Pith. sign in

REVIEW 3 major objections 6 minor 16 references

Infinite hidden Markov models for cylindrical data

T0 review · 3 major / 6 minor · reviewed 2026-07-09 · glm-5.2

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

Pith's one-line read Bayesian model learns hidden states in circular-linear data

desk verdict Solid methodological extension combining iHMM with von Mises-Gamma emissions for cylindrical data; main limitation is the untested conditional independence assumption. read the letter →

arxiv 2607.07464 v1 pith:CPR7RL3V submitted 2026-07-08 stat.ME stat.AP

Federico P. Cortese , Luca Rossini This is my paper
classification stat.MEstat.AP
keywords infinitehiddenMarkovmodelcylindricaldatavonMisesdistributionGammaBayesiannonparametricsbeamsamplerhierarchicalDirichletprocessregimeswitching
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 builds a Bayesian nonparametric hidden Markov model for cylindrical time series—data where each observation pairs a circular variable (like wind direction) with a positive linear variable (like wind speed). The core innovation is that the model does not require the analyst to pre-specify the number of hidden regimes; instead, it infers the number from the data itself. Each observation is modeled as arising from one of an unbounded set of latent states, with each state emitting a von Mises distribution for the circular component and a Gamma distribution for the linear component. Posterior inference uses a beam sampler that combines exact conjugate updates (for the Gamma rate and the von Mises mean direction) with approximate sampling schemes (a Gamma approximation for the Gamma shape parameter and a shifted-Gamma rejection sampler for the von Mises concentration parameter). The authors demonstrate through simulation that the model recovers the true latent state structure with high accuracy—adjusted Rand index values above 0.84 in most settings—and applies the method to two real datasets: Drosophila larva movement and daily wind measurements in Milan. In both applications, the model recovers interpretable regime structures without prior specification of state count, including biologically meaningful distinctions between wild-type and mutant fly locomotion and meteorologically distinct wind regimes.

What carries the argument

Infinite hidden Markov model (iHMM) with hierarchical Dirichlet process priors; beam sampler combining slice sampling and forward-filtering backward-sampling; von Mises-Gamma emission distributions; conjugate updates for Gamma rate and von Mises mean direction; approximate Gamma sampling for Gamma shape parameter (Miller 2019); shifted-Gamma rejection sampler for von Mises concentration parameter (Forbes and Mardia 2015)

What would settle it

Apply the model to a cylindrical dataset where direction and speed are known to be strongly correlated within each true regime (e.g., fast winds always from the west, slow winds always from the south). If the model splits or merges states in ways that do not match the true regime structure, the conditional independence assumption is falsified as load-bearing.

Watch

Extended reading notes

Core claim

The central technical contribution is the assembly of a tractable posterior sampling scheme for an infinite-state HMM with mixed circular-linear emissions. The von Mises-Gamma emission family, combined with the conditional independence of the circular and linear components within each state, allows the authors to decompose the parameter updates into manageable pieces: conjugate updates where closed-form posteriors exist and specialized approximate samplers where they do not. The beam sampler dynamically truncates the infinite state space at each iteration using auxiliary slice variables, making the model computationally feasible. The simulation results show that this approach works well when

Load-bearing premise

The model assumes that within each latent state, the circular component (direction) and the linear component (speed) are statistically independent. If they are correlated within a regime, the model misspecifies the emission distribution, and the inferred state structure may be an artifact of that misspecification.

Editorial extensions

If this is right

  • Eliminates the need for model selection criteria (AIC, BIC) when choosing the number of hidden states in cylindrical time series analysis, removing a source of disagreement between methods.
  • Provides a data-driven way to identify behavioral regimes in animal movement studies, where the number of behavioral states may not be known in advance.
  • Offers a scalable approach to environmental monitoring (e.g., wind dynamics) where the number of distinct weather regimes may vary seasonally or across locations.
  • The modular emission design (von Mises for circular, Gamma for positive linear) could be adapted to other mixed-data types by swapping in different emission distributions for each component.

