The exact contents of citations.db main_text.text for this paper — one flattened LaTeX string, title through conclusion, appendix excluded, unmodified except for removing email addresses. This is what our citation measures are computed over.
103,078 characters
Sandpile Economics: Theory, Identification, and Evidence
\title{Sandpile Economics: Theory, Identification, and Evidence}
\author{
Diego Vallarino\thanks{Counselor to the Board of Directors, Inter-American Development Bank and IDB Invest, Washington, D.C. 20577. Email: \texttt{[email removed]}. The views expressed here are those of the author and do not necessarily represent those of the IDB Group.}
}
\date{\small Inter-American Development Bank \& IDB Invest\\[0.3em]
\textit{This version: \today}}
\maketitle
\begin{abstract}
\noindent
Why do capitalist economies recurrently generate crises whose severity is disproportionate to the size of the triggering shock? This paper proposes a structural answer grounded in the \emph{evolutionary geometry} of production networks. As economies evolve through specialization, integration, and competitive selection, their inter-sectoral linkages drift toward configurations of increasing geometric fragility, eventually crossing a threshold beyond which small disturbances generate disproportionately large cascades.
We introduce \emph{Sandpile Economics}, a formal framework that interprets macroeconomic instability as an emergent property of disequilibrium production networks. The key state variable is the \emph{Forman--Ricci curvature} of the input--output graph, capturing local substitution possibilities when supply chains are disrupted. We show that when curvature falls below an endogenous threshold, the distribution of cascade sizes follows a power law with tail index $\alpha \in (1,2)$, implying a regime of unbounded amplification.
The underlying mechanism is evolutionary: specialization reduces input substitutability, pushing the economy toward criticality, while crisis episodes induce endogenous network reconfiguration and path dependence. These dynamics are inherently non-ergodic and cannot be captured by representative-agent frameworks.
Empirically, using global input--output data, we document that production networks operate in persistently negative curvature regimes and that curvature robustly predicts medium-run output dynamics. A one-standard-deviation increase in curvature is associated with higher cumulative growth over three-year horizons, and curvature systematically outperforms standard network metrics in explaining cross-country differences in resilience.
\medskip
\noindent\textbf{JEL Codes:} B52, C32, D57, E32, G01, L16, O33.\\
\textbf{Keywords:} self-organized criticality, structural change, evolutionary economics, Ricci curvature, production networks, endogenous instability, disequilibrium dynamics, power law, global value chains.
\end{abstract}
\onehalfspacing
\section{Introduction}
\label{sec:intro}
The Schumpeterian tradition in evolutionary economics holds that capitalism
is an engine of \emph{endogenous structural change}: firms, sectors, and
technologies are selected or discarded not by design but by competitive
pressures that continuously reshape the architecture of production. This
tradition has generated powerful insights into growth, innovation, and
industrial dynamics \citep{Nelson1982,Dosi1988,Metcalfe1998}. Yet it has
been largely silent on a related question: does the same evolutionary
process that drives structural change also generate, endogenously, the
conditions for systemic instability?
This paper argues that it does, and that the mechanism is geometric.
As economies evolve through specialization and global value chain
integration, the input-output linkages connecting sectors become
progressively less redundant: fewer alternative suppliers exist for
each input, fewer routing paths exist for each inter-sectoral flow.
This reduction in local redundancy—captured formally by the
\emph{Ricci curvature} of the weighted input-output network—is not
the result of any particular failure but the equilibrium outcome of
competitive selection favoring efficiency over resilience. When
aggregate curvature crosses a bifurcation threshold $\kappa^*$, the
system undergoes a qualitative change in its dynamics: the distribution
of cascade sizes transitions from a thin-tailed regime with finite
mean to a power-law regime with $\alpha \in (1,2)$, where the expected
size of any individual disruption is infinite. This is the structural
analogue of the \citet{Bak1987} sandpile: the economy self-organizes,
through the ordinary operation of markets, into a state of
\emph{permanent fragility}.
\textbf{The evolutionary mechanism.} The key departure from the standard
macroeconomic view is that instability is not produced by large exogenous
shocks acting on a stable system. It is produced by the endogenous
evolution of the system's structure toward a critical configuration. Two
evolutionary forces drive this trajectory. First, competitive selection
favors firms that minimize input costs through specialization and
concentration of sourcing---reducing the local redundancy of input
connections and pushing $\bar{\kappa}$ toward more negative values.
Second, global value chain deepening---the dominant structural change in
the global economy over 2000--2014 \citep{Baldwin2012}---increases the
geographic reach and complexity of supply chains while systematically
reducing the number of alternative routing paths for any given input flow.
Both forces are self-reinforcing: efficiency gains from specialization
induce further specialization, and network concentration generates further
concentration through preferential attachment dynamics
\citep{Barabasi1999}. The empirical evidence corroborates this
trajectory: network mean curvature deteriorated monotonically from
$-21.0$ in 2000 to $-27.0$ in 2014, with the secular decline accelerating
in the post-crisis period rather than reversing---a signature of
path-dependent structural change, not cyclical adjustment.
\textbf{What Ricci curvature measures that centrality metrics miss.}
The evolutionary economics tradition has emphasized the importance of
\emph{structural} rather than \emph{positional} heterogeneity
\citep{Dosi1988,Metcalfe1998}: what matters for selection dynamics is
not merely where an agent stands in a network but the structural
constraints that govern its ability to adapt. Standard network metrics
such as betweenness centrality, PageRank, or the Herfindahl concentration
index capture positional properties—they tell you how much flow passes
through a node or how concentrated its suppliers are. Ricci curvature
captures structural adaptability: it measures whether, when a link is
disrupted, alternative paths exist through which inputs can be
rerouted. Sectors with deeply negative curvature are structurally
constrained to single sourcing strategies regardless of their centrality.
When subjected to disruptions, they cannot adapt—they topple.
This is why, in the empirical analysis of Section~\ref{sec:empirics},
Ricci curvature explains 30 to 267 times more variation in output dynamics
than the classical alternatives: it captures the dimension of
\emph{structural adaptability} that is central to evolutionary
selection dynamics and that is invisible to positional metrics.
\textbf{Disequilibrium and bifurcation.} A central implication of the
framework is that production networks are not equilibrium objects.
The secular deterioration of curvature documented in our data reflects
an economy that is perpetually out of the resilience-efficiency frontier:
competitive pressures select for efficiency gains that simultaneously
erode structural resilience, generating a slow but persistent drift toward
the bifurcation threshold $\kappa^*$. The 2008 financial crisis did not
interrupt this drift---network curvature continued to deteriorate after
2009---but rather triggered a \emph{regime shift} in the topology of
the network, analogous to the phase transitions documented in evolutionary
complex systems \citep{Arthur1999,Kauffman1993}. Post-crisis, the global
network settled into a new basin of attraction characterized by lower
average connectivity but higher fragility per remaining link, precisely
as predicted by the curvature dynamics of Appendix~\ref{app:minsky}.
\textbf{Relation to Schumpeterian themes.} The framework connects to
three Schumpeterian themes that define the scope of this journal.
First, \emph{structural change}: the secular deterioration of curvature
is a measurable dimension of the structural transformation of the global
economy, alongside the more commonly tracked indicators of sectoral
composition and technological intensity. Second, \emph{selection and
imitation}: the spread of global value chain participation across
countries—from the Czech Republic at $\bar{\kappa} = -17.7$ to Greece
at $\bar{\kappa} = -27.8$---reflects differential selection into
integration strategies with systematically different fragility profiles.
Third, \emph{innovation and disruption}: the empirical ex ante ranking
of Greece and Portugal as the most fragile economies in the panel for
2001--2013 demonstrates that curvature-based indicators can detect
structural vulnerability before it manifests in visible financial
distress, precisely the early-warning role envisioned in evolutionary
macroeconomic policy \citep{Dosi2010}.
\textbf{Three formal results.}
First (Theorem~\ref{thm:powerlaw}): in a stochastic inter-sectoral model
with curvature below $\kappa^*$, the stationary distribution of cascade
sizes satisfies $\mathbb{P}(S>s)\sim s^{-(\alpha-1)}$ with
$\alpha = 1 + \beta(1-\rho(A))^{-1}(1+|\bar{\kappa}|\bar{d}\left\|L\right\|_2)^{-1}$.
Second (Proposition~\ref{prop:minsky}, Appendix~\ref{app:minsky}): the
power-law regime corresponds to Minsky's Ponzi-finance phase; the
endogenous drift toward $\kappa^*$ is the topological counterpart of
Minsky's financial instability hypothesis. Third
(Theorem~\ref{thm:amplification}, Appendix~\ref{app:ge}): in the
Baqaee--Farhi general equilibrium, more negative curvature multiplicatively
amplifies the second-order damage from disruptions.
\textbf{Empirical evidence.}
Section~\ref{sec:empirics} validates the framework on four fronts.
\emph{Power-law test}: maximum-likelihood tail estimation
(Section~\ref{sub:powerlaw_test}) confirms $\hat{\alpha} = 1.83$
for the full WIOD sample, falling to $1.51$ in the most fragile
curvature quartile and rising to $2.14$ in the most resilient---the
monotone ordering required by equation~(\ref{eq:alpha}). The KS test
rejects exponential tails at the 1\% level.
\emph{Calibrated simulation} (Section~\ref{sub:simulation}): the
Sandpile Economy calibrated to WIOD reproduces the empirical tail
exponent within 2\% and matches tail probabilities to within one
percentage point.
\emph{Impulse response}: the local projection IRF
\citep{Jorda2005} yields $\hat{\beta}_3 = 0.001540$ ($t = 2.86$,
$p < 0.01$) at a three-year horizon, amplification factor $11.2\times$
at five years.
\emph{Horse race}: Ricci curvature outperforms five classical metrics
by factors of 4.6 to 267 on adjusted $R^2$.
\textbf{Paper organization.}
Section~\ref{sec:literature} positions the paper.
Section~\ref{sec:theory} develops the framework.
Section~\ref{sec:econometrics} describes identification.
Section~\ref{sec:empirics} reports findings.
Section~\ref{sec:policy} derives policy implications.
Section~\ref{sec:conclusion} concludes.
Appendices~\ref{app:minsky}--\ref{app:A} contain the Minsky
correspondence, GE amplification, and all proofs.
\section{Literature Review}
\label{sec:literature}
\subsection{Evolutionary Economics: Structural Change, Selection, and Dynamics}
\label{sub:evolutionary}
The evolutionary approach to economics, rooted in \citet{Schumpeter1934} and
formalized by \citet{Nelson1982}, treats the economy as a population of
heterogeneous agents subject to selection, imitation, and mutation rather
than a system converging to a representative-agent equilibrium. This
tradition has generated a rich body of work on industrial dynamics
\citep{Dosi1988}, technological paradigms \citep{Dosi1982}, and the
micro-foundations of macroeconomic fluctuations \citep{Silverberg1988}.
A recurrent theme is that structural change is path-dependent: the selection
of more efficient production technologies and organizational forms alters the
topology of inter-firm and inter-sectoral relationships in ways that cannot be
reversed at will.
The present paper contributes a geometric dimension to this tradition.
We show that the evolutionary selection of efficient supply-chain configurations
alters a specific topological property---Ricci curvature---in a systematic
direction: toward lower redundancy and higher fragility. This is not a
side effect of structural change; it is its topological signature. The
secular deterioration of network curvature from $-21.0$ in 2000 to $-27.0$
in 2014 documented in Section~\ref{sub:structural_breaks} is, from an
evolutionary standpoint, a measure of how far the competitive selection
process has progressed toward the bifurcation threshold at which systemic
instability becomes structurally inevitable.
\citet{Dosi2010} and \citet{Fagiolo2008} have argued that evolutionary
macroeconomics requires new aggregate indicators that capture the
structural properties of the economy rather than merely its
compositional ones. Ricci curvature is such an indicator: it measures
a structural property of the input-output network—the local redundancy of
supply relationships—that is invisible to standard sectoral composition
statistics but is, as we show, a more powerful predictor of economic
resilience than any classical network metric.
\subsection{Complexity Economics, Self-Organization, and Bifurcation}
\label{sub:complexity}
The complexity approach to economics---associated with
\citet{Arthur1999}, \citet{Kirman2011}, \citet{Farmer2009}, and the
Santa Fe tradition---treats macroeconomic phenomena as emergent properties
of decentralized agent interactions rather than as solutions to optimization
problems. Key concepts include self-organization, far-from-equilibrium
dynamics, phase transitions, and the endogenous generation of instability
\citep{Kauffman1993}. \citet{Farmer2009} argue explicitly that the economy
should be modeled as a complex adaptive system, with power-law distributions
and cascading dynamics as generic properties rather than anomalies.
\citet{Arthur1999} emphasizes lock-in, path dependence, and increasing
returns as the drivers of structural change in technology-intensive industries—
mechanisms that translate directly into curvature dynamics in our framework,
since path-dependent technological lock-in reduces the diversity of
input-sourcing relationships and pushes curvature toward $\kappa^*$.
The concept of self-organized criticality (SOC) introduced by \citet{Bak1987}
provides the physical substrate for our economic framework. In the BTW
sandpile model, a driven dissipative system spontaneously organizes into
a critical state where the distribution of avalanche sizes follows a
power law $\mathbb{P}(s) \sim s^{-\alpha}$. \citet{Bak1996} argued that SOC
is the generic attractor of complex adaptive systems, and \citet{Scheinkman1994}
translated this intuition into an economic model of sectoral complementarities.
Our contribution is to provide, for the first time, a fully formal derivation
of the power-law exponent $\alpha$ as a closed-form function of observable
input-output network properties, closing the gap between the metaphorical
use of SOC in economics and a rigorous, testable theoretical result.
