SymQNet: Fast Adaptive Hamiltonian Learning

• SymQNet is a learned, amortized acquisition policy that shifts Bayesian design from online optimization to offline reinforcement learning, reducing computation time.
• It leverages a graph encoder, VAE for history compression, and transformer layers to effectively capture spatial and temporal features in the posterior.
• Empirical evaluations on transverse-field Ising model benchmarks demonstrate significant latency reductions of up to 83.7× while maintaining competitive parameter accuracy.

Searching arXiv for the SymQNet paper and closely related context papers. SymQNet is a learned, amortized acquisition policy for adaptive Hamiltonian learning, introduced to reduce the decision latency of Bayesian experimental design in quantum-device calibration and characterization (Tomar et al., 11 Jun 2026). Its central premise is to shift the computational burden of experiment selection from online posterior-conditioned search to offline reinforcement learning: a policy is trained across simulated Hamiltonian-learning episodes, then deployed online as a fast forward pass while retaining full Bayesian posterior feedback. In the setting studied, SymQNet targets one-dimensional transverse-field Ising model benchmarks, where repeated recomputation of Fisher-information or BALD-style design utilities can take seconds per adaptive step and thereby dominate wall-clock time.

1. Adaptive Hamiltonian learning and the latency problem

Adaptive Hamiltonian learning repeatedly interleaves experiment design, measurement, posterior updating, and redesign. After each shot or batch of shots, the controller uses the current posterior over Hamiltonian parameters to choose the next experiment. In principle, this posterior-aware loop improves estimation efficiency relative to fixed schedules because future measurements can be targeted toward regions of maximal uncertainty or expected information gain. In practice, the acquisition computation is itself a nontrivial inference-time workload.

The bottleneck addressed by SymQNet is the online recomputation of a Bayesian design rule after every posterior update. Standard utilities such as expected Fisher information, mutual information, or expected posterior-entropy reduction must be evaluated conditional on the current posterior, typically over multiple candidate experiments and, for look-ahead methods, over hypothetical future outcomes. Across hundreds of shots, these seconds-per-decision costs accumulate into a substantial wall-clock burden. SymQNet is designed specifically for this regime: it leaves Bayesian filtering in place, but amortizes the acquisition step by learning the map from posterior state to experimental decision offline (Tomar et al., 11 Jun 2026).

A common misconception is that a learned acquisition policy replaces Bayesian adaptivity with a purely heuristic controller. In SymQNet, that is not the case. The posterior remains explicit and is updated by Sequential Monte Carlo (SMC) after each observation; what is replaced is the online search or optimization used to score candidate experiments. The method therefore belongs to posterior-aware adaptive design rather than to fixed open-loop scheduling.

2. Formal problem specification

The paper studies an $N$-qubit one-dimensional transverse-field Ising model (TFIM) with Hamiltonian

$$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$

where the parameter vector is

$\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$

so the parameter dimension is $d=2N-1$ (Tomar et al., 11 Jun 2026). The prior is factorized and uniform:

$J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$

This defines a positive-coupling TFIM chain.

An acquisition decision at time $t$ is a discrete action

$a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$

where $i_t\in\{1,\dots,N\}$ is a qubit index, $\sigma_t\in\{X,Y,Z\}$ is a single-qubit Pauli measurement basis, and $\tau_t$ is one of $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$0 discrete evolution times. Accordingly,

$$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$1

Given $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$2 and $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$3, the observation model is

$$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$4

with measurement probabilities obtained from Schrödinger evolution under $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$5 followed by projective measurement and $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$6 symmetric readout-flip noise.

The Bayesian state is the posterior

$$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$7

for data $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$8. Because the model is nonlinear and high-dimensional, SymQNet uses SMC with $$H(\theta)=\sum_{i=1}^{N-1} J_i Z_i Z_{i+1}+\sum_{i=1}^{N} h_i X_i,$$9 particles. If the posterior approximation is written as $\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$0, then the weight update is

$\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$1

followed by normalization and systematic resampling when the effective sample size drops below $\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$2.

