Pith. sign in

REVIEW 3 minor 39 references

For attractively dependent targets, the independent product distribution is already the best DPP approximation.

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 14:07 UTC pith:QPDYBA5N

load-bearing objection The core reduction and the global optimality result under weak positive association are correct; a solid, accept-quality paper with two minor completeness gaps.

arxiv 2607.18849 v2 pith:QPDYBA5N submitted 2026-07-21 math.ST stat.TH

Determinantal Point Process Approximation under Positive and Negative Dependence

So Anzai , Hideitsu Hino This is my paper
classification math.ST stat.TH MSC 60G5562B10
keywords Determinantal point processesInformation geometryPositive and negative dependenceModel misspecificationForward KL divergenceQuality-diversity decompositionSupermodular orderInclusion probabilities
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 studies how well a determinantal point process (DPP) can approximate an arbitrary strictly positive distribution on random subsets, measuring error by forward Kullback-Leibler divergence. Its central reduction shows that the approximation splits cleanly: the quality matrix is always chosen uniquely to match all first-order inclusion probabilities, leaving the diversity matrix as the only free part. For targets with positive dependence, specifically weak positive association, the independent product distribution with the same marginals is globally optimal among all DPPs — repulsion cannot help. For negatively dependent targets, disjoint negatively correlated pairs give guaranteed improvements over independence, quantified by pairwise mutual information. The paper also gives local optimality conditions around the independent and block-diagonal diversity matrices.

Core claim

For any strictly positive target p*, the DPP approximation problem reduces to optimizing only the diversity component D of the L-ensemble kernel. A unique quality matrix always matches all first-order inclusion probabilities, so the error becomes inf over D of KL(p* || ν_D), where ν_D is the distribution sharing p*'s first-order marginals and D's higher-order interaction coordinates. If p* is weakly positively associated (or more generally p* ⪰_ism ν_I), then D = I is globally optimal and KL(p* || DPP) = KL(p* || ν_I), attained by the diagonal kernels K = Diag(η*_1,...,η*_N) and L = Diag(η*_i/(1−η*_i)). On the repulsive side, a matching of disjoint negatively correlated pairs yields KL(p* ||

What carries the argument

The load-bearing mechanism is the quality–diversity decomposition L = √Q D √Q together with a recursion expressing all higher-order interaction coordinates of a DPP distribution purely in terms of D: θ^I(p_L) = log det D_I − Σ_{J⊊I, |J|≥2} θ^J(p_L). This makes the first-order θ-coordinates depend only on Q and all higher-order coordinates only on D, enabling a 1-cut mixed-coordinate Pythagorean decomposition. The attractive-side result then uses the fact that S ↦ −log det D_S is increasing and supermodular, combined with a supermodular-order comparison between p* and the independent product ν_I.

Load-bearing premise

The entire reduction rests on a recursion, cited from the authors' earlier work rather than proved here, that expresses every higher-order interaction coordinate of a DPP distribution solely in terms of the diversity matrix; if that identity fails, quality and diversity do not separate and the global optimality results collapse.

What would settle it

Take a three-element ground set and search over strictly positive distributions satisfying weak positive association, computing min_D KL(p* || ν_D) over all feasible diversity matrices D. If any D ≠ I yields a smaller KL than KL(p* || ν_I), Theorem 2 is false. Alternatively, for a random positive-definite D with unit diagonal, compute the DPP's probabilities directly, apply Möbius inversion to get θ, and check whether the recursion θ^I(p_L) = log det D_I − Σ_{J⊊I} θ^J(p_L) holds; any mismatch would disprove the load-bearing identity.

