Cost Functions in Economic Complexity

TL;DR

This work reframes the Economic Complexity Index (ECI) and Economic Fitness and Complexity (EFC) as cost-minimization problems, revealing that ECI minimizes a Laplacian-based energy and EFC minimizes a strictly convex potential (in transformed variables $Z_i=-\log V_i$) with regularization $\delta$, thereby linking economic complexity to spectral graph theory and optimal transport. It proves global convergence for regular graphs and local convergence for general graphs, establishes the uniqueness of the EFC solution via convexity, and introduces a gradient-descent reformulation that accelerates computation. The authors validate the framework on synthetic and real networks (including Zachary’s karate club and UN COMTRADE data), showing distinct energy landscapes: ECI energy is smooth along the spectrum while EFC energy highlights structurally fragile links, informing potential attack or pruning strategies and offering a unified, interpretable view of network organization and resilience. The results advance theoretical foundations, enable faster algorithms (including Sinkhorn-type methods), and generalize economic complexity analyses to arbitrary graphs and broader network contexts, with implications for infrastructure, ecology, and regional development.

Abstract

Economic complexity algorithms aim to uncover the hidden capabilities that drive economic systems. Here, we present a fundamental reinterpretation of two of these algorithms, the Economic Complexity Index (ECI) and the Economic Fitness and Complexity (EFC), by reformulating them as optimization problems that minimize specific cost functions. We show that ECI computation is equivalent to finding eigenvectors of the network's transition matrix by minimizing the quadratic form associated with the network's Laplacian. For EFC, we derive a novel cost function that exploits the algorithm's intrinsic logarithmic structure and clarifies the role of the regularization parameter in its non-homogeneous version. Additionally, we establish the existence and uniqueness of its solution, providing theoretical foundations for its application. This optimization-based reformulation bridges economic complexity and established frameworks in spectral theory, network science, and optimization. The theoretical insights translate into practical computational advantages: we introduce a conservative, gradient-based update rule that substantially accelerates algorithmic convergence, with potential implications for a broader class of algorithms, including the Sinkhorn-Knopp method. Finally, we apply the energetic framework to a real-world trade network, demonstrating how link-wise energy provides a direct way to identify structurally relevant and vulnerable regions of the export matrix, thus complementing and enriching standard economic complexity analyses. Beyond advancing our theoretical understanding of economic complexity indicators, this work opens new pathways for algorithmic improvements and extends applicability to general network structures beyond traditional bipartite economic networks.

Figures (4)

• Figure 1: Energetic landscape of the fitness algorithm.a) Cost function from Eq. (\ref{['eq:efc_cost_general']}) for regular complete graphs with varying number of nodes $N$ and $K = N-1$. The potential $U(V)$ is plotted as a function of a single variable $V$, since by symmetry all node variables can be set equal. The curves for different values of $N$ demonstrate the convexity of the potential. Each curve exhibits a global minimum corresponding to the fixed point in Eq. (\ref{['eq:fixed_point_regular']}), which for small $\delta$ can be approximated as $V^* \approx \sqrt{K}$. b) Cost function around the fixed point $V^*$ for the ZKC network zachary1977information, plotted as a function of the fitness value $V_i$ for specific nodes. Each curve shows how the potential varies around the fixed point $V_i^*$ for that node. Nodes with higher fitness display larger values of $V_i^*$, i.e., their minima lie farther to the right. The legend reports the node degrees $d$ for comparison. As shown, the highest fitness nodes contribute most to the total potential, while low-fitness nodes provide only local corrections. Each curve exhibits a unique, global minimum. The curves are convex, though the logarithmic scale on the vertical axis may create a misleading appearance of concavity in some regions.
• Figure 2: Convergence of trajectories in the phase space.a) Bidimensional phase space of variables $V_1$ and $V_2$ for a complete graph with $N = 10$. The colormap shows the value of the potential, with the global minimum corresponding to the fixed point $(V_1^*, V_2^*) \approx (3, 3)$ (Eq.(\ref{['eq:fixed_point_regular']})). White points indicate several initial conditions, while the lines show the convergence of the original map defined in Eq. (\ref{['eq:explicit_nhefc']}). Trajectories are shown every ten steps. b) Convergence of several trajectories, starting from different initial conditions, toward the fixed point. Only one variable, $V$, is plotted, since by the symmetry of the graph, all node variables are equivalent. Trajectories are shown every two steps, due to the oscillatory behavior of the map. c) Bidimensional phase space of $V_1$ and $V_2$ for the ZKC network zachary1977information. The global minimum of the potential corresponds to the fixed point $(V_1^*, V_2^*) \approx (3.2, 2.2) \times 10^3$. White points indicate various initial conditions, and the lines show the trajectories under Eq. (\ref{['eq:explicit_nhefc']}), plotted every two steps. d) Convergence of several trajectories to the fixed point for variables $V_1$ and $V_2$, starting from different initial conditions.
• Figure 3: Efficient convergence of the conservative map compared to the original map.a) Trajectories of the potential $U(\vec{V})$ (Eq. (\ref{['eq:efc_cost_general']})) for a complete graph with $N = 10$. Following the original map, the dynamics exhibit strong period-2 oscillations, indicating algorithm instability. In contrast, the trajectory computed by descending the cost function gradient using the conservative map (Eq. (\ref{['eq:efc_map_z']})) shows no oscillatory behavior and converges more efficiently to the fixed point. It requires significantly fewer steps to reach a stable solution. b) Trajectories of the potential for the ZKC network. Also in this case, the conservative map converges considerably faster and avoids oscillations. Due to the formal analogy between EFC and the Sinkhorn--Knopp algorithm sinkhorn1967concerning, the conservative map offers a more efficient alternative for matrix renormalization.
• Figure 4: Energetic landscape of ECI and EFC on the trade network. Link-wise energy for the country--product matrix (UN COMTRADE, SITC Rev. 2, 2015), with rows and columns sorted by the corresponding scores. For visualization purposes, energy values are smoothly interpolated within the non-zero "frontier" of the matrix, i.e., the region to the left of the boundary beyond which all entries are zero. (a) ECI energy, $E_{cp}\propto M_{cp}(ECI_c-ECI_p)^2$, is lowest along the diagonal, consistent with its Laplacian structure, and increases for links between nodes with distant spectral positions. (b) EFC energy, $E_{cp} \propto M_{cp}/(F_c S_p)$, concentrates along the "frontier" of the matrix, highlighting structurally critical links involving low-fitness countries or low-simplicity products. ECI thus reflects smooth spectral variation, while EFC identifies points of fragility more relevant for targeted perturbations.