Post-Transfer Learning Statistical Inference in High-Dimensional Regression

Cao Huyen My; Nguyen Vu Khai Tam; Vo Nguyen Le Duy

REVIEW 4 major objections 6 minor 16 references

Post-Transfer Learning Statistical Inference in High-Dimensional Regression

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read PTL-SI computes valid p-values for features selected by transfer learning in high-dimensional regression, controlling the false positive rate at any chosen $\alpha$.

desk verdict First real attempt at selective inference for transfer-learned high-dimensional regression; the theory is standard and mostly sound, but the implementation's finite z-window and unstated full-rank assumptions mean the advertised exact FPR control has not actually been established. read the letter →

arxiv 2504.18212 v1 pith:ILTV7OWE submitted 2025-04-25 stat.ML cs.LG

Nguyen Vu Khai Tam , Cao Huyen My , Vo Nguyen Le Duy This is my paper · ORCID

classification stat.MLcs.LG

keywords transferlearninghigh-dimensionalregressionpost-selectioninferenceselectivep-valuefalsepositiverateTransFusionOracleTrans-LassoLasso

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

Transfer learning for high-dimensional regression can borrow strength from source data, but the features it selects are chosen by the data, so ordinary p-values computed as if the features were fixed beforehand are invalid. This paper introduces PTL-SI, a conditional-inference method that computes a selective p-value for every feature selected by the two-stage TransFusion estimator (and by Oracle Trans-Lasso). The paper claims these p-values are exactly valid under the selected-model null: for each selected feature, the probability of a false positive equals the chosen significance level $\alpha$ (e.g., 0.05) rather than being inflated by the selection step. A divide-and-conquer procedure makes the conditioning event tractable by characterizing, through systems of linear inequalities, the range of data perturbations that reproduce the same selected features. If the central claim holds, users of TL-HDR can attach error-controlled significance statements to individual selected features.

What carries the argument

The load-bearing object is the conditional-inference event, made concrete by a one-dimensional parametrization. For the test statistic $Z=\eta_j^{\top}Y$, the condition $Q(Y)=Q_{\mathrm{obs}}$ forces $Y$ onto the line $Y(z)=a+bz$; Lemma 2 identifies the truncation region $\mathcal{Z}$ as exactly the $z$-values for which TransFusion (with signs) returns the observed selected set. The divide-and-conquer step (Lemmas 3-5) rewrites $\mathcal{Z}$ as a union of intersections $\mathcal{Z}_u\cap\mathcal{Z}_v\cap\mathcal{Z}_t$, where the co-training active set and signs, the local-debias active set and signs, and the final coefficient active set and signs stay fixed; each such region is a system of linear inequalities obtained from Lasso KKT conditions. The selective p-value is then the two-sided tail probability of a truncated normal distribution over this region.

What would settle it

Simulate the null ($\beta_0=0$) with exact duplicate columns among candidate features so the selected design matrix can be rank-deficient; if the PTL-SI p-values are not uniform under $H_{0,j}$ or the empirical FPR at $\alpha=0.05$ departs from 0.05, the full-rank premise is shown to be load-bearing.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that valid post-selection inference is possible for transfer-learning high-dimensional regression. Conditioning on the selection event $M(Y)=M_{\mathrm{obs}}$ and on the sufficient statistic $Q(Y)=Q_{\mathrm{obs}}$ reduces the conditional data space to the line $Y(z)=a+bz$, and the truncation region $\mathcal{Z}=\{z: M(z)=M_{\mathrm{obs}}\}$ is a finite union of intervals; Lemma 1 then gives $P_{H_{0,j}}(p_j^{\mathrm{selective}}\leq\alpha)=\alpha$ for every $\alpha\in[0,1]$. The proof identifies each interval through KKT conditions of the co-training weighted Lasso and the local-debias Lasso, and the same construction is extended to Oracle Trans-Lasso. Thus each selected feature receives a p-value that is exact for testing whether its coefficient in the projection onto the selected target design is zero.

