{"id":"dabe7af0-6cd9-4a6b-84cf-b650f68b5401","arxiv_id":"2504.15156","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper demonstrates finite Markov chain imbedding for exact HMM summary-statistic posteriors and introduces an Artemis plot heuristic to calibrate hybrid Viterbi/posterior decoding.","lead":"The paper combines two tools for hidden Markov models: an exact method for computing posterior distributions of hidden-state summaries, and a hybrid decoder that balances global and per-position accuracy. A new heuristic, the Artemis plot, selects the balance weight, and the paper applies both tools to fetal movement and earthquake data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The block-wise decoding improvement is evaluated on the same simulated sequences that were used to select the Artemis tuning weight, so the reported gain may reflect in-sample tuning rather than a general property of hybrid decoding.","rationale":"The FMCI half of the paper is mathematically sound: Theorem 1 correctly establishes the conditional Markov property of the hidden sequence, and the imbedding construction follows Aston and Martin (2007). The truncation issue flagged by the reader is real but minor, since increasing the truncation level L makes the reported distributions exact up to an arbitrarily small tail probability. The hybrid risk derivation in Section 4.1 and the dynamic programming recursion in Appendix A.2 are also correct. The load-bearing weakness is the empirical support for the headline decoding improvement. Section 4.4 uses the same simulations to choose alpha via the Artemis rule and to evaluate block-wise accuracy, so the comparison lacks independence. Moreover, the 45-degree tangent rule is defined after min-max normalization of each axis, so the chosen alpha depends on arbitrary scaling and has no decision-theoretic justification. These are not internal inconsistencies, but they mean the central claim 'hybrid decoding shows increased performance' is supported only by an in-sample, scaling-dependent simulation study on a narrow family of models. The proposed split-sample test (and a raw-scale control) would settle whether the improvement generalizes. This does not change the reader's conditional verdict: the paper remains acceptable if the empirical claim is re-evaluated out-of-sample and the Artemis heuristic is either justified or clearly framed as an ad hoc guideline.","tokens_in":19595,"tokens_out":8353,"duration_ms":84999,"concrete_test":"Use a split-sample protocol for each of the nine models in Section 4.3: simulate, say, 50 pairs (x,y); on 25 simulations select alpha by the Artemis 45-degree rule exactly as in Section 4.2; on the remaining 25 (never used for tuning) compute average block-wise accuracy for posterior, Viterbi, and hybrid for block sizes 1-10, with standard errors. If hybrid does not beat both baselines for block sizes 2-10 on the held-out set, the central claim fails. As a secondary check, recompute the Artemis-selected alpha with the y-axis left in raw log-probability units (no min-max scaling); if the selected alpha or the held-out block gains change substantially, the 45-degree rule is scaling-dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central decoding claim rests on Section 4.4's comparison, but the protocol is not a clean out-of-sample test. In Section 4.2, the Artemis rule selects an optimal alpha separately for each of 10 simulated (x,y) pairs from the fitted 3-state Poisson model. Section 4.4 then reports block-wise accuracy 'based on 10 simulations for each block size' and states that 'The alpha's used for hybrid are the optimal alpha's found using the Artemis plots in Figure 10.' Thus the same simulated hidden states used to tune alpha are also used to measure accuracy. If the Artemis selection is noisy or tied to those particular y's, the hybrid gains in Figures 13-14 (often only 1-3% on the ratio scale) can be inflated; no error bars or cross-validation are provided. The problem is compounded by the 45-degree rule itself: the angle is taken after min-max scaling of each simulation's axes (Figure 10), so it has no invariant meaning and is not derived from any loss or utility function. The claim that hybrid decoding 'shows increased performance' is therefore only as strong as the transfer of this arbitrarily scaled heuristic to new data and to HMMs outside the tested three-state Poisson family.