Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Explaining Deep Network Classification of Matrices: A Case Study on Monotonicity

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A neural network's monotonicity verdict for 7×7 matrices reduces to a single ratio, |c0/c1|, which equals 1/tr(A^{-1}) for monotone matrices and stayed below 0.1755 in 18,000 samples.

desk verdict Nice XAI pipeline, but the headline bound is not a true necessary condition: an explicit one-parameter family of monotone 7x7 matrices in the cube has |c0/c1| arbitrarily large, so the EVT finite-endpoint claim is false. read the letter →

arxiv 2507.22570 v1 pith:EE3TOCUP submitted 2025-07-30 cs.LG cs.AIcs.NAmath.NA

classification cs.LGcs.AIcs.NAmath.NA MSC 15B4868T0715A1862G32
keywords monotonematricesinversenonnegativeexplainableAIintegratedgradientscharacteristicpolynomialcoefficientsextremevaluetheorygeneralizedParetodistributionmatrixclassification
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

Monotone matrices—those whose inverses are entrywise nonnegative—are easy to define but resist direct characterization from their entries. This paper tries to establish that a deep-network classifier can be explained, and that for 7×7 matrices drawn uniformly from (-1,1) the monotonicity signal condenses to the ratio |c0/c1| of the two lowest-order characteristic-polynomial coefficients. Since c0=(-1)^n det(A), c1=(-1)^{n-1}tr(adj(A)), and adj(A)=det(A)$A^{{-1}}$, the ratio equals 1/tr($A^{{-1}}$) for monotone matrices. The ratio alone classifies monotone versus non-monotone with 93.5% accuracy, and the two coefficients together reach 95.1%. In 18,000 monotone samples the ratio never exceeded 0.1755, and a generalized-Pareto tail fit puts the exceedance probability near 1.8×$10^{-4}$ (bootstrap interval up to 3.1×$10^{-4}$), supporting a proposed necessary condition tr($A^{{-1}}$)≥5.7. The paper presents this as a statistically derived bound, with the finite upper endpoint left as a conjecture.

What carries the argument

The load-bearing identity is |c0/c1| = 1/|tr($A^{{-1}}$)|, obtained from c0=(-1)^n det(A), c1=(-1)^{n-1}tr(adj(A)), and adj(A)=det(A)$A^{{-1}}$; for monotone matrices the inverse is entrywise nonnegative, so the trace is positive and the ratio is exactly 1/tr($A^{{-1}}$). This identity converts the algebraic condition $A^{{-1}}$≥0 into a one-dimensional diagnostic. Around it the paper assembles an interpretability pipeline—integrated gradients (a saliency method that integrates model gradients along a path from a baseline to the input), decision-tree surrogates, symbolic regression, and linear SVMs—to show the network relies on these coefficients, and a Peaks-Over-Threshold generalized Pareto model with threshold u=0.075 to estimate how unlikely it is that a monotone matrix exceeds the observed maximum.

What would settle it

Sample a far larger collection of random 7×7 monotone matrices and count the fraction with |c0/c1|>0.1755; if that fraction is not at or below roughly $10^{-4}$, the proposed necessary condition collapses. Alternatively, explicitly search the cube for a monotone A with tr($A^{{-1}}$)<5.7, which would disprove a literal reading of the bound.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the learned boundary for monotonicity is nearly one-dimensional. Integrated-gradient feature attribution, decision-tree surrogates, symbolic regression, and linear support-vector machines all converge on the magnitudes |c0| and |c1| of the lowest-order characteristic-polynomial coefficients; a two-feature network achieves 95.1% accuracy and a one-feature network using |c0/c1| achieves 93.5%, with validation curves that track or beat training curves, a sign the signal is robust rather than memorized. Using adj(A)=det(A)$A^{{-1}}$, the authors derive |c0/c1| = 1/tr($A^{{-1}}$) for monotone A, so large inverse traces correspond to small ratios. Among 18,000 monotone 7×7 matrices the largest observed ratio is T7=0.1755; fitting a generalized Pareto distribution to excesses above u=0.075 gives shape ξ=-0.028±0.006, indicating a finite tail, with exceedance probability estimated near 1.8×$10^{-4}$. They therefore propose the necessary condition |c0/c1|≤0.1755, equivalently tr($A^{{-1}}$)≥5.7, for monotone matrices in this sampling regime.

Load-bearing premise

The bound depends on treating 18,000 sampled monotone matrices—and a tail model whose threshold was chosen from those same samples—as representative of all monotone 7×7 matrices; the finite upper endpoint is explicitly conjectured, not proved.

Editorial extensions

If this is right

  • If the bound holds, monotonicity for uniform 7×7 matrices becomes a one-scalar test: |c0/c1|>0.1755, or tr(A^{-1})<5.7, marks a matrix non-monotone with probability above 99.98% by the paper's estimate.
  • The compact two-feature and one-feature networks generalize better than the 73-feature model, suggesting that the discovered ratio is a genuine property of the class rather than an artifact of overfitting.
  • Because c0 and c1 are similarity invariants, the non-monotonicity conclusion transfers to entire similarity classes of matrices even though monotonicity itself is not similarity-invariant.
  • In the ambiguous low-ratio region |c0/c1|<0.08, a subdomain-trained model reaches about 95.8% accuracy and shifts its attention to |c1|, the trace, and raw entries, enabling a hierarchical two-stage classifier.
  • The same finite-endpoint argument is size-dependent: observed maxima are T5≈0.5968 and T10≈0.104, but the generalized Pareto fits for n=5 and n=10 give positive shape parameters, so the n=7 bound may not generalize to other sizes.

Reading between the lines

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

  • The proposed threshold should not be read as a universal inequality: monotone matrices arbitrarily close to singular ones can have tr(A^{-1}) near zero, so the bound can fail on measure-zero configurations within the cube even if it holds for generic samples.
  • Under structured priors—such as M-matrices, Stieltjes matrices, or finite-element stiffness matrices—the threshold and even the dominant features could shift; the paper's own outlook asks whether the same low-order invariants survive outside the uniform distribution.
  • A natural testable extension is to run the same saliency-plus-extreme-value pipeline on other algebraic predicates (e.g., total positivity, Hurwitz stability) to see whether compact ratio criteria emerge there too.
  • The positive shape parameters for n=5 and n=10 suggest the n=7 finite-endpoint conclusion may be an accident of sample size; larger samples for those sizes would reveal whether heavy tails reappear.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents a deep-learning/XAI pipeline for distinguishing monotone from non-monotone matrices sampled uniformly from (-1,1)^{n×n}. Using integrated gradients, decision trees, and symbolic regression on a 73-feature model, the authors identify the characteristic-polynomial coefficient magnitudes |c0| and |c1| as the dominant features, and show that a two-feature network and a one-feature network using |c0/c1| achieve 95.1% and 93.5% accuracy respectively on their 7×7 dataset. The paper then derives the algebraic identity |c0/c1| = 1/tr(A^{-1}) for monotone A, and, on the basis of a sample maximum T7 = 0.1755 and a generalized Pareto tail fit, proposes the statistical necessary condition |c0/c1| ≤ 0.1755, equivalently tr(A^{-1}) ≥ 5.7, for monotone 7×7 matrices. The abstract states this bound holds with probability >99.98%.

Significance. The methodological idea of using saliency and surrogate models to compress a matrix classifier into an interpretable algebraic condition is timely and potentially useful, and the classification experiments are internally consistent: the two-feature model's 95.1% accuracy and the ratio-only model's 93.5% accuracy are plausible and well reported. The identity relating |c0/c1| to 1/tr(A^{-1}) is correct and is a nice observation. However, the paper's headline quantitative claim—the 99.98% probability bound and the derived lower bound on tr(A^{-1})—is not supported: the EVT finite-endpoint analysis is contradicted by an explicit family of monotone matrices with unbounded ratio, and the threshold T7 is the sample maximum, making the 'no violation' statement in-sample by construction. The ML/XAI findings may survive a careful reframing, but the statistical claim needs independent validation before it can be presented as a necessary condition.