\citet{DiGiovanni2014} provide direct empirical evidence that firm-level
shock propagation through supply chains exhibits superlinear amplification
consistent with SOC dynamics. \citet{Barrot2016} establish causal
evidence of upstream propagation using natural disasters as exogenous
sectoral disruptions. The power-law character of macroeconomic fluctuations
has been documented by \citet{Gabaix2011} for firm-size-driven aggregate
volatility and by \citet{Clauset2009} for a wide range of social and
economic phenomena.
\subsection{Network Economics and Production Networks}
\label{sub:networks}
The modern theory of production networks originates in the input-output
analysis of \citet{Leontief1941}. \citet{Acemoglu2012} provided the
foundational modern treatment, showing that network asymmetry causes the
law of large numbers to fail: microeconomic shocks to well-connected sectors
generate macroeconomic fluctuations. \citet{Baqaee2019} derived second-order
nonlinear terms showing that negative shocks are amplified more than positive
ones due to the curvature of the production possibilities frontier---a result
to which our Appendix~\ref{app:ge} provides a topological underpinning.
\citet{Carvalho2019} survey the literature on how network topology determines
whether sectoral disruptions become systemic events.
In financial networks, \citet{Allen2000} established the ``robust yet fragile''
property of dense interconnection, \citet{Gai2010} characterized threshold
dynamics in contagion, and \citet{Elliott2014} derived conditions for cascading
failures as phase transitions in network density. \citet{Glasserman2016}
provide upper and lower bounds on systemic loss under network clearing.
The regime-shift evidence from the global banking network---spectral radius
declining persistently from $\approx 0.075$ pre-2008 to $\approx 0.068$
post-2008 \citep{Minoiu2013}---corroborates the path-dependent structural
change predicted by our framework.
\subsection{Geometric Methods in Network Analysis}
\label{sub:geometry}
The application of Riemannian geometry to network analysis has accelerated
since \citet{Ollivier2009} introduced discrete Ricci curvature via optimal
transport. \citet{Lin2011} showed that negatively curved graphs exhibit
bottleneck edges through which information flow is concentrated and fragile.
\citet{Ni2019} demonstrated that aggregate Ricci curvature of equity
correlation networks falls sharply before market downturns, providing a
leading indicator of systemic events. \citet{Sandhu2016} showed that
Ricci curvature outperforms spectral methods in detecting phase transitions.
\citet{Saucan2019} and \citet{Weber2017} developed the Forman--Ricci
formulation that we employ throughout the paper.
The economic interpretation of negative curvature as \emph{structural
non-substitutability}---rather than merely as a topological property---is the
conceptual contribution of this paper to the geometric network literature.
In the evolutionary economics reading, negatively curved edges are those
for which competitive selection has eliminated alternative suppliers; they
are the topological footprint of specialization having proceeded past the
point of resilience.
\subsection{Evolutionary Mechanism Implied by the Curvature Trajectory}
\label{sub:mechanism}
The four strands of literature reviewed above converge on a single
evolutionary mechanism, which we state explicitly as the interpretive
framework for the empirical analysis.
\begin{enumerate}[label=(\arabic*)]
\item \textbf{Selection for efficiency reduces curvature.} Competitive
selection favors firms that minimize input costs through
specialization, reducing redundant sourcing relationships and
pushing edge-level curvature toward more negative values.
\item \textbf{Global value chain deepening accelerates the drift.}
Cross-border integration increases the reach of supply chains
while reducing the number of domestic substitutes for any
given input, compounding the curvature deterioration documented
by the WIOD data.
\item \textbf{The bifurcation threshold $\kappa^*$ is crossed endogenously.}
There is no coordination failure, no policy error, no exogenous
shock required for the transition. The economy drifts toward
$\kappa^*$ through the ordinary operation of market selection.
\item \textbf{Post-crisis dynamics are path-dependent, not cyclical.}
The evidence that curvature deterioration accelerated after 2009
rather than reversing indicates that crises do not restore the
pre-crisis structural configuration; instead they trigger
regime shifts to new topological basins of attraction with
altered fragility profiles.
\end{enumerate}
This mechanism is precisely what \citet{Schumpeter1939} called the
\emph{endogenous business cycle}---instability generated by the internal
dynamics of the capitalist economy rather than by external perturbations.
Sandpile Economics provides its first rigorous geometric formalization.
\subsection{Prior Work by the Author}
\label{sub:own}
\textbf{Nonlinear diffusion and spectral thresholds.}
\citet{Vallarino2026CNSNS} establishes a sharp spectral threshold
$\lambda_c = \gamma_p c_p/\rho(A+D)$ separating globally dissipative from
explosive regimes in stochastic diffusion-hazard systems on economic graphs.
In the present paper, this threshold is the structural counterpart of $\kappa^*$.
\textbf{Identification in nonlinear dynamic networks.}
\citet{Vallarino2026arXiv} shows that network interaction matrices are
identified if and only if their spectrum is sufficiently dispersed to generate
non-exchangeable covariance patterns. This identification condition
translates here into the requirement that the Leontief inverse have
heterogeneous eigenvalues---satisfied empirically by the degree heterogeneity
of WIOD production networks.
\textbf{Causal graph neural networks.}
\citet{Vallarino2025AIL} develops causal adjacency matrices $C_{vu}$
that disentangle genuine productive complementarity from historically
contingent concentration. In the Sandpile framework, negatively curved
edges are the topological signature of concentration that is historically
contingent---selection-driven bottlenecks, not technological necessity.
\textbf{Trade complexity in small open economies.}
\citet{Vallarino2025AEL} shows that tariff shocks trigger nonlinear
restructuring of trade networks in small open economies, exactly as the
sandpile model predicts: small perturbations generate disproportionate
structural change when the network is near criticality.
\section{Theoretical Framework}
\label{sec:theory}
\subsection{Production Network as a Weighted Directed Graph}
\label{sub:graph}
\begin{definition}[Production Network]
\label{def:network}
A \emph{production network} is a tuple $\mathcal{G} = (V, E, W, Z)$, where:
\begin{enumerate}[label=(\roman*)]
\item $V = \{1, \ldots, n\}$ is the set of sectors (nodes);
\item $E \subseteq V \times V$ is the set of input-output linkages (directed edges);
\item $W: E \to \mathbb{R}_{++}$ is the weight function with $W_{ij}$ denoting
the value of sector $j$'s inputs sourced from sector $i$ as a
fraction of $j$'s gross output;
\item $Z: V \to \mathbb{R}_{++}$ is the sectoral size function with $Z_i$
denoting gross output of sector $i$.
\end{enumerate}
The matrix $A = (W_{ij})_{n \times n}$, called the \emph{technical
coefficients matrix}, is assumed to be productive: $\rho(A) < 1$, where
$\rho(\cdot)$ denotes the spectral radius.
\end{definition}
Under \citeauthor{Leontief1941}'s (\citeyear{Leontief1941}) input-output
accounting identity, gross output satisfies:
\begin{equation}
\label{eq:leontief}
\bm{z} = A\bm{z} + \bm{d},
\end{equation}
where $\bm{z} = (z_1, \ldots, z_n)^\top$ is the vector of gross outputs
and $\bm{d} = (d_1, \ldots, d_n)^\top$ is final demand. Since $\rho(A) < 1$,
the Leontief inverse $L = (I - A)^{-1} = \sum_{k=0}^{\infty} A^k$ exists
and yields:
\begin{equation}
\label{eq:inverse}
\bm{z} = L\bm{d}.
\end{equation}
The $(i,j)$-th element $L_{ij}$ measures the total requirement of sector
$i$'s output (direct and indirect) per unit of sector $j$'s final demand.
Following \citet{Acemoglu2012}, define the \emph{influence vector}
$\bm{\ell} = \bm{1}^\top L / n$, where $\ell_j = \sum_i L_{ij}/n$ is
proportional to the first-order approximation of sector $j$'s contribution
to aggregate output.
\subsection{Discrete Ricci Curvature on Production Networks}
\label{sub:ricci}
We adapt the Ollivier--Ricci curvature \citep{Ollivier2009} to the
directed, weighted setting of production networks.
\begin{definition}[Ollivier--Ricci Curvature on $\mathcal{G}$]
\label{def:ricci}
For an edge $(i,j) \in E$, define the \emph{lazy random walk measure}
$\mu_i^{(\epsilon)}$ at node $i$ as:
\begin{equation}
\mu_i^{(\epsilon)}(k) =
\begin{cases}
\epsilon & \text{if } k = i, \\
(1-\epsilon)\, \dfrac{W_{ki}}{\sum_{\ell} W_{\ell i}} & \text{if } (k,i) \in E,\\
0 & \text{otherwise},
\end{cases}
\end{equation}
where $\epsilon \in [0,1)$ is the laziness parameter. The
\emph{Ollivier--Ricci curvature} of edge $(i,j)$ is:
\begin{equation}
\label{eq:ricci_edge}
\kappa(i,j) = 1 - \frac{W_1\!\left(\mu_i^{(\epsilon)}, \mu_j^{(\epsilon)}\right)}{d(i,j)},
\end{equation}
where $W_1(\cdot,\cdot)$ is the Wasserstein-1 distance and $d(i,j)$ is the
graph-theoretic shortest-path distance from $i$ to $j$. The
\emph{aggregate network Ricci curvature} is:
\begin{equation}
\label{eq:ricci_agg}
\bar{\kappa} = \frac{1}{|E|} \sum_{(i,j)\in E} W_{ij}\, \kappa(i,j).
\end{equation}
\end{definition}
\begin{remark}
Positive $\kappa(i,j)$ indicates that the neighborhoods of $i$ and $j$
are ``closer on average'' than $i$ and $j$ themselves—a signature of
robustness, local clustering, and redundancy. Negative $\kappa(i,j)$
indicates \emph{bottleneck} structure: the edge is a bridge that, if
severed, disconnects communities. In the production network context,
negative curvature on edge $(i,j)$ means that sector $j$'s suppliers
and sector $i$'s customers have few alternative trading paths, making
the link a systemic vulnerability.
\end{remark}
\begin{remark}[Economic interpretation: curvature as structural non-substitutability]
\label{rem:substitutability}
The mapping from Ricci curvature to economic fragility is most transparent
through three concrete sector-level contrasts from the WIOD data.
\textbf{Motor vehicles} ($\bar{\kappa} \approx -27.7$): a German car assembly
plant sources specialized steel stampings from a small cluster of tier-1
suppliers. No generic steel supplier can substitute at short notice, because
the stampings require proprietary tooling. The edge connecting the stamping
sector to assembly has few neighboring edges of comparable weight---the
penalization terms in equation~(\ref{eq:forman}) are small---and curvature is
deeply negative. When one stamping supplier is disrupted, the assembly plant
halts: a single-edge disruption generates an avalanche.
\textbf{Construction} ($\bar{\kappa} \approx -14.3$, most resilient sector):
a construction firm sources cement from multiple regional suppliers, sand and
gravel from dozens of quarries, and general labor from a thick local market.
Each input edge is embedded in a dense, weight-balanced neighborhood---many
alternative paths exist---and curvature approaches zero. A disruption to any
single supplier is absorbed by rerouting; no avalanche propagates.
\textbf{Forestry} ($\bar{\kappa} \approx -31.1$, most fragile sector):
primary wood sourcing is geographically concentrated and seasonally rigid;
no alternative exists for a specific species-grade-region combination at
short notice. The input network is a near-tree structure with curvature at
the bottom of the empirical distribution.
These contrasts illustrate the precise economic content of Definition~\ref{def:ricci}:
$\kappa_F(e) < 0$ measures the degree to which edge $e$ is
\emph{non-substitutable}---the degree to which no alternative routing path
exists in the Wasserstein sense. Theorem~\ref{thm:powerlaw} then establishes
that aggregate non-substitutability, $|\bar{\kappa}|$, drives the power-law
exponent $\alpha$: as $|\bar{\kappa}|$ grows, $\alpha$ falls, and the
probability of system-wide cascades increases without bound.
\end{remark}
\begin{remark}[Causal interpretation of curvature]
\label{rem:causal}
In the spirit of \citet{Vallarino2025AIL}, one can think of edges with
$\kappa(i,j) \ll 0$ as the structural analogues of spuriously
high-weight adjacencies in a correlational GNN: they concentrate stress
flows not because they represent genuine productive complementarity, but
because no alternative routing exists. Identifying which bottleneck edges
are causally load-bearing versus historically contingent is a natural
extension of the present framework.
\end{remark}
\subsection{Stochastic Shock Propagation and Avalanche Dynamics}
\label{sub:sandpile}
We embed the production network in a stochastic dynamical system inspired
by the BTW sandpile model \citep{Bak1987}. Let $h_i(t) \in \mathbb{R}_+$ denote
the \emph{stress level} of sector $i$ at discrete time $t$, interpreted
as the ratio of nominal debt-service obligations to operating cash flow—
the Minsky ratio introduced in Section~\ref{sub:minsky_theory}—with sector
$i$ toppling whenever this ratio crosses the critical threshold $h^*$.
\begin{definition}[Sandpile Economy]
\label{def:sandpile}
The \emph{Sandpile Economy} is the discrete-time stochastic dynamical
system on $\mathcal{G} = (V,E,W,Z)$:
\begin{equation}
\label{eq:sandpile_dynamics}
h_i(t+1) = h_i(t) + \xi_i(t)
- \mathbf{1}\bigl\{h_i(t) \geq h^*\bigr\}
\sum_{j:\,(i,j)\in E} W_{ij}\,\bigl(h_i(t) - h^*\bigr)
+ \eta_i(t), \quad i \in V,
\end{equation}
where: (i)~$\{\xi_i(t)\}_{i,t}$ are i.i.d.\ sector-specific input shocks
drawn from a distribution $F$ with mean $\mu_\xi > 0$ and upper tail
$\bar{F}(x) = \mathbb{P}(\xi > x)$ regularly varying with index $\beta > 1$;
(ii)~$h^* > 0$ is the \emph{critical threshold};
(iii)~when $h_i(t) \geq h^*$, sector $i$ \emph{fires}, redistributing
excess stress $W_{ij}(h_i-h^*)$ to each downstream sector $j$, which may
in turn fire; and (iv)~$\{\eta_i(t)\}$ is mean-zero idiosyncratic noise,
independent of $\{\xi_i(t)\}$.
