pith. sign in

arxiv: 1907.04008 · v1 · pith:KQXHL4OMnew · submitted 2019-07-09 · 📡 eess.SP · cs.RO· cs.SY· eess.SY· stat.CO

Decentralized Gaussian Mixture Fusion through Unified Quotient Approximations

Nisar R. Ahmed This is my paper

Pith reviewed 2026-05-25 00:26 UTC · model grok-4.3

classification 📡 eess.SP cs.ROcs.SYeess.SYstat.CO
keywords decentralized data fusionGaussian mixturequotient approximationimportance samplingtarget trackingBayesian fusionpeer-to-peercommon information
0
0 comments X

The pith

Decentralized Gaussian mixture fusion reduces to approximating non-Gaussian quotients via importance sampling to enable tractable recursive updates.

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

The paper shows that using finite Gaussian mixture PDFs for peer-to-peer decentralized data fusion leads to the same core problem whether exact or approximate methods are used: approximating the result of dividing a naive Bayes fusion mixture by another PDF that accounts for common information between sources. This division produces a mixture of non-Gaussian quotients that cannot be handled directly in recursive Bayesian filtering. Parallelizable importance sampling algorithms are introduced for both local direct approximation and global indirect approximation to recover usable Gaussian mixture representations. Examples in multi-platform static target search and range-based tracking of maneuverable targets illustrate that the resulting approximations achieve higher fidelity than prior GM DDF techniques while maintaining favorable run-time costs.

Core claim

Algorithms for both exact and approximate GM DDF lead to the same problem of finding a suitable GM approximation to a posterior fusion pdf resulting from the division of a naive Bayes fusion GM by another non-Gaussian pdf representing removal of common information. The resulting quotient pdf is naturally a mixture pdf with non-Gaussian and analytically intractable mixands. Parallelizable importance sampling algorithms for both direct local approximation and indirect global approximation of the quotient mixture are developed to obtain tractable GM approximations.

What carries the argument

Parallelizable importance sampling algorithms that perform direct local and indirect global approximation of the non-Gaussian quotient mixture

If this is right

  • The resulting GM approximations can be used directly in standard recursive Bayesian filters for multi-platform target search.
  • Higher fidelity is achieved in range-based tracking of maneuverable targets compared with existing GM DDF methods.
  • Favorable computational features are retained because the approximations support parallel execution.
  • Both static and dynamic target scenarios become amenable to the same unified quotient-based fusion procedure.

Where Pith is reading between the lines

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

  • The same quotient approximation strategy could be tested on fusion problems that begin with non-Gaussian source densities rather than Gaussian mixtures.
  • Long-horizon performance in very large networks would depend on how the parallel sampling scales when the number of platforms increases.
  • Integration with other local approximation methods such as variational inference might further reduce per-platform compute.
  • The approach leaves open whether similar quotient identities exist for fusion under different dependence structures beyond the naive Bayes case.

Load-bearing premise

The importance-sampling approximations of the non-Gaussian quotient mixtures remain sufficiently accurate across recursive Bayesian updates without the approximation errors accumulating to invalidate the fused posteriors.

What would settle it

Run the proposed fusion method recursively for many time steps on a multi-platform tracking scenario with known ground-truth centralized posterior and measure whether the decentralized fused density diverges from the centralized reference due to accumulated approximation error.

Figures

Figures reproduced from arXiv: 1907.04008 by Nisar R. Ahmed.

