pith. machine review for the scientific record. sign in

arxiv: 2603.24299 · v2 · submitted 2026-03-25 · 📊 stat.ME

Recognition: 2 theorem links

· Lean Theorem

Mortality Forecasting as a Flow Field in Tucker Decomposition Space

Samuel J. Clark
Authors on Pith no claims yet

Pith reviewed 2026-05-15 00:34 UTC · model grok-4.3

classification 📊 stat.ME
keywords mortality forecastingTucker decompositionlife expectancymortality schedulescross-validationLee-CarterBayesian demographic modelsHuman Mortality Database
0
0 comments X

The pith

Mortality forecasting improves by integrating a flow field through the one-dimensional score space of a Tucker decomposition of historical data, producing lower bias than time-series methods.

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

The paper reframes mortality forecasting as flowing through a low-dimensional space from Tucker tensor decomposition of age-sex-country mortality rates rather than extrapolating time trends. PCA on the scores shows the transition from high to low mortality is essentially one-dimensional, so a scalar speed function and fixed trajectory functions generate full future schedules. This structure keeps forecasts inside observed patterns instead of venturing into unseen territory, cutting long-horizon bias. Cross-validation on the Human Mortality Database confirms smaller bias than Lee-Carter or Hyndman-Ullah and lower error than the UN pipeline across millions of test points. Accurate long-range life expectancy matters directly for pension funding, healthcare planning, and social security solvency.

Core claim

Mortality forecasting is reframed as integrating a flow field through the low-dimensional score space of a Tucker tensor decomposition of the Human Mortality Database; PCA reduction shows the mortality transition is essentially one-dimensional, parameterized by a scalar speed function that advances level and trajectory functions that supply structural scores, with Tucker reconstruction yielding complete sex-specific single-year-of-age mortality schedules at each horizon.

What carries the argument

The flow field on the PCA-reduced Tucker score space, where a scalar speed function advances mortality level and trajectory functions supply the structural scores for reconstruction.

If this is right

  • Bias of only +1.058 years over 50-year horizons versus -3.2 years for Lee-Carter.
  • 2.7 times lower error than the UN pipeline on 1.66 million sex-age-specific test points at every age and horizon.
  • Complete age-specific schedules generated directly without separate model life tables.
  • Navigation stays within observed mortality structures instead of extrapolating temporal trends.

Where Pith is reading between the lines

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

  • Mortality schedules across populations may lie on a low-dimensional manifold traversable by simple flow rules.
  • The same flow-field logic could be tested on fertility or cause-specific mortality data.
  • Speed functions updated with recent observations might improve responsiveness to events like pandemics.

Load-bearing premise

The assumption that PCA reduction of the Tucker score space reveals the mortality transition to be essentially one-dimensional.

What would settle it

Future mortality data or held-out populations showing age-specific rates that deviate from the predicted one-dimensional trajectory in the Tucker score space.

Figures

Figures reproduced from arXiv: 2603.24299 by Samuel J. Clark.