\end{definition}
\begin{definition}[Avalanche and Avalanche Size]
\label{def:avalanche}
An \emph{avalanche} $\mathcal{A}(t_0,i_0)$ initiated at sector $i_0$ at
time $t_0$ is the minimal subset of $V$ such that: (a)~$i_0 \in \mathcal{A}$;
and (b)~if $i \in \mathcal{A}$ fires and $(i,j) \in E$ with
$h_j(t_0) + W_{ij}(h_i(t_0)-h^*) \geq h^*$, then $j \in \mathcal{A}$.
The \emph{avalanche size} is $S = |\mathcal{A}(t_0,i_0)|$.
\end{definition}
\noindent
The key object is the stationary distribution of $S$. We work under the
following standing assumptions.
\begin{assumption}
\label{ass:standing}
\begin{enumerate}[label=(A\arabic*)]
\item \label{ass:A1} $\rho(A) < 1$ (Leontief productivity);
\item \label{ass:A2} $\mathcal{G}$ is strongly connected (irreducibility);
\item \label{ass:A3} $\bar{F}(x) = x^{-\beta}\ell(x)$ for some $\beta > 1$
and slowly varying $\ell$;
\item \label{ass:A4} The aggregate Ricci curvature satisfies
$\bar{\kappa} \leq \kappa^* \equiv -\ln\rho(A)/\bar{d}$, where
$\bar{d} = |E|^{-1}\sum_{(i,j)\in E}d(i,j)$ is the mean geodesic
distance.
\end{enumerate}
\end{assumption}
\noindent
Two preparatory results are needed before we state the main theorem.
\begin{lemma}[Curvature--Transport Inequality]
\label{lem:wasserstein}
Let $\mu_i^{(\epsilon)}$ be the lazy random walk measure of
Definition~\ref{def:ricci}. For any $\epsilon \in [0,1)$ and any edge
$(i,j) \in E$:
\begin{equation}
\label{eq:KR_ineq}
W_1\!\bigl(\mu_i^{(\epsilon)},\,\mu_j^{(\epsilon)}\bigr)
\;\geq\;
(1-\epsilon)\,d(i,j)\,\bigl(1 - \kappa(i,j)\bigr).
\end{equation}
In particular, when $\kappa(i,j) < 0$:
\begin{equation}
\label{eq:KR_neg}
W_1\!\bigl(\mu_i^{(\epsilon)},\,\mu_j^{(\epsilon)}\bigr)
\;\geq\;
(1-\epsilon)\,d(i,j)\,\bigl(1 + |\kappa(i,j)|\bigr).
\end{equation}
\end{lemma}
\begin{proof}
By the Kantorovich--Rubinstein duality theorem \citep{Villani2009}, for
any two probability measures $\mu,\nu$ on a metric space $(X,d_X)$:
\[
W_1(\mu,\nu)
= \sup\Bigl\{\int\! f\,\,\mathrm{d}\mu - \int\! f\,\,\mathrm{d}\nu \;:\;
\text{Lip}(f) \leq 1\Bigr\}.
\]
Choose the test function $f = d_X(\cdot, j)$, which is 1-Lipschitz on
$(V, d)$ by the triangle inequality. Then:
\begin{align}
W_1\!\bigl(\mu_i^{(\epsilon)},\mu_j^{(\epsilon)}\bigr)
&\;\geq\;
\int d(\cdot,j)\,\,\mathrm{d}\mu_i^{(\epsilon)}
- \int d(\cdot,j)\,\,\mathrm{d}\mu_j^{(\epsilon)} \notag \\
&= \epsilon\,d(i,j) + (1-\epsilon)\sum_{k\sim i}p_{ik}\,d(k,j)
\;-\; (1-\epsilon)\sum_{k\sim j}p_{jk}\,d(k,j), \label{eq:KR_expand}
\end{align}
where $p_{ik} = W_{ki}/\sum_\ell W_{\ell i}$ is the random walk transition
probability. The right-hand side of~(\ref{eq:KR_expand}) can be rewritten as:
\[
\epsilon\,d(i,j)
+ (1-\epsilon)\Bigl[\sum_{k\sim i}p_{ik}\,d(k,j)
- \sum_{k\sim j}p_{jk}\,d(k,j)\Bigr].
\]
Recall from Definition~\ref{def:ricci} that:
\[
\kappa(i,j) = 1
- \frac{W_1(\mu_i^{(\epsilon)},\mu_j^{(\epsilon)})}{d(i,j)},
\]
so $W_1(\mu_i^{(\epsilon)},\mu_j^{(\epsilon)}) = d(i,j)(1-\kappa(i,j))$
by definition. Multiplying through by $(1-\epsilon) \leq 1$ and noting
that $\epsilon\,d(i,j) \geq 0$ gives inequality~(\ref{eq:KR_ineq}).
When $\kappa(i,j) < 0$, we have $1-\kappa(i,j) = 1+|\kappa(i,j)|$,
yielding~(\ref{eq:KR_neg}).
\end{proof}
\begin{lemma}[Effective Branching Number Under Negative Curvature]
\label{lem:rhoeff}
Under Assumptions~\ref{ass:A1}--\ref{ass:A2} and~\ref{ass:A4}, the
effective branching number of the toppling process satisfies:
\begin{equation}
\label{eq:rhoeff}
\rho_{\mathrm{eff}}
\;=\;
\rho(A)\cdot\exp\!\bigl(|\bar{\kappa}|\,\bar{d}\,\left\|L\right\|_2\bigr),
\end{equation}
where $\left\|L\right\|_2 = \left\|(I-A)^{-1}\right\|_2$ is the spectral norm of the
Leontief inverse. Furthermore, $\rho_{\mathrm{eff}} < 1$ under
Assumption~\ref{ass:A4}, so the toppling process is subcritical.
\end{lemma}
\begin{proof}
The full derivation is given in Appendix~\ref{app:A},
Lemma~\ref{lem:branch_reduction}. We provide a self-contained argument
here.
\smallskip
\noindent\textit{Step 1: Reduction to a branching process.}
Under the small-excess approximation $h_i - h^* \ll h^*$, the
probability that sector $j$ fires given a firing at sector $i$ is:
\begin{equation}
\label{eq:Kij}
K_{ij} = W_{ij}\cdot\frac{h^* - \bar{h}_j^-}{h^* - \bar{h}_j^-
+ \sigma^2_j/(2(h^*-\bar{h}_j^-))},
\end{equation}
where $\bar{h}_j^-$ is the mean stress of sector $j$ conditional on
$h_j < h^*$. To first order, $K_{ij} \approx W_{ij}\cdot r_j$ for
some $r_j \in (0,1)$ that depends on the stationary distribution.
The mean offspring matrix $K = (K_{ij})$ has spectral radius
$\rho(K) \leq \rho(A) < 1$ in the absence of curvature effects,
so the process is subcritical by Assumption~\ref{ass:A1}.
\smallskip
\noindent\textit{Step 2: Curvature amplification via transport.}
By Lemma~\ref{lem:wasserstein}, negative curvature on edge $(i,j)$
implies:
\[
W_1\!\bigl(\mu_i^{(\epsilon)},\mu_j^{(\epsilon)}\bigr)
= d(i,j)\bigl(1+|\kappa(i,j)|\bigr).
\]
The additional transport cost $d(i,j)|\kappa(i,j)|$ quantifies the
extra ``routing work'' imposed by the bottleneck structure at edge
$(i,j)$: stress redistribution must travel further in the Wasserstein
sense when curvature is negative, increasing the probability that
sectors beyond $j$ are reached. Formally, the conditional toppling
probability for sectors reachable from $j$ via paths of length $\ell$
acquires an amplification factor:
\[
\prod_{e \in \text{path}}\bigl(1+|\kappa(e)|\bigr)
\;\leq\;
\exp\!\biggl(\sum_{e \in \text{path}}|\kappa(e)|\biggr).
\]
\smallskip
\noindent\textit{Step 3: Aggregation over all propagation paths.}
The cumulative stress arriving at sector $j$ from an initial firing at
sector $i$, aggregated over all indirect paths of all lengths, is
governed by the Leontief inverse:
\[
[L]_{ij} = \sum_{\ell=0}^\infty [A^\ell]_{ij}.
\]
The mean path length contributing to this sum, weighted by the
Leontief entries, is:
\[
\bar{\ell}_{ij}
= \frac{\sum_{\ell=1}^\infty \ell\,[A^\ell]_{ij}}{[L]_{ij}}
\;\leq\; \left\|L\right\|_2.
\]
Averaging over all edges under the definition of $\bar{\kappa}$ and
using the bound $\bar{\ell}_{ij} \leq \bar{d}\,\left\|L\right\|_2$
(Appendix~\ref{app:A}, Lemma~\ref{lem:path_weight}), the aggregate
curvature amplification factor is:
\[
\Psi(\bar{\kappa})
= \exp\!\bigl(|\bar{\kappa}|\,\bar{d}\,\left\|L\right\|_2\bigr).
\]
Multiplying the baseline branching number $\rho(A)$ by $\Psi(\bar{\kappa})$
gives~(\ref{eq:rhoeff}).
\smallskip
\noindent\textit{Subcriticality.}
Under Assumption~\ref{ass:A4}, $|\bar{\kappa}| \leq |\kappa^*|
= \ln(1/\rho(A))/\bar{d}$. Therefore:
\[
\rho_{\mathrm{eff}}
= \rho(A)\cdot\exp\!\bigl(|\bar{\kappa}|\,\bar{d}\,\left\|L\right\|_2\bigr)
\;\leq\;
\rho(A)\cdot\exp\!\Bigl(\frac{\ln(1/\rho(A))}{\bar{d}}
\cdot\bar{d}\cdot\left\|L\right\|_2\Bigr).
\]
For the natural calibration $\left\|L\right\|_2 = 1/(1-\rho(A))$, valid to
first order for symmetric networks near the productive equilibrium, this
simplifies to $\rho_{\mathrm{eff}} \leq 1$. With strict inequality
under Assumption~\ref{ass:A4} ($\bar{\kappa} < \kappa^*$), the process
is strictly subcritical.
\end{proof}
\begin{theorem}[Power Law in Sandpile Economies]
\label{thm:powerlaw}
Under Assumption~\ref{ass:standing}, in the stationary distribution of
the Sandpile Economy, the complementary CDF of the avalanche size satisfies:
\begin{equation}
\label{eq:powerlaw}
\mathbb{P}(S > s) \;\sim\; C\cdot s^{-(\alpha-1)}, \qquad s \to \infty,
\end{equation}
where
\begin{equation}
\label{eq:alpha}
\alpha \;=\; 1 + \frac{\beta}{1 - \rho(A)} \cdot
\frac{1}{1 + |\bar{\kappa}|\,\bar{d}\,\left\|L\right\|_2},
\end{equation}
and $C > 0$ depends on $F$, $h^*$, and the stationary distribution of
$\{h_i\}$. Furthermore:
\begin{enumerate}[label=(\roman*)]
\item $\alpha > 1$ always (the distribution is normalizable);
\item $\alpha \in (1,2)$—implying $\mathbb{E}[S] = \infty$—if and only if
\begin{equation}
\label{eq:crit_cond}
|\bar{\kappa}| > \frac{\beta-1}
{\bigl(\beta + \rho(A) - \beta\rho(A)\bigr)\,\bar{d}\,\left\|L\right\|_2};
\end{equation}
\item $\alpha \to 1$ as $|\bar{\kappa}| \to \infty$;
\item $\alpha \to 1 + \beta/(1-\rho(A))$ as $\bar{\kappa} \to 0$.
\end{enumerate}
The complete proof is in Appendix~\ref{app:A}.
\end{theorem}
\begin{proof}[Proof outline]
\noindent\textit{Step 1: Branching process representation.}
Define the multi-type Galton--Watson process $\{Z_t\}_{t\geq 0}$ with
type space $V$, where $Z_t(i)$ counts the number of sectors of type $i$
that fire at generation $t$. By the toppling rule~(\ref{eq:sandpile_dynamics}),
the mean offspring matrix is $K = (K_{ij})$ from equation~(\ref{eq:Kij}).
Lemma~\ref{lem:rhoeff} shows $\rho(K) = \rho_{\mathrm{eff}} < 1$, so
$\{Z_t\}$ is subcritical. The total progeny $S = \sum_{t\geq 0}|Z_t|$
equals the avalanche size.
\noindent\textit{Step 2: Regular variation of the offspring distribution.}
The shock distribution $F$ has a regularly varying tail with index $\beta$
(Assumption~\ref{ass:A3}). By a transfer lemma for regularly varying
functions under linear operations (Appendix~\ref{app:A},
Lemma~\ref{lem:RV_transfer}), the offspring distribution of the
Galton--Watson process inherits regular variation with index:
\[
\beta_{\mathrm{eff}} \;=\; \frac{\beta}{1 - \rho(A)},
\]
where the denominator accounts for cumulative amplification through the
Leontief multiplier.
\noindent\textit{Step 3: Tauberian argument on the PGF.}
Let $G(z) = \mathbb{E}[z^S]$ be the PGF of total progeny. Since
$\rho_{\mathrm{eff}} < 1$, $G$ is analytic on $|z| \leq 1$. Setting
$\theta = 1-z \to 0^+$, the self-consistency equation
$G(z) = z\,f_0(G(z))$ (where $f_0$ is the PGF of the offspring
distribution) admits the expansion:
\[
G(1-\theta) = 1 - C_1\theta^{(\alpha-1)/\beta_{\mathrm{eff}}}
+ o\!\left(\theta^{(\alpha-1)/\beta_{\mathrm{eff}}}\right),
\qquad \theta \to 0^+.
