A sensor that sends its reading only when the controller's picture of the plant has drifted too far can save most of the channel's traffic. Such a "send-on-delta" rule is well understood for remote estimation. It is poorly understood once the controller must also respect hard safety limits and faces noise whose distribution is known only through samples. The tools that handle those two requirements assume the sensor transmits at every step. These are System Level Synthesis, which designs a controller by choosing its closed-loop responses to disturbances, and distributionally robust control, which guards against every noise distribution near the observed data. The threshold rule is an on/off switch, and inserting a switch into a closed loop seems to destroy the convexity that makes those tools work.
This note proposes a way around that apparent conflict. Suppose the sensor knows the control input, for example because the actuator sits at the plant. Then the gap between the true state and the controller's running prediction evolves independently of the controller, and the threshold rule guarantees that this gap stays inside a known box. From the controller's point of view, a threshold-triggered link is therefore an ordinary noisy measurement channel. Its "noise" has bounded support, and its probability law depends on the threshold and the disturbances but not on the controller. That reformulation keeps synthesis convex for every threshold. It also turns the transmission rate into a quantity that can be estimated before any controller is designed. Designing thresholds and controllers jointly then reduces to a low-dimensional search over thresholds wrapped around a convex robust program.
Core Claim
Consider the plant $x_{t+1} = A x_t + B u_t + w_t$. The remote controller keeps a prediction $\hat x_t$. When the sensor is silent, the prediction advances as $\hat x_{t+1} = A\hat x_t + B u_t$. When the sensor transmits, the prediction resets to $x_t$. The sensor tracks the error $e_t = x_t - \hat x_t$ and transmits whenever the candidate next error leaves the box $\{\|e\|_\infty \le \tau\}$. Because $u_t$ cancels, the error follows
$$e_{t+1} = \begin{cases} A e_t + w_t & \text{if } \|A e_t + w_t\|_\infty \le \tau,\\ 0 & \text{otherwise,}\end{cases}$$
where $\tau$ is the threshold vector, one entry per sensor channel. Two consequences follow directly.
First, the controller effectively observes $y_t = x_t - e_t$. This is the output-feedback setting of System Level Synthesis, which parameterizes the closed loop by its responses to both process noise and measurement noise under affine achievability constraints, with $-e_t$ playing the role of measurement noise. The switch has been moved out of the loop and into an exogenous signal whose support $[-\tau,\tau]$ holds by construction.
Second, the long-run transmission rate $R(\tau)$ is a property of the error chain alone. It can be computed from disturbance samples without reference to any controller.
The joint problem is
$$\min_{\tau}\; J^\star(\tau) + \lambda R(\tau),$$
where $J^\star(\tau)$ is the optimal value of a convex distributionally robust synthesis in the style of DRInC for fixed $\tau$, and $\lambda$ prices one transmission. For each $\tau$, the ambiguity set is a Wasserstein ball around empirical joint samples of $(w, e)$. We generate these samples by running the error chain on the observed disturbance pool, and we impose the exact box support on $e$.
What is new is the reduction itself. A dynamic, state-dependent transmission policy enters response-space synthesis as a controller-independent, bounded-support noise channel. This lets communication be traded against constrained, distributionally robust performance on one frontier.
Relation to Prior Work
Three papers combine distributional robustness with response-space synthesis, and all three assume the sensor transmits every step.
Distributionally Robust Infinite-horizon Control builds Wasserstein balls around a sample pool, truncates the responses to a finite horizon, and solves a semidefinite program. It proves strong duality when disturbances live on a bounded polytope. Data-Driven Distributionally Robust System Level Synthesis adds model mismatch and obtains a linear program with risk-measure constraints. Distributionally Robust SLS With Output Feedback Affine Control Policy extends this line to output feedback with affine policies.
We reuse their machinery directly. DRInC's polytope-support duality is exactly what the box-supported error $e$ needs, and the affine output-feedback form lets the controller exploit implicit information, as discussed below. We differ in giving these methods a measurement channel created by a transmission policy we also design.
Two papers show that structure-preserving linear control survives unreliable links, but treat packet loss as an external coin flip rather than a decision. These are Best Linear Controllers for Decentralized Linear Quadratic Systems and Optimal local and remote controllers with unreliable uplink channels. In both, the structure of the optimal controller is shown to hold for any zero-mean, finite-variance noise. Our transmission decisions are endogenous and depend on the state.
