📈 economics

Choice at Finite Capacity: The Bounded Agent as an Information Channel and the Recovery of Walrasian Demand

This paper redefines the consumer as a bounded information channel that compresses choices into probability distributions, demonstrating that downward-sloping demand emerges naturally from budget constraints and attention limits rather than perfect rationality, while recovering standard Walrasian demand as the limit of unlimited attention.

Original authors: Avishek Bhandari

Published 2026-07-15

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Avishek Bhandari

Original paper licensed under CC BY 4.0 (http://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

Imagine the classic economics textbook consumer as a superhero with a superpower: perfect, infinite brain power. This hero sees every single price tag in the store, calculates the absolute best possible combination of goods to maximize their happiness, and picks exactly one perfect bundle. They never make a mistake, never get confused, and never have to guess. In this world, the "Law of Demand" (the rule that says if a price goes up, you buy less) is a rigid law of physics, guaranteed because this superhero is perfectly rational.

But what if that superhero doesn't exist? What if real humans (and even AI robots) are more like bottled-up information channels?

This paper, written by Avishek Bhandari, suggests we stop pretending we have infinite brain power. Instead, imagine your mind is a Wi-Fi router with a strict data limit. You can't download the entire internet (the full state of the economy) at once. You have a "bit budget." You can only process a certain amount of information before you have to stop.

The Great Compression

Because your "router" has a limit, you can't see the world in high-definition 4K. You have to compress it. You take the messy, complex reality of prices and goods and squish it down into a lower-resolution image that fits your data budget. You then make your choices based on this blurry, compressed picture, not the perfect reality.

The paper argues that your choice isn't a single, sharp point like the textbook hero's. Instead, your choice is a cloud of possibilities, a probability distribution. You might buy a little bit of everything, with some items being more likely than others, because your "router" can't be 100% sure which single item is the absolute best.

The Two Budgets: Money and Bits

In the old story, you only have one budget: your money. You can't spend more than you have.
In this new story, you have two budgets:

Money: The usual cash constraint.
Bits (Attention): A limit on how much information you can process.

Just as you pay for goods with dollars, you pay for information with "bits." The paper introduces a concept called $\beta$ (beta), which acts like a "temperature" for your attention.

High $\beta$ (Infinite Capacity): Your router is super fast. You can process everything. The "cloud" of your choices collapses into a single, sharp point. You become the textbook superhero. The old laws of economics return perfectly.
Low $\beta$ (Zero Capacity): Your router is broken or turned off. You can't process any new info. You just stick to your habits (your "prior"). You buy the same things you always bought, regardless of price changes.
Middle $\beta$ (Real Life): You are somewhere in between. You are rational enough to notice price changes, but not perfect enough to calculate the absolute best answer every time. You act on a "blurred" version of the truth.

The Magic Formula: The Compressed Slutsky Matrix

The paper's biggest "aha!" moment is a mathematical discovery about how you react to price changes. In old economics, the rule that "when a price goes up, you buy less" (the compensated law of demand) was proven to work only because the consumer was perfectly rational.

This paper proves something surprising: You don't need to be rational for this rule to work.

Even if you are just a "noisy" channel making mistakes, as long as you are trying to get value within your limits, your behavior follows a specific pattern. The paper derives a new formula for how you react to prices, called the Compressed Slutsky Matrix.

Think of this matrix as a symmetry machine.

In the old theory, the symmetry (the idea that if apples get expensive, you buy more oranges, and if oranges get expensive, you buy more apples, in a perfectly balanced way) was a sign of a "good" rational mind.
In this paper, the symmetry comes from the math of the compression itself. Because your brain is a "channel" with a limit, your choices naturally form a pattern that looks symmetric and follows the law of demand. It's not because you are a genius; it's because you are a finite channel.

The paper shows that this "Compressed Slutsky Matrix" is always symmetric and always points in the "right" direction (negative), no matter how "dumb" or "habitual" the consumer is, as long as the budget constraint is binding. It's a property of the budget and the channel, not the consumer's soul.

