General Bounds on Functionals of the Lifetime under Life Table Constraints

Here is an explanation of the paper "General Bounds on Functionals of the Lifetime under Life Table Constraints," translated into simple, everyday language with creative analogies.

The Big Picture: The "Missing Puzzle Piece" Problem

Imagine you are an insurance company selling a complex product called a Variable Annuity. Think of this like a retirement savings account that grows with the stock market but has a safety net (a guarantee) if the market crashes or if you pass away.

To price this product correctly, the insurance company needs to know exactly when people are likely to die. If you die tomorrow, the company pays a death benefit. If you live for 50 more years, they pay you a monthly income.

The Problem:
Insurance companies have "Life Tables." These are like statistical maps that tell them: "Out of 1,000 people aged 40, 990 will survive to age 41."
However, these maps are blurry in the middle. They know the start and end of the year, but they don't know when during that year the 10 people died. Did they all die on January 1st? Did they all die on December 31st? Or did they die evenly throughout the year?

The Old Way (The "Guessing Game"):
Traditionally, actuaries (the math wizards who price insurance) would just guess a pattern. They might assume deaths happen evenly (Uniform Distribution), or that they happen faster at the start of the year (Balducci), or slower (Constant Force).

The Risk: If they guess wrong, they might charge too little (losing money) or too much (losing customers). It's like trying to drive a car by only looking at the road every 100 miles and guessing the bumps in between.

The New Way (This Paper's Solution):
Instead of guessing a specific pattern, the authors say: "Let's stop guessing. Let's calculate the worst-case and best-case scenarios that are still mathematically possible given our known data."

They built a framework to find the absolute highest price and the absolute lowest price for these insurance contracts, assuming any possible pattern of deaths that fits the known yearly survival numbers.

The Two Approaches: The "Strict" vs. The "Relaxed"

The paper offers two ways to do this, like two different levels of security checks.

1. The Strict Approach (The "Perfect Match" Rule)

The Analogy: Imagine a strict teacher who says, "Every single student in this class must get exactly 90% on the test."
How it works: The math forces the model to ensure that every single possible path of mortality matches the life table exactly.
The Result: This is very rigid. It often leads to "deterministic" results (the answer is fixed and predictable).
- Example: For a "Guaranteed Minimum Income Benefit" (where you get paid monthly while alive), the worst-case scenario turns out to be: "Everyone dies on the very last day of the year." This maximizes the time the insurer has to pay you, but minimizes the time they have to wait for you to die.
- Why it matters: It gives a "hard ceiling" and "hard floor" for prices. If your price is outside these bounds, your math is broken.

2. The Relaxed Approach (The "Average Match" Rule)

The Analogy: Imagine a coach who says, "The team's average score must be 90%, but individual players can score higher or lower."
How it works: This is more realistic. It allows the mortality rate to fluctuate wildly from year to year or day to day, as long as the average outcome matches the life table.
The Result: This creates a much wider range of possible prices. It accounts for the fact that in the real world, mortality isn't perfectly predictable.
The Math Magic: The authors used advanced "Stochastic Optimal Control" (think of it as a high-speed computer simulation that tries millions of different death schedules) to find the boundaries of this wider range. They used a "dual approach" (like looking at a problem from the inside out and the outside in) to solve the complex equations.

What Did They Find? (The "Aha!" Moments)

The authors ran computer simulations using real Belgian life tables and found some fascinating things:

The "Blind Spot" is Real: The difference between the "best guess" (using old assumptions) and the "worst-case scenario" can be huge, especially for older people or long-term contracts.
- Analogy: If you are 40, the guess doesn't matter much. But if you are 80, the difference between assuming people die on Jan 1st vs. Dec 31st changes the price of the insurance significantly.
Asymmetry is Key:
- For "Living" Benefits (Income): The lower bound (worst case for the insurer) is very low. If people die earlier than expected, the insurer saves a lot of money. The upper bound is closer to the guess.
- For "Death" Benefits (Payouts): The upper bound is very high. If people die earlier than expected, the insurer has to pay out the death benefit sooner, costing them more.
- Takeaway: The risk isn't symmetrical. You can't just add a small "safety margin"; you need to understand the direction of the risk.
The "Model-Free" Advantage:
- The authors didn't need to assume a specific formula for how people age (like the famous Gompertz law). They just used the raw data. This makes their results robust. It's like saying, "Regardless of how the engine works, here is the maximum speed this car can possibly go given its fuel tank size."