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses two tasks for hidden Markov models: computing exact posterior distributions of summary statistics of the hidden state sequence via finite Markov chain imbedding (FMCI), and improving decoding through hybrid decoding with a proposed tuning-parameter rule called the Artemis analysis. The authors prove Theorem 1, that the hidden state sequence conditional on the observed sequence is an inhomogeneous first-order Markov chain, and they use it to build FMCI transition matrices for the number of jumps, number of positions in a state, exact run lengths, and longest run. These FMCI results are validated against 1000 posterior samples on the fetal lamb movement data. For decoding, the paper rederives the hybrid risk as a weighted geometric mean, introduces the Artemis plot for choosing the tuning parameter alpha, and reports simulation results suggesting that hybrid decoding improves block-wise accuracy relative to Viterbi and posterior decoding for intermediate block sizes. Code is provided in a GitHub repository.","tokens_in":19861,"tokens_out":5934,"duration_ms":57205,"significance":"The FMCI part is sound and clearly connected to Theorem 1: the theorem is standard and the proof is correct, and the comparison with posterior samples is an appropriate internal consistency check rather than circular reasoning. The alternative weighted-geometric-mean derivation of hybrid decoding is clean and complements the existing decision-theoretic framework. The provision of reproducible code is a genuine strength. The main novel methodological contribution, however, is the Artemis rule for selecting alpha, and the current evidence for the central decoding claim rests on only ten simulations per setting, with tuning and evaluation performed on the same simulated data and no error bars. If the decoding claim can be supported by a more rigorous evaluation, the paper will be a useful methods contribution for practitioners; as it stands, the decoding part is not yet convincing enough to support the abstract's claim of increased performance.","major_comments":[{"comment":"The evaluation of hybrid decoding is in-sample. Section 4.4 states that \"The alpha's used for hybrid are the optimal alpha's found using the Artemis plots in Figure 10,\" and those optimal alpha values were determined in Section 4.2 from the same 10 simulated (x,y) pairs per setting that are subsequently used to compute block-wise accuracy. This protocol allows selection bias to inflate the reported gains, which are often modest (ratios from roughly 1.01 to 1.48 relative to Posterior decoding). With only 10 replicates and no error bars, the observed improvement could be within simulation noise. An out-of-sample or cross-validated evaluation, together with uncertainty estimates, is required before the abstract claim that \"hybrid decoding shows increased performance\" can be considered established.","section":"Section 4.4, Figures 13-14"},{"comment":"The 45-degree rule for choosing alpha is not well-defined as stated. The axes are \"scaled using the minimum and maximum value for each axis,\" independently for each simulation. The angle of the tangent to the bow-shaped curve is therefore not invariant under affine rescaling of the axes, and the reported optimal alpha could change simply by changing the plotting range. The authors need to either define alpha selection by an explicit criterion that does not depend on plotting scale, such as a point on a Pareto frontier defined by a meaningful trade-off, or demonstrate that the 45-degree rule is robust to reasonable changes in axis scaling. As written, the Artemis rule is an arbitrary heuristic rather than a principled estimator.","section":"Section 4.2, Figure 10"},{"comment":"The matrix displays for the FMCI transition matrices are incomplete and cannot be independently verified from the printed text. For example, in Section 3.2.1 the block B(a,b) is defined but the displayed matrix Lambda does not fill in the entries of the block structure explicitly; in Section 3.2.3 the row \"1-b 0 b\" leaves the first-column entry unspecified. Since these transition matrices are the core of the FMCI distributional claims, the authors should provide complete, unambiguous matrix definitions, either as displayed matrices or by explicit equation numbers, and should check that the rendering is not garbled.","section":"Sections 3.2.1-3.2.4"},{"comment":"The FMCI distributions are exact only conditional on the user-chosen truncation level ell. The last entry of the imbedding vector lumps all counts exceeding ell into one absorbing state, as stated in Section 3.2.1. This is a real limitation, and the manuscript should report the tail probability or a sensitivity analysis for the choice of ell, particularly for the longest-run distribution where the tail is of direct interest. The current text acknowledges the truncation but does not quantify its impact on the reported posterior