Robust Stability of Suboptimal Moving Horizon Estimation

20 concepts6 papers1 proposal

Moving Horizon Estimation (MHE) is a powerful framework for reconstructing the hidden states of a dynamic system from noisy physical measurements. By optimizing over a sliding window of recent data, MHE naturally incorporates physical constraints (such as non-negative concentrations or bounded velocities) and handles complex nonlinear dynamics.

However, MHE faces a critical bottleneck in real-time applications: it requires solving a constrained optimization problem at every single time step. For high-speed systems, the time budget allotted to the optimization solver might be mere milliseconds. If the solver fails to converge to the global optimum within this window, we are left with a suboptimal solution.

Historically, state estimation theory only guaranteed stability under the assumption of perfect, global optimality. If a solver was cut off early, the resulting estimate could cause the observer to diverge. To bridge this gap, recent research has established a "feasibility-implies-stability" paradigm, proving that MHE can remain robustly stable even when the underlying optimization is terminated prematurely.

The Arrival Cost and the Challenge of Suboptimality

To understand why suboptimality is dangerous, we must look at how MHE summarizes the past. An MHE window only looks at the last $M$ time steps. To ensure that information from before this window is not lost, MHE uses an arrival cost, which acts as a proxy for all historical data.

In their foundational work on constrained linear state estimation, Rao et al. demonstrated that the stability of an MHE framework depends heavily on how closely this approximate arrival cost tracks the true, full-information estimation problem (Constrained linear state estimation - a moving horizon approach). The authors showed that an exact analytical form for the constrained arrival cost is not strictly necessary; instead, asymptotic stability can be guaranteed if the approximated arrival cost satisfies a specific inequality condition relative to the true system dynamics.

When extending this to nonlinear systems or strict real-time constraints, maintaining this inequality via exact optimization becomes computationally intractable. If a real-time solver terminates early, the resulting suboptimal estimate may violate the theoretical bounds required to guarantee that the estimation error shrinks over time.

The Solution: Auxiliary Observers as a Safety Net

How do we guarantee stability when we cannot trust our optimization solver to find the absolute minimum? The core insight of modern suboptimal MHE is to anchor the optimization problem to a computationally cheap, pre-designed auxiliary observer that is already known to be stable (such as a standard nonlinear observer or an extended Kalman filter).

As proposed by Schiller et al., this auxiliary observer is used to generate a "warm start" candidate solution for the MHE solver at each step (Robust Stability of Suboptimal Moving Horizon Estimation using an Observer-Based Candidate Solution).

[System Measurements] ---> [Auxiliary Observer] ---> [Stable Candidate Solution]
                                                            |
                                                            v
[System Measurements] ---> [ Real-Time Solver ] <--- [Cost Comparison]
                                    |
                                    v
                       [Stable Suboptimal Estimate]

The mechanism works via a simple cost-comparison rule:

  1. The auxiliary observer computes a state trajectory estimate.
  2. This trajectory is evaluated using the MHE cost function to establish an upper bound on acceptable cost.
  3. The MHE solver initializes its search from this candidate solution.

Because the solver starts at a stable candidate, any iteration of the optimization algorithm that decreases—or simply maintains—the cost function will yield an estimate that is at least as accurate as the auxiliary observer. As a result, Müller et al. proved that the MHE framework inherits the robust stability properties of the auxiliary observer, completely independent of whether the solver reaches global convergence (Suboptimal Nonlinear Moving Horizon Estimation). This allows for early termination of the optimization problem under strict computational constraints without sacrificing safety.

Quantifying Stability via Time-Discounted $\delta$-IOSS

To mathematically formalize this inheritance of stability, researchers look at how errors propagate through the system. A key property used here is incremental Input/Output-to-State Stability ($\delta$-IOSS), which dictates that the difference between any two state trajectories can be bounded by the differences in their initial states, inputs, and outputs.

To make these bounds practical for real-time solvers, Knüfer et al. introduced a refined framework incorporating time-discounting and non-additive disturbances (Time-Discounted Incremental Input/Output-to-State Stability). Time-discounting places higher weight on recent measurements, allowing the estimator to discard the influence of old disturbances more rapidly.

By leveraging this time-discounted $\delta$-IOSS property, Schiller et al. constructed a direct, multi-step Lyapunov function for the MHE algorithm (A Lyapunov Function for Robust Stability of Moving Horizon Estimation). Rather than relying on complex contraction mappings or assuming infinitely long estimation horizons, the authors showed that the system's underlying $\delta$-IOSS Lyapunov function can be mapped directly to the MHE cost function. This mathematical bridge guarantees robust exponential stability using significantly shorter estimation horizons than previously required, reducing the computational load per optimization step.

Algorithm-Agnostic Guarantees

The ultimate practical benefit of this framework is that the stability guarantee becomes entirely decoupled from the choice of optimization algorithm.

In a simplified suboptimal MHE scheme developed by Schiller et al., the optimization is performed directly over the trajectory space of the pre-designed auxiliary observer (A Simple Suboptimal Moving Horizon Estimation Scheme With Guaranteed Robust Stability). Because the observer's Lyapunov function serves directly as the multi-step Lyapunov function for the MHE, the stability of the estimator is guaranteed for any horizon length $M$ that satisfies verifiable system conditions.

Whether the real-time system executes one iteration of a gradient descent solver or ten iterations of a sequential quadratic programming (SQP) solver, the estimation error is mathematically guaranteed to remain bounded. The optimization solver is transformed from a critical single point of failure into an engine that purely optimizes performance: more iterations yield cleaner estimates, but fewer iterations never compromise stability.

Key Foundational Work

Constrained linear state estimation - a moving horizon approach (2001) — contributed 4 concept elements to this direction.
Robust Stability of Suboptimal Moving Horizon Estimation using an Observer-Based Candidate Solution (2020) — contributed 1 concept elements to this direction.
Time-Discounted Incremental Input/Output-to-State Stability (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_real_time_stability_certification_9dffac5d,
  title = {real-time stability certification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/real-time-stability-certification/rdr_9dffac5df44248fc87d2968f3b6a39f8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_real_time_stability_certification_9dffac5d_v1,
  title = {real-time stability certification},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/real-time-stability-certification/rdr_9dffac5df44248fc87d2968f3b6a39f8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake