Bifurcation in optimal retirement

Bushra Shehnam Ashraf; Thomas S. Salisbury

REVIEW 5 major objections 5 minor 8 references

Bifurcation in optimal retirement

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read At a critical initial wealth, every retirement age gives the same utility, while above it retiring now is optimal and below it never retiring is optimal.

desk verdict A genuinely interesting numerical phenomenon in a simple retirement model, but the equal-utility continuum is not yet established because the verification theorem is explicitly omitted. read the letter →

arxiv 2506.02155 v1 pith:2YY2OL7H submitted 2025-06-02 q-fin.PM

Bushra Shehnam Ashraf , Thomas S. Salisbury This is my paper

classification q-fin.PM

keywords optimalretirementbifurcationCobb-DouglasutilitystoppingfreeboundaryGompertzmortalityHamilton-Jacobi-Bellmanequationlifecycleconsumption

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 asks when a worker should retire if utility comes from a Cobb-Douglas bundle of consumption and leisure, mortality follows a deterministic Gompertz hazard, and retirement wealth must fund all post-retirement spending. The paper claims the optimal policy bifurcates: high initial wealth makes immediate retirement optimal, low initial wealth makes never retiring optimal, and exactly one critical initial wealth connects the two. At that critical level, the worker may stay uncommitted for a while and then retire at any later time, and every such wealth path yields identical expected utility. The claim is supported by numerical solutions of the Hamilton-Jacobi-Bellman equations rather than by a verification theorem proving those solutions are the true optimum.

What carries the argument

The argument runs through two coupled Hamilton-Jacobi-Bellman equations, one for the pre-retirement regime and one for the post-retirement regime, linked by smooth pasting at the optimal retirement boundary. The post-retirement equation simplifies because the value function is scale-invariant: $V(t,w)=F(t)w^{1-\gamma}/(1-\gamma)$, which reduces the post-retirement problem to a linear ODE for $f(t)=F(t)^{\tilde\gamma}$. The pre-retirement equation is solved numerically with an upwind explicit finite-difference scheme after a log transform of low wealth. The scaling property is what makes the uncommitted curve appear: integrating forward from any retirement time gives trajectories that are scalar multiples of each other, while integrating backward makes them coalesce into the single curve where the continuum of equal-utility strategies lives.

What would settle it

Follow the uncommitted curve from the critical wealth to several different retirement ages, switch to post-retirement consumption at each, and compute lifetime utility; exact equality across all stopping ages is the paper's prediction, so any discrepancy larger than the numerical scheme's error would falsify the continuum, and a grid-refinement study near the curve would reveal whether the discrepancy is numerical.

Watch

Extended reading notes

Core claim

In the age-wealth plane the optimal stopping problem has three regions. Above a free boundary $\bar w_t$, immediate retirement is optimal; between that boundary and a second 'uncommitted' curve, the worker saves with retirement in mind; below the uncommitted curve, never retiring is optimal. The paper's central discovery is that the uncommitted curve begins at a critical initial wealth $\tilde w$ where optimality does not select a retirement age. An infinitesimal move upward from the curve commits the worker to a wealth path that leads to retirement, an infinitesimal move downward commits to never retiring, and all these paths give the same value. Because the post-retirement value function has the scaling form $V(t,w)=F(t)w^{1-\gamma}/(1-\gamma)$, forward wealth paths are scalar multiples of one another while backward paths from any chosen retirement time coalesce onto the same uncommitted curve.

Load-bearing premise

The claim collapses if the smooth solutions of the optimality equations do not give the true maximum of the worker's lifetime utility, and the paper does not prove that verification step, relying instead on numerical solutions it describes as being at the limits of stability.

Editorial extensions

If this is right

A worker whose initial wealth is exactly the critical value $\tilde w$ has no uniquely optimal retirement date, so models that predict a single retirement age will miss a set of equally good plans there.
A small change in initial wealth near $\tilde w$ flips the optimal plan between retiring and never retiring, so wealth-targeting advice is most sensitive exactly at the bifurcation point.
The deterministic-mortality Cobb-Douglas model can reproduce commonly cited retirement targets (wealth 7 to 12 times income and retirement in the mid-50s to mid-60s) with a single calibrated leisure parameter.
Adding an exogenous pension stream would make wealth hit zero near the end of life and is expected to mask the bifurcation, so the phenomenon is tied to retirement funded purely by savings.
The same qualitative sensitivity should appear in more complex life-cycle models even when those models do not exhibit a clean continuum.

