Find your next learning step
When a polymer remains in a kinetic regime for a characteristic time, forcing every exit to be memoryless can turn persistence into a misleading transition probability.
Open the chapter contents · Polymer Science root · Gold Standard of Education · How X Works
Chapter contents
Choose the route closest to the question, or read from Chapter 1 for the complete argument. The chapters remain open on the page; these grouped routes are an index.
Route 1 · Define persistent hidden states
Route 2 · Work the dwell-time ledgers
Route 3 · Decode and transfer regimes
Route 4 · Check limits and evidence
Route 5 · Ask and continue
CHAPTER 1 OF 12 · DEFINE PERSISTENT HIDDEN STATES
1. Why duration belongs inside the state model
A hidden Markov model can identify unobserved polymer regimes from noisy measurements, but its usual state process has a specific duration implication: once a state is entered, the chance of leaving is constant at every step. That produces a geometric dwell-time distribution.
Many kinetic regimes do not behave that way. Nucleation may require a minimum induction period; a metastable morphology may persist and then become easier to leave; instrument smoothing may create apparent plateaus. A hidden semi-Markov model separates the next-state choice from the time spent before that transition.
CHAPTER 2 OF 12 · DEFINE PERSISTENT HIDDEN STATES
2. Hidden states, observations and dwell times
Let the hidden state represent a kinetic regime such as induction, rapid ordering or slow relaxation. Measurements—scattering intensity, modulus or calorimetric signal—are generated from a state-specific observation model. The state remains active for a random duration, then moves according to a transition rule.
This separation is the core gain. The observation distribution answers “what does the regime look like?” The dwell distribution answers “how long does it persist?” The transition matrix answers “where can it go next?” Combining those questions in one self-transition probability can obscure the physical interpretation.
CHAPTER 3 OF 12 · WORK THE DWELL-TIME LEDGERS
3. A dwell-time ledger
Imagine observations recorded every minute. A fitted sequence contains three induction episodes with durations 4, 5 and 6 minutes, and three ordering episodes with durations 2, 2 and 7 minutes.
| Regime | Observed episode durations | Mean duration | Shape clue |
|---|---|---|---|
| Induction | 4, 5, 6 | 5.0 | Concentrated around a characteristic time |
| Ordering | 2, 2, 7 | 3.7 | Short episodes plus a long upper tail |
A geometric model can match a mean yet miss the absence of one-minute induction episodes or the mixture-like ordering tail. A semi-Markov duration family makes those discrepancies testable.
CHAPTER 4 OF 12 · WORK THE DWELL-TIME LEDGERS
4. Likelihood contributions by episode
Suppose an induction dwell distribution assigns probabilities 0.15, 0.50 and 0.35 to durations 4, 5 and 6. An observed five-minute episode contributes 0.50 to the duration factor before observation likelihoods are included. A geometric approximation with exit probability 0.20 assigns 0.2×0.8^4, about 0.082, to five steps.
| Duration model | Probability of 5-step induction | Memory statement | Scientific reading |
|---|---|---|---|
| Explicit 4–6 minute distribution | 0.50 | Exit depends on elapsed dwell | Characteristic induction window |
| Geometric with exit probability 0.20 | about 0.082 | Constant exit chance | Memoryless persistence |
The full likelihood also includes emissions, transitions and right-censoring. This single episode calculation is a diagnostic for what duration assumptions contribute.
CHAPTER 5 OF 12 · WORK THE DWELL-TIME LEDGERS
5. Right-censoring and the last regime
If an experiment stops while the polymer is still in a regime, the final dwell time is not observed exactly. Treating the run end as a true transition systematically shortens estimated persistence. The contribution should instead use the probability of surviving at least the observed duration.
Left-censoring matters when acquisition starts after the process has already entered a regime. Record trigger time, thermal history and acquisition window. A sophisticated duration family cannot reconstruct unrecorded entry times without assumptions.