Figure 1
Figure 1. Figure 1: Flow-field structure in Tucker PCA space. Top left: raw year-to-year e0 velocity (forward differences) vs e0 – the scatter is noisy but the LOWESS trend reveals level-dependent improvement; the production speed function uses per-country smoothed velocities in s1 space for a cleaner estimate (fig. 2). Top centre and right: derivative correlations ∆s1 vs ∆s2 and ∆s3 (raw forward differences); the tight linea… view at source ↗
Figure 2
Figure 2. Figure 2: Speed function denoising comparison in s1 space. Left: per-country LOWESS-smoothed forward differences pooled across countries (Method A, production) – the smoothing reveals the underlying improvement trend. Centre: raw forward differences pooled directly (Method B) – the cross-country LOWESS alone cannot fully denoise the year￾to-year noise. Right: overlay of the two LOWESS estimates, showing that per-cou… view at source ↗
Figure 3
Figure 3. Figure 3: shows the s1-to-surface-e0 mapping with and without the tail extension [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Validation of the joint tangent extrapolation in s1 space. Per-component score slopes dsk/ds1 for the LOWESS tangent at s ∗ 1 ≈ −12 (e0 ≈ 78, red) and for five frontier countries (Japan, Sweden, Switzerland, Spain, Italy) over their last 20 years. The cosine similarity between the LOWESS tangent and the frontier average is 0.94, confirming that the extrap￾olation direction agrees with observed frontier dyn… view at source ↗
Figure 5
Figure 5. Figure 5: Forecast e0 diagnostic for six countries under s1-space navigation (all-data flow field). Green: surface-derived e0 (raw, before bias correction). Red dash-dot: bias-corrected e0 (reported forecast). The annotation shows the 30-year e0 gain. Because navigation is in s1 space, there is no separate navigation e0 that can diverge from the surface e0. directly in the production forecaster; section 11.4 gives t… view at source ↗
Figure 6
Figure 6. Figure 6: Empirical convergence rates in s1 space. Left: autocorrelation of s1-velocity deviations from canonical. Centre: autocorrelation of structural score deviations by PC; PCs 2–5 have half-lives of 12–32 years. Right: speed convergence conditioned on mortality level. Dashed lines show fitted exponentials α h . horizons) – and the flow-field system’s constrained dynamics provide a distinctive advantage at these… view at source ↗
Figure 7
Figure 7. Figure 7: Four-method e0 comparison on 9,507 common test points, all evaluated against raw HMD life-table e0. Left: overall MAE with bias. Centre: MAE by horizon band – pyBayesLife has the lowest MAE at short horizons but accumulates the largest positive bias; Lee–Carter has the lowest overall MAE but substantial negative bias. Right: MAE by individual horizon year – the crossover near h = 12 is clearly visible. 21 … view at source ↗
Figure 8
Figure 8. Figure 8: shows leave-country-out forecasts from the 2000 origin for 18 selected countries. For each country, the flow field is built from the other 47 countries; the forecast (green dashed) with 80% and 95% prediction intervals is plotted against the held-out observations (red dots) that the model did not see during training. The fan opens with √ h scaling [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: 50-year production forecasts with calibrated prediction intervals for 18 selected countries (all-data flow field). Blue: observed e0. Green dashed: median forecast. Shaded: 80% (dark) and 95% (light) prediction intervals. These forecasts use the flow field trained on all 48 countries; see section 5.1 for the distinction between all-data and strict leave-country￾out evaluation. At long horizons (h = 26–50),… view at source ↗
Figure 10
Figure 10. Figure 10: Calibration diagnostics from strict leave-country-out CV (9,507 test points; each country’s flow field built excluding that country; evaluated against raw HMD life-table e0). Top left: forecast vs observed e0. Top right: MAE and bias by horizon. Bottom left: error distribution. Bottom right: error vs observed e0 coloured by horizon band. Year-to-year observed changes in logit(qx) are of order ±0.5/year (r… view at source ↗
Figure 11
Figure 11. Figure 11: Mortality surfaces: logit(qx) by age and year for Sweden, Japan, USA, and Russia (rows), female and male (columns). The vertical dashed line marks the forecast origin. The observed history (Tucker-reconstructed) flows seamlessly into the forecast, with smoothly evolving age-specific structure and no visible seam at the origin. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Rate of mortality improvement: year-over-year change in logit(qx) by age and year for the same four countries. Blue (negative) indicates improvement; red (positive) indicates deterioration. Year-to-year observed changes are 10–50× larger than the smooth forecast derivatives, so the observed region has been rescaled by dividing by the per-panel mean absolute ratio, placing both regions on a common colour s… view at source ↗
Figure 13
Figure 13. Figure 13: Sex differential in life expectancy (e0 female minus male) for Sweden, Japan, USA, and Russia. Solid: observed. Dashed: forecast. The differential continues the observed trend smoothly with no crossover. collapsing to a scalar intermediate. To evaluate whether this architectural difference translates into better age-specific forecasts, we need a benchmark that can also produce age-specific predictions – a… view at source ↗
Figure 14
Figure 14. Figure 14: Age-specific sex differential in mortality: logit(qx)male − logit(qx)female. Left: line plots at selected horizons. Right: heat map across age and time (observed + forecast). The differential is everywhere positive (male excess mortality) and evolves smoothly – no age-specific crossovers at any horizon. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Sex differential in e0 for Tier 1 external forecasts (South Africa, Brazil, India, Bangladesh) and two HMD validation countries (Poland, Japan). External countries enter the flow field at their current e0 level; the sex-specific schedules emerge entirely from Tucker reconstruction. The differential is everywhere positive and varies with mortality level in a pattern consistent with the HMD-wide empirical r… view at source ↗
Figure 16
Figure 16. Figure 16: Sex×age/horizon breakdown of lx-weighted log(1mx) MAE and bias. Top row: by age band. Bottom row: by horizon band. The flow-field’s near-zero bias across most ages and for both sexes contrasts with pyBayesLife’s strong age-dependent bias. 0 20 40 60 80 100 Age 0.2 0.4 0.6 0.8 1.0 lo g(m M x ) lo g(mFx ) Sex Differential by Age Observed (HMD) Flow-field pyBayesLife 0 20 40 60 80 100 Age 0.1 0.0 0.1 0.2 0.3… view at source ↗
Figure 17
Figure 17. Figure 17: Sex differential in age-specific mortality. Left: mean log(mM x ) − log(mF x ) by age – ob￾served (HMD, black), flow-field forecast (green), and pyBayesLife forecast (orange). Cen￾tre: error in the sex differential by age. Right: lx-weighted MAE of the sex differential by age band. The flow-field reproduces the observed sex differential far more accurately because it forecasts the full sex×age surface dir… view at source ↗
Figure 18
Figure 18. Figure 18: Error heatmaps: lx-weighted mean bias in log(1mx) by age (rows) and horizon (columns). Top row: flow-field (left: female, right: male). Bottom row: pyBayesLife. The flow-field errors are uniformly small (near-white) with no systematic age×horizon structure. py￾BayesLife shows strong structured bias reflecting the information bottleneck of recon￾structing an age schedule from scalar e0. The horizontal band… view at source ↗
Figure 19
Figure 19. Figure 19: Bias advantage across all methods. Left: e0 bias by horizon for all four methods on 9,507 common test points. The flow-field (green) stays near zero; Lee–Carter (red) and Hyndman–Ullah (purple) drift to −3 to −8 years; pyBayesLife (orange) drifts positive. Centre: |bias| by horizon band. Right: age-specific log(1mx) bias by age – the flow-field is near zero across ages while pyBayesLife shows strong age-d… view at source ↗
Figure 20
Figure 20. Figure 20: External country forecasts and method comparison. Top row: South Africa, Brazil, India – flow-field (green) vs WPP 2024 medium variant (red) using real WPP e0 estimates as input. Bottom row: Bangladesh, Poland, Japan. The Poland and Japan panels compare all methods: flow-field (green), Lee–Carter (orange), Hyndman–Ullah (purple), and WPP 2024 (red). At 50-year horizons, Lee–Carter and Hyndman–Ullah diverg… view at source ↗
Figure 21
Figure 21. Figure 21: Reconstructed forecast mortality schedules for Sweden, Japan, USA, and Russia. Black: last observed logit(qx). Coloured dashed: forecast at 5-year horizons. The Tucker reconstruction maintains smooth age profiles and coherent sex structure at all horizons. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Sex coherence of forecast mortality schedules. Female (solid, red) and male (dashed, blue) logit(qx) plotted on common axes for selected countries spanning a range of mortality levels, at the last observed year and three forecast horizons (+10, +30, +50 years). The Tucker reconstruction preserves mF x < mM x at every age and horizon without requiring explicit constraints – no crossovers occur. 8.4 Complet… view at source ↗
read the original abstract

