arxiv: 2604.24587 · v2 · submitted 2026-04-27 · 📊 stat.AP

Recognition: unknown

Bayesian inference for hidden Markov models under genuine multimodality with application to ecological time series

Jeffrey S. Rosenthal, Marco A. Gallegos-Herrada, Vianey Leos-Barajas

Pith reviewed 2026-05-07 17:04 UTC · model grok-4.3

classification 📊 stat.AP

keywords hidden Markov modelsBayesian inferenceparallel temperingmultimodalityecological time seriesblue whale divesposterior explorationstate transitions

0 comments

The pith

Modified parallel tempering enables full exploration of multimodal posteriors in hidden Markov models for whale dive data.

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

Bayesian inference for hidden Markov models often encounters genuine multimodality in the joint posterior that persists after label switching corrections. Standard parallel tempering implementations frequently fail to mix adequately across the resulting high-dimensional modes in HMM settings. This paper identifies specific pitfalls in PT for HMMs, introduces targeted algorithm modifications along with new non-informative priors that promote better exploration, and demonstrates the approach on two 3-state HMMs fitted to blue whale dive time series, one incorporating a categorical covariate for sound stimuli. A reader would care because incomplete mixing can produce misleading summaries of state transition probabilities and behavioral responses in ecological applications.

Core claim

The paper establishes that standard PT implementations commonly encounter insufficient exploration of multimodal posteriors in HMMs, but effective modifications to the PT algorithm combined with newly introduced non-informative priors enable reliable sampling across modes; when applied to the blue whale data, this full exploration alters the inferred movement patterns and the estimated effects of sound stimuli on transition probabilities between dive states.

What carries the argument

Modified parallel tempering algorithm with new non-informative priors that together promote mixing across modes in the multimodal posterior of an HMM.

If this is right

Full posterior exploration reveals multiple plausible parameter configurations for the three-state dive behaviors in the whale time series.
Accounting for the sound stimulus covariate under proper sampling yields more reliable estimates of its impact on state transition rates.
The remedies to PT pitfalls support consistent Bayesian inference across similar ecological HMM applications.
New priors reduce the risk of label-switching artifacts while aiding mode traversal in high-dimensional HMM posteriors.

Where Pith is reading between the lines

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

Similar PT modifications may prove useful in other state-space models that exhibit genuine multimodality beyond HMMs.
Analysts fitting HMMs to animal tracking data should routinely verify posterior multimodality before interpreting behavioral inferences.
The new priors could be tested on additional simulated multimodal HMMs to assess their general utility without bias.

Load-bearing premise

The assumption that the identified pitfalls are common to standard PT for HMMs and that the proposed remedies plus new priors enable reliable exploration without introducing new biases or computational problems.

What would settle it

A controlled simulation of an HMM with known multimodal posterior in which the modified PT still misses modes, or re-running the whale analysis and obtaining identical transition probability estimates to those from a standard PT run.

Figures

Figures reproduced from arXiv: 2604.24587 by Jeffrey S. Rosenthal, Marco A. Gallegos-Herrada, Vianey Leos-Barajas.

**Figure 1.** Figure 1: Graphical representation of an HMM. The likelihood of an HMM can be written as a matrix product of the initial state distribution vector δ, transition matrix Γ and state-dependent distributions, LT = δ ⊤D(y1)ΓD(y2)· · · ΓD(yT )1N , (1) where D(yt) is a diagonal matrix with entries f(yt | st), for t ∈ {1, . . . , T}, and 1N is an N-dimensional vector with 1 entries. Given the recursive nature of the likelih… view at source ↗

**Figure 2.** Figure 2: Illustration of a round trip occurrence at time view at source ↗

**Figure 3.** Figure 3: Joint distribution of (γi2, γi3) induced by different choices of priors for the working parameters α (ij) 0 . 7 view at source ↗

**Figure 4.** Figure 4: Induced joint distribution of (γi2, γi3) when the prior distribution of working parameters α (ij) 0 are tempered at different inverse temperature levels when number of hidden states is N = 3. For the within-temperature steps, we use the component-wise Metropolis–Hastings (CWMH) algorithm. The CWMH algorithm is a variant of Metropolis–Hastings that updates one sub-block of parameters at a time. The dimens… view at source ↗