Why Should You Care?

This paper gives insurance companies a new safety net.

Instead of saying, "We assume people die according to Formula X," they can now say:

"Even if our assumptions about when people die during the year are completely wrong, our prices will never be lower than $X** or higher than **$ Y. We have quantified the risk of our own ignorance."

It transforms mortality risk from a "black box" guess into a measurable, bounded range. It's the difference between driving blindfolded and driving with a radar that tells you exactly how close you are to the cliff, regardless of the fog.

Summary in One Sentence

This paper provides insurance companies with a mathematical "fence" that defines the absolute highest and lowest prices they should charge for life insurance, ensuring they are safe even if their guesses about when people die during the year turn out to be wrong.

Here is a detailed technical summary of the paper "General Bounds on Functionals of the Lifetime under Life Table Constraints" by Jean-Loup Dupret and Edouard Motte.

1. Problem Statement

In life insurance, actuaries rely on life tables which provide survival probabilities only at integer ages ( $_1p_x, _2p_x, \dots$ ). However, many complex actuarial products (e.g., variable annuities, Guaranteed Minimum Income Benefits) depend on the full continuous-time distribution of the lifetime, specifically the timing of deaths between integer ages.

To bridge this gap, practitioners traditionally use:

Fractional Age Assumptions: Interpolation methods like Uniform Distribution of Deaths (UDD), Constant Force of Mortality (CFM), or Balducci.
Parametric Mortality Models: Deterministic (Gompertz-Makeham) or Stochastic (Lee-Carter, CBD) models.

The Core Issue: These approaches introduce model risk. The valuation of liabilities is highly sensitive to the specific interpolation or model chosen, yet this sensitivity is rarely quantified. The paper addresses the need for a model-free framework to derive robust upper and lower bounds on actuarial functionals that are consistent with observed life table data, without imposing arbitrary structural assumptions on the mortality process within a year.

2. Methodology

The authors propose a stochastic optimal control framework to characterize the set of all possible mortality trajectories consistent with observed data. They analyze two distinct formulations:

A. The Strict (Pathwise) Setting (Section 4)

Constraint: The mortality trajectory must almost surely (a.s.) reproduce the one-year survival probabilities ( $\hat{p}_j$ ) given in the life table for every integer year.
Control: The force of mortality is treated as a control process $\mu = (\mu_{x,c}, \mu_{x,d})$ consisting of a continuous component and discrete jumps.
Mechanism: The constraint forces the total survival factor over any integer interval $[k, k+1]$ to equal $\hat{p}_k$ with probability 1.
Resulting Dynamics: This strict constraint effectively reduces the optimal control problem to a deterministic one, where the optimal mortality rates are piecewise constant or concentrated at specific points (endpoints of intervals).

B. The Relaxed (Expectation) Setting (Section 5)

Motivation: The strict setting is often too restrictive for real-world applications where mortality deviates from the table in expectation but not perfectly in every realization.
Constraint: The mortality intensity $\mu_x$ must satisfy the life table constraints in expectation:
$\mathbb{E}^P \left[ e^{-\int_0^j \mu_x(s) ds} \right] = \hat{p}_j, \quad j=1, \dots, n$
Regularization: The unconstrained expectation problem is ill-posed. To solve this, the authors introduce a regularized set of admissible controls $\mathcal{A}^\delta_{\text{bounded}}$ , where the mortality rate is bounded by time-dependent functions $\mu_{x,a}(t) \pm \delta$ .
Solution Approach:
1. Dual Formulation: The problem is transformed using Lagrange multipliers $\lambda \in \mathbb{R}^n$ to handle the expectation constraints.
2. Dynamic Programming: The inner optimization problem is solved via the Hamilton-Jacobi-Bellman (HJB) equation.
3. Convergence: The authors prove that as the bound parameter $\delta \to \infty$ , the solutions to the regularized problems converge to the solutions of the original unconstrained expectation problems.

