🔢 mathematics

Moral Hazard in LTI Dynamics: A Hypothesis Testing Approach

This paper proposes a hypothesis testing-based payment scheme to incentivize an agent to select a more efficient linear state-feedback controller in a moral hazard setting, where the agent bears control costs and is risk-averse, and demonstrates its effectiveness through applications in power system load frequency control and body weight loss interventions.

Original authors: Jaewon Jeong, Pan-Yang Su, S. Shankar Sastry, Anil Aswani

Published 2026-05-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jaewon Jeong, Pan-Yang Su, S. Shankar Sastry, Anil Aswani

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 a boss (the Principal) who needs a machine to run perfectly. But you can't see inside the machine, and you can't watch the person (the Agent) who is supposed to be controlling it. You know the machine has two settings: a "lazy" setting that is easy for the agent but makes the machine run poorly, and a "hard-working" setting that is tiring for the agent but makes the machine run great.

The problem is Moral Hazard: The agent might choose the lazy setting because it's easier for them, and you won't know the difference immediately because the machine's output is noisy (like a car engine that sputters sometimes even when it's running well).

This paper asks: How do you pay the agent so they choose the hard-working setting, even though you can't see them doing it?

The Core Idea: The "Coin Flip" vs. The "Pattern"

If you just paid the agent a flat fee, they would pick the lazy setting. If you paid them only when the machine looked perfect, they might get lucky with the lazy setting and still get paid.

The authors propose a clever solution based on detecting patterns over time.

Think of it like a detective trying to figure out if a coin is fair or weighted.

The Lazy Setting (Low Effort): The machine's behavior looks like a fair coin flip. It's random and messy.
The Hard-Working Setting (High Effort): The machine's behavior looks like a weighted coin. It still has some randomness, but there's a subtle, consistent pattern that leans one way.

If you only look at one flip (one second of data), you can't tell the difference. But if you look at 100 flips (a longer time period), the pattern of the "weighted coin" becomes obvious.

The Solution: The "Likelihood Ratio Test"

The paper proves that the best way to design a payment contract is to use a Likelihood Ratio Test.

In plain English, this means:

Wait and Watch: You don't pay the agent immediately. You let the system run for a specific amount of time (let's call this the "Horizon").
Do the Math: At the end of that time, you look at the history of the machine's output. You run a statistical test that asks: "Is this history more likely to have come from the 'Hard-Working' setting or the 'Lazy' setting?"
The Verdict:
- If the history looks like the Hard-Working pattern, you pay the agent a big bonus.
- If the history looks like the Lazy pattern, you pay them a small amount (or nothing).

The Tricky Part: How Long to Wait?

The paper discovers a very specific "Goldilocks" zone for how long you should wait before paying.

Waiting too short: The noise (randomness) is too strong. You can't tell if the agent is working hard or just got lucky. To convince the agent to work hard, you'd have to offer a massive, risky bonus just in case they were working hard. This is expensive for you.
Waiting too long: You can tell the difference perfectly, but you have to wait so long that the money you pay today is worth less tomorrow (because of "discounting"). Also, the agent gets impatient.
The Sweet Spot: There is a perfect middle ground (e.g., 1.8 seconds in one example, 80 weeks in another) where the pattern is clear enough to be fair, but you haven't waited so long that the payment loses its value.

Real-World Examples from the Paper

The authors tested this idea on two very different scenarios:

Power Grids (Load Frequency Control):
- The Boss: A power company.
- The Agent: A power generator.
- The Problem: The generator can choose to use a cheap, slow controller or an expensive, fast controller. The power company can't see which one is being used, only the frequency of the electricity.
- The Result: The power company should wait about 1.8 seconds of data, run the math, and then pay a bonus of about $31.65 if the data proves the generator was using the fast controller.
Weight Loss Programs (Wellness):
- The Boss: A health insurance company.
- The Agent: A person trying to lose weight.
- The Problem: The person can choose to exercise a lot (High Effort) or a little (Low Effort). The insurance company only sees the person's weight, which fluctuates due to water retention, food, etc.
- The Result: The insurance company should wait 80 weeks of data. If the weight loss pattern proves the person was exercising, the company pays a bonus of about $217.

The Bottom Line

You don't need to spy on your agent to get them to work hard. You just need to:

Wait long enough to see the difference between "working hard" and "getting lucky."
Pay based on a mathematical test that compares the data against what "working hard" looks like.
Stop waiting before the delay makes the payment too expensive.

The paper provides the exact math to find that perfect waiting time and the perfect payment amount, ensuring the agent is motivated to do the right thing without the boss needing to be a mind-reader.

1. Problem Statement

The paper addresses the incentive design problem in stochastic control systems under information asymmetry (specifically, moral hazard).

Context: A "Principal" (e.g., a power grid operator or insurance company) wishes to control a dynamical system but must rely on an "Agent" (e.g., a generator or an individual) to execute control actions.
The Asymmetry: The Principal cannot directly observe the Agent's effort level (the specific controller chosen). The Principal only observes noisy system outputs.
The Setup:
- The system is a discrete-time Linear Time-Invariant (LTI) system with process and measurement noise.
- The Agent chooses between two linear state-feedback controllers: a Low-Effort controller ( $K_0$ ) and a High-Effort controller ( $K_1$ ).
- Cost Structure: The High-Effort controller incurs a higher control cost for the Agent but results in a lower state cost for the Principal. The Agent is risk-averse regarding payments.
Objective: Design a payment contract $\pi$ based on observed output trajectories $Y_T = (y_1, \dots, y_T)$ that incentivizes the Agent to choose the High-Effort controller ( $K_1$ ) while minimizing the Principal's total expected cost (state cost + discounted payment).

