Pith. sign in

REVIEW 4 major objections 4 minor 31 references

This paper argues that every learning compositionality problem has a tangent lift, and that machine learning can be reformulated as the search for a fixed point where iterated infinitesimal non-compositionality stabilizes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 00:07 UTC pith:QN4K2YXZ

load-bearing objection New categorical idea; central obstruction undefined, so the theorems don't yet land on a concrete object. the 4 major comments →

arxiv 2607.15107 v1 pith:QN4K2YXZ submitted 2026-07-16 cs.LG math.CT

Learning in Infinitesimal Non-Compositional Sketches

Sridhar Mahadevan This is my paper
classification cs.LG math.CT MSC 18C1018D9968T01
keywords learning sketchestangent categoriescoalgebraic fixed pointsinfinitesimal non-compositionalityfactorization obstructionsdiagrammatic backpropagationequivariance diagnosticscategorical machine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes that machine-learning failures are not arithmetic errors but failures of a learned diagram to factor through the intended quotient of a learning sketch. It then introduces the core move: apply the tangent functor to the model and ask whether infinitesimal perturbations preserve the same factorization constraints, yielding a new obstruction called Infinitesimal Non-Compositionality (INC). Iterating this lift produces a tower of factorization problems, and the paper's central thesis is that learning is the search for a coalgebraic fixed point where this tower stabilizes, written νT_INC. A sympathetic reader would care because this reframes loss functions as scalarizations of factorization failures and offers a graded diagnostic for where and at what differential order a model departs from its intended structure.

Core claim

The central claim is that non-compositionality is the failure of a universal factorization problem, not a numeric residual, and that this failure has a canonical infinitesimal refinement: given a model D admissible for a learning sketch S, the tangent model TD poses the same factorization problem, and its obstruction INC(D) = Obs(Fact_S(TD)) is a new learning signal. The paper defines an endofunctor T_INC that iterates the tangent lift, producing a tower D, TD, T^2D, ... of factorization problems, and formulates ML as the search for a coalgebraic fixed point νT_INC. Using the Aczel–Mendler final coalgebra theorem, it proves existence of final INC coalgebras whenever T_INC admits a set-based

What carries the argument

The key mechanism is the INC endofunctor T_INC, which transports the obstruction to tangent factorization at level n to the obstruction at level n+1, making learning into the search for a fixed point of an unfolding operation. Supporting machinery includes: learning sketches (S,D,L,K) whose models factor through the commutativity quotient q_D; Cockett–Cruttwell tangent category axioms that make the tangent functor T well-behaved on admissible models; set-based and accessible class realizations used to invoke final coalgebra theorems; and complete, rho-contractive, bisimulation-conservative metric realizations that yield convergence of the INC tower and finite-error bounds for approximate unf

Load-bearing premise

The theorems hinge on the assumption that the iterated-tangent operation can be realized as a set-based or metric endofunctor with a final carrier and a contraction constant below one—a realization the paper defines but never exhibits for a concrete learning problem.

What would settle it

Take a concrete learning sketch, such as the permutation-equivariance sketch for a content-only Transformer, compute scalarized INC energies for a trained model and its tangent lifts, and check for geometric convergence of the tower; if the energies fail to converge or a set-based realization provably cannot create its final carrier for that sketch, the central stabilization claim collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the thesis is correct, every scalar loss in ML can be read as a scalarization of a factorization obstruction, and its gradient-like terms as scalarizations of tangent obstructions.
  • Exact equivariance—such as permutation equivariance of content-only Transformers—is inherited by all tangent orders, so the INC tower provides a graded diagnostic of where and at what differential order approximate equivariance breaks.
  • A stabilized INC tower (a state near νT_INC) would serve as a certificate that further tangent unfolding reveals no new higher-order or out-of-distribution obstruction not already captured.
  • Presentation descent for parameterized optimizers separates gauge or redundancy artifacts from genuine INC data, clarifying which parameter updates are true learning signals.

