🔬 condensed matter

Learning to erase quantum states: thermodynamic implications of quantum learning theory

This paper establishes a concrete connection between quantum learning theory and thermodynamics by demonstrating that efficient learning algorithms can acquire the knowledge necessary to erase unknown quantum states at optimal energy cost, thereby linking thermodynamic efficiency to state complexity while revealing fundamental computational limits under cryptographic assumptions.

Original authors: Haimeng Zhao, Yuzhen Zhang, John Preskill

Published 2026-08-25

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Haimeng Zhao, Yuzhen Zhang, John Preskill

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

In the physical world, information is not just a abstract concept; it is a tangible thing with weight and cost. For decades, physicists have understood that erasing information is an act that generates heat. This idea, known as Landauer's principle, states that if you have a system and you want to reset it to a blank, standard state, you must pay an energy price. The amount of energy required depends entirely on how much you know about the system. If you are completely ignorant of the system's current condition, the cost is high. But if you possess a detailed record of its state, you can reset it with almost no energy at all, essentially reversing the process that created it. This principle has long been a cornerstone of thermodynamics, linking the abstract world of data to the concrete reality of heat and work. Yet, a lingering question remained: does the act of learning itself—the process of gathering that crucial knowledge—carry its own hidden energy cost? If the cost of learning is too high, the savings from knowing the state might be lost before the erasure even begins.

A team of researchers has now answered this question with a definitive "no," revealing that learning can be a perfectly reversible process with no fundamental energy cost of its own. By combining the fields of quantum learning theory and thermodynamics, they have demonstrated a method where an agent can learn the identity of an unknown quantum state and then use that knowledge to erase countless copies of that state at the absolute minimum energy limit allowed by physics. Their work proves that if you can learn a state efficiently, you can erase it efficiently. However, they also uncovered a profound limitation: for certain complex quantum states, learning is so difficult that no efficient method exists to erase them cheaply, even though the laws of physics say it should be possible. This creates a strange gap where the energy cost of erasing a state depends not just on the state itself, but on the computational power of the agent trying to erase it.

The researchers began by imagining a scenario where a source repeatedly produces copies of an unknown quantum state. At first, the agent has no idea what this state is, so erasing each copy requires a significant amount of work. As the agent collects more copies, it can study them to figure out exactly what the state is. Once the state is identified, the agent can simply reverse the steps that created it, turning the copies back into a blank, standard state without spending any extra energy. The challenge was to prove that the process of studying the copies to learn the state did not itself consume enough energy to ruin the savings. To solve this, the team developed a way to make the learning process fully reversible. Instead of making irreversible measurements that destroy information and generate heat, they described a method where the learning algorithm uses quantum operations that can be run backward. This allows the agent to store the knowledge of the state in a memory register and then "unlearn" the process, clearing out any temporary junk data without paying an energy penalty. The only energy cost incurred is at the very end, when the agent finally erases its own memory of the state, a cost that is fixed and independent of how many copies of the state are being erased.

This approach allows the agent to erase a vast number of copies for a single, small energy price, effectively saturating the theoretical limit set by Landauer's principle. The researchers showed that this works beautifully for many types of quantum states that are relevant to current physics and computing, such as those generated by shallow circuits, states with specific patterns of entanglement, or states defined by simple mathematical functions. For these states, the energy cost of erasure is low and can be achieved quickly. However, the story changes dramatically when the researchers looked at more complex states, specifically those known as pseudorandom states. These are states that look so much like truly random noise that no efficient computer program can tell them apart from genuine randomness.

The team proved that for these pseudorandom states, a paradox arises. According to the laws of thermodynamics, if you knew the state, you could erase it cheaply. But because the state is designed to be computationally hard to learn, no efficient algorithm can figure out what it is. Consequently, any agent limited to efficient computation is forced to pay a massive amount of energy to erase these states, nearly the maximum possible cost, even though the state itself is not inherently complex in a physical sense. This result is a powerful "no-go" theorem, showing that for certain quantum systems, the ability to erase information cheaply is blocked not by the laws of physics, but by the limits of computation. It suggests that in the quantum world, the difficulty of learning a state can physically prevent an agent from accessing the energy savings that the state theoretically offers.

The implications of this work extend beyond just erasing states. The same principles apply to extracting work from quantum systems. Just as learning allows for cheap erasure, it also allows for the efficient extraction of maximum energy from a system. If a state is easy to learn, an agent can extract the optimal amount of work. If the state is hard to learn, the agent is stuck extracting only a tiny fraction of the available energy. This establishes a concrete link between the complexity of a quantum state and the physical resources required to manipulate it. The researchers suggest that this connection could lead to new ways of thinking about energy storage and security, such as the concept of an "encrypted battery" where the full energy is only accessible to those who possess the secret key to learn the state's structure.

Ultimately, this paper reshapes our understanding of the relationship between information and energy. It confirms that the act of learning is physically free, provided it is done with the right reversible tools. It also highlights a new kind of barrier in thermodynamics: a barrier created by computational difficulty. In the realm of quantum many-body systems, the complexity of a state can act as a shield, forcing even the most advanced agents to pay a heavy energy price to reset their systems. This work does not just solve a theoretical puzzle; it provides a blueprint for building future quantum technologies that are energy-efficient, while also warning that for the most complex quantum systems, the cost of ignorance may be measured in the very fuel we hope to save.

Technical Summary: Learning to Erase Quantum States

Problem Statement
The paper addresses the thermodynamic cost of erasing quantum states, specifically investigating how an agent's ability to learn about an unknown state influences the energy required for erasure. While Landauer's principle establishes that the work cost to erase a system depends on the entropy (ignorance) of the state, standard formulations often neglect the energy cost of acquiring the knowledge necessary to reduce that entropy. The authors ask whether learning processes themselves incur a fundamental energy cost and whether the ability to learn efficiently impacts the physical resources required to perform thermodynamic tasks like erasure or work extraction.