distributions.","section":"Section 3.2.1 and Section 5"}],"minor_comments":[{"comment":"There is a typo: \"we that have Lambda = Lambda(a,b)\" should read \"we have Lambda = Lambda(a,b)\".","section":"Section 3.2.1"},{"comment":"In the denominator of the conditional initial state probability, the notation \"Phi(x1|y1=i)\" should be \"Phi(x1|i)\" for consistency with the rest of the paper.","section":"Equation (4)"},{"comment":"The block matrix B(a,b) for the number of positions appears to have the first row as (1-b, a) and the second as (a, 1-a), but the accompanying text and Figure 7 should clarify which entries correspond to transitions that increment the count; currently the notation is confusing.","section":"Section 3.2.2"},{"comment":"The phrase \"blocks of size two any two pairs in the sequence\" is unclear and should be reworded to \"blocks of size two consist of any two consecutive positions in the sequence.\"","section":"Section 4.4, Figure 13 caption"},{"comment":"There is a typo: \"probablity\" should be \"probability\" in the description of the conditional probability used in the Artemis plot.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper contains two parts of quite different status. The FMCI portion is a clear exposition and application of known methodology, with a correct theorem, and it will be useful mainly as a didactic and reproducible resource. The hybrid decoding portion is the claimed novelty, but its validation is currently not convincing: the tuning rule is scale-dependent and the performance gain is evaluated in-sample with no error bars. These problems are fixable in principle through an out-of-sample protocol and a scale-invariant or explicitly justified selection criterion, so I am not recommending rejection, but the central decoding claim needs substantial additional work before publication in a serious journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the FMCI half is a clean, self-contained application of Aston and Martin (2007): Theorem 1 (posterior hidden sequence is an inhomogeneous Markov chain) is proved correctly, and the matrix constructions for jumps, dwell times, exact run lengths, and longest run are standard. The validation against posterior samples is a nice internal consistency check. Second, the hybrid decoding half introduces a genuinely new heuristic, the Artemis plot for choosing α, but the simulation evidence behind the 'increased performance' claim is weaker than the abstract implies.\n\nCredit where due: the weighted-geometric-mean re-derivation of hybrid risk in Section 4.1 is algebraically simple but didactically useful, connecting Lember–Koloydenko to a familiar mean. The paper ships code, which is good. The FMCI exposition is the strongest part.\n\nSoft spots, in increasing severity. Minor: the FMCI distributions are exact only up to the user-chosen truncation level L; the abstract's unqualified 'exact' should carry that caveat, and the garbled matrix displays in the text made independent verification harder than it should be. Major: the block-wise accuracy comparison in Section 4.4 uses the same 10 simulated sequences that were used to select α in Section 4.2. That is in-sample tuning. Combined with the fact that the 45-degree rule is applied after min-max scaling of the axes — so the angle has no invariant meaning — the reported 1–3% gains over posterior decoding in Figures 13–14 are not yet evidence of a generalizable improvement. There are no error bars, no cross-validation, and the tested family is narrow (3-state Poisson with symmetric transitions). The stress-test note is right to flag this.\n\nWho gets value: anyone doing posterior summary statistics for HMMs will be served well by the FMCI part and its code. The hybrid section is a useful conversation starter for practitioners who want to interpolate between Viterbi and posterior decoding, but the heuristic needs deeper justification or at least a clean out-of-sample evaluation.\n\nRecommendation: this deserves a serious referee. The FMCI part alone justifies it. Send it to review, ask the authors to fix the evaluation protocol (separate tuning from testing), add uncertainty measures, and clarify the 'exact' caveat. With those revisions it would be a useful paper.","headline":"A solid FMCI exposition with a genuinely new but empirically unvalidated tuning heuristic; the block-wise decoding claim rests on in-sample tuning, so the abstract overstates the result.","tokens_in":20393,"tokens_out":3152,"would_cite":true,"duration_ms":28884,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M05","60J10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that exact posterior distributions of hidden-state