Comparing sliding-mode, bang-bang and linear-quadratic-Gaussian for steering an atomic clock

This paper demonstrates through numerical simulations across multiple timescales that first-order SMC achieves slightly better time-domain accuracy than LQG while maintaining nearly identical stability, and substantially outperforms bang-bang (BB) steering in accuracy while avoiding BB's short-term instability.

The Gist

Imagine you are trying to keep a very sensitive, expensive grandfather clock perfectly in sync with a master clock located in a government observatory. The problem is that your clock is a bit "jittery"—it naturally drifts forward or backward due to tiny, random vibrations (noise). To fix this, you need a "steering wheel" that constantly nudges your clock back on track.

This paper compares three different "drivers" (control strategies) to see which one does the best job of keeping your clock accurate over time without making it shake too much.

Here is the breakdown of the three drivers and the race they ran:

The Three Drivers

1. The "Bang-Bang" Driver (BB):

   • How it works: This is the simplest approach. Imagine a driver who only looks at whether the clock is fast or slow. If it's even a tiny bit fast, they slam the brakes. If it's slow, they floor the gas. They only do two things: full speed or full stop.
   • The Problem: Because they are so aggressive, they overshoot constantly. It's like driving a car by only turning the steering wheel all the way left or all the way right. You eventually get to the destination, but the ride is bumpy, and the car swerves wildly in the short term.

2. The "Linear-Quadratic-Gaussian" Driver (LQG):

   • How it works: This is the "smart" driver. They use a complex mathematical formula (a computer brain) to calculate the perfect amount of gas or brake needed at every single moment. They weigh the cost of being wrong against the cost of making a big correction.
   • The Reputation: This has been the gold standard for years. It provides a very smooth, gentle ride.

3. The "Sliding-Mode" Driver (SMC):

   • How it works: This is the new challenger. It's a bit like a driver who keeps the car on a specific "rail" or path. If the car drifts off the rail, the driver makes a sharp correction to snap it back, but once it's back on the rail, they let it glide smoothly. It combines the simplicity of the "Bang-Bang" driver with the smoothness of the "Smart" driver.
   • The Goal: The authors wanted to see if this driver could be as smooth as the LQG driver but easier to build.

The Race (The Experiment)

The authors didn't just guess; they ran a massive simulation.

• The Track: They simulated a clock running for different lengths of time: one week, one month, one year, and even ten years.
• The Weather: They added "noise" (random jitters) to the clock to make it realistic.
• The Test: They ran the simulation 100 times with different random noise patterns to make sure the results weren't just a lucky fluke.

The Results

Here is what happened when they compared the drivers:

• Accuracy (How close is the time?):
  In this simulated benchmark, the Sliding-Mode (SMC) driver performed well, keeping the clock time closer to the master clock than the "Smart" (LQG) driver across all time periods (from one week to ten years). Both of these were much better than the "Bang-Bang" driver, which was often way off.

• Stability (How smooth is the ride?):

  • The Bang-Bang driver showed a distinct short-term instability — the clock wobbled in the short term before settling.
  • The LQG driver was very smooth.
  • The Sliding-Mode (SMC) driver was almost identical to the LQG driver in terms of smoothness. It did not have the jerky, swerving problems of the Bang-Bang driver.

The Bottom Line

The paper concludes that the Sliding-Mode (SMC) driver performed well in this simulated benchmark.

• It matched or beat the complex, math-heavy LQG driver in terms of accuracy within this test environment.
• It was much smoother than the simple, aggressive Bang-Bang driver.

The authors suggest that because SMC is simple to program (it doesn't need the heavy math machinery of LQG) and performed well in the simulation, it may be a promising candidate for experimental testing. However, the paper does not claim that SMC has already replaced existing methods or is proven as a real-world solution. Real-world validation and experimental testing remain future work to determine if these simulated results hold up when steering actual atomic clocks.

Technical Summary

Technical Summary: Comparing Sliding-Mode, Bang–Bang, and Linear-Quadratic-Gaussian for Steering an Atomic Clock

Problem Statement
Accurate timekeeping requires feedback loops that steer local atomic clocks toward a higher-grade reference to maintain time errors within nanoseconds over extended periods. While established methods exist, they present trade-offs: Linear-Quadratic-Gaussian (LQG) control offers optimal linear feedback but relies on complex model-based machinery, whereas Bang–Bang (BB) control (used in GPS) is simple but introduces characteristic short-term instability due to aggressive switching. Sliding-mode control (SMC) is proposed as a potential middle ground, offering the simplicity of BB with the robustness of LQG. However, published applications of SMC to atomic clocks have been limited to short-term acquisition stages rather than continuous long-term steering. This study addresses the gap by evaluating first-order SMC as a primary long-term steering policy against LQG and BB under identical stochastic conditions.

Methodology
The authors employ a standard two-state clock model driven by white-frequency noise (WFN) and random-walk-frequency noise (RWFN). The system is simulated using a discrete-time stochastic dynamics approach where the clock state vector consists of time offset ($X_1$) and fractional frequency error ($X_2$).

Three steering policies are implemented and compared:

1. LQG: Uses a linear state-feedback law derived from solving the discrete-time algebraic Riccati equation (DARE), minimizing a quadratic cost function weighted by state and control effort.
2. BB: A GPS-style policy that switches control based on the sign of a sliding surface defined by the time offset and frequency error.
3. SMC: A first-order sliding-mode controller using a linear sliding surface ($S = X_2 + \lambda X_1$) and a discontinuous feedback law to drive the system to the surface in finite time.

The evaluation utilizes two complementary metrics:

• Accuracy: Quantified by the standard deviation of the time-offset ($\sigma_{X_1}$). To ensure statistical robustness, the authors performed Monte Carlo simulations over 100 independent random seeds across four distinct time periods: one week, one month, one year, and ten years.
• Stability: Quantified by the overlapping Allan deviation (oADEV) over a range of averaging times ($10^5$ s to $10^8$ s), using a single long-duration simulation ($10^6$ days) to ensure clean spectral estimates.

All simulations use noise parameters and weight matrices consistent with established reference data (FGTB10) to ensure comparability.

Key Results

• Accuracy: Across all four time periods (week to decade), the SMC policy consistently achieved the lowest mean time-offset standard deviation in the simulations, performing comparably to or better than LQG. Both SMC and LQG substantially reduced error magnitudes relative to the BB policy, which exhibited significantly higher errors. The unsteered free-running clock (FRC) performed worst, with error growing dramatically over time.
• Stability: The oADEV analysis revealed that SMC and LQG exhibit nearly identical stability profiles across the full averaging-time range studied. Both show modest short-term instability that converges to the random-walk floor of the reference clock at longer timescales. In contrast, the BB policy displays a distinct "bump" in instability at short averaging times ($\tau \approx 2 \times 10^5$ s to $6 \times 10^5$ s), a characteristic artifact of its switching nature, before converging to the same long-term slope.
• Trajectory Behavior: Time-offset trajectories show that while BB acts like a relay with abrupt corrections, SMC and LQG follow the reference clock more smoothly from the onset.

Significance and Claims
The paper suggests that first-order SMC is a promising alternative for atomic-clock steering in the simulated benchmark, combining the accuracy benefits of LQG with the implementation simplicity of BB, without inheriting BB's short-term instability.

The authors emphasize that SMC achieves this performance without requiring the detailed model-based machinery associated with LQG. The control law relies on a simple sign-based nonlinear correction with only two tunable parameters ($\lambda$ and $K_{SMC}$). Consequently, the authors suggest that SMC could be implemented as a software-level modification to existing systems that already estimate time and frequency offsets, rather than requiring a hardware redesign.

The study concludes that while the simulated results indicate SMC is a promising candidate for experimental testing, the practical advantage must be validated in real-world systems accounting for measurement delays, command limits, and imperfect state estimation. The paper does not claim SMC should immediately replace existing policies but positions it as a robust, low-cost alternative worthy of operational investigation.