**Figure 5.** Figure 5: Running weight estimates ˆwB,K of mode B accross all 10 PT algorithm implementations. The chain states from the coldest chain corresponding to the PT algorithm implementation with ID 1 were extracted for the computation of the marginal posterior distributions, as well as the 66% and 95% credible intervals view at source ↗

**Figure 6.** Figure 6: Histograms of the marginal posterior distributions estimated from the coldest chain. view at source ↗

**Figure 7.** Figure 7: Running weight estimates of mode B˜ across all 10 PT algorithm implementations. credible intervals overlap are µ22, σ23, µ31, σ31, µ32, σ32, σ33, µ41, σ41, σ52. For the transition probabilities, the 95% credible intervals for γ22 and γ23 do not overlap, while those for γ11 overlap slightly. For the initial state probability estimates, all 95% credible intervals overlap. See view at source ↗

**Figure 8.** Figure 8: Histograms of the marginal posterior distributions estimated from the coldest chain. Row view at source ↗

**Figure 9.** Figure 9: Histograms of the marginal posterior distributions of the transition probabilities in the view at source ↗

**Figure 10.** Figure 10: Traceplot of information from the coldest replica moving across all tempered replicas view at source ↗

**Figure 11.** Figure 11: Traceplot of information from the coldest replica moving across all tempered replicas view at source ↗

read the original abstract

Bayesian inference in hidden Markov models (HMMs) can be challenging due to the presence of multimodality in the likelihood function, and consequently in the joint posterior distribution, even after correcting for label switching. The parallel tempering (PT) algorithm, a state-space augmentation method, is a widely used approach for dealing with multimodal distributions. Nevertheless, standard implementation of the PT algorithm may not always be sufficient to effectively explore the high-dimensional, complex multimodal posterior distributions that arise in HMMs. In this work, we demonstrate common pitfalls when implementing the PT algorithm for HMMs, approaches to remedy them, and introduce new non-informative prior distributions that facilitate effective posterior distribution exploration. We analyse time series of blue whale dive data with two 3-state HMMs in a Bayesian framework, one of which includes a categorical covariate in the transition probability matrix to account for the effect of sound stimuli on the whale's behavior. We demonstrate how effective implementation of the modified PT algorithm for Bayesian inference leads to effective exploration of the resultant multimodal posterior distribution and how that affects inference for the underlying movement patterns of the blue whales.

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.

Desk Editor's Note private letter to a colleague

The paper gives practical PT fixes and new non-informative priors for multimodal HMMs, but the priors need explicit checks to confirm they leave the target posterior unchanged on the whale data.

read the letter

The main things to know are that the authors flag real pitfalls in standard parallel tempering for HMMs and supply both remedies and new non-informative priors meant to improve exploration of the multimodal posterior. They then apply the approach to two 3-state models on blue whale dive time series, one with a sound-stimulus covariate in the transition matrix, and show how fuller mode exploration changes the inferred movement patterns and state dwell times. That application is the clearest part of the work and gives a concrete sense of why the computational issue matters for ecological inference. The remedies for PT look like straightforward implementation advice that practitioners could use. The new priors are presented as non-informative and helpful for mixing, which is the novel methodological piece. The soft spot is exactly the one the stress-test note flags. If the priors shift posterior mass on transition probabilities or state parameters even modestly, then differences in the reported whale inferences could be prior-driven rather than the result of better sampling. The abstract and framing claim the priors facilitate exploration without altering the target, but the paper needs to show direct comparisons—posterior means, credible intervals, or predictive checks—under the new priors versus standard choices to rule that out. Without those checks the central claim about reliable inference on latent behaviors rests on an unverified assumption. This is for readers who already fit Bayesian HMMs to time series and run into label-switching or multimodality problems, especially in ecology or animal movement studies. Someone facing the same computational headaches would pick up usable implementation tips. It is worth sending to peer review because the problem is genuine, the application is relevant, and the fixes are testable; referees can ask for the prior-comparison diagnostics that are currently missing.

Referee Report

2 major / 2 minor

Summary. The paper claims that standard parallel tempering (PT) implementations have common pitfalls when applied to hidden Markov models (HMMs) due to genuine multimodality in the posterior (even after label switching), proposes specific remedies plus new non-informative priors to enable effective exploration, and demonstrates the approach on two 3-state HMMs fitted to blue whale dive time series (one with a categorical sound-stimulus covariate in the transition matrix). Effective PT exploration is shown to change inferences about underlying movement patterns.

