Pith. sign in

REVIEW 4 major objections 5 minor 55 references

A Gaussian DP mechanism can keep mandated aggregates exactly fixed and still satisfy (epsilon, delta)-differential privacy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 04:42 UTC pith:OJNJPNM6

load-bearing objection A sound but under-instantiated reduction of Gaussian DP to invariant null spaces; the max-entropy correlation sampling is the real contribution, while the multi-constraint privacy guarantee remains conditional on an uncomputed sensitivity bound. the 4 major comments →

arxiv 2607.22450 v1 pith:OJNJPNM6 submitted 2026-07-24 cs.CR stat.ME

A Maximum Entropy Implementation of Differential Privacy Under Linear Invariants

Ryan Lafferty , Anindya Roy This is my paper
classification cs.CR stat.ME
keywords differential privacyaggregation invariantscorrelated Gaussian noisesingular covariancecorrelation matricesentropy maximizationscale mixturesconvex feasibility
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper develops a way to run differential privacy when some aggregate quantities, like state population totals in the census, must stay exactly unchanged. Instead of adding independent noise and then post-processing, it adds correlated Gaussian noise whose covariance is singular along the constraint directions, so the constraints hold with probability one. It proves that this 'invariant Gaussian mechanism' still satisfies (epsilon, delta)-DP on the universe of databases that honor the invariants, provided the noise variance is set from a sensitivity-like constant a_M. It also gives algorithms for finding the entropy-maximizing correlation matrix that satisfies the constraints, plus new characterizations of when such matrices exist.

Core claim

The paper's central claim is that differential privacy and exact linear invariants are compatible without post-processing, provided the noise is allowed to be correlated and singular. Formally, Theorem 1 states that on the invariant data universe D_{Q,A,c}, the mechanism M(Q(D)) = Q(D) + X with X ~ N(0, sigma^2 R) and R A^T = 0 satisfies P(M(Q(D)) in S) <= e^epsilon P(M(Q(D')) in S) + delta whenever sigma > t^{-1}_{a_M,epsilon}(kappa^{-1}(delta)), where a_M = sup{ v^T (M R M^T)^{-1} v : v = M(Q(D)-Q(D')), D ~ D' } and M is a semi-orthogonal matrix whose rows span the null space of A. The singular covariance makes the invariant hold almost surely: Var(A X) = sigma^2 A R A^T = 0. The proof tra

What carries the argument

The central object is a singular correlation matrix R satisfying R B = 0 (equivalently, placing the noise in the null space of the invariant constraints), so that the aggregation constraints hold with probability one. Around it revolve three tools: the reduced-sensitivity constant a_M, a substitute for the classic L2 sensitivity that controls the privacy-loss distribution; cyclic projection onto three convex sets (positive semidefinite matrices, unit-diagonal matrices, and matrices annihilating the constraint columns) to find a feasible R; and a log-det penalty -tau log det(K^T R K) that selects the max-entropy solution among feasible correlation matrices. For the most common case of a singl

Load-bearing premise

The privacy guarantee collapses if a_M cannot be bounded or computed for the constraint matrix in use, or if the correlation matrix returned by the projection algorithm leaves R B measurably different from zero.

What would settle it

For a fixed constraint matrix A and a chosen sigma, simulate many neighboring invariant database pairs and compute the privacy loss L = v^T Sigma^{-1} X + 0.5 v^T Sigma^{-1} v; if the empirical P(L > epsilon) exceeds delta at the theorem's threshold, the theorem's condition is violated. Alternatively, inspect the residual ||R B|| after Algorithm 1 or 2 — a nonzero residual means the invariant is not preserved with probability one.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If Theorem 1 is correct, a disclosure-avoidance system can add noise at block level while state and county totals stay exactly as true, using a single correlated draw instead of a projection post-processing step.
  • The result gives a formal (epsilon, delta) privacy statement on the invariant universe, addressing the open question of how DP semantics change when some queries are excluded from the privacy budget.
  • For disjoint invariants (one per county or state), the optimal correlation matrix is explicit and no numerical optimization is needed.
  • Because Laplace noise is a scale mixture of Gaussians, the same sampling scheme yields constrained Laplace noise, extending the approach to both standard DP noise families.
  • Nested hierarchies can be protected with a single block-level noise vector whose covariance is large enough to meet privacy budgets at both block and aggregated county levels.