Financial Framework

The valuation assumes a complete financial market with stochastic interest rates (Vasicek/Hull-White type) and stock prices. The pricing measure is the Minimal Martingale Measure (MMM), which is standard for locally risk-minimizing hedging in incomplete markets (where mortality risk is not fully hedgeable).

3. Key Contributions

Model-Free Bounds: The paper derives sharp upper and lower bounds for actuarial functionals (GMAB, GMIB, GMDB) that depend only on the observed life table and the financial model, eliminating the need for specific fractional age assumptions or parametric mortality models.
Two Complementary Frameworks:
- Strict Bounds: Provide the theoretical limits if mortality must perfectly match the table pathwise.
- Relaxed Bounds: Provide robust bounds allowing for mortality deviations, provided the average behavior matches the table.
Analytical Characterization of Optimal Controls:
- For GMAB (Guaranteed Minimum Accumulation Benefits), the authors prove that the bounds are trivial (equal to the standard price) because the payoff depends only on survival to maturity, not the timing of death within the year.
- For GMIB (Guaranteed Minimum Income Benefits) and GMDB (Guaranteed Minimum Death Benefits), the optimal mortality strategies are non-trivial.
  - Upper Bound (GMIB): Achieved by delaying deaths as much as possible (concentrating mortality at the end of each year).
  - Lower Bound (GMIB): Achieved by advancing deaths as much as possible (concentrating mortality at the start of each year).
  - GMDB: Similar logic applies but depends on the convexity of the death benefit payoff.
Numerical Validation: The authors implement the HJB equations to compute these bounds for Belgian female life tables, demonstrating the magnitude of model risk.

4. Key Results

Impact of Age and Maturity:
- For younger policyholders and short maturities, the difference between standard fractional age assumptions (UDD, CFM, Balducci) and the derived bounds is negligible.
- For older policyholders and long maturities, the bounds diverge significantly from standard assumptions. The "model risk" (the gap between the worst-case and best-case scenarios) becomes substantial.
Asymmetry of Bounds:
- Survival-contingent products (GMIB, GMAB): The lower bound is significantly lower than the baseline (Balducci) price. This reflects the high sensitivity of survival benefits to increased mortality (earlier deaths reduce the payout). The upper bound is closer to the baseline.
- Death-contingent products (GMDB): The upper bound is significantly higher than the baseline. Increased mortality increases the present value of death benefits. The lower bound remains close to the baseline.
Effect of Constraints:
- Expectation vs. Almost Sure: The relaxed (expectation) bounds are wider than the strict (almost sure) bounds because they allow for a broader set of mortality trajectories.
- Intermediate Constraints: Adding constraints at intermediate times (not just at maturity) significantly tightens the bounds, particularly for GMIB and GMDB, by ruling out extreme mortality paths that satisfy the final condition but produce extreme liabilities along the path.
Convergence: The numerical experiments confirm the theoretical result that the relaxed bounds converge rapidly as the control bounds ( $\delta$ ) increase.

5. Significance and Practical Implications

Risk Management Tool: The framework provides insurers with a robust, model-free tool to quantify the sensitivity of their liabilities to the "within-year" timing of deaths. It answers: "How much could my liability change if the actual mortality distribution within a year differs from my chosen assumption (e.g., UDD)?"
Pricing and Reserving: Instead of relying on a single point estimate based on an arbitrary assumption, insurers can calculate a price interval. Any price outside this interval is inconsistent with the observed life table.
Capital Adequacy: The bounds allow insurers to quantify the worst-case loss arising from model misspecification, aiding in capital allocation and stress testing without needing to specify a complex stochastic mortality model.
Generalizability: While the paper focuses on variable annuities, the methodology (HJB with expectation constraints) is applicable to any life-contingent liability where the timing of death within an interval matters.

In summary, Dupret and Motte provide a rigorous mathematical framework that replaces arbitrary interpolation assumptions with optimization-based bounds, offering a more robust approach to managing mortality risk in life insurance.