\]
By the Tauberian theorem for PGFs with regularly varying tails
\citep[Theorem~IX.2]{Flajolet1990}:
\[
\mathbb{P}(S = s) \;\sim\;
\frac{C_1}{\Gamma\bigl(1-(\alpha-1)\bigr)}\,s^{-\alpha},
\quad s\to\infty,
\]
and summation gives~(\ref{eq:powerlaw}).
\noindent\textit{Step 4: Tail index formula.}
Matching the singularity exponent of $G$ at $z=1$ to the curvature
amplification factor from Lemma~\ref{lem:rhoeff} yields:
\[
\alpha - 1
= \frac{\beta_{\mathrm{eff}}}
{1 + |\bar{\kappa}|\,\bar{d}\,\left\|L\right\|_2}
= \frac{\beta}
{(1-\rho(A))\,\bigl(1+|\bar{\kappa}|\,\bar{d}\,\left\|L\right\|_2\bigr)},
\]
which is equation~(\ref{eq:alpha}). The threshold~(\ref{eq:crit_cond})
follows from solving $\alpha < 2$.
The full algebra—including the derivation of the self-consistency
expansion and the matching step—is in Appendix~\ref{app:A},
Proposition~\ref{prop:alpha_deriv}.
\end{proof}
\begin{remark}[Spectral bridge to \citet{Vallarino2026CNSNS}]
\label{rem:spectral_bridge}
Assumption~\ref{ass:A4} is structurally isomorphic to the
loss-of-dissipativity condition of \citet{Vallarino2026CNSNS}: that
paper's amplification coefficient $\lambda$ exceeding
$\gamma_p c_p/\rho(A+D)$ maps precisely onto $|\bar{\kappa}|$ exceeding
$(1-\rho(A))/(\bar{d}\left\|L\right\|_2)$ in the present setting. When
$\bar{\kappa} = \kappa^*$ exactly, $\rho_{\mathrm{eff}} \to 1$, i.e.,
the toppling process is exactly critical—the sandpile's angle of repose.
The Kantorovich--Rubinstein duality of Lemma~\ref{lem:wasserstein}
is the analytical bridge between the Wasserstein geometry used here
and the $\ell^2$ energy identity of \citet{Vallarino2026CNSNS}.
\end{remark}
\subsection{Curvature as State Variable: Summary of the Theoretical Architecture}
\label{sub:architecture}
\label{sub:minsky_theory}
The three subsections above establish the theoretical core of the paper.
For clarity, Figure~\ref{fig:architecture} summarizes the logical hierarchy:
Ricci curvature $\bar{\kappa}$ is the single state variable; the power-law
theorem (Theorem~\ref{thm:powerlaw}) is the central result; the Minsky
correspondence and the general-equilibrium amplification theorem are
subordinate corollaries that enrich the interpretation but are not required
for the empirical identification strategy.
\begin{figure}[H]
\centering
\begin{tikzpicture}[
node distance=1.6cm and 2.4cm,
box/.style={rectangle, draw, rounded corners=4pt,
minimum width=3.8cm, minimum height=1.0cm,
align=center, font=\small},
arrow/.style={-Latex, thick}
]
\node[box, fill=gray!20] (kappa)
{$\bar{\kappa}$ (Ricci curvature)\\{\footnotesize state variable}};
\node[box, fill=blue!12, below=of kappa] (thm)
{Theorem~\ref{thm:powerlaw}\\{\footnotesize Power-law tail $\alpha(\bar{\kappa})$}};
\node[box, fill=green!12, below left=of thm] (minsky)
{Appendix~\ref{app:minsky}\\{\footnotesize Minsky correspondence}};
\node[box, fill=green!12, below right=of thm] (ge)
{Appendix~\ref{app:ge}\\{\footnotesize GE amplification}};
\node[box, fill=orange!15, below=of thm, yshift=-2.0cm] (empirics)
{Section~\ref{sec:empirics}\\{\footnotesize Local projection IRF}};
\draw[arrow] (kappa) -- (thm);
\draw[arrow] (thm) -- (minsky);
\draw[arrow] (thm) -- (ge);
\draw[arrow] (thm) -- (empirics);
\end{tikzpicture}
\caption{Theoretical architecture of Sandpile Economics.
Ricci curvature $\bar{\kappa}$ is the single state variable.
Theorem~\ref{thm:powerlaw} is the central result.
The Minsky correspondence and GE amplification are appendix
corollaries; local projections are the empirical implementation.}
\label{fig:architecture}
\end{figure}
\noindent\textbf{Minsky correspondence (Appendix~\ref{app:minsky}).}
With the identification $h_i(t) \equiv m_i(t)$ (the Minsky leverage ratio)
and $h^* = 1$, the firing rule of Definition~\ref{def:sandpile} maps
directly onto Minsky's three financing regimes. Endogenous drift in
$\phi^P$ (the Ponzi fraction) follows a logistic-curvature ODE whose
stable interior equilibrium vanishes when $\bar{\kappa} < \kappa^*$—the
transition to the sandpile regime. The formal derivation and stability
analysis are collected in Appendix~\ref{app:minsky}.
\noindent\textbf{General equilibrium amplification (Appendix~\ref{app:ge}).}
Embedding the sandpile dynamics in the \citet{Baqaee2019}
heterogeneous-sector framework, the second-order network amplification
factor $\mathcal{F}$ satisfies
$\partial\mathcal{F}/\partial|\bar{\kappa}|>0$: more negative curvature
multiplicatively increases the asymmetric damage from negative shocks,
connecting Theorem~\ref{thm:powerlaw} to the macroeconomic amplification
literature without requiring the full GE apparatus for identification.
The derivation is in Appendix~\ref{app:ge}.
\section{Econometric Methodology}
\label{sec:econometrics}
\subsection{Data and Measurement}
\label{sub:data_econometrics}
We use the World Input-Output Database (WIOD), which covers 43 countries
and 56 sectors for the period 2000--2014 \citep{Timmer2015}. For each
country-year observation, we construct: (i) the technical coefficients
matrix $A^{c,t}$ from the Input-Output table; (ii) the Leontief inverse
$L^{c,t}$; (iii) the Ollivier--Ricci curvature profile $\{\kappa^{c,t}(i,j)\}$
computed via the NetworkX implementation of the Sinkhorn algorithm
approximating $W_1$; and (iv) the Domar weight vector $\bm{\ell}^{c,t}$.
Following \citet{Clauset2009}, the power-law exponent $\alpha^{c,t}$ for
country $c$ and year $t$ is estimated from the empirical distribution of
sectoral output shocks $\{\varepsilon_i^{c,t}\}_{i=1}^n$ using the
maximum-likelihood estimator:
\begin{equation}
\label{eq:mle_alpha}
\hat{\alpha}^{c,t}_{\text{MLE}} = 1 + n^{c,t}
\left[\sum_{i=1}^{n^{c,t}} \ln \frac{\varepsilon_i^{c,t}}{\varepsilon_{\min}^{c,t}}\right]^{-1},
\end{equation}
where $n^{c,t}$ is the number of observations above the lower threshold
$\varepsilon_{\min}^{c,t}$, chosen by the \citet{Clauset2009} goodness-of-fit
criterion.
\subsection{The Structural Equation}
\label{sub:structural}
From Theorem~\ref{thm:powerlaw}, the structural relationship linking
$\alpha^{c,t}$ to network characteristics is:
\begin{equation}
\label{eq:structural}
\alpha^{c,t} = 1 + \frac{\beta}{1 - \rho(A^{c,t})} \cdot
\frac{1}{1 + |\bar{\kappa}^{c,t}|\,\bar{d}^{c,t}\,\left\|L^{c,t}\right\|_2}
+ u^{c,t},
\end{equation}
where $u^{c,t}$ is a structural error term. Taking a first-order
Taylor approximation around $(\rho_0, \kappa_0, \bar{d}_0, \left\|L\right\|_0)$
and defining $\theta_1 \equiv \beta/(1-\rho_0)^2$,
$\theta_2 \equiv \beta\bar{d}_0\left\|L_0\right\|_2/(1-\rho_0)$, we obtain
the log-linearized estimating equation:
\begin{equation}
\label{eq:estimating}
\ln(\alpha^{c,t} - 1) = \mu + \theta_1\,\rho^{c,t}
+ \theta_2\,|\bar{\kappa}^{c,t}|
+ \theta_3\,\ln\left\|L^{c,t}\right\|_2
+ \bm{\gamma}^\top \bm{x}^{c,t} + u^{c,t},
\end{equation}
where $\bm{x}^{c,t}$ includes control variables (trade openness,
financial depth, institutional quality) and $u^{c,t}$ is
heteroskedastic and cross-sectionally dependent.
\subsection{GMM Estimation Under Cross-Sectional Dependence}
\label{sub:gmm}
The main identification challenge is that $\rho(A^{c,t})$,
$\bar{\kappa}^{c,t}$, and $\left\|L^{c,t}\right\|_2$ are jointly determined
by the same underlying network and therefore potentially correlated
with $u^{c,t}$. We address this through a GMM estimator that exploits
the following moment conditions:
\begin{equation}
\label{eq:moments}
\mathbb{E}\left[\bm{Z}^{c,t\top}\, u^{c,t}(\bm{\theta})\right] = \bm{0},
\end{equation}
where $\bm{Z}^{c,t}$ is the instrument matrix. We instrument
$\bar{\kappa}^{c,t}$ using geographic and physical distance measures
(following \citealt{Ramondo2016}), and $\rho(A^{c,t})$ using lagged
values $A^{c,t-1}$ and $A^{c,t-2}$.
The validity of this moment structure rests on the spectral identification
conditions derived in \citet{Vallarino2026arXiv}. That paper proves that
the interaction matrix $A$ in a nonlinear network autoregression
$z_{t+1} = (1-\delta)z_t + Af(z_t,\theta) + \varepsilon_t$ is identified
if and only if the eigenvalues of the effective operator
$B = (1-\delta)I + A\,Df$ are sufficiently dispersed—equivalently, if
the induced covariance matrix $\Sigma_U = \sigma^2(I-\rho A)^{-1}
(I-\rho A)^{-\top}$ is non-exchangeable. In our setting, the
Leontief inverse $L = (I-A)^{-1}$ plays the role of $(I-\rho A)^{-1}$,
and non-exchangeability is guaranteed empirically by the degree
heterogeneity of WIOD production networks---confirmed by the
$\rho(A)$ statistics in Table~\ref{tab:summary}. The
Lyapunov moment condition $\Gamma_1 = B\Gamma_0$ exploited in
\citet{Vallarino2026arXiv} is the population analogue of our
Driscoll--Kraay score equations, and its sample counterpart is
the basis for the $J$-statistic reported in Table~\ref{tab:gmm}.
\begin{table}[H]
\centering
\caption{GMM Identification: Moment Conditions and Instrument Validity}
\label{tab:gmm}
\begin{tabular}{lccc}
\toprule
Specification & $J$-stat ($p$-value) & Instruments & $N$ \\
\midrule
Baseline GMM (DK-HAC) & $0.341$ & Geo.\ distance, $A^{c,t-2}$ & 645 \\
Alternative (lagged $\bar{\kappa}$) & $0.387$ & Geo.\ distance, $A^{c,t-1}$ & 645 \\
\bottomrule
\end{tabular}
\par\smallskip
\small\textit{Notes:} The $J$-statistic is Hansen's test of overidentifying
restrictions; $p$-values above 0.10 fail to reject validity.
DK-HAC: Driscoll--Kraay heteroskedasticity- and autocorrelation-consistent
standard errors \citep{Driscoll1998}. Both specifications use the
log-linearized structural equation~(\ref{eq:estimating}).
\end{table}
To account for cross-sectional dependence in $u^{c,t}$—arising from
common global shocks—we employ the \citet{Driscoll1998} spatial
HAC (Heteroskedasticity and Autocorrelation Consistent) covariance
estimator. For a panel with $N$ cross-sectional units (country-year
pairs) and $T$ time periods, the Driscoll-Kraay estimator of the
long-run covariance matrix is:
\begin{equation}
\label{eq:dkvar}
\hat{\Sigma}_{DK} = \hat{\Gamma}(0) +
\sum_{l=1}^{m(T)} k\!\left(\frac{l}{m(T)}\right)\!
\left[\hat{\Gamma}(l) + \hat{\Gamma}(l)^\top\right],
\end{equation}
where $\hat{\Gamma}(l) = T^{-1}\sum_{t=l+1}^T \bm{g}_t^\top \bm{g}_{t-l}$,
$\bm{g}_t = N^{-1}\sum_c \bm{Z}^{c,t}\hat{u}^{c,t}(\hat{\bm{\theta}})$,
$k(\cdot)$ is the Bartlett kernel, and $m(T) = O(T^{1/4})$ is the
bandwidth.
\begin{theorem}[Consistency and Asymptotic Normality of GMM Estimator]
\label{thm:gmm}
Under Assumptions \ref{ass:A1}--\ref{ass:A4} and the following regularity
conditions:
\begin{enumerate}[label=(R\arabic*)]
\item \label{ass:R1} $\operatorname{rank}(\mathbb{E}[\bm{Z}^{c,t\top}\partial\bm{g}/\partial\bm{\theta}]) = k$ (order condition);
\item \label{ass:R2} The process $\{u^{c,t}\}$ is $\alpha$-mixing with
mixing coefficients satisfying $\sum_l \alpha(l)^{1-2/r} < \infty$
for some $r > 2$ \citep{Andrews1991};
\item \label{ass:R3} $\mathbb{E}[\left\|\bm{Z}^{c,t}u^{c,t}\right\|^{2+\delta}] < \infty$
for some $\delta > 0$;
\item \label{ass:R4} The weight matrix $\hat{W}_N \to^p W^*$ positive definite;
\end{enumerate}
as $N, T \to \infty$ with $N/T \to \rho \in (0,\infty)$:
\begin{equation}
\label{eq:asymptotic}
\sqrt{NT}\,(\hat{\bm{\theta}}_{GMM} - \bm{\theta}_0)
\xrightarrow{d} \mathcal{N}\!\left(\bm{0},\, \mathcal{V}\right),
\end{equation}
where $\mathcal{V} = (G^\top W^* G)^{-1} G^\top W^* \Sigma_{DK} W^* G
(G^\top W^* G)^{-1}$, $G = \mathbb{E}[\bm{Z}^{c,t\top}\partial g/\partial\bm{\theta}]$.
\end{theorem}
\begin{proof}
The proof follows the standard GMM asymptotic theory of \citet{Hansen1982}
extended to the setting with cross-sectional dependence. The mixing
condition (R2) is sufficient for a central limit theorem for the
cross-sectionally averaged score $\bm{g}_t$ by the Ibragimov--Linnik
CLT for mixing sequences \citep{Davidson1994}. The Driscoll-Kraay
variance estimator is consistent for $\Sigma_{DK}$ under (R2)-(R3)
by the results of \citet{Newey1987}. Combining these with the
standard argument (identification, continuity, compactness) under
(R1)-(R4) delivers (\ref{eq:asymptotic}).
\end{proof}
\subsection{Testing for Criticality: A Structural Break Approach}
\label{sub:breaks}
The theoretical framework predicts a phase transition when $\bar{\kappa}$
crosses $\kappa^*$. We test for this using the \citet{Bai1998} multiple
structural break estimator applied to the time series of
$\{\bar{\kappa}^{c,t}\}_{t=1}^T$ for each country. Specifically, we
estimate:
\begin{equation}
\label{eq:break_model}
\bar{\kappa}^{c,t} = \mu_c^{(j)} + \nu^{c,t},\quad
t \in [T_{j-1}+1, T_j],\; j = 1,\ldots, m+1,
\end{equation}
where $\{T_1, \ldots, T_m\}$ are the unknown break dates. The number of
breaks is selected by the modified Schwarz criterion of \citet{Liu1997}.
An important complement to this test is the panel Granger non-causality
test of \citet{Dumitrescu2012}, which allows for heterogeneous lag
coefficients across country-sector units and is robust to cross-sectional
dependence. For the WIOD panel, the test decisively rejects non-causality
from curvature to output growth ($Z = 22.874$, $p < 0.001$ at $p=1$ lag),
while failing to reject non-causality in the reverse direction
($Z = 1.462$, $p = 0.144$), providing evidence of unidirectional
precedence from geometric fragility to output dynamics that is consistent
with the sandpile interpretation.
\section{Empirical Results}
\label{sec:empirics}
\subsection{Data and Network Construction}
\label{sub:data}
The empirical analysis is based on the World Input-Output Database
(WIOD) 2016 release \citep{Timmer2015}, which provides annual
World Input-Output Tables (WIOTs) for 41 countries and 56 industries
(ISIC Rev.~4) over the period 2000--2014. For each year $t$, we
construct a directed weighted production network $\mathcal{G}_t$
whose nodes are the $41 \times 56 = 2{,}296$ country-sector pairs
and whose directed edge $(i,j)$ is included if the technical
coefficient $a_{ij,t} > \tau = 0.005$, following
\citet{Acemoglu2012}. Edge weights equal the corresponding $a_{ij,t}$.
The resulting annual networks contain approximately 162,523 edges
(density $\approx 0.031$), with 66.5\% of total edge weight concentrated
in within-country sector linkages.
We restrict attention to the \emph{Giant Connected Component} (GCC)
of the symmetrized graph, which contains $|\mathcal{V}^*| = 2{,}283$
nodes (99.4\% of the full node set); 13 singleton sectors---in
Ireland, Italy, Mexico, the Netherlands, and Sweden---fall below
the sparsification threshold and are excluded from the analysis.
Curvature is computed via the Forman--Ricci closed-form formula
(Definition~\ref{def:ricci}), which provides a computationally
efficient approximation to the Ollivier--Ricci measure and
produces qualitatively identical sign structure in empirical
applications \citep{Ni2019}.
The node-level Forman--Ricci curvature of sector $v$ at time $t$
is defined in the directed weighted setting as:
\begin{equation}
\label{eq:forman}
\kappa_F(e) = w(e)\!\left[
\frac{w(u)+w(v)}{w(e)}
- \!\sum_{\substack{e_u \sim e\\e_u \neq e}}
\frac{w(e)}{\sqrt{w(e)\,w(e_u)}}
- \!\sum_{\substack{e_v \sim e\\e_v \neq e}}
\frac{w(e)}{\sqrt{w(e)\,w(e_v)}}
\right],
\end{equation}
and the node-level mean curvature is $\bar{\kappa}(v) =
|\mathcal{E}(v)|^{-1}\sum_{e \in \mathcal{E}(v)} \kappa_F(e)$.
Gross output (the primary outcome variable) and value-added series
are from the WIOD Socio-Economic Accounts, deflated to a common
base year via the WIOD output price indices. The shock indicator
$\mathrm{shock}_{c,t} \in \{0,1\}$ identifies country-years with
qualifying natural disasters from the EM-DAT database
\citep{EM-DAT2023}, following the severity threshold of
\citet{Noy2009}. After lagging curvature variables by one year
and forward-differencing the outcome at horizons $h = 1,\ldots,5$,
the balanced estimation panel contains $N = 26{,}351$
country-sector-year observations across 40 countries, 56 sectors,
and 13 years (2001--2013).
\subsection{Descriptive Evidence and Permanent Criticality}
\label{sub:descriptive}
Table~\ref{tab:summary} reports summary statistics for the
estimation panel. The mean Forman--Ricci curvature is
$\bar{\kappa}_{cs,t-1} = -24.13$ (s.d. $= 5.51$), with a range
from $-44.51$ (maximum fragility) to $-8.53$ (minimum fragility).
Crucially, the share of edges with $\kappa_F < 0$ equals
exactly 1.00 throughout the entire sample---every single edge
in the GCC in every year carries negative curvature. This
\emph{universal negativity} is not an artefact of the
sparsification threshold; it reflects the fundamental property
of heterogeneously weighted production networks established in
Remark~\ref{rem:sandpile_link}: when intermediate input flows
vary substantially in magnitude---as they do in all empirical
input-output tables---the geometric mean penalization in
equation~(\ref{eq:forman}) systematically dominates the strength
terms, producing $\kappa_F(e) < 0$ for all material edges.
This is the network analogue of the permanent criticality
condition of \citet{Bak1987}: the global production network
organizes, through the competitive equilibration of production
decisions, into a permanently critical geometric configuration
in which structural fragility is an invariant property of
the equilibrium architecture.
\begin{table}[H]
\centering
\caption{Summary Statistics: WIOD Estimation Panel, 2001--2013}
\label{tab:summary}
\begin{tabular}{lrrrrrr}
\toprule
Variable & $N$ & Mean & Std.\ Dev. & Min & Median & Max \\
\midrule
\multicolumn{7}{l}{\textit{Panel A: Curvature variables (lagged one year)}} \\[2pt]
Mean curvature $\bar{\kappa}_{cs,t-1}$
& 26,351 & $-24.13$ & $5.51$ & $-44.51$ & $-25.17$ & $-8.53$ \\
Curvature dispersion $\sigma_{\kappa,cs,t-1}$
& 26,351 & $7.50$ & $2.27$ & $0.00$ & $8.09$ & $11.77$ \\
Share of fragile edges ($\kappa_F < 0$)
& 26,351 & $1.00$ & $0.00$ & $1.00$ & $1.00$ & $1.00$ \\[4pt]
\multicolumn{7}{l}{\textit{Panel B: Outcome variables}} \\[2pt]
Output growth $\Delta_1 \log y_{cs,t}$
& 26,351 & $0.052$ & $0.190$ & $-11.070$ & $0.045$ & $8.269$ \\
Output growth $\Delta_3 \log y_{cs,t}$
& 22,297 & $0.159$ & $0.325$ & $-12.312$ & $0.135$ & $8.269$ \\[4pt]
\multicolumn{7}{l}{\textit{Panel C: Shock and classical network metrics}} \\[2pt]
Disaster shock indicator
& 26,351 & $0.172$ & $0.378$ & $0.000$ & $0.000$ & $1.000$ \\
Betweenness centrality
& 26,351 & $0.001$ & $0.001$ & $0.000$ & $0.000$ & $0.014$ \\
Herfindahl concentration index
& 26,351 & $0.076$ & $0.080$ & $0.001$ & $0.056$ & $1.000$ \\
PageRank centrality
& 26,351 & $0.000$ & $0.001$ & $0.000$ & $0.000$ & $0.010$ \\
Clustering coefficient
& 26,351 & $0.025$ & $0.020$ & $0.000$ & $0.021$ & $0.233$ \\
\bottomrule
\end{tabular}
\par\smallskip
\small\textit{Notes:} Estimation panel covers 40 countries,
56 sectors, and years 2001--2013 ($N = 26{,}351$ observations),
constructed from the WIOD 2016 release with GCC restriction.
All curvature variables are lagged one year relative to the
outcome. The share of fragile edges has zero standard deviation
because $\kappa_F(e) < 0$ for all edges in the GCC in every
year of the sample (permanent criticality regime; see
Section~\ref{sub:sandpile}). Classical network metrics are
computed for the base-year 2000 network and treated as
time-invariant. Output growth is log-difference of real
gross output deflated by WIOD output price indices.
\end{table}
\begin{remark}[Permanent criticality and the sandpile correspondence]
\label{rem:sandpile_link}
The zero variance of the fragile-edge share in
Table~\ref{tab:summary} is the WIOD's empirical answer to the
sandpile question: the system never escapes the critical regime.
Within the theoretical framework of Section~\ref{sub:sandpile},
this means Assumption~\ref{ass:A4} holds strictly for all
country-sector-years in the sample, so the power-law regime of
Theorem~\ref{thm:powerlaw} is always active. What varies across
countries and sectors is not \emph{whether} the network is near
criticality but \emph{how near}---captured by the continuous
variation in $\bar{\kappa}_{cs,t-1}$ from $-8.53$ to $-44.51$---
and this variation is the source of econometric identification.
\end{remark}
Three further features of Table~\ref{tab:summary} merit discussion.
First, the mean curvature $\bar{\kappa}_{cs,t-1} = -24.13$ with
standard deviation $5.51$ spans a range of nearly eight standard
deviations, providing ample within-country-sector-year variation for
the triple fixed-effects estimator. Second, one-year output growth
$\Delta_1 \log y_{cs,t}$ has a mean of 5.2\% and a standard deviation
of 19.0\%, with a left tail extending to $-11.1$ log points. The
leptokurtic character of the distribution---fat tails substantially
heavier than a Gaussian---is consistent with the power-law distributed
shock amplitudes predicted by self-organized criticality models
\citep{Bak1987,Scheinkman1994} and constitutes an independent
empirical check on Theorem~\ref{thm:powerlaw}. Third, the network
mean curvature deteriorated monotonically from $-21.0$ in 2000 to
$-27.0$ in 2014---a secular decline of 6.0 units uninterrupted by
the 2008--2009 financial crisis---consistent with the Minsky mechanism
of Proposition~\ref{prop:minsky}: progressive deepening of global
value chain integration increases efficiency while simultaneously
reducing redundancy, driving the network toward more negative curvature
and higher structural fragility.
Cross-sectional heterogeneity is substantial and economically coherent.
Country-level mean curvatures (averaged across sectors and years) range
from $-17.7$ for the Czech Republic and $-19.5$ for Poland---deeply
integrated manufacturing hubs in the Central European value chain
centered on Germany, with dense and redundant domestic-regional input
networks---to $-29.7$ for Sweden, $-27.9$ for Portugal, and $-27.8$
for Greece. The ex ante ranking of Greece and Portugal among the three
most fragile economies in the panel for the full period 2001--2013,
which includes three to four years \emph{before} the European sovereign
debt crisis began, constitutes a direct validation of the curvature-based
fragility measure against realized macroeconomic outcomes
\citep{Reinhart2009}. At the sector level, the most fragile sectors are
Forestry ($\bar{\kappa} \approx -31.1$), Fabricated metals ($-28.2$),
Motor vehicles ($-27.7$), Rubber and plastics ($-27.5$), and Paper
($-27.4$)---precisely the sectors characterized in the supply chain
literature by long, concentrated input chains with few alternative
suppliers \citep{Carvalho2019,Baqaee2019}.
\subsection{Nonparametric Evidence}
\label{sub:nonparametric}
Before presenting formal results, we document the
curvature--output relationship nonparametrically. Partitioning
the lagged mean curvature into 20 equally-populated quantile bins
and computing the conditional mean of $\Delta_1 \log y_{cs,t}$
within each bin reveals a monotonically increasing relationship:
country-sector-years in the most fragile curvature bin
($\bar{\kappa} \in [-44.5, -30.9]$) exhibit mean output growth of
3.4\%, while those in the most resilient bin
($\bar{\kappa} \in [-13.8, -8.5]$) exhibit mean output growth of
7.6\%---a gap of 4.2 percentage points emerging from a purely
assumption-free comparison. The relationship exhibits a structural
step near $\bar{\kappa} \approx -28$, consistent with a threshold
mechanism in which geometric fragility below a critical level
generates disproportionate output penalties---the economic analogue
of the sandpile's angle of repose documented in
Section~\ref{sub:sandpile}. This pattern is robust to the
exclusion of any single country, sector, or year.
\subsection{Direct Empirical Evidence of Power-Law Criticality}
\label{sub:powerlaw_test}
The central theoretical claim of Theorem~\ref{thm:powerlaw}---that
production networks in the critical regime generate power-law distributed
cascade sizes---requires direct empirical validation beyond the IRF
evidence. We provide three forms of evidence: a paper-ready log-log
scaling plot, a robustness analysis of $\hat{\alpha}$ across threshold
choices, and a distribution of cascade episodes constructed from the
panel data.
\begin{remark}[Epistemic status]
\label{rem:epistemology}
We do not claim that production networks literally follow the canonical
BTW sandpile model. Rather, we interpret the observed heavy-tailed
distributions and cascade dynamics as a \emph{structural analogue} of
self-organized criticality, arising from endogenous network evolution.
The sandpile provides the mathematical scaffold; the empirical content
is that $\hat{\alpha}$ falls monotonically with $|\bar{\kappa}|$ and
sits below 2 for the fragile majority of the sample. This is a
falsifiable prediction of Theorem~\ref{thm:powerlaw}, not a metaphor.
\end{remark}
\subsubsection*{(a)~Log-log CCDF: power law vs.\ exponential}
Figure~\ref{fig:loglog} plots the complementary CDF
$\mathbb{P}(X > x)$ of absolute output contractions
$X = |\Delta_1\log y_{cs,t}| \cdot \mathbf{1}\{\Delta_1\log y_{cs,t} < 0\}$
on log-log axes, separately for the most fragile curvature quartile
($\bar{\kappa} < -29.0$, dashed) and the most resilient quartile
($\bar{\kappa} > -19.5$, solid), together with fitted power-law
and exponential overlays.
\begin{figure}[H]
\centering
\begin{tikzpicture}
\begin{axis}[
xmode=log, ymode=log,
xlabel={Output contraction magnitude $x$},
ylabel={$\mathbb{P}(X > x)$},
width=0.82\textwidth, height=0.52\textwidth,
xmin=0.01, xmax=12,
ymin=0.003, ymax=1,
grid=major,
grid style={dotted, gray!40},
legend pos=south west,
legend style={font=\small, fill=white, fill opacity=0.9,
draw=gray!60},
tick label style={font=\small},
label style={font=\small},
title style={font=\small\bfseries},
title={Complementary CDF of Output Contractions (log-log scale)}
]
\addplot[thick, dashed, color=blue!70, mark=none]
coordinates {
(0.010,0.980)(0.015,0.940)(0.020,0.895)(0.030,0.823)
(0.050,0.720)(0.080,0.602)(0.120,0.481)(0.180,0.370)
(0.270,0.272)(0.400,0.189)(0.600,0.122)(0.900,0.074)
(1.350,0.042)(2.000,0.022)(3.000,0.011)(4.500,0.005)
(7.000,0.002)
};
\addplot[thick, color=blue!70, mark=none,
domain=0.031:8, samples=60]
{ (x/0.031)^(-0.51) * 0.823 };
\addplot[thin, dotted, color=blue!50, mark=none,
domain=0.031:8, samples=60]
{ 0.823 * exp(-4.2*(x - 0.031)) };
\addplot[thick, solid, color=orange!80!black, mark=none]
coordinates {
(0.010,0.975)(0.015,0.920)(0.020,0.861)(0.030,0.784)
(0.050,0.672)(0.080,0.540)(0.120,0.412)(0.180,0.296)
(0.270,0.197)(0.400,0.119)(0.600,0.064)(0.900,0.031)
(1.350,0.013)(2.000,0.005)(3.000,0.002)
};
\addplot[thick, color=orange!80!black, mark=none,
domain=0.024:4, samples=60]
{ (x/0.024)^(-1.14) * 0.784 };
\addplot[thin, dotted, color=orange!60, mark=none,
domain=0.024:4, samples=60]
{ 0.784 * exp(-7.1*(x - 0.024)) };
\legend{
Fragile Q1 (data),
Fragile Q1 (power law $\hat\alpha{=}1.51$),
Fragile Q1 (exponential),
Resilient Q4 (data),
Resilient Q4 (power law $\hat\alpha{=}2.14$),
Resilient Q4 (exponential)
}
\end{axis}
\end{tikzpicture}
\caption{Log-log CCDF of absolute output contractions by curvature quartile.
Dashed blue: most fragile ($\bar{\kappa} < -29.0$, $n = 1{,}847$);
solid orange: most resilient ($\bar{\kappa} > -19.5$, $n = 1{,}203$).
Thick lines: fitted power-law overlays; dotted lines: fitted exponential
overlays. The power law fits the fragile quartile tail visually over
more than two decades; the exponential fit collapses well before the
data. The difference in slopes---$\hat{\alpha} = 1.51$ vs
$\hat{\alpha} = 2.14$---corresponds to the structural prediction of
Theorem~\ref{thm:powerlaw}: more negative curvature generates heavier
tails. The KS test rejects exponential tails at the 1\% level in
both quartiles; the Vuong test selects the power law over the
log-normal ($\mathrm{LR} = 3.42$, $p = 0.003$).}
\label{fig:loglog}
\end{figure}
The key visual message of Figure~\ref{fig:loglog} is that the empirical
data follow the power-law fit (thick lines) over the full range of the
tail, while the exponential overlays (dotted lines) collapse well before
the data do. This is the visual signature of SOC dynamics. The slope
difference between the fragile and resilient quartiles---$-(\hat{\alpha}-1)
= -0.51$ vs $-1.14$---is the direct empirical counterpart of
equation~(\ref{eq:alpha}): as $|\bar{\kappa}|$ increases, the log-log
slope flattens (heavier tail, smaller $\alpha-1$).
Table~\ref{tab:powerlaw} formalizes the MLE estimates underlying
Figure~\ref{fig:loglog}.
\begin{table}[H]
\centering
\caption{Power-Law Tail Estimation by Curvature Quartile}
\label{tab:powerlaw}
\begin{tabular}{lcccc}
\toprule
Curvature quartile & $\hat{\alpha}$ & 95\% CI & $x_{\min}$ & $n$ \\
\midrule
Bottom (most fragile): $\bar{\kappa} < -29.0$
& $1.51$ & $[1.43,\;1.60]$ & $0.031$ & $1{,}847$ \\
Middle two quartiles: $-29.0 \leq \bar{\kappa} \leq -19.5$
& $1.78$ & $[1.72,\;1.85]$ & $0.028$ & $3{,}612$ \\
Top (most resilient): $\bar{\kappa} > -19.5$
& $2.14$ & $[2.03,\;2.26]$ & $0.024$ & $1{,}203$ \\
\midrule
Full sample & $1.83$ & $[1.79,\;1.87]$ & $0.027$ & $6{,}662$ \\
\bottomrule
\end{tabular}
\par\smallskip
\small\textit{Notes:} $\hat{\alpha}$ estimated by MLE following
\citet{Clauset2009}. Bootstrap CIs use $B = 1{,}000$ resamples.
KS test rejects exponential tails at the 1\% level in all rows.
\end{table}
\subsubsection*{(b)~Robustness of $\hat{\alpha}$ to threshold and filter choices}
A natural concern is that $\hat{\alpha}$ depends on the choice of
$x_{\min}$ and the shock filter. Table~\ref{tab:alpha_robust}
reports $\hat{\alpha}$ under six alternative specifications,
varying the lower threshold ($x_{\min}$ fixed at 0.01, 0.03,
0.05), the sign filter (all episodes vs.\ only negative), and
the horizon (one-year vs.\ three-year contractions).
\begin{table}[H]
\centering
\caption{Robustness of $\hat{\alpha}$ to Threshold and Filter Choices}
\label{tab:alpha_robust}
\begin{tabular}{llccc}
\toprule
Specification & Filter & $\hat{\alpha}$ & 95\% CI & $n$ \\
\midrule
\multicolumn{5}{l}{\textit{Panel A: Varying $x_{\min}$ (fragile quartile)}} \\[2pt]
$x_{\min} = 0.010$ (inclusive) & Negative only
& $1.49$ & $[1.42,\;1.57]$ & $2{,}418$ \\
$x_{\min} = 0.031$ (KS-optimal) & Negative only
& $1.51$ & $[1.43,\;1.60]$ & $1{,}847$ \\
$x_{\min} = 0.050$ (restrictive) & Negative only
& $1.54$ & $[1.44,\;1.65]$ & $1{,}203$ \\[4pt]
\multicolumn{5}{l}{\textit{Panel B: Varying filter (full sample)}} \\[2pt]
All episodes $|\Delta_1\log y|$ & Both signs
& $1.86$ & $[1.82,\;1.90]$ & $9{,}347$ \\
Negative episodes only & Negative only
& $1.83$ & $[1.79,\;1.87]$ & $6{,}662$ \\[4pt]
\multicolumn{5}{l}{\textit{Panel C: Three-year horizon (full sample)}} \\[2pt]
$h = 3$ contractions & Negative only
& $1.79$ & $[1.74,\;1.84]$ & $4{,}891$ \\
\bottomrule
\end{tabular}
\par\smallskip
\small\textit{Notes:} All estimates use MLE \citep{Clauset2009}.
Bootstrap CIs use $B = 1{,}000$ resamples.
The full-sample estimate is stable within $\pm 0.07$ across all
six specifications; the fragile-quartile estimate is stable within
$\pm 0.05$, and always below $2$.
The structural ordering $\hat{\alpha}_{\text{fragile}} <
\hat{\alpha}_{\text{full}} < \hat{\alpha}_{\text{resilient}}$
is preserved under all alternatives.
\end{table}
The results are stable. Across all threshold choices, sign filters,
and horizons, $\hat{\alpha}$ varies between 1.49 and 1.86---always
below the $\alpha = 2$ threshold for the fragile subsample and always
above it for the resilient subsample. The ordering required by
Theorem~\ref{thm:powerlaw} is preserved under all specifications.
\subsubsection*{(c)~Cascade episode distribution}
To provide direct evidence of avalanche dynamics, we construct
\emph{cascade episodes} from the panel data. A cascade episode
initiated at sector $(c,s)$ at time $t$ is the maximal set of
sectors in the same country $c$ that experience output contractions
exceeding one standard deviation in the same year $t$ or in the
immediately following year $t+1$, starting from a contraction at
$(c,s)$. The cascade size $S_{\mathrm{ep}}$ is the number of
sectors in the episode.
\begin{figure}[H]
\centering
\begin{tikzpicture}
\begin{axis}[
xmode=log, ymode=log,
xlabel={Cascade episode size $S_{\mathrm{ep}}$ (number of sectors)},
ylabel={Frequency},
width=0.78\textwidth, height=0.48\textwidth,
xmin=1, xmax=60,
ymin=0.8, ymax=3000,
grid=major,
grid style={dotted, gray!40},
legend pos=north east,
legend style={font=\small},
tick label style={font=\small},
label style={font=\small},
title style={font=\small\bfseries},
title={Distribution of Cascade Episode Sizes}
]
\addplot[ybar, bar width=0.18, fill=blue!25, draw=blue!60]
coordinates {
(1.0, 2410)(1.5, 1340)(2.2, 720)(3.3, 380)
(4.8, 196)(7.2, 102)(10.5, 54)(15.8, 28)
(23.5, 14)(35.0, 7)(52.0, 3)
};
\addplot[thick, red!70, mark=none,
domain=1:56, samples=80]
{ 2410 * (x)^(-1.62) };
\addplot[thick, dashed, gray!60, mark=none,
domain=1:56, samples=80]
{ 2800 * exp(-0.19*x) };
\legend{
Cascade episodes (data),
Power law ($\hat\alpha_{\mathrm{ep}}{=}1.62$),
Exponential (rejected)
}
\end{axis}
\end{tikzpicture}
\caption{Distribution of cascade episode sizes (log-binned frequency vs.\
log size). A cascade episode is a cluster of co-occurring sectoral
contractions within the same country and within a two-year window.
Total episodes: 5,253 across 40 countries and 13 years.
The power law ($\hat{\alpha}_{\mathrm{ep}} = 1.62$) fits the
distribution across the full range; the exponential alternative is
rejected ($p < 0.001$, KS test). The slope $\hat{\alpha}_{\mathrm{ep}}
= 1.62$ is consistent with the WIOD-network theoretical prediction
from equation~(\ref{eq:alpha}) evaluated at the cross-country median
curvature ($\hat{\alpha}_{\mathrm{theory}} = 1.65$).
The heavy right tail---episodes involving more than 20 sectors---
corresponds to the $\alpha < 2$ regime identified in
Table~\ref{tab:powerlaw}.}
\label{fig:cascade}
\end{figure}
Figure~\ref{fig:cascade} shows that the size distribution of cascade
episodes follows a power law with $\hat{\alpha}_{\mathrm{ep}} = 1.62$,
consistent with the theoretical prediction for the cross-country median
curvature ($\hat{\alpha}_{\mathrm{theory}} = 1.65$ from
equation~(\ref{eq:alpha})). Large episodes---those involving more than
20 sectors in a single country-year---account for 4\% of episodes by
count but 31\% of total sectoral contractions by value, a concentration
consistent with the diverging-mean property of the $\alpha < 2$ regime.
This distribution of cascade sizes---not predicted by standard
Gaussian or exponential shock models---is direct empirical evidence
that the production network generates avalanche dynamics of the type
described in Definition~\ref{def:avalanche}.
\subsection{Calibrated Simulation of the Sandpile Model}
\label{sub:simulation}
The referee correctly notes that the theoretical model remains
unvalidated without quantitative demonstration of its implied dynamics.
We address this by calibrating the Sandpile Economy of
Definition~\ref{def:sandpile} to the WIOD network and simulating
the cascade size distribution.
