arxiv: 2605.03393 · v1 · submitted 2026-05-05 · 📊 stat.ML · cs.LG

Recognition: unknown

Adaptive Estimation and Optimal Control in Offline Contextual MDPs without Stationarity

Riddhiman Bhattacharyya , Sayak Chakrabarty , Imon Banerjee

Authors on Pith no claims yet

Pith reviewed 2026-05-07 13:21 UTC · model grok-4.3

classification 📊 stat.ML cs.LG

keywords adaptive estimationcontextual MDPsoffline learningT-estimationoptimal controlnon-stationarityoracle risk bounds

0 comments

The pith

A T-estimation procedure selects an estimator for offline contextual MDPs that attains oracle risk bounds without stationarity assumptions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a method to choose an estimator from offline samples of a contextual MDP and shows it achieves oracle risk bounds under two loss functions even when the process is non-stationary or the model is irregular. It then uses that estimate to select an optimal control policy and proves finite-sample guarantees on the resulting cost. The approach relies on T-estimation to handle the endogenous features of contextual MDPs under complete generality. A reader would care because many real datasets from biostatistics and reinforcement learning come from environments whose dynamics change over time, yet most existing theory requires stationarity.

Core claim

By applying T-estimation the authors construct a procedure that, given a sample from a contextual MDP, produces an estimator whose risk is bounded by the oracle risk under two distinct loss functions; the same density estimate then yields a control whose expected cost satisfies finite-sample bounds, all without assuming stationarity or model regularity.

What carries the argument

T-estimation, a statistical technique that delivers estimators with oracle risk bounds in general non-parametric settings and is here used to select the density estimator for the contextual MDP.

If this is right

Density estimation for contextual MDPs becomes possible from offline data without stationarity.
Oracle risk bounds hold simultaneously for two different loss functions.
Optimal controls derived from the estimate come with explicit finite-sample cost guarantees.
The entire procedure works under complete generality, covering irregular models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same T-estimation route might extend to other endogenous structures such as non-stationary partially observable MDPs.
Practical testing on changing clinical or recommendation datasets could reveal how often the finite-sample cost bounds are tight.
If the bounds remain useful at moderate sample sizes, the method could replace heuristic offline RL pipelines that currently ignore non-stationarity.

Load-bearing premise

That T-estimation can be applied directly to the endogenous, non-stationary, and potentially irregular structure of contextual MDPs while preserving the stated oracle risk bounds and finite-sample cost guarantees under complete generality.

What would settle it

Generate samples from a simple non-stationary contextual MDP, apply the proposed estimator, and check whether its risk exceeds the oracle risk by more than the paper's bound or whether the derived control's cost exceeds the claimed finite-sample guarantee.

read the original abstract

Contextual MDPs are powerful tools with wide applicability in areas from biostatistics to machine learning. However, specializing them to offline datasets has been challenging due to a lack of robust, theoretically backed methods. Our work tackles this problem by introducing a new approach towards adaptive estimation and cost optimization of contextual MDPs. This estimator, to the best of our knowledge, is the first of its kind, and is endowed with strong optimality guarantees. We achieve this by overcoming the key technical challenges evolving from the endogenous properties of contextual MDPs; such as non-stationarity, or model irregularity. Our guarantees are established under complete generality by utilizing the relatively recent and powerful statistical technique of $T$-estimation (Baraud, 2011). We first provide a procedure for selecting an estimator given a sample from a contextual MDP and use it to derive oracle risk bounds under two distinct, but nevertheless meaningful, loss functions. We then consider the problem of determining the optimal control with the aid of the aforementioned density estimate and provide finite sample guarantees for the cost function.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper claims to be the first to get oracle bounds and finite-sample control guarantees for offline contextual MDPs without stationarity by importing T-estimation, but the transfer of that technique to dependent non-stationary trajectories is not shown in the abstract.