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

If this is right

  • When a target is weakly positively associated, fitting a DPP with any nontrivial diversity matrix cannot beat the simple independent product with the same first-order marginals; the optimal kernels are explicitly diagonal.
  • The approximation problem for any strictly positive target reduces to estimating first-order inclusion probabilities and then optimizing diversity, so kernel-learning algorithms can be split into these two stages.
  • If any pair of items is strictly positively correlated, the target cannot be a DPP, and the approximation error is at least the mutual information of that pair, which is bounded below by 8δ² where δ is the excess pairwise correlation.
  • For targets with disjoint negatively correlated pairs, a block-diagonal DPP kernel gives a guaranteed improvement over independence equal to the sum of the pairwise mutual informations of those pairs.
  • Local optimality of D = I is equivalent to pairwise positive correlation, and local optimality of a block-diagonal diversity matrix is governed by the sign of a second-order off-block perturbation term.

Where Pith is reading between the lines

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

  • Editorial inference: The results suggest a practical model-selection rule for DPPs — estimate pairwise inclusion probabilities first; if they are all at least the product of marginals, leave the diversity matrix as the identity and save the cost of learning off-diagonal parameters.
  • Editorial inference: The repulsive-side bound is built from disjoint pairs only and is not claimed to be sharp; extending it to interactions among three or more items is the natural next step, and the off-block perturbation criterion indicates how such higher-order structure could be optimized.
  • Editorial inference: The paper's strict-positivity assumption is technical for the log-linear coordinate system, but real subset data often include impossible combinations; a boundary-aware extension would likely require a different coordinate representation and may change the global optimality statement.
  • Editorial inference: A testable algorithmic consequence is that a maximum-weight repulsive matching, using estimated pairwise mutual informations as weights, would provide a computable, certified upper bound on DPP approximation error in practice.

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

0 major / 3 minor

Summary. The paper studies the population-level forward-KL approximation of a strictly positive target distribution p* on 2^[N] by DPPs. Using the quality–diversity decomposition L = sqrt(Q) D sqrt(Q) and the recursive θ-coordinate relation (7), the authors prove a structural reduction (Theorem 1): the DPP approximation problem reduces to minimizing KL(p* ∥ ν_D) over diversity matrices D, where ν_D is the unique distribution with p*'s first-order marginals and D's higher-order θ-coordinates. For weakly positively associated targets (in particular PLC or PA targets), Theorem 2 and Corollary 1 show that the independent product distribution ν_I with the same first-order marginals is globally optimal among all DPP approximations, with the explicit diagonal kernels (14)–(15). For targets with positively correlated pairs, Proposition 1 gives a positive lower bound on the approximation error in terms of pairwise mutual information, and Proposition 2 gives a local optimality characterization at D = I. On the repulsive side, Proposition 3 constructs block-diagonal DPP approximations that improve on ν_I by an amount bounded below by the sum of mutual informations over a matching of negatively correlated pairs, and Proposition 4 gives local criteria for whether off-block perturbations can further improve a block-structured minimizer. Numerical experiments for N = 3 use certified branch-and-bound bounds to illustrate the phase boundary between the attractive and repulsive regimes.

Significance. If the results hold, the paper provides a clean and broadly applicable answer to when the repulsive structure of a DPP can improve on the independent baseline under model misspecification. The main global optimality result (Theorem 2) is conceptually striking: for weakly positively associated targets, no DPP can beat independence, so the common intuition that DPPs are useful whenever repulsion is present is shown to fail. The reduction to the diversity component in Theorem 1 is simple, general, and does not rely on finite-sample heuristics. The repulsive-side matching bound is constructive and gives a quantitative improvement guarantee. The proofs are generally careful; the perturbation analysis in Section 4.3 and Section 5.2 controls remainder terms through Lemmas 5–7, and the numerical section provides certified bounds rather than only point estimates, which strengthens the empirical claims. The only soft spot is that two load-bearing identities are imported rather than proved in this manuscript: the L-ensemble restriction in Section 2.1 and the recursion (7) from Hino–Yano [26]. I verified both; they are correct, and the issues are local. The paper is not circular: Eq. (7) is a

minor comments (3)
  1. [Section 2.1, first paragraph] The sentence 'restricting the optimization to L-ensembles does not change the approximation error' is asserted without proof. The claim is true, but it should be justified: any DPP with an eigenvalue equal to 0 or 1 assigns zero probability to some subset, and for a strictly positive p* such a DPP has infinite forward KL divergence; hence the infimum is attained among kernels with 0 ≺ K ≺ I, i.e., among L-ensembles. Adding this one-line argument would make the reduction fully transparent.
  2. [Section 2.2, Eq. (7)] The recursion θ^{i}(p_L) = log Q_{ii}, θ^I(p_L) = log det D_I − Σ_{J⊊I, |J|≥2} θ^J(p_L) is cited from the authors' own Hino–Yano paper [26] and is load-bearing for Theorem 1 and all subsequent results. I checked that it follows by Möbius inversion of log det D_J in the θ-coordinates, so I do not doubt its correctness. However, because so much of the paper depends on it, the manuscript would be more self-contained if a derivation or at least a sketch were included in an appendix.
  3. [Proposition 2, proof] The proof of the necessary condition for D = I to be a local minimizer reads off the second-order coefficient after noting that the first-order term vanishes. Since perturbing with both positive and negative t is allowed, it is worth stating explicitly that local minimality forces Δ_{1,I}(E) = 0 before the t²-coefficient is interpreted; the current proof is implicitly correct but would benefit from this clarification.

Circularity Check

0 steps flagged

No significant circularity: the derivation is self-contained, and the only self-cited identity is parameter-free and independently verifiable.

full rationale

The central chain is not circular. Theorem 1 reduces the DPP approximation problem to the diversity component using the mixed-coordinate Pythagorean decomposition, which is proved in Appendix A.1 on the basis of the external extended Pythagorean theorem, and using Eq. (7). Eq. (7), cited from the authors' own Hino–Yano paper, is the only self-cited load-bearing ingredient, but it is a parameter-free algebraic identity following from the quality–diversity decomposition log det L_A = log det Q_A + log det D_A together with Möbius inversion; it does not contain the optimality result and can be verified independently, so under the stated rules it is real evidence rather than circularity. Theorem 2 is then derived from the extended Pythagorean theorem and Lemma 1, whose supermodularity claim rests on Koteljanskii's inequality; Corollary 1 uses the Christofides–Vaggelatou supermodular-order comparison, with a self-contained discrete proof in Appendix A.2. Proposition 1 uses the data-processing inequality and exact two-dimensional DPP characterizations; Proposition 3 uses block-factorization and the pairwise DPP construction; Proposition 4 uses perturbation expansions derived in Appendix A.3. No parameters are fitted to data and then renamed as predictions; the numerical section gives certified bounds rather than fitted claims. No circular reduction of a conclusion to an input was found.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper is self-contained theoretical work; no free parameters fitted to data and no invented entities. Its load-bearing inputs are standard information-geometric theorems, exponential-family bijectivity, the Koteljanskii inequality, and the strict-positivity domain restriction.

axioms (5)
  • standard math Extended Pythagorean theorem for KL divergence on the probability simplex (Amari–Nagaoka Theorem 3.7)
    Used in Section 2.2 and throughout to decompose KL(p*||ν_D) and to parameterize ν_D in mixed coordinates.
  • standard math Gradient map of a regular minimal exponential family is a bijection from natural-parameter space onto the interior of the mean-parameter space (Wainwright–Jordan Prop 3.2, Thm 3.3)
    Constitutes the existence proof of Q in Theorem 1 and of the mixed-coordinate point q in Appendix A.1.
  • standard math Koteljanskii (Hadamard–Fischer) inequality for principal minors of positive definite matrices
    Used in Lemma 1 to prove supermodularity and monotonicity of S↦−log det D_S.
  • domain assumption The target p* is strictly positive on 2^[N]
    Ensures η*(1)∈(0,1)^N so the quality-matching Q exists and the KL divergence is finite; the authors note in Section 7 that boundary targets would require modification.
  • domain assumption Restricting the DPP family to L-ensembles (0<eig(K)<1) does not change the infimum approximation error for strictly positive targets
    Stated in Section 2.1 without proof; boundary DPPs must be approximated by interior kernels, a standard density argument not spelled out.

pith-pipeline@v1.3.0-alltime-deepseek · 23242 in / 17649 out tokens · 141779 ms · 2026-08-01T14:07:58.054717+00:00 · methodology

0 comments
read the original abstract

Determinantal point processes (DPPs) are widely used as probabilistic models for diverse random subsets, but their approximation error under model misspecification has not been fully characterized. We study the population-level approximation of a strictly positive target distribution p* by DPPs under the forward Kullback-Leibler divergence. Using information-geometric analysis and the standard quality-diversity decomposition of an L-ensemble kernel, in which the diagonal quality component Q encodes item-specific weights and the diversity component D controls repulsive interactions among items, we show that the quality component can be chosen uniquely to match all first-order inclusion probabilities of p*. The DPP approximation problem therefore reduces to the optimization of the diversity component. This reduction yields a global optimality result for attractively dependent targets: under conditions including weak positive association, the independent product distribution with the same first-order marginals, corresponding to D=I, is an optimal DPP approximation. For more general target distributions, positively correlated pairs yield lower bounds on the approximation error. On the repulsive side, a matching of disjoint negatively correlated pairs yields an upper bound on the approximation error, or equivalently a guaranteed improvement over the independent approximation. We further study local optimality around D=I and, more generally, around block-diagonal diversity matrices by analyzing perturbations between their blocks.

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

39 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Foundations and Trends in Machine Learning, vol

    Kulesza, A., Taskar, B.: Determinantal Point Processes for Machine Learn- ing. Foundations and Trends in Machine Learning, vol. 5. Now Publishers Inc., Hanover, MA, USA (2012). https://doi.org/10.1561/2200000044

    doi:10.1561/2200000044 2012

  2. [2]

    Advances in Applied Probability7(1), 83–122 (1975) https://doi.org/10.2307/1425855

    Macchi, O.: The coincidence approach to stochastic point processes. Advances in Applied Probability7(1), 83–122 (1975) https://doi.org/10.2307/1425855

    doi:10.2307/1425855 1975

  3. [3]

    In: Cozman, F.G., Pfeffer, A

    Kulesza, A., Taskar, B.: Learning determinantal point processes. In: Cozman, F.G., Pfeffer, A. (eds.) Proceedings of the Twenty-Seventh Conference on Uncer- tainty in Artificial Intelligence, pp. 419–427. AUAI Press, Corvallis, OR, USA (2011)

    2011

  4. [4]

    In: Proceedings of the 10th ACM Conference on Recommender Sys- tems, pp

    Gartrell, M., Paquet, U., Koenigstein, N.: Bayesian low-rank determinantal point processes. In: Proceedings of the 10th ACM Conference on Recommender Sys- tems, pp. 349–356. Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2959100.2959178

    arXiv 2016

  5. [5]

    Statistica Sinica26(1), 97–117 (2016) https://doi

    Kojima, M., Komaki, F.: Determinantal point process priors for Bayesian variable selection in linear regression. Statistica Sinica26(1), 97–117 (2016) https://doi. org/10.5705/ss.202014.0161

    arXiv 2016

  6. [6]

    In: Advances in Neural Information Processing Systems, vol

    Snoek, J., Zemel, R.S., Adams, R.P.: A determinantal point process latent variable model for inhibition in neural spiking data. In: Advances in Neural Information Processing Systems, vol. 26, pp. 1932–1940. Curran Associates, Inc., Red Hook, NY, USA (2013)

    1932

  7. [7]

    Diversifying Sparsity Using Variational Determinantal Point Processes

    Batmanghelich, N.K., Quon, G., Kulesza, A., Kellis, M., Golland, P., Bornn, L.: Diversifying Sparsity Using Variational Determinantal Point Processes (2014). https://doi.org/10.48550/arXiv.1411.6307

