🧬 biology

The Curvature of Belief: Fisher–Rao Information Geometry and the Limits of Cognition

This paper argues that modeling belief spaces as curved Riemannian manifolds equipped with the Fisher–Rao metric provides a unified framework for distinguishing bounded rationality from irrationality, reinterpreting common judgment biases as Euclidean shortcuts on a curved geometry, and establishing hard limits on cognitive precision in uncertain environments.

Original authors: Bartosz W. Wojciechowski

Published 2026-10-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bartosz W. Wojciechowski

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

For decades, scientists studying how humans think have relied on a simple idea: that our beliefs are like numbers. When we learn something new, we adjust these numbers up or down, much like a student adding or subtracting on a chalkboard. This approach, known as Bayesian inference, has become the standard language for understanding everything from how we see the world to how juries decide guilt. It assumes that the space where our beliefs live is flat and straight, where the distance between two ideas is just the simple difference between their numbers. But this assumption leaves a crucial question unanswered: is the landscape of our minds actually flat, or is it curved? If the terrain of belief is curved, then the straight lines we use to navigate it are illusions, and the errors we make might not be failures of logic, but the inevitable result of walking in a straight line on a round world.

A new study by Bartosz W. Wojciechowski at the Jagiellonian University in Poland proposes that the space of beliefs is indeed curved, shaped by a mathematical principle called the Fisher–Rao metric. This metric measures the distance between two beliefs not by how different their numbers look, but by how easily evidence could tell them apart. In this curved landscape, moving from a 50% chance to a 55% chance is a tiny step, but moving from a 0% chance to a 0.1% chance is a massive journey, because the latter represents a shift from impossibility to possibility. The author argues that because our brains often treat this curved space as if it were flat, we make systematic errors that look like irrationality but are actually the price of using the wrong map.

The paper draws two major conclusions from this geometric insight. First, it establishes a hard limit on how precisely any mind can track a changing world. Just as a camera lens has a physical limit to how sharp an image can be, a human mind has a limit to how accurate a belief can be, depending on how much the environment is changing and how much information each new piece of evidence provides. The study derives a precise formula for this limit in a drifting environment, showing that in a volatile world, even the smartest agent must accept a certain amount of error. This creates a new way to judge human performance: if a person's errors match this mathematical floor, they are not being irrational; they are simply doing the best possible job given the chaos of the world. Only when errors exceed this floor is true irrationality at play.

Second, the study suggests that many common judgment biases—such as ignoring base rates, being too conservative, or overestimating our confidence—are the result of taking "Euclidean shortcuts." If a person updates their beliefs by simply averaging numbers on a flat scale, they will systematically land in the wrong place compared to someone who navigates the true curved geometry of the problem. The research demonstrates this with a specific example involving a simple yes-or-no probability. When an agent tries to update a belief from a low probability to a high one, a straight-line calculation lands them closer to the middle than the mathematically correct path would. This "centering bias" pulls extreme beliefs toward the average, explaining why people often underestimate rare events or fail to fully embrace strong evidence. The author illustrates this with a reproducible computer simulation showing exactly how these errors accumulate across the entire spectrum of probabilities, with the biggest mistakes happening near the extremes of certainty and impossibility.

However, the author is careful to distinguish between what is proven and what is still a hypothesis. The idea that belief space is curved and that it sets a hard limit on perception is well-established in the study of vision and neural coding, where it has been measured and confirmed repeatedly. In these sensory domains, the brain appears to respect this geometry perfectly. But when it comes to higher-level thinking, such as how people weigh probabilities in complex decisions, the curved-geometry explanation is still a promising research program rather than a finished fact. The paper admits that while the geometric account explains the shape of these biases, it does not yet perfectly match the specific numbers found in human experiments, and other explanations might fit the data just as well. The author calls for new, rigorous tests to see if the geometric model can predict human errors better than existing theories, particularly in situations involving rare events or high-stakes decisions like legal fact-finding or risk communication.

Ultimately, this work reframes the conversation about human error. It suggests that the mind is not necessarily broken when it makes mistakes in volatile environments or when it struggles with extreme probabilities; it may simply be navigating a curved landscape with a flat map. By recognizing the true shape of the space where beliefs live, scientists can better distinguish between the unavoidable limits of cognition and genuine irrationality. The study provides the tools to measure this curvature and offers a clear path forward for testing whether our minds truly follow the geodesics of a curved world or if they continue to stumble on the straight lines of our intuition.

Technical Summary: The Curvature of Belief: Fisher–Rao Information Geometry and the Limits of Cognition

Problem Statement
Cognitive science and the social sciences routinely model beliefs as probability distributions, treating the space of beliefs as a flat, Euclidean domain where distance is the simple arithmetic difference between numbers. However, this approach leaves a foundational question implicit: what is the natural geometry of the space of beliefs? If the space is curved, then standard Euclidean intuitions regarding distance, shortest paths (geodesics), and parallel transport are structurally incorrect. An agent relying on flat-space heuristics in a curved belief manifold would inevitably make systematic errors that mimic irrationality. The paper addresses the lack of a rigorous geometric framework for belief spaces and the consequent ambiguity in distinguishing between bounded rationality (optimal performance under constraints) and genuine irrationality.

Methodology
The paper employs Information Geometry, specifically the Fisher–Rao metric, to define the geometry of belief.

