⚛️ quantum physics

Thermodynamic Value of XOR-Game-Induced Side Information in a Szilard Engine

This paper demonstrates that while a fixed prediction accuracy in a binary side-information channel does not uniquely determine the thermodynamic value of a Szilard engine, the standard XOR-game embedding specifically yields the symmetric channel that realizes the minimum possible reversible work, whereas asymmetric channels with concentrated errors achieve the maximum.

Original authors: Piotr Ćwikliński

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Piotr Ćwikliński

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 a world where information is not just a string of data on a screen, but a physical substance that can be weighed, measured, and even turned into energy. This idea, which links the abstract world of logic to the tangible laws of heat and motion, has been a cornerstone of physics for over a century. It began with the realization that knowing the state of a tiny particle allows us to extract work from it, much like a skilled mechanic who can squeeze more power from an engine by understanding its inner workings. However, there is a catch: to use this information, one must eventually erase it, and that act of forgetting costs energy. This delicate balance between knowing and forgetting creates a bridge between the mind and the machine, suggesting that the very act of processing a bit of data has a thermodynamic price.

In a new study, researchers have explored the limits of this bridge, specifically asking how much energy can be harvested from a simple piece of information: a single yes-or-no answer about a physical system. The scientists focused on a classic thought experiment known as the Szilard engine, which involves a single particle trapped in a box. If an observer knows which side of the box the particle is on, they can use that knowledge to extract work as the particle expands. The question the researchers tackled was deceptively simple: if an observer receives a message that is correct most of the time, does that message always have the same energy value? They discovered that the answer is no. The same level of accuracy can hide vastly different amounts of usable energy, depending on how the errors in the message are distributed.

To understand this, picture a scenario where a controller receives a binary record—a single bit of data—that is supposed to match the state of a thermal bit, which is essentially a random coin flip representing the particle's position. If the controller is right 90% of the time, one might assume the energy value is fixed. However, the researchers found that the "shape" of the errors matters just as much as the total number of correct guesses. If the mistakes are spread out evenly, so that the controller is wrong about half the time when the particle is on the left and half the time when it is on the right, the energy value is at its lowest. But if the controller makes all its mistakes on just one side—say, it is perfect when the particle is on the left but consistently wrong when it is on the right—the same 90% accuracy yields a significantly higher amount of extractable work. The study proves that a single score of accuracy does not determine a unique energy value; instead, it defines a range of possibilities.

The researchers calculated the exact boundaries of this range. They showed that the minimum energy value is achieved only when the errors are perfectly balanced, a situation known as a binary symmetric channel. In this specific case, the energy value follows a well-known formula that has been used for decades in the context of games involving quantum mechanics. The study reveals that this familiar formula is not just a convenient approximation; it represents the absolute floor, the lowest possible energy value one can get for a given accuracy. At the other end of the spectrum, the maximum energy value is reached when the errors are completely unbalanced, concentrated entirely on one input. This means that a system that is flawless in one direction but flawed in the other is thermodynamically more valuable than a system that is consistently mediocre in both directions, even if both systems get the same number of answers right.

This finding has a direct connection to how scientists often test the limits of physics using "XOR games," which are puzzles where two players try to coordinate their answers based on separate clues. In the standard way these games are set up, the errors in the players' coordination are independent of the physical state of the system they are measuring. This setup naturally creates the perfectly balanced, symmetric error pattern. The study demonstrates that this standard approach always produces the minimum possible energy value. It is not that the game is inefficient; rather, the specific way the game is constructed forces it into the lowest-energy configuration. If one were to change the mechanism so that errors clustered on one side, the same game score would unlock more energy, but the standard rules of the game prevent this.

The researchers also contrasted this view with a different way of valuing information, often used in theoretical models where a successful task simply triggers a battery to charge. In that model, the energy gained depends only on the probability of success, regardless of how the errors are distributed. The study highlights a fundamental difference between these two perspectives. The battery model treats a successful event as a fixed prize, while the Szilard engine model values the information contained in the record itself. Because the engine cares about the specific details of the information—how the errors are distributed—the same success probability can correspond to a wide interval of energy values. The standard game setup happens to sit at the very bottom of this interval.

Ultimately, this work clarifies that the thermodynamic value of information is not a single number attached to a score. It is a property of the entire channel through which the information flows. The study confirms that while a single accuracy score might seem like a complete summary, it actually masks a hidden diversity of physical potential. The familiar formulas used to describe these systems are correct, but they describe only the most conservative case. By mapping out the full range of possibilities, the researchers have shown that the way errors are structured is just as important as the number of correct answers, revealing a deeper layer of complexity in the relationship between information and energy.

Technical Summary: Thermodynamic Value of XOR-Game-Induced Side Information in a Szilard Engine

Problem Statement
This work investigates the reversible work value extractable from a Szilard engine when the feedback controller possesses only a binary side-information record ( $G$ ) correlated with the physical state of the working system ( $X$ ). While the standard operational summary of such a record is its prediction accuracy, $p = P[G=X]$ , the paper addresses a critical gap: a fixed prediction accuracy does not uniquely determine the thermodynamic value of the information. Different binary channels can yield the same $p$ while carrying different amounts of mutual information, $I(X:G)$ . The central question is: for a fixed $p$ , what is the precise interval of reversible work values compatible with a binary side-information channel?

