Why Rational Controllers Beat Finite Lookahead When Disturbances Are Adversarial and the State Is Hidden

A distributionally robust controller is designed to work well against every disturbance distribution within some distance of a nominal one. The worst disturbances in that set are not white noise. They concentrate their energy at the frequencies where the plant is most fragile. Over an infinite horizon, the optimal controller against them turns out to be non-rational: its transfer function contains square roots of polynomial ratios. Such a controller has no finite state-space realization, so embedded hardware cannot run it exactly. Engineers have two ways to approximate it. They can truncate it in time, using finite memory, a finite lookahead, or a rolling finite-horizon optimization. Or they can fit a low-order rational filter to it.

This note argues that the choice between the two is predictable from one number we can compute. That number is how close the exact controller's singularities sit to the unit circle. As ambiguity grows, the worst-case disturbance spectrum sharpens into resonant peaks and pushes those singularities toward the circle. The benefit of each extra lookahead or memory step then decays more and more slowly, while a rational filter needs only one pole pair per resonance. We develop and test this claim for the setting the existing infinite-horizon theory avoids: output feedback, where the controller sees only noisy measurements rather than the full state.

Core Claim

Write $K_\star(z)$ for the exact infinite-horizon distributionally robust output-feedback controller. Let $\rho(r) < 1$ be the largest modulus of its singularities inside the unit disk (poles or square-root branch points), at ambiguity radius $r$. By the Cauchy–Hadamard relation, its impulse-response coefficients $h_n$ satisfy $|h_n| \lesssim \rho(r)^n$. Truncating the controller to $k$ steps of memory or lookahead therefore leaves a cost gap of order

$$\Delta J_{\text{trunc}}(k) \approx C\,\rho(r)^{2k},$$

because the extra quadratic cost scales with the squared tail energy of the discarded coefficients. This has the same form as the $\mathcal{O}(\lambda^{2k}T+1)$ regret bound in The Power of Predictions in Online Control, which proves exponential decay of the value of extra prediction steps for fully observed LQR. There, $\lambda$ is a closed-loop stability margin fixed by the plant.

Our claim is that, in the robust output-feedback setting, the governing rate is $\rho(r)$. That rate depends on three things: the adversary's worst-case spectrum, the plant, and the observer's spectral factor. It tends toward 1 as $r$ grows. The horizon needed for a target suboptimality $\epsilon$, roughly $k^\star \approx \log(1/\epsilon)/\log(1/\rho(r))$, should therefore blow up with $r$.

By contrast, the rational order needed should stay near $2N_{\text{res}} + n_{\text{obs}}$. Here $N_{\text{res}}$ is the number of separated near-circle singularity clusters, each of which a conjugate pole pair can absorb, and $n_{\text{obs}}$ accounts for the observer dynamics. The order should be nearly independent of how close those clusters are to the circle.

The full-state evidence already fits this picture. On the AC15 aircraft model, Distributionally Robust Kalman Filtering over an Infinite-Horizon reports results for the infinite-horizon full-state controller (labelled DR-LQR there):

  • At $r=1.5$, a first-order rational approximation costs 5789.1 against 153.2 for the exact controller.
  • At $r=10$, the first-order approximation costs 33262 against 3635.6.
  • A second-order fit stays within 1% at every radius.
  • A moving-window finite-horizon solver has a higher average cost than the infinite-horizon rational controller in 210-step simulations.

A single resonance needs a pole pair, and time truncation suffers when $\rho$ is near 1. What is new here is turning that pattern into a quantitative, falsifiable rule and testing it under partial observability.

Positioning

Three groups of prior work bear on this.

Methods we reuse. Two papers solve the full-state infinite-horizon problems in the frequency domain and fit rational approximations by minimizing $\|K_\star - \hat K\|_\infty$. The Distributionally Robust Infinite-Horizon LQR uses a fixed-point spectral factorization. Infinite-Horizon Distributionally Robust Regret-Optimal Control uses Frank–Wolfe. Both assume full state access, and the first names output feedback as open. Optimal Infinite-Horizon Mixed H2/ H∞ Control shows that the non-rational controller is fixed by a finite-dimensional anti-causal parameter. It also reports that a 4th-order fit is nearly indistinguishable from the optimum on the 8th-order REA4 reactor, so order tracks spectral structure rather than plant size. We reuse these techniques.