major comments (4)
  1. [§6.2, Eq. (6), Remark 6.1] The EVT finite-endpoint analysis is invalid. Let P be the 7×7 cyclic-shift permutation and define A_ε = (εI+P)^{-1} for ε>0. Then A_ε^{-1} = εI+P is entrywise nonnegative, so A_ε is monotone, and the geometric series expansion shows every entry of A_ε lies in (-1,1). Identity (5) gives |c0/c1| = 1/(7ε), which exceeds 0.1755 for every ε < 1/(7×0.1755) ≈ 0.814. Since A_ε^{-1} is strictly positive, sufficiently small entrywise perturbations of A_ε remain monotone and inside the cube, so the event {|c0/c1| > 0.1755} has positive probability under the uniform distribution. Consequently the fitted GPD with ξ < 0 and finite endpoint E7 ≈ 0.828 cannot describe the true tail of the ratio, and the abstract's 'probability >99.98%' cannot be justified by the EVT model as presented.
  2. [§6.1 and §6.2] The threshold T7 = 0.1755 is defined as the maximum observed ratio among the monotone training samples, so the statement that 'no monotone sample in the dataset violated it' is true by construction. The exceedance probability 1.8×10^{-4} is estimated from a generalized Pareto distribution fitted to the same data that produced the threshold, which is an in-sample procedure. To support the claimed population probability, the authors need an independent held-out or freshly generated monotone sample, or a nonparametric confidence bound, rather than a tail extrapolation from the same dataset.
  3. [Abstract and §6.2] The proposed condition is labeled a 'necessary condition,' but the explicit family A_ε above shows that the inequality |c0/c1| ≤ 0.1755 is not a mathematical necessary condition for monotonicity on (-1,1)^{7×7}; it is at most an empirical statement about the bulk of the distribution. The wording 'with high statistical probability' in Eq. (6) and the phrase 'the monotone class obeys ... with probability >99.98%' in the abstract should be revised to clearly distinguish an empirical tail estimate from a proven bound, and the equivalence with tr(A^{-1}) ≥ 5.7 should be presented as probabilistic rather than as a deterministic bound.
  4. [§6.2] Remark 6.1 already concedes that the finite endpoint is conjectured, but the preceding text calls the negative shape parameter 'strong statistical evidence' for a finite upper endpoint. These two statements are in tension, and the counterexample in the first major comment shows that the finite-endpoint conjecture is false. The authors should either remove the EVT-based justification or replace it with a direct estimate of the exceedance probability over the open neighborhoods of violating matrices.
minor comments (4)
  1. [§6.2] The bootstrap confidence interval is reported as [6.8×10^{-5}, 3.1×10^{-5}], whose lower endpoint exceeds its upper endpoint; the footnote gives [6.8×10^{-5}, 3.1×10^{-4}]. The inconsistency should be fixed.
  2. [§4.2] The phrase 'sufficient statistics' is used loosely; |c0| and |c1| are the features selected by the surrogate and saliency models, not sufficient statistics in the statistical sense. A more cautious wording would avoid implying a formal reduction.
  3. [§8] The concluding summary repeats the 'bounded random variable' and 'vanishing probability' claims without the caveats introduced in Remark 6.1, which overstates the strength of the EVT evidence.
  4. [Data and code availability] The data and training scripts are only 'available upon request'; for a data-driven paper whose central claim depends on a specific 18,000-sample dataset, depositing the data and code in a permanent repository would materially improve reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

The central statistical bound reduces to an in-sample fit: the threshold is the observed sample maximum, and the >99.98% probability is a GPD estimate from the same data, not an independent prediction.

  1. self definitional [Section 6.1, paragraph defining the Sample Maximum Ratio; Eq. (6) in Section 6.2]
    "The purpose of obtaining a potential necessary condition for monotonicity based on this dataset, made us search the maximum observed value of the ratio |c0/c1| among only the monotone matrices in our dataset. We term this value the Sample Maximum Ratio, denoted simply as Tn for each matrix size n. ... These Sample Maximum Ratios, T5 = 0.5968, T7 = 0.1755, and T10 = 0.104, define statistically derived necessary conditions based on our data: any matrix A in the respective dataset found to be monotone satisfied |c0/c1| ≤ Tn."

    The threshold in the proposed condition (6) is defined as the maximum of the training sample, so the statement that all observed monotone matrices satisfy it is true by construction. The paper then uses this tautology as the empirical basis for the abstract's claim that the monotone class obeys |c0/c1| ≤ 0.18 with probability >99.98%. A sample maximum cannot by itself establish a population probability, and the bound is not independent of the data used to define it.

  2. fitted input called prediction [Section 6.2, GPD fit and exceedance estimate]
    "Fitting a generalized Pareto distribution to the Nu = 562 excesses above this threshold yields ξ̂ = −0.028 ± 0.006, σ̂ = 0.021 ± 0.003 (MLE ± s.e.). ... We can use the fitted GPD model to estimate the probability of exceeding the maximum ratio observed in the data, T7 ≈ 0.1755. The model yields P(r > 0.1754 | monotone, n=7) ≈ 1.8 × 10−4."

    The GPD is fit to the same tail excesses of the monotone training data, and the 'predicted' quantity P(r > 0.1754) is evaluated at the sample maximum T7 from that same dataset. The small exceedance probability is thus an in-sample property of the fitted model, not an independent out-of-sample prediction. The finite-endpoint conclusion (E7 ≈ 0.828) is an artifact of the same fit, and Remark 6.1 concedes that the endpoint is only conjectured; the abstract's >99.98% probability is therefore not independently supported.

full rationale

The paper's algebraic identity |c0/c1| = 1/tr(A^{-1}) for monotone A is derived correctly from adj(A) = det(A)A^{-1} and is not circular. The ML feature-importance analysis, decision trees, SVMs, and symbolic regression are empirical explorations and are not themselves circular. The circularity is concentrated in the headline statistical claim: the threshold 0.1755 is the sample maximum of the monotone training set, so its satisfaction on that set is definitional, and the estimated exceedance probability comes from a generalized Pareto distribution fitted to the very same tail. This makes the abstract's 'probability >99.98%' an in-sample model output rather than an independent prediction. The paper is transparent in Remark 6.2 that the result is an 'empirically derived statistical bound, not a formally proven theorem,' and in Remark 6.1 that the finite endpoint is conjectured, which mitigates but does not remove the circularity. No load-bearing self-citations or imported uniqueness theorems were found; the cited works by the authors are background on discrete maximum principles and do not support the central statistical claim. Overall, the central claim partially reduces to its own fitting procedure, warranting a score of 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central statistical bound rests on the sample maximum and a fitted GPD tail, so the data-derived threshold and EVT parameters are the main free parameters. The mathematical identities are standard, but the GPD tail approximation and the use of the sample maximum as a bound are assumptions specific to this study.

free parameters (4)
  • T7 sample maximum ratio threshold = 0.1755
    Maximum observed |c0/c1| among 18,000 monotone 7x7 matrices; used directly as the necessary-condition threshold in Eq. (6).
  • EVT threshold u = 0.075
    Chosen from MRL and parameter stability plots in Section 6.2; the fitted GPD and estimated exceedance probability depend on this choice.
  • GPD shape and scale (xi, sigma) = -0.028 +/- 0.006 and 0.021 +/- 0.003
    Fitted by maximum likelihood to 562 excesses above u=0.075; the negative shape drives the finite-endpoint conclusion.
  • Decision tree threshold TDT,7 = 0.08
    Found by fitting a depth-1 CART tree on |c0/c1| to maximize overall accuracy; used to define the filtered subdomain Omega_0 in Section 7.
assumptions (5)
  • standard math c0 = (-1)^n det(A) and c1 = (-1)^{n-1} tr(adj(A))
    Vieta's formulas and the adjugate identity, used throughout Sections 4 and 6.
  • standard math adj(A) = det(A) A^{-1} for invertible A
    Used to derive |c0/c1| = 1/tr(A^{-1}) in Section 6.
  • domain assumption Entries are iid uniform(-1,1); monotone class is the conditional distribution on nonnegative inverse.
    All empirical thresholds and probabilities are with respect to this sampling prior, stated in Section 2.
  • domain assumption Excesses of the ratio over a high threshold follow a generalized Pareto distribution
    EVT analysis in Section 6.2 assumes a GPD tail approximation; the finite-endpoint conclusion follows only if this approximation is valid.
  • ad hoc to paper The sample maximum T7 can serve as a near-upper bound of the distribution
    The necessary condition uses the observed maximum directly; no independence from the data is established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explaining Deep Network Classification of Matrices: A Case Study on Monotonicity." pith.science (2026). https://pith.science/paper/EE3TOCUP

@misc{pith2026250722570,
  author       = {Pith},
  title        = {Pith review of: Explaining Deep Network Classification of Matrices: A Case Study on Monotonicity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EE3TOCUP}},
  note         = {Machine review of arXiv:2507.22570}
}
abstract