Where Pith is reading between the lines

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

  • The theorem's practical use requires a computable a_M; the paper gives a bound only for the block-diagonal single-constraint case, so deploying the general Algorithm 2 would need an additional sensitivity analysis before formal privacy parameters can be set.
  • Maximizing det(K^T R K) over feasible correlation matrices is plausibly equivalent to maximizing downstream query utility, since it minimizes entropy loss on the reduced data; an empirical comparison against simple post-projection noise would be a natural test.
  • The same singular-covariance construction transfers to domains like smart-meter or wearable-device data where linear consistency constraints are mandatory, though nonlinear or inequality constraints such as nonnegativity remain open extensions.
  • The max-rank/relative-interior result suggests that any feasible correlation matrix can be perturbed slightly to a max-entropy feasible solution, so even a poor initialization of the correlation search can be locally improved.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes an invariant Gaussian differential privacy mechanism for linear aggregation constraints. For a constraint matrix A, the noise X is drawn as N(0,Σ) with ΣA^T=0, so that AX=0 almost surely. Theorem 1 claims (ε,δ)-DP on the invariant data universe, with noise calibrated through the quantity a_M = sup v^T(M R M^T)^{-1}v over invariant neighbors, where M spans ker(A); the proof reduces the singular problem to the standard multivariate Gaussian mechanism. To construct the correlation matrix R, the paper proposes POCS (Algorithm 1) and an entropy-penalized projected-gradient method (Algorithm 2), and it provides results on existence and maximal rank of correlation matrices with prescribed null vectors (Theorems 2–3). Section 5 specializes to the single-constraint/block-diagonal case and gives a synthetic census-like illustration.

Significance. The coordinate-reduction idea is natural and the argument in the proof of Theorem 1 is plausible; if a_M were computable and the algorithms returned an exactly feasible R, the mechanism would give standard DP semantics under invariants without relying on post-processing. The correlation-matrix results are also of independent interest. However, as it stands, the central theorem is conditional on an unavailable sensitivity quantity and on exact feasibility that the numerical algorithms do not deliver. These are substantive, fixable gaps rather than a refutation of the underlying idea.

major comments (4)
  1. [§2.2, Theorem 1; §5.1] The theorem's noise calibration is σ > t^{-1}_{a_M,ε}(κ^{-1}(δ)). The only place where a_M is actually evaluated is the block-diagonal single-constraint case in §5.1. For the general multi-constraint setting targeted by Algorithms 1 and 2, the paper supplies neither a formula, an upper bound, nor a finiteness argument for a_M. Since a_M is the invariant-universe analogue of global sensitivity, Theorem 1 cannot be instantiated outside the worked special case; the formal guarantee is conditional in exactly the settings the algorithms are designed for.
  2. [§3.2–3.3, Algorithms 1–2] Theorem 1 assumes ΣA^T=0, so that the released X satisfies AX=0 almost surely and the privacy-loss calculation can be reduced to MX. Algorithm 1 stops at a tolerance TOL and then projects to the nearest correlation matrix; Algorithm 2 minimizes a penalized objective with trade-off τ. Neither step is shown to preserve RB=0 exactly. If RB≠0, then Var(AX)>0 and the invariant is violated with positive probability; the proof's reduction, which relies on X∈ker(A) almost surely, does not apply to the actual output. The abstract's 'probability one' claim is therefore not supported by the algorithmic implementations. A hard projection step or an explicit accounting of approximate feasibility in ε and δ is needed.
  3. [§3.2, §3.3, Algorithm pseudocode] There is a dimensional inconsistency in the definition of B. Section 3.2 defines B=DA^T (an n×k matrix) and uses the constraint RB=0. Algorithms 1 and 2 write B=AD (a k×n matrix), which makes RB undefined for the intended meaning and changes the objective from Tr(B^T R B) to an n×n trace. Algorithm 2 also obtains K from the SVD of A, whereas for B=DA^T the required K is a basis for the null space of B. As written, the algorithms cannot be executed without further clarification.
  4. [§5.1] The computation of a_M in the block case is presented with an inequality in the wrong direction for a usable bound. The text reads 'a_M ≥ Σ_i ...' and then concludes 'a_M = Σ_i sup ...'. If the displayed inequality is a lower bound, it cannot be used to choose σ; if it is meant to be an upper bound, it must be '≤' and the equality with a sum of per-block suprema requires the additional assumption that each neighboring pair changes only one block. Please correct and justify.
minor comments (5)
  1. [§2.2, Theorem 1 statement] The displayed definition of t_{a,ε}(σ) omits σ in the first term; it should be εσ√a − √a/(2σ), as used in the proof.
  2. [§3.2, Algorithm 1 return statement] The return line 'N(0, D R^1 D)' should presumably read 'N(0, D R D)'.
  3. [§2.1] The neighboring relation is ambiguous: 'differ by a single entry of D being swapped with a single entry of D′' is not a standard neighboring definition, and the state-level example can be misread. Please specify the exact swap operation and how it preserves the invariant.
  4. [§5.2] The text references Figure 2 for convergence metrics, but no Figure 2 appears in the manuscript.
  5. [Proof of Theorem 1] The proof uses 'max' for a_0 while the theorem uses sup; please state conditions under which the supremum is finite and attained, or adjust the notation consistently.

Circularity Check

0 steps flagged

No significant circularity: Theorem 1 is a standard Gaussian-mechanism reduction and no fitted parameter is relabeled as a prediction.

full rationale

I found no load-bearing circular step. Theorem 1 is a reduction of the invariant constrained Gaussian mechanism to the classical Gaussian mechanism after the linear change of variables M Q(D), with a_M defined as a sensitivity supremum over neighboring invariant databases. That sensitivity quantity is a premise of the guarantee, not a quantity fitted from data and then presented as a prediction; in the block-diagonal case it is computed analytically in Section 5.1. No equation in the paper is equivalent to its own input by construction. Algorithms 1 and 2 are numerical heuristics whose approximate satisfaction of RB=0 is a correctness/instantiation concern, not circularity. The references to Bailie et al. [2026] are contextual rather than load-bearing, and the mathematical tools cited (Dwork-Roth, Laurent-Poljak, Barrett-Pierce) are external. The paper does contain gaps and typos — the displayed t_{a,epsilon}(sigma) in Theorem 1 omits a sigma used in the proof, and no general method is given for computing a_M outside the block-diagonal case — but these affect completeness and rigor, not circularity. Since the derivation does not assume its conclusion and does not rename fitted inputs as predictions, the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The mechanism introduces no new physical or conceptual entities; its novel degrees of freedom are algorithmic tuning parameters (tau, step sizes, tolerances) and the unproven premise that the optimization loop returns an exactly feasible correlation matrix. The most consequential uncharged input is the computability of a_M (see red flags), which is a property of the problem rather than a fitted parameter.

free parameters (4)
  • tau (entropy penalty) = not specified ('typically small')
    Weight on -log det(K^T R K) in Eq. (6); controls trade-off between constraint feasibility and max entropy; no selection rule or sensitivity analysis is given.
  • alpha_t (projected gradient step sizes) = fixed alpha, value unspecified
    Algorithm 2 iterates R^{t+1}=P_{C1 cap C2}(R^t - alpha_t grad f(R^t)); convergence and final R depend on step size, but no value or line search is provided.
  • sigma_block = 46.041
    Noise scale in the Section 5.1 simulation; chosen to satisfy a max of two sigma_min bounds. Illustrative, not a method parameter.
  • TOL, MAX_CYCLE = unspecified
    Termination criteria in Algorithm 1; affect whether the returned R satisfies RB=0 exactly or only approximately.
axioms (6)
  • standard math Dwork-Roth Gaussian mechanism privacy-loss bound and Mill's inequality for the tail bound
    Used in the proof of Theorem 1 to convert a Gaussian privacy-loss bound into an (epsilon, delta) guarantee.
  • domain assumption Queries are linear and constraints are A Q(D)=c with A of full row rank
    The entire framework defines the invariant universe and the null-space requirement AX=0.
  • domain assumption Neighboring databases in the invariant universe differ by a single swapped entry that preserves the invariants, and DP is defined only over this restricted universe
    This is what makes the reduced sensitivity a_M finite; it is a modeling choice, not derived.
  • domain assumption Marginal noise densities are in the scale-mixture-of-normals class C_M
    Used to justify generating scales sigma_i and then X~N(0,D R D) so that requirement [R1] holds; excludes other marginals.
  • ad hoc to paper The algorithm returns an R that exactly satisfies RB=0 (or the theorem's Sigma with Sigma A^T=0 exists)
    Theorem 1 assumes exactness; Algorithms 1-2 use finite POCS/projected gradient plus a nearest-correlation projection, which only approximate the feasible set.
  • ad hoc to paper Arendarczyk et al. Proposition 1 tail behavior for Pareto risks
    Proposition 2's exponential bound is asserted from this citation without derivation; the shape-parameter-free bound is dubious for heavy-tailed entries.

pith-pipeline@v1.3.0-alltime-deepseek · 17078 in / 21434 out tokens · 239776 ms · 2026-08-01T04:42:50.499776+00:00 · methodology

0 comments
read the original abstract

Differential privacy is the standard for ensuring data privacy and is widely used in major data publications, including reporting results from the U.S. decennial census. Common implementation of differential privacy uses independent Gaussian or Laplace noise addition to the database. However, there could be aggregate (linear) queries to the database that are excluded from the privacy budget, for example, state totals that can not be perturbed due to constitutional mandates. Any implementation of a differential privacy is required to honor these constraints, also referred to as invariants. Under aggregation constraints, the noise vector is no longer independent and the traditional differential privacy guarantees have to be re-evaluated. We propose a high entropy differential privacy implementation that maintains the aggregation invariants with probability one or exponentially close to one and derive the privacy guarantees for the implementation under the invariants. The theoretical proof covers a partial solution to an open question about the null space of correlation matrices. Moreover, the methodology has general use in the context of sampling from normal mixture models under linear equality constraints.

Figures

Figures reproduced from arXiv: 2607.22450 by Anindya Roy, Ryan Lafferty.

Figure 1
Figure 1. Figure 1: Left: A Q-Q Plot shows agreement with the desired marginal distribution; Right: Abso [PITH_FULL_IMAGE:figures/full_fig_p020_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Algorithm 1 Convergence on simulated data. The blue color indicates the penalty term [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 2 linked inside Pith

  1. [1]

    Dwork and F

    C. Dwork and F. McSherry and K. Nissim and A. Smith , title =. Journal of Privacy and Confidentiality , volume =

  2. [2]

    Foundations and Trends in Theoretical Computer Science , volume=

    The Algorithmic Foundations of Differential Privacy , author=. Foundations and Trends in Theoretical Computer Science , volume=. 2014 , publisher=. doi:10.1561/0400000042 , url=

    doi:10.1561/0400000042 2014

  3. [3]

    Bunch and James W

    James R. Bunch and James W. Demmel and Charles F. Van Loan , title =. SIAM J. Matrix Anal. Appl. , volume =

  4. [4]

    Alternating Projection Methods , publisher =

    Ren\'. Alternating Projection Methods , publisher =

  5. [5]

    IMA Journal of Numerical Analysis , volume =

    Computing the nearest correlation matrix- a problem from finance , author =. IMA Journal of Numerical Analysis , volume =

  6. [6]

    Higham , title =

    Nicholas J. Higham , title =. Applications of Matrix Theory , year =

  7. [7]

    Higham , title =

    Nicholas J. Higham , title =

  8. [8]

    Higham and Natasa Strabi\'

    Nicholas J. Higham and Natasa Strabi\'. Anderson acceleration of the alternating projections method for computing the nearest correlation matrix , journal =

  9. [9]

    Delorme and S

    C. Delorme and S. Poljak , title =. European J. Combin. , year = 1993, pages =

    1993

  10. [10]

    Null spaces of correlation matrices , journal =

    Wayne Barrett and Stephen Pierc\'. Null spaces of correlation matrices , journal =

  11. [11]