read the letter

The main thing to know is that this paper gives a procedure for adaptive density estimation from offline contextual MDP data and then uses that estimate to get finite-sample guarantees on the cost of the resulting control policy, all without assuming stationarity. It positions the work as the first of its kind by using T-estimation from Baraud 2011 to handle the endogenous non-stationarity and model irregularity that come with these processes. The two-step structure—select an estimator from the sample under two loss functions, then optimize control—is laid out clearly in outline. That addresses a real gap, since most offline RL methods avoid non-stationary data by assumption and real applications in biostatistics often have changing environments. The authors correctly identify the dependence and irregularity as the technical obstacles. The soft spot is the soundness of the claimed guarantees. T-estimation requires the model class to meet entropy conditions and the observations to behave like draws from a fixed or mixing measure. MDP trajectories are sequentially dependent, with each context-action pair shaping the next distribution, and non-stationarity means the marginals shift. The abstract asserts the bounds hold under complete generality, but it gives no derivation, no explicit reduction showing the MDP likelihood satisfies Baraud's conditions, and no tailored argument for how the oracle inequality survives the dependence. Without those steps visible, the central claims rest on an unverified transfer. This is for theorists working on statistical estimation in sequential decisions or offline RL who want to relax stationarity. A reader following T-estimation extensions or non-stationary MDPs could extract value if the details hold up. It deserves peer review so referees can check whether the application actually delivers the stated bounds or needs extra regularity. I would send it out rather than desk reject.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a novel adaptive estimator for offline contextual MDPs that does not require stationarity. It applies T-estimation (Baraud 2011) to select a density estimator from trajectory data, derives oracle risk bounds under two loss functions, and then uses the resulting estimate to obtain finite-sample guarantees on the cost of the induced optimal policy. The central claims are that this is the first such method with strong optimality guarantees and that the bounds hold under complete generality despite endogenous sampling, non-stationarity, and model irregularity.

Significance. If the transfer of T-estimation to the dependent, non-stationary MDP setting can be made rigorous, the result would be significant: it would supply the first adaptive estimator with explicit oracle inequalities and finite-sample control guarantees for offline contextual MDPs without stationarity. The approach of importing a modern statistical selection technique to handle endogenous non-stationarity is conceptually attractive and could influence subsequent work on offline RL with general function classes.

major comments (2)

[Abstract and §3 (T-estimation procedure)] The abstract and introduction assert that T-estimation applies directly to yield oracle risk bounds 'under complete generality' for endogenous, non-stationary contextual MDPs, yet no explicit reduction is given showing that the MDP likelihood satisfies the entropy or covering conditions of Baraud (2011) or that the oracle inequality survives the dependence induced by the policy and transition kernel. This justification is load-bearing for the optimality claims.
[Section on optimal control guarantees] The finite-sample cost guarantees for the optimal control (derived from the density estimate) are stated without an accompanying list of assumptions on the context distribution, transition kernels, or reward function. It is therefore unclear whether the claimed bounds hold for arbitrary non-stationary MDPs or require hidden regularity that would contradict the 'complete generality' assertion.

minor comments (2)

[Introduction] The two loss functions used for the oracle risk bounds are not named or defined until late in the manuscript; early reference to them would improve readability.
[Method section] Citation to Baraud (2011) should include the precise theorem or corollary being invoked, together with a short statement of the conditions that are being verified for the MDP case.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments on our manuscript. We appreciate the feedback on the technical foundations of our T-estimation approach and the optimal control guarantees. We address each major comment point by point below, providing clarifications and committing to revisions where appropriate to strengthen the presentation.

read point-by-point responses

Referee: [Abstract and §3 (T-estimation procedure)] The abstract and introduction assert that T-estimation applies directly to yield oracle risk bounds 'under complete generality' for endogenous, non-stationary contextual MDPs, yet no explicit reduction is given showing that the MDP likelihood satisfies the entropy or covering conditions of Baraud (2011) or that the oracle inequality survives the dependence induced by the policy and transition kernel. This justification is load-bearing for the optimality claims.

Authors: We agree that an explicit reduction to the conditions of Baraud (2011) would improve clarity and rigor. In Section 3, the T-estimation procedure is applied by constructing a contrast function from the joint likelihood of observed trajectories under the contextual MDP, where the model class consists of densities for contexts, transitions, and rewards. The entropy and covering conditions are satisfied because we assume the relevant function classes admit finite entropy integrals (standard for nonparametric density estimation), and the non-stationarity is handled by allowing time-dependent kernels without requiring identical distributions across time steps. Dependence induced by the policy and transitions is addressed by noting that the contrast forms a martingale difference sequence with respect to the natural filtration of the MDP, permitting the concentration results from Baraud (2011) to apply directly. However, we acknowledge that this mapping is not spelled out in full detail. In the revised manuscript, we will add a dedicated subsection in Section 3 that explicitly verifies the entropy integrals and martingale property for the MDP likelihood, thereby making the reduction transparent. revision: yes
Referee: [Section on optimal control guarantees] The finite-sample cost guarantees for the optimal control (derived from the density estimate) are stated without an accompanying list of assumptions on the context distribution, transition kernels, or reward function. It is therefore unclear whether the claimed bounds hold for arbitrary non-stationary MDPs or require hidden regularity that would contradict the 'complete generality' assertion.

Authors: The finite-sample cost guarantees are derived under the same general setting as the density estimation step, without imposing stationarity, parametric forms, or extra regularity beyond what is needed for the expectations and integrals to be well-defined. Specifically, the context distribution may be arbitrary and time-varying, transition kernels are general (possibly non-stationary and endogenous), and rewards are assumed bounded (to ensure finite costs), but no Lipschitz continuity, smoothness, or other regularity is required. This does not contradict 'complete generality' because the only conditions are those necessary for any MDP to have a well-posed value function; the bounds hold as long as the T-estimation oracle inequality is available. To eliminate ambiguity, we will add an explicit 'Assumptions' paragraph at the start of the optimal control section that lists these minimal conditions and reiterates that they are compatible with arbitrary non-stationary behavior. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation relies on external T-estimation result

full rationale

The paper introduces an estimator for offline contextual MDPs and derives oracle risk bounds plus finite-sample cost guarantees by directly invoking the T-estimation framework of Baraud (2011). No equations or steps reduce a claimed prediction or bound to a quantity defined by the paper's own fitted parameters, self-citations, or ansatz. The central claims rest on the applicability of an independent, externally published statistical result rather than any self-referential construction. This is the normal case of a paper building on prior work without internal circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review prevents exhaustive enumeration; the central claims rest on the applicability of T-estimation to contextual MDPs and the existence of suitable loss functions and control mappings that admit the stated bounds.

axioms (1)

domain assumption T-estimation (Baraud 2011) applies to the density estimation problem induced by a contextual MDP under non-stationarity and model irregularity
Invoked to establish oracle risk bounds under complete generality

pith-pipeline@v0.9.0 · 5486 in / 1302 out tokens · 70181 ms · 2026-05-07T13:21:29.362429+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

300 extracted references · 88 canonical work pages · 1 internal anchor

[1]

Adaptive

Banerjee, Imon and Rao, Vinayak and Honnappa, Harsha , month = may, year =. Adaptive. doi:10.48550/arXiv.2505.14458 , abstract =

work page doi:10.48550/arxiv.2505.14458
[2]

https://arxiv.org/pdf/2505.14458 , url =

work page arXiv
[3]

Technometrics , author =

Powerful. Technometrics , author =. 2006 , note =. doi:10.1198/004017005000000328 , abstract =

work page doi:10.1198/004017005000000328 2006
[4]

Journal of the Royal Statistical Society

Powerful. Journal of the Royal Statistical Society. Series B (Statistical Methodology) , author =. 2002 , note =

2002
[5]

Signal Processing , author =

Selective review of offline change point detection methods , volume =. Signal Processing , author =. 2020 , note =. doi:10.1016/j.sigpro.2019.107299 , abstract =

work page doi:10.1016/j.sigpro.2019.107299 2020
[6]

Li, Shuang and Xie, Yao and Dai, Hanjun and Song, Le , year =. M-. Advances in
[7]

Wild binary segmentation for multiple change-point detection

Wild binary segmentation for multiple change-point detection , volume =. The Annals of Statistics , author =. 2014 , note =. doi:10.1214/14-AOS1245 , abstract =

work page doi:10.1214/14-aos1245 2014
[8]

The Annals of Statistics , author =

Wild. The Annals of Statistics , author =. 2014 , note =

2014
[9]

https://www-jstor-org.turing.library.northwestern.edu/stable/pdf/44114372.pdf?refreqid=fastly-default\

work page arXiv
[10]

https://www-jstor-org.turing.library.northwestern.edu/stable/pdf/43556493.pdf?refreqid=fastly-default\

work page arXiv
[11]