Mortality forecasting methods in the Lee-Carter tradition extrapolate temporal components via time-series models, often producing forecasts that systematically underpredict life expectancy at long horizons. This bias is consequential for planning pension funding, healthcare capacity, and social security solvency. The dominant alternative - the Bayesian double-logistic model underlying the UN World Population Prospects - forecasts scalar life expectancy and requires a separate model life table system to recover age-specific rates. We reframe forecasting as integrating a flow field through the low-dimensional score space of a Tucker tensor decomposition of the Human Mortality Database. PCA reduction reveals that the mortality transition is essentially a one-dimensional flow: a scalar speed function advances the level, trajectory functions supply the structural scores, and the Tucker reconstruction produces complete sex-specific, single-year-of-age mortality schedules at each horizon. In leave-country-out cross-validation (9,507 test points, 50-year horizon), the flow-field achieves bias of +1.058 years - substantially smaller than Lee-Carter (-3.2), Hyndman-Ullah (-3.5), and pyBayesLife (+3.3) - because it navigates a score space parameterised by mortality level rather than extrapolating temporal trends into unobserved territory. On 1.66 million sex-age-specific test points, it achieves 2.7x lower error than our de novo Python reimplementation of the UN pipeline trained on the same data - with lower error at every age, every forecast horizon, and for both sexes.

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.