Reading between the lines

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

The paper leaves implicit that the continuum of equal-utility trajectories is a degeneracy created by the deterministic hazard rate and the scale-invariant value function; a stochastic hazard rate may generically break this degeneracy into a unique threshold, so the bifurcation could be a boundary case rather than a robust feature.
Because the calibrated leisure parameter is reported at the limits of numerical stability, a testable extension is to recompute the uncommitted curve with a different solver or a finer grid near zero wealth and check whether the curve shifts materially.
A policy reading the paper does not state: a rule that says 'retire once wealth reaches X times income' behaves discontinuously at the critical wealth, so small measurement error in wealth can produce very different recommended behavior near the threshold.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know this paper for one reason: it claims a clean bifurcation in optimal retirement—above a boundary you retire immediately, below a lower curve you never retire, and on a middle “uncommitted” curve any retirement time yields the same utility. That last piece, a continuum of equal-utility plans at a critical wealth, is new compared to the cited Farhi-Panageas, Koo-Koo-Shin, and Menoncin-Vergalli work. The framework itself is standard Cobb-Douglas HJB, but the phenomenon is cleanly laid out.

What the paper does well: the setup is simple and readable. Cobb-Douglas utility with a leisure multiplier, no bequest motive, deterministic Gompertz mortality. The post-retirement problem reduces to an ODE through scaling, and the pre-retirement problem is a standard first-order HJB. The three-region decomposition (retire above, save between, never retire below) is a crisp way to see the discontinuity. The authors also deserve credit for being explicit that they do not prove a verification theorem and that their calibrated parameters sit at the numerical scheme's stability limit.

The soft spots are real. Section 3.2 explicitly defers the verification theorem. That is load-bearing: the central claim is that the coalescing backward trajectories in Figure 3 have identical lifetime utility. Without a proof that the smooth solution of the HJB equations is the value function (and that the free boundary is the optimal stopping rule), the continuum is a feature of the discretized equations, not of the optimization problem. The authors' own warning about stability at low wealth makes the numerical origin of the continuum a live possibility. The calibration of l=6.49 to Fidelity targets is also partly circular—they choose l to match observed retirement ages and wealth multiples, then present those outputs as the model reproducing reality. That weakens external-validity claims but does not bear on the qualitative existence of the bifurcation. No code or data is supplied, so the 48-hour computation is hard to check independently.

Still, the idea is plausible, the paper is honest, and the missing verification step is addressable. With a verification theorem (or a clear reframing as a numerical conjecture) this could be a solid contribution. It deserves a serious referee, who should push exactly on that point. My own verdict would be conditional: the math setup is coherent, the result is not yet proven, but the paper earns the space to be revised. I would bring it to a reading group as an example of why verification matters in optimal stopping.

Referee Report

5 major / 5 minor

Summary. This paper formulates a deterministic-mortality lifecycle model in which an individual earns one unit of labour income until a chosen retirement time, consumes from wealth, and receives post-retirement leisure utility through a Cobb-Douglas utility. The authors derive a post-retirement value function with a scaling reduction to an ODE, solve the pre-retirement HJB equation numerically with an upwind finite-difference scheme, and calibrate the leisure multiplier l so that a 30-year-old with wealth 1 retires at age 55-65 with wealth 7-12 times income. The central claim, stated in the abstract and Section 5, is that there is a bifurcation: low initial wealth makes never retiring optimal, high initial wealth makes immediate or planned retirement optimal, and at a critical initial wealth there is a continuum of wealth trajectories with identical utility. The paper identifies this structure numerically through the coalescence of backward trajectories from the retirement boundary in Figure 3.

