📈 economics

Clarke Differentials and the Envelope Theorem in Dynamic Programming

This paper establishes an envelope theorem for deterministic dynamic programming models using Clarke differentials, thereby removing the standard requirements of differentiability, convexity, and boundedness found in previous research.

Original authors: Yuhki Hosoya

Published 2026-06-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuhki Hosoya

Original paper licensed under CC BY 4.0 (http://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

Imagine you are the captain of a ship navigating a complex, winding river. Your goal is to reach a distant port with the most treasure possible. Every day, you have to decide how much fuel to burn today versus saving it for tomorrow. This is a classic "dynamic programming" problem: making a series of decisions where today's choice affects tomorrow's options.

In economics, this is often modeled as a "Value Function." Think of this function as a magic map that tells you the maximum treasure you can expect to find if you start from a specific location.

The Problem: The "Smoothness" Trap

For a long time, economists wanted to know: If I change my starting location just a tiny bit, how does my total treasure change?

To answer this, they used a tool called the Envelope Theorem. In simple terms, this theorem says: "If you are already making the best possible choices, you don't need to worry about how your future choices will change when you move slightly. You only need to look at how the immediate situation changes."

However, there was a catch. To use this theorem, the "magic map" (the value function) and the "rules of the river" (the constraints) had to be perfectly smooth and round, like a polished marble ball. In math, this means they had to be differentiable (having a clear, single slope at every point) and convex (shaped like a bowl, not a jagged mountain).

But real life isn't a polished marble ball.

The "Logarithm" Problem: Many economic models use a utility function that looks like a logarithm (like the CRRA function mentioned in the paper). This function shoots up to infinity or drops to negative infinity at the edges. It's not "bounded" (it doesn't stay within a fixed box), and it's not smooth at the very edge.
The "Jagged" Problem: Sometimes the rules of the game change abruptly. The map might have a sharp corner.

Previous versions of the Envelope Theorem (like the famous Benveniste-Scheinkman theorem) required the map to be smooth and the rules to be perfectly convex. If your economic model had a jagged edge or an unbounded function (like the logarithm), the old math tools broke down, and economists couldn't calculate how their treasure would change.

The Solution: The "Clarke Differential"

This paper introduces a new way to look at the map using a tool called the Clarke Differential.

Think of the Clarke Differential as a "fuzzy compass" or a "safety net" for slopes.

Old Way (Standard Derivative): If you stand on a sharp corner of a mountain, a standard compass spins wildly because there is no single "up" direction. The math says, "I can't calculate this."
New Way (Clarke Differential): The fuzzy compass doesn't look for one single direction. Instead, it looks at all the possible slopes that could exist at that sharp corner and gives you a "range" or a "cloud" of possible directions. It says, "The slope is somewhere between these two values."

By using this fuzzy compass, the author (Yuhki Hosoya) proves a new version of the Envelope Theorem that works even when:

The map has sharp corners (it's not smooth).
The treasure values can go to infinity (it's not bounded).
The shape of the river isn't a perfect bowl (it's not strictly convex).

The Main Result: A More Flexible Map

The paper's main achievement (Theorem 3) is showing that you can still apply the Envelope Theorem to these "messy," real-world economic models.

The author demonstrates this using the Ramsey-Cass-Koopmans (RCK) model, a standard model for capital accumulation (saving and investing).

Old Theorem: Would say, "This model is too messy because the utility function (happiness from consumption) can be infinite or undefined at zero. We can't use the Envelope Theorem here."
New Theorem: Says, "Even though the function is messy, jagged, and unbounded, we can still use our fuzzy compass to determine how the total treasure changes when you start with a little more or a little less capital."

The "Secret Sauce": Regularity

The paper also introduces a concept called Regularity. Imagine a mountain where the path going up is different from the path going down. That's "irregular." The paper shows that for many economic functions (like concave utility functions), the path up and the path down are actually consistent enough that the "fuzzy compass" gives a reliable answer. This allows the theorem to work even without assuming the function is perfectly smooth.

Summary

In everyday language, this paper is like upgrading a navigation system.

Before: The GPS only worked if the roads were perfectly straight, smooth, and within a specific city limit. If you tried to drive off-road or into a canyon, the GPS failed.
Now: The new GPS (using Clarke Differentials) can handle bumpy roads, sharp turns, and infinite distances. It tells the economist, "Even if your economic model is jagged and unbounded, you can still calculate the impact of small changes without getting stuck."

The paper doesn't claim to solve new economic problems or predict the future; it simply provides a more robust mathematical tool that allows economists to analyze models that were previously considered "too messy" to study with the Envelope Theorem.

