English

An abstract algorithm describes transformations of represented states. A physical computer must realize those states and transformations in matter: voltages, magnetic orientations, photons, or other systems. Physics therefore constrains energy, speed, noise, and which operations are available. Computer science provides ways to ask what information those operations preserve and what tasks they can perform.

The connection is useful when its levels are kept clear. A bit is a unit used to describe alternatives; a memory device is a physical system that represents those alternatives. The same abstract operation can have different implementations and practical costs. Thermodynamics and quantum information show how physical assumptions enter the analysis.

Background: Information theory (§3.8), computer architecture (§4.1), and quantum computing (§6.6). Probability and the distinction between a model and its implementation are the main starting points.

The Relationship: Information as a Physical Quantity

The Physicality of Information and Landauer’s Principle

Consider a reset operation that maps both zero and one to zero. Its output does not tell you which input was present, so the operation is logically irreversible. By contrast, flipping zero to one and one to zero preserves the distinction: applying the flip again recovers the input. Logical reversibility concerns this mapping, before we choose a device.

Landauer’s principle connects erasure to thermodynamics. In the standard model of a symmetric classical memory initially equally likely to hold zero or one, resetting it while coupled to a heat bath at temperature T has an ideal lower bound of k_B T ln 2 on average heat dissipated to that bath. Here k_B is Boltzmann’s constant. At 300 kelvin the value is about 2.9 × 10⁻²¹ joules per erased bit. The relevant quantity is lost uncertainty under the model; resetting a value already known to be zero is not the same initial situation.

This is not a claim that every logic gate or processor instruction must dissipate exactly that amount. Bennett showed how reversible computation can avoid the information loss that incurs this bound. A real implementation still faces switching losses, noise, error correction, finite speed, and the handling of intermediate records. Logical reversibility alone does not make a practical computer consume no energy.

The 2012 experiment by Bérut and colleagues used a trapped colloidal particle as a one-bit memory. Its mean dissipated heat approached the bound for long erasure cycles. That is evidence about a controlled physical implementation and limiting process, not a measurement that every commercial memory operates at the limit.

Maxwell’s demon illustrates the need to account for the whole cycle. A controller can use information about molecules to sort them, but its memory and environment are part of the physical system too. In the standard memory-reset account, restoring the controller for repeated use has an entropy cost. Ignoring that restoration makes the apparent violation possible; including it removes the free thermodynamic cycle.

Entropy: Shared Mathematics, Different Physical Assumptions

Shannon entropy describes uncertainty in a specified probability distribution: H = -Σ p_i log₂ p_i, measured in bits. A fair binary outcome has entropy one bit; a certain outcome has entropy zero. High entropy does not mean useful knowledge or meaningful text. It says that outcomes are less predictable under the distribution being used.

Statistical mechanics also uses probabilities over possible states. Gibbs entropy has the related form S = -k_B Σ p_i ln p_i; when the same probabilities refer to the relevant physical microstates, the formulas give S = k_B ln 2 × H. Boltzmann’s S = k_B ln W is the equally probable case with W accessible microstates. The physical interpretation requires identifying the states, probabilities, and constraints; a shared formula alone does not turn the entropy of any data file into its heat content.

Coarse descriptions matter. Knowing a gas’s temperature and volume leaves many microscopic configurations unspecified. Thermodynamic reasoning tracks macroscopic constraints and exchanges with an environment, while information theory can describe uncertainty about many non-thermal variables. These applications are connected, but the units and modeled quantities must remain explicit.

For example, compressing a file exploits predictability in its representation. It does not prove that the room’s thermodynamic entropy falls by the same number of bits. To discuss physical cost, include the computing device, memory operations, and environment. This distinction lets the mathematical connection be used without confusing description, meaning, and physical process.

Quantum Information: The Fusion of the Fields

A qubit is described by quantum state amplitudes, not by a classical choice whose value we simply do not know. Quantum gates change amplitudes, and measurement produces outcomes according to the chosen measurement. An algorithm must arrange interference so that useful outcomes become likely; the existence of many amplitudes does not let us read every possible answer at once.

