polyDAG: Polynomial Acyclicity Constraints for Efficient Continuous Causal Discovery in Visual Semantic Graphs

Kai Hwang; Minyi Guo; Ramin Ramezani; Tao Han; Wenhao Zhang

REVIEW 2 major objections 6 minor 52 references

polyDAG: Polynomial Acyclicity Constraints for Efficient Continuous Causal Discovery in Visual Semantic Graphs

T0 review · 2 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A finite polynomial trace sum can replace the matrix-exponential acyclicity constraint in continuous DAG learning.

desk verdict Correct math, but the experiments run a weaker constraint than the one proved, so take the speedup and recovery claims with salt until the truncation is reconciled. read the letter →

arxiv 2606.06908 v2 pith:FMVX4PS2 submitted 2026-06-05 cs.CV

Wenhao Zhang , Ramin Ramezani , Tao Han , Kai Hwang , Minyi Guo This is my paper · ORCID

classification cs.CV

keywords causaldiscoverydirectedacyclicgraphacyclicityconstraintcontinuousoptimizationpolynomialtracegeometricseriesvisualsemanticgraphsNOTEARS

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 tries to show that checking acyclicity in continuous DAG learning does not require the matrix exponential: a finite sum of traces of powers of the squared adjacency matrix, h_poly = Σ tr((W◦W)^k) for k=1..d, is zero exactly when the graph is acyclic. It proves this equivalence (Theorem 1) and shows h_poly, the NOTEARS exponential constraint, and nilpotency of the weight matrix define the same feasible set (Theorem 2). To make the polynomial practical, it derives a geometric-series evaluation h_geo = tr((I−A)^{-1}(A−A^{d+1})) that replaces d sequential matrix multiplications with one linear solve and two matrix powers, yielding 14–33% wall-clock speedups and better structure recovery on synthetic and CelebA visual-attribute experiments. If right, any NOTEARS-style solver can swap in the cheaper constraint without changing the solution set.

What carries the argument

The central object is the trace of powers of the nonnegative matrix A=W◦W, whose (i,i) entries count closed walks of length k through node i. h_poly sums these traces up to k=d, exploiting the fact that any simple cycle in a d-node graph has length at most d. The geometric-series identity h_geo = tr((I−A)^{-1}(A−A^{d+1})) is the computational workhorse: it converts the d-step power loop into a single matrix linear solve plus two matrix powers, both of which are highly optimized BLAS operations.

What would settle it

Construct a d-node graph containing a single directed cycle of length L>16 (e.g., d=100, L=50) with unit edge weights, set W to its adjacency matrix, and evaluate the implemented h_geo with K=16; the constraint will return zero, falsely signaling acyclicity. A direct check of whether any reported d=500 run ever contains a cycle longer than 16 at intermediate iterates would also settle whether the experiments exercised the proved equivalence regime.

Watch

Extended reading notes

Core claim

The paper's central claim is that acyclicity can be characterized by a finite polynomial: for A=W◦W, h_poly(W)=Σ_{k=1}^d tr(A^k) equals zero if and only if the directed graph encoded by W is acyclic (Theorem 1). Equivalently, h_poly, the exponential constraint hexp, and the nilpotency of A all define the same feasible set (Theorem 2). The paper further shows the polynomial sum can be evaluated via the geometric-series identity h_geo(W)=tr((I−A)^{-1}(A−A^{d+1})), which uses one linear solve and two matrix powers instead of a d-step loop, and reports that this variant improves SHD from 318.4 to 285.4 and F1 from 0.725 to 0.756 averaged over d∈{100,200,500}, while running 14–33% faster than the

Load-bearing premise

The equivalence theorem requires summing matrix powers up to k=d, but the code used for experiments truncates the geometric series at K=16, so for graph sizes above 16 the implemented constraint is not the one proved zero exactly on acyclic graphs and nonzero on graphs with long cycles.