The "Gaussian" Test Drive

To prove this isn't just a wild guess, the authors built a specific, simple example: a two-good quadratic consumer.

Imagine a shopper choosing between two goods (like apples and bananas).
They have a specific "happiness" function (a quadratic curve) and a fixed amount of money.
The authors ran the math for this shopper with different "bit budgets" (different levels of $\beta$ ).

The Results:

At Infinite Capacity ( $\beta \to \infty$ ): The shopper's choices became a single point, exactly matching the classic textbook "Walrasian demand." The math worked perfectly.
At Zero Capacity ( $\beta \to 0$ ): The shopper stopped thinking and just bought their "prior" bundle (pure habit).
At Finite Capacity: The shopper's choices were a Gaussian distribution (a bell curve). The paper calculated the exact numbers:
- With a flat prior, the mean demand was (1.25, 1.75).
- The "Substitution Matrix" (the rule for how they swap goods) had eigenvalues of -0.7883 and -0.2832.
- Crucially, the matrix was symmetric and negative (meaning the law of demand held) even though the shopper was "noisy."

The paper explicitly states that these results are exact for this specific quadratic case. For more complex, non-quadratic cases, the math suggests the same principles hold, but the numbers would only converge to the classic "perfect" answer as the capacity grows infinitely large.

What This Rules Out

The paper is very clear about what it rejects:

It rejects the idea that "Rationality" is the only source of order. You don't need a perfectly logical mind to get the "Law of Demand." The order comes from the constraint (the budget and the information limit).
It rejects the idea that the textbook consumer is the "real" one. The textbook consumer is just a special, extreme case (the "frictionless corner") where the price of thinking is zero. Real life is the "friction" in the middle.
It rejects the idea that "irrational" behavior is chaotic. Even a bounded, "irrational" agent follows strict, predictable laws (like symmetry) because of how information compression works.

The Takeaway

This paper doesn't say the old economics is "wrong." It says the old economics is a limit. It's what happens when you have infinite data and infinite time.

But for the rest of us—humans with limited attention spans, and AI models with limited processing power—we are finite channels. We compress the world. And when we do, our choices aren't random messes; they follow a beautiful, mathematically precise "Gibbs law" (a specific type of probability distribution).

The "Law of Demand" isn't a sign that we are geniuses. It's a sign that we are bottled-up information channels trying to do the best we can with the bits we have. Whether you are a human shopper, a robot, or a learning algorithm, if you have a limit on how much you can process, your choices will naturally fall into this symmetric, predictable pattern. The "rational" consumer is just the extreme case where the limit disappears.

Technical Summary: Choice at Finite Capacity

Problem Statement
Traditional microeconomic demand theory rests on the axiom of the rational maximiser: an agent with unlimited cognitive reach who selects a single optimal bundle from a budget set based on a fixed, complete, and transitive preference ordering. This paper challenges the premise that rationality is the primitive of choice. Instead, it posits that the consumer is a finite-capacity information channel. Such an agent cannot resolve the full state of the economy and must compress its environment to a resolution allowed by a "bit budget" before acting. The central problem is to derive the properties of demand and the Slutsky matrix from this information-constrained premise, rather than from utility maximisation, and to determine whether classical demand theory can be recovered as a limiting case.

Methodology
The paper models the agent using the framework of rate-distortion theory and free-energy minimisation:

The Agent as a Channel: The consumer is defined as a channel $p(\hat{x}|x)$ subject to a mutual information constraint $I(X; \hat{X}) \leq C$ , where $X$ is the economic state and $\hat{X}$ is the internal representation.
Dual Formulation: The constrained optimisation (maximise expected value subject to a rate ceiling) is treated via its Lagrangian dual. This introduces an "inverse temperature" parameter $\beta$ , which acts as the shadow price of information.
The Gibbs Law: The solution to the information-constrained problem is derived as a Gibbs distribution (or exponential tilt):
$q^*(a) \propto p_0(a) \exp(\beta U(a))$
Here, $p_0$ is a prior (default) policy, $U(a)$ is the value of action $a$ , and $\beta$ indexes the agent's capacity.
Compressed Demand: The consumer's choice is not a point but a probability distribution over the budget set. The mean demand is the expectation of this distribution.
Analytical Derivation: The paper derives the price response of the mean demand using the properties of exponential families. It specifically analyzes the compressed Slutsky decomposition to determine the symmetry and sign of the substitution matrix.

