Pith. sign in

REVIEW 3 major objections 5 minor 64 references

Achievable distributional robustness when the robust risk is only partially identified

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 When test shifts include unseen directions, anchor regression and pooled OLS are provably suboptimal, and the best predictor abstains on those directions.

desk verdict A solid theory paper that formalizes partial identifiability in robustness and shows anchor/OLS suboptimality for unseen shifts, but its quantitative minimax values hinge on a user-supplied norm bound C and one displayed limit looks wrong. read the letter →

arxiv 2502.02710 v1 pith:7WNED5GA submitted 2025-02-04 stat.ML cs.LG

Julia Kostin , Nicola Gnecco , Fanny Yang This is my paper · ORCID
classification stat.MLcs.LG MSC 62F3562J0562G35
keywords partialidentifiabilityworst-caserobustriskdistributionalrobustnessminimaxanchorregressionout-of-distributiongeneralizationadditivedistributionshiftsabstention
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

This paper studies distributional robustness when the robust risk is only partially identified, meaning the training data do not pin down the set of possible test distributions. It introduces the worst-case robust risk, defined as the supremum of the standard robust risk over all model parameters consistent with the training data, which is always well-defined even when the usual robust risk is not. For a linear model with additive shifts, the paper derives a closed form for this quantity and shows that its minimizer—the worst-case robust predictor—achieves an algorithm-independent minimax value. The central theoretical conclusion is that common robustness methods such as anchor regression and pooled OLS are strictly suboptimal whenever the test shift includes directions never seen during training, because they fail to account for the non-identifiable part of the parameter. Empirical results on gene expression data support the theory, showing that the test error of existing methods grows increasingly larger than the minimax value as the fraction of unseen environments increases.

What carries the argument

The central object is the observationally equivalent set Θeq of model parameters that induce the same training distributions, together with a norm bound C on the causal parameter (Assumption 3.2) that makes the supremum over Θeq finite. Decomposing the test-shift matrix as Mtest = γMseen + γ′RR⊤, with range(Mseen) ⊂ S and range(R) ⊂ S⊥, turns the worst-case robust risk into a sum of an identified robust risk and a non-identifiability penalty γ′(Cker + ‖R⊤β‖2)^2. Minimizing this objective produces the abstention rule on range(R) once γ′ exceeds the threshold γ′_th = (κ(Σ*_η)+1)‖RR⊤Σ^S_{η,ξ}‖ / Cker.

What would settle it

In the linear setting of Section 2.1, fix the true parameters, the training shift subspace S, and a chosen C, then compute the exact minimizer of sup_{α∈S⊥, ‖α‖≤Cker} Rrob(β; θ_α, γMseen + γ′RR⊤) by numerical optimization for γ′ above the stated threshold γ′_th; if any minimizer has ‖R⊤β‖2 > 0, the abstention claim of Theorem 3.1 is false.

Watch

Extended reading notes

Core claim

For the linear additive-shift model of Section 2.1, the worst-case robust risk equals γ′(Cker + ‖R⊤β‖2)^2 + γ(βS − β)⊤Mseen(βS − β) + R(β; P_{θ*,0}), where R spans the unseen test-shift directions and Cker bounds the non-identified part of the causal parameter. The minimax quantity is tight for γ′ ≥ γ′_th, and its minimizer sets R⊤β = 0, i.e. it abstains from prediction along the unseen shift directions; for smaller γ′ it interpolates between full prediction and abstention. Consequently, anchor regression, pooled OLS, and related invariance-based estimators have worst-case robust risk strictly larger than the minimax value whenever RR⊤(Σ*_η + γM_anchor)^{-1}Σ^S_{η,ξ} ≠ 0, proving these methods are suboptimal under partial identifiability even in the population limit.

Load-bearing premise

The results assume the practitioner knows a constant C that bounds the norm of the true causal parameter, and without this bound the worst-case robust risk is infinite; C is not identified from data, so a misspecified C changes the predicted minimax predictor and the claimed optimality gap.