Dembo, Amir and Zeitouni, Ofer , year =. Large. doi:10.1007/978-3-642-03311-7 , keywords =

work page doi:10.1007/978-3-642-03311-7
[12]

Nonlin- ear bayesian filtering with natural gradient gaussian approximation,

Cao, Wenhan and Zhang, Tianyi and Sun, Zeju and Liu, Chang and Yau, Stephen S.-T. and Li, Shengbo Eben , month = dec, year =. Nonlinear. doi:10.48550/arXiv.2410.15832 , abstract =

work page doi:10.48550/arxiv.2410.15832
[13]

Nonlinear

Liang, Xin and Jiang, Yi , month = apr, year =. Nonlinear. doi:10.48550/arXiv.2204.03485 , abstract =

work page doi:10.48550/arxiv.2204.03485
[14]

Electronic Journal of Probability , author =

A tail inequality for suprema of unbounded empirical processes with applications to. Electronic Journal of Probability , author =. 2008 , note =. doi:10.1214/EJP.v13-521 , abstract =

work page doi:10.1214/ejp.v13-521 2008
[15]

Electronic Journal of Statistics , author =

Optimal nonparametric change point analysis , volume =. Electronic Journal of Statistics , author =. 2021 , note =. doi:10.1214/21-EJS1809 , abstract =

work page doi:10.1214/21-ejs1809 2021
[16]

IEEE Transactions on Information Theory , author =

Optimal. IEEE Transactions on Information Theory , author =. 2022 , keywords =. doi:10.1109/TIT.2021.3130330 , abstract =

work page doi:10.1109/tit.2021.3130330 2022
[17]

Mutti, Mirco and Santi, Riccardo De and Restelli, Marcello , month = jun, year =. The. Proceedings of the 39th
[18]

https://proceedings.mlr.press/v162/mutti22a/mutti22a.pdf , url =
[19]

Advances in Applied Probability , author =

Integral. Advances in Applied Probability , author =. 2018 , note =

2018
[20]

how to add extension to toolbar firefox -
[21]

Advances in Neural Information Processing Systems , author =

On the. Advances in Neural Information Processing Systems , author =. 2023 , pages =

2023
[22]

Theory of Probability & Its Applications , author =

Central. Theory of Probability & Its Applications , author =. 1956 , note =. doi:10.1137/1101029 , abstract =

work page doi:10.1137/1101029 1956
[23]

Theory of Probability & Its Applications , author =

Central. Theory of Probability & Its Applications , author =. 1956 , note =. doi:10.1137/1101006 , abstract =

work page doi:10.1137/1101006 1956
[24]

The Annals of Statistics , author =

Note on the. The Annals of Statistics , author =. 1981 , note =. doi:10.1214/aos/1176345353 , abstract =

work page doi:10.1214/aos/1176345353 1981
[25]

2023 , note =

IEEE Transactions on Computational Social Systems , author =. 2023 , note =

2023
[26]

doi:10.48550/arXiv.2505.11725 , abstract =

Banerjee, Imon and Chakrabarty, Sayak , month = may, year =. doi:10.48550/arXiv.2505.11725 , abstract =

work page doi:10.48550/arxiv.2505.11725
[27]

Proceedings of the American Mathematical Society , author =

Products of. Proceedings of the American Mathematical Society , author =. 1963 , pages =. doi:10.2307/2034984 , number =

work page doi:10.2307/2034984 1963
[28]

On the Markov chain central limit theorem

On the. Probability Surveys , author =. 2004 , note =. doi:10.1214/154957804100000051 , abstract =

work page doi:10.1214/154957804100000051 2004
[29]

Markov chains and stochastic stability , publisher =

Meyn, Sean P and Tweedie, Richard L , year =. Markov chains and stochastic stability , publisher =
[30]

Cancer Discovery , author =

Mapping. Cancer Discovery , author =. 2022 , pmid =. doi:10.1158/2159-8290.CD-21-0282 , abstract =

work page doi:10.1158/2159-8290.cd-21-0282 2022
[31]

Electronic Journal of Statistics , author =

Markov chain. Electronic Journal of Statistics , author =. 2014 , note =. doi:10.1214/14-EJS957 , abstract =

work page doi:10.1214/14-ejs957 2014
[32]

