Pith. sign in

REVIEW 6 major objections 6 minor 41 references

Infinite Factorial Linear Dynamical Systems for Transient Signal Detection

T0 review · 6 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A Bayesian nonparametric model of an unknown number of background emitters, plus a recursive factorial Kalman filter, makes finite moving average detection of transient signals computable online with bounded missed-detection and…

desk verdict The model and experiments are worth a look, but the central claim of exact closed-form performance bounds is not established: FKFF is an unlabeled approximation and Eq. (43) has an internal inconsistency. read the letter →

arxiv 2501.04973 v1 pith:4WT4VGAN submitted 2025-01-09 eess.SP

Jiadi Bao , Yatong Wang , Yunjie Li , Mengtao Zhu , Shafei Wang This is my paper · ORCID
classification eess.SP
keywords infinitefactoriallineardynamicalsystemstickyIndianbuffetprocesstransientsignaldetectionfinitemovingaveragetestKalmanforwardfilteringBayesiannonparametricparticleGibbswithancestorsamplingslice
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper addresses transient signal detection when the background is not a single source but a superposition of signals from multiple emitters, with the number of emitters unknown. It proposes to learn the background offline with the infinite factorial linear dynamical system (IFLDS), a Bayesian nonparametric model built on a sticky Indian buffet process, and then detect the signal online with a finite moving average (FMA) stopping time. To make the stopping time computable, it derives a factorial Kalman forward filtering recursion and a dependence structure that turns the log-likelihood ratio into a recursive statistic; it also gives closed-form Gaussian bounds for the worst-case false alarm and missed detection probabilities. The significance is that a detector can handle time-correlated multi-source backgrounds without knowing how many sources are present, with calibrated performance guarantees.

What carries the argument

The central machinery has three pieces. First, the IFLDS, a conjugate Bayesian nonparametric model in which a sticky Indian buffet process (a prior over infinite binary activity matrices with persistence) gives each emitter a Markov activity state and an LDS, allowing unbounded source number and temporally persistent sources. Second, the factorial Kalman forward filtering (FKFF), a recursive Gaussian update for the latent-state distribution of the FLDS; the recursion is made tractable by assuming the filtered covariance of the latent state remains block diagonal (Eq. 27), and it supplies the predictive likelihoods (Eq. 30) used in the detection statistic. Third, the FMA stopping time with the reformulated log-likelihood ratio of Eq. (34), which uses two parallel FKFF runs (signal present versus signal absent) and whose threshold is set through the closed-form bounds of Eq. (43).

What would settle it

For a single time step with $M=2$, compute the exact Gaussian posterior over the two latent sources given the observation and compare the off-diagonal covariance block to zero; the FKFF equations assume that block is zero. If it is nonzero and large enough to shift the predictive likelihood or the FMA threshold by a non-negligible amount, the exactness claim of the recursive likelihood computation fails. Concretely, simulate from the FLDS, run both the FKFF and a full-covariance or particle-filter approximation, and check whether the empirical false alarm and missed detection probabilities violate the Eq. (43) bounds for a fixed threshold.

Watch

Extended reading notes

Core claim

The central claim is that a composite background can be represented as an unbounded factorial linear dynamical system: each emitter has its own linear-Gaussian latent state, and the observation is the sum of the active emitters' outputs plus noise, with a sticky Indian buffet process prior deciding which emitters are active. Given that representation, the paper claims the factorial Kalman forward filtering equations (Eqs. 28-30) provide a recursive way to compute predictive data likelihoods, and that under Assumptions 1 and 2 the finite moving average stopping time (Eq. 33) is a valid detector whose worst-case missed detection and false alarm probabilities are bounded by the closed-form Gaussian expressions in Eq. (43). If correct, the detector can run online after an offline parameter-learning stage, estimate the number of background sources automatically, and deliver lower missed detection than single-source LDS, Gaussian, CUSUM, and Shewhart baselines in multi-source scenarios.