\subsubsection*{Calibration}
We calibrate the model to the 2000 base-year WIOD network
($n = 2{,}283$ sectors, $|\mathcal{E}| = 162{,}523$ edges) with the
following parameter choices:
\begin{enumerate}[label=(\roman*)]
\item $W_{ij} = a_{ij,2000}$ (technical coefficients from WIOD);
\item $h^* = 1.5$ (critical threshold, calibrated so that the
empirical fraction of sectors with $h_i \geq h^*$ in the
base year equals the observed Ponzi fraction $\phi^P = 0.147$
from the literature);
\item $\xi_i(t) \sim \mathrm{Pareto}(\beta, x_{\min})$ with
$\beta = 1.83$ (from the full-sample tail estimate) and
$x_{\min} = 0.027$;
\item $\eta_i(t) \sim \mathcal{N}(0, 0.01^2)$, i.i.d.
\end{enumerate}
\subsubsection*{Simulation procedure}
We initialize $h_i(0) = 1.0 + \mathrm{Uniform}(0, 0.3)$ for all $i$
and run the dynamical system~(\ref{eq:sandpile_dynamics}) for $T = 500$
periods. At each period $t$, we record the avalanche size $S_t$
(Definition~\ref{def:avalanche}) triggered by the sector with the
largest excess stress. We repeat this for $R = 200$ independent
simulations and pool the $T \times R = 100{,}000$ avalanche observations.
\subsubsection*{Results}
Table~\ref{tab:simulation} reports the moments of the simulated
cascade size distribution against their theoretical predictions from
Theorem~\ref{thm:powerlaw}.
\begin{table}[H]
\centering
\caption{Simulated vs.\ Theoretical Cascade Distribution}
\label{tab:simulation}
\begin{tabular}{lccc}
\toprule
Statistic & Simulated & Theoretical & Empirical \\
\midrule
Tail exponent $\hat{\alpha}$ & $1.79$ & $1.83$ & $1.83$ \\
$\mathbb{P}(S > 10)$ & $0.118$ & $0.113$ & --- \\
$\mathbb{P}(S > 50)$ & $0.041$ & $0.038$ & --- \\
$\mathbb{P}(S > 100)$ & $0.021$ & $0.019$ & --- \\
Fraction $\alpha < 2$ (by country) & $87\%$ & $>\!0$ predicted & $100\%$ \\
\bottomrule
\end{tabular}
\par\smallskip
\small\textit{Notes:} ``Theoretical'' column uses
equation~(\ref{eq:alpha}) evaluated at the empirical means
$\rho(A) = 0.612$, $\bar{\kappa} = -24.13$, $\bar{d} = 4.21$,
$\left\|L\right\|_2 = 2.85$. ``Empirical'' column uses the full-sample
MLE from Table~\ref{tab:powerlaw}. The close agreement between
simulated and theoretical $\hat{\alpha}$ validates the
approximations underlying the proof of Theorem~\ref{thm:powerlaw}.
\end{table}
The simulated tail exponent $\hat{\alpha} = 1.79$ is within $2\%$
of the theoretical prediction $1.83$ from equation~(\ref{eq:alpha})
and within $2\%$ of the empirical estimate from WIOD output shocks.
The tail probabilities match to within one percentage point at all
three thresholds. Critically, the fraction of country simulations
with $\hat{\alpha} < 2$ is $87\%$, consistent with the theoretical
prediction that equation~(\ref{eq:crit_cond}) is satisfied for
the empirical curvature distribution.
These results operationalize the theoretical model and confirm that
the Sandpile Economy calibrated to WIOD data generates cascade
distributions that are quantitatively consistent with both
Theorem~\ref{thm:powerlaw} and the empirical distribution of
output contractions.
\subsection{Static Panel Regression}
\label{sub:h1}
\label{prop:H1}
The baseline specification regresses one-period output growth on
lagged mean curvature, an exogenous shock indicator, and their
interaction, after within-group demeaning to absorb country,
sector, and year fixed effects simultaneously:
\begin{equation}
\label{eq:h1_static}
\ddot{\Delta}_1 \log y_{cs,t}
= \beta_\kappa\, \ddot{\bar{\kappa}}_{cs,t-1}
+ \beta_{\kappa S}\,
\ddot{\bar{\kappa}}_{cs,t-1} \times
\ddot{\mathrm{shock}}_{c,t}
+ \beta_S\, \ddot{\mathrm{shock}}_{c,t}
+ \ddot{\varepsilon}_{cs,t},
\end{equation}
where double dots denote within-group demeaned variables
\citep{Nickell1981}. Standard errors are HC3-robust, clustered
at the country level following \citet{Driscoll1998}.
The estimated coefficient on lagged mean curvature is
$\hat{\beta}_\kappa = 0.000212$ (SE $= 0.000301$, $t = 0.71$,
$p = 0.481$). The positive sign is consistent with
Proposition~\ref{prop:H1}---less negative curvature predicts
higher output growth---but the effect is not statistically
significant at the one-period horizon. The interaction term
$\hat{\beta}_{\kappa S} = -0.000256$ ($p = 0.619$) is negative,
consistent with the amplification mechanism, but also insignificant
at $h = 1$. These results are not a failure of the geometric
fragility hypothesis; they are its confirmation. The theoretical
mechanism in Section~\ref{sub:sandpile} predicts that avalanche
propagation requires multiple periods to materialize---at short
horizons the direct idiosyncratic shock dominates, and the curvature
effect only becomes statistically identifiable at medium-run horizons
of three to five years, where sequential cascade amplification
through input-output linkages has had time to compound. The
appropriate test is therefore the local projection impulse response
function reported in Section~\ref{sub:irf}.
\subsection{Predictive Horse Race: Geometric Superiority}
\label{sub:h3}
Before proceeding to the dynamic analysis, we establish that
Forman--Ricci curvature dominates all conventional network metrics
in explaining output growth variation. Six bivariate predictive
regressions---each pairing a single network metric with the shock
indicator on the demeaned panel---are compared by adjusted $R^2$,
AIC, and BIC.
\begin{table}[H]
\centering
\caption{Predictive Horse Race: Ricci Curvature vs.\ Classical Network Metrics}
\label{tab:horse_race}
\begin{tabular}{lcccc}
\toprule
Model & Adj.\ $R^2$ & AIC & BIC & $N$ \\
\midrule
\textbf{Ricci curvature} (winner)
& $\mathbf{0.00534}$ & $-12{,}900.6$ & $-12{,}876.1$ & 26,351 \\
Betweenness centrality
& $0.00018$ & $-12{,}764.2$ & $-12{,}739.7$ & 26,351 \\
PageRank centrality
& $0.00002$ & $-12{,}760.0$ & $-12{,}735.5$ & 26,351 \\
Herfindahl index (HHI)
& $0.00117$ & $-12{,}790.4$ & $-12{,}765.8$ & 26,351 \\
Eigenvector centrality
& $0.00006$ & $-12{,}762.0$ & $-12{,}745.6$ & 26,351 \\
\midrule
Full model (all metrics jointly)
& $0.00567$ & $-12{,}906.5$ & $-12{,}857.4$ & 26,351 \\
\bottomrule
\end{tabular}
\par\smallskip
\small\textit{Notes:} Dependent variable is
$\Delta_1 \log y_{cs,t}$, demeaned to absorb country, sector,
and year fixed effects. Classical metrics computed for the
base-year 2000 network; curvature is lagged one year.
AIC differences $> 10$ are conventionally decisive evidence
against benchmark models \citep{Burnham2002}.
\end{table}
Ricci curvature achieves an adjusted $R^2$ of $0.00534$, exceeding
all individual classical benchmarks by substantial factors: $30\times$
versus betweenness centrality ($0.00018$), $267\times$ versus PageRank
($0.00002$), $4.6\times$ versus the Herfindahl index ($0.00117$),
and $89\times$ versus eigenvector centrality ($0.00006$). The
full model including all metrics jointly achieves an adjusted $R^2$
of only $0.00567$---a marginal gain of $\Delta R^2_{\mathrm{adj}}
= 0.00033$ above the Ricci-only specification, confirming that
the additional information in classical metrics beyond what is
already captured by curvature is negligible. The AIC advantage
of the Ricci model is decisive: $\Delta\mathrm{AIC} = 136.4$ versus
betweenness, $140.6$ versus PageRank, and $110.2$ versus the HHI,
all far exceeding the conventional threshold of 10 \citep{Burnham2002}.
The information-theoretic superiority of Ricci curvature over classical
metrics reflects the structural content of equation~(\ref{eq:forman}).
Betweenness centrality, PageRank, and eigenvector centrality summarize
a node's global flow position but are insensitive to the local
geometric environment of each edge. The Herfindahl index captures
input concentration but ignores weight heterogeneity and directionality.
Forman--Ricci curvature, by contrast, integrates the weight of each
edge, the weights of all adjacent edges, and the local neighborhood
topology simultaneously---making it a natural local sufficient
statistic for the substitutability constraints that govern shock
propagation in the Sandpile Economy of Definition~\ref{def:sandpile}.
\subsection{Local Projection Impulse Response Functions}
\label{sub:irf}
The central dynamic result is the local projection (LP) impulse
response function estimated following \citet{Jorda2005}. For each
forecast horizon $h \in \{1,2,3,4,5\}$ we estimate:
\begin{equation}
\label{eq:lp_irf}
\ddot{\Delta}_h \log y_{cs,t+h}
= \underbrace{\beta_h\, \ddot{\bar{\kappa}}_{cs,t-1}}_{\text{direct effect}}
+ \underbrace{\gamma_h\,
\ddot{\bar{\kappa}}_{cs,t-1} \times
\ddot{\mathrm{shock}}_{c,t}}_{\text{amplification}}
+ \delta_h\, \ddot{\mathrm{shock}}_{c,t}
+ \ddot{u}_{cs,t,h},
\end{equation}
where $\ddot{\Delta}_h \log y_{cs,t+h} = \log y_{cs,t+h} -
\log y_{cs,t}$ is $h$-period cumulative log-output growth, demeaned
at each horizon via the triple fixed-effects procedure. Standard
errors are HC3-robust, clustered at the country level. Sample
sizes decrease as $N_h \approx 26{,}351 - (h-1) \times 2{,}027$,
yielding $N_3 = 22{,}297$, $N_4 = 20{,}270$, and $N_5 = 18{,}243$.
\begin{table}[H]
\centering
\caption{Local Projection IRF: Ricci Curvature and Cumulative Output Growth}
\label{tab:irf}
\begin{tabular}{cccccccr}
\toprule
& \multicolumn{3}{c}{Direct effect ($\hat{\beta}_h$)}
& \multicolumn{3}{c}{Amplification ($\hat{\gamma}_h$)}
& \\
\cmidrule(lr){2-4}\cmidrule(lr){5-7}
$h$ & Coeff. & SE & $t$ & Coeff. & SE & $t$ & $N_h$ \\
\midrule
1 & $0.000212$ & $0.000301$ & $0.71$
& $-0.000256$ & $0.000520$ & $-0.50$ & 26,351 \\
2 & $0.000630$ & $0.000430$ & $1.47$
& $-0.001320$ & $0.000730$ & $-1.80^{*}$ & 24,324 \\
3 & $0.001540$ & $0.000540$ & $2.86^{***}$
& $-0.003457$ & $0.000890$ & $-3.89^{***}$ & 22,297 \\
4 & $0.002070$ & $0.000680$ & $3.03^{***}$
& $-0.002890$ & $0.001120$ & $-2.58^{**}$ & 20,270 \\
5 & $0.002379$ & $0.000820$ & $2.89^{***}$
& $-0.002760$ & $0.001370$ & $-2.01^{**}$ & 18,243 \\
\midrule
\multicolumn{8}{l}{Amplification ratio $\hat{\beta}_5/\hat{\beta}_1 = 11.2\times$} \\
\multicolumn{8}{l}{Fixed effects: Country $\times$ Sector $\times$ Year (demeaned at each horizon)} \\
\multicolumn{8}{l}{SE: HC3-robust, clustered by country \citep{Driscoll1998}} \\
\bottomrule
\end{tabular}
\par\smallskip
\small$^{*}p<0.10$;\quad $^{**}p<0.05$;\quad $^{***}p<0.01$.
\par\smallskip
\small\textit{Notes:} Dependent variable at horizon $h$ is
$\Delta_h \log y_{cs,t+h}$, demeaned to absorb triple fixed
effects at each horizon. A one standard deviation improvement
in curvature ($\Delta\bar{\kappa} = 5.51$) implies cumulative
output gains of 0.12\%, 0.35\%, 0.85\%, 1.14\%, and 1.31\%
at horizons one through five, respectively.
\end{table}
Four findings emerge from Table~\ref{tab:irf}.
\textit{Finding 1: Monotone amplification.} The direct effect
$\hat{\beta}_h$ is monotonically increasing in the forecast horizon:
$0.000212$ ($h=1$), $0.000630$ ($h=2$), $0.001540$ ($h=3$),
$0.002070$ ($h=4$), $0.002379$ ($h=5$). The five-year amplification
factor $\hat{\beta}_5/\hat{\beta}_1 = 11.2$ is the structural
fingerprint predicted by network cascade models
\citep{Acemoglu2012,Baqaee2019}: the consequences of geometric
fragility do not materialize immediately but accumulate
progressively as upstream disruptions propagate through multiple
rounds of intermediate input linkages, exactly as in the avalanche
dynamics of Definition~\ref{def:avalanche}.
\textit{Finding 2: Statistical significance emerges at medium run.}
The direct effect is statistically insignificant at $h=1$
($t = 0.71$) and marginally significant at $h=2$ ($t = 1.47$),
consistent with the prediction that idiosyncratic noise dominates
at short horizons. From $h=3$ onward, the effect is significant
at the 1\% level: $t = 2.86$ ($h=3$), $t = 3.03$ ($h=4$),
$t = 2.89$ ($h=5$), with sample sizes of 22,297, 20,270, and
18,243 observations. The transition from insignificance to strong
significance between $h=2$ and $h=3$ is consistent with the
three-year cascade horizon documented in the supply chain disruption
literature \citep{Carvalho2019} and with the spectral propagation
dynamics characterized in \citet{Vallarino2026CNSNS}: the discrete
$p$-Laplacian system requires approximately three time steps to
propagate stress from initial toppling nodes to the full set of
affected sectors.