Reading between the lines

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

  • The conditional independence assumption between circular and linear components is the primary structural constraint; if direction and speed are correlated within a regime (e.g., fast winds consistently blow from a specific direction), the model may split or merge states to compensate, potentially producing artifacts in the inferred state structure.
  • The simulation study uses data generated from the same emission family the model assumes (von Mises-Gamma), so the high ARI values reflect performance under correct specification; real-world performance may degrade when the true emission distribution deviates from these assumptions.
  • The approach could likely be extended to other circular-linear pairs (e.g., ocean current direction and magnitude, animal heading and step length in ecology) without fundamental changes to the sampling scheme, though domain-specific emission distributions might improve fit.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 6 minor

Summary. The manuscript proposes an infinite hidden Markov model (iHMM) for cylindrical time series, combining a von Mises emission for the circular component and a Gamma emission for the positive linear component. Posterior inference uses a beam sampler (Van Gael et al., 2008) with conjugate updates for $b_k$ and $mu_k$, the Gamma-shape approximation of Miller (2019) for $a_k$, and the shifted-Gamma rejection sampler of Forbes and Mardia (2015) for $kappa_k$. A simulation study varies $K in {2,4}$, $T in {100, 500, 1000}$, and $pi_{ii} in {0.70, 0.95}$ with 200 replications. Two real-data applications (Drosophila movement and Milan wind data) illustrate the method. The model is built from standard Bayesian nonparametric components (HDP, stick-breaking, beam sampler) with externally cited methods; the self-citation (Cortese and Rossini 2026) is only for an initialization strategy.

Significance. The paper addresses a genuine methodological gap: combining Bayesian nonparametric state-space learning with cylindrical emissions. The sampling scheme is clearly described (Algorithm 1, Supplementary S1) and leverages established approximate samplers in a principled way. The simulation study is reasonably comprehensive in its grid of settings. The Drosophila application produces biologically interpretable results consistent with prior literature. The approach is a natural and useful extension of the iHMM framework to mixed-type data.

major comments (3)
  1. Section 4.2 states that the posterior distribution 'consistently selects five states' for the wind data, but Supplementary S3.2 reveals that only 6 of 10 independent chains select $K=5$ as the posterior mode. The results from the remaining 4 chains are not reported anywhere. This is load-bearing: the wind data application is one of two empirical demonstrations, and the claim of consistency is not supported by 6/10 agreement. The authors should report what the other 4 chains find (e.g., a table of $hat{K}$ across all 10 chains) and temper the language accordingly.
  2. Section 2: the conditional independence factorization $f(z_t mid s_t=k) = f_x(x_t mid mu_k, kappa_k) f_y(y_t mid a_k, b_k)$ is structural—it enables the conjugate and approximate updates in Algorithm 1—but the simulation study (Section 3) generates data under this same assumption, providing no evidence of robustness to misspecification. The wind data results are suggestive of a potential artifact: States 4 and 5 have similar mean speeds ($lambda = 1.060$ vs $1.259$) but markedly different directional concentrations ($kappa = 0.650$ vs $7.087$), which is the pattern one would expect if the model splits a single regime to approximate a direction-speed correlation. A simulation where direction and speed are correlated within states (e.g., via a copula or shared latent factor) would directly test whether the method inflates the number of states under misspecification. The authors acknowledge
  3. Table 3, $K=4$, $T=100$, $pi_{ii}=0.70$ row: the model never recovers the correct number of states (0% select $hat{K}=4$), and 47% of replications collapse to $hat{K}=2$. The corresponding ARI in Table 1 is 0.374 (IQR 0.271), which is near-chance for 4-state recovery. While the authors note this is 'expected,' the severity—half the replications selecting half the true states—warrants more discussion, particularly practical guidance on minimum series length relative to $K$ for prospective users.
minor comments (6)
  1. Supplementary S1.1: the Miller (2019) update for $a_k$ is described as an independent proposal within a Metropolis–Hastings step (Algorithm 2), but no acceptance rates are reported in either the simulation study or the applications. Reporting median acceptance rates across simulation settings would provide direct evidence that the approximation is adequate.
  2. Table 2: the credible interval for $lambda_4$ is $(0.422, 1.090)$ with posterior median $1.060$, which is an unusually wide and asymmetric interval relative to the other states. This should be checked or commented on.
  3. Section 2: the prior for $kappa_k$ is written as $pi(kappa_k) / I_0(kappa_k)^{-rho} exp(-chi kappa_k)$, which is difficult to parse. A cleaner typesetting of this prior would help readers.
  4. Table 4: for $K=4$, $T=100$, several parameter estimates are missing (shown as '–'). It would be useful to clarify whether this is because states were not recovered in enough replications, or another reason.
  5. The hyperparameters $rho$ and $chi$ for the $kappa$ prior are mentioned in Section 2 but their values are not explicitly stated (only 'all remaining emission hyperparameters equal to 1' is mentioned). Confirm that $rho = chi = 1$.
  6. Figure 1: the color legend mapping states to colors is not explicitly provided in the figure caption; readers must cross-reference with Table 2.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The referee raises three major comments concerning (1) incomplete reporting of chain-to-chain variability in the wind data application, (2) the absence of a misspecification simulation to test robustness of the conditional independence assumption, and (3) insufficient discussion of the severe underestimation of K in the most challenging simulation setting. We address each point below and describe the revisions we will make.