Theoretical Framework: It utilizes the uniqueness theorem by Chentsov (1982), which establishes that the Fisher–Rao metric is the only Riemannian metric invariant under Markov morphisms (stochastic maps that do not create information). This forces the geometry of belief to be curved rather than a modeling choice.
Analytical Derivation: The author derives exact closed-form expressions for:
1. The Fisher–Rao distance on the Bernoulli manifold (binary beliefs), utilizing the arcsine-square-root transform ( $\phi = 2 \arcsin \sqrt{p}$ ) to map the curved manifold to a flat space.
2. The Cramér–Rao bound for a tracking agent in a volatile environment (random walk), solving the steady-state variance equation for a Bayesian tracker (Kalman filter approximation) to determine the irreducible error floor.
Computational Demonstration: A reproducible Python script (no fitted parameters) computes geodesic midpoints versus Euclidean midpoints across the Bernoulli manifold to visualize the "centring bias." It also calculates the exact steady-state tracking variance ( $V^*$ ) as a function of environmental volatility ( $\sigma$ ) and per-observation Fisher information ( $I$ ).

Key Contributions and Results

1. The Cognitive Cramér–Rao Bound
The paper reframes the statistical Cramér–Rao bound as a hard, epistemic limit on cognition.

The Limit: The precision of any agent's belief is floored by the inverse Fisher information of the evidence. No amount of intelligence or computation can bypass this limit.
Volatility-Dependent Floor: In changing environments where the tracked quantity drifts (variance $\sigma^2$ ), the bound is not just the inverse Fisher information but a joint function of information and volatility. The paper derives an exact expression for the steady-state posterior variance $V^*$ :
$V^* = \frac{I\sigma^2 + \sqrt{I^2\sigma^4 + 4I\sigma^2}}{2I} - \sigma^2$
Implication: This establishes a quantitative "yardstick" for bounded rationality. An agent operating at this floor in a high-volatility environment exhibits large errors (imprecision, lag) but is performing optimally. Only errors below this floor constitute genuine irrationality.

2. Biases as Curvature Errors
The paper proposes a mechanism for a family of judgment biases (base-rate neglect, conservatism, probability weighting, overconfidence) as the result of agents taking Euclidean shortcuts on a curved manifold.

Mechanism: Rational Bayesian updating corresponds to moving along a geodesic on the belief manifold. An agent using a standard Euclidean interpolation in probability coordinates deviates from this path.
Direction of Error: The Euclidean midpoint is systematically pulled toward the center ( $p=0.5$ ) relative to the geodesic midpoint. This "centring bias" mathematically mirrors base-rate neglect and conservatism.
Quantitative Findings:
- The discrepancy between Euclidean and geodesic midpoints is zero for symmetric pairs (e.g., 0.40 and 0.60).
- The discrepancy is maximal near the boundaries (e.g., a pair of 0.01 and 0.50 yields a displacement of ~0.071).
- The effect is real but modest in the interior of the manifold, becoming significant only near the boundaries (certainty/impossibility).

3. Empirical Standing and Discriminating Tests
The paper explicitly demarcates the framework's empirical validity:

Established: At the level of perceptual psychophysics and neural coding, the Fisher–Rao geometry is a measured constraint. Efficient coding, neural population decoding limits, and volatility-sensitive learning rates (Behrens et al., 2007; Nassar et al., 2010) are quantitatively confirmed.
Speculative (Research Programme): For higher-cognitive biases (e.g., probability weighting in decision making), the framework is a suggestive mechanism, not an established account. The geometric account currently yields qualitative matches (e.g., inverse-S shapes) but lacks the precise functional form of phenomenological models (e.g., Prelec, Tversky & Kahneman) and requires auxiliary assumptions about the specific heuristics used.
Proposed Tests: The paper specifies three discriminating tests to move the programme forward:
1. High-resolution elicitation of probability weighting near boundaries to compare geometric vs. phenomenological functional forms.
2. Belief-updating tasks that map error patterns against the predicted "centring bias" map (errors should vanish in symmetric regions and peak near boundaries).
3. Tracking tasks manipulating volatility and information to verify if human error rises according to the predicted $\sqrt{\sigma}/\sqrt{I}$ law and sits exactly on the derived floor.

Significance and Claims
The paper claims that the geometry of belief is curved and forced by information-theoretic invariance requirements, not a modeling convenience.

For Social Science: It provides a rigorous tool to distinguish between bounded rationality (performance limited by the environment's volatility and information) and irrationality. It suggests that legal fact-finding, forecasting, and risk communication models must account for the fact that "distance" in belief space is not linear; for instance, the difference between "impossible" and "one-in-a-thousand" is informationally vast, while the difference between 0.50 and 0.55 is slight.
Modesty: The author explicitly states that the framework is not a completed theory of the mind. For higher cognition, it remains a research programme awaiting discriminating tests. The paper does not claim to explain all biases but offers a geometrically natural candidate mechanism that unifies perceptual and cognitive phenomena under a single metric, provided the agent's representation aligns with the Fisher–Rao geometry.

Limitations Acknowledged

The quantitative demonstration is confined to the 1D Bernoulli manifold; higher-dimensional manifolds require separate computation.
The "Euclidean update" is a stylized heuristic; real agents may use other flat-space operations.
The exact tracking floor assumes a Gaussian random walk; real environments may involve jumps.
The paper presents no new empirical data; its contribution is theoretical and methodological, defining the predictions that future empirical work must test.

More like this