Key Contributions and Results

Recovery of Walrasian Demand: The paper demonstrates that classical demand theory is exactly recovered as the infinite-capacity limit ( $\beta \to \infty$ ). As capacity grows, the Gibbs distribution concentrates on the utility-maximising bundle, and the stochastic choice collapses to the deterministic Walrasian demand. Conversely, at zero capacity ( $\beta \to 0$ ), the agent reverts to pure habit (the prior $p_0$ ).
The Compressed Slutsky Matrix: The central technical result is a new decomposition of the substitution matrix for a bounded agent:
$S = -\beta \lambda \Sigma$
Where:
- $\beta$ is the inverse temperature (capacity price).
- $\lambda$ is the marginal value of wealth (shadow price of the budget).
- $\Sigma$ is the covariance matrix of the consumer's own choice law.
Symmetry and Negative Semidefiniteness without Rationality:
- The substitution matrix $S$ is symmetric and negative semidefinite at every finite capacity, not just in the limit.
- Crucially, these properties are derived solely from the fact that $S$ is a scalar multiple of a covariance matrix. They do not rely on the consumer having rational preferences (completeness, transitivity, or utility maximisation).
- The "compensated law of demand" (downward sloping compensated demand) is shown to be a property of the budget constraint and the channel structure, not of the agent's rationality.
Two-Dimensional Budget: The consumer faces two constraints: the wealth constraint (priced in money) and the information constraint (priced in bits). The inverse capacity $1/\beta$ acts as the price of attention.
Substrate Neutrality: The framework is "substrate-neutral." It applies equally to human consumers, machines running capacity-limited policies, deep neural networks (as cascades of rate-distortion channels), and reinforcement learners. An artificial agent is not a new type of participant but the same channel at a specific capacity.

Worked Example
The paper provides a closed-form solution for a two-good consumer with quadratic utility and a Gaussian prior.

The compressed demand is a Gaussian distribution.
The substitution matrix is explicitly calculated as $S = -\lambda Q^{-1}$ (independent of $\beta$ in the Gaussian case due to cancellation).
The example verifies that the matrix is symmetric and negative definite at all finite capacities and converges to the classical Hicksian substitution matrix as $\beta \to \infty$ .
It illustrates the collapse to habit as $\beta \to 0$ , where the mean demand converges to the prior mean.

Significance and Claims
The paper claims to provide a strict generalisation of demand theory rather than a refutation.

Rationality as a Limit: The classical rational maximiser is repositioned not as the fundamental axiom, but as the frictionless, infinite-capacity corner of a more general theory of bounded agents.
Origin of Integrability: The integrability conditions (symmetry and negativity of the Slutsky matrix), traditionally viewed as the signature of rational utility maximisation, are shown to be inherited from the exponential form imposed by the information constraint.
Decoupling Slope from Rationality: The paper argues that the compensated law of demand is secured by the budget constraint and the channel's exponential response, surviving even if the underlying value field is arbitrary or non-rational.
Unification: It unifies disparate literatures—rational inattention (Sims), random utility (McFadden), rate-distortion (Shannon), and machine learning (Information Bottleneck)—under a single primitive: the finite-capacity channel.

The paper concludes that the classical consumer is a limiting case of a "compressing, erring, and thought-economising" agent, and that the observable content of demand theory (specifically the structure of the substitution matrix) is robust to the relaxation of the rationality axiom, provided the agent operates under a finite information budget.