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1411.6307 2014

  8. [8]

    In: Advances in Neural Information Processing Systems, vol

    Gillenwater, J.A., Kulesza, A., Fox, E.B., Taskar, B.: Expectation-maximization for learning determinantal point processes. In: Advances in Neural Information Processing Systems, vol. 27, pp. 3149–3157. Curran Associates, Inc., Red Hook, NY, USA (2014)

    2014

  9. [9]

    In: Bach, F., Blei, D

    Mariet, Z., Sra, S.: Fixed-point algorithms for learning determinantal point processes. In: Bach, F., Blei, D. (eds.) Proceedings of the 32nd International Con- ference on Machine Learning. Proceedings of Machine Learning Research, vol. 37, pp. 2389–2397. PMLR, Lille, France (2015)

    2015

  10. [10]

    Transactions on Machine Learning Research (2023)

    Kawashima, T., Hino, H.: Minorization-maximization for learning determinantal 32 point processes. Transactions on Machine Learning Research (2023)

    2023

  11. [11]

    Transactions on Machine Learning Research (2026)

    Castella, M., Pesquet, J.-C.: Kernel matrix estimation of a determinantal point process from a finite set of samples: Properties and algorithms. Transactions on Machine Learning Research (2026)

    2026

  12. [12]

    In: Xing, E.P., Jebara, T

    Affandi, R.H., Fox, E., Adams, R., Taskar, B.: Learning the parameters of deter- minantal point process kernels. In: Xing, E.P., Jebara, T. (eds.) Proceedings of the 31st International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 32, pp. 1224–1232. PMLR, Beijing, China (2014)

    2014

  13. [13]

    In: Precup, D., Teh, Y.W

    Urschel, J., Brunel, V.-E., Moitra, A., Rigollet, P.: Learning determinantal point processes with moments and cycles. In: Precup, D., Teh, Y.W. (eds.) Proceed- ings of the 34th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 70, pp. 3511–3520. PMLR, Sydney, Australia (2017)

    2017

  14. [14]

    A simple estimator of the correlation kernel matrix of a determinantal point process

    Gouri´ eroux, C., Lu, Y.: A Simple Estimator of the Correlation Kernel Matrix of a Determinantal Point Process (2025). https://doi.org/10.48550/arXiv.2505.14529

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.14529 2025

  15. [15]

    In: Advances in Neural Information Processing Systems, vol

    Mariet, Z.E., Sra, S.: Kronecker determinantal point processes. In: Advances in Neural Information Processing Systems, vol. 29, pp. 2694–2702. Curran Associates, Inc., Red Hook, NY, USA (2016)

    2016

  16. [16]

    In: Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, vol

    Gartrell, M., Paquet, U., Koenigstein, N.: Low-rank factorization of determinantal point processes. In: Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, vol. 31, pp. 1912–1918 (2017). https://doi.org/10.1609/aaai.v31i1. 10869

    doi:10.1609/aaai.v31i1 1912

  17. [17]

    Probability Surveys3, 206–229 (2006) https://doi.org/10.1214/ 154957806000000078

    Hough, J.B., Krishnapur, M., Peres, Y., Vir´ ag, B.: Determinantal processes and independence. Probability Surveys3, 206–229 (2006) https://doi.org/10.1214/ 154957806000000078

    2006

  18. [18]

    In: Feldman, V., Rakhlin, A., Shamir, O

    Anari, N., Oveis Gharan, S., Rezaei, A.: Monte Carlo Markov Chain algorithms for sampling Strongly Rayleigh distributions and determinantal point processes. In: Feldman, V., Rakhlin, A., Shamir, O. (eds.) 29th Annual Conference on Learning Theory. Proceedings of Machine Learning Research, vol. 49, pp. 103–115. PMLR, Columbia University, New York, New York...