Referee Report

3 major / 2 minor

Summary. The paper reframes mortality forecasting as integration along a flow field in the low-dimensional score space of a Tucker tensor decomposition applied to the Human Mortality Database. PCA on the Tucker scores is used to argue that mortality dynamics are essentially one-dimensional, so that a scalar speed function combined with fixed trajectory functions suffices to generate future sex- and age-specific mortality schedules. Leave-country-out cross-validation on 9,507 test points at 50-year horizons reports a bias of +1.058 years (versus -3.2 for Lee-Carter, -3.5 for Hyndman-Ullah, and +3.3 for pyBayesLife) and 2.7× lower error than a reimplemented UN pipeline across 1.66 million sex-age-specific points.

Significance. If the one-dimensional flow assumption and cross-validation results prove robust, the approach would offer a meaningful advance over time-series extrapolation methods by keeping forecasts inside the empirically observed score manifold, thereby reducing the long-horizon bias that affects pension and social-security planning. The scale of the validation exercise and the direct head-to-head comparisons with established benchmarks constitute a clear empirical strength.

major comments (3)
  1. [PCA reduction of Tucker score space] Section describing the PCA step on Tucker scores: the claim that the mortality transition is 'essentially one-dimensional' is not supported by any reported eigenvalues, scree plot, or cumulative variance percentages. Without these quantities it is impossible to judge whether the first principal component captures the large majority of variation or whether higher-order components encode independent structural shifts (e.g., infant versus senescence patterns) that a scalar speed function cannot reproduce at 50-year horizons.
  2. [Leave-country-out cross-validation] Cross-validation section: the leave-country-out procedure must specify exactly how the speed function and trajectory functions are estimated from training countries only, and what data-exclusion rules are applied to avoid leakage when the held-out countries' schedules are reconstructed. The current description leaves open the possibility that the reported bias and error reductions partly reflect in-sample structure rather than genuine out-of-sample forecasting skill.
  3. [Flow-field construction and integration] Flow-field integration description: the scalar speed function is fitted on historical data; the manuscript does not demonstrate that this function remains valid when extrapolated to horizons beyond the training window, nor does it provide sensitivity checks on the integration step when the speed function is perturbed within its estimation uncertainty.
minor comments (2)
  1. [Abstract] The abstract refers to 'our de novo Python reimplementation of the UN pipeline' but the main text does not indicate whether the code or the exact parameter settings used for that baseline are supplied as supplementary material.
  2. [Notation and definitions] Notation for the Tucker core tensor and the subsequent PCA loadings should be introduced once and used consistently; occasional switches between score-space and original-age-space symbols reduce readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which highlight important areas for clarification and strengthening of the manuscript. We address each major comment below, indicating the revisions we will make.

read point-by-point responses

  1. Referee: Section describing the PCA step on Tucker scores: the claim that the mortality transition is 'essentially one-dimensional' is not supported by any reported eigenvalues, scree plot, or cumulative variance percentages. Without these quantities it is impossible to judge whether the first principal component captures the large majority of variation or whether higher-order components encode independent structural shifts (e.g., infant versus senescence patterns) that a scalar speed function cannot reproduce at 50-year horizons.

    Authors: We agree that the manuscript should provide quantitative support for the one-dimensional claim. In the revised version, we will add the eigenvalues of the PCA applied to the Tucker scores, a scree plot, and the cumulative variance percentages. These additions will allow readers to evaluate the dominance of the first principal component and assess whether higher-order components capture independent structural shifts that might require more than a scalar speed function. revision: yes

  2. Referee: Cross-validation section: the leave-country-out procedure must specify exactly how the speed function and trajectory functions are estimated from training countries only, and what data-exclusion rules are applied to avoid leakage when the held-out countries' schedules are reconstructed. The current description leaves open the possibility that the reported bias and error reductions partly reflect in-sample structure rather than genuine out-of-sample forecasting skill.

    Authors: We will revise the cross-validation section to provide a precise, step-by-step description of the procedure. The Tucker decomposition will be performed solely on the training countries, and both the speed function and trajectory functions will be estimated exclusively from those training data. Held-out countries will be used only for evaluation, with explicit data-exclusion rules to prevent any leakage. This clarification will confirm the out-of-sample nature of the reported results. revision: yes

  3. Referee: Flow-field integration description: the scalar speed function is fitted on historical data; the manuscript does not demonstrate that this function remains valid when extrapolated to horizons beyond the training window, nor does it provide sensitivity checks on the integration step when the speed function is perturbed within its estimation uncertainty.