Editorial extensions

If this is right

  • The minimax quantity M(Θeq, Mtest) grows linearly in the unseen-shift strength γ′, establishing an algorithm-independent limit on achievable robustness when parts of the robustness set are not identified.
  • Anchor regression and pooled OLS are provably suboptimal in the partially identifiable case, and their worst-case robust risk has a strictly larger slope in γ′ whenever confounding is aligned with the unseen shift directions.
  • For large unseen shifts, the optimal worst-case predictor abstains on the unseen span range(R); for vanishing unseen shifts it smoothly reduces to the anchor-type predictor and remains optimal in the γ′ → 0 limit.
  • In the identifiable case γ′ = 0, the framework reduces to the conventional robust risk and the advantages of invariance-based methods are recovered exactly.
  • Empirical comparisons on single-cell gene expression data show that the test loss of anchor regression, DRIG, ICP, and OLS grows increasingly suboptimal relative to the worst-case robust predictor as the proportion of unseen environments increases.

Reading between the lines

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

  • Extending beyond the paper: the abstention principle suggests that any predictor in a partially identified robustness problem should downweight or zero out components in directions where training data provide no information about shift behavior, a guideline that likely transfers to non-linear and latent-variable models.
  • The dependence on the user-supplied constant C implies that the exact minimax slopes and the abstention threshold are not fully data-determined; reported comparisons should include a sensitivity analysis over C, though the qualitative suboptimality ranking may persist.
  • The worst-case robust risk can serve as a tool for active intervention selection: by identifying which unseen directions most inflate the worst-case risk of a candidate estimator, one can choose the next intervention to collect that most reduces the minimax value.
  • A testable extension of the framework would replace the norm-ball bound C with a data-driven uncertainty set for the non-identified parameters, converting the minimax quantity into a more realistic finite-sample benchmark.
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 studies distributional robustness when the robust risk is only partially identified. In a linear model with additive shifts, it defines the worst-case robust risk as the supremum over observationally equivalent parameters, derives a closed-form expression (Theorem 3.1), lower-bounds the corresponding minimax quantity, and proves that anchor regression and pooled OLS are suboptimal when the test shift contains directions unseen during training. The theoretical results are supplemented by synthetic experiments and a single-cell gene-expression study.

Significance. The paper addresses a real gap: prior robustness analyses treat the robust risk as either identifiable or entirely unattainable, whereas partial identifiability is common in practice. The proposed framework is conceptually appealing, and the qualitative conclusion that anchor/OLS can be strictly suboptimal under unseen shifts is interesting and, as far as the proof goes, correct. The manuscript also provides an explicit finite-sample estimation procedure and a consistency result. However, the quantitative minimax claims are conditional on a user-supplied norm bound C and, moreover, one of the displayed limits (Eq. 17) is inconsistent with the proof, while the small-γ' limit is only obtained through a non-rigorous 'expected to be tight' approximation. These issues affect load-bearing parts of the contribution.