    2016

  19. [19]

    In: Advances in Neural Infor- mation Processing Systems, vol

    Derezi´ nski, M., Calandriello, D., Valko, M.: Exact sampling of determinantal point processes with sublinear time preprocessing. In: Advances in Neural Infor- mation Processing Systems, vol. 32. Curran Associates, Inc., Red Hook, NY, USA (2019)

    2019

  20. [20]

    Journal of the American Mathematical Society22(2), 521–567 (2009) https://doi.org/10.1090/S0894-0347-08-00618-8 33

    Borcea, J., Br¨ and´ en, P., Liggett, T.M.: Negative dependence and the geometry of polynomials. Journal of the American Mathematical Society22(2), 521–567 (2009) https://doi.org/10.1090/S0894-0347-08-00618-8 33

    doi:10.1090/s0894-0347-08-00618-8 2009

  21. [21]

    https://doi.org/10.48550/arXiv

    Brunel, V.-E., Moitra, A., Rigollet, P., Urschel, J.: Maximum Likelihood Estima- tion of Determinantal Point Processes (2017). https://doi.org/10.48550/arXiv. 1701.06501

    Pith/arXiv arXiv doi:10.48550/arxiv.1701.06501 2017

  22. [22]

    IEEE Transactions on Information Theory47(5), 1701–1711 (2001) https://doi.org/10

    Amari, S.: Information geometry on hierarchy of probability distributions. IEEE Transactions on Information Theory47(5), 1701–1711 (2001) https://doi.org/10. 1109/18.930911

    2001

  23. [23]

    Kybernetika50(2), 234–245 (2014) https://doi.org/10.14736/ kyb-2014-2-0234

    Mont´ ufar, G.F., Rauh, J.: Scaling of model approximation errors and expected entropy distances. Kybernetika50(2), 234–245 (2014) https://doi.org/10.14736/ kyb-2014-2-0234

    2014

  24. [24]

    In: Advances in Neural Information Processing Systems, vol

    Gatmiry, K., Aliakbarpour, M., Jegelka, S.: Testing determinantal point pro- cesses. In: Advances in Neural Information Processing Systems, vol. 33, pp. 12779–12791. Curran Associates, Inc., Red Hook, NY, USA (2020)

    2020

  25. [25]

    Translations of Math- ematical Monographs, vol

    Amari, S., Nagaoka, H.: Methods of Information Geometry. Translations of Math- ematical Monographs, vol. 191. American Mathematical Society and Oxford University Press, Providence, RI, USA and Oxford, UK (2000). https://doi.org/ 10.1090/mmono/191

    doi:10.1090/mmono/191 2000

  26. [26]

    Information Geometry7(2), 523–542 (2024) https://doi.org/10.1007/ s41884-024-00156-x

    Hino, H., Yano, K.: An embedding structure of determinantal point pro- cess. Information Geometry7(2), 523–542 (2024) https://doi.org/10.1007/ s41884-024-00156-x

    2024

  27. [27]

    The Annals of Mathematical Statistics38(5), 1466–1474 (1967) https://doi.org/10.1214/aoms/1177698701

    Esary, J.D., Proschan, F., Walkup, D.W.: Association of random variables, with applications. The Annals of Mathematical Statistics38(5), 1466–1474 (1967) https://doi.org/10.1214/aoms/1177698701

    arXiv 1967

  28. [28]

    Stochastic Processes and their Applications23(2), 301–306 (1986) https://doi.org/10.1016/0304-4149(86)90043-8

    Burton, R.M., Dabrowski, A.R., Dehling, H.: An invariance principle for weakly associated random vectors. Stochastic Processes and their Applications23(2), 301–306 (1986) https://doi.org/10.1016/0304-4149(86)90043-8

    work page doi:10.1016/0304-4149(86)90043-8 1986

  29. [29]

    Communications in Mathematical Physics22(2), 89–103 (1971) https://doi.org/10.1007/BF01651330