    Authors: The flow-field approach keeps forecasts within the empirically observed score manifold, which inherently limits extrapolation outside the range of historical dynamics. Nevertheless, we acknowledge the value of explicit sensitivity analysis. In the revision, we will add checks that perturb the speed function within its estimation uncertainty (e.g., via bootstrap resampling) and report the resulting variation in long-horizon forecasts, thereby demonstrating robustness. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's core derivation applies Tucker decomposition to the Human Mortality Database, performs PCA on the resulting scores to identify a one-dimensional structure, and constructs a flow-field model with scalar speed and trajectory functions for forecasting. This structure is discovered from training data and then used to generate out-of-sample forecasts. Performance is assessed via leave-country-out cross-validation on 9,507 held-out test points from separate countries, providing independent evaluation. No quoted step reduces the forecasting mechanism or reported metrics to the inputs by construction, self-definition, or self-citation load-bearing; the model is a data-driven parameterization evaluated externally to its fitting data.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 1 invented entities

The central claim rests on the Tucker decomposition of the Human Mortality Database, the identification of a one-dimensional flow via PCA, and fitted scalar speed and trajectory functions extracted from historical data patterns.

free parameters (2)
  • scalar speed function
    Determines the rate of advancement along the mortality level in score space and is extracted from the data.
  • trajectory functions
    Supply structural scores for Tucker reconstruction and are derived via PCA on the decomposition.
axioms (1)
  • domain assumption The mortality transition is essentially a one-dimensional flow in the low-dimensional Tucker score space.
    Identified through PCA reduction of the decomposed data.
invented entities (1)
  • flow field no independent evidence
    purpose: Models the temporal progression of mortality by integration through the score space to generate future schedules.
    Core new mechanism introduced to replace traditional time-series extrapolation.

pith-pipeline@v0.9.0 · 5559 in / 1502 out tokens · 50388 ms · 2026-05-15T00:34:52.746655+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 1 internal anchor

  1. [1]

    doi:10.1126/science.aat3119. U. Basellini, C. G. Camarda, and H. Booth. Thirty years on: A review of the Lee–Carter method for forecasting mortality.International Journal of Forecasting, 39(3):1033–1049,

    work page doi:10.1126/science.aat3119

  2. [2]

    doi:10.1016/j.ijforecast.2022.11.002. M. Betancourt. A conceptual introduction to Hamiltonian Monte Carlo.arXiv preprint,

    work page doi:10.1016/j.ijforecast.2022.11.002 2022

  3. [3]

    A Conceptual Introduction to Hamiltonian Monte Carlo

    doi:10.48550/arXiv.1701.02434. URLhttps://arxiv.org/abs/1701.02434. H. Booth and L. Tickle. Mortality modelling and forecasting: A review of methods.Annals of Actuarial Science, 3(1–2):3–43,

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1701.02434

  4. [4]

    doi:10.1017/S1748499500000440. H. Booth, L. Tickle, and L. Smith. Evaluation of the variants of the Lee–Carter method of forecasting mortality: A multicountry comparison.New Zealand Population Review, 31–32:13–34,

    work page doi:10.1017/s1748499500000440

  5. [5]

    doi:10.1007/s13524-019- 00785-3. S. J. Clark. Multi-dimensional mortality with exceptional mortality from armed conflict and pandemics (MDMx). arXiv preprint arXiv:2603.20518,

    work page doi:10.1007/s13524-019-

  6. [6]

    doi:10.1080/08898480500452109. X. Dong, B. Milholland, and J. Vijg. Evidence for a limit to human lifespan.Nature, 538(7624): 257–259,

    work page doi:10.1080/08898480500452109

  7. [7]

    doi:10.1038/nature19793. Y. Dong, F. Huang, and S. Haberman. Multi-population mortality forecasting using tensor decompo- sition.Scandinavian Actuarial Journal, 2020(8):754–775,

    work page doi:10.1038/nature19793 2020

  8. [8]

    doi:10.1080/03461238.2020.1740314. J. F. Fries. Aging, natural death, and the compression of morbidity.New England Journal of Medicine, 303(3):130–135,

    work page doi:10.1080/03461238.2020.1740314 2020

  9. [9]

    doi:10.1056/NEJM198007173030304. M. D. Hoffman and A. Gelman. The No-U-Turn sampler: Adaptively setting path lengths in Hamiltonian Monte Carlo.Journal of Machine Learning Research, 15(47):1593–1623,

    work page doi:10.1056/nejm198007173030304

  10. [10]

    doi:10.1016/j.csda.2006.07.028. R. J. Hyndman, H. Booth, and F. Yasmeen. Coherent mortality forecasting: The product-ratio method with functional time series models.Demography, 50(1):261–283,