The no-cloning theorem rules out a single physical operation that perfectly copies an arbitrary unknown quantum state. It does not forbid copying known classical basis values or preparing a known state again. Entanglement supplies correlations useful for information tasks, but does not by itself transmit a chosen message faster than light. For example, quantum teleportation requires a classical message as well as shared entanglement and local quantum operations.

Closed-system unitary evolution is reversible. Measurement, noise, discarded subsystems, and reset require a broader description, so a full quantum computation cannot simply be described as a sequence with no irreversible operations. Error correction and physical control are part of making the abstract algorithm work.

Shor’s algorithm gives polynomial-time integer factoring in the standard quantum model; no polynomial-time classical factoring algorithm is known. This is a difference in known efficient methods, not a proof that factoring is classically impossible. Standard quantum computation also does not solve the halting problem. Separate computability from complexity, and both from the feasibility of building a sufficiently reliable device.

Ideas about information in black holes and spacetime extend the discussion into theoretical physics. They require their own physical models and evidence. They do not follow merely from the fact that ordinary computers process bits, nor establish that the universe is literally a digital computer.

What This Perspective Changes

Ask which physical system represents the information, which operation is being performed, and which assumptions connect the abstract description to a physical cost or capability. This prevents an energy bound from becoming a claim about every instruction, or a quantum speedup from becoming a claim that all computation is easy.

A useful exercise is to compare a reset and a bit flip as mappings, then explain why only the first merges distinct inputs. Next compute the Shannon entropy of a fair and a biased binary variable. The exercises separate reversibility, uncertainty, and physical realization before combining them.

Resources

Foundational and Bridging Texts

Feynman’s lectures connect physical models with computation. Landauer and Bennett supply the thermodynamic arguments; Bennett’s later explanatory note also discusses common misinterpretations. The Bérut et al. experiment provides a concrete realization to compare with the idealized bound. Jaynes and MacKay develop connections among probability, inference, and physical entropy. Read the assumptions of a physical model alongside its equations.

Resource	Role	Type
Feynman, Feynman Lectures on Computation	Physical limits and models of computation	Depth
Landauer, “Irreversibility and Heat Generation” (1961)	Logical irreversibility and thermodynamic cost	Depth
Bennett, “The Thermodynamics of Computation” (1982)	Reversible computation and thermodynamic accounting	Depth
Jaynes, “Information Theory and Statistical Mechanics” (1957, free)	Information-theoretic methods in statistical mechanics	Depth
MacKay, Information Theory, Inference, and Learning Algorithms (free)	Probability, inference, coding, and physical connections	Depth
SEP, “Computation in Physical Systems” (free)	Conceptual accounts of physical computation	Reference

Quantum Information and the Deeper Connections

Nielsen and Chuang provide the technical quantum-information treatment. Susskind introduces the black-hole information problem for a broader audience. Wheeler and Lloyd propose broader interpretations of information and computation in physics; distinguish their proposals from established results about specific quantum or thermodynamic models.

Resource	Role	Type
Nielsen & Chuang, Quantum Computation and Quantum Information	Quantum information and algorithms	Depth
Susskind, The Black Hole War	A narrative account of black-hole information	Depth
Wheeler, “Information, Physics, Quantum” (1989)	A proposal about information in physical explanation	Depth
Lloyd, Programming the Universe	A broad computational interpretation of physics	Auxiliary

Traps

Trap	Why it misleads	Better response
Treating all entropy as the same physical quantity	The formula needs a specified distribution, state space, and units.	Identify the modeled states before making a thermodynamic claim.
Reading Landauer as a fixed cost for every instruction	The bound concerns erasure under stated physical conditions.	Distinguish lost information, logical reversibility, and device losses.
Treating quantum computation as unrestricted parallel answers	Measurement, interference, noise, and complexity still constrain tasks.	Explain the algorithm and the specific advantage it establishes.

中文

抽象算法描述表示出来的状态怎样变换。物理计算机必须用物质实现这些状态与变换，例如电压、磁性方向、光子或其他系统。因此，物理规律约束能量、速度、噪声和可用操作。计算机科学则提供方法，研究这些操作保留什么信息、能够完成哪些任务。

分清层次后，这种联系才有用。比特是描述不同可能性的单位，存储器件则是表示这些可能性的物理系统。同一个抽象操作，可以具有不同实现与实际成本。热力学和量子信息说明了物理假设怎样进入分析。

背景知识：信息论（§3.8）、计算机体系结构（§4.1）与量子计算（§6.6）。概率，以及模型与实现的区别，是主要起点。

这种关系：信息作为物理量

信息的物理性与 Landauer 原理

考虑一个复位操作：把零和一都变成零。输出不能告诉你原先是哪种输入，因此它在逻辑上不可逆。相比之下，把零翻转为一、把一翻转为零，保留了区别；再次翻转即可恢复输入。逻辑可逆性首先讨论这种映射，还没有涉及具体器件。

Landauer 原理把擦除与热力学联系起来。在标准模型中，一个对称的经典存储单元起初以相同概率保存零或一；它与温度为 T 的热浴接触时，复位过程向热浴耗散的平均热量，具有 k_B T ln 2 的理想下界。其中 k_B 是 Boltzmann 常数。在 300 开尔文时，每个被擦除比特对应约 2.9 × 10⁻²¹ 焦耳。相关量是模型中消除的不确定性；对已知为零的值复位，并不是相同的初始情况。

这不表示每个逻辑门或处理器指令都必须恰好耗散这一数值的热量。Bennett 展示了可逆计算怎样避免导致该下界的信息丢失。实际实现仍要处理开关损耗、噪声、纠错、有限速度和中间记录。仅有逻辑可逆性，不会使实际计算机不消耗能量。

Bérut 等人在 2012 年用被捕获的胶体粒子实现了一个比特的存储模型。擦除周期较长时，平均耗散热量接近该下界。这是关于受控物理实现与极限过程的证据，并不表示所有商用存储器都在该极限运行。

Maxwell 妖说明了为什么必须计算完整周期。控制器可以利用分子信息进行分类，但其存储器与环境也是物理系统的一部分。在标准的存储复位解释中，为了重复使用而恢复控制器，需要付出熵代价。忽略恢复过程，才会产生看似违反定律的结果；把它纳入，就不再得到免费的热力学循环。

熵：共同的数学形式与不同的物理假设

Shannon 熵描述指定概率分布中的不确定性：H = -Σ p_i log₂ p_i，单位为比特。公平二元结果的熵是一比特；确定结果的熵为零。高熵不意味着有用知识或有意义的文本，只说明在所用分布下，结果较难预测。

统计力学也使用可能状态上的概率。Gibbs 熵具有相关形式 S = -k_B Σ p_i ln p_i；当相同概率指向相关的物理微观状态时，两式给出 S = k_B ln 2 × H。Boltzmann 的 S = k_B ln W 是 W 个可达微观状态等概率的情形。物理解释需要明确状态、概率与约束；公式相同，不能让任意数据文件的熵直接等于其所含热量。

粗粒度描述很重要。知道气体的温度与体积后，仍有许多微观配置未确定。热力学推理追踪宏观约束及其与环境的交换；信息论则可以描述许多非热变量的不确定性。两种应用相互联系，但必须明确单位与建模对象。

例如，文件压缩利用表示中的可预测性，却不能证明房间的热力学熵减少了相同数量的比特。讨论物理成本时，还要包括计算器件、存储操作与环境。保持这种区别，才能使用数学联系，而不混淆描述、意义和物理过程。

量子信息：两个领域的融合

量子比特由量子态振幅描述，不是一个我们恰好不知道取值的经典选择。量子门改变振幅，测量则按照所选测量方式产生结果。算法必须安排干涉，使有用结果出现的概率提高；拥有许多振幅，并不能让我们一次读出所有可能答案。

不可克隆定理排除了这样一种统一物理操作：完美复制任意未知量子态。它并不禁止复制已知的经典基态值，也不禁止重新制备已知状态。纠缠提供有助于信息任务的关联，但它本身不能超光速传送任意指定消息。例如，量子隐形传态除了共享纠缠与局部量子操作，还需要传递经典消息。

封闭系统的酉演化是可逆的。测量、噪声、丢弃子系统和复位，需要更广的描述，因此完整量子计算不能简单说成没有不可逆操作的序列。纠错与物理控制，也是让抽象算法工作的组成部分。

Shor 算法在标准量子模型中提供多项式时间整数分解方法；目前没有已知的经典多项式时间分解算法。这是已知高效方法之间的差异，并不证明经典分解不可能。标准量子计算也不能解决停机问题。应区分可计算性、复杂度，以及能否构建足够可靠的器件。

黑洞与时空中的信息问题，把讨论扩展到理论物理。它们需要各自的物理模型与证据，不能仅从普通计算机处理比特这一事实推出，也不能据此认定宇宙字面上就是数字计算机。

这种视角会改变什么

先问哪个物理系统表示信息，执行什么操作，以及哪些假设把抽象描述与物理成本或能力联系起来。这样可以避免把能量下界变成关于每条指令的说法，或把量子加速说成所有计算都容易。

可以先把复位与比特翻转写成映射，解释为什么只有前者合并了不同输入；再计算公平与偏置二元变量的 Shannon 熵。这些练习先区分可逆性、不确定性和物理实现，再将它们联系起来。

资源

基础性与桥梁性文本

Feynman 的讲义联系物理模型与计算。Landauer 和 Bennett 提供热力学论证；Bennett 后来的说明文章也讨论常见误解。Bérut 等人的实验提供了可与理想下界比较的具体实现。Jaynes 和 MacKay 发展了概率、推断与物理熵之间的联系。阅读方程时，也要阅读物理模型的假设。

资源	作用	类型
Feynman, Feynman Lectures on Computation	计算的物理限制与模型	深入
Landauer, “Irreversibility and Heat Generation”（1961）	逻辑不可逆性与热力学成本	深入
Bennett, “The Thermodynamics of Computation”（1982）	可逆计算与热力学核算	深入
Jaynes, “Information Theory and Statistical Mechanics”（1957，免费）	统计力学中的信息论方法	深入
MacKay, Information Theory, Inference, and Learning Algorithms（免费）	概率、推断、编码及其物理联系	深入
SEP, “Computation in Physical Systems”（免费）	物理计算的概念解释	参考

量子信息与更深层连接

Nielsen 与 Chuang 提供量子信息的技术论述。Susskind 面向较广读者介绍黑洞信息问题。Wheeler 与 Lloyd 提出信息和计算在物理学中的更广泛解释；应区分这些主张与特定量子或热力学模型中已经确立的结果。

资源	作用	类型
Nielsen & Chuang, Quantum Computation and Quantum Information	量子信息与算法	深入
Susskind, The Black Hole War	黑洞信息问题的叙述性介绍	深入
Wheeler, “Information, Physics, Quantum”（1989）	关于信息在物理解释中作用的主张	深入
Lloyd, Programming the Universe	对物理学的广泛计算解释	辅助

陷阱

误区	为什么会误导	更好的做法
把所有熵当成同一个物理量	公式需要指定分布、状态空间与单位。	提出热力学结论前，先明确建模状态。
把 Landauer 下界当成每条指令的固定成本	该下界涉及给定物理条件下的擦除。	区分信息丢失、逻辑可逆性与器件损耗。
把量子计算理解为无限并行读取答案	测量、干涉、噪声和复杂度仍然约束任务。	解释算法及其能够确立的具体优势。