This work demonstrates a methodology for using deep learning to discover simple, practical criteria for classifying matrices based on abstract algebraic properties. By combining a high-performance neural network with explainable AI (XAI) techniques, we can distill a model's learned strategy into human-interpretable rules. We apply this approach to the challenging case of monotone matrices, defined by the condition that their inverses are entrywise nonnegative. Despite their simple definition, an easy characterization in terms of the matrix elements or the derived parameters is not known. Here, we present, to the best of our knowledge, the first systematic machine-learning approach for deriving a practical criterion that distinguishes monotone from non-monotone matrices. After establishing a labelled dataset by randomly generated monotone and non-monotone matrices uniformly on $(-1,1)$, we employ deep neural network algorithms for classifying the matrices as monotone or non-monotone, using both their entries and a comprehensive set of matrix features. By saliency methods, such as integrated gradients, we identify among all features, two matrix parameters which alone provide sufficient information for the matrix classification, with $95\%$ accuracy, namely the absolute values of the two lowest-order coefficients, $c_0$ and $c_1$ of the matrix's characteristic polynomial. A data-driven study of 18,000 random $7\times7$ matrices shows that the monotone class obeys $\lvert c_{0}/c_{1}\rvert\le0.18$ with probability $>99.98\%$; because $\lvert c_{0}/c_{1}\rvert = 1/\mathrm{tr}(A^{-1})$ for monotone $A$, this is equivalent to the simple bound $\mathrm{tr}(A^{-1})\ge5.7$.

Figures

Figures reproduced from arXiv: 2507.22570 by the authors.

Figure 1
Figure 1. Evolution of training and validation loss and accuracy over epochs for the 7 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Evolution of training and validation loss and accuracy over epochs for the 7 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Feature-importance ranking based on the mean absolute Integrated Gradients score [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Scatter plots of |c0| versus |c1| for monotone (blue) and non-monotone (red) matrices in the datasets for matrix sizes n = 5, 7, and 10. Note the concentration of monotone matrices near the origin. features might be needed for perfect classification, motivating explora…
Figure 5
Figure 5. Figure 5: Evolution of training and validation loss and accuracy over epochs for the 7 [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Scatter plot of |c0| vs |c1| for n = 7 matrices, showing monotone (blue) and non-monotone (red) points. The dashed line represents the linear separation boundary found by the SVM with a linear kernel (Equation 2). 11 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Evolution of training and validation loss and accuracy over epochs for the 7 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Line plots of frequency counts showing the distribution of the feature ratio [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: EVT diagnostic plots for the ratio |c0/c1| using the n = 7 monotone matrix dataset. (a) The downward trend in the MRL plot for u ≳ 0.1 suggests ξ < 0. (b) The stability plot shows relatively stable negative ξ estimates for u between approximately 0.05 and 0.075, before…
Figure 10
Figure 10. Figure 10: Evolution of training and validation loss and accuracy over epochs for the 7 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Feature importance for the FFN model trained on the filtered subdomain Ω [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Totally Positive Matrices and the Highest-Order Coefficients of the Characteristic Polynomial

    cs.LG 2026-07 conditional novelty 5.0 of 10

    The three highest characteristic-polynomial coefficients separate the tested totally-positive matrix families from non-totally-positive ones with >99.8% accuracy, and in higher dimensions the structured families form ...

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [1]

    Bouchon , Monotonicity of some perturbations of irreducibly diagonally dominant m- matrices, Numer

    F. Bouchon , Monotonicity of some perturbations of irreducibly diagonally dominant m- matrices, Numer. Math., 105 (2007), pp. 591–601

  2. [2]

    J. H. Bramble and B. E. Hubbard , New monotone type approximations for elliptic prob- lems, Math. Comp., 18 (1964), pp. 349–367

  3. [3]

    Brandts, S

    J. Brandts, S. Korotov, and M. K ˇr´ıˇzek, The discrete maximum principle for linear simplicial finite element approximations of a reaction-diffusion problem , Linear Algebra Appl., 429 (2008), pp. 2344–2357

  4. [4]

    Breiman, J

    L. Breiman, J. H. Friedman, R. A. Olshen, and C. J. Stone , Classification and Re- gression Trees, Wadsworth International Group, Belmont, CA, 1984

  5. [5]

    Buzzard, Mathematical reasoning and the computer, Bulletin of the American Mathemat- ical Society, 61 (2024), pp

    K. Buzzard, Mathematical reasoning and the computer, Bulletin of the American Mathemat- ical Society, 61 (2024), pp. 211–224

  6. [6]

    Charton, Linear algebra with transformers , Transactions on Machine Learning Research (TMLR), (2022), https://arxiv.org/abs/2112.01898

    F. Charton, Linear algebra with transformers , Transactions on Machine Learning Research (TMLR), (2022), https://arxiv.org/abs/2112.01898