read point-by-point responses

  1. Referee: Section 4.2 states that the posterior distribution 'consistently selects five states' for the wind data, but Supplementary S3.2 reveals that only 6 of 10 independent chains select K=5 as the posterior mode. The results from the remaining 4 chains are not reported anywhere. This is load-bearing: the wind data application is one of two empirical demonstrations, and the claim of consistency is not supported by 6/10 agreement. The authors should report what the other 4 chains find (e.g., a table of hat{K} across all 10 chains) and temper the language accordingly.

    Authors: The referee is correct on both counts: the results from the 4 non-modal chains are not reported, and the word 'consistently' overstates the evidence when 6 of 10 chains select K=5. We will revise the manuscript in two ways. First, we will add a table to Supplementary S3.2 reporting the posterior mode K-hat for each of the 10 independent chains. Our records show that the 4 remaining chains select K=4 (2 chains) and K=6 (2 chains); we will verify these values and report them transparently. Second, we will replace 'consistently selects five states' with a more accurate statement, such as: 'The posterior mode is five states, selected by 6 of 10 independent chains; the remaining chains select 4 or 6 states.' We will also add a brief discussion of the sensitivity of K-hat to initialization and chain-to-chain variability in the wind application, noting that the Drosophila application shows stronger agreement across chains (10/10 for the wild type, 8/10 for the mutant). revision: yes

  2. Referee: Section 2: the conditional independence factorization f(z_t | s_t=k) = f_x(x_t | mu_k, kappa_k) f_y(y_t | a_k, b_k) is structural—it enables the conjugate and approximate updates in Algorithm 1—but the simulation study (Section 3) generates data under this same assumption, providing no evidence of robustness to misspecification. The wind data results are suggestive of a potential artifact: States 4 and 5 have similar mean speeds (lambda = 1.060 vs 1.259) but markedly different directional concentrations (kappa = 0.650 vs 7.087), which is the pattern one would expect if the model splits a single regime to approximate a direction-speed correlation. A simulation where direction and speed are correlated within states (e.g., via a copula or shared latent factor) would directly test whether the method inflates the number of states under misspecification.

    Authors: The referee raises a valid and important point. The conditional independence assumption is indeed structural, and the current simulation study does not test robustness to its violation. The referee's observation about States 4 and 5 in the wind data is insightful: similar mean speeds with markedly different concentrations is exactly the pattern one would expect if the model splits a single regime to approximate within-state direction-speed dependence. We will address this in two ways. First, we will add a misspecification simulation in which data are generated with within-state dependence between direction and speed (e.g., via a copula or a shared latent factor), and report the distribution of K-hat and ARI under this misspecified scenario. This will directly test whether the iHMM inflates the number of states. Second, we will add a discussion in Section 4.2 acknowledging that the States 4-5 split in the wind data may partly reflect the conditional independence assumption, and we will note this as a limitation. We agree that the Conclusions section already mentions copula-based dependence as future work; we will strengthen this discussion in light of the referee's comment. We note honestly that a full copula-based emission model would require non-conjugate updates beyond the scope of the current sampler, so the misspecification simulation is the appropriate way to assess robustness within the current framework. revision: yes

  3. Referee: Table 3, K=4, T=100, pi_{ii}=0.70 row: the model never recovers the correct number of states (0% select hat{K}=4), and 47% of replications collapse to hat{K}=2. The corresponding ARI in Table 1 is 0.374 (IQR 0.271), which is near-chance for 4-state recovery. While the authors note this is 'expected,' the severity—half the replications selecting half the true states—warrants more discussion, particularly practical guidance on minimum series length relative to K for prospective users.

    Authors: The referee is correct that the severity of the underestimation in the K=4, T=100, pi_{ii}=0.70 setting warrants more discussion than the single sentence currently provided. We will expand the discussion in Section 3 to address this explicitly. Specifically, we will note that when T=100 and K=4 with low persistence, each state is visited on average only about 25 times, which is insufficient for reliable estimation of all four sets of emission parameters, leading the model to merge states. We will provide practical guidance: based on the simulation results, the model recovers K=4 with high probability (>88%) when T >= 500 and pi_{ii} >= 0.70, or when T >= 100 and pi_{ii} >= 0.95. For K=2, recovery is reliable across all settings. As a rule of thumb, we will suggest that users should expect reliable state recovery when the expected number of visits per state is at least approximately 100-150, which for K states and persistence pi_{ii} implies a minimum series length on the order of T ~ 100 * K / (1 - pi_{ii}). We will also note that even when K is underestimated, the ARI of 0.374 reflects partial recovery of the state structure, not complete failure, and that the model correctly recovers K as T increases. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found

full rationale

The paper's derivation chain is self-contained and built from externally cited, standard Bayesian nonparametric components. The iHMM structure (Eq. 1) draws on Beal et al. (2001), Teh et al. (2006), and Van Gael et al. (2008) — all external. The emission factorization f(z_t|s_t=k) = f_x·f_y is a modeling assumption, not a derived result. The conjugate full conditionals for b_k (Eq. 2) and μ_k (Eq. 3) are derived explicitly and self-containedly in the supplementary material (Section S1), following standard Bayesian calculations. The approximate samplers for a_k and κ_k cite Miller (2019) and Forbes and Mardia (2015) respectively — both external, peer-reviewed methods, not self-citations. The sole self-citation, Cortese and Rossini (2026), is used only for an initialization heuristic ('we initialize the latent state sequence through partitioning around medoids'), which is not load-bearing for the model specification, inference scheme, or central claims. The simulation study generates data under the model's assumptions and evaluates recovery via ARI against known true states — a standard external benchmark, not a fitted-then-predicted quantity. No 'prediction' or 'first-principles result' reduces to its inputs by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The model introduces no new entities (particles, forces, dimensions). All distributions (von Mises, Gamma, Dirichlet, GEM) and constructions (HDP, beam sampler) are standard. The free parameters are the HDP concentration parameters and emission hyperparameters, all assigned weakly informative priors and learned from data. The primary axioms are the conditional independence assumption and the validity of cited approximation methods.

free parameters (6)
  • gamma (HDP concentration)
    Assigned Gamma(1,5) prior, learned from data; controls number of states.
  • alpha (DP concentration)
    Assigned Gamma(1,5) prior, learned from data; controls state persistence.
  • delta, eta (Gamma shape prior) = 1, 1
    Weakly informative priors for a_k hyperparameters, set to 1.
  • zeta, xi (Gamma rate prior) = 1, 1
    Weakly informative priors for b_k hyperparameters, set to 1.
  • mu_0, kappa_0 (VM prior) = 0, 1
    Weakly informative prior for mean direction; mu_0=0, kappa_0=1.
  • rho, chi (kappa prior) = 1, 1
    Parameters of the prior on kappa_k; stated as 'all remaining emission hyperparameters equal to 1' but not individually named in the main text.
assumptions (4)
  • domain assumption Conditional independence of circular and linear components within each state
    Section 2: f(z_t|s_t=k) = f_x(x_t|...) f_y(y_t|...). This factorization is assumed, not derived; it enables the separate conjugate/approximate updates for each component.
  • standard math HDP-HMM framework (Teh et al. 2006, Beal et al. 2001)
    The infinite HMM construction via GEM and DP priors is a standard result invoked as the model backbone.
  • standard math Beam sampler validity (Van Gael et al. 2008)
    The correctness of the slice-augmented FFBS for truncating the infinite state space is assumed from prior work.
  • domain assumption Miller (2019) Gamma approximation accuracy
    The full conditional of a_k is approximated by matching derivatives; the accuracy of this approximation for the specific data regimes used here is assumed rather than verified with acceptance rates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Infinite hidden Markov models for cylindrical data." pith.science (2026). https://pith.science/paper/CPR7RL3V

@misc{pith2026260707464,
  author       = {Pith},
  title        = {Pith review of: Infinite hidden Markov models for cylindrical data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPR7RL3V}},
  note         = {Machine review of arXiv:2607.07464}
}
read the original abstract

We propose an infinite hidden Markov model for cylindrical time series with von Mises-Gamma emissions. Posterior inference is performed using a beam sampler combining conjugate updates and approximate sampling schemes. Simulation studies and two real data applications demonstrate the effectiveness of the proposed methodology.

Figures

Figures reproduced from arXiv: 2607.07464 by the authors.