The specific scenario involves a source producing $N$ copies of an unknown $n$ -qubit state $|\psi_x\rangle$ drawn from a known class of $m$ possible states $\mathcal{C}$ . The agent does not initially know which specific state $x$ is being produced (state-agnostic setting) and must design a protocol to erase these copies.

Methodology
The authors employ tools from quantum learning theory to construct a "learning-to-erase" protocol. The core methodology involves three main steps:

Reversible Learning: The authors demonstrate that any learning algorithm can be "lifted" to a fully reversible quantum algorithm. By replacing irreversible gates with reversible counterparts and converting projective measurements into coherent unitary operations (using ancilla qubits to record outcomes), they construct a unitary learning algorithm $\mathcal{L}$ . This algorithm takes $s$ copies of the unknown state as input and outputs a classical description (or index) $x$ into a memory register $M$ with high probability, without incurring an energy cost during the learning phase itself.
Erasure Protocol Construction: Using the reversible learning algorithm $\mathcal{L}$ $L$ , they construct an erasure protocol $\mathcal{E}(\mathcal{L})$ $E (L)$ .
- The agent uses $s$ copies to run $\mathcal{L}$ and store the learned index $x$ in memory $M$ .
- The index is copied to an auxiliary memory $M'$ .
- The learning process is "uncomputed" (reversed) on the $s$ copies and ancilla to restore them to their initial state, removing any entropy generated during learning.
- The remaining $N$ copies (and the $s$ learning copies) are erased by applying the inverse of the state preparation unitary $U_x^\dagger$ , conditioned on the index stored in $M'$ .
- Finally, the memory $M'$ is erased using standard classical Landauer erasure.
Complexity Analysis: The authors analyze the computational efficiency (time complexity) and thermodynamic cost (work) of these protocols. They relate the work cost to information-theoretic lower bounds (Landauer's limit) and compare them against computational hardness assumptions (e.g., the existence of one-way functions).

Key Contributions and Results

Reversibility of Learning: The paper proves that learning, as a physical process, has no fundamental energy cost if implemented coherently. The only energy cost arises when the agent must erase its own memory to reset for the next task. This allows the agent to acquire knowledge from a few copies and apply it to erase many additional copies at the optimal energy cost, saturating Landauer's limit.
Optimality and Complexity Relations: The work cost of erasing a class of states is shown to be proportional to the logarithm of the number of possible states ( $\log_2 m$ $lo g_{2} m$ ), which corresponds to the complexity of the state class. The authors explicitly relate this cost to four complexity measures:
- Circuit Depth ( $d$ ): For shallow-circuit states, $W \sim \Theta(nd)$ .
- Magic ( $t$ ): For $t$ -doped stabilizer states, $W \sim \Theta(n^2(t+1))$ .
- Entanglement ( $\mathcal{S}$ ): For Matrix Product States (MPS), $W \sim \exp(\tilde{\Theta}(\mathcal{S}))$ .
- Degree ( $k$ ): For low-degree phase states, $W \sim \Theta(n^k)$ .
  In all these physically relevant cases, if the learning algorithm and state preparation are computationally efficient, the erasure protocol is also efficient and achieves the optimal work cost.
Computational Hardness and No-Go Theorems: The paper establishes a "no-go" result for general states. Under standard cryptographic assumptions (specifically, the existence of one-way functions secure against sub-exponential quantum adversaries), there exist classes of pseudorandom states that are information-theoretically easy to erase (low work cost independent of $N$ $N$ ) but computationally hard to learn.
- For these states, any polynomial-time erasure protocol must pay a work cost nearly linear in $N$ (approaching the cost of erasing Haar random states), rather than the information-theoretic minimum which is independent of $N$ .
- This creates a massive gap between the capabilities of computationally bounded agents and information-theoretic limits, representing a genuine quantum many-body phenomenon.
Work Extraction: The authors extend these findings to work extraction. Efficient learning enables efficient extraction of the optimal amount of work. Conversely, for hard-to-learn states, computationally bounded agents can only extract a negligible fraction of the available work in polynomial time.

Significance
The paper claims to establish a concrete connection between quantum learning theory and thermodynamics. Its primary significance lies in demonstrating that the physical resources required for thermodynamic tasks are not solely determined by the state's entropy but are also fundamentally constrained by the computational complexity of learning that state.

Physical Significance of Learning: It highlights that learning processes can be physically reversible and energy-free, challenging the notion that acquiring information inherently costs energy.
Efficiency vs. Optimality: It provides provably efficient protocols for erasing structured states (e.g., shallow circuits, MPS) that achieve the theoretical thermodynamic limit, improving upon previous "compress-to-erase" methods that were information-theoretically optimal but computationally inefficient (exponential time).
Quantum Many-Body Phenomena: The results reveal a stronger limitation than the third law of thermodynamics: even with infinite resources (heat bath), computationally bounded agents cannot achieve the optimal thermodynamic efficiency for certain complex quantum states.
Cryptography and Thermodynamics: The work suggests a link between cryptographic security and thermodynamic costs, implying that "encrypted" quantum states (pseudorandom states) act as a barrier to efficient energy extraction or erasure for agents without the secret key.

The authors conclude that these findings enable the design of provably efficient, energy-optimal protocols for thermodynamic tasks in the quantum regime, with potential applications in quantum metrology initialization, qubit recycling, and the experimental realization of Maxwell's demon using learning agents.

Technical Summary: Learning to Erase Quantum States

More like this