    Goldstein , year = 1959, title =

    W Cheney and A. Goldstein , year = 1959, title =. Proceedings of the American Mathematical Society , pages =

    1959

  12. [12]

    1995 , author =

    On a positive semidefinite relaxation of the cut polytope , journal =. 1995 , author =

    1995

  13. [13]

    On the Facial Structure of the Set of Correlation Matrices , author=. SIAM J. Matrix Anal. Appl. , year=

  14. [14]

    USSR Computational Mathematics and Mathematical Physics , volume=

    The method of projections for finding the common point of convex sets , author=. USSR Computational Mathematics and Mathematical Physics , volume=. 1967 , publisher=

    1967

  15. [15]

    arXiv preprint arXiv:1802.07529 , year=

    Algorithms and convergence results of projection methods for inconsistent feasibility problems: A review , author=. arXiv preprint arXiv:1802.07529 , year=

    Pith/arXiv arXiv

  16. [16]

    Set-Valued Analysis , volume=

    On the convergence of von Neumann's alternating projection algorithm for two sets , author=. Set-Valued Analysis , volume=. 1993 , publisher=

    1993

  17. [17]

    Contemporary Mathematics , volume=

    The method of cyclic projections for closed convex sets in Hilbert space , author=. Contemporary Mathematics , volume=. 1997 , publisher=

    1997

  18. [18]

    and Chen, B

    Cong, Y. and Chen, B. and Zhou, M. , title =. Bayesian Analysis , volume =

  19. [19]

    Dykstra, R. L. , title =. J. Amer. Statist. Assoc , volume =

  20. [20]

    Boyle, J. P. and Dykstra, R. L. , title =. Advances in Order Restricted Statistical Inference , editor =

  21. [21]

    Sampling large hyperplane-truncated multivariate normal distributions , journal =

    Hassan Maatouk, Didier Rulli\'. Sampling large hyperplane-truncated multivariate normal distributions , journal =

  22. [22]

    Hoffman and E

    Y. Hoffman and E. Ribak. , title =. The Astrophysical Journal , volume =

  23. [23]

    Journel and C.J

    A.G. Journel and C.J. Huijbregts , title =

  24. [24]

    Efficient Simulation from the Multivariate Normal and Student-t Distributions Subject to Linear Constraints and the Evaluation of Constraint Probabilities , volume =

    Geweke, John , year =. Efficient Simulation from the Multivariate Normal and Student-t Distributions Subject to Linear Constraints and the Evaluation of Constraint Probabilities , volume =

  25. [25]

    Zaknoon , year =

    Y Censor and M. Zaknoon , year =

  26. [26]

    2022 , title =

    Abowd, John and Ashmead, Robert and Cumings-Menon, Ryan and Garfinkel, Simson and Heineck, Micah and Heiss, Christine and Johns, Robert and Kifer, Daniel and Leclerc, Philip and Machanavajjhala, Ashwin and Moran, Brett and Sexton, William and Spence, Matthew and Zhuravlev, Pavel , journal =. 2022 , title =

    2022

  27. [27]

    Chen , title =

    Wei Ma and George H. Chen , title =. NeurIPS , year =

  28. [28]

    An augmented Lagrangian dual approach for the H-weighted nearest correlation matrix problem , year=

    Qi, Houduo and Sun, Defeng , journal=. An augmented Lagrangian dual approach for the H-weighted nearest correlation matrix problem , year=

  29. [29]

    Moreau, Jean Jacques , journal=. D

  30. [30]

    , year =

    Ramana, Motakuri and Goldman, A. , year =. Some Geometric Results in Semidefinite Programming , volume =. Journal of Global Optimization , doi =

  31. [31]

    Tyrrell , biburl =

    Rockafellar, R. Tyrrell , biburl =

  32. [32]

    SIAM Journal on Computing , volume=

    What can we learn privately? , author=. SIAM Journal on Computing , volume=. 2011 , publisher=

    2011

  33. [33]

    Journal of multivariate analysis , volume=

    Parametric families of multivariate distributions with given margins , author=. Journal of multivariate analysis , volume=. 1993 , publisher=

    1993

  34. [34]