The adaptive learning process is cast as a finite-horizon belief-state MDP with state

$\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$3

where $\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$4 is the SMC posterior, $\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$5 is the measurement history, $\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$6 is the chain graph, and $\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$7 is the time index. The reward is the dense information gain between successive posteriors,

$\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$8

and the policy objective is

$\theta=(J_1,\ldots,J_{N-1},h_1,\ldots,h_N),$9

with horizon $d=2N-1$0.

3. Online Bayesian baselines and their computational burden

SymQNet is evaluated against two bounded online acquisition rules that instantiate standard Bayesian design principles while capping search complexity (Tomar et al., 11 Jun 2026). The first is bounded Fisher-information search. Its conceptual target is to maximize Fisher information for the next experiment, using a particle-linearized Fisher objective and a bounded candidate set of four actions per decision. Even in this restricted form, each step requires repeated likelihood or gradient-related computations across posterior particles.

The second is bounded two-step BALD, where BALD denotes the mutual information

$d=2N-1$1

The two-step variant introduces a short look-ahead tree. Because exact two-step BALD is too expensive, the bounded version scores up to three present candidates, retains two candidates for one-step look-ahead, and uses one predictive observation sample to approximate the future expectation. This still requires posterior prediction under candidate actions, hypothetical posterior updates, and repeated evaluation of information-theoretic quantities.

These baselines clarify the distinction between posterior update cost and acquisition cost. All methods share the same SMC filtering machinery; the extra latency arises from the design rule itself. SymQNet’s contribution is therefore not a faster Bayesian update, but the elimination of repeated online candidate scoring and explicit look-ahead planning. This makes the method especially relevant when the same adaptive workload is executed repeatedly, as in recalibration or drift-tracking regimes.

4. SymQNet architecture and amortized acquisition

SymQNet learns a stochastic policy $d=2N-1$2 with PPO, using offline simulation over many Hamiltonian-learning episodes (Tomar et al., 11 Jun 2026). Each episode samples a new $d=2N-1$3 from the task prior, simulates experiments with either a statevector backend for small $d=2N-1$4 or a matrix-product-state backend for larger $d=2N-1$5, injects $d=2N-1$6 symmetric flip noise, and updates the posterior with SMC. The policy is trained as a black-box controller: the simulator and SMC are not differentiated through.

The architectural design combines three forms of structure. First, a graph encoder operates on the chain graph $d=2N-1$7, taking qubit-local and possibly coupling-local posterior features as input. This exploits the spatial organization of the TFIM and supports representations that are permutation-equivariant with respect to relabelings consistent with the chain structure. Second, a VAE compresses the measurement history $d=2N-1$8 into a lower-dimensional latent representation, reflecting the fact that the sequential history can be long. Third, transformer layers summarize temporal structure in the sequence. These components feed a policy head that outputs logits over the full discrete action set $d=2N-1$9, defining

$J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$0

The architecture is described at a high level rather than layer-by-layer. The paper explicitly identifies graph layers, VAE-based history compression, transformer-based sequence processing, and a categorical action output, but does not enumerate hidden sizes or exact depth in the excerpt. This suggests that the central design claim is structural rather than micro-architectural: SymQNet is intended to encode posterior summaries, history, and graph symmetries in a form suitable for fast action selection.

The “Sym” in the name refers to this symmetry-aware encoding. By reusing local computations across sites and couplings on the chain graph, the policy need not learn a separate rule for every qubit index. A plausible implication is improved generalization across parameter draws and system sizes. The ablation results support the relevance of these components: removing the graph encoder causes a $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$1 increase in MSE, removing the VAE causes a $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$2 increase, and replacing the transformer with an MLP slightly improves 5-qubit MSE by $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$3, which the paper associates with possible overcapacity on small instances.

5. Training regime, deployment flow, and empirical results

Training uses PPO with the KL-based reward defined above (Tomar et al., 11 Jun 2026). For five-qubit policies, the reported setting is 2500 PPO updates with 64 rollout steps per update. For larger-$J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$4 scaling policies, training is reduced to 300 PPO updates with the same rollout horizon to keep training feasible. Final parameter quality is evaluated separately from the training reward using

$J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$5

where $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$6 is the SMC posterior mean.