Figure 1
Figure 1. Figure 1: Block diagram of IGS for GM-based DDF. generally consists of a two-step optimization process for either form of DDF. The first step generates importance samples over the global posterior fusion mixture in eq. (26). The second step probabilistically associates these im￾portance samples to the various posterior fusion pdf mixands to facilitate weighted maximum likelihood estimation of their individual zeroth… view at source ↗
Figure 2
Figure 2. Figure 2: Block diagram of DLS for GM-based DDF. Overall, DLS is more computationally intensive than the IGS approxima￾tion, since IS sampling must now be carried out separately for each mixand. However, the IS sampling steps can be easily parallelized across the mixands of (28) as well as across the samples generated for each mixand, making it possible to speed up implementation. Assuming u(xk) has been suitably id… view at source ↗
Figure 3
Figure 3. Figure 3: GMs for fusion example: (a)-(b) p i (xk) and p j (xk) with Mi = Mj = 14; (c) common information pdf p c,ij (xk) with 40 components. per mixand); the moment-matched Gaussian denominator (MMGD) approx￾imation of [15] (c, eq.13); DLS (d, using Ns = 1000 total samples); the mix￾ture Laplace approximation (e); and the IS technique of [14] (f, using p c,ij (xk) as the proposal pdf, followed by EM compression of … view at source ↗
Figure 4
Figure 4. Figure 4: Results for approximating p f,E(xk) for GMs in [PITH_FULL_IMAGE:figures/full_fig_p032_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results for approximating p f,W(xk) for GMs in [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Kullback-Leibler divergences (a)-(b) and execution times (c)-(d) for 100 randomly generated 2D GM Exact and WEP simulated fusion problems. time penalties for exact fusion compared to IGS. The time increases for IGS in the exact case can be attributed to the fact that IGS must sample from whole mixture, compute IS weights, and then compute SS-WEM responsibil￾ities via Algorithm 3 for 100-2000 samples 100-12… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of component-wise GM fusion approximation techniques for exact DDF (with KLD from true fusion result): (a) agent i GM pdf, (b) agent j GM pdf, (c) common information GM pdf, (d) exact DDF fusion result (approximated on high density grid); (e) INGIS approximation (KLD = 0.2555 nats); (f) direct component-wise Laplace approximation without sampling correction (KLD = 0.3763 nats); (g) LAGIS approxi… view at source ↗
Figure 8
Figure 8. Figure 8: (a) Single run component-wise ESS results for exact GM fusion problem in [PITH_FULL_IMAGE:figures/full_fig_p037_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of centralized and DDF GM fusion results for 5 robot target search scenario: (a) GM prior and binary visual target detection viewcones for search robots; (b) centralized GM fusion posterior pdf for robot trajectories (shown in cyan) after k = 50 time steps; (c) local GM fusion result for robot 3 (which started from lower right corner) after k = 49 steps, prior to DDF update; (d) local GM fusion … view at source ↗
Figure 10
Figure 10. Figure 10: Typical true target trajectory and true sensing platform locations for maneuvering range￾only tracking scenario. where Ri,ρ = 400 m2 and Ri,ρ˙ = 1 (m/s)2 . 5.4.2 Bayesian estimators for data fusion Interactive multiple model (IMM) filtering strategies are well-suited to the hybrid stochastic dynamics for this problem. In IMM filtering, recursive Bayesian estimates are sought for the joint posterior mode a… view at source ↗
Figure 11
Figure 11. Figure 11: Typical platform marginal GMs for independent non-DDF based tracking across all ma￾neuvering modes for target’s estimated E-N position (mixture component 2σ ellipses shown, with colors corresponding to platforms). DDF, where mixands are discarded if their weights are numerically indistin￾guishable from zero. Both IGS (with Ns =1000) and FOCI are separately implemented to approximate the resulting WEP fusi… view at source ↗
Figure 12
Figure 12. Figure 12: Typical platform marginal GMs for FOCI WEP DDF across all maneuvering modes for target’s estimated E-N position: (a)-(c) prior to DDF updates; (d)-(f) following DDF updates (mixture component 2σ ellipses shown, with colors corresponding to platforms). (a) (b) (c) (d) (e) (f) [PITH_FULL_IMAGE:figures/full_fig_p045_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Typical platform marginal GMs for IGS WEP DDF across all maneuvering modes for target’s estimated E-N position: (a)-(c) prior to DDF updates; (d)-(f) following DDF updates (mixture component 2σ ellipses shown, with colors corresponding to platforms). platform generally manages to keep some modal mixands close to the target’s true trajectory for a significant portion of the tracking run. However, the combi… view at source ↗
Figure 14
Figure 14. Figure 14: Platform tracking RMSEs and 2σ bounds vs. time for different fusion methodologies, averaged over 50 Monte Carlo trials (errors shown on log scale). 47 [PITH_FULL_IMAGE:figures/full_fig_p047_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Platform 2 distributions for log of absolute state error immediately following DDF events. fusion instance, as expected. It can also be seen that in all instances, both IGS and FOCI remain ‘conservative’ in the MSE sense relative to the cen￾tralized optimal fusion result (and hence statistically consistent), although IGS is less conservative overall, especially in the time windows immediately following DD… view at source ↗
read the original abstract