major comments (3)
  1. [Theorem 3.1, Eq. (17)] The small-γ' limit is stated as (Cker + ||RR^T Σ*_η^{-1} Σ^S_{η,ξ}||)^2, but the proof in Appendix F.2 derives the rate with (Σ*_η + γM)^{-1} in place of Σ*_η^{-1}, and Corollary 3.2 uses the γM-dependent expression. This is an internal inconsistency in a displayed theorem statement, and because this limit is used to claim that anchor/OLS are optimal as γ'→0, it is load-bearing.
  2. [Appendix F.2] The derivation of the small-γ' limit is not a proof. The text lower-bounds the non-differentiable term 2γ'Cker||R^T β|| by a scalar product, states that this is 'expected to be tight', and then derives only a lower bound on the γ'-rate. To establish the equality in Eq. (17) (or in Corollary 3.2), an upper bound or an exact calculation is needed. As written, the theorem's equality claim is unsupported.
  3. [Assumption 3.2] The constant C in Assumption 3.2 is a user-supplied bound on ||β*|| that is not identifiable from the training distributions. If the chosen C is smaller than ||β*||, the observationally equivalent set Θeq in Eq. (13) excludes the true parameter, so the worst-case robust risk in Eq. (9) is no longer an upper bound on the true robust risk; if C is too large, the predictor is unnecessarily conservative. Every quantitative output of Theorem 3.1 — Eq. (15), Eq. (16), the threshold γ'_th, and the empirical loss in Eq. (20) — depends on Cker = sqrt(C^2 − ||β_S||^2). The empirical implementation in Appendix D sets Cker = 1.0 with no estimation or sensitivity analysis. The qualitative suboptimality of anchor/OLS survives, but the claimed algorithm-independent minimax quantity is actually conditional on an arbitrary input.
minor comments (5)
  1. [Section 2.1] There is a typo in 'distrubtion shift setting'.
  2. [Section 3.1] The notation R(β; P^{X,Y}_{θ*,0}) in Eq. (15) is used without a definition; define the reference-environment risk when it is first introduced.
  3. [Corollary 3.2] The statement says c1(γ), c2(γ), c3(γ) are independent of γ', but later the text notes c2(γ)=Ω(γ) and c1(γ)≤(σ^S_ξ)^2; please clarify the exact dependence on γ and make the displayed rates consistent.
  4. [Appendix D, Algorithm 1] Step 3 asks to 'estimate' Ĉker, but Ĉker is a deterministic function of the user-chosen constant C and the estimate β̂_S; the algorithm should make this explicit and, ideally, include guidance or a sensitivity check for C.
  5. [Figure 3] The right panel shows a 'Lower bound' curve, but the text does not state whether this corresponds to Eq. (16), the small-γ' limit, or something else; please specify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the minimax values are derived from stated assumptions and are not fitted or defined in terms of the quantities they predict.

full rationale

The paper defines the worst-case robust risk as a supremum over the observationally equivalent set (Definition 2, Eq. (9)) and then computes it in closed form under Assumptions 3.1 and 3.2 (Theorem 3.1, Eq. (15)). This is a mathematical derivation from an explicit model, not a quantity fitted to data and then renamed as a prediction. The minimax quantity M(Theta_eq, Mtest) is a function of the model parameters and of the user-specified bound C from Assumption 3.2; the dependence on C is an explicit modeling assumption rather than a hidden reuse of the output. The suboptimality claims about anchor regression and pooled OLS (Corollary 3.2) compare the worst-case robust risk of those independently defined estimators with the derived minimax lower bound, so the comparison is a theorem, not an assumption dressed as a result. Appendix D notes that C is chosen by the practitioner and that beta_S is estimated using conjectured convergence of anchor/DRIG estimators, but these are nuisance-parameter choices in the empirical implementation and are not used to define the theoretical minimax. The only self-citation found is reference [46] in the related-work list, which is not load-bearing: no uniqueness theorem, no central premise, and no derivation step depends on it. The C-dependence of the abstention threshold and small-gamma-prime limit is a correctness or sensitivity concern, not a circularity, because the paper's claims are conditional on the stated assumptions. The derivation chain is therefore self-contained with respect to the framework's own definitions.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on the linear additive-shift model, the existence of a reference environment, knowledge of the test shift bound decomposition, and an externally supplied norm bound C on the true parameter. The real-world experiment adds an unvalidated linearity assumption. No new physical entities are introduced. The free parameters are the norm bound C, the seen-shift strength γ, and the test bound M_test, all chosen by the practitioner in the empirical demonstration.