Technical Summary: Clarke Differentials and the Envelope Theorem in Dynamic Programming

Problem Statement
The paper addresses the application of the envelope theorem to deterministic dynamic programming models. In standard optimization theory, the envelope theorem states that the derivative of the value function with respect to a parameter depends only on the direct effect of the parameter on the objective function, as the indirect effect via the optimal choice vanishes due to the first-order condition. However, in dynamic programming, the value function $\bar{V}(x)$ is derived endogenously, making its differentiability difficult to assume exogenously. Furthermore, the optimal policy function $y^*(x)$ may not be differentiable.

Previous literature has attempted to resolve this. Benveniste and Scheinkman [1] established an envelope theorem using convex analysis, but their results rely on strong assumptions, including the convexity of the domain and the concavity of the objective function. Mordukhovich and Sagara [4] extended these results using Clarke differentials to non-convex settings and stochastic models. However, their approach requires the objective function to be bounded and defined on the entire space. These assumptions are often violated in standard economic models, such as the Ramsey–Cass–Koopmans (RCK) model with Constant Relative Risk Aversion (CRRA) utility functions, which are unbounded and defined on restricted domains (e.g., positive reals).

Methodology
The author employs the Clarke differential, a tool from nonsmooth analysis, to derive an envelope theorem under significantly weaker assumptions. The methodology involves:

Framework: The paper utilizes a reduced-form deterministic dynamic programming model where the value function satisfies the Bellman equation:
$\bar{V}(x) = \max \{ w(x, y) + \delta \bar{V}(y) \mid y \in \Gamma(x) \}$
Here, $X$ is a Banach space, $w$ is the objective function, $\Gamma$ is the constraint correspondence, and $\delta \in (0, 1)$ .
Clarke Differential Tools: The paper defines the generalized directional derivative $f^\circ(x; v)$ and the Clarke differential $\partial^\circ f(x)$ for locally Lipschitz functions. It establishes fundamental properties, including the relationship between the Clarke differential and strong differentiability, and proves that for convex or concave functions, the Clarke differential coincides with the standard subdifferential.
Regularity Conditions: A key concept introduced is the regularity of a function, where the generalized directional derivative equals the classical directional derivative. The paper leverages the fact that concave functions on open convex sets are regular.
Proof Strategy: The proof of the main theorem (Theorem 3) proceeds by:
- Establishing that the value function $\bar{V}$ is locally Lipschitz around a point $\bar{x}$ without assuming global boundedness of $w$ .
- Utilizing the upper hemi-continuity of the policy correspondence $G(x)$ and the compactness of the constraint set to construct sequences converging to the optimal path.
- Applying the definition of the Clarke differential to the Bellman equation to bound the generalized derivative of $\bar{V}$ by the partial Clarke differential of $w$ with respect to the state variable.

Key Contributions and Results
The primary contribution is Theorem 3, which establishes the envelope theorem for deterministic dynamic programming models under conditions that are satisfied by many standard economic models but excluded by previous theorems.

Relaxed Assumptions: Unlike Mordukhovich and Sagara [4], Theorem 3 does not require the objective function $w$ to be bounded or defined on the entire space $X^2$ . Unlike Benveniste and Scheinkman [1], it does not require the domain to be convex or the objective function to be concave (though concavity implies local Lipschitz continuity and regularity, which are sufficient).
Main Result: Under the assumptions that $w$ is locally Lipschitz, regular at the optimal points, and that the policy correspondence satisfies specific continuity and non-empty intersection properties, the value function $\bar{V}$ is locally Lipschitz, and its Clarke differential satisfies:
$\partial^\circ \bar{V}(\bar{x}) \subset \partial^\circ_x w(\bar{x}, \bar{y})$
for some $\bar{y} \in \text{cl} G(\bar{x})$ .
Corollary 2: If the objective function is strongly differentiable, the value function is also strongly differentiable, and the inclusion becomes an equality involving the strong derivative.

Significance and Scope
The paper claims that its results are significant because they apply to a wider range of economic models than previously possible. Specifically, the author demonstrates that the theorem applies to the RCK model with CRRA utility functions (including the logarithmic case) and Cobb-Douglas or AK production technologies. These models feature unbounded utility functions and domains restricted to positive reals, which violate the boundedness and full-space assumptions of Mordukhovich–Sagara and the convexity requirements of Benveniste–Scheinkman.

The author notes that while Theorem 1 (Benveniste–Scheinkman) applies to standard concave models, Theorem 3 is necessary for models where the constraint set or objective function lacks convexity (e.g., models with government expenditure rules that break concavity) or where the objective function is unbounded. The paper concludes that Theorem 3 provides a robust tool for deriving the envelope theorem in "natural economic situations" where differentiability and convexity cannot be guaranteed.

The Problem: The "Smoothness" Trap

The Solution: The "Clarke Differential"

The Main Result: A More Flexible Map

The "Secret Sauce": Regularity

Summary

More like this