Most methods for controlling a noisy quantum system shape only the control pulse and take the environment as given. That environment is often not memoryless. In a structured surrounding, such as a photonic crystal or a resonator, information that leaks out of the qubit can later flow back into it, and this "environmental memory" can either help or hurt the operation. In several real platforms, the parameters that set this memory can themselves be tuned. Examples are the qubit's detuning from a photonic band edge or the cutoff frequency of an engineered resonator. We propose to optimize the pulse and these few environment parameters together, using alternating optimization: improve one block of variables while holding the other fixed, then swap. The pulse block is updated with Krotov's method, a monotone optimal-control algorithm. The environment block is low-dimensional and is searched globally. The result is a co-design procedure that never decreases fidelity from one round to the next and that decides, for each task, how much memory the bath should have.
The second point is the scientific payoff. The literature disagrees about whether memory is a resource or a nuisance. Memory lowers gate error and speeds evolution in some settings, yet it minimizes fidelity in a dephasing state-preparation channel. A procedure that chooses the bath's memory strength as an optimization variable turns this disagreement into a measurable quantity: the amount of memory the optimizer selects for each task.
Core Claim
Non-Markovian dynamics, meaning dynamics with environmental memory, can be rewritten as a larger memoryless system. The trick is to add auxiliary variables that represent the bath's correlation function as a sum of decaying exponentials ("poles"). Hwang and Goan did this in an extended auxiliary Liouville space and then ran Krotov iteration on a single $\sigma_z$ control field. They reached Z and identity gate errors below $10^{-5}$, and down to $10^{-8}$. They also showed that gate error grows sharply as the bath cutoff frequency $\omega_c$ moves toward the memoryless limit. Their bath was fixed in advance. Krotov's monotonicity derivation also assumes fixed equations of motion, so it cannot simply absorb the pole parameters as extra control variables.
This is where the alternating optimization used in wireless engineering fits. When a reconfigurable intelligent surface is added to a link, the transmit beamformer and the surface's reflection coefficients multiply each other (bilinear coupling). Practitioners handle this by fixing one block and solving the other with a tractable method, as in movable-antenna co-design (Jin et al., Pi et al.). Here the analogy is literal:
- The pulse $\epsilon(t)$ plays the role of the active beamformer.
- The bath parameters $\boldsymbol{\lambda}$ (pole frequencies, widths, and coupling strengths, constrained to a physically reachable set $\Lambda$) play the role of the passive surface.
- In the embedded equations, $\epsilon(t)$ enters the system Hamiltonian, while $\boldsymbol{\lambda}$ enters only the auxiliary-mode dynamics.
We alternate
$$\epsilon^{(k+1)} = \mathrm{Krotov}\big(\epsilon^{(k)};\,\boldsymbol{\lambda}^{(k)}\big),\qquad \boldsymbol{\lambda}^{(k+1)} = \arg\max_{\boldsymbol{\lambda}\in\Lambda} F\big(\epsilon^{(k+1)},\boldsymbol{\lambda}\big),$$
where $F$ is gate or state fidelity computed in the embedded space. The Krotov step does not decrease $F$ for fixed $\boldsymbol{\lambda}$. The $\boldsymbol{\lambda}$ step does not decrease $F$ for fixed $\epsilon$, provided we warm-start from the incumbent. Because $\Lambda$ has only two to six dimensions, that step can be a dense grid followed by local refinement. A global search is feasible in this block, but not over hundreds of pulse samples. The sequence of fidelities is therefore monotone and converges to a point that is stationary with respect to each block.
We claim two things:
- This co-design reaches lower error than the best fixed-bath Krotov solution at equal gate duration, especially in the short-duration and large-$\omega_c$ regime where fixed-bath control degrades.
- The memory strength the optimizer selects, measured by the trace-distance non-Markovianity $\mathcal{N}$, flips sign in usefulness between gate tasks and dephasing state-preparation tasks. This gives an operational resolution of the "memory helps or hurts" question.
