---
title: Floquet-Informed Learning Algorithm
url: https://www.emergentmind.com/topics/floquet-informed-learning-algorithm
type: topic
---

# Floquet-Informed Learning Algorithm

A Floquet-informed learning algorithm is a computational scheme for periodically driven quantum systems in which the learning component is constrained or structured by Floquet theory: periodicity, quasienergies, Floquet Hamiltonians, Floquet modes, micromotion, or Floquet-band structure enter the representation, target, cost function, similarity metric, or control loop. In current literature, the designation covers model-free reinforcement learning for Floquet state preparation, Hamiltonian learning of stroboscopic generators, Fourier-domain inverse problems for driven Hamiltonians, operator-learning surrogates, unsupervised phase-diagram inference, and hybrid quantum-classical eigensolvers [1808.08910] [2208.13837] [2509.02331] [2509.07084] [2106.11468] [2503.11613].

## 1. Floquet structure as an algorithmic prior

The common starting point is a Hamiltonian with temporal periodicity, \(H(t)=H(t+T)\), together with the one-period propagator \(U(T,0)\). Several papers use the standard Floquet relation
\[
e^{-iH_F T}=U(T,0),
\]
or its equivalent quasienergy formulation, as the mathematical object that organizes learning. In the kicked-top Hamiltonian-learning setting, the learned quantity is explicitly the stroboscopic generator \(H_F(\tau)\), expanded perturbatively through the Floquet-Magnus series [2208.13837]. In Floquet-ADAPT-VQE, the periodic problem is lifted into an extended Floquet-Hilbert space and converted into a time-independent eigenvalue problem for an enlarged Hamiltonian \(\hhat H_\mathrm{F}\) [2503.11613]. In unsupervised Floquet topology, the informative object is not merely \(U(T)\) but the periodic unitary
\[
U_\varepsilon(k,t)=U(k,t)\exp\!\big[iH^{\rm eff}_\varepsilon(k)t\big],
\]
which retains intra-period structure [2106.11468].

A central distinction in the literature is whether Floquet structure is used explicitly or implicitly. In explicit formulations, the algorithm reconstructs, approximates, or diagonalizes a Floquet Hamiltonian or a periodic unitary. In implicit formulations, Floquet theory defines the control target or the search landscape, but the learner itself does not manipulate an analytic \(H_F\). The Kapitza-oscillator reinforcement-learning study is the clearest example: the target is a Floquet eigenstate localized near the inverted position, yet the agent is deliberately model-free and is not given the Hamiltonian or the Floquet solution [1808.08910].

This separation is important because it prevents a common misreading: a Floquet-informed method need not “learn the Floquet Hamiltonian” in the narrow sense. The informing role of Floquet theory may instead reside in the target manifold, in the harmonic representation, in a symmetry-constrained similarity measure, or in an extended-space embedding.

## 2. Principal algorithmic families

The existing literature spans several distinct but related paradigms.

| Family | Floquet-informed ingredient | Representative example |
|---|---|---|
| Reinforcement learning control | Floquet target state, micromotion, periodic drive | Quantum Kapitza oscillator [1808.08910] |
| Reinforcement learning for bosonic codes | Engineered Floquet Hamiltonian and driven cavity controls | Cat-code preparation [2510.22227] |
| Hamiltonian learning | Floquet-Magnus-guided Ansatz for \(H_F\) | Kicked top [2208.13837] |
| Fourier inverse reconstruction | Truncated Fourier series and Floquet-band linear system | Periodically driven local Hamiltonians [2509.02331] |
| Neural operator surrogate | Fourier-space operator learning on periodic time domain | FNO for \(H_F\), observables, and operator growth [2509.07084] |
| Unsupervised phase discovery | Full intra-period unitary, adiabatic deformation, symmetry | Floquet topological boundaries [2106.11468] |
| Hybrid quantum-classical eigensolver | Extended Floquet-Hilbert-space Hamiltonian | Floquet-ADAPT-VQE [2503.11613] |

Across these families, two organizing axes recur. The first is **stroboscopic versus intra-period information**. Some methods target \(U(T)\) or \(H_F\) alone, whereas others require the full micromotion within a period. The second is **physics-guided representation versus black-box optimization**. Some methods build the model directly in Fourier or commutator bases; others leave the policy largely unconstrained and let Floquet structure enter only through rewards or targets.

