Pith. sign in

REVIEW 3 major objections 6 minor 24 references

Operator-Based Machine Intelligence: A Hilbert Space Framework for Spectral Learning and Symbolic Reasoning

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read This paper argues that machine learning and symbolic reasoning can both be carried out by bounded linear operators on Hilbert spaces, replacing deep neural networks with spectral, interpretable mappings.

desk verdict A clear, competent survey of Hilbert-space ML methods wrapped in an unsupported claim of a unified reasoning framework; no new experiments or theorems, so the central pitch does not hold. read the letter →

arxiv 2507.21189 v1 pith:UYWJ4OQD submitted 2025-07-27 cs.LG

classification cs.LG MSC 46E2247A0568T0542C40
keywords HilbertspaceoperatorlearningspectralmethodsscatteringtransformKoopmansymbolicreasoningRKHSinterpretablemachine
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that machine learning can be recast as sampling and computation in an infinite-dimensional Hilbert space, with learning, feature extraction, reasoning, and inference handled by the same family of operators rather than by deep neural networks. Its central proposal is that symbolic operations—implication, transitivity, analogy—can be implemented as bounded linear or spectral operators acting on function-valued embeddings, so that chaining relations is just composing operators. A sympathetic reader would care because, if the proposal holds, it offers a route to models that are interpretable through their spectra, stable under small perturbations, and competitive with neural networks, especially in low-data and safety-sensitive settings. The paper grounds this proposal in existing operator, kernel, scattering, and Koopman methods, and reports benchmark comparisons suggesting parity or better efficiency against neural baselines.

What carries the argument

The load-bearing mechanism is the bounded linear operator between Hilbert spaces, estimated as a regularized operator regression: given paired samples {(f_i, g_i)}, one solves min_T Σ_i ∥T f_i − g_i∥² + λ∥T∥_HS², with the Hilbert–Schmidt norm as regularizer. When T is a reasoning operator, the same least-squares form learns a map T_r with T_r f_A ≈ f_B; transitivity is carried by operator composition T_{r_2} T_{r_1}, and spectral reasoning by diagonal modulation coefficients γ_k^(r) multiplying basis coefficients. In the RKHS version, the mechanism is the reproducing kernel and operator-valued kernels over tensor products. This operator machinery unifies what are usually separate modules—feature extraction, regression, and symbolic inference—into one functional-analysis framework.

What would settle it

Train a fixed embedding, learn a linear operator T_r for each relation on a set of pairs, and compose operators for multi-step transitivity; if held-out composition accuracy on a standard analogy or relational benchmark is at chance, the claim that reasoning reduces to linear operator composition fails. A sharper test would use relations that require XOR-like structure, since no single bounded linear map on a fixed embedding can represent them.

Watch

Extended reading notes

Core claim

The paper's central claim is that learning, reasoning, and inference can be unified in a single operator-theoretic pipeline on a Hilbert space: inputs are embedded via a feature map, decomposed in a spectral basis, and processed by operators estimated through regularized least squares. Reasoning is treated as operator manipulation: a relation R is represented by an operator T_r with T_r f_A ≈ f_B, transitivity by composition T_{r_2} T_{r_1} f_A ≈ f_C, and analogy by spectral vector arithmetic such as f_king − f_man + f_woman. In the RKHS setting this becomes operator-valued kernels over relational tuples, and in the spectral setting it becomes relation-specific modulation coefficients applied to basis coefficients. The paper argues that because these operators are linear or compact, the resulting models are interpretable via their spectra and eigendecompositions, stable under perturbation, and expressive enough to match or exceed neural performance in domains such as texture classification, speech recognition, dynamical-system forecasting, and vision-language captioning.

Load-bearing premise

The reasoning half of the framework rests on the assumption that logical relations such as implication, transitivity, and analogy are faithfully captured by bounded linear operators or spectral modulations on the same embedding space used for perception.

