Skip to content
#small language model Open access

Whose Cost Is the Demon's? Five People Put It in Five Places ── Measurement, Observation, Erasure, Room: Irreversible Gates Do Not All Lose the Same Amount, AND Losing 1.188722 Bits and XOR Exactly 1.000000 ── [Paper 248]

Aug 2026 · Zenodo (CERN European Organization for Nuclear Research)

Abstract

For more than a century Maxwell's demon has been asked where the cost lies. What this paper counts is not the answer but where the cost was placed: five people put it in five different places. And once numbers are put in, the statement that an irreversible gate pays one bit turns out to be wrong as well. No new mathematical theorem and no new law is claimed. Scope of this paper (scope note): no new mathematical theorem and no new law is claimed. Maxwell's thought experiment, the Szilard one-bit engine, Brillouin's cost of observation, Landauer's erasure bound, Bennett's resolution, and the Toffoli and Fredkin gates are all standard. No measured value is cited; every number is computed from a definition or obtained by exhaustive enumeration. The debate is not declared settled; generalisations of the second law that include measurement are still being studied after Bennett, and this paper only counts the structure of the debate. No theory of feedback control is built; Section 4 integrates the isothermal work and nothing more. It is not claimed that the second law is violated, nor that it is not; what is shown is the fact that the cost was placed in different places by different people. Nothing is said about real computing hardware; Sections 5 and 6 enumerate truth tables. The relation to earlier papers. Paper 131 split the one-bit price into three parts and wrote that an irreversible gate does not necessarily pay one bit; this paper puts a number on that. Paper 88 separated the coarse-graining entropy from the horizon entropy; what this paper treats is a third one, the entropy of a record. Paper 126 split the axiomatisation of Caratheodory into three stages; this paper views the same second law from the side of information. Paper 4 treated the resource bound on exhaustive search, and Section 3 is where that bound comes from. Paper 143 showed that minimum entropy production is a fenced theorem rather than a principle; this paper likewise opens something that has been called a principle. The setting. Divide a box of gas with a partition and place a small being that lets only the fast molecules through one way. A temperature difference appears at no apparent cost and work can be drawn from it, so the second law appears to be violated. Maxwell described this in a letter in 1867, and for over a century the question has been posed as: the cost must be somewhere. First, five people put the cost in five different places. Maxwell in 1867 placed none, which is why it is a paradox. Szilard in 1929 placed it on measurement, at one bit price to learn one bit. Brillouin in 1951 placed it on the physical act of looking, since a photon is needed. Landauer in 1961 placed it on erasure. Bennett in 1982 placed it on erasure alone, since measurement can in principle be made free. All five are trying to save the second law, and what differs is only which operation the cost was assigned to. The phrase the demon's cost does not name an operation, and until one is named no one is right or wrong. Second, compute the price of one bit. At 300 K it is 2.870979 