## 3. Reinforcement learning for Floquet control and state preparation

In the quantum Kapitza oscillator, the control task is to steer the ground state of the undriven oscillator to a target state \( |\psi_\ast\rangle \), defined as the Floquet eigenstate localized near the inverted position \(\theta=\pi\), while maximizing
\[
F_h(t_f)=|\langle \psi(t_f)|\psi_\ast\rangle|^2.
\]
The control field is a bang-bang sequence \(h(t)\in\{0,\pm 4\}\). The agent uses a modified Watkins tabular Q-learning scheme with eligibility traces and replay; the RL state is the protocol history itself, the action set is \(\mathcal A=\{-4,0,+4\}\), and the reward is an estimated final fidelity from repeated binary measurements, \(r\approx n/m\). The study reports that the agent, starting from zero knowledge of the system, learns protocols with fidelity saturating around \(80\%\); it remains effective with noisy initial states and random control failures, yielding fidelities around \(\sim 73\%\) and \(\sim 71\%\), respectively. It also shows that non-stroboscopic protocols with 8 bangs per cycle outperform stroboscopic protocols with 1 bang per period at moderate drive frequencies, establishing micromotion as a usable control resource rather than a nuisance [1808.08910].

A more recent reinforcement-learning formulation targets bosonic code preparation in a periodically driven superconducting cavity. There the target state is \(\rho_0=|\psi_0\rangle\langle\psi_0|\), with \(|\psi_0\rangle=|0_{q,\Theta}\rangle\) for the \(q=4\) rotational code, and the drive parameters \(\beta(t)\) and \(\omega(t)\) are tuned by TD3. The system state \(s_t\) is the density matrix \(\rho(t)\), the action updates the driving amplitude and frequency, and the reward is shaped from the fidelity
\[
\mathcal{F}[\rho_0,\rho(t)] = \sqrt{\text{Tr}[\rho(t)\rho_0]}.
\]
The dynamics include dissipation and dephasing through a Lindblad master equation with photon-loss rate \(\kappa\) and dephasing rate \(\eta\). For the four-fold cat code, the reported preparation time is about \(50\times \frac{2\pi}{\omega_0}\), compared with roughly \(5000\times \frac{2\pi}{\omega_0}\) for an earlier adiabatic ramp, while the final fidelity remains around \(\mathcal{F}\sim 0.998\). The paper further reports transfer of a network trained at \(\kappa=\eta=10^{-7}\omega_0\) to stronger noise such as \(\kappa=\eta=10^{-4}\omega_0\), and recovery of high fidelity by retraining at \(\kappa=10^{-3}\omega_0\) [2510.22227].

Taken together, these studies establish two concrete meanings of “Floquet-informed” in reinforcement learning. In one, the Floquet aspect lies in the target eigenstate and in the exploitable micromotion, while the learner is model-free. In the other, Floquet engineering supplies the target Hamiltonian structure and the learner optimizes continuous drive parameters inside an open-system control loop. This suggests that the phrase denotes a family of control strategies rather than a single update rule.

## 4. Hamiltonian learning and reconstruction of driven generators

Hamiltonian learning provides a more explicit use of Floquet structure. In the kicked top, one period of Trotterized evolution is
\[
U_\tau=e^{-iH_z\tau}e^{-iH_x\tau}\equiv e^{-iH_{\mathrm F}(\tau)\tau},
\]
and for sufficiently small \(\tau\), \(H_{\mathrm F}(\tau)\) admits a Floquet-Magnus expansion. The learning protocol prepares several initial states, evolves them under the black-box Trotter block, and constructs a constraint matrix
\[
M_{i,j}=\langle\psi_i(t)|h_j|\psi_i(t)\rangle-\langle\psi_i(0)|h_j|\psi_i(0)\rangle.
\]
Assuming an Ansatz \(H(\mathbf c)=\sum_j c_j h_j\), the coefficients are reconstructed by
\[
\mathbf c^{\rm rec}=\arg\min_{\|\mathbf c\|=1}\|M\mathbf c\|,
\]
with the smallest singular value
\[
\lambda_1=\frac{\|M\mathbf c^{\rm rec}\|}{\|\mathbf c^{\rm rec}\|}
\]
serving as the central diagnostic. Below the Trotter threshold, if the Ansatz includes terms up to order \(k\), the residual scales as \(\lambda_1(\tau)=\mathcal O(\tau^{k+1})\). Near and beyond the threshold, \(\lambda_1(\tau)\) saturates to a \(\tau\)-independent plateau, the reconstructed coefficients become effectively random, and the same regime coincides with spectral statistics approaching the circular unitary ensemble, with \(r_{\mathrm{CUE}}\approx 0.5996\) rather than the Poisson value \(r_{\mathrm{POI}}\approx 0.39\) [2208.13837].

