📈 economics

Dynamic Programming Treatment of the Solow and Ramsey–Cass–Koopmans Growth Models

This paper bridges the Solow and Ramsey–Cass–Koopmans growth models to the Bellman equation to demonstrate that while the asymptotic convergence rate of fitted value function iteration is determined by the discount factor, selecting an initial value-function guess closer to the true solution significantly reduces the number of iterations required to reach a target error.

Original authors: Ahmad Y. A. Thabaineh, Adnan Daraghmeh, Alvaro H. Salas

Published 2026-10-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ahmad Y. A. Thabaineh, Adnan Daraghmeh, Alvaro H. Salas

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Economic growth is the story of how societies become richer over time, producing more goods and services for their people. For decades, economists have relied on two main stories to explain this process. The first story, known as the Solow model, treats the amount of money people save and invest as a fixed rule, like a tax rate set by a government. It assumes that if a country saves a certain percentage of its income every year, that decision is simply made for them. The second story, called the Ramsey model, is more complex. It suggests that people are smart planners who decide how much to save and how much to spend based on what will make them happiest over their entire lives. This second story is harder to solve because it requires calculating the best possible path for a household that lives forever, balancing the desire to enjoy life today against the need to prepare for tomorrow.

To solve these complex planning problems, economists use a mathematical tool called dynamic programming. Imagine trying to find the best route through a vast, foggy city. You cannot see the whole map at once, so you make a guess about the best path, check if it works, and then adjust your guess based on what you learned. You repeat this process, refining your route step by step, until you are confident you have found the shortest way. In economics, this process is called value function iteration. The computer starts with a rough guess about how much value a household gets from having a certain amount of wealth, then repeatedly updates that guess to see if it can do better. The question researchers often face is not whether this method eventually works, but how many steps it takes to get there. Does the starting guess matter, or does the computer just grind through the same number of steps regardless of where it begins?

A team of researchers from universities in Palestine and Colombia set out to answer this specific question using a well-known economic scenario where the correct answer is already known. They chose a simplified version of the Ramsey model where the math can be solved exactly by hand, giving them a perfect target to aim for. They wanted to see if starting the computer's guessing process with a specific type of educated guess would make it reach the correct answer faster than starting with a different, equally reasonable guess. They did not invent a new way to solve the problem; instead, they carefully measured how the choice of the first guess changed the speed of the solution.

The researchers ran their simulations on a computer using a grid of 150 points to represent different levels of wealth. They tested two different starting guesses. The first guess was a formula that looked somewhat like a logarithmic curve, while the second was a different curve that rose more slowly. Both guesses were reasonable starting points, but they were not the same. When they let the computer run its standard updating process, the results showed a clear difference. The second starting guess allowed the computer to reach a high level of accuracy in 28 steps. The first guess required 37 steps to reach that same level of accuracy. This means that simply choosing a better starting point saved the computer about 24 percent of the work.

However, the researchers found that this advantage was temporary. Once the computer got past the first few steps, both starting guesses began to improve at exactly the same speed. The gap between them did not grow or shrink; it stayed constant. The better starting guess was simply closer to the final answer to begin with, so it had less distance to travel. It was not a faster runner, just one that started nearer to the finish line. This distinction is important because it tells economists that while a good starting point saves time, it does not change the fundamental speed at which the computer learns.

The study also uncovered a hidden problem with how these computer models handle the very smallest amounts of wealth. When the researchers looked at the entire range of wealth, including the tiniest amounts near zero, the computer's accuracy hit a hard limit. No matter how many steps the computer took, the error near the bottom of the scale did not disappear. This happened because the computer was using a grid with evenly spaced points, which is like trying to draw a very sharp curve using only straight lines. Near zero, the curve is so steep that the straight lines cannot follow it closely, creating a permanent blur. The researchers showed that if they wanted to remove this blur, they could not just run the computer for more steps; they would have to use a finer grid with more points.

Furthermore, the team discovered that the computer often finds the best spending plan much faster than it finds the exact value of that plan. In many economic studies, the goal is to know exactly how much happiness a person gets from their wealth. But if the goal is simply to know how much a person should spend or save, the computer can find that answer very quickly, even while it is still struggling to pin down the exact value. This suggests that for many practical purposes, researchers might not need to wait for the computer to finish its full calculation.

Finally, the researchers compared their method of simply guessing and checking against a more advanced technique called policy iteration, which is like a smarter way of navigating the city. They found that using this advanced technique reduced the number of steps needed far more than simply picking a better starting guess did. While a good starting point is helpful, using a better algorithm is a much more powerful way to save time. The study concludes that when economists run these models, they should be careful to report not just the final result, but also how they started, how they measured their progress, and how finely they drew their grid. These details matter because they determine how much work the computer actually had to do to get the answer.