    Fortuin, C.M., Kasteleyn, P.W., Ginibre, J.: Correlation inequalities on some partially ordered sets. Communications in Mathematical Physics22(2), 89–103 (1971) https://doi.org/10.1007/BF01651330

    work page doi:10.1007/bf01651330 1971

  30. [30]

    The Annals of Statistics11(1), 286–295 (1983) https://doi.org/10.1214/ aos/1176346079

    Joag-Dev, K., Proschan, F.: Negative association of random variables with appli- cations. The Annals of Statistics11(1), 286–295 (1983) https://doi.org/10.1214/ aos/1176346079

    arXiv 1983

  31. [31]

    Journal of Mathematical Physics41(3), 1371–1390 (2000) https://doi.org/10.1063/1.533200

    Pemantle, R.: Towards a theory of negative dependence. Journal of Mathematical Physics41(3), 1371–1390 (2000) https://doi.org/10.1063/1.533200

    work page doi:10.1063/1.533200 2000

  32. [32]

    In: Lebanon, G., Vishwanathan, S.V.N

    Iyer, R., Bilmes, J.: Submodular point processes with applications to machine learning. In: Lebanon, G., Vishwanathan, S.V.N. (eds.) Proceedings of the 34 Eighteenth International Conference on Artificial Intelligence and Statistics. Pro- ceedings of Machine Learning Research, vol. 38, pp. 388–397. PMLR, San Diego, CA, USA (2015)

    2015

  33. [33]

    In: Li, Y., Mandt, S., Agrawal, S., Khan, E

    Kawashima, T., Hino, H.: A family of distributions of random subsets for control- ling positive and negative dependence. In: Li, Y., Mandt, S., Agrawal, S., Khan, E. (eds.) Proceedings of the 28th International Conference on Artificial Intelli- gence and Statistics. Proceedings of Machine Learning Research, vol. 258, pp. 64–72. PMLR, Mai Khao, Thailand (2025)

    2025

  34. [34]

    Foundations and Trends in Machine Learning1(1–2), 1–305 (2008) https://doi.org/10.1561/2200000001

    Wainwright, M.J., Jordan, M.I.: Graphical models, exponential families, and vari- ational inference. Foundations and Trends in Machine Learning1(1–2), 1–305 (2008) https://doi.org/10.1561/2200000001

    doi:10.1561/2200000001 2008

  35. [35]

    Wiley Series in Probability and Statistics

    M¨ uller, A., Stoyan, D.: Comparison Methods for Stochastic Models and Risks. Wiley Series in Probability and Statistics. John Wiley & Sons, Chichester, UK (2002)

    2002

  36. [36]

    Cambridge University Press, Cambridge, UK (2012)

    Horn, R.A., Johnson, C.R.: Matrix Analysis, 2nd edn. Cambridge University Press, Cambridge, UK (2012). https://doi.org/10.1017/CBO9781139020411

    doi:10.1017/cbo9781139020411 2012

  37. [37]

    Journal of Multivariate Analysis88(1), 138– 151 (2004) https://doi.org/10.1016/S0047-259X(03)00064-2

    Christofides, T.C., Vaggelatou, E.: A connection between supermodular ordering and positive/negative association. Journal of Multivariate Analysis88(1), 138– 151 (2004) https://doi.org/10.1016/S0047-259X(03)00064-2

    work page doi:10.1016/s0047-259x(03)00064-2 2004

  38. [38]

    John Wiley & Sons, Hoboken, NJ, USA (2006)

    Cover, T.M., Thomas, J.A.: Elements of Information Theory, 2nd edn. John Wiley & Sons, Hoboken, NJ, USA (2006). https://doi.org/10.1002/047174882X

    doi:10.1002/047174882x 2006

  39. [39]

    Journal of Economic Theory147(4), 1460–1489 (2012) https://doi.org/10

    Meyer, M., Strulovici, B.: Increasing interdependence of multivariate distribu- tions. Journal of Economic Theory147(4), 1460–1489 (2012) https://doi.org/10. 1016/j.jet.2011.09.001 35

    2012