Information Processing & Management , author =

Machine learning fairness notions:. Information Processing & Management , author =. 2021 , note =. doi:10.1016/j.ipm.2021.102642 , abstract =

work page doi:10.1016/j.ipm.2021.102642 2021
[33]

Logarithmic

Boucheron, Stéphane and Lugosi, Gábor and Massart, Pascal , editor =. Logarithmic. Concentration. 2013 , doi =

2013
[34]

Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete , author =

Limit theorems for the ratio of the empirical distribution function to the true distribution function , volume =. Zeitschrift für Wahrscheinlichkeitstheorie und Verwandte Gebiete , author =. 1978 , keywords =. doi:10.1007/BF00635964 , abstract =

work page doi:10.1007/bf00635964 1978
[35]

, year =

Dolgopyat, Dmitry and Sarig, Omri M. , year =. Local. doi:10.1007/978-3-031-32601-1 , keywords =

work page doi:10.1007/978-3-031-32601-1
[36]

arXiv preprint arXiv:1909.04176 , author =

Learning to learn and predict:. arXiv preprint arXiv:1909.04176 , author =

work page arXiv 1909
[37]

The annals of statistics , author =

Lasso-type recovery of sparse representations for high-dimensional data , volume =. The annals of statistics , author =. 2009 , note =

2009
[38]

Allisons.org , year =

Kullback. Allisons.org , year =
[39]

Pattern recognition , author =

Kernel. Pattern recognition , author =. 2007 , note =

2007
[40]

Athreya, Krishna B , year =. Kernel
[41]

Isoperimetry on the

Boucheron, Stéphane and Lugosi, Gábor and Massart, Pascal , editor =. Isoperimetry on the. Concentration. 2013 , doi =

2013
[42]

IEEE Photonics Journal , author =

Joint optical performance monitoring and modulation format/bit-rate identification by. IEEE Photonics Journal , author =. 2018 , note =

2018
[43]

Invexity and

Mishra, Shashi Kant and Giorgi, Giorgio , editor =. Invexity and. 2008 , doi =

2008
[44]

Introduction to the non-asymptotic analysis of random matrices , journal =

Vershynin, Roman , year =. Introduction to the non-asymptotic analysis of random matrices , journal =
[45]

Foundations and Trends in Machine Learning , author =

Introduction to. Foundations and Trends in Machine Learning , author =. 2008 , pages =

2008
[46]

and Leiserson, Charles E

Cormen, Thomas H. and Leiserson, Charles E. and Rivest, Ronald L. and Stein, Clifford , month = apr, year =. Introduction to
[47]

Introduction to

Tsybakov, Alexandre B , year =. Introduction to
[48]

Introduction to coding theory , volume =

Van Lint, Jacobus Hendricus , year =. Introduction to coding theory , volume =
[49]

Introduction , isbn =

Boucheron, Stéphane and Lugosi, Gábor and Massart, Pascal , editor =. Introduction , isbn =. Concentration. 2013 , doi =

2013
[50]

Interpolation

Bergh, Jöran and Löfström, Jörgen , editor =. Interpolation. 1976 , doi =

1976
[51]

Interpolation of

Bergh, Jöran and Löfström, Jörgen , editor =. Interpolation of. Interpolation. 1976 , doi =

1976
[52]

Machine learning , author =

Informing sequential clinical decision-making through reinforcement learning: an empirical study , volume =. Machine learning , author =. 2011 , note =

2011
[53]

Electronic Journal of Statistics , author =

Inhomogeneous and anisotropic conditional density estimation from dependent data , volume =. Electronic Journal of Statistics , author =. 2011 , note =. doi:10.1214/11-EJS653 , abstract =

work page doi:10.1214/11-ejs653 2011
[54]

Information-theoretically optimal sparse

Deshpande, Yash and Montanari, Andrea , year =. Information-theoretically optimal sparse. doi:10.1109/ISIT.2014.6875223 , booktitle =

work page doi:10.1109/isit.2014.6875223 2014
[55]

Information

Johnson, Oliver , year =. Information
[56]

Information

Akaike, Hirotogu , editor =. Information. Selected. 1998 , doi =

1998
[57]

Influences and

Boucheron, Stéphane and Lugosi, Gábor and Massart, Pascal , editor =. Influences and. Concentration. 2013 , doi =

