Semi-Quantum Restricted Boltzmann Machine

• Semi-Quantum RBMs are hybrid generative models that combine classical visible layers with quantum hidden layers to enhance expressivity and enable tractable optimization.
• They employ analytic techniques such as closed-form gradients and efficient sampling through visible layer commutativity, significantly reducing quantum resource requirements.
• Empirical benchmarks and reinforcement learning tests demonstrate that sqRBMs offer improved performance and stability while optimizing quantum hardware usage.

A Semi-Quantum Restricted Boltzmann Machine (sqRBM) is a hybrid generative model combining classical and quantum statistical mechanics within the bipartite architecture of the classical restricted Boltzmann machine (RBM), typically by preserving classical visible units and encoding quantum structure exclusively within the hidden layer. The principal motivations for this architecture are tractable optimization, improved expressivity compared to classical RBMs, and significant reductions in the quantum hardware resources required for physical implementation or quantum-inspired algorithms.

1. Model Architecture and Hamiltonian Structure

An sqRBM consists of a visible layer $V$ of classical variables and a hidden layer $H$ comprising quantum modes, which may be realized as qubits or fermionic modes depending on the model instantiation. The bipartite couplings between $V$ and $H$ generalize those in the classical RBM to include quantum operators.

Typical Hamiltonian forms include:

• Fermionic sqRBM (Lyakhova et al., 2020):

$$\hat H(v) = -\sum_{i=1}^n b_i v_i - \sum_{j,j'=1}^m c_{jj'}\,\hat h_j^\dagger \hat h_{j'} - \sum_{i=1}^{n} \sum_{j,j'=1}^m v_i w_{ijj'} \hat h_j^\dagger \hat h_{j'}$$

with $v_i\in\{0,1\}$ (classical), and $\hat h_j, \hat h_j^\dagger$ fermionic operators.

• Transverse-field sqRBM (Kimura et al., 29 Jul 2025):

$$H(\theta) = -\sum_{i} b_i\,\sigma_i^z - \sum_j (b_j\,\sigma_j^z + \Gamma_j\,\sigma_j^x) - \sum_{i,j} w_{ij}\,\sigma_i^z\,\sigma_j^z$$

with classical visible spins and quantum hidden qubits.

• Commuting-visible, non-commuting-hidden sqRBM (Demidik et al., 24 Feb 2025):

$H = H_v + H_h + H_{\text{int}}$

where $H_v$ involves only commuting $H$0 operators on visibles, $H$1 and $H$2 act non-commutatively on the hidden space.

A crucial property in these constructions is the commutativity in the visible subspace, enabling each visible configuration to induce a tractable "clamped" quantum Hamiltonian over the hidden sector, which can be diagonalized or sampled efficiently.

2. Exact Marginalization and Output Probability

For a given semi-quantum Hamiltonian, the marginal probability over visible configurations admits closed-form expressions due to the block-diagonal structure imposed by classical visibles: $H$3 where $H$4 traces over the quantum hidden layer and $H$5 is the partition function, typically

$H$6

Specifically, in models with non-interacting quantum hidden elements, "clamped" single-particle Hamiltonians can be diagonalized:

• Fermionic models yield conditional partition function

$H$7

• Qubit-based models yield per-hidden-unit factors such as

$H$8

where $H$9 is the hidden field vector parameterizing the qubit Hamiltonian (Demidik et al., 24 Feb 2025).

This analytic tractability contrasts with fully quantum RBMs, which feature positive phases with non-commuting operators and thus require costly quantum sampling or approximation.

3. Parameter Learning and Algorithmic Frameworks

A principal advantage of sqRBMs is the existence of closed-form gradients for maximum likelihood (or similar) training objectives, as the partition functions and their parameter derivatives are analytically accessible.

• Gradient expressions (fermionic sqRBM) (Lyakhova et al., 2020):

$V$0

$V$1

with $V$2 the Fermi–Dirac density matrix.

• EM-style Information-Geometric Training (sqRBM qubit model) (Kimura et al., 29 Jul 2025):
  • The E-step computes expectations over data; the M-step minimizes a convex combination of energy and partition function, guaranteeing monotonic KL divergence reduction and improved stability relative to plain gradient descent.
• Persistent Contrastive Divergence and Markov Chain Monte Carlo are common for models with intractable model marginals, with quantum steps necessary only for the hidden layer.

4. Expressivity, Resource Scaling, and Quantum Advantage

• Expressive equivalence results (Demidik et al., 24 Feb 2025) establish that, for a fully non-commuting hidden sector, an sqRBM with $V$3 quantum hidden units is as expressive as a classical RBM with $V$4 classical hidden units. Both models have identical parameter counts, but the sqRBM achieves the same representational power with one third as many hidden units, resulting in significant quantum resource reduction.
• Quantum generalization: Off-diagonal and non-commuting hidden couplings introduce quantum coherence and entanglement within the hidden layer, yielding richer model classes:
  • Fermionic sqRBMs with non-diagonal hidden couplings capture distributions not accessible to classical RBMs of the same size (Lyakhova et al., 2020).
  • Qubit-based sqRBMs with Hamiltonians commuting in the visible space avoid positive-phase intractability while enabling analytic gradients (Demidik et al., 24 Feb 2025).

This suggests that sqRBMs are optimal for near-term quantum hardware deployment, balancing tractable training, expressivity, and quantum resource demands.

5. Semi-Quantum Variants and Hybridization Strategies

Several variants and generalizations of the sqRBM paradigm have been developed:

• Feature-space semi-quantum RBMs (Wu et al., 2019): Parts of the model (e.g., parameter sets, bias terms) remain classical, while quantum routines are invoked only for expensive overlaps or feature summations, leveraging shallow circuits suitable for NISQ devices. Classical optimization (SPSA, MCMC) calls quantum subroutines, yielding exponential speedup in feature-space dimension for selected tasks.
• Continuous semi-quantum Boltzmann machines (CSQBMs) (Gerlach et al., 6 Nov 2025): The visible units are continuous and classical (exponential-family variables), while the hidden units comprise quantum qubits. CSQBMs admit analytic gradients, support both classical and quantum sampling phases, and enable direct integration into continuous-action Actor-Critic and Q-learning reinforcement learning frameworks with substantial savings in qubit count and preserved expressivity.

6. Applications, Empirical Results, and Benchmarking

sqRBMs have been empirically evaluated on a variety of generative modeling benchmarks. Key findings include:

• Empirical advantage:

On Bars & Stripes and Optdigits, sqRBMs surpass classical RBMs with equivalent hidden-layer cardinality and even outperform classical RBMs with hidden-layer sizes up to $V$5 for $V$6 quantum modes (Lyakhova et al., 2020).

• Simulation scaling:

sqRBMs achieve comparable or better total variation distance on a suite of structured and random-support datasets with fewer hidden units than classical RBMs. Expressive equivalence ratios ($V$7 or $V$8 hidden units for sqRBM:RBM) are confirmed numerically for up to 100 units (Demidik et al., 24 Feb 2025).

• Information-geometric EM:

Information-geometric EM on sqRBMs attains lower or comparable KL divergence to standard gradient descent, with increased stability and reduced variance across random initializations except in data distributions with maximal symmetry (e.g., cardinality) (Kimura et al., 29 Jul 2025).

• Reinforcement learning:

CSQBMs yield sample-efficient continuous-action control, stabilizing Q-learning by leveraging efficient quantum Gibbs sampling in the hidden layer (Gerlach et al., 6 Nov 2025).

7. Theoretical and Practical Implications for Quantum Machine Learning

sqRBMs are positioned as scalable quantum-enhanced generative models compatible with both classical and near-term quantum computational resources:

• Tractability: Closed-form gradients and analytically computable positive phases enable scalable optimization, circumnavigating the "barren plateau" and vanishing gradient problems that afflict fully quantum Boltzmann machines.
• Quantum resource minimization: By confining quantum structure to the hidden sector and commuting in the visible subspace, sqRBMs require fewer qubits, shallower circuits, and minimal entanglement between visible and hidden layers, directly benefiting NISQ hardware and hybrid quantum-classical algorithm design.
• Versatility: The sqRBM framework accommodates a range of learning paradigms, including classical Bayesian inference with quantum subroutines (Wu et al., 2019), quantum EM optimization (Kimura et al., 29 Jul 2025), and quantum-inspired continuous-variable models (Gerlach et al., 6 Nov 2025).

Practical realization depends on the scalability of quantum hidden-layer diagonalization or sampling, circuit depth control, and the effectiveness of hybrid learning schemes. The analytic tractability and proven empirical and expressive advantages of sqRBMs position them as a promising architecture for generative modeling and reinforcement learning tasks in quantum machine learning.