Editorial extensions

If this is right

  • A single embedding and operator family can serve both perception and reasoning, so symbolic inference no longer needs a separate rule-based layer.
  • Models such as scattering transforms and Koopman approximations can work without end-to-end backpropagation, lowering training cost and improving reproducibility.
  • Spectral and eigen-decompositions give a direct window into what a model has learned, making verification feasible in safety-critical applications.
  • In low-data and resource-constrained settings, Hilbert-space models could offer competitive accuracy with far fewer parameters than neural baselines.
  • Operator composition gives a natural way to chain relational knowledge, potentially enabling compositional generalization from few training pairs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the paper imports its benchmark numbers from earlier work rather than running one end-to-end experiment in its own pipeline, a direct test of embed → learn relation operators → compose → evaluate on a relational reasoning benchmark would be the cleanest way to see whether the reasoning half is doing real work.
  • The linear-representability assumption could be probed by adversarial relations: if XOR-like or disjunctive relations require nonlinear maps or a different embedding, then the unified claim holds only for a restricted class of relations.
  • A natural extension is a hybrid architecture where a neural network learns observables or dictionaries in a first stage and spectral reasoning operators act on those embeddings in a second stage; this would keep interpretability while relaxing the linearity constraint.
  • If spectral modulation works for analogy, the same mechanism could transfer to knowledge-base completion and multi-hop question answering, where transitivity is the core operation; these tasks are more demanding than the benchmarks reported here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript, arXiv:2507.21189, proposes a 'Hilbert space framework' for machine learning in which feature extraction, learning, reasoning, and inference are all expressed as operations in an infinite-dimensional Hilbert space. It reviews known material: RKHS theory, orthogonal projection, Parseval expansions, Fourier and wavelet transforms, scattering networks, Koopman operators, and spectral filtering. It then introduces a reasoning operator section (Section 8) that represents logical relations as linear operators T_r on embeddings, with transitive inference via composition T_{r2} T_{r1} f_A ≈ f_C. The paper claims in Section 2 that its contribution is a unified operator-theoretic pipeline, and the Conclusion asserts that Hilbert space methods 'can match or exceed the performance of neural models in several key domains.' The experimental section (Section 7) aggregates results from external papers, including the authors' own preprint [9], rather than reporting new experiments from the proposed pipeline.

Significance. The paper has some strengths: the exposition of Hilbert space basics, RKHS, and spectral filtering is mostly correct and is written in a clear, pedagogical style. It also correctly credits well-established results in scattering theory and Koopman analysis and offers a useful survey of those methods. However, the central research claim is not established. The proposed 'unified pipeline' is never instantiated or tested, and the reasoning operator in Section 8 is defined only as per-pair regularized least-squares regression, with the crucial compositionality property simply asserted. The empirical results in Section 7 are not original experiments; they are literature values with no protocol, baselines, error bars, or reproducibility details, and Table 1 even marks one entry as '(inferred)'. If the paper were repositioned as a survey, the lack of new experiments would be less problematic, but the title and the contribution statement in Section 2 make stronger research claims that the manuscript does not support.

major comments (3)
  1. [Section 7 and Table 1] The paper claims in Section 2 to 'demonstrate empirical competitiveness' of the proposed Hilbert space pipeline, but Section 7 reports no original experiments; the performance numbers are imported from external publications, including the authors' own preprint [9], with no dataset details, train/test splits, error bars, or baseline implementations. Table 1 explicitly marks VQAv2 as '(inferred)', which is not a measurement. Therefore the Conclusion's claim that Hilbert space methods 'can match or exceed the performance of neural models in several key domains' is unsupported by this manuscript.
  2. [Section 8, 'Learning to Reason in Hilbert Space'] The reasoning operator T_r is fitted per relation by minimizing the regularized least-squares objective over paired examples (A_i, B_i), and transitive inference is then asserted as T_{r2} T_{r1} f_A ≈ f_C. The objective contains no term involving (A, C) pairs, no consistency or commutation condition across relations, and no generalization bound. Nothing in the derivation ensures that the composed operator maps a novel A embedding to the correct C embedding. The paper provides no experiment, toy example, or theorem demonstrating compositional generalization. Section 8.1 lists limitations of Hilbert space models in general but does not address this gap, which is load-bearing for the paper's central claim of a unified reasoning pipeline.
  3. [Sections 2 and 6.3] The claimed novel components (learnable spectral modulations, differentiable soft-thresholding, reasoning operators) are described at a conceptual level without a concrete algorithmic specification, a precise objective, or an implementation. The paper does not distinguish these components from existing spectral operator methods such as Fourier Neural Operators [7] or wavelet-domain transformers [8], and it does not define how the 'unified pipeline' connects feature extraction, learning, and reasoning in a way that could be experimentally evaluated. As a result, the novelty of the unified operator-theoretic framework is asserted rather than demonstrated.
minor comments (6)
  1. [Section 6.2] The scattering transform definition is written as S[f] = {‖f∗ψ_{j1}‖, ‖|f∗ψ_{j1}|∗ψ_{j2}‖, ...}, which is not the standard Mallat scattering transform; the standard definition averages the wavelet modulus coefficients with a lowpass filter φ, not with the norm. Please correct the formula and the notation.
  2. [Section 4] The regularization term in the operator estimation objective is denoted ∥T∥_S, but the norm S is not defined in that paragraph; the subsequent text mentions the Hilbert-Schmidt norm. Use a consistent notation such as ∥T∥_HS throughout.
  3. [Section 3] For complex Hilbert spaces, the symmetry property of the inner product should read conjugate symmetry, and the linearity convention should be stated explicitly; the current list treats the field as if it were real in the symmetry condition.
  4. [Section 7.2] The statement that Koopman-based models 'outperform standard RNN baselines' on the Lorenz and Duffing systems is not supported by any table, figure, or quantitative metric in this manuscript; either add the comparison or qualify the claim as a summary of prior work.
  5. [Section 7.3 and Table 1] The entry 'VQAv2 Visual QA (inferred)' is unexplained; specify whether this result is taken from reference [9], estimated from related results, or measured, and avoid using inferred values in a table that presents experimental results.
  6. [Section 8, 'Reasoning with Kernelized Representations'] The composition formula K_{R2∘R1}(x,z) = ∫ K_{R1}(x,y) K_{R2}(y,z) dμ(y) is stated without conditions on μ or on the kernel operators; this identity is not valid for arbitrary kernels and should be derived or attributed to a specific framework.

Circularity Check

2 steps flagged · score 5.0 of 10

The unified framework's empirical support is carried by the authors' own preprint, and Section 8's reasoning operator is defined by the same paired-example residual it is then said to satisfy.

  1. self definitional [Section 8, 'Functional Composition as Reasoning' and 'Learning to Reason in Hilbert Space']
    "Given two concepts or entities fA,fB∈H , we define a reasoning operator T :H→H such that TfA≈fB, where T represents the effect of a logical or relational transition... We then aim to learn or define T such that the residual ∥TfA−fB∥H is minimized over a training set of logical pairs {(Ai,Bi)}... We learn a parameterized family of operators {T (r)θ}⊆B(H) such that minθ Σi ∥T (Ri)θ fxi−fyi∥2H + λR(θ)."

    The reasoning operator is selected by minimizing exactly the residual ∥T f_A − f_B∥ that is then presented as the relation being learned. The subsequent claim that 'transitive inference corresponds to Tr2Tr1fA≈fC' is not derived: fitting T_{r1} on (A,B) pairs and T_{r2} on (B,C) pairs imposes no constraint on the composed map applied to unseen (A,C) pairs, and no theorem or experiment is provided. Thus the central assertion that reasoning is implemented via operator composition reduces, in the paper's own formulation, to a training-objective fit rather than an independently validated prediction.