Research Program
Bath model. We build on Li's Markovian embedding. That work maps non-Markovian stochastic Schrödinger equations to time-local equations using sums of Ornstein–Uhlenbeck processes, which are simple mean-reverting random processes whose correlations decay exponentially. This cuts simulation cost from $O(N^2)$ to $O(N)$ in the number of time steps $N$. We use its linear (Type II) variant, which yields closed-form master equations for the density matrix, so Krotov runs without trajectory averaging. Li's correlated-OU construction is needed for baths whose low-frequency behavior matters, because it removes spurious long-time drift. We will use it in the sub-Ohmic spin-boson case. Li's paper only simulates; our contribution is to treat its pole parameters as decision variables.
Physical knobs. The bath block $\Lambda$ is restricted to knobs with a hardware story:
- In the photonic-crystal setting of Xu et al., the knob is the qubit's detuning from the band edge. That paper derived $\tau_{QSL}=\tau/(2\mathcal{N}/(1-P_\tau)+1)$ and proposed roughly 5 GHz Stark tuning of an N-V center to cross the edge.
- In a Lorentzian-bath setting, the knobs are the cutoff $\omega_c$ and coupling strength, as for a tunable filter resonator.
Bath parameters are static during an operation, or piecewise constant over at most three segments. This keeps the block low-dimensional and plausible to implement.
Mapping where memory helps or hurts. For every optimized solution we record $\mathcal{N}$ and the information-backflow intervals. We repeat this on a task where memory is known to be harmful: dephasing remote state preparation, after Xu et al.. In that work, a birefringent quartz plate acts as a dephasing bath, fidelity is minimal where $\mathcal{N}$ is maximal, and there are "immunity windows" where fidelity does not depend on memory strength. Our prediction is concrete: co-design should move toward high $\mathcal{N}$ for fast gates and toward low $\mathcal{N}$, or into immunity windows, for dephasing preparation. If it does not, the "dual-edged" picture is incomplete.
Scope. This proposal does not address the RIS-assisted quantum network of Chehimi et al., which optimized RIS position and entanglement rates with simulated annealing under a memoryless phase-damping model. Its block structure is an obvious AO target. However, that would be a classical resource-allocation exercise and not a question about memory. We use it only as evidence that metaheuristics are the default when active and passive quantum-link parameters are tuned together.
Baselines
Naive reference. Krotov pulse optimization on a memoryless Lindblad model, deployed on the true memory-bearing bath. The flux-qubit study designed pulses this way and reported open-system errors below $10^{-4}$ under Lindblad noise. Our evaluation measures how much of that design survives when the bath actually has memory.
Strongest existing method. Fixed-bath Krotov in the extended Liouville space, as in Hwang and Goan, which reports Z and identity gate errors below $10^{-5}$ and as low as $10^{-8}$. To make this baseline as strong as possible, we run it at every bath setting on the same grid our bath block searches and report the best result. This is "pulse-only optimization at the best fixed bath," and it isolates whether interleaving matters at all. It receives the same total number of propagations as AO, the same pulse discretization, and the same Krotov step-size tuning budget. The file gives Hwang's headline numbers but no per-$\omega_c$ error values, so we will regenerate that curve ourselves. AO beats this baseline if it reaches at least a factor-of-two lower gate error at matched gate duration in the regime where the best fixed bath is above $10^{-4}$. It should also not lose elsewhere.
Sequential physics heuristic. First set the bath to maximize $\mathcal{N}$ (for gates) or minimize it (for preparation), then run Krotov once. This is a single AO round with a hand-chosen bath step. It tests whether the coupling between blocks matters, or whether "pick the most non-Markovian bath" is enough.
Joint gradient descent. Automatic-differentiation gradient descent (GRAPE-style) over pulse and bath parameters together, with projection onto $\Lambda$ and the same number of random restarts our compute budget allows.
Black-box metaheuristic. Simulated annealing over a low-dimensional pulse parameterization plus bath parameters, mirroring Chehimi et al..
Ablation. We replace the global bath-block search with a single local gradient step. This isolates whether the gains come from searching the multimodal bath landscape globally, which is our hypothesis.
Why not just…?
Why not just optimize everything jointly with gradient descent? This is the real threat. With only a few bath parameters, a joint optimizer with enough restarts could match AO, and we will not argue otherwise in advance. Our reason to expect a gap is that fidelity as a function of band-edge detuning or cutoff should be multimodal, because backflow is oscillatory. Joint descent then tends to lock the bath parameter into whichever basin the initial pulse favors. The decisive experiment gives both methods equal propagation budgets and plots final error against the number of restarts. If joint descent with restarts matches AO, we will report that the contribution is the co-design itself, not the alternation. We would then recommend the simpler method.
Why not just tune the bath to maximal memory and then optimize the pulse? Hwang and Goan and Xu et al. suggest that more memory means faster, better operations. But Xu et al.'s remote-state-preparation result shows that maximal $\mathcal{N}$ can be the worst choice. Even for gates, the best bath plausibly depends on the gate duration and the pulse. The sequential baseline tests this directly. If it matches AO, the honest conclusion is that memory strength alone predicts the right bath, which would itself be a useful physics finding.
Why not just use a memoryless model? Hwang's own results show error rising as the bath approaches the memoryless limit. A Lindblad model also cannot represent the band-edge knob at all, because a memoryless bath has no memory to tune. The naive baseline quantifies the cost.
Why not just confine the system to a decoherence-free subspace? Moody et al. steer systems into such subspaces using path-space penalties derived for memoryless, diffusive noise. They state that non-Markovian kernels are outside their formulation, and a single qubit coupled to one bath has no such subspace. We do not compare against them numerically, because the settings do not overlap.
Proposed Experiments
We will run four benchmarks:
- The single-qubit non-Markovian gate system of Hwang and Goan: Z and identity gates, swept over gate duration $\Omega t_f$ and cutoff $\omega_c/\Omega$.
- The photonic-crystal qubit of Xu et al., with state-transfer fidelity at fixed time, using detuning as the bath knob within the cited roughly 5 GHz range.
- The dephasing remote-state-preparation channel of Xu et al., with plate thickness and frequency parameter as bath knobs.
- The Lorentzian and sub-Ohmic spin-boson model of Li.
Our metrics are gate error or state infidelity, wall-clock time and number of propagations to reach a target error, the selected $\mathcal{N}$, and variance across random initializations.
Decisive checks. There are three.
The first concerns model exploitation, a known risk when the environment is itself a decision variable. An optimizer allowed to move poles may exploit errors in the few-pole approximation. Every solution found with a 3-pole embedding is therefore re-evaluated with a 7-pole embedding, which Li showed matches the exact equation. Any gain that disappears under re-evaluation is discarded.
The second check is the memory-selection sign flip described above: high $\mathcal{N}$ chosen for gates, low $\mathcal{N}$ or immunity windows chosen for dephasing preparation.
The third check is the head-to-head against joint gradient descent at matched budget. Its outcome determines whether the alternation itself, or only the co-design, is the contribution.
References
- Optimal control for non-Markovian open quantum systems
- A General Optimization Framework for Tackling Distance Constraints in Movable Antenna-Aided Systems
- 6D Movable Antenna Enhanced Multi-Access Point Coordination via Position and Orientation Optimization
- Markovian embedding procedures for non-Markovian stochastic Schrödinger equations
- Quantum speedup in a memory environment
- Non-Markovian effect on remote state preparation
- Reconfigurable Intelligent Surface (RIS)-Assisted Entanglement Distribution in FSO Quantum Networks
- Optimal control for fast and high-fidelity quantum gates in coupled superconducting flux qubits
- QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem