Hypernetwork-Modulated Neural Quantum States

• The paper demonstrates that hypernetwork modulation in RBMs efficiently reconstructs an entire family of ground states, achieving fidelities of up to 99% across phase diagrams.
• It uses FiLM-style transformations to dynamically generate RBM bias vectors as smooth functions of the Hamiltonian parameter, enabling direct extraction of physical response functions.
• This unified, differentiable framework offers scalable quantum state tomography applicable to various many-body systems, eliminating the need for retraining at each parameter point.

Hypernetwork-modulated neural quantum states integrate neural-network quantum state ansätze with hypernetworks to create compact, differentiable models for quantum many-body wavefunctions conditioned on external control parameters. This approach allows a single model to represent an entire family of ground states for parametrized Hamiltonians, achieving efficient quantum state tomography (QST) across continuous regions of phase diagrams without the need for point-wise retraining. The framework has been validated on paradigmatic quantum models, offering high-fidelity reconstructions and enabling direct extraction of phase transition diagnostics from measurement data (Tonner et al., 28 Jan 2026).

1. Foundations: Neural-Network Quantum State Parametrization

Neural-network quantum states (NQS) employ neural architectures to parameterize pure-state wavefunctions over discrete basis configurations. The Restricted Boltzmann Machine (RBM) defines a wavefunction $\Psi(\mathbf{s};\theta)$ for $N$ spins $\mathbf{s}\in\{0,1\}^N$ and $M$ binary hidden units $\mathbf{h}$ as

$$\Psi(\mathbf{s};\theta) = \sum_{\mathbf{h}} \exp\left(\sum_{i=1}^N a_i s_i + \sum_{j=1}^M b_j h_j + \sum_{i=1}^N \sum_{j=1}^M W_{ij} s_i h_j\right),$$

where $a_i$ and $b_j$ denote visible and hidden biases, respectively, and $W_{ij}$ are the RBM weights. In practical QST applications, the probability $p_\theta(\mathbf{s})$ is normalized and used to represent stoquastic ground states via $$\Psi_\theta(\mathbf{s})=\sqrt{p_\theta(\mathbf{s})}$$.

Classical RBM-based QST is limited to single problem instances: each ground state corresponding to a different Hamiltonian parameter must be reconstructed via independent training. This point-wise strategy leads to inefficiencies when quantum systems are studied over an extended parameter regime, such as during phase transition analysis.

2. Hypernetwork Architecture for Quantum State Families

To overcome this limitation, hypernetwork-modulated RBMs ("HyperRBMs") incorporate a hypernetwork $H_\phi$ that dynamically generates certain RBM parameters as smooth functions of a control parameter $g$ (e.g., the transverse field in the transverse-field Ising model).

Specifically, $H_\phi$ is a multilayer perceptron (MLP) mapping $g \in \mathbb{R}$ to feature-wise scale $\boldsymbol\gamma(g)$ and shift $\boldsymbol\beta(g)$ vectors. Only the RBM bias vectors are modulated: $$\begin{aligned} \mathbf{a}(g) &= [1+\boldsymbol\gamma^a(g)]\odot\mathbf{a}_{\rm base} + \boldsymbol\beta^a(g), \ \mathbf{b}(g) &= [1+\boldsymbol\gamma^b(g)]\odot\mathbf{b}_{\rm base} + \boldsymbol\beta^b(g), \end{aligned}$$ with weights $W$ shared across all $g$. The expressive power derives from FiLM-style ("feature-wise linear modulation") transformations, keeping the hypernetwork's output dimension low (${O}(N+M)$) and scalable. The parameter map is $\theta(g)=H_\phi(g)$, yielding the conditional family $\Psi_\theta(\mathbf{s}|g)$.