This work examines the problem of using finite Gaussian mixtures (GM) probability density functions in recursive Bayesian peer-to-peer decentralized data fusion (DDF). It is shown that algorithms for both exact and approximate GM DDF lead to the same problem of finding a suitable GM approximation to a posterior fusion pdf resulting from the division of a `naive Bayes' fusion GM (representing direct combination of possibly dependent information sources) by another non-Gaussian pdf (representing removal of either the actual or estimated `common information' between the information sources). The resulting quotient pdf for general GM fusion is naturally a mixture pdf, although the fused mixands are non-Gaussian and are not analytically tractable for recursive Bayesian updates. Parallelizable importance sampling algorithms for both direct local approximation and indirect global approximation of the quotient mixture are developed to find tractable GM approximations to the non-Gaussian `sum of quotients' mixtures. Practical application examples for multi-platform static target search and maneuverable range-based target tracking demonstrate the higher fidelity of the resulting approximations compared to existing GM DDF techniques, as well as their favorable computational features.

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.

Referee Report

2 major / 2 minor

Summary. This paper examines the use of finite Gaussian mixtures (GMs) in recursive Bayesian peer-to-peer decentralized data fusion (DDF). It shows that both exact and approximate GM DDF reduce to the problem of finding a GM approximation to a quotient pdf formed by dividing a naive Bayes fusion GM by a non-Gaussian pdf representing common information. The resulting quotient is a mixture of non-Gaussian terms that are intractable for recursive updates. The authors develop parallelizable importance-sampling algorithms for direct local and indirect global approximation of this quotient mixture to obtain tractable GM representations. These methods are demonstrated on multi-platform static target search and maneuverable range-based target tracking, where they report higher fidelity and favorable computational features relative to existing GM DDF techniques.

Significance. If the results hold, the work supplies a unified treatment of the common-information problem for GM-based DDF that remains compatible with recursive Bayesian filtering. The explicit construction of the quotient mixture, the two parallelizable approximation algorithms, and the quantitative comparisons on recursive search and tracking scenarios are concrete strengths that could support practical deployment in distributed sensor networks.

major comments (2)
  1. [Application examples] Application examples: the central claim of higher fidelity rests on the importance-sampling approximations remaining accurate under recursion; while the two scenarios are recursive, the manuscript provides no explicit bounds or accumulated-error analysis, leaving generalization beyond the demonstrated regimes dependent on the specific Monte Carlo results shown.
  2. [Quotient-mixture construction] Quotient-mixture construction: the reduction of both exact and approximate DDF to the same quotient-mixture problem is load-bearing, yet the manuscript does not quantify how sensitive the final fused posterior is to the choice of common-information pdf (actual vs. estimated), which directly affects the fidelity gains reported.
minor comments (2)
  1. Notation for the sum-of-quotients mixture could be clarified with an explicit component-wise definition to aid readers implementing the sampling steps.
  2. A table comparing wall-clock time or flop counts of the direct versus indirect algorithms against the cited baseline GM DDF methods would make the computational-advantage claim easier to evaluate.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive review and recommendation of minor revision. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [Application examples] Application examples: the central claim of higher fidelity rests on the importance-sampling approximations remaining accurate under recursion; while the two scenarios are recursive, the manuscript provides no explicit bounds or accumulated-error analysis, leaving generalization beyond the demonstrated regimes dependent on the specific Monte Carlo results shown.

    Authors: We agree that explicit bounds or accumulated-error analysis would provide stronger guarantees for recursive application. Deriving such bounds for importance sampling on the non-Gaussian quotient mixtures is non-trivial and outside the scope of the present algorithmic and empirical contribution. The manuscript instead demonstrates stable recursive performance through Monte Carlo trials in the two scenarios. In revision we will add a brief discussion acknowledging the empirical nature of the validation and noting the lack of theoretical bounds as a limitation for future study. revision: partial

  2. Referee: [Quotient-mixture construction] Quotient-mixture construction: the reduction of both exact and approximate DDF to the same quotient-mixture problem is load-bearing, yet the manuscript does not quantify how sensitive the final fused posterior is to the choice of common-information pdf (actual vs. estimated), which directly affects the fidelity gains reported.

    Authors: The reduction to the quotient-mixture problem is obtained directly from the Bayesian fusion update and holds for any common-information pdf, whether the actual density or an estimate. The final posterior fidelity necessarily depends on the quality of that common-information input; this dependence is not a property of the quotient approximation algorithms themselves. The examples use estimated common information (standard in decentralized settings) and compare against prior GM DDF methods under the same modeling assumptions. A quantitative sensitivity study would require additional assumptions on estimation error and is not claimed or performed in the manuscript. revision: no

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper develops parallelizable importance-sampling algorithms to approximate non-Gaussian quotient mixtures arising in GM DDF, then validates them via explicit construction and quantitative comparisons on static-search and range-tracking scenarios. No load-bearing step reduces by the paper's own equations to a fitted parameter renamed as prediction, a self-citation chain, or an ansatz smuggled from prior author work; the central approximations are constructed directly from the quotient pdf definition and standard sampling methods without circular redefinition. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based solely on the abstract, the paper relies on standard Bayesian fusion assumptions and importance sampling without introducing new free parameters, axioms, or invented entities at the level of detail provided.

axioms (1)
  • domain assumption Recursive Bayesian updates remain valid when approximate GM representations replace exact non-Gaussian posteriors
    Implicit in the claim that the approximated quotients support further recursive fusion

pith-pipeline@v0.9.0 · 5728 in / 1188 out tokens · 22116 ms · 2026-05-25T00:26:41.045691+00:00 · methodology

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

57 extracted references · 57 canonical work pages

  1. [1]

    C.-Y. Chong, Hierarchical estimation, in: Proceedings of Second MIT/ONR Workshop on Distributed Information and Decision Systems Motivated by Naval Command Control Communication (C3) Problems, Monterey, CA, 2005

    work page 2005

  2. [2]

    Chong, S

    C.-Y. Chong, S. Mori, E. Tse, R. Wishner, Distributed estimation in distributed sensor networks, in: American Control Conference (ACC) 1982, Arlington, VA, 1982

    work page 1982

  3. [3]

    Chong, S

    C.-Y. Chong, S. Mori, E. Tse, Distributed estimation in networks, in: American Control Conference (ACC) 1983, San Francisco, CA, 1983

    work page 1983

  4. [4]

    Grime, H

    S. Grime, H. Durrant-Whyte, Data fusion in decentralized sensor net- works, Control Engineering Practice 2 (5) (1994) 849–863. 5interestingly, the closed-form solution derived by [53] bypasses the GM division prob- lem but does not yield a ‘true’ forward-backward algorithm as a result, since the sizes of the backward state messages grow over time rather t...

    work page 1994

  5. [5]

    M. E. Campbell, N. R. Ahmed, Distributed data fusion: Neighbors, ru- mors, and the art of collective knowledge, IEEE Control Systems 36 (4) (2016) 83–109

    work page 2016

  6. [6]

    J. R. Schoenberg, M. Campbell, Distributed terrain estimation using a mixture-model based algorithm, in: 12th Int’l Conf. on Information Fusion 2009 (FUSION’09), IEEE, 2009, pp. 960–967

    work page 2009

  7. [7]

    Ahmed, J

    N. Ahmed, J. Schoenberg, M. Campbell, Fast weighted exponential product rules for robust multi-robot data fusion, in: Robotics: Science and Systems 2012, 2012

    work page 2012

  8. [8]

    R. Tse, N. Ahmed, M. Campbell, Unified terrain mapping model with Markov random fields, IEEE Transactions on Robotics 31 (2) (2015) 290–306

    work page 2015

  9. [9]

    Ahmed, E

    N. Ahmed, E. Sample, M. Campbell, Bayesian Multicategorical Soft Data Fusion for Human-Robot Collaboration, IEEE Transactions on Robotics 29 (1) (2013) 189–206

    work page 2013

  10. [10]

    Julier, An empirical study into the use of chernoff information for robust, distributed fusion of Gaussian mixture models, in: FUSION 2006, 2006

    S. Julier, An empirical study into the use of chernoff information for robust, distributed fusion of Gaussian mixture models, in: FUSION 2006, 2006

    work page 2006

  11. [11]

    Ridley, B

    M. Ridley, B. Upcroft, L.-L. Ong, S. Kumar, S. Sukkarieh, Decentralised data fusion with parzen density estimates, in: Proc. of the 2004 In- tell. Sensors, Sensor Networks and Information Processing Conf., IEEE, 2004, pp. 161–166

    work page 2004

  12. [12]

    L.-L. Ong, B. Upcroft, M. Ridley, T. Bailey, S. Sukkarieh, H. Durrant- Whyte, Consistent Methods for Decentralised Data Fusion Using Par- ticle Filters, in: Int’l Conf. on Multisensor Fusion and Integration for Intell. Systems (MFI 2006), 2006, pp. 85–91

    work page 2006

  13. [13]

    L.-L. Ong, B. Upcroft, T. Bailey, M. Ridley, S. Sukkarieh, H. Durrant- Whyte, A decentralised particle filtering algorithm for multi-target tracking across multiple flight vehicles, in: 2006 Int’l Conf. on Intell. Robotics and Systems (IROS 2006), Beijing, China, 2006, pp. 4539– 4544. 51

    work page 2006

  14. [14]

    L.-L. Ong, T. Bailey, H. Durrant-Whyte, B. Upcroft, Decentralised Par- ticle Filtering for Multiple Target Tracking in Wireless Sensor Networks, in: 11th Int’l Conf. on Information Fusion (FUSION 2008), 2008, pp. 1–8

    work page 2008

  15. [15]

    K. C. Chang, W. Sun, Scalable Fusion with Mixture Distributions in Sensor Networks, in: 2010 Int’l Conf.s on Control, Automation, Robotics and Vision (ICARV), 2010, pp. 1251–1256

    work page 2010

  16. [16]

    Bailey, S

    T. Bailey, S. Julier, G. Agamennoni, On Conservative Fusion of Infor- mation with Unknown Non-Gaussian Dependence, in: 15th Int’l Conf. on Information Fusion Fusion 2012 (FUSION 2012), 2012, pp. 1–8

    work page 2012

  17. [17]

    N. R. Ahmed, What’s one mixture divided by another?: A unified ap- proach to high-fidelity distributed data fusion with mixture models, in: 2015 IEEE International Conference on Multisensor Fusion and Integra- tion for Intelligent Systems (MFI), IEEE, 2015, pp. 289–296

    work page 2015

  18. [18]

    Martin, K

    T. Martin, K. Chang, A distributed data fusion approach for mobile ad hoc networks, in: FUSION 2005, 2005, pp. 1062–1069

    work page 2005

  19. [19]

    Hurley, An information theoretic justification for covariance inter- section and its generalization, in: FUSION 2002, 2002, pp

    M. Hurley, An information theoretic justification for covariance inter- section and its generalization, in: FUSION 2002, 2002, pp. 505–511

    work page 2002

  20. [20]

    W. J. Farrell, C. Ganesh, Generalized Chernoff fusion approximation for practical distributed data fusion, in: 2009 12th International Conference on Information Fusion (FUSION 2009), IEEE, 2009, pp. 555–562

    work page 2009

  21. [21]

    C. N. Taylor, A. N. Bishop, Homogeneous functionals and Bayesian data fusion with unknown correlation, Information Fusion 45 (2019) 179–189

    work page 2019

  22. [22]

    J. Sijs, M. Lazar, P. Bosch, State fusion with unknown correlation: El- lipsoidal intersection, in: Proceedings of the 2010 American Control Conference, IEEE, 2010, pp. 3992–3997

    work page 2010

  23. [23]

    Rendas, J

    M.-J. Rendas, J. M. Leitao, Rumor-robust distributed data fusion, in: 2010 IEEE Conference on Multisensor Fusion and Integration (MFI 2010), IEEE, 2010, pp. 230–235

    work page 2010

  24. [24]

    S. J. Julier, J. K. Uhlmann, A non-divergent estimation algorithm in the presence of unknown correlations, in: Proceedings of the 1997 American Control Conference (ACC 1997), Vol. 4, IEEE, 1997, pp. 2369–2373. 52

    work page 1997

  25. [25]

    L. Chen, P. O. Arambel, R. K. Mehra, Fusion under unknown correlation-covariance intersection as a special case, in: Proceedings of the Fifth International Conference on Information Fusion (FUSION 2002), Vol. 2, IEEE, 2002, pp. 905–912

    work page 2002

  26. [26]