  2. self citation load bearing [Section 7.3, Table 1, and Section 9 Conclusion]
    "Spectral Dictionary Vision-Language Models (SDict-VLM) [9] present a recent advance in compositional learning via learned frequency atoms... On the MS-COCO captioning benchmark, SDict-VLM achieves BLEU-4 = 39.2, CIDEr = 127.5, and SPICE = 27.0, closing over 85% of the performance gap to BLIP-2... These results demonstrate that spectral learning in Hilbert spaces not only competes with modern transformer-based architectures..."

    Reference [9] is the authors' own arXiv preprint, 'From Pixels and Words to Waves: A Unified Framework for Spectral Dictionary vLLMs' (Kiruluta and Burity, 2025). The only empirical support offered for the paper's claim that Hilbert-space-based methods 'can match or exceed the performance of neural models' comes from this self-citation; the numbers are imported rather than produced by the pipeline described here. Table 1 even labels VQAv2 as '(inferred)'. The conclusion repeats the borrowed result as if it were a demonstration of the present framework, making the empirical pillar of the central claim load-bearing on the authors' own prior work.

full rationale

Most of the paper's mathematical content is standard, externally established material (RKHS, scattering transforms, Koopman operators, Fourier/wavelet analysis), and the paper appropriately cites the original sources for those results. No circularity arises from the Hilbert-space definitions themselves. However, two load-bearing moves deserve weight. First, Section 8 defines the reasoning operator by minimizing the paired residual it is then asserted to satisfy, and the compositional transitivity claim (T_{r2} T_{r1} f_A ≈ f_C) is stated without any constraint, theorem, or experiment connecting the composed fitted operators to unseen triples; in the paper's own equations, 'reasoning' and 'fitting' are the same operation. Second, the empirical competitiveness claim in the abstract, Section 7, and Conclusion rests on benchmark numbers imported from the authors' own preprint [9], with no independent evaluation in this paper. These are genuine circularity-adjacent defects: the first is a fit presented as a reasoning capability, and the second is a self-citation carrying the main empirical demonstration. The central framework still has independent conceptual content, so the score is moderate rather than extreme.

Assumptions & free parameters 4 free parameters · 7 assumptions · 1 invented entities

This ledger shows that the paper introduces no new fitted constants beyond illustrative learnable parameters, but it rests on multiple domain assumptions: operator representability of learning, linear representability of logic, and the validity of imported benchmark numbers. The reasoning operator has no independent falsifiable handle, so it appears as an invented entity with no independent evidence.

free parameters (4)
  • Learnable soft-threshold parameters theta_k = not fitted in this paper
    Introduced in Section 6.3 as sigma_theta(z) = z/(z + theta_k); a learnable per-frequency parameter for the proposed spectral models, with no values or training details.
  • Spectral multipliers gamma_k = not fitted in this paper
    Used to define spectral filtering in Section 6.1 and reasoning modulations in Section 8; the paper gives no values or training procedure.
  • Regularization coefficient lambda = not specified
    Appears in operator regression objectives in Sections 4 and 8; no selection rule is provided.
  • SDict-VLM learnable spectral atoms phi_i = learned in [9], not detailed here
    The central empirical model in Section 7.3 is based on learnable basis functions; no training or protocol details are given in this paper.
assumptions (7)
  • standard math Hilbert space axiomatics: inner product, completeness, orthonormal bases, Parseval identity, Riesz representation.
    Foundational material in Sections 3 through 6; accepted mathematical background.
  • standard math Representer theorem and RKHS properties.
    Section 5; standard result, invoked for kernel regression and for the kernelized reasoning discussion.
  • standard math Restricted Isometry Property for compressed sensing recovery.
    Section 6; cited from [19,20]; assumed for sparse reconstruction claims.
  • domain assumption Learning tasks can be represented as estimation of a bounded or compact operator T: H_X -> H_Y.
    Section 4; if real-world tasks are not reducible to such operator estimation, the framework's scope is empty.
  • domain assumption Logical relations can be represented by linear operators or spectral modulations on embedding coefficients.
    Section 8: T f_A approximately equals f_B and reasoning via gamma_k. No proof or evidence is given that arbitrary implication or conjunction is linearly representable.
  • domain assumption Word analogies behave as vector arithmetic in spectral coefficient space.
    Section 8 'king - man + woman = queen'; this is an approximate empirical property, not a general guarantee.
  • domain assumption The cited benchmark results for scattering, Koopman, and SDict-VLM are valid and comparable to neural baselines.
    Section 7 imports all numbers from references [2,3,4,9,10,12]; no experiment in this paper verifies them.
invented entities (1)
  • Reasoning operator R (relation-specific operators T_r)
    purpose: To perform logical or conceptual inference in Hilbert space via linear maps or spectral modulation; central to Section 8.
    No theorem, no dataset, and no falsifiable prediction. The operator is defined as any map that minimizes ||T f_A - f_B||^2, so success is guaranteed by fitting unless the map is constrained and tested independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Operator-Based Machine Intelligence: A Hilbert Space Framework for Spectral Learning and Symbolic Reasoning." pith.science (2026). https://pith.science/paper/UYWJ4OQD

@misc{pith2026250721189,
  author       = {Pith},
  title        = {Pith review of: Operator-Based Machine Intelligence: A Hilbert Space Framework for Spectral Learning and Symbolic Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UYWJ4OQD}},
  note         = {Machine review of arXiv:2507.21189}
}
read the original abstract

Traditional machine learning models, particularly neural networks, are rooted in finite-dimensional parameter spaces and nonlinear function approximations. This report explores an alternative formulation where learning tasks are expressed as sampling and computation in infinite dimensional Hilbert spaces, leveraging tools from functional analysis, signal processing, and spectral theory. We review foundational concepts such as Reproducing Kernel Hilbert Spaces (RKHS), spectral operator learning, and wavelet-domain representations. We present a rigorous mathematical formulation of learning in Hilbert spaces, highlight recent models based on scattering transforms and Koopman operators, and discuss advantages and limitations relative to conventional neural architectures. The report concludes by outlining directions for scalable and interpretable machine learning grounded in Hilbertian signal processing.

Figures

Figures reproduced from arXiv: 2507.21189 by the authors.

Figure 1
Figure 1. Hilbert space learning architecture with embedded reasoning. Inputs are mapped to [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages

  1. [9]

    From Pixels and Words to Waves: A Unified Framework for Spectral Dictionary vLLMs

    A. Kiruluta and P. Burity, “From Pixels and Words to Waves: A Unified Framework for Spectral Dictionary vLLMs,” arXiv preprint arXiv:2506.18943 , 2025

  2. [7]

    Fourier Neu- ral Operator for Parametric Partial Differential Equations,

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, A. Stuart, and A. Anandkumar, “Fourier Neu- ral Operator for Parametric Partial Differential Equations,” Advances in Neural Information Processing Systems (NeurIPS), 2020

  3. [8]

    Misspelling Semantics In Thai

    P. Fakhari, H. Aghdam, and A. Montanari, “Wavelet Vision Transformer,” arXiv preprint arXiv:2206.09680, 2022

  4. [1]

    Mallat, A Wavelet Tour of Signal Processing , Academic Press, 1999

    S. Mallat, A Wavelet Tour of Signal Processing , Academic Press, 1999

  5. [2]

    Group Invariant Scattering,

    S. Mallat, “Group Invariant Scattering,” Communications on Pure and Applied Mathematics , vol. 65, no. 10, pp. 1331–1398, 2012

  6. [3]

    Invariant Scattering Convolution Networks,

    J. Bruna and S. Mallat, “Invariant Scattering Convolution Networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 35, no. 8, pp. 1872–1886, 2013

  7. [4]

    Deep Scattering Spectrum,

    J. And´ en and S. Mallat, “Deep Scattering Spectrum,”IEEE Transactions on Signal Processing, vol. 62, no. 16, pp. 4114–4128, 2014

  8. [5]

    Spectral Networks and Locally Connected Networks on Graphs,

    J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun, “Spectral Networks and Locally Connected Networks on Graphs,” International Conference on Learning Representations (ICLR) , 2014