Load-bearing premise

The filtering equations assume that the latent sources stay uncorrelated with each other after each update, even though conditioning a sum of Gaussian sources on a shared observation generally makes their posterior correlated; if that approximation is poor, the likelihoods and the detection bounds inherit the error.

Editorial extensions

If this is right

  • The IFLDS parameter-learning method (slice sampling plus particle Gibbs with ancestor sampling) automatically estimates how many background sources are present, removing the need to specify the source count in advance.
  • FKFF gives an online recursive likelihood for a factorial state-space model, bypassing the NP-hard exact likelihood computation noted for factorial hidden Markov models.
  • The threshold-setting formula ties the window length (tolerable detection delay) to worst-case false alarm and missed detection bounds, so a system designer can set detection thresholds analytically rather than by simulation.
  • In multi-source simulations the proposed FLDS-FMA detector shows lower missed detection than LDS-FMA, Gaussian-FMA, CUSUM, and Shewhart baselines, with the gap widening as the number of background sources grows.
  • In the pulse-detection experiment under BPSK/QPSK communication interference, the method reports detection probabilities of 63% versus 41% for LDS-FMA and 33% for Gaussian-FMA at a fixed 15% false alarm rate.

Reading between the lines

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

  • A full-covariance correction to the FKFF would produce true predictive likelihoods; if the cross-source posterior covariance stays small in steady state the current bounds may be close, but this remains to be demonstrated.
  • The block-diagonal approximation may also explain the gap the authors report between the theoretical and empirical missed-detection curves under the signal-present hypothesis; an exact-filtering version could close that gap.
  • The same IFLDS representation could serve adjacent tasks such as emitter enumeration, source activity tracking, and change-point localization, not just binary transient detection.
  • A natural testable extension is to relax Assumption 2 and quantify how much violation of the independence assumption degrades the bounds; the paper states the assumption but does not probe its sensitivity.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

6 major / 6 minor

Summary. The paper proposes an infinite factorial linear dynamical system (IFLDS) with a sticky Indian buffet process prior to model an unknown number of background sources, a parameter-learning method based on slice sampling and particle Gibbs with ancestor sampling, and a finite moving average (FMA) stopping time for transient signal detection. The detection statistic uses a factorial Kalman forward filtering (FKFF) recursion to compute the null likelihood and a reformulated log-likelihood ratio. The paper claims closed-form Gaussian bounds on the worst-case false-alarm and missed-detection probabilities in Eq. (43), and validates the method on simulated data and on a pulse-detection experiment under communication interference.

Significance. The problem of detecting a finite-duration signal in a multi-source, time-correlated background with an unknown number of sources is well motivated, and the proposed pipeline is coherent at a high level: a Bayesian nonparametric prior over LDS sources, a PGAS/slice-sampling inference scheme, and an online FMA detector. The paper's recursive formulation of the detection statistic is a useful practical feature, and the experimental section covers both synthetic and communication-interference scenarios. However, the central advertised result is a closed-form performance guarantee for the FMA test, and that result is not supported as written: the FKFF likelihood is not exact under the stated model, the false-alarm bound relies on an unjustified independence assumption, and Eq. (43) contains an internal inconsistency. Because these issues affect the main theoretical claims, the contribution cannot be accepted in its current form.