Significance. The paper identifies a potentially interesting phenomenon in a simple optimal stopping model, and the post-retirement scaling argument leading to the ODE (18) is clean and useful. The authors are also candid about the paper's limitations, explicitly deferring a verification theorem and warning that the calibrated parameters are at the limits of numerical stability. However, the central bifurcation claim is not yet established: the value function is not proven to be represented by the numerical HJB solution, and the equal-utility property of the coalesced trajectories is asserted rather than verified. If the missing verification and numerical convergence analysis are supplied, the result would be a worthwhile contribution to the retirement timing literature.

major comments (5)

[Section 3.2] After Eq. (18) the authors write that one could prove a verification theorem but that instead they will focus on numerical solutions. This is the load-bearing gap for the central claim. Equations (13), (16), and (18) are only necessary optimality conditions; for a first-order HJB equation, smooth solutions are not unique without a comparison or verification argument, so the computed retirement boundary, the uncommitted curve, and the critical wealth are not yet shown to correspond to the value function in (7)-(8). A verification theorem, or at least a direct numerical check of the Bellman inequality dZ_t <= 0 along the proposed optimal paths, is needed before the optimality statements in the abstract and Section 5 can be accepted.
[Section 4, Figure 3] The observation that backward trajectories started at (t, \bar w_t) coalesce is a geometric statement about the numerical characteristics; it does not by itself imply that different retirement ages produce equal values of the objective in (7). The forward curves being scalar multiples of each other follows from the scaling property of the post-retirement value function, but scaling of wealth paths does not equate lifetime utilities across distinct retirement ages. To support the abstract's claim of a continuum of wealth trajectories with identical utilities, the paper should report the computed lifetime utility along the uncommitted curve for several retirement ages and show that it is the same and is maximal; otherwise the equality is inferred, not demonstrated.
[Section 3.3 and Section 5] The numerical evidence for the bifurcation is presented in a scheme that the authors themselves describe as unstable at low wealth: T=110 is chosen 'for the sake of numerical stability at low wealth levels', the calibration section reports a 'stability issue at small wealth values', and the conclusion states that the parameter values 'lie at the limits of stability for our numerical scheme'. Because the uncommitted curve and the critical wealth are only observed in this computation, the paper needs a grid-convergence study with error estimates, not just a statement that a coarser grid gave similar pictures, and a sensitivity analysis with respect to the discretization parameters dt, dw, and dy. Without this, the continuum can be a numerical artifact.
[Section 3.4] The calibration of l to the Fidelity targets (retirement age 55-65 and retirement wealth 7-12 times income) makes the later statement that the model reproduces realistic behaviour partly circular: the targets are inputs, not predictions. In addition, the chosen value l=6.49 lies at the upper end of the range 6.25-6.50 over which 'reasonable agreement' was obtained, and no sensitivity analysis is reported for the bifurcation or the uncommitted curve as l varies. Given that the central phenomenon may exist only over a narrow range of l, the paper should show how the critical wealth and the continuum depend on l and on the other calibrated parameters.
[Section 3.1] The free-boundary condition for \bar w_t is only described as 'smooth pasting' and is never written down. The numerical procedure 'compares' the pre-retirement solution with the post-retirement value to locate the boundary, but the value-matching and derivative conditions that define \bar w_t are not stated, nor is it explained how the finite-difference scheme enforces them. Since the coalescence in Figure 3 starts from points (t, \bar w_t), an unambiguous characterization of \bar w_t is needed to make the numerical experiment reproducible and the bifurcation claim well-posed.

minor comments (5)

[Section 3] The notation in (3) uses l both for the leisure process and for the constant post-retirement leisure parameter; the normalization l_1=1 is clear, but the distinction should be stated explicitly to avoid confusion.
[References] The in-text citation 'Francesco and Sergio [7]' does not match the reference list entry 'Menoncin and Vergalli'; the citation should be corrected.
[Figures and captions] Figure 2's caption contains a duplicated word ('for for'), and Figure 1 uses \tau=110 while the text uses T=110; the notation should be unified.
[Introduction and abstract] The term 'bifurcation' is used informally; a formal definition of the critical wealth and of the continuum of equal-utility trajectories would improve precision.
[Numerical reproducibility] No code or data availability statement is provided; since the central results are numerical, making the solver available would improve reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

Calibration of l to Fidelity targets makes the 'reproduce realistic behaviour' conclusion partly circular; the central bifurcation claim is not.