Load-bearing premise

The proof assumes that among the features selected at each stage, no column is a linear combination of the others; if collinearity makes the selected design matrix rank-deficient, the Lasso solution can be non-unique and the truncation-region description breaks down.

Editorial extensions

If this is right

A user can threshold a selected feature's PTL-SI p-value at $\alpha$ and know that, if the feature is truly null, it is declared relevant with probability exactly $\alpha$.
PTL-SI separates false positives from true positives in TL-HDR: null features receive large p-values while genuine signals receive small ones, which naive p-values fail to do.
The divide-and-conquer algorithm makes the truncation region computable by solving linear inequalities, so a valid p-value is obtained without re-running TransFusion over a continuum of datasets.
The same conditional-inference construction gives valid p-values for Oracle Trans-Lasso, not only for TransFusion.
Experiments on synthetic and real data indicate that FPR control survives non-Gaussian noise (Laplace, skew-normal, t20), while power remains higher than data splitting and Bonferroni baselines.

Reading between the lines

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

Editorial inference: the same KKT-interval machinery should transfer to any two-stage transfer estimator that is piecewise linear in $Y$; a concrete test is to run PTL-SI on a different TL-HDR estimator and check that null p-values are uniform.
Editorial inference: PTL-SI controls the per-feature false positive rate, not the error rate across the whole selected set; applying it to many selected features simultaneously still needs a multiplicity adjustment before global claims.
Editorial inference: the search is run over $[-20\sigma,20\sigma]$, so a truncation region lying almost entirely beyond those bounds would be missed; adaptively widening the interval until the computed region stabilizes would make the procedure fully exact.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

The paper is the first to give selective-inference p-values for features selected by a transfer-learning high-dimensional regression method (TransFusion, and also Oracle Trans-Lasso). That is genuinely new. The SI logic is textbook Lee et al. — condition on the selection event, reduce to a line, compute a truncated normal p-value — but the transfer-learning event is structurally harder because it involves two nested Lasso problems and a summation. The authors derive KKT-based truncation intervals for all three stages and assemble them with a divide-and-conquer scheme. The empirical FPR plots are consistent with the claim, and they ship code and run real-data experiments. Credit where due: this is a useful contribution to a niche that had no exact inference method.

Now the soft spots, in order of severity.

First, the gap between the stated lemma and the implementation. Lemma 1 promises exact uniform p-values using the exact truncation set Z = {z : M(z)=M_obs} on the whole real line. Algorithm 2 restricts to [-20σ,20σ]. The paper calls the outside mass negligible, but gives no bound; it also gives no proof that the interval enumeration covers Z. The p-values actually output by Algorithm 3 are therefore approximations, not the exact object in Lemma 1. For most Gaussian problems the error is tiny, but the claim as written is stronger than what is implemented.

Second, the algorithm can in principle stall at a Lasso breakpoint. If the active set returned at a breakpoint has a zero-length KKT interval, the right endpoint equals the current z, and the while loop doesn't advance. The paper does not analyze this case or provide an epsilon-perturbation safeguard. It's a fixable implementation detail, but it means the current code may miss intervals and condition on a smaller set.

Third — and this is the one a referee will insist on — Lemmas 3-5 require (X_Ou^T X_Ou)^{-1} and (X_Lv^T X_Lv)^{-1} to exist, i.e., full column rank of the selected design submatrices. The paper never states the general-position assumption. In the target regression, X_M has only n_T rows; if |M| > n_T, the inverse fails. In typical runs |M| is small, but the theory needs the condition on the table. It's a missing assumption, not a wrong derivation.

The central argument holds up. The missing pieces are all repairable: add the full-rank assumption, handle breakpoints, and explicitly state the finite-window approximation with an error bound. Who benefits: anyone doing inference after transfer learning in genomics or other data-scarce settings. The paper deserves a serious referee — I would not desk reject it — but I would send it back for revision before accepting. I'd bring it to reading group, mostly to argue about the implementation gap.