    Journal of multivariate analysis , volume=

    Constructing multivariate distributions with specific marginal distributions , author=. Journal of multivariate analysis , volume=. 1995 , publisher=

    1995

  35. [35]

    Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=

    Minimum correlation in construction of multivariate distributions , author=. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics , volume=. 2013 , publisher=

    2013

  36. [36]

    Journal of Applied Probability , volume=

    Multivariate distributions with fixed marginals and correlations , author=. Journal of Applied Probability , volume=. 2015 , publisher=

    2015

  37. [37]

    2014 IEEE International Symposium on Information Theory , pages=

    Entropy for singular distributions , author=. 2014 IEEE International Symposium on Information Theory , pages=. 2014 , organization=

    2014

  38. [38]

    1951 , publisher=

    Functional operators: The geometry of orthogonal spaces , author=. 1951 , publisher=

    1951

  39. [39]

    Reduction theory , author=

    On rings of operators. Reduction theory , author=. Annals of Mathematics , volume=. 1949 , publisher=

    1949

  40. [40]

    Mathematical programming , volume=

    Projected gradient methods for linearly constrained problems , author=. Mathematical programming , volume=. 1987 , publisher=

    1987

  41. [41]

    2021 , publisher=

    Introduction to continuous optimization , author=. 2021 , publisher=

    2021

  42. [42]

    2011 , publisher=

    Nonlinear optimization , author=. 2011 , publisher=

    2011

  43. [43]

    Journal of Multivariate Analysis , volume=

    The joint distribution of the sum and maximum of dependent Pareto risks , author=. Journal of Multivariate Analysis , volume=

  44. [44]

    Technometrics , volume=

    Some statistical characteristics of a peak to average ratio , author=. Technometrics , volume=. 1965 , publisher=

    1965

  45. [45]

    Theory of Probability & Its Applications , volume=

    The extreme terms of a sample and their role in the sum of independent variables , author=. Theory of Probability & Its Applications , volume=. 1960 , publisher=

    1960

  46. [46]

    Journal of Multivariate Analysis , volume=

    Scale mixtures of Kotz--Dirichlet distributions , author=. Journal of Multivariate Analysis , volume=. 2013 , publisher=

    2013

  47. [47]

    Theory of Probability & Its Applications , volume=

    On some limit theorems similar to the arc-sin law , author=. Theory of Probability & Its Applications , volume=. 1965 , publisher=

    1965

  48. [48]

    Transactions of the American Mathematical Society , volume=

    The influence of the maximum term in the addition of independent random variables , author=. Transactions of the American Mathematical Society , volume=. 1952 , publisher=

    1952

  49. [49]

    Journal of applied probability , volume=

    The maximum and mean of a random length sequence , author=. Journal of applied probability , volume=. 1992 , publisher=

    1992

  50. [50]

    Advances in Directional and Linear Statistics: A Festschrift for Sreenivasa Rao Jammalamadaka , pages=

    The distributions of the peak to average and peak to sum ratios under exponentiality , author=. Advances in Directional and Linear Statistics: A Festschrift for Sreenivasa Rao Jammalamadaka , pages=. 2010 , publisher=

    2010

  51. [51]

    Communications in Statistics-Theory and Methods , volume=

    The joint distribution of the sum and the maximum of IID exponential random variables , author=. Communications in Statistics-Theory and Methods , volume=. 2012 , publisher=

    2012

  52. [52]

    Journal of the American Statistical Association , volume =

    Andrew Gelman and Aki Vehtari , title =. Journal of the American Statistical Association , volume =. 2021 , publisher =. doi:10.1080/01621459.2021.1938081 , URL =

    arXiv 2021

  53. [53]

    Broman and Kara H

    Karl W. Broman and Kara H. Woo , title =. The American Statistician , volume =. 2018 , publisher =. doi:10.1080/00031305.2017.1375989 , URL =

    Pith/arXiv arXiv 2018

  54. [54]

    Bishop, Yvonne M. M. and Fienberg, Stephen E. and Holland, Paul W. , title =. 1975 , pages =

    1975

  55. [55]

    Bailie, James and Gong, Ruobin and Meng, Xiao-Li , journal =. A. 2026 , month =

    2026