Significance. If the central methodological claims hold, the work provides practical, field-specific guidance for Bayesian HMM inference in ecology and animal movement studies, where multimodality is common and reliable posterior exploration directly affects conclusions about behavioral responses to stimuli. The real-data application and focus on non-informative priors that preserve the target distribution would be a useful contribution.

major comments (2)

The load-bearing claim that the new non-informative priors (introduced to aid PT) leave the target posterior unchanged for HMM parameters must be verified explicitly. The manuscript should report posterior summaries (means, credible intervals, or mode locations) for transition probabilities and state dwell times under both the new priors and standard priors; any systematic shift would undermine the assertion that substantive conclusions on whale movement patterns remain reliable.
Results section (blue whale application): the paper must demonstrate that the modified PT plus new priors recovers the same multimodal structure and substantive inferences as would be obtained with standard priors once full exploration is achieved. Without this equivalence check, the reported effects of sound stimuli on transitions cannot be trusted as prior-independent.

minor comments (2)

Clarify the exact form of the new non-informative priors (e.g., on emission parameters, transition probabilities) and how they differ from common defaults; include the explicit prior densities or hyperparameters in an appendix or methods subsection.
Figure captions and text should explicitly state which modes correspond to which behavioral interpretations (e.g., surface vs. deep dives) to help readers connect posterior exploration to ecological conclusions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important aspects of our methodological claims. We respond to each major comment below and will incorporate revisions to address the concerns raised.

read point-by-point responses

Referee: The load-bearing claim that the new non-informative priors (introduced to aid PT) leave the target posterior unchanged for HMM parameters must be verified explicitly. The manuscript should report posterior summaries (means, credible intervals, or mode locations) for transition probabilities and state dwell times under both the new priors and standard priors; any systematic shift would undermine the assertion that substantive conclusions on whale movement patterns remain reliable.

Authors: We agree that an explicit empirical verification is necessary to support the claim that the new priors leave the target posterior unchanged. Although the new priors were derived to be non-informative and to preserve the target distribution for the HMM parameters (transition probabilities and emission parameters), the original manuscript did not include a side-by-side comparison of posterior summaries. In the revised manuscript we will add a dedicated comparison (in the main text or as supplementary material) reporting posterior means and 95% credible intervals for the transition probabilities and mean state dwell times under both the standard priors and the new non-informative priors, obtained with the modified PT algorithm. This will confirm the absence of systematic shifts and the robustness of the whale movement inferences. revision: yes
Referee: Results section (blue whale application): the paper must demonstrate that the modified PT plus new priors recovers the same multimodal structure and substantive inferences as would be obtained with standard priors once full exploration is achieved. Without this equivalence check, the reported effects of sound stimuli on transitions cannot be trusted as prior-independent.

Authors: We recognize the value of an explicit equivalence demonstration to ensure that the reported effects of sound stimuli are not artifacts of the new priors. The paper shows that standard PT implementations fail to explore the multimodal posterior fully, which is why the modified PT is introduced. To address the referee's point, the revised manuscript will include an additional analysis that applies the modified PT together with the standard priors to achieve fuller exploration, then compares the recovered multimodal structure, posterior summaries, and the inferred sound-stimulus effects on transitions against the results obtained with the new priors. This will establish that the substantive conclusions are prior-independent within the non-informative class. revision: yes

Circularity Check

0 steps flagged

No circularity: methodological fixes and new priors are presented as independent contributions demonstrated on data.

full rationale

The paper introduces modified parallel tempering implementations and new non-informative priors for HMMs to address multimodality, then applies them to blue whale dive data. No derivation reduces to its own inputs by construction, no fitted parameters are relabeled as predictions, and no load-bearing step relies on a self-citation chain that itself assumes the target result. The central claims rest on explicit algorithmic remedies and empirical exploration of the posterior on real time series, which are falsifiable against the data rather than tautological. Self-citations, if present, are not required for the validity of the new priors or PT modifications.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no free parameters, axioms, or invented entities are specified in the provided text.

pith-pipeline@v0.9.0 · 5509 in / 894 out tokens · 66077 ms · 2026-05-07T17:04:43.350619+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

6 extracted references · 2 canonical work pages

[1]

