📈 economics

When Finite Learning Looks Reliable: Equilibrium Consistency and Iteration Dynamics in Learning from Prices

This paper demonstrates that finite learning paths in a learning-from-prices model can appear numerically reliable despite underlying iteration dynamics failing to approximate equilibrium, revealing a critical distinction between code implementation errors, algebraic equilibrium validity, and the separate stability conditions governing finite iteration horizons.

Original authors: Mateo Bodon

Published 2026-09-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mateo Bodon

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

In the world of financial markets, prices are more than just numbers on a screen; they are messages. When traders buy and sell, the resulting price reveals a hidden mix of information: what people know about the future value of an asset, what they believe about supply, and how much risk they are willing to take. Economists have long studied how people learn from these price signals, a process where individuals update their beliefs based on what the market tells them. A central idea in this field is that if everyone keeps learning and adjusting their strategies, the market should eventually settle into a stable state, an equilibrium where supply matches demand and no one has an incentive to change their mind. However, reaching this stable state is not always a smooth journey. Sometimes, the path to stability is shaky, and the mathematical tools used to simulate these journeys can produce results that look correct on the surface but are actually built on a subtle flaw.

A new study by Mateo Bodon at Yale University investigates exactly this kind of hidden flaw in a specific model of how people learn from prices. The research focuses on a recent paper by Bastianello and Fontanier, which proposed a theory about how traders think when they only have partial information. The original authors showed that under certain conditions, the market would eventually stabilize. Bodon's work does not dispute the main economic ideas or the final conclusions of the original paper. Instead, he performs a meticulous forensic check of the computer code that was shared alongside the original research. He discovered that the code used to find the final stable point contained a single, small error in a denominator—a mathematical divisor—that was different from the equations described in the text. This tiny mismatch meant that the code was solving a slightly different problem than the one the author intended.

The most striking finding is that this error did not ruin the original paper's main results. The computer program that generated the graphs and figures for the original article never actually used the flawed output from the faulty code. It used a different, correct routine to draw the pictures. So, the economic story told in the original paper remains valid. However, the error created a distinct, incorrect solution that the code could have produced if it had been used differently. When Bodon compared this incorrect solution to the correct one, he found they were surprisingly different. At the very end of the learning process, the prices in the flawed solution were about 3.5 percent different from the correct ones. But the difference was even more dramatic at the very beginning of the process: after just one step of learning, the prices in the flawed system were nearly 17 percent away from the correct path. This shows that a system can look like it is heading toward a solution while actually drifting far away from it in the short term.

Beyond this specific error, the study explores a deeper question about how these learning systems behave. The researchers examined whether the path to stability is always smooth or if it can be jagged and unpredictable. They found that the system behaves like a projector that squashes information. In the first step of learning, this projection can actually make errors larger, even if the system eventually settles down. It is possible for a path to look very close to the correct answer after a few steps, only to reveal later that it is actually unstable and will never truly reach the intended destination. By testing thousands of different scenarios, the study identified that in 39 out of 100 cases in one specific sweep, and in 1,833 out of 8,100 cases in a broader variance grid, the learning path was unstable and would not settle down. In these unstable cases, the system might appear to be working well for a while, but it is actually circling a boundary where the math breaks down, rather than finding a true balance.

The study also mapped out exactly where these breakdowns happen. It identified a specific mathematical boundary, a "pole," where the calculations become undefined, similar to dividing by zero. The researchers showed that if a learning path gets too close to this boundary, it can explode or cycle endlessly. They found that for certain settings, the system is guaranteed to converge to a stable answer, but for others, it is destined to fail. Crucially, they demonstrated that a finite number of steps—what a human trader might actually experience—can look reliable even when the long-term math says it is not. A trader might see prices stabilizing after a few days and assume the market has found its balance, when in reality, the system is on an unstable path that will eventually diverge.

This work serves as a reminder that in complex economic models, the difference between a stable market and a chaotic one can depend on tiny details in the math. It highlights the importance of checking not just the final answer, but the entire journey the computer takes to get there. The study confirms that while the original economic theory holds up, the tools used to simulate it require careful scrutiny. It shows that a system can be locally unstable, amplifying errors in the first few steps, even if it eventually contracts toward a solution. Conversely, a system can look stable for a short time while heading toward a collapse. The research concludes that to truly understand market behavior, one must separate three things: the theoretical equilibrium, the stability of the learning process, and the reliability of the finite steps we actually observe. Only by keeping these distinct can we avoid being misled by numbers that look right but are fundamentally wrong.

Technical Summary: "When Finite Learning Looks Reliable: Equilibrium Consistency and Iteration Dynamics in Learning from Prices"

Problem Statement
The paper investigates a critical distinction in computational economics: the divergence between three objects that are often assumed to be identical in learning-from-prices models. These objects are:

The equilibrium conditions stated in the original article.
The residual equations solved by the deposited code (the "deposited- $D$ system").
The finite sequence of updates generated by the updating rule (the "article/updater- $H$ system").

The author examines the replication package for Bastianello and Fontanier [2025], which models partial-equilibrium thinking (PET) and learning from prices. The core problem is that a finite number of reasoning rounds may appear numerically reliable even when the underlying iteration does not approximate the intended equilibrium. Specifically, the paper identifies a discrepancy in the deposited private fixed-point routine where a single denominator differs from the theoretical derivation and the companion updating routine. This mismatch creates a distinct fixed point that fails to satisfy independent market-clearing conditions, raising questions about the reliability of finite learning paths when the iteration is unstable or near a pole.

Methodology
The analysis employs exact algebraic verification and deterministic parameter censuses rather than statistical sampling or rerunning the original simulation.

Source-Faithful Comparison: The author compares the "deposited- $D$ " system (using denominator $D$ ) against the "article/updater- $H$ " system (using denominator $H$ ) without altering either. The analysis relies on a recorded execution of the original MATLAB code and independent code that reconstructs the finite paths stored in the original figure objects.
Exact Algebraic Derivation: The paper derives exact roots for both systems at the benchmark calibration. It constructs independent Gaussian projection and market-clearing conditions to test whether the roots of the deposited code satisfy the economic equilibrium.
Jacobian and Stability Analysis: The local stability of the partially revealing map is analyzed by factoring the four-dimensional Jacobian into a scalar times a rank-two oblique projector ( $J = -\varepsilon \Pi$ ). This factorization separates the first-step projection dynamics from subsequent persistence.
Scalar Reduction and Pole Geometry: The shared core of the two private maps is reduced to a one-dimensional Möbius recurrence. This reduction allows for the exact identification of the pole (singularity), its finite preimages, and the classification of convergence, two-cycling, and explosive regimes based on a scalar parameter $t$ .
Finite-Horizon Census: The study performs an exact enumeration of stability across a prespecified grid of 100 cells (deposited sweep) and 8,100 cells (variance grid), tracking finite-horizon errors and pole proximity without filtering out undefined iterations.

Key Contributions and Results

Exact Root Divergence and Market Clearing Failure:
At the benchmark calibration, the two systems possess unique, denominator-safe fixed points that differ in exactly one coefficient: the mean-supply loading ( $q$ ).
- The article/updater- $H$ root satisfies independently reconstructed market-clearing conditions exactly (zero residuals).
- The deposited- $D$ root leaves a nonzero mean-supply residual of $5000/116781$ (approx. 0.0428) and a unit-mean-supply excess demand of $15625/61104$ .
- Impact: While the plotting driver in the original package never reads the discrepant fixed-point output (using the updater routine instead), the mismatch reveals that the deposited routine solves a system that does not represent the article's equilibrium. The price functions differ by 3.53% at the fixed point and 16.64% after the first update.
Local Stability and First-Step Amplification:
The Jacobian of the source update at the nonzero root factors as $J = -\varepsilon \Pi$ , where $\Pi$ is an oblique projector of rank two.
- Implication: The eigenvalues ( $\{0, 0, -\varepsilon, -\varepsilon\}$ ) govern persistence after the first step, but the oblique nature of $\Pi$ means the operator norm can exceed one. Consequently, the first update can amplify errors significantly even if the system is locally asymptotically stable ( $|\varepsilon| < 1$ ).
- Census Results: Exact enumeration identifies 39 unstable cells in the deposited 100-point sweep and 1,833 unstable cells in the 8,100-cell grid. The displayed benchmark ( $h=0.60$ ) is stable, but many nearby parameter configurations are not.
Global Geometry and Convergence Regimes:
Using the scalar reduction, the paper classifies the dynamics based on the parameter $t = A^2bc$ :
- $t > 1$ : Every finite, forward-defined orbit converges to the map's nonzero fixed point.
- $t = 1$ : The reduced coordinate alternates (two-cycle) unless starting at the fixed point.
- $t < 1$ : The magnitude expands unless starting at the fixed point.
- Convergence Sets: For $0 < t \le 1$ , the convergence set of the article/updater- $H$ system is a strict subset of the deposited- $D$ system. The article system requires specific initial conditions for the $q$ coefficient to converge, whereas the deposited system converges from any finite $q$ .
Finite-Horizon Misleading Accuracy:
The study demonstrates that "early closeness" is not proof of convergence.
- Unstable paths can pass finite tolerance bands before their limiting classification (explosive or singular boundary) becomes visible.
- At the benchmark, a large first-step discrepancy coexists with rapid eventual convergence.
- Unstable paths approach a singular boundary limit (not an equilibrium or attractor) rather than diverging to infinity in the traditional sense.

Significance and Claims
The paper claims a narrow but precise contribution to computational reproducibility and the theory of learning from prices:

Separation of Objects: It establishes that the equilibrium conditions, the fixed-point residual, and the finite iteration path must be evaluated separately. Agreement among them cannot be assumed.
No Correction of Economic Mechanism: The author explicitly states that the mismatch in the deposited fixed-point routine does not alter the economic mechanism or figures reported in the original article, as the plotting driver relies on the updater routine, not the discrepant fixed-point output.
Methodological Precedent: The work serves as an independent code check that distinguishes between "reproducibility" (reproducing the output) and "independent replication" (verifying the logic). It highlights that a finite learning path can look numerically reliable while failing to approximate the intended equilibrium.
Behavioral vs. Numerical Horizons: The results argue that a behavioral horizon (finite rounds of reasoning) and a numerical approximation horizon (convergence to a fixed point) are distinct concepts. A finite iterate may be close to a fixed point for a specific horizon while the underlying iteration is unstable or approaching a pole.

The paper concludes that responsible finite-reasoning claims must distinguish between the behavioral horizon and the numerical approximation horizon, and must report the stability geometry of the map, as oblique projections can amplify errors even in locally stable systems.

Technical Summary: "When Finite Learning Looks Reliable: Equilibrium Consistency and Iteration Dynamics in Learning from Prices"

More like this