fitted input called prediction [Section 3.4 (Calibrating l), validated in Section 5 (Conclusions)]
"We searched for an l such that with initial wealth w0 = 1 at age 30, we would see a retirement age in the range 55–65, and a wealth at retirement of 7 to 12 times annual labour income. We drew the latter numbers from an article 'How much do I need to save for retirement?' [4] at Fidelity.com. ... We indeed find parameter values for which this is the case, though they lie at the limits of stability for our numerical scheme."

l is the single calibrated parameter, chosen specifically so the model output (retirement age 55–65, retirement wealth 7–12 times income) matches the Fidelity targets. The concluding statement that the model 'indeed' finds parameter values reproducing realistic behavior is therefore not an independent test of the model: it restates the calibration target. The retirement ages and wealth multiples displayed in Figures 1–3 are outputs of a model containing this fitted l, so any claim that the model 'reproduce[s] realistic behaviour' with respect to those targets reduces to the fitting step. This does not affect the central bifurcation, which was not a calibration target, so the circularity is partial.

full rationale

The paper's derivation chain for the central bifurcation is self-contained: the value functions and HJB equations (7)–(18) are derived from the model primitives, and the three-region / continuum picture is a numerical output of those equations (Figures 2–3), not an input. The coalescence of backward trajectories from the retirement boundary is a mathematical property of the characteristics, not a fitted target. No load-bearing self-citation occurs: [1] merely identifies the thesis, and [2], [3], [5], [6], [7] are background. The explicit omission of a verification theorem in Section 3.2 ('One could prove a verification theorem, showing that a smooth solution of our HJB equations will indeed represent the solution to our optimization problem. Instead of doing this we will focus on the behaviour or solutions to our equations, obtained numerically.') is a load-bearing gap for the claim of identical utilities: without it, the numerical HJB solution is not rigorously tied to the original optimization problem. Similarly, Section 5's 'at the limits of stability for our numerical scheme' caveat weakens the numerical evidence. These are correctness risks, not circularity, because the HJB solution is not defined in terms of the claimed conclusion. The one genuine circular step is the calibration of l: l is chosen to match the Fidelity retirement-age and wealth targets, and the Conclusion then states the model 'indeed find[s] parameter values' reproducing realistic behaviour. That validation is equivalent to the fitting step. Since this does not affect the central non-uniqueness claim, the overall circularity is partial and limited to the calibration-validation framing.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or economic entities. The main free parameter is the leisure multiplier l, calibrated to external retirement wealth guidelines. The most consequential unstated input is the assumed validity of the HJB verification theorem, which the authors leave unproved.

free parameters (1)

leisure multiplier l = 6.49
Chosen in Section 3.4 so that with initial wealth w0=1 at age 30 the model gives retirement age 55-65 and retirement wealth 7-12 times labor income, matching Fidelity guidelines. All numerical results use this calibrated value.

assumptions (3)

domain assumption Cobb-Douglas utility of consumption and leisure with post-retirement leisure as a constant multiplicative factor.
Equations (4)-(6) define the utility structure; the entire analysis depends on this functional form.
domain assumption Deterministic Gompertz mortality, no bequest motive, no pension income, and investment only in risk-free assets.
Stated in Sections 1 and 3; these assumptions simplify the model and make the bifurcation visible.
ad hoc to paper A smooth solution of the HJB equations represents the true value function (verification theorem), despite being unproved.
Section 3.2 explicitly says the verification theorem could be proved but is not, while the central optimality and non-uniqueness claims rely on numerical HJB solutions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bifurcation in optimal retirement." pith.science (2026). https://pith.science/paper/2YY2OL7H

@misc{pith2026250602155,
  author       = {Pith},
  title        = {Pith review of: Bifurcation in optimal retirement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2YY2OL7H}},
  note         = {Machine review of arXiv:2506.02155}
}

read the original abstract

We study optimal consumption and retirement using a Cobb-Douglas utility and a simple model in which an interesting bifurcation arises. With high wealth, individuals plan to retire. With low wealth they plan to never retire. At a critical level of initial wealth they may choose to defer this decision, leading to a continuum of wealth trajectories with identical utilities.

Figures

Figures reproduced from arXiv: 2506.02155 by the authors.