summaries can be computed by finite Markov chain imbedding, and that hybrid decoding with an Artemis-chosen weight outperforms Viterbi and posterior decoding on…","keywords":["hidden Markov model","finite Markov chain imbedding","hybrid decoding","Artemis plot","pattern distributions","Viterbi decoding","posterior decoding","run length distribution"],"falsifier":"Simulate a hidden Markov model with misspecified emissions, such as overdispersed or heavy-tailed counts, compute the tuning weight chosen by the 45-degree Artemis tangent, and grid-search the weight that maximizes block-wise accuracy on independent test sequences; if the Artemis choice is not close to the accuracy-maximizing weight, the rule is not general.","tokens_in":19408,"feed_emoji":"🏹","tokens_out":8875,"duration_ms":72239,"temperature":0.7,"pith_summary":"Hidden Markov model users usually get one decoded path, from Viterbi or posterior decoding, and no distribution for summary statistics of the hidden chain. This paper combines two remedies. First, it proves and exploits the fact that the hidden chain conditional on the data is an inhomogeneous Markov chain, so finite Markov chain imbedding gives posterior distributions, exact up to a user-chosen truncation level, for the number of jumps, the time spent in a state, exact run lengths, and the longest run. Second, it treats decoding as a weighted compromise between pointwise accuracy and whole-sequence probability, and uses a bow-shaped Artemis plot to choose the weight by a 45-degree tangent. On simulated Poisson HMMs the resulting hybrid decoding beats both Viterbi and posterior decoding for blocks of size two or more.","feed_headline":"45-degree rule picks hybrid decoding that beats Viterbi and posterior","feed_subtitle":"A 45-degree Artemis rule improves block-wise accuracy, and FMCI gives exact run-length posteriors.","key_machinery":"The two mechanisms are the conditional hidden chain and the hybrid loss. Theorem 1 states that, conditional on the observations, the hidden sequence is an inhomogeneous first-order Markov chain with transition probabilities $$P(y_t \\mid y_{t-1}, x) = \\frac{\\beta_t(y_t)\\Gamma_{y_{t-1},y_t}\\Phi(x_t\\mid y_t)}{\\beta_{t-1}(y_{t-1})},$$ where $\\beta_t(y_t)=P(x_{t+1},\\ldots,x_n\\mid y_t)$ are backward probabilities. For a summary statistic such as the number of $1\\to2$ jumps, FMCI builds a larger Markov chain whose state tracks both the current hidden state and the accumulated count, with sparse block-structured transition matrices $\\Lambda(a_t,b_t)$ depending on the conditional staying probabilities; the product of these matrices over time gives the distribution of the count, with the last entry absorbing counts above a truncation level. For decoding, the hybrid loss is $$h(u)=(1-\\$\\alpha$)\\sum_t \\log P(y_t=u_t\\mid x)+\\$\\alpha$\\log P(y=u\\mid x),$$ which is equivalently the log of a weighted geometric mean of the pointwise and whole-sequence objectives, and the Artemis plot selects $\\alpha$ by the 45-degree tangent to the bow-shaped accuracy-versus-log-joint-probability curve.","core_discovery":"The central discovery is that the usual two tasks of HMM analysis, summarizing the hidden chain and decoding it, can both be improved by taking the whole set of hidden paths seriously. Conditional on the observed sequence, the hidden chain is an inhomogeneous Markov chain; this makes it possible to compute exact posterior distributions of pattern statistics through finite Markov chain imbedding rather than sampling alone. For decoding, the paper shows that a hybrid path maximizing a weighted sum of posterior pointwise log-probabilities and the log-probability of the whole path is a natural bridge between posterior and Viterbi decoding, and it introduces an Artemis plot, accuracy against log-joint probability for varying weights, whose bow shape suggests choosing the weight at a 45-degree tangent. The paper's simulations show the resulting hybrid decoder has higher block-wise accuracy than either classic decoder for intermediate block sizes, while posterior decoding remains best for single positions and Viterbi for very large blocks.","pith_inferences":["If the 45-degree heuristic is robust, the same calibration idea could be applied to other composite decoding losses, such as segmentation objectives with a penalty on the number of segments.","FMCI posterior distributions suggest a model-checking tool: compare the observed counts of jumps or runs with their posterior predictive distributions under the fitted HMM.","The block-accuracy results imply that evaluation of decoding should use a biologically or practically meaningful block size rather than pointwise accuracy alone.","For long genomic sequences, the truncation level L becomes a practical bottleneck, and the trade-off between matrix dimension and truncation bias will need explicit treatment in applications."],"forward_implications":["FMCI turns summary statistics of the hidden chain from point estimates into full posterior distributions, so users can state uncertainty about number of runs, state occupancy, and longest runs.","Hybrid decoding with the Artemis-chosen weight outperforms both Viterbi and posterior decoding for block sizes of two or more in the simulated Poisson HMMs studied here.","The optimal tuning weight depends on the model, so each data set needs its own Artemis calibration rather than a universal alpha.","Because alpha=0 and alpha=1 recover posterior and Viterbi decoding, hybrid decoding is a continuous interpolation between local and global decoding.","The weighted-geometric-mean derivation connects hybrid risk to conditional-risk formulations, placing the method in a broader decision-theoretic framework."],"supporting_citations":[{"why":"Supplies the finite Markov chain imbedding framework for distributions of runs and patterns in hidden Markov models.","marker":"[Aston and Martin, 2007]"},{"why":"Provides the original Markov chain imbedding approach for run distributions that the FMCI construction builds on.","marker":"[Fu and Koutras, 1994]"},{"why":"Is the source for the inhomogeneous Markov chain property of the posterior hidden sequence, proved here as Theorem 1.","marker":"[Lindgren, 1978]"},{"why":"Introduces the hybrid decoding or generalized risk approach that the paper applies and re-derives.","marker":"[Lember and Koloydenko, 2014]"},{"why":"Provides the pairwise Markov model formulation and the dynamic programming recursion used to compute hybrid paths.","marker":"[Kuljus and Lember, 2023]"},{"why":"Supplies the classical data sets, parameter estimates, and the Viterbi and posterior decoding baselines the paper compares against.","marker":"[Zucchini et al., 2016]"}],"fun_headline_variants":["Hybrid decoding beats Viterbi and posterior for block accuracy","Exact posterior runs via FMCI; hybrid decoder wins on blocks","45-degree Artemis pick for hybrid decoding beats classics","Exact posteriors and hybrid decode beat Viterbi and posterior","Hybrid decoding tuned by 45-degree rule improves block accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The practical benefit of hybrid decoding rests on the assumption that the point where a 45-degree arrow meets the bow-shaped accuracy-versus-probability curve identifies a good tuning weight for every hidden Markov model, even though the axes are rescaled before drawing the angle.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid decoding beats Viterbi and posterior for block accuracy","Exact posterior runs via FMCI; hybrid decoder wins on blocks","45-degree Artemis pick for hybrid decoding beats classics","Exact posteriors and hybrid decode beat Viterbi and posterior","Hybrid decoding tuned by 45-degree rule improves block accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3216,"prompt_tokens":940,"completion_tokens":2276,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2191}},"tokens_in":556,"tokens_out":2276,"duration_ms":14957,"temperature":1.0,"reasoning_tokens":2191,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:32:41.417705+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a hidden Markov model with misspecified emissions, such as overdispersed or heavy-tailed counts, compute the tuning weight chosen by the 45-degree Artemis tangent, and grid-search the weight that maximizes block-wise accuracy on independent test sequences; if the Artemis choice is not close to the accuracy-maximizing weight, the rule is not general.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the finite Markov chain imbedding framework for distributions of runs and patterns in hidden Markov models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the source for the inhomogeneous Markov chain property of the posterior hidden sequence, proved here as Theorem 1."},{"cited_title":"and Koloydenko, A","cited_arxiv_id":null,"evidence_quote":"Introduces the hybrid decoding or generalized risk approach that the paper applies and re-derives."},{"cited_title":"and Lember, J","cited_arxiv_id":null,"evidence_quote":"Provides the pairwise Markov model formulation and the dynamic programming recursion used to compute hybrid paths."},{"cited_title":"L., and Langrock, R","cited_arxiv_id":null,"evidence_quote":"Supplies the classical data sets, parameter estimates, and the Viterbi and posterior decoding baselines the paper compares against."}],"review_version":1}