free parameters (3)
  • C (causal parameter norm bound) = C_ker = 1.0 in real-data experiments
    Assumption 3.2 requires a known upper bound on ||β*||; the minimax value, abstention threshold, and the proposed estimator depend on C. In the real-data experiments C_ker is set to 1.0 without a data-driven procedure.
  • γ (strength of seen shifts) = 50 in real-data experiments
    The estimator and theory treat γ as a given shift strength; in experiments γ=50 is selected ad hoc for anchor, DRIG, and the worst-case robust predictor.
  • M_test (test shift upper bound) = I_d in real-data experiments
    The practitioner must specify the test shift bound; in the real-data experiment the identity matrix and γ=50 are used for the worst-case robust loss.
assumptions (6)
  • domain assumption Additive shift DGP: X_e = A_e + η, Y_e = β*^T X_e + ξ with A_e ⊥ (η,ξ) and invariant (η,ξ).
    Eq. (3) defines the entire data-generating process; all subsequent identifiability and minimax results rely on this linear additive structure and on first/second moments being finite.
  • domain assumption A reference environment with A_0 = 0 exists and is observed.
    Used to identify Σ*_η and the shift moments; Appendix B relaxes this but alters the expressions.
  • domain assumption Test shifts satisfy E[A_test A_test^T] ⪯ M_test for a known PSD M_test, and M_test decomposes as γM_seen + γ'RR^T with range(M_seen)⊂S, range(R)⊂S⊥.
    Eq. (4) and Assumption 3.1; the decomposition into seen/unseen directions is the crux of the partial-identifiability analysis and must be supplied externally.
  • domain assumption There exists a known C with ∥β*∥_2 ≤ C (Assumption 3.2).
    Without this bound the worst-case robust risk is infinite; C is an external input not identified from data.
  • ad hoc to paper In the real-data experiment, the linear additive-shift model approximately holds for the selected genes and environments.
    The paper applies the linear optimal predictor to gene-expression data but does not test linearity; Section 4 states the estimator is only provably optimal for the linear setting.
  • domain assumption Consistency of nuisance parameter estimators (Assumption D.1) and the conjecture that anchor regression converges to β_S on S.
    The consistency proof of the empirical worst-case predictor in Proposition 3 relies on Assumption D.1 and, for the anchor case, on a convergence result that the paper marks as conjectured in [45] (Appendix D.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Achievable distributional robustness when the robust risk is only partially identified." pith.science (2026). https://pith.science/paper/7WNED5GA

@misc{pith2026250202710,
  author       = {Pith},
  title        = {Pith review of: Achievable distributional robustness when the robust risk is only partially identified},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WNED5GA}},
  note         = {Machine review of arXiv:2502.02710}
}
read the original abstract

In safety-critical applications, machine learning models should generalize well under worst-case distribution shifts, that is, have a small robust risk. Invariance-based algorithms can provably take advantage of structural assumptions on the shifts when the training distributions are heterogeneous enough to identify the robust risk. However, in practice, such identifiability conditions are rarely satisfied -- a scenario so far underexplored in the theoretical literature. In this paper, we aim to fill the gap and propose to study the more general setting when the robust risk is only partially identifiable. In particular, we introduce the worst-case robust risk as a new measure of robustness that is always well-defined regardless of identifiability. Its minimum corresponds to an algorithm-independent (population) minimax quantity that measures the best achievable robustness under partial identifiability. While these concepts can be defined more broadly, in this paper we introduce and derive them explicitly for a linear model for concreteness of the presentation. First, we show that existing robustness methods are provably suboptimal in the partially identifiable case. We then evaluate these methods and the minimizer of the (empirical) worst-case robust risk on real-world gene expression data and find a similar trend: the test error of existing robustness methods grows increasingly suboptimal as the fraction of data from unseen environments increases, whereas accounting for partial identifiability allows for better generalization.

Figures

Figures reproduced from arXiv: 2502.02710 by the authors.

Figure 1
Figure 1. (Left) SCM with hidden confounding and (right) induced graph. The model allows for an arbitrary causal structure of the observed variables (X, Y ), as long as I − B is invertible, e.g. when the underlying graph is acyclic. The shifts across different distributions are captured via shift interventions on X. However, the model does not allow for interventions on the target variable Y or hidden confounders H. 2.2 The r… view at source ↗
Figure 2
Figure 2. Relationship between identifiability of the model parameters and identifiability of the robust risk. (a) The classical scenario where the test shift upper bound Mtest = Mseen is contained in the span of training shifts so that the robust risk is point-identified. (b) The more general scenario of this paper, where Mtest = Munseen contains new shift directions and where only a set can be identified in which the true r… view at source ↗
Figure 3
Figure 3. Worst-case robust risk of the baseline estimators βOLS, βanchor (using the "correct" γ), the worst-case robust predictor in (mean-shifted) multi-environment finite-sample experiments and theoretical population lower bound in the classical identified setting with varying shift strength γ (left) and the partially identifiable setting with fixed γ but varying γ ′ (right). The details of the experimental setting can be … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The figures show the performance of the worst-case robust predictor (Worst-case Rob.) compared to other methods as a function of perturbation strength s. Different panels correspond to the proportion of unseen shift directions at test time. For each panel and perturbat…
Figure 5
Figure 5. Figure 5: The figures illustrate the structure of the (a) training-time shifts and (b-c) test-time shifts for different perturbation [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 52 canonical work pages

  1. [26]

    Partial transportability for domain generalization

    Kasra Jalaldoust, Alexis Bellot, and Elias Bareinboim. Partial transportability for domain generalization. In Advances in Neural Information Processing Systems , 2024

    work page 2024

  2. [1]

    Robust solutions of optimization problems affected by uncertain probabilities

    Aharon Ben-Tal, Dick den Hertog, Anja De Waegenaere, Bertrand Melenberg, and Gijs Rennen. Robust solutions of optimization problems affected by uncertain probabilities. Management Science, 59(2):341–357, 2013

    2013

  3. [2]

    Learning models with uniform performance via distributionally robust optimization

    John C Duchi and Hongseok Namkoong. Learning models with uniform performance via distributionally robust optimization. The Annals of Statistics, 49(3):1378–1406, 2021

    2021

  4. [3]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In International Conference on Learning Representations , 2015

    work page 2015

  5. [4]

    Towards deep learning models resistant to adversarial attacks

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations , 2018

    2018

  6. [5]

    Domain adaptation with multiple sources

    Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation with multiple sources. In Advances in Neural Information Processing Systems , 2008

    work page 2008

  7. [6]

    Maximin effects in inhomogeneous large-scale data

    Nicolai Meinshausen and Peter Bühlmann. Maximin effects in inhomogeneous large-scale data. The Annals of Statistics, 43(4):1801–1830, 2015

    work page 2015

  8. [7]

    Hashimoto, and Percy Liang

    Shiori Sagawa*, Pang Wei Koh*, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust neural networks. In International Conference on Learning Representations , 2020

    work page 2020

Show all 64 references
  1. [8]

    Invariance, causality and robustness

    Peter Bühlmann. Invariance, causality and robustness. Statistical Science, 35(3):404–426, 2020

    2020

  2. [9]

    Causality from a distributional robustness point of view

    Nicolai Meinshausen. Causality from a distributional robustness point of view. In 2018 IEEE Data Science Workshop (DSW), pages 6–10. IEEE, 2018

    2018

  3. [10]

    Causality-oriented robustness: exploiting general additive interventions

    Xinwei Shen, Peter Bühlmann, and Armeen Taeb. Causality-oriented robustness: exploiting general additive interventions. arXiv preprint arXiv:2307.10299, 2023

    2023 arXiv

  4. [11]

    Adversarial training and robustness for multiple perturbations

    Florian Tramèr and Dan Boneh. Adversarial training and robustness for multiple perturbations. In Advances in Neural Information Processing Systems , 2019

    2019

  5. [12]

    Transfer of adversarial robustness between perturbation types

    Daniel Kang, Yi Sun, Tom Brown, Dan Hendrycks, and Jacob Steinhardt. Transfer of adversarial robustness between perturbation types. arXiv preprint arXiv:1905.01034, 2019. 10

    1905 arXiv

  6. [13]

    Causal inference by using invariant prediction: identifica- tion and confidence intervals

    Jonas Peters, Peter Bühlmann, and Nicolai Meinshausen. Causal inference by using invariant prediction: identifica- tion and confidence intervals. Journal of the Royal Statistical Society Series B: Statistical Methodology , 78(5):947–1012, 2016

    2016

  7. [14]

    Invariant models for causal transfer learning

    Mateo Rojas-Carulla, Bernhard Schölkopf, Richard Turner, and Jonas Peters. Invariant models for causal transfer learning. The Journal of Machine Learning Research , 19(1):1309–1342, 2018

    2018

  8. [15]

    Invariant risk minimization

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019

    1907 arXiv

  9. [16]

    Out-of-distribution generalization via risk extrapolation (REx)

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (REx). In International Conference on Machine Learning , pages 5815–5826. PMLR, 2021

    2021

  10. [17]

    Varshney

    Kartik Ahuja, Jun Wang, Amit Dhurandhar, Karthikeyan Shanmugam, and Kush R. Varshney. Empirical or invariant risk minimization? A sample complexity perspective. In International Conference on Learning Representations , 2021

    2021

  11. [18]

    Invariant risk minimization games

    Kartik Ahuja, Karthikeyan Shanmugam, Kush Varshney, and Amit Dhurandhar. Invariant risk minimization games. In International Conference on Machine Learning , pages 145–155. PMLR, 2020

    2020

  12. [19]

    Does invariant risk minimization capture invariance? In International Conference on Artificial Intelligence and Statistics , pages 4069–4077

    Pritish Kamath, Akilesh Tangella, Danica Sutherland, and Nathan Srebro. Does invariant risk minimization capture invariance? In International Conference on Artificial Intelligence and Statistics , pages 4069–4077. PMLR, 2021

    2021

  13. [20]

    The risks of invariant risk minimization

    Elan Rosenfeld, Pradeep Kumar Ravikumar, and Andrej Risteski. The risks of invariant risk minimization. In International Conference on Learning Representations , 2021

    2021

  14. [21]

    Anchor regression: Heterogeneous data meet causality

    Dominik Rothenhäusler, Nicolai Meinshausen, Peter Bühlmann, and Jonas Peters. Anchor regression: Heterogeneous data meet causality. Journal of the Royal Statistical Society Series B: Statistical Methodology , 83(2):215–246, 2021

    2021

  15. [22]

    Partial identification in econometrics

    Elie Tamer. Partial identification in econometrics. Annu. Rev. Econ., 2(1):167–195, 2010

    2010

  16. [23]

    From perfect to practical: Partial identification methods for causal inference in strategic management research

    Justin Frake, Anthony Gibbs, Brent Goldfarb, Takuya Hiraiwa, Evan Starr, and Shotaro Yamaguchi. From perfect to practical: Partial identification methods for causal inference in strategic management research. A vailable at SSRN 4228655, 2023

    2023

  17. [24]

    Assouad, Fano, and le Cam

    Bin Yu. Assouad, Fano, and le Cam. In Festschrift for Lucien Le Cam: Research papers in probability and statistics , pages 423–435. Springer, 1997

    1997

  18. [25]

    Exploiting independent instruments: Identification and distribution generalization

    Sorawit Saengkyongam, Leonard Henckel, Niklas Pfister, and Jonas Peters. Exploiting independent instruments: Identification and distribution generalization. In International Conference on Machine Learning , pages 18935–18958. PMLR, 2022

    2022

  19. [27]

    Improving predictive inference under covariate shift by weighting the log-likelihood function

    Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference , 90(2):227–244, 2000

    2000

  20. [28]

    Identification, pages 65–77

    Jean-Marie Dufour and Cheng Hsiao. Identification, pages 65–77. Palgrave Macmillan UK, London, 2010

    2010

  21. [29]

    Advanced Econometrics

    Takeshi Amemiya. Advanced Econometrics. Harvard University Press, 1985

    1985

  22. [30]

    Instrumental variables

    Roger J Bowden and Darrell A Turkington. Instrumental variables. Number 8. Cambridge University Press, 1990

    1990

  23. [31]

    Domain adaptation under structural causal models

    Yuansi Chen and Peter Bühlmann. Domain adaptation under structural causal models. Journal of Machine Learning Research, 22(261):1–80, 2021

    2021

  24. [32]

    Mapping information-rich genotype- phenotype landscapes with genome-scale Perturb-seq

    Joseph M Replogle, Reuben A Saunders, Angela N Pogson, Jeffrey A Hussmann, Alexander Lenail, Alina Guna, Lauren Mascibroda, Eric J Wagner, Karen Adelman, Gila Lithwick-Yanai, et al. Mapping information-rich genotype- phenotype landscapes with genome-scale Perturb-seq. Cell, 18...

    2022

  25. [33]

    Assessing the overall and partial causal well-specification of nonlinear additive noise models

    Christoph Schultheiss and Peter Bühlmann. Assessing the overall and partial causal well-specification of nonlinear additive noise models. Journal of Machine Learning Research , 25(159):1–41, 2024. 11

    2024

  26. [34]

    Evaluating robustness to dataset shift via parametric robustness sets

    Nikolaj Thams, Michael Oberst, and David Sontag. Evaluating robustness to dataset shift via parametric robustness sets. Advances in Neural Information Processing Systems , 2022

    2022

  27. [35]

    Learning linear causal representations from interventions under general nonlinear mixing

    Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Schölkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 2024

    2024

  28. [36]

    Active learning for optimal intervention design in causal models

    Jiaqi Zhang, Louis Cammarata, Chandler Squires, Themistoklis P Sapsis, and Caroline Uhler. Active learning for optimal intervention design in causal models. Nature Machine Intelligence, 5(10):1066–1075, 2023

    2023

  29. [37]

    Active invariant causal prediction: Experiment selection through stability

    Juan L Gamella and Christina Heinze-Deml. Active invariant causal prediction: Experiment selection through stability. Advances in Neural Information Processing Systems , 2020

    2020

  30. [38]

    Certifiable distributional robustness with principled adversarial training

    Aman Sinha, Hongseok Namkoong, and John Duchi. Certifiable distributional robustness with principled adversarial training. In International Conference on Learning Representations , 2018

    2018

  31. [39]

    Data-driven distributionally robust optimization using the Wasser- stein metric: Performance guarantees and tractable reformulations

    Peyman Mohajerin Esfahani and Daniel Kuhn. Data-driven distributionally robust optimization using the Wasser- stein metric: Performance guarantees and tractable reformulations. Mathematical Programming, 171(1):115–166, 2018

    2018

  32. [40]

    Environment invariant linear least squares

    Jianqing Fan, Cong Fang, Yihong Gu, and Tong Zhang. Environment invariant linear least squares. arXiv preprint arXiv:2303.03092, 2023

    2023 arXiv

  33. [41]

    Domain adaptation by using causal inference to predict invariant conditional distributions

    Sara Magliacane, Thijs Van Ommen, Tom Claassen, Stephan Bongers, Philip Versteeg, and Joris M Mooij. Domain adaptation by using causal inference to predict invariant conditional distributions. Advances in Neural Information Processing Systems, 2018

    2018

  34. [42]

    Gradient matching for domain generalization

    Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. In International Conference on Learning Representations , 2022

    2022

  35. [43]

    Risk variance penalization

    Chuanlong Xie, Haotian Ye, Fei Chen, Yue Liu, Rui Sun, and Zhenguo Li. Risk variance penalization. arXiv preprint arXiv:2006.07544, 2020

    2006 arXiv

  36. [44]

    Invariance principle meets information bottleneck for out-of-distribution generalization

    Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out-of-distribution generalization. Advances in Neural Information Processing Systems , 2021

    2021

  37. [45]

    Distributional robustness of K-class estimators and the PULSE

    Martin Emil Jakobsen and Jonas Peters. Distributional robustness of K-class estimators and the PULSE. The Econometrics Journal, 25(2):404–432, 2022

    2022

  38. [46]

    A causal framework for distribution generalization

    Rune Christiansen, Niklas Pfister, Martin Emil Jakobsen, Nicola Gnecco, and Jonas Peters. A causal framework for distribution generalization. IEEE Transactions on Pattern Analysis and Machine Intelligence , 44(10):6614–6630, 2021

    2021

  39. [47]

    Distributional anchor regression

    Lucas Kook, Beate Sick, and Peter Bühlmann. Distributional anchor regression. Statistics and Computing, 32(3):39, 2022

    2022

  40. [48]

    Incorporating unlabeled data into distribu- tionally robust learning

    Charlie Frogner, Sebastian Claici, Edward Chien, and Justin Solomon. Incorporating unlabeled data into distribu- tionally robust learning. Journal of Machine Learning Research , 22(56):1–46, 2021

    2021

  41. [49]

    Distributionally robust optimization with data geometry

    Jiashuo Liu, Jiayun Wu, Bo Li, and Peng Cui. Distributionally robust optimization with data geometry. Advances in Neural Information Processing Systems , 2022

    2022

  42. [50]

    Causality: Models, Reasoning and Inference

    Judea Pearl. Causality: Models, Reasoning and Inference . Cambridge University Press, USA, 2nd edition, 2009

    2009

  43. [51]

    Identification of causal effects using instrumental variables

    Joshua D Angrist, Guido W Imbens, and Donald B Rubin. Identification of causal effects using instrumental variables. Journal of the American Statistical Association , 91(434):444–455, 1996

    1996

  44. [52]

    Deep IV: A flexible approach for counterfactual prediction

    Jason Hartford, Greg Lewis, Kevin Leyton-Brown, and Matt Taddy. Deep IV: A flexible approach for counterfactual prediction. In International Conference on Machine Learning , pages 1414–1423. PMLR, 2017

    2017

  45. [53]

    Kernel instrumental variable regression

    Rahul Singh, Maneesh Sahani, and Arthur Gretton. Kernel instrumental variable regression. Advances in Neural Information Processing Systems, 2019

    2019

  46. [54]

    Deep generalized method of moments for instrumental variable analysis

    Andrew Bennett, Nathan Kallus, and Tobias Schnabel. Deep generalized method of moments for instrumental variable analysis. Advances in Neural Information Processing Systems , 2019. 12

    2019

  47. [55]

    Dual instrumental variable regression

    Krikamol Muandet, Arash Mehrjou, Si Kai Lee, and Anant Raj. Dual instrumental variable regression. Advances in Neural Information Processing Systems , 2020

    2020

  48. [56]

    In search of lost domain generalization

    Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021

    2021

  49. [57]

    Elements of causal inference: foundations and learning algorithms

    Jonas Peters, Dominik Janzing, and Bernhard Schölkopf. Elements of causal inference: foundations and learning algorithms. 2017

    2017

  50. [58]

    A useful variant of the Davis–Kahan theorem for statisticians

    Yi Yu, Tengyao Wang, and Richard J Samworth. A useful variant of the Davis–Kahan theorem for statisticians. Biometrika, 102(2):315–323, 2015

    2015

  51. [59]

    Asymptotic statistics, volume 3

    Aad W Van der Vaart. Asymptotic statistics, volume 3. Cambridge University Press, 2000

    2000

  52. [60]

    Causalbench: A large-scale benchmark for network inference from single-cell perturbation data

    Mathieu Chevalley, Yusuf Roohani, Arash Mehrjou, Jure Leskovec, and Patrick Schwab. Causalbench: A large-scale benchmark for network inference from single-cell perturbation data. arXiv preprint arXiv:2210.17283, 2022

    2022 arXiv

  53. [61]

    Repurposing CRISPR as an RNA-guided platform for sequence-specific control of gene expression

    Lei S Qi, Matthew H Larson, Luke A Gilbert, Jennifer A Doudna, Jonathan S Weissman, Adam P Arkin, and Wendell A Lim. Repurposing CRISPR as an RNA-guided platform for sequence-specific control of gene expression. Cell, 152(5):1173–1183, 2013. 13 Table 1: Comparison of various d...

    2013

  54. [62]

    dominated reference environment

    We define the test shift upper bound Mtest = γMseen + γ′RR⊤, where range(Mseen) ⊂ S and range(R) ⊂ S⊥. Again, we can decompose the parameter β⋆ as β⋆ = βS + βS⊥ . The projection βS of the causal parameter onto the relative training shifts induces the following observationally ...

  55. [63]

    Under the assumption of the consistency of the nuisance parameter estimators, we can now show that (21) is a consistent estimator of the worst-case robust predictor

    and Appendix H.3 therein) which implies that ˆβS := ΠˆSβ∞ anchor consistently estimates βS = ΠSβ⋆. Under the assumption of the consistency of the nuisance parameter estimators, we can now show that (21) is a consistent estimator of the worst-case robust predictor. Proposition ...

  56. [64]

    initial guess

    (29) Fix δ >0. Then, by (29), for all β ∈ B such that ∥β − β0∥2 > δit holds that L(β, φ0) ≥ L(β0, φ0) + mδ2 2 > L(β0, φ0). Since the inequality holds for all β ∈ B such that ∥β − β0∥2 > δ, we conclude that inf{L(β, φ0) : ∥β − β0∥2 > δ} > L(β0, φ0). Since δ >0 was arbitrary, th...

Pith tools

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