Referee Report

4 major / 6 minor

Summary. The paper introduces PTL-SI, a selective-inference framework for post-selection inference after transfer-learning-based high-dimensional regression, specifically for the TransFusion algorithm and its extension to Oracle Trans-Lasso. The method conditions on the selection event that the active sets and signs of the intermediate and final estimators match the observed values, reduces the conditional data space to a one-dimensional line, characterizes the truncation region through KKT-based linear inequalities, and computes a selective p-value from a truncated normal distribution. The central theoretical claim is Lemma 1, which states that under the selected-model null hypothesis the selective p-value is exactly uniform, thus controlling the per-feature false positive rate at any significance level α. Experiments on synthetic and real data compare FPR and TPR against naive, data-splitting, Bonferroni, and over-conditioned baselines, and include a robustness check under non-Gaussian noise.

Significance. If the exactness of the enumerated truncation region can be established, PTL-SI would be the first method to provide valid post-selection inference for transfer learning in high-dimensional regression, extending classical Lasso selective inference to a multi-source transfer setting. The paper's KKT-based interval characterizations cover two representative TL procedures, the experiments are reasonably broad, and a code repository is provided. These are genuine strengths. However, the central validity claim currently outruns the implementation: the exactness of Algorithm 2's truncation set and the required general-position assumptions are not proven, so the uniform-p-value guarantee is not yet established for the p-values that the algorithm actually outputs.

major comments (4)

[Section 3.3, Algorithm 2, Eq. (21)] The paper claims that Eq. (21) gives the exact truncation region Z, but no proof is provided that the union of intervals collected in A by Algorithm 2 equals Z ∩ [zmin, zmax]. The loop advances z to the computed right endpoint r_{u,v,t}; at a Lasso breakpoint the interval for the currently active set can have zero length, and nothing in the pseudocode or its description guarantees that the subsequent Lasso solve at that endpoint discovers the next interval. Since Lemma 1 is exact only for the set in Eq. (17), the p-values produced by Algorithm 3 are not shown to satisfy P(p_selective_j ≤ α) = α. The authors should either prove exact enumeration, for example by a parametric-programming or homotopy argument, or clearly label the implementation as approximate and quantify the error.
[Section 3.3 and Appendices A.3-A.5, Eqs. (A3), (A8), (A12)] The KKT characterizations invert matrices such as X_Ou^T X_Ou, (X_Lv^T X_Lv), and the corresponding matrices for the selected set M. The paper never states the general-position/full-column-rank assumption under which these inverses exist, the Lasso solution is unique, and the active set and signs are well defined. Without such an assumption, the linear-inequality descriptions of Z_u, Z_v, and Z_t in Lemmas 3-5 collapse, so the interval endpoints used by Algorithm 2 are not well defined. This assumption must be stated explicitly and its failure modes discussed.
[Section 3.3, footnote 1, and Eq. (18)] Eq. (18) is exact only when Z is the full truncation region on the real line. Restricting the search to [zmin, zmax] = [-20σ, 20σ] changes the conditioning event; the statement that the probability mass outside the interval is 'negligibly small' is not a proof of exact FPR control. If the bounded search is intended, the paper should state that the p-values are approximate and provide a bound on the deviation from uniformity; if exactness is claimed, the algorithm must be shown to cover all of Z.
[Appendix A.6, definition of ζ^{otl}_{uv}] The displayed formula for ζ^{otl}_{uv} contains ρ^{otl}_u where the analogous expression for ζ_{uv} in Appendix A.5 contains ι_u. This is not a purely notational difference, because it changes the intercept in the linear inequality characterizing Z^{otl}_t. The expression should be corrected, and the 'analogous' proof for Z^{otl}_v and Z^{otl}_t should either be written out or the exact substitutions specified, since Lemma 6 underlies the Oracle Trans-Lasso extension and the results in Fig. 8.

minor comments (6)