Figure 1
Figure 1. Daily wind speed (bottom) and state-conditional distributions of wind direction (top), colored according to the [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Boxplots of the adjusted Rand index (ARI, left) and the median autocorrelation time (ACT, right) obtained [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Observed circular direction and linear speed time series. Points are colored according to the maximum a [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Posterior distributions of the state-specific emission parameters [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Posterior distributions of the state-specific emission parameters [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Posterior distribution of the state-specific mean direction [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 16 canonical work pages

  1. [1]

    The infinite hidden Markov model

    Beal, M., Ghahramani, Z., Rasmussen, C., 2001. The infinite hidden Markov model. Adv. Neural Inf. Process. Syst. 14, 577–584

    work page 2001

  2. [2]

    Calculating posterior distributions and modal estimates in Markov mixture mod- els

    Chib, S., 1996. Calculating posterior distributions and modal estimates in Markov mixture mod- els. J. Econom. 75, 79–97

    work page 1996

  3. [3]

    A comparison between initialization strategies for the infinite hidden Markov model

    Cortese, F.P ., Rossini, L., 2026. A comparison between initialization strategies for the infinite hidden Markov model. Comput. Stat. Data Anal. , 108440

    work page 2026

  4. [4]

    A Bayesian analysis of some nonparametric problems

    Ferguson, T.S., 1973. A Bayesian analysis of some nonparametric problems. Ann. Stat. 1, 209–230

    work page 1973

  5. [5]

    A fast algorithm for sampling from the posterior of a von Mises distribution

    Forbes, P .G., Mardia, K.V ., 2015. A fast algorithm for sampling from the posterior of a von Mises distribution. J. Stat. Comput. Simul. 85, 2693–2701

    work page 2015

  6. [6]

    Evaluating the accuracy of sampling-based approaches to calculating poste- rior moments, in: Bernardo, J.M., Berger, J.O., Dawid, A.P ., Smith, A.F.M

    Geweke, J., 1992. Evaluating the accuracy of sampling-based approaches to calculating poste- rior moments, in: Bernardo, J.M., Berger, J.O., Dawid, A.P ., Smith, A.F.M. (Eds.), Bayesian Statistics 4. Clarendon Press, Oxford, UK, pp. 169–193

    work page 1992

  7. [7]

    A general coe fficient of similarity and some of its properties

    Gower, J.C., 1971. A general coe fficient of similarity and some of its properties. Biometrics , 857–871

    work page 1971

  8. [8]

    Regime switching models for circular and linear time series

    Harvey, A., Palumbo, D., 2023. Regime switching models for circular and linear time series. Journal of Time Series Analysis 44, 374–392

    work page 2023

Show all 16 references
  1. [9]

    Hidden markov models for circular and linear-circular time series

    Holzmann, H., Munk, A., Suster, M., Zucchini, W., 2006. Hidden markov models for circular and linear-circular time series. Environ. Ecol. Stat. 13, 325–347

    2006

  2. [10]

    Copula-based segmentation of cylindrical time series

    Lagona, F., 2019. Copula-based segmentation of cylindrical time series. Stat. & Prob. Letters 144, 16–22

    2019

  3. [11]

    A hidden markov approach to the analysis of space–time environmental data with linear and circular components

    Lagona, F., Picone, M., Maruotti, A., Cosoli, S., 2015. A hidden markov approach to the analysis of space–time environmental data with linear and circular components. Stoch. Environ. Res. Risk Assess. 29, 397–409

    2015

  4. [12]

    Fast and accurate approximation of the full conditional for gamma shape parameters

    Miller, J.W., 2019. Fast and accurate approximation of the full conditional for gamma shape parameters. J. Comput. Graph. Stat. 28, 476–480

    2019

  5. [13]

    A constructive definition of Dirichlet priors

    Sethuraman, J., 1994. A constructive definition of Dirichlet priors. Stat. Sin. 4, 639–650

    1994

  6. [14]

    Targeted expression of tetanus toxin reveals sets of neurons involved in larval locomotion in Drosophila

    Suster, M.L., Martin, J.R., Sung, C., Robinow, S., 2003. Targeted expression of tetanus toxin reveals sets of neurons involved in larval locomotion in Drosophila. J. Neurobiol. 55, 233–246

    2003

  7. [15]

    Hierarchical Dirichlet processes

    Teh, Y .W., Jordan, M.I., Beal, M.J., Blei, D.M., 2006. Hierarchical Dirichlet processes. J. Am. Stat. Assoc. 101, 1566–1581. V an Gael, J., Saatci, Y ., Teh, Y .W., Ghahramani, Z., 2008. Beam sampling for the infinite hidden Markov model, in: Proc. 25th Int. Conf. Mach. Learn....

    2006

  8. [16]

    Hidden Markov Models for Time Series: An Introduction Using R

    Zucchini, W., MacDonald, I.L., Langrock, R., 2017. Hidden Markov Models for Time Series: An Introduction Using R. CRC Press, Boca Raton, FL. 20

    2017

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.