Robust predictive controllers keep a real system safe by planning a nominal path and reserving a safety margin, called a "tube," big enough to contain every trajectory the disturbance could produce. Today that margin is computed from one number: a bound on how large the disturbance can get at any instant. This works poorly when the disturbance is a mix of a slow, small drift and a fast, large jitter, which is common for wind on a drone or vibration on a process line. A stable feedback loop smooths out fast jitter but lets slow drift build up. A single amplitude bound cannot express that difference, so the controller assumes the worst-case disturbance is a constant push at the full jitter amplitude. The margin inflates and the controller becomes sluggish or infeasible. We propose to describe the disturbance by separate bounds on its wavelet coefficients at each time scale. A wavelet transform splits a signal into localized slow and fast components, and it has a perfect-reconstruction property, so the scale-wise bounds together cover the whole disturbance with nothing left over. We then compute the tube margins of a Koopman-lifted linear controller directly from this scale-structured set.
The opportunity is that two bodies of work sit on either side of this idea without touching. Robust tube-based model predictive control with Koopman operators (r-KMPC) has the guarantees. It lifts a nonlinear plant into a linear surrogate, runs a convex tightened planner plus an offline ancillary feedback law, and proves recursive feasibility, meaning the optimization stays solvable at every future step. However, it lumps the external noise and the lifting residuals into one static set. On the other side, Wavelet Predictive Representations for Non-Stationary Reinforcement Learning (WISDOM) shows that splitting a signal into slow approximation coefficients and fast detail coefficients cleanly separates drift from shocks. That work does so inside a model-free reinforcement learner with no hard constraints and no stability proof. Our claim is that the wavelet split belongs in the disturbance model of the tube, not in the learned dynamics model, because that is where it changes a provable quantity: how much the constraints must be tightened.
Core Claim
Lift the nonlinear plant $x_{t+1}=f(x_t,u_t)+w_t$ into a Koopman surrogate. The Koopman operator is the linear operator that advances functions of the state, and we approximate it on a finite set of lifting functions. The error between the true lifted state and the nominal lifted state then evolves linearly under the ancillary feedback, $e_{t+1}=\Phi e_t + w_t$. Here $\Phi$ is the closed-loop error matrix, and $w_t$ now also absorbs the lifting residual, as in r-KMPC.
Instead of $\|w_t\|_\infty\le \bar w$, we assume a scale-wise set over any horizon window:
$$\mathcal W_\beta=\{\,w_{0:N-1} : |\langle w, \psi_{j,k}\rangle|\le \beta_j \ \text{for every scale } j \text{ and shift } k\,\}.$$
In this set, $\psi_{j,k}$ are atoms of an undecimated (shift-invariant) wavelet frame at scale $j$ and position $k$, and $\beta_j$ is the coefficient bound at that scale. A constraint $c^\top e \le$ margin at prediction step $i$ is then tightened by
$$h_i(\beta)=\max_{w\in\mathcal W_\beta}\; c^\top \sum_{k<i}\Phi^{\,i-1-k} w_k .$$
This is a linear program solved offline, once per step and per constraint.
The mechanism behind the expected gain is simple. If the error response $c^\top\Phi^m$ is positive and smooth, the worst disturbance under a single amplitude bound is a constant at $\bar w$, and the margin scales with the loop's steady-state gain times the full amplitude. Under $\mathcal W_\beta$, the steady-state gain multiplies only the small coarse-scale bound. The large fine-scale bound is multiplied by the loop's much smaller high-frequency gain.
We use the undecimated frame deliberately. The file notes that standard decimated transforms are shift-variant, and a shift-variant disturbance set would make the tightening depend on where the horizon window falls, which breaks the shifted-candidate argument behind recursive feasibility. Undecimated frames are larger than decimated ones, but here that cost is paid only offline.
Research Program
The work has three parts.
The first part is theoretical. We need recursive feasibility and input-to-state stability under $\mathcal W_\beta$. This is the real technical risk. The classical constraint-tightening proof needs the margin to grow by exactly the effect of one disturbance sample per step. With an amplitude bound that holds because the set treats each time step independently. With overlapping wavelet constraints it does not hold automatically. Our planned route borrows the indirect-feedback idea from Recursively Feasible Stochastic Model Predictive Control using Indirect Feedback. That paper keeps the planner's prediction anchored to the nominal state rather than the measured one, so feasibility never depends on realized disturbances. In our setting, the tightening is computed over the nominal predictions, and the realized error is handled by the ancillary law. If that argument fails, the fallback tightens with the sum, over scales, of each scale's individual worst case. This is still far tighter than a single amplitude bound when the $\beta_j$ differ, and recursive feasibility follows from standard arguments. We will report how much tightness the fallback gives up.
The second part is calibration. The $\beta_j$ come from wavelet coefficients of one-step lifting residuals on held-out data, which is a direct measurement. The alternative used in classical practice is a shaping filter, which models the disturbance as filtered bounded noise and requires first identifying that filter and then inverting it. We also study an online version. Here the $\beta_j$ are refined by a causal wavelet transform of observed residuals and are allowed only to shrink. This mirrors the shrink-only set-membership update of Linear robust adaptive model predictive control: Computational complexity and conservatism, which achieves fixed complexity with only 30 decision variables. We aim to keep that property: the online problem is a quadratic program the same size as r-KMPC's, and only the precomputed margins change.
The third part addresses the hand-tuning problem that WISDOM leaves open: choosing the decomposition depth $M$ and the wavelet family. We choose them by minimizing the total margin $\sum_i h_i$ on calibration data, which gives the depth selection a control-relevant objective.