Show all 24 references
  1. [6]

    Deeper Insights into Graph Convolutional Networks for Semi- Supervised Learning,

    Q. Li, Z. Han, and X. Wu, “Deeper Insights into Graph Convolutional Networks for Semi- Supervised Learning,” AAAI Conference on Artificial Intelligence , 2018

  2. [10]

    A Data–Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition,

    M. O. Williams, I. G. Kevrekidis, and C. W. Rowley, “A Data–Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition,” Journal of Nonlinear Science, vol. 25, pp. 1307–1346, 2015

  3. [11]

    J. N. Kutz, S. L. Brunton, B. W. Brunton, and J. L. Proctor, Dynamic Mode Decomposition: Data-Driven Modeling of Complex Systems , SIAM, 2016

  4. [12]

    Deep Learning for Universal Linear Embeddings of Nonlinear Dynamics,

    B. Lusch, J. N. Kutz, and S. L. Brunton, “Deep Learning for Universal Linear Embeddings of Nonlinear Dynamics,” Nature Communications, vol. 9, no. 1, pp. 1–10, 2018

  5. [13]

    Gradient-Based Learning Applied to Doc- ument Recognition,

    Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-Based Learning Applied to Doc- ument Recognition,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998. 17

  6. [14]

    Attention Is All You Need,

    A. Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems (NeurIPS), 2017

  7. [15]

    Reconciling Modern Machine-Learning Practice and the Classical Bias–Variance Trade-Off,

    M. Belkin, D. Hsu, S. Ma, and S. Mandal, “Reconciling Modern Machine-Learning Practice and the Classical Bias–Variance Trade-Off,” Proceedings of the National Academy of Sciences, vol. 116, no. 32, pp. 15849–15854, 2019

  8. [16]

    Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead,

    C. Rudin, “Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead,” Nature Machine Intelligence, vol. 1, pp. 206–215, 2019

  9. [17]

    Beyond Bandlimited Sampling: Nonlinear and Nonideal Sam- pling,

    Y. C. Eldar and T. Michaeli, “Beyond Bandlimited Sampling: Nonlinear and Nonideal Sam- pling,” IEEE Signal Processing Magazine , vol. 26, no. 3, pp. 48–57, 2009

  10. [18]

    Sampling—50 Years After Shannon,

    M. Unser, “Sampling—50 Years After Shannon,” Proceedings of the IEEE, vol. 88, no. 4, pp. 569–587, 2000

  11. [19]

    Robust Uncertainty Principles: Exact Signal Recon- struction from Highly Incomplete Frequency Information,

    E. J. Cand` es, J. Romberg, and T. Tao, “Robust Uncertainty Principles: Exact Signal Recon- struction from Highly Incomplete Frequency Information,” IEEE Transactions on Information Theory, vol. 52, no. 2, pp. 489–509, 2006

  12. [20]

    Compressed Sensing,

    D. L. Donoho, “Compressed Sensing,” IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1289–1306, 2006

  13. [21]

    Theory of Reproducing Kernels,

    N. Aronszajn, “Theory of Reproducing Kernels,” Transactions of the American Mathematical Society, vol. 68, no. 3, pp. 337–404, 1950

  14. [22]

    Sch¨ olkopf and A

    B. Sch¨ olkopf and A. J. Smola, Learning with Kernels: Support Vector Machines, Regulariza- tion, Optimization, and Beyond , MIT Press, 2001

  15. [23]

    Diffusion Maps,

    R. Coifman and S. Lafon, “Diffusion Maps,” Applied and Computational Harmonic Analysis , vol. 21, no. 1, pp. 5–30, 2006

  16. [24]

    Discovering Governing Equations from Data by Sparse Identification of Nonlinear Dynamical Systems,

    S. L. Brunton, B. W. Brunton, J. L. Proctor, and J. N. Kutz, “Discovering Governing Equations from Data by Sparse Identification of Nonlinear Dynamical Systems,” PNAS, vol. 113, no. 15, pp. 3932–3937, 2016. 18

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.