24 John M

URLhttps://www.sciencedirect.com/science/article/pii/S0009261409008604. 24 John M. Maheu and Thomas H. McCurdy. Identifying Bull and Bear Markets in Stock Returns.Journal of Business & Economic Statistics, 18(1):100–112, 2000. doi: 10.2307/1392140. URLhttps://www. jstor.org/stable/1392140. Publisher: [American Statistical Association, Taylor & Francis, Lt...

work page doi:10.2307/1392140 2000
[2]

URLhttps://link.aps.org/doi/10.1103/PhysRevE.100

doi: 10.1103/PhysRevE.100.043311. URLhttps://link.aps.org/doi/10.1103/PhysRevE.100. 043311. Giada Sacchi and Ben Swallow. Toward Efficient Bayesian Approaches to Inference in Hierarchical Hidden Markov Models for Inferring Animal Behavior.Frontiers in Ecology and Evolution, 9, May 2021. doi: 10.3389/fevo.2021.623731. URLhttps://www.frontiersin.org/journal...

work page doi:10.1103/physreve.100.043311 2021
[3]

, Ndo 6:Propose α(ij) 0 ∗ j ∼q α(ij) 0 j ,· 7:θ← α(ij) 0 ∗ 1 ,

Update baseline working parameters 5:fori= 1, . . . , Ndo 6:Propose α(ij) 0 ∗ j ∼q α(ij) 0 j ,· 7:θ← α(ij) 0 ∗ 1 , . . . , α(ij) 0 i , . . . α(ij) 0 N ,δ,α (ij) 1 ,ϕ 8:θ ∗ ← α(ij) 0 ∗ 1 , . . . , α(ij) 0 ∗ i , . . . α(ij) 0 N ,δ,α (ij) 1 ,ϕ 9:Compute: A= logp(θ ∗ |y 1:T )−logp(θ|y 1:T ) + logq α(ij) 0 ∗ j , α(ij) 0 j −logq α(ij) 0 j , α(ij) 0 ∗ j 10:U∼Uni...
[4]

Update initial distribution 20:Proposeδ ∗ ∼q(δ,·) 21:θ ∗ ← α(ij) 0 ∗ ,δ,α (ij) 1 ,ϕ 22:θ ∗ ← α(ij) 0 ∗ ,δ ∗,α (ij) 1 ,ϕ 30 23:Compute: A= logp(θ ∗ |y 1:T )−logp(θ|y 1:T ) + log q(δ ∗,δ) q(δ,δ ∗)
[5]

, Ndo 26:Propose α(ij) 1 ∗ j ∼q α(ij) 1 j ,· 27:θ← α(ij) 0 ∗ ,δ ∗, α(ij) 1 ∗ 1 ,

Update covariate parameters (if applicable) 24:ifincludeCovariatethen 25:fori= 1, . . . , Ndo 26:Propose α(ij) 1 ∗ j ∼q α(ij) 1 j ,· 27:θ← α(ij) 0 ∗ ,δ ∗, α(ij) 1 ∗ 1 , . . . , α(ij) 1 i , . . . α(ij) 1 N ,ϕ 28:θ ∗ ← α(ij) 0 ∗ ,δ ∗, α(ij) 1 ∗ 1 , . . . , α(ij) 1 ∗ i , . . . α(ij) 1 N ,ϕ 29:Compute: A= logp(θ ∗ |y 1:T )−logp(θ|y 1:T ) + log q α(ij) 1 ∗ j ,...
[6]

, Pdo 38:Proposeϕ ∗ p ∼q(ϕ p,·) 39:θ← α(ij) 0 ∗ ,δ ∗, α(ij) 1 ∗ ,ϕ ∗ 1,

Update state-dependent parameters 37:forp= 1, . . . , Pdo 38:Proposeϕ ∗ p ∼q(ϕ p,·) 39:θ← α(ij) 0 ∗ ,δ ∗, α(ij) 1 ∗ ,ϕ ∗ 1, . . . ,ϕp, . . . ,ϕP 40:θ ∗ ← α(ij) 0 ∗ ,δ ∗, α(ij) 1 ∗ ,ϕ ∗ 1, . . . ,ϕ∗ p, . . . ,ϕP 41:Compute: A= logp(θ ∗ |y 1:T )−logp(θ|y 1:T ) + log q(ϕ∗ k, ϕk) q(ϕk, ϕ∗ k) 42:U∼Unif(0,1) 43:iflog(U)> Athen 44:ϕ ∗ p ←ϕ p 45:end if 46:end for...

2017