[Section 3.1, Eq. (13)] The event {Q(Y) = Qobs} has probability zero under continuous Y, and the proof of Lemma 1 integrates over Qobs without defining a regular conditional distribution. This is standard in the selective-inference literature, so a short justification or a citation to the measure-theoretic treatment would suffice.
[Section 5.1] The method 'No inference' is listed and appears in the FPR plots, but no description is given of how an FPR is computed when no p-values are produced; please clarify.
[Figures 3-8] The axis label '# true beta' should be 'true coefficient Γ' to match the text, and 'FNR' should be defined where first used in the captions.
[Section 1, code link] The GitHub link in the introduction contains a space ('PTL SI') and should be a properly URL-encoded repository link.
[Appendix A.3] The operators ⊘ and ◦ and the dimensions of the vectors ψ, γ, ν, κ, ω, ρ should be defined explicitly for readers not familiar with the notation.
[Section 5.3] The sentence 'Although in certain cases the p-values obtained from PTL-SI-oc are smaller than those from PTL-SI, overall, PTL-SI consistently yields smaller p-values' is confusing because smaller p-values indicate more significance; please rephrase to state which method is more powerful.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: the selective p-value is derived from the truncated-normal law and KKT inequalities, not from fitted constants or self-citation.

full rationale

The central validity claim is Lemma 1, proved in Appendix A.1 by the standard probability-integral-transform argument: under H0,j, eta_j^T Y conditional on the selection event E follows a truncated normal, so the two-sided selective p-value is uniform conditional on E, and marginalizing over E gives P(p <= alpha) = alpha. This requires the exact truncation region Z = {z : M(z) = M_obs}, whose characterization is derived in Lemmas 3-5 from the KKT conditions of the TransFusion optimization problems (3) and (5), with proofs given in Appendices A.3-A.5. No parameter is fitted to the data to produce the p-values; Sigma is assumed known or estimable from independent data, and lambda0, lambda_tilde, and a_k are TransFusion inputs, not fitted by PTL-SI. The self-citations [13,16] are invoked only as inspiration for the divide-and-conquer/parametric-programming device, while the specific inequalities for TransFusion are derived in the paper; [14] is cited as related but different work. Thus no central claim reduces by construction to its inputs, and no prediction is a renamed fit. The main caveats are correctness risks rather than circularity: Algorithm 2 enumerates Z on [-20 sigma, 20 sigma] by solver-based interval jumps without a proof that the union equals Z intersect [z_min, z_max], and Lemmas 3-5 require full-column-rank active-set matrices (X_Ou^T X_Ou)^{-1} and (X_Lv^T X_Lv)^{-1}, a condition the paper does not state. These affect the exactness of the implemented p-values but are not examples of self-referential derivation.

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

The derivation introduces no new free constants beyond the z-window bounds; the mathematical framework relies on the standard SI assumptions plus an unstated full-rank condition for the Lasso active-set designs.

free parameters (1)

z_min, z_max = -20 sigma, 20 sigma
Hand-chosen bounds for the divide-and-conquer search. The paper says probability mass outside is negligible, but using a finite window makes the computed truncation region, and hence the p-value, an approximation to the exact conditioning event in Lemma 1.

assumptions (4)

domain assumption Response Y is multivariate normal with known covariance matrices Sigma(k) (or estimable from independent data)
Equations (1)-(2) and used in Lemma 1 proof to obtain the truncated normal distribution of the test statistic.
ad hoc to paper Every Lasso active-set submatrix encountered has full column rank, so inverses in the KKT characterizations exist
Appendix A.3-A.5 solve the KKT conditions using (X_Ou^T X_Ou)^{-1} and (X_Lv^T X_Lv)^{-1}; the assumption is never stated.
domain assumption The Lasso solutions are unique and active sets/signs are well-defined along the parametrized line
Needed for the decomposition into sub-problems in Section 3.3; follows from general position for continuous designs.
ad hoc to paper The truncation region Z is contained within [-20 sigma, 20 sigma]
Algorithm 2 only searches this range; the exact validity proof does not include this restriction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Post-Transfer Learning Statistical Inference in High-Dimensional Regression." pith.science (2026). https://pith.science/paper/ILTV7OWE

