Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Causal Abstraction Learning based on the Semantic Embedding Principle

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a linear causal abstraction between two Gaussian structural causal models can be learned from observational data and a partial structural prior by minimizing KL divergence on the Stiefel manifold.

desk verdict The paper's methods are sound for orthogonal linear causal abstractions, but the claim that SEP forces the Stiefel manifold is unsupported, making the advertised scope broader than what is actually solved. read the letter →

arxiv 2502.00407 v2 pith:LKDDGMCJ submitted 2025-02-01 cs.LG cs.AI

Gabriele D'Acunto , Fabio Massimo Zennaro , Yorgos Felekis , Paolo Di Lorenzo This is my paper · ORCID
classification cs.LGcs.AI
keywords causalabstractionsemanticembeddingprincipleStiefelmanifoldRiemannianoptimizationGaussianstructuralmodelslearningKLdivergencespectralinterlacing
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 is trying to establish that causal abstraction—a map between a fine-grained and a coarse-grained causal model—can be learned from purely observational data, without access to the model equations, without interventional data, and without aligned samples. The central move is the semantic embedding principle (SEP), which says the coarse (high-level) distribution lives inside the fine (low-level) one as a low-dimensional subspace; for linear maps this forces the abstraction matrix onto the Stiefel manifold of orthonormal projections. The learning problem then becomes Riemannian optimization of the KL divergence between the coarse distribution and the projected fine distribution, subject to a partial prior on which fine variables map to which coarse variables. If the paper is right, causal abstraction learning is tractable in realistic settings, and a spectral interlacing condition can tell you in advance whether any exact linear abstraction exists. The proposed algorithms—LinSEPAL-ADMM, LinSEPAL-PG, and CLinSEPAL—recover constructive abstractions on synthetic data and on resting-state fMRI brain networks.

What carries the argument

The load-bearing object is $V \in \mathrm{St}(\ell,h)$, the Stiefel manifold of matrices with orthonormal columns, which implements the semantic embedding principle for linear CA: requiring orthonormality makes the high-level geometry a genuine subspace embedding and gives a well-posed projection. The objective is the Gaussian KL divergence $D_{\mathrm{KL}}^V = \operatorname{Tr}\{(V^\top\Sigma_\ell V)^{-1}\Sigma_h\} + \log\det(V^\top\Sigma_\ell V) + C$, whose gradient the paper computes in closed form and projects onto the manifold tangent space. The spectral interlacing theorem (Theorem 4.3) is the paper's necessary-condition test, and the algorithms—MADMM- and ManPG-based solvers for the nonsmooth problem plus an SOC-ADMM-SCA solver for the smooth one—are the mechanisms that make the optimization practical.

What would settle it

Construct two Gaussian datasets whose covariance eigenvalues violate the interlacing bound, e.g. with $\kappa_1 < \lambda_1$ or $\kappa_h > \lambda_\ell$, and run CLinSEPAL: a returned low-KL exact reconstruction would contradict Theorem 4.3. Conversely, generate high-level data as a nonlinear function of low-level data plus independent noise; the method should show systematically degrading structural recovery (F1 drop, higher Frobenius distance) as noise grows, confirming that exact linear pushforward is the load-bearing premise.

Watch

Extended reading notes

Core claim

For zero-mean Gaussian SCMs, the paper shows that learning a linear constructive abstraction reduces to minimizing $D_{\mathrm{KL}}(\chi_h \,\|\, \varphi_{V^\#}(\chi_\ell))$ over $V \in \mathrm{St}(\ell,h)$, the Stiefel manifold of $\ell \times h$ matrices with orthonormal columns. Because SEP makes the high-level distribution the pushforward of the low-level one under $V^*$, the objective has a zero exactly at the true abstraction, and the search can be driven by a closed-form gradient. The paper also proves a necessary spectral condition: if a linear CA exists, the eigenvalues $\lambda_i$ of $\Sigma_\ell$ and $\kappa_i$ of $\Sigma_h$ must interlace as $\lambda_i \le \kappa_i \le \lambda_{i+\ell-h}$ for every $i \in [h]$. Two problem formulations follow—a nonsmooth penalized one (Problem 2) and a smooth constructive one (Problem 3) that learns the support of the abstraction—and the paper derives three algorithms from Riemannian optimization paradigms for them.