Where Pith is reading between the lines

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

  • If the stabilization claim is right, then computing successive tangent obstructions on a real network could yield a practical stopping criterion: the level at which the tower plateaus would indicate that no new infinitesimal defect directions remain, a testable and implementation-ready extension.
  • The framework implies that connection-symmetric second-order terms, not just antisymmetric Lie brackets, may carry predictive signal in adapter composition; one could ablate the symmetric acceleration coefficient in a LoRA-style setting to test this directly.
  • The paper's own LINCS–final-coalgebra comparison conjecture, if resolved, would yield a transfer principle: any sketch-preserving tangent functor would carry INC guarantees between learning domains, a consequence the paper does not develop.
  • A concrete way to stress the framework is to measure scalarized tower energies on small trained models; failure of geometric convergence would indicate that the contractive metric realization hypothesis does not hold in practice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes LINCS, a categorical framework that recasts machine learning as the repair of non-compositionality of diagrams. A learning sketch S = (S,D,L,K) is used to define a factorization problem Fact_S(D) for a model D, and the paper writes Obs(Fact_S(D)) for the obstruction to this problem. Applying a tangent functor T gives the tangent factorization problem Fact_S(TD), and the infinitesimal non-compositionality is defined as INC(D) := Obs(Fact_S(TD)). The paper then introduces an INC endofunctor T_INC, defines LINCS coalgebras, and states existence and stabilization theorems for final T_INC-coalgebras under set-based class, accessible set, or metric realizations. It also sketches connections to Lie brackets, connections, homotopical refinement, and several application domains, with a detailed comparison to νT_INC left as a conjecture.

Significance. If fully realized, the idea of using tangent lifts of factorization obstructions as a learning signal is a novel and potentially interesting synthesis of sketch semantics, tangent categories, and coalgebra. The paper is commendably explicit about the conditional nature of its theorems and about the remaining open problems. However, as written, the central mathematical object is not defined: Obs is never given a precise meaning, and the categories and endofunctors built from it are therefore under-specified. No concrete realization of the proposed structure is constructed, so the main thesis that ML is the search for a final INC coalgebra is not established. The significance of the framework cannot be assessed until these gaps are repaired.

major comments (4)
  1. [Defs. 3 and 10; Sec. 8] Obs(Fact_S(D)) is used as the central object from Definition 3 onward, and INC(D) is defined in Definition 10 as Obs(Fact_S(TD)). Yet Obs is never defined as a mathematical operation: what kind of object is an obstruction? In which category does it live? Is it a set, a class, a functor, a homotopy fiber, or an object of a to-be-constructed obstruction category? Without a definition, INC(D), the INC endofunctor T_INC (Definition 20), and the INC category (Definition 16) are not well-defined. Theorems 1-3 assume realizations of T_INC, but a realization of an undefined functor is not a well-formed hypothesis. This is the central load-bearing gap in the paper.
  2. [Sec. 3, Def. 5; Sec. 7, Axiom 2] The paper repeatedly quantifies over 'admissible models' without defining the class of admissible models. Definition 2 defines strict/compositional models, but admissibility is used as a stricter notion that is required to be preserved by the tangent functor. Axiom 2 simply states this preservation property. Consequently, tangent stability is an assumption built into the axiomatics, not a derived or verifiable property. To make the framework testable, the paper must specify what makes a model admissible and give conditions under which admissibility is preserved by T.
  3. [Sec. 8, Thms. 1-3; Sec. 12] Theorems 1-3 are conditional on the existence of set-based class, accessible set, or metric realizations of T_INC, but no such realization is constructed, nor is it shown that any concrete learning sketch satisfies the hypotheses. The paper's own Section 12 identifies the construction of the comparison LINCS(C) → νT_INC as the central open problem. Thus, even if Obs were defined, the main claim that ML can be formulated as the search for νT_INC would remain a research program rather than an established theorem. At least one verified instance of the realizations, or a precise class of learning sketches for which they exist, is needed.
  4. [Sec. 8, Def. 20; Sec. 6, Prop. 1] Definition 20 does not show that T_INC is a functor. It specifies an action on factorization problems of the form Fact_q(T^n D), but no assignment on morphisms of INC(C) is given, and INC(C) itself has no explicit morphisms. Proposition 1 additionally assumes naturality of the 'obstruction assignment' and of restriction maps, but since Obs and the category of obstruction objects are undefined, these naturality conditions are not meaningful as stated. The coalgebraic framework requires a genuine endofunctor on a genuine category; the present text provides only a suggestive operation on objects.
minor comments (4)
  1. [Notation throughout] There are many inconsistencies in math notation: 'Fact_S' and 'FactS', 'TD' and 'T D', and similar spacing issues. Please standardize the notation for readability.
  2. [Remark 1] The symbol D is used both for a learning model D:J→C and for the infinitesimal object D in the symmetric monoidal category. This is confusing; please use a different symbol for one of them.
  3. [Sec. 4] The statement that exact equivariance is inherited by all tangent orders is a standard differentiation fact, but the subsequent claim that the INC tower is 'informative' for approximate or restricted equivariance is not formalized. Please state precise assumptions or mark the stronger assertions as conjectures.
  4. [Sec. 4 and references] Concrete applications (GIRL, ALLORA, LASKO, DB) are deferred to companion papers. If these are not available to the reader, the examples cannot be verified. Either include the necessary details or describe them as planned applications.