Technical Summary: Dynamic Programming Treatment of the Solow and Ramsey–Cass–Koopmans Growth Models

Problem Statement
The paper addresses a specific implementation detail in the numerical solution of optimal growth models: the impact of the initial value-function guess on the computational effort required for Fitted Value Function Iteration (VFI). While contraction mapping theory guarantees that admissible starting values converge to the same unique fixed point, it does not specify the number of iterations (Bellman sweeps) required to reach a given tolerance. The authors investigate whether the choice of the initial guess significantly alters this computational cost in the context of the Ramsey–Cass–Koopmans model with Cobb–Douglas production and logarithmic utility.

Methodology
The study employs a self-contained expository framework followed by a rigorous numerical analysis:

Theoretical Foundation: The authors derive the Solow–Swan and Ramsey–Cass–Koopmans models, transitioning from continuous-time Hamiltonian formulations to a discrete-time Bellman equation. They establish the contraction mapping properties of the Bellman operator under bounded-return assumptions and discuss the specific challenges posed by unbounded log-utility (where the value function is unbounded below as capital approaches zero).
Analytic Benchmark: To enable precise error measurement, the authors utilize the closed-form solution for the Cobb–Douglas production function ( $f(k) = Ak^\alpha$ ) and log-utility ( $u(c) = \log c$ ) with full depreciation ( $\delta=1$ ). This provides an exact value function $v^*(k) = c_1 + c_2 \log k$ against which numerical approximations can be measured.
Numerical Experiment:
- Algorithm: Fitted VFI using piecewise-linear interpolation on a discrete grid of capital ( $I=150$ points).
- Parameters: $A=1, \alpha=0.65, \beta=0.95$ .
- Initial Guesses: The study compares two specific initial guesses used in prior literature: $w_A(k) = 5 \log k - 25$ and $w_B(k) = 2k^{0.25} - 35$ .
- Error Metrics: The authors distinguish between interior error (evaluated on $k \ge 0.05$ to avoid boundary discretization issues) and full-grid error. They also track policy-function convergence separately from value-function convergence.
- Comparisons: The study extends the comparison to a broader family of initializations and benchmarks the initialization effect against partial Howard policy iteration.

Key Results

Initialization Impact: The choice of initial guess materially affects the number of iterations required to reach a specific error tolerance. For an interior sup-norm error tolerance of 1.0:
- Guess $w_B$ required 28 Bellman sweeps.
- Guess $w_A$ required 37 Bellman sweeps.
- This represents a 24% reduction in computational effort for the superior guess.
Convergence Rates: Despite the difference in iteration counts, both sequences exhibit the same asymptotic geometric decay rate (approximately 0.9499), which matches the theoretical discount factor $\beta = 0.95$ . The advantage of $w_B$ is attributed entirely to a "starting-distance" or transient effect (it begins closer to the fixed point in the sup-norm), not a different asymptotic convergence rate.
Boundary Discretization Error: The study identifies a significant source of numerical error near the lower boundary ( $k \to 0$ ). Because the exact value function has a logarithmic singularity (unbounded slope) at zero, a uniform grid creates a large interpolation error in the first interval. This error creates a "discretization floor" that prevents the full-grid sup-norm error from vanishing, regardless of the number of iterations.
Policy vs. Value Convergence: The optimal policy function (consumption rule) converges to its numerical floor much faster than the value function. By iteration 20, policy errors are near the interpolation floor, while value-function errors remain distinct. This implies that for simulation purposes, strict value-function tolerances may be unnecessary.
Algorithmic Acceleration: When compared to partial Howard policy iteration, algorithmic acceleration proved to be a more potent factor than initialization. Howard's method reduced the sweep count to 14 (from $w_A$ ) and 6 (from $w_B$ ), outperforming the gains achieved solely by optimizing the initial guess.

Key Contributions

Quantification of Initialization Effects: The paper provides a reproducible, quantitative analysis of how initial value-function guesses affect computational cost, moving beyond visual inspection to exact sup-norm error trajectories.
Separation of Error Sources: It explicitly separates finite-iteration error (convergence to the fixed point) from interpolation error (grid resolution), demonstrating that the latter creates a hard floor for accuracy on uniform grids with log-utility.
Pedagogical Clarity: The paper offers a complete derivation linking the Solow and Ramsey models to the Bellman equation, including the finite-horizon recursion and the closed-form solution, making the numerical experiments auditable.
Reproducibility: The authors provide deterministic Python code that reproduces every numerical table and figure, ensuring the results are verifiable.

Significance and Claims
The paper modestly claims that its primary contribution is not a new algorithm, but a careful numerical analysis of an often-overlooked implementation choice: the initial value-function guess. It argues that while initialization is a useful margin for reducing computational effort, it is not a substitute for better algorithms (like Howard iteration) or better-resolved grids.

The authors emphasize that practitioners should report initialization, stopping criteria, evaluation domains (specifically excluding boundary regions where discretization errors dominate), and grid resolution together, rather than treating VFI convergence as a single scalar metric. The study highlights that for log-utility models, the "transient" distance to the fixed point is the primary driver of iteration counts, while the asymptotic rate remains invariant.

Technical Summary: Dynamic Programming Treatment of the Solow and Ramsey–Cass–Koopmans Growth Models

More like this