@misc{pith2026250418212,
  author       = {Pith},
  title        = {Pith review of: Post-Transfer Learning Statistical Inference in High-Dimensional Regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILTV7OWE}},
  note         = {Machine review of arXiv:2504.18212}
}

abstract

Transfer learning (TL) for high-dimensional regression (HDR) is an important problem in machine learning, particularly when dealing with limited sample size in the target task. However, there currently lacks a method to quantify the statistical significance of the relationship between features and the response in TL-HDR settings. In this paper, we introduce a novel statistical inference framework for assessing the reliability of feature selection in TL-HDR, called PTL-SI (Post-TL Statistical Inference). The core contribution of PTL-SI is its ability to provide valid $p$-values to features selected in TL-HDR, thereby rigorously controlling the false positive rate (FPR) at desired significance level $\alpha$ (e.g., 0.05). Furthermore, we enhance statistical power by incorporating a strategic divide-and-conquer approach into our framework. We demonstrate the validity and effectiveness of the proposed PTL-SI through extensive experiments on both synthetic and real-world high-dimensional datasets, confirming its theoretical properties and utility in testing the reliability of feature selection in TL scenarios.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages

[1]

Journal of the Royal Statistical Society Series B: Statistical Methodology 84(1), 149–173 (2022)

Li, S., Cai, T.T., Li, H.: Transfer learning for high-dimensional linear regression: Prediction, estimation and minimax optimality. Journal of the Royal Statistical Society Series B: Statistical Methodology 84(1), 149–173 (2022)

work page 2022
[2]

In: International Conference on Artificial Intelligence and Statistics, pp

He, Z., Sun, Y., Li, R.: Transfusion: Covariate-shift robust transfer learning for high-dimensional regression. In: International Conference on Artificial Intelligence and Statistics, pp. 703–711 (2024). PMLR

work page 2024
[3]

BMC bioinformatics 12, 1–12 (2011)

Mei, S., Fei, W., Zhou, S.: Gene ontology based transfer learning for protein subcellular localization. BMC bioinformatics 12, 1–12 (2011)

work page 2011
[4]

Journal of Statistical Planning and Inference 174, 52–67 (2016)

Yang, Y., Wu, L.: Nonnegative adaptive lasso for ultra-high dimensional regres- sion models and a two-stage method applied in financial modeling. Journal of Statistical Planning and Inference 174, 52–67 (2016)

work page 2016
[5]

IEEE transactions on medical imaging 35(5), 1285–1298 (2016)

Shin, H.-C., Roth, H.R., Gao, M., Lu, L., Xu, Z., Nogues, I., Yao, J., Mol- lura, D., Summers, R.M.: Deep convolutional neural networks for computer-aided detection: Cnn architectures, dataset characteristics and transfer learning. IEEE transactions on medical imaging 35(5), 1285–1298 (2016)

work page 2016
[6]

The Annals of Statistics 44(3), 907–927 (2016)

Lee, J.D., Sun, D.L., Sun, Y., Taylor, J.E., et al.: Exact post-selection inference, with application to the lasso. The Annals of Statistics 44(3), 907–927 (2016)

work page 2016
[7]

arXiv preprint arXiv:1405.3920 (2014)

Loftus, J.R., Taylor, J.E.: A significance test for forward stepwise model selection. arXiv preprint arXiv:1405.3920 (2014)

arXiv 2014
[8]

arXiv preprint arXiv:1410.2597 (2014)

Fithian, W., Sun, D., Taylor, J.: Optimal inference after model selection. arXiv preprint arXiv:1410.2597 (2014)

arXiv 2014

Show all 16 references

[9]

Journal of the American Statistical Association 111(514), 600–620 (2016)

Tibshirani, R.J., Taylor, J., Lockhart, R., Tibshirani, R.: Exact post-selection 20 inference for sequential regression procedures. Journal of the American Statistical Association 111(514), 600–620 (2016)

2016
[10]

In: Advances in Neural Information Processing Systems, pp

Yang, F., Barber, R.F., Jain, P., Lafferty, J.: Selective inference for group-sparse linear models. In: Advances in Neural Information Processing Systems, pp. 2469– 2477 (2016)

2016
[11]

In: Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp

Suzumura, S., Nakagawa, K., Umezu, Y., Tsuda, K., Takeuchi, I.: Selective inference for sparse high-order interaction models. In: Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 3338–3347 (2017). JMLR. org