    Kaupp, B

    T. Kaupp, B. Douillard, F. Ramos, A. Makarenko, B. Upcroft, Shared environment representation for a human-robot team performing infor- mation fusion, Journal of Field Robotics 24 (11) (2007) 911–942

    work page 2007

  27. [27]

    R. Tse, N. Ahmed, M. Campbell, Unified mixture-model based terrain estimation with Markov Random Fields, in: 2012 IEEE Int’l Conf. on Multisensor Fusion and Integration for Intell. Systems MFI, IEEE, 2012, pp. 238–243

    work page 2012

  28. [28]

    Brunskill, L

    E. Brunskill, L. P. Kaelbling, T. Lozano-Perez, N. Roy, Planning in partially-observable switching-mode continuous domains, Annals of Mathematics and Artificial Intelligence 58 (3) (2010) 185–216. doi: 10.1007/s10472-010-9202-1

    work page doi:10.1007/s10472-010-9202-1 2010

  29. [29]

    Porta, N

    J. Porta, N. Vlassis, M. Spaan, P. Poupart, Point-based value it- eration for continuous POMDPs, IJCAI International Joint Confer- ence on Artificial Intelligence 7 (2011) 1968–1974. doi:10.5591/ 978-1-57735-516-8/IJCAI11-329

    work page 2011

  30. [30]

    Burks, N

    L. Burks, N. Ahmed, Optimal continuous state pomdp planning with semantic observations, in: 2017 IEEE Conference on Decision and Con- trol, IEEE, 2017, pp. 1509–1516

    work page 2017

  31. [31]

    Burks, I

    L. Burks, I. Loefgren, L. Barbier, J. Muesing, J. McGinley, S. Vunnam, N. Ahmed, Closed-loop bayesian semantic data fusion for collaborative human-autonomy target search, in: 2018 International Conference on Information Fusion (FUSION 2018), IEEE, 2018

    work page 2018

  32. [32]

    Lesser, M

    K. Lesser, M. Oishi, Approximate safety verification and control of par- tially observable stochastic hybrid systems, IEEE Transactions on Au- tomatic Control 62 (1) (2017) 81–96

    work page 2017

  33. [33]

    M. J. Wainwright, E. P. Simoncelli, Scale mixtures of gaussians and the statistics of natural images, in: Advances in neural information process- ing systems, 2000, pp. 855–861. 53

    work page 2000

  34. [34]

    Portilla, V

    J. Portilla, V. Strela, M. J. Wainwright, E. P. Simoncelli, Image de- noising using scale mixtures of gaussians in the wavelet domain, IEEE Transactions on Image processing 12 (11) (2003) 1338–1351

    work page 2003

  35. [35]

    Goldberger, H

    J. Goldberger, H. Greenspan, J. Dreyfuss, Simplifying mixture models using the unscented transform, IEEE Transactions on Pattern Analysis and Machine Intelligence 30 (8) (2008) 1496–1502

    work page 2008

  36. [36]

    Lagrange, M

    A. Lagrange, M. Fauvel, M. Grizonnet, Large-scale feature selection with gaussian mixture models for the classification of high dimensional remote sensing images, IEEE Transactions on Computational Imaging (2017)

    work page 2017

  37. [37]

    Upcroft, L

    B. Upcroft, L. L. Ong, S. Kumar, M. Ridley, T. Bailey, S. Sukkarieh, H. Durrant-Whyte, Rich probabilistic representations for bearing only decentralised data fusion, in: 8th Int’l Conf. on Information Fusion (FU- SION 2005), 2005

    work page 2005

  38. [38]

    West, Approximating posterior distributions by mixture, Journal of the Royal Statistical Society

    M. West, Approximating posterior distributions by mixture, Journal of the Royal Statistical Society. Series B (Methodological) (1993) 409–422

    work page 1993

  39. [39]

    Ahmed, M

    N. Ahmed, M. Campbell, Fast Consistent Chernoff Fusion of Gaussian Mixtures for Ad Hoc Sensor Networks, IEEE Transactions on Signal Processing 60 (12) (2012) 6739–6745

    work page 2012

  40. [40]

    M. F. Huber, T. Bailey, H. Durrant-Whyte, U. D. Hanebeck, On en- tropy approximation for Gaussian mixture random vectors, in: 2008 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems, IEEE, 2008, pp. 181–188

    work page 2008

  41. [41]