3. Training Objectives and Optimization Strategies

The model is trained for parametric QST using projective measurements from multiple values of the Hamiltonian parameter $g$. For each $g_k$, measurement datasets $d_k$ are collected. The loss function sums negative log-likelihoods across all support fields: $$\mathcal{L}(\phi) = -\sum_k \sum_{\mathbf{s}\in d_k} \log\, p_{\theta(g_k)}(\mathbf{s}).$$ Evaluation of this loss requires approximating the partition function's gradients, accomplished via Contrastive Divergence (CD-$k$); $k=10$ or $k=20$ are used for the negative phase. The mix of Gibbs chains used in negative phase sampling includes a small fraction of randomly initialized visible states to improve mixing. Parameters $$(W, \mathbf{a}_{\rm base}, \mathbf{b}_{\rm base}, \phi)$$ are optimized jointly using Adam with an inverse–sigmoid learning rate schedule (from $10^{-2}$ to $10^{-4}$).

Physical symmetries are imposed during training when needed; for example, a $\mathbb{Z}_2$ spin-flip symmetry in the ferromagnetic regime is enforced by a symmetrized free-energy ansatz: $$F_{\rm sym}(\mathbf{s}|g) = -\ln\Big(e^{-F_\theta(\mathbf{s}|g)}+e^{-F_\theta(1-\mathbf{s}|g)}\Big).$$

4. Physical Results: Fidelity, Response Functions, Entanglement

The HyperRBM's performance was established for the transverse-field Ising model on both 1D ($N=16$ sites) and 2D ($3\times3$, $4\times4$ lattices):

• State Fidelity: Overlaps between the reconstructed wavefunction $\Psi_\theta(\mathbf{s}|g)$ and ground states from Lanczos exact diagonalization (ED) yield fidelities $F(g)\gtrsim98\%$ with $2\!~000$ shots per support point, and $F(g) > 99\%$ with $20\!~000$ shots, across the phase diagram.
• Fidelity Susceptibility: The model produces a smooth, differentiable family, enabling direct calculation of the fidelity susceptibility

$$\chi_F(g) := -\left.\frac{\partial^2}{\partial \delta^2}\ln|\langle\Psi_\theta(g)|\Psi_\theta(g+\delta)\rangle|\right|_{\delta=0} \approx \frac{1}{4}\textrm{Var}_{\mathbf{s}\sim p_\theta(\cdot|g)}[\partial_g F_\theta(\mathbf{s}|g)].$$

Peaks in $\chi_F$ accurately identify the critical field $g_c$ without prior input.

• Rényi-2 Entropy Surfaces: For the 1D chain, second Rényi entropy $S_2(\ell,g)$ is computed via the swap trick and RBM free energies, matching ED results for subsystems up to $\ell\leq8$, with smooth $g$-dependence and correct asymptotics in ferromagnetic/paramagnetic phases.

5. Advantages, Scalability, and Generalizations

Hypernetwork-modulated NQS provide a unified, differentiable ansatz capable of interpolating smoothly over an entire parameter regime, bypassing the need for retraining at every $g$. Key methodological advantages include:

• Low-dimensional hypernetwork modulation (scaling as $O(N+M)$), enabling practical training for sizable systems.
• Direct computation of physical response functions and entanglement diagnostics through $g$ gradients, supporting phase transition studies.
• Scalability to larger systems is feasible by replacing ED ground-truth data with QMC or tensor-network samples. The framework naturally generalizes to non-stoquastic models with amplitude + phase parametrizations.
• Extension to autoregressive or convolutional NQS is a potential avenue to enhance sampling efficiency and stability in high-dimensional Hilbert spaces.

6. Implications for Quantum Tomography and Phase Diagram Studies

The development of HyperRBMs establishes that a single neural-network ansatz can reconstruct quantum ground states and their phase transitions throughout the entire phase diagram from tomographic data, in contrast to conventional QST methods limited by exponential scaling and per-point retraining. The differentiability in the control parameter $g$ enables a suite of physical analyses—including fidelity susceptibility and entanglement entropy surfaces—within a single, unified framework. This suggests broad applicability to quantum device validation, quantum simulation, and many-body physics, particularly as data modalities and system sizes grow.

7. Outlook and Future Directions

Potential future developments include the use of QMC and tensor-network data for training on larger systems, the application to non-stoquastic and complex-valued quantum states via generalized RBMs, and integration with autoregressive or convolutional architectures to further improve scalability and performance in quantum state tomography and phase diagram diagnostics (Tonner et al., 28 Jan 2026).