Editorial extensions

If this is right

Any NOTEARS-style continuous DAG learner can replace its exponential acyclicity penalty with h_poly without altering the set of feasible graphs; only the optimization trajectory changes.
At graph sizes around 100–500 nodes, the geometric variant cuts wall-clock time by 14–33% relative to the exponential baseline, which directly reduces cost of repeated bootstrap and hyperparameter sweeps.
The post-threshold diagnostics in the paper show polyDAG-Geo producing DAG-valid outputs in 100% of runs at d=100 under the tested settings, versus 33.3% for the exponential baseline.
The polynomial form motivates sparse evaluation strategies for graphs with few edges, potentially scaling to thousands of nodes.

Reading between the lines

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

The reported implementation truncates the geometric series at K=16, not at K=d, so for graphs with cycles longer than 16 nodes the implemented constraint is not provably equivalent to acyclicity; the experimental speedups should be read as testing the truncated constraint, not the theorem.
Because Theorem 2 guarantees identical feasible sets, any observed SHD/F1 advantage is attributed by the authors to optimization dynamics rather than target change; a reviewer could test this by swapping constraints across solvers.
The geometric-series identity requires (I−A) invertible; near-singularity during optimization is flagged by the authors as future work, so the robustness of the reported runs to ill-conditioned intermediate iterates is an open empirical question.
A natural extension, suggested by the paper's own discussion, is training a learned surrogate to predict h_poly in O(d^2) or O(d) time, which would bypass the cubic bottleneck entirely.

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.

Desk Editor's Note

read the letter

The paper's core theorem is right, and the writing is refreshingly clear about what is and isn't claimed. The polynomial trace constraint h_poly = Σ tr((W◦W)^k) for k=1..d is a correct acyclicity characterization for nonnegative A with zero diagonal, and the equivalence to the NOTEARS exponential constraint is proved cleanly. That is a legitimate, if modest, theoretical contribution: it makes explicit that the exponential is not unique. The geometric-series identity is standard linear algebra, but the application as a drop-in constraint is reasonable, and the authors are honest that they are not claiming an asymptotic win, only a constant-factor speedup. Credit also for shipping code, reporting post-threshold DAG validity, and acknowledging the CelebA graph is an exploratory dependency structure, not a causal mechanism.

The problem is the implementation. Section 4.1 fixes K=16 for the geometric variant, while Theorem 2 requires the sum to run to k=d. The stress-test note is right: h_16 is still zero on acyclic graphs, so the reader's claim about false positives on DAGs was wrong, but h_16 is also zero on any graph whose shortest directed cycle exceeds 16. For d=100,200,500, a 20-node directed cycle gives h_16=0 while the graph is cyclic. The implemented constraint is strictly weaker than the proved one, and the paper never addresses this. Tables 4–6 therefore do not test h_geo as defined. That is a load-bearing flaw for the empirical claims, not a harmless constant.

The secondary issue is statistical: the main synthetic tables use 3 seeds, and at d=100 and d=500 the SHD standard deviations overlap substantially between methods. The extra 10-seed run at d=100 helps, but the headline aggregate in Table 6 is computed over 9 runs total. The speedup numbers are plausible but should be treated as preliminary until the constraint is fixed.

The audience for this paper is people working on continuous DAG constraints who might adopt a cheaper acyclicity surrogate. They will get value from the equivalence proof and the implementation notes, but only after the truncation issue is resolved. The fix is straightforward: run with K=d, or prove that smaller K preserves the zero set under the optimization trajectory, which they do not do.

I would not cite this in its current form. If the authors revise to use K=d and re-run the experiments, the paper becomes a decent incremental contribution. As is, it deserves a serious referee who will catch the truncation problem, so I would not desk-reject it—but acceptance should be conditional on redoing the experiments with the constraint actually proved.