2017
[12]

In: International Conference on Machine Learning, pp

Sugiyama, K., Le Duy, V.N., Takeuchi, I.: More powerful and general selective inference for stepwise feature selection using homotopy method. In: International Conference on Machine Learning, pp. 9891–9901 (2021). PMLR

2021
[13]

The Journal of Machine Learning Research 23(1), 13544–13580 (2022)

Duy, V.N.L., Takeuchi, I.: More powerful conditional selective inference for gen- eralized lasso by parametric programming. The Journal of Machine Learning Research 23(1), 13544–13580 (2022)

2022
[14]

arXiv preprint arXiv:2410.15022 (2024)

Loi, N.T., Loc, D.T., Duy, V.N.L.: Statistical inference for feature selection after optimal transport-based domain adaptation. arXiv preprint arXiv:2410.15022 (2024)

2024 arXiv
[15]

arXiv preprint arXiv:1801.09037 (2018)

Liu, K., Markovic, J., Tibshirani, R.: More powerful post-selection inference, with application to the lasso. arXiv preprint arXiv:1801.09037 (2018)

2018 arXiv
[16]

In: International Conference on Artificial Intelligence and Statistics, pp

Duy, V.N.L., Lin, H.-T., Takeuchi, I.: Cad-da: Controllable anomaly detection after domain adaptation by statistical inference. In: International Conference on Artificial Intelligence and Statistics, pp. 1828–1836 (2024). PMLR 21 Appendix A A.1 Proof of Lemma 1 We have η⊤ j Y ...

2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Journal of the Royal Statistical Society Series B: Statistical Methodology 84(1), 149–173 (2022)

Li, S., Cai, T.T., Li, H.: Transfer learning for high-dimensional linear regression: Prediction, estimation and minimax optimality. Journal of the Royal Statistical Society Series B: Statistical Methodology 84(1), 149–173 (2022)

work page 2022

[2] [2]

In: International Conference on Artificial Intelligence and Statistics, pp

He, Z., Sun, Y., Li, R.: Transfusion: Covariate-shift robust transfer learning for high-dimensional regression. In: International Conference on Artificial Intelligence and Statistics, pp. 703–711 (2024). PMLR

work page 2024

[3] [3]

BMC bioinformatics 12, 1–12 (2011)

Mei, S., Fei, W., Zhou, S.: Gene ontology based transfer learning for protein subcellular localization. BMC bioinformatics 12, 1–12 (2011)

work page 2011

[4] [4]

Journal of Statistical Planning and Inference 174, 52–67 (2016)

Yang, Y., Wu, L.: Nonnegative adaptive lasso for ultra-high dimensional regres- sion models and a two-stage method applied in financial modeling. Journal of Statistical Planning and Inference 174, 52–67 (2016)

work page 2016

[5] [5]

IEEE transactions on medical imaging 35(5), 1285–1298 (2016)

Shin, H.-C., Roth, H.R., Gao, M., Lu, L., Xu, Z., Nogues, I., Yao, J., Mol- lura, D., Summers, R.M.: Deep convolutional neural networks for computer-aided detection: Cnn architectures, dataset characteristics and transfer learning. IEEE transactions on medical imaging 35(5), 1285–1298 (2016)

work page 2016

[6] [6]

The Annals of Statistics 44(3), 907–927 (2016)

Lee, J.D., Sun, D.L., Sun, Y., Taylor, J.E., et al.: Exact post-selection inference, with application to the lasso. The Annals of Statistics 44(3), 907–927 (2016)

work page 2016

[7] [7]

arXiv preprint arXiv:1405.3920 (2014)

Loftus, J.R., Taylor, J.E.: A significance test for forward stepwise model selection. arXiv preprint arXiv:1405.3920 (2014)

arXiv 2014

[8] [8]

arXiv preprint arXiv:1410.2597 (2014)

Fithian, W., Sun, D., Taylor, J.: Optimal inference after model selection. arXiv preprint arXiv:1410.2597 (2014)

arXiv 2014