    work page doi:10.1016/j.csda.2006.07.028 2006

  11. [11]

    doi:10.1007/s13524- 012-0145-5. V . Kannisto.Development of Oldest-Old Mortality, 1950–1990: Evidence from 28 Developed Countries, volume 1 ofOdense Monographs on Population Aging. Odense University Press, Odense,

    work page doi:10.1007/s13524- 1950

  12. [12]

    doi:10.1353/dem.2001.0036. R. D. Lee and L. R. Carter. Modeling and forecasting U.S. mortality.Journal of the American Statistical Association, 87(419):659–671,

    work page doi:10.1353/dem.2001.0036 2001

  13. [13]

    doi:10.1080/01621459.1992.10475265. N. Li and R. Lee. Coherent mortality forecasts for a group of populations: An extension of the Lee–Carter method.Demography, 42(3):575–594,

    work page doi:10.1080/01621459.1992.10475265 1992

  14. [14]

    doi:10.1353/dem.2005.0021. N. Li, R. Lee, and P . Gerland. Extending the Lee–Carter method to model the rotation of age patterns of mortality decline for long-term projections.Demography, 50(6):2037–2051,

    work page doi:10.1353/dem.2005.0021 2005

  15. [15]

    doi:10.1007/s13524-013-0232-2. J. Oeppen and J. W. Vaupel. Broken limits to life expectancy.Science, 296(5570):1029–1031,

    work page doi:10.1007/s13524-013-0232-2

  16. [16]

    doi:10.1126/science.1069675. S. J. Olshansky, B. A. Carnes, and C. Cassel. In search of Methuselah: Estimating the upper limits to human longevity.Science, 250(4981):634–640,

    work page doi:10.1126/science.1069675

  17. [17]

    doi:10.1126/science.2237414. S. J. Olshansky, B. J. Willcox, L. Demetrius, and H. Beltrán-Sánchez. Implausibility of radical life extension in humans in the twenty-first century.Nature Aging, 4(11):1635–1642,

    work page doi:10.1126/science.2237414

  18. [18]

    doi:10.1038/s43587-024-00702-3. O. Papaspiliopoulos, G. O. Roberts, and M. Sköld. A general framework for the parametrization of hierarchical models.Statistical Science, 22(1):59–73,

    work page doi:10.1038/s43587-024-00702-3

  19. [19]

    doi:10.1214/088342307000000014. D. Phan, N. Pradhan, and M. Jankowiak. Composable effects for flexible and accelerated probabilis- tic programming in NumPyro. InAdvances in Neural Information Processing Systems,

    work page doi:10.1214/088342307000000014

  20. [20]

    doi:10.1007/s13524-012-0193-x. A. E. Raftery, N. Lali´ c, and P . Gerland. Joint probabilistic projection of female and male life expectancy.Demographic Research, 30:795–822,

    work page doi:10.1007/s13524-012-0193-x

  21. [21]

    doi:10.4054/DemRes.2014.30.27. M. Russolillo, G. Giordano, and S. Haberman. Extending the Lee–Carter model: A three-way decom- position.Scandinavian Actuarial Journal, 2011(2):96–117,

    work page doi:10.4054/demres.2014.30.27 2014

  22. [22]

    doi:10.1080/03461231003611933. H. Ševˇ cíková and A. E. Raftery.bayesLife: Bayesian Projection of Life Expectancy,

    work page doi:10.1080/03461231003611933

  23. [23]

    doi:10.1007/978-3-319-26603-9_15. H. Ševˇ cíková, L. Alkema, and A. E. Raftery.bayesTFR: Bayesian Fertility Projection, 2024a. URL https://CRAN.R-project.org/package=bayesTFR. R package version 7.4-4. H. Ševˇ cíková, N. Li, and P . Gerland.MortCast: Estimation and Projection of Age-Specific Mortality Rates, 2024b. URLhttps://CRAN.R-project.org/package=Mor...

    work page doi:10.1007/978-3-319-26603-9_15 2024

  24. [24]

    doi:10.1038/nature08984. X. Zhang, F. Huang, F. K. C. Hui, and S. Haberman. Cause-of-death mortality forecasting using adaptive penalized tensor decompositions.Insurance: Mathematics and Economics, 111:193–213,

    work page doi:10.1038/nature08984

  25. [25]

    doi:10.1016/j.insmatheco.2023.05.003. 68

    work page doi:10.1016/j.insmatheco.2023.05.003 2023