Referee Report

2 major / 6 minor

Summary. The paper proposes a polynomial acyclicity constraint for continuous DAG learning, h_poly(W) = Σ_{k=1}^d tr((W∘W)^k), and proves (Theorem 1) that it is zero exactly for acyclic graphs. Theorem 2 establishes equivalence with the NOTEARS exponential constraint and with nilpotency of W∘W. To reduce computational cost, the paper derives a geometric-series form h_geo(W) = tr((I-A)^{-1}(A-A^{d+1})) and evaluates it against NOTEARS-Exp on synthetic ER graphs (d = 100, 200, 500) and a CelebA visual-attribute experiment, reporting improved SHD/F1 and 14–33% wall-clock speedups. However, Section 4.1 fixes a truncated geometric order K=16 for all experiments, while the theoretical equivalence requires the sum to run to k=d. This mismatch means the implemented constraint is not the one analyzed in Theorems 1–2 for d>16, so the empirical validation does not currently support the paper's central claim.

Significance. Theorems 1 and 2 are correct for the exact h_poly and provide a clean, proof-based justification for substituting a finite polynomial trace constraint for the NOTEARS exponential. The geometric-series identity is standard and, when K=d, exactly reproduces h_poly. If validated, the proposed constraint would be a practically useful drop-in alternative with a smaller constant factor. The manuscript's reproducibility commitments are strong: public code, benchmark scripts, seeds, and post-threshold DAG-validity diagnostics. However, the experiments use K=16, which invalidates the zero-iff-acyclic equivalence for d>16; a 20-node directed cycle, for example, yields h_geo,K=0 while h_geo=20. Thus Tables 4–6 do not test the constraint proved equivalent, and the reported improvements may be due to a relaxed acyclicity condition. The gap is fixable by running with K=d or larger, but the empirical claims are not yet supported.

major comments (2)

[§4.1 (Hyperparameters) vs. Eq. (6), Theorem 2] The implementation fixes K=16, but Eq. (6) and Theorem 2 define h_geo with the sum to k=d. For d=100,200,500, the implemented function is h_geo,K = tr(Σ_{k=1}^{16} A^k), not h_geo = tr(Σ_{k=1}^d A^k). This constraint is zero on any directed cycle longer than 16 (e.g., a 20-cycle), so it is not zero-iff-acyclic. Consequently, Tables 4–6 do not evaluate the constraint proved equivalent in Theorem 2, and the statement in §4.2 that 'the geometric implementation preserves the theoretical acyclicity characterization' is unsupported. Please rerun with K=d (or K≥d) or explicitly reframe the experiments as a heuristic truncation and temper the equivalence claim.
[§4.2, Tables 4–6 (missing ablation)] The paper never reports results for h_geo with K=d, the only version covered by the theory. This leaves open the possibility that the observed improvements in SHD/F1 and runtime come from the relaxed constraint rather than from the geometric evaluation itself. An ablation comparing polyDAG-Geo with K=16 vs. K=d (and, if possible, with K=2d as a robustness check) is needed to separate the effect of the closed-form evaluation from the effect of truncation on the feasible set.

minor comments (6)

[Table 7, §4.6] The methods are listed as 'ASDAG-Poly' and 'ASDAG-Geo', but these names are not defined anywhere; they should presumably be 'polyDAG-Poly' and 'polyDAG-Geo'.
[Tables 2 and 4] Runtime values for the same settings disagree between the two tables (e.g., d=100: 3.70 vs. 3.44 s for polyDAG-Geo; 5.49 vs. 5.16 s for NOTEARS). Please reconcile and clarify what each table reports.
[§4.3] The title 'Reviewer-Requested Scalability Extensions' is inappropriate in a journal article; rephrase to a neutral description of the additional experiments.
[§3.5] The claim that repeated squaring 'features a smaller constant than d separate multiplications' is an empirical assertion; either provide supporting measurements or soften the wording.
[§4.6] The text says 'we construct a continuous attribute matrix' but CelebA attributes are binary; clarify how the semantic variables are extracted and whether they are treated as continuous or binary.
[§5.6 vs. Table 1] Section 5.6 states that both constraint variants rely on O(d^3) dense linear algebra, while Table 1 lists polyDAG-Geo as O(d^3 log d). Make the complexity statements consistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the acyclicity characterization and geometric-series evaluation are derived from first principles, and the K=16 truncation is a correctness concern, not a circular step.

full rationale

The paper's central derivation is self-contained. Theorem 1 defines h_poly(W) as a finite sum of traces of powers of A = W∘W and proves, via nonnegativity and standard walk-counting, that it is zero iff the encoded graph is acyclic. This is a mathematical derivation independent of any fitted data. Theorem 2's equivalence among h_poly, h_exp, and nilpotency of W∘W is proved in the text using nonnegative summands in the exponential series and the classical graph/nilpotency criterion, not imported from a self-citation. The geometric-series identity (I-A)S_d = A - A^{d+1} is an algebraic identity, and h_geo in Eq. (6) follows by inversion; no ansatz is smuggled in via citation. The empirical section compares polyDAG-Geo with NOTEARS-Exp on synthetic ER ground-truth graphs and CelebA without fitting any parameter to those outcomes, so the reported improvements are not forced by construction. The flagged implementation detail — Section 4.1 states 'For polyDAG-Geo, we use truncated geometric order K=16 in all reported experiments,' while Eq. (6) and Theorem 2 require the sum to run to d, and Appendix B.3 confirms the implementation computes trace(solve(I-A, A-A^{K+1})) — is a genuine gap between theorem and implementation: for d > 16, the implemented constraint can vanish on graphs whose only directed cycles are longer than 16. However, this is a correctness/validity issue, not circularity: the theorem is not defined in terms of the experimental result, and no fitted parameter is renamed as a prediction. The derivation chain itself is not circular, so the circularity score is 0; the K=16 mismatch belongs under correctness risk rather than under circularity analysis.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on classical graph theory (walk counts, nilpotency) and standard matrix identities. The two non-standard axioms are the invertibility of I-A at all iterates and the use of K=16, both of which are unverified or contradict the theory.

free parameters (1)

Geometric truncation order K = 16
The hyperparameters section fixes K=16 for polyDAG-Geo in all experiments. Theorem 2 requires K=d for exactness; for d>16 the implemented constraint is not the one proven equivalent. This is a hand-chosen value that changes the method.

assumptions (4)

standard math For a nonnegative matrix A, nilpotency (A^d=0) is equivalent to the directed graph being acyclic.
Invoked in Theorem 1 (second direction) and Theorem 2 to link acyclicity to nilpotency; cites Horn and Johnson (2012).
standard math (A^k)_{ij} counts walks of length k from i to j, so diagonal entries count closed walks.
Basis for the polynomial trace constraint; cites Harary (1962, 1965). Classical result.
domain assumption (I - A) is invertible at every optimization step where h_geo is evaluated.
Required to use S = (I-A)^{-1}(A-A^{d+1}). The paper notes I-A may be singular or ill-conditioned and states it does not regularize, so this is an unverified numerical assumption.
ad hoc to paper Truncating the geometric series at K=16 preserves the zero-iff-acyclic characterization.
The experiments set K=16, which is not consistent with Theorem 1 (which requires k up to d). No proof or discussion is provided for this truncation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of polyDAG: Polynomial Acyclicity Constraints for Efficient Continuous Causal Discovery in Visual Semantic Graphs." pith.science (2026). https://pith.science/paper/FMVX4PS2

@misc{pith2026260606908,
  author       = {Pith},
  title        = {Pith review of: polyDAG: Polynomial Acyclicity Constraints for Efficient Continuous Causal Discovery in Visual Semantic Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FMVX4PS2}},
  note         = {Machine review of arXiv:2606.06908}
}

read the original abstract

Modern image-analysis pipelines often convert images into structured semantic variables, such as facial attributes, object concepts, and scene descriptors. Learning directed dependencies among these variables can produce interpretable visual semantic graphs, but continuous directed acyclic graph learning is limited by the cost of enforcing acyclicity. We present polyDAG, a polynomial acyclicity framework for efficient continuous causal discovery in visual semantic graphs. polyDAG replaces the matrix-exponential acyclicity constraint with a finite polynomial trace constraint and proves that the new constraint is zero exactly for acyclic graphs. We further derive a geometric-series implementation that avoids the explicit summation loop while preserving the same acyclicity condition. Experiments on synthetic Erdos-Renyi graphs and CelebA facial visual attributes show that polyDAG improves efficiency and structure recovery. Averaged over the revised synthetic protocol with d in {100, 200, 500}, polyDAG reduces mean structural Hamming distance from 318.4 to 285.4 and improves mean F1 score from 0.725 to 0.756. At 100 nodes, the geometric variant runs in 3.44 seconds compared with 5.16 seconds for the exponential baseline, corresponding to a 33.4 percent speedup. Code and data are publicly available at https://github.com/wenhaoz-fengcai/polyDAG.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 4 linked inside Pith