A different reconstruction strategy exploits periodicity directly in Fourier space. The Hamiltonian is written as
\[
H(t)=\sum_{m=-M}^{M} e^{-im\omega t}\,H_m,\qquad H_m=\sum_{i=1}^{R} c_{m,i}\,P_{m,i},
\]
with total coefficient count \(K=(2M+1)R\). Substituting the Fourier expansions into Schrödinger’s equation gives the Floquet-band relation
\[
(\varepsilon_\alpha+k\omega)\ket{u_\alpha^k} =\sum_{m=k-M}^{k+M} H_{k-m}\ket{u_\alpha^m},
\]
which is converted into a linear inverse problem \(\mathbf A\,\mathbf c=\boldsymbol\beta\). The paper states that each correlator requires \(\mathcal{O}(N^2)\) time-domain measurements, that the total sample complexity is
\[
\mathcal{O}\!\big(N^2(2M+1)^2LR\big),
\]
often summarized as \(\mathcal{O}\!\big(N^2M^2\,\mathrm{poly}(\log d)\big)\), and that the classical least-squares post-processing runs in \(\mathcal{O}\!\big(M^3\,\mathrm{poly}(\log d)\big)\). It also introduces an adaptive cutoff rule that increments \(M\) and stops when successive reconstructions differ by less than a threshold [2509.02331].

These two lines of work share a common logic: the learning problem is made tractable by imposing the structural constraints already implied by periodic driving. In one case, the structure is the commutator hierarchy of the Floquet-Magnus expansion; in the other, it is the harmonic sparsity of a truncated Fourier representation.

## 5. Neural surrogates, unsupervised phase discovery, and periodic eigenproblems

Fourier Neural Operators extend Floquet-informed learning from reconstruction to surrogate modeling. In this formulation, the neural operator learns continuous maps between time-dependent functions, with Fourier layers
\[
u_{i+1}(x)=\sigma\!\left(\mathcal{F}^{-1}\!\big(R_i \cdot \mathcal{F}(u_i)\big)(x)+b_i(x)\cdot u_i(x)\right).
\]
The paper studies three tasks: \(H(t)\rightarrow H_F(t)\), \(H(t)\rightarrow \{\langle B_i(t)\rangle\}\), and \(H(t)\rightarrow c(t)\) for operator growth. Hamiltonians are encoded in a local Pauli basis, the output Floquet Hamiltonian is truncated to a 3-local basis, and the network is trained with \(M=3\) Fourier layers and \(k_{\max}=32\) retained modes. Reported results include relative error around \(2.4\times 10^{-3}\) for Floquet Hamiltonian learning, relative error below about \(0.8\%\) across the period in the benchmarked setting, good magnetization dynamics for hundreds of periods, and CPU inference times such as about \(0.728\) s exact versus \(0.0050\) s FNO for learning \(H_F\) at \(L=8\) [2509.07084].

Unsupervised Floquet phase identification uses a different representation of periodic structure. Samples are compared through the full intra-period unitary by the similarity
\[
S_{l,l'}=\frac{1}{N_d}\sum_d \frac{1}{2N_b}\operatorname{Tr}\!\left[U_{\varepsilon,l}^\dagger(d)U_{\varepsilon,l'}(d)+\text{h.c.}\right],
\]
then improved by symmetry-preserving adiabatic deformations, embedded by a diffusion map, and clustered with \(k\)-means. In the Floquet SSH model, the method recovers 4 phases when the 0 and \(\pi\) gaps are considered together; in the 2D bipartite model, it recovers 8 total Floquet topological phases, including anomalous Floquet insulators and an anomalous Floquet higher-order topological insulator [2106.11468]. A related unsupervised framework, designed for a periodically driven non-Hermitian lattice, uses amortized clustering and an algorithm selector over K-means++, hDBSCAN, Agglomerative clustering, and Mean-shift. Its input tasks are the \(2\times N\times N = 128\) eigenvectors of the finite lattice at each gain/loss value \(g\), and it classifies corner, edge, bulk, and hybrid corner-edge or edge-bulk Floquet modes in different non-Hermitian transition regimes [2508.00571].

The periodic-eigenproblem perspective also appears in a physics-informed neural solver for the Floquet-Bloch eigenproblem in 2D periodic potentials. There the networks \(\mathcal{N}_1(\mathbf{x},\mathbf{k};\theta)\) and \(\mathcal{N}_2(\mathbf{k};\phi)\) learn Bloch functions and band energies under the composite loss
\[
\mathcal{L} = \mathcal{L}_{\text{PDE} + \lambda_{\text{norm}\mathcal{L}_{\text{norm} + \lambda_{\text{BC}\mathcal{L}_{\text{BC},
\]
with \(\lambda_{\text{norm}}=100\) and \(\lambda_{\text{BC}}=10\). For the weak honeycomb potential \(V_0=1\), the reported final losses are total loss \(5.2\times 10^{-3}\), PDE residual \(3.3\times 10^{-3}\), normalization error \(3.0\times 10^{-7}\), and boundary-condition error \(1.9\times 10^{-4}\) [2512.21349]. That work concerns Floquet-Bloch periodic operators rather than temporal driving in the narrow sense, but it shows the same general pattern: periodic symmetry is built into the learning architecture rather than recovered afterward.

## 6. Quantum algorithms, mixed quantum–classical extensions, and recurring limitations

Floquet-ADAPT-VQE imports the same logic into hybrid quantum-classical simulation. The periodically driven problem is embedded into a time-independent eigenproblem for \(\hhat H_\mathrm{F}\), and the optimization target is the squared shifted Hamiltonian
\[
C_\lambda(\vec\theta)= \bra{\!\bra{\psi(\vec\theta)} \big(\hhat H_\mathrm{F}-\lambda \hhat I\big)^2 \ket{\!\ket{\psi(\vec\theta)}}.
\]
The ADAPT loop appends the operator with largest gradient and reoptimizes all parameters. A key design choice is the auxiliary initial state \( |\!\ket{\psi_\mathrm{in}} = |0\rangle \otimes |\phi_\mathrm{p}\rangle \), for which the initial squared Floquet energy is independent of both the number of auxiliary qubits and the drive frequency. The method also provides a fixed-depth circuit for time-dependent observables through single-qubit \(Z\)-rotations on the auxiliary register. The paper emphasizes, however, that finite auxiliary truncation makes the method approximate and that the number of Pauli terms in auxiliary operators grows as \(O(2^{n_a})\) [2503.11613].

In open-system nonadiabatic dynamics, the Floquet-informed algorithmic role is played by a master-equation reduction. The Floquet surface-hopping framework begins from a driven Anderson–Holstein impurity, derives a Floquet quantum master equation, takes a Wigner transform to obtain the Floquet classical master equation, and then solves it with three trajectory schemes: FSH, FaSH, and FaSH-density. The transition kernels are modified by the Floquet-dressed Fermi function \(\tilde f\), and the benchmark conclusion is that FaSH-density works best because it captures both fast oscillations due to the driving and correct steady-state observables [2303.00479].

A neighboring but non-ML computational development is the iterative algorithm for following individual Floquet states in high-dimensional Hilbert spaces. It uses repeated one-period propagation rather than diagonalization, supports maximum-likeness branch following under parameter changes, and was demonstrated for a periodically driven Bose-Hubbard chain with \(M=11\) sites, \(N=11\) bosons, unit filling, and Hilbert-space dimension \(d=352{,}716\). The drive-induced Mott-insulator-like target Floquet state can be populated with high efficiency when the amplitude is turned on smoothly but not too slowly, producing a window of pseudoadiabatic following [2111.12998].

Several recurrent limitations emerge across the literature. Some are representation-specific: low-order Floquet-Magnus truncations fail beyond a threshold, auxiliary-space truncations can require more resources at lower frequencies or larger amplitudes, and Fourier reconstructions depend on band limitation or sufficiently decaying harmonic content [2208.13837] [2503.11613] [2509.02331]. Others are methodological: unsupervised phase identification depends on dense enough sampling and optimization parameters, clustering-based mode classification is indirect and tied to the algorithm library, and neural solvers for periodic eigenproblems remain sensitive to loss weighting and degeneracies [2106.11468] [2508.00571] [2512.21349]. The cumulative picture nevertheless is coherent. A Floquet-informed learning algorithm is best understood not as a single method, but as a design principle: when temporal periodicity organizes the physics, the algorithm gains efficiency or robustness by representing that periodicity directly, whether through Floquet eigenstates, Floquet generators, Fourier harmonics, symmetry-constrained intra-period unitaries, or extended Floquet-Hilbert spaces.

Source: https://www.emergentmind.com/topics/floquet-informed-learning-algorithm