Robust Schrödinger Bridges for Linear Swarms under Unknown Noise

Suppose a swarm of drones has to move from one spatial formation to another, and the wind noise is only known to lie within a range. The controller has to choose feedback laws now, while the noise level keeps changing in ways nobody chose in advance. Recent work prices this kind of mass movement with the energy an optimal controller would spend, the "cost-to-go" of a linear-quadratic regulator (LQR), and it assumes a single, fully known noise model. Separately, mathematical finance has shown when a worst-case evaluation over many possible noise models stays logically consistent over time. The requirement is called pasting stability. It says that if you follow one model up to some random time and then switch to another, the result must still be an allowed model. We propose to join the two. The result is a continuous-time, control-priced transport problem in which an adversary sets the noise level moment by moment within a pasting-stable set. We will also build a solver that computes the resulting plan.

The central observation is that robustness here is not the same as assuming the worst noise level. For an ordinary LQR problem the cost-to-go is convex, so the adversary always picks maximum noise, and "robust" collapses to "pessimistic." That collapse breaks once the target distribution must be hit. The terminal Lagrange multiplier (the Kantorovich potential, the price the planner attaches to ending up at each place) has concave directions wherever the target is wider than the passive dynamics would produce. It also has concave directions between the modes of a multi-modal target. In those directions extra noise helps the controller, so a worst-case adversary switches to minimum noise there. The worst case therefore changes with state and time, and only a pasting-stable ambiguity set turns this into a plan that the controller would not want to revise halfway through.

Core Claim

Consider a linear system $dX_t = (AX_t + Bu_t)\,dt + \sigma_t\, dW_t$. Here $X_t$ is the agent state, $u_t$ is the control, $W_t$ is Brownian motion, and the adversary may choose the noise intensity $\sigma_t\sigma_t^\top$ adaptively within a matrix interval $[\Sigma_{lo}, \Sigma_{hi}]$. The set of laws generated this way is pasting-stable, in the sense of Superhedging and Dynamic Risk Measures under Volatility Uncertainty. The robust control-priced bridge value $V(t,x)$ then satisfies

$$\partial_t V + x^\top A^\top \nabla V - \tfrac12 \nabla V^\top B R^{-1} B^\top \nabla V + \tfrac12 x^\top Q x + \sup_{\Sigma_{lo} \preceq \Sigma \preceq \Sigma_{hi}} \tfrac12 \mathrm{tr}(\Sigma \nabla^2 V) = 0, \qquad V(T,\cdot) = \psi.$$

In this equation $R \succ 0$ penalizes control effort and $Q \succeq 0$ penalizes state deviation, as in The LQR-Schrödinger Bridge. The terminal cost $\psi$ is the dual potential that enforces the target distribution $\mu_1$. The supremum makes the noise term a G-type nonlinear expectation rather than a linear one. Its maximizer depends on the sign of the curvature $\nabla^2 V$ in each direction, which is where this problem departs from fixed-noise robustness.

We make three claims, each one checkable.

First, whenever $\psi$ is not convex, fixing the noise at $\Sigma_{hi}$ gives a plan that is strictly suboptimal and that misses the target under the true worst case. $\psi$ fails to be convex whenever the target requires spreading out or splitting into modes.

Second, a robust plan built on a set that is not pasting-stable is time-inconsistent. An example of such a set is "the noise level is constant but unknown." Re-solving the problem midway from the realized state gives a different policy.

Third, an exact robust marginal constraint, meaning every admissible law must end at exactly $\mu_1$, is generally infeasible for a nondegenerate noise interval. The right object is therefore a penalized or ball-constrained bridge, and we characterize it.

Positioning

The proposal draws on three groups of prior work, each in a different role.

Work that supplies the cost structure. Two papers show that LQR cost-to-go is quadratic in the start and end states and can be used directly as an optimal-transport ground cost:

The stochastic version comes from The LQR-Schrödinger Bridge. It solves a discrete-time entropic bridge through forward-backward Riccati recursions with Gaussian marginals, and it names continuous time as the open next step. All three papers assume one known noise model, and we drop that assumption.

Work that supplies the consistency theory. The pasting-stability results establish that time-consistency is equivalent to stability under concatenation. Dynamic Monetary Risk Measures for Bounded Discrete-Time Processes proves this in discrete time. Coherent and convex monetary risk measures for unbounded càdlàg processes extends it to processes. Constructing sublinear expectations on path space shows that the resulting conditional expectations are well defined pathwise. Superhedging and Dynamic Risk Measures under Volatility Uncertainty ties these value processes to second-order backward SDEs (2BSDEs). These results are existence theory, with no controls, no transport constraints, and no algorithms. We reuse the pasting construction and add the missing algorithm.

Work that supplies the solver. Likelihood Training of Schrödinger Bridge using Forward-Backward SDEs Theory gives a continuous-time Schrödinger-bridge solver built on forward-backward SDEs, but with a Brownian prior and no state cost or ambiguity. Trust Region Constrained Measure Transport in Path Space for Stochastic Optimal Control and Inference scales path-space stochastic optimal control to 200 dimensions under a single nominal diffusion. We borrow the forward-backward machinery from the first and the stabilization ideas from the second, and replace the linear expectation with the curvature-dependent supremum above.

Discrete scenario-tree methods, which we treat as constraints on the design. Risk-averse risk-constrained optimal control and Stochastic optimal control with dynamic, time-consistent risk constraints show that nested time-consistent risk can be handled exactly on scenario trees. Their discrete-state, tree-structured formulations become prohibitively large in continuous state spaces. That is why we work with the continuous-time partial differential equation instead.

Research Program

Theory in the Gaussian case. Restrict the controller to affine feedback $u = K_t x + v_t$ and the adversary to deterministic noise paths. The state covariance then evolves as $\dot P = (A + BK)P + P(A+BK)^\top + S_t$, with $S_t \in [\Sigma_{lo}, \Sigma_{hi}]$. We will show that the robust covariance-steering problem toward a ball around the target covariance is a semidefinite program. We will also prove that the adversary's optimal $S_t$ selects $\Sigma_{lo}$ along eigendirections in which the plan must expand the covariance.

Next, we will compare this deterministic adversary against the full adaptive, pasting-stable adversary. If the deterministic adversary does strictly better than the adaptive one, the gap measures how much the full pasting-stable adversary adds. We conjecture the gap is zero for Gaussian targets and positive for mixture targets.

We will also prove the infeasibility result for exact robust marginal constraints. Finally, we will prove a dynamic programming principle for the penalized robust bridge, relying on pasting stability of the closed-loop law set.

Solvers. In state dimension up to 4 (the double integrator), we solve the robust bridge equation on a grid. We alternate between the equation itself and Sinkhorn-style updates of the terminal potential $\psi$. This gives a reference solution.

For larger dimension, we build a deep forward-backward solver on the SB-FBSDE codebase ghliu/sb-fbsde. The noise term is evaluated through a learned Hessian head, and the inner supremum has a closed form per eigendirection of that Hessian. This is a numerical 2BSDE scheme, the kind of algorithm the volatility-uncertainty literature lists as missing.

Continuous-time bridge. Setting the noise interval to a single point recovers a continuous-time LQR-Schrödinger bridge. We will check it against The LQR-Schrödinger Bridge as the step size shrinks. This settles that paper's open question as a side result.

Baselines

All methods share the same data and budget. They see the same dynamics, marginals, and cost matrices. The same adversary evaluates every plan, and that adversary is computed as a best response from the grid solution or by trajectory optimization. Neural methods get equal wall-clock time on one GPU and equal hyperparameter search. The primary score is worst-case terminal discrepancy to $\mu_1$, measured by maximum mean discrepancy (MMD) and 2-Wasserstein distance. Worst-case control energy is reported next to it.

Naive reference: nominal plan. This is the nominal LQR-transport plan. For Gaussian cases it is computed in closed form as in Discrete-Time Linear-Quadratic Regulation via Optimal Transport. Otherwise it is the nominal continuous-time LQR bridge at the midpoint noise level. It shows how much ambiguity costs a method that ignores it.

Strongest simple competitor: fixed maximum noise. This is the same bridge solved at $\Sigma_{hi}$. It is the answer most reviewers would propose, and it is exactly optimal whenever the potential is convex. Beating it means a statistically significant reduction in worst-case terminal discrepancy on expansion and mixture targets. We predict ties on contraction targets and will report those ties as confirmation, not failure.

Strongest existing learned method. This is SB-FBSDE (Likelihood Training of Schrödinger Bridge using Forward-Backward SDEs Theory) with the LQR prior added and noise set to the nominal or maximum level. For higher dimensions we also run TR-LV from Trust Region Constrained Measure Transport in Path Space for Stochastic Optimal Control and Inference. These papers report CIFAR-10 FID 3.01 and a 43.95% Chignolin transition hit rate. Neither number transfers to our tasks, and no reported result found on Lacuna for any robust control-priced bridge benchmark.

Ablation isolating pasting stability. The robust plan is optimized against the non-pasting set of constant-but-unknown noise levels. It sees the same interval and is solved by min-max over constants. Comparing it with our method isolates exactly what time-varying, pasting-stable adversaries add.

Discrete-time Riccati bridge. We also run the discrete-time Riccati bridge at several step sizes, to measure the discretization gap.

Why not just…?

Why not just assume the worst noise level? For convex potentials this is exactly right, and our first experiment is designed to confirm that. The threat is real in that regime. Where the target requires expansion, the curvature-sign argument predicts that maximum noise overshoots the target spread, while an adversary at minimum noise leaves it undershooting. A plan tuned to $\Sigma_{hi}$ therefore fails against the low-noise adversary.

The diagnostic is to plot, along the trajectory, the fraction of mass sitting in regions where the Hessian of $V$ has a negative eigenvalue. The gain over the maximum-noise plan should track that fraction. If it does not, the idea is wrong.

Why not just take the continuous-time limit of the discrete Riccati bridge? That limit yields the nominal continuous-time bridge, which we recover as a special case. It contains no adversary. In addition, the curvature-dependent supremum makes the closed-loop law non-Gaussian under adaptive noise, so the Gaussian Riccati structure of The LQR-Schrödinger Bridge cannot represent the worst case. The Gaussian-versus-mixture comparison of the two adversaries tests this directly.

Why not just use nested risk on a scenario tree? Risk-averse risk-constrained optimal control is time-consistent by construction, and its conic program scales linearly in the number of tree nodes. The number of nodes, however, grows exponentially with horizon and branching. We will run it on the 1D and double-integrator tasks with matched compute, and report the branching factor it needs to reach our worst-case discrepancy.

Why not just re-plan online with a nominal model? Receding-horizon re-planning reacts to disturbances but offers no guarantee against an adversary that anticipates it. We include nominal re-planning at a fixed rate as an extra baseline. It also serves the time-consistency test below.

Proposed Experiments

1D bimodal transport. We transport $\mathcal N(0,1)$ to the two-mode mixture at $-8$ and $4$ from Learning Paths for Dynamic Measure Transport: A Control Perspective. That paper reports MMD 0.137 and 37.5% of mass in the left mode, against a 65.4% ground truth, for its learned path under known dynamics. That figure is our nominal-case sanity reference. The test is worst-case left-mode mass and worst-case MMD across noise intervals.

4D double integrator. We run the 4D double-integrator tasks from Optimal Transport of Linear Systems over Equilibrium Measures and Distribution Steering for Discrete-Time Linear Systems with General Disturbances using Characteristic Functions. The second paper reports $L_1$ characteristic-function error 0.0026 under known Gaussian noise. We will choose three target types:

  • contraction targets, where we predict a tie with the maximum-noise plan;
  • expansion targets, where we predict our method wins;
  • mixture-merge and mixture-split targets, where we predict the largest wins.

Swarm formation. We use a formation task modeled on the 10,000-agent "ETH Zurich" deployment in Discrete-Time Linear-Quadratic Regulation via Optimal Transport, which was computed in 9 to 12 seconds on a desktop CPU. This tests the neural solver in 8 to 12 state dimensions against TR-LV.

Time-consistency test. At times $T/4$, $T/2$ and $3T/4$ we re-solve each robust method from the realized empirical state distribution and measure the change in policy. Our method should show change near solver tolerance. The constant-noise ablation should show change that grows with interval width.

Discretization study. We shrink the time step of the discrete Riccati bridge and confirm that it converges to our nominal special case.

What would falsify the proposal. The proposal fails if any of three things happens:

  • the pasting-stable plan does not beat fixed maximum noise on expansion or mixture targets;
  • the constant-noise ablation turns out equally time-consistent;
  • the Gaussian deterministic-adversary semidefinite program matches the adaptive adversary on mixtures.

The third outcome would mean a simple semidefinite program suffices and no neural 2BSDE solver is needed, in which case we would publish that program instead.