Load-bearing premise

The method assumes the high-level distribution is exactly the linear projection of the low-level distribution through some orthonormal matrix $V^*$, so that a perfect zero-KL solution exists; if the high-level data carries its own noise or is produced by a nonlinear aggregation, the learned abstraction can be biased even when the optimization converges.

Editorial extensions

If this is right

  • A necessary and cheap pre-check: if the sorted eigenvalues of the fine-scale covariance do not bracket the coarse-scale eigenvalues, no orthonormal linear abstraction exists, so one can reject impossible granularities before running any optimization.
  • When full structural prior knowledge is available, the methods recover the true abstraction structure with near-perfect F1 score; with partial prior knowledge, CLinSEPAL is the only formulation that guarantees a constructive abstraction.
  • The method transfers to real brain data, recovering a ground-truth map from 45 brain-region covariances to 14 macro-regions (full prior) or 8 functional networks (partial prior) with low uncertainty.
  • The category-theoretic reformulation separates the semantic layer from the graphical layer, opening a route to translate SCM identifiability results into CA identifiability results, possibly achieving interventional consistency without interventional data.

Reading between the lines

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

  • The KL objective is zero only when the high-level data is exactly the linear pushforward of the low-level data, so under measurement noise or nonlinear aggregation the recovered $V$ will be the best linear projection rather than the true abstraction; a robust divergence or an explicit noise model would be a natural extension.
  • The interlacing theorem can be read as a model-selection tool on covariances alone: checking it before optimization tells which coarse granularities are even compatible with a given fine model.
  • The same SEP-plus-Stiefel mechanism should extend to other divergences, such as Wasserstein distances, and to exponential-family SCMs, where the projection geometry changes but the embedding principle remains the same.
  • In high coarse-graining settings the kernel of the KL objective grows, so recovering exact coefficients becomes harder while structure recovery stays reliable; downstream users should trust the learned support of the abstraction more than its precise weights.
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

3 major / 5 minor

Summary. The paper proposes a framework for learning causal abstractions (CAs) between structural causal models from observational data, based on a 'semantic embedding principle' (SEP). SCMs are formalized as functors into a category of probability spaces, SEP is defined via the existence of a right-inverse for the abstraction map, and the paper then specializes to linear Gaussian CAs, claiming that SEP naturally implies the abstraction matrix lies on the Stiefel manifold. Three optimization methods are developed (LinSEPAL-ADMM, LinSEPAL-PG, CLinSEPAL) for nonsmooth and smooth formulations with partial structural priors, together with a spectral interlacing necessary condition (Theorem 4.3). The methods are evaluated on synthetic data with full and partial priors and on resting-state fMRI data from the Human Connectome Project.

Significance. If the identification of SEP with the Stiefel manifold were valid, the paper would provide a principled optimization formulation for CA learning with partial priors, a spectral necessary condition for existence, and an open-source implementation. The gradient computations, optimization derivations, and algorithmic details are presented carefully, and the brain-imaging application is a valuable demonstration of the intended setting. However, the central conceptual step is a non-sequitur: SEP as defined does not imply orthonormality of the abstraction map, so the problems solved and the theorem proven characterize orthogonal linear abstractions rather than linear constructive abstractions generally. Thus the paper is a useful contribution to learning orthogonal linear CAs, but its claimed scope over general linear CA learning is not established as written.

major comments (3)
  1. [Section 4, Definition 4.2] The claim that 'SEP implies V∈St(ℓ,h)' is unsupported and in fact false. Definition 4.1 requires αX to have a right-inverse βX; for αX=V^⊤ this means there exists W with V^⊤W=I_h, which is equivalent to rank(V)=h, not to V^⊤V=I_h. A standard constructive linear abstraction, e.g., the cluster-average map with V_{i,j}=1/|C_j| for i∈C_j and 0 otherwise, satisfies SEP with W_{i,j}=1 for i∈C_j, yet V^⊤V=diag(1/|C_j|)≠I_h. Since Problems 2 and 3, the KL objective in Eq. (3), and Theorem 4.3 all impose the Stiefel constraint, the methods and theorem characterize only orthogonal linear CAs. The paper should either explicitly restrict its claims to orthogonal abstractions or present the Stiefel constraint as an additional modeling choice, revising Definition 4.2 and the abstract accordingly.
  2. [Section 6, partial prior experiments] The partial prior evaluation considers only a single dimensionality, (ℓ,h)=(4,2), with three fractions of known rows. The abstract and conclusion claim success 'with different degrees of prior information' and across 'different levels of prior knowledge', but no other ℓ,h settings are reported for the partial-prior case, and the figures do not include error bars or dispersion measures across the R initializations/simulations. This narrow evaluation does not support the general claim; please add additional settings or temper the conclusions.
  3. [Appendix K, Eq. (118)] The constructiveness metric is defined as constr = (number of rows with one nonzero entry)/ℓ + (number of columns with at least one nonzero entry)/h. For a fully constructive matrix, both fractions equal one, so constr=2. However, the text says a learned map is counted as constructive when constr=1, which is inconsistent and likely introduces a threshold that is not the one intended. This ambiguity affects the interpretation of the constructiveness plots in Section 6 and Section 7 and should be corrected.
minor comments (5)
  1. [Section 4, Fig. 1 and Def. 4.2] The notation for pushforwards is inconsistent: the paper defines φ_V as the map x↦V^⊤x, but Def. 4.2 writes χh = φ_{V∘V^⊤}#(χh), where V∘V^⊤ maps R^ℓ to R^ℓ while χh lives on R^h. Please clarify the domains and use a uniform notation.
  2. [Appendix G, Theorem 4.3 proof] In the proof, the notation 'eigvls(V^⊤V)_1 ≤ ϑ_i ≤ eigvls(V^⊤V)_h' should be defined explicitly, for example by specifying the ordering of the eigenvalues used in the Ostrowski bound.
  3. [Appendix K, metrics] The definition of the F1 score contains an incomplete sentence after 'Given' and before 'True positive rate', which makes the metric definition hard to parse. Please rewrite this item as a complete, self-contained definition.
  4. [Section 4, Problem 1] The formulation assumes that the high-level distribution is exactly the pushforward of the low-level distribution under the true CA, i.e., χh = φ_{αX}#(χℓ). Under model misspecification the KL objective cannot attain zero and the learned abstraction may be biased; a brief discussion of robustness to misspecification, or a simulation study, would strengthen the paper.
  5. [Appendix L, Fig. 7] The legend in Fig. 11 labels a method 'madmm-sca', which appears to refer to CLinSEPAL; please make the labels consistent with the method names used in the main text.

Circularity Check

1 steps flagged · score 4.0 of 10

The paper's linear SEP is stipulated rather than derived: Definition 4.2 replaces the right-inverse β_X from Definition 4.1 by V itself, turning a rank condition into the Stiefel constraint, while the downstream learning against external covariance data remains non-circular.

  1. self definitional [Section 4, paragraph before Eq. (2) and Definition 4.2]
    "Importantly, for linear CA, a simple yet principled way to satisfy SEP is via the geometry of the Stiefel manifold: St(ℓ, h) := {V ∈ Rℓ×h | V⊤V = Ih}. ... Definition 4.2 (Semantic embedding principle, linear case). Given the linear constructive CA, viz. V⊤, SEP implies that V ∈ St(ℓ, h). From Eq. (1) we get χh = φV◦V⊤ #(χh)."

    Definition 4.1 requires only that αX have some right-inverse βX, i.e., αX ◦ βX = Id; for αX = V⊤ this means V⊤W = Ih for some W, equivalent to rank(V⊤) = h, not to V⊤V = Ih. Definition 4.2 silently chooses W = V and asserts V ∈ St(ℓ,h). The Stiefel constraint is therefore not a consequence of SEP; it is an additional ansatz inserted as the definition of linear SEP. All subsequent results—Problems 2 and 3, Theorem 4.3, and the LinSEPAL/CLinSEPAL algorithms—solve this restricted orthogonal problem, so the claimed derivation 'SEP implies V∈St' is stipulated by the choice βX = V rather than derived from Definition 4.1.

full rationale

The empirical learning pipeline is not circular: KL divergence is evaluated between the abstracted low-level Gaussian φV#(χℓ) and the high-level Gaussian χh, using covariances Σℓ and Σh that are external inputs, and the optimization is run against those data and the prior B. The synthetic and brain experiments are likewise self-contained benchmarks, and the spectral interlacing theorem is a standard mathematical consequence of V⊤ΣℓV = Σh for V∈St(ℓ,h), not a fitted parameter renamed as a prediction. The only load-bearing reduction is Definition 4.2: under Definition 4.1, SEP is the existence of a right inverse for V⊤, which is a rank condition; Definition 4.2 converts this into orthonormality by substituting the right inverse with V itself. This makes the central restriction of the method definitional rather than derived, and it narrows the class of linear abstractions to orthogonal maps, excluding otherwise valid constructive abstractions such as cluster averaging with non-unit-norm columns. Because the paper's claimed first-principles link between SEP and the Stiefel manifold reduces by construction to this substitution, while the data-fitting and benchmark content remain independent, a moderate circularity score is appropriate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central assumption is SEP and the linear orthogonal CA form. The paper introduces SEP as a new principle, but it is not derived from external axioms. The free parameters are mostly optimization hyperparameters, which are typical. The spectral theorem relies on standard linear algebra. Overall, the framework is reasonable but rests on assumptions that may not hold in general real-world settings.

free parameters (3)
  • lambda = 1.0 in experiments
    Regularization weight for the nonsmooth penalty in LinSEPAL; chosen by hand, not tuned.
  • rho = 1.0 (CLinSEPAL, LinSEPAL-ADMM) or 1/(2||Sigma_l||_F^2) (LinSEPAL-PG)
    ADMM and proximal step size; set without a tuning procedure.
  • S = learned in [0,1]^l x h
    Structural support matrix learned in CLinSEPAL to enforce constructiveness; it is a variable optimized from data, not tied to external ground truth in partial prior settings.
assumptions (4)
  • domain assumption The high-level distribution is exactly the pushforward of the low-level distribution under the true abstraction: chi_h = phi_{V*^T}#(chi_l).
    This is the SEP principle in Problem 1 and is the target of the KL objective. If the high-level SCM is not an exact projection of the low-level SCM, the zero-minimum may not be achievable.
  • domain assumption The abstraction map is linear and orthogonal (V in St(ell,h)).
    Defines the class of CAs considered. The paper does not justify why real-world abstractions should be orthogonal; it follows from their SEP definition but may be restrictive.
  • standard math Covariances are positive definite and estimated from data.
    Needed for the KL divergence to be finite and for the gradient formulas. In practice, sample covariances may be singular, especially when h is close to ell or the number of samples is low.
  • ad hoc to paper The category-theoretic framework assumes SCMs can be represented as functors from Ind to Prob.
    This is a formalization choice; it does not add assumptions about the data generating process beyond those in standard SCM definitions.
invented entities (1)
  • Semantic Embedding Principle (SEP)
    purpose: A principle that constrains the abstraction map to admit a right inverse (for linear case, V in St(ell,h)).
    SEP is postulated as the defining property of a good abstraction. It gives a geometric constraint but is not independently testable outside the paper's framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Abstraction Learning based on the Semantic Embedding Principle." pith.science (2026). https://pith.science/paper/LKDDGMCJ

@misc{pith2026250200407,
  author       = {Pith},
  title        = {Pith review of: Causal Abstraction Learning based on the Semantic Embedding Principle},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LKDDGMCJ}},
  note         = {Machine review of arXiv:2502.00407}
}
read the original abstract

Structural causal models (SCMs) allow us to investigate complex systems at multiple levels of resolution. The causal abstraction (CA) framework formalizes the mapping between high- and low-level SCMs. We address CA learning in a challenging and realistic setting, where SCMs are inaccessible, interventional data is unavailable, and sample data is misaligned. A key principle of our framework is semantic embedding, formalized as the high-level distribution lying on a subspace of the low-level one. This principle naturally links linear CA to the geometry of the Stiefel manifold. We present a category-theoretic approach to SCMs that enables the learning of a CA by finding a morphism between the low- and high-level probability measures, adhering to the semantic embedding principle. Consequently, we formulate a general CA learning problem. As an application, we solve the latter problem for linear CA; considering Gaussian measures and the Kullback-Leibler divergence as an objective. Given the nonconvexity of the learning task, we develop three algorithms building upon existing paradigms for Riemannian optimization. We demonstrate that the proposed methods succeed on both synthetic and real-world brain data with different degrees of prior information about the structure of CA.

Figures

Figures reproduced from arXiv: 2502.00407 by the authors.

Figure 1
Figure 1. Pictorial representation of SEP for linear CA. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An abstraction as natural transformation, that [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Synthetic fp results for all settings (ℓ, h) and methods: (i) fraction of learned CAs that are construc￾tive, (ii) DKL Vb , (iii) normalized absolute Frobenius dis￾tance from V⋆ , and (iv) F1 score. has the correct structure (cf. F1 score). This suggests that under a high coarse-graining, the size of ker DKL grows, and it is more difficult for our methods to estimate V⋆ un￾der (NA1)-(NA5). Finally, the F1 score conf… view at source ↗
Figures from the paper (7 more)
Figure 5
Figure 5. Figure 5: An SCM is a functor (purple arrows) from [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Representation of Mℓ (blue), Mℓ ι (cyan), Mh (red), Mh ι (orange) as functors. An abstraction is just a natural transformation, that is, a set of commuting arrows in Prob (dashed black). Notice two commuting diagrams in Prob: the first observational one rooted on the e…
Figure 7
Figure 7. Figure 7: The figure shows the results in the pp setting where the [PITH_FULL_IMAGE:figures/full_fig_p033_7.png]
Figure 8
Figure 8. Figure 8: The figure shows (top) the ground truth linear CA and (bottom) the learned linear CA for the simulated [PITH_FULL_IMAGE:figures/full_fig_p034_8.png]
Figure 9
Figure 9. Figure 9: Starting from the top, the figure shows (i) the ground truth structure for linear CA, and the partial prior provided as input to CLinSEPALlearned linear CA for the simulated partial prior in the (ii) low, (iii) medium, and (iv) high uncertainty settings discussed in Se…
Figure 10
Figure 10. Figure 10: Starting from the top, the figure shows (i) the ground truth linear CA, and the learned linear CA for the simulated partial prior setting with (ii) low, (iii) medium, and (iv) high uncertainty in Sec. 7. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_10.png]
Figure 11
Figure 11. Figure 11: Starting from the left, the figure provides the [PITH_FULL_IMAGE:figures/full_fig_p037_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. Compositional Semantic Communication for Physical AI: Category Theory Meets Game Theory

    cs.IT 2026-07 reject novelty 6.0 of 10

    A category-theoretic and game-theoretic framework for compositional semantic communication is proposed, but its key measure is defined via learned functions and its existence theorems rest on unverified assumptions.

Reference graph

Works this paper leans on

56 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Optimization Algorithms on Matrix Manifolds

    Absil, P.-A., Mahony, R., and Sepulchre, R. Optimization Algorithms on Matrix Manifolds. Princeton University Press, 2008

    work page 2008

  3. [3]

    and Halpern, J

    Beckers, S. and Halpern, J. Y. Abstracting causal models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pp.\ 2678--2685, 2019

    work page 2019

  4. [4]

    and Gavves, S

    Bereska, L. and Gavves, S. Mechanistic interpretability for ai safety-a review. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=ePUVetPKu6

    work page 2024

  5. [5]

    Bollen, K. A. Structural equations with latent variables, volume 210. John Wiley & Sons, 1989

    work page 1989

  6. [6]

    An introduction to optimization on smooth manifolds

    Boumal, N. An introduction to optimization on smooth manifolds. Cambridge University Press, 2023

    2023

  7. [7]

    Manopt, a M atlab toolbox for optimization on manifolds

    Boumal, N., Mishra, B., Absil, P.-A., and Sepulchre, R. Manopt, a M atlab toolbox for optimization on manifolds. Journal of Machine Learning Research, 15 0 (1): 0 1455--1459, 2014

    work page 2014

  8. [8]

    and Vandenberghe, L

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

    2004

Show all 56 references
  1. [9]

    Distributed optimization and statistical learning via the alternating direction method of multipliers

    Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J., et al. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine learning , 3 0 (1): 0 1--122, 2011

    2011

  2. [10]

    The matrix reference manual

    Brookes, M. The matrix reference manual. http://www.ee.imperial.ac.uk/hp/staff/dmb/matrix/intro.html, 2020. Accessed: 2024-01-10

    2020

  3. [11]

    and Lim, L.-H

    Cai, Y. and Lim, L.-H. Distances between probability distributions of different dimensions. IEEE Transactions on Information Theory, 68 0 (6), 2022

    2022

  4. [12]

    Proximal gradient method for nonsmooth optimization over the S tiefel manifold

    Chen, S., Ma, S., Man-Cho So, A., and Zhang, T. Proximal gradient method for nonsmooth optimization over the S tiefel manifold. SIAM Journal on Optimization, 30 0 (1): 0 210--239, 2020. doi:10.1137/18M122457X. URL https://doi.org/10.1137/18M122457X

    2020 doi

  5. [13]

    D'Acunto, G., Bonchi, F., Morales, G. D. F., and Petri, G. Extracting the multiscale causal backbone of brain dynamics. In Causal Learning and Reasoning, pp.\ 265--295. PMLR, 2024

    2024

  6. [14]

    and Boyd, S

    Diamond, S. and Boyd, S. CVXPY : A python-embedded modeling language for convex optimization. Journal of Machine Learning Research, 17 0 (83): 0 1--5, 2016

    2016

  7. [15]

    G., Felekis, Y., Zennaro, F

    Dyer, J., Bishop, N. G., Felekis, Y., Zennaro, F. M., Calinescu, A., Damoulas, T., and Wooldridge, M. J. Interventionally consistent surrogates for complex simulation models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://ope...

    2024

  8. [16]

    A., and Smith, S

    Edelman, A., Arias, T. A., and Smith, S. T. The geometry of algorithms with orthogonality constraints. SIAM Journal on Matrix Analysis and Applications, 20 0 (2): 0 303--353, 1998

    1998

  9. [17]

    and Hoffman, A

    Fan, K. and Hoffman, A. J. Some metric inequalities in the space of matrices. Proceedings of the American Mathematical Society, 6 0 (1): 0 111--116, 1955

    1955

  10. [18]

    M., Branchini, N., and Damoulas, T

    Felekis, Y., Zennaro, F. M., Branchini, N., and Damoulas, T. Causal optimal transport of abstractions. In Causal Learning and Reasoning, pp.\ 462--498. PMLR, 2024

    2024

  11. [19]

    K., et al

    Ganguly, N., Fazlija, D., Badar, M., Fisichella, M., Sikdar, S., Schrader, J., Wallat, J., Rudra, K., Koubarakis, M., Patro, G. K., et al. A review of the role of causality in developing trustworthy AI systems. arXiv preprint arXiv:2302.06975, 2023

    2023 arXiv

  12. [20]

    Causal abstractions of neural networks

    Geiger, A., Lu, H., Icard, T., and Potts, C. Causal abstractions of neural networks. Advances in Neural Information Processing Systems, 34: 0 9574--9586, 2021

    2021

  13. [21]

    Finding alignments between interpretable causal variables and distributed neural representations

    Geiger, A., Wu, Z., Potts, C., Icard, T., and Goodman, N. Finding alignments between interpretable causal variables and distributed neural representations. In Causal Learning and Reasoning, pp.\ 160--187. PMLR, 2024

    2024

  14. [22]

    Higham, N. J. Computing the polar decomposition—with applications. SIAM Journal on Scientific and Statistical Computing, 7 0 (4): 0 1160--1174, 1986

    1986

  15. [23]

    Higham, N. J. and Cheng, S. H. Modifying the inertia of matrices arising in optimization. Linear Algebra and its Applications, 275: 0 261--279, 1998

    1998

  16. [24]

    Causal inference by string diagram surgery

    Jacobs, B., Kissinger, A., and Zanasi, F. Causal inference by string diagram surgery. In International Conference on Foundations of Software Science and Computation Structures, pp.\ 313--329. Springer, 2019

    2019

  17. [25]

    Targeted reduction of causal models

    Keki \'c , A., Sch \"o lkopf, B., and Besserve, M. Targeted reduction of causal models. In Uncertainty in Artificial Intelligence, pp.\ 1953--1980. PMLR, 2024

    1953

  18. [26]

    SCM - VAE : L earning identifiable causal representations via structural knowledge

    Komanduri, A., Wu, Y., Huang, W., Chen, F., and Wu, X. SCM - VAE : L earning identifiable causal representations via structural knowledge. In 2022 IEEE International Conference on Big Data (Big Data), pp.\ 1014--1023. IEEE, 2022

    2022

  19. [27]

    Kovnatsky, A., Glashoff, K., and Bronstein, M. M. MADMM : A generic algorithm for non-smooth optimization on manifolds. In Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V 14, pp.\ 680--696. Springer, 2016

    2016

  20. [28]

    and Osher, S

    Lai, R. and Osher, S. A splitting method for orthogonality constrained problems. Journal of Scientific Computing, 58: 0 431--449, 2014

    2014

  21. [29]

    D., Razaviyayn, M., and Hong, M

    Lu, S., Lee, J. D., Razaviyayn, M., and Hong, M. Linearized ADMM converges to second-order stationary points for non-convex problems. IEEE Transactions on Signal Processing, 69: 0 4859--4874, 2021

    2021

  22. [30]

    Categories for the Working Mathematician, volume 5

    Mac Lane, S. Categories for the Working Mathematician, volume 5. Springer Science & Business Media, 2013

    2013

  23. [31]

    Sampling via measure transport: A n introduction , pp.\ 1–41

    Marzouk, Y., Moselhy, T., Parno, M., and Spantini, A. Sampling via measure transport: A n introduction , pp.\ 1–41. Springer International Publishing, 2016. ISBN 9783319112596. doi:10.1007/978-3-319-11259-6_23-1. URL http://dx.doi.org/10.1007/978-3-319-11259-6_23-1

    2016 doi

  24. [32]

    Learning causal abstractions of linear structural causal models

    Massidda, R., Magliacane, S., and Bacciu, D. Learning causal abstractions of linear structural causal models. In Uncertainty in Artificial Intelligence, pp.\ 2486--2515. PMLR, 2024

    2024

  25. [33]

    Parallel and distributed successive convex approximation methods for big-data optimization

    Nedi \'c , A., Pang, J.-S., Scutari, G., Sun, Y., Scutari, G., and Sun, Y. Parallel and distributed successive convex approximation methods for big-data optimization. Multi-Agent Optimization: Cetraro, Italy 2014, pp.\ 141--308, 2018

    2014

  26. [34]

    and Saigo, H

    Otsuka, J. and Saigo, H. On the equivalence of causal models: A category-theoretic approach. In Conference on Causal Learning and Reasoning, pp.\ 634--646. PMLR, 2022

    2022

  27. [35]

    Proximal algorithms

    Parikh, N., Boyd, S., et al. Proximal algorithms. Foundations and Trends in Optimization , 1 0 (3): 0 127--239, 2014

    2014

  28. [36]

    Causality

    Pearl, J. Causality. Cambridge University Press, 2009

    2009

  29. [37]

    Starting Category Theory

    Perrone, P. Starting Category Theory. World Scientific, 2024

    2024

  30. [38]

    Causal responsibility attribution for human- AI collaboration

    Qi, Y., Sch \"o lkopf, B., and Jin, Z. Causal responsibility attribution for human- AI collaboration. arXiv preprint arXiv:2411.03275, 2024

    2024 arXiv

  31. [39]

    B., Sadler, B

    Rawal, A., Raglin, A., Rawat, D. B., Sadler, B. M., and McCoy, J. Causality for trustworthy artificial intelligence: Status, challenges and perspectives. ACM Computing Surveys, 2024

    2024

  32. [40]

    Rischel, E. F. The category theory of causal models. Master's thesis, University of Copenhagen, 2020

    2020

  33. [41]

    K., Weichwald, S., Bongers, S., Mooij, J

    Rubenstein, P. K., Weichwald, S., Bongers, S., Mooij, J. M., Janzing, D., Grosse-Wentrup, M., and Sch \"o lkopf, B. Causal consistency of structural equation models. In 33rd Conference on Uncertainty in Artificial Intelligence (UAI 2017), pp.\ 808--817. Curran Associates, Inc., 2017

    2017

  34. [42]

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

    Sch \"o lkopf, B., Locatello, F., Bauer, S., Ke, N. R., Kalchbrenner, N., Goyal, A., and Bengio, Y. Toward causal representation learning. Proceedings of the IEEE, 109 0 (5): 0 612--634, 2021

    2021

  35. [43]

    and Zennaro, F

    Schooltink, W. and Zennaro, F. M. Aligning graphical and functional causal abstractions. arXiv preprint arXiv:2412.17080, 2024

    2024 arXiv

  36. [44]

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

    Shimizu, S., Hoyer, P. O., Hyv \"a rinen, A., Kerminen, A., and Jordan, M. A linear non- G aussian acyclic model for causal discovery. Journal of Machine Learning Research, 7 0 (10), 2006

    2006

  37. [45]

    A R iemannian proximal N ewton method

    Si, W., Absil, P.-A., Huang, W., Jiang, R., and Vary, S. A R iemannian proximal N ewton method. SIAM Journal on Optimization, 34 0 (1): 0 654--681, 2024

    2024

  38. [46]

    M., Beckmann, C

    Smith, S. M., Beckmann, C. F., Andersson, J., Auerbach, E. J., Bijsterbosch, J., Douaud, G., Duff, E., Feinberg, D. A., Griffanti, L., Harms, M. P., et al. Resting-state fMRI in the H uman C onnectome P roject. NeuroImage, 80: 0 144--168, 2013

    2013

  39. [47]

    and Zhang, K

    Spirtes, P. and Zhang, K. Causal discovery and inference: C oncepts and recent methodological advances. In Applied Informatics, volume 3, pp.\ 1--28. Springer, 2016

    2016

  40. [48]

    OSQP : An operator splitting solver for quadratic programs

    Stellato, B., Banjac, G., Goulart, P., Bemporad, A., and Boyd, S. OSQP : An operator splitting solver for quadratic programs. Mathematical Programming Computation, 12 0 (4): 0 637--672, 2020. doi:10.1007/s12532-020-00179-2. URL https://doi.org/10.1007/s12532-020-00179-2

    2020 doi

  41. [49]

    C., Di Lorenzo, P., Sciancalepore, V., Aijaz, A., Kountouris, M., G \"u nd \"u z, D., Popovski, P., Sana, M., Stavrou, P

    Strinati, E. C., Di Lorenzo, P., Sciancalepore, V., Aijaz, A., Kountouris, M., G \"u nd \"u z, D., Popovski, P., Sana, M., Stavrou, P. A., Soret, B., et al. Goal-oriented and semantic communication in 6G AI -native networks: T he 6G - GOALS approach. In 2024 Joint European Con...

    2024

  42. [50]

    Thomas, C. K. and Saad, W. Neuro-symbolic causal reasoning meets signaling game for emergent semantic communications. IEEE Transactions on Wireless Communications, 23 0 (5): 0 4546--4563, 2023

    2023

  43. [51]

    Thomas, C. K. and Saad, W. Symbolic logic and category theory for reasoning-enabled semantic communications. In 2024 58th Asilomar Conference on Signals, Systems, and Computers, pp.\ 121--125. IEEE, 2024

    2024

  44. [52]

    K., Saad, W., and Xiao, Y

    Thomas, C. K., Saad, W., and Xiao, Y. Causal semantic communication for digital twins: A generalizable imitation learning approach. IEEE Journal on Selected Areas in Information Theory, 4: 0 698--717, 2023

    2023

  45. [53]

    A regularized semi-smooth N ewton method with projection steps for composite convex programs

    Xiao, X., Li, Y., Wen, Z., and Zhang, L. A regularized semi-smooth N ewton method with projection steps for composite convex programs. Journal of Scientific Computing, 76: 0 364--389, 2018

    2018

  46. [54]

    Causal VAE : Disentangled representation learning via neural structural causal models

    Yang, M., Liu, F., Chen, Z., Shen, X., Hao, J., and Wang, J. Causal VAE : Disentangled representation learning via neural structural causal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 9593--9602, 2021

    2021

  47. [55]

    M., Dr \'a vucz, M., Apachitei, G., Widanage, W

    Zennaro, F. M., Dr \'a vucz, M., Apachitei, G., Widanage, W. D., and Damoulas, T. Jointly learning consistent causal abstractions over multiple interventional distributions. In 2nd Conference on Causal Learning and Reasoning, 2023

    2023

  48. [56]

    and Hyv \"a rinen, A

    Zhang, K. and Hyv \"a rinen, A. On the identifiability of the post-nonlinear causal model. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, pp.\ 647--655, 2009

    2009

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.