major comments (6)
  1. [IV-A, Eq. (27)] The block-diagonal covariance parametrization is not preserved by exact Gaussian filtering for the FLDS observation model. Conditioning the joint state (x_t^1,...,x_t^M) on p_t couples the sources through the v-structure; with a block-diagonal prior, the exact posterior covariance has off-diagonal blocks of the form -P^{mm}(C^m)^T S^{-1} C^n P^{nn}, where S = sum_m C^m P^{mm}(C^m)^T + M R. Equations (28)-(30) therefore compute an approximate predictive likelihood, not the exact P_∞(p_t | p_1,...,p_{t-1}) used in the LLR (34). Lemma 1 establishes only Gaussianity, not the factorized covariance form, and the paper does not label FKFF as an approximation. Since Theorem 2, Corollary 1, and Eq. (43) are derived from this likelihood, the stated closed-form performance bounds are unsupported without an error analysis of the approximation.
  2. [IV-A, Eqs. (29)-(30)] The observation noise covariance used in the FKFF gain and predictive variance is inconsistent with model (2). In (2), n_t = sum_{m=1}^M (C^m x_t^m + v) with v ~ N(0,R); under the natural reading that each source contributes an independent noise term, the total observation noise covariance is M R, while Eq. (29) places M^2 R in the denominator and Eq. (30) adds M R inside the sum over m, again yielding M^2 R. If instead the same v is reused for all M sources, the covariance is M^2 R and this should be stated explicitly. Because the gain and predictive covariance determine the LLR numerically, this ambiguity changes the detection statistic and all subsequent bounds. The authors should specify the intended observation model and correct Eqs. (29)-(30) accordingly.
  3. [IV-C, Eq. (36)] The false-alarm bound α(h,w_α) = 1 - (P_∞(Ŵ_w < h))^{w_α} treats the overlapping window sums Ŵ_t as independent. The event in (5) is the existence of t in an interval of length w_α with Ŵ_t ≥ h, and consecutive Ŵ_t share w-1 observations, so they are strongly dependent. Without a mixing or independence argument, the product form is not a valid upper bound; a union bound would instead require summing marginal tail probabilities over t. Because Corollary 1 and the first line of Eq. (43) use this formula to set the threshold h, the false-alarm guarantee P_FA ≤ ᾱ is not established.
  4. [IV-C, Eq. (43), third line] The false-alarm expression uses μ_H1 in the Gaussian CDF, but under H0 the accumulated statistic Ŵ_t has mean w μ_H0, not w μ_H1. Thus the displayed formula is not a false-alarm bound. If the intended correction is to use μ_H0 and σ²_H0, the threshold h produced by Eq. (43) would change, and the numerical comparisons in Section V would need to be recomputed. This internal inconsistency independently prevents the stated formulas from being used for detector calibration.
  5. [IV-C, Lemma 2 and Eq. (43)] The convergence of e_t in (42) is asserted only on the basis of Fig. 4, which uses the same parameter settings as the evaluation in Section V-A, and the paper explicitly states that a theoretical investigation of the convergence is beyond its scope. The closed-form missed-detection bound therefore depends on a numerically calibrated steady-state value of e_t and μ_H1 rather than on a derived property of the model. In addition, the real-data experiment in Section VI sets h empirically to maintain a 15% false-alarm rate, so the theoretically predicted threshold from (39)-(43) is not actually validated. The dependence of the claimed bounds on simulation-calibrated quantities should be stated clearly, and the bounds should not be presented as closed-form model-based guarantees.
  6. [IV-B, IV-C; Supplemental Material] Several load-bearing proofs are deferred to an unavailable supplement: the derivation of the FKFF update (28), the proof of Theorem 1, the proof of Theorem 2, and the proof of Lemma 2 (Supplement Sections I-IV). The manuscript is not self-contained, and the central claims cannot be verified from the submitted text. In particular, Theorem 1 relies on Assumption 2, which states that the latent variable at arrival time ν is independent of the latent variable at ν-1 but depends on previous observations; this is not a consequence of the FLDS transition (2) or of Assumption 1, and no derivation is given in the main text. The omitted proofs and the unjustified assumption should be addressed before the theoretical claims can be evaluated.
minor comments (6)
  1. [IV-A, Eq. (28)] The update for μ̂_t^m uses x_{t-1}^m inside the innovation term where μ̂_{t-1}^m is evidently intended; as written, the recursion is not closed and cannot be implemented without additional definitions.
  2. [V-A, Eq. (44)] The same matrix C^m is given for all m, so the simulated background sources differ only in their transition matrices G^m; the text says the output matrices are defined for each source, so the formula should be reconciled with the statement.
  3. [V-C, Fig. 6] The x-axis is labeled SINR, but the conversion from the constant signal amplitude y_t (between 0.001 and 1.5) to SINR in dB is not defined; without this mapping the experimental curves cannot be reproduced.
  4. [V-B, Fig. 5] The text refers to 'MSE' while the defined metric is reconstruction error RE; please use consistent notation throughout the figure and the discussion.
  5. [VI, Fig. 9] The threshold h is said to be adjusted to maintain a 15% false-alarm probability, but the corresponding theoretical threshold from Section IV-C is not reported; reporting both would clarify the discrepancy between the empirical calibration and the theoretical prediction.
  6. [Figures 4-9] Several figure captions and equations contain garbled unicode strings (e.g., the caption lines beginning '/uni00000013...' near Figs. 4-9), indicating a rendering or encoding problem; the final version should be regenerated with correct math and text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the detection bounds use model-based approximations and numerically converged constants, but they are not defined in terms of the results they predict.

full rationale

The paper's derivation chain is not circular in the sense of Eq. X reducing to Eq. Y by construction or a fitted parameter being renamed as a prediction. The FKFF recursion (Eqs. 28-30) is presented as a likelihood computation for the FLDS; even though the block-diagonal covariance parametrization (Eq. 27) is an approximation rather than an exact Gaussian filter, the likelihood it produces is an input to the FMA statistic, not the target result, so this is a correctness/rigor concern rather than circularity. The FMA stopping time (Eq. 33) and the bounds in Theorem 2 and Corollary 1 follow from standard FMA arguments and from the definition of the detection statistic; they do not presuppose the reported detection performance. The H1 parameter μ_H1 in Eq. (43) is obtained from the recursion in Eq. (42), which is model-based and not fitted to the empirical missed-detection probabilities; the paper explicitly says a theoretical convergence analysis is beyond its scope, which is a limitation but not a circular reduction. The false-alarm formula in Eq. (43) appears to use μ_H1 where μ_H0 is needed, but that is an internal inconsistency, not circularity. The real-data threshold in Section VI is adjusted empirically, but the paper does not claim that threshold is set by the closed-form formula. Self-citations are peripheral and not load-bearing. The central claims therefore have independent content, though their exactness is questionable and several proofs are deferred to an unavailable supplement.

Assumptions & free parameters 4 free parameters · 8 assumptions · 2 invented entities

The central claim leans on several assumptions beyond the standard Bayesian machinery: the block-diagonal covariance in the FKFF, the arrival-time independence in Assumption 2, the convergence of e_t, and the independence of window statistics in the false alarm bound. These are either unstated approximations or empirically verified rather than proved.

free parameters (4)
  • IFLDS hyperparameters = alpha=1, beta0=2, beta1=0.1, gamma0=10, gamma1=1, M0=0, K0=I, n0=4, S0=0.75*Sbar
    Chosen by hand in Section V-B; the estimated source number and reconstruction error depend on these values.
  • Steady-state observation error e_t / mu_H1 = Converged value from simulation (Fig. 4)
    Used in Lemma 2 and Eq. (43) to define the theoretical missed-detection bound; convergence is verified numerically, not proved.
  • Detection threshold h = Adjusted to P_FA=15% in Section VI
    In the real-data experiment the threshold is set empirically to hit a target false alarm rate rather than from Eq. (39).
  • Window length w = 200 in simulation and experiment
    User-specified tolerable delay; performance curves in Fig. 7 treat it as a design knob rather than an estimated quantity.
assumptions (8)
  • ad hoc to paper The forward variable remains a multivariate Gaussian with block-diagonal covariance (Eq. 27).
    Needed for the FKFF update equations (28)-(29); not derived from the model and false in exact Gaussian filtering when observations sum several sources.
  • ad hoc to paper Assumption 2: the latent variable at arrival time nu is independent of the previous latent state but depends on earlier observations.
    Introduced in Section IV-B to obtain Theorem 1 and make the LLR recursive; it changes the generative model at the arrival time.
  • domain assumption Assumption 1: time index t approaches infinity so initial conditions are forgotten.
    Used with Assumption 2 in Theorem 1; reasonable for long records but not for short detection windows.
  • ad hoc to paper The observation noise summation uses per-source R, giving total covariance M R, but the FKFF formulas use M^2 R in several places.
    The covariance bookkeeping in Eqs. (29), (30) and (42) is inconsistent with Eq. (2) and affects the likelihood scaling.
  • ad hoc to paper The per-window statistics W_t are treated as independent for the false alarm bound in Eq. (36).
    Overlapping windows are dependent; no mixing or dependence correction is given.
  • ad hoc to paper The observation error e_t in Lemma 2 converges to a steady state.
    The paper states that a theoretical convergence investigation is beyond scope; Eq. (43) uses the numerically observed converged value.
  • standard math Gaussian densities are closed under product and convolution.
    Used in the proof outline of Lemma 1 to claim the forward variable is Gaussian.
  • standard math The algebraic Riccati equation has a stabilizing solution for the Kalman gain.
    Invoked in Section IV-C to justify convergence of the Kalman gain and covariance.
invented entities (2)
  • Sticky indicator variable z_t^m
    purpose: Controls whether the activity Markov chain s_t^m persists in its previous state (Eqs. 9-10).
    A model device with no external falsifiable prediction; it improves source-number estimation in simulations but is not independently testable.
  • Auxiliary slice variable theta
    purpose: Used in slice sampling to decide how many parallel chains are active (Eq. 14).
    Algorithmic auxiliary variable with no physical meaning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Infinite Factorial Linear Dynamical Systems for Transient Signal Detection." pith.science (2026). https://pith.science/paper/4WT4VGAN

@misc{pith2026250104973,
  author       = {Pith},
  title        = {Pith review of: Infinite Factorial Linear Dynamical Systems for Transient Signal Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4WT4VGAN}},
  note         = {Machine review of arXiv:2501.04973}
}
read the original abstract

Accurately detecting the transient signal of interest from the background signal is one of the fundamental tasks in signal processing. The most recent approaches assume the existence of a single background source and represent the background signal using a linear dynamical system (LDS). This assumption might fail to capture the complexities of modern electromagnetic environments with multiple sources. To address this limitation, this paper proposes a method for detecting the transient signal in a background composed of an unknown number of emitters. The proposed method consists of two main tasks. First, a Bayesian nonparametric model called the infinite factorial linear dynamical system (IFLDS) is developed. The developed model is based on the sticky Indian buffet process and enables the representation and parameter learning of the unbounded number of background sources. This study also designs a parameter learning method for the IFLDS using slice sampling and particle Gibbs with ancestor sampling. Second, the finite moving average (FMA) stopping time is introduced to minimize the worst-case probability of missed detection, and the statistical performance of the stopping time is investigated. To facilitate the computation of the FMA stopping time, this study derives the factorial Kalman forward filtering (FKFF) method and designs a dependence structure for the underlying model, allowing the stopping time to be defined by a recursive function. Numerical simulations demonstrate the effectiveness of the proposed method and the validity of the theoretical results. The experimental results of the pulse signal detection under the condition of communication interference confirm the effectiveness and superiority of the proposed method.

Figures

Figures reproduced from arXiv: 2501.04973 by the authors.

Figure 1
Figure 1. The diagram of the RPL and TSD tasks in the presence of two [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The graphical representation of the FLDS. (The colored nodes denote [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The graphical representation of the IFLDS; variable [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The observation error calculated by (42). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The estimation result of the background source number and the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Detection performance for different background source numbers. The red line in each figure represents the theoretical upper bound of the proposed [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: The PMD versus PF A curves of the FLDS-FMA method and baseline methods for various values of window length w. There were four background sources, and the SOI amplitude was 0.48, which corresponded to -12.03 dB; (a): w = 10, (b): w = 100, (c): w = 200. 200 600 1000 1400…
Figure 8
Figure 8. Figure 8: The distributions of the FLDS-FMA detection statistics for cases when the signal was absent and present. The theoretical distributions under the null [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: In the first and second panels of [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 9
Figure 9. Figure 9: The IQ data (the first and second panels) and the detection statistic (the third to fifth panels); the third panel: FLDS-FMA statistic; the fourth panel: [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 38 canonical work pages

  1. [1]

    Energy detection of unknown determin- istic signals,

    H. Urkowitz, “Energy detection of unknown determin- istic signals,” Proc. IEEE, vol. 55, no. 4, pp. 523–531, 1967

    work page 1967

  2. [2]

    Tartakovsky, I

    A. Tartakovsky, I. Nikiforov, and M. Basseville, Se- quential analysis: Hypothesis testing and changepoint detection. CRC press, 2014

    work page 2014

  3. [3]

    Detection in correlated impulsive noise using fourth-order cumulants,

    B. M. Sadler, “Detection in correlated impulsive noise using fourth-order cumulants,” IEEE Trans. Signal Pro- cess., vol. 44, no. 11, pp. 2793–2800, 1996

    work page 1996

  4. [4]

    Memoryless discrete-time signal detection in long-range dependent noise,

    X. Yang, H. V . Poor, and A. P. Petropulu, “Memoryless discrete-time signal detection in long-range dependent noise,” IEEE Trans. Signal Process. , vol. 52, no. 6, pp. 1607–1619, 2004

    work page 2004

  5. [5]

    S. M. Kay, Fundamentals of statistical signal processing: estimation theory. Prentice-Hall, Inc., 1993

    1993

  6. [6]

    Memoryless discrete-time de- tection of a constant signal in m-dependent noise,

    H. Poor and J. Thomas, “Memoryless discrete-time de- tection of a constant signal in m-dependent noise,” IEEE Trans. Inf. Theory., vol. 25, no. 1, pp. 54–61, 1979

    work page 1979

  7. [7]

    Unknown signal detection in linear dynamical system noise,

    G. Ford, T. Hu, D. Bucci, and B. Foster, “Unknown signal detection in linear dynamical system noise,” in Proc. IEEE 29th Int. Workshop Mach. Learn. Signal Process. IEEE, 2019, pp. 1–6

    work page 2019

  8. [8]

    Unknown signal detection in interference and noise using hidden markov models,

    G. Ford, B. J. Foster, M. J. Liston, and M. Kam, “Unknown signal detection in interference and noise using hidden markov models,” in Proc. IEEE Stat. Signal Process. Workshop. IEEE, 2021, pp. 406–410

    work page 2021

Show all 41 references
  1. [9]

    Unknown signal detection in switching linear dynamical system noise,

    G. Ford, B. J. Foster, S. A. Braun, and M. Kam, “Unknown signal detection in switching linear dynamical system noise,” IEEE Trans. Signal Process. , vol. 71, pp. 2220–2234, 2023. JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, APRIL 2024 13

    2023

  2. [10]

    The indian buffet process: An introduction and review

    T. L. Griffiths and Z. Ghahramani, “The indian buffet process: An introduction and review.” J. Mach. Learn. Res., vol. 12, no. 4, 2011

    2011

  3. [11]

    Stick-breaking construction for the indian buffet process,

    Y . W. Teh, D. Gr¨ur, and Z. Ghahramani, “Stick-breaking construction for the indian buffet process,” in Proc. 11th Conf. A.I. Statist. PMLR, 2007, pp. 556–563

    2007

  4. [12]

    Factorial hidden markov models,

    Z. Ghahramani and M. Jordan, “Factorial hidden markov models,” Adv. Neural Inf. Process. Syst. , vol. 8, 1995

    1995

  5. [13]

    The infinite fac- torial hidden markov model,

    J. Gael, Y . Teh, and Z. Ghahramani, “The infinite fac- torial hidden markov model,” Adv. Neural Inf. Process. Syst., vol. 21, 2008

    2008

  6. [14]

    Infinite factorial dynamical model,

    I. Valera, F. Ruiz, L. Svensson, and F. Perez-Cruz, “Infinite factorial dynamical model,” Adv. Neural Inf. Process. Syst., vol. 28, 2015

    2015

  7. [15]

    Sequential detection of transient changes,

    B. K. Gu ´epi´e, L. Fillatre, and I. Nikiforov, “Sequential detection of transient changes,” Seq. Anal., vol. 31, no. 4, pp. 528–547, 2012

    2012

  8. [16]

    Optimal sequential detection of signals with unknown appearance and disappearance points in time,

    A. G. Tartakovsky, N. R. Berenkov, A. E. Kolessa, and I. V . Nikiforov, “Optimal sequential detection of signals with unknown appearance and disappearance points in time,” IEEE Trans. Signal Process. , vol. 69, pp. 2653– 2662, 2021

    2021

  9. [17]

    Reliable detection of faults in measurement systems,

    B. Bakhache and I. Nikiforov, “Reliable detection of faults in measurement systems,” Int. J. Adapt. Control Signal Process., vol. 14, no. 7, pp. 683–700, 2000

    2000

  10. [18]

    Security of scada systems against cyber–physical attacks,

    L. Fillatre, I. Nikiforov, P. Willett et al. , “Security of scada systems against cyber–physical attacks,” IEEE Aerosp. Electron. Syst. Mag. , vol. 32, no. 5, pp. 28–45, 2017

    2017

  11. [19]

    Jdmr-net: Joint detection and modulation recognition networks for lpi radar signals,

    Z. Zhang, M. Zhu, Y . Li, and S. Wang, “Jdmr-net: Joint detection and modulation recognition networks for lpi radar signals,” IEEE Trans. Aerosp. Electron. Syst. , vol. 59, no. 6, pp. 7575–7589, 2023

    2023

  12. [20]

    Quickest change detection in au- toregressive models,

    Z. Sun and S. Zou, “Quickest change detection in au- toregressive models,” IEEE Trans. Inf. Theory. , 2024

    2024

  13. [21]

    Quickest change detection and kullback-leibler divergence for two-state hidden markov models,

    C.-D. Fuh and Y . Mei, “Quickest change detection and kullback-leibler divergence for two-state hidden markov models,” IEEE Trans. Signal Process. , vol. 63, no. 18, pp. 4866–4878, 2015

    2015

  14. [22]

    Asymptotically optimal change point detec- tion for composite hypothesis in state space models,

    C.-D. Fuh, “Asymptotically optimal change point detec- tion for composite hypothesis in state space models,” IEEE Trans. Inf. Theory. , vol. 67, no. 1, pp. 485–505, 2020

    2020

  15. [23]

    A review on kalman filter models,

    M. Khodarahmi and V . Maihami, “A review on kalman filter models,” Arch. Comput. Methods Eng. , vol. 30, no. 1, pp. 727–747, 2023

    2023

  16. [24]

    C. M. Bishop and N. M. Nasrabadi, Pattern recognition and machine learning . Springer, 2006, vol. 4, no. 4

    2006

  17. [25]

    A sticky hdp-hmm with application to speaker diarization,

    E. B. Fox, E. B. Sudderth, M. I. Jordan, and A. S. Willsky, “A sticky hdp-hmm with application to speaker diarization,” Ann. Appl. Stat. , pp. 1020–1056, 2011

    2011

  18. [26]

    Sticky hidden markov modeling of comparative genomic hybridiza- tion,

    L. Du, M. Chen, J. Lucas, and L. Carin, “Sticky hidden markov modeling of comparative genomic hybridiza- tion,” IEEE Trans. Signal Process. , vol. 58, no. 10, pp. 5353–5368, 2010

    2010

  19. [27]

    Slice sampling,

    R. M. Neal, “Slice sampling,” Ann. Stat., vol. 31, no. 3, pp. 705–767, 2003

    2003

  20. [28]

    Particle gibbs with ancestor sampling,

    F. Lindsten, M. I. Jordan, and T. B. Schon, “Particle gibbs with ancestor sampling,” J. Mach. Learn. Res. , vol. 15, pp. 2145–2184, 2014

    2014

  21. [29]

    Infinite factorial finite state machine for blind multiuser channel estimation,

    F. J. Ruiz, I. Valera, L. Svensson, and F. Perez-Cruz, “Infinite factorial finite state machine for blind multiuser channel estimation,” IEEE Trans. Cogn. Commun. Netw., vol. 4, no. 2, pp. 177–191, 2018

    2018

  22. [30]

    A note on particle gibbs method and its ex- tensions and variants,

    N. Gauraha, “A note on particle gibbs method and its ex- tensions and variants,” arXiv preprint arXiv:2007.15862, 2020

    2007 arXiv

  23. [31]

    Information bounds and quick detection of parameter changes in stochastic systems,

    T. L. Lai, “Information bounds and quick detection of parameter changes in stochastic systems,” IEEE Trans. Inf. Theory., vol. 44, no. 7, pp. 2917–2929, 1998

    1998

  24. [32]

    Performance bounds for finite moving average tests in transient change detection,

    D. Egea-Roca, J. A. L ´opez-Salcedo, G. Seco-Granados, and H. V . Poor, “Performance bounds for finite moving average tests in transient change detection,” IEEE Trans. Signal Process., vol. 66, no. 6, pp. 1594–1606, 2018

    2018

  25. [33]

    Two strategies in transient change detection,

    D. Egea-Roca, B. K. Gu ´epi´e, J. A. L ´opez-Salcedo, G. Seco-Granados, and I. V . Nikiforov, “Two strategies in transient change detection,” IEEE Trans. Signal Process., vol. 70, pp. 1418–1433, 2022

    2022

  26. [34]

    Chopin, O

    N. Chopin, O. Papaspiliopoulos et al., An introduction to sequential Monte Carlo . Springer, 2020, vol. 4

    2020

  27. [35]

    Koller and N

    D. Koller and N. Friedman, Probabilistic graphical mod- els: principles and techniques . MIT press, 2009

    2009

  28. [36]

    Lancaster, Algebraic Riccati Equations

    P. Lancaster, Algebraic Riccati Equations . Oxford Science Publications/The Clarendon Press, Oxford Uni- versity Press, 1995

    1995

  29. [37]

    Statis- tical and computational guarantees for the baum-welch algorithm,

    F. Yang, S. Balakrishnan, and M. J. Wainwright, “Statis- tical and computational guarantees for the baum-welch algorithm,” J. Mach. Learn. Res. , vol. 18, no. 125, pp. 1–53, 2017

    2017

  30. [38]

    Error bounds for convolutional codes and an asymptotically optimum decoding algorithm,

    A. Viterbi, “Error bounds for convolutional codes and an asymptotically optimum decoding algorithm,” IEEE Trans. Inf. Theory., vol. 13, no. 2, pp. 260–269, 1967

    1967

  31. [39]

    Parameter estimation for linear dynamical systems,

    Z. Ghahramani and G. E. Hinton, “Parameter estimation for linear dynamical systems,” Tech. Rep., 1996

    1996

  32. [40]

    Wald, Sequential Anal

    A. Wald, Sequential Anal. Courier Corporation, 2004

    2004

  33. [41]

    Multiple optimality properties of the shewhart test,

    G. V . Moustakides, “Multiple optimality properties of the shewhart test,” Sequential Anal., vol. 33, no. 3, pp. 318– 344, 2014

    2014

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.