The distributionally robust linear-quadratic results constrain our design. Distributionally Robust Linear Quadratic Control and Optimality of Linear Policies in DRLQ Control show the adversary's worst case is Gaussian for unconstrained linear feedback. That argument cannot apply here, because $e$ is bounded and so cannot be Gaussian. This is why we build the ambiguity set from samples with an explicit support constraint rather than using the Gelbrich covariance relaxation.
On the communication side, several papers motivate threshold rules but stop short of our setting. Fundamental limits of remote estimation proves that threshold rules are optimal for scalar autoregressive sources via renewal cycles. Remote Estimation over Costly Channels shows that silence carries information, about half the channel use of a randomized policy at matched accuracy. Feedback Control via Integrated Sensing and Communication finds threshold scheduling with a certainty-equivalent linear controller under Gaussian assumptions. None of these handles hard state constraints or distributional ambiguity.
The closest communication-minimizing synthesis is Safe Control of Multi-Agent Systems with Minimal Communication. It reduces messages by making off-diagonal controller blocks low-rank. That choice fixes which channels are used offline. Our design decides online when to transmit.
Research Program
The first piece is the reduction itself, written as a proposition. Under send-on-delta triggering with an ideal channel and a sensor that knows the control input, the closed loop is exactly an output-feedback SLS problem. Its measurement noise $-e$ satisfies $\|e_t\|_\infty \le \tau$ almost surely. The joint law of $(w,e)$ depends on $(A, \tau, P_w)$ and not on the controller. A consequence is that hard constraints can be enforced robustly over the box and over the Wasserstein ball of $w$, which yields a convex program for each $\tau$.
The second piece is the rate estimator. We will estimate $R(\tau)$ by running the error chain on resampled disturbances and bound its error with standard Markov-chain concentration. We will validate it against the renewal-theory closed form for scalar autoregressive sources from Fundamental limits of remote estimation.
The third piece is the outer search. For a scalar or low-dimensional $\tau$, we will use a grid followed by golden-section refinement. For per-channel thresholds, we will use coordinate descent, where each step requires one convex solve.
The fourth piece addresses implicit information. When noise is skewed, as with the Beta and bimodal distributions in DRInC's benchmark, the expected error given silence is not zero. An affine offset that depends on how long the sensor has been silent can then correct the prediction. The affine output-feedback form of DR-SLS With Output Feedback Affine Control Policy accommodates this offset without losing convexity. For symmetric noise and symmetric boxes, the offset should vanish by symmetry, which gives a built-in consistency check.
Finally, we will treat packet loss as a stress test, not a claim. A drop lets $e$ escape the box and breaks the guarantee, the same failure of the reset property flagged in the remote-estimation papers. We will measure how gracefully constraint satisfaction degrades as a function of drop probability.
Baselines
The naive reference point is time-triggered DRInC (Distributionally Robust Infinite-horizon Control) with transmission every step. It fixes the cost and safety achievable at full rate, the level any communication-saving method should approach. We run it with the same disturbance pool, Wasserstein radius selection procedure, and response-truncation horizon as our method.
The strongest single-agent competitor is periodic down-sampling inside the same DR-SLS program. The controller receives measurements every $K$ steps, which is also convex. We choose $K$ so that its average rate matches our $R(\tau)$, and we give it an identical radius-tuning budget. Beating it means lower out-of-sample cost at equal rate, with no more constraint violations, across the 1,000 test distributions of DRInC's double-integrator protocol. DRInC reports its costs only graphically, so there is no reported result found on Lacuna for a scalar DRInC cost on that benchmark. We will reproduce the full-rate curve ourselves.
The key ablation is emulation. We design the DR-SLS controller assuming full information, attach the same threshold trigger afterwards, and pick $\tau$ by the same search. This isolates whether designing the controller with the error channel in view matters. We expect emulation to violate constraints at useful thresholds. Prior evidence points that way: nominal SLS suffered "constraint violations in 100% of the simulations" under mismatch in DR-SLS With Output Feedback Affine Control Policy.