    G. N. Vanderplaats, Numerical optimization techniques for engineering design, Vanderplaats Research and Development, Inc., 2001

    work page 2001

  42. [42]

    Ahmed, Conditionally factorized DDF for general distributed Bayesian estimation, in: 2014 Int’l Conf

    N. Ahmed, Conditionally factorized DDF for general distributed Bayesian estimation, in: 2014 Int’l Conf. on Multisensor Fusion and Information Integration for Intell. Systems (MFI 2014), 2014, pp. 1–7

    work page 2014

  43. [43]

    Liu, Monte Carlo Strategies in Scientific Computing, Springer, New York, 2001

    J. Liu, Monte Carlo Strategies in Scientific Computing, Springer, New York, 2001. 54

    work page 2001

  44. [44]

    R. M. Neal, G. E. Hinton, A view of the EM algorithm that justifies in- cremental, sparse, and other variants, in: Learning in graphical models, Springer, 1998, pp. 355–368

    work page 1998

  45. [45]

    Azevedo-Filho, R

    A. Azevedo-Filho, R. D. Shachter, Laplace’s method approximations for probabilistic inference in belief networks with continuous variables, in: Uncertainty Proceedings 1994, Elsevier, 1994, pp. 28–36

    work page 1994

  46. [46]

    A. Y. C. Kuk, Laplace importance sampling for generalized linear mixed models, Journal of Statistical Computation and Simulation 63 (2) (1999) 143–158. doi:10.1080/00949659908548522

    work page doi:10.1080/00949659908548522 1999

  47. [47]

    T. J. DiCiccio, R. E. Kass, A. Raftery, L. Wasserman, Computing bayes factors by combining simulation and asymptotic approximations, Jour- nal of the American Statistical Association 92 (439) (1997) 903–915

    work page 1997

  48. [48]

    Runnalls, Kullback-Leibler approach to Gaussian mixture reduction, IEEE Trans

    A. Runnalls, Kullback-Leibler approach to Gaussian mixture reduction, IEEE Trans. on Aerospace and Electronic Sys. 43 (3) (2007) 989–999

    work page 2007

  49. [49]

    Bourgault, Decentralized control in a Bayesian world, Ph.D

    F. Bourgault, Decentralized control in a Bayesian world, Ph.D. thesis, University of Sydney (2005)

    work page 2005

  50. [50]

    Brown, P

    R. Brown, P. Hwang, Introduction to random signals and applied Kalman filtering: with MATLAB exercises, J. Wiley & Sons, 2012

    work page 2012

  51. [51]

    Bar-Shalom, X

    Y. Bar-Shalom, X. Li, T.Kirubarajan, Estimation with Applications to Navigation and Tracking, Wiley, New York, 2001

    work page 2001

  52. [52]

    Boers, J

    Y. Boers, J. Driessen, Interacting multiple model particle filter, IEE Proc.-Radar Sonar Navig. 150 (5) (2003) 344–349

    work page 2003

  53. [53]

    Vo, B.-T

    B.-N. Vo, B.-T. Vo, R. Mahler, Closed-form solutions to forward- backward smoothing, IEEE Transactions on Signal Processing 60 (1) (2012) 2–17

    work page 2012

  54. [54]

    D. J. Lee, M. E. Campbell, Smoothing algorithm for nonlinear systems using Gaussian mixture models, Journal of Guidance, Control, and Dy- namics (2015) 1–14

    work page 2015

  55. [55]

    Vo, W.-K

    B.-N. Vo, W.-K. Ma, The Gaussian mixture probability hypothesis den- sity filter, IEEE Transactions on Signal Processing 54 (11) (2006) 4091– 4104. 55

    work page 2006

  56. [56]

    Vo, B.-N

    B.-T. Vo, B.-N. Vo, A. Cantoni, Analytic implementations of the car- dinalized probability hypothesis density filter, IEEE Transactions on Signal Processing 55 (7) (2007) 3553–3567

    work page 2007

  57. [57]

    ¨Uney, D

    M. ¨Uney, D. E. Clark, S. J. Julier, Distributed fusion of PHD filters via exponential mixture densities, IEEE Journal of Selected Topics in Signal Processing 7 (3) (2013) 521–531. 56

    work page 2013