2013
[58]

Concentration

Index , isbn =. Concentration. 2013 , pages =

2013
[59]

Complex & Intelligent Systems , author =

Improving ant colony optimization algorithm with epsilon greedy and. Complex & Intelligent Systems , author =. 2021 , note =

2021
[60]

Journal of Visual Communication and Image Representation , author =

Image classification base on. Journal of Visual Communication and Image Representation , author =. 2019 , note =

2019
[61]

Interspeech , author =

Improved. Interspeech , author =. 2018 , pages =

2018
[62]

Biometrika , author =

Ideal spatial adaptation by wavelet shrinkage , volume =. Biometrika , author =. 1994 , pages =. doi:10.1093/biomet/81.3.425 , abstract =

work page doi:10.1093/biomet/81.3.425 1994
[63]

The annals of statistics , author =

High-dimensional graphs and variable selection with the lasso , volume =. The annals of statistics , author =. 2006 , note =

2006
[64]

Hitting-time and occupation-time bounds implied by drift analysis with applications , journal =

Hajek, Bruce , year =. Hitting-time and occupation-time bounds implied by drift analysis with applications , journal =
[65]

High-dimensional probability:

Vershynin, Roman , year =. High-dimensional probability:
[66]

Ibragimov, Marat and Ibragimov, Rustam and Walden, Johan , year =. Heavy-. doi:10.1007/978-3-319-16877-7 , keywords =

work page doi:10.1007/978-3-319-16877-7
[67]

High-dimensional analysis of semidefinite relaxations for sparse principal components , booktitle =

Amini, Arash A and Wainwright, Martin J , year =. High-dimensional analysis of semidefinite relaxations for sparse principal components , booktitle =
[68]

and Vos, Paul W

Kass, Robert E. and Vos, Paul W. , month = sep, year =. Geometrical
[69]

and Smyth, Gordon K

Dunn, Peter K. and Smyth, Gordon K. , year =. Generalized. doi:10.1007/978-1-4419-0118-7 , keywords =

work page doi:10.1007/978-1-4419-0118-7
[70]

Concentration

Gaussian. Concentration. 2007 , doi =

2007
[71]

From ads to interventions:

Tewari, Ambuj and Murphy, Susan A , year =. From ads to interventions:. Mobile
[72]

Indian Journal of Pure and Applied Mathematics , author =

Fuglede’s theorem , volume =. Indian Journal of Pure and Applied Mathematics , author =. 2015 , pages =. doi:10.1007/s13226-015-0143-6 , abstract =

work page doi:10.1007/s13226-015-0143-6 2015
[73]

Journal of the American Statistical Association , author =

Frequentist consistency of variational. Journal of the American Statistical Association , author =. 2019 , note =

2019
[74]

Probability Theory and Related Fields , author =

Fisher information and the central limit theorem , volume =. Probability Theory and Related Fields , author =. 2014 , keywords =. doi:10.1007/s00440-013-0500-5 , abstract =

work page doi:10.1007/s00440-013-0500-5 2014
[75]

Concentration

Foreword , isbn =. Concentration. 2013 , doi =

2013
[76]

Proceedings of the 2020

Black, Emily and Yeom, Samuel and Fredrikson, Matt , month = jan, year =. Proceedings of the 2020. doi:10.1145/3351095.3372845 , abstract =

work page doi:10.1145/3351095.3372845 2020
[77]

Fairness through awareness , isbn =

Dwork, Cynthia and Hardt, Moritz and Pitassi, Toniann and Reingold, Omer and Zemel, Richard , month = jan, year =. Fairness through awareness , isbn =. Proceedings of the 3rd. doi:10.1145/2090236.2090255 , abstract =

work page doi:10.1145/2090236.2090255
[78]

Advances in neural information processing systems , author =

Fantope projection and selection:. Advances in neural information processing systems , author =
[79]

The Annals of Statistics , author =

Finite sample approximation results for principal component analysis:. The Annals of Statistics , author =. 2008 , note =

2008
[80]

Journal of Mathematical Analysis and Applications , author =

Exponential convergence of products of stochastic matrices , volume =. Journal of Mathematical Analysis and Applications , author =. 1977 , note =

1977

Showing first 80 references.