At deployment, the online loop is simple. Starting from the prior $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$7, the controller forms the state $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$8, computes an action from the learned policy, executes the corresponding experiment on the device, records the outcome, updates the SMC posterior, appends the event to history, and repeats. The acquisition timer reported in the paper begins after the simulator and SMC state are available and ends when the acquisition rule outputs the action, so the measured decision latency isolates the policy computation rather than full experimental execution.

On the 5-qubit benchmark with 128 shots, SymQNet reduces acquisition-only decision latency by $J_i,h_i \sim \mathrm{Uniform}(0.5,1.5).$9 relative to bounded Fisher-information search and by $t$0 relative to bounded two-step BALD. The corresponding MSE ratios, defined as SymQNet MSE divided by baseline MSE, are $t$1 against bounded Fisher-information search and $t$2 against bounded two-step BALD. At 512 shots, the speedups are $t$3 and $t$4, with MSE ratios $t$5 and $t$6, respectively. These figures establish the paper’s main trade-off on small matched-prior problems: substantial latency reduction with some loss in final parameter accuracy.

The scaling results on 8-, 10-, and 12-qubit tasks are more favorable to SymQNet. At 12 qubits, the policy’s decision time is $t$7 ms per decision, while bounded two-step BALD requires $t$8 s per decision and bounded Fisher-information search yields a full simulated step of $t$9 s. SymQNet’s full simulated step is $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$0 s, compared with $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$1 s for bounded two-step BALD. Against bounded two-step BALD, the reported latency log-slope is $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$2 for SymQNet versus $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$3 for BALD, with latency slope ratio $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$4. The worst MSE ratio of SymQNet over BALD in this scaling regime is $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$5, meaning that SymQNet is sometimes slightly more accurate. The paper therefore portrays the large-$a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$6 regime as one in which learned amortized acquisition can approach baseline accuracy while dramatically lowering decision latency.

6. Interpretation, limitations, and broader significance

SymQNet’s principal significance lies in reframing adaptive Hamiltonian learning as an amortization problem (Tomar et al., 11 Jun 2026). Rather than solving a fresh Bayesian design problem after every posterior update, the method learns a posterior-conditioned acquisition rule offline and reuses it online. This preserves Bayesian feedback through SMC while decoupling the adaptive policy from expensive runtime optimization. For workloads involving repeated low-latency decision making, this is the paper’s central contribution.

The results also delimit the scope of the method. The study is confined to a 1D TFIM with positive couplings and a fixed prior range. The noise model includes only $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$7 symmetric readout-flip noise. Training and evaluation use matched priors and model class, which is an idealized setting for deployment. More realistic deviations—long-range couplings, higher-dimensional graphs, disorder, relaxation, dephasing, crosstalk, or prior mismatch—are identified as future work rather than established capabilities.

A second limitation concerns objective mismatch. SymQNet is trained on cumulative information gain,

$a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$8

but evaluated with final MSE. The reported Pearson correlation between cumulative information gain and final MSE ranges from $a_t=(i_t,\sigma_t,\tau_t)\in\mathcal{A}_N,$9 to $i_t\in\{1,\dots,N\}$0, indicating weak alignment. This is important for interpretation: high information-gain behavior does not automatically imply best final parameter accuracy under the chosen estimator. In the 5-qubit matched-prior benchmark, optimized fixed schedules and a DAD-style transformer baseline remain the most accurate policies, and SymQNet is positioned primarily as a latency-improving method for posterior-aware adaptive control rather than as a uniformly best estimator.

These caveats help resolve another common misconception, namely that SymQNet is chiefly an accuracy improvement over all alternatives. The evidence does not support that generalization. Its distinguishing property is the reduction of acquisition latency while keeping Bayesian posterior updates in the loop and maintaining competitive accuracy, especially as system size grows. This suggests a practical niche in settings where the benefits of adaptivity would otherwise be negated by the time required to choose the next experiment.