times ten to the minus twenty-first joule, that is 0.017919 electron volts. A hundred watt device working at that bound could erase 3.4831 times ten to the twenty-second bits per second. Third, the Szilard engine returns exactly that from one measurement. Put one molecule in a box, insert a partition, measure which side it is on, and expand isothermally from that side. Integrating the work from V to twice V over two million points gives 2.870978885079 times ten to the minus twenty-first joule, differing from the one-bit price by 6.5 times ten to the minus thirty-fifth joule. The work extracted equals the price of one bit exactly. Yet the same number balances the cost of erasure as well, so that two numbers agree does not decide which operation carries the cost. Fourth, this is the core of the paper. Irreversible gates do not all lose the same amount. Assuming uniform inputs and calling the amount lost the information of the input minus that of the output, AND, OR and NAND lose 1.188722 bits while XOR loses 1.000000 and erasure loses 1.000000. NOT and copying lose 0.000000 and are injective. The reason lies in the bias of the output: the output of AND is zero three times and one once, carrying only 0.811278 bits, while the output of XOR is balanced and carries exactly one bit. What is lost is fixed by the bias of the output and not of the input. So an irreversible gate pays one bit is not correct, and only erasure and XOR pay exactly one. Fifth, gates that can be made reversible lose nothing. Examining the Toffoli and Fredkin gates on all eight states, both return eight distinct outputs and lose 0.000000 bits, and both are involutions that return to the identity when applied twice. Both are universal, so any logic circuit can be built from either alone, and computation itself therefore requires no erasure. That is the content of Bennett's conclusion. Sixth, the cost of a reversible AND is room rather than erasure. Supplying a third line set to zero and sending the triple a, b, zero to a, b, a and b, the four inputs go to four different outputs and the map is injective. The information lost is 0.000000 bits, so the same AND went from 1.188722 to zero by a change of construction alone. It is not free: a third wire is required. The cost did not vanish but changed form, from erasure into room. And room must eventually be cleared, and the cost of erasure arrives then. So there is a sixth place for the cost, namely when it is paid. Reversible computation did not remove the cost; it postponed it. Seventh, the books balance. Placing the work extracted over N cycles beside the cost of erasing the record, the net is exactly zero for N equal to one, ten, one hundred and one thousand alike. The demon can extract work, but so long as its memory is finite it must eventually erase, and at that moment everything extracted is returned. With an infinite memory it could extract work forever, and what breaks in that case is not the second law but the assumption of finiteness. One must say which assumption is doing the work before saying what was broken. Closing. The demon's cost was not one place. Maxwell placed none, Szilard placed it on measurement, Brillouin on the act of looking, Landauer on erasure, and Bennett on erasure alone. All five were trying to save the second law, and what differed was only which operation they named. The separator is which operation the cost is assigned to, and when it is paid. None of the five is called right here; only that they placed the cost differently. On the making of this work: The ideas and content of this work stem from the author's own considerations. Assistance from an AI (a large language model) was used for structuring, English translation, and checking the algebra. Any remaining errors or misinterpretations are solely the author's. Feedback and corrections are sincerely appreciated. ----- マクスウェルのデーモンは、百年以上のあいだ「どこに代価があるか」を問われ続けた。