A second ablation treats $e$ as a worst-case box disturbance in the cost as well as the constraints, discarding its sample-based law. This isolates the distributional ingredient.
For multi-agent communication, the baseline is low-rank SLS from Safe Control of Multi-Agent Systems with Minimal Communication. It reports 44 messages against 94 for its baseline, a 53% reduction, on a four-vehicle chasing task, and 1 message against 21 on an asymmetric-noise task. We will reimplement the chasing task with per-agent threshold broadcasts and match safety constraints and horizon. Beating it means fewer than 44 messages at equal safety. The comparison may be imperfect if the reimplemented dynamics differ from theirs, so we will also rerun their method on our reimplementation.
Why not just…?
Why not just down-sample periodically? Down-sampling is convex and simple, so it is a real threat. The evidence against it is for estimation, not constrained control. Threshold sampling beats periodic sampling under tight budgets in Fundamental limits of remote estimation. Periodic and randomized policies use up to twice the channel in Remote Estimation over Costly Channels. Our matched-rate frontier comparison settles the question directly. If down-sampled DR-SLS matches our frontier on the double integrator, the event-triggered machinery is not earning its keep there.
Why not just design the controller first and add a trigger afterwards? This is the emulation ablation. The controller would never see that its measurements carry errors up to $\tau$, so the tightening its constraints need would be missing. We will report violation rates as a function of $\tau$. If a simple post-hoc tightening gives the same frontier, that would show the reduction is the useful contribution and joint optimization is not. We would say so.
Why not just treat the untransmitted error as a worst-case bounded disturbance, as robust control would? This keeps safety but prices every step as if the error sat at the corner of the box. That is the conservatism Distributionally Robust Linear Quadratic Control and DRInC criticize in worst-case robust control. The worst-case-cost ablation measures the price directly. If the gap is small, the distributional component should be dropped.
Why not just use the Gaussian worst case from distributionally robust linear-quadratic control? The error $e$ is bounded by construction, so a Gaussian adversary is outside the feasible set. The covariance-only reformulation of Optimality of Linear Policies in DRLQ Control would therefore be needlessly pessimistic about $e$ and blind to its support. We include it as a variant to quantify the gap.
Proposed Experiments
The first experiment checks the rate estimator. On scalar autoregressive sources, the estimated $R(\tau)$ must match the renewal closed form within its concentration bound.
The main experiment traces the communication–cost frontier on the double integrator, using DRInC's bimodal and Beta noise, 1,000 shifted test distributions, and safety constraints. We compare our co-design against the full-rate reference, down-sampled DR-SLS, emulation, and the worst-case-error ablation. The two metrics are out-of-sample cost and the rate of constraint violations. The decisive outcome is a frontier that dominates down-sampling at matched rate with zero violations where emulation violates.
On the small-data two-dimensional system with 20 trajectories of length 10 from Data-Driven DR-SLS, we test whether the reduction survives model mismatch. There, the controller-independence of $e$ holds only approximately, because the prediction uses the nominal model. We will measure how far the realized error leaves the box.
The implicit-information check compares the affine silence-age offset under symmetric and skewed noise. It should help only in the skewed case.
The packet-drop sweep reports violation rates as drop probability rises. This marks the boundary of the guarantee.
Finally, the four-vehicle chasing task compares message counts against the reported 44.
References
- System level synthesis
- Distributionally Robust Infinite-Horizon Control: From a Pool of Samples to the Design of Dependable Controllers
- Data-Driven Distributionally Robust System Level Synthesis
- Distributionally Robust System Level Synthesis With Output Feedback Affine Control Policy
- Best Linear Controllers for Decentralized Linear Quadratic Systems
- Optimal Local and Remote Controllers With Unreliable Uplink Channels: An Elementary Proof
- Distributionally Robust Linear Quadratic Control
- Optimality of Linear Policies in Distributionally Robust Linear Quadratic Control
- Fundamental Limits of Remote Estimation of Autoregressive Markov Processes Under Communication Constraints
- Remote Estimation of Markov Processes over Costly Channels: On Implicit Information Benefits
- Feedback Control via Integrated Sensing and Communication: Uncertainty Optimisation
- Safe Control of Multi-Agent Systems with Minimal Communication