2. Methodology

The authors formulate the problem as a constrained optimization and solve it using statistical hypothesis testing theory.

A. Contract Structure

The contract is restricted to a simple, interpretable form:

A fixed time horizon $T$ .
Two payment levels: $\pi_0$ (paid if the trajectory looks like Low Effort) and $\pi_1$ (paid if the trajectory looks like High Effort), with $\pi_1 \geq \pi_0$ .
A decision region $R$ in the observation space. If $Y_T \in R$ , the Agent receives $\pi_1$ ; otherwise, $\pi_0$ .

B. Optimization Formulation

The Principal minimizes their expected total cost:
$\min_{T, R, \pi_0, \pi_1} \left( J_P^1 + \gamma_p^T \cdot \mathbb{E}[\pi] \right)$
Subject to:

Incentive Compatibility (IC): The Agent's expected utility from choosing High Effort must exceed that of Low Effort.
Participation Constraint (PC): The Agent's expected utility from accepting the contract (and exerting High Effort) must be non-negative (or better than the outside option).
Limited Liability (Optional): A constraint $\pi_0 \geq 0$ ensuring the Agent is never fined.

C. Theoretical Characterization

The core theoretical contribution is proving that the optimal decision region $R^*$ is defined by a Likelihood Ratio Test (LRT).

Let $H_0$ and $H_1$ be the probability distributions of the output trajectory $Y_T$ under Low and High effort, respectively.
The optimal region is: $R^* = \{ Y_T : L_1(Y_T) \geq \kappa \cdot L_0(Y_T) \}$ , where $L_i$ is the likelihood under hypothesis $i$ .
Distribution Derivation: Since the system is LTI with Gaussian noise, the output vector $Y_T$ is Gaussian. The authors derive the distribution of the Log-Likelihood Ratio (LLR) statistic. They prove that under both hypotheses, the LLR follows a generalized chi-squared distribution (a linear combination of independent non-central chi-squared variables and a standard normal variable).

D. Computational Algorithm

The authors propose a numerical algorithm to compute the optimal contract:

Outer Loop: Iterate over possible time horizons $T$ .
Inner Loop: For a fixed $T$ , perform a line search over the LRT threshold $\eta$ (where $R = \{ \text{LLR} \geq \eta \}$ ).
Probability Calculation: Use Imhof's method (via the chi2comb Python package) to compute the probabilities $\alpha = P(Y_T \in R | H_0)$ and $\beta = P(Y_T \in R | H_1)$ based on the derived generalized chi-squared distributions.
Optimization: Solve for optimal payments $\pi_0, \pi_1$ using the IC and PC constraints, then select the $T$ and $\eta$ that minimize the Principal's cost.

3. Key Contributions

Optimality of LRT: The paper proves that for stochastic LTI systems with continuous outputs and two effort levels, the optimal contract structure relies on a Likelihood Ratio Test. This extends previous results known for discrete outputs to continuous dynamics.
Statistical Characterization: The authors explicitly derive the distribution of the log-likelihood ratio for LTI systems, showing it takes the form of a generalized sum of chi-squared distributions. This allows for precise calculation of Type I and Type II error rates ( $\alpha$ and $\beta$ ) without simulation.
Trade-off Analysis: The work identifies a fundamental trade-off in contract design:
- Longer Horizons ( $T$ ): Reduce uncertainty (noise) relative to the signal difference between controllers, allowing for smaller payments.
- Shorter Horizons: Reduce the discounting cost of the payment (since the Principal pays sooner).
- The optimal $T$ balances these opposing forces, resulting in a U-shaped cost curve.

4. Results and Applications

The methodology was validated using two distinct examples:

A. Load Frequency Control (LFC) in Power Systems

Scenario: An Independent System Operator (Principal) incentivizes a power generator (Agent) to use a high-performance controller for frequency regulation.
Parameters: Discrete-time system with 4 states, 30-second operating cycle.
Result: The optimal contract length was found to be 1.8 seconds ( $T=18$ $T = 18$ steps).
- At this horizon, the Principal's expected cost is minimized to $30.09.
- The required incentive payment ( $\pi_1$ ) is $31.65.
- Shorter horizons required significantly higher payments due to noise dominance; longer horizons increased costs due to discounting.

B. Wellness Incentive Program (Weight Loss)

Scenario: A health insurance company (Principal) incentivizes an individual (Agent) to increase physical activity (steps) to lose weight.
Parameters: Affine LTI model for weight dynamics over 200 weeks.
Result: The optimal contract length was 80 weeks.
- The minimized cost for the insurer was $121.58.
- The optimal high-effort payment was $217.02.
- Similar to the LFC case, the cost curve was U-shaped, demonstrating the robustness of the trade-off across different physical domains.

5. Significance

Bridging Economics and Control: The paper successfully integrates Principal-Agent theory (economics) with Stochastic Control (engineering), providing a rigorous framework for designing incentives in cyber-physical systems.
Practical Implementability: By reducing the infinite-dimensional optimization problem to a tractable 2D search (over $T$ and $\eta$ ) with closed-form statistical distributions, the approach is computationally feasible for real-world applications.
Handling Continuous Data: Unlike many economic models that assume discrete outcomes, this work handles continuous, noisy time-series data typical in engineering systems, proving that LRTs remain optimal even in these complex settings.
Risk-Aversion: The model explicitly accounts for the Agent's risk aversion, showing how the shape of the utility function influences the optimal payment structure.

In conclusion, the paper establishes that hypothesis testing is the mathematically optimal mechanism for resolving moral hazard in linear stochastic control systems, providing a clear prescription for how long to observe a system and how much to pay to ensure desired behavior.