation results, and lines are the empirical data. Error bars indicate standard error.
• Figure 3: Effects of dynamical inhomogeneity and structural correlation. (a) Outbreak size $C$ versus rescaled infection strength $\lambda$ on the empirical (Emp.) and randomized (Rand.) ITHs with $\alpha$ fixed at $-0.3$. Inset: Critical point $\lambda_\text{c}$ as a function of $\alpha$ on the empirical and randomized ITH for each year. Points are the Monte Carlo simulation results, and lines are the mean-field results. Error bars indicate standard error. (b) Mean normalized degree $k/\langle k \rangle$ of hyperedges of given normalized weight $w/\langle w \rangle$ in empirical data for each year. The dashed line represents the case where there is no correlation between hyperedge's degree and weight. Error bars indicate standard deviation.
• Figure B.1: Basic features of ITH. (a) Numbers of countries ($N_{\text{c}}$), product categories ($N_{\text{p}}$), and hyperedges ($N_{\text{h}}$) in the ITH for each year. (b) Distribution of the normalized weight ${w}/{\langle w\rangle}$ with the mean $\langle w\rangle = N_\text{h}^{-1} \sum_h w_h$ for each year. Inset: Mean weight $\langle w \rangle$ versus time. (c) Distribution of the normalized degree ${k}/{\langle k\rangle}$ with the mean $\langle k\rangle = N_\text{h}^{-1} \sum_h k_h$ for each year. Inset: Mean degree $\langle k \rangle$ versus time. (d) Proportion of collapsed trades under various thresholds for defining collapse in Eq. \ref{['eq:decline']}. (e) Distribution of the logarithmic change of trade volume $g$ shifted by the mean $\langle g\rangle_{\text{f}}$ for each year. The subscript 'f' denotes averaging over hyperedges exhibiting finite $g$. Inset: Mean logarithmic change $\langle g\rangle_{\text{f}}$ versus time. (f) The mean proportion $\tilde{c}_{\text{collapsed}}({w}/{\langle w\rangle})$ of collapsed neighbors around a collapsed hyperedge of given normalized weight ${w}/{\langle w\rangle}$. Error bars indicate standard error.
• Figure C.1: Proportion of collapsed hyperedges, $c(w / \langle w\rangle)$, among the hyperedges of given normalized weight $w / \langle w\rangle$ in empirical data. The collapse threshold in Eq. \ref{['eq:decline']}, set to $-1$ in the main text, is here varied as (a) $-0.5$, (b) $-2$, and (c) $-\infty$. Insets: The estimated power-law decay exponents $\zeta$ in the empirical data.