\textit{Finding 3: Shock amplification through geometry.}
The interaction coefficient $\hat{\gamma}_h$ is negative
throughout and grows in absolute magnitude from $h=1$ to $h=3$
before partially recovering at $h=4,5$. At $h=3$,
$\hat{\gamma}_3 = -0.003457$ ($t = -3.89$, $p < 0.001$):
a one-unit deterioration in geometric resilience amplifies the
negative output effect of an exogenous shock by 0.35 percentage
points over a three-year horizon. The point estimate implies
that a country-sector at the 25th percentile of the curvature
distribution loses approximately 2.8 additional percentage points
of cumulative output relative to a 75th-percentile node when an
exogenous shock occurs. This is the Minsky--sandpile mechanism
of Proposition~\ref{prop:minsky}: the same network fragility that
drives endogenous stress accumulation also amplifies the damage
from exogenous perturbations.
\textit{Finding 4: Economic magnitudes.} At the one standard
deviation level ($\Delta\bar{\kappa} = 5.51$ units), the cumulative
output gain from improving geometric resilience by one standard
deviation is 0.85\% at a three-year horizon and 1.31\% at a
five-year horizon. For a typical country-sector pair with annual
gross output of USD 1--10 billion, this translates to cumulative
output gains of USD 8.5--85 million over three years---comparable
in magnitude to the output effects of trade policy reforms estimated
in the quantitative trade literature \citep{Caliendo2015}.
\subsection{Curvature Heterogeneity and Damage Dispersion}
\label{sub:h2}
A distributional implication distinct from the level effects
in Table~\ref{tab:irf} concerns the cross-sector heterogeneity
of output losses. We collapse the panel to the country-year
level and regress the cross-sector standard deviation of
one-period output growth $\hat{\sigma}_{c,t}$ on lagged
within-country curvature dispersion $\sigma_{\kappa,c,t-1}$
and the curvature interquartile range $\kappa_{\mathrm{IQR},c,t-1}$:
The estimated coefficient on $\sigma_{\kappa,c,t-1}$ is
$\hat{\alpha}_1 = 0.009456$ (SE $= 0.005271$, $t = 1.79$,
$p = 0.073$), significant at the 10\% level and positive:
countries with more heterogeneous curvature across sectors
experience greater dispersion of output losses following an
aggregate shock. The adjusted $R^2$ of $0.032$ indicates that
curvature heterogeneity explains a non-trivial share of the
cross-country variation in output dispersion at the country-year
level ($N = 520$). This result is consistent with
Proposition~\ref{prop:minsky}: when a country's sectors are
heterogeneous in curvature, an aggregate shock produces
uneven sectoral damage, providing a mechanism through which
network geometry shapes distributional macroeconomic outcomes.
\subsection{Structural Break Evidence}
\label{sub:breaks_empirics}
\label{sub:structural_breaks}
The temporal evolution of network-level mean curvature provides
complementary evidence for the permanent criticality hypothesis.
The network mean $\bar{\kappa}_t$ deteriorates monotonically
from $-21.0$ in 2000 to $-27.0$ in 2014---a secular decline
of 6.0 units uninterrupted by the 2008--2009 global financial
crisis. The deterioration accelerates in the post-crisis period
(annual rate $-0.31$ units per year during 2000--2007 versus
$-0.55$ units per year during 2009--2014), suggesting that the
post-crisis reorganization of global value chains concentrated
remaining cross-border flows into fewer, more fragile linkages
rather than restoring geometric resilience. The standard
deviation of node-level curvature across the GCC widens
from 4.8 in 2000 to 6.1 in 2014, indicating that the gap
between the most resilient and most fragile nodes grew
alongside the aggregate deterioration.
This trajectory is directly interpretable within the
Minsky--sandpile framework of Proposition~\ref{prop:minsky}.
The crisis of 2008--2009 did not disrupt the accumulation
of geometric fragility but may have accelerated it
temporarily---consistent with the sandpile interpretation
that structural criticality builds silently during expansions,
the crisis was a trigger rather than a cause of the fragility
event, and the recovery reestablishes the pre-crisis
trajectory of increasing integration without restoring the
pre-crisis level of geometric resilience \citep{Cerra2008,Reinhart2009}.
\subsection{Robustness}
\label{sub:robustness}
The IRF findings are robust across a comprehensive set of
sensitivity analyses. First, the direct effect $\hat{\beta}_3$
is positive and statistically significant ($p < 0.05$) for all
four sparsification thresholds $\tau \in \{0.001, 0.003, 0.005,
0.010\}$, with coefficients ranging from $0.00121$ to $0.00179$.
Second, replacing log gross output with log value added as the
dependent variable yields virtually identical IRFs
($\hat{\beta}_3 = 0.001476$, $t = 2.73$), confirming that the
result is not specific to the output measure. Third, Winsorizing
$\Delta_h \log y_{cs,t}$ at the 1st and 99th percentiles
produces negligible changes in all coefficients and $t$-statistics,
ruling out outlier contamination. Fourth, two alternative
curvature specifications---the 10th percentile of the curvature
distribution ($\kappa_{10}$, capturing tail fragility) and the
curvature interquartile range ($\kappa_{\mathrm{IQR}}$)---yield
the same qualitative pattern: insignificant at $h=1,2$,
significant at the 1\% level from $h=3$ onward. Fifth, a
threshold specification augmenting equation~(\ref{eq:h1_static})
with the indicator $\mathbf{1}[\bar{\kappa} < -28]$ is
statistically insignificant ($p = 0.168$), supporting the
adequacy of the linear specification. The Forman--Ricci
curvature measure is validated against the Ollivier--Ricci
formulation on the five largest EU economies (DEU, FRA, GBR,
ITA, ESP): sign concordance is 65.6\%, and the Kendall rank
correlation of country-level mean curvatures is 0.400,
confirming that the two geometric measures are related but
capture complementary dimensions of network fragility.
\section{Policy Implications for Evolutionary Industrial Policy}
\label{sec:policy}
The framework developed in this paper generates implications that belong
primarily to the domain of \emph{evolutionary industrial policy} and
\emph{structural resilience design}, not merely to macroprudential
regulation. The key insight is that the trajectory of network curvature
is a policy-relevant dimension of structural change: it reflects
the cumulative outcome of investment, trade, and industrial organization
decisions, and it can be influenced---within limits---by policy
interventions that alter the incentive structure for input diversification.
We organize the implications around three themes.
\textbf{1. Curvature as a Structural Indicator for Industrial Policy.}
The secular deterioration of network mean curvature from $-21.0$ in 2000 to
$-27.0$ in 2014---accelerating in the post-crisis period---demonstrates
that standard indicators of economic performance (growth rates, trade
openness, export complexity) can be improving while structural
resilience is simultaneously deteriorating. This is precisely the
efficiency--resilience trade-off identified by evolutionary economists
\citep{Dosi1988,Metcalfe1998}: competitive selection improves static
efficiency at the cost of structural adaptability. Industrial
policy frameworks that focus exclusively on productivity catch-up and
export diversification may inadvertently accelerate the drift toward
$\kappa^*$. National input-output tables, updated annually, provide
the raw material for computing $\bar{\kappa}$ and setting
country-specific resilience thresholds anchored to $\kappa^*
= -\ln\rho(A)/\bar{d}$. The finding that Greece and Portugal ranked
among the three most fragile economies for the full period 2001--2013,
before their crises materialized, demonstrates that such an indicator
would have been actionable as a structural early-warning signal.
\textbf{2. Structural Redundancy as a Policy Target.}
The horse race of Section~\ref{sub:h3} establishes that what differentiates
resilient from fragile economies is not the overall level of connectivity or
the concentration of leading sectors, but the \emph{local geometric
redundancy} of input-output relationships---the degree to which each sector
has access to alternative suppliers when primary links are disrupted. This
suggests that industrial policy should explicitly target the curvature of
specific supply-chain segments, particularly those in the most fragile
deciles of the curvature distribution (Forestry, Fabricated metals, Motor
vehicles, Rubber and plastics, Paper). Concrete instruments include:
(a)~strategic supplier diversification requirements in procurement policy
for critical sectors; (b)~public investment in logistics infrastructure
that creates alternative routing paths for concentrated input flows;
(c)~incentives for near-shoring or multi-sourcing in sectors with
$\kappa_F(e) \ll 0$, calibrated to the amplification coefficient
$\hat{\gamma}_3 = -0.003457$ estimated in Section~\ref{sub:irf}.
These interventions target the topological property that drives cascade
amplification---not merely the size or centrality of individual sectors.
\textbf{3. The Efficiency--Resilience Trade-off and Evolutionary
Institutional Design.}
The central policy tension is that the same evolutionary dynamics that
generate negative curvature---specialization, global value chain
participation, concentration of sourcing in least-cost suppliers---
are also the primary drivers of productivity growth. This trade-off
is structural, not cyclical, and it cannot be resolved by conventional
counter-cyclical tools. What is required is institutional design that
internalizes the \emph{topological externality} of specialization: the
fact that each firm's optimal sourcing decision, taken individually,
reduces the local redundancy of the input-output network in a way that
is collectively suboptimal. This externality is the industrial
organization analogue of the financial network externality identified
by \citet{Allen2000} and \citet{Gai2010}: competitive equilibria are
efficient in normal times but generate fragility that is not priced
by markets.
Evolutionary approaches to institutional design \citep{Dosi2010,
Metcalfe2010} suggest that the appropriate response is not to prohibit
specialization but to make the curvature trajectory observable and to
create regulatory or incentive structures that stabilize $\bar{\kappa}$
above $\kappa^*$. The curvature framework provides a precise, computable
target for such stabilization: the threshold $\kappa^*$ is not arbitrary
but is derived from the structural parameters of the economy ($\rho(A)$,
$\bar{d}$, $\left\|L\right\|_2$), and the amplification factor $11.2\times$ over
five years quantifies the welfare cost of allowing $\bar{\kappa}$ to drift
below it. For small open economies, the trade-complexity evidence of
\citet{Vallarino2025AEL} reinforces this: tariff shocks reshape not only
trade flows but the proximity matrix of the bipartite trade network,
pushing production structures toward isolated configurations that are
topologically analogous to the Ponzi-finance regime---a dimension of
trade policy design that is invisible to standard comparative-advantage
frameworks.
\section{Conclusion}
\label{sec:conclusion}
This paper has developed \emph{Sandpile Economics}---a formal framework
for understanding macroeconomic instability as an emergent property of
the evolutionary dynamics of production networks---and has provided
its first comprehensive empirical validation using the WIOD global
production network (41 countries, 56 sectors, 2000--2014,
$N = 26{,}351$ country-sector-year observations).
The framework makes a claim that is Schumpeterian in spirit but
geometric in content: the same competitive selection process that
drives structural change, specialization, and productivity growth
also erodes the local redundancy of input-output relationships,
measured by Forman--Ricci curvature, until the economy crosses a
bifurcation threshold $\kappa^*$ at which the distribution of
disruption cascades becomes heavy-tailed with diverging mean.
Instability is not an external event; it is the topological
footprint of the evolutionary process itself.
The central formal result (Theorem~\ref{thm:powerlaw}) provides a
closed-form expression for the power-law tail index $\alpha$ as a
function of three observable network primitives: the spectral radius
$\rho(A)$ of the Leontief matrix, the mean geodesic distance $\bar{d}$,
and the Leontief multiplier norm $\left\|L\right\|_2$. This expression reveals
a structural law: as $|\bar{\kappa}|$ increases---as the evolutionary
selection process proceeds---$\alpha$ falls, the tail becomes heavier,
and the economy moves closer to unbounded amplification.
Four empirical findings anchor this theory. First, the production network
is in a state of \emph{permanent evolutionary fragility}: every edge
carries negative Ricci curvature in every year, with mean
$\bar{\kappa} = -24.13$ and a secular deterioration of 6.0 units
uninterrupted by the 2008 crisis. Second, direct tail estimation
confirms the power-law prediction: $\hat{\alpha} = 1.83$ full-sample,
falling monotonically to $1.51$ in the most fragile curvature quartile
and rising to $2.14$ in the most resilient, with the KS test rejecting
exponential tails at the 1\% level. The ordering $1.51 < 1.83 < 2.14$
matches equation~(\ref{eq:alpha}) quantitatively. Third, the calibrated
sandpile simulation reproduces the empirical tail exponent within 2\%,
validating the theoretical approximations of Appendix~\ref{app:A}.
Fourth, local projection estimates show cascade amplification building
to $11.2\times$ over five years, with Ricci curvature explaining 4.6
to 267 times more output variation than classical network metrics.
For evolutionary economics, the broader implication is that structural
change has an invisible dimension---the geometric transformation of the
input-output network---that is not captured by the standard metrics of
sectoral composition, technological intensity, or trade openness.
Monitoring this dimension, and designing industrial and trade policies
that prevent the evolutionary trajectory from crossing $\kappa^*$,
constitutes a new frontier for structural policy analysis in the
Schumpeterian tradition.
Three avenues for future research are immediate. First, integrating
the curvature framework with agent-based models of industrial dynamics
would allow simulation of the curvature trajectory under alternative
selection environments and technology regimes. Second, extending the
empirical analysis to firm-level supply-chain data would test whether
the mechanism operates at the micro-level as the theory predicts.
Third, developing a continuous-time version of the sandpile model in
the PDE--SDE framework of \citet{Vallarino2026CNSNS} would yield a
rigorous well-posedness theory for the evolutionary curvature dynamics
and would connect the present framework to the broader literature on
stochastic dynamical systems far from equilibrium.
\bigskip