Work that motivates the gap. Wasserstein Distributionally Robust Regret-Optimal Control under Partial Observability handles measurement feedback, but only through finite-horizon semidefinite programs. These are optimization problems over matrix inequalities whose cost grows roughly cubically in the horizon. It does not address the infinite-horizon limit or realization order. Distributed Truncated Predictive Control for Networked Systems under Uncertainty: Stability and Near-Optimality Guarantee proves exponential decay with horizon and spatial truncation, with forecast-error sensitivity decaying in lead time. It assumes locally observed states and no ambiguity set. Our contribution is to identify the decay rate in the regime that work excludes.

Work that constrains the design. Wasserstein Distributionally Robust Control of Partially Observable Linear Stochastic Systems and Wasserstein Distributionally Robust Control and State Estimation for Partially Observable Linear Systems recover a separation principle and rational Riccati controllers. They do this by limiting ambiguity to means and covariances through the Gelbrich bound. That is a strong, tractable competitor. If it captures resonant worst cases, the non-rational machinery is unnecessary, and our experiments are built to detect that outcome.

Research Program

Exact output-feedback controller. We first compute $K_\star$ for output feedback in two independent ways.

  • The primary route extends the full-state frequency-domain fixed point. It adds a measurement channel with a known Gaussian noise spectrum, keeping the known-measurement-noise assumption of the partially observable literature, and derives the output-feedback optimality conditions by the same Wiener–Hopf argument.
  • The convergence proof for multivariable disturbances is open even in the full-state case. We therefore cross-validate with a second route: solve the finite-horizon measurement-feedback program of the regret-optimal partial-observability paper at the largest tractable horizon, and read off the stationary middle rows of the resulting block-Toeplitz controller. Where the two routes disagree beyond discretization error, we report it rather than paper over it.

Singularity analysis. From $K_\star$ sampled on a fine frequency grid, we estimate $\rho(r)$ in two ways: from the geometric decay of its Fourier coefficients, and by locating branch points of the spectral factors through analytic continuation of the fixed-point iterates. The two estimates check each other. We also compute the Hankel singular values of $K_\star$; their decay is the standard predictor of rational approximation error. We expect them to show a sharp drop after $2N_{\text{res}} + n_{\text{obs}}$ values.

Two families of finite controllers. At matched online cost, we build:

  • time-truncated controllers: length-$k$ finite-impulse-response truncations of $K_\star$, and rolling $k$-step robust optimizations;
  • rational controllers of order $m$, by the convex $H_\infty$ feasibility fit.

We trace cost against $k$ and against $m$ for each $r$ and check whether the measured slopes match $\log \rho(r)$ and the predicted order.

The separation question. We test whether partial observability forces the joint controller to need more order than its parts. The comparison is between a joint rational fit of $K_\star$ and a cascade of separately fitted rational robust Kalman filter and robust state-feedback controllers.

Baselines

Naive reference: standard LQG. This is a Kalman filter plus LQR designed for the nominal Gaussian model. It uses the same plant, the same measurements, and no tuning beyond the nominal noise covariances. It shows how much the robust problem matters at all. Wasserstein Distributionally Robust Control of Partially Observable Linear Stochastic Systems reports, on the IEEE 39-bus grid, an average total cost of 2735.0 for LQG against 1842.6 for its Gelbrich-based controller.

Strongest existing tractable competitor: the Gelbrich-bound robust output-feedback controller. We use its infinite-horizon Riccati form from the same paper. It is the strongest competitor because it is already rational, already separated, and runs online at LQG speed. It is run with the same nominal model and the same radius $r$, with $r$ mapped consistently between the moment-based and full Wasserstein sets. It also gets the same out-of-sample disturbance draws.

Beating it means two things. First, lower worst-case expected cost under the true Wasserstein worst case computed for $K_\star$. Second, lower empirical cost under resonant, time-correlated non-Gaussian disturbances at radii where $\rho(r)$ exceeds 0.9. Both must hold at a rational order no larger than the Gelbrich controller's order plus two. On the IEEE 39-bus system its reported cost is 1842.6. For output feedback on AC15, AC17, REA4 or the Boeing 747 model, there is no reported Gelbrich-controller result on Lacuna, so we compute it ourselves.

Strongest time-truncated competitor: rolling finite-horizon robust measurement-feedback optimization. This follows Wasserstein Distributionally Robust Regret-Optimal Control under Partial Observability, with horizon $k$ swept up to its tractability limit. Its per-step online compute budget is matched against the rational controller's. On the Boeing 747 model that paper reports only qualitative interpolation between LQG and pure regret-optimal control; no numeric regret entries for it are on Lacuna.

Ablations. Two ablations isolate the proposal's key ingredients:

  • The finite-impulse-response truncation of $K_\star$ at length $k$ isolates time truncation from the approximation method.
  • The separated cascade of rational robust filter and robust controller isolates whether jointly approximating the output-feedback law matters.

For the full-state reference, Distributionally Robust Kalman Filtering over an Infinite-Horizon reports that a second-order fit is within 1% of exact on AC15. Its exact cost at $r=10$ is 3635.6, against 3645.8 for the second-order fit.

Why not just…?

Why not just use the Gelbrich-bound controller and keep classical separation? It is the most serious threat. The argument against it is structural: a mean-and-covariance ambiguity set cannot express where in frequency the adversary puts its energy. The full-state infinite-horizon papers show that the true worst case forms sharp resonant peaks that the optimal law must shape around. That argument has not yet been checked in the output-feedback regime. If the Gelbrich controller comes within a few percent of $K_\star$ across radii and plants, the conclusion is that non-rationality is practically irrelevant under partial observability. We will report that as a negative result. The experiment settles it directly by evaluating both controllers against the spectrally resonant worst case.

Why not just separate: approximate the robust Kalman filter and the robust state-feedback law independently and cascade them? Each half already has an infinite-horizon frequency-domain treatment, so the cascade is cheap to build. But no separation principle is known for the exact non-rational output-feedback problem; separation is recovered only under the Gelbrich proxy. The cascade ablation measures the cost of assuming it. If the cascade matches the joint fit at equal total order, the joint derivation is unnecessary for implementation and only the singularity-based order rule survives as a contribution.

Why not just extend the prediction horizon of a rolling robust MPC until it is good enough? The core claim predicts this fails exactly where robustness matters. When the adversary's peaks push $\rho(r)$ toward 1, the required horizon grows like $1/\log(1/\rho)$. Meanwhile the finite-horizon programs grow roughly cubically in horizon and are reported to become unstable beyond about 100 steps (Wasserstein Distributionally Robust Control and State Estimation for Partially Observable Linear Systems). If the measured horizon-cost slope stays flat as $r$ grows, our central claim is wrong.

Proposed Experiments and Decisive Checks

We use the COMPl_eb models from the infinite-horizon papers: AC15, AC17, REA4 and HE3. Each is converted to output feedback by keeping only a subset of the measured outputs and adding Gaussian measurement noise. We add the Boeing 747 longitudinal model, which is natively measurement-feedback, and the IEEE 39-bus grid to compare against the reported Gelbrich numbers. For each system we sweep $r$ from 0.01 to 10, matching the reported full-state grid.

Four metrics are recorded:

  • worst-case expected cost against the computed worst-case spectrum;
  • regret against the clairvoyant non-causal controller;
  • empirical cost under out-of-sample correlated, uniform and biased-Gaussian disturbances;
  • online operations per step.

Three checks decide the proposal.

  1. Rate check. On log-cost-gap versus $k$ plots, the fitted slope must equal $2\log\rho(r)$, with $\rho(r)$ estimated independently from the singularity analysis. The claim fails if the slope instead tracks the plant's closed-loop margin, as in the fully observed theory.
  2. Order check. The smallest rational order within 1% of $K_\star$ must match the predicted near-circle cluster count plus observer order. It must stay roughly constant while $k^\star$ grows with $r$, consistent with the reported full-state observation that order 2 works at every radius on AC15.
  3. Practical check. The rational output-feedback controller must beat both the Gelbrich controller and rolling robust optimization at matched online compute, at high $r$. If it does not, the non-rational theory is mathematically correct but practically unnecessary for these benchmarks.