[1]

and Dobbs, I

Beavis, B. and Dobbs, I. (1990). Optimisation and Stability Theory for Economic Analysis . Cambridge University Press

1990
[2]

Bello, K., Aragam, B., and Ravikumar, P. (2022). Dagma: Learning dags via m-matrices and a log-determinant acyclicity characterization. In Advances in Neural Information Processing Systems , volume 35, pages 8226--8239

2022
[3]

Bertsekas, D. P. (1997). Nonlinear programming. Journal of the Operational Research Society , 48:334--334

1997
[4]

and Vandenberghe, L

Boyd, S. and Vandenberghe, L. (2004). Convex Optimization . Cambridge University Press

2004
[5]

Brouillard, P., Lachapelle, S., Lacoste-Julien, S., and Drouin, A. (2020). Differentiable causal discovery from interventional data. Advances in Neural Information Processing Systems , 33:21865--21877

2020
[6]

C., Walker, I., and Glocker, B

Castro, D. C., Walker, I., and Glocker, B. (2020). Causality matters in medical imaging. Nature Communications , 11(1):3673

2020
[7]

Charpentier, B., Kibler, S., and G \"u nnemann, S. (2022). Differentiable dag sampling. In International Conference on Learning Representations

2022
[8]

and Varadhan, S

Chatterjee, S. and Varadhan, S. R. S. (2011). The large deviation principle for the erd o s-r \'e nyi random graph. European Journal of Combinatorics , 32(7):1000--1017

2011

Show all 52 references

[9]

Chickering, D. M. (2002). Optimal structure identification with greedy search. Journal of Machine Learning Research , 3:507--554

2002
[10]

Chickering, M., Heckerman, D., and Meek, C. (2004). Large-sample learning of bayesian networks is np-hard. Journal of Machine Learning Research , 5:1287--1330

2004
[11]

Cundy, C., Grover, A., and Ermon, S. (2021). Bcd nets: Scalable variational approaches for bayesian causal discovery. In Advances in Neural Information Processing Systems , volume 34

2021
[12]

Duong, B., Le, H., Huang, B., and Nguyen, T. (2025). Reinforcement learning for causal discovery without acyclicity constraints. Transactions on Machine Learning Research , 2025

2025
[13]

and Glowinski, R

Fortin, M. and Glowinski, R. (2000). Augmented Lagrangian Methods: Applications to the Numerical Solution of Boundary-Value Problems . Elsevier

2000
[14]

Harary, F. (1962). The determinant of the adjacency matrix of a graph. SIAM Review , 4(3):202--210

1962
[15]

Z., and Cartwright, D

Harary, F., Norman, R. Z., and Cartwright, D. (1965). Structural Models: An Introduction to the Theory of Directed Graphs . Wiley

1965
[16]

Hastie, T., Tibshirani, R., and Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction . Springer, 2 edition

2009
[17]

Horn, R. A. and Johnson, C. R. (2012). Matrix Analysis . Cambridge University Press, 2 edition

2012
[18]

M., Peters, J., and Sch \"o lkopf, B

Hoyer, P., Janzing, D., Mooij, J. M., Peters, J., and Sch \"o lkopf, B. (2008). Nonlinear causal discovery with additive noise models. In Advances in Neural Information Processing Systems , volume 21

2008
[19]

Kalainathan, D., Goudet, O., Guyon, I., Lopez-Paz, D., and Sebag, M. (2022). Structural agnostic modeling: Adversarial learning of causal graphs. Journal of Machine Learning Research , 23(219):1--62

2022
[20]

Kingma, D. P. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

2014 arXiv
[21]

and Friedman, N

Koller, D. and Friedman, N. (2009). Probabilistic Graphical Models: Principles and Techniques . MIT Press

2009
[22]

Lachapelle, S., Brouillard, P., Deleu, T., and Lacoste-Julien, S. (2019). Gradient-based neural dag learning. arXiv preprint arXiv:1906.02226

2019 arXiv
[23]

D., Hoang, T., Li, J., Liu, L., Liu, H., and Hu, S

Le, T. D., Hoang, T., Li, J., Liu, L., Liu, H., and Hu, S. (2016). A fast pc algorithm for high dimensional causal discovery with multi-core pcs. IEEE/ACM Transactions on Computational Biology and Bioinformatics , 16(5):1483--1495

2016
[24]

Lippe, P., Cohen, T., and Gavves, E. (2022). Efficient neural causal discovery without acyclicity constraints. In International Conference on Learning Representations

2022
[25]

Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages 3730--3738

2015
[26]

K., and Regev, A

Lopez, R., H \"u tter, J.-C., Pritchard, J. K., and Regev, A. (2022). Large-scale differentiable causal discovery of factor graphs. In Advances in Neural Information Processing Systems , volume 35

2022
[27]

Lorch, L., Rothfuss, J., Sch \"o lkopf, B., and Krause, A. (2021). Dibs: Differentiable bayesian structure learning. In Advances in Neural Information Processing Systems , volume 34

2021
[28]

Madras, D., Creager, E., Pitassi, T., and Zemel, R. (2019). Fairness through causal awareness: Learning causal latent-variable models for biased data. In Proceedings of the Conference on Fairness, Accountability, and Transparency , pages 349--358

2019
[29]

Ng, I., Ghassami, A., and Zhang, K. (2020). On the role of sparsity and dag constraints for learning linear dags. In Advances in Neural Information Processing Systems , volume 33, pages 17943--17954

2020
[30]

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019). Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems , volume 32

2019
[31]

Pearl, J. (2009). Causality: Models, Reasoning, and Inference . Cambridge University Press

2009
[32]

and B \"u hlmann, P

Peters, J. and B \"u hlmann, P. (2014). Identifiability of gaussian structural equation models with equal error variances. Biometrika , 101(1):219--228

2014
[33]

M., Janzing, D., and Sch \"o lkopf, B

Peters, J., Mooij, J. M., Janzing, D., and Sch \"o lkopf, B. (2017). Elements of Causal Inference: Foundations and Learning Algorithms . MIT Press

2017
[34]

Ramsey, J., Glymour, M., Sanchez-Romero, R., and Glymour, C. (2017). A million variables and more: the fast greedy equivalence search algorithm for learning high-dimensional graphical causal models. International Journal of Data Science and Analytics , 3(2):121--129

2017
[35]

G., Seiler, C., and Weichwald, S

Reisach, A. G., Seiler, C., and Weichwald, S. (2021). Beware of the simulated dag! varsortability and a standardized benchmark. In Advances in Neural Information Processing Systems , volume 34

2021
[36]

A., and Nolan, G

Sachs, K., Perez, O., Pe'er, D., Lauffenburger, D. A., and Nolan, G. P. (2005). Causal protein-signaling networks derived from multiparameter single-cell data. Science , 308(5721):523--529

2005
[37]

R., Kalchbrenner, N., Goyal, A., and Bengio, Y

Sch lkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., and Bengio, Y. (2021). Toward causal representation learning. Proceedings of the IEEE , 109(5):612--634

2021
[38]

O., Hyv \"a rinen, A., Kerminen, A., and Jordan, M

Shimizu, S., Hoyer, P. O., Hyv \"a rinen, A., Kerminen, A., and Jordan, M. (2006). A linear non-gaussian acyclic model for causal discovery. Journal of Machine Learning Research , 7:2003--2030

2006
[39]