**Figure 1.** Optimal retirement age and wealth for different values of annual risk free rate. Assuming γ = 2, ρ = 2.5%, α = 0.5, l = 6.49, b = 9.44, m = 88.82, τ = 110 [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

**Figure 2.** Wealth dynamics for different wealth levels at age 30. Assuming r = 2.5%, ρ = r, γ = 2, α = 0.5, l = 6.49, b = 9.44, m = 88.82, τ = 110 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Wealth dynamics for different retirement ages. Assuming r = 2.5%, ρ = r, γ = 2, α = 0.5, l = 6.49, b = 9.44, m = 88.82, τ = 110 We see that the time-wealth state space actually consists of three regions. Above the blue optimal boundary, optimal behaviour is to retire right away. Between the blue curve and the magenta curve, optimal behaviour is to save with the intention of retiring. Below the magenta uncommitted cu… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages

[1]

B. S. Ashraf. Voluntary retirement and optimal consumption in a stochastic mortality envi- ronment.Ph.D. Thesis, York University, 2023

work page 2023
[2]

Bodie, R

Z. Bodie, R. C. Merton, and W. F. Samuelson. Labor supply flexibility and portfolio choice in a life cycle model.Journal of Economic Dynamics and Control, 16(3-4):427–449, 1992

work page 1992
[3]

Farhi and S

E. Farhi and S. Panageas. Saving and investing for early retirement: A theoretical analysis. Journal of Financial Economics, 83(1):87–121, 2007

work page 2007
[4]

How much do I need to retire?

Fidelity Investments. How much do I need to retire? . www.fidelity.com/viewpoints/retirement/how-much-money-do-i-need-to-retire. Accessed: 2022-09

work page 2022
[5]

J. L. Koo, B. L. Koo, and Y. H. Shin. An optimal investment, consumption, leisure, and voluntary retirement problem with Cobb-Douglas utility: Dynamic programming approaches. Applied Mathematics Letters, 26(4):481–486, 2013

work page 2013
[6]

G. Kula. Optimal retirement decision.Ph.D. Thesis, 2003

work page 2003
[7]

Menoncin and S

F. Menoncin and S. Vergalli. Optimal stopping time, consumption, labour, and portfolio deci- sion for a pension scheme.Journal of Economics, 132:67–98, 2021

work page 2021
[8]

M. A. Milevsky.The calculus of retirement income : financial models for pension annuities and life insurance. Cambridge University Press, 2006. CANNEX Financial Exchanges Limited, bushra.sashraf@gmail.com Dept. of Mathematics and Statistics, York University, salt@yorku.ca 9

work page 2006

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

B. S. Ashraf. Voluntary retirement and optimal consumption in a stochastic mortality envi- ronment.Ph.D. Thesis, York University, 2023

work page 2023

[2] [2]

Bodie, R

Z. Bodie, R. C. Merton, and W. F. Samuelson. Labor supply flexibility and portfolio choice in a life cycle model.Journal of Economic Dynamics and Control, 16(3-4):427–449, 1992

work page 1992

[3] [3]

Farhi and S

E. Farhi and S. Panageas. Saving and investing for early retirement: A theoretical analysis. Journal of Financial Economics, 83(1):87–121, 2007

work page 2007

[4] [4]

How much do I need to retire?

Fidelity Investments. How much do I need to retire? . www.fidelity.com/viewpoints/retirement/how-much-money-do-i-need-to-retire. Accessed: 2022-09

work page 2022

[5] [5]

J. L. Koo, B. L. Koo, and Y. H. Shin. An optimal investment, consumption, leisure, and voluntary retirement problem with Cobb-Douglas utility: Dynamic programming approaches. Applied Mathematics Letters, 26(4):481–486, 2013

work page 2013

[6] [6]

G. Kula. Optimal retirement decision.Ph.D. Thesis, 2003

work page 2003

[7] [7]

Menoncin and S

F. Menoncin and S. Vergalli. Optimal stopping time, consumption, labour, and portfolio deci- sion for a pension scheme.Journal of Economics, 132:67–98, 2021

work page 2021

[8] [8]

M. A. Milevsky.The calculus of retirement income : financial models for pension annuities and life insurance. Cambridge University Press, 2006. CANNEX Financial Exchanges Limited, bushra.sashraf@gmail.com Dept. of Mathematics and Statistics, York University, salt@yorku.ca 9

work page 2006