Circularity Check

1 steps flagged

No load-bearing circularity; the core existence/convergence theorems are conditional transfers of external fixed-point results, with one definitional restatement and deferred validation.

specific steps
  1. self definitional [Definition 5 (Section 3) vs. Axiom 2 (Section 7.1)]
    "Axiom 2 (Tangent admissibility). If D:J→C is an admissible model of a tangent learning sketch S, then TD:J→C is also an admissible model of S, and the assignment D↦TD agrees with the tangent action specified by the sketch."

    Definition 5 has already defined a tangent learning sketch by exactly this stability condition: 'if J=Path(S) and D:J→C is admissible ... then TD:J→C is again admissible.' Axiom 2 repeats the definition as an axiom, so the paper's 'tangent stability' is built into the class of sketches studied rather than derived from Cockett-Cruttwell structure. This is a definitional redundancy; since the existence and convergence theorems are explicitly conditional on realization hypotheses, the redundancy does not by itself force the final-coalgebra conclusions.

full rationale

The paper is a conditional categorical framework rather than an empirical predictor. INC(D) and T_INC are built from factorization problems and tangent lift (Defs. 3, 10, 20); Theorems 1–3 are transparently conditional: they assume a set-based/accessible/metric realization and then transfer Aczel–Mendler, Barr, and Banach/Kozen–Ruozzi fixed-point results. No fitted parameter is renamed a prediction, and no theorem's conclusion is obtained solely by substituting its own definition. The principal weaknesses are completeness gaps, not circularity: Obs(Fact_S(D)) is never defined as a mathematical object (Defs. 3 and 10), so T_INC is not fully specified, and Sec. 12 openly lists the comparison LINCS(C)≃νT_INC as the central open problem. The same-author companion papers (GIRL, DB, ALLORA, LASKO) are cited for examples and pending experiments, not as load-bearing proof of existence or uniqueness. The only near-tautological move is Axiom 2 restating the admissibility-stability condition already present in Definition 5; it is a minor definitional redundancy rather than a circular derivation. Overall the derivation chain has no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 3 invented entities

The framework rests on several heavy assumptions: an undefined obstruction operation Obs, an axiom asserting tangent admissibility, and the unproven existence of set-based/accessible or metric realizations of T_INC. No free numerical parameters are fitted, but the theoretical scaffolding leans on in-preparation same-author companion papers for its examples and instantiations.

axioms (6)
  • ad hoc to paper Tangent admissibility: if D is an admissible model, then TD is also admissible (Def. 5, Axiom 2).
    This is the core stability property that makes T_INC and INC categories well-defined; it is assumed rather than derived.
  • ad hoc to paper The obstruction operation Obs exists and is a well-defined object/functor on factorization problems.
    Obs is used throughout (Def. 3, Def. 10) but never defined; INC(D) is literally Obs(Fact_S(TD)), so all subsequent results inherit this gap.
  • domain assumption Cockett-Cruttwell tangent category axioms on C.
    The tangent lift T is required to be a tangent functor; this is standard background structure, but its existence for a given learning category is not shown.
  • ad hoc to paper T_INC admits a set-based class realization or accessible set realization that creates the final carrier.
    Theorems 1 and 2 are conditional on this; no concrete realization is constructed for any learning sketch.
  • ad hoc to paper For Theorem 3, the metric realization (M_D, d_D, F_D) is complete, rho-contractive, and bisimulation-conservative.
    These hypotheses are assumed and not verified; they are needed for the fixed-point stabilization result.
  • standard math Aczel-Mendler final coalgebra theorem and Barr's accessible-coalgebra theorem.
    Used in the proofs of Theorems 1 and 2; standard results cited correctly, but they apply only under the assumed realizations.
invented entities (3)
  • INC (Infinitesimal Non-Compositionality) obstruction no independent evidence
    purpose: A new learning signal defined as the obstruction to the tangent-lifted factorization problem.
    No falsifiable prediction or measurement is attached; it is a categorical construction.
  • T_INC endofunctor and final INC coalgebra nu T_INC no independent evidence
    purpose: Iterates tangent lifts; the proposed fixed-point target of learning.
    Existence is conditional on unverified realizations; no concrete instance is given.
  • Tangent Learning Sketches and LINCS categories no independent evidence
    purpose: Bundle sketches with tangent structure so that factorization problems transport to the tangent level.
    A definitional layer; not yet tied to an experimental or observable consequence.

pith-pipeline@v1.3.0-alltime-deepseek · 22593 in / 13868 out tokens · 146969 ms · 2026-08-02T00:07:21.415095+00:00 · methodology

0 comments
read the original abstract

This paper develops a categorical framework -- Learning in Infinitesimal Non-Compositional Sketches (LINCS) -- as the repair of non-compositionality: failures of diagrams to factor through quotient sketches lifted to the tangent category setting. Machine learning problems are specified as sketches: graphs with commutativity conditions $\mathcal D$, limit cones $\mathcal L$, and colimit cocones $\mathcal K$, generalizing the usual scalarization of loss functions or vector space assumptions. Non-compositionality is defined purely as failure of a universal factorization problem, not as arithmetic error between the desired and actual predictions. Given a learning sketch $\mathbb S=(S,\mathcal D,\mathcal L,\mathcal K)$, whose underlying graph is $S$, and a model $D:J \rightarrow C$, the base defect is the obstruction to factorization $\mbox{Obs}(\mbox{Fact}_{\mathbb S}(D))$. The tangent lift applies the tangent functor $T$ to obtain $TD:J \rightarrow C$, and LINCS is defined as the obstruction $\mbox{Obs}(\mbox{Fact}_{\mathbb S}(TD))$ -- asking whether infinitesimal perturbations preserve the compositionality constraints.The paper also introduces Tangent Learning Sketches, which are sketches equipped with Cockett-Cruttwell tangent structure. The paper defines the INC endofunctor, which iterates the tangent lift, producing a tower $D,TD,T^2D, \cdots$ of factorization problems. ML is thereby formulated as the search for a coalgebraic fixed point where successive tangent unfoldings stabilize ($\nu T_{\mbox{INC}}$). Using the Aczel--Mendler theorem, we prove existence of a final INC coalgebra whenever $T_{\mbox{INC}}$ admits a set-based class realization that creates its final carrier. A detailed experimental evaluation of LINCS is underway in a number of concrete ML settings, including deep learning, large language models, and reinforcement learning, and is described in companion papers.

Figures

Figures reproduced from arXiv: 2607.15107 by Sridhar Mahadevan.

Figure 1
Figure 1. Figure 1: Every learning compositionality problem has a tangent lift. At left, a candidate model [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

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

31 extracted references · 3 canonical work pages

  1. [1]

    A final coalgebra theorem

    Peter Aczel and Nax Paul Mendler. A final coalgebra theorem. In Category Theory and Computer Science, volume 389 of Lecture Notes in Computer Science, pages 357--365. Springer, 1989

    1989

  2. [2]

    Locally Presentable and Accessible Categories, volume 189 of London Mathematical Society Lecture Note Series

    Ji r \'i Ad \'a mek and Ji r \'i Rosick \'y . Locally Presentable and Accessible Categories, volume 189 of London Mathematical Society Lecture Note Series. Cambridge University Press, 1994

    1994

  3. [3]

    Terminal coalgebras in well-founded set theory

    Michael Barr. Terminal coalgebras in well-founded set theory. Theoretical Computer Science, 114 0 (2): 0 299--315, 1993. doi:10.1016/0304-3975(93)90076-6

    doi:10.1016/0304-3975(93)90076-6 1993

  4. [4]

    On the limitations of sketches

    Michael Barr and Charles Wells. On the limitations of sketches. Canadian Mathematical Bulletin, 35 0 (3): 0 287--294, 1992. doi:10.4153/CMB-1992-040-7. URL https://doi.org/10.4153/CMB-1992-040-7

    work page doi:10.4153/cmb-1992-040-7 1992

  5. [5]

    Category Theory for Computing Science

    Michael Barr and Charles Wells. Category Theory for Computing Science. Centre de Recherches Math \'e matiques, 3 edition, 1999

    1999

  6. [6]

    J. R. B. Cockett and G. S. H. Cruttwell. Differential structure, tangent structure, and SDG . Applied Categorical Structures, 22 0 (2): 0 331--417, 2014

    2014

  7. [7]

    J. R. B. Cockett and G. S. H. Cruttwell. Connections in tangent categories. Theory and Applications of Categories, 32 0 (26): 0 835--888, 2017

    2017

  8. [8]

    J. R. B. Cockett and G. S. H. Cruttwell. Differential bundles and fibrations for tangent categories. Cahiers de Topologie et G \'e om \'e trie Diff \'e rentielle Cat \'e goriques , 59 0 (1): 0 10--92, 2018

    2018

  9. [9]

    Esquisses et types des structures alg \'e briques

    Charles Ehresmann. Esquisses et types des structures alg \'e briques. Buletinul Institutului Politehnic din Ia s i , 14: 0 1--14, 1968

    1968

  10. [10]

    Backprop as functor: A compositional perspective on supervised learning

    Brendan Fong, David Spivak, and R \'e my Tuy \'e ras. Backprop as functor: A compositional perspective on supervised learning. Proceedings of the 34th Annual ACM/IEEE Symposium on Logic in Computer Science, pages 1--13, 2019

    2019

  11. [11]

    The Elements of Statistical Learning: Data Mining, Inference, and Prediction

    Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer-Verlag, 2nd edition, 2009

    2009

  12. [12]

    Applications of metric coinduction

    Dexter Kozen and Nicholas Ruozzi. Applications of metric coinduction. Logical Methods in Computer Science, 5 0 (3:10): 0 1--19, 2009. doi:10.2168/LMCS-5(3:10)2009

    work page doi:10.2168/lmcs-5(3:10)2009 2009

  13. [13]

    Classifying tangent structures using Weil algebras

    Poon Leung. Classifying tangent structures using Weil algebras. Theory and Applications of Categories, 32 0 (9): 0 286--337, 2017. URL http://www.tac.mta.ca/tac/volumes/32/9/32-09.pdf

    2017

  14. [14]

    The Functorial Semantics of Lie Theory

    Benjamin MacAdam. The Functorial Semantics of Lie Theory. PhD thesis, University of Calgary, 2022. URL https://arxiv.org/abs/2301.00305

    Pith/arXiv arXiv 2022

  15. [15]

    Universal causality

    Sridhar Mahadevan. Universal causality. Entropy, 25 0 (4): 0 574, 2023. doi:10.3390/e25040574. URL https://doi.org/10.3390/e25040574

    work page doi:10.3390/e25040574 2023

  16. [16]

    Higher algebraic K -theory of causality

    Sridhar Mahadevan. Higher algebraic K -theory of causality. Entropy, 27 0 (5): 0 531, 2025. doi:10.3390/e27050531. URL https://doi.org/10.3390/e27050531

    doi:10.3390/e27050531 2025

  17. [17]

    ALLORA : A Lie -algebraic LoRA method for composable neural adapters, 2026 a

    Sridhar Mahadevan. ALLORA : A Lie -algebraic LoRA method for composable neural adapters, 2026 a . Manuscript in preparation

    2026

  18. [18]

    Latent confounded causal discovery via Lie bracket geometry

    Sridhar Mahadevan. Latent confounded causal discovery via Lie bracket geometry. arXiv preprint arXiv:2606.19610, 2026 b . URL https://arxiv.org/abs/2606.19610

    Pith/arXiv arXiv 2026

  19. [19]

    Categories for AGI , 2026 c

    Sridhar Mahadevan. Categories for AGI , 2026 c . URL https://people.cs.umass.edu/ mahadeva/papers/catagi.pdf. Textbook manuscript

    2026

  20. [20]

    Gradient infinitesimal reinforcement learning in tangent categories

    Sridhar Mahadevan. Gradient infinitesimal reinforcement learning in tangent categories. Manuscript in preparation, 2026 d

    2026

  21. [21]

    Infinitesimal causality

    Sridhar Mahadevan. Infinitesimal causality. arXiv preprint arXiv:2606.24621, 2026 e . URL https://arxiv.org/abs/2606.24621

    Pith/arXiv arXiv 2026

  22. [22]

    Kan extension transformers: A categorical unification of attention, diffusion, and predict-detach self-conditioning

    Sridhar Mahadevan. Kan extension transformers: A categorical unification of attention, diffusion, and predict-detach self-conditioning. arXiv preprint arXiv:2605.27259, 2026 f . URL https://arxiv.org/abs/2605.27259

    Pith/arXiv arXiv 2026

  23. [23]

    Agentic skill optimization over Lie algebroids

    Sridhar Mahadevan. Agentic skill optimization over Lie algebroids. arXiv preprint arXiv:2607.11493, 2026 g . URL https://arxiv.org/abs/2607.11493

    Pith/arXiv arXiv 2026

  24. [24]

    Universal decision learners

    Sridhar Mahadevan. Universal decision learners. arXiv preprint arXiv:2605.30694, 2026 h . URL https://arxiv.org/abs/2605.30694

    Pith/arXiv arXiv 2026

  25. [25]

    Accessible Categories: The Foundations of Categorical Model Theory, volume 104 of Contemporary Mathematics

    Michael Makkai and Robert Par \'e . Accessible Categories: The Foundations of Categorical Model Theory, volume 104 of Contemporary Mathematics. American Mathematical Society, 1989

    1989

  26. [26]

    Kevin P. Murphy. Machine Learning: A Probabilistic Perspective. The MIT Press, 2012. ISBN 978-0262018020

    2012

  27. [27]

    Abstract tangent functors

    Ji r \'i Rosick \'y . Abstract tangent functors. Diagrammes, 12: 0 1--11, 1984

    1984

  28. [28]

    Jan J. M. M. Rutten. Universal coalgebra: a theory of systems. Theoretical Computer Science, 249 0 (1): 0 3--80, 2000

    2000

  29. [29]

    David I. Spivak. Database queries and constraints via lifting problems. Mathematical Structures in Computer Science, 24 0 (6): 0 e240602, 2014. doi:10.1017/S0960129513000479. URL https://doi.org/10.1017/S0960129513000479

    doi:10.1017/s0960129513000479 2014

  30. [30]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. Adaptive Computation and Machine Learning Series. The MIT Press, second edition, 2018. ISBN 9780262039246. URL http://incompleteideas.net/book/the-book.html

    2018

  31. [31]

    Automated crossword solving

    Eric Wallace, Nicholas Tomlin, Albert Xu, Kevin Yang, Eshaan Pathak, Matthew Ginsberg, and Dan Klein. Automated crossword solving. arXiv preprint arXiv:2205.09665, 2022

    Pith/arXiv arXiv 2022