N., and Scheines, R

Spirtes, P., Glymour, C. N., and Scheines, R. (2000). Causation, Prediction, and Search . MIT Press, 2 edition

2000
[40]

Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B , 58(1):267--288

1996
[41]

E., and Aliferis, C

Tsamardinos, I., Brown, L. E., and Aliferis, C. F. (2006). The max-min hill-climbing bayesian network structure learning algorithm. Machine Learning , 65(1):31--78

2006
[42]

Wang, L., Huang, S., Wang, S., Liao, J., Li, T., and Liu, L. (2024). A survey of causal discovery based on functional causal model. Engineering Applications of Artificial Intelligence , 133:108258

2024
[43]

Williams, V. V. (2012). Multiplying matrices faster than coppersmith-winograd. In Proceedings of the 44th Annual ACM Symposium on Theory of Computing , pages 887--898. ACM

2012
[44]

Xia, Y., Zhang, H., Ren, Y., Guan, J., and Zhou, S. (2023). Causal discovery by continuous optimization with conditional independence constraint: Methodology and performance. In IEEE International Conference on Data Mining (ICDM) , pages 668--677

2023
[45]

Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., and Wang, J. (2021a). Causalvae: Disentangled representation learning via neural structural causal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9593--9602
[46]

Yang, X., Zhang, H., Qi, G., and Cai, J. (2021b). Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9847--9857
[47]

Yu, Y., Chen, J., Gao, T., and Yu, M. (2019). Dag-gnn: Dag structure learning with graph neural networks. In International Conference on Machine Learning , pages 7154--7163. PMLR

2019
[48]

Yu, Y., Gao, T., Yin, N., and Ji, Q. (2021). Dag learning on the permutohedron. arXiv preprint arXiv:2112.11555

2021 arXiv
[49]

Yue, Z., Sun, Q., Hua, X.-S., and Zhang, H. (2021). Transporting causal mechanisms for unsupervised domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 8599--8608

2021
[50]

Zheng, X., Aragam, B., Ravikumar, P., and Xing, E. P. (2018). Dags with no tears: Continuous optimization for structure learning. In Advances in Neural Information Processing Systems , volume 31

2018
[51]

Zheng, X., Dan, C., Aragam, B., Ravikumar, P., and Xing, E. P. (2020). Learning sparse nonparametric dags with reinforcement learning. In International Conference on Artificial Intelligence and Statistics , pages 3414--3425. PMLR

2020
[52]

Zhu, S., Ng, I., and Chen, Z. (2019). Causal discovery with reinforcement learning. arXiv preprint arXiv:1906.04477

2019 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

and Dobbs, I

Beavis, B. and Dobbs, I. (1990). Optimisation and Stability Theory for Economic Analysis . Cambridge University Press

1990

[2] [2]

Bello, K., Aragam, B., and Ravikumar, P. (2022). Dagma: Learning dags via m-matrices and a log-determinant acyclicity characterization. In Advances in Neural Information Processing Systems , volume 35, pages 8226--8239

2022

[3] [3]

Bertsekas, D. P. (1997). Nonlinear programming. Journal of the Operational Research Society , 48:334--334

1997

[4] [4]

and Vandenberghe, L

Boyd, S. and Vandenberghe, L. (2004). Convex Optimization . Cambridge University Press

2004

[5] [5]

Brouillard, P., Lachapelle, S., Lacoste-Julien, S., and Drouin, A. (2020). Differentiable causal discovery from interventional data. Advances in Neural Information Processing Systems , 33:21865--21877

2020

[6] [6]

C., Walker, I., and Glocker, B

Castro, D. C., Walker, I., and Glocker, B. (2020). Causality matters in medical imaging. Nature Communications , 11(1):3673

2020

[7] [7]

Charpentier, B., Kibler, S., and G \"u nnemann, S. (2022). Differentiable dag sampling. In International Conference on Learning Representations

2022

[8] [8]

and Varadhan, S

Chatterjee, S. and Varadhan, S. R. S. (2011). The large deviation principle for the erd o s-r \'e nyi random graph. European Journal of Combinatorics , 32(7):1000--1017

2011