CHAPTER 6 OF 12 · DECODE AND TRANSFER REGIMES
6. Decoding is not the same as fitting
Parameter estimation chooses observation, duration and transition parameters. Decoding then chooses a likely state path or computes posterior state probabilities. The single most likely path can look certain even when several nearby boundaries have similar posterior support.
Report posterior occupancy and boundary uncertainty beside the decoded path. Compare decoded dwell histograms with posterior predictive simulations. This extends Bayesian change-point maps by allowing multiple recurring regimes and state-specific persistence.
CHAPTER 7 OF 12 · DECODE AND TRANSFER REGIMES
7. Changed-context transfer: oscillatory shear
Transfer the model from a monotone thermal ramp to oscillatory shear with repeated ordering and disordering. Now the same regime may recur, and dwell time may depend on strain amplitude. A homogeneous duration distribution is no longer adequate if the protocol changes the exit hazard.
The transfer check is to include the protocol variable in transition or duration parameters, then test on a held-out amplitude. If states merely relabel themselves between runs, align them using physical emission features and transition context rather than numeric state labels.
CHAPTER 8 OF 12 · DECODE AND TRANSFER REGIMES
8. Replicates and kinetic persistence
Fit complete replicates rather than splicing runs into one artificial series. A hierarchical model can share information while allowing batch-specific baselines or duration shifts. Check whether a state retains comparable emission signatures and dwell distributions across specimens.
Persistence is strongest evidence when it survives reasonable changes in sampling interval, smoothing and initialisation. Down-sample a high-frequency series and refit: if a purported minimum dwell disappears, it may have been induced by the instrument or preprocessing.
CHAPTER 9 OF 12 · CHECK LIMITS AND EVIDENCE
9. Limits and common misconceptions
A hidden state is a statistical construct, not automatically a chemical species or morphology. Physical naming requires external evidence. Extra states can divide one broad emission distribution without discovering a new mechanism, while too few states can force heterogeneous episodes into one misleading duration law.
Semi-Markov flexibility does not remove identifiability problems. Duration and emission models can trade off; short noisy excursions may be explained as a new state or as observation error. Use constrained transition graphs only when chemistry justifies them, and compare plausible alternatives.
CHAPTER 10 OF 12 · CHECK LIMITS AND EVIDENCE
10. Practical evidence checklist
Retain raw time series, time resolution, preprocessing, protocol covariates, censoring indicators, initialisation strategy and the complete fitted parameter set. Publish state emission summaries, duration distributions, transition probabilities and posterior uncertainty.
Use simulated-data recovery, replicate holdouts, posterior predictive dwell checks and sensitivity to the maximum allowed duration. Compare with a simpler hidden Markov model and a direct change-point model. A more flexible model earns its place only when it improves relevant predictions or clarifies a reproducible kinetic question.
CHAPTER 11 OF 12 · ASK AND CONTINUE
11. Frequently asked questions
Why not expand each regime into many Markov substates?
That can approximate flexible durations and is computationally useful, but the expansion and truncation must be documented because they determine the implied dwell law.
Does a long dwell prove thermodynamic stability?
No. It shows persistence under the observed protocol and model. Kinetic trapping, measurement limits and unobserved covariates can also create long episodes.
Must every state have a different duration family?
No. Sharing a family can be parsimonious, but the equality should be checked rather than assumed for convenience.
CHAPTER 12 OF 12 · ASK AND CONTINUE
12. Sources and next reading
Yu’s review of hidden semi-Markov models explains explicit-duration formulations and inference. Langrock and Zucchini’s work on arbitrary state dwell-time distributions develops a flexible approximation through expanded-state hidden Markov models. The polymer episodes and probabilities above are original teaching examples.
Continue through the Polymer Science root, Bayesian change points and replicate-aligned maps, waiting-time spectra and robust subspace means. The Gold Standard of Education and How X Works provide the wider route.
Top · Contents · Previous · How X Works