  7. [7]

    P. G. Ciarlet, Discrete maximum principle for finite-difference operators, Aequationes Math, 4 (1970), pp. 338–352

  8. [8]

    P. G. Ciarlet and P.-A. Raviart , Maximum principle and uniform convergence for the finite element method , Comput. Methods Appl. Mech. Eng., 2 (1973), pp. 17–31

Show all 22 references
  1. [9]

    Coles, An Introduction to Statistical Modeling of Extreme Values, Springer Series in Statis- tics, Springer, London, 2001, https://doi.org/10.1007/978-1-4471-3675-0

    S. Coles, An Introduction to Statistical Modeling of Extreme Values, Springer Series in Statis- tics, Springer, London, 2001, https://doi.org/10.1007/978-1-4471-3675-0

  2. [10]

    Collatz, Aufgaben monotone art , Arch

    L. Collatz, Aufgaben monotone art , Arch. Math., 3 (1952), pp. 366–376

  3. [11]

    Cortes and V

    C. Cortes and V. N. V apnik , Support-vector networks , Machine Learning, 20 (1995), pp. 273–297

  4. [12]

    Cranmer, A

    M. Cranmer, A. Sanchez-Gonzalez, P. Battaglia, R. Xu, K. Cranmer, D. Spergel, and S. Ho , Discovering symbolic models from deep learning with inductive biases , NeurIPS, (2020), https://github.com/MilesCranmer/PySR. Introduces the PySR symbolic-regression package

  5. [13]

    Davies, P

    A. Davies, P. Veli ˇckovi´c, L. Buesing, S. Blackwell, D. Zheng, N. Toma ˇsev, R. Tanburn, P. Battaglia, C. Blundell, A. Juh´asz, M. Lackenby, G. Williamson, D. Hassabis, and P. Kohli , Advancing mathematics by guiding human intuition with AI , Nature, 600 (2021), pp. 70–74, h...

  6. [14]

    Dwivedi, D

    R. Dwivedi, D. Dave, H. Naik, S. Singhal, O. Rana, P. Patel, B. Qian, Z. Wen, T. Shah, G. Morgan, and R. Ranjan, Explainable AI (XAI): core ideas, techniques and so- lutions, ACM Computing Surveys, 55 (2023), pp. 1–36, https://doi.org/10.1145/3561048

  7. [15]

    Ghrist, J

    R. Ghrist, J. Gould, and M. Lopez , Lattice-valued bottleneck duality , (2024), https: //arxiv.org/abs/2410.00315

  8. [16]

    Kar´atson and S

    J. Kar´atson and S. Korotov, Discrete maximum principles for finite element solutions of nonlinear elliptic problems with mixed boundary conditions, Numer. Math., 99 (2005), pp. 669– 698

  9. [17]

    Lorenz, Zur inversmonotonie diskreter probleme , Numer

    J. Lorenz, Zur inversmonotonie diskreter probleme , Numer. Math., 27 (1977), pp. 227–238

  10. [18]

    O. L. Mangasarian , Characterization of real matrices of monotone kind , SIAM Review, 10 (1968), pp. 439–441

  11. [19]

    Novikov, N

    A. Novikov, N. V ˜u, M. Eisenberger, E. Dupont, P. Huang, A. Z. W agner, S. Shirobokov, B. Kozlovskii, F. J. R. Ruiz, A. Mehrabian, M. P. Kumar, A. See, S. Chaudhuri, G. Holland, A. Davies, S. Nowozin, P. Kohli, and M. Balog, AlphaE- volve: A coding agent for scientific and al...

  12. [20]

    Sundararajan, A

    M. Sundararajan, A. Taly, and Q. Yan , Axiomatic Attribution for Deep Networks , in Proceedings of the 34th International Conference on Machine Learning, vol. 70 of Proceed- ings of Machine Learning Research, PMLR, 2017, pp. 3319–3328, http://proceedings.mlr. press/v70/sundara...

  13. [21]

    V arga, Matrix Iterative Analysis , Prentice-Hall, New Jersey, 1962

    R. V arga, Matrix Iterative Analysis , Prentice-Hall, New Jersey, 1962

  14. [22]

    Williamson, Is deep learning a useful tool for the pure mathematician? , (2023), https: //arxiv.org/abs/2304.12602

    G. Williamson, Is deep learning a useful tool for the pure mathematician? , (2023), https: //arxiv.org/abs/2304.12602. 22

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.