Gradient Design of Nonlinear Coherent Quantum Controllers That Are Physical by Construction

In coherent feedback control, a quantum system is stabilized or steered by a second quantum device, such as an optical cavity, a squeezer, or a nonlinear crystal, instead of by a measure-then-compute classical loop. Measuring a quantum system disturbs it, and electronic processing adds delay. Coherent feedback avoids both. The cost is a design problem. The controller must itself be a physically possible open quantum system, which means its equations must preserve the quantum uncertainty relations (this requirement is called physical realizability). Today's design methods handle it in two ways. Some synthesize a linear controller and then patch realizability by attaching extra vacuum-noise inputs, which degrades performance. Others search over linear controller matrices with derivative-free evolutionary algorithms, which become expensive as the number of parameters grows. Both approaches treat real device nonlinearities, such as pump depletion or Kerr effects, as disturbances rather than as resources.

We propose to stop designing matrices and start designing devices. Every controller is written as a list of physical components: a scattering matrix, coupling operators, and a Hamiltonian, the "(S, L, H)" description of an open quantum system. Any such list is realizable automatically, including nonlinear ones. Plant and controller are wired together with the standard series-connection rules. The steady-state performance of the resulting closed loop is then optimized by exact gradients, computed by implicit differentiation of the closed loop's stationary state, with no simulation of long time horizons. The result is a first-order method for nonlinear coherent controllers that needs no noise-port repair. Its search cost should scale with controller parameter count far better than evolutionary search. It also makes the trade-off between controller ports and performance something we can measure directly.

Core Claim

Three claims, each testable. First, on the linear benchmarks where evolutionary search currently holds the record, gradient optimization over (S, L, H) parameters will match those costs at a fraction of the cost evaluations, and it will never violate realizability, because there is no constraint to violate. Second, on plants with genuinely nonlinear dynamics, controllers that are allowed nonlinear Hamiltonian terms will beat the best linear coherent controller designed for the same plant. Third, because the gradient comes from a stationary-state equation rather than a time integration, the method avoids the long-horizon error growth that limits trajectory-based optimal-control solvers.

The mechanism fits in two equations. Let $\theta$ be the controller's physical parameters: coupling rates, detunings, squeezing and Kerr strengths, and beam-splitter angles. Let $\mathcal{L}(\theta)$ be the closed-loop generator of plant plus controller, a superoperator acting on density matrices, and let $X$ be the cost observable, for example a quadrature variance or a photon number. The steady state $\rho_\theta$ and the cost satisfy

$$\mathcal{L}(\theta)\,\rho_\theta = 0,\quad \mathrm{Tr}\,\rho_\theta = 1,\qquad J(\theta) = \mathrm{Tr}(X\rho_\theta).$$

Differentiating the stationarity condition gives, for each parameter $\theta_k$,

$$\frac{\partial J}{\partial \theta_k} = -\,\mathrm{Tr}\!\left(\lambda^\dagger\, \frac{\partial \mathcal{L}}{\partial \theta_k}\,\rho_\theta\right),\qquad \mathcal{L}(\theta)^\dagger \lambda = X - J(\theta)\,\mathbb{1},$$

where $\lambda$ is an adjoint operator. The adjoint equation is solvable whenever the steady state is unique, because its right-hand side has zero expectation in $\rho_\theta$. One forward linear solve and one adjoint linear solve therefore give the full gradient, however many parameters there are. Both solves can be done matrix-free with Krylov methods, which only need repeated applications of $\mathcal{L}$. In the linear-quadratic special case, both equations reduce to small Lyapunov equations, so the linear benchmarks are solved exactly.

Relation to Prior Work

Three lines of work motivate this proposal and also define what it must beat.

The noise-port line uses Riccati-based coherent $H^\infty$ synthesis and then restores realizability by adding vacuum inputs. It appears in Fault-Tolerant Coherent $H^\infty$ Control for Linear Quantum Systems and in its application to optical parametric oscillators (OPOs) with fluctuating pumps in Fault-tolerant H∞ control for optical parametric oscillators with pumping fluctuations. It shows that realizability can be repaired, but the repair adds noise. That work also linearizes the OPO and ignores pump depletion. We do not repair anything after the fact, and we keep the depleting pump as a dynamical mode.

The (S, L, H) line, Static and dynamic coherent robust control for a class of uncertain quantum systems and Coherent robust H∞ control of uncertain linear quantum stochastic systems, already gains realizability by construction. This is the closest existing work. It restricts Hamiltonians to quadratic form and couplings to linear form so that design becomes a rank-constrained linear matrix inequality solved by alternating projections. We keep its parameterization, drop the quadratic restriction, and replace the matrix-inequality machinery with gradients through the stationary state.

Learning-Based Design of LQG Controllers in Quantum Coherent Feedback shows that standard gradient and Newton methods stall on the non-convex realizability constraints, and that a tailored differential evolution with relaxed feasibility and scheduled penalties does much better. We read that failure as evidence about the matrix parameterization, not about gradients as such. In (S, L, H) coordinates the constraint disappears, and the open question becomes whether the remaining cost landscape is benign enough for first-order methods.

From high-dimensional classical optimal control we borrow a viewpoint rather than an algorithm. A Schrödinger Eigenfunction Method for Long-Horizon Stochastic Optimal Control shows that long-horizon control is governed by a principal eigenfunction, and that learning it directly removes the horizon dependence that afflicts trajectory solvers. The coherent-feedback analogue is exact: the closed loop's infinite-horizon behaviour is the null vector of its generator. We differentiate that null vector instead of integrating trajectories. By contrast, Learning Deep Stochastic Optimal Control Policies Using Forward-Backward SDEs and Stochastic Optimal Control Matching learn classical feedback laws on commuting state variables. They have no notion of operator-valued states or realizability. We do not try to make them quantum, since a neural value function is not a device anyone can build.

Finally, Performance Bounds for Quantum Feedback Control computes certified upper bounds for measurement-based feedback on a monitored qubit in a cavity. Its bounds do not apply to coherent controllers. That makes them a sharp target: a coherent controller that exceeds them would show concretely what measurement costs.

We should be candid about which curse of dimensionality this addresses. The size of the simulated quantum state still grows exponentially with the number of optical modes. We do not claim to escape that. Matrix-free solvers push the practical limit to total Hilbert-space dimensions around $10^4$, which is a few truncated modes. The scaling we target is in the search. Evolutionary methods need many cost evaluations per parameter, while the adjoint gradient costs two linear solves regardless of parameter count. The horizon is also gone.

Research Program

We will build a differentiable library of (S, L, H) components: empty cavities, beam splitters, phase shifters, degenerate parametric amplifiers, Kerr cavities, and a two-mode χ(2) block with a dynamical pump. It will include the series and concatenation rules for wiring components into networks. For a given wiring diagram, the code assembles the closed-loop generator. For linear-quadratic diagrams it reduces automatically to a Lyapunov solve; otherwise it applies the generator matrix-free in a truncated photon-number basis. Forward and adjoint solves use preconditioned GMRES. Gradients are checked against finite differences, and cutoff convergence is checked by doubling the truncation.

Optimization uses Adam or L-BFGS over parameters reparameterized to log scale. This addresses the $10^{-5}$ to $10^4$ parameter spread reported in the differential-evolution work. We use 32 random restarts, and the restart count is reported as part of the compute budget.

A second strand of analysis maps the cost-versus-ports frontier. Because each controller input in (S, L, H) form is an actual port, we can fix the number of controller ports and optimize. This traces out, for the first time on a nonlinear plant, how much performance each additional vacuum-coupled port costs or buys. It gives a direct quantitative answer to the tension, raised by the noise-port papers, between restoring realizability and keeping noise out.

The third strand concerns failure modes. Two things can break the gradient. The closed-loop steady state may become non-unique or bistable, which is common in Kerr and pump-depleted OPO regimes, and the generator's spectral gap may shrink so the solves become ill-conditioned. We will monitor the smallest nonzero eigenvalue of the generator during optimization. When it collapses, we will switch to the time-averaged cost over a finite window and report how often this happens.

Baselines

The naive reference point is no controller at all, plus the plant with the best fixed passive mirror feedback that a one-parameter scan can find. Every reported gain is relative to these.

The strongest existing method on linear benchmarks is the tailored differential evolution of Learning-Based Design of LQG Controllers in Quantum Coherent Feedback. On the atom in a three-mirror cavity, it reports an LQG cost $J_\infty = 2.0064$ for direct coupling (previous benchmark 4.0000) and $J_\infty = 4.0801$ for indirect coupling (previous best 4.1793), with realizability violation near $10^{-20}$. We will reimplement it with its four modules: multi-start initialization, relaxed feasibility, scheduled penalties, and adaptive search ranges. It will run on our cost evaluator under the same budget of closed-loop solves and the same wall-clock limit. For the nonlinear plants we will apply it to the same (S, L, H) parameter vector, which removes its realizability handicap and isolates the search method. Beating it means reaching a cost at most 1% above its best with at least ten times fewer cost evaluations on the linear benchmark. On nonlinear plants with 20 or more parameters, it means a strictly lower cost at equal budget.

The strongest linear-theory competitor is the robust coherent $H^\infty$ design with noise ports of Fault-tolerant H∞ control for optical parametric oscillators with pumping fluctuations. It achieves attenuation $\gamma = 0.05$ and stability over the full $[-\pi, \pi]$ pump-phase range with 5% amplitude variation, but only on the linearized OPO. We will synthesize it on the linearization of our pump-depleted OPO and evaluate it on the full nonlinear model. For dynamic-versus-static comparisons we also use the degenerate parametric amplifier of Static and dynamic coherent robust control for a class of uncertain quantum systems. At $\kappa = 6.8$ that paper reports guaranteed cost bounds of 2.01 (dynamic) and 2.15 (static), with stability down to $\kappa > 6.65$. Those are upper bounds, not achieved costs. We will compare against the achieved costs of their controllers, which we recompute, not against the bounds.

The key ablation isolates nonlinearity. We run our identical gradient pipeline with controllers restricted to quadratic Hamiltonians and linear couplings. Any gain of the full method over this ablation on a nonlinear plant is attributable to nonlinear controller elements, not to better optimization. A second ablation replaces the steady-state adjoint with backpropagation through a finite-time master-equation integration over horizon $T$. This isolates the stationary-state gradient. For the nonlinear pump-depleted OPO and Kerr plants there is no reported result found on Lacuna, so all numbers there come from our reimplemented baselines.

Why not just…?

Why not just linearize the plant and use linear coherent $H^\infty$ or LQG? For small fluctuations this may be enough, and our first nonlinear experiment is built to find where it stops being enough. The linear OPO results already show that nominal linear designs go unstable beyond about ±1 radian of pump phase, and that only explicit robust synthesis reaches $[-\pi, \pi]$, within a model that drops pump depletion. We will sweep pump power from far below to above threshold. If the linear robust controller stays within a few percent of our nonlinear controller across the whole sweep, the nonlinear claim fails, and we will report it as a negative result.

Why not just run differential evolution over (S, L, H) parameters? This is a real threat. The realizability advantage we claim over matrix-space search vanishes once evolutionary search also uses (S, L, H) coordinates. What remains is search efficiency, which is why the equal-budget comparison is run in exactly that setting. We grow controller size from a single cavity (about 4 parameters) to three-mode nonlinear networks (about 40). The prediction is that the evaluation gap widens with parameter count. If it does not, the method's contribution reduces to the library and the port-frontier study.

Why not just backpropagate through a time-domain master-equation simulation? This is the simplest gradient alternative, and it may work well at short horizons. Our ablation runs it at horizons $T$ equal to 1, 5, and 20 closed-loop relaxation times. We expect two effects. If $T$ is short, the steady-state cost is biased. If $T$ is long, memory and time grow linearly in $T$ and gradient error accumulates, which is the horizon effect documented for classical solvers in A Schrödinger Eigenfunction Method for Long-Horizon Stochastic Optimal Control. If time-domain backpropagation at moderate $T$ matches our final costs at similar wall-clock, we will say so and recommend it.

Why not just use measurement-based feedback with a learned policy? That route avoids realizability entirely, but its ceilings are now known. On the monitored qubit in a cavity, Performance Bounds for Quantum Feedback Control reports a best achieved fidelity of 0.7750, with a certified upper bound of 0.7856, under homodyne detection (0.6547 and 0.6798 under photon counting). The decisive check below tests whether a coherent controller exceeds that ceiling on the same task.

Proposed Experiments and Decisive Checks

The first experiment is a sanity check on linear ground truth. We reproduce the three-mirror atom-cavity LQG benchmark in both coupling configurations. Success means $J_\infty$ within 1% of 2.0064 (direct) and 4.0801 (indirect), with realizability holding at machine precision by construction, using at least ten times fewer cost evaluations than our differential-evolution reimplementation. We report median and worst case over restarts. We repeat this on the degenerate parametric amplifier sweep over $\kappa$, reporting achieved cost and the smallest stabilizable $\kappa$.

The second experiment is the nonlinear test: an OPO whose pump is a dynamical mode, so depletion is modelled, plus a Kerr resonator plant. The target is steady-state squeezing (minimum quadrature variance) and output-field noise suppression across pump power and pump-phase fluctuation. We compare the full method against the linear-restricted ablation, the robust linear $H^\infty$ controller, and differential evolution at equal budget. The decisive outcome is a statistically separated gain of the nonlinear controller near and above threshold, where linearization is least accurate.

The third experiment is the port frontier. On the OPO plant, we optimize controllers with one to four ports and plot the best cost against port count. This directly quantifies the noise price of each port that post-hoc repair methods add implicitly.

The fourth is the measurement-versus-coherent check. We reimplement the monitored qubit-in-cavity task from its published description, verify that our implementation reproduces the 0.7750 homodyne figure with their controller class, and then design a coherent controller from our library, such as an auxiliary cavity and Kerr element coupled through the qubit's output field. Exceeding 0.7856 would be a clean demonstration that coherence buys what measurement cannot. Failing to exceed it while sitting near the bound would also be informative.

Throughout, we report Hilbert-space dimension, solver iterations, and wall-clock as functions of mode count and truncation. This states exactly where exponential state-space growth, rather than search, becomes the binding limit.