Baselines
The naive reference point is r-KMPC exactly as published. It uses the same EDMD (extended dynamic mode decomposition) lifting, the same ancillary gain $K$, and the same origin-preserving constraint $\Psi(0)=0$, with a single static disturbance bound fit to the same residual data. It is the strongest existing method for provable Koopman tube control. On the Van der Pol oscillator under sinusoidal noise it reports a cumulative cost of $J\approx 258$, against $J\approx 431$ for standard Koopman MPC (Robust tube-based model predictive control with Koopman operators). We first reproduce that setting until our r-KMPC is close to 258. Only then do we compare. Every method in a comparison shares the lifting, training data, prediction horizon, cost weights and solver. Beating r-KMPC means lower cumulative cost with zero constraint violations over the same disturbance realizations, together with a larger set of feasible initial states.
The second baseline is a shaped-disturbance tube. The disturbance is modeled as a bank of band-pass filters driven by bounded inputs. The filter states are appended to the error system, and a standard tube is built on the augmented system. Each band's input bound is fit from the same residuals with the same tuning budget, and the band edges are tuned on the same validation grid we use for $M$. This is the most serious classical competitor, because it also exploits frequency structure.
The third baseline is offset-free tube MPC. A disturbance observer with integral action removes slow drift, and a tube covers the remainder. It is the classical answer to "slow drift plus fast jitter."
For regularized data-driven control we include distributionally robust DeePC from Regularized and Distributionally Robust Data-Enabled Predictive Control. DeePC (data-enabled predictive control) predicts directly from stored trajectory data. We sweep its robustness radius on the validation set, because that paper documents a U-shaped cost curve in that radius.
Two ablations isolate the key ingredient. The first sets every $\beta_j$ equal, which removes the scale structure while keeping the transform; this tests whether the gain comes from per-scale bounds or merely from the extra machinery. The second swaps the undecimated frame for a decimated transform; this tests whether shift-invariance actually matters for feasibility in practice.
Why not just…?
Why not just shape the disturbance with frequency-weighting filters? This is the real threat, and we treat it as one. A filter bank with per-band bounds is close in spirit to our approach. We expect two differences. First, our bounds are read directly off residual coefficients with perfect reconstruction. A shaping model's bounds are valid only if the disturbance truly is filtered bounded noise, and recovering the filter input from residuals is ill-conditioned. Second, wavelet atoms are localized in time, so intermittent bursts inflate only the bounds near the burst's scale rather than a stationary band. If the shaped-filter baseline matches our tightening on all three benchmarks, the honest conclusion is that our contribution reduces to a calibration and depth-selection recipe. We would report it that way.
Why not just use a disturbance observer for the drift? An observer cancels slow drift well when the drift is smooth. It adds its own estimation error, however, and that error must itself be bounded inside the tube. Its transient after an abrupt regime shift is also exactly the large slow error our coarse-scale bound accounts for. The step-drift condition in our experiments is designed to show whether the observer's transient forces a larger tube than the scale-wise set.
Why not just regularize DeePC or use Fourier features? Norm penalties on the trajectory weights treat all frequencies the same, and the documented U-shaped sensitivity in Regularized and Distributionally Robust Data-Enabled Predictive Control suggests fragile tuning. Unconstrained data-driven MPC diverged on the Four-Tank system under noise in Data-Driven Model Predictive Control with Stability and Robustness Guarantees. A global Fourier bound loses timing, which WISDOM argues is essential for transient shifts. Neither alternative yields a margin that provably covers the disturbance. We nonetheless run both rather than assume they lose.
Proposed Experiments
We use the three benchmarks above, Van der Pol, CartPole and the Four-Tank system, each with disturbances built from a slow component (constant offset, ramp, or step drift), a fast bounded jitter, and optional intermittent bursts. We sweep the ratio of fast to slow amplitude.
On Van der Pol we first reproduce r-KMPC's published sinusoidal-noise setting with $8\cdot 10^5$ training points. Narrowband sinusoidal noise is a case where scale-wise bounds should already help. We then add the composite disturbances. CartPole tests an open-loop unstable plant, where the error system's steady-state gain is large and the predicted gap between the methods is widest. The Four-Tank system tests coupled, non-minimum-phase multivariable dynamics, using the noise level from Data-Driven Model Predictive Control with Stability and Robustness Guarantees.
The metrics are cumulative cost, constraint-violation count, the tightening margin $\sum_i h_i$, the volume of the feasible set of initial states, and online solve time. Solve time should match r-KMPC, since only the precomputed margins differ.
The decisive check has a clear prediction. The margin ratio between r-KMPC and our method should grow roughly as the ratio of the loop's steady-state gain to its high-frequency gain, weighted by the fast-to-slow amplitude ratio. The ratio of gains comes from the worst-case argument in the Core Claim and can be computed in advance. If measured margins track this prediction, if violations stay at zero under adversarially chosen disturbances inside $\mathcal W_\beta$, and if the equal-$\beta$ ablation erases the gain, the claim holds. If the shaped-filter tube matches us everywhere, or if the decimated variant shows no feasibility failures, those are informative negatives and the paper reports them as such.
References
- Robust tube-based model predictive control with Koopman operators
- Wavelet Predictive Representations for Non-Stationary Reinforcement Learning
- Recursively Feasible Stochastic Model Predictive Control using Indirect Feedback
- Linear robust adaptive model predictive control: Computational complexity and conservatism
- Regularized and Distributionally Robust Data-Enabled Predictive Control
- Data-Driven Model Predictive Control with Stability and Robustness Guarantees