Methodology
The author models the working system as a classical thermal bit (a single particle in a Szilard box) initially in thermal equilibrium at temperature $T$ with a uniform distribution over states $\{0, 1\}$ . The feedback controller receives a binary record $G$ and performs a reversible feedback protocol.

Thermodynamic Framework: The extractable work is identified with the mutual information $I(X:G)$ scaled by $k_B T \ln 2$ , following the standard Szilard engine analysis where the nonequilibrium free energy difference is converted to work.
Channel Parametrization: The binary channel is defined by conditional error probabilities $e_0 = P[G=1|X=0]$ and $e_1 = P[G=0|X=1]$ . The prediction accuracy is constrained by $p = 1 - (e_0 + e_1)/2$ . The author parametrizes the space of all channels with fixed $p$ using an asymmetry parameter $t$ , where $e_0 = (1-p) + t$ and $e_1 = (1-p) - t$ .
Optimization: The problem is reduced to finding the extrema of the mutual information function $I_p(t)$ over the valid range of the asymmetry parameter $t$ . The analysis relies on the convexity properties of the binary entropy function $h_2(\cdot)$ .
XOR-Game Embedding: The study specifically examines the standard embedding of binary-output Bell tasks (XOR games) into this thermodynamic setting. In this construction, the error bit is derived from the game outcome and is independent of the thermal bit $X$ .

Key Contributions and Results
The paper derives the exact minimum and maximum reversible Szilard work values for any fixed prediction accuracy $p \in [1/2, 1]$ :

The Minimum Work ( $W_{\min}$ ):
The minimum work is uniquely attained by the Binary Symmetric Channel (BSC), where errors are distributed equally ( $e_0 = e_1 = 1-p$ ).
$W_{\min}(p) = k_B T \ln 2 [1 - h_2(p)]$
This result confirms that the standard formula often associated with XOR games represents the lower thermodynamic bound.
The Maximum Work ( $W_{\max}$ ):
The maximum work is attained by two asymmetric "Z-type" channels, where all errors are concentrated on a single input value (e.g., $e_0 = 0$ and $e_1 = 2(1-p)$ , or vice versa).
$W_{\max}(p) = k_B T \ln 2 \left[ h_2\left(p - \frac{1}{2}\right) - \frac{1}{2}h_2(2p - 1) \right]$
Asymptotic Behavior:
Near random guessing ( $p = 1/2 + \delta$ ), the thermodynamic values diverge significantly. The minimum work scales quadratically with the excess accuracy ( $\sim \delta^2$ ), whereas the maximum work scales linearly ( $\sim \delta$ ). This demonstrates that records with identical small prediction accuracies can possess parametrically different thermodynamic values depending on the channel structure.
XOR-Game Specifics:
The paper demonstrates that the standard XOR-game embedding inherently produces a binary symmetric channel because the error bit is independent of the thermal bit $X$ . Consequently, the familiar work formula derived from XOR game scores ( $k_B T \ln 2 [1 - h_2(\omega_G)]$ ) is not merely a specific example but realizes the unique minimum-work configuration among all binary channels with that score.

Significance and Claims
The author positions this work as a refinement of the information-to-work valuation framework:

Channel-Level Valuation: The study establishes that a binary prediction score defines an interval of thermodynamic values rather than a unique value. The thermodynamic worth depends on the full conditional distribution $P(G|X)$ , not just the success probability.
Role of XOR Embedding: The paper clarifies that the standard XOR construction is thermodynamically "inefficient" in terms of work extraction relative to the information available; it sits at the lower envelope of the possible work interval. It does not provide a new Bell bound or a device-independent witness but rather contextualizes existing Bell values within a specific thermodynamic extremum.
Contrast with Event Routing: The work contrasts the Szilard feedback model with "battery-routing" models (where a successful event routes a fixed energy packet). In the routing model, work depends linearly on $p$ . In the Szilard model, work depends on the channel structure. The XOR embedding forces the symmetric channel, effectively removing the freedom to access the higher work values available in asymmetric channels with the same $p$ .
Scope: The author explicitly limits the scope to the "effective channel" $P(G|X)$ . The model does not account for the thermodynamic costs of the upstream mechanisms (e.g., the Bell experiment setup, question generation, or memory reset of the controller) unless they are part of the effective feedback module. Thus, the results represent the reversible work of the feedback stage itself, not necessarily the net work of a fully autonomous device.

In summary, the paper provides an exact characterization of the thermodynamic uncertainty inherent in binary side-information, showing that the standard XOR-game interpretation corresponds to the minimal work extraction scenario, while asymmetric channels with the same accuracy can yield significantly higher work.

Technical Summary: Thermodynamic Value of XOR-Game-Induced Side Information in a Szilard Engine

More like this