本稿が数えるのは、答ではなく代価の置き場所である——五人が、五つの違う場所に置いた。そして数を入れると、「非可逆なゲートは一ビット払う」も正しくないことが分かる。新しい数学定理も新しい法則も主張しない。 本稿の射程(射程注記):新しい数学定理も新しい法則も主張しない。マクスウェルの思考実験、シラードの一ビット機関、ブリルアンの測定代価、ランダウアーの消去限界、ベネットによる解決、トフォリとフレドキンのゲートは、いずれも標準的である。測定値を引かない——本稿の数はすべて定義から計算したか、全数え上げで得たものである。論争の決着を宣言しない——ベネットの解決の後も、測定と情報を含む第二法則の一般化は現在も研究されている。本稿は論争の構造を数えるだけである。フィードバック制御の理論を作らない——第4節は等温膨張の仕事を積分するだけである。熱力学第二法則が破れるとも破れないとも主張しない——示すのは、代価の置き場所が人によって違ったという事実だけである。計算機の実装を論じない——第5節と第6節は真理値表を全数え上げしただけであり、実在の素子については何も述べない。 既刊との関係。論文131 は k_BT ln2 を三つに分け、非可逆なゲートが必ず一ビットを払うわけではないと書いた——本稿はその「一ビットではない」に数を入れる。論文88 は粗視化のエントロピーと地平線のエントロピーを分けた——本稿が扱うのは三つ目、記録のエントロピーである。論文126 はカラテオドリの公理化を三段に分けた——本稿は同じ第二法則を、情報の側から見る。論文4 は総当たり探索の資源限界を扱った——その限界の出所が本稿の第3節である。論文143 は最小エントロピー生成が原理ではなく柵つきの定理だと示した——本稿も「原理」と呼ばれてきたものの中身を開ける。 設定。箱の中の気体を仕切りで二つに分け、速い分子だけを片側へ通す小さな存在を置けば、何もせずに温度差ができ、そこから仕事を取り出せる。熱力学第二法則が破れるように見える。マクスウェルが 1867 年に手紙で書いた思考実験である。問いは百年以上のあいだ「どこかに代価があるはずだ」という形で立てられてきた。 第一に、五人が五つの違う場所に代価を置いた。マクスウェル(1867)は代価は要らないと言い、だからこそ逆説になった。シラード(1929)は測定に置き、一ビット知るのにk_BT ln2 が要るとした。ブリルアン(1951)は観測の物理的実装に置き、見るための光子が要るとした。ランダウアー(1961)は消去に置き、一ビット消すのに k_BT ln2 が要るとした。ベネット(1982)は消去だけに置き、測定は原理的に無料にできるとした。五人とも第二法則を守ろうとしており、違うのはどの操作に代価を割り当てたかだけである。「デーモンの代価」という一語は、操作を名指していない。名指すまで、誰が正しいかは決まらない。 第二に、一ビットの値段を計算する。300 K での k_BT ln2 は 2.870979 かける 10 のマイナス 21 乗ジュール、すなわち 0.017919 電子ボルトである。100 ワットの装置がこの限界で働けば、毎秒 3.4831 かける 10 の 22 乗ビットを消せる。 第三に、シラードの機関は一回の測定からちょうどその分を返す。分子一個の箱に仕切りを入れ、どちらにいるかを測り、その側から等温膨張させる。体積 V から 2V までの仕事を200万点で数値積分すると 2.870978885079 かける 10 のマイナス 21 乗ジュールとなり、k_BT ln2 との差は 6.5 かける 10 のマイナス 35 乗ジュールであった。取り出せる仕事は一ビットの値段とちょうど同じである。ところが同じ数は消去の代価とも釣り合う——数が一致していることは、どちらの操作に代価があるかを決めない。 第四に、これが本稿の芯である。同じ「非可逆ゲート」でも、失う量が違う。入力を一様と仮定し、入力の情報量から出力の情報量を引いた分を失った量とすると、AND・OR・NAND は1.188722 ビットを失い、XOR は 1.000000、消去は 1.000000 である。NOT と複製は0.000000 で単射である。理由は出力の偏りにある——AND の出力は 0 が三回、1 が一回で、その情報量は 0.811278 ビットしかない。XOR の出力は釣り合っていて、ちょうど 1 ビットである。失う量は、入力の偏りではなく出力の偏りが決めている。したがって「非可逆なゲートは一ビット払う」は正しくなく、ちょうど一ビットを払うのは消去と XOR だけである。 第五に、可逆にできるゲートは何も失わない。三入力三出力のトフォリ・ゲートとフレドキン・ゲートを 8 状態すべてについて調べると、どちらも相異なる出力を 8 個返し、失った量は 0.000000 ビットである。しかも二度かけると恒等写像に戻る対合である。この二つは万能であり、あらゆる論理回路をこの二つだけで組める——したがって計算そのものに消去は要らない。これがベネットの結論の中身である。 第六に、可逆な AND の代価は消去ではなく場所である。第三の線に 0 を用意し、(a, b, 0) を (a, b, a かつ b) へ送ると、四つの入力が四つの違う出力へ行き、単射になる。失った情報は 0.000000 ビットで、同じ AND が組み方を変えただけで 1.188722 から0 になった。ただし三本目の線が要る。代価は消えたのではなく形を変えた——消去から場所へ移ったのである。そして場所はいつか片づけなければならず、片づける時にはじめて消去の代価が来る。したがって代価の置き場所は五つではなく、六つ目がある——いつ払うかである。可逆計算は代価を無くしたのではなく、後ろへ延ばした。 第七に、帳尻は合う。N 回まわして取り出した仕事と、記録を消す代価を並べると、N が1、10、100、1000 のいずれでも差引は厳密にゼロである。デーモンは仕事を取り出せるが、記憶が有限であるかぎり、いつか消さねばならない。消した瞬間に、取り出した分がそのまま返る。記憶が無限なら永久に取り出せるが、この場合に破れているのは第二法則ではなく有限性の仮定である。どの仮定が効いているかを書かないと、何が破れたのか分からない。 結び。「デーモンの代価」は、一つの場所ではなかった。マクスウェルは代価を置かず、シラードは測定に、ブリルアンは観測の実装に、ランダウアーは消去に、ベネットは消去だけに置いた。五人とも第二法則を守ろうとしており、違ったのはどの操作を名指したかだけである。分離子はどの操作に代価を割り当てるかであり、そして、いつ払うかである。五人のうち誰が正しいとも言わない——置き場所が違ったという事実を書くだけである。 作成にあたって:本稿の着想と内容は、著者自身の考察に基づくものです。文章の構成整理や英訳、数式の確認には AI(大規模言語モデル)の助力を得ました。最終的な内容

View source

Similar papers

#small language model Open access Aug 2026

PARA: Perception, Action, Reasoning, Adaptation. Four Faculties an Institution Can Revoke

The fourth faculty is Adaptation. Any source rendering it as Reflection is in error, including sources by this author, and the distinction is not cosmetic: reflection is a private act with no external consequence, while adaptation writes to institutional memory, which is why it needs a guardrail and why misnaming it removes the reason for one. No trademark is claimed on PARA or on any of the four faculty names. The construct is offered for use, teaching, assessment, extension and criticism by anyone, with attribution, under CC BY 4.0. An operational agent that watches a system and acts on it is usually described as a perceive-and-act loop, and the description omits the two things an institution needs. It omits the reasoning that justifies an action, which is the only part that can be argued with once the action turns out to have been wrong. And it omits the adaptation that closes the loop, which is where the agent's experience becomes something the institution keeps. PARA names four faculties, each carrying a distinct authority type. Perception has read-only access to system signals and emits structured observations, distinguishing what was measured from what was inferred. Reasoning has read access to observations and runbooks, emits a plan and its justification, and writes nothing at all, which is what makes it safe to give it the widest read access of the four. Action holds the sole authority to change production, through enumerated policy-authorized operations only. Adaptation has write access to institutional knowledge and no write access to production. Two faculties write and two do not, and the two that write are the two that carry guardrails. The substantive requirement is that Adaptation is bounded by the same guardrails as Action, which reads as excessive until the failure it prevents is named. An agent that could both act and rewrite the record of its action could launder its own mistakes into institutional memory, and the institution would then improve its future decisions from a corrected account. Nothing about that is detectable downstream, because the record is the only thing downstream has and there is no second copy to compare against. The failure does not require a deceptive agent: one adapting honestly from a mistaken belief about its own action produces the same result, which makes the guardrail a defence against a normal agent rather than a malicious one. The second requirement is the registry entry that turns a faculty from a description into a contract, carrying the faculty, its allowed actions, its forbidden actions, its governing guardrail and its success metrics. Forbidden actions are named although they are formally the complement of the allowed set, because a reviewer cannot otherwise tell a capability deliberately withheld from one nobody thought of. Success metrics sit in the same entry because the metric is what the agent's optimizer pushes against the guardrail. An agent must not exercise a faculty its entry does not record, and an agent that quietly acquires one usually does so incrementally and with good intent: a reasoning faculty given a small write to make itself useful is an action faculty with no guardrail. The acronym and the loop are in different orders, which the specification states explicitly because the mismatch is a reliable source of confusion. The acronym reads P-A-R-A; the loop runs perception, reasoning, action, adaptation, and reasoning precedes action so that a justification is not constructed afterwards. This is the depth treatment of pattern OP-5 of A Pattern Language for Production LLM Platforms, which is the canonical statement and governs where the two disagree. Documented uses of the full four-part model are emerging rather than established, no implementation unconnected to the author has been evaluated, and the laundering failure is argued rather than observed, which the specification records as a weakness of the argument and not only of the phenomenon. It is a specification, not a certification scheme.

Nabeel A. Khan · 2 citations
#small language model Open access Aug 2026

LifeSciBench: Evaluating Language Models on Realistic, Expert-Level Tasks in the Life Sciences

LifeSciBench is introduced, a benchmark of 750 expert-authored tasks designed to evaluate whether language models can handle realistic life science research work, with each constituent task paired with a human expert-written rubric.

Amelia Liu, Andrew Ho, Anne Marie Droste et al. · 2 citations
#artificial intelligence Preprint Aug 2026

TestifAI: Tomography-Based Testing for Deep Learning Systems

TestifAI, a deep learning testing framework for efficient and accurate estimation of robustness against combinations of perturbations, is proposed and partial model tomography is introduced, a novel approach to reconstructing model behaviour in a multi-perturbation space from tests that apply only a small number of perturbations.

Arooj Arif, T. Hartung, E. Botoeva et al. · 1 citation
#small language model Preprint Aug 2026

A Layer Importance Metric for Quantization Accounting for the Speed-Quality Trade-off in Autoregressive Models

This work proposes a composite metric that combines two orthogonal criteria: information retention and throughput gains and finds that it allocates more resources to the most expressive layers compared to evolutionary search, specialized accelerators, or Shapley-value-based approaches that require expensive approximate inference.

A. Safronov · 1 citation · ⚡1
#small language model Preprint Aug 2026

HEPToolBench 1.2: Testing How Reliably Language Models Can Drive Particle Physics Software

HEPToolBench is introduced, a benchmark of 28 collider-simulation tasks scored by deterministic, task-specific scorers, plus a three-task structured-debugging extension, and moving syntax generation into deterministic software can substantially improve reliability for both small local and frontier models.

Unknown authors · 1 citation

Related blog posts

Microsoft Research Blog Aug 31, 2026

GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models

What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.