pith. sign in

arxiv: 2604.20897 · v1 · submitted 2026-04-21 · 💻 cs.IT · cs.AI· math.IT· physics.comp-ph

Watts-per-Intelligence Part II: Algorithmic Catalysis

Elija Perrier This is my paper

Pith reviewed 2026-05-10 01:25 UTC · model grok-4.3

classification 💻 cs.IT cs.AImath.ITphysics.comp-ph
keywords algorithmic catalysismutual informationLandauer erasurethermodynamic costcomputational speed-updeployment horizontask classreusable structures
0
0 comments X

The pith

Reusable computational structures bound any class-specific speed-up by the algorithmic mutual information between substrate and task descriptor, at a minimum energy cost from Landauer erasure.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a thermodynamic account of algorithmic catalysis, in which reusable structures reduce the number of irreversible operations needed for tasks within a defined class. It establishes that the largest possible speed-up equals the algorithmic mutual information shared between the reusable structure and the description of the task class. Installing that shared information requires erasing at least as many bits as the mutual information value, incurring a minimum energy cost. These two results are combined into a coupling theorem that gives the shortest deployment time at which the catalyst saves more energy than it costs to prepare. The argument is illustrated on an affine SAT class and is used to place learned systems under the same information-thermodynamic limit.

Core claim

Algorithmic catalysis is defined by reusable computational structures that lower irreversible operations for a task class while meeting bounded restoration and structural selectivity constraints. Any resulting speed-up is upper-bounded by the algorithmic mutual information between the substrate and the class descriptor. Installing this information carries a minimum thermodynamic cost set by Landauer erasure. The two limits together produce a coupling theorem that lower-bounds the deployment horizon required for the catalyst to be energetically favourable.

What carries the argument

The algorithmic mutual information between the substrate and the class descriptor, which supplies both the upper bound on achievable speed-up and the minimum Landauer cost that must be amortized.

If this is right

  • The bound and theorem apply directly to an affine SAT class as a concrete illustration.
  • Contemporary learned systems are subject to the same information-thermodynamic constraint on intelligent computation.
  • A catalyst is energetically favourable only after its deployment horizon exceeds the lower bound given by the coupling theorem.
  • No class-specific speed-up can surpass the mutual information upper bound, regardless of the reusable structure chosen.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The bound could be tested by constructing reusable structures for SAT instances, measuring the resulting speed-up, and comparing it against the algorithmic mutual information between structure and class.
  • The same limit may constrain how much pre-computation or pre-training can accelerate later inference steps without a matching increase in shared information.
  • The coupling theorem suggests examining whether current AI systems already operate near or below the predicted deployment horizons for their effective task classes.

Load-bearing premise

Reusable computational structures exist that can meet bounded restoration and structural selectivity constraints for the task class while actually attaining the mutual information bound.

What would settle it

An observed speed-up for a task class that exceeds the measured algorithmic mutual information between the reusable structure and the class descriptor, or a catalyst that becomes net energy-positive before the horizon predicted by the coupling theorem.

read the original abstract

We develop a thermodynamic theory of algorithmic catalysis within the watts-per-intelligence framework, identifying reusable computational structures that reduce irreversible operations for a task class while satisfying bounded restoration and structural selectivity constraints. We prove that any class-specific speed-up is upper-bounded by the algorithmic mutual information between the substrate and the class descriptor, and that installing this information incurs a minimum thermodynamic cost via Landauer erasure. Combining these results yields a coupling theorem that lower-bounds the deployment horizon required for a catalyst to be energetically favourable. The framework is illustrated on an affine SAT class and situates contemporary learned systems within a unified information-thermodynamic constraint on intelligent computation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper develops a thermodynamic theory of algorithmic catalysis in the watts-per-intelligence framework. It identifies reusable computational structures satisfying bounded restoration and structural selectivity for a task class, proves that any class-specific speed-up is upper-bounded by the algorithmic mutual information between substrate and class descriptor, shows that installing this information incurs a minimum Landauer erasure cost, and combines these into a coupling theorem lower-bounding the deployment horizon for energetic favorability. The framework is illustrated on an affine SAT class and used to situate learned systems under a unified information-thermodynamic constraint.

Significance. If the central theorems and constructions can be rigorously established with explicit derivations and verifications, the work would provide a novel coupling between algorithmic mutual information and thermodynamic costs, yielding falsifiable bounds on when catalytic structures become energetically advantageous. This could unify constraints on intelligent computation across information theory and physics, with potential implications for evaluating the efficiency of learned systems.

major comments (3)
  1. [Abstract] Abstract and main theorems: the central coupling theorem lower-bounds deployment horizon by combining the mutual-information upper bound on speed-up with Landauer cost, but the manuscript states that proofs exist without providing any derivations, error analysis, or verification steps for either the mutual information bound or the Landauer installation cost.
  2. [Illustration on affine SAT] Affine SAT illustration: the framework claims to identify reusable structures achieving the mutual information bound under bounded restoration and selectivity, yet no explicit construction, restoration-cost calculation, or check that the information-theoretic bound is attained is supplied; this assumption is load-bearing for the theorem's applicability to real catalysts.
  3. [Coupling theorem] Coupling theorem: the lower bound on deployment horizon presupposes that reusable structures exist which realize the algorithmic mutual information without super-Landauer overhead, but the text provides no construction or quantitative check of this condition, leaving the bound's practical relevance unverified.
minor comments (2)
  1. Notation for 'algorithmic catalyst' and 'class descriptor' is introduced without a dedicated definitions section or table, making it difficult to track how these entities relate to standard mutual information quantities.
  2. [Abstract] The abstract claims the framework 'situates contemporary learned systems' but provides no specific examples or quantitative comparisons with existing systems.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The comments correctly identify places where the manuscript would benefit from expanded explicit derivations and constructions. We respond to each major comment below and will revise accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract and main theorems: the central coupling theorem lower-bounds deployment horizon by combining the mutual-information upper bound on speed-up with Landauer cost, but the manuscript states that proofs exist without providing any derivations, error analysis, or verification steps for either the mutual information bound or the Landauer installation cost.

    Authors: We agree that the main text states the theorems without supplying complete step-by-step derivations or error analysis. The proofs are outlined conceptually in Section 3, but we will add fully expanded derivations, including verification steps and any necessary error bounds, for both the algorithmic mutual-information upper bound on speed-up and the minimum Landauer erasure cost. These will appear in a new dedicated appendix in the revised manuscript. revision: yes

  2. Referee: [Illustration on affine SAT] Affine SAT illustration: the framework claims to identify reusable structures achieving the mutual information bound under bounded restoration and selectivity, yet no explicit construction, restoration-cost calculation, or check that the information-theoretic bound is attained is supplied; this assumption is load-bearing for the theorem's applicability to real catalysts.

    Authors: The affine SAT example is meant to illustrate the framework, but we acknowledge that it currently lacks an explicit reusable-structure construction, explicit restoration-cost calculations, and a direct check that the mutual-information bound is attained. In the revision we will supply a concrete catalyst construction for the affine SAT class, compute the associated restoration costs under bounded restoration, and verify that the information-theoretic bound is achieved. revision: yes

  3. Referee: [Coupling theorem] Coupling theorem: the lower bound on deployment horizon presupposes that reusable structures exist which realize the algorithmic mutual information without super-Landauer overhead, but the text provides no construction or quantitative check of this condition, leaving the bound's practical relevance unverified.

    Authors: The coupling theorem is stated under the assumption that suitable structures exist. The affine SAT illustration is intended to instantiate this assumption, yet quantitative verification is absent. We will augment the example with an explicit construction and quantitative checks confirming that the structure realizes the algorithmic mutual information without super-Landauer overhead, thereby establishing the practical relevance of the deployment-horizon bound. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper claims to prove an upper bound on class-specific speed-up via algorithmic mutual information, a minimum Landauer cost for installing that information, and a resulting coupling theorem for deployment horizon. These steps are presented as first-principles results from information theory and thermodynamics applied to reusable structures satisfying bounded restoration and selectivity. No equations or definitions are supplied that reduce the bounds or theorem to tautological re-statements of the inputs, fitted parameters renamed as predictions, or load-bearing self-citations whose content is unverified. The illustration on affine SAT is described as an application rather than a construction that forces the general result. The derivation chain therefore remains independent of its own outputs and is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

The central claims rest on the existence of reusable structures meeting restoration and selectivity constraints, the applicability of Landauer erasure to algorithmic information installation, and the definition of algorithmic mutual information for computation substrates. No explicit free parameters are stated in the abstract.

axioms (2)
  • domain assumption Landauer erasure sets a minimum thermodynamic cost for installing algorithmic information
    Invoked directly in the proof that installing mutual information incurs thermodynamic cost.
  • ad hoc to paper Reusable computational structures can satisfy bounded restoration and structural selectivity for a task class
    Stated as a constraint that the identified structures must meet; central to defining algorithmic catalysis.
invented entities (2)
  • algorithmic catalyst no independent evidence
    purpose: Reusable computational structure that reduces irreversible operations for a task class
    New concept introduced to unify reusable structures with thermodynamic bounds; no independent falsifiable evidence provided in abstract.
  • class descriptor no independent evidence
    purpose: Entity used to define algorithmic mutual information with the substrate
    Invented to bound speed-up; appears as part of the mutual information definition.

pith-pipeline@v0.9.0 · 5397 in / 1443 out tokens · 33205 ms · 2026-05-10T01:25:04.870553+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

  1. [1]

    IEEE Transactions on VLSI Systems2(4), 398–407 (1994)

    Athas, W.C., Svensson, L.J., Koller, J.G., Tzartzanis, N., Chou, E.Y.C.: Low-power digital systems based on adiabatic-switching principles. IEEE Transactions on VLSI Systems2(4), 398–407 (1994)

    work page 1994

  2. [2]

    Science 301(5637), 1196–1202 (2003)

    Benkovic, S.J., Hammes-Schiffer, S.: A perspective on enzyme catalysis. Science 301(5637), 1196–1202 (2003)

    work page 2003

  3. [3]

    IBM Journal of Research and Development17(6), 525–532 (1973)

    Bennett, C.H.: Logical reversibility of computation. IBM Journal of Research and Development17(6), 525–532 (1973)

    work page 1973

  4. [4]

    International Journal of Theoretical Physics21(12), 905–940 (1982)

    Bennett, C.H.: The thermodynamics of computation—a review. International Journal of Theoretical Physics21(12), 905–940 (1982)

    work page 1982

  5. [5]

    IBM Journal of Research and Development32(1), 16–23 (1988)

    Bennett, C.H.: Notes on the history of reversible computation. IBM Journal of Research and Development32(1), 16–23 (1988)

    work page 1988

  6. [6]

    SIAM Journal on Computing18(4), 766–776 (1989)

    Bennett, C.H.: Time/space trade-offs for reversible computation. SIAM Journal on Computing18(4), 766–776 (1989)

    work page 1989

  7. [7]

    In: Proceedings of the 46th Annual ACM Symposium on Theory of Computing (STOC 2014)

    Buhrman, H., Cleve, R., Koucký, M., Loff, B., Speelman, F.: Computing with a full memory: catalytic space. In: Proceedings of the 46th Annual ACM Symposium on Theory of Computing (STOC 2014). pp. 857–866. ACM (2014)

    work page 2014

  8. [8]

    Theory of Computing Systems62(1), 116–135 (2018)

    Buhrman, H., Koucký, M., Loff, B., Speelman, F.: Catalytic space: non-determinism and hierarchy. Theory of Computing Systems62(1), 116–135 (2018)

    work page 2018

  9. [9]

    In: Proceedings of the 56th Annual ACM Symposium on Theory of Computing (STOC 2024)

    Cook, J., Mertz, I.: Tree evaluation is in spaceO(logn·log logn ). In: Proceedings of the 56th Annual ACM Symposium on Theory of Computing (STOC 2024). pp. 1268–1278. ACM (2024)

    work page 2024

  10. [10]

    SIAM Journal on Computing (2025), journal version of STOC 2024

    Cook, J., Mertz, I.: Tree evaluation is in spaceO(logn·log logn ). SIAM Journal on Computing (2025), journal version of STOC 2024

    work page 2025

  11. [11]

    arXiv:2109.09709 (2021)

    Ebtekar, A.: Information dynamics and the arrow of time. arXiv:2109.09709 (2021)

    work page arXiv 2021

  12. [12]

    Physical Review E111(1), 014118 (2025)

    Ebtekar, A., Hutter, M.: Foundations of algorithmic thermodynamics. Physical Review E111(1), 014118 (2025)

    work page 2025

  13. [13]

    In: Proceedings of the 2nd ACM Conference on Computing Frontiers

    Frank, M.P.: Introduction to reversible computing: motivation, progress, and chal- lenges. In: Proceedings of the 2nd ACM Conference on Computing Frontiers. pp. 385–390. ACM (2005)

    work page 2005

  14. [14]

    Science303(5655), 186–195 (2004)

    Garcia-Viloca, M., Gao, J., Karplus, M., Truhlar, D.G.: How enzymes work: analysis by modern rate theory and computer simulations. Science303(5655), 186–195 (2004)

    work page 2004

  15. [15]

    Girard, V., Koucký, M., McKenzie, P.: Nonuniform catalytic space and the direct sum for space. Tech. Rep. TR15-138, Electronic Colloquium on Computational Complexity (ECCC) (2015),https://eccc.weizmann.ac.il/report/2015/138

    work page 2015

  16. [16]

    Goldreich, O.: Solving tree evaluation ino(logn·log logn )space. Tech. Rep. TR24- 124, Electronic Colloquium on Computational Complexity (ECCC) (2024),https: //eccc.weizmann.ac.il/report/2024/124

    work page 2024

  17. [17]

    Physical Review Letters118(1), 010601 (2017)

    Goldt, S., Seifert, U.: Stochastic thermodynamics of learning. Physical Review Letters118(1), 010601 (2017)

    work page 2017

  18. [18]

    MIT Press, Cam- bridge, MA (2007)

    Grünwald, P.D.: The Minimum Description Length Principle. MIT Press, Cam- bridge, MA (2007)

    work page 2007

  19. [19]

    Springer, Berlin (2004) Watts-per-Intelligence II: Algorithmic Catalysis 17

    Hutter, M.: Universal Artificial Intelligence: Sequential Decisions Based on Algo- rithmic Probability. Springer, Berlin (2004) Watts-per-Intelligence II: Algorithmic Catalysis 17

    work page 2004

  20. [20]

    In: Proceedings of the 2nd International Conference on Learning Representations (ICLR 2014) (2014)

    Kingma, D.P., Welling, M.: Auto-encoding variational Bayes. In: Proceedings of the 2nd International Conference on Learning Representations (ICLR 2014) (2014)

    work page 2014

  21. [21]

    Problems of Information Transmission1(1), 1–7 (1965)

    Kolmogorov, A.N.: Three approaches to the quantitative definition of information. Problems of Information Transmission1(1), 1–7 (1965)

    work page 1965

  22. [22]

    Bulletin of the EATCS (118) (2016)

    Koucký, M.: Catalytic computation. Bulletin of the EATCS (118) (2016)

    work page 2016

  23. [23]

    Harper and Row, New York, 3 edn

    Laidler, K.J.: Chemical Kinetics. Harper and Row, New York, 3 edn. (1987)

    work page 1987

  24. [24]

    Minds and Machines17(4), 391–444 (2007)

    Legg, S., Hutter, M.: Universal intelligence: a definition of machine intelligence. Minds and Machines17(4), 391–444 (2007)

    work page 2007

  25. [25]

    Information and Control61(1), 15–37 (1984)

    Levin, L.A.: Randomness conservation inequalities; information and independence in mathematical theories. Information and Control61(1), 15–37 (1984)

    work page 1984

  26. [26]

    Springer, 4th edn

    Li, M., Vitányi, P.M.B.: An Introduction to Kolmogorov Complexity and Its Applications. Springer, 4th edn. (2019)

    work page 2019

  27. [27]

    Machine Learning37, 355–363 (1999)

    McAllester, D.A.: PAC-Bayesian model averaging. Machine Learning37, 355–363 (1999)

    work page 1999

  28. [28]

    Nature Physics11(2), 131–139 (2015)

    Parrondo, J.M.R., Horowitz, J.M., Sagawa, T.: Thermodynamics of information. Nature Physics11(2), 131–139 (2015)

    work page 2015

  29. [29]

    In: International Conference on Artificial General Intelligence

    Perrier, E.: Watts-per-intelligence: Part I (energy efficiency). In: International Conference on Artificial General Intelligence. Lecture Notes in Computer Science, vol. 16058, pp. 46–57. Springer (2025)

    work page 2025

  30. [30]

    In: Proceedings of the 39th Computational Complexity Conference (CCC 2024)

    Pyne, E.: Derandomizing logspace with a small shared hard drive. In: Proceedings of the 39th Computational Complexity Conference (CCC 2024). LIPIcs, vol. 300, pp. 4:1–4:20. Schloss Dagstuhl (2024), preliminary version: ECCC TR23-168 (2023)

    work page 2024

  31. [31]

    In: Proceedings of the 31st International Conference on Machine Learning (ICML 2014)

    Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and approx- imate inference in deep generative models. In: Proceedings of the 31st International Conference on Machine Learning (ICML 2014). pp. 1278–1286 (2014)

    work page 2014

  32. [32]

    Automatica14(5), 465–471 (1978)

    Rissanen, J.: Modeling by shortest data description. Automatica14(5), 465–471 (1978)

    work page 1978

  33. [33]

    Information and Control 7(1–2), 1–22, 224–254 (1964)

    Solomonoff, R.J.: A formal theory of inductive inference. Information and Control 7(1–2), 1–22, 224–254 (1964)

    work page 1964

  34. [34]

    Physical Review Letters109(12), 120604 (2012)

    Still, S., Sivak, D.A., Bell, A.J., Crooks, G.E.: Thermodynamics of prediction. Physical Review Letters109(12), 120604 (2012)

    work page 2012

  35. [35]

    In: Advances in Neural Information Processing Systems 26 (NeurIPS 2013)

    Stuhlmüller, A., Taylor, J., Goodman, N.D.: Learning stochastic inverses. In: Advances in Neural Information Processing Systems 26 (NeurIPS 2013). pp. 3048– 3056 (2013)

    work page 2013

  36. [36]

    Accounts of Chemical Research34(12), 938–945 (2001)

    Wolfenden, R., Snider, M.J.: The depth of chemical time and the power of enzymes as catalysts. Accounts of Chemical Research34(12), 938–945 (2001)

    work page 2001

  37. [37]

    Journal of Physics A: Mathematical and Theoretical52(19), 193001 (2019)

    Wolpert, D.H.: The stochastic thermodynamics of computation. Journal of Physics A: Mathematical and Theoretical52(19), 193001 (2019)

    work page 2019

  38. [38]

    Physical Review A 40(8), 4731–4751 (1989)

    Zurek, W.H.: Algorithmic randomness and physical entropy. Physical Review A 40(8), 4731–4751 (1989)

    work page 1989